]>
Commit | Line | Data |
---|---|---|
ab835497 | 1 | /* Subroutines for insn-output.c for Sun SPARC. |
b73bf65f | 2 | Copyright (C) 1987, 88, 89, 92-98, 1999 Free Software Foundation, Inc. |
ab835497 | 3 | Contributed by Michael Tiemann (tiemann@cygnus.com) |
61a55e8b DE |
4 | 64 bit SPARC V9 support by Michael Tiemann, Jim Wilson, and Doug Evans, |
5 | at Cygnus Support. | |
ab835497 RK |
6 | |
7 | This file is part of GNU CC. | |
8 | ||
9 | GNU CC is free software; you can redistribute it and/or modify | |
10 | it under the terms of the GNU General Public License as published by | |
11 | the Free Software Foundation; either version 2, or (at your option) | |
12 | any later version. | |
13 | ||
14 | GNU CC is distributed in the hope that it will be useful, | |
15 | but WITHOUT ANY WARRANTY; without even the implied warranty of | |
16 | MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the | |
17 | GNU General Public License for more details. | |
18 | ||
19 | You should have received a copy of the GNU General Public License | |
20 | along with GNU CC; see the file COPYING. If not, write to | |
c15c9075 RK |
21 | the Free Software Foundation, 59 Temple Place - Suite 330, |
22 | Boston, MA 02111-1307, USA. */ | |
ab835497 | 23 | |
ab835497 | 24 | #include "config.h" |
284d86e9 | 25 | #include "system.h" |
210aa14a | 26 | #include "tree.h" |
ab835497 RK |
27 | #include "rtl.h" |
28 | #include "regs.h" | |
29 | #include "hard-reg-set.h" | |
30 | #include "real.h" | |
31 | #include "insn-config.h" | |
32 | #include "conditions.h" | |
33 | #include "insn-flags.h" | |
34 | #include "output.h" | |
35 | #include "insn-attr.h" | |
36 | #include "flags.h" | |
49ad7cfa | 37 | #include "function.h" |
ab835497 RK |
38 | #include "expr.h" |
39 | #include "recog.h" | |
487a6e06 | 40 | #include "toplev.h" |
d07d525a | 41 | #include "ggc.h" |
b1474bb7 | 42 | #include "tm_p.h" |
ab835497 | 43 | |
61a55e8b DE |
44 | /* 1 if the caller has placed an "unimp" insn immediately after the call. |
45 | This is used in v8 code when calling a function that returns a structure. | |
06b967f9 RK |
46 | v9 doesn't have this. Be careful to have this test be the same as that |
47 | used on the call. */ | |
48 | ||
49 | #define SKIP_CALLERS_UNIMP_P \ | |
50 | (!TARGET_ARCH64 && current_function_returns_struct \ | |
51 | && ! integer_zerop (DECL_SIZE (DECL_RESULT (current_function_decl))) \ | |
52 | && (TREE_CODE (DECL_SIZE (DECL_RESULT (current_function_decl))) \ | |
53 | == INTEGER_CST)) | |
61a55e8b | 54 | |
ab835497 RK |
55 | /* Global variables for machine-dependent things. */ |
56 | ||
61a55e8b DE |
57 | /* Size of frame. Need to know this to emit return insns from leaf procedures. |
58 | ACTUAL_FSIZE is set by compute_frame_size() which is called during the | |
59 | reload pass. This is important as the value is later used in insn | |
60 | scheduling (to see what can go in a delay slot). | |
61 | APPARENT_FSIZE is the size of the stack less the register save area and less | |
62 | the outgoing argument area. It is used when saving call preserved regs. */ | |
63 | static int apparent_fsize; | |
64 | static int actual_fsize; | |
65 | ||
ab835497 RK |
66 | /* Save the operands last given to a compare for use when we |
67 | generate a scc or bcc insn. */ | |
68 | ||
69 | rtx sparc_compare_op0, sparc_compare_op1; | |
70 | ||
71 | /* We may need an epilogue if we spill too many registers. | |
72 | If this is non-zero, then we branch here for the epilogue. */ | |
73 | static rtx leaf_label; | |
74 | ||
75 | #ifdef LEAF_REGISTERS | |
76 | ||
77 | /* Vector to say how input registers are mapped to output | |
78 | registers. FRAME_POINTER_REGNUM cannot be remapped by | |
79 | this function to eliminate it. You must use -fomit-frame-pointer | |
80 | to get that. */ | |
81 | char leaf_reg_remap[] = | |
82 | { 0, 1, 2, 3, 4, 5, 6, 7, | |
83 | -1, -1, -1, -1, -1, -1, 14, -1, | |
84 | -1, -1, -1, -1, -1, -1, -1, -1, | |
85 | 8, 9, 10, 11, 12, 13, -1, 15, | |
86 | ||
87 | 32, 33, 34, 35, 36, 37, 38, 39, | |
88 | 40, 41, 42, 43, 44, 45, 46, 47, | |
89 | 48, 49, 50, 51, 52, 53, 54, 55, | |
61a55e8b DE |
90 | 56, 57, 58, 59, 60, 61, 62, 63, |
91 | 64, 65, 66, 67, 68, 69, 70, 71, | |
92 | 72, 73, 74, 75, 76, 77, 78, 79, | |
93 | 80, 81, 82, 83, 84, 85, 86, 87, | |
94 | 88, 89, 90, 91, 92, 93, 94, 95, | |
c4ce6853 | 95 | 96, 97, 98, 99, 100}; |
ab835497 | 96 | |
6a4bb1fa | 97 | #endif |
ab835497 | 98 | |
ab835497 RK |
99 | /* Name of where we pretend to think the frame pointer points. |
100 | Normally, this is "%fp", but if we are in a leaf procedure, | |
61a55e8b DE |
101 | this is "%sp+something". We record "something" separately as it may be |
102 | too big for reg+constant addressing. */ | |
103 | ||
3bb5de61 | 104 | static const char *frame_base_name; |
61a55e8b | 105 | static int frame_base_offset; |
ab835497 | 106 | |
c85f7c16 | 107 | static rtx pic_setup_code PROTO((void)); |
c85f7c16 | 108 | static void sparc_init_modes PROTO((void)); |
3bb5de61 | 109 | static int save_regs PROTO((FILE *, int, int, const char *, |
c85f7c16 | 110 | int, int, int)); |
3bb5de61 KG |
111 | static int restore_regs PROTO((FILE *, int, int, const char *, int, int)); |
112 | static void build_big_number PROTO((FILE *, int, const char *)); | |
5e9defae | 113 | static int function_arg_slotno PROTO((const CUMULATIVE_ARGS *, |
c85f7c16 JL |
114 | enum machine_mode, tree, int, int, |
115 | int *, int *)); | |
8947065c RH |
116 | |
117 | static int supersparc_adjust_cost PROTO((rtx, rtx, rtx, int)); | |
118 | static int hypersparc_adjust_cost PROTO((rtx, rtx, rtx, int)); | |
119 | static int ultrasparc_adjust_cost PROTO((rtx, rtx, rtx, int)); | |
120 | ||
e0d80184 DM |
121 | static void sparc_output_addr_vec PROTO((rtx)); |
122 | static void sparc_output_addr_diff_vec PROTO((rtx)); | |
123 | static void sparc_output_deferred_case_vectors PROTO((void)); | |
d07d525a MM |
124 | static void sparc_add_gc_roots PROTO ((void)); |
125 | static void mark_ultrasparc_pipeline_state PROTO ((void *)); | |
b1474bb7 KG |
126 | static int check_return_regs PROTO ((rtx)); |
127 | static void epilogue_renumber PROTO ((rtx *)); | |
128 | static int ultra_cmove_results_ready_p PROTO ((rtx)); | |
129 | static int ultra_fpmode_conflict_exists PROTO ((enum machine_mode)); | |
130 | static rtx *ultra_find_type PROTO ((int, rtx *, int)); | |
131 | static void ultra_build_types_avail PROTO ((rtx *, int)); | |
132 | static void ultra_flush_pipeline PROTO ((void)); | |
133 | static void ultra_rescan_pipeline_state PROTO ((rtx *, int)); | |
134 | static int set_extends PROTO ((rtx, rtx)); | |
61a55e8b DE |
135 | \f |
136 | /* Option handling. */ | |
137 | ||
a0a301fc | 138 | /* Code model option as passed by user. */ |
3bb5de61 | 139 | const char *sparc_cmodel_string; |
a0a301fc DE |
140 | /* Parsed value. */ |
141 | enum cmodel sparc_cmodel; | |
142 | ||
1cb36a98 RH |
143 | char sparc_hard_reg_printed[8]; |
144 | ||
16956f6e DE |
145 | struct sparc_cpu_select sparc_select[] = |
146 | { | |
147 | /* switch name, tune arch */ | |
148 | { (char *)0, "default", 1, 1 }, | |
149 | { (char *)0, "-mcpu=", 1, 1 }, | |
150 | { (char *)0, "-mtune=", 1, 0 }, | |
db3d4438 | 151 | { 0, 0, 0, 0 } |
16956f6e | 152 | }; |
733f53f5 | 153 | |
16956f6e DE |
154 | /* CPU type. This is set from TARGET_CPU_DEFAULT and -m{cpu,tune}=xxx. */ |
155 | enum processor_type sparc_cpu; | |
733f53f5 | 156 | |
61a55e8b DE |
157 | /* Validate and override various options, and do some machine dependent |
158 | initialization. */ | |
159 | ||
160 | void | |
161 | sparc_override_options () | |
162 | { | |
a0a301fc | 163 | static struct code_model { |
3bb5de61 | 164 | const char *name; |
a0a301fc DE |
165 | int value; |
166 | } cmodels[] = { | |
167 | { "32", CM_32 }, | |
168 | { "medlow", CM_MEDLOW }, | |
169 | { "medmid", CM_MEDMID }, | |
170 | { "medany", CM_MEDANY }, | |
171 | { "embmedany", CM_EMBMEDANY }, | |
172 | { 0, 0 } | |
173 | }; | |
174 | struct code_model *cmodel; | |
16956f6e | 175 | /* Map TARGET_CPU_DEFAULT to value for -m{arch,tune}=. */ |
733f53f5 DE |
176 | static struct cpu_default { |
177 | int cpu; | |
3bb5de61 | 178 | const char *name; |
733f53f5 | 179 | } cpu_default[] = { |
a0a301fc | 180 | /* There must be one entry here for each TARGET_CPU value. */ |
16956f6e | 181 | { TARGET_CPU_sparc, "cypress" }, |
3592ea0d | 182 | { TARGET_CPU_sparclet, "tsc701" }, |
16956f6e | 183 | { TARGET_CPU_sparclite, "f930" }, |
a0a301fc | 184 | { TARGET_CPU_v8, "v8" }, |
8947065c RH |
185 | { TARGET_CPU_hypersparc, "hypersparc" }, |
186 | { TARGET_CPU_sparclite86x, "sparclite86x" }, | |
a0a301fc DE |
187 | { TARGET_CPU_supersparc, "supersparc" }, |
188 | { TARGET_CPU_v9, "v9" }, | |
ba542005 | 189 | { TARGET_CPU_ultrasparc, "ultrasparc" }, |
db3d4438 | 190 | { 0, 0 } |
733f53f5 | 191 | }; |
16956f6e DE |
192 | struct cpu_default *def; |
193 | /* Table of values for -m{cpu,tune}=. */ | |
733f53f5 | 194 | static struct cpu_table { |
3bb5de61 | 195 | const char *name; |
16956f6e | 196 | enum processor_type processor; |
733f53f5 DE |
197 | int disable; |
198 | int enable; | |
199 | } cpu_table[] = { | |
16956f6e DE |
200 | { "v7", PROCESSOR_V7, MASK_ISA, 0 }, |
201 | { "cypress", PROCESSOR_CYPRESS, MASK_ISA, 0 }, | |
202 | { "v8", PROCESSOR_V8, MASK_ISA, MASK_V8 }, | |
733f53f5 | 203 | /* TI TMS390Z55 supersparc */ |
16956f6e DE |
204 | { "supersparc", PROCESSOR_SUPERSPARC, MASK_ISA, MASK_V8 }, |
205 | { "sparclite", PROCESSOR_SPARCLITE, MASK_ISA, MASK_SPARCLITE }, | |
733f53f5 DE |
206 | /* The Fujitsu MB86930 is the original sparclite chip, with no fpu. |
207 | The Fujitsu MB86934 is the recent sparclite chip, with an fpu. */ | |
16956f6e DE |
208 | { "f930", PROCESSOR_F930, MASK_ISA|MASK_FPU, MASK_SPARCLITE }, |
209 | { "f934", PROCESSOR_F934, MASK_ISA, MASK_SPARCLITE|MASK_FPU }, | |
8947065c | 210 | { "hypersparc", PROCESSOR_HYPERSPARC, MASK_ISA, MASK_V8|MASK_FPU }, |
d3ec6b06 VM |
211 | { "sparclite86x", PROCESSOR_SPARCLITE86X, MASK_ISA|MASK_FPU, |
212 | MASK_SPARCLITE }, | |
16956f6e | 213 | { "sparclet", PROCESSOR_SPARCLET, MASK_ISA, MASK_SPARCLET }, |
733f53f5 | 214 | /* TEMIC sparclet */ |
3592ea0d | 215 | { "tsc701", PROCESSOR_TSC701, MASK_ISA, MASK_SPARCLET }, |
16956f6e | 216 | { "v9", PROCESSOR_V9, MASK_ISA, MASK_V9 }, |
5cb01b65 JJ |
217 | /* TI ultrasparc I, II, IIi */ |
218 | { "ultrasparc", PROCESSOR_ULTRASPARC, MASK_ISA, MASK_V9 | |
219 | /* Although insns using %y are deprecated, it is a clear win on current | |
220 | ultrasparcs. */ | |
221 | |MASK_DEPRECATED_V8_INSNS }, | |
db3d4438 | 222 | { 0, 0, 0, 0 } |
733f53f5 | 223 | }; |
16956f6e DE |
224 | struct cpu_table *cpu; |
225 | struct sparc_cpu_select *sel; | |
a0a301fc | 226 | int fpu; |
345a6161 | 227 | |
a0a301fc DE |
228 | #ifndef SPARC_BI_ARCH |
229 | /* Check for unsupported architecture size. */ | |
230 | if (! TARGET_64BIT != DEFAULT_ARCH32_P) | |
231 | { | |
232 | error ("%s is not supported by this configuration", | |
233 | DEFAULT_ARCH32_P ? "-m64" : "-m32"); | |
234 | } | |
235 | #endif | |
236 | ||
345a6161 DM |
237 | /* At the moment we don't allow different pointer size and architecture */ |
238 | if (! TARGET_64BIT != ! TARGET_PTR64) | |
239 | { | |
240 | error ("-mptr%d not allowed on -m%d", | |
241 | TARGET_PTR64 ? 64 : 32, TARGET_64BIT ? 64 : 32); | |
242 | if (TARGET_64BIT) | |
243 | target_flags |= MASK_PTR64; | |
244 | else | |
245 | target_flags &= ~MASK_PTR64; | |
246 | } | |
247 | ||
a0a301fc DE |
248 | /* Code model selection. */ |
249 | sparc_cmodel = SPARC_DEFAULT_CMODEL; | |
345a6161 DM |
250 | |
251 | #ifdef SPARC_BI_ARCH | |
252 | if (TARGET_ARCH32) | |
253 | sparc_cmodel = CM_32; | |
254 | #endif | |
255 | ||
a0a301fc DE |
256 | if (sparc_cmodel_string != NULL) |
257 | { | |
258 | if (TARGET_ARCH64) | |
259 | { | |
260 | for (cmodel = &cmodels[0]; cmodel->name; cmodel++) | |
261 | if (strcmp (sparc_cmodel_string, cmodel->name) == 0) | |
262 | break; | |
263 | if (cmodel->name == NULL) | |
264 | error ("bad value (%s) for -mcmodel= switch", sparc_cmodel_string); | |
265 | else | |
266 | sparc_cmodel = cmodel->value; | |
267 | } | |
268 | else | |
269 | error ("-mcmodel= is not supported on 32 bit systems"); | |
270 | } | |
16956f6e | 271 | |
a0a301fc | 272 | fpu = TARGET_FPU; /* save current -mfpu status */ |
1f1406b4 | 273 | |
a0a301fc | 274 | /* Set the default CPU. */ |
16956f6e DE |
275 | for (def = &cpu_default[0]; def->name; ++def) |
276 | if (def->cpu == TARGET_CPU_DEFAULT) | |
277 | break; | |
278 | if (! def->name) | |
279 | abort (); | |
280 | sparc_select[0].string = def->name; | |
281 | ||
282 | for (sel = &sparc_select[0]; sel->name; ++sel) | |
61a55e8b | 283 | { |
16956f6e | 284 | if (sel->string) |
733f53f5 | 285 | { |
16956f6e DE |
286 | for (cpu = &cpu_table[0]; cpu->name; ++cpu) |
287 | if (! strcmp (sel->string, cpu->name)) | |
288 | { | |
289 | if (sel->set_tune_p) | |
290 | sparc_cpu = cpu->processor; | |
291 | ||
292 | if (sel->set_arch_p) | |
293 | { | |
294 | target_flags &= ~cpu->disable; | |
295 | target_flags |= cpu->enable; | |
296 | } | |
297 | break; | |
298 | } | |
299 | ||
300 | if (! cpu->name) | |
301 | error ("bad value (%s) for %s switch", sel->string, sel->name); | |
733f53f5 | 302 | } |
61a55e8b | 303 | } |
97da85b7 | 304 | |
1f1406b4 | 305 | /* If -mfpu or -mno-fpu was explicitly used, don't override with |
46cc13b3 RO |
306 | the processor default. Clear MASK_FPU_SET to avoid confusing |
307 | the reverse mapping from switch values to names. */ | |
1f1406b4 | 308 | if (TARGET_FPU_SET) |
46cc13b3 RO |
309 | { |
310 | target_flags = (target_flags & ~MASK_FPU) | fpu; | |
311 | target_flags &= ~MASK_FPU_SET; | |
312 | } | |
1f1406b4 | 313 | |
16956f6e DE |
314 | /* Use the deprecated v8 insns for sparc64 in 32 bit mode. */ |
315 | if (TARGET_V9 && TARGET_ARCH32) | |
733f53f5 | 316 | target_flags |= MASK_DEPRECATED_V8_INSNS; |
61a55e8b | 317 | |
345a6161 DM |
318 | /* V8PLUS requires V9, makes no sense in 64 bit mode. */ |
319 | if (! TARGET_V9 || TARGET_ARCH64) | |
284d86e9 JC |
320 | target_flags &= ~MASK_V8PLUS; |
321 | ||
82d6b402 RH |
322 | /* Don't use stack biasing in 32 bit mode. */ |
323 | if (TARGET_ARCH32) | |
324 | target_flags &= ~MASK_STACK_BIAS; | |
f952a238 JJ |
325 | |
326 | /* Don't allow -mvis if FPU is disabled. */ | |
327 | if (! TARGET_FPU) | |
328 | target_flags &= ~MASK_VIS; | |
82d6b402 | 329 | |
efa3896a GK |
330 | /* Supply a default value for align_functions. */ |
331 | if (align_functions == 0 && sparc_cpu == PROCESSOR_ULTRASPARC) | |
332 | align_functions = 32; | |
bf62bbf1 | 333 | |
82d6b402 RH |
334 | /* Validate PCC_STRUCT_RETURN. */ |
335 | if (flag_pcc_struct_return == DEFAULT_PCC_STRUCT_RETURN) | |
336 | flag_pcc_struct_return = (TARGET_ARCH64 ? 0 : 1); | |
337 | ||
61a55e8b DE |
338 | /* Do various machine dependent initializations. */ |
339 | sparc_init_modes (); | |
2be15d0f RH |
340 | |
341 | if ((profile_flag || profile_block_flag) | |
342 | && sparc_cmodel != CM_MEDLOW) | |
343 | { | |
344 | error ("profiling does not support code models other than medlow"); | |
345 | } | |
d07d525a MM |
346 | |
347 | /* Register global variables with the garbage collector. */ | |
348 | sparc_add_gc_roots (); | |
61a55e8b DE |
349 | } |
350 | \f | |
cd5fb1ee DE |
351 | /* Miscellaneous utilities. */ |
352 | ||
353 | /* Nonzero if CODE, a comparison, is suitable for use in v9 conditional move | |
354 | or branch on register contents instructions. */ | |
355 | ||
356 | int | |
357 | v9_regcmp_p (code) | |
358 | enum rtx_code code; | |
359 | { | |
360 | return (code == EQ || code == NE || code == GE || code == LT | |
361 | || code == LE || code == GT); | |
362 | } | |
bfd6bc60 | 363 | |
cd5fb1ee DE |
364 | \f |
365 | /* Operand constraints. */ | |
366 | ||
ab835497 | 367 | /* Return non-zero only if OP is a register of mode MODE, |
c4ce6853 DE |
368 | or const0_rtx. Don't allow const0_rtx if TARGET_LIVE_G0 because |
369 | %g0 may contain anything. */ | |
370 | ||
ab835497 RK |
371 | int |
372 | reg_or_0_operand (op, mode) | |
373 | rtx op; | |
374 | enum machine_mode mode; | |
375 | { | |
c4ce6853 DE |
376 | if (register_operand (op, mode)) |
377 | return 1; | |
378 | if (TARGET_LIVE_G0) | |
379 | return 0; | |
380 | if (op == const0_rtx) | |
ab835497 | 381 | return 1; |
5b486ce0 | 382 | if (GET_MODE (op) == VOIDmode && GET_CODE (op) == CONST_DOUBLE |
ab835497 RK |
383 | && CONST_DOUBLE_HIGH (op) == 0 |
384 | && CONST_DOUBLE_LOW (op) == 0) | |
385 | return 1; | |
302484ff JW |
386 | if (GET_MODE_CLASS (GET_MODE (op)) == MODE_FLOAT |
387 | && GET_CODE (op) == CONST_DOUBLE | |
388 | && fp_zero_operand (op)) | |
389 | return 1; | |
ab835497 RK |
390 | return 0; |
391 | } | |
392 | ||
302484ff | 393 | /* Nonzero if OP is a floating point value with value 0.0. */ |
c4ce6853 | 394 | |
302484ff JW |
395 | int |
396 | fp_zero_operand (op) | |
397 | rtx op; | |
398 | { | |
399 | REAL_VALUE_TYPE r; | |
400 | ||
401 | REAL_VALUE_FROM_CONST_DOUBLE (r, op); | |
d61ba4f4 | 402 | return (REAL_VALUES_EQUAL (r, dconst0) && ! REAL_VALUE_MINUS_ZERO (r)); |
302484ff JW |
403 | } |
404 | ||
61a55e8b DE |
405 | /* Nonzero if OP is an integer register. */ |
406 | ||
407 | int | |
408 | intreg_operand (op, mode) | |
409 | rtx op; | |
487a6e06 | 410 | enum machine_mode mode ATTRIBUTE_UNUSED; |
61a55e8b DE |
411 | { |
412 | return (register_operand (op, SImode) | |
a9e27770 | 413 | || (TARGET_ARCH64 && register_operand (op, DImode))); |
61a55e8b DE |
414 | } |
415 | ||
416 | /* Nonzero if OP is a floating point condition code register. */ | |
417 | ||
418 | int | |
c4ce6853 | 419 | fcc_reg_operand (op, mode) |
61a55e8b DE |
420 | rtx op; |
421 | enum machine_mode mode; | |
422 | { | |
423 | /* This can happen when recog is called from combine. Op may be a MEM. | |
424 | Fail instead of calling abort in this case. */ | |
c4ce6853 | 425 | if (GET_CODE (op) != REG) |
61a55e8b | 426 | return 0; |
304b7a23 | 427 | |
c4ce6853 | 428 | if (mode != VOIDmode && mode != GET_MODE (op)) |
61a55e8b | 429 | return 0; |
304b7a23 DE |
430 | if (mode == VOIDmode |
431 | && (GET_MODE (op) != CCFPmode && GET_MODE (op) != CCFPEmode)) | |
432 | return 0; | |
61a55e8b | 433 | |
c4ce6853 | 434 | #if 0 /* ??? ==> 1 when %fcc0-3 are pseudos first. See gen_compare_reg(). */ |
61a55e8b DE |
435 | if (reg_renumber == 0) |
436 | return REGNO (op) >= FIRST_PSEUDO_REGISTER; | |
437 | return REGNO_OK_FOR_CCFP_P (REGNO (op)); | |
438 | #else | |
c4ce6853 | 439 | return (unsigned) REGNO (op) - SPARC_FIRST_V9_FCC_REG < 4; |
61a55e8b DE |
440 | #endif |
441 | } | |
442 | ||
c4ce6853 DE |
443 | /* Nonzero if OP is an integer or floating point condition code register. */ |
444 | ||
445 | int | |
446 | icc_or_fcc_reg_operand (op, mode) | |
447 | rtx op; | |
448 | enum machine_mode mode; | |
449 | { | |
450 | if (GET_CODE (op) == REG && REGNO (op) == SPARC_ICC_REG) | |
304b7a23 DE |
451 | { |
452 | if (mode != VOIDmode && mode != GET_MODE (op)) | |
453 | return 0; | |
454 | if (mode == VOIDmode | |
455 | && GET_MODE (op) != CCmode && GET_MODE (op) != CCXmode) | |
456 | return 0; | |
457 | return 1; | |
458 | } | |
459 | ||
c4ce6853 DE |
460 | return fcc_reg_operand (op, mode); |
461 | } | |
462 | ||
ab835497 RK |
463 | /* Nonzero if OP can appear as the dest of a RESTORE insn. */ |
464 | int | |
465 | restore_operand (op, mode) | |
466 | rtx op; | |
467 | enum machine_mode mode; | |
468 | { | |
469 | return (GET_CODE (op) == REG && GET_MODE (op) == mode | |
470 | && (REGNO (op) < 8 || (REGNO (op) >= 24 && REGNO (op) < 32))); | |
471 | } | |
472 | ||
401cec23 JW |
473 | /* Call insn on SPARC can take a PC-relative constant address, or any regular |
474 | memory address. */ | |
ab835497 RK |
475 | |
476 | int | |
477 | call_operand (op, mode) | |
478 | rtx op; | |
479 | enum machine_mode mode; | |
480 | { | |
481 | if (GET_CODE (op) != MEM) | |
482 | abort (); | |
483 | op = XEXP (op, 0); | |
2445f289 | 484 | return (symbolic_operand (op, mode) || memory_address_p (Pmode, op)); |
ab835497 RK |
485 | } |
486 | ||
487 | int | |
488 | call_operand_address (op, mode) | |
489 | rtx op; | |
490 | enum machine_mode mode; | |
491 | { | |
2445f289 | 492 | return (symbolic_operand (op, mode) || memory_address_p (Pmode, op)); |
ab835497 RK |
493 | } |
494 | ||
495 | /* Returns 1 if OP is either a symbol reference or a sum of a symbol | |
496 | reference and a constant. */ | |
497 | ||
498 | int | |
499 | symbolic_operand (op, mode) | |
500 | register rtx op; | |
501 | enum machine_mode mode; | |
502 | { | |
6871dd65 RH |
503 | enum machine_mode omode = GET_MODE (op); |
504 | ||
505 | if (omode != mode && omode != VOIDmode && mode != VOIDmode) | |
506 | return 0; | |
507 | ||
ab835497 RK |
508 | switch (GET_CODE (op)) |
509 | { | |
510 | case SYMBOL_REF: | |
511 | case LABEL_REF: | |
512 | return 1; | |
513 | ||
514 | case CONST: | |
515 | op = XEXP (op, 0); | |
516 | return ((GET_CODE (XEXP (op, 0)) == SYMBOL_REF | |
517 | || GET_CODE (XEXP (op, 0)) == LABEL_REF) | |
518 | && GET_CODE (XEXP (op, 1)) == CONST_INT); | |
519 | ||
ab835497 RK |
520 | default: |
521 | return 0; | |
522 | } | |
523 | } | |
524 | ||
525 | /* Return truth value of statement that OP is a symbolic memory | |
526 | operand of mode MODE. */ | |
527 | ||
528 | int | |
529 | symbolic_memory_operand (op, mode) | |
530 | rtx op; | |
487a6e06 | 531 | enum machine_mode mode ATTRIBUTE_UNUSED; |
ab835497 RK |
532 | { |
533 | if (GET_CODE (op) == SUBREG) | |
534 | op = SUBREG_REG (op); | |
535 | if (GET_CODE (op) != MEM) | |
536 | return 0; | |
537 | op = XEXP (op, 0); | |
538 | return (GET_CODE (op) == SYMBOL_REF || GET_CODE (op) == CONST | |
539 | || GET_CODE (op) == HIGH || GET_CODE (op) == LABEL_REF); | |
540 | } | |
541 | ||
5ab7138b DE |
542 | /* Return truth value of statement that OP is a LABEL_REF of mode MODE. */ |
543 | ||
544 | int | |
545 | label_ref_operand (op, mode) | |
546 | rtx op; | |
547 | enum machine_mode mode; | |
548 | { | |
549 | if (GET_CODE (op) != LABEL_REF) | |
550 | return 0; | |
551 | if (GET_MODE (op) != mode) | |
552 | return 0; | |
553 | return 1; | |
554 | } | |
555 | ||
95726648 DE |
556 | /* Return 1 if the operand is an argument used in generating pic references |
557 | in either the medium/low or medium/anywhere code models of sparc64. */ | |
558 | ||
559 | int | |
560 | sp64_medium_pic_operand (op, mode) | |
561 | rtx op; | |
487a6e06 | 562 | enum machine_mode mode ATTRIBUTE_UNUSED; |
95726648 DE |
563 | { |
564 | /* Check for (const (minus (symbol_ref:GOT) | |
565 | (const (minus (label) (pc))))). */ | |
566 | if (GET_CODE (op) != CONST) | |
567 | return 0; | |
568 | op = XEXP (op, 0); | |
569 | if (GET_CODE (op) != MINUS) | |
570 | return 0; | |
571 | if (GET_CODE (XEXP (op, 0)) != SYMBOL_REF) | |
572 | return 0; | |
573 | /* ??? Ensure symbol is GOT. */ | |
574 | if (GET_CODE (XEXP (op, 1)) != CONST) | |
575 | return 0; | |
576 | if (GET_CODE (XEXP (XEXP (op, 1), 0)) != MINUS) | |
577 | return 0; | |
578 | return 1; | |
579 | } | |
580 | ||
61a55e8b DE |
581 | /* Return 1 if the operand is a data segment reference. This includes |
582 | the readonly data segment, or in other words anything but the text segment. | |
583 | This is needed in the medium/anywhere code model on v9. These values | |
a0a301fc | 584 | are accessed with EMBMEDANY_BASE_REG. */ |
61a55e8b DE |
585 | |
586 | int | |
587 | data_segment_operand (op, mode) | |
588 | rtx op; | |
487a6e06 | 589 | enum machine_mode mode ATTRIBUTE_UNUSED; |
61a55e8b DE |
590 | { |
591 | switch (GET_CODE (op)) | |
592 | { | |
593 | case SYMBOL_REF : | |
594 | return ! SYMBOL_REF_FLAG (op); | |
595 | case PLUS : | |
95726648 DE |
596 | /* Assume canonical format of symbol + constant. |
597 | Fall through. */ | |
61a55e8b | 598 | case CONST : |
b1474bb7 | 599 | return data_segment_operand (XEXP (op, 0), VOIDmode); |
61a55e8b DE |
600 | default : |
601 | return 0; | |
602 | } | |
603 | } | |
604 | ||
605 | /* Return 1 if the operand is a text segment reference. | |
606 | This is needed in the medium/anywhere code model on v9. */ | |
607 | ||
608 | int | |
609 | text_segment_operand (op, mode) | |
610 | rtx op; | |
487a6e06 | 611 | enum machine_mode mode ATTRIBUTE_UNUSED; |
61a55e8b DE |
612 | { |
613 | switch (GET_CODE (op)) | |
614 | { | |
615 | case LABEL_REF : | |
616 | return 1; | |
617 | case SYMBOL_REF : | |
618 | return SYMBOL_REF_FLAG (op); | |
619 | case PLUS : | |
95726648 DE |
620 | /* Assume canonical format of symbol + constant. |
621 | Fall through. */ | |
61a55e8b | 622 | case CONST : |
b1474bb7 | 623 | return text_segment_operand (XEXP (op, 0), VOIDmode); |
61a55e8b DE |
624 | default : |
625 | return 0; | |
626 | } | |
627 | } | |
628 | ||
ab835497 RK |
629 | /* Return 1 if the operand is either a register or a memory operand that is |
630 | not symbolic. */ | |
631 | ||
632 | int | |
633 | reg_or_nonsymb_mem_operand (op, mode) | |
634 | register rtx op; | |
635 | enum machine_mode mode; | |
636 | { | |
637 | if (register_operand (op, mode)) | |
638 | return 1; | |
639 | ||
640 | if (memory_operand (op, mode) && ! symbolic_memory_operand (op, mode)) | |
641 | return 1; | |
642 | ||
643 | return 0; | |
644 | } | |
645 | ||
ab835497 | 646 | int |
6ac34277 | 647 | splittable_symbolic_memory_operand (op, mode) |
ab835497 | 648 | rtx op; |
487a6e06 | 649 | enum machine_mode mode ATTRIBUTE_UNUSED; |
ab835497 | 650 | { |
6ac34277 DE |
651 | if (GET_CODE (op) != MEM) |
652 | return 0; | |
653 | if (! symbolic_operand (XEXP (op, 0), Pmode)) | |
654 | return 0; | |
655 | return 1; | |
656 | } | |
657 | ||
658 | int | |
659 | splittable_immediate_memory_operand (op, mode) | |
660 | rtx op; | |
487a6e06 | 661 | enum machine_mode mode ATTRIBUTE_UNUSED; |
6ac34277 DE |
662 | { |
663 | if (GET_CODE (op) != MEM) | |
664 | return 0; | |
665 | if (! immediate_operand (XEXP (op, 0), Pmode)) | |
666 | return 0; | |
667 | return 1; | |
ab835497 RK |
668 | } |
669 | ||
670 | /* Return truth value of whether OP is EQ or NE. */ | |
671 | ||
672 | int | |
673 | eq_or_neq (op, mode) | |
674 | rtx op; | |
487a6e06 | 675 | enum machine_mode mode ATTRIBUTE_UNUSED; |
ab835497 RK |
676 | { |
677 | return (GET_CODE (op) == EQ || GET_CODE (op) == NE); | |
678 | } | |
679 | ||
680 | /* Return 1 if this is a comparison operator, but not an EQ, NE, GEU, | |
681 | or LTU for non-floating-point. We handle those specially. */ | |
682 | ||
683 | int | |
684 | normal_comp_operator (op, mode) | |
685 | rtx op; | |
487a6e06 | 686 | enum machine_mode mode ATTRIBUTE_UNUSED; |
ab835497 RK |
687 | { |
688 | enum rtx_code code = GET_CODE (op); | |
689 | ||
690 | if (GET_RTX_CLASS (code) != '<') | |
691 | return 0; | |
692 | ||
4d449554 JW |
693 | if (GET_MODE (XEXP (op, 0)) == CCFPmode |
694 | || GET_MODE (XEXP (op, 0)) == CCFPEmode) | |
ab835497 RK |
695 | return 1; |
696 | ||
697 | return (code != NE && code != EQ && code != GEU && code != LTU); | |
698 | } | |
699 | ||
700 | /* Return 1 if this is a comparison operator. This allows the use of | |
701 | MATCH_OPERATOR to recognize all the branch insns. */ | |
702 | ||
703 | int | |
704 | noov_compare_op (op, mode) | |
705 | register rtx op; | |
487a6e06 | 706 | enum machine_mode mode ATTRIBUTE_UNUSED; |
ab835497 RK |
707 | { |
708 | enum rtx_code code = GET_CODE (op); | |
709 | ||
710 | if (GET_RTX_CLASS (code) != '<') | |
711 | return 0; | |
712 | ||
713 | if (GET_MODE (XEXP (op, 0)) == CC_NOOVmode) | |
714 | /* These are the only branches which work with CC_NOOVmode. */ | |
715 | return (code == EQ || code == NE || code == GE || code == LT); | |
716 | return 1; | |
717 | } | |
718 | ||
61a55e8b DE |
719 | /* Nonzero if OP is a comparison operator suitable for use in v9 |
720 | conditional move or branch on register contents instructions. */ | |
721 | ||
722 | int | |
723 | v9_regcmp_op (op, mode) | |
724 | register rtx op; | |
487a6e06 | 725 | enum machine_mode mode ATTRIBUTE_UNUSED; |
61a55e8b DE |
726 | { |
727 | enum rtx_code code = GET_CODE (op); | |
728 | ||
729 | if (GET_RTX_CLASS (code) != '<') | |
730 | return 0; | |
731 | ||
cd5fb1ee | 732 | return v9_regcmp_p (code); |
61a55e8b DE |
733 | } |
734 | ||
ab835497 RK |
735 | /* Return 1 if this is a SIGN_EXTEND or ZERO_EXTEND operation. */ |
736 | ||
737 | int | |
738 | extend_op (op, mode) | |
739 | rtx op; | |
487a6e06 | 740 | enum machine_mode mode ATTRIBUTE_UNUSED; |
ab835497 RK |
741 | { |
742 | return GET_CODE (op) == SIGN_EXTEND || GET_CODE (op) == ZERO_EXTEND; | |
743 | } | |
744 | ||
745 | /* Return nonzero if OP is an operator of mode MODE which can set | |
746 | the condition codes explicitly. We do not include PLUS and MINUS | |
747 | because these require CC_NOOVmode, which we handle explicitly. */ | |
748 | ||
749 | int | |
750 | cc_arithop (op, mode) | |
751 | rtx op; | |
487a6e06 | 752 | enum machine_mode mode ATTRIBUTE_UNUSED; |
ab835497 RK |
753 | { |
754 | if (GET_CODE (op) == AND | |
755 | || GET_CODE (op) == IOR | |
756 | || GET_CODE (op) == XOR) | |
757 | return 1; | |
758 | ||
759 | return 0; | |
760 | } | |
761 | ||
762 | /* Return nonzero if OP is an operator of mode MODE which can bitwise | |
763 | complement its second operand and set the condition codes explicitly. */ | |
764 | ||
765 | int | |
766 | cc_arithopn (op, mode) | |
767 | rtx op; | |
487a6e06 | 768 | enum machine_mode mode ATTRIBUTE_UNUSED; |
ab835497 RK |
769 | { |
770 | /* XOR is not here because combine canonicalizes (xor (not ...) ...) | |
771 | and (xor ... (not ...)) to (not (xor ...)). */ | |
772 | return (GET_CODE (op) == AND | |
773 | || GET_CODE (op) == IOR); | |
774 | } | |
775 | \f | |
58cbf20e DE |
776 | /* Return true if OP is a register, or is a CONST_INT that can fit in a |
777 | signed 13 bit immediate field. This is an acceptable SImode operand for | |
778 | most 3 address instructions. */ | |
ab835497 RK |
779 | |
780 | int | |
781 | arith_operand (op, mode) | |
782 | rtx op; | |
783 | enum machine_mode mode; | |
784 | { | |
284d86e9 | 785 | int val; |
11301057 | 786 | if (register_operand (op, mode)) |
284d86e9 JC |
787 | return 1; |
788 | if (GET_CODE (op) != CONST_INT) | |
789 | return 0; | |
790 | val = INTVAL (op) & 0xffffffff; | |
791 | return SPARC_SIMM13_P (val); | |
ab835497 RK |
792 | } |
793 | ||
5d6d3339 JJ |
794 | /* Return true if OP is a constant 4096 */ |
795 | ||
796 | int | |
797 | arith_4096_operand (op, mode) | |
798 | rtx op; | |
ddf80874 | 799 | enum machine_mode mode ATTRIBUTE_UNUSED; |
5d6d3339 JJ |
800 | { |
801 | int val; | |
802 | if (GET_CODE (op) != CONST_INT) | |
803 | return 0; | |
804 | val = INTVAL (op) & 0xffffffff; | |
805 | return val == 4096; | |
806 | } | |
807 | ||
808 | /* Return true if OP is suitable as second operand for add/sub */ | |
809 | ||
810 | int | |
811 | arith_add_operand (op, mode) | |
812 | rtx op; | |
813 | enum machine_mode mode; | |
814 | { | |
815 | return arith_operand (op, mode) || arith_4096_operand (op, mode); | |
816 | } | |
817 | ||
89e65674 DM |
818 | /* Return true if OP is a CONST_INT or a CONST_DOUBLE which can fit in the |
819 | immediate field of OR and XOR instructions. Used for 64-bit | |
820 | constant formation patterns. */ | |
821 | int | |
822 | const64_operand (op, mode) | |
823 | rtx op; | |
ddf80874 | 824 | enum machine_mode mode ATTRIBUTE_UNUSED; |
89e65674 DM |
825 | { |
826 | return ((GET_CODE (op) == CONST_INT | |
827 | && SPARC_SIMM13_P (INTVAL (op))) | |
9208e4b2 | 828 | #if HOST_BITS_PER_WIDE_INT != 64 |
89e65674 | 829 | || (GET_CODE (op) == CONST_DOUBLE |
9208e4b2 DM |
830 | && SPARC_SIMM13_P (CONST_DOUBLE_LOW (op)) |
831 | && (CONST_DOUBLE_HIGH (op) == | |
832 | ((CONST_DOUBLE_LOW (op) & 0x80000000) != 0 ? | |
ddf80874 | 833 | (HOST_WIDE_INT)0xffffffff : 0))) |
9208e4b2 | 834 | #endif |
11301057 | 835 | ); |
89e65674 DM |
836 | } |
837 | ||
9208e4b2 | 838 | /* The same, but only for sethi instructions. */ |
89e65674 DM |
839 | int |
840 | const64_high_operand (op, mode) | |
841 | rtx op; | |
ddf80874 | 842 | enum machine_mode mode ATTRIBUTE_UNUSED; |
89e65674 DM |
843 | { |
844 | return ((GET_CODE (op) == CONST_INT | |
9208e4b2 DM |
845 | && (INTVAL (op) & 0xfffffc00) != 0 |
846 | && SPARC_SETHI_P (INTVAL (op)) | |
847 | #if HOST_BITS_PER_WIDE_INT != 64 | |
848 | /* Must be positive on non-64bit host else the | |
849 | optimizer is fooled into thinking that sethi | |
850 | sign extends, even though it does not. */ | |
851 | && INTVAL (op) >= 0 | |
852 | #endif | |
853 | ) | |
89e65674 DM |
854 | || (GET_CODE (op) == CONST_DOUBLE |
855 | && CONST_DOUBLE_HIGH (op) == 0 | |
9208e4b2 | 856 | && (CONST_DOUBLE_LOW (op) & 0xfffffc00) != 0 |
11301057 | 857 | && SPARC_SETHI_P (CONST_DOUBLE_LOW (op)))); |
89e65674 DM |
858 | } |
859 | ||
58cbf20e DE |
860 | /* Return true if OP is a register, or is a CONST_INT that can fit in a |
861 | signed 11 bit immediate field. This is an acceptable SImode operand for | |
862 | the movcc instructions. */ | |
61a55e8b DE |
863 | |
864 | int | |
865 | arith11_operand (op, mode) | |
866 | rtx op; | |
867 | enum machine_mode mode; | |
868 | { | |
869 | return (register_operand (op, mode) | |
58cbf20e | 870 | || (GET_CODE (op) == CONST_INT && SPARC_SIMM11_P (INTVAL (op)))); |
61a55e8b DE |
871 | } |
872 | ||
58cbf20e DE |
873 | /* Return true if OP is a register, or is a CONST_INT that can fit in a |
874 | signed 10 bit immediate field. This is an acceptable SImode operand for | |
875 | the movrcc instructions. */ | |
61a55e8b DE |
876 | |
877 | int | |
878 | arith10_operand (op, mode) | |
879 | rtx op; | |
880 | enum machine_mode mode; | |
881 | { | |
882 | return (register_operand (op, mode) | |
58cbf20e | 883 | || (GET_CODE (op) == CONST_INT && SPARC_SIMM10_P (INTVAL (op)))); |
61a55e8b DE |
884 | } |
885 | ||
23b8a89f JW |
886 | /* Return true if OP is a register, is a CONST_INT that fits in a 13 bit |
887 | immediate field, or is a CONST_DOUBLE whose both parts fit in a 13 bit | |
61a55e8b DE |
888 | immediate field. |
889 | v9: Return true if OP is a register, or is a CONST_INT or CONST_DOUBLE that | |
890 | can fit in a 13 bit immediate field. This is an acceptable DImode operand | |
891 | for most 3 address instructions. */ | |
ab835497 RK |
892 | |
893 | int | |
894 | arith_double_operand (op, mode) | |
895 | rtx op; | |
896 | enum machine_mode mode; | |
897 | { | |
898 | return (register_operand (op, mode) | |
23b8a89f | 899 | || (GET_CODE (op) == CONST_INT && SMALL_INT (op)) |
a9e27770 | 900 | || (! TARGET_ARCH64 |
61a55e8b | 901 | && GET_CODE (op) == CONST_DOUBLE |
bf62bbf1 DE |
902 | && (unsigned HOST_WIDE_INT) (CONST_DOUBLE_LOW (op) + 0x1000) < 0x2000 |
903 | && (unsigned HOST_WIDE_INT) (CONST_DOUBLE_HIGH (op) + 0x1000) < 0x2000) | |
a9e27770 | 904 | || (TARGET_ARCH64 |
61a55e8b | 905 | && GET_CODE (op) == CONST_DOUBLE |
bf62bbf1 | 906 | && (unsigned HOST_WIDE_INT) (CONST_DOUBLE_LOW (op) + 0x1000) < 0x2000 |
61a55e8b DE |
907 | && ((CONST_DOUBLE_HIGH (op) == -1 |
908 | && (CONST_DOUBLE_LOW (op) & 0x1000) == 0x1000) | |
909 | || (CONST_DOUBLE_HIGH (op) == 0 | |
910 | && (CONST_DOUBLE_LOW (op) & 0x1000) == 0)))); | |
911 | } | |
912 | ||
5d6d3339 JJ |
913 | /* Return true if OP is a constant 4096 for DImode on ARCH64 */ |
914 | ||
915 | int | |
916 | arith_double_4096_operand (op, mode) | |
917 | rtx op; | |
ddf80874 | 918 | enum machine_mode mode ATTRIBUTE_UNUSED; |
5d6d3339 JJ |
919 | { |
920 | return (TARGET_ARCH64 && | |
921 | ((GET_CODE (op) == CONST_INT && INTVAL (op) == 4096) || | |
922 | (GET_CODE (op) == CONST_DOUBLE && | |
923 | CONST_DOUBLE_LOW (op) == 4096 && | |
924 | CONST_DOUBLE_HIGH (op) == 0))); | |
925 | } | |
926 | ||
927 | /* Return true if OP is suitable as second operand for add/sub in DImode */ | |
928 | ||
929 | int | |
930 | arith_double_add_operand (op, mode) | |
931 | rtx op; | |
932 | enum machine_mode mode; | |
933 | { | |
934 | return arith_double_operand (op, mode) || arith_double_4096_operand (op, mode); | |
935 | } | |
936 | ||
61a55e8b DE |
937 | /* Return true if OP is a register, or is a CONST_INT or CONST_DOUBLE that |
938 | can fit in an 11 bit immediate field. This is an acceptable DImode | |
939 | operand for the movcc instructions. */ | |
940 | /* ??? Replace with arith11_operand? */ | |
941 | ||
942 | int | |
943 | arith11_double_operand (op, mode) | |
944 | rtx op; | |
945 | enum machine_mode mode; | |
946 | { | |
947 | return (register_operand (op, mode) | |
948 | || (GET_CODE (op) == CONST_DOUBLE | |
949 | && (GET_MODE (op) == mode || GET_MODE (op) == VOIDmode) | |
bf62bbf1 | 950 | && (unsigned HOST_WIDE_INT) (CONST_DOUBLE_LOW (op) + 0x400) < 0x800 |
61a55e8b DE |
951 | && ((CONST_DOUBLE_HIGH (op) == -1 |
952 | && (CONST_DOUBLE_LOW (op) & 0x400) == 0x400) | |
953 | || (CONST_DOUBLE_HIGH (op) == 0 | |
954 | && (CONST_DOUBLE_LOW (op) & 0x400) == 0))) | |
955 | || (GET_CODE (op) == CONST_INT | |
956 | && (GET_MODE (op) == mode || GET_MODE (op) == VOIDmode) | |
bf62bbf1 | 957 | && (unsigned HOST_WIDE_INT) (INTVAL (op) + 0x400) < 0x800)); |
61a55e8b DE |
958 | } |
959 | ||
960 | /* Return true if OP is a register, or is a CONST_INT or CONST_DOUBLE that | |
961 | can fit in an 10 bit immediate field. This is an acceptable DImode | |
962 | operand for the movrcc instructions. */ | |
963 | /* ??? Replace with arith10_operand? */ | |
964 | ||
965 | int | |
966 | arith10_double_operand (op, mode) | |
967 | rtx op; | |
968 | enum machine_mode mode; | |
969 | { | |
970 | return (register_operand (op, mode) | |
971 | || (GET_CODE (op) == CONST_DOUBLE | |
972 | && (GET_MODE (op) == mode || GET_MODE (op) == VOIDmode) | |
973 | && (unsigned) (CONST_DOUBLE_LOW (op) + 0x200) < 0x400 | |
974 | && ((CONST_DOUBLE_HIGH (op) == -1 | |
975 | && (CONST_DOUBLE_LOW (op) & 0x200) == 0x200) | |
976 | || (CONST_DOUBLE_HIGH (op) == 0 | |
977 | && (CONST_DOUBLE_LOW (op) & 0x200) == 0))) | |
978 | || (GET_CODE (op) == CONST_INT | |
979 | && (GET_MODE (op) == mode || GET_MODE (op) == VOIDmode) | |
bf62bbf1 | 980 | && (unsigned HOST_WIDE_INT) (INTVAL (op) + 0x200) < 0x400)); |
ab835497 RK |
981 | } |
982 | ||
983 | /* Return truth value of whether OP is a integer which fits the | |
228b4037 JW |
984 | range constraining immediate operands in most three-address insns, |
985 | which have a 13 bit immediate field. */ | |
ab835497 RK |
986 | |
987 | int | |
988 | small_int (op, mode) | |
989 | rtx op; | |
487a6e06 | 990 | enum machine_mode mode ATTRIBUTE_UNUSED; |
ab835497 | 991 | { |
11301057 | 992 | return (GET_CODE (op) == CONST_INT && SMALL_INT (op)); |
ab835497 RK |
993 | } |
994 | ||
e0d80184 DM |
995 | int |
996 | small_int_or_double (op, mode) | |
997 | rtx op; | |
998 | enum machine_mode mode ATTRIBUTE_UNUSED; | |
999 | { | |
1000 | return ((GET_CODE (op) == CONST_INT && SMALL_INT (op)) | |
1001 | || (GET_CODE (op) == CONST_DOUBLE | |
1002 | && CONST_DOUBLE_HIGH (op) == 0 | |
11301057 | 1003 | && SPARC_SIMM13_P (CONST_DOUBLE_LOW (op)))); |
e0d80184 DM |
1004 | } |
1005 | ||
9b3fad96 JW |
1006 | /* Recognize operand values for the umul instruction. That instruction sign |
1007 | extends immediate values just like all other sparc instructions, but | |
1008 | interprets the extended result as an unsigned number. */ | |
1009 | ||
1010 | int | |
1011 | uns_small_int (op, mode) | |
1012 | rtx op; | |
487a6e06 | 1013 | enum machine_mode mode ATTRIBUTE_UNUSED; |
9b3fad96 JW |
1014 | { |
1015 | #if HOST_BITS_PER_WIDE_INT > 32 | |
1016 | /* All allowed constants will fit a CONST_INT. */ | |
11301057 RH |
1017 | return (GET_CODE (op) == CONST_INT |
1018 | && ((INTVAL (op) >= 0 && INTVAL (op) < 0x1000) | |
1019 | || (INTVAL (op) >= 0xFFFFF000 | |
1020 | && INTVAL (op) < 0x100000000))); | |
9b3fad96 | 1021 | #else |
11301057 RH |
1022 | return ((GET_CODE (op) == CONST_INT && (unsigned) INTVAL (op) < 0x1000) |
1023 | || (GET_CODE (op) == CONST_DOUBLE | |
1024 | && CONST_DOUBLE_HIGH (op) == 0 | |
1025 | && (unsigned) CONST_DOUBLE_LOW (op) - 0xFFFFF000 < 0x1000)); | |
9b3fad96 JW |
1026 | #endif |
1027 | } | |
1028 | ||
1029 | int | |
1030 | uns_arith_operand (op, mode) | |
1031 | rtx op; | |
1032 | enum machine_mode mode; | |
1033 | { | |
1034 | return register_operand (op, mode) || uns_small_int (op, mode); | |
1035 | } | |
1036 | ||
ab835497 RK |
1037 | /* Return truth value of statement that OP is a call-clobbered register. */ |
1038 | int | |
1039 | clobbered_register (op, mode) | |
1040 | rtx op; | |
487a6e06 | 1041 | enum machine_mode mode ATTRIBUTE_UNUSED; |
ab835497 RK |
1042 | { |
1043 | return (GET_CODE (op) == REG && call_used_regs[REGNO (op)]); | |
1044 | } | |
ab835497 | 1045 | |
e0d80184 DM |
1046 | /* Return 1 if OP is const0_rtx, used for TARGET_LIVE_G0 insns. */ |
1047 | ||
1048 | int | |
1049 | zero_operand (op, mode) | |
1050 | rtx op; | |
1051 | enum machine_mode mode ATTRIBUTE_UNUSED; | |
ab835497 | 1052 | { |
11301057 | 1053 | return op == const0_rtx; |
e0d80184 | 1054 | } |
61a55e8b | 1055 | |
e0d80184 | 1056 | /* Return 1 if OP is a valid operand for the source of a move insn. */ |
61a55e8b | 1057 | |
e0d80184 DM |
1058 | int |
1059 | input_operand (op, mode) | |
1060 | rtx op; | |
1061 | enum machine_mode mode; | |
1062 | { | |
1063 | /* If both modes are non-void they must be the same. */ | |
1064 | if (mode != VOIDmode && GET_MODE (op) != VOIDmode && mode != GET_MODE (op)) | |
1065 | return 0; | |
61a55e8b | 1066 | |
11301057 RH |
1067 | /* Only a tiny bit of handling for CONSTANT_P_RTX is necessary. */ |
1068 | if (GET_CODE (op) == CONST && GET_CODE (XEXP (op, 0)) == CONSTANT_P_RTX) | |
1069 | return 1; | |
1070 | ||
e0d80184 DM |
1071 | /* Allow any one instruction integer constant, and all CONST_INT |
1072 | variants when we are working in DImode and !arch64. */ | |
1073 | if (GET_MODE_CLASS (mode) == MODE_INT | |
9208e4b2 DM |
1074 | && ((GET_CODE (op) == CONST_INT |
1075 | && ((SPARC_SETHI_P (INTVAL (op)) | |
1076 | && (! TARGET_ARCH64 | |
1077 | || (INTVAL (op) >= 0) | |
1078 | || mode == SImode)) | |
1079 | || SPARC_SIMM13_P (INTVAL (op)) | |
1080 | || (mode == DImode | |
1081 | && ! TARGET_ARCH64))) | |
1082 | || (TARGET_ARCH64 | |
1083 | && GET_CODE (op) == CONST_DOUBLE | |
1084 | && ((CONST_DOUBLE_HIGH (op) == 0 | |
1085 | && SPARC_SETHI_P (CONST_DOUBLE_LOW (op))) | |
1086 | || | |
1087 | #if HOST_BITS_PER_WIDE_INT == 64 | |
1088 | (CONST_DOUBLE_HIGH (op) == 0 | |
1089 | && SPARC_SIMM13_P (CONST_DOUBLE_LOW (op))) | |
1090 | #else | |
1091 | (SPARC_SIMM13_P (CONST_DOUBLE_LOW (op)) | |
1092 | && (((CONST_DOUBLE_LOW (op) & 0x80000000) == 0 | |
1093 | && CONST_DOUBLE_HIGH (op) == 0) | |
1094 | || (CONST_DOUBLE_HIGH (op) == -1))) | |
1095 | #endif | |
1096 | )))) | |
e0d80184 DM |
1097 | return 1; |
1098 | ||
e0d80184 DM |
1099 | /* If !arch64 and this is a DImode const, allow it so that |
1100 | the splits can be generated. */ | |
1101 | if (! TARGET_ARCH64 | |
1102 | && mode == DImode | |
1103 | && GET_CODE (op) == CONST_DOUBLE) | |
1104 | return 1; | |
1105 | ||
1106 | if (register_operand (op, mode)) | |
1107 | return 1; | |
1108 | ||
1109 | /* If this is a SUBREG, look inside so that we handle | |
1110 | paradoxical ones. */ | |
1111 | if (GET_CODE (op) == SUBREG) | |
1112 | op = SUBREG_REG (op); | |
1113 | ||
1114 | /* Check for valid MEM forms. */ | |
1115 | if (GET_CODE (op) == MEM) | |
61a55e8b | 1116 | { |
e0d80184 | 1117 | rtx inside = XEXP (op, 0); |
61a55e8b | 1118 | |
e0d80184 | 1119 | if (GET_CODE (inside) == LO_SUM) |
03ad6f4d DM |
1120 | { |
1121 | /* We can't allow these because all of the splits | |
1122 | (eventually as they trickle down into DFmode | |
1123 | splits) require offsettable memory references. */ | |
1124 | if (! TARGET_V9 | |
1125 | && GET_MODE (op) == TFmode) | |
1126 | return 0; | |
1127 | ||
1128 | return (register_operand (XEXP (inside, 0), Pmode) | |
1129 | && CONSTANT_P (XEXP (inside, 1))); | |
1130 | } | |
e0d80184 | 1131 | return memory_address_p (mode, inside); |
61a55e8b | 1132 | } |
ab835497 | 1133 | |
e0d80184 DM |
1134 | return 0; |
1135 | } | |
1136 | ||
1137 | \f | |
1138 | /* We know it can't be done in one insn when we get here, | |
1139 | the movsi expander guarentees this. */ | |
1140 | void | |
1141 | sparc_emit_set_const32 (op0, op1) | |
1142 | rtx op0; | |
1143 | rtx op1; | |
1144 | { | |
1145 | enum machine_mode mode = GET_MODE (op0); | |
1146 | rtx temp; | |
1147 | ||
1148 | if (GET_CODE (op1) == CONST_INT) | |
284d86e9 | 1149 | { |
2a01c939 | 1150 | HOST_WIDE_INT value = INTVAL (op1); |
e0d80184 DM |
1151 | |
1152 | if (SPARC_SETHI_P (value) | |
1153 | || SPARC_SIMM13_P (value)) | |
1154 | abort (); | |
284d86e9 | 1155 | } |
e0d80184 DM |
1156 | |
1157 | /* Full 2-insn decomposition is needed. */ | |
1158 | if (reload_in_progress || reload_completed) | |
1159 | temp = op0; | |
284d86e9 | 1160 | else |
e0d80184 DM |
1161 | temp = gen_reg_rtx (mode); |
1162 | ||
71648202 DM |
1163 | if (GET_CODE (op1) == CONST_INT) |
1164 | { | |
1165 | /* Emit them as real moves instead of a HIGH/LO_SUM, | |
1166 | this way CSE can see everything and reuse intermediate | |
1167 | values if it wants. */ | |
1168 | if (TARGET_ARCH64 | |
1169 | && HOST_BITS_PER_WIDE_INT != 64 | |
1170 | && (INTVAL (op1) & 0x80000000) != 0) | |
1171 | { | |
1172 | emit_insn (gen_rtx_SET (VOIDmode, | |
1173 | temp, | |
1174 | gen_rtx_CONST_DOUBLE (VOIDmode, const0_rtx, | |
1175 | INTVAL (op1) & 0xfffffc00, 0))); | |
1176 | } | |
1177 | else | |
1178 | { | |
1179 | emit_insn (gen_rtx_SET (VOIDmode, | |
1180 | temp, | |
1181 | GEN_INT (INTVAL (op1) & 0xfffffc00))); | |
1182 | } | |
1183 | emit_insn (gen_rtx_SET (VOIDmode, | |
1184 | op0, | |
1185 | gen_rtx_IOR (mode, | |
1186 | temp, | |
1187 | GEN_INT (INTVAL (op1) & 0x3ff)))); | |
1188 | } | |
1189 | else | |
1190 | { | |
1191 | /* A symbol, emit in the traditional way. */ | |
1192 | emit_insn (gen_rtx_SET (VOIDmode, | |
1193 | temp, | |
1194 | gen_rtx_HIGH (mode, | |
1195 | op1))); | |
1196 | emit_insn (gen_rtx_SET (VOIDmode, | |
1197 | op0, | |
1198 | gen_rtx_LO_SUM (mode, | |
1199 | temp, | |
1200 | op1))); | |
1201 | ||
1202 | } | |
e0d80184 DM |
1203 | } |
1204 | ||
1205 | \f | |
1206 | /* Sparc-v9 code-model support. */ | |
1207 | void | |
1208 | sparc_emit_set_symbolic_const64 (op0, op1, temp1) | |
1209 | rtx op0; | |
1210 | rtx op1; | |
1211 | rtx temp1; | |
1212 | { | |
1213 | switch (sparc_cmodel) | |
284d86e9 | 1214 | { |
e0d80184 DM |
1215 | case CM_MEDLOW: |
1216 | /* The range spanned by all instructions in the object is less | |
1217 | than 2^31 bytes (2GB) and the distance from any instruction | |
1218 | to the location of the label _GLOBAL_OFFSET_TABLE_ is less | |
1219 | than 2^31 bytes (2GB). | |
ab835497 | 1220 | |
e0d80184 DM |
1221 | The executable must be in the low 4TB of the virtual address |
1222 | space. | |
61a55e8b | 1223 | |
e0d80184 DM |
1224 | sethi %hi(symbol), %temp |
1225 | or %temp, %lo(symbol), %reg */ | |
9208e4b2 DM |
1226 | emit_insn (gen_rtx_SET (VOIDmode, temp1, gen_rtx_HIGH (DImode, op1))); |
1227 | emit_insn (gen_rtx_SET (VOIDmode, op0, gen_rtx_LO_SUM (DImode, temp1, op1))); | |
e0d80184 | 1228 | break; |
61a55e8b | 1229 | |
e0d80184 DM |
1230 | case CM_MEDMID: |
1231 | /* The range spanned by all instructions in the object is less | |
1232 | than 2^31 bytes (2GB) and the distance from any instruction | |
1233 | to the location of the label _GLOBAL_OFFSET_TABLE_ is less | |
1234 | than 2^31 bytes (2GB). | |
1235 | ||
1236 | The executable must be in the low 16TB of the virtual address | |
1237 | space. | |
1238 | ||
1239 | sethi %h44(symbol), %temp1 | |
1240 | or %temp1, %m44(symbol), %temp2 | |
1241 | sllx %temp2, 12, %temp3 | |
1242 | or %temp3, %l44(symbol), %reg */ | |
1243 | emit_insn (gen_seth44 (op0, op1)); | |
1244 | emit_insn (gen_setm44 (op0, op0, op1)); | |
9208e4b2 | 1245 | emit_insn (gen_rtx_SET (VOIDmode, temp1, |
e0d80184 DM |
1246 | gen_rtx_ASHIFT (DImode, op0, GEN_INT (12)))); |
1247 | emit_insn (gen_setl44 (op0, temp1, op1)); | |
1248 | break; | |
61a55e8b | 1249 | |
e0d80184 DM |
1250 | case CM_MEDANY: |
1251 | /* The range spanned by all instructions in the object is less | |
1252 | than 2^31 bytes (2GB) and the distance from any instruction | |
1253 | to the location of the label _GLOBAL_OFFSET_TABLE_ is less | |
1254 | than 2^31 bytes (2GB). | |
1255 | ||
1256 | The executable can be placed anywhere in the virtual address | |
1257 | space. | |
1258 | ||
1259 | sethi %hh(symbol), %temp1 | |
1260 | sethi %lm(symbol), %temp2 | |
1261 | or %temp1, %hm(symbol), %temp3 | |
1262 | or %temp2, %lo(symbol), %temp4 | |
1263 | sllx %temp3, 32, %temp5 | |
1264 | or %temp4, %temp5, %reg */ | |
1265 | ||
1266 | /* Getting this right wrt. reloading is really tricky. | |
5722d61c | 1267 | We _MUST_ have a separate temporary at this point, |
e0d80184 DM |
1268 | if we don't barf immediately instead of generating |
1269 | incorrect code. */ | |
1270 | if (temp1 == op0) | |
1271 | abort (); | |
61a55e8b | 1272 | |
e0d80184 DM |
1273 | emit_insn (gen_sethh (op0, op1)); |
1274 | emit_insn (gen_setlm (temp1, op1)); | |
1275 | emit_insn (gen_sethm (op0, op0, op1)); | |
9208e4b2 | 1276 | emit_insn (gen_rtx_SET (VOIDmode, op0, |
e0d80184 | 1277 | gen_rtx_ASHIFT (DImode, op0, GEN_INT (32)))); |
9208e4b2 | 1278 | emit_insn (gen_rtx_SET (VOIDmode, op0, |
e0d80184 DM |
1279 | gen_rtx_PLUS (DImode, op0, temp1))); |
1280 | emit_insn (gen_setlo (op0, op0, op1)); | |
1281 | break; | |
61a55e8b | 1282 | |
e0d80184 DM |
1283 | case CM_EMBMEDANY: |
1284 | /* Old old old backwards compatibility kruft here. | |
1285 | Essentially it is MEDLOW with a fixed 64-bit | |
1286 | virtual base added to all data segment addresses. | |
1287 | Text-segment stuff is computed like MEDANY, we can't | |
1288 | reuse the code above because the relocation knobs | |
1289 | look different. | |
1290 | ||
1291 | Data segment: sethi %hi(symbol), %temp1 | |
1292 | or %temp1, %lo(symbol), %temp2 | |
1293 | add %temp2, EMBMEDANY_BASE_REG, %reg | |
1294 | ||
1295 | Text segment: sethi %uhi(symbol), %temp1 | |
1296 | sethi %hi(symbol), %temp2 | |
1297 | or %temp1, %ulo(symbol), %temp3 | |
1298 | or %temp2, %lo(symbol), %temp4 | |
1299 | sllx %temp3, 32, %temp5 | |
1300 | or %temp4, %temp5, %reg */ | |
1301 | if (data_segment_operand (op1, GET_MODE (op1))) | |
1302 | { | |
1303 | emit_insn (gen_embmedany_sethi (temp1, op1)); | |
1304 | emit_insn (gen_embmedany_brsum (op0, temp1)); | |
1305 | emit_insn (gen_embmedany_losum (op0, op0, op1)); | |
1306 | } | |
1307 | else | |
1308 | { | |
1309 | /* Getting this right wrt. reloading is really tricky. | |
5722d61c | 1310 | We _MUST_ have a separate temporary at this point, |
f710f868 | 1311 | so we barf immediately instead of generating |
e0d80184 DM |
1312 | incorrect code. */ |
1313 | if (temp1 == op0) | |
1314 | abort (); | |
61a55e8b | 1315 | |
e0d80184 DM |
1316 | emit_insn (gen_embmedany_textuhi (op0, op1)); |
1317 | emit_insn (gen_embmedany_texthi (temp1, op1)); | |
1318 | emit_insn (gen_embmedany_textulo (op0, op0, op1)); | |
9208e4b2 | 1319 | emit_insn (gen_rtx_SET (VOIDmode, op0, |
e0d80184 | 1320 | gen_rtx_ASHIFT (DImode, op0, GEN_INT (32)))); |
9208e4b2 | 1321 | emit_insn (gen_rtx_SET (VOIDmode, op0, |
e0d80184 DM |
1322 | gen_rtx_PLUS (DImode, op0, temp1))); |
1323 | emit_insn (gen_embmedany_textlo (op0, op0, op1)); | |
1324 | } | |
1325 | break; | |
1326 | ||
1327 | default: | |
1328 | abort(); | |
1329 | } | |
1330 | } | |
1331 | ||
9208e4b2 DM |
1332 | /* These avoid problems when cross compiling. If we do not |
1333 | go through all this hair then the optimizer will see | |
1334 | invalid REG_EQUAL notes or in some cases none at all. */ | |
1335 | static void sparc_emit_set_safe_HIGH64 PROTO ((rtx, HOST_WIDE_INT)); | |
1336 | static rtx gen_safe_SET64 PROTO ((rtx, HOST_WIDE_INT)); | |
1337 | static rtx gen_safe_OR64 PROTO ((rtx, HOST_WIDE_INT)); | |
1338 | static rtx gen_safe_XOR64 PROTO ((rtx, HOST_WIDE_INT)); | |
1339 | ||
1340 | #if HOST_BITS_PER_WIDE_INT == 64 | |
1341 | #define GEN_HIGHINT64(__x) GEN_INT ((__x) & 0xfffffc00) | |
1342 | #define GEN_INT64(__x) GEN_INT (__x) | |
1343 | #else | |
1344 | #define GEN_HIGHINT64(__x) \ | |
1345 | gen_rtx_CONST_DOUBLE (VOIDmode, const0_rtx, \ | |
1346 | (__x) & 0xfffffc00, 0) | |
1347 | #define GEN_INT64(__x) \ | |
1348 | gen_rtx_CONST_DOUBLE (VOIDmode, const0_rtx, \ | |
1349 | (__x) & 0xffffffff, \ | |
1350 | ((__x) & 0x80000000 \ | |
1351 | ? 0xffffffff : 0)) | |
1352 | #endif | |
1353 | ||
1354 | /* The optimizer is not to assume anything about exactly | |
1355 | which bits are set for a HIGH, they are unspecified. | |
1356 | Unfortunately this leads to many missed optimizations | |
1357 | during CSE. We mask out the non-HIGH bits, and matches | |
1358 | a plain movdi, to alleviate this problem. */ | |
1359 | static void | |
1360 | sparc_emit_set_safe_HIGH64 (dest, val) | |
1361 | rtx dest; | |
1362 | HOST_WIDE_INT val; | |
1363 | { | |
1364 | emit_insn (gen_rtx_SET (VOIDmode, dest, GEN_HIGHINT64 (val))); | |
1365 | } | |
2a01c939 | 1366 | |
e0d80184 | 1367 | static rtx |
9208e4b2 DM |
1368 | gen_safe_SET64 (dest, val) |
1369 | rtx dest; | |
e0d80184 | 1370 | HOST_WIDE_INT val; |
61a55e8b | 1371 | { |
9208e4b2 DM |
1372 | return gen_rtx_SET (VOIDmode, dest, GEN_INT64 (val)); |
1373 | } | |
1374 | ||
1375 | static rtx | |
1376 | gen_safe_OR64 (src, val) | |
1377 | rtx src; | |
1378 | HOST_WIDE_INT val; | |
1379 | { | |
1380 | return gen_rtx_IOR (DImode, src, GEN_INT64 (val)); | |
1381 | } | |
1382 | ||
1383 | static rtx | |
1384 | gen_safe_XOR64 (src, val) | |
1385 | rtx src; | |
1386 | HOST_WIDE_INT val; | |
1387 | { | |
1388 | return gen_rtx_XOR (DImode, src, GEN_INT64 (val)); | |
e0d80184 | 1389 | } |
61a55e8b | 1390 | |
e0d80184 DM |
1391 | /* Worker routines for 64-bit constant formation on arch64. |
1392 | One of the key things to be doing in these emissions is | |
1393 | to create as many temp REGs as possible. This makes it | |
1394 | possible for half-built constants to be used later when | |
1395 | such values are similar to something required later on. | |
1396 | Without doing this, the optimizer cannot see such | |
1397 | opportunities. */ | |
2a01c939 DM |
1398 | |
1399 | static void sparc_emit_set_const64_quick1 | |
1400 | PROTO((rtx, rtx, unsigned HOST_WIDE_INT, int)); | |
1401 | ||
e0d80184 DM |
1402 | static void |
1403 | sparc_emit_set_const64_quick1 (op0, temp, low_bits, is_neg) | |
1404 | rtx op0; | |
1405 | rtx temp; | |
2a01c939 | 1406 | unsigned HOST_WIDE_INT low_bits; |
e0d80184 DM |
1407 | int is_neg; |
1408 | { | |
2a01c939 | 1409 | unsigned HOST_WIDE_INT high_bits; |
a9e27770 | 1410 | |
e0d80184 | 1411 | if (is_neg) |
2a01c939 | 1412 | high_bits = (~low_bits) & 0xffffffff; |
31bc9eba | 1413 | else |
e0d80184 | 1414 | high_bits = low_bits; |
31bc9eba | 1415 | |
9208e4b2 | 1416 | sparc_emit_set_safe_HIGH64 (temp, high_bits); |
e0d80184 | 1417 | if (!is_neg) |
5108f3e5 DM |
1418 | { |
1419 | emit_insn (gen_rtx_SET (VOIDmode, op0, | |
1420 | gen_safe_OR64 (temp, (high_bits & 0x3ff)))); | |
1421 | } | |
e0d80184 | 1422 | else |
5108f3e5 DM |
1423 | { |
1424 | /* If we are XOR'ing with -1, then we should emit a one's complement | |
1425 | instead. This way the combiner will notice logical operations | |
1426 | such as ANDN later on and substitute. */ | |
1427 | if ((low_bits & 0x3ff) == 0x3ff) | |
1428 | { | |
1429 | emit_insn (gen_rtx_SET (VOIDmode, op0, | |
1430 | gen_rtx_NOT (DImode, temp))); | |
1431 | } | |
1432 | else | |
1433 | { | |
1434 | emit_insn (gen_rtx_SET (VOIDmode, op0, | |
1435 | gen_safe_XOR64 (temp, | |
1436 | (-0x400 | (low_bits & 0x3ff))))); | |
1437 | } | |
1438 | } | |
e0d80184 | 1439 | } |
61a55e8b | 1440 | |
2a01c939 DM |
1441 | static void sparc_emit_set_const64_quick2 |
1442 | PROTO((rtx, rtx, unsigned HOST_WIDE_INT, | |
1443 | unsigned HOST_WIDE_INT, int)); | |
1444 | ||
e0d80184 DM |
1445 | static void |
1446 | sparc_emit_set_const64_quick2 (op0, temp, high_bits, low_immediate, shift_count) | |
1447 | rtx op0; | |
1448 | rtx temp; | |
2a01c939 DM |
1449 | unsigned HOST_WIDE_INT high_bits; |
1450 | unsigned HOST_WIDE_INT low_immediate; | |
e0d80184 DM |
1451 | int shift_count; |
1452 | { | |
1453 | rtx temp2 = op0; | |
61a55e8b | 1454 | |
e0d80184 DM |
1455 | if ((high_bits & 0xfffffc00) != 0) |
1456 | { | |
9208e4b2 | 1457 | sparc_emit_set_safe_HIGH64 (temp, high_bits); |
e0d80184 | 1458 | if ((high_bits & ~0xfffffc00) != 0) |
9208e4b2 DM |
1459 | emit_insn (gen_rtx_SET (VOIDmode, op0, |
1460 | gen_safe_OR64 (temp, (high_bits & 0x3ff)))); | |
61a55e8b | 1461 | else |
e0d80184 | 1462 | temp2 = temp; |
61a55e8b DE |
1463 | } |
1464 | else | |
1465 | { | |
f710f868 | 1466 | emit_insn (gen_safe_SET64 (temp, high_bits)); |
e0d80184 DM |
1467 | temp2 = temp; |
1468 | } | |
61a55e8b | 1469 | |
e0d80184 | 1470 | /* Now shift it up into place. */ |
9208e4b2 | 1471 | emit_insn (gen_rtx_SET (VOIDmode, op0, |
e0d80184 DM |
1472 | gen_rtx_ASHIFT (DImode, temp2, |
1473 | GEN_INT (shift_count)))); | |
61a55e8b | 1474 | |
e0d80184 DM |
1475 | /* If there is a low immediate part piece, finish up by |
1476 | putting that in as well. */ | |
1477 | if (low_immediate != 0) | |
9208e4b2 DM |
1478 | emit_insn (gen_rtx_SET (VOIDmode, op0, |
1479 | gen_safe_OR64 (op0, low_immediate))); | |
ab835497 RK |
1480 | } |
1481 | ||
2a01c939 DM |
1482 | static void sparc_emit_set_const64_longway |
1483 | PROTO((rtx, rtx, unsigned HOST_WIDE_INT, unsigned HOST_WIDE_INT)); | |
1484 | ||
e0d80184 DM |
1485 | /* Full 64-bit constant decomposition. Even though this is the |
1486 | 'worst' case, we still optimize a few things away. */ | |
1487 | static void | |
1488 | sparc_emit_set_const64_longway (op0, temp, high_bits, low_bits) | |
1489 | rtx op0; | |
1490 | rtx temp; | |
2a01c939 DM |
1491 | unsigned HOST_WIDE_INT high_bits; |
1492 | unsigned HOST_WIDE_INT low_bits; | |
ab835497 | 1493 | { |
e0d80184 | 1494 | rtx sub_temp; |
4bf7135b | 1495 | |
e0d80184 DM |
1496 | if (reload_in_progress || reload_completed) |
1497 | sub_temp = op0; | |
1498 | else | |
1499 | sub_temp = gen_reg_rtx (DImode); | |
1500 | ||
1501 | if ((high_bits & 0xfffffc00) != 0) | |
1502 | { | |
9208e4b2 | 1503 | sparc_emit_set_safe_HIGH64 (temp, high_bits); |
e0d80184 | 1504 | if ((high_bits & ~0xfffffc00) != 0) |
9208e4b2 | 1505 | emit_insn (gen_rtx_SET (VOIDmode, |
e0d80184 | 1506 | sub_temp, |
9208e4b2 | 1507 | gen_safe_OR64 (temp, (high_bits & 0x3ff)))); |
e0d80184 DM |
1508 | else |
1509 | sub_temp = temp; | |
1510 | } | |
1511 | else | |
ab835497 | 1512 | { |
f710f868 | 1513 | emit_insn (gen_safe_SET64 (temp, high_bits)); |
e0d80184 | 1514 | sub_temp = temp; |
ab835497 | 1515 | } |
4bf7135b | 1516 | |
e0d80184 | 1517 | if (!reload_in_progress && !reload_completed) |
284d86e9 | 1518 | { |
e0d80184 DM |
1519 | rtx temp2 = gen_reg_rtx (DImode); |
1520 | rtx temp3 = gen_reg_rtx (DImode); | |
1521 | rtx temp4 = gen_reg_rtx (DImode); | |
284d86e9 | 1522 | |
9208e4b2 | 1523 | emit_insn (gen_rtx_SET (VOIDmode, temp4, |
e0d80184 DM |
1524 | gen_rtx_ASHIFT (DImode, sub_temp, |
1525 | GEN_INT (32)))); | |
4bf7135b | 1526 | |
9208e4b2 | 1527 | sparc_emit_set_safe_HIGH64 (temp2, low_bits); |
e0d80184 | 1528 | if ((low_bits & ~0xfffffc00) != 0) |
f710f868 DM |
1529 | { |
1530 | emit_insn (gen_rtx_SET (VOIDmode, temp3, | |
1531 | gen_safe_OR64 (temp2, (low_bits & 0x3ff)))); | |
1532 | emit_insn (gen_rtx_SET (VOIDmode, op0, | |
1533 | gen_rtx_PLUS (DImode, temp4, temp3))); | |
1534 | } | |
1535 | else | |
1536 | { | |
1537 | emit_insn (gen_rtx_SET (VOIDmode, op0, | |
1538 | gen_rtx_PLUS (DImode, temp4, temp2))); | |
1539 | } | |
284d86e9 | 1540 | } |
e0d80184 DM |
1541 | else |
1542 | { | |
9208e4b2 DM |
1543 | rtx low1 = GEN_INT ((low_bits >> (32 - 12)) & 0xfff); |
1544 | rtx low2 = GEN_INT ((low_bits >> (32 - 12 - 12)) & 0xfff); | |
1545 | rtx low3 = GEN_INT ((low_bits >> (32 - 12 - 12 - 8)) & 0x0ff); | |
e0d80184 | 1546 | int to_shift = 12; |
284d86e9 | 1547 | |
e0d80184 DM |
1548 | /* We are in the middle of reload, so this is really |
1549 | painful. However we do still make an attempt to | |
5108f3e5 | 1550 | avoid emitting truly stupid code. */ |
e0d80184 DM |
1551 | if (low1 != const0_rtx) |
1552 | { | |
9208e4b2 | 1553 | emit_insn (gen_rtx_SET (VOIDmode, op0, |
e0d80184 | 1554 | gen_rtx_ASHIFT (DImode, sub_temp, |
2a01c939 | 1555 | GEN_INT (to_shift)))); |
9208e4b2 | 1556 | emit_insn (gen_rtx_SET (VOIDmode, op0, |
e0d80184 DM |
1557 | gen_rtx_IOR (DImode, op0, low1))); |
1558 | sub_temp = op0; | |
1559 | to_shift = 12; | |
1560 | } | |
1561 | else | |
1562 | { | |
1563 | to_shift += 12; | |
1564 | } | |
1565 | if (low2 != const0_rtx) | |
1566 | { | |
9208e4b2 | 1567 | emit_insn (gen_rtx_SET (VOIDmode, op0, |
e0d80184 DM |
1568 | gen_rtx_ASHIFT (DImode, sub_temp, |
1569 | GEN_INT (to_shift)))); | |
9208e4b2 | 1570 | emit_insn (gen_rtx_SET (VOIDmode, op0, |
e0d80184 DM |
1571 | gen_rtx_IOR (DImode, op0, low2))); |
1572 | sub_temp = op0; | |
1573 | to_shift = 8; | |
1574 | } | |
1575 | else | |
1576 | { | |
1577 | to_shift += 8; | |
1578 | } | |
9208e4b2 | 1579 | emit_insn (gen_rtx_SET (VOIDmode, op0, |
e0d80184 DM |
1580 | gen_rtx_ASHIFT (DImode, sub_temp, |
1581 | GEN_INT (to_shift)))); | |
1582 | if (low3 != const0_rtx) | |
9208e4b2 | 1583 | emit_insn (gen_rtx_SET (VOIDmode, op0, |
e0d80184 DM |
1584 | gen_rtx_IOR (DImode, op0, low3))); |
1585 | /* phew... */ | |
1586 | } | |
284d86e9 JC |
1587 | } |
1588 | ||
e0d80184 | 1589 | /* Analyze a 64-bit constant for certain properties. */ |
2a01c939 DM |
1590 | static void analyze_64bit_constant |
1591 | PROTO((unsigned HOST_WIDE_INT, | |
1592 | unsigned HOST_WIDE_INT, | |
1593 | int *, int *, int *)); | |
1594 | ||
e0d80184 DM |
1595 | static void |
1596 | analyze_64bit_constant (high_bits, low_bits, hbsp, lbsp, abbasp) | |
2a01c939 | 1597 | unsigned HOST_WIDE_INT high_bits, low_bits; |
e0d80184 | 1598 | int *hbsp, *lbsp, *abbasp; |
ab835497 | 1599 | { |
e0d80184 DM |
1600 | int lowest_bit_set, highest_bit_set, all_bits_between_are_set; |
1601 | int i; | |
ab835497 | 1602 | |
e0d80184 DM |
1603 | lowest_bit_set = highest_bit_set = -1; |
1604 | i = 0; | |
1605 | do | |
ab835497 | 1606 | { |
e0d80184 DM |
1607 | if ((lowest_bit_set == -1) |
1608 | && ((low_bits >> i) & 1)) | |
1609 | lowest_bit_set = i; | |
1610 | if ((highest_bit_set == -1) | |
1611 | && ((high_bits >> (32 - i - 1)) & 1)) | |
1612 | highest_bit_set = (64 - i - 1); | |
1613 | } | |
1614 | while (++i < 32 | |
1615 | && ((highest_bit_set == -1) | |
1616 | || (lowest_bit_set == -1))); | |
1617 | if (i == 32) | |
1618 | { | |
1619 | i = 0; | |
1620 | do | |
ab835497 | 1621 | { |
e0d80184 DM |
1622 | if ((lowest_bit_set == -1) |
1623 | && ((high_bits >> i) & 1)) | |
1624 | lowest_bit_set = i + 32; | |
1625 | if ((highest_bit_set == -1) | |
1626 | && ((low_bits >> (32 - i - 1)) & 1)) | |
1627 | highest_bit_set = 32 - i - 1; | |
ab835497 | 1628 | } |
e0d80184 DM |
1629 | while (++i < 32 |
1630 | && ((highest_bit_set == -1) | |
1631 | || (lowest_bit_set == -1))); | |
1632 | } | |
1633 | /* If there are no bits set this should have gone out | |
1634 | as one instruction! */ | |
1635 | if (lowest_bit_set == -1 | |
1636 | || highest_bit_set == -1) | |
9208e4b2 | 1637 | abort (); |
e0d80184 DM |
1638 | all_bits_between_are_set = 1; |
1639 | for (i = lowest_bit_set; i <= highest_bit_set; i++) | |
1640 | { | |
1641 | if (i < 32) | |
1642 | { | |
1643 | if ((low_bits & (1 << i)) != 0) | |
1644 | continue; | |
1645 | } | |
1646 | else | |
1647 | { | |
1648 | if ((high_bits & (1 << (i - 32))) != 0) | |
1649 | continue; | |
1650 | } | |
1651 | all_bits_between_are_set = 0; | |
1652 | break; | |
ab835497 | 1653 | } |
e0d80184 DM |
1654 | *hbsp = highest_bit_set; |
1655 | *lbsp = lowest_bit_set; | |
1656 | *abbasp = all_bits_between_are_set; | |
ab835497 | 1657 | } |
6a4bb1fa | 1658 | |
2a01c939 DM |
1659 | static int const64_is_2insns |
1660 | PROTO((unsigned HOST_WIDE_INT, unsigned HOST_WIDE_INT)); | |
1661 | ||
e0d80184 DM |
1662 | static int |
1663 | const64_is_2insns (high_bits, low_bits) | |
2a01c939 | 1664 | unsigned HOST_WIDE_INT high_bits, low_bits; |
e0d80184 DM |
1665 | { |
1666 | int highest_bit_set, lowest_bit_set, all_bits_between_are_set; | |
284d86e9 | 1667 | |
e0d80184 | 1668 | if (high_bits == 0 |
2a01c939 | 1669 | || high_bits == 0xffffffff) |
e0d80184 | 1670 | return 1; |
6a4bb1fa | 1671 | |
e0d80184 DM |
1672 | analyze_64bit_constant (high_bits, low_bits, |
1673 | &highest_bit_set, &lowest_bit_set, | |
1674 | &all_bits_between_are_set); | |
6a4bb1fa | 1675 | |
f710f868 DM |
1676 | if ((highest_bit_set == 63 |
1677 | || lowest_bit_set == 0) | |
e0d80184 DM |
1678 | && all_bits_between_are_set != 0) |
1679 | return 1; | |
6a4bb1fa | 1680 | |
9208e4b2 DM |
1681 | if ((highest_bit_set - lowest_bit_set) < 21) |
1682 | return 1; | |
1683 | ||
6a4bb1fa DE |
1684 | return 0; |
1685 | } | |
1686 | ||
2a01c939 DM |
1687 | static unsigned HOST_WIDE_INT create_simple_focus_bits |
1688 | PROTO((unsigned HOST_WIDE_INT, unsigned HOST_WIDE_INT, | |
ddf80874 | 1689 | int, int)); |
2a01c939 DM |
1690 | |
1691 | static unsigned HOST_WIDE_INT | |
ddf80874 | 1692 | create_simple_focus_bits (high_bits, low_bits, lowest_bit_set, shift) |
2a01c939 | 1693 | unsigned HOST_WIDE_INT high_bits, low_bits; |
ddf80874 | 1694 | int lowest_bit_set, shift; |
ab835497 | 1695 | { |
f710f868 | 1696 | HOST_WIDE_INT hi, lo; |
ab835497 | 1697 | |
e0d80184 DM |
1698 | if (lowest_bit_set < 32) |
1699 | { | |
1700 | lo = (low_bits >> lowest_bit_set) << shift; | |
1701 | hi = ((high_bits << (32 - lowest_bit_set)) << shift); | |
1702 | } | |
1703 | else | |
1704 | { | |
1705 | lo = 0; | |
1706 | hi = ((high_bits >> (lowest_bit_set - 32)) << shift); | |
1707 | } | |
1708 | if (hi & lo) | |
9208e4b2 | 1709 | abort (); |
e0d80184 DM |
1710 | return (hi | lo); |
1711 | } | |
ab835497 | 1712 | |
e0d80184 DM |
1713 | /* Here we are sure to be arch64 and this is an integer constant |
1714 | being loaded into a register. Emit the most efficient | |
1715 | insn sequence possible. Detection of all the 1-insn cases | |
1716 | has been done already. */ | |
1717 | void | |
1718 | sparc_emit_set_const64 (op0, op1) | |
1719 | rtx op0; | |
1720 | rtx op1; | |
1721 | { | |
2a01c939 | 1722 | unsigned HOST_WIDE_INT high_bits, low_bits; |
e0d80184 DM |
1723 | int lowest_bit_set, highest_bit_set; |
1724 | int all_bits_between_are_set; | |
e0d80184 | 1725 | rtx temp; |
ab835497 | 1726 | |
e0d80184 DM |
1727 | /* Sanity check that we know what we are working with. */ |
1728 | if (! TARGET_ARCH64 | |
1729 | || GET_CODE (op0) != REG | |
1730 | || (REGNO (op0) >= SPARC_FIRST_FP_REG | |
1731 | && REGNO (op0) <= SPARC_LAST_V9_FP_REG)) | |
9208e4b2 | 1732 | abort (); |
95726648 | 1733 | |
f710f868 DM |
1734 | if (reload_in_progress || reload_completed) |
1735 | temp = op0; | |
1736 | else | |
1737 | temp = gen_reg_rtx (DImode); | |
1738 | ||
e0d80184 DM |
1739 | if (GET_CODE (op1) != CONST_DOUBLE |
1740 | && GET_CODE (op1) != CONST_INT) | |
1741 | { | |
5f78aa30 JC |
1742 | sparc_emit_set_symbolic_const64 (op0, op1, temp); |
1743 | return; | |
e0d80184 | 1744 | } |
ab835497 | 1745 | |
e0d80184 DM |
1746 | if (GET_CODE (op1) == CONST_DOUBLE) |
1747 | { | |
1748 | #if HOST_BITS_PER_WIDE_INT == 64 | |
2a01c939 | 1749 | high_bits = (CONST_DOUBLE_LOW (op1) >> 32) & 0xffffffff; |
e0d80184 DM |
1750 | low_bits = CONST_DOUBLE_LOW (op1) & 0xffffffff; |
1751 | #else | |
1752 | high_bits = CONST_DOUBLE_HIGH (op1); | |
1753 | low_bits = CONST_DOUBLE_LOW (op1); | |
1754 | #endif | |
ab835497 | 1755 | } |
e0d80184 | 1756 | else |
ab835497 | 1757 | { |
e0d80184 | 1758 | #if HOST_BITS_PER_WIDE_INT == 64 |
2a01c939 | 1759 | high_bits = ((INTVAL (op1) >> 32) & 0xffffffff); |
e0d80184 DM |
1760 | low_bits = (INTVAL (op1) & 0xffffffff); |
1761 | #else | |
1762 | high_bits = ((INTVAL (op1) < 0) ? | |
1763 | 0xffffffff : | |
1764 | 0x00000000); | |
1765 | low_bits = INTVAL (op1); | |
1766 | #endif | |
1767 | } | |
ab835497 | 1768 | |
e0d80184 DM |
1769 | /* low_bits bits 0 --> 31 |
1770 | high_bits bits 32 --> 63 */ | |
ab835497 | 1771 | |
e0d80184 DM |
1772 | analyze_64bit_constant (high_bits, low_bits, |
1773 | &highest_bit_set, &lowest_bit_set, | |
1774 | &all_bits_between_are_set); | |
1775 | ||
1776 | /* First try for a 2-insn sequence. */ | |
1777 | ||
1778 | /* These situations are preferred because the optimizer can | |
1779 | * do more things with them: | |
1780 | * 1) mov -1, %reg | |
1781 | * sllx %reg, shift, %reg | |
1782 | * 2) mov -1, %reg | |
1783 | * srlx %reg, shift, %reg | |
1784 | * 3) mov some_small_const, %reg | |
1785 | * sllx %reg, shift, %reg | |
1786 | */ | |
1787 | if (((highest_bit_set == 63 | |
1788 | || lowest_bit_set == 0) | |
1789 | && all_bits_between_are_set != 0) | |
2a01c939 | 1790 | || ((highest_bit_set - lowest_bit_set) < 12)) |
e0d80184 | 1791 | { |
9208e4b2 | 1792 | HOST_WIDE_INT the_const = -1; |
e0d80184 DM |
1793 | int shift = lowest_bit_set; |
1794 | ||
f710f868 DM |
1795 | if ((highest_bit_set != 63 |
1796 | && lowest_bit_set != 0) | |
1797 | || all_bits_between_are_set == 0) | |
ab835497 | 1798 | { |
e0d80184 | 1799 | the_const = |
9208e4b2 | 1800 | create_simple_focus_bits (high_bits, low_bits, |
9208e4b2 | 1801 | lowest_bit_set, 0); |
ab835497 | 1802 | } |
e0d80184 | 1803 | else if (lowest_bit_set == 0) |
9208e4b2 | 1804 | shift = -(63 - highest_bit_set); |
e0d80184 | 1805 | |
f710f868 DM |
1806 | if (! SPARC_SIMM13_P (the_const)) |
1807 | abort (); | |
1808 | ||
9208e4b2 | 1809 | emit_insn (gen_safe_SET64 (temp, the_const)); |
e0d80184 | 1810 | if (shift > 0) |
9208e4b2 | 1811 | emit_insn (gen_rtx_SET (VOIDmode, |
e0d80184 DM |
1812 | op0, |
1813 | gen_rtx_ASHIFT (DImode, | |
1814 | temp, | |
1815 | GEN_INT (shift)))); | |
1816 | else if (shift < 0) | |
9208e4b2 | 1817 | emit_insn (gen_rtx_SET (VOIDmode, |
e0d80184 | 1818 | op0, |
9208e4b2 | 1819 | gen_rtx_LSHIFTRT (DImode, |
e0d80184 DM |
1820 | temp, |
1821 | GEN_INT (-shift)))); | |
ab835497 RK |
1822 | else |
1823 | abort (); | |
e0d80184 DM |
1824 | return; |
1825 | } | |
ab835497 | 1826 | |
e0d80184 DM |
1827 | /* Now a range of 22 or less bits set somewhere. |
1828 | * 1) sethi %hi(focus_bits), %reg | |
1829 | * sllx %reg, shift, %reg | |
1830 | * 2) sethi %hi(focus_bits), %reg | |
1831 | * srlx %reg, shift, %reg | |
1832 | */ | |
2a01c939 | 1833 | if ((highest_bit_set - lowest_bit_set) < 21) |
e0d80184 | 1834 | { |
2a01c939 | 1835 | unsigned HOST_WIDE_INT focus_bits = |
e0d80184 | 1836 | create_simple_focus_bits (high_bits, low_bits, |
ddf80874 | 1837 | lowest_bit_set, 10); |
f710f868 DM |
1838 | |
1839 | if (! SPARC_SETHI_P (focus_bits)) | |
1840 | abort (); | |
1841 | ||
9208e4b2 | 1842 | sparc_emit_set_safe_HIGH64 (temp, focus_bits); |
e0d80184 | 1843 | |
9208e4b2 | 1844 | /* If lowest_bit_set == 10 then a sethi alone could have done it. */ |
e0d80184 | 1845 | if (lowest_bit_set < 10) |
9208e4b2 | 1846 | emit_insn (gen_rtx_SET (VOIDmode, |
e0d80184 | 1847 | op0, |
9208e4b2 | 1848 | gen_rtx_LSHIFTRT (DImode, temp, |
e0d80184 | 1849 | GEN_INT (10 - lowest_bit_set)))); |
9208e4b2 DM |
1850 | else if (lowest_bit_set > 10) |
1851 | emit_insn (gen_rtx_SET (VOIDmode, | |
e0d80184 DM |
1852 | op0, |
1853 | gen_rtx_ASHIFT (DImode, temp, | |
1854 | GEN_INT (lowest_bit_set - 10)))); | |
1855 | else | |
9208e4b2 | 1856 | abort (); |
e0d80184 DM |
1857 | return; |
1858 | } | |
1859 | ||
1860 | /* 1) sethi %hi(low_bits), %reg | |
1861 | * or %reg, %lo(low_bits), %reg | |
1862 | * 2) sethi %hi(~low_bits), %reg | |
1863 | * xor %reg, %lo(-0x400 | (low_bits & 0x3ff)), %reg | |
1864 | */ | |
1865 | if (high_bits == 0 | |
2a01c939 | 1866 | || high_bits == 0xffffffff) |
5f78aa30 JC |
1867 | { |
1868 | sparc_emit_set_const64_quick1 (op0, temp, low_bits, | |
1869 | (high_bits == 0xffffffff)); | |
1870 | return; | |
1871 | } | |
e0d80184 | 1872 | |
f710f868 DM |
1873 | /* Now, try 3-insn sequences. */ |
1874 | ||
e0d80184 DM |
1875 | /* 1) sethi %hi(high_bits), %reg |
1876 | * or %reg, %lo(high_bits), %reg | |
1877 | * sllx %reg, 32, %reg | |
1878 | */ | |
f710f868 | 1879 | if (low_bits == 0) |
5f78aa30 | 1880 | { |
f710f868 | 1881 | sparc_emit_set_const64_quick2 (op0, temp, high_bits, 0, 32); |
5f78aa30 JC |
1882 | return; |
1883 | } | |
e0d80184 | 1884 | |
f710f868 DM |
1885 | /* We may be able to do something quick |
1886 | when the constant is negated, so try that. */ | |
e0d80184 DM |
1887 | if (const64_is_2insns ((~high_bits) & 0xffffffff, |
1888 | (~low_bits) & 0xfffffc00)) | |
1889 | { | |
5108f3e5 DM |
1890 | /* NOTE: The trailing bits get XOR'd so we need the |
1891 | non-negated bits, not the negated ones. */ | |
1892 | unsigned HOST_WIDE_INT trailing_bits = low_bits & 0x3ff; | |
e0d80184 DM |
1893 | |
1894 | if ((((~high_bits) & 0xffffffff) == 0 | |
1895 | && ((~low_bits) & 0x80000000) == 0) | |
1896 | || (((~high_bits) & 0xffffffff) == 0xffffffff | |
1897 | && ((~low_bits) & 0x80000000) != 0)) | |
ab835497 | 1898 | { |
5108f3e5 | 1899 | int fast_int = (~low_bits & 0xffffffff); |
e0d80184 | 1900 | |
5108f3e5 DM |
1901 | if ((SPARC_SETHI_P (fast_int) |
1902 | && (~high_bits & 0xffffffff) == 0) | |
1903 | || SPARC_SIMM13_P (fast_int)) | |
9208e4b2 | 1904 | emit_insn (gen_safe_SET64 (temp, fast_int)); |
ab835497 | 1905 | else |
9208e4b2 | 1906 | sparc_emit_set_const64 (temp, GEN_INT64 (fast_int)); |
ab835497 | 1907 | } |
e0d80184 DM |
1908 | else |
1909 | { | |
1910 | rtx negated_const; | |
1911 | #if HOST_BITS_PER_WIDE_INT == 64 | |
1912 | negated_const = GEN_INT (((~low_bits) & 0xfffffc00) | | |
1913 | (((HOST_WIDE_INT)((~high_bits) & 0xffffffff))<<32)); | |
1914 | #else | |
9208e4b2 | 1915 | negated_const = gen_rtx_CONST_DOUBLE (DImode, const0_rtx, |
e0d80184 DM |
1916 | (~low_bits) & 0xfffffc00, |
1917 | (~high_bits) & 0xffffffff); | |
1918 | #endif | |
1919 | sparc_emit_set_const64 (temp, negated_const); | |
1920 | } | |
5108f3e5 DM |
1921 | |
1922 | /* If we are XOR'ing with -1, then we should emit a one's complement | |
1923 | instead. This way the combiner will notice logical operations | |
1924 | such as ANDN later on and substitute. */ | |
1925 | if (trailing_bits == 0x3ff) | |
1926 | { | |
1927 | emit_insn (gen_rtx_SET (VOIDmode, op0, | |
1928 | gen_rtx_NOT (DImode, temp))); | |
1929 | } | |
1930 | else | |
1931 | { | |
1932 | emit_insn (gen_rtx_SET (VOIDmode, | |
1933 | op0, | |
1934 | gen_safe_XOR64 (temp, | |
1935 | (-0x400 | trailing_bits)))); | |
1936 | } | |
e0d80184 | 1937 | return; |
ab835497 | 1938 | } |
ab835497 | 1939 | |
e0d80184 DM |
1940 | /* 1) sethi %hi(xxx), %reg |
1941 | * or %reg, %lo(xxx), %reg | |
1942 | * sllx %reg, yyy, %reg | |
f710f868 DM |
1943 | * |
1944 | * ??? This is just a generalized version of the low_bits==0 | |
1945 | * thing above, FIXME... | |
e0d80184 DM |
1946 | */ |
1947 | if ((highest_bit_set - lowest_bit_set) < 32) | |
1948 | { | |
f710f868 DM |
1949 | unsigned HOST_WIDE_INT focus_bits = |
1950 | create_simple_focus_bits (high_bits, low_bits, | |
ddf80874 | 1951 | lowest_bit_set, 0); |
ab835497 | 1952 | |
e0d80184 DM |
1953 | /* We can't get here in this state. */ |
1954 | if (highest_bit_set < 32 | |
1955 | || lowest_bit_set >= 32) | |
9208e4b2 | 1956 | abort (); |
95726648 | 1957 | |
e0d80184 DM |
1958 | /* So what we know is that the set bits straddle the |
1959 | middle of the 64-bit word. */ | |
5f78aa30 JC |
1960 | sparc_emit_set_const64_quick2 (op0, temp, |
1961 | focus_bits, 0, | |
1962 | lowest_bit_set); | |
1963 | return; | |
e0d80184 | 1964 | } |
c85f7c16 | 1965 | |
f710f868 DM |
1966 | /* 1) sethi %hi(high_bits), %reg |
1967 | * or %reg, %lo(high_bits), %reg | |
1968 | * sllx %reg, 32, %reg | |
1969 | * or %reg, low_bits, %reg | |
1970 | */ | |
1971 | if (SPARC_SIMM13_P(low_bits) | |
1972 | && ((int)low_bits > 0)) | |
1973 | { | |
1974 | sparc_emit_set_const64_quick2 (op0, temp, high_bits, low_bits, 32); | |
1975 | return; | |
1976 | } | |
1977 | ||
e0d80184 DM |
1978 | /* The easiest way when all else fails, is full decomposition. */ |
1979 | #if 0 | |
1980 | printf ("sparc_emit_set_const64: Hard constant [%08lx%08lx] neg[%08lx%08lx]\n", | |
1981 | high_bits, low_bits, ~high_bits, ~low_bits); | |
1982 | #endif | |
1983 | sparc_emit_set_const64_longway (op0, temp, high_bits, low_bits); | |
c85f7c16 JL |
1984 | } |
1985 | ||
e0d80184 DM |
1986 | /* X and Y are two things to compare using CODE. Emit the compare insn and |
1987 | return the rtx for the cc reg in the proper mode. */ | |
c85f7c16 | 1988 | |
e0d80184 DM |
1989 | rtx |
1990 | gen_compare_reg (code, x, y) | |
1991 | enum rtx_code code; | |
1992 | rtx x, y; | |
c85f7c16 | 1993 | { |
e0d80184 DM |
1994 | enum machine_mode mode = SELECT_CC_MODE (code, x, y); |
1995 | rtx cc_reg; | |
c85f7c16 | 1996 | |
e0d80184 DM |
1997 | /* ??? We don't have movcc patterns so we cannot generate pseudo regs for the |
1998 | fcc regs (cse can't tell they're really call clobbered regs and will | |
1999 | remove a duplicate comparison even if there is an intervening function | |
2000 | call - it will then try to reload the cc reg via an int reg which is why | |
2001 | we need the movcc patterns). It is possible to provide the movcc | |
2002 | patterns by using the ldxfsr/stxfsr v9 insns. I tried it: you need two | |
2003 | registers (say %g1,%g5) and it takes about 6 insns. A better fix would be | |
2004 | to tell cse that CCFPE mode registers (even pseudos) are call | |
2005 | clobbered. */ | |
c85f7c16 | 2006 | |
e0d80184 DM |
2007 | /* ??? This is an experiment. Rather than making changes to cse which may |
2008 | or may not be easy/clean, we do our own cse. This is possible because | |
2009 | we will generate hard registers. Cse knows they're call clobbered (it | |
2010 | doesn't know the same thing about pseudos). If we guess wrong, no big | |
2011 | deal, but if we win, great! */ | |
c85f7c16 | 2012 | |
e0d80184 DM |
2013 | if (TARGET_V9 && GET_MODE_CLASS (GET_MODE (x)) == MODE_FLOAT) |
2014 | #if 1 /* experiment */ | |
284d86e9 | 2015 | { |
e0d80184 DM |
2016 | int reg; |
2017 | /* We cycle through the registers to ensure they're all exercised. */ | |
2018 | static int next_fcc_reg = 0; | |
2019 | /* Previous x,y for each fcc reg. */ | |
2020 | static rtx prev_args[4][2]; | |
284d86e9 | 2021 | |
e0d80184 DM |
2022 | /* Scan prev_args for x,y. */ |
2023 | for (reg = 0; reg < 4; reg++) | |
2024 | if (prev_args[reg][0] == x && prev_args[reg][1] == y) | |
2025 | break; | |
2026 | if (reg == 4) | |
2027 | { | |
2028 | reg = next_fcc_reg; | |
2029 | prev_args[reg][0] = x; | |
2030 | prev_args[reg][1] = y; | |
2031 | next_fcc_reg = (next_fcc_reg + 1) & 3; | |
2032 | } | |
2033 | cc_reg = gen_rtx_REG (mode, reg + SPARC_FIRST_V9_FCC_REG); | |
284d86e9 | 2034 | } |
e0d80184 DM |
2035 | #else |
2036 | cc_reg = gen_reg_rtx (mode); | |
2037 | #endif /* ! experiment */ | |
2038 | else if (GET_MODE_CLASS (GET_MODE (x)) == MODE_FLOAT) | |
2039 | cc_reg = gen_rtx_REG (mode, SPARC_FCC_REG); | |
2040 | else | |
2041 | cc_reg = gen_rtx_REG (mode, SPARC_ICC_REG); | |
284d86e9 | 2042 | |
e0d80184 DM |
2043 | emit_insn (gen_rtx_SET (VOIDmode, cc_reg, |
2044 | gen_rtx_COMPARE (mode, x, y))); | |
c85f7c16 | 2045 | |
e0d80184 DM |
2046 | return cc_reg; |
2047 | } | |
c85f7c16 | 2048 | |
e0d80184 DM |
2049 | /* This function is used for v9 only. |
2050 | CODE is the code for an Scc's comparison. | |
2051 | OPERANDS[0] is the target of the Scc insn. | |
2052 | OPERANDS[1] is the value we compare against const0_rtx (which hasn't | |
2053 | been generated yet). | |
c85f7c16 | 2054 | |
e0d80184 | 2055 | This function is needed to turn |
ab835497 | 2056 | |
e0d80184 DM |
2057 | (set (reg:SI 110) |
2058 | (gt (reg:CCX 100 %icc) | |
2059 | (const_int 0))) | |
2060 | into | |
2061 | (set (reg:SI 110) | |
2062 | (gt:DI (reg:CCX 100 %icc) | |
2063 | (const_int 0))) | |
2064 | ||
2065 | IE: The instruction recognizer needs to see the mode of the comparison to | |
2066 | find the right instruction. We could use "gt:DI" right in the | |
2067 | define_expand, but leaving it out allows us to handle DI, SI, etc. | |
ab835497 | 2068 | |
e0d80184 DM |
2069 | We refer to the global sparc compare operands sparc_compare_op0 and |
2070 | sparc_compare_op1. */ | |
ab835497 RK |
2071 | |
2072 | int | |
e0d80184 DM |
2073 | gen_v9_scc (compare_code, operands) |
2074 | enum rtx_code compare_code; | |
2075 | register rtx *operands; | |
ab835497 | 2076 | { |
e0d80184 | 2077 | rtx temp, op0, op1; |
ab835497 | 2078 | |
e0d80184 DM |
2079 | if (! TARGET_ARCH64 |
2080 | && (GET_MODE (sparc_compare_op0) == DImode | |
2081 | || GET_MODE (operands[0]) == DImode)) | |
2082 | return 0; | |
7e2feebf | 2083 | |
e0d80184 DM |
2084 | /* Handle the case where operands[0] == sparc_compare_op0. |
2085 | We "early clobber" the result. */ | |
2086 | if (REGNO (operands[0]) == REGNO (sparc_compare_op0)) | |
ab835497 | 2087 | { |
e0d80184 DM |
2088 | op0 = gen_reg_rtx (GET_MODE (sparc_compare_op0)); |
2089 | emit_move_insn (op0, sparc_compare_op0); | |
ab835497 | 2090 | } |
e0d80184 DM |
2091 | else |
2092 | op0 = sparc_compare_op0; | |
2093 | /* For consistency in the following. */ | |
2094 | op1 = sparc_compare_op1; | |
2095 | ||
2096 | /* Try to use the movrCC insns. */ | |
2097 | if (TARGET_ARCH64 | |
2098 | && GET_MODE_CLASS (GET_MODE (op0)) == MODE_INT | |
2099 | && op1 == const0_rtx | |
2100 | && v9_regcmp_p (compare_code)) | |
ab835497 | 2101 | { |
e0d80184 DM |
2102 | /* Special case for op0 != 0. This can be done with one instruction if |
2103 | operands[0] == sparc_compare_op0. We don't assume they are equal | |
2104 | now though. */ | |
2105 | ||
2106 | if (compare_code == NE | |
2107 | && GET_MODE (operands[0]) == DImode | |
2108 | && GET_MODE (op0) == DImode) | |
ab835497 | 2109 | { |
e0d80184 DM |
2110 | emit_insn (gen_rtx_SET (VOIDmode, operands[0], op0)); |
2111 | emit_insn (gen_rtx_SET (VOIDmode, operands[0], | |
2112 | gen_rtx_IF_THEN_ELSE (DImode, | |
2113 | gen_rtx_fmt_ee (compare_code, DImode, | |
2114 | op0, const0_rtx), | |
2115 | const1_rtx, | |
2116 | operands[0]))); | |
ab835497 RK |
2117 | return 1; |
2118 | } | |
e0d80184 DM |
2119 | |
2120 | emit_insn (gen_rtx_SET (VOIDmode, operands[0], const0_rtx)); | |
2121 | if (GET_MODE (op0) != DImode) | |
ab835497 | 2122 | { |
e0d80184 DM |
2123 | temp = gen_reg_rtx (DImode); |
2124 | convert_move (temp, op0, 0); | |
ab835497 | 2125 | } |
e0d80184 DM |
2126 | else |
2127 | temp = op0; | |
2128 | emit_insn (gen_rtx_SET (VOIDmode, operands[0], | |
2129 | gen_rtx_IF_THEN_ELSE (GET_MODE (operands[0]), | |
2130 | gen_rtx_fmt_ee (compare_code, DImode, | |
2131 | temp, const0_rtx), | |
2132 | const1_rtx, | |
2133 | operands[0]))); | |
95726648 DE |
2134 | return 1; |
2135 | } | |
e0d80184 | 2136 | else |
ab835497 | 2137 | { |
e0d80184 | 2138 | operands[1] = gen_compare_reg (compare_code, op0, op1); |
ab835497 | 2139 | |
e0d80184 | 2140 | switch (GET_MODE (operands[1])) |
ab835497 | 2141 | { |
e0d80184 DM |
2142 | case CCmode : |
2143 | case CCXmode : | |
2144 | case CCFPEmode : | |
2145 | case CCFPmode : | |
2146 | break; | |
2147 | default : | |
2148 | abort (); | |
ab835497 | 2149 | } |
e0d80184 DM |
2150 | emit_insn (gen_rtx_SET (VOIDmode, operands[0], const0_rtx)); |
2151 | emit_insn (gen_rtx_SET (VOIDmode, operands[0], | |
2152 | gen_rtx_IF_THEN_ELSE (GET_MODE (operands[0]), | |
2153 | gen_rtx_fmt_ee (compare_code, | |
2154 | GET_MODE (operands[1]), | |
2155 | operands[1], const0_rtx), | |
2156 | const1_rtx, operands[0]))); | |
2157 | return 1; | |
ab835497 | 2158 | } |
ab835497 | 2159 | } |
4fb4e4b8 | 2160 | |
e0d80184 DM |
2161 | /* Emit a conditional jump insn for the v9 architecture using comparison code |
2162 | CODE and jump target LABEL. | |
2163 | This function exists to take advantage of the v9 brxx insns. */ | |
ab835497 | 2164 | |
e0d80184 DM |
2165 | void |
2166 | emit_v9_brxx_insn (code, op0, label) | |
2167 | enum rtx_code code; | |
2168 | rtx op0, label; | |
ab835497 | 2169 | { |
e0d80184 DM |
2170 | emit_jump_insn (gen_rtx_SET (VOIDmode, |
2171 | pc_rtx, | |
2172 | gen_rtx_IF_THEN_ELSE (VOIDmode, | |
2173 | gen_rtx_fmt_ee (code, GET_MODE (op0), | |
2174 | op0, const0_rtx), | |
2175 | gen_rtx_LABEL_REF (VOIDmode, label), | |
2176 | pc_rtx))); | |
2177 | } | |
2178 | \f | |
2179 | /* Return nonzero if a return peephole merging return with | |
2180 | setting of output register is ok. */ | |
2181 | int | |
2182 | leaf_return_peephole_ok () | |
2183 | { | |
2184 | return (actual_fsize == 0); | |
2185 | } | |
284d86e9 | 2186 | |
e0d80184 DM |
2187 | /* Return nonzero if TRIAL can go into the function epilogue's |
2188 | delay slot. SLOT is the slot we are trying to fill. */ | |
ab835497 | 2189 | |
e0d80184 DM |
2190 | int |
2191 | eligible_for_epilogue_delay (trial, slot) | |
2192 | rtx trial; | |
2193 | int slot; | |
2194 | { | |
2195 | rtx pat, src; | |
4fb4e4b8 | 2196 | |
e0d80184 DM |
2197 | if (slot >= 1) |
2198 | return 0; | |
4fb4e4b8 | 2199 | |
e0d80184 DM |
2200 | if (GET_CODE (trial) != INSN || GET_CODE (PATTERN (trial)) != SET) |
2201 | return 0; | |
4fb4e4b8 | 2202 | |
e0d80184 DM |
2203 | if (get_attr_length (trial) != 1) |
2204 | return 0; | |
4fb4e4b8 | 2205 | |
e0d80184 DM |
2206 | /* If %g0 is live, there are lots of things we can't handle. |
2207 | Rather than trying to find them all now, let's punt and only | |
2208 | optimize things as necessary. */ | |
2209 | if (TARGET_LIVE_G0) | |
2210 | return 0; | |
4fb4e4b8 | 2211 | |
e0d80184 DM |
2212 | /* In the case of a true leaf function, anything can go into the delay slot. |
2213 | A delay slot only exists however if the frame size is zero, otherwise | |
2214 | we will put an insn to adjust the stack after the return. */ | |
54ff41b7 | 2215 | if (current_function_uses_only_leaf_regs) |
4fb4e4b8 | 2216 | { |
e0d80184 DM |
2217 | if (leaf_return_peephole_ok ()) |
2218 | return ((get_attr_in_uncond_branch_delay (trial) | |
2219 | == IN_BRANCH_DELAY_TRUE)); | |
2220 | return 0; | |
4fb4e4b8 | 2221 | } |
4fb4e4b8 | 2222 | |
e0d80184 DM |
2223 | /* If only trivial `restore' insns work, nothing can go in the |
2224 | delay slot. */ | |
2225 | else if (TARGET_BROKEN_SAVERESTORE) | |
2226 | return 0; | |
795068a4 | 2227 | |
e0d80184 | 2228 | pat = PATTERN (trial); |
795068a4 | 2229 | |
e0d80184 DM |
2230 | /* Otherwise, only operations which can be done in tandem with |
2231 | a `restore' insn can go into the delay slot. */ | |
2232 | if (GET_CODE (SET_DEST (pat)) != REG | |
2233 | || REGNO (SET_DEST (pat)) >= 32 | |
2234 | || REGNO (SET_DEST (pat)) < 24) | |
2235 | return 0; | |
795068a4 | 2236 | |
e0d80184 DM |
2237 | /* The set of insns matched here must agree precisely with the set of |
2238 | patterns paired with a RETURN in sparc.md. */ | |
795068a4 | 2239 | |
e0d80184 | 2240 | src = SET_SRC (pat); |
795068a4 | 2241 | |
3c6088d2 | 2242 | /* This matches "*return_[qhs]i" or even "*return_di" on TARGET_ARCH64. */ |
e0d80184 | 2243 | if (arith_operand (src, GET_MODE (src))) |
3c6088d2 JJ |
2244 | { |
2245 | if (TARGET_ARCH64) | |
2246 | return GET_MODE_SIZE (GET_MODE (src)) <= GET_MODE_SIZE (DImode); | |
2247 | else | |
2248 | return GET_MODE_SIZE (GET_MODE (src)) <= GET_MODE_SIZE (SImode); | |
2249 | } | |
e0d80184 DM |
2250 | |
2251 | /* This matches "*return_di". */ | |
2252 | else if (arith_double_operand (src, GET_MODE (src))) | |
2253 | return GET_MODE_SIZE (GET_MODE (src)) <= GET_MODE_SIZE (DImode); | |
795068a4 | 2254 | |
e0d80184 DM |
2255 | /* This matches "*return_sf_no_fpu". */ |
2256 | else if (! TARGET_FPU && restore_operand (SET_DEST (pat), SFmode) | |
2257 | && register_operand (src, SFmode)) | |
2258 | return 1; | |
f5da225e | 2259 | |
e0d80184 DM |
2260 | /* This matches "*return_addsi". */ |
2261 | else if (GET_CODE (src) == PLUS | |
2262 | && arith_operand (XEXP (src, 0), SImode) | |
2263 | && arith_operand (XEXP (src, 1), SImode) | |
2264 | && (register_operand (XEXP (src, 0), SImode) | |
2265 | || register_operand (XEXP (src, 1), SImode))) | |
2266 | return 1; | |
2267 | ||
2268 | /* This matches "*return_adddi". */ | |
2269 | else if (GET_CODE (src) == PLUS | |
2270 | && arith_double_operand (XEXP (src, 0), DImode) | |
2271 | && arith_double_operand (XEXP (src, 1), DImode) | |
2272 | && (register_operand (XEXP (src, 0), DImode) | |
2273 | || register_operand (XEXP (src, 1), DImode))) | |
795068a4 | 2274 | return 1; |
795068a4 | 2275 | |
795068a4 JW |
2276 | return 0; |
2277 | } | |
2278 | ||
e0d80184 DM |
2279 | static int |
2280 | check_return_regs (x) | |
2281 | rtx x; | |
ab835497 | 2282 | { |
e0d80184 DM |
2283 | switch (GET_CODE (x)) |
2284 | { | |
2285 | case REG: | |
2286 | return IN_OR_GLOBAL_P (x); | |
ab835497 | 2287 | |
e0d80184 DM |
2288 | case CONST_INT: |
2289 | case CONST_DOUBLE: | |
2290 | case CONST: | |
2291 | case SYMBOL_REF: | |
2292 | case LABEL_REF: | |
2293 | return 1; | |
ab835497 | 2294 | |
e0d80184 DM |
2295 | case SET: |
2296 | case IOR: | |
2297 | case AND: | |
2298 | case XOR: | |
2299 | case PLUS: | |
2300 | case MINUS: | |
2301 | if (check_return_regs (XEXP (x, 1)) == 0) | |
2302 | return 0; | |
2303 | case NOT: | |
2304 | case NEG: | |
2305 | case MEM: | |
2306 | return check_return_regs (XEXP (x, 0)); | |
2307 | ||
2308 | default: | |
2309 | return 0; | |
2310 | } | |
ab835497 | 2311 | |
e0d80184 | 2312 | } |
ab835497 | 2313 | |
e0d80184 DM |
2314 | /* Return 1 if TRIAL references only in and global registers. */ |
2315 | int | |
2316 | eligible_for_return_delay (trial) | |
2317 | rtx trial; | |
2318 | { | |
2319 | if (GET_CODE (PATTERN (trial)) != SET) | |
2320 | return 0; | |
ab835497 | 2321 | |
e0d80184 DM |
2322 | return check_return_regs (PATTERN (trial)); |
2323 | } | |
ab835497 | 2324 | |
e0d80184 DM |
2325 | int |
2326 | short_branch (uid1, uid2) | |
2327 | int uid1, uid2; | |
2328 | { | |
2329 | unsigned int delta = insn_addresses[uid1] - insn_addresses[uid2]; | |
2330 | if (delta + 1024 < 2048) | |
2331 | return 1; | |
2332 | /* warning ("long branch, distance %d", delta); */ | |
2333 | return 0; | |
2334 | } | |
ab835497 | 2335 | |
e0d80184 DM |
2336 | /* Return non-zero if REG is not used after INSN. |
2337 | We assume REG is a reload reg, and therefore does | |
2338 | not live past labels or calls or jumps. */ | |
2339 | int | |
2340 | reg_unused_after (reg, insn) | |
2341 | rtx reg; | |
2342 | rtx insn; | |
2343 | { | |
2344 | enum rtx_code code, prev_code = UNKNOWN; | |
ab835497 | 2345 | |
e0d80184 | 2346 | while ((insn = NEXT_INSN (insn))) |
61a55e8b | 2347 | { |
e0d80184 DM |
2348 | if (prev_code == CALL_INSN && call_used_regs[REGNO (reg)]) |
2349 | return 1; | |
ab835497 | 2350 | |
e0d80184 DM |
2351 | code = GET_CODE (insn); |
2352 | if (GET_CODE (insn) == CODE_LABEL) | |
2353 | return 1; | |
bc961ed7 | 2354 | |
e0d80184 | 2355 | if (GET_RTX_CLASS (code) == 'i') |
61a55e8b | 2356 | { |
e0d80184 DM |
2357 | rtx set = single_set (insn); |
2358 | int in_src = set && reg_overlap_mentioned_p (reg, SET_SRC (set)); | |
2359 | if (set && in_src) | |
2360 | return 0; | |
2361 | if (set && reg_overlap_mentioned_p (reg, SET_DEST (set))) | |
2362 | return 1; | |
2363 | if (set == 0 && reg_overlap_mentioned_p (reg, PATTERN (insn))) | |
2364 | return 0; | |
61a55e8b | 2365 | } |
e0d80184 | 2366 | prev_code = code; |
61a55e8b | 2367 | } |
e0d80184 DM |
2368 | return 1; |
2369 | } | |
2370 | \f | |
2371 | /* The table we use to reference PIC data. */ | |
2372 | static rtx global_offset_table; | |
bc961ed7 | 2373 | |
e0d80184 DM |
2374 | /* The function we use to get at it. */ |
2375 | static rtx get_pc_symbol; | |
2376 | static char get_pc_symbol_name[256]; | |
ab835497 | 2377 | |
e0d80184 | 2378 | /* Ensure that we are not using patterns that are not OK with PIC. */ |
795068a4 | 2379 | |
e0d80184 DM |
2380 | int |
2381 | check_pic (i) | |
2382 | int i; | |
2383 | { | |
2384 | switch (flag_pic) | |
ab835497 | 2385 | { |
e0d80184 | 2386 | case 1: |
1ccbefce RH |
2387 | if (GET_CODE (recog_data.operand[i]) == SYMBOL_REF |
2388 | || (GET_CODE (recog_data.operand[i]) == CONST | |
2389 | && ! (GET_CODE (XEXP (recog_data.operand[i], 0)) == MINUS | |
2390 | && (XEXP (XEXP (recog_data.operand[i], 0), 0) | |
e0d80184 | 2391 | == global_offset_table) |
1ccbefce | 2392 | && (GET_CODE (XEXP (XEXP (recog_data.operand[i], 0), 1)) |
e0d80184 DM |
2393 | == CONST)))) |
2394 | abort (); | |
2395 | case 2: | |
2396 | default: | |
2397 | return 1; | |
ab835497 | 2398 | } |
ab835497 | 2399 | } |
795068a4 | 2400 | |
e0d80184 DM |
2401 | /* Return true if X is an address which needs a temporary register when |
2402 | reloaded while generating PIC code. */ | |
795068a4 | 2403 | |
e0d80184 DM |
2404 | int |
2405 | pic_address_needs_scratch (x) | |
2406 | rtx x; | |
795068a4 | 2407 | { |
e0d80184 DM |
2408 | /* An address which is a symbolic plus a non SMALL_INT needs a temp reg. */ |
2409 | if (GET_CODE (x) == CONST && GET_CODE (XEXP (x, 0)) == PLUS | |
2410 | && GET_CODE (XEXP (XEXP (x, 0), 0)) == SYMBOL_REF | |
2411 | && GET_CODE (XEXP (XEXP (x, 0), 1)) == CONST_INT | |
2412 | && ! SMALL_INT (XEXP (XEXP (x, 0), 1))) | |
2413 | return 1; | |
795068a4 | 2414 | |
e0d80184 DM |
2415 | return 0; |
2416 | } | |
795068a4 | 2417 | |
e0d80184 DM |
2418 | /* Legitimize PIC addresses. If the address is already position-independent, |
2419 | we return ORIG. Newly generated position-independent addresses go into a | |
2420 | reg. This is REG if non zero, otherwise we allocate register(s) as | |
2421 | necessary. */ | |
795068a4 | 2422 | |
e0d80184 DM |
2423 | rtx |
2424 | legitimize_pic_address (orig, mode, reg) | |
2425 | rtx orig; | |
2426 | enum machine_mode mode ATTRIBUTE_UNUSED; | |
2427 | rtx reg; | |
2428 | { | |
2429 | if (GET_CODE (orig) == SYMBOL_REF) | |
82d6b402 | 2430 | { |
e0d80184 DM |
2431 | rtx pic_ref, address; |
2432 | rtx insn; | |
82d6b402 | 2433 | |
e0d80184 | 2434 | if (reg == 0) |
82d6b402 | 2435 | { |
e0d80184 DM |
2436 | if (reload_in_progress || reload_completed) |
2437 | abort (); | |
2438 | else | |
2439 | reg = gen_reg_rtx (Pmode); | |
82d6b402 | 2440 | } |
795068a4 | 2441 | |
e0d80184 | 2442 | if (flag_pic == 2) |
4fb4e4b8 | 2443 | { |
e0d80184 DM |
2444 | /* If not during reload, allocate another temp reg here for loading |
2445 | in the address, so that these instructions can be optimized | |
2446 | properly. */ | |
2447 | rtx temp_reg = ((reload_in_progress || reload_completed) | |
2448 | ? reg : gen_reg_rtx (Pmode)); | |
2449 | ||
2450 | /* Must put the SYMBOL_REF inside an UNSPEC here so that cse | |
2451 | won't get confused into thinking that these two instructions | |
2452 | are loading in the true address of the symbol. If in the | |
2453 | future a PIC rtx exists, that should be used instead. */ | |
be3f1ff5 DM |
2454 | if (Pmode == SImode) |
2455 | { | |
2456 | emit_insn (gen_movsi_high_pic (temp_reg, orig)); | |
2457 | emit_insn (gen_movsi_lo_sum_pic (temp_reg, temp_reg, orig)); | |
2458 | } | |
2459 | else | |
2460 | { | |
2461 | emit_insn (gen_movdi_high_pic (temp_reg, orig)); | |
2462 | emit_insn (gen_movdi_lo_sum_pic (temp_reg, temp_reg, orig)); | |
2463 | } | |
e0d80184 | 2464 | address = temp_reg; |
4fb4e4b8 DE |
2465 | } |
2466 | else | |
e0d80184 DM |
2467 | address = orig; |
2468 | ||
2469 | pic_ref = gen_rtx_MEM (Pmode, | |
2470 | gen_rtx_PLUS (Pmode, | |
2471 | pic_offset_table_rtx, address)); | |
2472 | current_function_uses_pic_offset_table = 1; | |
2473 | RTX_UNCHANGING_P (pic_ref) = 1; | |
2474 | insn = emit_move_insn (reg, pic_ref); | |
2475 | /* Put a REG_EQUAL note on this insn, so that it can be optimized | |
2476 | by loop. */ | |
2477 | REG_NOTES (insn) = gen_rtx_EXPR_LIST (REG_EQUAL, orig, | |
2478 | REG_NOTES (insn)); | |
2479 | return reg; | |
795068a4 | 2480 | } |
e0d80184 | 2481 | else if (GET_CODE (orig) == CONST) |
795068a4 | 2482 | { |
e0d80184 | 2483 | rtx base, offset; |
795068a4 | 2484 | |
e0d80184 DM |
2485 | if (GET_CODE (XEXP (orig, 0)) == PLUS |
2486 | && XEXP (XEXP (orig, 0), 0) == pic_offset_table_rtx) | |
2487 | return orig; | |
82d6b402 | 2488 | |
e0d80184 | 2489 | if (reg == 0) |
4fb4e4b8 | 2490 | { |
e0d80184 DM |
2491 | if (reload_in_progress || reload_completed) |
2492 | abort (); | |
2493 | else | |
2494 | reg = gen_reg_rtx (Pmode); | |
4fb4e4b8 | 2495 | } |
e0d80184 DM |
2496 | |
2497 | if (GET_CODE (XEXP (orig, 0)) == PLUS) | |
4fb4e4b8 | 2498 | { |
e0d80184 DM |
2499 | base = legitimize_pic_address (XEXP (XEXP (orig, 0), 0), Pmode, reg); |
2500 | offset = legitimize_pic_address (XEXP (XEXP (orig, 0), 1), Pmode, | |
2501 | base == reg ? 0 : reg); | |
4fb4e4b8 DE |
2502 | } |
2503 | else | |
61a66555 JW |
2504 | abort (); |
2505 | ||
e0d80184 DM |
2506 | if (GET_CODE (offset) == CONST_INT) |
2507 | { | |
2508 | if (SMALL_INT (offset)) | |
2509 | return plus_constant_for_output (base, INTVAL (offset)); | |
2510 | else if (! reload_in_progress && ! reload_completed) | |
2511 | offset = force_reg (Pmode, offset); | |
2512 | else | |
2513 | /* If we reach here, then something is seriously wrong. */ | |
2514 | abort (); | |
2515 | } | |
2516 | return gen_rtx_PLUS (Pmode, base, offset); | |
795068a4 | 2517 | } |
e0d80184 DM |
2518 | else if (GET_CODE (orig) == LABEL_REF) |
2519 | /* ??? Why do we do this? */ | |
2520 | /* Now movsi_pic_label_ref uses it, but we ought to be checking that | |
2521 | the register is live instead, in case it is eliminated. */ | |
2522 | current_function_uses_pic_offset_table = 1; | |
795068a4 | 2523 | |
e0d80184 DM |
2524 | return orig; |
2525 | } | |
795068a4 | 2526 | |
e0d80184 DM |
2527 | /* Return the RTX for insns to set the PIC register. */ |
2528 | ||
2529 | static rtx | |
2530 | pic_setup_code () | |
2531 | { | |
2532 | rtx seq; | |
2533 | ||
2534 | start_sequence (); | |
2535 | emit_insn (gen_get_pc (pic_offset_table_rtx, global_offset_table, | |
2536 | get_pc_symbol)); | |
2537 | seq = gen_sequence (); | |
2538 | end_sequence (); | |
82d6b402 | 2539 | |
e0d80184 DM |
2540 | return seq; |
2541 | } | |
795068a4 | 2542 | |
e0d80184 | 2543 | /* Emit special PIC prologues and epilogues. */ |
795068a4 | 2544 | |
e0d80184 DM |
2545 | void |
2546 | finalize_pic () | |
2547 | { | |
2548 | /* Labels to get the PC in the prologue of this function. */ | |
2549 | int orig_flag_pic = flag_pic; | |
2550 | rtx insn; | |
4fb4e4b8 | 2551 | |
e0d80184 DM |
2552 | if (current_function_uses_pic_offset_table == 0) |
2553 | return; | |
795068a4 | 2554 | |
e0d80184 DM |
2555 | if (! flag_pic) |
2556 | abort (); | |
2557 | ||
2558 | /* If we havn't emitted the special get_pc helper function, do so now. */ | |
2559 | if (get_pc_symbol_name[0] == 0) | |
4fb4e4b8 | 2560 | { |
d676da68 | 2561 | int align; |
82d6b402 | 2562 | |
d676da68 | 2563 | ASM_GENERATE_INTERNAL_LABEL (get_pc_symbol_name, "LGETPC", 0); |
e0d80184 | 2564 | text_section (); |
d676da68 DM |
2565 | |
2566 | align = floor_log2 (FUNCTION_BOUNDARY / BITS_PER_UNIT); | |
2567 | if (align > 0) | |
2568 | ASM_OUTPUT_ALIGN (asm_out_file, align); | |
e0d80184 DM |
2569 | ASM_OUTPUT_INTERNAL_LABEL (asm_out_file, "LGETPC", 0); |
2570 | fputs ("\tretl\n\tadd %o7,%l7,%l7\n", asm_out_file); | |
2571 | } | |
82d6b402 | 2572 | |
e0d80184 DM |
2573 | /* Initialize every time through, since we can't easily |
2574 | know this to be permanent. */ | |
2575 | global_offset_table = gen_rtx_SYMBOL_REF (Pmode, "_GLOBAL_OFFSET_TABLE_"); | |
2576 | get_pc_symbol = gen_rtx_SYMBOL_REF (Pmode, get_pc_symbol_name); | |
2577 | flag_pic = 0; | |
82d6b402 | 2578 | |
e0d80184 | 2579 | emit_insn_after (pic_setup_code (), get_insns ()); |
82d6b402 | 2580 | |
24a43d7b DM |
2581 | /* Insert the code in each nonlocal goto receiver. |
2582 | If you make changes here or to the nonlocal_goto_receiver | |
2583 | pattern, make sure the unspec_volatile numbers still | |
2584 | match. */ | |
e0d80184 DM |
2585 | for (insn = get_insns (); insn; insn = NEXT_INSN (insn)) |
2586 | if (GET_CODE (insn) == INSN && GET_CODE (PATTERN (insn)) == UNSPEC_VOLATILE | |
24a43d7b | 2587 | && XINT (PATTERN (insn), 1) == 5) |
e0d80184 | 2588 | emit_insn_after (pic_setup_code (), insn); |
82d6b402 | 2589 | |
e0d80184 | 2590 | flag_pic = orig_flag_pic; |
82d6b402 | 2591 | |
e0d80184 DM |
2592 | /* Need to emit this whether or not we obey regdecls, |
2593 | since setjmp/longjmp can cause life info to screw up. | |
2594 | ??? In the case where we don't obey regdecls, this is not sufficient | |
2595 | since we may not fall out the bottom. */ | |
2596 | emit_insn (gen_rtx_USE (VOIDmode, pic_offset_table_rtx)); | |
795068a4 | 2597 | } |
ab835497 | 2598 | \f |
e0d80184 DM |
2599 | /* Return 1 if RTX is a MEM which is known to be aligned to at |
2600 | least an 8 byte boundary. */ | |
a3ee5899 | 2601 | |
e0d80184 DM |
2602 | int |
2603 | mem_min_alignment (mem, desired) | |
2604 | rtx mem; | |
2605 | int desired; | |
ab835497 | 2606 | { |
e0d80184 DM |
2607 | rtx addr, base, offset; |
2608 | ||
2609 | /* If it's not a MEM we can't accept it. */ | |
2610 | if (GET_CODE (mem) != MEM) | |
2611 | return 0; | |
2612 | ||
2613 | addr = XEXP (mem, 0); | |
2614 | base = offset = NULL_RTX; | |
2615 | if (GET_CODE (addr) == PLUS) | |
ab835497 | 2616 | { |
e0d80184 | 2617 | if (GET_CODE (XEXP (addr, 0)) == REG) |
61a55e8b | 2618 | { |
e0d80184 DM |
2619 | base = XEXP (addr, 0); |
2620 | ||
2621 | /* What we are saying here is that if the base | |
2622 | REG is aligned properly, the compiler will make | |
2623 | sure any REG based index upon it will be so | |
2624 | as well. */ | |
2625 | if (GET_CODE (XEXP (addr, 1)) == CONST_INT) | |
2626 | offset = XEXP (addr, 1); | |
61a55e8b | 2627 | else |
e0d80184 | 2628 | offset = const0_rtx; |
61a55e8b | 2629 | } |
ab835497 | 2630 | } |
e0d80184 | 2631 | else if (GET_CODE (addr) == REG) |
ab835497 | 2632 | { |
e0d80184 DM |
2633 | base = addr; |
2634 | offset = const0_rtx; | |
ab835497 | 2635 | } |
82d6b402 | 2636 | |
e0d80184 | 2637 | if (base != NULL_RTX) |
795068a4 | 2638 | { |
e0d80184 DM |
2639 | int regno = REGNO (base); |
2640 | ||
2641 | if (regno != FRAME_POINTER_REGNUM | |
2642 | && regno != STACK_POINTER_REGNUM) | |
61a55e8b | 2643 | { |
e0d80184 DM |
2644 | /* Check if the compiler has recorded some information |
2645 | about the alignment of the base REG. If reload has | |
2646 | completed, we already matched with proper alignments. */ | |
49ad7cfa | 2647 | if (((current_function != 0 |
e0d80184 DM |
2648 | && REGNO_POINTER_ALIGN (regno) >= desired) |
2649 | || reload_completed) | |
2650 | && ((INTVAL (offset) & (desired - 1)) == 0)) | |
2651 | return 1; | |
61a55e8b | 2652 | } |
795068a4 | 2653 | else |
e0d80184 DM |
2654 | { |
2655 | if (((INTVAL (offset) - SPARC_STACK_BIAS) & (desired - 1)) == 0) | |
2656 | return 1; | |
2657 | } | |
795068a4 | 2658 | } |
e0d80184 DM |
2659 | else if (! TARGET_UNALIGNED_DOUBLES |
2660 | || CONSTANT_P (addr) | |
2661 | || GET_CODE (addr) == LO_SUM) | |
795068a4 | 2662 | { |
e0d80184 DM |
2663 | /* Anything else we know is properly aligned unless TARGET_UNALIGNED_DOUBLES |
2664 | is true, in which case we can only assume that an access is aligned if | |
2665 | it is to a constant address, or the address involves a LO_SUM. */ | |
2666 | return 1; | |
795068a4 | 2667 | } |
e0d80184 DM |
2668 | |
2669 | /* An obviously unaligned address. */ | |
2670 | return 0; | |
795068a4 | 2671 | } |
e0d80184 | 2672 | |
ab835497 | 2673 | \f |
61a55e8b DE |
2674 | /* Vectors to keep interesting information about registers where it can easily |
2675 | be got. We use to use the actual mode value as the bit number, but there | |
2676 | are more than 32 modes now. Instead we use two tables: one indexed by | |
2677 | hard register number, and one indexed by mode. */ | |
2678 | ||
2679 | /* The purpose of sparc_mode_class is to shrink the range of modes so that | |
2680 | they all fit (as bit numbers) in a 32 bit word (again). Each real mode is | |
2681 | mapped into one sparc_mode_class mode. */ | |
2682 | ||
2683 | enum sparc_mode_class { | |
61a55e8b | 2684 | S_MODE, D_MODE, T_MODE, O_MODE, |
c4ce6853 DE |
2685 | SF_MODE, DF_MODE, TF_MODE, OF_MODE, |
2686 | CC_MODE, CCFP_MODE | |
61a55e8b | 2687 | }; |
ab835497 | 2688 | |
61a55e8b DE |
2689 | /* Modes for single-word and smaller quantities. */ |
2690 | #define S_MODES ((1 << (int) S_MODE) | (1 << (int) SF_MODE)) | |
2691 | ||
2692 | /* Modes for double-word and smaller quantities. */ | |
2693 | #define D_MODES (S_MODES | (1 << (int) D_MODE) | (1 << DF_MODE)) | |
2694 | ||
2695 | /* Modes for quad-word and smaller quantities. */ | |
2696 | #define T_MODES (D_MODES | (1 << (int) T_MODE) | (1 << (int) TF_MODE)) | |
ab835497 | 2697 | |
1e4eb8d1 JW |
2698 | /* Modes for single-float quantities. We must allow any single word or |
2699 | smaller quantity. This is because the fix/float conversion instructions | |
2700 | take integer inputs/outputs from the float registers. */ | |
2701 | #define SF_MODES (S_MODES) | |
ab835497 | 2702 | |
61a55e8b DE |
2703 | /* Modes for double-float and smaller quantities. */ |
2704 | #define DF_MODES (S_MODES | D_MODES) | |
2705 | ||
284d86e9 | 2706 | #define DF_MODES64 DF_MODES |
61a55e8b DE |
2707 | |
2708 | /* Modes for double-float only quantities. */ | |
284d86e9 | 2709 | #define DF_ONLY_MODES ((1 << (int) DF_MODE) | (1 << (int) D_MODE)) |
61a55e8b DE |
2710 | |
2711 | /* Modes for double-float and larger quantities. */ | |
2712 | #define DF_UP_MODES (DF_ONLY_MODES | TF_ONLY_MODES) | |
ab835497 | 2713 | |
61a55e8b DE |
2714 | /* Modes for quad-float only quantities. */ |
2715 | #define TF_ONLY_MODES (1 << (int) TF_MODE) | |
2716 | ||
2717 | /* Modes for quad-float and smaller quantities. */ | |
2718 | #define TF_MODES (DF_MODES | TF_ONLY_MODES) | |
2719 | ||
61a55e8b | 2720 | #define TF_MODES64 (DF_MODES64 | TF_ONLY_MODES) |
ab835497 | 2721 | |
c4ce6853 DE |
2722 | /* Modes for condition codes. */ |
2723 | #define CC_MODES (1 << (int) CC_MODE) | |
2724 | #define CCFP_MODES (1 << (int) CCFP_MODE) | |
2725 | ||
ab835497 RK |
2726 | /* Value is 1 if register/mode pair is acceptable on sparc. |
2727 | The funny mixture of D and T modes is because integer operations | |
2728 | do not specially operate on tetra quantities, so non-quad-aligned | |
2729 | registers can hold quadword quantities (except %o4 and %i4 because | |
c4ce6853 | 2730 | they cross fixed registers). */ |
ab835497 | 2731 | |
61a55e8b DE |
2732 | /* This points to either the 32 bit or the 64 bit version. */ |
2733 | int *hard_regno_mode_classes; | |
2734 | ||
2735 | static int hard_32bit_mode_classes[] = { | |
c4ce6853 | 2736 | S_MODES, S_MODES, T_MODES, S_MODES, T_MODES, S_MODES, D_MODES, S_MODES, |
ab835497 RK |
2737 | T_MODES, S_MODES, T_MODES, S_MODES, D_MODES, S_MODES, D_MODES, S_MODES, |
2738 | T_MODES, S_MODES, T_MODES, S_MODES, T_MODES, S_MODES, D_MODES, S_MODES, | |
2739 | T_MODES, S_MODES, T_MODES, S_MODES, D_MODES, S_MODES, D_MODES, S_MODES, | |
2740 | ||
2741 | TF_MODES, SF_MODES, DF_MODES, SF_MODES, TF_MODES, SF_MODES, DF_MODES, SF_MODES, | |
2742 | TF_MODES, SF_MODES, DF_MODES, SF_MODES, TF_MODES, SF_MODES, DF_MODES, SF_MODES, | |
2743 | TF_MODES, SF_MODES, DF_MODES, SF_MODES, TF_MODES, SF_MODES, DF_MODES, SF_MODES, | |
61a55e8b | 2744 | TF_MODES, SF_MODES, DF_MODES, SF_MODES, TF_MODES, SF_MODES, DF_MODES, SF_MODES, |
a9e27770 DE |
2745 | |
2746 | /* FP regs f32 to f63. Only the even numbered registers actually exist, | |
2747 | and none can hold SFmode/SImode values. */ | |
2748 | DF_UP_MODES, 0, DF_ONLY_MODES, 0, DF_UP_MODES, 0, DF_ONLY_MODES, 0, | |
2749 | DF_UP_MODES, 0, DF_ONLY_MODES, 0, DF_UP_MODES, 0, DF_ONLY_MODES, 0, | |
2750 | DF_UP_MODES, 0, DF_ONLY_MODES, 0, DF_UP_MODES, 0, DF_ONLY_MODES, 0, | |
2751 | DF_UP_MODES, 0, DF_ONLY_MODES, 0, DF_UP_MODES, 0, DF_ONLY_MODES, 0, | |
2752 | ||
2753 | /* %fcc[0123] */ | |
c4ce6853 DE |
2754 | CCFP_MODES, CCFP_MODES, CCFP_MODES, CCFP_MODES, |
2755 | ||
2756 | /* %icc */ | |
2757 | CC_MODES | |
61a55e8b DE |
2758 | }; |
2759 | ||
2760 | static int hard_64bit_mode_classes[] = { | |
c4ce6853 | 2761 | D_MODES, D_MODES, T_MODES, D_MODES, T_MODES, D_MODES, T_MODES, D_MODES, |
61a55e8b DE |
2762 | T_MODES, D_MODES, T_MODES, D_MODES, T_MODES, D_MODES, T_MODES, D_MODES, |
2763 | T_MODES, D_MODES, T_MODES, D_MODES, T_MODES, D_MODES, T_MODES, D_MODES, | |
2764 | T_MODES, D_MODES, T_MODES, D_MODES, T_MODES, D_MODES, T_MODES, D_MODES, | |
2765 | ||
2766 | TF_MODES64, SF_MODES, DF_MODES64, SF_MODES, TF_MODES64, SF_MODES, DF_MODES64, SF_MODES, | |
2767 | TF_MODES64, SF_MODES, DF_MODES64, SF_MODES, TF_MODES64, SF_MODES, DF_MODES64, SF_MODES, | |
2768 | TF_MODES64, SF_MODES, DF_MODES64, SF_MODES, TF_MODES64, SF_MODES, DF_MODES64, SF_MODES, | |
2769 | TF_MODES64, SF_MODES, DF_MODES64, SF_MODES, TF_MODES64, SF_MODES, DF_MODES64, SF_MODES, | |
2770 | ||
a9e27770 | 2771 | /* FP regs f32 to f63. Only the even numbered registers actually exist, |
61a55e8b DE |
2772 | and none can hold SFmode/SImode values. */ |
2773 | DF_UP_MODES, 0, DF_ONLY_MODES, 0, DF_UP_MODES, 0, DF_ONLY_MODES, 0, | |
2774 | DF_UP_MODES, 0, DF_ONLY_MODES, 0, DF_UP_MODES, 0, DF_ONLY_MODES, 0, | |
2775 | DF_UP_MODES, 0, DF_ONLY_MODES, 0, DF_UP_MODES, 0, DF_ONLY_MODES, 0, | |
2776 | DF_UP_MODES, 0, DF_ONLY_MODES, 0, DF_UP_MODES, 0, DF_ONLY_MODES, 0, | |
2777 | ||
2778 | /* %fcc[0123] */ | |
c4ce6853 DE |
2779 | CCFP_MODES, CCFP_MODES, CCFP_MODES, CCFP_MODES, |
2780 | ||
2781 | /* %icc */ | |
2782 | CC_MODES | |
61a55e8b DE |
2783 | }; |
2784 | ||
2785 | int sparc_mode_class [NUM_MACHINE_MODES]; | |
2786 | ||
c4ce6853 DE |
2787 | enum reg_class sparc_regno_reg_class[FIRST_PSEUDO_REGISTER]; |
2788 | ||
61a55e8b DE |
2789 | static void |
2790 | sparc_init_modes () | |
2791 | { | |
2792 | int i; | |
2793 | ||
61a55e8b DE |
2794 | for (i = 0; i < NUM_MACHINE_MODES; i++) |
2795 | { | |
2796 | switch (GET_MODE_CLASS (i)) | |
2797 | { | |
2798 | case MODE_INT: | |
2799 | case MODE_PARTIAL_INT: | |
2800 | case MODE_COMPLEX_INT: | |
2801 | if (GET_MODE_SIZE (i) <= 4) | |
2802 | sparc_mode_class[i] = 1 << (int) S_MODE; | |
2803 | else if (GET_MODE_SIZE (i) == 8) | |
2804 | sparc_mode_class[i] = 1 << (int) D_MODE; | |
2805 | else if (GET_MODE_SIZE (i) == 16) | |
2806 | sparc_mode_class[i] = 1 << (int) T_MODE; | |
2807 | else if (GET_MODE_SIZE (i) == 32) | |
2808 | sparc_mode_class[i] = 1 << (int) O_MODE; | |
2809 | else | |
2810 | sparc_mode_class[i] = 0; | |
2811 | break; | |
2812 | case MODE_FLOAT: | |
2813 | case MODE_COMPLEX_FLOAT: | |
2814 | if (GET_MODE_SIZE (i) <= 4) | |
2815 | sparc_mode_class[i] = 1 << (int) SF_MODE; | |
2816 | else if (GET_MODE_SIZE (i) == 8) | |
2817 | sparc_mode_class[i] = 1 << (int) DF_MODE; | |
2818 | else if (GET_MODE_SIZE (i) == 16) | |
2819 | sparc_mode_class[i] = 1 << (int) TF_MODE; | |
2820 | else if (GET_MODE_SIZE (i) == 32) | |
2821 | sparc_mode_class[i] = 1 << (int) OF_MODE; | |
2822 | else | |
2823 | sparc_mode_class[i] = 0; | |
2824 | break; | |
2825 | case MODE_CC: | |
2826 | default: | |
2827 | /* mode_class hasn't been initialized yet for EXTRA_CC_MODES, so | |
2828 | we must explicitly check for them here. */ | |
2829 | if (i == (int) CCFPmode || i == (int) CCFPEmode) | |
2830 | sparc_mode_class[i] = 1 << (int) CCFP_MODE; | |
2831 | else if (i == (int) CCmode || i == (int) CC_NOOVmode | |
a9e27770 | 2832 | || i == (int) CCXmode || i == (int) CCX_NOOVmode) |
c4ce6853 | 2833 | sparc_mode_class[i] = 1 << (int) CC_MODE; |
61a55e8b DE |
2834 | else |
2835 | sparc_mode_class[i] = 0; | |
2836 | break; | |
2837 | } | |
2838 | } | |
2839 | ||
a9e27770 | 2840 | if (TARGET_ARCH64) |
61a55e8b DE |
2841 | hard_regno_mode_classes = hard_64bit_mode_classes; |
2842 | else | |
2843 | hard_regno_mode_classes = hard_32bit_mode_classes; | |
c4ce6853 DE |
2844 | |
2845 | /* Initialize the array used by REGNO_REG_CLASS. */ | |
2846 | for (i = 0; i < FIRST_PSEUDO_REGISTER; i++) | |
2847 | { | |
284d86e9 JC |
2848 | if (i < 16 && TARGET_V8PLUS) |
2849 | sparc_regno_reg_class[i] = I64_REGS; | |
2850 | else if (i < 32) | |
c4ce6853 DE |
2851 | sparc_regno_reg_class[i] = GENERAL_REGS; |
2852 | else if (i < 64) | |
2853 | sparc_regno_reg_class[i] = FP_REGS; | |
2854 | else if (i < 96) | |
2855 | sparc_regno_reg_class[i] = EXTRA_FP_REGS; | |
2856 | else if (i < 100) | |
2857 | sparc_regno_reg_class[i] = FPCC_REGS; | |
2858 | else | |
2859 | sparc_regno_reg_class[i] = NO_REGS; | |
2860 | } | |
61a55e8b | 2861 | } |
ab835497 | 2862 | \f |
61a55e8b DE |
2863 | /* Save non call used registers from LOW to HIGH at BASE+OFFSET. |
2864 | N_REGS is the number of 4-byte regs saved thus far. This applies even to | |
2865 | v9 int regs as it simplifies the code. */ | |
2866 | ||
ab835497 | 2867 | static int |
c53aa195 | 2868 | save_regs (file, low, high, base, offset, n_regs, real_offset) |
ab835497 RK |
2869 | FILE *file; |
2870 | int low, high; | |
3bb5de61 | 2871 | const char *base; |
ab835497 | 2872 | int offset; |
61a55e8b | 2873 | int n_regs; |
c53aa195 | 2874 | int real_offset; |
ab835497 RK |
2875 | { |
2876 | int i; | |
2877 | ||
a9e27770 | 2878 | if (TARGET_ARCH64 && high <= 32) |
ab835497 | 2879 | { |
61a55e8b DE |
2880 | for (i = low; i < high; i++) |
2881 | { | |
2882 | if (regs_ever_live[i] && ! call_used_regs[i]) | |
c53aa195 | 2883 | { |
e0d80184 | 2884 | fprintf (file, "\tstx\t%s, [%s+%d]\n", |
c53aa195 | 2885 | reg_names[i], base, offset + 4 * n_regs); |
0021b564 | 2886 | if (dwarf2out_do_frame ()) |
c53aa195 | 2887 | dwarf2out_reg_save ("", i, real_offset + 4 * n_regs); |
c53aa195 JM |
2888 | n_regs += 2; |
2889 | } | |
61a55e8b | 2890 | } |
ab835497 | 2891 | } |
61a55e8b DE |
2892 | else |
2893 | { | |
2894 | for (i = low; i < high; i += 2) | |
2895 | { | |
2896 | if (regs_ever_live[i] && ! call_used_regs[i]) | |
c53aa195 | 2897 | { |
5e9defae KG |
2898 | if (regs_ever_live[i+1] && ! call_used_regs[i+1]) |
2899 | { | |
e0d80184 | 2900 | fprintf (file, "\tstd\t%s, [%s+%d]\n", |
5e9defae KG |
2901 | reg_names[i], base, offset + 4 * n_regs); |
2902 | if (dwarf2out_do_frame ()) | |
2903 | { | |
2904 | char *l = dwarf2out_cfi_label (); | |
2905 | dwarf2out_reg_save (l, i, real_offset + 4 * n_regs); | |
2906 | dwarf2out_reg_save (l, i+1, real_offset + 4 * n_regs + 4); | |
2907 | } | |
2908 | n_regs += 2; | |
2909 | } | |
2910 | else | |
2911 | { | |
e0d80184 | 2912 | fprintf (file, "\tst\t%s, [%s+%d]\n", |
5e9defae KG |
2913 | reg_names[i], base, offset + 4 * n_regs); |
2914 | if (dwarf2out_do_frame ()) | |
2915 | dwarf2out_reg_save ("", i, real_offset + 4 * n_regs); | |
2916 | n_regs += 2; | |
2917 | } | |
2918 | } | |
2919 | else | |
2920 | { | |
2921 | if (regs_ever_live[i+1] && ! call_used_regs[i+1]) | |
2922 | { | |
e0d80184 | 2923 | fprintf (file, "\tst\t%s, [%s+%d]\n", |
5e9defae KG |
2924 | reg_names[i+1], base, offset + 4 * n_regs + 4); |
2925 | if (dwarf2out_do_frame ()) | |
2926 | dwarf2out_reg_save ("", i + 1, real_offset + 4 * n_regs + 4); | |
2927 | n_regs += 2; | |
2928 | } | |
c53aa195 | 2929 | } |
61a55e8b DE |
2930 | } |
2931 | } | |
2932 | return n_regs; | |
ab835497 RK |
2933 | } |
2934 | ||
61a55e8b DE |
2935 | /* Restore non call used registers from LOW to HIGH at BASE+OFFSET. |
2936 | ||
2937 | N_REGS is the number of 4-byte regs saved thus far. This applies even to | |
2938 | v9 int regs as it simplifies the code. */ | |
2939 | ||
ab835497 | 2940 | static int |
61a55e8b | 2941 | restore_regs (file, low, high, base, offset, n_regs) |
ab835497 RK |
2942 | FILE *file; |
2943 | int low, high; | |
3bb5de61 | 2944 | const char *base; |
ab835497 | 2945 | int offset; |
61a55e8b | 2946 | int n_regs; |
ab835497 RK |
2947 | { |
2948 | int i; | |
2949 | ||
a9e27770 | 2950 | if (TARGET_ARCH64 && high <= 32) |
61a55e8b DE |
2951 | { |
2952 | for (i = low; i < high; i++) | |
2953 | { | |
2954 | if (regs_ever_live[i] && ! call_used_regs[i]) | |
e0d80184 | 2955 | fprintf (file, "\tldx\t[%s+%d], %s\n", |
61a55e8b DE |
2956 | base, offset + 4 * n_regs, reg_names[i]), |
2957 | n_regs += 2; | |
2958 | } | |
2959 | } | |
2960 | else | |
ab835497 | 2961 | { |
61a55e8b DE |
2962 | for (i = low; i < high; i += 2) |
2963 | { | |
2964 | if (regs_ever_live[i] && ! call_used_regs[i]) | |
2965 | if (regs_ever_live[i+1] && ! call_used_regs[i+1]) | |
e0d80184 | 2966 | fprintf (file, "\tldd\t[%s+%d], %s\n", |
61a55e8b DE |
2967 | base, offset + 4 * n_regs, reg_names[i]), |
2968 | n_regs += 2; | |
2969 | else | |
e0d80184 | 2970 | fprintf (file, "\tld\t[%s+%d],%s\n", |
61a55e8b DE |
2971 | base, offset + 4 * n_regs, reg_names[i]), |
2972 | n_regs += 2; | |
2973 | else if (regs_ever_live[i+1] && ! call_used_regs[i+1]) | |
e0d80184 | 2974 | fprintf (file, "\tld\t[%s+%d],%s\n", |
61a55e8b DE |
2975 | base, offset + 4 * n_regs + 4, reg_names[i+1]), |
2976 | n_regs += 2; | |
2977 | } | |
ab835497 | 2978 | } |
61a55e8b | 2979 | return n_regs; |
ab835497 RK |
2980 | } |
2981 | ||
2982 | /* Static variables we want to share between prologue and epilogue. */ | |
2983 | ||
61a55e8b DE |
2984 | /* Number of live general or floating point registers needed to be saved |
2985 | (as 4-byte quantities). This is only done if TARGET_EPILOGUE. */ | |
2986 | static int num_gfregs; | |
2987 | ||
2988 | /* Compute the frame size required by the function. This function is called | |
2989 | during the reload pass and also by output_function_prologue(). */ | |
ab835497 | 2990 | |
ab835497 RK |
2991 | int |
2992 | compute_frame_size (size, leaf_function) | |
2993 | int size; | |
2994 | int leaf_function; | |
2995 | { | |
61a55e8b | 2996 | int n_regs = 0, i; |
ab835497 | 2997 | int outgoing_args_size = (current_function_outgoing_args_size |
4fb4e4b8 | 2998 | + REG_PARM_STACK_SPACE (current_function_decl)); |
ab835497 | 2999 | |
61a55e8b | 3000 | if (TARGET_EPILOGUE) |
ab835497 | 3001 | { |
61a55e8b DE |
3002 | /* N_REGS is the number of 4-byte regs saved thus far. This applies |
3003 | even to v9 int regs to be consistent with save_regs/restore_regs. */ | |
3004 | ||
a9e27770 | 3005 | if (TARGET_ARCH64) |
61a55e8b DE |
3006 | { |
3007 | for (i = 0; i < 8; i++) | |
3008 | if (regs_ever_live[i] && ! call_used_regs[i]) | |
3009 | n_regs += 2; | |
3010 | } | |
3011 | else | |
3012 | { | |
3013 | for (i = 0; i < 8; i += 2) | |
3014 | if ((regs_ever_live[i] && ! call_used_regs[i]) | |
3015 | || (regs_ever_live[i+1] && ! call_used_regs[i+1])) | |
3016 | n_regs += 2; | |
3017 | } | |
3018 | ||
3019 | for (i = 32; i < (TARGET_V9 ? 96 : 64); i += 2) | |
ab835497 RK |
3020 | if ((regs_ever_live[i] && ! call_used_regs[i]) |
3021 | || (regs_ever_live[i+1] && ! call_used_regs[i+1])) | |
61a55e8b | 3022 | n_regs += 2; |
ab835497 RK |
3023 | } |
3024 | ||
3025 | /* Set up values for use in `function_epilogue'. */ | |
61a55e8b | 3026 | num_gfregs = n_regs; |
ab835497 | 3027 | |
61a55e8b DE |
3028 | if (leaf_function && n_regs == 0 |
3029 | && size == 0 && current_function_outgoing_args_size == 0) | |
3030 | { | |
3031 | actual_fsize = apparent_fsize = 0; | |
3032 | } | |
3033 | else | |
3034 | { | |
3035 | /* We subtract STARTING_FRAME_OFFSET, remember it's negative. | |
3036 | The stack bias (if any) is taken out to undo its effects. */ | |
3037 | apparent_fsize = (size - STARTING_FRAME_OFFSET + SPARC_STACK_BIAS + 7) & -8; | |
3038 | apparent_fsize += n_regs * 4; | |
3039 | actual_fsize = apparent_fsize + ((outgoing_args_size + 7) & -8); | |
3040 | } | |
ab835497 RK |
3041 | |
3042 | /* Make sure nothing can clobber our register windows. | |
3043 | If a SAVE must be done, or there is a stack-local variable, | |
61a55e8b | 3044 | the register window area must be allocated. |
4fb4e4b8 | 3045 | ??? For v8 we apparently need an additional 8 bytes of reserved space. */ |
ab835497 | 3046 | if (leaf_function == 0 || size > 0) |
4fb4e4b8 | 3047 | actual_fsize += (16 * UNITS_PER_WORD) + (TARGET_ARCH64 ? 0 : 8); |
ab835497 | 3048 | |
61a55e8b DE |
3049 | return SPARC_STACK_ALIGN (actual_fsize); |
3050 | } | |
3051 | ||
3052 | /* Build a (32 bit) big number in a register. */ | |
86735b02 | 3053 | /* ??? We may be able to use the set macro here too. */ |
61a55e8b DE |
3054 | |
3055 | static void | |
3056 | build_big_number (file, num, reg) | |
3057 | FILE *file; | |
3058 | int num; | |
3bb5de61 | 3059 | const char *reg; |
61a55e8b | 3060 | { |
a9e27770 | 3061 | if (num >= 0 || ! TARGET_ARCH64) |
61a55e8b | 3062 | { |
e0d80184 | 3063 | fprintf (file, "\tsethi\t%%hi(%d), %s\n", num, reg); |
61a55e8b | 3064 | if ((num & 0x3ff) != 0) |
e0d80184 | 3065 | fprintf (file, "\tor\t%s, %%lo(%d), %s\n", reg, num, reg); |
61a55e8b | 3066 | } |
a9e27770 | 3067 | else /* num < 0 && TARGET_ARCH64 */ |
61a55e8b DE |
3068 | { |
3069 | /* Sethi does not sign extend, so we must use a little trickery | |
3070 | to use it for negative numbers. Invert the constant before | |
3071 | loading it in, then use xor immediate to invert the loaded bits | |
3072 | (along with the upper 32 bits) to the desired constant. This | |
3073 | works because the sethi and immediate fields overlap. */ | |
3074 | int asize = num; | |
3075 | int inv = ~asize; | |
3076 | int low = -0x400 + (asize & 0x3FF); | |
3077 | ||
e0d80184 | 3078 | fprintf (file, "\tsethi\t%%hi(%d), %s\n\txor\t%s, %d, %s\n", |
61a55e8b DE |
3079 | inv, reg, reg, low, reg); |
3080 | } | |
ab835497 RK |
3081 | } |
3082 | ||
1cb36a98 RH |
3083 | /* Output any necessary .register pseudo-ops. */ |
3084 | void | |
3085 | sparc_output_scratch_registers (file) | |
3086 | FILE *file; | |
3087 | { | |
3088 | #ifdef HAVE_AS_REGISTER_PSEUDO_OP | |
3089 | int i; | |
3090 | ||
3091 | if (TARGET_ARCH32) | |
3092 | return; | |
3093 | ||
3094 | /* Check if %g[2367] were used without | |
3095 | .register being printed for them already. */ | |
3096 | for (i = 2; i < 8; i++) | |
3097 | { | |
3098 | if (regs_ever_live [i] | |
3099 | && ! sparc_hard_reg_printed [i]) | |
3100 | { | |
3101 | sparc_hard_reg_printed [i] = 1; | |
3102 | fprintf (file, "\t.register\t%%g%d, #scratch\n", i); | |
3103 | } | |
3104 | if (i == 3) i = 5; | |
3105 | } | |
3106 | #endif | |
3107 | } | |
3108 | ||
915f619f JW |
3109 | /* Output code for the function prologue. */ |
3110 | ||
ab835497 RK |
3111 | void |
3112 | output_function_prologue (file, size, leaf_function) | |
3113 | FILE *file; | |
3114 | int size; | |
915f619f | 3115 | int leaf_function; |
ab835497 | 3116 | { |
1cb36a98 RH |
3117 | sparc_output_scratch_registers (file); |
3118 | ||
915f619f JW |
3119 | /* Need to use actual_fsize, since we are also allocating |
3120 | space for our callee (and our own register save area). */ | |
ab835497 RK |
3121 | actual_fsize = compute_frame_size (size, leaf_function); |
3122 | ||
61a55e8b DE |
3123 | if (leaf_function) |
3124 | { | |
3125 | frame_base_name = "%sp"; | |
3126 | frame_base_offset = actual_fsize + SPARC_STACK_BIAS; | |
3127 | } | |
3128 | else | |
3129 | { | |
3130 | frame_base_name = "%fp"; | |
3131 | frame_base_offset = SPARC_STACK_BIAS; | |
3132 | } | |
3133 | ||
c6aa9ce1 | 3134 | /* This is only for the human reader. */ |
bf62bbf1 | 3135 | fprintf (file, "\t%s#PROLOGUE# 0\n", ASM_COMMENT_START); |
c6aa9ce1 | 3136 | |
915f619f JW |
3137 | if (actual_fsize == 0) |
3138 | /* do nothing. */ ; | |
3592ea0d | 3139 | else if (! leaf_function && ! TARGET_BROKEN_SAVERESTORE) |
ab835497 | 3140 | { |
3592ea0d | 3141 | if (actual_fsize <= 4096) |
e0d80184 | 3142 | fprintf (file, "\tsave\t%%sp, -%d, %%sp\n", actual_fsize); |
3592ea0d | 3143 | else if (actual_fsize <= 8192) |
915f619f | 3144 | { |
e0d80184 DM |
3145 | fprintf (file, "\tsave\t%%sp, -4096, %%sp\n"); |
3146 | fprintf (file, "\tadd\t%%sp, -%d, %%sp\n", actual_fsize - 4096); | |
915f619f JW |
3147 | } |
3148 | else | |
3149 | { | |
3592ea0d | 3150 | build_big_number (file, -actual_fsize, "%g1"); |
e0d80184 | 3151 | fprintf (file, "\tsave\t%%sp, %%g1, %%sp\n"); |
915f619f | 3152 | } |
ab835497 | 3153 | } |
5f4241d5 | 3154 | else if (! leaf_function && TARGET_BROKEN_SAVERESTORE) |
ab835497 | 3155 | { |
5f4241d5 | 3156 | /* We assume the environment will properly handle or otherwise avoid |
956d6950 JL |
3157 | trouble associated with an interrupt occurring after the `save' or |
3158 | trap occurring during it. */ | |
5f4241d5 | 3159 | fprintf (file, "\tsave\n"); |
3592ea0d | 3160 | |
5f4241d5 | 3161 | if (actual_fsize <= 4096) |
e0d80184 | 3162 | fprintf (file, "\tadd\t%%fp, -%d, %%sp\n", actual_fsize); |
5f4241d5 DE |
3163 | else if (actual_fsize <= 8192) |
3164 | { | |
e0d80184 DM |
3165 | fprintf (file, "\tadd\t%%fp, -4096, %%sp\n"); |
3166 | fprintf (file, "\tadd\t%%fp, -%d, %%sp\n", actual_fsize - 4096); | |
5f4241d5 DE |
3167 | } |
3168 | else | |
3169 | { | |
3170 | build_big_number (file, -actual_fsize, "%g1"); | |
e0d80184 | 3171 | fprintf (file, "\tadd\t%%fp, %%g1, %%sp\n"); |
5f4241d5 DE |
3172 | } |
3173 | } | |
3174 | else /* leaf function */ | |
3175 | { | |
3592ea0d | 3176 | if (actual_fsize <= 4096) |
e0d80184 | 3177 | fprintf (file, "\tadd\t%%sp, -%d, %%sp\n", actual_fsize); |
3592ea0d DE |
3178 | else if (actual_fsize <= 8192) |
3179 | { | |
e0d80184 DM |
3180 | fprintf (file, "\tadd\t%%sp, -4096, %%sp\n"); |
3181 | fprintf (file, "\tadd\t%%sp, -%d, %%sp\n", actual_fsize - 4096); | |
3592ea0d | 3182 | } |
915f619f | 3183 | else |
3592ea0d DE |
3184 | { |
3185 | build_big_number (file, -actual_fsize, "%g1"); | |
e0d80184 | 3186 | fprintf (file, "\tadd\t%%sp, %%g1, %%sp\n"); |
3592ea0d | 3187 | } |
ab835497 RK |
3188 | } |
3189 | ||
0021b564 | 3190 | if (dwarf2out_do_frame () && actual_fsize) |
c53aa195 | 3191 | { |
96987e96 | 3192 | char *label = dwarf2out_cfi_label (); |
c53aa195 JM |
3193 | |
3194 | /* The canonical frame address refers to the top of the frame. */ | |
3195 | dwarf2out_def_cfa (label, (leaf_function ? STACK_POINTER_REGNUM | |
3196 | : FRAME_POINTER_REGNUM), | |
3197 | frame_base_offset); | |
3198 | ||
3199 | if (! leaf_function) | |
3200 | { | |
3201 | /* Note the register window save. This tells the unwinder that | |
3202 | it needs to restore the window registers from the previous | |
3203 | frame's window save area at 0(cfa). */ | |
3204 | dwarf2out_window_save (label); | |
3205 | ||
3206 | /* The return address (-8) is now in %i7. */ | |
3207 | dwarf2out_return_reg (label, 31); | |
3208 | } | |
3209 | } | |
c53aa195 | 3210 | |
ab835497 RK |
3211 | /* If doing anything with PIC, do it now. */ |
3212 | if (! flag_pic) | |
bf62bbf1 | 3213 | fprintf (file, "\t%s#PROLOGUE# 1\n", ASM_COMMENT_START); |
ab835497 | 3214 | |
61a55e8b DE |
3215 | /* Call saved registers are saved just above the outgoing argument area. */ |
3216 | if (num_gfregs) | |
ab835497 | 3217 | { |
c53aa195 | 3218 | int offset, real_offset, n_regs; |
3bb5de61 | 3219 | const char *base; |
ab835497 | 3220 | |
84e884e9 JM |
3221 | real_offset = -apparent_fsize; |
3222 | offset = -apparent_fsize + frame_base_offset; | |
61a55e8b DE |
3223 | if (offset < -4096 || offset + num_gfregs * 4 > 4096) |
3224 | { | |
3225 | /* ??? This might be optimized a little as %g1 might already have a | |
3226 | value close enough that a single add insn will do. */ | |
3227 | /* ??? Although, all of this is probably only a temporary fix | |
3228 | because if %g1 can hold a function result, then | |
3229 | output_function_epilogue will lose (the result will get | |
3230 | clobbered). */ | |
3231 | build_big_number (file, offset, "%g1"); | |
e0d80184 | 3232 | fprintf (file, "\tadd\t%s, %%g1, %%g1\n", frame_base_name); |
61a55e8b DE |
3233 | base = "%g1"; |
3234 | offset = 0; | |
3235 | } | |
ab835497 | 3236 | else |
61a55e8b DE |
3237 | { |
3238 | base = frame_base_name; | |
3239 | } | |
ab835497 | 3240 | |
db3d4438 | 3241 | n_regs = 0; |
ab835497 | 3242 | if (TARGET_EPILOGUE && ! leaf_function) |
61a55e8b | 3243 | /* ??? Originally saved regs 0-15 here. */ |
c53aa195 | 3244 | n_regs = save_regs (file, 0, 8, base, offset, 0, real_offset); |
ab835497 | 3245 | else if (leaf_function) |
61a55e8b | 3246 | /* ??? Originally saved regs 0-31 here. */ |
c53aa195 | 3247 | n_regs = save_regs (file, 0, 8, base, offset, 0, real_offset); |
ab835497 | 3248 | if (TARGET_EPILOGUE) |
c53aa195 JM |
3249 | save_regs (file, 32, TARGET_V9 ? 96 : 64, base, offset, n_regs, |
3250 | real_offset); | |
ab835497 RK |
3251 | } |
3252 | ||
ab835497 RK |
3253 | leaf_label = 0; |
3254 | if (leaf_function && actual_fsize != 0) | |
3255 | { | |
3256 | /* warning ("leaf procedure with frame size %d", actual_fsize); */ | |
3257 | if (! TARGET_EPILOGUE) | |
3258 | leaf_label = gen_label_rtx (); | |
3259 | } | |
3260 | } | |
3261 | ||
915f619f JW |
3262 | /* Output code for the function epilogue. */ |
3263 | ||
ab835497 | 3264 | void |
ef8200df | 3265 | output_function_epilogue (file, size, leaf_function) |
ab835497 | 3266 | FILE *file; |
db3d4438 | 3267 | int size ATTRIBUTE_UNUSED; |
ef8200df | 3268 | int leaf_function; |
ab835497 | 3269 | { |
3bb5de61 | 3270 | const char *ret; |
ab835497 RK |
3271 | |
3272 | if (leaf_label) | |
3273 | { | |
ab835497 RK |
3274 | emit_label_after (leaf_label, get_last_insn ()); |
3275 | final_scan_insn (get_last_insn (), file, 0, 0, 1); | |
3276 | } | |
3277 | ||
b69649e4 RK |
3278 | #ifdef FUNCTION_BLOCK_PROFILER_EXIT |
3279 | else if (profile_block_flag == 2) | |
3280 | { | |
3281 | FUNCTION_BLOCK_PROFILER_EXIT(file); | |
3282 | } | |
3283 | #endif | |
3284 | ||
bfd6bc60 JC |
3285 | else if (current_function_epilogue_delay_list == 0) |
3286 | { | |
e0d80184 DM |
3287 | /* If code does not drop into the epilogue, we need |
3288 | do nothing except output pending case vectors. */ | |
bfd6bc60 JC |
3289 | rtx insn = get_last_insn (); |
3290 | if (GET_CODE (insn) == NOTE) | |
3291 | insn = prev_nonnote_insn (insn); | |
3292 | if (insn && GET_CODE (insn) == BARRIER) | |
e0d80184 | 3293 | goto output_vectors; |
bfd6bc60 JC |
3294 | } |
3295 | ||
61a55e8b DE |
3296 | /* Restore any call saved registers. */ |
3297 | if (num_gfregs) | |
ab835497 | 3298 | { |
61a55e8b | 3299 | int offset, n_regs; |
3bb5de61 | 3300 | const char *base; |
ab835497 | 3301 | |
61a55e8b DE |
3302 | offset = -apparent_fsize + frame_base_offset; |
3303 | if (offset < -4096 || offset + num_gfregs * 4 > 4096 - 8 /*double*/) | |
3304 | { | |
3305 | build_big_number (file, offset, "%g1"); | |
e0d80184 | 3306 | fprintf (file, "\tadd\t%s, %%g1, %%g1\n", frame_base_name); |
61a55e8b DE |
3307 | base = "%g1"; |
3308 | offset = 0; | |
3309 | } | |
ab835497 | 3310 | else |
61a55e8b DE |
3311 | { |
3312 | base = frame_base_name; | |
3313 | } | |
ab835497 | 3314 | |
db3d4438 | 3315 | n_regs = 0; |
ab835497 | 3316 | if (TARGET_EPILOGUE && ! leaf_function) |
61a55e8b DE |
3317 | /* ??? Originally saved regs 0-15 here. */ |
3318 | n_regs = restore_regs (file, 0, 8, base, offset, 0); | |
ab835497 | 3319 | else if (leaf_function) |
61a55e8b DE |
3320 | /* ??? Originally saved regs 0-31 here. */ |
3321 | n_regs = restore_regs (file, 0, 8, base, offset, 0); | |
ab835497 | 3322 | if (TARGET_EPILOGUE) |
61a55e8b | 3323 | restore_regs (file, 32, TARGET_V9 ? 96 : 64, base, offset, n_regs); |
ab835497 RK |
3324 | } |
3325 | ||
3326 | /* Work out how to skip the caller's unimp instruction if required. */ | |
3327 | if (leaf_function) | |
e0d80184 | 3328 | ret = (SKIP_CALLERS_UNIMP_P ? "jmp\t%o7+12" : "retl"); |
ab835497 | 3329 | else |
e0d80184 | 3330 | ret = (SKIP_CALLERS_UNIMP_P ? "jmp\t%i7+12" : "ret"); |
ab835497 | 3331 | |
ef8200df | 3332 | if (TARGET_EPILOGUE || leaf_label) |
ab835497 | 3333 | { |
ef8200df JW |
3334 | int old_target_epilogue = TARGET_EPILOGUE; |
3335 | target_flags &= ~old_target_epilogue; | |
ab835497 | 3336 | |
ef8200df JW |
3337 | if (! leaf_function) |
3338 | { | |
3339 | /* If we wound up with things in our delay slot, flush them here. */ | |
3340 | if (current_function_epilogue_delay_list) | |
ab835497 | 3341 | { |
254110c2 | 3342 | rtx insn = emit_jump_insn_after (gen_rtx_RETURN (VOIDmode), |
ef8200df | 3343 | get_last_insn ()); |
254110c2 | 3344 | PATTERN (insn) = gen_rtx_PARALLEL (VOIDmode, |
ef8200df JW |
3345 | gen_rtvec (2, |
3346 | PATTERN (XEXP (current_function_epilogue_delay_list, 0)), | |
3347 | PATTERN (insn))); | |
3348 | final_scan_insn (insn, file, 1, 0, 1); | |
ab835497 | 3349 | } |
284d86e9 | 3350 | else if (TARGET_V9 && ! SKIP_CALLERS_UNIMP_P) |
e0d80184 | 3351 | fputs ("\treturn\t%i7+8\n\tnop\n", file); |
ab835497 | 3352 | else |
ef8200df JW |
3353 | fprintf (file, "\t%s\n\trestore\n", ret); |
3354 | } | |
915f619f JW |
3355 | /* All of the following cases are for leaf functions. */ |
3356 | else if (current_function_epilogue_delay_list) | |
ef8200df | 3357 | { |
915f619f JW |
3358 | /* eligible_for_epilogue_delay_slot ensures that if this is a |
3359 | leaf function, then we will only have insn in the delay slot | |
3360 | if the frame size is zero, thus no adjust for the stack is | |
3361 | needed here. */ | |
3362 | if (actual_fsize != 0) | |
ef8200df | 3363 | abort (); |
915f619f JW |
3364 | fprintf (file, "\t%s\n", ret); |
3365 | final_scan_insn (XEXP (current_function_epilogue_delay_list, 0), | |
3366 | file, 1, 0, 1); | |
ab835497 | 3367 | } |
9bb7ffda JW |
3368 | /* Output 'nop' instead of 'sub %sp,-0,%sp' when no frame, so as to |
3369 | avoid generating confusing assembly language output. */ | |
3370 | else if (actual_fsize == 0) | |
3371 | fprintf (file, "\t%s\n\tnop\n", ret); | |
915f619f | 3372 | else if (actual_fsize <= 4096) |
e0d80184 | 3373 | fprintf (file, "\t%s\n\tsub\t%%sp, -%d, %%sp\n", ret, actual_fsize); |
915f619f | 3374 | else if (actual_fsize <= 8192) |
e0d80184 | 3375 | fprintf (file, "\tsub\t%%sp, -4096, %%sp\n\t%s\n\tsub\t%%sp, -%d, %%sp\n", |
915f619f JW |
3376 | ret, actual_fsize - 4096); |
3377 | else if ((actual_fsize & 0x3ff) == 0) | |
e0d80184 | 3378 | fprintf (file, "\tsethi\t%%hi(%d), %%g1\n\t%s\n\tadd\t%%sp, %%g1, %%sp\n", |
915f619f JW |
3379 | actual_fsize, ret); |
3380 | else | |
e0d80184 | 3381 | fprintf (file, "\tsethi\t%%hi(%d), %%g1\n\tor\t%%g1, %%lo(%d), %%g1\n\t%s\n\tadd\t%%sp, %%g1, %%sp\n", |
915f619f | 3382 | actual_fsize, actual_fsize, ret); |
ef8200df | 3383 | target_flags |= old_target_epilogue; |
ab835497 | 3384 | } |
e0d80184 DM |
3385 | |
3386 | output_vectors: | |
3387 | sparc_output_deferred_case_vectors (); | |
ab835497 | 3388 | } |
4fb4e4b8 DE |
3389 | \f |
3390 | /* Functions for handling argument passing. | |
3391 | ||
3392 | For v8 the first six args are normally in registers and the rest are | |
3393 | pushed. Any arg that starts within the first 6 words is at least | |
3394 | partially passed in a register unless its data type forbids. | |
3395 | ||
3396 | For v9, the argument registers are laid out as an array of 16 elements | |
3397 | and arguments are added sequentially. The first 6 int args and up to the | |
3398 | first 16 fp args (depending on size) are passed in regs. | |
3399 | ||
3400 | Slot Stack Integral Float Float in structure Double Long Double | |
3401 | ---- ----- -------- ----- ------------------ ------ ----------- | |
3402 | 15 [SP+248] %f31 %f30,%f31 %d30 | |
3403 | 14 [SP+240] %f29 %f28,%f29 %d28 %q28 | |
3404 | 13 [SP+232] %f27 %f26,%f27 %d26 | |
3405 | 12 [SP+224] %f25 %f24,%f25 %d24 %q24 | |
3406 | 11 [SP+216] %f23 %f22,%f23 %d22 | |
3407 | 10 [SP+208] %f21 %f20,%f21 %d20 %q20 | |
3408 | 9 [SP+200] %f19 %f18,%f19 %d18 | |
3409 | 8 [SP+192] %f17 %f16,%f17 %d16 %q16 | |
3410 | 7 [SP+184] %f15 %f14,%f15 %d14 | |
3411 | 6 [SP+176] %f13 %f12,%f13 %d12 %q12 | |
3412 | 5 [SP+168] %o5 %f11 %f10,%f11 %d10 | |
3413 | 4 [SP+160] %o4 %f9 %f8,%f9 %d8 %q8 | |
3414 | 3 [SP+152] %o3 %f7 %f6,%f7 %d6 | |
3415 | 2 [SP+144] %o2 %f5 %f4,%f5 %d4 %q4 | |
3416 | 1 [SP+136] %o1 %f3 %f2,%f3 %d2 | |
3417 | 0 [SP+128] %o0 %f1 %f0,%f1 %d0 %q0 | |
3418 | ||
3419 | Here SP = %sp if -mno-stack-bias or %sp+stack_bias otherwise. | |
3420 | ||
3421 | Integral arguments are always passed as 64 bit quantities appropriately | |
3422 | extended. | |
3423 | ||
3424 | Passing of floating point values is handled as follows. | |
3425 | If a prototype is in scope: | |
3426 | If the value is in a named argument (i.e. not a stdarg function or a | |
3427 | value not part of the `...') then the value is passed in the appropriate | |
3428 | fp reg. | |
3429 | If the value is part of the `...' and is passed in one of the first 6 | |
3430 | slots then the value is passed in the appropriate int reg. | |
3431 | If the value is part of the `...' and is not passed in one of the first 6 | |
3432 | slots then the value is passed in memory. | |
3433 | If a prototype is not in scope: | |
3434 | If the value is one of the first 6 arguments the value is passed in the | |
3435 | appropriate integer reg and the appropriate fp reg. | |
3436 | If the value is not one of the first 6 arguments the value is passed in | |
3437 | the appropriate fp reg and in memory. | |
3438 | */ | |
3439 | ||
3440 | /* Maximum number of int regs for args. */ | |
3441 | #define SPARC_INT_ARG_MAX 6 | |
3442 | /* Maximum number of fp regs for args. */ | |
3443 | #define SPARC_FP_ARG_MAX 16 | |
3444 | ||
3445 | #define ROUND_ADVANCE(SIZE) (((SIZE) + UNITS_PER_WORD - 1) / UNITS_PER_WORD) | |
3446 | ||
3447 | /* Handle the INIT_CUMULATIVE_ARGS macro. | |
3448 | Initialize a variable CUM of type CUMULATIVE_ARGS | |
3449 | for a call to a function whose data type is FNTYPE. | |
3450 | For a library call, FNTYPE is 0. */ | |
3ea1fdd3 | 3451 | |
4fb4e4b8 DE |
3452 | void |
3453 | init_cumulative_args (cum, fntype, libname, indirect) | |
3454 | CUMULATIVE_ARGS *cum; | |
db3d4438 DM |
3455 | tree fntype; |
3456 | tree libname ATTRIBUTE_UNUSED; | |
3457 | int indirect ATTRIBUTE_UNUSED; | |
4fb4e4b8 DE |
3458 | { |
3459 | cum->words = 0; | |
3460 | cum->prototype_p = fntype && TYPE_ARG_TYPES (fntype); | |
3461 | cum->libcall_p = fntype == 0; | |
3462 | } | |
61a55e8b | 3463 | |
4fb4e4b8 DE |
3464 | /* Compute the slot number to pass an argument in. |
3465 | Returns the slot number or -1 if passing on the stack. | |
3466 | ||
3467 | CUM is a variable of type CUMULATIVE_ARGS which gives info about | |
3468 | the preceding args and about the function being called. | |
3469 | MODE is the argument's machine mode. | |
3470 | TYPE is the data type of the argument (as a tree). | |
3471 | This is null for libcalls where that information may | |
3472 | not be available. | |
3473 | NAMED is nonzero if this argument is a named parameter | |
3474 | (otherwise it is an extra parameter matching an ellipsis). | |
3475 | INCOMING_P is zero for FUNCTION_ARG, nonzero for FUNCTION_INCOMING_ARG. | |
3476 | *PREGNO records the register number to use if scalar type. | |
3477 | *PPADDING records the amount of padding needed in words. */ | |
3478 | ||
3479 | static int | |
3480 | function_arg_slotno (cum, mode, type, named, incoming_p, pregno, ppadding) | |
3481 | const CUMULATIVE_ARGS *cum; | |
3482 | enum machine_mode mode; | |
3483 | tree type; | |
3484 | int named; | |
3485 | int incoming_p; | |
3486 | int *pregno; | |
3487 | int *ppadding; | |
3488 | { | |
3489 | int regbase = (incoming_p | |
3490 | ? SPARC_INCOMING_INT_ARG_FIRST | |
3491 | : SPARC_OUTGOING_INT_ARG_FIRST); | |
3492 | int slotno = cum->words; | |
3493 | int regno; | |
3494 | ||
3495 | *ppadding = 0; | |
3496 | ||
3497 | if (type != 0 && TREE_ADDRESSABLE (type)) | |
3498 | return -1; | |
3499 | if (TARGET_ARCH32 | |
3500 | && type != 0 && mode == BLKmode | |
3501 | && TYPE_ALIGN (type) % PARM_BOUNDARY != 0) | |
3502 | return -1; | |
3503 | ||
3504 | switch (mode) | |
3505 | { | |
3506 | case VOIDmode : | |
3507 | /* MODE is VOIDmode when generating the actual call. | |
3508 | See emit_call_1. */ | |
3509 | return -1; | |
3510 | ||
3511 | case QImode : case CQImode : | |
3512 | case HImode : case CHImode : | |
3513 | case SImode : case CSImode : | |
3514 | case DImode : case CDImode : | |
3515 | if (slotno >= SPARC_INT_ARG_MAX) | |
3516 | return -1; | |
3517 | regno = regbase + slotno; | |
3518 | break; | |
3519 | ||
3520 | case SFmode : case SCmode : | |
3521 | case DFmode : case DCmode : | |
3522 | case TFmode : case TCmode : | |
3523 | if (TARGET_ARCH32) | |
3524 | { | |
3525 | if (slotno >= SPARC_INT_ARG_MAX) | |
3526 | return -1; | |
3527 | regno = regbase + slotno; | |
3528 | } | |
3529 | else | |
3530 | { | |
3531 | if ((mode == TFmode || mode == TCmode) | |
3532 | && (slotno & 1) != 0) | |
3533 | slotno++, *ppadding = 1; | |
3534 | if (TARGET_FPU && named) | |
3535 | { | |
3536 | if (slotno >= SPARC_FP_ARG_MAX) | |
82d6b402 | 3537 | return -1; |
4fb4e4b8 DE |
3538 | regno = SPARC_FP_ARG_FIRST + slotno * 2; |
3539 | if (mode == SFmode) | |
3540 | regno++; | |
3541 | } | |
3542 | else | |
3543 | { | |
3544 | if (slotno >= SPARC_INT_ARG_MAX) | |
3545 | return -1; | |
3546 | regno = regbase + slotno; | |
3547 | } | |
3548 | } | |
3549 | break; | |
3550 | ||
3551 | case BLKmode : | |
3552 | /* For sparc64, objects requiring 16 byte alignment get it. */ | |
3553 | if (TARGET_ARCH64) | |
3554 | { | |
3555 | if (type && TYPE_ALIGN (type) == 128 && (slotno & 1) != 0) | |
3556 | slotno++, *ppadding = 1; | |
3557 | } | |
3558 | ||
3559 | if (TARGET_ARCH32 | |
5e9defae | 3560 | || (type && TREE_CODE (type) == UNION_TYPE)) |
4fb4e4b8 DE |
3561 | { |
3562 | if (slotno >= SPARC_INT_ARG_MAX) | |
3563 | return -1; | |
3564 | regno = regbase + slotno; | |
3565 | } | |
3566 | else | |
3567 | { | |
3568 | tree field; | |
3569 | int intregs_p = 0, fpregs_p = 0; | |
3570 | /* The ABI obviously doesn't specify how packed | |
3571 | structures are passed. These are defined to be passed | |
3572 | in int regs if possible, otherwise memory. */ | |
3573 | int packed_p = 0; | |
3574 | ||
3575 | /* First see what kinds of registers we need. */ | |
3576 | for (field = TYPE_FIELDS (type); field; field = TREE_CHAIN (field)) | |
3577 | { | |
3578 | if (TREE_CODE (field) == FIELD_DECL) | |
3579 | { | |
3580 | if (TREE_CODE (TREE_TYPE (field)) == REAL_TYPE | |
3581 | && TARGET_FPU) | |
3582 | fpregs_p = 1; | |
3583 | else | |
3584 | intregs_p = 1; | |
3585 | if (DECL_PACKED (field)) | |
3586 | packed_p = 1; | |
3587 | } | |
3588 | } | |
3589 | if (packed_p || !named) | |
3590 | fpregs_p = 0, intregs_p = 1; | |
3591 | ||
3592 | /* If all arg slots are filled, then must pass on stack. */ | |
3593 | if (fpregs_p && slotno >= SPARC_FP_ARG_MAX) | |
3594 | return -1; | |
3595 | /* If there are only int args and all int arg slots are filled, | |
3596 | then must pass on stack. */ | |
3597 | if (!fpregs_p && intregs_p && slotno >= SPARC_INT_ARG_MAX) | |
3598 | return -1; | |
3599 | /* Note that even if all int arg slots are filled, fp members may | |
3600 | still be passed in regs if such regs are available. | |
3601 | *PREGNO isn't set because there may be more than one, it's up | |
3602 | to the caller to compute them. */ | |
3603 | return slotno; | |
3604 | } | |
3605 | break; | |
3606 | ||
3607 | default : | |
3608 | abort (); | |
3609 | } | |
3610 | ||
3611 | *pregno = regno; | |
3612 | return slotno; | |
3613 | } | |
3614 | ||
82d6b402 RH |
3615 | /* Handle recursive register counting for structure field layout. */ |
3616 | ||
3617 | struct function_arg_record_value_parms | |
3618 | { | |
3619 | rtx ret; | |
3620 | int slotno, named, regbase; | |
3621 | int nregs, intoffset; | |
3622 | }; | |
3623 | ||
2a01c939 DM |
3624 | static void function_arg_record_value_3 |
3625 | PROTO((int, struct function_arg_record_value_parms *)); | |
3626 | static void function_arg_record_value_2 | |
3627 | PROTO((tree, int, struct function_arg_record_value_parms *)); | |
b1474bb7 KG |
3628 | static void function_arg_record_value_1 |
3629 | PROTO((tree, int, struct function_arg_record_value_parms *)); | |
2a01c939 DM |
3630 | static rtx function_arg_record_value |
3631 | PROTO((tree, enum machine_mode, int, int, int)); | |
3632 | ||
82d6b402 RH |
3633 | static void |
3634 | function_arg_record_value_1 (type, startbitpos, parms) | |
3635 | tree type; | |
3636 | int startbitpos; | |
3637 | struct function_arg_record_value_parms *parms; | |
3638 | { | |
3639 | tree field; | |
3640 | ||
3641 | /* The ABI obviously doesn't specify how packed structures are | |
3642 | passed. These are defined to be passed in int regs if possible, | |
3643 | otherwise memory. */ | |
3644 | int packed_p = 0; | |
3645 | ||
3646 | /* We need to compute how many registers are needed so we can | |
3647 | allocate the PARALLEL but before we can do that we need to know | |
3648 | whether there are any packed fields. If there are, int regs are | |
3649 | used regardless of whether there are fp values present. */ | |
3650 | for (field = TYPE_FIELDS (type); field; field = TREE_CHAIN (field)) | |
3651 | { | |
3652 | if (TREE_CODE (field) == FIELD_DECL && DECL_PACKED (field)) | |
3653 | { | |
3654 | packed_p = 1; | |
3655 | break; | |
3656 | } | |
3657 | } | |
3658 | ||
3659 | /* Compute how many registers we need. */ | |
3660 | for (field = TYPE_FIELDS (type); field; field = TREE_CHAIN (field)) | |
3661 | { | |
3662 | if (TREE_CODE (field) == FIELD_DECL) | |
3663 | { | |
3664 | int bitpos = startbitpos; | |
3665 | if (DECL_FIELD_BITPOS (field)) | |
3666 | bitpos += TREE_INT_CST_LOW (DECL_FIELD_BITPOS (field)); | |
3667 | /* ??? FIXME: else assume zero offset. */ | |
3668 | ||
3669 | if (TREE_CODE (TREE_TYPE (field)) == RECORD_TYPE) | |
3670 | { | |
3671 | function_arg_record_value_1 (TREE_TYPE (field), bitpos, parms); | |
3672 | } | |
3673 | else if (TREE_CODE (TREE_TYPE (field)) == REAL_TYPE | |
3674 | && TARGET_FPU | |
3675 | && ! packed_p | |
3676 | && parms->named) | |
3677 | { | |
3678 | if (parms->intoffset != -1) | |
3679 | { | |
3680 | int intslots, this_slotno; | |
3681 | ||
3682 | intslots = (bitpos - parms->intoffset + BITS_PER_WORD - 1) | |
3683 | / BITS_PER_WORD; | |
3684 | this_slotno = parms->slotno + parms->intoffset | |
3685 | / BITS_PER_WORD; | |
3686 | ||
3687 | intslots = MIN (intslots, SPARC_INT_ARG_MAX - this_slotno); | |
3688 | intslots = MAX (intslots, 0); | |
3689 | parms->nregs += intslots; | |
3690 | parms->intoffset = -1; | |
3691 | } | |
3692 | ||
3693 | /* There's no need to check this_slotno < SPARC_FP_ARG MAX. | |
3694 | If it wasn't true we wouldn't be here. */ | |
3695 | parms->nregs += 1; | |
3696 | } | |
3697 | else | |
3698 | { | |
3699 | if (parms->intoffset == -1) | |
3700 | parms->intoffset = bitpos; | |
3701 | } | |
3702 | } | |
3703 | } | |
3704 | } | |
3705 | ||
3706 | /* Handle recursive structure field register assignment. */ | |
3707 | ||
3708 | static void | |
3709 | function_arg_record_value_3 (bitpos, parms) | |
3710 | int bitpos; | |
3711 | struct function_arg_record_value_parms *parms; | |
3712 | { | |
3713 | enum machine_mode mode; | |
3714 | int regno, this_slotno, intslots, intoffset; | |
3715 | rtx reg; | |
3716 | ||
3717 | if (parms->intoffset == -1) | |
3718 | return; | |
3719 | intoffset = parms->intoffset; | |
3720 | parms->intoffset = -1; | |
3721 | ||
3722 | intslots = (bitpos - intoffset + BITS_PER_WORD - 1) / BITS_PER_WORD; | |
3723 | this_slotno = parms->slotno + intoffset / BITS_PER_WORD; | |
3724 | ||
3725 | intslots = MIN (intslots, SPARC_INT_ARG_MAX - this_slotno); | |
3726 | if (intslots <= 0) | |
3727 | return; | |
3728 | ||
3729 | /* If this is the trailing part of a word, only load that much into | |
3730 | the register. Otherwise load the whole register. Note that in | |
3731 | the latter case we may pick up unwanted bits. It's not a problem | |
3732 | at the moment but may wish to revisit. */ | |
3733 | ||
3734 | if (intoffset % BITS_PER_WORD != 0) | |
3735 | { | |
3736 | mode = mode_for_size (BITS_PER_WORD - intoffset%BITS_PER_WORD, | |
3737 | MODE_INT, 0); | |
3738 | } | |
3739 | else | |
3740 | mode = word_mode; | |
3741 | ||
3742 | intoffset /= BITS_PER_UNIT; | |
3743 | do | |
3744 | { | |
3745 | regno = parms->regbase + this_slotno; | |
254110c2 | 3746 | reg = gen_rtx_REG (mode, regno); |
82d6b402 | 3747 | XVECEXP (parms->ret, 0, parms->nregs) |
254110c2 | 3748 | = gen_rtx_EXPR_LIST (VOIDmode, reg, GEN_INT (intoffset)); |
82d6b402 RH |
3749 | |
3750 | this_slotno += 1; | |
3751 | intoffset = (intoffset | (UNITS_PER_WORD-1)) + 1; | |
3752 | parms->nregs += 1; | |
3753 | intslots -= 1; | |
3754 | } | |
3755 | while (intslots > 0); | |
3756 | } | |
3757 | ||
3758 | static void | |
3759 | function_arg_record_value_2 (type, startbitpos, parms) | |
3760 | tree type; | |
3761 | int startbitpos; | |
3762 | struct function_arg_record_value_parms *parms; | |
3763 | { | |
3764 | tree field; | |
3765 | int packed_p = 0; | |
3766 | ||
3767 | for (field = TYPE_FIELDS (type); field; field = TREE_CHAIN (field)) | |
3768 | { | |
3769 | if (TREE_CODE (field) == FIELD_DECL && DECL_PACKED (field)) | |
3770 | { | |
3771 | packed_p = 1; | |
3772 | break; | |
3773 | } | |
3774 | } | |
3775 | ||
3776 | for (field = TYPE_FIELDS (type); field; field = TREE_CHAIN (field)) | |
3777 | { | |
3778 | if (TREE_CODE (field) == FIELD_DECL) | |
3779 | { | |
3780 | int bitpos = startbitpos; | |
3781 | if (DECL_FIELD_BITPOS (field)) | |
3782 | bitpos += TREE_INT_CST_LOW (DECL_FIELD_BITPOS (field)); | |
3783 | /* ??? FIXME: else assume zero offset. */ | |
3784 | ||
3785 | if (TREE_CODE (TREE_TYPE (field)) == RECORD_TYPE) | |
3786 | { | |
3787 | function_arg_record_value_2 (TREE_TYPE (field), bitpos, parms); | |
3788 | } | |
3789 | else if (TREE_CODE (TREE_TYPE (field)) == REAL_TYPE | |
3790 | && TARGET_FPU | |
3791 | && ! packed_p | |
3792 | && parms->named) | |
3793 | { | |
3794 | int this_slotno = parms->slotno + bitpos / BITS_PER_WORD; | |
3795 | rtx reg; | |
3796 | ||
3797 | function_arg_record_value_3 (bitpos, parms); | |
3798 | ||
254110c2 | 3799 | reg = gen_rtx_REG (DECL_MODE (field), |
82d6b402 RH |
3800 | (SPARC_FP_ARG_FIRST + this_slotno * 2 |
3801 | + (DECL_MODE (field) == SFmode | |
3802 | && (bitpos & 32) != 0))); | |
3803 | XVECEXP (parms->ret, 0, parms->nregs) | |
254110c2 | 3804 | = gen_rtx_EXPR_LIST (VOIDmode, reg, |
82d6b402 RH |
3805 | GEN_INT (bitpos / BITS_PER_UNIT)); |
3806 | parms->nregs += 1; | |
3807 | } | |
3808 | else | |
3809 | { | |
3810 | if (parms->intoffset == -1) | |
3811 | parms->intoffset = bitpos; | |
3812 | } | |
3813 | } | |
3814 | } | |
3815 | } | |
3816 | ||
3817 | static rtx | |
1eac9f59 | 3818 | function_arg_record_value (type, mode, slotno, named, regbase) |
82d6b402 | 3819 | tree type; |
1eac9f59 | 3820 | enum machine_mode mode; |
82d6b402 RH |
3821 | int slotno, named, regbase; |
3822 | { | |
3823 | HOST_WIDE_INT typesize = int_size_in_bytes (type); | |
3824 | struct function_arg_record_value_parms parms; | |
3825 | int nregs; | |
3826 | ||
3827 | parms.ret = NULL_RTX; | |
3828 | parms.slotno = slotno; | |
3829 | parms.named = named; | |
3830 | parms.regbase = regbase; | |
3831 | ||
3832 | /* Compute how many registers we need. */ | |
3833 | parms.nregs = 0; | |
3834 | parms.intoffset = 0; | |
3835 | function_arg_record_value_1 (type, 0, &parms); | |
3836 | ||
3837 | if (parms.intoffset != -1) | |
3838 | { | |
3839 | int intslots, this_slotno; | |
3840 | ||
3841 | intslots = (typesize*BITS_PER_UNIT - parms.intoffset + BITS_PER_WORD - 1) | |
3842 | / BITS_PER_WORD; | |
3843 | this_slotno = slotno + parms.intoffset / BITS_PER_WORD; | |
3844 | ||
3845 | intslots = MIN (intslots, SPARC_INT_ARG_MAX - this_slotno); | |
3846 | intslots = MAX (intslots, 0); | |
3847 | ||
3848 | parms.nregs += intslots; | |
3849 | } | |
3850 | nregs = parms.nregs; | |
3851 | ||
3852 | /* Allocate the vector and handle some annoying special cases. */ | |
3853 | if (nregs == 0) | |
3854 | { | |
3855 | /* ??? Empty structure has no value? Duh? */ | |
3856 | if (typesize <= 0) | |
3857 | { | |
3858 | /* Though there's nothing really to store, return a word register | |
3859 | anyway so the rest of gcc doesn't go nuts. Returning a PARALLEL | |
3860 | leads to breakage due to the fact that there are zero bytes to | |
3861 | load. */ | |
1eac9f59 | 3862 | return gen_rtx_REG (mode, regbase); |
82d6b402 RH |
3863 | } |
3864 | else | |
3865 | { | |
3866 | /* ??? C++ has structures with no fields, and yet a size. Give up | |
3867 | for now and pass everything back in integer registers. */ | |
3868 | nregs = (typesize + UNITS_PER_WORD - 1) / UNITS_PER_WORD; | |
3869 | } | |
3870 | if (nregs + slotno > SPARC_INT_ARG_MAX) | |
3871 | nregs = SPARC_INT_ARG_MAX - slotno; | |
3872 | } | |
3873 | if (nregs == 0) | |
9208e4b2 | 3874 | abort (); |
82d6b402 | 3875 | |
1eac9f59 | 3876 | parms.ret = gen_rtx_PARALLEL (mode, rtvec_alloc (nregs)); |
82d6b402 RH |
3877 | |
3878 | /* Fill in the entries. */ | |
3879 | parms.nregs = 0; | |
3880 | parms.intoffset = 0; | |
3881 | function_arg_record_value_2 (type, 0, &parms); | |
3882 | function_arg_record_value_3 (typesize * BITS_PER_UNIT, &parms); | |
3883 | ||
3884 | if (parms.nregs != nregs) | |
3885 | abort (); | |
3886 | ||
3887 | return parms.ret; | |
3888 | } | |
3889 | ||
4fb4e4b8 DE |
3890 | /* Handle the FUNCTION_ARG macro. |
3891 | Determine where to put an argument to a function. | |
3892 | Value is zero to push the argument on the stack, | |
3893 | or a hard register in which to store the argument. | |
3894 | ||
3895 | CUM is a variable of type CUMULATIVE_ARGS which gives info about | |
3896 | the preceding args and about the function being called. | |
3897 | MODE is the argument's machine mode. | |
3898 | TYPE is the data type of the argument (as a tree). | |
3899 | This is null for libcalls where that information may | |
3900 | not be available. | |
3901 | NAMED is nonzero if this argument is a named parameter | |
3902 | (otherwise it is an extra parameter matching an ellipsis). | |
3903 | INCOMING_P is zero for FUNCTION_ARG, nonzero for FUNCTION_INCOMING_ARG. */ | |
61a55e8b DE |
3904 | |
3905 | rtx | |
4fb4e4b8 DE |
3906 | function_arg (cum, mode, type, named, incoming_p) |
3907 | const CUMULATIVE_ARGS *cum; | |
3908 | enum machine_mode mode; | |
3909 | tree type; | |
3910 | int named; | |
3911 | int incoming_p; | |
61a55e8b | 3912 | { |
4fb4e4b8 DE |
3913 | int regbase = (incoming_p |
3914 | ? SPARC_INCOMING_INT_ARG_FIRST | |
3915 | : SPARC_OUTGOING_INT_ARG_FIRST); | |
3916 | int slotno, regno, padding; | |
3917 | rtx reg; | |
3918 | ||
3919 | slotno = function_arg_slotno (cum, mode, type, named, incoming_p, | |
3920 | ®no, &padding); | |
3921 | ||
3922 | if (slotno == -1) | |
3923 | return 0; | |
3924 | ||
3925 | if (TARGET_ARCH32) | |
3926 | { | |
254110c2 | 3927 | reg = gen_rtx_REG (mode, regno); |
4fb4e4b8 DE |
3928 | return reg; |
3929 | } | |
3930 | ||
3931 | /* v9 fp args in reg slots beyond the int reg slots get passed in regs | |
3932 | but also have the slot allocated for them. | |
3933 | If no prototype is in scope fp values in register slots get passed | |
3934 | in two places, either fp regs and int regs or fp regs and memory. */ | |
3935 | if ((GET_MODE_CLASS (mode) == MODE_FLOAT | |
3936 | || GET_MODE_CLASS (mode) == MODE_COMPLEX_FLOAT) | |
3937 | && SPARC_FP_REG_P (regno)) | |
3938 | { | |
254110c2 | 3939 | reg = gen_rtx_REG (mode, regno); |
4fb4e4b8 DE |
3940 | if (cum->prototype_p || cum->libcall_p) |
3941 | { | |
3942 | /* "* 2" because fp reg numbers are recorded in 4 byte | |
3943 | quantities. */ | |
82d6b402 | 3944 | #if 0 |
4fb4e4b8 DE |
3945 | /* ??? This will cause the value to be passed in the fp reg and |
3946 | in the stack. When a prototype exists we want to pass the | |
3947 | value in the reg but reserve space on the stack. That's an | |
956d6950 | 3948 | optimization, and is deferred [for a bit]. */ |
4fb4e4b8 | 3949 | if ((regno - SPARC_FP_ARG_FIRST) >= SPARC_INT_ARG_MAX * 2) |
254110c2 | 3950 | return gen_rtx_PARALLEL (mode, |
4fb4e4b8 | 3951 | gen_rtvec (2, |
254110c2 | 3952 | gen_rtx_EXPR_LIST (VOIDmode, |
4fb4e4b8 | 3953 | NULL_RTX, const0_rtx), |
254110c2 | 3954 | gen_rtx_EXPR_LIST (VOIDmode, |
4fb4e4b8 DE |
3955 | reg, const0_rtx))); |
3956 | else | |
82d6b402 RH |
3957 | #else |
3958 | /* ??? It seems that passing back a register even when past | |
3959 | the area declared by REG_PARM_STACK_SPACE will allocate | |
3960 | space appropriately, and will not copy the data onto the | |
3961 | stack, exactly as we desire. | |
3962 | ||
3963 | This is due to locate_and_pad_parm being called in | |
3964 | expand_call whenever reg_parm_stack_space > 0, which | |
3965 | while benefical to our example here, would seem to be | |
3966 | in error from what had been intended. Ho hum... -- r~ */ | |
3967 | #endif | |
4fb4e4b8 DE |
3968 | return reg; |
3969 | } | |
3970 | else | |
3971 | { | |
82d6b402 RH |
3972 | rtx v0, v1; |
3973 | ||
4fb4e4b8 DE |
3974 | if ((regno - SPARC_FP_ARG_FIRST) < SPARC_INT_ARG_MAX * 2) |
3975 | { | |
82d6b402 RH |
3976 | int intreg; |
3977 | ||
3978 | /* On incoming, we don't need to know that the value | |
3979 | is passed in %f0 and %i0, and it confuses other parts | |
3980 | causing needless spillage even on the simplest cases. */ | |
3981 | if (incoming_p) | |
3982 | return reg; | |
3983 | ||
3984 | intreg = (SPARC_OUTGOING_INT_ARG_FIRST | |
3985 | + (regno - SPARC_FP_ARG_FIRST) / 2); | |
3986 | ||
3987 | v0 = gen_rtx_EXPR_LIST (VOIDmode, reg, const0_rtx); | |
3988 | v1 = gen_rtx_EXPR_LIST (VOIDmode, gen_rtx_REG (mode, intreg), | |
3989 | const0_rtx); | |
3990 | return gen_rtx_PARALLEL (mode, gen_rtvec (2, v0, v1)); | |
4fb4e4b8 DE |
3991 | } |
3992 | else | |
82d6b402 RH |
3993 | { |
3994 | v0 = gen_rtx_EXPR_LIST (VOIDmode, NULL_RTX, const0_rtx); | |
3995 | v1 = gen_rtx_EXPR_LIST (VOIDmode, reg, const0_rtx); | |
3996 | return gen_rtx_PARALLEL (mode, gen_rtvec (2, v0, v1)); | |
3997 | } | |
4fb4e4b8 DE |
3998 | } |
3999 | } | |
4000 | else if (type && TREE_CODE (type) == RECORD_TYPE) | |
4001 | { | |
4002 | /* Structures up to 16 bytes in size are passed in arg slots on the | |
4003 | stack and are promoted to registers where possible. */ | |
4fb4e4b8 DE |
4004 | |
4005 | if (int_size_in_bytes (type) > 16) | |
4006 | abort (); /* shouldn't get here */ | |
4007 | ||
1eac9f59 | 4008 | return function_arg_record_value (type, mode, slotno, named, regbase); |
4fb4e4b8 DE |
4009 | } |
4010 | else if (type && TREE_CODE (type) == UNION_TYPE) | |
4011 | { | |
4012 | enum machine_mode mode; | |
4013 | int bytes = int_size_in_bytes (type); | |
4014 | ||
4015 | if (bytes > 16) | |
4016 | abort (); | |
4017 | ||
4018 | mode = mode_for_size (bytes * BITS_PER_UNIT, MODE_INT, 0); | |
254110c2 | 4019 | reg = gen_rtx_REG (mode, regno); |
4fb4e4b8 | 4020 | } |
a7acd911 | 4021 | else |
4fb4e4b8 DE |
4022 | { |
4023 | /* Scalar or complex int. */ | |
254110c2 | 4024 | reg = gen_rtx_REG (mode, regno); |
4fb4e4b8 DE |
4025 | } |
4026 | ||
4027 | return reg; | |
4028 | } | |
a7acd911 | 4029 | |
4fb4e4b8 DE |
4030 | /* Handle the FUNCTION_ARG_PARTIAL_NREGS macro. |
4031 | For an arg passed partly in registers and partly in memory, | |
4032 | this is the number of registers used. | |
4033 | For args passed entirely in registers or entirely in memory, zero. | |
61a55e8b | 4034 | |
4fb4e4b8 DE |
4035 | Any arg that starts in the first 6 regs but won't entirely fit in them |
4036 | needs partial registers on v8. On v9, structures with integer | |
4037 | values in arg slots 5,6 will be passed in %o5 and SP+176, and complex fp | |
4038 | values that begin in the last fp reg [where "last fp reg" varies with the | |
4039 | mode] will be split between that reg and memory. */ | |
61a55e8b | 4040 | |
4fb4e4b8 DE |
4041 | int |
4042 | function_arg_partial_nregs (cum, mode, type, named) | |
4043 | const CUMULATIVE_ARGS *cum; | |
4044 | enum machine_mode mode; | |
4045 | tree type; | |
4046 | int named; | |
4047 | { | |
4048 | int slotno, regno, padding; | |
61a55e8b | 4049 | |
4fb4e4b8 DE |
4050 | /* We pass 0 for incoming_p here, it doesn't matter. */ |
4051 | slotno = function_arg_slotno (cum, mode, type, named, 0, ®no, &padding); | |
4052 | ||
4053 | if (slotno == -1) | |
4054 | return 0; | |
4055 | ||
4056 | if (TARGET_ARCH32) | |
bf62bbf1 | 4057 | { |
4fb4e4b8 DE |
4058 | if ((slotno + (mode == BLKmode |
4059 | ? ROUND_ADVANCE (int_size_in_bytes (type)) | |
4060 | : ROUND_ADVANCE (GET_MODE_SIZE (mode)))) | |
4061 | > NPARM_REGS (SImode)) | |
4062 | return NPARM_REGS (SImode) - slotno; | |
4063 | return 0; | |
4064 | } | |
4065 | else | |
4066 | { | |
4067 | if (type && AGGREGATE_TYPE_P (type)) | |
4068 | { | |
4069 | int size = int_size_in_bytes (type); | |
4070 | int align = TYPE_ALIGN (type); | |
bf62bbf1 | 4071 | |
4fb4e4b8 DE |
4072 | if (align == 16) |
4073 | slotno += slotno & 1; | |
4074 | if (size > 8 && size <= 16 | |
4075 | && slotno == SPARC_INT_ARG_MAX - 1) | |
4076 | return 1; | |
4077 | } | |
4078 | else if (GET_MODE_CLASS (mode) == MODE_COMPLEX_INT | |
4079 | || (GET_MODE_CLASS (mode) == MODE_COMPLEX_FLOAT | |
4080 | && ! TARGET_FPU)) | |
4081 | { | |
4082 | if (GET_MODE_ALIGNMENT (mode) == 128) | |
4083 | { | |
4084 | slotno += slotno & 1; | |
4085 | if (slotno == SPARC_INT_ARG_MAX - 2) | |
4086 | return 1; | |
4087 | } | |
4088 | else | |
4089 | { | |
4090 | if (slotno == SPARC_INT_ARG_MAX - 1) | |
4091 | return 1; | |
4092 | } | |
4093 | } | |
4094 | else if (GET_MODE_CLASS (mode) == MODE_COMPLEX_FLOAT) | |
4095 | { | |
4096 | if (GET_MODE_ALIGNMENT (mode) == 128) | |
4097 | slotno += slotno & 1; | |
4098 | if ((slotno + GET_MODE_SIZE (mode) / UNITS_PER_WORD) | |
4099 | > SPARC_FP_ARG_MAX) | |
4100 | return 1; | |
4101 | } | |
4102 | return 0; | |
bf62bbf1 | 4103 | } |
4fb4e4b8 | 4104 | } |
61a55e8b | 4105 | |
4fb4e4b8 DE |
4106 | /* Handle the FUNCTION_ARG_PASS_BY_REFERENCE macro. |
4107 | !v9: The SPARC ABI stipulates passing struct arguments (of any size) and | |
4108 | quad-precision floats by invisible reference. | |
82d6b402 | 4109 | v9: Aggregates greater than 16 bytes are passed by reference. |
4fb4e4b8 DE |
4110 | For Pascal, also pass arrays by reference. */ |
4111 | ||
4112 | int | |
4113 | function_arg_pass_by_reference (cum, mode, type, named) | |
db3d4438 | 4114 | const CUMULATIVE_ARGS *cum ATTRIBUTE_UNUSED; |
4fb4e4b8 DE |
4115 | enum machine_mode mode; |
4116 | tree type; | |
db3d4438 | 4117 | int named ATTRIBUTE_UNUSED; |
4fb4e4b8 DE |
4118 | { |
4119 | if (TARGET_ARCH32) | |
eadceb59 | 4120 | { |
5e9defae | 4121 | return ((type && AGGREGATE_TYPE_P (type)) |
4fb4e4b8 DE |
4122 | || mode == TFmode || mode == TCmode); |
4123 | } | |
4124 | else | |
4125 | { | |
4126 | return ((type && TREE_CODE (type) == ARRAY_TYPE) | |
82d6b402 RH |
4127 | /* Consider complex values as aggregates, so care for TCmode. */ |
4128 | || GET_MODE_SIZE (mode) > 16 | |
4fb4e4b8 DE |
4129 | || (type && AGGREGATE_TYPE_P (type) |
4130 | && int_size_in_bytes (type) > 16)); | |
4131 | } | |
4132 | } | |
4133 | ||
4134 | /* Handle the FUNCTION_ARG_ADVANCE macro. | |
4135 | Update the data in CUM to advance over an argument | |
4136 | of mode MODE and data type TYPE. | |
4137 | TYPE is null for libcalls where that information may not be available. */ | |
4138 | ||
4139 | void | |
4140 | function_arg_advance (cum, mode, type, named) | |
4141 | CUMULATIVE_ARGS *cum; | |
4142 | enum machine_mode mode; | |
4143 | tree type; | |
4144 | int named; | |
4145 | { | |
4146 | int slotno, regno, padding; | |
4147 | ||
4148 | /* We pass 0 for incoming_p here, it doesn't matter. */ | |
4149 | slotno = function_arg_slotno (cum, mode, type, named, 0, ®no, &padding); | |
4150 | ||
4151 | /* If register required leading padding, add it. */ | |
4152 | if (slotno != -1) | |
4153 | cum->words += padding; | |
eadceb59 | 4154 | |
4fb4e4b8 DE |
4155 | if (TARGET_ARCH32) |
4156 | { | |
4157 | cum->words += (mode != BLKmode | |
4158 | ? ROUND_ADVANCE (GET_MODE_SIZE (mode)) | |
4159 | : ROUND_ADVANCE (int_size_in_bytes (type))); | |
4160 | } | |
4161 | else | |
4162 | { | |
4163 | if (type && AGGREGATE_TYPE_P (type)) | |
4164 | { | |
4165 | int size = int_size_in_bytes (type); | |
4166 | ||
4167 | if (size <= 8) | |
4168 | ++cum->words; | |
4169 | else if (size <= 16) | |
4170 | cum->words += 2; | |
4171 | else /* passed by reference */ | |
4172 | ++cum->words; | |
4173 | } | |
4174 | else if (GET_MODE_CLASS (mode) == MODE_COMPLEX_INT) | |
4175 | { | |
4176 | cum->words += 2; | |
4177 | } | |
4178 | else if (GET_MODE_CLASS (mode) == MODE_COMPLEX_FLOAT) | |
4179 | { | |
4180 | cum->words += GET_MODE_SIZE (mode) / UNITS_PER_WORD; | |
4181 | } | |
4182 | else | |
4183 | { | |
4184 | cum->words += (mode != BLKmode | |
4185 | ? ROUND_ADVANCE (GET_MODE_SIZE (mode)) | |
4186 | : ROUND_ADVANCE (int_size_in_bytes (type))); | |
4187 | } | |
eadceb59 | 4188 | } |
4fb4e4b8 | 4189 | } |
eadceb59 | 4190 | |
4fb4e4b8 DE |
4191 | /* Handle the FUNCTION_ARG_PADDING macro. |
4192 | For the 64 bit ABI structs are always stored left shifted in their | |
4193 | argument slot. */ | |
4194 | ||
4195 | enum direction | |
4196 | function_arg_padding (mode, type) | |
4197 | enum machine_mode mode; | |
4198 | tree type; | |
4199 | { | |
c85f7c16 JL |
4200 | if (TARGET_ARCH64 && type != 0 && AGGREGATE_TYPE_P (type)) |
4201 | return upward; | |
61a55e8b | 4202 | |
4fb4e4b8 DE |
4203 | /* This is the default definition. */ |
4204 | return (! BYTES_BIG_ENDIAN | |
4205 | ? upward | |
4206 | : ((mode == BLKmode | |
4207 | ? (type && TREE_CODE (TYPE_SIZE (type)) == INTEGER_CST | |
4208 | && int_size_in_bytes (type) < (PARM_BOUNDARY / BITS_PER_UNIT)) | |
4209 | : GET_MODE_BITSIZE (mode) < PARM_BOUNDARY) | |
4210 | ? downward : upward)); | |
61a55e8b | 4211 | } |
82d6b402 RH |
4212 | |
4213 | /* Handle FUNCTION_VALUE, FUNCTION_OUTGOING_VALUE, and LIBCALL_VALUE macros. | |
4214 | For v9, function return values are subject to the same rules as arguments, | |
4215 | except that up to 32-bytes may be returned in registers. */ | |
4216 | ||
4217 | rtx | |
4218 | function_value (type, mode, incoming_p) | |
4219 | tree type; | |
4220 | enum machine_mode mode; | |
4221 | int incoming_p; | |
4222 | { | |
4223 | int regno; | |
4224 | int regbase = (incoming_p | |
4225 | ? SPARC_OUTGOING_INT_ARG_FIRST | |
4226 | : SPARC_INCOMING_INT_ARG_FIRST); | |
4227 | ||
4228 | if (TARGET_ARCH64 && type) | |
4229 | { | |
4230 | if (TREE_CODE (type) == RECORD_TYPE) | |
4231 | { | |
4232 | /* Structures up to 32 bytes in size are passed in registers, | |
4233 | promoted to fp registers where possible. */ | |
4234 | ||
4235 | if (int_size_in_bytes (type) > 32) | |
4236 | abort (); /* shouldn't get here */ | |
4237 | ||
1eac9f59 | 4238 | return function_arg_record_value (type, mode, 0, 1, regbase); |
82d6b402 RH |
4239 | } |
4240 | else if (TREE_CODE (type) == UNION_TYPE) | |
4241 | { | |
4242 | int bytes = int_size_in_bytes (type); | |
4243 | ||
4244 | if (bytes > 32) | |
4245 | abort (); | |
4246 | ||
4247 | mode = mode_for_size (bytes * BITS_PER_UNIT, MODE_INT, 0); | |
4248 | } | |
4249 | } | |
3c6088d2 JJ |
4250 | |
4251 | if (TARGET_ARCH64 | |
4252 | && GET_MODE_CLASS (mode) == MODE_INT | |
4253 | && GET_MODE_SIZE (mode) < UNITS_PER_WORD | |
4254 | && type && TREE_CODE (type) != UNION_TYPE) | |
4255 | mode = DImode; | |
82d6b402 RH |
4256 | |
4257 | if (incoming_p) | |
4258 | regno = BASE_RETURN_VALUE_REG (mode); | |
4259 | else | |
4260 | regno = BASE_OUTGOING_VALUE_REG (mode); | |
4261 | ||
254110c2 | 4262 | return gen_rtx_REG (mode, regno); |
82d6b402 RH |
4263 | } |
4264 | ||
648d2ffc RH |
4265 | /* Do what is necessary for `va_start'. We look at the current function |
4266 | to determine if stdarg or varargs is used and return the address of | |
4267 | the first unnamed parameter. */ | |
3ea1fdd3 JW |
4268 | |
4269 | rtx | |
648d2ffc | 4270 | sparc_builtin_saveregs () |
3ea1fdd3 | 4271 | { |
4fb4e4b8 | 4272 | int first_reg = current_function_args_info.words; |
3ea1fdd3 JW |
4273 | rtx address; |
4274 | int regno; | |
4275 | ||
4fb4e4b8 | 4276 | for (regno = first_reg; regno < NPARM_REGS (word_mode); regno++) |
254110c2 DM |
4277 | emit_move_insn (gen_rtx_MEM (word_mode, |
4278 | gen_rtx_PLUS (Pmode, | |
3ea1fdd3 JW |
4279 | frame_pointer_rtx, |
4280 | GEN_INT (STACK_POINTER_OFFSET | |
4281 | + UNITS_PER_WORD * regno))), | |
254110c2 | 4282 | gen_rtx_REG (word_mode, |
4fb4e4b8 | 4283 | BASE_INCOMING_ARG_REG (word_mode) + regno)); |
3ea1fdd3 | 4284 | |
254110c2 | 4285 | address = gen_rtx_PLUS (Pmode, |
3ea1fdd3 JW |
4286 | frame_pointer_rtx, |
4287 | GEN_INT (STACK_POINTER_OFFSET | |
4288 | + UNITS_PER_WORD * first_reg)); | |
4289 | ||
7d384cc0 | 4290 | if (current_function_check_memory_usage |
4fb4e4b8 | 4291 | && first_reg < NPARM_REGS (word_mode)) |
eadceb59 RK |
4292 | emit_library_call (chkr_set_right_libfunc, 1, VOIDmode, 3, |
4293 | address, ptr_mode, | |
4294 | GEN_INT (UNITS_PER_WORD | |
4fb4e4b8 | 4295 | * (NPARM_REGS (word_mode) - first_reg)), |
956d6950 JL |
4296 | TYPE_MODE (sizetype), GEN_INT (MEMORY_USE_RW), |
4297 | TYPE_MODE (integer_type_node)); | |
eadceb59 | 4298 | |
3ea1fdd3 JW |
4299 | return address; |
4300 | } | |
a8b2c8a1 RH |
4301 | |
4302 | /* Implement `va_start' for varargs and stdarg. */ | |
4303 | ||
4304 | void | |
4305 | sparc_va_start (stdarg_p, valist, nextarg) | |
4306 | int stdarg_p ATTRIBUTE_UNUSED; | |
4307 | tree valist; | |
4308 | rtx nextarg; | |
4309 | { | |
4310 | nextarg = expand_builtin_saveregs (); | |
4311 | std_expand_builtin_va_start (1, valist, nextarg); | |
4312 | } | |
4313 | ||
4314 | /* Implement `va_arg'. */ | |
4315 | ||
4316 | rtx | |
4317 | sparc_va_arg (valist, type) | |
4318 | tree valist, type; | |
4319 | { | |
4320 | HOST_WIDE_INT size, rsize, align; | |
da09e317 | 4321 | tree addr, incr; |
a8b2c8a1 RH |
4322 | rtx addr_rtx; |
4323 | int indirect = 0; | |
4324 | ||
4325 | /* Round up sizeof(type) to a word. */ | |
4326 | size = int_size_in_bytes (type); | |
4327 | rsize = (size + UNITS_PER_WORD - 1) & -UNITS_PER_WORD; | |
4328 | align = 0; | |
4329 | ||
4330 | if (TARGET_ARCH64) | |
4331 | { | |
4332 | if (TYPE_ALIGN (type) >= 2 * BITS_PER_WORD) | |
4333 | align = 2 * UNITS_PER_WORD; | |
4334 | ||
4335 | if (AGGREGATE_TYPE_P (type) && size > 16) | |
4336 | { | |
4337 | indirect = 1; | |
4338 | size = rsize = UNITS_PER_WORD; | |
4339 | } | |
4340 | } | |
4341 | else | |
4342 | { | |
4343 | if (AGGREGATE_TYPE_P (type) | |
4344 | || TYPE_MODE (type) == TFmode | |
4345 | || TYPE_MODE (type) == TCmode) | |
4346 | { | |
4347 | indirect = 1; | |
4348 | size = rsize = UNITS_PER_WORD; | |
4349 | } | |
4350 | else | |
4351 | { | |
4352 | /* ??? The old va-sparc.h implementation, for 8 byte objects | |
4353 | copied stuff to a temporary -- I don't see that that | |
4354 | provides any more alignment than the stack slot did. */ | |
4355 | } | |
4356 | } | |
4357 | ||
4358 | incr = valist; | |
4359 | if (align) | |
4360 | { | |
4361 | incr = fold (build (PLUS_EXPR, ptr_type_node, incr, | |
4362 | build_int_2 (align - 1, 0))); | |
4363 | incr = fold (build (BIT_AND_EXPR, ptr_type_node, incr, | |
4364 | build_int_2 (-align, -1))); | |
4365 | } | |
4366 | ||
4367 | addr = incr = save_expr (incr); | |
4368 | if (BYTES_BIG_ENDIAN && size < rsize) | |
4369 | { | |
4370 | addr = fold (build (PLUS_EXPR, ptr_type_node, incr, | |
4371 | build_int_2 (rsize - size, 0))); | |
4372 | } | |
4373 | incr = fold (build (PLUS_EXPR, ptr_type_node, incr, | |
4374 | build_int_2 (rsize, 0))); | |
4375 | ||
4376 | incr = build (MODIFY_EXPR, ptr_type_node, valist, incr); | |
4377 | TREE_SIDE_EFFECTS (incr) = 1; | |
4378 | expand_expr (incr, const0_rtx, VOIDmode, EXPAND_NORMAL); | |
4379 | ||
4380 | addr_rtx = expand_expr (addr, NULL, Pmode, EXPAND_NORMAL); | |
4381 | ||
4382 | if (indirect) | |
4383 | { | |
4384 | addr_rtx = force_reg (Pmode, addr_rtx); | |
4385 | addr_rtx = gen_rtx_MEM (Pmode, addr_rtx); | |
4386 | MEM_ALIAS_SET (addr_rtx) = get_varargs_alias_set (); | |
4387 | } | |
4388 | ||
4389 | return addr_rtx; | |
4390 | } | |
ab835497 RK |
4391 | \f |
4392 | /* Return the string to output a conditional branch to LABEL, which is | |
c4ce6853 DE |
4393 | the operand number of the label. OP is the conditional expression. |
4394 | XEXP (OP, 0) is assumed to be a condition code register (integer or | |
4395 | floating point) and its mode specifies what kind of comparison we made. | |
61a55e8b | 4396 | |
ab835497 RK |
4397 | REVERSED is non-zero if we should reverse the sense of the comparison. |
4398 | ||
4399 | ANNUL is non-zero if we should generate an annulling branch. | |
4400 | ||
c6b0465b JC |
4401 | NOOP is non-zero if we have to follow this branch by a noop. |
4402 | ||
4403 | INSN, if set, is the insn. */ | |
ab835497 RK |
4404 | |
4405 | char * | |
c6b0465b | 4406 | output_cbranch (op, label, reversed, annul, noop, insn) |
c4ce6853 | 4407 | rtx op; |
ab835497 RK |
4408 | int label; |
4409 | int reversed, annul, noop; | |
c6b0465b | 4410 | rtx insn; |
ab835497 | 4411 | { |
e0d80184 | 4412 | static char string[32]; |
ab835497 | 4413 | enum rtx_code code = GET_CODE (op); |
c4ce6853 DE |
4414 | rtx cc_reg = XEXP (op, 0); |
4415 | enum machine_mode mode = GET_MODE (cc_reg); | |
e0d80184 DM |
4416 | static char v8_labelno[] = "%lX"; |
4417 | static char v9_icc_labelno[] = "%%icc, %lX"; | |
4418 | static char v9_xcc_labelno[] = "%%xcc, %lX"; | |
4419 | static char v9_fcc_labelno[] = "%%fccX, %lY"; | |
61a55e8b | 4420 | char *labelno; |
e0d80184 | 4421 | int labeloff, spaces = 8; |
61a55e8b DE |
4422 | |
4423 | /* ??? !v9: FP branches cannot be preceded by another floating point insn. | |
ab835497 RK |
4424 | Because there is currently no concept of pre-delay slots, we can fix |
4425 | this only by always emitting a nop before a floating point branch. */ | |
4426 | ||
61a55e8b | 4427 | if ((mode == CCFPmode || mode == CCFPEmode) && ! TARGET_V9) |
ab835497 | 4428 | strcpy (string, "nop\n\t"); |
61a55e8b DE |
4429 | else |
4430 | string[0] = '\0'; | |
ab835497 RK |
4431 | |
4432 | /* If not floating-point or if EQ or NE, we can just reverse the code. */ | |
4d449554 JW |
4433 | if (reversed |
4434 | && ((mode != CCFPmode && mode != CCFPEmode) || code == EQ || code == NE)) | |
ab835497 RK |
4435 | code = reverse_condition (code), reversed = 0; |
4436 | ||
4437 | /* Start by writing the branch condition. */ | |
4438 | switch (code) | |
4439 | { | |
4440 | case NE: | |
4d449554 | 4441 | if (mode == CCFPmode || mode == CCFPEmode) |
e0d80184 DM |
4442 | { |
4443 | strcat (string, "fbne"); | |
4444 | spaces -= 4; | |
4445 | } | |
ab835497 | 4446 | else |
e0d80184 DM |
4447 | { |
4448 | strcpy (string, "bne"); | |
4449 | spaces -= 3; | |
4450 | } | |
ab835497 RK |
4451 | break; |
4452 | ||
4453 | case EQ: | |
4d449554 | 4454 | if (mode == CCFPmode || mode == CCFPEmode) |
e0d80184 DM |
4455 | { |
4456 | strcat (string, "fbe"); | |
4457 | spaces -= 3; | |
4458 | } | |
ab835497 | 4459 | else |
e0d80184 DM |
4460 | { |
4461 | strcpy (string, "be"); | |
4462 | spaces -= 2; | |
4463 | } | |
ab835497 RK |
4464 | break; |
4465 | ||
4466 | case GE: | |
4d449554 | 4467 | if (mode == CCFPmode || mode == CCFPEmode) |
ab835497 RK |
4468 | { |
4469 | if (reversed) | |
4470 | strcat (string, "fbul"); | |
4471 | else | |
4472 | strcat (string, "fbge"); | |
e0d80184 | 4473 | spaces -= 4; |
ab835497 RK |
4474 | } |
4475 | else if (mode == CC_NOOVmode) | |
e0d80184 DM |
4476 | { |
4477 | strcpy (string, "bpos"); | |
4478 | spaces -= 4; | |
4479 | } | |
ab835497 | 4480 | else |
e0d80184 DM |
4481 | { |
4482 | strcpy (string, "bge"); | |
4483 | spaces -= 3; | |
4484 | } | |
ab835497 RK |
4485 | break; |
4486 | ||
4487 | case GT: | |
4d449554 | 4488 | if (mode == CCFPmode || mode == CCFPEmode) |
ab835497 RK |
4489 | { |
4490 | if (reversed) | |
e0d80184 DM |
4491 | { |
4492 | strcat (string, "fbule"); | |
4493 | spaces -= 5; | |
4494 | } | |
ab835497 | 4495 | else |
e0d80184 DM |
4496 | { |
4497 | strcat (string, "fbg"); | |
4498 | spaces -= 3; | |
4499 | } | |
ab835497 RK |
4500 | } |
4501 | else | |
e0d80184 DM |
4502 | { |
4503 | strcpy (string, "bg"); | |
4504 | spaces -= 2; | |
4505 | } | |
ab835497 RK |
4506 | break; |
4507 | ||
4508 | case LE: | |
4d449554 | 4509 | if (mode == CCFPmode || mode == CCFPEmode) |
ab835497 RK |
4510 | { |
4511 | if (reversed) | |
4512 | strcat (string, "fbug"); | |
4513 | else | |
4514 | strcat (string, "fble"); | |
e0d80184 | 4515 | spaces -= 4; |
ab835497 RK |
4516 | } |
4517 | else | |
e0d80184 DM |
4518 | { |
4519 | strcpy (string, "ble"); | |
4520 | spaces -= 3; | |
4521 | } | |
ab835497 RK |
4522 | break; |
4523 | ||
4524 | case LT: | |
4d449554 | 4525 | if (mode == CCFPmode || mode == CCFPEmode) |
ab835497 RK |
4526 | { |
4527 | if (reversed) | |
e0d80184 DM |
4528 | { |
4529 | strcat (string, "fbuge"); | |
4530 | spaces -= 5; | |
4531 | } | |
ab835497 | 4532 | else |
e0d80184 DM |
4533 | { |
4534 | strcat (string, "fbl"); | |
4535 | spaces -= 3; | |
4536 | } | |
ab835497 RK |
4537 | } |
4538 | else if (mode == CC_NOOVmode) | |
e0d80184 DM |
4539 | { |
4540 | strcpy (string, "bneg"); | |
4541 | spaces -= 4; | |
4542 | } | |
ab835497 | 4543 | else |
e0d80184 DM |
4544 | { |
4545 | strcpy (string, "bl"); | |
4546 | spaces -= 2; | |
4547 | } | |
ab835497 RK |
4548 | break; |
4549 | ||
4550 | case GEU: | |
4551 | strcpy (string, "bgeu"); | |
e0d80184 | 4552 | spaces -= 4; |
ab835497 RK |
4553 | break; |
4554 | ||
4555 | case GTU: | |
4556 | strcpy (string, "bgu"); | |
e0d80184 | 4557 | spaces -= 3; |
ab835497 RK |
4558 | break; |
4559 | ||
4560 | case LEU: | |
4561 | strcpy (string, "bleu"); | |
e0d80184 | 4562 | spaces -= 4; |
ab835497 RK |
4563 | break; |
4564 | ||
4565 | case LTU: | |
4566 | strcpy (string, "blu"); | |
e0d80184 | 4567 | spaces -= 3; |
ab835497 | 4568 | break; |
91a51951 KG |
4569 | |
4570 | default: | |
e0d80184 | 4571 | abort (); |
ab835497 RK |
4572 | } |
4573 | ||
4574 | /* Now add the annulling, the label, and a possible noop. */ | |
4575 | if (annul) | |
e0d80184 DM |
4576 | { |
4577 | strcat (string, ",a"); | |
4578 | spaces -= 2; | |
4579 | } | |
ab835497 | 4580 | |
61a55e8b DE |
4581 | if (! TARGET_V9) |
4582 | { | |
e0d80184 | 4583 | labeloff = 2; |
61a55e8b DE |
4584 | labelno = v8_labelno; |
4585 | } | |
4586 | else | |
4587 | { | |
c6b0465b JC |
4588 | rtx note; |
4589 | ||
4590 | if (insn && (note = find_reg_note (insn, REG_BR_PRED, NULL_RTX))) | |
e0d80184 DM |
4591 | { |
4592 | strcat (string, | |
4593 | INTVAL (XEXP (note, 0)) & ATTR_FLAG_likely ? ",pt" : ",pn"); | |
4594 | spaces -= 3; | |
4595 | } | |
c6b0465b | 4596 | |
61a55e8b DE |
4597 | labeloff = 9; |
4598 | if (mode == CCFPmode || mode == CCFPEmode) | |
4599 | { | |
4600 | labeloff = 10; | |
4601 | labelno = v9_fcc_labelno; | |
4602 | /* Set the char indicating the number of the fcc reg to use. */ | |
e0d80184 | 4603 | labelno[5] = REGNO (cc_reg) - SPARC_FIRST_V9_FCC_REG + '0'; |
61a55e8b DE |
4604 | } |
4605 | else if (mode == CCXmode || mode == CCX_NOOVmode) | |
4606 | labelno = v9_xcc_labelno; | |
4607 | else | |
4608 | labelno = v9_icc_labelno; | |
4609 | } | |
4610 | /* Set the char indicating the number of the operand containing the | |
4611 | label_ref. */ | |
4612 | labelno[labeloff] = label + '0'; | |
e0d80184 DM |
4613 | if (spaces > 0) |
4614 | strcat (string, "\t"); | |
4615 | else | |
4616 | strcat (string, " "); | |
61a55e8b DE |
4617 | strcat (string, labelno); |
4618 | ||
4619 | if (noop) | |
4620 | strcat (string, "\n\tnop"); | |
4621 | ||
4622 | return string; | |
4623 | } | |
4624 | ||
4625 | /* Return the string to output a conditional branch to LABEL, testing | |
4626 | register REG. LABEL is the operand number of the label; REG is the | |
4627 | operand number of the reg. OP is the conditional expression. The mode | |
4628 | of REG says what kind of comparison we made. | |
4629 | ||
4630 | REVERSED is non-zero if we should reverse the sense of the comparison. | |
4631 | ||
4632 | ANNUL is non-zero if we should generate an annulling branch. | |
4633 | ||
4634 | NOOP is non-zero if we have to follow this branch by a noop. */ | |
4635 | ||
4636 | char * | |
e0d80184 | 4637 | output_v9branch (op, reg, label, reversed, annul, noop, insn) |
61a55e8b DE |
4638 | rtx op; |
4639 | int reg, label; | |
4640 | int reversed, annul, noop; | |
e0d80184 | 4641 | rtx insn; |
61a55e8b DE |
4642 | { |
4643 | static char string[20]; | |
4644 | enum rtx_code code = GET_CODE (op); | |
4645 | enum machine_mode mode = GET_MODE (XEXP (op, 0)); | |
e0d80184 DM |
4646 | static char labelno[] = "%X, %lX"; |
4647 | rtx note; | |
4648 | int spaces = 8; | |
61a55e8b DE |
4649 | |
4650 | /* If not floating-point or if EQ or NE, we can just reverse the code. */ | |
4651 | if (reversed) | |
4652 | code = reverse_condition (code), reversed = 0; | |
4653 | ||
4654 | /* Only 64 bit versions of these instructions exist. */ | |
4655 | if (mode != DImode) | |
4656 | abort (); | |
4657 | ||
4658 | /* Start by writing the branch condition. */ | |
4659 | ||
4660 | switch (code) | |
4661 | { | |
4662 | case NE: | |
4663 | strcpy (string, "brnz"); | |
e0d80184 | 4664 | spaces -= 4; |
61a55e8b DE |
4665 | break; |
4666 | ||
4667 | case EQ: | |
4668 | strcpy (string, "brz"); | |
e0d80184 | 4669 | spaces -= 3; |
61a55e8b DE |
4670 | break; |
4671 | ||
4672 | case GE: | |
4673 | strcpy (string, "brgez"); | |
e0d80184 | 4674 | spaces -= 5; |
61a55e8b DE |
4675 | break; |
4676 | ||
4677 | case LT: | |
4678 | strcpy (string, "brlz"); | |
e0d80184 | 4679 | spaces -= 4; |
61a55e8b DE |
4680 | break; |
4681 | ||
4682 | case LE: | |
4683 | strcpy (string, "brlez"); | |
e0d80184 | 4684 | spaces -= 5; |
61a55e8b DE |
4685 | break; |
4686 | ||
4687 | case GT: | |
4688 | strcpy (string, "brgz"); | |
e0d80184 | 4689 | spaces -= 4; |
61a55e8b DE |
4690 | break; |
4691 | ||
4692 | default: | |
4693 | abort (); | |
4694 | } | |
4695 | ||
4696 | /* Now add the annulling, reg, label, and nop. */ | |
4697 | if (annul) | |
e0d80184 DM |
4698 | { |
4699 | strcat (string, ",a"); | |
4700 | spaces -= 2; | |
4701 | } | |
61a55e8b | 4702 | |
e0d80184 DM |
4703 | if (insn && (note = find_reg_note (insn, REG_BR_PRED, NULL_RTX))) |
4704 | { | |
4705 | strcat (string, | |
4706 | INTVAL (XEXP (note, 0)) & ATTR_FLAG_likely ? ",pt" : ",pn"); | |
4707 | spaces -= 3; | |
4708 | } | |
61a55e8b | 4709 | |
e0d80184 | 4710 | labelno[1] = reg + '0'; |
61a55e8b | 4711 | labelno[6] = label + '0'; |
e0d80184 DM |
4712 | if (spaces > 0) |
4713 | strcat (string, "\t"); | |
4714 | else | |
4715 | strcat (string, " "); | |
ab835497 RK |
4716 | strcat (string, labelno); |
4717 | ||
4718 | if (noop) | |
4719 | strcat (string, "\n\tnop"); | |
4720 | ||
4721 | return string; | |
4722 | } | |
4723 | ||
284d86e9 JC |
4724 | /* Renumber registers in delay slot. Replace registers instead of |
4725 | renumbering because they may be shared. | |
795068a4 | 4726 | |
284d86e9 JC |
4727 | This does not handle instructions other than move. */ |
4728 | ||
4729 | static void | |
4730 | epilogue_renumber (where) | |
4731 | rtx *where; | |
4732 | { | |
4733 | rtx x = *where; | |
4734 | enum rtx_code code = GET_CODE (x); | |
4735 | ||
4736 | switch (code) | |
4737 | { | |
4738 | case MEM: | |
4739 | *where = x = copy_rtx (x); | |
4740 | epilogue_renumber (&XEXP (x, 0)); | |
4741 | return; | |
4742 | ||
4743 | case REG: | |
4744 | { | |
4745 | int regno = REGNO (x); | |
4746 | if (regno > 8 && regno < 24) | |
4747 | abort (); | |
4748 | if (regno >= 24 && regno < 32) | |
4749 | *where = gen_rtx_REG (GET_MODE (x), regno - 16); | |
4750 | return; | |
4751 | } | |
4752 | case CONST_INT: | |
4753 | case CONST_DOUBLE: | |
4754 | case CONST: | |
4755 | case SYMBOL_REF: | |
4756 | case LABEL_REF: | |
4757 | return; | |
4758 | ||
4759 | case IOR: | |
4760 | case AND: | |
4761 | case XOR: | |
4762 | case PLUS: | |
4763 | case MINUS: | |
4764 | epilogue_renumber (&XEXP (x, 1)); | |
4765 | case NEG: | |
4766 | case NOT: | |
4767 | epilogue_renumber (&XEXP (x, 0)); | |
4768 | return; | |
4769 | ||
4770 | default: | |
4771 | debug_rtx (*where); | |
9208e4b2 | 4772 | abort (); |
284d86e9 JC |
4773 | } |
4774 | } | |
4775 | ||
4776 | /* Output assembler code to return from a function. */ | |
61a55e8b | 4777 | |
3bb5de61 | 4778 | const char * |
ab835497 RK |
4779 | output_return (operands) |
4780 | rtx *operands; | |
4781 | { | |
284d86e9 JC |
4782 | rtx delay = final_sequence ? XVECEXP (final_sequence, 0, 1) : 0; |
4783 | ||
ab835497 RK |
4784 | if (leaf_label) |
4785 | { | |
4786 | operands[0] = leaf_label; | |
284d86e9 | 4787 | return "b%* %l0%("; |
ab835497 | 4788 | } |
54ff41b7 | 4789 | else if (current_function_uses_only_leaf_regs) |
ab835497 | 4790 | { |
284d86e9 JC |
4791 | /* No delay slot in a leaf function. */ |
4792 | if (delay) | |
4793 | abort (); | |
4794 | ||
915f619f JW |
4795 | /* If we didn't allocate a frame pointer for the current function, |
4796 | the stack pointer might have been adjusted. Output code to | |
4797 | restore it now. */ | |
4798 | ||
3a598fbe | 4799 | operands[0] = GEN_INT (actual_fsize); |
915f619f JW |
4800 | |
4801 | /* Use sub of negated value in first two cases instead of add to | |
4802 | allow actual_fsize == 4096. */ | |
4803 | ||
4804 | if (actual_fsize <= 4096) | |
ab835497 | 4805 | { |
61a55e8b | 4806 | if (SKIP_CALLERS_UNIMP_P) |
e0d80184 | 4807 | return "jmp\t%%o7+12\n\tsub\t%%sp, -%0, %%sp"; |
ab835497 | 4808 | else |
e0d80184 | 4809 | return "retl\n\tsub\t%%sp, -%0, %%sp"; |
ab835497 | 4810 | } |
915f619f | 4811 | else if (actual_fsize <= 8192) |
ab835497 | 4812 | { |
3a598fbe | 4813 | operands[0] = GEN_INT (actual_fsize - 4096); |
61a55e8b | 4814 | if (SKIP_CALLERS_UNIMP_P) |
e0d80184 | 4815 | return "sub\t%%sp, -4096, %%sp\n\tjmp\t%%o7+12\n\tsub\t%%sp, -%0, %%sp"; |
915f619f | 4816 | else |
e0d80184 | 4817 | return "sub\t%%sp, -4096, %%sp\n\tretl\n\tsub\t%%sp, -%0, %%sp"; |
915f619f | 4818 | } |
61a55e8b | 4819 | else if (SKIP_CALLERS_UNIMP_P) |
915f619f JW |
4820 | { |
4821 | if ((actual_fsize & 0x3ff) != 0) | |
e0d80184 | 4822 | return "sethi\t%%hi(%a0), %%g1\n\tor\t%%g1, %%lo(%a0), %%g1\n\tjmp\t%%o7+12\n\tadd\t%%sp, %%g1, %%sp"; |
ab835497 | 4823 | else |
e0d80184 | 4824 | return "sethi\t%%hi(%a0), %%g1\n\tjmp\t%%o7+12\n\tadd\t%%sp, %%g1, %%sp"; |
915f619f JW |
4825 | } |
4826 | else | |
4827 | { | |
4828 | if ((actual_fsize & 0x3ff) != 0) | |
ab835497 | 4829 | return "sethi %%hi(%a0),%%g1\n\tor %%g1,%%lo(%a0),%%g1\n\tretl\n\tadd %%sp,%%g1,%%sp"; |
915f619f JW |
4830 | else |
4831 | return "sethi %%hi(%a0),%%g1\n\tretl\n\tadd %%sp,%%g1,%%sp"; | |
ab835497 RK |
4832 | } |
4833 | } | |
284d86e9 JC |
4834 | else if (TARGET_V9) |
4835 | { | |
4836 | if (delay) | |
4837 | { | |
4838 | epilogue_renumber (&SET_DEST (PATTERN (delay))); | |
4839 | epilogue_renumber (&SET_SRC (PATTERN (delay))); | |
4840 | } | |
4841 | if (SKIP_CALLERS_UNIMP_P) | |
e0d80184 | 4842 | return "return\t%%i7+12%#"; |
284d86e9 | 4843 | else |
e0d80184 | 4844 | return "return\t%%i7+8%#"; |
284d86e9 | 4845 | } |
ab835497 RK |
4846 | else |
4847 | { | |
284d86e9 JC |
4848 | if (delay) |
4849 | abort (); | |
61a55e8b | 4850 | if (SKIP_CALLERS_UNIMP_P) |
e0d80184 | 4851 | return "jmp\t%%i7+12\n\trestore"; |
ab835497 RK |
4852 | else |
4853 | return "ret\n\trestore"; | |
4854 | } | |
4855 | } | |
ab835497 RK |
4856 | \f |
4857 | /* Leaf functions and non-leaf functions have different needs. */ | |
4858 | ||
4859 | static int | |
4860 | reg_leaf_alloc_order[] = REG_LEAF_ALLOC_ORDER; | |
4861 | ||
4862 | static int | |
4863 | reg_nonleaf_alloc_order[] = REG_ALLOC_ORDER; | |
4864 | ||
4865 | static int *reg_alloc_orders[] = { | |
4866 | reg_leaf_alloc_order, | |
4867 | reg_nonleaf_alloc_order}; | |
4868 | ||
4869 | void | |
4870 | order_regs_for_local_alloc () | |
4871 | { | |
4872 | static int last_order_nonleaf = 1; | |
4873 | ||
4874 | if (regs_ever_live[15] != last_order_nonleaf) | |
4875 | { | |
4876 | last_order_nonleaf = !last_order_nonleaf; | |
d2889939 RK |
4877 | bcopy ((char *) reg_alloc_orders[last_order_nonleaf], |
4878 | (char *) reg_alloc_order, FIRST_PSEUDO_REGISTER * sizeof (int)); | |
ab835497 RK |
4879 | } |
4880 | } | |
4881 | \f | |
e0d80184 DM |
4882 | /* Return 1 if REG and MEM are legitimate enough to allow the various |
4883 | mem<-->reg splits to be run. */ | |
4884 | ||
4885 | int | |
e61c29e9 | 4886 | sparc_splitdi_legitimate (reg, mem) |
e0d80184 DM |
4887 | rtx reg; |
4888 | rtx mem; | |
4889 | { | |
e0d80184 DM |
4890 | /* Punt if we are here by mistake. */ |
4891 | if (! reload_completed) | |
9208e4b2 | 4892 | abort (); |
e0d80184 DM |
4893 | |
4894 | /* We must have an offsettable memory reference. */ | |
4895 | if (! offsettable_memref_p (mem)) | |
4896 | return 0; | |
4897 | ||
4898 | /* If we have legitimate args for ldd/std, we do not want | |
4899 | the split to happen. */ | |
4900 | if ((REGNO (reg) % 2) == 0 | |
4901 | && mem_min_alignment (mem, 8)) | |
4902 | return 0; | |
4903 | ||
4904 | /* Success. */ | |
4905 | return 1; | |
4906 | } | |
4907 | ||
e61c29e9 DM |
4908 | /* Return 1 if x and y are some kind of REG and they refer to |
4909 | different hard registers. This test is guarenteed to be | |
4910 | run after reload. */ | |
4911 | ||
4912 | int | |
4913 | sparc_absnegfloat_split_legitimate (x, y) | |
4914 | rtx x, y; | |
4915 | { | |
4916 | if (GET_CODE (x) == SUBREG) | |
4917 | x = alter_subreg (x); | |
4918 | if (GET_CODE (x) != REG) | |
4919 | return 0; | |
4920 | if (GET_CODE (y) == SUBREG) | |
4921 | y = alter_subreg (y); | |
4922 | if (GET_CODE (y) != REG) | |
4923 | return 0; | |
4924 | if (REGNO (x) == REGNO (y)) | |
4925 | return 0; | |
4926 | return 1; | |
4927 | } | |
4928 | ||
35016322 JW |
4929 | /* Return 1 if REGNO (reg1) is even and REGNO (reg1) == REGNO (reg2) - 1. |
4930 | This makes them candidates for using ldd and std insns. | |
4931 | ||
bfd6bc60 | 4932 | Note reg1 and reg2 *must* be hard registers. */ |
35016322 JW |
4933 | |
4934 | int | |
7c56249d | 4935 | registers_ok_for_ldd_peep (reg1, reg2) |
35016322 JW |
4936 | rtx reg1, reg2; |
4937 | { | |
35016322 JW |
4938 | /* We might have been passed a SUBREG. */ |
4939 | if (GET_CODE (reg1) != REG || GET_CODE (reg2) != REG) | |
4940 | return 0; | |
4941 | ||
35016322 JW |
4942 | if (REGNO (reg1) % 2 != 0) |
4943 | return 0; | |
4944 | ||
bfd6bc60 JC |
4945 | /* Integer ldd is deprecated in SPARC V9 */ |
4946 | if (TARGET_V9 && REGNO (reg1) < 32) | |
4947 | return 0; | |
4948 | ||
35016322 | 4949 | return (REGNO (reg1) == REGNO (reg2) - 1); |
35016322 JW |
4950 | } |
4951 | ||
4952 | /* Return 1 if addr1 and addr2 are suitable for use in an ldd or | |
4953 | std insn. | |
4954 | ||
4955 | This can only happen when addr1 and addr2 are consecutive memory | |
4956 | locations (addr1 + 4 == addr2). addr1 must also be aligned on a | |
4957 | 64 bit boundary (addr1 % 8 == 0). | |
4958 | ||
4959 | We know %sp and %fp are kept aligned on a 64 bit boundary. Other | |
4960 | registers are assumed to *never* be properly aligned and are | |
4961 | rejected. | |
4962 | ||
4963 | Knowing %sp and %fp are kept aligned on a 64 bit boundary, we | |
4964 | need only check that the offset for addr1 % 8 == 0. */ | |
4965 | ||
4966 | int | |
7c56249d | 4967 | addrs_ok_for_ldd_peep (addr1, addr2) |
35016322 JW |
4968 | rtx addr1, addr2; |
4969 | { | |
4970 | int reg1, offset1; | |
4971 | ||
4972 | /* Extract a register number and offset (if used) from the first addr. */ | |
4973 | if (GET_CODE (addr1) == PLUS) | |
4974 | { | |
4975 | /* If not a REG, return zero. */ | |
4976 | if (GET_CODE (XEXP (addr1, 0)) != REG) | |
4977 | return 0; | |
4978 | else | |
4979 | { | |
4980 | reg1 = REGNO (XEXP (addr1, 0)); | |
4981 | /* The offset must be constant! */ | |
4982 | if (GET_CODE (XEXP (addr1, 1)) != CONST_INT) | |
4983 | return 0; | |
4984 | offset1 = INTVAL (XEXP (addr1, 1)); | |
4985 | } | |
4986 | } | |
4987 | else if (GET_CODE (addr1) != REG) | |
4988 | return 0; | |
4989 | else | |
4990 | { | |
4991 | reg1 = REGNO (addr1); | |
4992 | /* This was a simple (mem (reg)) expression. Offset is 0. */ | |
4993 | offset1 = 0; | |
4994 | } | |
4995 | ||
4996 | /* Make sure the second address is a (mem (plus (reg) (const_int). */ | |
4997 | if (GET_CODE (addr2) != PLUS) | |
4998 | return 0; | |
4999 | ||
5000 | if (GET_CODE (XEXP (addr2, 0)) != REG | |
5001 | || GET_CODE (XEXP (addr2, 1)) != CONST_INT) | |
5002 | return 0; | |
5003 | ||
5004 | /* Only %fp and %sp are allowed. Additionally both addresses must | |
5005 | use the same register. */ | |
5006 | if (reg1 != FRAME_POINTER_REGNUM && reg1 != STACK_POINTER_REGNUM) | |
5007 | return 0; | |
5008 | ||
5009 | if (reg1 != REGNO (XEXP (addr2, 0))) | |
5010 | return 0; | |
5011 | ||
2296cba3 | 5012 | /* The first offset must be evenly divisible by 8 to ensure the |
35016322 JW |
5013 | address is 64 bit aligned. */ |
5014 | if (offset1 % 8 != 0) | |
5015 | return 0; | |
5016 | ||
5017 | /* The offset for the second addr must be 4 more than the first addr. */ | |
5018 | if (INTVAL (XEXP (addr2, 1)) != offset1 + 4) | |
5019 | return 0; | |
5020 | ||
5021 | /* All the tests passed. addr1 and addr2 are valid for ldd and std | |
5022 | instructions. */ | |
5023 | return 1; | |
5024 | } | |
7c56249d JL |
5025 | |
5026 | /* Return 1 if reg is a pseudo, or is the first register in | |
5027 | a hard register pair. This makes it a candidate for use in | |
5028 | ldd and std insns. */ | |
5029 | ||
5030 | int | |
5031 | register_ok_for_ldd (reg) | |
5032 | rtx reg; | |
5033 | { | |
7c56249d JL |
5034 | /* We might have been passed a SUBREG. */ |
5035 | if (GET_CODE (reg) != REG) | |
5036 | return 0; | |
5037 | ||
5038 | if (REGNO (reg) < FIRST_PSEUDO_REGISTER) | |
5039 | return (REGNO (reg) % 2 == 0); | |
5040 | else | |
5041 | return 1; | |
7c56249d | 5042 | } |
ab835497 | 5043 | \f |
ab835497 RK |
5044 | /* Print operand X (an rtx) in assembler syntax to file FILE. |
5045 | CODE is a letter or dot (`z' in `%z0') or 0 if no letter was specified. | |
5046 | For `%' followed by punctuation, CODE is the punctuation and X is null. */ | |
5047 | ||
5048 | void | |
5049 | print_operand (file, x, code) | |
5050 | FILE *file; | |
5051 | rtx x; | |
5052 | int code; | |
5053 | { | |
5054 | switch (code) | |
5055 | { | |
5056 | case '#': | |
5057 | /* Output a 'nop' if there's nothing for the delay slot. */ | |
5058 | if (dbr_sequence_length () == 0) | |
e0d80184 | 5059 | fputs ("\n\t nop", file); |
ab835497 RK |
5060 | return; |
5061 | case '*': | |
c6ce0969 JW |
5062 | /* Output an annul flag if there's nothing for the delay slot and we |
5063 | are optimizing. This is always used with '(' below. */ | |
5064 | /* Sun OS 4.1.1 dbx can't handle an annulled unconditional branch; | |
5065 | this is a dbx bug. So, we only do this when optimizing. */ | |
bfd6bc60 JC |
5066 | /* On UltraSPARC, a branch in a delay slot causes a pipeline flush. |
5067 | Always emit a nop in case the next instruction is a branch. */ | |
5068 | if (dbr_sequence_length () == 0 | |
284d86e9 | 5069 | && (optimize && (int)sparc_cpu < PROCESSOR_V9)) |
c6ce0969 JW |
5070 | fputs (",a", file); |
5071 | return; | |
5072 | case '(': | |
5073 | /* Output a 'nop' if there's nothing for the delay slot and we are | |
5074 | not optimizing. This is always used with '*' above. */ | |
bfd6bc60 | 5075 | if (dbr_sequence_length () == 0 |
284d86e9 | 5076 | && ! (optimize && (int)sparc_cpu < PROCESSOR_V9)) |
e0d80184 | 5077 | fputs ("\n\t nop", file); |
ab835497 | 5078 | return; |
61a55e8b | 5079 | case '_': |
a0a301fc DE |
5080 | /* Output the Embedded Medium/Anywhere code model base register. */ |
5081 | fputs (EMBMEDANY_BASE_REG, file); | |
61a55e8b DE |
5082 | return; |
5083 | case '@': | |
5084 | /* Print out what we are using as the frame pointer. This might | |
5085 | be %fp, or might be %sp+offset. */ | |
5086 | /* ??? What if offset is too big? Perhaps the caller knows it isn't? */ | |
5087 | fprintf (file, "%s+%d", frame_base_name, frame_base_offset); | |
5088 | return; | |
ab835497 RK |
5089 | case 'Y': |
5090 | /* Adjust the operand to take into account a RESTORE operation. */ | |
bfd6bc60 JC |
5091 | if (GET_CODE (x) == CONST_INT) |
5092 | break; | |
5093 | else if (GET_CODE (x) != REG) | |
b3b1e8bd JW |
5094 | output_operand_lossage ("Invalid %%Y operand"); |
5095 | else if (REGNO (x) < 8) | |
ab835497 RK |
5096 | fputs (reg_names[REGNO (x)], file); |
5097 | else if (REGNO (x) >= 24 && REGNO (x) < 32) | |
5098 | fputs (reg_names[REGNO (x)-16], file); | |
5099 | else | |
b3b1e8bd | 5100 | output_operand_lossage ("Invalid %%Y operand"); |
ab835497 | 5101 | return; |
0f39a9aa DE |
5102 | case 'L': |
5103 | /* Print out the low order register name of a register pair. */ | |
5104 | if (WORDS_BIG_ENDIAN) | |
5105 | fputs (reg_names[REGNO (x)+1], file); | |
5106 | else | |
5107 | fputs (reg_names[REGNO (x)], file); | |
5108 | return; | |
5109 | case 'H': | |
5110 | /* Print out the high order register name of a register pair. */ | |
5111 | if (WORDS_BIG_ENDIAN) | |
5112 | fputs (reg_names[REGNO (x)], file); | |
5113 | else | |
5114 | fputs (reg_names[REGNO (x)+1], file); | |
5115 | return; | |
ab835497 | 5116 | case 'R': |
795068a4 | 5117 | /* Print out the second register name of a register pair or quad. |
ab835497 RK |
5118 | I.e., R (%o0) => %o1. */ |
5119 | fputs (reg_names[REGNO (x)+1], file); | |
5120 | return; | |
795068a4 JW |
5121 | case 'S': |
5122 | /* Print out the third register name of a register quad. | |
5123 | I.e., S (%o0) => %o2. */ | |
5124 | fputs (reg_names[REGNO (x)+2], file); | |
5125 | return; | |
5126 | case 'T': | |
5127 | /* Print out the fourth register name of a register quad. | |
5128 | I.e., T (%o0) => %o3. */ | |
5129 | fputs (reg_names[REGNO (x)+3], file); | |
5130 | return; | |
304b7a23 DE |
5131 | case 'x': |
5132 | /* Print a condition code register. */ | |
5133 | if (REGNO (x) == SPARC_ICC_REG) | |
5134 | { | |
5135 | /* We don't handle CC[X]_NOOVmode because they're not supposed | |
5136 | to occur here. */ | |
5137 | if (GET_MODE (x) == CCmode) | |
5138 | fputs ("%icc", file); | |
5139 | else if (GET_MODE (x) == CCXmode) | |
5140 | fputs ("%xcc", file); | |
5141 | else | |
5142 | abort (); | |
5143 | } | |
5144 | else | |
5145 | /* %fccN register */ | |
5146 | fputs (reg_names[REGNO (x)], file); | |
5147 | return; | |
ab835497 RK |
5148 | case 'm': |
5149 | /* Print the operand's address only. */ | |
5150 | output_address (XEXP (x, 0)); | |
5151 | return; | |
5152 | case 'r': | |
5153 | /* In this case we need a register. Use %g0 if the | |
77a02b01 | 5154 | operand is const0_rtx. */ |
76052e74 RS |
5155 | if (x == const0_rtx |
5156 | || (GET_MODE (x) != VOIDmode && x == CONST0_RTX (GET_MODE (x)))) | |
ab835497 RK |
5157 | { |
5158 | fputs ("%g0", file); | |
5159 | return; | |
5160 | } | |
5161 | else | |
5162 | break; | |
5163 | ||
61a55e8b | 5164 | case 'A': |
ab835497 RK |
5165 | switch (GET_CODE (x)) |
5166 | { | |
5167 | case IOR: fputs ("or", file); break; | |
5168 | case AND: fputs ("and", file); break; | |
5169 | case XOR: fputs ("xor", file); break; | |
b3b1e8bd | 5170 | default: output_operand_lossage ("Invalid %%A operand"); |
ab835497 RK |
5171 | } |
5172 | return; | |
5173 | ||
5174 | case 'B': | |
5175 | switch (GET_CODE (x)) | |
5176 | { | |
5177 | case IOR: fputs ("orn", file); break; | |
5178 | case AND: fputs ("andn", file); break; | |
5179 | case XOR: fputs ("xnor", file); break; | |
b3b1e8bd | 5180 | default: output_operand_lossage ("Invalid %%B operand"); |
ab835497 RK |
5181 | } |
5182 | return; | |
5183 | ||
304b7a23 DE |
5184 | /* These are used by the conditional move instructions. */ |
5185 | case 'c' : | |
61a55e8b | 5186 | case 'C': |
304b7a23 DE |
5187 | { |
5188 | enum rtx_code rc = (code == 'c' | |
5189 | ? reverse_condition (GET_CODE (x)) | |
5190 | : GET_CODE (x)); | |
5191 | switch (rc) | |
5192 | { | |
5193 | case NE: fputs ("ne", file); break; | |
5194 | case EQ: fputs ("e", file); break; | |
5195 | case GE: fputs ("ge", file); break; | |
5196 | case GT: fputs ("g", file); break; | |
5197 | case LE: fputs ("le", file); break; | |
5198 | case LT: fputs ("l", file); break; | |
5199 | case GEU: fputs ("geu", file); break; | |
5200 | case GTU: fputs ("gu", file); break; | |
5201 | case LEU: fputs ("leu", file); break; | |
5202 | case LTU: fputs ("lu", file); break; | |
5203 | default: output_operand_lossage (code == 'c' | |
5204 | ? "Invalid %%c operand" | |
5205 | : "Invalid %%C operand"); | |
5206 | } | |
5207 | return; | |
5208 | } | |
5209 | ||
5210 | /* These are used by the movr instruction pattern. */ | |
5211 | case 'd': | |
61a55e8b | 5212 | case 'D': |
304b7a23 DE |
5213 | { |
5214 | enum rtx_code rc = (code == 'd' | |
5215 | ? reverse_condition (GET_CODE (x)) | |
5216 | : GET_CODE (x)); | |
5217 | switch (rc) | |
5218 | { | |
5219 | case NE: fputs ("ne", file); break; | |
5220 | case EQ: fputs ("e", file); break; | |
5221 | case GE: fputs ("gez", file); break; | |
5222 | case LT: fputs ("lz", file); break; | |
5223 | case LE: fputs ("lez", file); break; | |
5224 | case GT: fputs ("gz", file); break; | |
5225 | default: output_operand_lossage (code == 'd' | |
5226 | ? "Invalid %%d operand" | |
5227 | : "Invalid %%D operand"); | |
5228 | } | |
5229 | return; | |
5230 | } | |
61a55e8b | 5231 | |
ab835497 RK |
5232 | case 'b': |
5233 | { | |
5234 | /* Print a sign-extended character. */ | |
5235 | int i = INTVAL (x) & 0xff; | |
5236 | if (i & 0x80) | |
5237 | i |= 0xffffff00; | |
5238 | fprintf (file, "%d", i); | |
5239 | return; | |
5240 | } | |
5241 | ||
d2889939 RK |
5242 | case 'f': |
5243 | /* Operand must be a MEM; write its address. */ | |
5244 | if (GET_CODE (x) != MEM) | |
5245 | output_operand_lossage ("Invalid %%f operand"); | |
5246 | output_address (XEXP (x, 0)); | |
5247 | return; | |
5248 | ||
ab835497 RK |
5249 | case 0: |
5250 | /* Do nothing special. */ | |
5251 | break; | |
5252 | ||
5253 | default: | |
5254 | /* Undocumented flag. */ | |
415f583e | 5255 | output_operand_lossage ("invalid operand output code"); |
ab835497 RK |
5256 | } |
5257 | ||
5258 | if (GET_CODE (x) == REG) | |
5259 | fputs (reg_names[REGNO (x)], file); | |
5260 | else if (GET_CODE (x) == MEM) | |
5261 | { | |
5262 | fputc ('[', file); | |
ab835497 | 5263 | /* Poor Sun assembler doesn't understand absolute addressing. */ |
c4ce6853 DE |
5264 | if (CONSTANT_P (XEXP (x, 0)) |
5265 | && ! TARGET_LIVE_G0) | |
ab835497 RK |
5266 | fputs ("%g0+", file); |
5267 | output_address (XEXP (x, 0)); | |
5268 | fputc (']', file); | |
5269 | } | |
5270 | else if (GET_CODE (x) == HIGH) | |
5271 | { | |
5272 | fputs ("%hi(", file); | |
5273 | output_addr_const (file, XEXP (x, 0)); | |
5274 | fputc (')', file); | |
5275 | } | |
5276 | else if (GET_CODE (x) == LO_SUM) | |
5277 | { | |
5278 | print_operand (file, XEXP (x, 0), 0); | |
e0d80184 DM |
5279 | if (TARGET_CM_MEDMID) |
5280 | fputs ("+%l44(", file); | |
5281 | else | |
5282 | fputs ("+%lo(", file); | |
ab835497 RK |
5283 | output_addr_const (file, XEXP (x, 1)); |
5284 | fputc (')', file); | |
5285 | } | |
e601abce JW |
5286 | else if (GET_CODE (x) == CONST_DOUBLE |
5287 | && (GET_MODE (x) == VOIDmode | |
5288 | || GET_MODE_CLASS (GET_MODE (x)) == MODE_INT)) | |
ab835497 RK |
5289 | { |
5290 | if (CONST_DOUBLE_HIGH (x) == 0) | |
5291 | fprintf (file, "%u", CONST_DOUBLE_LOW (x)); | |
5292 | else if (CONST_DOUBLE_HIGH (x) == -1 | |
5293 | && CONST_DOUBLE_LOW (x) < 0) | |
5294 | fprintf (file, "%d", CONST_DOUBLE_LOW (x)); | |
5295 | else | |
e601abce | 5296 | output_operand_lossage ("long long constant not a valid immediate operand"); |
ab835497 | 5297 | } |
e601abce JW |
5298 | else if (GET_CODE (x) == CONST_DOUBLE) |
5299 | output_operand_lossage ("floating point constant not a valid immediate operand"); | |
ab835497 RK |
5300 | else { output_addr_const (file, x); } |
5301 | } | |
5302 | \f | |
5303 | /* This function outputs assembler code for VALUE to FILE, where VALUE is | |
5304 | a 64 bit (DImode) value. */ | |
5305 | ||
5306 | /* ??? If there is a 64 bit counterpart to .word that the assembler | |
5307 | understands, then using that would simply this code greatly. */ | |
61a55e8b DE |
5308 | /* ??? We only output .xword's for symbols and only then in environments |
5309 | where the assembler can handle them. */ | |
ab835497 RK |
5310 | |
5311 | void | |
5312 | output_double_int (file, value) | |
5313 | FILE *file; | |
5314 | rtx value; | |
5315 | { | |
5316 | if (GET_CODE (value) == CONST_INT) | |
5317 | { | |
f4f8e370 DE |
5318 | /* ??? This has endianness issues. */ |
5319 | #if HOST_BITS_PER_WIDE_INT == 64 | |
5320 | HOST_WIDE_INT xword = INTVAL (value); | |
5321 | HOST_WIDE_INT high, low; | |
5322 | ||
5323 | high = (xword >> 32) & 0xffffffff; | |
5324 | low = xword & 0xffffffff; | |
3a598fbe JL |
5325 | ASM_OUTPUT_INT (file, GEN_INT (high)); |
5326 | ASM_OUTPUT_INT (file, GEN_INT (low)); | |
f4f8e370 | 5327 | #else |
ab835497 RK |
5328 | if (INTVAL (value) < 0) |
5329 | ASM_OUTPUT_INT (file, constm1_rtx); | |
5330 | else | |
5331 | ASM_OUTPUT_INT (file, const0_rtx); | |
5332 | ASM_OUTPUT_INT (file, value); | |
f4f8e370 | 5333 | #endif |
ab835497 RK |
5334 | } |
5335 | else if (GET_CODE (value) == CONST_DOUBLE) | |
5336 | { | |
3a598fbe JL |
5337 | ASM_OUTPUT_INT (file, GEN_INT (CONST_DOUBLE_HIGH (value))); |
5338 | ASM_OUTPUT_INT (file, GEN_INT (CONST_DOUBLE_LOW (value))); | |
ab835497 RK |
5339 | } |
5340 | else if (GET_CODE (value) == SYMBOL_REF | |
5341 | || GET_CODE (value) == CONST | |
61a55e8b | 5342 | || GET_CODE (value) == PLUS |
a9e27770 | 5343 | || (TARGET_ARCH64 && |
61a55e8b | 5344 | (GET_CODE (value) == LABEL_REF |
9fa9626b | 5345 | || GET_CODE (value) == CODE_LABEL |
61a55e8b | 5346 | || GET_CODE (value) == MINUS))) |
ab835497 | 5347 | { |
f3b8847b | 5348 | if (! TARGET_V9) |
61a55e8b DE |
5349 | { |
5350 | ASM_OUTPUT_INT (file, const0_rtx); | |
5351 | ASM_OUTPUT_INT (file, value); | |
5352 | } | |
5353 | else | |
5354 | { | |
5355 | fprintf (file, "\t%s\t", ASM_LONGLONG); | |
5356 | output_addr_const (file, value); | |
5357 | fprintf (file, "\n"); | |
5358 | } | |
ab835497 RK |
5359 | } |
5360 | else | |
5361 | abort (); | |
5362 | } | |
210aa14a | 5363 | \f |
b0468b84 RK |
5364 | /* Return the value of a code used in the .proc pseudo-op that says |
5365 | what kind of result this function returns. For non-C types, we pick | |
5366 | the closest C type. */ | |
5367 | ||
77a02b01 JW |
5368 | #ifndef CHAR_TYPE_SIZE |
5369 | #define CHAR_TYPE_SIZE BITS_PER_UNIT | |
5370 | #endif | |
5371 | ||
5372 | #ifndef SHORT_TYPE_SIZE | |
5373 | #define SHORT_TYPE_SIZE (BITS_PER_UNIT * 2) | |
5374 | #endif | |
5375 | ||
5376 | #ifndef INT_TYPE_SIZE | |
5377 | #define INT_TYPE_SIZE BITS_PER_WORD | |
5378 | #endif | |
5379 | ||
5380 | #ifndef LONG_TYPE_SIZE | |
5381 | #define LONG_TYPE_SIZE BITS_PER_WORD | |
5382 | #endif | |
5383 | ||
5384 | #ifndef LONG_LONG_TYPE_SIZE | |
5385 | #define LONG_LONG_TYPE_SIZE (BITS_PER_WORD * 2) | |
5386 | #endif | |
5387 | ||
5388 | #ifndef FLOAT_TYPE_SIZE | |
5389 | #define FLOAT_TYPE_SIZE BITS_PER_WORD | |
5390 | #endif | |
5391 | ||
5392 | #ifndef DOUBLE_TYPE_SIZE | |
5393 | #define DOUBLE_TYPE_SIZE (BITS_PER_WORD * 2) | |
5394 | #endif | |
5395 | ||
5396 | #ifndef LONG_DOUBLE_TYPE_SIZE | |
5397 | #define LONG_DOUBLE_TYPE_SIZE (BITS_PER_WORD * 2) | |
5398 | #endif | |
210aa14a RS |
5399 | |
5400 | unsigned long | |
5401 | sparc_type_code (type) | |
5402 | register tree type; | |
5403 | { | |
5404 | register unsigned long qualifiers = 0; | |
e5e809f4 | 5405 | register unsigned shift; |
210aa14a | 5406 | |
abc95ed3 | 5407 | /* Only the first 30 bits of the qualifier are valid. We must refrain from |
aee2c3c5 JW |
5408 | setting more, since some assemblers will give an error for this. Also, |
5409 | we must be careful to avoid shifts of 32 bits or more to avoid getting | |
5410 | unpredictable results. */ | |
5411 | ||
e5e809f4 | 5412 | for (shift = 6; shift < 30; shift += 2, type = TREE_TYPE (type)) |
210aa14a RS |
5413 | { |
5414 | switch (TREE_CODE (type)) | |
5415 | { | |
5416 | case ERROR_MARK: | |
5417 | return qualifiers; | |
5418 | ||
5419 | case ARRAY_TYPE: | |
e5e809f4 | 5420 | qualifiers |= (3 << shift); |
210aa14a RS |
5421 | break; |
5422 | ||
5423 | case FUNCTION_TYPE: | |
5424 | case METHOD_TYPE: | |
e5e809f4 | 5425 | qualifiers |= (2 << shift); |
210aa14a RS |
5426 | break; |
5427 | ||
5428 | case POINTER_TYPE: | |
5429 | case REFERENCE_TYPE: | |
5430 | case OFFSET_TYPE: | |
e5e809f4 | 5431 | qualifiers |= (1 << shift); |
210aa14a | 5432 | break; |
ab835497 | 5433 | |
210aa14a RS |
5434 | case RECORD_TYPE: |
5435 | return (qualifiers | 8); | |
5436 | ||
5437 | case UNION_TYPE: | |
b0468b84 | 5438 | case QUAL_UNION_TYPE: |
210aa14a RS |
5439 | return (qualifiers | 9); |
5440 | ||
5441 | case ENUMERAL_TYPE: | |
5442 | return (qualifiers | 10); | |
5443 | ||
5444 | case VOID_TYPE: | |
5445 | return (qualifiers | 16); | |
5446 | ||
5447 | case INTEGER_TYPE: | |
654209e6 JW |
5448 | /* If this is a range type, consider it to be the underlying |
5449 | type. */ | |
5450 | if (TREE_TYPE (type) != 0) | |
e5e809f4 | 5451 | break; |
654209e6 | 5452 | |
77a02b01 | 5453 | /* Carefully distinguish all the standard types of C, |
b0468b84 RK |
5454 | without messing up if the language is not C. We do this by |
5455 | testing TYPE_PRECISION and TREE_UNSIGNED. The old code used to | |
5456 | look at both the names and the above fields, but that's redundant. | |
5457 | Any type whose size is between two C types will be considered | |
5458 | to be the wider of the two types. Also, we do not have a | |
5459 | special code to use for "long long", so anything wider than | |
5460 | long is treated the same. Note that we can't distinguish | |
5461 | between "int" and "long" in this code if they are the same | |
5462 | size, but that's fine, since neither can the assembler. */ | |
5463 | ||
5464 | if (TYPE_PRECISION (type) <= CHAR_TYPE_SIZE) | |
5465 | return (qualifiers | (TREE_UNSIGNED (type) ? 12 : 2)); | |
77a02b01 | 5466 | |
b0468b84 RK |
5467 | else if (TYPE_PRECISION (type) <= SHORT_TYPE_SIZE) |
5468 | return (qualifiers | (TREE_UNSIGNED (type) ? 13 : 3)); | |
77a02b01 | 5469 | |
b0468b84 | 5470 | else if (TYPE_PRECISION (type) <= INT_TYPE_SIZE) |
77a02b01 JW |
5471 | return (qualifiers | (TREE_UNSIGNED (type) ? 14 : 4)); |
5472 | ||
b0468b84 | 5473 | else |
77a02b01 JW |
5474 | return (qualifiers | (TREE_UNSIGNED (type) ? 15 : 5)); |
5475 | ||
210aa14a | 5476 | case REAL_TYPE: |
e5e809f4 JL |
5477 | /* If this is a range type, consider it to be the underlying |
5478 | type. */ | |
5479 | if (TREE_TYPE (type) != 0) | |
5480 | break; | |
5481 | ||
77a02b01 JW |
5482 | /* Carefully distinguish all the standard types of C, |
5483 | without messing up if the language is not C. */ | |
b0468b84 | 5484 | |
77a02b01 | 5485 | if (TYPE_PRECISION (type) == FLOAT_TYPE_SIZE) |
210aa14a | 5486 | return (qualifiers | 6); |
b0468b84 RK |
5487 | |
5488 | else | |
5489 | return (qualifiers | 7); | |
210aa14a RS |
5490 | |
5491 | case COMPLEX_TYPE: /* GNU Fortran COMPLEX type. */ | |
13d39dbc | 5492 | /* ??? We need to distinguish between double and float complex types, |
c82aa69a JW |
5493 | but I don't know how yet because I can't reach this code from |
5494 | existing front-ends. */ | |
5495 | return (qualifiers | 7); /* Who knows? */ | |
5496 | ||
210aa14a RS |
5497 | case CHAR_TYPE: /* GNU Pascal CHAR type. Not used in C. */ |
5498 | case BOOLEAN_TYPE: /* GNU Fortran BOOLEAN type. */ | |
5499 | case FILE_TYPE: /* GNU Pascal FILE type. */ | |
26902ae0 | 5500 | case SET_TYPE: /* GNU Pascal SET type. */ |
210aa14a | 5501 | case LANG_TYPE: /* ? */ |
26902ae0 | 5502 | return qualifiers; |
210aa14a RS |
5503 | |
5504 | default: | |
5505 | abort (); /* Not a type! */ | |
5506 | } | |
5507 | } | |
e5e809f4 JL |
5508 | |
5509 | return qualifiers; | |
210aa14a | 5510 | } |
ead69aea | 5511 | \f |
61a55e8b DE |
5512 | /* Nested function support. */ |
5513 | ||
5514 | /* Emit RTL insns to initialize the variable parts of a trampoline. | |
5515 | FNADDR is an RTX for the address of the function's pure code. | |
5516 | CXT is an RTX for the static chain value for the function. | |
5517 | ||
5518 | This takes 16 insns: 2 shifts & 2 ands (to split up addresses), 4 sethi | |
5519 | (to load in opcodes), 4 iors (to merge address and opcodes), and 4 writes | |
5520 | (to store insns). This is a bit excessive. Perhaps a different | |
5521 | mechanism would be better here. | |
5522 | ||
849a528d | 5523 | Emit enough FLUSH insns to synchronize the data and instruction caches. */ |
61a55e8b DE |
5524 | |
5525 | void | |
5526 | sparc_initialize_trampoline (tramp, fnaddr, cxt) | |
5527 | rtx tramp, fnaddr, cxt; | |
5528 | { | |
c6b0465b JC |
5529 | /* SPARC 32 bit trampoline: |
5530 | ||
e0d80184 DM |
5531 | sethi %hi(fn), %g1 |
5532 | sethi %hi(static), %g2 | |
5533 | jmp %g1+%lo(fn) | |
5534 | or %g2, %lo(static), %g2 | |
c6b0465b JC |
5535 | |
5536 | SETHI i,r = 00rr rrr1 00ii iiii iiii iiii iiii iiii | |
5537 | JMPL r+i,d = 10dd ddd1 1100 0rrr rr1i iiii iiii iiii | |
5538 | */ | |
959eb758 DM |
5539 | #ifdef TRANSFER_FROM_TRAMPOLINE |
5540 | emit_library_call (gen_rtx (SYMBOL_REF, Pmode, "__enable_execute_stack"), | |
5541 | 0, VOIDmode, 1, tramp, Pmode); | |
5542 | #endif | |
c6b0465b JC |
5543 | |
5544 | emit_move_insn (gen_rtx_MEM (SImode, plus_constant (tramp, 0)), | |
5545 | expand_binop (SImode, ior_optab, | |
5546 | expand_shift (RSHIFT_EXPR, SImode, fnaddr, | |
5547 | size_int (10), 0, 1), | |
5548 | GEN_INT (0x03000000), | |
5549 | NULL_RTX, 1, OPTAB_DIRECT)); | |
5550 | ||
5551 | emit_move_insn (gen_rtx_MEM (SImode, plus_constant (tramp, 4)), | |
5552 | expand_binop (SImode, ior_optab, | |
5553 | expand_shift (RSHIFT_EXPR, SImode, cxt, | |
5554 | size_int (10), 0, 1), | |
5555 | GEN_INT (0x05000000), | |
5556 | NULL_RTX, 1, OPTAB_DIRECT)); | |
5557 | ||
5558 | emit_move_insn (gen_rtx_MEM (SImode, plus_constant (tramp, 8)), | |
5559 | expand_binop (SImode, ior_optab, | |
5560 | expand_and (fnaddr, GEN_INT (0x3ff), NULL_RTX), | |
5561 | GEN_INT (0x81c06000), | |
5562 | NULL_RTX, 1, OPTAB_DIRECT)); | |
5563 | ||
5564 | emit_move_insn (gen_rtx_MEM (SImode, plus_constant (tramp, 12)), | |
5565 | expand_binop (SImode, ior_optab, | |
5566 | expand_and (cxt, GEN_INT (0x3ff), NULL_RTX), | |
5567 | GEN_INT (0x8410a000), | |
5568 | NULL_RTX, 1, OPTAB_DIRECT)); | |
5569 | ||
254110c2 | 5570 | emit_insn (gen_flush (validize_mem (gen_rtx_MEM (SImode, tramp)))); |
c6b0465b JC |
5571 | /* On UltraSPARC a flush flushes an entire cache line. The trampoline is |
5572 | aligned on a 16 byte boundary so one flush clears it all. */ | |
5573 | if (sparc_cpu != PROCESSOR_ULTRASPARC) | |
5574 | emit_insn (gen_flush (validize_mem (gen_rtx_MEM (SImode, | |
5575 | plus_constant (tramp, 8))))); | |
61a55e8b DE |
5576 | } |
5577 | ||
849a528d RK |
5578 | /* The 64 bit version is simpler because it makes more sense to load the |
5579 | values as "immediate" data out of the trampoline. It's also easier since | |
5580 | we can read the PC without clobbering a register. */ | |
5581 | ||
61a55e8b DE |
5582 | void |
5583 | sparc64_initialize_trampoline (tramp, fnaddr, cxt) | |
5584 | rtx tramp, fnaddr, cxt; | |
5585 | { | |
345a6161 DM |
5586 | #ifdef TRANSFER_FROM_TRAMPOLINE |
5587 | emit_library_call (gen_rtx (SYMBOL_REF, Pmode, "__enable_execute_stack"), | |
5588 | 0, VOIDmode, 1, tramp, Pmode); | |
5589 | #endif | |
5590 | ||
c6b0465b | 5591 | /* |
e0d80184 DM |
5592 | rd %pc, %g1 |
5593 | ldx [%g1+24], %g5 | |
5594 | jmp %g5 | |
5595 | ldx [%g1+16], %g5 | |
c6b0465b JC |
5596 | +16 bytes data |
5597 | */ | |
5598 | ||
5599 | emit_move_insn (gen_rtx_MEM (SImode, tramp), | |
5600 | GEN_INT (0x83414000)); | |
5601 | emit_move_insn (gen_rtx_MEM (SImode, plus_constant (tramp, 4)), | |
5602 | GEN_INT (0xca586018)); | |
5603 | emit_move_insn (gen_rtx_MEM (SImode, plus_constant (tramp, 8)), | |
345a6161 | 5604 | GEN_INT (0x81c14000)); |
c6b0465b JC |
5605 | emit_move_insn (gen_rtx_MEM (SImode, plus_constant (tramp, 12)), |
5606 | GEN_INT (0xca586010)); | |
5607 | emit_move_insn (gen_rtx_MEM (DImode, plus_constant (tramp, 16)), cxt); | |
345a6161 | 5608 | emit_move_insn (gen_rtx_MEM (DImode, plus_constant (tramp, 24)), fnaddr); |
254110c2 | 5609 | emit_insn (gen_flush (validize_mem (gen_rtx_MEM (DImode, tramp)))); |
345a6161 | 5610 | |
c6b0465b JC |
5611 | if (sparc_cpu != PROCESSOR_ULTRASPARC) |
5612 | emit_insn (gen_flush (validize_mem (gen_rtx_MEM (DImode, plus_constant (tramp, 8))))); | |
61a55e8b DE |
5613 | } |
5614 | \f | |
c819be5b JW |
5615 | /* Subroutines to support a flat (single) register window calling |
5616 | convention. */ | |
5617 | ||
5618 | /* Single-register window sparc stack frames look like: | |
5619 | ||
5620 | Before call After call | |
5621 | +-----------------------+ +-----------------------+ | |
86735b02 DE |
5622 | high | | | | |
5623 | mem | caller's temps. | | caller's temps. | | |
c819be5b JW |
5624 | | | | | |
5625 | +-----------------------+ +-----------------------+ | |
5626 | | | | | | |
5627 | | arguments on stack. | | arguments on stack. | | |
86735b02 DE |
5628 | | | | | |
5629 | +-----------------------+FP+92->+-----------------------+ | |
c819be5b JW |
5630 | | 6 words to save | | 6 words to save | |
5631 | | arguments passed | | arguments passed | | |
5632 | | in registers, even | | in registers, even | | |
86735b02 DE |
5633 | | if not passed. | | if not passed. | |
5634 | SP+68->+-----------------------+FP+68->+-----------------------+ | |
5635 | | 1 word struct addr | | 1 word struct addr | | |
5636 | +-----------------------+FP+64->+-----------------------+ | |
5637 | | | | | | |
5638 | | 16 word reg save area | | 16 word reg save area | | |
5639 | | | | | | |
5640 | SP->+-----------------------+ FP->+-----------------------+ | |
5641 | | 4 word area for | | |
5642 | | fp/alu reg moves | | |
5643 | FP-16->+-----------------------+ | |
5644 | | | | |
5645 | | local variables | | |
5646 | | | | |
5647 | +-----------------------+ | |
5648 | | | | |
c819be5b | 5649 | | fp register save | |
86735b02 DE |
5650 | | | |
5651 | +-----------------------+ | |
5652 | | | | |
c819be5b JW |
5653 | | gp register save | |
5654 | | | | |
86735b02 DE |
5655 | +-----------------------+ |
5656 | | | | |
c819be5b | 5657 | | alloca allocations | |
86735b02 DE |
5658 | | | |
5659 | +-----------------------+ | |
5660 | | | | |
c819be5b | 5661 | | arguments on stack | |
86735b02 DE |
5662 | | | |
5663 | SP+92->+-----------------------+ | |
c819be5b | 5664 | | 6 words to save | |
86735b02 | 5665 | | arguments passed | |
c819be5b | 5666 | | in registers, even | |
86735b02 DE |
5667 | low | if not passed. | |
5668 | memory SP+68->+-----------------------+ | |
5669 | | 1 word struct addr | | |
5670 | SP+64->+-----------------------+ | |
5671 | | | | |
5672 | I 16 word reg save area | | |
5673 | | | | |
5674 | SP->+-----------------------+ */ | |
c819be5b | 5675 | |
c6aa9ce1 | 5676 | /* Structure to be filled in by sparc_flat_compute_frame_size with register |
c819be5b JW |
5677 | save masks, and offsets for the current function. */ |
5678 | ||
5679 | struct sparc_frame_info | |
5680 | { | |
5681 | unsigned long total_size; /* # bytes that the entire frame takes up. */ | |
5682 | unsigned long var_size; /* # bytes that variables take up. */ | |
5683 | unsigned long args_size; /* # bytes that outgoing arguments take up. */ | |
5684 | unsigned long extra_size; /* # bytes of extra gunk. */ | |
5685 | unsigned int gp_reg_size; /* # bytes needed to store gp regs. */ | |
5686 | unsigned int fp_reg_size; /* # bytes needed to store fp regs. */ | |
86735b02 | 5687 | unsigned long gmask; /* Mask of saved gp registers. */ |
c819be5b | 5688 | unsigned long fmask; /* Mask of saved fp registers. */ |
86735b02 | 5689 | unsigned long reg_offset; /* Offset from new sp to store regs. */ |
c819be5b JW |
5690 | int initialized; /* Nonzero if frame size already calculated. */ |
5691 | }; | |
5692 | ||
c6aa9ce1 | 5693 | /* Current frame information calculated by sparc_flat_compute_frame_size. */ |
c819be5b JW |
5694 | struct sparc_frame_info current_frame_info; |
5695 | ||
5696 | /* Zero structure to initialize current_frame_info. */ | |
5697 | struct sparc_frame_info zero_frame_info; | |
5698 | ||
5699 | /* Tell prologue and epilogue if register REGNO should be saved / restored. */ | |
5700 | ||
86735b02 DE |
5701 | #define RETURN_ADDR_REGNUM 15 |
5702 | #define FRAME_POINTER_MASK (1 << (FRAME_POINTER_REGNUM)) | |
5703 | #define RETURN_ADDR_MASK (1 << (RETURN_ADDR_REGNUM)) | |
5704 | ||
c819be5b JW |
5705 | #define MUST_SAVE_REGISTER(regno) \ |
5706 | ((regs_ever_live[regno] && !call_used_regs[regno]) \ | |
5707 | || (regno == FRAME_POINTER_REGNUM && frame_pointer_needed) \ | |
86735b02 | 5708 | || (regno == RETURN_ADDR_REGNUM && regs_ever_live[RETURN_ADDR_REGNUM])) |
c819be5b JW |
5709 | |
5710 | /* Return the bytes needed to compute the frame pointer from the current | |
5711 | stack pointer. */ | |
5712 | ||
5713 | unsigned long | |
c6aa9ce1 | 5714 | sparc_flat_compute_frame_size (size) |
c819be5b JW |
5715 | int size; /* # of var. bytes allocated. */ |
5716 | { | |
5717 | int regno; | |
5718 | unsigned long total_size; /* # bytes that the entire frame takes up. */ | |
5719 | unsigned long var_size; /* # bytes that variables take up. */ | |
5720 | unsigned long args_size; /* # bytes that outgoing arguments take up. */ | |
5721 | unsigned long extra_size; /* # extra bytes. */ | |
5722 | unsigned int gp_reg_size; /* # bytes needed to store gp regs. */ | |
5723 | unsigned int fp_reg_size; /* # bytes needed to store fp regs. */ | |
86735b02 | 5724 | unsigned long gmask; /* Mask of saved gp registers. */ |
c819be5b | 5725 | unsigned long fmask; /* Mask of saved fp registers. */ |
86735b02 DE |
5726 | unsigned long reg_offset; /* Offset to register save area. */ |
5727 | int need_aligned_p; /* 1 if need the save area 8 byte aligned. */ | |
c819be5b JW |
5728 | |
5729 | /* This is the size of the 16 word reg save area, 1 word struct addr | |
5730 | area, and 4 word fp/alu register copy area. */ | |
8c64be75 SC |
5731 | extra_size = -STARTING_FRAME_OFFSET + FIRST_PARM_OFFSET(0); |
5732 | var_size = size; | |
5733 | gp_reg_size = 0; | |
5734 | fp_reg_size = 0; | |
5735 | gmask = 0; | |
5736 | fmask = 0; | |
5737 | reg_offset = 0; | |
86735b02 | 5738 | need_aligned_p = 0; |
c819be5b | 5739 | |
8c64be75 SC |
5740 | args_size = 0; |
5741 | if (!leaf_function_p ()) | |
5742 | { | |
5743 | /* Also include the size needed for the 6 parameter registers. */ | |
5744 | args_size = current_function_outgoing_args_size + 24; | |
5745 | } | |
5746 | total_size = var_size + args_size; | |
5747 | ||
c819be5b JW |
5748 | /* Calculate space needed for gp registers. */ |
5749 | for (regno = 1; regno <= 31; regno++) | |
5750 | { | |
5751 | if (MUST_SAVE_REGISTER (regno)) | |
5752 | { | |
86735b02 DE |
5753 | /* If we need to save two regs in a row, ensure there's room to bump |
5754 | up the address to align it to a doubleword boundary. */ | |
c819be5b JW |
5755 | if ((regno & 0x1) == 0 && MUST_SAVE_REGISTER (regno+1)) |
5756 | { | |
5757 | if (gp_reg_size % 8 != 0) | |
86735b02 | 5758 | gp_reg_size += 4; |
c819be5b | 5759 | gp_reg_size += 2 * UNITS_PER_WORD; |
86735b02 | 5760 | gmask |= 3 << regno; |
c819be5b | 5761 | regno++; |
86735b02 | 5762 | need_aligned_p = 1; |
c819be5b JW |
5763 | } |
5764 | else | |
5765 | { | |
5766 | gp_reg_size += UNITS_PER_WORD; | |
86735b02 | 5767 | gmask |= 1 << regno; |
c819be5b JW |
5768 | } |
5769 | } | |
5770 | } | |
c819be5b JW |
5771 | |
5772 | /* Calculate space needed for fp registers. */ | |
5773 | for (regno = 32; regno <= 63; regno++) | |
5774 | { | |
5775 | if (regs_ever_live[regno] && !call_used_regs[regno]) | |
5776 | { | |
5777 | fp_reg_size += UNITS_PER_WORD; | |
5778 | fmask |= 1 << (regno - 32); | |
5779 | } | |
5780 | } | |
5781 | ||
86735b02 DE |
5782 | if (gmask || fmask) |
5783 | { | |
5784 | int n; | |
5785 | reg_offset = FIRST_PARM_OFFSET(0) + args_size; | |
5786 | /* Ensure save area is 8 byte aligned if we need it. */ | |
5787 | n = reg_offset % 8; | |
5788 | if (need_aligned_p && n != 0) | |
5789 | { | |
5790 | total_size += 8 - n; | |
5791 | reg_offset += 8 - n; | |
5792 | } | |
5793 | total_size += gp_reg_size + fp_reg_size; | |
5794 | } | |
c819be5b | 5795 | |
8c64be75 SC |
5796 | /* If we must allocate a stack frame at all, we must also allocate |
5797 | room for register window spillage, so as to be binary compatible | |
5798 | with libraries and operating systems that do not use -mflat. */ | |
5799 | if (total_size > 0) | |
5800 | total_size += extra_size; | |
5801 | else | |
5802 | extra_size = 0; | |
c819be5b JW |
5803 | |
5804 | total_size = SPARC_STACK_ALIGN (total_size); | |
5805 | ||
5806 | /* Save other computed information. */ | |
5807 | current_frame_info.total_size = total_size; | |
5808 | current_frame_info.var_size = var_size; | |
5809 | current_frame_info.args_size = args_size; | |
5810 | current_frame_info.extra_size = extra_size; | |
5811 | current_frame_info.gp_reg_size = gp_reg_size; | |
5812 | current_frame_info.fp_reg_size = fp_reg_size; | |
86735b02 | 5813 | current_frame_info.gmask = gmask; |
c819be5b | 5814 | current_frame_info.fmask = fmask; |
86735b02 | 5815 | current_frame_info.reg_offset = reg_offset; |
c819be5b JW |
5816 | current_frame_info.initialized = reload_completed; |
5817 | ||
c819be5b JW |
5818 | /* Ok, we're done. */ |
5819 | return total_size; | |
5820 | } | |
5821 | \f | |
86735b02 DE |
5822 | /* Save/restore registers in GMASK and FMASK at register BASE_REG plus offset |
5823 | OFFSET. | |
5824 | ||
5825 | BASE_REG must be 8 byte aligned. This allows us to test OFFSET for | |
5826 | appropriate alignment and use DOUBLEWORD_OP when we can. We assume | |
5827 | [BASE_REG+OFFSET] will always be a valid address. | |
5828 | ||
5829 | WORD_OP is either "st" for save, "ld" for restore. | |
5830 | DOUBLEWORD_OP is either "std" for save, "ldd" for restore. */ | |
c819be5b JW |
5831 | |
5832 | void | |
96987e96 DE |
5833 | sparc_flat_save_restore (file, base_reg, offset, gmask, fmask, word_op, |
5834 | doubleword_op, base_offset) | |
86735b02 | 5835 | FILE *file; |
b1474bb7 | 5836 | const char *base_reg; |
86735b02 DE |
5837 | unsigned int offset; |
5838 | unsigned long gmask; | |
5839 | unsigned long fmask; | |
b1474bb7 KG |
5840 | const char *word_op; |
5841 | const char *doubleword_op; | |
c53aa195 | 5842 | unsigned long base_offset; |
c819be5b JW |
5843 | { |
5844 | int regno; | |
c819be5b | 5845 | |
86735b02 DE |
5846 | if (gmask == 0 && fmask == 0) |
5847 | return; | |
c819be5b | 5848 | |
86735b02 DE |
5849 | /* Save registers starting from high to low. We've already saved the |
5850 | previous frame pointer and previous return address for the debugger's | |
5851 | sake. The debugger allows us to not need a nop in the epilog if at least | |
5852 | one register is reloaded in addition to return address. */ | |
c819be5b | 5853 | |
86735b02 | 5854 | if (gmask) |
c819be5b JW |
5855 | { |
5856 | for (regno = 1; regno <= 31; regno++) | |
5857 | { | |
86735b02 | 5858 | if ((gmask & (1L << regno)) != 0) |
c819be5b | 5859 | { |
86735b02 | 5860 | if ((regno & 0x1) == 0 && ((gmask & (1L << (regno+1))) != 0)) |
c819be5b | 5861 | { |
86735b02 DE |
5862 | /* We can save two registers in a row. If we're not at a |
5863 | double word boundary, move to one. | |
5864 | sparc_flat_compute_frame_size ensures there's room to do | |
5865 | this. */ | |
5866 | if (offset % 8 != 0) | |
5867 | offset += UNITS_PER_WORD; | |
5868 | ||
c819be5b | 5869 | if (word_op[0] == 's') |
c53aa195 | 5870 | { |
e0d80184 | 5871 | fprintf (file, "\t%s\t%s, [%s+%d]\n", |
c53aa195 JM |
5872 | doubleword_op, reg_names[regno], |
5873 | base_reg, offset); | |
0021b564 | 5874 | if (dwarf2out_do_frame ()) |
c53aa195 | 5875 | { |
96987e96 | 5876 | char *l = dwarf2out_cfi_label (); |
c53aa195 JM |
5877 | dwarf2out_reg_save (l, regno, offset + base_offset); |
5878 | dwarf2out_reg_save | |
5879 | (l, regno+1, offset+base_offset + UNITS_PER_WORD); | |
5880 | } | |
c53aa195 | 5881 | } |
c819be5b | 5882 | else |
e0d80184 | 5883 | fprintf (file, "\t%s\t[%s+%d], %s\n", |
86735b02 | 5884 | doubleword_op, base_reg, offset, |
c819be5b JW |
5885 | reg_names[regno]); |
5886 | ||
86735b02 | 5887 | offset += 2 * UNITS_PER_WORD; |
c819be5b JW |
5888 | regno++; |
5889 | } | |
5890 | else | |
5891 | { | |
5892 | if (word_op[0] == 's') | |
c53aa195 | 5893 | { |
e0d80184 | 5894 | fprintf (file, "\t%s\t%s, [%s+%d]\n", |
c53aa195 JM |
5895 | word_op, reg_names[regno], |
5896 | base_reg, offset); | |
0021b564 | 5897 | if (dwarf2out_do_frame ()) |
c53aa195 | 5898 | dwarf2out_reg_save ("", regno, offset + base_offset); |
c53aa195 | 5899 | } |
c819be5b | 5900 | else |
e0d80184 | 5901 | fprintf (file, "\t%s\t[%s+%d], %s\n", |
86735b02 | 5902 | word_op, base_reg, offset, reg_names[regno]); |
c819be5b | 5903 | |
86735b02 | 5904 | offset += UNITS_PER_WORD; |
c819be5b JW |
5905 | } |
5906 | } | |
5907 | } | |
5908 | } | |
5909 | ||
5910 | if (fmask) | |
5911 | { | |
5912 | for (regno = 32; regno <= 63; regno++) | |
5913 | { | |
5914 | if ((fmask & (1L << (regno - 32))) != 0) | |
5915 | { | |
5916 | if (word_op[0] == 's') | |
c53aa195 | 5917 | { |
e0d80184 | 5918 | fprintf (file, "\t%s\t%s, [%s+%d]\n", |
c53aa195 JM |
5919 | word_op, reg_names[regno], |
5920 | base_reg, offset); | |
0021b564 | 5921 | if (dwarf2out_do_frame ()) |
c53aa195 | 5922 | dwarf2out_reg_save ("", regno, offset + base_offset); |
c53aa195 | 5923 | } |
c819be5b | 5924 | else |
e0d80184 | 5925 | fprintf (file, "\t%s\t[%s+%d], %s\n", |
86735b02 | 5926 | word_op, base_reg, offset, reg_names[regno]); |
c819be5b | 5927 | |
86735b02 | 5928 | offset += UNITS_PER_WORD; |
c819be5b JW |
5929 | } |
5930 | } | |
5931 | } | |
5932 | } | |
5933 | \f | |
5934 | /* Set up the stack and frame (if desired) for the function. */ | |
5935 | ||
5936 | void | |
c6aa9ce1 | 5937 | sparc_flat_output_function_prologue (file, size) |
c819be5b JW |
5938 | FILE *file; |
5939 | int size; | |
5940 | { | |
c819be5b | 5941 | char *sp_str = reg_names[STACK_POINTER_REGNUM]; |
86735b02 | 5942 | unsigned long gmask = current_frame_info.gmask; |
a07c1915 | 5943 | |
1cb36a98 RH |
5944 | sparc_output_scratch_registers (file); |
5945 | ||
c6aa9ce1 | 5946 | /* This is only for the human reader. */ |
bf62bbf1 | 5947 | fprintf (file, "\t%s#PROLOGUE# 0\n", ASM_COMMENT_START); |
5e9defae | 5948 | fprintf (file, "\t%s# vars= %ld, regs= %d/%d, args= %d, extra= %ld\n", |
bf62bbf1 | 5949 | ASM_COMMENT_START, |
86735b02 DE |
5950 | current_frame_info.var_size, |
5951 | current_frame_info.gp_reg_size / 4, | |
9f8b85b8 | 5952 | current_frame_info.fp_reg_size / 4, |
86735b02 DE |
5953 | current_function_outgoing_args_size, |
5954 | current_frame_info.extra_size); | |
c819be5b JW |
5955 | |
5956 | size = SPARC_STACK_ALIGN (size); | |
86735b02 DE |
5957 | size = (! current_frame_info.initialized |
5958 | ? sparc_flat_compute_frame_size (size) | |
5959 | : current_frame_info.total_size); | |
c819be5b | 5960 | |
86735b02 DE |
5961 | /* These cases shouldn't happen. Catch them now. */ |
5962 | if (size == 0 && (gmask || current_frame_info.fmask)) | |
5963 | abort (); | |
c819be5b | 5964 | |
86735b02 DE |
5965 | /* Allocate our stack frame by decrementing %sp. |
5966 | At present, the only algorithm gdb can use to determine if this is a | |
5967 | flat frame is if we always set %i7 if we set %sp. This can be optimized | |
5968 | in the future by putting in some sort of debugging information that says | |
5969 | this is a `flat' function. However, there is still the case of debugging | |
5970 | code without such debugging information (including cases where most fns | |
5971 | have such info, but there is one that doesn't). So, always do this now | |
5972 | so we don't get a lot of code out there that gdb can't handle. | |
5973 | If the frame pointer isn't needn't then that's ok - gdb won't be able to | |
5974 | distinguish us from a non-flat function but there won't (and shouldn't) | |
5975 | be any differences anyway. The return pc is saved (if necessary) right | |
5976 | after %i7 so gdb won't have to look too far to find it. */ | |
5977 | if (size > 0) | |
c819be5b | 5978 | { |
86735b02 | 5979 | unsigned int reg_offset = current_frame_info.reg_offset; |
c6aa9ce1 | 5980 | char *fp_str = reg_names[FRAME_POINTER_REGNUM]; |
3bb5de61 | 5981 | const char *t1_str = "%g1"; |
86735b02 DE |
5982 | |
5983 | /* Things get a little tricky if local variables take up more than ~4096 | |
5984 | bytes and outgoing arguments take up more than ~4096 bytes. When that | |
5985 | happens, the register save area can't be accessed from either end of | |
5986 | the frame. Handle this by decrementing %sp to the start of the gp | |
5987 | register save area, save the regs, update %i7, and then set %sp to its | |
5988 | final value. Given that we only have one scratch register to play | |
5989 | with it is the cheapest solution, and it helps gdb out as it won't | |
5990 | slow down recognition of flat functions. | |
5991 | Don't change the order of insns emitted here without checking with | |
5992 | the gdb folk first. */ | |
5993 | ||
ddd5a7c1 | 5994 | /* Is the entire register save area offsettable from %sp? */ |
86735b02 DE |
5995 | if (reg_offset < 4096 - 64 * UNITS_PER_WORD) |
5996 | { | |
5997 | if (size <= 4096) | |
5998 | { | |
e0d80184 | 5999 | fprintf (file, "\tadd\t%s, %d, %s\n", |
86735b02 DE |
6000 | sp_str, -size, sp_str); |
6001 | if (gmask & FRAME_POINTER_MASK) | |
6002 | { | |
e0d80184 | 6003 | fprintf (file, "\tst\t%s, [%s+%d]\n", |
86735b02 | 6004 | fp_str, sp_str, reg_offset); |
e0d80184 | 6005 | fprintf (file, "\tsub\t%s, %d, %s\t%s# set up frame pointer\n", |
bf62bbf1 | 6006 | sp_str, -size, fp_str, ASM_COMMENT_START); |
86735b02 DE |
6007 | reg_offset += 4; |
6008 | } | |
6009 | } | |
6010 | else | |
6011 | { | |
e0d80184 | 6012 | fprintf (file, "\tset\t%d, %s\n\tsub\t%s, %s, %s\n", |
86735b02 DE |
6013 | size, t1_str, sp_str, t1_str, sp_str); |
6014 | if (gmask & FRAME_POINTER_MASK) | |
6015 | { | |
e0d80184 | 6016 | fprintf (file, "\tst\t%s, [%s+%d]\n", |
86735b02 | 6017 | fp_str, sp_str, reg_offset); |
e0d80184 | 6018 | fprintf (file, "\tadd\t%s, %s, %s\t%s# set up frame pointer\n", |
bf62bbf1 | 6019 | sp_str, t1_str, fp_str, ASM_COMMENT_START); |
86735b02 DE |
6020 | reg_offset += 4; |
6021 | } | |
6022 | } | |
0021b564 | 6023 | if (dwarf2out_do_frame ()) |
c53aa195 | 6024 | { |
96987e96 | 6025 | char *l = dwarf2out_cfi_label (); |
c53aa195 JM |
6026 | if (gmask & FRAME_POINTER_MASK) |
6027 | { | |
6028 | dwarf2out_reg_save (l, FRAME_POINTER_REGNUM, | |
6029 | reg_offset - 4 - size); | |
6030 | dwarf2out_def_cfa (l, FRAME_POINTER_REGNUM, 0); | |
6031 | } | |
6032 | else | |
6033 | dwarf2out_def_cfa (l, STACK_POINTER_REGNUM, size); | |
6034 | } | |
86735b02 DE |
6035 | if (gmask & RETURN_ADDR_MASK) |
6036 | { | |
e0d80184 | 6037 | fprintf (file, "\tst\t%s, [%s+%d]\n", |
86735b02 | 6038 | reg_names[RETURN_ADDR_REGNUM], sp_str, reg_offset); |
0021b564 | 6039 | if (dwarf2out_do_frame ()) |
c53aa195 | 6040 | dwarf2out_return_save ("", reg_offset - size); |
86735b02 DE |
6041 | reg_offset += 4; |
6042 | } | |
6043 | sparc_flat_save_restore (file, sp_str, reg_offset, | |
6044 | gmask & ~(FRAME_POINTER_MASK | RETURN_ADDR_MASK), | |
6045 | current_frame_info.fmask, | |
84e884e9 | 6046 | "st", "std", -size); |
86735b02 | 6047 | } |
c819be5b | 6048 | else |
86735b02 DE |
6049 | { |
6050 | /* Subtract %sp in two steps, but make sure there is always a | |
6051 | 64 byte register save area, and %sp is properly aligned. */ | |
6052 | /* Amount to decrement %sp by, the first time. */ | |
6053 | unsigned int size1 = ((size - reg_offset + 64) + 15) & -16; | |
6054 | /* Offset to register save area from %sp. */ | |
6055 | unsigned int offset = size1 - (size - reg_offset); | |
6056 | ||
6057 | if (size1 <= 4096) | |
6058 | { | |
e0d80184 | 6059 | fprintf (file, "\tadd\t%s, %d, %s\n", |
86735b02 DE |
6060 | sp_str, -size1, sp_str); |
6061 | if (gmask & FRAME_POINTER_MASK) | |
6062 | { | |
e0d80184 | 6063 | fprintf (file, "\tst\t%s, [%s+%d]\n\tsub\t%s, %d, %s\t%s# set up frame pointer\n", |
bf62bbf1 DE |
6064 | fp_str, sp_str, offset, sp_str, -size1, fp_str, |
6065 | ASM_COMMENT_START); | |
86735b02 DE |
6066 | offset += 4; |
6067 | } | |
6068 | } | |
6069 | else | |
6070 | { | |
e0d80184 | 6071 | fprintf (file, "\tset\t%d, %s\n\tsub\t%s, %s, %s\n", |
86735b02 DE |
6072 | size1, t1_str, sp_str, t1_str, sp_str); |
6073 | if (gmask & FRAME_POINTER_MASK) | |
6074 | { | |
e0d80184 | 6075 | fprintf (file, "\tst\t%s, [%s+%d]\n\tadd\t%s, %s, %s\t%s# set up frame pointer\n", |
bf62bbf1 DE |
6076 | fp_str, sp_str, offset, sp_str, t1_str, fp_str, |
6077 | ASM_COMMENT_START); | |
86735b02 DE |
6078 | offset += 4; |
6079 | } | |
6080 | } | |
0021b564 | 6081 | if (dwarf2out_do_frame ()) |
c53aa195 | 6082 | { |
96987e96 | 6083 | char *l = dwarf2out_cfi_label (); |
c53aa195 JM |
6084 | if (gmask & FRAME_POINTER_MASK) |
6085 | { | |
6086 | dwarf2out_reg_save (l, FRAME_POINTER_REGNUM, | |
6087 | offset - 4 - size1); | |
6088 | dwarf2out_def_cfa (l, FRAME_POINTER_REGNUM, 0); | |
6089 | } | |
6090 | else | |
6091 | dwarf2out_def_cfa (l, STACK_POINTER_REGNUM, size1); | |
6092 | } | |
86735b02 DE |
6093 | if (gmask & RETURN_ADDR_MASK) |
6094 | { | |
e0d80184 | 6095 | fprintf (file, "\tst\t%s, [%s+%d]\n", |
86735b02 | 6096 | reg_names[RETURN_ADDR_REGNUM], sp_str, offset); |
0021b564 | 6097 | if (dwarf2out_do_frame ()) |
c53aa195 JM |
6098 | /* offset - size1 == reg_offset - size |
6099 | if reg_offset were updated above like offset. */ | |
6100 | dwarf2out_return_save ("", offset - size1); | |
86735b02 DE |
6101 | offset += 4; |
6102 | } | |
6103 | sparc_flat_save_restore (file, sp_str, offset, | |
6104 | gmask & ~(FRAME_POINTER_MASK | RETURN_ADDR_MASK), | |
6105 | current_frame_info.fmask, | |
84e884e9 | 6106 | "st", "std", -size1); |
e0d80184 | 6107 | fprintf (file, "\tset\t%d, %s\n\tsub\t%s, %s, %s\n", |
86735b02 | 6108 | size - size1, t1_str, sp_str, t1_str, sp_str); |
0021b564 | 6109 | if (dwarf2out_do_frame ()) |
c53aa195 JM |
6110 | if (! (gmask & FRAME_POINTER_MASK)) |
6111 | dwarf2out_def_cfa ("", STACK_POINTER_REGNUM, size); | |
86735b02 | 6112 | } |
c819be5b | 6113 | } |
c6aa9ce1 | 6114 | |
bf62bbf1 | 6115 | fprintf (file, "\t%s#PROLOGUE# 1\n", ASM_COMMENT_START); |
c819be5b JW |
6116 | } |
6117 | \f | |
6118 | /* Do any necessary cleanup after a function to restore stack, frame, | |
6119 | and regs. */ | |
6120 | ||
6121 | void | |
c6aa9ce1 | 6122 | sparc_flat_output_function_epilogue (file, size) |
c819be5b JW |
6123 | FILE *file; |
6124 | int size; | |
6125 | { | |
c819be5b JW |
6126 | rtx epilogue_delay = current_function_epilogue_delay_list; |
6127 | int noepilogue = FALSE; | |
c819be5b | 6128 | |
c6aa9ce1 | 6129 | /* This is only for the human reader. */ |
bf62bbf1 | 6130 | fprintf (file, "\t%s#EPILOGUE#\n", ASM_COMMENT_START); |
c6aa9ce1 | 6131 | |
c819be5b JW |
6132 | /* The epilogue does not depend on any registers, but the stack |
6133 | registers, so we assume that if we have 1 pending nop, it can be | |
6134 | ignored, and 2 it must be filled (2 nops occur for integer | |
6135 | multiply and divide). */ | |
6136 | ||
6137 | size = SPARC_STACK_ALIGN (size); | |
86735b02 | 6138 | size = (!current_frame_info.initialized |
c6aa9ce1 | 6139 | ? sparc_flat_compute_frame_size (size) |
c819be5b JW |
6140 | : current_frame_info.total_size); |
6141 | ||
86735b02 | 6142 | if (size == 0 && epilogue_delay == 0) |
c819be5b JW |
6143 | { |
6144 | rtx insn = get_last_insn (); | |
6145 | ||
6146 | /* If the last insn was a BARRIER, we don't have to write any code | |
6147 | because a jump (aka return) was put there. */ | |
6148 | if (GET_CODE (insn) == NOTE) | |
6149 | insn = prev_nonnote_insn (insn); | |
6150 | if (insn && GET_CODE (insn) == BARRIER) | |
6151 | noepilogue = TRUE; | |
6152 | } | |
6153 | ||
6154 | if (!noepilogue) | |
6155 | { | |
86735b02 DE |
6156 | unsigned int reg_offset = current_frame_info.reg_offset; |
6157 | unsigned int size1; | |
6158 | char *sp_str = reg_names[STACK_POINTER_REGNUM]; | |
6159 | char *fp_str = reg_names[FRAME_POINTER_REGNUM]; | |
3bb5de61 | 6160 | const char *t1_str = "%g1"; |
86735b02 | 6161 | |
c819be5b JW |
6162 | /* In the reload sequence, we don't need to fill the load delay |
6163 | slots for most of the loads, also see if we can fill the final | |
6164 | delay slot if not otherwise filled by the reload sequence. */ | |
6165 | ||
86735b02 | 6166 | if (size > 4095) |
e0d80184 | 6167 | fprintf (file, "\tset\t%d, %s\n", size, t1_str); |
c819be5b JW |
6168 | |
6169 | if (frame_pointer_needed) | |
6170 | { | |
86735b02 | 6171 | if (size > 4095) |
e0d80184 | 6172 | fprintf (file,"\tsub\t%s, %s, %s\t\t%s# sp not trusted here\n", |
bf62bbf1 | 6173 | fp_str, t1_str, sp_str, ASM_COMMENT_START); |
c819be5b | 6174 | else |
e0d80184 | 6175 | fprintf (file,"\tsub\t%s, %d, %s\t\t%s# sp not trusted here\n", |
bf62bbf1 | 6176 | fp_str, size, sp_str, ASM_COMMENT_START); |
c819be5b JW |
6177 | } |
6178 | ||
ddd5a7c1 | 6179 | /* Is the entire register save area offsettable from %sp? */ |
86735b02 DE |
6180 | if (reg_offset < 4096 - 64 * UNITS_PER_WORD) |
6181 | { | |
6182 | size1 = 0; | |
6183 | } | |
6184 | else | |
6185 | { | |
6186 | /* Restore %sp in two steps, but make sure there is always a | |
6187 | 64 byte register save area, and %sp is properly aligned. */ | |
6188 | /* Amount to increment %sp by, the first time. */ | |
6189 | size1 = ((reg_offset - 64 - 16) + 15) & -16; | |
6190 | /* Offset to register save area from %sp. */ | |
6191 | reg_offset = size1 - reg_offset; | |
6192 | ||
e0d80184 | 6193 | fprintf (file, "\tset\t%d, %s\n\tadd\t%s, %s, %s\n", |
86735b02 DE |
6194 | size1, t1_str, sp_str, t1_str, sp_str); |
6195 | } | |
6196 | ||
6197 | /* We must restore the frame pointer and return address reg first | |
6198 | because they are treated specially by the prologue output code. */ | |
6199 | if (current_frame_info.gmask & FRAME_POINTER_MASK) | |
6200 | { | |
e0d80184 | 6201 | fprintf (file, "\tld\t[%s+%d], %s\n", |
86735b02 DE |
6202 | sp_str, reg_offset, fp_str); |
6203 | reg_offset += 4; | |
6204 | } | |
6205 | if (current_frame_info.gmask & RETURN_ADDR_MASK) | |
6206 | { | |
e0d80184 | 6207 | fprintf (file, "\tld\t[%s+%d], %s\n", |
86735b02 DE |
6208 | sp_str, reg_offset, reg_names[RETURN_ADDR_REGNUM]); |
6209 | reg_offset += 4; | |
6210 | } | |
6211 | ||
6212 | /* Restore any remaining saved registers. */ | |
6213 | sparc_flat_save_restore (file, sp_str, reg_offset, | |
6214 | current_frame_info.gmask & ~(FRAME_POINTER_MASK | RETURN_ADDR_MASK), | |
6215 | current_frame_info.fmask, | |
c53aa195 | 6216 | "ld", "ldd", 0); |
86735b02 DE |
6217 | |
6218 | /* If we had to increment %sp in two steps, record it so the second | |
6219 | restoration in the epilogue finishes up. */ | |
6220 | if (size1 > 0) | |
6221 | { | |
6222 | size -= size1; | |
6223 | if (size > 4095) | |
e0d80184 | 6224 | fprintf (file, "\tset\t%d, %s\n", |
86735b02 DE |
6225 | size, t1_str); |
6226 | } | |
c819be5b | 6227 | |
c819be5b | 6228 | if (current_function_returns_struct) |
e0d80184 | 6229 | fprintf (file, "\tjmp\t%%o7+12\n"); |
c819be5b JW |
6230 | else |
6231 | fprintf (file, "\tretl\n"); | |
6232 | ||
6233 | /* If the only register saved is the return address, we need a | |
6234 | nop, unless we have an instruction to put into it. Otherwise | |
6235 | we don't since reloading multiple registers doesn't reference | |
6236 | the register being loaded. */ | |
6237 | ||
6238 | if (epilogue_delay) | |
6239 | { | |
86735b02 | 6240 | if (size) |
c819be5b JW |
6241 | abort (); |
6242 | final_scan_insn (XEXP (epilogue_delay, 0), file, 1, -2, 1); | |
6243 | } | |
6244 | ||
86735b02 | 6245 | else if (size > 4095) |
e0d80184 | 6246 | fprintf (file, "\tadd\t%s, %s, %s\n", sp_str, t1_str, sp_str); |
c819be5b | 6247 | |
86735b02 | 6248 | else if (size > 0) |
e0d80184 | 6249 | fprintf (file, "\tadd\t%s, %d, %s\n", sp_str, size, sp_str); |
c819be5b JW |
6250 | |
6251 | else | |
6252 | fprintf (file, "\tnop\n"); | |
6253 | } | |
6254 | ||
6255 | /* Reset state info for each function. */ | |
6256 | current_frame_info = zero_frame_info; | |
e0d80184 DM |
6257 | |
6258 | sparc_output_deferred_case_vectors (); | |
c819be5b JW |
6259 | } |
6260 | \f | |
6261 | /* Define the number of delay slots needed for the function epilogue. | |
6262 | ||
6263 | On the sparc, we need a slot if either no stack has been allocated, | |
6264 | or the only register saved is the return register. */ | |
6265 | ||
6266 | int | |
c6aa9ce1 | 6267 | sparc_flat_epilogue_delay_slots () |
c819be5b JW |
6268 | { |
6269 | if (!current_frame_info.initialized) | |
c6aa9ce1 | 6270 | (void) sparc_flat_compute_frame_size (get_frame_size ()); |
c819be5b JW |
6271 | |
6272 | if (current_frame_info.total_size == 0) | |
6273 | return 1; | |
6274 | ||
c819be5b JW |
6275 | return 0; |
6276 | } | |
6277 | ||
6278 | /* Return true is TRIAL is a valid insn for the epilogue delay slot. | |
6279 | Any single length instruction which doesn't reference the stack or frame | |
6280 | pointer is OK. */ | |
6281 | ||
6282 | int | |
c6aa9ce1 | 6283 | sparc_flat_eligible_for_epilogue_delay (trial, slot) |
c819be5b | 6284 | rtx trial; |
db3d4438 | 6285 | int slot ATTRIBUTE_UNUSED; |
c819be5b | 6286 | { |
c4ce6853 DE |
6287 | rtx pat = PATTERN (trial); |
6288 | ||
6289 | if (get_attr_length (trial) != 1) | |
6290 | return 0; | |
6291 | ||
6292 | /* If %g0 is live, there are lots of things we can't handle. | |
6293 | Rather than trying to find them all now, let's punt and only | |
6294 | optimize things as necessary. */ | |
6295 | if (TARGET_LIVE_G0) | |
6296 | return 0; | |
6297 | ||
6298 | if (! reg_mentioned_p (stack_pointer_rtx, pat) | |
6299 | && ! reg_mentioned_p (frame_pointer_rtx, pat)) | |
c819be5b | 6300 | return 1; |
c4ce6853 | 6301 | |
c819be5b JW |
6302 | return 0; |
6303 | } | |
dbb54862 JW |
6304 | \f |
6305 | /* Adjust the cost of a scheduling dependency. Return the new cost of | |
6306 | a dependency LINK or INSN on DEP_INSN. COST is the current cost. */ | |
6307 | ||
8947065c | 6308 | static int |
dbb54862 JW |
6309 | supersparc_adjust_cost (insn, link, dep_insn, cost) |
6310 | rtx insn; | |
6311 | rtx link; | |
6312 | rtx dep_insn; | |
6313 | int cost; | |
6314 | { | |
6315 | enum attr_type insn_type; | |
6316 | ||
6317 | if (! recog_memoized (insn)) | |
6318 | return 0; | |
6319 | ||
6320 | insn_type = get_attr_type (insn); | |
6321 | ||
6322 | if (REG_NOTE_KIND (link) == 0) | |
6323 | { | |
6324 | /* Data dependency; DEP_INSN writes a register that INSN reads some | |
6325 | cycles later. */ | |
6326 | ||
6327 | /* if a load, then the dependence must be on the memory address; | |
3bc8b61e DM |
6328 | add an extra "cycle". Note that the cost could be two cycles |
6329 | if the reg was written late in an instruction group; we ca not tell | |
dbb54862 JW |
6330 | here. */ |
6331 | if (insn_type == TYPE_LOAD || insn_type == TYPE_FPLOAD) | |
6332 | return cost + 3; | |
6333 | ||
6334 | /* Get the delay only if the address of the store is the dependence. */ | |
6335 | if (insn_type == TYPE_STORE || insn_type == TYPE_FPSTORE) | |
6336 | { | |
6337 | rtx pat = PATTERN(insn); | |
6338 | rtx dep_pat = PATTERN (dep_insn); | |
6339 | ||
6340 | if (GET_CODE (pat) != SET || GET_CODE (dep_pat) != SET) | |
3bc8b61e | 6341 | return cost; /* This should not happen! */ |
dbb54862 JW |
6342 | |
6343 | /* The dependency between the two instructions was on the data that | |
6344 | is being stored. Assume that this implies that the address of the | |
6345 | store is not dependent. */ | |
6346 | if (rtx_equal_p (SET_DEST (dep_pat), SET_SRC (pat))) | |
6347 | return cost; | |
6348 | ||
6349 | return cost + 3; /* An approximation. */ | |
6350 | } | |
6351 | ||
6352 | /* A shift instruction cannot receive its data from an instruction | |
6353 | in the same cycle; add a one cycle penalty. */ | |
6354 | if (insn_type == TYPE_SHIFT) | |
6355 | return cost + 3; /* Split before cascade into shift. */ | |
6356 | } | |
6357 | else | |
6358 | { | |
6359 | /* Anti- or output- dependency; DEP_INSN reads/writes a register that | |
6360 | INSN writes some cycles later. */ | |
6361 | ||
6362 | /* These are only significant for the fpu unit; writing a fp reg before | |
6363 | the fpu has finished with it stalls the processor. */ | |
6364 | ||
6365 | /* Reusing an integer register causes no problems. */ | |
6366 | if (insn_type == TYPE_IALU || insn_type == TYPE_SHIFT) | |
6367 | return 0; | |
6368 | } | |
6369 | ||
6370 | return cost; | |
6371 | } | |
bfd6bc60 | 6372 | |
8947065c RH |
6373 | static int |
6374 | hypersparc_adjust_cost (insn, link, dep_insn, cost) | |
6375 | rtx insn; | |
6376 | rtx link; | |
6377 | rtx dep_insn; | |
6378 | int cost; | |
6379 | { | |
6380 | enum attr_type insn_type, dep_type; | |
6381 | rtx pat = PATTERN(insn); | |
6382 | rtx dep_pat = PATTERN (dep_insn); | |
6383 | ||
6384 | if (recog_memoized (insn) < 0 || recog_memoized (dep_insn) < 0) | |
6385 | return cost; | |
6386 | ||
6387 | insn_type = get_attr_type (insn); | |
6388 | dep_type = get_attr_type (dep_insn); | |
6389 | ||
6390 | switch (REG_NOTE_KIND (link)) | |
6391 | { | |
6392 | case 0: | |
6393 | /* Data dependency; DEP_INSN writes a register that INSN reads some | |
6394 | cycles later. */ | |
6395 | ||
6396 | switch (insn_type) | |
6397 | { | |
6398 | case TYPE_STORE: | |
6399 | case TYPE_FPSTORE: | |
6400 | /* Get the delay iff the address of the store is the dependence. */ | |
6401 | if (GET_CODE (pat) != SET || GET_CODE (dep_pat) != SET) | |
6402 | return cost; | |
6403 | ||
6404 | if (rtx_equal_p (SET_DEST (dep_pat), SET_SRC (pat))) | |
6405 | return cost; | |
6406 | return cost + 3; | |
6407 | ||
6408 | case TYPE_LOAD: | |
6409 | case TYPE_SLOAD: | |
6410 | case TYPE_FPLOAD: | |
6411 | /* If a load, then the dependence must be on the memory address. If | |
6412 | the addresses aren't equal, then it might be a false dependency */ | |
6413 | if (dep_type == TYPE_STORE || dep_type == TYPE_FPSTORE) | |
6414 | { | |
6415 | if (GET_CODE (pat) != SET || GET_CODE (dep_pat) != SET | |
6416 | || GET_CODE (SET_DEST (dep_pat)) != MEM | |
6417 | || GET_CODE (SET_SRC (pat)) != MEM | |
6418 | || ! rtx_equal_p (XEXP (SET_DEST (dep_pat), 0), | |
6419 | XEXP (SET_SRC (pat), 0))) | |
6420 | return cost + 2; | |
6421 | ||
6422 | return cost + 8; | |
6423 | } | |
6424 | break; | |
6425 | ||
6426 | case TYPE_BRANCH: | |
6427 | /* Compare to branch latency is 0. There is no benefit from | |
6428 | separating compare and branch. */ | |
6429 | if (dep_type == TYPE_COMPARE) | |
6430 | return 0; | |
6431 | /* Floating point compare to branch latency is less than | |
6432 | compare to conditional move. */ | |
6433 | if (dep_type == TYPE_FPCMP) | |
6434 | return cost - 1; | |
6435 | break; | |
3bb5de61 KG |
6436 | default: |
6437 | break; | |
8947065c RH |
6438 | } |
6439 | break; | |
6440 | ||
6441 | case REG_DEP_ANTI: | |
6442 | /* Anti-dependencies only penalize the fpu unit. */ | |
6443 | if (insn_type == TYPE_IALU || insn_type == TYPE_SHIFT) | |
6444 | return 0; | |
6445 | break; | |
6446 | ||
6447 | default: | |
6448 | break; | |
6449 | } | |
6450 | ||
6451 | return cost; | |
6452 | } | |
6453 | ||
6454 | static int | |
6455 | ultrasparc_adjust_cost (insn, link, dep_insn, cost) | |
6456 | rtx insn; | |
6457 | rtx link; | |
6458 | rtx dep_insn; | |
6459 | int cost; | |
6460 | { | |
6461 | enum attr_type insn_type, dep_type; | |
6462 | rtx pat = PATTERN(insn); | |
6463 | rtx dep_pat = PATTERN (dep_insn); | |
6464 | ||
6465 | if (recog_memoized (insn) < 0 || recog_memoized (dep_insn) < 0) | |
6466 | return cost; | |
6467 | ||
6468 | insn_type = get_attr_type (insn); | |
6469 | dep_type = get_attr_type (dep_insn); | |
6470 | ||
6471 | /* Nothing issues in parallel with integer multiplies, so | |
6472 | mark as zero cost since the scheduler can not do anything | |
6473 | about it. */ | |
6474 | if (insn_type == TYPE_IMUL) | |
6475 | return 0; | |
6476 | ||
6477 | #define SLOW_FP(dep_type) \ | |
6478 | (dep_type == TYPE_FPSQRT || dep_type == TYPE_FPDIVS || dep_type == TYPE_FPDIVD) | |
6479 | ||
6480 | switch (REG_NOTE_KIND (link)) | |
6481 | { | |
6482 | case 0: | |
6483 | /* Data dependency; DEP_INSN writes a register that INSN reads some | |
6484 | cycles later. */ | |
6485 | ||
6486 | if (dep_type == TYPE_CMOVE) | |
6487 | { | |
6488 | /* Instructions that read the result of conditional moves cannot | |
6489 | be in the same group or the following group. */ | |
6490 | return cost + 1; | |
6491 | } | |
6492 | ||
6493 | switch (insn_type) | |
6494 | { | |
6495 | /* UltraSPARC can dual issue a store and an instruction setting | |
6496 | the value stored, except for divide and square root. */ | |
6497 | case TYPE_FPSTORE: | |
6498 | if (! SLOW_FP (dep_type)) | |
6499 | return 0; | |
6500 | return cost; | |
6501 | ||
6502 | case TYPE_STORE: | |
6503 | if (GET_CODE (pat) != SET || GET_CODE (dep_pat) != SET) | |
6504 | return cost; | |
6505 | ||
6506 | if (rtx_equal_p (SET_DEST (dep_pat), SET_SRC (pat))) | |
6507 | /* The dependency between the two instructions is on the data | |
6508 | that is being stored. Assume that the address of the store | |
6509 | is not also dependent. */ | |
6510 | return 0; | |
6511 | return cost; | |
6512 | ||
6513 | case TYPE_LOAD: | |
6514 | case TYPE_SLOAD: | |
6515 | case TYPE_FPLOAD: | |
6516 | /* A load does not return data until at least 11 cycles after | |
6517 | a store to the same location. 3 cycles are accounted for | |
6518 | in the load latency; add the other 8 here. */ | |
6519 | if (dep_type == TYPE_STORE || dep_type == TYPE_FPSTORE) | |
6520 | { | |
6521 | /* If the addresses are not equal this may be a false | |
6522 | dependency because pointer aliasing could not be | |
6523 | determined. Add only 2 cycles in that case. 2 is | |
6524 | an arbitrary compromise between 8, which would cause | |
6525 | the scheduler to generate worse code elsewhere to | |
6526 | compensate for a dependency which might not really | |
6527 | exist, and 0. */ | |
6528 | if (GET_CODE (pat) != SET || GET_CODE (dep_pat) != SET | |
6529 | || GET_CODE (SET_SRC (pat)) != MEM | |
6530 | || GET_CODE (SET_DEST (dep_pat)) != MEM | |
6531 | || ! rtx_equal_p (XEXP (SET_SRC (pat), 0), | |
6532 | XEXP (SET_DEST (dep_pat), 0))) | |
6533 | return cost + 2; | |
6534 | ||
6535 | return cost + 8; | |
6536 | } | |
6537 | return cost; | |
6538 | ||
6539 | case TYPE_BRANCH: | |
6540 | /* Compare to branch latency is 0. There is no benefit from | |
6541 | separating compare and branch. */ | |
6542 | if (dep_type == TYPE_COMPARE) | |
6543 | return 0; | |
6544 | /* Floating point compare to branch latency is less than | |
6545 | compare to conditional move. */ | |
6546 | if (dep_type == TYPE_FPCMP) | |
6547 | return cost - 1; | |
6548 | return cost; | |
6549 | ||
6550 | case TYPE_FPCMOVE: | |
6551 | /* FMOVR class instructions can not issue in the same cycle | |
6552 | or the cycle after an instruction which writes any | |
6553 | integer register. Model this as cost 2 for dependent | |
6554 | instructions. */ | |
6555 | if ((dep_type == TYPE_IALU || dep_type == TYPE_UNARY | |
6556 | || dep_type == TYPE_BINARY) | |
6557 | && cost < 2) | |
6558 | return 2; | |
6559 | /* Otherwise check as for integer conditional moves. */ | |
6560 | ||
6561 | case TYPE_CMOVE: | |
6562 | /* Conditional moves involving integer registers wait until | |
6563 | 3 cycles after loads return data. The interlock applies | |
6564 | to all loads, not just dependent loads, but that is hard | |
6565 | to model. */ | |
6566 | if (dep_type == TYPE_LOAD || dep_type == TYPE_SLOAD) | |
6567 | return cost + 3; | |
6568 | return cost; | |
6569 | ||
6570 | default: | |
6571 | break; | |
6572 | } | |
6573 | break; | |
6574 | ||
6575 | case REG_DEP_ANTI: | |
6576 | /* Divide and square root lock destination registers for full latency. */ | |
6577 | if (! SLOW_FP (dep_type)) | |
6578 | return 0; | |
6579 | break; | |
6580 | ||
6581 | case REG_DEP_OUTPUT: | |
6582 | /* IEU and FPU instruction that have the same destination | |
6583 | register cannot be grouped together. */ | |
6584 | return cost + 1; | |
6585 | ||
6586 | default: | |
6587 | break; | |
6588 | } | |
6589 | ||
6590 | /* Other costs not accounted for: | |
6591 | - Single precision floating point loads lock the other half of | |
6592 | the even/odd register pair. | |
6593 | - Several hazards associated with ldd/std are ignored because these | |
6594 | instructions are rarely generated for V9. | |
6595 | - The floating point pipeline can not have both a single and double | |
6596 | precision operation active at the same time. Format conversions | |
6597 | and graphics instructions are given honorary double precision status. | |
6598 | - call and jmpl are always the first instruction in a group. */ | |
6599 | ||
6600 | return cost; | |
6601 | ||
6602 | #undef SLOW_FP | |
6603 | } | |
6604 | ||
6605 | int | |
6606 | sparc_adjust_cost(insn, link, dep, cost) | |
6607 | rtx insn; | |
6608 | rtx link; | |
6609 | rtx dep; | |
6610 | int cost; | |
6611 | { | |
6612 | switch (sparc_cpu) | |
6613 | { | |
6614 | case PROCESSOR_SUPERSPARC: | |
6615 | cost = supersparc_adjust_cost (insn, link, dep, cost); | |
6616 | break; | |
6617 | case PROCESSOR_HYPERSPARC: | |
6618 | case PROCESSOR_SPARCLITE86X: | |
6619 | cost = hypersparc_adjust_cost (insn, link, dep, cost); | |
6620 | break; | |
6621 | case PROCESSOR_ULTRASPARC: | |
6622 | cost = ultrasparc_adjust_cost (insn, link, dep, cost); | |
6623 | break; | |
6624 | default: | |
6625 | break; | |
6626 | } | |
6627 | return cost; | |
6628 | } | |
6629 | ||
3bc8b61e DM |
6630 | /* This describes the state of the UltraSPARC pipeline during |
6631 | instruction scheduling. */ | |
6632 | ||
b96a5ea5 KG |
6633 | #define TMASK(__x) ((unsigned)1 << ((int)(__x))) |
6634 | #define UMASK(__x) ((unsigned)1 << ((int)(__x))) | |
3bc8b61e DM |
6635 | |
6636 | enum ultra_code { NONE=0, /* no insn at all */ | |
6637 | IEU0, /* shifts and conditional moves */ | |
6638 | IEU1, /* condition code setting insns, calls+jumps */ | |
6639 | IEUN, /* all other single cycle ieu insns */ | |
6640 | LSU, /* loads and stores */ | |
6641 | CTI, /* branches */ | |
6642 | FPM, /* FPU pipeline 1, multiplies and divides */ | |
6643 | FPA, /* FPU pipeline 2, all other operations */ | |
6644 | SINGLE, /* single issue instructions */ | |
6645 | NUM_ULTRA_CODES }; | |
6646 | ||
b1474bb7 KG |
6647 | static enum ultra_code ultra_code_from_mask PROTO ((int)); |
6648 | static void ultra_schedule_insn PROTO ((rtx *, rtx *, int, enum ultra_code)); | |
6649 | ||
3bb5de61 | 6650 | static const char *ultra_code_names[NUM_ULTRA_CODES] = { |
3bc8b61e DM |
6651 | "NONE", "IEU0", "IEU1", "IEUN", "LSU", "CTI", |
6652 | "FPM", "FPA", "SINGLE" }; | |
6653 | ||
6654 | struct ultrasparc_pipeline_state { | |
6655 | /* The insns in this group. */ | |
6656 | rtx group[4]; | |
6657 | ||
6658 | /* The code for each insn. */ | |
6659 | enum ultra_code codes[4]; | |
6660 | ||
6661 | /* Which insns in this group have been committed by the | |
6662 | scheduler. This is how we determine how many more | |
6663 | can issue this cycle. */ | |
6664 | char commit[4]; | |
6665 | ||
6666 | /* How many insns in this group. */ | |
6667 | char group_size; | |
6668 | ||
6669 | /* Mask of free slots still in this group. */ | |
6670 | char free_slot_mask; | |
6671 | ||
6672 | /* The slotter uses the following to determine what other | |
6673 | insn types can still make their way into this group. */ | |
6674 | char contents [NUM_ULTRA_CODES]; | |
6675 | char num_ieu_insns; | |
6676 | }; | |
6677 | ||
6678 | #define ULTRA_NUM_HIST 8 | |
6679 | static struct ultrasparc_pipeline_state ultra_pipe_hist[ULTRA_NUM_HIST]; | |
6680 | static int ultra_cur_hist; | |
6681 | static int ultra_cycles_elapsed; | |
6682 | ||
6683 | #define ultra_pipe (ultra_pipe_hist[ultra_cur_hist]) | |
6684 | ||
6685 | /* Given TYPE_MASK compute the ultra_code it has. */ | |
6686 | static enum ultra_code | |
6687 | ultra_code_from_mask (type_mask) | |
6688 | int type_mask; | |
6689 | { | |
3bc8b61e DM |
6690 | if (type_mask & (TMASK (TYPE_SHIFT) | TMASK (TYPE_CMOVE))) |
6691 | return IEU0; | |
6692 | else if (type_mask & (TMASK (TYPE_COMPARE) | | |
6693 | TMASK (TYPE_CALL) | | |
6694 | TMASK (TYPE_UNCOND_BRANCH))) | |
6695 | return IEU1; | |
6696 | else if (type_mask & (TMASK (TYPE_IALU) | TMASK (TYPE_BINARY) | | |
6697 | TMASK (TYPE_MOVE) | TMASK (TYPE_UNARY))) | |
6698 | return IEUN; | |
6699 | else if (type_mask & (TMASK (TYPE_LOAD) | TMASK (TYPE_SLOAD) | | |
6700 | TMASK (TYPE_STORE) | TMASK (TYPE_FPLOAD) | | |
6701 | TMASK (TYPE_FPSTORE))) | |
6702 | return LSU; | |
6703 | else if (type_mask & (TMASK (TYPE_FPMUL) | TMASK (TYPE_FPDIVS) | | |
6704 | TMASK (TYPE_FPDIVD) | TMASK (TYPE_FPSQRT))) | |
6705 | return FPM; | |
6706 | else if (type_mask & (TMASK (TYPE_FPMOVE) | TMASK (TYPE_FPCMOVE) | | |
6707 | TMASK (TYPE_FP) | TMASK (TYPE_FPCMP))) | |
6708 | return FPA; | |
6709 | else if (type_mask & TMASK (TYPE_BRANCH)) | |
6710 | return CTI; | |
6711 | ||
6712 | return SINGLE; | |
6713 | } | |
6714 | ||
6715 | /* Check INSN (a conditional move) and make sure that it's | |
6716 | results are available at this cycle. Return 1 if the | |
6717 | results are in fact ready. */ | |
6718 | static int | |
6719 | ultra_cmove_results_ready_p (insn) | |
6720 | rtx insn; | |
6721 | { | |
6722 | struct ultrasparc_pipeline_state *up; | |
6723 | int entry, slot; | |
6724 | ||
6725 | /* If this got dispatched in the previous | |
6726 | group, the results are not ready. */ | |
ddf80874 | 6727 | entry = (ultra_cur_hist - 1) % (ULTRA_NUM_HIST - 1); |
3bc8b61e DM |
6728 | up = &ultra_pipe_hist[entry]; |
6729 | slot = 4; | |
6730 | while (--slot >= 0) | |
6731 | if (up->group[slot] == insn) | |
c305b1a7 | 6732 | return 0; |
3bc8b61e | 6733 | |
c305b1a7 | 6734 | return 1; |
3bc8b61e DM |
6735 | } |
6736 | ||
6737 | /* Walk backwards in pipeline history looking for FPU | |
6738 | operations which use a mode different than FPMODE and | |
6739 | will create a stall if an insn using FPMODE were to be | |
6740 | dispatched this cycle. */ | |
6741 | static int | |
6742 | ultra_fpmode_conflict_exists (fpmode) | |
6743 | enum machine_mode fpmode; | |
6744 | { | |
6745 | int hist_ent; | |
6746 | int hist_lim; | |
6747 | ||
ddf80874 | 6748 | hist_ent = (ultra_cur_hist - 1) % (ULTRA_NUM_HIST - 1); |
3bc8b61e DM |
6749 | if (ultra_cycles_elapsed < 4) |
6750 | hist_lim = ultra_cycles_elapsed; | |
6751 | else | |
6752 | hist_lim = 4; | |
6753 | while (hist_lim > 0) | |
6754 | { | |
6755 | struct ultrasparc_pipeline_state *up = &ultra_pipe_hist[hist_ent]; | |
6756 | int slot = 4; | |
6757 | ||
6758 | while (--slot >= 0) | |
6759 | { | |
6760 | rtx insn = up->group[slot]; | |
6761 | enum machine_mode this_mode; | |
3bc8b61e DM |
6762 | rtx pat; |
6763 | ||
6764 | if (! insn | |
6765 | || GET_CODE (insn) != INSN | |
6766 | || (pat = PATTERN (insn)) == 0 | |
6767 | || GET_CODE (pat) != SET) | |
6768 | continue; | |
6769 | ||
6770 | this_mode = GET_MODE (SET_DEST (pat)); | |
6771 | if ((this_mode != SFmode | |
6772 | && this_mode != DFmode) | |
6773 | || this_mode == fpmode) | |
6774 | continue; | |
6775 | ||
6776 | /* If it is not FMOV, FABS, FNEG, FDIV, or FSQRT then | |
ddf80874 DM |
6777 | we will get a stall. Loads and stores are independant |
6778 | of these rules. */ | |
3bc8b61e DM |
6779 | if (GET_CODE (SET_SRC (pat)) != ABS |
6780 | && GET_CODE (SET_SRC (pat)) != NEG | |
6781 | && ((TMASK (get_attr_type (insn)) & | |
6782 | (TMASK (TYPE_FPDIVS) | TMASK (TYPE_FPDIVD) | | |
ddf80874 DM |
6783 | TMASK (TYPE_FPMOVE) | TMASK (TYPE_FPSQRT) | |
6784 | TMASK (TYPE_LOAD) | TMASK (TYPE_STORE))) == 0)) | |
3bc8b61e DM |
6785 | return 1; |
6786 | } | |
6787 | hist_lim--; | |
ddf80874 | 6788 | hist_ent = (hist_ent - 1) % (ULTRA_NUM_HIST - 1); |
3bc8b61e DM |
6789 | } |
6790 | ||
6791 | /* No conflicts, safe to dispatch. */ | |
6792 | return 0; | |
6793 | } | |
6794 | ||
6795 | /* Find an instruction in LIST which has one of the | |
6796 | type attributes enumerated in TYPE_MASK. START | |
6797 | says where to begin the search. | |
6798 | ||
6799 | NOTE: This scheme depends upon the fact that we | |
6800 | have less than 32 distinct type attributes. */ | |
5a5325cb DM |
6801 | |
6802 | static int ultra_types_avail; | |
6803 | ||
3bc8b61e DM |
6804 | static rtx * |
6805 | ultra_find_type (type_mask, list, start) | |
6806 | int type_mask; | |
6807 | rtx *list; | |
6808 | int start; | |
6809 | { | |
6810 | int i; | |
6811 | ||
5a5325cb DM |
6812 | /* Short circuit if no such insn exists in the ready |
6813 | at the moment. */ | |
6814 | if ((type_mask & ultra_types_avail) == 0) | |
6815 | return 0; | |
6816 | ||
3bc8b61e DM |
6817 | for (i = start; i >= 0; i--) |
6818 | { | |
6819 | rtx insn = list[i]; | |
6820 | ||
6821 | if (recog_memoized (insn) >= 0 | |
6822 | && (TMASK(get_attr_type (insn)) & type_mask)) | |
6823 | { | |
ddf80874 | 6824 | enum machine_mode fpmode = SFmode; |
3bc8b61e DM |
6825 | rtx pat = 0; |
6826 | int slot; | |
6827 | int check_depend = 0; | |
6828 | int check_fpmode_conflict = 0; | |
6829 | ||
6830 | if (GET_CODE (insn) == INSN | |
6831 | && (pat = PATTERN(insn)) != 0 | |
6832 | && GET_CODE (pat) == SET | |
6833 | && !(type_mask & (TMASK (TYPE_STORE) | | |
6834 | TMASK (TYPE_FPSTORE)))) | |
6835 | { | |
6836 | check_depend = 1; | |
6837 | if (GET_MODE (SET_DEST (pat)) == SFmode | |
6838 | || GET_MODE (SET_DEST (pat)) == DFmode) | |
6839 | { | |
6840 | fpmode = GET_MODE (SET_DEST (pat)); | |
6841 | check_fpmode_conflict = 1; | |
6842 | } | |
6843 | } | |
6844 | ||
6845 | slot = 4; | |
6846 | while(--slot >= 0) | |
6847 | { | |
6848 | rtx slot_insn = ultra_pipe.group[slot]; | |
6849 | rtx slot_pat; | |
6850 | ||
6851 | /* Already issued, bad dependency, or FPU | |
6852 | mode conflict. */ | |
6853 | if (slot_insn != 0 | |
6854 | && (slot_pat = PATTERN (slot_insn)) != 0 | |
6855 | && ((insn == slot_insn) | |
6856 | || (check_depend == 1 | |
6857 | && GET_CODE (slot_insn) == INSN | |
6858 | && GET_CODE (slot_pat) == SET | |
5a5325cb DM |
6859 | && ((GET_CODE (SET_DEST (slot_pat)) == REG |
6860 | && GET_CODE (SET_SRC (pat)) == REG | |
6861 | && REGNO (SET_DEST (slot_pat)) == | |
6862 | REGNO (SET_SRC (pat))) | |
6863 | || (GET_CODE (SET_DEST (slot_pat)) == SUBREG | |
6864 | && GET_CODE (SET_SRC (pat)) == SUBREG | |
6865 | && REGNO (SUBREG_REG (SET_DEST (slot_pat))) == | |
6866 | REGNO (SUBREG_REG (SET_SRC (pat))) | |
6867 | && SUBREG_WORD (SET_DEST (slot_pat)) == | |
ddf80874 | 6868 | SUBREG_WORD (SET_SRC (pat))))) |
3bc8b61e DM |
6869 | || (check_fpmode_conflict == 1 |
6870 | && GET_CODE (slot_insn) == INSN | |
6871 | && GET_CODE (slot_pat) == SET | |
ddf80874 DM |
6872 | && (GET_MODE (SET_DEST (slot_pat)) == SFmode |
6873 | || GET_MODE (SET_DEST (slot_pat)) == DFmode) | |
6874 | && GET_MODE (SET_DEST (slot_pat)) != fpmode))) | |
3bc8b61e DM |
6875 | goto next; |
6876 | } | |
6877 | ||
6878 | /* Check for peculiar result availability and dispatch | |
6879 | interference situations. */ | |
6880 | if (pat != 0 | |
6881 | && ultra_cycles_elapsed > 0) | |
6882 | { | |
6883 | rtx link; | |
6884 | ||
6885 | for (link = LOG_LINKS (insn); link; link = XEXP (link, 1)) | |
6886 | { | |
6887 | rtx link_insn = XEXP (link, 0); | |
6888 | if (GET_CODE (link_insn) == INSN | |
6889 | && recog_memoized (link_insn) >= 0 | |
6890 | && (TMASK (get_attr_type (link_insn)) & | |
6891 | (TMASK (TYPE_CMOVE) | TMASK (TYPE_FPCMOVE))) | |
6892 | && ! ultra_cmove_results_ready_p (link_insn)) | |
6893 | goto next; | |
6894 | } | |
6895 | ||
6896 | if (check_fpmode_conflict | |
6897 | && ultra_fpmode_conflict_exists (fpmode)) | |
6898 | goto next; | |
6899 | } | |
6900 | ||
6901 | return &list[i]; | |
6902 | } | |
6903 | next: | |
5a5325cb | 6904 | ; |
3bc8b61e DM |
6905 | } |
6906 | return 0; | |
6907 | } | |
6908 | ||
5a5325cb DM |
6909 | static void |
6910 | ultra_build_types_avail (ready, n_ready) | |
6911 | rtx *ready; | |
6912 | int n_ready; | |
6913 | { | |
6914 | int i = n_ready - 1; | |
6915 | ||
6916 | ultra_types_avail = 0; | |
6917 | while(i >= 0) | |
6918 | { | |
6919 | rtx insn = ready[i]; | |
6920 | ||
6921 | if (recog_memoized (insn) >= 0) | |
6922 | ultra_types_avail |= TMASK (get_attr_type (insn)); | |
6923 | ||
6924 | i -= 1; | |
6925 | } | |
6926 | } | |
6927 | ||
3bc8b61e DM |
6928 | /* Place insn pointed to my IP into the pipeline. |
6929 | Make element THIS of READY be that insn if it | |
6930 | is not already. TYPE indicates the pipeline class | |
6931 | this insn falls into. */ | |
6932 | static void | |
6933 | ultra_schedule_insn (ip, ready, this, type) | |
6934 | rtx *ip; | |
6935 | rtx *ready; | |
6936 | int this; | |
6937 | enum ultra_code type; | |
6938 | { | |
6939 | int pipe_slot; | |
6940 | char mask = ultra_pipe.free_slot_mask; | |
6941 | ||
6942 | /* Obtain free slot. */ | |
6943 | for (pipe_slot = 0; pipe_slot < 4; pipe_slot++) | |
6944 | if ((mask & (1 << pipe_slot)) != 0) | |
6945 | break; | |
6946 | if (pipe_slot == 4) | |
6947 | abort (); | |
6948 | ||
6949 | /* In it goes, and it hasn't been committed yet. */ | |
6950 | ultra_pipe.group[pipe_slot] = *ip; | |
6951 | ultra_pipe.codes[pipe_slot] = type; | |
6952 | ultra_pipe.contents[type] = 1; | |
6953 | if (UMASK (type) & | |
6954 | (UMASK (IEUN) | UMASK (IEU0) | UMASK (IEU1))) | |
6955 | ultra_pipe.num_ieu_insns += 1; | |
6956 | ||
6957 | ultra_pipe.free_slot_mask = (mask & ~(1 << pipe_slot)); | |
6958 | ultra_pipe.group_size += 1; | |
6959 | ultra_pipe.commit[pipe_slot] = 0; | |
6960 | ||
6961 | /* Update ready list. */ | |
6962 | if (ip != &ready[this]) | |
6963 | { | |
6964 | rtx temp = *ip; | |
6965 | ||
6966 | *ip = ready[this]; | |
6967 | ready[this] = temp; | |
6968 | } | |
6969 | } | |
6970 | ||
6971 | /* Advance to the next pipeline group. */ | |
6972 | static void | |
6973 | ultra_flush_pipeline () | |
6974 | { | |
ddf80874 | 6975 | ultra_cur_hist = (ultra_cur_hist + 1) % (ULTRA_NUM_HIST - 1); |
3bc8b61e DM |
6976 | ultra_cycles_elapsed += 1; |
6977 | bzero ((char *) &ultra_pipe, sizeof ultra_pipe); | |
6978 | ultra_pipe.free_slot_mask = 0xf; | |
6979 | } | |
6980 | ||
3bc8b61e DM |
6981 | /* Init our data structures for this current block. */ |
6982 | void | |
6983 | ultrasparc_sched_init (dump, sched_verbose) | |
ddf80874 DM |
6984 | FILE *dump ATTRIBUTE_UNUSED; |
6985 | int sched_verbose ATTRIBUTE_UNUSED; | |
3bc8b61e | 6986 | { |
b96a5ea5 | 6987 | bzero ((char *) ultra_pipe_hist, sizeof ultra_pipe_hist); |
3bc8b61e DM |
6988 | ultra_cur_hist = 0; |
6989 | ultra_cycles_elapsed = 0; | |
ddf80874 | 6990 | ultra_pipe.free_slot_mask = 0xf; |
3bc8b61e DM |
6991 | } |
6992 | ||
6993 | /* INSN has been scheduled, update pipeline commit state | |
6994 | and return how many instructions are still to be | |
6995 | scheduled in this group. */ | |
bfd6bc60 | 6996 | int |
3bc8b61e DM |
6997 | ultrasparc_variable_issue (insn) |
6998 | rtx insn; | |
6999 | { | |
7000 | struct ultrasparc_pipeline_state *up = &ultra_pipe; | |
7001 | int i, left_to_fire; | |
7002 | ||
7003 | left_to_fire = 0; | |
7004 | for (i = 0; i < 4; i++) | |
7005 | { | |
7006 | if (up->group[i] == 0) | |
7007 | continue; | |
7008 | ||
7009 | if (up->group[i] == insn) | |
7010 | { | |
7011 | up->commit[i] = 1; | |
7012 | } | |
7013 | else if (! up->commit[i]) | |
7014 | left_to_fire++; | |
7015 | } | |
7016 | ||
7017 | return left_to_fire; | |
7018 | } | |
7019 | ||
7020 | /* In actual_hazard_this_instance, we may have yanked some | |
7021 | instructions from the ready list due to conflict cost | |
7022 | adjustments. If so, and such an insn was in our pipeline | |
7023 | group, remove it and update state. */ | |
7024 | static void | |
7025 | ultra_rescan_pipeline_state (ready, n_ready) | |
7026 | rtx *ready; | |
7027 | int n_ready; | |
7028 | { | |
7029 | struct ultrasparc_pipeline_state *up = &ultra_pipe; | |
7030 | int i; | |
7031 | ||
7032 | for (i = 0; i < 4; i++) | |
7033 | { | |
7034 | rtx insn = up->group[i]; | |
3bc8b61e DM |
7035 | int j; |
7036 | ||
7037 | if (! insn) | |
7038 | continue; | |
7039 | ||
7040 | /* If it has been committed, then it was removed from | |
7041 | the ready list because it was actually scheduled, | |
7042 | and that is not the case we are searching for here. */ | |
7043 | if (up->commit[i] != 0) | |
7044 | continue; | |
7045 | ||
7046 | for (j = n_ready - 1; j >= 0; j--) | |
7047 | if (ready[j] == insn) | |
7048 | break; | |
7049 | ||
7050 | /* If we didn't find it, toss it. */ | |
7051 | if (j < 0) | |
7052 | { | |
7053 | enum ultra_code ucode = up->codes[i]; | |
7054 | ||
7055 | up->group[i] = 0; | |
7056 | up->codes[i] = NONE; | |
7057 | up->contents[ucode] = 0; | |
7058 | if (UMASK (ucode) & | |
7059 | (UMASK (IEUN) | UMASK (IEU0) | UMASK (IEU1))) | |
7060 | up->num_ieu_insns -= 1; | |
7061 | ||
7062 | up->free_slot_mask |= (1 << i); | |
7063 | up->group_size -= 1; | |
7064 | up->commit[i] = 0; | |
7065 | } | |
7066 | } | |
7067 | } | |
7068 | ||
7069 | void | |
7070 | ultrasparc_sched_reorder (dump, sched_verbose, ready, n_ready) | |
7071 | FILE *dump; | |
7072 | int sched_verbose; | |
7073 | rtx *ready; | |
7074 | int n_ready; | |
7075 | { | |
7076 | struct ultrasparc_pipeline_state *up = &ultra_pipe; | |
7077 | int i, this_insn; | |
7078 | ||
3bc8b61e DM |
7079 | if (sched_verbose) |
7080 | { | |
7081 | int n; | |
7082 | ||
7083 | fprintf (dump, "\n;;\tUltraSPARC Looking at ["); | |
7084 | for (n = n_ready - 1; n >= 0; n--) | |
7085 | { | |
7086 | rtx insn = ready[n]; | |
7087 | enum ultra_code ucode; | |
7088 | ||
7089 | if (recog_memoized (insn) < 0) | |
7090 | continue; | |
7091 | ucode = ultra_code_from_mask (TMASK (get_attr_type (insn))); | |
7092 | if (n != 0) | |
7093 | fprintf (dump, "%s(%d) ", | |
7094 | ultra_code_names[ucode], | |
7095 | INSN_UID (insn)); | |
7096 | else | |
7097 | fprintf (dump, "%s(%d)", | |
7098 | ultra_code_names[ucode], | |
7099 | INSN_UID (insn)); | |
7100 | } | |
7101 | fprintf (dump, "]\n"); | |
7102 | } | |
7103 | ||
7104 | this_insn = n_ready - 1; | |
7105 | ||
7106 | /* Skip over junk we don't understand. */ | |
7107 | while ((this_insn >= 0) | |
7108 | && recog_memoized (ready[this_insn]) < 0) | |
7109 | this_insn--; | |
7110 | ||
5a5325cb DM |
7111 | ultra_build_types_avail (ready, this_insn + 1); |
7112 | ||
3bc8b61e DM |
7113 | while (this_insn >= 0) { |
7114 | int old_group_size = up->group_size; | |
7115 | ||
7116 | if (up->group_size != 0) | |
7117 | { | |
7118 | int num_committed; | |
7119 | ||
7120 | num_committed = (up->commit[0] + up->commit[1] + | |
7121 | up->commit[2] + up->commit[3]); | |
7122 | /* If nothing has been commited from our group, or all of | |
7123 | them have. Clear out the (current cycle's) pipeline | |
7124 | state and start afresh. */ | |
7125 | if (num_committed == 0 | |
7126 | || num_committed == up->group_size) | |
7127 | { | |
ddf80874 DM |
7128 | ultra_flush_pipeline (); |
7129 | up = &ultra_pipe; | |
3bc8b61e DM |
7130 | old_group_size = 0; |
7131 | } | |
7132 | else | |
7133 | { | |
7134 | /* OK, some ready list insns got requeued and thus removed | |
7135 | from the ready list. Account for this fact. */ | |
7136 | ultra_rescan_pipeline_state (ready, n_ready); | |
7137 | ||
7138 | /* Something "changed", make this look like a newly | |
7139 | formed group so the code at the end of the loop | |
7140 | knows that progress was in fact made. */ | |
7141 | if (up->group_size != old_group_size) | |
ddf80874 | 7142 | old_group_size = 0; |
3bc8b61e DM |
7143 | } |
7144 | } | |
7145 | ||
7146 | if (up->group_size == 0) | |
7147 | { | |
7148 | /* If the pipeline is (still) empty and we have any single | |
7149 | group insns, get them out now as this is a good time. */ | |
7150 | rtx *ip = ultra_find_type ((TMASK (TYPE_RETURN) | TMASK (TYPE_ADDRESS) | | |
7151 | TMASK (TYPE_IMUL) | TMASK (TYPE_CMOVE) | | |
7152 | TMASK (TYPE_MULTI) | TMASK (TYPE_MISC)), | |
7153 | ready, this_insn); | |
7154 | if (ip) | |
7155 | { | |
7156 | ultra_schedule_insn (ip, ready, this_insn, SINGLE); | |
7157 | break; | |
7158 | } | |
7159 | ||
7160 | /* If we are not in the process of emptying out the pipe, try to | |
7161 | obtain an instruction which must be the first in it's group. */ | |
7162 | ip = ultra_find_type ((TMASK (TYPE_CALL) | | |
7163 | TMASK (TYPE_CALL_NO_DELAY_SLOT) | | |
7164 | TMASK (TYPE_UNCOND_BRANCH)), | |
7165 | ready, this_insn); | |
7166 | if (ip) | |
7167 | { | |
7168 | ultra_schedule_insn (ip, ready, this_insn, IEU1); | |
7169 | this_insn--; | |
7170 | } | |
7171 | else if ((ip = ultra_find_type ((TMASK (TYPE_FPDIVS) | | |
7172 | TMASK (TYPE_FPDIVD) | | |
7173 | TMASK (TYPE_FPSQRT)), | |
7174 | ready, this_insn)) != 0) | |
7175 | { | |
7176 | ultra_schedule_insn (ip, ready, this_insn, FPM); | |
7177 | this_insn--; | |
7178 | } | |
7179 | } | |
7180 | ||
7181 | /* Try to fill the integer pipeline. First, look for an IEU0 specific | |
7182 | operation. We can't do more IEU operations if the first 3 slots are | |
7183 | all full or we have dispatched two IEU insns already. */ | |
7184 | if ((up->free_slot_mask & 0x7) != 0 | |
7185 | && up->num_ieu_insns < 2 | |
7186 | && up->contents[IEU0] == 0 | |
7187 | && up->contents[IEUN] == 0) | |
7188 | { | |
7189 | rtx *ip = ultra_find_type (TMASK(TYPE_SHIFT), ready, this_insn); | |
7190 | if (ip) | |
7191 | { | |
7192 | ultra_schedule_insn (ip, ready, this_insn, IEU0); | |
7193 | this_insn--; | |
7194 | } | |
7195 | } | |
7196 | ||
7197 | /* If we can, try to find an IEU1 specific or an unnamed | |
7198 | IEU instruction. */ | |
7199 | if ((up->free_slot_mask & 0x7) != 0 | |
7200 | && up->num_ieu_insns < 2) | |
7201 | { | |
7202 | rtx *ip = ultra_find_type ((TMASK (TYPE_IALU) | TMASK (TYPE_BINARY) | | |
7203 | TMASK (TYPE_MOVE) | TMASK (TYPE_UNARY) | | |
7204 | (up->contents[IEU1] == 0 ? TMASK (TYPE_COMPARE) : 0)), | |
7205 | ready, this_insn); | |
7206 | if (ip) | |
7207 | { | |
7208 | rtx insn = *ip; | |
7209 | ||
7210 | ultra_schedule_insn (ip, ready, this_insn, | |
7211 | (!up->contents[IEU1] | |
7212 | && get_attr_type (insn) == TYPE_COMPARE) | |
7213 | ? IEU1 : IEUN); | |
7214 | this_insn--; | |
7215 | } | |
7216 | } | |
7217 | ||
7218 | /* If only one IEU insn has been found, try to find another unnamed | |
7219 | IEU operation or an IEU1 specific one. */ | |
7220 | if ((up->free_slot_mask & 0x7) != 0 | |
7221 | && up->num_ieu_insns < 2) | |
7222 | { | |
7223 | rtx *ip; | |
7224 | int tmask = (TMASK (TYPE_IALU) | TMASK (TYPE_BINARY) | | |
7225 | TMASK (TYPE_MOVE) | TMASK (TYPE_UNARY)); | |
7226 | ||
7227 | if (!up->contents[IEU1]) | |
7228 | tmask |= TMASK (TYPE_COMPARE); | |
7229 | ip = ultra_find_type (tmask, ready, this_insn); | |
7230 | if (ip) | |
7231 | { | |
7232 | rtx insn = *ip; | |
7233 | ||
7234 | ultra_schedule_insn (ip, ready, this_insn, | |
7235 | (!up->contents[IEU1] | |
7236 | && get_attr_type (insn) == TYPE_COMPARE) | |
7237 | ? IEU1 : IEUN); | |
7238 | this_insn--; | |
7239 | } | |
7240 | } | |
7241 | ||
7242 | /* Try for a load or store, but such an insn can only be issued | |
7243 | if it is within' one of the first 3 slots. */ | |
7244 | if ((up->free_slot_mask & 0x7) != 0 | |
7245 | && up->contents[LSU] == 0) | |
7246 | { | |
7247 | rtx *ip = ultra_find_type ((TMASK (TYPE_LOAD) | TMASK (TYPE_SLOAD) | | |
7248 | TMASK (TYPE_STORE) | TMASK (TYPE_FPLOAD) | | |
7249 | TMASK (TYPE_FPSTORE)), ready, this_insn); | |
7250 | if (ip) | |
7251 | { | |
7252 | ultra_schedule_insn (ip, ready, this_insn, LSU); | |
7253 | this_insn--; | |
7254 | } | |
7255 | } | |
7256 | ||
7257 | /* Now find FPU operations, first FPM class. But not divisions or | |
7258 | square-roots because those will break the group up. Unlike all | |
7259 | the previous types, these can go in any slot. */ | |
7260 | if (up->free_slot_mask != 0 | |
7261 | && up->contents[FPM] == 0) | |
7262 | { | |
7263 | rtx *ip = ultra_find_type (TMASK (TYPE_FPMUL), ready, this_insn); | |
7264 | if (ip) | |
7265 | { | |
7266 | ultra_schedule_insn (ip, ready, this_insn, FPM); | |
7267 | this_insn--; | |
7268 | } | |
7269 | } | |
7270 | ||
7271 | /* Continue on with FPA class if we have not filled the group already. */ | |
7272 | if (up->free_slot_mask != 0 | |
7273 | && up->contents[FPA] == 0) | |
7274 | { | |
7275 | rtx *ip = ultra_find_type ((TMASK (TYPE_FPMOVE) | TMASK (TYPE_FPCMOVE) | | |
7276 | TMASK (TYPE_FP) | TMASK (TYPE_FPCMP)), | |
7277 | ready, this_insn); | |
7278 | if (ip) | |
7279 | { | |
7280 | ultra_schedule_insn (ip, ready, this_insn, FPA); | |
7281 | this_insn--; | |
7282 | } | |
7283 | } | |
7284 | ||
7285 | /* Finally, maybe stick a branch in here. */ | |
7286 | if (up->free_slot_mask != 0 | |
7287 | && up->contents[CTI] == 0) | |
7288 | { | |
7289 | rtx *ip = ultra_find_type (TMASK (TYPE_BRANCH), ready, this_insn); | |
7290 | ||
7291 | /* Try to slip in a branch only if it is one of the | |
7292 | next 2 in the ready list. */ | |
7293 | if (ip && ((&ready[this_insn] - ip) < 2)) | |
7294 | { | |
7295 | ultra_schedule_insn (ip, ready, this_insn, CTI); | |
7296 | this_insn--; | |
7297 | } | |
7298 | } | |
7299 | ||
7300 | up->group_size = 0; | |
7301 | for (i = 0; i < 4; i++) | |
7302 | if ((up->free_slot_mask & (1 << i)) == 0) | |
7303 | up->group_size++; | |
7304 | ||
7305 | /* See if we made any progress... */ | |
7306 | if (old_group_size != up->group_size) | |
7307 | break; | |
7308 | ||
7309 | /* Clean out the (current cycle's) pipeline state | |
ddf80874 DM |
7310 | and try once more. If we placed no instructions |
7311 | into the pipeline at all, it means a real hard | |
7312 | conflict exists with some earlier issued instruction | |
7313 | so we must advance to the next cycle to clear it up. */ | |
7314 | if (up->group_size == 0) | |
7315 | { | |
7316 | ultra_flush_pipeline (); | |
7317 | up = &ultra_pipe; | |
7318 | } | |
7319 | else | |
7320 | { | |
7321 | bzero ((char *) &ultra_pipe, sizeof ultra_pipe); | |
7322 | ultra_pipe.free_slot_mask = 0xf; | |
7323 | } | |
3bc8b61e DM |
7324 | } |
7325 | ||
7326 | if (sched_verbose) | |
7327 | { | |
7328 | int n, gsize; | |
7329 | ||
7330 | fprintf (dump, ";;\tUltraSPARC Launched ["); | |
7331 | gsize = up->group_size; | |
7332 | for (n = 0; n < 4; n++) | |
7333 | { | |
7334 | rtx insn = up->group[n]; | |
7335 | ||
7336 | if (! insn) | |
7337 | continue; | |
7338 | ||
7339 | gsize -= 1; | |
7340 | if (gsize != 0) | |
7341 | fprintf (dump, "%s(%d) ", | |
7342 | ultra_code_names[up->codes[n]], | |
7343 | INSN_UID (insn)); | |
7344 | else | |
7345 | fprintf (dump, "%s(%d)", | |
7346 | ultra_code_names[up->codes[n]], | |
7347 | INSN_UID (insn)); | |
7348 | } | |
7349 | fprintf (dump, "]\n"); | |
7350 | } | |
7351 | } | |
7352 | ||
bfd6bc60 JC |
7353 | int |
7354 | sparc_issue_rate () | |
7355 | { | |
7356 | switch (sparc_cpu) | |
7357 | { | |
7358 | default: | |
7359 | return 1; | |
bfd6bc60 | 7360 | case PROCESSOR_V9: |
284d86e9 | 7361 | /* Assume V9 processors are capable of at least dual-issue. */ |
bfd6bc60 JC |
7362 | return 2; |
7363 | case PROCESSOR_SUPERSPARC: | |
7364 | return 3; | |
393049a9 RH |
7365 | case PROCESSOR_HYPERSPARC: |
7366 | case PROCESSOR_SPARCLITE86X: | |
7367 | return 2; | |
bfd6bc60 JC |
7368 | case PROCESSOR_ULTRASPARC: |
7369 | return 4; | |
7370 | } | |
7371 | } | |
284d86e9 JC |
7372 | |
7373 | static int | |
7374 | set_extends(x, insn) | |
7375 | rtx x, insn; | |
7376 | { | |
7377 | register rtx pat = PATTERN (insn); | |
7378 | ||
7379 | switch (GET_CODE (SET_SRC (pat))) | |
7380 | { | |
7381 | /* Load and some shift instructions zero extend. */ | |
7382 | case MEM: | |
7383 | case ZERO_EXTEND: | |
7384 | /* sethi clears the high bits */ | |
7385 | case HIGH: | |
7386 | /* LO_SUM is used with sethi. sethi cleared the high | |
7387 | bits and the values used with lo_sum are positive */ | |
7388 | case LO_SUM: | |
284d86e9 JC |
7389 | /* Store flag stores 0 or 1 */ |
7390 | case LT: case LTU: | |
7391 | case GT: case GTU: | |
7392 | case LE: case LEU: | |
7393 | case GE: case GEU: | |
7394 | case EQ: | |
7395 | case NE: | |
7396 | return 1; | |
7397 | case AND: | |
7398 | { | |
7399 | rtx op1 = XEXP (SET_SRC (pat), 1); | |
7400 | if (GET_CODE (op1) == CONST_INT) | |
7401 | return INTVAL (op1) >= 0; | |
7402 | if (GET_CODE (XEXP (SET_SRC (pat), 0)) == REG | |
7403 | && sparc_check_64 (XEXP (SET_SRC (pat), 0), insn) == 1) | |
7404 | return 1; | |
7405 | if (GET_CODE (op1) == REG | |
7406 | && sparc_check_64 ((op1), insn) == 1) | |
7407 | return 1; | |
7408 | } | |
7409 | case ASHIFT: | |
7410 | case LSHIFTRT: | |
7411 | return GET_MODE (SET_SRC (pat)) == SImode; | |
7412 | /* Positive integers leave the high bits zero. */ | |
7413 | case CONST_DOUBLE: | |
7414 | return ! (CONST_DOUBLE_LOW (x) & 0x80000000); | |
7415 | case CONST_INT: | |
7416 | return ! (INTVAL (x) & 0x80000000); | |
7417 | case ASHIFTRT: | |
7418 | case SIGN_EXTEND: | |
7419 | return - (GET_MODE (SET_SRC (pat)) == SImode); | |
7420 | default: | |
7421 | return 0; | |
7422 | } | |
7423 | } | |
7424 | ||
e0d80184 DM |
7425 | /* We _ought_ to have only one kind per function, but... */ |
7426 | static rtx sparc_addr_diff_list; | |
7427 | static rtx sparc_addr_list; | |
7428 | ||
7429 | void | |
7430 | sparc_defer_case_vector (lab, vec, diff) | |
7431 | rtx lab, vec; | |
7432 | int diff; | |
7433 | { | |
7434 | vec = gen_rtx_EXPR_LIST (VOIDmode, lab, vec); | |
7435 | if (diff) | |
7436 | sparc_addr_diff_list | |
7437 | = gen_rtx_EXPR_LIST (VOIDmode, vec, sparc_addr_diff_list); | |
7438 | else | |
7439 | sparc_addr_list = gen_rtx_EXPR_LIST (VOIDmode, vec, sparc_addr_list); | |
7440 | } | |
7441 | ||
7442 | static void | |
7443 | sparc_output_addr_vec (vec) | |
7444 | rtx vec; | |
7445 | { | |
7446 | rtx lab = XEXP (vec, 0), body = XEXP (vec, 1); | |
7447 | int idx, vlen = XVECLEN (body, 0); | |
7448 | ||
d1accaa3 JJ |
7449 | #ifdef ASM_OUTPUT_ADDR_VEC_START |
7450 | ASM_OUTPUT_ADDR_VEC_START (asm_out_file); | |
7451 | #endif | |
7452 | ||
e0d80184 DM |
7453 | #ifdef ASM_OUTPUT_CASE_LABEL |
7454 | ASM_OUTPUT_CASE_LABEL (asm_out_file, "L", CODE_LABEL_NUMBER (lab), | |
7455 | NEXT_INSN (lab)); | |
7456 | #else | |
7457 | ASM_OUTPUT_INTERNAL_LABEL (asm_out_file, "L", CODE_LABEL_NUMBER (lab)); | |
7458 | #endif | |
7459 | ||
7460 | for (idx = 0; idx < vlen; idx++) | |
7461 | { | |
7462 | ASM_OUTPUT_ADDR_VEC_ELT | |
7463 | (asm_out_file, CODE_LABEL_NUMBER (XEXP (XVECEXP (body, 0, idx), 0))); | |
7464 | } | |
d1accaa3 JJ |
7465 | |
7466 | #ifdef ASM_OUTPUT_ADDR_VEC_END | |
7467 | ASM_OUTPUT_ADDR_VEC_END (asm_out_file); | |
7468 | #endif | |
e0d80184 DM |
7469 | } |
7470 | ||
7471 | static void | |
7472 | sparc_output_addr_diff_vec (vec) | |
7473 | rtx vec; | |
7474 | { | |
7475 | rtx lab = XEXP (vec, 0), body = XEXP (vec, 1); | |
7476 | rtx base = XEXP (XEXP (body, 0), 0); | |
7477 | int idx, vlen = XVECLEN (body, 1); | |
7478 | ||
d1accaa3 JJ |
7479 | #ifdef ASM_OUTPUT_ADDR_VEC_START |
7480 | ASM_OUTPUT_ADDR_VEC_START (asm_out_file); | |
7481 | #endif | |
7482 | ||
e0d80184 DM |
7483 | #ifdef ASM_OUTPUT_CASE_LABEL |
7484 | ASM_OUTPUT_CASE_LABEL (asm_out_file, "L", CODE_LABEL_NUMBER (lab), | |
7485 | NEXT_INSN (lab)); | |
7486 | #else | |
7487 | ASM_OUTPUT_INTERNAL_LABEL (asm_out_file, "L", CODE_LABEL_NUMBER (lab)); | |
7488 | #endif | |
7489 | ||
7490 | for (idx = 0; idx < vlen; idx++) | |
7491 | { | |
7492 | ASM_OUTPUT_ADDR_DIFF_ELT | |
7493 | (asm_out_file, | |
7494 | body, | |
7495 | CODE_LABEL_NUMBER (XEXP (XVECEXP (body, 1, idx), 0)), | |
7496 | CODE_LABEL_NUMBER (base)); | |
7497 | } | |
d1accaa3 JJ |
7498 | |
7499 | #ifdef ASM_OUTPUT_ADDR_VEC_END | |
7500 | ASM_OUTPUT_ADDR_VEC_END (asm_out_file); | |
7501 | #endif | |
e0d80184 DM |
7502 | } |
7503 | ||
7504 | static void | |
7505 | sparc_output_deferred_case_vectors () | |
7506 | { | |
7507 | rtx t; | |
d676da68 | 7508 | int align; |
e0d80184 | 7509 | |
f3b8847b DM |
7510 | if (sparc_addr_list == NULL_RTX |
7511 | && sparc_addr_diff_list == NULL_RTX) | |
7512 | return; | |
7513 | ||
e0d80184 DM |
7514 | /* Align to cache line in the function's code section. */ |
7515 | function_section (current_function_decl); | |
d676da68 DM |
7516 | |
7517 | align = floor_log2 (FUNCTION_BOUNDARY / BITS_PER_UNIT); | |
7518 | if (align > 0) | |
f3b8847b | 7519 | ASM_OUTPUT_ALIGN (asm_out_file, align); |
e0d80184 DM |
7520 | |
7521 | for (t = sparc_addr_list; t ; t = XEXP (t, 1)) | |
7522 | sparc_output_addr_vec (XEXP (t, 0)); | |
7523 | for (t = sparc_addr_diff_list; t ; t = XEXP (t, 1)) | |
7524 | sparc_output_addr_diff_vec (XEXP (t, 0)); | |
7525 | ||
7526 | sparc_addr_list = sparc_addr_diff_list = NULL_RTX; | |
7527 | } | |
7528 | ||
284d86e9 JC |
7529 | /* Return 0 if the high 32 bits of X (the low word of X, if DImode) are |
7530 | unknown. Return 1 if the high bits are zero, -1 if the register is | |
7531 | sign extended. */ | |
7532 | int | |
7533 | sparc_check_64 (x, insn) | |
7534 | rtx x, insn; | |
7535 | { | |
7536 | /* If a register is set only once it is safe to ignore insns this | |
7537 | code does not know how to handle. The loop will either recognize | |
7538 | the single set and return the correct value or fail to recognize | |
7539 | it and return 0. */ | |
7540 | int set_once = 0; | |
7541 | ||
7542 | if (GET_CODE (x) == REG | |
7543 | && flag_expensive_optimizations | |
7544 | && REG_N_SETS (REGNO (x)) == 1) | |
7545 | set_once = 1; | |
7546 | ||
7547 | if (insn == 0) | |
d6f4ec51 KG |
7548 | { |
7549 | if (set_once) | |
7550 | insn = get_last_insn_anywhere (); | |
7551 | else | |
7552 | return 0; | |
7553 | } | |
284d86e9 | 7554 | |
d6f4ec51 | 7555 | while ((insn = PREV_INSN (insn))) |
284d86e9 JC |
7556 | { |
7557 | switch (GET_CODE (insn)) | |
7558 | { | |
7559 | case JUMP_INSN: | |
7560 | case NOTE: | |
7561 | break; | |
7562 | case CODE_LABEL: | |
7563 | case CALL_INSN: | |
7564 | default: | |
7565 | if (! set_once) | |
7566 | return 0; | |
7567 | break; | |
7568 | case INSN: | |
7569 | { | |
7570 | rtx pat = PATTERN (insn); | |
7571 | if (GET_CODE (pat) != SET) | |
7572 | return 0; | |
7573 | if (rtx_equal_p (x, SET_DEST (pat))) | |
7574 | return set_extends (x, insn); | |
7575 | if (reg_overlap_mentioned_p (SET_DEST (pat), x)) | |
7576 | return 0; | |
7577 | } | |
7578 | } | |
7579 | } | |
7580 | return 0; | |
7581 | } | |
7582 | ||
7583 | char * | |
7584 | sparc_v8plus_shift (operands, insn, opcode) | |
7585 | rtx *operands; | |
7586 | rtx insn; | |
b1474bb7 | 7587 | const char *opcode; |
284d86e9 JC |
7588 | { |
7589 | static char asm_code[60]; | |
7590 | ||
7591 | if (GET_CODE (operands[3]) == SCRATCH) | |
7592 | operands[3] = operands[0]; | |
c6b0465b JC |
7593 | if (GET_CODE (operands[1]) == CONST_INT) |
7594 | { | |
7595 | output_asm_insn ("mov %1,%3", operands); | |
7596 | } | |
7597 | else | |
7598 | { | |
7599 | output_asm_insn ("sllx %H1,32,%3", operands); | |
7600 | if (sparc_check_64 (operands[1], insn) <= 0) | |
7601 | output_asm_insn ("srl %L1,0,%L1", operands); | |
7602 | output_asm_insn ("or %L1,%3,%3", operands); | |
7603 | } | |
284d86e9 JC |
7604 | |
7605 | strcpy(asm_code, opcode); | |
7606 | if (which_alternative != 2) | |
7607 | return strcat (asm_code, " %0,%2,%L0\n\tsrlx %L0,32,%H0"); | |
7608 | else | |
7609 | return strcat (asm_code, " %3,%2,%3\n\tsrlx %3,32,%H0\n\tmov %3,%L0"); | |
7610 | } | |
7611 | ||
7612 | ||
7613 | /* Return 1 if DEST and SRC reference only global and in registers. */ | |
7614 | ||
7615 | int | |
7616 | sparc_return_peephole_ok (dest, src) | |
7617 | rtx dest, src; | |
7618 | { | |
7619 | if (! TARGET_V9) | |
7620 | return 0; | |
54ff41b7 | 7621 | if (current_function_uses_only_leaf_regs) |
284d86e9 JC |
7622 | return 0; |
7623 | if (GET_CODE (src) != CONST_INT | |
7624 | && (GET_CODE (src) != REG || ! IN_OR_GLOBAL_P (src))) | |
7625 | return 0; | |
7626 | return IN_OR_GLOBAL_P (dest); | |
7627 | } | |
2be15d0f RH |
7628 | \f |
7629 | /* Output assembler code to FILE to increment profiler label # LABELNO | |
7630 | for profiling a function entry. | |
7631 | ||
7632 | 32 bit sparc uses %g2 as the STATIC_CHAIN_REGNUM which gets clobbered | |
7633 | during profiling so we need to save/restore it around the call to mcount. | |
7634 | We're guaranteed that a save has just been done, and we use the space | |
7635 | allocated for intreg/fpreg value passing. */ | |
7636 | ||
7637 | void | |
7638 | sparc_function_profiler (file, labelno) | |
7639 | FILE *file; | |
7640 | int labelno; | |
7641 | { | |
7642 | char buf[32]; | |
7643 | ASM_GENERATE_INTERNAL_LABEL (buf, "LP", labelno); | |
7644 | ||
7645 | if (! TARGET_ARCH64) | |
7646 | fputs ("\tst\t%g2,[%fp-4]\n", file); | |
7647 | ||
7648 | fputs ("\tsethi\t%hi(", file); | |
7649 | assemble_name (file, buf); | |
7650 | fputs ("),%o0\n", file); | |
7651 | ||
7652 | fputs ("\tcall\t", file); | |
7653 | assemble_name (file, MCOUNT_FUNCTION); | |
7654 | putc ('\n', file); | |
7655 | ||
7656 | fputs ("\t or\t%o0,%lo(", file); | |
7657 | assemble_name (file, buf); | |
7658 | fputs ("),%o0\n", file); | |
7659 | ||
7660 | if (! TARGET_ARCH64) | |
7661 | fputs ("\tld\t[%fp-4],%g2\n", file); | |
7662 | } | |
7663 | ||
7664 | ||
7665 | /* The following macro shall output assembler code to FILE | |
7666 | to initialize basic-block profiling. | |
7667 | ||
7668 | If profile_block_flag == 2 | |
7669 | ||
7670 | Output code to call the subroutine `__bb_init_trace_func' | |
7671 | and pass two parameters to it. The first parameter is | |
7672 | the address of a block allocated in the object module. | |
7673 | The second parameter is the number of the first basic block | |
7674 | of the function. | |
7675 | ||
7676 | The name of the block is a local symbol made with this statement: | |
7677 | ||
7678 | ASM_GENERATE_INTERNAL_LABEL (BUFFER, "LPBX", 0); | |
7679 | ||
7680 | Of course, since you are writing the definition of | |
7681 | `ASM_GENERATE_INTERNAL_LABEL' as well as that of this macro, you | |
7682 | can take a short cut in the definition of this macro and use the | |
7683 | name that you know will result. | |
7684 | ||
7685 | The number of the first basic block of the function is | |
7686 | passed to the macro in BLOCK_OR_LABEL. | |
7687 | ||
7688 | If described in a virtual assembler language the code to be | |
7689 | output looks like: | |
7690 | ||
7691 | parameter1 <- LPBX0 | |
7692 | parameter2 <- BLOCK_OR_LABEL | |
7693 | call __bb_init_trace_func | |
7694 | ||
7695 | else if profile_block_flag != 0 | |
7696 | ||
7697 | Output code to call the subroutine `__bb_init_func' | |
7698 | and pass one single parameter to it, which is the same | |
7699 | as the first parameter to `__bb_init_trace_func'. | |
7700 | ||
7701 | The first word of this parameter is a flag which will be nonzero if | |
7702 | the object module has already been initialized. So test this word | |
7703 | first, and do not call `__bb_init_func' if the flag is nonzero. | |
7704 | Note: When profile_block_flag == 2 the test need not be done | |
7705 | but `__bb_init_trace_func' *must* be called. | |
7706 | ||
7707 | BLOCK_OR_LABEL may be used to generate a label number as a | |
7708 | branch destination in case `__bb_init_func' will not be called. | |
7709 | ||
7710 | If described in a virtual assembler language the code to be | |
7711 | output looks like: | |
7712 | ||
7713 | cmp (LPBX0),0 | |
7714 | jne local_label | |
7715 | parameter1 <- LPBX0 | |
7716 | call __bb_init_func | |
7717 | local_label: | |
7718 | ||
7719 | */ | |
7720 | ||
7721 | void | |
7722 | sparc_function_block_profiler(file, block_or_label) | |
7723 | FILE *file; | |
7724 | int block_or_label; | |
7725 | { | |
7726 | char LPBX[32]; | |
7727 | ASM_GENERATE_INTERNAL_LABEL (LPBX, "LPBX", 0); | |
7728 | ||
7729 | if (profile_block_flag == 2) | |
7730 | { | |
7731 | fputs ("\tsethi\t%hi(", file); | |
7732 | assemble_name (file, LPBX); | |
7733 | fputs ("),%o0\n", file); | |
7734 | ||
7735 | fprintf (file, "\tsethi\t%%hi(%d),%%o1\n", block_or_label); | |
7736 | ||
7737 | fputs ("\tor\t%o0,%lo(", file); | |
7738 | assemble_name (file, LPBX); | |
7739 | fputs ("),%o0\n", file); | |
7740 | ||
7741 | fprintf (file, "\tcall\t%s__bb_init_trace_func\n", user_label_prefix); | |
7742 | ||
7743 | fprintf (file, "\t or\t%%o1,%%lo(%d),%%o1\n", block_or_label); | |
7744 | } | |
7745 | else if (profile_block_flag != 0) | |
7746 | { | |
7747 | char LPBY[32]; | |
7748 | ASM_GENERATE_INTERNAL_LABEL (LPBY, "LPBY", block_or_label); | |
7749 | ||
7750 | fputs ("\tsethi\t%hi(", file); | |
7751 | assemble_name (file, LPBX); | |
7752 | fputs ("),%o0\n", file); | |
7753 | ||
7754 | fputs ("\tld\t[%lo(", file); | |
7755 | assemble_name (file, LPBX); | |
7756 | fputs (")+%o0],%o1\n", file); | |
7757 | ||
7758 | fputs ("\ttst\t%o1\n", file); | |
7759 | ||
7760 | if (TARGET_V9) | |
7761 | { | |
7762 | fputs ("\tbne,pn\t%icc,", file); | |
7763 | assemble_name (file, LPBY); | |
7764 | putc ('\n', file); | |
7765 | } | |
7766 | else | |
7767 | { | |
7768 | fputs ("\tbne\t", file); | |
7769 | assemble_name (file, LPBY); | |
7770 | putc ('\n', file); | |
7771 | } | |
7772 | ||
7773 | fputs ("\t or\t%o0,%lo(", file); | |
7774 | assemble_name (file, LPBX); | |
7775 | fputs ("),%o0\n", file); | |
7776 | ||
7777 | fprintf (file, "\tcall\t%s__bb_init_func\n\t nop\n", user_label_prefix); | |
7778 | ||
7779 | ASM_OUTPUT_INTERNAL_LABEL (file, "LPBY", block_or_label); | |
7780 | } | |
7781 | } | |
7782 | ||
7783 | /* The following macro shall output assembler code to FILE | |
7784 | to increment a counter associated with basic block number BLOCKNO. | |
7785 | ||
7786 | If profile_block_flag == 2 | |
7787 | ||
7788 | Output code to initialize the global structure `__bb' and | |
7789 | call the function `__bb_trace_func' which will increment the | |
7790 | counter. | |
7791 | ||
7792 | `__bb' consists of two words. In the first word the number | |
7793 | of the basic block has to be stored. In the second word | |
7794 | the address of a block allocated in the object module | |
7795 | has to be stored. | |
7796 | ||
7797 | The basic block number is given by BLOCKNO. | |
7798 | ||
7799 | The address of the block is given by the label created with | |
7800 | ||
7801 | ASM_GENERATE_INTERNAL_LABEL (BUFFER, "LPBX", 0); | |
7802 | ||
7803 | by FUNCTION_BLOCK_PROFILER. | |
7804 | ||
7805 | Of course, since you are writing the definition of | |
7806 | `ASM_GENERATE_INTERNAL_LABEL' as well as that of this macro, you | |
7807 | can take a short cut in the definition of this macro and use the | |
7808 | name that you know will result. | |
7809 | ||
7810 | If described in a virtual assembler language the code to be | |
7811 | output looks like: | |
7812 | ||
7813 | move BLOCKNO -> (__bb) | |
7814 | move LPBX0 -> (__bb+4) | |
7815 | call __bb_trace_func | |
7816 | ||
7817 | Note that function `__bb_trace_func' must not change the | |
7818 | machine state, especially the flag register. To grant | |
7819 | this, you must output code to save and restore registers | |
7820 | either in this macro or in the macros MACHINE_STATE_SAVE | |
7821 | and MACHINE_STATE_RESTORE. The last two macros will be | |
7822 | used in the function `__bb_trace_func', so you must make | |
7823 | sure that the function prologue does not change any | |
7824 | register prior to saving it with MACHINE_STATE_SAVE. | |
7825 | ||
7826 | else if profile_block_flag != 0 | |
7827 | ||
7828 | Output code to increment the counter directly. | |
7829 | Basic blocks are numbered separately from zero within each | |
7830 | compiled object module. The count associated with block number | |
7831 | BLOCKNO is at index BLOCKNO in an array of words; the name of | |
7832 | this array is a local symbol made with this statement: | |
7833 | ||
7834 | ASM_GENERATE_INTERNAL_LABEL (BUFFER, "LPBX", 2); | |
7835 | ||
7836 | Of course, since you are writing the definition of | |
7837 | `ASM_GENERATE_INTERNAL_LABEL' as well as that of this macro, you | |
7838 | can take a short cut in the definition of this macro and use the | |
7839 | name that you know will result. | |
7840 | ||
7841 | If described in a virtual assembler language, the code to be | |
7842 | output looks like: | |
7843 | ||
7844 | inc (LPBX2+4*BLOCKNO) | |
7845 | ||
7846 | */ | |
7847 | ||
7848 | void | |
7849 | sparc_block_profiler(file, blockno) | |
7850 | FILE *file; | |
7851 | int blockno; | |
7852 | { | |
7853 | char LPBX[32]; | |
7854 | ||
7855 | if (profile_block_flag == 2) | |
7856 | { | |
7857 | ASM_GENERATE_INTERNAL_LABEL (LPBX, "LPBX", 0); | |
7858 | ||
7859 | fprintf (file, "\tsethi\t%%hi(%s__bb),%%g1\n", user_label_prefix); | |
7860 | fprintf (file, "\tsethi\t%%hi(%d),%%g2\n", blockno); | |
7861 | fprintf (file, "\tor\t%%g1,%%lo(%s__bb),%%g1\n", user_label_prefix); | |
7862 | fprintf (file, "\tor\t%%g2,%%lo(%d),%%g2\n", blockno); | |
7863 | ||
7864 | fputs ("\tst\t%g2,[%g1]\n", file); | |
7865 | ||
7866 | fputs ("\tsethi\t%hi(", file); | |
7867 | assemble_name (file, LPBX); | |
7868 | fputs ("),%g2\n", file); | |
7869 | ||
110b0302 | 7870 | fputs ("\tor\t%o2,%lo(", file); |
2be15d0f RH |
7871 | assemble_name (file, LPBX); |
7872 | fputs ("),%g2\n", file); | |
7873 | ||
7874 | fputs ("\tst\t%g2,[%g1+4]\n", file); | |
7875 | fputs ("\tmov\t%o7,%g2\n", file); | |
7876 | ||
7877 | fprintf (file, "\tcall\t%s__bb_trace_func\n\t nop\n", user_label_prefix); | |
7878 | ||
7879 | fputs ("\tmov\t%g2,%o7\n", file); | |
7880 | } | |
7881 | else if (profile_block_flag != 0) | |
7882 | { | |
7883 | ASM_GENERATE_INTERNAL_LABEL (LPBX, "LPBX", 2); | |
7884 | ||
7885 | fputs ("\tsethi\t%hi(", file); | |
7886 | assemble_name (file, LPBX); | |
7887 | fprintf (file, "+%d),%%g1\n", blockno*4); | |
7888 | ||
7889 | fputs ("\tld\t[%g1+%lo(", file); | |
7890 | assemble_name (file, LPBX); | |
7891 | fprintf (file, "+%d)],%%g2\n", blockno*4); | |
7892 | ||
7893 | fputs ("\tadd\t%g2,1,%g2\n", file); | |
7894 | ||
7895 | fputs ("\tst\t%g2,[%g1+%lo(", file); | |
7896 | assemble_name (file, LPBX); | |
7897 | fprintf (file, "+%d)]\n", blockno*4); | |
7898 | } | |
7899 | } | |
7900 | ||
7901 | /* The following macro shall output assembler code to FILE | |
7902 | to indicate a return from function during basic-block profiling. | |
7903 | ||
7904 | If profile_block_flag == 2: | |
7905 | ||
7906 | Output assembler code to call function `__bb_trace_ret'. | |
7907 | ||
7908 | Note that function `__bb_trace_ret' must not change the | |
7909 | machine state, especially the flag register. To grant | |
7910 | this, you must output code to save and restore registers | |
7911 | either in this macro or in the macros MACHINE_STATE_SAVE_RET | |
7912 | and MACHINE_STATE_RESTORE_RET. The last two macros will be | |
7913 | used in the function `__bb_trace_ret', so you must make | |
7914 | sure that the function prologue does not change any | |
7915 | register prior to saving it with MACHINE_STATE_SAVE_RET. | |
7916 | ||
7917 | else if profile_block_flag != 0: | |
7918 | ||
7919 | The macro will not be used, so it need not distinguish | |
7920 | these cases. | |
7921 | */ | |
7922 | ||
7923 | void | |
7924 | sparc_function_block_profiler_exit(file) | |
7925 | FILE *file; | |
7926 | { | |
7927 | if (profile_block_flag == 2) | |
7928 | fprintf (file, "\tcall\t%s__bb_trace_ret\n\t nop\n", user_label_prefix); | |
7929 | else | |
7930 | abort (); | |
7931 | } | |
d07d525a MM |
7932 | |
7933 | /* Mark ARG, which is really a struct ultrasparc_pipline_state *, for | |
7934 | GC. */ | |
7935 | ||
7936 | static void | |
7937 | mark_ultrasparc_pipeline_state (arg) | |
7938 | void *arg; | |
7939 | { | |
7940 | struct ultrasparc_pipeline_state *ups; | |
7941 | size_t i; | |
7942 | ||
7943 | ups = (struct ultrasparc_pipeline_state *) arg; | |
7944 | for (i = 0; i < sizeof (ups->group) / sizeof (rtx); ++i) | |
7945 | ggc_mark_rtx (ups->group[i]); | |
7946 | } | |
7947 | ||
7948 | /* Called to register all of our global variables with the garbage | |
7949 | collector. */ | |
7950 | ||
7951 | static void | |
7952 | sparc_add_gc_roots () | |
7953 | { | |
7954 | ggc_add_rtx_root (&sparc_compare_op0, 1); | |
7955 | ggc_add_rtx_root (&sparc_compare_op1, 1); | |
7956 | ggc_add_rtx_root (&leaf_label, 1); | |
7957 | ggc_add_rtx_root (&global_offset_table, 1); | |
7958 | ggc_add_rtx_root (&get_pc_symbol, 1); | |
7959 | ggc_add_rtx_root (&sparc_addr_diff_list, 1); | |
7960 | ggc_add_rtx_root (&sparc_addr_list, 1); | |
7961 | ggc_add_root (ultra_pipe_hist, | |
7962 | sizeof (ultra_pipe_hist) / sizeof (ultra_pipe_hist[0]), | |
7963 | sizeof (ultra_pipe_hist[0]), | |
7964 | &mark_ultrasparc_pipeline_state); | |
7965 | } |