1 /* Loop unrolling and peeling.
2 Copyright (C) 2002, 2003, 2004, 2005, 2007, 2008, 2010, 2011
3 Free Software Foundation, Inc.
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 3, 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 COPYING3. If not see
19 <http://www.gnu.org/licenses/>. */
23 #include "coretypes.h"
26 #include "hard-reg-set.h"
28 #include "basic-block.h"
30 #include "cfglayout.h"
38 /* This pass performs loop unrolling and peeling. We only perform these
39 optimizations on innermost loops (with single exception) because
40 the impact on performance is greatest here, and we want to avoid
41 unnecessary code size growth. The gain is caused by greater sequentiality
42 of code, better code to optimize for further passes and in some cases
43 by fewer testings of exit conditions. The main problem is code growth,
44 that impacts performance negatively due to effect of caches.
48 -- complete peeling of once-rolling loops; this is the above mentioned
49 exception, as this causes loop to be cancelled completely and
50 does not cause code growth
51 -- complete peeling of loops that roll (small) constant times.
52 -- simple peeling of first iterations of loops that do not roll much
53 (according to profile feedback)
54 -- unrolling of loops that roll constant times; this is almost always
55 win, as we get rid of exit condition tests.
56 -- unrolling of loops that roll number of times that we can compute
57 in runtime; we also get rid of exit condition tests here, but there
58 is the extra expense for calculating the number of iterations
59 -- simple unrolling of remaining loops; this is performed only if we
60 are asked to, as the gain is questionable in this case and often
61 it may even slow down the code
62 For more detailed descriptions of each of those, see comments at
63 appropriate function below.
65 There is a lot of parameters (defined and described in params.def) that
66 control how much we unroll/peel.
68 ??? A great problem is that we don't have a good way how to determine
69 how many times we should unroll the loop; the experiments I have made
70 showed that this choice may affect performance in order of several %.
73 /* Information about induction variables to split. */
77 rtx insn
; /* The insn in that the induction variable occurs. */
78 rtx base_var
; /* The variable on that the values in the further
79 iterations are based. */
80 rtx step
; /* Step of the induction variable. */
81 struct iv_to_split
*next
; /* Next entry in walking order. */
83 unsigned loc
[3]; /* Location where the definition of the induction
84 variable occurs in the insn. For example if
85 N_LOC is 2, the expression is located at
86 XEXP (XEXP (single_set, loc[0]), loc[1]). */
89 /* Information about accumulators to expand. */
93 rtx insn
; /* The insn in that the variable expansion occurs. */
94 rtx reg
; /* The accumulator which is expanded. */
95 VEC(rtx
,heap
) *var_expansions
; /* The copies of the accumulator which is expanded. */
96 struct var_to_expand
*next
; /* Next entry in walking order. */
97 enum rtx_code op
; /* The type of the accumulation - addition, subtraction
99 int expansion_count
; /* Count the number of expansions generated so far. */
100 int reuse_expansion
; /* The expansion we intend to reuse to expand
101 the accumulator. If REUSE_EXPANSION is 0 reuse
102 the original accumulator. Else use
103 var_expansions[REUSE_EXPANSION - 1]. */
104 unsigned accum_pos
; /* The position in which the accumulator is placed in
105 the insn src. For example in x = x + something
106 accum_pos is 0 while in x = something + x accum_pos
110 /* Information about optimization applied in
111 the unrolled loop. */
115 htab_t insns_to_split
; /* A hashtable of insns to split. */
116 struct iv_to_split
*iv_to_split_head
; /* The first iv to split. */
117 struct iv_to_split
**iv_to_split_tail
; /* Pointer to the tail of the list. */
118 htab_t insns_with_var_to_expand
; /* A hashtable of insns with accumulators
120 struct var_to_expand
*var_to_expand_head
; /* The first var to expand. */
121 struct var_to_expand
**var_to_expand_tail
; /* Pointer to the tail of the list. */
122 unsigned first_new_block
; /* The first basic block that was
124 basic_block loop_exit
; /* The loop exit basic block. */
125 basic_block loop_preheader
; /* The loop preheader basic block. */
128 static void decide_unrolling_and_peeling (int);
129 static void peel_loops_completely (int);
130 static void decide_peel_simple (struct loop
*, int);
131 static void decide_peel_once_rolling (struct loop
*, int);
132 static void decide_peel_completely (struct loop
*, int);
133 static void decide_unroll_stupid (struct loop
*, int);
134 static void decide_unroll_constant_iterations (struct loop
*, int);
135 static void decide_unroll_runtime_iterations (struct loop
*, int);
136 static void peel_loop_simple (struct loop
*);
137 static void peel_loop_completely (struct loop
*);
138 static void unroll_loop_stupid (struct loop
*);
139 static void unroll_loop_constant_iterations (struct loop
*);
140 static void unroll_loop_runtime_iterations (struct loop
*);
141 static struct opt_info
*analyze_insns_in_loop (struct loop
*);
142 static void opt_info_start_duplication (struct opt_info
*);
143 static void apply_opt_in_copies (struct opt_info
*, unsigned, bool, bool);
144 static void free_opt_info (struct opt_info
*);
145 static struct var_to_expand
*analyze_insn_to_expand_var (struct loop
*, rtx
);
146 static bool referenced_in_one_insn_in_loop_p (struct loop
*, rtx
, int *);
147 static struct iv_to_split
*analyze_iv_to_split_insn (rtx
);
148 static void expand_var_during_unrolling (struct var_to_expand
*, rtx
);
149 static void insert_var_expansion_initialization (struct var_to_expand
*,
151 static void combine_var_copies_in_loop_exit (struct var_to_expand
*,
153 static rtx
get_expansion (struct var_to_expand
*);
155 /* Unroll and/or peel (depending on FLAGS) LOOPS. */
157 unroll_and_peel_loops (int flags
)
163 /* First perform complete loop peeling (it is almost surely a win,
164 and affects parameters for further decision a lot). */
165 peel_loops_completely (flags
);
167 /* Now decide rest of unrolling and peeling. */
168 decide_unrolling_and_peeling (flags
);
170 /* Scan the loops, inner ones first. */
171 FOR_EACH_LOOP (li
, loop
, LI_FROM_INNERMOST
)
174 /* And perform the appropriate transformations. */
175 switch (loop
->lpt_decision
.decision
)
177 case LPT_PEEL_COMPLETELY
:
180 case LPT_PEEL_SIMPLE
:
181 peel_loop_simple (loop
);
183 case LPT_UNROLL_CONSTANT
:
184 unroll_loop_constant_iterations (loop
);
186 case LPT_UNROLL_RUNTIME
:
187 unroll_loop_runtime_iterations (loop
);
189 case LPT_UNROLL_STUPID
:
190 unroll_loop_stupid (loop
);
200 #ifdef ENABLE_CHECKING
201 verify_loop_structure ();
209 /* Check whether exit of the LOOP is at the end of loop body. */
212 loop_exit_at_end_p (struct loop
*loop
)
214 struct niter_desc
*desc
= get_simple_loop_desc (loop
);
217 if (desc
->in_edge
->dest
!= loop
->latch
)
220 /* Check that the latch is empty. */
221 FOR_BB_INSNS (loop
->latch
, insn
)
230 /* Depending on FLAGS, check whether to peel loops completely and do so. */
232 peel_loops_completely (int flags
)
237 /* Scan the loops, the inner ones first. */
238 FOR_EACH_LOOP (li
, loop
, LI_FROM_INNERMOST
)
240 loop
->lpt_decision
.decision
= LPT_NONE
;
244 "\n;; *** Considering loop %d for complete peeling ***\n",
247 loop
->ninsns
= num_loop_insns (loop
);
249 decide_peel_once_rolling (loop
, flags
);
250 if (loop
->lpt_decision
.decision
== LPT_NONE
)
251 decide_peel_completely (loop
, flags
);
253 if (loop
->lpt_decision
.decision
== LPT_PEEL_COMPLETELY
)
255 peel_loop_completely (loop
);
256 #ifdef ENABLE_CHECKING
257 verify_loop_structure ();
263 /* Decide whether unroll or peel loops (depending on FLAGS) and how much. */
265 decide_unrolling_and_peeling (int flags
)
270 /* Scan the loops, inner ones first. */
271 FOR_EACH_LOOP (li
, loop
, LI_FROM_INNERMOST
)
273 loop
->lpt_decision
.decision
= LPT_NONE
;
276 fprintf (dump_file
, "\n;; *** Considering loop %d ***\n", loop
->num
);
278 /* Do not peel cold areas. */
279 if (optimize_loop_for_size_p (loop
))
282 fprintf (dump_file
, ";; Not considering loop, cold area\n");
286 /* Can the loop be manipulated? */
287 if (!can_duplicate_loop_p (loop
))
291 ";; Not considering loop, cannot duplicate\n");
295 /* Skip non-innermost loops. */
299 fprintf (dump_file
, ";; Not considering loop, is not innermost\n");
303 loop
->ninsns
= num_loop_insns (loop
);
304 loop
->av_ninsns
= average_num_loop_insns (loop
);
306 /* Try transformations one by one in decreasing order of
309 decide_unroll_constant_iterations (loop
, flags
);
310 if (loop
->lpt_decision
.decision
== LPT_NONE
)
311 decide_unroll_runtime_iterations (loop
, flags
);
312 if (loop
->lpt_decision
.decision
== LPT_NONE
)
313 decide_unroll_stupid (loop
, flags
);
314 if (loop
->lpt_decision
.decision
== LPT_NONE
)
315 decide_peel_simple (loop
, flags
);
319 /* Decide whether the LOOP is once rolling and suitable for complete
322 decide_peel_once_rolling (struct loop
*loop
, int flags ATTRIBUTE_UNUSED
)
324 struct niter_desc
*desc
;
327 fprintf (dump_file
, "\n;; Considering peeling once rolling loop\n");
329 /* Is the loop small enough? */
330 if ((unsigned) PARAM_VALUE (PARAM_MAX_ONCE_PEELED_INSNS
) < loop
->ninsns
)
333 fprintf (dump_file
, ";; Not considering loop, is too big\n");
337 /* Check for simple loops. */
338 desc
= get_simple_loop_desc (loop
);
340 /* Check number of iterations. */
349 ";; Unable to prove that the loop rolls exactly once\n");
355 fprintf (dump_file
, ";; Decided to peel exactly once rolling loop\n");
356 loop
->lpt_decision
.decision
= LPT_PEEL_COMPLETELY
;
359 /* Decide whether the LOOP is suitable for complete peeling. */
361 decide_peel_completely (struct loop
*loop
, int flags ATTRIBUTE_UNUSED
)
364 struct niter_desc
*desc
;
367 fprintf (dump_file
, "\n;; Considering peeling completely\n");
369 /* Skip non-innermost loops. */
373 fprintf (dump_file
, ";; Not considering loop, is not innermost\n");
377 /* Do not peel cold areas. */
378 if (optimize_loop_for_size_p (loop
))
381 fprintf (dump_file
, ";; Not considering loop, cold area\n");
385 /* Can the loop be manipulated? */
386 if (!can_duplicate_loop_p (loop
))
390 ";; Not considering loop, cannot duplicate\n");
394 /* npeel = number of iterations to peel. */
395 npeel
= PARAM_VALUE (PARAM_MAX_COMPLETELY_PEELED_INSNS
) / loop
->ninsns
;
396 if (npeel
> (unsigned) PARAM_VALUE (PARAM_MAX_COMPLETELY_PEEL_TIMES
))
397 npeel
= PARAM_VALUE (PARAM_MAX_COMPLETELY_PEEL_TIMES
);
399 /* Is the loop small enough? */
403 fprintf (dump_file
, ";; Not considering loop, is too big\n");
407 /* Check for simple loops. */
408 desc
= get_simple_loop_desc (loop
);
410 /* Check number of iterations. */
418 ";; Unable to prove that the loop iterates constant times\n");
422 if (desc
->niter
> npeel
- 1)
427 ";; Not peeling loop completely, rolls too much (");
428 fprintf (dump_file
, HOST_WIDEST_INT_PRINT_DEC
, desc
->niter
);
429 fprintf (dump_file
, " iterations > %d [maximum peelings])\n", npeel
);
436 fprintf (dump_file
, ";; Decided to peel loop completely\n");
437 loop
->lpt_decision
.decision
= LPT_PEEL_COMPLETELY
;
440 /* Peel all iterations of LOOP, remove exit edges and cancel the loop
441 completely. The transformation done:
443 for (i = 0; i < 4; i++)
455 peel_loop_completely (struct loop
*loop
)
458 unsigned HOST_WIDE_INT npeel
;
460 VEC (edge
, heap
) *remove_edges
;
462 struct niter_desc
*desc
= get_simple_loop_desc (loop
);
463 struct opt_info
*opt_info
= NULL
;
471 wont_exit
= sbitmap_alloc (npeel
+ 1);
472 sbitmap_ones (wont_exit
);
473 RESET_BIT (wont_exit
, 0);
474 if (desc
->noloop_assumptions
)
475 RESET_BIT (wont_exit
, 1);
479 if (flag_split_ivs_in_unroller
)
480 opt_info
= analyze_insns_in_loop (loop
);
482 opt_info_start_duplication (opt_info
);
483 ok
= duplicate_loop_to_header_edge (loop
, loop_preheader_edge (loop
),
485 wont_exit
, desc
->out_edge
,
487 DLTHE_FLAG_UPDATE_FREQ
488 | DLTHE_FLAG_COMPLETTE_PEEL
490 ? DLTHE_RECORD_COPY_NUMBER
: 0));
497 apply_opt_in_copies (opt_info
, npeel
, false, true);
498 free_opt_info (opt_info
);
501 /* Remove the exit edges. */
502 FOR_EACH_VEC_ELT (edge
, remove_edges
, i
, ein
)
504 VEC_free (edge
, heap
, remove_edges
);
508 free_simple_loop_desc (loop
);
510 /* Now remove the unreachable part of the last iteration and cancel
515 fprintf (dump_file
, ";; Peeled loop completely, %d times\n", (int) npeel
);
518 /* Decide whether to unroll LOOP iterating constant number of times
522 decide_unroll_constant_iterations (struct loop
*loop
, int flags
)
524 unsigned nunroll
, nunroll_by_av
, best_copies
, best_unroll
= 0, n_copies
, i
;
525 struct niter_desc
*desc
;
527 if (!(flags
& UAP_UNROLL
))
529 /* We were not asked to, just return back silently. */
535 "\n;; Considering unrolling loop with constant "
536 "number of iterations\n");
538 /* nunroll = total number of copies of the original loop body in
539 unrolled loop (i.e. if it is 2, we have to duplicate loop body once. */
540 nunroll
= PARAM_VALUE (PARAM_MAX_UNROLLED_INSNS
) / loop
->ninsns
;
542 = PARAM_VALUE (PARAM_MAX_AVERAGE_UNROLLED_INSNS
) / loop
->av_ninsns
;
543 if (nunroll
> nunroll_by_av
)
544 nunroll
= nunroll_by_av
;
545 if (nunroll
> (unsigned) PARAM_VALUE (PARAM_MAX_UNROLL_TIMES
))
546 nunroll
= PARAM_VALUE (PARAM_MAX_UNROLL_TIMES
);
548 /* Skip big loops. */
552 fprintf (dump_file
, ";; Not considering loop, is too big\n");
556 /* Check for simple loops. */
557 desc
= get_simple_loop_desc (loop
);
559 /* Check number of iterations. */
560 if (!desc
->simple_p
|| !desc
->const_iter
|| desc
->assumptions
)
564 ";; Unable to prove that the loop iterates constant times\n");
568 /* Check whether the loop rolls enough to consider. */
569 if (desc
->niter
< 2 * nunroll
)
572 fprintf (dump_file
, ";; Not unrolling loop, doesn't roll\n");
576 /* Success; now compute number of iterations to unroll. We alter
577 nunroll so that as few as possible copies of loop body are
578 necessary, while still not decreasing the number of unrollings
579 too much (at most by 1). */
580 best_copies
= 2 * nunroll
+ 10;
583 if (i
- 1 >= desc
->niter
)
586 for (; i
>= nunroll
- 1; i
--)
588 unsigned exit_mod
= desc
->niter
% (i
+ 1);
590 if (!loop_exit_at_end_p (loop
))
591 n_copies
= exit_mod
+ i
+ 1;
592 else if (exit_mod
!= (unsigned) i
593 || desc
->noloop_assumptions
!= NULL_RTX
)
594 n_copies
= exit_mod
+ i
+ 2;
598 if (n_copies
< best_copies
)
600 best_copies
= n_copies
;
606 fprintf (dump_file
, ";; max_unroll %d (%d copies, initial %d).\n",
607 best_unroll
+ 1, best_copies
, nunroll
);
609 loop
->lpt_decision
.decision
= LPT_UNROLL_CONSTANT
;
610 loop
->lpt_decision
.times
= best_unroll
;
614 ";; Decided to unroll the constant times rolling loop, %d times.\n",
615 loop
->lpt_decision
.times
);
618 /* Unroll LOOP with constant number of iterations LOOP->LPT_DECISION.TIMES + 1
619 times. The transformation does this:
621 for (i = 0; i < 102; i++)
638 unroll_loop_constant_iterations (struct loop
*loop
)
640 unsigned HOST_WIDE_INT niter
;
644 VEC (edge
, heap
) *remove_edges
;
646 unsigned max_unroll
= loop
->lpt_decision
.times
;
647 struct niter_desc
*desc
= get_simple_loop_desc (loop
);
648 bool exit_at_end
= loop_exit_at_end_p (loop
);
649 struct opt_info
*opt_info
= NULL
;
654 /* Should not get here (such loop should be peeled instead). */
655 gcc_assert (niter
> max_unroll
+ 1);
657 exit_mod
= niter
% (max_unroll
+ 1);
659 wont_exit
= sbitmap_alloc (max_unroll
+ 1);
660 sbitmap_ones (wont_exit
);
663 if (flag_split_ivs_in_unroller
664 || flag_variable_expansion_in_unroller
)
665 opt_info
= analyze_insns_in_loop (loop
);
669 /* The exit is not at the end of the loop; leave exit test
670 in the first copy, so that the loops that start with test
671 of exit condition have continuous body after unrolling. */
674 fprintf (dump_file
, ";; Condition on beginning of loop.\n");
676 /* Peel exit_mod iterations. */
677 RESET_BIT (wont_exit
, 0);
678 if (desc
->noloop_assumptions
)
679 RESET_BIT (wont_exit
, 1);
683 opt_info_start_duplication (opt_info
);
684 ok
= duplicate_loop_to_header_edge (loop
, loop_preheader_edge (loop
),
686 wont_exit
, desc
->out_edge
,
688 DLTHE_FLAG_UPDATE_FREQ
689 | (opt_info
&& exit_mod
> 1
690 ? DLTHE_RECORD_COPY_NUMBER
694 if (opt_info
&& exit_mod
> 1)
695 apply_opt_in_copies (opt_info
, exit_mod
, false, false);
697 desc
->noloop_assumptions
= NULL_RTX
;
698 desc
->niter
-= exit_mod
;
699 desc
->niter_max
-= exit_mod
;
702 SET_BIT (wont_exit
, 1);
706 /* Leave exit test in last copy, for the same reason as above if
707 the loop tests the condition at the end of loop body. */
710 fprintf (dump_file
, ";; Condition on end of loop.\n");
712 /* We know that niter >= max_unroll + 2; so we do not need to care of
713 case when we would exit before reaching the loop. So just peel
714 exit_mod + 1 iterations. */
715 if (exit_mod
!= max_unroll
716 || desc
->noloop_assumptions
)
718 RESET_BIT (wont_exit
, 0);
719 if (desc
->noloop_assumptions
)
720 RESET_BIT (wont_exit
, 1);
722 opt_info_start_duplication (opt_info
);
723 ok
= duplicate_loop_to_header_edge (loop
, loop_preheader_edge (loop
),
725 wont_exit
, desc
->out_edge
,
727 DLTHE_FLAG_UPDATE_FREQ
728 | (opt_info
&& exit_mod
> 0
729 ? DLTHE_RECORD_COPY_NUMBER
733 if (opt_info
&& exit_mod
> 0)
734 apply_opt_in_copies (opt_info
, exit_mod
+ 1, false, false);
736 desc
->niter
-= exit_mod
+ 1;
737 desc
->niter_max
-= exit_mod
+ 1;
738 desc
->noloop_assumptions
= NULL_RTX
;
740 SET_BIT (wont_exit
, 0);
741 SET_BIT (wont_exit
, 1);
744 RESET_BIT (wont_exit
, max_unroll
);
747 /* Now unroll the loop. */
749 opt_info_start_duplication (opt_info
);
750 ok
= duplicate_loop_to_header_edge (loop
, loop_latch_edge (loop
),
752 wont_exit
, desc
->out_edge
,
754 DLTHE_FLAG_UPDATE_FREQ
756 ? DLTHE_RECORD_COPY_NUMBER
762 apply_opt_in_copies (opt_info
, max_unroll
, true, true);
763 free_opt_info (opt_info
);
770 basic_block exit_block
= get_bb_copy (desc
->in_edge
->src
);
771 /* Find a new in and out edge; they are in the last copy we have made. */
773 if (EDGE_SUCC (exit_block
, 0)->dest
== desc
->out_edge
->dest
)
775 desc
->out_edge
= EDGE_SUCC (exit_block
, 0);
776 desc
->in_edge
= EDGE_SUCC (exit_block
, 1);
780 desc
->out_edge
= EDGE_SUCC (exit_block
, 1);
781 desc
->in_edge
= EDGE_SUCC (exit_block
, 0);
785 desc
->niter
/= max_unroll
+ 1;
786 desc
->niter_max
/= max_unroll
+ 1;
787 desc
->niter_expr
= GEN_INT (desc
->niter
);
789 /* Remove the edges. */
790 FOR_EACH_VEC_ELT (edge
, remove_edges
, i
, e
)
792 VEC_free (edge
, heap
, remove_edges
);
796 ";; Unrolled loop %d times, constant # of iterations %i insns\n",
797 max_unroll
, num_loop_insns (loop
));
800 /* Decide whether to unroll LOOP iterating runtime computable number of times
803 decide_unroll_runtime_iterations (struct loop
*loop
, int flags
)
805 unsigned nunroll
, nunroll_by_av
, i
;
806 struct niter_desc
*desc
;
808 if (!(flags
& UAP_UNROLL
))
810 /* We were not asked to, just return back silently. */
816 "\n;; Considering unrolling loop with runtime "
817 "computable number of iterations\n");
819 /* nunroll = total number of copies of the original loop body in
820 unrolled loop (i.e. if it is 2, we have to duplicate loop body once. */
821 nunroll
= PARAM_VALUE (PARAM_MAX_UNROLLED_INSNS
) / loop
->ninsns
;
822 nunroll_by_av
= PARAM_VALUE (PARAM_MAX_AVERAGE_UNROLLED_INSNS
) / loop
->av_ninsns
;
823 if (nunroll
> nunroll_by_av
)
824 nunroll
= nunroll_by_av
;
825 if (nunroll
> (unsigned) PARAM_VALUE (PARAM_MAX_UNROLL_TIMES
))
826 nunroll
= PARAM_VALUE (PARAM_MAX_UNROLL_TIMES
);
828 if (targetm
.loop_unroll_adjust
)
829 nunroll
= targetm
.loop_unroll_adjust (nunroll
, loop
);
831 /* Skip big loops. */
835 fprintf (dump_file
, ";; Not considering loop, is too big\n");
839 /* Check for simple loops. */
840 desc
= get_simple_loop_desc (loop
);
842 /* Check simpleness. */
843 if (!desc
->simple_p
|| desc
->assumptions
)
847 ";; Unable to prove that the number of iterations "
848 "can be counted in runtime\n");
852 if (desc
->const_iter
)
855 fprintf (dump_file
, ";; Loop iterates constant times\n");
859 /* If we have profile feedback, check whether the loop rolls. */
860 if (loop
->header
->count
&& expected_loop_iterations (loop
) < 2 * nunroll
)
863 fprintf (dump_file
, ";; Not unrolling loop, doesn't roll\n");
867 /* Success; now force nunroll to be power of 2, as we are unable to
868 cope with overflows in computation of number of iterations. */
869 for (i
= 1; 2 * i
<= nunroll
; i
*= 2)
872 loop
->lpt_decision
.decision
= LPT_UNROLL_RUNTIME
;
873 loop
->lpt_decision
.times
= i
- 1;
877 ";; Decided to unroll the runtime computable "
878 "times rolling loop, %d times.\n",
879 loop
->lpt_decision
.times
);
882 /* Splits edge E and inserts the sequence of instructions INSNS on it, and
883 returns the newly created block. If INSNS is NULL_RTX, nothing is changed
884 and NULL is returned instead. */
887 split_edge_and_insert (edge e
, rtx insns
)
894 emit_insn_after (insns
, BB_END (bb
));
896 /* ??? We used to assume that INSNS can contain control flow insns, and
897 that we had to try to find sub basic blocks in BB to maintain a valid
898 CFG. For this purpose we used to set the BB_SUPERBLOCK flag on BB
899 and call break_superblocks when going out of cfglayout mode. But it
900 turns out that this never happens; and that if it does ever happen,
901 the TODO_verify_flow at the end of the RTL loop passes would fail.
903 There are two reasons why we expected we could have control flow insns
904 in INSNS. The first is when a comparison has to be done in parts, and
905 the second is when the number of iterations is computed for loops with
906 the number of iterations known at runtime. In both cases, test cases
907 to get control flow in INSNS appear to be impossible to construct:
909 * If do_compare_rtx_and_jump needs several branches to do comparison
910 in a mode that needs comparison by parts, we cannot analyze the
911 number of iterations of the loop, and we never get to unrolling it.
913 * The code in expand_divmod that was suspected to cause creation of
914 branching code seems to be only accessed for signed division. The
915 divisions used by # of iterations analysis are always unsigned.
916 Problems might arise on architectures that emits branching code
917 for some operations that may appear in the unroller (especially
918 for division), but we have no such architectures.
920 Considering all this, it was decided that we should for now assume
921 that INSNS can in theory contain control flow insns, but in practice
922 it never does. So we don't handle the theoretical case, and should
923 a real failure ever show up, we have a pretty good clue for how to
929 /* Unroll LOOP for that we are able to count number of iterations in runtime
930 LOOP->LPT_DECISION.TIMES + 1 times. The transformation does this (with some
931 extra care for case n < 0):
933 for (i = 0; i < n; i++)
961 unroll_loop_runtime_iterations (struct loop
*loop
)
963 rtx old_niter
, niter
, init_code
, branch_code
, tmp
;
965 basic_block preheader
, *body
, swtch
, ezc_swtch
;
966 VEC (basic_block
, heap
) *dom_bbs
;
970 VEC (edge
, heap
) *remove_edges
;
972 bool extra_zero_check
, last_may_exit
;
973 unsigned max_unroll
= loop
->lpt_decision
.times
;
974 struct niter_desc
*desc
= get_simple_loop_desc (loop
);
975 bool exit_at_end
= loop_exit_at_end_p (loop
);
976 struct opt_info
*opt_info
= NULL
;
979 if (flag_split_ivs_in_unroller
980 || flag_variable_expansion_in_unroller
)
981 opt_info
= analyze_insns_in_loop (loop
);
983 /* Remember blocks whose dominators will have to be updated. */
986 body
= get_loop_body (loop
);
987 for (i
= 0; i
< loop
->num_nodes
; i
++)
989 VEC (basic_block
, heap
) *ldom
;
992 ldom
= get_dominated_by (CDI_DOMINATORS
, body
[i
]);
993 FOR_EACH_VEC_ELT (basic_block
, ldom
, j
, bb
)
994 if (!flow_bb_inside_loop_p (loop
, bb
))
995 VEC_safe_push (basic_block
, heap
, dom_bbs
, bb
);
997 VEC_free (basic_block
, heap
, ldom
);
1003 /* Leave exit in first copy (for explanation why see comment in
1004 unroll_loop_constant_iterations). */
1006 n_peel
= max_unroll
- 1;
1007 extra_zero_check
= true;
1008 last_may_exit
= false;
1012 /* Leave exit in last copy (for explanation why see comment in
1013 unroll_loop_constant_iterations). */
1014 may_exit_copy
= max_unroll
;
1015 n_peel
= max_unroll
;
1016 extra_zero_check
= false;
1017 last_may_exit
= true;
1020 /* Get expression for number of iterations. */
1022 old_niter
= niter
= gen_reg_rtx (desc
->mode
);
1023 tmp
= force_operand (copy_rtx (desc
->niter_expr
), niter
);
1025 emit_move_insn (niter
, tmp
);
1027 /* Count modulo by ANDing it with max_unroll; we use the fact that
1028 the number of unrollings is a power of two, and thus this is correct
1029 even if there is overflow in the computation. */
1030 niter
= expand_simple_binop (desc
->mode
, AND
,
1032 GEN_INT (max_unroll
),
1033 NULL_RTX
, 0, OPTAB_LIB_WIDEN
);
1035 init_code
= get_insns ();
1037 unshare_all_rtl_in_chain (init_code
);
1039 /* Precondition the loop. */
1040 split_edge_and_insert (loop_preheader_edge (loop
), init_code
);
1042 remove_edges
= NULL
;
1044 wont_exit
= sbitmap_alloc (max_unroll
+ 2);
1046 /* Peel the first copy of loop body (almost always we must leave exit test
1047 here; the only exception is when we have extra zero check and the number
1048 of iterations is reliable. Also record the place of (possible) extra
1050 sbitmap_zero (wont_exit
);
1051 if (extra_zero_check
1052 && !desc
->noloop_assumptions
)
1053 SET_BIT (wont_exit
, 1);
1054 ezc_swtch
= loop_preheader_edge (loop
)->src
;
1055 ok
= duplicate_loop_to_header_edge (loop
, loop_preheader_edge (loop
),
1056 1, wont_exit
, desc
->out_edge
,
1058 DLTHE_FLAG_UPDATE_FREQ
);
1061 /* Record the place where switch will be built for preconditioning. */
1062 swtch
= split_edge (loop_preheader_edge (loop
));
1064 for (i
= 0; i
< n_peel
; i
++)
1066 /* Peel the copy. */
1067 sbitmap_zero (wont_exit
);
1068 if (i
!= n_peel
- 1 || !last_may_exit
)
1069 SET_BIT (wont_exit
, 1);
1070 ok
= duplicate_loop_to_header_edge (loop
, loop_preheader_edge (loop
),
1071 1, wont_exit
, desc
->out_edge
,
1073 DLTHE_FLAG_UPDATE_FREQ
);
1076 /* Create item for switch. */
1077 j
= n_peel
- i
- (extra_zero_check
? 0 : 1);
1078 p
= REG_BR_PROB_BASE
/ (i
+ 2);
1080 preheader
= split_edge (loop_preheader_edge (loop
));
1081 branch_code
= compare_and_jump_seq (copy_rtx (niter
), GEN_INT (j
), EQ
,
1082 block_label (preheader
), p
,
1085 /* We rely on the fact that the compare and jump cannot be optimized out,
1086 and hence the cfg we create is correct. */
1087 gcc_assert (branch_code
!= NULL_RTX
);
1089 swtch
= split_edge_and_insert (single_pred_edge (swtch
), branch_code
);
1090 set_immediate_dominator (CDI_DOMINATORS
, preheader
, swtch
);
1091 single_pred_edge (swtch
)->probability
= REG_BR_PROB_BASE
- p
;
1092 e
= make_edge (swtch
, preheader
,
1093 single_succ_edge (swtch
)->flags
& EDGE_IRREDUCIBLE_LOOP
);
1097 if (extra_zero_check
)
1099 /* Add branch for zero iterations. */
1100 p
= REG_BR_PROB_BASE
/ (max_unroll
+ 1);
1102 preheader
= split_edge (loop_preheader_edge (loop
));
1103 branch_code
= compare_and_jump_seq (copy_rtx (niter
), const0_rtx
, EQ
,
1104 block_label (preheader
), p
,
1106 gcc_assert (branch_code
!= NULL_RTX
);
1108 swtch
= split_edge_and_insert (single_succ_edge (swtch
), branch_code
);
1109 set_immediate_dominator (CDI_DOMINATORS
, preheader
, swtch
);
1110 single_succ_edge (swtch
)->probability
= REG_BR_PROB_BASE
- p
;
1111 e
= make_edge (swtch
, preheader
,
1112 single_succ_edge (swtch
)->flags
& EDGE_IRREDUCIBLE_LOOP
);
1116 /* Recount dominators for outer blocks. */
1117 iterate_fix_dominators (CDI_DOMINATORS
, dom_bbs
, false);
1119 /* And unroll loop. */
1121 sbitmap_ones (wont_exit
);
1122 RESET_BIT (wont_exit
, may_exit_copy
);
1123 opt_info_start_duplication (opt_info
);
1125 ok
= duplicate_loop_to_header_edge (loop
, loop_latch_edge (loop
),
1127 wont_exit
, desc
->out_edge
,
1129 DLTHE_FLAG_UPDATE_FREQ
1131 ? DLTHE_RECORD_COPY_NUMBER
1137 apply_opt_in_copies (opt_info
, max_unroll
, true, true);
1138 free_opt_info (opt_info
);
1145 basic_block exit_block
= get_bb_copy (desc
->in_edge
->src
);
1146 /* Find a new in and out edge; they are in the last copy we have
1149 if (EDGE_SUCC (exit_block
, 0)->dest
== desc
->out_edge
->dest
)
1151 desc
->out_edge
= EDGE_SUCC (exit_block
, 0);
1152 desc
->in_edge
= EDGE_SUCC (exit_block
, 1);
1156 desc
->out_edge
= EDGE_SUCC (exit_block
, 1);
1157 desc
->in_edge
= EDGE_SUCC (exit_block
, 0);
1161 /* Remove the edges. */
1162 FOR_EACH_VEC_ELT (edge
, remove_edges
, i
, e
)
1164 VEC_free (edge
, heap
, remove_edges
);
1166 /* We must be careful when updating the number of iterations due to
1167 preconditioning and the fact that the value must be valid at entry
1168 of the loop. After passing through the above code, we see that
1169 the correct new number of iterations is this: */
1170 gcc_assert (!desc
->const_iter
);
1172 simplify_gen_binary (UDIV
, desc
->mode
, old_niter
,
1173 GEN_INT (max_unroll
+ 1));
1174 desc
->niter_max
/= max_unroll
+ 1;
1178 simplify_gen_binary (MINUS
, desc
->mode
, desc
->niter_expr
, const1_rtx
);
1179 desc
->noloop_assumptions
= NULL_RTX
;
1185 ";; Unrolled loop %d times, counting # of iterations "
1186 "in runtime, %i insns\n",
1187 max_unroll
, num_loop_insns (loop
));
1189 VEC_free (basic_block
, heap
, dom_bbs
);
1192 /* Decide whether to simply peel LOOP and how much. */
1194 decide_peel_simple (struct loop
*loop
, int flags
)
1197 struct niter_desc
*desc
;
1199 if (!(flags
& UAP_PEEL
))
1201 /* We were not asked to, just return back silently. */
1206 fprintf (dump_file
, "\n;; Considering simply peeling loop\n");
1208 /* npeel = number of iterations to peel. */
1209 npeel
= PARAM_VALUE (PARAM_MAX_PEELED_INSNS
) / loop
->ninsns
;
1210 if (npeel
> (unsigned) PARAM_VALUE (PARAM_MAX_PEEL_TIMES
))
1211 npeel
= PARAM_VALUE (PARAM_MAX_PEEL_TIMES
);
1213 /* Skip big loops. */
1217 fprintf (dump_file
, ";; Not considering loop, is too big\n");
1221 /* Check for simple loops. */
1222 desc
= get_simple_loop_desc (loop
);
1224 /* Check number of iterations. */
1225 if (desc
->simple_p
&& !desc
->assumptions
&& desc
->const_iter
)
1228 fprintf (dump_file
, ";; Loop iterates constant times\n");
1232 /* Do not simply peel loops with branches inside -- it increases number
1234 if (num_loop_branches (loop
) > 1)
1237 fprintf (dump_file
, ";; Not peeling, contains branches\n");
1241 if (loop
->header
->count
)
1243 unsigned niter
= expected_loop_iterations (loop
);
1244 if (niter
+ 1 > npeel
)
1248 fprintf (dump_file
, ";; Not peeling loop, rolls too much (");
1249 fprintf (dump_file
, HOST_WIDEST_INT_PRINT_DEC
,
1250 (HOST_WIDEST_INT
) (niter
+ 1));
1251 fprintf (dump_file
, " iterations > %d [maximum peelings])\n",
1260 /* For now we have no good heuristics to decide whether loop peeling
1261 will be effective, so disable it. */
1264 ";; Not peeling loop, no evidence it will be profitable\n");
1269 loop
->lpt_decision
.decision
= LPT_PEEL_SIMPLE
;
1270 loop
->lpt_decision
.times
= npeel
;
1273 fprintf (dump_file
, ";; Decided to simply peel the loop, %d times.\n",
1274 loop
->lpt_decision
.times
);
1277 /* Peel a LOOP LOOP->LPT_DECISION.TIMES times. The transformation:
1283 if (!cond) goto end;
1285 if (!cond) goto end;
1292 peel_loop_simple (struct loop
*loop
)
1295 unsigned npeel
= loop
->lpt_decision
.times
;
1296 struct niter_desc
*desc
= get_simple_loop_desc (loop
);
1297 struct opt_info
*opt_info
= NULL
;
1300 if (flag_split_ivs_in_unroller
&& npeel
> 1)
1301 opt_info
= analyze_insns_in_loop (loop
);
1303 wont_exit
= sbitmap_alloc (npeel
+ 1);
1304 sbitmap_zero (wont_exit
);
1306 opt_info_start_duplication (opt_info
);
1308 ok
= duplicate_loop_to_header_edge (loop
, loop_preheader_edge (loop
),
1309 npeel
, wont_exit
, NULL
,
1310 NULL
, DLTHE_FLAG_UPDATE_FREQ
1312 ? DLTHE_RECORD_COPY_NUMBER
1320 apply_opt_in_copies (opt_info
, npeel
, false, false);
1321 free_opt_info (opt_info
);
1326 if (desc
->const_iter
)
1328 desc
->niter
-= npeel
;
1329 desc
->niter_expr
= GEN_INT (desc
->niter
);
1330 desc
->noloop_assumptions
= NULL_RTX
;
1334 /* We cannot just update niter_expr, as its value might be clobbered
1335 inside loop. We could handle this by counting the number into
1336 temporary just like we do in runtime unrolling, but it does not
1338 free_simple_loop_desc (loop
);
1342 fprintf (dump_file
, ";; Peeling loop %d times\n", npeel
);
1345 /* Decide whether to unroll LOOP stupidly and how much. */
1347 decide_unroll_stupid (struct loop
*loop
, int flags
)
1349 unsigned nunroll
, nunroll_by_av
, i
;
1350 struct niter_desc
*desc
;
1352 if (!(flags
& UAP_UNROLL_ALL
))
1354 /* We were not asked to, just return back silently. */
1359 fprintf (dump_file
, "\n;; Considering unrolling loop stupidly\n");
1361 /* nunroll = total number of copies of the original loop body in
1362 unrolled loop (i.e. if it is 2, we have to duplicate loop body once. */
1363 nunroll
= PARAM_VALUE (PARAM_MAX_UNROLLED_INSNS
) / loop
->ninsns
;
1365 = PARAM_VALUE (PARAM_MAX_AVERAGE_UNROLLED_INSNS
) / loop
->av_ninsns
;
1366 if (nunroll
> nunroll_by_av
)
1367 nunroll
= nunroll_by_av
;
1368 if (nunroll
> (unsigned) PARAM_VALUE (PARAM_MAX_UNROLL_TIMES
))
1369 nunroll
= PARAM_VALUE (PARAM_MAX_UNROLL_TIMES
);
1371 if (targetm
.loop_unroll_adjust
)
1372 nunroll
= targetm
.loop_unroll_adjust (nunroll
, loop
);
1374 /* Skip big loops. */
1378 fprintf (dump_file
, ";; Not considering loop, is too big\n");
1382 /* Check for simple loops. */
1383 desc
= get_simple_loop_desc (loop
);
1385 /* Check simpleness. */
1386 if (desc
->simple_p
&& !desc
->assumptions
)
1389 fprintf (dump_file
, ";; The loop is simple\n");
1393 /* Do not unroll loops with branches inside -- it increases number
1395 if (num_loop_branches (loop
) > 1)
1398 fprintf (dump_file
, ";; Not unrolling, contains branches\n");
1402 /* If we have profile feedback, check whether the loop rolls. */
1403 if (loop
->header
->count
1404 && expected_loop_iterations (loop
) < 2 * nunroll
)
1407 fprintf (dump_file
, ";; Not unrolling loop, doesn't roll\n");
1411 /* Success. Now force nunroll to be power of 2, as it seems that this
1412 improves results (partially because of better alignments, partially
1413 because of some dark magic). */
1414 for (i
= 1; 2 * i
<= nunroll
; i
*= 2)
1417 loop
->lpt_decision
.decision
= LPT_UNROLL_STUPID
;
1418 loop
->lpt_decision
.times
= i
- 1;
1422 ";; Decided to unroll the loop stupidly, %d times.\n",
1423 loop
->lpt_decision
.times
);
1426 /* Unroll a LOOP LOOP->LPT_DECISION.TIMES times. The transformation:
1444 unroll_loop_stupid (struct loop
*loop
)
1447 unsigned nunroll
= loop
->lpt_decision
.times
;
1448 struct niter_desc
*desc
= get_simple_loop_desc (loop
);
1449 struct opt_info
*opt_info
= NULL
;
1452 if (flag_split_ivs_in_unroller
1453 || flag_variable_expansion_in_unroller
)
1454 opt_info
= analyze_insns_in_loop (loop
);
1457 wont_exit
= sbitmap_alloc (nunroll
+ 1);
1458 sbitmap_zero (wont_exit
);
1459 opt_info_start_duplication (opt_info
);
1461 ok
= duplicate_loop_to_header_edge (loop
, loop_latch_edge (loop
),
1464 DLTHE_FLAG_UPDATE_FREQ
1466 ? DLTHE_RECORD_COPY_NUMBER
1472 apply_opt_in_copies (opt_info
, nunroll
, true, true);
1473 free_opt_info (opt_info
);
1480 /* We indeed may get here provided that there are nontrivial assumptions
1481 for a loop to be really simple. We could update the counts, but the
1482 problem is that we are unable to decide which exit will be taken
1483 (not really true in case the number of iterations is constant,
1484 but noone will do anything with this information, so we do not
1486 desc
->simple_p
= false;
1490 fprintf (dump_file
, ";; Unrolled loop %d times, %i insns\n",
1491 nunroll
, num_loop_insns (loop
));
1494 /* A hash function for information about insns to split. */
1497 si_info_hash (const void *ivts
)
1499 return (hashval_t
) INSN_UID (((const struct iv_to_split
*) ivts
)->insn
);
1502 /* An equality functions for information about insns to split. */
1505 si_info_eq (const void *ivts1
, const void *ivts2
)
1507 const struct iv_to_split
*const i1
= (const struct iv_to_split
*) ivts1
;
1508 const struct iv_to_split
*const i2
= (const struct iv_to_split
*) ivts2
;
1510 return i1
->insn
== i2
->insn
;
1513 /* Return a hash for VES, which is really a "var_to_expand *". */
1516 ve_info_hash (const void *ves
)
1518 return (hashval_t
) INSN_UID (((const struct var_to_expand
*) ves
)->insn
);
1521 /* Return true if IVTS1 and IVTS2 (which are really both of type
1522 "var_to_expand *") refer to the same instruction. */
1525 ve_info_eq (const void *ivts1
, const void *ivts2
)
1527 const struct var_to_expand
*const i1
= (const struct var_to_expand
*) ivts1
;
1528 const struct var_to_expand
*const i2
= (const struct var_to_expand
*) ivts2
;
1530 return i1
->insn
== i2
->insn
;
1533 /* Returns true if REG is referenced in one nondebug insn in LOOP.
1534 Set *DEBUG_USES to the number of debug insns that reference the
1538 referenced_in_one_insn_in_loop_p (struct loop
*loop
, rtx reg
,
1541 basic_block
*body
, bb
;
1546 body
= get_loop_body (loop
);
1547 for (i
= 0; i
< loop
->num_nodes
; i
++)
1551 FOR_BB_INSNS (bb
, insn
)
1552 if (!rtx_referenced_p (reg
, insn
))
1554 else if (DEBUG_INSN_P (insn
))
1556 else if (++count_ref
> 1)
1560 return (count_ref
== 1);
1563 /* Reset the DEBUG_USES debug insns in LOOP that reference REG. */
1566 reset_debug_uses_in_loop (struct loop
*loop
, rtx reg
, int debug_uses
)
1568 basic_block
*body
, bb
;
1572 body
= get_loop_body (loop
);
1573 for (i
= 0; debug_uses
&& i
< loop
->num_nodes
; i
++)
1577 FOR_BB_INSNS (bb
, insn
)
1578 if (!DEBUG_INSN_P (insn
) || !rtx_referenced_p (reg
, insn
))
1582 validate_change (insn
, &INSN_VAR_LOCATION_LOC (insn
),
1583 gen_rtx_UNKNOWN_VAR_LOC (), 0);
1591 /* Determine whether INSN contains an accumulator
1592 which can be expanded into separate copies,
1593 one for each copy of the LOOP body.
1595 for (i = 0 ; i < n; i++)
1609 Return NULL if INSN contains no opportunity for expansion of accumulator.
1610 Otherwise, allocate a VAR_TO_EXPAND structure, fill it with the relevant
1611 information and return a pointer to it.
1614 static struct var_to_expand
*
1615 analyze_insn_to_expand_var (struct loop
*loop
, rtx insn
)
1618 struct var_to_expand
*ves
;
1623 set
= single_set (insn
);
1627 dest
= SET_DEST (set
);
1628 src
= SET_SRC (set
);
1629 code
= GET_CODE (src
);
1631 if (code
!= PLUS
&& code
!= MINUS
&& code
!= MULT
&& code
!= FMA
)
1634 if (FLOAT_MODE_P (GET_MODE (dest
)))
1636 if (!flag_associative_math
)
1638 /* In the case of FMA, we're also changing the rounding. */
1639 if (code
== FMA
&& !flag_unsafe_math_optimizations
)
1643 /* Hmm, this is a bit paradoxical. We know that INSN is a valid insn
1644 in MD. But if there is no optab to generate the insn, we can not
1645 perform the variable expansion. This can happen if an MD provides
1646 an insn but not a named pattern to generate it, for example to avoid
1647 producing code that needs additional mode switches like for x87/mmx.
1649 So we check have_insn_for which looks for an optab for the operation
1650 in SRC. If it doesn't exist, we can't perform the expansion even
1651 though INSN is valid. */
1652 if (!have_insn_for (code
, GET_MODE (src
)))
1656 && !(GET_CODE (dest
) == SUBREG
1657 && REG_P (SUBREG_REG (dest
))))
1660 /* Find the accumulator use within the operation. */
1663 /* We only support accumulation via FMA in the ADD position. */
1664 if (!rtx_equal_p (dest
, XEXP (src
, 2)))
1668 else if (rtx_equal_p (dest
, XEXP (src
, 0)))
1670 else if (rtx_equal_p (dest
, XEXP (src
, 1)))
1672 /* The method of expansion that we are using; which includes the
1673 initialization of the expansions with zero and the summation of
1674 the expansions at the end of the computation will yield wrong
1675 results for (x = something - x) thus avoid using it in that case. */
1683 /* It must not otherwise be used. */
1686 if (rtx_referenced_p (dest
, XEXP (src
, 0))
1687 || rtx_referenced_p (dest
, XEXP (src
, 1)))
1690 else if (rtx_referenced_p (dest
, XEXP (src
, 1 - accum_pos
)))
1693 /* It must be used in exactly one insn. */
1694 if (!referenced_in_one_insn_in_loop_p (loop
, dest
, &debug_uses
))
1699 fprintf (dump_file
, "\n;; Expanding Accumulator ");
1700 print_rtl (dump_file
, dest
);
1701 fprintf (dump_file
, "\n");
1705 /* Instead of resetting the debug insns, we could replace each
1706 debug use in the loop with the sum or product of all expanded
1707 accummulators. Since we'll only know of all expansions at the
1708 end, we'd have to keep track of which vars_to_expand a debug
1709 insn in the loop references, take note of each copy of the
1710 debug insn during unrolling, and when it's all done, compute
1711 the sum or product of each variable and adjust the original
1712 debug insn and each copy thereof. What a pain! */
1713 reset_debug_uses_in_loop (loop
, dest
, debug_uses
);
1715 /* Record the accumulator to expand. */
1716 ves
= XNEW (struct var_to_expand
);
1718 ves
->reg
= copy_rtx (dest
);
1719 ves
->var_expansions
= VEC_alloc (rtx
, heap
, 1);
1721 ves
->op
= GET_CODE (src
);
1722 ves
->expansion_count
= 0;
1723 ves
->reuse_expansion
= 0;
1724 ves
->accum_pos
= accum_pos
;
1728 /* Determine whether there is an induction variable in INSN that
1729 we would like to split during unrolling.
1749 Return NULL if INSN contains no interesting IVs. Otherwise, allocate
1750 an IV_TO_SPLIT structure, fill it with the relevant information and return a
1753 static struct iv_to_split
*
1754 analyze_iv_to_split_insn (rtx insn
)
1758 struct iv_to_split
*ivts
;
1761 /* For now we just split the basic induction variables. Later this may be
1762 extended for example by selecting also addresses of memory references. */
1763 set
= single_set (insn
);
1767 dest
= SET_DEST (set
);
1771 if (!biv_p (insn
, dest
))
1774 ok
= iv_analyze_result (insn
, dest
, &iv
);
1776 /* This used to be an assert under the assumption that if biv_p returns
1777 true that iv_analyze_result must also return true. However, that
1778 assumption is not strictly correct as evidenced by pr25569.
1780 Returning NULL when iv_analyze_result returns false is safe and
1781 avoids the problems in pr25569 until the iv_analyze_* routines
1782 can be fixed, which is apparently hard and time consuming
1783 according to their author. */
1787 if (iv
.step
== const0_rtx
1788 || iv
.mode
!= iv
.extend_mode
)
1791 /* Record the insn to split. */
1792 ivts
= XNEW (struct iv_to_split
);
1794 ivts
->base_var
= NULL_RTX
;
1795 ivts
->step
= iv
.step
;
1803 /* Determines which of insns in LOOP can be optimized.
1804 Return a OPT_INFO struct with the relevant hash tables filled
1805 with all insns to be optimized. The FIRST_NEW_BLOCK field
1806 is undefined for the return value. */
1808 static struct opt_info
*
1809 analyze_insns_in_loop (struct loop
*loop
)
1811 basic_block
*body
, bb
;
1813 struct opt_info
*opt_info
= XCNEW (struct opt_info
);
1815 struct iv_to_split
*ivts
= NULL
;
1816 struct var_to_expand
*ves
= NULL
;
1819 VEC (edge
, heap
) *edges
= get_loop_exit_edges (loop
);
1821 bool can_apply
= false;
1823 iv_analysis_loop_init (loop
);
1825 body
= get_loop_body (loop
);
1827 if (flag_split_ivs_in_unroller
)
1829 opt_info
->insns_to_split
= htab_create (5 * loop
->num_nodes
,
1830 si_info_hash
, si_info_eq
, free
);
1831 opt_info
->iv_to_split_head
= NULL
;
1832 opt_info
->iv_to_split_tail
= &opt_info
->iv_to_split_head
;
1835 /* Record the loop exit bb and loop preheader before the unrolling. */
1836 opt_info
->loop_preheader
= loop_preheader_edge (loop
)->src
;
1838 if (VEC_length (edge
, edges
) == 1)
1840 exit
= VEC_index (edge
, edges
, 0);
1841 if (!(exit
->flags
& EDGE_COMPLEX
))
1843 opt_info
->loop_exit
= split_edge (exit
);
1848 if (flag_variable_expansion_in_unroller
1851 opt_info
->insns_with_var_to_expand
= htab_create (5 * loop
->num_nodes
,
1854 opt_info
->var_to_expand_head
= NULL
;
1855 opt_info
->var_to_expand_tail
= &opt_info
->var_to_expand_head
;
1858 for (i
= 0; i
< loop
->num_nodes
; i
++)
1861 if (!dominated_by_p (CDI_DOMINATORS
, loop
->latch
, bb
))
1864 FOR_BB_INSNS (bb
, insn
)
1869 if (opt_info
->insns_to_split
)
1870 ivts
= analyze_iv_to_split_insn (insn
);
1874 slot1
= htab_find_slot (opt_info
->insns_to_split
, ivts
, INSERT
);
1875 gcc_assert (*slot1
== NULL
);
1877 *opt_info
->iv_to_split_tail
= ivts
;
1878 opt_info
->iv_to_split_tail
= &ivts
->next
;
1882 if (opt_info
->insns_with_var_to_expand
)
1883 ves
= analyze_insn_to_expand_var (loop
, insn
);
1887 slot2
= htab_find_slot (opt_info
->insns_with_var_to_expand
, ves
, INSERT
);
1888 gcc_assert (*slot2
== NULL
);
1890 *opt_info
->var_to_expand_tail
= ves
;
1891 opt_info
->var_to_expand_tail
= &ves
->next
;
1896 VEC_free (edge
, heap
, edges
);
1901 /* Called just before loop duplication. Records start of duplicated area
1905 opt_info_start_duplication (struct opt_info
*opt_info
)
1908 opt_info
->first_new_block
= last_basic_block
;
1911 /* Determine the number of iterations between initialization of the base
1912 variable and the current copy (N_COPY). N_COPIES is the total number
1913 of newly created copies. UNROLLING is true if we are unrolling
1914 (not peeling) the loop. */
1917 determine_split_iv_delta (unsigned n_copy
, unsigned n_copies
, bool unrolling
)
1921 /* If we are unrolling, initialization is done in the original loop
1927 /* If we are peeling, the copy in that the initialization occurs has
1928 number 1. The original loop (number 0) is the last. */
1936 /* Locate in EXPR the expression corresponding to the location recorded
1937 in IVTS, and return a pointer to the RTX for this location. */
1940 get_ivts_expr (rtx expr
, struct iv_to_split
*ivts
)
1945 for (i
= 0; i
< ivts
->n_loc
; i
++)
1946 ret
= &XEXP (*ret
, ivts
->loc
[i
]);
1951 /* Allocate basic variable for the induction variable chain. */
1954 allocate_basic_variable (struct iv_to_split
*ivts
)
1956 rtx expr
= *get_ivts_expr (single_set (ivts
->insn
), ivts
);
1958 ivts
->base_var
= gen_reg_rtx (GET_MODE (expr
));
1961 /* Insert initialization of basic variable of IVTS before INSN, taking
1962 the initial value from INSN. */
1965 insert_base_initialization (struct iv_to_split
*ivts
, rtx insn
)
1967 rtx expr
= copy_rtx (*get_ivts_expr (single_set (insn
), ivts
));
1971 expr
= force_operand (expr
, ivts
->base_var
);
1972 if (expr
!= ivts
->base_var
)
1973 emit_move_insn (ivts
->base_var
, expr
);
1977 emit_insn_before (seq
, insn
);
1980 /* Replace the use of induction variable described in IVTS in INSN
1981 by base variable + DELTA * step. */
1984 split_iv (struct iv_to_split
*ivts
, rtx insn
, unsigned delta
)
1986 rtx expr
, *loc
, seq
, incr
, var
;
1987 enum machine_mode mode
= GET_MODE (ivts
->base_var
);
1990 /* Construct base + DELTA * step. */
1992 expr
= ivts
->base_var
;
1995 incr
= simplify_gen_binary (MULT
, mode
,
1996 ivts
->step
, gen_int_mode (delta
, mode
));
1997 expr
= simplify_gen_binary (PLUS
, GET_MODE (ivts
->base_var
),
1998 ivts
->base_var
, incr
);
2001 /* Figure out where to do the replacement. */
2002 loc
= get_ivts_expr (single_set (insn
), ivts
);
2004 /* If we can make the replacement right away, we're done. */
2005 if (validate_change (insn
, loc
, expr
, 0))
2008 /* Otherwise, force EXPR into a register and try again. */
2010 var
= gen_reg_rtx (mode
);
2011 expr
= force_operand (expr
, var
);
2013 emit_move_insn (var
, expr
);
2016 emit_insn_before (seq
, insn
);
2018 if (validate_change (insn
, loc
, var
, 0))
2021 /* The last chance. Try recreating the assignment in insn
2022 completely from scratch. */
2023 set
= single_set (insn
);
2028 src
= copy_rtx (SET_SRC (set
));
2029 dest
= copy_rtx (SET_DEST (set
));
2030 src
= force_operand (src
, dest
);
2032 emit_move_insn (dest
, src
);
2036 emit_insn_before (seq
, insn
);
2041 /* Return one expansion of the accumulator recorded in struct VE. */
2044 get_expansion (struct var_to_expand
*ve
)
2048 if (ve
->reuse_expansion
== 0)
2051 reg
= VEC_index (rtx
, ve
->var_expansions
, ve
->reuse_expansion
- 1);
2053 if (VEC_length (rtx
, ve
->var_expansions
) == (unsigned) ve
->reuse_expansion
)
2054 ve
->reuse_expansion
= 0;
2056 ve
->reuse_expansion
++;
2062 /* Given INSN replace the uses of the accumulator recorded in VE
2063 with a new register. */
2066 expand_var_during_unrolling (struct var_to_expand
*ve
, rtx insn
)
2069 bool really_new_expansion
= false;
2071 set
= single_set (insn
);
2074 /* Generate a new register only if the expansion limit has not been
2075 reached. Else reuse an already existing expansion. */
2076 if (PARAM_VALUE (PARAM_MAX_VARIABLE_EXPANSIONS
) > ve
->expansion_count
)
2078 really_new_expansion
= true;
2079 new_reg
= gen_reg_rtx (GET_MODE (ve
->reg
));
2082 new_reg
= get_expansion (ve
);
2084 validate_change (insn
, &SET_DEST (set
), new_reg
, 1);
2085 validate_change (insn
, &XEXP (SET_SRC (set
), ve
->accum_pos
), new_reg
, 1);
2087 if (apply_change_group ())
2088 if (really_new_expansion
)
2090 VEC_safe_push (rtx
, heap
, ve
->var_expansions
, new_reg
);
2091 ve
->expansion_count
++;
2095 /* Initialize the variable expansions in loop preheader. PLACE is the
2096 loop-preheader basic block where the initialization of the
2097 expansions should take place. The expansions are initialized with
2098 (-0) when the operation is plus or minus to honor sign zero. This
2099 way we can prevent cases where the sign of the final result is
2100 effected by the sign of the expansion. Here is an example to
2103 for (i = 0 ; i < n; i++)
2117 When SUM is initialized with -zero and SOMETHING is also -zero; the
2118 final result of sum should be -zero thus the expansions sum1 and sum2
2119 should be initialized with -zero as well (otherwise we will get +zero
2120 as the final result). */
2123 insert_var_expansion_initialization (struct var_to_expand
*ve
,
2126 rtx seq
, var
, zero_init
, insn
;
2128 enum machine_mode mode
= GET_MODE (ve
->reg
);
2129 bool honor_signed_zero_p
= HONOR_SIGNED_ZEROS (mode
);
2131 if (VEC_length (rtx
, ve
->var_expansions
) == 0)
2138 /* Note that we only accumulate FMA via the ADD operand. */
2141 FOR_EACH_VEC_ELT (rtx
, ve
->var_expansions
, i
, var
)
2143 if (honor_signed_zero_p
)
2144 zero_init
= simplify_gen_unary (NEG
, mode
, CONST0_RTX (mode
), mode
);
2146 zero_init
= CONST0_RTX (mode
);
2147 emit_move_insn (var
, zero_init
);
2152 FOR_EACH_VEC_ELT (rtx
, ve
->var_expansions
, i
, var
)
2154 zero_init
= CONST1_RTX (GET_MODE (var
));
2155 emit_move_insn (var
, zero_init
);
2166 insn
= BB_HEAD (place
);
2167 while (!NOTE_INSN_BASIC_BLOCK_P (insn
))
2168 insn
= NEXT_INSN (insn
);
2170 emit_insn_after (seq
, insn
);
2173 /* Combine the variable expansions at the loop exit. PLACE is the
2174 loop exit basic block where the summation of the expansions should
2178 combine_var_copies_in_loop_exit (struct var_to_expand
*ve
, basic_block place
)
2181 rtx expr
, seq
, var
, insn
;
2184 if (VEC_length (rtx
, ve
->var_expansions
) == 0)
2191 /* Note that we only accumulate FMA via the ADD operand. */
2194 FOR_EACH_VEC_ELT (rtx
, ve
->var_expansions
, i
, var
)
2195 sum
= simplify_gen_binary (PLUS
, GET_MODE (ve
->reg
), var
, sum
);
2199 FOR_EACH_VEC_ELT (rtx
, ve
->var_expansions
, i
, var
)
2200 sum
= simplify_gen_binary (MULT
, GET_MODE (ve
->reg
), var
, sum
);
2207 expr
= force_operand (sum
, ve
->reg
);
2208 if (expr
!= ve
->reg
)
2209 emit_move_insn (ve
->reg
, expr
);
2213 insn
= BB_HEAD (place
);
2214 while (!NOTE_INSN_BASIC_BLOCK_P (insn
))
2215 insn
= NEXT_INSN (insn
);
2217 emit_insn_after (seq
, insn
);
2220 /* Apply loop optimizations in loop copies using the
2221 data which gathered during the unrolling. Structure
2222 OPT_INFO record that data.
2224 UNROLLING is true if we unrolled (not peeled) the loop.
2225 REWRITE_ORIGINAL_BODY is true if we should also rewrite the original body of
2226 the loop (as it should happen in complete unrolling, but not in ordinary
2227 peeling of the loop). */
2230 apply_opt_in_copies (struct opt_info
*opt_info
,
2231 unsigned n_copies
, bool unrolling
,
2232 bool rewrite_original_loop
)
2235 basic_block bb
, orig_bb
;
2236 rtx insn
, orig_insn
, next
;
2237 struct iv_to_split ivts_templ
, *ivts
;
2238 struct var_to_expand ve_templ
, *ves
;
2240 /* Sanity check -- we need to put initialization in the original loop
2242 gcc_assert (!unrolling
|| rewrite_original_loop
);
2244 /* Allocate the basic variables (i0). */
2245 if (opt_info
->insns_to_split
)
2246 for (ivts
= opt_info
->iv_to_split_head
; ivts
; ivts
= ivts
->next
)
2247 allocate_basic_variable (ivts
);
2249 for (i
= opt_info
->first_new_block
; i
< (unsigned) last_basic_block
; i
++)
2251 bb
= BASIC_BLOCK (i
);
2252 orig_bb
= get_bb_original (bb
);
2254 /* bb->aux holds position in copy sequence initialized by
2255 duplicate_loop_to_header_edge. */
2256 delta
= determine_split_iv_delta ((size_t)bb
->aux
, n_copies
,
2259 orig_insn
= BB_HEAD (orig_bb
);
2260 for (insn
= BB_HEAD (bb
); insn
!= NEXT_INSN (BB_END (bb
)); insn
= next
)
2262 next
= NEXT_INSN (insn
);
2264 || (DEBUG_INSN_P (insn
)
2265 && TREE_CODE (INSN_VAR_LOCATION_DECL (insn
)) == LABEL_DECL
))
2268 while (!INSN_P (orig_insn
)
2269 || (DEBUG_INSN_P (orig_insn
)
2270 && (TREE_CODE (INSN_VAR_LOCATION_DECL (orig_insn
))
2272 orig_insn
= NEXT_INSN (orig_insn
);
2274 ivts_templ
.insn
= orig_insn
;
2275 ve_templ
.insn
= orig_insn
;
2277 /* Apply splitting iv optimization. */
2278 if (opt_info
->insns_to_split
)
2280 ivts
= (struct iv_to_split
*)
2281 htab_find (opt_info
->insns_to_split
, &ivts_templ
);
2285 gcc_assert (GET_CODE (PATTERN (insn
))
2286 == GET_CODE (PATTERN (orig_insn
)));
2289 insert_base_initialization (ivts
, insn
);
2290 split_iv (ivts
, insn
, delta
);
2293 /* Apply variable expansion optimization. */
2294 if (unrolling
&& opt_info
->insns_with_var_to_expand
)
2296 ves
= (struct var_to_expand
*)
2297 htab_find (opt_info
->insns_with_var_to_expand
, &ve_templ
);
2300 gcc_assert (GET_CODE (PATTERN (insn
))
2301 == GET_CODE (PATTERN (orig_insn
)));
2302 expand_var_during_unrolling (ves
, insn
);
2305 orig_insn
= NEXT_INSN (orig_insn
);
2309 if (!rewrite_original_loop
)
2312 /* Initialize the variable expansions in the loop preheader
2313 and take care of combining them at the loop exit. */
2314 if (opt_info
->insns_with_var_to_expand
)
2316 for (ves
= opt_info
->var_to_expand_head
; ves
; ves
= ves
->next
)
2317 insert_var_expansion_initialization (ves
, opt_info
->loop_preheader
);
2318 for (ves
= opt_info
->var_to_expand_head
; ves
; ves
= ves
->next
)
2319 combine_var_copies_in_loop_exit (ves
, opt_info
->loop_exit
);
2322 /* Rewrite also the original loop body. Find them as originals of the blocks
2323 in the last copied iteration, i.e. those that have
2324 get_bb_copy (get_bb_original (bb)) == bb. */
2325 for (i
= opt_info
->first_new_block
; i
< (unsigned) last_basic_block
; i
++)
2327 bb
= BASIC_BLOCK (i
);
2328 orig_bb
= get_bb_original (bb
);
2329 if (get_bb_copy (orig_bb
) != bb
)
2332 delta
= determine_split_iv_delta (0, n_copies
, unrolling
);
2333 for (orig_insn
= BB_HEAD (orig_bb
);
2334 orig_insn
!= NEXT_INSN (BB_END (bb
));
2337 next
= NEXT_INSN (orig_insn
);
2339 if (!INSN_P (orig_insn
))
2342 ivts_templ
.insn
= orig_insn
;
2343 if (opt_info
->insns_to_split
)
2345 ivts
= (struct iv_to_split
*)
2346 htab_find (opt_info
->insns_to_split
, &ivts_templ
);
2350 insert_base_initialization (ivts
, orig_insn
);
2351 split_iv (ivts
, orig_insn
, delta
);
2360 /* Release OPT_INFO. */
2363 free_opt_info (struct opt_info
*opt_info
)
2365 if (opt_info
->insns_to_split
)
2366 htab_delete (opt_info
->insns_to_split
);
2367 if (opt_info
->insns_with_var_to_expand
)
2369 struct var_to_expand
*ves
;
2371 for (ves
= opt_info
->var_to_expand_head
; ves
; ves
= ves
->next
)
2372 VEC_free (rtx
, heap
, ves
->var_expansions
);
2373 htab_delete (opt_info
->insns_with_var_to_expand
);