1 /* RTL-level loop invariant motion.
2 Copyright (C) 2004, 2005, 2006, 2007, 2008 Free Software Foundation, Inc.
4 This file is part of GCC.
6 GCC is free software; you can redistribute it and/or modify it
7 under the terms of the GNU General Public License as published by the
8 Free Software Foundation; either version 3, or (at your option) any
11 GCC is distributed in the hope that it will be useful, but WITHOUT
12 ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
13 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
16 You should have received a copy of the GNU General Public License
17 along with GCC; see the file COPYING3. If not see
18 <http://www.gnu.org/licenses/>. */
20 /* This implements the loop invariant motion pass. It is very simple
21 (no calls, no loads/stores, etc.). This should be sufficient to cleanup
22 things like address arithmetics -- other more complicated invariants should
23 be eliminated on GIMPLE either in tree-ssa-loop-im.c or in tree-ssa-pre.c.
25 We proceed loop by loop -- it is simpler than trying to handle things
26 globally and should not lose much. First we inspect all sets inside loop
27 and create a dependency graph on insns (saying "to move this insn, you must
28 also move the following insns").
30 We then need to determine what to move. We estimate the number of registers
31 used and move as many invariants as possible while we still have enough free
32 registers. We prefer the expensive invariants.
34 Then we move the selected invariants out of the loop, creating a new
35 temporaries for them if necessary. */
39 #include "coretypes.h"
43 #include "hard-reg-set.h"
45 #include "basic-block.h"
56 /* The data stored for the loop. */
60 struct loop *outermost_exit; /* The outermost exit of the loop. */
61 bool has_call; /* True if the loop contains a call. */
64 #define LOOP_DATA(LOOP) ((struct loop_data *) (LOOP)->aux)
66 /* The description of an use. */
70 rtx *pos; /* Position of the use. */
71 rtx insn; /* The insn in that the use occurs. */
73 struct use *next; /* Next use in the list. */
76 /* The description of a def. */
80 struct use *uses; /* The list of uses that are uniquely reached
82 unsigned n_uses; /* Number of such uses. */
83 unsigned invno; /* The corresponding invariant. */
86 /* The data stored for each invariant. */
90 /* The number of the invariant. */
93 /* The number of the invariant with the same value. */
96 /* If we moved the invariant out of the loop, the register that contains its
100 /* The definition of the invariant. */
103 /* The insn in that it is defined. */
106 /* Whether it is always executed. */
107 bool always_executed;
109 /* Whether to move the invariant. */
112 /* Cost of the invariant. */
115 /* The invariants it depends on. */
118 /* Used for detecting already visited invariants during determining
119 costs of movements. */
123 /* Table of invariants indexed by the df_ref uid field. */
125 static unsigned int invariant_table_size = 0;
126 static struct invariant ** invariant_table;
128 /* Entry for hash table of invariant expressions. */
130 struct invariant_expr_entry
133 struct invariant *inv;
139 enum machine_mode mode;
145 /* The actual stamp for marking already visited invariants during determining
146 costs of movements. */
148 static unsigned actual_stamp;
150 typedef struct invariant *invariant_p;
152 DEF_VEC_P(invariant_p);
153 DEF_VEC_ALLOC_P(invariant_p, heap);
155 /* The invariants. */
157 static VEC(invariant_p,heap) *invariants;
159 /* Check the size of the invariant table and realloc if necessary. */
162 check_invariant_table_size (void)
164 if (invariant_table_size < DF_DEFS_TABLE_SIZE())
166 unsigned int new_size = DF_DEFS_TABLE_SIZE () + (DF_DEFS_TABLE_SIZE () / 4);
167 invariant_table = xrealloc (invariant_table,
168 sizeof (struct rtx_iv *) * new_size);
169 memset (&invariant_table[invariant_table_size], 0,
170 (new_size - invariant_table_size) * sizeof (struct rtx_iv *));
171 invariant_table_size = new_size;
175 /* Test for possibility of invariantness of X. */
178 check_maybe_invariant (rtx x)
180 enum rtx_code code = GET_CODE (x);
196 case UNSPEC_VOLATILE:
204 /* Load/store motion is done elsewhere. ??? Perhaps also add it here?
205 It should not be hard, and might be faster than "elsewhere". */
207 /* Just handle the most trivial case where we load from an unchanging
208 location (most importantly, pic tables). */
209 if (MEM_READONLY_P (x) && !MEM_VOLATILE_P (x))
215 /* Don't mess with insns declared volatile. */
216 if (MEM_VOLATILE_P (x))
224 fmt = GET_RTX_FORMAT (code);
225 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
229 if (!check_maybe_invariant (XEXP (x, i)))
232 else if (fmt[i] == 'E')
234 for (j = 0; j < XVECLEN (x, i); j++)
235 if (!check_maybe_invariant (XVECEXP (x, i, j)))
243 /* Returns the invariant definition for USE, or NULL if USE is not
246 static struct invariant *
247 invariant_for_use (struct df_ref *use)
249 struct df_link *defs;
251 basic_block bb = BLOCK_FOR_INSN (use->insn), def_bb;
253 if (use->flags & DF_REF_READ_WRITE)
256 defs = DF_REF_CHAIN (use);
257 if (!defs || defs->next)
260 check_invariant_table_size ();
261 if (!invariant_table[DF_REF_ID(def)])
264 def_bb = DF_REF_BB (def);
265 if (!dominated_by_p (CDI_DOMINATORS, bb, def_bb))
267 return invariant_table[DF_REF_ID(def)];
270 /* Computes hash value for invariant expression X in INSN. */
273 hash_invariant_expr_1 (rtx insn, rtx x)
275 enum rtx_code code = GET_CODE (x);
278 hashval_t val = code;
281 struct invariant *inv;
291 return hash_rtx (x, GET_MODE (x), &do_not_record_p, NULL, false);
294 use = df_find_use (insn, x);
296 return hash_rtx (x, GET_MODE (x), &do_not_record_p, NULL, false);
297 inv = invariant_for_use (use);
299 return hash_rtx (x, GET_MODE (x), &do_not_record_p, NULL, false);
301 gcc_assert (inv->eqto != ~0u);
308 fmt = GET_RTX_FORMAT (code);
309 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
312 val ^= hash_invariant_expr_1 (insn, XEXP (x, i));
313 else if (fmt[i] == 'E')
315 for (j = 0; j < XVECLEN (x, i); j++)
316 val ^= hash_invariant_expr_1 (insn, XVECEXP (x, i, j));
318 else if (fmt[i] == 'i' || fmt[i] == 'n')
325 /* Returns true if the invariant expressions E1 and E2 used in insns INSN1
326 and INSN2 have always the same value. */
329 invariant_expr_equal_p (rtx insn1, rtx e1, rtx insn2, rtx e2)
331 enum rtx_code code = GET_CODE (e1);
334 struct df_ref *use1, *use2;
335 struct invariant *inv1 = NULL, *inv2 = NULL;
338 /* If mode of only one of the operands is VOIDmode, it is not equivalent to
339 the other one. If both are VOIDmode, we rely on the caller of this
340 function to verify that their modes are the same. */
341 if (code != GET_CODE (e2) || GET_MODE (e1) != GET_MODE (e2))
352 return rtx_equal_p (e1, e2);
355 use1 = df_find_use (insn1, e1);
356 use2 = df_find_use (insn2, e2);
358 inv1 = invariant_for_use (use1);
360 inv2 = invariant_for_use (use2);
363 return rtx_equal_p (e1, e2);
368 gcc_assert (inv1->eqto != ~0u);
369 gcc_assert (inv2->eqto != ~0u);
370 return inv1->eqto == inv2->eqto;
376 fmt = GET_RTX_FORMAT (code);
377 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
384 if (!invariant_expr_equal_p (insn1, sub1, insn2, sub2))
388 else if (fmt[i] == 'E')
390 if (XVECLEN (e1, i) != XVECLEN (e2, i))
393 for (j = 0; j < XVECLEN (e1, i); j++)
395 sub1 = XVECEXP (e1, i, j);
396 sub2 = XVECEXP (e2, i, j);
398 if (!invariant_expr_equal_p (insn1, sub1, insn2, sub2))
402 else if (fmt[i] == 'i' || fmt[i] == 'n')
404 if (XINT (e1, i) != XINT (e2, i))
407 /* Unhandled type of subexpression, we fail conservatively. */
415 /* Returns hash value for invariant expression entry E. */
418 hash_invariant_expr (const void *e)
420 const struct invariant_expr_entry *entry = e;
425 /* Compares invariant expression entries E1 and E2. */
428 eq_invariant_expr (const void *e1, const void *e2)
430 const struct invariant_expr_entry *entry1 = e1;
431 const struct invariant_expr_entry *entry2 = e2;
433 if (entry1->mode != entry2->mode)
436 return invariant_expr_equal_p (entry1->inv->insn, entry1->expr,
437 entry2->inv->insn, entry2->expr);
440 /* Checks whether invariant with value EXPR in machine mode MODE is
441 recorded in EQ. If this is the case, return the invariant. Otherwise
442 insert INV to the table for this expression and return INV. */
444 static struct invariant *
445 find_or_insert_inv (htab_t eq, rtx expr, enum machine_mode mode,
446 struct invariant *inv)
448 hashval_t hash = hash_invariant_expr_1 (inv->insn, expr);
449 struct invariant_expr_entry *entry;
450 struct invariant_expr_entry pentry;
456 slot = htab_find_slot_with_hash (eq, &pentry, hash, INSERT);
462 entry = XNEW (struct invariant_expr_entry);
472 /* Finds invariants identical to INV and records the equivalence. EQ is the
473 hash table of the invariants. */
476 find_identical_invariants (htab_t eq, struct invariant *inv)
480 struct invariant *dep;
482 enum machine_mode mode;
484 if (inv->eqto != ~0u)
487 EXECUTE_IF_SET_IN_BITMAP (inv->depends_on, 0, depno, bi)
489 dep = VEC_index (invariant_p, invariants, depno);
490 find_identical_invariants (eq, dep);
493 set = single_set (inv->insn);
494 expr = SET_SRC (set);
495 mode = GET_MODE (expr);
496 if (mode == VOIDmode)
497 mode = GET_MODE (SET_DEST (set));
498 inv->eqto = find_or_insert_inv (eq, expr, mode, inv)->invno;
500 if (dump_file && inv->eqto != inv->invno)
502 "Invariant %d is equivalent to invariant %d.\n",
503 inv->invno, inv->eqto);
506 /* Find invariants with the same value and record the equivalences. */
509 merge_identical_invariants (void)
512 struct invariant *inv;
513 htab_t eq = htab_create (VEC_length (invariant_p, invariants),
514 hash_invariant_expr, eq_invariant_expr, free);
516 for (i = 0; VEC_iterate (invariant_p, invariants, i, inv); i++)
517 find_identical_invariants (eq, inv);
522 /* Determines the basic blocks inside LOOP that are always executed and
523 stores their bitmap to ALWAYS_REACHED. MAY_EXIT is a bitmap of
524 basic blocks that may either exit the loop, or contain the call that
525 does not have to return. BODY is body of the loop obtained by
526 get_loop_body_in_dom_order. */
529 compute_always_reached (struct loop *loop, basic_block *body,
530 bitmap may_exit, bitmap always_reached)
534 for (i = 0; i < loop->num_nodes; i++)
536 if (dominated_by_p (CDI_DOMINATORS, loop->latch, body[i]))
537 bitmap_set_bit (always_reached, i);
539 if (bitmap_bit_p (may_exit, i))
544 /* Finds exits out of the LOOP with body BODY. Marks blocks in that we may
545 exit the loop by cfg edge to HAS_EXIT and MAY_EXIT. In MAY_EXIT
546 additionally mark blocks that may exit due to a call. */
549 find_exits (struct loop *loop, basic_block *body,
550 bitmap may_exit, bitmap has_exit)
555 struct loop *outermost_exit = loop, *aexit;
556 bool has_call = false;
559 for (i = 0; i < loop->num_nodes; i++)
561 if (body[i]->loop_father == loop)
563 FOR_BB_INSNS (body[i], insn)
566 && (RTL_LOOPING_CONST_OR_PURE_CALL_P (insn)
567 || !RTL_CONST_OR_PURE_CALL_P (insn)))
570 bitmap_set_bit (may_exit, i);
575 FOR_EACH_EDGE (e, ei, body[i]->succs)
577 if (flow_bb_inside_loop_p (loop, e->dest))
580 bitmap_set_bit (may_exit, i);
581 bitmap_set_bit (has_exit, i);
582 outermost_exit = find_common_loop (outermost_exit,
583 e->dest->loop_father);
588 /* Use the data stored for the subloop to decide whether we may exit
589 through it. It is sufficient to do this for header of the loop,
590 as other basic blocks inside it must be dominated by it. */
591 if (body[i]->loop_father->header != body[i])
594 if (LOOP_DATA (body[i]->loop_father)->has_call)
597 bitmap_set_bit (may_exit, i);
599 aexit = LOOP_DATA (body[i]->loop_father)->outermost_exit;
602 bitmap_set_bit (may_exit, i);
603 bitmap_set_bit (has_exit, i);
605 if (flow_loop_nested_p (aexit, outermost_exit))
606 outermost_exit = aexit;
610 loop->aux = xcalloc (1, sizeof (struct loop_data));
611 LOOP_DATA (loop)->outermost_exit = outermost_exit;
612 LOOP_DATA (loop)->has_call = has_call;
615 /* Check whether we may assign a value to X from a register. */
618 may_assign_reg_p (rtx x)
620 return (GET_MODE (x) != VOIDmode
621 && GET_MODE (x) != BLKmode
622 && can_copy_p (GET_MODE (x))
624 || !HARD_REGISTER_P (x)
625 || REGNO_REG_CLASS (REGNO (x)) != NO_REGS));
628 /* Finds definitions that may correspond to invariants in LOOP with body
632 find_defs (struct loop *loop, basic_block *body)
635 bitmap blocks = BITMAP_ALLOC (NULL);
637 for (i = 0; i < loop->num_nodes; i++)
638 bitmap_set_bit (blocks, body[i]->index);
640 df_remove_problem (df_chain);
641 df_process_deferred_rescans ();
642 df_chain_add_problem (DF_UD_CHAIN);
643 df_set_blocks (blocks);
648 df_dump_region (dump_file);
649 fprintf (dump_file, "*****starting processing of loop ******\n");
650 print_rtl_with_bb (dump_file, get_insns ());
651 fprintf (dump_file, "*****ending processing of loop ******\n");
653 check_invariant_table_size ();
655 BITMAP_FREE (blocks);
658 /* Creates a new invariant for definition DEF in INSN, depending on invariants
659 in DEPENDS_ON. ALWAYS_EXECUTED is true if the insn is always executed,
660 unless the program ends due to a function call. The newly created invariant
663 static struct invariant *
664 create_new_invariant (struct def *def, rtx insn, bitmap depends_on,
665 bool always_executed)
667 struct invariant *inv = XNEW (struct invariant);
668 rtx set = single_set (insn);
671 inv->always_executed = always_executed;
672 inv->depends_on = depends_on;
674 /* If the set is simple, usually by moving it we move the whole store out of
675 the loop. Otherwise we save only cost of the computation. */
677 inv->cost = rtx_cost (set, SET);
679 inv->cost = rtx_cost (SET_SRC (set), SET);
686 inv->invno = VEC_length (invariant_p, invariants);
689 def->invno = inv->invno;
690 VEC_safe_push (invariant_p, heap, invariants, inv);
695 "Set in insn %d is invariant (%d), cost %d, depends on ",
696 INSN_UID (insn), inv->invno, inv->cost);
697 dump_bitmap (dump_file, inv->depends_on);
703 /* Record USE at DEF. */
706 record_use (struct def *def, rtx *use, rtx insn)
708 struct use *u = XNEW (struct use);
710 gcc_assert (REG_P (*use));
719 /* Finds the invariants USE depends on and store them to the DEPENDS_ON
720 bitmap. Returns true if all dependencies of USE are known to be
721 loop invariants, false otherwise. */
724 check_dependency (basic_block bb, struct df_ref *use, bitmap depends_on)
728 struct df_link *defs;
729 struct def *def_data;
730 struct invariant *inv;
732 if (use->flags & DF_REF_READ_WRITE)
735 defs = DF_REF_CHAIN (use);
743 check_invariant_table_size ();
744 inv = invariant_table[DF_REF_ID(def)];
749 gcc_assert (def_data != NULL);
751 def_bb = DF_REF_BB (def);
752 /* Note that in case bb == def_bb, we know that the definition
753 dominates insn, because def has invariant_table[DF_REF_ID(def)]
754 defined and we process the insns in the basic block bb
756 if (!dominated_by_p (CDI_DOMINATORS, bb, def_bb))
759 bitmap_set_bit (depends_on, def_data->invno);
764 /* Finds the invariants INSN depends on and store them to the DEPENDS_ON
765 bitmap. Returns true if all dependencies of INSN are known to be
766 loop invariants, false otherwise. */
769 check_dependencies (rtx insn, bitmap depends_on)
771 struct df_ref **use_rec;
772 basic_block bb = BLOCK_FOR_INSN (insn);
774 for (use_rec = DF_INSN_USES (insn); *use_rec; use_rec++)
775 if (!check_dependency (bb, *use_rec, depends_on))
777 for (use_rec = DF_INSN_EQ_USES (insn); *use_rec; use_rec++)
778 if (!check_dependency (bb, *use_rec, depends_on))
784 /* Finds invariant in INSN. ALWAYS_REACHED is true if the insn is always
785 executed. ALWAYS_EXECUTED is true if the insn is always executed,
786 unless the program ends due to a function call. */
789 find_invariant_insn (rtx insn, bool always_reached, bool always_executed)
796 struct invariant *inv;
799 /* We can't move a CC0 setter without the user. */
800 if (sets_cc0_p (insn))
804 set = single_set (insn);
807 dest = SET_DEST (set);
810 || HARD_REGISTER_P (dest))
813 if (!may_assign_reg_p (SET_DEST (set))
814 || !check_maybe_invariant (SET_SRC (set)))
817 /* If the insn can throw exception, we cannot move it at all without changing
819 if (can_throw_internal (insn))
822 /* We cannot make trapping insn executed, unless it was executed before. */
823 if (may_trap_after_code_motion_p (PATTERN (insn)) && !always_reached)
826 depends_on = BITMAP_ALLOC (NULL);
827 if (!check_dependencies (insn, depends_on))
829 BITMAP_FREE (depends_on);
834 def = XCNEW (struct def);
838 inv = create_new_invariant (def, insn, depends_on, always_executed);
842 ref = df_find_def (insn, dest);
843 check_invariant_table_size ();
844 invariant_table[DF_REF_ID(ref)] = inv;
848 /* Record registers used in INSN that have a unique invariant definition. */
851 record_uses (rtx insn)
853 struct df_ref **use_rec;
854 struct invariant *inv;
856 for (use_rec = DF_INSN_USES (insn); *use_rec; use_rec++)
858 struct df_ref *use = *use_rec;
859 inv = invariant_for_use (use);
861 record_use (inv->def, DF_REF_REAL_LOC (use), DF_REF_INSN (use));
863 for (use_rec = DF_INSN_EQ_USES (insn); *use_rec; use_rec++)
865 struct df_ref *use = *use_rec;
866 inv = invariant_for_use (use);
868 record_use (inv->def, DF_REF_REAL_LOC (use), DF_REF_INSN (use));
872 /* Finds invariants in INSN. ALWAYS_REACHED is true if the insn is always
873 executed. ALWAYS_EXECUTED is true if the insn is always executed,
874 unless the program ends due to a function call. */
877 find_invariants_insn (rtx insn, bool always_reached, bool always_executed)
879 find_invariant_insn (insn, always_reached, always_executed);
883 /* Finds invariants in basic block BB. ALWAYS_REACHED is true if the
884 basic block is always executed. ALWAYS_EXECUTED is true if the basic
885 block is always executed, unless the program ends due to a function
889 find_invariants_bb (basic_block bb, bool always_reached, bool always_executed)
893 FOR_BB_INSNS (bb, insn)
898 find_invariants_insn (insn, always_reached, always_executed);
902 && (RTL_LOOPING_CONST_OR_PURE_CALL_P (insn)
903 || ! RTL_CONST_OR_PURE_CALL_P (insn)))
904 always_reached = false;
908 /* Finds invariants in LOOP with body BODY. ALWAYS_REACHED is the bitmap of
909 basic blocks in BODY that are always executed. ALWAYS_EXECUTED is the
910 bitmap of basic blocks in BODY that are always executed unless the program
911 ends due to a function call. */
914 find_invariants_body (struct loop *loop, basic_block *body,
915 bitmap always_reached, bitmap always_executed)
919 for (i = 0; i < loop->num_nodes; i++)
920 find_invariants_bb (body[i],
921 bitmap_bit_p (always_reached, i),
922 bitmap_bit_p (always_executed, i));
925 /* Finds invariants in LOOP. */
928 find_invariants (struct loop *loop)
930 bitmap may_exit = BITMAP_ALLOC (NULL);
931 bitmap always_reached = BITMAP_ALLOC (NULL);
932 bitmap has_exit = BITMAP_ALLOC (NULL);
933 bitmap always_executed = BITMAP_ALLOC (NULL);
934 basic_block *body = get_loop_body_in_dom_order (loop);
936 find_exits (loop, body, may_exit, has_exit);
937 compute_always_reached (loop, body, may_exit, always_reached);
938 compute_always_reached (loop, body, has_exit, always_executed);
940 find_defs (loop, body);
941 find_invariants_body (loop, body, always_reached, always_executed);
942 merge_identical_invariants ();
944 BITMAP_FREE (always_reached);
945 BITMAP_FREE (always_executed);
946 BITMAP_FREE (may_exit);
947 BITMAP_FREE (has_exit);
951 /* Frees a list of uses USE. */
954 free_use_list (struct use *use)
958 for (; use; use = next)
965 /* Calculates cost and number of registers needed for moving invariant INV
966 out of the loop and stores them to *COST and *REGS_NEEDED. */
969 get_inv_cost (struct invariant *inv, int *comp_cost, unsigned *regs_needed)
972 unsigned aregs_needed;
974 struct invariant *dep;
977 /* Find the representative of the class of the equivalent invariants. */
978 inv = VEC_index (invariant_p, invariants, inv->eqto);
983 || inv->stamp == actual_stamp)
985 inv->stamp = actual_stamp;
988 (*comp_cost) += inv->cost;
992 /* Hoisting constant pool constants into stack regs may cost more than
993 just single register. On x87, the balance is affected both by the
994 small number of FP registers, and by its register stack organization,
995 that forces us to add compensation code in and around the loop to
996 shuffle the operands to the top of stack before use, and pop them
997 from the stack after the loop finishes.
999 To model this effect, we increase the number of registers needed for
1000 stack registers by two: one register push, and one register pop.
1001 This usually has the effect that FP constant loads from the constant
1002 pool are not moved out of the loop.
1004 Note that this also means that dependent invariants can not be moved.
1005 However, the primary purpose of this pass is to move loop invariant
1006 address arithmetic out of loops, and address arithmetic that depends
1007 on floating point constants is unlikely to ever occur. */
1008 rtx set = single_set (inv->insn);
1010 && IS_STACK_MODE (GET_MODE (SET_SRC (set)))
1011 && constant_pool_constant_p (SET_SRC (set)))
1012 (*regs_needed) += 2;
1016 EXECUTE_IF_SET_IN_BITMAP (inv->depends_on, 0, depno, bi)
1018 dep = VEC_index (invariant_p, invariants, depno);
1020 get_inv_cost (dep, &acomp_cost, &aregs_needed);
1023 /* We need to check always_executed, since if the original value of
1024 the invariant may be preserved, we may need to keep it in a
1025 separate register. TODO check whether the register has an
1026 use outside of the loop. */
1027 && dep->always_executed
1028 && !dep->def->uses->next)
1030 /* If this is a single use, after moving the dependency we will not
1031 need a new register. */
1035 (*regs_needed) += aregs_needed;
1036 (*comp_cost) += acomp_cost;
1040 /* Calculates gain for eliminating invariant INV. REGS_USED is the number
1041 of registers used in the loop, NEW_REGS is the number of new variables
1042 already added due to the invariant motion. The number of registers needed
1043 for it is stored in *REGS_NEEDED. */
1046 gain_for_invariant (struct invariant *inv, unsigned *regs_needed,
1047 unsigned new_regs, unsigned regs_used)
1049 int comp_cost, size_cost;
1051 get_inv_cost (inv, &comp_cost, regs_needed);
1054 size_cost = (estimate_reg_pressure_cost (new_regs + *regs_needed, regs_used)
1055 - estimate_reg_pressure_cost (new_regs, regs_used));
1057 return comp_cost - size_cost;
1060 /* Finds invariant with best gain for moving. Returns the gain, stores
1061 the invariant in *BEST and number of registers needed for it to
1062 *REGS_NEEDED. REGS_USED is the number of registers used in the loop.
1063 NEW_REGS is the number of new variables already added due to invariant
1067 best_gain_for_invariant (struct invariant **best, unsigned *regs_needed,
1068 unsigned new_regs, unsigned regs_used)
1070 struct invariant *inv;
1071 int gain = 0, again;
1072 unsigned aregs_needed, invno;
1074 for (invno = 0; VEC_iterate (invariant_p, invariants, invno, inv); invno++)
1079 /* Only consider the "representatives" of equivalent invariants. */
1080 if (inv->eqto != inv->invno)
1083 again = gain_for_invariant (inv, &aregs_needed, new_regs, regs_used);
1088 *regs_needed = aregs_needed;
1095 /* Marks invariant INVNO and all its dependencies for moving. */
1098 set_move_mark (unsigned invno)
1100 struct invariant *inv = VEC_index (invariant_p, invariants, invno);
1103 /* Find the representative of the class of the equivalent invariants. */
1104 inv = VEC_index (invariant_p, invariants, inv->eqto);
1111 fprintf (dump_file, "Decided to move invariant %d\n", invno);
1113 EXECUTE_IF_SET_IN_BITMAP (inv->depends_on, 0, invno, bi)
1115 set_move_mark (invno);
1119 /* Determines which invariants to move. */
1122 find_invariants_to_move (void)
1124 unsigned i, regs_used, regs_needed = 0, new_regs;
1125 struct invariant *inv = NULL;
1126 unsigned int n_regs = DF_REG_SIZE (df);
1128 if (!VEC_length (invariant_p, invariants))
1131 /* We do not really do a good job in estimating number of registers used;
1132 we put some initial bound here to stand for induction variables etc.
1133 that we do not detect. */
1136 for (i = 0; i < n_regs; i++)
1138 if (!DF_REGNO_FIRST_DEF (i) && DF_REGNO_LAST_USE (i))
1140 /* This is a value that is used but not changed inside loop. */
1146 while (best_gain_for_invariant (&inv, ®s_needed, new_regs, regs_used) > 0)
1148 set_move_mark (inv->invno);
1149 new_regs += regs_needed;
1153 /* Move invariant INVNO out of the LOOP. Returns true if this succeeds, false
1157 move_invariant_reg (struct loop *loop, unsigned invno)
1159 struct invariant *inv = VEC_index (invariant_p, invariants, invno);
1160 struct invariant *repr = VEC_index (invariant_p, invariants, inv->eqto);
1162 basic_block preheader = loop_preheader_edge (loop)->src;
1163 rtx reg, set, dest, note;
1171 /* If this is a representative of the class of equivalent invariants,
1172 really move the invariant. Otherwise just replace its use with
1173 the register used for the representative. */
1176 if (inv->depends_on)
1178 EXECUTE_IF_SET_IN_BITMAP (inv->depends_on, 0, i, bi)
1180 if (!move_invariant_reg (loop, i))
1185 /* Move the set out of the loop. If the set is always executed (we could
1186 omit this condition if we know that the register is unused outside of the
1187 loop, but it does not seem worth finding out) and it has no uses that
1188 would not be dominated by it, we may just move it (TODO). Otherwise we
1189 need to create a temporary register. */
1190 set = single_set (inv->insn);
1191 dest = SET_DEST (set);
1192 reg = gen_reg_rtx_and_attrs (dest);
1194 /* Try replacing the destination by a new pseudoregister. */
1195 if (!validate_change (inv->insn, &SET_DEST (set), reg, false))
1197 df_insn_rescan (inv->insn);
1199 emit_insn_after (gen_move_insn (dest, reg), inv->insn);
1200 reorder_insns (inv->insn, inv->insn, BB_END (preheader));
1202 /* If there is a REG_EQUAL note on the insn we just moved, and
1203 insn is in a basic block that is not always executed, the note
1204 may no longer be valid after we move the insn.
1205 Note that uses in REG_EQUAL notes are taken into account in
1206 the computation of invariants. Hence it is safe to retain the
1207 note even if the note contains register references. */
1208 if (! inv->always_executed
1209 && (note = find_reg_note (inv->insn, REG_EQUAL, NULL_RTX)))
1210 remove_note (inv->insn, note);
1214 if (!move_invariant_reg (loop, repr->invno))
1217 set = single_set (inv->insn);
1218 emit_insn_after (gen_move_insn (SET_DEST (set), reg), inv->insn);
1219 delete_insn (inv->insn);
1225 /* Replace the uses we know to be dominated. It saves work for copy
1226 propagation, and also it is necessary so that dependent invariants
1227 are computed right. */
1230 for (use = inv->def->uses; use; use = use->next)
1233 df_insn_rescan (use->insn);
1240 /* If we failed, clear move flag, so that we do not try to move inv
1243 fprintf (dump_file, "Failed to move invariant %d\n", invno);
1245 inv->reg = NULL_RTX;
1250 /* Move selected invariant out of the LOOP. Newly created regs are marked
1251 in TEMPORARY_REGS. */
1254 move_invariants (struct loop *loop)
1256 struct invariant *inv;
1259 for (i = 0; VEC_iterate (invariant_p, invariants, i, inv); i++)
1260 move_invariant_reg (loop, i);
1263 /* Initializes invariant motion data. */
1266 init_inv_motion_data (void)
1270 invariants = VEC_alloc (invariant_p, heap, 100);
1273 /* Frees the data allocated by invariant motion. */
1276 free_inv_motion_data (void)
1280 struct invariant *inv;
1282 check_invariant_table_size ();
1283 for (i = 0; i < DF_DEFS_TABLE_SIZE (); i++)
1285 inv = invariant_table[i];
1289 gcc_assert (def != NULL);
1291 free_use_list (def->uses);
1293 invariant_table[i] = NULL;
1297 for (i = 0; VEC_iterate (invariant_p, invariants, i, inv); i++)
1299 BITMAP_FREE (inv->depends_on);
1302 VEC_free (invariant_p, heap, invariants);
1305 /* Move the invariants out of the LOOP. */
1308 move_single_loop_invariants (struct loop *loop)
1310 init_inv_motion_data ();
1312 find_invariants (loop);
1313 find_invariants_to_move ();
1314 move_invariants (loop);
1316 free_inv_motion_data ();
1319 /* Releases the auxiliary data for LOOP. */
1322 free_loop_data (struct loop *loop)
1324 struct loop_data *data = LOOP_DATA (loop);
1330 /* Move the invariants out of the loops. */
1333 move_loop_invariants (void)
1338 df_set_flags (DF_EQ_NOTES + DF_DEFER_INSN_RESCAN);
1339 /* Process the loops, innermost first. */
1340 FOR_EACH_LOOP (li, loop, LI_FROM_INNERMOST)
1342 move_single_loop_invariants (loop);
1345 FOR_EACH_LOOP (li, loop, 0)
1347 free_loop_data (loop);
1350 free (invariant_table);
1351 invariant_table = NULL;
1352 invariant_table_size = 0;
1354 #ifdef ENABLE_CHECKING
1355 verify_flow_info ();