* c-common.c (get_priority): Add check for
[official-gcc.git] / gcc / fwprop.c
blob6ce91836d450f39a4bb78aef36acf58d838ab95d
1 /* RTL-based forward propagation pass for GNU compiler.
2 Copyright (C) 2005, 2006 Free Software Foundation, Inc.
3 Contributed by Paolo Bonzini and Steven Bosscher.
5 This file is part of GCC.
7 GCC is free software; you can redistribute it and/or modify it under
8 the terms of the GNU General Public License as published by the Free
9 Software Foundation; either version 2, or (at your option) any later
10 version.
12 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
13 WARRANTY; without even the implied warranty of MERCHANTABILITY or
14 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
15 for more details.
17 You should have received a copy of the GNU General Public License
18 along with GCC; see the file COPYING. If not, write to the Free
19 Software Foundation, 51 Franklin Street, Fifth Floor, Boston, MA
20 02110-1301, USA. */
22 #include "config.h"
23 #include "system.h"
24 #include "coretypes.h"
25 #include "tm.h"
26 #include "toplev.h"
28 #include "timevar.h"
29 #include "rtl.h"
30 #include "tm_p.h"
31 #include "emit-rtl.h"
32 #include "insn-config.h"
33 #include "recog.h"
34 #include "flags.h"
35 #include "obstack.h"
36 #include "basic-block.h"
37 #include "output.h"
38 #include "df.h"
39 #include "target.h"
40 #include "cfgloop.h"
41 #include "tree-pass.h"
44 /* This pass does simple forward propagation and simplification when an
45 operand of an insn can only come from a single def. This pass uses
46 df.c, so it is global. However, we only do limited analysis of
47 available expressions.
49 1) The pass tries to propagate the source of the def into the use,
50 and checks if the result is independent of the substituted value.
51 For example, the high word of a (zero_extend:DI (reg:SI M)) is always
52 zero, independent of the source register.
54 In particular, we propagate constants into the use site. Sometimes
55 RTL expansion did not put the constant in the same insn on purpose,
56 to satisfy a predicate, and the result will fail to be recognized;
57 but this happens rarely and in this case we can still create a
58 REG_EQUAL note. For multi-word operations, this
60 (set (subreg:SI (reg:DI 120) 0) (const_int 0))
61 (set (subreg:SI (reg:DI 120) 4) (const_int -1))
62 (set (subreg:SI (reg:DI 122) 0)
63 (ior:SI (subreg:SI (reg:DI 119) 0) (subreg:SI (reg:DI 120) 0)))
64 (set (subreg:SI (reg:DI 122) 4)
65 (ior:SI (subreg:SI (reg:DI 119) 4) (subreg:SI (reg:DI 120) 4)))
67 can be simplified to the much simpler
69 (set (subreg:SI (reg:DI 122) 0) (subreg:SI (reg:DI 119)))
70 (set (subreg:SI (reg:DI 122) 4) (const_int -1))
72 This particular propagation is also effective at putting together
73 complex addressing modes. We are more aggressive inside MEMs, in
74 that all definitions are propagated if the use is in a MEM; if the
75 result is a valid memory address we check address_cost to decide
76 whether the substitution is worthwhile.
78 2) The pass propagates register copies. This is not as effective as
79 the copy propagation done by CSE's canon_reg, which works by walking
80 the instruction chain, it can help the other transformations.
82 We should consider removing this optimization, and instead reorder the
83 RTL passes, because GCSE does this transformation too. With some luck,
84 the CSE pass at the end of rest_of_handle_gcse could also go away.
86 3) The pass looks for paradoxical subregs that are actually unnecessary.
87 Things like this:
89 (set (reg:QI 120) (subreg:QI (reg:SI 118) 0))
90 (set (reg:QI 121) (subreg:QI (reg:SI 119) 0))
91 (set (reg:SI 122) (plus:SI (subreg:SI (reg:QI 120) 0)
92 (subreg:SI (reg:QI 121) 0)))
94 are very common on machines that can only do word-sized operations.
95 For each use of a paradoxical subreg (subreg:WIDER (reg:NARROW N) 0),
96 if it has a single def and it is (subreg:NARROW (reg:WIDE M) 0),
97 we can replace the paradoxical subreg with simply (reg:WIDE M). The
98 above will simplify this to
100 (set (reg:QI 120) (subreg:QI (reg:SI 118) 0))
101 (set (reg:QI 121) (subreg:QI (reg:SI 119) 0))
102 (set (reg:SI 122) (plus:SI (reg:SI 118) (reg:SI 119)))
104 where the first two insns are now dead. */
107 static struct df *df;
108 static int num_changes;
111 /* Do not try to replace constant addresses or addresses of local and
112 argument slots. These MEM expressions are made only once and inserted
113 in many instructions, as well as being used to control symbol table
114 output. It is not safe to clobber them.
116 There are some uncommon cases where the address is already in a register
117 for some reason, but we cannot take advantage of that because we have
118 no easy way to unshare the MEM. In addition, looking up all stack
119 addresses is costly. */
121 static bool
122 can_simplify_addr (rtx addr)
124 rtx reg;
126 if (CONSTANT_ADDRESS_P (addr))
127 return false;
129 if (GET_CODE (addr) == PLUS)
130 reg = XEXP (addr, 0);
131 else
132 reg = addr;
134 return (!REG_P (reg)
135 || (REGNO (reg) != FRAME_POINTER_REGNUM
136 && REGNO (reg) != HARD_FRAME_POINTER_REGNUM
137 && REGNO (reg) != ARG_POINTER_REGNUM));
140 /* Returns a canonical version of X for the address, from the point of view,
141 that all multiplications are represented as MULT instead of the multiply
142 by a power of 2 being represented as ASHIFT.
144 Every ASHIFT we find has been made by simplify_gen_binary and was not
145 there before, so it is not shared. So we can do this in place. */
147 static void
148 canonicalize_address (rtx x)
150 for (;;)
151 switch (GET_CODE (x))
153 case ASHIFT:
154 if (GET_CODE (XEXP (x, 1)) == CONST_INT
155 && INTVAL (XEXP (x, 1)) < GET_MODE_BITSIZE (GET_MODE (x))
156 && INTVAL (XEXP (x, 1)) >= 0)
158 HOST_WIDE_INT shift = INTVAL (XEXP (x, 1));
159 PUT_CODE (x, MULT);
160 XEXP (x, 1) = gen_int_mode ((HOST_WIDE_INT) 1 << shift,
161 GET_MODE (x));
164 x = XEXP (x, 0);
165 break;
167 case PLUS:
168 if (GET_CODE (XEXP (x, 0)) == PLUS
169 || GET_CODE (XEXP (x, 0)) == ASHIFT
170 || GET_CODE (XEXP (x, 0)) == CONST)
171 canonicalize_address (XEXP (x, 0));
173 x = XEXP (x, 1);
174 break;
176 case CONST:
177 x = XEXP (x, 0);
178 break;
180 default:
181 return;
185 /* OLD is a memory address. Return whether it is good to use NEW instead,
186 for a memory access in the given MODE. */
188 static bool
189 should_replace_address (rtx old, rtx new, enum machine_mode mode)
191 int gain;
193 if (rtx_equal_p (old, new) || !memory_address_p (mode, new))
194 return false;
196 /* Copy propagation is always ok. */
197 if (REG_P (old) && REG_P (new))
198 return true;
200 /* Prefer the new address if it is less expensive. */
201 gain = address_cost (old, mode) - address_cost (new, mode);
203 /* If the addresses have equivalent cost, prefer the new address
204 if it has the highest `rtx_cost'. That has the potential of
205 eliminating the most insns without additional costs, and it
206 is the same that cse.c used to do. */
207 if (gain == 0)
208 gain = rtx_cost (new, SET) - rtx_cost (old, SET);
210 return (gain > 0);
213 /* Replace all occurrences of OLD in *PX with NEW and try to simplify the
214 resulting expression. Replace *PX with a new RTL expression if an
215 occurrence of OLD was found.
217 If CAN_APPEAR is true, we always return true; if it is false, we
218 can return false if, for at least one occurrence OLD, we failed to
219 collapse the result to a constant. For example, (mult:M (reg:M A)
220 (minus:M (reg:M B) (reg:M A))) may collapse to zero if replacing
221 (reg:M B) with (reg:M A).
223 CAN_APPEAR is disregarded inside MEMs: in that case, we always return
224 true if the simplification is a cheaper and valid memory address.
226 This is only a wrapper around simplify-rtx.c: do not add any pattern
227 matching code here. (The sole exception is the handling of LO_SUM, but
228 that is because there is no simplify_gen_* function for LO_SUM). */
230 static bool
231 propagate_rtx_1 (rtx *px, rtx old, rtx new, bool can_appear)
233 rtx x = *px, tem = NULL_RTX, op0, op1, op2;
234 enum rtx_code code = GET_CODE (x);
235 enum machine_mode mode = GET_MODE (x);
236 enum machine_mode op_mode;
237 bool valid_ops = true;
239 /* If X is OLD_RTX, return NEW_RTX. Otherwise, if this is an expression,
240 try to build a new expression from recursive substitution. */
242 if (x == old)
244 *px = new;
245 return can_appear;
248 switch (GET_RTX_CLASS (code))
250 case RTX_UNARY:
251 op0 = XEXP (x, 0);
252 op_mode = GET_MODE (op0);
253 valid_ops &= propagate_rtx_1 (&op0, old, new, can_appear);
254 if (op0 == XEXP (x, 0))
255 return true;
256 tem = simplify_gen_unary (code, mode, op0, op_mode);
257 break;
259 case RTX_BIN_ARITH:
260 case RTX_COMM_ARITH:
261 op0 = XEXP (x, 0);
262 op1 = XEXP (x, 1);
263 valid_ops &= propagate_rtx_1 (&op0, old, new, can_appear);
264 valid_ops &= propagate_rtx_1 (&op1, old, new, can_appear);
265 if (op0 == XEXP (x, 0) && op1 == XEXP (x, 1))
266 return true;
267 tem = simplify_gen_binary (code, mode, op0, op1);
268 break;
270 case RTX_COMPARE:
271 case RTX_COMM_COMPARE:
272 op0 = XEXP (x, 0);
273 op1 = XEXP (x, 1);
274 op_mode = GET_MODE (op0) != VOIDmode ? GET_MODE (op0) : GET_MODE (op1);
275 valid_ops &= propagate_rtx_1 (&op0, old, new, can_appear);
276 valid_ops &= propagate_rtx_1 (&op1, old, new, can_appear);
277 if (op0 == XEXP (x, 0) && op1 == XEXP (x, 1))
278 return true;
279 tem = simplify_gen_relational (code, mode, op_mode, op0, op1);
280 break;
282 case RTX_TERNARY:
283 case RTX_BITFIELD_OPS:
284 op0 = XEXP (x, 0);
285 op1 = XEXP (x, 1);
286 op2 = XEXP (x, 2);
287 op_mode = GET_MODE (op0);
288 valid_ops &= propagate_rtx_1 (&op0, old, new, can_appear);
289 valid_ops &= propagate_rtx_1 (&op1, old, new, can_appear);
290 valid_ops &= propagate_rtx_1 (&op2, old, new, can_appear);
291 if (op0 == XEXP (x, 0) && op1 == XEXP (x, 1) && op2 == XEXP (x, 2))
292 return true;
293 if (op_mode == VOIDmode)
294 op_mode = GET_MODE (op0);
295 tem = simplify_gen_ternary (code, mode, op_mode, op0, op1, op2);
296 break;
298 case RTX_EXTRA:
299 /* The only case we try to handle is a SUBREG. */
300 if (code == SUBREG)
302 op0 = XEXP (x, 0);
303 valid_ops &= propagate_rtx_1 (&op0, old, new, can_appear);
304 if (op0 == XEXP (x, 0))
305 return true;
306 tem = simplify_gen_subreg (mode, op0, GET_MODE (SUBREG_REG (x)),
307 SUBREG_BYTE (x));
309 break;
311 case RTX_OBJ:
312 if (code == MEM && x != new)
314 rtx new_op0;
315 op0 = XEXP (x, 0);
317 /* There are some addresses that we cannot work on. */
318 if (!can_simplify_addr (op0))
319 return true;
321 op0 = new_op0 = targetm.delegitimize_address (op0);
322 valid_ops &= propagate_rtx_1 (&new_op0, old, new, true);
324 /* Dismiss transformation that we do not want to carry on. */
325 if (!valid_ops
326 || new_op0 == op0
327 || !(GET_MODE (new_op0) == GET_MODE (op0)
328 || GET_MODE (new_op0) == VOIDmode))
329 return true;
331 canonicalize_address (new_op0);
333 /* Copy propagations are always ok. Otherwise check the costs. */
334 if (!(REG_P (old) && REG_P (new))
335 && !should_replace_address (op0, new_op0, GET_MODE (x)))
336 return true;
338 tem = replace_equiv_address_nv (x, new_op0);
341 else if (code == LO_SUM)
343 op0 = XEXP (x, 0);
344 op1 = XEXP (x, 1);
346 /* The only simplification we do attempts to remove references to op0
347 or make it constant -- in both cases, op0's invalidity will not
348 make the result invalid. */
349 propagate_rtx_1 (&op0, old, new, true);
350 valid_ops &= propagate_rtx_1 (&op1, old, new, can_appear);
351 if (op0 == XEXP (x, 0) && op1 == XEXP (x, 1))
352 return true;
354 /* (lo_sum (high x) x) -> x */
355 if (GET_CODE (op0) == HIGH && rtx_equal_p (XEXP (op0, 0), op1))
356 tem = op1;
357 else
358 tem = gen_rtx_LO_SUM (mode, op0, op1);
360 /* OP1 is likely not a legitimate address, otherwise there would have
361 been no LO_SUM. We want it to disappear if it is invalid, return
362 false in that case. */
363 return memory_address_p (mode, tem);
366 else if (code == REG)
368 if (rtx_equal_p (x, old))
370 *px = new;
371 return can_appear;
374 break;
376 default:
377 break;
380 /* No change, no trouble. */
381 if (tem == NULL_RTX)
382 return true;
384 *px = tem;
386 /* The replacement we made so far is valid, if all of the recursive
387 replacements were valid, or we could simplify everything to
388 a constant. */
389 return valid_ops || can_appear || CONSTANT_P (tem);
392 /* Replace all occurrences of OLD in X with NEW and try to simplify the
393 resulting expression (in mode MODE). Return a new expression if it is
394 a constant, otherwise X.
396 Simplifications where occurrences of NEW collapse to a constant are always
397 accepted. All simplifications are accepted if NEW is a pseudo too.
398 Otherwise, we accept simplifications that have a lower or equal cost. */
400 static rtx
401 propagate_rtx (rtx x, enum machine_mode mode, rtx old, rtx new)
403 rtx tem;
404 bool collapsed;
406 if (REG_P (new) && REGNO (new) < FIRST_PSEUDO_REGISTER)
407 return NULL_RTX;
409 new = copy_rtx (new);
411 tem = x;
412 collapsed = propagate_rtx_1 (&tem, old, new, REG_P (new) || CONSTANT_P (new));
413 if (tem == x || !collapsed)
414 return NULL_RTX;
416 /* gen_lowpart_common will not be able to process VOIDmode entities other
417 than CONST_INTs. */
418 if (GET_MODE (tem) == VOIDmode && GET_CODE (tem) != CONST_INT)
419 return NULL_RTX;
421 if (GET_MODE (tem) == VOIDmode)
422 tem = rtl_hooks.gen_lowpart_no_emit (mode, tem);
423 else
424 gcc_assert (GET_MODE (tem) == mode);
426 return tem;
432 /* Return true if the register from reference REF is killed
433 between FROM to (but not including) TO. */
435 static bool
436 local_ref_killed_between_p (struct df_ref * ref, rtx from, rtx to)
438 rtx insn;
439 struct df_ref *def;
441 for (insn = from; insn != to; insn = NEXT_INSN (insn))
443 if (!INSN_P (insn))
444 continue;
446 def = DF_INSN_DEFS (df, insn);
447 while (def)
449 if (DF_REF_REGNO (ref) == DF_REF_REGNO (def))
450 return true;
451 def = def->next_ref;
454 return false;
458 /* Check if the given DEF is available in INSN. This would require full
459 computation of available expressions; we check only restricted conditions:
460 - if DEF is the sole definition of its register, go ahead;
461 - in the same basic block, we check for no definitions killing the
462 definition of DEF_INSN;
463 - if USE's basic block has DEF's basic block as the sole predecessor,
464 we check if the definition is killed after DEF_INSN or before
465 TARGET_INSN insn, in their respective basic blocks. */
466 static bool
467 use_killed_between (struct df_ref *use, rtx def_insn, rtx target_insn)
469 basic_block def_bb = BLOCK_FOR_INSN (def_insn);
470 basic_block target_bb = BLOCK_FOR_INSN (target_insn);
471 int regno;
472 struct df_ref * def;
474 /* In some obscure situations we can have a def reaching a use
475 that is _before_ the def. In other words the def does not
476 dominate the use even though the use and def are in the same
477 basic block. This can happen when a register may be used
478 uninitialized in a loop. In such cases, we must assume that
479 DEF is not available. */
480 if (def_bb == target_bb
481 ? DF_INSN_LUID (df, def_insn) >= DF_INSN_LUID (df, target_insn)
482 : !dominated_by_p (CDI_DOMINATORS, target_bb, def_bb))
483 return true;
485 /* Check if the reg in USE has only one definition. We already
486 know that this definition reaches use, or we wouldn't be here. */
487 regno = DF_REF_REGNO (use);
488 def = DF_REG_DEF_GET (df, regno)->reg_chain;
489 if (def && (def->next_reg == NULL))
490 return false;
492 /* Check locally if we are in the same basic block. */
493 if (def_bb == target_bb)
494 return local_ref_killed_between_p (use, def_insn, target_insn);
496 /* Finally, if DEF_BB is the sole predecessor of TARGET_BB. */
497 if (single_pred_p (target_bb)
498 && single_pred (target_bb) == def_bb)
500 struct df_ref *x;
502 /* See if USE is killed between DEF_INSN and the last insn in the
503 basic block containing DEF_INSN. */
504 x = df_bb_regno_last_def_find (df, def_bb, regno);
505 if (x && DF_INSN_LUID (df, x->insn) >= DF_INSN_LUID (df, def_insn))
506 return true;
508 /* See if USE is killed between TARGET_INSN and the first insn in the
509 basic block containing TARGET_INSN. */
510 x = df_bb_regno_first_def_find (df, target_bb, regno);
511 if (x && DF_INSN_LUID (df, x->insn) < DF_INSN_LUID (df, target_insn))
512 return true;
514 return false;
517 /* Otherwise assume the worst case. */
518 return true;
522 /* for_each_rtx traversal function that returns 1 if BODY points to
523 a non-constant mem. */
525 static int
526 varying_mem_p (rtx *body, void *data ATTRIBUTE_UNUSED)
528 rtx x = *body;
529 return MEM_P (x) && !MEM_READONLY_P (x);
532 /* Check if all uses in DEF_INSN can be used in TARGET_INSN. This
533 would require full computation of available expressions;
534 we check only restricted conditions, see use_killed_between. */
535 static bool
536 all_uses_available_at (rtx def_insn, rtx target_insn)
538 struct df_ref * use;
539 rtx def_set = single_set (def_insn);
541 gcc_assert (def_set);
543 /* If target_insn comes right after def_insn, which is very common
544 for addresses, we can use a quicker test. */
545 if (NEXT_INSN (def_insn) == target_insn
546 && REG_P (SET_DEST (def_set)))
548 rtx def_reg = SET_DEST (def_set);
550 /* If the insn uses the reg that it defines, the substitution is
551 invalid. */
552 for (use = DF_INSN_USES (df, def_insn); use; use = use->next_ref)
553 if (rtx_equal_p (use->reg, def_reg))
554 return false;
556 else
558 /* Look at all the uses of DEF_INSN, and see if they are not
559 killed between DEF_INSN and TARGET_INSN. */
560 for (use = DF_INSN_USES (df, def_insn); use; use = use->next_ref)
561 if (use_killed_between (use, def_insn, target_insn))
562 return false;
565 /* We don't do any analysis of memories or aliasing. Reject any
566 instruction that involves references to non-constant memory. */
567 return !for_each_rtx (&SET_SRC (def_set), varying_mem_p, NULL);
571 struct find_occurrence_data
573 rtx find;
574 rtx *retval;
577 /* Callback for for_each_rtx, used in find_occurrence.
578 See if PX is the rtx we have to find. Return 1 to stop for_each_rtx
579 if successful, or 0 to continue traversing otherwise. */
581 static int
582 find_occurrence_callback (rtx *px, void *data)
584 struct find_occurrence_data *fod = (struct find_occurrence_data *) data;
585 rtx x = *px;
586 rtx find = fod->find;
588 if (x == find)
590 fod->retval = px;
591 return 1;
594 return 0;
597 /* Return a pointer to one of the occurrences of register FIND in *PX. */
599 static rtx *
600 find_occurrence (rtx *px, rtx find)
602 struct find_occurrence_data data;
604 gcc_assert (REG_P (find)
605 || (GET_CODE (find) == SUBREG
606 && REG_P (SUBREG_REG (find))));
608 data.find = find;
609 data.retval = NULL;
610 for_each_rtx (px, find_occurrence_callback, &data);
611 return data.retval;
615 /* Inside INSN, the expression rooted at *LOC has been changed, moving some
616 uses from ORIG_USES. Find those that are present, and create new items
617 in the data flow object of the pass. Mark any new uses as having the
618 given TYPE. */
619 static void
620 update_df (rtx insn, rtx *loc, struct df_ref *orig_uses, enum df_ref_type type,
621 int new_flags)
623 struct df_ref *use;
625 /* Add a use for the registers that were propagated. */
626 for (use = orig_uses; use; use = use->next_ref)
628 struct df_ref *orig_use = use, *new_use;
629 rtx *new_loc = find_occurrence (loc, DF_REF_REG (orig_use));
631 if (!new_loc)
632 continue;
634 /* Add a new insn use. Use the original type, because it says if the
635 use was within a MEM. */
636 new_use = df_ref_create (df, DF_REF_REG (orig_use), new_loc,
637 insn, BLOCK_FOR_INSN (insn),
638 type, DF_REF_FLAGS (orig_use) | new_flags);
640 /* Set up the use-def chain. */
641 df_chain_copy (df->problems_by_index[DF_CHAIN],
642 new_use, DF_REF_CHAIN (orig_use));
647 /* Try substituting NEW into LOC, which originated from forward propagation
648 of USE's value from DEF_INSN. SET_REG_EQUAL says whether we are
649 substituting the whole SET_SRC, so we can set a REG_EQUAL note if the
650 new insn is not recognized. Return whether the substitution was
651 performed. */
653 static bool
654 try_fwprop_subst (struct df_ref *use, rtx *loc, rtx new, rtx def_insn, bool set_reg_equal)
656 rtx insn = DF_REF_INSN (use);
657 enum df_ref_type type = DF_REF_TYPE (use);
658 int flags = DF_REF_FLAGS (use);
660 if (dump_file)
662 fprintf (dump_file, "\nIn insn %d, replacing\n ", INSN_UID (insn));
663 print_inline_rtx (dump_file, *loc, 2);
664 fprintf (dump_file, "\n with ");
665 print_inline_rtx (dump_file, new, 2);
666 fprintf (dump_file, "\n");
669 if (validate_change (insn, loc, new, false))
671 num_changes++;
672 if (dump_file)
673 fprintf (dump_file, "Changed insn %d\n", INSN_UID (insn));
675 /* Unlink the use that we changed. */
676 df_ref_remove (df, use);
677 if (!CONSTANT_P (new))
678 update_df (insn, loc, DF_INSN_USES (df, def_insn), type, flags);
680 return true;
682 else
684 if (dump_file)
685 fprintf (dump_file, "Changes to insn %d not recognized\n",
686 INSN_UID (insn));
688 /* Can also record a simplified value in a REG_EQUAL note, making a
689 new one if one does not already exist. */
690 if (set_reg_equal)
692 if (dump_file)
693 fprintf (dump_file, " Setting REG_EQUAL note\n");
695 set_unique_reg_note (insn, REG_EQUAL, copy_rtx (new));
697 /* ??? Is this still necessary if we add the note through
698 set_unique_reg_note? */
699 if (!CONSTANT_P (new))
700 update_df (insn, loc, DF_INSN_USES (df, def_insn),
701 type, DF_REF_IN_NOTE);
704 return false;
709 /* If USE is a paradoxical subreg, see if it can be replaced by a pseudo. */
711 static bool
712 forward_propagate_subreg (struct df_ref *use, rtx def_insn, rtx def_set)
714 rtx use_reg = DF_REF_REG (use);
715 rtx use_insn, src;
717 /* Only consider paradoxical subregs... */
718 enum machine_mode use_mode = GET_MODE (use_reg);
719 if (GET_CODE (use_reg) != SUBREG
720 || !REG_P (SET_DEST (def_set))
721 || GET_MODE_SIZE (use_mode)
722 <= GET_MODE_SIZE (GET_MODE (SUBREG_REG (use_reg))))
723 return false;
725 /* If this is a paradoxical SUBREG, we have no idea what value the
726 extra bits would have. However, if the operand is equivalent to
727 a SUBREG whose operand is the same as our mode, and all the modes
728 are within a word, we can just use the inner operand because
729 these SUBREGs just say how to treat the register. */
730 use_insn = DF_REF_INSN (use);
731 src = SET_SRC (def_set);
732 if (GET_CODE (src) == SUBREG
733 && REG_P (SUBREG_REG (src))
734 && GET_MODE (SUBREG_REG (src)) == use_mode
735 && subreg_lowpart_p (src)
736 && all_uses_available_at (def_insn, use_insn))
737 return try_fwprop_subst (use, DF_REF_LOC (use), SUBREG_REG (src),
738 def_insn, false);
739 else
740 return false;
743 /* Try to replace USE with SRC (defined in DEF_INSN) and simplify the
744 result. */
746 static bool
747 forward_propagate_and_simplify (struct df_ref *use, rtx def_insn, rtx def_set)
749 rtx use_insn = DF_REF_INSN (use);
750 rtx use_set = single_set (use_insn);
751 rtx src, reg, new, *loc;
752 bool set_reg_equal;
753 enum machine_mode mode;
755 if (!use_set)
756 return false;
758 /* Do not propagate into PC, CC0, etc. */
759 if (GET_MODE (SET_DEST (use_set)) == VOIDmode)
760 return false;
762 /* If def and use are subreg, check if they match. */
763 reg = DF_REF_REG (use);
764 if (GET_CODE (reg) == SUBREG
765 && GET_CODE (SET_DEST (def_set)) == SUBREG
766 && (SUBREG_BYTE (SET_DEST (def_set)) != SUBREG_BYTE (reg)
767 || GET_MODE (SET_DEST (def_set)) != GET_MODE (reg)))
768 return false;
770 /* Check if the def had a subreg, but the use has the whole reg. */
771 if (REG_P (reg) && GET_CODE (SET_DEST (def_set)) == SUBREG)
772 return false;
774 /* Check if the use has a subreg, but the def had the whole reg. Unlike the
775 previous case, the optimization is possible and often useful indeed. */
776 if (GET_CODE (reg) == SUBREG && REG_P (SET_DEST (def_set)))
777 reg = SUBREG_REG (reg);
779 /* Check if the substitution is valid (last, because it's the most
780 expensive check!). */
781 src = SET_SRC (def_set);
782 if (!CONSTANT_P (src) && !all_uses_available_at (def_insn, use_insn))
783 return false;
785 /* Check if the def is loading something from the constant pool; in this
786 case we would undo optimization such as compress_float_constant.
787 Still, we can set a REG_EQUAL note. */
788 if (MEM_P (src) && MEM_READONLY_P (src))
790 rtx x = avoid_constant_pool_reference (src);
791 if (x != src)
793 rtx note = find_reg_note (use_insn, REG_EQUAL, NULL_RTX);
794 rtx old = note ? XEXP (note, 0) : SET_SRC (use_set);
795 rtx new = simplify_replace_rtx (old, src, x);
796 if (old != new)
797 set_unique_reg_note (use_insn, REG_EQUAL, copy_rtx (new));
799 return false;
802 /* Else try simplifying. */
804 if (DF_REF_TYPE (use) == DF_REF_REG_MEM_STORE)
806 loc = &SET_DEST (use_set);
807 set_reg_equal = false;
809 else
811 rtx note = find_reg_note (use_insn, REG_EQUAL, NULL_RTX);
812 if (DF_REF_FLAGS (use) & DF_REF_IN_NOTE)
813 loc = &XEXP (note, 0);
814 else
815 loc = &SET_SRC (use_set);
817 /* Do not replace an existing REG_EQUAL note if the insn is not
818 recognized. Either we're already replacing in the note, or
819 we'll separately try plugging the definition in the note and
820 simplifying. */
821 set_reg_equal = (note == NULL_RTX);
824 if (GET_MODE (*loc) == VOIDmode)
825 mode = GET_MODE (SET_DEST (use_set));
826 else
827 mode = GET_MODE (*loc);
829 new = propagate_rtx (*loc, mode, reg, src);
831 if (!new)
832 return false;
834 return try_fwprop_subst (use, loc, new, def_insn, set_reg_equal);
838 /* Given a use USE of an insn, if it has a single reaching
839 definition, try to forward propagate it into that insn. */
841 static void
842 forward_propagate_into (struct df_ref *use)
844 struct df_link *defs;
845 struct df_ref *def;
846 rtx def_insn, def_set, use_insn;
847 rtx parent;
849 if (DF_REF_FLAGS (use) & DF_REF_READ_WRITE)
850 return;
851 if (DF_REF_FLAGS (use) & DF_REF_ARTIFICIAL)
852 return;
854 /* Only consider uses that have a single definition. */
855 defs = DF_REF_CHAIN (use);
856 if (!defs || defs->next)
857 return;
859 def = defs->ref;
860 if (DF_REF_FLAGS (def) & DF_REF_READ_WRITE)
861 return;
862 if (DF_REF_FLAGS (def) & DF_REF_ARTIFICIAL)
863 return;
865 /* Do not propagate loop invariant definitions inside the loop if
866 we are going to unroll. */
867 if (current_loops
868 && DF_REF_BB (def)->loop_father != DF_REF_BB (use)->loop_father)
869 return;
871 /* Check if the use is still present in the insn! */
872 use_insn = DF_REF_INSN (use);
873 if (DF_REF_FLAGS (use) & DF_REF_IN_NOTE)
874 parent = find_reg_note (use_insn, REG_EQUAL, NULL_RTX);
875 else
876 parent = PATTERN (use_insn);
878 if (!loc_mentioned_in_p (DF_REF_LOC (use), parent))
879 return;
881 def_insn = DF_REF_INSN (def);
882 def_set = single_set (def_insn);
883 if (!def_set)
884 return;
886 /* Only try one kind of propagation. If two are possible, we'll
887 do it on the following iterations. */
888 if (!forward_propagate_and_simplify (use, def_insn, def_set))
889 forward_propagate_subreg (use, def_insn, def_set);
893 static void
894 fwprop_init (void)
896 num_changes = 0;
897 calculate_dominance_info (CDI_DOMINATORS);
899 /* We do not always want to propagate into loops, so we have to find
900 loops and be careful about them. But we have to call flow_loops_find
901 before df_analyze, because flow_loops_find may introduce new jump
902 insns (sadly) if we are not working in cfglayout mode. */
903 if (flag_rerun_cse_after_loop && (flag_unroll_loops || flag_peel_loops))
904 loop_optimizer_init (0);
906 /* Now set up the dataflow problem (we only want use-def chains) and
907 put the dataflow solver to work. */
908 df = df_init (DF_HARD_REGS | DF_SUBREGS | DF_EQUIV_NOTES);
909 df_chain_add_problem (df, DF_UD_CHAIN);
910 df_analyze (df);
911 df_dump (df, dump_file);
914 static void
915 fwprop_done (void)
917 df_finish (df);
919 if (flag_rerun_cse_after_loop && (flag_unroll_loops || flag_peel_loops))
920 loop_optimizer_finalize ();
922 free_dominance_info (CDI_DOMINATORS);
923 cleanup_cfg (0);
924 delete_trivially_dead_insns (get_insns (), max_reg_num ());
926 if (dump_file)
927 fprintf (dump_file,
928 "\nNumber of successful forward propagations: %d\n\n",
929 num_changes);
934 /* Main entry point. */
936 static bool
937 gate_fwprop (void)
939 return optimize > 0 && flag_forward_propagate;
942 static unsigned int
943 fwprop (void)
945 unsigned i;
947 fwprop_init ();
949 /* Go through all the uses. update_df will create new ones at the
950 end, and we'll go through them as well.
952 Do not forward propagate addresses into loops until after unrolling.
953 CSE did so because it was able to fix its own mess, but we are not. */
955 df_reorganize_refs (&df->use_info);
956 for (i = 0; i < DF_USES_SIZE (df); i++)
958 struct df_ref *use = DF_USES_GET (df, i);
959 if (use)
960 if (!current_loops
961 || DF_REF_TYPE (use) == DF_REF_REG_USE
962 || DF_REF_BB (use)->loop_father == NULL)
963 forward_propagate_into (use);
966 fwprop_done ();
968 return 0;
971 struct tree_opt_pass pass_rtl_fwprop =
973 "fwprop1", /* name */
974 gate_fwprop, /* gate */
975 fwprop, /* execute */
976 NULL, /* sub */
977 NULL, /* next */
978 0, /* static_pass_number */
979 TV_FWPROP, /* tv_id */
980 0, /* properties_required */
981 0, /* properties_provided */
982 0, /* properties_destroyed */
983 0, /* todo_flags_start */
984 TODO_dump_func, /* todo_flags_finish */
985 0 /* letter */
988 static bool
989 gate_fwprop_addr (void)
991 return optimize > 0 && flag_forward_propagate && flag_rerun_cse_after_loop
992 && (flag_unroll_loops || flag_peel_loops);
995 static unsigned int
996 fwprop_addr (void)
998 unsigned i;
999 fwprop_init ();
1001 /* Go through all the uses. update_df will create new ones at the
1002 end, and we'll go through them as well. */
1003 df_reorganize_refs (&df->use_info);
1004 for (i = 0; i < DF_USES_SIZE (df); i++)
1006 struct df_ref *use = DF_USES_GET (df, i);
1007 if (use)
1008 if (DF_REF_TYPE (use) != DF_REF_REG_USE
1009 && DF_REF_BB (use)->loop_father != NULL)
1010 forward_propagate_into (use);
1013 fwprop_done ();
1015 return 0;
1018 struct tree_opt_pass pass_rtl_fwprop_addr =
1020 "fwprop2", /* name */
1021 gate_fwprop_addr, /* gate */
1022 fwprop_addr, /* execute */
1023 NULL, /* sub */
1024 NULL, /* next */
1025 0, /* static_pass_number */
1026 TV_FWPROP, /* tv_id */
1027 0, /* properties_required */
1028 0, /* properties_provided */
1029 0, /* properties_destroyed */
1030 0, /* todo_flags_start */
1031 TODO_dump_func, /* todo_flags_finish */
1032 0 /* letter */