1 /* RTL-level loop invariant motion.
2 Copyright (C) 2004, 2005 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 2, 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 COPYING. If not, write to the Free
18 Software Foundation, 51 Franklin Street, Fifth Floor, Boston, MA
21 /* This implements the loop invariant motion pass. It is very simple
22 (no calls, libcalls, etc.). This should be sufficient to cleanup things
23 like address arithmetics -- other more complicated invariants should be
24 eliminated on tree level either in tree-ssa-loop-im.c or in tree-ssa-pre.c.
26 We proceed loop by loop -- it is simpler than trying to handle things
27 globally and should not lose much. First we inspect all sets inside loop
28 and create a dependency graph on insns (saying "to move this insn, you must
29 also move the following insns").
31 We then need to determine what to move. We estimate the number of registers
32 used and move as many invariants as possible while we still have enough free
33 registers. We prefer the expensive invariants.
35 Then we move the selected invariants out of the loop, creating a new
36 temporaries for them if necessary. */
40 #include "coretypes.h"
44 #include "hard-reg-set.h"
46 #include "basic-block.h"
54 /* The data stored for the loop. */
58 struct loop
*outermost_exit
; /* The outermost exit of the loop. */
59 bool has_call
; /* True if the loop contains a call. */
62 #define LOOP_DATA(LOOP) ((struct loop_data *) (LOOP)->aux)
64 /* The description of an use. */
68 rtx
*pos
; /* Position of the use. */
69 rtx insn
; /* The insn in that the use occurs. */
71 struct use
*next
; /* Next use in the list. */
74 /* The description of a def. */
78 struct use
*uses
; /* The list of uses that are uniquely reached
80 unsigned n_uses
; /* Number of such uses. */
81 unsigned invno
; /* The corresponding invariant. */
84 /* The data stored for each invariant. */
88 /* The number of the invariant. */
91 /* Whether we already processed the invariant. */
94 /* The definition of the invariant. */
97 /* The insn in that it is defined. */
100 /* Whether it is always executed. */
101 bool always_executed
;
103 /* Whether to move the invariant. */
106 /* Cost of the invariant. */
109 /* The invariants it depends on. */
112 /* Used for detecting already visited invariants during determining
113 costs of movements. */
117 /* The actual stamp for marking already visited invariants during determining
118 costs of movements. */
120 static unsigned actual_stamp
;
122 typedef struct invariant
*invariant_p
;
124 DEF_VEC_P(invariant_p
);
125 DEF_VEC_ALLOC_P(invariant_p
, heap
);
127 /* The invariants. */
129 static VEC(invariant_p
,heap
) *invariants
;
131 /* The dataflow object. */
133 static struct df
*df
;
135 /* Test for possibility of invariantness of X. */
138 check_maybe_invariant (rtx x
)
140 enum rtx_code code
= GET_CODE (x
);
155 case UNSPEC_VOLATILE
:
163 /* Load/store motion is done elsewhere. ??? Perhaps also add it here?
164 It should not be hard, and might be faster than "elsewhere". */
166 /* Just handle the most trivial case where we load from an unchanging
167 location (most importantly, pic tables). */
168 if (MEM_READONLY_P (x
))
174 /* Don't mess with insns declared volatile. */
175 if (MEM_VOLATILE_P (x
))
183 fmt
= GET_RTX_FORMAT (code
);
184 for (i
= GET_RTX_LENGTH (code
) - 1; i
>= 0; i
--)
188 if (!check_maybe_invariant (XEXP (x
, i
)))
191 else if (fmt
[i
] == 'E')
193 for (j
= 0; j
< XVECLEN (x
, i
); j
++)
194 if (!check_maybe_invariant (XVECEXP (x
, i
, j
)))
202 /* Determines the basic blocks inside LOOP that are always executed and
203 stores their bitmap to ALWAYS_REACHED. MAY_EXIT is a bitmap of
204 basic blocks that may either exit the loop, or contain the call that
205 does not have to return. BODY is body of the loop obtained by
206 get_loop_body_in_dom_order. */
209 compute_always_reached (struct loop
*loop
, basic_block
*body
,
210 bitmap may_exit
, bitmap always_reached
)
214 for (i
= 0; i
< loop
->num_nodes
; i
++)
216 if (dominated_by_p (CDI_DOMINATORS
, loop
->latch
, body
[i
]))
217 bitmap_set_bit (always_reached
, i
);
219 if (bitmap_bit_p (may_exit
, i
))
224 /* Finds exits out of the LOOP with body BODY. Marks blocks in that we may
225 exit the loop by cfg edge to HAS_EXIT and MAY_EXIT. In MAY_EXIT
226 additionally mark blocks that may exit due to a call. */
229 find_exits (struct loop
*loop
, basic_block
*body
,
230 bitmap may_exit
, bitmap has_exit
)
235 struct loop
*outermost_exit
= loop
, *aexit
;
236 bool has_call
= false;
239 for (i
= 0; i
< loop
->num_nodes
; i
++)
241 if (body
[i
]->loop_father
== loop
)
243 FOR_BB_INSNS (body
[i
], insn
)
246 && !CONST_OR_PURE_CALL_P (insn
))
249 bitmap_set_bit (may_exit
, i
);
254 FOR_EACH_EDGE (e
, ei
, body
[i
]->succs
)
256 if (flow_bb_inside_loop_p (loop
, e
->dest
))
259 bitmap_set_bit (may_exit
, i
);
260 bitmap_set_bit (has_exit
, i
);
261 outermost_exit
= find_common_loop (outermost_exit
,
262 e
->dest
->loop_father
);
267 /* Use the data stored for the subloop to decide whether we may exit
268 through it. It is sufficient to do this for header of the loop,
269 as other basic blocks inside it must be dominated by it. */
270 if (body
[i
]->loop_father
->header
!= body
[i
])
273 if (LOOP_DATA (body
[i
]->loop_father
)->has_call
)
276 bitmap_set_bit (may_exit
, i
);
278 aexit
= LOOP_DATA (body
[i
]->loop_father
)->outermost_exit
;
281 bitmap_set_bit (may_exit
, i
);
282 bitmap_set_bit (has_exit
, i
);
284 if (flow_loop_nested_p (aexit
, outermost_exit
))
285 outermost_exit
= aexit
;
289 loop
->aux
= xcalloc (1, sizeof (struct loop_data
));
290 LOOP_DATA (loop
)->outermost_exit
= outermost_exit
;
291 LOOP_DATA (loop
)->has_call
= has_call
;
294 /* Check whether we may assign a value to X from a register. */
297 may_assign_reg_p (rtx x
)
299 return (can_copy_p (GET_MODE (x
))
301 || !HARD_REGISTER_P (x
)
302 || REGNO_REG_CLASS (REGNO (x
)) != NO_REGS
));
305 /* Finds definitions that may correspond to invariants in LOOP with body
309 find_defs (struct loop
*loop
, basic_block
*body
)
312 bitmap blocks
= BITMAP_ALLOC (NULL
);
314 for (i
= 0; i
< loop
->num_nodes
; i
++)
315 bitmap_set_bit (blocks
, body
[i
]->index
);
317 df_analyze_subcfg (df
, blocks
, DF_UD_CHAIN
| DF_HARD_REGS
| DF_EQUIV_NOTES
);
318 BITMAP_FREE (blocks
);
321 /* Creates a new invariant for definition DEF in INSN, depending on invariants
322 in DEPENDS_ON. ALWAYS_EXECUTED is true if the insn is always executed,
323 unless the program ends due to a function call. */
326 create_new_invariant (struct def
*def
, rtx insn
, bitmap depends_on
,
327 bool always_executed
)
329 struct invariant
*inv
= xmalloc (sizeof (struct invariant
));
330 rtx set
= single_set (insn
);
333 inv
->always_executed
= always_executed
;
334 inv
->depends_on
= depends_on
;
336 /* If the set is simple, usually by moving it we move the whole store out of
337 the loop. Otherwise we save only cost of the computation. */
339 inv
->cost
= rtx_cost (set
, SET
);
341 inv
->cost
= rtx_cost (SET_SRC (set
), SET
);
344 inv
->processed
= false;
348 inv
->invno
= VEC_length (invariant_p
, invariants
);
350 def
->invno
= inv
->invno
;
351 VEC_safe_push (invariant_p
, heap
, invariants
, inv
);
356 "Set in insn %d is invariant (%d), cost %d, depends on ",
357 INSN_UID (insn
), inv
->invno
, inv
->cost
);
358 dump_bitmap (dump_file
, inv
->depends_on
);
362 /* Record USE at DEF. */
365 record_use (struct def
*def
, rtx
*use
, rtx insn
)
367 struct use
*u
= xmalloc (sizeof (struct use
));
369 if (GET_CODE (*use
) == SUBREG
)
370 use
= &SUBREG_REG (*use
);
371 gcc_assert (REG_P (*use
));
380 /* Finds the invariants INSN depends on and store them to the DEPENDS_ON
384 check_dependencies (rtx insn
, bitmap depends_on
)
386 struct df_link
*uses
, *defs
;
387 struct ref
*use
, *def
;
388 basic_block bb
= BLOCK_FOR_INSN (insn
), def_bb
;
389 struct def
*def_data
;
391 for (uses
= DF_INSN_USES (df
, insn
); uses
; uses
= uses
->next
)
395 defs
= DF_REF_CHAIN (use
);
403 def_data
= DF_REF_DATA (def
);
407 def_bb
= DF_REF_BB (def
);
408 if (!dominated_by_p (CDI_DOMINATORS
, bb
, def_bb
))
411 bitmap_set_bit (depends_on
, def_data
->invno
);
417 /* Finds invariant in INSN. ALWAYS_REACHED is true if the insn is always
418 executed. ALWAYS_EXECUTED is true if the insn is always executed,
419 unless the program ends due to a function call. */
422 find_invariant_insn (rtx insn
, bool always_reached
, bool always_executed
)
430 /* Until we get rid of LIBCALLS. */
431 if (find_reg_note (insn
, REG_RETVAL
, NULL_RTX
)
432 || find_reg_note (insn
, REG_LIBCALL
, NULL_RTX
)
433 || find_reg_note (insn
, REG_NO_CONFLICT
, NULL_RTX
))
436 set
= single_set (insn
);
439 dest
= SET_DEST (set
);
442 || HARD_REGISTER_P (dest
))
445 if (!may_assign_reg_p (SET_DEST (set
))
446 || !check_maybe_invariant (SET_SRC (set
)))
449 if (may_trap_p (PATTERN (insn
)))
454 /* Unless the exceptions are handled, the behavior is undefined
455 if the trap occurs. */
456 if (flag_non_call_exceptions
)
460 depends_on
= BITMAP_ALLOC (NULL
);
461 if (!check_dependencies (insn
, depends_on
))
463 BITMAP_FREE (depends_on
);
469 ref
= df_find_def (df
, insn
, dest
);
470 def
= xcalloc (1, sizeof (struct def
));
471 DF_REF_DATA (ref
) = def
;
476 create_new_invariant (def
, insn
, depends_on
, always_executed
);
479 /* Record registers used in INSN that have a unique invariant definition. */
482 record_uses (rtx insn
)
484 struct df_link
*uses
, *defs
;
485 struct ref
*use
, *def
;
486 basic_block bb
= BLOCK_FOR_INSN (insn
), def_bb
;
488 for (uses
= DF_INSN_USES (df
, insn
); uses
; uses
= uses
->next
)
492 defs
= DF_REF_CHAIN (use
);
493 if (!defs
|| defs
->next
)
496 if (!DF_REF_DATA (def
))
499 def_bb
= DF_REF_BB (def
);
500 if (!dominated_by_p (CDI_DOMINATORS
, bb
, def_bb
))
503 record_use (DF_REF_DATA (def
), DF_REF_LOC (use
), DF_REF_INSN (use
));
507 /* Finds invariants in INSN. ALWAYS_REACHED is true if the insn is always
508 executed. ALWAYS_EXECUTED is true if the insn is always executed,
509 unless the program ends due to a function call. */
512 find_invariants_insn (rtx insn
, bool always_reached
, bool always_executed
)
514 find_invariant_insn (insn
, always_reached
, always_executed
);
518 /* Finds invariants in basic block BB. ALWAYS_REACHED is true if the
519 basic block is always executed. ALWAYS_EXECUTED is true if the basic
520 block is always executed, unless the program ends due to a function
524 find_invariants_bb (basic_block bb
, bool always_reached
, bool always_executed
)
528 FOR_BB_INSNS (bb
, insn
)
533 find_invariants_insn (insn
, always_reached
, always_executed
);
537 && !CONST_OR_PURE_CALL_P (insn
))
538 always_reached
= false;
542 /* Finds invariants in LOOP with body BODY. ALWAYS_REACHED is the bitmap of
543 basic blocks in BODY that are always executed. ALWAYS_EXECUTED is the
544 bitmap of basic blocks in BODY that are always executed unless the program
545 ends due to a function call. */
548 find_invariants_body (struct loop
*loop
, basic_block
*body
,
549 bitmap always_reached
, bitmap always_executed
)
553 for (i
= 0; i
< loop
->num_nodes
; i
++)
554 find_invariants_bb (body
[i
],
555 bitmap_bit_p (always_reached
, i
),
556 bitmap_bit_p (always_executed
, i
));
559 /* Finds invariants in LOOP. */
562 find_invariants (struct loop
*loop
)
564 bitmap may_exit
= BITMAP_ALLOC (NULL
);
565 bitmap always_reached
= BITMAP_ALLOC (NULL
);
566 bitmap has_exit
= BITMAP_ALLOC (NULL
);
567 bitmap always_executed
= BITMAP_ALLOC (NULL
);
568 basic_block
*body
= get_loop_body_in_dom_order (loop
);
570 find_exits (loop
, body
, may_exit
, has_exit
);
571 compute_always_reached (loop
, body
, may_exit
, always_reached
);
572 compute_always_reached (loop
, body
, has_exit
, always_executed
);
574 find_defs (loop
, body
);
575 find_invariants_body (loop
, body
, always_reached
, always_executed
);
577 BITMAP_FREE (always_reached
);
578 BITMAP_FREE (always_executed
);
579 BITMAP_FREE (may_exit
);
580 BITMAP_FREE (has_exit
);
584 /* Frees a list of uses USE. */
587 free_use_list (struct use
*use
)
591 for (; use
; use
= next
)
598 /* Calculates cost and number of registers needed for moving invariant INV
599 out of the loop and stores them to *COST and *REGS_NEEDED. */
602 get_inv_cost (struct invariant
*inv
, int *comp_cost
, unsigned *regs_needed
)
605 unsigned aregs_needed
;
607 struct invariant
*dep
;
613 || inv
->stamp
== actual_stamp
)
615 inv
->stamp
= actual_stamp
;
618 (*comp_cost
) += inv
->cost
;
620 EXECUTE_IF_SET_IN_BITMAP (inv
->depends_on
, 0, depno
, bi
)
622 dep
= VEC_index (invariant_p
, invariants
, depno
);
624 get_inv_cost (dep
, &acomp_cost
, &aregs_needed
);
627 /* We need to check always_executed, since if the original value of
628 the invariant may be preserved, we may need to keep it in a
629 separate register. TODO check whether the register has an
630 use outside of the loop. */
631 && dep
->always_executed
632 && !dep
->def
->uses
->next
)
634 /* If this is a single use, after moving the dependency we will not
635 need a new register. */
639 (*regs_needed
) += aregs_needed
;
640 (*comp_cost
) += acomp_cost
;
644 /* Calculates gain for eliminating invariant INV. REGS_USED is the number
645 of registers used in the loop, N_INV_USES is the number of uses of
646 invariants, NEW_REGS is the number of new variables already added due to
647 the invariant motion. The number of registers needed for it is stored in
651 gain_for_invariant (struct invariant
*inv
, unsigned *regs_needed
,
652 unsigned new_regs
, unsigned regs_used
, unsigned n_inv_uses
)
654 int comp_cost
, size_cost
;
656 get_inv_cost (inv
, &comp_cost
, regs_needed
);
659 size_cost
= (global_cost_for_size (new_regs
+ *regs_needed
,
660 regs_used
, n_inv_uses
)
661 - global_cost_for_size (new_regs
, regs_used
, n_inv_uses
));
663 return comp_cost
- size_cost
;
666 /* Finds invariant with best gain for moving. Returns the gain, stores
667 the invariant in *BEST and number of registers needed for it to
668 *REGS_NEEDED. REGS_USED is the number of registers used in
669 the loop, N_INV_USES is the number of uses of invariants. NEW_REGS
670 is the number of new variables already added due to invariant motion. */
673 best_gain_for_invariant (struct invariant
**best
, unsigned *regs_needed
,
674 unsigned new_regs
, unsigned regs_used
,
677 struct invariant
*inv
;
679 unsigned aregs_needed
, invno
;
681 for (invno
= 0; VEC_iterate (invariant_p
, invariants
, invno
, inv
); invno
++)
686 again
= gain_for_invariant (inv
, &aregs_needed
,
687 new_regs
, regs_used
, n_inv_uses
);
692 *regs_needed
= aregs_needed
;
699 /* Marks invariant INVNO and all its dependencies for moving. */
702 set_move_mark (unsigned invno
)
704 struct invariant
*inv
= VEC_index (invariant_p
, invariants
, invno
);
712 fprintf (dump_file
, "Decided to move invariant %d\n", invno
);
714 EXECUTE_IF_SET_IN_BITMAP (inv
->depends_on
, 0, invno
, bi
)
716 set_move_mark (invno
);
720 /* Determines which invariants to move. */
723 find_invariants_to_move (void)
725 unsigned i
, regs_used
, n_inv_uses
, regs_needed
= 0, new_regs
;
726 struct invariant
*inv
= NULL
;
728 if (!VEC_length (invariant_p
, invariants
))
731 /* Now something slightly more involved. First estimate the number of used
735 /* We do not really do a good job in this estimation; put some initial bound
736 here to stand for induction variables etc. that we do not detect. */
739 for (i
= 0; i
< df
->n_regs
; i
++)
741 if (!DF_REGNO_FIRST_DEF (df
, i
) && DF_REGNO_LAST_USE (df
, i
))
743 /* This is a value that is used but not changed inside loop. */
748 for (i
= 0; VEC_iterate (invariant_p
, invariants
, i
, inv
); i
++)
751 n_inv_uses
+= inv
->def
->n_uses
;
755 while (best_gain_for_invariant (&inv
, ®s_needed
,
756 new_regs
, regs_used
, n_inv_uses
) > 0)
758 set_move_mark (inv
->invno
);
759 new_regs
+= regs_needed
;
763 /* Move invariant INVNO out of the LOOP. */
766 move_invariant_reg (struct loop
*loop
, unsigned invno
)
768 struct invariant
*inv
= VEC_index (invariant_p
, invariants
, invno
);
770 basic_block preheader
= loop_preheader_edge (loop
)->src
;
777 inv
->processed
= true;
781 EXECUTE_IF_SET_IN_BITMAP (inv
->depends_on
, 0, i
, bi
)
783 move_invariant_reg (loop
, i
);
787 /* Move the set out of the loop. If the set is always executed (we could
788 omit this condition if we know that the register is unused outside of the
789 loop, but it does not seem worth finding out) and it has no uses that
790 would not be dominated by it, we may just move it (TODO). Otherwise we
791 need to create a temporary register. */
792 set
= single_set (inv
->insn
);
793 reg
= gen_reg_rtx (GET_MODE (SET_DEST (set
)));
794 df_pattern_emit_after (df
, gen_move_insn (SET_DEST (set
), reg
),
795 BLOCK_FOR_INSN (inv
->insn
), inv
->insn
);
797 /* If the SET_DEST of the invariant insn is a reg, we can just move
798 the insn out of the loop. Otherwise, we have to use gen_move_insn
799 to let emit_move_insn produce a valid instruction stream. */
800 if (REG_P (SET_DEST (set
)))
802 SET_DEST (set
) = reg
;
803 reorder_insns (inv
->insn
, inv
->insn
, BB_END (preheader
));
804 df_insn_modify (df
, preheader
, inv
->insn
);
808 df_pattern_emit_after (df
, gen_move_insn (reg
, SET_SRC (set
)),
809 preheader
, BB_END (preheader
));
810 df_insn_delete (df
, BLOCK_FOR_INSN (inv
->insn
), inv
->insn
);
813 /* Replace the uses we know to be dominated. It saves work for copy
814 propagation, and also it is necessary so that dependent invariants
815 are computed right. */
818 for (use
= inv
->def
->uses
; use
; use
= use
->next
)
821 df_insn_modify (df
, BLOCK_FOR_INSN (use
->insn
), use
->insn
);
826 /* Move selected invariant out of the LOOP. Newly created regs are marked
827 in TEMPORARY_REGS. */
830 move_invariants (struct loop
*loop
)
832 struct invariant
*inv
;
835 for (i
= 0; VEC_iterate (invariant_p
, invariants
, i
, inv
); i
++)
838 move_invariant_reg (loop
, i
);
842 /* Initializes invariant motion data. */
845 init_inv_motion_data (void)
849 invariants
= VEC_alloc (invariant_p
, heap
, 100);
852 /* Frees the data allocated by invariant motion. */
855 free_inv_motion_data (void)
859 struct invariant
*inv
;
861 for (i
= 0; i
< df
->n_defs
; i
++)
866 def
= DF_REF_DATA (df
->defs
[i
]);
870 free_use_list (def
->uses
);
872 DF_REF_DATA (df
->defs
[i
]) = NULL
;
875 for (i
= 0; VEC_iterate (invariant_p
, invariants
, i
, inv
); i
++)
877 BITMAP_FREE (inv
->depends_on
);
880 VEC_free (invariant_p
, heap
, invariants
);
883 /* Move the invariants out of the LOOP. */
886 move_single_loop_invariants (struct loop
*loop
)
888 init_inv_motion_data ();
890 find_invariants (loop
);
891 find_invariants_to_move ();
892 move_invariants (loop
);
894 free_inv_motion_data ();
897 /* Releases the auxiliary data for LOOP. */
900 free_loop_data (struct loop
*loop
)
902 struct loop_data
*data
= LOOP_DATA (loop
);
908 /* Move the invariants out of the LOOPS. */
911 move_loop_invariants (struct loops
*loops
)
918 /* Process the loops, innermost first. */
919 loop
= loops
->tree_root
;
923 while (loop
!= loops
->tree_root
)
925 move_single_loop_invariants (loop
);
937 for (i
= 1; i
< loops
->num
; i
++)
938 if (loops
->parray
[i
])
939 free_loop_data (loops
->parray
[i
]);
944 #ifdef ENABLE_CHECKING