1 /* RTL-level loop invariant motion.
2 Copyright (C) 2004, 2005, 2006, 2007 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, libcalls, etc.). This should be sufficient to cleanup things
22 like address arithmetics -- other more complicated invariants should be
23 eliminated on tree level 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))
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 && !CONST_OR_PURE_CALL_P (insn))
569 bitmap_set_bit (may_exit, i);
574 FOR_EACH_EDGE (e, ei, body[i]->succs)
576 if (flow_bb_inside_loop_p (loop, e->dest))
579 bitmap_set_bit (may_exit, i);
580 bitmap_set_bit (has_exit, i);
581 outermost_exit = find_common_loop (outermost_exit,
582 e->dest->loop_father);
587 /* Use the data stored for the subloop to decide whether we may exit
588 through it. It is sufficient to do this for header of the loop,
589 as other basic blocks inside it must be dominated by it. */
590 if (body[i]->loop_father->header != body[i])
593 if (LOOP_DATA (body[i]->loop_father)->has_call)
596 bitmap_set_bit (may_exit, i);
598 aexit = LOOP_DATA (body[i]->loop_father)->outermost_exit;
601 bitmap_set_bit (may_exit, i);
602 bitmap_set_bit (has_exit, i);
604 if (flow_loop_nested_p (aexit, outermost_exit))
605 outermost_exit = aexit;
609 loop->aux = xcalloc (1, sizeof (struct loop_data));
610 LOOP_DATA (loop)->outermost_exit = outermost_exit;
611 LOOP_DATA (loop)->has_call = has_call;
614 /* Check whether we may assign a value to X from a register. */
617 may_assign_reg_p (rtx x)
619 return (GET_MODE (x) != VOIDmode
620 && GET_MODE (x) != BLKmode
621 && can_copy_p (GET_MODE (x))
623 || !HARD_REGISTER_P (x)
624 || REGNO_REG_CLASS (REGNO (x)) != NO_REGS));
627 /* Finds definitions that may correspond to invariants in LOOP with body
631 find_defs (struct loop *loop, basic_block *body)
634 bitmap blocks = BITMAP_ALLOC (NULL);
636 for (i = 0; i < loop->num_nodes; i++)
637 bitmap_set_bit (blocks, body[i]->index);
639 df_remove_problem (df_chain);
640 df_process_deferred_rescans ();
641 df_chain_add_problem (DF_UD_CHAIN);
642 df_set_blocks (blocks);
647 fprintf (dump_file, "*****starting processing of loop ******\n");
648 print_rtl_with_bb (dump_file, get_insns ());
649 fprintf (dump_file, "*****ending processing of loop ******\n");
651 check_invariant_table_size ();
653 BITMAP_FREE (blocks);
656 /* Creates a new invariant for definition DEF in INSN, depending on invariants
657 in DEPENDS_ON. ALWAYS_EXECUTED is true if the insn is always executed,
658 unless the program ends due to a function call. The newly created invariant
661 static struct invariant *
662 create_new_invariant (struct def *def, rtx insn, bitmap depends_on,
663 bool always_executed)
665 struct invariant *inv = XNEW (struct invariant);
666 rtx set = single_set (insn);
669 inv->always_executed = always_executed;
670 inv->depends_on = depends_on;
672 /* If the set is simple, usually by moving it we move the whole store out of
673 the loop. Otherwise we save only cost of the computation. */
675 inv->cost = rtx_cost (set, SET);
677 inv->cost = rtx_cost (SET_SRC (set), SET);
684 inv->invno = VEC_length (invariant_p, invariants);
687 def->invno = inv->invno;
688 VEC_safe_push (invariant_p, heap, invariants, inv);
693 "Set in insn %d is invariant (%d), cost %d, depends on ",
694 INSN_UID (insn), inv->invno, inv->cost);
695 dump_bitmap (dump_file, inv->depends_on);
701 /* Record USE at DEF. */
704 record_use (struct def *def, rtx *use, rtx insn)
706 struct use *u = XNEW (struct use);
708 gcc_assert (REG_P (*use));
717 /* Finds the invariants USE depends on and store them to the DEPENDS_ON
718 bitmap. Returns true if all dependencies of USE are known to be
719 loop invariants, false otherwise. */
722 check_dependency (basic_block bb, struct df_ref *use, bitmap depends_on)
726 struct df_link *defs;
727 struct def *def_data;
728 struct invariant *inv;
730 if (use->flags & DF_REF_READ_WRITE)
733 defs = DF_REF_CHAIN (use);
741 check_invariant_table_size ();
742 inv = invariant_table[DF_REF_ID(def)];
747 gcc_assert (def_data != NULL);
749 def_bb = DF_REF_BB (def);
750 /* Note that in case bb == def_bb, we know that the definition
751 dominates insn, because def has invariant_table[DF_REF_ID(def)]
752 defined and we process the insns in the basic block bb
754 if (!dominated_by_p (CDI_DOMINATORS, bb, def_bb))
757 bitmap_set_bit (depends_on, def_data->invno);
762 /* Finds the invariants INSN depends on and store them to the DEPENDS_ON
763 bitmap. Returns true if all dependencies of INSN are known to be
764 loop invariants, false otherwise. */
767 check_dependencies (rtx insn, bitmap depends_on)
769 struct df_ref **use_rec;
770 basic_block bb = BLOCK_FOR_INSN (insn);
772 for (use_rec = DF_INSN_USES (insn); *use_rec; use_rec++)
773 if (!check_dependency (bb, *use_rec, depends_on))
775 for (use_rec = DF_INSN_EQ_USES (insn); *use_rec; use_rec++)
776 if (!check_dependency (bb, *use_rec, depends_on))
782 /* Finds invariant in INSN. ALWAYS_REACHED is true if the insn is always
783 executed. ALWAYS_EXECUTED is true if the insn is always executed,
784 unless the program ends due to a function call. */
787 find_invariant_insn (rtx insn, bool always_reached, bool always_executed)
794 struct invariant *inv;
796 /* Until we get rid of LIBCALLS. */
797 if (find_reg_note (insn, REG_RETVAL, NULL_RTX)
798 || find_reg_note (insn, REG_LIBCALL, NULL_RTX)
799 || find_reg_note (insn, REG_NO_CONFLICT, NULL_RTX))
803 /* We can't move a CC0 setter without the user. */
804 if (sets_cc0_p (insn))
808 set = single_set (insn);
811 dest = SET_DEST (set);
814 || HARD_REGISTER_P (dest))
817 if (!may_assign_reg_p (SET_DEST (set))
818 || !check_maybe_invariant (SET_SRC (set)))
821 /* If the insn can throw exception, we cannot move it at all without changing
823 if (can_throw_internal (insn))
826 /* We cannot make trapping insn executed, unless it was executed before. */
827 if (may_trap_after_code_motion_p (PATTERN (insn)) && !always_reached)
830 depends_on = BITMAP_ALLOC (NULL);
831 if (!check_dependencies (insn, depends_on))
833 BITMAP_FREE (depends_on);
838 def = XCNEW (struct def);
842 inv = create_new_invariant (def, insn, depends_on, always_executed);
846 ref = df_find_def (insn, dest);
847 check_invariant_table_size ();
848 invariant_table[DF_REF_ID(ref)] = inv;
852 /* Record registers used in INSN that have a unique invariant definition. */
855 record_uses (rtx insn)
857 struct df_ref **use_rec;
858 struct invariant *inv;
860 for (use_rec = DF_INSN_USES (insn); *use_rec; use_rec++)
862 struct df_ref *use = *use_rec;
863 inv = invariant_for_use (use);
865 record_use (inv->def, DF_REF_REAL_LOC (use), DF_REF_INSN (use));
867 for (use_rec = DF_INSN_EQ_USES (insn); *use_rec; use_rec++)
869 struct df_ref *use = *use_rec;
870 inv = invariant_for_use (use);
872 record_use (inv->def, DF_REF_REAL_LOC (use), DF_REF_INSN (use));
876 /* Finds invariants in INSN. ALWAYS_REACHED is true if the insn is always
877 executed. ALWAYS_EXECUTED is true if the insn is always executed,
878 unless the program ends due to a function call. */
881 find_invariants_insn (rtx insn, bool always_reached, bool always_executed)
883 find_invariant_insn (insn, always_reached, always_executed);
887 /* Finds invariants in basic block BB. ALWAYS_REACHED is true if the
888 basic block is always executed. ALWAYS_EXECUTED is true if the basic
889 block is always executed, unless the program ends due to a function
893 find_invariants_bb (basic_block bb, bool always_reached, bool always_executed)
897 FOR_BB_INSNS (bb, insn)
902 find_invariants_insn (insn, always_reached, always_executed);
906 && !CONST_OR_PURE_CALL_P (insn))
907 always_reached = false;
911 /* Finds invariants in LOOP with body BODY. ALWAYS_REACHED is the bitmap of
912 basic blocks in BODY that are always executed. ALWAYS_EXECUTED is the
913 bitmap of basic blocks in BODY that are always executed unless the program
914 ends due to a function call. */
917 find_invariants_body (struct loop *loop, basic_block *body,
918 bitmap always_reached, bitmap always_executed)
922 for (i = 0; i < loop->num_nodes; i++)
923 find_invariants_bb (body[i],
924 bitmap_bit_p (always_reached, i),
925 bitmap_bit_p (always_executed, i));
928 /* Finds invariants in LOOP. */
931 find_invariants (struct loop *loop)
933 bitmap may_exit = BITMAP_ALLOC (NULL);
934 bitmap always_reached = BITMAP_ALLOC (NULL);
935 bitmap has_exit = BITMAP_ALLOC (NULL);
936 bitmap always_executed = BITMAP_ALLOC (NULL);
937 basic_block *body = get_loop_body_in_dom_order (loop);
939 find_exits (loop, body, may_exit, has_exit);
940 compute_always_reached (loop, body, may_exit, always_reached);
941 compute_always_reached (loop, body, has_exit, always_executed);
943 find_defs (loop, body);
944 find_invariants_body (loop, body, always_reached, always_executed);
945 merge_identical_invariants ();
947 BITMAP_FREE (always_reached);
948 BITMAP_FREE (always_executed);
949 BITMAP_FREE (may_exit);
950 BITMAP_FREE (has_exit);
954 /* Frees a list of uses USE. */
957 free_use_list (struct use *use)
961 for (; use; use = next)
968 /* Calculates cost and number of registers needed for moving invariant INV
969 out of the loop and stores them to *COST and *REGS_NEEDED. */
972 get_inv_cost (struct invariant *inv, int *comp_cost, unsigned *regs_needed)
975 unsigned aregs_needed;
977 struct invariant *dep;
980 /* Find the representative of the class of the equivalent invariants. */
981 inv = VEC_index (invariant_p, invariants, inv->eqto);
986 || inv->stamp == actual_stamp)
988 inv->stamp = actual_stamp;
991 (*comp_cost) += inv->cost;
995 /* Hoisting constant pool constants into stack regs may cost more than
996 just single register. On x87, the balance is affected both by the
997 small number of FP registers, and by its register stack organization,
998 that forces us to add compensation code in and around the loop to
999 shuffle the operands to the top of stack before use, and pop them
1000 from the stack after the loop finishes.
1002 To model this effect, we increase the number of registers needed for
1003 stack registers by two: one register push, and one register pop.
1004 This usually has the effect that FP constant loads from the constant
1005 pool are not moved out of the loop.
1007 Note that this also means that dependent invariants can not be moved.
1008 However, the primary purpose of this pass is to move loop invariant
1009 address arithmetic out of loops, and address arithmetic that depends
1010 on floating point constants is unlikely to ever occur. */
1011 rtx set = single_set (inv->insn);
1013 && IS_STACK_MODE (GET_MODE (SET_SRC (set)))
1014 && constant_pool_constant_p (SET_SRC (set)))
1015 (*regs_needed) += 2;
1019 EXECUTE_IF_SET_IN_BITMAP (inv->depends_on, 0, depno, bi)
1021 dep = VEC_index (invariant_p, invariants, depno);
1023 get_inv_cost (dep, &acomp_cost, &aregs_needed);
1026 /* We need to check always_executed, since if the original value of
1027 the invariant may be preserved, we may need to keep it in a
1028 separate register. TODO check whether the register has an
1029 use outside of the loop. */
1030 && dep->always_executed
1031 && !dep->def->uses->next)
1033 /* If this is a single use, after moving the dependency we will not
1034 need a new register. */
1038 (*regs_needed) += aregs_needed;
1039 (*comp_cost) += acomp_cost;
1043 /* Calculates gain for eliminating invariant INV. REGS_USED is the number
1044 of registers used in the loop, NEW_REGS is the number of new variables
1045 already added due to the invariant motion. The number of registers needed
1046 for it is stored in *REGS_NEEDED. */
1049 gain_for_invariant (struct invariant *inv, unsigned *regs_needed,
1050 unsigned new_regs, unsigned regs_used)
1052 int comp_cost, size_cost;
1054 get_inv_cost (inv, &comp_cost, regs_needed);
1057 size_cost = (estimate_reg_pressure_cost (new_regs + *regs_needed, regs_used)
1058 - estimate_reg_pressure_cost (new_regs, regs_used));
1060 return comp_cost - size_cost;
1063 /* Finds invariant with best gain for moving. Returns the gain, stores
1064 the invariant in *BEST and number of registers needed for it to
1065 *REGS_NEEDED. REGS_USED is the number of registers used in the loop.
1066 NEW_REGS is the number of new variables already added due to invariant
1070 best_gain_for_invariant (struct invariant **best, unsigned *regs_needed,
1071 unsigned new_regs, unsigned regs_used)
1073 struct invariant *inv;
1074 int gain = 0, again;
1075 unsigned aregs_needed, invno;
1077 for (invno = 0; VEC_iterate (invariant_p, invariants, invno, inv); invno++)
1082 /* Only consider the "representatives" of equivalent invariants. */
1083 if (inv->eqto != inv->invno)
1086 again = gain_for_invariant (inv, &aregs_needed, new_regs, regs_used);
1091 *regs_needed = aregs_needed;
1098 /* Marks invariant INVNO and all its dependencies for moving. */
1101 set_move_mark (unsigned invno)
1103 struct invariant *inv = VEC_index (invariant_p, invariants, invno);
1106 /* Find the representative of the class of the equivalent invariants. */
1107 inv = VEC_index (invariant_p, invariants, inv->eqto);
1114 fprintf (dump_file, "Decided to move invariant %d\n", invno);
1116 EXECUTE_IF_SET_IN_BITMAP (inv->depends_on, 0, invno, bi)
1118 set_move_mark (invno);
1122 /* Determines which invariants to move. */
1125 find_invariants_to_move (void)
1127 unsigned i, regs_used, regs_needed = 0, new_regs;
1128 struct invariant *inv = NULL;
1129 unsigned int n_regs = DF_REG_SIZE (df);
1131 if (!VEC_length (invariant_p, invariants))
1134 /* We do not really do a good job in estimating number of registers used;
1135 we put some initial bound here to stand for induction variables etc.
1136 that we do not detect. */
1139 for (i = 0; i < n_regs; i++)
1141 if (!DF_REGNO_FIRST_DEF (i) && DF_REGNO_LAST_USE (i))
1143 /* This is a value that is used but not changed inside loop. */
1149 while (best_gain_for_invariant (&inv, ®s_needed, new_regs, regs_used) > 0)
1151 set_move_mark (inv->invno);
1152 new_regs += regs_needed;
1156 /* Returns true if all insns in SEQ are valid. */
1159 seq_insns_valid_p (rtx seq)
1163 for (x = seq; x; x = NEXT_INSN (x))
1164 if (insn_invalid_p (x))
1170 /* Move invariant INVNO out of the LOOP. Returns true if this succeeds, false
1174 move_invariant_reg (struct loop *loop, unsigned invno)
1176 struct invariant *inv = VEC_index (invariant_p, invariants, invno);
1177 struct invariant *repr = VEC_index (invariant_p, invariants, inv->eqto);
1179 basic_block preheader = loop_preheader_edge (loop)->src;
1180 rtx reg, set, dest, seq, op;
1188 /* If this is a representative of the class of equivalent invariants,
1189 really move the invariant. Otherwise just replace its use with
1190 the register used for the representative. */
1193 if (inv->depends_on)
1195 EXECUTE_IF_SET_IN_BITMAP (inv->depends_on, 0, i, bi)
1197 if (!move_invariant_reg (loop, i))
1202 /* Move the set out of the loop. If the set is always executed (we could
1203 omit this condition if we know that the register is unused outside of the
1204 loop, but it does not seem worth finding out) and it has no uses that
1205 would not be dominated by it, we may just move it (TODO). Otherwise we
1206 need to create a temporary register. */
1207 set = single_set (inv->insn);
1208 dest = SET_DEST (set);
1209 reg = gen_reg_rtx (GET_MODE (dest));
1211 /* If the SET_DEST of the invariant insn is a pseudo, we can just move
1212 the insn out of the loop. Otherwise, we have to use gen_move_insn
1213 to let emit_move_insn produce a valid instruction stream. */
1214 if (REG_P (dest) && !HARD_REGISTER_P (dest))
1218 emit_insn_after (gen_move_insn (dest, reg), inv->insn);
1219 SET_DEST (set) = reg;
1220 df_insn_rescan (inv->insn);
1221 reorder_insns (inv->insn, inv->insn, BB_END (preheader));
1223 /* If there is a REG_EQUAL note on the insn we just moved, and
1224 insn is in a basic block that is not always executed, the note
1225 may no longer be valid after we move the insn.
1226 Note that uses in REG_EQUAL notes are taken into account in
1227 the computation of invariants. Hence it is safe to retain the
1228 note even if the note contains register references. */
1229 if (! inv->always_executed
1230 && (note = find_reg_note (inv->insn, REG_EQUAL, NULL_RTX)))
1231 remove_note (inv->insn, note);
1236 op = force_operand (SET_SRC (set), reg);
1243 emit_move_insn (reg, op);
1247 if (!seq_insns_valid_p (seq))
1249 emit_insn_after (seq, BB_END (preheader));
1251 emit_insn_after (gen_move_insn (dest, reg), inv->insn);
1252 delete_insn (inv->insn);
1257 if (!move_invariant_reg (loop, repr->invno))
1260 set = single_set (inv->insn);
1261 emit_insn_after (gen_move_insn (SET_DEST (set), reg), inv->insn);
1262 delete_insn (inv->insn);
1268 /* Replace the uses we know to be dominated. It saves work for copy
1269 propagation, and also it is necessary so that dependent invariants
1270 are computed right. */
1273 for (use = inv->def->uses; use; use = use->next)
1276 df_insn_rescan (use->insn);
1283 /* If we failed, clear move flag, so that we do not try to move inv
1286 fprintf (dump_file, "Failed to move invariant %d\n", invno);
1288 inv->reg = NULL_RTX;
1293 /* Move selected invariant out of the LOOP. Newly created regs are marked
1294 in TEMPORARY_REGS. */
1297 move_invariants (struct loop *loop)
1299 struct invariant *inv;
1302 for (i = 0; VEC_iterate (invariant_p, invariants, i, inv); i++)
1303 move_invariant_reg (loop, i);
1306 /* Initializes invariant motion data. */
1309 init_inv_motion_data (void)
1313 invariants = VEC_alloc (invariant_p, heap, 100);
1316 /* Frees the data allocated by invariant motion. */
1319 free_inv_motion_data (void)
1323 struct invariant *inv;
1325 check_invariant_table_size ();
1326 for (i = 0; i < DF_DEFS_TABLE_SIZE (); i++)
1328 inv = invariant_table[i];
1332 gcc_assert (def != NULL);
1334 free_use_list (def->uses);
1336 invariant_table[i] = NULL;
1340 for (i = 0; VEC_iterate (invariant_p, invariants, i, inv); i++)
1342 BITMAP_FREE (inv->depends_on);
1345 VEC_free (invariant_p, heap, invariants);
1348 /* Move the invariants out of the LOOP. */
1351 move_single_loop_invariants (struct loop *loop)
1353 init_inv_motion_data ();
1355 find_invariants (loop);
1356 find_invariants_to_move ();
1357 move_invariants (loop);
1359 free_inv_motion_data ();
1362 /* Releases the auxiliary data for LOOP. */
1365 free_loop_data (struct loop *loop)
1367 struct loop_data *data = LOOP_DATA (loop);
1373 /* Move the invariants out of the loops. */
1376 move_loop_invariants (void)
1381 df_set_flags (DF_EQ_NOTES + DF_DEFER_INSN_RESCAN);
1382 /* Process the loops, innermost first. */
1383 FOR_EACH_LOOP (li, loop, LI_FROM_INNERMOST)
1385 move_single_loop_invariants (loop);
1388 FOR_EACH_LOOP (li, loop, 0)
1390 free_loop_data (loop);
1393 free (invariant_table);
1394 invariant_table = NULL;
1395 invariant_table_size = 0;
1397 #ifdef ENABLE_CHECKING
1398 verify_flow_info ();