]>
Commit | Line | Data |
---|---|---|
1 | /* Subroutines used for code generation on IA-32. | |
2 | Copyright (C) 1988, 1992, 1994, 1995, 1996, 1997, 1998, 1999, 2000 | |
3 | Free Software Foundation, Inc. | |
4 | ||
5 | This file is part of GNU CC. | |
6 | ||
7 | GNU CC is free software; you can redistribute it and/or modify | |
8 | it under the terms of the GNU General Public License as published by | |
9 | the Free Software Foundation; either version 2, or (at your option) | |
10 | any later version. | |
11 | ||
12 | GNU CC is distributed in the hope that it will be useful, | |
13 | but WITHOUT ANY WARRANTY; without even the implied warranty of | |
14 | MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the | |
15 | GNU General Public License for more details. | |
16 | ||
17 | You should have received a copy of the GNU General Public License | |
18 | along with GNU CC; see the file COPYING. If not, write to | |
19 | the Free Software Foundation, 59 Temple Place - Suite 330, | |
20 | Boston, MA 02111-1307, USA. */ | |
21 | ||
22 | #include <setjmp.h> | |
23 | #include "config.h" | |
24 | #include "system.h" | |
25 | #include "rtl.h" | |
26 | #include "tree.h" | |
27 | #include "tm_p.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" | |
37 | #include "except.h" | |
38 | #include "function.h" | |
39 | #include "recog.h" | |
40 | #include "expr.h" | |
41 | #include "toplev.h" | |
42 | #include "basic-block.h" | |
43 | #include "ggc.h" | |
44 | ||
45 | #ifdef EXTRA_CONSTRAINT | |
46 | /* If EXTRA_CONSTRAINT is defined, then the 'S' | |
47 | constraint in REG_CLASS_FROM_LETTER will no longer work, and various | |
48 | asm statements that need 'S' for class SIREG will break. */ | |
49 | error EXTRA_CONSTRAINT conflicts with S constraint letter | |
50 | /* The previous line used to be #error, but some compilers barf | |
51 | even if the conditional was untrue. */ | |
52 | #endif | |
53 | ||
54 | #ifndef CHECK_STACK_LIMIT | |
55 | #define CHECK_STACK_LIMIT -1 | |
56 | #endif | |
57 | ||
58 | /* Processor costs (relative to an add) */ | |
59 | struct processor_costs i386_cost = { /* 386 specific costs */ | |
60 | 1, /* cost of an add instruction */ | |
61 | 1, /* cost of a lea instruction */ | |
62 | 3, /* variable shift costs */ | |
63 | 2, /* constant shift costs */ | |
64 | 6, /* cost of starting a multiply */ | |
65 | 1, /* cost of multiply per each bit set */ | |
66 | 23, /* cost of a divide/mod */ | |
67 | 15, /* "large" insn */ | |
68 | 3, /* MOVE_RATIO */ | |
69 | 4, /* cost for loading QImode using movzbl */ | |
70 | {2, 4, 2}, /* cost of loading integer registers | |
71 | in QImode, HImode and SImode. | |
72 | Relative to reg-reg move (2). */ | |
73 | {2, 4, 2}, /* cost of storing integer registers */ | |
74 | 2, /* cost of reg,reg fld/fst */ | |
75 | {8, 8, 8}, /* cost of loading fp registers | |
76 | in SFmode, DFmode and XFmode */ | |
77 | {8, 8, 8} /* cost of loading integer registers */ | |
78 | }; | |
79 | ||
80 | struct processor_costs i486_cost = { /* 486 specific costs */ | |
81 | 1, /* cost of an add instruction */ | |
82 | 1, /* cost of a lea instruction */ | |
83 | 3, /* variable shift costs */ | |
84 | 2, /* constant shift costs */ | |
85 | 12, /* cost of starting a multiply */ | |
86 | 1, /* cost of multiply per each bit set */ | |
87 | 40, /* cost of a divide/mod */ | |
88 | 15, /* "large" insn */ | |
89 | 3, /* MOVE_RATIO */ | |
90 | 4, /* cost for loading QImode using movzbl */ | |
91 | {2, 4, 2}, /* cost of loading integer registers | |
92 | in QImode, HImode and SImode. | |
93 | Relative to reg-reg move (2). */ | |
94 | {2, 4, 2}, /* cost of storing integer registers */ | |
95 | 2, /* cost of reg,reg fld/fst */ | |
96 | {8, 8, 8}, /* cost of loading fp registers | |
97 | in SFmode, DFmode and XFmode */ | |
98 | {8, 8, 8} /* cost of loading integer registers */ | |
99 | }; | |
100 | ||
101 | struct processor_costs pentium_cost = { | |
102 | 1, /* cost of an add instruction */ | |
103 | 1, /* cost of a lea instruction */ | |
104 | 4, /* variable shift costs */ | |
105 | 1, /* constant shift costs */ | |
106 | 11, /* cost of starting a multiply */ | |
107 | 0, /* cost of multiply per each bit set */ | |
108 | 25, /* cost of a divide/mod */ | |
109 | 8, /* "large" insn */ | |
110 | 6, /* MOVE_RATIO */ | |
111 | 6, /* cost for loading QImode using movzbl */ | |
112 | {2, 4, 2}, /* cost of loading integer registers | |
113 | in QImode, HImode and SImode. | |
114 | Relative to reg-reg move (2). */ | |
115 | {2, 4, 2}, /* cost of storing integer registers */ | |
116 | 2, /* cost of reg,reg fld/fst */ | |
117 | {2, 2, 6}, /* cost of loading fp registers | |
118 | in SFmode, DFmode and XFmode */ | |
119 | {4, 4, 6} /* cost of loading integer registers */ | |
120 | }; | |
121 | ||
122 | struct processor_costs pentiumpro_cost = { | |
123 | 1, /* cost of an add instruction */ | |
124 | 1, /* cost of a lea instruction */ | |
125 | 1, /* variable shift costs */ | |
126 | 1, /* constant shift costs */ | |
127 | 4, /* cost of starting a multiply */ | |
128 | 0, /* cost of multiply per each bit set */ | |
129 | 17, /* cost of a divide/mod */ | |
130 | 8, /* "large" insn */ | |
131 | 6, /* MOVE_RATIO */ | |
132 | 2, /* cost for loading QImode using movzbl */ | |
133 | {4, 4, 4}, /* cost of loading integer registers | |
134 | in QImode, HImode and SImode. | |
135 | Relative to reg-reg move (2). */ | |
136 | {2, 2, 2}, /* cost of storing integer registers */ | |
137 | 2, /* cost of reg,reg fld/fst */ | |
138 | {2, 2, 6}, /* cost of loading fp registers | |
139 | in SFmode, DFmode and XFmode */ | |
140 | {4, 4, 6} /* cost of loading integer registers */ | |
141 | }; | |
142 | ||
143 | struct processor_costs k6_cost = { | |
144 | 1, /* cost of an add instruction */ | |
145 | 2, /* cost of a lea instruction */ | |
146 | 1, /* variable shift costs */ | |
147 | 1, /* constant shift costs */ | |
148 | 3, /* cost of starting a multiply */ | |
149 | 0, /* cost of multiply per each bit set */ | |
150 | 18, /* cost of a divide/mod */ | |
151 | 8, /* "large" insn */ | |
152 | 4, /* MOVE_RATIO */ | |
153 | 3, /* cost for loading QImode using movzbl */ | |
154 | {4, 5, 4}, /* cost of loading integer registers | |
155 | in QImode, HImode and SImode. | |
156 | Relative to reg-reg move (2). */ | |
157 | {2, 3, 2}, /* cost of storing integer registers */ | |
158 | 4, /* cost of reg,reg fld/fst */ | |
159 | {6, 6, 6}, /* cost of loading fp registers | |
160 | in SFmode, DFmode and XFmode */ | |
161 | {4, 4, 4} /* cost of loading integer registers */ | |
162 | }; | |
163 | ||
164 | struct processor_costs athlon_cost = { | |
165 | 1, /* cost of an add instruction */ | |
166 | 1, /* cost of a lea instruction */ | |
167 | 1, /* variable shift costs */ | |
168 | 1, /* constant shift costs */ | |
169 | 5, /* cost of starting a multiply */ | |
170 | 0, /* cost of multiply per each bit set */ | |
171 | 19, /* cost of a divide/mod */ | |
172 | 8, /* "large" insn */ | |
173 | 9, /* MOVE_RATIO */ | |
174 | 4, /* cost for loading QImode using movzbl */ | |
175 | {4, 5, 4}, /* cost of loading integer registers | |
176 | in QImode, HImode and SImode. | |
177 | Relative to reg-reg move (2). */ | |
178 | {2, 3, 2}, /* cost of storing integer registers */ | |
179 | 4, /* cost of reg,reg fld/fst */ | |
180 | {6, 6, 6}, /* cost of loading fp registers | |
181 | in SFmode, DFmode and XFmode */ | |
182 | {4, 4, 4} /* cost of loading integer registers */ | |
183 | }; | |
184 | ||
185 | struct processor_costs *ix86_cost = &pentium_cost; | |
186 | ||
187 | /* Processor feature/optimization bitmasks. */ | |
188 | #define m_386 (1<<PROCESSOR_I386) | |
189 | #define m_486 (1<<PROCESSOR_I486) | |
190 | #define m_PENT (1<<PROCESSOR_PENTIUM) | |
191 | #define m_PPRO (1<<PROCESSOR_PENTIUMPRO) | |
192 | #define m_K6 (1<<PROCESSOR_K6) | |
193 | #define m_ATHLON (1<<PROCESSOR_ATHLON) | |
194 | ||
195 | const int x86_use_leave = m_386 | m_K6 | m_ATHLON; | |
196 | const int x86_push_memory = m_386 | m_K6 | m_ATHLON; | |
197 | const int x86_zero_extend_with_and = m_486 | m_PENT; | |
198 | const int x86_movx = m_ATHLON | m_PPRO /* m_386 | m_K6 */; | |
199 | const int x86_double_with_add = ~m_386; | |
200 | const int x86_use_bit_test = m_386; | |
201 | const int x86_unroll_strlen = m_486 | m_PENT | m_PPRO | m_ATHLON | m_K6; | |
202 | const int x86_use_q_reg = m_PENT | m_PPRO | m_K6; | |
203 | const int x86_use_any_reg = m_486; | |
204 | const int x86_cmove = m_PPRO | m_ATHLON; | |
205 | const int x86_deep_branch = m_PPRO | m_K6 | m_ATHLON; | |
206 | const int x86_use_sahf = m_PPRO | m_K6 | m_ATHLON; | |
207 | const int x86_partial_reg_stall = m_PPRO; | |
208 | const int x86_use_loop = m_K6; | |
209 | const int x86_use_fiop = ~(m_PPRO | m_ATHLON | m_PENT); | |
210 | const int x86_use_mov0 = m_K6; | |
211 | const int x86_use_cltd = ~(m_PENT | m_K6); | |
212 | const int x86_read_modify_write = ~m_PENT; | |
213 | const int x86_read_modify = ~(m_PENT | m_PPRO); | |
214 | const int x86_split_long_moves = m_PPRO; | |
215 | const int x86_promote_QImode = m_K6 | m_PENT | m_386 | m_486; | |
216 | const int x86_single_stringop = m_386; | |
217 | const int x86_qimode_math = ~(0); | |
218 | const int x86_promote_qi_regs = 0; | |
219 | const int x86_himode_math = ~(m_PPRO); | |
220 | const int x86_promote_hi_regs = m_PPRO; | |
221 | ||
222 | #define AT_BP(mode) (gen_rtx_MEM ((mode), hard_frame_pointer_rtx)) | |
223 | ||
224 | const char * const hi_reg_name[] = HI_REGISTER_NAMES; | |
225 | const char * const qi_reg_name[] = QI_REGISTER_NAMES; | |
226 | const char * const qi_high_reg_name[] = QI_HIGH_REGISTER_NAMES; | |
227 | ||
228 | /* Array of the smallest class containing reg number REGNO, indexed by | |
229 | REGNO. Used by REGNO_REG_CLASS in i386.h. */ | |
230 | ||
231 | enum reg_class const regclass_map[FIRST_PSEUDO_REGISTER] = | |
232 | { | |
233 | /* ax, dx, cx, bx */ | |
234 | AREG, DREG, CREG, BREG, | |
235 | /* si, di, bp, sp */ | |
236 | SIREG, DIREG, NON_Q_REGS, NON_Q_REGS, | |
237 | /* FP registers */ | |
238 | FP_TOP_REG, FP_SECOND_REG, FLOAT_REGS, FLOAT_REGS, | |
239 | FLOAT_REGS, FLOAT_REGS, FLOAT_REGS, FLOAT_REGS, | |
240 | /* arg pointer */ | |
241 | NON_Q_REGS, | |
242 | /* flags, fpsr, dirflag, frame */ | |
243 | NO_REGS, NO_REGS, NO_REGS, NON_Q_REGS | |
244 | }; | |
245 | ||
246 | /* The "default" register map. */ | |
247 | ||
248 | int const dbx_register_map[FIRST_PSEUDO_REGISTER] = | |
249 | { | |
250 | 0, 2, 1, 3, 6, 7, 4, 5, /* general regs */ | |
251 | 12, 13, 14, 15, 16, 17, 18, 19, /* fp regs */ | |
252 | -1, -1, -1, -1, /* arg, flags, fpsr, dir */ | |
253 | }; | |
254 | ||
255 | /* Define the register numbers to be used in Dwarf debugging information. | |
256 | The SVR4 reference port C compiler uses the following register numbers | |
257 | in its Dwarf output code: | |
258 | 0 for %eax (gcc regno = 0) | |
259 | 1 for %ecx (gcc regno = 2) | |
260 | 2 for %edx (gcc regno = 1) | |
261 | 3 for %ebx (gcc regno = 3) | |
262 | 4 for %esp (gcc regno = 7) | |
263 | 5 for %ebp (gcc regno = 6) | |
264 | 6 for %esi (gcc regno = 4) | |
265 | 7 for %edi (gcc regno = 5) | |
266 | The following three DWARF register numbers are never generated by | |
267 | the SVR4 C compiler or by the GNU compilers, but SDB on x86/svr4 | |
268 | believes these numbers have these meanings. | |
269 | 8 for %eip (no gcc equivalent) | |
270 | 9 for %eflags (gcc regno = 17) | |
271 | 10 for %trapno (no gcc equivalent) | |
272 | It is not at all clear how we should number the FP stack registers | |
273 | for the x86 architecture. If the version of SDB on x86/svr4 were | |
274 | a bit less brain dead with respect to floating-point then we would | |
275 | have a precedent to follow with respect to DWARF register numbers | |
276 | for x86 FP registers, but the SDB on x86/svr4 is so completely | |
277 | broken with respect to FP registers that it is hardly worth thinking | |
278 | of it as something to strive for compatibility with. | |
279 | The version of x86/svr4 SDB I have at the moment does (partially) | |
280 | seem to believe that DWARF register number 11 is associated with | |
281 | the x86 register %st(0), but that's about all. Higher DWARF | |
282 | register numbers don't seem to be associated with anything in | |
283 | particular, and even for DWARF regno 11, SDB only seems to under- | |
284 | stand that it should say that a variable lives in %st(0) (when | |
285 | asked via an `=' command) if we said it was in DWARF regno 11, | |
286 | but SDB still prints garbage when asked for the value of the | |
287 | variable in question (via a `/' command). | |
288 | (Also note that the labels SDB prints for various FP stack regs | |
289 | when doing an `x' command are all wrong.) | |
290 | Note that these problems generally don't affect the native SVR4 | |
291 | C compiler because it doesn't allow the use of -O with -g and | |
292 | because when it is *not* optimizing, it allocates a memory | |
293 | location for each floating-point variable, and the memory | |
294 | location is what gets described in the DWARF AT_location | |
295 | attribute for the variable in question. | |
296 | Regardless of the severe mental illness of the x86/svr4 SDB, we | |
297 | do something sensible here and we use the following DWARF | |
298 | register numbers. Note that these are all stack-top-relative | |
299 | numbers. | |
300 | 11 for %st(0) (gcc regno = 8) | |
301 | 12 for %st(1) (gcc regno = 9) | |
302 | 13 for %st(2) (gcc regno = 10) | |
303 | 14 for %st(3) (gcc regno = 11) | |
304 | 15 for %st(4) (gcc regno = 12) | |
305 | 16 for %st(5) (gcc regno = 13) | |
306 | 17 for %st(6) (gcc regno = 14) | |
307 | 18 for %st(7) (gcc regno = 15) | |
308 | */ | |
309 | int const svr4_dbx_register_map[FIRST_PSEUDO_REGISTER] = | |
310 | { | |
311 | 0, 2, 1, 3, 6, 7, 5, 4, /* general regs */ | |
312 | 11, 12, 13, 14, 15, 16, 17, 18, /* fp regs */ | |
313 | -1, 9, -1, -1, /* arg, flags, fpsr, dir */ | |
314 | }; | |
315 | ||
316 | ||
317 | ||
318 | /* Test and compare insns in i386.md store the information needed to | |
319 | generate branch and scc insns here. */ | |
320 | ||
321 | struct rtx_def *ix86_compare_op0 = NULL_RTX; | |
322 | struct rtx_def *ix86_compare_op1 = NULL_RTX; | |
323 | ||
324 | #define MAX_386_STACK_LOCALS 2 | |
325 | ||
326 | /* Define the structure for the machine field in struct function. */ | |
327 | struct machine_function | |
328 | { | |
329 | rtx stack_locals[(int) MAX_MACHINE_MODE][MAX_386_STACK_LOCALS]; | |
330 | }; | |
331 | ||
332 | #define ix86_stack_locals (cfun->machine->stack_locals) | |
333 | ||
334 | /* which cpu are we scheduling for */ | |
335 | enum processor_type ix86_cpu; | |
336 | ||
337 | /* which instruction set architecture to use. */ | |
338 | int ix86_arch; | |
339 | ||
340 | /* Strings to hold which cpu and instruction set architecture to use. */ | |
341 | const char *ix86_cpu_string; /* for -mcpu=<xxx> */ | |
342 | const char *ix86_arch_string; /* for -march=<xxx> */ | |
343 | ||
344 | /* Register allocation order */ | |
345 | const char *ix86_reg_alloc_order; | |
346 | static char regs_allocated[FIRST_PSEUDO_REGISTER]; | |
347 | ||
348 | /* # of registers to use to pass arguments. */ | |
349 | const char *ix86_regparm_string; | |
350 | ||
351 | /* ix86_regparm_string as a number */ | |
352 | int ix86_regparm; | |
353 | ||
354 | /* Alignment to use for loops and jumps: */ | |
355 | ||
356 | /* Power of two alignment for loops. */ | |
357 | const char *ix86_align_loops_string; | |
358 | ||
359 | /* Power of two alignment for non-loop jumps. */ | |
360 | const char *ix86_align_jumps_string; | |
361 | ||
362 | /* Power of two alignment for stack boundary in bytes. */ | |
363 | const char *ix86_preferred_stack_boundary_string; | |
364 | ||
365 | /* Preferred alignment for stack boundary in bits. */ | |
366 | int ix86_preferred_stack_boundary; | |
367 | ||
368 | /* Values 1-5: see jump.c */ | |
369 | int ix86_branch_cost; | |
370 | const char *ix86_branch_cost_string; | |
371 | ||
372 | /* Power of two alignment for functions. */ | |
373 | int ix86_align_funcs; | |
374 | const char *ix86_align_funcs_string; | |
375 | ||
376 | /* Power of two alignment for loops. */ | |
377 | int ix86_align_loops; | |
378 | ||
379 | /* Power of two alignment for non-loop jumps. */ | |
380 | int ix86_align_jumps; | |
381 | \f | |
382 | static void output_pic_addr_const PARAMS ((FILE *, rtx, int)); | |
383 | static void put_condition_code PARAMS ((enum rtx_code, enum machine_mode, | |
384 | int, int, FILE *)); | |
385 | static enum rtx_code unsigned_comparison PARAMS ((enum rtx_code code)); | |
386 | static rtx ix86_expand_int_compare PARAMS ((enum rtx_code, rtx, rtx)); | |
387 | static rtx ix86_expand_fp_compare PARAMS ((enum rtx_code, rtx, rtx, int)); | |
388 | static rtx ix86_expand_compare PARAMS ((enum rtx_code, int)); | |
389 | static rtx gen_push PARAMS ((rtx)); | |
390 | static int memory_address_length PARAMS ((rtx addr)); | |
391 | static int ix86_flags_dependant PARAMS ((rtx, rtx, enum attr_type)); | |
392 | static int ix86_agi_dependant PARAMS ((rtx, rtx, enum attr_type)); | |
393 | static int ix86_safe_length PARAMS ((rtx)); | |
394 | static enum attr_memory ix86_safe_memory PARAMS ((rtx)); | |
395 | static enum attr_pent_pair ix86_safe_pent_pair PARAMS ((rtx)); | |
396 | static enum attr_ppro_uops ix86_safe_ppro_uops PARAMS ((rtx)); | |
397 | static void ix86_dump_ppro_packet PARAMS ((FILE *)); | |
398 | static void ix86_reorder_insn PARAMS ((rtx *, rtx *)); | |
399 | static rtx * ix86_pent_find_pair PARAMS ((rtx *, rtx *, enum attr_pent_pair, | |
400 | rtx)); | |
401 | static void ix86_init_machine_status PARAMS ((struct function *)); | |
402 | static void ix86_mark_machine_status PARAMS ((struct function *)); | |
403 | static void ix86_split_to_parts PARAMS ((rtx, rtx *, enum machine_mode)); | |
404 | static int ix86_safe_length_prefix PARAMS ((rtx)); | |
405 | static HOST_WIDE_INT ix86_compute_frame_size PARAMS((HOST_WIDE_INT, | |
406 | int *, int *, int *)); | |
407 | static int ix86_nsaved_regs PARAMS((void)); | |
408 | static void ix86_emit_save_regs PARAMS((void)); | |
409 | static void ix86_emit_restore_regs_using_mov PARAMS ((rtx, int)); | |
410 | static void ix86_emit_epilogue_esp_adjustment PARAMS((int)); | |
411 | static void ix86_sched_reorder_pentium PARAMS((rtx *, rtx *)); | |
412 | static void ix86_sched_reorder_ppro PARAMS((rtx *, rtx *)); | |
413 | ||
414 | struct ix86_address | |
415 | { | |
416 | rtx base, index, disp; | |
417 | HOST_WIDE_INT scale; | |
418 | }; | |
419 | ||
420 | static int ix86_decompose_address PARAMS ((rtx, struct ix86_address *)); | |
421 | \f | |
422 | /* Sometimes certain combinations of command options do not make | |
423 | sense on a particular target machine. You can define a macro | |
424 | `OVERRIDE_OPTIONS' to take account of this. This macro, if | |
425 | defined, is executed once just after all the command options have | |
426 | been parsed. | |
427 | ||
428 | Don't use this macro to turn on various extra optimizations for | |
429 | `-O'. That is what `OPTIMIZATION_OPTIONS' is for. */ | |
430 | ||
431 | void | |
432 | override_options () | |
433 | { | |
434 | /* Comes from final.c -- no real reason to change it. */ | |
435 | #define MAX_CODE_ALIGN 16 | |
436 | ||
437 | static struct ptt | |
438 | { | |
439 | struct processor_costs *cost; /* Processor costs */ | |
440 | int target_enable; /* Target flags to enable. */ | |
441 | int target_disable; /* Target flags to disable. */ | |
442 | int align_loop; /* Default alignments. */ | |
443 | int align_jump; | |
444 | int align_func; | |
445 | int branch_cost; | |
446 | } | |
447 | const processor_target_table[PROCESSOR_max] = | |
448 | { | |
449 | {&i386_cost, 0, 0, 2, 2, 2, 1}, | |
450 | {&i486_cost, 0, 0, 4, 4, 4, 1}, | |
451 | {&pentium_cost, 0, 0, -4, -4, -4, 1}, | |
452 | {&pentiumpro_cost, 0, 0, 4, -4, 4, 1}, | |
453 | {&k6_cost, 0, 0, -5, -5, 4, 1}, | |
454 | {&athlon_cost, 0, 0, 4, -4, 4, 1} | |
455 | }; | |
456 | ||
457 | static struct pta | |
458 | { | |
459 | const char *name; /* processor name or nickname. */ | |
460 | enum processor_type processor; | |
461 | } | |
462 | const processor_alias_table[] = | |
463 | { | |
464 | {"i386", PROCESSOR_I386}, | |
465 | {"i486", PROCESSOR_I486}, | |
466 | {"i586", PROCESSOR_PENTIUM}, | |
467 | {"pentium", PROCESSOR_PENTIUM}, | |
468 | {"i686", PROCESSOR_PENTIUMPRO}, | |
469 | {"pentiumpro", PROCESSOR_PENTIUMPRO}, | |
470 | {"k6", PROCESSOR_K6}, | |
471 | {"athlon", PROCESSOR_ATHLON}, | |
472 | }; | |
473 | ||
474 | int const pta_size = sizeof(processor_alias_table)/sizeof(struct pta); | |
475 | ||
476 | #ifdef SUBTARGET_OVERRIDE_OPTIONS | |
477 | SUBTARGET_OVERRIDE_OPTIONS; | |
478 | #endif | |
479 | ||
480 | ix86_arch = PROCESSOR_I386; | |
481 | ix86_cpu = (enum processor_type) TARGET_CPU_DEFAULT; | |
482 | ||
483 | if (ix86_arch_string != 0) | |
484 | { | |
485 | int i; | |
486 | for (i = 0; i < pta_size; i++) | |
487 | if (! strcmp (ix86_arch_string, processor_alias_table[i].name)) | |
488 | { | |
489 | ix86_arch = processor_alias_table[i].processor; | |
490 | /* Default cpu tuning to the architecture. */ | |
491 | ix86_cpu = ix86_arch; | |
492 | break; | |
493 | } | |
494 | if (i == pta_size) | |
495 | error ("bad value (%s) for -march= switch", ix86_arch_string); | |
496 | } | |
497 | ||
498 | if (ix86_cpu_string != 0) | |
499 | { | |
500 | int i; | |
501 | for (i = 0; i < pta_size; i++) | |
502 | if (! strcmp (ix86_cpu_string, processor_alias_table[i].name)) | |
503 | { | |
504 | ix86_cpu = processor_alias_table[i].processor; | |
505 | break; | |
506 | } | |
507 | if (i == pta_size) | |
508 | error ("bad value (%s) for -mcpu= switch", ix86_cpu_string); | |
509 | } | |
510 | ||
511 | ix86_cost = processor_target_table[ix86_cpu].cost; | |
512 | target_flags |= processor_target_table[ix86_cpu].target_enable; | |
513 | target_flags &= ~processor_target_table[ix86_cpu].target_disable; | |
514 | ||
515 | /* Arrange to set up i386_stack_locals for all functions. */ | |
516 | init_machine_status = ix86_init_machine_status; | |
517 | mark_machine_status = ix86_mark_machine_status; | |
518 | ||
519 | /* Validate registers in register allocation order. */ | |
520 | if (ix86_reg_alloc_order) | |
521 | { | |
522 | int i, ch; | |
523 | for (i = 0; (ch = ix86_reg_alloc_order[i]) != '\0'; i++) | |
524 | { | |
525 | int regno = 0; | |
526 | ||
527 | switch (ch) | |
528 | { | |
529 | case 'a': regno = 0; break; | |
530 | case 'd': regno = 1; break; | |
531 | case 'c': regno = 2; break; | |
532 | case 'b': regno = 3; break; | |
533 | case 'S': regno = 4; break; | |
534 | case 'D': regno = 5; break; | |
535 | case 'B': regno = 6; break; | |
536 | ||
537 | default: fatal ("Register '%c' is unknown", ch); | |
538 | } | |
539 | ||
540 | if (regs_allocated[regno]) | |
541 | fatal ("Register '%c' already specified in allocation order", ch); | |
542 | ||
543 | regs_allocated[regno] = 1; | |
544 | } | |
545 | } | |
546 | ||
547 | /* Validate -mregparm= value. */ | |
548 | if (ix86_regparm_string) | |
549 | { | |
550 | ix86_regparm = atoi (ix86_regparm_string); | |
551 | if (ix86_regparm < 0 || ix86_regparm > REGPARM_MAX) | |
552 | fatal ("-mregparm=%d is not between 0 and %d", | |
553 | ix86_regparm, REGPARM_MAX); | |
554 | } | |
555 | ||
556 | /* Validate -malign-loops= value, or provide default. */ | |
557 | ix86_align_loops = processor_target_table[ix86_cpu].align_loop; | |
558 | if (ix86_align_loops_string) | |
559 | { | |
560 | ix86_align_loops = atoi (ix86_align_loops_string); | |
561 | if (ix86_align_loops < 0 || ix86_align_loops > MAX_CODE_ALIGN) | |
562 | fatal ("-malign-loops=%d is not between 0 and %d", | |
563 | ix86_align_loops, MAX_CODE_ALIGN); | |
564 | } | |
565 | ||
566 | /* Validate -malign-jumps= value, or provide default. */ | |
567 | ix86_align_jumps = processor_target_table[ix86_cpu].align_jump; | |
568 | if (ix86_align_jumps_string) | |
569 | { | |
570 | ix86_align_jumps = atoi (ix86_align_jumps_string); | |
571 | if (ix86_align_jumps < 0 || ix86_align_jumps > MAX_CODE_ALIGN) | |
572 | fatal ("-malign-jumps=%d is not between 0 and %d", | |
573 | ix86_align_jumps, MAX_CODE_ALIGN); | |
574 | } | |
575 | ||
576 | /* Validate -malign-functions= value, or provide default. */ | |
577 | ix86_align_funcs = processor_target_table[ix86_cpu].align_func; | |
578 | if (ix86_align_funcs_string) | |
579 | { | |
580 | ix86_align_funcs = atoi (ix86_align_funcs_string); | |
581 | if (ix86_align_funcs < 0 || ix86_align_funcs > MAX_CODE_ALIGN) | |
582 | fatal ("-malign-functions=%d is not between 0 and %d", | |
583 | ix86_align_funcs, MAX_CODE_ALIGN); | |
584 | } | |
585 | ||
586 | /* Validate -mpreferred-stack-boundary= value, or provide default. | |
587 | The default of 128 bits is for Pentium III's SSE __m128. */ | |
588 | ix86_preferred_stack_boundary = 128; | |
589 | if (ix86_preferred_stack_boundary_string) | |
590 | { | |
591 | int i = atoi (ix86_preferred_stack_boundary_string); | |
592 | if (i < 2 || i > 31) | |
593 | fatal ("-mpreferred-stack-boundary=%d is not between 2 and 31", i); | |
594 | ix86_preferred_stack_boundary = (1 << i) * BITS_PER_UNIT; | |
595 | } | |
596 | ||
597 | /* Validate -mbranch-cost= value, or provide default. */ | |
598 | ix86_branch_cost = processor_target_table[ix86_cpu].branch_cost; | |
599 | if (ix86_branch_cost_string) | |
600 | { | |
601 | ix86_branch_cost = atoi (ix86_branch_cost_string); | |
602 | if (ix86_branch_cost < 0 || ix86_branch_cost > 5) | |
603 | fatal ("-mbranch-cost=%d is not between 0 and 5", | |
604 | ix86_branch_cost); | |
605 | } | |
606 | ||
607 | /* Keep nonleaf frame pointers. */ | |
608 | if (TARGET_OMIT_LEAF_FRAME_POINTER) | |
609 | flag_omit_frame_pointer = 1; | |
610 | ||
611 | /* If we're doing fast math, we don't care about comparison order | |
612 | wrt NaNs. This lets us use a shorter comparison sequence. */ | |
613 | if (flag_fast_math) | |
614 | target_flags &= ~MASK_IEEE_FP; | |
615 | ||
616 | /* If we're planning on using `loop', use it. */ | |
617 | if (TARGET_USE_LOOP && optimize) | |
618 | flag_branch_on_count_reg = 1; | |
619 | } | |
620 | \f | |
621 | /* A C statement (sans semicolon) to choose the order in which to | |
622 | allocate hard registers for pseudo-registers local to a basic | |
623 | block. | |
624 | ||
625 | Store the desired register order in the array `reg_alloc_order'. | |
626 | Element 0 should be the register to allocate first; element 1, the | |
627 | next register; and so on. | |
628 | ||
629 | The macro body should not assume anything about the contents of | |
630 | `reg_alloc_order' before execution of the macro. | |
631 | ||
632 | On most machines, it is not necessary to define this macro. */ | |
633 | ||
634 | void | |
635 | order_regs_for_local_alloc () | |
636 | { | |
637 | int i, ch, order; | |
638 | ||
639 | /* User specified the register allocation order. */ | |
640 | ||
641 | if (ix86_reg_alloc_order) | |
642 | { | |
643 | for (i = order = 0; (ch = ix86_reg_alloc_order[i]) != '\0'; i++) | |
644 | { | |
645 | int regno = 0; | |
646 | ||
647 | switch (ch) | |
648 | { | |
649 | case 'a': regno = 0; break; | |
650 | case 'd': regno = 1; break; | |
651 | case 'c': regno = 2; break; | |
652 | case 'b': regno = 3; break; | |
653 | case 'S': regno = 4; break; | |
654 | case 'D': regno = 5; break; | |
655 | case 'B': regno = 6; break; | |
656 | } | |
657 | ||
658 | reg_alloc_order[order++] = regno; | |
659 | } | |
660 | ||
661 | for (i = 0; i < FIRST_PSEUDO_REGISTER; i++) | |
662 | { | |
663 | if (! regs_allocated[i]) | |
664 | reg_alloc_order[order++] = i; | |
665 | } | |
666 | } | |
667 | ||
668 | /* If user did not specify a register allocation order, use natural order. */ | |
669 | else | |
670 | { | |
671 | for (i = 0; i < FIRST_PSEUDO_REGISTER; i++) | |
672 | reg_alloc_order[i] = i; | |
673 | } | |
674 | } | |
675 | \f | |
676 | void | |
677 | optimization_options (level, size) | |
678 | int level; | |
679 | int size ATTRIBUTE_UNUSED; | |
680 | { | |
681 | /* For -O2 and beyond, turn off -fschedule-insns by default. It tends to | |
682 | make the problem with not enough registers even worse. */ | |
683 | #ifdef INSN_SCHEDULING | |
684 | if (level > 1) | |
685 | flag_schedule_insns = 0; | |
686 | #endif | |
687 | } | |
688 | \f | |
689 | /* Return nonzero if IDENTIFIER with arguments ARGS is a valid machine specific | |
690 | attribute for DECL. The attributes in ATTRIBUTES have previously been | |
691 | assigned to DECL. */ | |
692 | ||
693 | int | |
694 | ix86_valid_decl_attribute_p (decl, attributes, identifier, args) | |
695 | tree decl ATTRIBUTE_UNUSED; | |
696 | tree attributes ATTRIBUTE_UNUSED; | |
697 | tree identifier ATTRIBUTE_UNUSED; | |
698 | tree args ATTRIBUTE_UNUSED; | |
699 | { | |
700 | return 0; | |
701 | } | |
702 | ||
703 | /* Return nonzero if IDENTIFIER with arguments ARGS is a valid machine specific | |
704 | attribute for TYPE. The attributes in ATTRIBUTES have previously been | |
705 | assigned to TYPE. */ | |
706 | ||
707 | int | |
708 | ix86_valid_type_attribute_p (type, attributes, identifier, args) | |
709 | tree type; | |
710 | tree attributes ATTRIBUTE_UNUSED; | |
711 | tree identifier; | |
712 | tree args; | |
713 | { | |
714 | if (TREE_CODE (type) != FUNCTION_TYPE | |
715 | && TREE_CODE (type) != METHOD_TYPE | |
716 | && TREE_CODE (type) != FIELD_DECL | |
717 | && TREE_CODE (type) != TYPE_DECL) | |
718 | return 0; | |
719 | ||
720 | /* Stdcall attribute says callee is responsible for popping arguments | |
721 | if they are not variable. */ | |
722 | if (is_attribute_p ("stdcall", identifier)) | |
723 | return (args == NULL_TREE); | |
724 | ||
725 | /* Cdecl attribute says the callee is a normal C declaration. */ | |
726 | if (is_attribute_p ("cdecl", identifier)) | |
727 | return (args == NULL_TREE); | |
728 | ||
729 | /* Regparm attribute specifies how many integer arguments are to be | |
730 | passed in registers. */ | |
731 | if (is_attribute_p ("regparm", identifier)) | |
732 | { | |
733 | tree cst; | |
734 | ||
735 | if (! args || TREE_CODE (args) != TREE_LIST | |
736 | || TREE_CHAIN (args) != NULL_TREE | |
737 | || TREE_VALUE (args) == NULL_TREE) | |
738 | return 0; | |
739 | ||
740 | cst = TREE_VALUE (args); | |
741 | if (TREE_CODE (cst) != INTEGER_CST) | |
742 | return 0; | |
743 | ||
744 | if (compare_tree_int (cst, REGPARM_MAX) > 0) | |
745 | return 0; | |
746 | ||
747 | return 1; | |
748 | } | |
749 | ||
750 | return 0; | |
751 | } | |
752 | ||
753 | /* Return 0 if the attributes for two types are incompatible, 1 if they | |
754 | are compatible, and 2 if they are nearly compatible (which causes a | |
755 | warning to be generated). */ | |
756 | ||
757 | int | |
758 | ix86_comp_type_attributes (type1, type2) | |
759 | tree type1; | |
760 | tree type2; | |
761 | { | |
762 | /* Check for mismatch of non-default calling convention. */ | |
763 | const char *rtdstr = TARGET_RTD ? "cdecl" : "stdcall"; | |
764 | ||
765 | if (TREE_CODE (type1) != FUNCTION_TYPE) | |
766 | return 1; | |
767 | ||
768 | /* Check for mismatched return types (cdecl vs stdcall). */ | |
769 | if (!lookup_attribute (rtdstr, TYPE_ATTRIBUTES (type1)) | |
770 | != !lookup_attribute (rtdstr, TYPE_ATTRIBUTES (type2))) | |
771 | return 0; | |
772 | return 1; | |
773 | } | |
774 | \f | |
775 | /* Value is the number of bytes of arguments automatically | |
776 | popped when returning from a subroutine call. | |
777 | FUNDECL is the declaration node of the function (as a tree), | |
778 | FUNTYPE is the data type of the function (as a tree), | |
779 | or for a library call it is an identifier node for the subroutine name. | |
780 | SIZE is the number of bytes of arguments passed on the stack. | |
781 | ||
782 | On the 80386, the RTD insn may be used to pop them if the number | |
783 | of args is fixed, but if the number is variable then the caller | |
784 | must pop them all. RTD can't be used for library calls now | |
785 | because the library is compiled with the Unix compiler. | |
786 | Use of RTD is a selectable option, since it is incompatible with | |
787 | standard Unix calling sequences. If the option is not selected, | |
788 | the caller must always pop the args. | |
789 | ||
790 | The attribute stdcall is equivalent to RTD on a per module basis. */ | |
791 | ||
792 | int | |
793 | ix86_return_pops_args (fundecl, funtype, size) | |
794 | tree fundecl; | |
795 | tree funtype; | |
796 | int size; | |
797 | { | |
798 | int rtd = TARGET_RTD && (!fundecl || TREE_CODE (fundecl) != IDENTIFIER_NODE); | |
799 | ||
800 | /* Cdecl functions override -mrtd, and never pop the stack. */ | |
801 | if (! lookup_attribute ("cdecl", TYPE_ATTRIBUTES (funtype))) { | |
802 | ||
803 | /* Stdcall functions will pop the stack if not variable args. */ | |
804 | if (lookup_attribute ("stdcall", TYPE_ATTRIBUTES (funtype))) | |
805 | rtd = 1; | |
806 | ||
807 | if (rtd | |
808 | && (TYPE_ARG_TYPES (funtype) == NULL_TREE | |
809 | || (TREE_VALUE (tree_last (TYPE_ARG_TYPES (funtype))) | |
810 | == void_type_node))) | |
811 | return size; | |
812 | } | |
813 | ||
814 | /* Lose any fake structure return argument. */ | |
815 | if (aggregate_value_p (TREE_TYPE (funtype))) | |
816 | return GET_MODE_SIZE (Pmode); | |
817 | ||
818 | return 0; | |
819 | } | |
820 | \f | |
821 | /* Argument support functions. */ | |
822 | ||
823 | /* Initialize a variable CUM of type CUMULATIVE_ARGS | |
824 | for a call to a function whose data type is FNTYPE. | |
825 | For a library call, FNTYPE is 0. */ | |
826 | ||
827 | void | |
828 | init_cumulative_args (cum, fntype, libname) | |
829 | CUMULATIVE_ARGS *cum; /* Argument info to initialize */ | |
830 | tree fntype; /* tree ptr for function decl */ | |
831 | rtx libname; /* SYMBOL_REF of library name or 0 */ | |
832 | { | |
833 | static CUMULATIVE_ARGS zero_cum; | |
834 | tree param, next_param; | |
835 | ||
836 | if (TARGET_DEBUG_ARG) | |
837 | { | |
838 | fprintf (stderr, "\ninit_cumulative_args ("); | |
839 | if (fntype) | |
840 | fprintf (stderr, "fntype code = %s, ret code = %s", | |
841 | tree_code_name[(int) TREE_CODE (fntype)], | |
842 | tree_code_name[(int) TREE_CODE (TREE_TYPE (fntype))]); | |
843 | else | |
844 | fprintf (stderr, "no fntype"); | |
845 | ||
846 | if (libname) | |
847 | fprintf (stderr, ", libname = %s", XSTR (libname, 0)); | |
848 | } | |
849 | ||
850 | *cum = zero_cum; | |
851 | ||
852 | /* Set up the number of registers to use for passing arguments. */ | |
853 | cum->nregs = ix86_regparm; | |
854 | if (fntype) | |
855 | { | |
856 | tree attr = lookup_attribute ("regparm", TYPE_ATTRIBUTES (fntype)); | |
857 | ||
858 | if (attr) | |
859 | cum->nregs = TREE_INT_CST_LOW (TREE_VALUE (TREE_VALUE (attr))); | |
860 | } | |
861 | ||
862 | /* Determine if this function has variable arguments. This is | |
863 | indicated by the last argument being 'void_type_mode' if there | |
864 | are no variable arguments. If there are variable arguments, then | |
865 | we won't pass anything in registers */ | |
866 | ||
867 | if (cum->nregs) | |
868 | { | |
869 | for (param = (fntype) ? TYPE_ARG_TYPES (fntype) : 0; | |
870 | param != 0; param = next_param) | |
871 | { | |
872 | next_param = TREE_CHAIN (param); | |
873 | if (next_param == 0 && TREE_VALUE (param) != void_type_node) | |
874 | cum->nregs = 0; | |
875 | } | |
876 | } | |
877 | ||
878 | if (TARGET_DEBUG_ARG) | |
879 | fprintf (stderr, ", nregs=%d )\n", cum->nregs); | |
880 | ||
881 | return; | |
882 | } | |
883 | ||
884 | /* Update the data in CUM to advance over an argument | |
885 | of mode MODE and data type TYPE. | |
886 | (TYPE is null for libcalls where that information may not be available.) */ | |
887 | ||
888 | void | |
889 | function_arg_advance (cum, mode, type, named) | |
890 | CUMULATIVE_ARGS *cum; /* current arg information */ | |
891 | enum machine_mode mode; /* current arg mode */ | |
892 | tree type; /* type of the argument or 0 if lib support */ | |
893 | int named; /* whether or not the argument was named */ | |
894 | { | |
895 | int bytes | |
896 | = (mode == BLKmode) ? int_size_in_bytes (type) : GET_MODE_SIZE (mode); | |
897 | int words = (bytes + UNITS_PER_WORD - 1) / UNITS_PER_WORD; | |
898 | ||
899 | if (TARGET_DEBUG_ARG) | |
900 | fprintf (stderr, | |
901 | "function_adv (sz=%d, wds=%2d, nregs=%d, mode=%s, named=%d)\n\n", | |
902 | words, cum->words, cum->nregs, GET_MODE_NAME (mode), named); | |
903 | ||
904 | cum->words += words; | |
905 | cum->nregs -= words; | |
906 | cum->regno += words; | |
907 | ||
908 | if (cum->nregs <= 0) | |
909 | { | |
910 | cum->nregs = 0; | |
911 | cum->regno = 0; | |
912 | } | |
913 | ||
914 | return; | |
915 | } | |
916 | ||
917 | /* Define where to put the arguments to a function. | |
918 | Value is zero to push the argument on the stack, | |
919 | or a hard register in which to store the argument. | |
920 | ||
921 | MODE is the argument's machine mode. | |
922 | TYPE is the data type of the argument (as a tree). | |
923 | This is null for libcalls where that information may | |
924 | not be available. | |
925 | CUM is a variable of type CUMULATIVE_ARGS which gives info about | |
926 | the preceding args and about the function being called. | |
927 | NAMED is nonzero if this argument is a named parameter | |
928 | (otherwise it is an extra parameter matching an ellipsis). */ | |
929 | ||
930 | struct rtx_def * | |
931 | function_arg (cum, mode, type, named) | |
932 | CUMULATIVE_ARGS *cum; /* current arg information */ | |
933 | enum machine_mode mode; /* current arg mode */ | |
934 | tree type; /* type of the argument or 0 if lib support */ | |
935 | int named; /* != 0 for normal args, == 0 for ... args */ | |
936 | { | |
937 | rtx ret = NULL_RTX; | |
938 | int bytes | |
939 | = (mode == BLKmode) ? int_size_in_bytes (type) : GET_MODE_SIZE (mode); | |
940 | int words = (bytes + UNITS_PER_WORD - 1) / UNITS_PER_WORD; | |
941 | ||
942 | switch (mode) | |
943 | { | |
944 | /* For now, pass fp/complex values on the stack. */ | |
945 | default: | |
946 | break; | |
947 | ||
948 | case BLKmode: | |
949 | case DImode: | |
950 | case SImode: | |
951 | case HImode: | |
952 | case QImode: | |
953 | if (words <= cum->nregs) | |
954 | ret = gen_rtx_REG (mode, cum->regno); | |
955 | break; | |
956 | } | |
957 | ||
958 | if (TARGET_DEBUG_ARG) | |
959 | { | |
960 | fprintf (stderr, | |
961 | "function_arg (size=%d, wds=%2d, nregs=%d, mode=%4s, named=%d", | |
962 | words, cum->words, cum->nregs, GET_MODE_NAME (mode), named); | |
963 | ||
964 | if (ret) | |
965 | fprintf (stderr, ", reg=%%e%s", reg_names[ REGNO(ret) ]); | |
966 | else | |
967 | fprintf (stderr, ", stack"); | |
968 | ||
969 | fprintf (stderr, " )\n"); | |
970 | } | |
971 | ||
972 | return ret; | |
973 | } | |
974 | \f | |
975 | /* Returns 1 if OP is either a symbol reference or a sum of a symbol | |
976 | reference and a constant. */ | |
977 | ||
978 | int | |
979 | symbolic_operand (op, mode) | |
980 | register rtx op; | |
981 | enum machine_mode mode ATTRIBUTE_UNUSED; | |
982 | { | |
983 | switch (GET_CODE (op)) | |
984 | { | |
985 | case SYMBOL_REF: | |
986 | case LABEL_REF: | |
987 | return 1; | |
988 | ||
989 | case CONST: | |
990 | op = XEXP (op, 0); | |
991 | if (GET_CODE (op) == SYMBOL_REF | |
992 | || GET_CODE (op) == LABEL_REF | |
993 | || (GET_CODE (op) == UNSPEC | |
994 | && XINT (op, 1) >= 6 | |
995 | && XINT (op, 1) <= 7)) | |
996 | return 1; | |
997 | if (GET_CODE (op) != PLUS | |
998 | || GET_CODE (XEXP (op, 1)) != CONST_INT) | |
999 | return 0; | |
1000 | ||
1001 | op = XEXP (op, 0); | |
1002 | if (GET_CODE (op) == SYMBOL_REF | |
1003 | || GET_CODE (op) == LABEL_REF) | |
1004 | return 1; | |
1005 | /* Only @GOTOFF gets offsets. */ | |
1006 | if (GET_CODE (op) != UNSPEC | |
1007 | || XINT (op, 1) != 7) | |
1008 | return 0; | |
1009 | ||
1010 | op = XVECEXP (op, 0, 0); | |
1011 | if (GET_CODE (op) == SYMBOL_REF | |
1012 | || GET_CODE (op) == LABEL_REF) | |
1013 | return 1; | |
1014 | return 0; | |
1015 | ||
1016 | default: | |
1017 | return 0; | |
1018 | } | |
1019 | } | |
1020 | ||
1021 | /* Return true if the operand contains a @GOT or @GOTOFF reference. */ | |
1022 | ||
1023 | int | |
1024 | pic_symbolic_operand (op, mode) | |
1025 | register rtx op; | |
1026 | enum machine_mode mode ATTRIBUTE_UNUSED; | |
1027 | { | |
1028 | if (GET_CODE (op) == CONST) | |
1029 | { | |
1030 | op = XEXP (op, 0); | |
1031 | if (GET_CODE (op) == UNSPEC) | |
1032 | return 1; | |
1033 | if (GET_CODE (op) != PLUS | |
1034 | || GET_CODE (XEXP (op, 1)) != CONST_INT) | |
1035 | return 0; | |
1036 | op = XEXP (op, 0); | |
1037 | if (GET_CODE (op) == UNSPEC) | |
1038 | return 1; | |
1039 | } | |
1040 | return 0; | |
1041 | } | |
1042 | ||
1043 | /* Test for a valid operand for a call instruction. Don't allow the | |
1044 | arg pointer register or virtual regs since they may decay into | |
1045 | reg + const, which the patterns can't handle. */ | |
1046 | ||
1047 | int | |
1048 | call_insn_operand (op, mode) | |
1049 | rtx op; | |
1050 | enum machine_mode mode ATTRIBUTE_UNUSED; | |
1051 | { | |
1052 | if (GET_CODE (op) != MEM) | |
1053 | return 0; | |
1054 | op = XEXP (op, 0); | |
1055 | ||
1056 | /* Disallow indirect through a virtual register. This leads to | |
1057 | compiler aborts when trying to eliminate them. */ | |
1058 | if (GET_CODE (op) == REG | |
1059 | && (op == arg_pointer_rtx | |
1060 | || op == frame_pointer_rtx | |
1061 | || (REGNO (op) >= FIRST_PSEUDO_REGISTER | |
1062 | && REGNO (op) <= LAST_VIRTUAL_REGISTER))) | |
1063 | return 0; | |
1064 | ||
1065 | /* Disallow `call 1234'. Due to varying assembler lameness this | |
1066 | gets either rejected or translated to `call .+1234'. */ | |
1067 | if (GET_CODE (op) == CONST_INT) | |
1068 | return 0; | |
1069 | ||
1070 | /* Otherwise we can allow any general_operand in the address. */ | |
1071 | return general_operand (op, Pmode); | |
1072 | } | |
1073 | ||
1074 | /* Like call_insn_operand but allow (mem (symbol_ref ...)) even if pic. */ | |
1075 | ||
1076 | int | |
1077 | expander_call_insn_operand (op, mode) | |
1078 | rtx op; | |
1079 | enum machine_mode mode; | |
1080 | { | |
1081 | if (GET_CODE (op) == MEM | |
1082 | && GET_CODE (XEXP (op, 0)) == SYMBOL_REF) | |
1083 | return 1; | |
1084 | ||
1085 | return call_insn_operand (op, mode); | |
1086 | } | |
1087 | ||
1088 | int | |
1089 | constant_call_address_operand (op, mode) | |
1090 | rtx op; | |
1091 | enum machine_mode mode ATTRIBUTE_UNUSED; | |
1092 | { | |
1093 | return GET_CODE (op) == MEM && | |
1094 | CONSTANT_ADDRESS_P (XEXP (op, 0)) && | |
1095 | GET_CODE (XEXP (op, 0)) != CONST_INT; | |
1096 | } | |
1097 | ||
1098 | /* Match exactly zero and one. */ | |
1099 | ||
1100 | int | |
1101 | const0_operand (op, mode) | |
1102 | register rtx op; | |
1103 | enum machine_mode mode; | |
1104 | { | |
1105 | return op == CONST0_RTX (mode); | |
1106 | } | |
1107 | ||
1108 | int | |
1109 | const1_operand (op, mode) | |
1110 | register rtx op; | |
1111 | enum machine_mode mode ATTRIBUTE_UNUSED; | |
1112 | { | |
1113 | return op == const1_rtx; | |
1114 | } | |
1115 | ||
1116 | /* Match 2, 4, or 8. Used for leal multiplicands. */ | |
1117 | ||
1118 | int | |
1119 | const248_operand (op, mode) | |
1120 | register rtx op; | |
1121 | enum machine_mode mode ATTRIBUTE_UNUSED; | |
1122 | { | |
1123 | return (GET_CODE (op) == CONST_INT | |
1124 | && (INTVAL (op) == 2 || INTVAL (op) == 4 || INTVAL (op) == 8)); | |
1125 | } | |
1126 | ||
1127 | /* True if this is a constant appropriate for an increment or decremenmt. */ | |
1128 | ||
1129 | int | |
1130 | incdec_operand (op, mode) | |
1131 | register rtx op; | |
1132 | enum machine_mode mode; | |
1133 | { | |
1134 | if (op == const1_rtx || op == constm1_rtx) | |
1135 | return 1; | |
1136 | if (GET_CODE (op) != CONST_INT) | |
1137 | return 0; | |
1138 | if (mode == SImode && INTVAL (op) == (HOST_WIDE_INT) 0xffffffff) | |
1139 | return 1; | |
1140 | if (mode == HImode && INTVAL (op) == (HOST_WIDE_INT) 0xffff) | |
1141 | return 1; | |
1142 | if (mode == QImode && INTVAL (op) == (HOST_WIDE_INT) 0xff) | |
1143 | return 1; | |
1144 | return 0; | |
1145 | } | |
1146 | ||
1147 | /* Return false if this is the stack pointer, or any other fake | |
1148 | register eliminable to the stack pointer. Otherwise, this is | |
1149 | a register operand. | |
1150 | ||
1151 | This is used to prevent esp from being used as an index reg. | |
1152 | Which would only happen in pathological cases. */ | |
1153 | ||
1154 | int | |
1155 | reg_no_sp_operand (op, mode) | |
1156 | register rtx op; | |
1157 | enum machine_mode mode; | |
1158 | { | |
1159 | rtx t = op; | |
1160 | if (GET_CODE (t) == SUBREG) | |
1161 | t = SUBREG_REG (t); | |
1162 | if (t == stack_pointer_rtx || t == arg_pointer_rtx || t == frame_pointer_rtx) | |
1163 | return 0; | |
1164 | ||
1165 | return register_operand (op, mode); | |
1166 | } | |
1167 | ||
1168 | /* Return false if this is any eliminable register. Otherwise | |
1169 | general_operand. */ | |
1170 | ||
1171 | int | |
1172 | general_no_elim_operand (op, mode) | |
1173 | register rtx op; | |
1174 | enum machine_mode mode; | |
1175 | { | |
1176 | rtx t = op; | |
1177 | if (GET_CODE (t) == SUBREG) | |
1178 | t = SUBREG_REG (t); | |
1179 | if (t == arg_pointer_rtx || t == frame_pointer_rtx | |
1180 | || t == virtual_incoming_args_rtx || t == virtual_stack_vars_rtx | |
1181 | || t == virtual_stack_dynamic_rtx) | |
1182 | return 0; | |
1183 | ||
1184 | return general_operand (op, mode); | |
1185 | } | |
1186 | ||
1187 | /* Return false if this is any eliminable register. Otherwise | |
1188 | register_operand or const_int. */ | |
1189 | ||
1190 | int | |
1191 | nonmemory_no_elim_operand (op, mode) | |
1192 | register rtx op; | |
1193 | enum machine_mode mode; | |
1194 | { | |
1195 | rtx t = op; | |
1196 | if (GET_CODE (t) == SUBREG) | |
1197 | t = SUBREG_REG (t); | |
1198 | if (t == arg_pointer_rtx || t == frame_pointer_rtx | |
1199 | || t == virtual_incoming_args_rtx || t == virtual_stack_vars_rtx | |
1200 | || t == virtual_stack_dynamic_rtx) | |
1201 | return 0; | |
1202 | ||
1203 | return GET_CODE (op) == CONST_INT || register_operand (op, mode); | |
1204 | } | |
1205 | ||
1206 | /* Return true if op is a Q_REGS class register. */ | |
1207 | ||
1208 | int | |
1209 | q_regs_operand (op, mode) | |
1210 | register rtx op; | |
1211 | enum machine_mode mode; | |
1212 | { | |
1213 | if (mode != VOIDmode && GET_MODE (op) != mode) | |
1214 | return 0; | |
1215 | if (GET_CODE (op) == SUBREG) | |
1216 | op = SUBREG_REG (op); | |
1217 | return QI_REG_P (op); | |
1218 | } | |
1219 | ||
1220 | /* Return true if op is a NON_Q_REGS class register. */ | |
1221 | ||
1222 | int | |
1223 | non_q_regs_operand (op, mode) | |
1224 | register rtx op; | |
1225 | enum machine_mode mode; | |
1226 | { | |
1227 | if (mode != VOIDmode && GET_MODE (op) != mode) | |
1228 | return 0; | |
1229 | if (GET_CODE (op) == SUBREG) | |
1230 | op = SUBREG_REG (op); | |
1231 | return NON_QI_REG_P (op); | |
1232 | } | |
1233 | ||
1234 | /* Return 1 if OP is a comparison operator that can use the condition code | |
1235 | generated by a logical operation, which characteristicly does not set | |
1236 | overflow or carry. To be used with CCNOmode. */ | |
1237 | ||
1238 | int | |
1239 | no_comparison_operator (op, mode) | |
1240 | register rtx op; | |
1241 | enum machine_mode mode; | |
1242 | { | |
1243 | return ((mode == VOIDmode || GET_MODE (op) == mode) | |
1244 | && GET_RTX_CLASS (GET_CODE (op)) == '<' | |
1245 | && GET_CODE (op) != LE | |
1246 | && GET_CODE (op) != GT); | |
1247 | } | |
1248 | ||
1249 | /* Return 1 if OP is a comparison operator that can be issued by fcmov. */ | |
1250 | ||
1251 | int | |
1252 | fcmov_comparison_operator (op, mode) | |
1253 | register rtx op; | |
1254 | enum machine_mode mode; | |
1255 | { | |
1256 | return ((mode == VOIDmode || GET_MODE (op) == mode) | |
1257 | && GET_RTX_CLASS (GET_CODE (op)) == '<' | |
1258 | && GET_CODE (op) == unsigned_condition (GET_CODE (op))); | |
1259 | } | |
1260 | ||
1261 | /* Return 1 if OP is a binary operator that can be promoted to wider mode. */ | |
1262 | ||
1263 | int | |
1264 | promotable_binary_operator (op, mode) | |
1265 | register rtx op; | |
1266 | enum machine_mode mode ATTRIBUTE_UNUSED; | |
1267 | { | |
1268 | switch (GET_CODE (op)) | |
1269 | { | |
1270 | case MULT: | |
1271 | /* Modern CPUs have same latency for HImode and SImode multiply, | |
1272 | but 386 and 486 do HImode multiply faster. */ | |
1273 | return ix86_cpu > PROCESSOR_I486; | |
1274 | case PLUS: | |
1275 | case AND: | |
1276 | case IOR: | |
1277 | case XOR: | |
1278 | case ASHIFT: | |
1279 | return 1; | |
1280 | default: | |
1281 | return 0; | |
1282 | } | |
1283 | } | |
1284 | ||
1285 | /* Nearly general operand, but accept any const_double, since we wish | |
1286 | to be able to drop them into memory rather than have them get pulled | |
1287 | into registers. */ | |
1288 | ||
1289 | int | |
1290 | cmp_fp_expander_operand (op, mode) | |
1291 | register rtx op; | |
1292 | enum machine_mode mode; | |
1293 | { | |
1294 | if (mode != VOIDmode && mode != GET_MODE (op)) | |
1295 | return 0; | |
1296 | if (GET_CODE (op) == CONST_DOUBLE) | |
1297 | return 1; | |
1298 | return general_operand (op, mode); | |
1299 | } | |
1300 | ||
1301 | /* Match an SI or HImode register for a zero_extract. */ | |
1302 | ||
1303 | int | |
1304 | ext_register_operand (op, mode) | |
1305 | register rtx op; | |
1306 | enum machine_mode mode ATTRIBUTE_UNUSED; | |
1307 | { | |
1308 | if (GET_MODE (op) != SImode && GET_MODE (op) != HImode) | |
1309 | return 0; | |
1310 | return register_operand (op, VOIDmode); | |
1311 | } | |
1312 | ||
1313 | /* Return 1 if this is a valid binary floating-point operation. | |
1314 | OP is the expression matched, and MODE is its mode. */ | |
1315 | ||
1316 | int | |
1317 | binary_fp_operator (op, mode) | |
1318 | register rtx op; | |
1319 | enum machine_mode mode; | |
1320 | { | |
1321 | if (mode != VOIDmode && mode != GET_MODE (op)) | |
1322 | return 0; | |
1323 | ||
1324 | switch (GET_CODE (op)) | |
1325 | { | |
1326 | case PLUS: | |
1327 | case MINUS: | |
1328 | case MULT: | |
1329 | case DIV: | |
1330 | return GET_MODE_CLASS (GET_MODE (op)) == MODE_FLOAT; | |
1331 | ||
1332 | default: | |
1333 | return 0; | |
1334 | } | |
1335 | } | |
1336 | ||
1337 | int | |
1338 | mult_operator(op, mode) | |
1339 | register rtx op; | |
1340 | enum machine_mode mode ATTRIBUTE_UNUSED; | |
1341 | { | |
1342 | return GET_CODE (op) == MULT; | |
1343 | } | |
1344 | ||
1345 | int | |
1346 | div_operator(op, mode) | |
1347 | register rtx op; | |
1348 | enum machine_mode mode ATTRIBUTE_UNUSED; | |
1349 | { | |
1350 | return GET_CODE (op) == DIV; | |
1351 | } | |
1352 | ||
1353 | int | |
1354 | arith_or_logical_operator (op, mode) | |
1355 | rtx op; | |
1356 | enum machine_mode mode; | |
1357 | { | |
1358 | return ((mode == VOIDmode || GET_MODE (op) == mode) | |
1359 | && (GET_RTX_CLASS (GET_CODE (op)) == 'c' | |
1360 | || GET_RTX_CLASS (GET_CODE (op)) == '2')); | |
1361 | } | |
1362 | ||
1363 | /* Returns 1 if OP is memory operand with a displacement. */ | |
1364 | ||
1365 | int | |
1366 | memory_displacement_operand (op, mode) | |
1367 | register rtx op; | |
1368 | enum machine_mode mode; | |
1369 | { | |
1370 | struct ix86_address parts; | |
1371 | ||
1372 | if (! memory_operand (op, mode)) | |
1373 | return 0; | |
1374 | ||
1375 | if (! ix86_decompose_address (XEXP (op, 0), &parts)) | |
1376 | abort (); | |
1377 | ||
1378 | return parts.disp != NULL_RTX; | |
1379 | } | |
1380 | ||
1381 | /* To avoid problems when jump re-emits comparisons like testqi_ext_0, | |
1382 | re-recognize the operand to avoid a copy_to_mode_reg that will fail. | |
1383 | ||
1384 | ??? It seems likely that this will only work because cmpsi is an | |
1385 | expander, and no actual insns use this. */ | |
1386 | ||
1387 | int | |
1388 | cmpsi_operand (op, mode) | |
1389 | rtx op; | |
1390 | enum machine_mode mode; | |
1391 | { | |
1392 | if (general_operand (op, mode)) | |
1393 | return 1; | |
1394 | ||
1395 | if (GET_CODE (op) == AND | |
1396 | && GET_MODE (op) == SImode | |
1397 | && GET_CODE (XEXP (op, 0)) == ZERO_EXTRACT | |
1398 | && GET_CODE (XEXP (XEXP (op, 0), 1)) == CONST_INT | |
1399 | && GET_CODE (XEXP (XEXP (op, 0), 2)) == CONST_INT | |
1400 | && INTVAL (XEXP (XEXP (op, 0), 1)) == 8 | |
1401 | && INTVAL (XEXP (XEXP (op, 0), 2)) == 8 | |
1402 | && GET_CODE (XEXP (op, 1)) == CONST_INT) | |
1403 | return 1; | |
1404 | ||
1405 | return 0; | |
1406 | } | |
1407 | ||
1408 | /* Returns 1 if OP is memory operand that can not be represented by the | |
1409 | modRM array. */ | |
1410 | ||
1411 | int | |
1412 | long_memory_operand (op, mode) | |
1413 | register rtx op; | |
1414 | enum machine_mode mode; | |
1415 | { | |
1416 | if (! memory_operand (op, mode)) | |
1417 | return 0; | |
1418 | ||
1419 | return memory_address_length (op) != 0; | |
1420 | } | |
1421 | ||
1422 | /* Return nonzero if the rtx is known aligned. */ | |
1423 | ||
1424 | int | |
1425 | aligned_operand (op, mode) | |
1426 | rtx op; | |
1427 | enum machine_mode mode; | |
1428 | { | |
1429 | struct ix86_address parts; | |
1430 | ||
1431 | if (!general_operand (op, mode)) | |
1432 | return 0; | |
1433 | ||
1434 | /* Registers and immediate operands are always "aligned". */ | |
1435 | if (GET_CODE (op) != MEM) | |
1436 | return 1; | |
1437 | ||
1438 | /* Don't even try to do any aligned optimizations with volatiles. */ | |
1439 | if (MEM_VOLATILE_P (op)) | |
1440 | return 0; | |
1441 | ||
1442 | op = XEXP (op, 0); | |
1443 | ||
1444 | /* Pushes and pops are only valid on the stack pointer. */ | |
1445 | if (GET_CODE (op) == PRE_DEC | |
1446 | || GET_CODE (op) == POST_INC) | |
1447 | return 1; | |
1448 | ||
1449 | /* Decode the address. */ | |
1450 | if (! ix86_decompose_address (op, &parts)) | |
1451 | abort (); | |
1452 | ||
1453 | /* Look for some component that isn't known to be aligned. */ | |
1454 | if (parts.index) | |
1455 | { | |
1456 | if (parts.scale < 4 | |
1457 | && REGNO_POINTER_ALIGN (REGNO (parts.index)) < 4) | |
1458 | return 0; | |
1459 | } | |
1460 | if (parts.base) | |
1461 | { | |
1462 | if (REGNO_POINTER_ALIGN (REGNO (parts.base)) < 4) | |
1463 | return 0; | |
1464 | } | |
1465 | if (parts.disp) | |
1466 | { | |
1467 | if (GET_CODE (parts.disp) != CONST_INT | |
1468 | || (INTVAL (parts.disp) & 3) != 0) | |
1469 | return 0; | |
1470 | } | |
1471 | ||
1472 | /* Didn't find one -- this must be an aligned address. */ | |
1473 | return 1; | |
1474 | } | |
1475 | \f | |
1476 | /* Return true if the constant is something that can be loaded with | |
1477 | a special instruction. Only handle 0.0 and 1.0; others are less | |
1478 | worthwhile. */ | |
1479 | ||
1480 | int | |
1481 | standard_80387_constant_p (x) | |
1482 | rtx x; | |
1483 | { | |
1484 | if (GET_CODE (x) != CONST_DOUBLE) | |
1485 | return -1; | |
1486 | ||
1487 | #if ! defined (REAL_IS_NOT_DOUBLE) || defined (REAL_ARITHMETIC) | |
1488 | { | |
1489 | REAL_VALUE_TYPE d; | |
1490 | jmp_buf handler; | |
1491 | int is0, is1; | |
1492 | ||
1493 | if (setjmp (handler)) | |
1494 | return 0; | |
1495 | ||
1496 | set_float_handler (handler); | |
1497 | REAL_VALUE_FROM_CONST_DOUBLE (d, x); | |
1498 | is0 = REAL_VALUES_EQUAL (d, dconst0) && !REAL_VALUE_MINUS_ZERO (d); | |
1499 | is1 = REAL_VALUES_EQUAL (d, dconst1); | |
1500 | set_float_handler (NULL_PTR); | |
1501 | ||
1502 | if (is0) | |
1503 | return 1; | |
1504 | ||
1505 | if (is1) | |
1506 | return 2; | |
1507 | ||
1508 | /* Note that on the 80387, other constants, such as pi, | |
1509 | are much slower to load as standard constants | |
1510 | than to load from doubles in memory! */ | |
1511 | /* ??? Not true on K6: all constants are equal cost. */ | |
1512 | } | |
1513 | #endif | |
1514 | ||
1515 | return 0; | |
1516 | } | |
1517 | ||
1518 | /* Returns 1 if OP contains a symbol reference */ | |
1519 | ||
1520 | int | |
1521 | symbolic_reference_mentioned_p (op) | |
1522 | rtx op; | |
1523 | { | |
1524 | register const char *fmt; | |
1525 | register int i; | |
1526 | ||
1527 | if (GET_CODE (op) == SYMBOL_REF || GET_CODE (op) == LABEL_REF) | |
1528 | return 1; | |
1529 | ||
1530 | fmt = GET_RTX_FORMAT (GET_CODE (op)); | |
1531 | for (i = GET_RTX_LENGTH (GET_CODE (op)) - 1; i >= 0; i--) | |
1532 | { | |
1533 | if (fmt[i] == 'E') | |
1534 | { | |
1535 | register int j; | |
1536 | ||
1537 | for (j = XVECLEN (op, i) - 1; j >= 0; j--) | |
1538 | if (symbolic_reference_mentioned_p (XVECEXP (op, i, j))) | |
1539 | return 1; | |
1540 | } | |
1541 | ||
1542 | else if (fmt[i] == 'e' && symbolic_reference_mentioned_p (XEXP (op, i))) | |
1543 | return 1; | |
1544 | } | |
1545 | ||
1546 | return 0; | |
1547 | } | |
1548 | ||
1549 | /* Return 1 if it is appropriate to emit `ret' instructions in the | |
1550 | body of a function. Do this only if the epilogue is simple, needing a | |
1551 | couple of insns. Prior to reloading, we can't tell how many registers | |
1552 | must be saved, so return 0 then. Return 0 if there is no frame | |
1553 | marker to de-allocate. | |
1554 | ||
1555 | If NON_SAVING_SETJMP is defined and true, then it is not possible | |
1556 | for the epilogue to be simple, so return 0. This is a special case | |
1557 | since NON_SAVING_SETJMP will not cause regs_ever_live to change | |
1558 | until final, but jump_optimize may need to know sooner if a | |
1559 | `return' is OK. */ | |
1560 | ||
1561 | int | |
1562 | ix86_can_use_return_insn_p () | |
1563 | { | |
1564 | HOST_WIDE_INT tsize; | |
1565 | int nregs; | |
1566 | ||
1567 | #ifdef NON_SAVING_SETJMP | |
1568 | if (NON_SAVING_SETJMP && current_function_calls_setjmp) | |
1569 | return 0; | |
1570 | #endif | |
1571 | #ifdef FUNCTION_BLOCK_PROFILER_EXIT | |
1572 | if (profile_block_flag == 2) | |
1573 | return 0; | |
1574 | #endif | |
1575 | ||
1576 | if (! reload_completed || frame_pointer_needed) | |
1577 | return 0; | |
1578 | ||
1579 | /* Don't allow more than 32 pop, since that's all we can do | |
1580 | with one instruction. */ | |
1581 | if (current_function_pops_args | |
1582 | && current_function_args_size >= 32768) | |
1583 | return 0; | |
1584 | ||
1585 | tsize = ix86_compute_frame_size (get_frame_size (), &nregs, NULL, NULL); | |
1586 | return tsize == 0 && nregs == 0; | |
1587 | } | |
1588 | \f | |
1589 | static char *pic_label_name; | |
1590 | static int pic_label_output; | |
1591 | static char *global_offset_table_name; | |
1592 | ||
1593 | /* This function generates code for -fpic that loads %ebx with | |
1594 | the return address of the caller and then returns. */ | |
1595 | ||
1596 | void | |
1597 | asm_output_function_prefix (file, name) | |
1598 | FILE *file; | |
1599 | const char *name ATTRIBUTE_UNUSED; | |
1600 | { | |
1601 | rtx xops[2]; | |
1602 | int pic_reg_used = flag_pic && (current_function_uses_pic_offset_table | |
1603 | || current_function_uses_const_pool); | |
1604 | xops[0] = pic_offset_table_rtx; | |
1605 | xops[1] = stack_pointer_rtx; | |
1606 | ||
1607 | /* Deep branch prediction favors having a return for every call. */ | |
1608 | if (pic_reg_used && TARGET_DEEP_BRANCH_PREDICTION) | |
1609 | { | |
1610 | if (!pic_label_output) | |
1611 | { | |
1612 | /* This used to call ASM_DECLARE_FUNCTION_NAME() but since it's an | |
1613 | internal (non-global) label that's being emitted, it didn't make | |
1614 | sense to have .type information for local labels. This caused | |
1615 | the SCO OpenServer 5.0.4 ELF assembler grief (why are you giving | |
1616 | me debug info for a label that you're declaring non-global?) this | |
1617 | was changed to call ASM_OUTPUT_LABEL() instead. */ | |
1618 | ||
1619 | ASM_OUTPUT_LABEL (file, pic_label_name); | |
1620 | ||
1621 | xops[1] = gen_rtx_MEM (SImode, xops[1]); | |
1622 | output_asm_insn ("mov{l}\t{%1, %0|%0, %1}", xops); | |
1623 | output_asm_insn ("ret", xops); | |
1624 | ||
1625 | pic_label_output = 1; | |
1626 | } | |
1627 | } | |
1628 | } | |
1629 | ||
1630 | void | |
1631 | load_pic_register () | |
1632 | { | |
1633 | rtx gotsym, pclab; | |
1634 | ||
1635 | if (global_offset_table_name == NULL) | |
1636 | { | |
1637 | global_offset_table_name = | |
1638 | ggc_alloc_string ("_GLOBAL_OFFSET_TABLE_", 21); | |
1639 | ggc_add_string_root (&global_offset_table_name, 1); | |
1640 | } | |
1641 | gotsym = gen_rtx_SYMBOL_REF (Pmode, global_offset_table_name); | |
1642 | ||
1643 | if (TARGET_DEEP_BRANCH_PREDICTION) | |
1644 | { | |
1645 | if (pic_label_name == NULL) | |
1646 | { | |
1647 | pic_label_name = ggc_alloc_string (NULL, 32); | |
1648 | ggc_add_string_root (&pic_label_name, 1); | |
1649 | ASM_GENERATE_INTERNAL_LABEL (pic_label_name, "LPR", 0); | |
1650 | } | |
1651 | pclab = gen_rtx_MEM (QImode, gen_rtx_SYMBOL_REF (Pmode, pic_label_name)); | |
1652 | } | |
1653 | else | |
1654 | { | |
1655 | pclab = gen_rtx_LABEL_REF (VOIDmode, gen_label_rtx ()); | |
1656 | } | |
1657 | ||
1658 | emit_insn (gen_prologue_get_pc (pic_offset_table_rtx, pclab)); | |
1659 | ||
1660 | if (! TARGET_DEEP_BRANCH_PREDICTION) | |
1661 | emit_insn (gen_popsi1 (pic_offset_table_rtx)); | |
1662 | ||
1663 | emit_insn (gen_prologue_set_got (pic_offset_table_rtx, gotsym, pclab)); | |
1664 | } | |
1665 | ||
1666 | /* Generate an SImode "push" pattern for input ARG. */ | |
1667 | ||
1668 | static rtx | |
1669 | gen_push (arg) | |
1670 | rtx arg; | |
1671 | { | |
1672 | return gen_rtx_SET (VOIDmode, | |
1673 | gen_rtx_MEM (SImode, | |
1674 | gen_rtx_PRE_DEC (SImode, | |
1675 | stack_pointer_rtx)), | |
1676 | arg); | |
1677 | } | |
1678 | ||
1679 | /* Return number of registers to be saved on the stack. */ | |
1680 | ||
1681 | static int | |
1682 | ix86_nsaved_regs () | |
1683 | { | |
1684 | int nregs = 0; | |
1685 | int pic_reg_used = flag_pic && (current_function_uses_pic_offset_table | |
1686 | || current_function_uses_const_pool); | |
1687 | int limit = (frame_pointer_needed | |
1688 | ? HARD_FRAME_POINTER_REGNUM : STACK_POINTER_REGNUM); | |
1689 | int regno; | |
1690 | ||
1691 | for (regno = limit - 1; regno >= 0; regno--) | |
1692 | if ((regs_ever_live[regno] && ! call_used_regs[regno]) | |
1693 | || (regno == PIC_OFFSET_TABLE_REGNUM && pic_reg_used)) | |
1694 | { | |
1695 | nregs ++; | |
1696 | } | |
1697 | return nregs; | |
1698 | } | |
1699 | ||
1700 | /* Return the offset between two registers, one to be eliminated, and the other | |
1701 | its replacement, at the start of a routine. */ | |
1702 | ||
1703 | HOST_WIDE_INT | |
1704 | ix86_initial_elimination_offset (from, to) | |
1705 | int from; | |
1706 | int to; | |
1707 | { | |
1708 | int padding1; | |
1709 | int nregs; | |
1710 | ||
1711 | /* Stack grows downward: | |
1712 | ||
1713 | [arguments] | |
1714 | <- ARG_POINTER | |
1715 | saved pc | |
1716 | ||
1717 | saved frame pointer if frame_pointer_needed | |
1718 | <- HARD_FRAME_POINTER | |
1719 | [saved regs] | |
1720 | ||
1721 | [padding1] \ | |
1722 | | <- FRAME_POINTER | |
1723 | [frame] > tsize | |
1724 | | | |
1725 | [padding2] / | |
1726 | */ | |
1727 | ||
1728 | if (from == ARG_POINTER_REGNUM && to == HARD_FRAME_POINTER_REGNUM) | |
1729 | /* Skip saved PC and previous frame pointer. | |
1730 | Executed only when frame_pointer_needed. */ | |
1731 | return 8; | |
1732 | else if (from == FRAME_POINTER_REGNUM | |
1733 | && to == HARD_FRAME_POINTER_REGNUM) | |
1734 | { | |
1735 | ix86_compute_frame_size (get_frame_size (), &nregs, &padding1, (int *)0); | |
1736 | padding1 += nregs * UNITS_PER_WORD; | |
1737 | return -padding1; | |
1738 | } | |
1739 | else | |
1740 | { | |
1741 | /* ARG_POINTER or FRAME_POINTER to STACK_POINTER elimination. */ | |
1742 | int frame_size = frame_pointer_needed ? 8 : 4; | |
1743 | HOST_WIDE_INT tsize = ix86_compute_frame_size (get_frame_size (), | |
1744 | &nregs, &padding1, (int *)0); | |
1745 | ||
1746 | ||
1747 | if (to != STACK_POINTER_REGNUM) | |
1748 | abort (); | |
1749 | else if (from == ARG_POINTER_REGNUM) | |
1750 | return tsize + nregs * UNITS_PER_WORD + frame_size; | |
1751 | else if (from != FRAME_POINTER_REGNUM) | |
1752 | abort (); | |
1753 | else | |
1754 | return tsize - padding1; | |
1755 | } | |
1756 | } | |
1757 | ||
1758 | /* Compute the size of local storage taking into consideration the | |
1759 | desired stack alignment which is to be maintained. Also determine | |
1760 | the number of registers saved below the local storage. | |
1761 | ||
1762 | PADDING1 returns padding before stack frame and PADDING2 returns | |
1763 | padding after stack frame; | |
1764 | */ | |
1765 | ||
1766 | static HOST_WIDE_INT | |
1767 | ix86_compute_frame_size (size, nregs_on_stack, rpadding1, rpadding2) | |
1768 | HOST_WIDE_INT size; | |
1769 | int *nregs_on_stack; | |
1770 | int *rpadding1; | |
1771 | int *rpadding2; | |
1772 | { | |
1773 | int nregs; | |
1774 | int padding1 = 0; | |
1775 | int padding2 = 0; | |
1776 | HOST_WIDE_INT total_size; | |
1777 | int stack_alignment_needed = cfun->stack_alignment_needed / BITS_PER_UNIT; | |
1778 | int offset; | |
1779 | int preferred_alignment = cfun->preferred_stack_boundary / BITS_PER_UNIT; | |
1780 | ||
1781 | nregs = ix86_nsaved_regs (); | |
1782 | total_size = size; | |
1783 | ||
1784 | offset = frame_pointer_needed ? 8 : 4; | |
1785 | ||
1786 | /* Do some sanity checking of stack_alignment_needed and preferred_alignment, | |
1787 | since i386 port is the only using those features that may break easilly. */ | |
1788 | ||
1789 | if (size && !stack_alignment_needed) | |
1790 | abort (); | |
1791 | if (!size && stack_alignment_needed != STACK_BOUNDARY / BITS_PER_UNIT) | |
1792 | abort (); | |
1793 | if (preferred_alignment < STACK_BOUNDARY / BITS_PER_UNIT) | |
1794 | abort (); | |
1795 | if (preferred_alignment > PREFERRED_STACK_BOUNDARY / BITS_PER_UNIT) | |
1796 | abort (); | |
1797 | if (stack_alignment_needed > PREFERRED_STACK_BOUNDARY / BITS_PER_UNIT) | |
1798 | abort (); | |
1799 | ||
1800 | if (stack_alignment_needed < 4) | |
1801 | stack_alignment_needed = 4; | |
1802 | ||
1803 | offset += nregs * UNITS_PER_WORD; | |
1804 | ||
1805 | total_size += offset; | |
1806 | ||
1807 | /* Align start of frame for local function. */ | |
1808 | padding1 = ((offset + stack_alignment_needed - 1) | |
1809 | & -stack_alignment_needed) - offset; | |
1810 | total_size += padding1; | |
1811 | ||
1812 | /* Align stack boundary. */ | |
1813 | padding2 = ((total_size + preferred_alignment - 1) | |
1814 | & -preferred_alignment) - total_size; | |
1815 | ||
1816 | if (nregs_on_stack) | |
1817 | *nregs_on_stack = nregs; | |
1818 | if (rpadding1) | |
1819 | *rpadding1 = padding1; | |
1820 | if (rpadding2) | |
1821 | *rpadding2 = padding2; | |
1822 | ||
1823 | return size + padding1 + padding2; | |
1824 | } | |
1825 | ||
1826 | /* Emit code to save registers in the prologue. */ | |
1827 | ||
1828 | static void | |
1829 | ix86_emit_save_regs () | |
1830 | { | |
1831 | register int regno; | |
1832 | int limit; | |
1833 | rtx insn; | |
1834 | int pic_reg_used = flag_pic && (current_function_uses_pic_offset_table | |
1835 | || current_function_uses_const_pool); | |
1836 | limit = (frame_pointer_needed | |
1837 | ? HARD_FRAME_POINTER_REGNUM : STACK_POINTER_REGNUM); | |
1838 | ||
1839 | for (regno = limit - 1; regno >= 0; regno--) | |
1840 | if ((regs_ever_live[regno] && !call_used_regs[regno]) | |
1841 | || (regno == PIC_OFFSET_TABLE_REGNUM && pic_reg_used)) | |
1842 | { | |
1843 | insn = emit_insn (gen_push (gen_rtx_REG (SImode, regno))); | |
1844 | RTX_FRAME_RELATED_P (insn) = 1; | |
1845 | } | |
1846 | } | |
1847 | ||
1848 | /* Expand the prologue into a bunch of separate insns. */ | |
1849 | ||
1850 | void | |
1851 | ix86_expand_prologue () | |
1852 | { | |
1853 | HOST_WIDE_INT tsize = ix86_compute_frame_size (get_frame_size (), (int *)0, (int *)0, | |
1854 | (int *)0); | |
1855 | rtx insn; | |
1856 | int pic_reg_used = flag_pic && (current_function_uses_pic_offset_table | |
1857 | || current_function_uses_const_pool); | |
1858 | ||
1859 | /* Note: AT&T enter does NOT have reversed args. Enter is probably | |
1860 | slower on all targets. Also sdb doesn't like it. */ | |
1861 | ||
1862 | if (frame_pointer_needed) | |
1863 | { | |
1864 | insn = emit_insn (gen_push (hard_frame_pointer_rtx)); | |
1865 | RTX_FRAME_RELATED_P (insn) = 1; | |
1866 | ||
1867 | insn = emit_move_insn (hard_frame_pointer_rtx, stack_pointer_rtx); | |
1868 | RTX_FRAME_RELATED_P (insn) = 1; | |
1869 | } | |
1870 | ||
1871 | ix86_emit_save_regs (); | |
1872 | ||
1873 | if (tsize == 0) | |
1874 | ; | |
1875 | else if (! TARGET_STACK_PROBE || tsize < CHECK_STACK_LIMIT) | |
1876 | { | |
1877 | if (frame_pointer_needed) | |
1878 | insn = emit_insn (gen_pro_epilogue_adjust_stack | |
1879 | (stack_pointer_rtx, stack_pointer_rtx, | |
1880 | GEN_INT (-tsize), hard_frame_pointer_rtx)); | |
1881 | else | |
1882 | insn = emit_insn (gen_addsi3 (stack_pointer_rtx, stack_pointer_rtx, | |
1883 | GEN_INT (-tsize))); | |
1884 | RTX_FRAME_RELATED_P (insn) = 1; | |
1885 | } | |
1886 | else | |
1887 | { | |
1888 | /* ??? Is this only valid for Win32? */ | |
1889 | ||
1890 | rtx arg0, sym; | |
1891 | ||
1892 | arg0 = gen_rtx_REG (SImode, 0); | |
1893 | emit_move_insn (arg0, GEN_INT (tsize)); | |
1894 | ||
1895 | sym = gen_rtx_MEM (FUNCTION_MODE, | |
1896 | gen_rtx_SYMBOL_REF (Pmode, "_alloca")); | |
1897 | insn = emit_call_insn (gen_call (sym, const0_rtx)); | |
1898 | ||
1899 | CALL_INSN_FUNCTION_USAGE (insn) | |
1900 | = gen_rtx_EXPR_LIST (VOIDmode, gen_rtx_USE (VOIDmode, arg0), | |
1901 | CALL_INSN_FUNCTION_USAGE (insn)); | |
1902 | } | |
1903 | ||
1904 | #ifdef SUBTARGET_PROLOGUE | |
1905 | SUBTARGET_PROLOGUE; | |
1906 | #endif | |
1907 | ||
1908 | if (pic_reg_used) | |
1909 | load_pic_register (); | |
1910 | ||
1911 | /* If we are profiling, make sure no instructions are scheduled before | |
1912 | the call to mcount. However, if -fpic, the above call will have | |
1913 | done that. */ | |
1914 | if ((profile_flag || profile_block_flag) && ! pic_reg_used) | |
1915 | emit_insn (gen_blockage ()); | |
1916 | } | |
1917 | ||
1918 | /* Emit code to add TSIZE to esp value. Use POP instruction when | |
1919 | profitable. */ | |
1920 | ||
1921 | static void | |
1922 | ix86_emit_epilogue_esp_adjustment (tsize) | |
1923 | int tsize; | |
1924 | { | |
1925 | /* Intel's docs say that for 4 or 8 bytes of stack frame one should | |
1926 | use `pop' and not `add'. */ | |
1927 | int use_pop = tsize == 4; | |
1928 | rtx edx = 0, ecx; | |
1929 | ||
1930 | /* Use two pops only for the Pentium processors. */ | |
1931 | if (tsize == 8 && !TARGET_386 && !TARGET_486) | |
1932 | { | |
1933 | rtx retval = current_function_return_rtx; | |
1934 | ||
1935 | edx = gen_rtx_REG (SImode, 1); | |
1936 | ||
1937 | /* This case is a bit more complex. Since we cannot pop into | |
1938 | %ecx twice we need a second register. But this is only | |
1939 | available if the return value is not of DImode in which | |
1940 | case the %edx register is not available. */ | |
1941 | use_pop = (retval == NULL | |
1942 | || !reg_overlap_mentioned_p (edx, retval)); | |
1943 | } | |
1944 | ||
1945 | if (use_pop) | |
1946 | { | |
1947 | ecx = gen_rtx_REG (SImode, 2); | |
1948 | ||
1949 | /* We have to prevent the two pops here from being scheduled. | |
1950 | GCC otherwise would try in some situation to put other | |
1951 | instructions in between them which has a bad effect. */ | |
1952 | emit_insn (gen_blockage ()); | |
1953 | emit_insn (gen_popsi1 (ecx)); | |
1954 | if (tsize == 8) | |
1955 | emit_insn (gen_popsi1 (edx)); | |
1956 | } | |
1957 | else | |
1958 | { | |
1959 | /* If a frame pointer is present, we must be sure to tie the sp | |
1960 | to the fp so that we don't mis-schedule. */ | |
1961 | if (frame_pointer_needed) | |
1962 | emit_insn (gen_pro_epilogue_adjust_stack (stack_pointer_rtx, | |
1963 | stack_pointer_rtx, | |
1964 | GEN_INT (tsize), | |
1965 | hard_frame_pointer_rtx)); | |
1966 | else | |
1967 | emit_insn (gen_addsi3 (stack_pointer_rtx, stack_pointer_rtx, | |
1968 | GEN_INT (tsize))); | |
1969 | } | |
1970 | } | |
1971 | ||
1972 | /* Emit code to restore saved registers using MOV insns. First register | |
1973 | is restored from POINTER + OFFSET. */ | |
1974 | static void | |
1975 | ix86_emit_restore_regs_using_mov (pointer, offset) | |
1976 | rtx pointer; | |
1977 | int offset; | |
1978 | { | |
1979 | int regno; | |
1980 | int pic_reg_used = flag_pic && (current_function_uses_pic_offset_table | |
1981 | || current_function_uses_const_pool); | |
1982 | int limit = (frame_pointer_needed | |
1983 | ? HARD_FRAME_POINTER_REGNUM : STACK_POINTER_REGNUM); | |
1984 | ||
1985 | for (regno = 0; regno < limit; regno++) | |
1986 | if ((regs_ever_live[regno] && !call_used_regs[regno]) | |
1987 | || (regno == PIC_OFFSET_TABLE_REGNUM && pic_reg_used)) | |
1988 | { | |
1989 | emit_move_insn (gen_rtx_REG (SImode, regno), | |
1990 | adj_offsettable_operand (gen_rtx_MEM (SImode, | |
1991 | pointer), | |
1992 | offset)); | |
1993 | offset += 4; | |
1994 | } | |
1995 | } | |
1996 | ||
1997 | /* Restore function stack, frame, and registers. */ | |
1998 | ||
1999 | void | |
2000 | ix86_expand_epilogue () | |
2001 | { | |
2002 | int nregs; | |
2003 | int regno; | |
2004 | ||
2005 | int pic_reg_used = flag_pic && (current_function_uses_pic_offset_table | |
2006 | || current_function_uses_const_pool); | |
2007 | int sp_valid = !frame_pointer_needed || current_function_sp_is_unchanging; | |
2008 | HOST_WIDE_INT offset; | |
2009 | HOST_WIDE_INT tsize = ix86_compute_frame_size (get_frame_size (), &nregs, | |
2010 | (int *)0, (int *)0); | |
2011 | ||
2012 | ||
2013 | /* Calculate start of saved registers relative to ebp. */ | |
2014 | offset = -nregs * UNITS_PER_WORD; | |
2015 | ||
2016 | #ifdef FUNCTION_BLOCK_PROFILER_EXIT | |
2017 | if (profile_block_flag == 2) | |
2018 | { | |
2019 | FUNCTION_BLOCK_PROFILER_EXIT; | |
2020 | } | |
2021 | #endif | |
2022 | ||
2023 | /* If we're only restoring one register and sp is not valid then | |
2024 | using a move instruction to restore the register since it's | |
2025 | less work than reloading sp and popping the register. | |
2026 | ||
2027 | The default code result in stack adjustment using add/lea instruction, | |
2028 | while this code results in LEAVE instruction (or discrete equivalent), | |
2029 | so it is profitable in some other cases as well. Especially when there | |
2030 | are no registers to restore. We also use this code when TARGET_USE_LEAVE | |
2031 | and there is exactly one register to pop. This heruistic may need some | |
2032 | tuning in future. */ | |
2033 | if ((!sp_valid && nregs <= 1) | |
2034 | || (frame_pointer_needed && !nregs && tsize) | |
2035 | || (frame_pointer_needed && TARGET_USE_LEAVE && !optimize_size | |
2036 | && nregs == 1)) | |
2037 | { | |
2038 | /* Restore registers. We can use ebp or esp to address the memory | |
2039 | locations. If both are available, default to ebp, since offsets | |
2040 | are known to be small. Only exception is esp pointing directly to the | |
2041 | end of block of saved registers, where we may simplify addressing | |
2042 | mode. */ | |
2043 | ||
2044 | if (!frame_pointer_needed || (sp_valid && !tsize)) | |
2045 | ix86_emit_restore_regs_using_mov (stack_pointer_rtx, tsize); | |
2046 | else | |
2047 | ix86_emit_restore_regs_using_mov (hard_frame_pointer_rtx, offset); | |
2048 | ||
2049 | if (!frame_pointer_needed) | |
2050 | ix86_emit_epilogue_esp_adjustment (tsize + nregs * UNITS_PER_WORD); | |
2051 | /* If not an i386, mov & pop is faster than "leave". */ | |
2052 | else if (TARGET_USE_LEAVE || optimize_size) | |
2053 | emit_insn (gen_leave ()); | |
2054 | else | |
2055 | { | |
2056 | emit_insn (gen_pro_epilogue_adjust_stack (stack_pointer_rtx, | |
2057 | hard_frame_pointer_rtx, | |
2058 | const0_rtx, | |
2059 | hard_frame_pointer_rtx)); | |
2060 | emit_insn (gen_popsi1 (hard_frame_pointer_rtx)); | |
2061 | } | |
2062 | } | |
2063 | else | |
2064 | { | |
2065 | /* First step is to deallocate the stack frame so that we can | |
2066 | pop the registers. */ | |
2067 | if (!sp_valid) | |
2068 | { | |
2069 | if (!frame_pointer_needed) | |
2070 | abort (); | |
2071 | emit_insn (gen_pro_epilogue_adjust_stack (stack_pointer_rtx, | |
2072 | hard_frame_pointer_rtx, | |
2073 | GEN_INT (offset), | |
2074 | hard_frame_pointer_rtx)); | |
2075 | } | |
2076 | else if (tsize) | |
2077 | ix86_emit_epilogue_esp_adjustment (tsize); | |
2078 | ||
2079 | for (regno = 0; regno < STACK_POINTER_REGNUM; regno++) | |
2080 | if ((regs_ever_live[regno] && !call_used_regs[regno]) | |
2081 | || (regno == PIC_OFFSET_TABLE_REGNUM && pic_reg_used)) | |
2082 | emit_insn (gen_popsi1 (gen_rtx_REG (SImode, regno))); | |
2083 | } | |
2084 | ||
2085 | if (current_function_pops_args && current_function_args_size) | |
2086 | { | |
2087 | rtx popc = GEN_INT (current_function_pops_args); | |
2088 | ||
2089 | /* i386 can only pop 32K bytes (maybe 64K? Is it signed?). If | |
2090 | asked to pop more, pop return address, do explicit add, and jump | |
2091 | indirectly to the caller. */ | |
2092 | ||
2093 | if (current_function_pops_args >= 32768) | |
2094 | { | |
2095 | rtx ecx = gen_rtx_REG (SImode, 2); | |
2096 | ||
2097 | emit_insn (gen_popsi1 (ecx)); | |
2098 | emit_insn (gen_addsi3 (stack_pointer_rtx, stack_pointer_rtx, popc)); | |
2099 | emit_indirect_jump (ecx); | |
2100 | } | |
2101 | else | |
2102 | emit_jump_insn (gen_return_pop_internal (popc)); | |
2103 | } | |
2104 | else | |
2105 | emit_jump_insn (gen_return_internal ()); | |
2106 | } | |
2107 | \f | |
2108 | /* Extract the parts of an RTL expression that is a valid memory address | |
2109 | for an instruction. Return false if the structure of the address is | |
2110 | grossly off. */ | |
2111 | ||
2112 | static int | |
2113 | ix86_decompose_address (addr, out) | |
2114 | register rtx addr; | |
2115 | struct ix86_address *out; | |
2116 | { | |
2117 | rtx base = NULL_RTX; | |
2118 | rtx index = NULL_RTX; | |
2119 | rtx disp = NULL_RTX; | |
2120 | HOST_WIDE_INT scale = 1; | |
2121 | rtx scale_rtx = NULL_RTX; | |
2122 | ||
2123 | if (GET_CODE (addr) == REG || GET_CODE (addr) == SUBREG) | |
2124 | base = addr; | |
2125 | else if (GET_CODE (addr) == PLUS) | |
2126 | { | |
2127 | rtx op0 = XEXP (addr, 0); | |
2128 | rtx op1 = XEXP (addr, 1); | |
2129 | enum rtx_code code0 = GET_CODE (op0); | |
2130 | enum rtx_code code1 = GET_CODE (op1); | |
2131 | ||
2132 | if (code0 == REG || code0 == SUBREG) | |
2133 | { | |
2134 | if (code1 == REG || code1 == SUBREG) | |
2135 | index = op0, base = op1; /* index + base */ | |
2136 | else | |
2137 | base = op0, disp = op1; /* base + displacement */ | |
2138 | } | |
2139 | else if (code0 == MULT) | |
2140 | { | |
2141 | index = XEXP (op0, 0); | |
2142 | scale_rtx = XEXP (op0, 1); | |
2143 | if (code1 == REG || code1 == SUBREG) | |
2144 | base = op1; /* index*scale + base */ | |
2145 | else | |
2146 | disp = op1; /* index*scale + disp */ | |
2147 | } | |
2148 | else if (code0 == PLUS && GET_CODE (XEXP (op0, 0)) == MULT) | |
2149 | { | |
2150 | index = XEXP (XEXP (op0, 0), 0); /* index*scale + base + disp */ | |
2151 | scale_rtx = XEXP (XEXP (op0, 0), 1); | |
2152 | base = XEXP (op0, 1); | |
2153 | disp = op1; | |
2154 | } | |
2155 | else if (code0 == PLUS) | |
2156 | { | |
2157 | index = XEXP (op0, 0); /* index + base + disp */ | |
2158 | base = XEXP (op0, 1); | |
2159 | disp = op1; | |
2160 | } | |
2161 | else | |
2162 | return FALSE; | |
2163 | } | |
2164 | else if (GET_CODE (addr) == MULT) | |
2165 | { | |
2166 | index = XEXP (addr, 0); /* index*scale */ | |
2167 | scale_rtx = XEXP (addr, 1); | |
2168 | } | |
2169 | else if (GET_CODE (addr) == ASHIFT) | |
2170 | { | |
2171 | rtx tmp; | |
2172 | ||
2173 | /* We're called for lea too, which implements ashift on occasion. */ | |
2174 | index = XEXP (addr, 0); | |
2175 | tmp = XEXP (addr, 1); | |
2176 | if (GET_CODE (tmp) != CONST_INT) | |
2177 | return FALSE; | |
2178 | scale = INTVAL (tmp); | |
2179 | if ((unsigned HOST_WIDE_INT) scale > 3) | |
2180 | return FALSE; | |
2181 | scale = 1 << scale; | |
2182 | } | |
2183 | else | |
2184 | disp = addr; /* displacement */ | |
2185 | ||
2186 | /* Extract the integral value of scale. */ | |
2187 | if (scale_rtx) | |
2188 | { | |
2189 | if (GET_CODE (scale_rtx) != CONST_INT) | |
2190 | return FALSE; | |
2191 | scale = INTVAL (scale_rtx); | |
2192 | } | |
2193 | ||
2194 | /* Allow arg pointer and stack pointer as index if there is not scaling */ | |
2195 | if (base && index && scale == 1 | |
2196 | && (index == arg_pointer_rtx || index == frame_pointer_rtx | |
2197 | || index == stack_pointer_rtx)) | |
2198 | { | |
2199 | rtx tmp = base; | |
2200 | base = index; | |
2201 | index = tmp; | |
2202 | } | |
2203 | ||
2204 | /* Special case: %ebp cannot be encoded as a base without a displacement. */ | |
2205 | if ((base == hard_frame_pointer_rtx | |
2206 | || base == frame_pointer_rtx | |
2207 | || base == arg_pointer_rtx) && !disp) | |
2208 | disp = const0_rtx; | |
2209 | ||
2210 | /* Special case: on K6, [%esi] makes the instruction vector decoded. | |
2211 | Avoid this by transforming to [%esi+0]. */ | |
2212 | if (ix86_cpu == PROCESSOR_K6 && !optimize_size | |
2213 | && base && !index && !disp | |
2214 | && REG_P (base) | |
2215 | && REGNO_REG_CLASS (REGNO (base)) == SIREG) | |
2216 | disp = const0_rtx; | |
2217 | ||
2218 | /* Special case: encode reg+reg instead of reg*2. */ | |
2219 | if (!base && index && scale && scale == 2) | |
2220 | base = index, scale = 1; | |
2221 | ||
2222 | /* Special case: scaling cannot be encoded without base or displacement. */ | |
2223 | if (!base && !disp && index && scale != 1) | |
2224 | disp = const0_rtx; | |
2225 | ||
2226 | out->base = base; | |
2227 | out->index = index; | |
2228 | out->disp = disp; | |
2229 | out->scale = scale; | |
2230 | ||
2231 | return TRUE; | |
2232 | } | |
2233 | ||
2234 | /* Determine if a given CONST RTX is a valid memory displacement | |
2235 | in PIC mode. */ | |
2236 | ||
2237 | int | |
2238 | legitimate_pic_address_disp_p (disp) | |
2239 | register rtx disp; | |
2240 | { | |
2241 | if (GET_CODE (disp) != CONST) | |
2242 | return 0; | |
2243 | disp = XEXP (disp, 0); | |
2244 | ||
2245 | if (GET_CODE (disp) == PLUS) | |
2246 | { | |
2247 | if (GET_CODE (XEXP (disp, 1)) != CONST_INT) | |
2248 | return 0; | |
2249 | disp = XEXP (disp, 0); | |
2250 | } | |
2251 | ||
2252 | if (GET_CODE (disp) != UNSPEC | |
2253 | || XVECLEN (disp, 0) != 1) | |
2254 | return 0; | |
2255 | ||
2256 | /* Must be @GOT or @GOTOFF. */ | |
2257 | if (XINT (disp, 1) != 6 | |
2258 | && XINT (disp, 1) != 7) | |
2259 | return 0; | |
2260 | ||
2261 | if (GET_CODE (XVECEXP (disp, 0, 0)) != SYMBOL_REF | |
2262 | && GET_CODE (XVECEXP (disp, 0, 0)) != LABEL_REF) | |
2263 | return 0; | |
2264 | ||
2265 | return 1; | |
2266 | } | |
2267 | ||
2268 | /* GO_IF_LEGITIMATE_ADDRESS recognizes an RTL expression that is a valid | |
2269 | memory address for an instruction. The MODE argument is the machine mode | |
2270 | for the MEM expression that wants to use this address. | |
2271 | ||
2272 | It only recognizes address in canonical form. LEGITIMIZE_ADDRESS should | |
2273 | convert common non-canonical forms to canonical form so that they will | |
2274 | be recognized. */ | |
2275 | ||
2276 | int | |
2277 | legitimate_address_p (mode, addr, strict) | |
2278 | enum machine_mode mode; | |
2279 | register rtx addr; | |
2280 | int strict; | |
2281 | { | |
2282 | struct ix86_address parts; | |
2283 | rtx base, index, disp; | |
2284 | HOST_WIDE_INT scale; | |
2285 | const char *reason = NULL; | |
2286 | rtx reason_rtx = NULL_RTX; | |
2287 | ||
2288 | if (TARGET_DEBUG_ADDR) | |
2289 | { | |
2290 | fprintf (stderr, | |
2291 | "\n======\nGO_IF_LEGITIMATE_ADDRESS, mode = %s, strict = %d\n", | |
2292 | GET_MODE_NAME (mode), strict); | |
2293 | debug_rtx (addr); | |
2294 | } | |
2295 | ||
2296 | if (! ix86_decompose_address (addr, &parts)) | |
2297 | { | |
2298 | reason = "decomposition failed"; | |
2299 | goto error; | |
2300 | } | |
2301 | ||
2302 | base = parts.base; | |
2303 | index = parts.index; | |
2304 | disp = parts.disp; | |
2305 | scale = parts.scale; | |
2306 | ||
2307 | /* Validate base register. | |
2308 | ||
2309 | Don't allow SUBREG's here, it can lead to spill failures when the base | |
2310 | is one word out of a two word structure, which is represented internally | |
2311 | as a DImode int. */ | |
2312 | ||
2313 | if (base) | |
2314 | { | |
2315 | reason_rtx = base; | |
2316 | ||
2317 | if (GET_CODE (base) != REG) | |
2318 | { | |
2319 | reason = "base is not a register"; | |
2320 | goto error; | |
2321 | } | |
2322 | ||
2323 | if (GET_MODE (base) != Pmode) | |
2324 | { | |
2325 | reason = "base is not in Pmode"; | |
2326 | goto error; | |
2327 | } | |
2328 | ||
2329 | if ((strict && ! REG_OK_FOR_BASE_STRICT_P (base)) | |
2330 | || (! strict && ! REG_OK_FOR_BASE_NONSTRICT_P (base))) | |
2331 | { | |
2332 | reason = "base is not valid"; | |
2333 | goto error; | |
2334 | } | |
2335 | } | |
2336 | ||
2337 | /* Validate index register. | |
2338 | ||
2339 | Don't allow SUBREG's here, it can lead to spill failures when the index | |
2340 | is one word out of a two word structure, which is represented internally | |
2341 | as a DImode int. */ | |
2342 | ||
2343 | if (index) | |
2344 | { | |
2345 | reason_rtx = index; | |
2346 | ||
2347 | if (GET_CODE (index) != REG) | |
2348 | { | |
2349 | reason = "index is not a register"; | |
2350 | goto error; | |
2351 | } | |
2352 | ||
2353 | if (GET_MODE (index) != Pmode) | |
2354 | { | |
2355 | reason = "index is not in Pmode"; | |
2356 | goto error; | |
2357 | } | |
2358 | ||
2359 | if ((strict && ! REG_OK_FOR_INDEX_STRICT_P (index)) | |
2360 | || (! strict && ! REG_OK_FOR_INDEX_NONSTRICT_P (index))) | |
2361 | { | |
2362 | reason = "index is not valid"; | |
2363 | goto error; | |
2364 | } | |
2365 | } | |
2366 | ||
2367 | /* Validate scale factor. */ | |
2368 | if (scale != 1) | |
2369 | { | |
2370 | reason_rtx = GEN_INT (scale); | |
2371 | if (!index) | |
2372 | { | |
2373 | reason = "scale without index"; | |
2374 | goto error; | |
2375 | } | |
2376 | ||
2377 | if (scale != 2 && scale != 4 && scale != 8) | |
2378 | { | |
2379 | reason = "scale is not a valid multiplier"; | |
2380 | goto error; | |
2381 | } | |
2382 | } | |
2383 | ||
2384 | /* Validate displacement. */ | |
2385 | if (disp) | |
2386 | { | |
2387 | reason_rtx = disp; | |
2388 | ||
2389 | if (!CONSTANT_ADDRESS_P (disp)) | |
2390 | { | |
2391 | reason = "displacement is not constant"; | |
2392 | goto error; | |
2393 | } | |
2394 | ||
2395 | if (GET_CODE (disp) == CONST_DOUBLE) | |
2396 | { | |
2397 | reason = "displacement is a const_double"; | |
2398 | goto error; | |
2399 | } | |
2400 | ||
2401 | if (flag_pic && SYMBOLIC_CONST (disp)) | |
2402 | { | |
2403 | if (! legitimate_pic_address_disp_p (disp)) | |
2404 | { | |
2405 | reason = "displacement is an invalid pic construct"; | |
2406 | goto error; | |
2407 | } | |
2408 | ||
2409 | /* Verify that a symbolic pic displacement includes | |
2410 | the pic_offset_table_rtx register. */ | |
2411 | if (base != pic_offset_table_rtx | |
2412 | && (index != pic_offset_table_rtx || scale != 1)) | |
2413 | { | |
2414 | reason = "pic displacement against invalid base"; | |
2415 | goto error; | |
2416 | } | |
2417 | } | |
2418 | else if (HALF_PIC_P ()) | |
2419 | { | |
2420 | if (! HALF_PIC_ADDRESS_P (disp) | |
2421 | || (base != NULL_RTX || index != NULL_RTX)) | |
2422 | { | |
2423 | reason = "displacement is an invalid half-pic reference"; | |
2424 | goto error; | |
2425 | } | |
2426 | } | |
2427 | } | |
2428 | ||
2429 | /* Everything looks valid. */ | |
2430 | if (TARGET_DEBUG_ADDR) | |
2431 | fprintf (stderr, "Success.\n"); | |
2432 | return TRUE; | |
2433 | ||
2434 | error: | |
2435 | if (TARGET_DEBUG_ADDR) | |
2436 | { | |
2437 | fprintf (stderr, "Error: %s\n", reason); | |
2438 | debug_rtx (reason_rtx); | |
2439 | } | |
2440 | return FALSE; | |
2441 | } | |
2442 | \f | |
2443 | /* Return a legitimate reference for ORIG (an address) using the | |
2444 | register REG. If REG is 0, a new pseudo is generated. | |
2445 | ||
2446 | There are two types of references that must be handled: | |
2447 | ||
2448 | 1. Global data references must load the address from the GOT, via | |
2449 | the PIC reg. An insn is emitted to do this load, and the reg is | |
2450 | returned. | |
2451 | ||
2452 | 2. Static data references, constant pool addresses, and code labels | |
2453 | compute the address as an offset from the GOT, whose base is in | |
2454 | the PIC reg. Static data objects have SYMBOL_REF_FLAG set to | |
2455 | differentiate them from global data objects. The returned | |
2456 | address is the PIC reg + an unspec constant. | |
2457 | ||
2458 | GO_IF_LEGITIMATE_ADDRESS rejects symbolic references unless the PIC | |
2459 | reg also appears in the address. */ | |
2460 | ||
2461 | rtx | |
2462 | legitimize_pic_address (orig, reg) | |
2463 | rtx orig; | |
2464 | rtx reg; | |
2465 | { | |
2466 | rtx addr = orig; | |
2467 | rtx new = orig; | |
2468 | rtx base; | |
2469 | ||
2470 | if (GET_CODE (addr) == LABEL_REF | |
2471 | || (GET_CODE (addr) == SYMBOL_REF | |
2472 | && (CONSTANT_POOL_ADDRESS_P (addr) | |
2473 | || SYMBOL_REF_FLAG (addr)))) | |
2474 | { | |
2475 | /* This symbol may be referenced via a displacement from the PIC | |
2476 | base address (@GOTOFF). */ | |
2477 | ||
2478 | current_function_uses_pic_offset_table = 1; | |
2479 | new = gen_rtx_UNSPEC (VOIDmode, gen_rtvec (1, addr), 7); | |
2480 | new = gen_rtx_CONST (VOIDmode, new); | |
2481 | new = gen_rtx_PLUS (Pmode, pic_offset_table_rtx, new); | |
2482 | ||
2483 | if (reg != 0) | |
2484 | { | |
2485 | emit_move_insn (reg, new); | |
2486 | new = reg; | |
2487 | } | |
2488 | } | |
2489 | else if (GET_CODE (addr) == SYMBOL_REF) | |
2490 | { | |
2491 | /* This symbol must be referenced via a load from the | |
2492 | Global Offset Table (@GOT). */ | |
2493 | ||
2494 | current_function_uses_pic_offset_table = 1; | |
2495 | new = gen_rtx_UNSPEC (VOIDmode, gen_rtvec (1, addr), 6); | |
2496 | new = gen_rtx_CONST (VOIDmode, new); | |
2497 | new = gen_rtx_PLUS (Pmode, pic_offset_table_rtx, new); | |
2498 | new = gen_rtx_MEM (Pmode, new); | |
2499 | RTX_UNCHANGING_P (new) = 1; | |
2500 | ||
2501 | if (reg == 0) | |
2502 | reg = gen_reg_rtx (Pmode); | |
2503 | emit_move_insn (reg, new); | |
2504 | new = reg; | |
2505 | } | |
2506 | else | |
2507 | { | |
2508 | if (GET_CODE (addr) == CONST) | |
2509 | { | |
2510 | addr = XEXP (addr, 0); | |
2511 | if (GET_CODE (addr) == UNSPEC) | |
2512 | { | |
2513 | /* Check that the unspec is one of the ones we generate? */ | |
2514 | } | |
2515 | else if (GET_CODE (addr) != PLUS) | |
2516 | abort (); | |
2517 | } | |
2518 | if (GET_CODE (addr) == PLUS) | |
2519 | { | |
2520 | rtx op0 = XEXP (addr, 0), op1 = XEXP (addr, 1); | |
2521 | ||
2522 | /* Check first to see if this is a constant offset from a @GOTOFF | |
2523 | symbol reference. */ | |
2524 | if ((GET_CODE (op0) == LABEL_REF | |
2525 | || (GET_CODE (op0) == SYMBOL_REF | |
2526 | && (CONSTANT_POOL_ADDRESS_P (op0) | |
2527 | || SYMBOL_REF_FLAG (op0)))) | |
2528 | && GET_CODE (op1) == CONST_INT) | |
2529 | { | |
2530 | current_function_uses_pic_offset_table = 1; | |
2531 | new = gen_rtx_UNSPEC (VOIDmode, gen_rtvec (1, op0), 7); | |
2532 | new = gen_rtx_PLUS (VOIDmode, new, op1); | |
2533 | new = gen_rtx_CONST (VOIDmode, new); | |
2534 | new = gen_rtx_PLUS (Pmode, pic_offset_table_rtx, new); | |
2535 | ||
2536 | if (reg != 0) | |
2537 | { | |
2538 | emit_move_insn (reg, new); | |
2539 | new = reg; | |
2540 | } | |
2541 | } | |
2542 | else | |
2543 | { | |
2544 | base = legitimize_pic_address (XEXP (addr, 0), reg); | |
2545 | new = legitimize_pic_address (XEXP (addr, 1), | |
2546 | base == reg ? NULL_RTX : reg); | |
2547 | ||
2548 | if (GET_CODE (new) == CONST_INT) | |
2549 | new = plus_constant (base, INTVAL (new)); | |
2550 | else | |
2551 | { | |
2552 | if (GET_CODE (new) == PLUS && CONSTANT_P (XEXP (new, 1))) | |
2553 | { | |
2554 | base = gen_rtx_PLUS (Pmode, base, XEXP (new, 0)); | |
2555 | new = XEXP (new, 1); | |
2556 | } | |
2557 | new = gen_rtx_PLUS (Pmode, base, new); | |
2558 | } | |
2559 | } | |
2560 | } | |
2561 | } | |
2562 | return new; | |
2563 | } | |
2564 | \f | |
2565 | /* Try machine-dependent ways of modifying an illegitimate address | |
2566 | to be legitimate. If we find one, return the new, valid address. | |
2567 | This macro is used in only one place: `memory_address' in explow.c. | |
2568 | ||
2569 | OLDX is the address as it was before break_out_memory_refs was called. | |
2570 | In some cases it is useful to look at this to decide what needs to be done. | |
2571 | ||
2572 | MODE and WIN are passed so that this macro can use | |
2573 | GO_IF_LEGITIMATE_ADDRESS. | |
2574 | ||
2575 | It is always safe for this macro to do nothing. It exists to recognize | |
2576 | opportunities to optimize the output. | |
2577 | ||
2578 | For the 80386, we handle X+REG by loading X into a register R and | |
2579 | using R+REG. R will go in a general reg and indexing will be used. | |
2580 | However, if REG is a broken-out memory address or multiplication, | |
2581 | nothing needs to be done because REG can certainly go in a general reg. | |
2582 | ||
2583 | When -fpic is used, special handling is needed for symbolic references. | |
2584 | See comments by legitimize_pic_address in i386.c for details. */ | |
2585 | ||
2586 | rtx | |
2587 | legitimize_address (x, oldx, mode) | |
2588 | register rtx x; | |
2589 | register rtx oldx ATTRIBUTE_UNUSED; | |
2590 | enum machine_mode mode; | |
2591 | { | |
2592 | int changed = 0; | |
2593 | unsigned log; | |
2594 | ||
2595 | if (TARGET_DEBUG_ADDR) | |
2596 | { | |
2597 | fprintf (stderr, "\n==========\nLEGITIMIZE_ADDRESS, mode = %s\n", | |
2598 | GET_MODE_NAME (mode)); | |
2599 | debug_rtx (x); | |
2600 | } | |
2601 | ||
2602 | if (flag_pic && SYMBOLIC_CONST (x)) | |
2603 | return legitimize_pic_address (x, 0); | |
2604 | ||
2605 | /* Canonicalize shifts by 0, 1, 2, 3 into multiply */ | |
2606 | if (GET_CODE (x) == ASHIFT | |
2607 | && GET_CODE (XEXP (x, 1)) == CONST_INT | |
2608 | && (log = (unsigned)exact_log2 (INTVAL (XEXP (x, 1)))) < 4) | |
2609 | { | |
2610 | changed = 1; | |
2611 | x = gen_rtx_MULT (Pmode, force_reg (Pmode, XEXP (x, 0)), | |
2612 | GEN_INT (1 << log)); | |
2613 | } | |
2614 | ||
2615 | if (GET_CODE (x) == PLUS) | |
2616 | { | |
2617 | /* Canonicalize shifts by 0, 1, 2, 3 into multiply. */ | |
2618 | ||
2619 | if (GET_CODE (XEXP (x, 0)) == ASHIFT | |
2620 | && GET_CODE (XEXP (XEXP (x, 0), 1)) == CONST_INT | |
2621 | && (log = (unsigned)exact_log2 (INTVAL (XEXP (XEXP (x, 0), 1)))) < 4) | |
2622 | { | |
2623 | changed = 1; | |
2624 | XEXP (x, 0) = gen_rtx_MULT (Pmode, | |
2625 | force_reg (Pmode, XEXP (XEXP (x, 0), 0)), | |
2626 | GEN_INT (1 << log)); | |
2627 | } | |
2628 | ||
2629 | if (GET_CODE (XEXP (x, 1)) == ASHIFT | |
2630 | && GET_CODE (XEXP (XEXP (x, 1), 1)) == CONST_INT | |
2631 | && (log = (unsigned)exact_log2 (INTVAL (XEXP (XEXP (x, 1), 1)))) < 4) | |
2632 | { | |
2633 | changed = 1; | |
2634 | XEXP (x, 1) = gen_rtx_MULT (Pmode, | |
2635 | force_reg (Pmode, XEXP (XEXP (x, 1), 0)), | |
2636 | GEN_INT (1 << log)); | |
2637 | } | |
2638 | ||
2639 | /* Put multiply first if it isn't already. */ | |
2640 | if (GET_CODE (XEXP (x, 1)) == MULT) | |
2641 | { | |
2642 | rtx tmp = XEXP (x, 0); | |
2643 | XEXP (x, 0) = XEXP (x, 1); | |
2644 | XEXP (x, 1) = tmp; | |
2645 | changed = 1; | |
2646 | } | |
2647 | ||
2648 | /* Canonicalize (plus (mult (reg) (const)) (plus (reg) (const))) | |
2649 | into (plus (plus (mult (reg) (const)) (reg)) (const)). This can be | |
2650 | created by virtual register instantiation, register elimination, and | |
2651 | similar optimizations. */ | |
2652 | if (GET_CODE (XEXP (x, 0)) == MULT && GET_CODE (XEXP (x, 1)) == PLUS) | |
2653 | { | |
2654 | changed = 1; | |
2655 | x = gen_rtx_PLUS (Pmode, | |
2656 | gen_rtx_PLUS (Pmode, XEXP (x, 0), | |
2657 | XEXP (XEXP (x, 1), 0)), | |
2658 | XEXP (XEXP (x, 1), 1)); | |
2659 | } | |
2660 | ||
2661 | /* Canonicalize | |
2662 | (plus (plus (mult (reg) (const)) (plus (reg) (const))) const) | |
2663 | into (plus (plus (mult (reg) (const)) (reg)) (const)). */ | |
2664 | else if (GET_CODE (x) == PLUS && GET_CODE (XEXP (x, 0)) == PLUS | |
2665 | && GET_CODE (XEXP (XEXP (x, 0), 0)) == MULT | |
2666 | && GET_CODE (XEXP (XEXP (x, 0), 1)) == PLUS | |
2667 | && CONSTANT_P (XEXP (x, 1))) | |
2668 | { | |
2669 | rtx constant; | |
2670 | rtx other = NULL_RTX; | |
2671 | ||
2672 | if (GET_CODE (XEXP (x, 1)) == CONST_INT) | |
2673 | { | |
2674 | constant = XEXP (x, 1); | |
2675 | other = XEXP (XEXP (XEXP (x, 0), 1), 1); | |
2676 | } | |
2677 | else if (GET_CODE (XEXP (XEXP (XEXP (x, 0), 1), 1)) == CONST_INT) | |
2678 | { | |
2679 | constant = XEXP (XEXP (XEXP (x, 0), 1), 1); | |
2680 | other = XEXP (x, 1); | |
2681 | } | |
2682 | else | |
2683 | constant = 0; | |
2684 | ||
2685 | if (constant) | |
2686 | { | |
2687 | changed = 1; | |
2688 | x = gen_rtx_PLUS (Pmode, | |
2689 | gen_rtx_PLUS (Pmode, XEXP (XEXP (x, 0), 0), | |
2690 | XEXP (XEXP (XEXP (x, 0), 1), 0)), | |
2691 | plus_constant (other, INTVAL (constant))); | |
2692 | } | |
2693 | } | |
2694 | ||
2695 | if (changed && legitimate_address_p (mode, x, FALSE)) | |
2696 | return x; | |
2697 | ||
2698 | if (GET_CODE (XEXP (x, 0)) == MULT) | |
2699 | { | |
2700 | changed = 1; | |
2701 | XEXP (x, 0) = force_operand (XEXP (x, 0), 0); | |
2702 | } | |
2703 | ||
2704 | if (GET_CODE (XEXP (x, 1)) == MULT) | |
2705 | { | |
2706 | changed = 1; | |
2707 | XEXP (x, 1) = force_operand (XEXP (x, 1), 0); | |
2708 | } | |
2709 | ||
2710 | if (changed | |
2711 | && GET_CODE (XEXP (x, 1)) == REG | |
2712 | && GET_CODE (XEXP (x, 0)) == REG) | |
2713 | return x; | |
2714 | ||
2715 | if (flag_pic && SYMBOLIC_CONST (XEXP (x, 1))) | |
2716 | { | |
2717 | changed = 1; | |
2718 | x = legitimize_pic_address (x, 0); | |
2719 | } | |
2720 | ||
2721 | if (changed && legitimate_address_p (mode, x, FALSE)) | |
2722 | return x; | |
2723 | ||
2724 | if (GET_CODE (XEXP (x, 0)) == REG) | |
2725 | { | |
2726 | register rtx temp = gen_reg_rtx (Pmode); | |
2727 | register rtx val = force_operand (XEXP (x, 1), temp); | |
2728 | if (val != temp) | |
2729 | emit_move_insn (temp, val); | |
2730 | ||
2731 | XEXP (x, 1) = temp; | |
2732 | return x; | |
2733 | } | |
2734 | ||
2735 | else if (GET_CODE (XEXP (x, 1)) == REG) | |
2736 | { | |
2737 | register rtx temp = gen_reg_rtx (Pmode); | |
2738 | register rtx val = force_operand (XEXP (x, 0), temp); | |
2739 | if (val != temp) | |
2740 | emit_move_insn (temp, val); | |
2741 | ||
2742 | XEXP (x, 0) = temp; | |
2743 | return x; | |
2744 | } | |
2745 | } | |
2746 | ||
2747 | return x; | |
2748 | } | |
2749 | \f | |
2750 | /* Print an integer constant expression in assembler syntax. Addition | |
2751 | and subtraction are the only arithmetic that may appear in these | |
2752 | expressions. FILE is the stdio stream to write to, X is the rtx, and | |
2753 | CODE is the operand print code from the output string. */ | |
2754 | ||
2755 | static void | |
2756 | output_pic_addr_const (file, x, code) | |
2757 | FILE *file; | |
2758 | rtx x; | |
2759 | int code; | |
2760 | { | |
2761 | char buf[256]; | |
2762 | ||
2763 | switch (GET_CODE (x)) | |
2764 | { | |
2765 | case PC: | |
2766 | if (flag_pic) | |
2767 | putc ('.', file); | |
2768 | else | |
2769 | abort (); | |
2770 | break; | |
2771 | ||
2772 | case SYMBOL_REF: | |
2773 | assemble_name (file, XSTR (x, 0)); | |
2774 | if (code == 'P' && ! SYMBOL_REF_FLAG (x)) | |
2775 | fputs ("@PLT", file); | |
2776 | break; | |
2777 | ||
2778 | case LABEL_REF: | |
2779 | x = XEXP (x, 0); | |
2780 | /* FALLTHRU */ | |
2781 | case CODE_LABEL: | |
2782 | ASM_GENERATE_INTERNAL_LABEL (buf, "L", CODE_LABEL_NUMBER (x)); | |
2783 | assemble_name (asm_out_file, buf); | |
2784 | break; | |
2785 | ||
2786 | case CONST_INT: | |
2787 | fprintf (file, HOST_WIDE_INT_PRINT_DEC, INTVAL (x)); | |
2788 | break; | |
2789 | ||
2790 | case CONST: | |
2791 | /* This used to output parentheses around the expression, | |
2792 | but that does not work on the 386 (either ATT or BSD assembler). */ | |
2793 | output_pic_addr_const (file, XEXP (x, 0), code); | |
2794 | break; | |
2795 | ||
2796 | case CONST_DOUBLE: | |
2797 | if (GET_MODE (x) == VOIDmode) | |
2798 | { | |
2799 | /* We can use %d if the number is <32 bits and positive. */ | |
2800 | if (CONST_DOUBLE_HIGH (x) || CONST_DOUBLE_LOW (x) < 0) | |
2801 | fprintf (file, "0x%lx%08lx", | |
2802 | (unsigned long) CONST_DOUBLE_HIGH (x), | |
2803 | (unsigned long) CONST_DOUBLE_LOW (x)); | |
2804 | else | |
2805 | fprintf (file, HOST_WIDE_INT_PRINT_DEC, CONST_DOUBLE_LOW (x)); | |
2806 | } | |
2807 | else | |
2808 | /* We can't handle floating point constants; | |
2809 | PRINT_OPERAND must handle them. */ | |
2810 | output_operand_lossage ("floating constant misused"); | |
2811 | break; | |
2812 | ||
2813 | case PLUS: | |
2814 | /* Some assemblers need integer constants to appear first. */ | |
2815 | if (GET_CODE (XEXP (x, 0)) == CONST_INT) | |
2816 | { | |
2817 | output_pic_addr_const (file, XEXP (x, 0), code); | |
2818 | putc ('+', file); | |
2819 | output_pic_addr_const (file, XEXP (x, 1), code); | |
2820 | } | |
2821 | else if (GET_CODE (XEXP (x, 1)) == CONST_INT) | |
2822 | { | |
2823 | output_pic_addr_const (file, XEXP (x, 1), code); | |
2824 | putc ('+', file); | |
2825 | output_pic_addr_const (file, XEXP (x, 0), code); | |
2826 | } | |
2827 | else | |
2828 | abort (); | |
2829 | break; | |
2830 | ||
2831 | case MINUS: | |
2832 | putc (ASSEMBLER_DIALECT ? '(' : '[', file); | |
2833 | output_pic_addr_const (file, XEXP (x, 0), code); | |
2834 | putc ('-', file); | |
2835 | output_pic_addr_const (file, XEXP (x, 1), code); | |
2836 | putc (ASSEMBLER_DIALECT ? ')' : ']', file); | |
2837 | break; | |
2838 | ||
2839 | case UNSPEC: | |
2840 | if (XVECLEN (x, 0) != 1) | |
2841 | abort (); | |
2842 | output_pic_addr_const (file, XVECEXP (x, 0, 0), code); | |
2843 | switch (XINT (x, 1)) | |
2844 | { | |
2845 | case 6: | |
2846 | fputs ("@GOT", file); | |
2847 | break; | |
2848 | case 7: | |
2849 | fputs ("@GOTOFF", file); | |
2850 | break; | |
2851 | case 8: | |
2852 | fputs ("@PLT", file); | |
2853 | break; | |
2854 | default: | |
2855 | output_operand_lossage ("invalid UNSPEC as operand"); | |
2856 | break; | |
2857 | } | |
2858 | break; | |
2859 | ||
2860 | default: | |
2861 | output_operand_lossage ("invalid expression as operand"); | |
2862 | } | |
2863 | } | |
2864 | ||
2865 | /* This is called from dwarfout.c via ASM_OUTPUT_DWARF_ADDR_CONST. | |
2866 | We need to handle our special PIC relocations. */ | |
2867 | ||
2868 | void | |
2869 | i386_dwarf_output_addr_const (file, x) | |
2870 | FILE *file; | |
2871 | rtx x; | |
2872 | { | |
2873 | fprintf (file, "\t%s\t", INT_ASM_OP); | |
2874 | if (flag_pic) | |
2875 | output_pic_addr_const (file, x, '\0'); | |
2876 | else | |
2877 | output_addr_const (file, x); | |
2878 | fputc ('\n', file); | |
2879 | } | |
2880 | ||
2881 | /* In the name of slightly smaller debug output, and to cater to | |
2882 | general assembler losage, recognize PIC+GOTOFF and turn it back | |
2883 | into a direct symbol reference. */ | |
2884 | ||
2885 | rtx | |
2886 | i386_simplify_dwarf_addr (orig_x) | |
2887 | rtx orig_x; | |
2888 | { | |
2889 | rtx x = orig_x; | |
2890 | ||
2891 | if (GET_CODE (x) != PLUS | |
2892 | || GET_CODE (XEXP (x, 0)) != REG | |
2893 | || GET_CODE (XEXP (x, 1)) != CONST) | |
2894 | return orig_x; | |
2895 | ||
2896 | x = XEXP (XEXP (x, 1), 0); | |
2897 | if (GET_CODE (x) == UNSPEC | |
2898 | && XINT (x, 1) == 7) | |
2899 | return XVECEXP (x, 0, 0); | |
2900 | ||
2901 | if (GET_CODE (x) == PLUS | |
2902 | && GET_CODE (XEXP (x, 0)) == UNSPEC | |
2903 | && GET_CODE (XEXP (x, 1)) == CONST_INT | |
2904 | && XINT (XEXP (x, 0), 1) == 7) | |
2905 | return gen_rtx_PLUS (VOIDmode, XVECEXP (XEXP (x, 0), 0, 0), XEXP (x, 1)); | |
2906 | ||
2907 | return orig_x; | |
2908 | } | |
2909 | \f | |
2910 | static void | |
2911 | put_condition_code (code, mode, reverse, fp, file) | |
2912 | enum rtx_code code; | |
2913 | enum machine_mode mode; | |
2914 | int reverse, fp; | |
2915 | FILE *file; | |
2916 | { | |
2917 | const char *suffix; | |
2918 | ||
2919 | if (reverse) | |
2920 | code = reverse_condition (code); | |
2921 | ||
2922 | switch (code) | |
2923 | { | |
2924 | case EQ: | |
2925 | suffix = "e"; | |
2926 | break; | |
2927 | case NE: | |
2928 | suffix = "ne"; | |
2929 | break; | |
2930 | case GT: | |
2931 | if (mode == CCNOmode) | |
2932 | abort (); | |
2933 | suffix = "g"; | |
2934 | break; | |
2935 | case GTU: | |
2936 | /* ??? Use "nbe" instead of "a" for fcmov losage on some assemblers. | |
2937 | Those same assemblers have the same but opposite losage on cmov. */ | |
2938 | suffix = fp ? "nbe" : "a"; | |
2939 | break; | |
2940 | case LT: | |
2941 | if (mode == CCNOmode) | |
2942 | suffix = "s"; | |
2943 | else | |
2944 | suffix = "l"; | |
2945 | break; | |
2946 | case LTU: | |
2947 | suffix = "b"; | |
2948 | break; | |
2949 | case GE: | |
2950 | if (mode == CCNOmode) | |
2951 | suffix = "ns"; | |
2952 | else | |
2953 | suffix = "ge"; | |
2954 | break; | |
2955 | case GEU: | |
2956 | /* ??? As above. */ | |
2957 | suffix = fp ? "nb" : "ae"; | |
2958 | break; | |
2959 | case LE: | |
2960 | if (mode == CCNOmode) | |
2961 | abort (); | |
2962 | suffix = "le"; | |
2963 | break; | |
2964 | case LEU: | |
2965 | suffix = "be"; | |
2966 | break; | |
2967 | default: | |
2968 | abort (); | |
2969 | } | |
2970 | fputs (suffix, file); | |
2971 | } | |
2972 | ||
2973 | void | |
2974 | print_reg (x, code, file) | |
2975 | rtx x; | |
2976 | int code; | |
2977 | FILE *file; | |
2978 | { | |
2979 | if (REGNO (x) == ARG_POINTER_REGNUM | |
2980 | || REGNO (x) == FRAME_POINTER_REGNUM | |
2981 | || REGNO (x) == FLAGS_REG | |
2982 | || REGNO (x) == FPSR_REG) | |
2983 | abort (); | |
2984 | ||
2985 | if (ASSEMBLER_DIALECT == 0 || USER_LABEL_PREFIX[0] == 0) | |
2986 | putc ('%', file); | |
2987 | ||
2988 | if (code == 'w') | |
2989 | code = 2; | |
2990 | else if (code == 'b') | |
2991 | code = 1; | |
2992 | else if (code == 'k') | |
2993 | code = 4; | |
2994 | else if (code == 'y') | |
2995 | code = 3; | |
2996 | else if (code == 'h') | |
2997 | code = 0; | |
2998 | else | |
2999 | code = GET_MODE_SIZE (GET_MODE (x)); | |
3000 | ||
3001 | switch (code) | |
3002 | { | |
3003 | case 3: | |
3004 | if (STACK_TOP_P (x)) | |
3005 | { | |
3006 | fputs ("st(0)", file); | |
3007 | break; | |
3008 | } | |
3009 | /* FALLTHRU */ | |
3010 | case 4: | |
3011 | case 8: | |
3012 | case 12: | |
3013 | if (! FP_REG_P (x)) | |
3014 | putc ('e', file); | |
3015 | /* FALLTHRU */ | |
3016 | case 2: | |
3017 | fputs (hi_reg_name[REGNO (x)], file); | |
3018 | break; | |
3019 | case 1: | |
3020 | fputs (qi_reg_name[REGNO (x)], file); | |
3021 | break; | |
3022 | case 0: | |
3023 | fputs (qi_high_reg_name[REGNO (x)], file); | |
3024 | break; | |
3025 | default: | |
3026 | abort (); | |
3027 | } | |
3028 | } | |
3029 | ||
3030 | /* Meaning of CODE: | |
3031 | L,W,B,Q,S,T -- print the opcode suffix for specified size of operand. | |
3032 | C -- print opcode suffix for set/cmov insn. | |
3033 | c -- like C, but print reversed condition | |
3034 | R -- print the prefix for register names. | |
3035 | z -- print the opcode suffix for the size of the current operand. | |
3036 | * -- print a star (in certain assembler syntax) | |
3037 | w -- print the operand as if it's a "word" (HImode) even if it isn't. | |
3038 | s -- print a shift double count, followed by the assemblers argument | |
3039 | delimiter. | |
3040 | b -- print the QImode name of the register for the indicated operand. | |
3041 | %b0 would print %al if operands[0] is reg 0. | |
3042 | w -- likewise, print the HImode name of the register. | |
3043 | k -- likewise, print the SImode name of the register. | |
3044 | h -- print the QImode name for a "high" register, either ah, bh, ch or dh. | |
3045 | y -- print "st(0)" instead of "st" as a register. */ | |
3046 | ||
3047 | void | |
3048 | print_operand (file, x, code) | |
3049 | FILE *file; | |
3050 | rtx x; | |
3051 | int code; | |
3052 | { | |
3053 | if (code) | |
3054 | { | |
3055 | switch (code) | |
3056 | { | |
3057 | case '*': | |
3058 | if (ASSEMBLER_DIALECT == 0) | |
3059 | putc ('*', file); | |
3060 | return; | |
3061 | ||
3062 | case 'L': | |
3063 | if (ASSEMBLER_DIALECT == 0) | |
3064 | putc ('l', file); | |
3065 | return; | |
3066 | ||
3067 | case 'W': | |
3068 | if (ASSEMBLER_DIALECT == 0) | |
3069 | putc ('w', file); | |
3070 | return; | |
3071 | ||
3072 | case 'B': | |
3073 | if (ASSEMBLER_DIALECT == 0) | |
3074 | putc ('b', file); | |
3075 | return; | |
3076 | ||
3077 | case 'Q': | |
3078 | if (ASSEMBLER_DIALECT == 0) | |
3079 | putc ('l', file); | |
3080 | return; | |
3081 | ||
3082 | case 'S': | |
3083 | if (ASSEMBLER_DIALECT == 0) | |
3084 | putc ('s', file); | |
3085 | return; | |
3086 | ||
3087 | case 'T': | |
3088 | if (ASSEMBLER_DIALECT == 0) | |
3089 | putc ('t', file); | |
3090 | return; | |
3091 | ||
3092 | case 'z': | |
3093 | /* 387 opcodes don't get size suffixes if the operands are | |
3094 | registers. */ | |
3095 | ||
3096 | if (STACK_REG_P (x)) | |
3097 | return; | |
3098 | ||
3099 | /* Intel syntax has no truck with instruction suffixes. */ | |
3100 | if (ASSEMBLER_DIALECT != 0) | |
3101 | return; | |
3102 | ||
3103 | /* this is the size of op from size of operand */ | |
3104 | switch (GET_MODE_SIZE (GET_MODE (x))) | |
3105 | { | |
3106 | case 1: | |
3107 | putc ('b', file); | |
3108 | return; | |
3109 | ||
3110 | case 2: | |
3111 | putc ('w', file); | |
3112 | return; | |
3113 | ||
3114 | case 4: | |
3115 | if (GET_MODE (x) == SFmode) | |
3116 | { | |
3117 | putc ('s', file); | |
3118 | return; | |
3119 | } | |
3120 | else | |
3121 | putc ('l', file); | |
3122 | return; | |
3123 | ||
3124 | case 12: | |
3125 | putc ('t', file); | |
3126 | return; | |
3127 | ||
3128 | case 8: | |
3129 | if (GET_MODE_CLASS (GET_MODE (x)) == MODE_INT) | |
3130 | { | |
3131 | #ifdef GAS_MNEMONICS | |
3132 | putc ('q', file); | |
3133 | #else | |
3134 | putc ('l', file); | |
3135 | putc ('l', file); | |
3136 | #endif | |
3137 | } | |
3138 | else | |
3139 | putc ('l', file); | |
3140 | return; | |
3141 | } | |
3142 | ||
3143 | case 'b': | |
3144 | case 'w': | |
3145 | case 'k': | |
3146 | case 'h': | |
3147 | case 'y': | |
3148 | case 'X': | |
3149 | case 'P': | |
3150 | break; | |
3151 | ||
3152 | case 's': | |
3153 | if (GET_CODE (x) == CONST_INT || ! SHIFT_DOUBLE_OMITS_COUNT) | |
3154 | { | |
3155 | PRINT_OPERAND (file, x, 0); | |
3156 | putc (',', file); | |
3157 | } | |
3158 | return; | |
3159 | ||
3160 | case 'C': | |
3161 | put_condition_code (GET_CODE (x), GET_MODE (XEXP (x, 0)), 0, 0, file); | |
3162 | return; | |
3163 | case 'F': | |
3164 | put_condition_code (GET_CODE (x), GET_MODE (XEXP (x, 0)), 0, 1, file); | |
3165 | return; | |
3166 | ||
3167 | /* Like above, but reverse condition */ | |
3168 | case 'c': | |
3169 | put_condition_code (GET_CODE (x), GET_MODE (XEXP (x, 0)), 1, 0, file); | |
3170 | return; | |
3171 | case 'f': | |
3172 | put_condition_code (GET_CODE (x), GET_MODE (XEXP (x, 0)), 1, 1, file); | |
3173 | return; | |
3174 | ||
3175 | default: | |
3176 | { | |
3177 | char str[50]; | |
3178 | sprintf (str, "invalid operand code `%c'", code); | |
3179 | output_operand_lossage (str); | |
3180 | } | |
3181 | } | |
3182 | } | |
3183 | ||
3184 | if (GET_CODE (x) == REG) | |
3185 | { | |
3186 | PRINT_REG (x, code, file); | |
3187 | } | |
3188 | ||
3189 | else if (GET_CODE (x) == MEM) | |
3190 | { | |
3191 | /* No `byte ptr' prefix for call instructions. */ | |
3192 | if (ASSEMBLER_DIALECT != 0 && code != 'X' && code != 'P') | |
3193 | { | |
3194 | const char * size; | |
3195 | switch (GET_MODE_SIZE (GET_MODE (x))) | |
3196 | { | |
3197 | case 1: size = "BYTE"; break; | |
3198 | case 2: size = "WORD"; break; | |
3199 | case 4: size = "DWORD"; break; | |
3200 | case 8: size = "QWORD"; break; | |
3201 | case 12: size = "XWORD"; break; | |
3202 | default: | |
3203 | abort (); | |
3204 | } | |
3205 | fputs (size, file); | |
3206 | fputs (" PTR ", file); | |
3207 | } | |
3208 | ||
3209 | x = XEXP (x, 0); | |
3210 | if (flag_pic && CONSTANT_ADDRESS_P (x)) | |
3211 | output_pic_addr_const (file, x, code); | |
3212 | else | |
3213 | output_address (x); | |
3214 | } | |
3215 | ||
3216 | else if (GET_CODE (x) == CONST_DOUBLE && GET_MODE (x) == SFmode) | |
3217 | { | |
3218 | REAL_VALUE_TYPE r; | |
3219 | long l; | |
3220 | ||
3221 | REAL_VALUE_FROM_CONST_DOUBLE (r, x); | |
3222 | REAL_VALUE_TO_TARGET_SINGLE (r, l); | |
3223 | ||
3224 | if (ASSEMBLER_DIALECT == 0) | |
3225 | putc ('$', file); | |
3226 | fprintf (file, "0x%lx", l); | |
3227 | } | |
3228 | ||
3229 | /* These float cases don't actually occur as immediate operands. */ | |
3230 | else if (GET_CODE (x) == CONST_DOUBLE && GET_MODE (x) == DFmode) | |
3231 | { | |
3232 | REAL_VALUE_TYPE r; | |
3233 | char dstr[30]; | |
3234 | ||
3235 | REAL_VALUE_FROM_CONST_DOUBLE (r, x); | |
3236 | REAL_VALUE_TO_DECIMAL (r, "%.22e", dstr); | |
3237 | fprintf (file, "%s", dstr); | |
3238 | } | |
3239 | ||
3240 | else if (GET_CODE (x) == CONST_DOUBLE && GET_MODE (x) == XFmode) | |
3241 | { | |
3242 | REAL_VALUE_TYPE r; | |
3243 | char dstr[30]; | |
3244 | ||
3245 | REAL_VALUE_FROM_CONST_DOUBLE (r, x); | |
3246 | REAL_VALUE_TO_DECIMAL (r, "%.22e", dstr); | |
3247 | fprintf (file, "%s", dstr); | |
3248 | } | |
3249 | else | |
3250 | { | |
3251 | if (code != 'P') | |
3252 | { | |
3253 | if (GET_CODE (x) == CONST_INT || GET_CODE (x) == CONST_DOUBLE) | |
3254 | { | |
3255 | if (ASSEMBLER_DIALECT == 0) | |
3256 | putc ('$', file); | |
3257 | } | |
3258 | else if (GET_CODE (x) == CONST || GET_CODE (x) == SYMBOL_REF | |
3259 | || GET_CODE (x) == LABEL_REF) | |
3260 | { | |
3261 | if (ASSEMBLER_DIALECT == 0) | |
3262 | putc ('$', file); | |
3263 | else | |
3264 | fputs ("OFFSET FLAT:", file); | |
3265 | } | |
3266 | } | |
3267 | if (GET_CODE (x) == CONST_INT) | |
3268 | fprintf (file, HOST_WIDE_INT_PRINT_DEC, INTVAL (x)); | |
3269 | else if (flag_pic) | |
3270 | output_pic_addr_const (file, x, code); | |
3271 | else | |
3272 | output_addr_const (file, x); | |
3273 | } | |
3274 | } | |
3275 | \f | |
3276 | /* Print a memory operand whose address is ADDR. */ | |
3277 | ||
3278 | void | |
3279 | print_operand_address (file, addr) | |
3280 | FILE *file; | |
3281 | register rtx addr; | |
3282 | { | |
3283 | struct ix86_address parts; | |
3284 | rtx base, index, disp; | |
3285 | int scale; | |
3286 | ||
3287 | if (! ix86_decompose_address (addr, &parts)) | |
3288 | abort (); | |
3289 | ||
3290 | base = parts.base; | |
3291 | index = parts.index; | |
3292 | disp = parts.disp; | |
3293 | scale = parts.scale; | |
3294 | ||
3295 | if (!base && !index) | |
3296 | { | |
3297 | /* Displacement only requires special attention. */ | |
3298 | ||
3299 | if (GET_CODE (disp) == CONST_INT) | |
3300 | { | |
3301 | if (ASSEMBLER_DIALECT != 0) | |
3302 | fputs ("ds:", file); | |
3303 | fprintf (file, HOST_WIDE_INT_PRINT_DEC, INTVAL (addr)); | |
3304 | } | |
3305 | else if (flag_pic) | |
3306 | output_pic_addr_const (file, addr, 0); | |
3307 | else | |
3308 | output_addr_const (file, addr); | |
3309 | } | |
3310 | else | |
3311 | { | |
3312 | if (ASSEMBLER_DIALECT == 0) | |
3313 | { | |
3314 | if (disp) | |
3315 | { | |
3316 | if (flag_pic) | |
3317 | output_pic_addr_const (file, disp, 0); | |
3318 | else if (GET_CODE (disp) == LABEL_REF) | |
3319 | output_asm_label (disp); | |
3320 | else | |
3321 | output_addr_const (file, disp); | |
3322 | } | |
3323 | ||
3324 | putc ('(', file); | |
3325 | if (base) | |
3326 | PRINT_REG (base, 0, file); | |
3327 | if (index) | |
3328 | { | |
3329 | putc (',', file); | |
3330 | PRINT_REG (index, 0, file); | |
3331 | if (scale != 1) | |
3332 | fprintf (file, ",%d", scale); | |
3333 | } | |
3334 | putc (')', file); | |
3335 | } | |
3336 | else | |
3337 | { | |
3338 | rtx offset = NULL_RTX; | |
3339 | ||
3340 | if (disp) | |
3341 | { | |
3342 | /* Pull out the offset of a symbol; print any symbol itself. */ | |
3343 | if (GET_CODE (disp) == CONST | |
3344 | && GET_CODE (XEXP (disp, 0)) == PLUS | |
3345 | && GET_CODE (XEXP (XEXP (disp, 0), 1)) == CONST_INT) | |
3346 | { | |
3347 | offset = XEXP (XEXP (disp, 0), 1); | |
3348 | disp = gen_rtx_CONST (VOIDmode, | |
3349 | XEXP (XEXP (disp, 0), 0)); | |
3350 | } | |
3351 | ||
3352 | if (flag_pic) | |
3353 | output_pic_addr_const (file, disp, 0); | |
3354 | else if (GET_CODE (disp) == LABEL_REF) | |
3355 | output_asm_label (disp); | |
3356 | else if (GET_CODE (disp) == CONST_INT) | |
3357 | offset = disp; | |
3358 | else | |
3359 | output_addr_const (file, disp); | |
3360 | } | |
3361 | ||
3362 | putc ('[', file); | |
3363 | if (base) | |
3364 | { | |
3365 | PRINT_REG (base, 0, file); | |
3366 | if (offset) | |
3367 | { | |
3368 | if (INTVAL (offset) >= 0) | |
3369 | putc ('+', file); | |
3370 | fprintf (file, HOST_WIDE_INT_PRINT_DEC, INTVAL (offset)); | |
3371 | } | |
3372 | } | |
3373 | else if (offset) | |
3374 | fprintf (file, HOST_WIDE_INT_PRINT_DEC, INTVAL (offset)); | |
3375 | else | |
3376 | putc ('0', file); | |
3377 | ||
3378 | if (index) | |
3379 | { | |
3380 | putc ('+', file); | |
3381 | PRINT_REG (index, 0, file); | |
3382 | if (scale != 1) | |
3383 | fprintf (file, "*%d", scale); | |
3384 | } | |
3385 | putc (']', file); | |
3386 | } | |
3387 | } | |
3388 | } | |
3389 | \f | |
3390 | /* Split one or more DImode RTL references into pairs of SImode | |
3391 | references. The RTL can be REG, offsettable MEM, integer constant, or | |
3392 | CONST_DOUBLE. "operands" is a pointer to an array of DImode RTL to | |
3393 | split and "num" is its length. lo_half and hi_half are output arrays | |
3394 | that parallel "operands". */ | |
3395 | ||
3396 | void | |
3397 | split_di (operands, num, lo_half, hi_half) | |
3398 | rtx operands[]; | |
3399 | int num; | |
3400 | rtx lo_half[], hi_half[]; | |
3401 | { | |
3402 | while (num--) | |
3403 | { | |
3404 | rtx op = operands[num]; | |
3405 | if (CONSTANT_P (op)) | |
3406 | split_double (op, &lo_half[num], &hi_half[num]); | |
3407 | else if (! reload_completed) | |
3408 | { | |
3409 | lo_half[num] = gen_lowpart (SImode, op); | |
3410 | hi_half[num] = gen_highpart (SImode, op); | |
3411 | } | |
3412 | else if (GET_CODE (op) == REG) | |
3413 | { | |
3414 | lo_half[num] = gen_rtx_REG (SImode, REGNO (op)); | |
3415 | hi_half[num] = gen_rtx_REG (SImode, REGNO (op) + 1); | |
3416 | } | |
3417 | else if (offsettable_memref_p (op)) | |
3418 | { | |
3419 | rtx lo_addr = XEXP (op, 0); | |
3420 | rtx hi_addr = XEXP (adj_offsettable_operand (op, 4), 0); | |
3421 | lo_half[num] = change_address (op, SImode, lo_addr); | |
3422 | hi_half[num] = change_address (op, SImode, hi_addr); | |
3423 | } | |
3424 | else | |
3425 | abort (); | |
3426 | } | |
3427 | } | |
3428 | \f | |
3429 | /* Output code to perform a 387 binary operation in INSN, one of PLUS, | |
3430 | MINUS, MULT or DIV. OPERANDS are the insn operands, where operands[3] | |
3431 | is the expression of the binary operation. The output may either be | |
3432 | emitted here, or returned to the caller, like all output_* functions. | |
3433 | ||
3434 | There is no guarantee that the operands are the same mode, as they | |
3435 | might be within FLOAT or FLOAT_EXTEND expressions. */ | |
3436 | ||
3437 | const char * | |
3438 | output_387_binary_op (insn, operands) | |
3439 | rtx insn; | |
3440 | rtx *operands; | |
3441 | { | |
3442 | static char buf[100]; | |
3443 | rtx temp; | |
3444 | const char *p; | |
3445 | ||
3446 | switch (GET_CODE (operands[3])) | |
3447 | { | |
3448 | case PLUS: | |
3449 | if (GET_MODE_CLASS (GET_MODE (operands[1])) == MODE_INT | |
3450 | || GET_MODE_CLASS (GET_MODE (operands[2])) == MODE_INT) | |
3451 | p = "fiadd"; | |
3452 | else | |
3453 | p = "fadd"; | |
3454 | break; | |
3455 | ||
3456 | case MINUS: | |
3457 | if (GET_MODE_CLASS (GET_MODE (operands[1])) == MODE_INT | |
3458 | || GET_MODE_CLASS (GET_MODE (operands[2])) == MODE_INT) | |
3459 | p = "fisub"; | |
3460 | else | |
3461 | p = "fsub"; | |
3462 | break; | |
3463 | ||
3464 | case MULT: | |
3465 | if (GET_MODE_CLASS (GET_MODE (operands[1])) == MODE_INT | |
3466 | || GET_MODE_CLASS (GET_MODE (operands[2])) == MODE_INT) | |
3467 | p = "fimul"; | |
3468 | else | |
3469 | p = "fmul"; | |
3470 | break; | |
3471 | ||
3472 | case DIV: | |
3473 | if (GET_MODE_CLASS (GET_MODE (operands[1])) == MODE_INT | |
3474 | || GET_MODE_CLASS (GET_MODE (operands[2])) == MODE_INT) | |
3475 | p = "fidiv"; | |
3476 | else | |
3477 | p = "fdiv"; | |
3478 | break; | |
3479 | ||
3480 | default: | |
3481 | abort (); | |
3482 | } | |
3483 | ||
3484 | strcpy (buf, p); | |
3485 | ||
3486 | switch (GET_CODE (operands[3])) | |
3487 | { | |
3488 | case MULT: | |
3489 | case PLUS: | |
3490 | if (REG_P (operands[2]) && REGNO (operands[0]) == REGNO (operands[2])) | |
3491 | { | |
3492 | temp = operands[2]; | |
3493 | operands[2] = operands[1]; | |
3494 | operands[1] = temp; | |
3495 | } | |
3496 | ||
3497 | if (GET_CODE (operands[2]) == MEM) | |
3498 | { | |
3499 | p = "%z2\t%2"; | |
3500 | break; | |
3501 | } | |
3502 | ||
3503 | if (find_regno_note (insn, REG_DEAD, REGNO (operands[2]))) | |
3504 | { | |
3505 | if (STACK_TOP_P (operands[0])) | |
3506 | p = "p\t{%0,%2|%2, %0}"; | |
3507 | else | |
3508 | p = "p\t{%2,%0|%0, %2}"; | |
3509 | break; | |
3510 | } | |
3511 | ||
3512 | if (STACK_TOP_P (operands[0])) | |
3513 | p = "\t{%y2,%0|%0, %y2}"; | |
3514 | else | |
3515 | p = "\t{%2,%0|%0, %2}"; | |
3516 | break; | |
3517 | ||
3518 | case MINUS: | |
3519 | case DIV: | |
3520 | if (GET_CODE (operands[1]) == MEM) | |
3521 | { | |
3522 | p = "r%z1\t%1"; | |
3523 | break; | |
3524 | } | |
3525 | ||
3526 | if (GET_CODE (operands[2]) == MEM) | |
3527 | { | |
3528 | p = "%z2\t%2"; | |
3529 | break; | |
3530 | } | |
3531 | ||
3532 | if (! STACK_REG_P (operands[1]) || ! STACK_REG_P (operands[2])) | |
3533 | abort (); | |
3534 | ||
3535 | /* Note that the Unixware assembler, and the AT&T assembler before | |
3536 | that, are confusingly not reversed from Intel syntax in this | |
3537 | area. */ | |
3538 | if (find_regno_note (insn, REG_DEAD, REGNO (operands[2]))) | |
3539 | { | |
3540 | if (STACK_TOP_P (operands[0])) | |
3541 | p = "p\t%0,%2"; | |
3542 | else | |
3543 | p = "rp\t%2,%0"; | |
3544 | break; | |
3545 | } | |
3546 | ||
3547 | if (find_regno_note (insn, REG_DEAD, REGNO (operands[1]))) | |
3548 | { | |
3549 | if (STACK_TOP_P (operands[0])) | |
3550 | p = "rp\t%0,%1"; | |
3551 | else | |
3552 | p = "p\t%1,%0"; | |
3553 | break; | |
3554 | } | |
3555 | ||
3556 | if (STACK_TOP_P (operands[0])) | |
3557 | { | |
3558 | if (STACK_TOP_P (operands[1])) | |
3559 | p = "\t%y2,%0"; | |
3560 | else | |
3561 | p = "r\t%y1,%0"; | |
3562 | break; | |
3563 | } | |
3564 | else if (STACK_TOP_P (operands[1])) | |
3565 | p = "\t%1,%0"; | |
3566 | else | |
3567 | p = "r\t%2,%0"; | |
3568 | break; | |
3569 | ||
3570 | default: | |
3571 | abort (); | |
3572 | } | |
3573 | ||
3574 | strcat (buf, p); | |
3575 | return buf; | |
3576 | } | |
3577 | ||
3578 | /* Output code for INSN to convert a float to a signed int. OPERANDS | |
3579 | are the insn operands. The output may be [SD]Imode and the input | |
3580 | operand may be [SDX]Fmode. */ | |
3581 | ||
3582 | const char * | |
3583 | output_fix_trunc (insn, operands) | |
3584 | rtx insn; | |
3585 | rtx *operands; | |
3586 | { | |
3587 | int stack_top_dies = find_regno_note (insn, REG_DEAD, FIRST_STACK_REG) != 0; | |
3588 | int dimode_p = GET_MODE (operands[0]) == DImode; | |
3589 | rtx xops[4]; | |
3590 | ||
3591 | /* Jump through a hoop or two for DImode, since the hardware has no | |
3592 | non-popping instruction. We used to do this a different way, but | |
3593 | that was somewhat fragile and broke with post-reload splitters. */ | |
3594 | if (dimode_p && !stack_top_dies) | |
3595 | output_asm_insn ("fld\t%y1", operands); | |
3596 | ||
3597 | if (! STACK_TOP_P (operands[1])) | |
3598 | abort (); | |
3599 | ||
3600 | xops[0] = GEN_INT (12); | |
3601 | xops[1] = adj_offsettable_operand (operands[2], 1); | |
3602 | xops[1] = change_address (xops[1], QImode, NULL_RTX); | |
3603 | ||
3604 | xops[2] = operands[0]; | |
3605 | if (GET_CODE (operands[0]) != MEM) | |
3606 | xops[2] = operands[3]; | |
3607 | ||
3608 | output_asm_insn ("fnstcw\t%2", operands); | |
3609 | output_asm_insn ("mov{l}\t{%2, %4|%4, %2}", operands); | |
3610 | output_asm_insn ("mov{b}\t{%0, %1|%1, %0}", xops); | |
3611 | output_asm_insn ("fldcw\t%2", operands); | |
3612 | output_asm_insn ("mov{l}\t{%4, %2|%2, %4}", operands); | |
3613 | ||
3614 | if (stack_top_dies || dimode_p) | |
3615 | output_asm_insn ("fistp%z2\t%2", xops); | |
3616 | else | |
3617 | output_asm_insn ("fist%z2\t%2", xops); | |
3618 | ||
3619 | output_asm_insn ("fldcw\t%2", operands); | |
3620 | ||
3621 | if (GET_CODE (operands[0]) != MEM) | |
3622 | { | |
3623 | if (dimode_p) | |
3624 | { | |
3625 | split_di (operands+0, 1, xops+0, xops+1); | |
3626 | split_di (operands+3, 1, xops+2, xops+3); | |
3627 | output_asm_insn ("mov{l}\t{%2, %0|%0, %2}", xops); | |
3628 | output_asm_insn ("mov{l}\t{%3, %1|%1, %3}", xops); | |
3629 | } | |
3630 | else | |
3631 | output_asm_insn ("mov{l}\t{%3,%0|%0, %3}", operands); | |
3632 | } | |
3633 | ||
3634 | return ""; | |
3635 | } | |
3636 | ||
3637 | /* Output code for INSN to compare OPERANDS. EFLAGS_P is 1 when fcomi | |
3638 | should be used and 2 when fnstsw should be used. UNORDERED_P is true | |
3639 | when fucom should be used. */ | |
3640 | ||
3641 | const char * | |
3642 | output_fp_compare (insn, operands, eflags_p, unordered_p) | |
3643 | rtx insn; | |
3644 | rtx *operands; | |
3645 | int eflags_p, unordered_p; | |
3646 | { | |
3647 | int stack_top_dies; | |
3648 | rtx cmp_op0 = operands[0]; | |
3649 | rtx cmp_op1 = operands[1]; | |
3650 | ||
3651 | if (eflags_p == 2) | |
3652 | { | |
3653 | cmp_op0 = cmp_op1; | |
3654 | cmp_op1 = operands[2]; | |
3655 | } | |
3656 | ||
3657 | if (! STACK_TOP_P (cmp_op0)) | |
3658 | abort (); | |
3659 | ||
3660 | stack_top_dies = find_regno_note (insn, REG_DEAD, FIRST_STACK_REG) != 0; | |
3661 | ||
3662 | if (STACK_REG_P (cmp_op1) | |
3663 | && stack_top_dies | |
3664 | && find_regno_note (insn, REG_DEAD, REGNO (cmp_op1)) | |
3665 | && REGNO (cmp_op1) != FIRST_STACK_REG) | |
3666 | { | |
3667 | /* If both the top of the 387 stack dies, and the other operand | |
3668 | is also a stack register that dies, then this must be a | |
3669 | `fcompp' float compare */ | |
3670 | ||
3671 | if (eflags_p == 1) | |
3672 | { | |
3673 | /* There is no double popping fcomi variant. Fortunately, | |
3674 | eflags is immune from the fstp's cc clobbering. */ | |
3675 | if (unordered_p) | |
3676 | output_asm_insn ("fucomip\t{%y1, %0|%0, %y1}", operands); | |
3677 | else | |
3678 | output_asm_insn ("fcomip\t{%y1, %0|%0, %y1}", operands); | |
3679 | return "fstp\t%y0"; | |
3680 | } | |
3681 | else | |
3682 | { | |
3683 | if (eflags_p == 2) | |
3684 | { | |
3685 | if (unordered_p) | |
3686 | return "fucompp\n\tfnstsw\t%0"; | |
3687 | else | |
3688 | return "fcompp\n\tfnstsw\t%0"; | |
3689 | } | |
3690 | else | |
3691 | { | |
3692 | if (unordered_p) | |
3693 | return "fucompp"; | |
3694 | else | |
3695 | return "fcompp"; | |
3696 | } | |
3697 | } | |
3698 | } | |
3699 | else | |
3700 | { | |
3701 | /* Encoded here as eflags_p | intmode | unordered_p | stack_top_dies. */ | |
3702 | ||
3703 | static const char * const alt[24] = | |
3704 | { | |
3705 | "fcom%z1\t%y1", | |
3706 | "fcomp%z1\t%y1", | |
3707 | "fucom%z1\t%y1", | |
3708 | "fucomp%z1\t%y1", | |
3709 | ||
3710 | "ficom%z1\t%y1", | |
3711 | "ficomp%z1\t%y1", | |
3712 | NULL, | |
3713 | NULL, | |
3714 | ||
3715 | "fcomi\t{%y1, %0|%0, %y1}", | |
3716 | "fcomip\t{%y1, %0|%0, %y1}", | |
3717 | "fucomi\t{%y1, %0|%0, %y1}", | |
3718 | "fucomip\t{%y1, %0|%0, %y1}", | |
3719 | ||
3720 | NULL, | |
3721 | NULL, | |
3722 | NULL, | |
3723 | NULL, | |
3724 | ||
3725 | "fcom%z2\t%y2\n\tfnstsw\t%0", | |
3726 | "fcomp%z2\t%y2\n\tfnstsw\t%0", | |
3727 | "fucom%z2\t%y2\n\tfnstsw\t%0", | |
3728 | "fucomp%z2\t%y2\n\tfnstsw\t%0", | |
3729 | ||
3730 | "ficom%z2\t%y2\n\tfnstsw\t%0", | |
3731 | "ficomp%z2\t%y2\n\tfnstsw\t%0", | |
3732 | NULL, | |
3733 | NULL | |
3734 | }; | |
3735 | ||
3736 | int mask; | |
3737 | const char *ret; | |
3738 | ||
3739 | mask = eflags_p << 3; | |
3740 | mask |= (GET_MODE_CLASS (GET_MODE (operands[1])) == MODE_INT) << 2; | |
3741 | mask |= unordered_p << 1; | |
3742 | mask |= stack_top_dies; | |
3743 | ||
3744 | if (mask >= 24) | |
3745 | abort (); | |
3746 | ret = alt[mask]; | |
3747 | if (ret == NULL) | |
3748 | abort (); | |
3749 | ||
3750 | return ret; | |
3751 | } | |
3752 | } | |
3753 | ||
3754 | /* Output assembler code to FILE to initialize basic-block profiling. | |
3755 | ||
3756 | If profile_block_flag == 2 | |
3757 | ||
3758 | Output code to call the subroutine `__bb_init_trace_func' | |
3759 | and pass two parameters to it. The first parameter is | |
3760 | the address of a block allocated in the object module. | |
3761 | The second parameter is the number of the first basic block | |
3762 | of the function. | |
3763 | ||
3764 | The name of the block is a local symbol made with this statement: | |
3765 | ||
3766 | ASM_GENERATE_INTERNAL_LABEL (BUFFER, "LPBX", 0); | |
3767 | ||
3768 | Of course, since you are writing the definition of | |
3769 | `ASM_GENERATE_INTERNAL_LABEL' as well as that of this macro, you | |
3770 | can take a short cut in the definition of this macro and use the | |
3771 | name that you know will result. | |
3772 | ||
3773 | The number of the first basic block of the function is | |
3774 | passed to the macro in BLOCK_OR_LABEL. | |
3775 | ||
3776 | If described in a virtual assembler language the code to be | |
3777 | output looks like: | |
3778 | ||
3779 | parameter1 <- LPBX0 | |
3780 | parameter2 <- BLOCK_OR_LABEL | |
3781 | call __bb_init_trace_func | |
3782 | ||
3783 | else if profile_block_flag != 0 | |
3784 | ||
3785 | Output code to call the subroutine `__bb_init_func' | |
3786 | and pass one single parameter to it, which is the same | |
3787 | as the first parameter to `__bb_init_trace_func'. | |
3788 | ||
3789 | The first word of this parameter is a flag which will be nonzero if | |
3790 | the object module has already been initialized. So test this word | |
3791 | first, and do not call `__bb_init_func' if the flag is nonzero. | |
3792 | Note: When profile_block_flag == 2 the test need not be done | |
3793 | but `__bb_init_trace_func' *must* be called. | |
3794 | ||
3795 | BLOCK_OR_LABEL may be used to generate a label number as a | |
3796 | branch destination in case `__bb_init_func' will not be called. | |
3797 | ||
3798 | If described in a virtual assembler language the code to be | |
3799 | output looks like: | |
3800 | ||
3801 | cmp (LPBX0),0 | |
3802 | jne local_label | |
3803 | parameter1 <- LPBX0 | |
3804 | call __bb_init_func | |
3805 | local_label: | |
3806 | */ | |
3807 | ||
3808 | void | |
3809 | ix86_output_function_block_profiler (file, block_or_label) | |
3810 | FILE *file; | |
3811 | int block_or_label; | |
3812 | { | |
3813 | static int num_func = 0; | |
3814 | rtx xops[8]; | |
3815 | char block_table[80], false_label[80]; | |
3816 | ||
3817 | ASM_GENERATE_INTERNAL_LABEL (block_table, "LPBX", 0); | |
3818 | ||
3819 | xops[1] = gen_rtx_SYMBOL_REF (VOIDmode, block_table); | |
3820 | xops[5] = stack_pointer_rtx; | |
3821 | xops[7] = gen_rtx_REG (Pmode, 0); /* eax */ | |
3822 | ||
3823 | CONSTANT_POOL_ADDRESS_P (xops[1]) = TRUE; | |
3824 | ||
3825 | switch (profile_block_flag) | |
3826 | { | |
3827 | case 2: | |
3828 | xops[2] = GEN_INT (block_or_label); | |
3829 | xops[3] = gen_rtx_MEM (Pmode, | |
3830 | gen_rtx_SYMBOL_REF (VOIDmode, "__bb_init_trace_func")); | |
3831 | xops[6] = GEN_INT (8); | |
3832 | ||
3833 | output_asm_insn ("push{l}\t%2", xops); | |
3834 | if (!flag_pic) | |
3835 | output_asm_insn ("push{l}\t%1", xops); | |
3836 | else | |
3837 | { | |
3838 | output_asm_insn ("lea{l}\t{%a1, %7|%7, %a1}", xops); | |
3839 | output_asm_insn ("push{l}\t%7", xops); | |
3840 | } | |
3841 | output_asm_insn ("call\t%P3", xops); | |
3842 | output_asm_insn ("add{l}\t{%6, %5|%5, %6}", xops); | |
3843 | break; | |
3844 | ||
3845 | default: | |
3846 | ASM_GENERATE_INTERNAL_LABEL (false_label, "LPBZ", num_func); | |
3847 | ||
3848 | xops[0] = const0_rtx; | |
3849 | xops[2] = gen_rtx_MEM (Pmode, | |
3850 | gen_rtx_SYMBOL_REF (VOIDmode, false_label)); | |
3851 | xops[3] = gen_rtx_MEM (Pmode, | |
3852 | gen_rtx_SYMBOL_REF (VOIDmode, "__bb_init_func")); | |
3853 | xops[4] = gen_rtx_MEM (Pmode, xops[1]); | |
3854 | xops[6] = GEN_INT (4); | |
3855 | ||
3856 | CONSTANT_POOL_ADDRESS_P (xops[2]) = TRUE; | |
3857 | ||
3858 | output_asm_insn ("cmp{l}\t{%0, %4|%4, %0}", xops); | |
3859 | output_asm_insn ("jne\t%2", xops); | |
3860 | ||
3861 | if (!flag_pic) | |
3862 | output_asm_insn ("push{l}\t%1", xops); | |
3863 | else | |
3864 | { | |
3865 | output_asm_insn ("lea{l}\t{%a1, %7|%7, %a2}", xops); | |
3866 | output_asm_insn ("push{l}\t%7", xops); | |
3867 | } | |
3868 | output_asm_insn ("call\t%P3", xops); | |
3869 | output_asm_insn ("add{l}\t{%6, %5|%5, %6}", xops); | |
3870 | ASM_OUTPUT_INTERNAL_LABEL (file, "LPBZ", num_func); | |
3871 | num_func++; | |
3872 | break; | |
3873 | } | |
3874 | } | |
3875 | ||
3876 | /* Output assembler code to FILE to increment a counter associated | |
3877 | with basic block number BLOCKNO. | |
3878 | ||
3879 | If profile_block_flag == 2 | |
3880 | ||
3881 | Output code to initialize the global structure `__bb' and | |
3882 | call the function `__bb_trace_func' which will increment the | |
3883 | counter. | |
3884 | ||
3885 | `__bb' consists of two words. In the first word the number | |
3886 | of the basic block has to be stored. In the second word | |
3887 | the address of a block allocated in the object module | |
3888 | has to be stored. | |
3889 | ||
3890 | The basic block number is given by BLOCKNO. | |
3891 | ||
3892 | The address of the block is given by the label created with | |
3893 | ||
3894 | ASM_GENERATE_INTERNAL_LABEL (BUFFER, "LPBX", 0); | |
3895 | ||
3896 | by FUNCTION_BLOCK_PROFILER. | |
3897 | ||
3898 | Of course, since you are writing the definition of | |
3899 | `ASM_GENERATE_INTERNAL_LABEL' as well as that of this macro, you | |
3900 | can take a short cut in the definition of this macro and use the | |
3901 | name that you know will result. | |
3902 | ||
3903 | If described in a virtual assembler language the code to be | |
3904 | output looks like: | |
3905 | ||
3906 | move BLOCKNO -> (__bb) | |
3907 | move LPBX0 -> (__bb+4) | |
3908 | call __bb_trace_func | |
3909 | ||
3910 | Note that function `__bb_trace_func' must not change the | |
3911 | machine state, especially the flag register. To grant | |
3912 | this, you must output code to save and restore registers | |
3913 | either in this macro or in the macros MACHINE_STATE_SAVE | |
3914 | and MACHINE_STATE_RESTORE. The last two macros will be | |
3915 | used in the function `__bb_trace_func', so you must make | |
3916 | sure that the function prologue does not change any | |
3917 | register prior to saving it with MACHINE_STATE_SAVE. | |
3918 | ||
3919 | else if profile_block_flag != 0 | |
3920 | ||
3921 | Output code to increment the counter directly. | |
3922 | Basic blocks are numbered separately from zero within each | |
3923 | compiled object module. The count associated with block number | |
3924 | BLOCKNO is at index BLOCKNO in an array of words; the name of | |
3925 | this array is a local symbol made with this statement: | |
3926 | ||
3927 | ASM_GENERATE_INTERNAL_LABEL (BUFFER, "LPBX", 2); | |
3928 | ||
3929 | Of course, since you are writing the definition of | |
3930 | `ASM_GENERATE_INTERNAL_LABEL' as well as that of this macro, you | |
3931 | can take a short cut in the definition of this macro and use the | |
3932 | name that you know will result. | |
3933 | ||
3934 | If described in a virtual assembler language the code to be | |
3935 | output looks like: | |
3936 | ||
3937 | inc (LPBX2+4*BLOCKNO) | |
3938 | */ | |
3939 | ||
3940 | void | |
3941 | ix86_output_block_profiler (file, blockno) | |
3942 | FILE *file ATTRIBUTE_UNUSED; | |
3943 | int blockno; | |
3944 | { | |
3945 | rtx xops[8], cnt_rtx; | |
3946 | char counts[80]; | |
3947 | char *block_table = counts; | |
3948 | ||
3949 | switch (profile_block_flag) | |
3950 | { | |
3951 | case 2: | |
3952 | ASM_GENERATE_INTERNAL_LABEL (block_table, "LPBX", 0); | |
3953 | ||
3954 | xops[1] = gen_rtx_SYMBOL_REF (VOIDmode, block_table); | |
3955 | xops[2] = GEN_INT (blockno); | |
3956 | xops[3] = gen_rtx_MEM (Pmode, | |
3957 | gen_rtx_SYMBOL_REF (VOIDmode, "__bb_trace_func")); | |
3958 | xops[4] = gen_rtx_SYMBOL_REF (VOIDmode, "__bb"); | |
3959 | xops[5] = plus_constant (xops[4], 4); | |
3960 | xops[0] = gen_rtx_MEM (SImode, xops[4]); | |
3961 | xops[6] = gen_rtx_MEM (SImode, xops[5]); | |
3962 | ||
3963 | CONSTANT_POOL_ADDRESS_P (xops[1]) = TRUE; | |
3964 | ||
3965 | output_asm_insn ("pushf", xops); | |
3966 | output_asm_insn ("mov{l}\t{%2, %0|%0, %2}", xops); | |
3967 | if (flag_pic) | |
3968 | { | |
3969 | xops[7] = gen_rtx_REG (Pmode, 0); /* eax */ | |
3970 | output_asm_insn ("push{l}\t%7", xops); | |
3971 | output_asm_insn ("lea{l}\t{%a1, %7|%7, %a1}", xops); | |
3972 | output_asm_insn ("mov{l}\t{%7, %6|%6, %7}", xops); | |
3973 | output_asm_insn ("pop{l}\t%7", xops); | |
3974 | } | |
3975 | else | |
3976 | output_asm_insn ("mov{l}\t{%1, %6|%6, %1}", xops); | |
3977 | output_asm_insn ("call\t%P3", xops); | |
3978 | output_asm_insn ("popf", xops); | |
3979 | ||
3980 | break; | |
3981 | ||
3982 | default: | |
3983 | ASM_GENERATE_INTERNAL_LABEL (counts, "LPBX", 2); | |
3984 | cnt_rtx = gen_rtx_SYMBOL_REF (VOIDmode, counts); | |
3985 | SYMBOL_REF_FLAG (cnt_rtx) = TRUE; | |
3986 | ||
3987 | if (blockno) | |
3988 | cnt_rtx = plus_constant (cnt_rtx, blockno*4); | |
3989 | ||
3990 | if (flag_pic) | |
3991 | cnt_rtx = gen_rtx_PLUS (Pmode, pic_offset_table_rtx, cnt_rtx); | |
3992 | ||
3993 | xops[0] = gen_rtx_MEM (SImode, cnt_rtx); | |
3994 | output_asm_insn ("inc{l}\t%0", xops); | |
3995 | ||
3996 | break; | |
3997 | } | |
3998 | } | |
3999 | \f | |
4000 | void | |
4001 | ix86_expand_move (mode, operands) | |
4002 | enum machine_mode mode; | |
4003 | rtx operands[]; | |
4004 | { | |
4005 | int strict = (reload_in_progress || reload_completed); | |
4006 | rtx insn; | |
4007 | ||
4008 | if (flag_pic && mode == Pmode && symbolic_operand (operands[1], Pmode)) | |
4009 | { | |
4010 | /* Emit insns to move operands[1] into operands[0]. */ | |
4011 | ||
4012 | if (GET_CODE (operands[0]) == MEM) | |
4013 | operands[1] = force_reg (Pmode, operands[1]); | |
4014 | else | |
4015 | { | |
4016 | rtx temp = operands[0]; | |
4017 | if (GET_CODE (temp) != REG) | |
4018 | temp = gen_reg_rtx (Pmode); | |
4019 | temp = legitimize_pic_address (operands[1], temp); | |
4020 | if (temp == operands[0]) | |
4021 | return; | |
4022 | operands[1] = temp; | |
4023 | } | |
4024 | } | |
4025 | else | |
4026 | { | |
4027 | if (GET_CODE (operands[0]) == MEM | |
4028 | && (GET_MODE (operands[0]) == QImode | |
4029 | || !push_operand (operands[0], mode)) | |
4030 | && GET_CODE (operands[1]) == MEM) | |
4031 | operands[1] = force_reg (mode, operands[1]); | |
4032 | ||
4033 | if (push_operand (operands[0], mode) | |
4034 | && ! general_no_elim_operand (operands[1], mode)) | |
4035 | operands[1] = copy_to_mode_reg (mode, operands[1]); | |
4036 | ||
4037 | if (FLOAT_MODE_P (mode)) | |
4038 | { | |
4039 | /* If we are loading a floating point constant to a register, | |
4040 | force the value to memory now, since we'll get better code | |
4041 | out the back end. */ | |
4042 | ||
4043 | if (strict) | |
4044 | ; | |
4045 | else if (GET_CODE (operands[1]) == CONST_DOUBLE | |
4046 | && register_operand (operands[0], mode)) | |
4047 | operands[1] = validize_mem (force_const_mem (mode, operands[1])); | |
4048 | } | |
4049 | } | |
4050 | ||
4051 | insn = gen_rtx_SET (VOIDmode, operands[0], operands[1]); | |
4052 | ||
4053 | emit_insn (insn); | |
4054 | } | |
4055 | ||
4056 | /* Attempt to expand a binary operator. Make the expansion closer to the | |
4057 | actual machine, then just general_operand, which will allow 3 separate | |
4058 | memory references (one output, two input) in a single insn. */ | |
4059 | ||
4060 | void | |
4061 | ix86_expand_binary_operator (code, mode, operands) | |
4062 | enum rtx_code code; | |
4063 | enum machine_mode mode; | |
4064 | rtx operands[]; | |
4065 | { | |
4066 | int matching_memory; | |
4067 | rtx src1, src2, dst, op, clob; | |
4068 | ||
4069 | dst = operands[0]; | |
4070 | src1 = operands[1]; | |
4071 | src2 = operands[2]; | |
4072 | ||
4073 | /* Recognize <var1> = <value> <op> <var1> for commutative operators */ | |
4074 | if (GET_RTX_CLASS (code) == 'c' | |
4075 | && (rtx_equal_p (dst, src2) | |
4076 | || immediate_operand (src1, mode))) | |
4077 | { | |
4078 | rtx temp = src1; | |
4079 | src1 = src2; | |
4080 | src2 = temp; | |
4081 | } | |
4082 | ||
4083 | /* If the destination is memory, and we do not have matching source | |
4084 | operands, do things in registers. */ | |
4085 | matching_memory = 0; | |
4086 | if (GET_CODE (dst) == MEM) | |
4087 | { | |
4088 | if (rtx_equal_p (dst, src1)) | |
4089 | matching_memory = 1; | |
4090 | else if (GET_RTX_CLASS (code) == 'c' | |
4091 | && rtx_equal_p (dst, src2)) | |
4092 | matching_memory = 2; | |
4093 | else | |
4094 | dst = gen_reg_rtx (mode); | |
4095 | } | |
4096 | ||
4097 | /* Both source operands cannot be in memory. */ | |
4098 | if (GET_CODE (src1) == MEM && GET_CODE (src2) == MEM) | |
4099 | { | |
4100 | if (matching_memory != 2) | |
4101 | src2 = force_reg (mode, src2); | |
4102 | else | |
4103 | src1 = force_reg (mode, src1); | |
4104 | } | |
4105 | ||
4106 | /* If the operation is not commutable, source 1 cannot be a constant | |
4107 | or non-matching memory. */ | |
4108 | if ((CONSTANT_P (src1) | |
4109 | || (!matching_memory && GET_CODE (src1) == MEM)) | |
4110 | && GET_RTX_CLASS (code) != 'c') | |
4111 | src1 = force_reg (mode, src1); | |
4112 | ||
4113 | /* If optimizing, copy to regs to improve CSE */ | |
4114 | if (optimize && !reload_in_progress && !reload_completed) | |
4115 | { | |
4116 | if (GET_CODE (dst) == MEM) | |
4117 | dst = gen_reg_rtx (mode); | |
4118 | if (GET_CODE (src1) == MEM) | |
4119 | src1 = force_reg (mode, src1); | |
4120 | if (GET_CODE (src2) == MEM) | |
4121 | src2 = force_reg (mode, src2); | |
4122 | } | |
4123 | ||
4124 | /* Emit the instruction. */ | |
4125 | ||
4126 | op = gen_rtx_SET (VOIDmode, dst, gen_rtx_fmt_ee (code, mode, src1, src2)); | |
4127 | if (reload_in_progress) | |
4128 | { | |
4129 | /* Reload doesn't know about the flags register, and doesn't know that | |
4130 | it doesn't want to clobber it. We can only do this with PLUS. */ | |
4131 | if (code != PLUS) | |
4132 | abort (); | |
4133 | emit_insn (op); | |
4134 | } | |
4135 | else | |
4136 | { | |
4137 | clob = gen_rtx_CLOBBER (VOIDmode, gen_rtx_REG (CCmode, FLAGS_REG)); | |
4138 | emit_insn (gen_rtx_PARALLEL (VOIDmode, gen_rtvec (2, op, clob))); | |
4139 | } | |
4140 | ||
4141 | /* Fix up the destination if needed. */ | |
4142 | if (dst != operands[0]) | |
4143 | emit_move_insn (operands[0], dst); | |
4144 | } | |
4145 | ||
4146 | /* Return TRUE or FALSE depending on whether the binary operator meets the | |
4147 | appropriate constraints. */ | |
4148 | ||
4149 | int | |
4150 | ix86_binary_operator_ok (code, mode, operands) | |
4151 | enum rtx_code code; | |
4152 | enum machine_mode mode ATTRIBUTE_UNUSED; | |
4153 | rtx operands[3]; | |
4154 | { | |
4155 | /* Both source operands cannot be in memory. */ | |
4156 | if (GET_CODE (operands[1]) == MEM && GET_CODE (operands[2]) == MEM) | |
4157 | return 0; | |
4158 | /* If the operation is not commutable, source 1 cannot be a constant. */ | |
4159 | if (CONSTANT_P (operands[1]) && GET_RTX_CLASS (code) != 'c') | |
4160 | return 0; | |
4161 | /* If the destination is memory, we must have a matching source operand. */ | |
4162 | if (GET_CODE (operands[0]) == MEM | |
4163 | && ! (rtx_equal_p (operands[0], operands[1]) | |
4164 | || (GET_RTX_CLASS (code) == 'c' | |
4165 | && rtx_equal_p (operands[0], operands[2])))) | |
4166 | return 0; | |
4167 | /* If the operation is not commutable and the source 1 is memory, we must | |
4168 | have a matching destionation. */ | |
4169 | if (GET_CODE (operands[1]) == MEM | |
4170 | && GET_RTX_CLASS (code) != 'c' | |
4171 | && ! rtx_equal_p (operands[0], operands[1])) | |
4172 | return 0; | |
4173 | return 1; | |
4174 | } | |
4175 | ||
4176 | /* Attempt to expand a unary operator. Make the expansion closer to the | |
4177 | actual machine, then just general_operand, which will allow 2 separate | |
4178 | memory references (one output, one input) in a single insn. */ | |
4179 | ||
4180 | void | |
4181 | ix86_expand_unary_operator (code, mode, operands) | |
4182 | enum rtx_code code; | |
4183 | enum machine_mode mode; | |
4184 | rtx operands[]; | |
4185 | { | |
4186 | int matching_memory; | |
4187 | rtx src, dst, op, clob; | |
4188 | ||
4189 | dst = operands[0]; | |
4190 | src = operands[1]; | |
4191 | ||
4192 | /* If the destination is memory, and we do not have matching source | |
4193 | operands, do things in registers. */ | |
4194 | matching_memory = 0; | |
4195 | if (GET_CODE (dst) == MEM) | |
4196 | { | |
4197 | if (rtx_equal_p (dst, src)) | |
4198 | matching_memory = 1; | |
4199 | else | |
4200 | dst = gen_reg_rtx (mode); | |
4201 | } | |
4202 | ||
4203 | /* When source operand is memory, destination must match. */ | |
4204 | if (!matching_memory && GET_CODE (src) == MEM) | |
4205 | src = force_reg (mode, src); | |
4206 | ||
4207 | /* If optimizing, copy to regs to improve CSE */ | |
4208 | if (optimize && !reload_in_progress && !reload_completed) | |
4209 | { | |
4210 | if (GET_CODE (dst) == MEM) | |
4211 | dst = gen_reg_rtx (mode); | |
4212 | if (GET_CODE (src) == MEM) | |
4213 | src = force_reg (mode, src); | |
4214 | } | |
4215 | ||
4216 | /* Emit the instruction. */ | |
4217 | ||
4218 | op = gen_rtx_SET (VOIDmode, dst, gen_rtx_fmt_e (code, mode, src)); | |
4219 | if (reload_in_progress || code == NOT) | |
4220 | { | |
4221 | /* Reload doesn't know about the flags register, and doesn't know that | |
4222 | it doesn't want to clobber it. */ | |
4223 | if (code != NOT) | |
4224 | abort (); | |
4225 | emit_insn (op); | |
4226 | } | |
4227 | else | |
4228 | { | |
4229 | clob = gen_rtx_CLOBBER (VOIDmode, gen_rtx_REG (CCmode, FLAGS_REG)); | |
4230 | emit_insn (gen_rtx_PARALLEL (VOIDmode, gen_rtvec (2, op, clob))); | |
4231 | } | |
4232 | ||
4233 | /* Fix up the destination if needed. */ | |
4234 | if (dst != operands[0]) | |
4235 | emit_move_insn (operands[0], dst); | |
4236 | } | |
4237 | ||
4238 | /* Return TRUE or FALSE depending on whether the unary operator meets the | |
4239 | appropriate constraints. */ | |
4240 | ||
4241 | int | |
4242 | ix86_unary_operator_ok (code, mode, operands) | |
4243 | enum rtx_code code ATTRIBUTE_UNUSED; | |
4244 | enum machine_mode mode ATTRIBUTE_UNUSED; | |
4245 | rtx operands[2] ATTRIBUTE_UNUSED; | |
4246 | { | |
4247 | /* If one of operands is memory, source and destination must match. */ | |
4248 | if ((GET_CODE (operands[0]) == MEM | |
4249 | || GET_CODE (operands[1]) == MEM) | |
4250 | && ! rtx_equal_p (operands[0], operands[1])) | |
4251 | return FALSE; | |
4252 | return TRUE; | |
4253 | } | |
4254 | ||
4255 | /* Produce an unsigned comparison for a given signed comparison. */ | |
4256 | ||
4257 | static enum rtx_code | |
4258 | unsigned_comparison (code) | |
4259 | enum rtx_code code; | |
4260 | { | |
4261 | switch (code) | |
4262 | { | |
4263 | case GT: | |
4264 | code = GTU; | |
4265 | break; | |
4266 | case LT: | |
4267 | code = LTU; | |
4268 | break; | |
4269 | case GE: | |
4270 | code = GEU; | |
4271 | break; | |
4272 | case LE: | |
4273 | code = LEU; | |
4274 | break; | |
4275 | case EQ: | |
4276 | case NE: | |
4277 | case LEU: | |
4278 | case LTU: | |
4279 | case GEU: | |
4280 | case GTU: | |
4281 | break; | |
4282 | default: | |
4283 | abort (); | |
4284 | } | |
4285 | return code; | |
4286 | } | |
4287 | ||
4288 | /* Generate insn patterns to do an integer compare of OPERANDS. */ | |
4289 | ||
4290 | static rtx | |
4291 | ix86_expand_int_compare (code, op0, op1) | |
4292 | enum rtx_code code; | |
4293 | rtx op0, op1; | |
4294 | { | |
4295 | enum machine_mode cmpmode; | |
4296 | rtx tmp, flags; | |
4297 | ||
4298 | cmpmode = SELECT_CC_MODE (code, op0, op1); | |
4299 | flags = gen_rtx_REG (cmpmode, FLAGS_REG); | |
4300 | ||
4301 | /* This is very simple, but making the interface the same as in the | |
4302 | FP case makes the rest of the code easier. */ | |
4303 | tmp = gen_rtx_COMPARE (cmpmode, op0, op1); | |
4304 | emit_insn (gen_rtx_SET (VOIDmode, flags, tmp)); | |
4305 | ||
4306 | /* Return the test that should be put into the flags user, i.e. | |
4307 | the bcc, scc, or cmov instruction. */ | |
4308 | return gen_rtx_fmt_ee (code, VOIDmode, flags, const0_rtx); | |
4309 | } | |
4310 | ||
4311 | /* Generate insn patterns to do a floating point compare of OPERANDS. | |
4312 | If UNORDERED, allow for unordered compares. */ | |
4313 | ||
4314 | static rtx | |
4315 | ix86_expand_fp_compare (code, op0, op1, unordered) | |
4316 | enum rtx_code code; | |
4317 | rtx op0, op1; | |
4318 | int unordered; | |
4319 | { | |
4320 | enum machine_mode fpcmp_mode; | |
4321 | enum machine_mode intcmp_mode; | |
4322 | rtx tmp; | |
4323 | ||
4324 | /* When not doing IEEE compliant compares, disable unordered. */ | |
4325 | if (! TARGET_IEEE_FP) | |
4326 | unordered = 0; | |
4327 | fpcmp_mode = unordered ? CCFPUmode : CCFPmode; | |
4328 | ||
4329 | /* ??? If we knew whether invalid-operand exceptions were masked, | |
4330 | we could rely on fcom to raise an exception and take care of | |
4331 | NaNs. But we don't. We could know this from c9x math bits. */ | |
4332 | if (TARGET_IEEE_FP) | |
4333 | unordered = 1; | |
4334 | ||
4335 | /* All of the unordered compare instructions only work on registers. | |
4336 | The same is true of the XFmode compare instructions. */ | |
4337 | if (unordered || GET_MODE (op0) == XFmode) | |
4338 | { | |
4339 | op0 = force_reg (GET_MODE (op0), op0); | |
4340 | op1 = force_reg (GET_MODE (op1), op1); | |
4341 | } | |
4342 | else | |
4343 | { | |
4344 | /* %%% We only allow op1 in memory; op0 must be st(0). So swap | |
4345 | things around if they appear profitable, otherwise force op0 | |
4346 | into a register. */ | |
4347 | ||
4348 | if (standard_80387_constant_p (op0) == 0 | |
4349 | || (GET_CODE (op0) == MEM | |
4350 | && ! (standard_80387_constant_p (op1) == 0 | |
4351 | || GET_CODE (op1) == MEM))) | |
4352 | { | |
4353 | rtx tmp; | |
4354 | tmp = op0, op0 = op1, op1 = tmp; | |
4355 | code = swap_condition (code); | |
4356 | } | |
4357 | ||
4358 | if (GET_CODE (op0) != REG) | |
4359 | op0 = force_reg (GET_MODE (op0), op0); | |
4360 | ||
4361 | if (CONSTANT_P (op1)) | |
4362 | { | |
4363 | if (standard_80387_constant_p (op1)) | |
4364 | op1 = force_reg (GET_MODE (op1), op1); | |
4365 | else | |
4366 | op1 = validize_mem (force_const_mem (GET_MODE (op1), op1)); | |
4367 | } | |
4368 | } | |
4369 | ||
4370 | /* %%% fcomi is probably always faster, even when dealing with memory, | |
4371 | since compare-and-branch would be three insns instead of four. */ | |
4372 | if (TARGET_CMOVE && !unordered) | |
4373 | { | |
4374 | if (GET_CODE (op0) != REG) | |
4375 | op0 = force_reg (GET_MODE (op0), op0); | |
4376 | if (GET_CODE (op1) != REG) | |
4377 | op1 = force_reg (GET_MODE (op1), op1); | |
4378 | ||
4379 | tmp = gen_rtx_COMPARE (fpcmp_mode, op0, op1); | |
4380 | tmp = gen_rtx_SET (VOIDmode, gen_rtx_REG (fpcmp_mode, FLAGS_REG), tmp); | |
4381 | emit_insn (tmp); | |
4382 | ||
4383 | /* The FP codes work out to act like unsigned. */ | |
4384 | code = unsigned_comparison (code); | |
4385 | intcmp_mode = fpcmp_mode; | |
4386 | } | |
4387 | else | |
4388 | { | |
4389 | /* Sadness wrt reg-stack pops killing fpsr -- gotta get fnstsw first. */ | |
4390 | ||
4391 | rtx tmp2; | |
4392 | tmp = gen_rtx_COMPARE (fpcmp_mode, op0, op1); | |
4393 | tmp2 = gen_rtx_UNSPEC (HImode, gen_rtvec (1, tmp), 9); | |
4394 | tmp = gen_reg_rtx (HImode); | |
4395 | emit_insn (gen_rtx_SET (VOIDmode, tmp, tmp2)); | |
4396 | ||
4397 | if (! unordered) | |
4398 | { | |
4399 | /* We have two options here -- use sahf, or testing bits of ah | |
4400 | directly. On PPRO, they are equivalent, sahf being one byte | |
4401 | smaller. On Pentium, sahf is non-pairable while test is UV | |
4402 | pairable. */ | |
4403 | ||
4404 | if (TARGET_USE_SAHF || optimize_size) | |
4405 | { | |
4406 | do_sahf: | |
4407 | ||
4408 | /* The FP codes work out to act like unsigned. */ | |
4409 | code = unsigned_comparison (code); | |
4410 | emit_insn (gen_x86_sahf_1 (tmp)); | |
4411 | intcmp_mode = CCmode; | |
4412 | } | |
4413 | else | |
4414 | { | |
4415 | /* | |
4416 | * The numbers below correspond to the bits of the FPSW in AH. | |
4417 | * C3, C2, and C0 are in bits 0x40, 0x4, and 0x01 respectively. | |
4418 | * | |
4419 | * cmp C3 C2 C0 | |
4420 | * > 0 0 0 | |
4421 | * < 0 0 1 | |
4422 | * = 1 0 0 | |
4423 | * un 1 1 1 | |
4424 | */ | |
4425 | ||
4426 | int mask; | |
4427 | ||
4428 | switch (code) | |
4429 | { | |
4430 | case GT: | |
4431 | mask = 0x41; | |
4432 | code = EQ; | |
4433 | break; | |
4434 | case LT: | |
4435 | mask = 0x01; | |
4436 | code = NE; | |
4437 | break; | |
4438 | case GE: | |
4439 | /* We'd have to use `xorb 1,ah; andb 0x41,ah', so it's | |
4440 | faster in all cases to just fall back on sahf. */ | |
4441 | goto do_sahf; | |
4442 | case LE: | |
4443 | mask = 0x41; | |
4444 | code = NE; | |
4445 | break; | |
4446 | case EQ: | |
4447 | mask = 0x40; | |
4448 | code = NE; | |
4449 | break; | |
4450 | case NE: | |
4451 | mask = 0x40; | |
4452 | code = EQ; | |
4453 | break; | |
4454 | default: | |
4455 | abort (); | |
4456 | } | |
4457 | ||
4458 | emit_insn (gen_testqi_ext_0 (tmp, GEN_INT (mask))); | |
4459 | intcmp_mode = CCNOmode; | |
4460 | } | |
4461 | } | |
4462 | else | |
4463 | { | |
4464 | /* In the unordered case, we have to check C2 for NaN's, which | |
4465 | doesn't happen to work out to anything nice combination-wise. | |
4466 | So do some bit twiddling on the value we've got in AH to come | |
4467 | up with an appropriate set of condition codes. */ | |
4468 | ||
4469 | intcmp_mode = CCNOmode; | |
4470 | switch (code) | |
4471 | { | |
4472 | case GT: | |
4473 | emit_insn (gen_testqi_ext_0 (tmp, GEN_INT (0x45))); | |
4474 | code = EQ; | |
4475 | break; | |
4476 | case LT: | |
4477 | emit_insn (gen_andqi_ext_0 (tmp, tmp, GEN_INT (0x45))); | |
4478 | emit_insn (gen_cmpqi_ext_3 (tmp, GEN_INT (0x01))); | |
4479 | intcmp_mode = CCmode; | |
4480 | code = EQ; | |
4481 | break; | |
4482 | case GE: | |
4483 | emit_insn (gen_testqi_ext_0 (tmp, GEN_INT (0x05))); | |
4484 | code = EQ; | |
4485 | break; | |
4486 | case LE: | |
4487 | emit_insn (gen_andqi_ext_0 (tmp, tmp, GEN_INT (0x45))); | |
4488 | emit_insn (gen_addqi_ext_1 (tmp, tmp, constm1_rtx)); | |
4489 | emit_insn (gen_cmpqi_ext_3 (tmp, GEN_INT (0x40))); | |
4490 | intcmp_mode = CCmode; | |
4491 | code = LTU; | |
4492 | break; | |
4493 | case EQ: | |
4494 | emit_insn (gen_andqi_ext_0 (tmp, tmp, GEN_INT (0x45))); | |
4495 | emit_insn (gen_cmpqi_ext_3 (tmp, GEN_INT (0x40))); | |
4496 | intcmp_mode = CCmode; | |
4497 | code = EQ; | |
4498 | break; | |
4499 | case NE: | |
4500 | emit_insn (gen_andqi_ext_0 (tmp, tmp, GEN_INT (0x45))); | |
4501 | emit_insn (gen_xorqi_cc_ext_1 (tmp, tmp, GEN_INT (0x40))); | |
4502 | code = NE; | |
4503 | break; | |
4504 | default: | |
4505 | abort (); | |
4506 | } | |
4507 | } | |
4508 | } | |
4509 | ||
4510 | /* Return the test that should be put into the flags user, i.e. | |
4511 | the bcc, scc, or cmov instruction. */ | |
4512 | return gen_rtx_fmt_ee (code, VOIDmode, | |
4513 | gen_rtx_REG (intcmp_mode, FLAGS_REG), | |
4514 | const0_rtx); | |
4515 | } | |
4516 | ||
4517 | static rtx | |
4518 | ix86_expand_compare (code, unordered) | |
4519 | enum rtx_code code; | |
4520 | int unordered; | |
4521 | { | |
4522 | rtx op0, op1, ret; | |
4523 | op0 = ix86_compare_op0; | |
4524 | op1 = ix86_compare_op1; | |
4525 | ||
4526 | if (GET_MODE_CLASS (GET_MODE (op0)) == MODE_FLOAT) | |
4527 | ret = ix86_expand_fp_compare (code, op0, op1, unordered); | |
4528 | else | |
4529 | ret = ix86_expand_int_compare (code, op0, op1); | |
4530 | ||
4531 | return ret; | |
4532 | } | |
4533 | ||
4534 | void | |
4535 | ix86_expand_branch (code, unordered, label) | |
4536 | enum rtx_code code; | |
4537 | int unordered; | |
4538 | rtx label; | |
4539 | { | |
4540 | rtx tmp, lo[2], hi[2], label2; | |
4541 | enum rtx_code code1, code2, code3; | |
4542 | ||
4543 | if (GET_MODE (ix86_compare_op0) != DImode) | |
4544 | { | |
4545 | tmp = ix86_expand_compare (code, unordered); | |
4546 | tmp = gen_rtx_IF_THEN_ELSE (VOIDmode, tmp, | |
4547 | gen_rtx_LABEL_REF (VOIDmode, label), | |
4548 | pc_rtx); | |
4549 | emit_jump_insn (gen_rtx_SET (VOIDmode, pc_rtx, tmp)); | |
4550 | return; | |
4551 | } | |
4552 | ||
4553 | /* Expand DImode branch into multiple compare+branch. */ | |
4554 | ||
4555 | if (CONSTANT_P (ix86_compare_op0) && ! CONSTANT_P (ix86_compare_op1)) | |
4556 | { | |
4557 | tmp = ix86_compare_op0; | |
4558 | ix86_compare_op0 = ix86_compare_op1; | |
4559 | ix86_compare_op1 = tmp; | |
4560 | code = swap_condition (code); | |
4561 | } | |
4562 | split_di (&ix86_compare_op0, 1, lo+0, hi+0); | |
4563 | split_di (&ix86_compare_op1, 1, lo+1, hi+1); | |
4564 | ||
4565 | /* When comparing for equality, we can use (hi0^hi1)|(lo0^lo1) to avoid | |
4566 | two branches. This costs one extra insn, so disable when optimizing | |
4567 | for size. */ | |
4568 | ||
4569 | if ((code == EQ || code == NE) | |
4570 | && (!optimize_size | |
4571 | || hi[1] == const0_rtx || lo[1] == const0_rtx)) | |
4572 | { | |
4573 | rtx xor0, xor1; | |
4574 | ||
4575 | xor1 = hi[0]; | |
4576 | if (hi[1] != const0_rtx) | |
4577 | { | |
4578 | xor1 = expand_binop (SImode, xor_optab, xor1, hi[1], | |
4579 | NULL_RTX, 0, OPTAB_WIDEN); | |
4580 | } | |
4581 | ||
4582 | xor0 = lo[0]; | |
4583 | if (lo[1] != const0_rtx) | |
4584 | { | |
4585 | xor0 = expand_binop (SImode, xor_optab, xor0, lo[1], | |
4586 | NULL_RTX, 0, OPTAB_WIDEN); | |
4587 | } | |
4588 | ||
4589 | tmp = expand_binop (SImode, ior_optab, xor1, xor0, | |
4590 | NULL_RTX, 0, OPTAB_WIDEN); | |
4591 | ||
4592 | ix86_compare_op0 = tmp; | |
4593 | ix86_compare_op1 = const0_rtx; | |
4594 | ix86_expand_branch (code, unordered, label); | |
4595 | return; | |
4596 | } | |
4597 | ||
4598 | /* Otherwise, if we are doing less-than, op1 is a constant and the | |
4599 | low word is zero, then we can just examine the high word. */ | |
4600 | ||
4601 | if (GET_CODE (hi[1]) == CONST_INT && lo[1] == const0_rtx | |
4602 | && (code == LT || code == LTU)) | |
4603 | { | |
4604 | ix86_compare_op0 = hi[0]; | |
4605 | ix86_compare_op1 = hi[1]; | |
4606 | ix86_expand_branch (code, unordered, label); | |
4607 | return; | |
4608 | } | |
4609 | ||
4610 | /* Otherwise, we need two or three jumps. */ | |
4611 | ||
4612 | label2 = gen_label_rtx (); | |
4613 | ||
4614 | code1 = code; | |
4615 | code2 = swap_condition (code); | |
4616 | code3 = unsigned_condition (code); | |
4617 | ||
4618 | switch (code) | |
4619 | { | |
4620 | case LT: case GT: case LTU: case GTU: | |
4621 | break; | |
4622 | ||
4623 | case LE: code1 = LT; code2 = GT; break; | |
4624 | case GE: code1 = GT; code2 = LT; break; | |
4625 | case LEU: code1 = LTU; code2 = GTU; break; | |
4626 | case GEU: code1 = GTU; code2 = LTU; break; | |
4627 | ||
4628 | case EQ: code1 = NIL; code2 = NE; break; | |
4629 | case NE: code2 = NIL; break; | |
4630 | ||
4631 | default: | |
4632 | abort (); | |
4633 | } | |
4634 | ||
4635 | /* | |
4636 | * a < b => | |
4637 | * if (hi(a) < hi(b)) goto true; | |
4638 | * if (hi(a) > hi(b)) goto false; | |
4639 | * if (lo(a) < lo(b)) goto true; | |
4640 | * false: | |
4641 | */ | |
4642 | ||
4643 | ix86_compare_op0 = hi[0]; | |
4644 | ix86_compare_op1 = hi[1]; | |
4645 | ||
4646 | if (code1 != NIL) | |
4647 | ix86_expand_branch (code1, unordered, label); | |
4648 | if (code2 != NIL) | |
4649 | ix86_expand_branch (code2, unordered, label2); | |
4650 | ||
4651 | ix86_compare_op0 = lo[0]; | |
4652 | ix86_compare_op1 = lo[1]; | |
4653 | ix86_expand_branch (code3, unordered, label); | |
4654 | ||
4655 | if (code2 != NIL) | |
4656 | emit_label (label2); | |
4657 | } | |
4658 | ||
4659 | int | |
4660 | ix86_expand_setcc (code, unordered, dest) | |
4661 | enum rtx_code code; | |
4662 | int unordered; | |
4663 | rtx dest; | |
4664 | { | |
4665 | rtx ret, tmp; | |
4666 | int type; | |
4667 | ||
4668 | if (GET_MODE (ix86_compare_op0) == DImode) | |
4669 | return 0; /* FAIL */ | |
4670 | ||
4671 | /* Three modes of generation: | |
4672 | 0 -- destination does not overlap compare sources: | |
4673 | clear dest first, emit strict_low_part setcc. | |
4674 | 1 -- destination does overlap compare sources: | |
4675 | emit subreg setcc, zero extend. | |
4676 | 2 -- destination is in QImode: | |
4677 | emit setcc only. | |
4678 | */ | |
4679 | ||
4680 | type = 0; | |
4681 | ||
4682 | if (GET_MODE (dest) == QImode) | |
4683 | type = 2; | |
4684 | else if (reg_overlap_mentioned_p (dest, ix86_compare_op0) | |
4685 | || reg_overlap_mentioned_p (dest, ix86_compare_op1)) | |
4686 | type = 1; | |
4687 | ||
4688 | if (type == 0) | |
4689 | emit_move_insn (dest, const0_rtx); | |
4690 | ||
4691 | ret = ix86_expand_compare (code, unordered); | |
4692 | PUT_MODE (ret, QImode); | |
4693 | ||
4694 | tmp = dest; | |
4695 | if (type == 0) | |
4696 | { | |
4697 | tmp = gen_lowpart (QImode, dest); | |
4698 | tmp = gen_rtx_STRICT_LOW_PART (VOIDmode, tmp); | |
4699 | } | |
4700 | else if (type == 1) | |
4701 | { | |
4702 | if (!cse_not_expected) | |
4703 | tmp = gen_reg_rtx (QImode); | |
4704 | else | |
4705 | tmp = gen_lowpart (QImode, dest); | |
4706 | } | |
4707 | ||
4708 | emit_insn (gen_rtx_SET (VOIDmode, tmp, ret)); | |
4709 | ||
4710 | if (type == 1) | |
4711 | { | |
4712 | rtx clob; | |
4713 | ||
4714 | tmp = gen_rtx_ZERO_EXTEND (GET_MODE (dest), tmp); | |
4715 | tmp = gen_rtx_SET (VOIDmode, dest, tmp); | |
4716 | clob = gen_rtx_CLOBBER (VOIDmode, gen_rtx_REG (CCmode, FLAGS_REG)); | |
4717 | tmp = gen_rtx_PARALLEL (VOIDmode, gen_rtvec (2, tmp, clob)); | |
4718 | emit_insn (tmp); | |
4719 | } | |
4720 | ||
4721 | return 1; /* DONE */ | |
4722 | } | |
4723 | ||
4724 | int | |
4725 | ix86_expand_int_movcc (operands) | |
4726 | rtx operands[]; | |
4727 | { | |
4728 | enum rtx_code code = GET_CODE (operands[1]), compare_code; | |
4729 | rtx compare_seq, compare_op; | |
4730 | ||
4731 | /* When the compare code is not LTU or GEU, we can not use sbbl case. | |
4732 | In case comparsion is done with immediate, we can convert it to LTU or | |
4733 | GEU by altering the integer. */ | |
4734 | ||
4735 | if ((code == LEU || code == GTU) | |
4736 | && GET_CODE (ix86_compare_op1) == CONST_INT | |
4737 | && GET_MODE (operands[0]) != HImode | |
4738 | && (unsigned int)INTVAL (ix86_compare_op1) != 0xffffffff | |
4739 | && GET_CODE (operands[2]) == CONST_INT | |
4740 | && GET_CODE (operands[3]) == CONST_INT) | |
4741 | { | |
4742 | if (code == LEU) | |
4743 | code = LTU; | |
4744 | else | |
4745 | code = GEU; | |
4746 | ix86_compare_op1 = GEN_INT (INTVAL (ix86_compare_op1) + 1); | |
4747 | } | |
4748 | start_sequence (); | |
4749 | compare_op = ix86_expand_compare (code, code == EQ || code == NE); | |
4750 | compare_seq = gen_sequence (); | |
4751 | end_sequence (); | |
4752 | ||
4753 | compare_code = GET_CODE (compare_op); | |
4754 | ||
4755 | /* Don't attempt mode expansion here -- if we had to expand 5 or 6 | |
4756 | HImode insns, we'd be swallowed in word prefix ops. */ | |
4757 | ||
4758 | if (GET_MODE (operands[0]) != HImode | |
4759 | && GET_CODE (operands[2]) == CONST_INT | |
4760 | && GET_CODE (operands[3]) == CONST_INT) | |
4761 | { | |
4762 | rtx out = operands[0]; | |
4763 | HOST_WIDE_INT ct = INTVAL (operands[2]); | |
4764 | HOST_WIDE_INT cf = INTVAL (operands[3]); | |
4765 | HOST_WIDE_INT diff; | |
4766 | ||
4767 | if (compare_code == LTU || compare_code == GEU) | |
4768 | { | |
4769 | ||
4770 | /* Detect overlap between destination and compare sources. */ | |
4771 | rtx tmp = out; | |
4772 | ||
4773 | /* To simplify rest of code, restrict to the GEU case. */ | |
4774 | if (compare_code == LTU) | |
4775 | { | |
4776 | int tmp = ct; | |
4777 | ct = cf; | |
4778 | cf = tmp; | |
4779 | compare_code = reverse_condition (compare_code); | |
4780 | code = reverse_condition (code); | |
4781 | } | |
4782 | diff = ct - cf; | |
4783 | ||
4784 | if (reg_overlap_mentioned_p (out, ix86_compare_op0) | |
4785 | || reg_overlap_mentioned_p (out, ix86_compare_op1)) | |
4786 | tmp = gen_reg_rtx (SImode); | |
4787 | ||
4788 | emit_insn (compare_seq); | |
4789 | emit_insn (gen_x86_movsicc_0_m1 (tmp)); | |
4790 | ||
4791 | if (diff == 1) | |
4792 | { | |
4793 | /* | |
4794 | * cmpl op0,op1 | |
4795 | * sbbl dest,dest | |
4796 | * [addl dest, ct] | |
4797 | * | |
4798 | * Size 5 - 8. | |
4799 | */ | |
4800 | if (ct) | |
4801 | emit_insn (gen_addsi3 (out, out, GEN_INT (ct))); | |
4802 | } | |
4803 | else if (cf == -1) | |
4804 | { | |
4805 | /* | |
4806 | * cmpl op0,op1 | |
4807 | * sbbl dest,dest | |
4808 | * orl $ct, dest | |
4809 | * | |
4810 | * Size 8. | |
4811 | */ | |
4812 | emit_insn (gen_iorsi3 (out, out, GEN_INT (ct))); | |
4813 | } | |
4814 | else if (diff == -1 && ct) | |
4815 | { | |
4816 | /* | |
4817 | * cmpl op0,op1 | |
4818 | * sbbl dest,dest | |
4819 | * xorl $-1, dest | |
4820 | * [addl dest, cf] | |
4821 | * | |
4822 | * Size 8 - 11. | |
4823 | */ | |
4824 | emit_insn (gen_one_cmplsi2 (tmp, tmp)); | |
4825 | if (cf) | |
4826 | emit_insn (gen_addsi3 (out, out, GEN_INT (cf))); | |
4827 | } | |
4828 | else | |
4829 | { | |
4830 | /* | |
4831 | * cmpl op0,op1 | |
4832 | * sbbl dest,dest | |
4833 | * andl cf - ct, dest | |
4834 | * [addl dest, ct] | |
4835 | * | |
4836 | * Size 8 - 11. | |
4837 | */ | |
4838 | emit_insn (gen_andsi3 (out, out, GEN_INT (cf - ct))); | |
4839 | if (ct) | |
4840 | emit_insn (gen_addsi3 (out, out, GEN_INT (ct))); | |
4841 | } | |
4842 | ||
4843 | if (tmp != out) | |
4844 | emit_move_insn (out, tmp); | |
4845 | ||
4846 | return 1; /* DONE */ | |
4847 | } | |
4848 | ||
4849 | diff = ct - cf; | |
4850 | if (diff < 0) | |
4851 | { | |
4852 | HOST_WIDE_INT tmp; | |
4853 | tmp = ct, ct = cf, cf = tmp; | |
4854 | diff = -diff; | |
4855 | compare_code = reverse_condition (compare_code); | |
4856 | code = reverse_condition (code); | |
4857 | } | |
4858 | if (diff == 1 || diff == 2 || diff == 4 || diff == 8 | |
4859 | || diff == 3 || diff == 5 || diff == 9) | |
4860 | { | |
4861 | /* | |
4862 | * xorl dest,dest | |
4863 | * cmpl op1,op2 | |
4864 | * setcc dest | |
4865 | * lea cf(dest*(ct-cf)),dest | |
4866 | * | |
4867 | * Size 14. | |
4868 | * | |
4869 | * This also catches the degenerate setcc-only case. | |
4870 | */ | |
4871 | ||
4872 | rtx tmp; | |
4873 | int nops; | |
4874 | ||
4875 | out = emit_store_flag (out, code, ix86_compare_op0, | |
4876 | ix86_compare_op1, VOIDmode, 0, 1); | |
4877 | ||
4878 | nops = 0; | |
4879 | if (diff == 1) | |
4880 | tmp = out; | |
4881 | else | |
4882 | { | |
4883 | tmp = gen_rtx_MULT (SImode, out, GEN_INT (diff & ~1)); | |
4884 | nops++; | |
4885 | if (diff & 1) | |
4886 | { | |
4887 | tmp = gen_rtx_PLUS (SImode, tmp, out); | |
4888 | nops++; | |
4889 | } | |
4890 | } | |
4891 | if (cf != 0) | |
4892 | { | |
4893 | tmp = gen_rtx_PLUS (SImode, tmp, GEN_INT (cf)); | |
4894 | nops++; | |
4895 | } | |
4896 | if (tmp != out) | |
4897 | { | |
4898 | if (nops == 0) | |
4899 | emit_move_insn (out, tmp); | |
4900 | else if (nops == 1) | |
4901 | { | |
4902 | rtx clob; | |
4903 | ||
4904 | clob = gen_rtx_REG (CCmode, FLAGS_REG); | |
4905 | clob = gen_rtx_CLOBBER (VOIDmode, clob); | |
4906 | ||
4907 | tmp = gen_rtx_SET (VOIDmode, out, tmp); | |
4908 | tmp = gen_rtx_PARALLEL (VOIDmode, gen_rtvec (2, tmp, clob)); | |
4909 | emit_insn (tmp); | |
4910 | } | |
4911 | else | |
4912 | emit_insn (gen_rtx_SET (VOIDmode, out, tmp)); | |
4913 | } | |
4914 | if (out != operands[0]) | |
4915 | emit_move_insn (operands[0], out); | |
4916 | ||
4917 | return 1; /* DONE */ | |
4918 | } | |
4919 | ||
4920 | /* | |
4921 | * General case: Jumpful: | |
4922 | * xorl dest,dest cmpl op1, op2 | |
4923 | * cmpl op1, op2 movl ct, dest | |
4924 | * setcc dest jcc 1f | |
4925 | * decl dest movl cf, dest | |
4926 | * andl (cf-ct),dest 1: | |
4927 | * addl ct,dest | |
4928 | * | |
4929 | * Size 20. Size 14. | |
4930 | * | |
4931 | * This is reasonably steep, but branch mispredict costs are | |
4932 | * high on modern cpus, so consider failing only if optimizing | |
4933 | * for space. | |
4934 | * | |
4935 | * %%% Parameterize branch_cost on the tuning architecture, then | |
4936 | * use that. The 80386 couldn't care less about mispredicts. | |
4937 | */ | |
4938 | ||
4939 | if (!optimize_size && !TARGET_CMOVE) | |
4940 | { | |
4941 | if (ct == 0) | |
4942 | { | |
4943 | ct = cf; | |
4944 | cf = 0; | |
4945 | compare_code = reverse_condition (compare_code); | |
4946 | code = reverse_condition (code); | |
4947 | } | |
4948 | ||
4949 | out = emit_store_flag (out, code, ix86_compare_op0, | |
4950 | ix86_compare_op1, VOIDmode, 0, 1); | |
4951 | ||
4952 | emit_insn (gen_addsi3 (out, out, constm1_rtx)); | |
4953 | emit_insn (gen_andsi3 (out, out, GEN_INT (cf-ct))); | |
4954 | if (ct != 0) | |
4955 | emit_insn (gen_addsi3 (out, out, GEN_INT (ct))); | |
4956 | if (out != operands[0]) | |
4957 | emit_move_insn (operands[0], out); | |
4958 | ||
4959 | return 1; /* DONE */ | |
4960 | } | |
4961 | } | |
4962 | ||
4963 | if (!TARGET_CMOVE) | |
4964 | { | |
4965 | /* Try a few things more with specific constants and a variable. */ | |
4966 | ||
4967 | optab op; | |
4968 | rtx var, orig_out, out, tmp; | |
4969 | ||
4970 | if (optimize_size) | |
4971 | return 0; /* FAIL */ | |
4972 | ||
4973 | /* If one of the two operands is an interesting constant, load a | |
4974 | constant with the above and mask it in with a logical operation. */ | |
4975 | ||
4976 | if (GET_CODE (operands[2]) == CONST_INT) | |
4977 | { | |
4978 | var = operands[3]; | |
4979 | if (INTVAL (operands[2]) == 0) | |
4980 | operands[3] = constm1_rtx, op = and_optab; | |
4981 | else if (INTVAL (operands[2]) == -1) | |
4982 | operands[3] = const0_rtx, op = ior_optab; | |
4983 | else | |
4984 | return 0; /* FAIL */ | |
4985 | } | |
4986 | else if (GET_CODE (operands[3]) == CONST_INT) | |
4987 | { | |
4988 | var = operands[2]; | |
4989 | if (INTVAL (operands[3]) == 0) | |
4990 | operands[2] = constm1_rtx, op = and_optab; | |
4991 | else if (INTVAL (operands[3]) == -1) | |
4992 | operands[2] = const0_rtx, op = ior_optab; | |
4993 | else | |
4994 | return 0; /* FAIL */ | |
4995 | } | |
4996 | else | |
4997 | return 0; /* FAIL */ | |
4998 | ||
4999 | orig_out = operands[0]; | |
5000 | tmp = gen_reg_rtx (GET_MODE (orig_out)); | |
5001 | operands[0] = tmp; | |
5002 | ||
5003 | /* Recurse to get the constant loaded. */ | |
5004 | if (ix86_expand_int_movcc (operands) == 0) | |
5005 | return 0; /* FAIL */ | |
5006 | ||
5007 | /* Mask in the interesting variable. */ | |
5008 | out = expand_binop (GET_MODE (orig_out), op, var, tmp, orig_out, 0, | |
5009 | OPTAB_WIDEN); | |
5010 | if (out != orig_out) | |
5011 | emit_move_insn (orig_out, out); | |
5012 | ||
5013 | return 1; /* DONE */ | |
5014 | } | |
5015 | ||
5016 | /* | |
5017 | * For comparison with above, | |
5018 | * | |
5019 | * movl cf,dest | |
5020 | * movl ct,tmp | |
5021 | * cmpl op1,op2 | |
5022 | * cmovcc tmp,dest | |
5023 | * | |
5024 | * Size 15. | |
5025 | */ | |
5026 | ||
5027 | if (! nonimmediate_operand (operands[2], GET_MODE (operands[0]))) | |
5028 | operands[2] = force_reg (GET_MODE (operands[0]), operands[2]); | |
5029 | if (! nonimmediate_operand (operands[3], GET_MODE (operands[0]))) | |
5030 | operands[3] = force_reg (GET_MODE (operands[0]), operands[3]); | |
5031 | ||
5032 | emit_insn (compare_seq); | |
5033 | emit_insn (gen_rtx_SET (VOIDmode, operands[0], | |
5034 | gen_rtx_IF_THEN_ELSE (GET_MODE (operands[0]), | |
5035 | compare_op, operands[2], | |
5036 | operands[3]))); | |
5037 | ||
5038 | return 1; /* DONE */ | |
5039 | } | |
5040 | ||
5041 | int | |
5042 | ix86_expand_fp_movcc (operands) | |
5043 | rtx operands[]; | |
5044 | { | |
5045 | enum rtx_code code; | |
5046 | enum machine_mode mode; | |
5047 | rtx tmp; | |
5048 | ||
5049 | /* The floating point conditional move instructions don't directly | |
5050 | support conditions resulting from a signed integer comparison. */ | |
5051 | ||
5052 | code = GET_CODE (operands[1]); | |
5053 | switch (code) | |
5054 | { | |
5055 | case LT: | |
5056 | case LE: | |
5057 | case GE: | |
5058 | case GT: | |
5059 | tmp = gen_reg_rtx (QImode); | |
5060 | ix86_expand_setcc (code, 0, tmp); | |
5061 | code = NE; | |
5062 | ix86_compare_op0 = tmp; | |
5063 | ix86_compare_op1 = const0_rtx; | |
5064 | break; | |
5065 | ||
5066 | default: | |
5067 | break; | |
5068 | } | |
5069 | ||
5070 | mode = SELECT_CC_MODE (code, ix86_compare_op0, ix86_compare_op1); | |
5071 | emit_insn (gen_rtx_SET (VOIDmode, gen_rtx_REG (mode, FLAGS_REG), | |
5072 | gen_rtx_COMPARE (mode, | |
5073 | ix86_compare_op0, | |
5074 | ix86_compare_op1))); | |
5075 | emit_insn (gen_rtx_SET (VOIDmode, operands[0], | |
5076 | gen_rtx_IF_THEN_ELSE (GET_MODE (operands[0]), | |
5077 | gen_rtx_fmt_ee (code, VOIDmode, | |
5078 | gen_rtx_REG (mode, FLAGS_REG), | |
5079 | const0_rtx), | |
5080 | operands[2], | |
5081 | operands[3]))); | |
5082 | ||
5083 | return 1; | |
5084 | } | |
5085 | ||
5086 | /* Split operands 0 and 1 into SImode parts. Similar to split_di, but | |
5087 | works for floating pointer parameters and nonoffsetable memories. | |
5088 | For pushes, it returns just stack offsets; the values will be saved | |
5089 | in the right order. Maximally three parts are generated. */ | |
5090 | ||
5091 | static void | |
5092 | ix86_split_to_parts (operand, parts, mode) | |
5093 | rtx operand; | |
5094 | rtx *parts; | |
5095 | enum machine_mode mode; | |
5096 | { | |
5097 | int size = GET_MODE_SIZE (mode) / 4; | |
5098 | ||
5099 | if (size < 2 || size > 3) | |
5100 | abort (); | |
5101 | ||
5102 | /* Optimize constant pool reference to immediates. This is used by fp moves, | |
5103 | that force all constants to memory to allow combining. */ | |
5104 | ||
5105 | if (GET_CODE (operand) == MEM | |
5106 | && GET_CODE (XEXP (operand, 0)) == SYMBOL_REF | |
5107 | && CONSTANT_POOL_ADDRESS_P (XEXP (operand, 0))) | |
5108 | operand = get_pool_constant (XEXP (operand, 0)); | |
5109 | ||
5110 | if (GET_CODE (operand) == MEM && !offsettable_memref_p (operand)) | |
5111 | { | |
5112 | /* The only non-offsetable memories we handle are pushes. */ | |
5113 | if (! push_operand (operand, VOIDmode)) | |
5114 | abort (); | |
5115 | ||
5116 | PUT_MODE (operand, SImode); | |
5117 | parts[0] = parts[1] = parts[2] = operand; | |
5118 | } | |
5119 | else | |
5120 | { | |
5121 | if (mode == DImode) | |
5122 | split_di (&operand, 1, &parts[0], &parts[1]); | |
5123 | else | |
5124 | { | |
5125 | if (REG_P (operand)) | |
5126 | { | |
5127 | if (!reload_completed) | |
5128 | abort (); | |
5129 | parts[0] = gen_rtx_REG (SImode, REGNO (operand) + 0); | |
5130 | parts[1] = gen_rtx_REG (SImode, REGNO (operand) + 1); | |
5131 | if (size == 3) | |
5132 | parts[2] = gen_rtx_REG (SImode, REGNO (operand) + 2); | |
5133 | } | |
5134 | else if (offsettable_memref_p (operand)) | |
5135 | { | |
5136 | PUT_MODE (operand, SImode); | |
5137 | parts[0] = operand; | |
5138 | parts[1] = adj_offsettable_operand (operand, 4); | |
5139 | if (size == 3) | |
5140 | parts[2] = adj_offsettable_operand (operand, 8); | |
5141 | } | |
5142 | else if (GET_CODE (operand) == CONST_DOUBLE) | |
5143 | { | |
5144 | REAL_VALUE_TYPE r; | |
5145 | long l[3]; | |
5146 | ||
5147 | REAL_VALUE_FROM_CONST_DOUBLE (r, operand); | |
5148 | switch (mode) | |
5149 | { | |
5150 | case XFmode: | |
5151 | REAL_VALUE_TO_TARGET_LONG_DOUBLE (r, l); | |
5152 | parts[2] = GEN_INT (l[2]); | |
5153 | break; | |
5154 | case DFmode: | |
5155 | REAL_VALUE_TO_TARGET_DOUBLE (r, l); | |
5156 | break; | |
5157 | default: | |
5158 | abort (); | |
5159 | } | |
5160 | parts[1] = GEN_INT (l[1]); | |
5161 | parts[0] = GEN_INT (l[0]); | |
5162 | } | |
5163 | else | |
5164 | abort (); | |
5165 | } | |
5166 | } | |
5167 | ||
5168 | return; | |
5169 | } | |
5170 | ||
5171 | /* Emit insns to perform a move or push of DI, DF, and XF values. | |
5172 | Return false when normal moves are needed; true when all required | |
5173 | insns have been emitted. Operands 2-4 contain the input values | |
5174 | int the correct order; operands 5-7 contain the output values. */ | |
5175 | ||
5176 | int | |
5177 | ix86_split_long_move (operands1) | |
5178 | rtx operands1[]; | |
5179 | { | |
5180 | rtx part[2][3]; | |
5181 | rtx operands[2]; | |
5182 | int size = GET_MODE_SIZE (GET_MODE (operands1[0])) / 4; | |
5183 | int push = 0; | |
5184 | int collisions = 0; | |
5185 | ||
5186 | /* Make our own copy to avoid clobbering the operands. */ | |
5187 | operands[0] = copy_rtx (operands1[0]); | |
5188 | operands[1] = copy_rtx (operands1[1]); | |
5189 | ||
5190 | if (size < 2 || size > 3) | |
5191 | abort (); | |
5192 | ||
5193 | /* The only non-offsettable memory we handle is push. */ | |
5194 | if (push_operand (operands[0], VOIDmode)) | |
5195 | push = 1; | |
5196 | else if (GET_CODE (operands[0]) == MEM | |
5197 | && ! offsettable_memref_p (operands[0])) | |
5198 | abort (); | |
5199 | ||
5200 | ix86_split_to_parts (operands[0], part[0], GET_MODE (operands1[0])); | |
5201 | ix86_split_to_parts (operands[1], part[1], GET_MODE (operands1[0])); | |
5202 | ||
5203 | /* When emitting push, take care for source operands on the stack. */ | |
5204 | if (push && GET_CODE (operands[1]) == MEM | |
5205 | && reg_overlap_mentioned_p (stack_pointer_rtx, operands[1])) | |
5206 | { | |
5207 | if (size == 3) | |
5208 | part[1][1] = part[1][2]; | |
5209 | part[1][0] = part[1][1]; | |
5210 | } | |
5211 | ||
5212 | /* We need to do copy in the right order in case an address register | |
5213 | of the source overlaps the destination. */ | |
5214 | if (REG_P (part[0][0]) && GET_CODE (part[1][0]) == MEM) | |
5215 | { | |
5216 | if (reg_overlap_mentioned_p (part[0][0], XEXP (part[1][0], 0))) | |
5217 | collisions++; | |
5218 | if (reg_overlap_mentioned_p (part[0][1], XEXP (part[1][0], 0))) | |
5219 | collisions++; | |
5220 | if (size == 3 | |
5221 | && reg_overlap_mentioned_p (part[0][2], XEXP (part[1][0], 0))) | |
5222 | collisions++; | |
5223 | ||
5224 | /* Collision in the middle part can be handled by reordering. */ | |
5225 | if (collisions == 1 && size == 3 | |
5226 | && reg_overlap_mentioned_p (part[0][1], XEXP (part[1][0], 0))) | |
5227 | { | |
5228 | rtx tmp; | |
5229 | tmp = part[0][1]; part[0][1] = part[0][2]; part[0][2] = tmp; | |
5230 | tmp = part[1][1]; part[1][1] = part[1][2]; part[1][2] = tmp; | |
5231 | } | |
5232 | ||
5233 | /* If there are more collisions, we can't handle it by reordering. | |
5234 | Do an lea to the last part and use only one colliding move. */ | |
5235 | else if (collisions > 1) | |
5236 | { | |
5237 | collisions = 1; | |
5238 | emit_insn (gen_rtx_SET (VOIDmode, part[0][size - 1], | |
5239 | XEXP (part[1][0], 0))); | |
5240 | part[1][0] = change_address (part[1][0], SImode, part[0][size - 1]); | |
5241 | part[1][1] = adj_offsettable_operand (part[1][0], 4); | |
5242 | if (size == 3) | |
5243 | part[1][2] = adj_offsettable_operand (part[1][0], 8); | |
5244 | } | |
5245 | } | |
5246 | ||
5247 | if (push) | |
5248 | { | |
5249 | if (size == 3) | |
5250 | emit_insn (gen_push (part[1][2])); | |
5251 | emit_insn (gen_push (part[1][1])); | |
5252 | emit_insn (gen_push (part[1][0])); | |
5253 | return 1; | |
5254 | } | |
5255 | ||
5256 | /* Choose correct order to not overwrite the source before it is copied. */ | |
5257 | if ((REG_P (part[0][0]) | |
5258 | && REG_P (part[1][1]) | |
5259 | && (REGNO (part[0][0]) == REGNO (part[1][1]) | |
5260 | || (size == 3 | |
5261 | && REGNO (part[0][0]) == REGNO (part[1][2])))) | |
5262 | || (collisions > 0 | |
5263 | && reg_overlap_mentioned_p (part[0][0], XEXP (part[1][0], 0)))) | |
5264 | { | |
5265 | if (size == 3) | |
5266 | { | |
5267 | operands1[2] = part[0][2]; | |
5268 | operands1[3] = part[0][1]; | |
5269 | operands1[4] = part[0][0]; | |
5270 | operands1[5] = part[1][2]; | |
5271 | operands1[6] = part[1][1]; | |
5272 | operands1[7] = part[1][0]; | |
5273 | } | |
5274 | else | |
5275 | { | |
5276 | operands1[2] = part[0][1]; | |
5277 | operands1[3] = part[0][0]; | |
5278 | operands1[5] = part[1][1]; | |
5279 | operands1[6] = part[1][0]; | |
5280 | } | |
5281 | } | |
5282 | else | |
5283 | { | |
5284 | if (size == 3) | |
5285 | { | |
5286 | operands1[2] = part[0][0]; | |
5287 | operands1[3] = part[0][1]; | |
5288 | operands1[4] = part[0][2]; | |
5289 | operands1[5] = part[1][0]; | |
5290 | operands1[6] = part[1][1]; | |
5291 | operands1[7] = part[1][2]; | |
5292 | } | |
5293 | else | |
5294 | { | |
5295 | operands1[2] = part[0][0]; | |
5296 | operands1[3] = part[0][1]; | |
5297 | operands1[5] = part[1][0]; | |
5298 | operands1[6] = part[1][1]; | |
5299 | } | |
5300 | } | |
5301 | ||
5302 | return 0; | |
5303 | } | |
5304 | ||
5305 | void | |
5306 | ix86_split_ashldi (operands, scratch) | |
5307 | rtx *operands, scratch; | |
5308 | { | |
5309 | rtx low[2], high[2]; | |
5310 | int count; | |
5311 | ||
5312 | if (GET_CODE (operands[2]) == CONST_INT) | |
5313 | { | |
5314 | split_di (operands, 2, low, high); | |
5315 | count = INTVAL (operands[2]) & 63; | |
5316 | ||
5317 | if (count >= 32) | |
5318 | { | |
5319 | emit_move_insn (high[0], low[1]); | |
5320 | emit_move_insn (low[0], const0_rtx); | |
5321 | ||
5322 | if (count > 32) | |
5323 | emit_insn (gen_ashlsi3 (high[0], high[0], GEN_INT (count - 32))); | |
5324 | } | |
5325 | else | |
5326 | { | |
5327 | if (!rtx_equal_p (operands[0], operands[1])) | |
5328 | emit_move_insn (operands[0], operands[1]); | |
5329 | emit_insn (gen_x86_shld_1 (high[0], low[0], GEN_INT (count))); | |
5330 | emit_insn (gen_ashlsi3 (low[0], low[0], GEN_INT (count))); | |
5331 | } | |
5332 | } | |
5333 | else | |
5334 | { | |
5335 | if (!rtx_equal_p (operands[0], operands[1])) | |
5336 | emit_move_insn (operands[0], operands[1]); | |
5337 | ||
5338 | split_di (operands, 1, low, high); | |
5339 | ||
5340 | emit_insn (gen_x86_shld_1 (high[0], low[0], operands[2])); | |
5341 | emit_insn (gen_ashlsi3 (low[0], low[0], operands[2])); | |
5342 | ||
5343 | if (TARGET_CMOVE && (! reload_completed || scratch)) | |
5344 | { | |
5345 | if (! reload_completed) | |
5346 | scratch = force_reg (SImode, const0_rtx); | |
5347 | else | |
5348 | emit_move_insn (scratch, const0_rtx); | |
5349 | ||
5350 | emit_insn (gen_x86_shift_adj_1 (high[0], low[0], operands[2], | |
5351 | scratch)); | |
5352 | } | |
5353 | else | |
5354 | emit_insn (gen_x86_shift_adj_2 (high[0], low[0], operands[2])); | |
5355 | } | |
5356 | } | |
5357 | ||
5358 | void | |
5359 | ix86_split_ashrdi (operands, scratch) | |
5360 | rtx *operands, scratch; | |
5361 | { | |
5362 | rtx low[2], high[2]; | |
5363 | int count; | |
5364 | ||
5365 | if (GET_CODE (operands[2]) == CONST_INT) | |
5366 | { | |
5367 | split_di (operands, 2, low, high); | |
5368 | count = INTVAL (operands[2]) & 63; | |
5369 | ||
5370 | if (count >= 32) | |
5371 | { | |
5372 | emit_move_insn (low[0], high[1]); | |
5373 | ||
5374 | if (! reload_completed) | |
5375 | emit_insn (gen_ashrsi3 (high[0], low[0], GEN_INT (31))); | |
5376 | else | |
5377 | { | |
5378 | emit_move_insn (high[0], low[0]); | |
5379 | emit_insn (gen_ashrsi3 (high[0], high[0], GEN_INT (31))); | |
5380 | } | |
5381 | ||
5382 | if (count > 32) | |
5383 | emit_insn (gen_ashrsi3 (low[0], low[0], GEN_INT (count - 32))); | |
5384 | } | |
5385 | else | |
5386 | { | |
5387 | if (!rtx_equal_p (operands[0], operands[1])) | |
5388 | emit_move_insn (operands[0], operands[1]); | |
5389 | emit_insn (gen_x86_shrd_1 (low[0], high[0], GEN_INT (count))); | |
5390 | emit_insn (gen_ashrsi3 (high[0], high[0], GEN_INT (count))); | |
5391 | } | |
5392 | } | |
5393 | else | |
5394 | { | |
5395 | if (!rtx_equal_p (operands[0], operands[1])) | |
5396 | emit_move_insn (operands[0], operands[1]); | |
5397 | ||
5398 | split_di (operands, 1, low, high); | |
5399 | ||
5400 | emit_insn (gen_x86_shrd_1 (low[0], high[0], operands[2])); | |
5401 | emit_insn (gen_ashrsi3 (high[0], high[0], operands[2])); | |
5402 | ||
5403 | if (TARGET_CMOVE && (!reload_completed || scratch)) | |
5404 | { | |
5405 | if (! reload_completed) | |
5406 | scratch = gen_reg_rtx (SImode); | |
5407 | emit_move_insn (scratch, high[0]); | |
5408 | emit_insn (gen_ashrsi3 (scratch, scratch, GEN_INT (31))); | |
5409 | emit_insn (gen_x86_shift_adj_1 (low[0], high[0], operands[2], | |
5410 | scratch)); | |
5411 | } | |
5412 | else | |
5413 | emit_insn (gen_x86_shift_adj_3 (low[0], high[0], operands[2])); | |
5414 | } | |
5415 | } | |
5416 | ||
5417 | void | |
5418 | ix86_split_lshrdi (operands, scratch) | |
5419 | rtx *operands, scratch; | |
5420 | { | |
5421 | rtx low[2], high[2]; | |
5422 | int count; | |
5423 | ||
5424 | if (GET_CODE (operands[2]) == CONST_INT) | |
5425 | { | |
5426 | split_di (operands, 2, low, high); | |
5427 | count = INTVAL (operands[2]) & 63; | |
5428 | ||
5429 | if (count >= 32) | |
5430 | { | |
5431 | emit_move_insn (low[0], high[1]); | |
5432 | emit_move_insn (high[0], const0_rtx); | |
5433 | ||
5434 | if (count > 32) | |
5435 | emit_insn (gen_lshrsi3 (low[0], low[0], GEN_INT (count - 32))); | |
5436 | } | |
5437 | else | |
5438 | { | |
5439 | if (!rtx_equal_p (operands[0], operands[1])) | |
5440 | emit_move_insn (operands[0], operands[1]); | |
5441 | emit_insn (gen_x86_shrd_1 (low[0], high[0], GEN_INT (count))); | |
5442 | emit_insn (gen_lshrsi3 (high[0], high[0], GEN_INT (count))); | |
5443 | } | |
5444 | } | |
5445 | else | |
5446 | { | |
5447 | if (!rtx_equal_p (operands[0], operands[1])) | |
5448 | emit_move_insn (operands[0], operands[1]); | |
5449 | ||
5450 | split_di (operands, 1, low, high); | |
5451 | ||
5452 | emit_insn (gen_x86_shrd_1 (low[0], high[0], operands[2])); | |
5453 | emit_insn (gen_lshrsi3 (high[0], high[0], operands[2])); | |
5454 | ||
5455 | /* Heh. By reversing the arguments, we can reuse this pattern. */ | |
5456 | if (TARGET_CMOVE && (! reload_completed || scratch)) | |
5457 | { | |
5458 | if (! reload_completed) | |
5459 | scratch = force_reg (SImode, const0_rtx); | |
5460 | else | |
5461 | emit_move_insn (scratch, const0_rtx); | |
5462 | ||
5463 | emit_insn (gen_x86_shift_adj_1 (low[0], high[0], operands[2], | |
5464 | scratch)); | |
5465 | } | |
5466 | else | |
5467 | emit_insn (gen_x86_shift_adj_2 (low[0], high[0], operands[2])); | |
5468 | } | |
5469 | } | |
5470 | ||
5471 | /* Expand the appropriate insns for doing strlen if not just doing | |
5472 | repnz; scasb | |
5473 | ||
5474 | out = result, initialized with the start address | |
5475 | align_rtx = alignment of the address. | |
5476 | scratch = scratch register, initialized with the startaddress when | |
5477 | not aligned, otherwise undefined | |
5478 | ||
5479 | This is just the body. It needs the initialisations mentioned above and | |
5480 | some address computing at the end. These things are done in i386.md. */ | |
5481 | ||
5482 | void | |
5483 | ix86_expand_strlensi_unroll_1 (out, align_rtx, scratch) | |
5484 | rtx out, align_rtx, scratch; | |
5485 | { | |
5486 | int align; | |
5487 | rtx tmp; | |
5488 | rtx align_2_label = NULL_RTX; | |
5489 | rtx align_3_label = NULL_RTX; | |
5490 | rtx align_4_label = gen_label_rtx (); | |
5491 | rtx end_0_label = gen_label_rtx (); | |
5492 | rtx mem; | |
5493 | rtx flags = gen_rtx_REG (CCNOmode, FLAGS_REG); | |
5494 | rtx tmpreg = gen_reg_rtx (SImode); | |
5495 | ||
5496 | align = 0; | |
5497 | if (GET_CODE (align_rtx) == CONST_INT) | |
5498 | align = INTVAL (align_rtx); | |
5499 | ||
5500 | /* Loop to check 1..3 bytes for null to get an aligned pointer. */ | |
5501 | ||
5502 | /* Is there a known alignment and is it less than 4? */ | |
5503 | if (align < 4) | |
5504 | { | |
5505 | /* Is there a known alignment and is it not 2? */ | |
5506 | if (align != 2) | |
5507 | { | |
5508 | align_3_label = gen_label_rtx (); /* Label when aligned to 3-byte */ | |
5509 | align_2_label = gen_label_rtx (); /* Label when aligned to 2-byte */ | |
5510 | ||
5511 | /* Leave just the 3 lower bits. */ | |
5512 | align_rtx = expand_binop (SImode, and_optab, scratch, GEN_INT (3), | |
5513 | NULL_RTX, 0, OPTAB_WIDEN); | |
5514 | ||
5515 | emit_insn (gen_cmpsi_0 (align_rtx, const0_rtx)); | |
5516 | ||
5517 | tmp = gen_rtx_EQ (VOIDmode, flags, const0_rtx); | |
5518 | tmp = gen_rtx_IF_THEN_ELSE (VOIDmode, tmp, | |
5519 | gen_rtx_LABEL_REF (VOIDmode, | |
5520 | align_4_label), | |
5521 | pc_rtx); | |
5522 | emit_jump_insn (gen_rtx_SET (VOIDmode, pc_rtx, tmp)); | |
5523 | ||
5524 | emit_insn (gen_cmpsi_1 (align_rtx, GEN_INT (2))); | |
5525 | ||
5526 | tmp = gen_rtx_EQ (VOIDmode, flags, const0_rtx); | |
5527 | tmp = gen_rtx_IF_THEN_ELSE (VOIDmode, tmp, | |
5528 | gen_rtx_LABEL_REF (VOIDmode, | |
5529 | align_2_label), | |
5530 | pc_rtx); | |
5531 | emit_jump_insn (gen_rtx_SET (VOIDmode, pc_rtx, tmp)); | |
5532 | ||
5533 | tmp = gen_rtx_GTU (VOIDmode, flags, const0_rtx); | |
5534 | tmp = gen_rtx_IF_THEN_ELSE (VOIDmode, tmp, | |
5535 | gen_rtx_LABEL_REF (VOIDmode, | |
5536 | align_3_label), | |
5537 | pc_rtx); | |
5538 | emit_jump_insn (gen_rtx_SET (VOIDmode, pc_rtx, tmp)); | |
5539 | } | |
5540 | else | |
5541 | { | |
5542 | /* Since the alignment is 2, we have to check 2 or 0 bytes; | |
5543 | check if is aligned to 4 - byte. */ | |
5544 | ||
5545 | align_rtx = expand_binop (SImode, and_optab, scratch, GEN_INT (2), | |
5546 | NULL_RTX, 0, OPTAB_WIDEN); | |
5547 | ||
5548 | emit_insn (gen_cmpsi_0 (align_rtx, const0_rtx)); | |
5549 | ||
5550 | tmp = gen_rtx_EQ (VOIDmode, flags, const0_rtx); | |
5551 | tmp = gen_rtx_IF_THEN_ELSE (VOIDmode, tmp, | |
5552 | gen_rtx_LABEL_REF (VOIDmode, | |
5553 | align_4_label), | |
5554 | pc_rtx); | |
5555 | emit_jump_insn (gen_rtx_SET (VOIDmode, pc_rtx, tmp)); | |
5556 | } | |
5557 | ||
5558 | mem = gen_rtx_MEM (QImode, out); | |
5559 | ||
5560 | /* Now compare the bytes. */ | |
5561 | ||
5562 | /* Compare the first n unaligned byte on a byte per byte basis. */ | |
5563 | emit_insn (gen_cmpqi_0 (mem, const0_rtx)); | |
5564 | ||
5565 | tmp = gen_rtx_EQ (VOIDmode, flags, const0_rtx); | |
5566 | tmp = gen_rtx_IF_THEN_ELSE (VOIDmode, tmp, | |
5567 | gen_rtx_LABEL_REF (VOIDmode, end_0_label), | |
5568 | pc_rtx); | |
5569 | emit_jump_insn (gen_rtx_SET (VOIDmode, pc_rtx, tmp)); | |
5570 | ||
5571 | /* Increment the address. */ | |
5572 | emit_insn (gen_addsi3 (out, out, const1_rtx)); | |
5573 | ||
5574 | /* Not needed with an alignment of 2 */ | |
5575 | if (align != 2) | |
5576 | { | |
5577 | emit_label (align_2_label); | |
5578 | ||
5579 | emit_insn (gen_cmpqi_0 (mem, const0_rtx)); | |
5580 | ||
5581 | tmp = gen_rtx_EQ (VOIDmode, flags, const0_rtx); | |
5582 | tmp = gen_rtx_IF_THEN_ELSE (VOIDmode, tmp, | |
5583 | gen_rtx_LABEL_REF (VOIDmode, | |
5584 | end_0_label), | |
5585 | pc_rtx); | |
5586 | emit_jump_insn (gen_rtx_SET (VOIDmode, pc_rtx, tmp)); | |
5587 | ||
5588 | emit_insn (gen_addsi3 (out, out, const1_rtx)); | |
5589 | ||
5590 | emit_label (align_3_label); | |
5591 | } | |
5592 | ||
5593 | emit_insn (gen_cmpqi_0 (mem, const0_rtx)); | |
5594 | ||
5595 | tmp = gen_rtx_EQ (VOIDmode, flags, const0_rtx); | |
5596 | tmp = gen_rtx_IF_THEN_ELSE (VOIDmode, tmp, | |
5597 | gen_rtx_LABEL_REF (VOIDmode, end_0_label), | |
5598 | pc_rtx); | |
5599 | emit_jump_insn (gen_rtx_SET (VOIDmode, pc_rtx, tmp)); | |
5600 | ||
5601 | emit_insn (gen_addsi3 (out, out, const1_rtx)); | |
5602 | } | |
5603 | ||
5604 | /* Generate loop to check 4 bytes at a time. It is not a good idea to | |
5605 | align this loop. It gives only huge programs, but does not help to | |
5606 | speed up. */ | |
5607 | emit_label (align_4_label); | |
5608 | ||
5609 | mem = gen_rtx_MEM (SImode, out); | |
5610 | emit_move_insn (scratch, mem); | |
5611 | emit_insn (gen_addsi3 (out, out, GEN_INT (4))); | |
5612 | ||
5613 | /* This formula yields a nonzero result iff one of the bytes is zero. | |
5614 | This saves three branches inside loop and many cycles. */ | |
5615 | ||
5616 | emit_insn (gen_addsi3 (tmpreg, scratch, GEN_INT (-0x01010101))); | |
5617 | emit_insn (gen_one_cmplsi2 (scratch, scratch)); | |
5618 | emit_insn (gen_andsi3 (tmpreg, tmpreg, scratch)); | |
5619 | emit_insn (gen_andsi3 (tmpreg, tmpreg, GEN_INT (0x80808080))); | |
5620 | emit_cmp_and_jump_insns (tmpreg, const0_rtx, EQ, 0, SImode, 1, 0, align_4_label); | |
5621 | ||
5622 | if (TARGET_CMOVE) | |
5623 | { | |
5624 | rtx reg = gen_reg_rtx (SImode); | |
5625 | emit_move_insn (reg, tmpreg); | |
5626 | emit_insn (gen_lshrsi3 (reg, reg, GEN_INT (16))); | |
5627 | ||
5628 | /* If zero is not in the first two bytes, move two bytes forward. */ | |
5629 | emit_insn (gen_testsi_1 (tmpreg, GEN_INT (0x8080))); | |
5630 | tmp = gen_rtx_REG (CCNOmode, FLAGS_REG); | |
5631 | tmp = gen_rtx_EQ (VOIDmode, tmp, const0_rtx); | |
5632 | emit_insn (gen_rtx_SET (VOIDmode, tmpreg, | |
5633 | gen_rtx_IF_THEN_ELSE (SImode, tmp, | |
5634 | reg, | |
5635 | tmpreg))); | |
5636 | /* Emit lea manually to avoid clobbering of flags. */ | |
5637 | emit_insn (gen_rtx_SET (SImode, reg, | |
5638 | gen_rtx_PLUS (SImode, out, GEN_INT (2)))); | |
5639 | ||
5640 | tmp = gen_rtx_REG (CCNOmode, FLAGS_REG); | |
5641 | tmp = gen_rtx_EQ (VOIDmode, tmp, const0_rtx); | |
5642 | emit_insn (gen_rtx_SET (VOIDmode, out, | |
5643 | gen_rtx_IF_THEN_ELSE (SImode, tmp, | |
5644 | reg, | |
5645 | out))); | |
5646 | ||
5647 | } | |
5648 | else | |
5649 | { | |
5650 | rtx end_2_label = gen_label_rtx (); | |
5651 | /* Is zero in the first two bytes? */ | |
5652 | ||
5653 | emit_insn (gen_testsi_1 (tmpreg, GEN_INT (0x8080))); | |
5654 | tmp = gen_rtx_REG (CCNOmode, FLAGS_REG); | |
5655 | tmp = gen_rtx_NE (VOIDmode, tmp, const0_rtx); | |
5656 | tmp = gen_rtx_IF_THEN_ELSE (VOIDmode, tmp, | |
5657 | gen_rtx_LABEL_REF (VOIDmode, end_2_label), | |
5658 | pc_rtx); | |
5659 | tmp = emit_jump_insn (gen_rtx_SET (VOIDmode, pc_rtx, tmp)); | |
5660 | JUMP_LABEL (tmp) = end_2_label; | |
5661 | ||
5662 | /* Not in the first two. Move two bytes forward. */ | |
5663 | emit_insn (gen_lshrsi3 (tmpreg, tmpreg, GEN_INT (16))); | |
5664 | emit_insn (gen_addsi3 (out, out, GEN_INT (2))); | |
5665 | ||
5666 | emit_label (end_2_label); | |
5667 | ||
5668 | } | |
5669 | ||
5670 | /* Avoid branch in fixing the byte. */ | |
5671 | tmpreg = gen_lowpart (QImode, tmpreg); | |
5672 | emit_insn (gen_addqi3_cc (tmpreg, tmpreg, tmpreg)); | |
5673 | emit_insn (gen_subsi3_carry (out, out, GEN_INT (3))); | |
5674 | ||
5675 | emit_label (end_0_label); | |
5676 | } | |
5677 | \f | |
5678 | /* Clear stack slot assignments remembered from previous functions. | |
5679 | This is called from INIT_EXPANDERS once before RTL is emitted for each | |
5680 | function. */ | |
5681 | ||
5682 | static void | |
5683 | ix86_init_machine_status (p) | |
5684 | struct function *p; | |
5685 | { | |
5686 | enum machine_mode mode; | |
5687 | int n; | |
5688 | p->machine | |
5689 | = (struct machine_function *) xmalloc (sizeof (struct machine_function)); | |
5690 | ||
5691 | for (mode = VOIDmode; (int) mode < (int) MAX_MACHINE_MODE; | |
5692 | mode = (enum machine_mode) ((int) mode + 1)) | |
5693 | for (n = 0; n < MAX_386_STACK_LOCALS; n++) | |
5694 | ix86_stack_locals[(int) mode][n] = NULL_RTX; | |
5695 | } | |
5696 | ||
5697 | /* Mark machine specific bits of P for GC. */ | |
5698 | static void | |
5699 | ix86_mark_machine_status (p) | |
5700 | struct function *p; | |
5701 | { | |
5702 | enum machine_mode mode; | |
5703 | int n; | |
5704 | ||
5705 | for (mode = VOIDmode; (int) mode < (int) MAX_MACHINE_MODE; | |
5706 | mode = (enum machine_mode) ((int) mode + 1)) | |
5707 | for (n = 0; n < MAX_386_STACK_LOCALS; n++) | |
5708 | ggc_mark_rtx (p->machine->stack_locals[(int) mode][n]); | |
5709 | } | |
5710 | ||
5711 | /* Return a MEM corresponding to a stack slot with mode MODE. | |
5712 | Allocate a new slot if necessary. | |
5713 | ||
5714 | The RTL for a function can have several slots available: N is | |
5715 | which slot to use. */ | |
5716 | ||
5717 | rtx | |
5718 | assign_386_stack_local (mode, n) | |
5719 | enum machine_mode mode; | |
5720 | int n; | |
5721 | { | |
5722 | if (n < 0 || n >= MAX_386_STACK_LOCALS) | |
5723 | abort (); | |
5724 | ||
5725 | if (ix86_stack_locals[(int) mode][n] == NULL_RTX) | |
5726 | ix86_stack_locals[(int) mode][n] | |
5727 | = assign_stack_local (mode, GET_MODE_SIZE (mode), 0); | |
5728 | ||
5729 | return ix86_stack_locals[(int) mode][n]; | |
5730 | } | |
5731 | \f | |
5732 | /* Calculate the length of the memory address in the instruction | |
5733 | encoding. Does not include the one-byte modrm, opcode, or prefix. */ | |
5734 | ||
5735 | static int | |
5736 | memory_address_length (addr) | |
5737 | rtx addr; | |
5738 | { | |
5739 | struct ix86_address parts; | |
5740 | rtx base, index, disp; | |
5741 | int len; | |
5742 | ||
5743 | if (GET_CODE (addr) == PRE_DEC | |
5744 | || GET_CODE (addr) == POST_INC) | |
5745 | return 0; | |
5746 | ||
5747 | if (! ix86_decompose_address (addr, &parts)) | |
5748 | abort (); | |
5749 | ||
5750 | base = parts.base; | |
5751 | index = parts.index; | |
5752 | disp = parts.disp; | |
5753 | len = 0; | |
5754 | ||
5755 | /* Register Indirect. */ | |
5756 | if (base && !index && !disp) | |
5757 | { | |
5758 | /* Special cases: ebp and esp need the two-byte modrm form. */ | |
5759 | if (addr == stack_pointer_rtx | |
5760 | || addr == arg_pointer_rtx | |
5761 | || addr == frame_pointer_rtx | |
5762 | || addr == hard_frame_pointer_rtx) | |
5763 | len = 1; | |
5764 | } | |
5765 | ||
5766 | /* Direct Addressing. */ | |
5767 | else if (disp && !base && !index) | |
5768 | len = 4; | |
5769 | ||
5770 | else | |
5771 | { | |
5772 | /* Find the length of the displacement constant. */ | |
5773 | if (disp) | |
5774 | { | |
5775 | if (GET_CODE (disp) == CONST_INT | |
5776 | && CONST_OK_FOR_LETTER_P (INTVAL (disp), 'K')) | |
5777 | len = 1; | |
5778 | else | |
5779 | len = 4; | |
5780 | } | |
5781 | ||
5782 | /* An index requires the two-byte modrm form. */ | |
5783 | if (index) | |
5784 | len += 1; | |
5785 | } | |
5786 | ||
5787 | return len; | |
5788 | } | |
5789 | ||
5790 | int | |
5791 | ix86_attr_length_default (insn) | |
5792 | rtx insn; | |
5793 | { | |
5794 | enum attr_type type; | |
5795 | int len = 0, i; | |
5796 | ||
5797 | type = get_attr_type (insn); | |
5798 | extract_insn (insn); | |
5799 | switch (type) | |
5800 | { | |
5801 | case TYPE_INCDEC: | |
5802 | case TYPE_SETCC: | |
5803 | case TYPE_ICMOV: | |
5804 | case TYPE_FMOV: | |
5805 | case TYPE_FOP: | |
5806 | case TYPE_FCMP: | |
5807 | case TYPE_FOP1: | |
5808 | case TYPE_FMUL: | |
5809 | case TYPE_FDIV: | |
5810 | case TYPE_FSGN: | |
5811 | case TYPE_FPSPC: | |
5812 | case TYPE_FCMOV: | |
5813 | case TYPE_IBR: | |
5814 | break; | |
5815 | case TYPE_STR: | |
5816 | case TYPE_CLD: | |
5817 | len = 0; | |
5818 | ||
5819 | case TYPE_ALU1: | |
5820 | case TYPE_NEGNOT: | |
5821 | case TYPE_ALU: | |
5822 | case TYPE_ICMP: | |
5823 | case TYPE_IMOVX: | |
5824 | case TYPE_ISHIFT: | |
5825 | case TYPE_IMUL: | |
5826 | case TYPE_IDIV: | |
5827 | case TYPE_PUSH: | |
5828 | case TYPE_POP: | |
5829 | for (i = recog_data.n_operands - 1; i >= 0; --i) | |
5830 | if (CONSTANT_P (recog_data.operand[i])) | |
5831 | { | |
5832 | if (GET_CODE (recog_data.operand[i]) == CONST_INT | |
5833 | && CONST_OK_FOR_LETTER_P (INTVAL (recog_data.operand[i]), 'K')) | |
5834 | len += 1; | |
5835 | else | |
5836 | len += GET_MODE_SIZE (GET_MODE (recog_data.operand[0])); | |
5837 | } | |
5838 | break; | |
5839 | ||
5840 | case TYPE_IMOV: | |
5841 | if (CONSTANT_P (recog_data.operand[1])) | |
5842 | len += GET_MODE_SIZE (GET_MODE (recog_data.operand[0])); | |
5843 | break; | |
5844 | ||
5845 | case TYPE_CALL: | |
5846 | if (constant_call_address_operand (recog_data.operand[0], | |
5847 | GET_MODE (recog_data.operand[0]))) | |
5848 | return 5; | |
5849 | break; | |
5850 | ||
5851 | case TYPE_CALLV: | |
5852 | if (constant_call_address_operand (recog_data.operand[1], | |
5853 | GET_MODE (recog_data.operand[1]))) | |
5854 | return 5; | |
5855 | break; | |
5856 | ||
5857 | case TYPE_LEA: | |
5858 | { | |
5859 | /* Irritatingly, single_set doesn't work with REG_UNUSED present, | |
5860 | as we'll get from running life_analysis during reg-stack when | |
5861 | not optimizing. Not that it matters anyway, now that | |
5862 | pro_epilogue_adjust_stack uses lea, and is by design not | |
5863 | single_set. */ | |
5864 | rtx set = PATTERN (insn); | |
5865 | if (GET_CODE (set) == SET) | |
5866 | ; | |
5867 | else if (GET_CODE (set) == PARALLEL | |
5868 | && GET_CODE (XVECEXP (set, 0, 0)) == SET) | |
5869 | set = XVECEXP (set, 0, 0); | |
5870 | else | |
5871 | abort (); | |
5872 | ||
5873 | len += memory_address_length (SET_SRC (set)); | |
5874 | goto just_opcode; | |
5875 | } | |
5876 | ||
5877 | case TYPE_OTHER: | |
5878 | case TYPE_MULTI: | |
5879 | return 15; | |
5880 | ||
5881 | case TYPE_FXCH: | |
5882 | if (STACK_TOP_P (recog_data.operand[0])) | |
5883 | return 2 + (REGNO (recog_data.operand[1]) != FIRST_STACK_REG + 1); | |
5884 | else | |
5885 | return 2 + (REGNO (recog_data.operand[0]) != FIRST_STACK_REG + 1); | |
5886 | ||
5887 | default: | |
5888 | abort (); | |
5889 | } | |
5890 | ||
5891 | for (i = recog_data.n_operands - 1; i >= 0; --i) | |
5892 | if (GET_CODE (recog_data.operand[i]) == MEM) | |
5893 | { | |
5894 | len += memory_address_length (XEXP (recog_data.operand[i], 0)); | |
5895 | break; | |
5896 | } | |
5897 | ||
5898 | just_opcode: | |
5899 | len += get_attr_length_opcode (insn); | |
5900 | len += get_attr_length_prefix (insn); | |
5901 | ||
5902 | return len; | |
5903 | } | |
5904 | \f | |
5905 | /* Return the maximum number of instructions a cpu can issue. */ | |
5906 | ||
5907 | int | |
5908 | ix86_issue_rate () | |
5909 | { | |
5910 | switch (ix86_cpu) | |
5911 | { | |
5912 | case PROCESSOR_PENTIUM: | |
5913 | case PROCESSOR_K6: | |
5914 | return 2; | |
5915 | ||
5916 | case PROCESSOR_PENTIUMPRO: | |
5917 | return 3; | |
5918 | ||
5919 | default: | |
5920 | return 1; | |
5921 | } | |
5922 | } | |
5923 | ||
5924 | /* A subroutine of ix86_adjust_cost -- return true iff INSN reads flags set | |
5925 | by DEP_INSN and nothing set by DEP_INSN. */ | |
5926 | ||
5927 | static int | |
5928 | ix86_flags_dependant (insn, dep_insn, insn_type) | |
5929 | rtx insn, dep_insn; | |
5930 | enum attr_type insn_type; | |
5931 | { | |
5932 | rtx set, set2; | |
5933 | ||
5934 | /* Simplify the test for uninteresting insns. */ | |
5935 | if (insn_type != TYPE_SETCC | |
5936 | && insn_type != TYPE_ICMOV | |
5937 | && insn_type != TYPE_FCMOV | |
5938 | && insn_type != TYPE_IBR) | |
5939 | return 0; | |
5940 | ||
5941 | if ((set = single_set (dep_insn)) != 0) | |
5942 | { | |
5943 | set = SET_DEST (set); | |
5944 | set2 = NULL_RTX; | |
5945 | } | |
5946 | else if (GET_CODE (PATTERN (dep_insn)) == PARALLEL | |
5947 | && XVECLEN (PATTERN (dep_insn), 0) == 2 | |
5948 | && GET_CODE (XVECEXP (PATTERN (dep_insn), 0, 0)) == SET | |
5949 | && GET_CODE (XVECEXP (PATTERN (dep_insn), 0, 1)) == SET) | |
5950 | { | |
5951 | set = SET_DEST (XVECEXP (PATTERN (dep_insn), 0, 0)); | |
5952 | set2 = SET_DEST (XVECEXP (PATTERN (dep_insn), 0, 0)); | |
5953 | } | |
5954 | else | |
5955 | return 0; | |
5956 | ||
5957 | if (GET_CODE (set) != REG || REGNO (set) != FLAGS_REG) | |
5958 | return 0; | |
5959 | ||
5960 | /* This test is true if the dependant insn reads the flags but | |
5961 | not any other potentially set register. */ | |
5962 | if (!reg_overlap_mentioned_p (set, PATTERN (insn))) | |
5963 | return 0; | |
5964 | ||
5965 | if (set2 && reg_overlap_mentioned_p (set2, PATTERN (insn))) | |
5966 | return 0; | |
5967 | ||
5968 | return 1; | |
5969 | } | |
5970 | ||
5971 | /* A subroutine of ix86_adjust_cost -- return true iff INSN has a memory | |
5972 | address with operands set by DEP_INSN. */ | |
5973 | ||
5974 | static int | |
5975 | ix86_agi_dependant (insn, dep_insn, insn_type) | |
5976 | rtx insn, dep_insn; | |
5977 | enum attr_type insn_type; | |
5978 | { | |
5979 | rtx addr; | |
5980 | ||
5981 | if (insn_type == TYPE_LEA) | |
5982 | { | |
5983 | addr = PATTERN (insn); | |
5984 | if (GET_CODE (addr) == SET) | |
5985 | ; | |
5986 | else if (GET_CODE (addr) == PARALLEL | |
5987 | && GET_CODE (XVECEXP (addr, 0, 0)) == SET) | |
5988 | addr = XVECEXP (addr, 0, 0); | |
5989 | else | |
5990 | abort (); | |
5991 | addr = SET_SRC (addr); | |
5992 | } | |
5993 | else | |
5994 | { | |
5995 | int i; | |
5996 | extract_insn (insn); | |
5997 | for (i = recog_data.n_operands - 1; i >= 0; --i) | |
5998 | if (GET_CODE (recog_data.operand[i]) == MEM) | |
5999 | { | |
6000 | addr = XEXP (recog_data.operand[i], 0); | |
6001 | goto found; | |
6002 | } | |
6003 | return 0; | |
6004 | found:; | |
6005 | } | |
6006 | ||
6007 | return modified_in_p (addr, dep_insn); | |
6008 | } | |
6009 | ||
6010 | int | |
6011 | ix86_adjust_cost (insn, link, dep_insn, cost) | |
6012 | rtx insn, link, dep_insn; | |
6013 | int cost; | |
6014 | { | |
6015 | enum attr_type insn_type, dep_insn_type; | |
6016 | rtx set, set2; | |
6017 | int dep_insn_code_number; | |
6018 | ||
6019 | /* Anti and output depenancies have zero cost on all CPUs. */ | |
6020 | if (REG_NOTE_KIND (link) != 0) | |
6021 | return 0; | |
6022 | ||
6023 | dep_insn_code_number = recog_memoized (dep_insn); | |
6024 | ||
6025 | /* If we can't recognize the insns, we can't really do anything. */ | |
6026 | if (dep_insn_code_number < 0 || recog_memoized (insn) < 0) | |
6027 | return cost; | |
6028 | ||
6029 | insn_type = get_attr_type (insn); | |
6030 | dep_insn_type = get_attr_type (dep_insn); | |
6031 | ||
6032 | /* Prologue and epilogue allocators can have a false dependency on ebp. | |
6033 | This results in one cycle extra stall on Pentium prologue scheduling, | |
6034 | so handle this important case manually. */ | |
6035 | if (dep_insn_code_number == CODE_FOR_pro_epilogue_adjust_stack | |
6036 | && dep_insn_type == TYPE_ALU | |
6037 | && !reg_mentioned_p (stack_pointer_rtx, insn)) | |
6038 | return 0; | |
6039 | ||
6040 | switch (ix86_cpu) | |
6041 | { | |
6042 | case PROCESSOR_PENTIUM: | |
6043 | /* Address Generation Interlock adds a cycle of latency. */ | |
6044 | if (ix86_agi_dependant (insn, dep_insn, insn_type)) | |
6045 | cost += 1; | |
6046 | ||
6047 | /* ??? Compares pair with jump/setcc. */ | |
6048 | if (ix86_flags_dependant (insn, dep_insn, insn_type)) | |
6049 | cost = 0; | |
6050 | ||
6051 | /* Floating point stores require value to be ready one cycle ealier. */ | |
6052 | if (insn_type == TYPE_FMOV | |
6053 | && get_attr_memory (insn) == MEMORY_STORE | |
6054 | && !ix86_agi_dependant (insn, dep_insn, insn_type)) | |
6055 | cost += 1; | |
6056 | break; | |
6057 | ||
6058 | case PROCESSOR_PENTIUMPRO: | |
6059 | /* Since we can't represent delayed latencies of load+operation, | |
6060 | increase the cost here for non-imov insns. */ | |
6061 | if (dep_insn_type != TYPE_IMOV | |
6062 | && dep_insn_type != TYPE_FMOV | |
6063 | && get_attr_memory (dep_insn) == MEMORY_LOAD) | |
6064 | cost += 1; | |
6065 | ||
6066 | /* INT->FP conversion is expensive. */ | |
6067 | if (get_attr_fp_int_src (dep_insn)) | |
6068 | cost += 5; | |
6069 | ||
6070 | /* There is one cycle extra latency between an FP op and a store. */ | |
6071 | if (insn_type == TYPE_FMOV | |
6072 | && (set = single_set (dep_insn)) != NULL_RTX | |
6073 | && (set2 = single_set (insn)) != NULL_RTX | |
6074 | && rtx_equal_p (SET_DEST (set), SET_SRC (set2)) | |
6075 | && GET_CODE (SET_DEST (set2)) == MEM) | |
6076 | cost += 1; | |
6077 | break; | |
6078 | ||
6079 | case PROCESSOR_K6: | |
6080 | /* The esp dependency is resolved before the instruction is really | |
6081 | finished. */ | |
6082 | if ((insn_type == TYPE_PUSH || insn_type == TYPE_POP) | |
6083 | && (dep_insn_type == TYPE_PUSH || dep_insn_type == TYPE_POP)) | |
6084 | return 1; | |
6085 | ||
6086 | /* Since we can't represent delayed latencies of load+operation, | |
6087 | increase the cost here for non-imov insns. */ | |
6088 | if (get_attr_memory (dep_insn) == MEMORY_LOAD) | |
6089 | cost += (dep_insn_type != TYPE_IMOV) ? 2 : 1; | |
6090 | ||
6091 | /* INT->FP conversion is expensive. */ | |
6092 | if (get_attr_fp_int_src (dep_insn)) | |
6093 | cost += 5; | |
6094 | break; | |
6095 | ||
6096 | case PROCESSOR_ATHLON: | |
6097 | /* Address Generation Interlock cause problems on the Athlon CPU because | |
6098 | the loads and stores are done in order so once one load or store has | |
6099 | to wait, others must too, so penalize the AGIs slightly by one cycle. | |
6100 | We might experiment with this value later. */ | |
6101 | if (ix86_agi_dependant (insn, dep_insn, insn_type)) | |
6102 | cost += 1; | |
6103 | ||
6104 | /* Since we can't represent delayed latencies of load+operation, | |
6105 | increase the cost here for non-imov insns. */ | |
6106 | if (dep_insn_type != TYPE_IMOV | |
6107 | && dep_insn_type != TYPE_FMOV | |
6108 | && get_attr_memory (dep_insn) == MEMORY_LOAD) | |
6109 | cost += 2; | |
6110 | default: | |
6111 | break; | |
6112 | } | |
6113 | ||
6114 | return cost; | |
6115 | } | |
6116 | ||
6117 | static union | |
6118 | { | |
6119 | struct ppro_sched_data | |
6120 | { | |
6121 | rtx decode[3]; | |
6122 | int issued_this_cycle; | |
6123 | } ppro; | |
6124 | } ix86_sched_data; | |
6125 | ||
6126 | static int | |
6127 | ix86_safe_length (insn) | |
6128 | rtx insn; | |
6129 | { | |
6130 | if (recog_memoized (insn) >= 0) | |
6131 | return get_attr_length(insn); | |
6132 | else | |
6133 | return 128; | |
6134 | } | |
6135 | ||
6136 | static int | |
6137 | ix86_safe_length_prefix (insn) | |
6138 | rtx insn; | |
6139 | { | |
6140 | if (recog_memoized (insn) >= 0) | |
6141 | return get_attr_length(insn); | |
6142 | else | |
6143 | return 0; | |
6144 | } | |
6145 | ||
6146 | static enum attr_memory | |
6147 | ix86_safe_memory (insn) | |
6148 | rtx insn; | |
6149 | { | |
6150 | if (recog_memoized (insn) >= 0) | |
6151 | return get_attr_memory(insn); | |
6152 | else | |
6153 | return MEMORY_UNKNOWN; | |
6154 | } | |
6155 | ||
6156 | static enum attr_pent_pair | |
6157 | ix86_safe_pent_pair (insn) | |
6158 | rtx insn; | |
6159 | { | |
6160 | if (recog_memoized (insn) >= 0) | |
6161 | return get_attr_pent_pair(insn); | |
6162 | else | |
6163 | return PENT_PAIR_NP; | |
6164 | } | |
6165 | ||
6166 | static enum attr_ppro_uops | |
6167 | ix86_safe_ppro_uops (insn) | |
6168 | rtx insn; | |
6169 | { | |
6170 | if (recog_memoized (insn) >= 0) | |
6171 | return get_attr_ppro_uops (insn); | |
6172 | else | |
6173 | return PPRO_UOPS_MANY; | |
6174 | } | |
6175 | ||
6176 | static void | |
6177 | ix86_dump_ppro_packet (dump) | |
6178 | FILE *dump; | |
6179 | { | |
6180 | if (ix86_sched_data.ppro.decode[0]) | |
6181 | { | |
6182 | fprintf (dump, "PPRO packet: %d", | |
6183 | INSN_UID (ix86_sched_data.ppro.decode[0])); | |
6184 | if (ix86_sched_data.ppro.decode[1]) | |
6185 | fprintf (dump, " %d", INSN_UID (ix86_sched_data.ppro.decode[1])); | |
6186 | if (ix86_sched_data.ppro.decode[2]) | |
6187 | fprintf (dump, " %d", INSN_UID (ix86_sched_data.ppro.decode[2])); | |
6188 | fputc ('\n', dump); | |
6189 | } | |
6190 | } | |
6191 | ||
6192 | /* We're beginning a new block. Initialize data structures as necessary. */ | |
6193 | ||
6194 | void | |
6195 | ix86_sched_init (dump, sched_verbose) | |
6196 | FILE *dump ATTRIBUTE_UNUSED; | |
6197 | int sched_verbose ATTRIBUTE_UNUSED; | |
6198 | { | |
6199 | memset (&ix86_sched_data, 0, sizeof (ix86_sched_data)); | |
6200 | } | |
6201 | ||
6202 | /* Shift INSN to SLOT, and shift everything else down. */ | |
6203 | ||
6204 | static void | |
6205 | ix86_reorder_insn (insnp, slot) | |
6206 | rtx *insnp, *slot; | |
6207 | { | |
6208 | if (insnp != slot) | |
6209 | { | |
6210 | rtx insn = *insnp; | |
6211 | do | |
6212 | insnp[0] = insnp[1]; | |
6213 | while (++insnp != slot); | |
6214 | *insnp = insn; | |
6215 | } | |
6216 | } | |
6217 | ||
6218 | /* Find an instruction with given pairability and minimal amount of cycles | |
6219 | lost by the fact that the CPU waits for both pipelines to finish before | |
6220 | reading next instructions. Also take care that both instructions together | |
6221 | can not exceed 7 bytes. */ | |
6222 | ||
6223 | static rtx * | |
6224 | ix86_pent_find_pair (e_ready, ready, type, first) | |
6225 | rtx *e_ready; | |
6226 | rtx *ready; | |
6227 | enum attr_pent_pair type; | |
6228 | rtx first; | |
6229 | { | |
6230 | int mincycles, cycles; | |
6231 | enum attr_pent_pair tmp; | |
6232 | enum attr_memory memory; | |
6233 | rtx *insnp, *bestinsnp = NULL; | |
6234 | ||
6235 | if (ix86_safe_length (first) > 7 + ix86_safe_length_prefix (first)) | |
6236 | return NULL; | |
6237 | ||
6238 | memory = ix86_safe_memory (first); | |
6239 | cycles = result_ready_cost (first); | |
6240 | mincycles = INT_MAX; | |
6241 | ||
6242 | for (insnp = e_ready; insnp >= ready && mincycles; --insnp) | |
6243 | if ((tmp = ix86_safe_pent_pair (*insnp)) == type | |
6244 | && ix86_safe_length (*insnp) <= 7 + ix86_safe_length_prefix (*insnp)) | |
6245 | { | |
6246 | enum attr_memory second_memory; | |
6247 | int secondcycles, currentcycles; | |
6248 | ||
6249 | second_memory = ix86_safe_memory (*insnp); | |
6250 | secondcycles = result_ready_cost (*insnp); | |
6251 | currentcycles = abs (cycles - secondcycles); | |
6252 | ||
6253 | if (secondcycles >= 1 && cycles >= 1) | |
6254 | { | |
6255 | /* Two read/modify/write instructions together takes two | |
6256 | cycles longer. */ | |
6257 | if (memory == MEMORY_BOTH && second_memory == MEMORY_BOTH) | |
6258 | currentcycles += 2; | |
6259 | ||
6260 | /* Read modify/write instruction followed by read/modify | |
6261 | takes one cycle longer. */ | |
6262 | if (memory == MEMORY_BOTH && second_memory == MEMORY_LOAD | |
6263 | && tmp != PENT_PAIR_UV | |
6264 | && ix86_safe_pent_pair (first) != PENT_PAIR_UV) | |
6265 | currentcycles += 1; | |
6266 | } | |
6267 | if (currentcycles < mincycles) | |
6268 | bestinsnp = insnp, mincycles = currentcycles; | |
6269 | } | |
6270 | ||
6271 | return bestinsnp; | |
6272 | } | |
6273 | ||
6274 | /* Subroutines of ix86_sched_reorder. */ | |
6275 | ||
6276 | static void | |
6277 | ix86_sched_reorder_pentium (ready, e_ready) | |
6278 | rtx *ready; | |
6279 | rtx *e_ready; | |
6280 | { | |
6281 | enum attr_pent_pair pair1, pair2; | |
6282 | rtx *insnp; | |
6283 | ||
6284 | /* This wouldn't be necessary if Haifa knew that static insn ordering | |
6285 | is important to which pipe an insn is issued to. So we have to make | |
6286 | some minor rearrangements. */ | |
6287 | ||
6288 | pair1 = ix86_safe_pent_pair (*e_ready); | |
6289 | ||
6290 | /* If the first insn is non-pairable, let it be. */ | |
6291 | if (pair1 == PENT_PAIR_NP) | |
6292 | return; | |
6293 | ||
6294 | pair2 = PENT_PAIR_NP; | |
6295 | insnp = 0; | |
6296 | ||
6297 | /* If the first insn is UV or PV pairable, search for a PU | |
6298 | insn to go with. */ | |
6299 | if (pair1 == PENT_PAIR_UV || pair1 == PENT_PAIR_PV) | |
6300 | { | |
6301 | insnp = ix86_pent_find_pair (e_ready-1, ready, | |
6302 | PENT_PAIR_PU, *e_ready); | |
6303 | if (insnp) | |
6304 | pair2 = PENT_PAIR_PU; | |
6305 | } | |
6306 | ||
6307 | /* If the first insn is PU or UV pairable, search for a PV | |
6308 | insn to go with. */ | |
6309 | if (pair2 == PENT_PAIR_NP | |
6310 | && (pair1 == PENT_PAIR_PU || pair1 == PENT_PAIR_UV)) | |
6311 | { | |
6312 | insnp = ix86_pent_find_pair (e_ready-1, ready, | |
6313 | PENT_PAIR_PV, *e_ready); | |
6314 | if (insnp) | |
6315 | pair2 = PENT_PAIR_PV; | |
6316 | } | |
6317 | ||
6318 | /* If the first insn is pairable, search for a UV | |
6319 | insn to go with. */ | |
6320 | if (pair2 == PENT_PAIR_NP) | |
6321 | { | |
6322 | insnp = ix86_pent_find_pair (e_ready-1, ready, | |
6323 | PENT_PAIR_UV, *e_ready); | |
6324 | if (insnp) | |
6325 | pair2 = PENT_PAIR_UV; | |
6326 | } | |
6327 | ||
6328 | if (pair2 == PENT_PAIR_NP) | |
6329 | return; | |
6330 | ||
6331 | /* Found something! Decide if we need to swap the order. */ | |
6332 | if (pair1 == PENT_PAIR_PV || pair2 == PENT_PAIR_PU | |
6333 | || (pair1 == PENT_PAIR_UV && pair2 == PENT_PAIR_UV | |
6334 | && ix86_safe_memory (*e_ready) == MEMORY_BOTH | |
6335 | && ix86_safe_memory (*insnp) == MEMORY_LOAD)) | |
6336 | ix86_reorder_insn (insnp, e_ready); | |
6337 | else | |
6338 | ix86_reorder_insn (insnp, e_ready - 1); | |
6339 | } | |
6340 | ||
6341 | static void | |
6342 | ix86_sched_reorder_ppro (ready, e_ready) | |
6343 | rtx *ready; | |
6344 | rtx *e_ready; | |
6345 | { | |
6346 | rtx decode[3]; | |
6347 | enum attr_ppro_uops cur_uops; | |
6348 | int issued_this_cycle; | |
6349 | rtx *insnp; | |
6350 | int i; | |
6351 | ||
6352 | /* At this point .ppro.decode contains the state of the three | |
6353 | decoders from last "cycle". That is, those insns that were | |
6354 | actually independent. But here we're scheduling for the | |
6355 | decoder, and we may find things that are decodable in the | |
6356 | same cycle. */ | |
6357 | ||
6358 | memcpy (decode, ix86_sched_data.ppro.decode, sizeof(decode)); | |
6359 | issued_this_cycle = 0; | |
6360 | ||
6361 | insnp = e_ready; | |
6362 | cur_uops = ix86_safe_ppro_uops (*insnp); | |
6363 | ||
6364 | /* If the decoders are empty, and we've a complex insn at the | |
6365 | head of the priority queue, let it issue without complaint. */ | |
6366 | if (decode[0] == NULL) | |
6367 | { | |
6368 | if (cur_uops == PPRO_UOPS_MANY) | |
6369 | { | |
6370 | decode[0] = *insnp; | |
6371 | goto ppro_done; | |
6372 | } | |
6373 | ||
6374 | /* Otherwise, search for a 2-4 uop unsn to issue. */ | |
6375 | while (cur_uops != PPRO_UOPS_FEW) | |
6376 | { | |
6377 | if (insnp == ready) | |
6378 | break; | |
6379 | cur_uops = ix86_safe_ppro_uops (*--insnp); | |
6380 | } | |
6381 | ||
6382 | /* If so, move it to the head of the line. */ | |
6383 | if (cur_uops == PPRO_UOPS_FEW) | |
6384 | ix86_reorder_insn (insnp, e_ready); | |
6385 | ||
6386 | /* Issue the head of the queue. */ | |
6387 | issued_this_cycle = 1; | |
6388 | decode[0] = *e_ready--; | |
6389 | } | |
6390 | ||
6391 | /* Look for simple insns to fill in the other two slots. */ | |
6392 | for (i = 1; i < 3; ++i) | |
6393 | if (decode[i] == NULL) | |
6394 | { | |
6395 | if (ready >= e_ready) | |
6396 | goto ppro_done; | |
6397 | ||
6398 | insnp = e_ready; | |
6399 | cur_uops = ix86_safe_ppro_uops (*insnp); | |
6400 | while (cur_uops != PPRO_UOPS_ONE) | |
6401 | { | |
6402 | if (insnp == ready) | |
6403 | break; | |
6404 | cur_uops = ix86_safe_ppro_uops (*--insnp); | |
6405 | } | |
6406 | ||
6407 | /* Found one. Move it to the head of the queue and issue it. */ | |
6408 | if (cur_uops == PPRO_UOPS_ONE) | |
6409 | { | |
6410 | ix86_reorder_insn (insnp, e_ready); | |
6411 | decode[i] = *e_ready--; | |
6412 | issued_this_cycle++; | |
6413 | continue; | |
6414 | } | |
6415 | ||
6416 | /* ??? Didn't find one. Ideally, here we would do a lazy split | |
6417 | of 2-uop insns, issue one and queue the other. */ | |
6418 | } | |
6419 | ||
6420 | ppro_done: | |
6421 | if (issued_this_cycle == 0) | |
6422 | issued_this_cycle = 1; | |
6423 | ix86_sched_data.ppro.issued_this_cycle = issued_this_cycle; | |
6424 | } | |
6425 | ||
6426 | ||
6427 | /* We are about to being issuing insns for this clock cycle. | |
6428 | Override the default sort algorithm to better slot instructions. */ | |
6429 | int | |
6430 | ix86_sched_reorder (dump, sched_verbose, ready, n_ready, clock_var) | |
6431 | FILE *dump ATTRIBUTE_UNUSED; | |
6432 | int sched_verbose ATTRIBUTE_UNUSED; | |
6433 | rtx *ready; | |
6434 | int n_ready; | |
6435 | int clock_var ATTRIBUTE_UNUSED; | |
6436 | { | |
6437 | rtx *e_ready = ready + n_ready - 1; | |
6438 | ||
6439 | if (n_ready < 2) | |
6440 | goto out; | |
6441 | ||
6442 | switch (ix86_cpu) | |
6443 | { | |
6444 | default: | |
6445 | break; | |
6446 | ||
6447 | case PROCESSOR_PENTIUM: | |
6448 | ix86_sched_reorder_pentium (ready, e_ready); | |
6449 | break; | |
6450 | ||
6451 | case PROCESSOR_PENTIUMPRO: | |
6452 | ix86_sched_reorder_ppro (ready, e_ready); | |
6453 | break; | |
6454 | } | |
6455 | ||
6456 | out: | |
6457 | return ix86_issue_rate (); | |
6458 | } | |
6459 | ||
6460 | /* We are about to issue INSN. Return the number of insns left on the | |
6461 | ready queue that can be issued this cycle. */ | |
6462 | ||
6463 | int | |
6464 | ix86_variable_issue (dump, sched_verbose, insn, can_issue_more) | |
6465 | FILE *dump; | |
6466 | int sched_verbose; | |
6467 | rtx insn; | |
6468 | int can_issue_more; | |
6469 | { | |
6470 | int i; | |
6471 | switch (ix86_cpu) | |
6472 | { | |
6473 | default: | |
6474 | return can_issue_more - 1; | |
6475 | ||
6476 | case PROCESSOR_PENTIUMPRO: | |
6477 | { | |
6478 | enum attr_ppro_uops uops = ix86_safe_ppro_uops (insn); | |
6479 | ||
6480 | if (uops == PPRO_UOPS_MANY) | |
6481 | { | |
6482 | if (sched_verbose) | |
6483 | ix86_dump_ppro_packet (dump); | |
6484 | ix86_sched_data.ppro.decode[0] = insn; | |
6485 | ix86_sched_data.ppro.decode[1] = NULL; | |
6486 | ix86_sched_data.ppro.decode[2] = NULL; | |
6487 | if (sched_verbose) | |
6488 | ix86_dump_ppro_packet (dump); | |
6489 | ix86_sched_data.ppro.decode[0] = NULL; | |
6490 | } | |
6491 | else if (uops == PPRO_UOPS_FEW) | |
6492 | { | |
6493 | if (sched_verbose) | |
6494 | ix86_dump_ppro_packet (dump); | |
6495 | ix86_sched_data.ppro.decode[0] = insn; | |
6496 | ix86_sched_data.ppro.decode[1] = NULL; | |
6497 | ix86_sched_data.ppro.decode[2] = NULL; | |
6498 | } | |
6499 | else | |
6500 | { | |
6501 | for (i = 0; i < 3; ++i) | |
6502 | if (ix86_sched_data.ppro.decode[i] == NULL) | |
6503 | { | |
6504 | ix86_sched_data.ppro.decode[i] = insn; | |
6505 | break; | |
6506 | } | |
6507 | if (i == 3) | |
6508 | abort (); | |
6509 | if (i == 2) | |
6510 | { | |
6511 | if (sched_verbose) | |
6512 | ix86_dump_ppro_packet (dump); | |
6513 | ix86_sched_data.ppro.decode[0] = NULL; | |
6514 | ix86_sched_data.ppro.decode[1] = NULL; | |
6515 | ix86_sched_data.ppro.decode[2] = NULL; | |
6516 | } | |
6517 | } | |
6518 | } | |
6519 | return --ix86_sched_data.ppro.issued_this_cycle; | |
6520 | } | |
6521 | } |