]>
Commit | Line | Data |
---|---|---|
c7a0240a | 1 | /* Combine stack adjustments. |
a5544970 | 2 | Copyright (C) 1987-2019 Free Software Foundation, Inc. |
c7a0240a SB |
3 | |
4 | This file is part of GCC. | |
5 | ||
6 | GCC is free software; you can redistribute it and/or modify it under | |
7 | the terms of the GNU General Public License as published by the Free | |
9dcd6f09 | 8 | Software Foundation; either version 3, or (at your option) any later |
c7a0240a SB |
9 | version. |
10 | ||
11 | GCC is distributed in the hope that it will be useful, but WITHOUT ANY | |
12 | WARRANTY; without even the implied warranty of MERCHANTABILITY or | |
13 | FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License | |
14 | for more details. | |
15 | ||
16 | You should have received a copy of the GNU General Public License | |
9dcd6f09 NC |
17 | along with GCC; see the file COPYING3. If not see |
18 | <http://www.gnu.org/licenses/>. */ | |
c7a0240a SB |
19 | |
20 | /* Track stack adjustments and stack memory references. Attempt to | |
21 | reduce the number of stack adjustments by back-propagating across | |
22 | the memory references. | |
23 | ||
24 | This is intended primarily for use with targets that do not define | |
25 | ACCUMULATE_OUTGOING_ARGS. It is of significantly more value to | |
26 | targets that define PREFERRED_STACK_BOUNDARY more aligned than | |
27 | STACK_BOUNDARY (e.g. x86), or if not all registers can be pushed | |
28 | (e.g. x86 fp regs) which would ordinarily have to be implemented | |
29 | as a sub/mov pair due to restrictions in calls.c. | |
30 | ||
31 | Propagation stops when any of the insns that need adjusting are | |
32 | (a) no longer valid because we've exceeded their range, (b) a | |
33 | non-trivial push instruction, or (c) a call instruction. | |
34 | ||
35 | Restriction B is based on the assumption that push instructions | |
36 | are smaller or faster. If a port really wants to remove all | |
37 | pushes, it should have defined ACCUMULATE_OUTGOING_ARGS. The | |
38 | one exception that is made is for an add immediately followed | |
39 | by a push. */ | |
40 | ||
41 | #include "config.h" | |
42 | #include "system.h" | |
43 | #include "coretypes.h" | |
c7131fb2 | 44 | #include "backend.h" |
c7a0240a | 45 | #include "rtl.h" |
c7131fb2 | 46 | #include "df.h" |
c7a0240a | 47 | #include "insn-config.h" |
4d0cdd0c | 48 | #include "memmodel.h" |
957060b5 AM |
49 | #include "emit-rtl.h" |
50 | #include "recog.h" | |
60393bbc | 51 | #include "cfgrtl.h" |
c7a0240a | 52 | #include "tree-pass.h" |
f8305d18 | 53 | #include "rtl-iter.h" |
c7a0240a SB |
54 | |
55 | \f | |
90588a10 JJ |
56 | /* This structure records two kinds of stack references between stack |
57 | adjusting instructions: stack references in memory addresses for | |
58 | regular insns and all stack references for debug insns. */ | |
c7a0240a | 59 | |
90588a10 | 60 | struct csa_reflist |
c7a0240a SB |
61 | { |
62 | HOST_WIDE_INT sp_offset; | |
71e88baf DM |
63 | rtx_insn *insn; |
64 | rtx *ref; | |
90588a10 | 65 | struct csa_reflist *next; |
c7a0240a SB |
66 | }; |
67 | ||
68 | static int stack_memref_p (rtx); | |
71e88baf | 69 | static rtx single_set_for_csa (rtx_insn *); |
90588a10 | 70 | static void free_csa_reflist (struct csa_reflist *); |
71e88baf | 71 | static struct csa_reflist *record_one_stack_ref (rtx_insn *, rtx *, |
90588a10 | 72 | struct csa_reflist *); |
71e88baf | 73 | static int try_apply_stack_adjustment (rtx_insn *, struct csa_reflist *, |
c7a0240a SB |
74 | HOST_WIDE_INT, HOST_WIDE_INT); |
75 | static void combine_stack_adjustments_for_block (basic_block); | |
c7a0240a SB |
76 | |
77 | ||
78 | /* Main entry point for stack adjustment combination. */ | |
79 | ||
80 | static void | |
81 | combine_stack_adjustments (void) | |
82 | { | |
83 | basic_block bb; | |
84 | ||
11cd3bed | 85 | FOR_EACH_BB_FN (bb, cfun) |
c7a0240a SB |
86 | combine_stack_adjustments_for_block (bb); |
87 | } | |
88 | ||
89 | /* Recognize a MEM of the form (sp) or (plus sp const). */ | |
90 | ||
91 | static int | |
92 | stack_memref_p (rtx x) | |
93 | { | |
94 | if (!MEM_P (x)) | |
95 | return 0; | |
96 | x = XEXP (x, 0); | |
97 | ||
98 | if (x == stack_pointer_rtx) | |
99 | return 1; | |
100 | if (GET_CODE (x) == PLUS | |
101 | && XEXP (x, 0) == stack_pointer_rtx | |
481683e1 | 102 | && CONST_INT_P (XEXP (x, 1))) |
c7a0240a SB |
103 | return 1; |
104 | ||
105 | return 0; | |
106 | } | |
107 | ||
108 | /* Recognize either normal single_set or the hack in i386.md for | |
109 | tying fp and sp adjustments. */ | |
110 | ||
111 | static rtx | |
71e88baf | 112 | single_set_for_csa (rtx_insn *insn) |
c7a0240a SB |
113 | { |
114 | int i; | |
115 | rtx tmp = single_set (insn); | |
116 | if (tmp) | |
117 | return tmp; | |
118 | ||
119 | if (!NONJUMP_INSN_P (insn) | |
120 | || GET_CODE (PATTERN (insn)) != PARALLEL) | |
121 | return NULL_RTX; | |
122 | ||
123 | tmp = PATTERN (insn); | |
124 | if (GET_CODE (XVECEXP (tmp, 0, 0)) != SET) | |
125 | return NULL_RTX; | |
126 | ||
127 | for (i = 1; i < XVECLEN (tmp, 0); ++i) | |
128 | { | |
48c54229 | 129 | rtx this_rtx = XVECEXP (tmp, 0, i); |
c7a0240a SB |
130 | |
131 | /* The special case is allowing a no-op set. */ | |
48c54229 KG |
132 | if (GET_CODE (this_rtx) == SET |
133 | && SET_SRC (this_rtx) == SET_DEST (this_rtx)) | |
c7a0240a | 134 | ; |
48c54229 | 135 | else if (GET_CODE (this_rtx) != CLOBBER |
8df47bdf | 136 | && GET_CODE (this_rtx) != CLOBBER_HIGH |
48c54229 | 137 | && GET_CODE (this_rtx) != USE) |
c7a0240a SB |
138 | return NULL_RTX; |
139 | } | |
140 | ||
141 | return XVECEXP (tmp, 0, 0); | |
142 | } | |
143 | ||
90588a10 | 144 | /* Free the list of csa_reflist nodes. */ |
c7a0240a SB |
145 | |
146 | static void | |
90588a10 | 147 | free_csa_reflist (struct csa_reflist *reflist) |
c7a0240a | 148 | { |
90588a10 JJ |
149 | struct csa_reflist *next; |
150 | for (; reflist ; reflist = next) | |
c7a0240a | 151 | { |
90588a10 JJ |
152 | next = reflist->next; |
153 | free (reflist); | |
c7a0240a SB |
154 | } |
155 | } | |
156 | ||
90588a10 JJ |
157 | /* Create a new csa_reflist node from the given stack reference. |
158 | It is already known that the reference is either a MEM satisfying the | |
159 | predicate stack_memref_p or a REG representing the stack pointer. */ | |
c7a0240a | 160 | |
90588a10 | 161 | static struct csa_reflist * |
71e88baf | 162 | record_one_stack_ref (rtx_insn *insn, rtx *ref, struct csa_reflist *next_reflist) |
c7a0240a | 163 | { |
90588a10 | 164 | struct csa_reflist *ml; |
c7a0240a | 165 | |
90588a10 | 166 | ml = XNEW (struct csa_reflist); |
c7a0240a | 167 | |
90588a10 | 168 | if (REG_P (*ref) || XEXP (*ref, 0) == stack_pointer_rtx) |
c7a0240a SB |
169 | ml->sp_offset = 0; |
170 | else | |
90588a10 | 171 | ml->sp_offset = INTVAL (XEXP (XEXP (*ref, 0), 1)); |
c7a0240a SB |
172 | |
173 | ml->insn = insn; | |
90588a10 JJ |
174 | ml->ref = ref; |
175 | ml->next = next_reflist; | |
c7a0240a SB |
176 | |
177 | return ml; | |
178 | } | |
179 | ||
3dce0964 RH |
180 | /* We only know how to adjust the CFA; no other frame-related changes |
181 | may appear in any insn to be deleted. */ | |
182 | ||
183 | static bool | |
184 | no_unhandled_cfa (rtx_insn *insn) | |
185 | { | |
186 | if (!RTX_FRAME_RELATED_P (insn)) | |
187 | return true; | |
188 | ||
189 | /* No CFA notes at all is a legacy interpretation like | |
190 | FRAME_RELATED_EXPR, and is context sensitive within | |
191 | the prologue state machine. We can't handle that here. */ | |
192 | bool has_cfa_adjust = false; | |
193 | ||
194 | for (rtx link = REG_NOTES (insn); link; link = XEXP (link, 1)) | |
195 | switch (REG_NOTE_KIND (link)) | |
196 | { | |
197 | default: | |
198 | break; | |
199 | case REG_CFA_ADJUST_CFA: | |
200 | has_cfa_adjust = true; | |
201 | break; | |
202 | ||
203 | case REG_FRAME_RELATED_EXPR: | |
204 | case REG_CFA_DEF_CFA: | |
205 | case REG_CFA_OFFSET: | |
206 | case REG_CFA_REGISTER: | |
207 | case REG_CFA_EXPRESSION: | |
208 | case REG_CFA_RESTORE: | |
209 | case REG_CFA_SET_VDRAP: | |
210 | case REG_CFA_WINDOW_SAVE: | |
211 | case REG_CFA_FLUSH_QUEUE: | |
27169e45 | 212 | case REG_CFA_TOGGLE_RA_MANGLE: |
3dce0964 RH |
213 | return false; |
214 | } | |
215 | ||
216 | return has_cfa_adjust; | |
217 | } | |
218 | ||
c7a0240a | 219 | /* Attempt to apply ADJUST to the stack adjusting insn INSN, as well |
90588a10 JJ |
220 | as each of the memories and stack references in REFLIST. Return true |
221 | on success. */ | |
c7a0240a SB |
222 | |
223 | static int | |
71e88baf | 224 | try_apply_stack_adjustment (rtx_insn *insn, struct csa_reflist *reflist, |
90588a10 | 225 | HOST_WIDE_INT new_adjust, HOST_WIDE_INT delta) |
c7a0240a | 226 | { |
90588a10 | 227 | struct csa_reflist *ml; |
c7a0240a SB |
228 | rtx set; |
229 | ||
230 | set = single_set_for_csa (insn); | |
1362aa31 EB |
231 | if (MEM_P (SET_DEST (set))) |
232 | validate_change (insn, &SET_DEST (set), | |
233 | replace_equiv_address (SET_DEST (set), stack_pointer_rtx), | |
234 | 1); | |
235 | else | |
236 | validate_change (insn, &XEXP (SET_SRC (set), 1), GEN_INT (new_adjust), 1); | |
c7a0240a | 237 | |
90588a10 JJ |
238 | for (ml = reflist; ml ; ml = ml->next) |
239 | { | |
0a81f074 RS |
240 | rtx new_addr = plus_constant (Pmode, stack_pointer_rtx, |
241 | ml->sp_offset - delta); | |
90588a10 JJ |
242 | rtx new_val; |
243 | ||
244 | if (MEM_P (*ml->ref)) | |
245 | new_val = replace_equiv_address_nv (*ml->ref, new_addr); | |
246 | else if (GET_MODE (*ml->ref) == GET_MODE (stack_pointer_rtx)) | |
247 | new_val = new_addr; | |
248 | else | |
249 | new_val = lowpart_subreg (GET_MODE (*ml->ref), new_addr, | |
250 | GET_MODE (new_addr)); | |
251 | validate_change (ml->insn, ml->ref, new_val, 1); | |
252 | } | |
c7a0240a SB |
253 | |
254 | if (apply_change_group ()) | |
255 | { | |
90588a10 JJ |
256 | /* Succeeded. Update our knowledge of the stack references. */ |
257 | for (ml = reflist; ml ; ml = ml->next) | |
c7a0240a SB |
258 | ml->sp_offset -= delta; |
259 | ||
260 | return 1; | |
261 | } | |
262 | else | |
263 | return 0; | |
264 | } | |
265 | ||
f8305d18 RS |
266 | /* For non-debug insns, record all stack memory references in INSN |
267 | and return true if there were no other (unrecorded) references to the | |
268 | stack pointer. For debug insns, record all stack references regardless | |
269 | of context and unconditionally return true. */ | |
c7a0240a | 270 | |
f8305d18 RS |
271 | static bool |
272 | record_stack_refs (rtx_insn *insn, struct csa_reflist **reflist) | |
c7a0240a | 273 | { |
f8305d18 RS |
274 | subrtx_ptr_iterator::array_type array; |
275 | FOR_EACH_SUBRTX_PTR (iter, array, &PATTERN (insn), NONCONST) | |
c7a0240a | 276 | { |
f8305d18 RS |
277 | rtx *loc = *iter; |
278 | rtx x = *loc; | |
279 | switch (GET_CODE (x)) | |
90588a10 | 280 | { |
f8305d18 RS |
281 | case MEM: |
282 | if (!reg_mentioned_p (stack_pointer_rtx, x)) | |
283 | iter.skip_subrtxes (); | |
284 | /* We are not able to handle correctly all possible memrefs | |
285 | containing stack pointer, so this check is necessary. */ | |
286 | else if (stack_memref_p (x)) | |
287 | { | |
288 | *reflist = record_one_stack_ref (insn, loc, *reflist); | |
289 | iter.skip_subrtxes (); | |
290 | } | |
291 | /* Try harder for DEBUG_INSNs, handle e.g. | |
292 | (mem (mem (sp + 16) + 4). */ | |
293 | else if (!DEBUG_INSN_P (insn)) | |
294 | return false; | |
295 | break; | |
296 | ||
297 | case REG: | |
298 | /* ??? We want be able to handle non-memory stack pointer | |
299 | references later. For now just discard all insns referring to | |
300 | stack pointer outside mem expressions. We would probably | |
301 | want to teach validate_replace to simplify expressions first. | |
302 | ||
303 | We can't just compare with STACK_POINTER_RTX because the | |
304 | reference to the stack pointer might be in some other mode. | |
305 | In particular, an explicit clobber in an asm statement will | |
306 | result in a QImode clobber. | |
307 | ||
308 | In DEBUG_INSNs, we want to replace all occurrences, otherwise | |
309 | they will cause -fcompare-debug failures. */ | |
310 | if (REGNO (x) == STACK_POINTER_REGNUM) | |
311 | { | |
312 | if (!DEBUG_INSN_P (insn)) | |
313 | return false; | |
314 | *reflist = record_one_stack_ref (insn, loc, *reflist); | |
315 | } | |
316 | break; | |
317 | ||
318 | default: | |
319 | break; | |
90588a10 | 320 | } |
c7a0240a | 321 | } |
f8305d18 | 322 | return true; |
c7a0240a SB |
323 | } |
324 | ||
4efb91df RH |
325 | /* If INSN has a REG_ARGS_SIZE note, move it to LAST. |
326 | AFTER is true iff LAST follows INSN in the instruction stream. */ | |
a182fb6b JJ |
327 | |
328 | static void | |
71e88baf | 329 | maybe_move_args_size_note (rtx_insn *last, rtx_insn *insn, bool after) |
a182fb6b | 330 | { |
9a08d230 | 331 | rtx note, last_note; |
a182fb6b | 332 | |
9a08d230 RH |
333 | note = find_reg_note (insn, REG_ARGS_SIZE, NULL_RTX); |
334 | if (note == NULL) | |
a182fb6b JJ |
335 | return; |
336 | ||
9a08d230 RH |
337 | last_note = find_reg_note (last, REG_ARGS_SIZE, NULL_RTX); |
338 | if (last_note) | |
4efb91df RH |
339 | { |
340 | /* The ARGS_SIZE notes are *not* cumulative. They represent an | |
341 | absolute value, and the "most recent" note wins. */ | |
342 | if (!after) | |
343 | XEXP (last_note, 0) = XEXP (note, 0); | |
344 | } | |
a182fb6b | 345 | else |
9a08d230 | 346 | add_reg_note (last, REG_ARGS_SIZE, XEXP (note, 0)); |
a182fb6b JJ |
347 | } |
348 | ||
3dce0964 RH |
349 | /* Merge any REG_CFA_ADJUST_CFA note from SRC into DST. |
350 | AFTER is true iff DST follows SRC in the instruction stream. */ | |
351 | ||
352 | static void | |
353 | maybe_merge_cfa_adjust (rtx_insn *dst, rtx_insn *src, bool after) | |
354 | { | |
355 | rtx snote = NULL, dnote = NULL; | |
356 | rtx sexp, dexp; | |
357 | rtx exp1, exp2; | |
358 | ||
359 | if (RTX_FRAME_RELATED_P (src)) | |
360 | snote = find_reg_note (src, REG_CFA_ADJUST_CFA, NULL_RTX); | |
361 | if (snote == NULL) | |
362 | return; | |
363 | sexp = XEXP (snote, 0); | |
364 | ||
365 | if (RTX_FRAME_RELATED_P (dst)) | |
366 | dnote = find_reg_note (dst, REG_CFA_ADJUST_CFA, NULL_RTX); | |
367 | if (dnote == NULL) | |
368 | { | |
369 | add_reg_note (dst, REG_CFA_ADJUST_CFA, sexp); | |
370 | return; | |
371 | } | |
372 | dexp = XEXP (dnote, 0); | |
373 | ||
374 | gcc_assert (GET_CODE (sexp) == SET); | |
375 | gcc_assert (GET_CODE (dexp) == SET); | |
376 | ||
377 | if (after) | |
378 | exp1 = dexp, exp2 = sexp; | |
379 | else | |
380 | exp1 = sexp, exp2 = dexp; | |
381 | ||
382 | SET_SRC (exp1) = simplify_replace_rtx (SET_SRC (exp1), SET_DEST (exp2), | |
383 | SET_SRC (exp2)); | |
384 | XEXP (dnote, 0) = exp1; | |
385 | } | |
386 | ||
3825692d RH |
387 | /* Return the next (or previous) active insn within BB. */ |
388 | ||
71e88baf DM |
389 | static rtx_insn * |
390 | prev_active_insn_bb (basic_block bb, rtx_insn *insn) | |
3825692d RH |
391 | { |
392 | for (insn = PREV_INSN (insn); | |
393 | insn != PREV_INSN (BB_HEAD (bb)); | |
394 | insn = PREV_INSN (insn)) | |
395 | if (active_insn_p (insn)) | |
396 | return insn; | |
71e88baf | 397 | return NULL; |
3825692d RH |
398 | } |
399 | ||
71e88baf DM |
400 | static rtx_insn * |
401 | next_active_insn_bb (basic_block bb, rtx_insn *insn) | |
3825692d RH |
402 | { |
403 | for (insn = NEXT_INSN (insn); | |
404 | insn != NEXT_INSN (BB_END (bb)); | |
405 | insn = NEXT_INSN (insn)) | |
406 | if (active_insn_p (insn)) | |
407 | return insn; | |
71e88baf | 408 | return NULL; |
3825692d RH |
409 | } |
410 | ||
411 | /* If INSN has a REG_ARGS_SIZE note, if possible move it to PREV. Otherwise | |
412 | search for a nearby candidate within BB where we can stick the note. */ | |
413 | ||
414 | static void | |
71e88baf | 415 | force_move_args_size_note (basic_block bb, rtx_insn *prev, rtx_insn *insn) |
3825692d | 416 | { |
71e88baf DM |
417 | rtx note; |
418 | rtx_insn *test, *next_candidate, *prev_candidate; | |
3825692d RH |
419 | |
420 | /* If PREV exists, tail-call to the logic in the other function. */ | |
421 | if (prev) | |
422 | { | |
423 | maybe_move_args_size_note (prev, insn, false); | |
424 | return; | |
425 | } | |
426 | ||
427 | /* First, make sure there's anything that needs doing. */ | |
428 | note = find_reg_note (insn, REG_ARGS_SIZE, NULL_RTX); | |
429 | if (note == NULL) | |
430 | return; | |
431 | ||
432 | /* We need to find a spot between the previous and next exception points | |
433 | where we can place the note and "properly" deallocate the arguments. */ | |
434 | next_candidate = prev_candidate = NULL; | |
435 | ||
436 | /* It is often the case that we have insns in the order: | |
437 | call | |
438 | add sp (previous deallocation) | |
439 | sub sp (align for next arglist) | |
440 | push arg | |
441 | and the add/sub cancel. Therefore we begin by searching forward. */ | |
442 | ||
443 | test = insn; | |
444 | while ((test = next_active_insn_bb (bb, test)) != NULL) | |
445 | { | |
446 | /* Found an existing note: nothing to do. */ | |
447 | if (find_reg_note (test, REG_ARGS_SIZE, NULL_RTX)) | |
448 | return; | |
449 | /* Found something that affects unwinding. Stop searching. */ | |
450 | if (CALL_P (test) || !insn_nothrow_p (test)) | |
451 | break; | |
452 | if (next_candidate == NULL) | |
453 | next_candidate = test; | |
454 | } | |
455 | ||
456 | test = insn; | |
457 | while ((test = prev_active_insn_bb (bb, test)) != NULL) | |
458 | { | |
459 | rtx tnote; | |
460 | /* Found a place that seems logical to adjust the stack. */ | |
461 | tnote = find_reg_note (test, REG_ARGS_SIZE, NULL_RTX); | |
462 | if (tnote) | |
463 | { | |
464 | XEXP (tnote, 0) = XEXP (note, 0); | |
465 | return; | |
466 | } | |
467 | if (prev_candidate == NULL) | |
468 | prev_candidate = test; | |
469 | /* Found something that affects unwinding. Stop searching. */ | |
470 | if (CALL_P (test) || !insn_nothrow_p (test)) | |
471 | break; | |
472 | } | |
473 | ||
474 | if (prev_candidate) | |
475 | test = prev_candidate; | |
476 | else if (next_candidate) | |
477 | test = next_candidate; | |
478 | else | |
479 | { | |
480 | /* ??? We *must* have a place, lest we ICE on the lost adjustment. | |
481 | Options are: dummy clobber insn, nop, or prevent the removal of | |
65c2e636 RH |
482 | the sp += 0 insn. */ |
483 | /* TODO: Find another way to indicate to the dwarf2 code that we | |
484 | have not in fact lost an adjustment. */ | |
485 | test = emit_insn_before (gen_rtx_CLOBBER (VOIDmode, const0_rtx), insn); | |
3825692d RH |
486 | } |
487 | add_reg_note (test, REG_ARGS_SIZE, XEXP (note, 0)); | |
488 | } | |
489 | ||
c7a0240a SB |
490 | /* Subroutine of combine_stack_adjustments, called for each basic block. */ |
491 | ||
492 | static void | |
493 | combine_stack_adjustments_for_block (basic_block bb) | |
494 | { | |
495 | HOST_WIDE_INT last_sp_adjust = 0; | |
71e88baf DM |
496 | rtx_insn *last_sp_set = NULL; |
497 | rtx_insn *last2_sp_set = NULL; | |
90588a10 | 498 | struct csa_reflist *reflist = NULL; |
71e88baf DM |
499 | rtx_insn *insn, *next; |
500 | rtx set; | |
c7a0240a SB |
501 | bool end_of_block = false; |
502 | ||
503 | for (insn = BB_HEAD (bb); !end_of_block ; insn = next) | |
504 | { | |
505 | end_of_block = insn == BB_END (bb); | |
506 | next = NEXT_INSN (insn); | |
507 | ||
508 | if (! INSN_P (insn)) | |
509 | continue; | |
510 | ||
511 | set = single_set_for_csa (insn); | |
ac9c032e JL |
512 | if (set && find_reg_note (insn, REG_STACK_CHECK, NULL_RTX)) |
513 | set = NULL_RTX; | |
c7a0240a SB |
514 | if (set) |
515 | { | |
516 | rtx dest = SET_DEST (set); | |
517 | rtx src = SET_SRC (set); | |
518 | ||
519 | /* Find constant additions to the stack pointer. */ | |
520 | if (dest == stack_pointer_rtx | |
521 | && GET_CODE (src) == PLUS | |
522 | && XEXP (src, 0) == stack_pointer_rtx | |
481683e1 | 523 | && CONST_INT_P (XEXP (src, 1))) |
c7a0240a SB |
524 | { |
525 | HOST_WIDE_INT this_adjust = INTVAL (XEXP (src, 1)); | |
526 | ||
527 | /* If we've not seen an adjustment previously, record | |
528 | it now and continue. */ | |
529 | if (! last_sp_set) | |
530 | { | |
531 | last_sp_set = insn; | |
532 | last_sp_adjust = this_adjust; | |
533 | continue; | |
534 | } | |
535 | ||
90588a10 | 536 | /* If not all recorded refs can be adjusted, or the |
c7a0240a SB |
537 | adjustment is now too large for a constant addition, |
538 | we cannot merge the two stack adjustments. | |
539 | ||
540 | Also we need to be careful to not move stack pointer | |
541 | such that we create stack accesses outside the allocated | |
542 | area. We can combine an allocation into the first insn, | |
543 | or a deallocation into the second insn. We can not | |
544 | combine an allocation followed by a deallocation. | |
545 | ||
546 | The only somewhat frequent occurrence of the later is when | |
547 | a function allocates a stack frame but does not use it. | |
548 | For this case, we would need to analyze rtl stream to be | |
549 | sure that allocated area is really unused. This means not | |
550 | only checking the memory references, but also all registers | |
551 | or global memory references possibly containing a stack | |
552 | frame address. | |
553 | ||
554 | Perhaps the best way to address this problem is to teach | |
555 | gcc not to allocate stack for objects never used. */ | |
556 | ||
557 | /* Combine an allocation into the first instruction. */ | |
558 | if (STACK_GROWS_DOWNWARD ? this_adjust <= 0 : this_adjust >= 0) | |
559 | { | |
3dce0964 RH |
560 | if (no_unhandled_cfa (insn) |
561 | && try_apply_stack_adjustment (last_sp_set, reflist, | |
562 | last_sp_adjust | |
563 | + this_adjust, | |
564 | this_adjust)) | |
c7a0240a SB |
565 | { |
566 | /* It worked! */ | |
3825692d | 567 | maybe_move_args_size_note (last_sp_set, insn, false); |
3dce0964 | 568 | maybe_merge_cfa_adjust (last_sp_set, insn, false); |
c7a0240a SB |
569 | delete_insn (insn); |
570 | last_sp_adjust += this_adjust; | |
571 | continue; | |
572 | } | |
573 | } | |
574 | ||
575 | /* Otherwise we have a deallocation. Do not combine with | |
576 | a previous allocation. Combine into the second insn. */ | |
577 | else if (STACK_GROWS_DOWNWARD | |
578 | ? last_sp_adjust >= 0 : last_sp_adjust <= 0) | |
579 | { | |
3dce0964 RH |
580 | if (no_unhandled_cfa (last_sp_set) |
581 | && try_apply_stack_adjustment (insn, reflist, | |
582 | last_sp_adjust | |
583 | + this_adjust, | |
584 | -last_sp_adjust)) | |
c7a0240a SB |
585 | { |
586 | /* It worked! */ | |
3825692d | 587 | maybe_move_args_size_note (insn, last_sp_set, true); |
3dce0964 | 588 | maybe_merge_cfa_adjust (insn, last_sp_set, true); |
c7a0240a SB |
589 | delete_insn (last_sp_set); |
590 | last_sp_set = insn; | |
591 | last_sp_adjust += this_adjust; | |
90588a10 JJ |
592 | free_csa_reflist (reflist); |
593 | reflist = NULL; | |
c7a0240a SB |
594 | continue; |
595 | } | |
596 | } | |
597 | ||
598 | /* Combination failed. Restart processing from here. If | |
599 | deallocation+allocation conspired to cancel, we can | |
600 | delete the old deallocation insn. */ | |
3825692d RH |
601 | if (last_sp_set) |
602 | { | |
3dce0964 | 603 | if (last_sp_adjust == 0 && no_unhandled_cfa (last_sp_set)) |
3825692d RH |
604 | { |
605 | maybe_move_args_size_note (insn, last_sp_set, true); | |
3dce0964 | 606 | maybe_merge_cfa_adjust (insn, last_sp_set, true); |
3825692d RH |
607 | delete_insn (last_sp_set); |
608 | } | |
609 | else | |
610 | last2_sp_set = last_sp_set; | |
611 | } | |
90588a10 JJ |
612 | free_csa_reflist (reflist); |
613 | reflist = NULL; | |
c7a0240a SB |
614 | last_sp_set = insn; |
615 | last_sp_adjust = this_adjust; | |
616 | continue; | |
617 | } | |
618 | ||
1362aa31 EB |
619 | /* Find a store with pre-(dec|inc)rement or pre-modify of exactly |
620 | the previous adjustment and turn it into a simple store. This | |
621 | is equivalent to anticipating the stack adjustment so this must | |
622 | be an allocation. */ | |
623 | if (MEM_P (dest) | |
624 | && ((STACK_GROWS_DOWNWARD | |
625 | ? (GET_CODE (XEXP (dest, 0)) == PRE_DEC | |
cf098191 RS |
626 | && known_eq (last_sp_adjust, |
627 | GET_MODE_SIZE (GET_MODE (dest)))) | |
1362aa31 | 628 | : (GET_CODE (XEXP (dest, 0)) == PRE_INC |
cf098191 RS |
629 | && known_eq (-last_sp_adjust, |
630 | GET_MODE_SIZE (GET_MODE (dest))))) | |
1362aa31 EB |
631 | || ((STACK_GROWS_DOWNWARD |
632 | ? last_sp_adjust >= 0 : last_sp_adjust <= 0) | |
633 | && GET_CODE (XEXP (dest, 0)) == PRE_MODIFY | |
c7a0240a | 634 | && GET_CODE (XEXP (XEXP (dest, 0), 1)) == PLUS |
1362aa31 EB |
635 | && XEXP (XEXP (XEXP (dest, 0), 1), 0) |
636 | == stack_pointer_rtx | |
637 | && GET_CODE (XEXP (XEXP (XEXP (dest, 0), 1), 1)) | |
638 | == CONST_INT | |
639 | && INTVAL (XEXP (XEXP (XEXP (dest, 0), 1), 1)) | |
640 | == -last_sp_adjust)) | |
c7a0240a | 641 | && XEXP (XEXP (dest, 0), 0) == stack_pointer_rtx |
1362aa31 | 642 | && !reg_mentioned_p (stack_pointer_rtx, src) |
c7a0240a | 643 | && memory_address_p (GET_MODE (dest), stack_pointer_rtx) |
1362aa31 EB |
644 | && try_apply_stack_adjustment (insn, reflist, 0, |
645 | -last_sp_adjust)) | |
c7a0240a | 646 | { |
3825692d RH |
647 | if (last2_sp_set) |
648 | maybe_move_args_size_note (last2_sp_set, last_sp_set, false); | |
649 | else | |
650 | maybe_move_args_size_note (insn, last_sp_set, true); | |
c7a0240a | 651 | delete_insn (last_sp_set); |
90588a10 JJ |
652 | free_csa_reflist (reflist); |
653 | reflist = NULL; | |
71e88baf | 654 | last_sp_set = NULL; |
c7a0240a SB |
655 | last_sp_adjust = 0; |
656 | continue; | |
657 | } | |
658 | } | |
659 | ||
c7a0240a | 660 | if (!CALL_P (insn) && last_sp_set |
f8305d18 RS |
661 | && record_stack_refs (insn, &reflist)) |
662 | continue; | |
c7a0240a SB |
663 | |
664 | /* Otherwise, we were not able to process the instruction. | |
665 | Do not continue collecting data across such a one. */ | |
666 | if (last_sp_set | |
667 | && (CALL_P (insn) | |
668 | || reg_mentioned_p (stack_pointer_rtx, PATTERN (insn)))) | |
669 | { | |
670 | if (last_sp_set && last_sp_adjust == 0) | |
3825692d RH |
671 | { |
672 | force_move_args_size_note (bb, last2_sp_set, last_sp_set); | |
673 | delete_insn (last_sp_set); | |
674 | } | |
90588a10 JJ |
675 | free_csa_reflist (reflist); |
676 | reflist = NULL; | |
71e88baf DM |
677 | last2_sp_set = NULL; |
678 | last_sp_set = NULL; | |
c7a0240a SB |
679 | last_sp_adjust = 0; |
680 | } | |
681 | } | |
682 | ||
683 | if (last_sp_set && last_sp_adjust == 0) | |
3825692d RH |
684 | { |
685 | force_move_args_size_note (bb, last2_sp_set, last_sp_set); | |
686 | delete_insn (last_sp_set); | |
687 | } | |
c7a0240a | 688 | |
90588a10 JJ |
689 | if (reflist) |
690 | free_csa_reflist (reflist); | |
c7a0240a SB |
691 | } |
692 | \f | |
11a687e7 RH |
693 | static unsigned int |
694 | rest_of_handle_stack_adjustments (void) | |
695 | { | |
696 | df_note_add_problem (); | |
697 | df_analyze (); | |
698 | combine_stack_adjustments (); | |
c7a0240a SB |
699 | return 0; |
700 | } | |
701 | ||
27a4cd48 DM |
702 | namespace { |
703 | ||
704 | const pass_data pass_data_stack_adjustments = | |
c7a0240a | 705 | { |
27a4cd48 DM |
706 | RTL_PASS, /* type */ |
707 | "csa", /* name */ | |
708 | OPTGROUP_NONE, /* optinfo_flags */ | |
27a4cd48 DM |
709 | TV_COMBINE_STACK_ADJUST, /* tv_id */ |
710 | 0, /* properties_required */ | |
711 | 0, /* properties_provided */ | |
712 | 0, /* properties_destroyed */ | |
713 | 0, /* todo_flags_start */ | |
3bea341f | 714 | TODO_df_finish, /* todo_flags_finish */ |
c7a0240a | 715 | }; |
27a4cd48 DM |
716 | |
717 | class pass_stack_adjustments : public rtl_opt_pass | |
718 | { | |
719 | public: | |
c3284718 RS |
720 | pass_stack_adjustments (gcc::context *ctxt) |
721 | : rtl_opt_pass (pass_data_stack_adjustments, ctxt) | |
27a4cd48 DM |
722 | {} |
723 | ||
724 | /* opt_pass methods: */ | |
1a3d085c | 725 | virtual bool gate (function *); |
be55bfe6 TS |
726 | virtual unsigned int execute (function *) |
727 | { | |
728 | return rest_of_handle_stack_adjustments (); | |
729 | } | |
27a4cd48 DM |
730 | |
731 | }; // class pass_stack_adjustments | |
732 | ||
1a3d085c TS |
733 | bool |
734 | pass_stack_adjustments::gate (function *) | |
735 | { | |
736 | /* This is kind of a heuristic. We need to run combine_stack_adjustments | |
737 | even for machines with possibly nonzero TARGET_RETURN_POPS_ARGS | |
738 | and ACCUMULATE_OUTGOING_ARGS. We expect that only ports having | |
739 | push instructions will have popping returns. */ | |
740 | #ifndef PUSH_ROUNDING | |
741 | if (ACCUMULATE_OUTGOING_ARGS) | |
742 | return false; | |
743 | #endif | |
744 | return flag_combine_stack_adjustments; | |
745 | } | |
746 | ||
27a4cd48 DM |
747 | } // anon namespace |
748 | ||
749 | rtl_opt_pass * | |
750 | make_pass_stack_adjustments (gcc::context *ctxt) | |
751 | { | |
752 | return new pass_stack_adjustments (ctxt); | |
753 | } |