1 /* Perform doloop optimizations
2 Copyright (C) 2004-2015 Free Software Foundation, Inc.
3 Based on code by Michael P. Hayes (m.hayes@elec.canterbury.ac.nz)
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"
32 #include "hard-reg-set.h"
35 #include "statistics.h"
36 #include "double-int.h"
38 #include "fixed-value.h"
43 #include "insn-config.h"
52 #include "diagnostic-core.h"
55 #include "dominance.h"
59 #include "basic-block.h"
63 #include "loop-unroll.h"
65 /* This module is used to modify loops with a determinable number of
66 iterations to use special low-overhead looping instructions.
68 It first validates whether the loop is well behaved and has a
69 determinable number of iterations (either at compile or run-time).
70 It then modifies the loop to use a low-overhead looping pattern as
73 1. A pseudo register is allocated as the loop iteration counter.
75 2. The number of loop iterations is calculated and is stored
78 3. At the end of the loop, the jump insn is replaced by the
79 doloop_end pattern. The compare must remain because it might be
80 used elsewhere. If the loop-variable or condition register are
81 used elsewhere, they will be eliminated by flow.
83 4. An optional doloop_begin pattern is inserted at the top of the
86 TODO The optimization should only performed when either the biv used for exit
87 condition is unused at all except for the exit test, or if we do not have to
88 change its value, since otherwise we have to add a new induction variable,
89 which usually will not pay up (unless the cost of the doloop pattern is
90 somehow extremely lower than the cost of compare & jump, or unless the bct
91 register cannot be used for anything else but doloop -- ??? detect these
94 #ifdef HAVE_doloop_end
96 /* Return the loop termination condition for PATTERN or zero
97 if it is not a decrement and branch jump insn. */
100 doloop_condition_get (rtx doloop_pat
)
108 rtx cc_reg
= NULL_RTX
;
109 rtx reg_orig
= NULL_RTX
;
111 /* The canonical doloop pattern we expect has one of the following
114 1) (parallel [(set (pc) (if_then_else (condition)
117 (set (reg) (plus (reg) (const_int -1)))
118 (additional clobbers and uses)])
120 The branch must be the first entry of the parallel (also required
121 by jump.c), and the second entry of the parallel must be a set of
122 the loop counter register. Some targets (IA-64) wrap the set of
123 the loop counter in an if_then_else too.
125 2) (set (reg) (plus (reg) (const_int -1))
126 (set (pc) (if_then_else (reg != 0)
130 Some targets (ARM) do the comparison before the branch, as in the
133 3) (parallel [(set (cc) (compare ((plus (reg) (const_int -1), 0)))
134 (set (reg) (plus (reg) (const_int -1)))])
135 (set (pc) (if_then_else (cc == NE)
139 pattern
= PATTERN (doloop_pat
);
141 if (GET_CODE (pattern
) != PARALLEL
)
144 rtx_insn
*prev_insn
= prev_nondebug_insn (doloop_pat
);
145 rtx cmp_arg1
, cmp_arg2
;
148 /* In case the pattern is not PARALLEL we expect two forms
149 of doloop which are cases 2) and 3) above: in case 2) the
150 decrement immediately precedes the branch, while in case 3)
151 the compare and decrement instructions immediately precede
154 if (prev_insn
== NULL_RTX
|| !INSN_P (prev_insn
))
158 if (GET_CODE (PATTERN (prev_insn
)) == PARALLEL
)
160 /* The third case: the compare and decrement instructions
161 immediately precede the branch. */
162 cmp_orig
= XVECEXP (PATTERN (prev_insn
), 0, 0);
163 if (GET_CODE (cmp_orig
) != SET
)
165 if (GET_CODE (SET_SRC (cmp_orig
)) != COMPARE
)
167 cmp_arg1
= XEXP (SET_SRC (cmp_orig
), 0);
168 cmp_arg2
= XEXP (SET_SRC (cmp_orig
), 1);
169 if (cmp_arg2
!= const0_rtx
170 || GET_CODE (cmp_arg1
) != PLUS
)
172 reg_orig
= XEXP (cmp_arg1
, 0);
173 if (XEXP (cmp_arg1
, 1) != GEN_INT (-1)
174 || !REG_P (reg_orig
))
176 cc_reg
= SET_DEST (cmp_orig
);
178 inc
= XVECEXP (PATTERN (prev_insn
), 0, 1);
181 inc
= PATTERN (prev_insn
);
182 /* We expect the condition to be of the form (reg != 0) */
183 cond
= XEXP (SET_SRC (cmp
), 0);
184 if (GET_CODE (cond
) != NE
|| XEXP (cond
, 1) != const0_rtx
)
189 cmp
= XVECEXP (pattern
, 0, 0);
190 inc
= XVECEXP (pattern
, 0, 1);
193 /* Check for (set (reg) (something)). */
194 if (GET_CODE (inc
) != SET
)
196 reg
= SET_DEST (inc
);
200 /* Check if something = (plus (reg) (const_int -1)).
201 On IA-64, this decrement is wrapped in an if_then_else. */
202 inc_src
= SET_SRC (inc
);
203 if (GET_CODE (inc_src
) == IF_THEN_ELSE
)
204 inc_src
= XEXP (inc_src
, 1);
205 if (GET_CODE (inc_src
) != PLUS
206 || XEXP (inc_src
, 0) != reg
207 || XEXP (inc_src
, 1) != constm1_rtx
)
210 /* Check for (set (pc) (if_then_else (condition)
213 if (GET_CODE (cmp
) != SET
214 || SET_DEST (cmp
) != pc_rtx
215 || GET_CODE (SET_SRC (cmp
)) != IF_THEN_ELSE
216 || GET_CODE (XEXP (SET_SRC (cmp
), 1)) != LABEL_REF
217 || XEXP (SET_SRC (cmp
), 2) != pc_rtx
)
220 /* Extract loop termination condition. */
221 condition
= XEXP (SET_SRC (cmp
), 0);
223 /* We expect a GE or NE comparison with 0 or 1. */
224 if ((GET_CODE (condition
) != GE
225 && GET_CODE (condition
) != NE
)
226 || (XEXP (condition
, 1) != const0_rtx
227 && XEXP (condition
, 1) != const1_rtx
))
230 if ((XEXP (condition
, 0) == reg
)
231 /* For the third case: */
232 || ((cc_reg
!= NULL_RTX
)
233 && (XEXP (condition
, 0) == cc_reg
)
234 && (reg_orig
== reg
))
235 || (GET_CODE (XEXP (condition
, 0)) == PLUS
236 && XEXP (XEXP (condition
, 0), 0) == reg
))
238 if (GET_CODE (pattern
) != PARALLEL
)
239 /* For the second form we expect:
241 (set (reg) (plus (reg) (const_int -1))
242 (set (pc) (if_then_else (reg != 0)
246 is equivalent to the following:
248 (parallel [(set (pc) (if_then_else (reg != 1)
251 (set (reg) (plus (reg) (const_int -1)))
252 (additional clobbers and uses)])
254 For the third form we expect:
256 (parallel [(set (cc) (compare ((plus (reg) (const_int -1)), 0))
257 (set (reg) (plus (reg) (const_int -1)))])
258 (set (pc) (if_then_else (cc == NE)
262 which is equivalent to the following:
264 (parallel [(set (cc) (compare (reg, 1))
265 (set (reg) (plus (reg) (const_int -1)))
266 (set (pc) (if_then_else (NE == cc)
270 So we return the second form instead for the two cases.
273 condition
= gen_rtx_fmt_ee (NE
, VOIDmode
, inc_src
, const1_rtx
);
278 /* ??? If a machine uses a funny comparison, we could return a
279 canonicalized form here. */
284 /* Return nonzero if the loop specified by LOOP is suitable for
285 the use of special low-overhead looping instructions. DESC
286 describes the number of iterations of the loop. */
289 doloop_valid_p (struct loop
*loop
, struct niter_desc
*desc
)
291 basic_block
*body
= get_loop_body (loop
), bb
;
296 /* Check for loops that may not terminate under special conditions. */
301 /* There are some cases that would require a special attention.
302 For example if the comparison is LEU and the comparison value
303 is UINT_MAX then the loop will not terminate. Similarly, if the
304 comparison code is GEU and the comparison value is 0, the
305 loop will not terminate.
307 If the absolute increment is not 1, the loop can be infinite
308 even with LTU/GTU, e.g. for (i = 3; i > 0; i -= 2)
310 ??? We could compute these conditions at run-time and have a
311 additional jump around the loop to ensure an infinite loop.
312 However, it is very unlikely that this is the intended
313 behavior of the loop and checking for these rare boundary
314 conditions would pessimize all other code.
316 If the loop is executed only a few times an extra check to
317 restart the loop could use up most of the benefits of using a
318 count register loop. Note however, that normally, this
319 restart branch would never execute, so it could be predicted
320 well by the CPU. We should generate the pessimistic code by
321 default, and have an option, e.g. -funsafe-loops that would
322 enable count-register loops in this case. */
324 fprintf (dump_file
, "Doloop: Possible infinite iteration case.\n");
329 for (i
= 0; i
< loop
->num_nodes
; i
++)
333 for (insn
= BB_HEAD (bb
);
334 insn
!= NEXT_INSN (BB_END (bb
));
335 insn
= NEXT_INSN (insn
))
337 /* Different targets have different necessities for low-overhead
338 looping. Call the back end for each instruction within the loop
339 to let it decide whether the insn prohibits a low-overhead loop.
340 It will then return the cause for it to emit to the dump file. */
341 const char * invalid
= targetm
.invalid_within_doloop (insn
);
345 fprintf (dump_file
, "Doloop: %s\n", invalid
);
359 /* Adds test of COND jumping to DEST on edge *E and set *E to the new fallthru
360 edge. If the condition is always false, do not do anything. If it is always
361 true, redirect E to DEST and return false. In all other cases, true is
365 add_test (rtx cond
, edge
*e
, basic_block dest
)
367 rtx_insn
*seq
, *jump
;
368 rtx_code_label
*label
;
370 rtx op0
= XEXP (cond
, 0), op1
= XEXP (cond
, 1);
371 enum rtx_code code
= GET_CODE (cond
);
374 mode
= GET_MODE (XEXP (cond
, 0));
375 if (mode
== VOIDmode
)
376 mode
= GET_MODE (XEXP (cond
, 1));
379 op0
= force_operand (op0
, NULL_RTX
);
380 op1
= force_operand (op1
, NULL_RTX
);
381 label
= block_label (dest
);
382 do_compare_rtx_and_jump (op0
, op1
, code
, 0, mode
, NULL_RTX
, NULL
, label
, -1);
384 jump
= get_last_insn ();
385 if (!jump
|| !JUMP_P (jump
))
387 /* The condition is always false and the jump was optimized out. */
395 /* There always is at least the jump insn in the sequence. */
396 gcc_assert (seq
!= NULL_RTX
);
398 bb
= split_edge_and_insert (*e
, seq
);
399 *e
= single_succ_edge (bb
);
401 if (any_uncondjump_p (jump
))
403 /* The condition is always true. */
405 redirect_edge_and_branch_force (*e
, dest
);
409 JUMP_LABEL (jump
) = label
;
411 /* The jump is supposed to handle an unlikely special case. */
412 add_int_reg_note (jump
, REG_BR_PROB
, 0);
414 LABEL_NUSES (label
)++;
416 make_edge (bb
, dest
, (*e
)->flags
& ~EDGE_FALLTHRU
);
420 /* Modify the loop to use the low-overhead looping insn where LOOP
421 describes the loop, DESC describes the number of iterations of the
422 loop, and DOLOOP_INSN is the low-overhead looping insn to emit at the
423 end of the loop. CONDITION is the condition separated from the
424 DOLOOP_SEQ. COUNT is the number of iterations of the LOOP. */
427 doloop_modify (struct loop
*loop
, struct niter_desc
*desc
,
428 rtx doloop_seq
, rtx condition
, rtx count
)
431 rtx tmp
, noloop
= NULL_RTX
;
434 rtx_code_label
*jump_label
;
436 bool increment_count
;
437 basic_block loop_end
= desc
->out_edge
->src
;
440 widest_int iterations
;
442 jump_insn
= BB_END (loop_end
);
446 fprintf (dump_file
, "Doloop: Inserting doloop pattern (");
447 if (desc
->const_iter
)
448 fprintf (dump_file
, "%" PRId64
, desc
->niter
);
450 fputs ("runtime", dump_file
);
451 fputs (" iterations).\n", dump_file
);
454 /* Get the probability of the original branch. If it exists we would
455 need to update REG_BR_PROB of the new jump_insn. */
456 true_prob_val
= find_reg_note (jump_insn
, REG_BR_PROB
, NULL_RTX
);
458 /* Discard original jump to continue loop. The original compare
459 result may still be live, so it cannot be discarded explicitly. */
460 delete_insn (jump_insn
);
462 counter_reg
= XEXP (condition
, 0);
463 if (GET_CODE (counter_reg
) == PLUS
)
464 counter_reg
= XEXP (counter_reg
, 0);
465 mode
= GET_MODE (counter_reg
);
467 increment_count
= false;
468 switch (GET_CODE (condition
))
471 /* Currently only NE tests against zero and one are supported. */
472 noloop
= XEXP (condition
, 1);
473 if (noloop
!= const0_rtx
)
475 gcc_assert (noloop
== const1_rtx
);
476 increment_count
= true;
481 /* Currently only GE tests against zero are supported. */
482 gcc_assert (XEXP (condition
, 1) == const0_rtx
);
484 noloop
= constm1_rtx
;
486 /* The iteration count does not need incrementing for a GE test. */
487 increment_count
= false;
489 /* Determine if the iteration counter will be non-negative.
490 Note that the maximum value loaded is iterations_max - 1. */
491 if (get_max_loop_iterations (loop
, &iterations
)
492 && wi::leu_p (iterations
,
493 wi::set_bit_in_zero
<widest_int
>
494 (GET_MODE_PRECISION (mode
) - 1)))
498 /* Abort if an invalid doloop pattern has been generated. */
504 count
= simplify_gen_binary (PLUS
, mode
, count
, const1_rtx
);
506 /* Insert initialization of the count register into the loop header. */
508 tmp
= force_operand (count
, counter_reg
);
509 convert_move (counter_reg
, tmp
, 1);
510 sequence
= get_insns ();
512 emit_insn_after (sequence
, BB_END (loop_preheader_edge (loop
)->src
));
514 if (desc
->noloop_assumptions
)
516 rtx ass
= copy_rtx (desc
->noloop_assumptions
);
517 basic_block preheader
= loop_preheader_edge (loop
)->src
;
519 = split_edge (loop_preheader_edge (loop
));
520 basic_block new_preheader
521 = split_edge (loop_preheader_edge (loop
));
524 /* Expand the condition testing the assumptions and if it does not pass,
525 reset the count register to 0. */
526 redirect_edge_and_branch_force (single_succ_edge (preheader
), new_preheader
);
527 set_immediate_dominator (CDI_DOMINATORS
, new_preheader
, preheader
);
530 set_zero
->frequency
= 0;
532 te
= single_succ_edge (preheader
);
533 for (; ass
; ass
= XEXP (ass
, 1))
534 if (!add_test (XEXP (ass
, 0), &te
, set_zero
))
539 /* We reached a condition that is always true. This is very hard to
540 reproduce (such a loop does not roll, and thus it would most
541 likely get optimized out by some of the preceding optimizations).
542 In fact, I do not have any testcase for it. However, it would
543 also be very hard to show that it is impossible, so we must
545 set_zero
->count
= preheader
->count
;
546 set_zero
->frequency
= preheader
->frequency
;
549 if (EDGE_COUNT (set_zero
->preds
) == 0)
551 /* All the conditions were simplified to false, remove the
552 unreachable set_zero block. */
553 delete_basic_block (set_zero
);
557 /* Reset the counter to zero in the set_zero block. */
559 convert_move (counter_reg
, noloop
, 0);
560 sequence
= get_insns ();
562 emit_insn_after (sequence
, BB_END (set_zero
));
564 set_immediate_dominator (CDI_DOMINATORS
, set_zero
,
565 recompute_dominator (CDI_DOMINATORS
,
569 set_immediate_dominator (CDI_DOMINATORS
, new_preheader
,
570 recompute_dominator (CDI_DOMINATORS
,
574 /* Some targets (eg, C4x) need to initialize special looping
576 #ifdef HAVE_doloop_begin
580 init
= gen_doloop_begin (counter_reg
, doloop_seq
);
585 sequence
= get_insns ();
587 emit_insn_after (sequence
, BB_END (loop_preheader_edge (loop
)->src
));
592 /* Insert the new low-overhead looping insn. */
593 emit_jump_insn_after (doloop_seq
, BB_END (loop_end
));
594 jump_insn
= BB_END (loop_end
);
595 jump_label
= block_label (desc
->in_edge
->dest
);
596 JUMP_LABEL (jump_insn
) = jump_label
;
597 LABEL_NUSES (jump_label
)++;
599 /* Ensure the right fallthru edge is marked, for case we have reversed
601 desc
->in_edge
->flags
&= ~EDGE_FALLTHRU
;
602 desc
->out_edge
->flags
|= EDGE_FALLTHRU
;
604 /* Add a REG_NONNEG note if the actual or estimated maximum number
605 of iterations is non-negative. */
607 add_reg_note (jump_insn
, REG_NONNEG
, NULL_RTX
);
609 /* Update the REG_BR_PROB note. */
612 /* Seems safer to use the branch probability. */
613 add_int_reg_note (jump_insn
, REG_BR_PROB
, desc
->in_edge
->probability
);
617 /* Process loop described by LOOP validating that the loop is suitable for
618 conversion to use a low overhead looping instruction, replacing the jump
619 insn where suitable. Returns true if the loop was successfully
623 doloop_optimize (struct loop
*loop
)
626 rtx doloop_seq
, doloop_pat
, doloop_reg
;
628 widest_int iterations
, iterations_max
;
629 rtx_code_label
*start_label
;
631 unsigned level
, est_niter
;
633 struct niter_desc
*desc
;
634 unsigned word_mode_size
;
635 unsigned HOST_WIDE_INT word_mode_max
;
639 fprintf (dump_file
, "Doloop: Processing loop %d.\n", loop
->num
);
641 iv_analysis_loop_init (loop
);
643 /* Find the simple exit of a LOOP. */
644 desc
= get_simple_loop_desc (loop
);
646 /* Check that loop is a candidate for a low-overhead looping insn. */
647 if (!doloop_valid_p (loop
, desc
))
651 "Doloop: The loop is not suitable.\n");
657 if (desc
->const_iter
)
658 est_niter
= desc
->niter
;
659 /* If the estimate on number of iterations is reliable (comes from profile
660 feedback), use it. Do not use it normally, since the expected number
661 of iterations of an unrolled loop is 2. */
662 if (loop
->header
->count
)
663 est_niter
= expected_loop_iterations (loop
);
669 "Doloop: Too few iterations (%u) to be profitable.\n",
675 = COSTS_N_INSNS (PARAM_VALUE (PARAM_MAX_ITERATIONS_COMPUTATION_COST
));
676 if (set_src_cost (desc
->niter_expr
, optimize_loop_for_speed_p (loop
))
681 "Doloop: number of iterations too costly to compute.\n");
685 if (desc
->const_iter
)
686 iterations
= widest_int::from (std::make_pair (desc
->niter_expr
, mode
),
690 if (!get_max_loop_iterations (loop
, &iterations_max
))
692 level
= get_loop_level (loop
) + 1;
693 entered_at_top
= (loop
->latch
== desc
->in_edge
->dest
694 && contains_no_active_insn_p (loop
->latch
));
695 if (!targetm
.can_use_doloop_p (iterations
, iterations_max
, level
,
699 fprintf (dump_file
, "Loop rejected by can_use_doloop_p.\n");
703 /* Generate looping insn. If the pattern FAILs then give up trying
704 to modify the loop since there is some aspect the back-end does
706 count
= copy_rtx (desc
->niter_expr
);
707 start_label
= block_label (desc
->in_edge
->dest
);
708 doloop_reg
= gen_reg_rtx (mode
);
709 doloop_seq
= gen_doloop_end (doloop_reg
, start_label
);
711 word_mode_size
= GET_MODE_PRECISION (word_mode
);
713 = ((unsigned HOST_WIDE_INT
) 1 << (word_mode_size
- 1) << 1) - 1;
716 /* Before trying mode different from the one in that # of iterations is
717 computed, we must be sure that the number of iterations fits into
719 && (word_mode_size
>= GET_MODE_PRECISION (mode
)
720 || wi::leu_p (iterations_max
, word_mode_max
)))
722 if (word_mode_size
> GET_MODE_PRECISION (mode
))
723 count
= simplify_gen_unary (ZERO_EXTEND
, word_mode
, count
, mode
);
725 count
= lowpart_subreg (word_mode
, count
, mode
);
726 PUT_MODE (doloop_reg
, word_mode
);
727 doloop_seq
= gen_doloop_end (doloop_reg
, start_label
);
733 "Doloop: Target unwilling to use doloop pattern!\n");
737 /* If multiple instructions were created, the last must be the
738 jump instruction. Also, a raw define_insn may yield a plain
740 doloop_pat
= doloop_seq
;
741 if (INSN_P (doloop_pat
))
743 rtx_insn
*doloop_insn
= as_a
<rtx_insn
*> (doloop_pat
);
744 while (NEXT_INSN (doloop_insn
) != NULL_RTX
)
745 doloop_insn
= NEXT_INSN (doloop_insn
);
746 if (!JUMP_P (doloop_insn
))
748 doloop_pat
= doloop_insn
;
752 || ! (condition
= doloop_condition_get (doloop_pat
)))
755 fprintf (dump_file
, "Doloop: Unrecognizable doloop pattern!\n");
759 doloop_modify (loop
, desc
, doloop_seq
, condition
, count
);
763 /* This is the main entry point. Process all loops using doloop_optimize. */
766 doloop_optimize_loops (void)
770 FOR_EACH_LOOP (loop
, 0)
772 doloop_optimize (loop
);
777 #ifdef ENABLE_CHECKING
778 verify_loop_structure ();
781 #endif /* HAVE_doloop_end */