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
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
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
24 #include "coretypes.h"
32 #include "insn-config.h"
36 #include "basic-block.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.
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 loops loops;
108 static struct df *df;
109 static int num_changes;
112 /* Do not try to replace constant addresses or addresses of local and
113 argument slots. These MEM expressions are made only once and inserted
114 in many instructions, as well as being used to control symbol table
115 output. It is not safe to clobber them.
117 There are some uncommon cases where the address is already in a register
118 for some reason, but we cannot take advantage of that because we have
119 no easy way to unshare the MEM. In addition, looking up all stack
120 addresses is costly. */
123 can_simplify_addr (rtx addr)
127 if (CONSTANT_ADDRESS_P (addr))
130 if (GET_CODE (addr) == PLUS)
131 reg = XEXP (addr, 0);
136 || (REGNO (reg) != FRAME_POINTER_REGNUM
137 && REGNO (reg) != HARD_FRAME_POINTER_REGNUM
138 && REGNO (reg) != ARG_POINTER_REGNUM));
141 /* Returns a canonical version of X for the address, from the point of view,
142 that all multiplications are represented as MULT instead of the multiply
143 by a power of 2 being represented as ASHIFT.
145 Every ASHIFT we find has been made by simplify_gen_binary and was not
146 there before, so it is not shared. So we can do this in place. */
149 canonicalize_address (rtx x)
152 switch (GET_CODE (x))
155 if (GET_CODE (XEXP (x, 1)) == CONST_INT
156 && INTVAL (XEXP (x, 1)) < GET_MODE_BITSIZE (GET_MODE (x))
157 && INTVAL (XEXP (x, 1)) >= 0)
159 HOST_WIDE_INT shift = INTVAL (XEXP (x, 1));
161 XEXP (x, 1) = gen_int_mode ((HOST_WIDE_INT) 1 << shift,
169 if (GET_CODE (XEXP (x, 0)) == PLUS
170 || GET_CODE (XEXP (x, 0)) == ASHIFT
171 || GET_CODE (XEXP (x, 0)) == CONST)
172 canonicalize_address (XEXP (x, 0));
186 /* OLD is a memory address. Return whether it is good to use NEW instead,
187 for a memory access in the given MODE. */
190 should_replace_address (rtx old, rtx new, enum machine_mode mode)
194 if (rtx_equal_p (old, new) || !memory_address_p (mode, new))
197 /* Copy propagation is always ok. */
198 if (REG_P (old) && REG_P (new))
201 /* Prefer the new address if it is less expensive. */
202 gain = address_cost (old, mode) - address_cost (new, mode);
204 /* If the addresses have equivalent cost, prefer the new address
205 if it has the highest `rtx_cost'. That has the potential of
206 eliminating the most insns without additional costs, and it
207 is the same that cse.c used to do. */
209 gain = rtx_cost (new, SET) - rtx_cost (old, SET);
214 /* Replace all occurrences of OLD in *PX with NEW and try to simplify the
215 resulting expression. Replace *PX with a new RTL expression if an
216 occurrence of OLD was found.
218 If CAN_APPEAR is true, we always return true; if it is false, we
219 can return false if, for at least one occurrence OLD, we failed to
220 collapse the result to a constant. For example, (mult:M (reg:M A)
221 (minus:M (reg:M B) (reg:M A))) may collapse to zero if replacing
222 (reg:M B) with (reg:M A).
224 CAN_APPEAR is disregarded inside MEMs: in that case, we always return
225 true if the simplification is a cheaper and valid memory address.
227 This is only a wrapper around simplify-rtx.c: do not add any pattern
228 matching code here. (The sole exception is the handling of LO_SUM, but
229 that is because there is no simplify_gen_* function for LO_SUM). */
232 propagate_rtx_1 (rtx *px, rtx old, rtx new, bool can_appear)
234 rtx x = *px, tem = NULL_RTX, op0, op1, op2;
235 enum rtx_code code = GET_CODE (x);
236 enum machine_mode mode = GET_MODE (x);
237 enum machine_mode op_mode;
238 bool valid_ops = true;
240 /* If X is OLD_RTX, return NEW_RTX. Otherwise, if this is an expression,
241 try to build a new expression from recursive substitution. */
249 switch (GET_RTX_CLASS (code))
253 op_mode = GET_MODE (op0);
254 valid_ops &= propagate_rtx_1 (&op0, old, new, can_appear);
255 if (op0 == XEXP (x, 0))
257 tem = simplify_gen_unary (code, mode, op0, op_mode);
264 valid_ops &= propagate_rtx_1 (&op0, old, new, can_appear);
265 valid_ops &= propagate_rtx_1 (&op1, old, new, can_appear);
266 if (op0 == XEXP (x, 0) && op1 == XEXP (x, 1))
268 tem = simplify_gen_binary (code, mode, op0, op1);
272 case RTX_COMM_COMPARE:
275 op_mode = GET_MODE (op0) != VOIDmode ? GET_MODE (op0) : GET_MODE (op1);
276 valid_ops &= propagate_rtx_1 (&op0, old, new, can_appear);
277 valid_ops &= propagate_rtx_1 (&op1, old, new, can_appear);
278 if (op0 == XEXP (x, 0) && op1 == XEXP (x, 1))
280 tem = simplify_gen_relational (code, mode, op_mode, op0, op1);
284 case RTX_BITFIELD_OPS:
288 op_mode = GET_MODE (op0);
289 valid_ops &= propagate_rtx_1 (&op0, old, new, can_appear);
290 valid_ops &= propagate_rtx_1 (&op1, old, new, can_appear);
291 valid_ops &= propagate_rtx_1 (&op2, old, new, can_appear);
292 if (op0 == XEXP (x, 0) && op1 == XEXP (x, 1) && op2 == XEXP (x, 2))
294 if (op_mode == VOIDmode)
295 op_mode = GET_MODE (op0);
296 tem = simplify_gen_ternary (code, mode, op_mode, op0, op1, op2);
300 /* The only case we try to handle is a SUBREG. */
304 valid_ops &= propagate_rtx_1 (&op0, old, new, can_appear);
305 if (op0 == XEXP (x, 0))
307 tem = simplify_gen_subreg (mode, op0, GET_MODE (SUBREG_REG (x)),
313 if (code == MEM && x != new)
318 /* There are some addresses that we cannot work on. */
319 if (!can_simplify_addr (op0))
322 op0 = new_op0 = targetm.delegitimize_address (op0);
323 valid_ops &= propagate_rtx_1 (&new_op0, old, new, true);
325 /* Dismiss transformation that we do not want to carry on. */
328 || GET_MODE (new_op0) != GET_MODE (op0))
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)))
338 tem = replace_equiv_address_nv (x, new_op0);
341 else if (code == LO_SUM)
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))
354 /* (lo_sum (high x) x) -> x */
355 if (GET_CODE (op0) == HIGH && rtx_equal_p (XEXP (op0, 0), op1))
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))
380 /* No change, no trouble. */
386 /* The replacement we made so far is valid, if all of the recursive
387 replacements were valid, or we could simplify everything to
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 expresion 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. */
401 propagate_rtx (rtx x, enum machine_mode mode, rtx old, rtx new)
406 if (REG_P (new) && REGNO (new) < FIRST_PSEUDO_REGISTER)
409 new = copy_rtx (new);
412 collapsed = propagate_rtx_1 (&tem, old, new, REG_P (new) || CONSTANT_P (new));
413 if (tem == x || !collapsed)
416 /* gen_lowpart_common will not be able to process VOIDmode entities other
418 if (GET_MODE (tem) == VOIDmode && GET_CODE (tem) != CONST_INT)
421 if (GET_MODE (tem) == VOIDmode)
422 tem = rtl_hooks.gen_lowpart_no_emit (mode, tem);
424 gcc_assert (GET_MODE (tem) == mode);
432 /* Return true if the register from reference REF is killed
433 between FROM to (but not including) TO. */
436 local_ref_killed_between_p (struct df_ref * ref, rtx from, rtx to)
441 for (insn = from; insn != to; insn = NEXT_INSN (insn))
446 def = DF_INSN_DEFS (df, insn);
449 if (DF_REF_REGNO (ref) == DF_REF_REGNO (def))
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. */
467 use_killed_between (struct df_ref *use, rtx def_insn, rtx target_insn)
469 basic_block def_bb, target_bb;
473 /* Check if the reg in USE has only one definition. We already
474 know that this definition reaches use, or we wouldn't be here. */
475 regno = DF_REF_REGNO (use);
476 def = DF_REG_DEF_GET (df, regno)->reg_chain;
477 if (def && (def->next_reg == NULL))
480 /* Check if we are in the same basic block. */
481 def_bb = BLOCK_FOR_INSN (def_insn);
482 target_bb = BLOCK_FOR_INSN (target_insn);
483 if (def_bb == target_bb)
485 /* In some obscure situations we can have a def reaching a use
486 that is _before_ the def. In other words the def does not
487 dominate the use even though the use and def are in the same
488 basic block. This can happen when a register may be used
489 uninitialized in a loop. In such cases, we must assume that
490 DEF is not available. */
491 if (DF_INSN_LUID (df, def_insn) >= DF_INSN_LUID (df, target_insn))
494 return local_ref_killed_between_p (use, def_insn, target_insn);
497 /* Finally, if DEF_BB is the sole predecessor of TARGET_BB. */
498 if (single_pred_p (target_bb)
499 && single_pred (target_bb) == def_bb)
503 /* See if USE is killed between DEF_INSN and the last insn in the
504 basic block containing DEF_INSN. */
505 x = df_bb_regno_last_def_find (df, def_bb, regno);
506 if (x && DF_INSN_LUID (df, x->insn) >= DF_INSN_LUID (df, def_insn))
509 /* See if USE is killed between TARGET_INSN and the first insn in the
510 basic block containing TARGET_INSN. */
511 x = df_bb_regno_first_def_find (df, target_bb, regno);
512 if (x && DF_INSN_LUID (df, x->insn) < DF_INSN_LUID (df, target_insn))
518 /* Otherwise assume the worst case. */
523 /* for_each_rtx traversal function that returns 1 if BODY points to
524 a non-constant mem. */
527 varying_mem_p (rtx *body, void *data ATTRIBUTE_UNUSED)
530 return MEM_P (x) && !MEM_READONLY_P (x);
533 /* Check if all uses in DEF_INSN can be used in TARGET_INSN. This
534 would require full computation of available expressions;
535 we check only restricted conditions, see use_killed_between. */
537 all_uses_available_at (rtx def_insn, rtx target_insn)
540 rtx def_set = single_set (def_insn);
542 gcc_assert (def_set);
544 /* If target_insn comes right after def_insn, which is very common
545 for addresses, we can use a quicker test. */
546 if (NEXT_INSN (def_insn) == target_insn
547 && REG_P (SET_DEST (def_set)))
549 rtx def_reg = SET_DEST (def_set);
551 /* If the insn uses the reg that it defines, the substitution is
553 for (use = DF_INSN_USES (df, def_insn); use; use = use->next_ref)
554 if (rtx_equal_p (use->reg, def_reg))
559 /* Look at all the uses of DEF_INSN, and see if they are not
560 killed between DEF_INSN and TARGET_INSN. */
561 for (use = DF_INSN_USES (df, def_insn); use; use = use->next_ref)
562 if (use_killed_between (use, def_insn, target_insn))
566 /* We don't do any analysis of memories or aliasing. Reject any
567 instruction that involves references to non-constant memory. */
568 return !for_each_rtx (&SET_SRC (def_set), varying_mem_p, NULL);
572 struct find_occurrence_data
578 /* Callback for for_each_rtx, used in find_occurrence.
579 See if PX is the rtx we have to find. Return 1 to stop for_each_rtx
580 if successful, or 0 to continue traversing otherwise. */
583 find_occurrence_callback (rtx *px, void *data)
585 struct find_occurrence_data *fod = (struct find_occurrence_data *) data;
587 rtx find = fod->find;
598 /* Return a pointer to one of the occurrences of register FIND in *PX. */
601 find_occurrence (rtx *px, rtx find)
603 struct find_occurrence_data data;
605 gcc_assert (REG_P (find)
606 || (GET_CODE (find) == SUBREG
607 && REG_P (SUBREG_REG (find))));
611 for_each_rtx (px, find_occurrence_callback, &data);
616 /* Inside INSN, the expression rooted at *LOC has been changed, moving some
617 uses from ORIG_USES. Find those that are present, and create new items
618 in the data flow object of the pass. Mark any new uses as having the
621 update_df (rtx insn, rtx *loc, struct df_ref *orig_uses, enum df_ref_type type,
626 /* Add a use for the registers that were propagated. */
627 for (use = orig_uses; use; use = use->next_ref)
629 struct df_ref *orig_use = use, *new_use;
630 rtx *new_loc = find_occurrence (loc, DF_REF_REG (orig_use));
635 /* Add a new insn use. Use the original type, because it says if the
636 use was within a MEM. */
637 new_use = df_ref_create (df, DF_REF_REG (orig_use), new_loc,
638 insn, BLOCK_FOR_INSN (insn),
639 type, DF_REF_FLAGS (orig_use) | new_flags);
641 /* Set up the use-def chain. */
642 df_chain_copy (df->problems_by_index[DF_CHAIN],
643 new_use, DF_REF_CHAIN (orig_use));
648 /* Try substituting NEW into LOC, which originated from forward propagation
649 of USE's value from DEF_INSN. SET_REG_EQUAL says whether we are
650 substituting the whole SET_SRC, so we can set a REG_EQUAL note if the
651 new insn is not recognized. Return whether the substitution was
655 try_fwprop_subst (struct df_ref *use, rtx *loc, rtx new, rtx def_insn, bool set_reg_equal)
657 rtx insn = DF_REF_INSN (use);
658 enum df_ref_type type = DF_REF_TYPE (use);
659 int flags = DF_REF_FLAGS (use);
663 fprintf (dump_file, "\nIn insn %d, replacing\n ", INSN_UID (insn));
664 print_inline_rtx (dump_file, *loc, 2);
665 fprintf (dump_file, "\n with ");
666 print_inline_rtx (dump_file, new, 2);
667 fprintf (dump_file, "\n");
670 if (validate_change (insn, loc, new, false))
674 fprintf (dump_file, "Changed insn %d\n", INSN_UID (insn));
676 /* Unlink the use that we changed. */
677 df_ref_remove (df, use);
678 if (!CONSTANT_P (new))
679 update_df (insn, loc, DF_INSN_USES (df, def_insn), type, flags);
686 fprintf (dump_file, "Changes to insn %d not recognized\n",
689 /* Can also record a simplified value in a REG_EQUAL note, making a
690 new one if one does not already exist. */
694 fprintf (dump_file, " Setting REG_EQUAL note\n");
696 REG_NOTES (insn) = gen_rtx_EXPR_LIST (REG_EQUAL, copy_rtx (new),
699 if (!CONSTANT_P (new))
700 update_df (insn, loc, DF_INSN_USES (df, def_insn),
701 type, DF_REF_IN_NOTE);
709 /* If USE is a paradoxical subreg, see if it can be replaced by a pseudo. */
712 forward_propagate_subreg (struct df_ref *use, rtx def_insn, rtx def_set)
714 rtx use_reg = DF_REF_REG (use);
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))))
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),
743 /* Try to replace USE with SRC (defined in DEF_INSN) and simplify the
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;
753 enum machine_mode mode;
758 /* Do not propagate into PC, CC0, etc. */
759 if (GET_MODE (SET_DEST (use_set)) == VOIDmode)
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)))
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)
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))
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);
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);
797 set_unique_reg_note (use_insn, REG_EQUAL, copy_rtx (new));
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;
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);
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
821 set_reg_equal = (note == NULL_RTX);
824 if (GET_MODE (*loc) == VOIDmode)
825 mode = GET_MODE (SET_DEST (use_set));
827 mode = GET_MODE (*loc);
829 new = propagate_rtx (*loc, mode, reg, src);
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. */
842 forward_propagate_into (struct df_ref *use)
844 struct df_link *defs;
846 rtx def_insn, def_set, use_insn;
849 if (DF_REF_FLAGS (use) & DF_REF_READ_WRITE)
852 /* Only consider uses that have a single definition. */
853 defs = DF_REF_CHAIN (use);
854 if (!defs || defs->next)
858 if (DF_REF_FLAGS (def) & DF_REF_READ_WRITE)
861 /* Do not propagate loop invariant definitions inside the loop if
862 we are going to unroll. */
864 && DF_REF_BB (def)->loop_father != DF_REF_BB (use)->loop_father)
867 /* Check if the use is still present in the insn! */
868 use_insn = DF_REF_INSN (use);
869 if (DF_REF_FLAGS (use) & DF_REF_IN_NOTE)
870 parent = find_reg_note (use_insn, REG_EQUAL, NULL_RTX);
872 parent = PATTERN (use_insn);
874 if (!loc_mentioned_in_p (DF_REF_LOC (use), parent))
877 def_insn = DF_REF_INSN (def);
878 def_set = single_set (def_insn);
882 /* Only try one kind of propagation. If two are possible, we'll
883 do it on the following iterations. */
884 if (!forward_propagate_and_simplify (use, def_insn, def_set))
885 forward_propagate_subreg (use, def_insn, def_set);
894 /* We do not always want to propagate into loops, so we have to find
895 loops and be careful about them. But we have to call flow_loops_find
896 before df_analyze, because flow_loops_find may introduce new jump
897 insns (sadly) if we are not working in cfglayout mode. */
898 if (flag_rerun_cse_after_loop && (flag_unroll_loops || flag_peel_loops))
900 calculate_dominance_info (CDI_DOMINATORS);
901 flow_loops_find (&loops);
904 /* Now set up the dataflow problem (we only want use-def chains) and
905 put the dataflow solver to work. */
906 df = df_init (DF_SUBREGS | DF_EQUIV_NOTES);
907 df_chain_add_problem (df, DF_UD_CHAIN);
909 df_dump (df, dump_file);
917 if (flag_rerun_cse_after_loop && (flag_unroll_loops || flag_peel_loops))
919 flow_loops_free (&loops);
920 free_dominance_info (CDI_DOMINATORS);
925 delete_trivially_dead_insns (get_insns (), max_reg_num ());
929 "\nNumber of successful forward propagations: %d\n\n",
935 /* Main entry point. */
940 return optimize > 0 && flag_forward_propagate;
950 /* Go through all the uses. update_df will create new ones at the
951 end, and we'll go through them as well.
953 Do not forward propagate addresses into loops until after unrolling.
954 CSE did so because it was able to fix its own mess, but we are not. */
956 df_reorganize_refs (&df->use_info);
957 for (i = 0; i < DF_USES_SIZE (df); i++)
959 struct df_ref *use = DF_USES_GET (df, i);
962 || DF_REF_TYPE (use) == DF_REF_REG_USE
963 || DF_REF_BB (use)->loop_father == NULL)
964 forward_propagate_into (use);
972 struct tree_opt_pass pass_rtl_fwprop =
974 "fwprop1", /* name */
975 gate_fwprop, /* gate */
976 fwprop, /* execute */
979 0, /* static_pass_number */
980 TV_FWPROP, /* tv_id */
981 0, /* properties_required */
982 0, /* properties_provided */
983 0, /* properties_destroyed */
984 0, /* todo_flags_start */
985 TODO_dump_func, /* todo_flags_finish */
990 gate_fwprop_addr (void)
992 return optimize > 0 && flag_forward_propagate && flag_rerun_cse_after_loop
993 && (flag_unroll_loops || flag_peel_loops);
1002 /* Go through all the uses. update_df will create new ones at the
1003 end, and we'll go through them as well. */
1004 df_reorganize_refs (&df->use_info);
1005 for (i = 0; i < DF_USES_SIZE (df); i++)
1007 struct df_ref *use = DF_USES_GET (df, i);
1009 if (DF_REF_TYPE (use) != DF_REF_REG_USE
1010 && DF_REF_BB (use)->loop_father != NULL)
1011 forward_propagate_into (use);
1019 struct tree_opt_pass pass_rtl_fwprop_addr =
1021 "fwprop2", /* name */
1022 gate_fwprop_addr, /* gate */
1023 fwprop_addr, /* execute */
1026 0, /* static_pass_number */
1027 TV_FWPROP, /* tv_id */
1028 0, /* properties_required */
1029 0, /* properties_provided */
1030 0, /* properties_destroyed */
1031 0, /* todo_flags_start */
1032 TODO_dump_func, /* todo_flags_finish */