Mercurial > hg > CbC > CbC_gcc
comparison gcc/reg-stack.c @ 0:a06113de4d67
first commit
author | kent <kent@cr.ie.u-ryukyu.ac.jp> |
---|---|
date | Fri, 17 Jul 2009 14:47:48 +0900 |
parents | |
children | 58ad6c70ea60 |
comparison
equal
deleted
inserted
replaced
-1:000000000000 | 0:a06113de4d67 |
---|---|
1 /* Register to Stack convert for GNU compiler. | |
2 Copyright (C) 1992, 1993, 1994, 1995, 1996, 1997, 1998, 1999, 2000, | |
3 2001, 2002, 2003, 2004, 2005, 2006, 2007, 2008 | |
4 Free Software Foundation, Inc. | |
5 | |
6 This file is part of GCC. | |
7 | |
8 GCC is free software; you can redistribute it and/or modify it | |
9 under the terms of the GNU General Public License as published by | |
10 the Free Software Foundation; either version 3, or (at your option) | |
11 any later version. | |
12 | |
13 GCC is distributed in the hope that it will be useful, but WITHOUT | |
14 ANY WARRANTY; without even the implied warranty of MERCHANTABILITY | |
15 or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public | |
16 License for more details. | |
17 | |
18 You should have received a copy of the GNU General Public License | |
19 along with GCC; see the file COPYING3. If not see | |
20 <http://www.gnu.org/licenses/>. */ | |
21 | |
22 /* This pass converts stack-like registers from the "flat register | |
23 file" model that gcc uses, to a stack convention that the 387 uses. | |
24 | |
25 * The form of the input: | |
26 | |
27 On input, the function consists of insn that have had their | |
28 registers fully allocated to a set of "virtual" registers. Note that | |
29 the word "virtual" is used differently here than elsewhere in gcc: for | |
30 each virtual stack reg, there is a hard reg, but the mapping between | |
31 them is not known until this pass is run. On output, hard register | |
32 numbers have been substituted, and various pop and exchange insns have | |
33 been emitted. The hard register numbers and the virtual register | |
34 numbers completely overlap - before this pass, all stack register | |
35 numbers are virtual, and afterward they are all hard. | |
36 | |
37 The virtual registers can be manipulated normally by gcc, and their | |
38 semantics are the same as for normal registers. After the hard | |
39 register numbers are substituted, the semantics of an insn containing | |
40 stack-like regs are not the same as for an insn with normal regs: for | |
41 instance, it is not safe to delete an insn that appears to be a no-op | |
42 move. In general, no insn containing hard regs should be changed | |
43 after this pass is done. | |
44 | |
45 * The form of the output: | |
46 | |
47 After this pass, hard register numbers represent the distance from | |
48 the current top of stack to the desired register. A reference to | |
49 FIRST_STACK_REG references the top of stack, FIRST_STACK_REG + 1, | |
50 represents the register just below that, and so forth. Also, REG_DEAD | |
51 notes indicate whether or not a stack register should be popped. | |
52 | |
53 A "swap" insn looks like a parallel of two patterns, where each | |
54 pattern is a SET: one sets A to B, the other B to A. | |
55 | |
56 A "push" or "load" insn is a SET whose SET_DEST is FIRST_STACK_REG | |
57 and whose SET_DEST is REG or MEM. Any other SET_DEST, such as PLUS, | |
58 will replace the existing stack top, not push a new value. | |
59 | |
60 A store insn is a SET whose SET_DEST is FIRST_STACK_REG, and whose | |
61 SET_SRC is REG or MEM. | |
62 | |
63 The case where the SET_SRC and SET_DEST are both FIRST_STACK_REG | |
64 appears ambiguous. As a special case, the presence of a REG_DEAD note | |
65 for FIRST_STACK_REG differentiates between a load insn and a pop. | |
66 | |
67 If a REG_DEAD is present, the insn represents a "pop" that discards | |
68 the top of the register stack. If there is no REG_DEAD note, then the | |
69 insn represents a "dup" or a push of the current top of stack onto the | |
70 stack. | |
71 | |
72 * Methodology: | |
73 | |
74 Existing REG_DEAD and REG_UNUSED notes for stack registers are | |
75 deleted and recreated from scratch. REG_DEAD is never created for a | |
76 SET_DEST, only REG_UNUSED. | |
77 | |
78 * asm_operands: | |
79 | |
80 There are several rules on the usage of stack-like regs in | |
81 asm_operands insns. These rules apply only to the operands that are | |
82 stack-like regs: | |
83 | |
84 1. Given a set of input regs that die in an asm_operands, it is | |
85 necessary to know which are implicitly popped by the asm, and | |
86 which must be explicitly popped by gcc. | |
87 | |
88 An input reg that is implicitly popped by the asm must be | |
89 explicitly clobbered, unless it is constrained to match an | |
90 output operand. | |
91 | |
92 2. For any input reg that is implicitly popped by an asm, it is | |
93 necessary to know how to adjust the stack to compensate for the pop. | |
94 If any non-popped input is closer to the top of the reg-stack than | |
95 the implicitly popped reg, it would not be possible to know what the | |
96 stack looked like - it's not clear how the rest of the stack "slides | |
97 up". | |
98 | |
99 All implicitly popped input regs must be closer to the top of | |
100 the reg-stack than any input that is not implicitly popped. | |
101 | |
102 3. It is possible that if an input dies in an insn, reload might | |
103 use the input reg for an output reload. Consider this example: | |
104 | |
105 asm ("foo" : "=t" (a) : "f" (b)); | |
106 | |
107 This asm says that input B is not popped by the asm, and that | |
108 the asm pushes a result onto the reg-stack, i.e., the stack is one | |
109 deeper after the asm than it was before. But, it is possible that | |
110 reload will think that it can use the same reg for both the input and | |
111 the output, if input B dies in this insn. | |
112 | |
113 If any input operand uses the "f" constraint, all output reg | |
114 constraints must use the "&" earlyclobber. | |
115 | |
116 The asm above would be written as | |
117 | |
118 asm ("foo" : "=&t" (a) : "f" (b)); | |
119 | |
120 4. Some operands need to be in particular places on the stack. All | |
121 output operands fall in this category - there is no other way to | |
122 know which regs the outputs appear in unless the user indicates | |
123 this in the constraints. | |
124 | |
125 Output operands must specifically indicate which reg an output | |
126 appears in after an asm. "=f" is not allowed: the operand | |
127 constraints must select a class with a single reg. | |
128 | |
129 5. Output operands may not be "inserted" between existing stack regs. | |
130 Since no 387 opcode uses a read/write operand, all output operands | |
131 are dead before the asm_operands, and are pushed by the asm_operands. | |
132 It makes no sense to push anywhere but the top of the reg-stack. | |
133 | |
134 Output operands must start at the top of the reg-stack: output | |
135 operands may not "skip" a reg. | |
136 | |
137 6. Some asm statements may need extra stack space for internal | |
138 calculations. This can be guaranteed by clobbering stack registers | |
139 unrelated to the inputs and outputs. | |
140 | |
141 Here are a couple of reasonable asms to want to write. This asm | |
142 takes one input, which is internally popped, and produces two outputs. | |
143 | |
144 asm ("fsincos" : "=t" (cos), "=u" (sin) : "0" (inp)); | |
145 | |
146 This asm takes two inputs, which are popped by the fyl2xp1 opcode, | |
147 and replaces them with one output. The user must code the "st(1)" | |
148 clobber for reg-stack.c to know that fyl2xp1 pops both inputs. | |
149 | |
150 asm ("fyl2xp1" : "=t" (result) : "0" (x), "u" (y) : "st(1)"); | |
151 | |
152 */ | |
153 | |
154 #include "config.h" | |
155 #include "system.h" | |
156 #include "coretypes.h" | |
157 #include "tm.h" | |
158 #include "tree.h" | |
159 #include "rtl.h" | |
160 #include "tm_p.h" | |
161 #include "function.h" | |
162 #include "insn-config.h" | |
163 #include "regs.h" | |
164 #include "hard-reg-set.h" | |
165 #include "flags.h" | |
166 #include "toplev.h" | |
167 #include "recog.h" | |
168 #include "output.h" | |
169 #include "basic-block.h" | |
170 #include "cfglayout.h" | |
171 #include "varray.h" | |
172 #include "reload.h" | |
173 #include "ggc.h" | |
174 #include "timevar.h" | |
175 #include "tree-pass.h" | |
176 #include "target.h" | |
177 #include "df.h" | |
178 #include "vecprim.h" | |
179 | |
180 #ifdef STACK_REGS | |
181 | |
182 /* We use this array to cache info about insns, because otherwise we | |
183 spend too much time in stack_regs_mentioned_p. | |
184 | |
185 Indexed by insn UIDs. A value of zero is uninitialized, one indicates | |
186 the insn uses stack registers, two indicates the insn does not use | |
187 stack registers. */ | |
188 static VEC(char,heap) *stack_regs_mentioned_data; | |
189 | |
190 #define REG_STACK_SIZE (LAST_STACK_REG - FIRST_STACK_REG + 1) | |
191 | |
192 int regstack_completed = 0; | |
193 | |
194 /* This is the basic stack record. TOP is an index into REG[] such | |
195 that REG[TOP] is the top of stack. If TOP is -1 the stack is empty. | |
196 | |
197 If TOP is -2, REG[] is not yet initialized. Stack initialization | |
198 consists of placing each live reg in array `reg' and setting `top' | |
199 appropriately. | |
200 | |
201 REG_SET indicates which registers are live. */ | |
202 | |
203 typedef struct stack_def | |
204 { | |
205 int top; /* index to top stack element */ | |
206 HARD_REG_SET reg_set; /* set of live registers */ | |
207 unsigned char reg[REG_STACK_SIZE];/* register - stack mapping */ | |
208 } *stack; | |
209 | |
210 /* This is used to carry information about basic blocks. It is | |
211 attached to the AUX field of the standard CFG block. */ | |
212 | |
213 typedef struct block_info_def | |
214 { | |
215 struct stack_def stack_in; /* Input stack configuration. */ | |
216 struct stack_def stack_out; /* Output stack configuration. */ | |
217 HARD_REG_SET out_reg_set; /* Stack regs live on output. */ | |
218 int done; /* True if block already converted. */ | |
219 int predecessors; /* Number of predecessors that need | |
220 to be visited. */ | |
221 } *block_info; | |
222 | |
223 #define BLOCK_INFO(B) ((block_info) (B)->aux) | |
224 | |
225 /* Passed to change_stack to indicate where to emit insns. */ | |
226 enum emit_where | |
227 { | |
228 EMIT_AFTER, | |
229 EMIT_BEFORE | |
230 }; | |
231 | |
232 /* The block we're currently working on. */ | |
233 static basic_block current_block; | |
234 | |
235 /* In the current_block, whether we're processing the first register | |
236 stack or call instruction, i.e. the regstack is currently the | |
237 same as BLOCK_INFO(current_block)->stack_in. */ | |
238 static bool starting_stack_p; | |
239 | |
240 /* This is the register file for all register after conversion. */ | |
241 static rtx | |
242 FP_mode_reg[LAST_STACK_REG+1-FIRST_STACK_REG][(int) MAX_MACHINE_MODE]; | |
243 | |
244 #define FP_MODE_REG(regno,mode) \ | |
245 (FP_mode_reg[(regno)-FIRST_STACK_REG][(int) (mode)]) | |
246 | |
247 /* Used to initialize uninitialized registers. */ | |
248 static rtx not_a_num; | |
249 | |
250 /* Forward declarations */ | |
251 | |
252 static int stack_regs_mentioned_p (const_rtx pat); | |
253 static void pop_stack (stack, int); | |
254 static rtx *get_true_reg (rtx *); | |
255 | |
256 static int check_asm_stack_operands (rtx); | |
257 static int get_asm_operand_n_inputs (rtx); | |
258 static rtx stack_result (tree); | |
259 static void replace_reg (rtx *, int); | |
260 static void remove_regno_note (rtx, enum reg_note, unsigned int); | |
261 static int get_hard_regnum (stack, rtx); | |
262 static rtx emit_pop_insn (rtx, stack, rtx, enum emit_where); | |
263 static void swap_to_top(rtx, stack, rtx, rtx); | |
264 static bool move_for_stack_reg (rtx, stack, rtx); | |
265 static bool move_nan_for_stack_reg (rtx, stack, rtx); | |
266 static int swap_rtx_condition_1 (rtx); | |
267 static int swap_rtx_condition (rtx); | |
268 static void compare_for_stack_reg (rtx, stack, rtx); | |
269 static bool subst_stack_regs_pat (rtx, stack, rtx); | |
270 static void subst_asm_stack_regs (rtx, stack); | |
271 static bool subst_stack_regs (rtx, stack); | |
272 static void change_stack (rtx, stack, stack, enum emit_where); | |
273 static void print_stack (FILE *, stack); | |
274 static rtx next_flags_user (rtx); | |
275 | |
276 /* Return nonzero if any stack register is mentioned somewhere within PAT. */ | |
277 | |
278 static int | |
279 stack_regs_mentioned_p (const_rtx pat) | |
280 { | |
281 const char *fmt; | |
282 int i; | |
283 | |
284 if (STACK_REG_P (pat)) | |
285 return 1; | |
286 | |
287 fmt = GET_RTX_FORMAT (GET_CODE (pat)); | |
288 for (i = GET_RTX_LENGTH (GET_CODE (pat)) - 1; i >= 0; i--) | |
289 { | |
290 if (fmt[i] == 'E') | |
291 { | |
292 int j; | |
293 | |
294 for (j = XVECLEN (pat, i) - 1; j >= 0; j--) | |
295 if (stack_regs_mentioned_p (XVECEXP (pat, i, j))) | |
296 return 1; | |
297 } | |
298 else if (fmt[i] == 'e' && stack_regs_mentioned_p (XEXP (pat, i))) | |
299 return 1; | |
300 } | |
301 | |
302 return 0; | |
303 } | |
304 | |
305 /* Return nonzero if INSN mentions stacked registers, else return zero. */ | |
306 | |
307 int | |
308 stack_regs_mentioned (const_rtx insn) | |
309 { | |
310 unsigned int uid, max; | |
311 int test; | |
312 | |
313 if (! INSN_P (insn) || !stack_regs_mentioned_data) | |
314 return 0; | |
315 | |
316 uid = INSN_UID (insn); | |
317 max = VEC_length (char, stack_regs_mentioned_data); | |
318 if (uid >= max) | |
319 { | |
320 /* Allocate some extra size to avoid too many reallocs, but | |
321 do not grow too quickly. */ | |
322 max = uid + uid / 20 + 1; | |
323 VEC_safe_grow_cleared (char, heap, stack_regs_mentioned_data, max); | |
324 } | |
325 | |
326 test = VEC_index (char, stack_regs_mentioned_data, uid); | |
327 if (test == 0) | |
328 { | |
329 /* This insn has yet to be examined. Do so now. */ | |
330 test = stack_regs_mentioned_p (PATTERN (insn)) ? 1 : 2; | |
331 VEC_replace (char, stack_regs_mentioned_data, uid, test); | |
332 } | |
333 | |
334 return test == 1; | |
335 } | |
336 | |
337 static rtx ix86_flags_rtx; | |
338 | |
339 static rtx | |
340 next_flags_user (rtx insn) | |
341 { | |
342 /* Search forward looking for the first use of this value. | |
343 Stop at block boundaries. */ | |
344 | |
345 while (insn != BB_END (current_block)) | |
346 { | |
347 insn = NEXT_INSN (insn); | |
348 | |
349 if (INSN_P (insn) && reg_mentioned_p (ix86_flags_rtx, PATTERN (insn))) | |
350 return insn; | |
351 | |
352 if (CALL_P (insn)) | |
353 return NULL_RTX; | |
354 } | |
355 return NULL_RTX; | |
356 } | |
357 | |
358 /* Reorganize the stack into ascending numbers, before this insn. */ | |
359 | |
360 static void | |
361 straighten_stack (rtx insn, stack regstack) | |
362 { | |
363 struct stack_def temp_stack; | |
364 int top; | |
365 | |
366 /* If there is only a single register on the stack, then the stack is | |
367 already in increasing order and no reorganization is needed. | |
368 | |
369 Similarly if the stack is empty. */ | |
370 if (regstack->top <= 0) | |
371 return; | |
372 | |
373 COPY_HARD_REG_SET (temp_stack.reg_set, regstack->reg_set); | |
374 | |
375 for (top = temp_stack.top = regstack->top; top >= 0; top--) | |
376 temp_stack.reg[top] = FIRST_STACK_REG + temp_stack.top - top; | |
377 | |
378 change_stack (insn, regstack, &temp_stack, EMIT_BEFORE); | |
379 } | |
380 | |
381 /* Pop a register from the stack. */ | |
382 | |
383 static void | |
384 pop_stack (stack regstack, int regno) | |
385 { | |
386 int top = regstack->top; | |
387 | |
388 CLEAR_HARD_REG_BIT (regstack->reg_set, regno); | |
389 regstack->top--; | |
390 /* If regno was not at the top of stack then adjust stack. */ | |
391 if (regstack->reg [top] != regno) | |
392 { | |
393 int i; | |
394 for (i = regstack->top; i >= 0; i--) | |
395 if (regstack->reg [i] == regno) | |
396 { | |
397 int j; | |
398 for (j = i; j < top; j++) | |
399 regstack->reg [j] = regstack->reg [j + 1]; | |
400 break; | |
401 } | |
402 } | |
403 } | |
404 | |
405 /* Return a pointer to the REG expression within PAT. If PAT is not a | |
406 REG, possible enclosed by a conversion rtx, return the inner part of | |
407 PAT that stopped the search. */ | |
408 | |
409 static rtx * | |
410 get_true_reg (rtx *pat) | |
411 { | |
412 for (;;) | |
413 switch (GET_CODE (*pat)) | |
414 { | |
415 case SUBREG: | |
416 /* Eliminate FP subregister accesses in favor of the | |
417 actual FP register in use. */ | |
418 { | |
419 rtx subreg; | |
420 if (FP_REG_P (subreg = SUBREG_REG (*pat))) | |
421 { | |
422 int regno_off = subreg_regno_offset (REGNO (subreg), | |
423 GET_MODE (subreg), | |
424 SUBREG_BYTE (*pat), | |
425 GET_MODE (*pat)); | |
426 *pat = FP_MODE_REG (REGNO (subreg) + regno_off, | |
427 GET_MODE (subreg)); | |
428 return pat; | |
429 } | |
430 } | |
431 case FLOAT: | |
432 case FIX: | |
433 case FLOAT_EXTEND: | |
434 pat = & XEXP (*pat, 0); | |
435 break; | |
436 | |
437 case UNSPEC: | |
438 if (XINT (*pat, 1) == UNSPEC_TRUNC_NOOP) | |
439 pat = & XVECEXP (*pat, 0, 0); | |
440 return pat; | |
441 | |
442 case FLOAT_TRUNCATE: | |
443 if (!flag_unsafe_math_optimizations) | |
444 return pat; | |
445 pat = & XEXP (*pat, 0); | |
446 break; | |
447 | |
448 default: | |
449 return pat; | |
450 } | |
451 } | |
452 | |
453 /* Set if we find any malformed asms in a block. */ | |
454 static bool any_malformed_asm; | |
455 | |
456 /* There are many rules that an asm statement for stack-like regs must | |
457 follow. Those rules are explained at the top of this file: the rule | |
458 numbers below refer to that explanation. */ | |
459 | |
460 static int | |
461 check_asm_stack_operands (rtx insn) | |
462 { | |
463 int i; | |
464 int n_clobbers; | |
465 int malformed_asm = 0; | |
466 rtx body = PATTERN (insn); | |
467 | |
468 char reg_used_as_output[FIRST_PSEUDO_REGISTER]; | |
469 char implicitly_dies[FIRST_PSEUDO_REGISTER]; | |
470 int alt; | |
471 | |
472 rtx *clobber_reg = 0; | |
473 int n_inputs, n_outputs; | |
474 | |
475 /* Find out what the constraints require. If no constraint | |
476 alternative matches, this asm is malformed. */ | |
477 extract_insn (insn); | |
478 constrain_operands (1); | |
479 alt = which_alternative; | |
480 | |
481 preprocess_constraints (); | |
482 | |
483 n_inputs = get_asm_operand_n_inputs (body); | |
484 n_outputs = recog_data.n_operands - n_inputs; | |
485 | |
486 if (alt < 0) | |
487 { | |
488 malformed_asm = 1; | |
489 /* Avoid further trouble with this insn. */ | |
490 PATTERN (insn) = gen_rtx_USE (VOIDmode, const0_rtx); | |
491 return 0; | |
492 } | |
493 | |
494 /* Strip SUBREGs here to make the following code simpler. */ | |
495 for (i = 0; i < recog_data.n_operands; i++) | |
496 if (GET_CODE (recog_data.operand[i]) == SUBREG | |
497 && REG_P (SUBREG_REG (recog_data.operand[i]))) | |
498 recog_data.operand[i] = SUBREG_REG (recog_data.operand[i]); | |
499 | |
500 /* Set up CLOBBER_REG. */ | |
501 | |
502 n_clobbers = 0; | |
503 | |
504 if (GET_CODE (body) == PARALLEL) | |
505 { | |
506 clobber_reg = XALLOCAVEC (rtx, XVECLEN (body, 0)); | |
507 | |
508 for (i = 0; i < XVECLEN (body, 0); i++) | |
509 if (GET_CODE (XVECEXP (body, 0, i)) == CLOBBER) | |
510 { | |
511 rtx clobber = XVECEXP (body, 0, i); | |
512 rtx reg = XEXP (clobber, 0); | |
513 | |
514 if (GET_CODE (reg) == SUBREG && REG_P (SUBREG_REG (reg))) | |
515 reg = SUBREG_REG (reg); | |
516 | |
517 if (STACK_REG_P (reg)) | |
518 { | |
519 clobber_reg[n_clobbers] = reg; | |
520 n_clobbers++; | |
521 } | |
522 } | |
523 } | |
524 | |
525 /* Enforce rule #4: Output operands must specifically indicate which | |
526 reg an output appears in after an asm. "=f" is not allowed: the | |
527 operand constraints must select a class with a single reg. | |
528 | |
529 Also enforce rule #5: Output operands must start at the top of | |
530 the reg-stack: output operands may not "skip" a reg. */ | |
531 | |
532 memset (reg_used_as_output, 0, sizeof (reg_used_as_output)); | |
533 for (i = 0; i < n_outputs; i++) | |
534 if (STACK_REG_P (recog_data.operand[i])) | |
535 { | |
536 if (reg_class_size[(int) recog_op_alt[i][alt].cl] != 1) | |
537 { | |
538 error_for_asm (insn, "output constraint %d must specify a single register", i); | |
539 malformed_asm = 1; | |
540 } | |
541 else | |
542 { | |
543 int j; | |
544 | |
545 for (j = 0; j < n_clobbers; j++) | |
546 if (REGNO (recog_data.operand[i]) == REGNO (clobber_reg[j])) | |
547 { | |
548 error_for_asm (insn, "output constraint %d cannot be specified together with \"%s\" clobber", | |
549 i, reg_names [REGNO (clobber_reg[j])]); | |
550 malformed_asm = 1; | |
551 break; | |
552 } | |
553 if (j == n_clobbers) | |
554 reg_used_as_output[REGNO (recog_data.operand[i])] = 1; | |
555 } | |
556 } | |
557 | |
558 | |
559 /* Search for first non-popped reg. */ | |
560 for (i = FIRST_STACK_REG; i < LAST_STACK_REG + 1; i++) | |
561 if (! reg_used_as_output[i]) | |
562 break; | |
563 | |
564 /* If there are any other popped regs, that's an error. */ | |
565 for (; i < LAST_STACK_REG + 1; i++) | |
566 if (reg_used_as_output[i]) | |
567 break; | |
568 | |
569 if (i != LAST_STACK_REG + 1) | |
570 { | |
571 error_for_asm (insn, "output regs must be grouped at top of stack"); | |
572 malformed_asm = 1; | |
573 } | |
574 | |
575 /* Enforce rule #2: All implicitly popped input regs must be closer | |
576 to the top of the reg-stack than any input that is not implicitly | |
577 popped. */ | |
578 | |
579 memset (implicitly_dies, 0, sizeof (implicitly_dies)); | |
580 for (i = n_outputs; i < n_outputs + n_inputs; i++) | |
581 if (STACK_REG_P (recog_data.operand[i])) | |
582 { | |
583 /* An input reg is implicitly popped if it is tied to an | |
584 output, or if there is a CLOBBER for it. */ | |
585 int j; | |
586 | |
587 for (j = 0; j < n_clobbers; j++) | |
588 if (operands_match_p (clobber_reg[j], recog_data.operand[i])) | |
589 break; | |
590 | |
591 if (j < n_clobbers || recog_op_alt[i][alt].matches >= 0) | |
592 implicitly_dies[REGNO (recog_data.operand[i])] = 1; | |
593 } | |
594 | |
595 /* Search for first non-popped reg. */ | |
596 for (i = FIRST_STACK_REG; i < LAST_STACK_REG + 1; i++) | |
597 if (! implicitly_dies[i]) | |
598 break; | |
599 | |
600 /* If there are any other popped regs, that's an error. */ | |
601 for (; i < LAST_STACK_REG + 1; i++) | |
602 if (implicitly_dies[i]) | |
603 break; | |
604 | |
605 if (i != LAST_STACK_REG + 1) | |
606 { | |
607 error_for_asm (insn, | |
608 "implicitly popped regs must be grouped at top of stack"); | |
609 malformed_asm = 1; | |
610 } | |
611 | |
612 /* Enforce rule #3: If any input operand uses the "f" constraint, all | |
613 output constraints must use the "&" earlyclobber. | |
614 | |
615 ??? Detect this more deterministically by having constrain_asm_operands | |
616 record any earlyclobber. */ | |
617 | |
618 for (i = n_outputs; i < n_outputs + n_inputs; i++) | |
619 if (recog_op_alt[i][alt].matches == -1) | |
620 { | |
621 int j; | |
622 | |
623 for (j = 0; j < n_outputs; j++) | |
624 if (operands_match_p (recog_data.operand[j], recog_data.operand[i])) | |
625 { | |
626 error_for_asm (insn, | |
627 "output operand %d must use %<&%> constraint", j); | |
628 malformed_asm = 1; | |
629 } | |
630 } | |
631 | |
632 if (malformed_asm) | |
633 { | |
634 /* Avoid further trouble with this insn. */ | |
635 PATTERN (insn) = gen_rtx_USE (VOIDmode, const0_rtx); | |
636 any_malformed_asm = true; | |
637 return 0; | |
638 } | |
639 | |
640 return 1; | |
641 } | |
642 | |
643 /* Calculate the number of inputs and outputs in BODY, an | |
644 asm_operands. N_OPERANDS is the total number of operands, and | |
645 N_INPUTS and N_OUTPUTS are pointers to ints into which the results are | |
646 placed. */ | |
647 | |
648 static int | |
649 get_asm_operand_n_inputs (rtx body) | |
650 { | |
651 switch (GET_CODE (body)) | |
652 { | |
653 case SET: | |
654 gcc_assert (GET_CODE (SET_SRC (body)) == ASM_OPERANDS); | |
655 return ASM_OPERANDS_INPUT_LENGTH (SET_SRC (body)); | |
656 | |
657 case ASM_OPERANDS: | |
658 return ASM_OPERANDS_INPUT_LENGTH (body); | |
659 | |
660 case PARALLEL: | |
661 return get_asm_operand_n_inputs (XVECEXP (body, 0, 0)); | |
662 | |
663 default: | |
664 gcc_unreachable (); | |
665 } | |
666 } | |
667 | |
668 /* If current function returns its result in an fp stack register, | |
669 return the REG. Otherwise, return 0. */ | |
670 | |
671 static rtx | |
672 stack_result (tree decl) | |
673 { | |
674 rtx result; | |
675 | |
676 /* If the value is supposed to be returned in memory, then clearly | |
677 it is not returned in a stack register. */ | |
678 if (aggregate_value_p (DECL_RESULT (decl), decl)) | |
679 return 0; | |
680 | |
681 result = DECL_RTL_IF_SET (DECL_RESULT (decl)); | |
682 if (result != 0) | |
683 result = targetm.calls.function_value (TREE_TYPE (DECL_RESULT (decl)), | |
684 decl, true); | |
685 | |
686 return result != 0 && STACK_REG_P (result) ? result : 0; | |
687 } | |
688 | |
689 | |
690 /* | |
691 * This section deals with stack register substitution, and forms the second | |
692 * pass over the RTL. | |
693 */ | |
694 | |
695 /* Replace REG, which is a pointer to a stack reg RTX, with an RTX for | |
696 the desired hard REGNO. */ | |
697 | |
698 static void | |
699 replace_reg (rtx *reg, int regno) | |
700 { | |
701 gcc_assert (IN_RANGE (regno, FIRST_STACK_REG, LAST_STACK_REG)); | |
702 gcc_assert (STACK_REG_P (*reg)); | |
703 | |
704 gcc_assert (SCALAR_FLOAT_MODE_P (GET_MODE (*reg)) | |
705 || GET_MODE_CLASS (GET_MODE (*reg)) == MODE_COMPLEX_FLOAT); | |
706 | |
707 *reg = FP_MODE_REG (regno, GET_MODE (*reg)); | |
708 } | |
709 | |
710 /* Remove a note of type NOTE, which must be found, for register | |
711 number REGNO from INSN. Remove only one such note. */ | |
712 | |
713 static void | |
714 remove_regno_note (rtx insn, enum reg_note note, unsigned int regno) | |
715 { | |
716 rtx *note_link, this_rtx; | |
717 | |
718 note_link = ®_NOTES (insn); | |
719 for (this_rtx = *note_link; this_rtx; this_rtx = XEXP (this_rtx, 1)) | |
720 if (REG_NOTE_KIND (this_rtx) == note | |
721 && REG_P (XEXP (this_rtx, 0)) && REGNO (XEXP (this_rtx, 0)) == regno) | |
722 { | |
723 *note_link = XEXP (this_rtx, 1); | |
724 return; | |
725 } | |
726 else | |
727 note_link = &XEXP (this_rtx, 1); | |
728 | |
729 gcc_unreachable (); | |
730 } | |
731 | |
732 /* Find the hard register number of virtual register REG in REGSTACK. | |
733 The hard register number is relative to the top of the stack. -1 is | |
734 returned if the register is not found. */ | |
735 | |
736 static int | |
737 get_hard_regnum (stack regstack, rtx reg) | |
738 { | |
739 int i; | |
740 | |
741 gcc_assert (STACK_REG_P (reg)); | |
742 | |
743 for (i = regstack->top; i >= 0; i--) | |
744 if (regstack->reg[i] == REGNO (reg)) | |
745 break; | |
746 | |
747 return i >= 0 ? (FIRST_STACK_REG + regstack->top - i) : -1; | |
748 } | |
749 | |
750 /* Emit an insn to pop virtual register REG before or after INSN. | |
751 REGSTACK is the stack state after INSN and is updated to reflect this | |
752 pop. WHEN is either emit_insn_before or emit_insn_after. A pop insn | |
753 is represented as a SET whose destination is the register to be popped | |
754 and source is the top of stack. A death note for the top of stack | |
755 cases the movdf pattern to pop. */ | |
756 | |
757 static rtx | |
758 emit_pop_insn (rtx insn, stack regstack, rtx reg, enum emit_where where) | |
759 { | |
760 rtx pop_insn, pop_rtx; | |
761 int hard_regno; | |
762 | |
763 /* For complex types take care to pop both halves. These may survive in | |
764 CLOBBER and USE expressions. */ | |
765 if (COMPLEX_MODE_P (GET_MODE (reg))) | |
766 { | |
767 rtx reg1 = FP_MODE_REG (REGNO (reg), DFmode); | |
768 rtx reg2 = FP_MODE_REG (REGNO (reg) + 1, DFmode); | |
769 | |
770 pop_insn = NULL_RTX; | |
771 if (get_hard_regnum (regstack, reg1) >= 0) | |
772 pop_insn = emit_pop_insn (insn, regstack, reg1, where); | |
773 if (get_hard_regnum (regstack, reg2) >= 0) | |
774 pop_insn = emit_pop_insn (insn, regstack, reg2, where); | |
775 gcc_assert (pop_insn); | |
776 return pop_insn; | |
777 } | |
778 | |
779 hard_regno = get_hard_regnum (regstack, reg); | |
780 | |
781 gcc_assert (hard_regno >= FIRST_STACK_REG); | |
782 | |
783 pop_rtx = gen_rtx_SET (VOIDmode, FP_MODE_REG (hard_regno, DFmode), | |
784 FP_MODE_REG (FIRST_STACK_REG, DFmode)); | |
785 | |
786 if (where == EMIT_AFTER) | |
787 pop_insn = emit_insn_after (pop_rtx, insn); | |
788 else | |
789 pop_insn = emit_insn_before (pop_rtx, insn); | |
790 | |
791 add_reg_note (pop_insn, REG_DEAD, FP_MODE_REG (FIRST_STACK_REG, DFmode)); | |
792 | |
793 regstack->reg[regstack->top - (hard_regno - FIRST_STACK_REG)] | |
794 = regstack->reg[regstack->top]; | |
795 regstack->top -= 1; | |
796 CLEAR_HARD_REG_BIT (regstack->reg_set, REGNO (reg)); | |
797 | |
798 return pop_insn; | |
799 } | |
800 | |
801 /* Emit an insn before or after INSN to swap virtual register REG with | |
802 the top of stack. REGSTACK is the stack state before the swap, and | |
803 is updated to reflect the swap. A swap insn is represented as a | |
804 PARALLEL of two patterns: each pattern moves one reg to the other. | |
805 | |
806 If REG is already at the top of the stack, no insn is emitted. */ | |
807 | |
808 static void | |
809 emit_swap_insn (rtx insn, stack regstack, rtx reg) | |
810 { | |
811 int hard_regno; | |
812 rtx swap_rtx; | |
813 int tmp, other_reg; /* swap regno temps */ | |
814 rtx i1; /* the stack-reg insn prior to INSN */ | |
815 rtx i1set = NULL_RTX; /* the SET rtx within I1 */ | |
816 | |
817 hard_regno = get_hard_regnum (regstack, reg); | |
818 | |
819 if (hard_regno == FIRST_STACK_REG) | |
820 return; | |
821 if (hard_regno == -1) | |
822 { | |
823 /* Something failed if the register wasn't on the stack. If we had | |
824 malformed asms, we zapped the instruction itself, but that didn't | |
825 produce the same pattern of register sets as before. To prevent | |
826 further failure, adjust REGSTACK to include REG at TOP. */ | |
827 gcc_assert (any_malformed_asm); | |
828 regstack->reg[++regstack->top] = REGNO (reg); | |
829 return; | |
830 } | |
831 gcc_assert (hard_regno >= FIRST_STACK_REG); | |
832 | |
833 other_reg = regstack->top - (hard_regno - FIRST_STACK_REG); | |
834 | |
835 tmp = regstack->reg[other_reg]; | |
836 regstack->reg[other_reg] = regstack->reg[regstack->top]; | |
837 regstack->reg[regstack->top] = tmp; | |
838 | |
839 /* Find the previous insn involving stack regs, but don't pass a | |
840 block boundary. */ | |
841 i1 = NULL; | |
842 if (current_block && insn != BB_HEAD (current_block)) | |
843 { | |
844 rtx tmp = PREV_INSN (insn); | |
845 rtx limit = PREV_INSN (BB_HEAD (current_block)); | |
846 while (tmp != limit) | |
847 { | |
848 if (LABEL_P (tmp) | |
849 || CALL_P (tmp) | |
850 || NOTE_INSN_BASIC_BLOCK_P (tmp) | |
851 || (NONJUMP_INSN_P (tmp) | |
852 && stack_regs_mentioned (tmp))) | |
853 { | |
854 i1 = tmp; | |
855 break; | |
856 } | |
857 tmp = PREV_INSN (tmp); | |
858 } | |
859 } | |
860 | |
861 if (i1 != NULL_RTX | |
862 && (i1set = single_set (i1)) != NULL_RTX) | |
863 { | |
864 rtx i1src = *get_true_reg (&SET_SRC (i1set)); | |
865 rtx i1dest = *get_true_reg (&SET_DEST (i1set)); | |
866 | |
867 /* If the previous register stack push was from the reg we are to | |
868 swap with, omit the swap. */ | |
869 | |
870 if (REG_P (i1dest) && REGNO (i1dest) == FIRST_STACK_REG | |
871 && REG_P (i1src) | |
872 && REGNO (i1src) == (unsigned) hard_regno - 1 | |
873 && find_regno_note (i1, REG_DEAD, FIRST_STACK_REG) == NULL_RTX) | |
874 return; | |
875 | |
876 /* If the previous insn wrote to the reg we are to swap with, | |
877 omit the swap. */ | |
878 | |
879 if (REG_P (i1dest) && REGNO (i1dest) == (unsigned) hard_regno | |
880 && REG_P (i1src) && REGNO (i1src) == FIRST_STACK_REG | |
881 && find_regno_note (i1, REG_DEAD, FIRST_STACK_REG) == NULL_RTX) | |
882 return; | |
883 } | |
884 | |
885 /* Avoid emitting the swap if this is the first register stack insn | |
886 of the current_block. Instead update the current_block's stack_in | |
887 and let compensate edges take care of this for us. */ | |
888 if (current_block && starting_stack_p) | |
889 { | |
890 BLOCK_INFO (current_block)->stack_in = *regstack; | |
891 starting_stack_p = false; | |
892 return; | |
893 } | |
894 | |
895 swap_rtx = gen_swapxf (FP_MODE_REG (hard_regno, XFmode), | |
896 FP_MODE_REG (FIRST_STACK_REG, XFmode)); | |
897 | |
898 if (i1) | |
899 emit_insn_after (swap_rtx, i1); | |
900 else if (current_block) | |
901 emit_insn_before (swap_rtx, BB_HEAD (current_block)); | |
902 else | |
903 emit_insn_before (swap_rtx, insn); | |
904 } | |
905 | |
906 /* Emit an insns before INSN to swap virtual register SRC1 with | |
907 the top of stack and virtual register SRC2 with second stack | |
908 slot. REGSTACK is the stack state before the swaps, and | |
909 is updated to reflect the swaps. A swap insn is represented as a | |
910 PARALLEL of two patterns: each pattern moves one reg to the other. | |
911 | |
912 If SRC1 and/or SRC2 are already at the right place, no swap insn | |
913 is emitted. */ | |
914 | |
915 static void | |
916 swap_to_top (rtx insn, stack regstack, rtx src1, rtx src2) | |
917 { | |
918 struct stack_def temp_stack; | |
919 int regno, j, k, temp; | |
920 | |
921 temp_stack = *regstack; | |
922 | |
923 /* Place operand 1 at the top of stack. */ | |
924 regno = get_hard_regnum (&temp_stack, src1); | |
925 gcc_assert (regno >= 0); | |
926 if (regno != FIRST_STACK_REG) | |
927 { | |
928 k = temp_stack.top - (regno - FIRST_STACK_REG); | |
929 j = temp_stack.top; | |
930 | |
931 temp = temp_stack.reg[k]; | |
932 temp_stack.reg[k] = temp_stack.reg[j]; | |
933 temp_stack.reg[j] = temp; | |
934 } | |
935 | |
936 /* Place operand 2 next on the stack. */ | |
937 regno = get_hard_regnum (&temp_stack, src2); | |
938 gcc_assert (regno >= 0); | |
939 if (regno != FIRST_STACK_REG + 1) | |
940 { | |
941 k = temp_stack.top - (regno - FIRST_STACK_REG); | |
942 j = temp_stack.top - 1; | |
943 | |
944 temp = temp_stack.reg[k]; | |
945 temp_stack.reg[k] = temp_stack.reg[j]; | |
946 temp_stack.reg[j] = temp; | |
947 } | |
948 | |
949 change_stack (insn, regstack, &temp_stack, EMIT_BEFORE); | |
950 } | |
951 | |
952 /* Handle a move to or from a stack register in PAT, which is in INSN. | |
953 REGSTACK is the current stack. Return whether a control flow insn | |
954 was deleted in the process. */ | |
955 | |
956 static bool | |
957 move_for_stack_reg (rtx insn, stack regstack, rtx pat) | |
958 { | |
959 rtx *psrc = get_true_reg (&SET_SRC (pat)); | |
960 rtx *pdest = get_true_reg (&SET_DEST (pat)); | |
961 rtx src, dest; | |
962 rtx note; | |
963 bool control_flow_insn_deleted = false; | |
964 | |
965 src = *psrc; dest = *pdest; | |
966 | |
967 if (STACK_REG_P (src) && STACK_REG_P (dest)) | |
968 { | |
969 /* Write from one stack reg to another. If SRC dies here, then | |
970 just change the register mapping and delete the insn. */ | |
971 | |
972 note = find_regno_note (insn, REG_DEAD, REGNO (src)); | |
973 if (note) | |
974 { | |
975 int i; | |
976 | |
977 /* If this is a no-op move, there must not be a REG_DEAD note. */ | |
978 gcc_assert (REGNO (src) != REGNO (dest)); | |
979 | |
980 for (i = regstack->top; i >= 0; i--) | |
981 if (regstack->reg[i] == REGNO (src)) | |
982 break; | |
983 | |
984 /* The destination must be dead, or life analysis is borked. */ | |
985 gcc_assert (get_hard_regnum (regstack, dest) < FIRST_STACK_REG); | |
986 | |
987 /* If the source is not live, this is yet another case of | |
988 uninitialized variables. Load up a NaN instead. */ | |
989 if (i < 0) | |
990 return move_nan_for_stack_reg (insn, regstack, dest); | |
991 | |
992 /* It is possible that the dest is unused after this insn. | |
993 If so, just pop the src. */ | |
994 | |
995 if (find_regno_note (insn, REG_UNUSED, REGNO (dest))) | |
996 emit_pop_insn (insn, regstack, src, EMIT_AFTER); | |
997 else | |
998 { | |
999 regstack->reg[i] = REGNO (dest); | |
1000 SET_HARD_REG_BIT (regstack->reg_set, REGNO (dest)); | |
1001 CLEAR_HARD_REG_BIT (regstack->reg_set, REGNO (src)); | |
1002 } | |
1003 | |
1004 control_flow_insn_deleted |= control_flow_insn_p (insn); | |
1005 delete_insn (insn); | |
1006 return control_flow_insn_deleted; | |
1007 } | |
1008 | |
1009 /* The source reg does not die. */ | |
1010 | |
1011 /* If this appears to be a no-op move, delete it, or else it | |
1012 will confuse the machine description output patterns. But if | |
1013 it is REG_UNUSED, we must pop the reg now, as per-insn processing | |
1014 for REG_UNUSED will not work for deleted insns. */ | |
1015 | |
1016 if (REGNO (src) == REGNO (dest)) | |
1017 { | |
1018 if (find_regno_note (insn, REG_UNUSED, REGNO (dest))) | |
1019 emit_pop_insn (insn, regstack, dest, EMIT_AFTER); | |
1020 | |
1021 control_flow_insn_deleted |= control_flow_insn_p (insn); | |
1022 delete_insn (insn); | |
1023 return control_flow_insn_deleted; | |
1024 } | |
1025 | |
1026 /* The destination ought to be dead. */ | |
1027 gcc_assert (get_hard_regnum (regstack, dest) < FIRST_STACK_REG); | |
1028 | |
1029 replace_reg (psrc, get_hard_regnum (regstack, src)); | |
1030 | |
1031 regstack->reg[++regstack->top] = REGNO (dest); | |
1032 SET_HARD_REG_BIT (regstack->reg_set, REGNO (dest)); | |
1033 replace_reg (pdest, FIRST_STACK_REG); | |
1034 } | |
1035 else if (STACK_REG_P (src)) | |
1036 { | |
1037 /* Save from a stack reg to MEM, or possibly integer reg. Since | |
1038 only top of stack may be saved, emit an exchange first if | |
1039 needs be. */ | |
1040 | |
1041 emit_swap_insn (insn, regstack, src); | |
1042 | |
1043 note = find_regno_note (insn, REG_DEAD, REGNO (src)); | |
1044 if (note) | |
1045 { | |
1046 replace_reg (&XEXP (note, 0), FIRST_STACK_REG); | |
1047 regstack->top--; | |
1048 CLEAR_HARD_REG_BIT (regstack->reg_set, REGNO (src)); | |
1049 } | |
1050 else if ((GET_MODE (src) == XFmode) | |
1051 && regstack->top < REG_STACK_SIZE - 1) | |
1052 { | |
1053 /* A 387 cannot write an XFmode value to a MEM without | |
1054 clobbering the source reg. The output code can handle | |
1055 this by reading back the value from the MEM. | |
1056 But it is more efficient to use a temp register if one is | |
1057 available. Push the source value here if the register | |
1058 stack is not full, and then write the value to memory via | |
1059 a pop. */ | |
1060 rtx push_rtx; | |
1061 rtx top_stack_reg = FP_MODE_REG (FIRST_STACK_REG, GET_MODE (src)); | |
1062 | |
1063 push_rtx = gen_movxf (top_stack_reg, top_stack_reg); | |
1064 emit_insn_before (push_rtx, insn); | |
1065 add_reg_note (insn, REG_DEAD, top_stack_reg); | |
1066 } | |
1067 | |
1068 replace_reg (psrc, FIRST_STACK_REG); | |
1069 } | |
1070 else | |
1071 { | |
1072 rtx pat = PATTERN (insn); | |
1073 | |
1074 gcc_assert (STACK_REG_P (dest)); | |
1075 | |
1076 /* Load from MEM, or possibly integer REG or constant, into the | |
1077 stack regs. The actual target is always the top of the | |
1078 stack. The stack mapping is changed to reflect that DEST is | |
1079 now at top of stack. */ | |
1080 | |
1081 /* The destination ought to be dead. However, there is a | |
1082 special case with i387 UNSPEC_TAN, where destination is live | |
1083 (an argument to fptan) but inherent load of 1.0 is modelled | |
1084 as a load from a constant. */ | |
1085 if (GET_CODE (pat) == PARALLEL | |
1086 && XVECLEN (pat, 0) == 2 | |
1087 && GET_CODE (XVECEXP (pat, 0, 1)) == SET | |
1088 && GET_CODE (SET_SRC (XVECEXP (pat, 0, 1))) == UNSPEC | |
1089 && XINT (SET_SRC (XVECEXP (pat, 0, 1)), 1) == UNSPEC_TAN) | |
1090 emit_swap_insn (insn, regstack, dest); | |
1091 else | |
1092 gcc_assert (get_hard_regnum (regstack, dest) < FIRST_STACK_REG); | |
1093 | |
1094 gcc_assert (regstack->top < REG_STACK_SIZE); | |
1095 | |
1096 regstack->reg[++regstack->top] = REGNO (dest); | |
1097 SET_HARD_REG_BIT (regstack->reg_set, REGNO (dest)); | |
1098 replace_reg (pdest, FIRST_STACK_REG); | |
1099 } | |
1100 | |
1101 return control_flow_insn_deleted; | |
1102 } | |
1103 | |
1104 /* A helper function which replaces INSN with a pattern that loads up | |
1105 a NaN into DEST, then invokes move_for_stack_reg. */ | |
1106 | |
1107 static bool | |
1108 move_nan_for_stack_reg (rtx insn, stack regstack, rtx dest) | |
1109 { | |
1110 rtx pat; | |
1111 | |
1112 dest = FP_MODE_REG (REGNO (dest), SFmode); | |
1113 pat = gen_rtx_SET (VOIDmode, dest, not_a_num); | |
1114 PATTERN (insn) = pat; | |
1115 INSN_CODE (insn) = -1; | |
1116 | |
1117 return move_for_stack_reg (insn, regstack, pat); | |
1118 } | |
1119 | |
1120 /* Swap the condition on a branch, if there is one. Return true if we | |
1121 found a condition to swap. False if the condition was not used as | |
1122 such. */ | |
1123 | |
1124 static int | |
1125 swap_rtx_condition_1 (rtx pat) | |
1126 { | |
1127 const char *fmt; | |
1128 int i, r = 0; | |
1129 | |
1130 if (COMPARISON_P (pat)) | |
1131 { | |
1132 PUT_CODE (pat, swap_condition (GET_CODE (pat))); | |
1133 r = 1; | |
1134 } | |
1135 else | |
1136 { | |
1137 fmt = GET_RTX_FORMAT (GET_CODE (pat)); | |
1138 for (i = GET_RTX_LENGTH (GET_CODE (pat)) - 1; i >= 0; i--) | |
1139 { | |
1140 if (fmt[i] == 'E') | |
1141 { | |
1142 int j; | |
1143 | |
1144 for (j = XVECLEN (pat, i) - 1; j >= 0; j--) | |
1145 r |= swap_rtx_condition_1 (XVECEXP (pat, i, j)); | |
1146 } | |
1147 else if (fmt[i] == 'e') | |
1148 r |= swap_rtx_condition_1 (XEXP (pat, i)); | |
1149 } | |
1150 } | |
1151 | |
1152 return r; | |
1153 } | |
1154 | |
1155 static int | |
1156 swap_rtx_condition (rtx insn) | |
1157 { | |
1158 rtx pat = PATTERN (insn); | |
1159 | |
1160 /* We're looking for a single set to cc0 or an HImode temporary. */ | |
1161 | |
1162 if (GET_CODE (pat) == SET | |
1163 && REG_P (SET_DEST (pat)) | |
1164 && REGNO (SET_DEST (pat)) == FLAGS_REG) | |
1165 { | |
1166 insn = next_flags_user (insn); | |
1167 if (insn == NULL_RTX) | |
1168 return 0; | |
1169 pat = PATTERN (insn); | |
1170 } | |
1171 | |
1172 /* See if this is, or ends in, a fnstsw. If so, we're not doing anything | |
1173 with the cc value right now. We may be able to search for one | |
1174 though. */ | |
1175 | |
1176 if (GET_CODE (pat) == SET | |
1177 && GET_CODE (SET_SRC (pat)) == UNSPEC | |
1178 && XINT (SET_SRC (pat), 1) == UNSPEC_FNSTSW) | |
1179 { | |
1180 rtx dest = SET_DEST (pat); | |
1181 | |
1182 /* Search forward looking for the first use of this value. | |
1183 Stop at block boundaries. */ | |
1184 while (insn != BB_END (current_block)) | |
1185 { | |
1186 insn = NEXT_INSN (insn); | |
1187 if (INSN_P (insn) && reg_mentioned_p (dest, insn)) | |
1188 break; | |
1189 if (CALL_P (insn)) | |
1190 return 0; | |
1191 } | |
1192 | |
1193 /* We haven't found it. */ | |
1194 if (insn == BB_END (current_block)) | |
1195 return 0; | |
1196 | |
1197 /* So we've found the insn using this value. If it is anything | |
1198 other than sahf or the value does not die (meaning we'd have | |
1199 to search further), then we must give up. */ | |
1200 pat = PATTERN (insn); | |
1201 if (GET_CODE (pat) != SET | |
1202 || GET_CODE (SET_SRC (pat)) != UNSPEC | |
1203 || XINT (SET_SRC (pat), 1) != UNSPEC_SAHF | |
1204 || ! dead_or_set_p (insn, dest)) | |
1205 return 0; | |
1206 | |
1207 /* Now we are prepared to handle this as a normal cc0 setter. */ | |
1208 insn = next_flags_user (insn); | |
1209 if (insn == NULL_RTX) | |
1210 return 0; | |
1211 pat = PATTERN (insn); | |
1212 } | |
1213 | |
1214 if (swap_rtx_condition_1 (pat)) | |
1215 { | |
1216 int fail = 0; | |
1217 INSN_CODE (insn) = -1; | |
1218 if (recog_memoized (insn) == -1) | |
1219 fail = 1; | |
1220 /* In case the flags don't die here, recurse to try fix | |
1221 following user too. */ | |
1222 else if (! dead_or_set_p (insn, ix86_flags_rtx)) | |
1223 { | |
1224 insn = next_flags_user (insn); | |
1225 if (!insn || !swap_rtx_condition (insn)) | |
1226 fail = 1; | |
1227 } | |
1228 if (fail) | |
1229 { | |
1230 swap_rtx_condition_1 (pat); | |
1231 return 0; | |
1232 } | |
1233 return 1; | |
1234 } | |
1235 return 0; | |
1236 } | |
1237 | |
1238 /* Handle a comparison. Special care needs to be taken to avoid | |
1239 causing comparisons that a 387 cannot do correctly, such as EQ. | |
1240 | |
1241 Also, a pop insn may need to be emitted. The 387 does have an | |
1242 `fcompp' insn that can pop two regs, but it is sometimes too expensive | |
1243 to do this - a `fcomp' followed by a `fstpl %st(0)' may be easier to | |
1244 set up. */ | |
1245 | |
1246 static void | |
1247 compare_for_stack_reg (rtx insn, stack regstack, rtx pat_src) | |
1248 { | |
1249 rtx *src1, *src2; | |
1250 rtx src1_note, src2_note; | |
1251 | |
1252 src1 = get_true_reg (&XEXP (pat_src, 0)); | |
1253 src2 = get_true_reg (&XEXP (pat_src, 1)); | |
1254 | |
1255 /* ??? If fxch turns out to be cheaper than fstp, give priority to | |
1256 registers that die in this insn - move those to stack top first. */ | |
1257 if ((! STACK_REG_P (*src1) | |
1258 || (STACK_REG_P (*src2) | |
1259 && get_hard_regnum (regstack, *src2) == FIRST_STACK_REG)) | |
1260 && swap_rtx_condition (insn)) | |
1261 { | |
1262 rtx temp; | |
1263 temp = XEXP (pat_src, 0); | |
1264 XEXP (pat_src, 0) = XEXP (pat_src, 1); | |
1265 XEXP (pat_src, 1) = temp; | |
1266 | |
1267 src1 = get_true_reg (&XEXP (pat_src, 0)); | |
1268 src2 = get_true_reg (&XEXP (pat_src, 1)); | |
1269 | |
1270 INSN_CODE (insn) = -1; | |
1271 } | |
1272 | |
1273 /* We will fix any death note later. */ | |
1274 | |
1275 src1_note = find_regno_note (insn, REG_DEAD, REGNO (*src1)); | |
1276 | |
1277 if (STACK_REG_P (*src2)) | |
1278 src2_note = find_regno_note (insn, REG_DEAD, REGNO (*src2)); | |
1279 else | |
1280 src2_note = NULL_RTX; | |
1281 | |
1282 emit_swap_insn (insn, regstack, *src1); | |
1283 | |
1284 replace_reg (src1, FIRST_STACK_REG); | |
1285 | |
1286 if (STACK_REG_P (*src2)) | |
1287 replace_reg (src2, get_hard_regnum (regstack, *src2)); | |
1288 | |
1289 if (src1_note) | |
1290 { | |
1291 pop_stack (regstack, REGNO (XEXP (src1_note, 0))); | |
1292 replace_reg (&XEXP (src1_note, 0), FIRST_STACK_REG); | |
1293 } | |
1294 | |
1295 /* If the second operand dies, handle that. But if the operands are | |
1296 the same stack register, don't bother, because only one death is | |
1297 needed, and it was just handled. */ | |
1298 | |
1299 if (src2_note | |
1300 && ! (STACK_REG_P (*src1) && STACK_REG_P (*src2) | |
1301 && REGNO (*src1) == REGNO (*src2))) | |
1302 { | |
1303 /* As a special case, two regs may die in this insn if src2 is | |
1304 next to top of stack and the top of stack also dies. Since | |
1305 we have already popped src1, "next to top of stack" is really | |
1306 at top (FIRST_STACK_REG) now. */ | |
1307 | |
1308 if (get_hard_regnum (regstack, XEXP (src2_note, 0)) == FIRST_STACK_REG | |
1309 && src1_note) | |
1310 { | |
1311 pop_stack (regstack, REGNO (XEXP (src2_note, 0))); | |
1312 replace_reg (&XEXP (src2_note, 0), FIRST_STACK_REG + 1); | |
1313 } | |
1314 else | |
1315 { | |
1316 /* The 386 can only represent death of the first operand in | |
1317 the case handled above. In all other cases, emit a separate | |
1318 pop and remove the death note from here. */ | |
1319 | |
1320 /* link_cc0_insns (insn); */ | |
1321 | |
1322 remove_regno_note (insn, REG_DEAD, REGNO (XEXP (src2_note, 0))); | |
1323 | |
1324 emit_pop_insn (insn, regstack, XEXP (src2_note, 0), | |
1325 EMIT_AFTER); | |
1326 } | |
1327 } | |
1328 } | |
1329 | |
1330 /* Substitute new registers in PAT, which is part of INSN. REGSTACK | |
1331 is the current register layout. Return whether a control flow insn | |
1332 was deleted in the process. */ | |
1333 | |
1334 static bool | |
1335 subst_stack_regs_pat (rtx insn, stack regstack, rtx pat) | |
1336 { | |
1337 rtx *dest, *src; | |
1338 bool control_flow_insn_deleted = false; | |
1339 | |
1340 switch (GET_CODE (pat)) | |
1341 { | |
1342 case USE: | |
1343 /* Deaths in USE insns can happen in non optimizing compilation. | |
1344 Handle them by popping the dying register. */ | |
1345 src = get_true_reg (&XEXP (pat, 0)); | |
1346 if (STACK_REG_P (*src) | |
1347 && find_regno_note (insn, REG_DEAD, REGNO (*src))) | |
1348 { | |
1349 /* USEs are ignored for liveness information so USEs of dead | |
1350 register might happen. */ | |
1351 if (TEST_HARD_REG_BIT (regstack->reg_set, REGNO (*src))) | |
1352 emit_pop_insn (insn, regstack, *src, EMIT_AFTER); | |
1353 return control_flow_insn_deleted; | |
1354 } | |
1355 /* Uninitialized USE might happen for functions returning uninitialized | |
1356 value. We will properly initialize the USE on the edge to EXIT_BLOCK, | |
1357 so it is safe to ignore the use here. This is consistent with behavior | |
1358 of dataflow analyzer that ignores USE too. (This also imply that | |
1359 forcibly initializing the register to NaN here would lead to ICE later, | |
1360 since the REG_DEAD notes are not issued.) */ | |
1361 break; | |
1362 | |
1363 case CLOBBER: | |
1364 { | |
1365 rtx note; | |
1366 | |
1367 dest = get_true_reg (&XEXP (pat, 0)); | |
1368 if (STACK_REG_P (*dest)) | |
1369 { | |
1370 note = find_reg_note (insn, REG_DEAD, *dest); | |
1371 | |
1372 if (pat != PATTERN (insn)) | |
1373 { | |
1374 /* The fix_truncdi_1 pattern wants to be able to allocate | |
1375 its own scratch register. It does this by clobbering | |
1376 an fp reg so that it is assured of an empty reg-stack | |
1377 register. If the register is live, kill it now. | |
1378 Remove the DEAD/UNUSED note so we don't try to kill it | |
1379 later too. */ | |
1380 | |
1381 if (note) | |
1382 emit_pop_insn (insn, regstack, *dest, EMIT_BEFORE); | |
1383 else | |
1384 { | |
1385 note = find_reg_note (insn, REG_UNUSED, *dest); | |
1386 gcc_assert (note); | |
1387 } | |
1388 remove_note (insn, note); | |
1389 replace_reg (dest, FIRST_STACK_REG + 1); | |
1390 } | |
1391 else | |
1392 { | |
1393 /* A top-level clobber with no REG_DEAD, and no hard-regnum | |
1394 indicates an uninitialized value. Because reload removed | |
1395 all other clobbers, this must be due to a function | |
1396 returning without a value. Load up a NaN. */ | |
1397 | |
1398 if (!note) | |
1399 { | |
1400 rtx t = *dest; | |
1401 if (COMPLEX_MODE_P (GET_MODE (t))) | |
1402 { | |
1403 rtx u = FP_MODE_REG (REGNO (t) + 1, SFmode); | |
1404 if (get_hard_regnum (regstack, u) == -1) | |
1405 { | |
1406 rtx pat2 = gen_rtx_CLOBBER (VOIDmode, u); | |
1407 rtx insn2 = emit_insn_before (pat2, insn); | |
1408 control_flow_insn_deleted | |
1409 |= move_nan_for_stack_reg (insn2, regstack, u); | |
1410 } | |
1411 } | |
1412 if (get_hard_regnum (regstack, t) == -1) | |
1413 control_flow_insn_deleted | |
1414 |= move_nan_for_stack_reg (insn, regstack, t); | |
1415 } | |
1416 } | |
1417 } | |
1418 break; | |
1419 } | |
1420 | |
1421 case SET: | |
1422 { | |
1423 rtx *src1 = (rtx *) 0, *src2; | |
1424 rtx src1_note, src2_note; | |
1425 rtx pat_src; | |
1426 | |
1427 dest = get_true_reg (&SET_DEST (pat)); | |
1428 src = get_true_reg (&SET_SRC (pat)); | |
1429 pat_src = SET_SRC (pat); | |
1430 | |
1431 /* See if this is a `movM' pattern, and handle elsewhere if so. */ | |
1432 if (STACK_REG_P (*src) | |
1433 || (STACK_REG_P (*dest) | |
1434 && (REG_P (*src) || MEM_P (*src) | |
1435 || GET_CODE (*src) == CONST_DOUBLE))) | |
1436 { | |
1437 control_flow_insn_deleted |= move_for_stack_reg (insn, regstack, pat); | |
1438 break; | |
1439 } | |
1440 | |
1441 switch (GET_CODE (pat_src)) | |
1442 { | |
1443 case COMPARE: | |
1444 compare_for_stack_reg (insn, regstack, pat_src); | |
1445 break; | |
1446 | |
1447 case CALL: | |
1448 { | |
1449 int count; | |
1450 for (count = hard_regno_nregs[REGNO (*dest)][GET_MODE (*dest)]; | |
1451 --count >= 0;) | |
1452 { | |
1453 regstack->reg[++regstack->top] = REGNO (*dest) + count; | |
1454 SET_HARD_REG_BIT (regstack->reg_set, REGNO (*dest) + count); | |
1455 } | |
1456 } | |
1457 replace_reg (dest, FIRST_STACK_REG); | |
1458 break; | |
1459 | |
1460 case REG: | |
1461 /* This is a `tstM2' case. */ | |
1462 gcc_assert (*dest == cc0_rtx); | |
1463 src1 = src; | |
1464 | |
1465 /* Fall through. */ | |
1466 | |
1467 case FLOAT_TRUNCATE: | |
1468 case SQRT: | |
1469 case ABS: | |
1470 case NEG: | |
1471 /* These insns only operate on the top of the stack. DEST might | |
1472 be cc0_rtx if we're processing a tstM pattern. Also, it's | |
1473 possible that the tstM case results in a REG_DEAD note on the | |
1474 source. */ | |
1475 | |
1476 if (src1 == 0) | |
1477 src1 = get_true_reg (&XEXP (pat_src, 0)); | |
1478 | |
1479 emit_swap_insn (insn, regstack, *src1); | |
1480 | |
1481 src1_note = find_regno_note (insn, REG_DEAD, REGNO (*src1)); | |
1482 | |
1483 if (STACK_REG_P (*dest)) | |
1484 replace_reg (dest, FIRST_STACK_REG); | |
1485 | |
1486 if (src1_note) | |
1487 { | |
1488 replace_reg (&XEXP (src1_note, 0), FIRST_STACK_REG); | |
1489 regstack->top--; | |
1490 CLEAR_HARD_REG_BIT (regstack->reg_set, REGNO (*src1)); | |
1491 } | |
1492 | |
1493 replace_reg (src1, FIRST_STACK_REG); | |
1494 break; | |
1495 | |
1496 case MINUS: | |
1497 case DIV: | |
1498 /* On i386, reversed forms of subM3 and divM3 exist for | |
1499 MODE_FLOAT, so the same code that works for addM3 and mulM3 | |
1500 can be used. */ | |
1501 case MULT: | |
1502 case PLUS: | |
1503 /* These insns can accept the top of stack as a destination | |
1504 from a stack reg or mem, or can use the top of stack as a | |
1505 source and some other stack register (possibly top of stack) | |
1506 as a destination. */ | |
1507 | |
1508 src1 = get_true_reg (&XEXP (pat_src, 0)); | |
1509 src2 = get_true_reg (&XEXP (pat_src, 1)); | |
1510 | |
1511 /* We will fix any death note later. */ | |
1512 | |
1513 if (STACK_REG_P (*src1)) | |
1514 src1_note = find_regno_note (insn, REG_DEAD, REGNO (*src1)); | |
1515 else | |
1516 src1_note = NULL_RTX; | |
1517 if (STACK_REG_P (*src2)) | |
1518 src2_note = find_regno_note (insn, REG_DEAD, REGNO (*src2)); | |
1519 else | |
1520 src2_note = NULL_RTX; | |
1521 | |
1522 /* If either operand is not a stack register, then the dest | |
1523 must be top of stack. */ | |
1524 | |
1525 if (! STACK_REG_P (*src1) || ! STACK_REG_P (*src2)) | |
1526 emit_swap_insn (insn, regstack, *dest); | |
1527 else | |
1528 { | |
1529 /* Both operands are REG. If neither operand is already | |
1530 at the top of stack, choose to make the one that is the | |
1531 dest the new top of stack. */ | |
1532 | |
1533 int src1_hard_regnum, src2_hard_regnum; | |
1534 | |
1535 src1_hard_regnum = get_hard_regnum (regstack, *src1); | |
1536 src2_hard_regnum = get_hard_regnum (regstack, *src2); | |
1537 | |
1538 /* If the source is not live, this is yet another case of | |
1539 uninitialized variables. Load up a NaN instead. */ | |
1540 if (src1_hard_regnum == -1) | |
1541 { | |
1542 rtx pat2 = gen_rtx_CLOBBER (VOIDmode, *src1); | |
1543 rtx insn2 = emit_insn_before (pat2, insn); | |
1544 control_flow_insn_deleted | |
1545 |= move_nan_for_stack_reg (insn2, regstack, *src1); | |
1546 } | |
1547 if (src2_hard_regnum == -1) | |
1548 { | |
1549 rtx pat2 = gen_rtx_CLOBBER (VOIDmode, *src2); | |
1550 rtx insn2 = emit_insn_before (pat2, insn); | |
1551 control_flow_insn_deleted | |
1552 |= move_nan_for_stack_reg (insn2, regstack, *src2); | |
1553 } | |
1554 | |
1555 if (src1_hard_regnum != FIRST_STACK_REG | |
1556 && src2_hard_regnum != FIRST_STACK_REG) | |
1557 emit_swap_insn (insn, regstack, *dest); | |
1558 } | |
1559 | |
1560 if (STACK_REG_P (*src1)) | |
1561 replace_reg (src1, get_hard_regnum (regstack, *src1)); | |
1562 if (STACK_REG_P (*src2)) | |
1563 replace_reg (src2, get_hard_regnum (regstack, *src2)); | |
1564 | |
1565 if (src1_note) | |
1566 { | |
1567 rtx src1_reg = XEXP (src1_note, 0); | |
1568 | |
1569 /* If the register that dies is at the top of stack, then | |
1570 the destination is somewhere else - merely substitute it. | |
1571 But if the reg that dies is not at top of stack, then | |
1572 move the top of stack to the dead reg, as though we had | |
1573 done the insn and then a store-with-pop. */ | |
1574 | |
1575 if (REGNO (src1_reg) == regstack->reg[regstack->top]) | |
1576 { | |
1577 SET_HARD_REG_BIT (regstack->reg_set, REGNO (*dest)); | |
1578 replace_reg (dest, get_hard_regnum (regstack, *dest)); | |
1579 } | |
1580 else | |
1581 { | |
1582 int regno = get_hard_regnum (regstack, src1_reg); | |
1583 | |
1584 SET_HARD_REG_BIT (regstack->reg_set, REGNO (*dest)); | |
1585 replace_reg (dest, regno); | |
1586 | |
1587 regstack->reg[regstack->top - (regno - FIRST_STACK_REG)] | |
1588 = regstack->reg[regstack->top]; | |
1589 } | |
1590 | |
1591 CLEAR_HARD_REG_BIT (regstack->reg_set, | |
1592 REGNO (XEXP (src1_note, 0))); | |
1593 replace_reg (&XEXP (src1_note, 0), FIRST_STACK_REG); | |
1594 regstack->top--; | |
1595 } | |
1596 else if (src2_note) | |
1597 { | |
1598 rtx src2_reg = XEXP (src2_note, 0); | |
1599 if (REGNO (src2_reg) == regstack->reg[regstack->top]) | |
1600 { | |
1601 SET_HARD_REG_BIT (regstack->reg_set, REGNO (*dest)); | |
1602 replace_reg (dest, get_hard_regnum (regstack, *dest)); | |
1603 } | |
1604 else | |
1605 { | |
1606 int regno = get_hard_regnum (regstack, src2_reg); | |
1607 | |
1608 SET_HARD_REG_BIT (regstack->reg_set, REGNO (*dest)); | |
1609 replace_reg (dest, regno); | |
1610 | |
1611 regstack->reg[regstack->top - (regno - FIRST_STACK_REG)] | |
1612 = regstack->reg[regstack->top]; | |
1613 } | |
1614 | |
1615 CLEAR_HARD_REG_BIT (regstack->reg_set, | |
1616 REGNO (XEXP (src2_note, 0))); | |
1617 replace_reg (&XEXP (src2_note, 0), FIRST_STACK_REG); | |
1618 regstack->top--; | |
1619 } | |
1620 else | |
1621 { | |
1622 SET_HARD_REG_BIT (regstack->reg_set, REGNO (*dest)); | |
1623 replace_reg (dest, get_hard_regnum (regstack, *dest)); | |
1624 } | |
1625 | |
1626 /* Keep operand 1 matching with destination. */ | |
1627 if (COMMUTATIVE_ARITH_P (pat_src) | |
1628 && REG_P (*src1) && REG_P (*src2) | |
1629 && REGNO (*src1) != REGNO (*dest)) | |
1630 { | |
1631 int tmp = REGNO (*src1); | |
1632 replace_reg (src1, REGNO (*src2)); | |
1633 replace_reg (src2, tmp); | |
1634 } | |
1635 break; | |
1636 | |
1637 case UNSPEC: | |
1638 switch (XINT (pat_src, 1)) | |
1639 { | |
1640 case UNSPEC_FIST: | |
1641 | |
1642 case UNSPEC_FIST_FLOOR: | |
1643 case UNSPEC_FIST_CEIL: | |
1644 | |
1645 /* These insns only operate on the top of the stack. */ | |
1646 | |
1647 src1 = get_true_reg (&XVECEXP (pat_src, 0, 0)); | |
1648 emit_swap_insn (insn, regstack, *src1); | |
1649 | |
1650 src1_note = find_regno_note (insn, REG_DEAD, REGNO (*src1)); | |
1651 | |
1652 if (STACK_REG_P (*dest)) | |
1653 replace_reg (dest, FIRST_STACK_REG); | |
1654 | |
1655 if (src1_note) | |
1656 { | |
1657 replace_reg (&XEXP (src1_note, 0), FIRST_STACK_REG); | |
1658 regstack->top--; | |
1659 CLEAR_HARD_REG_BIT (regstack->reg_set, REGNO (*src1)); | |
1660 } | |
1661 | |
1662 replace_reg (src1, FIRST_STACK_REG); | |
1663 break; | |
1664 | |
1665 case UNSPEC_FXAM: | |
1666 | |
1667 /* This insn only operate on the top of the stack. */ | |
1668 | |
1669 src1 = get_true_reg (&XVECEXP (pat_src, 0, 0)); | |
1670 emit_swap_insn (insn, regstack, *src1); | |
1671 | |
1672 src1_note = find_regno_note (insn, REG_DEAD, REGNO (*src1)); | |
1673 | |
1674 replace_reg (src1, FIRST_STACK_REG); | |
1675 | |
1676 if (src1_note) | |
1677 { | |
1678 remove_regno_note (insn, REG_DEAD, | |
1679 REGNO (XEXP (src1_note, 0))); | |
1680 emit_pop_insn (insn, regstack, XEXP (src1_note, 0), | |
1681 EMIT_AFTER); | |
1682 } | |
1683 | |
1684 break; | |
1685 | |
1686 case UNSPEC_SIN: | |
1687 case UNSPEC_COS: | |
1688 case UNSPEC_FRNDINT: | |
1689 case UNSPEC_F2XM1: | |
1690 | |
1691 case UNSPEC_FRNDINT_FLOOR: | |
1692 case UNSPEC_FRNDINT_CEIL: | |
1693 case UNSPEC_FRNDINT_TRUNC: | |
1694 case UNSPEC_FRNDINT_MASK_PM: | |
1695 | |
1696 /* Above insns operate on the top of the stack. */ | |
1697 | |
1698 case UNSPEC_SINCOS_COS: | |
1699 case UNSPEC_XTRACT_FRACT: | |
1700 | |
1701 /* Above insns operate on the top two stack slots, | |
1702 first part of one input, double output insn. */ | |
1703 | |
1704 src1 = get_true_reg (&XVECEXP (pat_src, 0, 0)); | |
1705 | |
1706 emit_swap_insn (insn, regstack, *src1); | |
1707 | |
1708 /* Input should never die, it is replaced with output. */ | |
1709 src1_note = find_regno_note (insn, REG_DEAD, REGNO (*src1)); | |
1710 gcc_assert (!src1_note); | |
1711 | |
1712 if (STACK_REG_P (*dest)) | |
1713 replace_reg (dest, FIRST_STACK_REG); | |
1714 | |
1715 replace_reg (src1, FIRST_STACK_REG); | |
1716 break; | |
1717 | |
1718 case UNSPEC_SINCOS_SIN: | |
1719 case UNSPEC_XTRACT_EXP: | |
1720 | |
1721 /* These insns operate on the top two stack slots, | |
1722 second part of one input, double output insn. */ | |
1723 | |
1724 regstack->top++; | |
1725 /* FALLTHRU */ | |
1726 | |
1727 case UNSPEC_TAN: | |
1728 | |
1729 /* For UNSPEC_TAN, regstack->top is already increased | |
1730 by inherent load of constant 1.0. */ | |
1731 | |
1732 /* Output value is generated in the second stack slot. | |
1733 Move current value from second slot to the top. */ | |
1734 regstack->reg[regstack->top] | |
1735 = regstack->reg[regstack->top - 1]; | |
1736 | |
1737 gcc_assert (STACK_REG_P (*dest)); | |
1738 | |
1739 regstack->reg[regstack->top - 1] = REGNO (*dest); | |
1740 SET_HARD_REG_BIT (regstack->reg_set, REGNO (*dest)); | |
1741 replace_reg (dest, FIRST_STACK_REG + 1); | |
1742 | |
1743 src1 = get_true_reg (&XVECEXP (pat_src, 0, 0)); | |
1744 | |
1745 replace_reg (src1, FIRST_STACK_REG); | |
1746 break; | |
1747 | |
1748 case UNSPEC_FPATAN: | |
1749 case UNSPEC_FYL2X: | |
1750 case UNSPEC_FYL2XP1: | |
1751 /* These insns operate on the top two stack slots. */ | |
1752 | |
1753 src1 = get_true_reg (&XVECEXP (pat_src, 0, 0)); | |
1754 src2 = get_true_reg (&XVECEXP (pat_src, 0, 1)); | |
1755 | |
1756 src1_note = find_regno_note (insn, REG_DEAD, REGNO (*src1)); | |
1757 src2_note = find_regno_note (insn, REG_DEAD, REGNO (*src2)); | |
1758 | |
1759 swap_to_top (insn, regstack, *src1, *src2); | |
1760 | |
1761 replace_reg (src1, FIRST_STACK_REG); | |
1762 replace_reg (src2, FIRST_STACK_REG + 1); | |
1763 | |
1764 if (src1_note) | |
1765 replace_reg (&XEXP (src1_note, 0), FIRST_STACK_REG); | |
1766 if (src2_note) | |
1767 replace_reg (&XEXP (src2_note, 0), FIRST_STACK_REG + 1); | |
1768 | |
1769 /* Pop both input operands from the stack. */ | |
1770 CLEAR_HARD_REG_BIT (regstack->reg_set, | |
1771 regstack->reg[regstack->top]); | |
1772 CLEAR_HARD_REG_BIT (regstack->reg_set, | |
1773 regstack->reg[regstack->top - 1]); | |
1774 regstack->top -= 2; | |
1775 | |
1776 /* Push the result back onto the stack. */ | |
1777 regstack->reg[++regstack->top] = REGNO (*dest); | |
1778 SET_HARD_REG_BIT (regstack->reg_set, REGNO (*dest)); | |
1779 replace_reg (dest, FIRST_STACK_REG); | |
1780 break; | |
1781 | |
1782 case UNSPEC_FSCALE_FRACT: | |
1783 case UNSPEC_FPREM_F: | |
1784 case UNSPEC_FPREM1_F: | |
1785 /* These insns operate on the top two stack slots, | |
1786 first part of double input, double output insn. */ | |
1787 | |
1788 src1 = get_true_reg (&XVECEXP (pat_src, 0, 0)); | |
1789 src2 = get_true_reg (&XVECEXP (pat_src, 0, 1)); | |
1790 | |
1791 src1_note = find_regno_note (insn, REG_DEAD, REGNO (*src1)); | |
1792 src2_note = find_regno_note (insn, REG_DEAD, REGNO (*src2)); | |
1793 | |
1794 /* Inputs should never die, they are | |
1795 replaced with outputs. */ | |
1796 gcc_assert (!src1_note); | |
1797 gcc_assert (!src2_note); | |
1798 | |
1799 swap_to_top (insn, regstack, *src1, *src2); | |
1800 | |
1801 /* Push the result back onto stack. Empty stack slot | |
1802 will be filled in second part of insn. */ | |
1803 if (STACK_REG_P (*dest)) | |
1804 { | |
1805 regstack->reg[regstack->top] = REGNO (*dest); | |
1806 SET_HARD_REG_BIT (regstack->reg_set, REGNO (*dest)); | |
1807 replace_reg (dest, FIRST_STACK_REG); | |
1808 } | |
1809 | |
1810 replace_reg (src1, FIRST_STACK_REG); | |
1811 replace_reg (src2, FIRST_STACK_REG + 1); | |
1812 break; | |
1813 | |
1814 case UNSPEC_FSCALE_EXP: | |
1815 case UNSPEC_FPREM_U: | |
1816 case UNSPEC_FPREM1_U: | |
1817 /* These insns operate on the top two stack slots, | |
1818 second part of double input, double output insn. */ | |
1819 | |
1820 src1 = get_true_reg (&XVECEXP (pat_src, 0, 0)); | |
1821 src2 = get_true_reg (&XVECEXP (pat_src, 0, 1)); | |
1822 | |
1823 /* Push the result back onto stack. Fill empty slot from | |
1824 first part of insn and fix top of stack pointer. */ | |
1825 if (STACK_REG_P (*dest)) | |
1826 { | |
1827 regstack->reg[regstack->top - 1] = REGNO (*dest); | |
1828 SET_HARD_REG_BIT (regstack->reg_set, REGNO (*dest)); | |
1829 replace_reg (dest, FIRST_STACK_REG + 1); | |
1830 } | |
1831 | |
1832 replace_reg (src1, FIRST_STACK_REG); | |
1833 replace_reg (src2, FIRST_STACK_REG + 1); | |
1834 break; | |
1835 | |
1836 case UNSPEC_C2_FLAG: | |
1837 /* This insn operates on the top two stack slots, | |
1838 third part of C2 setting double input insn. */ | |
1839 | |
1840 src1 = get_true_reg (&XVECEXP (pat_src, 0, 0)); | |
1841 src2 = get_true_reg (&XVECEXP (pat_src, 0, 1)); | |
1842 | |
1843 replace_reg (src1, FIRST_STACK_REG); | |
1844 replace_reg (src2, FIRST_STACK_REG + 1); | |
1845 break; | |
1846 | |
1847 case UNSPEC_SAHF: | |
1848 /* (unspec [(unspec [(compare)] UNSPEC_FNSTSW)] UNSPEC_SAHF) | |
1849 The combination matches the PPRO fcomi instruction. */ | |
1850 | |
1851 pat_src = XVECEXP (pat_src, 0, 0); | |
1852 gcc_assert (GET_CODE (pat_src) == UNSPEC); | |
1853 gcc_assert (XINT (pat_src, 1) == UNSPEC_FNSTSW); | |
1854 /* Fall through. */ | |
1855 | |
1856 case UNSPEC_FNSTSW: | |
1857 /* Combined fcomp+fnstsw generated for doing well with | |
1858 CSE. When optimizing this would have been broken | |
1859 up before now. */ | |
1860 | |
1861 pat_src = XVECEXP (pat_src, 0, 0); | |
1862 gcc_assert (GET_CODE (pat_src) == COMPARE); | |
1863 | |
1864 compare_for_stack_reg (insn, regstack, pat_src); | |
1865 break; | |
1866 | |
1867 default: | |
1868 gcc_unreachable (); | |
1869 } | |
1870 break; | |
1871 | |
1872 case IF_THEN_ELSE: | |
1873 /* This insn requires the top of stack to be the destination. */ | |
1874 | |
1875 src1 = get_true_reg (&XEXP (pat_src, 1)); | |
1876 src2 = get_true_reg (&XEXP (pat_src, 2)); | |
1877 | |
1878 src1_note = find_regno_note (insn, REG_DEAD, REGNO (*src1)); | |
1879 src2_note = find_regno_note (insn, REG_DEAD, REGNO (*src2)); | |
1880 | |
1881 /* If the comparison operator is an FP comparison operator, | |
1882 it is handled correctly by compare_for_stack_reg () who | |
1883 will move the destination to the top of stack. But if the | |
1884 comparison operator is not an FP comparison operator, we | |
1885 have to handle it here. */ | |
1886 if (get_hard_regnum (regstack, *dest) >= FIRST_STACK_REG | |
1887 && REGNO (*dest) != regstack->reg[regstack->top]) | |
1888 { | |
1889 /* In case one of operands is the top of stack and the operands | |
1890 dies, it is safe to make it the destination operand by | |
1891 reversing the direction of cmove and avoid fxch. */ | |
1892 if ((REGNO (*src1) == regstack->reg[regstack->top] | |
1893 && src1_note) | |
1894 || (REGNO (*src2) == regstack->reg[regstack->top] | |
1895 && src2_note)) | |
1896 { | |
1897 int idx1 = (get_hard_regnum (regstack, *src1) | |
1898 - FIRST_STACK_REG); | |
1899 int idx2 = (get_hard_regnum (regstack, *src2) | |
1900 - FIRST_STACK_REG); | |
1901 | |
1902 /* Make reg-stack believe that the operands are already | |
1903 swapped on the stack */ | |
1904 regstack->reg[regstack->top - idx1] = REGNO (*src2); | |
1905 regstack->reg[regstack->top - idx2] = REGNO (*src1); | |
1906 | |
1907 /* Reverse condition to compensate the operand swap. | |
1908 i386 do have comparison always reversible. */ | |
1909 PUT_CODE (XEXP (pat_src, 0), | |
1910 reversed_comparison_code (XEXP (pat_src, 0), insn)); | |
1911 } | |
1912 else | |
1913 emit_swap_insn (insn, regstack, *dest); | |
1914 } | |
1915 | |
1916 { | |
1917 rtx src_note [3]; | |
1918 int i; | |
1919 | |
1920 src_note[0] = 0; | |
1921 src_note[1] = src1_note; | |
1922 src_note[2] = src2_note; | |
1923 | |
1924 if (STACK_REG_P (*src1)) | |
1925 replace_reg (src1, get_hard_regnum (regstack, *src1)); | |
1926 if (STACK_REG_P (*src2)) | |
1927 replace_reg (src2, get_hard_regnum (regstack, *src2)); | |
1928 | |
1929 for (i = 1; i <= 2; i++) | |
1930 if (src_note [i]) | |
1931 { | |
1932 int regno = REGNO (XEXP (src_note[i], 0)); | |
1933 | |
1934 /* If the register that dies is not at the top of | |
1935 stack, then move the top of stack to the dead reg. | |
1936 Top of stack should never die, as it is the | |
1937 destination. */ | |
1938 gcc_assert (regno != regstack->reg[regstack->top]); | |
1939 remove_regno_note (insn, REG_DEAD, regno); | |
1940 emit_pop_insn (insn, regstack, XEXP (src_note[i], 0), | |
1941 EMIT_AFTER); | |
1942 } | |
1943 } | |
1944 | |
1945 /* Make dest the top of stack. Add dest to regstack if | |
1946 not present. */ | |
1947 if (get_hard_regnum (regstack, *dest) < FIRST_STACK_REG) | |
1948 regstack->reg[++regstack->top] = REGNO (*dest); | |
1949 SET_HARD_REG_BIT (regstack->reg_set, REGNO (*dest)); | |
1950 replace_reg (dest, FIRST_STACK_REG); | |
1951 break; | |
1952 | |
1953 default: | |
1954 gcc_unreachable (); | |
1955 } | |
1956 break; | |
1957 } | |
1958 | |
1959 default: | |
1960 break; | |
1961 } | |
1962 | |
1963 return control_flow_insn_deleted; | |
1964 } | |
1965 | |
1966 /* Substitute hard regnums for any stack regs in INSN, which has | |
1967 N_INPUTS inputs and N_OUTPUTS outputs. REGSTACK is the stack info | |
1968 before the insn, and is updated with changes made here. | |
1969 | |
1970 There are several requirements and assumptions about the use of | |
1971 stack-like regs in asm statements. These rules are enforced by | |
1972 record_asm_stack_regs; see comments there for details. Any | |
1973 asm_operands left in the RTL at this point may be assume to meet the | |
1974 requirements, since record_asm_stack_regs removes any problem asm. */ | |
1975 | |
1976 static void | |
1977 subst_asm_stack_regs (rtx insn, stack regstack) | |
1978 { | |
1979 rtx body = PATTERN (insn); | |
1980 int alt; | |
1981 | |
1982 rtx *note_reg; /* Array of note contents */ | |
1983 rtx **note_loc; /* Address of REG field of each note */ | |
1984 enum reg_note *note_kind; /* The type of each note */ | |
1985 | |
1986 rtx *clobber_reg = 0; | |
1987 rtx **clobber_loc = 0; | |
1988 | |
1989 struct stack_def temp_stack; | |
1990 int n_notes; | |
1991 int n_clobbers; | |
1992 rtx note; | |
1993 int i; | |
1994 int n_inputs, n_outputs; | |
1995 | |
1996 if (! check_asm_stack_operands (insn)) | |
1997 return; | |
1998 | |
1999 /* Find out what the constraints required. If no constraint | |
2000 alternative matches, that is a compiler bug: we should have caught | |
2001 such an insn in check_asm_stack_operands. */ | |
2002 extract_insn (insn); | |
2003 constrain_operands (1); | |
2004 alt = which_alternative; | |
2005 | |
2006 preprocess_constraints (); | |
2007 | |
2008 n_inputs = get_asm_operand_n_inputs (body); | |
2009 n_outputs = recog_data.n_operands - n_inputs; | |
2010 | |
2011 gcc_assert (alt >= 0); | |
2012 | |
2013 /* Strip SUBREGs here to make the following code simpler. */ | |
2014 for (i = 0; i < recog_data.n_operands; i++) | |
2015 if (GET_CODE (recog_data.operand[i]) == SUBREG | |
2016 && REG_P (SUBREG_REG (recog_data.operand[i]))) | |
2017 { | |
2018 recog_data.operand_loc[i] = & SUBREG_REG (recog_data.operand[i]); | |
2019 recog_data.operand[i] = SUBREG_REG (recog_data.operand[i]); | |
2020 } | |
2021 | |
2022 /* Set up NOTE_REG, NOTE_LOC and NOTE_KIND. */ | |
2023 | |
2024 for (i = 0, note = REG_NOTES (insn); note; note = XEXP (note, 1)) | |
2025 i++; | |
2026 | |
2027 note_reg = XALLOCAVEC (rtx, i); | |
2028 note_loc = XALLOCAVEC (rtx *, i); | |
2029 note_kind = XALLOCAVEC (enum reg_note, i); | |
2030 | |
2031 n_notes = 0; | |
2032 for (note = REG_NOTES (insn); note; note = XEXP (note, 1)) | |
2033 { | |
2034 rtx reg = XEXP (note, 0); | |
2035 rtx *loc = & XEXP (note, 0); | |
2036 | |
2037 if (GET_CODE (reg) == SUBREG && REG_P (SUBREG_REG (reg))) | |
2038 { | |
2039 loc = & SUBREG_REG (reg); | |
2040 reg = SUBREG_REG (reg); | |
2041 } | |
2042 | |
2043 if (STACK_REG_P (reg) | |
2044 && (REG_NOTE_KIND (note) == REG_DEAD | |
2045 || REG_NOTE_KIND (note) == REG_UNUSED)) | |
2046 { | |
2047 note_reg[n_notes] = reg; | |
2048 note_loc[n_notes] = loc; | |
2049 note_kind[n_notes] = REG_NOTE_KIND (note); | |
2050 n_notes++; | |
2051 } | |
2052 } | |
2053 | |
2054 /* Set up CLOBBER_REG and CLOBBER_LOC. */ | |
2055 | |
2056 n_clobbers = 0; | |
2057 | |
2058 if (GET_CODE (body) == PARALLEL) | |
2059 { | |
2060 clobber_reg = XALLOCAVEC (rtx, XVECLEN (body, 0)); | |
2061 clobber_loc = XALLOCAVEC (rtx *, XVECLEN (body, 0)); | |
2062 | |
2063 for (i = 0; i < XVECLEN (body, 0); i++) | |
2064 if (GET_CODE (XVECEXP (body, 0, i)) == CLOBBER) | |
2065 { | |
2066 rtx clobber = XVECEXP (body, 0, i); | |
2067 rtx reg = XEXP (clobber, 0); | |
2068 rtx *loc = & XEXP (clobber, 0); | |
2069 | |
2070 if (GET_CODE (reg) == SUBREG && REG_P (SUBREG_REG (reg))) | |
2071 { | |
2072 loc = & SUBREG_REG (reg); | |
2073 reg = SUBREG_REG (reg); | |
2074 } | |
2075 | |
2076 if (STACK_REG_P (reg)) | |
2077 { | |
2078 clobber_reg[n_clobbers] = reg; | |
2079 clobber_loc[n_clobbers] = loc; | |
2080 n_clobbers++; | |
2081 } | |
2082 } | |
2083 } | |
2084 | |
2085 temp_stack = *regstack; | |
2086 | |
2087 /* Put the input regs into the desired place in TEMP_STACK. */ | |
2088 | |
2089 for (i = n_outputs; i < n_outputs + n_inputs; i++) | |
2090 if (STACK_REG_P (recog_data.operand[i]) | |
2091 && reg_class_subset_p (recog_op_alt[i][alt].cl, | |
2092 FLOAT_REGS) | |
2093 && recog_op_alt[i][alt].cl != FLOAT_REGS) | |
2094 { | |
2095 /* If an operand needs to be in a particular reg in | |
2096 FLOAT_REGS, the constraint was either 't' or 'u'. Since | |
2097 these constraints are for single register classes, and | |
2098 reload guaranteed that operand[i] is already in that class, | |
2099 we can just use REGNO (recog_data.operand[i]) to know which | |
2100 actual reg this operand needs to be in. */ | |
2101 | |
2102 int regno = get_hard_regnum (&temp_stack, recog_data.operand[i]); | |
2103 | |
2104 gcc_assert (regno >= 0); | |
2105 | |
2106 if ((unsigned int) regno != REGNO (recog_data.operand[i])) | |
2107 { | |
2108 /* recog_data.operand[i] is not in the right place. Find | |
2109 it and swap it with whatever is already in I's place. | |
2110 K is where recog_data.operand[i] is now. J is where it | |
2111 should be. */ | |
2112 int j, k, temp; | |
2113 | |
2114 k = temp_stack.top - (regno - FIRST_STACK_REG); | |
2115 j = (temp_stack.top | |
2116 - (REGNO (recog_data.operand[i]) - FIRST_STACK_REG)); | |
2117 | |
2118 temp = temp_stack.reg[k]; | |
2119 temp_stack.reg[k] = temp_stack.reg[j]; | |
2120 temp_stack.reg[j] = temp; | |
2121 } | |
2122 } | |
2123 | |
2124 /* Emit insns before INSN to make sure the reg-stack is in the right | |
2125 order. */ | |
2126 | |
2127 change_stack (insn, regstack, &temp_stack, EMIT_BEFORE); | |
2128 | |
2129 /* Make the needed input register substitutions. Do death notes and | |
2130 clobbers too, because these are for inputs, not outputs. */ | |
2131 | |
2132 for (i = n_outputs; i < n_outputs + n_inputs; i++) | |
2133 if (STACK_REG_P (recog_data.operand[i])) | |
2134 { | |
2135 int regnum = get_hard_regnum (regstack, recog_data.operand[i]); | |
2136 | |
2137 gcc_assert (regnum >= 0); | |
2138 | |
2139 replace_reg (recog_data.operand_loc[i], regnum); | |
2140 } | |
2141 | |
2142 for (i = 0; i < n_notes; i++) | |
2143 if (note_kind[i] == REG_DEAD) | |
2144 { | |
2145 int regnum = get_hard_regnum (regstack, note_reg[i]); | |
2146 | |
2147 gcc_assert (regnum >= 0); | |
2148 | |
2149 replace_reg (note_loc[i], regnum); | |
2150 } | |
2151 | |
2152 for (i = 0; i < n_clobbers; i++) | |
2153 { | |
2154 /* It's OK for a CLOBBER to reference a reg that is not live. | |
2155 Don't try to replace it in that case. */ | |
2156 int regnum = get_hard_regnum (regstack, clobber_reg[i]); | |
2157 | |
2158 if (regnum >= 0) | |
2159 { | |
2160 /* Sigh - clobbers always have QImode. But replace_reg knows | |
2161 that these regs can't be MODE_INT and will assert. Just put | |
2162 the right reg there without calling replace_reg. */ | |
2163 | |
2164 *clobber_loc[i] = FP_MODE_REG (regnum, DFmode); | |
2165 } | |
2166 } | |
2167 | |
2168 /* Now remove from REGSTACK any inputs that the asm implicitly popped. */ | |
2169 | |
2170 for (i = n_outputs; i < n_outputs + n_inputs; i++) | |
2171 if (STACK_REG_P (recog_data.operand[i])) | |
2172 { | |
2173 /* An input reg is implicitly popped if it is tied to an | |
2174 output, or if there is a CLOBBER for it. */ | |
2175 int j; | |
2176 | |
2177 for (j = 0; j < n_clobbers; j++) | |
2178 if (operands_match_p (clobber_reg[j], recog_data.operand[i])) | |
2179 break; | |
2180 | |
2181 if (j < n_clobbers || recog_op_alt[i][alt].matches >= 0) | |
2182 { | |
2183 /* recog_data.operand[i] might not be at the top of stack. | |
2184 But that's OK, because all we need to do is pop the | |
2185 right number of regs off of the top of the reg-stack. | |
2186 record_asm_stack_regs guaranteed that all implicitly | |
2187 popped regs were grouped at the top of the reg-stack. */ | |
2188 | |
2189 CLEAR_HARD_REG_BIT (regstack->reg_set, | |
2190 regstack->reg[regstack->top]); | |
2191 regstack->top--; | |
2192 } | |
2193 } | |
2194 | |
2195 /* Now add to REGSTACK any outputs that the asm implicitly pushed. | |
2196 Note that there isn't any need to substitute register numbers. | |
2197 ??? Explain why this is true. */ | |
2198 | |
2199 for (i = LAST_STACK_REG; i >= FIRST_STACK_REG; i--) | |
2200 { | |
2201 /* See if there is an output for this hard reg. */ | |
2202 int j; | |
2203 | |
2204 for (j = 0; j < n_outputs; j++) | |
2205 if (STACK_REG_P (recog_data.operand[j]) | |
2206 && REGNO (recog_data.operand[j]) == (unsigned) i) | |
2207 { | |
2208 regstack->reg[++regstack->top] = i; | |
2209 SET_HARD_REG_BIT (regstack->reg_set, i); | |
2210 break; | |
2211 } | |
2212 } | |
2213 | |
2214 /* Now emit a pop insn for any REG_UNUSED output, or any REG_DEAD | |
2215 input that the asm didn't implicitly pop. If the asm didn't | |
2216 implicitly pop an input reg, that reg will still be live. | |
2217 | |
2218 Note that we can't use find_regno_note here: the register numbers | |
2219 in the death notes have already been substituted. */ | |
2220 | |
2221 for (i = 0; i < n_outputs; i++) | |
2222 if (STACK_REG_P (recog_data.operand[i])) | |
2223 { | |
2224 int j; | |
2225 | |
2226 for (j = 0; j < n_notes; j++) | |
2227 if (REGNO (recog_data.operand[i]) == REGNO (note_reg[j]) | |
2228 && note_kind[j] == REG_UNUSED) | |
2229 { | |
2230 insn = emit_pop_insn (insn, regstack, recog_data.operand[i], | |
2231 EMIT_AFTER); | |
2232 break; | |
2233 } | |
2234 } | |
2235 | |
2236 for (i = n_outputs; i < n_outputs + n_inputs; i++) | |
2237 if (STACK_REG_P (recog_data.operand[i])) | |
2238 { | |
2239 int j; | |
2240 | |
2241 for (j = 0; j < n_notes; j++) | |
2242 if (REGNO (recog_data.operand[i]) == REGNO (note_reg[j]) | |
2243 && note_kind[j] == REG_DEAD | |
2244 && TEST_HARD_REG_BIT (regstack->reg_set, | |
2245 REGNO (recog_data.operand[i]))) | |
2246 { | |
2247 insn = emit_pop_insn (insn, regstack, recog_data.operand[i], | |
2248 EMIT_AFTER); | |
2249 break; | |
2250 } | |
2251 } | |
2252 } | |
2253 | |
2254 /* Substitute stack hard reg numbers for stack virtual registers in | |
2255 INSN. Non-stack register numbers are not changed. REGSTACK is the | |
2256 current stack content. Insns may be emitted as needed to arrange the | |
2257 stack for the 387 based on the contents of the insn. Return whether | |
2258 a control flow insn was deleted in the process. */ | |
2259 | |
2260 static bool | |
2261 subst_stack_regs (rtx insn, stack regstack) | |
2262 { | |
2263 rtx *note_link, note; | |
2264 bool control_flow_insn_deleted = false; | |
2265 int i; | |
2266 | |
2267 if (CALL_P (insn)) | |
2268 { | |
2269 int top = regstack->top; | |
2270 | |
2271 /* If there are any floating point parameters to be passed in | |
2272 registers for this call, make sure they are in the right | |
2273 order. */ | |
2274 | |
2275 if (top >= 0) | |
2276 { | |
2277 straighten_stack (insn, regstack); | |
2278 | |
2279 /* Now mark the arguments as dead after the call. */ | |
2280 | |
2281 while (regstack->top >= 0) | |
2282 { | |
2283 CLEAR_HARD_REG_BIT (regstack->reg_set, FIRST_STACK_REG + regstack->top); | |
2284 regstack->top--; | |
2285 } | |
2286 } | |
2287 } | |
2288 | |
2289 /* Do the actual substitution if any stack regs are mentioned. | |
2290 Since we only record whether entire insn mentions stack regs, and | |
2291 subst_stack_regs_pat only works for patterns that contain stack regs, | |
2292 we must check each pattern in a parallel here. A call_value_pop could | |
2293 fail otherwise. */ | |
2294 | |
2295 if (stack_regs_mentioned (insn)) | |
2296 { | |
2297 int n_operands = asm_noperands (PATTERN (insn)); | |
2298 if (n_operands >= 0) | |
2299 { | |
2300 /* This insn is an `asm' with operands. Decode the operands, | |
2301 decide how many are inputs, and do register substitution. | |
2302 Any REG_UNUSED notes will be handled by subst_asm_stack_regs. */ | |
2303 | |
2304 subst_asm_stack_regs (insn, regstack); | |
2305 return control_flow_insn_deleted; | |
2306 } | |
2307 | |
2308 if (GET_CODE (PATTERN (insn)) == PARALLEL) | |
2309 for (i = 0; i < XVECLEN (PATTERN (insn), 0); i++) | |
2310 { | |
2311 if (stack_regs_mentioned_p (XVECEXP (PATTERN (insn), 0, i))) | |
2312 { | |
2313 if (GET_CODE (XVECEXP (PATTERN (insn), 0, i)) == CLOBBER) | |
2314 XVECEXP (PATTERN (insn), 0, i) | |
2315 = shallow_copy_rtx (XVECEXP (PATTERN (insn), 0, i)); | |
2316 control_flow_insn_deleted | |
2317 |= subst_stack_regs_pat (insn, regstack, | |
2318 XVECEXP (PATTERN (insn), 0, i)); | |
2319 } | |
2320 } | |
2321 else | |
2322 control_flow_insn_deleted | |
2323 |= subst_stack_regs_pat (insn, regstack, PATTERN (insn)); | |
2324 } | |
2325 | |
2326 /* subst_stack_regs_pat may have deleted a no-op insn. If so, any | |
2327 REG_UNUSED will already have been dealt with, so just return. */ | |
2328 | |
2329 if (NOTE_P (insn) || INSN_DELETED_P (insn)) | |
2330 return control_flow_insn_deleted; | |
2331 | |
2332 /* If this a noreturn call, we can't insert pop insns after it. | |
2333 Instead, reset the stack state to empty. */ | |
2334 if (CALL_P (insn) | |
2335 && find_reg_note (insn, REG_NORETURN, NULL)) | |
2336 { | |
2337 regstack->top = -1; | |
2338 CLEAR_HARD_REG_SET (regstack->reg_set); | |
2339 return control_flow_insn_deleted; | |
2340 } | |
2341 | |
2342 /* If there is a REG_UNUSED note on a stack register on this insn, | |
2343 the indicated reg must be popped. The REG_UNUSED note is removed, | |
2344 since the form of the newly emitted pop insn references the reg, | |
2345 making it no longer `unset'. */ | |
2346 | |
2347 note_link = ®_NOTES (insn); | |
2348 for (note = *note_link; note; note = XEXP (note, 1)) | |
2349 if (REG_NOTE_KIND (note) == REG_UNUSED && STACK_REG_P (XEXP (note, 0))) | |
2350 { | |
2351 *note_link = XEXP (note, 1); | |
2352 insn = emit_pop_insn (insn, regstack, XEXP (note, 0), EMIT_AFTER); | |
2353 } | |
2354 else | |
2355 note_link = &XEXP (note, 1); | |
2356 | |
2357 return control_flow_insn_deleted; | |
2358 } | |
2359 | |
2360 /* Change the organization of the stack so that it fits a new basic | |
2361 block. Some registers might have to be popped, but there can never be | |
2362 a register live in the new block that is not now live. | |
2363 | |
2364 Insert any needed insns before or after INSN, as indicated by | |
2365 WHERE. OLD is the original stack layout, and NEW is the desired | |
2366 form. OLD is updated to reflect the code emitted, i.e., it will be | |
2367 the same as NEW upon return. | |
2368 | |
2369 This function will not preserve block_end[]. But that information | |
2370 is no longer needed once this has executed. */ | |
2371 | |
2372 static void | |
2373 change_stack (rtx insn, stack old, stack new_stack, enum emit_where where) | |
2374 { | |
2375 int reg; | |
2376 int update_end = 0; | |
2377 int i; | |
2378 | |
2379 /* Stack adjustments for the first insn in a block update the | |
2380 current_block's stack_in instead of inserting insns directly. | |
2381 compensate_edges will add the necessary code later. */ | |
2382 if (current_block | |
2383 && starting_stack_p | |
2384 && where == EMIT_BEFORE) | |
2385 { | |
2386 BLOCK_INFO (current_block)->stack_in = *new_stack; | |
2387 starting_stack_p = false; | |
2388 *old = *new_stack; | |
2389 return; | |
2390 } | |
2391 | |
2392 /* We will be inserting new insns "backwards". If we are to insert | |
2393 after INSN, find the next insn, and insert before it. */ | |
2394 | |
2395 if (where == EMIT_AFTER) | |
2396 { | |
2397 if (current_block && BB_END (current_block) == insn) | |
2398 update_end = 1; | |
2399 insn = NEXT_INSN (insn); | |
2400 } | |
2401 | |
2402 /* Initialize partially dead variables. */ | |
2403 for (i = FIRST_STACK_REG; i < LAST_STACK_REG + 1; i++) | |
2404 if (TEST_HARD_REG_BIT (new_stack->reg_set, i) | |
2405 && !TEST_HARD_REG_BIT (old->reg_set, i)) | |
2406 { | |
2407 old->reg[++old->top] = i; | |
2408 SET_HARD_REG_BIT (old->reg_set, i); | |
2409 emit_insn_before (gen_rtx_SET (VOIDmode, | |
2410 FP_MODE_REG (i, SFmode), not_a_num), insn); | |
2411 } | |
2412 | |
2413 /* Pop any registers that are not needed in the new block. */ | |
2414 | |
2415 /* If the destination block's stack already has a specified layout | |
2416 and contains two or more registers, use a more intelligent algorithm | |
2417 to pop registers that minimizes the number number of fxchs below. */ | |
2418 if (new_stack->top > 0) | |
2419 { | |
2420 bool slots[REG_STACK_SIZE]; | |
2421 int pops[REG_STACK_SIZE]; | |
2422 int next, dest, topsrc; | |
2423 | |
2424 /* First pass to determine the free slots. */ | |
2425 for (reg = 0; reg <= new_stack->top; reg++) | |
2426 slots[reg] = TEST_HARD_REG_BIT (new_stack->reg_set, old->reg[reg]); | |
2427 | |
2428 /* Second pass to allocate preferred slots. */ | |
2429 topsrc = -1; | |
2430 for (reg = old->top; reg > new_stack->top; reg--) | |
2431 if (TEST_HARD_REG_BIT (new_stack->reg_set, old->reg[reg])) | |
2432 { | |
2433 dest = -1; | |
2434 for (next = 0; next <= new_stack->top; next++) | |
2435 if (!slots[next] && new_stack->reg[next] == old->reg[reg]) | |
2436 { | |
2437 /* If this is a preference for the new top of stack, record | |
2438 the fact by remembering it's old->reg in topsrc. */ | |
2439 if (next == new_stack->top) | |
2440 topsrc = reg; | |
2441 slots[next] = true; | |
2442 dest = next; | |
2443 break; | |
2444 } | |
2445 pops[reg] = dest; | |
2446 } | |
2447 else | |
2448 pops[reg] = reg; | |
2449 | |
2450 /* Intentionally, avoid placing the top of stack in it's correct | |
2451 location, if we still need to permute the stack below and we | |
2452 can usefully place it somewhere else. This is the case if any | |
2453 slot is still unallocated, in which case we should place the | |
2454 top of stack there. */ | |
2455 if (topsrc != -1) | |
2456 for (reg = 0; reg < new_stack->top; reg++) | |
2457 if (!slots[reg]) | |
2458 { | |
2459 pops[topsrc] = reg; | |
2460 slots[new_stack->top] = false; | |
2461 slots[reg] = true; | |
2462 break; | |
2463 } | |
2464 | |
2465 /* Third pass allocates remaining slots and emits pop insns. */ | |
2466 next = new_stack->top; | |
2467 for (reg = old->top; reg > new_stack->top; reg--) | |
2468 { | |
2469 dest = pops[reg]; | |
2470 if (dest == -1) | |
2471 { | |
2472 /* Find next free slot. */ | |
2473 while (slots[next]) | |
2474 next--; | |
2475 dest = next--; | |
2476 } | |
2477 emit_pop_insn (insn, old, FP_MODE_REG (old->reg[dest], DFmode), | |
2478 EMIT_BEFORE); | |
2479 } | |
2480 } | |
2481 else | |
2482 { | |
2483 /* The following loop attempts to maximize the number of times we | |
2484 pop the top of the stack, as this permits the use of the faster | |
2485 ffreep instruction on platforms that support it. */ | |
2486 int live, next; | |
2487 | |
2488 live = 0; | |
2489 for (reg = 0; reg <= old->top; reg++) | |
2490 if (TEST_HARD_REG_BIT (new_stack->reg_set, old->reg[reg])) | |
2491 live++; | |
2492 | |
2493 next = live; | |
2494 while (old->top >= live) | |
2495 if (TEST_HARD_REG_BIT (new_stack->reg_set, old->reg[old->top])) | |
2496 { | |
2497 while (TEST_HARD_REG_BIT (new_stack->reg_set, old->reg[next])) | |
2498 next--; | |
2499 emit_pop_insn (insn, old, FP_MODE_REG (old->reg[next], DFmode), | |
2500 EMIT_BEFORE); | |
2501 } | |
2502 else | |
2503 emit_pop_insn (insn, old, FP_MODE_REG (old->reg[old->top], DFmode), | |
2504 EMIT_BEFORE); | |
2505 } | |
2506 | |
2507 if (new_stack->top == -2) | |
2508 { | |
2509 /* If the new block has never been processed, then it can inherit | |
2510 the old stack order. */ | |
2511 | |
2512 new_stack->top = old->top; | |
2513 memcpy (new_stack->reg, old->reg, sizeof (new_stack->reg)); | |
2514 } | |
2515 else | |
2516 { | |
2517 /* This block has been entered before, and we must match the | |
2518 previously selected stack order. */ | |
2519 | |
2520 /* By now, the only difference should be the order of the stack, | |
2521 not their depth or liveliness. */ | |
2522 | |
2523 gcc_assert (hard_reg_set_equal_p (old->reg_set, new_stack->reg_set)); | |
2524 gcc_assert (old->top == new_stack->top); | |
2525 | |
2526 /* If the stack is not empty (new_stack->top != -1), loop here emitting | |
2527 swaps until the stack is correct. | |
2528 | |
2529 The worst case number of swaps emitted is N + 2, where N is the | |
2530 depth of the stack. In some cases, the reg at the top of | |
2531 stack may be correct, but swapped anyway in order to fix | |
2532 other regs. But since we never swap any other reg away from | |
2533 its correct slot, this algorithm will converge. */ | |
2534 | |
2535 if (new_stack->top != -1) | |
2536 do | |
2537 { | |
2538 /* Swap the reg at top of stack into the position it is | |
2539 supposed to be in, until the correct top of stack appears. */ | |
2540 | |
2541 while (old->reg[old->top] != new_stack->reg[new_stack->top]) | |
2542 { | |
2543 for (reg = new_stack->top; reg >= 0; reg--) | |
2544 if (new_stack->reg[reg] == old->reg[old->top]) | |
2545 break; | |
2546 | |
2547 gcc_assert (reg != -1); | |
2548 | |
2549 emit_swap_insn (insn, old, | |
2550 FP_MODE_REG (old->reg[reg], DFmode)); | |
2551 } | |
2552 | |
2553 /* See if any regs remain incorrect. If so, bring an | |
2554 incorrect reg to the top of stack, and let the while loop | |
2555 above fix it. */ | |
2556 | |
2557 for (reg = new_stack->top; reg >= 0; reg--) | |
2558 if (new_stack->reg[reg] != old->reg[reg]) | |
2559 { | |
2560 emit_swap_insn (insn, old, | |
2561 FP_MODE_REG (old->reg[reg], DFmode)); | |
2562 break; | |
2563 } | |
2564 } while (reg >= 0); | |
2565 | |
2566 /* At this point there must be no differences. */ | |
2567 | |
2568 for (reg = old->top; reg >= 0; reg--) | |
2569 gcc_assert (old->reg[reg] == new_stack->reg[reg]); | |
2570 } | |
2571 | |
2572 if (update_end) | |
2573 BB_END (current_block) = PREV_INSN (insn); | |
2574 } | |
2575 | |
2576 /* Print stack configuration. */ | |
2577 | |
2578 static void | |
2579 print_stack (FILE *file, stack s) | |
2580 { | |
2581 if (! file) | |
2582 return; | |
2583 | |
2584 if (s->top == -2) | |
2585 fprintf (file, "uninitialized\n"); | |
2586 else if (s->top == -1) | |
2587 fprintf (file, "empty\n"); | |
2588 else | |
2589 { | |
2590 int i; | |
2591 fputs ("[ ", file); | |
2592 for (i = 0; i <= s->top; ++i) | |
2593 fprintf (file, "%d ", s->reg[i]); | |
2594 fputs ("]\n", file); | |
2595 } | |
2596 } | |
2597 | |
2598 /* This function was doing life analysis. We now let the regular live | |
2599 code do it's job, so we only need to check some extra invariants | |
2600 that reg-stack expects. Primary among these being that all registers | |
2601 are initialized before use. | |
2602 | |
2603 The function returns true when code was emitted to CFG edges and | |
2604 commit_edge_insertions needs to be called. */ | |
2605 | |
2606 static int | |
2607 convert_regs_entry (void) | |
2608 { | |
2609 int inserted = 0; | |
2610 edge e; | |
2611 edge_iterator ei; | |
2612 | |
2613 /* Load something into each stack register live at function entry. | |
2614 Such live registers can be caused by uninitialized variables or | |
2615 functions not returning values on all paths. In order to keep | |
2616 the push/pop code happy, and to not scrog the register stack, we | |
2617 must put something in these registers. Use a QNaN. | |
2618 | |
2619 Note that we are inserting converted code here. This code is | |
2620 never seen by the convert_regs pass. */ | |
2621 | |
2622 FOR_EACH_EDGE (e, ei, ENTRY_BLOCK_PTR->succs) | |
2623 { | |
2624 basic_block block = e->dest; | |
2625 block_info bi = BLOCK_INFO (block); | |
2626 int reg, top = -1; | |
2627 | |
2628 for (reg = LAST_STACK_REG; reg >= FIRST_STACK_REG; --reg) | |
2629 if (TEST_HARD_REG_BIT (bi->stack_in.reg_set, reg)) | |
2630 { | |
2631 rtx init; | |
2632 | |
2633 bi->stack_in.reg[++top] = reg; | |
2634 | |
2635 init = gen_rtx_SET (VOIDmode, | |
2636 FP_MODE_REG (FIRST_STACK_REG, SFmode), | |
2637 not_a_num); | |
2638 insert_insn_on_edge (init, e); | |
2639 inserted = 1; | |
2640 } | |
2641 | |
2642 bi->stack_in.top = top; | |
2643 } | |
2644 | |
2645 return inserted; | |
2646 } | |
2647 | |
2648 /* Construct the desired stack for function exit. This will either | |
2649 be `empty', or the function return value at top-of-stack. */ | |
2650 | |
2651 static void | |
2652 convert_regs_exit (void) | |
2653 { | |
2654 int value_reg_low, value_reg_high; | |
2655 stack output_stack; | |
2656 rtx retvalue; | |
2657 | |
2658 retvalue = stack_result (current_function_decl); | |
2659 value_reg_low = value_reg_high = -1; | |
2660 if (retvalue) | |
2661 { | |
2662 value_reg_low = REGNO (retvalue); | |
2663 value_reg_high = END_HARD_REGNO (retvalue) - 1; | |
2664 } | |
2665 | |
2666 output_stack = &BLOCK_INFO (EXIT_BLOCK_PTR)->stack_in; | |
2667 if (value_reg_low == -1) | |
2668 output_stack->top = -1; | |
2669 else | |
2670 { | |
2671 int reg; | |
2672 | |
2673 output_stack->top = value_reg_high - value_reg_low; | |
2674 for (reg = value_reg_low; reg <= value_reg_high; ++reg) | |
2675 { | |
2676 output_stack->reg[value_reg_high - reg] = reg; | |
2677 SET_HARD_REG_BIT (output_stack->reg_set, reg); | |
2678 } | |
2679 } | |
2680 } | |
2681 | |
2682 /* Copy the stack info from the end of edge E's source block to the | |
2683 start of E's destination block. */ | |
2684 | |
2685 static void | |
2686 propagate_stack (edge e) | |
2687 { | |
2688 stack src_stack = &BLOCK_INFO (e->src)->stack_out; | |
2689 stack dest_stack = &BLOCK_INFO (e->dest)->stack_in; | |
2690 int reg; | |
2691 | |
2692 /* Preserve the order of the original stack, but check whether | |
2693 any pops are needed. */ | |
2694 dest_stack->top = -1; | |
2695 for (reg = 0; reg <= src_stack->top; ++reg) | |
2696 if (TEST_HARD_REG_BIT (dest_stack->reg_set, src_stack->reg[reg])) | |
2697 dest_stack->reg[++dest_stack->top] = src_stack->reg[reg]; | |
2698 | |
2699 /* Push in any partially dead values. */ | |
2700 for (reg = FIRST_STACK_REG; reg < LAST_STACK_REG + 1; reg++) | |
2701 if (TEST_HARD_REG_BIT (dest_stack->reg_set, reg) | |
2702 && !TEST_HARD_REG_BIT (src_stack->reg_set, reg)) | |
2703 dest_stack->reg[++dest_stack->top] = reg; | |
2704 } | |
2705 | |
2706 | |
2707 /* Adjust the stack of edge E's source block on exit to match the stack | |
2708 of it's target block upon input. The stack layouts of both blocks | |
2709 should have been defined by now. */ | |
2710 | |
2711 static bool | |
2712 compensate_edge (edge e) | |
2713 { | |
2714 basic_block source = e->src, target = e->dest; | |
2715 stack target_stack = &BLOCK_INFO (target)->stack_in; | |
2716 stack source_stack = &BLOCK_INFO (source)->stack_out; | |
2717 struct stack_def regstack; | |
2718 int reg; | |
2719 | |
2720 if (dump_file) | |
2721 fprintf (dump_file, "Edge %d->%d: ", source->index, target->index); | |
2722 | |
2723 gcc_assert (target_stack->top != -2); | |
2724 | |
2725 /* Check whether stacks are identical. */ | |
2726 if (target_stack->top == source_stack->top) | |
2727 { | |
2728 for (reg = target_stack->top; reg >= 0; --reg) | |
2729 if (target_stack->reg[reg] != source_stack->reg[reg]) | |
2730 break; | |
2731 | |
2732 if (reg == -1) | |
2733 { | |
2734 if (dump_file) | |
2735 fprintf (dump_file, "no changes needed\n"); | |
2736 return false; | |
2737 } | |
2738 } | |
2739 | |
2740 if (dump_file) | |
2741 { | |
2742 fprintf (dump_file, "correcting stack to "); | |
2743 print_stack (dump_file, target_stack); | |
2744 } | |
2745 | |
2746 /* Abnormal calls may appear to have values live in st(0), but the | |
2747 abnormal return path will not have actually loaded the values. */ | |
2748 if (e->flags & EDGE_ABNORMAL_CALL) | |
2749 { | |
2750 /* Assert that the lifetimes are as we expect -- one value | |
2751 live at st(0) on the end of the source block, and no | |
2752 values live at the beginning of the destination block. | |
2753 For complex return values, we may have st(1) live as well. */ | |
2754 gcc_assert (source_stack->top == 0 || source_stack->top == 1); | |
2755 gcc_assert (target_stack->top == -1); | |
2756 return false; | |
2757 } | |
2758 | |
2759 /* Handle non-call EH edges specially. The normal return path have | |
2760 values in registers. These will be popped en masse by the unwind | |
2761 library. */ | |
2762 if (e->flags & EDGE_EH) | |
2763 { | |
2764 gcc_assert (target_stack->top == -1); | |
2765 return false; | |
2766 } | |
2767 | |
2768 /* We don't support abnormal edges. Global takes care to | |
2769 avoid any live register across them, so we should never | |
2770 have to insert instructions on such edges. */ | |
2771 gcc_assert (! (e->flags & EDGE_ABNORMAL)); | |
2772 | |
2773 /* Make a copy of source_stack as change_stack is destructive. */ | |
2774 regstack = *source_stack; | |
2775 | |
2776 /* It is better to output directly to the end of the block | |
2777 instead of to the edge, because emit_swap can do minimal | |
2778 insn scheduling. We can do this when there is only one | |
2779 edge out, and it is not abnormal. */ | |
2780 if (EDGE_COUNT (source->succs) == 1) | |
2781 { | |
2782 current_block = source; | |
2783 change_stack (BB_END (source), ®stack, target_stack, | |
2784 (JUMP_P (BB_END (source)) ? EMIT_BEFORE : EMIT_AFTER)); | |
2785 } | |
2786 else | |
2787 { | |
2788 rtx seq, after; | |
2789 | |
2790 current_block = NULL; | |
2791 start_sequence (); | |
2792 | |
2793 /* ??? change_stack needs some point to emit insns after. */ | |
2794 after = emit_note (NOTE_INSN_DELETED); | |
2795 | |
2796 change_stack (after, ®stack, target_stack, EMIT_BEFORE); | |
2797 | |
2798 seq = get_insns (); | |
2799 end_sequence (); | |
2800 | |
2801 insert_insn_on_edge (seq, e); | |
2802 return true; | |
2803 } | |
2804 return false; | |
2805 } | |
2806 | |
2807 /* Traverse all non-entry edges in the CFG, and emit the necessary | |
2808 edge compensation code to change the stack from stack_out of the | |
2809 source block to the stack_in of the destination block. */ | |
2810 | |
2811 static bool | |
2812 compensate_edges (void) | |
2813 { | |
2814 bool inserted = false; | |
2815 basic_block bb; | |
2816 | |
2817 starting_stack_p = false; | |
2818 | |
2819 FOR_EACH_BB (bb) | |
2820 if (bb != ENTRY_BLOCK_PTR) | |
2821 { | |
2822 edge e; | |
2823 edge_iterator ei; | |
2824 | |
2825 FOR_EACH_EDGE (e, ei, bb->succs) | |
2826 inserted |= compensate_edge (e); | |
2827 } | |
2828 return inserted; | |
2829 } | |
2830 | |
2831 /* Select the better of two edges E1 and E2 to use to determine the | |
2832 stack layout for their shared destination basic block. This is | |
2833 typically the more frequently executed. The edge E1 may be NULL | |
2834 (in which case E2 is returned), but E2 is always non-NULL. */ | |
2835 | |
2836 static edge | |
2837 better_edge (edge e1, edge e2) | |
2838 { | |
2839 if (!e1) | |
2840 return e2; | |
2841 | |
2842 if (EDGE_FREQUENCY (e1) > EDGE_FREQUENCY (e2)) | |
2843 return e1; | |
2844 if (EDGE_FREQUENCY (e1) < EDGE_FREQUENCY (e2)) | |
2845 return e2; | |
2846 | |
2847 if (e1->count > e2->count) | |
2848 return e1; | |
2849 if (e1->count < e2->count) | |
2850 return e2; | |
2851 | |
2852 /* Prefer critical edges to minimize inserting compensation code on | |
2853 critical edges. */ | |
2854 | |
2855 if (EDGE_CRITICAL_P (e1) != EDGE_CRITICAL_P (e2)) | |
2856 return EDGE_CRITICAL_P (e1) ? e1 : e2; | |
2857 | |
2858 /* Avoid non-deterministic behavior. */ | |
2859 return (e1->src->index < e2->src->index) ? e1 : e2; | |
2860 } | |
2861 | |
2862 /* Convert stack register references in one block. */ | |
2863 | |
2864 static void | |
2865 convert_regs_1 (basic_block block) | |
2866 { | |
2867 struct stack_def regstack; | |
2868 block_info bi = BLOCK_INFO (block); | |
2869 int reg; | |
2870 rtx insn, next; | |
2871 bool control_flow_insn_deleted = false; | |
2872 | |
2873 any_malformed_asm = false; | |
2874 | |
2875 /* Choose an initial stack layout, if one hasn't already been chosen. */ | |
2876 if (bi->stack_in.top == -2) | |
2877 { | |
2878 edge e, beste = NULL; | |
2879 edge_iterator ei; | |
2880 | |
2881 /* Select the best incoming edge (typically the most frequent) to | |
2882 use as a template for this basic block. */ | |
2883 FOR_EACH_EDGE (e, ei, block->preds) | |
2884 if (BLOCK_INFO (e->src)->done) | |
2885 beste = better_edge (beste, e); | |
2886 | |
2887 if (beste) | |
2888 propagate_stack (beste); | |
2889 else | |
2890 { | |
2891 /* No predecessors. Create an arbitrary input stack. */ | |
2892 bi->stack_in.top = -1; | |
2893 for (reg = LAST_STACK_REG; reg >= FIRST_STACK_REG; --reg) | |
2894 if (TEST_HARD_REG_BIT (bi->stack_in.reg_set, reg)) | |
2895 bi->stack_in.reg[++bi->stack_in.top] = reg; | |
2896 } | |
2897 } | |
2898 | |
2899 if (dump_file) | |
2900 { | |
2901 fprintf (dump_file, "\nBasic block %d\nInput stack: ", block->index); | |
2902 print_stack (dump_file, &bi->stack_in); | |
2903 } | |
2904 | |
2905 /* Process all insns in this block. Keep track of NEXT so that we | |
2906 don't process insns emitted while substituting in INSN. */ | |
2907 current_block = block; | |
2908 next = BB_HEAD (block); | |
2909 regstack = bi->stack_in; | |
2910 starting_stack_p = true; | |
2911 | |
2912 do | |
2913 { | |
2914 insn = next; | |
2915 next = NEXT_INSN (insn); | |
2916 | |
2917 /* Ensure we have not missed a block boundary. */ | |
2918 gcc_assert (next); | |
2919 if (insn == BB_END (block)) | |
2920 next = NULL; | |
2921 | |
2922 /* Don't bother processing unless there is a stack reg | |
2923 mentioned or if it's a CALL_INSN. */ | |
2924 if (stack_regs_mentioned (insn) | |
2925 || CALL_P (insn)) | |
2926 { | |
2927 if (dump_file) | |
2928 { | |
2929 fprintf (dump_file, " insn %d input stack: ", | |
2930 INSN_UID (insn)); | |
2931 print_stack (dump_file, ®stack); | |
2932 } | |
2933 control_flow_insn_deleted |= subst_stack_regs (insn, ®stack); | |
2934 starting_stack_p = false; | |
2935 } | |
2936 } | |
2937 while (next); | |
2938 | |
2939 if (dump_file) | |
2940 { | |
2941 fprintf (dump_file, "Expected live registers ["); | |
2942 for (reg = FIRST_STACK_REG; reg <= LAST_STACK_REG; ++reg) | |
2943 if (TEST_HARD_REG_BIT (bi->out_reg_set, reg)) | |
2944 fprintf (dump_file, " %d", reg); | |
2945 fprintf (dump_file, " ]\nOutput stack: "); | |
2946 print_stack (dump_file, ®stack); | |
2947 } | |
2948 | |
2949 insn = BB_END (block); | |
2950 if (JUMP_P (insn)) | |
2951 insn = PREV_INSN (insn); | |
2952 | |
2953 /* If the function is declared to return a value, but it returns one | |
2954 in only some cases, some registers might come live here. Emit | |
2955 necessary moves for them. */ | |
2956 | |
2957 for (reg = FIRST_STACK_REG; reg <= LAST_STACK_REG; ++reg) | |
2958 { | |
2959 if (TEST_HARD_REG_BIT (bi->out_reg_set, reg) | |
2960 && ! TEST_HARD_REG_BIT (regstack.reg_set, reg)) | |
2961 { | |
2962 rtx set; | |
2963 | |
2964 if (dump_file) | |
2965 fprintf (dump_file, "Emitting insn initializing reg %d\n", reg); | |
2966 | |
2967 set = gen_rtx_SET (VOIDmode, FP_MODE_REG (reg, SFmode), not_a_num); | |
2968 insn = emit_insn_after (set, insn); | |
2969 control_flow_insn_deleted |= subst_stack_regs (insn, ®stack); | |
2970 } | |
2971 } | |
2972 | |
2973 /* Amongst the insns possibly deleted during the substitution process above, | |
2974 might have been the only trapping insn in the block. We purge the now | |
2975 possibly dead EH edges here to avoid an ICE from fixup_abnormal_edges, | |
2976 called at the end of convert_regs. The order in which we process the | |
2977 blocks ensures that we never delete an already processed edge. | |
2978 | |
2979 Note that, at this point, the CFG may have been damaged by the emission | |
2980 of instructions after an abnormal call, which moves the basic block end | |
2981 (and is the reason why we call fixup_abnormal_edges later). So we must | |
2982 be sure that the trapping insn has been deleted before trying to purge | |
2983 dead edges, otherwise we risk purging valid edges. | |
2984 | |
2985 ??? We are normally supposed not to delete trapping insns, so we pretend | |
2986 that the insns deleted above don't actually trap. It would have been | |
2987 better to detect this earlier and avoid creating the EH edge in the first | |
2988 place, still, but we don't have enough information at that time. */ | |
2989 | |
2990 if (control_flow_insn_deleted) | |
2991 purge_dead_edges (block); | |
2992 | |
2993 /* Something failed if the stack lives don't match. If we had malformed | |
2994 asms, we zapped the instruction itself, but that didn't produce the | |
2995 same pattern of register kills as before. */ | |
2996 | |
2997 gcc_assert (hard_reg_set_equal_p (regstack.reg_set, bi->out_reg_set) | |
2998 || any_malformed_asm); | |
2999 bi->stack_out = regstack; | |
3000 bi->done = true; | |
3001 } | |
3002 | |
3003 /* Convert registers in all blocks reachable from BLOCK. */ | |
3004 | |
3005 static void | |
3006 convert_regs_2 (basic_block block) | |
3007 { | |
3008 basic_block *stack, *sp; | |
3009 | |
3010 /* We process the blocks in a top-down manner, in a way such that one block | |
3011 is only processed after all its predecessors. The number of predecessors | |
3012 of every block has already been computed. */ | |
3013 | |
3014 stack = XNEWVEC (basic_block, n_basic_blocks); | |
3015 sp = stack; | |
3016 | |
3017 *sp++ = block; | |
3018 | |
3019 do | |
3020 { | |
3021 edge e; | |
3022 edge_iterator ei; | |
3023 | |
3024 block = *--sp; | |
3025 | |
3026 /* Processing BLOCK is achieved by convert_regs_1, which may purge | |
3027 some dead EH outgoing edge after the deletion of the trapping | |
3028 insn inside the block. Since the number of predecessors of | |
3029 BLOCK's successors was computed based on the initial edge set, | |
3030 we check the necessity to process some of these successors | |
3031 before such an edge deletion may happen. However, there is | |
3032 a pitfall: if BLOCK is the only predecessor of a successor and | |
3033 the edge between them happens to be deleted, the successor | |
3034 becomes unreachable and should not be processed. The problem | |
3035 is that there is no way to preventively detect this case so we | |
3036 stack the successor in all cases and hand over the task of | |
3037 fixing up the discrepancy to convert_regs_1. */ | |
3038 | |
3039 FOR_EACH_EDGE (e, ei, block->succs) | |
3040 if (! (e->flags & EDGE_DFS_BACK)) | |
3041 { | |
3042 BLOCK_INFO (e->dest)->predecessors--; | |
3043 if (!BLOCK_INFO (e->dest)->predecessors) | |
3044 *sp++ = e->dest; | |
3045 } | |
3046 | |
3047 convert_regs_1 (block); | |
3048 } | |
3049 while (sp != stack); | |
3050 | |
3051 free (stack); | |
3052 } | |
3053 | |
3054 /* Traverse all basic blocks in a function, converting the register | |
3055 references in each insn from the "flat" register file that gcc uses, | |
3056 to the stack-like registers the 387 uses. */ | |
3057 | |
3058 static void | |
3059 convert_regs (void) | |
3060 { | |
3061 int inserted; | |
3062 basic_block b; | |
3063 edge e; | |
3064 edge_iterator ei; | |
3065 | |
3066 /* Initialize uninitialized registers on function entry. */ | |
3067 inserted = convert_regs_entry (); | |
3068 | |
3069 /* Construct the desired stack for function exit. */ | |
3070 convert_regs_exit (); | |
3071 BLOCK_INFO (EXIT_BLOCK_PTR)->done = 1; | |
3072 | |
3073 /* ??? Future: process inner loops first, and give them arbitrary | |
3074 initial stacks which emit_swap_insn can modify. This ought to | |
3075 prevent double fxch that often appears at the head of a loop. */ | |
3076 | |
3077 /* Process all blocks reachable from all entry points. */ | |
3078 FOR_EACH_EDGE (e, ei, ENTRY_BLOCK_PTR->succs) | |
3079 convert_regs_2 (e->dest); | |
3080 | |
3081 /* ??? Process all unreachable blocks. Though there's no excuse | |
3082 for keeping these even when not optimizing. */ | |
3083 FOR_EACH_BB (b) | |
3084 { | |
3085 block_info bi = BLOCK_INFO (b); | |
3086 | |
3087 if (! bi->done) | |
3088 convert_regs_2 (b); | |
3089 } | |
3090 | |
3091 inserted |= compensate_edges (); | |
3092 | |
3093 clear_aux_for_blocks (); | |
3094 | |
3095 fixup_abnormal_edges (); | |
3096 if (inserted) | |
3097 commit_edge_insertions (); | |
3098 | |
3099 if (dump_file) | |
3100 fputc ('\n', dump_file); | |
3101 } | |
3102 | |
3103 /* Convert register usage from "flat" register file usage to a "stack | |
3104 register file. FILE is the dump file, if used. | |
3105 | |
3106 Construct a CFG and run life analysis. Then convert each insn one | |
3107 by one. Run a last cleanup_cfg pass, if optimizing, to eliminate | |
3108 code duplication created when the converter inserts pop insns on | |
3109 the edges. */ | |
3110 | |
3111 static bool | |
3112 reg_to_stack (void) | |
3113 { | |
3114 basic_block bb; | |
3115 int i; | |
3116 int max_uid; | |
3117 | |
3118 /* Clean up previous run. */ | |
3119 if (stack_regs_mentioned_data != NULL) | |
3120 VEC_free (char, heap, stack_regs_mentioned_data); | |
3121 | |
3122 /* See if there is something to do. Flow analysis is quite | |
3123 expensive so we might save some compilation time. */ | |
3124 for (i = FIRST_STACK_REG; i <= LAST_STACK_REG; i++) | |
3125 if (df_regs_ever_live_p (i)) | |
3126 break; | |
3127 if (i > LAST_STACK_REG) | |
3128 return false; | |
3129 | |
3130 df_note_add_problem (); | |
3131 df_analyze (); | |
3132 | |
3133 mark_dfs_back_edges (); | |
3134 | |
3135 /* Set up block info for each basic block. */ | |
3136 alloc_aux_for_blocks (sizeof (struct block_info_def)); | |
3137 FOR_EACH_BB (bb) | |
3138 { | |
3139 block_info bi = BLOCK_INFO (bb); | |
3140 edge_iterator ei; | |
3141 edge e; | |
3142 int reg; | |
3143 | |
3144 FOR_EACH_EDGE (e, ei, bb->preds) | |
3145 if (!(e->flags & EDGE_DFS_BACK) | |
3146 && e->src != ENTRY_BLOCK_PTR) | |
3147 bi->predecessors++; | |
3148 | |
3149 /* Set current register status at last instruction `uninitialized'. */ | |
3150 bi->stack_in.top = -2; | |
3151 | |
3152 /* Copy live_at_end and live_at_start into temporaries. */ | |
3153 for (reg = FIRST_STACK_REG; reg <= LAST_STACK_REG; reg++) | |
3154 { | |
3155 if (REGNO_REG_SET_P (DF_LR_OUT (bb), reg)) | |
3156 SET_HARD_REG_BIT (bi->out_reg_set, reg); | |
3157 if (REGNO_REG_SET_P (DF_LR_IN (bb), reg)) | |
3158 SET_HARD_REG_BIT (bi->stack_in.reg_set, reg); | |
3159 } | |
3160 } | |
3161 | |
3162 /* Create the replacement registers up front. */ | |
3163 for (i = FIRST_STACK_REG; i <= LAST_STACK_REG; i++) | |
3164 { | |
3165 enum machine_mode mode; | |
3166 for (mode = GET_CLASS_NARROWEST_MODE (MODE_FLOAT); | |
3167 mode != VOIDmode; | |
3168 mode = GET_MODE_WIDER_MODE (mode)) | |
3169 FP_MODE_REG (i, mode) = gen_rtx_REG (mode, i); | |
3170 for (mode = GET_CLASS_NARROWEST_MODE (MODE_COMPLEX_FLOAT); | |
3171 mode != VOIDmode; | |
3172 mode = GET_MODE_WIDER_MODE (mode)) | |
3173 FP_MODE_REG (i, mode) = gen_rtx_REG (mode, i); | |
3174 } | |
3175 | |
3176 ix86_flags_rtx = gen_rtx_REG (CCmode, FLAGS_REG); | |
3177 | |
3178 /* A QNaN for initializing uninitialized variables. | |
3179 | |
3180 ??? We can't load from constant memory in PIC mode, because | |
3181 we're inserting these instructions before the prologue and | |
3182 the PIC register hasn't been set up. In that case, fall back | |
3183 on zero, which we can get from `fldz'. */ | |
3184 | |
3185 if ((flag_pic && !TARGET_64BIT) | |
3186 || ix86_cmodel == CM_LARGE || ix86_cmodel == CM_LARGE_PIC) | |
3187 not_a_num = CONST0_RTX (SFmode); | |
3188 else | |
3189 { | |
3190 REAL_VALUE_TYPE r; | |
3191 | |
3192 real_nan (&r, "", 1, SFmode); | |
3193 not_a_num = CONST_DOUBLE_FROM_REAL_VALUE (r, SFmode); | |
3194 not_a_num = force_const_mem (SFmode, not_a_num); | |
3195 } | |
3196 | |
3197 /* Allocate a cache for stack_regs_mentioned. */ | |
3198 max_uid = get_max_uid (); | |
3199 stack_regs_mentioned_data = VEC_alloc (char, heap, max_uid + 1); | |
3200 memset (VEC_address (char, stack_regs_mentioned_data), | |
3201 0, sizeof (char) * (max_uid + 1)); | |
3202 | |
3203 convert_regs (); | |
3204 | |
3205 free_aux_for_blocks (); | |
3206 return true; | |
3207 } | |
3208 #endif /* STACK_REGS */ | |
3209 | |
3210 static bool | |
3211 gate_handle_stack_regs (void) | |
3212 { | |
3213 #ifdef STACK_REGS | |
3214 return 1; | |
3215 #else | |
3216 return 0; | |
3217 #endif | |
3218 } | |
3219 | |
3220 struct rtl_opt_pass pass_stack_regs = | |
3221 { | |
3222 { | |
3223 RTL_PASS, | |
3224 NULL, /* name */ | |
3225 gate_handle_stack_regs, /* gate */ | |
3226 NULL, /* execute */ | |
3227 NULL, /* sub */ | |
3228 NULL, /* next */ | |
3229 0, /* static_pass_number */ | |
3230 TV_REG_STACK, /* tv_id */ | |
3231 0, /* properties_required */ | |
3232 0, /* properties_provided */ | |
3233 0, /* properties_destroyed */ | |
3234 0, /* todo_flags_start */ | |
3235 0 /* todo_flags_finish */ | |
3236 } | |
3237 }; | |
3238 | |
3239 /* Convert register usage from flat register file usage to a stack | |
3240 register file. */ | |
3241 static unsigned int | |
3242 rest_of_handle_stack_regs (void) | |
3243 { | |
3244 #ifdef STACK_REGS | |
3245 reg_to_stack (); | |
3246 regstack_completed = 1; | |
3247 #endif | |
3248 return 0; | |
3249 } | |
3250 | |
3251 struct rtl_opt_pass pass_stack_regs_run = | |
3252 { | |
3253 { | |
3254 RTL_PASS, | |
3255 "stack", /* name */ | |
3256 NULL, /* gate */ | |
3257 rest_of_handle_stack_regs, /* execute */ | |
3258 NULL, /* sub */ | |
3259 NULL, /* next */ | |
3260 0, /* static_pass_number */ | |
3261 TV_REG_STACK, /* tv_id */ | |
3262 0, /* properties_required */ | |
3263 0, /* properties_provided */ | |
3264 0, /* properties_destroyed */ | |
3265 0, /* todo_flags_start */ | |
3266 TODO_df_finish | TODO_verify_rtl_sharing | | |
3267 TODO_dump_func | | |
3268 TODO_ggc_collect /* todo_flags_finish */ | |
3269 } | |
3270 }; |