]>
Commit | Line | Data |
---|---|---|
b36ba79f | 1 | /* Output routines for GCC for ARM. |
f954388e | 2 | Copyright (C) 1991, 1993, 1994, 1995, 1996, 1997, 1998, 1999, 2000, 2001, |
bf98ec6c | 3 | 2002, 2003, 2004, 2005, 2006, 2007, 2008 Free Software Foundation, Inc. |
cce8749e | 4 | Contributed by Pieter `Tiggr' Schoenmakers (rcpieter@win.tue.nl) |
956d6950 | 5 | and Martin Simmons (@harleqn.co.uk). |
b36ba79f | 6 | More major hacks by Richard Earnshaw (rearnsha@arm.com). |
cce8749e | 7 | |
4f448245 | 8 | This file is part of GCC. |
cce8749e | 9 | |
4f448245 NC |
10 | GCC is free software; you can redistribute it and/or modify it |
11 | under the terms of the GNU General Public License as published | |
2f83c7d6 | 12 | by the Free Software Foundation; either version 3, or (at your |
4f448245 | 13 | option) any later version. |
cce8749e | 14 | |
4f448245 NC |
15 | GCC is distributed in the hope that it will be useful, but WITHOUT |
16 | ANY WARRANTY; without even the implied warranty of MERCHANTABILITY | |
17 | or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public | |
18 | License for more details. | |
cce8749e | 19 | |
4f448245 | 20 | You should have received a copy of the GNU General Public License |
2f83c7d6 NC |
21 | along with GCC; see the file COPYING3. If not see |
22 | <http://www.gnu.org/licenses/>. */ | |
f676971a | 23 | |
56636818 | 24 | #include "config.h" |
43cffd11 | 25 | #include "system.h" |
4977bab6 ZW |
26 | #include "coretypes.h" |
27 | #include "tm.h" | |
cce8749e | 28 | #include "rtl.h" |
d5b7b3ae | 29 | #include "tree.h" |
c7319d87 | 30 | #include "obstack.h" |
cce8749e CH |
31 | #include "regs.h" |
32 | #include "hard-reg-set.h" | |
33 | #include "real.h" | |
34 | #include "insn-config.h" | |
35 | #include "conditions.h" | |
cce8749e CH |
36 | #include "output.h" |
37 | #include "insn-attr.h" | |
38 | #include "flags.h" | |
af48348a | 39 | #include "reload.h" |
49ad7cfa | 40 | #include "function.h" |
bee06f3d | 41 | #include "expr.h" |
e78d8e51 | 42 | #include "optabs.h" |
ad076f4e | 43 | #include "toplev.h" |
aec3cfba | 44 | #include "recog.h" |
92a432f4 | 45 | #include "ggc.h" |
d5b7b3ae | 46 | #include "except.h" |
8b97c5f8 | 47 | #include "c-pragma.h" |
7b8b8ade | 48 | #include "integrate.h" |
c27ba912 | 49 | #include "tm_p.h" |
672a6f42 NB |
50 | #include "target.h" |
51 | #include "target-def.h" | |
980e61bb | 52 | #include "debug.h" |
6e34d3a3 | 53 | #include "langhooks.h" |
6fb5fa3c | 54 | #include "df.h" |
cce8749e | 55 | |
d5b7b3ae RE |
56 | /* Forward definitions of types. */ |
57 | typedef struct minipool_node Mnode; | |
58 | typedef struct minipool_fixup Mfix; | |
59 | ||
1d6e90ac NC |
60 | const struct attribute_spec arm_attribute_table[]; |
61 | ||
b76c3c4b PB |
62 | void (*arm_lang_output_object_attributes_hook)(void); |
63 | ||
d5b7b3ae | 64 | /* Forward function declarations. */ |
35596784 | 65 | static int arm_compute_static_chain_stack_bytes (void); |
5848830f | 66 | static arm_stack_offsets *arm_get_frame_offsets (void); |
e32bac5b | 67 | static void arm_add_gc_roots (void); |
a406f566 MM |
68 | static int arm_gen_constant (enum rtx_code, enum machine_mode, rtx, |
69 | HOST_WIDE_INT, rtx, rtx, int, int); | |
e32bac5b RE |
70 | static unsigned bit_count (unsigned long); |
71 | static int arm_address_register_rtx_p (rtx, int); | |
1e1ab407 | 72 | static int arm_legitimate_index_p (enum machine_mode, rtx, RTX_CODE, int); |
5b3e6663 PB |
73 | static int thumb2_legitimate_index_p (enum machine_mode, rtx, int); |
74 | static int thumb1_base_register_rtx_p (rtx, enum machine_mode, int); | |
75 | inline static int thumb1_index_register_rtx_p (rtx, int); | |
5848830f | 76 | static int thumb_far_jump_used_p (void); |
57934c39 | 77 | static bool thumb_force_lr_save (void); |
e32bac5b | 78 | static int const_ok_for_op (HOST_WIDE_INT, enum rtx_code); |
e32bac5b | 79 | static rtx emit_sfm (int, int); |
466e4b7a | 80 | static unsigned arm_size_return_regs (void); |
e32bac5b | 81 | static bool arm_assemble_integer (rtx, unsigned int, int); |
e32bac5b RE |
82 | static const char *fp_const_from_val (REAL_VALUE_TYPE *); |
83 | static arm_cc get_arm_condition_code (rtx); | |
e32bac5b RE |
84 | static HOST_WIDE_INT int_log2 (HOST_WIDE_INT); |
85 | static rtx is_jump_table (rtx); | |
86 | static const char *output_multi_immediate (rtx *, const char *, const char *, | |
87 | int, HOST_WIDE_INT); | |
e32bac5b RE |
88 | static const char *shift_op (rtx, HOST_WIDE_INT *); |
89 | static struct machine_function *arm_init_machine_status (void); | |
c9ca9b88 | 90 | static void thumb_exit (FILE *, int); |
e32bac5b RE |
91 | static rtx is_jump_table (rtx); |
92 | static HOST_WIDE_INT get_jump_table_size (rtx); | |
93 | static Mnode *move_minipool_fix_forward_ref (Mnode *, Mnode *, HOST_WIDE_INT); | |
94 | static Mnode *add_minipool_forward_ref (Mfix *); | |
95 | static Mnode *move_minipool_fix_backward_ref (Mnode *, Mnode *, HOST_WIDE_INT); | |
96 | static Mnode *add_minipool_backward_ref (Mfix *); | |
97 | static void assign_minipool_offsets (Mfix *); | |
98 | static void arm_print_value (FILE *, rtx); | |
99 | static void dump_minipool (rtx); | |
100 | static int arm_barrier_cost (rtx); | |
101 | static Mfix *create_fix_barrier (Mfix *, HOST_WIDE_INT); | |
102 | static void push_minipool_barrier (rtx, HOST_WIDE_INT); | |
103 | static void push_minipool_fix (rtx, HOST_WIDE_INT, rtx *, enum machine_mode, | |
104 | rtx); | |
105 | static void arm_reorg (void); | |
106 | static bool note_invalid_constants (rtx, HOST_WIDE_INT, int); | |
e32bac5b RE |
107 | static unsigned long arm_compute_save_reg0_reg12_mask (void); |
108 | static unsigned long arm_compute_save_reg_mask (void); | |
109 | static unsigned long arm_isr_value (tree); | |
110 | static unsigned long arm_compute_func_type (void); | |
111 | static tree arm_handle_fndecl_attribute (tree *, tree, tree, int, bool *); | |
112 | static tree arm_handle_isr_attribute (tree *, tree, tree, int, bool *); | |
7bff66a7 | 113 | #if TARGET_DLLIMPORT_DECL_ATTRIBUTES |
04fb56d5 | 114 | static tree arm_handle_notshared_attribute (tree *, tree, tree, int, bool *); |
7bff66a7 | 115 | #endif |
e32bac5b RE |
116 | static void arm_output_function_epilogue (FILE *, HOST_WIDE_INT); |
117 | static void arm_output_function_prologue (FILE *, HOST_WIDE_INT); | |
5b3e6663 | 118 | static void thumb1_output_function_prologue (FILE *, HOST_WIDE_INT); |
3101faab | 119 | static int arm_comp_type_attributes (const_tree, const_tree); |
e32bac5b RE |
120 | static void arm_set_default_type_attributes (tree); |
121 | static int arm_adjust_cost (rtx, rtx, rtx, int); | |
e32bac5b RE |
122 | static int count_insns_for_constant (HOST_WIDE_INT, int); |
123 | static int arm_get_strip_length (int); | |
124 | static bool arm_function_ok_for_sibcall (tree, tree); | |
125 | static void arm_internal_label (FILE *, const char *, unsigned long); | |
126 | static void arm_output_mi_thunk (FILE *, tree, HOST_WIDE_INT, HOST_WIDE_INT, | |
127 | tree); | |
128 | static int arm_rtx_costs_1 (rtx, enum rtx_code, enum rtx_code); | |
21b5653c | 129 | static bool arm_size_rtx_costs (rtx, int, int, int *); |
9b66ebb1 PB |
130 | static bool arm_slowmul_rtx_costs (rtx, int, int, int *); |
131 | static bool arm_fastmul_rtx_costs (rtx, int, int, int *); | |
132 | static bool arm_xscale_rtx_costs (rtx, int, int, int *); | |
133 | static bool arm_9e_rtx_costs (rtx, int, int, int *); | |
e32bac5b RE |
134 | static int arm_address_cost (rtx); |
135 | static bool arm_memory_load_p (rtx); | |
136 | static bool arm_cirrus_insn_p (rtx); | |
137 | static void cirrus_reorg (rtx); | |
5a9335ef NC |
138 | static void arm_init_builtins (void); |
139 | static rtx arm_expand_builtin (tree, rtx, rtx, enum machine_mode, int); | |
140 | static void arm_init_iwmmxt_builtins (void); | |
141 | static rtx safe_vector_operand (rtx, enum machine_mode); | |
142 | static rtx arm_expand_binop_builtin (enum insn_code, tree, rtx); | |
143 | static rtx arm_expand_unop_builtin (enum insn_code, tree, rtx, int); | |
144 | static rtx arm_expand_builtin (tree, rtx, rtx, enum machine_mode, int); | |
a406f566 | 145 | static void emit_constant_insn (rtx cond, rtx pattern); |
d66437c5 | 146 | static rtx emit_set_insn (rtx, rtx); |
78a52f11 RH |
147 | static int arm_arg_partial_bytes (CUMULATIVE_ARGS *, enum machine_mode, |
148 | tree, bool); | |
5a9335ef | 149 | |
7abc66b1 | 150 | #ifdef OBJECT_FORMAT_ELF |
9403b7f7 RS |
151 | static void arm_elf_asm_constructor (rtx, int) ATTRIBUTE_UNUSED; |
152 | static void arm_elf_asm_destructor (rtx, int) ATTRIBUTE_UNUSED; | |
7abc66b1 | 153 | #endif |
fb49053f | 154 | #ifndef ARM_PE |
e32bac5b | 155 | static void arm_encode_section_info (tree, rtx, int); |
fb49053f | 156 | #endif |
b12a00f1 RE |
157 | |
158 | static void arm_file_end (void); | |
6c6aa1af | 159 | static void arm_file_start (void); |
b12a00f1 | 160 | |
1cc9f5f5 KH |
161 | static void arm_setup_incoming_varargs (CUMULATIVE_ARGS *, enum machine_mode, |
162 | tree, int *, int); | |
8cd5a4e0 | 163 | static bool arm_pass_by_reference (CUMULATIVE_ARGS *, |
586de218 KG |
164 | enum machine_mode, const_tree, bool); |
165 | static bool arm_promote_prototypes (const_tree); | |
6b045785 | 166 | static bool arm_default_short_enums (void); |
13c1cd82 | 167 | static bool arm_align_anon_bitfield (void); |
586de218 KG |
168 | static bool arm_return_in_msb (const_tree); |
169 | static bool arm_must_pass_in_stack (enum machine_mode, const_tree); | |
23668cf7 | 170 | static bool arm_return_in_memory (const_tree, const_tree); |
617a1b71 PB |
171 | #ifdef TARGET_UNWIND_INFO |
172 | static void arm_unwind_emit (FILE *, rtx); | |
173 | static bool arm_output_ttype (rtx); | |
174 | #endif | |
5b3e6663 | 175 | static void arm_dwarf_handle_frame_unspec (const char *, rtx, int); |
c237e94a | 176 | |
4185ae53 PB |
177 | static tree arm_cxx_guard_type (void); |
178 | static bool arm_cxx_guard_mask_bit (void); | |
46e995e0 PB |
179 | static tree arm_get_cookie_size (tree); |
180 | static bool arm_cookie_has_size (void); | |
44d10c10 | 181 | static bool arm_cxx_cdtor_returns_this (void); |
505970fc | 182 | static bool arm_cxx_key_method_may_be_inline (void); |
1e731102 MM |
183 | static void arm_cxx_determine_class_data_visibility (tree); |
184 | static bool arm_cxx_class_data_always_comdat (void); | |
9f62c3e3 | 185 | static bool arm_cxx_use_aeabi_atexit (void); |
b3f8d95d | 186 | static void arm_init_libfuncs (void); |
c54c7322 | 187 | static bool arm_handle_option (size_t, const char *, int); |
67e6ba46 | 188 | static void arm_target_help (void); |
273a2526 | 189 | static unsigned HOST_WIDE_INT arm_shift_truncation_mask (enum machine_mode); |
d3585b76 DJ |
190 | static bool arm_cannot_copy_insn_p (rtx); |
191 | static bool arm_tls_symbol_p (rtx x); | |
bd4dc3cd | 192 | static int arm_issue_rate (void); |
afcc986d | 193 | static void arm_output_dwarf_dtprel (FILE *, int, rtx) ATTRIBUTE_UNUSED; |
007e61c2 | 194 | static bool arm_allocate_stack_slots_for_args (void); |
d3585b76 | 195 | |
672a6f42 NB |
196 | \f |
197 | /* Initialize the GCC target structure. */ | |
b2ca3702 | 198 | #if TARGET_DLLIMPORT_DECL_ATTRIBUTES |
1d6e90ac | 199 | #undef TARGET_MERGE_DECL_ATTRIBUTES |
672a6f42 NB |
200 | #define TARGET_MERGE_DECL_ATTRIBUTES merge_dllimport_decl_attributes |
201 | #endif | |
f3bb6135 | 202 | |
1d6e90ac | 203 | #undef TARGET_ATTRIBUTE_TABLE |
91d231cb | 204 | #define TARGET_ATTRIBUTE_TABLE arm_attribute_table |
672a6f42 | 205 | |
6c6aa1af PB |
206 | #undef TARGET_ASM_FILE_START |
207 | #define TARGET_ASM_FILE_START arm_file_start | |
b12a00f1 RE |
208 | #undef TARGET_ASM_FILE_END |
209 | #define TARGET_ASM_FILE_END arm_file_end | |
210 | ||
1d6e90ac | 211 | #undef TARGET_ASM_ALIGNED_SI_OP |
301d03af | 212 | #define TARGET_ASM_ALIGNED_SI_OP NULL |
1d6e90ac | 213 | #undef TARGET_ASM_INTEGER |
301d03af | 214 | #define TARGET_ASM_INTEGER arm_assemble_integer |
301d03af | 215 | |
1d6e90ac | 216 | #undef TARGET_ASM_FUNCTION_PROLOGUE |
08c148a8 NB |
217 | #define TARGET_ASM_FUNCTION_PROLOGUE arm_output_function_prologue |
218 | ||
1d6e90ac | 219 | #undef TARGET_ASM_FUNCTION_EPILOGUE |
08c148a8 NB |
220 | #define TARGET_ASM_FUNCTION_EPILOGUE arm_output_function_epilogue |
221 | ||
c54c7322 RS |
222 | #undef TARGET_DEFAULT_TARGET_FLAGS |
223 | #define TARGET_DEFAULT_TARGET_FLAGS (TARGET_DEFAULT | MASK_SCHED_PROLOG) | |
224 | #undef TARGET_HANDLE_OPTION | |
225 | #define TARGET_HANDLE_OPTION arm_handle_option | |
67e6ba46 NC |
226 | #undef TARGET_HELP |
227 | #define TARGET_HELP arm_target_help | |
c54c7322 | 228 | |
1d6e90ac | 229 | #undef TARGET_COMP_TYPE_ATTRIBUTES |
8d8e52be JM |
230 | #define TARGET_COMP_TYPE_ATTRIBUTES arm_comp_type_attributes |
231 | ||
1d6e90ac | 232 | #undef TARGET_SET_DEFAULT_TYPE_ATTRIBUTES |
8d8e52be JM |
233 | #define TARGET_SET_DEFAULT_TYPE_ATTRIBUTES arm_set_default_type_attributes |
234 | ||
1d6e90ac | 235 | #undef TARGET_SCHED_ADJUST_COST |
c237e94a ZW |
236 | #define TARGET_SCHED_ADJUST_COST arm_adjust_cost |
237 | ||
fb49053f RH |
238 | #undef TARGET_ENCODE_SECTION_INFO |
239 | #ifdef ARM_PE | |
240 | #define TARGET_ENCODE_SECTION_INFO arm_pe_encode_section_info | |
241 | #else | |
242 | #define TARGET_ENCODE_SECTION_INFO arm_encode_section_info | |
243 | #endif | |
244 | ||
5a9335ef | 245 | #undef TARGET_STRIP_NAME_ENCODING |
772c5265 RH |
246 | #define TARGET_STRIP_NAME_ENCODING arm_strip_name_encoding |
247 | ||
5a9335ef | 248 | #undef TARGET_ASM_INTERNAL_LABEL |
4977bab6 ZW |
249 | #define TARGET_ASM_INTERNAL_LABEL arm_internal_label |
250 | ||
5a9335ef | 251 | #undef TARGET_FUNCTION_OK_FOR_SIBCALL |
4977bab6 ZW |
252 | #define TARGET_FUNCTION_OK_FOR_SIBCALL arm_function_ok_for_sibcall |
253 | ||
5a9335ef | 254 | #undef TARGET_ASM_OUTPUT_MI_THUNK |
c590b625 | 255 | #define TARGET_ASM_OUTPUT_MI_THUNK arm_output_mi_thunk |
5a9335ef | 256 | #undef TARGET_ASM_CAN_OUTPUT_MI_THUNK |
3961e8fe | 257 | #define TARGET_ASM_CAN_OUTPUT_MI_THUNK default_can_output_mi_thunk_no_vcall |
c590b625 | 258 | |
9b66ebb1 | 259 | /* This will be overridden in arm_override_options. */ |
5a9335ef | 260 | #undef TARGET_RTX_COSTS |
9b66ebb1 | 261 | #define TARGET_RTX_COSTS arm_slowmul_rtx_costs |
5a9335ef | 262 | #undef TARGET_ADDRESS_COST |
dcefdf67 | 263 | #define TARGET_ADDRESS_COST arm_address_cost |
3c50106f | 264 | |
273a2526 RS |
265 | #undef TARGET_SHIFT_TRUNCATION_MASK |
266 | #define TARGET_SHIFT_TRUNCATION_MASK arm_shift_truncation_mask | |
f676971a EC |
267 | #undef TARGET_VECTOR_MODE_SUPPORTED_P |
268 | #define TARGET_VECTOR_MODE_SUPPORTED_P arm_vector_mode_supported_p | |
269 | ||
5a9335ef | 270 | #undef TARGET_MACHINE_DEPENDENT_REORG |
18dbd950 RS |
271 | #define TARGET_MACHINE_DEPENDENT_REORG arm_reorg |
272 | ||
5a9335ef NC |
273 | #undef TARGET_INIT_BUILTINS |
274 | #define TARGET_INIT_BUILTINS arm_init_builtins | |
275 | #undef TARGET_EXPAND_BUILTIN | |
276 | #define TARGET_EXPAND_BUILTIN arm_expand_builtin | |
277 | ||
b3f8d95d MM |
278 | #undef TARGET_INIT_LIBFUNCS |
279 | #define TARGET_INIT_LIBFUNCS arm_init_libfuncs | |
280 | ||
f9ba5949 | 281 | #undef TARGET_PROMOTE_FUNCTION_ARGS |
586de218 | 282 | #define TARGET_PROMOTE_FUNCTION_ARGS hook_bool_const_tree_true |
d4453b7a | 283 | #undef TARGET_PROMOTE_FUNCTION_RETURN |
586de218 | 284 | #define TARGET_PROMOTE_FUNCTION_RETURN hook_bool_const_tree_true |
f9ba5949 | 285 | #undef TARGET_PROMOTE_PROTOTYPES |
70301b45 | 286 | #define TARGET_PROMOTE_PROTOTYPES arm_promote_prototypes |
8cd5a4e0 RH |
287 | #undef TARGET_PASS_BY_REFERENCE |
288 | #define TARGET_PASS_BY_REFERENCE arm_pass_by_reference | |
78a52f11 RH |
289 | #undef TARGET_ARG_PARTIAL_BYTES |
290 | #define TARGET_ARG_PARTIAL_BYTES arm_arg_partial_bytes | |
f9ba5949 | 291 | |
1cc9f5f5 KH |
292 | #undef TARGET_SETUP_INCOMING_VARARGS |
293 | #define TARGET_SETUP_INCOMING_VARARGS arm_setup_incoming_varargs | |
294 | ||
007e61c2 PB |
295 | #undef TARGET_ALLOCATE_STACK_SLOTS_FOR_ARGS |
296 | #define TARGET_ALLOCATE_STACK_SLOTS_FOR_ARGS arm_allocate_stack_slots_for_args | |
297 | ||
6b045785 PB |
298 | #undef TARGET_DEFAULT_SHORT_ENUMS |
299 | #define TARGET_DEFAULT_SHORT_ENUMS arm_default_short_enums | |
300 | ||
13c1cd82 PB |
301 | #undef TARGET_ALIGN_ANON_BITFIELD |
302 | #define TARGET_ALIGN_ANON_BITFIELD arm_align_anon_bitfield | |
303 | ||
c2a64439 PB |
304 | #undef TARGET_NARROW_VOLATILE_BITFIELD |
305 | #define TARGET_NARROW_VOLATILE_BITFIELD hook_bool_void_false | |
306 | ||
4185ae53 PB |
307 | #undef TARGET_CXX_GUARD_TYPE |
308 | #define TARGET_CXX_GUARD_TYPE arm_cxx_guard_type | |
309 | ||
310 | #undef TARGET_CXX_GUARD_MASK_BIT | |
311 | #define TARGET_CXX_GUARD_MASK_BIT arm_cxx_guard_mask_bit | |
312 | ||
46e995e0 PB |
313 | #undef TARGET_CXX_GET_COOKIE_SIZE |
314 | #define TARGET_CXX_GET_COOKIE_SIZE arm_get_cookie_size | |
315 | ||
316 | #undef TARGET_CXX_COOKIE_HAS_SIZE | |
317 | #define TARGET_CXX_COOKIE_HAS_SIZE arm_cookie_has_size | |
318 | ||
44d10c10 PB |
319 | #undef TARGET_CXX_CDTOR_RETURNS_THIS |
320 | #define TARGET_CXX_CDTOR_RETURNS_THIS arm_cxx_cdtor_returns_this | |
321 | ||
505970fc MM |
322 | #undef TARGET_CXX_KEY_METHOD_MAY_BE_INLINE |
323 | #define TARGET_CXX_KEY_METHOD_MAY_BE_INLINE arm_cxx_key_method_may_be_inline | |
324 | ||
9f62c3e3 PB |
325 | #undef TARGET_CXX_USE_AEABI_ATEXIT |
326 | #define TARGET_CXX_USE_AEABI_ATEXIT arm_cxx_use_aeabi_atexit | |
327 | ||
1e731102 MM |
328 | #undef TARGET_CXX_DETERMINE_CLASS_DATA_VISIBILITY |
329 | #define TARGET_CXX_DETERMINE_CLASS_DATA_VISIBILITY \ | |
330 | arm_cxx_determine_class_data_visibility | |
331 | ||
332 | #undef TARGET_CXX_CLASS_DATA_ALWAYS_COMDAT | |
333 | #define TARGET_CXX_CLASS_DATA_ALWAYS_COMDAT arm_cxx_class_data_always_comdat | |
505970fc | 334 | |
866af8a9 JB |
335 | #undef TARGET_RETURN_IN_MSB |
336 | #define TARGET_RETURN_IN_MSB arm_return_in_msb | |
337 | ||
23668cf7 CLT |
338 | #undef TARGET_RETURN_IN_MEMORY |
339 | #define TARGET_RETURN_IN_MEMORY arm_return_in_memory | |
340 | ||
866af8a9 JB |
341 | #undef TARGET_MUST_PASS_IN_STACK |
342 | #define TARGET_MUST_PASS_IN_STACK arm_must_pass_in_stack | |
343 | ||
617a1b71 PB |
344 | #ifdef TARGET_UNWIND_INFO |
345 | #undef TARGET_UNWIND_EMIT | |
346 | #define TARGET_UNWIND_EMIT arm_unwind_emit | |
347 | ||
348 | /* EABI unwinding tables use a different format for the typeinfo tables. */ | |
349 | #undef TARGET_ASM_TTYPE | |
350 | #define TARGET_ASM_TTYPE arm_output_ttype | |
351 | ||
352 | #undef TARGET_ARM_EABI_UNWINDER | |
353 | #define TARGET_ARM_EABI_UNWINDER true | |
354 | #endif /* TARGET_UNWIND_INFO */ | |
355 | ||
5b3e6663 PB |
356 | #undef TARGET_DWARF_HANDLE_FRAME_UNSPEC |
357 | #define TARGET_DWARF_HANDLE_FRAME_UNSPEC arm_dwarf_handle_frame_unspec | |
358 | ||
d3585b76 DJ |
359 | #undef TARGET_CANNOT_COPY_INSN_P |
360 | #define TARGET_CANNOT_COPY_INSN_P arm_cannot_copy_insn_p | |
361 | ||
362 | #ifdef HAVE_AS_TLS | |
363 | #undef TARGET_HAVE_TLS | |
364 | #define TARGET_HAVE_TLS true | |
365 | #endif | |
366 | ||
367 | #undef TARGET_CANNOT_FORCE_CONST_MEM | |
8426b956 | 368 | #define TARGET_CANNOT_FORCE_CONST_MEM arm_cannot_force_const_mem |
d3585b76 | 369 | |
bd4dc3cd PB |
370 | #undef TARGET_SCHED_ISSUE_RATE |
371 | #define TARGET_SCHED_ISSUE_RATE arm_issue_rate | |
372 | ||
608063c3 JB |
373 | #undef TARGET_MANGLE_TYPE |
374 | #define TARGET_MANGLE_TYPE arm_mangle_type | |
375 | ||
afcc986d JM |
376 | #ifdef HAVE_AS_TLS |
377 | #undef TARGET_ASM_OUTPUT_DWARF_DTPREL | |
378 | #define TARGET_ASM_OUTPUT_DWARF_DTPREL arm_output_dwarf_dtprel | |
379 | #endif | |
380 | ||
f6897b10 | 381 | struct gcc_target targetm = TARGET_INITIALIZER; |
672a6f42 | 382 | \f |
c7319d87 RE |
383 | /* Obstack for minipool constant handling. */ |
384 | static struct obstack minipool_obstack; | |
1d6e90ac | 385 | static char * minipool_startobj; |
c7319d87 | 386 | |
1d6e90ac NC |
387 | /* The maximum number of insns skipped which |
388 | will be conditionalised if possible. */ | |
c27ba912 DM |
389 | static int max_insns_skipped = 5; |
390 | ||
391 | extern FILE * asm_out_file; | |
392 | ||
6354dc9b | 393 | /* True if we are currently building a constant table. */ |
13bd191d PB |
394 | int making_const_table; |
395 | ||
60d0536b | 396 | /* Define the information needed to generate branch insns. This is |
6354dc9b | 397 | stored from the compare operation. */ |
ff9940b0 | 398 | rtx arm_compare_op0, arm_compare_op1; |
ff9940b0 | 399 | |
9b66ebb1 PB |
400 | /* The processor for which instructions should be scheduled. */ |
401 | enum processor_type arm_tune = arm_none; | |
402 | ||
2f8e468b | 403 | /* The default processor used if not overridden by commandline. */ |
6c6aa1af PB |
404 | static enum processor_type arm_default_cpu = arm_none; |
405 | ||
9b66ebb1 PB |
406 | /* Which floating point model to use. */ |
407 | enum arm_fp_model arm_fp_model; | |
bee06f3d | 408 | |
9b66ebb1 | 409 | /* Which floating point hardware is available. */ |
29ad9694 | 410 | enum fputype arm_fpu_arch; |
b111229a | 411 | |
9b66ebb1 PB |
412 | /* Which floating point hardware to schedule for. */ |
413 | enum fputype arm_fpu_tune; | |
414 | ||
415 | /* Whether to use floating point hardware. */ | |
416 | enum float_abi_type arm_float_abi; | |
417 | ||
5848830f PB |
418 | /* Which ABI to use. */ |
419 | enum arm_abi_type arm_abi; | |
420 | ||
d3585b76 DJ |
421 | /* Which thread pointer model to use. */ |
422 | enum arm_tp_type target_thread_pointer = TP_AUTO; | |
423 | ||
b355a481 | 424 | /* Used to parse -mstructure_size_boundary command line option. */ |
723ae7c1 | 425 | int arm_structure_size_boundary = DEFAULT_STRUCTURE_SIZE_BOUNDARY; |
b355a481 | 426 | |
b12a00f1 | 427 | /* Used for Thumb call_via trampolines. */ |
57ecec57 | 428 | rtx thumb_call_via_label[14]; |
b12a00f1 RE |
429 | static int thumb_call_reg_needed; |
430 | ||
aec3cfba | 431 | /* Bit values used to identify processor capabilities. */ |
62b10bbc | 432 | #define FL_CO_PROC (1 << 0) /* Has external co-processor bus */ |
9b66ebb1 | 433 | #define FL_ARCH3M (1 << 1) /* Extended multiply */ |
62b10bbc NC |
434 | #define FL_MODE26 (1 << 2) /* 26-bit mode support */ |
435 | #define FL_MODE32 (1 << 3) /* 32-bit mode support */ | |
436 | #define FL_ARCH4 (1 << 4) /* Architecture rel 4 */ | |
437 | #define FL_ARCH5 (1 << 5) /* Architecture rel 5 */ | |
438 | #define FL_THUMB (1 << 6) /* Thumb aware */ | |
439 | #define FL_LDSCHED (1 << 7) /* Load scheduling necessary */ | |
440 | #define FL_STRONG (1 << 8) /* StrongARM */ | |
6bc82793 | 441 | #define FL_ARCH5E (1 << 9) /* DSP extensions to v5 */ |
d19fb8e3 | 442 | #define FL_XSCALE (1 << 10) /* XScale */ |
9b6b54e2 | 443 | #define FL_CIRRUS (1 << 11) /* Cirrus/DSP. */ |
9b66ebb1 | 444 | #define FL_ARCH6 (1 << 12) /* Architecture rel 6. Adds |
81f9037c MM |
445 | media instructions. */ |
446 | #define FL_VFPV2 (1 << 13) /* Vector Floating Point V2. */ | |
abac3b49 RE |
447 | #define FL_WBUF (1 << 14) /* Schedule for write buffer ops. |
448 | Note: ARM6 & 7 derivatives only. */ | |
d3585b76 | 449 | #define FL_ARCH6K (1 << 15) /* Architecture rel 6 K extensions. */ |
5b3e6663 PB |
450 | #define FL_THUMB2 (1 << 16) /* Thumb-2. */ |
451 | #define FL_NOTM (1 << 17) /* Instructions not present in the 'M' | |
452 | profile. */ | |
7a085dce | 453 | #define FL_DIV (1 << 18) /* Hardware divide. */ |
f1adb0a9 | 454 | #define FL_VFPV3 (1 << 19) /* Vector Floating Point V3. */ |
88f77cba | 455 | #define FL_NEON (1 << 20) /* Neon instructions. */ |
aec3cfba | 456 | |
9b66ebb1 PB |
457 | #define FL_IWMMXT (1 << 29) /* XScale v2 or "Intel Wireless MMX technology". */ |
458 | ||
5b3e6663 PB |
459 | #define FL_FOR_ARCH2 FL_NOTM |
460 | #define FL_FOR_ARCH3 (FL_FOR_ARCH2 | FL_MODE32) | |
78011587 PB |
461 | #define FL_FOR_ARCH3M (FL_FOR_ARCH3 | FL_ARCH3M) |
462 | #define FL_FOR_ARCH4 (FL_FOR_ARCH3M | FL_ARCH4) | |
463 | #define FL_FOR_ARCH4T (FL_FOR_ARCH4 | FL_THUMB) | |
464 | #define FL_FOR_ARCH5 (FL_FOR_ARCH4 | FL_ARCH5) | |
465 | #define FL_FOR_ARCH5T (FL_FOR_ARCH5 | FL_THUMB) | |
466 | #define FL_FOR_ARCH5E (FL_FOR_ARCH5 | FL_ARCH5E) | |
467 | #define FL_FOR_ARCH5TE (FL_FOR_ARCH5E | FL_THUMB) | |
468 | #define FL_FOR_ARCH5TEJ FL_FOR_ARCH5TE | |
469 | #define FL_FOR_ARCH6 (FL_FOR_ARCH5TE | FL_ARCH6) | |
470 | #define FL_FOR_ARCH6J FL_FOR_ARCH6 | |
d3585b76 | 471 | #define FL_FOR_ARCH6K (FL_FOR_ARCH6 | FL_ARCH6K) |
fa91adc6 | 472 | #define FL_FOR_ARCH6Z FL_FOR_ARCH6 |
d3585b76 | 473 | #define FL_FOR_ARCH6ZK FL_FOR_ARCH6K |
5b3e6663 | 474 | #define FL_FOR_ARCH6T2 (FL_FOR_ARCH6 | FL_THUMB2) |
bf98ec6c | 475 | #define FL_FOR_ARCH6M (FL_FOR_ARCH6 & ~FL_NOTM) |
5b3e6663 PB |
476 | #define FL_FOR_ARCH7 (FL_FOR_ARCH6T2 &~ FL_NOTM) |
477 | #define FL_FOR_ARCH7A (FL_FOR_ARCH7 | FL_NOTM) | |
478 | #define FL_FOR_ARCH7R (FL_FOR_ARCH7A | FL_DIV) | |
479 | #define FL_FOR_ARCH7M (FL_FOR_ARCH7 | FL_DIV) | |
78011587 | 480 | |
1d6e90ac NC |
481 | /* The bits in this mask specify which |
482 | instructions we are allowed to generate. */ | |
0977774b | 483 | static unsigned long insn_flags = 0; |
d5b7b3ae | 484 | |
aec3cfba | 485 | /* The bits in this mask specify which instruction scheduling options should |
9b66ebb1 | 486 | be used. */ |
0977774b | 487 | static unsigned long tune_flags = 0; |
aec3cfba NC |
488 | |
489 | /* The following are used in the arm.md file as equivalents to bits | |
490 | in the above two flag variables. */ | |
491 | ||
9b66ebb1 PB |
492 | /* Nonzero if this chip supports the ARM Architecture 3M extensions. */ |
493 | int arm_arch3m = 0; | |
2b835d68 | 494 | |
6354dc9b | 495 | /* Nonzero if this chip supports the ARM Architecture 4 extensions. */ |
2b835d68 RE |
496 | int arm_arch4 = 0; |
497 | ||
68d560d4 RE |
498 | /* Nonzero if this chip supports the ARM Architecture 4t extensions. */ |
499 | int arm_arch4t = 0; | |
500 | ||
6354dc9b | 501 | /* Nonzero if this chip supports the ARM Architecture 5 extensions. */ |
62b10bbc NC |
502 | int arm_arch5 = 0; |
503 | ||
b15bca31 RE |
504 | /* Nonzero if this chip supports the ARM Architecture 5E extensions. */ |
505 | int arm_arch5e = 0; | |
506 | ||
9b66ebb1 PB |
507 | /* Nonzero if this chip supports the ARM Architecture 6 extensions. */ |
508 | int arm_arch6 = 0; | |
509 | ||
d3585b76 DJ |
510 | /* Nonzero if this chip supports the ARM 6K extensions. */ |
511 | int arm_arch6k = 0; | |
512 | ||
5b3e6663 PB |
513 | /* Nonzero if instructions not present in the 'M' profile can be used. */ |
514 | int arm_arch_notm = 0; | |
515 | ||
aec3cfba | 516 | /* Nonzero if this chip can benefit from load scheduling. */ |
f5a1b0d2 NC |
517 | int arm_ld_sched = 0; |
518 | ||
519 | /* Nonzero if this chip is a StrongARM. */ | |
abac3b49 | 520 | int arm_tune_strongarm = 0; |
f5a1b0d2 | 521 | |
78011587 PB |
522 | /* Nonzero if this chip is a Cirrus variant. */ |
523 | int arm_arch_cirrus = 0; | |
524 | ||
5a9335ef NC |
525 | /* Nonzero if this chip supports Intel Wireless MMX technology. */ |
526 | int arm_arch_iwmmxt = 0; | |
527 | ||
d19fb8e3 | 528 | /* Nonzero if this chip is an XScale. */ |
4b3c2e48 PB |
529 | int arm_arch_xscale = 0; |
530 | ||
531 | /* Nonzero if tuning for XScale */ | |
532 | int arm_tune_xscale = 0; | |
d19fb8e3 | 533 | |
e0b92319 | 534 | /* Nonzero if we want to tune for stores that access the write-buffer. |
c5d34bb2 | 535 | This typically means an ARM6 or ARM7 with MMU or MPU. */ |
abac3b49 | 536 | int arm_tune_wbuf = 0; |
b111229a | 537 | |
0616531f RE |
538 | /* Nonzero if generating Thumb instructions. */ |
539 | int thumb_code = 0; | |
540 | ||
2ad4dcf9 | 541 | /* Nonzero if we should define __THUMB_INTERWORK__ in the |
f676971a | 542 | preprocessor. |
2ad4dcf9 RE |
543 | XXX This is a bit of a hack, it's intended to help work around |
544 | problems in GLD which doesn't understand that armv5t code is | |
545 | interworking clean. */ | |
546 | int arm_cpp_interwork = 0; | |
547 | ||
5b3e6663 PB |
548 | /* Nonzero if chip supports Thumb 2. */ |
549 | int arm_arch_thumb2; | |
550 | ||
551 | /* Nonzero if chip supports integer division instruction. */ | |
552 | int arm_arch_hwdiv; | |
553 | ||
cce8749e CH |
554 | /* In case of a PRE_INC, POST_INC, PRE_DEC, POST_DEC memory reference, we |
555 | must report the mode of the memory reference from PRINT_OPERAND to | |
556 | PRINT_OPERAND_ADDRESS. */ | |
f3bb6135 | 557 | enum machine_mode output_memory_reference_mode; |
cce8749e | 558 | |
32de079a | 559 | /* The register number to be used for the PIC offset register. */ |
020a4035 | 560 | unsigned arm_pic_register = INVALID_REGNUM; |
32de079a | 561 | |
ff9940b0 | 562 | /* Set to 1 when a return insn is output, this means that the epilogue |
6354dc9b | 563 | is not needed. */ |
d5b7b3ae | 564 | int return_used_this_function; |
ff9940b0 | 565 | |
aec3cfba NC |
566 | /* Set to 1 after arm_reorg has started. Reset to start at the start of |
567 | the next function. */ | |
4b632bf1 RE |
568 | static int after_arm_reorg = 0; |
569 | ||
aec3cfba | 570 | /* The maximum number of insns to be used when loading a constant. */ |
2b835d68 RE |
571 | static int arm_constant_limit = 3; |
572 | ||
cce8749e CH |
573 | /* For an explanation of these variables, see final_prescan_insn below. */ |
574 | int arm_ccfsm_state; | |
5b3e6663 | 575 | /* arm_current_cc is also used for Thumb-2 cond_exec blocks. */ |
84ed5e79 | 576 | enum arm_cond_code arm_current_cc; |
cce8749e CH |
577 | rtx arm_target_insn; |
578 | int arm_target_label; | |
5b3e6663 PB |
579 | /* The number of conditionally executed insns, including the current insn. */ |
580 | int arm_condexec_count = 0; | |
581 | /* A bitmask specifying the patterns for the IT block. | |
582 | Zero means do not output an IT block before this insn. */ | |
583 | int arm_condexec_mask = 0; | |
584 | /* The number of bits used in arm_condexec_mask. */ | |
585 | int arm_condexec_masklen = 0; | |
9997d19d RE |
586 | |
587 | /* The condition codes of the ARM, and the inverse function. */ | |
1d6e90ac | 588 | static const char * const arm_condition_codes[] = |
9997d19d RE |
589 | { |
590 | "eq", "ne", "cs", "cc", "mi", "pl", "vs", "vc", | |
591 | "hi", "ls", "ge", "lt", "gt", "le", "al", "nv" | |
592 | }; | |
593 | ||
5b3e6663 | 594 | #define ARM_LSL_NAME (TARGET_UNIFIED_ASM ? "lsl" : "asl") |
f5a1b0d2 | 595 | #define streq(string1, string2) (strcmp (string1, string2) == 0) |
5b3e6663 PB |
596 | |
597 | #define THUMB2_WORK_REGS (0xff & ~( (1 << THUMB_HARD_FRAME_POINTER_REGNUM) \ | |
598 | | (1 << SP_REGNUM) | (1 << PC_REGNUM) \ | |
599 | | (1 << PIC_OFFSET_TABLE_REGNUM))) | |
2b835d68 | 600 | \f |
6354dc9b | 601 | /* Initialization code. */ |
2b835d68 | 602 | |
2b835d68 RE |
603 | struct processors |
604 | { | |
8b60264b | 605 | const char *const name; |
9b66ebb1 | 606 | enum processor_type core; |
78011587 | 607 | const char *arch; |
0977774b | 608 | const unsigned long flags; |
9b66ebb1 | 609 | bool (* rtx_costs) (rtx, int, int, int *); |
2b835d68 RE |
610 | }; |
611 | ||
612 | /* Not all of these give usefully different compilation alternatives, | |
613 | but there is no simple way of generalizing them. */ | |
8b60264b | 614 | static const struct processors all_cores[] = |
f5a1b0d2 NC |
615 | { |
616 | /* ARM Cores */ | |
d98a72fd RE |
617 | #define ARM_CORE(NAME, IDENT, ARCH, FLAGS, COSTS) \ |
618 | {NAME, arm_none, #ARCH, FLAGS | FL_FOR_ARCH##ARCH, arm_##COSTS##_rtx_costs}, | |
9b66ebb1 PB |
619 | #include "arm-cores.def" |
620 | #undef ARM_CORE | |
78011587 | 621 | {NULL, arm_none, NULL, 0, NULL} |
f5a1b0d2 NC |
622 | }; |
623 | ||
8b60264b | 624 | static const struct processors all_architectures[] = |
2b835d68 | 625 | { |
f5a1b0d2 | 626 | /* ARM Architectures */ |
9b66ebb1 PB |
627 | /* We don't specify rtx_costs here as it will be figured out |
628 | from the core. */ | |
f676971a | 629 | |
78011587 PB |
630 | {"armv2", arm2, "2", FL_CO_PROC | FL_MODE26 | FL_FOR_ARCH2, NULL}, |
631 | {"armv2a", arm2, "2", FL_CO_PROC | FL_MODE26 | FL_FOR_ARCH2, NULL}, | |
632 | {"armv3", arm6, "3", FL_CO_PROC | FL_MODE26 | FL_FOR_ARCH3, NULL}, | |
633 | {"armv3m", arm7m, "3M", FL_CO_PROC | FL_MODE26 | FL_FOR_ARCH3M, NULL}, | |
634 | {"armv4", arm7tdmi, "4", FL_CO_PROC | FL_MODE26 | FL_FOR_ARCH4, NULL}, | |
b111229a RE |
635 | /* Strictly, FL_MODE26 is a permitted option for v4t, but there are no |
636 | implementations that support it, so we will leave it out for now. */ | |
78011587 PB |
637 | {"armv4t", arm7tdmi, "4T", FL_CO_PROC | FL_FOR_ARCH4T, NULL}, |
638 | {"armv5", arm10tdmi, "5", FL_CO_PROC | FL_FOR_ARCH5, NULL}, | |
639 | {"armv5t", arm10tdmi, "5T", FL_CO_PROC | FL_FOR_ARCH5T, NULL}, | |
640 | {"armv5e", arm1026ejs, "5E", FL_CO_PROC | FL_FOR_ARCH5E, NULL}, | |
641 | {"armv5te", arm1026ejs, "5TE", FL_CO_PROC | FL_FOR_ARCH5TE, NULL}, | |
642 | {"armv6", arm1136js, "6", FL_CO_PROC | FL_FOR_ARCH6, NULL}, | |
643 | {"armv6j", arm1136js, "6J", FL_CO_PROC | FL_FOR_ARCH6J, NULL}, | |
fa91adc6 PB |
644 | {"armv6k", mpcore, "6K", FL_CO_PROC | FL_FOR_ARCH6K, NULL}, |
645 | {"armv6z", arm1176jzs, "6Z", FL_CO_PROC | FL_FOR_ARCH6Z, NULL}, | |
646 | {"armv6zk", arm1176jzs, "6ZK", FL_CO_PROC | FL_FOR_ARCH6ZK, NULL}, | |
5b3e6663 | 647 | {"armv6t2", arm1156t2s, "6T2", FL_CO_PROC | FL_FOR_ARCH6T2, NULL}, |
bf98ec6c | 648 | {"armv6-m", cortexm1, "6M", FL_FOR_ARCH6M, NULL}, |
5b3e6663 PB |
649 | {"armv7", cortexa8, "7", FL_CO_PROC | FL_FOR_ARCH7, NULL}, |
650 | {"armv7-a", cortexa8, "7A", FL_CO_PROC | FL_FOR_ARCH7A, NULL}, | |
651 | {"armv7-r", cortexr4, "7R", FL_CO_PROC | FL_FOR_ARCH7R, NULL}, | |
652 | {"armv7-m", cortexm3, "7M", FL_CO_PROC | FL_FOR_ARCH7M, NULL}, | |
78011587 PB |
653 | {"ep9312", ep9312, "4T", FL_LDSCHED | FL_CIRRUS | FL_FOR_ARCH4, NULL}, |
654 | {"iwmmxt", iwmmxt, "5TE", FL_LDSCHED | FL_STRONG | FL_FOR_ARCH5TE | FL_XSCALE | FL_IWMMXT , NULL}, | |
655 | {NULL, arm_none, NULL, 0 , NULL} | |
f5a1b0d2 NC |
656 | }; |
657 | ||
c54c7322 RS |
658 | struct arm_cpu_select |
659 | { | |
660 | const char * string; | |
661 | const char * name; | |
662 | const struct processors * processors; | |
663 | }; | |
664 | ||
9a9f7594 | 665 | /* This is a magic structure. The 'string' field is magically filled in |
f5a1b0d2 NC |
666 | with a pointer to the value specified by the user on the command line |
667 | assuming that the user has specified such a value. */ | |
668 | ||
c54c7322 | 669 | static struct arm_cpu_select arm_select[] = |
f5a1b0d2 | 670 | { |
f676971a | 671 | /* string name processors */ |
f5a1b0d2 NC |
672 | { NULL, "-mcpu=", all_cores }, |
673 | { NULL, "-march=", all_architectures }, | |
674 | { NULL, "-mtune=", all_cores } | |
2b835d68 RE |
675 | }; |
676 | ||
e6fffade RE |
677 | /* Defines representing the indexes into the above table. */ |
678 | #define ARM_OPT_SET_CPU 0 | |
679 | #define ARM_OPT_SET_ARCH 1 | |
680 | #define ARM_OPT_SET_TUNE 2 | |
78011587 | 681 | |
afc0a4ba | 682 | /* The name of the preprocessor macro to define for this architecture. */ |
78011587 PB |
683 | |
684 | char arm_arch_name[] = "__ARM_ARCH_0UNK__"; | |
685 | ||
9b66ebb1 PB |
686 | struct fpu_desc |
687 | { | |
688 | const char * name; | |
689 | enum fputype fpu; | |
690 | }; | |
691 | ||
692 | ||
56f42830 | 693 | /* Available values for -mfpu=. */ |
9b66ebb1 PB |
694 | |
695 | static const struct fpu_desc all_fpus[] = | |
696 | { | |
697 | {"fpa", FPUTYPE_FPA}, | |
698 | {"fpe2", FPUTYPE_FPA_EMU2}, | |
699 | {"fpe3", FPUTYPE_FPA_EMU2}, | |
700 | {"maverick", FPUTYPE_MAVERICK}, | |
f1adb0a9 JB |
701 | {"vfp", FPUTYPE_VFP}, |
702 | {"vfp3", FPUTYPE_VFP3}, | |
88f77cba | 703 | {"neon", FPUTYPE_NEON} |
9b66ebb1 PB |
704 | }; |
705 | ||
706 | ||
707 | /* Floating point models used by the different hardware. | |
708 | See fputype in arm.h. */ | |
709 | ||
710 | static const enum fputype fp_model_for_fpu[] = | |
711 | { | |
712 | /* No FP hardware. */ | |
713 | ARM_FP_MODEL_UNKNOWN, /* FPUTYPE_NONE */ | |
714 | ARM_FP_MODEL_FPA, /* FPUTYPE_FPA */ | |
715 | ARM_FP_MODEL_FPA, /* FPUTYPE_FPA_EMU2 */ | |
716 | ARM_FP_MODEL_FPA, /* FPUTYPE_FPA_EMU3 */ | |
717 | ARM_FP_MODEL_MAVERICK, /* FPUTYPE_MAVERICK */ | |
f1adb0a9 | 718 | ARM_FP_MODEL_VFP, /* FPUTYPE_VFP */ |
88f77cba JB |
719 | ARM_FP_MODEL_VFP, /* FPUTYPE_VFP3 */ |
720 | ARM_FP_MODEL_VFP /* FPUTYPE_NEON */ | |
9b66ebb1 PB |
721 | }; |
722 | ||
723 | ||
724 | struct float_abi | |
725 | { | |
726 | const char * name; | |
727 | enum float_abi_type abi_type; | |
728 | }; | |
729 | ||
730 | ||
731 | /* Available values for -mfloat-abi=. */ | |
732 | ||
733 | static const struct float_abi all_float_abis[] = | |
734 | { | |
735 | {"soft", ARM_FLOAT_ABI_SOFT}, | |
736 | {"softfp", ARM_FLOAT_ABI_SOFTFP}, | |
737 | {"hard", ARM_FLOAT_ABI_HARD} | |
738 | }; | |
739 | ||
740 | ||
5848830f PB |
741 | struct abi_name |
742 | { | |
743 | const char *name; | |
744 | enum arm_abi_type abi_type; | |
745 | }; | |
746 | ||
747 | ||
748 | /* Available values for -mabi=. */ | |
749 | ||
750 | static const struct abi_name arm_all_abis[] = | |
751 | { | |
752 | {"apcs-gnu", ARM_ABI_APCS}, | |
753 | {"atpcs", ARM_ABI_ATPCS}, | |
754 | {"aapcs", ARM_ABI_AAPCS}, | |
077fc835 KH |
755 | {"iwmmxt", ARM_ABI_IWMMXT}, |
756 | {"aapcs-linux", ARM_ABI_AAPCS_LINUX} | |
5848830f PB |
757 | }; |
758 | ||
d3585b76 DJ |
759 | /* Supported TLS relocations. */ |
760 | ||
761 | enum tls_reloc { | |
762 | TLS_GD32, | |
763 | TLS_LDM32, | |
764 | TLS_LDO32, | |
765 | TLS_IE32, | |
766 | TLS_LE32 | |
767 | }; | |
768 | ||
d66437c5 RE |
769 | /* Emit an insn that's a simple single-set. Both the operands must be known |
770 | to be valid. */ | |
771 | inline static rtx | |
772 | emit_set_insn (rtx x, rtx y) | |
773 | { | |
774 | return emit_insn (gen_rtx_SET (VOIDmode, x, y)); | |
775 | } | |
776 | ||
0977774b JT |
777 | /* Return the number of bits set in VALUE. */ |
778 | static unsigned | |
e32bac5b | 779 | bit_count (unsigned long value) |
aec3cfba | 780 | { |
d5b7b3ae | 781 | unsigned long count = 0; |
f676971a | 782 | |
aec3cfba NC |
783 | while (value) |
784 | { | |
0977774b JT |
785 | count++; |
786 | value &= value - 1; /* Clear the least-significant set bit. */ | |
aec3cfba NC |
787 | } |
788 | ||
789 | return count; | |
790 | } | |
791 | ||
c112cf2b | 792 | /* Set up library functions unique to ARM. */ |
b3f8d95d MM |
793 | |
794 | static void | |
795 | arm_init_libfuncs (void) | |
796 | { | |
797 | /* There are no special library functions unless we are using the | |
798 | ARM BPABI. */ | |
799 | if (!TARGET_BPABI) | |
800 | return; | |
801 | ||
802 | /* The functions below are described in Section 4 of the "Run-Time | |
803 | ABI for the ARM architecture", Version 1.0. */ | |
804 | ||
805 | /* Double-precision floating-point arithmetic. Table 2. */ | |
806 | set_optab_libfunc (add_optab, DFmode, "__aeabi_dadd"); | |
807 | set_optab_libfunc (sdiv_optab, DFmode, "__aeabi_ddiv"); | |
808 | set_optab_libfunc (smul_optab, DFmode, "__aeabi_dmul"); | |
809 | set_optab_libfunc (neg_optab, DFmode, "__aeabi_dneg"); | |
810 | set_optab_libfunc (sub_optab, DFmode, "__aeabi_dsub"); | |
811 | ||
c112cf2b | 812 | /* Double-precision comparisons. Table 3. */ |
b3f8d95d MM |
813 | set_optab_libfunc (eq_optab, DFmode, "__aeabi_dcmpeq"); |
814 | set_optab_libfunc (ne_optab, DFmode, NULL); | |
815 | set_optab_libfunc (lt_optab, DFmode, "__aeabi_dcmplt"); | |
816 | set_optab_libfunc (le_optab, DFmode, "__aeabi_dcmple"); | |
817 | set_optab_libfunc (ge_optab, DFmode, "__aeabi_dcmpge"); | |
818 | set_optab_libfunc (gt_optab, DFmode, "__aeabi_dcmpgt"); | |
819 | set_optab_libfunc (unord_optab, DFmode, "__aeabi_dcmpun"); | |
820 | ||
821 | /* Single-precision floating-point arithmetic. Table 4. */ | |
822 | set_optab_libfunc (add_optab, SFmode, "__aeabi_fadd"); | |
823 | set_optab_libfunc (sdiv_optab, SFmode, "__aeabi_fdiv"); | |
824 | set_optab_libfunc (smul_optab, SFmode, "__aeabi_fmul"); | |
825 | set_optab_libfunc (neg_optab, SFmode, "__aeabi_fneg"); | |
826 | set_optab_libfunc (sub_optab, SFmode, "__aeabi_fsub"); | |
f676971a | 827 | |
c112cf2b | 828 | /* Single-precision comparisons. Table 5. */ |
b3f8d95d MM |
829 | set_optab_libfunc (eq_optab, SFmode, "__aeabi_fcmpeq"); |
830 | set_optab_libfunc (ne_optab, SFmode, NULL); | |
831 | set_optab_libfunc (lt_optab, SFmode, "__aeabi_fcmplt"); | |
832 | set_optab_libfunc (le_optab, SFmode, "__aeabi_fcmple"); | |
833 | set_optab_libfunc (ge_optab, SFmode, "__aeabi_fcmpge"); | |
834 | set_optab_libfunc (gt_optab, SFmode, "__aeabi_fcmpgt"); | |
835 | set_optab_libfunc (unord_optab, SFmode, "__aeabi_fcmpun"); | |
836 | ||
837 | /* Floating-point to integer conversions. Table 6. */ | |
838 | set_conv_libfunc (sfix_optab, SImode, DFmode, "__aeabi_d2iz"); | |
839 | set_conv_libfunc (ufix_optab, SImode, DFmode, "__aeabi_d2uiz"); | |
840 | set_conv_libfunc (sfix_optab, DImode, DFmode, "__aeabi_d2lz"); | |
841 | set_conv_libfunc (ufix_optab, DImode, DFmode, "__aeabi_d2ulz"); | |
842 | set_conv_libfunc (sfix_optab, SImode, SFmode, "__aeabi_f2iz"); | |
843 | set_conv_libfunc (ufix_optab, SImode, SFmode, "__aeabi_f2uiz"); | |
844 | set_conv_libfunc (sfix_optab, DImode, SFmode, "__aeabi_f2lz"); | |
845 | set_conv_libfunc (ufix_optab, DImode, SFmode, "__aeabi_f2ulz"); | |
846 | ||
847 | /* Conversions between floating types. Table 7. */ | |
848 | set_conv_libfunc (trunc_optab, SFmode, DFmode, "__aeabi_d2f"); | |
849 | set_conv_libfunc (sext_optab, DFmode, SFmode, "__aeabi_f2d"); | |
850 | ||
c112cf2b | 851 | /* Integer to floating-point conversions. Table 8. */ |
b3f8d95d MM |
852 | set_conv_libfunc (sfloat_optab, DFmode, SImode, "__aeabi_i2d"); |
853 | set_conv_libfunc (ufloat_optab, DFmode, SImode, "__aeabi_ui2d"); | |
854 | set_conv_libfunc (sfloat_optab, DFmode, DImode, "__aeabi_l2d"); | |
855 | set_conv_libfunc (ufloat_optab, DFmode, DImode, "__aeabi_ul2d"); | |
856 | set_conv_libfunc (sfloat_optab, SFmode, SImode, "__aeabi_i2f"); | |
857 | set_conv_libfunc (ufloat_optab, SFmode, SImode, "__aeabi_ui2f"); | |
858 | set_conv_libfunc (sfloat_optab, SFmode, DImode, "__aeabi_l2f"); | |
859 | set_conv_libfunc (ufloat_optab, SFmode, DImode, "__aeabi_ul2f"); | |
860 | ||
861 | /* Long long. Table 9. */ | |
862 | set_optab_libfunc (smul_optab, DImode, "__aeabi_lmul"); | |
863 | set_optab_libfunc (sdivmod_optab, DImode, "__aeabi_ldivmod"); | |
864 | set_optab_libfunc (udivmod_optab, DImode, "__aeabi_uldivmod"); | |
865 | set_optab_libfunc (ashl_optab, DImode, "__aeabi_llsl"); | |
866 | set_optab_libfunc (lshr_optab, DImode, "__aeabi_llsr"); | |
867 | set_optab_libfunc (ashr_optab, DImode, "__aeabi_lasr"); | |
868 | set_optab_libfunc (cmp_optab, DImode, "__aeabi_lcmp"); | |
869 | set_optab_libfunc (ucmp_optab, DImode, "__aeabi_ulcmp"); | |
870 | ||
871 | /* Integer (32/32->32) division. \S 4.3.1. */ | |
872 | set_optab_libfunc (sdivmod_optab, SImode, "__aeabi_idivmod"); | |
873 | set_optab_libfunc (udivmod_optab, SImode, "__aeabi_uidivmod"); | |
874 | ||
875 | /* The divmod functions are designed so that they can be used for | |
876 | plain division, even though they return both the quotient and the | |
877 | remainder. The quotient is returned in the usual location (i.e., | |
878 | r0 for SImode, {r0, r1} for DImode), just as would be expected | |
879 | for an ordinary division routine. Because the AAPCS calling | |
880 | conventions specify that all of { r0, r1, r2, r3 } are | |
881 | callee-saved registers, there is no need to tell the compiler | |
882 | explicitly that those registers are clobbered by these | |
883 | routines. */ | |
884 | set_optab_libfunc (sdiv_optab, DImode, "__aeabi_ldivmod"); | |
885 | set_optab_libfunc (udiv_optab, DImode, "__aeabi_uldivmod"); | |
e993ba8f DJ |
886 | |
887 | /* For SImode division the ABI provides div-without-mod routines, | |
888 | which are faster. */ | |
889 | set_optab_libfunc (sdiv_optab, SImode, "__aeabi_idiv"); | |
890 | set_optab_libfunc (udiv_optab, SImode, "__aeabi_uidiv"); | |
01c19d47 PB |
891 | |
892 | /* We don't have mod libcalls. Fortunately gcc knows how to use the | |
893 | divmod libcalls instead. */ | |
894 | set_optab_libfunc (smod_optab, DImode, NULL); | |
895 | set_optab_libfunc (umod_optab, DImode, NULL); | |
896 | set_optab_libfunc (smod_optab, SImode, NULL); | |
897 | set_optab_libfunc (umod_optab, SImode, NULL); | |
b3f8d95d MM |
898 | } |
899 | ||
c54c7322 RS |
900 | /* Implement TARGET_HANDLE_OPTION. */ |
901 | ||
902 | static bool | |
903 | arm_handle_option (size_t code, const char *arg, int value ATTRIBUTE_UNUSED) | |
904 | { | |
905 | switch (code) | |
906 | { | |
c54c7322 RS |
907 | case OPT_march_: |
908 | arm_select[1].string = arg; | |
909 | return true; | |
910 | ||
911 | case OPT_mcpu_: | |
912 | arm_select[0].string = arg; | |
913 | return true; | |
914 | ||
c54c7322 RS |
915 | case OPT_mhard_float: |
916 | target_float_abi_name = "hard"; | |
917 | return true; | |
918 | ||
c54c7322 RS |
919 | case OPT_msoft_float: |
920 | target_float_abi_name = "soft"; | |
921 | return true; | |
922 | ||
c54c7322 RS |
923 | case OPT_mtune_: |
924 | arm_select[2].string = arg; | |
925 | return true; | |
926 | ||
927 | default: | |
928 | return true; | |
929 | } | |
930 | } | |
931 | ||
67e6ba46 NC |
932 | static void |
933 | arm_target_help (void) | |
934 | { | |
935 | int i; | |
936 | static int columns = 0; | |
937 | int remaining; | |
938 | ||
939 | /* If we have not done so already, obtain the desired maximum width of | |
940 | the output. Note - this is a duplication of the code at the start of | |
941 | gcc/opts.c:print_specific_help() - the two copies should probably be | |
942 | replaced by a single function. */ | |
943 | if (columns == 0) | |
944 | { | |
945 | const char *p; | |
946 | ||
947 | GET_ENVIRONMENT (p, "COLUMNS"); | |
948 | if (p != NULL) | |
949 | { | |
950 | int value = atoi (p); | |
951 | ||
952 | if (value > 0) | |
953 | columns = value; | |
954 | } | |
955 | ||
956 | if (columns == 0) | |
957 | /* Use a reasonable default. */ | |
958 | columns = 80; | |
959 | } | |
960 | ||
961 | printf (" Known ARM CPUs (for use with the -mcpu= and -mtune= options):\n"); | |
962 | ||
963 | /* The - 2 is because we know that the last entry in the array is NULL. */ | |
964 | i = ARRAY_SIZE (all_cores) - 2; | |
965 | gcc_assert (i > 0); | |
966 | printf (" %s", all_cores[i].name); | |
967 | remaining = columns - (strlen (all_cores[i].name) + 4); | |
968 | gcc_assert (remaining >= 0); | |
969 | ||
970 | while (i--) | |
971 | { | |
972 | int len = strlen (all_cores[i].name); | |
973 | ||
974 | if (remaining > len + 2) | |
975 | { | |
976 | printf (", %s", all_cores[i].name); | |
977 | remaining -= len + 2; | |
978 | } | |
979 | else | |
980 | { | |
981 | if (remaining > 0) | |
982 | printf (","); | |
983 | printf ("\n %s", all_cores[i].name); | |
984 | remaining = columns - (len + 4); | |
985 | } | |
986 | } | |
987 | ||
988 | printf ("\n\n Known ARM architectures (for use with the -march= option):\n"); | |
989 | ||
990 | i = ARRAY_SIZE (all_architectures) - 2; | |
991 | gcc_assert (i > 0); | |
992 | ||
993 | printf (" %s", all_architectures[i].name); | |
994 | remaining = columns - (strlen (all_architectures[i].name) + 4); | |
995 | gcc_assert (remaining >= 0); | |
996 | ||
997 | while (i--) | |
998 | { | |
999 | int len = strlen (all_architectures[i].name); | |
1000 | ||
1001 | if (remaining > len + 2) | |
1002 | { | |
1003 | printf (", %s", all_architectures[i].name); | |
1004 | remaining -= len + 2; | |
1005 | } | |
1006 | else | |
1007 | { | |
1008 | if (remaining > 0) | |
1009 | printf (","); | |
1010 | printf ("\n %s", all_architectures[i].name); | |
1011 | remaining = columns - (len + 4); | |
1012 | } | |
1013 | } | |
1014 | printf ("\n"); | |
1015 | ||
1016 | } | |
1017 | ||
2b835d68 RE |
1018 | /* Fix up any incompatible options that the user has specified. |
1019 | This has now turned into a maze. */ | |
1020 | void | |
e32bac5b | 1021 | arm_override_options (void) |
2b835d68 | 1022 | { |
ed4c4348 | 1023 | unsigned i; |
e6fffade | 1024 | enum processor_type target_arch_cpu = arm_none; |
9b66ebb1 | 1025 | |
f5a1b0d2 | 1026 | /* Set up the flags based on the cpu/architecture selected by the user. */ |
b6a1cbae | 1027 | for (i = ARRAY_SIZE (arm_select); i--;) |
bd9c7e23 | 1028 | { |
f5a1b0d2 | 1029 | struct arm_cpu_select * ptr = arm_select + i; |
f676971a | 1030 | |
f5a1b0d2 | 1031 | if (ptr->string != NULL && ptr->string[0] != '\0') |
bd9c7e23 | 1032 | { |
13bd191d | 1033 | const struct processors * sel; |
bd9c7e23 | 1034 | |
5895f793 | 1035 | for (sel = ptr->processors; sel->name != NULL; sel++) |
f5a1b0d2 | 1036 | if (streq (ptr->string, sel->name)) |
bd9c7e23 | 1037 | { |
78011587 | 1038 | /* Set the architecture define. */ |
e6fffade | 1039 | if (i != ARM_OPT_SET_TUNE) |
78011587 PB |
1040 | sprintf (arm_arch_name, "__ARM_ARCH_%s__", sel->arch); |
1041 | ||
9b66ebb1 PB |
1042 | /* Determine the processor core for which we should |
1043 | tune code-generation. */ | |
1044 | if (/* -mcpu= is a sensible default. */ | |
e6fffade | 1045 | i == ARM_OPT_SET_CPU |
9b66ebb1 | 1046 | /* -mtune= overrides -mcpu= and -march=. */ |
e6fffade | 1047 | || i == ARM_OPT_SET_TUNE) |
9b66ebb1 PB |
1048 | arm_tune = (enum processor_type) (sel - ptr->processors); |
1049 | ||
e6fffade RE |
1050 | /* Remember the CPU associated with this architecture. |
1051 | If no other option is used to set the CPU type, | |
1052 | we'll use this to guess the most suitable tuning | |
1053 | options. */ | |
1054 | if (i == ARM_OPT_SET_ARCH) | |
1055 | target_arch_cpu = sel->core; | |
e0b92319 | 1056 | |
e6fffade | 1057 | if (i != ARM_OPT_SET_TUNE) |
b111229a | 1058 | { |
aec3cfba NC |
1059 | /* If we have been given an architecture and a processor |
1060 | make sure that they are compatible. We only generate | |
1061 | a warning though, and we prefer the CPU over the | |
6354dc9b | 1062 | architecture. */ |
aec3cfba | 1063 | if (insn_flags != 0 && (insn_flags ^ sel->flags)) |
d4ee4d25 | 1064 | warning (0, "switch -mcpu=%s conflicts with -march= switch", |
aec3cfba | 1065 | ptr->string); |
f676971a | 1066 | |
aec3cfba | 1067 | insn_flags = sel->flags; |
b111229a | 1068 | } |
f676971a | 1069 | |
bd9c7e23 RE |
1070 | break; |
1071 | } | |
1072 | ||
1073 | if (sel->name == NULL) | |
1074 | error ("bad value (%s) for %s switch", ptr->string, ptr->name); | |
1075 | } | |
1076 | } | |
f676971a | 1077 | |
e6fffade RE |
1078 | /* Guess the tuning options from the architecture if necessary. */ |
1079 | if (arm_tune == arm_none) | |
1080 | arm_tune = target_arch_cpu; | |
1081 | ||
f5a1b0d2 | 1082 | /* If the user did not specify a processor, choose one for them. */ |
aec3cfba | 1083 | if (insn_flags == 0) |
f5a1b0d2 | 1084 | { |
8b60264b | 1085 | const struct processors * sel; |
aec3cfba | 1086 | unsigned int sought; |
78011587 | 1087 | enum processor_type cpu; |
aec3cfba | 1088 | |
78011587 PB |
1089 | cpu = TARGET_CPU_DEFAULT; |
1090 | if (cpu == arm_none) | |
1091 | { | |
1092 | #ifdef SUBTARGET_CPU_DEFAULT | |
1093 | /* Use the subtarget default CPU if none was specified by | |
1094 | configure. */ | |
1095 | cpu = SUBTARGET_CPU_DEFAULT; | |
1096 | #endif | |
1097 | /* Default to ARM6. */ | |
1098 | if (cpu == arm_none) | |
1099 | cpu = arm6; | |
1100 | } | |
1101 | sel = &all_cores[cpu]; | |
aec3cfba NC |
1102 | |
1103 | insn_flags = sel->flags; | |
9b66ebb1 | 1104 | |
aec3cfba NC |
1105 | /* Now check to see if the user has specified some command line |
1106 | switch that require certain abilities from the cpu. */ | |
1107 | sought = 0; | |
f676971a | 1108 | |
d5b7b3ae | 1109 | if (TARGET_INTERWORK || TARGET_THUMB) |
f5a1b0d2 | 1110 | { |
aec3cfba | 1111 | sought |= (FL_THUMB | FL_MODE32); |
f676971a | 1112 | |
d5b7b3ae | 1113 | /* There are no ARM processors that support both APCS-26 and |
aec3cfba NC |
1114 | interworking. Therefore we force FL_MODE26 to be removed |
1115 | from insn_flags here (if it was set), so that the search | |
1116 | below will always be able to find a compatible processor. */ | |
5895f793 | 1117 | insn_flags &= ~FL_MODE26; |
f5a1b0d2 | 1118 | } |
f676971a | 1119 | |
aec3cfba | 1120 | if (sought != 0 && ((sought & insn_flags) != sought)) |
f5a1b0d2 | 1121 | { |
aec3cfba NC |
1122 | /* Try to locate a CPU type that supports all of the abilities |
1123 | of the default CPU, plus the extra abilities requested by | |
1124 | the user. */ | |
5895f793 | 1125 | for (sel = all_cores; sel->name != NULL; sel++) |
aec3cfba | 1126 | if ((sel->flags & sought) == (sought | insn_flags)) |
f5a1b0d2 NC |
1127 | break; |
1128 | ||
1129 | if (sel->name == NULL) | |
aec3cfba | 1130 | { |
0977774b | 1131 | unsigned current_bit_count = 0; |
8b60264b | 1132 | const struct processors * best_fit = NULL; |
f676971a | 1133 | |
aec3cfba NC |
1134 | /* Ideally we would like to issue an error message here |
1135 | saying that it was not possible to find a CPU compatible | |
1136 | with the default CPU, but which also supports the command | |
1137 | line options specified by the programmer, and so they | |
1138 | ought to use the -mcpu=<name> command line option to | |
1139 | override the default CPU type. | |
1140 | ||
61f0ccff RE |
1141 | If we cannot find a cpu that has both the |
1142 | characteristics of the default cpu and the given | |
1143 | command line options we scan the array again looking | |
1144 | for a best match. */ | |
5895f793 | 1145 | for (sel = all_cores; sel->name != NULL; sel++) |
aec3cfba NC |
1146 | if ((sel->flags & sought) == sought) |
1147 | { | |
0977774b | 1148 | unsigned count; |
aec3cfba NC |
1149 | |
1150 | count = bit_count (sel->flags & insn_flags); | |
1151 | ||
1152 | if (count >= current_bit_count) | |
1153 | { | |
1154 | best_fit = sel; | |
1155 | current_bit_count = count; | |
1156 | } | |
1157 | } | |
f5a1b0d2 | 1158 | |
e6d29d15 NS |
1159 | gcc_assert (best_fit); |
1160 | sel = best_fit; | |
aec3cfba NC |
1161 | } |
1162 | ||
1163 | insn_flags = sel->flags; | |
f5a1b0d2 | 1164 | } |
78011587 | 1165 | sprintf (arm_arch_name, "__ARM_ARCH_%s__", sel->arch); |
6c6aa1af | 1166 | arm_default_cpu = (enum processor_type) (sel - all_cores); |
9b66ebb1 | 1167 | if (arm_tune == arm_none) |
6c6aa1af | 1168 | arm_tune = arm_default_cpu; |
f5a1b0d2 | 1169 | } |
f676971a | 1170 | |
9b66ebb1 PB |
1171 | /* The processor for which we should tune should now have been |
1172 | chosen. */ | |
e6d29d15 | 1173 | gcc_assert (arm_tune != arm_none); |
f676971a | 1174 | |
9b66ebb1 | 1175 | tune_flags = all_cores[(int)arm_tune].flags; |
21b5653c RE |
1176 | if (optimize_size) |
1177 | targetm.rtx_costs = arm_size_rtx_costs; | |
1178 | else | |
1179 | targetm.rtx_costs = all_cores[(int)arm_tune].rtx_costs; | |
e26053d1 | 1180 | |
f5a1b0d2 NC |
1181 | /* Make sure that the processor choice does not conflict with any of the |
1182 | other command line choices. */ | |
5b3e6663 PB |
1183 | if (TARGET_ARM && !(insn_flags & FL_NOTM)) |
1184 | error ("target CPU does not support ARM mode"); | |
1185 | ||
6cfc7210 | 1186 | if (TARGET_INTERWORK && !(insn_flags & FL_THUMB)) |
f5a1b0d2 | 1187 | { |
d4ee4d25 | 1188 | warning (0, "target CPU does not support interworking" ); |
c54c7322 | 1189 | target_flags &= ~MASK_INTERWORK; |
f5a1b0d2 | 1190 | } |
f676971a | 1191 | |
d5b7b3ae RE |
1192 | if (TARGET_THUMB && !(insn_flags & FL_THUMB)) |
1193 | { | |
d4ee4d25 | 1194 | warning (0, "target CPU does not support THUMB instructions"); |
c54c7322 | 1195 | target_flags &= ~MASK_THUMB; |
d5b7b3ae RE |
1196 | } |
1197 | ||
1198 | if (TARGET_APCS_FRAME && TARGET_THUMB) | |
1199 | { | |
d4ee4d25 | 1200 | /* warning (0, "ignoring -mapcs-frame because -mthumb was used"); */ |
c54c7322 | 1201 | target_flags &= ~MASK_APCS_FRAME; |
d5b7b3ae | 1202 | } |
d19fb8e3 | 1203 | |
da8ce8be RE |
1204 | /* Callee super interworking implies thumb interworking. Adding |
1205 | this to the flags here simplifies the logic elsewhere. */ | |
1206 | if (TARGET_THUMB && TARGET_CALLEE_INTERWORKING) | |
1207 | target_flags |= MASK_INTERWORK; | |
1208 | ||
d5b7b3ae RE |
1209 | /* TARGET_BACKTRACE calls leaf_function_p, which causes a crash if done |
1210 | from here where no function is being compiled currently. */ | |
c54c7322 | 1211 | if ((TARGET_TPCS_FRAME || TARGET_TPCS_LEAF_FRAME) && TARGET_ARM) |
d4ee4d25 | 1212 | warning (0, "enabling backtrace support is only meaningful when compiling for the Thumb"); |
d5b7b3ae RE |
1213 | |
1214 | if (TARGET_ARM && TARGET_CALLEE_INTERWORKING) | |
d4ee4d25 | 1215 | warning (0, "enabling callee interworking support is only meaningful when compiling for the Thumb"); |
d5b7b3ae RE |
1216 | |
1217 | if (TARGET_ARM && TARGET_CALLER_INTERWORKING) | |
d4ee4d25 | 1218 | warning (0, "enabling caller interworking support is only meaningful when compiling for the Thumb"); |
d5b7b3ae | 1219 | |
5895f793 | 1220 | if (TARGET_APCS_STACK && !TARGET_APCS_FRAME) |
f5a1b0d2 | 1221 | { |
d4ee4d25 | 1222 | warning (0, "-mapcs-stack-check incompatible with -mno-apcs-frame"); |
c54c7322 | 1223 | target_flags |= MASK_APCS_FRAME; |
f5a1b0d2 | 1224 | } |
f676971a | 1225 | |
2b835d68 | 1226 | if (TARGET_POKE_FUNCTION_NAME) |
c54c7322 | 1227 | target_flags |= MASK_APCS_FRAME; |
f676971a | 1228 | |
2b835d68 | 1229 | if (TARGET_APCS_REENT && flag_pic) |
400500c4 | 1230 | error ("-fpic and -mapcs-reent are incompatible"); |
f676971a | 1231 | |
2b835d68 | 1232 | if (TARGET_APCS_REENT) |
d4ee4d25 | 1233 | warning (0, "APCS reentrant code not supported. Ignored"); |
f676971a | 1234 | |
d5b7b3ae RE |
1235 | /* If this target is normally configured to use APCS frames, warn if they |
1236 | are turned off and debugging is turned on. */ | |
1237 | if (TARGET_ARM | |
1238 | && write_symbols != NO_DEBUG | |
5895f793 | 1239 | && !TARGET_APCS_FRAME |
c54c7322 | 1240 | && (TARGET_DEFAULT & MASK_APCS_FRAME)) |
d4ee4d25 | 1241 | warning (0, "-g with -mno-apcs-frame may not give sensible debugging"); |
f676971a | 1242 | |
2b835d68 | 1243 | if (TARGET_APCS_FLOAT) |
d4ee4d25 | 1244 | warning (0, "passing floating point arguments in fp regs not yet supported"); |
f676971a | 1245 | |
4912a07c | 1246 | /* Initialize boolean versions of the flags, for use in the arm.md file. */ |
9b66ebb1 PB |
1247 | arm_arch3m = (insn_flags & FL_ARCH3M) != 0; |
1248 | arm_arch4 = (insn_flags & FL_ARCH4) != 0; | |
68d560d4 | 1249 | arm_arch4t = arm_arch4 & ((insn_flags & FL_THUMB) != 0); |
9b66ebb1 PB |
1250 | arm_arch5 = (insn_flags & FL_ARCH5) != 0; |
1251 | arm_arch5e = (insn_flags & FL_ARCH5E) != 0; | |
1252 | arm_arch6 = (insn_flags & FL_ARCH6) != 0; | |
d3585b76 | 1253 | arm_arch6k = (insn_flags & FL_ARCH6K) != 0; |
5b3e6663 PB |
1254 | arm_arch_notm = (insn_flags & FL_NOTM) != 0; |
1255 | arm_arch_thumb2 = (insn_flags & FL_THUMB2) != 0; | |
9b66ebb1 | 1256 | arm_arch_xscale = (insn_flags & FL_XSCALE) != 0; |
78011587 | 1257 | arm_arch_cirrus = (insn_flags & FL_CIRRUS) != 0; |
9b66ebb1 PB |
1258 | |
1259 | arm_ld_sched = (tune_flags & FL_LDSCHED) != 0; | |
abac3b49 | 1260 | arm_tune_strongarm = (tune_flags & FL_STRONG) != 0; |
9b66ebb1 | 1261 | thumb_code = (TARGET_ARM == 0); |
abac3b49 | 1262 | arm_tune_wbuf = (tune_flags & FL_WBUF) != 0; |
9b66ebb1 PB |
1263 | arm_tune_xscale = (tune_flags & FL_XSCALE) != 0; |
1264 | arm_arch_iwmmxt = (insn_flags & FL_IWMMXT) != 0; | |
5b3e6663 | 1265 | arm_arch_hwdiv = (insn_flags & FL_DIV) != 0; |
5a9335ef | 1266 | |
68d560d4 RE |
1267 | /* V5 code we generate is completely interworking capable, so we turn off |
1268 | TARGET_INTERWORK here to avoid many tests later on. */ | |
2ad4dcf9 RE |
1269 | |
1270 | /* XXX However, we must pass the right pre-processor defines to CPP | |
1271 | or GLD can get confused. This is a hack. */ | |
1272 | if (TARGET_INTERWORK) | |
1273 | arm_cpp_interwork = 1; | |
1274 | ||
68d560d4 | 1275 | if (arm_arch5) |
c54c7322 | 1276 | target_flags &= ~MASK_INTERWORK; |
68d560d4 | 1277 | |
5848830f PB |
1278 | if (target_abi_name) |
1279 | { | |
1280 | for (i = 0; i < ARRAY_SIZE (arm_all_abis); i++) | |
1281 | { | |
1282 | if (streq (arm_all_abis[i].name, target_abi_name)) | |
1283 | { | |
1284 | arm_abi = arm_all_abis[i].abi_type; | |
1285 | break; | |
1286 | } | |
1287 | } | |
1288 | if (i == ARRAY_SIZE (arm_all_abis)) | |
1289 | error ("invalid ABI option: -mabi=%s", target_abi_name); | |
1290 | } | |
1291 | else | |
c805f22e | 1292 | arm_abi = ARM_DEFAULT_ABI; |
5848830f PB |
1293 | |
1294 | if (TARGET_IWMMXT && !ARM_DOUBLEWORD_ALIGN) | |
1295 | error ("iwmmxt requires an AAPCS compatible ABI for proper operation"); | |
1296 | ||
1297 | if (TARGET_IWMMXT_ABI && !TARGET_IWMMXT) | |
1298 | error ("iwmmxt abi requires an iwmmxt capable cpu"); | |
6f7ebcbb | 1299 | |
9b66ebb1 PB |
1300 | arm_fp_model = ARM_FP_MODEL_UNKNOWN; |
1301 | if (target_fpu_name == NULL && target_fpe_name != NULL) | |
9b6b54e2 | 1302 | { |
9b66ebb1 PB |
1303 | if (streq (target_fpe_name, "2")) |
1304 | target_fpu_name = "fpe2"; | |
1305 | else if (streq (target_fpe_name, "3")) | |
1306 | target_fpu_name = "fpe3"; | |
1307 | else | |
1308 | error ("invalid floating point emulation option: -mfpe=%s", | |
1309 | target_fpe_name); | |
1310 | } | |
1311 | if (target_fpu_name != NULL) | |
1312 | { | |
1313 | /* The user specified a FPU. */ | |
1314 | for (i = 0; i < ARRAY_SIZE (all_fpus); i++) | |
1315 | { | |
1316 | if (streq (all_fpus[i].name, target_fpu_name)) | |
1317 | { | |
1318 | arm_fpu_arch = all_fpus[i].fpu; | |
1319 | arm_fpu_tune = arm_fpu_arch; | |
1320 | arm_fp_model = fp_model_for_fpu[arm_fpu_arch]; | |
1321 | break; | |
1322 | } | |
1323 | } | |
1324 | if (arm_fp_model == ARM_FP_MODEL_UNKNOWN) | |
1325 | error ("invalid floating point option: -mfpu=%s", target_fpu_name); | |
9b6b54e2 NC |
1326 | } |
1327 | else | |
2b835d68 | 1328 | { |
9b66ebb1 | 1329 | #ifdef FPUTYPE_DEFAULT |
78011587 | 1330 | /* Use the default if it is specified for this platform. */ |
9b66ebb1 PB |
1331 | arm_fpu_arch = FPUTYPE_DEFAULT; |
1332 | arm_fpu_tune = FPUTYPE_DEFAULT; | |
1333 | #else | |
1334 | /* Pick one based on CPU type. */ | |
78011587 | 1335 | /* ??? Some targets assume FPA is the default. |
9b66ebb1 PB |
1336 | if ((insn_flags & FL_VFP) != 0) |
1337 | arm_fpu_arch = FPUTYPE_VFP; | |
78011587 PB |
1338 | else |
1339 | */ | |
1340 | if (arm_arch_cirrus) | |
9b66ebb1 PB |
1341 | arm_fpu_arch = FPUTYPE_MAVERICK; |
1342 | else | |
29ad9694 | 1343 | arm_fpu_arch = FPUTYPE_FPA_EMU2; |
9b66ebb1 PB |
1344 | #endif |
1345 | if (tune_flags & FL_CO_PROC && arm_fpu_arch == FPUTYPE_FPA_EMU2) | |
1346 | arm_fpu_tune = FPUTYPE_FPA; | |
2b835d68 | 1347 | else |
9b66ebb1 PB |
1348 | arm_fpu_tune = arm_fpu_arch; |
1349 | arm_fp_model = fp_model_for_fpu[arm_fpu_arch]; | |
e6d29d15 | 1350 | gcc_assert (arm_fp_model != ARM_FP_MODEL_UNKNOWN); |
9b66ebb1 PB |
1351 | } |
1352 | ||
1353 | if (target_float_abi_name != NULL) | |
1354 | { | |
1355 | /* The user specified a FP ABI. */ | |
1356 | for (i = 0; i < ARRAY_SIZE (all_float_abis); i++) | |
1357 | { | |
1358 | if (streq (all_float_abis[i].name, target_float_abi_name)) | |
1359 | { | |
1360 | arm_float_abi = all_float_abis[i].abi_type; | |
1361 | break; | |
1362 | } | |
1363 | } | |
1364 | if (i == ARRAY_SIZE (all_float_abis)) | |
1365 | error ("invalid floating point abi: -mfloat-abi=%s", | |
1366 | target_float_abi_name); | |
2b835d68 | 1367 | } |
3d8532aa PB |
1368 | else |
1369 | arm_float_abi = TARGET_DEFAULT_FLOAT_ABI; | |
9b66ebb1 | 1370 | |
72cdc543 PB |
1371 | if (arm_float_abi == ARM_FLOAT_ABI_HARD && TARGET_VFP) |
1372 | sorry ("-mfloat-abi=hard and VFP"); | |
1373 | ||
87b24aaf PB |
1374 | /* FPA and iWMMXt are incompatible because the insn encodings overlap. |
1375 | VFP and iWMMXt can theoretically coexist, but it's unlikely such silicon | |
1376 | will ever exist. GCC makes no attempt to support this combination. */ | |
1377 | if (TARGET_IWMMXT && !TARGET_SOFT_FLOAT) | |
1378 | sorry ("iWMMXt and hardware floating point"); | |
1379 | ||
5b3e6663 PB |
1380 | /* ??? iWMMXt insn patterns need auditing for Thumb-2. */ |
1381 | if (TARGET_THUMB2 && TARGET_IWMMXT) | |
1382 | sorry ("Thumb-2 iWMMXt"); | |
1383 | ||
9b66ebb1 PB |
1384 | /* If soft-float is specified then don't use FPU. */ |
1385 | if (TARGET_SOFT_FLOAT) | |
1386 | arm_fpu_arch = FPUTYPE_NONE; | |
f676971a | 1387 | |
f5a1b0d2 NC |
1388 | /* For arm2/3 there is no need to do any scheduling if there is only |
1389 | a floating point emulator, or we are doing software floating-point. */ | |
9b66ebb1 PB |
1390 | if ((TARGET_SOFT_FLOAT |
1391 | || arm_fpu_tune == FPUTYPE_FPA_EMU2 | |
1392 | || arm_fpu_tune == FPUTYPE_FPA_EMU3) | |
ed0e6530 | 1393 | && (tune_flags & FL_MODE32) == 0) |
f5a1b0d2 | 1394 | flag_schedule_insns = flag_schedule_insns_after_reload = 0; |
f676971a | 1395 | |
d3585b76 DJ |
1396 | if (target_thread_switch) |
1397 | { | |
1398 | if (strcmp (target_thread_switch, "soft") == 0) | |
1399 | target_thread_pointer = TP_SOFT; | |
1400 | else if (strcmp (target_thread_switch, "auto") == 0) | |
1401 | target_thread_pointer = TP_AUTO; | |
1402 | else if (strcmp (target_thread_switch, "cp15") == 0) | |
1403 | target_thread_pointer = TP_CP15; | |
1404 | else | |
1405 | error ("invalid thread pointer option: -mtp=%s", target_thread_switch); | |
1406 | } | |
1407 | ||
1408 | /* Use the cp15 method if it is available. */ | |
1409 | if (target_thread_pointer == TP_AUTO) | |
1410 | { | |
1411 | if (arm_arch6k && !TARGET_THUMB) | |
1412 | target_thread_pointer = TP_CP15; | |
1413 | else | |
1414 | target_thread_pointer = TP_SOFT; | |
1415 | } | |
1416 | ||
5b3e6663 PB |
1417 | if (TARGET_HARD_TP && TARGET_THUMB1) |
1418 | error ("can not use -mtp=cp15 with 16-bit Thumb"); | |
d3585b76 | 1419 | |
5848830f | 1420 | /* Override the default structure alignment for AAPCS ABI. */ |
077fc835 | 1421 | if (TARGET_AAPCS_BASED) |
5848830f PB |
1422 | arm_structure_size_boundary = 8; |
1423 | ||
b355a481 NC |
1424 | if (structure_size_string != NULL) |
1425 | { | |
1426 | int size = strtol (structure_size_string, NULL, 0); | |
5848830f PB |
1427 | |
1428 | if (size == 8 || size == 32 | |
1429 | || (ARM_DOUBLEWORD_ALIGN && size == 64)) | |
b355a481 NC |
1430 | arm_structure_size_boundary = size; |
1431 | else | |
d4ee4d25 | 1432 | warning (0, "structure size boundary can only be set to %s", |
5848830f | 1433 | ARM_DOUBLEWORD_ALIGN ? "8, 32 or 64": "8 or 32"); |
b355a481 | 1434 | } |
ed0e6530 | 1435 | |
9403b7f7 RS |
1436 | if (!TARGET_ARM && TARGET_VXWORKS_RTP && flag_pic) |
1437 | { | |
1438 | error ("RTP PIC is incompatible with Thumb"); | |
1439 | flag_pic = 0; | |
1440 | } | |
1441 | ||
c147eacb PB |
1442 | /* If stack checking is disabled, we can use r10 as the PIC register, |
1443 | which keeps r9 available. The EABI specifies r9 as the PIC register. */ | |
1444 | if (flag_pic && TARGET_SINGLE_PIC_BASE) | |
9403b7f7 RS |
1445 | { |
1446 | if (TARGET_VXWORKS_RTP) | |
1447 | warning (0, "RTP PIC is incompatible with -msingle-pic-base"); | |
1448 | arm_pic_register = (TARGET_APCS_STACK || TARGET_AAPCS_BASED) ? 9 : 10; | |
1449 | } | |
1450 | ||
1451 | if (flag_pic && TARGET_VXWORKS_RTP) | |
1452 | arm_pic_register = 9; | |
c147eacb | 1453 | |
ed0e6530 PB |
1454 | if (arm_pic_register_string != NULL) |
1455 | { | |
5b43fed1 | 1456 | int pic_register = decode_reg_name (arm_pic_register_string); |
e26053d1 | 1457 | |
5895f793 | 1458 | if (!flag_pic) |
d4ee4d25 | 1459 | warning (0, "-mpic-register= is useless without -fpic"); |
ed0e6530 | 1460 | |
ed0e6530 | 1461 | /* Prevent the user from choosing an obviously stupid PIC register. */ |
5b43fed1 RH |
1462 | else if (pic_register < 0 || call_used_regs[pic_register] |
1463 | || pic_register == HARD_FRAME_POINTER_REGNUM | |
1464 | || pic_register == STACK_POINTER_REGNUM | |
9403b7f7 RS |
1465 | || pic_register >= PC_REGNUM |
1466 | || (TARGET_VXWORKS_RTP | |
1467 | && (unsigned int) pic_register != arm_pic_register)) | |
c725bd79 | 1468 | error ("unable to use '%s' for PIC register", arm_pic_register_string); |
ed0e6530 PB |
1469 | else |
1470 | arm_pic_register = pic_register; | |
1471 | } | |
d5b7b3ae | 1472 | |
5b3e6663 | 1473 | /* ??? We might want scheduling for thumb2. */ |
d5b7b3ae RE |
1474 | if (TARGET_THUMB && flag_schedule_insns) |
1475 | { | |
1476 | /* Don't warn since it's on by default in -O2. */ | |
1477 | flag_schedule_insns = 0; | |
1478 | } | |
1479 | ||
f5a1b0d2 | 1480 | if (optimize_size) |
be03ccc9 | 1481 | { |
577d6328 | 1482 | arm_constant_limit = 1; |
be03ccc9 NP |
1483 | |
1484 | /* If optimizing for size, bump the number of instructions that we | |
d6b4baa4 | 1485 | are prepared to conditionally execute (even on a StrongARM). */ |
be03ccc9 NP |
1486 | max_insns_skipped = 6; |
1487 | } | |
1488 | else | |
1489 | { | |
1490 | /* For processors with load scheduling, it never costs more than | |
1491 | 2 cycles to load a constant, and the load scheduler may well | |
1492 | reduce that to 1. */ | |
2075b05d | 1493 | if (arm_ld_sched) |
be03ccc9 NP |
1494 | arm_constant_limit = 1; |
1495 | ||
1496 | /* On XScale the longer latency of a load makes it more difficult | |
1497 | to achieve a good schedule, so it's faster to synthesize | |
d6b4baa4 | 1498 | constants that can be done in two insns. */ |
be03ccc9 NP |
1499 | if (arm_tune_xscale) |
1500 | arm_constant_limit = 2; | |
1501 | ||
1502 | /* StrongARM has early execution of branches, so a sequence | |
1503 | that is worth skipping is shorter. */ | |
abac3b49 | 1504 | if (arm_tune_strongarm) |
be03ccc9 NP |
1505 | max_insns_skipped = 3; |
1506 | } | |
92a432f4 RE |
1507 | |
1508 | /* Register global variables with the garbage collector. */ | |
1509 | arm_add_gc_roots (); | |
1510 | } | |
1511 | ||
1512 | static void | |
e32bac5b | 1513 | arm_add_gc_roots (void) |
92a432f4 | 1514 | { |
c7319d87 RE |
1515 | gcc_obstack_init(&minipool_obstack); |
1516 | minipool_startobj = (char *) obstack_alloc (&minipool_obstack, 0); | |
2b835d68 | 1517 | } |
cce8749e | 1518 | \f |
6d3d9133 NC |
1519 | /* A table of known ARM exception types. |
1520 | For use with the interrupt function attribute. */ | |
1521 | ||
1522 | typedef struct | |
1523 | { | |
8b60264b KG |
1524 | const char *const arg; |
1525 | const unsigned long return_value; | |
6d3d9133 NC |
1526 | } |
1527 | isr_attribute_arg; | |
1528 | ||
8b60264b | 1529 | static const isr_attribute_arg isr_attribute_args [] = |
6d3d9133 NC |
1530 | { |
1531 | { "IRQ", ARM_FT_ISR }, | |
1532 | { "irq", ARM_FT_ISR }, | |
1533 | { "FIQ", ARM_FT_FIQ }, | |
1534 | { "fiq", ARM_FT_FIQ }, | |
1535 | { "ABORT", ARM_FT_ISR }, | |
1536 | { "abort", ARM_FT_ISR }, | |
1537 | { "ABORT", ARM_FT_ISR }, | |
1538 | { "abort", ARM_FT_ISR }, | |
1539 | { "UNDEF", ARM_FT_EXCEPTION }, | |
1540 | { "undef", ARM_FT_EXCEPTION }, | |
1541 | { "SWI", ARM_FT_EXCEPTION }, | |
1542 | { "swi", ARM_FT_EXCEPTION }, | |
1543 | { NULL, ARM_FT_NORMAL } | |
1544 | }; | |
1545 | ||
1546 | /* Returns the (interrupt) function type of the current | |
1547 | function, or ARM_FT_UNKNOWN if the type cannot be determined. */ | |
1548 | ||
1549 | static unsigned long | |
e32bac5b | 1550 | arm_isr_value (tree argument) |
6d3d9133 | 1551 | { |
8b60264b | 1552 | const isr_attribute_arg * ptr; |
1d6e90ac | 1553 | const char * arg; |
6d3d9133 | 1554 | |
5b3e6663 PB |
1555 | if (!arm_arch_notm) |
1556 | return ARM_FT_NORMAL | ARM_FT_STACKALIGN; | |
1557 | ||
6d3d9133 NC |
1558 | /* No argument - default to IRQ. */ |
1559 | if (argument == NULL_TREE) | |
1560 | return ARM_FT_ISR; | |
1561 | ||
1562 | /* Get the value of the argument. */ | |
1563 | if (TREE_VALUE (argument) == NULL_TREE | |
1564 | || TREE_CODE (TREE_VALUE (argument)) != STRING_CST) | |
1565 | return ARM_FT_UNKNOWN; | |
1566 | ||
1567 | arg = TREE_STRING_POINTER (TREE_VALUE (argument)); | |
1568 | ||
1569 | /* Check it against the list of known arguments. */ | |
5a9335ef | 1570 | for (ptr = isr_attribute_args; ptr->arg != NULL; ptr++) |
1d6e90ac NC |
1571 | if (streq (arg, ptr->arg)) |
1572 | return ptr->return_value; | |
6d3d9133 | 1573 | |
05713b80 | 1574 | /* An unrecognized interrupt type. */ |
6d3d9133 NC |
1575 | return ARM_FT_UNKNOWN; |
1576 | } | |
1577 | ||
1578 | /* Computes the type of the current function. */ | |
1579 | ||
1580 | static unsigned long | |
e32bac5b | 1581 | arm_compute_func_type (void) |
6d3d9133 NC |
1582 | { |
1583 | unsigned long type = ARM_FT_UNKNOWN; | |
1584 | tree a; | |
1585 | tree attr; | |
f676971a | 1586 | |
e6d29d15 | 1587 | gcc_assert (TREE_CODE (current_function_decl) == FUNCTION_DECL); |
6d3d9133 NC |
1588 | |
1589 | /* Decide if the current function is volatile. Such functions | |
1590 | never return, and many memory cycles can be saved by not storing | |
1591 | register values that will never be needed again. This optimization | |
1592 | was added to speed up context switching in a kernel application. */ | |
1593 | if (optimize > 0 | |
cf1955dc PB |
1594 | && (TREE_NOTHROW (current_function_decl) |
1595 | || !(flag_unwind_tables | |
1596 | || (flag_exceptions && !USING_SJLJ_EXCEPTIONS))) | |
6d3d9133 NC |
1597 | && TREE_THIS_VOLATILE (current_function_decl)) |
1598 | type |= ARM_FT_VOLATILE; | |
f676971a | 1599 | |
6de9cd9a | 1600 | if (cfun->static_chain_decl != NULL) |
6d3d9133 NC |
1601 | type |= ARM_FT_NESTED; |
1602 | ||
91d231cb | 1603 | attr = DECL_ATTRIBUTES (current_function_decl); |
f676971a | 1604 | |
6d3d9133 NC |
1605 | a = lookup_attribute ("naked", attr); |
1606 | if (a != NULL_TREE) | |
1607 | type |= ARM_FT_NAKED; | |
1608 | ||
c9ca9b88 PB |
1609 | a = lookup_attribute ("isr", attr); |
1610 | if (a == NULL_TREE) | |
1611 | a = lookup_attribute ("interrupt", attr); | |
f676971a | 1612 | |
c9ca9b88 PB |
1613 | if (a == NULL_TREE) |
1614 | type |= TARGET_INTERWORK ? ARM_FT_INTERWORKED : ARM_FT_NORMAL; | |
6d3d9133 | 1615 | else |
c9ca9b88 | 1616 | type |= arm_isr_value (TREE_VALUE (a)); |
f676971a | 1617 | |
6d3d9133 NC |
1618 | return type; |
1619 | } | |
1620 | ||
1621 | /* Returns the type of the current function. */ | |
1622 | ||
1623 | unsigned long | |
e32bac5b | 1624 | arm_current_func_type (void) |
6d3d9133 NC |
1625 | { |
1626 | if (ARM_FUNC_TYPE (cfun->machine->func_type) == ARM_FT_UNKNOWN) | |
1627 | cfun->machine->func_type = arm_compute_func_type (); | |
1628 | ||
1629 | return cfun->machine->func_type; | |
1630 | } | |
007e61c2 PB |
1631 | |
1632 | bool | |
1633 | arm_allocate_stack_slots_for_args (void) | |
1634 | { | |
1635 | /* Naked functions should not allocate stack slots for arguments. */ | |
1636 | return !IS_NAKED (arm_current_func_type ()); | |
1637 | } | |
1638 | ||
6d3d9133 | 1639 | \f |
f676971a | 1640 | /* Return 1 if it is possible to return using a single instruction. |
a72d4945 RE |
1641 | If SIBLING is non-null, this is a test for a return before a sibling |
1642 | call. SIBLING is the call insn, so we can examine its register usage. */ | |
6d3d9133 | 1643 | |
ff9940b0 | 1644 | int |
a72d4945 | 1645 | use_return_insn (int iscond, rtx sibling) |
ff9940b0 RE |
1646 | { |
1647 | int regno; | |
9b598fa0 | 1648 | unsigned int func_type; |
d5db54a1 | 1649 | unsigned long saved_int_regs; |
a72d4945 | 1650 | unsigned HOST_WIDE_INT stack_adjust; |
5848830f | 1651 | arm_stack_offsets *offsets; |
ff9940b0 | 1652 | |
d5b7b3ae | 1653 | /* Never use a return instruction before reload has run. */ |
6d3d9133 NC |
1654 | if (!reload_completed) |
1655 | return 0; | |
efc2515b | 1656 | |
9b598fa0 RE |
1657 | func_type = arm_current_func_type (); |
1658 | ||
5b3e6663 | 1659 | /* Naked, volatile and stack alignment functions need special |
3a7731fd | 1660 | consideration. */ |
5b3e6663 | 1661 | if (func_type & (ARM_FT_VOLATILE | ARM_FT_NAKED | ARM_FT_STACKALIGN)) |
6d3d9133 | 1662 | return 0; |
06bea5aa | 1663 | |
a15908a4 PB |
1664 | /* So do interrupt functions that use the frame pointer and Thumb |
1665 | interrupt functions. */ | |
1666 | if (IS_INTERRUPT (func_type) && (frame_pointer_needed || TARGET_THUMB)) | |
06bea5aa | 1667 | return 0; |
a72d4945 | 1668 | |
5848830f PB |
1669 | offsets = arm_get_frame_offsets (); |
1670 | stack_adjust = offsets->outgoing_args - offsets->saved_regs; | |
a72d4945 | 1671 | |
6d3d9133 | 1672 | /* As do variadic functions. */ |
38173d38 | 1673 | if (crtl->args.pretend_args_size |
3cb66fd7 | 1674 | || cfun->machine->uses_anonymous_args |
699a4925 | 1675 | /* Or if the function calls __builtin_eh_return () */ |
e3b5732b | 1676 | || crtl->calls_eh_return |
699a4925 | 1677 | /* Or if the function calls alloca */ |
e3b5732b | 1678 | || cfun->calls_alloca |
a72d4945 RE |
1679 | /* Or if there is a stack adjustment. However, if the stack pointer |
1680 | is saved on the stack, we can use a pre-incrementing stack load. */ | |
ec6237e4 PB |
1681 | || !(stack_adjust == 0 || (TARGET_APCS_FRAME && frame_pointer_needed |
1682 | && stack_adjust == 4))) | |
ff9940b0 RE |
1683 | return 0; |
1684 | ||
954954d1 | 1685 | saved_int_regs = offsets->saved_regs_mask; |
d5db54a1 | 1686 | |
a72d4945 RE |
1687 | /* Unfortunately, the insn |
1688 | ||
1689 | ldmib sp, {..., sp, ...} | |
1690 | ||
1691 | triggers a bug on most SA-110 based devices, such that the stack | |
1692 | pointer won't be correctly restored if the instruction takes a | |
839a4992 | 1693 | page fault. We work around this problem by popping r3 along with |
a72d4945 | 1694 | the other registers, since that is never slower than executing |
f676971a | 1695 | another instruction. |
a72d4945 RE |
1696 | |
1697 | We test for !arm_arch5 here, because code for any architecture | |
1698 | less than this could potentially be run on one of the buggy | |
1699 | chips. */ | |
5b3e6663 | 1700 | if (stack_adjust == 4 && !arm_arch5 && TARGET_ARM) |
a72d4945 RE |
1701 | { |
1702 | /* Validate that r3 is a call-clobbered register (always true in | |
d6b4baa4 | 1703 | the default abi) ... */ |
a72d4945 RE |
1704 | if (!call_used_regs[3]) |
1705 | return 0; | |
1706 | ||
4f5dfed0 JC |
1707 | /* ... that it isn't being used for a return value ... */ |
1708 | if (arm_size_return_regs () >= (4 * UNITS_PER_WORD)) | |
1709 | return 0; | |
1710 | ||
1711 | /* ... or for a tail-call argument ... */ | |
a72d4945 RE |
1712 | if (sibling) |
1713 | { | |
e6d29d15 | 1714 | gcc_assert (GET_CODE (sibling) == CALL_INSN); |
a72d4945 RE |
1715 | |
1716 | if (find_regno_fusage (sibling, USE, 3)) | |
1717 | return 0; | |
1718 | } | |
1719 | ||
1720 | /* ... and that there are no call-saved registers in r0-r2 | |
1721 | (always true in the default ABI). */ | |
1722 | if (saved_int_regs & 0x7) | |
1723 | return 0; | |
1724 | } | |
1725 | ||
b111229a | 1726 | /* Can't be done if interworking with Thumb, and any registers have been |
d5db54a1 | 1727 | stacked. */ |
a15908a4 | 1728 | if (TARGET_INTERWORK && saved_int_regs != 0 && !IS_INTERRUPT(func_type)) |
b36ba79f | 1729 | return 0; |
d5db54a1 RE |
1730 | |
1731 | /* On StrongARM, conditional returns are expensive if they aren't | |
1732 | taken and multiple registers have been stacked. */ | |
abac3b49 | 1733 | if (iscond && arm_tune_strongarm) |
6ed30148 | 1734 | { |
f676971a | 1735 | /* Conditional return when just the LR is stored is a simple |
d5db54a1 RE |
1736 | conditional-load instruction, that's not expensive. */ |
1737 | if (saved_int_regs != 0 && saved_int_regs != (1 << LR_REGNUM)) | |
1738 | return 0; | |
6ed30148 | 1739 | |
020a4035 RE |
1740 | if (flag_pic |
1741 | && arm_pic_register != INVALID_REGNUM | |
6fb5fa3c | 1742 | && df_regs_ever_live_p (PIC_OFFSET_TABLE_REGNUM)) |
b111229a | 1743 | return 0; |
6ed30148 | 1744 | } |
d5db54a1 RE |
1745 | |
1746 | /* If there are saved registers but the LR isn't saved, then we need | |
1747 | two instructions for the return. */ | |
1748 | if (saved_int_regs && !(saved_int_regs & (1 << LR_REGNUM))) | |
1749 | return 0; | |
1750 | ||
3b684012 | 1751 | /* Can't be done if any of the FPA regs are pushed, |
6d3d9133 | 1752 | since this also requires an insn. */ |
9b66ebb1 PB |
1753 | if (TARGET_HARD_FLOAT && TARGET_FPA) |
1754 | for (regno = FIRST_FPA_REGNUM; regno <= LAST_FPA_REGNUM; regno++) | |
6fb5fa3c | 1755 | if (df_regs_ever_live_p (regno) && !call_used_regs[regno]) |
9b66ebb1 PB |
1756 | return 0; |
1757 | ||
1758 | /* Likewise VFP regs. */ | |
1759 | if (TARGET_HARD_FLOAT && TARGET_VFP) | |
1760 | for (regno = FIRST_VFP_REGNUM; regno <= LAST_VFP_REGNUM; regno++) | |
6fb5fa3c | 1761 | if (df_regs_ever_live_p (regno) && !call_used_regs[regno]) |
d5b7b3ae | 1762 | return 0; |
ff9940b0 | 1763 | |
5a9335ef NC |
1764 | if (TARGET_REALLY_IWMMXT) |
1765 | for (regno = FIRST_IWMMXT_REGNUM; regno <= LAST_IWMMXT_REGNUM; regno++) | |
6fb5fa3c | 1766 | if (df_regs_ever_live_p (regno) && ! call_used_regs[regno]) |
5a9335ef NC |
1767 | return 0; |
1768 | ||
ff9940b0 RE |
1769 | return 1; |
1770 | } | |
1771 | ||
cce8749e CH |
1772 | /* Return TRUE if int I is a valid immediate ARM constant. */ |
1773 | ||
1774 | int | |
e32bac5b | 1775 | const_ok_for_arm (HOST_WIDE_INT i) |
cce8749e | 1776 | { |
4642ccb1 | 1777 | int lowbit; |
e0b92319 | 1778 | |
f676971a | 1779 | /* For machines with >32 bit HOST_WIDE_INT, the bits above bit 31 must |
56636818 | 1780 | be all zero, or all one. */ |
30cf4896 KG |
1781 | if ((i & ~(unsigned HOST_WIDE_INT) 0xffffffff) != 0 |
1782 | && ((i & ~(unsigned HOST_WIDE_INT) 0xffffffff) | |
1783 | != ((~(unsigned HOST_WIDE_INT) 0) | |
1784 | & ~(unsigned HOST_WIDE_INT) 0xffffffff))) | |
56636818 | 1785 | return FALSE; |
f676971a | 1786 | |
4642ccb1 | 1787 | i &= (unsigned HOST_WIDE_INT) 0xffffffff; |
e0b92319 | 1788 | |
4642ccb1 RE |
1789 | /* Fast return for 0 and small values. We must do this for zero, since |
1790 | the code below can't handle that one case. */ | |
1791 | if ((i & ~(unsigned HOST_WIDE_INT) 0xff) == 0) | |
e2c671ba RE |
1792 | return TRUE; |
1793 | ||
5b3e6663 PB |
1794 | /* Get the number of trailing zeros. */ |
1795 | lowbit = ffs((int) i) - 1; | |
1796 | ||
1797 | /* Only even shifts are allowed in ARM mode so round down to the | |
1798 | nearest even number. */ | |
1799 | if (TARGET_ARM) | |
1800 | lowbit &= ~1; | |
4642ccb1 RE |
1801 | |
1802 | if ((i & ~(((unsigned HOST_WIDE_INT) 0xff) << lowbit)) == 0) | |
1803 | return TRUE; | |
5b3e6663 PB |
1804 | |
1805 | if (TARGET_ARM) | |
1806 | { | |
1807 | /* Allow rotated constants in ARM mode. */ | |
1808 | if (lowbit <= 4 | |
4642ccb1 RE |
1809 | && ((i & ~0xc000003f) == 0 |
1810 | || (i & ~0xf000000f) == 0 | |
1811 | || (i & ~0xfc000003) == 0)) | |
5b3e6663 PB |
1812 | return TRUE; |
1813 | } | |
1814 | else | |
1815 | { | |
1816 | HOST_WIDE_INT v; | |
1817 | ||
1818 | /* Allow repeated pattern. */ | |
1819 | v = i & 0xff; | |
1820 | v |= v << 16; | |
1821 | if (i == v || i == (v | (v << 8))) | |
1822 | return TRUE; | |
1823 | } | |
cce8749e | 1824 | |
f3bb6135 RE |
1825 | return FALSE; |
1826 | } | |
cce8749e | 1827 | |
6354dc9b | 1828 | /* Return true if I is a valid constant for the operation CODE. */ |
74bbc178 | 1829 | static int |
e32bac5b | 1830 | const_ok_for_op (HOST_WIDE_INT i, enum rtx_code code) |
e2c671ba RE |
1831 | { |
1832 | if (const_ok_for_arm (i)) | |
1833 | return 1; | |
1834 | ||
1835 | switch (code) | |
1836 | { | |
1837 | case PLUS: | |
1838 | return const_ok_for_arm (ARM_SIGN_EXTEND (-i)); | |
1839 | ||
1840 | case MINUS: /* Should only occur with (MINUS I reg) => rsb */ | |
1841 | case XOR: | |
1842 | case IOR: | |
1843 | return 0; | |
1844 | ||
1845 | case AND: | |
1846 | return const_ok_for_arm (ARM_SIGN_EXTEND (~i)); | |
1847 | ||
1848 | default: | |
e6d29d15 | 1849 | gcc_unreachable (); |
e2c671ba RE |
1850 | } |
1851 | } | |
1852 | ||
1853 | /* Emit a sequence of insns to handle a large constant. | |
1854 | CODE is the code of the operation required, it can be any of SET, PLUS, | |
1855 | IOR, AND, XOR, MINUS; | |
1856 | MODE is the mode in which the operation is being performed; | |
1857 | VAL is the integer to operate on; | |
1858 | SOURCE is the other operand (a register, or a null-pointer for SET); | |
1859 | SUBTARGETS means it is safe to create scratch registers if that will | |
2b835d68 RE |
1860 | either produce a simpler sequence, or we will want to cse the values. |
1861 | Return value is the number of insns emitted. */ | |
e2c671ba | 1862 | |
5b3e6663 | 1863 | /* ??? Tweak this for thumb2. */ |
e2c671ba | 1864 | int |
a406f566 | 1865 | arm_split_constant (enum rtx_code code, enum machine_mode mode, rtx insn, |
e32bac5b | 1866 | HOST_WIDE_INT val, rtx target, rtx source, int subtargets) |
2b835d68 | 1867 | { |
a406f566 MM |
1868 | rtx cond; |
1869 | ||
1870 | if (insn && GET_CODE (PATTERN (insn)) == COND_EXEC) | |
1871 | cond = COND_EXEC_TEST (PATTERN (insn)); | |
1872 | else | |
1873 | cond = NULL_RTX; | |
1874 | ||
2b835d68 RE |
1875 | if (subtargets || code == SET |
1876 | || (GET_CODE (target) == REG && GET_CODE (source) == REG | |
1877 | && REGNO (target) != REGNO (source))) | |
1878 | { | |
4b632bf1 | 1879 | /* After arm_reorg has been called, we can't fix up expensive |
05713b80 | 1880 | constants by pushing them into memory so we must synthesize |
4b632bf1 RE |
1881 | them in-line, regardless of the cost. This is only likely to |
1882 | be more costly on chips that have load delay slots and we are | |
1883 | compiling without running the scheduler (so no splitting | |
aec3cfba NC |
1884 | occurred before the final instruction emission). |
1885 | ||
1886 | Ref: gcc -O1 -mcpu=strongarm gcc.c-torture/compile/980506-2.c | |
aec3cfba | 1887 | */ |
5895f793 | 1888 | if (!after_arm_reorg |
a406f566 | 1889 | && !cond |
f676971a | 1890 | && (arm_gen_constant (code, mode, NULL_RTX, val, target, source, |
a406f566 | 1891 | 1, 0) |
4b632bf1 | 1892 | > arm_constant_limit + (code != SET))) |
2b835d68 RE |
1893 | { |
1894 | if (code == SET) | |
1895 | { | |
1896 | /* Currently SET is the only monadic value for CODE, all | |
1897 | the rest are diadic. */ | |
d66437c5 | 1898 | emit_set_insn (target, GEN_INT (val)); |
2b835d68 RE |
1899 | return 1; |
1900 | } | |
1901 | else | |
1902 | { | |
1903 | rtx temp = subtargets ? gen_reg_rtx (mode) : target; | |
1904 | ||
d66437c5 | 1905 | emit_set_insn (temp, GEN_INT (val)); |
2b835d68 RE |
1906 | /* For MINUS, the value is subtracted from, since we never |
1907 | have subtraction of a constant. */ | |
1908 | if (code == MINUS) | |
d66437c5 | 1909 | emit_set_insn (target, gen_rtx_MINUS (mode, temp, source)); |
2b835d68 | 1910 | else |
d66437c5 RE |
1911 | emit_set_insn (target, |
1912 | gen_rtx_fmt_ee (code, mode, source, temp)); | |
2b835d68 RE |
1913 | return 2; |
1914 | } | |
1915 | } | |
1916 | } | |
1917 | ||
f676971a | 1918 | return arm_gen_constant (code, mode, cond, val, target, source, subtargets, |
a406f566 | 1919 | 1); |
2b835d68 RE |
1920 | } |
1921 | ||
5b3e6663 PB |
1922 | /* Return the number of ARM instructions required to synthesize the given |
1923 | constant. */ | |
ceebdb09 | 1924 | static int |
e32bac5b | 1925 | count_insns_for_constant (HOST_WIDE_INT remainder, int i) |
ceebdb09 PB |
1926 | { |
1927 | HOST_WIDE_INT temp1; | |
1928 | int num_insns = 0; | |
1929 | do | |
1930 | { | |
1931 | int end; | |
f676971a | 1932 | |
ceebdb09 PB |
1933 | if (i <= 0) |
1934 | i += 32; | |
1935 | if (remainder & (3 << (i - 2))) | |
1936 | { | |
1937 | end = i - 8; | |
1938 | if (end < 0) | |
1939 | end += 32; | |
1940 | temp1 = remainder & ((0x0ff << end) | |
1941 | | ((i < end) ? (0xff >> (32 - end)) : 0)); | |
1942 | remainder &= ~temp1; | |
1943 | num_insns++; | |
1944 | i -= 6; | |
1945 | } | |
1946 | i -= 2; | |
1947 | } while (remainder); | |
1948 | return num_insns; | |
1949 | } | |
1950 | ||
a406f566 MM |
1951 | /* Emit an instruction with the indicated PATTERN. If COND is |
1952 | non-NULL, conditionalize the execution of the instruction on COND | |
1953 | being true. */ | |
1954 | ||
1955 | static void | |
1956 | emit_constant_insn (rtx cond, rtx pattern) | |
1957 | { | |
1958 | if (cond) | |
1959 | pattern = gen_rtx_COND_EXEC (VOIDmode, copy_rtx (cond), pattern); | |
1960 | emit_insn (pattern); | |
1961 | } | |
1962 | ||
2b835d68 RE |
1963 | /* As above, but extra parameter GENERATE which, if clear, suppresses |
1964 | RTL generation. */ | |
5b3e6663 | 1965 | /* ??? This needs more work for thumb2. */ |
1d6e90ac | 1966 | |
d5b7b3ae | 1967 | static int |
a406f566 | 1968 | arm_gen_constant (enum rtx_code code, enum machine_mode mode, rtx cond, |
e32bac5b RE |
1969 | HOST_WIDE_INT val, rtx target, rtx source, int subtargets, |
1970 | int generate) | |
e2c671ba | 1971 | { |
e2c671ba RE |
1972 | int can_invert = 0; |
1973 | int can_negate = 0; | |
1974 | int can_negate_initial = 0; | |
1975 | int can_shift = 0; | |
1976 | int i; | |
1977 | int num_bits_set = 0; | |
1978 | int set_sign_bit_copies = 0; | |
1979 | int clear_sign_bit_copies = 0; | |
1980 | int clear_zero_bit_copies = 0; | |
1981 | int set_zero_bit_copies = 0; | |
1982 | int insns = 0; | |
e2c671ba | 1983 | unsigned HOST_WIDE_INT temp1, temp2; |
30cf4896 | 1984 | unsigned HOST_WIDE_INT remainder = val & 0xffffffff; |
e2c671ba | 1985 | |
d5b7b3ae | 1986 | /* Find out which operations are safe for a given CODE. Also do a quick |
e2c671ba RE |
1987 | check for degenerate cases; these can occur when DImode operations |
1988 | are split. */ | |
1989 | switch (code) | |
1990 | { | |
1991 | case SET: | |
1992 | can_invert = 1; | |
1993 | can_shift = 1; | |
1994 | can_negate = 1; | |
1995 | break; | |
1996 | ||
1997 | case PLUS: | |
1998 | can_negate = 1; | |
1999 | can_negate_initial = 1; | |
2000 | break; | |
2001 | ||
2002 | case IOR: | |
30cf4896 | 2003 | if (remainder == 0xffffffff) |
e2c671ba | 2004 | { |
2b835d68 | 2005 | if (generate) |
a406f566 MM |
2006 | emit_constant_insn (cond, |
2007 | gen_rtx_SET (VOIDmode, target, | |
2008 | GEN_INT (ARM_SIGN_EXTEND (val)))); | |
e2c671ba RE |
2009 | return 1; |
2010 | } | |
2011 | if (remainder == 0) | |
2012 | { | |
2013 | if (reload_completed && rtx_equal_p (target, source)) | |
2014 | return 0; | |
2b835d68 | 2015 | if (generate) |
a406f566 MM |
2016 | emit_constant_insn (cond, |
2017 | gen_rtx_SET (VOIDmode, target, source)); | |
e2c671ba RE |
2018 | return 1; |
2019 | } | |
2020 | break; | |
2021 | ||
2022 | case AND: | |
2023 | if (remainder == 0) | |
2024 | { | |
2b835d68 | 2025 | if (generate) |
a406f566 MM |
2026 | emit_constant_insn (cond, |
2027 | gen_rtx_SET (VOIDmode, target, const0_rtx)); | |
e2c671ba RE |
2028 | return 1; |
2029 | } | |
30cf4896 | 2030 | if (remainder == 0xffffffff) |
e2c671ba RE |
2031 | { |
2032 | if (reload_completed && rtx_equal_p (target, source)) | |
2033 | return 0; | |
2b835d68 | 2034 | if (generate) |
a406f566 MM |
2035 | emit_constant_insn (cond, |
2036 | gen_rtx_SET (VOIDmode, target, source)); | |
e2c671ba RE |
2037 | return 1; |
2038 | } | |
2039 | can_invert = 1; | |
2040 | break; | |
2041 | ||
2042 | case XOR: | |
2043 | if (remainder == 0) | |
2044 | { | |
2045 | if (reload_completed && rtx_equal_p (target, source)) | |
2046 | return 0; | |
2b835d68 | 2047 | if (generate) |
a406f566 MM |
2048 | emit_constant_insn (cond, |
2049 | gen_rtx_SET (VOIDmode, target, source)); | |
e2c671ba RE |
2050 | return 1; |
2051 | } | |
e0b92319 | 2052 | |
e6d29d15 NS |
2053 | /* We don't know how to handle other cases yet. */ |
2054 | gcc_assert (remainder == 0xffffffff); | |
e0b92319 | 2055 | |
e6d29d15 NS |
2056 | if (generate) |
2057 | emit_constant_insn (cond, | |
2058 | gen_rtx_SET (VOIDmode, target, | |
2059 | gen_rtx_NOT (mode, source))); | |
2060 | return 1; | |
e2c671ba RE |
2061 | |
2062 | case MINUS: | |
2063 | /* We treat MINUS as (val - source), since (source - val) is always | |
2064 | passed as (source + (-val)). */ | |
2065 | if (remainder == 0) | |
2066 | { | |
2b835d68 | 2067 | if (generate) |
a406f566 MM |
2068 | emit_constant_insn (cond, |
2069 | gen_rtx_SET (VOIDmode, target, | |
2070 | gen_rtx_NEG (mode, source))); | |
e2c671ba RE |
2071 | return 1; |
2072 | } | |
2073 | if (const_ok_for_arm (val)) | |
2074 | { | |
2b835d68 | 2075 | if (generate) |
a406f566 | 2076 | emit_constant_insn (cond, |
f676971a | 2077 | gen_rtx_SET (VOIDmode, target, |
a406f566 MM |
2078 | gen_rtx_MINUS (mode, GEN_INT (val), |
2079 | source))); | |
e2c671ba RE |
2080 | return 1; |
2081 | } | |
2082 | can_negate = 1; | |
2083 | ||
2084 | break; | |
2085 | ||
2086 | default: | |
e6d29d15 | 2087 | gcc_unreachable (); |
e2c671ba RE |
2088 | } |
2089 | ||
6354dc9b | 2090 | /* If we can do it in one insn get out quickly. */ |
e2c671ba RE |
2091 | if (const_ok_for_arm (val) |
2092 | || (can_negate_initial && const_ok_for_arm (-val)) | |
2093 | || (can_invert && const_ok_for_arm (~val))) | |
2094 | { | |
2b835d68 | 2095 | if (generate) |
a406f566 MM |
2096 | emit_constant_insn (cond, |
2097 | gen_rtx_SET (VOIDmode, target, | |
f676971a | 2098 | (source |
a406f566 MM |
2099 | ? gen_rtx_fmt_ee (code, mode, source, |
2100 | GEN_INT (val)) | |
2101 | : GEN_INT (val)))); | |
e2c671ba RE |
2102 | return 1; |
2103 | } | |
2104 | ||
e2c671ba | 2105 | /* Calculate a few attributes that may be useful for specific |
6354dc9b | 2106 | optimizations. */ |
e2c671ba RE |
2107 | for (i = 31; i >= 0; i--) |
2108 | { | |
2109 | if ((remainder & (1 << i)) == 0) | |
2110 | clear_sign_bit_copies++; | |
2111 | else | |
2112 | break; | |
2113 | } | |
2114 | ||
2115 | for (i = 31; i >= 0; i--) | |
2116 | { | |
2117 | if ((remainder & (1 << i)) != 0) | |
2118 | set_sign_bit_copies++; | |
2119 | else | |
2120 | break; | |
2121 | } | |
2122 | ||
2123 | for (i = 0; i <= 31; i++) | |
2124 | { | |
2125 | if ((remainder & (1 << i)) == 0) | |
2126 | clear_zero_bit_copies++; | |
2127 | else | |
2128 | break; | |
2129 | } | |
2130 | ||
2131 | for (i = 0; i <= 31; i++) | |
2132 | { | |
2133 | if ((remainder & (1 << i)) != 0) | |
2134 | set_zero_bit_copies++; | |
2135 | else | |
2136 | break; | |
2137 | } | |
2138 | ||
2139 | switch (code) | |
2140 | { | |
2141 | case SET: | |
5b3e6663 PB |
2142 | /* See if we can use movw. */ |
2143 | if (arm_arch_thumb2 && (remainder & 0xffff0000) == 0) | |
2144 | { | |
2145 | if (generate) | |
2146 | emit_constant_insn (cond, gen_rtx_SET (VOIDmode, target, | |
2147 | GEN_INT (val))); | |
2148 | return 1; | |
2149 | } | |
2150 | ||
e2c671ba RE |
2151 | /* See if we can do this by sign_extending a constant that is known |
2152 | to be negative. This is a good, way of doing it, since the shift | |
2153 | may well merge into a subsequent insn. */ | |
2154 | if (set_sign_bit_copies > 1) | |
2155 | { | |
2156 | if (const_ok_for_arm | |
f676971a | 2157 | (temp1 = ARM_SIGN_EXTEND (remainder |
e2c671ba RE |
2158 | << (set_sign_bit_copies - 1)))) |
2159 | { | |
2b835d68 RE |
2160 | if (generate) |
2161 | { | |
d499463f | 2162 | rtx new_src = subtargets ? gen_reg_rtx (mode) : target; |
a406f566 | 2163 | emit_constant_insn (cond, |
f676971a | 2164 | gen_rtx_SET (VOIDmode, new_src, |
a406f566 MM |
2165 | GEN_INT (temp1))); |
2166 | emit_constant_insn (cond, | |
f676971a | 2167 | gen_ashrsi3 (target, new_src, |
a406f566 | 2168 | GEN_INT (set_sign_bit_copies - 1))); |
2b835d68 | 2169 | } |
e2c671ba RE |
2170 | return 2; |
2171 | } | |
2172 | /* For an inverted constant, we will need to set the low bits, | |
2173 | these will be shifted out of harm's way. */ | |
2174 | temp1 |= (1 << (set_sign_bit_copies - 1)) - 1; | |
2175 | if (const_ok_for_arm (~temp1)) | |
2176 | { | |
2b835d68 RE |
2177 | if (generate) |
2178 | { | |
d499463f | 2179 | rtx new_src = subtargets ? gen_reg_rtx (mode) : target; |
a406f566 MM |
2180 | emit_constant_insn (cond, |
2181 | gen_rtx_SET (VOIDmode, new_src, | |
2182 | GEN_INT (temp1))); | |
2183 | emit_constant_insn (cond, | |
f676971a | 2184 | gen_ashrsi3 (target, new_src, |
a406f566 | 2185 | GEN_INT (set_sign_bit_copies - 1))); |
2b835d68 | 2186 | } |
e2c671ba RE |
2187 | return 2; |
2188 | } | |
2189 | } | |
2190 | ||
c87e6352 RE |
2191 | /* See if we can calculate the value as the difference between two |
2192 | valid immediates. */ | |
2193 | if (clear_sign_bit_copies + clear_zero_bit_copies <= 16) | |
2194 | { | |
2195 | int topshift = clear_sign_bit_copies & ~1; | |
2196 | ||
fa2c88a0 RE |
2197 | temp1 = ARM_SIGN_EXTEND ((remainder + (0x00800000 >> topshift)) |
2198 | & (0xff000000 >> topshift)); | |
c87e6352 RE |
2199 | |
2200 | /* If temp1 is zero, then that means the 9 most significant | |
2201 | bits of remainder were 1 and we've caused it to overflow. | |
2202 | When topshift is 0 we don't need to do anything since we | |
2203 | can borrow from 'bit 32'. */ | |
2204 | if (temp1 == 0 && topshift != 0) | |
2205 | temp1 = 0x80000000 >> (topshift - 1); | |
2206 | ||
fa2c88a0 | 2207 | temp2 = ARM_SIGN_EXTEND (temp1 - remainder); |
e0b92319 | 2208 | |
c87e6352 RE |
2209 | if (const_ok_for_arm (temp2)) |
2210 | { | |
2211 | if (generate) | |
2212 | { | |
2213 | rtx new_src = subtargets ? gen_reg_rtx (mode) : target; | |
2214 | emit_constant_insn (cond, | |
2215 | gen_rtx_SET (VOIDmode, new_src, | |
2216 | GEN_INT (temp1))); | |
2217 | emit_constant_insn (cond, | |
2218 | gen_addsi3 (target, new_src, | |
2219 | GEN_INT (-temp2))); | |
2220 | } | |
2221 | ||
2222 | return 2; | |
2223 | } | |
2224 | } | |
2225 | ||
e2c671ba RE |
2226 | /* See if we can generate this by setting the bottom (or the top) |
2227 | 16 bits, and then shifting these into the other half of the | |
2228 | word. We only look for the simplest cases, to do more would cost | |
2229 | too much. Be careful, however, not to generate this when the | |
2230 | alternative would take fewer insns. */ | |
30cf4896 | 2231 | if (val & 0xffff0000) |
e2c671ba | 2232 | { |
30cf4896 | 2233 | temp1 = remainder & 0xffff0000; |
e2c671ba RE |
2234 | temp2 = remainder & 0x0000ffff; |
2235 | ||
6354dc9b | 2236 | /* Overlaps outside this range are best done using other methods. */ |
e2c671ba RE |
2237 | for (i = 9; i < 24; i++) |
2238 | { | |
30cf4896 | 2239 | if ((((temp2 | (temp2 << i)) & 0xffffffff) == remainder) |
5895f793 | 2240 | && !const_ok_for_arm (temp2)) |
e2c671ba | 2241 | { |
d499463f RE |
2242 | rtx new_src = (subtargets |
2243 | ? (generate ? gen_reg_rtx (mode) : NULL_RTX) | |
2244 | : target); | |
a406f566 | 2245 | insns = arm_gen_constant (code, mode, cond, temp2, new_src, |
2b835d68 | 2246 | source, subtargets, generate); |
e2c671ba | 2247 | source = new_src; |
2b835d68 | 2248 | if (generate) |
f676971a | 2249 | emit_constant_insn |
a406f566 MM |
2250 | (cond, |
2251 | gen_rtx_SET | |
2252 | (VOIDmode, target, | |
2253 | gen_rtx_IOR (mode, | |
2254 | gen_rtx_ASHIFT (mode, source, | |
2255 | GEN_INT (i)), | |
2256 | source))); | |
e2c671ba RE |
2257 | return insns + 1; |
2258 | } | |
2259 | } | |
2260 | ||
6354dc9b | 2261 | /* Don't duplicate cases already considered. */ |
e2c671ba RE |
2262 | for (i = 17; i < 24; i++) |
2263 | { | |
2264 | if (((temp1 | (temp1 >> i)) == remainder) | |
5895f793 | 2265 | && !const_ok_for_arm (temp1)) |
e2c671ba | 2266 | { |
d499463f RE |
2267 | rtx new_src = (subtargets |
2268 | ? (generate ? gen_reg_rtx (mode) : NULL_RTX) | |
2269 | : target); | |
a406f566 | 2270 | insns = arm_gen_constant (code, mode, cond, temp1, new_src, |
2b835d68 | 2271 | source, subtargets, generate); |
e2c671ba | 2272 | source = new_src; |
2b835d68 | 2273 | if (generate) |
a406f566 MM |
2274 | emit_constant_insn |
2275 | (cond, | |
2276 | gen_rtx_SET (VOIDmode, target, | |
43cffd11 RE |
2277 | gen_rtx_IOR |
2278 | (mode, | |
2279 | gen_rtx_LSHIFTRT (mode, source, | |
2280 | GEN_INT (i)), | |
2281 | source))); | |
e2c671ba RE |
2282 | return insns + 1; |
2283 | } | |
2284 | } | |
2285 | } | |
2286 | break; | |
2287 | ||
2288 | case IOR: | |
2289 | case XOR: | |
7b64da89 RE |
2290 | /* If we have IOR or XOR, and the constant can be loaded in a |
2291 | single instruction, and we can find a temporary to put it in, | |
e2c671ba RE |
2292 | then this can be done in two instructions instead of 3-4. */ |
2293 | if (subtargets | |
d499463f | 2294 | /* TARGET can't be NULL if SUBTARGETS is 0 */ |
5895f793 | 2295 | || (reload_completed && !reg_mentioned_p (target, source))) |
e2c671ba | 2296 | { |
5895f793 | 2297 | if (const_ok_for_arm (ARM_SIGN_EXTEND (~val))) |
e2c671ba | 2298 | { |
2b835d68 RE |
2299 | if (generate) |
2300 | { | |
2301 | rtx sub = subtargets ? gen_reg_rtx (mode) : target; | |
e2c671ba | 2302 | |
a406f566 | 2303 | emit_constant_insn (cond, |
f676971a | 2304 | gen_rtx_SET (VOIDmode, sub, |
a406f566 MM |
2305 | GEN_INT (val))); |
2306 | emit_constant_insn (cond, | |
f676971a | 2307 | gen_rtx_SET (VOIDmode, target, |
a406f566 MM |
2308 | gen_rtx_fmt_ee (code, mode, |
2309 | source, sub))); | |
2b835d68 | 2310 | } |
e2c671ba RE |
2311 | return 2; |
2312 | } | |
2313 | } | |
2314 | ||
2315 | if (code == XOR) | |
2316 | break; | |
2317 | ||
2318 | if (set_sign_bit_copies > 8 | |
2319 | && (val & (-1 << (32 - set_sign_bit_copies))) == val) | |
2320 | { | |
2b835d68 RE |
2321 | if (generate) |
2322 | { | |
2323 | rtx sub = subtargets ? gen_reg_rtx (mode) : target; | |
2324 | rtx shift = GEN_INT (set_sign_bit_copies); | |
2325 | ||
f676971a | 2326 | emit_constant_insn |
a406f566 MM |
2327 | (cond, |
2328 | gen_rtx_SET (VOIDmode, sub, | |
f676971a | 2329 | gen_rtx_NOT (mode, |
a406f566 | 2330 | gen_rtx_ASHIFT (mode, |
f676971a | 2331 | source, |
a406f566 | 2332 | shift)))); |
f676971a | 2333 | emit_constant_insn |
a406f566 MM |
2334 | (cond, |
2335 | gen_rtx_SET (VOIDmode, target, | |
2336 | gen_rtx_NOT (mode, | |
2337 | gen_rtx_LSHIFTRT (mode, sub, | |
2338 | shift)))); | |
2b835d68 | 2339 | } |
e2c671ba RE |
2340 | return 2; |
2341 | } | |
2342 | ||
2343 | if (set_zero_bit_copies > 8 | |
2344 | && (remainder & ((1 << set_zero_bit_copies) - 1)) == remainder) | |
2345 | { | |
2b835d68 RE |
2346 | if (generate) |
2347 | { | |
2348 | rtx sub = subtargets ? gen_reg_rtx (mode) : target; | |
2349 | rtx shift = GEN_INT (set_zero_bit_copies); | |
2350 | ||
a406f566 MM |
2351 | emit_constant_insn |
2352 | (cond, | |
2353 | gen_rtx_SET (VOIDmode, sub, | |
2354 | gen_rtx_NOT (mode, | |
2355 | gen_rtx_LSHIFTRT (mode, | |
2356 | source, | |
2357 | shift)))); | |
f676971a | 2358 | emit_constant_insn |
a406f566 MM |
2359 | (cond, |
2360 | gen_rtx_SET (VOIDmode, target, | |
2361 | gen_rtx_NOT (mode, | |
2362 | gen_rtx_ASHIFT (mode, sub, | |
2363 | shift)))); | |
2b835d68 | 2364 | } |
e2c671ba RE |
2365 | return 2; |
2366 | } | |
2367 | ||
5895f793 | 2368 | if (const_ok_for_arm (temp1 = ARM_SIGN_EXTEND (~val))) |
e2c671ba | 2369 | { |
2b835d68 RE |
2370 | if (generate) |
2371 | { | |
2372 | rtx sub = subtargets ? gen_reg_rtx (mode) : target; | |
a406f566 MM |
2373 | emit_constant_insn (cond, |
2374 | gen_rtx_SET (VOIDmode, sub, | |
2375 | gen_rtx_NOT (mode, source))); | |
2b835d68 RE |
2376 | source = sub; |
2377 | if (subtargets) | |
2378 | sub = gen_reg_rtx (mode); | |
a406f566 MM |
2379 | emit_constant_insn (cond, |
2380 | gen_rtx_SET (VOIDmode, sub, | |
f676971a | 2381 | gen_rtx_AND (mode, source, |
a406f566 MM |
2382 | GEN_INT (temp1)))); |
2383 | emit_constant_insn (cond, | |
2384 | gen_rtx_SET (VOIDmode, target, | |
2385 | gen_rtx_NOT (mode, sub))); | |
2b835d68 | 2386 | } |
e2c671ba RE |
2387 | return 3; |
2388 | } | |
2389 | break; | |
2390 | ||
2391 | case AND: | |
2392 | /* See if two shifts will do 2 or more insn's worth of work. */ | |
2393 | if (clear_sign_bit_copies >= 16 && clear_sign_bit_copies < 24) | |
2394 | { | |
30cf4896 | 2395 | HOST_WIDE_INT shift_mask = ((0xffffffff |
e2c671ba | 2396 | << (32 - clear_sign_bit_copies)) |
30cf4896 | 2397 | & 0xffffffff); |
e2c671ba | 2398 | |
30cf4896 | 2399 | if ((remainder | shift_mask) != 0xffffffff) |
e2c671ba | 2400 | { |
2b835d68 RE |
2401 | if (generate) |
2402 | { | |
d499463f | 2403 | rtx new_src = subtargets ? gen_reg_rtx (mode) : target; |
f676971a | 2404 | insns = arm_gen_constant (AND, mode, cond, |
a406f566 | 2405 | remainder | shift_mask, |
d499463f RE |
2406 | new_src, source, subtargets, 1); |
2407 | source = new_src; | |
2b835d68 RE |
2408 | } |
2409 | else | |
d499463f RE |
2410 | { |
2411 | rtx targ = subtargets ? NULL_RTX : target; | |
a406f566 MM |
2412 | insns = arm_gen_constant (AND, mode, cond, |
2413 | remainder | shift_mask, | |
d499463f RE |
2414 | targ, source, subtargets, 0); |
2415 | } | |
2b835d68 RE |
2416 | } |
2417 | ||
2418 | if (generate) | |
2419 | { | |
d499463f RE |
2420 | rtx new_src = subtargets ? gen_reg_rtx (mode) : target; |
2421 | rtx shift = GEN_INT (clear_sign_bit_copies); | |
2422 | ||
2423 | emit_insn (gen_ashlsi3 (new_src, source, shift)); | |
2424 | emit_insn (gen_lshrsi3 (target, new_src, shift)); | |
e2c671ba RE |
2425 | } |
2426 | ||
e2c671ba RE |
2427 | return insns + 2; |
2428 | } | |
2429 | ||
2430 | if (clear_zero_bit_copies >= 16 && clear_zero_bit_copies < 24) | |
2431 | { | |
2432 | HOST_WIDE_INT shift_mask = (1 << clear_zero_bit_copies) - 1; | |
f676971a | 2433 | |
30cf4896 | 2434 | if ((remainder | shift_mask) != 0xffffffff) |
e2c671ba | 2435 | { |
2b835d68 RE |
2436 | if (generate) |
2437 | { | |
d499463f RE |
2438 | rtx new_src = subtargets ? gen_reg_rtx (mode) : target; |
2439 | ||
a406f566 MM |
2440 | insns = arm_gen_constant (AND, mode, cond, |
2441 | remainder | shift_mask, | |
d499463f RE |
2442 | new_src, source, subtargets, 1); |
2443 | source = new_src; | |
2b835d68 RE |
2444 | } |
2445 | else | |
d499463f RE |
2446 | { |
2447 | rtx targ = subtargets ? NULL_RTX : target; | |
2448 | ||
a406f566 MM |
2449 | insns = arm_gen_constant (AND, mode, cond, |
2450 | remainder | shift_mask, | |
d499463f RE |
2451 | targ, source, subtargets, 0); |
2452 | } | |
2b835d68 RE |
2453 | } |
2454 | ||
2455 | if (generate) | |
2456 | { | |
d499463f RE |
2457 | rtx new_src = subtargets ? gen_reg_rtx (mode) : target; |
2458 | rtx shift = GEN_INT (clear_zero_bit_copies); | |
2459 | ||
2460 | emit_insn (gen_lshrsi3 (new_src, source, shift)); | |
2461 | emit_insn (gen_ashlsi3 (target, new_src, shift)); | |
e2c671ba RE |
2462 | } |
2463 | ||
e2c671ba RE |
2464 | return insns + 2; |
2465 | } | |
2466 | ||
2467 | break; | |
2468 | ||
2469 | default: | |
2470 | break; | |
2471 | } | |
2472 | ||
2473 | for (i = 0; i < 32; i++) | |
2474 | if (remainder & (1 << i)) | |
2475 | num_bits_set++; | |
2476 | ||
2477 | if (code == AND || (can_invert && num_bits_set > 16)) | |
30cf4896 | 2478 | remainder = (~remainder) & 0xffffffff; |
e2c671ba | 2479 | else if (code == PLUS && num_bits_set > 16) |
30cf4896 | 2480 | remainder = (-remainder) & 0xffffffff; |
e2c671ba RE |
2481 | else |
2482 | { | |
2483 | can_invert = 0; | |
2484 | can_negate = 0; | |
2485 | } | |
2486 | ||
2487 | /* Now try and find a way of doing the job in either two or three | |
2488 | instructions. | |
2489 | We start by looking for the largest block of zeros that are aligned on | |
2490 | a 2-bit boundary, we then fill up the temps, wrapping around to the | |
2491 | top of the word when we drop off the bottom. | |
5b3e6663 PB |
2492 | In the worst case this code should produce no more than four insns. |
2493 | Thumb-2 constants are shifted, not rotated, so the MSB is always the | |
2494 | best place to start. */ | |
2495 | ||
2496 | /* ??? Use thumb2 replicated constants when the high and low halfwords are | |
2497 | the same. */ | |
e2c671ba RE |
2498 | { |
2499 | int best_start = 0; | |
5b3e6663 | 2500 | if (!TARGET_THUMB2) |
e2c671ba | 2501 | { |
5b3e6663 | 2502 | int best_consecutive_zeros = 0; |
e2c671ba | 2503 | |
5b3e6663 | 2504 | for (i = 0; i < 32; i += 2) |
e2c671ba | 2505 | { |
5b3e6663 PB |
2506 | int consecutive_zeros = 0; |
2507 | ||
2508 | if (!(remainder & (3 << i))) | |
e2c671ba | 2509 | { |
5b3e6663 PB |
2510 | while ((i < 32) && !(remainder & (3 << i))) |
2511 | { | |
2512 | consecutive_zeros += 2; | |
2513 | i += 2; | |
2514 | } | |
2515 | if (consecutive_zeros > best_consecutive_zeros) | |
2516 | { | |
2517 | best_consecutive_zeros = consecutive_zeros; | |
2518 | best_start = i - consecutive_zeros; | |
2519 | } | |
2520 | i -= 2; | |
e2c671ba | 2521 | } |
e2c671ba | 2522 | } |
ceebdb09 | 2523 | |
5b3e6663 PB |
2524 | /* So long as it won't require any more insns to do so, it's |
2525 | desirable to emit a small constant (in bits 0...9) in the last | |
2526 | insn. This way there is more chance that it can be combined with | |
2527 | a later addressing insn to form a pre-indexed load or store | |
2528 | operation. Consider: | |
2529 | ||
2530 | *((volatile int *)0xe0000100) = 1; | |
2531 | *((volatile int *)0xe0000110) = 2; | |
2532 | ||
2533 | We want this to wind up as: | |
2534 | ||
2535 | mov rA, #0xe0000000 | |
2536 | mov rB, #1 | |
2537 | str rB, [rA, #0x100] | |
2538 | mov rB, #2 | |
2539 | str rB, [rA, #0x110] | |
2540 | ||
2541 | rather than having to synthesize both large constants from scratch. | |
2542 | ||
2543 | Therefore, we calculate how many insns would be required to emit | |
2544 | the constant starting from `best_start', and also starting from | |
2545 | zero (i.e. with bit 31 first to be output). If `best_start' doesn't | |
2546 | yield a shorter sequence, we may as well use zero. */ | |
2547 | if (best_start != 0 | |
2548 | && ((((unsigned HOST_WIDE_INT) 1) << best_start) < remainder) | |
2549 | && (count_insns_for_constant (remainder, 0) <= | |
2550 | count_insns_for_constant (remainder, best_start))) | |
2551 | best_start = 0; | |
2552 | } | |
ceebdb09 PB |
2553 | |
2554 | /* Now start emitting the insns. */ | |
e2c671ba RE |
2555 | i = best_start; |
2556 | do | |
2557 | { | |
2558 | int end; | |
2559 | ||
2560 | if (i <= 0) | |
2561 | i += 32; | |
2562 | if (remainder & (3 << (i - 2))) | |
2563 | { | |
2564 | end = i - 8; | |
2565 | if (end < 0) | |
2566 | end += 32; | |
2567 | temp1 = remainder & ((0x0ff << end) | |
2568 | | ((i < end) ? (0xff >> (32 - end)) : 0)); | |
2569 | remainder &= ~temp1; | |
2570 | ||
d499463f | 2571 | if (generate) |
e2c671ba | 2572 | { |
9503f3d1 RH |
2573 | rtx new_src, temp1_rtx; |
2574 | ||
2575 | if (code == SET || code == MINUS) | |
2576 | { | |
2577 | new_src = (subtargets ? gen_reg_rtx (mode) : target); | |
96ae8197 | 2578 | if (can_invert && code != MINUS) |
9503f3d1 RH |
2579 | temp1 = ~temp1; |
2580 | } | |
2581 | else | |
2582 | { | |
96ae8197 | 2583 | if (remainder && subtargets) |
9503f3d1 | 2584 | new_src = gen_reg_rtx (mode); |
96ae8197 NC |
2585 | else |
2586 | new_src = target; | |
9503f3d1 RH |
2587 | if (can_invert) |
2588 | temp1 = ~temp1; | |
2589 | else if (can_negate) | |
2590 | temp1 = -temp1; | |
2591 | } | |
2592 | ||
2593 | temp1 = trunc_int_for_mode (temp1, mode); | |
2594 | temp1_rtx = GEN_INT (temp1); | |
d499463f RE |
2595 | |
2596 | if (code == SET) | |
9503f3d1 | 2597 | ; |
d499463f | 2598 | else if (code == MINUS) |
9503f3d1 | 2599 | temp1_rtx = gen_rtx_MINUS (mode, temp1_rtx, source); |
d499463f | 2600 | else |
9503f3d1 RH |
2601 | temp1_rtx = gen_rtx_fmt_ee (code, mode, source, temp1_rtx); |
2602 | ||
a406f566 | 2603 | emit_constant_insn (cond, |
f676971a | 2604 | gen_rtx_SET (VOIDmode, new_src, |
a406f566 | 2605 | temp1_rtx)); |
d499463f | 2606 | source = new_src; |
e2c671ba RE |
2607 | } |
2608 | ||
d499463f RE |
2609 | if (code == SET) |
2610 | { | |
2611 | can_invert = 0; | |
2612 | code = PLUS; | |
2613 | } | |
2614 | else if (code == MINUS) | |
2615 | code = PLUS; | |
2616 | ||
e2c671ba | 2617 | insns++; |
5b3e6663 PB |
2618 | if (TARGET_ARM) |
2619 | i -= 6; | |
2620 | else | |
2621 | i -= 7; | |
e2c671ba | 2622 | } |
7a085dce | 2623 | /* Arm allows rotates by a multiple of two. Thumb-2 allows arbitrary |
5b3e6663 PB |
2624 | shifts. */ |
2625 | if (TARGET_ARM) | |
2626 | i -= 2; | |
2627 | else | |
2628 | i--; | |
1d6e90ac NC |
2629 | } |
2630 | while (remainder); | |
e2c671ba | 2631 | } |
1d6e90ac | 2632 | |
e2c671ba RE |
2633 | return insns; |
2634 | } | |
2635 | ||
bd9c7e23 RE |
2636 | /* Canonicalize a comparison so that we are more likely to recognize it. |
2637 | This can be done for a few constant compares, where we can make the | |
2638 | immediate value easier to load. */ | |
1d6e90ac | 2639 | |
bd9c7e23 | 2640 | enum rtx_code |
a14b88bb PB |
2641 | arm_canonicalize_comparison (enum rtx_code code, enum machine_mode mode, |
2642 | rtx * op1) | |
bd9c7e23 | 2643 | { |
ad076f4e | 2644 | unsigned HOST_WIDE_INT i = INTVAL (*op1); |
a14b88bb PB |
2645 | unsigned HOST_WIDE_INT maxval; |
2646 | maxval = (((unsigned HOST_WIDE_INT) 1) << (GET_MODE_BITSIZE(mode) - 1)) - 1; | |
bd9c7e23 RE |
2647 | |
2648 | switch (code) | |
2649 | { | |
2650 | case EQ: | |
2651 | case NE: | |
2652 | return code; | |
2653 | ||
2654 | case GT: | |
2655 | case LE: | |
a14b88bb | 2656 | if (i != maxval |
5895f793 | 2657 | && (const_ok_for_arm (i + 1) || const_ok_for_arm (-(i + 1)))) |
bd9c7e23 | 2658 | { |
5895f793 | 2659 | *op1 = GEN_INT (i + 1); |
bd9c7e23 RE |
2660 | return code == GT ? GE : LT; |
2661 | } | |
2662 | break; | |
2663 | ||
2664 | case GE: | |
2665 | case LT: | |
a14b88bb | 2666 | if (i != ~maxval |
5895f793 | 2667 | && (const_ok_for_arm (i - 1) || const_ok_for_arm (-(i - 1)))) |
bd9c7e23 | 2668 | { |
5895f793 | 2669 | *op1 = GEN_INT (i - 1); |
bd9c7e23 RE |
2670 | return code == GE ? GT : LE; |
2671 | } | |
2672 | break; | |
2673 | ||
2674 | case GTU: | |
2675 | case LEU: | |
30cf4896 | 2676 | if (i != ~((unsigned HOST_WIDE_INT) 0) |
5895f793 | 2677 | && (const_ok_for_arm (i + 1) || const_ok_for_arm (-(i + 1)))) |
bd9c7e23 RE |
2678 | { |
2679 | *op1 = GEN_INT (i + 1); | |
2680 | return code == GTU ? GEU : LTU; | |
2681 | } | |
2682 | break; | |
2683 | ||
2684 | case GEU: | |
2685 | case LTU: | |
2686 | if (i != 0 | |
5895f793 | 2687 | && (const_ok_for_arm (i - 1) || const_ok_for_arm (-(i - 1)))) |
bd9c7e23 RE |
2688 | { |
2689 | *op1 = GEN_INT (i - 1); | |
2690 | return code == GEU ? GTU : LEU; | |
2691 | } | |
2692 | break; | |
2693 | ||
2694 | default: | |
e6d29d15 | 2695 | gcc_unreachable (); |
bd9c7e23 RE |
2696 | } |
2697 | ||
2698 | return code; | |
2699 | } | |
bd9c7e23 | 2700 | |
d4453b7a PB |
2701 | |
2702 | /* Define how to find the value returned by a function. */ | |
2703 | ||
9f7bf991 | 2704 | rtx |
586de218 | 2705 | arm_function_value(const_tree type, const_tree func ATTRIBUTE_UNUSED) |
d4453b7a PB |
2706 | { |
2707 | enum machine_mode mode; | |
2708 | int unsignedp ATTRIBUTE_UNUSED; | |
2709 | rtx r ATTRIBUTE_UNUSED; | |
2710 | ||
d4453b7a PB |
2711 | mode = TYPE_MODE (type); |
2712 | /* Promote integer types. */ | |
2713 | if (INTEGRAL_TYPE_P (type)) | |
2714 | PROMOTE_FUNCTION_MODE (mode, unsignedp, type); | |
866af8a9 JB |
2715 | |
2716 | /* Promotes small structs returned in a register to full-word size | |
2717 | for big-endian AAPCS. */ | |
2718 | if (arm_return_in_msb (type)) | |
2719 | { | |
2720 | HOST_WIDE_INT size = int_size_in_bytes (type); | |
2721 | if (size % UNITS_PER_WORD != 0) | |
2722 | { | |
2723 | size += UNITS_PER_WORD - size % UNITS_PER_WORD; | |
2724 | mode = mode_for_size (size * BITS_PER_UNIT, MODE_INT, 0); | |
2725 | } | |
2726 | } | |
e0b92319 | 2727 | |
d4453b7a PB |
2728 | return LIBCALL_VALUE(mode); |
2729 | } | |
2730 | ||
e0b92319 | 2731 | /* Determine the amount of memory needed to store the possible return |
9f7bf991 RE |
2732 | registers of an untyped call. */ |
2733 | int | |
2734 | arm_apply_result_size (void) | |
2735 | { | |
2736 | int size = 16; | |
2737 | ||
2738 | if (TARGET_ARM) | |
2739 | { | |
2740 | if (TARGET_HARD_FLOAT_ABI) | |
2741 | { | |
2742 | if (TARGET_FPA) | |
2743 | size += 12; | |
2744 | if (TARGET_MAVERICK) | |
2745 | size += 8; | |
2746 | } | |
2747 | if (TARGET_IWMMXT_ABI) | |
2748 | size += 8; | |
2749 | } | |
2750 | ||
2751 | return size; | |
2752 | } | |
d4453b7a | 2753 | |
f5a1b0d2 | 2754 | /* Decide whether a type should be returned in memory (true) |
23668cf7 | 2755 | or in a register (false). This is called as the target hook |
81464b2c | 2756 | TARGET_RETURN_IN_MEMORY. */ |
23668cf7 | 2757 | static bool |
81464b2c | 2758 | arm_return_in_memory (const_tree type, const_tree fntype ATTRIBUTE_UNUSED) |
2b835d68 | 2759 | { |
dc0ba55a JT |
2760 | HOST_WIDE_INT size; |
2761 | ||
88f77cba JB |
2762 | size = int_size_in_bytes (type); |
2763 | ||
2764 | /* Vector values should be returned using ARM registers, not memory (unless | |
2765 | they're over 16 bytes, which will break since we only have four | |
2766 | call-clobbered registers to play with). */ | |
2767 | if (TREE_CODE (type) == VECTOR_TYPE) | |
2768 | return (size < 0 || size > (4 * UNITS_PER_WORD)); | |
2769 | ||
3dd7ab65 JB |
2770 | if (!AGGREGATE_TYPE_P (type) && |
2771 | !(TARGET_AAPCS_BASED && TREE_CODE (type) == COMPLEX_TYPE)) | |
2772 | /* All simple types are returned in registers. | |
2773 | For AAPCS, complex types are treated the same as aggregates. */ | |
d7d01975 | 2774 | return 0; |
dc0ba55a | 2775 | |
5848830f | 2776 | if (arm_abi != ARM_ABI_APCS) |
dc0ba55a | 2777 | { |
5848830f | 2778 | /* ATPCS and later return aggregate types in memory only if they are |
dc0ba55a JT |
2779 | larger than a word (or are variable size). */ |
2780 | return (size < 0 || size > UNITS_PER_WORD); | |
2781 | } | |
f676971a | 2782 | |
6bc82793 | 2783 | /* For the arm-wince targets we choose to be compatible with Microsoft's |
d5b7b3ae RE |
2784 | ARM and Thumb compilers, which always return aggregates in memory. */ |
2785 | #ifndef ARM_WINCE | |
e529bd42 NC |
2786 | /* All structures/unions bigger than one word are returned in memory. |
2787 | Also catch the case where int_size_in_bytes returns -1. In this case | |
6bc82793 | 2788 | the aggregate is either huge or of variable size, and in either case |
e529bd42 | 2789 | we will want to return it via memory and not in a register. */ |
dc0ba55a | 2790 | if (size < 0 || size > UNITS_PER_WORD) |
d7d01975 | 2791 | return 1; |
f676971a | 2792 | |
d7d01975 | 2793 | if (TREE_CODE (type) == RECORD_TYPE) |
2b835d68 RE |
2794 | { |
2795 | tree field; | |
2796 | ||
3a2ea258 RE |
2797 | /* For a struct the APCS says that we only return in a register |
2798 | if the type is 'integer like' and every addressable element | |
2799 | has an offset of zero. For practical purposes this means | |
2800 | that the structure can have at most one non bit-field element | |
2801 | and that this element must be the first one in the structure. */ | |
f676971a | 2802 | |
f5a1b0d2 NC |
2803 | /* Find the first field, ignoring non FIELD_DECL things which will |
2804 | have been created by C++. */ | |
2805 | for (field = TYPE_FIELDS (type); | |
2806 | field && TREE_CODE (field) != FIELD_DECL; | |
2807 | field = TREE_CHAIN (field)) | |
2808 | continue; | |
f676971a | 2809 | |
f5a1b0d2 | 2810 | if (field == NULL) |
9e291dbe | 2811 | return 0; /* An empty structure. Allowed by an extension to ANSI C. */ |
f5a1b0d2 | 2812 | |
d5b7b3ae RE |
2813 | /* Check that the first field is valid for returning in a register. */ |
2814 | ||
2815 | /* ... Floats are not allowed */ | |
9e291dbe | 2816 | if (FLOAT_TYPE_P (TREE_TYPE (field))) |
3a2ea258 RE |
2817 | return 1; |
2818 | ||
d5b7b3ae RE |
2819 | /* ... Aggregates that are not themselves valid for returning in |
2820 | a register are not allowed. */ | |
81464b2c | 2821 | if (arm_return_in_memory (TREE_TYPE (field), NULL_TREE)) |
3a2ea258 | 2822 | return 1; |
6f7ebcbb | 2823 | |
3a2ea258 RE |
2824 | /* Now check the remaining fields, if any. Only bitfields are allowed, |
2825 | since they are not addressable. */ | |
f5a1b0d2 NC |
2826 | for (field = TREE_CHAIN (field); |
2827 | field; | |
2828 | field = TREE_CHAIN (field)) | |
2829 | { | |
2830 | if (TREE_CODE (field) != FIELD_DECL) | |
2831 | continue; | |
f676971a | 2832 | |
5895f793 | 2833 | if (!DECL_BIT_FIELD_TYPE (field)) |
f5a1b0d2 NC |
2834 | return 1; |
2835 | } | |
2b835d68 RE |
2836 | |
2837 | return 0; | |
2838 | } | |
f676971a | 2839 | |
d7d01975 | 2840 | if (TREE_CODE (type) == UNION_TYPE) |
2b835d68 RE |
2841 | { |
2842 | tree field; | |
2843 | ||
2844 | /* Unions can be returned in registers if every element is | |
2845 | integral, or can be returned in an integer register. */ | |
f5a1b0d2 NC |
2846 | for (field = TYPE_FIELDS (type); |
2847 | field; | |
2848 | field = TREE_CHAIN (field)) | |
2b835d68 | 2849 | { |
f5a1b0d2 NC |
2850 | if (TREE_CODE (field) != FIELD_DECL) |
2851 | continue; | |
2852 | ||
6cc8c0b3 NC |
2853 | if (FLOAT_TYPE_P (TREE_TYPE (field))) |
2854 | return 1; | |
f676971a | 2855 | |
81464b2c | 2856 | if (arm_return_in_memory (TREE_TYPE (field), NULL_TREE)) |
2b835d68 RE |
2857 | return 1; |
2858 | } | |
f676971a | 2859 | |
2b835d68 RE |
2860 | return 0; |
2861 | } | |
f676971a EC |
2862 | #endif /* not ARM_WINCE */ |
2863 | ||
d5b7b3ae | 2864 | /* Return all other types in memory. */ |
2b835d68 RE |
2865 | return 1; |
2866 | } | |
2867 | ||
d6b4baa4 | 2868 | /* Indicate whether or not words of a double are in big-endian order. */ |
3717da94 JT |
2869 | |
2870 | int | |
e32bac5b | 2871 | arm_float_words_big_endian (void) |
3717da94 | 2872 | { |
9b66ebb1 | 2873 | if (TARGET_MAVERICK) |
9b6b54e2 | 2874 | return 0; |
3717da94 JT |
2875 | |
2876 | /* For FPA, float words are always big-endian. For VFP, floats words | |
2877 | follow the memory system mode. */ | |
2878 | ||
9b66ebb1 | 2879 | if (TARGET_FPA) |
3717da94 | 2880 | { |
3717da94 JT |
2881 | return 1; |
2882 | } | |
2883 | ||
2884 | if (TARGET_VFP) | |
2885 | return (TARGET_BIG_END ? 1 : 0); | |
2886 | ||
2887 | return 1; | |
2888 | } | |
2889 | ||
82e9d970 PB |
2890 | /* Initialize a variable CUM of type CUMULATIVE_ARGS |
2891 | for a call to a function whose data type is FNTYPE. | |
2892 | For a library call, FNTYPE is NULL. */ | |
2893 | void | |
f676971a | 2894 | arm_init_cumulative_args (CUMULATIVE_ARGS *pcum, tree fntype, |
e32bac5b RE |
2895 | rtx libname ATTRIBUTE_UNUSED, |
2896 | tree fndecl ATTRIBUTE_UNUSED) | |
82e9d970 PB |
2897 | { |
2898 | /* On the ARM, the offset starts at 0. */ | |
29e339b9 | 2899 | pcum->nregs = 0; |
5a9335ef | 2900 | pcum->iwmmxt_nregs = 0; |
5848830f | 2901 | pcum->can_split = true; |
f676971a | 2902 | |
5a9335ef NC |
2903 | /* Varargs vectors are treated the same as long long. |
2904 | named_count avoids having to change the way arm handles 'named' */ | |
2905 | pcum->named_count = 0; | |
2906 | pcum->nargs = 0; | |
2907 | ||
2908 | if (TARGET_REALLY_IWMMXT && fntype) | |
2909 | { | |
2910 | tree fn_arg; | |
2911 | ||
2912 | for (fn_arg = TYPE_ARG_TYPES (fntype); | |
2913 | fn_arg; | |
2914 | fn_arg = TREE_CHAIN (fn_arg)) | |
2915 | pcum->named_count += 1; | |
2916 | ||
2917 | if (! pcum->named_count) | |
2918 | pcum->named_count = INT_MAX; | |
2919 | } | |
82e9d970 PB |
2920 | } |
2921 | ||
5848830f PB |
2922 | |
2923 | /* Return true if mode/type need doubleword alignment. */ | |
2924 | bool | |
2925 | arm_needs_doubleword_align (enum machine_mode mode, tree type) | |
2926 | { | |
65a939f7 PB |
2927 | return (GET_MODE_ALIGNMENT (mode) > PARM_BOUNDARY |
2928 | || (type && TYPE_ALIGN (type) > PARM_BOUNDARY)); | |
5848830f PB |
2929 | } |
2930 | ||
2931 | ||
82e9d970 PB |
2932 | /* Determine where to put an argument to a function. |
2933 | Value is zero to push the argument on the stack, | |
2934 | or a hard register in which to store the argument. | |
2935 | ||
2936 | MODE is the argument's machine mode. | |
2937 | TYPE is the data type of the argument (as a tree). | |
2938 | This is null for libcalls where that information may | |
2939 | not be available. | |
2940 | CUM is a variable of type CUMULATIVE_ARGS which gives info about | |
2941 | the preceding args and about the function being called. | |
2942 | NAMED is nonzero if this argument is a named parameter | |
2943 | (otherwise it is an extra parameter matching an ellipsis). */ | |
1d6e90ac | 2944 | |
82e9d970 | 2945 | rtx |
e32bac5b | 2946 | arm_function_arg (CUMULATIVE_ARGS *pcum, enum machine_mode mode, |
5848830f | 2947 | tree type, int named) |
82e9d970 | 2948 | { |
5848830f PB |
2949 | int nregs; |
2950 | ||
2951 | /* Varargs vectors are treated the same as long long. | |
2952 | named_count avoids having to change the way arm handles 'named' */ | |
2953 | if (TARGET_IWMMXT_ABI | |
f676971a | 2954 | && arm_vector_mode_supported_p (mode) |
5848830f | 2955 | && pcum->named_count > pcum->nargs + 1) |
5a9335ef | 2956 | { |
5848830f PB |
2957 | if (pcum->iwmmxt_nregs <= 9) |
2958 | return gen_rtx_REG (mode, pcum->iwmmxt_nregs + FIRST_IWMMXT_REGNUM); | |
2959 | else | |
5a9335ef | 2960 | { |
5848830f PB |
2961 | pcum->can_split = false; |
2962 | return NULL_RTX; | |
5a9335ef | 2963 | } |
5a9335ef NC |
2964 | } |
2965 | ||
5848830f PB |
2966 | /* Put doubleword aligned quantities in even register pairs. */ |
2967 | if (pcum->nregs & 1 | |
2968 | && ARM_DOUBLEWORD_ALIGN | |
2969 | && arm_needs_doubleword_align (mode, type)) | |
2970 | pcum->nregs++; | |
2971 | ||
82e9d970 | 2972 | if (mode == VOIDmode) |
25a65198 RS |
2973 | /* Pick an arbitrary value for operand 2 of the call insn. */ |
2974 | return const0_rtx; | |
5848830f | 2975 | |
666c27b9 | 2976 | /* Only allow splitting an arg between regs and memory if all preceding |
5848830f PB |
2977 | args were allocated to regs. For args passed by reference we only count |
2978 | the reference pointer. */ | |
2979 | if (pcum->can_split) | |
2980 | nregs = 1; | |
2981 | else | |
2982 | nregs = ARM_NUM_REGS2 (mode, type); | |
2983 | ||
2984 | if (!named || pcum->nregs + nregs > NUM_ARG_REGS) | |
82e9d970 | 2985 | return NULL_RTX; |
f676971a | 2986 | |
82e9d970 PB |
2987 | return gen_rtx_REG (mode, pcum->nregs); |
2988 | } | |
1741620c | 2989 | |
78a52f11 RH |
2990 | static int |
2991 | arm_arg_partial_bytes (CUMULATIVE_ARGS *pcum, enum machine_mode mode, | |
2992 | tree type, bool named ATTRIBUTE_UNUSED) | |
2993 | { | |
2994 | int nregs = pcum->nregs; | |
2995 | ||
88f77cba | 2996 | if (TARGET_IWMMXT_ABI && arm_vector_mode_supported_p (mode)) |
78a52f11 RH |
2997 | return 0; |
2998 | ||
2999 | if (NUM_ARG_REGS > nregs | |
3000 | && (NUM_ARG_REGS < nregs + ARM_NUM_REGS2 (mode, type)) | |
3001 | && pcum->can_split) | |
3002 | return (NUM_ARG_REGS - nregs) * UNITS_PER_WORD; | |
3003 | ||
3004 | return 0; | |
3005 | } | |
3006 | ||
1741620c JD |
3007 | /* Variable sized types are passed by reference. This is a GCC |
3008 | extension to the ARM ABI. */ | |
3009 | ||
8cd5a4e0 RH |
3010 | static bool |
3011 | arm_pass_by_reference (CUMULATIVE_ARGS *cum ATTRIBUTE_UNUSED, | |
3012 | enum machine_mode mode ATTRIBUTE_UNUSED, | |
586de218 | 3013 | const_tree type, bool named ATTRIBUTE_UNUSED) |
1741620c JD |
3014 | { |
3015 | return type && TREE_CODE (TYPE_SIZE (type)) != INTEGER_CST; | |
3016 | } | |
82e9d970 | 3017 | \f |
c27ba912 DM |
3018 | /* Encode the current state of the #pragma [no_]long_calls. */ |
3019 | typedef enum | |
82e9d970 | 3020 | { |
6fc0bb99 | 3021 | OFF, /* No #pragma [no_]long_calls is in effect. */ |
c27ba912 DM |
3022 | LONG, /* #pragma long_calls is in effect. */ |
3023 | SHORT /* #pragma no_long_calls is in effect. */ | |
3024 | } arm_pragma_enum; | |
82e9d970 | 3025 | |
c27ba912 | 3026 | static arm_pragma_enum arm_pragma_long_calls = OFF; |
82e9d970 | 3027 | |
8b97c5f8 | 3028 | void |
e32bac5b | 3029 | arm_pr_long_calls (struct cpp_reader * pfile ATTRIBUTE_UNUSED) |
82e9d970 | 3030 | { |
8b97c5f8 ZW |
3031 | arm_pragma_long_calls = LONG; |
3032 | } | |
3033 | ||
3034 | void | |
e32bac5b | 3035 | arm_pr_no_long_calls (struct cpp_reader * pfile ATTRIBUTE_UNUSED) |
8b97c5f8 ZW |
3036 | { |
3037 | arm_pragma_long_calls = SHORT; | |
3038 | } | |
3039 | ||
3040 | void | |
e32bac5b | 3041 | arm_pr_long_calls_off (struct cpp_reader * pfile ATTRIBUTE_UNUSED) |
8b97c5f8 ZW |
3042 | { |
3043 | arm_pragma_long_calls = OFF; | |
82e9d970 PB |
3044 | } |
3045 | \f | |
91d231cb JM |
3046 | /* Table of machine attributes. */ |
3047 | const struct attribute_spec arm_attribute_table[] = | |
82e9d970 | 3048 | { |
91d231cb | 3049 | /* { name, min_len, max_len, decl_req, type_req, fn_type_req, handler } */ |
82e9d970 PB |
3050 | /* Function calls made to this symbol must be done indirectly, because |
3051 | it may lie outside of the 26 bit addressing range of a normal function | |
3052 | call. */ | |
91d231cb | 3053 | { "long_call", 0, 0, false, true, true, NULL }, |
82e9d970 PB |
3054 | /* Whereas these functions are always known to reside within the 26 bit |
3055 | addressing range. */ | |
91d231cb | 3056 | { "short_call", 0, 0, false, true, true, NULL }, |
f676971a | 3057 | /* Interrupt Service Routines have special prologue and epilogue requirements. */ |
91d231cb JM |
3058 | { "isr", 0, 1, false, false, false, arm_handle_isr_attribute }, |
3059 | { "interrupt", 0, 1, false, false, false, arm_handle_isr_attribute }, | |
3060 | { "naked", 0, 0, true, false, false, arm_handle_fndecl_attribute }, | |
3061 | #ifdef ARM_PE | |
3062 | /* ARM/PE has three new attributes: | |
3063 | interfacearm - ? | |
3064 | dllexport - for exporting a function/variable that will live in a dll | |
3065 | dllimport - for importing a function/variable from a dll | |
3066 | ||
3067 | Microsoft allows multiple declspecs in one __declspec, separating | |
3068 | them with spaces. We do NOT support this. Instead, use __declspec | |
3069 | multiple times. | |
3070 | */ | |
3071 | { "dllimport", 0, 0, true, false, false, NULL }, | |
3072 | { "dllexport", 0, 0, true, false, false, NULL }, | |
3073 | { "interfacearm", 0, 0, true, false, false, arm_handle_fndecl_attribute }, | |
b2ca3702 MM |
3074 | #elif TARGET_DLLIMPORT_DECL_ATTRIBUTES |
3075 | { "dllimport", 0, 0, false, false, false, handle_dll_attribute }, | |
3076 | { "dllexport", 0, 0, false, false, false, handle_dll_attribute }, | |
04fb56d5 | 3077 | { "notshared", 0, 0, false, true, false, arm_handle_notshared_attribute }, |
91d231cb JM |
3078 | #endif |
3079 | { NULL, 0, 0, false, false, false, NULL } | |
3080 | }; | |
6d3d9133 | 3081 | |
91d231cb JM |
3082 | /* Handle an attribute requiring a FUNCTION_DECL; |
3083 | arguments as in struct attribute_spec.handler. */ | |
3084 | static tree | |
e32bac5b RE |
3085 | arm_handle_fndecl_attribute (tree *node, tree name, tree args ATTRIBUTE_UNUSED, |
3086 | int flags ATTRIBUTE_UNUSED, bool *no_add_attrs) | |
91d231cb JM |
3087 | { |
3088 | if (TREE_CODE (*node) != FUNCTION_DECL) | |
3089 | { | |
5c498b10 | 3090 | warning (OPT_Wattributes, "%qs attribute only applies to functions", |
91d231cb JM |
3091 | IDENTIFIER_POINTER (name)); |
3092 | *no_add_attrs = true; | |
3093 | } | |
3094 | ||
3095 | return NULL_TREE; | |
3096 | } | |
3097 | ||
3098 | /* Handle an "interrupt" or "isr" attribute; | |
3099 | arguments as in struct attribute_spec.handler. */ | |
3100 | static tree | |
e32bac5b RE |
3101 | arm_handle_isr_attribute (tree *node, tree name, tree args, int flags, |
3102 | bool *no_add_attrs) | |
91d231cb JM |
3103 | { |
3104 | if (DECL_P (*node)) | |
3105 | { | |
3106 | if (TREE_CODE (*node) != FUNCTION_DECL) | |
3107 | { | |
5c498b10 | 3108 | warning (OPT_Wattributes, "%qs attribute only applies to functions", |
91d231cb JM |
3109 | IDENTIFIER_POINTER (name)); |
3110 | *no_add_attrs = true; | |
3111 | } | |
3112 | /* FIXME: the argument if any is checked for type attributes; | |
3113 | should it be checked for decl ones? */ | |
3114 | } | |
3115 | else | |
3116 | { | |
3117 | if (TREE_CODE (*node) == FUNCTION_TYPE | |
3118 | || TREE_CODE (*node) == METHOD_TYPE) | |
3119 | { | |
3120 | if (arm_isr_value (args) == ARM_FT_UNKNOWN) | |
3121 | { | |
5c498b10 DD |
3122 | warning (OPT_Wattributes, "%qs attribute ignored", |
3123 | IDENTIFIER_POINTER (name)); | |
91d231cb JM |
3124 | *no_add_attrs = true; |
3125 | } | |
3126 | } | |
3127 | else if (TREE_CODE (*node) == POINTER_TYPE | |
3128 | && (TREE_CODE (TREE_TYPE (*node)) == FUNCTION_TYPE | |
3129 | || TREE_CODE (TREE_TYPE (*node)) == METHOD_TYPE) | |
3130 | && arm_isr_value (args) != ARM_FT_UNKNOWN) | |
3131 | { | |
8dd16ecc | 3132 | *node = build_variant_type_copy (*node); |
1d6e90ac NC |
3133 | TREE_TYPE (*node) = build_type_attribute_variant |
3134 | (TREE_TYPE (*node), | |
3135 | tree_cons (name, args, TYPE_ATTRIBUTES (TREE_TYPE (*node)))); | |
91d231cb JM |
3136 | *no_add_attrs = true; |
3137 | } | |
3138 | else | |
3139 | { | |
3140 | /* Possibly pass this attribute on from the type to a decl. */ | |
3141 | if (flags & ((int) ATTR_FLAG_DECL_NEXT | |
3142 | | (int) ATTR_FLAG_FUNCTION_NEXT | |
3143 | | (int) ATTR_FLAG_ARRAY_NEXT)) | |
3144 | { | |
3145 | *no_add_attrs = true; | |
3146 | return tree_cons (name, args, NULL_TREE); | |
3147 | } | |
3148 | else | |
3149 | { | |
5c498b10 DD |
3150 | warning (OPT_Wattributes, "%qs attribute ignored", |
3151 | IDENTIFIER_POINTER (name)); | |
91d231cb JM |
3152 | } |
3153 | } | |
3154 | } | |
3155 | ||
3156 | return NULL_TREE; | |
82e9d970 PB |
3157 | } |
3158 | ||
7bff66a7 | 3159 | #if TARGET_DLLIMPORT_DECL_ATTRIBUTES |
04fb56d5 MM |
3160 | /* Handle the "notshared" attribute. This attribute is another way of |
3161 | requesting hidden visibility. ARM's compiler supports | |
3162 | "__declspec(notshared)"; we support the same thing via an | |
3163 | attribute. */ | |
3164 | ||
3165 | static tree | |
e0b92319 NC |
3166 | arm_handle_notshared_attribute (tree *node, |
3167 | tree name ATTRIBUTE_UNUSED, | |
3168 | tree args ATTRIBUTE_UNUSED, | |
3169 | int flags ATTRIBUTE_UNUSED, | |
04fb56d5 MM |
3170 | bool *no_add_attrs) |
3171 | { | |
3172 | tree decl = TYPE_NAME (*node); | |
3173 | ||
3174 | if (decl) | |
3175 | { | |
3176 | DECL_VISIBILITY (decl) = VISIBILITY_HIDDEN; | |
3177 | DECL_VISIBILITY_SPECIFIED (decl) = 1; | |
3178 | *no_add_attrs = false; | |
3179 | } | |
3180 | return NULL_TREE; | |
3181 | } | |
7bff66a7 | 3182 | #endif |
04fb56d5 | 3183 | |
82e9d970 PB |
3184 | /* Return 0 if the attributes for two types are incompatible, 1 if they |
3185 | are compatible, and 2 if they are nearly compatible (which causes a | |
3186 | warning to be generated). */ | |
8d8e52be | 3187 | static int |
3101faab | 3188 | arm_comp_type_attributes (const_tree type1, const_tree type2) |
82e9d970 | 3189 | { |
1cb8d58a | 3190 | int l1, l2, s1, s2; |
f676971a | 3191 | |
82e9d970 PB |
3192 | /* Check for mismatch of non-default calling convention. */ |
3193 | if (TREE_CODE (type1) != FUNCTION_TYPE) | |
3194 | return 1; | |
3195 | ||
3196 | /* Check for mismatched call attributes. */ | |
1cb8d58a NC |
3197 | l1 = lookup_attribute ("long_call", TYPE_ATTRIBUTES (type1)) != NULL; |
3198 | l2 = lookup_attribute ("long_call", TYPE_ATTRIBUTES (type2)) != NULL; | |
3199 | s1 = lookup_attribute ("short_call", TYPE_ATTRIBUTES (type1)) != NULL; | |
3200 | s2 = lookup_attribute ("short_call", TYPE_ATTRIBUTES (type2)) != NULL; | |
bd7fc26f NC |
3201 | |
3202 | /* Only bother to check if an attribute is defined. */ | |
3203 | if (l1 | l2 | s1 | s2) | |
3204 | { | |
3205 | /* If one type has an attribute, the other must have the same attribute. */ | |
1cb8d58a | 3206 | if ((l1 != l2) || (s1 != s2)) |
bd7fc26f | 3207 | return 0; |
82e9d970 | 3208 | |
bd7fc26f NC |
3209 | /* Disallow mixed attributes. */ |
3210 | if ((l1 & s2) || (l2 & s1)) | |
3211 | return 0; | |
3212 | } | |
f676971a | 3213 | |
6d3d9133 NC |
3214 | /* Check for mismatched ISR attribute. */ |
3215 | l1 = lookup_attribute ("isr", TYPE_ATTRIBUTES (type1)) != NULL; | |
3216 | if (! l1) | |
3217 | l1 = lookup_attribute ("interrupt", TYPE_ATTRIBUTES (type1)) != NULL; | |
3218 | l2 = lookup_attribute ("isr", TYPE_ATTRIBUTES (type2)) != NULL; | |
3219 | if (! l2) | |
3220 | l1 = lookup_attribute ("interrupt", TYPE_ATTRIBUTES (type2)) != NULL; | |
3221 | if (l1 != l2) | |
3222 | return 0; | |
3223 | ||
bd7fc26f | 3224 | return 1; |
82e9d970 PB |
3225 | } |
3226 | ||
c27ba912 DM |
3227 | /* Assigns default attributes to newly defined type. This is used to |
3228 | set short_call/long_call attributes for function types of | |
3229 | functions defined inside corresponding #pragma scopes. */ | |
8d8e52be | 3230 | static void |
e32bac5b | 3231 | arm_set_default_type_attributes (tree type) |
c27ba912 DM |
3232 | { |
3233 | /* Add __attribute__ ((long_call)) to all functions, when | |
3234 | inside #pragma long_calls or __attribute__ ((short_call)), | |
3235 | when inside #pragma no_long_calls. */ | |
3236 | if (TREE_CODE (type) == FUNCTION_TYPE || TREE_CODE (type) == METHOD_TYPE) | |
3237 | { | |
3238 | tree type_attr_list, attr_name; | |
3239 | type_attr_list = TYPE_ATTRIBUTES (type); | |
3240 | ||
3241 | if (arm_pragma_long_calls == LONG) | |
3242 | attr_name = get_identifier ("long_call"); | |
3243 | else if (arm_pragma_long_calls == SHORT) | |
3244 | attr_name = get_identifier ("short_call"); | |
3245 | else | |
3246 | return; | |
3247 | ||
3248 | type_attr_list = tree_cons (attr_name, NULL_TREE, type_attr_list); | |
3249 | TYPE_ATTRIBUTES (type) = type_attr_list; | |
3250 | } | |
3251 | } | |
3252 | \f | |
25a65198 RS |
3253 | /* Return true if DECL is known to be linked into section SECTION. */ |
3254 | ||
3255 | static bool | |
3256 | arm_function_in_section_p (tree decl, section *section) | |
c27ba912 | 3257 | { |
25a65198 RS |
3258 | /* We can only be certain about functions defined in the same |
3259 | compilation unit. */ | |
3260 | if (!TREE_STATIC (decl)) | |
3261 | return false; | |
c27ba912 | 3262 | |
25a65198 RS |
3263 | /* Make sure that SYMBOL always binds to the definition in this |
3264 | compilation unit. */ | |
3265 | if (!targetm.binds_local_p (decl)) | |
3266 | return false; | |
c27ba912 | 3267 | |
25a65198 RS |
3268 | /* If DECL_SECTION_NAME is set, assume it is trustworthy. */ |
3269 | if (!DECL_SECTION_NAME (decl)) | |
3270 | { | |
25a65198 RS |
3271 | /* Make sure that we will not create a unique section for DECL. */ |
3272 | if (flag_function_sections || DECL_ONE_ONLY (decl)) | |
3273 | return false; | |
3274 | } | |
3275 | ||
3276 | return function_section (decl) == section; | |
c27ba912 DM |
3277 | } |
3278 | ||
a50aa827 | 3279 | /* Return nonzero if a 32-bit "long_call" should be generated for |
25a65198 RS |
3280 | a call from the current function to DECL. We generate a long_call |
3281 | if the function: | |
c27ba912 DM |
3282 | |
3283 | a. has an __attribute__((long call)) | |
3284 | or b. is within the scope of a #pragma long_calls | |
3285 | or c. the -mlong-calls command line switch has been specified | |
3286 | ||
3287 | However we do not generate a long call if the function: | |
f676971a | 3288 | |
c27ba912 DM |
3289 | d. has an __attribute__ ((short_call)) |
3290 | or e. is inside the scope of a #pragma no_long_calls | |
25a65198 | 3291 | or f. is defined in the same section as the current function. */ |
c27ba912 | 3292 | |
25a65198 RS |
3293 | bool |
3294 | arm_is_long_call_p (tree decl) | |
3295 | { | |
3296 | tree attrs; | |
c27ba912 | 3297 | |
25a65198 RS |
3298 | if (!decl) |
3299 | return TARGET_LONG_CALLS; | |
c27ba912 | 3300 | |
25a65198 RS |
3301 | attrs = TYPE_ATTRIBUTES (TREE_TYPE (decl)); |
3302 | if (lookup_attribute ("short_call", attrs)) | |
3303 | return false; | |
c27ba912 | 3304 | |
25a65198 RS |
3305 | /* For "f", be conservative, and only cater for cases in which the |
3306 | whole of the current function is placed in the same section. */ | |
3307 | if (!flag_reorder_blocks_and_partition | |
3308 | && arm_function_in_section_p (decl, current_function_section ())) | |
3309 | return false; | |
a77655b1 | 3310 | |
25a65198 RS |
3311 | if (lookup_attribute ("long_call", attrs)) |
3312 | return true; | |
f676971a | 3313 | |
25a65198 | 3314 | return TARGET_LONG_CALLS; |
c27ba912 | 3315 | } |
f99fce0c | 3316 | |
825dda42 | 3317 | /* Return nonzero if it is ok to make a tail-call to DECL. */ |
4977bab6 | 3318 | static bool |
e32bac5b | 3319 | arm_function_ok_for_sibcall (tree decl, tree exp ATTRIBUTE_UNUSED) |
f99fce0c | 3320 | { |
5b3e6663 | 3321 | unsigned long func_type; |
f99fce0c | 3322 | |
5a9335ef NC |
3323 | if (cfun->machine->sibcall_blocked) |
3324 | return false; | |
3325 | ||
f99fce0c RE |
3326 | /* Never tailcall something for which we have no decl, or if we |
3327 | are in Thumb mode. */ | |
3328 | if (decl == NULL || TARGET_THUMB) | |
4977bab6 | 3329 | return false; |
f99fce0c | 3330 | |
9403b7f7 RS |
3331 | /* The PIC register is live on entry to VxWorks PLT entries, so we |
3332 | must make the call before restoring the PIC register. */ | |
3333 | if (TARGET_VXWORKS_RTP && flag_pic && !targetm.binds_local_p (decl)) | |
3334 | return false; | |
3335 | ||
f99fce0c | 3336 | /* Cannot tail-call to long calls, since these are out of range of |
25a65198 RS |
3337 | a branch instruction. */ |
3338 | if (arm_is_long_call_p (decl)) | |
4977bab6 | 3339 | return false; |
f99fce0c RE |
3340 | |
3341 | /* If we are interworking and the function is not declared static | |
f676971a | 3342 | then we can't tail-call it unless we know that it exists in this |
f99fce0c | 3343 | compilation unit (since it might be a Thumb routine). */ |
5895f793 | 3344 | if (TARGET_INTERWORK && TREE_PUBLIC (decl) && !TREE_ASM_WRITTEN (decl)) |
4977bab6 | 3345 | return false; |
f99fce0c | 3346 | |
5b3e6663 | 3347 | func_type = arm_current_func_type (); |
6d3d9133 | 3348 | /* Never tailcall from an ISR routine - it needs a special exit sequence. */ |
5b3e6663 PB |
3349 | if (IS_INTERRUPT (func_type)) |
3350 | return false; | |
3351 | ||
3352 | /* Never tailcall if function may be called with a misaligned SP. */ | |
3353 | if (IS_STACKALIGN (func_type)) | |
4977bab6 | 3354 | return false; |
6d3d9133 | 3355 | |
f99fce0c | 3356 | /* Everything else is ok. */ |
4977bab6 | 3357 | return true; |
f99fce0c RE |
3358 | } |
3359 | ||
82e9d970 | 3360 | \f |
6b990f6b RE |
3361 | /* Addressing mode support functions. */ |
3362 | ||
0b4be7de | 3363 | /* Return nonzero if X is a legitimate immediate operand when compiling |
020a4035 | 3364 | for PIC. We know that X satisfies CONSTANT_P and flag_pic is true. */ |
32de079a | 3365 | int |
e32bac5b | 3366 | legitimate_pic_operand_p (rtx x) |
32de079a | 3367 | { |
020a4035 RE |
3368 | if (GET_CODE (x) == SYMBOL_REF |
3369 | || (GET_CODE (x) == CONST | |
3370 | && GET_CODE (XEXP (x, 0)) == PLUS | |
3371 | && GET_CODE (XEXP (XEXP (x, 0), 0)) == SYMBOL_REF)) | |
32de079a RE |
3372 | return 0; |
3373 | ||
3374 | return 1; | |
3375 | } | |
3376 | ||
9403b7f7 RS |
3377 | /* Record that the current function needs a PIC register. Initialize |
3378 | cfun->machine->pic_reg if we have not already done so. */ | |
3379 | ||
3380 | static void | |
3381 | require_pic_register (void) | |
3382 | { | |
3383 | /* A lot of the logic here is made obscure by the fact that this | |
3384 | routine gets called as part of the rtx cost estimation process. | |
3385 | We don't want those calls to affect any assumptions about the real | |
3386 | function; and further, we can't call entry_of_function() until we | |
3387 | start the real expansion process. */ | |
e3b5732b | 3388 | if (!crtl->uses_pic_offset_table) |
9403b7f7 | 3389 | { |
b3a13419 | 3390 | gcc_assert (can_create_pseudo_p ()); |
9403b7f7 RS |
3391 | if (arm_pic_register != INVALID_REGNUM) |
3392 | { | |
3393 | cfun->machine->pic_reg = gen_rtx_REG (Pmode, arm_pic_register); | |
3394 | ||
3395 | /* Play games to avoid marking the function as needing pic | |
3396 | if we are being called as part of the cost-estimation | |
3397 | process. */ | |
3398 | if (current_ir_type () != IR_GIMPLE) | |
e3b5732b | 3399 | crtl->uses_pic_offset_table = 1; |
9403b7f7 RS |
3400 | } |
3401 | else | |
3402 | { | |
3403 | rtx seq; | |
3404 | ||
3405 | cfun->machine->pic_reg = gen_reg_rtx (Pmode); | |
3406 | ||
3407 | /* Play games to avoid marking the function as needing pic | |
3408 | if we are being called as part of the cost-estimation | |
3409 | process. */ | |
3410 | if (current_ir_type () != IR_GIMPLE) | |
3411 | { | |
e3b5732b | 3412 | crtl->uses_pic_offset_table = 1; |
9403b7f7 RS |
3413 | start_sequence (); |
3414 | ||
3415 | arm_load_pic_register (0UL); | |
3416 | ||
3417 | seq = get_insns (); | |
3418 | end_sequence (); | |
3419 | emit_insn_after (seq, entry_of_function ()); | |
3420 | } | |
3421 | } | |
3422 | } | |
3423 | } | |
3424 | ||
32de079a | 3425 | rtx |
e32bac5b | 3426 | legitimize_pic_address (rtx orig, enum machine_mode mode, rtx reg) |
32de079a | 3427 | { |
a3c48721 RE |
3428 | if (GET_CODE (orig) == SYMBOL_REF |
3429 | || GET_CODE (orig) == LABEL_REF) | |
32de079a RE |
3430 | { |
3431 | rtx pic_ref, address; | |
3432 | rtx insn; | |
3433 | int subregs = 0; | |
3434 | ||
9403b7f7 RS |
3435 | /* If this function doesn't have a pic register, create one now. */ |
3436 | require_pic_register (); | |
020a4035 | 3437 | |
32de079a RE |
3438 | if (reg == 0) |
3439 | { | |
b3a13419 | 3440 | gcc_assert (can_create_pseudo_p ()); |
e6d29d15 | 3441 | reg = gen_reg_rtx (Pmode); |
32de079a RE |
3442 | |
3443 | subregs = 1; | |
3444 | } | |
3445 | ||
32de079a RE |
3446 | if (subregs) |
3447 | address = gen_reg_rtx (Pmode); | |
3448 | else | |
3449 | address = reg; | |
3450 | ||
4bec9f7d NC |
3451 | if (TARGET_ARM) |
3452 | emit_insn (gen_pic_load_addr_arm (address, orig)); | |
5b3e6663 PB |
3453 | else if (TARGET_THUMB2) |
3454 | emit_insn (gen_pic_load_addr_thumb2 (address, orig)); | |
3455 | else /* TARGET_THUMB1 */ | |
3456 | emit_insn (gen_pic_load_addr_thumb1 (address, orig)); | |
32de079a | 3457 | |
9403b7f7 RS |
3458 | /* VxWorks does not impose a fixed gap between segments; the run-time |
3459 | gap can be different from the object-file gap. We therefore can't | |
3460 | use GOTOFF unless we are absolutely sure that the symbol is in the | |
3461 | same segment as the GOT. Unfortunately, the flexibility of linker | |
3462 | scripts means that we can't be sure of that in general, so assume | |
3463 | that GOTOFF is never valid on VxWorks. */ | |
14f583b8 | 3464 | if ((GET_CODE (orig) == LABEL_REF |
f676971a | 3465 | || (GET_CODE (orig) == SYMBOL_REF && |
94428622 | 3466 | SYMBOL_REF_LOCAL_P (orig))) |
9403b7f7 RS |
3467 | && NEED_GOT_RELOC |
3468 | && !TARGET_VXWORKS_RTP) | |
020a4035 | 3469 | pic_ref = gen_rtx_PLUS (Pmode, cfun->machine->pic_reg, address); |
a3c48721 RE |
3470 | else |
3471 | { | |
542a8afa | 3472 | pic_ref = gen_const_mem (Pmode, |
020a4035 | 3473 | gen_rtx_PLUS (Pmode, cfun->machine->pic_reg, |
542a8afa | 3474 | address)); |
a3c48721 RE |
3475 | } |
3476 | ||
32de079a | 3477 | insn = emit_move_insn (reg, pic_ref); |
96a3900d | 3478 | |
32de079a RE |
3479 | /* Put a REG_EQUAL note on this insn, so that it can be optimized |
3480 | by loop. */ | |
bd94cb6e SB |
3481 | set_unique_reg_note (insn, REG_EQUAL, orig); |
3482 | ||
32de079a RE |
3483 | return reg; |
3484 | } | |
3485 | else if (GET_CODE (orig) == CONST) | |
3486 | { | |
3487 | rtx base, offset; | |
3488 | ||
3489 | if (GET_CODE (XEXP (orig, 0)) == PLUS | |
020a4035 | 3490 | && XEXP (XEXP (orig, 0), 0) == cfun->machine->pic_reg) |
32de079a RE |
3491 | return orig; |
3492 | ||
d3585b76 DJ |
3493 | if (GET_CODE (XEXP (orig, 0)) == UNSPEC |
3494 | && XINT (XEXP (orig, 0), 1) == UNSPEC_TLS) | |
3495 | return orig; | |
3496 | ||
32de079a RE |
3497 | if (reg == 0) |
3498 | { | |
b3a13419 | 3499 | gcc_assert (can_create_pseudo_p ()); |
e6d29d15 | 3500 | reg = gen_reg_rtx (Pmode); |
32de079a RE |
3501 | } |
3502 | ||
e6d29d15 | 3503 | gcc_assert (GET_CODE (XEXP (orig, 0)) == PLUS); |
e0b92319 | 3504 | |
e6d29d15 NS |
3505 | base = legitimize_pic_address (XEXP (XEXP (orig, 0), 0), Pmode, reg); |
3506 | offset = legitimize_pic_address (XEXP (XEXP (orig, 0), 1), Pmode, | |
3507 | base == reg ? 0 : reg); | |
32de079a RE |
3508 | |
3509 | if (GET_CODE (offset) == CONST_INT) | |
3510 | { | |
3511 | /* The base register doesn't really matter, we only want to | |
3512 | test the index for the appropriate mode. */ | |
1e1ab407 | 3513 | if (!arm_legitimate_index_p (mode, offset, SET, 0)) |
6b990f6b | 3514 | { |
b3a13419 | 3515 | gcc_assert (can_create_pseudo_p ()); |
e6d29d15 | 3516 | offset = force_reg (Pmode, offset); |
6b990f6b | 3517 | } |
32de079a | 3518 | |
32de079a | 3519 | if (GET_CODE (offset) == CONST_INT) |
ed8908e7 | 3520 | return plus_constant (base, INTVAL (offset)); |
32de079a RE |
3521 | } |
3522 | ||
3523 | if (GET_MODE_SIZE (mode) > 4 | |
3524 | && (GET_MODE_CLASS (mode) == MODE_INT | |
3525 | || TARGET_SOFT_FLOAT)) | |
3526 | { | |
3527 | emit_insn (gen_addsi3 (reg, base, offset)); | |
3528 | return reg; | |
3529 | } | |
3530 | ||
43cffd11 | 3531 | return gen_rtx_PLUS (Pmode, base, offset); |
32de079a | 3532 | } |
32de079a RE |
3533 | |
3534 | return orig; | |
3535 | } | |
3536 | ||
57934c39 | 3537 | |
5b3e6663 | 3538 | /* Find a spare register to use during the prolog of a function. */ |
57934c39 PB |
3539 | |
3540 | static int | |
b279b20a | 3541 | thumb_find_work_register (unsigned long pushed_regs_mask) |
57934c39 PB |
3542 | { |
3543 | int reg; | |
3544 | ||
b279b20a NC |
3545 | /* Check the argument registers first as these are call-used. The |
3546 | register allocation order means that sometimes r3 might be used | |
3547 | but earlier argument registers might not, so check them all. */ | |
3548 | for (reg = LAST_ARG_REGNUM; reg >= 0; reg --) | |
6fb5fa3c | 3549 | if (!df_regs_ever_live_p (reg)) |
b279b20a NC |
3550 | return reg; |
3551 | ||
3552 | /* Before going on to check the call-saved registers we can try a couple | |
3553 | more ways of deducing that r3 is available. The first is when we are | |
3554 | pushing anonymous arguments onto the stack and we have less than 4 | |
3555 | registers worth of fixed arguments(*). In this case r3 will be part of | |
3556 | the variable argument list and so we can be sure that it will be | |
3557 | pushed right at the start of the function. Hence it will be available | |
3558 | for the rest of the prologue. | |
38173d38 | 3559 | (*): ie crtl->args.pretend_args_size is greater than 0. */ |
b279b20a | 3560 | if (cfun->machine->uses_anonymous_args |
38173d38 | 3561 | && crtl->args.pretend_args_size > 0) |
57934c39 PB |
3562 | return LAST_ARG_REGNUM; |
3563 | ||
b279b20a NC |
3564 | /* The other case is when we have fixed arguments but less than 4 registers |
3565 | worth. In this case r3 might be used in the body of the function, but | |
3566 | it is not being used to convey an argument into the function. In theory | |
38173d38 | 3567 | we could just check crtl->args.size to see how many bytes are |
b279b20a NC |
3568 | being passed in argument registers, but it seems that it is unreliable. |
3569 | Sometimes it will have the value 0 when in fact arguments are being | |
3570 | passed. (See testcase execute/20021111-1.c for an example). So we also | |
3571 | check the args_info.nregs field as well. The problem with this field is | |
3572 | that it makes no allowances for arguments that are passed to the | |
3573 | function but which are not used. Hence we could miss an opportunity | |
3574 | when a function has an unused argument in r3. But it is better to be | |
3575 | safe than to be sorry. */ | |
3576 | if (! cfun->machine->uses_anonymous_args | |
38173d38 JH |
3577 | && crtl->args.size >= 0 |
3578 | && crtl->args.size <= (LAST_ARG_REGNUM * UNITS_PER_WORD) | |
f6d2671e | 3579 | && crtl->args.info.nregs < 4) |
b279b20a | 3580 | return LAST_ARG_REGNUM; |
e0b92319 | 3581 | |
b279b20a NC |
3582 | /* Otherwise look for a call-saved register that is going to be pushed. */ |
3583 | for (reg = LAST_LO_REGNUM; reg > LAST_ARG_REGNUM; reg --) | |
3584 | if (pushed_regs_mask & (1 << reg)) | |
57934c39 PB |
3585 | return reg; |
3586 | ||
5b3e6663 PB |
3587 | if (TARGET_THUMB2) |
3588 | { | |
3589 | /* Thumb-2 can use high regs. */ | |
3590 | for (reg = FIRST_HI_REGNUM; reg < 15; reg ++) | |
3591 | if (pushed_regs_mask & (1 << reg)) | |
3592 | return reg; | |
3593 | } | |
b279b20a NC |
3594 | /* Something went wrong - thumb_compute_save_reg_mask() |
3595 | should have arranged for a suitable register to be pushed. */ | |
e6d29d15 | 3596 | gcc_unreachable (); |
57934c39 PB |
3597 | } |
3598 | ||
f16fe45f | 3599 | static GTY(()) int pic_labelno; |
876f13b0 | 3600 | |
fe013435 PB |
3601 | /* Generate code to load the PIC register. In thumb mode SCRATCH is a |
3602 | low register. */ | |
876f13b0 | 3603 | |
32de079a | 3604 | void |
e55ef7f4 | 3605 | arm_load_pic_register (unsigned long saved_regs ATTRIBUTE_UNUSED) |
32de079a | 3606 | { |
9403b7f7 | 3607 | rtx l1, labelno, pic_tmp, pic_tmp2, pic_rtx, pic_reg; |
32de079a RE |
3608 | rtx global_offset_table; |
3609 | ||
e3b5732b | 3610 | if (crtl->uses_pic_offset_table == 0 || TARGET_SINGLE_PIC_BASE) |
32de079a RE |
3611 | return; |
3612 | ||
e6d29d15 | 3613 | gcc_assert (flag_pic); |
32de079a | 3614 | |
9403b7f7 RS |
3615 | pic_reg = cfun->machine->pic_reg; |
3616 | if (TARGET_VXWORKS_RTP) | |
3617 | { | |
3618 | pic_rtx = gen_rtx_SYMBOL_REF (Pmode, VXWORKS_GOTT_BASE); | |
3619 | pic_rtx = gen_rtx_CONST (Pmode, pic_rtx); | |
3620 | emit_insn (gen_pic_load_addr_arm (pic_reg, pic_rtx)); | |
43cffd11 | 3621 | |
9403b7f7 | 3622 | emit_insn (gen_rtx_SET (Pmode, pic_reg, gen_rtx_MEM (Pmode, pic_reg))); |
f676971a | 3623 | |
9403b7f7 RS |
3624 | pic_tmp = gen_rtx_SYMBOL_REF (Pmode, VXWORKS_GOTT_INDEX); |
3625 | emit_insn (gen_pic_offset_arm (pic_reg, pic_reg, pic_tmp)); | |
4bec9f7d | 3626 | } |
9403b7f7 | 3627 | else |
5b3e6663 | 3628 | { |
9403b7f7 RS |
3629 | /* We use an UNSPEC rather than a LABEL_REF because this label |
3630 | never appears in the code stream. */ | |
3631 | ||
3632 | labelno = GEN_INT (pic_labelno++); | |
3633 | l1 = gen_rtx_UNSPEC (Pmode, gen_rtvec (1, labelno), UNSPEC_PIC_LABEL); | |
3634 | l1 = gen_rtx_CONST (VOIDmode, l1); | |
3635 | ||
3636 | global_offset_table | |
3637 | = gen_rtx_SYMBOL_REF (Pmode, "_GLOBAL_OFFSET_TABLE_"); | |
3638 | /* On the ARM the PC register contains 'dot + 8' at the time of the | |
3639 | addition, on the Thumb it is 'dot + 4'. */ | |
3640 | pic_tmp = plus_constant (l1, TARGET_ARM ? 8 : 4); | |
3641 | if (GOT_PCREL) | |
5b3e6663 | 3642 | { |
9403b7f7 RS |
3643 | pic_tmp2 = gen_rtx_PLUS (Pmode, global_offset_table, pc_rtx); |
3644 | pic_tmp2 = gen_rtx_CONST (VOIDmode, pic_tmp2); | |
5b3e6663 PB |
3645 | } |
3646 | else | |
9403b7f7 RS |
3647 | pic_tmp2 = gen_rtx_CONST (VOIDmode, global_offset_table); |
3648 | ||
3649 | pic_rtx = gen_rtx_MINUS (Pmode, pic_tmp2, pic_tmp); | |
3650 | pic_rtx = gen_rtx_CONST (Pmode, pic_rtx); | |
3651 | ||
3652 | if (TARGET_ARM) | |
5b3e6663 | 3653 | { |
9403b7f7 RS |
3654 | emit_insn (gen_pic_load_addr_arm (pic_reg, pic_rtx)); |
3655 | emit_insn (gen_pic_add_dot_plus_eight (pic_reg, pic_reg, labelno)); | |
5b3e6663 | 3656 | } |
9403b7f7 RS |
3657 | else if (TARGET_THUMB2) |
3658 | { | |
3659 | /* Thumb-2 only allows very limited access to the PC. Calculate the | |
3660 | address in a temporary register. */ | |
3661 | if (arm_pic_register != INVALID_REGNUM) | |
3662 | { | |
3663 | pic_tmp = gen_rtx_REG (SImode, | |
3664 | thumb_find_work_register (saved_regs)); | |
3665 | } | |
3666 | else | |
3667 | { | |
b3a13419 | 3668 | gcc_assert (can_create_pseudo_p ()); |
9403b7f7 RS |
3669 | pic_tmp = gen_reg_rtx (Pmode); |
3670 | } | |
5b3e6663 | 3671 | |
9403b7f7 RS |
3672 | emit_insn (gen_pic_load_addr_thumb2 (pic_reg, pic_rtx)); |
3673 | emit_insn (gen_pic_load_dot_plus_four (pic_tmp, labelno)); | |
3674 | emit_insn (gen_addsi3 (pic_reg, pic_reg, pic_tmp)); | |
3675 | } | |
3676 | else /* TARGET_THUMB1 */ | |
876f13b0 | 3677 | { |
9403b7f7 RS |
3678 | if (arm_pic_register != INVALID_REGNUM |
3679 | && REGNO (pic_reg) > LAST_LO_REGNUM) | |
3680 | { | |
3681 | /* We will have pushed the pic register, so we should always be | |
3682 | able to find a work register. */ | |
3683 | pic_tmp = gen_rtx_REG (SImode, | |
3684 | thumb_find_work_register (saved_regs)); | |
3685 | emit_insn (gen_pic_load_addr_thumb1 (pic_tmp, pic_rtx)); | |
3686 | emit_insn (gen_movsi (pic_offset_table_rtx, pic_tmp)); | |
3687 | } | |
3688 | else | |
3689 | emit_insn (gen_pic_load_addr_thumb1 (pic_reg, pic_rtx)); | |
3690 | emit_insn (gen_pic_add_dot_plus_four (pic_reg, pic_reg, labelno)); | |
876f13b0 | 3691 | } |
4bec9f7d | 3692 | } |
32de079a | 3693 | |
32de079a RE |
3694 | /* Need to emit this whether or not we obey regdecls, |
3695 | since setjmp/longjmp can cause life info to screw up. */ | |
c41c1387 | 3696 | emit_use (pic_reg); |
32de079a RE |
3697 | } |
3698 | ||
876f13b0 | 3699 | |
6b990f6b RE |
3700 | /* Return nonzero if X is valid as an ARM state addressing register. */ |
3701 | static int | |
e32bac5b | 3702 | arm_address_register_rtx_p (rtx x, int strict_p) |
6b990f6b RE |
3703 | { |
3704 | int regno; | |
3705 | ||
3706 | if (GET_CODE (x) != REG) | |
3707 | return 0; | |
3708 | ||
3709 | regno = REGNO (x); | |
3710 | ||
3711 | if (strict_p) | |
3712 | return ARM_REGNO_OK_FOR_BASE_P (regno); | |
3713 | ||
3714 | return (regno <= LAST_ARM_REGNUM | |
3715 | || regno >= FIRST_PSEUDO_REGISTER | |
3716 | || regno == FRAME_POINTER_REGNUM | |
3717 | || regno == ARG_POINTER_REGNUM); | |
3718 | } | |
3719 | ||
d3585b76 DJ |
3720 | /* Return TRUE if this rtx is the difference of a symbol and a label, |
3721 | and will reduce to a PC-relative relocation in the object file. | |
3722 | Expressions like this can be left alone when generating PIC, rather | |
3723 | than forced through the GOT. */ | |
3724 | static int | |
3725 | pcrel_constant_p (rtx x) | |
3726 | { | |
3727 | if (GET_CODE (x) == MINUS) | |
3728 | return symbol_mentioned_p (XEXP (x, 0)) && label_mentioned_p (XEXP (x, 1)); | |
3729 | ||
3730 | return FALSE; | |
3731 | } | |
3732 | ||
6b990f6b RE |
3733 | /* Return nonzero if X is a valid ARM state address operand. */ |
3734 | int | |
1e1ab407 RE |
3735 | arm_legitimate_address_p (enum machine_mode mode, rtx x, RTX_CODE outer, |
3736 | int strict_p) | |
6b990f6b | 3737 | { |
fdd695fd PB |
3738 | bool use_ldrd; |
3739 | enum rtx_code code = GET_CODE (x); | |
f676971a | 3740 | |
6b990f6b RE |
3741 | if (arm_address_register_rtx_p (x, strict_p)) |
3742 | return 1; | |
3743 | ||
fdd695fd PB |
3744 | use_ldrd = (TARGET_LDRD |
3745 | && (mode == DImode | |
3746 | || (mode == DFmode && (TARGET_SOFT_FLOAT || TARGET_VFP)))); | |
3747 | ||
3748 | if (code == POST_INC || code == PRE_DEC | |
3749 | || ((code == PRE_INC || code == POST_DEC) | |
3750 | && (use_ldrd || GET_MODE_SIZE (mode) <= 4))) | |
6b990f6b RE |
3751 | return arm_address_register_rtx_p (XEXP (x, 0), strict_p); |
3752 | ||
fdd695fd | 3753 | else if ((code == POST_MODIFY || code == PRE_MODIFY) |
6b990f6b RE |
3754 | && arm_address_register_rtx_p (XEXP (x, 0), strict_p) |
3755 | && GET_CODE (XEXP (x, 1)) == PLUS | |
386d3a16 | 3756 | && rtx_equal_p (XEXP (XEXP (x, 1), 0), XEXP (x, 0))) |
fdd695fd PB |
3757 | { |
3758 | rtx addend = XEXP (XEXP (x, 1), 1); | |
3759 | ||
112cdef5 | 3760 | /* Don't allow ldrd post increment by register because it's hard |
fdd695fd PB |
3761 | to fixup invalid register choices. */ |
3762 | if (use_ldrd | |
3763 | && GET_CODE (x) == POST_MODIFY | |
3764 | && GET_CODE (addend) == REG) | |
3765 | return 0; | |
3766 | ||
3767 | return ((use_ldrd || GET_MODE_SIZE (mode) <= 4) | |
3768 | && arm_legitimate_index_p (mode, addend, outer, strict_p)); | |
3769 | } | |
6b990f6b RE |
3770 | |
3771 | /* After reload constants split into minipools will have addresses | |
3772 | from a LABEL_REF. */ | |
0bfb39ef | 3773 | else if (reload_completed |
fdd695fd PB |
3774 | && (code == LABEL_REF |
3775 | || (code == CONST | |
6b990f6b RE |
3776 | && GET_CODE (XEXP (x, 0)) == PLUS |
3777 | && GET_CODE (XEXP (XEXP (x, 0), 0)) == LABEL_REF | |
3778 | && GET_CODE (XEXP (XEXP (x, 0), 1)) == CONST_INT))) | |
3779 | return 1; | |
3780 | ||
88f77cba | 3781 | else if (mode == TImode || (TARGET_NEON && VALID_NEON_STRUCT_MODE (mode))) |
6b990f6b RE |
3782 | return 0; |
3783 | ||
fdd695fd | 3784 | else if (code == PLUS) |
6b990f6b RE |
3785 | { |
3786 | rtx xop0 = XEXP (x, 0); | |
3787 | rtx xop1 = XEXP (x, 1); | |
3788 | ||
3789 | return ((arm_address_register_rtx_p (xop0, strict_p) | |
1e1ab407 | 3790 | && arm_legitimate_index_p (mode, xop1, outer, strict_p)) |
6b990f6b | 3791 | || (arm_address_register_rtx_p (xop1, strict_p) |
1e1ab407 | 3792 | && arm_legitimate_index_p (mode, xop0, outer, strict_p))); |
6b990f6b RE |
3793 | } |
3794 | ||
3795 | #if 0 | |
3796 | /* Reload currently can't handle MINUS, so disable this for now */ | |
3797 | else if (GET_CODE (x) == MINUS) | |
3798 | { | |
3799 | rtx xop0 = XEXP (x, 0); | |
3800 | rtx xop1 = XEXP (x, 1); | |
3801 | ||
3802 | return (arm_address_register_rtx_p (xop0, strict_p) | |
1e1ab407 | 3803 | && arm_legitimate_index_p (mode, xop1, outer, strict_p)); |
6b990f6b RE |
3804 | } |
3805 | #endif | |
3806 | ||
3807 | else if (GET_MODE_CLASS (mode) != MODE_FLOAT | |
fdd695fd | 3808 | && code == SYMBOL_REF |
6b990f6b RE |
3809 | && CONSTANT_POOL_ADDRESS_P (x) |
3810 | && ! (flag_pic | |
d3585b76 DJ |
3811 | && symbol_mentioned_p (get_pool_constant (x)) |
3812 | && ! pcrel_constant_p (get_pool_constant (x)))) | |
6b990f6b RE |
3813 | return 1; |
3814 | ||
6b990f6b RE |
3815 | return 0; |
3816 | } | |
3817 | ||
5b3e6663 PB |
3818 | /* Return nonzero if X is a valid Thumb-2 address operand. */ |
3819 | int | |
3820 | thumb2_legitimate_address_p (enum machine_mode mode, rtx x, int strict_p) | |
3821 | { | |
3822 | bool use_ldrd; | |
3823 | enum rtx_code code = GET_CODE (x); | |
3824 | ||
3825 | if (arm_address_register_rtx_p (x, strict_p)) | |
3826 | return 1; | |
3827 | ||
3828 | use_ldrd = (TARGET_LDRD | |
3829 | && (mode == DImode | |
3830 | || (mode == DFmode && (TARGET_SOFT_FLOAT || TARGET_VFP)))); | |
3831 | ||
3832 | if (code == POST_INC || code == PRE_DEC | |
3833 | || ((code == PRE_INC || code == POST_DEC) | |
3834 | && (use_ldrd || GET_MODE_SIZE (mode) <= 4))) | |
3835 | return arm_address_register_rtx_p (XEXP (x, 0), strict_p); | |
3836 | ||
3837 | else if ((code == POST_MODIFY || code == PRE_MODIFY) | |
3838 | && arm_address_register_rtx_p (XEXP (x, 0), strict_p) | |
3839 | && GET_CODE (XEXP (x, 1)) == PLUS | |
3840 | && rtx_equal_p (XEXP (XEXP (x, 1), 0), XEXP (x, 0))) | |
3841 | { | |
3842 | /* Thumb-2 only has autoincrement by constant. */ | |
3843 | rtx addend = XEXP (XEXP (x, 1), 1); | |
3844 | HOST_WIDE_INT offset; | |
3845 | ||
3846 | if (GET_CODE (addend) != CONST_INT) | |
3847 | return 0; | |
3848 | ||
3849 | offset = INTVAL(addend); | |
3850 | if (GET_MODE_SIZE (mode) <= 4) | |
3851 | return (offset > -256 && offset < 256); | |
3852 | ||
3853 | return (use_ldrd && offset > -1024 && offset < 1024 | |
3854 | && (offset & 3) == 0); | |
3855 | } | |
3856 | ||
3857 | /* After reload constants split into minipools will have addresses | |
3858 | from a LABEL_REF. */ | |
3859 | else if (reload_completed | |
3860 | && (code == LABEL_REF | |
3861 | || (code == CONST | |
3862 | && GET_CODE (XEXP (x, 0)) == PLUS | |
3863 | && GET_CODE (XEXP (XEXP (x, 0), 0)) == LABEL_REF | |
3864 | && GET_CODE (XEXP (XEXP (x, 0), 1)) == CONST_INT))) | |
3865 | return 1; | |
3866 | ||
88f77cba | 3867 | else if (mode == TImode || (TARGET_NEON && VALID_NEON_STRUCT_MODE (mode))) |
5b3e6663 PB |
3868 | return 0; |
3869 | ||
3870 | else if (code == PLUS) | |
3871 | { | |
3872 | rtx xop0 = XEXP (x, 0); | |
3873 | rtx xop1 = XEXP (x, 1); | |
3874 | ||
3875 | return ((arm_address_register_rtx_p (xop0, strict_p) | |
3876 | && thumb2_legitimate_index_p (mode, xop1, strict_p)) | |
3877 | || (arm_address_register_rtx_p (xop1, strict_p) | |
3878 | && thumb2_legitimate_index_p (mode, xop0, strict_p))); | |
3879 | } | |
3880 | ||
3881 | else if (GET_MODE_CLASS (mode) != MODE_FLOAT | |
3882 | && code == SYMBOL_REF | |
3883 | && CONSTANT_POOL_ADDRESS_P (x) | |
3884 | && ! (flag_pic | |
3885 | && symbol_mentioned_p (get_pool_constant (x)) | |
3886 | && ! pcrel_constant_p (get_pool_constant (x)))) | |
3887 | return 1; | |
3888 | ||
3889 | return 0; | |
3890 | } | |
3891 | ||
6b990f6b RE |
3892 | /* Return nonzero if INDEX is valid for an address index operand in |
3893 | ARM state. */ | |
3894 | static int | |
1e1ab407 RE |
3895 | arm_legitimate_index_p (enum machine_mode mode, rtx index, RTX_CODE outer, |
3896 | int strict_p) | |
6b990f6b RE |
3897 | { |
3898 | HOST_WIDE_INT range; | |
3899 | enum rtx_code code = GET_CODE (index); | |
3900 | ||
778ebdd9 PB |
3901 | /* Standard coprocessor addressing modes. */ |
3902 | if (TARGET_HARD_FLOAT | |
3903 | && (TARGET_FPA || TARGET_MAVERICK) | |
3904 | && (GET_MODE_CLASS (mode) == MODE_FLOAT | |
3905 | || (TARGET_MAVERICK && mode == DImode))) | |
6b990f6b RE |
3906 | return (code == CONST_INT && INTVAL (index) < 1024 |
3907 | && INTVAL (index) > -1024 | |
3908 | && (INTVAL (index) & 3) == 0); | |
3909 | ||
88f77cba JB |
3910 | if (TARGET_NEON |
3911 | && (VALID_NEON_DREG_MODE (mode) || VALID_NEON_QREG_MODE (mode))) | |
3912 | return (code == CONST_INT | |
3913 | && INTVAL (index) < 1016 | |
3914 | && INTVAL (index) > -1024 | |
3915 | && (INTVAL (index) & 3) == 0); | |
3916 | ||
5a9335ef NC |
3917 | if (TARGET_REALLY_IWMMXT && VALID_IWMMXT_REG_MODE (mode)) |
3918 | return (code == CONST_INT | |
3657dc3e PB |
3919 | && INTVAL (index) < 1024 |
3920 | && INTVAL (index) > -1024 | |
3921 | && (INTVAL (index) & 3) == 0); | |
5a9335ef | 3922 | |
fdd695fd PB |
3923 | if (arm_address_register_rtx_p (index, strict_p) |
3924 | && (GET_MODE_SIZE (mode) <= 4)) | |
3925 | return 1; | |
3926 | ||
3927 | if (mode == DImode || mode == DFmode) | |
3928 | { | |
3929 | if (code == CONST_INT) | |
3930 | { | |
3931 | HOST_WIDE_INT val = INTVAL (index); | |
3932 | ||
3933 | if (TARGET_LDRD) | |
3934 | return val > -256 && val < 256; | |
3935 | else | |
f372c932 | 3936 | return val > -4096 && val < 4092; |
fdd695fd PB |
3937 | } |
3938 | ||
3939 | return TARGET_LDRD && arm_address_register_rtx_p (index, strict_p); | |
3940 | } | |
3941 | ||
6b990f6b | 3942 | if (GET_MODE_SIZE (mode) <= 4 |
1e1ab407 RE |
3943 | && ! (arm_arch4 |
3944 | && (mode == HImode | |
3945 | || (mode == QImode && outer == SIGN_EXTEND)))) | |
6b990f6b | 3946 | { |
1e1ab407 RE |
3947 | if (code == MULT) |
3948 | { | |
3949 | rtx xiop0 = XEXP (index, 0); | |
3950 | rtx xiop1 = XEXP (index, 1); | |
3951 | ||
3952 | return ((arm_address_register_rtx_p (xiop0, strict_p) | |
3953 | && power_of_two_operand (xiop1, SImode)) | |
3954 | || (arm_address_register_rtx_p (xiop1, strict_p) | |
3955 | && power_of_two_operand (xiop0, SImode))); | |
3956 | } | |
3957 | else if (code == LSHIFTRT || code == ASHIFTRT | |
3958 | || code == ASHIFT || code == ROTATERT) | |
3959 | { | |
3960 | rtx op = XEXP (index, 1); | |
6b990f6b | 3961 | |
1e1ab407 RE |
3962 | return (arm_address_register_rtx_p (XEXP (index, 0), strict_p) |
3963 | && GET_CODE (op) == CONST_INT | |
3964 | && INTVAL (op) > 0 | |
3965 | && INTVAL (op) <= 31); | |
3966 | } | |
6b990f6b RE |
3967 | } |
3968 | ||
1e1ab407 RE |
3969 | /* For ARM v4 we may be doing a sign-extend operation during the |
3970 | load. */ | |
e1471c91 | 3971 | if (arm_arch4) |
1e1ab407 RE |
3972 | { |
3973 | if (mode == HImode || (outer == SIGN_EXTEND && mode == QImode)) | |
3974 | range = 256; | |
3975 | else | |
3976 | range = 4096; | |
3977 | } | |
e1471c91 RE |
3978 | else |
3979 | range = (mode == HImode) ? 4095 : 4096; | |
6b990f6b RE |
3980 | |
3981 | return (code == CONST_INT | |
3982 | && INTVAL (index) < range | |
3983 | && INTVAL (index) > -range); | |
76a318e9 RE |
3984 | } |
3985 | ||
5b3e6663 PB |
3986 | /* Return true if OP is a valid index scaling factor for Thumb-2 address |
3987 | index operand. i.e. 1, 2, 4 or 8. */ | |
3988 | static bool | |
3989 | thumb2_index_mul_operand (rtx op) | |
3990 | { | |
3991 | HOST_WIDE_INT val; | |
3992 | ||
3993 | if (GET_CODE(op) != CONST_INT) | |
3994 | return false; | |
3995 | ||
3996 | val = INTVAL(op); | |
3997 | return (val == 1 || val == 2 || val == 4 || val == 8); | |
3998 | } | |
3999 | ||
4000 | /* Return nonzero if INDEX is a valid Thumb-2 address index operand. */ | |
4001 | static int | |
4002 | thumb2_legitimate_index_p (enum machine_mode mode, rtx index, int strict_p) | |
4003 | { | |
4004 | enum rtx_code code = GET_CODE (index); | |
4005 | ||
4006 | /* ??? Combine arm and thumb2 coprocessor addressing modes. */ | |
4007 | /* Standard coprocessor addressing modes. */ | |
4008 | if (TARGET_HARD_FLOAT | |
4009 | && (TARGET_FPA || TARGET_MAVERICK) | |
4010 | && (GET_MODE_CLASS (mode) == MODE_FLOAT | |
4011 | || (TARGET_MAVERICK && mode == DImode))) | |
4012 | return (code == CONST_INT && INTVAL (index) < 1024 | |
4013 | && INTVAL (index) > -1024 | |
4014 | && (INTVAL (index) & 3) == 0); | |
4015 | ||
4016 | if (TARGET_REALLY_IWMMXT && VALID_IWMMXT_REG_MODE (mode)) | |
fe2d934b PB |
4017 | { |
4018 | /* For DImode assume values will usually live in core regs | |
4019 | and only allow LDRD addressing modes. */ | |
4020 | if (!TARGET_LDRD || mode != DImode) | |
4021 | return (code == CONST_INT | |
4022 | && INTVAL (index) < 1024 | |
4023 | && INTVAL (index) > -1024 | |
4024 | && (INTVAL (index) & 3) == 0); | |
4025 | } | |
5b3e6663 | 4026 | |
88f77cba JB |
4027 | if (TARGET_NEON |
4028 | && (VALID_NEON_DREG_MODE (mode) || VALID_NEON_QREG_MODE (mode))) | |
4029 | return (code == CONST_INT | |
4030 | && INTVAL (index) < 1016 | |
4031 | && INTVAL (index) > -1024 | |
4032 | && (INTVAL (index) & 3) == 0); | |
4033 | ||
5b3e6663 PB |
4034 | if (arm_address_register_rtx_p (index, strict_p) |
4035 | && (GET_MODE_SIZE (mode) <= 4)) | |
4036 | return 1; | |
4037 | ||
4038 | if (mode == DImode || mode == DFmode) | |
4039 | { | |
4040 | HOST_WIDE_INT val = INTVAL (index); | |
4041 | /* ??? Can we assume ldrd for thumb2? */ | |
4042 | /* Thumb-2 ldrd only has reg+const addressing modes. */ | |
4043 | if (code != CONST_INT) | |
4044 | return 0; | |
4045 | ||
4046 | /* ldrd supports offsets of +-1020. | |
4047 | However the ldr fallback does not. */ | |
4048 | return val > -256 && val < 256 && (val & 3) == 0; | |
4049 | } | |
4050 | ||
4051 | if (code == MULT) | |
4052 | { | |
4053 | rtx xiop0 = XEXP (index, 0); | |
4054 | rtx xiop1 = XEXP (index, 1); | |
4055 | ||
4056 | return ((arm_address_register_rtx_p (xiop0, strict_p) | |
4057 | && thumb2_index_mul_operand (xiop1)) | |
4058 | || (arm_address_register_rtx_p (xiop1, strict_p) | |
4059 | && thumb2_index_mul_operand (xiop0))); | |
4060 | } | |
4061 | else if (code == ASHIFT) | |
4062 | { | |
4063 | rtx op = XEXP (index, 1); | |
4064 | ||
4065 | return (arm_address_register_rtx_p (XEXP (index, 0), strict_p) | |
4066 | && GET_CODE (op) == CONST_INT | |
4067 | && INTVAL (op) > 0 | |
4068 | && INTVAL (op) <= 3); | |
4069 | } | |
4070 | ||
4071 | return (code == CONST_INT | |
4072 | && INTVAL (index) < 4096 | |
4073 | && INTVAL (index) > -256); | |
4074 | } | |
4075 | ||
4076 | /* Return nonzero if X is valid as a 16-bit Thumb state base register. */ | |
76a318e9 | 4077 | static int |
5b3e6663 | 4078 | thumb1_base_register_rtx_p (rtx x, enum machine_mode mode, int strict_p) |
76a318e9 RE |
4079 | { |
4080 | int regno; | |
4081 | ||
4082 | if (GET_CODE (x) != REG) | |
4083 | return 0; | |
4084 | ||
4085 | regno = REGNO (x); | |
4086 | ||
4087 | if (strict_p) | |
5b3e6663 | 4088 | return THUMB1_REGNO_MODE_OK_FOR_BASE_P (regno, mode); |
76a318e9 RE |
4089 | |
4090 | return (regno <= LAST_LO_REGNUM | |
07e58265 | 4091 | || regno > LAST_VIRTUAL_REGISTER |
76a318e9 RE |
4092 | || regno == FRAME_POINTER_REGNUM |
4093 | || (GET_MODE_SIZE (mode) >= 4 | |
4094 | && (regno == STACK_POINTER_REGNUM | |
edf7cee8 | 4095 | || regno >= FIRST_PSEUDO_REGISTER |
76a318e9 RE |
4096 | || x == hard_frame_pointer_rtx |
4097 | || x == arg_pointer_rtx))); | |
4098 | } | |
4099 | ||
4100 | /* Return nonzero if x is a legitimate index register. This is the case | |
4101 | for any base register that can access a QImode object. */ | |
4102 | inline static int | |
5b3e6663 | 4103 | thumb1_index_register_rtx_p (rtx x, int strict_p) |
76a318e9 | 4104 | { |
5b3e6663 | 4105 | return thumb1_base_register_rtx_p (x, QImode, strict_p); |
76a318e9 RE |
4106 | } |
4107 | ||
5b3e6663 | 4108 | /* Return nonzero if x is a legitimate 16-bit Thumb-state address. |
f676971a | 4109 | |
76a318e9 RE |
4110 | The AP may be eliminated to either the SP or the FP, so we use the |
4111 | least common denominator, e.g. SImode, and offsets from 0 to 64. | |
4112 | ||
4113 | ??? Verify whether the above is the right approach. | |
4114 | ||
4115 | ??? Also, the FP may be eliminated to the SP, so perhaps that | |
4116 | needs special handling also. | |
4117 | ||
4118 | ??? Look at how the mips16 port solves this problem. It probably uses | |
4119 | better ways to solve some of these problems. | |
4120 | ||
4121 | Although it is not incorrect, we don't accept QImode and HImode | |
4122 | addresses based on the frame pointer or arg pointer until the | |
4123 | reload pass starts. This is so that eliminating such addresses | |
4124 | into stack based ones won't produce impossible code. */ | |
4125 | int | |
5b3e6663 | 4126 | thumb1_legitimate_address_p (enum machine_mode mode, rtx x, int strict_p) |
76a318e9 RE |
4127 | { |
4128 | /* ??? Not clear if this is right. Experiment. */ | |
4129 | if (GET_MODE_SIZE (mode) < 4 | |
4130 | && !(reload_in_progress || reload_completed) | |
4131 | && (reg_mentioned_p (frame_pointer_rtx, x) | |
4132 | || reg_mentioned_p (arg_pointer_rtx, x) | |
4133 | || reg_mentioned_p (virtual_incoming_args_rtx, x) | |
4134 | || reg_mentioned_p (virtual_outgoing_args_rtx, x) | |
4135 | || reg_mentioned_p (virtual_stack_dynamic_rtx, x) | |
4136 | || reg_mentioned_p (virtual_stack_vars_rtx, x))) | |
4137 | return 0; | |
4138 | ||
4139 | /* Accept any base register. SP only in SImode or larger. */ | |
5b3e6663 | 4140 | else if (thumb1_base_register_rtx_p (x, mode, strict_p)) |
76a318e9 RE |
4141 | return 1; |
4142 | ||
18dbd950 | 4143 | /* This is PC relative data before arm_reorg runs. */ |
76a318e9 RE |
4144 | else if (GET_MODE_SIZE (mode) >= 4 && CONSTANT_P (x) |
4145 | && GET_CODE (x) == SYMBOL_REF | |
020a4035 | 4146 | && CONSTANT_POOL_ADDRESS_P (x) && !flag_pic) |
76a318e9 RE |
4147 | return 1; |
4148 | ||
18dbd950 | 4149 | /* This is PC relative data after arm_reorg runs. */ |
76a318e9 RE |
4150 | else if (GET_MODE_SIZE (mode) >= 4 && reload_completed |
4151 | && (GET_CODE (x) == LABEL_REF | |
4152 | || (GET_CODE (x) == CONST | |
4153 | && GET_CODE (XEXP (x, 0)) == PLUS | |
4154 | && GET_CODE (XEXP (XEXP (x, 0), 0)) == LABEL_REF | |
4155 | && GET_CODE (XEXP (XEXP (x, 0), 1)) == CONST_INT))) | |
4156 | return 1; | |
4157 | ||
4158 | /* Post-inc indexing only supported for SImode and larger. */ | |
4159 | else if (GET_CODE (x) == POST_INC && GET_MODE_SIZE (mode) >= 4 | |
5b3e6663 | 4160 | && thumb1_index_register_rtx_p (XEXP (x, 0), strict_p)) |
76a318e9 RE |
4161 | return 1; |
4162 | ||
4163 | else if (GET_CODE (x) == PLUS) | |
4164 | { | |
4165 | /* REG+REG address can be any two index registers. */ | |
4166 | /* We disallow FRAME+REG addressing since we know that FRAME | |
4167 | will be replaced with STACK, and SP relative addressing only | |
4168 | permits SP+OFFSET. */ | |
4169 | if (GET_MODE_SIZE (mode) <= 4 | |
4170 | && XEXP (x, 0) != frame_pointer_rtx | |
4171 | && XEXP (x, 1) != frame_pointer_rtx | |
5b3e6663 PB |
4172 | && thumb1_index_register_rtx_p (XEXP (x, 0), strict_p) |
4173 | && thumb1_index_register_rtx_p (XEXP (x, 1), strict_p)) | |
76a318e9 RE |
4174 | return 1; |
4175 | ||
4176 | /* REG+const has 5-7 bit offset for non-SP registers. */ | |
5b3e6663 | 4177 | else if ((thumb1_index_register_rtx_p (XEXP (x, 0), strict_p) |
76a318e9 RE |
4178 | || XEXP (x, 0) == arg_pointer_rtx) |
4179 | && GET_CODE (XEXP (x, 1)) == CONST_INT | |
4180 | && thumb_legitimate_offset_p (mode, INTVAL (XEXP (x, 1)))) | |
4181 | return 1; | |
4182 | ||
a50aa827 | 4183 | /* REG+const has 10-bit offset for SP, but only SImode and |
76a318e9 RE |
4184 | larger is supported. */ |
4185 | /* ??? Should probably check for DI/DFmode overflow here | |
4186 | just like GO_IF_LEGITIMATE_OFFSET does. */ | |
4187 | else if (GET_CODE (XEXP (x, 0)) == REG | |
4188 | && REGNO (XEXP (x, 0)) == STACK_POINTER_REGNUM | |
4189 | && GET_MODE_SIZE (mode) >= 4 | |
4190 | && GET_CODE (XEXP (x, 1)) == CONST_INT | |
4191 | && INTVAL (XEXP (x, 1)) >= 0 | |
4192 | && INTVAL (XEXP (x, 1)) + GET_MODE_SIZE (mode) <= 1024 | |
4193 | && (INTVAL (XEXP (x, 1)) & 3) == 0) | |
4194 | return 1; | |
4195 | ||
4196 | else if (GET_CODE (XEXP (x, 0)) == REG | |
c5289e45 RE |
4197 | && (REGNO (XEXP (x, 0)) == FRAME_POINTER_REGNUM |
4198 | || REGNO (XEXP (x, 0)) == ARG_POINTER_REGNUM | |
4199 | || (REGNO (XEXP (x, 0)) >= FIRST_VIRTUAL_REGISTER | |
4200 | && REGNO (XEXP (x, 0)) <= LAST_VIRTUAL_REGISTER)) | |
76a318e9 RE |
4201 | && GET_MODE_SIZE (mode) >= 4 |
4202 | && GET_CODE (XEXP (x, 1)) == CONST_INT | |
4203 | && (INTVAL (XEXP (x, 1)) & 3) == 0) | |
4204 | return 1; | |
4205 | } | |
4206 | ||
4207 | else if (GET_MODE_CLASS (mode) != MODE_FLOAT | |
f954388e | 4208 | && GET_MODE_SIZE (mode) == 4 |
76a318e9 RE |
4209 | && GET_CODE (x) == SYMBOL_REF |
4210 | && CONSTANT_POOL_ADDRESS_P (x) | |
d3585b76 DJ |
4211 | && ! (flag_pic |
4212 | && symbol_mentioned_p (get_pool_constant (x)) | |
4213 | && ! pcrel_constant_p (get_pool_constant (x)))) | |
76a318e9 RE |
4214 | return 1; |
4215 | ||
4216 | return 0; | |
4217 | } | |
4218 | ||
4219 | /* Return nonzero if VAL can be used as an offset in a Thumb-state address | |
4220 | instruction of mode MODE. */ | |
4221 | int | |
e32bac5b | 4222 | thumb_legitimate_offset_p (enum machine_mode mode, HOST_WIDE_INT val) |
76a318e9 RE |
4223 | { |
4224 | switch (GET_MODE_SIZE (mode)) | |
4225 | { | |
4226 | case 1: | |
4227 | return val >= 0 && val < 32; | |
4228 | ||
4229 | case 2: | |
4230 | return val >= 0 && val < 64 && (val & 1) == 0; | |
4231 | ||
4232 | default: | |
4233 | return (val >= 0 | |
4234 | && (val + GET_MODE_SIZE (mode)) <= 128 | |
4235 | && (val & 3) == 0); | |
4236 | } | |
4237 | } | |
4238 | ||
d3585b76 DJ |
4239 | /* Build the SYMBOL_REF for __tls_get_addr. */ |
4240 | ||
4241 | static GTY(()) rtx tls_get_addr_libfunc; | |
4242 | ||
4243 | static rtx | |
4244 | get_tls_get_addr (void) | |
4245 | { | |
4246 | if (!tls_get_addr_libfunc) | |
4247 | tls_get_addr_libfunc = init_one_libfunc ("__tls_get_addr"); | |
4248 | return tls_get_addr_libfunc; | |
4249 | } | |
4250 | ||
4251 | static rtx | |
4252 | arm_load_tp (rtx target) | |
4253 | { | |
4254 | if (!target) | |
4255 | target = gen_reg_rtx (SImode); | |
4256 | ||
4257 | if (TARGET_HARD_TP) | |
4258 | { | |
4259 | /* Can return in any reg. */ | |
4260 | emit_insn (gen_load_tp_hard (target)); | |
4261 | } | |
4262 | else | |
4263 | { | |
4264 | /* Always returned in r0. Immediately copy the result into a pseudo, | |
4265 | otherwise other uses of r0 (e.g. setting up function arguments) may | |
4266 | clobber the value. */ | |
4267 | ||
4268 | rtx tmp; | |
4269 | ||
4270 | emit_insn (gen_load_tp_soft ()); | |
4271 | ||
4272 | tmp = gen_rtx_REG (SImode, 0); | |
4273 | emit_move_insn (target, tmp); | |
4274 | } | |
4275 | return target; | |
4276 | } | |
4277 | ||
4278 | static rtx | |
4279 | load_tls_operand (rtx x, rtx reg) | |
4280 | { | |
4281 | rtx tmp; | |
4282 | ||
4283 | if (reg == NULL_RTX) | |
4284 | reg = gen_reg_rtx (SImode); | |
4285 | ||
4286 | tmp = gen_rtx_CONST (SImode, x); | |
4287 | ||
4288 | emit_move_insn (reg, tmp); | |
4289 | ||
4290 | return reg; | |
4291 | } | |
4292 | ||
4293 | static rtx | |
4294 | arm_call_tls_get_addr (rtx x, rtx reg, rtx *valuep, int reloc) | |
4295 | { | |
f16fe45f | 4296 | rtx insns, label, labelno, sum; |
d3585b76 DJ |
4297 | |
4298 | start_sequence (); | |
4299 | ||
f16fe45f DJ |
4300 | labelno = GEN_INT (pic_labelno++); |
4301 | label = gen_rtx_UNSPEC (Pmode, gen_rtvec (1, labelno), UNSPEC_PIC_LABEL); | |
4302 | label = gen_rtx_CONST (VOIDmode, label); | |
4303 | ||
d3585b76 | 4304 | sum = gen_rtx_UNSPEC (Pmode, |
f16fe45f | 4305 | gen_rtvec (4, x, GEN_INT (reloc), label, |
d3585b76 DJ |
4306 | GEN_INT (TARGET_ARM ? 8 : 4)), |
4307 | UNSPEC_TLS); | |
4308 | reg = load_tls_operand (sum, reg); | |
4309 | ||
4310 | if (TARGET_ARM) | |
f16fe45f | 4311 | emit_insn (gen_pic_add_dot_plus_eight (reg, reg, labelno)); |
5b3e6663 PB |
4312 | else if (TARGET_THUMB2) |
4313 | { | |
4314 | rtx tmp; | |
4315 | /* Thumb-2 only allows very limited access to the PC. Calculate | |
4316 | the address in a temporary register. */ | |
4317 | tmp = gen_reg_rtx (SImode); | |
4318 | emit_insn (gen_pic_load_dot_plus_four (tmp, labelno)); | |
4319 | emit_insn (gen_addsi3(reg, reg, tmp)); | |
4320 | } | |
4321 | else /* TARGET_THUMB1 */ | |
f16fe45f | 4322 | emit_insn (gen_pic_add_dot_plus_four (reg, reg, labelno)); |
d3585b76 DJ |
4323 | |
4324 | *valuep = emit_library_call_value (get_tls_get_addr (), NULL_RTX, LCT_PURE, /* LCT_CONST? */ | |
4325 | Pmode, 1, reg, Pmode); | |
4326 | ||
4327 | insns = get_insns (); | |
4328 | end_sequence (); | |
4329 | ||
4330 | return insns; | |
4331 | } | |
4332 | ||
4333 | rtx | |
4334 | legitimize_tls_address (rtx x, rtx reg) | |
4335 | { | |
f16fe45f | 4336 | rtx dest, tp, label, labelno, sum, insns, ret, eqv, addend; |
d3585b76 DJ |
4337 | unsigned int model = SYMBOL_REF_TLS_MODEL (x); |
4338 | ||
4339 | switch (model) | |
4340 | { | |
4341 | case TLS_MODEL_GLOBAL_DYNAMIC: | |
4342 | insns = arm_call_tls_get_addr (x, reg, &ret, TLS_GD32); | |
4343 | dest = gen_reg_rtx (Pmode); | |
4344 | emit_libcall_block (insns, dest, ret, x); | |
4345 | return dest; | |
4346 | ||
4347 | case TLS_MODEL_LOCAL_DYNAMIC: | |
4348 | insns = arm_call_tls_get_addr (x, reg, &ret, TLS_LDM32); | |
4349 | ||
4350 | /* Attach a unique REG_EQUIV, to allow the RTL optimizers to | |
4351 | share the LDM result with other LD model accesses. */ | |
4352 | eqv = gen_rtx_UNSPEC (Pmode, gen_rtvec (1, const1_rtx), | |
4353 | UNSPEC_TLS); | |
4354 | dest = gen_reg_rtx (Pmode); | |
e66e1c68 | 4355 | emit_libcall_block (insns, dest, ret, eqv); |
d3585b76 DJ |
4356 | |
4357 | /* Load the addend. */ | |
4358 | addend = gen_rtx_UNSPEC (Pmode, gen_rtvec (2, x, GEN_INT (TLS_LDO32)), | |
4359 | UNSPEC_TLS); | |
4360 | addend = force_reg (SImode, gen_rtx_CONST (SImode, addend)); | |
4361 | return gen_rtx_PLUS (Pmode, dest, addend); | |
4362 | ||
4363 | case TLS_MODEL_INITIAL_EXEC: | |
f16fe45f DJ |
4364 | labelno = GEN_INT (pic_labelno++); |
4365 | label = gen_rtx_UNSPEC (Pmode, gen_rtvec (1, labelno), UNSPEC_PIC_LABEL); | |
4366 | label = gen_rtx_CONST (VOIDmode, label); | |
d3585b76 | 4367 | sum = gen_rtx_UNSPEC (Pmode, |
f16fe45f | 4368 | gen_rtvec (4, x, GEN_INT (TLS_IE32), label, |
d3585b76 DJ |
4369 | GEN_INT (TARGET_ARM ? 8 : 4)), |
4370 | UNSPEC_TLS); | |
4371 | reg = load_tls_operand (sum, reg); | |
4372 | ||
4373 | if (TARGET_ARM) | |
f16fe45f | 4374 | emit_insn (gen_tls_load_dot_plus_eight (reg, reg, labelno)); |
5b3e6663 PB |
4375 | else if (TARGET_THUMB2) |
4376 | { | |
4377 | rtx tmp; | |
4378 | /* Thumb-2 only allows very limited access to the PC. Calculate | |
4379 | the address in a temporary register. */ | |
4380 | tmp = gen_reg_rtx (SImode); | |
4381 | emit_insn (gen_pic_load_dot_plus_four (tmp, labelno)); | |
4382 | emit_insn (gen_addsi3(reg, reg, tmp)); | |
4383 | emit_move_insn (reg, gen_const_mem (SImode, reg)); | |
4384 | } | |
d3585b76 DJ |
4385 | else |
4386 | { | |
f16fe45f | 4387 | emit_insn (gen_pic_add_dot_plus_four (reg, reg, labelno)); |
d3585b76 DJ |
4388 | emit_move_insn (reg, gen_const_mem (SImode, reg)); |
4389 | } | |
4390 | ||
4391 | tp = arm_load_tp (NULL_RTX); | |
4392 | ||
4393 | return gen_rtx_PLUS (Pmode, tp, reg); | |
4394 | ||
4395 | case TLS_MODEL_LOCAL_EXEC: | |
4396 | tp = arm_load_tp (NULL_RTX); | |
4397 | ||
4398 | reg = gen_rtx_UNSPEC (Pmode, | |
4399 | gen_rtvec (2, x, GEN_INT (TLS_LE32)), | |
4400 | UNSPEC_TLS); | |
4401 | reg = force_reg (SImode, gen_rtx_CONST (SImode, reg)); | |
4402 | ||
4403 | return gen_rtx_PLUS (Pmode, tp, reg); | |
4404 | ||
4405 | default: | |
4406 | abort (); | |
4407 | } | |
4408 | } | |
4409 | ||
ccf4d512 RE |
4410 | /* Try machine-dependent ways of modifying an illegitimate address |
4411 | to be legitimate. If we find one, return the new, valid address. */ | |
ccf4d512 | 4412 | rtx |
e32bac5b | 4413 | arm_legitimize_address (rtx x, rtx orig_x, enum machine_mode mode) |
ccf4d512 | 4414 | { |
d3585b76 DJ |
4415 | if (arm_tls_symbol_p (x)) |
4416 | return legitimize_tls_address (x, NULL_RTX); | |
4417 | ||
ccf4d512 RE |
4418 | if (GET_CODE (x) == PLUS) |
4419 | { | |
4420 | rtx xop0 = XEXP (x, 0); | |
4421 | rtx xop1 = XEXP (x, 1); | |
4422 | ||
4423 | if (CONSTANT_P (xop0) && !symbol_mentioned_p (xop0)) | |
4424 | xop0 = force_reg (SImode, xop0); | |
4425 | ||
4426 | if (CONSTANT_P (xop1) && !symbol_mentioned_p (xop1)) | |
4427 | xop1 = force_reg (SImode, xop1); | |
4428 | ||
4429 | if (ARM_BASE_REGISTER_RTX_P (xop0) | |
4430 | && GET_CODE (xop1) == CONST_INT) | |
4431 | { | |
4432 | HOST_WIDE_INT n, low_n; | |
4433 | rtx base_reg, val; | |
4434 | n = INTVAL (xop1); | |
4435 | ||
9b66ebb1 PB |
4436 | /* VFP addressing modes actually allow greater offsets, but for |
4437 | now we just stick with the lowest common denominator. */ | |
4438 | if (mode == DImode | |
4439 | || ((TARGET_SOFT_FLOAT || TARGET_VFP) && mode == DFmode)) | |
ccf4d512 RE |
4440 | { |
4441 | low_n = n & 0x0f; | |
4442 | n &= ~0x0f; | |
4443 | if (low_n > 4) | |
4444 | { | |
4445 | n += 16; | |
4446 | low_n -= 16; | |
4447 | } | |
4448 | } | |
4449 | else | |
4450 | { | |
4451 | low_n = ((mode) == TImode ? 0 | |
4452 | : n >= 0 ? (n & 0xfff) : -((-n) & 0xfff)); | |
4453 | n -= low_n; | |
4454 | } | |
4455 | ||
4456 | base_reg = gen_reg_rtx (SImode); | |
d66437c5 | 4457 | val = force_operand (plus_constant (xop0, n), NULL_RTX); |
ccf4d512 | 4458 | emit_move_insn (base_reg, val); |
d66437c5 | 4459 | x = plus_constant (base_reg, low_n); |
ccf4d512 RE |
4460 | } |
4461 | else if (xop0 != XEXP (x, 0) || xop1 != XEXP (x, 1)) | |
4462 | x = gen_rtx_PLUS (SImode, xop0, xop1); | |
4463 | } | |
4464 | ||
4465 | /* XXX We don't allow MINUS any more -- see comment in | |
4466 | arm_legitimate_address_p (). */ | |
4467 | else if (GET_CODE (x) == MINUS) | |
4468 | { | |
4469 | rtx xop0 = XEXP (x, 0); | |
4470 | rtx xop1 = XEXP (x, 1); | |
4471 | ||
4472 | if (CONSTANT_P (xop0)) | |
4473 | xop0 = force_reg (SImode, xop0); | |
4474 | ||
4475 | if (CONSTANT_P (xop1) && ! symbol_mentioned_p (xop1)) | |
4476 | xop1 = force_reg (SImode, xop1); | |
4477 | ||
4478 | if (xop0 != XEXP (x, 0) || xop1 != XEXP (x, 1)) | |
4479 | x = gen_rtx_MINUS (SImode, xop0, xop1); | |
4480 | } | |
4481 | ||
86805759 NP |
4482 | /* Make sure to take full advantage of the pre-indexed addressing mode |
4483 | with absolute addresses which often allows for the base register to | |
4484 | be factorized for multiple adjacent memory references, and it might | |
4485 | even allows for the mini pool to be avoided entirely. */ | |
4486 | else if (GET_CODE (x) == CONST_INT && optimize > 0) | |
4487 | { | |
4488 | unsigned int bits; | |
4489 | HOST_WIDE_INT mask, base, index; | |
4490 | rtx base_reg; | |
4491 | ||
a50aa827 KH |
4492 | /* ldr and ldrb can use a 12-bit index, ldrsb and the rest can only |
4493 | use a 8-bit index. So let's use a 12-bit index for SImode only and | |
86805759 NP |
4494 | hope that arm_gen_constant will enable ldrb to use more bits. */ |
4495 | bits = (mode == SImode) ? 12 : 8; | |
4496 | mask = (1 << bits) - 1; | |
4497 | base = INTVAL (x) & ~mask; | |
4498 | index = INTVAL (x) & mask; | |
b107d743 | 4499 | if (bit_count (base & 0xffffffff) > (32 - bits)/2) |
86805759 NP |
4500 | { |
4501 | /* It'll most probably be more efficient to generate the base | |
4502 | with more bits set and use a negative index instead. */ | |
4503 | base |= mask; | |
4504 | index -= mask; | |
4505 | } | |
4506 | base_reg = force_reg (SImode, GEN_INT (base)); | |
d66437c5 | 4507 | x = plus_constant (base_reg, index); |
86805759 NP |
4508 | } |
4509 | ||
ccf4d512 RE |
4510 | if (flag_pic) |
4511 | { | |
4512 | /* We need to find and carefully transform any SYMBOL and LABEL | |
4513 | references; so go back to the original address expression. */ | |
4514 | rtx new_x = legitimize_pic_address (orig_x, mode, NULL_RTX); | |
4515 | ||
4516 | if (new_x != orig_x) | |
4517 | x = new_x; | |
4518 | } | |
4519 | ||
4520 | return x; | |
4521 | } | |
4522 | ||
6f5b4f3e RE |
4523 | |
4524 | /* Try machine-dependent ways of modifying an illegitimate Thumb address | |
4525 | to be legitimate. If we find one, return the new, valid address. */ | |
4526 | rtx | |
4527 | thumb_legitimize_address (rtx x, rtx orig_x, enum machine_mode mode) | |
4528 | { | |
d3585b76 DJ |
4529 | if (arm_tls_symbol_p (x)) |
4530 | return legitimize_tls_address (x, NULL_RTX); | |
4531 | ||
6f5b4f3e RE |
4532 | if (GET_CODE (x) == PLUS |
4533 | && GET_CODE (XEXP (x, 1)) == CONST_INT | |
4534 | && (INTVAL (XEXP (x, 1)) >= 32 * GET_MODE_SIZE (mode) | |
4535 | || INTVAL (XEXP (x, 1)) < 0)) | |
4536 | { | |
4537 | rtx xop0 = XEXP (x, 0); | |
4538 | rtx xop1 = XEXP (x, 1); | |
4539 | HOST_WIDE_INT offset = INTVAL (xop1); | |
4540 | ||
4541 | /* Try and fold the offset into a biasing of the base register and | |
4542 | then offsetting that. Don't do this when optimizing for space | |
4543 | since it can cause too many CSEs. */ | |
4544 | if (optimize_size && offset >= 0 | |
4545 | && offset < 256 + 31 * GET_MODE_SIZE (mode)) | |
4546 | { | |
4547 | HOST_WIDE_INT delta; | |
4548 | ||
4549 | if (offset >= 256) | |
4550 | delta = offset - (256 - GET_MODE_SIZE (mode)); | |
4551 | else if (offset < 32 * GET_MODE_SIZE (mode) + 8) | |
4552 | delta = 31 * GET_MODE_SIZE (mode); | |
4553 | else | |
4554 | delta = offset & (~31 * GET_MODE_SIZE (mode)); | |
4555 | ||
4556 | xop0 = force_operand (plus_constant (xop0, offset - delta), | |
4557 | NULL_RTX); | |
4558 | x = plus_constant (xop0, delta); | |
4559 | } | |
4560 | else if (offset < 0 && offset > -256) | |
4561 | /* Small negative offsets are best done with a subtract before the | |
4562 | dereference, forcing these into a register normally takes two | |
4563 | instructions. */ | |
4564 | x = force_operand (x, NULL_RTX); | |
4565 | else | |
4566 | { | |
4567 | /* For the remaining cases, force the constant into a register. */ | |
4568 | xop1 = force_reg (SImode, xop1); | |
4569 | x = gen_rtx_PLUS (SImode, xop0, xop1); | |
4570 | } | |
4571 | } | |
4572 | else if (GET_CODE (x) == PLUS | |
4573 | && s_register_operand (XEXP (x, 1), SImode) | |
4574 | && !s_register_operand (XEXP (x, 0), SImode)) | |
4575 | { | |
4576 | rtx xop0 = force_operand (XEXP (x, 0), NULL_RTX); | |
4577 | ||
4578 | x = gen_rtx_PLUS (SImode, xop0, XEXP (x, 1)); | |
4579 | } | |
4580 | ||
4581 | if (flag_pic) | |
4582 | { | |
4583 | /* We need to find and carefully transform any SYMBOL and LABEL | |
4584 | references; so go back to the original address expression. */ | |
4585 | rtx new_x = legitimize_pic_address (orig_x, mode, NULL_RTX); | |
4586 | ||
4587 | if (new_x != orig_x) | |
4588 | x = new_x; | |
4589 | } | |
4590 | ||
4591 | return x; | |
4592 | } | |
4593 | ||
a132dad6 | 4594 | rtx |
e0b92319 NC |
4595 | thumb_legitimize_reload_address (rtx *x_p, |
4596 | enum machine_mode mode, | |
4597 | int opnum, int type, | |
4598 | int ind_levels ATTRIBUTE_UNUSED) | |
a132dad6 RE |
4599 | { |
4600 | rtx x = *x_p; | |
e0b92319 | 4601 | |
a132dad6 RE |
4602 | if (GET_CODE (x) == PLUS |
4603 | && GET_MODE_SIZE (mode) < 4 | |
4604 | && REG_P (XEXP (x, 0)) | |
4605 | && XEXP (x, 0) == stack_pointer_rtx | |
4606 | && GET_CODE (XEXP (x, 1)) == CONST_INT | |
4607 | && !thumb_legitimate_offset_p (mode, INTVAL (XEXP (x, 1)))) | |
4608 | { | |
4609 | rtx orig_x = x; | |
4610 | ||
4611 | x = copy_rtx (x); | |
4612 | push_reload (orig_x, NULL_RTX, x_p, NULL, MODE_BASE_REG_CLASS (mode), | |
4613 | Pmode, VOIDmode, 0, 0, opnum, type); | |
4614 | return x; | |
4615 | } | |
4616 | ||
4617 | /* If both registers are hi-regs, then it's better to reload the | |
4618 | entire expression rather than each register individually. That | |
4619 | only requires one reload register rather than two. */ | |
4620 | if (GET_CODE (x) == PLUS | |
4621 | && REG_P (XEXP (x, 0)) | |
4622 | && REG_P (XEXP (x, 1)) | |
4623 | && !REG_MODE_OK_FOR_REG_BASE_P (XEXP (x, 0), mode) | |
4624 | && !REG_MODE_OK_FOR_REG_BASE_P (XEXP (x, 1), mode)) | |
4625 | { | |
4626 | rtx orig_x = x; | |
4627 | ||
4628 | x = copy_rtx (x); | |
4629 | push_reload (orig_x, NULL_RTX, x_p, NULL, MODE_BASE_REG_CLASS (mode), | |
4630 | Pmode, VOIDmode, 0, 0, opnum, type); | |
4631 | return x; | |
4632 | } | |
4633 | ||
4634 | return NULL; | |
4635 | } | |
d3585b76 DJ |
4636 | |
4637 | /* Test for various thread-local symbols. */ | |
4638 | ||
4639 | /* Return TRUE if X is a thread-local symbol. */ | |
4640 | ||
4641 | static bool | |
4642 | arm_tls_symbol_p (rtx x) | |
4643 | { | |
4644 | if (! TARGET_HAVE_TLS) | |
4645 | return false; | |
4646 | ||
4647 | if (GET_CODE (x) != SYMBOL_REF) | |
4648 | return false; | |
4649 | ||
4650 | return SYMBOL_REF_TLS_MODEL (x) != 0; | |
4651 | } | |
4652 | ||
4653 | /* Helper for arm_tls_referenced_p. */ | |
4654 | ||
4655 | static int | |
4656 | arm_tls_operand_p_1 (rtx *x, void *data ATTRIBUTE_UNUSED) | |
4657 | { | |
4658 | if (GET_CODE (*x) == SYMBOL_REF) | |
4659 | return SYMBOL_REF_TLS_MODEL (*x) != 0; | |
4660 | ||
4661 | /* Don't recurse into UNSPEC_TLS looking for TLS symbols; these are | |
4662 | TLS offsets, not real symbol references. */ | |
4663 | if (GET_CODE (*x) == UNSPEC | |
4664 | && XINT (*x, 1) == UNSPEC_TLS) | |
4665 | return -1; | |
4666 | ||
4667 | return 0; | |
4668 | } | |
4669 | ||
4670 | /* Return TRUE if X contains any TLS symbol references. */ | |
4671 | ||
4672 | bool | |
4673 | arm_tls_referenced_p (rtx x) | |
4674 | { | |
4675 | if (! TARGET_HAVE_TLS) | |
4676 | return false; | |
4677 | ||
4678 | return for_each_rtx (&x, arm_tls_operand_p_1, NULL); | |
4679 | } | |
8426b956 RS |
4680 | |
4681 | /* Implement TARGET_CANNOT_FORCE_CONST_MEM. */ | |
4682 | ||
4683 | bool | |
4684 | arm_cannot_force_const_mem (rtx x) | |
4685 | { | |
4686 | rtx base, offset; | |
4687 | ||
4688 | if (ARM_OFFSETS_MUST_BE_WITHIN_SECTIONS_P) | |
4689 | { | |
4690 | split_const (x, &base, &offset); | |
4691 | if (GET_CODE (base) == SYMBOL_REF | |
4692 | && !offset_within_block_p (base, INTVAL (offset))) | |
4693 | return true; | |
4694 | } | |
4695 | return arm_tls_referenced_p (x); | |
4696 | } | |
6b990f6b | 4697 | \f |
e2c671ba RE |
4698 | #define REG_OR_SUBREG_REG(X) \ |
4699 | (GET_CODE (X) == REG \ | |
4700 | || (GET_CODE (X) == SUBREG && GET_CODE (SUBREG_REG (X)) == REG)) | |
4701 | ||
4702 | #define REG_OR_SUBREG_RTX(X) \ | |
4703 | (GET_CODE (X) == REG ? (X) : SUBREG_REG (X)) | |
4704 | ||
d5b7b3ae RE |
4705 | #ifndef COSTS_N_INSNS |
4706 | #define COSTS_N_INSNS(N) ((N) * 4 - 2) | |
4707 | #endif | |
3c50106f | 4708 | static inline int |
5b3e6663 | 4709 | thumb1_rtx_costs (rtx x, enum rtx_code code, enum rtx_code outer) |
e2c671ba RE |
4710 | { |
4711 | enum machine_mode mode = GET_MODE (x); | |
e2c671ba | 4712 | |
9b66ebb1 | 4713 | switch (code) |
d5b7b3ae | 4714 | { |
9b66ebb1 PB |
4715 | case ASHIFT: |
4716 | case ASHIFTRT: | |
4717 | case LSHIFTRT: | |
f676971a | 4718 | case ROTATERT: |
9b66ebb1 PB |
4719 | case PLUS: |
4720 | case MINUS: | |
4721 | case COMPARE: | |
4722 | case NEG: | |
f676971a | 4723 | case NOT: |
9b66ebb1 | 4724 | return COSTS_N_INSNS (1); |
f676971a EC |
4725 | |
4726 | case MULT: | |
4727 | if (GET_CODE (XEXP (x, 1)) == CONST_INT) | |
4728 | { | |
4729 | int cycles = 0; | |
9b66ebb1 | 4730 | unsigned HOST_WIDE_INT i = INTVAL (XEXP (x, 1)); |
f676971a EC |
4731 | |
4732 | while (i) | |
4733 | { | |
4734 | i >>= 2; | |
4735 | cycles++; | |
4736 | } | |
4737 | return COSTS_N_INSNS (2) + cycles; | |
9b66ebb1 PB |
4738 | } |
4739 | return COSTS_N_INSNS (1) + 16; | |
f676971a EC |
4740 | |
4741 | case SET: | |
4742 | return (COSTS_N_INSNS (1) | |
4743 | + 4 * ((GET_CODE (SET_SRC (x)) == MEM) | |
9b66ebb1 | 4744 | + GET_CODE (SET_DEST (x)) == MEM)); |
f676971a EC |
4745 | |
4746 | case CONST_INT: | |
4747 | if (outer == SET) | |
4748 | { | |
4749 | if ((unsigned HOST_WIDE_INT) INTVAL (x) < 256) | |
4750 | return 0; | |
4751 | if (thumb_shiftable_const (INTVAL (x))) | |
4752 | return COSTS_N_INSNS (2); | |
4753 | return COSTS_N_INSNS (3); | |
4754 | } | |
9b66ebb1 | 4755 | else if ((outer == PLUS || outer == COMPARE) |
f676971a | 4756 | && INTVAL (x) < 256 && INTVAL (x) > -256) |
9b66ebb1 PB |
4757 | return 0; |
4758 | else if (outer == AND | |
4759 | && INTVAL (x) < 256 && INTVAL (x) >= -256) | |
4760 | return COSTS_N_INSNS (1); | |
f676971a EC |
4761 | else if (outer == ASHIFT || outer == ASHIFTRT |
4762 | || outer == LSHIFTRT) | |
4763 | return 0; | |
9b66ebb1 | 4764 | return COSTS_N_INSNS (2); |
f676971a EC |
4765 | |
4766 | case CONST: | |
4767 | case CONST_DOUBLE: | |
4768 | case LABEL_REF: | |
4769 | case SYMBOL_REF: | |
9b66ebb1 | 4770 | return COSTS_N_INSNS (3); |
f676971a | 4771 | |
9b66ebb1 PB |
4772 | case UDIV: |
4773 | case UMOD: | |
4774 | case DIV: | |
4775 | case MOD: | |
4776 | return 100; | |
d5b7b3ae | 4777 | |
9b66ebb1 PB |
4778 | case TRUNCATE: |
4779 | return 99; | |
d5b7b3ae | 4780 | |
9b66ebb1 PB |
4781 | case AND: |
4782 | case XOR: | |
f676971a | 4783 | case IOR: |
ff482c8d | 4784 | /* XXX guess. */ |
9b66ebb1 | 4785 | return 8; |
d5b7b3ae | 4786 | |
9b66ebb1 PB |
4787 | case MEM: |
4788 | /* XXX another guess. */ | |
4789 | /* Memory costs quite a lot for the first word, but subsequent words | |
4790 | load at the equivalent of a single insn each. */ | |
4791 | return (10 + 4 * ((GET_MODE_SIZE (mode) - 1) / UNITS_PER_WORD) | |
4792 | + ((GET_CODE (x) == SYMBOL_REF && CONSTANT_POOL_ADDRESS_P (x)) | |
4793 | ? 4 : 0)); | |
4794 | ||
4795 | case IF_THEN_ELSE: | |
ff482c8d | 4796 | /* XXX a guess. */ |
9b66ebb1 PB |
4797 | if (GET_CODE (XEXP (x, 1)) == PC || GET_CODE (XEXP (x, 2)) == PC) |
4798 | return 14; | |
4799 | return 2; | |
4800 | ||
4801 | case ZERO_EXTEND: | |
4802 | /* XXX still guessing. */ | |
4803 | switch (GET_MODE (XEXP (x, 0))) | |
4804 | { | |
4805 | case QImode: | |
4806 | return (1 + (mode == DImode ? 4 : 0) | |
4807 | + (GET_CODE (XEXP (x, 0)) == MEM ? 10 : 0)); | |
f676971a | 4808 | |
9b66ebb1 PB |
4809 | case HImode: |
4810 | return (4 + (mode == DImode ? 4 : 0) | |
4811 | + (GET_CODE (XEXP (x, 0)) == MEM ? 10 : 0)); | |
f676971a | 4812 | |
9b66ebb1 PB |
4813 | case SImode: |
4814 | return (1 + (GET_CODE (XEXP (x, 0)) == MEM ? 10 : 0)); | |
f676971a | 4815 | |
d5b7b3ae RE |
4816 | default: |
4817 | return 99; | |
d5b7b3ae | 4818 | } |
f676971a | 4819 | |
9b66ebb1 PB |
4820 | default: |
4821 | return 99; | |
d5b7b3ae | 4822 | } |
9b66ebb1 PB |
4823 | } |
4824 | ||
4825 | ||
4826 | /* Worker routine for arm_rtx_costs. */ | |
5b3e6663 | 4827 | /* ??? This needs updating for thumb2. */ |
9b66ebb1 PB |
4828 | static inline int |
4829 | arm_rtx_costs_1 (rtx x, enum rtx_code code, enum rtx_code outer) | |
4830 | { | |
4831 | enum machine_mode mode = GET_MODE (x); | |
4832 | enum rtx_code subcode; | |
4833 | int extra_cost; | |
4834 | ||
e2c671ba RE |
4835 | switch (code) |
4836 | { | |
4837 | case MEM: | |
4838 | /* Memory costs quite a lot for the first word, but subsequent words | |
4839 | load at the equivalent of a single insn each. */ | |
4840 | return (10 + 4 * ((GET_MODE_SIZE (mode) - 1) / UNITS_PER_WORD) | |
48f6efae NC |
4841 | + (GET_CODE (x) == SYMBOL_REF |
4842 | && CONSTANT_POOL_ADDRESS_P (x) ? 4 : 0)); | |
e2c671ba RE |
4843 | |
4844 | case DIV: | |
4845 | case MOD: | |
b9c53150 RS |
4846 | case UDIV: |
4847 | case UMOD: | |
4848 | return optimize_size ? COSTS_N_INSNS (2) : 100; | |
e2c671ba RE |
4849 | |
4850 | case ROTATE: | |
4851 | if (mode == SImode && GET_CODE (XEXP (x, 1)) == REG) | |
4852 | return 4; | |
4853 | /* Fall through */ | |
4854 | case ROTATERT: | |
4855 | if (mode != SImode) | |
4856 | return 8; | |
4857 | /* Fall through */ | |
4858 | case ASHIFT: case LSHIFTRT: case ASHIFTRT: | |
4859 | if (mode == DImode) | |
4860 | return (8 + (GET_CODE (XEXP (x, 1)) == CONST_INT ? 0 : 8) | |
f676971a | 4861 | + ((GET_CODE (XEXP (x, 0)) == REG |
e2c671ba RE |
4862 | || (GET_CODE (XEXP (x, 0)) == SUBREG |
4863 | && GET_CODE (SUBREG_REG (XEXP (x, 0))) == REG)) | |
4864 | ? 0 : 8)); | |
4865 | return (1 + ((GET_CODE (XEXP (x, 0)) == REG | |
4866 | || (GET_CODE (XEXP (x, 0)) == SUBREG | |
4867 | && GET_CODE (SUBREG_REG (XEXP (x, 0))) == REG)) | |
4868 | ? 0 : 4) | |
4869 | + ((GET_CODE (XEXP (x, 1)) == REG | |
4870 | || (GET_CODE (XEXP (x, 1)) == SUBREG | |
4871 | && GET_CODE (SUBREG_REG (XEXP (x, 1))) == REG) | |
4872 | || (GET_CODE (XEXP (x, 1)) == CONST_INT)) | |
4873 | ? 0 : 4)); | |
4874 | ||
4875 | case MINUS: | |
26b53f80 PB |
4876 | if (GET_CODE (XEXP (x, 1)) == MULT && mode == SImode && arm_arch_thumb2) |
4877 | { | |
4878 | extra_cost = rtx_cost (XEXP (x, 1), code); | |
4879 | if (!REG_OR_SUBREG_REG (XEXP (x, 0))) | |
4880 | extra_cost += 4 * ARM_NUM_REGS (mode); | |
4881 | return extra_cost; | |
4882 | } | |
4883 | ||
e2c671ba RE |
4884 | if (mode == DImode) |
4885 | return (4 + (REG_OR_SUBREG_REG (XEXP (x, 1)) ? 0 : 8) | |
4886 | + ((REG_OR_SUBREG_REG (XEXP (x, 0)) | |
4887 | || (GET_CODE (XEXP (x, 0)) == CONST_INT | |
4888 | && const_ok_for_arm (INTVAL (XEXP (x, 0))))) | |
4889 | ? 0 : 8)); | |
4890 | ||
4891 | if (GET_MODE_CLASS (mode) == MODE_FLOAT) | |
4892 | return (2 + ((REG_OR_SUBREG_REG (XEXP (x, 1)) | |
4893 | || (GET_CODE (XEXP (x, 1)) == CONST_DOUBLE | |
9b66ebb1 | 4894 | && arm_const_double_rtx (XEXP (x, 1)))) |
e2c671ba RE |
4895 | ? 0 : 8) |
4896 | + ((REG_OR_SUBREG_REG (XEXP (x, 0)) | |
4897 | || (GET_CODE (XEXP (x, 0)) == CONST_DOUBLE | |
9b66ebb1 | 4898 | && arm_const_double_rtx (XEXP (x, 0)))) |
e2c671ba RE |
4899 | ? 0 : 8)); |
4900 | ||
4901 | if (((GET_CODE (XEXP (x, 0)) == CONST_INT | |
4902 | && const_ok_for_arm (INTVAL (XEXP (x, 0))) | |
4903 | && REG_OR_SUBREG_REG (XEXP (x, 1)))) | |
4904 | || (((subcode = GET_CODE (XEXP (x, 1))) == ASHIFT | |
4905 | || subcode == ASHIFTRT || subcode == LSHIFTRT | |
4906 | || subcode == ROTATE || subcode == ROTATERT | |
4907 | || (subcode == MULT | |
4908 | && GET_CODE (XEXP (XEXP (x, 1), 1)) == CONST_INT | |
4909 | && ((INTVAL (XEXP (XEXP (x, 1), 1)) & | |
4910 | (INTVAL (XEXP (XEXP (x, 1), 1)) - 1)) == 0))) | |
4911 | && REG_OR_SUBREG_REG (XEXP (XEXP (x, 1), 0)) | |
4912 | && (REG_OR_SUBREG_REG (XEXP (XEXP (x, 1), 1)) | |
4913 | || GET_CODE (XEXP (XEXP (x, 1), 1)) == CONST_INT) | |
4914 | && REG_OR_SUBREG_REG (XEXP (x, 0)))) | |
4915 | return 1; | |
4916 | /* Fall through */ | |
4917 | ||
f676971a | 4918 | case PLUS: |
ff069900 PB |
4919 | if (arm_arch6 && mode == SImode |
4920 | && (GET_CODE (XEXP (x, 0)) == ZERO_EXTEND | |
4921 | || GET_CODE (XEXP (x, 0)) == SIGN_EXTEND)) | |
4922 | return 1 + (GET_CODE (XEXP (XEXP (x, 0), 0)) == MEM ? 10 : 0) | |
4923 | + (GET_CODE (XEXP (x, 1)) == MEM ? 10 : 0); | |
4924 | ||
26da58dd PB |
4925 | if (GET_CODE (XEXP (x, 0)) == MULT) |
4926 | { | |
4927 | extra_cost = rtx_cost (XEXP (x, 0), code); | |
4928 | if (!REG_OR_SUBREG_REG (XEXP (x, 1))) | |
4929 | extra_cost += 4 * ARM_NUM_REGS (mode); | |
4930 | return extra_cost; | |
4931 | } | |
4932 | ||
e2c671ba RE |
4933 | if (GET_MODE_CLASS (mode) == MODE_FLOAT) |
4934 | return (2 + (REG_OR_SUBREG_REG (XEXP (x, 0)) ? 0 : 8) | |
4935 | + ((REG_OR_SUBREG_REG (XEXP (x, 1)) | |
4936 | || (GET_CODE (XEXP (x, 1)) == CONST_DOUBLE | |
9b66ebb1 | 4937 | && arm_const_double_rtx (XEXP (x, 1)))) |
e2c671ba RE |
4938 | ? 0 : 8)); |
4939 | ||
4940 | /* Fall through */ | |
f676971a | 4941 | case AND: case XOR: case IOR: |
e2c671ba RE |
4942 | extra_cost = 0; |
4943 | ||
4944 | /* Normally the frame registers will be spilt into reg+const during | |
4945 | reload, so it is a bad idea to combine them with other instructions, | |
4946 | since then they might not be moved outside of loops. As a compromise | |
4947 | we allow integration with ops that have a constant as their second | |
4948 | operand. */ | |
4949 | if ((REG_OR_SUBREG_REG (XEXP (x, 0)) | |
4950 | && ARM_FRAME_RTX (REG_OR_SUBREG_RTX (XEXP (x, 0))) | |
4951 | && GET_CODE (XEXP (x, 1)) != CONST_INT) | |
4952 | || (REG_OR_SUBREG_REG (XEXP (x, 0)) | |
4953 | && ARM_FRAME_RTX (REG_OR_SUBREG_RTX (XEXP (x, 0))))) | |
4954 | extra_cost = 4; | |
4955 | ||
4956 | if (mode == DImode) | |
4957 | return (4 + extra_cost + (REG_OR_SUBREG_REG (XEXP (x, 0)) ? 0 : 8) | |
4958 | + ((REG_OR_SUBREG_REG (XEXP (x, 1)) | |
4959 | || (GET_CODE (XEXP (x, 1)) == CONST_INT | |
74bbc178 | 4960 | && const_ok_for_op (INTVAL (XEXP (x, 1)), code))) |
e2c671ba RE |
4961 | ? 0 : 8)); |
4962 | ||
4963 | if (REG_OR_SUBREG_REG (XEXP (x, 0))) | |
4964 | return (1 + (GET_CODE (XEXP (x, 1)) == CONST_INT ? 0 : extra_cost) | |
4965 | + ((REG_OR_SUBREG_REG (XEXP (x, 1)) | |
4966 | || (GET_CODE (XEXP (x, 1)) == CONST_INT | |
74bbc178 | 4967 | && const_ok_for_op (INTVAL (XEXP (x, 1)), code))) |
e2c671ba RE |
4968 | ? 0 : 4)); |
4969 | ||
4970 | else if (REG_OR_SUBREG_REG (XEXP (x, 1))) | |
4971 | return (1 + extra_cost | |
4972 | + ((((subcode = GET_CODE (XEXP (x, 0))) == ASHIFT | |
4973 | || subcode == LSHIFTRT || subcode == ASHIFTRT | |
4974 | || subcode == ROTATE || subcode == ROTATERT | |
4975 | || (subcode == MULT | |
4976 | && GET_CODE (XEXP (XEXP (x, 0), 1)) == CONST_INT | |
4977 | && ((INTVAL (XEXP (XEXP (x, 0), 1)) & | |
ad076f4e | 4978 | (INTVAL (XEXP (XEXP (x, 0), 1)) - 1)) == 0))) |
e2c671ba RE |
4979 | && (REG_OR_SUBREG_REG (XEXP (XEXP (x, 0), 0))) |
4980 | && ((REG_OR_SUBREG_REG (XEXP (XEXP (x, 0), 1))) | |
ad076f4e | 4981 | || GET_CODE (XEXP (XEXP (x, 0), 1)) == CONST_INT)) |
e2c671ba RE |
4982 | ? 0 : 4)); |
4983 | ||
4984 | return 8; | |
4985 | ||
4986 | case MULT: | |
9b66ebb1 | 4987 | /* This should have been handled by the CPU specific routines. */ |
e6d29d15 | 4988 | gcc_unreachable (); |
e2c671ba | 4989 | |
56636818 | 4990 | case TRUNCATE: |
9b66ebb1 | 4991 | if (arm_arch3m && mode == SImode |
56636818 JL |
4992 | && GET_CODE (XEXP (x, 0)) == LSHIFTRT |
4993 | && GET_CODE (XEXP (XEXP (x, 0), 0)) == MULT | |
4994 | && (GET_CODE (XEXP (XEXP (XEXP (x, 0), 0), 0)) | |
4995 | == GET_CODE (XEXP (XEXP (XEXP (x, 0), 0), 1))) | |
4996 | && (GET_CODE (XEXP (XEXP (XEXP (x, 0), 0), 0)) == ZERO_EXTEND | |
4997 | || GET_CODE (XEXP (XEXP (XEXP (x, 0), 0), 0)) == SIGN_EXTEND)) | |
4998 | return 8; | |
4999 | return 99; | |
5000 | ||
e2c671ba RE |
5001 | case NEG: |
5002 | if (GET_MODE_CLASS (mode) == MODE_FLOAT) | |
5003 | return 4 + (REG_OR_SUBREG_REG (XEXP (x, 0)) ? 0 : 6); | |
5004 | /* Fall through */ | |
5005 | case NOT: | |
5006 | if (mode == DImode) | |
5007 | return 4 + (REG_OR_SUBREG_REG (XEXP (x, 0)) ? 0 : 4); | |
5008 | ||
5009 | return 1 + (REG_OR_SUBREG_REG (XEXP (x, 0)) ? 0 : 4); | |
5010 | ||
5011 | case IF_THEN_ELSE: | |
5012 | if (GET_CODE (XEXP (x, 1)) == PC || GET_CODE (XEXP (x, 2)) == PC) | |
5013 | return 14; | |
5014 | return 2; | |
5015 | ||
5016 | case COMPARE: | |
5017 | return 1; | |
5018 | ||
5019 | case ABS: | |
5020 | return 4 + (mode == DImode ? 4 : 0); | |
5021 | ||
5022 | case SIGN_EXTEND: | |
ff069900 PB |
5023 | if (arm_arch_thumb2 && mode == SImode) |
5024 | return 1 + (GET_CODE (XEXP (x, 0)) == MEM ? 10 : 0); | |
5025 | ||
e2c671ba RE |
5026 | if (GET_MODE (XEXP (x, 0)) == QImode) |
5027 | return (4 + (mode == DImode ? 4 : 0) | |
5028 | + (GET_CODE (XEXP (x, 0)) == MEM ? 10 : 0)); | |
5029 | /* Fall through */ | |
5030 | case ZERO_EXTEND: | |
ff069900 PB |
5031 | if (arm_arch6 && mode == SImode) |
5032 | return 1 + (GET_CODE (XEXP (x, 0)) == MEM ? 10 : 0); | |
5033 | ||
e2c671ba RE |
5034 | switch (GET_MODE (XEXP (x, 0))) |
5035 | { | |
5036 | case QImode: | |
5037 | return (1 + (mode == DImode ? 4 : 0) | |
5038 | + (GET_CODE (XEXP (x, 0)) == MEM ? 10 : 0)); | |
5039 | ||
5040 | case HImode: | |
5041 | return (4 + (mode == DImode ? 4 : 0) | |
5042 | + (GET_CODE (XEXP (x, 0)) == MEM ? 10 : 0)); | |
5043 | ||
5044 | case SImode: | |
5045 | return (1 + (GET_CODE (XEXP (x, 0)) == MEM ? 10 : 0)); | |
ad076f4e | 5046 | |
5a9335ef NC |
5047 | case V8QImode: |
5048 | case V4HImode: | |
5049 | case V2SImode: | |
5050 | case V4QImode: | |
5051 | case V2HImode: | |
5052 | return 1; | |
5053 | ||
ad076f4e | 5054 | default: |
e6d29d15 | 5055 | gcc_unreachable (); |
e2c671ba | 5056 | } |
e6d29d15 | 5057 | gcc_unreachable (); |
e2c671ba | 5058 | |
f676971a EC |
5059 | case CONST_INT: |
5060 | if (const_ok_for_arm (INTVAL (x))) | |
5061 | return outer == SET ? 2 : -1; | |
5062 | else if (outer == AND | |
5063 | && const_ok_for_arm (~INTVAL (x))) | |
5064 | return -1; | |
5065 | else if ((outer == COMPARE | |
5066 | || outer == PLUS || outer == MINUS) | |
5067 | && const_ok_for_arm (-INTVAL (x))) | |
5068 | return -1; | |
5069 | else | |
d5b7b3ae | 5070 | return 5; |
f676971a EC |
5071 | |
5072 | case CONST: | |
5073 | case LABEL_REF: | |
5074 | case SYMBOL_REF: | |
d5b7b3ae | 5075 | return 6; |
f676971a EC |
5076 | |
5077 | case CONST_DOUBLE: | |
f1adb0a9 | 5078 | if (arm_const_double_rtx (x) || vfp3_const_double_rtx (x)) |
f676971a EC |
5079 | return outer == SET ? 2 : -1; |
5080 | else if ((outer == COMPARE || outer == PLUS) | |
5081 | && neg_const_double_rtx_ok_for_fpa (x)) | |
5082 | return -1; | |
d5b7b3ae | 5083 | return 7; |
f676971a | 5084 | |
e2c671ba RE |
5085 | default: |
5086 | return 99; | |
5087 | } | |
5088 | } | |
32de079a | 5089 | |
21b5653c RE |
5090 | /* RTX costs when optimizing for size. */ |
5091 | static bool | |
5092 | arm_size_rtx_costs (rtx x, int code, int outer_code, int *total) | |
5093 | { | |
5094 | enum machine_mode mode = GET_MODE (x); | |
5095 | ||
09754904 | 5096 | if (TARGET_THUMB1) |
21b5653c RE |
5097 | { |
5098 | /* XXX TBD. For now, use the standard costs. */ | |
5b3e6663 | 5099 | *total = thumb1_rtx_costs (x, code, outer_code); |
21b5653c RE |
5100 | return true; |
5101 | } | |
5102 | ||
09754904 | 5103 | /* FIXME: This makes no attempt to prefer narrow Thumb-2 instructions. */ |
21b5653c RE |
5104 | switch (code) |
5105 | { | |
5106 | case MEM: | |
f676971a | 5107 | /* A memory access costs 1 insn if the mode is small, or the address is |
21b5653c RE |
5108 | a single register, otherwise it costs one insn per word. */ |
5109 | if (REG_P (XEXP (x, 0))) | |
5110 | *total = COSTS_N_INSNS (1); | |
5111 | else | |
5112 | *total = COSTS_N_INSNS (ARM_NUM_REGS (mode)); | |
5113 | return true; | |
5114 | ||
5115 | case DIV: | |
5116 | case MOD: | |
5117 | case UDIV: | |
5118 | case UMOD: | |
5119 | /* Needs a libcall, so it costs about this. */ | |
5120 | *total = COSTS_N_INSNS (2); | |
5121 | return false; | |
5122 | ||
5123 | case ROTATE: | |
5124 | if (mode == SImode && GET_CODE (XEXP (x, 1)) == REG) | |
5125 | { | |
5126 | *total = COSTS_N_INSNS (2) + rtx_cost (XEXP (x, 0), code); | |
5127 | return true; | |
5128 | } | |
5129 | /* Fall through */ | |
5130 | case ROTATERT: | |
5131 | case ASHIFT: | |
5132 | case LSHIFTRT: | |
5133 | case ASHIFTRT: | |
5134 | if (mode == DImode && GET_CODE (XEXP (x, 1)) == CONST_INT) | |
5135 | { | |
5136 | *total = COSTS_N_INSNS (3) + rtx_cost (XEXP (x, 0), code); | |
5137 | return true; | |
5138 | } | |
5139 | else if (mode == SImode) | |
5140 | { | |
5141 | *total = COSTS_N_INSNS (1) + rtx_cost (XEXP (x, 0), code); | |
5142 | /* Slightly disparage register shifts, but not by much. */ | |
5143 | if (GET_CODE (XEXP (x, 1)) != CONST_INT) | |
5144 | *total += 1 + rtx_cost (XEXP (x, 1), code); | |
5145 | return true; | |
5146 | } | |
5147 | ||
5148 | /* Needs a libcall. */ | |
5149 | *total = COSTS_N_INSNS (2); | |
5150 | return false; | |
5151 | ||
5152 | case MINUS: | |
5153 | if (TARGET_HARD_FLOAT && GET_MODE_CLASS (mode) == MODE_FLOAT) | |
5154 | { | |
5155 | *total = COSTS_N_INSNS (1); | |
5156 | return false; | |
5157 | } | |
5158 | ||
5159 | if (mode == SImode) | |
5160 | { | |
5161 | enum rtx_code subcode0 = GET_CODE (XEXP (x, 0)); | |
5162 | enum rtx_code subcode1 = GET_CODE (XEXP (x, 1)); | |
5163 | ||
5164 | if (subcode0 == ROTATE || subcode0 == ROTATERT || subcode0 == ASHIFT | |
5165 | || subcode0 == LSHIFTRT || subcode0 == ASHIFTRT | |
5166 | || subcode1 == ROTATE || subcode1 == ROTATERT | |
5167 | || subcode1 == ASHIFT || subcode1 == LSHIFTRT | |
5168 | || subcode1 == ASHIFTRT) | |
5169 | { | |
5170 | /* It's just the cost of the two operands. */ | |
5171 | *total = 0; | |
5172 | return false; | |
5173 | } | |
5174 | ||
5175 | *total = COSTS_N_INSNS (1); | |
5176 | return false; | |
5177 | } | |
5178 | ||
5179 | *total = COSTS_N_INSNS (ARM_NUM_REGS (mode)); | |
5180 | return false; | |
5181 | ||
f676971a | 5182 | case PLUS: |
21b5653c RE |
5183 | if (TARGET_HARD_FLOAT && GET_MODE_CLASS (mode) == MODE_FLOAT) |
5184 | { | |
5185 | *total = COSTS_N_INSNS (1); | |
5186 | return false; | |
5187 | } | |
5188 | ||
5189 | /* Fall through */ | |
5190 | case AND: case XOR: case IOR: | |
5191 | if (mode == SImode) | |
5192 | { | |
5193 | enum rtx_code subcode = GET_CODE (XEXP (x, 0)); | |
5194 | ||
5195 | if (subcode == ROTATE || subcode == ROTATERT || subcode == ASHIFT | |
5196 | || subcode == LSHIFTRT || subcode == ASHIFTRT | |
5197 | || (code == AND && subcode == NOT)) | |
5198 | { | |
5199 | /* It's just the cost of the two operands. */ | |
5200 | *total = 0; | |
5201 | return false; | |
5202 | } | |
5203 | } | |
5204 | ||
5205 | *total = COSTS_N_INSNS (ARM_NUM_REGS (mode)); | |
5206 | return false; | |
5207 | ||
5208 | case MULT: | |
5209 | *total = COSTS_N_INSNS (ARM_NUM_REGS (mode)); | |
5210 | return false; | |
5211 | ||
5212 | case NEG: | |
5213 | if (TARGET_HARD_FLOAT && GET_MODE_CLASS (mode) == MODE_FLOAT) | |
5214 | *total = COSTS_N_INSNS (1); | |
5215 | /* Fall through */ | |
5216 | case NOT: | |
5217 | *total = COSTS_N_INSNS (ARM_NUM_REGS (mode)); | |
5218 | ||
5219 | return false; | |
5220 | ||
5221 | case IF_THEN_ELSE: | |
5222 | *total = 0; | |
5223 | return false; | |
5224 | ||
5225 | case COMPARE: | |
5226 | if (cc_register (XEXP (x, 0), VOIDmode)) | |
5227 | * total = 0; | |
5228 | else | |
5229 | *total = COSTS_N_INSNS (1); | |
5230 | return false; | |
5231 | ||
5232 | case ABS: | |
5233 | if (TARGET_HARD_FLOAT && GET_MODE_CLASS (mode) == MODE_FLOAT) | |
5234 | *total = COSTS_N_INSNS (1); | |
5235 | else | |
5236 | *total = COSTS_N_INSNS (1 + ARM_NUM_REGS (mode)); | |
5237 | return false; | |
5238 | ||
5239 | case SIGN_EXTEND: | |
5240 | *total = 0; | |
5241 | if (GET_MODE_SIZE (GET_MODE (XEXP (x, 0))) < 4) | |
5242 | { | |
5243 | if (!(arm_arch4 && MEM_P (XEXP (x, 0)))) | |
5244 | *total += COSTS_N_INSNS (arm_arch6 ? 1 : 2); | |
5245 | } | |
5246 | if (mode == DImode) | |
5247 | *total += COSTS_N_INSNS (1); | |
5248 | return false; | |
5249 | ||
5250 | case ZERO_EXTEND: | |
5251 | *total = 0; | |
5252 | if (!(arm_arch4 && MEM_P (XEXP (x, 0)))) | |
5253 | { | |
5254 | switch (GET_MODE (XEXP (x, 0))) | |
5255 | { | |
5256 | case QImode: | |
5257 | *total += COSTS_N_INSNS (1); | |
5258 | break; | |
5259 | ||
5260 | case HImode: | |
5261 | *total += COSTS_N_INSNS (arm_arch6 ? 1 : 2); | |
f676971a | 5262 | |
21b5653c RE |
5263 | case SImode: |
5264 | break; | |
5265 | ||
5266 | default: | |
5267 | *total += COSTS_N_INSNS (2); | |
5268 | } | |
5269 | } | |
5270 | ||
5271 | if (mode == DImode) | |
5272 | *total += COSTS_N_INSNS (1); | |
5273 | ||
5274 | return false; | |
5275 | ||
f676971a EC |
5276 | case CONST_INT: |
5277 | if (const_ok_for_arm (INTVAL (x))) | |
21b5653c RE |
5278 | *total = COSTS_N_INSNS (outer_code == SET ? 1 : 0); |
5279 | else if (const_ok_for_arm (~INTVAL (x))) | |
5280 | *total = COSTS_N_INSNS (outer_code == AND ? 0 : 1); | |
5281 | else if (const_ok_for_arm (-INTVAL (x))) | |
5282 | { | |
5283 | if (outer_code == COMPARE || outer_code == PLUS | |
5284 | || outer_code == MINUS) | |
5285 | *total = 0; | |
5286 | else | |
5287 | *total = COSTS_N_INSNS (1); | |
5288 | } | |
5289 | else | |
5290 | *total = COSTS_N_INSNS (2); | |
5291 | return true; | |
f676971a EC |
5292 | |
5293 | case CONST: | |
5294 | case LABEL_REF: | |
5295 | case SYMBOL_REF: | |
21b5653c RE |
5296 | *total = COSTS_N_INSNS (2); |
5297 | return true; | |
f676971a | 5298 | |
21b5653c RE |
5299 | case CONST_DOUBLE: |
5300 | *total = COSTS_N_INSNS (4); | |
5301 | return true; | |
5302 | ||
5303 | default: | |
5304 | if (mode != VOIDmode) | |
5305 | *total = COSTS_N_INSNS (ARM_NUM_REGS (mode)); | |
5306 | else | |
5307 | *total = COSTS_N_INSNS (4); /* How knows? */ | |
5308 | return false; | |
5309 | } | |
5310 | } | |
5311 | ||
5b3e6663 PB |
5312 | /* RTX costs for cores with a slow MUL implementation. Thumb-2 is not |
5313 | supported on any "slowmul" cores, so it can be ignored. */ | |
9b66ebb1 | 5314 | |
3c50106f | 5315 | static bool |
9b66ebb1 | 5316 | arm_slowmul_rtx_costs (rtx x, int code, int outer_code, int *total) |
3c50106f | 5317 | { |
9b66ebb1 PB |
5318 | enum machine_mode mode = GET_MODE (x); |
5319 | ||
5320 | if (TARGET_THUMB) | |
5321 | { | |
5b3e6663 | 5322 | *total = thumb1_rtx_costs (x, code, outer_code); |
9b66ebb1 PB |
5323 | return true; |
5324 | } | |
f676971a | 5325 | |
9b66ebb1 PB |
5326 | switch (code) |
5327 | { | |
5328 | case MULT: | |
5329 | if (GET_MODE_CLASS (mode) == MODE_FLOAT | |
5330 | || mode == DImode) | |
5331 | { | |
5332 | *total = 30; | |
5333 | return true; | |
5334 | } | |
5335 | ||
5336 | if (GET_CODE (XEXP (x, 1)) == CONST_INT) | |
5337 | { | |
5338 | unsigned HOST_WIDE_INT i = (INTVAL (XEXP (x, 1)) | |
5339 | & (unsigned HOST_WIDE_INT) 0xffffffff); | |
5340 | int cost, const_ok = const_ok_for_arm (i); | |
5341 | int j, booth_unit_size; | |
5342 | ||
f676971a | 5343 | /* Tune as appropriate. */ |
9b66ebb1 PB |
5344 | cost = const_ok ? 4 : 8; |
5345 | booth_unit_size = 2; | |
5346 | for (j = 0; i && j < 32; j += booth_unit_size) | |
5347 | { | |
5348 | i >>= booth_unit_size; | |
5349 | cost += 2; | |
5350 | } | |
5351 | ||
5352 | *total = cost; | |
5353 | return true; | |
5354 | } | |
5355 | ||
5356 | *total = 30 + (REG_OR_SUBREG_REG (XEXP (x, 0)) ? 0 : 4) | |
5357 | + (REG_OR_SUBREG_REG (XEXP (x, 1)) ? 0 : 4); | |
5358 | return true; | |
f676971a | 5359 | |
9b66ebb1 PB |
5360 | default: |
5361 | *total = arm_rtx_costs_1 (x, code, outer_code); | |
5362 | return true; | |
5363 | } | |
3c50106f RH |
5364 | } |
5365 | ||
9b66ebb1 PB |
5366 | |
5367 | /* RTX cost for cores with a fast multiply unit (M variants). */ | |
5368 | ||
5369 | static bool | |
5370 | arm_fastmul_rtx_costs (rtx x, int code, int outer_code, int *total) | |
5371 | { | |
5372 | enum machine_mode mode = GET_MODE (x); | |
5373 | ||
5b3e6663 | 5374 | if (TARGET_THUMB1) |
9b66ebb1 | 5375 | { |
5b3e6663 | 5376 | *total = thumb1_rtx_costs (x, code, outer_code); |
9b66ebb1 PB |
5377 | return true; |
5378 | } | |
f676971a | 5379 | |
5b3e6663 | 5380 | /* ??? should thumb2 use different costs? */ |
9b66ebb1 PB |
5381 | switch (code) |
5382 | { | |
5383 | case MULT: | |
5384 | /* There is no point basing this on the tuning, since it is always the | |
5385 | fast variant if it exists at all. */ | |
5386 | if (mode == DImode | |
5387 | && (GET_CODE (XEXP (x, 0)) == GET_CODE (XEXP (x, 1))) | |
5388 | && (GET_CODE (XEXP (x, 0)) == ZERO_EXTEND | |
5389 | || GET_CODE (XEXP (x, 0)) == SIGN_EXTEND)) | |
5390 | { | |
5391 | *total = 8; | |
5392 | return true; | |
5393 | } | |
f676971a | 5394 | |
9b66ebb1 PB |
5395 | |
5396 | if (GET_MODE_CLASS (mode) == MODE_FLOAT | |
5397 | || mode == DImode) | |
5398 | { | |
5399 | *total = 30; | |
5400 | return true; | |
5401 | } | |
5402 | ||
5403 | if (GET_CODE (XEXP (x, 1)) == CONST_INT) | |
5404 | { | |
5405 | unsigned HOST_WIDE_INT i = (INTVAL (XEXP (x, 1)) | |
5406 | & (unsigned HOST_WIDE_INT) 0xffffffff); | |
5407 | int cost, const_ok = const_ok_for_arm (i); | |
5408 | int j, booth_unit_size; | |
5409 | ||
f676971a | 5410 | /* Tune as appropriate. */ |
9b66ebb1 PB |
5411 | cost = const_ok ? 4 : 8; |
5412 | booth_unit_size = 8; | |
5413 | for (j = 0; i && j < 32; j += booth_unit_size) | |
5414 | { | |
5415 | i >>= booth_unit_size; | |
5416 | cost += 2; | |
5417 | } | |
5418 | ||
5419 | *total = cost; | |
5420 | return true; | |
5421 | } | |
5422 | ||
5423 | *total = 8 + (REG_OR_SUBREG_REG (XEXP (x, 0)) ? 0 : 4) | |
5424 | + (REG_OR_SUBREG_REG (XEXP (x, 1)) ? 0 : 4); | |
5425 | return true; | |
f676971a | 5426 | |
9b66ebb1 PB |
5427 | default: |
5428 | *total = arm_rtx_costs_1 (x, code, outer_code); | |
5429 | return true; | |
5430 | } | |
5431 | } | |
5432 | ||
5433 | ||
5b3e6663 PB |
5434 | /* RTX cost for XScale CPUs. Thumb-2 is not supported on any xscale cores, |
5435 | so it can be ignored. */ | |
9b66ebb1 PB |
5436 | |
5437 | static bool | |
5438 | arm_xscale_rtx_costs (rtx x, int code, int outer_code, int *total) | |
5439 | { | |
5440 | enum machine_mode mode = GET_MODE (x); | |
5441 | ||
5442 | if (TARGET_THUMB) | |
5443 | { | |
5b3e6663 | 5444 | *total = thumb1_rtx_costs (x, code, outer_code); |
9b66ebb1 PB |
5445 | return true; |
5446 | } | |
f676971a | 5447 | |
9b66ebb1 PB |
5448 | switch (code) |
5449 | { | |
5450 | case MULT: | |
5451 | /* There is no point basing this on the tuning, since it is always the | |
5452 | fast variant if it exists at all. */ | |
5453 | if (mode == DImode | |
5454 | && (GET_CODE (XEXP (x, 0)) == GET_CODE (XEXP (x, 1))) | |
5455 | && (GET_CODE (XEXP (x, 0)) == ZERO_EXTEND | |
5456 | || GET_CODE (XEXP (x, 0)) == SIGN_EXTEND)) | |
5457 | { | |
5458 | *total = 8; | |
5459 | return true; | |
5460 | } | |
f676971a | 5461 | |
9b66ebb1 PB |
5462 | |
5463 | if (GET_MODE_CLASS (mode) == MODE_FLOAT | |
5464 | || mode == DImode) | |
5465 | { | |
5466 | *total = 30; | |
5467 | return true; | |
5468 | } | |
5469 | ||
5470 | if (GET_CODE (XEXP (x, 1)) == CONST_INT) | |
5471 | { | |
5472 | unsigned HOST_WIDE_INT i = (INTVAL (XEXP (x, 1)) | |
5473 | & (unsigned HOST_WIDE_INT) 0xffffffff); | |
5474 | int cost, const_ok = const_ok_for_arm (i); | |
5475 | unsigned HOST_WIDE_INT masked_const; | |
5476 | ||
5477 | /* The cost will be related to two insns. | |
ff482c8d | 5478 | First a load of the constant (MOV or LDR), then a multiply. */ |
9b66ebb1 PB |
5479 | cost = 2; |
5480 | if (! const_ok) | |
5481 | cost += 1; /* LDR is probably more expensive because | |
ff482c8d | 5482 | of longer result latency. */ |
9b66ebb1 PB |
5483 | masked_const = i & 0xffff8000; |
5484 | if (masked_const != 0 && masked_const != 0xffff8000) | |
5485 | { | |
5486 | masked_const = i & 0xf8000000; | |
5487 | if (masked_const == 0 || masked_const == 0xf8000000) | |
5488 | cost += 1; | |
5489 | else | |
5490 | cost += 2; | |
5491 | } | |
5492 | *total = cost; | |
5493 | return true; | |
5494 | } | |
5495 | ||
5496 | *total = 8 + (REG_OR_SUBREG_REG (XEXP (x, 0)) ? 0 : 4) | |
5497 | + (REG_OR_SUBREG_REG (XEXP (x, 1)) ? 0 : 4); | |
5498 | return true; | |
f676971a | 5499 | |
06d5588c ILT |
5500 | case COMPARE: |
5501 | /* A COMPARE of a MULT is slow on XScale; the muls instruction | |
5502 | will stall until the multiplication is complete. */ | |
5503 | if (GET_CODE (XEXP (x, 0)) == MULT) | |
5504 | *total = 4 + rtx_cost (XEXP (x, 0), code); | |
5505 | else | |
5506 | *total = arm_rtx_costs_1 (x, code, outer_code); | |
5507 | return true; | |
5508 | ||
9b66ebb1 PB |
5509 | default: |
5510 | *total = arm_rtx_costs_1 (x, code, outer_code); | |
5511 | return true; | |
5512 | } | |
5513 | } | |
5514 | ||
5515 | ||
5516 | /* RTX costs for 9e (and later) cores. */ | |
5517 | ||
5518 | static bool | |
5519 | arm_9e_rtx_costs (rtx x, int code, int outer_code, int *total) | |
5520 | { | |
5521 | enum machine_mode mode = GET_MODE (x); | |
5522 | int nonreg_cost; | |
5523 | int cost; | |
f676971a | 5524 | |
5b3e6663 | 5525 | if (TARGET_THUMB1) |
9b66ebb1 PB |
5526 | { |
5527 | switch (code) | |
5528 | { | |
5529 | case MULT: | |
5530 | *total = COSTS_N_INSNS (3); | |
5531 | return true; | |
f676971a | 5532 | |
9b66ebb1 | 5533 | default: |
5b3e6663 | 5534 | *total = thumb1_rtx_costs (x, code, outer_code); |
9b66ebb1 PB |
5535 | return true; |
5536 | } | |
5537 | } | |
f676971a | 5538 | |
9b66ebb1 PB |
5539 | switch (code) |
5540 | { | |
5541 | case MULT: | |
5542 | /* There is no point basing this on the tuning, since it is always the | |
5543 | fast variant if it exists at all. */ | |
5544 | if (mode == DImode | |
5545 | && (GET_CODE (XEXP (x, 0)) == GET_CODE (XEXP (x, 1))) | |
5546 | && (GET_CODE (XEXP (x, 0)) == ZERO_EXTEND | |
5547 | || GET_CODE (XEXP (x, 0)) == SIGN_EXTEND)) | |
5548 | { | |
5549 | *total = 3; | |
5550 | return true; | |
5551 | } | |
f676971a | 5552 | |
9b66ebb1 PB |
5553 | |
5554 | if (GET_MODE_CLASS (mode) == MODE_FLOAT) | |
5555 | { | |
5556 | *total = 30; | |
5557 | return true; | |
5558 | } | |
5559 | if (mode == DImode) | |
5560 | { | |
5561 | cost = 7; | |
5562 | nonreg_cost = 8; | |
5563 | } | |
5564 | else | |
5565 | { | |
5566 | cost = 2; | |
5567 | nonreg_cost = 4; | |
5568 | } | |
5569 | ||
5570 | ||
5571 | *total = cost + (REG_OR_SUBREG_REG (XEXP (x, 0)) ? 0 : nonreg_cost) | |
5572 | + (REG_OR_SUBREG_REG (XEXP (x, 1)) ? 0 : nonreg_cost); | |
5573 | return true; | |
f676971a | 5574 | |
9b66ebb1 PB |
5575 | default: |
5576 | *total = arm_rtx_costs_1 (x, code, outer_code); | |
5577 | return true; | |
5578 | } | |
5579 | } | |
dcefdf67 RH |
5580 | /* All address computations that can be done are free, but rtx cost returns |
5581 | the same for practically all of them. So we weight the different types | |
5582 | of address here in the order (most pref first): | |
d6b4baa4 | 5583 | PRE/POST_INC/DEC, SHIFT or NON-INT sum, INT sum, REG, MEM or LABEL. */ |
d2b6eb76 ZW |
5584 | static inline int |
5585 | arm_arm_address_cost (rtx x) | |
5586 | { | |
5587 | enum rtx_code c = GET_CODE (x); | |
5588 | ||
5589 | if (c == PRE_INC || c == PRE_DEC || c == POST_INC || c == POST_DEC) | |
5590 | return 0; | |
5591 | if (c == MEM || c == LABEL_REF || c == SYMBOL_REF) | |
5592 | return 10; | |
5593 | ||
5594 | if (c == PLUS || c == MINUS) | |
5595 | { | |
d2b6eb76 ZW |
5596 | if (GET_CODE (XEXP (x, 0)) == CONST_INT) |
5597 | return 2; | |
5598 | ||
ec8e098d | 5599 | if (ARITHMETIC_P (XEXP (x, 0)) || ARITHMETIC_P (XEXP (x, 1))) |
d2b6eb76 ZW |
5600 | return 3; |
5601 | ||
5602 | return 4; | |
5603 | } | |
5604 | ||
5605 | return 6; | |
5606 | } | |
5607 | ||
5608 | static inline int | |
5609 | arm_thumb_address_cost (rtx x) | |
5610 | { | |
5611 | enum rtx_code c = GET_CODE (x); | |
5612 | ||
5613 | if (c == REG) | |
5614 | return 1; | |
5615 | if (c == PLUS | |
5616 | && GET_CODE (XEXP (x, 0)) == REG | |
5617 | && GET_CODE (XEXP (x, 1)) == CONST_INT) | |
5618 | return 1; | |
5619 | ||
5620 | return 2; | |
5621 | } | |
5622 | ||
dcefdf67 | 5623 | static int |
e32bac5b | 5624 | arm_address_cost (rtx x) |
dcefdf67 | 5625 | { |
5b3e6663 | 5626 | return TARGET_32BIT ? arm_arm_address_cost (x) : arm_thumb_address_cost (x); |
dcefdf67 RH |
5627 | } |
5628 | ||
c237e94a | 5629 | static int |
e32bac5b | 5630 | arm_adjust_cost (rtx insn, rtx link, rtx dep, int cost) |
32de079a RE |
5631 | { |
5632 | rtx i_pat, d_pat; | |
5633 | ||
d19fb8e3 NC |
5634 | /* Some true dependencies can have a higher cost depending |
5635 | on precisely how certain input operands are used. */ | |
4b3c2e48 | 5636 | if (arm_tune_xscale |
d19fb8e3 | 5637 | && REG_NOTE_KIND (link) == 0 |
eda833e3 BE |
5638 | && recog_memoized (insn) >= 0 |
5639 | && recog_memoized (dep) >= 0) | |
d19fb8e3 NC |
5640 | { |
5641 | int shift_opnum = get_attr_shift (insn); | |
5642 | enum attr_type attr_type = get_attr_type (dep); | |
5643 | ||
5644 | /* If nonzero, SHIFT_OPNUM contains the operand number of a shifted | |
5645 | operand for INSN. If we have a shifted input operand and the | |
5646 | instruction we depend on is another ALU instruction, then we may | |
5647 | have to account for an additional stall. */ | |
9b66ebb1 PB |
5648 | if (shift_opnum != 0 |
5649 | && (attr_type == TYPE_ALU_SHIFT || attr_type == TYPE_ALU_SHIFT_REG)) | |
d19fb8e3 NC |
5650 | { |
5651 | rtx shifted_operand; | |
5652 | int opno; | |
f676971a | 5653 | |
d19fb8e3 NC |
5654 | /* Get the shifted operand. */ |
5655 | extract_insn (insn); | |
5656 | shifted_operand = recog_data.operand[shift_opnum]; | |
5657 | ||
5658 | /* Iterate over all the operands in DEP. If we write an operand | |
5659 | that overlaps with SHIFTED_OPERAND, then we have increase the | |
5660 | cost of this dependency. */ | |
5661 | extract_insn (dep); | |
5662 | preprocess_constraints (); | |
5663 | for (opno = 0; opno < recog_data.n_operands; opno++) | |
5664 | { | |
5665 | /* We can ignore strict inputs. */ | |
5666 | if (recog_data.operand_type[opno] == OP_IN) | |
5667 | continue; | |
5668 | ||
5669 | if (reg_overlap_mentioned_p (recog_data.operand[opno], | |
5670 | shifted_operand)) | |
5671 | return 2; | |
5672 | } | |
5673 | } | |
5674 | } | |
5675 | ||
6354dc9b | 5676 | /* XXX This is not strictly true for the FPA. */ |
d5b7b3ae RE |
5677 | if (REG_NOTE_KIND (link) == REG_DEP_ANTI |
5678 | || REG_NOTE_KIND (link) == REG_DEP_OUTPUT) | |
b36ba79f RE |
5679 | return 0; |
5680 | ||
d5b7b3ae RE |
5681 | /* Call insns don't incur a stall, even if they follow a load. */ |
5682 | if (REG_NOTE_KIND (link) == 0 | |
5683 | && GET_CODE (insn) == CALL_INSN) | |
5684 | return 1; | |
5685 | ||
32de079a RE |
5686 | if ((i_pat = single_set (insn)) != NULL |
5687 | && GET_CODE (SET_SRC (i_pat)) == MEM | |
5688 | && (d_pat = single_set (dep)) != NULL | |
5689 | && GET_CODE (SET_DEST (d_pat)) == MEM) | |
5690 | { | |
48f6efae | 5691 | rtx src_mem = XEXP (SET_SRC (i_pat), 0); |
32de079a RE |
5692 | /* This is a load after a store, there is no conflict if the load reads |
5693 | from a cached area. Assume that loads from the stack, and from the | |
f676971a | 5694 | constant pool are cached, and that others will miss. This is a |
6354dc9b | 5695 | hack. */ |
f676971a | 5696 | |
48f6efae NC |
5697 | if ((GET_CODE (src_mem) == SYMBOL_REF && CONSTANT_POOL_ADDRESS_P (src_mem)) |
5698 | || reg_mentioned_p (stack_pointer_rtx, src_mem) | |
5699 | || reg_mentioned_p (frame_pointer_rtx, src_mem) | |
5700 | || reg_mentioned_p (hard_frame_pointer_rtx, src_mem)) | |
949d79eb | 5701 | return 1; |
32de079a RE |
5702 | } |
5703 | ||
5704 | return cost; | |
5705 | } | |
5706 | ||
9b66ebb1 | 5707 | static int fp_consts_inited = 0; |
ff9940b0 | 5708 | |
9b66ebb1 PB |
5709 | /* Only zero is valid for VFP. Other values are also valid for FPA. */ |
5710 | static const char * const strings_fp[8] = | |
62b10bbc | 5711 | { |
2b835d68 RE |
5712 | "0", "1", "2", "3", |
5713 | "4", "5", "0.5", "10" | |
5714 | }; | |
ff9940b0 | 5715 | |
9b66ebb1 | 5716 | static REAL_VALUE_TYPE values_fp[8]; |
ff9940b0 RE |
5717 | |
5718 | static void | |
9b66ebb1 | 5719 | init_fp_table (void) |
ff9940b0 RE |
5720 | { |
5721 | int i; | |
5722 | REAL_VALUE_TYPE r; | |
5723 | ||
9b66ebb1 PB |
5724 | if (TARGET_VFP) |
5725 | fp_consts_inited = 1; | |
5726 | else | |
5727 | fp_consts_inited = 8; | |
5728 | ||
5729 | for (i = 0; i < fp_consts_inited; i++) | |
ff9940b0 | 5730 | { |
9b66ebb1 PB |
5731 | r = REAL_VALUE_ATOF (strings_fp[i], DFmode); |
5732 | values_fp[i] = r; | |
ff9940b0 | 5733 | } |
ff9940b0 RE |
5734 | } |
5735 | ||
9b66ebb1 | 5736 | /* Return TRUE if rtx X is a valid immediate FP constant. */ |
cce8749e | 5737 | int |
9b66ebb1 | 5738 | arm_const_double_rtx (rtx x) |
cce8749e | 5739 | { |
ff9940b0 RE |
5740 | REAL_VALUE_TYPE r; |
5741 | int i; | |
f676971a | 5742 | |
9b66ebb1 PB |
5743 | if (!fp_consts_inited) |
5744 | init_fp_table (); | |
f676971a | 5745 | |
ff9940b0 RE |
5746 | REAL_VALUE_FROM_CONST_DOUBLE (r, x); |
5747 | if (REAL_VALUE_MINUS_ZERO (r)) | |
5748 | return 0; | |
f3bb6135 | 5749 | |
9b66ebb1 PB |
5750 | for (i = 0; i < fp_consts_inited; i++) |
5751 | if (REAL_VALUES_EQUAL (r, values_fp[i])) | |
ff9940b0 | 5752 | return 1; |
f3bb6135 | 5753 | |
ff9940b0 | 5754 | return 0; |
f3bb6135 | 5755 | } |
ff9940b0 | 5756 | |
3b684012 | 5757 | /* Return TRUE if rtx X is a valid immediate FPA constant. */ |
ff9940b0 | 5758 | int |
e32bac5b | 5759 | neg_const_double_rtx_ok_for_fpa (rtx x) |
ff9940b0 RE |
5760 | { |
5761 | REAL_VALUE_TYPE r; | |
5762 | int i; | |
f676971a | 5763 | |
9b66ebb1 PB |
5764 | if (!fp_consts_inited) |
5765 | init_fp_table (); | |
f676971a | 5766 | |
ff9940b0 RE |
5767 | REAL_VALUE_FROM_CONST_DOUBLE (r, x); |
5768 | r = REAL_VALUE_NEGATE (r); | |
5769 | if (REAL_VALUE_MINUS_ZERO (r)) | |
5770 | return 0; | |
f3bb6135 | 5771 | |
ff9940b0 | 5772 | for (i = 0; i < 8; i++) |
9b66ebb1 | 5773 | if (REAL_VALUES_EQUAL (r, values_fp[i])) |
ff9940b0 | 5774 | return 1; |
f3bb6135 | 5775 | |
ff9940b0 | 5776 | return 0; |
f3bb6135 | 5777 | } |
f1adb0a9 JB |
5778 | |
5779 | ||
5780 | /* VFPv3 has a fairly wide range of representable immediates, formed from | |
5781 | "quarter-precision" floating-point values. These can be evaluated using this | |
5782 | formula (with ^ for exponentiation): | |
5783 | ||
5784 | -1^s * n * 2^-r | |
5785 | ||
5786 | Where 's' is a sign bit (0/1), 'n' and 'r' are integers such that | |
5787 | 16 <= n <= 31 and 0 <= r <= 7. | |
5788 | ||
5789 | These values are mapped onto an 8-bit integer ABCDEFGH s.t. | |
5790 | ||
5791 | - A (most-significant) is the sign bit. | |
5792 | - BCD are the exponent (encoded as r XOR 3). | |
5793 | - EFGH are the mantissa (encoded as n - 16). | |
5794 | */ | |
5795 | ||
5796 | /* Return an integer index for a VFPv3 immediate operand X suitable for the | |
5797 | fconst[sd] instruction, or -1 if X isn't suitable. */ | |
5798 | static int | |
5799 | vfp3_const_double_index (rtx x) | |
5800 | { | |
5801 | REAL_VALUE_TYPE r, m; | |
5802 | int sign, exponent; | |
5803 | unsigned HOST_WIDE_INT mantissa, mant_hi; | |
5804 | unsigned HOST_WIDE_INT mask; | |
8e39e9af | 5805 | HOST_WIDE_INT m1, m2; |
f1adb0a9 JB |
5806 | int point_pos = 2 * HOST_BITS_PER_WIDE_INT - 1; |
5807 | ||
5808 | if (!TARGET_VFP3 || GET_CODE (x) != CONST_DOUBLE) | |
5809 | return -1; | |
5810 | ||
5811 | REAL_VALUE_FROM_CONST_DOUBLE (r, x); | |
5812 | ||
5813 | /* We can't represent these things, so detect them first. */ | |
5814 | if (REAL_VALUE_ISINF (r) || REAL_VALUE_ISNAN (r) || REAL_VALUE_MINUS_ZERO (r)) | |
5815 | return -1; | |
5816 | ||
5817 | /* Extract sign, exponent and mantissa. */ | |
5818 | sign = REAL_VALUE_NEGATIVE (r) ? 1 : 0; | |
5819 | r = REAL_VALUE_ABS (r); | |
5820 | exponent = REAL_EXP (&r); | |
5821 | /* For the mantissa, we expand into two HOST_WIDE_INTS, apart from the | |
5822 | highest (sign) bit, with a fixed binary point at bit point_pos. | |
5823 | WARNING: If there's ever a VFP version which uses more than 2 * H_W_I - 1 | |
5824 | bits for the mantissa, this may fail (low bits would be lost). */ | |
5825 | real_ldexp (&m, &r, point_pos - exponent); | |
8e39e9af RE |
5826 | REAL_VALUE_TO_INT (&m1, &m2, m); |
5827 | mantissa = m1; | |
5828 | mant_hi = m2; | |
f1adb0a9 JB |
5829 | |
5830 | /* If there are bits set in the low part of the mantissa, we can't | |
5831 | represent this value. */ | |
5832 | if (mantissa != 0) | |
5833 | return -1; | |
5834 | ||
5835 | /* Now make it so that mantissa contains the most-significant bits, and move | |
5836 | the point_pos to indicate that the least-significant bits have been | |
5837 | discarded. */ | |
5838 | point_pos -= HOST_BITS_PER_WIDE_INT; | |
5839 | mantissa = mant_hi; | |
5840 | ||
5841 | /* We can permit four significant bits of mantissa only, plus a high bit | |
5842 | which is always 1. */ | |
5843 | mask = ((unsigned HOST_WIDE_INT)1 << (point_pos - 5)) - 1; | |
5844 | if ((mantissa & mask) != 0) | |
5845 | return -1; | |
5846 | ||
5847 | /* Now we know the mantissa is in range, chop off the unneeded bits. */ | |
5848 | mantissa >>= point_pos - 5; | |
5849 | ||
5850 | /* The mantissa may be zero. Disallow that case. (It's possible to load the | |
5851 | floating-point immediate zero with Neon using an integer-zero load, but | |
5852 | that case is handled elsewhere.) */ | |
5853 | if (mantissa == 0) | |
5854 | return -1; | |
5855 | ||
5856 | gcc_assert (mantissa >= 16 && mantissa <= 31); | |
5857 | ||
5858 | /* The value of 5 here would be 4 if GCC used IEEE754-like encoding (where | |
6ed3da00 KH |
5859 | normalized significands are in the range [1, 2). (Our mantissa is shifted |
5860 | left 4 places at this point relative to normalized IEEE754 values). GCC | |
f1adb0a9 JB |
5861 | internally uses [0.5, 1) (see real.c), so the exponent returned from |
5862 | REAL_EXP must be altered. */ | |
5863 | exponent = 5 - exponent; | |
5864 | ||
5865 | if (exponent < 0 || exponent > 7) | |
5866 | return -1; | |
5867 | ||
5868 | /* Sign, mantissa and exponent are now in the correct form to plug into the | |
15dc95cb | 5869 | formula described in the comment above. */ |
f1adb0a9 JB |
5870 | return (sign << 7) | ((exponent ^ 3) << 4) | (mantissa - 16); |
5871 | } | |
5872 | ||
5873 | /* Return TRUE if rtx X is a valid immediate VFPv3 constant. */ | |
5874 | int | |
5875 | vfp3_const_double_rtx (rtx x) | |
5876 | { | |
5877 | if (!TARGET_VFP3) | |
5878 | return 0; | |
5879 | ||
5880 | return vfp3_const_double_index (x) != -1; | |
5881 | } | |
5882 | ||
88f77cba JB |
5883 | /* Recognize immediates which can be used in various Neon instructions. Legal |
5884 | immediates are described by the following table (for VMVN variants, the | |
5885 | bitwise inverse of the constant shown is recognized. In either case, VMOV | |
5886 | is output and the correct instruction to use for a given constant is chosen | |
5887 | by the assembler). The constant shown is replicated across all elements of | |
5888 | the destination vector. | |
5889 | ||
5890 | insn elems variant constant (binary) | |
5891 | ---- ----- ------- ----------------- | |
5892 | vmov i32 0 00000000 00000000 00000000 abcdefgh | |
5893 | vmov i32 1 00000000 00000000 abcdefgh 00000000 | |
5894 | vmov i32 2 00000000 abcdefgh 00000000 00000000 | |
5895 | vmov i32 3 abcdefgh 00000000 00000000 00000000 | |
5896 | vmov i16 4 00000000 abcdefgh | |
5897 | vmov i16 5 abcdefgh 00000000 | |
5898 | vmvn i32 6 00000000 00000000 00000000 abcdefgh | |
5899 | vmvn i32 7 00000000 00000000 abcdefgh 00000000 | |
5900 | vmvn i32 8 00000000 abcdefgh 00000000 00000000 | |
5901 | vmvn i32 9 abcdefgh 00000000 00000000 00000000 | |
5902 | vmvn i16 10 00000000 abcdefgh | |
5903 | vmvn i16 11 abcdefgh 00000000 | |
5904 | vmov i32 12 00000000 00000000 abcdefgh 11111111 | |
5905 | vmvn i32 13 00000000 00000000 abcdefgh 11111111 | |
5906 | vmov i32 14 00000000 abcdefgh 11111111 11111111 | |
5907 | vmvn i32 15 00000000 abcdefgh 11111111 11111111 | |
5908 | vmov i8 16 abcdefgh | |
5909 | vmov i64 17 aaaaaaaa bbbbbbbb cccccccc dddddddd | |
5910 | eeeeeeee ffffffff gggggggg hhhhhhhh | |
5911 | vmov f32 18 aBbbbbbc defgh000 00000000 00000000 | |
5912 | ||
5913 | For case 18, B = !b. Representable values are exactly those accepted by | |
5914 | vfp3_const_double_index, but are output as floating-point numbers rather | |
5915 | than indices. | |
5916 | ||
5917 | Variants 0-5 (inclusive) may also be used as immediates for the second | |
5918 | operand of VORR/VBIC instructions. | |
5919 | ||
5920 | The INVERSE argument causes the bitwise inverse of the given operand to be | |
5921 | recognized instead (used for recognizing legal immediates for the VAND/VORN | |
5922 | pseudo-instructions). If INVERSE is true, the value placed in *MODCONST is | |
5923 | *not* inverted (i.e. the pseudo-instruction forms vand/vorn should still be | |
5924 | output, rather than the real insns vbic/vorr). | |
5925 | ||
5926 | INVERSE makes no difference to the recognition of float vectors. | |
5927 | ||
5928 | The return value is the variant of immediate as shown in the above table, or | |
5929 | -1 if the given value doesn't match any of the listed patterns. | |
5930 | */ | |
5931 | static int | |
5932 | neon_valid_immediate (rtx op, enum machine_mode mode, int inverse, | |
5933 | rtx *modconst, int *elementwidth) | |
5934 | { | |
5935 | #define CHECK(STRIDE, ELSIZE, CLASS, TEST) \ | |
5936 | matches = 1; \ | |
5937 | for (i = 0; i < idx; i += (STRIDE)) \ | |
5938 | if (!(TEST)) \ | |
5939 | matches = 0; \ | |
5940 | if (matches) \ | |
5941 | { \ | |
5942 | immtype = (CLASS); \ | |
5943 | elsize = (ELSIZE); \ | |
5944 | break; \ | |
5945 | } | |
5946 | ||
5947 | unsigned int i, elsize, idx = 0, n_elts = CONST_VECTOR_NUNITS (op); | |
5948 | unsigned int innersize = GET_MODE_SIZE (GET_MODE_INNER (mode)); | |
5949 | unsigned char bytes[16]; | |
5950 | int immtype = -1, matches; | |
5951 | unsigned int invmask = inverse ? 0xff : 0; | |
5952 | ||
5953 | /* Vectors of float constants. */ | |
5954 | if (GET_MODE_CLASS (mode) == MODE_VECTOR_FLOAT) | |
5955 | { | |
5956 | rtx el0 = CONST_VECTOR_ELT (op, 0); | |
5957 | REAL_VALUE_TYPE r0; | |
5958 | ||
5959 | if (!vfp3_const_double_rtx (el0)) | |
5960 | return -1; | |
5961 | ||
5962 | REAL_VALUE_FROM_CONST_DOUBLE (r0, el0); | |
5963 | ||
5964 | for (i = 1; i < n_elts; i++) | |
5965 | { | |
5966 | rtx elt = CONST_VECTOR_ELT (op, i); | |
5967 | REAL_VALUE_TYPE re; | |
5968 | ||
5969 | REAL_VALUE_FROM_CONST_DOUBLE (re, elt); | |
5970 | ||
5971 | if (!REAL_VALUES_EQUAL (r0, re)) | |
5972 | return -1; | |
5973 | } | |
5974 | ||
5975 | if (modconst) | |
5976 | *modconst = CONST_VECTOR_ELT (op, 0); | |
5977 | ||
5978 | if (elementwidth) | |
5979 | *elementwidth = 0; | |
5980 | ||
5981 | return 18; | |
5982 | } | |
5983 | ||
5984 | /* Splat vector constant out into a byte vector. */ | |
5985 | for (i = 0; i < n_elts; i++) | |
5986 | { | |
5987 | rtx el = CONST_VECTOR_ELT (op, i); | |
5988 | unsigned HOST_WIDE_INT elpart; | |
5989 | unsigned int part, parts; | |
5990 | ||
5991 | if (GET_CODE (el) == CONST_INT) | |
5992 | { | |
5993 | elpart = INTVAL (el); | |
5994 | parts = 1; | |
5995 | } | |
5996 | else if (GET_CODE (el) == CONST_DOUBLE) | |
5997 | { | |
5998 | elpart = CONST_DOUBLE_LOW (el); | |
5999 | parts = 2; | |
6000 | } | |
6001 | else | |
6002 | gcc_unreachable (); | |
6003 | ||
6004 | for (part = 0; part < parts; part++) | |
6005 | { | |
6006 | unsigned int byte; | |
6007 | for (byte = 0; byte < innersize; byte++) | |
6008 | { | |
6009 | bytes[idx++] = (elpart & 0xff) ^ invmask; | |
6010 | elpart >>= BITS_PER_UNIT; | |
6011 | } | |
6012 | if (GET_CODE (el) == CONST_DOUBLE) | |
6013 | elpart = CONST_DOUBLE_HIGH (el); | |
6014 | } | |
6015 | } | |
6016 | ||
6017 | /* Sanity check. */ | |
6018 | gcc_assert (idx == GET_MODE_SIZE (mode)); | |
6019 | ||
6020 | do | |
6021 | { | |
6022 | CHECK (4, 32, 0, bytes[i] == bytes[0] && bytes[i + 1] == 0 | |
6023 | && bytes[i + 2] == 0 && bytes[i + 3] == 0); | |
6024 | ||
6025 | CHECK (4, 32, 1, bytes[i] == 0 && bytes[i + 1] == bytes[1] | |
6026 | && bytes[i + 2] == 0 && bytes[i + 3] == 0); | |
6027 | ||
6028 | CHECK (4, 32, 2, bytes[i] == 0 && bytes[i + 1] == 0 | |
6029 | && bytes[i + 2] == bytes[2] && bytes[i + 3] == 0); | |
6030 | ||
6031 | CHECK (4, 32, 3, bytes[i] == 0 && bytes[i + 1] == 0 | |
6032 | && bytes[i + 2] == 0 && bytes[i + 3] == bytes[3]); | |
6033 | ||
6034 | CHECK (2, 16, 4, bytes[i] == bytes[0] && bytes[i + 1] == 0); | |
6035 | ||
6036 | CHECK (2, 16, 5, bytes[i] == 0 && bytes[i + 1] == bytes[1]); | |
6037 | ||
6038 | CHECK (4, 32, 6, bytes[i] == bytes[0] && bytes[i + 1] == 0xff | |
6039 | && bytes[i + 2] == 0xff && bytes[i + 3] == 0xff); | |
6040 | ||
6041 | CHECK (4, 32, 7, bytes[i] == 0xff && bytes[i + 1] == bytes[1] | |
6042 | && bytes[i + 2] == 0xff && bytes[i + 3] == 0xff); | |
6043 | ||
6044 | CHECK (4, 32, 8, bytes[i] == 0xff && bytes[i + 1] == 0xff | |
6045 | && bytes[i + 2] == bytes[2] && bytes[i + 3] == 0xff); | |
6046 | ||
6047 | CHECK (4, 32, 9, bytes[i] == 0xff && bytes[i + 1] == 0xff | |
6048 | && bytes[i + 2] == 0xff && bytes[i + 3] == bytes[3]); | |
6049 | ||
6050 | CHECK (2, 16, 10, bytes[i] == bytes[0] && bytes[i + 1] == 0xff); | |
6051 | ||
6052 | CHECK (2, 16, 11, bytes[i] == 0xff && bytes[i + 1] == bytes[1]); | |
6053 | ||
6054 | CHECK (4, 32, 12, bytes[i] == 0xff && bytes[i + 1] == bytes[1] | |
6055 | && bytes[i + 2] == 0 && bytes[i + 3] == 0); | |
6056 | ||
6057 | CHECK (4, 32, 13, bytes[i] == 0 && bytes[i + 1] == bytes[1] | |
6058 | && bytes[i + 2] == 0xff && bytes[i + 3] == 0xff); | |
6059 | ||
6060 | CHECK (4, 32, 14, bytes[i] == 0xff && bytes[i + 1] == 0xff | |
6061 | && bytes[i + 2] == bytes[2] && bytes[i + 3] == 0); | |
6062 | ||
6063 | CHECK (4, 32, 15, bytes[i] == 0 && bytes[i + 1] == 0 | |
6064 | && bytes[i + 2] == bytes[2] && bytes[i + 3] == 0xff); | |
6065 | ||
6066 | CHECK (1, 8, 16, bytes[i] == bytes[0]); | |
6067 | ||
6068 | CHECK (1, 64, 17, (bytes[i] == 0 || bytes[i] == 0xff) | |
6069 | && bytes[i] == bytes[(i + 8) % idx]); | |
6070 | } | |
6071 | while (0); | |
6072 | ||
6073 | if (immtype == -1) | |
6074 | return -1; | |
6075 | ||
6076 | if (elementwidth) | |
6077 | *elementwidth = elsize; | |
6078 | ||
6079 | if (modconst) | |
6080 | { | |
6081 | unsigned HOST_WIDE_INT imm = 0; | |
6082 | ||
cea618ac | 6083 | /* Un-invert bytes of recognized vector, if necessary. */ |
88f77cba JB |
6084 | if (invmask != 0) |
6085 | for (i = 0; i < idx; i++) | |
6086 | bytes[i] ^= invmask; | |
6087 | ||
6088 | if (immtype == 17) | |
6089 | { | |
6090 | /* FIXME: Broken on 32-bit H_W_I hosts. */ | |
6091 | gcc_assert (sizeof (HOST_WIDE_INT) == 8); | |
6092 | ||
6093 | for (i = 0; i < 8; i++) | |
6094 | imm |= (unsigned HOST_WIDE_INT) (bytes[i] ? 0xff : 0) | |
6095 | << (i * BITS_PER_UNIT); | |
6096 | ||
6097 | *modconst = GEN_INT (imm); | |
6098 | } | |
6099 | else | |
6100 | { | |
6101 | unsigned HOST_WIDE_INT imm = 0; | |
6102 | ||
6103 | for (i = 0; i < elsize / BITS_PER_UNIT; i++) | |
6104 | imm |= (unsigned HOST_WIDE_INT) bytes[i] << (i * BITS_PER_UNIT); | |
6105 | ||
6106 | *modconst = GEN_INT (imm); | |
6107 | } | |
6108 | } | |
6109 | ||
6110 | return immtype; | |
6111 | #undef CHECK | |
6112 | } | |
6113 | ||
6114 | /* Return TRUE if rtx X is legal for use as either a Neon VMOV (or, implicitly, | |
6115 | VMVN) immediate. Write back width per element to *ELEMENTWIDTH (or zero for | |
6116 | float elements), and a modified constant (whatever should be output for a | |
6117 | VMOV) in *MODCONST. */ | |
6118 | ||
6119 | int | |
6120 | neon_immediate_valid_for_move (rtx op, enum machine_mode mode, | |
6121 | rtx *modconst, int *elementwidth) | |
6122 | { | |
6123 | rtx tmpconst; | |
6124 | int tmpwidth; | |
6125 | int retval = neon_valid_immediate (op, mode, 0, &tmpconst, &tmpwidth); | |
6126 | ||
6127 | if (retval == -1) | |
6128 | return 0; | |
6129 | ||
6130 | if (modconst) | |
6131 | *modconst = tmpconst; | |
6132 | ||
6133 | if (elementwidth) | |
6134 | *elementwidth = tmpwidth; | |
6135 | ||
6136 | return 1; | |
6137 | } | |
6138 | ||
6139 | /* Return TRUE if rtx X is legal for use in a VORR or VBIC instruction. If | |
6140 | the immediate is valid, write a constant suitable for using as an operand | |
6141 | to VORR/VBIC/VAND/VORN to *MODCONST and the corresponding element width to | |
6142 | *ELEMENTWIDTH. See neon_valid_immediate for description of INVERSE. */ | |
6143 | ||
6144 | int | |
6145 | neon_immediate_valid_for_logic (rtx op, enum machine_mode mode, int inverse, | |
6146 | rtx *modconst, int *elementwidth) | |
6147 | { | |
6148 | rtx tmpconst; | |
6149 | int tmpwidth; | |
6150 | int retval = neon_valid_immediate (op, mode, inverse, &tmpconst, &tmpwidth); | |
6151 | ||
6152 | if (retval < 0 || retval > 5) | |
6153 | return 0; | |
6154 | ||
6155 | if (modconst) | |
6156 | *modconst = tmpconst; | |
6157 | ||
6158 | if (elementwidth) | |
6159 | *elementwidth = tmpwidth; | |
6160 | ||
6161 | return 1; | |
6162 | } | |
6163 | ||
6164 | /* Return a string suitable for output of Neon immediate logic operation | |
6165 | MNEM. */ | |
6166 | ||
6167 | char * | |
6168 | neon_output_logic_immediate (const char *mnem, rtx *op2, enum machine_mode mode, | |
6169 | int inverse, int quad) | |
6170 | { | |
6171 | int width, is_valid; | |
6172 | static char templ[40]; | |
6173 | ||
6174 | is_valid = neon_immediate_valid_for_logic (*op2, mode, inverse, op2, &width); | |
6175 | ||
6176 | gcc_assert (is_valid != 0); | |
6177 | ||
6178 | if (quad) | |
6179 | sprintf (templ, "%s.i%d\t%%q0, %%2", mnem, width); | |
6180 | else | |
6181 | sprintf (templ, "%s.i%d\t%%P0, %%2", mnem, width); | |
6182 | ||
6183 | return templ; | |
6184 | } | |
6185 | ||
6186 | /* Output a sequence of pairwise operations to implement a reduction. | |
6187 | NOTE: We do "too much work" here, because pairwise operations work on two | |
6188 | registers-worth of operands in one go. Unfortunately we can't exploit those | |
6189 | extra calculations to do the full operation in fewer steps, I don't think. | |
6190 | Although all vector elements of the result but the first are ignored, we | |
6191 | actually calculate the same result in each of the elements. An alternative | |
6192 | such as initially loading a vector with zero to use as each of the second | |
6193 | operands would use up an additional register and take an extra instruction, | |
6194 | for no particular gain. */ | |
6195 | ||
6196 | void | |
6197 | neon_pairwise_reduce (rtx op0, rtx op1, enum machine_mode mode, | |
6198 | rtx (*reduc) (rtx, rtx, rtx)) | |
6199 | { | |
6200 | enum machine_mode inner = GET_MODE_INNER (mode); | |
6201 | unsigned int i, parts = GET_MODE_SIZE (mode) / GET_MODE_SIZE (inner); | |
6202 | rtx tmpsum = op1; | |
6203 | ||
6204 | for (i = parts / 2; i >= 1; i /= 2) | |
6205 | { | |
6206 | rtx dest = (i == 1) ? op0 : gen_reg_rtx (mode); | |
6207 | emit_insn (reduc (dest, tmpsum, tmpsum)); | |
6208 | tmpsum = dest; | |
6209 | } | |
6210 | } | |
6211 | ||
cea618ac | 6212 | /* Initialize a vector with non-constant elements. FIXME: We can do better |
88f77cba JB |
6213 | than the current implementation (building a vector on the stack and then |
6214 | loading it) in many cases. See rs6000.c. */ | |
6215 | ||
6216 | void | |
6217 | neon_expand_vector_init (rtx target, rtx vals) | |
6218 | { | |
6219 | enum machine_mode mode = GET_MODE (target); | |
6220 | enum machine_mode inner = GET_MODE_INNER (mode); | |
6221 | unsigned int i, n_elts = GET_MODE_NUNITS (mode); | |
6222 | rtx mem; | |
6223 | ||
6224 | gcc_assert (VECTOR_MODE_P (mode)); | |
6225 | ||
6226 | mem = assign_stack_temp (mode, GET_MODE_SIZE (mode), 0); | |
6227 | for (i = 0; i < n_elts; i++) | |
6228 | emit_move_insn (adjust_address_nv (mem, inner, i * GET_MODE_SIZE (inner)), | |
6229 | XVECEXP (vals, 0, i)); | |
6230 | ||
6231 | emit_move_insn (target, mem); | |
6232 | } | |
6233 | ||
b617fc71 JB |
6234 | /* Ensure OPERAND lies between LOW (inclusive) and HIGH (exclusive). Raise |
6235 | ERR if it doesn't. FIXME: NEON bounds checks occur late in compilation, so | |
6236 | reported source locations are bogus. */ | |
6237 | ||
6238 | static void | |
6239 | bounds_check (rtx operand, HOST_WIDE_INT low, HOST_WIDE_INT high, | |
6240 | const char *err) | |
6241 | { | |
6242 | HOST_WIDE_INT lane; | |
6243 | ||
6244 | gcc_assert (GET_CODE (operand) == CONST_INT); | |
6245 | ||
6246 | lane = INTVAL (operand); | |
6247 | ||
6248 | if (lane < low || lane >= high) | |
6249 | error (err); | |
6250 | } | |
6251 | ||
6252 | /* Bounds-check lanes. */ | |
6253 | ||
6254 | void | |
6255 | neon_lane_bounds (rtx operand, HOST_WIDE_INT low, HOST_WIDE_INT high) | |
6256 | { | |
6257 | bounds_check (operand, low, high, "lane out of range"); | |
6258 | } | |
6259 | ||
6260 | /* Bounds-check constants. */ | |
6261 | ||
6262 | void | |
6263 | neon_const_bounds (rtx operand, HOST_WIDE_INT low, HOST_WIDE_INT high) | |
6264 | { | |
6265 | bounds_check (operand, low, high, "constant out of range"); | |
6266 | } | |
6267 | ||
6268 | HOST_WIDE_INT | |
6269 | neon_element_bits (enum machine_mode mode) | |
6270 | { | |
6271 | if (mode == DImode) | |
6272 | return GET_MODE_BITSIZE (mode); | |
6273 | else | |
6274 | return GET_MODE_BITSIZE (GET_MODE_INNER (mode)); | |
6275 | } | |
6276 | ||
cce8749e CH |
6277 | \f |
6278 | /* Predicates for `match_operand' and `match_operator'. */ | |
6279 | ||
9b6b54e2 | 6280 | /* Return nonzero if OP is a valid Cirrus memory address pattern. */ |
9b6b54e2 | 6281 | int |
e32bac5b | 6282 | cirrus_memory_offset (rtx op) |
9b6b54e2 NC |
6283 | { |
6284 | /* Reject eliminable registers. */ | |
6285 | if (! (reload_in_progress || reload_completed) | |
6286 | && ( reg_mentioned_p (frame_pointer_rtx, op) | |
6287 | || reg_mentioned_p (arg_pointer_rtx, op) | |
6288 | || reg_mentioned_p (virtual_incoming_args_rtx, op) | |
6289 | || reg_mentioned_p (virtual_outgoing_args_rtx, op) | |
6290 | || reg_mentioned_p (virtual_stack_dynamic_rtx, op) | |
6291 | || reg_mentioned_p (virtual_stack_vars_rtx, op))) | |
6292 | return 0; | |
6293 | ||
6294 | if (GET_CODE (op) == MEM) | |
6295 | { | |
6296 | rtx ind; | |
6297 | ||
6298 | ind = XEXP (op, 0); | |
6299 | ||
6300 | /* Match: (mem (reg)). */ | |
6301 | if (GET_CODE (ind) == REG) | |
6302 | return 1; | |
6303 | ||
6304 | /* Match: | |
6305 | (mem (plus (reg) | |
6306 | (const))). */ | |
6307 | if (GET_CODE (ind) == PLUS | |
6308 | && GET_CODE (XEXP (ind, 0)) == REG | |
6309 | && REG_MODE_OK_FOR_BASE_P (XEXP (ind, 0), VOIDmode) | |
6310 | && GET_CODE (XEXP (ind, 1)) == CONST_INT) | |
6311 | return 1; | |
6312 | } | |
6313 | ||
6314 | return 0; | |
6315 | } | |
6316 | ||
f26b8ec9 | 6317 | /* Return TRUE if OP is a valid coprocessor memory address pattern. |
5b3e6663 PB |
6318 | WB is true if full writeback address modes are allowed and is false |
6319 | if limited writeback address modes (POST_INC and PRE_DEC) are | |
6320 | allowed. */ | |
9b66ebb1 PB |
6321 | |
6322 | int | |
fdd695fd | 6323 | arm_coproc_mem_operand (rtx op, bool wb) |
9b66ebb1 | 6324 | { |
fdd695fd | 6325 | rtx ind; |
9b66ebb1 | 6326 | |
fdd695fd | 6327 | /* Reject eliminable registers. */ |
9b66ebb1 PB |
6328 | if (! (reload_in_progress || reload_completed) |
6329 | && ( reg_mentioned_p (frame_pointer_rtx, op) | |
6330 | || reg_mentioned_p (arg_pointer_rtx, op) | |
6331 | || reg_mentioned_p (virtual_incoming_args_rtx, op) | |
6332 | || reg_mentioned_p (virtual_outgoing_args_rtx, op) | |
6333 | || reg_mentioned_p (virtual_stack_dynamic_rtx, op) | |
6334 | || reg_mentioned_p (virtual_stack_vars_rtx, op))) | |
6335 | return FALSE; | |
6336 | ||
59b9a953 | 6337 | /* Constants are converted into offsets from labels. */ |
fdd695fd PB |
6338 | if (GET_CODE (op) != MEM) |
6339 | return FALSE; | |
9b66ebb1 | 6340 | |
fdd695fd | 6341 | ind = XEXP (op, 0); |
9b66ebb1 | 6342 | |
fdd695fd PB |
6343 | if (reload_completed |
6344 | && (GET_CODE (ind) == LABEL_REF | |
6345 | || (GET_CODE (ind) == CONST | |
6346 | && GET_CODE (XEXP (ind, 0)) == PLUS | |
6347 | && GET_CODE (XEXP (XEXP (ind, 0), 0)) == LABEL_REF | |
6348 | && GET_CODE (XEXP (XEXP (ind, 0), 1)) == CONST_INT))) | |
6349 | return TRUE; | |
9b66ebb1 | 6350 | |
fdd695fd PB |
6351 | /* Match: (mem (reg)). */ |
6352 | if (GET_CODE (ind) == REG) | |
6353 | return arm_address_register_rtx_p (ind, 0); | |
6354 | ||
5b3e6663 PB |
6355 | /* Autoincremment addressing modes. POST_INC and PRE_DEC are |
6356 | acceptable in any case (subject to verification by | |
6357 | arm_address_register_rtx_p). We need WB to be true to accept | |
6358 | PRE_INC and POST_DEC. */ | |
6359 | if (GET_CODE (ind) == POST_INC | |
6360 | || GET_CODE (ind) == PRE_DEC | |
6361 | || (wb | |
6362 | && (GET_CODE (ind) == PRE_INC | |
6363 | || GET_CODE (ind) == POST_DEC))) | |
fdd695fd PB |
6364 | return arm_address_register_rtx_p (XEXP (ind, 0), 0); |
6365 | ||
6366 | if (wb | |
6367 | && (GET_CODE (ind) == POST_MODIFY || GET_CODE (ind) == PRE_MODIFY) | |
6368 | && arm_address_register_rtx_p (XEXP (ind, 0), 0) | |
6369 | && GET_CODE (XEXP (ind, 1)) == PLUS | |
6370 | && rtx_equal_p (XEXP (XEXP (ind, 1), 0), XEXP (ind, 0))) | |
6371 | ind = XEXP (ind, 1); | |
6372 | ||
6373 | /* Match: | |
6374 | (plus (reg) | |
6375 | (const)). */ | |
6376 | if (GET_CODE (ind) == PLUS | |
6377 | && GET_CODE (XEXP (ind, 0)) == REG | |
6378 | && REG_MODE_OK_FOR_BASE_P (XEXP (ind, 0), VOIDmode) | |
6379 | && GET_CODE (XEXP (ind, 1)) == CONST_INT | |
6380 | && INTVAL (XEXP (ind, 1)) > -1024 | |
6381 | && INTVAL (XEXP (ind, 1)) < 1024 | |
6382 | && (INTVAL (XEXP (ind, 1)) & 3) == 0) | |
6383 | return TRUE; | |
9b66ebb1 PB |
6384 | |
6385 | return FALSE; | |
6386 | } | |
6387 | ||
88f77cba JB |
6388 | /* Return TRUE if OP is a memory operand which we can load or store a vector |
6389 | to/from. If CORE is true, we're moving from ARM registers not Neon | |
6390 | registers. */ | |
6391 | int | |
6392 | neon_vector_mem_operand (rtx op, bool core) | |
6393 | { | |
6394 | rtx ind; | |
6395 | ||
6396 | /* Reject eliminable registers. */ | |
6397 | if (! (reload_in_progress || reload_completed) | |
6398 | && ( reg_mentioned_p (frame_pointer_rtx, op) | |
6399 | || reg_mentioned_p (arg_pointer_rtx, op) | |
6400 | || reg_mentioned_p (virtual_incoming_args_rtx, op) | |
6401 | || reg_mentioned_p (virtual_outgoing_args_rtx, op) | |
6402 | || reg_mentioned_p (virtual_stack_dynamic_rtx, op) | |
6403 | || reg_mentioned_p (virtual_stack_vars_rtx, op))) | |
6404 | return FALSE; | |
6405 | ||
6406 | /* Constants are converted into offsets from labels. */ | |
6407 | if (GET_CODE (op) != MEM) | |
6408 | return FALSE; | |
6409 | ||
6410 | ind = XEXP (op, 0); | |
6411 | ||
6412 | if (reload_completed | |
6413 | && (GET_CODE (ind) == LABEL_REF | |
6414 | || (GET_CODE (ind) == CONST | |
6415 | && GET_CODE (XEXP (ind, 0)) == PLUS | |
6416 | && GET_CODE (XEXP (XEXP (ind, 0), 0)) == LABEL_REF | |
6417 | && GET_CODE (XEXP (XEXP (ind, 0), 1)) == CONST_INT))) | |
6418 | return TRUE; | |
6419 | ||
6420 | /* Match: (mem (reg)). */ | |
6421 | if (GET_CODE (ind) == REG) | |
6422 | return arm_address_register_rtx_p (ind, 0); | |
6423 | ||
6424 | /* Allow post-increment with Neon registers. */ | |
6425 | if (!core && GET_CODE (ind) == POST_INC) | |
6426 | return arm_address_register_rtx_p (XEXP (ind, 0), 0); | |
6427 | ||
6428 | #if 0 | |
6429 | /* FIXME: We can support this too if we use VLD1/VST1. */ | |
6430 | if (!core | |
6431 | && GET_CODE (ind) == POST_MODIFY | |
6432 | && arm_address_register_rtx_p (XEXP (ind, 0), 0) | |
6433 | && GET_CODE (XEXP (ind, 1)) == PLUS | |
6434 | && rtx_equal_p (XEXP (XEXP (ind, 1), 0), XEXP (ind, 0))) | |
6435 | ind = XEXP (ind, 1); | |
6436 | #endif | |
6437 | ||
6438 | /* Match: | |
6439 | (plus (reg) | |
6440 | (const)). */ | |
6441 | if (!core | |
6442 | && GET_CODE (ind) == PLUS | |
6443 | && GET_CODE (XEXP (ind, 0)) == REG | |
6444 | && REG_MODE_OK_FOR_BASE_P (XEXP (ind, 0), VOIDmode) | |
6445 | && GET_CODE (XEXP (ind, 1)) == CONST_INT | |
6446 | && INTVAL (XEXP (ind, 1)) > -1024 | |
6447 | && INTVAL (XEXP (ind, 1)) < 1016 | |
6448 | && (INTVAL (XEXP (ind, 1)) & 3) == 0) | |
6449 | return TRUE; | |
6450 | ||
6451 | return FALSE; | |
6452 | } | |
6453 | ||
6454 | /* Return TRUE if OP is a mem suitable for loading/storing a Neon struct | |
6455 | type. */ | |
6456 | int | |
6457 | neon_struct_mem_operand (rtx op) | |
6458 | { | |
6459 | rtx ind; | |
6460 | ||
6461 | /* Reject eliminable registers. */ | |
6462 | if (! (reload_in_progress || reload_completed) | |
6463 | && ( reg_mentioned_p (frame_pointer_rtx, op) | |
6464 | || reg_mentioned_p (arg_pointer_rtx, op) | |
6465 | || reg_mentioned_p (virtual_incoming_args_rtx, op) | |
6466 | || reg_mentioned_p (virtual_outgoing_args_rtx, op) | |
6467 | || reg_mentioned_p (virtual_stack_dynamic_rtx, op) | |
6468 | || reg_mentioned_p (virtual_stack_vars_rtx, op))) | |
6469 | return FALSE; | |
6470 | ||
6471 | /* Constants are converted into offsets from labels. */ | |
6472 | if (GET_CODE (op) != MEM) | |
6473 | return FALSE; | |
6474 | ||
6475 | ind = XEXP (op, 0); | |
6476 | ||
6477 | if (reload_completed | |
6478 | && (GET_CODE (ind) == LABEL_REF | |
6479 | || (GET_CODE (ind) == CONST | |
6480 | && GET_CODE (XEXP (ind, 0)) == PLUS | |
6481 | && GET_CODE (XEXP (XEXP (ind, 0), 0)) == LABEL_REF | |
6482 | && GET_CODE (XEXP (XEXP (ind, 0), 1)) == CONST_INT))) | |
6483 | return TRUE; | |
6484 | ||
6485 | /* Match: (mem (reg)). */ | |
6486 | if (GET_CODE (ind) == REG) | |
6487 | return arm_address_register_rtx_p (ind, 0); | |
6488 | ||
6489 | return FALSE; | |
6490 | } | |
6491 | ||
6555b6bd RE |
6492 | /* Return true if X is a register that will be eliminated later on. */ |
6493 | int | |
6494 | arm_eliminable_register (rtx x) | |
6495 | { | |
6496 | return REG_P (x) && (REGNO (x) == FRAME_POINTER_REGNUM | |
6497 | || REGNO (x) == ARG_POINTER_REGNUM | |
6498 | || (REGNO (x) >= FIRST_VIRTUAL_REGISTER | |
6499 | && REGNO (x) <= LAST_VIRTUAL_REGISTER)); | |
6500 | } | |
9b66ebb1 | 6501 | |
9b66ebb1 | 6502 | /* Return GENERAL_REGS if a scratch register required to reload x to/from |
fe2d934b | 6503 | coprocessor registers. Otherwise return NO_REGS. */ |
9b66ebb1 PB |
6504 | |
6505 | enum reg_class | |
fe2d934b | 6506 | coproc_secondary_reload_class (enum machine_mode mode, rtx x, bool wb) |
9b66ebb1 | 6507 | { |
88f77cba JB |
6508 | if (TARGET_NEON |
6509 | && (GET_MODE_CLASS (mode) == MODE_VECTOR_INT | |
6510 | || GET_MODE_CLASS (mode) == MODE_VECTOR_FLOAT) | |
6511 | && neon_vector_mem_operand (x, FALSE)) | |
6512 | return NO_REGS; | |
6513 | ||
fe2d934b | 6514 | if (arm_coproc_mem_operand (x, wb) || s_register_operand (x, mode)) |
9b66ebb1 PB |
6515 | return NO_REGS; |
6516 | ||
6517 | return GENERAL_REGS; | |
6518 | } | |
6519 | ||
866af8a9 JB |
6520 | /* Values which must be returned in the most-significant end of the return |
6521 | register. */ | |
6522 | ||
6523 | static bool | |
586de218 | 6524 | arm_return_in_msb (const_tree valtype) |
866af8a9 JB |
6525 | { |
6526 | return (TARGET_AAPCS_BASED | |
6527 | && BYTES_BIG_ENDIAN | |
6528 | && (AGGREGATE_TYPE_P (valtype) | |
6529 | || TREE_CODE (valtype) == COMPLEX_TYPE)); | |
6530 | } | |
9b66ebb1 | 6531 | |
f0375c66 NC |
6532 | /* Returns TRUE if INSN is an "LDR REG, ADDR" instruction. |
6533 | Use by the Cirrus Maverick code which has to workaround | |
6534 | a hardware bug triggered by such instructions. */ | |
f0375c66 | 6535 | static bool |
e32bac5b | 6536 | arm_memory_load_p (rtx insn) |
9b6b54e2 NC |
6537 | { |
6538 | rtx body, lhs, rhs;; | |
6539 | ||
f0375c66 NC |
6540 | if (insn == NULL_RTX || GET_CODE (insn) != INSN) |
6541 | return false; | |
9b6b54e2 NC |
6542 | |
6543 | body = PATTERN (insn); | |
6544 | ||
6545 | if (GET_CODE (body) != SET) | |
f0375c66 | 6546 | return false; |
9b6b54e2 NC |
6547 | |
6548 | lhs = XEXP (body, 0); | |
6549 | rhs = XEXP (body, 1); | |
6550 | ||
f0375c66 NC |
6551 | lhs = REG_OR_SUBREG_RTX (lhs); |
6552 | ||
6553 | /* If the destination is not a general purpose | |
6554 | register we do not have to worry. */ | |
6555 | if (GET_CODE (lhs) != REG | |
6556 | || REGNO_REG_CLASS (REGNO (lhs)) != GENERAL_REGS) | |
6557 | return false; | |
6558 | ||
6559 | /* As well as loads from memory we also have to react | |
6560 | to loads of invalid constants which will be turned | |
6561 | into loads from the minipool. */ | |
6562 | return (GET_CODE (rhs) == MEM | |
6563 | || GET_CODE (rhs) == SYMBOL_REF | |
6564 | || note_invalid_constants (insn, -1, false)); | |
9b6b54e2 NC |
6565 | } |
6566 | ||
f0375c66 | 6567 | /* Return TRUE if INSN is a Cirrus instruction. */ |
f0375c66 | 6568 | static bool |
e32bac5b | 6569 | arm_cirrus_insn_p (rtx insn) |
9b6b54e2 NC |
6570 | { |
6571 | enum attr_cirrus attr; | |
6572 | ||
e6d29d15 | 6573 | /* get_attr cannot accept USE or CLOBBER. */ |
9b6b54e2 NC |
6574 | if (!insn |
6575 | || GET_CODE (insn) != INSN | |
6576 | || GET_CODE (PATTERN (insn)) == USE | |
6577 | || GET_CODE (PATTERN (insn)) == CLOBBER) | |
6578 | return 0; | |
6579 | ||
6580 | attr = get_attr_cirrus (insn); | |
6581 | ||
f0375c66 | 6582 | return attr != CIRRUS_NOT; |
9b6b54e2 NC |
6583 | } |
6584 | ||
6585 | /* Cirrus reorg for invalid instruction combinations. */ | |
9b6b54e2 | 6586 | static void |
e32bac5b | 6587 | cirrus_reorg (rtx first) |
9b6b54e2 NC |
6588 | { |
6589 | enum attr_cirrus attr; | |
6590 | rtx body = PATTERN (first); | |
6591 | rtx t; | |
6592 | int nops; | |
6593 | ||
6594 | /* Any branch must be followed by 2 non Cirrus instructions. */ | |
6595 | if (GET_CODE (first) == JUMP_INSN && GET_CODE (body) != RETURN) | |
6596 | { | |
6597 | nops = 0; | |
6598 | t = next_nonnote_insn (first); | |
6599 | ||
f0375c66 | 6600 | if (arm_cirrus_insn_p (t)) |
9b6b54e2 NC |
6601 | ++ nops; |
6602 | ||
f0375c66 | 6603 | if (arm_cirrus_insn_p (next_nonnote_insn (t))) |
9b6b54e2 NC |
6604 | ++ nops; |
6605 | ||
6606 | while (nops --) | |
6607 | emit_insn_after (gen_nop (), first); | |
6608 | ||
6609 | return; | |
6610 | } | |
6611 | ||
6612 | /* (float (blah)) is in parallel with a clobber. */ | |
6613 | if (GET_CODE (body) == PARALLEL && XVECLEN (body, 0) > 0) | |
6614 | body = XVECEXP (body, 0, 0); | |
6615 | ||
6616 | if (GET_CODE (body) == SET) | |
6617 | { | |
6618 | rtx lhs = XEXP (body, 0), rhs = XEXP (body, 1); | |
6619 | ||
6620 | /* cfldrd, cfldr64, cfstrd, cfstr64 must | |
6621 | be followed by a non Cirrus insn. */ | |
6622 | if (get_attr_cirrus (first) == CIRRUS_DOUBLE) | |
6623 | { | |
f0375c66 | 6624 | if (arm_cirrus_insn_p (next_nonnote_insn (first))) |
9b6b54e2 NC |
6625 | emit_insn_after (gen_nop (), first); |
6626 | ||
6627 | return; | |
6628 | } | |
f0375c66 | 6629 | else if (arm_memory_load_p (first)) |
9b6b54e2 NC |
6630 | { |
6631 | unsigned int arm_regno; | |
6632 | ||
6633 | /* Any ldr/cfmvdlr, ldr/cfmvdhr, ldr/cfmvsr, ldr/cfmv64lr, | |
6634 | ldr/cfmv64hr combination where the Rd field is the same | |
6635 | in both instructions must be split with a non Cirrus | |
6636 | insn. Example: | |
6637 | ||
6638 | ldr r0, blah | |
6639 | nop | |
6640 | cfmvsr mvf0, r0. */ | |
6641 | ||
6642 | /* Get Arm register number for ldr insn. */ | |
6643 | if (GET_CODE (lhs) == REG) | |
6644 | arm_regno = REGNO (lhs); | |
9b6b54e2 | 6645 | else |
e6d29d15 NS |
6646 | { |
6647 | gcc_assert (GET_CODE (rhs) == REG); | |
6648 | arm_regno = REGNO (rhs); | |
6649 | } | |
9b6b54e2 NC |
6650 | |
6651 | /* Next insn. */ | |
6652 | first = next_nonnote_insn (first); | |
6653 | ||
f0375c66 | 6654 | if (! arm_cirrus_insn_p (first)) |
9b6b54e2 NC |
6655 | return; |
6656 | ||
6657 | body = PATTERN (first); | |
6658 | ||
6659 | /* (float (blah)) is in parallel with a clobber. */ | |
6660 | if (GET_CODE (body) == PARALLEL && XVECLEN (body, 0)) | |
6661 | body = XVECEXP (body, 0, 0); | |
6662 | ||
6663 | if (GET_CODE (body) == FLOAT) | |
6664 | body = XEXP (body, 0); | |
6665 | ||
6666 | if (get_attr_cirrus (first) == CIRRUS_MOVE | |
6667 | && GET_CODE (XEXP (body, 1)) == REG | |
6668 | && arm_regno == REGNO (XEXP (body, 1))) | |
6669 | emit_insn_after (gen_nop (), first); | |
6670 | ||
6671 | return; | |
6672 | } | |
6673 | } | |
6674 | ||
e6d29d15 | 6675 | /* get_attr cannot accept USE or CLOBBER. */ |
9b6b54e2 NC |
6676 | if (!first |
6677 | || GET_CODE (first) != INSN | |
6678 | || GET_CODE (PATTERN (first)) == USE | |
6679 | || GET_CODE (PATTERN (first)) == CLOBBER) | |
6680 | return; | |
6681 | ||
6682 | attr = get_attr_cirrus (first); | |
6683 | ||
6684 | /* Any coprocessor compare instruction (cfcmps, cfcmpd, ...) | |
6685 | must be followed by a non-coprocessor instruction. */ | |
6686 | if (attr == CIRRUS_COMPARE) | |
6687 | { | |
6688 | nops = 0; | |
6689 | ||
6690 | t = next_nonnote_insn (first); | |
6691 | ||
f0375c66 | 6692 | if (arm_cirrus_insn_p (t)) |
9b6b54e2 NC |
6693 | ++ nops; |
6694 | ||
f0375c66 | 6695 | if (arm_cirrus_insn_p (next_nonnote_insn (t))) |
9b6b54e2 NC |
6696 | ++ nops; |
6697 | ||
6698 | while (nops --) | |
6699 | emit_insn_after (gen_nop (), first); | |
6700 | ||
6701 | return; | |
6702 | } | |
6703 | } | |
6704 | ||
2b835d68 RE |
6705 | /* Return TRUE if X references a SYMBOL_REF. */ |
6706 | int | |
e32bac5b | 6707 | symbol_mentioned_p (rtx x) |
2b835d68 | 6708 | { |
1d6e90ac NC |
6709 | const char * fmt; |
6710 | int i; | |
2b835d68 RE |
6711 | |
6712 | if (GET_CODE (x) == SYMBOL_REF) | |
6713 | return 1; | |
6714 | ||
d3585b76 DJ |
6715 | /* UNSPEC_TLS entries for a symbol include the SYMBOL_REF, but they |
6716 | are constant offsets, not symbols. */ | |
6717 | if (GET_CODE (x) == UNSPEC && XINT (x, 1) == UNSPEC_TLS) | |
6718 | return 0; | |
6719 | ||
2b835d68 | 6720 | fmt = GET_RTX_FORMAT (GET_CODE (x)); |
f676971a | 6721 | |
2b835d68 RE |
6722 | for (i = GET_RTX_LENGTH (GET_CODE (x)) - 1; i >= 0; i--) |
6723 | { | |
6724 | if (fmt[i] == 'E') | |
6725 | { | |
1d6e90ac | 6726 | int j; |
2b835d68 RE |
6727 | |
6728 | for (j = XVECLEN (x, i) - 1; j >= 0; j--) | |
6729 | if (symbol_mentioned_p (XVECEXP (x, i, j))) | |
6730 | return 1; | |
6731 | } | |
6732 | else if (fmt[i] == 'e' && symbol_mentioned_p (XEXP (x, i))) | |
6733 | return 1; | |
6734 | } | |
6735 | ||
6736 | return 0; | |
6737 | } | |
6738 | ||
6739 | /* Return TRUE if X references a LABEL_REF. */ | |
6740 | int | |
e32bac5b | 6741 | label_mentioned_p (rtx x) |
2b835d68 | 6742 | { |
1d6e90ac NC |
6743 | const char * fmt; |
6744 | int i; | |
2b835d68 RE |
6745 | |
6746 | if (GET_CODE (x) == LABEL_REF) | |
6747 | return 1; | |
6748 | ||
d3585b76 DJ |
6749 | /* UNSPEC_TLS entries for a symbol include a LABEL_REF for the referencing |
6750 | instruction, but they are constant offsets, not symbols. */ | |
6751 | if (GET_CODE (x) == UNSPEC && XINT (x, 1) == UNSPEC_TLS) | |
6752 | return 0; | |
6753 | ||
2b835d68 RE |
6754 | fmt = GET_RTX_FORMAT (GET_CODE (x)); |
6755 | for (i = GET_RTX_LENGTH (GET_CODE (x)) - 1; i >= 0; i--) | |
6756 | { | |
6757 | if (fmt[i] == 'E') | |
6758 | { | |
1d6e90ac | 6759 | int j; |
2b835d68 RE |
6760 | |
6761 | for (j = XVECLEN (x, i) - 1; j >= 0; j--) | |
6762 | if (label_mentioned_p (XVECEXP (x, i, j))) | |
6763 | return 1; | |
6764 | } | |
6765 | else if (fmt[i] == 'e' && label_mentioned_p (XEXP (x, i))) | |
6766 | return 1; | |
6767 | } | |
6768 | ||
6769 | return 0; | |
6770 | } | |
6771 | ||
d3585b76 DJ |
6772 | int |
6773 | tls_mentioned_p (rtx x) | |
6774 | { | |
6775 | switch (GET_CODE (x)) | |
6776 | { | |
6777 | case CONST: | |
6778 | return tls_mentioned_p (XEXP (x, 0)); | |
6779 | ||
6780 | case UNSPEC: | |
6781 | if (XINT (x, 1) == UNSPEC_TLS) | |
6782 | return 1; | |
6783 | ||
6784 | default: | |
6785 | return 0; | |
6786 | } | |
6787 | } | |
6788 | ||
6789 | /* Must not copy a SET whose source operand is PC-relative. */ | |
6790 | ||
6791 | static bool | |
6792 | arm_cannot_copy_insn_p (rtx insn) | |
6793 | { | |
6794 | rtx pat = PATTERN (insn); | |
6795 | ||
10ae7c7b | 6796 | if (GET_CODE (pat) == SET) |
d3585b76 | 6797 | { |
10ae7c7b | 6798 | rtx rhs = SET_SRC (pat); |
d3585b76 DJ |
6799 | |
6800 | if (GET_CODE (rhs) == UNSPEC | |
6801 | && XINT (rhs, 1) == UNSPEC_PIC_BASE) | |
6802 | return TRUE; | |
6803 | ||
6804 | if (GET_CODE (rhs) == MEM | |
6805 | && GET_CODE (XEXP (rhs, 0)) == UNSPEC | |
6806 | && XINT (XEXP (rhs, 0), 1) == UNSPEC_PIC_BASE) | |
6807 | return TRUE; | |
6808 | } | |
6809 | ||
6810 | return FALSE; | |
6811 | } | |
6812 | ||
ff9940b0 | 6813 | enum rtx_code |
e32bac5b | 6814 | minmax_code (rtx x) |
ff9940b0 RE |
6815 | { |
6816 | enum rtx_code code = GET_CODE (x); | |
6817 | ||
e6d29d15 NS |
6818 | switch (code) |
6819 | { | |
6820 | case SMAX: | |
6821 | return GE; | |
6822 | case SMIN: | |
6823 | return LE; | |
6824 | case UMIN: | |
6825 | return LEU; | |
6826 | case UMAX: | |
6827 | return GEU; | |
6828 | default: | |
6829 | gcc_unreachable (); | |
6830 | } | |
ff9940b0 RE |
6831 | } |
6832 | ||
6354dc9b | 6833 | /* Return 1 if memory locations are adjacent. */ |
f3bb6135 | 6834 | int |
e32bac5b | 6835 | adjacent_mem_locations (rtx a, rtx b) |
ff9940b0 | 6836 | { |
15b5c4c1 RE |
6837 | /* We don't guarantee to preserve the order of these memory refs. */ |
6838 | if (volatile_refs_p (a) || volatile_refs_p (b)) | |
6839 | return 0; | |
6840 | ||
ff9940b0 RE |
6841 | if ((GET_CODE (XEXP (a, 0)) == REG |
6842 | || (GET_CODE (XEXP (a, 0)) == PLUS | |
6843 | && GET_CODE (XEXP (XEXP (a, 0), 1)) == CONST_INT)) | |
6844 | && (GET_CODE (XEXP (b, 0)) == REG | |
6845 | || (GET_CODE (XEXP (b, 0)) == PLUS | |
6846 | && GET_CODE (XEXP (XEXP (b, 0), 1)) == CONST_INT))) | |
6847 | { | |
6555b6bd RE |
6848 | HOST_WIDE_INT val0 = 0, val1 = 0; |
6849 | rtx reg0, reg1; | |
6850 | int val_diff; | |
f676971a | 6851 | |
ff9940b0 RE |
6852 | if (GET_CODE (XEXP (a, 0)) == PLUS) |
6853 | { | |
6555b6bd | 6854 | reg0 = XEXP (XEXP (a, 0), 0); |
ff9940b0 RE |
6855 | val0 = INTVAL (XEXP (XEXP (a, 0), 1)); |
6856 | } | |
6857 | else | |
6555b6bd | 6858 | reg0 = XEXP (a, 0); |
1d6e90ac | 6859 | |
ff9940b0 RE |
6860 | if (GET_CODE (XEXP (b, 0)) == PLUS) |
6861 | { | |
6555b6bd | 6862 | reg1 = XEXP (XEXP (b, 0), 0); |
ff9940b0 RE |
6863 | val1 = INTVAL (XEXP (XEXP (b, 0), 1)); |
6864 | } | |
6865 | else | |
6555b6bd | 6866 | reg1 = XEXP (b, 0); |
1d6e90ac | 6867 | |
e32bac5b RE |
6868 | /* Don't accept any offset that will require multiple |
6869 | instructions to handle, since this would cause the | |
6870 | arith_adjacentmem pattern to output an overlong sequence. */ | |
c75a3ddc PB |
6871 | if (!const_ok_for_op (PLUS, val0) || !const_ok_for_op (PLUS, val1)) |
6872 | return 0; | |
f676971a | 6873 | |
6555b6bd RE |
6874 | /* Don't allow an eliminable register: register elimination can make |
6875 | the offset too large. */ | |
6876 | if (arm_eliminable_register (reg0)) | |
6877 | return 0; | |
6878 | ||
6879 | val_diff = val1 - val0; | |
15b5c4c1 RE |
6880 | |
6881 | if (arm_ld_sched) | |
6882 | { | |
6883 | /* If the target has load delay slots, then there's no benefit | |
6884 | to using an ldm instruction unless the offset is zero and | |
6885 | we are optimizing for size. */ | |
6886 | return (optimize_size && (REGNO (reg0) == REGNO (reg1)) | |
6887 | && (val0 == 0 || val1 == 0 || val0 == 4 || val1 == 4) | |
6888 | && (val_diff == 4 || val_diff == -4)); | |
6889 | } | |
6890 | ||
6555b6bd RE |
6891 | return ((REGNO (reg0) == REGNO (reg1)) |
6892 | && (val_diff == 4 || val_diff == -4)); | |
ff9940b0 | 6893 | } |
6555b6bd | 6894 | |
ff9940b0 RE |
6895 | return 0; |
6896 | } | |
6897 | ||
84ed5e79 | 6898 | int |
e32bac5b RE |
6899 | load_multiple_sequence (rtx *operands, int nops, int *regs, int *base, |
6900 | HOST_WIDE_INT *load_offset) | |
84ed5e79 RE |
6901 | { |
6902 | int unsorted_regs[4]; | |
6903 | HOST_WIDE_INT unsorted_offsets[4]; | |
6904 | int order[4]; | |
ad076f4e | 6905 | int base_reg = -1; |
84ed5e79 RE |
6906 | int i; |
6907 | ||
1d6e90ac NC |
6908 | /* Can only handle 2, 3, or 4 insns at present, |
6909 | though could be easily extended if required. */ | |
e6d29d15 | 6910 | gcc_assert (nops >= 2 && nops <= 4); |
84ed5e79 RE |
6911 | |
6912 | /* Loop over the operands and check that the memory references are | |
112cdef5 | 6913 | suitable (i.e. immediate offsets from the same base register). At |
84ed5e79 RE |
6914 | the same time, extract the target register, and the memory |
6915 | offsets. */ | |
6916 | for (i = 0; i < nops; i++) | |
6917 | { | |
6918 | rtx reg; | |
6919 | rtx offset; | |
6920 | ||
56636818 JL |
6921 | /* Convert a subreg of a mem into the mem itself. */ |
6922 | if (GET_CODE (operands[nops + i]) == SUBREG) | |
4e26a7af | 6923 | operands[nops + i] = alter_subreg (operands + (nops + i)); |
56636818 | 6924 | |
e6d29d15 | 6925 | gcc_assert (GET_CODE (operands[nops + i]) == MEM); |
84ed5e79 RE |
6926 | |
6927 | /* Don't reorder volatile memory references; it doesn't seem worth | |
6928 | looking for the case where the order is ok anyway. */ | |
6929 | if (MEM_VOLATILE_P (operands[nops + i])) | |
6930 | return 0; | |
6931 | ||
6932 | offset = const0_rtx; | |
6933 | ||
6934 | if ((GET_CODE (reg = XEXP (operands[nops + i], 0)) == REG | |
6935 | || (GET_CODE (reg) == SUBREG | |
6936 | && GET_CODE (reg = SUBREG_REG (reg)) == REG)) | |
6937 | || (GET_CODE (XEXP (operands[nops + i], 0)) == PLUS | |
6938 | && ((GET_CODE (reg = XEXP (XEXP (operands[nops + i], 0), 0)) | |
6939 | == REG) | |
6940 | || (GET_CODE (reg) == SUBREG | |
6941 | && GET_CODE (reg = SUBREG_REG (reg)) == REG)) | |
6942 | && (GET_CODE (offset = XEXP (XEXP (operands[nops + i], 0), 1)) | |
6943 | == CONST_INT))) | |
6944 | { | |
6945 | if (i == 0) | |
6946 | { | |
d5b7b3ae | 6947 | base_reg = REGNO (reg); |
84ed5e79 RE |
6948 | unsorted_regs[0] = (GET_CODE (operands[i]) == REG |
6949 | ? REGNO (operands[i]) | |
6950 | : REGNO (SUBREG_REG (operands[i]))); | |
6951 | order[0] = 0; | |
6952 | } | |
f676971a | 6953 | else |
84ed5e79 | 6954 | { |
6354dc9b | 6955 | if (base_reg != (int) REGNO (reg)) |
84ed5e79 RE |
6956 | /* Not addressed from the same base register. */ |
6957 | return 0; | |
6958 | ||
6959 | unsorted_regs[i] = (GET_CODE (operands[i]) == REG | |
6960 | ? REGNO (operands[i]) | |
6961 | : REGNO (SUBREG_REG (operands[i]))); | |
6962 | if (unsorted_regs[i] < unsorted_regs[order[0]]) | |
6963 | order[0] = i; | |
6964 | } | |
6965 | ||
6966 | /* If it isn't an integer register, or if it overwrites the | |
6967 | base register but isn't the last insn in the list, then | |
6968 | we can't do this. */ | |
6969 | if (unsorted_regs[i] < 0 || unsorted_regs[i] > 14 | |
6970 | || (i != nops - 1 && unsorted_regs[i] == base_reg)) | |
6971 | return 0; | |
6972 | ||
6973 | unsorted_offsets[i] = INTVAL (offset); | |
6974 | } | |
6975 | else | |
6976 | /* Not a suitable memory address. */ | |
6977 | return 0; | |
6978 | } | |
6979 | ||
6980 | /* All the useful information has now been extracted from the | |
6981 | operands into unsorted_regs and unsorted_offsets; additionally, | |
6982 | order[0] has been set to the lowest numbered register in the | |
6983 | list. Sort the registers into order, and check that the memory | |
6984 | offsets are ascending and adjacent. */ | |
6985 | ||
6986 | for (i = 1; i < nops; i++) | |
6987 | { | |
6988 | int j; | |
6989 | ||
6990 | order[i] = order[i - 1]; | |
6991 | for (j = 0; j < nops; j++) | |
6992 | if (unsorted_regs[j] > unsorted_regs[order[i - 1]] | |
6993 | && (order[i] == order[i - 1] | |
6994 | || unsorted_regs[j] < unsorted_regs[order[i]])) | |
6995 | order[i] = j; | |
6996 | ||
6997 | /* Have we found a suitable register? if not, one must be used more | |
6998 | than once. */ | |
6999 | if (order[i] == order[i - 1]) | |
7000 | return 0; | |
7001 | ||
7002 | /* Is the memory address adjacent and ascending? */ | |
7003 | if (unsorted_offsets[order[i]] != unsorted_offsets[order[i - 1]] + 4) | |
7004 | return 0; | |
7005 | } | |
7006 | ||
7007 | if (base) | |
7008 | { | |
7009 | *base = base_reg; | |
7010 | ||
7011 | for (i = 0; i < nops; i++) | |
7012 | regs[i] = unsorted_regs[order[i]]; | |
7013 | ||
7014 | *load_offset = unsorted_offsets[order[0]]; | |
7015 | } | |
7016 | ||
7017 | if (unsorted_offsets[order[0]] == 0) | |
7018 | return 1; /* ldmia */ | |
7019 | ||
5b3e6663 | 7020 | if (TARGET_ARM && unsorted_offsets[order[0]] == 4) |
84ed5e79 RE |
7021 | return 2; /* ldmib */ |
7022 | ||
5b3e6663 | 7023 | if (TARGET_ARM && unsorted_offsets[order[nops - 1]] == 0) |
84ed5e79 RE |
7024 | return 3; /* ldmda */ |
7025 | ||
7026 | if (unsorted_offsets[order[nops - 1]] == -4) | |
7027 | return 4; /* ldmdb */ | |
7028 | ||
949d79eb RE |
7029 | /* For ARM8,9 & StrongARM, 2 ldr instructions are faster than an ldm |
7030 | if the offset isn't small enough. The reason 2 ldrs are faster | |
7031 | is because these ARMs are able to do more than one cache access | |
7032 | in a single cycle. The ARM9 and StrongARM have Harvard caches, | |
7033 | whilst the ARM8 has a double bandwidth cache. This means that | |
7034 | these cores can do both an instruction fetch and a data fetch in | |
7035 | a single cycle, so the trick of calculating the address into a | |
7036 | scratch register (one of the result regs) and then doing a load | |
7037 | multiple actually becomes slower (and no smaller in code size). | |
7038 | That is the transformation | |
f676971a | 7039 | |
6cc8c0b3 NC |
7040 | ldr rd1, [rbase + offset] |
7041 | ldr rd2, [rbase + offset + 4] | |
f676971a | 7042 | |
6cc8c0b3 | 7043 | to |
f676971a | 7044 | |
6cc8c0b3 NC |
7045 | add rd1, rbase, offset |
7046 | ldmia rd1, {rd1, rd2} | |
f676971a | 7047 | |
949d79eb RE |
7048 | produces worse code -- '3 cycles + any stalls on rd2' instead of |
7049 | '2 cycles + any stalls on rd2'. On ARMs with only one cache | |
7050 | access per cycle, the first sequence could never complete in less | |
7051 | than 6 cycles, whereas the ldm sequence would only take 5 and | |
7052 | would make better use of sequential accesses if not hitting the | |
7053 | cache. | |
7054 | ||
7055 | We cheat here and test 'arm_ld_sched' which we currently know to | |
7056 | only be true for the ARM8, ARM9 and StrongARM. If this ever | |
7057 | changes, then the test below needs to be reworked. */ | |
f5a1b0d2 | 7058 | if (nops == 2 && arm_ld_sched) |
b36ba79f RE |
7059 | return 0; |
7060 | ||
84ed5e79 RE |
7061 | /* Can't do it without setting up the offset, only do this if it takes |
7062 | no more than one insn. */ | |
f676971a | 7063 | return (const_ok_for_arm (unsorted_offsets[order[0]]) |
84ed5e79 RE |
7064 | || const_ok_for_arm (-unsorted_offsets[order[0]])) ? 5 : 0; |
7065 | } | |
7066 | ||
cd2b33d0 | 7067 | const char * |
e32bac5b | 7068 | emit_ldm_seq (rtx *operands, int nops) |
84ed5e79 RE |
7069 | { |
7070 | int regs[4]; | |
7071 | int base_reg; | |
7072 | HOST_WIDE_INT offset; | |
7073 | char buf[100]; | |
7074 | int i; | |
7075 | ||
7076 | switch (load_multiple_sequence (operands, nops, regs, &base_reg, &offset)) | |
7077 | { | |
7078 | case 1: | |
5b3e6663 | 7079 | strcpy (buf, "ldm%(ia%)\t"); |
84ed5e79 RE |
7080 | break; |
7081 | ||
7082 | case 2: | |
5b3e6663 | 7083 | strcpy (buf, "ldm%(ib%)\t"); |
84ed5e79 RE |
7084 | break; |
7085 | ||
7086 | case 3: | |
5b3e6663 | 7087 | strcpy (buf, "ldm%(da%)\t"); |
84ed5e79 RE |
7088 | break; |
7089 | ||
7090 | case 4: | |
5b3e6663 | 7091 | strcpy (buf, "ldm%(db%)\t"); |
84ed5e79 RE |
7092 | break; |
7093 | ||
7094 | case 5: | |
7095 | if (offset >= 0) | |
7096 | sprintf (buf, "add%%?\t%s%s, %s%s, #%ld", REGISTER_PREFIX, | |
7097 | reg_names[regs[0]], REGISTER_PREFIX, reg_names[base_reg], | |
7098 | (long) offset); | |
7099 | else | |
7100 | sprintf (buf, "sub%%?\t%s%s, %s%s, #%ld", REGISTER_PREFIX, | |
7101 | reg_names[regs[0]], REGISTER_PREFIX, reg_names[base_reg], | |
7102 | (long) -offset); | |
7103 | output_asm_insn (buf, operands); | |
7104 | base_reg = regs[0]; | |
5b3e6663 | 7105 | strcpy (buf, "ldm%(ia%)\t"); |
84ed5e79 RE |
7106 | break; |
7107 | ||
7108 | default: | |
e6d29d15 | 7109 | gcc_unreachable (); |
84ed5e79 RE |
7110 | } |
7111 | ||
f676971a | 7112 | sprintf (buf + strlen (buf), "%s%s, {%s%s", REGISTER_PREFIX, |
84ed5e79 RE |
7113 | reg_names[base_reg], REGISTER_PREFIX, reg_names[regs[0]]); |
7114 | ||
7115 | for (i = 1; i < nops; i++) | |
7116 | sprintf (buf + strlen (buf), ", %s%s", REGISTER_PREFIX, | |
7117 | reg_names[regs[i]]); | |
7118 | ||
7119 | strcat (buf, "}\t%@ phole ldm"); | |
7120 | ||
7121 | output_asm_insn (buf, operands); | |
7122 | return ""; | |
7123 | } | |
7124 | ||
7125 | int | |
e32bac5b RE |
7126 | store_multiple_sequence (rtx *operands, int nops, int *regs, int *base, |
7127 | HOST_WIDE_INT * load_offset) | |
84ed5e79 RE |
7128 | { |
7129 | int unsorted_regs[4]; | |
7130 | HOST_WIDE_INT unsorted_offsets[4]; | |
7131 | int order[4]; | |
ad076f4e | 7132 | int base_reg = -1; |
84ed5e79 RE |
7133 | int i; |
7134 | ||
7135 | /* Can only handle 2, 3, or 4 insns at present, though could be easily | |
7136 | extended if required. */ | |
e6d29d15 | 7137 | gcc_assert (nops >= 2 && nops <= 4); |
84ed5e79 RE |
7138 | |
7139 | /* Loop over the operands and check that the memory references are | |
112cdef5 | 7140 | suitable (i.e. immediate offsets from the same base register). At |
84ed5e79 RE |
7141 | the same time, extract the target register, and the memory |
7142 | offsets. */ | |
7143 | for (i = 0; i < nops; i++) | |
7144 | { | |
7145 | rtx reg; | |
7146 | rtx offset; | |
7147 | ||
56636818 JL |
7148 | /* Convert a subreg of a mem into the mem itself. */ |
7149 | if (GET_CODE (operands[nops + i]) == SUBREG) | |
4e26a7af | 7150 | operands[nops + i] = alter_subreg (operands + (nops + i)); |
56636818 | 7151 | |
e6d29d15 | 7152 | gcc_assert (GET_CODE (operands[nops + i]) == MEM); |
84ed5e79 RE |
7153 | |
7154 | /* Don't reorder volatile memory references; it doesn't seem worth | |
7155 | looking for the case where the order is ok anyway. */ | |
7156 | if (MEM_VOLATILE_P (operands[nops + i])) | |
7157 | return 0; | |
7158 | ||
7159 | offset = const0_rtx; | |
7160 | ||
7161 | if ((GET_CODE (reg = XEXP (operands[nops + i], 0)) == REG | |
7162 | || (GET_CODE (reg) == SUBREG | |
7163 | && GET_CODE (reg = SUBREG_REG (reg)) == REG)) | |
7164 | || (GET_CODE (XEXP (operands[nops + i], 0)) == PLUS | |
7165 | && ((GET_CODE (reg = XEXP (XEXP (operands[nops + i], 0), 0)) | |
7166 | == REG) | |
7167 | || (GET_CODE (reg) == SUBREG | |
7168 | && GET_CODE (reg = SUBREG_REG (reg)) == REG)) | |
7169 | && (GET_CODE (offset = XEXP (XEXP (operands[nops + i], 0), 1)) | |
7170 | == CONST_INT))) | |
7171 | { | |
7172 | if (i == 0) | |
7173 | { | |
62b10bbc | 7174 | base_reg = REGNO (reg); |
84ed5e79 RE |
7175 | unsorted_regs[0] = (GET_CODE (operands[i]) == REG |
7176 | ? REGNO (operands[i]) | |
7177 | : REGNO (SUBREG_REG (operands[i]))); | |
7178 | order[0] = 0; | |
7179 | } | |
f676971a | 7180 | else |
84ed5e79 | 7181 | { |
6354dc9b | 7182 | if (base_reg != (int) REGNO (reg)) |
84ed5e79 RE |
7183 | /* Not addressed from the same base register. */ |
7184 | return 0; | |
7185 | ||
7186 | unsorted_regs[i] = (GET_CODE (operands[i]) == REG | |
7187 | ? REGNO (operands[i]) | |
7188 | : REGNO (SUBREG_REG (operands[i]))); | |
7189 | if (unsorted_regs[i] < unsorted_regs[order[0]]) | |
7190 | order[0] = i; | |
7191 | } | |
7192 | ||
7193 | /* If it isn't an integer register, then we can't do this. */ | |
7194 | if (unsorted_regs[i] < 0 || unsorted_regs[i] > 14) | |
7195 | return 0; | |
7196 | ||
7197 | unsorted_offsets[i] = INTVAL (offset); | |
7198 | } | |
7199 | else | |
7200 | /* Not a suitable memory address. */ | |
7201 | return 0; | |
7202 | } | |
7203 | ||
7204 | /* All the useful information has now been extracted from the | |
7205 | operands into unsorted_regs and unsorted_offsets; additionally, | |
7206 | order[0] has been set to the lowest numbered register in the | |
7207 | list. Sort the registers into order, and check that the memory | |
7208 | offsets are ascending and adjacent. */ | |
7209 | ||
7210 | for (i = 1; i < nops; i++) | |
7211 | { | |
7212 | int j; | |
7213 | ||
7214 | order[i] = order[i - 1]; | |
7215 | for (j = 0; j < nops; j++) | |
7216 | if (unsorted_regs[j] > unsorted_regs[order[i - 1]] | |
7217 | && (order[i] == order[i - 1] | |
7218 | || unsorted_regs[j] < unsorted_regs[order[i]])) | |
7219 | order[i] = j; | |
7220 | ||
7221 | /* Have we found a suitable register? if not, one must be used more | |
7222 | than once. */ | |
7223 | if (order[i] == order[i - 1]) | |
7224 | return 0; | |
7225 | ||
7226 | /* Is the memory address adjacent and ascending? */ | |
7227 | if (unsorted_offsets[order[i]] != unsorted_offsets[order[i - 1]] + 4) | |
7228 | return 0; | |
7229 | } | |
7230 | ||
7231 | if (base) | |
7232 | { | |
7233 | *base = base_reg; | |
7234 | ||
7235 | for (i = 0; i < nops; i++) | |
7236 | regs[i] = unsorted_regs[order[i]]; | |
7237 | ||
7238 | *load_offset = unsorted_offsets[order[0]]; | |
7239 | } | |
7240 | ||
7241 | if (unsorted_offsets[order[0]] == 0) | |
7242 | return 1; /* stmia */ | |
7243 | ||
7244 | if (unsorted_offsets[order[0]] == 4) | |
7245 | return 2; /* stmib */ | |
7246 | ||
7247 | if (unsorted_offsets[order[nops - 1]] == 0) | |
7248 | return 3; /* stmda */ | |
7249 | ||
7250 | if (unsorted_offsets[order[nops - 1]] == -4) | |
7251 | return 4; /* stmdb */ | |
7252 | ||
7253 | return 0; | |
7254 | } | |
7255 | ||
cd2b33d0 | 7256 | const char * |
e32bac5b | 7257 | emit_stm_seq (rtx *operands, int nops) |
84ed5e79 RE |
7258 | { |
7259 | int regs[4]; | |
7260 | int base_reg; | |
7261 | HOST_WIDE_INT offset; | |
7262 | char buf[100]; | |
7263 | int i; | |
7264 | ||
7265 | switch (store_multiple_sequence (operands, nops, regs, &base_reg, &offset)) | |
7266 | { | |
7267 | case 1: | |
5b3e6663 | 7268 | strcpy (buf, "stm%(ia%)\t"); |
84ed5e79 RE |
7269 | break; |
7270 | ||
7271 | case 2: | |
5b3e6663 | 7272 | strcpy (buf, "stm%(ib%)\t"); |
84ed5e79 RE |
7273 | break; |
7274 | ||
7275 | case 3: | |
5b3e6663 | 7276 | strcpy (buf, "stm%(da%)\t"); |
84ed5e79 RE |
7277 | break; |
7278 | ||
7279 | case 4: | |
5b3e6663 | 7280 | strcpy (buf, "stm%(db%)\t"); |
84ed5e79 RE |
7281 | break; |
7282 | ||
7283 | default: | |
e6d29d15 | 7284 | gcc_unreachable (); |
84ed5e79 RE |
7285 | } |
7286 | ||
f676971a | 7287 | sprintf (buf + strlen (buf), "%s%s, {%s%s", REGISTER_PREFIX, |
84ed5e79 RE |
7288 | reg_names[base_reg], REGISTER_PREFIX, reg_names[regs[0]]); |
7289 | ||
7290 | for (i = 1; i < nops; i++) | |
7291 | sprintf (buf + strlen (buf), ", %s%s", REGISTER_PREFIX, | |
7292 | reg_names[regs[i]]); | |
7293 | ||
7294 | strcat (buf, "}\t%@ phole stm"); | |
7295 | ||
7296 | output_asm_insn (buf, operands); | |
7297 | return ""; | |
7298 | } | |
ff9940b0 | 7299 | \f |
6354dc9b | 7300 | /* Routines for use in generating RTL. */ |
1d6e90ac | 7301 | |
f3bb6135 | 7302 | rtx |
e32bac5b | 7303 | arm_gen_load_multiple (int base_regno, int count, rtx from, int up, |
50ed9cea | 7304 | int write_back, rtx basemem, HOST_WIDE_INT *offsetp) |
ff9940b0 | 7305 | { |
50ed9cea | 7306 | HOST_WIDE_INT offset = *offsetp; |
ff9940b0 RE |
7307 | int i = 0, j; |
7308 | rtx result; | |
7309 | int sign = up ? 1 : -1; | |
50ed9cea | 7310 | rtx mem, addr; |
ff9940b0 | 7311 | |
d19fb8e3 | 7312 | /* XScale has load-store double instructions, but they have stricter |
1e5f1716 | 7313 | alignment requirements than load-store multiple, so we cannot |
d19fb8e3 NC |
7314 | use them. |
7315 | ||
7316 | For XScale ldm requires 2 + NREGS cycles to complete and blocks | |
7317 | the pipeline until completion. | |
7318 | ||
7319 | NREGS CYCLES | |
7320 | 1 3 | |
7321 | 2 4 | |
7322 | 3 5 | |
7323 | 4 6 | |
7324 | ||
7325 | An ldr instruction takes 1-3 cycles, but does not block the | |
7326 | pipeline. | |
7327 | ||
7328 | NREGS CYCLES | |
7329 | 1 1-3 | |
7330 | 2 2-6 | |
7331 | 3 3-9 | |
7332 | 4 4-12 | |
7333 | ||
7334 | Best case ldr will always win. However, the more ldr instructions | |
7335 | we issue, the less likely we are to be able to schedule them well. | |
7336 | Using ldr instructions also increases code size. | |
7337 | ||
7338 | As a compromise, we use ldr for counts of 1 or 2 regs, and ldm | |
7339 | for counts of 3 or 4 regs. */ | |
4b3c2e48 | 7340 | if (arm_tune_xscale && count <= 2 && ! optimize_size) |
d19fb8e3 NC |
7341 | { |
7342 | rtx seq; | |
f676971a | 7343 | |
d19fb8e3 | 7344 | start_sequence (); |
f676971a | 7345 | |
d19fb8e3 NC |
7346 | for (i = 0; i < count; i++) |
7347 | { | |
50ed9cea RH |
7348 | addr = plus_constant (from, i * 4 * sign); |
7349 | mem = adjust_automodify_address (basemem, SImode, addr, offset); | |
d19fb8e3 | 7350 | emit_move_insn (gen_rtx_REG (SImode, base_regno + i), mem); |
50ed9cea | 7351 | offset += 4 * sign; |
d19fb8e3 NC |
7352 | } |
7353 | ||
7354 | if (write_back) | |
50ed9cea RH |
7355 | { |
7356 | emit_move_insn (from, plus_constant (from, count * 4 * sign)); | |
7357 | *offsetp = offset; | |
7358 | } | |
d19fb8e3 | 7359 | |
2f937369 | 7360 | seq = get_insns (); |
d19fb8e3 | 7361 | end_sequence (); |
f676971a | 7362 | |
d19fb8e3 NC |
7363 | return seq; |
7364 | } | |
7365 | ||
43cffd11 | 7366 | result = gen_rtx_PARALLEL (VOIDmode, |
41e3f998 | 7367 | rtvec_alloc (count + (write_back ? 1 : 0))); |
ff9940b0 | 7368 | if (write_back) |
f3bb6135 | 7369 | { |
ff9940b0 | 7370 | XVECEXP (result, 0, 0) |
d66437c5 | 7371 | = gen_rtx_SET (VOIDmode, from, plus_constant (from, count * 4 * sign)); |
ff9940b0 RE |
7372 | i = 1; |
7373 | count++; | |
f3bb6135 RE |
7374 | } |
7375 | ||
ff9940b0 | 7376 | for (j = 0; i < count; i++, j++) |
f3bb6135 | 7377 | { |
50ed9cea RH |
7378 | addr = plus_constant (from, j * 4 * sign); |
7379 | mem = adjust_automodify_address_nv (basemem, SImode, addr, offset); | |
43cffd11 RE |
7380 | XVECEXP (result, 0, i) |
7381 | = gen_rtx_SET (VOIDmode, gen_rtx_REG (SImode, base_regno + j), mem); | |
50ed9cea | 7382 | offset += 4 * sign; |
f3bb6135 RE |
7383 | } |
7384 | ||
50ed9cea RH |
7385 | if (write_back) |
7386 | *offsetp = offset; | |
7387 | ||
ff9940b0 RE |
7388 | return result; |
7389 | } | |
7390 | ||
f3bb6135 | 7391 | rtx |
e32bac5b | 7392 | arm_gen_store_multiple (int base_regno, int count, rtx to, int up, |
50ed9cea | 7393 | int write_back, rtx basemem, HOST_WIDE_INT *offsetp) |
ff9940b0 | 7394 | { |
50ed9cea | 7395 | HOST_WIDE_INT offset = *offsetp; |
ff9940b0 RE |
7396 | int i = 0, j; |
7397 | rtx result; | |
7398 | int sign = up ? 1 : -1; | |
50ed9cea | 7399 | rtx mem, addr; |
ff9940b0 | 7400 | |
d19fb8e3 NC |
7401 | /* See arm_gen_load_multiple for discussion of |
7402 | the pros/cons of ldm/stm usage for XScale. */ | |
4b3c2e48 | 7403 | if (arm_tune_xscale && count <= 2 && ! optimize_size) |
d19fb8e3 NC |
7404 | { |
7405 | rtx seq; | |
f676971a | 7406 | |
d19fb8e3 | 7407 | start_sequence (); |
f676971a | 7408 | |
d19fb8e3 NC |
7409 | for (i = 0; i < count; i++) |
7410 | { | |
50ed9cea RH |
7411 | addr = plus_constant (to, i * 4 * sign); |
7412 | mem = adjust_automodify_address (basemem, SImode, addr, offset); | |
d19fb8e3 | 7413 | emit_move_insn (mem, gen_rtx_REG (SImode, base_regno + i)); |
50ed9cea | 7414 | offset += 4 * sign; |
d19fb8e3 NC |
7415 | } |
7416 | ||
7417 | if (write_back) | |
50ed9cea RH |
7418 | { |
7419 | emit_move_insn (to, plus_constant (to, count * 4 * sign)); | |
7420 | *offsetp = offset; | |
7421 | } | |
d19fb8e3 | 7422 | |
2f937369 | 7423 | seq = get_insns (); |
d19fb8e3 | 7424 | end_sequence (); |
f676971a | 7425 | |
d19fb8e3 NC |
7426 | return seq; |
7427 | } | |
7428 | ||
43cffd11 | 7429 | result = gen_rtx_PARALLEL (VOIDmode, |
41e3f998 | 7430 | rtvec_alloc (count + (write_back ? 1 : 0))); |
ff9940b0 | 7431 | if (write_back) |
f3bb6135 | 7432 | { |
ff9940b0 | 7433 | XVECEXP (result, 0, 0) |
d66437c5 | 7434 | = gen_rtx_SET (VOIDmode, to, |
43cffd11 | 7435 | plus_constant (to, count * 4 * sign)); |
ff9940b0 RE |
7436 | i = 1; |
7437 | count++; | |
f3bb6135 RE |
7438 | } |
7439 | ||
ff9940b0 | 7440 | for (j = 0; i < count; i++, j++) |
f3bb6135 | 7441 | { |
50ed9cea RH |
7442 | addr = plus_constant (to, j * 4 * sign); |
7443 | mem = adjust_automodify_address_nv (basemem, SImode, addr, offset); | |
43cffd11 RE |
7444 | XVECEXP (result, 0, i) |
7445 | = gen_rtx_SET (VOIDmode, mem, gen_rtx_REG (SImode, base_regno + j)); | |
50ed9cea | 7446 | offset += 4 * sign; |
f3bb6135 RE |
7447 | } |
7448 | ||
50ed9cea RH |
7449 | if (write_back) |
7450 | *offsetp = offset; | |
7451 | ||
ff9940b0 RE |
7452 | return result; |
7453 | } | |
7454 | ||
880e2516 | 7455 | int |
70128ad9 | 7456 | arm_gen_movmemqi (rtx *operands) |
880e2516 RE |
7457 | { |
7458 | HOST_WIDE_INT in_words_to_go, out_words_to_go, last_bytes; | |
50ed9cea | 7459 | HOST_WIDE_INT srcoffset, dstoffset; |
ad076f4e | 7460 | int i; |
50ed9cea | 7461 | rtx src, dst, srcbase, dstbase; |
880e2516 | 7462 | rtx part_bytes_reg = NULL; |
56636818 | 7463 | rtx mem; |
880e2516 RE |
7464 | |
7465 | if (GET_CODE (operands[2]) != CONST_INT | |
7466 | || GET_CODE (operands[3]) != CONST_INT | |
7467 | || INTVAL (operands[2]) > 64 | |
7468 | || INTVAL (operands[3]) & 3) | |
7469 | return 0; | |
7470 | ||
50ed9cea RH |
7471 | dstbase = operands[0]; |
7472 | srcbase = operands[1]; | |
56636818 | 7473 | |
50ed9cea RH |
7474 | dst = copy_to_mode_reg (SImode, XEXP (dstbase, 0)); |
7475 | src = copy_to_mode_reg (SImode, XEXP (srcbase, 0)); | |
880e2516 | 7476 | |
e9d7b180 | 7477 | in_words_to_go = ARM_NUM_INTS (INTVAL (operands[2])); |
880e2516 RE |
7478 | out_words_to_go = INTVAL (operands[2]) / 4; |
7479 | last_bytes = INTVAL (operands[2]) & 3; | |
50ed9cea | 7480 | dstoffset = srcoffset = 0; |
880e2516 RE |
7481 | |
7482 | if (out_words_to_go != in_words_to_go && ((in_words_to_go - 1) & 3) != 0) | |
43cffd11 | 7483 | part_bytes_reg = gen_rtx_REG (SImode, (in_words_to_go - 1) & 3); |
880e2516 RE |
7484 | |
7485 | for (i = 0; in_words_to_go >= 2; i+=4) | |
7486 | { | |
bd9c7e23 | 7487 | if (in_words_to_go > 4) |
56636818 | 7488 | emit_insn (arm_gen_load_multiple (0, 4, src, TRUE, TRUE, |
50ed9cea | 7489 | srcbase, &srcoffset)); |
bd9c7e23 | 7490 | else |
f676971a | 7491 | emit_insn (arm_gen_load_multiple (0, in_words_to_go, src, TRUE, |
50ed9cea | 7492 | FALSE, srcbase, &srcoffset)); |
bd9c7e23 | 7493 | |
880e2516 RE |
7494 | if (out_words_to_go) |
7495 | { | |
bd9c7e23 | 7496 | if (out_words_to_go > 4) |
56636818 | 7497 | emit_insn (arm_gen_store_multiple (0, 4, dst, TRUE, TRUE, |
50ed9cea | 7498 | dstbase, &dstoffset)); |
bd9c7e23 RE |
7499 | else if (out_words_to_go != 1) |
7500 | emit_insn (arm_gen_store_multiple (0, out_words_to_go, | |
f676971a | 7501 | dst, TRUE, |
bd9c7e23 | 7502 | (last_bytes == 0 |
56636818 | 7503 | ? FALSE : TRUE), |
50ed9cea | 7504 | dstbase, &dstoffset)); |
880e2516 RE |
7505 | else |
7506 | { | |
50ed9cea | 7507 | mem = adjust_automodify_address (dstbase, SImode, dst, dstoffset); |
43cffd11 | 7508 | emit_move_insn (mem, gen_rtx_REG (SImode, 0)); |
bd9c7e23 | 7509 | if (last_bytes != 0) |
50ed9cea RH |
7510 | { |
7511 | emit_insn (gen_addsi3 (dst, dst, GEN_INT (4))); | |
7512 | dstoffset += 4; | |
7513 | } | |
880e2516 RE |
7514 | } |
7515 | } | |
7516 | ||
7517 | in_words_to_go -= in_words_to_go < 4 ? in_words_to_go : 4; | |
7518 | out_words_to_go -= out_words_to_go < 4 ? out_words_to_go : 4; | |
7519 | } | |
7520 | ||
7521 | /* OUT_WORDS_TO_GO will be zero here if there are byte stores to do. */ | |
7522 | if (out_words_to_go) | |
62b10bbc NC |
7523 | { |
7524 | rtx sreg; | |
f676971a | 7525 | |
50ed9cea RH |
7526 | mem = adjust_automodify_address (srcbase, SImode, src, srcoffset); |
7527 | sreg = copy_to_reg (mem); | |
7528 | ||
7529 | mem = adjust_automodify_address (dstbase, SImode, dst, dstoffset); | |
62b10bbc | 7530 | emit_move_insn (mem, sreg); |
62b10bbc | 7531 | in_words_to_go--; |
f676971a | 7532 | |
e6d29d15 | 7533 | gcc_assert (!in_words_to_go); /* Sanity check */ |
62b10bbc | 7534 | } |
880e2516 RE |
7535 | |
7536 | if (in_words_to_go) | |
7537 | { | |
e6d29d15 | 7538 | gcc_assert (in_words_to_go > 0); |
880e2516 | 7539 | |
50ed9cea | 7540 | mem = adjust_automodify_address (srcbase, SImode, src, srcoffset); |
56636818 | 7541 | part_bytes_reg = copy_to_mode_reg (SImode, mem); |
880e2516 RE |
7542 | } |
7543 | ||
e6d29d15 | 7544 | gcc_assert (!last_bytes || part_bytes_reg); |
d5b7b3ae | 7545 | |
880e2516 RE |
7546 | if (BYTES_BIG_ENDIAN && last_bytes) |
7547 | { | |
7548 | rtx tmp = gen_reg_rtx (SImode); | |
7549 | ||
6354dc9b | 7550 | /* The bytes we want are in the top end of the word. */ |
bee06f3d RE |
7551 | emit_insn (gen_lshrsi3 (tmp, part_bytes_reg, |
7552 | GEN_INT (8 * (4 - last_bytes)))); | |
880e2516 | 7553 | part_bytes_reg = tmp; |
f676971a | 7554 | |
880e2516 RE |
7555 | while (last_bytes) |
7556 | { | |
50ed9cea RH |
7557 | mem = adjust_automodify_address (dstbase, QImode, |
7558 | plus_constant (dst, last_bytes - 1), | |
7559 | dstoffset + last_bytes - 1); | |
5d5603e2 BS |
7560 | emit_move_insn (mem, gen_lowpart (QImode, part_bytes_reg)); |
7561 | ||
880e2516 RE |
7562 | if (--last_bytes) |
7563 | { | |
7564 | tmp = gen_reg_rtx (SImode); | |
7565 | emit_insn (gen_lshrsi3 (tmp, part_bytes_reg, GEN_INT (8))); | |
7566 | part_bytes_reg = tmp; | |
7567 | } | |
7568 | } | |
f676971a | 7569 | |
880e2516 RE |
7570 | } |
7571 | else | |
7572 | { | |
d5b7b3ae | 7573 | if (last_bytes > 1) |
880e2516 | 7574 | { |
50ed9cea | 7575 | mem = adjust_automodify_address (dstbase, HImode, dst, dstoffset); |
5d5603e2 | 7576 | emit_move_insn (mem, gen_lowpart (HImode, part_bytes_reg)); |
d5b7b3ae RE |
7577 | last_bytes -= 2; |
7578 | if (last_bytes) | |
880e2516 RE |
7579 | { |
7580 | rtx tmp = gen_reg_rtx (SImode); | |
a556fd39 | 7581 | emit_insn (gen_addsi3 (dst, dst, const2_rtx)); |
d5b7b3ae | 7582 | emit_insn (gen_lshrsi3 (tmp, part_bytes_reg, GEN_INT (16))); |
880e2516 | 7583 | part_bytes_reg = tmp; |
50ed9cea | 7584 | dstoffset += 2; |
880e2516 RE |
7585 | } |
7586 | } | |
f676971a | 7587 | |
d5b7b3ae RE |
7588 | if (last_bytes) |
7589 | { | |
50ed9cea | 7590 | mem = adjust_automodify_address (dstbase, QImode, dst, dstoffset); |
5d5603e2 | 7591 | emit_move_insn (mem, gen_lowpart (QImode, part_bytes_reg)); |
d5b7b3ae | 7592 | } |
880e2516 RE |
7593 | } |
7594 | ||
7595 | return 1; | |
7596 | } | |
7597 | ||
03f1640c RE |
7598 | /* Select a dominance comparison mode if possible for a test of the general |
7599 | form (OP (COND_OR (X) (Y)) (const_int 0)). We support three forms. | |
f676971a | 7600 | COND_OR == DOM_CC_X_AND_Y => (X && Y) |
03f1640c | 7601 | COND_OR == DOM_CC_NX_OR_Y => ((! X) || Y) |
f676971a | 7602 | COND_OR == DOM_CC_X_OR_Y => (X || Y) |
03f1640c | 7603 | In all cases OP will be either EQ or NE, but we don't need to know which |
f676971a | 7604 | here. If we are unable to support a dominance comparison we return |
03f1640c RE |
7605 | CC mode. This will then fail to match for the RTL expressions that |
7606 | generate this call. */ | |
03f1640c | 7607 | enum machine_mode |
e32bac5b | 7608 | arm_select_dominance_cc_mode (rtx x, rtx y, HOST_WIDE_INT cond_or) |
84ed5e79 RE |
7609 | { |
7610 | enum rtx_code cond1, cond2; | |
7611 | int swapped = 0; | |
7612 | ||
7613 | /* Currently we will probably get the wrong result if the individual | |
7614 | comparisons are not simple. This also ensures that it is safe to | |
956d6950 | 7615 | reverse a comparison if necessary. */ |
84ed5e79 RE |
7616 | if ((arm_select_cc_mode (cond1 = GET_CODE (x), XEXP (x, 0), XEXP (x, 1)) |
7617 | != CCmode) | |
7618 | || (arm_select_cc_mode (cond2 = GET_CODE (y), XEXP (y, 0), XEXP (y, 1)) | |
7619 | != CCmode)) | |
7620 | return CCmode; | |
7621 | ||
1646cf41 RE |
7622 | /* The if_then_else variant of this tests the second condition if the |
7623 | first passes, but is true if the first fails. Reverse the first | |
7624 | condition to get a true "inclusive-or" expression. */ | |
03f1640c | 7625 | if (cond_or == DOM_CC_NX_OR_Y) |
84ed5e79 RE |
7626 | cond1 = reverse_condition (cond1); |
7627 | ||
7628 | /* If the comparisons are not equal, and one doesn't dominate the other, | |
7629 | then we can't do this. */ | |
f676971a | 7630 | if (cond1 != cond2 |
5895f793 RE |
7631 | && !comparison_dominates_p (cond1, cond2) |
7632 | && (swapped = 1, !comparison_dominates_p (cond2, cond1))) | |
84ed5e79 RE |
7633 | return CCmode; |
7634 | ||
7635 | if (swapped) | |
7636 | { | |
7637 | enum rtx_code temp = cond1; | |
7638 | cond1 = cond2; | |
7639 | cond2 = temp; | |
7640 | } | |
7641 | ||
7642 | switch (cond1) | |
7643 | { | |
7644 | case EQ: | |
e6d29d15 | 7645 | if (cond_or == DOM_CC_X_AND_Y) |
84ed5e79 RE |
7646 | return CC_DEQmode; |
7647 | ||
7648 | switch (cond2) | |
7649 | { | |
e6d29d15 | 7650 | case EQ: return CC_DEQmode; |
84ed5e79 RE |
7651 | case LE: return CC_DLEmode; |
7652 | case LEU: return CC_DLEUmode; | |
7653 | case GE: return CC_DGEmode; | |
7654 | case GEU: return CC_DGEUmode; | |
e6d29d15 | 7655 | default: gcc_unreachable (); |
84ed5e79 RE |
7656 | } |
7657 | ||
84ed5e79 | 7658 | case LT: |
e6d29d15 | 7659 | if (cond_or == DOM_CC_X_AND_Y) |
84ed5e79 | 7660 | return CC_DLTmode; |
e0b92319 | 7661 | |
e6d29d15 NS |
7662 | switch (cond2) |
7663 | { | |
7664 | case LT: | |
7665 | return CC_DLTmode; | |
7666 | case LE: | |
7667 | return CC_DLEmode; | |
7668 | case NE: | |
7669 | return CC_DNEmode; | |
7670 | default: | |
7671 | gcc_unreachable (); | |
7672 | } | |
84ed5e79 RE |
7673 | |
7674 | case GT: | |
e6d29d15 | 7675 | if (cond_or == DOM_CC_X_AND_Y) |
84ed5e79 | 7676 | return CC_DGTmode; |
e6d29d15 NS |
7677 | |
7678 | switch (cond2) | |
7679 | { | |
7680 | case GT: | |
7681 | return CC_DGTmode; | |
7682 | case GE: | |
7683 | return CC_DGEmode; | |
7684 | case NE: | |
7685 | return CC_DNEmode; | |
7686 | default: | |
7687 | gcc_unreachable (); | |
7688 | } | |
f676971a | 7689 | |
84ed5e79 | 7690 | case LTU: |
e6d29d15 | 7691 | if (cond_or == DOM_CC_X_AND_Y) |
84ed5e79 | 7692 | return CC_DLTUmode; |
e6d29d15 NS |
7693 | |
7694 | switch (cond2) | |
7695 | { | |
7696 | case LTU: | |
7697 | return CC_DLTUmode; | |
7698 | case LEU: | |
7699 | return CC_DLEUmode; | |
7700 | case NE: | |
7701 | return CC_DNEmode; | |
7702 | default: | |
7703 | gcc_unreachable (); | |
7704 | } | |
84ed5e79 RE |
7705 | |
7706 | case GTU: | |
e6d29d15 | 7707 | if (cond_or == DOM_CC_X_AND_Y) |
84ed5e79 | 7708 | return CC_DGTUmode; |
e0b92319 | 7709 | |
e6d29d15 NS |
7710 | switch (cond2) |
7711 | { | |
7712 | case GTU: | |
7713 | return CC_DGTUmode; | |
7714 | case GEU: | |
7715 | return CC_DGEUmode; | |
7716 | case NE: | |
7717 | return CC_DNEmode; | |
7718 | default: | |
7719 | gcc_unreachable (); | |
7720 | } | |
84ed5e79 RE |
7721 | |
7722 | /* The remaining cases only occur when both comparisons are the | |
7723 | same. */ | |
7724 | case NE: | |
e6d29d15 | 7725 | gcc_assert (cond1 == cond2); |
84ed5e79 RE |
7726 | return CC_DNEmode; |
7727 | ||
7728 | case LE: | |
e6d29d15 | 7729 | gcc_assert (cond1 == cond2); |
84ed5e79 RE |
7730 | return CC_DLEmode; |
7731 | ||
7732 | case GE: | |
e6d29d15 | 7733 | gcc_assert (cond1 == cond2); |
84ed5e79 RE |
7734 | return CC_DGEmode; |
7735 | ||
7736 | case LEU: | |
e6d29d15 | 7737 | gcc_assert (cond1 == cond2); |
84ed5e79 RE |
7738 | return CC_DLEUmode; |
7739 | ||
7740 | case GEU: | |
e6d29d15 | 7741 | gcc_assert (cond1 == cond2); |
84ed5e79 | 7742 | return CC_DGEUmode; |
ad076f4e RE |
7743 | |
7744 | default: | |
e6d29d15 | 7745 | gcc_unreachable (); |
84ed5e79 | 7746 | } |
84ed5e79 RE |
7747 | } |
7748 | ||
7749 | enum machine_mode | |
e32bac5b | 7750 | arm_select_cc_mode (enum rtx_code op, rtx x, rtx y) |
84ed5e79 RE |
7751 | { |
7752 | /* All floating point compares return CCFP if it is an equality | |
7753 | comparison, and CCFPE otherwise. */ | |
7754 | if (GET_MODE_CLASS (GET_MODE (x)) == MODE_FLOAT) | |
e45b72c4 RE |
7755 | { |
7756 | switch (op) | |
7757 | { | |
7758 | case EQ: | |
7759 | case NE: | |
7760 | case UNORDERED: | |
7761 | case ORDERED: | |
7762 | case UNLT: | |
7763 | case UNLE: | |
7764 | case UNGT: | |
7765 | case UNGE: | |
7766 | case UNEQ: | |
7767 | case LTGT: | |
7768 | return CCFPmode; | |
7769 | ||
7770 | case LT: | |
7771 | case LE: | |
7772 | case GT: | |
7773 | case GE: | |
9b66ebb1 | 7774 | if (TARGET_HARD_FLOAT && TARGET_MAVERICK) |
9b6b54e2 | 7775 | return CCFPmode; |
e45b72c4 RE |
7776 | return CCFPEmode; |
7777 | ||
7778 | default: | |
e6d29d15 | 7779 | gcc_unreachable (); |
e45b72c4 RE |
7780 | } |
7781 | } | |
f676971a | 7782 | |
84ed5e79 RE |
7783 | /* A compare with a shifted operand. Because of canonicalization, the |
7784 | comparison will have to be swapped when we emit the assembler. */ | |
7785 | if (GET_MODE (y) == SImode && GET_CODE (y) == REG | |
7786 | && (GET_CODE (x) == ASHIFT || GET_CODE (x) == ASHIFTRT | |
7787 | || GET_CODE (x) == LSHIFTRT || GET_CODE (x) == ROTATE | |
7788 | || GET_CODE (x) == ROTATERT)) | |
7789 | return CC_SWPmode; | |
7790 | ||
04d8b819 RE |
7791 | /* This operation is performed swapped, but since we only rely on the Z |
7792 | flag we don't need an additional mode. */ | |
7793 | if (GET_MODE (y) == SImode && REG_P (y) | |
7794 | && GET_CODE (x) == NEG | |
7795 | && (op == EQ || op == NE)) | |
7796 | return CC_Zmode; | |
7797 | ||
f676971a | 7798 | /* This is a special case that is used by combine to allow a |
956d6950 | 7799 | comparison of a shifted byte load to be split into a zero-extend |
84ed5e79 | 7800 | followed by a comparison of the shifted integer (only valid for |
956d6950 | 7801 | equalities and unsigned inequalities). */ |
84ed5e79 RE |
7802 | if (GET_MODE (x) == SImode |
7803 | && GET_CODE (x) == ASHIFT | |
7804 | && GET_CODE (XEXP (x, 1)) == CONST_INT && INTVAL (XEXP (x, 1)) == 24 | |
7805 | && GET_CODE (XEXP (x, 0)) == SUBREG | |
7806 | && GET_CODE (SUBREG_REG (XEXP (x, 0))) == MEM | |
7807 | && GET_MODE (SUBREG_REG (XEXP (x, 0))) == QImode | |
7808 | && (op == EQ || op == NE | |
7809 | || op == GEU || op == GTU || op == LTU || op == LEU) | |
7810 | && GET_CODE (y) == CONST_INT) | |
7811 | return CC_Zmode; | |
7812 | ||
1646cf41 RE |
7813 | /* A construct for a conditional compare, if the false arm contains |
7814 | 0, then both conditions must be true, otherwise either condition | |
7815 | must be true. Not all conditions are possible, so CCmode is | |
7816 | returned if it can't be done. */ | |
7817 | if (GET_CODE (x) == IF_THEN_ELSE | |
7818 | && (XEXP (x, 2) == const0_rtx | |
7819 | || XEXP (x, 2) == const1_rtx) | |
ec8e098d PB |
7820 | && COMPARISON_P (XEXP (x, 0)) |
7821 | && COMPARISON_P (XEXP (x, 1))) | |
f676971a | 7822 | return arm_select_dominance_cc_mode (XEXP (x, 0), XEXP (x, 1), |
03f1640c | 7823 | INTVAL (XEXP (x, 2))); |
1646cf41 RE |
7824 | |
7825 | /* Alternate canonicalizations of the above. These are somewhat cleaner. */ | |
7826 | if (GET_CODE (x) == AND | |
ec8e098d PB |
7827 | && COMPARISON_P (XEXP (x, 0)) |
7828 | && COMPARISON_P (XEXP (x, 1))) | |
03f1640c RE |
7829 | return arm_select_dominance_cc_mode (XEXP (x, 0), XEXP (x, 1), |
7830 | DOM_CC_X_AND_Y); | |
1646cf41 RE |
7831 | |
7832 | if (GET_CODE (x) == IOR | |
ec8e098d PB |
7833 | && COMPARISON_P (XEXP (x, 0)) |
7834 | && COMPARISON_P (XEXP (x, 1))) | |
03f1640c RE |
7835 | return arm_select_dominance_cc_mode (XEXP (x, 0), XEXP (x, 1), |
7836 | DOM_CC_X_OR_Y); | |
1646cf41 | 7837 | |
defc0463 RE |
7838 | /* An operation (on Thumb) where we want to test for a single bit. |
7839 | This is done by shifting that bit up into the top bit of a | |
7840 | scratch register; we can then branch on the sign bit. */ | |
5b3e6663 | 7841 | if (TARGET_THUMB1 |
defc0463 RE |
7842 | && GET_MODE (x) == SImode |
7843 | && (op == EQ || op == NE) | |
f9fa4363 RE |
7844 | && GET_CODE (x) == ZERO_EXTRACT |
7845 | && XEXP (x, 1) == const1_rtx) | |
defc0463 RE |
7846 | return CC_Nmode; |
7847 | ||
84ed5e79 RE |
7848 | /* An operation that sets the condition codes as a side-effect, the |
7849 | V flag is not set correctly, so we can only use comparisons where | |
7850 | this doesn't matter. (For LT and GE we can use "mi" and "pl" | |
defc0463 | 7851 | instead.) */ |
5b3e6663 | 7852 | /* ??? Does the ZERO_EXTRACT case really apply to thumb2? */ |
84ed5e79 RE |
7853 | if (GET_MODE (x) == SImode |
7854 | && y == const0_rtx | |
7855 | && (op == EQ || op == NE || op == LT || op == GE) | |
7856 | && (GET_CODE (x) == PLUS || GET_CODE (x) == MINUS | |
7857 | || GET_CODE (x) == AND || GET_CODE (x) == IOR | |
7858 | || GET_CODE (x) == XOR || GET_CODE (x) == MULT | |
7859 | || GET_CODE (x) == NOT || GET_CODE (x) == NEG | |
7860 | || GET_CODE (x) == LSHIFTRT | |
7861 | || GET_CODE (x) == ASHIFT || GET_CODE (x) == ASHIFTRT | |
defc0463 | 7862 | || GET_CODE (x) == ROTATERT |
5b3e6663 | 7863 | || (TARGET_32BIT && GET_CODE (x) == ZERO_EXTRACT))) |
84ed5e79 RE |
7864 | return CC_NOOVmode; |
7865 | ||
84ed5e79 RE |
7866 | if (GET_MODE (x) == QImode && (op == EQ || op == NE)) |
7867 | return CC_Zmode; | |
7868 | ||
bd9c7e23 RE |
7869 | if (GET_MODE (x) == SImode && (op == LTU || op == GEU) |
7870 | && GET_CODE (x) == PLUS | |
7871 | && (rtx_equal_p (XEXP (x, 0), y) || rtx_equal_p (XEXP (x, 1), y))) | |
7872 | return CC_Cmode; | |
7873 | ||
84ed5e79 RE |
7874 | return CCmode; |
7875 | } | |
7876 | ||
ff9940b0 RE |
7877 | /* X and Y are two things to compare using CODE. Emit the compare insn and |
7878 | return the rtx for register 0 in the proper mode. FP means this is a | |
7879 | floating point compare: I don't think that it is needed on the arm. */ | |
ff9940b0 | 7880 | rtx |
e32bac5b | 7881 | arm_gen_compare_reg (enum rtx_code code, rtx x, rtx y) |
ff9940b0 RE |
7882 | { |
7883 | enum machine_mode mode = SELECT_CC_MODE (code, x, y); | |
d5b7b3ae | 7884 | rtx cc_reg = gen_rtx_REG (mode, CC_REGNUM); |
ff9940b0 | 7885 | |
d66437c5 | 7886 | emit_set_insn (cc_reg, gen_rtx_COMPARE (mode, x, y)); |
ff9940b0 RE |
7887 | |
7888 | return cc_reg; | |
7889 | } | |
7890 | ||
fcd53748 JT |
7891 | /* Generate a sequence of insns that will generate the correct return |
7892 | address mask depending on the physical architecture that the program | |
7893 | is running on. */ | |
fcd53748 | 7894 | rtx |
e32bac5b | 7895 | arm_gen_return_addr_mask (void) |
fcd53748 JT |
7896 | { |
7897 | rtx reg = gen_reg_rtx (Pmode); | |
7898 | ||
7899 | emit_insn (gen_return_addr_mask (reg)); | |
7900 | return reg; | |
7901 | } | |
7902 | ||
0a81f500 | 7903 | void |
e32bac5b | 7904 | arm_reload_in_hi (rtx *operands) |
0a81f500 | 7905 | { |
f9cc092a RE |
7906 | rtx ref = operands[1]; |
7907 | rtx base, scratch; | |
7908 | HOST_WIDE_INT offset = 0; | |
7909 | ||
7910 | if (GET_CODE (ref) == SUBREG) | |
7911 | { | |
ddef6bc7 | 7912 | offset = SUBREG_BYTE (ref); |
f9cc092a RE |
7913 | ref = SUBREG_REG (ref); |
7914 | } | |
7915 | ||
7916 | if (GET_CODE (ref) == REG) | |
7917 | { | |
7918 | /* We have a pseudo which has been spilt onto the stack; there | |
7919 | are two cases here: the first where there is a simple | |
7920 | stack-slot replacement and a second where the stack-slot is | |
7921 | out of range, or is used as a subreg. */ | |
7922 | if (reg_equiv_mem[REGNO (ref)]) | |
7923 | { | |
7924 | ref = reg_equiv_mem[REGNO (ref)]; | |
7925 | base = find_replacement (&XEXP (ref, 0)); | |
7926 | } | |
7927 | else | |
6354dc9b | 7928 | /* The slot is out of range, or was dressed up in a SUBREG. */ |
f9cc092a RE |
7929 | base = reg_equiv_address[REGNO (ref)]; |
7930 | } | |
7931 | else | |
7932 | base = find_replacement (&XEXP (ref, 0)); | |
0a81f500 | 7933 | |
e5e809f4 JL |
7934 | /* Handle the case where the address is too complex to be offset by 1. */ |
7935 | if (GET_CODE (base) == MINUS | |
7936 | || (GET_CODE (base) == PLUS && GET_CODE (XEXP (base, 1)) != CONST_INT)) | |
7937 | { | |
f9cc092a | 7938 | rtx base_plus = gen_rtx_REG (SImode, REGNO (operands[2]) + 1); |
e5e809f4 | 7939 | |
d66437c5 | 7940 | emit_set_insn (base_plus, base); |
e5e809f4 JL |
7941 | base = base_plus; |
7942 | } | |
f9cc092a RE |
7943 | else if (GET_CODE (base) == PLUS) |
7944 | { | |
6354dc9b | 7945 | /* The addend must be CONST_INT, or we would have dealt with it above. */ |
f9cc092a RE |
7946 | HOST_WIDE_INT hi, lo; |
7947 | ||
7948 | offset += INTVAL (XEXP (base, 1)); | |
7949 | base = XEXP (base, 0); | |
7950 | ||
6354dc9b | 7951 | /* Rework the address into a legal sequence of insns. */ |
f9cc092a RE |
7952 | /* Valid range for lo is -4095 -> 4095 */ |
7953 | lo = (offset >= 0 | |
7954 | ? (offset & 0xfff) | |
7955 | : -((-offset) & 0xfff)); | |
7956 | ||
7957 | /* Corner case, if lo is the max offset then we would be out of range | |
7958 | once we have added the additional 1 below, so bump the msb into the | |
7959 | pre-loading insn(s). */ | |
7960 | if (lo == 4095) | |
7961 | lo &= 0x7ff; | |
7962 | ||
30cf4896 KG |
7963 | hi = ((((offset - lo) & (HOST_WIDE_INT) 0xffffffff) |
7964 | ^ (HOST_WIDE_INT) 0x80000000) | |
7965 | - (HOST_WIDE_INT) 0x80000000); | |
f9cc092a | 7966 | |
e6d29d15 | 7967 | gcc_assert (hi + lo == offset); |
f9cc092a RE |
7968 | |
7969 | if (hi != 0) | |
7970 | { | |
7971 | rtx base_plus = gen_rtx_REG (SImode, REGNO (operands[2]) + 1); | |
7972 | ||
7973 | /* Get the base address; addsi3 knows how to handle constants | |
6354dc9b | 7974 | that require more than one insn. */ |
f9cc092a RE |
7975 | emit_insn (gen_addsi3 (base_plus, base, GEN_INT (hi))); |
7976 | base = base_plus; | |
7977 | offset = lo; | |
7978 | } | |
7979 | } | |
e5e809f4 | 7980 | |
3a1944a6 RE |
7981 | /* Operands[2] may overlap operands[0] (though it won't overlap |
7982 | operands[1]), that's why we asked for a DImode reg -- so we can | |
7983 | use the bit that does not overlap. */ | |
7984 | if (REGNO (operands[2]) == REGNO (operands[0])) | |
7985 | scratch = gen_rtx_REG (SImode, REGNO (operands[2]) + 1); | |
7986 | else | |
7987 | scratch = gen_rtx_REG (SImode, REGNO (operands[2])); | |
7988 | ||
f9cc092a RE |
7989 | emit_insn (gen_zero_extendqisi2 (scratch, |
7990 | gen_rtx_MEM (QImode, | |
7991 | plus_constant (base, | |
7992 | offset)))); | |
43cffd11 | 7993 | emit_insn (gen_zero_extendqisi2 (gen_rtx_SUBREG (SImode, operands[0], 0), |
f676971a | 7994 | gen_rtx_MEM (QImode, |
f9cc092a RE |
7995 | plus_constant (base, |
7996 | offset + 1)))); | |
5895f793 | 7997 | if (!BYTES_BIG_ENDIAN) |
d66437c5 RE |
7998 | emit_set_insn (gen_rtx_SUBREG (SImode, operands[0], 0), |
7999 | gen_rtx_IOR (SImode, | |
8000 | gen_rtx_ASHIFT | |
8001 | (SImode, | |
8002 | gen_rtx_SUBREG (SImode, operands[0], 0), | |
8003 | GEN_INT (8)), | |
8004 | scratch)); | |
0a81f500 | 8005 | else |
d66437c5 RE |
8006 | emit_set_insn (gen_rtx_SUBREG (SImode, operands[0], 0), |
8007 | gen_rtx_IOR (SImode, | |
8008 | gen_rtx_ASHIFT (SImode, scratch, | |
8009 | GEN_INT (8)), | |
8010 | gen_rtx_SUBREG (SImode, operands[0], 0))); | |
0a81f500 RE |
8011 | } |
8012 | ||
72ac76be | 8013 | /* Handle storing a half-word to memory during reload by synthesizing as two |
f9cc092a RE |
8014 | byte stores. Take care not to clobber the input values until after we |
8015 | have moved them somewhere safe. This code assumes that if the DImode | |
8016 | scratch in operands[2] overlaps either the input value or output address | |
8017 | in some way, then that value must die in this insn (we absolutely need | |
8018 | two scratch registers for some corner cases). */ | |
f3bb6135 | 8019 | void |
e32bac5b | 8020 | arm_reload_out_hi (rtx *operands) |
af48348a | 8021 | { |
f9cc092a RE |
8022 | rtx ref = operands[0]; |
8023 | rtx outval = operands[1]; | |
8024 | rtx base, scratch; | |
8025 | HOST_WIDE_INT offset = 0; | |
8026 | ||
8027 | if (GET_CODE (ref) == SUBREG) | |
8028 | { | |
ddef6bc7 | 8029 | offset = SUBREG_BYTE (ref); |
f9cc092a RE |
8030 | ref = SUBREG_REG (ref); |
8031 | } | |
8032 | ||
f9cc092a RE |
8033 | if (GET_CODE (ref) == REG) |
8034 | { | |
8035 | /* We have a pseudo which has been spilt onto the stack; there | |
8036 | are two cases here: the first where there is a simple | |
8037 | stack-slot replacement and a second where the stack-slot is | |
8038 | out of range, or is used as a subreg. */ | |
8039 | if (reg_equiv_mem[REGNO (ref)]) | |
8040 | { | |
8041 | ref = reg_equiv_mem[REGNO (ref)]; | |
8042 | base = find_replacement (&XEXP (ref, 0)); | |
8043 | } | |
8044 | else | |
6354dc9b | 8045 | /* The slot is out of range, or was dressed up in a SUBREG. */ |
f9cc092a RE |
8046 | base = reg_equiv_address[REGNO (ref)]; |
8047 | } | |
8048 | else | |
8049 | base = find_replacement (&XEXP (ref, 0)); | |
8050 | ||
8051 | scratch = gen_rtx_REG (SImode, REGNO (operands[2])); | |
8052 | ||
8053 | /* Handle the case where the address is too complex to be offset by 1. */ | |
8054 | if (GET_CODE (base) == MINUS | |
8055 | || (GET_CODE (base) == PLUS && GET_CODE (XEXP (base, 1)) != CONST_INT)) | |
8056 | { | |
8057 | rtx base_plus = gen_rtx_REG (SImode, REGNO (operands[2]) + 1); | |
8058 | ||
8059 | /* Be careful not to destroy OUTVAL. */ | |
8060 | if (reg_overlap_mentioned_p (base_plus, outval)) | |
8061 | { | |
8062 | /* Updating base_plus might destroy outval, see if we can | |
8063 | swap the scratch and base_plus. */ | |
5895f793 | 8064 | if (!reg_overlap_mentioned_p (scratch, outval)) |
f9cc092a RE |
8065 | { |
8066 | rtx tmp = scratch; | |
8067 | scratch = base_plus; | |
8068 | base_plus = tmp; | |
8069 | } | |
8070 | else | |
8071 | { | |
8072 | rtx scratch_hi = gen_rtx_REG (HImode, REGNO (operands[2])); | |
8073 | ||
8074 | /* Be conservative and copy OUTVAL into the scratch now, | |
8075 | this should only be necessary if outval is a subreg | |
8076 | of something larger than a word. */ | |
8077 | /* XXX Might this clobber base? I can't see how it can, | |
8078 | since scratch is known to overlap with OUTVAL, and | |
8079 | must be wider than a word. */ | |
8080 | emit_insn (gen_movhi (scratch_hi, outval)); | |
8081 | outval = scratch_hi; | |
8082 | } | |
8083 | } | |
8084 | ||
d66437c5 | 8085 | emit_set_insn (base_plus, base); |
f9cc092a RE |
8086 | base = base_plus; |
8087 | } | |
8088 | else if (GET_CODE (base) == PLUS) | |
8089 | { | |
6354dc9b | 8090 | /* The addend must be CONST_INT, or we would have dealt with it above. */ |
f9cc092a RE |
8091 | HOST_WIDE_INT hi, lo; |
8092 | ||
8093 | offset += INTVAL (XEXP (base, 1)); | |
8094 | base = XEXP (base, 0); | |
8095 | ||
6354dc9b | 8096 | /* Rework the address into a legal sequence of insns. */ |
f9cc092a RE |
8097 | /* Valid range for lo is -4095 -> 4095 */ |
8098 | lo = (offset >= 0 | |
8099 | ? (offset & 0xfff) | |
8100 | : -((-offset) & 0xfff)); | |
8101 | ||
8102 | /* Corner case, if lo is the max offset then we would be out of range | |
8103 | once we have added the additional 1 below, so bump the msb into the | |
8104 | pre-loading insn(s). */ | |
8105 | if (lo == 4095) | |
8106 | lo &= 0x7ff; | |
8107 | ||
30cf4896 KG |
8108 | hi = ((((offset - lo) & (HOST_WIDE_INT) 0xffffffff) |
8109 | ^ (HOST_WIDE_INT) 0x80000000) | |
8110 | - (HOST_WIDE_INT) 0x80000000); | |
f9cc092a | 8111 | |
e6d29d15 | 8112 | gcc_assert (hi + lo == offset); |
f9cc092a RE |
8113 | |
8114 | if (hi != 0) | |
8115 | { | |
8116 | rtx base_plus = gen_rtx_REG (SImode, REGNO (operands[2]) + 1); | |
8117 | ||
8118 | /* Be careful not to destroy OUTVAL. */ | |
8119 | if (reg_overlap_mentioned_p (base_plus, outval)) | |
8120 | { | |
8121 | /* Updating base_plus might destroy outval, see if we | |
8122 | can swap the scratch and base_plus. */ | |
5895f793 | 8123 | if (!reg_overlap_mentioned_p (scratch, outval)) |
f9cc092a RE |
8124 | { |
8125 | rtx tmp = scratch; | |
8126 | scratch = base_plus; | |
8127 | base_plus = tmp; | |
8128 | } | |
8129 | else | |
8130 | { | |
8131 | rtx scratch_hi = gen_rtx_REG (HImode, REGNO (operands[2])); | |
8132 | ||
8133 | /* Be conservative and copy outval into scratch now, | |
8134 | this should only be necessary if outval is a | |
8135 | subreg of something larger than a word. */ | |
8136 | /* XXX Might this clobber base? I can't see how it | |
8137 | can, since scratch is known to overlap with | |
8138 | outval. */ | |
8139 | emit_insn (gen_movhi (scratch_hi, outval)); | |
8140 | outval = scratch_hi; | |
8141 | } | |
8142 | } | |
8143 | ||
8144 | /* Get the base address; addsi3 knows how to handle constants | |
6354dc9b | 8145 | that require more than one insn. */ |
f9cc092a RE |
8146 | emit_insn (gen_addsi3 (base_plus, base, GEN_INT (hi))); |
8147 | base = base_plus; | |
8148 | offset = lo; | |
8149 | } | |
8150 | } | |
af48348a | 8151 | |
b5cc037f RE |
8152 | if (BYTES_BIG_ENDIAN) |
8153 | { | |
f676971a | 8154 | emit_insn (gen_movqi (gen_rtx_MEM (QImode, |
f9cc092a | 8155 | plus_constant (base, offset + 1)), |
5d5603e2 | 8156 | gen_lowpart (QImode, outval))); |
f9cc092a RE |
8157 | emit_insn (gen_lshrsi3 (scratch, |
8158 | gen_rtx_SUBREG (SImode, outval, 0), | |
b5cc037f | 8159 | GEN_INT (8))); |
f9cc092a | 8160 | emit_insn (gen_movqi (gen_rtx_MEM (QImode, plus_constant (base, offset)), |
5d5603e2 | 8161 | gen_lowpart (QImode, scratch))); |
b5cc037f RE |
8162 | } |
8163 | else | |
8164 | { | |
f9cc092a | 8165 | emit_insn (gen_movqi (gen_rtx_MEM (QImode, plus_constant (base, offset)), |
5d5603e2 | 8166 | gen_lowpart (QImode, outval))); |
f9cc092a RE |
8167 | emit_insn (gen_lshrsi3 (scratch, |
8168 | gen_rtx_SUBREG (SImode, outval, 0), | |
b5cc037f | 8169 | GEN_INT (8))); |
f9cc092a RE |
8170 | emit_insn (gen_movqi (gen_rtx_MEM (QImode, |
8171 | plus_constant (base, offset + 1)), | |
5d5603e2 | 8172 | gen_lowpart (QImode, scratch))); |
b5cc037f | 8173 | } |
af48348a | 8174 | } |
866af8a9 JB |
8175 | |
8176 | /* Return true if a type must be passed in memory. For AAPCS, small aggregates | |
8177 | (padded to the size of a word) should be passed in a register. */ | |
8178 | ||
8179 | static bool | |
586de218 | 8180 | arm_must_pass_in_stack (enum machine_mode mode, const_tree type) |
866af8a9 JB |
8181 | { |
8182 | if (TARGET_AAPCS_BASED) | |
8183 | return must_pass_in_stack_var_size (mode, type); | |
8184 | else | |
8185 | return must_pass_in_stack_var_size_or_pad (mode, type); | |
8186 | } | |
8187 | ||
8188 | ||
8189 | /* For use by FUNCTION_ARG_PADDING (MODE, TYPE). | |
8190 | Return true if an argument passed on the stack should be padded upwards, | |
5a29b385 PB |
8191 | i.e. if the least-significant byte has useful data. |
8192 | For legacy APCS ABIs we use the default. For AAPCS based ABIs small | |
8193 | aggregate types are placed in the lowest memory address. */ | |
866af8a9 JB |
8194 | |
8195 | bool | |
586de218 | 8196 | arm_pad_arg_upward (enum machine_mode mode, const_tree type) |
866af8a9 JB |
8197 | { |
8198 | if (!TARGET_AAPCS_BASED) | |
5a29b385 | 8199 | return DEFAULT_FUNCTION_ARG_PADDING(mode, type) == upward; |
866af8a9 JB |
8200 | |
8201 | if (type && BYTES_BIG_ENDIAN && INTEGRAL_TYPE_P (type)) | |
8202 | return false; | |
8203 | ||
8204 | return true; | |
8205 | } | |
8206 | ||
8207 | ||
8208 | /* Similarly, for use by BLOCK_REG_PADDING (MODE, TYPE, FIRST). | |
8209 | For non-AAPCS, return !BYTES_BIG_ENDIAN if the least significant | |
8210 | byte of the register has useful data, and return the opposite if the | |
8211 | most significant byte does. | |
8212 | For AAPCS, small aggregates and small complex types are always padded | |
8213 | upwards. */ | |
8214 | ||
8215 | bool | |
8216 | arm_pad_reg_upward (enum machine_mode mode ATTRIBUTE_UNUSED, | |
8217 | tree type, int first ATTRIBUTE_UNUSED) | |
8218 | { | |
8219 | if (TARGET_AAPCS_BASED | |
8220 | && BYTES_BIG_ENDIAN | |
8221 | && (AGGREGATE_TYPE_P (type) || TREE_CODE (type) == COMPLEX_TYPE) | |
8222 | && int_size_in_bytes (type) <= 4) | |
8223 | return true; | |
8224 | ||
8225 | /* Otherwise, use default padding. */ | |
8226 | return !BYTES_BIG_ENDIAN; | |
8227 | } | |
8228 | ||
2b835d68 | 8229 | \f |
d5b7b3ae RE |
8230 | /* Print a symbolic form of X to the debug file, F. */ |
8231 | static void | |
e32bac5b | 8232 | arm_print_value (FILE *f, rtx x) |
d5b7b3ae RE |
8233 | { |
8234 | switch (GET_CODE (x)) | |
8235 | { | |
8236 | case CONST_INT: | |
8237 | fprintf (f, HOST_WIDE_INT_PRINT_HEX, INTVAL (x)); | |
8238 | return; | |
8239 | ||
8240 | case CONST_DOUBLE: | |
8241 | fprintf (f, "<0x%lx,0x%lx>", (long)XWINT (x, 2), (long)XWINT (x, 3)); | |
8242 | return; | |
8243 | ||
5a9335ef NC |
8244 | case CONST_VECTOR: |
8245 | { | |
8246 | int i; | |
8247 | ||
8248 | fprintf (f, "<"); | |
8249 | for (i = 0; i < CONST_VECTOR_NUNITS (x); i++) | |
8250 | { | |
8251 | fprintf (f, HOST_WIDE_INT_PRINT_HEX, INTVAL (CONST_VECTOR_ELT (x, i))); | |
8252 | if (i < (CONST_VECTOR_NUNITS (x) - 1)) | |
8253 | fputc (',', f); | |
8254 | } | |
8255 | fprintf (f, ">"); | |
8256 | } | |
8257 | return; | |
8258 | ||
d5b7b3ae RE |
8259 | case CONST_STRING: |
8260 | fprintf (f, "\"%s\"", XSTR (x, 0)); | |
8261 | return; | |
8262 | ||
8263 | case SYMBOL_REF: | |
8264 | fprintf (f, "`%s'", XSTR (x, 0)); | |
8265 | return; | |
8266 | ||
8267 | case LABEL_REF: | |
8268 | fprintf (f, "L%d", INSN_UID (XEXP (x, 0))); | |
8269 | return; | |
8270 | ||
8271 | case CONST: | |
8272 | arm_print_value (f, XEXP (x, 0)); | |
8273 | return; | |
8274 | ||
8275 | case PLUS: | |
8276 | arm_print_value (f, XEXP (x, 0)); | |
8277 | fprintf (f, "+"); | |
8278 | arm_print_value (f, XEXP (x, 1)); | |
8279 | return; | |
8280 | ||
8281 | case PC: | |
8282 | fprintf (f, "pc"); | |
8283 | return; | |
8284 | ||
8285 | default: | |
8286 | fprintf (f, "????"); | |
8287 | return; | |
8288 | } | |
8289 | } | |
8290 | \f | |
2b835d68 | 8291 | /* Routines for manipulation of the constant pool. */ |
2b835d68 | 8292 | |
949d79eb RE |
8293 | /* Arm instructions cannot load a large constant directly into a |
8294 | register; they have to come from a pc relative load. The constant | |
8295 | must therefore be placed in the addressable range of the pc | |
8296 | relative load. Depending on the precise pc relative load | |
8297 | instruction the range is somewhere between 256 bytes and 4k. This | |
8298 | means that we often have to dump a constant inside a function, and | |
2b835d68 RE |
8299 | generate code to branch around it. |
8300 | ||
949d79eb RE |
8301 | It is important to minimize this, since the branches will slow |
8302 | things down and make the code larger. | |
2b835d68 | 8303 | |
949d79eb RE |
8304 | Normally we can hide the table after an existing unconditional |
8305 | branch so that there is no interruption of the flow, but in the | |
8306 | worst case the code looks like this: | |
2b835d68 RE |
8307 | |
8308 | ldr rn, L1 | |
949d79eb | 8309 | ... |
2b835d68 RE |
8310 | b L2 |
8311 | align | |
8312 | L1: .long value | |
8313 | L2: | |
949d79eb | 8314 | ... |
2b835d68 | 8315 | |
2b835d68 | 8316 | ldr rn, L3 |
949d79eb | 8317 | ... |
2b835d68 RE |
8318 | b L4 |
8319 | align | |
2b835d68 RE |
8320 | L3: .long value |
8321 | L4: | |
949d79eb RE |
8322 | ... |
8323 | ||
8324 | We fix this by performing a scan after scheduling, which notices | |
8325 | which instructions need to have their operands fetched from the | |
8326 | constant table and builds the table. | |
8327 | ||
8328 | The algorithm starts by building a table of all the constants that | |
8329 | need fixing up and all the natural barriers in the function (places | |
8330 | where a constant table can be dropped without breaking the flow). | |
8331 | For each fixup we note how far the pc-relative replacement will be | |
8332 | able to reach and the offset of the instruction into the function. | |
8333 | ||
8334 | Having built the table we then group the fixes together to form | |
8335 | tables that are as large as possible (subject to addressing | |
8336 | constraints) and emit each table of constants after the last | |
8337 | barrier that is within range of all the instructions in the group. | |
8338 | If a group does not contain a barrier, then we forcibly create one | |
8339 | by inserting a jump instruction into the flow. Once the table has | |
8340 | been inserted, the insns are then modified to reference the | |
8341 | relevant entry in the pool. | |
8342 | ||
6354dc9b | 8343 | Possible enhancements to the algorithm (not implemented) are: |
949d79eb | 8344 | |
d5b7b3ae | 8345 | 1) For some processors and object formats, there may be benefit in |
949d79eb RE |
8346 | aligning the pools to the start of cache lines; this alignment |
8347 | would need to be taken into account when calculating addressability | |
6354dc9b | 8348 | of a pool. */ |
2b835d68 | 8349 | |
d5b7b3ae RE |
8350 | /* These typedefs are located at the start of this file, so that |
8351 | they can be used in the prototypes there. This comment is to | |
8352 | remind readers of that fact so that the following structures | |
8353 | can be understood more easily. | |
8354 | ||
8355 | typedef struct minipool_node Mnode; | |
8356 | typedef struct minipool_fixup Mfix; */ | |
8357 | ||
8358 | struct minipool_node | |
8359 | { | |
8360 | /* Doubly linked chain of entries. */ | |
8361 | Mnode * next; | |
8362 | Mnode * prev; | |
8363 | /* The maximum offset into the code that this entry can be placed. While | |
8364 | pushing fixes for forward references, all entries are sorted in order | |
8365 | of increasing max_address. */ | |
8366 | HOST_WIDE_INT max_address; | |
5519a4f9 | 8367 | /* Similarly for an entry inserted for a backwards ref. */ |
d5b7b3ae RE |
8368 | HOST_WIDE_INT min_address; |
8369 | /* The number of fixes referencing this entry. This can become zero | |
8370 | if we "unpush" an entry. In this case we ignore the entry when we | |
8371 | come to emit the code. */ | |
8372 | int refcount; | |
8373 | /* The offset from the start of the minipool. */ | |
8374 | HOST_WIDE_INT offset; | |
8375 | /* The value in table. */ | |
8376 | rtx value; | |
8377 | /* The mode of value. */ | |
8378 | enum machine_mode mode; | |
5a9335ef NC |
8379 | /* The size of the value. With iWMMXt enabled |
8380 | sizes > 4 also imply an alignment of 8-bytes. */ | |
d5b7b3ae RE |
8381 | int fix_size; |
8382 | }; | |
8383 | ||
8384 | struct minipool_fixup | |
2b835d68 | 8385 | { |
d5b7b3ae RE |
8386 | Mfix * next; |
8387 | rtx insn; | |
8388 | HOST_WIDE_INT address; | |
8389 | rtx * loc; | |
8390 | enum machine_mode mode; | |
8391 | int fix_size; | |
8392 | rtx value; | |
8393 | Mnode * minipool; | |
8394 | HOST_WIDE_INT forwards; | |
8395 | HOST_WIDE_INT backwards; | |
8396 | }; | |
2b835d68 | 8397 | |
d5b7b3ae RE |
8398 | /* Fixes less than a word need padding out to a word boundary. */ |
8399 | #define MINIPOOL_FIX_SIZE(mode) \ | |
8400 | (GET_MODE_SIZE ((mode)) >= 4 ? GET_MODE_SIZE ((mode)) : 4) | |
2b835d68 | 8401 | |
d5b7b3ae RE |
8402 | static Mnode * minipool_vector_head; |
8403 | static Mnode * minipool_vector_tail; | |
8404 | static rtx minipool_vector_label; | |
34a9f549 | 8405 | static int minipool_pad; |
332072db | 8406 | |
d5b7b3ae RE |
8407 | /* The linked list of all minipool fixes required for this function. */ |
8408 | Mfix * minipool_fix_head; | |
8409 | Mfix * minipool_fix_tail; | |
8410 | /* The fix entry for the current minipool, once it has been placed. */ | |
8411 | Mfix * minipool_barrier; | |
8412 | ||
8413 | /* Determines if INSN is the start of a jump table. Returns the end | |
8414 | of the TABLE or NULL_RTX. */ | |
8415 | static rtx | |
e32bac5b | 8416 | is_jump_table (rtx insn) |
2b835d68 | 8417 | { |
d5b7b3ae | 8418 | rtx table; |
f676971a | 8419 | |
d5b7b3ae RE |
8420 | if (GET_CODE (insn) == JUMP_INSN |
8421 | && JUMP_LABEL (insn) != NULL | |
8422 | && ((table = next_real_insn (JUMP_LABEL (insn))) | |
8423 | == next_real_insn (insn)) | |
8424 | && table != NULL | |
8425 | && GET_CODE (table) == JUMP_INSN | |
8426 | && (GET_CODE (PATTERN (table)) == ADDR_VEC | |
8427 | || GET_CODE (PATTERN (table)) == ADDR_DIFF_VEC)) | |
8428 | return table; | |
8429 | ||
8430 | return NULL_RTX; | |
2b835d68 RE |
8431 | } |
8432 | ||
657d9449 RE |
8433 | #ifndef JUMP_TABLES_IN_TEXT_SECTION |
8434 | #define JUMP_TABLES_IN_TEXT_SECTION 0 | |
8435 | #endif | |
8436 | ||
d5b7b3ae | 8437 | static HOST_WIDE_INT |
e32bac5b | 8438 | get_jump_table_size (rtx insn) |
2b835d68 | 8439 | { |
657d9449 RE |
8440 | /* ADDR_VECs only take room if read-only data does into the text |
8441 | section. */ | |
d6b5193b | 8442 | if (JUMP_TABLES_IN_TEXT_SECTION || readonly_data_section == text_section) |
657d9449 RE |
8443 | { |
8444 | rtx body = PATTERN (insn); | |
8445 | int elt = GET_CODE (body) == ADDR_DIFF_VEC ? 1 : 0; | |
5b3e6663 PB |
8446 | HOST_WIDE_INT size; |
8447 | HOST_WIDE_INT modesize; | |
2b835d68 | 8448 | |
5b3e6663 PB |
8449 | modesize = GET_MODE_SIZE (GET_MODE (body)); |
8450 | size = modesize * XVECLEN (body, elt); | |
8451 | switch (modesize) | |
8452 | { | |
8453 | case 1: | |
88512ba0 | 8454 | /* Round up size of TBB table to a halfword boundary. */ |
5b3e6663 PB |
8455 | size = (size + 1) & ~(HOST_WIDE_INT)1; |
8456 | break; | |
8457 | case 2: | |
7a085dce | 8458 | /* No padding necessary for TBH. */ |
5b3e6663 PB |
8459 | break; |
8460 | case 4: | |
8461 | /* Add two bytes for alignment on Thumb. */ | |
8462 | if (TARGET_THUMB) | |
8463 | size += 2; | |
8464 | break; | |
8465 | default: | |
8466 | gcc_unreachable (); | |
8467 | } | |
8468 | return size; | |
657d9449 RE |
8469 | } |
8470 | ||
8471 | return 0; | |
d5b7b3ae | 8472 | } |
2b835d68 | 8473 | |
d5b7b3ae RE |
8474 | /* Move a minipool fix MP from its current location to before MAX_MP. |
8475 | If MAX_MP is NULL, then MP doesn't need moving, but the addressing | |
093354e0 | 8476 | constraints may need updating. */ |
d5b7b3ae | 8477 | static Mnode * |
e32bac5b RE |
8478 | move_minipool_fix_forward_ref (Mnode *mp, Mnode *max_mp, |
8479 | HOST_WIDE_INT max_address) | |
d5b7b3ae | 8480 | { |
e6d29d15 NS |
8481 | /* The code below assumes these are different. */ |
8482 | gcc_assert (mp != max_mp); | |
d5b7b3ae RE |
8483 | |
8484 | if (max_mp == NULL) | |
8485 | { | |
8486 | if (max_address < mp->max_address) | |
8487 | mp->max_address = max_address; | |
8488 | } | |
8489 | else | |
2b835d68 | 8490 | { |
d5b7b3ae RE |
8491 | if (max_address > max_mp->max_address - mp->fix_size) |
8492 | mp->max_address = max_mp->max_address - mp->fix_size; | |
8493 | else | |
8494 | mp->max_address = max_address; | |
2b835d68 | 8495 | |
d5b7b3ae RE |
8496 | /* Unlink MP from its current position. Since max_mp is non-null, |
8497 | mp->prev must be non-null. */ | |
8498 | mp->prev->next = mp->next; | |
8499 | if (mp->next != NULL) | |
8500 | mp->next->prev = mp->prev; | |
8501 | else | |
8502 | minipool_vector_tail = mp->prev; | |
2b835d68 | 8503 | |
d5b7b3ae RE |
8504 | /* Re-insert it before MAX_MP. */ |
8505 | mp->next = max_mp; | |
8506 | mp->prev = max_mp->prev; | |
8507 | max_mp->prev = mp; | |
f676971a | 8508 | |
d5b7b3ae RE |
8509 | if (mp->prev != NULL) |
8510 | mp->prev->next = mp; | |
8511 | else | |
8512 | minipool_vector_head = mp; | |
8513 | } | |
2b835d68 | 8514 | |
d5b7b3ae RE |
8515 | /* Save the new entry. */ |
8516 | max_mp = mp; | |
8517 | ||
d6a7951f | 8518 | /* Scan over the preceding entries and adjust their addresses as |
d5b7b3ae RE |
8519 | required. */ |
8520 | while (mp->prev != NULL | |
8521 | && mp->prev->max_address > mp->max_address - mp->prev->fix_size) | |
8522 | { | |
8523 | mp->prev->max_address = mp->max_address - mp->prev->fix_size; | |
8524 | mp = mp->prev; | |
2b835d68 RE |
8525 | } |
8526 | ||
d5b7b3ae | 8527 | return max_mp; |
2b835d68 RE |
8528 | } |
8529 | ||
d5b7b3ae RE |
8530 | /* Add a constant to the minipool for a forward reference. Returns the |
8531 | node added or NULL if the constant will not fit in this pool. */ | |
8532 | static Mnode * | |
e32bac5b | 8533 | add_minipool_forward_ref (Mfix *fix) |
d5b7b3ae RE |
8534 | { |
8535 | /* If set, max_mp is the first pool_entry that has a lower | |
8536 | constraint than the one we are trying to add. */ | |
8537 | Mnode * max_mp = NULL; | |
34a9f549 | 8538 | HOST_WIDE_INT max_address = fix->address + fix->forwards - minipool_pad; |
d5b7b3ae | 8539 | Mnode * mp; |
f676971a | 8540 | |
7a7017bc PB |
8541 | /* If the minipool starts before the end of FIX->INSN then this FIX |
8542 | can not be placed into the current pool. Furthermore, adding the | |
8543 | new constant pool entry may cause the pool to start FIX_SIZE bytes | |
8544 | earlier. */ | |
d5b7b3ae | 8545 | if (minipool_vector_head && |
7a7017bc PB |
8546 | (fix->address + get_attr_length (fix->insn) |
8547 | >= minipool_vector_head->max_address - fix->fix_size)) | |
d5b7b3ae | 8548 | return NULL; |
2b835d68 | 8549 | |
d5b7b3ae RE |
8550 | /* Scan the pool to see if a constant with the same value has |
8551 | already been added. While we are doing this, also note the | |
8552 | location where we must insert the constant if it doesn't already | |
8553 | exist. */ | |
8554 | for (mp = minipool_vector_head; mp != NULL; mp = mp->next) | |
8555 | { | |
8556 | if (GET_CODE (fix->value) == GET_CODE (mp->value) | |
8557 | && fix->mode == mp->mode | |
8558 | && (GET_CODE (fix->value) != CODE_LABEL | |
8559 | || (CODE_LABEL_NUMBER (fix->value) | |
8560 | == CODE_LABEL_NUMBER (mp->value))) | |
8561 | && rtx_equal_p (fix->value, mp->value)) | |
8562 | { | |
8563 | /* More than one fix references this entry. */ | |
8564 | mp->refcount++; | |
8565 | return move_minipool_fix_forward_ref (mp, max_mp, max_address); | |
8566 | } | |
8567 | ||
8568 | /* Note the insertion point if necessary. */ | |
8569 | if (max_mp == NULL | |
8570 | && mp->max_address > max_address) | |
8571 | max_mp = mp; | |
5a9335ef NC |
8572 | |
8573 | /* If we are inserting an 8-bytes aligned quantity and | |
8574 | we have not already found an insertion point, then | |
8575 | make sure that all such 8-byte aligned quantities are | |
8576 | placed at the start of the pool. */ | |
5848830f | 8577 | if (ARM_DOUBLEWORD_ALIGN |
5a9335ef | 8578 | && max_mp == NULL |
88f77cba JB |
8579 | && fix->fix_size >= 8 |
8580 | && mp->fix_size < 8) | |
5a9335ef NC |
8581 | { |
8582 | max_mp = mp; | |
8583 | max_address = mp->max_address; | |
8584 | } | |
d5b7b3ae RE |
8585 | } |
8586 | ||
8587 | /* The value is not currently in the minipool, so we need to create | |
8588 | a new entry for it. If MAX_MP is NULL, the entry will be put on | |
8589 | the end of the list since the placement is less constrained than | |
8590 | any existing entry. Otherwise, we insert the new fix before | |
6bc82793 | 8591 | MAX_MP and, if necessary, adjust the constraints on the other |
d5b7b3ae | 8592 | entries. */ |
5ed6ace5 | 8593 | mp = XNEW (Mnode); |
d5b7b3ae RE |
8594 | mp->fix_size = fix->fix_size; |
8595 | mp->mode = fix->mode; | |
8596 | mp->value = fix->value; | |
8597 | mp->refcount = 1; | |
8598 | /* Not yet required for a backwards ref. */ | |
8599 | mp->min_address = -65536; | |
8600 | ||
8601 | if (max_mp == NULL) | |
8602 | { | |
8603 | mp->max_address = max_address; | |
8604 | mp->next = NULL; | |
8605 | mp->prev = minipool_vector_tail; | |
8606 | ||
8607 | if (mp->prev == NULL) | |
8608 | { | |
8609 | minipool_vector_head = mp; | |
8610 | minipool_vector_label = gen_label_rtx (); | |
7551cbc7 | 8611 | } |
2b835d68 | 8612 | else |
d5b7b3ae | 8613 | mp->prev->next = mp; |
2b835d68 | 8614 | |
d5b7b3ae RE |
8615 | minipool_vector_tail = mp; |
8616 | } | |
8617 | else | |
8618 | { | |
8619 | if (max_address > max_mp->max_address - mp->fix_size) | |
8620 | mp->max_address = max_mp->max_address - mp->fix_size; | |
8621 | else | |
8622 | mp->max_address = max_address; | |
8623 | ||
8624 | mp->next = max_mp; | |
8625 | mp->prev = max_mp->prev; | |
8626 | max_mp->prev = mp; | |
8627 | if (mp->prev != NULL) | |
8628 | mp->prev->next = mp; | |
8629 | else | |
8630 | minipool_vector_head = mp; | |
8631 | } | |
8632 | ||
8633 | /* Save the new entry. */ | |
8634 | max_mp = mp; | |
8635 | ||
d6a7951f | 8636 | /* Scan over the preceding entries and adjust their addresses as |
d5b7b3ae RE |
8637 | required. */ |
8638 | while (mp->prev != NULL | |
8639 | && mp->prev->max_address > mp->max_address - mp->prev->fix_size) | |
8640 | { | |
8641 | mp->prev->max_address = mp->max_address - mp->prev->fix_size; | |
8642 | mp = mp->prev; | |
2b835d68 RE |
8643 | } |
8644 | ||
d5b7b3ae RE |
8645 | return max_mp; |
8646 | } | |
8647 | ||
8648 | static Mnode * | |
e32bac5b RE |
8649 | move_minipool_fix_backward_ref (Mnode *mp, Mnode *min_mp, |
8650 | HOST_WIDE_INT min_address) | |
d5b7b3ae RE |
8651 | { |
8652 | HOST_WIDE_INT offset; | |
8653 | ||
e6d29d15 NS |
8654 | /* The code below assumes these are different. */ |
8655 | gcc_assert (mp != min_mp); | |
d5b7b3ae RE |
8656 | |
8657 | if (min_mp == NULL) | |
2b835d68 | 8658 | { |
d5b7b3ae RE |
8659 | if (min_address > mp->min_address) |
8660 | mp->min_address = min_address; | |
8661 | } | |
8662 | else | |
8663 | { | |
8664 | /* We will adjust this below if it is too loose. */ | |
8665 | mp->min_address = min_address; | |
8666 | ||
8667 | /* Unlink MP from its current position. Since min_mp is non-null, | |
8668 | mp->next must be non-null. */ | |
8669 | mp->next->prev = mp->prev; | |
8670 | if (mp->prev != NULL) | |
8671 | mp->prev->next = mp->next; | |
8672 | else | |
8673 | minipool_vector_head = mp->next; | |
8674 | ||
8675 | /* Reinsert it after MIN_MP. */ | |
8676 | mp->prev = min_mp; | |
8677 | mp->next = min_mp->next; | |
8678 | min_mp->next = mp; | |
8679 | if (mp->next != NULL) | |
8680 | mp->next->prev = mp; | |
2b835d68 | 8681 | else |
d5b7b3ae RE |
8682 | minipool_vector_tail = mp; |
8683 | } | |
8684 | ||
8685 | min_mp = mp; | |
8686 | ||
8687 | offset = 0; | |
8688 | for (mp = minipool_vector_head; mp != NULL; mp = mp->next) | |
8689 | { | |
8690 | mp->offset = offset; | |
8691 | if (mp->refcount > 0) | |
8692 | offset += mp->fix_size; | |
8693 | ||
8694 | if (mp->next && mp->next->min_address < mp->min_address + mp->fix_size) | |
8695 | mp->next->min_address = mp->min_address + mp->fix_size; | |
8696 | } | |
8697 | ||
8698 | return min_mp; | |
f676971a | 8699 | } |
d5b7b3ae RE |
8700 | |
8701 | /* Add a constant to the minipool for a backward reference. Returns the | |
f676971a | 8702 | node added or NULL if the constant will not fit in this pool. |
d5b7b3ae RE |
8703 | |
8704 | Note that the code for insertion for a backwards reference can be | |
8705 | somewhat confusing because the calculated offsets for each fix do | |
8706 | not take into account the size of the pool (which is still under | |
8707 | construction. */ | |
8708 | static Mnode * | |
e32bac5b | 8709 | add_minipool_backward_ref (Mfix *fix) |
d5b7b3ae RE |
8710 | { |
8711 | /* If set, min_mp is the last pool_entry that has a lower constraint | |
8712 | than the one we are trying to add. */ | |
e32bac5b | 8713 | Mnode *min_mp = NULL; |
d5b7b3ae RE |
8714 | /* This can be negative, since it is only a constraint. */ |
8715 | HOST_WIDE_INT min_address = fix->address - fix->backwards; | |
e32bac5b | 8716 | Mnode *mp; |
d5b7b3ae RE |
8717 | |
8718 | /* If we can't reach the current pool from this insn, or if we can't | |
8719 | insert this entry at the end of the pool without pushing other | |
8720 | fixes out of range, then we don't try. This ensures that we | |
8721 | can't fail later on. */ | |
8722 | if (min_address >= minipool_barrier->address | |
8723 | || (minipool_vector_tail->min_address + fix->fix_size | |
8724 | >= minipool_barrier->address)) | |
8725 | return NULL; | |
8726 | ||
8727 | /* Scan the pool to see if a constant with the same value has | |
8728 | already been added. While we are doing this, also note the | |
8729 | location where we must insert the constant if it doesn't already | |
8730 | exist. */ | |
8731 | for (mp = minipool_vector_tail; mp != NULL; mp = mp->prev) | |
8732 | { | |
8733 | if (GET_CODE (fix->value) == GET_CODE (mp->value) | |
8734 | && fix->mode == mp->mode | |
8735 | && (GET_CODE (fix->value) != CODE_LABEL | |
8736 | || (CODE_LABEL_NUMBER (fix->value) | |
8737 | == CODE_LABEL_NUMBER (mp->value))) | |
8738 | && rtx_equal_p (fix->value, mp->value) | |
8739 | /* Check that there is enough slack to move this entry to the | |
8740 | end of the table (this is conservative). */ | |
f676971a EC |
8741 | && (mp->max_address |
8742 | > (minipool_barrier->address | |
d5b7b3ae RE |
8743 | + minipool_vector_tail->offset |
8744 | + minipool_vector_tail->fix_size))) | |
8745 | { | |
8746 | mp->refcount++; | |
8747 | return move_minipool_fix_backward_ref (mp, min_mp, min_address); | |
8748 | } | |
8749 | ||
8750 | if (min_mp != NULL) | |
8751 | mp->min_address += fix->fix_size; | |
8752 | else | |
8753 | { | |
8754 | /* Note the insertion point if necessary. */ | |
8755 | if (mp->min_address < min_address) | |
5a9335ef NC |
8756 | { |
8757 | /* For now, we do not allow the insertion of 8-byte alignment | |
8758 | requiring nodes anywhere but at the start of the pool. */ | |
5848830f | 8759 | if (ARM_DOUBLEWORD_ALIGN |
88f77cba | 8760 | && fix->fix_size >= 8 && mp->fix_size < 8) |
5a9335ef NC |
8761 | return NULL; |
8762 | else | |
8763 | min_mp = mp; | |
8764 | } | |
d5b7b3ae RE |
8765 | else if (mp->max_address |
8766 | < minipool_barrier->address + mp->offset + fix->fix_size) | |
8767 | { | |
8768 | /* Inserting before this entry would push the fix beyond | |
8769 | its maximum address (which can happen if we have | |
8770 | re-located a forwards fix); force the new fix to come | |
8771 | after it. */ | |
8772 | min_mp = mp; | |
8773 | min_address = mp->min_address + fix->fix_size; | |
8774 | } | |
5a9335ef NC |
8775 | /* If we are inserting an 8-bytes aligned quantity and |
8776 | we have not already found an insertion point, then | |
8777 | make sure that all such 8-byte aligned quantities are | |
8778 | placed at the start of the pool. */ | |
5848830f | 8779 | else if (ARM_DOUBLEWORD_ALIGN |
5a9335ef | 8780 | && min_mp == NULL |
88f77cba | 8781 | && fix->fix_size >= 8 |
5a9335ef NC |
8782 | && mp->fix_size < 8) |
8783 | { | |
8784 | min_mp = mp; | |
8785 | min_address = mp->min_address + fix->fix_size; | |
8786 | } | |
d5b7b3ae RE |
8787 | } |
8788 | } | |
8789 | ||
8790 | /* We need to create a new entry. */ | |
5ed6ace5 | 8791 | mp = XNEW (Mnode); |
d5b7b3ae RE |
8792 | mp->fix_size = fix->fix_size; |
8793 | mp->mode = fix->mode; | |
8794 | mp->value = fix->value; | |
8795 | mp->refcount = 1; | |
8796 | mp->max_address = minipool_barrier->address + 65536; | |
8797 | ||
8798 | mp->min_address = min_address; | |
8799 | ||
8800 | if (min_mp == NULL) | |
8801 | { | |
8802 | mp->prev = NULL; | |
8803 | mp->next = minipool_vector_head; | |
8804 | ||
8805 | if (mp->next == NULL) | |
8806 | { | |
8807 | minipool_vector_tail = mp; | |
8808 | minipool_vector_label = gen_label_rtx (); | |
8809 | } | |
8810 | else | |
8811 | mp->next->prev = mp; | |
8812 | ||
8813 | minipool_vector_head = mp; | |
8814 | } | |
8815 | else | |
8816 | { | |
8817 | mp->next = min_mp->next; | |
8818 | mp->prev = min_mp; | |
8819 | min_mp->next = mp; | |
f676971a | 8820 | |
d5b7b3ae RE |
8821 | if (mp->next != NULL) |
8822 | mp->next->prev = mp; | |
8823 | else | |
8824 | minipool_vector_tail = mp; | |
8825 | } | |
8826 | ||
8827 | /* Save the new entry. */ | |
8828 | min_mp = mp; | |
8829 | ||
8830 | if (mp->prev) | |
8831 | mp = mp->prev; | |
8832 | else | |
8833 | mp->offset = 0; | |
8834 | ||
8835 | /* Scan over the following entries and adjust their offsets. */ | |
8836 | while (mp->next != NULL) | |
8837 | { | |
8838 | if (mp->next->min_address < mp->min_address + mp->fix_size) | |
8839 | mp->next->min_address = mp->min_address + mp->fix_size; | |
8840 | ||
8841 | if (mp->refcount) | |
8842 | mp->next->offset = mp->offset + mp->fix_size; | |
8843 | else | |
8844 | mp->next->offset = mp->offset; | |
8845 | ||
8846 | mp = mp->next; | |
8847 | } | |
8848 | ||
8849 | return min_mp; | |
8850 | } | |
8851 | ||
8852 | static void | |
e32bac5b | 8853 | assign_minipool_offsets (Mfix *barrier) |
d5b7b3ae RE |
8854 | { |
8855 | HOST_WIDE_INT offset = 0; | |
e32bac5b | 8856 | Mnode *mp; |
d5b7b3ae RE |
8857 | |
8858 | minipool_barrier = barrier; | |
8859 | ||
8860 | for (mp = minipool_vector_head; mp != NULL; mp = mp->next) | |
8861 | { | |
8862 | mp->offset = offset; | |
f676971a | 8863 | |
d5b7b3ae RE |
8864 | if (mp->refcount > 0) |
8865 | offset += mp->fix_size; | |
8866 | } | |
8867 | } | |
8868 | ||
8869 | /* Output the literal table */ | |
8870 | static void | |
e32bac5b | 8871 | dump_minipool (rtx scan) |
d5b7b3ae | 8872 | { |
5a9335ef NC |
8873 | Mnode * mp; |
8874 | Mnode * nmp; | |
8875 | int align64 = 0; | |
8876 | ||
5848830f | 8877 | if (ARM_DOUBLEWORD_ALIGN) |
5a9335ef | 8878 | for (mp = minipool_vector_head; mp != NULL; mp = mp->next) |
88f77cba | 8879 | if (mp->refcount > 0 && mp->fix_size >= 8) |
5a9335ef NC |
8880 | { |
8881 | align64 = 1; | |
8882 | break; | |
8883 | } | |
d5b7b3ae | 8884 | |
c263766c RH |
8885 | if (dump_file) |
8886 | fprintf (dump_file, | |
5a9335ef NC |
8887 | ";; Emitting minipool after insn %u; address %ld; align %d (bytes)\n", |
8888 | INSN_UID (scan), (unsigned long) minipool_barrier->address, align64 ? 8 : 4); | |
d5b7b3ae RE |
8889 | |
8890 | scan = emit_label_after (gen_label_rtx (), scan); | |
5a9335ef | 8891 | scan = emit_insn_after (align64 ? gen_align_8 () : gen_align_4 (), scan); |
d5b7b3ae RE |
8892 | scan = emit_label_after (minipool_vector_label, scan); |
8893 | ||
8894 | for (mp = minipool_vector_head; mp != NULL; mp = nmp) | |
8895 | { | |
8896 | if (mp->refcount > 0) | |
8897 | { | |
c263766c | 8898 | if (dump_file) |
d5b7b3ae | 8899 | { |
f676971a | 8900 | fprintf (dump_file, |
d5b7b3ae RE |
8901 | ";; Offset %u, min %ld, max %ld ", |
8902 | (unsigned) mp->offset, (unsigned long) mp->min_address, | |
8903 | (unsigned long) mp->max_address); | |
c263766c RH |
8904 | arm_print_value (dump_file, mp->value); |
8905 | fputc ('\n', dump_file); | |
d5b7b3ae RE |
8906 | } |
8907 | ||
8908 | switch (mp->fix_size) | |
8909 | { | |
8910 | #ifdef HAVE_consttable_1 | |
8911 | case 1: | |
8912 | scan = emit_insn_after (gen_consttable_1 (mp->value), scan); | |
8913 | break; | |
8914 | ||
8915 | #endif | |
8916 | #ifdef HAVE_consttable_2 | |
8917 | case 2: | |
8918 | scan = emit_insn_after (gen_consttable_2 (mp->value), scan); | |
8919 | break; | |
8920 | ||
8921 | #endif | |
8922 | #ifdef HAVE_consttable_4 | |
8923 | case 4: | |
8924 | scan = emit_insn_after (gen_consttable_4 (mp->value), scan); | |
8925 | break; | |
8926 | ||
8927 | #endif | |
8928 | #ifdef HAVE_consttable_8 | |
8929 | case 8: | |
8930 | scan = emit_insn_after (gen_consttable_8 (mp->value), scan); | |
8931 | break; | |
8932 | ||
88f77cba JB |
8933 | #endif |
8934 | #ifdef HAVE_consttable_16 | |
8935 | case 16: | |
8936 | scan = emit_insn_after (gen_consttable_16 (mp->value), scan); | |
8937 | break; | |
8938 | ||
d5b7b3ae RE |
8939 | #endif |
8940 | default: | |
e6d29d15 | 8941 | gcc_unreachable (); |
d5b7b3ae RE |
8942 | } |
8943 | } | |
8944 | ||
8945 | nmp = mp->next; | |
8946 | free (mp); | |
2b835d68 RE |
8947 | } |
8948 | ||
d5b7b3ae RE |
8949 | minipool_vector_head = minipool_vector_tail = NULL; |
8950 | scan = emit_insn_after (gen_consttable_end (), scan); | |
8951 | scan = emit_barrier_after (scan); | |
2b835d68 RE |
8952 | } |
8953 | ||
d5b7b3ae RE |
8954 | /* Return the cost of forcibly inserting a barrier after INSN. */ |
8955 | static int | |
e32bac5b | 8956 | arm_barrier_cost (rtx insn) |
949d79eb | 8957 | { |
d5b7b3ae RE |
8958 | /* Basing the location of the pool on the loop depth is preferable, |
8959 | but at the moment, the basic block information seems to be | |
8960 | corrupt by this stage of the compilation. */ | |
8961 | int base_cost = 50; | |
8962 | rtx next = next_nonnote_insn (insn); | |
8963 | ||
8964 | if (next != NULL && GET_CODE (next) == CODE_LABEL) | |
8965 | base_cost -= 20; | |
8966 | ||
8967 | switch (GET_CODE (insn)) | |
8968 | { | |
8969 | case CODE_LABEL: | |
8970 | /* It will always be better to place the table before the label, rather | |
8971 | than after it. */ | |
f676971a | 8972 | return 50; |
949d79eb | 8973 | |
d5b7b3ae RE |
8974 | case INSN: |
8975 | case CALL_INSN: | |
8976 | return base_cost; | |
8977 | ||
8978 | case JUMP_INSN: | |
8979 | return base_cost - 10; | |
8980 | ||
8981 | default: | |
8982 | return base_cost + 10; | |
8983 | } | |
8984 | } | |
8985 | ||
8986 | /* Find the best place in the insn stream in the range | |
8987 | (FIX->address,MAX_ADDRESS) to forcibly insert a minipool barrier. | |
8988 | Create the barrier by inserting a jump and add a new fix entry for | |
8989 | it. */ | |
8990 | static Mfix * | |
e32bac5b | 8991 | create_fix_barrier (Mfix *fix, HOST_WIDE_INT max_address) |
d5b7b3ae RE |
8992 | { |
8993 | HOST_WIDE_INT count = 0; | |
8994 | rtx barrier; | |
8995 | rtx from = fix->insn; | |
7a7017bc PB |
8996 | /* The instruction after which we will insert the jump. */ |
8997 | rtx selected = NULL; | |
d5b7b3ae | 8998 | int selected_cost; |
7a7017bc | 8999 | /* The address at which the jump instruction will be placed. */ |
d5b7b3ae RE |
9000 | HOST_WIDE_INT selected_address; |
9001 | Mfix * new_fix; | |
9002 | HOST_WIDE_INT max_count = max_address - fix->address; | |
9003 | rtx label = gen_label_rtx (); | |
9004 | ||
9005 | selected_cost = arm_barrier_cost (from); | |
9006 | selected_address = fix->address; | |
9007 | ||
9008 | while (from && count < max_count) | |
9009 | { | |
9010 | rtx tmp; | |
9011 | int new_cost; | |
9012 | ||
9013 | /* This code shouldn't have been called if there was a natural barrier | |
9014 | within range. */ | |
e6d29d15 | 9015 | gcc_assert (GET_CODE (from) != BARRIER); |
d5b7b3ae RE |
9016 | |
9017 | /* Count the length of this insn. */ | |
9018 | count += get_attr_length (from); | |
9019 | ||
9020 | /* If there is a jump table, add its length. */ | |
9021 | tmp = is_jump_table (from); | |
9022 | if (tmp != NULL) | |
9023 | { | |
9024 | count += get_jump_table_size (tmp); | |
9025 | ||
9026 | /* Jump tables aren't in a basic block, so base the cost on | |
9027 | the dispatch insn. If we select this location, we will | |
9028 | still put the pool after the table. */ | |
9029 | new_cost = arm_barrier_cost (from); | |
9030 | ||
7a7017bc PB |
9031 | if (count < max_count |
9032 | && (!selected || new_cost <= selected_cost)) | |
d5b7b3ae RE |
9033 | { |
9034 | selected = tmp; | |
9035 | selected_cost = new_cost; | |
9036 | selected_address = fix->address + count; | |
9037 | } | |
9038 | ||
9039 | /* Continue after the dispatch table. */ | |
9040 | from = NEXT_INSN (tmp); | |
9041 | continue; | |
9042 | } | |
9043 | ||
9044 | new_cost = arm_barrier_cost (from); | |
f676971a | 9045 | |
7a7017bc PB |
9046 | if (count < max_count |
9047 | && (!selected || new_cost <= selected_cost)) | |
d5b7b3ae RE |
9048 | { |
9049 | selected = from; | |
9050 | selected_cost = new_cost; | |
9051 | selected_address = fix->address + count; | |
9052 | } | |
9053 | ||
9054 | from = NEXT_INSN (from); | |
9055 | } | |
9056 | ||
7a7017bc PB |
9057 | /* Make sure that we found a place to insert the jump. */ |
9058 | gcc_assert (selected); | |
9059 | ||
d5b7b3ae RE |
9060 | /* Create a new JUMP_INSN that branches around a barrier. */ |
9061 | from = emit_jump_insn_after (gen_jump (label), selected); | |
9062 | JUMP_LABEL (from) = label; | |
9063 | barrier = emit_barrier_after (from); | |
9064 | emit_label_after (label, barrier); | |
9065 | ||
9066 | /* Create a minipool barrier entry for the new barrier. */ | |
c7319d87 | 9067 | new_fix = (Mfix *) obstack_alloc (&minipool_obstack, sizeof (* new_fix)); |
d5b7b3ae RE |
9068 | new_fix->insn = barrier; |
9069 | new_fix->address = selected_address; | |
9070 | new_fix->next = fix->next; | |
9071 | fix->next = new_fix; | |
9072 | ||
9073 | return new_fix; | |
9074 | } | |
9075 | ||
9076 | /* Record that there is a natural barrier in the insn stream at | |
9077 | ADDRESS. */ | |
949d79eb | 9078 | static void |
e32bac5b | 9079 | push_minipool_barrier (rtx insn, HOST_WIDE_INT address) |
2b835d68 | 9080 | { |
c7319d87 | 9081 | Mfix * fix = (Mfix *) obstack_alloc (&minipool_obstack, sizeof (* fix)); |
ad076f4e | 9082 | |
949d79eb RE |
9083 | fix->insn = insn; |
9084 | fix->address = address; | |
2b835d68 | 9085 | |
949d79eb RE |
9086 | fix->next = NULL; |
9087 | if (minipool_fix_head != NULL) | |
9088 | minipool_fix_tail->next = fix; | |
9089 | else | |
9090 | minipool_fix_head = fix; | |
9091 | ||
9092 | minipool_fix_tail = fix; | |
9093 | } | |
2b835d68 | 9094 | |
d5b7b3ae RE |
9095 | /* Record INSN, which will need fixing up to load a value from the |
9096 | minipool. ADDRESS is the offset of the insn since the start of the | |
9097 | function; LOC is a pointer to the part of the insn which requires | |
9098 | fixing; VALUE is the constant that must be loaded, which is of type | |
9099 | MODE. */ | |
949d79eb | 9100 | static void |
e32bac5b RE |
9101 | push_minipool_fix (rtx insn, HOST_WIDE_INT address, rtx *loc, |
9102 | enum machine_mode mode, rtx value) | |
949d79eb | 9103 | { |
c7319d87 | 9104 | Mfix * fix = (Mfix *) obstack_alloc (&minipool_obstack, sizeof (* fix)); |
949d79eb | 9105 | |
949d79eb RE |
9106 | fix->insn = insn; |
9107 | fix->address = address; | |
9108 | fix->loc = loc; | |
9109 | fix->mode = mode; | |
d5b7b3ae | 9110 | fix->fix_size = MINIPOOL_FIX_SIZE (mode); |
949d79eb | 9111 | fix->value = value; |
d5b7b3ae RE |
9112 | fix->forwards = get_attr_pool_range (insn); |
9113 | fix->backwards = get_attr_neg_pool_range (insn); | |
9114 | fix->minipool = NULL; | |
949d79eb RE |
9115 | |
9116 | /* If an insn doesn't have a range defined for it, then it isn't | |
e6d29d15 | 9117 | expecting to be reworked by this code. Better to stop now than |
949d79eb | 9118 | to generate duff assembly code. */ |
e6d29d15 | 9119 | gcc_assert (fix->forwards || fix->backwards); |
949d79eb | 9120 | |
34a9f549 PB |
9121 | /* If an entry requires 8-byte alignment then assume all constant pools |
9122 | require 4 bytes of padding. Trying to do this later on a per-pool | |
917f1b7e | 9123 | basis is awkward because existing pool entries have to be modified. */ |
88f77cba | 9124 | if (ARM_DOUBLEWORD_ALIGN && fix->fix_size >= 8) |
34a9f549 | 9125 | minipool_pad = 4; |
5a9335ef | 9126 | |
c263766c | 9127 | if (dump_file) |
d5b7b3ae | 9128 | { |
c263766c | 9129 | fprintf (dump_file, |
d5b7b3ae RE |
9130 | ";; %smode fixup for i%d; addr %lu, range (%ld,%ld): ", |
9131 | GET_MODE_NAME (mode), | |
f676971a | 9132 | INSN_UID (insn), (unsigned long) address, |
d5b7b3ae | 9133 | -1 * (long)fix->backwards, (long)fix->forwards); |
c263766c RH |
9134 | arm_print_value (dump_file, fix->value); |
9135 | fprintf (dump_file, "\n"); | |
d5b7b3ae RE |
9136 | } |
9137 | ||
6354dc9b | 9138 | /* Add it to the chain of fixes. */ |
949d79eb | 9139 | fix->next = NULL; |
f676971a | 9140 | |
949d79eb RE |
9141 | if (minipool_fix_head != NULL) |
9142 | minipool_fix_tail->next = fix; | |
9143 | else | |
9144 | minipool_fix_head = fix; | |
9145 | ||
9146 | minipool_fix_tail = fix; | |
9147 | } | |
9148 | ||
9b901d50 | 9149 | /* Return the cost of synthesizing a 64-bit constant VAL inline. |
2075b05d RE |
9150 | Returns the number of insns needed, or 99 if we don't know how to |
9151 | do it. */ | |
9152 | int | |
9153 | arm_const_double_inline_cost (rtx val) | |
b9e8bfda | 9154 | { |
9b901d50 RE |
9155 | rtx lowpart, highpart; |
9156 | enum machine_mode mode; | |
e0b92319 | 9157 | |
9b901d50 | 9158 | mode = GET_MODE (val); |
b9e8bfda | 9159 | |
9b901d50 RE |
9160 | if (mode == VOIDmode) |
9161 | mode = DImode; | |
9162 | ||
9163 | gcc_assert (GET_MODE_SIZE (mode) == 8); | |
e0b92319 | 9164 | |
9b901d50 RE |
9165 | lowpart = gen_lowpart (SImode, val); |
9166 | highpart = gen_highpart_mode (SImode, mode, val); | |
e0b92319 | 9167 | |
9b901d50 RE |
9168 | gcc_assert (GET_CODE (lowpart) == CONST_INT); |
9169 | gcc_assert (GET_CODE (highpart) == CONST_INT); | |
9170 | ||
9171 | return (arm_gen_constant (SET, SImode, NULL_RTX, INTVAL (lowpart), | |
2075b05d | 9172 | NULL_RTX, NULL_RTX, 0, 0) |
9b901d50 | 9173 | + arm_gen_constant (SET, SImode, NULL_RTX, INTVAL (highpart), |
2075b05d RE |
9174 | NULL_RTX, NULL_RTX, 0, 0)); |
9175 | } | |
9176 | ||
f3b569ca | 9177 | /* Return true if it is worthwhile to split a 64-bit constant into two |
b4a58f80 RE |
9178 | 32-bit operations. This is the case if optimizing for size, or |
9179 | if we have load delay slots, or if one 32-bit part can be done with | |
9180 | a single data operation. */ | |
9181 | bool | |
9182 | arm_const_double_by_parts (rtx val) | |
9183 | { | |
9184 | enum machine_mode mode = GET_MODE (val); | |
9185 | rtx part; | |
9186 | ||
9187 | if (optimize_size || arm_ld_sched) | |
9188 | return true; | |
9189 | ||
9190 | if (mode == VOIDmode) | |
9191 | mode = DImode; | |
e0b92319 | 9192 | |
b4a58f80 | 9193 | part = gen_highpart_mode (SImode, mode, val); |
e0b92319 | 9194 | |
b4a58f80 | 9195 | gcc_assert (GET_CODE (part) == CONST_INT); |
e0b92319 | 9196 | |
b4a58f80 RE |
9197 | if (const_ok_for_arm (INTVAL (part)) |
9198 | || const_ok_for_arm (~INTVAL (part))) | |
9199 | return true; | |
e0b92319 | 9200 | |
b4a58f80 | 9201 | part = gen_lowpart (SImode, val); |
e0b92319 | 9202 | |
b4a58f80 | 9203 | gcc_assert (GET_CODE (part) == CONST_INT); |
e0b92319 | 9204 | |
b4a58f80 RE |
9205 | if (const_ok_for_arm (INTVAL (part)) |
9206 | || const_ok_for_arm (~INTVAL (part))) | |
9207 | return true; | |
e0b92319 | 9208 | |
b4a58f80 RE |
9209 | return false; |
9210 | } | |
9211 | ||
f0375c66 NC |
9212 | /* Scan INSN and note any of its operands that need fixing. |
9213 | If DO_PUSHES is false we do not actually push any of the fixups | |
9b901d50 | 9214 | needed. The function returns TRUE if any fixups were needed/pushed. |
f0375c66 NC |
9215 | This is used by arm_memory_load_p() which needs to know about loads |
9216 | of constants that will be converted into minipool loads. */ | |
f0375c66 | 9217 | static bool |
e32bac5b | 9218 | note_invalid_constants (rtx insn, HOST_WIDE_INT address, int do_pushes) |
949d79eb | 9219 | { |
f0375c66 | 9220 | bool result = false; |
949d79eb RE |
9221 | int opno; |
9222 | ||
d5b7b3ae | 9223 | extract_insn (insn); |
949d79eb | 9224 | |
5895f793 | 9225 | if (!constrain_operands (1)) |
949d79eb RE |
9226 | fatal_insn_not_found (insn); |
9227 | ||
8c2a5582 RE |
9228 | if (recog_data.n_alternatives == 0) |
9229 | return false; | |
9230 | ||
9b901d50 RE |
9231 | /* Fill in recog_op_alt with information about the constraints of |
9232 | this insn. */ | |
949d79eb RE |
9233 | preprocess_constraints (); |
9234 | ||
1ccbefce | 9235 | for (opno = 0; opno < recog_data.n_operands; opno++) |
949d79eb | 9236 | { |
6354dc9b | 9237 | /* Things we need to fix can only occur in inputs. */ |
36ab44c7 | 9238 | if (recog_data.operand_type[opno] != OP_IN) |
949d79eb RE |
9239 | continue; |
9240 | ||
9241 | /* If this alternative is a memory reference, then any mention | |
9242 | of constants in this alternative is really to fool reload | |
9243 | into allowing us to accept one there. We need to fix them up | |
9244 | now so that we output the right code. */ | |
9245 | if (recog_op_alt[opno][which_alternative].memory_ok) | |
9246 | { | |
1ccbefce | 9247 | rtx op = recog_data.operand[opno]; |
949d79eb | 9248 | |
9b901d50 | 9249 | if (CONSTANT_P (op)) |
f0375c66 NC |
9250 | { |
9251 | if (do_pushes) | |
9252 | push_minipool_fix (insn, address, recog_data.operand_loc[opno], | |
9253 | recog_data.operand_mode[opno], op); | |
9254 | result = true; | |
9255 | } | |
d5b7b3ae | 9256 | else if (GET_CODE (op) == MEM |
949d79eb RE |
9257 | && GET_CODE (XEXP (op, 0)) == SYMBOL_REF |
9258 | && CONSTANT_POOL_ADDRESS_P (XEXP (op, 0))) | |
f0375c66 NC |
9259 | { |
9260 | if (do_pushes) | |
244b1afb RE |
9261 | { |
9262 | rtx cop = avoid_constant_pool_reference (op); | |
9263 | ||
9264 | /* Casting the address of something to a mode narrower | |
9265 | than a word can cause avoid_constant_pool_reference() | |
9266 | to return the pool reference itself. That's no good to | |
f676971a | 9267 | us here. Lets just hope that we can use the |
244b1afb RE |
9268 | constant pool value directly. */ |
9269 | if (op == cop) | |
c769a35d | 9270 | cop = get_pool_constant (XEXP (op, 0)); |
244b1afb RE |
9271 | |
9272 | push_minipool_fix (insn, address, | |
9273 | recog_data.operand_loc[opno], | |
c769a35d | 9274 | recog_data.operand_mode[opno], cop); |
244b1afb | 9275 | } |
f0375c66 NC |
9276 | |
9277 | result = true; | |
9278 | } | |
949d79eb | 9279 | } |
2b835d68 | 9280 | } |
f0375c66 NC |
9281 | |
9282 | return result; | |
2b835d68 RE |
9283 | } |
9284 | ||
18dbd950 RS |
9285 | /* Gcc puts the pool in the wrong place for ARM, since we can only |
9286 | load addresses a limited distance around the pc. We do some | |
9287 | special munging to move the constant pool values to the correct | |
9288 | point in the code. */ | |
18dbd950 | 9289 | static void |
e32bac5b | 9290 | arm_reorg (void) |
2b835d68 RE |
9291 | { |
9292 | rtx insn; | |
d5b7b3ae RE |
9293 | HOST_WIDE_INT address = 0; |
9294 | Mfix * fix; | |
ad076f4e | 9295 | |
949d79eb | 9296 | minipool_fix_head = minipool_fix_tail = NULL; |
2b835d68 | 9297 | |
949d79eb RE |
9298 | /* The first insn must always be a note, or the code below won't |
9299 | scan it properly. */ | |
18dbd950 | 9300 | insn = get_insns (); |
e6d29d15 | 9301 | gcc_assert (GET_CODE (insn) == NOTE); |
34a9f549 | 9302 | minipool_pad = 0; |
949d79eb RE |
9303 | |
9304 | /* Scan all the insns and record the operands that will need fixing. */ | |
18dbd950 | 9305 | for (insn = next_nonnote_insn (insn); insn; insn = next_nonnote_insn (insn)) |
2b835d68 | 9306 | { |
9b6b54e2 | 9307 | if (TARGET_CIRRUS_FIX_INVALID_INSNS |
f0375c66 | 9308 | && (arm_cirrus_insn_p (insn) |
9b6b54e2 | 9309 | || GET_CODE (insn) == JUMP_INSN |
f0375c66 | 9310 | || arm_memory_load_p (insn))) |
9b6b54e2 NC |
9311 | cirrus_reorg (insn); |
9312 | ||
949d79eb | 9313 | if (GET_CODE (insn) == BARRIER) |
d5b7b3ae | 9314 | push_minipool_barrier (insn, address); |
f0375c66 | 9315 | else if (INSN_P (insn)) |
949d79eb RE |
9316 | { |
9317 | rtx table; | |
9318 | ||
f0375c66 | 9319 | note_invalid_constants (insn, address, true); |
949d79eb | 9320 | address += get_attr_length (insn); |
d5b7b3ae | 9321 | |
949d79eb RE |
9322 | /* If the insn is a vector jump, add the size of the table |
9323 | and skip the table. */ | |
d5b7b3ae | 9324 | if ((table = is_jump_table (insn)) != NULL) |
2b835d68 | 9325 | { |
d5b7b3ae | 9326 | address += get_jump_table_size (table); |
949d79eb RE |
9327 | insn = table; |
9328 | } | |
9329 | } | |
9330 | } | |
332072db | 9331 | |
d5b7b3ae | 9332 | fix = minipool_fix_head; |
f676971a | 9333 | |
949d79eb | 9334 | /* Now scan the fixups and perform the required changes. */ |
d5b7b3ae | 9335 | while (fix) |
949d79eb | 9336 | { |
d5b7b3ae RE |
9337 | Mfix * ftmp; |
9338 | Mfix * fdel; | |
9339 | Mfix * last_added_fix; | |
9340 | Mfix * last_barrier = NULL; | |
9341 | Mfix * this_fix; | |
949d79eb RE |
9342 | |
9343 | /* Skip any further barriers before the next fix. */ | |
9344 | while (fix && GET_CODE (fix->insn) == BARRIER) | |
9345 | fix = fix->next; | |
9346 | ||
d5b7b3ae | 9347 | /* No more fixes. */ |
949d79eb RE |
9348 | if (fix == NULL) |
9349 | break; | |
332072db | 9350 | |
d5b7b3ae | 9351 | last_added_fix = NULL; |
2b835d68 | 9352 | |
d5b7b3ae | 9353 | for (ftmp = fix; ftmp; ftmp = ftmp->next) |
949d79eb | 9354 | { |
949d79eb | 9355 | if (GET_CODE (ftmp->insn) == BARRIER) |
949d79eb | 9356 | { |
d5b7b3ae RE |
9357 | if (ftmp->address >= minipool_vector_head->max_address) |
9358 | break; | |
2b835d68 | 9359 | |
d5b7b3ae | 9360 | last_barrier = ftmp; |
2b835d68 | 9361 | } |
d5b7b3ae RE |
9362 | else if ((ftmp->minipool = add_minipool_forward_ref (ftmp)) == NULL) |
9363 | break; | |
9364 | ||
9365 | last_added_fix = ftmp; /* Keep track of the last fix added. */ | |
2b835d68 | 9366 | } |
949d79eb | 9367 | |
d5b7b3ae RE |
9368 | /* If we found a barrier, drop back to that; any fixes that we |
9369 | could have reached but come after the barrier will now go in | |
9370 | the next mini-pool. */ | |
949d79eb RE |
9371 | if (last_barrier != NULL) |
9372 | { | |
f676971a | 9373 | /* Reduce the refcount for those fixes that won't go into this |
d5b7b3ae RE |
9374 | pool after all. */ |
9375 | for (fdel = last_barrier->next; | |
9376 | fdel && fdel != ftmp; | |
9377 | fdel = fdel->next) | |
9378 | { | |
9379 | fdel->minipool->refcount--; | |
9380 | fdel->minipool = NULL; | |
9381 | } | |
9382 | ||
949d79eb RE |
9383 | ftmp = last_barrier; |
9384 | } | |
9385 | else | |
2bfa88dc | 9386 | { |
d5b7b3ae RE |
9387 | /* ftmp is first fix that we can't fit into this pool and |
9388 | there no natural barriers that we could use. Insert a | |
9389 | new barrier in the code somewhere between the previous | |
9390 | fix and this one, and arrange to jump around it. */ | |
9391 | HOST_WIDE_INT max_address; | |
9392 | ||
9393 | /* The last item on the list of fixes must be a barrier, so | |
9394 | we can never run off the end of the list of fixes without | |
9395 | last_barrier being set. */ | |
e6d29d15 | 9396 | gcc_assert (ftmp); |
d5b7b3ae RE |
9397 | |
9398 | max_address = minipool_vector_head->max_address; | |
2bfa88dc RE |
9399 | /* Check that there isn't another fix that is in range that |
9400 | we couldn't fit into this pool because the pool was | |
9401 | already too large: we need to put the pool before such an | |
7a7017bc PB |
9402 | instruction. The pool itself may come just after the |
9403 | fix because create_fix_barrier also allows space for a | |
9404 | jump instruction. */ | |
d5b7b3ae | 9405 | if (ftmp->address < max_address) |
7a7017bc | 9406 | max_address = ftmp->address + 1; |
d5b7b3ae RE |
9407 | |
9408 | last_barrier = create_fix_barrier (last_added_fix, max_address); | |
9409 | } | |
9410 | ||
9411 | assign_minipool_offsets (last_barrier); | |
9412 | ||
9413 | while (ftmp) | |
9414 | { | |
9415 | if (GET_CODE (ftmp->insn) != BARRIER | |
9416 | && ((ftmp->minipool = add_minipool_backward_ref (ftmp)) | |
9417 | == NULL)) | |
9418 | break; | |
2bfa88dc | 9419 | |
d5b7b3ae | 9420 | ftmp = ftmp->next; |
2bfa88dc | 9421 | } |
949d79eb RE |
9422 | |
9423 | /* Scan over the fixes we have identified for this pool, fixing them | |
9424 | up and adding the constants to the pool itself. */ | |
d5b7b3ae | 9425 | for (this_fix = fix; this_fix && ftmp != this_fix; |
949d79eb RE |
9426 | this_fix = this_fix->next) |
9427 | if (GET_CODE (this_fix->insn) != BARRIER) | |
9428 | { | |
949d79eb | 9429 | rtx addr |
f676971a | 9430 | = plus_constant (gen_rtx_LABEL_REF (VOIDmode, |
949d79eb | 9431 | minipool_vector_label), |
d5b7b3ae | 9432 | this_fix->minipool->offset); |
949d79eb RE |
9433 | *this_fix->loc = gen_rtx_MEM (this_fix->mode, addr); |
9434 | } | |
9435 | ||
d5b7b3ae | 9436 | dump_minipool (last_barrier->insn); |
949d79eb | 9437 | fix = ftmp; |
2b835d68 | 9438 | } |
4b632bf1 | 9439 | |
949d79eb RE |
9440 | /* From now on we must synthesize any constants that we can't handle |
9441 | directly. This can happen if the RTL gets split during final | |
9442 | instruction generation. */ | |
4b632bf1 | 9443 | after_arm_reorg = 1; |
c7319d87 RE |
9444 | |
9445 | /* Free the minipool memory. */ | |
9446 | obstack_free (&minipool_obstack, minipool_startobj); | |
2b835d68 | 9447 | } |
cce8749e CH |
9448 | \f |
9449 | /* Routines to output assembly language. */ | |
9450 | ||
f3bb6135 | 9451 | /* If the rtx is the correct value then return the string of the number. |
ff9940b0 | 9452 | In this way we can ensure that valid double constants are generated even |
6354dc9b | 9453 | when cross compiling. */ |
cd2b33d0 | 9454 | const char * |
e32bac5b | 9455 | fp_immediate_constant (rtx x) |
ff9940b0 RE |
9456 | { |
9457 | REAL_VALUE_TYPE r; | |
9458 | int i; | |
f676971a | 9459 | |
9b66ebb1 PB |
9460 | if (!fp_consts_inited) |
9461 | init_fp_table (); | |
f676971a | 9462 | |
ff9940b0 RE |
9463 | REAL_VALUE_FROM_CONST_DOUBLE (r, x); |
9464 | for (i = 0; i < 8; i++) | |
9b66ebb1 PB |
9465 | if (REAL_VALUES_EQUAL (r, values_fp[i])) |
9466 | return strings_fp[i]; | |
f3bb6135 | 9467 | |
e6d29d15 | 9468 | gcc_unreachable (); |
ff9940b0 RE |
9469 | } |
9470 | ||
9997d19d | 9471 | /* As for fp_immediate_constant, but value is passed directly, not in rtx. */ |
cd2b33d0 | 9472 | static const char * |
e32bac5b | 9473 | fp_const_from_val (REAL_VALUE_TYPE *r) |
9997d19d RE |
9474 | { |
9475 | int i; | |
9476 | ||
9b66ebb1 PB |
9477 | if (!fp_consts_inited) |
9478 | init_fp_table (); | |
9997d19d RE |
9479 | |
9480 | for (i = 0; i < 8; i++) | |
9b66ebb1 PB |
9481 | if (REAL_VALUES_EQUAL (*r, values_fp[i])) |
9482 | return strings_fp[i]; | |
9997d19d | 9483 | |
e6d29d15 | 9484 | gcc_unreachable (); |
9997d19d | 9485 | } |
ff9940b0 | 9486 | |
cce8749e CH |
9487 | /* Output the operands of a LDM/STM instruction to STREAM. |
9488 | MASK is the ARM register set mask of which only bits 0-15 are important. | |
6d3d9133 | 9489 | REG is the base register, either the frame pointer or the stack pointer, |
a15908a4 PB |
9490 | INSTR is the possibly suffixed load or store instruction. |
9491 | RFE is nonzero if the instruction should also copy spsr to cpsr. */ | |
b279b20a | 9492 | |
d5b7b3ae | 9493 | static void |
b279b20a | 9494 | print_multi_reg (FILE *stream, const char *instr, unsigned reg, |
a15908a4 | 9495 | unsigned long mask, int rfe) |
cce8749e | 9496 | { |
b279b20a NC |
9497 | unsigned i; |
9498 | bool not_first = FALSE; | |
cce8749e | 9499 | |
a15908a4 | 9500 | gcc_assert (!rfe || (mask & (1 << PC_REGNUM))); |
1d5473cb | 9501 | fputc ('\t', stream); |
dd18ae56 | 9502 | asm_fprintf (stream, instr, reg); |
5b3e6663 | 9503 | fputc ('{', stream); |
f676971a | 9504 | |
d5b7b3ae | 9505 | for (i = 0; i <= LAST_ARM_REGNUM; i++) |
cce8749e CH |
9506 | if (mask & (1 << i)) |
9507 | { | |
9508 | if (not_first) | |
9509 | fprintf (stream, ", "); | |
f676971a | 9510 | |
dd18ae56 | 9511 | asm_fprintf (stream, "%r", i); |
cce8749e CH |
9512 | not_first = TRUE; |
9513 | } | |
f3bb6135 | 9514 | |
a15908a4 PB |
9515 | if (rfe) |
9516 | fprintf (stream, "}^\n"); | |
9517 | else | |
9518 | fprintf (stream, "}\n"); | |
f3bb6135 | 9519 | } |
cce8749e | 9520 | |
9b66ebb1 | 9521 | |
8edfc4cc | 9522 | /* Output a FLDMD instruction to STREAM. |
9728c9d1 PB |
9523 | BASE if the register containing the address. |
9524 | REG and COUNT specify the register range. | |
8edfc4cc MS |
9525 | Extra registers may be added to avoid hardware bugs. |
9526 | ||
9527 | We output FLDMD even for ARMv5 VFP implementations. Although | |
9528 | FLDMD is technically not supported until ARMv6, it is believed | |
9529 | that all VFP implementations support its use in this context. */ | |
9b66ebb1 PB |
9530 | |
9531 | static void | |
8edfc4cc | 9532 | vfp_output_fldmd (FILE * stream, unsigned int base, int reg, int count) |
9b66ebb1 PB |
9533 | { |
9534 | int i; | |
9535 | ||
9728c9d1 PB |
9536 | /* Workaround ARM10 VFPr1 bug. */ |
9537 | if (count == 2 && !arm_arch6) | |
9538 | { | |
9539 | if (reg == 15) | |
9540 | reg--; | |
9541 | count++; | |
9542 | } | |
9543 | ||
f1adb0a9 JB |
9544 | /* FLDMD may not load more than 16 doubleword registers at a time. Split the |
9545 | load into multiple parts if we have to handle more than 16 registers. */ | |
9546 | if (count > 16) | |
9547 | { | |
9548 | vfp_output_fldmd (stream, base, reg, 16); | |
9549 | vfp_output_fldmd (stream, base, reg + 16, count - 16); | |
9550 | return; | |
9551 | } | |
9552 | ||
9b66ebb1 | 9553 | fputc ('\t', stream); |
8edfc4cc | 9554 | asm_fprintf (stream, "fldmfdd\t%r!, {", base); |
9b66ebb1 | 9555 | |
9728c9d1 | 9556 | for (i = reg; i < reg + count; i++) |
9b66ebb1 | 9557 | { |
9728c9d1 | 9558 | if (i > reg) |
9b66ebb1 | 9559 | fputs (", ", stream); |
9728c9d1 | 9560 | asm_fprintf (stream, "d%d", i); |
9b66ebb1 PB |
9561 | } |
9562 | fputs ("}\n", stream); | |
9728c9d1 | 9563 | |
9b66ebb1 PB |
9564 | } |
9565 | ||
9566 | ||
9567 | /* Output the assembly for a store multiple. */ | |
9568 | ||
9569 | const char * | |
8edfc4cc | 9570 | vfp_output_fstmd (rtx * operands) |
9b66ebb1 PB |
9571 | { |
9572 | char pattern[100]; | |
9573 | int p; | |
9574 | int base; | |
9575 | int i; | |
9576 | ||
8edfc4cc | 9577 | strcpy (pattern, "fstmfdd\t%m0!, {%P1"); |
9b66ebb1 PB |
9578 | p = strlen (pattern); |
9579 | ||
e6d29d15 | 9580 | gcc_assert (GET_CODE (operands[1]) == REG); |
9b66ebb1 PB |
9581 | |
9582 | base = (REGNO (operands[1]) - FIRST_VFP_REGNUM) / 2; | |
9583 | for (i = 1; i < XVECLEN (operands[2], 0); i++) | |
9584 | { | |
9585 | p += sprintf (&pattern[p], ", d%d", base + i); | |
9586 | } | |
9587 | strcpy (&pattern[p], "}"); | |
9588 | ||
9589 | output_asm_insn (pattern, operands); | |
9590 | return ""; | |
9591 | } | |
9592 | ||
9593 | ||
9728c9d1 PB |
9594 | /* Emit RTL to save block of VFP register pairs to the stack. Returns the |
9595 | number of bytes pushed. */ | |
9b66ebb1 | 9596 | |
9728c9d1 | 9597 | static int |
8edfc4cc | 9598 | vfp_emit_fstmd (int base_reg, int count) |
9b66ebb1 PB |
9599 | { |
9600 | rtx par; | |
9601 | rtx dwarf; | |
9602 | rtx tmp, reg; | |
9603 | int i; | |
9604 | ||
9728c9d1 PB |
9605 | /* Workaround ARM10 VFPr1 bug. Data corruption can occur when exactly two |
9606 | register pairs are stored by a store multiple insn. We avoid this | |
9607 | by pushing an extra pair. */ | |
9608 | if (count == 2 && !arm_arch6) | |
9609 | { | |
9610 | if (base_reg == LAST_VFP_REGNUM - 3) | |
9611 | base_reg -= 2; | |
9612 | count++; | |
9613 | } | |
9614 | ||
f1adb0a9 JB |
9615 | /* FSTMD may not store more than 16 doubleword registers at once. Split |
9616 | larger stores into multiple parts (up to a maximum of two, in | |
9617 | practice). */ | |
9618 | if (count > 16) | |
9619 | { | |
9620 | int saved; | |
9621 | /* NOTE: base_reg is an internal register number, so each D register | |
9622 | counts as 2. */ | |
9623 | saved = vfp_emit_fstmd (base_reg + 32, count - 16); | |
9624 | saved += vfp_emit_fstmd (base_reg, 16); | |
9625 | return saved; | |
9626 | } | |
9627 | ||
9b66ebb1 PB |
9628 | par = gen_rtx_PARALLEL (VOIDmode, rtvec_alloc (count)); |
9629 | dwarf = gen_rtx_SEQUENCE (VOIDmode, rtvec_alloc (count + 1)); | |
9630 | ||
9631 | reg = gen_rtx_REG (DFmode, base_reg); | |
9632 | base_reg += 2; | |
9633 | ||
9634 | XVECEXP (par, 0, 0) | |
9635 | = gen_rtx_SET (VOIDmode, | |
31fa16b6 RE |
9636 | gen_frame_mem (BLKmode, |
9637 | gen_rtx_PRE_DEC (BLKmode, | |
9638 | stack_pointer_rtx)), | |
9b66ebb1 PB |
9639 | gen_rtx_UNSPEC (BLKmode, |
9640 | gen_rtvec (1, reg), | |
9641 | UNSPEC_PUSH_MULT)); | |
9642 | ||
9643 | tmp = gen_rtx_SET (VOIDmode, stack_pointer_rtx, | |
8edfc4cc | 9644 | plus_constant (stack_pointer_rtx, -(count * 8))); |
9b66ebb1 PB |
9645 | RTX_FRAME_RELATED_P (tmp) = 1; |
9646 | XVECEXP (dwarf, 0, 0) = tmp; | |
9647 | ||
9648 | tmp = gen_rtx_SET (VOIDmode, | |
31fa16b6 | 9649 | gen_frame_mem (DFmode, stack_pointer_rtx), |
9b66ebb1 PB |
9650 | reg); |
9651 | RTX_FRAME_RELATED_P (tmp) = 1; | |
9652 | XVECEXP (dwarf, 0, 1) = tmp; | |
9653 | ||
9654 | for (i = 1; i < count; i++) | |
9655 | { | |
9656 | reg = gen_rtx_REG (DFmode, base_reg); | |
9657 | base_reg += 2; | |
9658 | XVECEXP (par, 0, i) = gen_rtx_USE (VOIDmode, reg); | |
9659 | ||
9660 | tmp = gen_rtx_SET (VOIDmode, | |
31fa16b6 | 9661 | gen_frame_mem (DFmode, |
d66437c5 RE |
9662 | plus_constant (stack_pointer_rtx, |
9663 | i * 8)), | |
9b66ebb1 PB |
9664 | reg); |
9665 | RTX_FRAME_RELATED_P (tmp) = 1; | |
9666 | XVECEXP (dwarf, 0, i + 1) = tmp; | |
9667 | } | |
9668 | ||
9669 | par = emit_insn (par); | |
9670 | REG_NOTES (par) = gen_rtx_EXPR_LIST (REG_FRAME_RELATED_EXPR, dwarf, | |
9671 | REG_NOTES (par)); | |
9728c9d1 PB |
9672 | RTX_FRAME_RELATED_P (par) = 1; |
9673 | ||
8edfc4cc | 9674 | return count * 8; |
9b66ebb1 PB |
9675 | } |
9676 | ||
9403b7f7 RS |
9677 | /* Emit a call instruction with pattern PAT. ADDR is the address of |
9678 | the call target. */ | |
9679 | ||
9680 | void | |
9681 | arm_emit_call_insn (rtx pat, rtx addr) | |
9682 | { | |
9683 | rtx insn; | |
9684 | ||
9685 | insn = emit_call_insn (pat); | |
9686 | ||
9687 | /* The PIC register is live on entry to VxWorks PIC PLT entries. | |
9688 | If the call might use such an entry, add a use of the PIC register | |
9689 | to the instruction's CALL_INSN_FUNCTION_USAGE. */ | |
9690 | if (TARGET_VXWORKS_RTP | |
9691 | && flag_pic | |
9692 | && GET_CODE (addr) == SYMBOL_REF | |
9693 | && (SYMBOL_REF_DECL (addr) | |
9694 | ? !targetm.binds_local_p (SYMBOL_REF_DECL (addr)) | |
9695 | : !SYMBOL_REF_LOCAL_P (addr))) | |
9696 | { | |
9697 | require_pic_register (); | |
9698 | use_reg (&CALL_INSN_FUNCTION_USAGE (insn), cfun->machine->pic_reg); | |
9699 | } | |
9700 | } | |
9b66ebb1 | 9701 | |
6354dc9b | 9702 | /* Output a 'call' insn. */ |
cd2b33d0 | 9703 | const char * |
e32bac5b | 9704 | output_call (rtx *operands) |
cce8749e | 9705 | { |
e6d29d15 | 9706 | gcc_assert (!arm_arch5); /* Patterns should call blx <reg> directly. */ |
cce8749e | 9707 | |
68d560d4 | 9708 | /* Handle calls to lr using ip (which may be clobbered in subr anyway). */ |
62b10bbc | 9709 | if (REGNO (operands[0]) == LR_REGNUM) |
cce8749e | 9710 | { |
62b10bbc | 9711 | operands[0] = gen_rtx_REG (SImode, IP_REGNUM); |
1d5473cb | 9712 | output_asm_insn ("mov%?\t%0, %|lr", operands); |
cce8749e | 9713 | } |
f676971a | 9714 | |
1d5473cb | 9715 | output_asm_insn ("mov%?\t%|lr, %|pc", operands); |
f676971a | 9716 | |
68d560d4 | 9717 | if (TARGET_INTERWORK || arm_arch4t) |
da6558fd NC |
9718 | output_asm_insn ("bx%?\t%0", operands); |
9719 | else | |
9720 | output_asm_insn ("mov%?\t%|pc, %0", operands); | |
f676971a | 9721 | |
f3bb6135 RE |
9722 | return ""; |
9723 | } | |
cce8749e | 9724 | |
6354dc9b | 9725 | /* Output a 'call' insn that is a reference in memory. */ |
cd2b33d0 | 9726 | const char * |
e32bac5b | 9727 | output_call_mem (rtx *operands) |
ff9940b0 | 9728 | { |
68d560d4 | 9729 | if (TARGET_INTERWORK && !arm_arch5) |
da6558fd NC |
9730 | { |
9731 | output_asm_insn ("ldr%?\t%|ip, %0", operands); | |
9732 | output_asm_insn ("mov%?\t%|lr, %|pc", operands); | |
9733 | output_asm_insn ("bx%?\t%|ip", operands); | |
9734 | } | |
6ab5da80 RE |
9735 | else if (regno_use_in (LR_REGNUM, operands[0])) |
9736 | { | |
9737 | /* LR is used in the memory address. We load the address in the | |
9738 | first instruction. It's safe to use IP as the target of the | |
9739 | load since the call will kill it anyway. */ | |
9740 | output_asm_insn ("ldr%?\t%|ip, %0", operands); | |
68d560d4 | 9741 | if (arm_arch5) |
07ec3810 | 9742 | output_asm_insn ("blx%?\t%|ip", operands); |
68d560d4 RE |
9743 | else |
9744 | { | |
9745 | output_asm_insn ("mov%?\t%|lr, %|pc", operands); | |
9746 | if (arm_arch4t) | |
9747 | output_asm_insn ("bx%?\t%|ip", operands); | |
9748 | else | |
9749 | output_asm_insn ("mov%?\t%|pc, %|ip", operands); | |
9750 | } | |
6ab5da80 | 9751 | } |
da6558fd NC |
9752 | else |
9753 | { | |
9754 | output_asm_insn ("mov%?\t%|lr, %|pc", operands); | |
9755 | output_asm_insn ("ldr%?\t%|pc, %0", operands); | |
9756 | } | |
9757 | ||
f3bb6135 RE |
9758 | return ""; |
9759 | } | |
ff9940b0 RE |
9760 | |
9761 | ||
3b684012 RE |
9762 | /* Output a move from arm registers to an fpa registers. |
9763 | OPERANDS[0] is an fpa register. | |
ff9940b0 | 9764 | OPERANDS[1] is the first registers of an arm register pair. */ |
cd2b33d0 | 9765 | const char * |
e32bac5b | 9766 | output_mov_long_double_fpa_from_arm (rtx *operands) |
ff9940b0 RE |
9767 | { |
9768 | int arm_reg0 = REGNO (operands[1]); | |
9769 | rtx ops[3]; | |
9770 | ||
e6d29d15 | 9771 | gcc_assert (arm_reg0 != IP_REGNUM); |
f3bb6135 | 9772 | |
43cffd11 RE |
9773 | ops[0] = gen_rtx_REG (SImode, arm_reg0); |
9774 | ops[1] = gen_rtx_REG (SImode, 1 + arm_reg0); | |
9775 | ops[2] = gen_rtx_REG (SImode, 2 + arm_reg0); | |
f676971a | 9776 | |
5b3e6663 | 9777 | output_asm_insn ("stm%(fd%)\t%|sp!, {%0, %1, %2}", ops); |
1d5473cb | 9778 | output_asm_insn ("ldf%?e\t%0, [%|sp], #12", operands); |
f676971a | 9779 | |
f3bb6135 RE |
9780 | return ""; |
9781 | } | |
ff9940b0 | 9782 | |
3b684012 | 9783 | /* Output a move from an fpa register to arm registers. |
ff9940b0 | 9784 | OPERANDS[0] is the first registers of an arm register pair. |
3b684012 | 9785 | OPERANDS[1] is an fpa register. */ |
cd2b33d0 | 9786 | const char * |
e32bac5b | 9787 | output_mov_long_double_arm_from_fpa (rtx *operands) |
ff9940b0 RE |
9788 | { |
9789 | int arm_reg0 = REGNO (operands[0]); | |
9790 | rtx ops[3]; | |
9791 | ||
e6d29d15 | 9792 | gcc_assert (arm_reg0 != IP_REGNUM); |
f3bb6135 | 9793 | |
43cffd11 RE |
9794 | ops[0] = gen_rtx_REG (SImode, arm_reg0); |
9795 | ops[1] = gen_rtx_REG (SImode, 1 + arm_reg0); | |
9796 | ops[2] = gen_rtx_REG (SImode, 2 + arm_reg0); | |
ff9940b0 | 9797 | |
1d5473cb | 9798 | output_asm_insn ("stf%?e\t%1, [%|sp, #-12]!", operands); |
5b3e6663 | 9799 | output_asm_insn ("ldm%(fd%)\t%|sp!, {%0, %1, %2}", ops); |
f3bb6135 RE |
9800 | return ""; |
9801 | } | |
ff9940b0 RE |
9802 | |
9803 | /* Output a move from arm registers to arm registers of a long double | |
9804 | OPERANDS[0] is the destination. | |
9805 | OPERANDS[1] is the source. */ | |
cd2b33d0 | 9806 | const char * |
e32bac5b | 9807 | output_mov_long_double_arm_from_arm (rtx *operands) |
ff9940b0 | 9808 | { |
6354dc9b | 9809 | /* We have to be careful here because the two might overlap. */ |
ff9940b0 RE |
9810 | int dest_start = REGNO (operands[0]); |
9811 | int src_start = REGNO (operands[1]); | |
9812 | rtx ops[2]; | |
9813 | int i; | |
9814 | ||
9815 | if (dest_start < src_start) | |
9816 | { | |
9817 | for (i = 0; i < 3; i++) | |
9818 | { | |
43cffd11 RE |
9819 | ops[0] = gen_rtx_REG (SImode, dest_start + i); |
9820 | ops[1] = gen_rtx_REG (SImode, src_start + i); | |
9997d19d | 9821 | output_asm_insn ("mov%?\t%0, %1", ops); |
ff9940b0 RE |
9822 | } |
9823 | } | |
9824 | else | |
9825 | { | |
9826 | for (i = 2; i >= 0; i--) | |
9827 | { | |
43cffd11 RE |
9828 | ops[0] = gen_rtx_REG (SImode, dest_start + i); |
9829 | ops[1] = gen_rtx_REG (SImode, src_start + i); | |
9997d19d | 9830 | output_asm_insn ("mov%?\t%0, %1", ops); |
ff9940b0 RE |
9831 | } |
9832 | } | |
f3bb6135 | 9833 | |
ff9940b0 RE |
9834 | return ""; |
9835 | } | |
9836 | ||
9837 | ||
3b684012 RE |
9838 | /* Output a move from arm registers to an fpa registers. |
9839 | OPERANDS[0] is an fpa register. | |
cce8749e | 9840 | OPERANDS[1] is the first registers of an arm register pair. */ |
cd2b33d0 | 9841 | const char * |
e32bac5b | 9842 | output_mov_double_fpa_from_arm (rtx *operands) |
cce8749e CH |
9843 | { |
9844 | int arm_reg0 = REGNO (operands[1]); | |
9845 | rtx ops[2]; | |
9846 | ||
e6d29d15 | 9847 | gcc_assert (arm_reg0 != IP_REGNUM); |
f676971a | 9848 | |
43cffd11 RE |
9849 | ops[0] = gen_rtx_REG (SImode, arm_reg0); |
9850 | ops[1] = gen_rtx_REG (SImode, 1 + arm_reg0); | |
5b3e6663 | 9851 | output_asm_insn ("stm%(fd%)\t%|sp!, {%0, %1}", ops); |
1d5473cb | 9852 | output_asm_insn ("ldf%?d\t%0, [%|sp], #8", operands); |
f3bb6135 RE |
9853 | return ""; |
9854 | } | |
cce8749e | 9855 | |
3b684012 | 9856 | /* Output a move from an fpa register to arm registers. |
cce8749e | 9857 | OPERANDS[0] is the first registers of an arm register pair. |
3b684012 | 9858 | OPERANDS[1] is an fpa register. */ |
cd2b33d0 | 9859 | const char * |
e32bac5b | 9860 | output_mov_double_arm_from_fpa (rtx *operands) |
cce8749e CH |
9861 | { |
9862 | int arm_reg0 = REGNO (operands[0]); | |
9863 | rtx ops[2]; | |
9864 | ||
e6d29d15 | 9865 | gcc_assert (arm_reg0 != IP_REGNUM); |
f3bb6135 | 9866 | |
43cffd11 RE |
9867 | ops[0] = gen_rtx_REG (SImode, arm_reg0); |
9868 | ops[1] = gen_rtx_REG (SImode, 1 + arm_reg0); | |
1d5473cb | 9869 | output_asm_insn ("stf%?d\t%1, [%|sp, #-8]!", operands); |
5b3e6663 | 9870 | output_asm_insn ("ldm%(fd%)\t%|sp!, {%0, %1}", ops); |
f3bb6135 RE |
9871 | return ""; |
9872 | } | |
cce8749e CH |
9873 | |
9874 | /* Output a move between double words. | |
9875 | It must be REG<-REG, REG<-CONST_DOUBLE, REG<-CONST_INT, REG<-MEM | |
9876 | or MEM<-REG and all MEMs must be offsettable addresses. */ | |
cd2b33d0 | 9877 | const char * |
e32bac5b | 9878 | output_move_double (rtx *operands) |
cce8749e CH |
9879 | { |
9880 | enum rtx_code code0 = GET_CODE (operands[0]); | |
9881 | enum rtx_code code1 = GET_CODE (operands[1]); | |
56636818 | 9882 | rtx otherops[3]; |
cce8749e CH |
9883 | |
9884 | if (code0 == REG) | |
9885 | { | |
9886 | int reg0 = REGNO (operands[0]); | |
9887 | ||
43cffd11 | 9888 | otherops[0] = gen_rtx_REG (SImode, 1 + reg0); |
f676971a | 9889 | |
e6d29d15 NS |
9890 | gcc_assert (code1 == MEM); /* Constraints should ensure this. */ |
9891 | ||
9892 | switch (GET_CODE (XEXP (operands[1], 0))) | |
cce8749e | 9893 | { |
e6d29d15 | 9894 | case REG: |
5dea0c19 PB |
9895 | if (TARGET_LDRD) |
9896 | output_asm_insn ("ldr%(d%)\t%0, [%m1]", operands); | |
9897 | else | |
9898 | output_asm_insn ("ldm%(ia%)\t%m1, %M0", operands); | |
e6d29d15 | 9899 | break; |
e0b92319 | 9900 | |
e6d29d15 NS |
9901 | case PRE_INC: |
9902 | gcc_assert (TARGET_LDRD); | |
5b3e6663 | 9903 | output_asm_insn ("ldr%(d%)\t%0, [%m1, #8]!", operands); |
e6d29d15 | 9904 | break; |
e0b92319 | 9905 | |
e6d29d15 | 9906 | case PRE_DEC: |
5b3e6663 PB |
9907 | if (TARGET_LDRD) |
9908 | output_asm_insn ("ldr%(d%)\t%0, [%m1, #-8]!", operands); | |
9909 | else | |
9910 | output_asm_insn ("ldm%(db%)\t%m1!, %M0", operands); | |
e6d29d15 | 9911 | break; |
e0b92319 | 9912 | |
e6d29d15 | 9913 | case POST_INC: |
5dea0c19 PB |
9914 | if (TARGET_LDRD) |
9915 | output_asm_insn ("ldr%(d%)\t%0, [%m1], #8", operands); | |
9916 | else | |
9917 | output_asm_insn ("ldm%(ia%)\t%m1!, %M0", operands); | |
e6d29d15 | 9918 | break; |
e0b92319 | 9919 | |
e6d29d15 NS |
9920 | case POST_DEC: |
9921 | gcc_assert (TARGET_LDRD); | |
5b3e6663 | 9922 | output_asm_insn ("ldr%(d%)\t%0, [%m1], #-8", operands); |
e6d29d15 | 9923 | break; |
e0b92319 | 9924 | |
e6d29d15 NS |
9925 | case PRE_MODIFY: |
9926 | case POST_MODIFY: | |
9927 | otherops[0] = operands[0]; | |
9928 | otherops[1] = XEXP (XEXP (XEXP (operands[1], 0), 1), 0); | |
9929 | otherops[2] = XEXP (XEXP (XEXP (operands[1], 0), 1), 1); | |
e0b92319 | 9930 | |
e6d29d15 | 9931 | if (GET_CODE (XEXP (operands[1], 0)) == PRE_MODIFY) |
cce8749e | 9932 | { |
e6d29d15 | 9933 | if (reg_overlap_mentioned_p (otherops[0], otherops[2])) |
fdd695fd | 9934 | { |
e6d29d15 NS |
9935 | /* Registers overlap so split out the increment. */ |
9936 | output_asm_insn ("add%?\t%1, %1, %2", otherops); | |
5b3e6663 | 9937 | output_asm_insn ("ldr%(d%)\t%0, [%1] @split", otherops); |
fdd695fd PB |
9938 | } |
9939 | else | |
fe2d934b PB |
9940 | { |
9941 | /* IWMMXT allows offsets larger than ldrd can handle, | |
9942 | fix these up with a pair of ldr. */ | |
9943 | if (GET_CODE (otherops[2]) == CONST_INT | |
9944 | && (INTVAL(otherops[2]) <= -256 | |
9945 | || INTVAL(otherops[2]) >= 256)) | |
9946 | { | |
9947 | output_asm_insn ("ldr%?\t%0, [%1, %2]!", otherops); | |
9948 | otherops[0] = gen_rtx_REG (SImode, 1 + reg0); | |
9949 | output_asm_insn ("ldr%?\t%0, [%1, #4]", otherops); | |
9950 | } | |
9951 | else | |
9952 | output_asm_insn ("ldr%(d%)\t%0, [%1, %2]!", otherops); | |
9953 | } | |
e6d29d15 NS |
9954 | } |
9955 | else | |
9956 | { | |
fe2d934b PB |
9957 | /* IWMMXT allows offsets larger than ldrd can handle, |
9958 | fix these up with a pair of ldr. */ | |
9959 | if (GET_CODE (otherops[2]) == CONST_INT | |
9960 | && (INTVAL(otherops[2]) <= -256 | |
9961 | || INTVAL(otherops[2]) >= 256)) | |
9962 | { | |
9963 | otherops[0] = gen_rtx_REG (SImode, 1 + reg0); | |
9964 | output_asm_insn ("ldr%?\t%0, [%1, #4]", otherops); | |
9965 | otherops[0] = operands[0]; | |
9966 | output_asm_insn ("ldr%?\t%0, [%1], %2", otherops); | |
9967 | } | |
9968 | else | |
9969 | /* We only allow constant increments, so this is safe. */ | |
9970 | output_asm_insn ("ldr%(d%)\t%0, [%1], %2", otherops); | |
e6d29d15 NS |
9971 | } |
9972 | break; | |
e0b92319 | 9973 | |
e6d29d15 NS |
9974 | case LABEL_REF: |
9975 | case CONST: | |
5dea0c19 PB |
9976 | /* We might be able to use ldrd %0, %1 here. However the range is |
9977 | different to ldr/adr, and it is broken on some ARMv7-M | |
9978 | implementations. */ | |
e6d29d15 | 9979 | output_asm_insn ("adr%?\t%0, %1", operands); |
5dea0c19 PB |
9980 | if (TARGET_LDRD) |
9981 | output_asm_insn ("ldr%(d%)\t%0, [%0]", operands); | |
9982 | else | |
9983 | output_asm_insn ("ldm%(ia%)\t%0, %M0", operands); | |
e6d29d15 | 9984 | break; |
e0b92319 | 9985 | |
5b3e6663 | 9986 | /* ??? This needs checking for thumb2. */ |
e6d29d15 NS |
9987 | default: |
9988 | if (arm_add_operand (XEXP (XEXP (operands[1], 0), 1), | |
9989 | GET_MODE (XEXP (XEXP (operands[1], 0), 1)))) | |
9990 | { | |
9991 | otherops[0] = operands[0]; | |
9992 | otherops[1] = XEXP (XEXP (operands[1], 0), 0); | |
9993 | otherops[2] = XEXP (XEXP (operands[1], 0), 1); | |
e0b92319 | 9994 | |
e6d29d15 | 9995 | if (GET_CODE (XEXP (operands[1], 0)) == PLUS) |
fdd695fd | 9996 | { |
5dea0c19 | 9997 | if (GET_CODE (otherops[2]) == CONST_INT && !TARGET_LDRD) |
2b835d68 | 9998 | { |
e6d29d15 | 9999 | switch ((int) INTVAL (otherops[2])) |
2b835d68 | 10000 | { |
e6d29d15 | 10001 | case -8: |
5b3e6663 | 10002 | output_asm_insn ("ldm%(db%)\t%1, %M0", otherops); |
e6d29d15 NS |
10003 | return ""; |
10004 | case -4: | |
5b3e6663 PB |
10005 | if (TARGET_THUMB2) |
10006 | break; | |
10007 | output_asm_insn ("ldm%(da%)\t%1, %M0", otherops); | |
e6d29d15 NS |
10008 | return ""; |
10009 | case 4: | |
5b3e6663 PB |
10010 | if (TARGET_THUMB2) |
10011 | break; | |
10012 | output_asm_insn ("ldm%(ib%)\t%1, %M0", otherops); | |
e6d29d15 | 10013 | return ""; |
fdd695fd | 10014 | } |
e6d29d15 NS |
10015 | } |
10016 | if (TARGET_LDRD | |
10017 | && (GET_CODE (otherops[2]) == REG | |
10018 | || (GET_CODE (otherops[2]) == CONST_INT | |
10019 | && INTVAL (otherops[2]) > -256 | |
10020 | && INTVAL (otherops[2]) < 256))) | |
10021 | { | |
10022 | if (reg_overlap_mentioned_p (otherops[0], | |
10023 | otherops[2])) | |
fdd695fd | 10024 | { |
e6d29d15 NS |
10025 | /* Swap base and index registers over to |
10026 | avoid a conflict. */ | |
10027 | otherops[1] = XEXP (XEXP (operands[1], 0), 1); | |
10028 | otherops[2] = XEXP (XEXP (operands[1], 0), 0); | |
fdd695fd | 10029 | } |
e6d29d15 NS |
10030 | /* If both registers conflict, it will usually |
10031 | have been fixed by a splitter. */ | |
10032 | if (reg_overlap_mentioned_p (otherops[0], otherops[2])) | |
fdd695fd | 10033 | { |
e6d29d15 | 10034 | output_asm_insn ("add%?\t%1, %1, %2", otherops); |
5b3e6663 | 10035 | output_asm_insn ("ldr%(d%)\t%0, [%1]", |
e6d29d15 | 10036 | otherops); |
2b835d68 RE |
10037 | } |
10038 | else | |
5b3e6663 | 10039 | output_asm_insn ("ldr%(d%)\t%0, [%1, %2]", otherops); |
e6d29d15 | 10040 | return ""; |
2b835d68 | 10041 | } |
e0b92319 | 10042 | |
e6d29d15 | 10043 | if (GET_CODE (otherops[2]) == CONST_INT) |
2b835d68 | 10044 | { |
e6d29d15 NS |
10045 | if (!(const_ok_for_arm (INTVAL (otherops[2])))) |
10046 | output_asm_insn ("sub%?\t%0, %1, #%n2", otherops); | |
10047 | else | |
10048 | output_asm_insn ("add%?\t%0, %1, %2", otherops); | |
2b835d68 RE |
10049 | } |
10050 | else | |
e6d29d15 NS |
10051 | output_asm_insn ("add%?\t%0, %1, %2", otherops); |
10052 | } | |
10053 | else | |
10054 | output_asm_insn ("sub%?\t%0, %1, %2", otherops); | |
10055 | ||
5dea0c19 PB |
10056 | if (TARGET_LDRD) |
10057 | return "ldr%(d%)\t%0, [%0]"; | |
10058 | ||
5b3e6663 | 10059 | return "ldm%(ia%)\t%0, %M0"; |
e6d29d15 NS |
10060 | } |
10061 | else | |
10062 | { | |
10063 | otherops[1] = adjust_address (operands[1], SImode, 4); | |
10064 | /* Take care of overlapping base/data reg. */ | |
10065 | if (reg_mentioned_p (operands[0], operands[1])) | |
10066 | { | |
10067 | output_asm_insn ("ldr%?\t%0, %1", otherops); | |
10068 | output_asm_insn ("ldr%?\t%0, %1", operands); | |
10069 | } | |
10070 | else | |
10071 | { | |
10072 | output_asm_insn ("ldr%?\t%0, %1", operands); | |
10073 | output_asm_insn ("ldr%?\t%0, %1", otherops); | |
cce8749e CH |
10074 | } |
10075 | } | |
10076 | } | |
cce8749e | 10077 | } |
e6d29d15 | 10078 | else |
cce8749e | 10079 | { |
e6d29d15 NS |
10080 | /* Constraints should ensure this. */ |
10081 | gcc_assert (code0 == MEM && code1 == REG); | |
10082 | gcc_assert (REGNO (operands[1]) != IP_REGNUM); | |
2b835d68 | 10083 | |
ff9940b0 RE |
10084 | switch (GET_CODE (XEXP (operands[0], 0))) |
10085 | { | |
10086 | case REG: | |
5dea0c19 PB |
10087 | if (TARGET_LDRD) |
10088 | output_asm_insn ("str%(d%)\t%1, [%m0]", operands); | |
10089 | else | |
10090 | output_asm_insn ("stm%(ia%)\t%m0, %M1", operands); | |
ff9940b0 | 10091 | break; |
2b835d68 | 10092 | |
ff9940b0 | 10093 | case PRE_INC: |
e6d29d15 | 10094 | gcc_assert (TARGET_LDRD); |
5b3e6663 | 10095 | output_asm_insn ("str%(d%)\t%1, [%m0, #8]!", operands); |
ff9940b0 | 10096 | break; |
2b835d68 | 10097 | |
ff9940b0 | 10098 | case PRE_DEC: |
5b3e6663 PB |
10099 | if (TARGET_LDRD) |
10100 | output_asm_insn ("str%(d%)\t%1, [%m0, #-8]!", operands); | |
10101 | else | |
10102 | output_asm_insn ("stm%(db%)\t%m0!, %M1", operands); | |
ff9940b0 | 10103 | break; |
2b835d68 | 10104 | |
ff9940b0 | 10105 | case POST_INC: |
5dea0c19 PB |
10106 | if (TARGET_LDRD) |
10107 | output_asm_insn ("str%(d%)\t%1, [%m0], #8", operands); | |
10108 | else | |
10109 | output_asm_insn ("stm%(ia%)\t%m0!, %M1", operands); | |
ff9940b0 | 10110 | break; |
2b835d68 | 10111 | |
ff9940b0 | 10112 | case POST_DEC: |
e6d29d15 | 10113 | gcc_assert (TARGET_LDRD); |
5b3e6663 | 10114 | output_asm_insn ("str%(d%)\t%1, [%m0], #-8", operands); |
fdd695fd PB |
10115 | break; |
10116 | ||
10117 | case PRE_MODIFY: | |
10118 | case POST_MODIFY: | |
10119 | otherops[0] = operands[1]; | |
10120 | otherops[1] = XEXP (XEXP (XEXP (operands[0], 0), 1), 0); | |
10121 | otherops[2] = XEXP (XEXP (XEXP (operands[0], 0), 1), 1); | |
10122 | ||
fe2d934b PB |
10123 | /* IWMMXT allows offsets larger than ldrd can handle, |
10124 | fix these up with a pair of ldr. */ | |
10125 | if (GET_CODE (otherops[2]) == CONST_INT | |
10126 | && (INTVAL(otherops[2]) <= -256 | |
10127 | || INTVAL(otherops[2]) >= 256)) | |
10128 | { | |
10129 | rtx reg1; | |
10130 | reg1 = gen_rtx_REG (SImode, 1 + REGNO (operands[1])); | |
10131 | if (GET_CODE (XEXP (operands[0], 0)) == PRE_MODIFY) | |
10132 | { | |
10133 | output_asm_insn ("ldr%?\t%0, [%1, %2]!", otherops); | |
10134 | otherops[0] = reg1; | |
10135 | output_asm_insn ("ldr%?\t%0, [%1, #4]", otherops); | |
10136 | } | |
10137 | else | |
10138 | { | |
10139 | otherops[0] = reg1; | |
10140 | output_asm_insn ("ldr%?\t%0, [%1, #4]", otherops); | |
10141 | otherops[0] = operands[1]; | |
10142 | output_asm_insn ("ldr%?\t%0, [%1], %2", otherops); | |
10143 | } | |
10144 | } | |
10145 | else if (GET_CODE (XEXP (operands[0], 0)) == PRE_MODIFY) | |
5b3e6663 | 10146 | output_asm_insn ("str%(d%)\t%0, [%1, %2]!", otherops); |
fdd695fd | 10147 | else |
5b3e6663 | 10148 | output_asm_insn ("str%(d%)\t%0, [%1], %2", otherops); |
ff9940b0 | 10149 | break; |
2b835d68 RE |
10150 | |
10151 | case PLUS: | |
fdd695fd | 10152 | otherops[2] = XEXP (XEXP (operands[0], 0), 1); |
5dea0c19 | 10153 | if (GET_CODE (otherops[2]) == CONST_INT && !TARGET_LDRD) |
2b835d68 | 10154 | { |
06bea5aa | 10155 | switch ((int) INTVAL (XEXP (XEXP (operands[0], 0), 1))) |
2b835d68 RE |
10156 | { |
10157 | case -8: | |
5b3e6663 | 10158 | output_asm_insn ("stm%(db%)\t%m0, %M1", operands); |
2b835d68 RE |
10159 | return ""; |
10160 | ||
10161 | case -4: | |
5b3e6663 PB |
10162 | if (TARGET_THUMB2) |
10163 | break; | |
10164 | output_asm_insn ("stm%(da%)\t%m0, %M1", operands); | |
2b835d68 RE |
10165 | return ""; |
10166 | ||
10167 | case 4: | |
5b3e6663 PB |
10168 | if (TARGET_THUMB2) |
10169 | break; | |
10170 | output_asm_insn ("stm%(ib%)\t%m0, %M1", operands); | |
2b835d68 RE |
10171 | return ""; |
10172 | } | |
10173 | } | |
fdd695fd PB |
10174 | if (TARGET_LDRD |
10175 | && (GET_CODE (otherops[2]) == REG | |
10176 | || (GET_CODE (otherops[2]) == CONST_INT | |
10177 | && INTVAL (otherops[2]) > -256 | |
10178 | && INTVAL (otherops[2]) < 256))) | |
10179 | { | |
10180 | otherops[0] = operands[1]; | |
10181 | otherops[1] = XEXP (XEXP (operands[0], 0), 0); | |
5b3e6663 | 10182 | output_asm_insn ("str%(d%)\t%0, [%1, %2]", otherops); |
fdd695fd PB |
10183 | return ""; |
10184 | } | |
2b835d68 RE |
10185 | /* Fall through */ |
10186 | ||
ff9940b0 | 10187 | default: |
a4a37b30 | 10188 | otherops[0] = adjust_address (operands[0], SImode, 4); |
43cffd11 | 10189 | otherops[1] = gen_rtx_REG (SImode, 1 + REGNO (operands[1])); |
9997d19d RE |
10190 | output_asm_insn ("str%?\t%1, %0", operands); |
10191 | output_asm_insn ("str%?\t%1, %0", otherops); | |
cce8749e CH |
10192 | } |
10193 | } | |
cce8749e | 10194 | |
9997d19d RE |
10195 | return ""; |
10196 | } | |
cce8749e | 10197 | |
88f77cba JB |
10198 | /* Output a move, load or store for quad-word vectors in ARM registers. Only |
10199 | handles MEMs accepted by neon_vector_mem_operand with CORE=true. */ | |
5b3e6663 PB |
10200 | |
10201 | const char * | |
88f77cba | 10202 | output_move_quad (rtx *operands) |
5b3e6663 | 10203 | { |
88f77cba JB |
10204 | if (REG_P (operands[0])) |
10205 | { | |
10206 | /* Load, or reg->reg move. */ | |
5b3e6663 | 10207 | |
88f77cba JB |
10208 | if (MEM_P (operands[1])) |
10209 | { | |
10210 | switch (GET_CODE (XEXP (operands[1], 0))) | |
10211 | { | |
10212 | case REG: | |
10213 | output_asm_insn ("ldm%(ia%)\t%m1, %M0", operands); | |
10214 | break; | |
10215 | ||
10216 | case LABEL_REF: | |
10217 | case CONST: | |
10218 | output_asm_insn ("adr%?\t%0, %1", operands); | |
10219 | output_asm_insn ("ldm%(ia%)\t%0, %M0", operands); | |
10220 | break; | |
10221 | ||
10222 | default: | |
10223 | gcc_unreachable (); | |
10224 | } | |
10225 | } | |
10226 | else | |
10227 | { | |
10228 | rtx ops[2]; | |
10229 | int dest, src, i; | |
5b3e6663 | 10230 | |
88f77cba | 10231 | gcc_assert (REG_P (operands[1])); |
5b3e6663 | 10232 | |
88f77cba JB |
10233 | dest = REGNO (operands[0]); |
10234 | src = REGNO (operands[1]); | |
5b3e6663 | 10235 | |
88f77cba JB |
10236 | /* This seems pretty dumb, but hopefully GCC won't try to do it |
10237 | very often. */ | |
10238 | if (dest < src) | |
10239 | for (i = 0; i < 4; i++) | |
10240 | { | |
10241 | ops[0] = gen_rtx_REG (SImode, dest + i); | |
10242 | ops[1] = gen_rtx_REG (SImode, src + i); | |
10243 | output_asm_insn ("mov%?\t%0, %1", ops); | |
10244 | } | |
10245 | else | |
10246 | for (i = 3; i >= 0; i--) | |
10247 | { | |
10248 | ops[0] = gen_rtx_REG (SImode, dest + i); | |
10249 | ops[1] = gen_rtx_REG (SImode, src + i); | |
10250 | output_asm_insn ("mov%?\t%0, %1", ops); | |
10251 | } | |
10252 | } | |
10253 | } | |
10254 | else | |
10255 | { | |
10256 | gcc_assert (MEM_P (operands[0])); | |
10257 | gcc_assert (REG_P (operands[1])); | |
10258 | gcc_assert (!reg_overlap_mentioned_p (operands[1], operands[0])); | |
10259 | ||
10260 | switch (GET_CODE (XEXP (operands[0], 0))) | |
10261 | { | |
10262 | case REG: | |
10263 | output_asm_insn ("stm%(ia%)\t%m0, %M1", operands); | |
10264 | break; | |
10265 | ||
10266 | default: | |
10267 | gcc_unreachable (); | |
10268 | } | |
10269 | } | |
10270 | ||
10271 | return ""; | |
10272 | } | |
10273 | ||
10274 | /* Output a VFP load or store instruction. */ | |
10275 | ||
10276 | const char * | |
10277 | output_move_vfp (rtx *operands) | |
10278 | { | |
10279 | rtx reg, mem, addr, ops[2]; | |
10280 | int load = REG_P (operands[0]); | |
10281 | int dp = GET_MODE_SIZE (GET_MODE (operands[0])) == 8; | |
10282 | int integer_p = GET_MODE_CLASS (GET_MODE (operands[0])) == MODE_INT; | |
10283 | const char *template; | |
10284 | char buff[50]; | |
10285 | enum machine_mode mode; | |
10286 | ||
10287 | reg = operands[!load]; | |
10288 | mem = operands[load]; | |
10289 | ||
10290 | mode = GET_MODE (reg); | |
10291 | ||
10292 | gcc_assert (REG_P (reg)); | |
10293 | gcc_assert (IS_VFP_REGNUM (REGNO (reg))); | |
10294 | gcc_assert (mode == SFmode | |
10295 | || mode == DFmode | |
10296 | || mode == SImode | |
10297 | || mode == DImode | |
10298 | || (TARGET_NEON && VALID_NEON_DREG_MODE (mode))); | |
10299 | gcc_assert (MEM_P (mem)); | |
10300 | ||
10301 | addr = XEXP (mem, 0); | |
10302 | ||
10303 | switch (GET_CODE (addr)) | |
10304 | { | |
10305 | case PRE_DEC: | |
10306 | template = "f%smdb%c%%?\t%%0!, {%%%s1}%s"; | |
10307 | ops[0] = XEXP (addr, 0); | |
10308 | ops[1] = reg; | |
5b3e6663 PB |
10309 | break; |
10310 | ||
10311 | case POST_INC: | |
10312 | template = "f%smia%c%%?\t%%0!, {%%%s1}%s"; | |
10313 | ops[0] = XEXP (addr, 0); | |
10314 | ops[1] = reg; | |
10315 | break; | |
10316 | ||
10317 | default: | |
10318 | template = "f%s%c%%?\t%%%s0, %%1%s"; | |
10319 | ops[0] = reg; | |
10320 | ops[1] = mem; | |
10321 | break; | |
10322 | } | |
10323 | ||
10324 | sprintf (buff, template, | |
10325 | load ? "ld" : "st", | |
10326 | dp ? 'd' : 's', | |
10327 | dp ? "P" : "", | |
10328 | integer_p ? "\t%@ int" : ""); | |
10329 | output_asm_insn (buff, ops); | |
10330 | ||
10331 | return ""; | |
10332 | } | |
10333 | ||
88f77cba JB |
10334 | /* Output a Neon quad-word load or store, or a load or store for |
10335 | larger structure modes. We could also support post-modify forms using | |
10336 | VLD1/VST1 (for the vectorizer, and perhaps otherwise), but we don't do that | |
10337 | yet. | |
10338 | WARNING: The ordering of elements in memory is weird in big-endian mode, | |
10339 | because we use VSTM instead of VST1, to make it easy to make vector stores | |
10340 | via ARM registers write values in the same order as stores direct from Neon | |
10341 | registers. For example, the byte ordering of a quadword vector with 16-byte | |
10342 | elements like this: | |
10343 | ||
10344 | [e7:e6:e5:e4:e3:e2:e1:e0] (highest-numbered element first) | |
10345 | ||
10346 | will be (with lowest address first, h = most-significant byte, | |
10347 | l = least-significant byte of element): | |
10348 | ||
10349 | [e3h, e3l, e2h, e2l, e1h, e1l, e0h, e0l, | |
10350 | e7h, e7l, e6h, e6l, e5h, e5l, e4h, e4l] | |
10351 | ||
10352 | When necessary, quadword registers (dN, dN+1) are moved to ARM registers from | |
10353 | rN in the order: | |
10354 | ||
10355 | dN -> (rN+1, rN), dN+1 -> (rN+3, rN+2) | |
10356 | ||
10357 | So that STM/LDM can be used on vectors in ARM registers, and the same memory | |
10358 | layout will result as if VSTM/VLDM were used. */ | |
10359 | ||
10360 | const char * | |
10361 | output_move_neon (rtx *operands) | |
10362 | { | |
10363 | rtx reg, mem, addr, ops[2]; | |
10364 | int regno, load = REG_P (operands[0]); | |
10365 | const char *template; | |
10366 | char buff[50]; | |
10367 | enum machine_mode mode; | |
10368 | ||
10369 | reg = operands[!load]; | |
10370 | mem = operands[load]; | |
10371 | ||
10372 | mode = GET_MODE (reg); | |
10373 | ||
10374 | gcc_assert (REG_P (reg)); | |
10375 | regno = REGNO (reg); | |
10376 | gcc_assert (VFP_REGNO_OK_FOR_DOUBLE (regno) | |
10377 | || NEON_REGNO_OK_FOR_QUAD (regno)); | |
10378 | gcc_assert (VALID_NEON_DREG_MODE (mode) | |
10379 | || VALID_NEON_QREG_MODE (mode) | |
10380 | || VALID_NEON_STRUCT_MODE (mode)); | |
10381 | gcc_assert (MEM_P (mem)); | |
10382 | ||
10383 | addr = XEXP (mem, 0); | |
10384 | ||
10385 | /* Strip off const from addresses like (const (plus (...))). */ | |
10386 | if (GET_CODE (addr) == CONST && GET_CODE (XEXP (addr, 0)) == PLUS) | |
10387 | addr = XEXP (addr, 0); | |
10388 | ||
10389 | switch (GET_CODE (addr)) | |
10390 | { | |
10391 | case POST_INC: | |
10392 | template = "v%smia%%?\t%%0!, %%h1"; | |
10393 | ops[0] = XEXP (addr, 0); | |
10394 | ops[1] = reg; | |
10395 | break; | |
10396 | ||
10397 | case POST_MODIFY: | |
10398 | /* FIXME: Not currently enabled in neon_vector_mem_operand. */ | |
10399 | gcc_unreachable (); | |
10400 | ||
10401 | case LABEL_REF: | |
10402 | case PLUS: | |
10403 | { | |
10404 | int nregs = HARD_REGNO_NREGS (REGNO (reg), mode) / 2; | |
10405 | int i; | |
10406 | int overlap = -1; | |
10407 | for (i = 0; i < nregs; i++) | |
10408 | { | |
10409 | /* We're only using DImode here because it's a convenient size. */ | |
10410 | ops[0] = gen_rtx_REG (DImode, REGNO (reg) + 2 * i); | |
10411 | ops[1] = adjust_address (mem, SImode, 8 * i); | |
10412 | if (reg_overlap_mentioned_p (ops[0], mem)) | |
10413 | { | |
10414 | gcc_assert (overlap == -1); | |
10415 | overlap = i; | |
10416 | } | |
10417 | else | |
10418 | { | |
10419 | sprintf (buff, "v%sr%%?\t%%P0, %%1", load ? "ld" : "st"); | |
10420 | output_asm_insn (buff, ops); | |
10421 | } | |
10422 | } | |
10423 | if (overlap != -1) | |
10424 | { | |
10425 | ops[0] = gen_rtx_REG (DImode, REGNO (reg) + 2 * overlap); | |
10426 | ops[1] = adjust_address (mem, SImode, 8 * overlap); | |
10427 | sprintf (buff, "v%sr%%?\t%%P0, %%1", load ? "ld" : "st"); | |
10428 | output_asm_insn (buff, ops); | |
10429 | } | |
10430 | ||
10431 | return ""; | |
10432 | } | |
10433 | ||
10434 | default: | |
10435 | template = "v%smia%%?\t%%m0, %%h1"; | |
10436 | ops[0] = mem; | |
10437 | ops[1] = reg; | |
10438 | } | |
10439 | ||
10440 | sprintf (buff, template, load ? "ld" : "st"); | |
10441 | output_asm_insn (buff, ops); | |
10442 | ||
10443 | return ""; | |
10444 | } | |
10445 | ||
1d6e90ac NC |
10446 | /* Output an ADD r, s, #n where n may be too big for one instruction. |
10447 | If adding zero to one register, output nothing. */ | |
cd2b33d0 | 10448 | const char * |
e32bac5b | 10449 | output_add_immediate (rtx *operands) |
cce8749e | 10450 | { |
f3bb6135 | 10451 | HOST_WIDE_INT n = INTVAL (operands[2]); |
cce8749e CH |
10452 | |
10453 | if (n != 0 || REGNO (operands[0]) != REGNO (operands[1])) | |
10454 | { | |
10455 | if (n < 0) | |
10456 | output_multi_immediate (operands, | |
9997d19d RE |
10457 | "sub%?\t%0, %1, %2", "sub%?\t%0, %0, %2", 2, |
10458 | -n); | |
cce8749e CH |
10459 | else |
10460 | output_multi_immediate (operands, | |
9997d19d RE |
10461 | "add%?\t%0, %1, %2", "add%?\t%0, %0, %2", 2, |
10462 | n); | |
cce8749e | 10463 | } |
f3bb6135 RE |
10464 | |
10465 | return ""; | |
10466 | } | |
cce8749e | 10467 | |
cce8749e CH |
10468 | /* Output a multiple immediate operation. |
10469 | OPERANDS is the vector of operands referred to in the output patterns. | |
10470 | INSTR1 is the output pattern to use for the first constant. | |
10471 | INSTR2 is the output pattern to use for subsequent constants. | |
10472 | IMMED_OP is the index of the constant slot in OPERANDS. | |
10473 | N is the constant value. */ | |
cd2b33d0 | 10474 | static const char * |
e32bac5b RE |
10475 | output_multi_immediate (rtx *operands, const char *instr1, const char *instr2, |
10476 | int immed_op, HOST_WIDE_INT n) | |
cce8749e | 10477 | { |
f3bb6135 | 10478 | #if HOST_BITS_PER_WIDE_INT > 32 |
30cf4896 | 10479 | n &= 0xffffffff; |
f3bb6135 RE |
10480 | #endif |
10481 | ||
cce8749e CH |
10482 | if (n == 0) |
10483 | { | |
1d6e90ac | 10484 | /* Quick and easy output. */ |
cce8749e | 10485 | operands[immed_op] = const0_rtx; |
1d6e90ac | 10486 | output_asm_insn (instr1, operands); |
cce8749e CH |
10487 | } |
10488 | else | |
10489 | { | |
10490 | int i; | |
cd2b33d0 | 10491 | const char * instr = instr1; |
cce8749e | 10492 | |
6354dc9b | 10493 | /* Note that n is never zero here (which would give no output). */ |
cce8749e CH |
10494 | for (i = 0; i < 32; i += 2) |
10495 | { | |
10496 | if (n & (3 << i)) | |
10497 | { | |
f3bb6135 RE |
10498 | operands[immed_op] = GEN_INT (n & (255 << i)); |
10499 | output_asm_insn (instr, operands); | |
cce8749e CH |
10500 | instr = instr2; |
10501 | i += 6; | |
10502 | } | |
10503 | } | |
10504 | } | |
f676971a | 10505 | |
f3bb6135 | 10506 | return ""; |
9997d19d | 10507 | } |
cce8749e | 10508 | |
5b3e6663 PB |
10509 | /* Return the name of a shifter operation. */ |
10510 | static const char * | |
10511 | arm_shift_nmem(enum rtx_code code) | |
10512 | { | |
10513 | switch (code) | |
10514 | { | |
10515 | case ASHIFT: | |
10516 | return ARM_LSL_NAME; | |
10517 | ||
10518 | case ASHIFTRT: | |
10519 | return "asr"; | |
10520 | ||
10521 | case LSHIFTRT: | |
10522 | return "lsr"; | |
10523 | ||
10524 | case ROTATERT: | |
10525 | return "ror"; | |
10526 | ||
10527 | default: | |
10528 | abort(); | |
10529 | } | |
10530 | } | |
10531 | ||
cce8749e CH |
10532 | /* Return the appropriate ARM instruction for the operation code. |
10533 | The returned result should not be overwritten. OP is the rtx of the | |
10534 | operation. SHIFT_FIRST_ARG is TRUE if the first argument of the operator | |
10535 | was shifted. */ | |
cd2b33d0 | 10536 | const char * |
e32bac5b | 10537 | arithmetic_instr (rtx op, int shift_first_arg) |
cce8749e | 10538 | { |
9997d19d | 10539 | switch (GET_CODE (op)) |
cce8749e CH |
10540 | { |
10541 | case PLUS: | |
f3bb6135 RE |
10542 | return "add"; |
10543 | ||
cce8749e | 10544 | case MINUS: |
f3bb6135 RE |
10545 | return shift_first_arg ? "rsb" : "sub"; |
10546 | ||
cce8749e | 10547 | case IOR: |
f3bb6135 RE |
10548 | return "orr"; |
10549 | ||
cce8749e | 10550 | case XOR: |
f3bb6135 RE |
10551 | return "eor"; |
10552 | ||
cce8749e | 10553 | case AND: |
f3bb6135 RE |
10554 | return "and"; |
10555 | ||
5b3e6663 PB |
10556 | case ASHIFT: |
10557 | case ASHIFTRT: | |
10558 | case LSHIFTRT: | |
10559 | case ROTATERT: | |
10560 | return arm_shift_nmem(GET_CODE(op)); | |
10561 | ||
cce8749e | 10562 | default: |
e6d29d15 | 10563 | gcc_unreachable (); |
cce8749e | 10564 | } |
f3bb6135 | 10565 | } |
cce8749e | 10566 | |
cce8749e CH |
10567 | /* Ensure valid constant shifts and return the appropriate shift mnemonic |
10568 | for the operation code. The returned result should not be overwritten. | |
10569 | OP is the rtx code of the shift. | |
9997d19d | 10570 | On exit, *AMOUNTP will be -1 if the shift is by a register, or a constant |
6354dc9b | 10571 | shift. */ |
cd2b33d0 | 10572 | static const char * |
e32bac5b | 10573 | shift_op (rtx op, HOST_WIDE_INT *amountp) |
cce8749e | 10574 | { |
cd2b33d0 | 10575 | const char * mnem; |
e2c671ba | 10576 | enum rtx_code code = GET_CODE (op); |
cce8749e | 10577 | |
e6d29d15 NS |
10578 | switch (GET_CODE (XEXP (op, 1))) |
10579 | { | |
10580 | case REG: | |
10581 | case SUBREG: | |
10582 | *amountp = -1; | |
10583 | break; | |
10584 | ||
10585 | case CONST_INT: | |
10586 | *amountp = INTVAL (XEXP (op, 1)); | |
10587 | break; | |
10588 | ||
10589 | default: | |
10590 | gcc_unreachable (); | |
10591 | } | |
9997d19d | 10592 | |
e2c671ba | 10593 | switch (code) |
cce8749e | 10594 | { |
a657c98a | 10595 | case ROTATE: |
e6d29d15 | 10596 | gcc_assert (*amountp != -1); |
a657c98a | 10597 | *amountp = 32 - *amountp; |
5b3e6663 | 10598 | code = ROTATERT; |
a657c98a RE |
10599 | |
10600 | /* Fall through. */ | |
10601 | ||
5b3e6663 PB |
10602 | case ASHIFT: |
10603 | case ASHIFTRT: | |
10604 | case LSHIFTRT: | |
9997d19d | 10605 | case ROTATERT: |
5b3e6663 | 10606 | mnem = arm_shift_nmem(code); |
9997d19d RE |
10607 | break; |
10608 | ||
ff9940b0 | 10609 | case MULT: |
e2c671ba RE |
10610 | /* We never have to worry about the amount being other than a |
10611 | power of 2, since this case can never be reloaded from a reg. */ | |
e6d29d15 NS |
10612 | gcc_assert (*amountp != -1); |
10613 | *amountp = int_log2 (*amountp); | |
5b3e6663 | 10614 | return ARM_LSL_NAME; |
f3bb6135 | 10615 | |
cce8749e | 10616 | default: |
e6d29d15 | 10617 | gcc_unreachable (); |
cce8749e CH |
10618 | } |
10619 | ||
e2c671ba RE |
10620 | if (*amountp != -1) |
10621 | { | |
10622 | /* This is not 100% correct, but follows from the desire to merge | |
10623 | multiplication by a power of 2 with the recognizer for a | |
5b3e6663 | 10624 | shift. >=32 is not a valid shift for "lsl", so we must try and |
e2c671ba | 10625 | output a shift that produces the correct arithmetical result. |
ddd5a7c1 | 10626 | Using lsr #32 is identical except for the fact that the carry bit |
f676971a | 10627 | is not set correctly if we set the flags; but we never use the |
e2c671ba RE |
10628 | carry bit from such an operation, so we can ignore that. */ |
10629 | if (code == ROTATERT) | |
1d6e90ac NC |
10630 | /* Rotate is just modulo 32. */ |
10631 | *amountp &= 31; | |
e2c671ba RE |
10632 | else if (*amountp != (*amountp & 31)) |
10633 | { | |
10634 | if (code == ASHIFT) | |
10635 | mnem = "lsr"; | |
10636 | *amountp = 32; | |
10637 | } | |
10638 | ||
10639 | /* Shifts of 0 are no-ops. */ | |
10640 | if (*amountp == 0) | |
10641 | return NULL; | |
f676971a | 10642 | } |
e2c671ba | 10643 | |
9997d19d RE |
10644 | return mnem; |
10645 | } | |
cce8749e | 10646 | |
6354dc9b | 10647 | /* Obtain the shift from the POWER of two. */ |
1d6e90ac | 10648 | |
18af7313 | 10649 | static HOST_WIDE_INT |
e32bac5b | 10650 | int_log2 (HOST_WIDE_INT power) |
cce8749e | 10651 | { |
f3bb6135 | 10652 | HOST_WIDE_INT shift = 0; |
cce8749e | 10653 | |
30cf4896 | 10654 | while ((((HOST_WIDE_INT) 1 << shift) & power) == 0) |
cce8749e | 10655 | { |
e6d29d15 | 10656 | gcc_assert (shift <= 31); |
e32bac5b | 10657 | shift++; |
cce8749e | 10658 | } |
f3bb6135 RE |
10659 | |
10660 | return shift; | |
10661 | } | |
cce8749e | 10662 | |
c5ff069d ZW |
10663 | /* Output a .ascii pseudo-op, keeping track of lengths. This is |
10664 | because /bin/as is horribly restrictive. The judgement about | |
10665 | whether or not each character is 'printable' (and can be output as | |
10666 | is) or not (and must be printed with an octal escape) must be made | |
10667 | with reference to the *host* character set -- the situation is | |
10668 | similar to that discussed in the comments above pp_c_char in | |
10669 | c-pretty-print.c. */ | |
10670 | ||
6cfc7210 | 10671 | #define MAX_ASCII_LEN 51 |
cce8749e CH |
10672 | |
10673 | void | |
e32bac5b | 10674 | output_ascii_pseudo_op (FILE *stream, const unsigned char *p, int len) |
cce8749e CH |
10675 | { |
10676 | int i; | |
6cfc7210 | 10677 | int len_so_far = 0; |
cce8749e | 10678 | |
6cfc7210 | 10679 | fputs ("\t.ascii\t\"", stream); |
f676971a | 10680 | |
cce8749e CH |
10681 | for (i = 0; i < len; i++) |
10682 | { | |
1d6e90ac | 10683 | int c = p[i]; |
cce8749e | 10684 | |
6cfc7210 | 10685 | if (len_so_far >= MAX_ASCII_LEN) |
cce8749e | 10686 | { |
6cfc7210 | 10687 | fputs ("\"\n\t.ascii\t\"", stream); |
cce8749e | 10688 | len_so_far = 0; |
cce8749e CH |
10689 | } |
10690 | ||
c5ff069d | 10691 | if (ISPRINT (c)) |
cce8749e | 10692 | { |
c5ff069d | 10693 | if (c == '\\' || c == '\"') |
6cfc7210 | 10694 | { |
c5ff069d | 10695 | putc ('\\', stream); |
5895f793 | 10696 | len_so_far++; |
6cfc7210 | 10697 | } |
c5ff069d ZW |
10698 | putc (c, stream); |
10699 | len_so_far++; | |
10700 | } | |
10701 | else | |
10702 | { | |
10703 | fprintf (stream, "\\%03o", c); | |
10704 | len_so_far += 4; | |
cce8749e | 10705 | } |
cce8749e | 10706 | } |
f3bb6135 | 10707 | |
cce8749e | 10708 | fputs ("\"\n", stream); |
f3bb6135 | 10709 | } |
cce8749e | 10710 | \f |
c9ca9b88 | 10711 | /* Compute the register save mask for registers 0 through 12 |
5848830f | 10712 | inclusive. This code is used by arm_compute_save_reg_mask. */ |
b279b20a | 10713 | |
6d3d9133 | 10714 | static unsigned long |
e32bac5b | 10715 | arm_compute_save_reg0_reg12_mask (void) |
6d3d9133 | 10716 | { |
121308d4 | 10717 | unsigned long func_type = arm_current_func_type (); |
b279b20a | 10718 | unsigned long save_reg_mask = 0; |
6d3d9133 | 10719 | unsigned int reg; |
6d3d9133 | 10720 | |
7b8b8ade | 10721 | if (IS_INTERRUPT (func_type)) |
6d3d9133 | 10722 | { |
7b8b8ade | 10723 | unsigned int max_reg; |
7b8b8ade NC |
10724 | /* Interrupt functions must not corrupt any registers, |
10725 | even call clobbered ones. If this is a leaf function | |
10726 | we can just examine the registers used by the RTL, but | |
10727 | otherwise we have to assume that whatever function is | |
10728 | called might clobber anything, and so we have to save | |
10729 | all the call-clobbered registers as well. */ | |
10730 | if (ARM_FUNC_TYPE (func_type) == ARM_FT_FIQ) | |
10731 | /* FIQ handlers have registers r8 - r12 banked, so | |
10732 | we only need to check r0 - r7, Normal ISRs only | |
121308d4 | 10733 | bank r14 and r15, so we must check up to r12. |
7b8b8ade NC |
10734 | r13 is the stack pointer which is always preserved, |
10735 | so we do not need to consider it here. */ | |
10736 | max_reg = 7; | |
10737 | else | |
10738 | max_reg = 12; | |
f676971a | 10739 | |
7b8b8ade | 10740 | for (reg = 0; reg <= max_reg; reg++) |
6fb5fa3c DB |
10741 | if (df_regs_ever_live_p (reg) |
10742 | || (! current_function_is_leaf && call_used_regs[reg])) | |
6d3d9133 | 10743 | save_reg_mask |= (1 << reg); |
cfa01aab | 10744 | |
286d28c3 | 10745 | /* Also save the pic base register if necessary. */ |
cfa01aab PB |
10746 | if (flag_pic |
10747 | && !TARGET_SINGLE_PIC_BASE | |
020a4035 | 10748 | && arm_pic_register != INVALID_REGNUM |
e3b5732b | 10749 | && crtl->uses_pic_offset_table) |
cfa01aab | 10750 | save_reg_mask |= 1 << PIC_OFFSET_TABLE_REGNUM; |
6d3d9133 NC |
10751 | } |
10752 | else | |
10753 | { | |
10754 | /* In the normal case we only need to save those registers | |
10755 | which are call saved and which are used by this function. */ | |
ec6237e4 | 10756 | for (reg = 0; reg <= 11; reg++) |
6fb5fa3c | 10757 | if (df_regs_ever_live_p (reg) && ! call_used_regs[reg]) |
6d3d9133 NC |
10758 | save_reg_mask |= (1 << reg); |
10759 | ||
10760 | /* Handle the frame pointer as a special case. */ | |
ec6237e4 | 10761 | if (frame_pointer_needed) |
6d3d9133 NC |
10762 | save_reg_mask |= 1 << HARD_FRAME_POINTER_REGNUM; |
10763 | ||
10764 | /* If we aren't loading the PIC register, | |
10765 | don't stack it even though it may be live. */ | |
10766 | if (flag_pic | |
e0b92319 | 10767 | && !TARGET_SINGLE_PIC_BASE |
020a4035 | 10768 | && arm_pic_register != INVALID_REGNUM |
6fb5fa3c | 10769 | && (df_regs_ever_live_p (PIC_OFFSET_TABLE_REGNUM) |
e3b5732b | 10770 | || crtl->uses_pic_offset_table)) |
6d3d9133 | 10771 | save_reg_mask |= 1 << PIC_OFFSET_TABLE_REGNUM; |
5b3e6663 PB |
10772 | |
10773 | /* The prologue will copy SP into R0, so save it. */ | |
10774 | if (IS_STACKALIGN (func_type)) | |
10775 | save_reg_mask |= 1; | |
6d3d9133 NC |
10776 | } |
10777 | ||
c9ca9b88 | 10778 | /* Save registers so the exception handler can modify them. */ |
e3b5732b | 10779 | if (crtl->calls_eh_return) |
c9ca9b88 PB |
10780 | { |
10781 | unsigned int i; | |
f676971a | 10782 | |
c9ca9b88 PB |
10783 | for (i = 0; ; i++) |
10784 | { | |
10785 | reg = EH_RETURN_DATA_REGNO (i); | |
10786 | if (reg == INVALID_REGNUM) | |
10787 | break; | |
10788 | save_reg_mask |= 1 << reg; | |
10789 | } | |
10790 | } | |
10791 | ||
121308d4 NC |
10792 | return save_reg_mask; |
10793 | } | |
10794 | ||
5b3e6663 | 10795 | |
35596784 AJ |
10796 | /* Compute the number of bytes used to store the static chain register on the |
10797 | stack, above the stack frame. We need to know this accurately to get the | |
10798 | alignment of the rest of the stack frame correct. */ | |
10799 | ||
10800 | static int arm_compute_static_chain_stack_bytes (void) | |
10801 | { | |
10802 | unsigned long func_type = arm_current_func_type (); | |
10803 | int static_chain_stack_bytes = 0; | |
10804 | ||
10805 | if (TARGET_APCS_FRAME && frame_pointer_needed && TARGET_ARM && | |
10806 | IS_NESTED (func_type) && | |
10807 | df_regs_ever_live_p (3) && crtl->args.pretend_args_size == 0) | |
10808 | static_chain_stack_bytes = 4; | |
10809 | ||
10810 | return static_chain_stack_bytes; | |
10811 | } | |
10812 | ||
10813 | ||
121308d4 | 10814 | /* Compute a bit mask of which registers need to be |
954954d1 PB |
10815 | saved on the stack for the current function. |
10816 | This is used by arm_get_frame_offsets, which may add extra registers. */ | |
121308d4 NC |
10817 | |
10818 | static unsigned long | |
e32bac5b | 10819 | arm_compute_save_reg_mask (void) |
121308d4 NC |
10820 | { |
10821 | unsigned int save_reg_mask = 0; | |
10822 | unsigned long func_type = arm_current_func_type (); | |
5b3e6663 | 10823 | unsigned int reg; |
121308d4 NC |
10824 | |
10825 | if (IS_NAKED (func_type)) | |
10826 | /* This should never really happen. */ | |
10827 | return 0; | |
10828 | ||
10829 | /* If we are creating a stack frame, then we must save the frame pointer, | |
10830 | IP (which will hold the old stack pointer), LR and the PC. */ | |
ec6237e4 | 10831 | if (TARGET_APCS_FRAME && frame_pointer_needed && TARGET_ARM) |
121308d4 NC |
10832 | save_reg_mask |= |
10833 | (1 << ARM_HARD_FRAME_POINTER_REGNUM) | |
10834 | | (1 << IP_REGNUM) | |
10835 | | (1 << LR_REGNUM) | |
10836 | | (1 << PC_REGNUM); | |
10837 | ||
10838 | /* Volatile functions do not return, so there | |
10839 | is no need to save any other registers. */ | |
10840 | if (IS_VOLATILE (func_type)) | |
10841 | return save_reg_mask; | |
10842 | ||
10843 | save_reg_mask |= arm_compute_save_reg0_reg12_mask (); | |
10844 | ||
6d3d9133 NC |
10845 | /* Decide if we need to save the link register. |
10846 | Interrupt routines have their own banked link register, | |
10847 | so they never need to save it. | |
1768c26f | 10848 | Otherwise if we do not use the link register we do not need to save |
6d3d9133 NC |
10849 | it. If we are pushing other registers onto the stack however, we |
10850 | can save an instruction in the epilogue by pushing the link register | |
10851 | now and then popping it back into the PC. This incurs extra memory | |
72ac76be | 10852 | accesses though, so we only do it when optimizing for size, and only |
6d3d9133 | 10853 | if we know that we will not need a fancy return sequence. */ |
6fb5fa3c DB |
10854 | if (df_regs_ever_live_p (LR_REGNUM) |
10855 | || (save_reg_mask | |
10856 | && optimize_size | |
10857 | && ARM_FUNC_TYPE (func_type) == ARM_FT_NORMAL | |
e3b5732b | 10858 | && !crtl->calls_eh_return)) |
6d3d9133 NC |
10859 | save_reg_mask |= 1 << LR_REGNUM; |
10860 | ||
6f7ebcbb NC |
10861 | if (cfun->machine->lr_save_eliminated) |
10862 | save_reg_mask &= ~ (1 << LR_REGNUM); | |
10863 | ||
5a9335ef NC |
10864 | if (TARGET_REALLY_IWMMXT |
10865 | && ((bit_count (save_reg_mask) | |
35596784 AJ |
10866 | + ARM_NUM_INTS (crtl->args.pretend_args_size + |
10867 | arm_compute_static_chain_stack_bytes()) | |
10868 | ) % 2) != 0) | |
5a9335ef | 10869 | { |
5a9335ef NC |
10870 | /* The total number of registers that are going to be pushed |
10871 | onto the stack is odd. We need to ensure that the stack | |
10872 | is 64-bit aligned before we start to save iWMMXt registers, | |
10873 | and also before we start to create locals. (A local variable | |
10874 | might be a double or long long which we will load/store using | |
10875 | an iWMMXt instruction). Therefore we need to push another | |
10876 | ARM register, so that the stack will be 64-bit aligned. We | |
10877 | try to avoid using the arg registers (r0 -r3) as they might be | |
10878 | used to pass values in a tail call. */ | |
10879 | for (reg = 4; reg <= 12; reg++) | |
10880 | if ((save_reg_mask & (1 << reg)) == 0) | |
10881 | break; | |
10882 | ||
10883 | if (reg <= 12) | |
10884 | save_reg_mask |= (1 << reg); | |
10885 | else | |
10886 | { | |
10887 | cfun->machine->sibcall_blocked = 1; | |
10888 | save_reg_mask |= (1 << 3); | |
10889 | } | |
10890 | } | |
10891 | ||
5b3e6663 PB |
10892 | /* We may need to push an additional register for use initializing the |
10893 | PIC base register. */ | |
10894 | if (TARGET_THUMB2 && IS_NESTED (func_type) && flag_pic | |
10895 | && (save_reg_mask & THUMB2_WORK_REGS) == 0) | |
10896 | { | |
10897 | reg = thumb_find_work_register (1 << 4); | |
10898 | if (!call_used_regs[reg]) | |
10899 | save_reg_mask |= (1 << reg); | |
10900 | } | |
10901 | ||
6d3d9133 NC |
10902 | return save_reg_mask; |
10903 | } | |
10904 | ||
9728c9d1 | 10905 | |
57934c39 PB |
10906 | /* Compute a bit mask of which registers need to be |
10907 | saved on the stack for the current function. */ | |
10908 | static unsigned long | |
5b3e6663 | 10909 | thumb1_compute_save_reg_mask (void) |
57934c39 PB |
10910 | { |
10911 | unsigned long mask; | |
b279b20a | 10912 | unsigned reg; |
57934c39 PB |
10913 | |
10914 | mask = 0; | |
10915 | for (reg = 0; reg < 12; reg ++) | |
6fb5fa3c | 10916 | if (df_regs_ever_live_p (reg) && !call_used_regs[reg]) |
b279b20a | 10917 | mask |= 1 << reg; |
57934c39 | 10918 | |
39c39be0 RE |
10919 | if (flag_pic |
10920 | && !TARGET_SINGLE_PIC_BASE | |
020a4035 | 10921 | && arm_pic_register != INVALID_REGNUM |
e3b5732b | 10922 | && crtl->uses_pic_offset_table) |
39c39be0 | 10923 | mask |= 1 << PIC_OFFSET_TABLE_REGNUM; |
b279b20a | 10924 | |
a2503645 RS |
10925 | /* See if we might need r11 for calls to _interwork_r11_call_via_rN(). */ |
10926 | if (!frame_pointer_needed && CALLER_INTERWORKING_SLOT_SIZE > 0) | |
10927 | mask |= 1 << ARM_HARD_FRAME_POINTER_REGNUM; | |
57934c39 | 10928 | |
b279b20a | 10929 | /* LR will also be pushed if any lo regs are pushed. */ |
57934c39 PB |
10930 | if (mask & 0xff || thumb_force_lr_save ()) |
10931 | mask |= (1 << LR_REGNUM); | |
10932 | ||
b279b20a NC |
10933 | /* Make sure we have a low work register if we need one. |
10934 | We will need one if we are going to push a high register, | |
10935 | but we are not currently intending to push a low register. */ | |
10936 | if ((mask & 0xff) == 0 | |
57934c39 | 10937 | && ((mask & 0x0f00) || TARGET_BACKTRACE)) |
b279b20a NC |
10938 | { |
10939 | /* Use thumb_find_work_register to choose which register | |
10940 | we will use. If the register is live then we will | |
10941 | have to push it. Use LAST_LO_REGNUM as our fallback | |
10942 | choice for the register to select. */ | |
10943 | reg = thumb_find_work_register (1 << LAST_LO_REGNUM); | |
19e723f4 PB |
10944 | /* Make sure the register returned by thumb_find_work_register is |
10945 | not part of the return value. */ | |
954954d1 | 10946 | if (reg * UNITS_PER_WORD <= (unsigned) arm_size_return_regs ()) |
19e723f4 | 10947 | reg = LAST_LO_REGNUM; |
b279b20a NC |
10948 | |
10949 | if (! call_used_regs[reg]) | |
10950 | mask |= 1 << reg; | |
10951 | } | |
57934c39 | 10952 | |
35596784 AJ |
10953 | /* The 504 below is 8 bytes less than 512 because there are two possible |
10954 | alignment words. We can't tell here if they will be present or not so we | |
10955 | have to play it safe and assume that they are. */ | |
10956 | if ((CALLER_INTERWORKING_SLOT_SIZE + | |
10957 | ROUND_UP_WORD (get_frame_size ()) + | |
10958 | crtl->outgoing_args_size) >= 504) | |
10959 | { | |
10960 | /* This is the same as the code in thumb1_expand_prologue() which | |
10961 | determines which register to use for stack decrement. */ | |
10962 | for (reg = LAST_ARG_REGNUM + 1; reg <= LAST_LO_REGNUM; reg++) | |
10963 | if (mask & (1 << reg)) | |
10964 | break; | |
10965 | ||
10966 | if (reg > LAST_LO_REGNUM) | |
10967 | { | |
10968 | /* Make sure we have a register available for stack decrement. */ | |
10969 | mask |= 1 << LAST_LO_REGNUM; | |
10970 | } | |
10971 | } | |
10972 | ||
57934c39 PB |
10973 | return mask; |
10974 | } | |
10975 | ||
10976 | ||
9728c9d1 PB |
10977 | /* Return the number of bytes required to save VFP registers. */ |
10978 | static int | |
10979 | arm_get_vfp_saved_size (void) | |
10980 | { | |
10981 | unsigned int regno; | |
10982 | int count; | |
10983 | int saved; | |
10984 | ||
10985 | saved = 0; | |
10986 | /* Space for saved VFP registers. */ | |
10987 | if (TARGET_HARD_FLOAT && TARGET_VFP) | |
10988 | { | |
10989 | count = 0; | |
10990 | for (regno = FIRST_VFP_REGNUM; | |
10991 | regno < LAST_VFP_REGNUM; | |
10992 | regno += 2) | |
10993 | { | |
6fb5fa3c DB |
10994 | if ((!df_regs_ever_live_p (regno) || call_used_regs[regno]) |
10995 | && (!df_regs_ever_live_p (regno + 1) || call_used_regs[regno + 1])) | |
9728c9d1 PB |
10996 | { |
10997 | if (count > 0) | |
10998 | { | |
10999 | /* Workaround ARM10 VFPr1 bug. */ | |
11000 | if (count == 2 && !arm_arch6) | |
11001 | count++; | |
8edfc4cc | 11002 | saved += count * 8; |
9728c9d1 PB |
11003 | } |
11004 | count = 0; | |
11005 | } | |
11006 | else | |
11007 | count++; | |
11008 | } | |
11009 | if (count > 0) | |
11010 | { | |
11011 | if (count == 2 && !arm_arch6) | |
11012 | count++; | |
8edfc4cc | 11013 | saved += count * 8; |
9728c9d1 PB |
11014 | } |
11015 | } | |
11016 | return saved; | |
11017 | } | |
11018 | ||
11019 | ||
699a4925 | 11020 | /* Generate a function exit sequence. If REALLY_RETURN is false, then do |
6d3d9133 | 11021 | everything bar the final return instruction. */ |
cd2b33d0 | 11022 | const char * |
e32bac5b | 11023 | output_return_instruction (rtx operand, int really_return, int reverse) |
ff9940b0 | 11024 | { |
6d3d9133 | 11025 | char conditional[10]; |
ff9940b0 | 11026 | char instr[100]; |
b279b20a | 11027 | unsigned reg; |
6d3d9133 NC |
11028 | unsigned long live_regs_mask; |
11029 | unsigned long func_type; | |
5848830f | 11030 | arm_stack_offsets *offsets; |
e26053d1 | 11031 | |
6d3d9133 | 11032 | func_type = arm_current_func_type (); |
e2c671ba | 11033 | |
6d3d9133 | 11034 | if (IS_NAKED (func_type)) |
d5b7b3ae | 11035 | return ""; |
6d3d9133 NC |
11036 | |
11037 | if (IS_VOLATILE (func_type) && TARGET_ABORT_NORETURN) | |
e2c671ba | 11038 | { |
699a4925 RE |
11039 | /* If this function was declared non-returning, and we have |
11040 | found a tail call, then we have to trust that the called | |
11041 | function won't return. */ | |
3a5a4282 PB |
11042 | if (really_return) |
11043 | { | |
11044 | rtx ops[2]; | |
f676971a | 11045 | |
3a5a4282 PB |
11046 | /* Otherwise, trap an attempted return by aborting. */ |
11047 | ops[0] = operand; | |
f676971a | 11048 | ops[1] = gen_rtx_SYMBOL_REF (Pmode, NEED_PLT_RELOC ? "abort(PLT)" |
3a5a4282 PB |
11049 | : "abort"); |
11050 | assemble_external_libcall (ops[1]); | |
11051 | output_asm_insn (reverse ? "bl%D0\t%a1" : "bl%d0\t%a1", ops); | |
11052 | } | |
f676971a | 11053 | |
e2c671ba RE |
11054 | return ""; |
11055 | } | |
6d3d9133 | 11056 | |
e3b5732b | 11057 | gcc_assert (!cfun->calls_alloca || really_return); |
ff9940b0 | 11058 | |
c414f8a9 | 11059 | sprintf (conditional, "%%?%%%c0", reverse ? 'D' : 'd'); |
d5b7b3ae | 11060 | |
6d3d9133 | 11061 | return_used_this_function = 1; |
ff9940b0 | 11062 | |
954954d1 PB |
11063 | offsets = arm_get_frame_offsets (); |
11064 | live_regs_mask = offsets->saved_regs_mask; | |
ff9940b0 | 11065 | |
1768c26f | 11066 | if (live_regs_mask) |
6d3d9133 | 11067 | { |
1768c26f PB |
11068 | const char * return_reg; |
11069 | ||
f676971a | 11070 | /* If we do not have any special requirements for function exit |
a15908a4 | 11071 | (e.g. interworking) then we can load the return address |
1768c26f PB |
11072 | directly into the PC. Otherwise we must load it into LR. */ |
11073 | if (really_return | |
a15908a4 | 11074 | && (IS_INTERRUPT (func_type) || !TARGET_INTERWORK)) |
1768c26f | 11075 | return_reg = reg_names[PC_REGNUM]; |
6d3d9133 | 11076 | else |
1768c26f PB |
11077 | return_reg = reg_names[LR_REGNUM]; |
11078 | ||
6d3d9133 | 11079 | if ((live_regs_mask & (1 << IP_REGNUM)) == (1 << IP_REGNUM)) |
b034930f ILT |
11080 | { |
11081 | /* There are three possible reasons for the IP register | |
11082 | being saved. 1) a stack frame was created, in which case | |
11083 | IP contains the old stack pointer, or 2) an ISR routine | |
11084 | corrupted it, or 3) it was saved to align the stack on | |
11085 | iWMMXt. In case 1, restore IP into SP, otherwise just | |
11086 | restore IP. */ | |
11087 | if (frame_pointer_needed) | |
11088 | { | |
11089 | live_regs_mask &= ~ (1 << IP_REGNUM); | |
11090 | live_regs_mask |= (1 << SP_REGNUM); | |
11091 | } | |
11092 | else | |
e6d29d15 | 11093 | gcc_assert (IS_INTERRUPT (func_type) || TARGET_REALLY_IWMMXT); |
b034930f | 11094 | } |
f3bb6135 | 11095 | |
3a7731fd PB |
11096 | /* On some ARM architectures it is faster to use LDR rather than |
11097 | LDM to load a single register. On other architectures, the | |
11098 | cost is the same. In 26 bit mode, or for exception handlers, | |
11099 | we have to use LDM to load the PC so that the CPSR is also | |
11100 | restored. */ | |
11101 | for (reg = 0; reg <= LAST_ARM_REGNUM; reg++) | |
b279b20a NC |
11102 | if (live_regs_mask == (1U << reg)) |
11103 | break; | |
11104 | ||
3a7731fd PB |
11105 | if (reg <= LAST_ARM_REGNUM |
11106 | && (reg != LR_REGNUM | |
f676971a | 11107 | || ! really_return |
61f0ccff | 11108 | || ! IS_INTERRUPT (func_type))) |
3a7731fd | 11109 | { |
f676971a | 11110 | sprintf (instr, "ldr%s\t%%|%s, [%%|sp], #4", conditional, |
3a7731fd | 11111 | (reg == LR_REGNUM) ? return_reg : reg_names[reg]); |
6d3d9133 | 11112 | } |
ff9940b0 | 11113 | else |
1d5473cb | 11114 | { |
1768c26f PB |
11115 | char *p; |
11116 | int first = 1; | |
6d3d9133 | 11117 | |
699a4925 RE |
11118 | /* Generate the load multiple instruction to restore the |
11119 | registers. Note we can get here, even if | |
11120 | frame_pointer_needed is true, but only if sp already | |
11121 | points to the base of the saved core registers. */ | |
11122 | if (live_regs_mask & (1 << SP_REGNUM)) | |
a72d4945 | 11123 | { |
5848830f PB |
11124 | unsigned HOST_WIDE_INT stack_adjust; |
11125 | ||
5848830f | 11126 | stack_adjust = offsets->outgoing_args - offsets->saved_regs; |
e6d29d15 | 11127 | gcc_assert (stack_adjust == 0 || stack_adjust == 4); |
a72d4945 | 11128 | |
5b3e6663 | 11129 | if (stack_adjust && arm_arch5 && TARGET_ARM) |
a72d4945 RE |
11130 | sprintf (instr, "ldm%sib\t%%|sp, {", conditional); |
11131 | else | |
11132 | { | |
b279b20a NC |
11133 | /* If we can't use ldmib (SA110 bug), |
11134 | then try to pop r3 instead. */ | |
a72d4945 RE |
11135 | if (stack_adjust) |
11136 | live_regs_mask |= 1 << 3; | |
11137 | sprintf (instr, "ldm%sfd\t%%|sp, {", conditional); | |
11138 | } | |
11139 | } | |
da6558fd | 11140 | else |
1768c26f PB |
11141 | sprintf (instr, "ldm%sfd\t%%|sp!, {", conditional); |
11142 | ||
11143 | p = instr + strlen (instr); | |
6d3d9133 | 11144 | |
1768c26f PB |
11145 | for (reg = 0; reg <= SP_REGNUM; reg++) |
11146 | if (live_regs_mask & (1 << reg)) | |
11147 | { | |
11148 | int l = strlen (reg_names[reg]); | |
11149 | ||
11150 | if (first) | |
11151 | first = 0; | |
11152 | else | |
11153 | { | |
11154 | memcpy (p, ", ", 2); | |
11155 | p += 2; | |
11156 | } | |
11157 | ||
11158 | memcpy (p, "%|", 2); | |
11159 | memcpy (p + 2, reg_names[reg], l); | |
11160 | p += l + 2; | |
11161 | } | |
f676971a | 11162 | |
1768c26f PB |
11163 | if (live_regs_mask & (1 << LR_REGNUM)) |
11164 | { | |
b17fe233 | 11165 | sprintf (p, "%s%%|%s}", first ? "" : ", ", return_reg); |
61f0ccff RE |
11166 | /* If returning from an interrupt, restore the CPSR. */ |
11167 | if (IS_INTERRUPT (func_type)) | |
b17fe233 | 11168 | strcat (p, "^"); |
1768c26f PB |
11169 | } |
11170 | else | |
11171 | strcpy (p, "}"); | |
1d5473cb | 11172 | } |
da6558fd | 11173 | |
1768c26f PB |
11174 | output_asm_insn (instr, & operand); |
11175 | ||
3a7731fd PB |
11176 | /* See if we need to generate an extra instruction to |
11177 | perform the actual function return. */ | |
11178 | if (really_return | |
11179 | && func_type != ARM_FT_INTERWORKED | |
11180 | && (live_regs_mask & (1 << LR_REGNUM)) != 0) | |
da6558fd | 11181 | { |
3a7731fd PB |
11182 | /* The return has already been handled |
11183 | by loading the LR into the PC. */ | |
11184 | really_return = 0; | |
da6558fd | 11185 | } |
ff9940b0 | 11186 | } |
e26053d1 | 11187 | |
1768c26f | 11188 | if (really_return) |
ff9940b0 | 11189 | { |
6d3d9133 NC |
11190 | switch ((int) ARM_FUNC_TYPE (func_type)) |
11191 | { | |
11192 | case ARM_FT_ISR: | |
11193 | case ARM_FT_FIQ: | |
5b3e6663 | 11194 | /* ??? This is wrong for unified assembly syntax. */ |
6d3d9133 NC |
11195 | sprintf (instr, "sub%ss\t%%|pc, %%|lr, #4", conditional); |
11196 | break; | |
11197 | ||
11198 | case ARM_FT_INTERWORKED: | |
11199 | sprintf (instr, "bx%s\t%%|lr", conditional); | |
11200 | break; | |
11201 | ||
11202 | case ARM_FT_EXCEPTION: | |
5b3e6663 | 11203 | /* ??? This is wrong for unified assembly syntax. */ |
6d3d9133 NC |
11204 | sprintf (instr, "mov%ss\t%%|pc, %%|lr", conditional); |
11205 | break; | |
11206 | ||
11207 | default: | |
68d560d4 RE |
11208 | /* Use bx if it's available. */ |
11209 | if (arm_arch5 || arm_arch4t) | |
f676971a | 11210 | sprintf (instr, "bx%s\t%%|lr", conditional); |
1768c26f | 11211 | else |
61f0ccff | 11212 | sprintf (instr, "mov%s\t%%|pc, %%|lr", conditional); |
6d3d9133 NC |
11213 | break; |
11214 | } | |
1768c26f PB |
11215 | |
11216 | output_asm_insn (instr, & operand); | |
ff9940b0 | 11217 | } |
f3bb6135 | 11218 | |
ff9940b0 RE |
11219 | return ""; |
11220 | } | |
11221 | ||
ef179a26 NC |
11222 | /* Write the function name into the code section, directly preceding |
11223 | the function prologue. | |
11224 | ||
11225 | Code will be output similar to this: | |
11226 | t0 | |
11227 | .ascii "arm_poke_function_name", 0 | |
11228 | .align | |
11229 | t1 | |
11230 | .word 0xff000000 + (t1 - t0) | |
11231 | arm_poke_function_name | |
11232 | mov ip, sp | |
11233 | stmfd sp!, {fp, ip, lr, pc} | |
11234 | sub fp, ip, #4 | |
11235 | ||
11236 | When performing a stack backtrace, code can inspect the value | |
11237 | of 'pc' stored at 'fp' + 0. If the trace function then looks | |
11238 | at location pc - 12 and the top 8 bits are set, then we know | |
11239 | that there is a function name embedded immediately preceding this | |
11240 | location and has length ((pc[-3]) & 0xff000000). | |
11241 | ||
11242 | We assume that pc is declared as a pointer to an unsigned long. | |
11243 | ||
11244 | It is of no benefit to output the function name if we are assembling | |
11245 | a leaf function. These function types will not contain a stack | |
11246 | backtrace structure, therefore it is not possible to determine the | |
11247 | function name. */ | |
ef179a26 | 11248 | void |
e32bac5b | 11249 | arm_poke_function_name (FILE *stream, const char *name) |
ef179a26 NC |
11250 | { |
11251 | unsigned long alignlength; | |
11252 | unsigned long length; | |
11253 | rtx x; | |
11254 | ||
d5b7b3ae | 11255 | length = strlen (name) + 1; |
0c2ca901 | 11256 | alignlength = ROUND_UP_WORD (length); |
f676971a | 11257 | |
949d79eb | 11258 | ASM_OUTPUT_ASCII (stream, name, length); |
ef179a26 | 11259 | ASM_OUTPUT_ALIGN (stream, 2); |
30cf4896 | 11260 | x = GEN_INT ((unsigned HOST_WIDE_INT) 0xff000000 + alignlength); |
301d03af | 11261 | assemble_aligned_integer (UNITS_PER_WORD, x); |
ef179a26 NC |
11262 | } |
11263 | ||
6d3d9133 NC |
11264 | /* Place some comments into the assembler stream |
11265 | describing the current function. */ | |
08c148a8 | 11266 | static void |
e32bac5b | 11267 | arm_output_function_prologue (FILE *f, HOST_WIDE_INT frame_size) |
cce8749e | 11268 | { |
6d3d9133 | 11269 | unsigned long func_type; |
08c148a8 | 11270 | |
5b3e6663 | 11271 | if (TARGET_THUMB1) |
08c148a8 | 11272 | { |
5b3e6663 | 11273 | thumb1_output_function_prologue (f, frame_size); |
08c148a8 NB |
11274 | return; |
11275 | } | |
f676971a | 11276 | |
6d3d9133 | 11277 | /* Sanity check. */ |
e6d29d15 | 11278 | gcc_assert (!arm_ccfsm_state && !arm_target_insn); |
31fdb4d5 | 11279 | |
6d3d9133 | 11280 | func_type = arm_current_func_type (); |
f676971a | 11281 | |
6d3d9133 NC |
11282 | switch ((int) ARM_FUNC_TYPE (func_type)) |
11283 | { | |
11284 | default: | |
11285 | case ARM_FT_NORMAL: | |
11286 | break; | |
11287 | case ARM_FT_INTERWORKED: | |
11288 | asm_fprintf (f, "\t%@ Function supports interworking.\n"); | |
11289 | break; | |
6d3d9133 NC |
11290 | case ARM_FT_ISR: |
11291 | asm_fprintf (f, "\t%@ Interrupt Service Routine.\n"); | |
11292 | break; | |
11293 | case ARM_FT_FIQ: | |
11294 | asm_fprintf (f, "\t%@ Fast Interrupt Service Routine.\n"); | |
11295 | break; | |
11296 | case ARM_FT_EXCEPTION: | |
11297 | asm_fprintf (f, "\t%@ ARM Exception Handler.\n"); | |
11298 | break; | |
11299 | } | |
f676971a | 11300 | |
6d3d9133 NC |
11301 | if (IS_NAKED (func_type)) |
11302 | asm_fprintf (f, "\t%@ Naked Function: prologue and epilogue provided by programmer.\n"); | |
11303 | ||
11304 | if (IS_VOLATILE (func_type)) | |
11305 | asm_fprintf (f, "\t%@ Volatile: function does not return.\n"); | |
11306 | ||
11307 | if (IS_NESTED (func_type)) | |
11308 | asm_fprintf (f, "\t%@ Nested: function declared inside another function.\n"); | |
5b3e6663 PB |
11309 | if (IS_STACKALIGN (func_type)) |
11310 | asm_fprintf (f, "\t%@ Stack Align: May be called with mis-aligned SP.\n"); | |
f676971a | 11311 | |
c53dddc2 | 11312 | asm_fprintf (f, "\t%@ args = %d, pretend = %d, frame = %wd\n", |
38173d38 JH |
11313 | crtl->args.size, |
11314 | crtl->args.pretend_args_size, frame_size); | |
6d3d9133 | 11315 | |
3cb66fd7 | 11316 | asm_fprintf (f, "\t%@ frame_needed = %d, uses_anonymous_args = %d\n", |
dd18ae56 | 11317 | frame_pointer_needed, |
3cb66fd7 | 11318 | cfun->machine->uses_anonymous_args); |
cce8749e | 11319 | |
6f7ebcbb NC |
11320 | if (cfun->machine->lr_save_eliminated) |
11321 | asm_fprintf (f, "\t%@ link register save eliminated.\n"); | |
11322 | ||
e3b5732b | 11323 | if (crtl->calls_eh_return) |
c9ca9b88 PB |
11324 | asm_fprintf (f, "\t@ Calls __builtin_eh_return.\n"); |
11325 | ||
f676971a | 11326 | return_used_this_function = 0; |
f3bb6135 | 11327 | } |
cce8749e | 11328 | |
cd2b33d0 | 11329 | const char * |
a72d4945 | 11330 | arm_output_epilogue (rtx sibling) |
cce8749e | 11331 | { |
949d79eb | 11332 | int reg; |
6f7ebcbb | 11333 | unsigned long saved_regs_mask; |
6d3d9133 | 11334 | unsigned long func_type; |
f676971a | 11335 | /* Floats_offset is the offset from the "virtual" frame. In an APCS |
c882c7ac RE |
11336 | frame that is $fp + 4 for a non-variadic function. */ |
11337 | int floats_offset = 0; | |
cce8749e | 11338 | rtx operands[3]; |
d5b7b3ae | 11339 | FILE * f = asm_out_file; |
5a9335ef | 11340 | unsigned int lrm_count = 0; |
a72d4945 | 11341 | int really_return = (sibling == NULL); |
9b66ebb1 | 11342 | int start_reg; |
5848830f | 11343 | arm_stack_offsets *offsets; |
cce8749e | 11344 | |
6d3d9133 NC |
11345 | /* If we have already generated the return instruction |
11346 | then it is futile to generate anything else. */ | |
a72d4945 | 11347 | if (use_return_insn (FALSE, sibling) && return_used_this_function) |
949d79eb | 11348 | return ""; |
cce8749e | 11349 | |
6d3d9133 | 11350 | func_type = arm_current_func_type (); |
d5b7b3ae | 11351 | |
6d3d9133 NC |
11352 | if (IS_NAKED (func_type)) |
11353 | /* Naked functions don't have epilogues. */ | |
11354 | return ""; | |
0616531f | 11355 | |
6d3d9133 | 11356 | if (IS_VOLATILE (func_type) && TARGET_ABORT_NORETURN) |
e2c671ba | 11357 | { |
86efdc8e | 11358 | rtx op; |
f676971a | 11359 | |
6d3d9133 | 11360 | /* A volatile function should never return. Call abort. */ |
ed0e6530 | 11361 | op = gen_rtx_SYMBOL_REF (Pmode, NEED_PLT_RELOC ? "abort(PLT)" : "abort"); |
2b835d68 | 11362 | assemble_external_libcall (op); |
e2c671ba | 11363 | output_asm_insn ("bl\t%a0", &op); |
f676971a | 11364 | |
949d79eb | 11365 | return ""; |
e2c671ba RE |
11366 | } |
11367 | ||
e6d29d15 NS |
11368 | /* If we are throwing an exception, then we really must be doing a |
11369 | return, so we can't tail-call. */ | |
e3b5732b | 11370 | gcc_assert (!crtl->calls_eh_return || really_return); |
f676971a | 11371 | |
5848830f | 11372 | offsets = arm_get_frame_offsets (); |
954954d1 | 11373 | saved_regs_mask = offsets->saved_regs_mask; |
5a9335ef NC |
11374 | |
11375 | if (TARGET_IWMMXT) | |
11376 | lrm_count = bit_count (saved_regs_mask); | |
11377 | ||
5848830f | 11378 | floats_offset = offsets->saved_args; |
6d3d9133 | 11379 | /* Compute how far away the floats will be. */ |
5a9335ef | 11380 | for (reg = 0; reg <= LAST_ARM_REGNUM; reg++) |
6f7ebcbb | 11381 | if (saved_regs_mask & (1 << reg)) |
6ed30148 | 11382 | floats_offset += 4; |
f676971a | 11383 | |
ec6237e4 | 11384 | if (TARGET_APCS_FRAME && frame_pointer_needed && TARGET_ARM) |
cce8749e | 11385 | { |
9b66ebb1 | 11386 | /* This variable is for the Virtual Frame Pointer, not VFP regs. */ |
5848830f | 11387 | int vfp_offset = offsets->frame; |
c882c7ac | 11388 | |
29ad9694 | 11389 | if (arm_fpu_arch == FPUTYPE_FPA_EMU2) |
b111229a | 11390 | { |
9b66ebb1 | 11391 | for (reg = LAST_FPA_REGNUM; reg >= FIRST_FPA_REGNUM; reg--) |
6fb5fa3c | 11392 | if (df_regs_ever_live_p (reg) && !call_used_regs[reg]) |
b111229a RE |
11393 | { |
11394 | floats_offset += 12; | |
f676971a | 11395 | asm_fprintf (f, "\tldfe\t%r, [%r, #-%d]\n", |
c882c7ac | 11396 | reg, FP_REGNUM, floats_offset - vfp_offset); |
b111229a RE |
11397 | } |
11398 | } | |
11399 | else | |
11400 | { | |
9b66ebb1 | 11401 | start_reg = LAST_FPA_REGNUM; |
b111229a | 11402 | |
9b66ebb1 | 11403 | for (reg = LAST_FPA_REGNUM; reg >= FIRST_FPA_REGNUM; reg--) |
b111229a | 11404 | { |
6fb5fa3c | 11405 | if (df_regs_ever_live_p (reg) && !call_used_regs[reg]) |
b111229a RE |
11406 | { |
11407 | floats_offset += 12; | |
f676971a | 11408 | |
6354dc9b | 11409 | /* We can't unstack more than four registers at once. */ |
b111229a RE |
11410 | if (start_reg - reg == 3) |
11411 | { | |
dd18ae56 | 11412 | asm_fprintf (f, "\tlfm\t%r, 4, [%r, #-%d]\n", |
c882c7ac | 11413 | reg, FP_REGNUM, floats_offset - vfp_offset); |
b111229a RE |
11414 | start_reg = reg - 1; |
11415 | } | |
11416 | } | |
11417 | else | |
11418 | { | |
11419 | if (reg != start_reg) | |
dd18ae56 NC |
11420 | asm_fprintf (f, "\tlfm\t%r, %d, [%r, #-%d]\n", |
11421 | reg + 1, start_reg - reg, | |
c882c7ac | 11422 | FP_REGNUM, floats_offset - vfp_offset); |
b111229a RE |
11423 | start_reg = reg - 1; |
11424 | } | |
11425 | } | |
11426 | ||
11427 | /* Just in case the last register checked also needs unstacking. */ | |
11428 | if (reg != start_reg) | |
dd18ae56 NC |
11429 | asm_fprintf (f, "\tlfm\t%r, %d, [%r, #-%d]\n", |
11430 | reg + 1, start_reg - reg, | |
c882c7ac | 11431 | FP_REGNUM, floats_offset - vfp_offset); |
b111229a | 11432 | } |
6d3d9133 | 11433 | |
9b66ebb1 PB |
11434 | if (TARGET_HARD_FLOAT && TARGET_VFP) |
11435 | { | |
9728c9d1 | 11436 | int saved_size; |
9b66ebb1 | 11437 | |
8edfc4cc MS |
11438 | /* The fldmd insns do not have base+offset addressing |
11439 | modes, so we use IP to hold the address. */ | |
9728c9d1 | 11440 | saved_size = arm_get_vfp_saved_size (); |
9b66ebb1 | 11441 | |
9728c9d1 | 11442 | if (saved_size > 0) |
9b66ebb1 | 11443 | { |
9728c9d1 | 11444 | floats_offset += saved_size; |
9b66ebb1 PB |
11445 | asm_fprintf (f, "\tsub\t%r, %r, #%d\n", IP_REGNUM, |
11446 | FP_REGNUM, floats_offset - vfp_offset); | |
11447 | } | |
11448 | start_reg = FIRST_VFP_REGNUM; | |
11449 | for (reg = FIRST_VFP_REGNUM; reg < LAST_VFP_REGNUM; reg += 2) | |
11450 | { | |
6fb5fa3c DB |
11451 | if ((!df_regs_ever_live_p (reg) || call_used_regs[reg]) |
11452 | && (!df_regs_ever_live_p (reg + 1) || call_used_regs[reg + 1])) | |
9b66ebb1 PB |
11453 | { |
11454 | if (start_reg != reg) | |
8edfc4cc | 11455 | vfp_output_fldmd (f, IP_REGNUM, |
9728c9d1 PB |
11456 | (start_reg - FIRST_VFP_REGNUM) / 2, |
11457 | (reg - start_reg) / 2); | |
9b66ebb1 PB |
11458 | start_reg = reg + 2; |
11459 | } | |
11460 | } | |
11461 | if (start_reg != reg) | |
8edfc4cc | 11462 | vfp_output_fldmd (f, IP_REGNUM, |
9728c9d1 PB |
11463 | (start_reg - FIRST_VFP_REGNUM) / 2, |
11464 | (reg - start_reg) / 2); | |
9b66ebb1 PB |
11465 | } |
11466 | ||
5a9335ef NC |
11467 | if (TARGET_IWMMXT) |
11468 | { | |
11469 | /* The frame pointer is guaranteed to be non-double-word aligned. | |
11470 | This is because it is set to (old_stack_pointer - 4) and the | |
11471 | old_stack_pointer was double word aligned. Thus the offset to | |
11472 | the iWMMXt registers to be loaded must also be non-double-word | |
11473 | sized, so that the resultant address *is* double-word aligned. | |
11474 | We can ignore floats_offset since that was already included in | |
11475 | the live_regs_mask. */ | |
11476 | lrm_count += (lrm_count % 2 ? 2 : 1); | |
f676971a | 11477 | |
01d4c813 | 11478 | for (reg = LAST_IWMMXT_REGNUM; reg >= FIRST_IWMMXT_REGNUM; reg--) |
6fb5fa3c | 11479 | if (df_regs_ever_live_p (reg) && !call_used_regs[reg]) |
5a9335ef | 11480 | { |
f676971a | 11481 | asm_fprintf (f, "\twldrd\t%r, [%r, #-%d]\n", |
5a9335ef | 11482 | reg, FP_REGNUM, lrm_count * 4); |
f676971a | 11483 | lrm_count += 2; |
5a9335ef NC |
11484 | } |
11485 | } | |
11486 | ||
6f7ebcbb | 11487 | /* saved_regs_mask should contain the IP, which at the time of stack |
6d3d9133 NC |
11488 | frame generation actually contains the old stack pointer. So a |
11489 | quick way to unwind the stack is just pop the IP register directly | |
11490 | into the stack pointer. */ | |
e6d29d15 | 11491 | gcc_assert (saved_regs_mask & (1 << IP_REGNUM)); |
6f7ebcbb NC |
11492 | saved_regs_mask &= ~ (1 << IP_REGNUM); |
11493 | saved_regs_mask |= (1 << SP_REGNUM); | |
6d3d9133 | 11494 | |
6f7ebcbb | 11495 | /* There are two registers left in saved_regs_mask - LR and PC. We |
6d3d9133 NC |
11496 | only need to restore the LR register (the return address), but to |
11497 | save time we can load it directly into the PC, unless we need a | |
11498 | special function exit sequence, or we are not really returning. */ | |
c9ca9b88 PB |
11499 | if (really_return |
11500 | && ARM_FUNC_TYPE (func_type) == ARM_FT_NORMAL | |
e3b5732b | 11501 | && !crtl->calls_eh_return) |
6d3d9133 NC |
11502 | /* Delete the LR from the register mask, so that the LR on |
11503 | the stack is loaded into the PC in the register mask. */ | |
6f7ebcbb | 11504 | saved_regs_mask &= ~ (1 << LR_REGNUM); |
b111229a | 11505 | else |
6f7ebcbb | 11506 | saved_regs_mask &= ~ (1 << PC_REGNUM); |
efc2515b RE |
11507 | |
11508 | /* We must use SP as the base register, because SP is one of the | |
11509 | registers being restored. If an interrupt or page fault | |
11510 | happens in the ldm instruction, the SP might or might not | |
11511 | have been restored. That would be bad, as then SP will no | |
11512 | longer indicate the safe area of stack, and we can get stack | |
11513 | corruption. Using SP as the base register means that it will | |
11514 | be reset correctly to the original value, should an interrupt | |
699a4925 RE |
11515 | occur. If the stack pointer already points at the right |
11516 | place, then omit the subtraction. */ | |
5848830f | 11517 | if (offsets->outgoing_args != (1 + (int) bit_count (saved_regs_mask)) |
e3b5732b | 11518 | || cfun->calls_alloca) |
699a4925 RE |
11519 | asm_fprintf (f, "\tsub\t%r, %r, #%d\n", SP_REGNUM, FP_REGNUM, |
11520 | 4 * bit_count (saved_regs_mask)); | |
a15908a4 | 11521 | print_multi_reg (f, "ldmfd\t%r, ", SP_REGNUM, saved_regs_mask, 0); |
7b8b8ade NC |
11522 | |
11523 | if (IS_INTERRUPT (func_type)) | |
11524 | /* Interrupt handlers will have pushed the | |
11525 | IP onto the stack, so restore it now. */ | |
a15908a4 | 11526 | print_multi_reg (f, "ldmfd\t%r!, ", SP_REGNUM, 1 << IP_REGNUM, 0); |
cce8749e CH |
11527 | } |
11528 | else | |
11529 | { | |
ec6237e4 PB |
11530 | /* This branch is executed for ARM mode (non-apcs frames) and |
11531 | Thumb-2 mode. Frame layout is essentially the same for those | |
11532 | cases, except that in ARM mode frame pointer points to the | |
11533 | first saved register, while in Thumb-2 mode the frame pointer points | |
11534 | to the last saved register. | |
11535 | ||
11536 | It is possible to make frame pointer point to last saved | |
11537 | register in both cases, and remove some conditionals below. | |
11538 | That means that fp setup in prologue would be just "mov fp, sp" | |
11539 | and sp restore in epilogue would be just "mov sp, fp", whereas | |
11540 | now we have to use add/sub in those cases. However, the value | |
11541 | of that would be marginal, as both mov and add/sub are 32-bit | |
11542 | in ARM mode, and it would require extra conditionals | |
11543 | in arm_expand_prologue to distingish ARM-apcs-frame case | |
11544 | (where frame pointer is required to point at first register) | |
11545 | and ARM-non-apcs-frame. Therefore, such change is postponed | |
11546 | until real need arise. */ | |
5b3e6663 | 11547 | HOST_WIDE_INT amount; |
a15908a4 | 11548 | int rfe; |
d2288d8d | 11549 | /* Restore stack pointer if necessary. */ |
ec6237e4 | 11550 | if (TARGET_ARM && frame_pointer_needed) |
5b3e6663 PB |
11551 | { |
11552 | operands[0] = stack_pointer_rtx; | |
ec6237e4 PB |
11553 | operands[1] = hard_frame_pointer_rtx; |
11554 | ||
11555 | operands[2] = GEN_INT (offsets->frame - offsets->saved_regs); | |
11556 | output_add_immediate (operands); | |
5b3e6663 | 11557 | } |
ec6237e4 | 11558 | else |
5b3e6663 | 11559 | { |
ec6237e4 PB |
11560 | if (frame_pointer_needed) |
11561 | { | |
11562 | /* For Thumb-2 restore sp from the frame pointer. | |
11563 | Operand restrictions mean we have to incrememnt FP, then copy | |
11564 | to SP. */ | |
11565 | amount = offsets->locals_base - offsets->saved_regs; | |
11566 | operands[0] = hard_frame_pointer_rtx; | |
11567 | } | |
11568 | else | |
11569 | { | |
954954d1 | 11570 | unsigned long count; |
ec6237e4 PB |
11571 | operands[0] = stack_pointer_rtx; |
11572 | amount = offsets->outgoing_args - offsets->saved_regs; | |
954954d1 PB |
11573 | /* pop call clobbered registers if it avoids a |
11574 | separate stack adjustment. */ | |
11575 | count = offsets->saved_regs - offsets->saved_args; | |
11576 | if (optimize_size | |
11577 | && count != 0 | |
e3b5732b | 11578 | && !crtl->calls_eh_return |
954954d1 PB |
11579 | && bit_count(saved_regs_mask) * 4 == count |
11580 | && !IS_INTERRUPT (func_type) | |
e3b5732b | 11581 | && !crtl->tail_call_emit) |
954954d1 PB |
11582 | { |
11583 | unsigned long mask; | |
11584 | mask = (1 << (arm_size_return_regs() / 4)) - 1; | |
11585 | mask ^= 0xf; | |
11586 | mask &= ~saved_regs_mask; | |
11587 | reg = 0; | |
11588 | while (bit_count (mask) * 4 > amount) | |
11589 | { | |
11590 | while ((mask & (1 << reg)) == 0) | |
11591 | reg++; | |
11592 | mask &= ~(1 << reg); | |
11593 | } | |
11594 | if (bit_count (mask) * 4 == amount) { | |
11595 | amount = 0; | |
11596 | saved_regs_mask |= mask; | |
11597 | } | |
11598 | } | |
ec6237e4 PB |
11599 | } |
11600 | ||
11601 | if (amount) | |
11602 | { | |
11603 | operands[1] = operands[0]; | |
11604 | operands[2] = GEN_INT (amount); | |
11605 | output_add_immediate (operands); | |
11606 | } | |
11607 | if (frame_pointer_needed) | |
11608 | asm_fprintf (f, "\tmov\t%r, %r\n", | |
11609 | SP_REGNUM, HARD_FRAME_POINTER_REGNUM); | |
d2288d8d TG |
11610 | } |
11611 | ||
29ad9694 | 11612 | if (arm_fpu_arch == FPUTYPE_FPA_EMU2) |
b111229a | 11613 | { |
9b66ebb1 | 11614 | for (reg = FIRST_FPA_REGNUM; reg <= LAST_FPA_REGNUM; reg++) |
6fb5fa3c | 11615 | if (df_regs_ever_live_p (reg) && !call_used_regs[reg]) |
dd18ae56 NC |
11616 | asm_fprintf (f, "\tldfe\t%r, [%r], #12\n", |
11617 | reg, SP_REGNUM); | |
b111229a RE |
11618 | } |
11619 | else | |
11620 | { | |
9b66ebb1 | 11621 | start_reg = FIRST_FPA_REGNUM; |
b111229a | 11622 | |
9b66ebb1 | 11623 | for (reg = FIRST_FPA_REGNUM; reg <= LAST_FPA_REGNUM; reg++) |
b111229a | 11624 | { |
6fb5fa3c | 11625 | if (df_regs_ever_live_p (reg) && !call_used_regs[reg]) |
b111229a RE |
11626 | { |
11627 | if (reg - start_reg == 3) | |
11628 | { | |
dd18ae56 NC |
11629 | asm_fprintf (f, "\tlfmfd\t%r, 4, [%r]!\n", |
11630 | start_reg, SP_REGNUM); | |
b111229a RE |
11631 | start_reg = reg + 1; |
11632 | } | |
11633 | } | |
11634 | else | |
11635 | { | |
11636 | if (reg != start_reg) | |
dd18ae56 NC |
11637 | asm_fprintf (f, "\tlfmfd\t%r, %d, [%r]!\n", |
11638 | start_reg, reg - start_reg, | |
11639 | SP_REGNUM); | |
f676971a | 11640 | |
b111229a RE |
11641 | start_reg = reg + 1; |
11642 | } | |
11643 | } | |
11644 | ||
11645 | /* Just in case the last register checked also needs unstacking. */ | |
11646 | if (reg != start_reg) | |
dd18ae56 NC |
11647 | asm_fprintf (f, "\tlfmfd\t%r, %d, [%r]!\n", |
11648 | start_reg, reg - start_reg, SP_REGNUM); | |
b111229a RE |
11649 | } |
11650 | ||
9b66ebb1 PB |
11651 | if (TARGET_HARD_FLOAT && TARGET_VFP) |
11652 | { | |
11653 | start_reg = FIRST_VFP_REGNUM; | |
11654 | for (reg = FIRST_VFP_REGNUM; reg < LAST_VFP_REGNUM; reg += 2) | |
11655 | { | |
6fb5fa3c DB |
11656 | if ((!df_regs_ever_live_p (reg) || call_used_regs[reg]) |
11657 | && (!df_regs_ever_live_p (reg + 1) || call_used_regs[reg + 1])) | |
9b66ebb1 PB |
11658 | { |
11659 | if (start_reg != reg) | |
8edfc4cc | 11660 | vfp_output_fldmd (f, SP_REGNUM, |
9728c9d1 PB |
11661 | (start_reg - FIRST_VFP_REGNUM) / 2, |
11662 | (reg - start_reg) / 2); | |
9b66ebb1 PB |
11663 | start_reg = reg + 2; |
11664 | } | |
11665 | } | |
11666 | if (start_reg != reg) | |
8edfc4cc | 11667 | vfp_output_fldmd (f, SP_REGNUM, |
9728c9d1 PB |
11668 | (start_reg - FIRST_VFP_REGNUM) / 2, |
11669 | (reg - start_reg) / 2); | |
9b66ebb1 | 11670 | } |
5a9335ef NC |
11671 | if (TARGET_IWMMXT) |
11672 | for (reg = FIRST_IWMMXT_REGNUM; reg <= LAST_IWMMXT_REGNUM; reg++) | |
6fb5fa3c | 11673 | if (df_regs_ever_live_p (reg) && !call_used_regs[reg]) |
01d4c813 | 11674 | asm_fprintf (f, "\twldrd\t%r, [%r], #8\n", reg, SP_REGNUM); |
5a9335ef | 11675 | |
6d3d9133 | 11676 | /* If we can, restore the LR into the PC. */ |
a15908a4 PB |
11677 | if (ARM_FUNC_TYPE (func_type) != ARM_FT_INTERWORKED |
11678 | && (TARGET_ARM || ARM_FUNC_TYPE (func_type) == ARM_FT_NORMAL) | |
5b3e6663 | 11679 | && !IS_STACKALIGN (func_type) |
6d3d9133 | 11680 | && really_return |
38173d38 | 11681 | && crtl->args.pretend_args_size == 0 |
c9ca9b88 | 11682 | && saved_regs_mask & (1 << LR_REGNUM) |
e3b5732b | 11683 | && !crtl->calls_eh_return) |
cce8749e | 11684 | { |
6f7ebcbb NC |
11685 | saved_regs_mask &= ~ (1 << LR_REGNUM); |
11686 | saved_regs_mask |= (1 << PC_REGNUM); | |
a15908a4 | 11687 | rfe = IS_INTERRUPT (func_type); |
6d3d9133 | 11688 | } |
a15908a4 PB |
11689 | else |
11690 | rfe = 0; | |
d5b7b3ae | 11691 | |
6d3d9133 | 11692 | /* Load the registers off the stack. If we only have one register |
5b3e6663 PB |
11693 | to load use the LDR instruction - it is faster. For Thumb-2 |
11694 | always use pop and the assembler will pick the best instruction.*/ | |
a15908a4 PB |
11695 | if (TARGET_ARM && saved_regs_mask == (1 << LR_REGNUM) |
11696 | && !IS_INTERRUPT(func_type)) | |
6d3d9133 | 11697 | { |
c9ca9b88 | 11698 | asm_fprintf (f, "\tldr\t%r, [%r], #4\n", LR_REGNUM, SP_REGNUM); |
cce8749e | 11699 | } |
6f7ebcbb | 11700 | else if (saved_regs_mask) |
f1acdf8b NC |
11701 | { |
11702 | if (saved_regs_mask & (1 << SP_REGNUM)) | |
11703 | /* Note - write back to the stack register is not enabled | |
112cdef5 | 11704 | (i.e. "ldmfd sp!..."). We know that the stack pointer is |
f1acdf8b NC |
11705 | in the list of registers and if we add writeback the |
11706 | instruction becomes UNPREDICTABLE. */ | |
a15908a4 PB |
11707 | print_multi_reg (f, "ldmfd\t%r, ", SP_REGNUM, saved_regs_mask, |
11708 | rfe); | |
5b3e6663 | 11709 | else if (TARGET_ARM) |
a15908a4 PB |
11710 | print_multi_reg (f, "ldmfd\t%r!, ", SP_REGNUM, saved_regs_mask, |
11711 | rfe); | |
f1acdf8b | 11712 | else |
a15908a4 | 11713 | print_multi_reg (f, "pop\t", SP_REGNUM, saved_regs_mask, 0); |
f1acdf8b | 11714 | } |
6d3d9133 | 11715 | |
38173d38 | 11716 | if (crtl->args.pretend_args_size) |
cce8749e | 11717 | { |
6d3d9133 NC |
11718 | /* Unwind the pre-pushed regs. */ |
11719 | operands[0] = operands[1] = stack_pointer_rtx; | |
38173d38 | 11720 | operands[2] = GEN_INT (crtl->args.pretend_args_size); |
6d3d9133 NC |
11721 | output_add_immediate (operands); |
11722 | } | |
11723 | } | |
32de079a | 11724 | |
2966b00e | 11725 | /* We may have already restored PC directly from the stack. */ |
0cc3dda8 | 11726 | if (!really_return || saved_regs_mask & (1 << PC_REGNUM)) |
6d3d9133 | 11727 | return ""; |
d5b7b3ae | 11728 | |
c9ca9b88 | 11729 | /* Stack adjustment for exception handler. */ |
e3b5732b | 11730 | if (crtl->calls_eh_return) |
f676971a | 11731 | asm_fprintf (f, "\tadd\t%r, %r, %r\n", SP_REGNUM, SP_REGNUM, |
c9ca9b88 PB |
11732 | ARM_EH_STACKADJ_REGNUM); |
11733 | ||
6d3d9133 NC |
11734 | /* Generate the return instruction. */ |
11735 | switch ((int) ARM_FUNC_TYPE (func_type)) | |
11736 | { | |
6d3d9133 NC |
11737 | case ARM_FT_ISR: |
11738 | case ARM_FT_FIQ: | |
11739 | asm_fprintf (f, "\tsubs\t%r, %r, #4\n", PC_REGNUM, LR_REGNUM); | |
11740 | break; | |
11741 | ||
11742 | case ARM_FT_EXCEPTION: | |
11743 | asm_fprintf (f, "\tmovs\t%r, %r\n", PC_REGNUM, LR_REGNUM); | |
11744 | break; | |
11745 | ||
11746 | case ARM_FT_INTERWORKED: | |
11747 | asm_fprintf (f, "\tbx\t%r\n", LR_REGNUM); | |
11748 | break; | |
11749 | ||
11750 | default: | |
5b3e6663 PB |
11751 | if (IS_STACKALIGN (func_type)) |
11752 | { | |
11753 | /* See comment in arm_expand_prologue. */ | |
11754 | asm_fprintf (f, "\tmov\t%r, %r\n", SP_REGNUM, 0); | |
11755 | } | |
68d560d4 RE |
11756 | if (arm_arch5 || arm_arch4t) |
11757 | asm_fprintf (f, "\tbx\t%r\n", LR_REGNUM); | |
11758 | else | |
11759 | asm_fprintf (f, "\tmov\t%r, %r\n", PC_REGNUM, LR_REGNUM); | |
6d3d9133 | 11760 | break; |
cce8749e | 11761 | } |
f3bb6135 | 11762 | |
949d79eb RE |
11763 | return ""; |
11764 | } | |
11765 | ||
08c148a8 | 11766 | static void |
e32bac5b | 11767 | arm_output_function_epilogue (FILE *file ATTRIBUTE_UNUSED, |
5848830f | 11768 | HOST_WIDE_INT frame_size ATTRIBUTE_UNUSED) |
949d79eb | 11769 | { |
5848830f PB |
11770 | arm_stack_offsets *offsets; |
11771 | ||
5b3e6663 | 11772 | if (TARGET_THUMB1) |
d5b7b3ae | 11773 | { |
b12a00f1 RE |
11774 | int regno; |
11775 | ||
11776 | /* Emit any call-via-reg trampolines that are needed for v4t support | |
11777 | of call_reg and call_value_reg type insns. */ | |
57ecec57 | 11778 | for (regno = 0; regno < LR_REGNUM; regno++) |
b12a00f1 RE |
11779 | { |
11780 | rtx label = cfun->machine->call_via[regno]; | |
11781 | ||
11782 | if (label != NULL) | |
11783 | { | |
d6b5193b | 11784 | switch_to_section (function_section (current_function_decl)); |
b12a00f1 RE |
11785 | targetm.asm_out.internal_label (asm_out_file, "L", |
11786 | CODE_LABEL_NUMBER (label)); | |
11787 | asm_fprintf (asm_out_file, "\tbx\t%r\n", regno); | |
11788 | } | |
11789 | } | |
11790 | ||
d5b7b3ae RE |
11791 | /* ??? Probably not safe to set this here, since it assumes that a |
11792 | function will be emitted as assembly immediately after we generate | |
11793 | RTL for it. This does not happen for inline functions. */ | |
11794 | return_used_this_function = 0; | |
11795 | } | |
5b3e6663 | 11796 | else /* TARGET_32BIT */ |
d5b7b3ae | 11797 | { |
0977774b | 11798 | /* We need to take into account any stack-frame rounding. */ |
5848830f | 11799 | offsets = arm_get_frame_offsets (); |
0977774b | 11800 | |
e6d29d15 NS |
11801 | gcc_assert (!use_return_insn (FALSE, NULL) |
11802 | || !return_used_this_function | |
11803 | || offsets->saved_regs == offsets->outgoing_args | |
11804 | || frame_pointer_needed); | |
f3bb6135 | 11805 | |
d5b7b3ae | 11806 | /* Reset the ARM-specific per-function variables. */ |
d5b7b3ae RE |
11807 | after_arm_reorg = 0; |
11808 | } | |
f3bb6135 | 11809 | } |
e2c671ba | 11810 | |
2c849145 JM |
11811 | /* Generate and emit an insn that we will recognize as a push_multi. |
11812 | Unfortunately, since this insn does not reflect very well the actual | |
11813 | semantics of the operation, we need to annotate the insn for the benefit | |
11814 | of DWARF2 frame unwind information. */ | |
2c849145 | 11815 | static rtx |
b279b20a | 11816 | emit_multi_reg_push (unsigned long mask) |
e2c671ba RE |
11817 | { |
11818 | int num_regs = 0; | |
9b598fa0 | 11819 | int num_dwarf_regs; |
e2c671ba RE |
11820 | int i, j; |
11821 | rtx par; | |
2c849145 | 11822 | rtx dwarf; |
87e27392 | 11823 | int dwarf_par_index; |
2c849145 | 11824 | rtx tmp, reg; |
e2c671ba | 11825 | |
d5b7b3ae | 11826 | for (i = 0; i <= LAST_ARM_REGNUM; i++) |
e2c671ba | 11827 | if (mask & (1 << i)) |
5895f793 | 11828 | num_regs++; |
e2c671ba | 11829 | |
e6d29d15 | 11830 | gcc_assert (num_regs && num_regs <= 16); |
e2c671ba | 11831 | |
9b598fa0 RE |
11832 | /* We don't record the PC in the dwarf frame information. */ |
11833 | num_dwarf_regs = num_regs; | |
11834 | if (mask & (1 << PC_REGNUM)) | |
11835 | num_dwarf_regs--; | |
11836 | ||
87e27392 | 11837 | /* For the body of the insn we are going to generate an UNSPEC in |
05713b80 | 11838 | parallel with several USEs. This allows the insn to be recognized |
87e27392 NC |
11839 | by the push_multi pattern in the arm.md file. The insn looks |
11840 | something like this: | |
11841 | ||
f676971a | 11842 | (parallel [ |
b15bca31 RE |
11843 | (set (mem:BLK (pre_dec:BLK (reg:SI sp))) |
11844 | (unspec:BLK [(reg:SI r4)] UNSPEC_PUSH_MULT)) | |
87e27392 NC |
11845 | (use (reg:SI 11 fp)) |
11846 | (use (reg:SI 12 ip)) | |
11847 | (use (reg:SI 14 lr)) | |
11848 | (use (reg:SI 15 pc)) | |
11849 | ]) | |
11850 | ||
11851 | For the frame note however, we try to be more explicit and actually | |
11852 | show each register being stored into the stack frame, plus a (single) | |
11853 | decrement of the stack pointer. We do it this way in order to be | |
11854 | friendly to the stack unwinding code, which only wants to see a single | |
11855 | stack decrement per instruction. The RTL we generate for the note looks | |
11856 | something like this: | |
11857 | ||
f676971a | 11858 | (sequence [ |
87e27392 NC |
11859 | (set (reg:SI sp) (plus:SI (reg:SI sp) (const_int -20))) |
11860 | (set (mem:SI (reg:SI sp)) (reg:SI r4)) | |
11861 | (set (mem:SI (plus:SI (reg:SI sp) (const_int 4))) (reg:SI fp)) | |
11862 | (set (mem:SI (plus:SI (reg:SI sp) (const_int 8))) (reg:SI ip)) | |
11863 | (set (mem:SI (plus:SI (reg:SI sp) (const_int 12))) (reg:SI lr)) | |
87e27392 NC |
11864 | ]) |
11865 | ||
11866 | This sequence is used both by the code to support stack unwinding for | |
11867 | exceptions handlers and the code to generate dwarf2 frame debugging. */ | |
f676971a | 11868 | |
43cffd11 | 11869 | par = gen_rtx_PARALLEL (VOIDmode, rtvec_alloc (num_regs)); |
9b598fa0 | 11870 | dwarf = gen_rtx_SEQUENCE (VOIDmode, rtvec_alloc (num_dwarf_regs + 1)); |
87e27392 | 11871 | dwarf_par_index = 1; |
e2c671ba | 11872 | |
d5b7b3ae | 11873 | for (i = 0; i <= LAST_ARM_REGNUM; i++) |
e2c671ba RE |
11874 | { |
11875 | if (mask & (1 << i)) | |
11876 | { | |
2c849145 JM |
11877 | reg = gen_rtx_REG (SImode, i); |
11878 | ||
e2c671ba | 11879 | XVECEXP (par, 0, 0) |
43cffd11 | 11880 | = gen_rtx_SET (VOIDmode, |
31fa16b6 RE |
11881 | gen_frame_mem (BLKmode, |
11882 | gen_rtx_PRE_DEC (BLKmode, | |
11883 | stack_pointer_rtx)), | |
43cffd11 | 11884 | gen_rtx_UNSPEC (BLKmode, |
2c849145 | 11885 | gen_rtvec (1, reg), |
9b598fa0 | 11886 | UNSPEC_PUSH_MULT)); |
2c849145 | 11887 | |
9b598fa0 RE |
11888 | if (i != PC_REGNUM) |
11889 | { | |
11890 | tmp = gen_rtx_SET (VOIDmode, | |
31fa16b6 | 11891 | gen_frame_mem (SImode, stack_pointer_rtx), |
9b598fa0 RE |
11892 | reg); |
11893 | RTX_FRAME_RELATED_P (tmp) = 1; | |
11894 | XVECEXP (dwarf, 0, dwarf_par_index) = tmp; | |
11895 | dwarf_par_index++; | |
11896 | } | |
2c849145 | 11897 | |
e2c671ba RE |
11898 | break; |
11899 | } | |
11900 | } | |
11901 | ||
11902 | for (j = 1, i++; j < num_regs; i++) | |
11903 | { | |
11904 | if (mask & (1 << i)) | |
11905 | { | |
2c849145 JM |
11906 | reg = gen_rtx_REG (SImode, i); |
11907 | ||
11908 | XVECEXP (par, 0, j) = gen_rtx_USE (VOIDmode, reg); | |
11909 | ||
9b598fa0 RE |
11910 | if (i != PC_REGNUM) |
11911 | { | |
31fa16b6 RE |
11912 | tmp |
11913 | = gen_rtx_SET (VOIDmode, | |
11914 | gen_frame_mem (SImode, | |
9b598fa0 RE |
11915 | plus_constant (stack_pointer_rtx, |
11916 | 4 * j)), | |
31fa16b6 | 11917 | reg); |
9b598fa0 RE |
11918 | RTX_FRAME_RELATED_P (tmp) = 1; |
11919 | XVECEXP (dwarf, 0, dwarf_par_index++) = tmp; | |
11920 | } | |
11921 | ||
e2c671ba RE |
11922 | j++; |
11923 | } | |
11924 | } | |
b111229a | 11925 | |
2c849145 | 11926 | par = emit_insn (par); |
f676971a | 11927 | |
d66437c5 | 11928 | tmp = gen_rtx_SET (VOIDmode, |
87e27392 | 11929 | stack_pointer_rtx, |
d66437c5 | 11930 | plus_constant (stack_pointer_rtx, -4 * num_regs)); |
87e27392 NC |
11931 | RTX_FRAME_RELATED_P (tmp) = 1; |
11932 | XVECEXP (dwarf, 0, 0) = tmp; | |
f676971a | 11933 | |
2c849145 JM |
11934 | REG_NOTES (par) = gen_rtx_EXPR_LIST (REG_FRAME_RELATED_EXPR, dwarf, |
11935 | REG_NOTES (par)); | |
11936 | return par; | |
b111229a RE |
11937 | } |
11938 | ||
4f5dfed0 | 11939 | /* Calculate the size of the return value that is passed in registers. */ |
466e4b7a | 11940 | static unsigned |
4f5dfed0 JC |
11941 | arm_size_return_regs (void) |
11942 | { | |
11943 | enum machine_mode mode; | |
11944 | ||
38173d38 JH |
11945 | if (crtl->return_rtx != 0) |
11946 | mode = GET_MODE (crtl->return_rtx); | |
4f5dfed0 JC |
11947 | else |
11948 | mode = DECL_MODE (DECL_RESULT (current_function_decl)); | |
11949 | ||
11950 | return GET_MODE_SIZE (mode); | |
11951 | } | |
11952 | ||
2c849145 | 11953 | static rtx |
e32bac5b | 11954 | emit_sfm (int base_reg, int count) |
b111229a RE |
11955 | { |
11956 | rtx par; | |
2c849145 JM |
11957 | rtx dwarf; |
11958 | rtx tmp, reg; | |
b111229a RE |
11959 | int i; |
11960 | ||
43cffd11 | 11961 | par = gen_rtx_PARALLEL (VOIDmode, rtvec_alloc (count)); |
8ee6eb4e | 11962 | dwarf = gen_rtx_SEQUENCE (VOIDmode, rtvec_alloc (count + 1)); |
2c849145 JM |
11963 | |
11964 | reg = gen_rtx_REG (XFmode, base_reg++); | |
43cffd11 RE |
11965 | |
11966 | XVECEXP (par, 0, 0) | |
f676971a | 11967 | = gen_rtx_SET (VOIDmode, |
31fa16b6 RE |
11968 | gen_frame_mem (BLKmode, |
11969 | gen_rtx_PRE_DEC (BLKmode, | |
11970 | stack_pointer_rtx)), | |
43cffd11 | 11971 | gen_rtx_UNSPEC (BLKmode, |
2c849145 | 11972 | gen_rtvec (1, reg), |
b15bca31 | 11973 | UNSPEC_PUSH_MULT)); |
f676971a | 11974 | tmp = gen_rtx_SET (VOIDmode, |
31fa16b6 | 11975 | gen_frame_mem (XFmode, stack_pointer_rtx), reg); |
2c849145 | 11976 | RTX_FRAME_RELATED_P (tmp) = 1; |
f676971a EC |
11977 | XVECEXP (dwarf, 0, 1) = tmp; |
11978 | ||
b111229a | 11979 | for (i = 1; i < count; i++) |
2c849145 JM |
11980 | { |
11981 | reg = gen_rtx_REG (XFmode, base_reg++); | |
11982 | XVECEXP (par, 0, i) = gen_rtx_USE (VOIDmode, reg); | |
11983 | ||
f676971a | 11984 | tmp = gen_rtx_SET (VOIDmode, |
31fa16b6 RE |
11985 | gen_frame_mem (XFmode, |
11986 | plus_constant (stack_pointer_rtx, | |
11987 | i * 12)), | |
2c849145 JM |
11988 | reg); |
11989 | RTX_FRAME_RELATED_P (tmp) = 1; | |
f676971a | 11990 | XVECEXP (dwarf, 0, i + 1) = tmp; |
2c849145 | 11991 | } |
b111229a | 11992 | |
8ee6eb4e PB |
11993 | tmp = gen_rtx_SET (VOIDmode, |
11994 | stack_pointer_rtx, | |
d66437c5 RE |
11995 | plus_constant (stack_pointer_rtx, -12 * count)); |
11996 | ||
8ee6eb4e PB |
11997 | RTX_FRAME_RELATED_P (tmp) = 1; |
11998 | XVECEXP (dwarf, 0, 0) = tmp; | |
11999 | ||
2c849145 JM |
12000 | par = emit_insn (par); |
12001 | REG_NOTES (par) = gen_rtx_EXPR_LIST (REG_FRAME_RELATED_EXPR, dwarf, | |
12002 | REG_NOTES (par)); | |
12003 | return par; | |
e2c671ba RE |
12004 | } |
12005 | ||
9b66ebb1 | 12006 | |
3c7ad43e PB |
12007 | /* Return true if the current function needs to save/restore LR. */ |
12008 | ||
12009 | static bool | |
12010 | thumb_force_lr_save (void) | |
12011 | { | |
12012 | return !cfun->machine->lr_save_eliminated | |
12013 | && (!leaf_function_p () | |
12014 | || thumb_far_jump_used_p () | |
6fb5fa3c | 12015 | || df_regs_ever_live_p (LR_REGNUM)); |
3c7ad43e PB |
12016 | } |
12017 | ||
12018 | ||
095bb276 NC |
12019 | /* Compute the distance from register FROM to register TO. |
12020 | These can be the arg pointer (26), the soft frame pointer (25), | |
12021 | the stack pointer (13) or the hard frame pointer (11). | |
c9ca9b88 | 12022 | In thumb mode r7 is used as the soft frame pointer, if needed. |
095bb276 NC |
12023 | Typical stack layout looks like this: |
12024 | ||
12025 | old stack pointer -> | | | |
12026 | ---- | |
12027 | | | \ | |
12028 | | | saved arguments for | |
12029 | | | vararg functions | |
12030 | | | / | |
12031 | -- | |
12032 | hard FP & arg pointer -> | | \ | |
12033 | | | stack | |
12034 | | | frame | |
12035 | | | / | |
12036 | -- | |
12037 | | | \ | |
12038 | | | call saved | |
12039 | | | registers | |
12040 | soft frame pointer -> | | / | |
12041 | -- | |
12042 | | | \ | |
12043 | | | local | |
12044 | | | variables | |
2591db65 | 12045 | locals base pointer -> | | / |
095bb276 NC |
12046 | -- |
12047 | | | \ | |
12048 | | | outgoing | |
12049 | | | arguments | |
12050 | current stack pointer -> | | / | |
12051 | -- | |
12052 | ||
43aa4e05 | 12053 | For a given function some or all of these stack components |
095bb276 NC |
12054 | may not be needed, giving rise to the possibility of |
12055 | eliminating some of the registers. | |
12056 | ||
825dda42 | 12057 | The values returned by this function must reflect the behavior |
095bb276 NC |
12058 | of arm_expand_prologue() and arm_compute_save_reg_mask(). |
12059 | ||
12060 | The sign of the number returned reflects the direction of stack | |
12061 | growth, so the values are positive for all eliminations except | |
5848830f PB |
12062 | from the soft frame pointer to the hard frame pointer. |
12063 | ||
12064 | SFP may point just inside the local variables block to ensure correct | |
12065 | alignment. */ | |
12066 | ||
12067 | ||
12068 | /* Calculate stack offsets. These are used to calculate register elimination | |
954954d1 PB |
12069 | offsets and in prologue/epilogue code. Also calculates which registers |
12070 | should be saved. */ | |
5848830f PB |
12071 | |
12072 | static arm_stack_offsets * | |
12073 | arm_get_frame_offsets (void) | |
095bb276 | 12074 | { |
5848830f | 12075 | struct arm_stack_offsets *offsets; |
095bb276 | 12076 | unsigned long func_type; |
5848830f | 12077 | int leaf; |
5848830f | 12078 | int saved; |
954954d1 | 12079 | int core_saved; |
5848830f | 12080 | HOST_WIDE_INT frame_size; |
954954d1 | 12081 | int i; |
5848830f PB |
12082 | |
12083 | offsets = &cfun->machine->stack_offsets; | |
f676971a | 12084 | |
5848830f PB |
12085 | /* We need to know if we are a leaf function. Unfortunately, it |
12086 | is possible to be called after start_sequence has been called, | |
12087 | which causes get_insns to return the insns for the sequence, | |
12088 | not the function, which will cause leaf_function_p to return | |
12089 | the incorrect result. | |
095bb276 | 12090 | |
5848830f PB |
12091 | to know about leaf functions once reload has completed, and the |
12092 | frame size cannot be changed after that time, so we can safely | |
12093 | use the cached value. */ | |
12094 | ||
12095 | if (reload_completed) | |
12096 | return offsets; | |
12097 | ||
666c27b9 KH |
12098 | /* Initially this is the size of the local variables. It will translated |
12099 | into an offset once we have determined the size of preceding data. */ | |
5848830f PB |
12100 | frame_size = ROUND_UP_WORD (get_frame_size ()); |
12101 | ||
12102 | leaf = leaf_function_p (); | |
12103 | ||
12104 | /* Space for variadic functions. */ | |
38173d38 | 12105 | offsets->saved_args = crtl->args.pretend_args_size; |
5848830f | 12106 | |
5b3e6663 | 12107 | /* In Thumb mode this is incorrect, but never used. */ |
35596784 AJ |
12108 | offsets->frame = offsets->saved_args + (frame_pointer_needed ? 4 : 0) + |
12109 | arm_compute_static_chain_stack_bytes(); | |
5848830f | 12110 | |
5b3e6663 | 12111 | if (TARGET_32BIT) |
095bb276 | 12112 | { |
5848830f | 12113 | unsigned int regno; |
ef7112de | 12114 | |
954954d1 PB |
12115 | offsets->saved_regs_mask = arm_compute_save_reg_mask (); |
12116 | core_saved = bit_count (offsets->saved_regs_mask) * 4; | |
12117 | saved = core_saved; | |
5a9335ef | 12118 | |
5848830f PB |
12119 | /* We know that SP will be doubleword aligned on entry, and we must |
12120 | preserve that condition at any subroutine call. We also require the | |
12121 | soft frame pointer to be doubleword aligned. */ | |
12122 | ||
12123 | if (TARGET_REALLY_IWMMXT) | |
9b66ebb1 | 12124 | { |
5848830f PB |
12125 | /* Check for the call-saved iWMMXt registers. */ |
12126 | for (regno = FIRST_IWMMXT_REGNUM; | |
12127 | regno <= LAST_IWMMXT_REGNUM; | |
12128 | regno++) | |
6fb5fa3c | 12129 | if (df_regs_ever_live_p (regno) && ! call_used_regs[regno]) |
5848830f PB |
12130 | saved += 8; |
12131 | } | |
12132 | ||
12133 | func_type = arm_current_func_type (); | |
12134 | if (! IS_VOLATILE (func_type)) | |
12135 | { | |
12136 | /* Space for saved FPA registers. */ | |
12137 | for (regno = FIRST_FPA_REGNUM; regno <= LAST_FPA_REGNUM; regno++) | |
6fb5fa3c | 12138 | if (df_regs_ever_live_p (regno) && ! call_used_regs[regno]) |
5848830f PB |
12139 | saved += 12; |
12140 | ||
12141 | /* Space for saved VFP registers. */ | |
12142 | if (TARGET_HARD_FLOAT && TARGET_VFP) | |
9728c9d1 | 12143 | saved += arm_get_vfp_saved_size (); |
9b66ebb1 | 12144 | } |
5848830f | 12145 | } |
5b3e6663 | 12146 | else /* TARGET_THUMB1 */ |
5848830f | 12147 | { |
954954d1 PB |
12148 | offsets->saved_regs_mask = thumb1_compute_save_reg_mask (); |
12149 | core_saved = bit_count (offsets->saved_regs_mask) * 4; | |
12150 | saved = core_saved; | |
5848830f | 12151 | if (TARGET_BACKTRACE) |
57934c39 | 12152 | saved += 16; |
5848830f | 12153 | } |
9b66ebb1 | 12154 | |
5848830f | 12155 | /* Saved registers include the stack frame. */ |
35596784 AJ |
12156 | offsets->saved_regs = offsets->saved_args + saved + |
12157 | arm_compute_static_chain_stack_bytes(); | |
a2503645 | 12158 | offsets->soft_frame = offsets->saved_regs + CALLER_INTERWORKING_SLOT_SIZE; |
5848830f PB |
12159 | /* A leaf function does not need any stack alignment if it has nothing |
12160 | on the stack. */ | |
12161 | if (leaf && frame_size == 0) | |
12162 | { | |
12163 | offsets->outgoing_args = offsets->soft_frame; | |
a3a531ec | 12164 | offsets->locals_base = offsets->soft_frame; |
5848830f PB |
12165 | return offsets; |
12166 | } | |
12167 | ||
12168 | /* Ensure SFP has the correct alignment. */ | |
12169 | if (ARM_DOUBLEWORD_ALIGN | |
12170 | && (offsets->soft_frame & 7)) | |
954954d1 PB |
12171 | { |
12172 | offsets->soft_frame += 4; | |
12173 | /* Try to align stack by pushing an extra reg. Don't bother doing this | |
12174 | when there is a stack frame as the alignment will be rolled into | |
12175 | the normal stack adjustment. */ | |
38173d38 | 12176 | if (frame_size + crtl->outgoing_args_size == 0) |
954954d1 PB |
12177 | { |
12178 | int reg = -1; | |
12179 | ||
12180 | for (i = 4; i <= (TARGET_THUMB1 ? LAST_LO_REGNUM : 11); i++) | |
12181 | { | |
12182 | if ((offsets->saved_regs_mask & (1 << i)) == 0) | |
12183 | { | |
12184 | reg = i; | |
12185 | break; | |
12186 | } | |
12187 | } | |
12188 | ||
12189 | if (reg == -1 && arm_size_return_regs () <= 12 | |
e3b5732b | 12190 | && !crtl->tail_call_emit) |
954954d1 PB |
12191 | { |
12192 | /* Push/pop an argument register (r3) if all callee saved | |
12193 | registers are already being pushed. */ | |
12194 | reg = 3; | |
12195 | } | |
12196 | ||
12197 | if (reg != -1) | |
12198 | { | |
12199 | offsets->saved_regs += 4; | |
12200 | offsets->saved_regs_mask |= (1 << reg); | |
12201 | } | |
12202 | } | |
12203 | } | |
5848830f | 12204 | |
2591db65 RE |
12205 | offsets->locals_base = offsets->soft_frame + frame_size; |
12206 | offsets->outgoing_args = (offsets->locals_base | |
38173d38 | 12207 | + crtl->outgoing_args_size); |
5848830f PB |
12208 | |
12209 | if (ARM_DOUBLEWORD_ALIGN) | |
12210 | { | |
12211 | /* Ensure SP remains doubleword aligned. */ | |
12212 | if (offsets->outgoing_args & 7) | |
12213 | offsets->outgoing_args += 4; | |
e6d29d15 | 12214 | gcc_assert (!(offsets->outgoing_args & 7)); |
095bb276 NC |
12215 | } |
12216 | ||
5848830f PB |
12217 | return offsets; |
12218 | } | |
12219 | ||
12220 | ||
666c27b9 | 12221 | /* Calculate the relative offsets for the different stack pointers. Positive |
5848830f PB |
12222 | offsets are in the direction of stack growth. */ |
12223 | ||
b3f8d95d | 12224 | HOST_WIDE_INT |
5848830f PB |
12225 | arm_compute_initial_elimination_offset (unsigned int from, unsigned int to) |
12226 | { | |
12227 | arm_stack_offsets *offsets; | |
12228 | ||
12229 | offsets = arm_get_frame_offsets (); | |
095bb276 | 12230 | |
095bb276 NC |
12231 | /* OK, now we have enough information to compute the distances. |
12232 | There must be an entry in these switch tables for each pair | |
12233 | of registers in ELIMINABLE_REGS, even if some of the entries | |
12234 | seem to be redundant or useless. */ | |
12235 | switch (from) | |
12236 | { | |
12237 | case ARG_POINTER_REGNUM: | |
12238 | switch (to) | |
12239 | { | |
12240 | case THUMB_HARD_FRAME_POINTER_REGNUM: | |
12241 | return 0; | |
12242 | ||
12243 | case FRAME_POINTER_REGNUM: | |
12244 | /* This is the reverse of the soft frame pointer | |
12245 | to hard frame pointer elimination below. */ | |
5848830f | 12246 | return offsets->soft_frame - offsets->saved_args; |
095bb276 NC |
12247 | |
12248 | case ARM_HARD_FRAME_POINTER_REGNUM: | |
35596784 AJ |
12249 | /* This is only non-zero in the case where the static chain register |
12250 | is stored above the frame. */ | |
12251 | return offsets->frame - offsets->saved_args - 4; | |
095bb276 NC |
12252 | |
12253 | case STACK_POINTER_REGNUM: | |
12254 | /* If nothing has been pushed on the stack at all | |
12255 | then this will return -4. This *is* correct! */ | |
5848830f | 12256 | return offsets->outgoing_args - (offsets->saved_args + 4); |
095bb276 NC |
12257 | |
12258 | default: | |
e6d29d15 | 12259 | gcc_unreachable (); |
095bb276 | 12260 | } |
e6d29d15 | 12261 | gcc_unreachable (); |
095bb276 NC |
12262 | |
12263 | case FRAME_POINTER_REGNUM: | |
12264 | switch (to) | |
12265 | { | |
12266 | case THUMB_HARD_FRAME_POINTER_REGNUM: | |
12267 | return 0; | |
12268 | ||
12269 | case ARM_HARD_FRAME_POINTER_REGNUM: | |
12270 | /* The hard frame pointer points to the top entry in the | |
12271 | stack frame. The soft frame pointer to the bottom entry | |
12272 | in the stack frame. If there is no stack frame at all, | |
12273 | then they are identical. */ | |
5848830f PB |
12274 | |
12275 | return offsets->frame - offsets->soft_frame; | |
095bb276 NC |
12276 | |
12277 | case STACK_POINTER_REGNUM: | |
5848830f | 12278 | return offsets->outgoing_args - offsets->soft_frame; |
095bb276 NC |
12279 | |
12280 | default: | |
e6d29d15 | 12281 | gcc_unreachable (); |
095bb276 | 12282 | } |
e6d29d15 | 12283 | gcc_unreachable (); |
095bb276 NC |
12284 | |
12285 | default: | |
12286 | /* You cannot eliminate from the stack pointer. | |
12287 | In theory you could eliminate from the hard frame | |
12288 | pointer to the stack pointer, but this will never | |
12289 | happen, since if a stack frame is not needed the | |
12290 | hard frame pointer will never be used. */ | |
e6d29d15 | 12291 | gcc_unreachable (); |
095bb276 NC |
12292 | } |
12293 | } | |
12294 | ||
0977774b | 12295 | |
7a085dce | 12296 | /* Emit RTL to save coprocessor registers on function entry. Returns the |
5b3e6663 PB |
12297 | number of bytes pushed. */ |
12298 | ||
12299 | static int | |
12300 | arm_save_coproc_regs(void) | |
12301 | { | |
12302 | int saved_size = 0; | |
12303 | unsigned reg; | |
12304 | unsigned start_reg; | |
12305 | rtx insn; | |
12306 | ||
12307 | for (reg = LAST_IWMMXT_REGNUM; reg >= FIRST_IWMMXT_REGNUM; reg--) | |
6fb5fa3c | 12308 | if (df_regs_ever_live_p (reg) && ! call_used_regs[reg]) |
5b3e6663 PB |
12309 | { |
12310 | insn = gen_rtx_PRE_DEC (V2SImode, stack_pointer_rtx); | |
12311 | insn = gen_rtx_MEM (V2SImode, insn); | |
12312 | insn = emit_set_insn (insn, gen_rtx_REG (V2SImode, reg)); | |
12313 | RTX_FRAME_RELATED_P (insn) = 1; | |
12314 | saved_size += 8; | |
12315 | } | |
12316 | ||
12317 | /* Save any floating point call-saved registers used by this | |
12318 | function. */ | |
12319 | if (arm_fpu_arch == FPUTYPE_FPA_EMU2) | |
12320 | { | |
12321 | for (reg = LAST_FPA_REGNUM; reg >= FIRST_FPA_REGNUM; reg--) | |
6fb5fa3c | 12322 | if (df_regs_ever_live_p (reg) && !call_used_regs[reg]) |
5b3e6663 PB |
12323 | { |
12324 | insn = gen_rtx_PRE_DEC (XFmode, stack_pointer_rtx); | |
12325 | insn = gen_rtx_MEM (XFmode, insn); | |
12326 | insn = emit_set_insn (insn, gen_rtx_REG (XFmode, reg)); | |
12327 | RTX_FRAME_RELATED_P (insn) = 1; | |
12328 | saved_size += 12; | |
12329 | } | |
12330 | } | |
12331 | else | |
12332 | { | |
12333 | start_reg = LAST_FPA_REGNUM; | |
12334 | ||
12335 | for (reg = LAST_FPA_REGNUM; reg >= FIRST_FPA_REGNUM; reg--) | |
12336 | { | |
6fb5fa3c | 12337 | if (df_regs_ever_live_p (reg) && !call_used_regs[reg]) |
5b3e6663 PB |
12338 | { |
12339 | if (start_reg - reg == 3) | |
12340 | { | |
12341 | insn = emit_sfm (reg, 4); | |
12342 | RTX_FRAME_RELATED_P (insn) = 1; | |
12343 | saved_size += 48; | |
12344 | start_reg = reg - 1; | |
12345 | } | |
12346 | } | |
12347 | else | |
12348 | { | |
12349 | if (start_reg != reg) | |
12350 | { | |
12351 | insn = emit_sfm (reg + 1, start_reg - reg); | |
12352 | RTX_FRAME_RELATED_P (insn) = 1; | |
12353 | saved_size += (start_reg - reg) * 12; | |
12354 | } | |
12355 | start_reg = reg - 1; | |
12356 | } | |
12357 | } | |
12358 | ||
12359 | if (start_reg != reg) | |
12360 | { | |
12361 | insn = emit_sfm (reg + 1, start_reg - reg); | |
12362 | saved_size += (start_reg - reg) * 12; | |
12363 | RTX_FRAME_RELATED_P (insn) = 1; | |
12364 | } | |
12365 | } | |
12366 | if (TARGET_HARD_FLOAT && TARGET_VFP) | |
12367 | { | |
12368 | start_reg = FIRST_VFP_REGNUM; | |
12369 | ||
12370 | for (reg = FIRST_VFP_REGNUM; reg < LAST_VFP_REGNUM; reg += 2) | |
12371 | { | |
6fb5fa3c DB |
12372 | if ((!df_regs_ever_live_p (reg) || call_used_regs[reg]) |
12373 | && (!df_regs_ever_live_p (reg + 1) || call_used_regs[reg + 1])) | |
5b3e6663 PB |
12374 | { |
12375 | if (start_reg != reg) | |
12376 | saved_size += vfp_emit_fstmd (start_reg, | |
12377 | (reg - start_reg) / 2); | |
12378 | start_reg = reg + 2; | |
12379 | } | |
12380 | } | |
12381 | if (start_reg != reg) | |
12382 | saved_size += vfp_emit_fstmd (start_reg, | |
12383 | (reg - start_reg) / 2); | |
12384 | } | |
12385 | return saved_size; | |
12386 | } | |
12387 | ||
12388 | ||
12389 | /* Set the Thumb frame pointer from the stack pointer. */ | |
12390 | ||
12391 | static void | |
12392 | thumb_set_frame_pointer (arm_stack_offsets *offsets) | |
12393 | { | |
12394 | HOST_WIDE_INT amount; | |
12395 | rtx insn, dwarf; | |
12396 | ||
12397 | amount = offsets->outgoing_args - offsets->locals_base; | |
12398 | if (amount < 1024) | |
12399 | insn = emit_insn (gen_addsi3 (hard_frame_pointer_rtx, | |
12400 | stack_pointer_rtx, GEN_INT (amount))); | |
12401 | else | |
12402 | { | |
12403 | emit_insn (gen_movsi (hard_frame_pointer_rtx, GEN_INT (amount))); | |
f5c630c3 PB |
12404 | /* Thumb-2 RTL patterns expect sp as the first input. Thumb-1 |
12405 | expects the first two operands to be the same. */ | |
12406 | if (TARGET_THUMB2) | |
12407 | { | |
12408 | insn = emit_insn (gen_addsi3 (hard_frame_pointer_rtx, | |
12409 | stack_pointer_rtx, | |
12410 | hard_frame_pointer_rtx)); | |
12411 | } | |
12412 | else | |
12413 | { | |
12414 | insn = emit_insn (gen_addsi3 (hard_frame_pointer_rtx, | |
12415 | hard_frame_pointer_rtx, | |
12416 | stack_pointer_rtx)); | |
12417 | } | |
5b3e6663 PB |
12418 | dwarf = gen_rtx_SET (VOIDmode, hard_frame_pointer_rtx, |
12419 | plus_constant (stack_pointer_rtx, amount)); | |
12420 | RTX_FRAME_RELATED_P (dwarf) = 1; | |
12421 | REG_NOTES (insn) = gen_rtx_EXPR_LIST (REG_FRAME_RELATED_EXPR, dwarf, | |
12422 | REG_NOTES (insn)); | |
12423 | } | |
12424 | ||
12425 | RTX_FRAME_RELATED_P (insn) = 1; | |
12426 | } | |
12427 | ||
12428 | /* Generate the prologue instructions for entry into an ARM or Thumb-2 | |
12429 | function. */ | |
e2c671ba | 12430 | void |
e32bac5b | 12431 | arm_expand_prologue (void) |
e2c671ba | 12432 | { |
6d3d9133 | 12433 | rtx amount; |
2c849145 | 12434 | rtx insn; |
68dfd979 | 12435 | rtx ip_rtx; |
6d3d9133 NC |
12436 | unsigned long live_regs_mask; |
12437 | unsigned long func_type; | |
68dfd979 | 12438 | int fp_offset = 0; |
095bb276 | 12439 | int saved_pretend_args = 0; |
5848830f | 12440 | int saved_regs = 0; |
b9705a21 | 12441 | unsigned HOST_WIDE_INT args_to_push; |
5848830f | 12442 | arm_stack_offsets *offsets; |
d3236b4d | 12443 | |
6d3d9133 | 12444 | func_type = arm_current_func_type (); |
e2c671ba | 12445 | |
31fdb4d5 | 12446 | /* Naked functions don't have prologues. */ |
6d3d9133 | 12447 | if (IS_NAKED (func_type)) |
31fdb4d5 DE |
12448 | return; |
12449 | ||
095bb276 | 12450 | /* Make a copy of c_f_p_a_s as we may need to modify it locally. */ |
38173d38 | 12451 | args_to_push = crtl->args.pretend_args_size; |
f676971a | 12452 | |
6d3d9133 | 12453 | /* Compute which register we will have to save onto the stack. */ |
954954d1 PB |
12454 | offsets = arm_get_frame_offsets (); |
12455 | live_regs_mask = offsets->saved_regs_mask; | |
e2c671ba | 12456 | |
68dfd979 | 12457 | ip_rtx = gen_rtx_REG (SImode, IP_REGNUM); |
d3236b4d | 12458 | |
5b3e6663 PB |
12459 | if (IS_STACKALIGN (func_type)) |
12460 | { | |
12461 | rtx dwarf; | |
12462 | rtx r0; | |
12463 | rtx r1; | |
12464 | /* Handle a word-aligned stack pointer. We generate the following: | |
12465 | ||
12466 | mov r0, sp | |
12467 | bic r1, r0, #7 | |
12468 | mov sp, r1 | |
12469 | <save and restore r0 in normal prologue/epilogue> | |
12470 | mov sp, r0 | |
12471 | bx lr | |
12472 | ||
12473 | The unwinder doesn't need to know about the stack realignment. | |
12474 | Just tell it we saved SP in r0. */ | |
12475 | gcc_assert (TARGET_THUMB2 && !arm_arch_notm && args_to_push == 0); | |
12476 | ||
12477 | r0 = gen_rtx_REG (SImode, 0); | |
12478 | r1 = gen_rtx_REG (SImode, 1); | |
44bfa35b NF |
12479 | /* Use a real rtvec rather than NULL_RTVEC so the rest of the |
12480 | compiler won't choke. */ | |
12481 | dwarf = gen_rtx_UNSPEC (SImode, rtvec_alloc (0), UNSPEC_STACK_ALIGN); | |
5b3e6663 PB |
12482 | dwarf = gen_rtx_SET (VOIDmode, r0, dwarf); |
12483 | insn = gen_movsi (r0, stack_pointer_rtx); | |
12484 | RTX_FRAME_RELATED_P (insn) = 1; | |
12485 | REG_NOTES (insn) = gen_rtx_EXPR_LIST (REG_FRAME_RELATED_EXPR, | |
12486 | dwarf, REG_NOTES (insn)); | |
12487 | emit_insn (insn); | |
12488 | emit_insn (gen_andsi3 (r1, r0, GEN_INT (~(HOST_WIDE_INT)7))); | |
12489 | emit_insn (gen_movsi (stack_pointer_rtx, r1)); | |
12490 | } | |
12491 | ||
ec6237e4 PB |
12492 | /* For APCS frames, if IP register is clobbered |
12493 | when creating frame, save that register in a special | |
12494 | way. */ | |
12495 | if (TARGET_APCS_FRAME && frame_pointer_needed && TARGET_ARM) | |
e2c671ba | 12496 | { |
7b8b8ade NC |
12497 | if (IS_INTERRUPT (func_type)) |
12498 | { | |
12499 | /* Interrupt functions must not corrupt any registers. | |
12500 | Creating a frame pointer however, corrupts the IP | |
12501 | register, so we must push it first. */ | |
12502 | insn = emit_multi_reg_push (1 << IP_REGNUM); | |
121308d4 NC |
12503 | |
12504 | /* Do not set RTX_FRAME_RELATED_P on this insn. | |
12505 | The dwarf stack unwinding code only wants to see one | |
12506 | stack decrement per function, and this is not it. If | |
12507 | this instruction is labeled as being part of the frame | |
12508 | creation sequence then dwarf2out_frame_debug_expr will | |
e6d29d15 | 12509 | die when it encounters the assignment of IP to FP |
121308d4 NC |
12510 | later on, since the use of SP here establishes SP as |
12511 | the CFA register and not IP. | |
12512 | ||
12513 | Anyway this instruction is not really part of the stack | |
12514 | frame creation although it is part of the prologue. */ | |
7b8b8ade NC |
12515 | } |
12516 | else if (IS_NESTED (func_type)) | |
68dfd979 NC |
12517 | { |
12518 | /* The Static chain register is the same as the IP register | |
12519 | used as a scratch register during stack frame creation. | |
12520 | To get around this need to find somewhere to store IP | |
12521 | whilst the frame is being created. We try the following | |
12522 | places in order: | |
f676971a | 12523 | |
6d3d9133 | 12524 | 1. The last argument register. |
68dfd979 NC |
12525 | 2. A slot on the stack above the frame. (This only |
12526 | works if the function is not a varargs function). | |
095bb276 NC |
12527 | 3. Register r3, after pushing the argument registers |
12528 | onto the stack. | |
6d3d9133 | 12529 | |
34ce3d7b JM |
12530 | Note - we only need to tell the dwarf2 backend about the SP |
12531 | adjustment in the second variant; the static chain register | |
12532 | doesn't need to be unwound, as it doesn't contain a value | |
12533 | inherited from the caller. */ | |
d3236b4d | 12534 | |
6fb5fa3c | 12535 | if (df_regs_ever_live_p (3) == false) |
d66437c5 | 12536 | insn = emit_set_insn (gen_rtx_REG (SImode, 3), ip_rtx); |
095bb276 | 12537 | else if (args_to_push == 0) |
68dfd979 | 12538 | { |
35596784 AJ |
12539 | gcc_assert(arm_compute_static_chain_stack_bytes() == 4); |
12540 | saved_regs += 4; | |
12541 | ||
34ce3d7b | 12542 | rtx dwarf; |
34ce3d7b | 12543 | |
d66437c5 RE |
12544 | insn = gen_rtx_PRE_DEC (SImode, stack_pointer_rtx); |
12545 | insn = emit_set_insn (gen_frame_mem (SImode, insn), ip_rtx); | |
68dfd979 | 12546 | fp_offset = 4; |
34ce3d7b JM |
12547 | |
12548 | /* Just tell the dwarf backend that we adjusted SP. */ | |
12549 | dwarf = gen_rtx_SET (VOIDmode, stack_pointer_rtx, | |
d66437c5 RE |
12550 | plus_constant (stack_pointer_rtx, |
12551 | -fp_offset)); | |
34ce3d7b JM |
12552 | RTX_FRAME_RELATED_P (insn) = 1; |
12553 | REG_NOTES (insn) = gen_rtx_EXPR_LIST (REG_FRAME_RELATED_EXPR, | |
12554 | dwarf, REG_NOTES (insn)); | |
68dfd979 NC |
12555 | } |
12556 | else | |
095bb276 NC |
12557 | { |
12558 | /* Store the args on the stack. */ | |
3cb66fd7 | 12559 | if (cfun->machine->uses_anonymous_args) |
095bb276 NC |
12560 | insn = emit_multi_reg_push |
12561 | ((0xf0 >> (args_to_push / 4)) & 0xf); | |
12562 | else | |
12563 | insn = emit_insn | |
f676971a | 12564 | (gen_addsi3 (stack_pointer_rtx, stack_pointer_rtx, |
095bb276 NC |
12565 | GEN_INT (- args_to_push))); |
12566 | ||
12567 | RTX_FRAME_RELATED_P (insn) = 1; | |
12568 | ||
12569 | saved_pretend_args = 1; | |
12570 | fp_offset = args_to_push; | |
12571 | args_to_push = 0; | |
12572 | ||
12573 | /* Now reuse r3 to preserve IP. */ | |
d66437c5 | 12574 | emit_set_insn (gen_rtx_REG (SImode, 3), ip_rtx); |
095bb276 | 12575 | } |
68dfd979 NC |
12576 | } |
12577 | ||
d66437c5 RE |
12578 | insn = emit_set_insn (ip_rtx, |
12579 | plus_constant (stack_pointer_rtx, fp_offset)); | |
8e56560e | 12580 | RTX_FRAME_RELATED_P (insn) = 1; |
e2c671ba RE |
12581 | } |
12582 | ||
095bb276 | 12583 | if (args_to_push) |
e2c671ba | 12584 | { |
6d3d9133 | 12585 | /* Push the argument registers, or reserve space for them. */ |
3cb66fd7 | 12586 | if (cfun->machine->uses_anonymous_args) |
2c849145 | 12587 | insn = emit_multi_reg_push |
095bb276 | 12588 | ((0xf0 >> (args_to_push / 4)) & 0xf); |
e2c671ba | 12589 | else |
2c849145 | 12590 | insn = emit_insn |
f676971a | 12591 | (gen_addsi3 (stack_pointer_rtx, stack_pointer_rtx, |
095bb276 | 12592 | GEN_INT (- args_to_push))); |
2c849145 | 12593 | RTX_FRAME_RELATED_P (insn) = 1; |
e2c671ba RE |
12594 | } |
12595 | ||
06bea5aa | 12596 | /* If this is an interrupt service routine, and the link register |
ec6237e4 PB |
12597 | is going to be pushed, and we're not generating extra |
12598 | push of IP (needed when frame is needed and frame layout if apcs), | |
06bea5aa NC |
12599 | subtracting four from LR now will mean that the function return |
12600 | can be done with a single instruction. */ | |
3a7731fd | 12601 | if ((func_type == ARM_FT_ISR || func_type == ARM_FT_FIQ) |
06bea5aa | 12602 | && (live_regs_mask & (1 << LR_REGNUM)) != 0 |
ec6237e4 | 12603 | && !(frame_pointer_needed && TARGET_APCS_FRAME) |
a15908a4 | 12604 | && TARGET_ARM) |
d66437c5 RE |
12605 | { |
12606 | rtx lr = gen_rtx_REG (SImode, LR_REGNUM); | |
12607 | ||
12608 | emit_set_insn (lr, plus_constant (lr, -4)); | |
12609 | } | |
3a7731fd | 12610 | |
e2c671ba RE |
12611 | if (live_regs_mask) |
12612 | { | |
5848830f | 12613 | saved_regs += bit_count (live_regs_mask) * 4; |
954954d1 PB |
12614 | if (optimize_size && !frame_pointer_needed |
12615 | && saved_regs == offsets->saved_regs - offsets->saved_args) | |
12616 | { | |
12617 | /* If no coprocessor registers are being pushed and we don't have | |
12618 | to worry about a frame pointer then push extra registers to | |
12619 | create the stack frame. This is done is a way that does not | |
12620 | alter the frame layout, so is independent of the epilogue. */ | |
12621 | int n; | |
12622 | int frame; | |
12623 | n = 0; | |
12624 | while (n < 8 && (live_regs_mask & (1 << n)) == 0) | |
12625 | n++; | |
12626 | frame = offsets->outgoing_args - (offsets->saved_args + saved_regs); | |
12627 | if (frame && n * 4 >= frame) | |
12628 | { | |
12629 | n = frame / 4; | |
12630 | live_regs_mask |= (1 << n) - 1; | |
12631 | saved_regs += frame; | |
12632 | } | |
12633 | } | |
12634 | insn = emit_multi_reg_push (live_regs_mask); | |
2c849145 | 12635 | RTX_FRAME_RELATED_P (insn) = 1; |
e2c671ba | 12636 | } |
d5b7b3ae | 12637 | |
6d3d9133 | 12638 | if (! IS_VOLATILE (func_type)) |
5b3e6663 | 12639 | saved_regs += arm_save_coproc_regs (); |
b111229a | 12640 | |
5b3e6663 PB |
12641 | if (frame_pointer_needed && TARGET_ARM) |
12642 | { | |
12643 | /* Create the new frame pointer. */ | |
ec6237e4 | 12644 | if (TARGET_APCS_FRAME) |
9b66ebb1 | 12645 | { |
5b3e6663 PB |
12646 | insn = GEN_INT (-(4 + args_to_push + fp_offset)); |
12647 | insn = emit_insn (gen_addsi3 (hard_frame_pointer_rtx, ip_rtx, insn)); | |
12648 | RTX_FRAME_RELATED_P (insn) = 1; | |
9b66ebb1 | 12649 | |
5b3e6663 | 12650 | if (IS_NESTED (func_type)) |
9b66ebb1 | 12651 | { |
5b3e6663 | 12652 | /* Recover the static chain register. */ |
6fb5fa3c | 12653 | if (!df_regs_ever_live_p (3) |
5b3e6663 PB |
12654 | || saved_pretend_args) |
12655 | insn = gen_rtx_REG (SImode, 3); | |
38173d38 | 12656 | else /* if (crtl->args.pretend_args_size == 0) */ |
9b66ebb1 | 12657 | { |
5b3e6663 PB |
12658 | insn = plus_constant (hard_frame_pointer_rtx, 4); |
12659 | insn = gen_frame_mem (SImode, insn); | |
9b66ebb1 | 12660 | } |
5b3e6663 PB |
12661 | emit_set_insn (ip_rtx, insn); |
12662 | /* Add a USE to stop propagate_one_insn() from barfing. */ | |
12663 | emit_insn (gen_prologue_use (ip_rtx)); | |
9b66ebb1 | 12664 | } |
68dfd979 | 12665 | } |
ec6237e4 PB |
12666 | else |
12667 | { | |
12668 | insn = GEN_INT (saved_regs - 4); | |
12669 | insn = emit_insn (gen_addsi3 (hard_frame_pointer_rtx, | |
12670 | stack_pointer_rtx, insn)); | |
12671 | RTX_FRAME_RELATED_P (insn) = 1; | |
12672 | } | |
2c849145 | 12673 | } |
e2c671ba | 12674 | |
5848830f | 12675 | if (offsets->outgoing_args != offsets->saved_args + saved_regs) |
e2c671ba | 12676 | { |
745b9093 JM |
12677 | /* This add can produce multiple insns for a large constant, so we |
12678 | need to get tricky. */ | |
12679 | rtx last = get_last_insn (); | |
5848830f PB |
12680 | |
12681 | amount = GEN_INT (offsets->saved_args + saved_regs | |
12682 | - offsets->outgoing_args); | |
12683 | ||
2c849145 JM |
12684 | insn = emit_insn (gen_addsi3 (stack_pointer_rtx, stack_pointer_rtx, |
12685 | amount)); | |
745b9093 JM |
12686 | do |
12687 | { | |
12688 | last = last ? NEXT_INSN (last) : get_insns (); | |
12689 | RTX_FRAME_RELATED_P (last) = 1; | |
12690 | } | |
12691 | while (last != insn); | |
e04c2d6c RE |
12692 | |
12693 | /* If the frame pointer is needed, emit a special barrier that | |
12694 | will prevent the scheduler from moving stores to the frame | |
12695 | before the stack adjustment. */ | |
12696 | if (frame_pointer_needed) | |
3894f59e RE |
12697 | insn = emit_insn (gen_stack_tie (stack_pointer_rtx, |
12698 | hard_frame_pointer_rtx)); | |
e2c671ba RE |
12699 | } |
12700 | ||
876f13b0 | 12701 | |
5b3e6663 PB |
12702 | if (frame_pointer_needed && TARGET_THUMB2) |
12703 | thumb_set_frame_pointer (offsets); | |
12704 | ||
020a4035 | 12705 | if (flag_pic && arm_pic_register != INVALID_REGNUM) |
5b3e6663 PB |
12706 | { |
12707 | unsigned long mask; | |
12708 | ||
12709 | mask = live_regs_mask; | |
12710 | mask &= THUMB2_WORK_REGS; | |
12711 | if (!IS_NESTED (func_type)) | |
12712 | mask |= (1 << IP_REGNUM); | |
12713 | arm_load_pic_register (mask); | |
12714 | } | |
876f13b0 | 12715 | |
e2c671ba | 12716 | /* If we are profiling, make sure no instructions are scheduled before |
f5a1b0d2 | 12717 | the call to mcount. Similarly if the user has requested no |
74d9c39f DJ |
12718 | scheduling in the prolog. Similarly if we want non-call exceptions |
12719 | using the EABI unwinder, to prevent faulting instructions from being | |
12720 | swapped with a stack adjustment. */ | |
e3b5732b | 12721 | if (crtl->profile || !TARGET_SCHED_PROLOG |
74d9c39f | 12722 | || (ARM_EABI_UNWIND_TABLES && flag_non_call_exceptions)) |
e2c671ba | 12723 | emit_insn (gen_blockage ()); |
6f7ebcbb NC |
12724 | |
12725 | /* If the link register is being kept alive, with the return address in it, | |
12726 | then make sure that it does not get reused by the ce2 pass. */ | |
12727 | if ((live_regs_mask & (1 << LR_REGNUM)) == 0) | |
6fb5fa3c | 12728 | cfun->machine->lr_save_eliminated = 1; |
e2c671ba | 12729 | } |
cce8749e | 12730 | \f |
5b3e6663 PB |
12731 | /* Print condition code to STREAM. Helper function for arm_print_operand. */ |
12732 | static void | |
12733 | arm_print_condition (FILE *stream) | |
12734 | { | |
12735 | if (arm_ccfsm_state == 3 || arm_ccfsm_state == 4) | |
12736 | { | |
12737 | /* Branch conversion is not implemented for Thumb-2. */ | |
12738 | if (TARGET_THUMB) | |
12739 | { | |
12740 | output_operand_lossage ("predicated Thumb instruction"); | |
12741 | return; | |
12742 | } | |
12743 | if (current_insn_predicate != NULL) | |
12744 | { | |
12745 | output_operand_lossage | |
12746 | ("predicated instruction in conditional sequence"); | |
12747 | return; | |
12748 | } | |
12749 | ||
12750 | fputs (arm_condition_codes[arm_current_cc], stream); | |
12751 | } | |
12752 | else if (current_insn_predicate) | |
12753 | { | |
12754 | enum arm_cond_code code; | |
12755 | ||
12756 | if (TARGET_THUMB1) | |
12757 | { | |
12758 | output_operand_lossage ("predicated Thumb instruction"); | |
12759 | return; | |
12760 | } | |
12761 | ||
12762 | code = get_arm_condition_code (current_insn_predicate); | |
12763 | fputs (arm_condition_codes[code], stream); | |
12764 | } | |
12765 | } | |
12766 | ||
12767 | ||
9997d19d RE |
12768 | /* If CODE is 'd', then the X is a condition operand and the instruction |
12769 | should only be executed if the condition is true. | |
ddd5a7c1 | 12770 | if CODE is 'D', then the X is a condition operand and the instruction |
9997d19d RE |
12771 | should only be executed if the condition is false: however, if the mode |
12772 | of the comparison is CCFPEmode, then always execute the instruction -- we | |
12773 | do this because in these circumstances !GE does not necessarily imply LT; | |
12774 | in these cases the instruction pattern will take care to make sure that | |
12775 | an instruction containing %d will follow, thereby undoing the effects of | |
ddd5a7c1 | 12776 | doing this instruction unconditionally. |
9997d19d RE |
12777 | If CODE is 'N' then X is a floating point operand that must be negated |
12778 | before output. | |
12779 | If CODE is 'B' then output a bitwise inverted value of X (a const int). | |
12780 | If X is a REG and CODE is `M', output a ldm/stm style multi-reg. */ | |
9997d19d | 12781 | void |
e32bac5b | 12782 | arm_print_operand (FILE *stream, rtx x, int code) |
9997d19d RE |
12783 | { |
12784 | switch (code) | |
12785 | { | |
12786 | case '@': | |
f3139301 | 12787 | fputs (ASM_COMMENT_START, stream); |
9997d19d RE |
12788 | return; |
12789 | ||
d5b7b3ae RE |
12790 | case '_': |
12791 | fputs (user_label_prefix, stream); | |
12792 | return; | |
f676971a | 12793 | |
9997d19d | 12794 | case '|': |
f3139301 | 12795 | fputs (REGISTER_PREFIX, stream); |
9997d19d RE |
12796 | return; |
12797 | ||
12798 | case '?': | |
5b3e6663 PB |
12799 | arm_print_condition (stream); |
12800 | return; | |
cca0a211 | 12801 | |
5b3e6663 PB |
12802 | case '(': |
12803 | /* Nothing in unified syntax, otherwise the current condition code. */ | |
12804 | if (!TARGET_UNIFIED_ASM) | |
12805 | arm_print_condition (stream); | |
12806 | break; | |
12807 | ||
12808 | case ')': | |
12809 | /* The current condition code in unified syntax, otherwise nothing. */ | |
12810 | if (TARGET_UNIFIED_ASM) | |
12811 | arm_print_condition (stream); | |
12812 | break; | |
12813 | ||
12814 | case '.': | |
12815 | /* The current condition code for a condition code setting instruction. | |
7a085dce | 12816 | Preceded by 's' in unified syntax, otherwise followed by 's'. */ |
5b3e6663 PB |
12817 | if (TARGET_UNIFIED_ASM) |
12818 | { | |
12819 | fputc('s', stream); | |
12820 | arm_print_condition (stream); | |
cca0a211 | 12821 | } |
5b3e6663 | 12822 | else |
cca0a211 | 12823 | { |
5b3e6663 PB |
12824 | arm_print_condition (stream); |
12825 | fputc('s', stream); | |
cca0a211 | 12826 | } |
9997d19d RE |
12827 | return; |
12828 | ||
5b3e6663 PB |
12829 | case '!': |
12830 | /* If the instruction is conditionally executed then print | |
12831 | the current condition code, otherwise print 's'. */ | |
12832 | gcc_assert (TARGET_THUMB2 && TARGET_UNIFIED_ASM); | |
12833 | if (current_insn_predicate) | |
12834 | arm_print_condition (stream); | |
12835 | else | |
12836 | fputc('s', stream); | |
12837 | break; | |
12838 | ||
88f77cba | 12839 | /* %# is a "break" sequence. It doesn't output anything, but is used to |
cea618ac | 12840 | separate e.g. operand numbers from following text, if that text consists |
88f77cba JB |
12841 | of further digits which we don't want to be part of the operand |
12842 | number. */ | |
12843 | case '#': | |
12844 | return; | |
12845 | ||
9997d19d RE |
12846 | case 'N': |
12847 | { | |
12848 | REAL_VALUE_TYPE r; | |
12849 | REAL_VALUE_FROM_CONST_DOUBLE (r, x); | |
12850 | r = REAL_VALUE_NEGATE (r); | |
12851 | fprintf (stream, "%s", fp_const_from_val (&r)); | |
12852 | } | |
12853 | return; | |
12854 | ||
88f77cba JB |
12855 | /* An integer without a preceding # sign. */ |
12856 | case 'c': | |
12857 | gcc_assert (GET_CODE (x) == CONST_INT); | |
12858 | fprintf (stream, HOST_WIDE_INT_PRINT_DEC, INTVAL (x)); | |
12859 | return; | |
12860 | ||
9997d19d RE |
12861 | case 'B': |
12862 | if (GET_CODE (x) == CONST_INT) | |
4bc74ece NC |
12863 | { |
12864 | HOST_WIDE_INT val; | |
5895f793 | 12865 | val = ARM_SIGN_EXTEND (~INTVAL (x)); |
36ba9cb8 | 12866 | fprintf (stream, HOST_WIDE_INT_PRINT_DEC, val); |
4bc74ece | 12867 | } |
9997d19d RE |
12868 | else |
12869 | { | |
12870 | putc ('~', stream); | |
12871 | output_addr_const (stream, x); | |
12872 | } | |
12873 | return; | |
12874 | ||
5b3e6663 PB |
12875 | case 'L': |
12876 | /* The low 16 bits of an immediate constant. */ | |
12877 | fprintf (stream, HOST_WIDE_INT_PRINT_DEC, INTVAL(x) & 0xffff); | |
12878 | return; | |
12879 | ||
9997d19d RE |
12880 | case 'i': |
12881 | fprintf (stream, "%s", arithmetic_instr (x, 1)); | |
12882 | return; | |
12883 | ||
9b6b54e2 NC |
12884 | /* Truncate Cirrus shift counts. */ |
12885 | case 's': | |
12886 | if (GET_CODE (x) == CONST_INT) | |
12887 | { | |
12888 | fprintf (stream, HOST_WIDE_INT_PRINT_DEC, INTVAL (x) & 0x3f); | |
12889 | return; | |
12890 | } | |
12891 | arm_print_operand (stream, x, 0); | |
12892 | return; | |
12893 | ||
9997d19d RE |
12894 | case 'I': |
12895 | fprintf (stream, "%s", arithmetic_instr (x, 0)); | |
12896 | return; | |
12897 | ||
12898 | case 'S': | |
12899 | { | |
12900 | HOST_WIDE_INT val; | |
beed32b8 RE |
12901 | const char *shift; |
12902 | ||
12903 | if (!shift_operator (x, SImode)) | |
12904 | { | |
12905 | output_operand_lossage ("invalid shift operand"); | |
12906 | break; | |
12907 | } | |
12908 | ||
12909 | shift = shift_op (x, &val); | |
9997d19d | 12910 | |
e2c671ba RE |
12911 | if (shift) |
12912 | { | |
beed32b8 | 12913 | fprintf (stream, ", %s ", shift); |
e2c671ba RE |
12914 | if (val == -1) |
12915 | arm_print_operand (stream, XEXP (x, 1), 0); | |
12916 | else | |
4a0a75dd | 12917 | fprintf (stream, "#" HOST_WIDE_INT_PRINT_DEC, val); |
e2c671ba | 12918 | } |
9997d19d RE |
12919 | } |
12920 | return; | |
12921 | ||
d5b7b3ae | 12922 | /* An explanation of the 'Q', 'R' and 'H' register operands: |
f676971a | 12923 | |
d5b7b3ae RE |
12924 | In a pair of registers containing a DI or DF value the 'Q' |
12925 | operand returns the register number of the register containing | |
093354e0 | 12926 | the least significant part of the value. The 'R' operand returns |
d5b7b3ae RE |
12927 | the register number of the register containing the most |
12928 | significant part of the value. | |
f676971a | 12929 | |
d5b7b3ae RE |
12930 | The 'H' operand returns the higher of the two register numbers. |
12931 | On a run where WORDS_BIG_ENDIAN is true the 'H' operand is the | |
093354e0 | 12932 | same as the 'Q' operand, since the most significant part of the |
d5b7b3ae RE |
12933 | value is held in the lower number register. The reverse is true |
12934 | on systems where WORDS_BIG_ENDIAN is false. | |
f676971a | 12935 | |
d5b7b3ae RE |
12936 | The purpose of these operands is to distinguish between cases |
12937 | where the endian-ness of the values is important (for example | |
12938 | when they are added together), and cases where the endian-ness | |
12939 | is irrelevant, but the order of register operations is important. | |
12940 | For example when loading a value from memory into a register | |
12941 | pair, the endian-ness does not matter. Provided that the value | |
12942 | from the lower memory address is put into the lower numbered | |
12943 | register, and the value from the higher address is put into the | |
12944 | higher numbered register, the load will work regardless of whether | |
12945 | the value being loaded is big-wordian or little-wordian. The | |
12946 | order of the two register loads can matter however, if the address | |
12947 | of the memory location is actually held in one of the registers | |
12948 | being overwritten by the load. */ | |
c1c2bc04 | 12949 | case 'Q': |
22de4c3d RE |
12950 | if (GET_CODE (x) != REG || REGNO (x) > LAST_ARM_REGNUM) |
12951 | { | |
12952 | output_operand_lossage ("invalid operand for code '%c'", code); | |
12953 | return; | |
12954 | } | |
12955 | ||
d5b7b3ae | 12956 | asm_fprintf (stream, "%r", REGNO (x) + (WORDS_BIG_ENDIAN ? 1 : 0)); |
c1c2bc04 RE |
12957 | return; |
12958 | ||
9997d19d | 12959 | case 'R': |
22de4c3d RE |
12960 | if (GET_CODE (x) != REG || REGNO (x) > LAST_ARM_REGNUM) |
12961 | { | |
12962 | output_operand_lossage ("invalid operand for code '%c'", code); | |
12963 | return; | |
12964 | } | |
12965 | ||
d5b7b3ae RE |
12966 | asm_fprintf (stream, "%r", REGNO (x) + (WORDS_BIG_ENDIAN ? 0 : 1)); |
12967 | return; | |
12968 | ||
12969 | case 'H': | |
22de4c3d RE |
12970 | if (GET_CODE (x) != REG || REGNO (x) > LAST_ARM_REGNUM) |
12971 | { | |
12972 | output_operand_lossage ("invalid operand for code '%c'", code); | |
12973 | return; | |
12974 | } | |
12975 | ||
d5b7b3ae | 12976 | asm_fprintf (stream, "%r", REGNO (x) + 1); |
9997d19d RE |
12977 | return; |
12978 | ||
88f77cba JB |
12979 | case 'J': |
12980 | if (GET_CODE (x) != REG || REGNO (x) > LAST_ARM_REGNUM) | |
12981 | { | |
12982 | output_operand_lossage ("invalid operand for code '%c'", code); | |
12983 | return; | |
12984 | } | |
12985 | ||
12986 | asm_fprintf (stream, "%r", REGNO (x) + (WORDS_BIG_ENDIAN ? 3 : 2)); | |
12987 | return; | |
12988 | ||
12989 | case 'K': | |
12990 | if (GET_CODE (x) != REG || REGNO (x) > LAST_ARM_REGNUM) | |
12991 | { | |
12992 | output_operand_lossage ("invalid operand for code '%c'", code); | |
12993 | return; | |
12994 | } | |
12995 | ||
12996 | asm_fprintf (stream, "%r", REGNO (x) + (WORDS_BIG_ENDIAN ? 2 : 3)); | |
12997 | return; | |
12998 | ||
9997d19d | 12999 | case 'm': |
f676971a | 13000 | asm_fprintf (stream, "%r", |
d5b7b3ae RE |
13001 | GET_CODE (XEXP (x, 0)) == REG |
13002 | ? REGNO (XEXP (x, 0)) : REGNO (XEXP (XEXP (x, 0), 0))); | |
9997d19d RE |
13003 | return; |
13004 | ||
13005 | case 'M': | |
dd18ae56 | 13006 | asm_fprintf (stream, "{%r-%r}", |
d5b7b3ae | 13007 | REGNO (x), |
e9d7b180 | 13008 | REGNO (x) + ARM_NUM_REGS (GET_MODE (x)) - 1); |
9997d19d RE |
13009 | return; |
13010 | ||
88f77cba JB |
13011 | /* Like 'M', but writing doubleword vector registers, for use by Neon |
13012 | insns. */ | |
13013 | case 'h': | |
13014 | { | |
13015 | int regno = (REGNO (x) - FIRST_VFP_REGNUM) / 2; | |
13016 | int numregs = ARM_NUM_REGS (GET_MODE (x)) / 2; | |
13017 | if (numregs == 1) | |
13018 | asm_fprintf (stream, "{d%d}", regno); | |
13019 | else | |
13020 | asm_fprintf (stream, "{d%d-d%d}", regno, regno + numregs - 1); | |
13021 | } | |
13022 | return; | |
13023 | ||
9997d19d | 13024 | case 'd': |
64e92a26 RE |
13025 | /* CONST_TRUE_RTX means always -- that's the default. */ |
13026 | if (x == const_true_rtx) | |
d5b7b3ae | 13027 | return; |
f676971a | 13028 | |
22de4c3d RE |
13029 | if (!COMPARISON_P (x)) |
13030 | { | |
13031 | output_operand_lossage ("invalid operand for code '%c'", code); | |
13032 | return; | |
13033 | } | |
13034 | ||
defc0463 RE |
13035 | fputs (arm_condition_codes[get_arm_condition_code (x)], |
13036 | stream); | |
9997d19d RE |
13037 | return; |
13038 | ||
13039 | case 'D': | |
112cdef5 | 13040 | /* CONST_TRUE_RTX means not always -- i.e. never. We shouldn't ever |
64e92a26 RE |
13041 | want to do that. */ |
13042 | if (x == const_true_rtx) | |
22de4c3d | 13043 | { |
4dad0aca | 13044 | output_operand_lossage ("instruction never executed"); |
22de4c3d RE |
13045 | return; |
13046 | } | |
13047 | if (!COMPARISON_P (x)) | |
13048 | { | |
13049 | output_operand_lossage ("invalid operand for code '%c'", code); | |
13050 | return; | |
13051 | } | |
d5b7b3ae | 13052 | |
defc0463 RE |
13053 | fputs (arm_condition_codes[ARM_INVERSE_CONDITION_CODE |
13054 | (get_arm_condition_code (x))], | |
13055 | stream); | |
9997d19d RE |
13056 | return; |
13057 | ||
9b6b54e2 NC |
13058 | /* Cirrus registers can be accessed in a variety of ways: |
13059 | single floating point (f) | |
13060 | double floating point (d) | |
13061 | 32bit integer (fx) | |
13062 | 64bit integer (dx). */ | |
13063 | case 'W': /* Cirrus register in F mode. */ | |
13064 | case 'X': /* Cirrus register in D mode. */ | |
13065 | case 'Y': /* Cirrus register in FX mode. */ | |
13066 | case 'Z': /* Cirrus register in DX mode. */ | |
e6d29d15 NS |
13067 | gcc_assert (GET_CODE (x) == REG |
13068 | && REGNO_REG_CLASS (REGNO (x)) == CIRRUS_REGS); | |
9b6b54e2 NC |
13069 | |
13070 | fprintf (stream, "mv%s%s", | |
13071 | code == 'W' ? "f" | |
13072 | : code == 'X' ? "d" | |
13073 | : code == 'Y' ? "fx" : "dx", reg_names[REGNO (x)] + 2); | |
13074 | ||
13075 | return; | |
13076 | ||
13077 | /* Print cirrus register in the mode specified by the register's mode. */ | |
13078 | case 'V': | |
13079 | { | |
13080 | int mode = GET_MODE (x); | |
13081 | ||
13082 | if (GET_CODE (x) != REG || REGNO_REG_CLASS (REGNO (x)) != CIRRUS_REGS) | |
22de4c3d RE |
13083 | { |
13084 | output_operand_lossage ("invalid operand for code '%c'", code); | |
13085 | return; | |
13086 | } | |
9b6b54e2 NC |
13087 | |
13088 | fprintf (stream, "mv%s%s", | |
13089 | mode == DFmode ? "d" | |
13090 | : mode == SImode ? "fx" | |
13091 | : mode == DImode ? "dx" | |
13092 | : "f", reg_names[REGNO (x)] + 2); | |
13093 | ||
13094 | return; | |
13095 | } | |
13096 | ||
5a9335ef NC |
13097 | case 'U': |
13098 | if (GET_CODE (x) != REG | |
13099 | || REGNO (x) < FIRST_IWMMXT_GR_REGNUM | |
13100 | || REGNO (x) > LAST_IWMMXT_GR_REGNUM) | |
13101 | /* Bad value for wCG register number. */ | |
22de4c3d RE |
13102 | { |
13103 | output_operand_lossage ("invalid operand for code '%c'", code); | |
13104 | return; | |
13105 | } | |
13106 | ||
5a9335ef NC |
13107 | else |
13108 | fprintf (stream, "%d", REGNO (x) - FIRST_IWMMXT_GR_REGNUM); | |
13109 | return; | |
13110 | ||
13111 | /* Print an iWMMXt control register name. */ | |
13112 | case 'w': | |
13113 | if (GET_CODE (x) != CONST_INT | |
13114 | || INTVAL (x) < 0 | |
13115 | || INTVAL (x) >= 16) | |
13116 | /* Bad value for wC register number. */ | |
22de4c3d RE |
13117 | { |
13118 | output_operand_lossage ("invalid operand for code '%c'", code); | |
13119 | return; | |
13120 | } | |
13121 | ||
5a9335ef NC |
13122 | else |
13123 | { | |
13124 | static const char * wc_reg_names [16] = | |
13125 | { | |
13126 | "wCID", "wCon", "wCSSF", "wCASF", | |
13127 | "wC4", "wC5", "wC6", "wC7", | |
13128 | "wCGR0", "wCGR1", "wCGR2", "wCGR3", | |
13129 | "wC12", "wC13", "wC14", "wC15" | |
13130 | }; | |
f676971a | 13131 | |
5a9335ef NC |
13132 | fprintf (stream, wc_reg_names [INTVAL (x)]); |
13133 | } | |
13134 | return; | |
13135 | ||
88f77cba | 13136 | /* Print a VFP/Neon double precision or quad precision register name. */ |
9b66ebb1 | 13137 | case 'P': |
88f77cba | 13138 | case 'q': |
9b66ebb1 PB |
13139 | { |
13140 | int mode = GET_MODE (x); | |
88f77cba JB |
13141 | int is_quad = (code == 'q'); |
13142 | int regno; | |
9b66ebb1 | 13143 | |
88f77cba | 13144 | if (GET_MODE_SIZE (mode) != (is_quad ? 16 : 8)) |
22de4c3d RE |
13145 | { |
13146 | output_operand_lossage ("invalid operand for code '%c'", code); | |
13147 | return; | |
13148 | } | |
9b66ebb1 PB |
13149 | |
13150 | if (GET_CODE (x) != REG | |
13151 | || !IS_VFP_REGNUM (REGNO (x))) | |
22de4c3d RE |
13152 | { |
13153 | output_operand_lossage ("invalid operand for code '%c'", code); | |
13154 | return; | |
13155 | } | |
9b66ebb1 | 13156 | |
88f77cba JB |
13157 | regno = REGNO (x); |
13158 | if ((is_quad && !NEON_REGNO_OK_FOR_QUAD (regno)) | |
13159 | || (!is_quad && !VFP_REGNO_OK_FOR_DOUBLE (regno))) | |
22de4c3d RE |
13160 | { |
13161 | output_operand_lossage ("invalid operand for code '%c'", code); | |
13162 | return; | |
13163 | } | |
9b66ebb1 | 13164 | |
88f77cba JB |
13165 | fprintf (stream, "%c%d", is_quad ? 'q' : 'd', |
13166 | (regno - FIRST_VFP_REGNUM) >> (is_quad ? 2 : 1)); | |
13167 | } | |
13168 | return; | |
13169 | ||
13170 | /* These two codes print the low/high doubleword register of a Neon quad | |
13171 | register, respectively. For pair-structure types, can also print | |
13172 | low/high quadword registers. */ | |
13173 | case 'e': | |
13174 | case 'f': | |
13175 | { | |
13176 | int mode = GET_MODE (x); | |
13177 | int regno; | |
13178 | ||
13179 | if ((GET_MODE_SIZE (mode) != 16 | |
13180 | && GET_MODE_SIZE (mode) != 32) || GET_CODE (x) != REG) | |
13181 | { | |
13182 | output_operand_lossage ("invalid operand for code '%c'", code); | |
13183 | return; | |
13184 | } | |
13185 | ||
13186 | regno = REGNO (x); | |
13187 | if (!NEON_REGNO_OK_FOR_QUAD (regno)) | |
13188 | { | |
13189 | output_operand_lossage ("invalid operand for code '%c'", code); | |
13190 | return; | |
13191 | } | |
13192 | ||
13193 | if (GET_MODE_SIZE (mode) == 16) | |
13194 | fprintf (stream, "d%d", ((regno - FIRST_VFP_REGNUM) >> 1) | |
13195 | + (code == 'f' ? 1 : 0)); | |
13196 | else | |
13197 | fprintf (stream, "q%d", ((regno - FIRST_VFP_REGNUM) >> 2) | |
13198 | + (code == 'f' ? 1 : 0)); | |
9b66ebb1 PB |
13199 | } |
13200 | return; | |
13201 | ||
f1adb0a9 JB |
13202 | /* Print a VFPv3 floating-point constant, represented as an integer |
13203 | index. */ | |
13204 | case 'G': | |
13205 | { | |
13206 | int index = vfp3_const_double_index (x); | |
13207 | gcc_assert (index != -1); | |
13208 | fprintf (stream, "%d", index); | |
13209 | } | |
13210 | return; | |
13211 | ||
88f77cba JB |
13212 | /* Print bits representing opcode features for Neon. |
13213 | ||
13214 | Bit 0 is 1 for signed, 0 for unsigned. Floats count as signed | |
13215 | and polynomials as unsigned. | |
13216 | ||
13217 | Bit 1 is 1 for floats and polynomials, 0 for ordinary integers. | |
13218 | ||
13219 | Bit 2 is 1 for rounding functions, 0 otherwise. */ | |
13220 | ||
13221 | /* Identify the type as 's', 'u', 'p' or 'f'. */ | |
13222 | case 'T': | |
13223 | { | |
13224 | HOST_WIDE_INT bits = INTVAL (x); | |
13225 | fputc ("uspf"[bits & 3], stream); | |
13226 | } | |
13227 | return; | |
13228 | ||
13229 | /* Likewise, but signed and unsigned integers are both 'i'. */ | |
13230 | case 'F': | |
13231 | { | |
13232 | HOST_WIDE_INT bits = INTVAL (x); | |
13233 | fputc ("iipf"[bits & 3], stream); | |
13234 | } | |
13235 | return; | |
13236 | ||
13237 | /* As for 'T', but emit 'u' instead of 'p'. */ | |
13238 | case 't': | |
13239 | { | |
13240 | HOST_WIDE_INT bits = INTVAL (x); | |
13241 | fputc ("usuf"[bits & 3], stream); | |
13242 | } | |
13243 | return; | |
13244 | ||
13245 | /* Bit 2: rounding (vs none). */ | |
13246 | case 'O': | |
13247 | { | |
13248 | HOST_WIDE_INT bits = INTVAL (x); | |
13249 | fputs ((bits & 4) != 0 ? "r" : "", stream); | |
13250 | } | |
13251 | return; | |
13252 | ||
9997d19d RE |
13253 | default: |
13254 | if (x == 0) | |
22de4c3d RE |
13255 | { |
13256 | output_operand_lossage ("missing operand"); | |
13257 | return; | |
13258 | } | |
9997d19d | 13259 | |
e6d29d15 | 13260 | switch (GET_CODE (x)) |
9997d19d | 13261 | { |
e6d29d15 NS |
13262 | case REG: |
13263 | asm_fprintf (stream, "%r", REGNO (x)); | |
13264 | break; | |
13265 | ||
13266 | case MEM: | |
9997d19d RE |
13267 | output_memory_reference_mode = GET_MODE (x); |
13268 | output_address (XEXP (x, 0)); | |
e6d29d15 NS |
13269 | break; |
13270 | ||
13271 | case CONST_DOUBLE: | |
88f77cba JB |
13272 | if (TARGET_NEON) |
13273 | { | |
13274 | char fpstr[20]; | |
13275 | real_to_decimal (fpstr, CONST_DOUBLE_REAL_VALUE (x), | |
13276 | sizeof (fpstr), 0, 1); | |
13277 | fprintf (stream, "#%s", fpstr); | |
13278 | } | |
13279 | else | |
13280 | fprintf (stream, "#%s", fp_immediate_constant (x)); | |
e6d29d15 NS |
13281 | break; |
13282 | ||
13283 | default: | |
13284 | gcc_assert (GET_CODE (x) != NEG); | |
9997d19d RE |
13285 | fputc ('#', stream); |
13286 | output_addr_const (stream, x); | |
e6d29d15 | 13287 | break; |
9997d19d RE |
13288 | } |
13289 | } | |
13290 | } | |
cce8749e | 13291 | \f |
301d03af RS |
13292 | /* Target hook for assembling integer objects. The ARM version needs to |
13293 | handle word-sized values specially. */ | |
301d03af | 13294 | static bool |
e32bac5b | 13295 | arm_assemble_integer (rtx x, unsigned int size, int aligned_p) |
301d03af | 13296 | { |
88f77cba JB |
13297 | enum machine_mode mode; |
13298 | ||
301d03af RS |
13299 | if (size == UNITS_PER_WORD && aligned_p) |
13300 | { | |
13301 | fputs ("\t.word\t", asm_out_file); | |
13302 | output_addr_const (asm_out_file, x); | |
13303 | ||
13304 | /* Mark symbols as position independent. We only do this in the | |
d6b4baa4 | 13305 | .text segment, not in the .data segment. */ |
301d03af RS |
13306 | if (NEED_GOT_RELOC && flag_pic && making_const_table && |
13307 | (GET_CODE (x) == SYMBOL_REF || GET_CODE (x) == LABEL_REF)) | |
13308 | { | |
9403b7f7 RS |
13309 | /* See legitimize_pic_address for an explanation of the |
13310 | TARGET_VXWORKS_RTP check. */ | |
13311 | if (TARGET_VXWORKS_RTP | |
13312 | || (GET_CODE (x) == SYMBOL_REF && !SYMBOL_REF_LOCAL_P (x))) | |
301d03af | 13313 | fputs ("(GOT)", asm_out_file); |
9403b7f7 RS |
13314 | else |
13315 | fputs ("(GOTOFF)", asm_out_file); | |
301d03af RS |
13316 | } |
13317 | fputc ('\n', asm_out_file); | |
13318 | return true; | |
13319 | } | |
1d6e90ac | 13320 | |
88f77cba JB |
13321 | mode = GET_MODE (x); |
13322 | ||
13323 | if (arm_vector_mode_supported_p (mode)) | |
5a9335ef NC |
13324 | { |
13325 | int i, units; | |
88f77cba | 13326 | unsigned int invmask = 0, parts_per_word; |
5a9335ef | 13327 | |
e6d29d15 | 13328 | gcc_assert (GET_CODE (x) == CONST_VECTOR); |
5a9335ef NC |
13329 | |
13330 | units = CONST_VECTOR_NUNITS (x); | |
88f77cba | 13331 | size = GET_MODE_SIZE (GET_MODE_INNER (mode)); |
5a9335ef | 13332 | |
88f77cba JB |
13333 | /* For big-endian Neon vectors, we must permute the vector to the form |
13334 | which, when loaded by a VLDR or VLDM instruction, will give a vector | |
13335 | with the elements in the right order. */ | |
13336 | if (TARGET_NEON && WORDS_BIG_ENDIAN) | |
13337 | { | |
13338 | parts_per_word = UNITS_PER_WORD / size; | |
13339 | /* FIXME: This might be wrong for 64-bit vector elements, but we don't | |
13340 | support those anywhere yet. */ | |
13341 | invmask = (parts_per_word == 0) ? 0 : (1 << (parts_per_word - 1)) - 1; | |
13342 | } | |
5a9335ef | 13343 | |
88f77cba JB |
13344 | if (GET_MODE_CLASS (mode) == MODE_VECTOR_INT) |
13345 | for (i = 0; i < units; i++) | |
13346 | { | |
13347 | rtx elt = CONST_VECTOR_ELT (x, i ^ invmask); | |
13348 | assemble_integer | |
13349 | (elt, size, i == 0 ? BIGGEST_ALIGNMENT : size * BITS_PER_UNIT, 1); | |
13350 | } | |
13351 | else | |
13352 | for (i = 0; i < units; i++) | |
13353 | { | |
13354 | rtx elt = CONST_VECTOR_ELT (x, i); | |
13355 | REAL_VALUE_TYPE rval; | |
5a9335ef | 13356 | |
88f77cba JB |
13357 | REAL_VALUE_FROM_CONST_DOUBLE (rval, elt); |
13358 | ||
13359 | assemble_real | |
13360 | (rval, GET_MODE_INNER (mode), | |
13361 | i == 0 ? BIGGEST_ALIGNMENT : size * BITS_PER_UNIT); | |
13362 | } | |
5a9335ef NC |
13363 | |
13364 | return true; | |
13365 | } | |
13366 | ||
301d03af RS |
13367 | return default_assemble_integer (x, size, aligned_p); |
13368 | } | |
7abc66b1 | 13369 | |
7abc66b1 | 13370 | static void |
9f296620 | 13371 | arm_elf_asm_cdtor (rtx symbol, int priority, bool is_ctor) |
7abc66b1 | 13372 | { |
50603eed PB |
13373 | section *s; |
13374 | ||
7abc66b1 JB |
13375 | if (!TARGET_AAPCS_BASED) |
13376 | { | |
9f296620 MM |
13377 | (is_ctor ? |
13378 | default_named_section_asm_out_constructor | |
13379 | : default_named_section_asm_out_destructor) (symbol, priority); | |
7abc66b1 JB |
13380 | return; |
13381 | } | |
13382 | ||
13383 | /* Put these in the .init_array section, using a special relocation. */ | |
50603eed PB |
13384 | if (priority != DEFAULT_INIT_PRIORITY) |
13385 | { | |
13386 | char buf[18]; | |
9f296620 MM |
13387 | sprintf (buf, "%s.%.5u", |
13388 | is_ctor ? ".init_array" : ".fini_array", | |
13389 | priority); | |
50603eed PB |
13390 | s = get_section (buf, SECTION_WRITE, NULL_TREE); |
13391 | } | |
9f296620 | 13392 | else if (is_ctor) |
50603eed | 13393 | s = ctors_section; |
9f296620 MM |
13394 | else |
13395 | s = dtors_section; | |
50603eed PB |
13396 | |
13397 | switch_to_section (s); | |
7abc66b1 JB |
13398 | assemble_align (POINTER_SIZE); |
13399 | fputs ("\t.word\t", asm_out_file); | |
13400 | output_addr_const (asm_out_file, symbol); | |
13401 | fputs ("(target1)\n", asm_out_file); | |
13402 | } | |
9f296620 MM |
13403 | |
13404 | /* Add a function to the list of static constructors. */ | |
13405 | ||
13406 | static void | |
13407 | arm_elf_asm_constructor (rtx symbol, int priority) | |
13408 | { | |
13409 | arm_elf_asm_cdtor (symbol, priority, /*is_ctor=*/true); | |
13410 | } | |
13411 | ||
13412 | /* Add a function to the list of static destructors. */ | |
13413 | ||
13414 | static void | |
13415 | arm_elf_asm_destructor (rtx symbol, int priority) | |
13416 | { | |
13417 | arm_elf_asm_cdtor (symbol, priority, /*is_ctor=*/false); | |
13418 | } | |
301d03af | 13419 | \f |
cce8749e CH |
13420 | /* A finite state machine takes care of noticing whether or not instructions |
13421 | can be conditionally executed, and thus decrease execution time and code | |
13422 | size by deleting branch instructions. The fsm is controlled by | |
13423 | final_prescan_insn, and controls the actions of ASM_OUTPUT_OPCODE. */ | |
13424 | ||
13425 | /* The state of the fsm controlling condition codes are: | |
13426 | 0: normal, do nothing special | |
13427 | 1: make ASM_OUTPUT_OPCODE not output this instruction | |
13428 | 2: make ASM_OUTPUT_OPCODE not output this instruction | |
13429 | 3: make instructions conditional | |
13430 | 4: make instructions conditional | |
13431 | ||
13432 | State transitions (state->state by whom under condition): | |
13433 | 0 -> 1 final_prescan_insn if the `target' is a label | |
13434 | 0 -> 2 final_prescan_insn if the `target' is an unconditional branch | |
13435 | 1 -> 3 ASM_OUTPUT_OPCODE after not having output the conditional branch | |
13436 | 2 -> 4 ASM_OUTPUT_OPCODE after not having output the conditional branch | |
4977bab6 | 13437 | 3 -> 0 (*targetm.asm_out.internal_label) if the `target' label is reached |
cce8749e CH |
13438 | (the target label has CODE_LABEL_NUMBER equal to arm_target_label). |
13439 | 4 -> 0 final_prescan_insn if the `target' unconditional branch is reached | |
13440 | (the target insn is arm_target_insn). | |
13441 | ||
ff9940b0 RE |
13442 | If the jump clobbers the conditions then we use states 2 and 4. |
13443 | ||
13444 | A similar thing can be done with conditional return insns. | |
13445 | ||
cce8749e CH |
13446 | XXX In case the `target' is an unconditional branch, this conditionalising |
13447 | of the instructions always reduces code size, but not always execution | |
13448 | time. But then, I want to reduce the code size to somewhere near what | |
13449 | /bin/cc produces. */ | |
13450 | ||
5b3e6663 PB |
13451 | /* In addition to this, state is maintained for Thumb-2 COND_EXEC |
13452 | instructions. When a COND_EXEC instruction is seen the subsequent | |
13453 | instructions are scanned so that multiple conditional instructions can be | |
13454 | combined into a single IT block. arm_condexec_count and arm_condexec_mask | |
13455 | specify the length and true/false mask for the IT block. These will be | |
13456 | decremented/zeroed by arm_asm_output_opcode as the insns are output. */ | |
13457 | ||
cce8749e CH |
13458 | /* Returns the index of the ARM condition code string in |
13459 | `arm_condition_codes'. COMPARISON should be an rtx like | |
13460 | `(eq (...) (...))'. */ | |
84ed5e79 | 13461 | static enum arm_cond_code |
e32bac5b | 13462 | get_arm_condition_code (rtx comparison) |
cce8749e | 13463 | { |
5165176d | 13464 | enum machine_mode mode = GET_MODE (XEXP (comparison, 0)); |
1d6e90ac NC |
13465 | int code; |
13466 | enum rtx_code comp_code = GET_CODE (comparison); | |
5165176d RE |
13467 | |
13468 | if (GET_MODE_CLASS (mode) != MODE_CC) | |
84ed5e79 | 13469 | mode = SELECT_CC_MODE (comp_code, XEXP (comparison, 0), |
5165176d RE |
13470 | XEXP (comparison, 1)); |
13471 | ||
13472 | switch (mode) | |
cce8749e | 13473 | { |
84ed5e79 RE |
13474 | case CC_DNEmode: code = ARM_NE; goto dominance; |
13475 | case CC_DEQmode: code = ARM_EQ; goto dominance; | |
13476 | case CC_DGEmode: code = ARM_GE; goto dominance; | |
13477 | case CC_DGTmode: code = ARM_GT; goto dominance; | |
13478 | case CC_DLEmode: code = ARM_LE; goto dominance; | |
13479 | case CC_DLTmode: code = ARM_LT; goto dominance; | |
13480 | case CC_DGEUmode: code = ARM_CS; goto dominance; | |
13481 | case CC_DGTUmode: code = ARM_HI; goto dominance; | |
13482 | case CC_DLEUmode: code = ARM_LS; goto dominance; | |
13483 | case CC_DLTUmode: code = ARM_CC; | |
13484 | ||
13485 | dominance: | |
e6d29d15 | 13486 | gcc_assert (comp_code == EQ || comp_code == NE); |
84ed5e79 RE |
13487 | |
13488 | if (comp_code == EQ) | |
13489 | return ARM_INVERSE_CONDITION_CODE (code); | |
13490 | return code; | |
13491 | ||
5165176d | 13492 | case CC_NOOVmode: |
84ed5e79 | 13493 | switch (comp_code) |
5165176d | 13494 | { |
84ed5e79 RE |
13495 | case NE: return ARM_NE; |
13496 | case EQ: return ARM_EQ; | |
13497 | case GE: return ARM_PL; | |
13498 | case LT: return ARM_MI; | |
e6d29d15 | 13499 | default: gcc_unreachable (); |
5165176d RE |
13500 | } |
13501 | ||
13502 | case CC_Zmode: | |
84ed5e79 | 13503 | switch (comp_code) |
5165176d | 13504 | { |
84ed5e79 RE |
13505 | case NE: return ARM_NE; |
13506 | case EQ: return ARM_EQ; | |
e6d29d15 | 13507 | default: gcc_unreachable (); |
5165176d RE |
13508 | } |
13509 | ||
defc0463 RE |
13510 | case CC_Nmode: |
13511 | switch (comp_code) | |
13512 | { | |
13513 | case NE: return ARM_MI; | |
13514 | case EQ: return ARM_PL; | |
e6d29d15 | 13515 | default: gcc_unreachable (); |
defc0463 RE |
13516 | } |
13517 | ||
5165176d | 13518 | case CCFPEmode: |
e45b72c4 RE |
13519 | case CCFPmode: |
13520 | /* These encodings assume that AC=1 in the FPA system control | |
13521 | byte. This allows us to handle all cases except UNEQ and | |
13522 | LTGT. */ | |
84ed5e79 RE |
13523 | switch (comp_code) |
13524 | { | |
13525 | case GE: return ARM_GE; | |
13526 | case GT: return ARM_GT; | |
13527 | case LE: return ARM_LS; | |
13528 | case LT: return ARM_MI; | |
e45b72c4 RE |
13529 | case NE: return ARM_NE; |
13530 | case EQ: return ARM_EQ; | |
13531 | case ORDERED: return ARM_VC; | |
13532 | case UNORDERED: return ARM_VS; | |
13533 | case UNLT: return ARM_LT; | |
13534 | case UNLE: return ARM_LE; | |
13535 | case UNGT: return ARM_HI; | |
13536 | case UNGE: return ARM_PL; | |
13537 | /* UNEQ and LTGT do not have a representation. */ | |
13538 | case UNEQ: /* Fall through. */ | |
13539 | case LTGT: /* Fall through. */ | |
e6d29d15 | 13540 | default: gcc_unreachable (); |
84ed5e79 RE |
13541 | } |
13542 | ||
13543 | case CC_SWPmode: | |
13544 | switch (comp_code) | |
13545 | { | |
13546 | case NE: return ARM_NE; | |
13547 | case EQ: return ARM_EQ; | |
13548 | case GE: return ARM_LE; | |
13549 | case GT: return ARM_LT; | |
13550 | case LE: return ARM_GE; | |
13551 | case LT: return ARM_GT; | |
13552 | case GEU: return ARM_LS; | |
13553 | case GTU: return ARM_CC; | |
13554 | case LEU: return ARM_CS; | |
13555 | case LTU: return ARM_HI; | |
e6d29d15 | 13556 | default: gcc_unreachable (); |
84ed5e79 RE |
13557 | } |
13558 | ||
bd9c7e23 RE |
13559 | case CC_Cmode: |
13560 | switch (comp_code) | |
13561 | { | |
13562 | case LTU: return ARM_CS; | |
13563 | case GEU: return ARM_CC; | |
e6d29d15 | 13564 | default: gcc_unreachable (); |
bd9c7e23 | 13565 | } |
f676971a | 13566 | |
5165176d | 13567 | case CCmode: |
84ed5e79 | 13568 | switch (comp_code) |
5165176d | 13569 | { |
84ed5e79 RE |
13570 | case NE: return ARM_NE; |
13571 | case EQ: return ARM_EQ; | |
13572 | case GE: return ARM_GE; | |
13573 | case GT: return ARM_GT; | |
13574 | case LE: return ARM_LE; | |
13575 | case LT: return ARM_LT; | |
13576 | case GEU: return ARM_CS; | |
13577 | case GTU: return ARM_HI; | |
13578 | case LEU: return ARM_LS; | |
13579 | case LTU: return ARM_CC; | |
e6d29d15 | 13580 | default: gcc_unreachable (); |
5165176d RE |
13581 | } |
13582 | ||
e6d29d15 | 13583 | default: gcc_unreachable (); |
cce8749e | 13584 | } |
f3bb6135 | 13585 | } |
cce8749e | 13586 | |
44c7bd63 | 13587 | /* Tell arm_asm_output_opcode to output IT blocks for conditionally executed |
5b3e6663 PB |
13588 | instructions. */ |
13589 | void | |
13590 | thumb2_final_prescan_insn (rtx insn) | |
13591 | { | |
13592 | rtx first_insn = insn; | |
13593 | rtx body = PATTERN (insn); | |
13594 | rtx predicate; | |
13595 | enum arm_cond_code code; | |
13596 | int n; | |
13597 | int mask; | |
13598 | ||
13599 | /* Remove the previous insn from the count of insns to be output. */ | |
13600 | if (arm_condexec_count) | |
13601 | arm_condexec_count--; | |
13602 | ||
13603 | /* Nothing to do if we are already inside a conditional block. */ | |
13604 | if (arm_condexec_count) | |
13605 | return; | |
13606 | ||
13607 | if (GET_CODE (body) != COND_EXEC) | |
13608 | return; | |
13609 | ||
13610 | /* Conditional jumps are implemented directly. */ | |
13611 | if (GET_CODE (insn) == JUMP_INSN) | |
13612 | return; | |
13613 | ||
13614 | predicate = COND_EXEC_TEST (body); | |
13615 | arm_current_cc = get_arm_condition_code (predicate); | |
13616 | ||
13617 | n = get_attr_ce_count (insn); | |
13618 | arm_condexec_count = 1; | |
13619 | arm_condexec_mask = (1 << n) - 1; | |
13620 | arm_condexec_masklen = n; | |
13621 | /* See if subsequent instructions can be combined into the same block. */ | |
13622 | for (;;) | |
13623 | { | |
13624 | insn = next_nonnote_insn (insn); | |
13625 | ||
13626 | /* Jumping into the middle of an IT block is illegal, so a label or | |
13627 | barrier terminates the block. */ | |
13628 | if (GET_CODE (insn) != INSN && GET_CODE(insn) != JUMP_INSN) | |
13629 | break; | |
13630 | ||
13631 | body = PATTERN (insn); | |
13632 | /* USE and CLOBBER aren't really insns, so just skip them. */ | |
13633 | if (GET_CODE (body) == USE | |
13634 | || GET_CODE (body) == CLOBBER) | |
5b0202af | 13635 | continue; |
5b3e6663 | 13636 | |
7a085dce | 13637 | /* ??? Recognize conditional jumps, and combine them with IT blocks. */ |
5b3e6663 PB |
13638 | if (GET_CODE (body) != COND_EXEC) |
13639 | break; | |
13640 | /* Allow up to 4 conditionally executed instructions in a block. */ | |
13641 | n = get_attr_ce_count (insn); | |
13642 | if (arm_condexec_masklen + n > 4) | |
13643 | break; | |
13644 | ||
13645 | predicate = COND_EXEC_TEST (body); | |
13646 | code = get_arm_condition_code (predicate); | |
13647 | mask = (1 << n) - 1; | |
13648 | if (arm_current_cc == code) | |
13649 | arm_condexec_mask |= (mask << arm_condexec_masklen); | |
13650 | else if (arm_current_cc != ARM_INVERSE_CONDITION_CODE(code)) | |
13651 | break; | |
13652 | ||
13653 | arm_condexec_count++; | |
13654 | arm_condexec_masklen += n; | |
13655 | ||
13656 | /* A jump must be the last instruction in a conditional block. */ | |
13657 | if (GET_CODE(insn) == JUMP_INSN) | |
13658 | break; | |
13659 | } | |
13660 | /* Restore recog_data (getting the attributes of other insns can | |
13661 | destroy this array, but final.c assumes that it remains intact | |
13662 | across this call). */ | |
13663 | extract_constrain_insn_cached (first_insn); | |
13664 | } | |
13665 | ||
cce8749e | 13666 | void |
e32bac5b | 13667 | arm_final_prescan_insn (rtx insn) |
cce8749e CH |
13668 | { |
13669 | /* BODY will hold the body of INSN. */ | |
1d6e90ac | 13670 | rtx body = PATTERN (insn); |
cce8749e CH |
13671 | |
13672 | /* This will be 1 if trying to repeat the trick, and things need to be | |
13673 | reversed if it appears to fail. */ | |
13674 | int reverse = 0; | |
13675 | ||
ff9940b0 RE |
13676 | /* JUMP_CLOBBERS will be one implies that the conditions if a branch is |
13677 | taken are clobbered, even if the rtl suggests otherwise. It also | |
13678 | means that we have to grub around within the jump expression to find | |
13679 | out what the conditions are when the jump isn't taken. */ | |
13680 | int jump_clobbers = 0; | |
f676971a | 13681 | |
6354dc9b | 13682 | /* If we start with a return insn, we only succeed if we find another one. */ |
ff9940b0 | 13683 | int seeking_return = 0; |
f676971a | 13684 | |
cce8749e CH |
13685 | /* START_INSN will hold the insn from where we start looking. This is the |
13686 | first insn after the following code_label if REVERSE is true. */ | |
13687 | rtx start_insn = insn; | |
13688 | ||
13689 | /* If in state 4, check if the target branch is reached, in order to | |
13690 | change back to state 0. */ | |
13691 | if (arm_ccfsm_state == 4) | |
13692 | { | |
13693 | if (insn == arm_target_insn) | |
f5a1b0d2 NC |
13694 | { |
13695 | arm_target_insn = NULL; | |
13696 | arm_ccfsm_state = 0; | |
13697 | } | |
cce8749e CH |
13698 | return; |
13699 | } | |
13700 | ||
13701 | /* If in state 3, it is possible to repeat the trick, if this insn is an | |
13702 | unconditional branch to a label, and immediately following this branch | |
13703 | is the previous target label which is only used once, and the label this | |
13704 | branch jumps to is not too far off. */ | |
13705 | if (arm_ccfsm_state == 3) | |
13706 | { | |
13707 | if (simplejump_p (insn)) | |
13708 | { | |
13709 | start_insn = next_nonnote_insn (start_insn); | |
13710 | if (GET_CODE (start_insn) == BARRIER) | |
13711 | { | |
13712 | /* XXX Isn't this always a barrier? */ | |
13713 | start_insn = next_nonnote_insn (start_insn); | |
13714 | } | |
13715 | if (GET_CODE (start_insn) == CODE_LABEL | |
13716 | && CODE_LABEL_NUMBER (start_insn) == arm_target_label | |
13717 | && LABEL_NUSES (start_insn) == 1) | |
13718 | reverse = TRUE; | |
13719 | else | |
13720 | return; | |
13721 | } | |
ff9940b0 RE |
13722 | else if (GET_CODE (body) == RETURN) |
13723 | { | |
13724 | start_insn = next_nonnote_insn (start_insn); | |
13725 | if (GET_CODE (start_insn) == BARRIER) | |
13726 | start_insn = next_nonnote_insn (start_insn); | |
13727 | if (GET_CODE (start_insn) == CODE_LABEL | |
13728 | && CODE_LABEL_NUMBER (start_insn) == arm_target_label | |
13729 | && LABEL_NUSES (start_insn) == 1) | |
13730 | { | |
13731 | reverse = TRUE; | |
13732 | seeking_return = 1; | |
13733 | } | |
13734 | else | |
13735 | return; | |
13736 | } | |
cce8749e CH |
13737 | else |
13738 | return; | |
13739 | } | |
13740 | ||
e6d29d15 | 13741 | gcc_assert (!arm_ccfsm_state || reverse); |
cce8749e CH |
13742 | if (GET_CODE (insn) != JUMP_INSN) |
13743 | return; | |
13744 | ||
f676971a | 13745 | /* This jump might be paralleled with a clobber of the condition codes |
ff9940b0 RE |
13746 | the jump should always come first */ |
13747 | if (GET_CODE (body) == PARALLEL && XVECLEN (body, 0) > 0) | |
13748 | body = XVECEXP (body, 0, 0); | |
13749 | ||
cce8749e CH |
13750 | if (reverse |
13751 | || (GET_CODE (body) == SET && GET_CODE (SET_DEST (body)) == PC | |
13752 | && GET_CODE (SET_SRC (body)) == IF_THEN_ELSE)) | |
13753 | { | |
bd9c7e23 RE |
13754 | int insns_skipped; |
13755 | int fail = FALSE, succeed = FALSE; | |
cce8749e CH |
13756 | /* Flag which part of the IF_THEN_ELSE is the LABEL_REF. */ |
13757 | int then_not_else = TRUE; | |
ff9940b0 | 13758 | rtx this_insn = start_insn, label = 0; |
cce8749e | 13759 | |
f676971a | 13760 | /* If the jump cannot be done with one instruction, we cannot |
e45b72c4 | 13761 | conditionally execute the instruction in the inverse case. */ |
ff9940b0 | 13762 | if (get_attr_conds (insn) == CONDS_JUMP_CLOB) |
5bbe2d40 | 13763 | { |
5bbe2d40 RE |
13764 | jump_clobbers = 1; |
13765 | return; | |
13766 | } | |
f676971a | 13767 | |
cce8749e CH |
13768 | /* Register the insn jumped to. */ |
13769 | if (reverse) | |
ff9940b0 RE |
13770 | { |
13771 | if (!seeking_return) | |
13772 | label = XEXP (SET_SRC (body), 0); | |
13773 | } | |
cce8749e CH |
13774 | else if (GET_CODE (XEXP (SET_SRC (body), 1)) == LABEL_REF) |
13775 | label = XEXP (XEXP (SET_SRC (body), 1), 0); | |
13776 | else if (GET_CODE (XEXP (SET_SRC (body), 2)) == LABEL_REF) | |
13777 | { | |
13778 | label = XEXP (XEXP (SET_SRC (body), 2), 0); | |
13779 | then_not_else = FALSE; | |
13780 | } | |
ff9940b0 RE |
13781 | else if (GET_CODE (XEXP (SET_SRC (body), 1)) == RETURN) |
13782 | seeking_return = 1; | |
13783 | else if (GET_CODE (XEXP (SET_SRC (body), 2)) == RETURN) | |
13784 | { | |
13785 | seeking_return = 1; | |
13786 | then_not_else = FALSE; | |
13787 | } | |
cce8749e | 13788 | else |
e6d29d15 | 13789 | gcc_unreachable (); |
cce8749e CH |
13790 | |
13791 | /* See how many insns this branch skips, and what kind of insns. If all | |
13792 | insns are okay, and the label or unconditional branch to the same | |
13793 | label is not too far away, succeed. */ | |
13794 | for (insns_skipped = 0; | |
b36ba79f | 13795 | !fail && !succeed && insns_skipped++ < max_insns_skipped;) |
cce8749e CH |
13796 | { |
13797 | rtx scanbody; | |
13798 | ||
13799 | this_insn = next_nonnote_insn (this_insn); | |
13800 | if (!this_insn) | |
13801 | break; | |
13802 | ||
cce8749e CH |
13803 | switch (GET_CODE (this_insn)) |
13804 | { | |
13805 | case CODE_LABEL: | |
13806 | /* Succeed if it is the target label, otherwise fail since | |
13807 | control falls in from somewhere else. */ | |
13808 | if (this_insn == label) | |
13809 | { | |
ff9940b0 RE |
13810 | if (jump_clobbers) |
13811 | { | |
13812 | arm_ccfsm_state = 2; | |
13813 | this_insn = next_nonnote_insn (this_insn); | |
13814 | } | |
13815 | else | |
13816 | arm_ccfsm_state = 1; | |
cce8749e CH |
13817 | succeed = TRUE; |
13818 | } | |
13819 | else | |
13820 | fail = TRUE; | |
13821 | break; | |
13822 | ||
ff9940b0 | 13823 | case BARRIER: |
cce8749e | 13824 | /* Succeed if the following insn is the target label. |
f676971a EC |
13825 | Otherwise fail. |
13826 | If return insns are used then the last insn in a function | |
6354dc9b | 13827 | will be a barrier. */ |
cce8749e | 13828 | this_insn = next_nonnote_insn (this_insn); |
ff9940b0 | 13829 | if (this_insn && this_insn == label) |
cce8749e | 13830 | { |
ff9940b0 RE |
13831 | if (jump_clobbers) |
13832 | { | |
13833 | arm_ccfsm_state = 2; | |
13834 | this_insn = next_nonnote_insn (this_insn); | |
13835 | } | |
13836 | else | |
13837 | arm_ccfsm_state = 1; | |
cce8749e CH |
13838 | succeed = TRUE; |
13839 | } | |
13840 | else | |
13841 | fail = TRUE; | |
13842 | break; | |
13843 | ||
ff9940b0 | 13844 | case CALL_INSN: |
68d560d4 RE |
13845 | /* The AAPCS says that conditional calls should not be |
13846 | used since they make interworking inefficient (the | |
13847 | linker can't transform BL<cond> into BLX). That's | |
13848 | only a problem if the machine has BLX. */ | |
13849 | if (arm_arch5) | |
13850 | { | |
13851 | fail = TRUE; | |
13852 | break; | |
13853 | } | |
13854 | ||
61f0ccff RE |
13855 | /* Succeed if the following insn is the target label, or |
13856 | if the following two insns are a barrier and the | |
13857 | target label. */ | |
13858 | this_insn = next_nonnote_insn (this_insn); | |
13859 | if (this_insn && GET_CODE (this_insn) == BARRIER) | |
13860 | this_insn = next_nonnote_insn (this_insn); | |
bd9c7e23 | 13861 | |
61f0ccff RE |
13862 | if (this_insn && this_insn == label |
13863 | && insns_skipped < max_insns_skipped) | |
13864 | { | |
13865 | if (jump_clobbers) | |
bd9c7e23 | 13866 | { |
61f0ccff RE |
13867 | arm_ccfsm_state = 2; |
13868 | this_insn = next_nonnote_insn (this_insn); | |
bd9c7e23 RE |
13869 | } |
13870 | else | |
61f0ccff RE |
13871 | arm_ccfsm_state = 1; |
13872 | succeed = TRUE; | |
bd9c7e23 | 13873 | } |
61f0ccff RE |
13874 | else |
13875 | fail = TRUE; | |
ff9940b0 | 13876 | break; |
2b835d68 | 13877 | |
cce8749e CH |
13878 | case JUMP_INSN: |
13879 | /* If this is an unconditional branch to the same label, succeed. | |
13880 | If it is to another label, do nothing. If it is conditional, | |
13881 | fail. */ | |
e32bac5b RE |
13882 | /* XXX Probably, the tests for SET and the PC are |
13883 | unnecessary. */ | |
cce8749e | 13884 | |
ed4c4348 | 13885 | scanbody = PATTERN (this_insn); |
ff9940b0 RE |
13886 | if (GET_CODE (scanbody) == SET |
13887 | && GET_CODE (SET_DEST (scanbody)) == PC) | |
cce8749e CH |
13888 | { |
13889 | if (GET_CODE (SET_SRC (scanbody)) == LABEL_REF | |
13890 | && XEXP (SET_SRC (scanbody), 0) == label && !reverse) | |
13891 | { | |
13892 | arm_ccfsm_state = 2; | |
13893 | succeed = TRUE; | |
13894 | } | |
13895 | else if (GET_CODE (SET_SRC (scanbody)) == IF_THEN_ELSE) | |
13896 | fail = TRUE; | |
13897 | } | |
112cdef5 | 13898 | /* Fail if a conditional return is undesirable (e.g. on a |
b36ba79f RE |
13899 | StrongARM), but still allow this if optimizing for size. */ |
13900 | else if (GET_CODE (scanbody) == RETURN | |
a72d4945 | 13901 | && !use_return_insn (TRUE, NULL) |
5895f793 | 13902 | && !optimize_size) |
b36ba79f | 13903 | fail = TRUE; |
ff9940b0 RE |
13904 | else if (GET_CODE (scanbody) == RETURN |
13905 | && seeking_return) | |
13906 | { | |
13907 | arm_ccfsm_state = 2; | |
13908 | succeed = TRUE; | |
13909 | } | |
13910 | else if (GET_CODE (scanbody) == PARALLEL) | |
13911 | { | |
13912 | switch (get_attr_conds (this_insn)) | |
13913 | { | |
13914 | case CONDS_NOCOND: | |
13915 | break; | |
13916 | default: | |
13917 | fail = TRUE; | |
13918 | break; | |
13919 | } | |
13920 | } | |
4e67550b | 13921 | else |
112cdef5 | 13922 | fail = TRUE; /* Unrecognized jump (e.g. epilogue). */ |
4e67550b | 13923 | |
cce8749e CH |
13924 | break; |
13925 | ||
13926 | case INSN: | |
ff9940b0 RE |
13927 | /* Instructions using or affecting the condition codes make it |
13928 | fail. */ | |
ed4c4348 | 13929 | scanbody = PATTERN (this_insn); |
5895f793 RE |
13930 | if (!(GET_CODE (scanbody) == SET |
13931 | || GET_CODE (scanbody) == PARALLEL) | |
74641843 | 13932 | || get_attr_conds (this_insn) != CONDS_NOCOND) |
cce8749e | 13933 | fail = TRUE; |
9b6b54e2 NC |
13934 | |
13935 | /* A conditional cirrus instruction must be followed by | |
13936 | a non Cirrus instruction. However, since we | |
13937 | conditionalize instructions in this function and by | |
13938 | the time we get here we can't add instructions | |
13939 | (nops), because shorten_branches() has already been | |
13940 | called, we will disable conditionalizing Cirrus | |
13941 | instructions to be safe. */ | |
13942 | if (GET_CODE (scanbody) != USE | |
13943 | && GET_CODE (scanbody) != CLOBBER | |
f0375c66 | 13944 | && get_attr_cirrus (this_insn) != CIRRUS_NOT) |
9b6b54e2 | 13945 | fail = TRUE; |
cce8749e CH |
13946 | break; |
13947 | ||
13948 | default: | |
13949 | break; | |
13950 | } | |
13951 | } | |
13952 | if (succeed) | |
13953 | { | |
ff9940b0 | 13954 | if ((!seeking_return) && (arm_ccfsm_state == 1 || reverse)) |
cce8749e | 13955 | arm_target_label = CODE_LABEL_NUMBER (label); |
e6d29d15 | 13956 | else |
ff9940b0 | 13957 | { |
e6d29d15 | 13958 | gcc_assert (seeking_return || arm_ccfsm_state == 2); |
e0b92319 | 13959 | |
ff9940b0 RE |
13960 | while (this_insn && GET_CODE (PATTERN (this_insn)) == USE) |
13961 | { | |
13962 | this_insn = next_nonnote_insn (this_insn); | |
e6d29d15 NS |
13963 | gcc_assert (!this_insn |
13964 | || (GET_CODE (this_insn) != BARRIER | |
13965 | && GET_CODE (this_insn) != CODE_LABEL)); | |
ff9940b0 RE |
13966 | } |
13967 | if (!this_insn) | |
13968 | { | |
d6b4baa4 | 13969 | /* Oh, dear! we ran off the end.. give up. */ |
5b3e6663 | 13970 | extract_constrain_insn_cached (insn); |
ff9940b0 | 13971 | arm_ccfsm_state = 0; |
abaa26e5 | 13972 | arm_target_insn = NULL; |
ff9940b0 RE |
13973 | return; |
13974 | } | |
13975 | arm_target_insn = this_insn; | |
13976 | } | |
ff9940b0 RE |
13977 | if (jump_clobbers) |
13978 | { | |
e6d29d15 | 13979 | gcc_assert (!reverse); |
f676971a | 13980 | arm_current_cc = |
ff9940b0 RE |
13981 | get_arm_condition_code (XEXP (XEXP (XEXP (SET_SRC (body), |
13982 | 0), 0), 1)); | |
13983 | if (GET_CODE (XEXP (XEXP (SET_SRC (body), 0), 0)) == AND) | |
13984 | arm_current_cc = ARM_INVERSE_CONDITION_CODE (arm_current_cc); | |
13985 | if (GET_CODE (XEXP (SET_SRC (body), 0)) == NE) | |
13986 | arm_current_cc = ARM_INVERSE_CONDITION_CODE (arm_current_cc); | |
13987 | } | |
13988 | else | |
13989 | { | |
13990 | /* If REVERSE is true, ARM_CURRENT_CC needs to be inverted from | |
13991 | what it was. */ | |
13992 | if (!reverse) | |
13993 | arm_current_cc = get_arm_condition_code (XEXP (SET_SRC (body), | |
13994 | 0)); | |
13995 | } | |
cce8749e | 13996 | |
cce8749e CH |
13997 | if (reverse || then_not_else) |
13998 | arm_current_cc = ARM_INVERSE_CONDITION_CODE (arm_current_cc); | |
13999 | } | |
f676971a | 14000 | |
1ccbefce | 14001 | /* Restore recog_data (getting the attributes of other insns can |
ff9940b0 | 14002 | destroy this array, but final.c assumes that it remains intact |
5b3e6663 PB |
14003 | across this call. */ |
14004 | extract_constrain_insn_cached (insn); | |
14005 | } | |
14006 | } | |
14007 | ||
14008 | /* Output IT instructions. */ | |
14009 | void | |
14010 | thumb2_asm_output_opcode (FILE * stream) | |
14011 | { | |
14012 | char buff[5]; | |
14013 | int n; | |
14014 | ||
14015 | if (arm_condexec_mask) | |
14016 | { | |
14017 | for (n = 0; n < arm_condexec_masklen; n++) | |
14018 | buff[n] = (arm_condexec_mask & (1 << n)) ? 't' : 'e'; | |
14019 | buff[n] = 0; | |
14020 | asm_fprintf(stream, "i%s\t%s\n\t", buff, | |
14021 | arm_condition_codes[arm_current_cc]); | |
14022 | arm_condexec_mask = 0; | |
cce8749e | 14023 | } |
f3bb6135 | 14024 | } |
cce8749e | 14025 | |
4b02997f | 14026 | /* Returns true if REGNO is a valid register |
21b5653c | 14027 | for holding a quantity of type MODE. */ |
4b02997f | 14028 | int |
e32bac5b | 14029 | arm_hard_regno_mode_ok (unsigned int regno, enum machine_mode mode) |
4b02997f NC |
14030 | { |
14031 | if (GET_MODE_CLASS (mode) == MODE_CC) | |
a6a5de04 RE |
14032 | return (regno == CC_REGNUM |
14033 | || (TARGET_HARD_FLOAT && TARGET_VFP | |
14034 | && regno == VFPCC_REGNUM)); | |
f676971a | 14035 | |
5b3e6663 | 14036 | if (TARGET_THUMB1) |
4b02997f NC |
14037 | /* For the Thumb we only allow values bigger than SImode in |
14038 | registers 0 - 6, so that there is always a second low | |
14039 | register available to hold the upper part of the value. | |
14040 | We probably we ought to ensure that the register is the | |
14041 | start of an even numbered register pair. */ | |
e9d7b180 | 14042 | return (ARM_NUM_REGS (mode) < 2) || (regno < LAST_LO_REGNUM); |
4b02997f | 14043 | |
a6a5de04 RE |
14044 | if (TARGET_HARD_FLOAT && TARGET_MAVERICK |
14045 | && IS_CIRRUS_REGNUM (regno)) | |
9b6b54e2 NC |
14046 | /* We have outlawed SI values in Cirrus registers because they |
14047 | reside in the lower 32 bits, but SF values reside in the | |
14048 | upper 32 bits. This causes gcc all sorts of grief. We can't | |
14049 | even split the registers into pairs because Cirrus SI values | |
14050 | get sign extended to 64bits-- aldyh. */ | |
14051 | return (GET_MODE_CLASS (mode) == MODE_FLOAT) || (mode == DImode); | |
14052 | ||
a6a5de04 RE |
14053 | if (TARGET_HARD_FLOAT && TARGET_VFP |
14054 | && IS_VFP_REGNUM (regno)) | |
9b66ebb1 PB |
14055 | { |
14056 | if (mode == SFmode || mode == SImode) | |
f1adb0a9 | 14057 | return VFP_REGNO_OK_FOR_SINGLE (regno); |
9b66ebb1 | 14058 | |
9b66ebb1 | 14059 | if (mode == DFmode) |
f1adb0a9 | 14060 | return VFP_REGNO_OK_FOR_DOUBLE (regno); |
88f77cba JB |
14061 | |
14062 | if (TARGET_NEON) | |
14063 | return (VALID_NEON_DREG_MODE (mode) && VFP_REGNO_OK_FOR_DOUBLE (regno)) | |
14064 | || (VALID_NEON_QREG_MODE (mode) | |
14065 | && NEON_REGNO_OK_FOR_QUAD (regno)) | |
14066 | || (mode == TImode && NEON_REGNO_OK_FOR_NREGS (regno, 2)) | |
14067 | || (mode == EImode && NEON_REGNO_OK_FOR_NREGS (regno, 3)) | |
14068 | || (mode == OImode && NEON_REGNO_OK_FOR_NREGS (regno, 4)) | |
14069 | || (mode == CImode && NEON_REGNO_OK_FOR_NREGS (regno, 6)) | |
14070 | || (mode == XImode && NEON_REGNO_OK_FOR_NREGS (regno, 8)); | |
14071 | ||
9b66ebb1 PB |
14072 | return FALSE; |
14073 | } | |
14074 | ||
a6a5de04 RE |
14075 | if (TARGET_REALLY_IWMMXT) |
14076 | { | |
14077 | if (IS_IWMMXT_GR_REGNUM (regno)) | |
14078 | return mode == SImode; | |
5a9335ef | 14079 | |
a6a5de04 RE |
14080 | if (IS_IWMMXT_REGNUM (regno)) |
14081 | return VALID_IWMMXT_REG_MODE (mode); | |
14082 | } | |
14083 | ||
fdd695fd PB |
14084 | /* We allow any value to be stored in the general registers. |
14085 | Restrict doubleword quantities to even register pairs so that we can | |
88f77cba JB |
14086 | use ldrd. Do not allow Neon structure opaque modes in general registers; |
14087 | they would use too many. */ | |
4b02997f | 14088 | if (regno <= LAST_ARM_REGNUM) |
88f77cba JB |
14089 | return !(TARGET_LDRD && GET_MODE_SIZE (mode) > 4 && (regno & 1) != 0) |
14090 | && !VALID_NEON_STRUCT_MODE (mode); | |
4b02997f | 14091 | |
a6a5de04 | 14092 | if (regno == FRAME_POINTER_REGNUM |
4b02997f NC |
14093 | || regno == ARG_POINTER_REGNUM) |
14094 | /* We only allow integers in the fake hard registers. */ | |
14095 | return GET_MODE_CLASS (mode) == MODE_INT; | |
14096 | ||
3b684012 | 14097 | /* The only registers left are the FPA registers |
4b02997f | 14098 | which we only allow to hold FP values. */ |
a6a5de04 RE |
14099 | return (TARGET_HARD_FLOAT && TARGET_FPA |
14100 | && GET_MODE_CLASS (mode) == MODE_FLOAT | |
14101 | && regno >= FIRST_FPA_REGNUM | |
14102 | && regno <= LAST_FPA_REGNUM); | |
4b02997f NC |
14103 | } |
14104 | ||
5b3e6663 PB |
14105 | /* For efficiency and historical reasons LO_REGS, HI_REGS and CC_REGS are |
14106 | not used in arm mode. */ | |
d5b7b3ae | 14107 | int |
e32bac5b | 14108 | arm_regno_class (int regno) |
d5b7b3ae | 14109 | { |
5b3e6663 | 14110 | if (TARGET_THUMB1) |
d5b7b3ae RE |
14111 | { |
14112 | if (regno == STACK_POINTER_REGNUM) | |
14113 | return STACK_REG; | |
14114 | if (regno == CC_REGNUM) | |
14115 | return CC_REG; | |
14116 | if (regno < 8) | |
14117 | return LO_REGS; | |
14118 | return HI_REGS; | |
14119 | } | |
14120 | ||
5b3e6663 PB |
14121 | if (TARGET_THUMB2 && regno < 8) |
14122 | return LO_REGS; | |
14123 | ||
d5b7b3ae RE |
14124 | if ( regno <= LAST_ARM_REGNUM |
14125 | || regno == FRAME_POINTER_REGNUM | |
14126 | || regno == ARG_POINTER_REGNUM) | |
5b3e6663 | 14127 | return TARGET_THUMB2 ? HI_REGS : GENERAL_REGS; |
f676971a | 14128 | |
9b66ebb1 | 14129 | if (regno == CC_REGNUM || regno == VFPCC_REGNUM) |
5b3e6663 | 14130 | return TARGET_THUMB2 ? CC_REG : NO_REGS; |
d5b7b3ae | 14131 | |
9b6b54e2 NC |
14132 | if (IS_CIRRUS_REGNUM (regno)) |
14133 | return CIRRUS_REGS; | |
14134 | ||
9b66ebb1 | 14135 | if (IS_VFP_REGNUM (regno)) |
f1adb0a9 JB |
14136 | { |
14137 | if (regno <= D7_VFP_REGNUM) | |
14138 | return VFP_D0_D7_REGS; | |
14139 | else if (regno <= LAST_LO_VFP_REGNUM) | |
14140 | return VFP_LO_REGS; | |
14141 | else | |
14142 | return VFP_HI_REGS; | |
14143 | } | |
9b66ebb1 | 14144 | |
5a9335ef NC |
14145 | if (IS_IWMMXT_REGNUM (regno)) |
14146 | return IWMMXT_REGS; | |
14147 | ||
e99faaaa ILT |
14148 | if (IS_IWMMXT_GR_REGNUM (regno)) |
14149 | return IWMMXT_GR_REGS; | |
14150 | ||
3b684012 | 14151 | return FPA_REGS; |
d5b7b3ae RE |
14152 | } |
14153 | ||
14154 | /* Handle a special case when computing the offset | |
14155 | of an argument from the frame pointer. */ | |
14156 | int | |
e32bac5b | 14157 | arm_debugger_arg_offset (int value, rtx addr) |
d5b7b3ae RE |
14158 | { |
14159 | rtx insn; | |
14160 | ||
14161 | /* We are only interested if dbxout_parms() failed to compute the offset. */ | |
14162 | if (value != 0) | |
14163 | return 0; | |
14164 | ||
14165 | /* We can only cope with the case where the address is held in a register. */ | |
14166 | if (GET_CODE (addr) != REG) | |
14167 | return 0; | |
14168 | ||
14169 | /* If we are using the frame pointer to point at the argument, then | |
14170 | an offset of 0 is correct. */ | |
cd2b33d0 | 14171 | if (REGNO (addr) == (unsigned) HARD_FRAME_POINTER_REGNUM) |
d5b7b3ae | 14172 | return 0; |
f676971a | 14173 | |
d5b7b3ae RE |
14174 | /* If we are using the stack pointer to point at the |
14175 | argument, then an offset of 0 is correct. */ | |
5b3e6663 | 14176 | /* ??? Check this is consistent with thumb2 frame layout. */ |
5895f793 | 14177 | if ((TARGET_THUMB || !frame_pointer_needed) |
d5b7b3ae RE |
14178 | && REGNO (addr) == SP_REGNUM) |
14179 | return 0; | |
f676971a | 14180 | |
d5b7b3ae RE |
14181 | /* Oh dear. The argument is pointed to by a register rather |
14182 | than being held in a register, or being stored at a known | |
14183 | offset from the frame pointer. Since GDB only understands | |
14184 | those two kinds of argument we must translate the address | |
14185 | held in the register into an offset from the frame pointer. | |
14186 | We do this by searching through the insns for the function | |
14187 | looking to see where this register gets its value. If the | |
4912a07c | 14188 | register is initialized from the frame pointer plus an offset |
d5b7b3ae | 14189 | then we are in luck and we can continue, otherwise we give up. |
f676971a | 14190 | |
d5b7b3ae RE |
14191 | This code is exercised by producing debugging information |
14192 | for a function with arguments like this: | |
f676971a | 14193 | |
d5b7b3ae | 14194 | double func (double a, double b, int c, double d) {return d;} |
f676971a | 14195 | |
d5b7b3ae RE |
14196 | Without this code the stab for parameter 'd' will be set to |
14197 | an offset of 0 from the frame pointer, rather than 8. */ | |
14198 | ||
14199 | /* The if() statement says: | |
14200 | ||
14201 | If the insn is a normal instruction | |
14202 | and if the insn is setting the value in a register | |
14203 | and if the register being set is the register holding the address of the argument | |
14204 | and if the address is computing by an addition | |
14205 | that involves adding to a register | |
14206 | which is the frame pointer | |
14207 | a constant integer | |
14208 | ||
d6b4baa4 | 14209 | then... */ |
f676971a | 14210 | |
d5b7b3ae RE |
14211 | for (insn = get_insns (); insn; insn = NEXT_INSN (insn)) |
14212 | { | |
f676971a | 14213 | if ( GET_CODE (insn) == INSN |
d5b7b3ae RE |
14214 | && GET_CODE (PATTERN (insn)) == SET |
14215 | && REGNO (XEXP (PATTERN (insn), 0)) == REGNO (addr) | |
14216 | && GET_CODE (XEXP (PATTERN (insn), 1)) == PLUS | |
14217 | && GET_CODE (XEXP (XEXP (PATTERN (insn), 1), 0)) == REG | |
cd2b33d0 | 14218 | && REGNO (XEXP (XEXP (PATTERN (insn), 1), 0)) == (unsigned) HARD_FRAME_POINTER_REGNUM |
d5b7b3ae RE |
14219 | && GET_CODE (XEXP (XEXP (PATTERN (insn), 1), 1)) == CONST_INT |
14220 | ) | |
14221 | { | |
14222 | value = INTVAL (XEXP (XEXP (PATTERN (insn), 1), 1)); | |
f676971a | 14223 | |
d5b7b3ae RE |
14224 | break; |
14225 | } | |
14226 | } | |
f676971a | 14227 | |
d5b7b3ae RE |
14228 | if (value == 0) |
14229 | { | |
14230 | debug_rtx (addr); | |
d4ee4d25 | 14231 | warning (0, "unable to compute real location of stacked parameter"); |
d5b7b3ae RE |
14232 | value = 8; /* XXX magic hack */ |
14233 | } | |
14234 | ||
14235 | return value; | |
14236 | } | |
d5b7b3ae | 14237 | \f |
5a9335ef NC |
14238 | #define def_mbuiltin(MASK, NAME, TYPE, CODE) \ |
14239 | do \ | |
14240 | { \ | |
14241 | if ((MASK) & insn_flags) \ | |
c79efc4d RÁE |
14242 | add_builtin_function ((NAME), (TYPE), (CODE), \ |
14243 | BUILT_IN_MD, NULL, NULL_TREE); \ | |
5a9335ef NC |
14244 | } \ |
14245 | while (0) | |
14246 | ||
14247 | struct builtin_description | |
14248 | { | |
14249 | const unsigned int mask; | |
14250 | const enum insn_code icode; | |
14251 | const char * const name; | |
14252 | const enum arm_builtins code; | |
14253 | const enum rtx_code comparison; | |
14254 | const unsigned int flag; | |
14255 | }; | |
14256 | ||
14257 | static const struct builtin_description bdesc_2arg[] = | |
14258 | { | |
14259 | #define IWMMXT_BUILTIN(code, string, builtin) \ | |
14260 | { FL_IWMMXT, CODE_FOR_##code, "__builtin_arm_" string, \ | |
14261 | ARM_BUILTIN_##builtin, 0, 0 }, | |
14262 | ||
14263 | IWMMXT_BUILTIN (addv8qi3, "waddb", WADDB) | |
14264 | IWMMXT_BUILTIN (addv4hi3, "waddh", WADDH) | |
14265 | IWMMXT_BUILTIN (addv2si3, "waddw", WADDW) | |
14266 | IWMMXT_BUILTIN (subv8qi3, "wsubb", WSUBB) | |
14267 | IWMMXT_BUILTIN (subv4hi3, "wsubh", WSUBH) | |
14268 | IWMMXT_BUILTIN (subv2si3, "wsubw", WSUBW) | |
14269 | IWMMXT_BUILTIN (ssaddv8qi3, "waddbss", WADDSSB) | |
14270 | IWMMXT_BUILTIN (ssaddv4hi3, "waddhss", WADDSSH) | |
14271 | IWMMXT_BUILTIN (ssaddv2si3, "waddwss", WADDSSW) | |
14272 | IWMMXT_BUILTIN (sssubv8qi3, "wsubbss", WSUBSSB) | |
14273 | IWMMXT_BUILTIN (sssubv4hi3, "wsubhss", WSUBSSH) | |
14274 | IWMMXT_BUILTIN (sssubv2si3, "wsubwss", WSUBSSW) | |
14275 | IWMMXT_BUILTIN (usaddv8qi3, "waddbus", WADDUSB) | |
14276 | IWMMXT_BUILTIN (usaddv4hi3, "waddhus", WADDUSH) | |
14277 | IWMMXT_BUILTIN (usaddv2si3, "waddwus", WADDUSW) | |
14278 | IWMMXT_BUILTIN (ussubv8qi3, "wsubbus", WSUBUSB) | |
14279 | IWMMXT_BUILTIN (ussubv4hi3, "wsubhus", WSUBUSH) | |
14280 | IWMMXT_BUILTIN (ussubv2si3, "wsubwus", WSUBUSW) | |
14281 | IWMMXT_BUILTIN (mulv4hi3, "wmulul", WMULUL) | |
f07a6b21 BE |
14282 | IWMMXT_BUILTIN (smulv4hi3_highpart, "wmulsm", WMULSM) |
14283 | IWMMXT_BUILTIN (umulv4hi3_highpart, "wmulum", WMULUM) | |
5a9335ef NC |
14284 | IWMMXT_BUILTIN (eqv8qi3, "wcmpeqb", WCMPEQB) |
14285 | IWMMXT_BUILTIN (eqv4hi3, "wcmpeqh", WCMPEQH) | |
14286 | IWMMXT_BUILTIN (eqv2si3, "wcmpeqw", WCMPEQW) | |
14287 | IWMMXT_BUILTIN (gtuv8qi3, "wcmpgtub", WCMPGTUB) | |
14288 | IWMMXT_BUILTIN (gtuv4hi3, "wcmpgtuh", WCMPGTUH) | |
14289 | IWMMXT_BUILTIN (gtuv2si3, "wcmpgtuw", WCMPGTUW) | |
14290 | IWMMXT_BUILTIN (gtv8qi3, "wcmpgtsb", WCMPGTSB) | |
14291 | IWMMXT_BUILTIN (gtv4hi3, "wcmpgtsh", WCMPGTSH) | |
14292 | IWMMXT_BUILTIN (gtv2si3, "wcmpgtsw", WCMPGTSW) | |
14293 | IWMMXT_BUILTIN (umaxv8qi3, "wmaxub", WMAXUB) | |
14294 | IWMMXT_BUILTIN (smaxv8qi3, "wmaxsb", WMAXSB) | |
14295 | IWMMXT_BUILTIN (umaxv4hi3, "wmaxuh", WMAXUH) | |
14296 | IWMMXT_BUILTIN (smaxv4hi3, "wmaxsh", WMAXSH) | |
14297 | IWMMXT_BUILTIN (umaxv2si3, "wmaxuw", WMAXUW) | |
14298 | IWMMXT_BUILTIN (smaxv2si3, "wmaxsw", WMAXSW) | |
14299 | IWMMXT_BUILTIN (uminv8qi3, "wminub", WMINUB) | |
14300 | IWMMXT_BUILTIN (sminv8qi3, "wminsb", WMINSB) | |
14301 | IWMMXT_BUILTIN (uminv4hi3, "wminuh", WMINUH) | |
14302 | IWMMXT_BUILTIN (sminv4hi3, "wminsh", WMINSH) | |
14303 | IWMMXT_BUILTIN (uminv2si3, "wminuw", WMINUW) | |
14304 | IWMMXT_BUILTIN (sminv2si3, "wminsw", WMINSW) | |
14305 | IWMMXT_BUILTIN (iwmmxt_anddi3, "wand", WAND) | |
14306 | IWMMXT_BUILTIN (iwmmxt_nanddi3, "wandn", WANDN) | |
14307 | IWMMXT_BUILTIN (iwmmxt_iordi3, "wor", WOR) | |
14308 | IWMMXT_BUILTIN (iwmmxt_xordi3, "wxor", WXOR) | |
14309 | IWMMXT_BUILTIN (iwmmxt_uavgv8qi3, "wavg2b", WAVG2B) | |
14310 | IWMMXT_BUILTIN (iwmmxt_uavgv4hi3, "wavg2h", WAVG2H) | |
14311 | IWMMXT_BUILTIN (iwmmxt_uavgrndv8qi3, "wavg2br", WAVG2BR) | |
14312 | IWMMXT_BUILTIN (iwmmxt_uavgrndv4hi3, "wavg2hr", WAVG2HR) | |
14313 | IWMMXT_BUILTIN (iwmmxt_wunpckilb, "wunpckilb", WUNPCKILB) | |
14314 | IWMMXT_BUILTIN (iwmmxt_wunpckilh, "wunpckilh", WUNPCKILH) | |
14315 | IWMMXT_BUILTIN (iwmmxt_wunpckilw, "wunpckilw", WUNPCKILW) | |
14316 | IWMMXT_BUILTIN (iwmmxt_wunpckihb, "wunpckihb", WUNPCKIHB) | |
14317 | IWMMXT_BUILTIN (iwmmxt_wunpckihh, "wunpckihh", WUNPCKIHH) | |
14318 | IWMMXT_BUILTIN (iwmmxt_wunpckihw, "wunpckihw", WUNPCKIHW) | |
14319 | IWMMXT_BUILTIN (iwmmxt_wmadds, "wmadds", WMADDS) | |
14320 | IWMMXT_BUILTIN (iwmmxt_wmaddu, "wmaddu", WMADDU) | |
14321 | ||
14322 | #define IWMMXT_BUILTIN2(code, builtin) \ | |
14323 | { FL_IWMMXT, CODE_FOR_##code, NULL, ARM_BUILTIN_##builtin, 0, 0 }, | |
f676971a | 14324 | |
5a9335ef NC |
14325 | IWMMXT_BUILTIN2 (iwmmxt_wpackhss, WPACKHSS) |
14326 | IWMMXT_BUILTIN2 (iwmmxt_wpackwss, WPACKWSS) | |
14327 | IWMMXT_BUILTIN2 (iwmmxt_wpackdss, WPACKDSS) | |
14328 | IWMMXT_BUILTIN2 (iwmmxt_wpackhus, WPACKHUS) | |
14329 | IWMMXT_BUILTIN2 (iwmmxt_wpackwus, WPACKWUS) | |
14330 | IWMMXT_BUILTIN2 (iwmmxt_wpackdus, WPACKDUS) | |
14331 | IWMMXT_BUILTIN2 (ashlv4hi3_di, WSLLH) | |
88f77cba | 14332 | IWMMXT_BUILTIN2 (ashlv4hi3_iwmmxt, WSLLHI) |
5a9335ef | 14333 | IWMMXT_BUILTIN2 (ashlv2si3_di, WSLLW) |
88f77cba | 14334 | IWMMXT_BUILTIN2 (ashlv2si3_iwmmxt, WSLLWI) |
5a9335ef NC |
14335 | IWMMXT_BUILTIN2 (ashldi3_di, WSLLD) |
14336 | IWMMXT_BUILTIN2 (ashldi3_iwmmxt, WSLLDI) | |
14337 | IWMMXT_BUILTIN2 (lshrv4hi3_di, WSRLH) | |
88f77cba | 14338 | IWMMXT_BUILTIN2 (lshrv4hi3_iwmmxt, WSRLHI) |
5a9335ef | 14339 | IWMMXT_BUILTIN2 (lshrv2si3_di, WSRLW) |
88f77cba | 14340 | IWMMXT_BUILTIN2 (lshrv2si3_iwmmxt, WSRLWI) |
5a9335ef | 14341 | IWMMXT_BUILTIN2 (lshrdi3_di, WSRLD) |
9b66ebb1 | 14342 | IWMMXT_BUILTIN2 (lshrdi3_iwmmxt, WSRLDI) |
5a9335ef | 14343 | IWMMXT_BUILTIN2 (ashrv4hi3_di, WSRAH) |
88f77cba | 14344 | IWMMXT_BUILTIN2 (ashrv4hi3_iwmmxt, WSRAHI) |
5a9335ef | 14345 | IWMMXT_BUILTIN2 (ashrv2si3_di, WSRAW) |
88f77cba | 14346 | IWMMXT_BUILTIN2 (ashrv2si3_iwmmxt, WSRAWI) |
5a9335ef | 14347 | IWMMXT_BUILTIN2 (ashrdi3_di, WSRAD) |
9b66ebb1 | 14348 | IWMMXT_BUILTIN2 (ashrdi3_iwmmxt, WSRADI) |
5a9335ef NC |
14349 | IWMMXT_BUILTIN2 (rorv4hi3_di, WRORH) |
14350 | IWMMXT_BUILTIN2 (rorv4hi3, WRORHI) | |
14351 | IWMMXT_BUILTIN2 (rorv2si3_di, WRORW) | |
14352 | IWMMXT_BUILTIN2 (rorv2si3, WRORWI) | |
14353 | IWMMXT_BUILTIN2 (rordi3_di, WRORD) | |
14354 | IWMMXT_BUILTIN2 (rordi3, WRORDI) | |
14355 | IWMMXT_BUILTIN2 (iwmmxt_wmacuz, WMACUZ) | |
14356 | IWMMXT_BUILTIN2 (iwmmxt_wmacsz, WMACSZ) | |
14357 | }; | |
14358 | ||
14359 | static const struct builtin_description bdesc_1arg[] = | |
14360 | { | |
14361 | IWMMXT_BUILTIN (iwmmxt_tmovmskb, "tmovmskb", TMOVMSKB) | |
14362 | IWMMXT_BUILTIN (iwmmxt_tmovmskh, "tmovmskh", TMOVMSKH) | |
14363 | IWMMXT_BUILTIN (iwmmxt_tmovmskw, "tmovmskw", TMOVMSKW) | |
14364 | IWMMXT_BUILTIN (iwmmxt_waccb, "waccb", WACCB) | |
14365 | IWMMXT_BUILTIN (iwmmxt_wacch, "wacch", WACCH) | |
14366 | IWMMXT_BUILTIN (iwmmxt_waccw, "waccw", WACCW) | |
14367 | IWMMXT_BUILTIN (iwmmxt_wunpckehub, "wunpckehub", WUNPCKEHUB) | |
14368 | IWMMXT_BUILTIN (iwmmxt_wunpckehuh, "wunpckehuh", WUNPCKEHUH) | |
14369 | IWMMXT_BUILTIN (iwmmxt_wunpckehuw, "wunpckehuw", WUNPCKEHUW) | |
14370 | IWMMXT_BUILTIN (iwmmxt_wunpckehsb, "wunpckehsb", WUNPCKEHSB) | |
14371 | IWMMXT_BUILTIN (iwmmxt_wunpckehsh, "wunpckehsh", WUNPCKEHSH) | |
14372 | IWMMXT_BUILTIN (iwmmxt_wunpckehsw, "wunpckehsw", WUNPCKEHSW) | |
14373 | IWMMXT_BUILTIN (iwmmxt_wunpckelub, "wunpckelub", WUNPCKELUB) | |
14374 | IWMMXT_BUILTIN (iwmmxt_wunpckeluh, "wunpckeluh", WUNPCKELUH) | |
14375 | IWMMXT_BUILTIN (iwmmxt_wunpckeluw, "wunpckeluw", WUNPCKELUW) | |
14376 | IWMMXT_BUILTIN (iwmmxt_wunpckelsb, "wunpckelsb", WUNPCKELSB) | |
14377 | IWMMXT_BUILTIN (iwmmxt_wunpckelsh, "wunpckelsh", WUNPCKELSH) | |
14378 | IWMMXT_BUILTIN (iwmmxt_wunpckelsw, "wunpckelsw", WUNPCKELSW) | |
14379 | }; | |
14380 | ||
14381 | /* Set up all the iWMMXt builtins. This is | |
14382 | not called if TARGET_IWMMXT is zero. */ | |
14383 | ||
14384 | static void | |
14385 | arm_init_iwmmxt_builtins (void) | |
14386 | { | |
14387 | const struct builtin_description * d; | |
14388 | size_t i; | |
14389 | tree endlink = void_list_node; | |
14390 | ||
4a5eab38 PB |
14391 | tree V2SI_type_node = build_vector_type_for_mode (intSI_type_node, V2SImode); |
14392 | tree V4HI_type_node = build_vector_type_for_mode (intHI_type_node, V4HImode); | |
14393 | tree V8QI_type_node = build_vector_type_for_mode (intQI_type_node, V8QImode); | |
14394 | ||
5a9335ef NC |
14395 | tree int_ftype_int |
14396 | = build_function_type (integer_type_node, | |
14397 | tree_cons (NULL_TREE, integer_type_node, endlink)); | |
14398 | tree v8qi_ftype_v8qi_v8qi_int | |
14399 | = build_function_type (V8QI_type_node, | |
14400 | tree_cons (NULL_TREE, V8QI_type_node, | |
14401 | tree_cons (NULL_TREE, V8QI_type_node, | |
14402 | tree_cons (NULL_TREE, | |
14403 | integer_type_node, | |
14404 | endlink)))); | |
14405 | tree v4hi_ftype_v4hi_int | |
14406 | = build_function_type (V4HI_type_node, | |
14407 | tree_cons (NULL_TREE, V4HI_type_node, | |
14408 | tree_cons (NULL_TREE, integer_type_node, | |
14409 | endlink))); | |
14410 | tree v2si_ftype_v2si_int | |
14411 | = build_function_type (V2SI_type_node, | |
14412 | tree_cons (NULL_TREE, V2SI_type_node, | |
14413 | tree_cons (NULL_TREE, integer_type_node, | |
14414 | endlink))); | |
14415 | tree v2si_ftype_di_di | |
14416 | = build_function_type (V2SI_type_node, | |
14417 | tree_cons (NULL_TREE, long_long_integer_type_node, | |
14418 | tree_cons (NULL_TREE, long_long_integer_type_node, | |
14419 | endlink))); | |
14420 | tree di_ftype_di_int | |
14421 | = build_function_type (long_long_integer_type_node, | |
14422 | tree_cons (NULL_TREE, long_long_integer_type_node, | |
14423 | tree_cons (NULL_TREE, integer_type_node, | |
14424 | endlink))); | |
14425 | tree di_ftype_di_int_int | |
14426 | = build_function_type (long_long_integer_type_node, | |
14427 | tree_cons (NULL_TREE, long_long_integer_type_node, | |
14428 | tree_cons (NULL_TREE, integer_type_node, | |
14429 | tree_cons (NULL_TREE, | |
14430 | integer_type_node, | |
14431 | endlink)))); | |
14432 | tree int_ftype_v8qi | |
14433 | = build_function_type (integer_type_node, | |
14434 | tree_cons (NULL_TREE, V8QI_type_node, | |
14435 | endlink)); | |
14436 | tree int_ftype_v4hi | |
14437 | = build_function_type (integer_type_node, | |
14438 | tree_cons (NULL_TREE, V4HI_type_node, | |
14439 | endlink)); | |
14440 | tree int_ftype_v2si | |
14441 | = build_function_type (integer_type_node, | |
14442 | tree_cons (NULL_TREE, V2SI_type_node, | |
14443 | endlink)); | |
14444 | tree int_ftype_v8qi_int | |
14445 | = build_function_type (integer_type_node, | |
14446 | tree_cons (NULL_TREE, V8QI_type_node, | |
14447 | tree_cons (NULL_TREE, integer_type_node, | |
14448 | endlink))); | |
14449 | tree int_ftype_v4hi_int | |
14450 | = build_function_type (integer_type_node, | |
14451 | tree_cons (NULL_TREE, V4HI_type_node, | |
14452 | tree_cons (NULL_TREE, integer_type_node, | |
14453 | endlink))); | |
14454 | tree int_ftype_v2si_int | |
14455 | = build_function_type (integer_type_node, | |
14456 | tree_cons (NULL_TREE, V2SI_type_node, | |
14457 | tree_cons (NULL_TREE, integer_type_node, | |
14458 | endlink))); | |
14459 | tree v8qi_ftype_v8qi_int_int | |
14460 | = build_function_type (V8QI_type_node, | |
14461 | tree_cons (NULL_TREE, V8QI_type_node, | |
14462 | tree_cons (NULL_TREE, integer_type_node, | |
14463 | tree_cons (NULL_TREE, | |
14464 | integer_type_node, | |
14465 | endlink)))); | |
14466 | tree v4hi_ftype_v4hi_int_int | |
14467 | = build_function_type (V4HI_type_node, | |
14468 | tree_cons (NULL_TREE, V4HI_type_node, | |
14469 | tree_cons (NULL_TREE, integer_type_node, | |
14470 | tree_cons (NULL_TREE, | |
14471 | integer_type_node, | |
14472 | endlink)))); | |
14473 | tree v2si_ftype_v2si_int_int | |
14474 | = build_function_type (V2SI_type_node, | |
14475 | tree_cons (NULL_TREE, V2SI_type_node, | |
14476 | tree_cons (NULL_TREE, integer_type_node, | |
14477 | tree_cons (NULL_TREE, | |
14478 | integer_type_node, | |
14479 | endlink)))); | |
14480 | /* Miscellaneous. */ | |
14481 | tree v8qi_ftype_v4hi_v4hi | |
14482 | = build_function_type (V8QI_type_node, | |
14483 | tree_cons (NULL_TREE, V4HI_type_node, | |
14484 | tree_cons (NULL_TREE, V4HI_type_node, | |
14485 | endlink))); | |
14486 | tree v4hi_ftype_v2si_v2si | |
14487 | = build_function_type (V4HI_type_node, | |
14488 | tree_cons (NULL_TREE, V2SI_type_node, | |
14489 | tree_cons (NULL_TREE, V2SI_type_node, | |
14490 | endlink))); | |
14491 | tree v2si_ftype_v4hi_v4hi | |
14492 | = build_function_type (V2SI_type_node, | |
14493 | tree_cons (NULL_TREE, V4HI_type_node, | |
14494 | tree_cons (NULL_TREE, V4HI_type_node, | |
14495 | endlink))); | |
14496 | tree v2si_ftype_v8qi_v8qi | |
14497 | = build_function_type (V2SI_type_node, | |
14498 | tree_cons (NULL_TREE, V8QI_type_node, | |
14499 | tree_cons (NULL_TREE, V8QI_type_node, | |
14500 | endlink))); | |
14501 | tree v4hi_ftype_v4hi_di | |
14502 | = build_function_type (V4HI_type_node, | |
14503 | tree_cons (NULL_TREE, V4HI_type_node, | |
14504 | tree_cons (NULL_TREE, | |
14505 | long_long_integer_type_node, | |
14506 | endlink))); | |
14507 | tree v2si_ftype_v2si_di | |
14508 | = build_function_type (V2SI_type_node, | |
14509 | tree_cons (NULL_TREE, V2SI_type_node, | |
14510 | tree_cons (NULL_TREE, | |
14511 | long_long_integer_type_node, | |
14512 | endlink))); | |
14513 | tree void_ftype_int_int | |
14514 | = build_function_type (void_type_node, | |
14515 | tree_cons (NULL_TREE, integer_type_node, | |
14516 | tree_cons (NULL_TREE, integer_type_node, | |
14517 | endlink))); | |
14518 | tree di_ftype_void | |
14519 | = build_function_type (long_long_unsigned_type_node, endlink); | |
14520 | tree di_ftype_v8qi | |
14521 | = build_function_type (long_long_integer_type_node, | |
14522 | tree_cons (NULL_TREE, V8QI_type_node, | |
14523 | endlink)); | |
14524 | tree di_ftype_v4hi | |
14525 | = build_function_type (long_long_integer_type_node, | |
14526 | tree_cons (NULL_TREE, V4HI_type_node, | |
14527 | endlink)); | |
14528 | tree di_ftype_v2si | |
14529 | = build_function_type (long_long_integer_type_node, | |
14530 | tree_cons (NULL_TREE, V2SI_type_node, | |
14531 | endlink)); | |
14532 | tree v2si_ftype_v4hi | |
14533 | = build_function_type (V2SI_type_node, | |
14534 | tree_cons (NULL_TREE, V4HI_type_node, | |
14535 | endlink)); | |
14536 | tree v4hi_ftype_v8qi | |
14537 | = build_function_type (V4HI_type_node, | |
14538 | tree_cons (NULL_TREE, V8QI_type_node, | |
14539 | endlink)); | |
14540 | ||
14541 | tree di_ftype_di_v4hi_v4hi | |
14542 | = build_function_type (long_long_unsigned_type_node, | |
14543 | tree_cons (NULL_TREE, | |
14544 | long_long_unsigned_type_node, | |
14545 | tree_cons (NULL_TREE, V4HI_type_node, | |
14546 | tree_cons (NULL_TREE, | |
14547 | V4HI_type_node, | |
14548 | endlink)))); | |
14549 | ||
14550 | tree di_ftype_v4hi_v4hi | |
14551 | = build_function_type (long_long_unsigned_type_node, | |
14552 | tree_cons (NULL_TREE, V4HI_type_node, | |
14553 | tree_cons (NULL_TREE, V4HI_type_node, | |
14554 | endlink))); | |
14555 | ||
14556 | /* Normal vector binops. */ | |
14557 | tree v8qi_ftype_v8qi_v8qi | |
14558 | = build_function_type (V8QI_type_node, | |
14559 | tree_cons (NULL_TREE, V8QI_type_node, | |
14560 | tree_cons (NULL_TREE, V8QI_type_node, | |
14561 | endlink))); | |
14562 | tree v4hi_ftype_v4hi_v4hi | |
14563 | = build_function_type (V4HI_type_node, | |
14564 | tree_cons (NULL_TREE, V4HI_type_node, | |
14565 | tree_cons (NULL_TREE, V4HI_type_node, | |
14566 | endlink))); | |
14567 | tree v2si_ftype_v2si_v2si | |
14568 | = build_function_type (V2SI_type_node, | |
14569 | tree_cons (NULL_TREE, V2SI_type_node, | |
14570 | tree_cons (NULL_TREE, V2SI_type_node, | |
14571 | endlink))); | |
14572 | tree di_ftype_di_di | |
14573 | = build_function_type (long_long_unsigned_type_node, | |
14574 | tree_cons (NULL_TREE, long_long_unsigned_type_node, | |
14575 | tree_cons (NULL_TREE, | |
14576 | long_long_unsigned_type_node, | |
14577 | endlink))); | |
14578 | ||
14579 | /* Add all builtins that are more or less simple operations on two | |
14580 | operands. */ | |
e97a46ce | 14581 | for (i = 0, d = bdesc_2arg; i < ARRAY_SIZE (bdesc_2arg); i++, d++) |
5a9335ef NC |
14582 | { |
14583 | /* Use one of the operands; the target can have a different mode for | |
14584 | mask-generating compares. */ | |
14585 | enum machine_mode mode; | |
14586 | tree type; | |
14587 | ||
14588 | if (d->name == 0) | |
14589 | continue; | |
14590 | ||
14591 | mode = insn_data[d->icode].operand[1].mode; | |
14592 | ||
14593 | switch (mode) | |
14594 | { | |
14595 | case V8QImode: | |
14596 | type = v8qi_ftype_v8qi_v8qi; | |
14597 | break; | |
14598 | case V4HImode: | |
14599 | type = v4hi_ftype_v4hi_v4hi; | |
14600 | break; | |
14601 | case V2SImode: | |
14602 | type = v2si_ftype_v2si_v2si; | |
14603 | break; | |
14604 | case DImode: | |
14605 | type = di_ftype_di_di; | |
14606 | break; | |
14607 | ||
14608 | default: | |
e6d29d15 | 14609 | gcc_unreachable (); |
5a9335ef NC |
14610 | } |
14611 | ||
14612 | def_mbuiltin (d->mask, d->name, type, d->code); | |
14613 | } | |
14614 | ||
14615 | /* Add the remaining MMX insns with somewhat more complicated types. */ | |
14616 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_wzero", di_ftype_void, ARM_BUILTIN_WZERO); | |
14617 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_setwcx", void_ftype_int_int, ARM_BUILTIN_SETWCX); | |
14618 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_getwcx", int_ftype_int, ARM_BUILTIN_GETWCX); | |
14619 | ||
14620 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_wsllh", v4hi_ftype_v4hi_di, ARM_BUILTIN_WSLLH); | |
14621 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_wsllw", v2si_ftype_v2si_di, ARM_BUILTIN_WSLLW); | |
14622 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_wslld", di_ftype_di_di, ARM_BUILTIN_WSLLD); | |
14623 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_wsllhi", v4hi_ftype_v4hi_int, ARM_BUILTIN_WSLLHI); | |
14624 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_wsllwi", v2si_ftype_v2si_int, ARM_BUILTIN_WSLLWI); | |
14625 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_wslldi", di_ftype_di_int, ARM_BUILTIN_WSLLDI); | |
14626 | ||
14627 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_wsrlh", v4hi_ftype_v4hi_di, ARM_BUILTIN_WSRLH); | |
14628 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_wsrlw", v2si_ftype_v2si_di, ARM_BUILTIN_WSRLW); | |
14629 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_wsrld", di_ftype_di_di, ARM_BUILTIN_WSRLD); | |
14630 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_wsrlhi", v4hi_ftype_v4hi_int, ARM_BUILTIN_WSRLHI); | |
14631 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_wsrlwi", v2si_ftype_v2si_int, ARM_BUILTIN_WSRLWI); | |
14632 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_wsrldi", di_ftype_di_int, ARM_BUILTIN_WSRLDI); | |
14633 | ||
14634 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_wsrah", v4hi_ftype_v4hi_di, ARM_BUILTIN_WSRAH); | |
14635 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_wsraw", v2si_ftype_v2si_di, ARM_BUILTIN_WSRAW); | |
14636 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_wsrad", di_ftype_di_di, ARM_BUILTIN_WSRAD); | |
14637 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_wsrahi", v4hi_ftype_v4hi_int, ARM_BUILTIN_WSRAHI); | |
14638 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_wsrawi", v2si_ftype_v2si_int, ARM_BUILTIN_WSRAWI); | |
14639 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_wsradi", di_ftype_di_int, ARM_BUILTIN_WSRADI); | |
14640 | ||
14641 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_wrorh", v4hi_ftype_v4hi_di, ARM_BUILTIN_WRORH); | |
14642 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_wrorw", v2si_ftype_v2si_di, ARM_BUILTIN_WRORW); | |
14643 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_wrord", di_ftype_di_di, ARM_BUILTIN_WRORD); | |
14644 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_wrorhi", v4hi_ftype_v4hi_int, ARM_BUILTIN_WRORHI); | |
14645 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_wrorwi", v2si_ftype_v2si_int, ARM_BUILTIN_WRORWI); | |
14646 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_wrordi", di_ftype_di_int, ARM_BUILTIN_WRORDI); | |
14647 | ||
14648 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_wshufh", v4hi_ftype_v4hi_int, ARM_BUILTIN_WSHUFH); | |
14649 | ||
14650 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_wsadb", v2si_ftype_v8qi_v8qi, ARM_BUILTIN_WSADB); | |
14651 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_wsadh", v2si_ftype_v4hi_v4hi, ARM_BUILTIN_WSADH); | |
14652 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_wsadbz", v2si_ftype_v8qi_v8qi, ARM_BUILTIN_WSADBZ); | |
14653 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_wsadhz", v2si_ftype_v4hi_v4hi, ARM_BUILTIN_WSADHZ); | |
14654 | ||
14655 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_textrmsb", int_ftype_v8qi_int, ARM_BUILTIN_TEXTRMSB); | |
14656 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_textrmsh", int_ftype_v4hi_int, ARM_BUILTIN_TEXTRMSH); | |
14657 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_textrmsw", int_ftype_v2si_int, ARM_BUILTIN_TEXTRMSW); | |
14658 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_textrmub", int_ftype_v8qi_int, ARM_BUILTIN_TEXTRMUB); | |
14659 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_textrmuh", int_ftype_v4hi_int, ARM_BUILTIN_TEXTRMUH); | |
14660 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_textrmuw", int_ftype_v2si_int, ARM_BUILTIN_TEXTRMUW); | |
14661 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_tinsrb", v8qi_ftype_v8qi_int_int, ARM_BUILTIN_TINSRB); | |
14662 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_tinsrh", v4hi_ftype_v4hi_int_int, ARM_BUILTIN_TINSRH); | |
14663 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_tinsrw", v2si_ftype_v2si_int_int, ARM_BUILTIN_TINSRW); | |
14664 | ||
14665 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_waccb", di_ftype_v8qi, ARM_BUILTIN_WACCB); | |
14666 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_wacch", di_ftype_v4hi, ARM_BUILTIN_WACCH); | |
14667 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_waccw", di_ftype_v2si, ARM_BUILTIN_WACCW); | |
14668 | ||
14669 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_tmovmskb", int_ftype_v8qi, ARM_BUILTIN_TMOVMSKB); | |
14670 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_tmovmskh", int_ftype_v4hi, ARM_BUILTIN_TMOVMSKH); | |
14671 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_tmovmskw", int_ftype_v2si, ARM_BUILTIN_TMOVMSKW); | |
14672 | ||
14673 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_wpackhss", v8qi_ftype_v4hi_v4hi, ARM_BUILTIN_WPACKHSS); | |
14674 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_wpackhus", v8qi_ftype_v4hi_v4hi, ARM_BUILTIN_WPACKHUS); | |
14675 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_wpackwus", v4hi_ftype_v2si_v2si, ARM_BUILTIN_WPACKWUS); | |
14676 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_wpackwss", v4hi_ftype_v2si_v2si, ARM_BUILTIN_WPACKWSS); | |
14677 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_wpackdus", v2si_ftype_di_di, ARM_BUILTIN_WPACKDUS); | |
14678 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_wpackdss", v2si_ftype_di_di, ARM_BUILTIN_WPACKDSS); | |
14679 | ||
14680 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_wunpckehub", v4hi_ftype_v8qi, ARM_BUILTIN_WUNPCKEHUB); | |
14681 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_wunpckehuh", v2si_ftype_v4hi, ARM_BUILTIN_WUNPCKEHUH); | |
14682 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_wunpckehuw", di_ftype_v2si, ARM_BUILTIN_WUNPCKEHUW); | |
14683 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_wunpckehsb", v4hi_ftype_v8qi, ARM_BUILTIN_WUNPCKEHSB); | |
14684 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_wunpckehsh", v2si_ftype_v4hi, ARM_BUILTIN_WUNPCKEHSH); | |
14685 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_wunpckehsw", di_ftype_v2si, ARM_BUILTIN_WUNPCKEHSW); | |
14686 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_wunpckelub", v4hi_ftype_v8qi, ARM_BUILTIN_WUNPCKELUB); | |
14687 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_wunpckeluh", v2si_ftype_v4hi, ARM_BUILTIN_WUNPCKELUH); | |
14688 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_wunpckeluw", di_ftype_v2si, ARM_BUILTIN_WUNPCKELUW); | |
14689 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_wunpckelsb", v4hi_ftype_v8qi, ARM_BUILTIN_WUNPCKELSB); | |
14690 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_wunpckelsh", v2si_ftype_v4hi, ARM_BUILTIN_WUNPCKELSH); | |
14691 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_wunpckelsw", di_ftype_v2si, ARM_BUILTIN_WUNPCKELSW); | |
14692 | ||
14693 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_wmacs", di_ftype_di_v4hi_v4hi, ARM_BUILTIN_WMACS); | |
14694 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_wmacsz", di_ftype_v4hi_v4hi, ARM_BUILTIN_WMACSZ); | |
14695 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_wmacu", di_ftype_di_v4hi_v4hi, ARM_BUILTIN_WMACU); | |
14696 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_wmacuz", di_ftype_v4hi_v4hi, ARM_BUILTIN_WMACUZ); | |
14697 | ||
14698 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_walign", v8qi_ftype_v8qi_v8qi_int, ARM_BUILTIN_WALIGN); | |
14699 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_tmia", di_ftype_di_int_int, ARM_BUILTIN_TMIA); | |
14700 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_tmiaph", di_ftype_di_int_int, ARM_BUILTIN_TMIAPH); | |
14701 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_tmiabb", di_ftype_di_int_int, ARM_BUILTIN_TMIABB); | |
14702 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_tmiabt", di_ftype_di_int_int, ARM_BUILTIN_TMIABT); | |
14703 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_tmiatb", di_ftype_di_int_int, ARM_BUILTIN_TMIATB); | |
14704 | def_mbuiltin (FL_IWMMXT, "__builtin_arm_tmiatt", di_ftype_di_int_int, ARM_BUILTIN_TMIATT); | |
14705 | } | |
14706 | ||
d3585b76 DJ |
14707 | static void |
14708 | arm_init_tls_builtins (void) | |
14709 | { | |
ebfe65a3 | 14710 | tree ftype, decl; |
d3585b76 DJ |
14711 | |
14712 | ftype = build_function_type (ptr_type_node, void_list_node); | |
ebfe65a3 JJ |
14713 | decl = add_builtin_function ("__builtin_thread_pointer", ftype, |
14714 | ARM_BUILTIN_THREAD_POINTER, BUILT_IN_MD, | |
14715 | NULL, NULL_TREE); | |
14716 | TREE_NOTHROW (decl) = 1; | |
14717 | TREE_READONLY (decl) = 1; | |
d3585b76 DJ |
14718 | } |
14719 | ||
88f77cba JB |
14720 | typedef enum { |
14721 | T_V8QI = 0x0001, | |
14722 | T_V4HI = 0x0002, | |
14723 | T_V2SI = 0x0004, | |
14724 | T_V2SF = 0x0008, | |
14725 | T_DI = 0x0010, | |
14726 | T_V16QI = 0x0020, | |
14727 | T_V8HI = 0x0040, | |
14728 | T_V4SI = 0x0080, | |
14729 | T_V4SF = 0x0100, | |
14730 | T_V2DI = 0x0200, | |
14731 | T_TI = 0x0400, | |
14732 | T_EI = 0x0800, | |
14733 | T_OI = 0x1000 | |
14734 | } neon_builtin_type_bits; | |
14735 | ||
14736 | #define v8qi_UP T_V8QI | |
14737 | #define v4hi_UP T_V4HI | |
14738 | #define v2si_UP T_V2SI | |
14739 | #define v2sf_UP T_V2SF | |
14740 | #define di_UP T_DI | |
14741 | #define v16qi_UP T_V16QI | |
14742 | #define v8hi_UP T_V8HI | |
14743 | #define v4si_UP T_V4SI | |
14744 | #define v4sf_UP T_V4SF | |
14745 | #define v2di_UP T_V2DI | |
14746 | #define ti_UP T_TI | |
14747 | #define ei_UP T_EI | |
14748 | #define oi_UP T_OI | |
14749 | ||
14750 | #define UP(X) X##_UP | |
14751 | ||
14752 | #define T_MAX 13 | |
14753 | ||
14754 | typedef enum { | |
14755 | NEON_BINOP, | |
14756 | NEON_TERNOP, | |
14757 | NEON_UNOP, | |
14758 | NEON_GETLANE, | |
14759 | NEON_SETLANE, | |
14760 | NEON_CREATE, | |
14761 | NEON_DUP, | |
14762 | NEON_DUPLANE, | |
14763 | NEON_COMBINE, | |
14764 | NEON_SPLIT, | |
14765 | NEON_LANEMUL, | |
14766 | NEON_LANEMULL, | |
14767 | NEON_LANEMULH, | |
14768 | NEON_LANEMAC, | |
14769 | NEON_SCALARMUL, | |
14770 | NEON_SCALARMULL, | |
14771 | NEON_SCALARMULH, | |
14772 | NEON_SCALARMAC, | |
14773 | NEON_CONVERT, | |
14774 | NEON_FIXCONV, | |
14775 | NEON_SELECT, | |
14776 | NEON_RESULTPAIR, | |
14777 | NEON_REINTERP, | |
14778 | NEON_VTBL, | |
14779 | NEON_VTBX, | |
14780 | NEON_LOAD1, | |
14781 | NEON_LOAD1LANE, | |
14782 | NEON_STORE1, | |
14783 | NEON_STORE1LANE, | |
14784 | NEON_LOADSTRUCT, | |
14785 | NEON_LOADSTRUCTLANE, | |
14786 | NEON_STORESTRUCT, | |
14787 | NEON_STORESTRUCTLANE, | |
14788 | NEON_LOGICBINOP, | |
14789 | NEON_SHIFTINSERT, | |
14790 | NEON_SHIFTIMM, | |
14791 | NEON_SHIFTACC | |
14792 | } neon_itype; | |
14793 | ||
14794 | typedef struct { | |
14795 | const char *name; | |
14796 | const neon_itype itype; | |
14797 | const neon_builtin_type_bits bits; | |
14798 | const enum insn_code codes[T_MAX]; | |
14799 | const unsigned int num_vars; | |
14800 | unsigned int base_fcode; | |
14801 | } neon_builtin_datum; | |
14802 | ||
14803 | #define CF(N,X) CODE_FOR_neon_##N##X | |
14804 | ||
14805 | #define VAR1(T, N, A) \ | |
14806 | #N, NEON_##T, UP (A), { CF (N, A) }, 1, 0 | |
14807 | #define VAR2(T, N, A, B) \ | |
14808 | #N, NEON_##T, UP (A) | UP (B), { CF (N, A), CF (N, B) }, 2, 0 | |
14809 | #define VAR3(T, N, A, B, C) \ | |
14810 | #N, NEON_##T, UP (A) | UP (B) | UP (C), \ | |
14811 | { CF (N, A), CF (N, B), CF (N, C) }, 3, 0 | |
14812 | #define VAR4(T, N, A, B, C, D) \ | |
14813 | #N, NEON_##T, UP (A) | UP (B) | UP (C) | UP (D), \ | |
14814 | { CF (N, A), CF (N, B), CF (N, C), CF (N, D) }, 4, 0 | |
14815 | #define VAR5(T, N, A, B, C, D, E) \ | |
14816 | #N, NEON_##T, UP (A) | UP (B) | UP (C) | UP (D) | UP (E), \ | |
14817 | { CF (N, A), CF (N, B), CF (N, C), CF (N, D), CF (N, E) }, 5, 0 | |
14818 | #define VAR6(T, N, A, B, C, D, E, F) \ | |
14819 | #N, NEON_##T, UP (A) | UP (B) | UP (C) | UP (D) | UP (E) | UP (F), \ | |
14820 | { CF (N, A), CF (N, B), CF (N, C), CF (N, D), CF (N, E), CF (N, F) }, 6, 0 | |
14821 | #define VAR7(T, N, A, B, C, D, E, F, G) \ | |
14822 | #N, NEON_##T, UP (A) | UP (B) | UP (C) | UP (D) | UP (E) | UP (F) | UP (G), \ | |
14823 | { CF (N, A), CF (N, B), CF (N, C), CF (N, D), CF (N, E), CF (N, F), \ | |
14824 | CF (N, G) }, 7, 0 | |
14825 | #define VAR8(T, N, A, B, C, D, E, F, G, H) \ | |
14826 | #N, NEON_##T, UP (A) | UP (B) | UP (C) | UP (D) | UP (E) | UP (F) | UP (G) \ | |
14827 | | UP (H), \ | |
14828 | { CF (N, A), CF (N, B), CF (N, C), CF (N, D), CF (N, E), CF (N, F), \ | |
14829 | CF (N, G), CF (N, H) }, 8, 0 | |
14830 | #define VAR9(T, N, A, B, C, D, E, F, G, H, I) \ | |
14831 | #N, NEON_##T, UP (A) | UP (B) | UP (C) | UP (D) | UP (E) | UP (F) | UP (G) \ | |
14832 | | UP (H) | UP (I), \ | |
14833 | { CF (N, A), CF (N, B), CF (N, C), CF (N, D), CF (N, E), CF (N, F), \ | |
14834 | CF (N, G), CF (N, H), CF (N, I) }, 9, 0 | |
14835 | #define VAR10(T, N, A, B, C, D, E, F, G, H, I, J) \ | |
14836 | #N, NEON_##T, UP (A) | UP (B) | UP (C) | UP (D) | UP (E) | UP (F) | UP (G) \ | |
14837 | | UP (H) | UP (I) | UP (J), \ | |
14838 | { CF (N, A), CF (N, B), CF (N, C), CF (N, D), CF (N, E), CF (N, F), \ | |
14839 | CF (N, G), CF (N, H), CF (N, I), CF (N, J) }, 10, 0 | |
14840 | ||
14841 | /* The mode entries in the following table correspond to the "key" type of the | |
14842 | instruction variant, i.e. equivalent to that which would be specified after | |
14843 | the assembler mnemonic, which usually refers to the last vector operand. | |
14844 | (Signed/unsigned/polynomial types are not differentiated between though, and | |
14845 | are all mapped onto the same mode for a given element size.) The modes | |
14846 | listed per instruction should be the same as those defined for that | |
14847 | instruction's pattern in neon.md. | |
14848 | WARNING: Variants should be listed in the same increasing order as | |
14849 | neon_builtin_type_bits. */ | |
14850 | ||
14851 | static neon_builtin_datum neon_builtin_data[] = | |
14852 | { | |
14853 | { VAR10 (BINOP, vadd, | |
14854 | v8qi, v4hi, v2si, v2sf, di, v16qi, v8hi, v4si, v4sf, v2di) }, | |
14855 | { VAR3 (BINOP, vaddl, v8qi, v4hi, v2si) }, | |
14856 | { VAR3 (BINOP, vaddw, v8qi, v4hi, v2si) }, | |
14857 | { VAR6 (BINOP, vhadd, v8qi, v4hi, v2si, v16qi, v8hi, v4si) }, | |
14858 | { VAR8 (BINOP, vqadd, v8qi, v4hi, v2si, di, v16qi, v8hi, v4si, v2di) }, | |
14859 | { VAR3 (BINOP, vaddhn, v8hi, v4si, v2di) }, | |
14860 | { VAR8 (BINOP, vmul, v8qi, v4hi, v2si, v2sf, v16qi, v8hi, v4si, v4sf) }, | |
14861 | { VAR8 (TERNOP, vmla, v8qi, v4hi, v2si, v2sf, v16qi, v8hi, v4si, v4sf) }, | |
14862 | { VAR3 (TERNOP, vmlal, v8qi, v4hi, v2si) }, | |
14863 | { VAR8 (TERNOP, vmls, v8qi, v4hi, v2si, v2sf, v16qi, v8hi, v4si, v4sf) }, | |
14864 | { VAR3 (TERNOP, vmlsl, v8qi, v4hi, v2si) }, | |
14865 | { VAR4 (BINOP, vqdmulh, v4hi, v2si, v8hi, v4si) }, | |
14866 | { VAR2 (TERNOP, vqdmlal, v4hi, v2si) }, | |
14867 | { VAR2 (TERNOP, vqdmlsl, v4hi, v2si) }, | |
14868 | { VAR3 (BINOP, vmull, v8qi, v4hi, v2si) }, | |
14869 | { VAR2 (SCALARMULL, vmull_n, v4hi, v2si) }, | |
14870 | { VAR2 (LANEMULL, vmull_lane, v4hi, v2si) }, | |
14871 | { VAR2 (SCALARMULL, vqdmull_n, v4hi, v2si) }, | |
14872 | { VAR2 (LANEMULL, vqdmull_lane, v4hi, v2si) }, | |
14873 | { VAR4 (SCALARMULH, vqdmulh_n, v4hi, v2si, v8hi, v4si) }, | |
14874 | { VAR4 (LANEMULH, vqdmulh_lane, v4hi, v2si, v8hi, v4si) }, | |
14875 | { VAR2 (BINOP, vqdmull, v4hi, v2si) }, | |
14876 | { VAR8 (BINOP, vshl, v8qi, v4hi, v2si, di, v16qi, v8hi, v4si, v2di) }, | |
14877 | { VAR8 (BINOP, vqshl, v8qi, v4hi, v2si, di, v16qi, v8hi, v4si, v2di) }, | |
14878 | { VAR8 (SHIFTIMM, vshr_n, v8qi, v4hi, v2si, di, v16qi, v8hi, v4si, v2di) }, | |
14879 | { VAR3 (SHIFTIMM, vshrn_n, v8hi, v4si, v2di) }, | |
14880 | { VAR3 (SHIFTIMM, vqshrn_n, v8hi, v4si, v2di) }, | |
14881 | { VAR3 (SHIFTIMM, vqshrun_n, v8hi, v4si, v2di) }, | |
14882 | { VAR8 (SHIFTIMM, vshl_n, v8qi, v4hi, v2si, di, v16qi, v8hi, v4si, v2di) }, | |
14883 | { VAR8 (SHIFTIMM, vqshl_n, v8qi, v4hi, v2si, di, v16qi, v8hi, v4si, v2di) }, | |
14884 | { VAR8 (SHIFTIMM, vqshlu_n, v8qi, v4hi, v2si, di, v16qi, v8hi, v4si, v2di) }, | |
14885 | { VAR3 (SHIFTIMM, vshll_n, v8qi, v4hi, v2si) }, | |
14886 | { VAR8 (SHIFTACC, vsra_n, v8qi, v4hi, v2si, di, v16qi, v8hi, v4si, v2di) }, | |
14887 | { VAR10 (BINOP, vsub, | |
14888 | v8qi, v4hi, v2si, v2sf, di, v16qi, v8hi, v4si, v4sf, v2di) }, | |
14889 | { VAR3 (BINOP, vsubl, v8qi, v4hi, v2si) }, | |
14890 | { VAR3 (BINOP, vsubw, v8qi, v4hi, v2si) }, | |
14891 | { VAR8 (BINOP, vqsub, v8qi, v4hi, v2si, di, v16qi, v8hi, v4si, v2di) }, | |
14892 | { VAR6 (BINOP, vhsub, v8qi, v4hi, v2si, v16qi, v8hi, v4si) }, | |
14893 | { VAR3 (BINOP, vsubhn, v8hi, v4si, v2di) }, | |
14894 | { VAR8 (BINOP, vceq, v8qi, v4hi, v2si, v2sf, v16qi, v8hi, v4si, v4sf) }, | |
14895 | { VAR8 (BINOP, vcge, v8qi, v4hi, v2si, v2sf, v16qi, v8hi, v4si, v4sf) }, | |
14896 | { VAR8 (BINOP, vcgt, v8qi, v4hi, v2si, v2sf, v16qi, v8hi, v4si, v4sf) }, | |
14897 | { VAR2 (BINOP, vcage, v2sf, v4sf) }, | |
14898 | { VAR2 (BINOP, vcagt, v2sf, v4sf) }, | |
14899 | { VAR6 (BINOP, vtst, v8qi, v4hi, v2si, v16qi, v8hi, v4si) }, | |
14900 | { VAR8 (BINOP, vabd, v8qi, v4hi, v2si, v2sf, v16qi, v8hi, v4si, v4sf) }, | |
14901 | { VAR3 (BINOP, vabdl, v8qi, v4hi, v2si) }, | |
14902 | { VAR6 (TERNOP, vaba, v8qi, v4hi, v2si, v16qi, v8hi, v4si) }, | |
14903 | { VAR3 (TERNOP, vabal, v8qi, v4hi, v2si) }, | |
14904 | { VAR8 (BINOP, vmax, v8qi, v4hi, v2si, v2sf, v16qi, v8hi, v4si, v4sf) }, | |
14905 | { VAR8 (BINOP, vmin, v8qi, v4hi, v2si, v2sf, v16qi, v8hi, v4si, v4sf) }, | |
14906 | { VAR4 (BINOP, vpadd, v8qi, v4hi, v2si, v2sf) }, | |
14907 | { VAR6 (UNOP, vpaddl, v8qi, v4hi, v2si, v16qi, v8hi, v4si) }, | |
14908 | { VAR6 (BINOP, vpadal, v8qi, v4hi, v2si, v16qi, v8hi, v4si) }, | |
14909 | { VAR4 (BINOP, vpmax, v8qi, v4hi, v2si, v2sf) }, | |
14910 | { VAR4 (BINOP, vpmin, v8qi, v4hi, v2si, v2sf) }, | |
14911 | { VAR2 (BINOP, vrecps, v2sf, v4sf) }, | |
14912 | { VAR2 (BINOP, vrsqrts, v2sf, v4sf) }, | |
14913 | { VAR8 (SHIFTINSERT, vsri_n, v8qi, v4hi, v2si, di, v16qi, v8hi, v4si, v2di) }, | |
14914 | { VAR8 (SHIFTINSERT, vsli_n, v8qi, v4hi, v2si, di, v16qi, v8hi, v4si, v2di) }, | |
14915 | { VAR8 (UNOP, vabs, v8qi, v4hi, v2si, v2sf, v16qi, v8hi, v4si, v4sf) }, | |
14916 | { VAR6 (UNOP, vqabs, v8qi, v4hi, v2si, v16qi, v8hi, v4si) }, | |
14917 | { VAR8 (UNOP, vneg, v8qi, v4hi, v2si, v2sf, v16qi, v8hi, v4si, v4sf) }, | |
14918 | { VAR6 (UNOP, vqneg, v8qi, v4hi, v2si, v16qi, v8hi, v4si) }, | |
14919 | { VAR6 (UNOP, vcls, v8qi, v4hi, v2si, v16qi, v8hi, v4si) }, | |
14920 | { VAR6 (UNOP, vclz, v8qi, v4hi, v2si, v16qi, v8hi, v4si) }, | |
14921 | { VAR2 (UNOP, vcnt, v8qi, v16qi) }, | |
14922 | { VAR4 (UNOP, vrecpe, v2si, v2sf, v4si, v4sf) }, | |
14923 | { VAR4 (UNOP, vrsqrte, v2si, v2sf, v4si, v4sf) }, | |
14924 | { VAR6 (UNOP, vmvn, v8qi, v4hi, v2si, v16qi, v8hi, v4si) }, | |
14925 | /* FIXME: vget_lane supports more variants than this! */ | |
14926 | { VAR10 (GETLANE, vget_lane, | |
14927 | v8qi, v4hi, v2si, v2sf, di, v16qi, v8hi, v4si, v4sf, v2di) }, | |
14928 | { VAR10 (SETLANE, vset_lane, | |
14929 | v8qi, v4hi, v2si, v2sf, di, v16qi, v8hi, v4si, v4sf, v2di) }, | |
14930 | { VAR5 (CREATE, vcreate, v8qi, v4hi, v2si, v2sf, di) }, | |
14931 | { VAR10 (DUP, vdup_n, | |
14932 | v8qi, v4hi, v2si, v2sf, di, v16qi, v8hi, v4si, v4sf, v2di) }, | |
14933 | { VAR10 (DUPLANE, vdup_lane, | |
14934 | v8qi, v4hi, v2si, v2sf, di, v16qi, v8hi, v4si, v4sf, v2di) }, | |
14935 | { VAR5 (COMBINE, vcombine, v8qi, v4hi, v2si, v2sf, di) }, | |
14936 | { VAR5 (SPLIT, vget_high, v16qi, v8hi, v4si, v4sf, v2di) }, | |
14937 | { VAR5 (SPLIT, vget_low, v16qi, v8hi, v4si, v4sf, v2di) }, | |
14938 | { VAR3 (UNOP, vmovn, v8hi, v4si, v2di) }, | |
14939 | { VAR3 (UNOP, vqmovn, v8hi, v4si, v2di) }, | |
14940 | { VAR3 (UNOP, vqmovun, v8hi, v4si, v2di) }, | |
14941 | { VAR3 (UNOP, vmovl, v8qi, v4hi, v2si) }, | |
14942 | { VAR6 (LANEMUL, vmul_lane, v4hi, v2si, v2sf, v8hi, v4si, v4sf) }, | |
14943 | { VAR6 (LANEMAC, vmla_lane, v4hi, v2si, v2sf, v8hi, v4si, v4sf) }, | |
14944 | { VAR2 (LANEMAC, vmlal_lane, v4hi, v2si) }, | |
14945 | { VAR2 (LANEMAC, vqdmlal_lane, v4hi, v2si) }, | |
14946 | { VAR6 (LANEMAC, vmls_lane, v4hi, v2si, v2sf, v8hi, v4si, v4sf) }, | |
14947 | { VAR2 (LANEMAC, vmlsl_lane, v4hi, v2si) }, | |
14948 | { VAR2 (LANEMAC, vqdmlsl_lane, v4hi, v2si) }, | |
14949 | { VAR6 (SCALARMUL, vmul_n, v4hi, v2si, v2sf, v8hi, v4si, v4sf) }, | |
14950 | { VAR6 (SCALARMAC, vmla_n, v4hi, v2si, v2sf, v8hi, v4si, v4sf) }, | |
14951 | { VAR2 (SCALARMAC, vmlal_n, v4hi, v2si) }, | |
14952 | { VAR2 (SCALARMAC, vqdmlal_n, v4hi, v2si) }, | |
14953 | { VAR6 (SCALARMAC, vmls_n, v4hi, v2si, v2sf, v8hi, v4si, v4sf) }, | |
14954 | { VAR2 (SCALARMAC, vmlsl_n, v4hi, v2si) }, | |
14955 | { VAR2 (SCALARMAC, vqdmlsl_n, v4hi, v2si) }, | |
14956 | { VAR10 (BINOP, vext, | |
14957 | v8qi, v4hi, v2si, v2sf, di, v16qi, v8hi, v4si, v4sf, v2di) }, | |
14958 | { VAR8 (UNOP, vrev64, v8qi, v4hi, v2si, v2sf, v16qi, v8hi, v4si, v4sf) }, | |
14959 | { VAR4 (UNOP, vrev32, v8qi, v4hi, v16qi, v8hi) }, | |
14960 | { VAR2 (UNOP, vrev16, v8qi, v16qi) }, | |
14961 | { VAR4 (CONVERT, vcvt, v2si, v2sf, v4si, v4sf) }, | |
14962 | { VAR4 (FIXCONV, vcvt_n, v2si, v2sf, v4si, v4sf) }, | |
14963 | { VAR10 (SELECT, vbsl, | |
14964 | v8qi, v4hi, v2si, v2sf, di, v16qi, v8hi, v4si, v4sf, v2di) }, | |
14965 | { VAR1 (VTBL, vtbl1, v8qi) }, | |
14966 | { VAR1 (VTBL, vtbl2, v8qi) }, | |
14967 | { VAR1 (VTBL, vtbl3, v8qi) }, | |
14968 | { VAR1 (VTBL, vtbl4, v8qi) }, | |
14969 | { VAR1 (VTBX, vtbx1, v8qi) }, | |
14970 | { VAR1 (VTBX, vtbx2, v8qi) }, | |
14971 | { VAR1 (VTBX, vtbx3, v8qi) }, | |
14972 | { VAR1 (VTBX, vtbx4, v8qi) }, | |
14973 | { VAR8 (RESULTPAIR, vtrn, v8qi, v4hi, v2si, v2sf, v16qi, v8hi, v4si, v4sf) }, | |
14974 | { VAR8 (RESULTPAIR, vzip, v8qi, v4hi, v2si, v2sf, v16qi, v8hi, v4si, v4sf) }, | |
14975 | { VAR8 (RESULTPAIR, vuzp, v8qi, v4hi, v2si, v2sf, v16qi, v8hi, v4si, v4sf) }, | |
14976 | { VAR5 (REINTERP, vreinterpretv8qi, v8qi, v4hi, v2si, v2sf, di) }, | |
14977 | { VAR5 (REINTERP, vreinterpretv4hi, v8qi, v4hi, v2si, v2sf, di) }, | |
14978 | { VAR5 (REINTERP, vreinterpretv2si, v8qi, v4hi, v2si, v2sf, di) }, | |
14979 | { VAR5 (REINTERP, vreinterpretv2sf, v8qi, v4hi, v2si, v2sf, di) }, | |
14980 | { VAR5 (REINTERP, vreinterpretdi, v8qi, v4hi, v2si, v2sf, di) }, | |
14981 | { VAR5 (REINTERP, vreinterpretv16qi, v16qi, v8hi, v4si, v4sf, v2di) }, | |
14982 | { VAR5 (REINTERP, vreinterpretv8hi, v16qi, v8hi, v4si, v4sf, v2di) }, | |
14983 | { VAR5 (REINTERP, vreinterpretv4si, v16qi, v8hi, v4si, v4sf, v2di) }, | |
14984 | { VAR5 (REINTERP, vreinterpretv4sf, v16qi, v8hi, v4si, v4sf, v2di) }, | |
14985 | { VAR5 (REINTERP, vreinterpretv2di, v16qi, v8hi, v4si, v4sf, v2di) }, | |
14986 | { VAR10 (LOAD1, vld1, | |
14987 | v8qi, v4hi, v2si, v2sf, di, v16qi, v8hi, v4si, v4sf, v2di) }, | |
14988 | { VAR10 (LOAD1LANE, vld1_lane, | |
14989 | v8qi, v4hi, v2si, v2sf, di, v16qi, v8hi, v4si, v4sf, v2di) }, | |
14990 | { VAR10 (LOAD1, vld1_dup, | |
14991 | v8qi, v4hi, v2si, v2sf, di, v16qi, v8hi, v4si, v4sf, v2di) }, | |
14992 | { VAR10 (STORE1, vst1, | |
14993 | v8qi, v4hi, v2si, v2sf, di, v16qi, v8hi, v4si, v4sf, v2di) }, | |
14994 | { VAR10 (STORE1LANE, vst1_lane, | |
14995 | v8qi, v4hi, v2si, v2sf, di, v16qi, v8hi, v4si, v4sf, v2di) }, | |
14996 | { VAR9 (LOADSTRUCT, | |
14997 | vld2, v8qi, v4hi, v2si, v2sf, di, v16qi, v8hi, v4si, v4sf) }, | |
14998 | { VAR7 (LOADSTRUCTLANE, vld2_lane, | |
14999 | v8qi, v4hi, v2si, v2sf, v8hi, v4si, v4sf) }, | |
15000 | { VAR5 (LOADSTRUCT, vld2_dup, v8qi, v4hi, v2si, v2sf, di) }, | |
15001 | { VAR9 (STORESTRUCT, vst2, | |
15002 | v8qi, v4hi, v2si, v2sf, di, v16qi, v8hi, v4si, v4sf) }, | |
15003 | { VAR7 (STORESTRUCTLANE, vst2_lane, | |
15004 | v8qi, v4hi, v2si, v2sf, v8hi, v4si, v4sf) }, | |
15005 | { VAR9 (LOADSTRUCT, | |
15006 | vld3, v8qi, v4hi, v2si, v2sf, di, v16qi, v8hi, v4si, v4sf) }, | |
15007 | { VAR7 (LOADSTRUCTLANE, vld3_lane, | |
15008 | v8qi, v4hi, v2si, v2sf, v8hi, v4si, v4sf) }, | |
15009 | { VAR5 (LOADSTRUCT, vld3_dup, v8qi, v4hi, v2si, v2sf, di) }, | |
15010 | { VAR9 (STORESTRUCT, vst3, | |
15011 | v8qi, v4hi, v2si, v2sf, di, v16qi, v8hi, v4si, v4sf) }, | |
15012 | { VAR7 (STORESTRUCTLANE, vst3_lane, | |
15013 | v8qi, v4hi, v2si, v2sf, v8hi, v4si, v4sf) }, | |
15014 | { VAR9 (LOADSTRUCT, vld4, | |
15015 | v8qi, v4hi, v2si, v2sf, di, v16qi, v8hi, v4si, v4sf) }, | |
15016 | { VAR7 (LOADSTRUCTLANE, vld4_lane, | |
15017 | v8qi, v4hi, v2si, v2sf, v8hi, v4si, v4sf) }, | |
15018 | { VAR5 (LOADSTRUCT, vld4_dup, v8qi, v4hi, v2si, v2sf, di) }, | |
15019 | { VAR9 (STORESTRUCT, vst4, | |
15020 | v8qi, v4hi, v2si, v2sf, di, v16qi, v8hi, v4si, v4sf) }, | |
15021 | { VAR7 (STORESTRUCTLANE, vst4_lane, | |
15022 | v8qi, v4hi, v2si, v2sf, v8hi, v4si, v4sf) }, | |
15023 | { VAR10 (LOGICBINOP, vand, | |
15024 | v8qi, v4hi, v2si, v2sf, di, v16qi, v8hi, v4si, v4sf, v2di) }, | |
15025 | { VAR10 (LOGICBINOP, vorr, | |
15026 | v8qi, v4hi, v2si, v2sf, di, v16qi, v8hi, v4si, v4sf, v2di) }, | |
15027 | { VAR10 (BINOP, veor, | |
15028 | v8qi, v4hi, v2si, v2sf, di, v16qi, v8hi, v4si, v4sf, v2di) }, | |
15029 | { VAR10 (LOGICBINOP, vbic, | |
15030 | v8qi, v4hi, v2si, v2sf, di, v16qi, v8hi, v4si, v4sf, v2di) }, | |
15031 | { VAR10 (LOGICBINOP, vorn, | |
15032 | v8qi, v4hi, v2si, v2sf, di, v16qi, v8hi, v4si, v4sf, v2di) } | |
15033 | }; | |
15034 | ||
15035 | #undef CF | |
15036 | #undef VAR1 | |
15037 | #undef VAR2 | |
15038 | #undef VAR3 | |
15039 | #undef VAR4 | |
15040 | #undef VAR5 | |
15041 | #undef VAR6 | |
15042 | #undef VAR7 | |
15043 | #undef VAR8 | |
15044 | #undef VAR9 | |
15045 | #undef VAR10 | |
15046 | ||
15047 | static void | |
15048 | arm_init_neon_builtins (void) | |
15049 | { | |
15050 | unsigned int i, fcode = ARM_BUILTIN_NEON_BASE; | |
15051 | ||
af06585a JM |
15052 | tree neon_intQI_type_node; |
15053 | tree neon_intHI_type_node; | |
15054 | tree neon_polyQI_type_node; | |
15055 | tree neon_polyHI_type_node; | |
15056 | tree neon_intSI_type_node; | |
15057 | tree neon_intDI_type_node; | |
15058 | tree neon_float_type_node; | |
15059 | ||
15060 | tree intQI_pointer_node; | |
15061 | tree intHI_pointer_node; | |
15062 | tree intSI_pointer_node; | |
15063 | tree intDI_pointer_node; | |
15064 | tree float_pointer_node; | |
15065 | ||
15066 | tree const_intQI_node; | |
15067 | tree const_intHI_node; | |
15068 | tree const_intSI_node; | |
15069 | tree const_intDI_node; | |
15070 | tree const_float_node; | |
15071 | ||
15072 | tree const_intQI_pointer_node; | |
15073 | tree const_intHI_pointer_node; | |
15074 | tree const_intSI_pointer_node; | |
15075 | tree const_intDI_pointer_node; | |
15076 | tree const_float_pointer_node; | |
15077 | ||
15078 | tree V8QI_type_node; | |
15079 | tree V4HI_type_node; | |
15080 | tree V2SI_type_node; | |
15081 | tree V2SF_type_node; | |
15082 | tree V16QI_type_node; | |
15083 | tree V8HI_type_node; | |
15084 | tree V4SI_type_node; | |
15085 | tree V4SF_type_node; | |
15086 | tree V2DI_type_node; | |
15087 | ||
15088 | tree intUQI_type_node; | |
15089 | tree intUHI_type_node; | |
15090 | tree intUSI_type_node; | |
15091 | tree intUDI_type_node; | |
15092 | ||
15093 | tree intEI_type_node; | |
15094 | tree intOI_type_node; | |
15095 | tree intCI_type_node; | |
15096 | tree intXI_type_node; | |
15097 | ||
15098 | tree V8QI_pointer_node; | |
15099 | tree V4HI_pointer_node; | |
15100 | tree V2SI_pointer_node; | |
15101 | tree V2SF_pointer_node; | |
15102 | tree V16QI_pointer_node; | |
15103 | tree V8HI_pointer_node; | |
15104 | tree V4SI_pointer_node; | |
15105 | tree V4SF_pointer_node; | |
15106 | tree V2DI_pointer_node; | |
15107 | ||
15108 | tree void_ftype_pv8qi_v8qi_v8qi; | |
15109 | tree void_ftype_pv4hi_v4hi_v4hi; | |
15110 | tree void_ftype_pv2si_v2si_v2si; | |
15111 | tree void_ftype_pv2sf_v2sf_v2sf; | |
15112 | tree void_ftype_pdi_di_di; | |
15113 | tree void_ftype_pv16qi_v16qi_v16qi; | |
15114 | tree void_ftype_pv8hi_v8hi_v8hi; | |
15115 | tree void_ftype_pv4si_v4si_v4si; | |
15116 | tree void_ftype_pv4sf_v4sf_v4sf; | |
15117 | tree void_ftype_pv2di_v2di_v2di; | |
15118 | ||
15119 | tree reinterp_ftype_dreg[5][5]; | |
15120 | tree reinterp_ftype_qreg[5][5]; | |
15121 | tree dreg_types[5], qreg_types[5]; | |
15122 | ||
88f77cba JB |
15123 | /* Create distinguished type nodes for NEON vector element types, |
15124 | and pointers to values of such types, so we can detect them later. */ | |
af06585a JM |
15125 | neon_intQI_type_node = make_signed_type (GET_MODE_PRECISION (QImode)); |
15126 | neon_intHI_type_node = make_signed_type (GET_MODE_PRECISION (HImode)); | |
15127 | neon_polyQI_type_node = make_signed_type (GET_MODE_PRECISION (QImode)); | |
15128 | neon_polyHI_type_node = make_signed_type (GET_MODE_PRECISION (HImode)); | |
15129 | neon_intSI_type_node = make_signed_type (GET_MODE_PRECISION (SImode)); | |
15130 | neon_intDI_type_node = make_signed_type (GET_MODE_PRECISION (DImode)); | |
15131 | neon_float_type_node = make_node (REAL_TYPE); | |
15132 | TYPE_PRECISION (neon_float_type_node) = FLOAT_TYPE_SIZE; | |
15133 | layout_type (neon_float_type_node); | |
15134 | ||
bcbdbbb0 JM |
15135 | /* Define typedefs which exactly correspond to the modes we are basing vector |
15136 | types on. If you change these names you'll need to change | |
15137 | the table used by arm_mangle_type too. */ | |
15138 | (*lang_hooks.types.register_builtin_type) (neon_intQI_type_node, | |
15139 | "__builtin_neon_qi"); | |
15140 | (*lang_hooks.types.register_builtin_type) (neon_intHI_type_node, | |
15141 | "__builtin_neon_hi"); | |
15142 | (*lang_hooks.types.register_builtin_type) (neon_intSI_type_node, | |
15143 | "__builtin_neon_si"); | |
15144 | (*lang_hooks.types.register_builtin_type) (neon_float_type_node, | |
15145 | "__builtin_neon_sf"); | |
15146 | (*lang_hooks.types.register_builtin_type) (neon_intDI_type_node, | |
15147 | "__builtin_neon_di"); | |
15148 | (*lang_hooks.types.register_builtin_type) (neon_polyQI_type_node, | |
15149 | "__builtin_neon_poly8"); | |
15150 | (*lang_hooks.types.register_builtin_type) (neon_polyHI_type_node, | |
15151 | "__builtin_neon_poly16"); | |
15152 | ||
af06585a JM |
15153 | intQI_pointer_node = build_pointer_type (neon_intQI_type_node); |
15154 | intHI_pointer_node = build_pointer_type (neon_intHI_type_node); | |
15155 | intSI_pointer_node = build_pointer_type (neon_intSI_type_node); | |
15156 | intDI_pointer_node = build_pointer_type (neon_intDI_type_node); | |
15157 | float_pointer_node = build_pointer_type (neon_float_type_node); | |
88f77cba JB |
15158 | |
15159 | /* Next create constant-qualified versions of the above types. */ | |
af06585a JM |
15160 | const_intQI_node = build_qualified_type (neon_intQI_type_node, |
15161 | TYPE_QUAL_CONST); | |
15162 | const_intHI_node = build_qualified_type (neon_intHI_type_node, | |
15163 | TYPE_QUAL_CONST); | |
15164 | const_intSI_node = build_qualified_type (neon_intSI_type_node, | |
15165 | TYPE_QUAL_CONST); | |
15166 | const_intDI_node = build_qualified_type (neon_intDI_type_node, | |
15167 | TYPE_QUAL_CONST); | |
15168 | const_float_node = build_qualified_type (neon_float_type_node, | |
15169 | TYPE_QUAL_CONST); | |
15170 | ||
15171 | const_intQI_pointer_node = build_pointer_type (const_intQI_node); | |
15172 | const_intHI_pointer_node = build_pointer_type (const_intHI_node); | |
15173 | const_intSI_pointer_node = build_pointer_type (const_intSI_node); | |
15174 | const_intDI_pointer_node = build_pointer_type (const_intDI_node); | |
15175 | const_float_pointer_node = build_pointer_type (const_float_node); | |
88f77cba JB |
15176 | |
15177 | /* Now create vector types based on our NEON element types. */ | |
15178 | /* 64-bit vectors. */ | |
af06585a | 15179 | V8QI_type_node = |
88f77cba | 15180 | build_vector_type_for_mode (neon_intQI_type_node, V8QImode); |
af06585a | 15181 | V4HI_type_node = |
88f77cba | 15182 | build_vector_type_for_mode (neon_intHI_type_node, V4HImode); |
af06585a | 15183 | V2SI_type_node = |
88f77cba | 15184 | build_vector_type_for_mode (neon_intSI_type_node, V2SImode); |
af06585a | 15185 | V2SF_type_node = |
88f77cba JB |
15186 | build_vector_type_for_mode (neon_float_type_node, V2SFmode); |
15187 | /* 128-bit vectors. */ | |
af06585a | 15188 | V16QI_type_node = |
88f77cba | 15189 | build_vector_type_for_mode (neon_intQI_type_node, V16QImode); |
af06585a | 15190 | V8HI_type_node = |
88f77cba | 15191 | build_vector_type_for_mode (neon_intHI_type_node, V8HImode); |
af06585a | 15192 | V4SI_type_node = |
88f77cba | 15193 | build_vector_type_for_mode (neon_intSI_type_node, V4SImode); |
af06585a | 15194 | V4SF_type_node = |
88f77cba | 15195 | build_vector_type_for_mode (neon_float_type_node, V4SFmode); |
af06585a | 15196 | V2DI_type_node = |
88f77cba JB |
15197 | build_vector_type_for_mode (neon_intDI_type_node, V2DImode); |
15198 | ||
15199 | /* Unsigned integer types for various mode sizes. */ | |
af06585a JM |
15200 | intUQI_type_node = make_unsigned_type (GET_MODE_PRECISION (QImode)); |
15201 | intUHI_type_node = make_unsigned_type (GET_MODE_PRECISION (HImode)); | |
15202 | intUSI_type_node = make_unsigned_type (GET_MODE_PRECISION (SImode)); | |
15203 | intUDI_type_node = make_unsigned_type (GET_MODE_PRECISION (DImode)); | |
88f77cba | 15204 | |
bcbdbbb0 JM |
15205 | (*lang_hooks.types.register_builtin_type) (intUQI_type_node, |
15206 | "__builtin_neon_uqi"); | |
15207 | (*lang_hooks.types.register_builtin_type) (intUHI_type_node, | |
15208 | "__builtin_neon_uhi"); | |
15209 | (*lang_hooks.types.register_builtin_type) (intUSI_type_node, | |
15210 | "__builtin_neon_usi"); | |
15211 | (*lang_hooks.types.register_builtin_type) (intUDI_type_node, | |
15212 | "__builtin_neon_udi"); | |
15213 | ||
88f77cba | 15214 | /* Opaque integer types for structures of vectors. */ |
af06585a JM |
15215 | intEI_type_node = make_signed_type (GET_MODE_PRECISION (EImode)); |
15216 | intOI_type_node = make_signed_type (GET_MODE_PRECISION (OImode)); | |
15217 | intCI_type_node = make_signed_type (GET_MODE_PRECISION (CImode)); | |
15218 | intXI_type_node = make_signed_type (GET_MODE_PRECISION (XImode)); | |
88f77cba | 15219 | |
bcbdbbb0 JM |
15220 | (*lang_hooks.types.register_builtin_type) (intTI_type_node, |
15221 | "__builtin_neon_ti"); | |
15222 | (*lang_hooks.types.register_builtin_type) (intEI_type_node, | |
15223 | "__builtin_neon_ei"); | |
15224 | (*lang_hooks.types.register_builtin_type) (intOI_type_node, | |
15225 | "__builtin_neon_oi"); | |
15226 | (*lang_hooks.types.register_builtin_type) (intCI_type_node, | |
15227 | "__builtin_neon_ci"); | |
15228 | (*lang_hooks.types.register_builtin_type) (intXI_type_node, | |
15229 | "__builtin_neon_xi"); | |
15230 | ||
88f77cba | 15231 | /* Pointers to vector types. */ |
af06585a JM |
15232 | V8QI_pointer_node = build_pointer_type (V8QI_type_node); |
15233 | V4HI_pointer_node = build_pointer_type (V4HI_type_node); | |
15234 | V2SI_pointer_node = build_pointer_type (V2SI_type_node); | |
15235 | V2SF_pointer_node = build_pointer_type (V2SF_type_node); | |
15236 | V16QI_pointer_node = build_pointer_type (V16QI_type_node); | |
15237 | V8HI_pointer_node = build_pointer_type (V8HI_type_node); | |
15238 | V4SI_pointer_node = build_pointer_type (V4SI_type_node); | |
15239 | V4SF_pointer_node = build_pointer_type (V4SF_type_node); | |
15240 | V2DI_pointer_node = build_pointer_type (V2DI_type_node); | |
88f77cba JB |
15241 | |
15242 | /* Operations which return results as pairs. */ | |
af06585a | 15243 | void_ftype_pv8qi_v8qi_v8qi = |
88f77cba JB |
15244 | build_function_type_list (void_type_node, V8QI_pointer_node, V8QI_type_node, |
15245 | V8QI_type_node, NULL); | |
af06585a | 15246 | void_ftype_pv4hi_v4hi_v4hi = |
88f77cba JB |
15247 | build_function_type_list (void_type_node, V4HI_pointer_node, V4HI_type_node, |
15248 | V4HI_type_node, NULL); | |
af06585a | 15249 | void_ftype_pv2si_v2si_v2si = |
88f77cba JB |
15250 | build_function_type_list (void_type_node, V2SI_pointer_node, V2SI_type_node, |
15251 | V2SI_type_node, NULL); | |
af06585a | 15252 | void_ftype_pv2sf_v2sf_v2sf = |
88f77cba JB |
15253 | build_function_type_list (void_type_node, V2SF_pointer_node, V2SF_type_node, |
15254 | V2SF_type_node, NULL); | |
af06585a | 15255 | void_ftype_pdi_di_di = |
88f77cba JB |
15256 | build_function_type_list (void_type_node, intDI_pointer_node, |
15257 | neon_intDI_type_node, neon_intDI_type_node, NULL); | |
af06585a | 15258 | void_ftype_pv16qi_v16qi_v16qi = |
88f77cba JB |
15259 | build_function_type_list (void_type_node, V16QI_pointer_node, |
15260 | V16QI_type_node, V16QI_type_node, NULL); | |
af06585a | 15261 | void_ftype_pv8hi_v8hi_v8hi = |
88f77cba JB |
15262 | build_function_type_list (void_type_node, V8HI_pointer_node, V8HI_type_node, |
15263 | V8HI_type_node, NULL); | |
af06585a | 15264 | void_ftype_pv4si_v4si_v4si = |
88f77cba JB |
15265 | build_function_type_list (void_type_node, V4SI_pointer_node, V4SI_type_node, |
15266 | V4SI_type_node, NULL); | |
af06585a | 15267 | void_ftype_pv4sf_v4sf_v4sf = |
88f77cba JB |
15268 | build_function_type_list (void_type_node, V4SF_pointer_node, V4SF_type_node, |
15269 | V4SF_type_node, NULL); | |
af06585a | 15270 | void_ftype_pv2di_v2di_v2di = |
88f77cba JB |
15271 | build_function_type_list (void_type_node, V2DI_pointer_node, V2DI_type_node, |
15272 | V2DI_type_node, NULL); | |
15273 | ||
88f77cba JB |
15274 | dreg_types[0] = V8QI_type_node; |
15275 | dreg_types[1] = V4HI_type_node; | |
15276 | dreg_types[2] = V2SI_type_node; | |
15277 | dreg_types[3] = V2SF_type_node; | |
15278 | dreg_types[4] = neon_intDI_type_node; | |
15279 | ||
15280 | qreg_types[0] = V16QI_type_node; | |
15281 | qreg_types[1] = V8HI_type_node; | |
15282 | qreg_types[2] = V4SI_type_node; | |
15283 | qreg_types[3] = V4SF_type_node; | |
15284 | qreg_types[4] = V2DI_type_node; | |
15285 | ||
15286 | for (i = 0; i < 5; i++) | |
15287 | { | |
15288 | int j; | |
15289 | for (j = 0; j < 5; j++) | |
15290 | { | |
15291 | reinterp_ftype_dreg[i][j] | |
15292 | = build_function_type_list (dreg_types[i], dreg_types[j], NULL); | |
15293 | reinterp_ftype_qreg[i][j] | |
15294 | = build_function_type_list (qreg_types[i], qreg_types[j], NULL); | |
15295 | } | |
15296 | } | |
15297 | ||
15298 | for (i = 0; i < ARRAY_SIZE (neon_builtin_data); i++) | |
15299 | { | |
15300 | neon_builtin_datum *d = &neon_builtin_data[i]; | |
15301 | unsigned int j, codeidx = 0; | |
15302 | ||
15303 | d->base_fcode = fcode; | |
15304 | ||
15305 | for (j = 0; j < T_MAX; j++) | |
15306 | { | |
15307 | const char* const modenames[] = { | |
15308 | "v8qi", "v4hi", "v2si", "v2sf", "di", | |
15309 | "v16qi", "v8hi", "v4si", "v4sf", "v2di" | |
15310 | }; | |
15311 | char namebuf[60]; | |
15312 | tree ftype = NULL; | |
15313 | enum insn_code icode; | |
15314 | int is_load = 0, is_store = 0; | |
15315 | ||
15316 | if ((d->bits & (1 << j)) == 0) | |
15317 | continue; | |
15318 | ||
15319 | icode = d->codes[codeidx++]; | |
15320 | ||
15321 | switch (d->itype) | |
15322 | { | |
15323 | case NEON_LOAD1: | |
15324 | case NEON_LOAD1LANE: | |
15325 | case NEON_LOADSTRUCT: | |
15326 | case NEON_LOADSTRUCTLANE: | |
15327 | is_load = 1; | |
15328 | /* Fall through. */ | |
15329 | case NEON_STORE1: | |
15330 | case NEON_STORE1LANE: | |
15331 | case NEON_STORESTRUCT: | |
15332 | case NEON_STORESTRUCTLANE: | |
15333 | if (!is_load) | |
15334 | is_store = 1; | |
15335 | /* Fall through. */ | |
15336 | case NEON_UNOP: | |
15337 | case NEON_BINOP: | |
15338 | case NEON_LOGICBINOP: | |
15339 | case NEON_SHIFTINSERT: | |
15340 | case NEON_TERNOP: | |
15341 | case NEON_GETLANE: | |
15342 | case NEON_SETLANE: | |
15343 | case NEON_CREATE: | |
15344 | case NEON_DUP: | |
15345 | case NEON_DUPLANE: | |
15346 | case NEON_SHIFTIMM: | |
15347 | case NEON_SHIFTACC: | |
15348 | case NEON_COMBINE: | |
15349 | case NEON_SPLIT: | |
15350 | case NEON_CONVERT: | |
15351 | case NEON_FIXCONV: | |
15352 | case NEON_LANEMUL: | |
15353 | case NEON_LANEMULL: | |
15354 | case NEON_LANEMULH: | |
15355 | case NEON_LANEMAC: | |
15356 | case NEON_SCALARMUL: | |
15357 | case NEON_SCALARMULL: | |
15358 | case NEON_SCALARMULH: | |
15359 | case NEON_SCALARMAC: | |
15360 | case NEON_SELECT: | |
15361 | case NEON_VTBL: | |
15362 | case NEON_VTBX: | |
15363 | { | |
15364 | int k; | |
15365 | tree return_type = void_type_node, args = void_list_node; | |
15366 | ||
15367 | /* Build a function type directly from the insn_data for this | |
15368 | builtin. The build_function_type() function takes care of | |
15369 | removing duplicates for us. */ | |
15370 | for (k = insn_data[icode].n_operands - 1; k >= 0; k--) | |
15371 | { | |
15372 | tree eltype; | |
15373 | ||
15374 | if (is_load && k == 1) | |
15375 | { | |
15376 | /* Neon load patterns always have the memory operand | |
15377 | (a SImode pointer) in the operand 1 position. We | |
15378 | want a const pointer to the element type in that | |
15379 | position. */ | |
15380 | gcc_assert (insn_data[icode].operand[k].mode == SImode); | |
15381 | ||
15382 | switch (1 << j) | |
15383 | { | |
15384 | case T_V8QI: | |
15385 | case T_V16QI: | |
15386 | eltype = const_intQI_pointer_node; | |
15387 | break; | |
15388 | ||
15389 | case T_V4HI: | |
15390 | case T_V8HI: | |
15391 | eltype = const_intHI_pointer_node; | |
15392 | break; | |
15393 | ||
15394 | case T_V2SI: | |
15395 | case T_V4SI: | |
15396 | eltype = const_intSI_pointer_node; | |
15397 | break; | |
15398 | ||
15399 | case T_V2SF: | |
15400 | case T_V4SF: | |
15401 | eltype = const_float_pointer_node; | |
15402 | break; | |
15403 | ||
15404 | case T_DI: | |
15405 | case T_V2DI: | |
15406 | eltype = const_intDI_pointer_node; | |
15407 | break; | |
15408 | ||
15409 | default: gcc_unreachable (); | |
15410 | } | |
15411 | } | |
15412 | else if (is_store && k == 0) | |
15413 | { | |
15414 | /* Similarly, Neon store patterns use operand 0 as | |
15415 | the memory location to store to (a SImode pointer). | |
15416 | Use a pointer to the element type of the store in | |
15417 | that position. */ | |
15418 | gcc_assert (insn_data[icode].operand[k].mode == SImode); | |
15419 | ||
15420 | switch (1 << j) | |
15421 | { | |
15422 | case T_V8QI: | |
15423 | case T_V16QI: | |
15424 | eltype = intQI_pointer_node; | |
15425 | break; | |
15426 | ||
15427 | case T_V4HI: | |
15428 | case T_V8HI: | |
15429 | eltype = intHI_pointer_node; | |
15430 | break; | |
15431 | ||
15432 | case T_V2SI: | |
15433 | case T_V4SI: | |
15434 | eltype = intSI_pointer_node; | |
15435 | break; | |
15436 | ||
15437 | case T_V2SF: | |
15438 | case T_V4SF: | |
15439 | eltype = float_pointer_node; | |
15440 | break; | |
15441 | ||
15442 | case T_DI: | |
15443 | case T_V2DI: | |
15444 | eltype = intDI_pointer_node; | |
15445 | break; | |
15446 | ||
15447 | default: gcc_unreachable (); | |
15448 | } | |
15449 | } | |
15450 | else | |
15451 | { | |
15452 | switch (insn_data[icode].operand[k].mode) | |
15453 | { | |
15454 | case VOIDmode: eltype = void_type_node; break; | |
15455 | /* Scalars. */ | |
15456 | case QImode: eltype = neon_intQI_type_node; break; | |
15457 | case HImode: eltype = neon_intHI_type_node; break; | |
15458 | case SImode: eltype = neon_intSI_type_node; break; | |
15459 | case SFmode: eltype = neon_float_type_node; break; | |
15460 | case DImode: eltype = neon_intDI_type_node; break; | |
15461 | case TImode: eltype = intTI_type_node; break; | |
15462 | case EImode: eltype = intEI_type_node; break; | |
15463 | case OImode: eltype = intOI_type_node; break; | |
15464 | case CImode: eltype = intCI_type_node; break; | |
15465 | case XImode: eltype = intXI_type_node; break; | |
15466 | /* 64-bit vectors. */ | |
15467 | case V8QImode: eltype = V8QI_type_node; break; | |
15468 | case V4HImode: eltype = V4HI_type_node; break; | |
15469 | case V2SImode: eltype = V2SI_type_node; break; | |
15470 | case V2SFmode: eltype = V2SF_type_node; break; | |
15471 | /* 128-bit vectors. */ | |
15472 | case V16QImode: eltype = V16QI_type_node; break; | |
15473 | case V8HImode: eltype = V8HI_type_node; break; | |
15474 | case V4SImode: eltype = V4SI_type_node; break; | |
15475 | case V4SFmode: eltype = V4SF_type_node; break; | |
15476 | case V2DImode: eltype = V2DI_type_node; break; | |
15477 | default: gcc_unreachable (); | |
15478 | } | |
15479 | } | |
15480 | ||
15481 | if (k == 0 && !is_store) | |
15482 | return_type = eltype; | |
15483 | else | |
15484 | args = tree_cons (NULL_TREE, eltype, args); | |
15485 | } | |
15486 | ||
15487 | ftype = build_function_type (return_type, args); | |
15488 | } | |
15489 | break; | |
15490 | ||
15491 | case NEON_RESULTPAIR: | |
15492 | { | |
15493 | switch (insn_data[icode].operand[1].mode) | |
15494 | { | |
15495 | case V8QImode: ftype = void_ftype_pv8qi_v8qi_v8qi; break; | |
15496 | case V4HImode: ftype = void_ftype_pv4hi_v4hi_v4hi; break; | |
15497 | case V2SImode: ftype = void_ftype_pv2si_v2si_v2si; break; | |
15498 | case V2SFmode: ftype = void_ftype_pv2sf_v2sf_v2sf; break; | |
15499 | case DImode: ftype = void_ftype_pdi_di_di; break; | |
15500 | case V16QImode: ftype = void_ftype_pv16qi_v16qi_v16qi; break; | |
15501 | case V8HImode: ftype = void_ftype_pv8hi_v8hi_v8hi; break; | |
15502 | case V4SImode: ftype = void_ftype_pv4si_v4si_v4si; break; | |
15503 | case V4SFmode: ftype = void_ftype_pv4sf_v4sf_v4sf; break; | |
15504 | case V2DImode: ftype = void_ftype_pv2di_v2di_v2di; break; | |
15505 | default: gcc_unreachable (); | |
15506 | } | |
15507 | } | |
15508 | break; | |
15509 | ||
15510 | case NEON_REINTERP: | |
15511 | { | |
15512 | /* We iterate over 5 doubleword types, then 5 quadword | |
15513 | types. */ | |
15514 | int rhs = j % 5; | |
15515 | switch (insn_data[icode].operand[0].mode) | |
15516 | { | |
15517 | case V8QImode: ftype = reinterp_ftype_dreg[0][rhs]; break; | |
15518 | case V4HImode: ftype = reinterp_ftype_dreg[1][rhs]; break; | |
15519 | case V2SImode: ftype = reinterp_ftype_dreg[2][rhs]; break; | |
15520 | case V2SFmode: ftype = reinterp_ftype_dreg[3][rhs]; break; | |
15521 | case DImode: ftype = reinterp_ftype_dreg[4][rhs]; break; | |
15522 | case V16QImode: ftype = reinterp_ftype_qreg[0][rhs]; break; | |
15523 | case V8HImode: ftype = reinterp_ftype_qreg[1][rhs]; break; | |
15524 | case V4SImode: ftype = reinterp_ftype_qreg[2][rhs]; break; | |
15525 | case V4SFmode: ftype = reinterp_ftype_qreg[3][rhs]; break; | |
15526 | case V2DImode: ftype = reinterp_ftype_qreg[4][rhs]; break; | |
15527 | default: gcc_unreachable (); | |
15528 | } | |
15529 | } | |
15530 | break; | |
15531 | ||
15532 | default: | |
15533 | gcc_unreachable (); | |
15534 | } | |
15535 | ||
15536 | gcc_assert (ftype != NULL); | |
15537 | ||
15538 | sprintf (namebuf, "__builtin_neon_%s%s", d->name, modenames[j]); | |
15539 | ||
15540 | add_builtin_function (namebuf, ftype, fcode++, BUILT_IN_MD, NULL, | |
15541 | NULL_TREE); | |
15542 | } | |
15543 | } | |
15544 | } | |
15545 | ||
5a9335ef NC |
15546 | static void |
15547 | arm_init_builtins (void) | |
15548 | { | |
d3585b76 DJ |
15549 | arm_init_tls_builtins (); |
15550 | ||
5a9335ef NC |
15551 | if (TARGET_REALLY_IWMMXT) |
15552 | arm_init_iwmmxt_builtins (); | |
88f77cba JB |
15553 | |
15554 | if (TARGET_NEON) | |
15555 | arm_init_neon_builtins (); | |
5a9335ef NC |
15556 | } |
15557 | ||
15558 | /* Errors in the source file can cause expand_expr to return const0_rtx | |
15559 | where we expect a vector. To avoid crashing, use one of the vector | |
15560 | clear instructions. */ | |
15561 | ||
15562 | static rtx | |
15563 | safe_vector_operand (rtx x, enum machine_mode mode) | |
15564 | { | |
15565 | if (x != const0_rtx) | |
15566 | return x; | |
15567 | x = gen_reg_rtx (mode); | |
15568 | ||
15569 | emit_insn (gen_iwmmxt_clrdi (mode == DImode ? x | |
15570 | : gen_rtx_SUBREG (DImode, x, 0))); | |
15571 | return x; | |
15572 | } | |
15573 | ||
15574 | /* Subroutine of arm_expand_builtin to take care of binop insns. */ | |
15575 | ||
15576 | static rtx | |
15577 | arm_expand_binop_builtin (enum insn_code icode, | |
5039610b | 15578 | tree exp, rtx target) |
5a9335ef NC |
15579 | { |
15580 | rtx pat; | |
5039610b SL |
15581 | tree arg0 = CALL_EXPR_ARG (exp, 0); |
15582 | tree arg1 = CALL_EXPR_ARG (exp, 1); | |
84217346 MD |
15583 | rtx op0 = expand_normal (arg0); |
15584 | rtx op1 = expand_normal (arg1); | |
5a9335ef NC |
15585 | enum machine_mode tmode = insn_data[icode].operand[0].mode; |
15586 | enum machine_mode mode0 = insn_data[icode].operand[1].mode; | |
15587 | enum machine_mode mode1 = insn_data[icode].operand[2].mode; | |
15588 | ||
15589 | if (VECTOR_MODE_P (mode0)) | |
15590 | op0 = safe_vector_operand (op0, mode0); | |
15591 | if (VECTOR_MODE_P (mode1)) | |
15592 | op1 = safe_vector_operand (op1, mode1); | |
15593 | ||
15594 | if (! target | |
15595 | || GET_MODE (target) != tmode | |
15596 | || ! (*insn_data[icode].operand[0].predicate) (target, tmode)) | |
15597 | target = gen_reg_rtx (tmode); | |
15598 | ||
e6d29d15 | 15599 | gcc_assert (GET_MODE (op0) == mode0 && GET_MODE (op1) == mode1); |
5a9335ef NC |
15600 | |
15601 | if (! (*insn_data[icode].operand[1].predicate) (op0, mode0)) | |
15602 | op0 = copy_to_mode_reg (mode0, op0); | |
15603 | if (! (*insn_data[icode].operand[2].predicate) (op1, mode1)) | |
15604 | op1 = copy_to_mode_reg (mode1, op1); | |
15605 | ||
15606 | pat = GEN_FCN (icode) (target, op0, op1); | |
15607 | if (! pat) | |
15608 | return 0; | |
15609 | emit_insn (pat); | |
15610 | return target; | |
15611 | } | |
15612 | ||
15613 | /* Subroutine of arm_expand_builtin to take care of unop insns. */ | |
15614 | ||
15615 | static rtx | |
15616 | arm_expand_unop_builtin (enum insn_code icode, | |
5039610b | 15617 | tree exp, rtx target, int do_load) |
5a9335ef NC |
15618 | { |
15619 | rtx pat; | |
5039610b | 15620 | tree arg0 = CALL_EXPR_ARG (exp, 0); |
84217346 | 15621 | rtx op0 = expand_normal (arg0); |
5a9335ef NC |
15622 | enum machine_mode tmode = insn_data[icode].operand[0].mode; |
15623 | enum machine_mode mode0 = insn_data[icode].operand[1].mode; | |
15624 | ||
15625 | if (! target | |
15626 | || GET_MODE (target) != tmode | |
15627 | || ! (*insn_data[icode].operand[0].predicate) (target, tmode)) | |
15628 | target = gen_reg_rtx (tmode); | |
15629 | if (do_load) | |
15630 | op0 = gen_rtx_MEM (mode0, copy_to_mode_reg (Pmode, op0)); | |
15631 | else | |
15632 | { | |
15633 | if (VECTOR_MODE_P (mode0)) | |
15634 | op0 = safe_vector_operand (op0, mode0); | |
15635 | ||
15636 | if (! (*insn_data[icode].operand[1].predicate) (op0, mode0)) | |
15637 | op0 = copy_to_mode_reg (mode0, op0); | |
15638 | } | |
15639 | ||
15640 | pat = GEN_FCN (icode) (target, op0); | |
15641 | if (! pat) | |
15642 | return 0; | |
15643 | emit_insn (pat); | |
15644 | return target; | |
15645 | } | |
15646 | ||
88f77cba JB |
15647 | static int |
15648 | neon_builtin_compare (const void *a, const void *b) | |
15649 | { | |
5ead67f6 KG |
15650 | const neon_builtin_datum *const key = (const neon_builtin_datum *) a; |
15651 | const neon_builtin_datum *const memb = (const neon_builtin_datum *) b; | |
88f77cba JB |
15652 | unsigned int soughtcode = key->base_fcode; |
15653 | ||
15654 | if (soughtcode >= memb->base_fcode | |
15655 | && soughtcode < memb->base_fcode + memb->num_vars) | |
15656 | return 0; | |
15657 | else if (soughtcode < memb->base_fcode) | |
15658 | return -1; | |
15659 | else | |
15660 | return 1; | |
15661 | } | |
15662 | ||
15663 | static enum insn_code | |
15664 | locate_neon_builtin_icode (int fcode, neon_itype *itype) | |
15665 | { | |
15666 | neon_builtin_datum key, *found; | |
15667 | int idx; | |
15668 | ||
15669 | key.base_fcode = fcode; | |
5ead67f6 KG |
15670 | found = (neon_builtin_datum *) |
15671 | bsearch (&key, &neon_builtin_data[0], ARRAY_SIZE (neon_builtin_data), | |
88f77cba JB |
15672 | sizeof (neon_builtin_data[0]), neon_builtin_compare); |
15673 | gcc_assert (found); | |
15674 | idx = fcode - (int) found->base_fcode; | |
15675 | gcc_assert (idx >= 0 && idx < T_MAX && idx < (int)found->num_vars); | |
15676 | ||
15677 | if (itype) | |
15678 | *itype = found->itype; | |
15679 | ||
15680 | return found->codes[idx]; | |
15681 | } | |
15682 | ||
15683 | typedef enum { | |
15684 | NEON_ARG_COPY_TO_REG, | |
15685 | NEON_ARG_CONSTANT, | |
15686 | NEON_ARG_STOP | |
15687 | } builtin_arg; | |
15688 | ||
15689 | #define NEON_MAX_BUILTIN_ARGS 5 | |
15690 | ||
15691 | /* Expand a Neon builtin. */ | |
15692 | static rtx | |
15693 | arm_expand_neon_args (rtx target, int icode, int have_retval, | |
15694 | tree exp, ...) | |
15695 | { | |
15696 | va_list ap; | |
15697 | rtx pat; | |
15698 | tree arg[NEON_MAX_BUILTIN_ARGS]; | |
15699 | rtx op[NEON_MAX_BUILTIN_ARGS]; | |
15700 | enum machine_mode tmode = insn_data[icode].operand[0].mode; | |
15701 | enum machine_mode mode[NEON_MAX_BUILTIN_ARGS]; | |
15702 | int argc = 0; | |
15703 | ||
15704 | if (have_retval | |
15705 | && (!target | |
15706 | || GET_MODE (target) != tmode | |
15707 | || !(*insn_data[icode].operand[0].predicate) (target, tmode))) | |
15708 | target = gen_reg_rtx (tmode); | |
15709 | ||
15710 | va_start (ap, exp); | |
15711 | ||
15712 | for (;;) | |
15713 | { | |
15714 | builtin_arg thisarg = va_arg (ap, int); | |
15715 | ||
15716 | if (thisarg == NEON_ARG_STOP) | |
15717 | break; | |
15718 | else | |
15719 | { | |
15720 | arg[argc] = CALL_EXPR_ARG (exp, argc); | |
15721 | op[argc] = expand_normal (arg[argc]); | |
15722 | mode[argc] = insn_data[icode].operand[argc + have_retval].mode; | |
15723 | ||
15724 | switch (thisarg) | |
15725 | { | |
15726 | case NEON_ARG_COPY_TO_REG: | |
15727 | /*gcc_assert (GET_MODE (op[argc]) == mode[argc]);*/ | |
15728 | if (!(*insn_data[icode].operand[argc + have_retval].predicate) | |
15729 | (op[argc], mode[argc])) | |
15730 | op[argc] = copy_to_mode_reg (mode[argc], op[argc]); | |
15731 | break; | |
15732 | ||
15733 | case NEON_ARG_CONSTANT: | |
15734 | /* FIXME: This error message is somewhat unhelpful. */ | |
15735 | if (!(*insn_data[icode].operand[argc + have_retval].predicate) | |
15736 | (op[argc], mode[argc])) | |
15737 | error ("argument must be a constant"); | |
15738 | break; | |
15739 | ||
15740 | case NEON_ARG_STOP: | |
15741 | gcc_unreachable (); | |
15742 | } | |
15743 | ||
15744 | argc++; | |
15745 | } | |
15746 | } | |
15747 | ||
15748 | va_end (ap); | |
15749 | ||
15750 | if (have_retval) | |
15751 | switch (argc) | |
15752 | { | |
15753 | case 1: | |
15754 | pat = GEN_FCN (icode) (target, op[0]); | |
15755 | break; | |
15756 | ||
15757 | case 2: | |
15758 | pat = GEN_FCN (icode) (target, op[0], op[1]); | |
15759 | break; | |
15760 | ||
15761 | case 3: | |
15762 | pat = GEN_FCN (icode) (target, op[0], op[1], op[2]); | |
15763 | break; | |
15764 | ||
15765 | case 4: | |
15766 | pat = GEN_FCN (icode) (target, op[0], op[1], op[2], op[3]); | |
15767 | break; | |
15768 | ||
15769 | case 5: | |
15770 | pat = GEN_FCN (icode) (target, op[0], op[1], op[2], op[3], op[4]); | |
15771 | break; | |
15772 | ||
15773 | default: | |
15774 | gcc_unreachable (); | |
15775 | } | |
15776 | else | |
15777 | switch (argc) | |
15778 | { | |
15779 | case 1: | |
15780 | pat = GEN_FCN (icode) (op[0]); | |
15781 | break; | |
15782 | ||
15783 | case 2: | |
15784 | pat = GEN_FCN (icode) (op[0], op[1]); | |
15785 | break; | |
15786 | ||
15787 | case 3: | |
15788 | pat = GEN_FCN (icode) (op[0], op[1], op[2]); | |
15789 | break; | |
15790 | ||
15791 | case 4: | |
15792 | pat = GEN_FCN (icode) (op[0], op[1], op[2], op[3]); | |
15793 | break; | |
15794 | ||
15795 | case 5: | |
15796 | pat = GEN_FCN (icode) (op[0], op[1], op[2], op[3], op[4]); | |
15797 | break; | |
15798 | ||
15799 | default: | |
15800 | gcc_unreachable (); | |
15801 | } | |
15802 | ||
15803 | if (!pat) | |
15804 | return 0; | |
15805 | ||
15806 | emit_insn (pat); | |
15807 | ||
15808 | return target; | |
15809 | } | |
15810 | ||
15811 | /* Expand a Neon builtin. These are "special" because they don't have symbolic | |
15812 | constants defined per-instruction or per instruction-variant. Instead, the | |
15813 | required info is looked up in the table neon_builtin_data. */ | |
15814 | static rtx | |
15815 | arm_expand_neon_builtin (int fcode, tree exp, rtx target) | |
15816 | { | |
15817 | neon_itype itype; | |
15818 | enum insn_code icode = locate_neon_builtin_icode (fcode, &itype); | |
15819 | ||
15820 | switch (itype) | |
15821 | { | |
15822 | case NEON_UNOP: | |
15823 | case NEON_CONVERT: | |
15824 | case NEON_DUPLANE: | |
15825 | return arm_expand_neon_args (target, icode, 1, exp, | |
15826 | NEON_ARG_COPY_TO_REG, NEON_ARG_CONSTANT, NEON_ARG_STOP); | |
15827 | ||
15828 | case NEON_BINOP: | |
15829 | case NEON_SETLANE: | |
15830 | case NEON_SCALARMUL: | |
15831 | case NEON_SCALARMULL: | |
15832 | case NEON_SCALARMULH: | |
15833 | case NEON_SHIFTINSERT: | |
15834 | case NEON_LOGICBINOP: | |
15835 | return arm_expand_neon_args (target, icode, 1, exp, | |
15836 | NEON_ARG_COPY_TO_REG, NEON_ARG_COPY_TO_REG, NEON_ARG_CONSTANT, | |
15837 | NEON_ARG_STOP); | |
15838 | ||
15839 | case NEON_TERNOP: | |
15840 | return arm_expand_neon_args (target, icode, 1, exp, | |
15841 | NEON_ARG_COPY_TO_REG, NEON_ARG_COPY_TO_REG, NEON_ARG_COPY_TO_REG, | |
15842 | NEON_ARG_CONSTANT, NEON_ARG_STOP); | |
15843 | ||
15844 | case NEON_GETLANE: | |
15845 | case NEON_FIXCONV: | |
15846 | case NEON_SHIFTIMM: | |
15847 | return arm_expand_neon_args (target, icode, 1, exp, | |
15848 | NEON_ARG_COPY_TO_REG, NEON_ARG_CONSTANT, NEON_ARG_CONSTANT, | |
15849 | NEON_ARG_STOP); | |
15850 | ||
15851 | case NEON_CREATE: | |
15852 | return arm_expand_neon_args (target, icode, 1, exp, | |
15853 | NEON_ARG_COPY_TO_REG, NEON_ARG_STOP); | |
15854 | ||
15855 | case NEON_DUP: | |
15856 | case NEON_SPLIT: | |
15857 | case NEON_REINTERP: | |
15858 | return arm_expand_neon_args (target, icode, 1, exp, | |
15859 | NEON_ARG_COPY_TO_REG, NEON_ARG_STOP); | |
15860 | ||
15861 | case NEON_COMBINE: | |
15862 | case NEON_VTBL: | |
15863 | return arm_expand_neon_args (target, icode, 1, exp, | |
15864 | NEON_ARG_COPY_TO_REG, NEON_ARG_COPY_TO_REG, NEON_ARG_STOP); | |
15865 | ||
15866 | case NEON_RESULTPAIR: | |
15867 | return arm_expand_neon_args (target, icode, 0, exp, | |
15868 | NEON_ARG_COPY_TO_REG, NEON_ARG_COPY_TO_REG, NEON_ARG_COPY_TO_REG, | |
15869 | NEON_ARG_STOP); | |
15870 | ||
15871 | case NEON_LANEMUL: | |
15872 | case NEON_LANEMULL: | |
15873 | case NEON_LANEMULH: | |
15874 | return arm_expand_neon_args (target, icode, 1, exp, | |
15875 | NEON_ARG_COPY_TO_REG, NEON_ARG_COPY_TO_REG, NEON_ARG_CONSTANT, | |
15876 | NEON_ARG_CONSTANT, NEON_ARG_STOP); | |
15877 | ||
15878 | case NEON_LANEMAC: | |
15879 | return arm_expand_neon_args (target, icode, 1, exp, | |
15880 | NEON_ARG_COPY_TO_REG, NEON_ARG_COPY_TO_REG, NEON_ARG_COPY_TO_REG, | |
15881 | NEON_ARG_CONSTANT, NEON_ARG_CONSTANT, NEON_ARG_STOP); | |
15882 | ||
15883 | case NEON_SHIFTACC: | |
15884 | return arm_expand_neon_args (target, icode, 1, exp, | |
15885 | NEON_ARG_COPY_TO_REG, NEON_ARG_COPY_TO_REG, NEON_ARG_CONSTANT, | |
15886 | NEON_ARG_CONSTANT, NEON_ARG_STOP); | |
15887 | ||
15888 | case NEON_SCALARMAC: | |
15889 | return arm_expand_neon_args (target, icode, 1, exp, | |
15890 | NEON_ARG_COPY_TO_REG, NEON_ARG_COPY_TO_REG, NEON_ARG_COPY_TO_REG, | |
15891 | NEON_ARG_CONSTANT, NEON_ARG_STOP); | |
15892 | ||
15893 | case NEON_SELECT: | |
15894 | case NEON_VTBX: | |
15895 | return arm_expand_neon_args (target, icode, 1, exp, | |
15896 | NEON_ARG_COPY_TO_REG, NEON_ARG_COPY_TO_REG, NEON_ARG_COPY_TO_REG, | |
15897 | NEON_ARG_STOP); | |
15898 | ||
15899 | case NEON_LOAD1: | |
15900 | case NEON_LOADSTRUCT: | |
15901 | return arm_expand_neon_args (target, icode, 1, exp, | |
15902 | NEON_ARG_COPY_TO_REG, NEON_ARG_STOP); | |
15903 | ||
15904 | case NEON_LOAD1LANE: | |
15905 | case NEON_LOADSTRUCTLANE: | |
15906 | return arm_expand_neon_args (target, icode, 1, exp, | |
15907 | NEON_ARG_COPY_TO_REG, NEON_ARG_COPY_TO_REG, NEON_ARG_CONSTANT, | |
15908 | NEON_ARG_STOP); | |
15909 | ||
15910 | case NEON_STORE1: | |
15911 | case NEON_STORESTRUCT: | |
15912 | return arm_expand_neon_args (target, icode, 0, exp, | |
15913 | NEON_ARG_COPY_TO_REG, NEON_ARG_COPY_TO_REG, NEON_ARG_STOP); | |
15914 | ||
15915 | case NEON_STORE1LANE: | |
15916 | case NEON_STORESTRUCTLANE: | |
15917 | return arm_expand_neon_args (target, icode, 0, exp, | |
15918 | NEON_ARG_COPY_TO_REG, NEON_ARG_COPY_TO_REG, NEON_ARG_CONSTANT, | |
15919 | NEON_ARG_STOP); | |
15920 | } | |
15921 | ||
15922 | gcc_unreachable (); | |
15923 | } | |
15924 | ||
15925 | /* Emit code to reinterpret one Neon type as another, without altering bits. */ | |
15926 | void | |
15927 | neon_reinterpret (rtx dest, rtx src) | |
15928 | { | |
15929 | emit_move_insn (dest, gen_lowpart (GET_MODE (dest), src)); | |
15930 | } | |
15931 | ||
15932 | /* Emit code to place a Neon pair result in memory locations (with equal | |
15933 | registers). */ | |
15934 | void | |
15935 | neon_emit_pair_result_insn (enum machine_mode mode, | |
15936 | rtx (*intfn) (rtx, rtx, rtx, rtx), rtx destaddr, | |
15937 | rtx op1, rtx op2) | |
15938 | { | |
15939 | rtx mem = gen_rtx_MEM (mode, destaddr); | |
15940 | rtx tmp1 = gen_reg_rtx (mode); | |
15941 | rtx tmp2 = gen_reg_rtx (mode); | |
15942 | ||
15943 | emit_insn (intfn (tmp1, op1, tmp2, op2)); | |
15944 | ||
15945 | emit_move_insn (mem, tmp1); | |
15946 | mem = adjust_address (mem, mode, GET_MODE_SIZE (mode)); | |
15947 | emit_move_insn (mem, tmp2); | |
15948 | } | |
15949 | ||
15950 | /* Set up operands for a register copy from src to dest, taking care not to | |
15951 | clobber registers in the process. | |
15952 | FIXME: This has rather high polynomial complexity (O(n^3)?) but shouldn't | |
15953 | be called with a large N, so that should be OK. */ | |
15954 | ||
15955 | void | |
15956 | neon_disambiguate_copy (rtx *operands, rtx *dest, rtx *src, unsigned int count) | |
15957 | { | |
15958 | unsigned int copied = 0, opctr = 0; | |
15959 | unsigned int done = (1 << count) - 1; | |
15960 | unsigned int i, j; | |
15961 | ||
15962 | while (copied != done) | |
15963 | { | |
15964 | for (i = 0; i < count; i++) | |
15965 | { | |
15966 | int good = 1; | |
15967 | ||
15968 | for (j = 0; good && j < count; j++) | |
15969 | if (i != j && (copied & (1 << j)) == 0 | |
15970 | && reg_overlap_mentioned_p (src[j], dest[i])) | |
15971 | good = 0; | |
15972 | ||
15973 | if (good) | |
15974 | { | |
15975 | operands[opctr++] = dest[i]; | |
15976 | operands[opctr++] = src[i]; | |
15977 | copied |= 1 << i; | |
15978 | } | |
15979 | } | |
15980 | } | |
15981 | ||
15982 | gcc_assert (opctr == count * 2); | |
15983 | } | |
15984 | ||
5a9335ef NC |
15985 | /* Expand an expression EXP that calls a built-in function, |
15986 | with result going to TARGET if that's convenient | |
15987 | (and in mode MODE if that's convenient). | |
15988 | SUBTARGET may be used as the target for computing one of EXP's operands. | |
15989 | IGNORE is nonzero if the value is to be ignored. */ | |
15990 | ||
15991 | static rtx | |
15992 | arm_expand_builtin (tree exp, | |
15993 | rtx target, | |
15994 | rtx subtarget ATTRIBUTE_UNUSED, | |
15995 | enum machine_mode mode ATTRIBUTE_UNUSED, | |
15996 | int ignore ATTRIBUTE_UNUSED) | |
15997 | { | |
15998 | const struct builtin_description * d; | |
15999 | enum insn_code icode; | |
5039610b | 16000 | tree fndecl = TREE_OPERAND (CALL_EXPR_FN (exp), 0); |
5a9335ef NC |
16001 | tree arg0; |
16002 | tree arg1; | |
16003 | tree arg2; | |
16004 | rtx op0; | |
16005 | rtx op1; | |
16006 | rtx op2; | |
16007 | rtx pat; | |
16008 | int fcode = DECL_FUNCTION_CODE (fndecl); | |
16009 | size_t i; | |
16010 | enum machine_mode tmode; | |
16011 | enum machine_mode mode0; | |
16012 | enum machine_mode mode1; | |
16013 | enum machine_mode mode2; | |
16014 | ||
88f77cba JB |
16015 | if (fcode >= ARM_BUILTIN_NEON_BASE) |
16016 | return arm_expand_neon_builtin (fcode, exp, target); | |
16017 | ||
5a9335ef NC |
16018 | switch (fcode) |
16019 | { | |
16020 | case ARM_BUILTIN_TEXTRMSB: | |
16021 | case ARM_BUILTIN_TEXTRMUB: | |
16022 | case ARM_BUILTIN_TEXTRMSH: | |
16023 | case ARM_BUILTIN_TEXTRMUH: | |
16024 | case ARM_BUILTIN_TEXTRMSW: | |
16025 | case ARM_BUILTIN_TEXTRMUW: | |
16026 | icode = (fcode == ARM_BUILTIN_TEXTRMSB ? CODE_FOR_iwmmxt_textrmsb | |
16027 | : fcode == ARM_BUILTIN_TEXTRMUB ? CODE_FOR_iwmmxt_textrmub | |
16028 | : fcode == ARM_BUILTIN_TEXTRMSH ? CODE_FOR_iwmmxt_textrmsh | |
16029 | : fcode == ARM_BUILTIN_TEXTRMUH ? CODE_FOR_iwmmxt_textrmuh | |
16030 | : CODE_FOR_iwmmxt_textrmw); | |
16031 | ||
5039610b SL |
16032 | arg0 = CALL_EXPR_ARG (exp, 0); |
16033 | arg1 = CALL_EXPR_ARG (exp, 1); | |
84217346 MD |
16034 | op0 = expand_normal (arg0); |
16035 | op1 = expand_normal (arg1); | |
5a9335ef NC |
16036 | tmode = insn_data[icode].operand[0].mode; |
16037 | mode0 = insn_data[icode].operand[1].mode; | |
16038 | mode1 = insn_data[icode].operand[2].mode; | |
16039 | ||
16040 | if (! (*insn_data[icode].operand[1].predicate) (op0, mode0)) | |
16041 | op0 = copy_to_mode_reg (mode0, op0); | |
16042 | if (! (*insn_data[icode].operand[2].predicate) (op1, mode1)) | |
16043 | { | |
16044 | /* @@@ better error message */ | |
16045 | error ("selector must be an immediate"); | |
16046 | return gen_reg_rtx (tmode); | |
16047 | } | |
16048 | if (target == 0 | |
16049 | || GET_MODE (target) != tmode | |
16050 | || ! (*insn_data[icode].operand[0].predicate) (target, tmode)) | |
16051 | target = gen_reg_rtx (tmode); | |
16052 | pat = GEN_FCN (icode) (target, op0, op1); | |
16053 | if (! pat) | |
16054 | return 0; | |
16055 | emit_insn (pat); | |
16056 | return target; | |
16057 | ||
16058 | case ARM_BUILTIN_TINSRB: | |
16059 | case ARM_BUILTIN_TINSRH: | |
16060 | case ARM_BUILTIN_TINSRW: | |
16061 | icode = (fcode == ARM_BUILTIN_TINSRB ? CODE_FOR_iwmmxt_tinsrb | |
16062 | : fcode == ARM_BUILTIN_TINSRH ? CODE_FOR_iwmmxt_tinsrh | |
16063 | : CODE_FOR_iwmmxt_tinsrw); | |
5039610b SL |
16064 | arg0 = CALL_EXPR_ARG (exp, 0); |
16065 | arg1 = CALL_EXPR_ARG (exp, 1); | |
16066 | arg2 = CALL_EXPR_ARG (exp, 2); | |
84217346 MD |
16067 | op0 = expand_normal (arg0); |
16068 | op1 = expand_normal (arg1); | |
16069 | op2 = expand_normal (arg2); | |
5a9335ef NC |
16070 | tmode = insn_data[icode].operand[0].mode; |
16071 | mode0 = insn_data[icode].operand[1].mode; | |
16072 | mode1 = insn_data[icode].operand[2].mode; | |
16073 | mode2 = insn_data[icode].operand[3].mode; | |
16074 | ||
16075 | if (! (*insn_data[icode].operand[1].predicate) (op0, mode0)) | |
16076 | op0 = copy_to_mode_reg (mode0, op0); | |
16077 | if (! (*insn_data[icode].operand[2].predicate) (op1, mode1)) | |
16078 | op1 = copy_to_mode_reg (mode1, op1); | |
16079 | if (! (*insn_data[icode].operand[3].predicate) (op2, mode2)) | |
16080 | { | |
16081 | /* @@@ better error message */ | |
16082 | error ("selector must be an immediate"); | |
16083 | return const0_rtx; | |
16084 | } | |
16085 | if (target == 0 | |
16086 | || GET_MODE (target) != tmode | |
16087 | || ! (*insn_data[icode].operand[0].predicate) (target, tmode)) | |
16088 | target = gen_reg_rtx (tmode); | |
16089 | pat = GEN_FCN (icode) (target, op0, op1, op2); | |
16090 | if (! pat) | |
16091 | return 0; | |
16092 | emit_insn (pat); | |
16093 | return target; | |
16094 | ||
16095 | case ARM_BUILTIN_SETWCX: | |
5039610b SL |
16096 | arg0 = CALL_EXPR_ARG (exp, 0); |
16097 | arg1 = CALL_EXPR_ARG (exp, 1); | |
84217346 MD |
16098 | op0 = force_reg (SImode, expand_normal (arg0)); |
16099 | op1 = expand_normal (arg1); | |
f07a6b21 | 16100 | emit_insn (gen_iwmmxt_tmcr (op1, op0)); |
5a9335ef NC |
16101 | return 0; |
16102 | ||
16103 | case ARM_BUILTIN_GETWCX: | |
5039610b | 16104 | arg0 = CALL_EXPR_ARG (exp, 0); |
84217346 | 16105 | op0 = expand_normal (arg0); |
5a9335ef NC |
16106 | target = gen_reg_rtx (SImode); |
16107 | emit_insn (gen_iwmmxt_tmrc (target, op0)); | |
16108 | return target; | |
16109 | ||
16110 | case ARM_BUILTIN_WSHUFH: | |
16111 | icode = CODE_FOR_iwmmxt_wshufh; | |
5039610b SL |
16112 | arg0 = CALL_EXPR_ARG (exp, 0); |
16113 | arg1 = CALL_EXPR_ARG (exp, 1); | |
84217346 MD |
16114 | op0 = expand_normal (arg0); |
16115 | op1 = expand_normal (arg1); | |
5a9335ef NC |
16116 | tmode = insn_data[icode].operand[0].mode; |
16117 | mode1 = insn_data[icode].operand[1].mode; | |
16118 | mode2 = insn_data[icode].operand[2].mode; | |
16119 | ||
16120 | if (! (*insn_data[icode].operand[1].predicate) (op0, mode1)) | |
16121 | op0 = copy_to_mode_reg (mode1, op0); | |
16122 | if (! (*insn_data[icode].operand[2].predicate) (op1, mode2)) | |
16123 | { | |
16124 | /* @@@ better error message */ | |
16125 | error ("mask must be an immediate"); | |
16126 | return const0_rtx; | |
16127 | } | |
16128 | if (target == 0 | |
16129 | || GET_MODE (target) != tmode | |
16130 | || ! (*insn_data[icode].operand[0].predicate) (target, tmode)) | |
16131 | target = gen_reg_rtx (tmode); | |
16132 | pat = GEN_FCN (icode) (target, op0, op1); | |
16133 | if (! pat) | |
16134 | return 0; | |
16135 | emit_insn (pat); | |
16136 | return target; | |
16137 | ||
16138 | case ARM_BUILTIN_WSADB: | |
5039610b | 16139 | return arm_expand_binop_builtin (CODE_FOR_iwmmxt_wsadb, exp, target); |
5a9335ef | 16140 | case ARM_BUILTIN_WSADH: |
5039610b | 16141 | return arm_expand_binop_builtin (CODE_FOR_iwmmxt_wsadh, exp, target); |
5a9335ef | 16142 | case ARM_BUILTIN_WSADBZ: |
5039610b | 16143 | return arm_expand_binop_builtin (CODE_FOR_iwmmxt_wsadbz, exp, target); |
5a9335ef | 16144 | case ARM_BUILTIN_WSADHZ: |
5039610b | 16145 | return arm_expand_binop_builtin (CODE_FOR_iwmmxt_wsadhz, exp, target); |
5a9335ef NC |
16146 | |
16147 | /* Several three-argument builtins. */ | |
16148 | case ARM_BUILTIN_WMACS: | |
16149 | case ARM_BUILTIN_WMACU: | |
16150 | case ARM_BUILTIN_WALIGN: | |
16151 | case ARM_BUILTIN_TMIA: | |
16152 | case ARM_BUILTIN_TMIAPH: | |
16153 | case ARM_BUILTIN_TMIATT: | |
16154 | case ARM_BUILTIN_TMIATB: | |
16155 | case ARM_BUILTIN_TMIABT: | |
16156 | case ARM_BUILTIN_TMIABB: | |
16157 | icode = (fcode == ARM_BUILTIN_WMACS ? CODE_FOR_iwmmxt_wmacs | |
16158 | : fcode == ARM_BUILTIN_WMACU ? CODE_FOR_iwmmxt_wmacu | |
16159 | : fcode == ARM_BUILTIN_TMIA ? CODE_FOR_iwmmxt_tmia | |
16160 | : fcode == ARM_BUILTIN_TMIAPH ? CODE_FOR_iwmmxt_tmiaph | |
16161 | : fcode == ARM_BUILTIN_TMIABB ? CODE_FOR_iwmmxt_tmiabb | |
16162 | : fcode == ARM_BUILTIN_TMIABT ? CODE_FOR_iwmmxt_tmiabt | |
16163 | : fcode == ARM_BUILTIN_TMIATB ? CODE_FOR_iwmmxt_tmiatb | |
16164 | : fcode == ARM_BUILTIN_TMIATT ? CODE_FOR_iwmmxt_tmiatt | |
16165 | : CODE_FOR_iwmmxt_walign); | |
5039610b SL |
16166 | arg0 = CALL_EXPR_ARG (exp, 0); |
16167 | arg1 = CALL_EXPR_ARG (exp, 1); | |
16168 | arg2 = CALL_EXPR_ARG (exp, 2); | |
84217346 MD |
16169 | op0 = expand_normal (arg0); |
16170 | op1 = expand_normal (arg1); | |
16171 | op2 = expand_normal (arg2); | |
5a9335ef NC |
16172 | tmode = insn_data[icode].operand[0].mode; |
16173 | mode0 = insn_data[icode].operand[1].mode; | |
16174 | mode1 = insn_data[icode].operand[2].mode; | |
16175 | mode2 = insn_data[icode].operand[3].mode; | |
16176 | ||
16177 | if (! (*insn_data[icode].operand[1].predicate) (op0, mode0)) | |
16178 | op0 = copy_to_mode_reg (mode0, op0); | |
16179 | if (! (*insn_data[icode].operand[2].predicate) (op1, mode1)) | |
16180 | op1 = copy_to_mode_reg (mode1, op1); | |
16181 | if (! (*insn_data[icode].operand[3].predicate) (op2, mode2)) | |
16182 | op2 = copy_to_mode_reg (mode2, op2); | |
16183 | if (target == 0 | |
16184 | || GET_MODE (target) != tmode | |
16185 | || ! (*insn_data[icode].operand[0].predicate) (target, tmode)) | |
16186 | target = gen_reg_rtx (tmode); | |
16187 | pat = GEN_FCN (icode) (target, op0, op1, op2); | |
16188 | if (! pat) | |
16189 | return 0; | |
16190 | emit_insn (pat); | |
16191 | return target; | |
f676971a | 16192 | |
5a9335ef NC |
16193 | case ARM_BUILTIN_WZERO: |
16194 | target = gen_reg_rtx (DImode); | |
16195 | emit_insn (gen_iwmmxt_clrdi (target)); | |
16196 | return target; | |
16197 | ||
d3585b76 DJ |
16198 | case ARM_BUILTIN_THREAD_POINTER: |
16199 | return arm_load_tp (target); | |
16200 | ||
5a9335ef NC |
16201 | default: |
16202 | break; | |
16203 | } | |
16204 | ||
e97a46ce | 16205 | for (i = 0, d = bdesc_2arg; i < ARRAY_SIZE (bdesc_2arg); i++, d++) |
5a9335ef | 16206 | if (d->code == (const enum arm_builtins) fcode) |
5039610b | 16207 | return arm_expand_binop_builtin (d->icode, exp, target); |
5a9335ef | 16208 | |
e97a46ce | 16209 | for (i = 0, d = bdesc_1arg; i < ARRAY_SIZE (bdesc_1arg); i++, d++) |
5a9335ef | 16210 | if (d->code == (const enum arm_builtins) fcode) |
5039610b | 16211 | return arm_expand_unop_builtin (d->icode, exp, target, 0); |
5a9335ef NC |
16212 | |
16213 | /* @@@ Should really do something sensible here. */ | |
16214 | return NULL_RTX; | |
16215 | } | |
16216 | \f | |
1d6e90ac NC |
16217 | /* Return the number (counting from 0) of |
16218 | the least significant set bit in MASK. */ | |
16219 | ||
e32bac5b | 16220 | inline static int |
b279b20a | 16221 | number_of_first_bit_set (unsigned mask) |
d5b7b3ae RE |
16222 | { |
16223 | int bit; | |
16224 | ||
16225 | for (bit = 0; | |
16226 | (mask & (1 << bit)) == 0; | |
5895f793 | 16227 | ++bit) |
d5b7b3ae RE |
16228 | continue; |
16229 | ||
16230 | return bit; | |
16231 | } | |
16232 | ||
b279b20a NC |
16233 | /* Emit code to push or pop registers to or from the stack. F is the |
16234 | assembly file. MASK is the registers to push or pop. PUSH is | |
16235 | nonzero if we should push, and zero if we should pop. For debugging | |
16236 | output, if pushing, adjust CFA_OFFSET by the amount of space added | |
16237 | to the stack. REAL_REGS should have the same number of bits set as | |
16238 | MASK, and will be used instead (in the same order) to describe which | |
16239 | registers were saved - this is used to mark the save slots when we | |
16240 | push high registers after moving them to low registers. */ | |
16241 | static void | |
16242 | thumb_pushpop (FILE *f, unsigned long mask, int push, int *cfa_offset, | |
16243 | unsigned long real_regs) | |
16244 | { | |
16245 | int regno; | |
16246 | int lo_mask = mask & 0xFF; | |
16247 | int pushed_words = 0; | |
16248 | ||
e6d29d15 | 16249 | gcc_assert (mask); |
b279b20a NC |
16250 | |
16251 | if (lo_mask == 0 && !push && (mask & (1 << PC_REGNUM))) | |
16252 | { | |
16253 | /* Special case. Do not generate a POP PC statement here, do it in | |
16254 | thumb_exit() */ | |
16255 | thumb_exit (f, -1); | |
16256 | return; | |
16257 | } | |
16258 | ||
617a1b71 PB |
16259 | if (ARM_EABI_UNWIND_TABLES && push) |
16260 | { | |
16261 | fprintf (f, "\t.save\t{"); | |
16262 | for (regno = 0; regno < 15; regno++) | |
16263 | { | |
16264 | if (real_regs & (1 << regno)) | |
16265 | { | |
16266 | if (real_regs & ((1 << regno) -1)) | |
16267 | fprintf (f, ", "); | |
16268 | asm_fprintf (f, "%r", regno); | |
16269 | } | |
16270 | } | |
16271 | fprintf (f, "}\n"); | |
16272 | } | |
16273 | ||
b279b20a NC |
16274 | fprintf (f, "\t%s\t{", push ? "push" : "pop"); |
16275 | ||
16276 | /* Look at the low registers first. */ | |
16277 | for (regno = 0; regno <= LAST_LO_REGNUM; regno++, lo_mask >>= 1) | |
16278 | { | |
16279 | if (lo_mask & 1) | |
16280 | { | |
16281 | asm_fprintf (f, "%r", regno); | |
16282 | ||
16283 | if ((lo_mask & ~1) != 0) | |
16284 | fprintf (f, ", "); | |
16285 | ||
16286 | pushed_words++; | |
16287 | } | |
16288 | } | |
16289 | ||
16290 | if (push && (mask & (1 << LR_REGNUM))) | |
16291 | { | |
16292 | /* Catch pushing the LR. */ | |
16293 | if (mask & 0xFF) | |
16294 | fprintf (f, ", "); | |
16295 | ||
16296 | asm_fprintf (f, "%r", LR_REGNUM); | |
16297 | ||
16298 | pushed_words++; | |
16299 | } | |
16300 | else if (!push && (mask & (1 << PC_REGNUM))) | |
16301 | { | |
16302 | /* Catch popping the PC. */ | |
16303 | if (TARGET_INTERWORK || TARGET_BACKTRACE | |
e3b5732b | 16304 | || crtl->calls_eh_return) |
b279b20a NC |
16305 | { |
16306 | /* The PC is never poped directly, instead | |
16307 | it is popped into r3 and then BX is used. */ | |
16308 | fprintf (f, "}\n"); | |
16309 | ||
16310 | thumb_exit (f, -1); | |
16311 | ||
16312 | return; | |
16313 | } | |
16314 | else | |
16315 | { | |
16316 | if (mask & 0xFF) | |
16317 | fprintf (f, ", "); | |
16318 | ||
16319 | asm_fprintf (f, "%r", PC_REGNUM); | |
16320 | } | |
16321 | } | |
16322 | ||
16323 | fprintf (f, "}\n"); | |
16324 | ||
16325 | if (push && pushed_words && dwarf2out_do_frame ()) | |
16326 | { | |
16327 | char *l = dwarf2out_cfi_label (); | |
16328 | int pushed_mask = real_regs; | |
16329 | ||
16330 | *cfa_offset += pushed_words * 4; | |
16331 | dwarf2out_def_cfa (l, SP_REGNUM, *cfa_offset); | |
16332 | ||
16333 | pushed_words = 0; | |
16334 | pushed_mask = real_regs; | |
16335 | for (regno = 0; regno <= 14; regno++, pushed_mask >>= 1) | |
16336 | { | |
16337 | if (pushed_mask & 1) | |
16338 | dwarf2out_reg_save (l, regno, 4 * pushed_words++ - *cfa_offset); | |
16339 | } | |
16340 | } | |
16341 | } | |
16342 | ||
d5b7b3ae RE |
16343 | /* Generate code to return from a thumb function. |
16344 | If 'reg_containing_return_addr' is -1, then the return address is | |
16345 | actually on the stack, at the stack pointer. */ | |
16346 | static void | |
c9ca9b88 | 16347 | thumb_exit (FILE *f, int reg_containing_return_addr) |
d5b7b3ae RE |
16348 | { |
16349 | unsigned regs_available_for_popping; | |
16350 | unsigned regs_to_pop; | |
16351 | int pops_needed; | |
16352 | unsigned available; | |
16353 | unsigned required; | |
16354 | int mode; | |
16355 | int size; | |
16356 | int restore_a4 = FALSE; | |
16357 | ||
16358 | /* Compute the registers we need to pop. */ | |
16359 | regs_to_pop = 0; | |
16360 | pops_needed = 0; | |
16361 | ||
c9ca9b88 | 16362 | if (reg_containing_return_addr == -1) |
d5b7b3ae | 16363 | { |
d5b7b3ae | 16364 | regs_to_pop |= 1 << LR_REGNUM; |
5895f793 | 16365 | ++pops_needed; |
d5b7b3ae RE |
16366 | } |
16367 | ||
16368 | if (TARGET_BACKTRACE) | |
16369 | { | |
16370 | /* Restore the (ARM) frame pointer and stack pointer. */ | |
16371 | regs_to_pop |= (1 << ARM_HARD_FRAME_POINTER_REGNUM) | (1 << SP_REGNUM); | |
16372 | pops_needed += 2; | |
16373 | } | |
16374 | ||
16375 | /* If there is nothing to pop then just emit the BX instruction and | |
16376 | return. */ | |
16377 | if (pops_needed == 0) | |
16378 | { | |
e3b5732b | 16379 | if (crtl->calls_eh_return) |
c9ca9b88 | 16380 | asm_fprintf (f, "\tadd\t%r, %r\n", SP_REGNUM, ARM_EH_STACKADJ_REGNUM); |
d5b7b3ae RE |
16381 | |
16382 | asm_fprintf (f, "\tbx\t%r\n", reg_containing_return_addr); | |
16383 | return; | |
16384 | } | |
16385 | /* Otherwise if we are not supporting interworking and we have not created | |
16386 | a backtrace structure and the function was not entered in ARM mode then | |
16387 | just pop the return address straight into the PC. */ | |
5895f793 RE |
16388 | else if (!TARGET_INTERWORK |
16389 | && !TARGET_BACKTRACE | |
c9ca9b88 | 16390 | && !is_called_in_ARM_mode (current_function_decl) |
e3b5732b | 16391 | && !crtl->calls_eh_return) |
d5b7b3ae | 16392 | { |
c9ca9b88 | 16393 | asm_fprintf (f, "\tpop\t{%r}\n", PC_REGNUM); |
d5b7b3ae RE |
16394 | return; |
16395 | } | |
16396 | ||
16397 | /* Find out how many of the (return) argument registers we can corrupt. */ | |
16398 | regs_available_for_popping = 0; | |
16399 | ||
16400 | /* If returning via __builtin_eh_return, the bottom three registers | |
16401 | all contain information needed for the return. */ | |
e3b5732b | 16402 | if (crtl->calls_eh_return) |
d5b7b3ae RE |
16403 | size = 12; |
16404 | else | |
16405 | { | |
d5b7b3ae RE |
16406 | /* If we can deduce the registers used from the function's |
16407 | return value. This is more reliable that examining | |
6fb5fa3c | 16408 | df_regs_ever_live_p () because that will be set if the register is |
d5b7b3ae RE |
16409 | ever used in the function, not just if the register is used |
16410 | to hold a return value. */ | |
16411 | ||
38173d38 JH |
16412 | if (crtl->return_rtx != 0) |
16413 | mode = GET_MODE (crtl->return_rtx); | |
d5b7b3ae | 16414 | else |
d5b7b3ae RE |
16415 | mode = DECL_MODE (DECL_RESULT (current_function_decl)); |
16416 | ||
16417 | size = GET_MODE_SIZE (mode); | |
16418 | ||
16419 | if (size == 0) | |
16420 | { | |
16421 | /* In a void function we can use any argument register. | |
16422 | In a function that returns a structure on the stack | |
16423 | we can use the second and third argument registers. */ | |
16424 | if (mode == VOIDmode) | |
16425 | regs_available_for_popping = | |
16426 | (1 << ARG_REGISTER (1)) | |
16427 | | (1 << ARG_REGISTER (2)) | |
16428 | | (1 << ARG_REGISTER (3)); | |
16429 | else | |
16430 | regs_available_for_popping = | |
16431 | (1 << ARG_REGISTER (2)) | |
16432 | | (1 << ARG_REGISTER (3)); | |
16433 | } | |
16434 | else if (size <= 4) | |
16435 | regs_available_for_popping = | |
16436 | (1 << ARG_REGISTER (2)) | |
16437 | | (1 << ARG_REGISTER (3)); | |
16438 | else if (size <= 8) | |
16439 | regs_available_for_popping = | |
16440 | (1 << ARG_REGISTER (3)); | |
16441 | } | |
16442 | ||
16443 | /* Match registers to be popped with registers into which we pop them. */ | |
16444 | for (available = regs_available_for_popping, | |
16445 | required = regs_to_pop; | |
16446 | required != 0 && available != 0; | |
16447 | available &= ~(available & - available), | |
16448 | required &= ~(required & - required)) | |
16449 | -- pops_needed; | |
16450 | ||
16451 | /* If we have any popping registers left over, remove them. */ | |
16452 | if (available > 0) | |
5895f793 | 16453 | regs_available_for_popping &= ~available; |
f676971a | 16454 | |
d5b7b3ae RE |
16455 | /* Otherwise if we need another popping register we can use |
16456 | the fourth argument register. */ | |
16457 | else if (pops_needed) | |
16458 | { | |
16459 | /* If we have not found any free argument registers and | |
16460 | reg a4 contains the return address, we must move it. */ | |
16461 | if (regs_available_for_popping == 0 | |
16462 | && reg_containing_return_addr == LAST_ARG_REGNUM) | |
16463 | { | |
16464 | asm_fprintf (f, "\tmov\t%r, %r\n", LR_REGNUM, LAST_ARG_REGNUM); | |
16465 | reg_containing_return_addr = LR_REGNUM; | |
16466 | } | |
16467 | else if (size > 12) | |
16468 | { | |
16469 | /* Register a4 is being used to hold part of the return value, | |
16470 | but we have dire need of a free, low register. */ | |
16471 | restore_a4 = TRUE; | |
f676971a | 16472 | |
d5b7b3ae RE |
16473 | asm_fprintf (f, "\tmov\t%r, %r\n",IP_REGNUM, LAST_ARG_REGNUM); |
16474 | } | |
f676971a | 16475 | |
d5b7b3ae RE |
16476 | if (reg_containing_return_addr != LAST_ARG_REGNUM) |
16477 | { | |
16478 | /* The fourth argument register is available. */ | |
16479 | regs_available_for_popping |= 1 << LAST_ARG_REGNUM; | |
f676971a | 16480 | |
5895f793 | 16481 | --pops_needed; |
d5b7b3ae RE |
16482 | } |
16483 | } | |
16484 | ||
16485 | /* Pop as many registers as we can. */ | |
980e61bb DJ |
16486 | thumb_pushpop (f, regs_available_for_popping, FALSE, NULL, |
16487 | regs_available_for_popping); | |
d5b7b3ae RE |
16488 | |
16489 | /* Process the registers we popped. */ | |
16490 | if (reg_containing_return_addr == -1) | |
16491 | { | |
16492 | /* The return address was popped into the lowest numbered register. */ | |
5895f793 | 16493 | regs_to_pop &= ~(1 << LR_REGNUM); |
f676971a | 16494 | |
d5b7b3ae RE |
16495 | reg_containing_return_addr = |
16496 | number_of_first_bit_set (regs_available_for_popping); | |
16497 | ||
16498 | /* Remove this register for the mask of available registers, so that | |
6bc82793 | 16499 | the return address will not be corrupted by further pops. */ |
5895f793 | 16500 | regs_available_for_popping &= ~(1 << reg_containing_return_addr); |
d5b7b3ae RE |
16501 | } |
16502 | ||
16503 | /* If we popped other registers then handle them here. */ | |
16504 | if (regs_available_for_popping) | |
16505 | { | |
16506 | int frame_pointer; | |
f676971a | 16507 | |
d5b7b3ae RE |
16508 | /* Work out which register currently contains the frame pointer. */ |
16509 | frame_pointer = number_of_first_bit_set (regs_available_for_popping); | |
16510 | ||
16511 | /* Move it into the correct place. */ | |
16512 | asm_fprintf (f, "\tmov\t%r, %r\n", | |
16513 | ARM_HARD_FRAME_POINTER_REGNUM, frame_pointer); | |
16514 | ||
16515 | /* (Temporarily) remove it from the mask of popped registers. */ | |
5895f793 RE |
16516 | regs_available_for_popping &= ~(1 << frame_pointer); |
16517 | regs_to_pop &= ~(1 << ARM_HARD_FRAME_POINTER_REGNUM); | |
f676971a | 16518 | |
d5b7b3ae RE |
16519 | if (regs_available_for_popping) |
16520 | { | |
16521 | int stack_pointer; | |
f676971a | 16522 | |
d5b7b3ae RE |
16523 | /* We popped the stack pointer as well, |
16524 | find the register that contains it. */ | |
16525 | stack_pointer = number_of_first_bit_set (regs_available_for_popping); | |
16526 | ||
16527 | /* Move it into the stack register. */ | |
16528 | asm_fprintf (f, "\tmov\t%r, %r\n", SP_REGNUM, stack_pointer); | |
f676971a | 16529 | |
d5b7b3ae RE |
16530 | /* At this point we have popped all necessary registers, so |
16531 | do not worry about restoring regs_available_for_popping | |
16532 | to its correct value: | |
16533 | ||
16534 | assert (pops_needed == 0) | |
16535 | assert (regs_available_for_popping == (1 << frame_pointer)) | |
16536 | assert (regs_to_pop == (1 << STACK_POINTER)) */ | |
16537 | } | |
16538 | else | |
16539 | { | |
16540 | /* Since we have just move the popped value into the frame | |
16541 | pointer, the popping register is available for reuse, and | |
16542 | we know that we still have the stack pointer left to pop. */ | |
16543 | regs_available_for_popping |= (1 << frame_pointer); | |
16544 | } | |
16545 | } | |
f676971a | 16546 | |
d5b7b3ae RE |
16547 | /* If we still have registers left on the stack, but we no longer have |
16548 | any registers into which we can pop them, then we must move the return | |
16549 | address into the link register and make available the register that | |
16550 | contained it. */ | |
16551 | if (regs_available_for_popping == 0 && pops_needed > 0) | |
16552 | { | |
16553 | regs_available_for_popping |= 1 << reg_containing_return_addr; | |
f676971a | 16554 | |
d5b7b3ae RE |
16555 | asm_fprintf (f, "\tmov\t%r, %r\n", LR_REGNUM, |
16556 | reg_containing_return_addr); | |
f676971a | 16557 | |
d5b7b3ae RE |
16558 | reg_containing_return_addr = LR_REGNUM; |
16559 | } | |
16560 | ||
16561 | /* If we have registers left on the stack then pop some more. | |
16562 | We know that at most we will want to pop FP and SP. */ | |
16563 | if (pops_needed > 0) | |
16564 | { | |
16565 | int popped_into; | |
16566 | int move_to; | |
f676971a | 16567 | |
980e61bb DJ |
16568 | thumb_pushpop (f, regs_available_for_popping, FALSE, NULL, |
16569 | regs_available_for_popping); | |
d5b7b3ae RE |
16570 | |
16571 | /* We have popped either FP or SP. | |
16572 | Move whichever one it is into the correct register. */ | |
16573 | popped_into = number_of_first_bit_set (regs_available_for_popping); | |
16574 | move_to = number_of_first_bit_set (regs_to_pop); | |
16575 | ||
16576 | asm_fprintf (f, "\tmov\t%r, %r\n", move_to, popped_into); | |
16577 | ||
5895f793 | 16578 | regs_to_pop &= ~(1 << move_to); |
d5b7b3ae | 16579 | |
5895f793 | 16580 | --pops_needed; |
d5b7b3ae | 16581 | } |
f676971a | 16582 | |
d5b7b3ae RE |
16583 | /* If we still have not popped everything then we must have only |
16584 | had one register available to us and we are now popping the SP. */ | |
16585 | if (pops_needed > 0) | |
16586 | { | |
16587 | int popped_into; | |
f676971a | 16588 | |
980e61bb DJ |
16589 | thumb_pushpop (f, regs_available_for_popping, FALSE, NULL, |
16590 | regs_available_for_popping); | |
d5b7b3ae RE |
16591 | |
16592 | popped_into = number_of_first_bit_set (regs_available_for_popping); | |
16593 | ||
16594 | asm_fprintf (f, "\tmov\t%r, %r\n", SP_REGNUM, popped_into); | |
16595 | /* | |
16596 | assert (regs_to_pop == (1 << STACK_POINTER)) | |
16597 | assert (pops_needed == 1) | |
16598 | */ | |
16599 | } | |
16600 | ||
16601 | /* If necessary restore the a4 register. */ | |
16602 | if (restore_a4) | |
16603 | { | |
16604 | if (reg_containing_return_addr != LR_REGNUM) | |
16605 | { | |
16606 | asm_fprintf (f, "\tmov\t%r, %r\n", LR_REGNUM, LAST_ARG_REGNUM); | |
16607 | reg_containing_return_addr = LR_REGNUM; | |
16608 | } | |
f676971a | 16609 | |
d5b7b3ae RE |
16610 | asm_fprintf (f, "\tmov\t%r, %r\n", LAST_ARG_REGNUM, IP_REGNUM); |
16611 | } | |
16612 | ||
e3b5732b | 16613 | if (crtl->calls_eh_return) |
c9ca9b88 | 16614 | asm_fprintf (f, "\tadd\t%r, %r\n", SP_REGNUM, ARM_EH_STACKADJ_REGNUM); |
d5b7b3ae RE |
16615 | |
16616 | /* Return to caller. */ | |
16617 | asm_fprintf (f, "\tbx\t%r\n", reg_containing_return_addr); | |
16618 | } | |
16619 | ||
d5b7b3ae RE |
16620 | \f |
16621 | void | |
5b3e6663 | 16622 | thumb1_final_prescan_insn (rtx insn) |
d5b7b3ae | 16623 | { |
d5b7b3ae | 16624 | if (flag_print_asm_name) |
9d98a694 AO |
16625 | asm_fprintf (asm_out_file, "%@ 0x%04x\n", |
16626 | INSN_ADDRESSES (INSN_UID (insn))); | |
d5b7b3ae RE |
16627 | } |
16628 | ||
16629 | int | |
e32bac5b | 16630 | thumb_shiftable_const (unsigned HOST_WIDE_INT val) |
d5b7b3ae RE |
16631 | { |
16632 | unsigned HOST_WIDE_INT mask = 0xff; | |
16633 | int i; | |
16634 | ||
16635 | if (val == 0) /* XXX */ | |
16636 | return 0; | |
f676971a | 16637 | |
d5b7b3ae RE |
16638 | for (i = 0; i < 25; i++) |
16639 | if ((val & (mask << i)) == val) | |
16640 | return 1; | |
16641 | ||
16642 | return 0; | |
16643 | } | |
16644 | ||
825dda42 | 16645 | /* Returns nonzero if the current function contains, |
d5b7b3ae | 16646 | or might contain a far jump. */ |
5848830f PB |
16647 | static int |
16648 | thumb_far_jump_used_p (void) | |
d5b7b3ae RE |
16649 | { |
16650 | rtx insn; | |
16651 | ||
16652 | /* This test is only important for leaf functions. */ | |
5895f793 | 16653 | /* assert (!leaf_function_p ()); */ |
f676971a | 16654 | |
d5b7b3ae RE |
16655 | /* If we have already decided that far jumps may be used, |
16656 | do not bother checking again, and always return true even if | |
16657 | it turns out that they are not being used. Once we have made | |
16658 | the decision that far jumps are present (and that hence the link | |
16659 | register will be pushed onto the stack) we cannot go back on it. */ | |
16660 | if (cfun->machine->far_jump_used) | |
16661 | return 1; | |
16662 | ||
16663 | /* If this function is not being called from the prologue/epilogue | |
16664 | generation code then it must be being called from the | |
16665 | INITIAL_ELIMINATION_OFFSET macro. */ | |
5848830f | 16666 | if (!(ARM_DOUBLEWORD_ALIGN || reload_completed)) |
d5b7b3ae RE |
16667 | { |
16668 | /* In this case we know that we are being asked about the elimination | |
16669 | of the arg pointer register. If that register is not being used, | |
16670 | then there are no arguments on the stack, and we do not have to | |
16671 | worry that a far jump might force the prologue to push the link | |
16672 | register, changing the stack offsets. In this case we can just | |
16673 | return false, since the presence of far jumps in the function will | |
16674 | not affect stack offsets. | |
16675 | ||
16676 | If the arg pointer is live (or if it was live, but has now been | |
16677 | eliminated and so set to dead) then we do have to test to see if | |
16678 | the function might contain a far jump. This test can lead to some | |
16679 | false negatives, since before reload is completed, then length of | |
16680 | branch instructions is not known, so gcc defaults to returning their | |
16681 | longest length, which in turn sets the far jump attribute to true. | |
16682 | ||
16683 | A false negative will not result in bad code being generated, but it | |
16684 | will result in a needless push and pop of the link register. We | |
5848830f PB |
16685 | hope that this does not occur too often. |
16686 | ||
16687 | If we need doubleword stack alignment this could affect the other | |
16688 | elimination offsets so we can't risk getting it wrong. */ | |
6fb5fa3c | 16689 | if (df_regs_ever_live_p (ARG_POINTER_REGNUM)) |
d5b7b3ae | 16690 | cfun->machine->arg_pointer_live = 1; |
5895f793 | 16691 | else if (!cfun->machine->arg_pointer_live) |
d5b7b3ae RE |
16692 | return 0; |
16693 | } | |
16694 | ||
16695 | /* Check to see if the function contains a branch | |
16696 | insn with the far jump attribute set. */ | |
16697 | for (insn = get_insns (); insn; insn = NEXT_INSN (insn)) | |
16698 | { | |
16699 | if (GET_CODE (insn) == JUMP_INSN | |
16700 | /* Ignore tablejump patterns. */ | |
16701 | && GET_CODE (PATTERN (insn)) != ADDR_VEC | |
16702 | && GET_CODE (PATTERN (insn)) != ADDR_DIFF_VEC | |
16703 | && get_attr_far_jump (insn) == FAR_JUMP_YES | |
16704 | ) | |
16705 | { | |
9a9f7594 | 16706 | /* Record the fact that we have decided that |
d5b7b3ae RE |
16707 | the function does use far jumps. */ |
16708 | cfun->machine->far_jump_used = 1; | |
16709 | return 1; | |
16710 | } | |
16711 | } | |
f676971a | 16712 | |
d5b7b3ae RE |
16713 | return 0; |
16714 | } | |
16715 | ||
825dda42 | 16716 | /* Return nonzero if FUNC must be entered in ARM mode. */ |
d5b7b3ae | 16717 | int |
e32bac5b | 16718 | is_called_in_ARM_mode (tree func) |
d5b7b3ae | 16719 | { |
e6d29d15 | 16720 | gcc_assert (TREE_CODE (func) == FUNCTION_DECL); |
d5b7b3ae | 16721 | |
696e78bf | 16722 | /* Ignore the problem about functions whose address is taken. */ |
d5b7b3ae RE |
16723 | if (TARGET_CALLEE_INTERWORKING && TREE_PUBLIC (func)) |
16724 | return TRUE; | |
16725 | ||
f676971a | 16726 | #ifdef ARM_PE |
91d231cb | 16727 | return lookup_attribute ("interfacearm", DECL_ATTRIBUTES (func)) != NULL_TREE; |
d5b7b3ae RE |
16728 | #else |
16729 | return FALSE; | |
16730 | #endif | |
16731 | } | |
16732 | ||
d6b4baa4 | 16733 | /* The bits which aren't usefully expanded as rtl. */ |
cd2b33d0 | 16734 | const char * |
e32bac5b | 16735 | thumb_unexpanded_epilogue (void) |
d5b7b3ae | 16736 | { |
954954d1 | 16737 | arm_stack_offsets *offsets; |
d5b7b3ae | 16738 | int regno; |
b279b20a | 16739 | unsigned long live_regs_mask = 0; |
d5b7b3ae | 16740 | int high_regs_pushed = 0; |
d5b7b3ae | 16741 | int had_to_push_lr; |
57934c39 | 16742 | int size; |
d5b7b3ae RE |
16743 | |
16744 | if (return_used_this_function) | |
16745 | return ""; | |
16746 | ||
58e60158 AN |
16747 | if (IS_NAKED (arm_current_func_type ())) |
16748 | return ""; | |
16749 | ||
954954d1 PB |
16750 | offsets = arm_get_frame_offsets (); |
16751 | live_regs_mask = offsets->saved_regs_mask; | |
57934c39 PB |
16752 | high_regs_pushed = bit_count (live_regs_mask & 0x0f00); |
16753 | ||
16754 | /* If we can deduce the registers used from the function's return value. | |
6fb5fa3c | 16755 | This is more reliable that examining df_regs_ever_live_p () because that |
57934c39 PB |
16756 | will be set if the register is ever used in the function, not just if |
16757 | the register is used to hold a return value. */ | |
4f5dfed0 | 16758 | size = arm_size_return_regs (); |
d5b7b3ae RE |
16759 | |
16760 | /* The prolog may have pushed some high registers to use as | |
112cdef5 | 16761 | work registers. e.g. the testsuite file: |
d5b7b3ae RE |
16762 | gcc/testsuite/gcc/gcc.c-torture/execute/complex-2.c |
16763 | compiles to produce: | |
16764 | push {r4, r5, r6, r7, lr} | |
16765 | mov r7, r9 | |
16766 | mov r6, r8 | |
16767 | push {r6, r7} | |
16768 | as part of the prolog. We have to undo that pushing here. */ | |
f676971a | 16769 | |
d5b7b3ae RE |
16770 | if (high_regs_pushed) |
16771 | { | |
b279b20a | 16772 | unsigned long mask = live_regs_mask & 0xff; |
d5b7b3ae | 16773 | int next_hi_reg; |
d5b7b3ae | 16774 | |
57934c39 PB |
16775 | /* The available low registers depend on the size of the value we are |
16776 | returning. */ | |
16777 | if (size <= 12) | |
d5b7b3ae | 16778 | mask |= 1 << 3; |
57934c39 PB |
16779 | if (size <= 8) |
16780 | mask |= 1 << 2; | |
d5b7b3ae RE |
16781 | |
16782 | if (mask == 0) | |
16783 | /* Oh dear! We have no low registers into which we can pop | |
16784 | high registers! */ | |
400500c4 RK |
16785 | internal_error |
16786 | ("no low registers available for popping high registers"); | |
f676971a | 16787 | |
d5b7b3ae | 16788 | for (next_hi_reg = 8; next_hi_reg < 13; next_hi_reg++) |
57934c39 | 16789 | if (live_regs_mask & (1 << next_hi_reg)) |
d5b7b3ae RE |
16790 | break; |
16791 | ||
16792 | while (high_regs_pushed) | |
16793 | { | |
16794 | /* Find lo register(s) into which the high register(s) can | |
16795 | be popped. */ | |
16796 | for (regno = 0; regno <= LAST_LO_REGNUM; regno++) | |
16797 | { | |
16798 | if (mask & (1 << regno)) | |
16799 | high_regs_pushed--; | |
16800 | if (high_regs_pushed == 0) | |
16801 | break; | |
16802 | } | |
16803 | ||
16804 | mask &= (2 << regno) - 1; /* A noop if regno == 8 */ | |
16805 | ||
d6b4baa4 | 16806 | /* Pop the values into the low register(s). */ |
980e61bb | 16807 | thumb_pushpop (asm_out_file, mask, 0, NULL, mask); |
d5b7b3ae RE |
16808 | |
16809 | /* Move the value(s) into the high registers. */ | |
16810 | for (regno = 0; regno <= LAST_LO_REGNUM; regno++) | |
16811 | { | |
16812 | if (mask & (1 << regno)) | |
16813 | { | |
16814 | asm_fprintf (asm_out_file, "\tmov\t%r, %r\n", next_hi_reg, | |
16815 | regno); | |
f676971a | 16816 | |
d5b7b3ae | 16817 | for (next_hi_reg++; next_hi_reg < 13; next_hi_reg++) |
57934c39 | 16818 | if (live_regs_mask & (1 << next_hi_reg)) |
d5b7b3ae RE |
16819 | break; |
16820 | } | |
16821 | } | |
16822 | } | |
57934c39 | 16823 | live_regs_mask &= ~0x0f00; |
d5b7b3ae RE |
16824 | } |
16825 | ||
57934c39 PB |
16826 | had_to_push_lr = (live_regs_mask & (1 << LR_REGNUM)) != 0; |
16827 | live_regs_mask &= 0xff; | |
16828 | ||
38173d38 | 16829 | if (crtl->args.pretend_args_size == 0 || TARGET_BACKTRACE) |
d5b7b3ae | 16830 | { |
f676971a | 16831 | /* Pop the return address into the PC. */ |
57934c39 | 16832 | if (had_to_push_lr) |
d5b7b3ae RE |
16833 | live_regs_mask |= 1 << PC_REGNUM; |
16834 | ||
16835 | /* Either no argument registers were pushed or a backtrace | |
16836 | structure was created which includes an adjusted stack | |
16837 | pointer, so just pop everything. */ | |
16838 | if (live_regs_mask) | |
980e61bb DJ |
16839 | thumb_pushpop (asm_out_file, live_regs_mask, FALSE, NULL, |
16840 | live_regs_mask); | |
57934c39 | 16841 | |
d5b7b3ae | 16842 | /* We have either just popped the return address into the |
57934c39 PB |
16843 | PC or it is was kept in LR for the entire function. */ |
16844 | if (!had_to_push_lr) | |
16845 | thumb_exit (asm_out_file, LR_REGNUM); | |
d5b7b3ae RE |
16846 | } |
16847 | else | |
16848 | { | |
16849 | /* Pop everything but the return address. */ | |
d5b7b3ae | 16850 | if (live_regs_mask) |
980e61bb DJ |
16851 | thumb_pushpop (asm_out_file, live_regs_mask, FALSE, NULL, |
16852 | live_regs_mask); | |
d5b7b3ae RE |
16853 | |
16854 | if (had_to_push_lr) | |
57934c39 PB |
16855 | { |
16856 | if (size > 12) | |
16857 | { | |
16858 | /* We have no free low regs, so save one. */ | |
16859 | asm_fprintf (asm_out_file, "\tmov\t%r, %r\n", IP_REGNUM, | |
16860 | LAST_ARG_REGNUM); | |
16861 | } | |
16862 | ||
16863 | /* Get the return address into a temporary register. */ | |
16864 | thumb_pushpop (asm_out_file, 1 << LAST_ARG_REGNUM, 0, NULL, | |
16865 | 1 << LAST_ARG_REGNUM); | |
16866 | ||
16867 | if (size > 12) | |
16868 | { | |
16869 | /* Move the return address to lr. */ | |
16870 | asm_fprintf (asm_out_file, "\tmov\t%r, %r\n", LR_REGNUM, | |
16871 | LAST_ARG_REGNUM); | |
16872 | /* Restore the low register. */ | |
16873 | asm_fprintf (asm_out_file, "\tmov\t%r, %r\n", LAST_ARG_REGNUM, | |
16874 | IP_REGNUM); | |
16875 | regno = LR_REGNUM; | |
16876 | } | |
16877 | else | |
16878 | regno = LAST_ARG_REGNUM; | |
16879 | } | |
16880 | else | |
16881 | regno = LR_REGNUM; | |
f676971a | 16882 | |
d5b7b3ae RE |
16883 | /* Remove the argument registers that were pushed onto the stack. */ |
16884 | asm_fprintf (asm_out_file, "\tadd\t%r, %r, #%d\n", | |
16885 | SP_REGNUM, SP_REGNUM, | |
38173d38 | 16886 | crtl->args.pretend_args_size); |
f676971a | 16887 | |
57934c39 | 16888 | thumb_exit (asm_out_file, regno); |
d5b7b3ae RE |
16889 | } |
16890 | ||
16891 | return ""; | |
16892 | } | |
16893 | ||
16894 | /* Functions to save and restore machine-specific function data. */ | |
e2500fed | 16895 | static struct machine_function * |
e32bac5b | 16896 | arm_init_machine_status (void) |
d5b7b3ae | 16897 | { |
e2500fed GK |
16898 | struct machine_function *machine; |
16899 | machine = (machine_function *) ggc_alloc_cleared (sizeof (machine_function)); | |
6d3d9133 | 16900 | |
f676971a | 16901 | #if ARM_FT_UNKNOWN != 0 |
e2500fed | 16902 | machine->func_type = ARM_FT_UNKNOWN; |
6d3d9133 | 16903 | #endif |
e2500fed | 16904 | return machine; |
f7a80099 NC |
16905 | } |
16906 | ||
d5b7b3ae RE |
16907 | /* Return an RTX indicating where the return address to the |
16908 | calling function can be found. */ | |
16909 | rtx | |
e32bac5b | 16910 | arm_return_addr (int count, rtx frame ATTRIBUTE_UNUSED) |
d5b7b3ae | 16911 | { |
d5b7b3ae RE |
16912 | if (count != 0) |
16913 | return NULL_RTX; | |
16914 | ||
61f0ccff | 16915 | return get_hard_reg_initial_val (Pmode, LR_REGNUM); |
d5b7b3ae RE |
16916 | } |
16917 | ||
16918 | /* Do anything needed before RTL is emitted for each function. */ | |
16919 | void | |
e32bac5b | 16920 | arm_init_expanders (void) |
d5b7b3ae RE |
16921 | { |
16922 | /* Arrange to initialize and mark the machine per-function status. */ | |
16923 | init_machine_status = arm_init_machine_status; | |
3ac5ea7c RH |
16924 | |
16925 | /* This is to stop the combine pass optimizing away the alignment | |
16926 | adjustment of va_arg. */ | |
16927 | /* ??? It is claimed that this should not be necessary. */ | |
16928 | if (cfun) | |
16929 | mark_reg_pointer (arg_pointer_rtx, PARM_BOUNDARY); | |
d5b7b3ae RE |
16930 | } |
16931 | ||
0977774b | 16932 | |
2591db65 RE |
16933 | /* Like arm_compute_initial_elimination offset. Simpler because there |
16934 | isn't an ABI specified frame pointer for Thumb. Instead, we set it | |
16935 | to point at the base of the local variables after static stack | |
16936 | space for a function has been allocated. */ | |
0977774b | 16937 | |
5848830f PB |
16938 | HOST_WIDE_INT |
16939 | thumb_compute_initial_elimination_offset (unsigned int from, unsigned int to) | |
16940 | { | |
16941 | arm_stack_offsets *offsets; | |
0977774b | 16942 | |
5848830f | 16943 | offsets = arm_get_frame_offsets (); |
0977774b | 16944 | |
5848830f | 16945 | switch (from) |
0977774b | 16946 | { |
5848830f PB |
16947 | case ARG_POINTER_REGNUM: |
16948 | switch (to) | |
16949 | { | |
16950 | case STACK_POINTER_REGNUM: | |
16951 | return offsets->outgoing_args - offsets->saved_args; | |
0977774b | 16952 | |
5848830f PB |
16953 | case FRAME_POINTER_REGNUM: |
16954 | return offsets->soft_frame - offsets->saved_args; | |
0977774b | 16955 | |
5848830f PB |
16956 | case ARM_HARD_FRAME_POINTER_REGNUM: |
16957 | return offsets->saved_regs - offsets->saved_args; | |
0977774b | 16958 | |
2591db65 RE |
16959 | case THUMB_HARD_FRAME_POINTER_REGNUM: |
16960 | return offsets->locals_base - offsets->saved_args; | |
16961 | ||
5848830f | 16962 | default: |
e6d29d15 | 16963 | gcc_unreachable (); |
5848830f PB |
16964 | } |
16965 | break; | |
0977774b | 16966 | |
5848830f PB |
16967 | case FRAME_POINTER_REGNUM: |
16968 | switch (to) | |
16969 | { | |
16970 | case STACK_POINTER_REGNUM: | |
16971 | return offsets->outgoing_args - offsets->soft_frame; | |
0977774b | 16972 | |
5848830f PB |
16973 | case ARM_HARD_FRAME_POINTER_REGNUM: |
16974 | return offsets->saved_regs - offsets->soft_frame; | |
0977774b | 16975 | |
2591db65 RE |
16976 | case THUMB_HARD_FRAME_POINTER_REGNUM: |
16977 | return offsets->locals_base - offsets->soft_frame; | |
16978 | ||
5848830f | 16979 | default: |
e6d29d15 | 16980 | gcc_unreachable (); |
5848830f PB |
16981 | } |
16982 | break; | |
0977774b | 16983 | |
5848830f | 16984 | default: |
e6d29d15 | 16985 | gcc_unreachable (); |
5848830f | 16986 | } |
0977774b JT |
16987 | } |
16988 | ||
d5b7b3ae RE |
16989 | /* Generate the rest of a function's prologue. */ |
16990 | void | |
5b3e6663 | 16991 | thumb1_expand_prologue (void) |
d5b7b3ae | 16992 | { |
980e61bb DJ |
16993 | rtx insn, dwarf; |
16994 | ||
5848830f PB |
16995 | HOST_WIDE_INT amount; |
16996 | arm_stack_offsets *offsets; | |
6d3d9133 | 16997 | unsigned long func_type; |
3c7ad43e | 16998 | int regno; |
57934c39 | 16999 | unsigned long live_regs_mask; |
6d3d9133 NC |
17000 | |
17001 | func_type = arm_current_func_type (); | |
f676971a | 17002 | |
d5b7b3ae | 17003 | /* Naked functions don't have prologues. */ |
6d3d9133 | 17004 | if (IS_NAKED (func_type)) |
d5b7b3ae RE |
17005 | return; |
17006 | ||
6d3d9133 NC |
17007 | if (IS_INTERRUPT (func_type)) |
17008 | { | |
c725bd79 | 17009 | error ("interrupt Service Routines cannot be coded in Thumb mode"); |
6d3d9133 NC |
17010 | return; |
17011 | } | |
17012 | ||
954954d1 PB |
17013 | offsets = arm_get_frame_offsets (); |
17014 | live_regs_mask = offsets->saved_regs_mask; | |
b279b20a NC |
17015 | /* Load the pic register before setting the frame pointer, |
17016 | so we can use r7 as a temporary work register. */ | |
020a4035 | 17017 | if (flag_pic && arm_pic_register != INVALID_REGNUM) |
e55ef7f4 | 17018 | arm_load_pic_register (live_regs_mask); |
876f13b0 | 17019 | |
2591db65 | 17020 | if (!frame_pointer_needed && CALLER_INTERWORKING_SLOT_SIZE > 0) |
a2503645 RS |
17021 | emit_move_insn (gen_rtx_REG (Pmode, ARM_HARD_FRAME_POINTER_REGNUM), |
17022 | stack_pointer_rtx); | |
d5b7b3ae | 17023 | |
5848830f | 17024 | amount = offsets->outgoing_args - offsets->saved_regs; |
d5b7b3ae RE |
17025 | if (amount) |
17026 | { | |
d5b7b3ae | 17027 | if (amount < 512) |
980e61bb DJ |
17028 | { |
17029 | insn = emit_insn (gen_addsi3 (stack_pointer_rtx, stack_pointer_rtx, | |
17030 | GEN_INT (- amount))); | |
17031 | RTX_FRAME_RELATED_P (insn) = 1; | |
17032 | } | |
d5b7b3ae RE |
17033 | else |
17034 | { | |
d5b7b3ae RE |
17035 | rtx reg; |
17036 | ||
17037 | /* The stack decrement is too big for an immediate value in a single | |
17038 | insn. In theory we could issue multiple subtracts, but after | |
17039 | three of them it becomes more space efficient to place the full | |
17040 | value in the constant pool and load into a register. (Also the | |
17041 | ARM debugger really likes to see only one stack decrement per | |
17042 | function). So instead we look for a scratch register into which | |
17043 | we can load the decrement, and then we subtract this from the | |
17044 | stack pointer. Unfortunately on the thumb the only available | |
17045 | scratch registers are the argument registers, and we cannot use | |
17046 | these as they may hold arguments to the function. Instead we | |
17047 | attempt to locate a call preserved register which is used by this | |
17048 | function. If we can find one, then we know that it will have | |
17049 | been pushed at the start of the prologue and so we can corrupt | |
17050 | it now. */ | |
17051 | for (regno = LAST_ARG_REGNUM + 1; regno <= LAST_LO_REGNUM; regno++) | |
35596784 | 17052 | if (live_regs_mask & (1 << regno)) |
d5b7b3ae RE |
17053 | break; |
17054 | ||
35596784 | 17055 | gcc_assert(regno <= LAST_LO_REGNUM); |
d5b7b3ae | 17056 | |
35596784 | 17057 | reg = gen_rtx_REG (SImode, regno); |
d5b7b3ae | 17058 | |
35596784 | 17059 | emit_insn (gen_movsi (reg, GEN_INT (- amount))); |
980e61bb | 17060 | |
35596784 AJ |
17061 | insn = emit_insn (gen_addsi3 (stack_pointer_rtx, |
17062 | stack_pointer_rtx, reg)); | |
17063 | RTX_FRAME_RELATED_P (insn) = 1; | |
17064 | dwarf = gen_rtx_SET (VOIDmode, stack_pointer_rtx, | |
17065 | plus_constant (stack_pointer_rtx, | |
17066 | -amount)); | |
17067 | RTX_FRAME_RELATED_P (dwarf) = 1; | |
17068 | REG_NOTES (insn) | |
17069 | = gen_rtx_EXPR_LIST (REG_FRAME_RELATED_EXPR, dwarf, | |
17070 | REG_NOTES (insn)); | |
d5b7b3ae | 17071 | } |
2591db65 RE |
17072 | } |
17073 | ||
17074 | if (frame_pointer_needed) | |
5b3e6663 | 17075 | thumb_set_frame_pointer (offsets); |
f676971a | 17076 | |
74d9c39f DJ |
17077 | /* If we are profiling, make sure no instructions are scheduled before |
17078 | the call to mcount. Similarly if the user has requested no | |
17079 | scheduling in the prolog. Similarly if we want non-call exceptions | |
17080 | using the EABI unwinder, to prevent faulting instructions from being | |
17081 | swapped with a stack adjustment. */ | |
e3b5732b | 17082 | if (crtl->profile || !TARGET_SCHED_PROLOG |
74d9c39f | 17083 | || (ARM_EABI_UNWIND_TABLES && flag_non_call_exceptions)) |
d5b7b3ae | 17084 | emit_insn (gen_blockage ()); |
3c7ad43e PB |
17085 | |
17086 | cfun->machine->lr_save_eliminated = !thumb_force_lr_save (); | |
57934c39 PB |
17087 | if (live_regs_mask & 0xff) |
17088 | cfun->machine->lr_save_eliminated = 0; | |
d5b7b3ae RE |
17089 | } |
17090 | ||
57934c39 | 17091 | |
d5b7b3ae | 17092 | void |
5b3e6663 | 17093 | thumb1_expand_epilogue (void) |
d5b7b3ae | 17094 | { |
5848830f PB |
17095 | HOST_WIDE_INT amount; |
17096 | arm_stack_offsets *offsets; | |
defc0463 RE |
17097 | int regno; |
17098 | ||
6d3d9133 NC |
17099 | /* Naked functions don't have prologues. */ |
17100 | if (IS_NAKED (arm_current_func_type ())) | |
d5b7b3ae RE |
17101 | return; |
17102 | ||
5848830f PB |
17103 | offsets = arm_get_frame_offsets (); |
17104 | amount = offsets->outgoing_args - offsets->saved_regs; | |
17105 | ||
d5b7b3ae | 17106 | if (frame_pointer_needed) |
2591db65 RE |
17107 | { |
17108 | emit_insn (gen_movsi (stack_pointer_rtx, hard_frame_pointer_rtx)); | |
17109 | amount = offsets->locals_base - offsets->saved_regs; | |
17110 | } | |
e0b92319 | 17111 | |
a3a531ec | 17112 | gcc_assert (amount >= 0); |
2591db65 | 17113 | if (amount) |
d5b7b3ae | 17114 | { |
d5b7b3ae RE |
17115 | if (amount < 512) |
17116 | emit_insn (gen_addsi3 (stack_pointer_rtx, stack_pointer_rtx, | |
17117 | GEN_INT (amount))); | |
17118 | else | |
17119 | { | |
17120 | /* r3 is always free in the epilogue. */ | |
f1c25d3b | 17121 | rtx reg = gen_rtx_REG (SImode, LAST_ARG_REGNUM); |
d5b7b3ae RE |
17122 | |
17123 | emit_insn (gen_movsi (reg, GEN_INT (amount))); | |
17124 | emit_insn (gen_addsi3 (stack_pointer_rtx, stack_pointer_rtx, reg)); | |
17125 | } | |
17126 | } | |
f676971a | 17127 | |
d5b7b3ae RE |
17128 | /* Emit a USE (stack_pointer_rtx), so that |
17129 | the stack adjustment will not be deleted. */ | |
6bacc7b0 | 17130 | emit_insn (gen_prologue_use (stack_pointer_rtx)); |
d5b7b3ae | 17131 | |
e3b5732b | 17132 | if (crtl->profile || !TARGET_SCHED_PROLOG) |
d5b7b3ae | 17133 | emit_insn (gen_blockage ()); |
defc0463 RE |
17134 | |
17135 | /* Emit a clobber for each insn that will be restored in the epilogue, | |
17136 | so that flow2 will get register lifetimes correct. */ | |
17137 | for (regno = 0; regno < 13; regno++) | |
6fb5fa3c | 17138 | if (df_regs_ever_live_p (regno) && !call_used_regs[regno]) |
c41c1387 | 17139 | emit_clobber (gen_rtx_REG (SImode, regno)); |
defc0463 | 17140 | |
6fb5fa3c | 17141 | if (! df_regs_ever_live_p (LR_REGNUM)) |
c41c1387 | 17142 | emit_use (gen_rtx_REG (SImode, LR_REGNUM)); |
d5b7b3ae RE |
17143 | } |
17144 | ||
08c148a8 | 17145 | static void |
5b3e6663 | 17146 | thumb1_output_function_prologue (FILE *f, HOST_WIDE_INT size ATTRIBUTE_UNUSED) |
d5b7b3ae | 17147 | { |
954954d1 | 17148 | arm_stack_offsets *offsets; |
b279b20a NC |
17149 | unsigned long live_regs_mask = 0; |
17150 | unsigned long l_mask; | |
17151 | unsigned high_regs_pushed = 0; | |
980e61bb | 17152 | int cfa_offset = 0; |
d5b7b3ae RE |
17153 | int regno; |
17154 | ||
6d3d9133 | 17155 | if (IS_NAKED (arm_current_func_type ())) |
d5b7b3ae RE |
17156 | return; |
17157 | ||
17158 | if (is_called_in_ARM_mode (current_function_decl)) | |
17159 | { | |
17160 | const char * name; | |
17161 | ||
e6d29d15 NS |
17162 | gcc_assert (GET_CODE (DECL_RTL (current_function_decl)) == MEM); |
17163 | gcc_assert (GET_CODE (XEXP (DECL_RTL (current_function_decl), 0)) | |
17164 | == SYMBOL_REF); | |
d5b7b3ae | 17165 | name = XSTR (XEXP (DECL_RTL (current_function_decl), 0), 0); |
f676971a | 17166 | |
d5b7b3ae RE |
17167 | /* Generate code sequence to switch us into Thumb mode. */ |
17168 | /* The .code 32 directive has already been emitted by | |
6d77b53e | 17169 | ASM_DECLARE_FUNCTION_NAME. */ |
d5b7b3ae RE |
17170 | asm_fprintf (f, "\torr\t%r, %r, #1\n", IP_REGNUM, PC_REGNUM); |
17171 | asm_fprintf (f, "\tbx\t%r\n", IP_REGNUM); | |
17172 | ||
17173 | /* Generate a label, so that the debugger will notice the | |
17174 | change in instruction sets. This label is also used by | |
17175 | the assembler to bypass the ARM code when this function | |
17176 | is called from a Thumb encoded function elsewhere in the | |
17177 | same file. Hence the definition of STUB_NAME here must | |
d6b4baa4 | 17178 | agree with the definition in gas/config/tc-arm.c. */ |
f676971a | 17179 | |
d5b7b3ae | 17180 | #define STUB_NAME ".real_start_of" |
f676971a | 17181 | |
761c70aa | 17182 | fprintf (f, "\t.code\t16\n"); |
d5b7b3ae RE |
17183 | #ifdef ARM_PE |
17184 | if (arm_dllexport_name_p (name)) | |
e5951263 | 17185 | name = arm_strip_name_encoding (name); |
f676971a | 17186 | #endif |
d5b7b3ae | 17187 | asm_fprintf (f, "\t.globl %s%U%s\n", STUB_NAME, name); |
761c70aa | 17188 | fprintf (f, "\t.thumb_func\n"); |
d5b7b3ae RE |
17189 | asm_fprintf (f, "%s%U%s:\n", STUB_NAME, name); |
17190 | } | |
f676971a | 17191 | |
38173d38 | 17192 | if (crtl->args.pretend_args_size) |
d5b7b3ae | 17193 | { |
617a1b71 PB |
17194 | /* Output unwind directive for the stack adjustment. */ |
17195 | if (ARM_EABI_UNWIND_TABLES) | |
17196 | fprintf (f, "\t.pad #%d\n", | |
38173d38 | 17197 | crtl->args.pretend_args_size); |
617a1b71 | 17198 | |
3cb66fd7 | 17199 | if (cfun->machine->uses_anonymous_args) |
d5b7b3ae RE |
17200 | { |
17201 | int num_pushes; | |
f676971a | 17202 | |
761c70aa | 17203 | fprintf (f, "\tpush\t{"); |
d5b7b3ae | 17204 | |
38173d38 | 17205 | num_pushes = ARM_NUM_INTS (crtl->args.pretend_args_size); |
f676971a | 17206 | |
d5b7b3ae RE |
17207 | for (regno = LAST_ARG_REGNUM + 1 - num_pushes; |
17208 | regno <= LAST_ARG_REGNUM; | |
5895f793 | 17209 | regno++) |
d5b7b3ae RE |
17210 | asm_fprintf (f, "%r%s", regno, |
17211 | regno == LAST_ARG_REGNUM ? "" : ", "); | |
17212 | ||
761c70aa | 17213 | fprintf (f, "}\n"); |
d5b7b3ae RE |
17214 | } |
17215 | else | |
f676971a | 17216 | asm_fprintf (f, "\tsub\t%r, %r, #%d\n", |
d5b7b3ae | 17217 | SP_REGNUM, SP_REGNUM, |
38173d38 | 17218 | crtl->args.pretend_args_size); |
980e61bb DJ |
17219 | |
17220 | /* We don't need to record the stores for unwinding (would it | |
17221 | help the debugger any if we did?), but record the change in | |
17222 | the stack pointer. */ | |
17223 | if (dwarf2out_do_frame ()) | |
17224 | { | |
17225 | char *l = dwarf2out_cfi_label (); | |
b279b20a | 17226 | |
38173d38 | 17227 | cfa_offset = cfa_offset + crtl->args.pretend_args_size; |
980e61bb DJ |
17228 | dwarf2out_def_cfa (l, SP_REGNUM, cfa_offset); |
17229 | } | |
d5b7b3ae RE |
17230 | } |
17231 | ||
b279b20a | 17232 | /* Get the registers we are going to push. */ |
954954d1 PB |
17233 | offsets = arm_get_frame_offsets (); |
17234 | live_regs_mask = offsets->saved_regs_mask; | |
b279b20a | 17235 | /* Extract a mask of the ones we can give to the Thumb's push instruction. */ |
57934c39 | 17236 | l_mask = live_regs_mask & 0x40ff; |
b279b20a NC |
17237 | /* Then count how many other high registers will need to be pushed. */ |
17238 | high_regs_pushed = bit_count (live_regs_mask & 0x0f00); | |
d5b7b3ae RE |
17239 | |
17240 | if (TARGET_BACKTRACE) | |
17241 | { | |
b279b20a NC |
17242 | unsigned offset; |
17243 | unsigned work_register; | |
f676971a | 17244 | |
d5b7b3ae RE |
17245 | /* We have been asked to create a stack backtrace structure. |
17246 | The code looks like this: | |
f676971a | 17247 | |
d5b7b3ae RE |
17248 | 0 .align 2 |
17249 | 0 func: | |
17250 | 0 sub SP, #16 Reserve space for 4 registers. | |
57934c39 | 17251 | 2 push {R7} Push low registers. |
d5b7b3ae RE |
17252 | 4 add R7, SP, #20 Get the stack pointer before the push. |
17253 | 6 str R7, [SP, #8] Store the stack pointer (before reserving the space). | |
17254 | 8 mov R7, PC Get hold of the start of this code plus 12. | |
17255 | 10 str R7, [SP, #16] Store it. | |
17256 | 12 mov R7, FP Get hold of the current frame pointer. | |
17257 | 14 str R7, [SP, #4] Store it. | |
17258 | 16 mov R7, LR Get hold of the current return address. | |
17259 | 18 str R7, [SP, #12] Store it. | |
17260 | 20 add R7, SP, #16 Point at the start of the backtrace structure. | |
17261 | 22 mov FP, R7 Put this value into the frame pointer. */ | |
17262 | ||
57934c39 | 17263 | work_register = thumb_find_work_register (live_regs_mask); |
f676971a | 17264 | |
617a1b71 PB |
17265 | if (ARM_EABI_UNWIND_TABLES) |
17266 | asm_fprintf (f, "\t.pad #16\n"); | |
17267 | ||
d5b7b3ae RE |
17268 | asm_fprintf |
17269 | (f, "\tsub\t%r, %r, #16\t%@ Create stack backtrace structure\n", | |
17270 | SP_REGNUM, SP_REGNUM); | |
980e61bb DJ |
17271 | |
17272 | if (dwarf2out_do_frame ()) | |
17273 | { | |
17274 | char *l = dwarf2out_cfi_label (); | |
b279b20a | 17275 | |
980e61bb DJ |
17276 | cfa_offset = cfa_offset + 16; |
17277 | dwarf2out_def_cfa (l, SP_REGNUM, cfa_offset); | |
17278 | } | |
17279 | ||
57934c39 PB |
17280 | if (l_mask) |
17281 | { | |
17282 | thumb_pushpop (f, l_mask, 1, &cfa_offset, l_mask); | |
1a59548b | 17283 | offset = bit_count (l_mask) * UNITS_PER_WORD; |
57934c39 PB |
17284 | } |
17285 | else | |
17286 | offset = 0; | |
f676971a | 17287 | |
d5b7b3ae | 17288 | asm_fprintf (f, "\tadd\t%r, %r, #%d\n", work_register, SP_REGNUM, |
38173d38 | 17289 | offset + 16 + crtl->args.pretend_args_size); |
f676971a | 17290 | |
d5b7b3ae RE |
17291 | asm_fprintf (f, "\tstr\t%r, [%r, #%d]\n", work_register, SP_REGNUM, |
17292 | offset + 4); | |
17293 | ||
17294 | /* Make sure that the instruction fetching the PC is in the right place | |
17295 | to calculate "start of backtrace creation code + 12". */ | |
57934c39 | 17296 | if (l_mask) |
d5b7b3ae RE |
17297 | { |
17298 | asm_fprintf (f, "\tmov\t%r, %r\n", work_register, PC_REGNUM); | |
17299 | asm_fprintf (f, "\tstr\t%r, [%r, #%d]\n", work_register, SP_REGNUM, | |
17300 | offset + 12); | |
17301 | asm_fprintf (f, "\tmov\t%r, %r\n", work_register, | |
17302 | ARM_HARD_FRAME_POINTER_REGNUM); | |
17303 | asm_fprintf (f, "\tstr\t%r, [%r, #%d]\n", work_register, SP_REGNUM, | |
17304 | offset); | |
17305 | } | |
17306 | else | |
17307 | { | |
17308 | asm_fprintf (f, "\tmov\t%r, %r\n", work_register, | |
17309 | ARM_HARD_FRAME_POINTER_REGNUM); | |
17310 | asm_fprintf (f, "\tstr\t%r, [%r, #%d]\n", work_register, SP_REGNUM, | |
17311 | offset); | |
17312 | asm_fprintf (f, "\tmov\t%r, %r\n", work_register, PC_REGNUM); | |
17313 | asm_fprintf (f, "\tstr\t%r, [%r, #%d]\n", work_register, SP_REGNUM, | |
17314 | offset + 12); | |
17315 | } | |
f676971a | 17316 | |
d5b7b3ae RE |
17317 | asm_fprintf (f, "\tmov\t%r, %r\n", work_register, LR_REGNUM); |
17318 | asm_fprintf (f, "\tstr\t%r, [%r, #%d]\n", work_register, SP_REGNUM, | |
17319 | offset + 8); | |
17320 | asm_fprintf (f, "\tadd\t%r, %r, #%d\n", work_register, SP_REGNUM, | |
17321 | offset + 12); | |
17322 | asm_fprintf (f, "\tmov\t%r, %r\t\t%@ Backtrace structure created\n", | |
17323 | ARM_HARD_FRAME_POINTER_REGNUM, work_register); | |
17324 | } | |
0fa2e4df | 17325 | /* Optimization: If we are not pushing any low registers but we are going |
b279b20a NC |
17326 | to push some high registers then delay our first push. This will just |
17327 | be a push of LR and we can combine it with the push of the first high | |
17328 | register. */ | |
17329 | else if ((l_mask & 0xff) != 0 | |
17330 | || (high_regs_pushed == 0 && l_mask)) | |
57934c39 | 17331 | thumb_pushpop (f, l_mask, 1, &cfa_offset, l_mask); |
d5b7b3ae | 17332 | |
d5b7b3ae RE |
17333 | if (high_regs_pushed) |
17334 | { | |
b279b20a NC |
17335 | unsigned pushable_regs; |
17336 | unsigned next_hi_reg; | |
d5b7b3ae RE |
17337 | |
17338 | for (next_hi_reg = 12; next_hi_reg > LAST_LO_REGNUM; next_hi_reg--) | |
57934c39 | 17339 | if (live_regs_mask & (1 << next_hi_reg)) |
e26053d1 | 17340 | break; |
d5b7b3ae | 17341 | |
57934c39 | 17342 | pushable_regs = l_mask & 0xff; |
d5b7b3ae RE |
17343 | |
17344 | if (pushable_regs == 0) | |
57934c39 | 17345 | pushable_regs = 1 << thumb_find_work_register (live_regs_mask); |
d5b7b3ae RE |
17346 | |
17347 | while (high_regs_pushed > 0) | |
17348 | { | |
b279b20a | 17349 | unsigned long real_regs_mask = 0; |
980e61bb | 17350 | |
b279b20a | 17351 | for (regno = LAST_LO_REGNUM; regno >= 0; regno --) |
d5b7b3ae | 17352 | { |
57934c39 | 17353 | if (pushable_regs & (1 << regno)) |
d5b7b3ae RE |
17354 | { |
17355 | asm_fprintf (f, "\tmov\t%r, %r\n", regno, next_hi_reg); | |
f676971a | 17356 | |
b279b20a | 17357 | high_regs_pushed --; |
980e61bb | 17358 | real_regs_mask |= (1 << next_hi_reg); |
f676971a | 17359 | |
d5b7b3ae | 17360 | if (high_regs_pushed) |
aeaf4d25 | 17361 | { |
b279b20a NC |
17362 | for (next_hi_reg --; next_hi_reg > LAST_LO_REGNUM; |
17363 | next_hi_reg --) | |
57934c39 | 17364 | if (live_regs_mask & (1 << next_hi_reg)) |
d5b7b3ae | 17365 | break; |
aeaf4d25 | 17366 | } |
d5b7b3ae RE |
17367 | else |
17368 | { | |
57934c39 | 17369 | pushable_regs &= ~((1 << regno) - 1); |
d5b7b3ae RE |
17370 | break; |
17371 | } | |
17372 | } | |
17373 | } | |
980e61bb | 17374 | |
b279b20a NC |
17375 | /* If we had to find a work register and we have not yet |
17376 | saved the LR then add it to the list of regs to push. */ | |
17377 | if (l_mask == (1 << LR_REGNUM)) | |
17378 | { | |
17379 | thumb_pushpop (f, pushable_regs | (1 << LR_REGNUM), | |
17380 | 1, &cfa_offset, | |
17381 | real_regs_mask | (1 << LR_REGNUM)); | |
17382 | l_mask = 0; | |
17383 | } | |
17384 | else | |
17385 | thumb_pushpop (f, pushable_regs, 1, &cfa_offset, real_regs_mask); | |
d5b7b3ae | 17386 | } |
d5b7b3ae RE |
17387 | } |
17388 | } | |
17389 | ||
17390 | /* Handle the case of a double word load into a low register from | |
17391 | a computed memory address. The computed address may involve a | |
17392 | register which is overwritten by the load. */ | |
cd2b33d0 | 17393 | const char * |
e32bac5b | 17394 | thumb_load_double_from_address (rtx *operands) |
d5b7b3ae RE |
17395 | { |
17396 | rtx addr; | |
17397 | rtx base; | |
17398 | rtx offset; | |
17399 | rtx arg1; | |
17400 | rtx arg2; | |
f676971a | 17401 | |
e6d29d15 NS |
17402 | gcc_assert (GET_CODE (operands[0]) == REG); |
17403 | gcc_assert (GET_CODE (operands[1]) == MEM); | |
d5b7b3ae RE |
17404 | |
17405 | /* Get the memory address. */ | |
17406 | addr = XEXP (operands[1], 0); | |
f676971a | 17407 | |
d5b7b3ae RE |
17408 | /* Work out how the memory address is computed. */ |
17409 | switch (GET_CODE (addr)) | |
17410 | { | |
17411 | case REG: | |
31fa16b6 | 17412 | operands[2] = adjust_address (operands[1], SImode, 4); |
e0b92319 | 17413 | |
d5b7b3ae RE |
17414 | if (REGNO (operands[0]) == REGNO (addr)) |
17415 | { | |
17416 | output_asm_insn ("ldr\t%H0, %2", operands); | |
17417 | output_asm_insn ("ldr\t%0, %1", operands); | |
17418 | } | |
17419 | else | |
17420 | { | |
17421 | output_asm_insn ("ldr\t%0, %1", operands); | |
17422 | output_asm_insn ("ldr\t%H0, %2", operands); | |
17423 | } | |
17424 | break; | |
f676971a | 17425 | |
d5b7b3ae RE |
17426 | case CONST: |
17427 | /* Compute <address> + 4 for the high order load. */ | |
31fa16b6 | 17428 | operands[2] = adjust_address (operands[1], SImode, 4); |
e0b92319 | 17429 | |
d5b7b3ae RE |
17430 | output_asm_insn ("ldr\t%0, %1", operands); |
17431 | output_asm_insn ("ldr\t%H0, %2", operands); | |
17432 | break; | |
f676971a | 17433 | |
d5b7b3ae RE |
17434 | case PLUS: |
17435 | arg1 = XEXP (addr, 0); | |
17436 | arg2 = XEXP (addr, 1); | |
f676971a | 17437 | |
d5b7b3ae RE |
17438 | if (CONSTANT_P (arg1)) |
17439 | base = arg2, offset = arg1; | |
17440 | else | |
17441 | base = arg1, offset = arg2; | |
f676971a | 17442 | |
e6d29d15 | 17443 | gcc_assert (GET_CODE (base) == REG); |
d5b7b3ae RE |
17444 | |
17445 | /* Catch the case of <address> = <reg> + <reg> */ | |
17446 | if (GET_CODE (offset) == REG) | |
17447 | { | |
17448 | int reg_offset = REGNO (offset); | |
17449 | int reg_base = REGNO (base); | |
17450 | int reg_dest = REGNO (operands[0]); | |
f676971a | 17451 | |
d5b7b3ae RE |
17452 | /* Add the base and offset registers together into the |
17453 | higher destination register. */ | |
17454 | asm_fprintf (asm_out_file, "\tadd\t%r, %r, %r", | |
17455 | reg_dest + 1, reg_base, reg_offset); | |
f676971a | 17456 | |
d5b7b3ae RE |
17457 | /* Load the lower destination register from the address in |
17458 | the higher destination register. */ | |
17459 | asm_fprintf (asm_out_file, "\tldr\t%r, [%r, #0]", | |
17460 | reg_dest, reg_dest + 1); | |
f676971a | 17461 | |
d5b7b3ae RE |
17462 | /* Load the higher destination register from its own address |
17463 | plus 4. */ | |
17464 | asm_fprintf (asm_out_file, "\tldr\t%r, [%r, #4]", | |
17465 | reg_dest + 1, reg_dest + 1); | |
17466 | } | |
17467 | else | |
17468 | { | |
17469 | /* Compute <address> + 4 for the high order load. */ | |
31fa16b6 | 17470 | operands[2] = adjust_address (operands[1], SImode, 4); |
f676971a | 17471 | |
d5b7b3ae RE |
17472 | /* If the computed address is held in the low order register |
17473 | then load the high order register first, otherwise always | |
17474 | load the low order register first. */ | |
17475 | if (REGNO (operands[0]) == REGNO (base)) | |
17476 | { | |
17477 | output_asm_insn ("ldr\t%H0, %2", operands); | |
17478 | output_asm_insn ("ldr\t%0, %1", operands); | |
17479 | } | |
17480 | else | |
17481 | { | |
17482 | output_asm_insn ("ldr\t%0, %1", operands); | |
17483 | output_asm_insn ("ldr\t%H0, %2", operands); | |
17484 | } | |
17485 | } | |
17486 | break; | |
17487 | ||
17488 | case LABEL_REF: | |
17489 | /* With no registers to worry about we can just load the value | |
17490 | directly. */ | |
31fa16b6 | 17491 | operands[2] = adjust_address (operands[1], SImode, 4); |
f676971a | 17492 | |
d5b7b3ae RE |
17493 | output_asm_insn ("ldr\t%H0, %2", operands); |
17494 | output_asm_insn ("ldr\t%0, %1", operands); | |
17495 | break; | |
f676971a | 17496 | |
d5b7b3ae | 17497 | default: |
e6d29d15 | 17498 | gcc_unreachable (); |
d5b7b3ae | 17499 | } |
f676971a | 17500 | |
d5b7b3ae RE |
17501 | return ""; |
17502 | } | |
17503 | ||
cd2b33d0 | 17504 | const char * |
e32bac5b | 17505 | thumb_output_move_mem_multiple (int n, rtx *operands) |
d5b7b3ae RE |
17506 | { |
17507 | rtx tmp; | |
17508 | ||
17509 | switch (n) | |
17510 | { | |
17511 | case 2: | |
ca356f3a | 17512 | if (REGNO (operands[4]) > REGNO (operands[5])) |
d5b7b3ae | 17513 | { |
ca356f3a RE |
17514 | tmp = operands[4]; |
17515 | operands[4] = operands[5]; | |
17516 | operands[5] = tmp; | |
d5b7b3ae | 17517 | } |
ca356f3a RE |
17518 | output_asm_insn ("ldmia\t%1!, {%4, %5}", operands); |
17519 | output_asm_insn ("stmia\t%0!, {%4, %5}", operands); | |
d5b7b3ae RE |
17520 | break; |
17521 | ||
17522 | case 3: | |
ca356f3a | 17523 | if (REGNO (operands[4]) > REGNO (operands[5])) |
d5b7b3ae | 17524 | { |
ca356f3a RE |
17525 | tmp = operands[4]; |
17526 | operands[4] = operands[5]; | |
17527 | operands[5] = tmp; | |
d5b7b3ae | 17528 | } |
ca356f3a | 17529 | if (REGNO (operands[5]) > REGNO (operands[6])) |
d5b7b3ae | 17530 | { |
ca356f3a RE |
17531 | tmp = operands[5]; |
17532 | operands[5] = operands[6]; | |
17533 | operands[6] = tmp; | |
d5b7b3ae | 17534 | } |
ca356f3a | 17535 | if (REGNO (operands[4]) > REGNO (operands[5])) |
d5b7b3ae | 17536 | { |
ca356f3a RE |
17537 | tmp = operands[4]; |
17538 | operands[4] = operands[5]; | |
17539 | operands[5] = tmp; | |
d5b7b3ae | 17540 | } |
f676971a | 17541 | |
ca356f3a RE |
17542 | output_asm_insn ("ldmia\t%1!, {%4, %5, %6}", operands); |
17543 | output_asm_insn ("stmia\t%0!, {%4, %5, %6}", operands); | |
d5b7b3ae RE |
17544 | break; |
17545 | ||
17546 | default: | |
e6d29d15 | 17547 | gcc_unreachable (); |
d5b7b3ae RE |
17548 | } |
17549 | ||
17550 | return ""; | |
17551 | } | |
17552 | ||
b12a00f1 RE |
17553 | /* Output a call-via instruction for thumb state. */ |
17554 | const char * | |
17555 | thumb_call_via_reg (rtx reg) | |
17556 | { | |
17557 | int regno = REGNO (reg); | |
17558 | rtx *labelp; | |
17559 | ||
57ecec57 | 17560 | gcc_assert (regno < LR_REGNUM); |
b12a00f1 RE |
17561 | |
17562 | /* If we are in the normal text section we can use a single instance | |
17563 | per compilation unit. If we are doing function sections, then we need | |
17564 | an entry per section, since we can't rely on reachability. */ | |
d6b5193b | 17565 | if (in_section == text_section) |
b12a00f1 RE |
17566 | { |
17567 | thumb_call_reg_needed = 1; | |
17568 | ||
17569 | if (thumb_call_via_label[regno] == NULL) | |
17570 | thumb_call_via_label[regno] = gen_label_rtx (); | |
17571 | labelp = thumb_call_via_label + regno; | |
17572 | } | |
17573 | else | |
17574 | { | |
17575 | if (cfun->machine->call_via[regno] == NULL) | |
17576 | cfun->machine->call_via[regno] = gen_label_rtx (); | |
17577 | labelp = cfun->machine->call_via + regno; | |
17578 | } | |
17579 | ||
17580 | output_asm_insn ("bl\t%a0", labelp); | |
17581 | return ""; | |
17582 | } | |
17583 | ||
1d6e90ac | 17584 | /* Routines for generating rtl. */ |
d5b7b3ae | 17585 | void |
70128ad9 | 17586 | thumb_expand_movmemqi (rtx *operands) |
d5b7b3ae RE |
17587 | { |
17588 | rtx out = copy_to_mode_reg (SImode, XEXP (operands[0], 0)); | |
17589 | rtx in = copy_to_mode_reg (SImode, XEXP (operands[1], 0)); | |
17590 | HOST_WIDE_INT len = INTVAL (operands[2]); | |
17591 | HOST_WIDE_INT offset = 0; | |
17592 | ||
17593 | while (len >= 12) | |
17594 | { | |
ca356f3a | 17595 | emit_insn (gen_movmem12b (out, in, out, in)); |
d5b7b3ae RE |
17596 | len -= 12; |
17597 | } | |
f676971a | 17598 | |
d5b7b3ae RE |
17599 | if (len >= 8) |
17600 | { | |
ca356f3a | 17601 | emit_insn (gen_movmem8b (out, in, out, in)); |
d5b7b3ae RE |
17602 | len -= 8; |
17603 | } | |
f676971a | 17604 | |
d5b7b3ae RE |
17605 | if (len >= 4) |
17606 | { | |
17607 | rtx reg = gen_reg_rtx (SImode); | |
f1c25d3b KH |
17608 | emit_insn (gen_movsi (reg, gen_rtx_MEM (SImode, in))); |
17609 | emit_insn (gen_movsi (gen_rtx_MEM (SImode, out), reg)); | |
d5b7b3ae RE |
17610 | len -= 4; |
17611 | offset += 4; | |
17612 | } | |
f676971a | 17613 | |
d5b7b3ae RE |
17614 | if (len >= 2) |
17615 | { | |
17616 | rtx reg = gen_reg_rtx (HImode); | |
f676971a | 17617 | emit_insn (gen_movhi (reg, gen_rtx_MEM (HImode, |
f1c25d3b KH |
17618 | plus_constant (in, offset)))); |
17619 | emit_insn (gen_movhi (gen_rtx_MEM (HImode, plus_constant (out, offset)), | |
d5b7b3ae RE |
17620 | reg)); |
17621 | len -= 2; | |
17622 | offset += 2; | |
17623 | } | |
f676971a | 17624 | |
d5b7b3ae RE |
17625 | if (len) |
17626 | { | |
17627 | rtx reg = gen_reg_rtx (QImode); | |
f1c25d3b KH |
17628 | emit_insn (gen_movqi (reg, gen_rtx_MEM (QImode, |
17629 | plus_constant (in, offset)))); | |
17630 | emit_insn (gen_movqi (gen_rtx_MEM (QImode, plus_constant (out, offset)), | |
d5b7b3ae RE |
17631 | reg)); |
17632 | } | |
17633 | } | |
17634 | ||
d5b7b3ae | 17635 | void |
e32bac5b | 17636 | thumb_reload_out_hi (rtx *operands) |
d5b7b3ae RE |
17637 | { |
17638 | emit_insn (gen_thumb_movhi_clobber (operands[0], operands[1], operands[2])); | |
17639 | } | |
17640 | ||
f676971a | 17641 | /* Handle reading a half-word from memory during reload. */ |
d5b7b3ae | 17642 | void |
e32bac5b | 17643 | thumb_reload_in_hi (rtx *operands ATTRIBUTE_UNUSED) |
d5b7b3ae | 17644 | { |
e6d29d15 | 17645 | gcc_unreachable (); |
d5b7b3ae RE |
17646 | } |
17647 | ||
c27ba912 DM |
17648 | /* Return the length of a function name prefix |
17649 | that starts with the character 'c'. */ | |
17650 | static int | |
e32bac5b | 17651 | arm_get_strip_length (int c) |
c27ba912 DM |
17652 | { |
17653 | switch (c) | |
17654 | { | |
17655 | ARM_NAME_ENCODING_LENGTHS | |
f676971a | 17656 | default: return 0; |
c27ba912 DM |
17657 | } |
17658 | } | |
17659 | ||
17660 | /* Return a pointer to a function's name with any | |
17661 | and all prefix encodings stripped from it. */ | |
17662 | const char * | |
e32bac5b | 17663 | arm_strip_name_encoding (const char *name) |
c27ba912 DM |
17664 | { |
17665 | int skip; | |
f676971a | 17666 | |
c27ba912 DM |
17667 | while ((skip = arm_get_strip_length (* name))) |
17668 | name += skip; | |
17669 | ||
17670 | return name; | |
17671 | } | |
17672 | ||
e1944073 KW |
17673 | /* If there is a '*' anywhere in the name's prefix, then |
17674 | emit the stripped name verbatim, otherwise prepend an | |
17675 | underscore if leading underscores are being used. */ | |
e1944073 | 17676 | void |
e32bac5b | 17677 | arm_asm_output_labelref (FILE *stream, const char *name) |
e1944073 KW |
17678 | { |
17679 | int skip; | |
17680 | int verbatim = 0; | |
17681 | ||
17682 | while ((skip = arm_get_strip_length (* name))) | |
17683 | { | |
17684 | verbatim |= (*name == '*'); | |
17685 | name += skip; | |
17686 | } | |
17687 | ||
17688 | if (verbatim) | |
17689 | fputs (name, stream); | |
17690 | else | |
17691 | asm_fprintf (stream, "%U%s", name); | |
17692 | } | |
17693 | ||
6c6aa1af PB |
17694 | static void |
17695 | arm_file_start (void) | |
17696 | { | |
17697 | int val; | |
17698 | ||
5b3e6663 PB |
17699 | if (TARGET_UNIFIED_ASM) |
17700 | asm_fprintf (asm_out_file, "\t.syntax unified\n"); | |
17701 | ||
6c6aa1af PB |
17702 | if (TARGET_BPABI) |
17703 | { | |
17704 | const char *fpu_name; | |
17705 | if (arm_select[0].string) | |
17706 | asm_fprintf (asm_out_file, "\t.cpu %s\n", arm_select[0].string); | |
17707 | else if (arm_select[1].string) | |
17708 | asm_fprintf (asm_out_file, "\t.arch %s\n", arm_select[1].string); | |
17709 | else | |
17710 | asm_fprintf (asm_out_file, "\t.cpu %s\n", | |
17711 | all_cores[arm_default_cpu].name); | |
17712 | ||
17713 | if (TARGET_SOFT_FLOAT) | |
17714 | { | |
17715 | if (TARGET_VFP) | |
17716 | fpu_name = "softvfp"; | |
17717 | else | |
17718 | fpu_name = "softfpa"; | |
17719 | } | |
17720 | else | |
17721 | { | |
f1adb0a9 | 17722 | int set_float_abi_attributes = 0; |
6c6aa1af PB |
17723 | switch (arm_fpu_arch) |
17724 | { | |
17725 | case FPUTYPE_FPA: | |
17726 | fpu_name = "fpa"; | |
17727 | break; | |
17728 | case FPUTYPE_FPA_EMU2: | |
17729 | fpu_name = "fpe2"; | |
17730 | break; | |
17731 | case FPUTYPE_FPA_EMU3: | |
17732 | fpu_name = "fpe3"; | |
17733 | break; | |
17734 | case FPUTYPE_MAVERICK: | |
17735 | fpu_name = "maverick"; | |
17736 | break; | |
17737 | case FPUTYPE_VFP: | |
6c6aa1af | 17738 | fpu_name = "vfp"; |
f1adb0a9 JB |
17739 | set_float_abi_attributes = 1; |
17740 | break; | |
17741 | case FPUTYPE_VFP3: | |
17742 | fpu_name = "vfp3"; | |
17743 | set_float_abi_attributes = 1; | |
6c6aa1af | 17744 | break; |
88f77cba JB |
17745 | case FPUTYPE_NEON: |
17746 | fpu_name = "neon"; | |
17747 | set_float_abi_attributes = 1; | |
17748 | break; | |
6c6aa1af PB |
17749 | default: |
17750 | abort(); | |
17751 | } | |
f1adb0a9 JB |
17752 | if (set_float_abi_attributes) |
17753 | { | |
17754 | if (TARGET_HARD_FLOAT) | |
17755 | asm_fprintf (asm_out_file, "\t.eabi_attribute 27, 3\n"); | |
17756 | if (TARGET_HARD_FLOAT_ABI) | |
17757 | asm_fprintf (asm_out_file, "\t.eabi_attribute 28, 1\n"); | |
17758 | } | |
6c6aa1af PB |
17759 | } |
17760 | asm_fprintf (asm_out_file, "\t.fpu %s\n", fpu_name); | |
17761 | ||
17762 | /* Some of these attributes only apply when the corresponding features | |
17763 | are used. However we don't have any easy way of figuring this out. | |
17764 | Conservatively record the setting that would have been used. */ | |
17765 | ||
6c6aa1af PB |
17766 | /* Tag_ABI_FP_rounding. */ |
17767 | if (flag_rounding_math) | |
17768 | asm_fprintf (asm_out_file, "\t.eabi_attribute 19, 1\n"); | |
17769 | if (!flag_unsafe_math_optimizations) | |
17770 | { | |
17771 | /* Tag_ABI_FP_denomal. */ | |
17772 | asm_fprintf (asm_out_file, "\t.eabi_attribute 20, 1\n"); | |
17773 | /* Tag_ABI_FP_exceptions. */ | |
17774 | asm_fprintf (asm_out_file, "\t.eabi_attribute 21, 1\n"); | |
17775 | } | |
17776 | /* Tag_ABI_FP_user_exceptions. */ | |
17777 | if (flag_signaling_nans) | |
17778 | asm_fprintf (asm_out_file, "\t.eabi_attribute 22, 1\n"); | |
17779 | /* Tag_ABI_FP_number_model. */ | |
17780 | asm_fprintf (asm_out_file, "\t.eabi_attribute 23, %d\n", | |
17781 | flag_finite_math_only ? 1 : 3); | |
17782 | ||
17783 | /* Tag_ABI_align8_needed. */ | |
17784 | asm_fprintf (asm_out_file, "\t.eabi_attribute 24, 1\n"); | |
17785 | /* Tag_ABI_align8_preserved. */ | |
17786 | asm_fprintf (asm_out_file, "\t.eabi_attribute 25, 1\n"); | |
17787 | /* Tag_ABI_enum_size. */ | |
17788 | asm_fprintf (asm_out_file, "\t.eabi_attribute 26, %d\n", | |
17789 | flag_short_enums ? 1 : 2); | |
17790 | ||
17791 | /* Tag_ABI_optimization_goals. */ | |
17792 | if (optimize_size) | |
17793 | val = 4; | |
17794 | else if (optimize >= 2) | |
17795 | val = 2; | |
17796 | else if (optimize) | |
17797 | val = 1; | |
17798 | else | |
17799 | val = 6; | |
17800 | asm_fprintf (asm_out_file, "\t.eabi_attribute 30, %d\n", val); | |
b76c3c4b PB |
17801 | |
17802 | if (arm_lang_output_object_attributes_hook) | |
17803 | arm_lang_output_object_attributes_hook(); | |
6c6aa1af PB |
17804 | } |
17805 | default_file_start(); | |
17806 | } | |
17807 | ||
b12a00f1 RE |
17808 | static void |
17809 | arm_file_end (void) | |
17810 | { | |
17811 | int regno; | |
17812 | ||
978e411f CD |
17813 | if (NEED_INDICATE_EXEC_STACK) |
17814 | /* Add .note.GNU-stack. */ | |
17815 | file_end_indicate_exec_stack (); | |
17816 | ||
b12a00f1 RE |
17817 | if (! thumb_call_reg_needed) |
17818 | return; | |
17819 | ||
d6b5193b | 17820 | switch_to_section (text_section); |
b12a00f1 RE |
17821 | asm_fprintf (asm_out_file, "\t.code 16\n"); |
17822 | ASM_OUTPUT_ALIGN (asm_out_file, 1); | |
17823 | ||
57ecec57 | 17824 | for (regno = 0; regno < LR_REGNUM; regno++) |
b12a00f1 RE |
17825 | { |
17826 | rtx label = thumb_call_via_label[regno]; | |
17827 | ||
17828 | if (label != 0) | |
17829 | { | |
17830 | targetm.asm_out.internal_label (asm_out_file, "L", | |
17831 | CODE_LABEL_NUMBER (label)); | |
17832 | asm_fprintf (asm_out_file, "\tbx\t%r\n", regno); | |
17833 | } | |
17834 | } | |
17835 | } | |
17836 | ||
fb49053f RH |
17837 | #ifndef ARM_PE |
17838 | /* Symbols in the text segment can be accessed without indirecting via the | |
17839 | constant pool; it may take an extra binary operation, but this is still | |
17840 | faster than indirecting via memory. Don't do this when not optimizing, | |
17841 | since we won't be calculating al of the offsets necessary to do this | |
17842 | simplification. */ | |
17843 | ||
17844 | static void | |
e32bac5b | 17845 | arm_encode_section_info (tree decl, rtx rtl, int first) |
fb49053f | 17846 | { |
3521b33c | 17847 | if (optimize > 0 && TREE_CONSTANT (decl)) |
c6a2438a | 17848 | SYMBOL_REF_FLAG (XEXP (rtl, 0)) = 1; |
fb49053f | 17849 | |
d3585b76 | 17850 | default_encode_section_info (decl, rtl, first); |
fb49053f RH |
17851 | } |
17852 | #endif /* !ARM_PE */ | |
483ab821 | 17853 | |
4977bab6 | 17854 | static void |
e32bac5b | 17855 | arm_internal_label (FILE *stream, const char *prefix, unsigned long labelno) |
4977bab6 ZW |
17856 | { |
17857 | if (arm_ccfsm_state == 3 && (unsigned) arm_target_label == labelno | |
17858 | && !strcmp (prefix, "L")) | |
17859 | { | |
17860 | arm_ccfsm_state = 0; | |
17861 | arm_target_insn = NULL; | |
17862 | } | |
17863 | default_internal_label (stream, prefix, labelno); | |
17864 | } | |
17865 | ||
c590b625 RH |
17866 | /* Output code to add DELTA to the first argument, and then jump |
17867 | to FUNCTION. Used for C++ multiple inheritance. */ | |
c590b625 | 17868 | static void |
e32bac5b RE |
17869 | arm_output_mi_thunk (FILE *file, tree thunk ATTRIBUTE_UNUSED, |
17870 | HOST_WIDE_INT delta, | |
17871 | HOST_WIDE_INT vcall_offset ATTRIBUTE_UNUSED, | |
17872 | tree function) | |
483ab821 | 17873 | { |
9b66ebb1 PB |
17874 | static int thunk_label = 0; |
17875 | char label[256]; | |
54b9e939 | 17876 | char labelpc[256]; |
483ab821 MM |
17877 | int mi_delta = delta; |
17878 | const char *const mi_op = mi_delta < 0 ? "sub" : "add"; | |
17879 | int shift = 0; | |
61f71b34 | 17880 | int this_regno = (aggregate_value_p (TREE_TYPE (TREE_TYPE (function)), function) |
483ab821 MM |
17881 | ? 1 : 0); |
17882 | if (mi_delta < 0) | |
17883 | mi_delta = - mi_delta; | |
bf98ec6c | 17884 | |
5b3e6663 | 17885 | if (TARGET_THUMB1) |
9b66ebb1 PB |
17886 | { |
17887 | int labelno = thunk_label++; | |
17888 | ASM_GENERATE_INTERNAL_LABEL (label, "LTHUMBFUNC", labelno); | |
bf98ec6c PB |
17889 | /* Thunks are entered in arm mode when avaiable. */ |
17890 | if (TARGET_THUMB1_ONLY) | |
17891 | { | |
17892 | /* push r3 so we can use it as a temporary. */ | |
17893 | /* TODO: Omit this save if r3 is not used. */ | |
17894 | fputs ("\tpush {r3}\n", file); | |
17895 | fputs ("\tldr\tr3, ", file); | |
17896 | } | |
17897 | else | |
17898 | { | |
17899 | fputs ("\tldr\tr12, ", file); | |
17900 | } | |
9b66ebb1 PB |
17901 | assemble_name (file, label); |
17902 | fputc ('\n', file); | |
54b9e939 KH |
17903 | if (flag_pic) |
17904 | { | |
17905 | /* If we are generating PIC, the ldr instruction below loads | |
17906 | "(target - 7) - .LTHUNKPCn" into r12. The pc reads as | |
17907 | the address of the add + 8, so we have: | |
17908 | ||
17909 | r12 = (target - 7) - .LTHUNKPCn + (.LTHUNKPCn + 8) | |
17910 | = target + 1. | |
17911 | ||
17912 | Note that we have "+ 1" because some versions of GNU ld | |
17913 | don't set the low bit of the result for R_ARM_REL32 | |
bf98ec6c PB |
17914 | relocations against thumb function symbols. |
17915 | On ARMv6M this is +4, not +8. */ | |
54b9e939 KH |
17916 | ASM_GENERATE_INTERNAL_LABEL (labelpc, "LTHUNKPC", labelno); |
17917 | assemble_name (file, labelpc); | |
17918 | fputs (":\n", file); | |
bf98ec6c PB |
17919 | if (TARGET_THUMB1_ONLY) |
17920 | { | |
17921 | /* This is 2 insns after the start of the thunk, so we know it | |
17922 | is 4-byte aligned. */ | |
17923 | fputs ("\tadd\tr3, pc, r3\n", file); | |
17924 | fputs ("\tmov r12, r3\n", file); | |
17925 | } | |
17926 | else | |
17927 | fputs ("\tadd\tr12, pc, r12\n", file); | |
54b9e939 | 17928 | } |
bf98ec6c PB |
17929 | else if (TARGET_THUMB1_ONLY) |
17930 | fputs ("\tmov r12, r3\n", file); | |
9b66ebb1 | 17931 | } |
bf98ec6c | 17932 | if (TARGET_THUMB1_ONLY) |
483ab821 | 17933 | { |
bf98ec6c PB |
17934 | if (mi_delta > 255) |
17935 | { | |
17936 | fputs ("\tldr\tr3, ", file); | |
17937 | assemble_name (file, label); | |
17938 | fputs ("+4\n", file); | |
17939 | asm_fprintf (file, "\t%s\t%r, %r, r3\n", | |
17940 | mi_op, this_regno, this_regno); | |
17941 | } | |
17942 | else if (mi_delta != 0) | |
17943 | { | |
17944 | asm_fprintf (file, "\t%s\t%r, %r, #%d\n", | |
17945 | mi_op, this_regno, this_regno, | |
17946 | mi_delta); | |
17947 | } | |
17948 | } | |
17949 | else | |
17950 | { | |
17951 | /* TODO: Use movw/movt for large constants when available. */ | |
17952 | while (mi_delta != 0) | |
17953 | { | |
17954 | if ((mi_delta & (3 << shift)) == 0) | |
17955 | shift += 2; | |
17956 | else | |
17957 | { | |
17958 | asm_fprintf (file, "\t%s\t%r, %r, #%d\n", | |
17959 | mi_op, this_regno, this_regno, | |
17960 | mi_delta & (0xff << shift)); | |
17961 | mi_delta &= ~(0xff << shift); | |
17962 | shift += 8; | |
17963 | } | |
17964 | } | |
483ab821 | 17965 | } |
5b3e6663 | 17966 | if (TARGET_THUMB1) |
9b66ebb1 | 17967 | { |
bf98ec6c PB |
17968 | if (TARGET_THUMB1_ONLY) |
17969 | fputs ("\tpop\t{r3}\n", file); | |
17970 | ||
9b66ebb1 PB |
17971 | fprintf (file, "\tbx\tr12\n"); |
17972 | ASM_OUTPUT_ALIGN (file, 2); | |
17973 | assemble_name (file, label); | |
17974 | fputs (":\n", file); | |
54b9e939 KH |
17975 | if (flag_pic) |
17976 | { | |
17977 | /* Output ".word .LTHUNKn-7-.LTHUNKPCn". */ | |
17978 | rtx tem = XEXP (DECL_RTL (function), 0); | |
17979 | tem = gen_rtx_PLUS (GET_MODE (tem), tem, GEN_INT (-7)); | |
17980 | tem = gen_rtx_MINUS (GET_MODE (tem), | |
17981 | tem, | |
17982 | gen_rtx_SYMBOL_REF (Pmode, | |
17983 | ggc_strdup (labelpc))); | |
17984 | assemble_integer (tem, 4, BITS_PER_WORD, 1); | |
17985 | } | |
17986 | else | |
17987 | /* Output ".word .LTHUNKn". */ | |
17988 | assemble_integer (XEXP (DECL_RTL (function), 0), 4, BITS_PER_WORD, 1); | |
bf98ec6c PB |
17989 | |
17990 | if (TARGET_THUMB1_ONLY && mi_delta > 255) | |
17991 | assemble_integer (GEN_INT(mi_delta), 4, BITS_PER_WORD, 1); | |
9b66ebb1 PB |
17992 | } |
17993 | else | |
17994 | { | |
17995 | fputs ("\tb\t", file); | |
17996 | assemble_name (file, XSTR (XEXP (DECL_RTL (function), 0), 0)); | |
17997 | if (NEED_PLT_RELOC) | |
17998 | fputs ("(PLT)", file); | |
17999 | fputc ('\n', file); | |
18000 | } | |
483ab821 | 18001 | } |
5a9335ef NC |
18002 | |
18003 | int | |
6f5f2481 | 18004 | arm_emit_vector_const (FILE *file, rtx x) |
5a9335ef NC |
18005 | { |
18006 | int i; | |
18007 | const char * pattern; | |
18008 | ||
e6d29d15 | 18009 | gcc_assert (GET_CODE (x) == CONST_VECTOR); |
5a9335ef NC |
18010 | |
18011 | switch (GET_MODE (x)) | |
18012 | { | |
18013 | case V2SImode: pattern = "%08x"; break; | |
18014 | case V4HImode: pattern = "%04x"; break; | |
18015 | case V8QImode: pattern = "%02x"; break; | |
e6d29d15 | 18016 | default: gcc_unreachable (); |
5a9335ef NC |
18017 | } |
18018 | ||
18019 | fprintf (file, "0x"); | |
18020 | for (i = CONST_VECTOR_NUNITS (x); i--;) | |
18021 | { | |
18022 | rtx element; | |
18023 | ||
18024 | element = CONST_VECTOR_ELT (x, i); | |
18025 | fprintf (file, pattern, INTVAL (element)); | |
18026 | } | |
18027 | ||
18028 | return 1; | |
18029 | } | |
18030 | ||
18031 | const char * | |
6f5f2481 | 18032 | arm_output_load_gr (rtx *operands) |
5a9335ef NC |
18033 | { |
18034 | rtx reg; | |
18035 | rtx offset; | |
18036 | rtx wcgr; | |
18037 | rtx sum; | |
f676971a | 18038 | |
5a9335ef NC |
18039 | if (GET_CODE (operands [1]) != MEM |
18040 | || GET_CODE (sum = XEXP (operands [1], 0)) != PLUS | |
18041 | || GET_CODE (reg = XEXP (sum, 0)) != REG | |
18042 | || GET_CODE (offset = XEXP (sum, 1)) != CONST_INT | |
18043 | || ((INTVAL (offset) < 1024) && (INTVAL (offset) > -1024))) | |
18044 | return "wldrw%?\t%0, %1"; | |
f676971a EC |
18045 | |
18046 | /* Fix up an out-of-range load of a GR register. */ | |
5a9335ef NC |
18047 | output_asm_insn ("str%?\t%0, [sp, #-4]!\t@ Start of GR load expansion", & reg); |
18048 | wcgr = operands[0]; | |
18049 | operands[0] = reg; | |
18050 | output_asm_insn ("ldr%?\t%0, %1", operands); | |
18051 | ||
18052 | operands[0] = wcgr; | |
18053 | operands[1] = reg; | |
18054 | output_asm_insn ("tmcr%?\t%0, %1", operands); | |
18055 | output_asm_insn ("ldr%?\t%0, [sp], #4\t@ End of GR load expansion", & reg); | |
18056 | ||
18057 | return ""; | |
18058 | } | |
f9ba5949 | 18059 | |
1cc9f5f5 KH |
18060 | /* Worker function for TARGET_SETUP_INCOMING_VARARGS. |
18061 | ||
18062 | On the ARM, PRETEND_SIZE is set in order to have the prologue push the last | |
18063 | named arg and all anonymous args onto the stack. | |
18064 | XXX I know the prologue shouldn't be pushing registers, but it is faster | |
18065 | that way. */ | |
18066 | ||
18067 | static void | |
18068 | arm_setup_incoming_varargs (CUMULATIVE_ARGS *cum, | |
22ccaaee JJ |
18069 | enum machine_mode mode, |
18070 | tree type, | |
1cc9f5f5 KH |
18071 | int *pretend_size, |
18072 | int second_time ATTRIBUTE_UNUSED) | |
18073 | { | |
22ccaaee JJ |
18074 | int nregs = cum->nregs; |
18075 | if (nregs & 1 | |
18076 | && ARM_DOUBLEWORD_ALIGN | |
18077 | && arm_needs_doubleword_align (mode, type)) | |
18078 | nregs++; | |
18079 | ||
1cc9f5f5 | 18080 | cfun->machine->uses_anonymous_args = 1; |
22ccaaee JJ |
18081 | if (nregs < NUM_ARG_REGS) |
18082 | *pretend_size = (NUM_ARG_REGS - nregs) * UNITS_PER_WORD; | |
1cc9f5f5 | 18083 | } |
9b66ebb1 | 18084 | |
59b9a953 | 18085 | /* Return nonzero if the CONSUMER instruction (a store) does not need |
9b66ebb1 PB |
18086 | PRODUCER's value to calculate the address. */ |
18087 | ||
18088 | int | |
18089 | arm_no_early_store_addr_dep (rtx producer, rtx consumer) | |
18090 | { | |
18091 | rtx value = PATTERN (producer); | |
18092 | rtx addr = PATTERN (consumer); | |
18093 | ||
18094 | if (GET_CODE (value) == COND_EXEC) | |
18095 | value = COND_EXEC_CODE (value); | |
18096 | if (GET_CODE (value) == PARALLEL) | |
18097 | value = XVECEXP (value, 0, 0); | |
18098 | value = XEXP (value, 0); | |
18099 | if (GET_CODE (addr) == COND_EXEC) | |
18100 | addr = COND_EXEC_CODE (addr); | |
18101 | if (GET_CODE (addr) == PARALLEL) | |
18102 | addr = XVECEXP (addr, 0, 0); | |
18103 | addr = XEXP (addr, 0); | |
f676971a | 18104 | |
9b66ebb1 PB |
18105 | return !reg_overlap_mentioned_p (value, addr); |
18106 | } | |
18107 | ||
59b9a953 | 18108 | /* Return nonzero if the CONSUMER instruction (an ALU op) does not |
9b66ebb1 PB |
18109 | have an early register shift value or amount dependency on the |
18110 | result of PRODUCER. */ | |
18111 | ||
18112 | int | |
18113 | arm_no_early_alu_shift_dep (rtx producer, rtx consumer) | |
18114 | { | |
18115 | rtx value = PATTERN (producer); | |
18116 | rtx op = PATTERN (consumer); | |
18117 | rtx early_op; | |
18118 | ||
18119 | if (GET_CODE (value) == COND_EXEC) | |
18120 | value = COND_EXEC_CODE (value); | |
18121 | if (GET_CODE (value) == PARALLEL) | |
18122 | value = XVECEXP (value, 0, 0); | |
18123 | value = XEXP (value, 0); | |
18124 | if (GET_CODE (op) == COND_EXEC) | |
18125 | op = COND_EXEC_CODE (op); | |
18126 | if (GET_CODE (op) == PARALLEL) | |
18127 | op = XVECEXP (op, 0, 0); | |
18128 | op = XEXP (op, 1); | |
f676971a | 18129 | |
9b66ebb1 PB |
18130 | early_op = XEXP (op, 0); |
18131 | /* This is either an actual independent shift, or a shift applied to | |
18132 | the first operand of another operation. We want the whole shift | |
18133 | operation. */ | |
18134 | if (GET_CODE (early_op) == REG) | |
18135 | early_op = op; | |
18136 | ||
18137 | return !reg_overlap_mentioned_p (value, early_op); | |
18138 | } | |
18139 | ||
59b9a953 | 18140 | /* Return nonzero if the CONSUMER instruction (an ALU op) does not |
9b66ebb1 PB |
18141 | have an early register shift value dependency on the result of |
18142 | PRODUCER. */ | |
18143 | ||
18144 | int | |
18145 | arm_no_early_alu_shift_value_dep (rtx producer, rtx consumer) | |
18146 | { | |
18147 | rtx value = PATTERN (producer); | |
18148 | rtx op = PATTERN (consumer); | |
18149 | rtx early_op; | |
18150 | ||
18151 | if (GET_CODE (value) == COND_EXEC) | |
18152 | value = COND_EXEC_CODE (value); | |
18153 | if (GET_CODE (value) == PARALLEL) | |
18154 | value = XVECEXP (value, 0, 0); | |
18155 | value = XEXP (value, 0); | |
18156 | if (GET_CODE (op) == COND_EXEC) | |
18157 | op = COND_EXEC_CODE (op); | |
18158 | if (GET_CODE (op) == PARALLEL) | |
18159 | op = XVECEXP (op, 0, 0); | |
18160 | op = XEXP (op, 1); | |
f676971a | 18161 | |
9b66ebb1 PB |
18162 | early_op = XEXP (op, 0); |
18163 | ||
18164 | /* This is either an actual independent shift, or a shift applied to | |
18165 | the first operand of another operation. We want the value being | |
18166 | shifted, in either case. */ | |
18167 | if (GET_CODE (early_op) != REG) | |
18168 | early_op = XEXP (early_op, 0); | |
f676971a | 18169 | |
9b66ebb1 PB |
18170 | return !reg_overlap_mentioned_p (value, early_op); |
18171 | } | |
18172 | ||
59b9a953 | 18173 | /* Return nonzero if the CONSUMER (a mul or mac op) does not |
9b66ebb1 PB |
18174 | have an early register mult dependency on the result of |
18175 | PRODUCER. */ | |
18176 | ||
18177 | int | |
18178 | arm_no_early_mul_dep (rtx producer, rtx consumer) | |
18179 | { | |
18180 | rtx value = PATTERN (producer); | |
18181 | rtx op = PATTERN (consumer); | |
18182 | ||
18183 | if (GET_CODE (value) == COND_EXEC) | |
18184 | value = COND_EXEC_CODE (value); | |
18185 | if (GET_CODE (value) == PARALLEL) | |
18186 | value = XVECEXP (value, 0, 0); | |
18187 | value = XEXP (value, 0); | |
18188 | if (GET_CODE (op) == COND_EXEC) | |
18189 | op = COND_EXEC_CODE (op); | |
18190 | if (GET_CODE (op) == PARALLEL) | |
18191 | op = XVECEXP (op, 0, 0); | |
18192 | op = XEXP (op, 1); | |
f676971a | 18193 | |
9b66ebb1 PB |
18194 | return (GET_CODE (op) == PLUS |
18195 | && !reg_overlap_mentioned_p (value, XEXP (op, 0))); | |
18196 | } | |
18197 | ||
70301b45 PB |
18198 | /* We can't rely on the caller doing the proper promotion when |
18199 | using APCS or ATPCS. */ | |
18200 | ||
18201 | static bool | |
586de218 | 18202 | arm_promote_prototypes (const_tree t ATTRIBUTE_UNUSED) |
70301b45 | 18203 | { |
b6685939 | 18204 | return !TARGET_AAPCS_BASED; |
70301b45 PB |
18205 | } |
18206 | ||
6b045785 PB |
18207 | |
18208 | /* AAPCS based ABIs use short enums by default. */ | |
18209 | ||
18210 | static bool | |
18211 | arm_default_short_enums (void) | |
18212 | { | |
077fc835 | 18213 | return TARGET_AAPCS_BASED && arm_abi != ARM_ABI_AAPCS_LINUX; |
6b045785 | 18214 | } |
13c1cd82 PB |
18215 | |
18216 | ||
18217 | /* AAPCS requires that anonymous bitfields affect structure alignment. */ | |
18218 | ||
18219 | static bool | |
18220 | arm_align_anon_bitfield (void) | |
18221 | { | |
18222 | return TARGET_AAPCS_BASED; | |
18223 | } | |
4185ae53 PB |
18224 | |
18225 | ||
18226 | /* The generic C++ ABI says 64-bit (long long). The EABI says 32-bit. */ | |
18227 | ||
18228 | static tree | |
18229 | arm_cxx_guard_type (void) | |
18230 | { | |
18231 | return TARGET_AAPCS_BASED ? integer_type_node : long_long_integer_type_node; | |
18232 | } | |
18233 | ||
c956e102 MS |
18234 | /* Return non-zero if the consumer (a multiply-accumulate instruction) |
18235 | has an accumulator dependency on the result of the producer (a | |
18236 | multiplication instruction) and no other dependency on that result. */ | |
18237 | int | |
18238 | arm_mac_accumulator_is_mul_result (rtx producer, rtx consumer) | |
18239 | { | |
18240 | rtx mul = PATTERN (producer); | |
18241 | rtx mac = PATTERN (consumer); | |
18242 | rtx mul_result; | |
18243 | rtx mac_op0, mac_op1, mac_acc; | |
18244 | ||
18245 | if (GET_CODE (mul) == COND_EXEC) | |
18246 | mul = COND_EXEC_CODE (mul); | |
18247 | if (GET_CODE (mac) == COND_EXEC) | |
18248 | mac = COND_EXEC_CODE (mac); | |
18249 | ||
18250 | /* Check that mul is of the form (set (...) (mult ...)) | |
18251 | and mla is of the form (set (...) (plus (mult ...) (...))). */ | |
18252 | if ((GET_CODE (mul) != SET || GET_CODE (XEXP (mul, 1)) != MULT) | |
18253 | || (GET_CODE (mac) != SET || GET_CODE (XEXP (mac, 1)) != PLUS | |
18254 | || GET_CODE (XEXP (XEXP (mac, 1), 0)) != MULT)) | |
18255 | return 0; | |
18256 | ||
18257 | mul_result = XEXP (mul, 0); | |
18258 | mac_op0 = XEXP (XEXP (XEXP (mac, 1), 0), 0); | |
18259 | mac_op1 = XEXP (XEXP (XEXP (mac, 1), 0), 1); | |
18260 | mac_acc = XEXP (XEXP (mac, 1), 1); | |
18261 | ||
18262 | return (reg_overlap_mentioned_p (mul_result, mac_acc) | |
18263 | && !reg_overlap_mentioned_p (mul_result, mac_op0) | |
18264 | && !reg_overlap_mentioned_p (mul_result, mac_op1)); | |
18265 | } | |
18266 | ||
4185ae53 | 18267 | |
0fa2e4df | 18268 | /* The EABI says test the least significant bit of a guard variable. */ |
4185ae53 PB |
18269 | |
18270 | static bool | |
18271 | arm_cxx_guard_mask_bit (void) | |
18272 | { | |
18273 | return TARGET_AAPCS_BASED; | |
18274 | } | |
46e995e0 PB |
18275 | |
18276 | ||
18277 | /* The EABI specifies that all array cookies are 8 bytes long. */ | |
18278 | ||
18279 | static tree | |
18280 | arm_get_cookie_size (tree type) | |
18281 | { | |
18282 | tree size; | |
18283 | ||
18284 | if (!TARGET_AAPCS_BASED) | |
18285 | return default_cxx_get_cookie_size (type); | |
18286 | ||
7d60be94 | 18287 | size = build_int_cst (sizetype, 8); |
46e995e0 PB |
18288 | return size; |
18289 | } | |
18290 | ||
18291 | ||
18292 | /* The EABI says that array cookies should also contain the element size. */ | |
18293 | ||
18294 | static bool | |
18295 | arm_cookie_has_size (void) | |
18296 | { | |
18297 | return TARGET_AAPCS_BASED; | |
18298 | } | |
44d10c10 PB |
18299 | |
18300 | ||
18301 | /* The EABI says constructors and destructors should return a pointer to | |
18302 | the object constructed/destroyed. */ | |
18303 | ||
18304 | static bool | |
18305 | arm_cxx_cdtor_returns_this (void) | |
18306 | { | |
18307 | return TARGET_AAPCS_BASED; | |
18308 | } | |
c9ca9b88 | 18309 | |
505970fc MM |
18310 | /* The EABI says that an inline function may never be the key |
18311 | method. */ | |
18312 | ||
18313 | static bool | |
18314 | arm_cxx_key_method_may_be_inline (void) | |
18315 | { | |
18316 | return !TARGET_AAPCS_BASED; | |
18317 | } | |
18318 | ||
1e731102 MM |
18319 | static void |
18320 | arm_cxx_determine_class_data_visibility (tree decl) | |
18321 | { | |
711b2998 JB |
18322 | if (!TARGET_AAPCS_BASED |
18323 | || !TARGET_DLLIMPORT_DECL_ATTRIBUTES) | |
1e731102 | 18324 | return; |
505970fc | 18325 | |
1e731102 MM |
18326 | /* In general, \S 3.2.5.5 of the ARM EABI requires that class data |
18327 | is exported. However, on systems without dynamic vague linkage, | |
18328 | \S 3.2.5.6 says that COMDAT class data has hidden linkage. */ | |
18329 | if (!TARGET_ARM_DYNAMIC_VAGUE_LINKAGE_P && DECL_COMDAT (decl)) | |
18330 | DECL_VISIBILITY (decl) = VISIBILITY_HIDDEN; | |
18331 | else | |
18332 | DECL_VISIBILITY (decl) = VISIBILITY_DEFAULT; | |
18333 | DECL_VISIBILITY_SPECIFIED (decl) = 1; | |
18334 | } | |
e0b92319 | 18335 | |
505970fc | 18336 | static bool |
1e731102 | 18337 | arm_cxx_class_data_always_comdat (void) |
505970fc | 18338 | { |
1e731102 MM |
18339 | /* \S 3.2.5.4 of the ARM C++ ABI says that class data only have |
18340 | vague linkage if the class has no key function. */ | |
18341 | return !TARGET_AAPCS_BASED; | |
505970fc | 18342 | } |
c9ca9b88 | 18343 | |
9f62c3e3 PB |
18344 | |
18345 | /* The EABI says __aeabi_atexit should be used to register static | |
18346 | destructors. */ | |
18347 | ||
18348 | static bool | |
18349 | arm_cxx_use_aeabi_atexit (void) | |
18350 | { | |
18351 | return TARGET_AAPCS_BASED; | |
18352 | } | |
18353 | ||
18354 | ||
c9ca9b88 PB |
18355 | void |
18356 | arm_set_return_address (rtx source, rtx scratch) | |
18357 | { | |
18358 | arm_stack_offsets *offsets; | |
18359 | HOST_WIDE_INT delta; | |
18360 | rtx addr; | |
18361 | unsigned long saved_regs; | |
18362 | ||
954954d1 PB |
18363 | offsets = arm_get_frame_offsets (); |
18364 | saved_regs = offsets->saved_regs_mask; | |
c9ca9b88 PB |
18365 | |
18366 | if ((saved_regs & (1 << LR_REGNUM)) == 0) | |
18367 | emit_move_insn (gen_rtx_REG (Pmode, LR_REGNUM), source); | |
18368 | else | |
18369 | { | |
18370 | if (frame_pointer_needed) | |
18371 | addr = plus_constant(hard_frame_pointer_rtx, -4); | |
18372 | else | |
18373 | { | |
18374 | /* LR will be the first saved register. */ | |
c9ca9b88 PB |
18375 | delta = offsets->outgoing_args - (offsets->frame + 4); |
18376 | ||
f676971a | 18377 | |
c9ca9b88 PB |
18378 | if (delta >= 4096) |
18379 | { | |
18380 | emit_insn (gen_addsi3 (scratch, stack_pointer_rtx, | |
18381 | GEN_INT (delta & ~4095))); | |
18382 | addr = scratch; | |
18383 | delta &= 4095; | |
18384 | } | |
18385 | else | |
18386 | addr = stack_pointer_rtx; | |
18387 | ||
18388 | addr = plus_constant (addr, delta); | |
18389 | } | |
31fa16b6 | 18390 | emit_move_insn (gen_frame_mem (Pmode, addr), source); |
c9ca9b88 PB |
18391 | } |
18392 | } | |
18393 | ||
18394 | ||
18395 | void | |
18396 | thumb_set_return_address (rtx source, rtx scratch) | |
18397 | { | |
18398 | arm_stack_offsets *offsets; | |
c9ca9b88 | 18399 | HOST_WIDE_INT delta; |
5b3e6663 | 18400 | HOST_WIDE_INT limit; |
c9ca9b88 PB |
18401 | int reg; |
18402 | rtx addr; | |
57934c39 | 18403 | unsigned long mask; |
c9ca9b88 | 18404 | |
c41c1387 | 18405 | emit_use (source); |
c9ca9b88 | 18406 | |
954954d1 PB |
18407 | offsets = arm_get_frame_offsets (); |
18408 | mask = offsets->saved_regs_mask; | |
57934c39 | 18409 | if (mask & (1 << LR_REGNUM)) |
c9ca9b88 | 18410 | { |
5b3e6663 | 18411 | limit = 1024; |
c9ca9b88 PB |
18412 | /* Find the saved regs. */ |
18413 | if (frame_pointer_needed) | |
18414 | { | |
18415 | delta = offsets->soft_frame - offsets->saved_args; | |
18416 | reg = THUMB_HARD_FRAME_POINTER_REGNUM; | |
5b3e6663 PB |
18417 | if (TARGET_THUMB1) |
18418 | limit = 128; | |
c9ca9b88 PB |
18419 | } |
18420 | else | |
18421 | { | |
18422 | delta = offsets->outgoing_args - offsets->saved_args; | |
18423 | reg = SP_REGNUM; | |
18424 | } | |
18425 | /* Allow for the stack frame. */ | |
5b3e6663 | 18426 | if (TARGET_THUMB1 && TARGET_BACKTRACE) |
c9ca9b88 PB |
18427 | delta -= 16; |
18428 | /* The link register is always the first saved register. */ | |
18429 | delta -= 4; | |
f676971a | 18430 | |
c9ca9b88 PB |
18431 | /* Construct the address. */ |
18432 | addr = gen_rtx_REG (SImode, reg); | |
5b3e6663 | 18433 | if (delta > limit) |
c9ca9b88 PB |
18434 | { |
18435 | emit_insn (gen_movsi (scratch, GEN_INT (delta))); | |
18436 | emit_insn (gen_addsi3 (scratch, scratch, stack_pointer_rtx)); | |
18437 | addr = scratch; | |
18438 | } | |
18439 | else | |
18440 | addr = plus_constant (addr, delta); | |
18441 | ||
31fa16b6 | 18442 | emit_move_insn (gen_frame_mem (Pmode, addr), source); |
c9ca9b88 PB |
18443 | } |
18444 | else | |
18445 | emit_move_insn (gen_rtx_REG (Pmode, LR_REGNUM), source); | |
18446 | } | |
18447 | ||
f676971a EC |
18448 | /* Implements target hook vector_mode_supported_p. */ |
18449 | bool | |
18450 | arm_vector_mode_supported_p (enum machine_mode mode) | |
18451 | { | |
88f77cba JB |
18452 | /* Neon also supports V2SImode, etc. listed in the clause below. */ |
18453 | if (TARGET_NEON && (mode == V2SFmode || mode == V4SImode || mode == V8HImode | |
18454 | || mode == V16QImode || mode == V4SFmode || mode == V2DImode)) | |
18455 | return true; | |
18456 | ||
f676971a EC |
18457 | if ((mode == V2SImode) |
18458 | || (mode == V4HImode) | |
18459 | || (mode == V8QImode)) | |
18460 | return true; | |
18461 | ||
18462 | return false; | |
18463 | } | |
273a2526 RS |
18464 | |
18465 | /* Implement TARGET_SHIFT_TRUNCATION_MASK. SImode shifts use normal | |
18466 | ARM insns and therefore guarantee that the shift count is modulo 256. | |
18467 | DImode shifts (those implemented by lib1funcs.asm or by optabs.c) | |
18468 | guarantee no particular behavior for out-of-range counts. */ | |
18469 | ||
18470 | static unsigned HOST_WIDE_INT | |
18471 | arm_shift_truncation_mask (enum machine_mode mode) | |
18472 | { | |
18473 | return mode == SImode ? 255 : 0; | |
18474 | } | |
2fa330b2 PB |
18475 | |
18476 | ||
18477 | /* Map internal gcc register numbers to DWARF2 register numbers. */ | |
18478 | ||
18479 | unsigned int | |
18480 | arm_dbx_register_number (unsigned int regno) | |
18481 | { | |
18482 | if (regno < 16) | |
18483 | return regno; | |
18484 | ||
18485 | /* TODO: Legacy targets output FPA regs as registers 16-23 for backwards | |
18486 | compatibility. The EABI defines them as registers 96-103. */ | |
18487 | if (IS_FPA_REGNUM (regno)) | |
18488 | return (TARGET_AAPCS_BASED ? 96 : 16) + regno - FIRST_FPA_REGNUM; | |
18489 | ||
f1adb0a9 | 18490 | /* FIXME: VFPv3 register numbering. */ |
2fa330b2 PB |
18491 | if (IS_VFP_REGNUM (regno)) |
18492 | return 64 + regno - FIRST_VFP_REGNUM; | |
18493 | ||
18494 | if (IS_IWMMXT_GR_REGNUM (regno)) | |
18495 | return 104 + regno - FIRST_IWMMXT_GR_REGNUM; | |
18496 | ||
18497 | if (IS_IWMMXT_REGNUM (regno)) | |
18498 | return 112 + regno - FIRST_IWMMXT_REGNUM; | |
18499 | ||
e6d29d15 | 18500 | gcc_unreachable (); |
2fa330b2 PB |
18501 | } |
18502 | ||
617a1b71 PB |
18503 | |
18504 | #ifdef TARGET_UNWIND_INFO | |
5b3e6663 PB |
18505 | /* Emit unwind directives for a store-multiple instruction or stack pointer |
18506 | push during alignment. | |
18507 | These should only ever be generated by the function prologue code, so | |
18508 | expect them to have a particular form. */ | |
617a1b71 PB |
18509 | |
18510 | static void | |
5b3e6663 | 18511 | arm_unwind_emit_sequence (FILE * asm_out_file, rtx p) |
617a1b71 PB |
18512 | { |
18513 | int i; | |
18514 | HOST_WIDE_INT offset; | |
18515 | HOST_WIDE_INT nregs; | |
18516 | int reg_size; | |
18517 | unsigned reg; | |
18518 | unsigned lastreg; | |
18519 | rtx e; | |
18520 | ||
617a1b71 | 18521 | e = XVECEXP (p, 0, 0); |
5b3e6663 PB |
18522 | if (GET_CODE (e) != SET) |
18523 | abort (); | |
18524 | ||
18525 | /* First insn will adjust the stack pointer. */ | |
617a1b71 PB |
18526 | if (GET_CODE (e) != SET |
18527 | || GET_CODE (XEXP (e, 0)) != REG | |
18528 | || REGNO (XEXP (e, 0)) != SP_REGNUM | |
18529 | || GET_CODE (XEXP (e, 1)) != PLUS) | |
18530 | abort (); | |
18531 | ||
18532 | offset = -INTVAL (XEXP (XEXP (e, 1), 1)); | |
18533 | nregs = XVECLEN (p, 0) - 1; | |
18534 | ||
18535 | reg = REGNO (XEXP (XVECEXP (p, 0, 1), 1)); | |
18536 | if (reg < 16) | |
18537 | { | |
18538 | /* The function prologue may also push pc, but not annotate it as it is | |
569b7f6a | 18539 | never restored. We turn this into a stack pointer adjustment. */ |
617a1b71 PB |
18540 | if (nregs * 4 == offset - 4) |
18541 | { | |
18542 | fprintf (asm_out_file, "\t.pad #4\n"); | |
18543 | offset -= 4; | |
18544 | } | |
18545 | reg_size = 4; | |
8edfc4cc | 18546 | fprintf (asm_out_file, "\t.save {"); |
617a1b71 PB |
18547 | } |
18548 | else if (IS_VFP_REGNUM (reg)) | |
18549 | { | |
617a1b71 | 18550 | reg_size = 8; |
8edfc4cc | 18551 | fprintf (asm_out_file, "\t.vsave {"); |
617a1b71 PB |
18552 | } |
18553 | else if (reg >= FIRST_FPA_REGNUM && reg <= LAST_FPA_REGNUM) | |
18554 | { | |
18555 | /* FPA registers are done differently. */ | |
ea40ba9c | 18556 | asm_fprintf (asm_out_file, "\t.save %r, %wd\n", reg, nregs); |
617a1b71 PB |
18557 | return; |
18558 | } | |
18559 | else | |
18560 | /* Unknown register type. */ | |
18561 | abort (); | |
18562 | ||
18563 | /* If the stack increment doesn't match the size of the saved registers, | |
18564 | something has gone horribly wrong. */ | |
18565 | if (offset != nregs * reg_size) | |
18566 | abort (); | |
18567 | ||
617a1b71 PB |
18568 | offset = 0; |
18569 | lastreg = 0; | |
18570 | /* The remaining insns will describe the stores. */ | |
18571 | for (i = 1; i <= nregs; i++) | |
18572 | { | |
18573 | /* Expect (set (mem <addr>) (reg)). | |
18574 | Where <addr> is (reg:SP) or (plus (reg:SP) (const_int)). */ | |
18575 | e = XVECEXP (p, 0, i); | |
18576 | if (GET_CODE (e) != SET | |
18577 | || GET_CODE (XEXP (e, 0)) != MEM | |
18578 | || GET_CODE (XEXP (e, 1)) != REG) | |
18579 | abort (); | |
e0b92319 | 18580 | |
617a1b71 PB |
18581 | reg = REGNO (XEXP (e, 1)); |
18582 | if (reg < lastreg) | |
18583 | abort (); | |
e0b92319 | 18584 | |
617a1b71 PB |
18585 | if (i != 1) |
18586 | fprintf (asm_out_file, ", "); | |
18587 | /* We can't use %r for vfp because we need to use the | |
18588 | double precision register names. */ | |
18589 | if (IS_VFP_REGNUM (reg)) | |
18590 | asm_fprintf (asm_out_file, "d%d", (reg - FIRST_VFP_REGNUM) / 2); | |
18591 | else | |
18592 | asm_fprintf (asm_out_file, "%r", reg); | |
18593 | ||
18594 | #ifdef ENABLE_CHECKING | |
18595 | /* Check that the addresses are consecutive. */ | |
18596 | e = XEXP (XEXP (e, 0), 0); | |
18597 | if (GET_CODE (e) == PLUS) | |
18598 | { | |
18599 | offset += reg_size; | |
18600 | if (GET_CODE (XEXP (e, 0)) != REG | |
18601 | || REGNO (XEXP (e, 0)) != SP_REGNUM | |
18602 | || GET_CODE (XEXP (e, 1)) != CONST_INT | |
18603 | || offset != INTVAL (XEXP (e, 1))) | |
18604 | abort (); | |
18605 | } | |
18606 | else if (i != 1 | |
18607 | || GET_CODE (e) != REG | |
18608 | || REGNO (e) != SP_REGNUM) | |
18609 | abort (); | |
18610 | #endif | |
18611 | } | |
18612 | fprintf (asm_out_file, "}\n"); | |
18613 | } | |
18614 | ||
18615 | /* Emit unwind directives for a SET. */ | |
18616 | ||
18617 | static void | |
18618 | arm_unwind_emit_set (FILE * asm_out_file, rtx p) | |
18619 | { | |
18620 | rtx e0; | |
18621 | rtx e1; | |
5b3e6663 | 18622 | unsigned reg; |
617a1b71 PB |
18623 | |
18624 | e0 = XEXP (p, 0); | |
18625 | e1 = XEXP (p, 1); | |
18626 | switch (GET_CODE (e0)) | |
18627 | { | |
18628 | case MEM: | |
18629 | /* Pushing a single register. */ | |
18630 | if (GET_CODE (XEXP (e0, 0)) != PRE_DEC | |
18631 | || GET_CODE (XEXP (XEXP (e0, 0), 0)) != REG | |
18632 | || REGNO (XEXP (XEXP (e0, 0), 0)) != SP_REGNUM) | |
18633 | abort (); | |
18634 | ||
18635 | asm_fprintf (asm_out_file, "\t.save "); | |
18636 | if (IS_VFP_REGNUM (REGNO (e1))) | |
18637 | asm_fprintf(asm_out_file, "{d%d}\n", | |
18638 | (REGNO (e1) - FIRST_VFP_REGNUM) / 2); | |
18639 | else | |
18640 | asm_fprintf(asm_out_file, "{%r}\n", REGNO (e1)); | |
18641 | break; | |
18642 | ||
18643 | case REG: | |
18644 | if (REGNO (e0) == SP_REGNUM) | |
18645 | { | |
18646 | /* A stack increment. */ | |
18647 | if (GET_CODE (e1) != PLUS | |
18648 | || GET_CODE (XEXP (e1, 0)) != REG | |
18649 | || REGNO (XEXP (e1, 0)) != SP_REGNUM | |
18650 | || GET_CODE (XEXP (e1, 1)) != CONST_INT) | |
18651 | abort (); | |
18652 | ||
ea40ba9c | 18653 | asm_fprintf (asm_out_file, "\t.pad #%wd\n", |
617a1b71 PB |
18654 | -INTVAL (XEXP (e1, 1))); |
18655 | } | |
18656 | else if (REGNO (e0) == HARD_FRAME_POINTER_REGNUM) | |
18657 | { | |
18658 | HOST_WIDE_INT offset; | |
e0b92319 | 18659 | |
617a1b71 PB |
18660 | if (GET_CODE (e1) == PLUS) |
18661 | { | |
18662 | if (GET_CODE (XEXP (e1, 0)) != REG | |
18663 | || GET_CODE (XEXP (e1, 1)) != CONST_INT) | |
18664 | abort (); | |
18665 | reg = REGNO (XEXP (e1, 0)); | |
18666 | offset = INTVAL (XEXP (e1, 1)); | |
ea40ba9c | 18667 | asm_fprintf (asm_out_file, "\t.setfp %r, %r, #%wd\n", |
617a1b71 PB |
18668 | HARD_FRAME_POINTER_REGNUM, reg, |
18669 | INTVAL (XEXP (e1, 1))); | |
18670 | } | |
18671 | else if (GET_CODE (e1) == REG) | |
18672 | { | |
18673 | reg = REGNO (e1); | |
18674 | asm_fprintf (asm_out_file, "\t.setfp %r, %r\n", | |
18675 | HARD_FRAME_POINTER_REGNUM, reg); | |
18676 | } | |
18677 | else | |
18678 | abort (); | |
18679 | } | |
18680 | else if (GET_CODE (e1) == REG && REGNO (e1) == SP_REGNUM) | |
18681 | { | |
18682 | /* Move from sp to reg. */ | |
18683 | asm_fprintf (asm_out_file, "\t.movsp %r\n", REGNO (e0)); | |
18684 | } | |
758ed9b2 PB |
18685 | else if (GET_CODE (e1) == PLUS |
18686 | && GET_CODE (XEXP (e1, 0)) == REG | |
18687 | && REGNO (XEXP (e1, 0)) == SP_REGNUM | |
18688 | && GET_CODE (XEXP (e1, 1)) == CONST_INT) | |
18689 | { | |
18690 | /* Set reg to offset from sp. */ | |
18691 | asm_fprintf (asm_out_file, "\t.movsp %r, #%d\n", | |
18692 | REGNO (e0), (int)INTVAL(XEXP (e1, 1))); | |
18693 | } | |
5b3e6663 PB |
18694 | else if (GET_CODE (e1) == UNSPEC && XINT (e1, 1) == UNSPEC_STACK_ALIGN) |
18695 | { | |
18696 | /* Stack pointer save before alignment. */ | |
18697 | reg = REGNO (e0); | |
18698 | asm_fprintf (asm_out_file, "\t.unwind_raw 0, 0x%x @ vsp = r%d\n", | |
18699 | reg + 0x90, reg); | |
18700 | } | |
617a1b71 PB |
18701 | else |
18702 | abort (); | |
18703 | break; | |
18704 | ||
18705 | default: | |
18706 | abort (); | |
18707 | } | |
18708 | } | |
18709 | ||
18710 | ||
18711 | /* Emit unwind directives for the given insn. */ | |
18712 | ||
18713 | static void | |
18714 | arm_unwind_emit (FILE * asm_out_file, rtx insn) | |
18715 | { | |
18716 | rtx pat; | |
18717 | ||
18718 | if (!ARM_EABI_UNWIND_TABLES) | |
18719 | return; | |
18720 | ||
e3b5732b | 18721 | if (!(flag_unwind_tables || crtl->uses_eh_lsda) |
80efdb6a | 18722 | && (TREE_NOTHROW (current_function_decl) |
ad516a74 | 18723 | || crtl->all_throwers_are_sibcalls)) |
80efdb6a PB |
18724 | return; |
18725 | ||
617a1b71 PB |
18726 | if (GET_CODE (insn) == NOTE || !RTX_FRAME_RELATED_P (insn)) |
18727 | return; | |
18728 | ||
18729 | pat = find_reg_note (insn, REG_FRAME_RELATED_EXPR, NULL_RTX); | |
18730 | if (pat) | |
18731 | pat = XEXP (pat, 0); | |
18732 | else | |
18733 | pat = PATTERN (insn); | |
18734 | ||
18735 | switch (GET_CODE (pat)) | |
18736 | { | |
18737 | case SET: | |
18738 | arm_unwind_emit_set (asm_out_file, pat); | |
18739 | break; | |
18740 | ||
18741 | case SEQUENCE: | |
18742 | /* Store multiple. */ | |
5b3e6663 | 18743 | arm_unwind_emit_sequence (asm_out_file, pat); |
617a1b71 PB |
18744 | break; |
18745 | ||
18746 | default: | |
18747 | abort(); | |
18748 | } | |
18749 | } | |
18750 | ||
18751 | ||
18752 | /* Output a reference from a function exception table to the type_info | |
18753 | object X. The EABI specifies that the symbol should be relocated by | |
18754 | an R_ARM_TARGET2 relocation. */ | |
18755 | ||
18756 | static bool | |
18757 | arm_output_ttype (rtx x) | |
18758 | { | |
18759 | fputs ("\t.word\t", asm_out_file); | |
18760 | output_addr_const (asm_out_file, x); | |
18761 | /* Use special relocations for symbol references. */ | |
18762 | if (GET_CODE (x) != CONST_INT) | |
18763 | fputs ("(TARGET2)", asm_out_file); | |
18764 | fputc ('\n', asm_out_file); | |
18765 | ||
18766 | return TRUE; | |
18767 | } | |
18768 | #endif /* TARGET_UNWIND_INFO */ | |
18769 | ||
18770 | ||
5b3e6663 PB |
18771 | /* Handle UNSPEC DWARF call frame instructions. These are needed for dynamic |
18772 | stack alignment. */ | |
18773 | ||
18774 | static void | |
18775 | arm_dwarf_handle_frame_unspec (const char *label, rtx pattern, int index) | |
18776 | { | |
18777 | rtx unspec = SET_SRC (pattern); | |
18778 | gcc_assert (GET_CODE (unspec) == UNSPEC); | |
18779 | ||
18780 | switch (index) | |
18781 | { | |
18782 | case UNSPEC_STACK_ALIGN: | |
18783 | /* ??? We should set the CFA = (SP & ~7). At this point we haven't | |
18784 | put anything on the stack, so hopefully it won't matter. | |
18785 | CFA = SP will be correct after alignment. */ | |
18786 | dwarf2out_reg_save_reg (label, stack_pointer_rtx, | |
18787 | SET_DEST (pattern)); | |
18788 | break; | |
18789 | default: | |
18790 | gcc_unreachable (); | |
18791 | } | |
18792 | } | |
18793 | ||
18794 | ||
617a1b71 PB |
18795 | /* Output unwind directives for the start/end of a function. */ |
18796 | ||
18797 | void | |
18798 | arm_output_fn_unwind (FILE * f, bool prologue) | |
18799 | { | |
18800 | if (!ARM_EABI_UNWIND_TABLES) | |
18801 | return; | |
18802 | ||
18803 | if (prologue) | |
18804 | fputs ("\t.fnstart\n", f); | |
18805 | else | |
80efdb6a PB |
18806 | { |
18807 | /* If this function will never be unwound, then mark it as such. | |
18808 | The came condition is used in arm_unwind_emit to suppress | |
18809 | the frame annotations. */ | |
e3b5732b | 18810 | if (!(flag_unwind_tables || crtl->uses_eh_lsda) |
80efdb6a | 18811 | && (TREE_NOTHROW (current_function_decl) |
ad516a74 | 18812 | || crtl->all_throwers_are_sibcalls)) |
80efdb6a PB |
18813 | fputs("\t.cantunwind\n", f); |
18814 | ||
18815 | fputs ("\t.fnend\n", f); | |
18816 | } | |
617a1b71 | 18817 | } |
d3585b76 DJ |
18818 | |
18819 | static bool | |
18820 | arm_emit_tls_decoration (FILE *fp, rtx x) | |
18821 | { | |
18822 | enum tls_reloc reloc; | |
18823 | rtx val; | |
18824 | ||
18825 | val = XVECEXP (x, 0, 0); | |
18826 | reloc = INTVAL (XVECEXP (x, 0, 1)); | |
18827 | ||
18828 | output_addr_const (fp, val); | |
18829 | ||
18830 | switch (reloc) | |
18831 | { | |
18832 | case TLS_GD32: | |
18833 | fputs ("(tlsgd)", fp); | |
18834 | break; | |
18835 | case TLS_LDM32: | |
18836 | fputs ("(tlsldm)", fp); | |
18837 | break; | |
18838 | case TLS_LDO32: | |
18839 | fputs ("(tlsldo)", fp); | |
18840 | break; | |
18841 | case TLS_IE32: | |
18842 | fputs ("(gottpoff)", fp); | |
18843 | break; | |
18844 | case TLS_LE32: | |
18845 | fputs ("(tpoff)", fp); | |
18846 | break; | |
18847 | default: | |
18848 | gcc_unreachable (); | |
18849 | } | |
18850 | ||
18851 | switch (reloc) | |
18852 | { | |
18853 | case TLS_GD32: | |
18854 | case TLS_LDM32: | |
18855 | case TLS_IE32: | |
18856 | fputs (" + (. - ", fp); | |
18857 | output_addr_const (fp, XVECEXP (x, 0, 2)); | |
18858 | fputs (" - ", fp); | |
18859 | output_addr_const (fp, XVECEXP (x, 0, 3)); | |
18860 | fputc (')', fp); | |
18861 | break; | |
18862 | default: | |
18863 | break; | |
18864 | } | |
18865 | ||
18866 | return TRUE; | |
18867 | } | |
18868 | ||
afcc986d JM |
18869 | /* ARM implementation of TARGET_ASM_OUTPUT_DWARF_DTPREL. */ |
18870 | ||
18871 | static void | |
18872 | arm_output_dwarf_dtprel (FILE *file, int size, rtx x) | |
18873 | { | |
18874 | gcc_assert (size == 4); | |
18875 | fputs ("\t.word\t", file); | |
18876 | output_addr_const (file, x); | |
18877 | fputs ("(tlsldo)", file); | |
18878 | } | |
18879 | ||
d3585b76 DJ |
18880 | bool |
18881 | arm_output_addr_const_extra (FILE *fp, rtx x) | |
18882 | { | |
18883 | if (GET_CODE (x) == UNSPEC && XINT (x, 1) == UNSPEC_TLS) | |
18884 | return arm_emit_tls_decoration (fp, x); | |
f16fe45f DJ |
18885 | else if (GET_CODE (x) == UNSPEC && XINT (x, 1) == UNSPEC_PIC_LABEL) |
18886 | { | |
18887 | char label[256]; | |
18888 | int labelno = INTVAL (XVECEXP (x, 0, 0)); | |
18889 | ||
18890 | ASM_GENERATE_INTERNAL_LABEL (label, "LPIC", labelno); | |
18891 | assemble_name_raw (fp, label); | |
18892 | ||
18893 | return TRUE; | |
18894 | } | |
d3585b76 DJ |
18895 | else if (GET_CODE (x) == CONST_VECTOR) |
18896 | return arm_emit_vector_const (fp, x); | |
18897 | ||
18898 | return FALSE; | |
18899 | } | |
18900 | ||
5b3e6663 PB |
18901 | /* Output assembly for a shift instruction. |
18902 | SET_FLAGS determines how the instruction modifies the condition codes. | |
7a085dce | 18903 | 0 - Do not set condition codes. |
5b3e6663 PB |
18904 | 1 - Set condition codes. |
18905 | 2 - Use smallest instruction. */ | |
18906 | const char * | |
18907 | arm_output_shift(rtx * operands, int set_flags) | |
18908 | { | |
18909 | char pattern[100]; | |
18910 | static const char flag_chars[3] = {'?', '.', '!'}; | |
18911 | const char *shift; | |
18912 | HOST_WIDE_INT val; | |
18913 | char c; | |
18914 | ||
18915 | c = flag_chars[set_flags]; | |
18916 | if (TARGET_UNIFIED_ASM) | |
18917 | { | |
18918 | shift = shift_op(operands[3], &val); | |
18919 | if (shift) | |
18920 | { | |
18921 | if (val != -1) | |
18922 | operands[2] = GEN_INT(val); | |
18923 | sprintf (pattern, "%s%%%c\t%%0, %%1, %%2", shift, c); | |
18924 | } | |
18925 | else | |
18926 | sprintf (pattern, "mov%%%c\t%%0, %%1", c); | |
18927 | } | |
18928 | else | |
18929 | sprintf (pattern, "mov%%%c\t%%0, %%1%%S3", c); | |
18930 | output_asm_insn (pattern, operands); | |
18931 | return ""; | |
18932 | } | |
18933 | ||
18934 | /* Output a Thumb-2 casesi instruction. */ | |
18935 | const char * | |
18936 | thumb2_output_casesi (rtx *operands) | |
18937 | { | |
18938 | rtx diff_vec = PATTERN (next_real_insn (operands[2])); | |
18939 | ||
18940 | gcc_assert (GET_CODE (diff_vec) == ADDR_DIFF_VEC); | |
18941 | ||
18942 | output_asm_insn ("cmp\t%0, %1", operands); | |
18943 | output_asm_insn ("bhi\t%l3", operands); | |
18944 | switch (GET_MODE(diff_vec)) | |
18945 | { | |
18946 | case QImode: | |
18947 | return "tbb\t[%|pc, %0]"; | |
18948 | case HImode: | |
18949 | return "tbh\t[%|pc, %0, lsl #1]"; | |
18950 | case SImode: | |
18951 | if (flag_pic) | |
18952 | { | |
18953 | output_asm_insn ("adr\t%4, %l2", operands); | |
18954 | output_asm_insn ("ldr\t%5, [%4, %0, lsl #2]", operands); | |
18955 | output_asm_insn ("add\t%4, %4, %5", operands); | |
18956 | return "bx\t%4"; | |
18957 | } | |
18958 | else | |
18959 | { | |
18960 | output_asm_insn ("adr\t%4, %l2", operands); | |
18961 | return "ldr\t%|pc, [%4, %0, lsl #2]"; | |
18962 | } | |
18963 | default: | |
18964 | gcc_unreachable (); | |
18965 | } | |
18966 | } | |
18967 | ||
bd4dc3cd PB |
18968 | /* Most ARM cores are single issue, but some newer ones can dual issue. |
18969 | The scheduler descriptions rely on this being correct. */ | |
18970 | static int | |
18971 | arm_issue_rate (void) | |
18972 | { | |
18973 | switch (arm_tune) | |
18974 | { | |
18975 | case cortexr4: | |
18976 | case cortexa8: | |
18977 | return 2; | |
18978 | ||
18979 | default: | |
18980 | return 1; | |
18981 | } | |
18982 | } | |
18983 | ||
608063c3 JB |
18984 | /* A table and a function to perform ARM-specific name mangling for |
18985 | NEON vector types in order to conform to the AAPCS (see "Procedure | |
18986 | Call Standard for the ARM Architecture", Appendix A). To qualify | |
18987 | for emission with the mangled names defined in that document, a | |
18988 | vector type must not only be of the correct mode but also be | |
18989 | composed of NEON vector element types (e.g. __builtin_neon_qi). */ | |
18990 | typedef struct | |
18991 | { | |
18992 | enum machine_mode mode; | |
18993 | const char *element_type_name; | |
18994 | const char *aapcs_name; | |
18995 | } arm_mangle_map_entry; | |
18996 | ||
18997 | static arm_mangle_map_entry arm_mangle_map[] = { | |
18998 | /* 64-bit containerized types. */ | |
18999 | { V8QImode, "__builtin_neon_qi", "15__simd64_int8_t" }, | |
19000 | { V8QImode, "__builtin_neon_uqi", "16__simd64_uint8_t" }, | |
19001 | { V4HImode, "__builtin_neon_hi", "16__simd64_int16_t" }, | |
19002 | { V4HImode, "__builtin_neon_uhi", "17__simd64_uint16_t" }, | |
19003 | { V2SImode, "__builtin_neon_si", "16__simd64_int32_t" }, | |
19004 | { V2SImode, "__builtin_neon_usi", "17__simd64_uint32_t" }, | |
19005 | { V2SFmode, "__builtin_neon_sf", "18__simd64_float32_t" }, | |
19006 | { V8QImode, "__builtin_neon_poly8", "16__simd64_poly8_t" }, | |
19007 | { V4HImode, "__builtin_neon_poly16", "17__simd64_poly16_t" }, | |
19008 | /* 128-bit containerized types. */ | |
19009 | { V16QImode, "__builtin_neon_qi", "16__simd128_int8_t" }, | |
19010 | { V16QImode, "__builtin_neon_uqi", "17__simd128_uint8_t" }, | |
19011 | { V8HImode, "__builtin_neon_hi", "17__simd128_int16_t" }, | |
19012 | { V8HImode, "__builtin_neon_uhi", "18__simd128_uint16_t" }, | |
19013 | { V4SImode, "__builtin_neon_si", "17__simd128_int32_t" }, | |
19014 | { V4SImode, "__builtin_neon_usi", "18__simd128_uint32_t" }, | |
19015 | { V4SFmode, "__builtin_neon_sf", "19__simd128_float32_t" }, | |
19016 | { V16QImode, "__builtin_neon_poly8", "17__simd128_poly8_t" }, | |
19017 | { V8HImode, "__builtin_neon_poly16", "18__simd128_poly16_t" }, | |
19018 | { VOIDmode, NULL, NULL } | |
19019 | }; | |
19020 | ||
19021 | const char * | |
3101faab | 19022 | arm_mangle_type (const_tree type) |
608063c3 JB |
19023 | { |
19024 | arm_mangle_map_entry *pos = arm_mangle_map; | |
19025 | ||
19026 | if (TREE_CODE (type) != VECTOR_TYPE) | |
19027 | return NULL; | |
19028 | ||
19029 | /* Check the mode of the vector type, and the name of the vector | |
19030 | element type, against the table. */ | |
19031 | while (pos->mode != VOIDmode) | |
19032 | { | |
19033 | tree elt_type = TREE_TYPE (type); | |
19034 | ||
19035 | if (pos->mode == TYPE_MODE (type) | |
19036 | && TREE_CODE (TYPE_NAME (elt_type)) == TYPE_DECL | |
19037 | && !strcmp (IDENTIFIER_POINTER (DECL_NAME (TYPE_NAME (elt_type))), | |
19038 | pos->element_type_name)) | |
19039 | return pos->aapcs_name; | |
19040 | ||
19041 | pos++; | |
19042 | } | |
19043 | ||
19044 | /* Use the default mangling for unrecognized (possibly user-defined) | |
19045 | vector types. */ | |
19046 | return NULL; | |
19047 | } | |
19048 | ||
795dc4fc PB |
19049 | /* Order of allocation of core registers for Thumb: this allocation is |
19050 | written over the corresponding initial entries of the array | |
19051 | initialized with REG_ALLOC_ORDER. We allocate all low registers | |
19052 | first. Saving and restoring a low register is usually cheaper than | |
19053 | using a call-clobbered high register. */ | |
19054 | ||
19055 | static const int thumb_core_reg_alloc_order[] = | |
19056 | { | |
19057 | 3, 2, 1, 0, 4, 5, 6, 7, | |
19058 | 14, 12, 8, 9, 10, 11, 13, 15 | |
19059 | }; | |
19060 | ||
19061 | /* Adjust register allocation order when compiling for Thumb. */ | |
19062 | ||
19063 | void | |
19064 | arm_order_regs_for_local_alloc (void) | |
19065 | { | |
19066 | const int arm_reg_alloc_order[] = REG_ALLOC_ORDER; | |
19067 | memcpy(reg_alloc_order, arm_reg_alloc_order, sizeof (reg_alloc_order)); | |
19068 | if (TARGET_THUMB) | |
19069 | memcpy (reg_alloc_order, thumb_core_reg_alloc_order, | |
19070 | sizeof (thumb_core_reg_alloc_order)); | |
19071 | } | |
19072 | ||
d3585b76 | 19073 | #include "gt-arm.h" |