2 Copyright (C) 2003-2013 Free Software Foundation, Inc.
3 Contributed by Dorit Naishlos <dorit@il.ibm.com> and
4 Ira Rosen <irar@il.ibm.com>
6 This file is part of GCC.
8 GCC is free software; you can redistribute it and/or modify it under
9 the terms of the GNU General Public License as published by the Free
10 Software Foundation; either version 3, or (at your option) any later
13 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
14 WARRANTY; without even the implied warranty of MERCHANTABILITY or
15 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
18 You should have received a copy of the GNU General Public License
19 along with GCC; see the file COPYING3. If not see
20 <http://www.gnu.org/licenses/>. */
24 #include "coretypes.h"
29 #include "basic-block.h"
30 #include "gimple-pretty-print.h"
31 #include "tree-flow.h"
32 #include "tree-pass.h"
38 #include "diagnostic-core.h"
39 #include "tree-chrec.h"
40 #include "tree-scalar-evolution.h"
41 #include "tree-vectorizer.h"
44 /* Loop Vectorization Pass.
46 This pass tries to vectorize loops.
48 For example, the vectorizer transforms the following simple loop:
50 short a[N]; short b[N]; short c[N]; int i;
56 as if it was manually vectorized by rewriting the source code into:
58 typedef int __attribute__((mode(V8HI))) v8hi;
59 short a[N]; short b[N]; short c[N]; int i;
60 v8hi *pa = (v8hi*)a, *pb = (v8hi*)b, *pc = (v8hi*)c;
63 for (i=0; i<N/8; i++){
70 The main entry to this pass is vectorize_loops(), in which
71 the vectorizer applies a set of analyses on a given set of loops,
72 followed by the actual vectorization transformation for the loops that
73 had successfully passed the analysis phase.
74 Throughout this pass we make a distinction between two types of
75 data: scalars (which are represented by SSA_NAMES), and memory references
76 ("data-refs"). These two types of data require different handling both
77 during analysis and transformation. The types of data-refs that the
78 vectorizer currently supports are ARRAY_REFS which base is an array DECL
79 (not a pointer), and INDIRECT_REFS through pointers; both array and pointer
80 accesses are required to have a simple (consecutive) access pattern.
84 The driver for the analysis phase is vect_analyze_loop().
85 It applies a set of analyses, some of which rely on the scalar evolution
86 analyzer (scev) developed by Sebastian Pop.
88 During the analysis phase the vectorizer records some information
89 per stmt in a "stmt_vec_info" struct which is attached to each stmt in the
90 loop, as well as general information about the loop as a whole, which is
91 recorded in a "loop_vec_info" struct attached to each loop.
95 The loop transformation phase scans all the stmts in the loop, and
96 creates a vector stmt (or a sequence of stmts) for each scalar stmt S in
97 the loop that needs to be vectorized. It inserts the vector code sequence
98 just before the scalar stmt S, and records a pointer to the vector code
99 in STMT_VINFO_VEC_STMT (stmt_info) (stmt_info is the stmt_vec_info struct
100 attached to S). This pointer will be used for the vectorization of following
101 stmts which use the def of stmt S. Stmt S is removed if it writes to memory;
102 otherwise, we rely on dead code elimination for removing it.
104 For example, say stmt S1 was vectorized into stmt VS1:
107 S1: b = x[i]; STMT_VINFO_VEC_STMT (stmt_info (S1)) = VS1
110 To vectorize stmt S2, the vectorizer first finds the stmt that defines
111 the operand 'b' (S1), and gets the relevant vector def 'vb' from the
112 vector stmt VS1 pointed to by STMT_VINFO_VEC_STMT (stmt_info (S1)). The
113 resulting sequence would be:
116 S1: b = x[i]; STMT_VINFO_VEC_STMT (stmt_info (S1)) = VS1
118 S2: a = b; STMT_VINFO_VEC_STMT (stmt_info (S2)) = VS2
120 Operands that are not SSA_NAMEs, are data-refs that appear in
121 load/store operations (like 'x[i]' in S1), and are handled differently.
125 Currently the only target specific information that is used is the
126 size of the vector (in bytes) - "TARGET_VECTORIZE_UNITS_PER_SIMD_WORD".
127 Targets that can support different sizes of vectors, for now will need
128 to specify one value for "TARGET_VECTORIZE_UNITS_PER_SIMD_WORD". More
129 flexibility will be added in the future.
131 Since we only vectorize operations which vector form can be
132 expressed using existing tree codes, to verify that an operation is
133 supported, the vectorizer checks the relevant optab at the relevant
134 machine_mode (e.g, optab_handler (add_optab, V8HImode)). If
135 the value found is CODE_FOR_nothing, then there's no target support, and
136 we can't vectorize the stmt.
138 For additional information on this project see:
139 http://gcc.gnu.org/projects/tree-ssa/vectorization.html
142 static void vect_estimate_min_profitable_iters (loop_vec_info
, int *, int *);
144 /* Function vect_determine_vectorization_factor
146 Determine the vectorization factor (VF). VF is the number of data elements
147 that are operated upon in parallel in a single iteration of the vectorized
148 loop. For example, when vectorizing a loop that operates on 4byte elements,
149 on a target with vector size (VS) 16byte, the VF is set to 4, since 4
150 elements can fit in a single vector register.
152 We currently support vectorization of loops in which all types operated upon
153 are of the same size. Therefore this function currently sets VF according to
154 the size of the types operated upon, and fails if there are multiple sizes
157 VF is also the factor by which the loop iterations are strip-mined, e.g.:
164 for (i=0; i<N; i+=VF){
165 a[i:VF] = b[i:VF] + c[i:VF];
170 vect_determine_vectorization_factor (loop_vec_info loop_vinfo
)
172 struct loop
*loop
= LOOP_VINFO_LOOP (loop_vinfo
);
173 basic_block
*bbs
= LOOP_VINFO_BBS (loop_vinfo
);
174 int nbbs
= loop
->num_nodes
;
175 gimple_stmt_iterator si
;
176 unsigned int vectorization_factor
= 0;
181 stmt_vec_info stmt_info
;
184 gimple stmt
, pattern_stmt
= NULL
;
185 gimple_seq pattern_def_seq
= NULL
;
186 gimple_stmt_iterator pattern_def_si
= gsi_none ();
187 bool analyze_pattern_stmt
= false;
189 if (dump_enabled_p ())
190 dump_printf_loc (MSG_NOTE
, vect_location
,
191 "=== vect_determine_vectorization_factor ===");
193 for (i
= 0; i
< nbbs
; i
++)
195 basic_block bb
= bbs
[i
];
197 for (si
= gsi_start_phis (bb
); !gsi_end_p (si
); gsi_next (&si
))
200 stmt_info
= vinfo_for_stmt (phi
);
201 if (dump_enabled_p ())
203 dump_printf_loc (MSG_NOTE
, vect_location
, "==> examining phi: ");
204 dump_gimple_stmt (MSG_NOTE
, TDF_SLIM
, phi
, 0);
207 gcc_assert (stmt_info
);
209 if (STMT_VINFO_RELEVANT_P (stmt_info
))
211 gcc_assert (!STMT_VINFO_VECTYPE (stmt_info
));
212 scalar_type
= TREE_TYPE (PHI_RESULT (phi
));
214 if (dump_enabled_p ())
216 dump_printf_loc (MSG_NOTE
, vect_location
,
217 "get vectype for scalar type: ");
218 dump_generic_expr (MSG_NOTE
, TDF_SLIM
, scalar_type
);
221 vectype
= get_vectype_for_scalar_type (scalar_type
);
224 if (dump_enabled_p ())
226 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
227 "not vectorized: unsupported "
229 dump_generic_expr (MSG_MISSED_OPTIMIZATION
, TDF_SLIM
,
234 STMT_VINFO_VECTYPE (stmt_info
) = vectype
;
236 if (dump_enabled_p ())
238 dump_printf_loc (MSG_NOTE
, vect_location
, "vectype: ");
239 dump_generic_expr (MSG_NOTE
, TDF_SLIM
, vectype
);
242 nunits
= TYPE_VECTOR_SUBPARTS (vectype
);
243 if (dump_enabled_p ())
244 dump_printf_loc (MSG_NOTE
, vect_location
, "nunits = %d", nunits
);
246 if (!vectorization_factor
247 || (nunits
> vectorization_factor
))
248 vectorization_factor
= nunits
;
252 for (si
= gsi_start_bb (bb
); !gsi_end_p (si
) || analyze_pattern_stmt
;)
256 if (analyze_pattern_stmt
)
259 stmt
= gsi_stmt (si
);
261 stmt_info
= vinfo_for_stmt (stmt
);
263 if (dump_enabled_p ())
265 dump_printf_loc (MSG_NOTE
, vect_location
,
266 "==> examining statement: ");
267 dump_gimple_stmt (MSG_NOTE
, TDF_SLIM
, stmt
, 0);
270 gcc_assert (stmt_info
);
272 /* Skip stmts which do not need to be vectorized. */
273 if (!STMT_VINFO_RELEVANT_P (stmt_info
)
274 && !STMT_VINFO_LIVE_P (stmt_info
))
276 if (STMT_VINFO_IN_PATTERN_P (stmt_info
)
277 && (pattern_stmt
= STMT_VINFO_RELATED_STMT (stmt_info
))
278 && (STMT_VINFO_RELEVANT_P (vinfo_for_stmt (pattern_stmt
))
279 || STMT_VINFO_LIVE_P (vinfo_for_stmt (pattern_stmt
))))
282 stmt_info
= vinfo_for_stmt (pattern_stmt
);
283 if (dump_enabled_p ())
285 dump_printf_loc (MSG_NOTE
, vect_location
,
286 "==> examining pattern statement: ");
287 dump_gimple_stmt (MSG_NOTE
, TDF_SLIM
, stmt
, 0);
292 if (dump_enabled_p ())
293 dump_printf_loc (MSG_NOTE
, vect_location
, "skip.");
298 else if (STMT_VINFO_IN_PATTERN_P (stmt_info
)
299 && (pattern_stmt
= STMT_VINFO_RELATED_STMT (stmt_info
))
300 && (STMT_VINFO_RELEVANT_P (vinfo_for_stmt (pattern_stmt
))
301 || STMT_VINFO_LIVE_P (vinfo_for_stmt (pattern_stmt
))))
302 analyze_pattern_stmt
= true;
304 /* If a pattern statement has def stmts, analyze them too. */
305 if (is_pattern_stmt_p (stmt_info
))
307 if (pattern_def_seq
== NULL
)
309 pattern_def_seq
= STMT_VINFO_PATTERN_DEF_SEQ (stmt_info
);
310 pattern_def_si
= gsi_start (pattern_def_seq
);
312 else if (!gsi_end_p (pattern_def_si
))
313 gsi_next (&pattern_def_si
);
314 if (pattern_def_seq
!= NULL
)
316 gimple pattern_def_stmt
= NULL
;
317 stmt_vec_info pattern_def_stmt_info
= NULL
;
319 while (!gsi_end_p (pattern_def_si
))
321 pattern_def_stmt
= gsi_stmt (pattern_def_si
);
322 pattern_def_stmt_info
323 = vinfo_for_stmt (pattern_def_stmt
);
324 if (STMT_VINFO_RELEVANT_P (pattern_def_stmt_info
)
325 || STMT_VINFO_LIVE_P (pattern_def_stmt_info
))
327 gsi_next (&pattern_def_si
);
330 if (!gsi_end_p (pattern_def_si
))
332 if (dump_enabled_p ())
334 dump_printf_loc (MSG_NOTE
, vect_location
,
335 "==> examining pattern def stmt: ");
336 dump_gimple_stmt (MSG_NOTE
, TDF_SLIM
,
337 pattern_def_stmt
, 0);
340 stmt
= pattern_def_stmt
;
341 stmt_info
= pattern_def_stmt_info
;
345 pattern_def_si
= gsi_none ();
346 analyze_pattern_stmt
= false;
350 analyze_pattern_stmt
= false;
353 if (gimple_get_lhs (stmt
) == NULL_TREE
)
355 if (dump_enabled_p ())
357 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
358 "not vectorized: irregular stmt.");
359 dump_gimple_stmt (MSG_MISSED_OPTIMIZATION
, TDF_SLIM
, stmt
,
365 if (VECTOR_MODE_P (TYPE_MODE (gimple_expr_type (stmt
))))
367 if (dump_enabled_p ())
369 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
370 "not vectorized: vector stmt in loop:");
371 dump_gimple_stmt (MSG_MISSED_OPTIMIZATION
, TDF_SLIM
, stmt
, 0);
376 if (STMT_VINFO_VECTYPE (stmt_info
))
378 /* The only case when a vectype had been already set is for stmts
379 that contain a dataref, or for "pattern-stmts" (stmts
380 generated by the vectorizer to represent/replace a certain
382 gcc_assert (STMT_VINFO_DATA_REF (stmt_info
)
383 || is_pattern_stmt_p (stmt_info
)
384 || !gsi_end_p (pattern_def_si
));
385 vectype
= STMT_VINFO_VECTYPE (stmt_info
);
389 gcc_assert (!STMT_VINFO_DATA_REF (stmt_info
));
390 scalar_type
= TREE_TYPE (gimple_get_lhs (stmt
));
391 if (dump_enabled_p ())
393 dump_printf_loc (MSG_NOTE
, vect_location
,
394 "get vectype for scalar type: ");
395 dump_generic_expr (MSG_NOTE
, TDF_SLIM
, scalar_type
);
397 vectype
= get_vectype_for_scalar_type (scalar_type
);
400 if (dump_enabled_p ())
402 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
403 "not vectorized: unsupported "
405 dump_generic_expr (MSG_MISSED_OPTIMIZATION
, TDF_SLIM
,
411 STMT_VINFO_VECTYPE (stmt_info
) = vectype
;
414 /* The vectorization factor is according to the smallest
415 scalar type (or the largest vector size, but we only
416 support one vector size per loop). */
417 scalar_type
= vect_get_smallest_scalar_type (stmt
, &dummy
,
419 if (dump_enabled_p ())
421 dump_printf_loc (MSG_NOTE
, vect_location
,
422 "get vectype for scalar type: ");
423 dump_generic_expr (MSG_NOTE
, TDF_SLIM
, scalar_type
);
425 vf_vectype
= get_vectype_for_scalar_type (scalar_type
);
428 if (dump_enabled_p ())
430 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
431 "not vectorized: unsupported data-type ");
432 dump_generic_expr (MSG_MISSED_OPTIMIZATION
, TDF_SLIM
,
438 if ((GET_MODE_SIZE (TYPE_MODE (vectype
))
439 != GET_MODE_SIZE (TYPE_MODE (vf_vectype
))))
441 if (dump_enabled_p ())
443 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
444 "not vectorized: different sized vector "
445 "types in statement, ");
446 dump_generic_expr (MSG_MISSED_OPTIMIZATION
, TDF_SLIM
,
448 dump_printf (MSG_MISSED_OPTIMIZATION
, " and ");
449 dump_generic_expr (MSG_MISSED_OPTIMIZATION
, TDF_SLIM
,
455 if (dump_enabled_p ())
457 dump_printf_loc (MSG_NOTE
, vect_location
, "vectype: ");
458 dump_generic_expr (MSG_NOTE
, TDF_SLIM
, vf_vectype
);
461 nunits
= TYPE_VECTOR_SUBPARTS (vf_vectype
);
462 if (dump_enabled_p ())
463 dump_printf_loc (MSG_NOTE
, vect_location
, "nunits = %d", nunits
);
464 if (!vectorization_factor
465 || (nunits
> vectorization_factor
))
466 vectorization_factor
= nunits
;
468 if (!analyze_pattern_stmt
&& gsi_end_p (pattern_def_si
))
470 pattern_def_seq
= NULL
;
476 /* TODO: Analyze cost. Decide if worth while to vectorize. */
477 if (dump_enabled_p ())
478 dump_printf_loc (MSG_NOTE
, vect_location
, "vectorization factor = %d",
479 vectorization_factor
);
480 if (vectorization_factor
<= 1)
482 if (dump_enabled_p ())
483 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
484 "not vectorized: unsupported data-type");
487 LOOP_VINFO_VECT_FACTOR (loop_vinfo
) = vectorization_factor
;
493 /* Function vect_is_simple_iv_evolution.
495 FORNOW: A simple evolution of an induction variables in the loop is
496 considered a polynomial evolution with constant step. */
499 vect_is_simple_iv_evolution (unsigned loop_nb
, tree access_fn
, tree
* init
,
504 tree evolution_part
= evolution_part_in_loop_num (access_fn
, loop_nb
);
506 /* When there is no evolution in this loop, the evolution function
508 if (evolution_part
== NULL_TREE
)
511 /* When the evolution is a polynomial of degree >= 2
512 the evolution function is not "simple". */
513 if (tree_is_chrec (evolution_part
))
516 step_expr
= evolution_part
;
517 init_expr
= unshare_expr (initial_condition_in_loop_num (access_fn
, loop_nb
));
519 if (dump_enabled_p ())
521 dump_printf_loc (MSG_NOTE
, vect_location
, "step: ");
522 dump_generic_expr (MSG_NOTE
, TDF_SLIM
, step_expr
);
523 dump_printf (MSG_NOTE
, ", init: ");
524 dump_generic_expr (MSG_NOTE
, TDF_SLIM
, init_expr
);
530 if (TREE_CODE (step_expr
) != INTEGER_CST
)
532 if (dump_enabled_p ())
533 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
541 /* Function vect_analyze_scalar_cycles_1.
543 Examine the cross iteration def-use cycles of scalar variables
544 in LOOP. LOOP_VINFO represents the loop that is now being
545 considered for vectorization (can be LOOP, or an outer-loop
549 vect_analyze_scalar_cycles_1 (loop_vec_info loop_vinfo
, struct loop
*loop
)
551 basic_block bb
= loop
->header
;
553 vec
<gimple
> worklist
;
554 worklist
.create (64);
555 gimple_stmt_iterator gsi
;
558 if (dump_enabled_p ())
559 dump_printf_loc (MSG_NOTE
, vect_location
,
560 "=== vect_analyze_scalar_cycles ===");
562 /* First - identify all inductions. Reduction detection assumes that all the
563 inductions have been identified, therefore, this order must not be
565 for (gsi
= gsi_start_phis (bb
); !gsi_end_p (gsi
); gsi_next (&gsi
))
567 gimple phi
= gsi_stmt (gsi
);
568 tree access_fn
= NULL
;
569 tree def
= PHI_RESULT (phi
);
570 stmt_vec_info stmt_vinfo
= vinfo_for_stmt (phi
);
572 if (dump_enabled_p ())
574 dump_printf_loc (MSG_NOTE
, vect_location
, "Analyze phi: ");
575 dump_gimple_stmt (MSG_NOTE
, TDF_SLIM
, phi
, 0);
578 /* Skip virtual phi's. The data dependences that are associated with
579 virtual defs/uses (i.e., memory accesses) are analyzed elsewhere. */
580 if (virtual_operand_p (def
))
583 STMT_VINFO_DEF_TYPE (stmt_vinfo
) = vect_unknown_def_type
;
585 /* Analyze the evolution function. */
586 access_fn
= analyze_scalar_evolution (loop
, def
);
589 STRIP_NOPS (access_fn
);
590 if (dump_enabled_p ())
592 dump_printf_loc (MSG_NOTE
, vect_location
,
593 "Access function of PHI: ");
594 dump_generic_expr (MSG_NOTE
, TDF_SLIM
, access_fn
);
596 STMT_VINFO_LOOP_PHI_EVOLUTION_PART (stmt_vinfo
)
597 = evolution_part_in_loop_num (access_fn
, loop
->num
);
601 || !vect_is_simple_iv_evolution (loop
->num
, access_fn
, &dumy
, &dumy
))
603 worklist
.safe_push (phi
);
607 gcc_assert (STMT_VINFO_LOOP_PHI_EVOLUTION_PART (stmt_vinfo
) != NULL_TREE
);
609 if (dump_enabled_p ())
610 dump_printf_loc (MSG_NOTE
, vect_location
, "Detected induction.");
611 STMT_VINFO_DEF_TYPE (stmt_vinfo
) = vect_induction_def
;
615 /* Second - identify all reductions and nested cycles. */
616 while (worklist
.length () > 0)
618 gimple phi
= worklist
.pop ();
619 tree def
= PHI_RESULT (phi
);
620 stmt_vec_info stmt_vinfo
= vinfo_for_stmt (phi
);
624 if (dump_enabled_p ())
626 dump_printf_loc (MSG_NOTE
, vect_location
, "Analyze phi: ");
627 dump_gimple_stmt (MSG_NOTE
, TDF_SLIM
, phi
, 0);
630 gcc_assert (!virtual_operand_p (def
)
631 && STMT_VINFO_DEF_TYPE (stmt_vinfo
) == vect_unknown_def_type
);
633 nested_cycle
= (loop
!= LOOP_VINFO_LOOP (loop_vinfo
));
634 reduc_stmt
= vect_force_simple_reduction (loop_vinfo
, phi
, !nested_cycle
,
640 if (dump_enabled_p ())
641 dump_printf_loc (MSG_NOTE
, vect_location
,
642 "Detected double reduction.");
644 STMT_VINFO_DEF_TYPE (stmt_vinfo
) = vect_double_reduction_def
;
645 STMT_VINFO_DEF_TYPE (vinfo_for_stmt (reduc_stmt
)) =
646 vect_double_reduction_def
;
652 if (dump_enabled_p ())
653 dump_printf_loc (MSG_NOTE
, vect_location
,
654 "Detected vectorizable nested cycle.");
656 STMT_VINFO_DEF_TYPE (stmt_vinfo
) = vect_nested_cycle
;
657 STMT_VINFO_DEF_TYPE (vinfo_for_stmt (reduc_stmt
)) =
662 if (dump_enabled_p ())
663 dump_printf_loc (MSG_NOTE
, vect_location
,
664 "Detected reduction.");
666 STMT_VINFO_DEF_TYPE (stmt_vinfo
) = vect_reduction_def
;
667 STMT_VINFO_DEF_TYPE (vinfo_for_stmt (reduc_stmt
)) =
669 /* Store the reduction cycles for possible vectorization in
671 LOOP_VINFO_REDUCTIONS (loop_vinfo
).safe_push (reduc_stmt
);
676 if (dump_enabled_p ())
677 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
678 "Unknown def-use cycle pattern.");
685 /* Function vect_analyze_scalar_cycles.
687 Examine the cross iteration def-use cycles of scalar variables, by
688 analyzing the loop-header PHIs of scalar variables. Classify each
689 cycle as one of the following: invariant, induction, reduction, unknown.
690 We do that for the loop represented by LOOP_VINFO, and also to its
691 inner-loop, if exists.
692 Examples for scalar cycles:
707 vect_analyze_scalar_cycles (loop_vec_info loop_vinfo
)
709 struct loop
*loop
= LOOP_VINFO_LOOP (loop_vinfo
);
711 vect_analyze_scalar_cycles_1 (loop_vinfo
, loop
);
713 /* When vectorizing an outer-loop, the inner-loop is executed sequentially.
714 Reductions in such inner-loop therefore have different properties than
715 the reductions in the nest that gets vectorized:
716 1. When vectorized, they are executed in the same order as in the original
717 scalar loop, so we can't change the order of computation when
719 2. FIXME: Inner-loop reductions can be used in the inner-loop, so the
720 current checks are too strict. */
723 vect_analyze_scalar_cycles_1 (loop_vinfo
, loop
->inner
);
726 /* Function vect_get_loop_niters.
728 Determine how many iterations the loop is executed.
729 If an expression that represents the number of iterations
730 can be constructed, place it in NUMBER_OF_ITERATIONS.
731 Return the loop exit condition. */
734 vect_get_loop_niters (struct loop
*loop
, tree
*number_of_iterations
)
738 if (dump_enabled_p ())
739 dump_printf_loc (MSG_NOTE
, vect_location
,
740 "=== get_loop_niters ===");
741 niters
= number_of_exit_cond_executions (loop
);
743 if (niters
!= NULL_TREE
744 && niters
!= chrec_dont_know
)
746 *number_of_iterations
= niters
;
748 if (dump_enabled_p ())
750 dump_printf_loc (MSG_NOTE
, vect_location
, "==> get_loop_niters:");
751 dump_generic_expr (MSG_NOTE
, TDF_SLIM
, *number_of_iterations
);
755 return get_loop_exit_condition (loop
);
759 /* Function bb_in_loop_p
761 Used as predicate for dfs order traversal of the loop bbs. */
764 bb_in_loop_p (const_basic_block bb
, const void *data
)
766 const struct loop
*const loop
= (const struct loop
*)data
;
767 if (flow_bb_inside_loop_p (loop
, bb
))
773 /* Function new_loop_vec_info.
775 Create and initialize a new loop_vec_info struct for LOOP, as well as
776 stmt_vec_info structs for all the stmts in LOOP. */
779 new_loop_vec_info (struct loop
*loop
)
783 gimple_stmt_iterator si
;
784 unsigned int i
, nbbs
;
786 res
= (loop_vec_info
) xcalloc (1, sizeof (struct _loop_vec_info
));
787 LOOP_VINFO_LOOP (res
) = loop
;
789 bbs
= get_loop_body (loop
);
791 /* Create/Update stmt_info for all stmts in the loop. */
792 for (i
= 0; i
< loop
->num_nodes
; i
++)
794 basic_block bb
= bbs
[i
];
796 /* BBs in a nested inner-loop will have been already processed (because
797 we will have called vect_analyze_loop_form for any nested inner-loop).
798 Therefore, for stmts in an inner-loop we just want to update the
799 STMT_VINFO_LOOP_VINFO field of their stmt_info to point to the new
800 loop_info of the outer-loop we are currently considering to vectorize
801 (instead of the loop_info of the inner-loop).
802 For stmts in other BBs we need to create a stmt_info from scratch. */
803 if (bb
->loop_father
!= loop
)
806 gcc_assert (loop
->inner
&& bb
->loop_father
== loop
->inner
);
807 for (si
= gsi_start_phis (bb
); !gsi_end_p (si
); gsi_next (&si
))
809 gimple phi
= gsi_stmt (si
);
810 stmt_vec_info stmt_info
= vinfo_for_stmt (phi
);
811 loop_vec_info inner_loop_vinfo
=
812 STMT_VINFO_LOOP_VINFO (stmt_info
);
813 gcc_assert (loop
->inner
== LOOP_VINFO_LOOP (inner_loop_vinfo
));
814 STMT_VINFO_LOOP_VINFO (stmt_info
) = res
;
816 for (si
= gsi_start_bb (bb
); !gsi_end_p (si
); gsi_next (&si
))
818 gimple stmt
= gsi_stmt (si
);
819 stmt_vec_info stmt_info
= vinfo_for_stmt (stmt
);
820 loop_vec_info inner_loop_vinfo
=
821 STMT_VINFO_LOOP_VINFO (stmt_info
);
822 gcc_assert (loop
->inner
== LOOP_VINFO_LOOP (inner_loop_vinfo
));
823 STMT_VINFO_LOOP_VINFO (stmt_info
) = res
;
828 /* bb in current nest. */
829 for (si
= gsi_start_phis (bb
); !gsi_end_p (si
); gsi_next (&si
))
831 gimple phi
= gsi_stmt (si
);
832 gimple_set_uid (phi
, 0);
833 set_vinfo_for_stmt (phi
, new_stmt_vec_info (phi
, res
, NULL
));
836 for (si
= gsi_start_bb (bb
); !gsi_end_p (si
); gsi_next (&si
))
838 gimple stmt
= gsi_stmt (si
);
839 gimple_set_uid (stmt
, 0);
840 set_vinfo_for_stmt (stmt
, new_stmt_vec_info (stmt
, res
, NULL
));
845 /* CHECKME: We want to visit all BBs before their successors (except for
846 latch blocks, for which this assertion wouldn't hold). In the simple
847 case of the loop forms we allow, a dfs order of the BBs would the same
848 as reversed postorder traversal, so we are safe. */
851 bbs
= XCNEWVEC (basic_block
, loop
->num_nodes
);
852 nbbs
= dfs_enumerate_from (loop
->header
, 0, bb_in_loop_p
,
853 bbs
, loop
->num_nodes
, loop
);
854 gcc_assert (nbbs
== loop
->num_nodes
);
856 LOOP_VINFO_BBS (res
) = bbs
;
857 LOOP_VINFO_NITERS (res
) = NULL
;
858 LOOP_VINFO_NITERS_UNCHANGED (res
) = NULL
;
859 LOOP_VINFO_COST_MODEL_MIN_ITERS (res
) = 0;
860 LOOP_VINFO_VECTORIZABLE_P (res
) = 0;
861 LOOP_PEELING_FOR_ALIGNMENT (res
) = 0;
862 LOOP_VINFO_VECT_FACTOR (res
) = 0;
863 LOOP_VINFO_LOOP_NEST (res
).create (3);
864 LOOP_VINFO_DATAREFS (res
).create (10);
865 LOOP_VINFO_DDRS (res
).create (10 * 10);
866 LOOP_VINFO_UNALIGNED_DR (res
) = NULL
;
867 LOOP_VINFO_MAY_MISALIGN_STMTS (res
).create (
868 PARAM_VALUE (PARAM_VECT_MAX_VERSION_FOR_ALIGNMENT_CHECKS
));
869 LOOP_VINFO_MAY_ALIAS_DDRS (res
).create (
870 PARAM_VALUE (PARAM_VECT_MAX_VERSION_FOR_ALIAS_CHECKS
));
871 LOOP_VINFO_GROUPED_STORES (res
).create (10);
872 LOOP_VINFO_REDUCTIONS (res
).create (10);
873 LOOP_VINFO_REDUCTION_CHAINS (res
).create (10);
874 LOOP_VINFO_SLP_INSTANCES (res
).create (10);
875 LOOP_VINFO_SLP_UNROLLING_FACTOR (res
) = 1;
876 LOOP_VINFO_PEELING_HTAB (res
) = NULL
;
877 LOOP_VINFO_TARGET_COST_DATA (res
) = init_cost (loop
);
878 LOOP_VINFO_PEELING_FOR_GAPS (res
) = false;
879 LOOP_VINFO_OPERANDS_SWAPPED (res
) = false;
885 /* Function destroy_loop_vec_info.
887 Free LOOP_VINFO struct, as well as all the stmt_vec_info structs of all the
888 stmts in the loop. */
891 destroy_loop_vec_info (loop_vec_info loop_vinfo
, bool clean_stmts
)
896 gimple_stmt_iterator si
;
898 vec
<slp_instance
> slp_instances
;
899 slp_instance instance
;
905 loop
= LOOP_VINFO_LOOP (loop_vinfo
);
907 bbs
= LOOP_VINFO_BBS (loop_vinfo
);
908 nbbs
= loop
->num_nodes
;
909 swapped
= LOOP_VINFO_OPERANDS_SWAPPED (loop_vinfo
);
913 free (LOOP_VINFO_BBS (loop_vinfo
));
914 free_data_refs (LOOP_VINFO_DATAREFS (loop_vinfo
));
915 free_dependence_relations (LOOP_VINFO_DDRS (loop_vinfo
));
916 LOOP_VINFO_LOOP_NEST (loop_vinfo
).release ();
917 LOOP_VINFO_MAY_MISALIGN_STMTS (loop_vinfo
).release ();
918 LOOP_VINFO_MAY_ALIAS_DDRS (loop_vinfo
).release ();
925 for (j
= 0; j
< nbbs
; j
++)
927 basic_block bb
= bbs
[j
];
928 for (si
= gsi_start_phis (bb
); !gsi_end_p (si
); gsi_next (&si
))
929 free_stmt_vec_info (gsi_stmt (si
));
931 for (si
= gsi_start_bb (bb
); !gsi_end_p (si
); )
933 gimple stmt
= gsi_stmt (si
);
935 /* We may have broken canonical form by moving a constant
936 into RHS1 of a commutative op. Fix such occurrences. */
937 if (swapped
&& is_gimple_assign (stmt
))
939 enum tree_code code
= gimple_assign_rhs_code (stmt
);
941 if ((code
== PLUS_EXPR
942 || code
== POINTER_PLUS_EXPR
943 || code
== MULT_EXPR
)
944 && CONSTANT_CLASS_P (gimple_assign_rhs1 (stmt
)))
945 swap_tree_operands (stmt
,
946 gimple_assign_rhs1_ptr (stmt
),
947 gimple_assign_rhs2_ptr (stmt
));
950 /* Free stmt_vec_info. */
951 free_stmt_vec_info (stmt
);
956 free (LOOP_VINFO_BBS (loop_vinfo
));
957 free_data_refs (LOOP_VINFO_DATAREFS (loop_vinfo
));
958 free_dependence_relations (LOOP_VINFO_DDRS (loop_vinfo
));
959 LOOP_VINFO_LOOP_NEST (loop_vinfo
).release ();
960 LOOP_VINFO_MAY_MISALIGN_STMTS (loop_vinfo
).release ();
961 LOOP_VINFO_MAY_ALIAS_DDRS (loop_vinfo
).release ();
962 slp_instances
= LOOP_VINFO_SLP_INSTANCES (loop_vinfo
);
963 FOR_EACH_VEC_ELT (slp_instances
, j
, instance
)
964 vect_free_slp_instance (instance
);
966 LOOP_VINFO_SLP_INSTANCES (loop_vinfo
).release ();
967 LOOP_VINFO_GROUPED_STORES (loop_vinfo
).release ();
968 LOOP_VINFO_REDUCTIONS (loop_vinfo
).release ();
969 LOOP_VINFO_REDUCTION_CHAINS (loop_vinfo
).release ();
971 if (LOOP_VINFO_PEELING_HTAB (loop_vinfo
))
972 htab_delete (LOOP_VINFO_PEELING_HTAB (loop_vinfo
));
974 destroy_cost_data (LOOP_VINFO_TARGET_COST_DATA (loop_vinfo
));
981 /* Function vect_analyze_loop_1.
983 Apply a set of analyses on LOOP, and create a loop_vec_info struct
984 for it. The different analyses will record information in the
985 loop_vec_info struct. This is a subset of the analyses applied in
986 vect_analyze_loop, to be applied on an inner-loop nested in the loop
987 that is now considered for (outer-loop) vectorization. */
990 vect_analyze_loop_1 (struct loop
*loop
)
992 loop_vec_info loop_vinfo
;
994 if (dump_enabled_p ())
995 dump_printf_loc (MSG_NOTE
, vect_location
,
996 "===== analyze_loop_nest_1 =====");
998 /* Check the CFG characteristics of the loop (nesting, entry/exit, etc. */
1000 loop_vinfo
= vect_analyze_loop_form (loop
);
1003 if (dump_enabled_p ())
1004 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
1005 "bad inner-loop form.");
1013 /* Function vect_analyze_loop_form.
1015 Verify that certain CFG restrictions hold, including:
1016 - the loop has a pre-header
1017 - the loop has a single entry and exit
1018 - the loop exit condition is simple enough, and the number of iterations
1019 can be analyzed (a countable loop). */
1022 vect_analyze_loop_form (struct loop
*loop
)
1024 loop_vec_info loop_vinfo
;
1026 tree number_of_iterations
= NULL
;
1027 loop_vec_info inner_loop_vinfo
= NULL
;
1029 if (dump_enabled_p ())
1030 dump_printf_loc (MSG_NOTE
, vect_location
,
1031 "=== vect_analyze_loop_form ===");
1033 /* Different restrictions apply when we are considering an inner-most loop,
1034 vs. an outer (nested) loop.
1035 (FORNOW. May want to relax some of these restrictions in the future). */
1039 /* Inner-most loop. We currently require that the number of BBs is
1040 exactly 2 (the header and latch). Vectorizable inner-most loops
1051 if (loop
->num_nodes
!= 2)
1053 if (dump_enabled_p ())
1054 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
1055 "not vectorized: control flow in loop.");
1059 if (empty_block_p (loop
->header
))
1061 if (dump_enabled_p ())
1062 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
1063 "not vectorized: empty loop.");
1069 struct loop
*innerloop
= loop
->inner
;
1072 /* Nested loop. We currently require that the loop is doubly-nested,
1073 contains a single inner loop, and the number of BBs is exactly 5.
1074 Vectorizable outer-loops look like this:
1086 The inner-loop has the properties expected of inner-most loops
1087 as described above. */
1089 if ((loop
->inner
)->inner
|| (loop
->inner
)->next
)
1091 if (dump_enabled_p ())
1092 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
1093 "not vectorized: multiple nested loops.");
1097 /* Analyze the inner-loop. */
1098 inner_loop_vinfo
= vect_analyze_loop_1 (loop
->inner
);
1099 if (!inner_loop_vinfo
)
1101 if (dump_enabled_p ())
1102 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
1103 "not vectorized: Bad inner loop.");
1107 if (!expr_invariant_in_loop_p (loop
,
1108 LOOP_VINFO_NITERS (inner_loop_vinfo
)))
1110 if (dump_enabled_p ())
1111 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
1112 "not vectorized: inner-loop count not invariant.");
1113 destroy_loop_vec_info (inner_loop_vinfo
, true);
1117 if (loop
->num_nodes
!= 5)
1119 if (dump_enabled_p ())
1120 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
1121 "not vectorized: control flow in loop.");
1122 destroy_loop_vec_info (inner_loop_vinfo
, true);
1126 gcc_assert (EDGE_COUNT (innerloop
->header
->preds
) == 2);
1127 entryedge
= EDGE_PRED (innerloop
->header
, 0);
1128 if (EDGE_PRED (innerloop
->header
, 0)->src
== innerloop
->latch
)
1129 entryedge
= EDGE_PRED (innerloop
->header
, 1);
1131 if (entryedge
->src
!= loop
->header
1132 || !single_exit (innerloop
)
1133 || single_exit (innerloop
)->dest
!= EDGE_PRED (loop
->latch
, 0)->src
)
1135 if (dump_enabled_p ())
1136 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
1137 "not vectorized: unsupported outerloop form.");
1138 destroy_loop_vec_info (inner_loop_vinfo
, true);
1142 if (dump_enabled_p ())
1143 dump_printf_loc (MSG_NOTE
, vect_location
,
1144 "Considering outer-loop vectorization.");
1147 if (!single_exit (loop
)
1148 || EDGE_COUNT (loop
->header
->preds
) != 2)
1150 if (dump_enabled_p ())
1152 if (!single_exit (loop
))
1153 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
1154 "not vectorized: multiple exits.");
1155 else if (EDGE_COUNT (loop
->header
->preds
) != 2)
1156 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
1157 "not vectorized: too many incoming edges.");
1159 if (inner_loop_vinfo
)
1160 destroy_loop_vec_info (inner_loop_vinfo
, true);
1164 /* We assume that the loop exit condition is at the end of the loop. i.e,
1165 that the loop is represented as a do-while (with a proper if-guard
1166 before the loop if needed), where the loop header contains all the
1167 executable statements, and the latch is empty. */
1168 if (!empty_block_p (loop
->latch
)
1169 || !gimple_seq_empty_p (phi_nodes (loop
->latch
)))
1171 if (dump_enabled_p ())
1172 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
1173 "not vectorized: latch block not empty.");
1174 if (inner_loop_vinfo
)
1175 destroy_loop_vec_info (inner_loop_vinfo
, true);
1179 /* Make sure there exists a single-predecessor exit bb: */
1180 if (!single_pred_p (single_exit (loop
)->dest
))
1182 edge e
= single_exit (loop
);
1183 if (!(e
->flags
& EDGE_ABNORMAL
))
1185 split_loop_exit_edge (e
);
1186 if (dump_enabled_p ())
1187 dump_printf (MSG_NOTE
, "split exit edge.");
1191 if (dump_enabled_p ())
1192 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
1193 "not vectorized: abnormal loop exit edge.");
1194 if (inner_loop_vinfo
)
1195 destroy_loop_vec_info (inner_loop_vinfo
, true);
1200 loop_cond
= vect_get_loop_niters (loop
, &number_of_iterations
);
1203 if (dump_enabled_p ())
1204 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
1205 "not vectorized: complicated exit condition.");
1206 if (inner_loop_vinfo
)
1207 destroy_loop_vec_info (inner_loop_vinfo
, true);
1211 if (!number_of_iterations
)
1213 if (dump_enabled_p ())
1214 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
1215 "not vectorized: number of iterations cannot be "
1217 if (inner_loop_vinfo
)
1218 destroy_loop_vec_info (inner_loop_vinfo
, true);
1222 if (chrec_contains_undetermined (number_of_iterations
))
1224 if (dump_enabled_p ())
1225 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
1226 "Infinite number of iterations.");
1227 if (inner_loop_vinfo
)
1228 destroy_loop_vec_info (inner_loop_vinfo
, true);
1232 if (!NITERS_KNOWN_P (number_of_iterations
))
1234 if (dump_enabled_p ())
1236 dump_printf_loc (MSG_NOTE
, vect_location
,
1237 "Symbolic number of iterations is ");
1238 dump_generic_expr (MSG_NOTE
, TDF_DETAILS
, number_of_iterations
);
1241 else if (TREE_INT_CST_LOW (number_of_iterations
) == 0)
1243 if (dump_enabled_p ())
1244 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
1245 "not vectorized: number of iterations = 0.");
1246 if (inner_loop_vinfo
)
1247 destroy_loop_vec_info (inner_loop_vinfo
, false);
1251 loop_vinfo
= new_loop_vec_info (loop
);
1252 LOOP_VINFO_NITERS (loop_vinfo
) = number_of_iterations
;
1253 LOOP_VINFO_NITERS_UNCHANGED (loop_vinfo
) = number_of_iterations
;
1255 STMT_VINFO_TYPE (vinfo_for_stmt (loop_cond
)) = loop_exit_ctrl_vec_info_type
;
1257 /* CHECKME: May want to keep it around it in the future. */
1258 if (inner_loop_vinfo
)
1259 destroy_loop_vec_info (inner_loop_vinfo
, false);
1261 gcc_assert (!loop
->aux
);
1262 loop
->aux
= loop_vinfo
;
1267 /* Function vect_analyze_loop_operations.
1269 Scan the loop stmts and make sure they are all vectorizable. */
1272 vect_analyze_loop_operations (loop_vec_info loop_vinfo
, bool slp
)
1274 struct loop
*loop
= LOOP_VINFO_LOOP (loop_vinfo
);
1275 basic_block
*bbs
= LOOP_VINFO_BBS (loop_vinfo
);
1276 int nbbs
= loop
->num_nodes
;
1277 gimple_stmt_iterator si
;
1278 unsigned int vectorization_factor
= 0;
1281 stmt_vec_info stmt_info
;
1282 bool need_to_vectorize
= false;
1283 int min_profitable_iters
;
1284 int min_scalar_loop_bound
;
1286 bool only_slp_in_loop
= true, ok
;
1287 HOST_WIDE_INT max_niter
;
1288 HOST_WIDE_INT estimated_niter
;
1289 int min_profitable_estimate
;
1291 if (dump_enabled_p ())
1292 dump_printf_loc (MSG_NOTE
, vect_location
,
1293 "=== vect_analyze_loop_operations ===");
1295 gcc_assert (LOOP_VINFO_VECT_FACTOR (loop_vinfo
));
1296 vectorization_factor
= LOOP_VINFO_VECT_FACTOR (loop_vinfo
);
1299 /* If all the stmts in the loop can be SLPed, we perform only SLP, and
1300 vectorization factor of the loop is the unrolling factor required by
1301 the SLP instances. If that unrolling factor is 1, we say, that we
1302 perform pure SLP on loop - cross iteration parallelism is not
1304 for (i
= 0; i
< nbbs
; i
++)
1306 basic_block bb
= bbs
[i
];
1307 for (si
= gsi_start_bb (bb
); !gsi_end_p (si
); gsi_next (&si
))
1309 gimple stmt
= gsi_stmt (si
);
1310 stmt_vec_info stmt_info
= vinfo_for_stmt (stmt
);
1311 gcc_assert (stmt_info
);
1312 if ((STMT_VINFO_RELEVANT_P (stmt_info
)
1313 || VECTORIZABLE_CYCLE_DEF (STMT_VINFO_DEF_TYPE (stmt_info
)))
1314 && !PURE_SLP_STMT (stmt_info
))
1315 /* STMT needs both SLP and loop-based vectorization. */
1316 only_slp_in_loop
= false;
1320 if (only_slp_in_loop
)
1321 vectorization_factor
= LOOP_VINFO_SLP_UNROLLING_FACTOR (loop_vinfo
);
1323 vectorization_factor
= least_common_multiple (vectorization_factor
,
1324 LOOP_VINFO_SLP_UNROLLING_FACTOR (loop_vinfo
));
1326 LOOP_VINFO_VECT_FACTOR (loop_vinfo
) = vectorization_factor
;
1327 if (dump_enabled_p ())
1328 dump_printf_loc (MSG_NOTE
, vect_location
,
1329 "Updating vectorization factor to %d ",
1330 vectorization_factor
);
1333 for (i
= 0; i
< nbbs
; i
++)
1335 basic_block bb
= bbs
[i
];
1337 for (si
= gsi_start_phis (bb
); !gsi_end_p (si
); gsi_next (&si
))
1339 phi
= gsi_stmt (si
);
1342 stmt_info
= vinfo_for_stmt (phi
);
1343 if (dump_enabled_p ())
1345 dump_printf_loc (MSG_NOTE
, vect_location
, "examining phi: ");
1346 dump_gimple_stmt (MSG_NOTE
, TDF_SLIM
, phi
, 0);
1349 /* Inner-loop loop-closed exit phi in outer-loop vectorization
1350 (i.e., a phi in the tail of the outer-loop). */
1351 if (! is_loop_header_bb_p (bb
))
1353 /* FORNOW: we currently don't support the case that these phis
1354 are not used in the outerloop (unless it is double reduction,
1355 i.e., this phi is vect_reduction_def), cause this case
1356 requires to actually do something here. */
1357 if ((!STMT_VINFO_RELEVANT_P (stmt_info
)
1358 || STMT_VINFO_LIVE_P (stmt_info
))
1359 && STMT_VINFO_DEF_TYPE (stmt_info
)
1360 != vect_double_reduction_def
)
1362 if (dump_enabled_p ())
1363 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
1364 "Unsupported loop-closed phi in "
1369 /* If PHI is used in the outer loop, we check that its operand
1370 is defined in the inner loop. */
1371 if (STMT_VINFO_RELEVANT_P (stmt_info
))
1376 if (gimple_phi_num_args (phi
) != 1)
1379 phi_op
= PHI_ARG_DEF (phi
, 0);
1380 if (TREE_CODE (phi_op
) != SSA_NAME
)
1383 op_def_stmt
= SSA_NAME_DEF_STMT (phi_op
);
1385 || !flow_bb_inside_loop_p (loop
, gimple_bb (op_def_stmt
))
1386 || !vinfo_for_stmt (op_def_stmt
))
1389 if (STMT_VINFO_RELEVANT (vinfo_for_stmt (op_def_stmt
))
1390 != vect_used_in_outer
1391 && STMT_VINFO_RELEVANT (vinfo_for_stmt (op_def_stmt
))
1392 != vect_used_in_outer_by_reduction
)
1399 gcc_assert (stmt_info
);
1401 if (STMT_VINFO_LIVE_P (stmt_info
))
1403 /* FORNOW: not yet supported. */
1404 if (dump_enabled_p ())
1405 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
1406 "not vectorized: value used after loop.");
1410 if (STMT_VINFO_RELEVANT (stmt_info
) == vect_used_in_scope
1411 && STMT_VINFO_DEF_TYPE (stmt_info
) != vect_induction_def
)
1413 /* A scalar-dependence cycle that we don't support. */
1414 if (dump_enabled_p ())
1415 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
1416 "not vectorized: scalar dependence cycle.");
1420 if (STMT_VINFO_RELEVANT_P (stmt_info
))
1422 need_to_vectorize
= true;
1423 if (STMT_VINFO_DEF_TYPE (stmt_info
) == vect_induction_def
)
1424 ok
= vectorizable_induction (phi
, NULL
, NULL
);
1429 if (dump_enabled_p ())
1431 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
1432 "not vectorized: relevant phi not "
1434 dump_gimple_stmt (MSG_MISSED_OPTIMIZATION
, TDF_SLIM
, phi
, 0);
1440 for (si
= gsi_start_bb (bb
); !gsi_end_p (si
); gsi_next (&si
))
1442 gimple stmt
= gsi_stmt (si
);
1443 if (!vect_analyze_stmt (stmt
, &need_to_vectorize
, NULL
))
1448 /* All operations in the loop are either irrelevant (deal with loop
1449 control, or dead), or only used outside the loop and can be moved
1450 out of the loop (e.g. invariants, inductions). The loop can be
1451 optimized away by scalar optimizations. We're better off not
1452 touching this loop. */
1453 if (!need_to_vectorize
)
1455 if (dump_enabled_p ())
1456 dump_printf_loc (MSG_NOTE
, vect_location
,
1457 "All the computation can be taken out of the loop.");
1458 if (dump_enabled_p ())
1459 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
1460 "not vectorized: redundant loop. no profit to "
1465 if (LOOP_VINFO_NITERS_KNOWN_P (loop_vinfo
) && dump_enabled_p ())
1466 dump_printf_loc (MSG_NOTE
, vect_location
,
1467 "vectorization_factor = %d, niters = "
1468 HOST_WIDE_INT_PRINT_DEC
, vectorization_factor
,
1469 LOOP_VINFO_INT_NITERS (loop_vinfo
));
1471 if ((LOOP_VINFO_NITERS_KNOWN_P (loop_vinfo
)
1472 && (LOOP_VINFO_INT_NITERS (loop_vinfo
) < vectorization_factor
))
1473 || ((max_niter
= max_stmt_executions_int (loop
)) != -1
1474 && (unsigned HOST_WIDE_INT
) max_niter
< vectorization_factor
))
1476 if (dump_enabled_p ())
1477 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
1478 "not vectorized: iteration count too small.");
1479 if (dump_enabled_p ())
1480 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
1481 "not vectorized: iteration count smaller than "
1482 "vectorization factor.");
1486 /* Analyze cost. Decide if worth while to vectorize. */
1488 /* Once VF is set, SLP costs should be updated since the number of created
1489 vector stmts depends on VF. */
1490 vect_update_slp_costs_according_to_vf (loop_vinfo
);
1492 vect_estimate_min_profitable_iters (loop_vinfo
, &min_profitable_iters
,
1493 &min_profitable_estimate
);
1494 LOOP_VINFO_COST_MODEL_MIN_ITERS (loop_vinfo
) = min_profitable_iters
;
1496 if (min_profitable_iters
< 0)
1498 if (dump_enabled_p ())
1499 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
1500 "not vectorized: vectorization not profitable.");
1501 if (dump_enabled_p ())
1502 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
1503 "not vectorized: vector version will never be "
1508 min_scalar_loop_bound
= ((PARAM_VALUE (PARAM_MIN_VECT_LOOP_BOUND
)
1509 * vectorization_factor
) - 1);
1512 /* Use the cost model only if it is more conservative than user specified
1515 th
= (unsigned) min_scalar_loop_bound
;
1516 if (min_profitable_iters
1517 && (!min_scalar_loop_bound
1518 || min_profitable_iters
> min_scalar_loop_bound
))
1519 th
= (unsigned) min_profitable_iters
;
1521 if (LOOP_VINFO_NITERS_KNOWN_P (loop_vinfo
)
1522 && LOOP_VINFO_INT_NITERS (loop_vinfo
) <= th
)
1524 if (dump_enabled_p ())
1525 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
1526 "not vectorized: vectorization not profitable.");
1527 if (dump_enabled_p ())
1528 dump_printf_loc (MSG_NOTE
, vect_location
,
1529 "not vectorized: iteration count smaller than user "
1530 "specified loop bound parameter or minimum profitable "
1531 "iterations (whichever is more conservative).");
1535 if ((estimated_niter
= estimated_stmt_executions_int (loop
)) != -1
1536 && ((unsigned HOST_WIDE_INT
) estimated_niter
1537 <= MAX (th
, (unsigned)min_profitable_estimate
)))
1539 if (dump_enabled_p ())
1540 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
1541 "not vectorized: estimated iteration count too "
1543 if (dump_enabled_p ())
1544 dump_printf_loc (MSG_NOTE
, vect_location
,
1545 "not vectorized: estimated iteration count smaller "
1546 "than specified loop bound parameter or minimum "
1547 "profitable iterations (whichever is more "
1552 if (!LOOP_VINFO_NITERS_KNOWN_P (loop_vinfo
)
1553 || LOOP_VINFO_INT_NITERS (loop_vinfo
) % vectorization_factor
!= 0
1554 || LOOP_PEELING_FOR_ALIGNMENT (loop_vinfo
))
1556 if (dump_enabled_p ())
1557 dump_printf_loc (MSG_NOTE
, vect_location
, "epilog loop required.");
1558 if (!vect_can_advance_ivs_p (loop_vinfo
))
1560 if (dump_enabled_p ())
1561 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
1562 "not vectorized: can't create epilog loop 1.");
1565 if (!slpeel_can_duplicate_loop_p (loop
, single_exit (loop
)))
1567 if (dump_enabled_p ())
1568 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
1569 "not vectorized: can't create epilog loop 2.");
1578 /* Function vect_analyze_loop_2.
1580 Apply a set of analyses on LOOP, and create a loop_vec_info struct
1581 for it. The different analyses will record information in the
1582 loop_vec_info struct. */
1584 vect_analyze_loop_2 (loop_vec_info loop_vinfo
)
1586 bool ok
, slp
= false;
1587 int max_vf
= MAX_VECTORIZATION_FACTOR
;
1590 /* Find all data references in the loop (which correspond to vdefs/vuses)
1591 and analyze their evolution in the loop. Also adjust the minimal
1592 vectorization factor according to the loads and stores.
1594 FORNOW: Handle only simple, array references, which
1595 alignment can be forced, and aligned pointer-references. */
1597 ok
= vect_analyze_data_refs (loop_vinfo
, NULL
, &min_vf
);
1600 if (dump_enabled_p ())
1601 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
1602 "bad data references.");
1606 /* Classify all cross-iteration scalar data-flow cycles.
1607 Cross-iteration cycles caused by virtual phis are analyzed separately. */
1609 vect_analyze_scalar_cycles (loop_vinfo
);
1611 vect_pattern_recog (loop_vinfo
, NULL
);
1613 /* Data-flow analysis to detect stmts that do not need to be vectorized. */
1615 ok
= vect_mark_stmts_to_be_vectorized (loop_vinfo
);
1618 if (dump_enabled_p ())
1619 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
1620 "unexpected pattern.");
1624 /* Analyze data dependences between the data-refs in the loop
1625 and adjust the maximum vectorization factor according to
1627 FORNOW: fail at the first data dependence that we encounter. */
1629 ok
= vect_analyze_data_ref_dependences (loop_vinfo
, NULL
, &max_vf
);
1633 if (dump_enabled_p ())
1634 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
1635 "bad data dependence.");
1639 ok
= vect_determine_vectorization_factor (loop_vinfo
);
1642 if (dump_enabled_p ())
1643 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
1644 "can't determine vectorization factor.");
1647 if (max_vf
< LOOP_VINFO_VECT_FACTOR (loop_vinfo
))
1649 if (dump_enabled_p ())
1650 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
1651 "bad data dependence.");
1655 /* Analyze the alignment of the data-refs in the loop.
1656 Fail if a data reference is found that cannot be vectorized. */
1658 ok
= vect_analyze_data_refs_alignment (loop_vinfo
, NULL
);
1661 if (dump_enabled_p ())
1662 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
1663 "bad data alignment.");
1667 /* Analyze the access patterns of the data-refs in the loop (consecutive,
1668 complex, etc.). FORNOW: Only handle consecutive access pattern. */
1670 ok
= vect_analyze_data_ref_accesses (loop_vinfo
, NULL
);
1673 if (dump_enabled_p ())
1674 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
1675 "bad data access.");
1679 /* Prune the list of ddrs to be tested at run-time by versioning for alias.
1680 It is important to call pruning after vect_analyze_data_ref_accesses,
1681 since we use grouping information gathered by interleaving analysis. */
1682 ok
= vect_prune_runtime_alias_test_list (loop_vinfo
);
1685 if (dump_enabled_p ())
1686 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
1687 "too long list of versioning for alias "
1692 /* This pass will decide on using loop versioning and/or loop peeling in
1693 order to enhance the alignment of data references in the loop. */
1695 ok
= vect_enhance_data_refs_alignment (loop_vinfo
);
1698 if (dump_enabled_p ())
1699 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
1700 "bad data alignment.");
1704 /* Check the SLP opportunities in the loop, analyze and build SLP trees. */
1705 ok
= vect_analyze_slp (loop_vinfo
, NULL
);
1708 /* Decide which possible SLP instances to SLP. */
1709 slp
= vect_make_slp_decision (loop_vinfo
);
1711 /* Find stmts that need to be both vectorized and SLPed. */
1712 vect_detect_hybrid_slp (loop_vinfo
);
1717 /* Scan all the operations in the loop and make sure they are
1720 ok
= vect_analyze_loop_operations (loop_vinfo
, slp
);
1723 if (dump_enabled_p ())
1724 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
1725 "bad operation or unsupported loop bound.");
1732 /* Function vect_analyze_loop.
1734 Apply a set of analyses on LOOP, and create a loop_vec_info struct
1735 for it. The different analyses will record information in the
1736 loop_vec_info struct. */
1738 vect_analyze_loop (struct loop
*loop
)
1740 loop_vec_info loop_vinfo
;
1741 unsigned int vector_sizes
;
1743 /* Autodetect first vector size we try. */
1744 current_vector_size
= 0;
1745 vector_sizes
= targetm
.vectorize
.autovectorize_vector_sizes ();
1747 if (dump_enabled_p ())
1748 dump_printf_loc (MSG_NOTE
, vect_location
,
1749 "===== analyze_loop_nest =====");
1751 if (loop_outer (loop
)
1752 && loop_vec_info_for_loop (loop_outer (loop
))
1753 && LOOP_VINFO_VECTORIZABLE_P (loop_vec_info_for_loop (loop_outer (loop
))))
1755 if (dump_enabled_p ())
1756 dump_printf_loc (MSG_NOTE
, vect_location
,
1757 "outer-loop already vectorized.");
1763 /* Check the CFG characteristics of the loop (nesting, entry/exit). */
1764 loop_vinfo
= vect_analyze_loop_form (loop
);
1767 if (dump_enabled_p ())
1768 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
1773 if (vect_analyze_loop_2 (loop_vinfo
))
1775 LOOP_VINFO_VECTORIZABLE_P (loop_vinfo
) = 1;
1780 destroy_loop_vec_info (loop_vinfo
, true);
1782 vector_sizes
&= ~current_vector_size
;
1783 if (vector_sizes
== 0
1784 || current_vector_size
== 0)
1787 /* Try the next biggest vector size. */
1788 current_vector_size
= 1 << floor_log2 (vector_sizes
);
1789 if (dump_enabled_p ())
1790 dump_printf_loc (MSG_NOTE
, vect_location
,
1791 "***** Re-trying analysis with "
1792 "vector size %d\n", current_vector_size
);
1797 /* Function reduction_code_for_scalar_code
1800 CODE - tree_code of a reduction operations.
1803 REDUC_CODE - the corresponding tree-code to be used to reduce the
1804 vector of partial results into a single scalar result (which
1805 will also reside in a vector) or ERROR_MARK if the operation is
1806 a supported reduction operation, but does not have such tree-code.
1808 Return FALSE if CODE currently cannot be vectorized as reduction. */
1811 reduction_code_for_scalar_code (enum tree_code code
,
1812 enum tree_code
*reduc_code
)
1817 *reduc_code
= REDUC_MAX_EXPR
;
1821 *reduc_code
= REDUC_MIN_EXPR
;
1825 *reduc_code
= REDUC_PLUS_EXPR
;
1833 *reduc_code
= ERROR_MARK
;
1842 /* Error reporting helper for vect_is_simple_reduction below. GIMPLE statement
1843 STMT is printed with a message MSG. */
1846 report_vect_op (int msg_type
, gimple stmt
, const char *msg
)
1848 dump_printf_loc (msg_type
, vect_location
, "%s", msg
);
1849 dump_gimple_stmt (msg_type
, TDF_SLIM
, stmt
, 0);
1853 /* Detect SLP reduction of the form:
1863 PHI is the reduction phi node (#a1 = phi <a5, a0> above)
1864 FIRST_STMT is the first reduction stmt in the chain
1865 (a2 = operation (a1)).
1867 Return TRUE if a reduction chain was detected. */
1870 vect_is_slp_reduction (loop_vec_info loop_info
, gimple phi
, gimple first_stmt
)
1872 struct loop
*loop
= (gimple_bb (phi
))->loop_father
;
1873 struct loop
*vect_loop
= LOOP_VINFO_LOOP (loop_info
);
1874 enum tree_code code
;
1875 gimple current_stmt
= NULL
, loop_use_stmt
= NULL
, first
, next_stmt
;
1876 stmt_vec_info use_stmt_info
, current_stmt_info
;
1878 imm_use_iterator imm_iter
;
1879 use_operand_p use_p
;
1880 int nloop_uses
, size
= 0, n_out_of_loop_uses
;
1883 if (loop
!= vect_loop
)
1886 lhs
= PHI_RESULT (phi
);
1887 code
= gimple_assign_rhs_code (first_stmt
);
1891 n_out_of_loop_uses
= 0;
1892 FOR_EACH_IMM_USE_FAST (use_p
, imm_iter
, lhs
)
1894 gimple use_stmt
= USE_STMT (use_p
);
1895 if (is_gimple_debug (use_stmt
))
1898 use_stmt
= USE_STMT (use_p
);
1900 /* Check if we got back to the reduction phi. */
1901 if (use_stmt
== phi
)
1903 loop_use_stmt
= use_stmt
;
1908 if (flow_bb_inside_loop_p (loop
, gimple_bb (use_stmt
)))
1910 if (vinfo_for_stmt (use_stmt
)
1911 && !STMT_VINFO_IN_PATTERN_P (vinfo_for_stmt (use_stmt
)))
1913 loop_use_stmt
= use_stmt
;
1918 n_out_of_loop_uses
++;
1920 /* There are can be either a single use in the loop or two uses in
1922 if (nloop_uses
> 1 || (n_out_of_loop_uses
&& nloop_uses
))
1929 /* We reached a statement with no loop uses. */
1930 if (nloop_uses
== 0)
1933 /* This is a loop exit phi, and we haven't reached the reduction phi. */
1934 if (gimple_code (loop_use_stmt
) == GIMPLE_PHI
)
1937 if (!is_gimple_assign (loop_use_stmt
)
1938 || code
!= gimple_assign_rhs_code (loop_use_stmt
)
1939 || !flow_bb_inside_loop_p (loop
, gimple_bb (loop_use_stmt
)))
1942 /* Insert USE_STMT into reduction chain. */
1943 use_stmt_info
= vinfo_for_stmt (loop_use_stmt
);
1946 current_stmt_info
= vinfo_for_stmt (current_stmt
);
1947 GROUP_NEXT_ELEMENT (current_stmt_info
) = loop_use_stmt
;
1948 GROUP_FIRST_ELEMENT (use_stmt_info
)
1949 = GROUP_FIRST_ELEMENT (current_stmt_info
);
1952 GROUP_FIRST_ELEMENT (use_stmt_info
) = loop_use_stmt
;
1954 lhs
= gimple_assign_lhs (loop_use_stmt
);
1955 current_stmt
= loop_use_stmt
;
1959 if (!found
|| loop_use_stmt
!= phi
|| size
< 2)
1962 /* Swap the operands, if needed, to make the reduction operand be the second
1964 lhs
= PHI_RESULT (phi
);
1965 next_stmt
= GROUP_FIRST_ELEMENT (vinfo_for_stmt (current_stmt
));
1968 if (gimple_assign_rhs2 (next_stmt
) == lhs
)
1970 tree op
= gimple_assign_rhs1 (next_stmt
);
1971 gimple def_stmt
= NULL
;
1973 if (TREE_CODE (op
) == SSA_NAME
)
1974 def_stmt
= SSA_NAME_DEF_STMT (op
);
1976 /* Check that the other def is either defined in the loop
1977 ("vect_internal_def"), or it's an induction (defined by a
1978 loop-header phi-node). */
1980 && gimple_bb (def_stmt
)
1981 && flow_bb_inside_loop_p (loop
, gimple_bb (def_stmt
))
1982 && (is_gimple_assign (def_stmt
)
1983 || is_gimple_call (def_stmt
)
1984 || STMT_VINFO_DEF_TYPE (vinfo_for_stmt (def_stmt
))
1985 == vect_induction_def
1986 || (gimple_code (def_stmt
) == GIMPLE_PHI
1987 && STMT_VINFO_DEF_TYPE (vinfo_for_stmt (def_stmt
))
1988 == vect_internal_def
1989 && !is_loop_header_bb_p (gimple_bb (def_stmt
)))))
1991 lhs
= gimple_assign_lhs (next_stmt
);
1992 next_stmt
= GROUP_NEXT_ELEMENT (vinfo_for_stmt (next_stmt
));
2000 tree op
= gimple_assign_rhs2 (next_stmt
);
2001 gimple def_stmt
= NULL
;
2003 if (TREE_CODE (op
) == SSA_NAME
)
2004 def_stmt
= SSA_NAME_DEF_STMT (op
);
2006 /* Check that the other def is either defined in the loop
2007 ("vect_internal_def"), or it's an induction (defined by a
2008 loop-header phi-node). */
2010 && gimple_bb (def_stmt
)
2011 && flow_bb_inside_loop_p (loop
, gimple_bb (def_stmt
))
2012 && (is_gimple_assign (def_stmt
)
2013 || is_gimple_call (def_stmt
)
2014 || STMT_VINFO_DEF_TYPE (vinfo_for_stmt (def_stmt
))
2015 == vect_induction_def
2016 || (gimple_code (def_stmt
) == GIMPLE_PHI
2017 && STMT_VINFO_DEF_TYPE (vinfo_for_stmt (def_stmt
))
2018 == vect_internal_def
2019 && !is_loop_header_bb_p (gimple_bb (def_stmt
)))))
2021 if (dump_enabled_p ())
2023 dump_printf_loc (MSG_NOTE
, vect_location
, "swapping oprnds: ");
2024 dump_gimple_stmt (MSG_NOTE
, TDF_SLIM
, next_stmt
, 0);
2027 swap_tree_operands (next_stmt
,
2028 gimple_assign_rhs1_ptr (next_stmt
),
2029 gimple_assign_rhs2_ptr (next_stmt
));
2030 update_stmt (next_stmt
);
2032 if (CONSTANT_CLASS_P (gimple_assign_rhs1 (next_stmt
)))
2033 LOOP_VINFO_OPERANDS_SWAPPED (loop_info
) = true;
2039 lhs
= gimple_assign_lhs (next_stmt
);
2040 next_stmt
= GROUP_NEXT_ELEMENT (vinfo_for_stmt (next_stmt
));
2043 /* Save the chain for further analysis in SLP detection. */
2044 first
= GROUP_FIRST_ELEMENT (vinfo_for_stmt (current_stmt
));
2045 LOOP_VINFO_REDUCTION_CHAINS (loop_info
).safe_push (first
);
2046 GROUP_SIZE (vinfo_for_stmt (first
)) = size
;
2052 /* Function vect_is_simple_reduction_1
2054 (1) Detect a cross-iteration def-use cycle that represents a simple
2055 reduction computation. We look for the following pattern:
2060 a2 = operation (a3, a1)
2063 1. operation is commutative and associative and it is safe to
2064 change the order of the computation (if CHECK_REDUCTION is true)
2065 2. no uses for a2 in the loop (a2 is used out of the loop)
2066 3. no uses of a1 in the loop besides the reduction operation
2067 4. no uses of a1 outside the loop.
2069 Conditions 1,4 are tested here.
2070 Conditions 2,3 are tested in vect_mark_stmts_to_be_vectorized.
2072 (2) Detect a cross-iteration def-use cycle in nested loops, i.e.,
2073 nested cycles, if CHECK_REDUCTION is false.
2075 (3) Detect cycles of phi nodes in outer-loop vectorization, i.e., double
2079 inner loop (def of a3)
2082 If MODIFY is true it tries also to rework the code in-place to enable
2083 detection of more reduction patterns. For the time being we rewrite
2084 "res -= RHS" into "rhs += -RHS" when it seems worthwhile.
2088 vect_is_simple_reduction_1 (loop_vec_info loop_info
, gimple phi
,
2089 bool check_reduction
, bool *double_reduc
,
2092 struct loop
*loop
= (gimple_bb (phi
))->loop_father
;
2093 struct loop
*vect_loop
= LOOP_VINFO_LOOP (loop_info
);
2094 edge latch_e
= loop_latch_edge (loop
);
2095 tree loop_arg
= PHI_ARG_DEF_FROM_EDGE (phi
, latch_e
);
2096 gimple def_stmt
, def1
= NULL
, def2
= NULL
;
2097 enum tree_code orig_code
, code
;
2098 tree op1
, op2
, op3
= NULL_TREE
, op4
= NULL_TREE
;
2102 imm_use_iterator imm_iter
;
2103 use_operand_p use_p
;
2106 *double_reduc
= false;
2108 /* If CHECK_REDUCTION is true, we assume inner-most loop vectorization,
2109 otherwise, we assume outer loop vectorization. */
2110 gcc_assert ((check_reduction
&& loop
== vect_loop
)
2111 || (!check_reduction
&& flow_loop_nested_p (vect_loop
, loop
)));
2113 name
= PHI_RESULT (phi
);
2115 FOR_EACH_IMM_USE_FAST (use_p
, imm_iter
, name
)
2117 gimple use_stmt
= USE_STMT (use_p
);
2118 if (is_gimple_debug (use_stmt
))
2121 if (!flow_bb_inside_loop_p (loop
, gimple_bb (use_stmt
)))
2123 if (dump_enabled_p ())
2124 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
2125 "intermediate value used outside loop.");
2130 if (vinfo_for_stmt (use_stmt
)
2131 && !is_pattern_stmt_p (vinfo_for_stmt (use_stmt
)))
2135 if (dump_enabled_p ())
2136 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
2137 "reduction used in loop.");
2142 if (TREE_CODE (loop_arg
) != SSA_NAME
)
2144 if (dump_enabled_p ())
2146 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
2147 "reduction: not ssa_name: ");
2148 dump_generic_expr (MSG_MISSED_OPTIMIZATION
, TDF_SLIM
, loop_arg
);
2153 def_stmt
= SSA_NAME_DEF_STMT (loop_arg
);
2156 if (dump_enabled_p ())
2157 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
2158 "reduction: no def_stmt.");
2162 if (!is_gimple_assign (def_stmt
) && gimple_code (def_stmt
) != GIMPLE_PHI
)
2164 if (dump_enabled_p ())
2165 dump_gimple_stmt (MSG_NOTE
, TDF_SLIM
, def_stmt
, 0);
2169 if (is_gimple_assign (def_stmt
))
2171 name
= gimple_assign_lhs (def_stmt
);
2176 name
= PHI_RESULT (def_stmt
);
2181 FOR_EACH_IMM_USE_FAST (use_p
, imm_iter
, name
)
2183 gimple use_stmt
= USE_STMT (use_p
);
2184 if (is_gimple_debug (use_stmt
))
2186 if (flow_bb_inside_loop_p (loop
, gimple_bb (use_stmt
))
2187 && vinfo_for_stmt (use_stmt
)
2188 && !is_pattern_stmt_p (vinfo_for_stmt (use_stmt
)))
2192 if (dump_enabled_p ())
2193 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
2194 "reduction used in loop.");
2199 /* If DEF_STMT is a phi node itself, we expect it to have a single argument
2200 defined in the inner loop. */
2203 op1
= PHI_ARG_DEF (def_stmt
, 0);
2205 if (gimple_phi_num_args (def_stmt
) != 1
2206 || TREE_CODE (op1
) != SSA_NAME
)
2208 if (dump_enabled_p ())
2209 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
2210 "unsupported phi node definition.");
2215 def1
= SSA_NAME_DEF_STMT (op1
);
2216 if (flow_bb_inside_loop_p (loop
, gimple_bb (def_stmt
))
2218 && flow_bb_inside_loop_p (loop
->inner
, gimple_bb (def1
))
2219 && is_gimple_assign (def1
))
2221 if (dump_enabled_p ())
2222 report_vect_op (MSG_NOTE
, def_stmt
,
2223 "detected double reduction: ");
2225 *double_reduc
= true;
2232 code
= orig_code
= gimple_assign_rhs_code (def_stmt
);
2234 /* We can handle "res -= x[i]", which is non-associative by
2235 simply rewriting this into "res += -x[i]". Avoid changing
2236 gimple instruction for the first simple tests and only do this
2237 if we're allowed to change code at all. */
2238 if (code
== MINUS_EXPR
2240 && (op1
= gimple_assign_rhs1 (def_stmt
))
2241 && TREE_CODE (op1
) == SSA_NAME
2242 && SSA_NAME_DEF_STMT (op1
) == phi
)
2246 && (!commutative_tree_code (code
) || !associative_tree_code (code
)))
2248 if (dump_enabled_p ())
2249 report_vect_op (MSG_MISSED_OPTIMIZATION
, def_stmt
,
2250 "reduction: not commutative/associative: ");
2254 if (get_gimple_rhs_class (code
) != GIMPLE_BINARY_RHS
)
2256 if (code
!= COND_EXPR
)
2258 if (dump_enabled_p ())
2259 report_vect_op (MSG_MISSED_OPTIMIZATION
, def_stmt
,
2260 "reduction: not binary operation: ");
2265 op3
= gimple_assign_rhs1 (def_stmt
);
2266 if (COMPARISON_CLASS_P (op3
))
2268 op4
= TREE_OPERAND (op3
, 1);
2269 op3
= TREE_OPERAND (op3
, 0);
2272 op1
= gimple_assign_rhs2 (def_stmt
);
2273 op2
= gimple_assign_rhs3 (def_stmt
);
2275 if (TREE_CODE (op1
) != SSA_NAME
&& TREE_CODE (op2
) != SSA_NAME
)
2277 if (dump_enabled_p ())
2278 report_vect_op (MSG_MISSED_OPTIMIZATION
, def_stmt
,
2279 "reduction: uses not ssa_names: ");
2286 op1
= gimple_assign_rhs1 (def_stmt
);
2287 op2
= gimple_assign_rhs2 (def_stmt
);
2289 if (TREE_CODE (op1
) != SSA_NAME
&& TREE_CODE (op2
) != SSA_NAME
)
2291 if (dump_enabled_p ())
2292 report_vect_op (MSG_MISSED_OPTIMIZATION
, def_stmt
,
2293 "reduction: uses not ssa_names: ");
2299 type
= TREE_TYPE (gimple_assign_lhs (def_stmt
));
2300 if ((TREE_CODE (op1
) == SSA_NAME
2301 && !types_compatible_p (type
,TREE_TYPE (op1
)))
2302 || (TREE_CODE (op2
) == SSA_NAME
2303 && !types_compatible_p (type
, TREE_TYPE (op2
)))
2304 || (op3
&& TREE_CODE (op3
) == SSA_NAME
2305 && !types_compatible_p (type
, TREE_TYPE (op3
)))
2306 || (op4
&& TREE_CODE (op4
) == SSA_NAME
2307 && !types_compatible_p (type
, TREE_TYPE (op4
))))
2309 if (dump_enabled_p ())
2311 dump_printf_loc (MSG_NOTE
, vect_location
,
2312 "reduction: multiple types: operation type: ");
2313 dump_generic_expr (MSG_NOTE
, TDF_SLIM
, type
);
2314 dump_printf (MSG_NOTE
, ", operands types: ");
2315 dump_generic_expr (MSG_NOTE
, TDF_SLIM
,
2317 dump_printf (MSG_NOTE
, ",");
2318 dump_generic_expr (MSG_NOTE
, TDF_SLIM
,
2322 dump_printf (MSG_NOTE
, ",");
2323 dump_generic_expr (MSG_NOTE
, TDF_SLIM
,
2329 dump_printf (MSG_NOTE
, ",");
2330 dump_generic_expr (MSG_NOTE
, TDF_SLIM
,
2338 /* Check that it's ok to change the order of the computation.
2339 Generally, when vectorizing a reduction we change the order of the
2340 computation. This may change the behavior of the program in some
2341 cases, so we need to check that this is ok. One exception is when
2342 vectorizing an outer-loop: the inner-loop is executed sequentially,
2343 and therefore vectorizing reductions in the inner-loop during
2344 outer-loop vectorization is safe. */
2346 /* CHECKME: check for !flag_finite_math_only too? */
2347 if (SCALAR_FLOAT_TYPE_P (type
) && !flag_associative_math
2350 /* Changing the order of operations changes the semantics. */
2351 if (dump_enabled_p ())
2352 report_vect_op (MSG_MISSED_OPTIMIZATION
, def_stmt
,
2353 "reduction: unsafe fp math optimization: ");
2356 else if (INTEGRAL_TYPE_P (type
) && TYPE_OVERFLOW_TRAPS (type
)
2359 /* Changing the order of operations changes the semantics. */
2360 if (dump_enabled_p ())
2361 report_vect_op (MSG_MISSED_OPTIMIZATION
, def_stmt
,
2362 "reduction: unsafe int math optimization: ");
2365 else if (SAT_FIXED_POINT_TYPE_P (type
) && check_reduction
)
2367 /* Changing the order of operations changes the semantics. */
2368 if (dump_enabled_p ())
2369 report_vect_op (MSG_MISSED_OPTIMIZATION
, def_stmt
,
2370 "reduction: unsafe fixed-point math optimization: ");
2374 /* If we detected "res -= x[i]" earlier, rewrite it into
2375 "res += -x[i]" now. If this turns out to be useless reassoc
2376 will clean it up again. */
2377 if (orig_code
== MINUS_EXPR
)
2379 tree rhs
= gimple_assign_rhs2 (def_stmt
);
2380 tree negrhs
= make_ssa_name (TREE_TYPE (rhs
), NULL
);
2381 gimple negate_stmt
= gimple_build_assign_with_ops (NEGATE_EXPR
, negrhs
,
2383 gimple_stmt_iterator gsi
= gsi_for_stmt (def_stmt
);
2384 set_vinfo_for_stmt (negate_stmt
, new_stmt_vec_info (negate_stmt
,
2386 gsi_insert_before (&gsi
, negate_stmt
, GSI_NEW_STMT
);
2387 gimple_assign_set_rhs2 (def_stmt
, negrhs
);
2388 gimple_assign_set_rhs_code (def_stmt
, PLUS_EXPR
);
2389 update_stmt (def_stmt
);
2392 /* Reduction is safe. We're dealing with one of the following:
2393 1) integer arithmetic and no trapv
2394 2) floating point arithmetic, and special flags permit this optimization
2395 3) nested cycle (i.e., outer loop vectorization). */
2396 if (TREE_CODE (op1
) == SSA_NAME
)
2397 def1
= SSA_NAME_DEF_STMT (op1
);
2399 if (TREE_CODE (op2
) == SSA_NAME
)
2400 def2
= SSA_NAME_DEF_STMT (op2
);
2402 if (code
!= COND_EXPR
2403 && ((!def1
|| gimple_nop_p (def1
)) && (!def2
|| gimple_nop_p (def2
))))
2405 if (dump_enabled_p ())
2406 report_vect_op (MSG_NOTE
, def_stmt
, "reduction: no defs for operands: ");
2410 /* Check that one def is the reduction def, defined by PHI,
2411 the other def is either defined in the loop ("vect_internal_def"),
2412 or it's an induction (defined by a loop-header phi-node). */
2414 if (def2
&& def2
== phi
2415 && (code
== COND_EXPR
2416 || !def1
|| gimple_nop_p (def1
)
2417 || (def1
&& flow_bb_inside_loop_p (loop
, gimple_bb (def1
))
2418 && (is_gimple_assign (def1
)
2419 || is_gimple_call (def1
)
2420 || STMT_VINFO_DEF_TYPE (vinfo_for_stmt (def1
))
2421 == vect_induction_def
2422 || (gimple_code (def1
) == GIMPLE_PHI
2423 && STMT_VINFO_DEF_TYPE (vinfo_for_stmt (def1
))
2424 == vect_internal_def
2425 && !is_loop_header_bb_p (gimple_bb (def1
)))))))
2427 if (dump_enabled_p ())
2428 report_vect_op (MSG_NOTE
, def_stmt
, "detected reduction: ");
2432 if (def1
&& def1
== phi
2433 && (code
== COND_EXPR
2434 || !def2
|| gimple_nop_p (def2
)
2435 || (def2
&& flow_bb_inside_loop_p (loop
, gimple_bb (def2
))
2436 && (is_gimple_assign (def2
)
2437 || is_gimple_call (def2
)
2438 || STMT_VINFO_DEF_TYPE (vinfo_for_stmt (def2
))
2439 == vect_induction_def
2440 || (gimple_code (def2
) == GIMPLE_PHI
2441 && STMT_VINFO_DEF_TYPE (vinfo_for_stmt (def2
))
2442 == vect_internal_def
2443 && !is_loop_header_bb_p (gimple_bb (def2
)))))))
2445 if (check_reduction
)
2447 /* Swap operands (just for simplicity - so that the rest of the code
2448 can assume that the reduction variable is always the last (second)
2450 if (dump_enabled_p ())
2451 report_vect_op (MSG_NOTE
, def_stmt
,
2452 "detected reduction: need to swap operands: ");
2454 swap_tree_operands (def_stmt
, gimple_assign_rhs1_ptr (def_stmt
),
2455 gimple_assign_rhs2_ptr (def_stmt
));
2457 if (CONSTANT_CLASS_P (gimple_assign_rhs1 (def_stmt
)))
2458 LOOP_VINFO_OPERANDS_SWAPPED (loop_info
) = true;
2462 if (dump_enabled_p ())
2463 report_vect_op (MSG_NOTE
, def_stmt
, "detected reduction: ");
2469 /* Try to find SLP reduction chain. */
2470 if (check_reduction
&& vect_is_slp_reduction (loop_info
, phi
, def_stmt
))
2472 if (dump_enabled_p ())
2473 report_vect_op (MSG_NOTE
, def_stmt
,
2474 "reduction: detected reduction chain: ");
2479 if (dump_enabled_p ())
2480 report_vect_op (MSG_MISSED_OPTIMIZATION
, def_stmt
,
2481 "reduction: unknown pattern: ");
2486 /* Wrapper around vect_is_simple_reduction_1, that won't modify code
2487 in-place. Arguments as there. */
2490 vect_is_simple_reduction (loop_vec_info loop_info
, gimple phi
,
2491 bool check_reduction
, bool *double_reduc
)
2493 return vect_is_simple_reduction_1 (loop_info
, phi
, check_reduction
,
2494 double_reduc
, false);
2497 /* Wrapper around vect_is_simple_reduction_1, which will modify code
2498 in-place if it enables detection of more reductions. Arguments
2502 vect_force_simple_reduction (loop_vec_info loop_info
, gimple phi
,
2503 bool check_reduction
, bool *double_reduc
)
2505 return vect_is_simple_reduction_1 (loop_info
, phi
, check_reduction
,
2506 double_reduc
, true);
2509 /* Calculate the cost of one scalar iteration of the loop. */
2511 vect_get_single_scalar_iteration_cost (loop_vec_info loop_vinfo
)
2513 struct loop
*loop
= LOOP_VINFO_LOOP (loop_vinfo
);
2514 basic_block
*bbs
= LOOP_VINFO_BBS (loop_vinfo
);
2515 int nbbs
= loop
->num_nodes
, factor
, scalar_single_iter_cost
= 0;
2516 int innerloop_iters
, i
, stmt_cost
;
2518 /* Count statements in scalar loop. Using this as scalar cost for a single
2521 TODO: Add outer loop support.
2523 TODO: Consider assigning different costs to different scalar
2527 innerloop_iters
= 1;
2529 innerloop_iters
= 50; /* FIXME */
2531 for (i
= 0; i
< nbbs
; i
++)
2533 gimple_stmt_iterator si
;
2534 basic_block bb
= bbs
[i
];
2536 if (bb
->loop_father
== loop
->inner
)
2537 factor
= innerloop_iters
;
2541 for (si
= gsi_start_bb (bb
); !gsi_end_p (si
); gsi_next (&si
))
2543 gimple stmt
= gsi_stmt (si
);
2544 stmt_vec_info stmt_info
= vinfo_for_stmt (stmt
);
2546 if (!is_gimple_assign (stmt
) && !is_gimple_call (stmt
))
2549 /* Skip stmts that are not vectorized inside the loop. */
2551 && !STMT_VINFO_RELEVANT_P (stmt_info
)
2552 && (!STMT_VINFO_LIVE_P (stmt_info
)
2553 || !VECTORIZABLE_CYCLE_DEF (STMT_VINFO_DEF_TYPE (stmt_info
)))
2554 && !STMT_VINFO_IN_PATTERN_P (stmt_info
))
2557 if (STMT_VINFO_DATA_REF (vinfo_for_stmt (stmt
)))
2559 if (DR_IS_READ (STMT_VINFO_DATA_REF (vinfo_for_stmt (stmt
))))
2560 stmt_cost
= vect_get_stmt_cost (scalar_load
);
2562 stmt_cost
= vect_get_stmt_cost (scalar_store
);
2565 stmt_cost
= vect_get_stmt_cost (scalar_stmt
);
2567 scalar_single_iter_cost
+= stmt_cost
* factor
;
2570 return scalar_single_iter_cost
;
2573 /* Calculate cost of peeling the loop PEEL_ITERS_PROLOGUE times. */
2575 vect_get_known_peeling_cost (loop_vec_info loop_vinfo
, int peel_iters_prologue
,
2576 int *peel_iters_epilogue
,
2577 int scalar_single_iter_cost
,
2578 stmt_vector_for_cost
*prologue_cost_vec
,
2579 stmt_vector_for_cost
*epilogue_cost_vec
)
2582 int vf
= LOOP_VINFO_VECT_FACTOR (loop_vinfo
);
2584 if (!LOOP_VINFO_NITERS_KNOWN_P (loop_vinfo
))
2586 *peel_iters_epilogue
= vf
/2;
2587 if (dump_enabled_p ())
2588 dump_printf_loc (MSG_NOTE
, vect_location
,
2589 "cost model: epilogue peel iters set to vf/2 "
2590 "because loop iterations are unknown .");
2592 /* If peeled iterations are known but number of scalar loop
2593 iterations are unknown, count a taken branch per peeled loop. */
2594 retval
= record_stmt_cost (prologue_cost_vec
, 2, cond_branch_taken
,
2595 NULL
, 0, vect_prologue
);
2599 int niters
= LOOP_VINFO_INT_NITERS (loop_vinfo
);
2600 peel_iters_prologue
= niters
< peel_iters_prologue
?
2601 niters
: peel_iters_prologue
;
2602 *peel_iters_epilogue
= (niters
- peel_iters_prologue
) % vf
;
2603 /* If we need to peel for gaps, but no peeling is required, we have to
2604 peel VF iterations. */
2605 if (LOOP_VINFO_PEELING_FOR_GAPS (loop_vinfo
) && !*peel_iters_epilogue
)
2606 *peel_iters_epilogue
= vf
;
2609 if (peel_iters_prologue
)
2610 retval
+= record_stmt_cost (prologue_cost_vec
,
2611 peel_iters_prologue
* scalar_single_iter_cost
,
2612 scalar_stmt
, NULL
, 0, vect_prologue
);
2613 if (*peel_iters_epilogue
)
2614 retval
+= record_stmt_cost (epilogue_cost_vec
,
2615 *peel_iters_epilogue
* scalar_single_iter_cost
,
2616 scalar_stmt
, NULL
, 0, vect_epilogue
);
2620 /* Function vect_estimate_min_profitable_iters
2622 Return the number of iterations required for the vector version of the
2623 loop to be profitable relative to the cost of the scalar version of the
2627 vect_estimate_min_profitable_iters (loop_vec_info loop_vinfo
,
2628 int *ret_min_profitable_niters
,
2629 int *ret_min_profitable_estimate
)
2631 int min_profitable_iters
;
2632 int min_profitable_estimate
;
2633 int peel_iters_prologue
;
2634 int peel_iters_epilogue
;
2635 unsigned vec_inside_cost
= 0;
2636 int vec_outside_cost
= 0;
2637 unsigned vec_prologue_cost
= 0;
2638 unsigned vec_epilogue_cost
= 0;
2639 int scalar_single_iter_cost
= 0;
2640 int scalar_outside_cost
= 0;
2641 int vf
= LOOP_VINFO_VECT_FACTOR (loop_vinfo
);
2642 int npeel
= LOOP_PEELING_FOR_ALIGNMENT (loop_vinfo
);
2643 void *target_cost_data
= LOOP_VINFO_TARGET_COST_DATA (loop_vinfo
);
2645 /* Cost model disabled. */
2646 if (!flag_vect_cost_model
)
2648 dump_printf_loc (MSG_NOTE
, vect_location
, "cost model disabled.");
2649 *ret_min_profitable_niters
= 0;
2650 *ret_min_profitable_estimate
= 0;
2654 /* Requires loop versioning tests to handle misalignment. */
2655 if (LOOP_REQUIRES_VERSIONING_FOR_ALIGNMENT (loop_vinfo
))
2657 /* FIXME: Make cost depend on complexity of individual check. */
2658 unsigned len
= LOOP_VINFO_MAY_MISALIGN_STMTS (loop_vinfo
).length ();
2659 (void) add_stmt_cost (target_cost_data
, len
, vector_stmt
, NULL
, 0,
2661 dump_printf (MSG_NOTE
,
2662 "cost model: Adding cost of checks for loop "
2663 "versioning to treat misalignment.\n");
2666 /* Requires loop versioning with alias checks. */
2667 if (LOOP_REQUIRES_VERSIONING_FOR_ALIAS (loop_vinfo
))
2669 /* FIXME: Make cost depend on complexity of individual check. */
2670 unsigned len
= LOOP_VINFO_MAY_ALIAS_DDRS (loop_vinfo
).length ();
2671 (void) add_stmt_cost (target_cost_data
, len
, vector_stmt
, NULL
, 0,
2673 dump_printf (MSG_NOTE
,
2674 "cost model: Adding cost of checks for loop "
2675 "versioning aliasing.\n");
2678 if (LOOP_REQUIRES_VERSIONING_FOR_ALIGNMENT (loop_vinfo
)
2679 || LOOP_REQUIRES_VERSIONING_FOR_ALIAS (loop_vinfo
))
2680 (void) add_stmt_cost (target_cost_data
, 1, cond_branch_taken
, NULL
, 0,
2683 /* Count statements in scalar loop. Using this as scalar cost for a single
2686 TODO: Add outer loop support.
2688 TODO: Consider assigning different costs to different scalar
2691 scalar_single_iter_cost
= vect_get_single_scalar_iteration_cost (loop_vinfo
);
2693 /* Add additional cost for the peeled instructions in prologue and epilogue
2696 FORNOW: If we don't know the value of peel_iters for prologue or epilogue
2697 at compile-time - we assume it's vf/2 (the worst would be vf-1).
2699 TODO: Build an expression that represents peel_iters for prologue and
2700 epilogue to be used in a run-time test. */
2704 peel_iters_prologue
= vf
/2;
2705 dump_printf (MSG_NOTE
, "cost model: "
2706 "prologue peel iters set to vf/2.");
2708 /* If peeling for alignment is unknown, loop bound of main loop becomes
2710 peel_iters_epilogue
= vf
/2;
2711 dump_printf (MSG_NOTE
, "cost model: "
2712 "epilogue peel iters set to vf/2 because "
2713 "peeling for alignment is unknown.");
2715 /* If peeled iterations are unknown, count a taken branch and a not taken
2716 branch per peeled loop. Even if scalar loop iterations are known,
2717 vector iterations are not known since peeled prologue iterations are
2718 not known. Hence guards remain the same. */
2719 (void) add_stmt_cost (target_cost_data
, 2, cond_branch_taken
,
2720 NULL
, 0, vect_prologue
);
2721 (void) add_stmt_cost (target_cost_data
, 2, cond_branch_not_taken
,
2722 NULL
, 0, vect_prologue
);
2723 /* FORNOW: Don't attempt to pass individual scalar instructions to
2724 the model; just assume linear cost for scalar iterations. */
2725 (void) add_stmt_cost (target_cost_data
,
2726 peel_iters_prologue
* scalar_single_iter_cost
,
2727 scalar_stmt
, NULL
, 0, vect_prologue
);
2728 (void) add_stmt_cost (target_cost_data
,
2729 peel_iters_epilogue
* scalar_single_iter_cost
,
2730 scalar_stmt
, NULL
, 0, vect_epilogue
);
2734 stmt_vector_for_cost prologue_cost_vec
, epilogue_cost_vec
;
2735 stmt_info_for_cost
*si
;
2737 void *data
= LOOP_VINFO_TARGET_COST_DATA (loop_vinfo
);
2739 prologue_cost_vec
.create (2);
2740 epilogue_cost_vec
.create (2);
2741 peel_iters_prologue
= npeel
;
2743 (void) vect_get_known_peeling_cost (loop_vinfo
, peel_iters_prologue
,
2744 &peel_iters_epilogue
,
2745 scalar_single_iter_cost
,
2747 &epilogue_cost_vec
);
2749 FOR_EACH_VEC_ELT (prologue_cost_vec
, j
, si
)
2751 struct _stmt_vec_info
*stmt_info
2752 = si
->stmt
? vinfo_for_stmt (si
->stmt
) : NULL
;
2753 (void) add_stmt_cost (data
, si
->count
, si
->kind
, stmt_info
,
2754 si
->misalign
, vect_prologue
);
2757 FOR_EACH_VEC_ELT (epilogue_cost_vec
, j
, si
)
2759 struct _stmt_vec_info
*stmt_info
2760 = si
->stmt
? vinfo_for_stmt (si
->stmt
) : NULL
;
2761 (void) add_stmt_cost (data
, si
->count
, si
->kind
, stmt_info
,
2762 si
->misalign
, vect_epilogue
);
2765 prologue_cost_vec
.release ();
2766 epilogue_cost_vec
.release ();
2769 /* FORNOW: The scalar outside cost is incremented in one of the
2772 1. The vectorizer checks for alignment and aliasing and generates
2773 a condition that allows dynamic vectorization. A cost model
2774 check is ANDED with the versioning condition. Hence scalar code
2775 path now has the added cost of the versioning check.
2777 if (cost > th & versioning_check)
2780 Hence run-time scalar is incremented by not-taken branch cost.
2782 2. The vectorizer then checks if a prologue is required. If the
2783 cost model check was not done before during versioning, it has to
2784 be done before the prologue check.
2787 prologue = scalar_iters
2792 if (prologue == num_iters)
2795 Hence the run-time scalar cost is incremented by a taken branch,
2796 plus a not-taken branch, plus a taken branch cost.
2798 3. The vectorizer then checks if an epilogue is required. If the
2799 cost model check was not done before during prologue check, it
2800 has to be done with the epilogue check.
2806 if (prologue == num_iters)
2809 if ((cost <= th) | (scalar_iters-prologue-epilogue == 0))
2812 Hence the run-time scalar cost should be incremented by 2 taken
2815 TODO: The back end may reorder the BBS's differently and reverse
2816 conditions/branch directions. Change the estimates below to
2817 something more reasonable. */
2819 /* If the number of iterations is known and we do not do versioning, we can
2820 decide whether to vectorize at compile time. Hence the scalar version
2821 do not carry cost model guard costs. */
2822 if (!LOOP_VINFO_NITERS_KNOWN_P (loop_vinfo
)
2823 || LOOP_REQUIRES_VERSIONING_FOR_ALIGNMENT (loop_vinfo
)
2824 || LOOP_REQUIRES_VERSIONING_FOR_ALIAS (loop_vinfo
))
2826 /* Cost model check occurs at versioning. */
2827 if (LOOP_REQUIRES_VERSIONING_FOR_ALIGNMENT (loop_vinfo
)
2828 || LOOP_REQUIRES_VERSIONING_FOR_ALIAS (loop_vinfo
))
2829 scalar_outside_cost
+= vect_get_stmt_cost (cond_branch_not_taken
);
2832 /* Cost model check occurs at prologue generation. */
2833 if (LOOP_PEELING_FOR_ALIGNMENT (loop_vinfo
) < 0)
2834 scalar_outside_cost
+= 2 * vect_get_stmt_cost (cond_branch_taken
)
2835 + vect_get_stmt_cost (cond_branch_not_taken
);
2836 /* Cost model check occurs at epilogue generation. */
2838 scalar_outside_cost
+= 2 * vect_get_stmt_cost (cond_branch_taken
);
2842 /* Complete the target-specific cost calculations. */
2843 finish_cost (LOOP_VINFO_TARGET_COST_DATA (loop_vinfo
), &vec_prologue_cost
,
2844 &vec_inside_cost
, &vec_epilogue_cost
);
2846 vec_outside_cost
= (int)(vec_prologue_cost
+ vec_epilogue_cost
);
2848 /* Calculate number of iterations required to make the vector version
2849 profitable, relative to the loop bodies only. The following condition
2851 SIC * niters + SOC > VIC * ((niters-PL_ITERS-EP_ITERS)/VF) + VOC
2853 SIC = scalar iteration cost, VIC = vector iteration cost,
2854 VOC = vector outside cost, VF = vectorization factor,
2855 PL_ITERS = prologue iterations, EP_ITERS= epilogue iterations
2856 SOC = scalar outside cost for run time cost model check. */
2858 if ((scalar_single_iter_cost
* vf
) > (int) vec_inside_cost
)
2860 if (vec_outside_cost
<= 0)
2861 min_profitable_iters
= 1;
2864 min_profitable_iters
= ((vec_outside_cost
- scalar_outside_cost
) * vf
2865 - vec_inside_cost
* peel_iters_prologue
2866 - vec_inside_cost
* peel_iters_epilogue
)
2867 / ((scalar_single_iter_cost
* vf
)
2870 if ((scalar_single_iter_cost
* vf
* min_profitable_iters
)
2871 <= (((int) vec_inside_cost
* min_profitable_iters
)
2872 + (((int) vec_outside_cost
- scalar_outside_cost
) * vf
)))
2873 min_profitable_iters
++;
2876 /* vector version will never be profitable. */
2879 if (dump_enabled_p ())
2880 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
2881 "cost model: the vector iteration cost = %d "
2882 "divided by the scalar iteration cost = %d "
2883 "is greater or equal to the vectorization factor = %d.",
2884 vec_inside_cost
, scalar_single_iter_cost
, vf
);
2885 *ret_min_profitable_niters
= -1;
2886 *ret_min_profitable_estimate
= -1;
2890 if (dump_enabled_p ())
2892 dump_printf_loc (MSG_NOTE
, vect_location
, "Cost model analysis: \n");
2893 dump_printf (MSG_NOTE
, " Vector inside of loop cost: %d\n",
2895 dump_printf (MSG_NOTE
, " Vector prologue cost: %d\n",
2897 dump_printf (MSG_NOTE
, " Vector epilogue cost: %d\n",
2899 dump_printf (MSG_NOTE
, " Scalar iteration cost: %d\n",
2900 scalar_single_iter_cost
);
2901 dump_printf (MSG_NOTE
, " Scalar outside cost: %d\n",
2902 scalar_outside_cost
);
2903 dump_printf (MSG_NOTE
, " Vector outside cost: %d\n",
2905 dump_printf (MSG_NOTE
, " prologue iterations: %d\n",
2906 peel_iters_prologue
);
2907 dump_printf (MSG_NOTE
, " epilogue iterations: %d\n",
2908 peel_iters_epilogue
);
2909 dump_printf (MSG_NOTE
,
2910 " Calculated minimum iters for profitability: %d\n",
2911 min_profitable_iters
);
2914 min_profitable_iters
=
2915 min_profitable_iters
< vf
? vf
: min_profitable_iters
;
2917 /* Because the condition we create is:
2918 if (niters <= min_profitable_iters)
2919 then skip the vectorized loop. */
2920 min_profitable_iters
--;
2922 if (dump_enabled_p ())
2923 dump_printf_loc (MSG_NOTE
, vect_location
,
2924 " Runtime profitability threshold = %d\n", min_profitable_iters
);
2926 *ret_min_profitable_niters
= min_profitable_iters
;
2928 /* Calculate number of iterations required to make the vector version
2929 profitable, relative to the loop bodies only.
2931 Non-vectorized variant is SIC * niters and it must win over vector
2932 variant on the expected loop trip count. The following condition must hold true:
2933 SIC * niters > VIC * ((niters-PL_ITERS-EP_ITERS)/VF) + VOC + SOC */
2935 if (vec_outside_cost
<= 0)
2936 min_profitable_estimate
= 1;
2939 min_profitable_estimate
= ((vec_outside_cost
+ scalar_outside_cost
) * vf
2940 - vec_inside_cost
* peel_iters_prologue
2941 - vec_inside_cost
* peel_iters_epilogue
)
2942 / ((scalar_single_iter_cost
* vf
)
2945 min_profitable_estimate
--;
2946 min_profitable_estimate
= MAX (min_profitable_estimate
, min_profitable_iters
);
2947 if (dump_enabled_p ())
2948 dump_printf_loc (MSG_NOTE
, vect_location
,
2949 " Static estimate profitability threshold = %d\n",
2950 min_profitable_iters
);
2952 *ret_min_profitable_estimate
= min_profitable_estimate
;
2956 /* TODO: Close dependency between vect_model_*_cost and vectorizable_*
2957 functions. Design better to avoid maintenance issues. */
2959 /* Function vect_model_reduction_cost.
2961 Models cost for a reduction operation, including the vector ops
2962 generated within the strip-mine loop, the initial definition before
2963 the loop, and the epilogue code that must be generated. */
2966 vect_model_reduction_cost (stmt_vec_info stmt_info
, enum tree_code reduc_code
,
2969 int prologue_cost
= 0, epilogue_cost
= 0;
2970 enum tree_code code
;
2973 gimple stmt
, orig_stmt
;
2975 enum machine_mode mode
;
2976 loop_vec_info loop_vinfo
= STMT_VINFO_LOOP_VINFO (stmt_info
);
2977 struct loop
*loop
= LOOP_VINFO_LOOP (loop_vinfo
);
2978 void *target_cost_data
= LOOP_VINFO_TARGET_COST_DATA (loop_vinfo
);
2980 /* Cost of reduction op inside loop. */
2981 unsigned inside_cost
= add_stmt_cost (target_cost_data
, ncopies
, vector_stmt
,
2982 stmt_info
, 0, vect_body
);
2983 stmt
= STMT_VINFO_STMT (stmt_info
);
2985 switch (get_gimple_rhs_class (gimple_assign_rhs_code (stmt
)))
2987 case GIMPLE_SINGLE_RHS
:
2988 gcc_assert (TREE_OPERAND_LENGTH (gimple_assign_rhs1 (stmt
)) == ternary_op
);
2989 reduction_op
= TREE_OPERAND (gimple_assign_rhs1 (stmt
), 2);
2991 case GIMPLE_UNARY_RHS
:
2992 reduction_op
= gimple_assign_rhs1 (stmt
);
2994 case GIMPLE_BINARY_RHS
:
2995 reduction_op
= gimple_assign_rhs2 (stmt
);
2997 case GIMPLE_TERNARY_RHS
:
2998 reduction_op
= gimple_assign_rhs3 (stmt
);
3004 vectype
= get_vectype_for_scalar_type (TREE_TYPE (reduction_op
));
3007 if (dump_enabled_p ())
3009 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
3010 "unsupported data-type ");
3011 dump_generic_expr (MSG_MISSED_OPTIMIZATION
, TDF_SLIM
,
3012 TREE_TYPE (reduction_op
));
3017 mode
= TYPE_MODE (vectype
);
3018 orig_stmt
= STMT_VINFO_RELATED_STMT (stmt_info
);
3021 orig_stmt
= STMT_VINFO_STMT (stmt_info
);
3023 code
= gimple_assign_rhs_code (orig_stmt
);
3025 /* Add in cost for initial definition. */
3026 prologue_cost
+= add_stmt_cost (target_cost_data
, 1, scalar_to_vec
,
3027 stmt_info
, 0, vect_prologue
);
3029 /* Determine cost of epilogue code.
3031 We have a reduction operator that will reduce the vector in one statement.
3032 Also requires scalar extract. */
3034 if (!nested_in_vect_loop_p (loop
, orig_stmt
))
3036 if (reduc_code
!= ERROR_MARK
)
3038 epilogue_cost
+= add_stmt_cost (target_cost_data
, 1, vector_stmt
,
3039 stmt_info
, 0, vect_epilogue
);
3040 epilogue_cost
+= add_stmt_cost (target_cost_data
, 1, vec_to_scalar
,
3041 stmt_info
, 0, vect_epilogue
);
3045 int vec_size_in_bits
= tree_low_cst (TYPE_SIZE (vectype
), 1);
3047 TYPE_SIZE (TREE_TYPE (gimple_assign_lhs (orig_stmt
)));
3048 int element_bitsize
= tree_low_cst (bitsize
, 1);
3049 int nelements
= vec_size_in_bits
/ element_bitsize
;
3051 optab
= optab_for_tree_code (code
, vectype
, optab_default
);
3053 /* We have a whole vector shift available. */
3054 if (VECTOR_MODE_P (mode
)
3055 && optab_handler (optab
, mode
) != CODE_FOR_nothing
3056 && optab_handler (vec_shr_optab
, mode
) != CODE_FOR_nothing
)
3058 /* Final reduction via vector shifts and the reduction operator.
3059 Also requires scalar extract. */
3060 epilogue_cost
+= add_stmt_cost (target_cost_data
,
3061 exact_log2 (nelements
) * 2,
3062 vector_stmt
, stmt_info
, 0,
3064 epilogue_cost
+= add_stmt_cost (target_cost_data
, 1,
3065 vec_to_scalar
, stmt_info
, 0,
3069 /* Use extracts and reduction op for final reduction. For N
3070 elements, we have N extracts and N-1 reduction ops. */
3071 epilogue_cost
+= add_stmt_cost (target_cost_data
,
3072 nelements
+ nelements
- 1,
3073 vector_stmt
, stmt_info
, 0,
3078 if (dump_enabled_p ())
3079 dump_printf (MSG_NOTE
,
3080 "vect_model_reduction_cost: inside_cost = %d, "
3081 "prologue_cost = %d, epilogue_cost = %d .", inside_cost
,
3082 prologue_cost
, epilogue_cost
);
3088 /* Function vect_model_induction_cost.
3090 Models cost for induction operations. */
3093 vect_model_induction_cost (stmt_vec_info stmt_info
, int ncopies
)
3095 loop_vec_info loop_vinfo
= STMT_VINFO_LOOP_VINFO (stmt_info
);
3096 void *target_cost_data
= LOOP_VINFO_TARGET_COST_DATA (loop_vinfo
);
3097 unsigned inside_cost
, prologue_cost
;
3099 /* loop cost for vec_loop. */
3100 inside_cost
= add_stmt_cost (target_cost_data
, ncopies
, vector_stmt
,
3101 stmt_info
, 0, vect_body
);
3103 /* prologue cost for vec_init and vec_step. */
3104 prologue_cost
= add_stmt_cost (target_cost_data
, 2, scalar_to_vec
,
3105 stmt_info
, 0, vect_prologue
);
3107 if (dump_enabled_p ())
3108 dump_printf_loc (MSG_NOTE
, vect_location
,
3109 "vect_model_induction_cost: inside_cost = %d, "
3110 "prologue_cost = %d .", inside_cost
, prologue_cost
);
3114 /* Function get_initial_def_for_induction
3117 STMT - a stmt that performs an induction operation in the loop.
3118 IV_PHI - the initial value of the induction variable
3121 Return a vector variable, initialized with the first VF values of
3122 the induction variable. E.g., for an iv with IV_PHI='X' and
3123 evolution S, for a vector of 4 units, we want to return:
3124 [X, X + S, X + 2*S, X + 3*S]. */
3127 get_initial_def_for_induction (gimple iv_phi
)
3129 stmt_vec_info stmt_vinfo
= vinfo_for_stmt (iv_phi
);
3130 loop_vec_info loop_vinfo
= STMT_VINFO_LOOP_VINFO (stmt_vinfo
);
3131 struct loop
*loop
= LOOP_VINFO_LOOP (loop_vinfo
);
3135 edge pe
= loop_preheader_edge (loop
);
3136 struct loop
*iv_loop
;
3138 tree new_vec
, vec_init
, vec_step
, t
;
3142 gimple init_stmt
, induction_phi
, new_stmt
;
3143 tree induc_def
, vec_def
, vec_dest
;
3144 tree init_expr
, step_expr
;
3145 int vf
= LOOP_VINFO_VECT_FACTOR (loop_vinfo
);
3150 stmt_vec_info phi_info
= vinfo_for_stmt (iv_phi
);
3151 bool nested_in_vect_loop
= false;
3152 gimple_seq stmts
= NULL
;
3153 imm_use_iterator imm_iter
;
3154 use_operand_p use_p
;
3158 gimple_stmt_iterator si
;
3159 basic_block bb
= gimple_bb (iv_phi
);
3163 /* Is phi in an inner-loop, while vectorizing an enclosing outer-loop? */
3164 if (nested_in_vect_loop_p (loop
, iv_phi
))
3166 nested_in_vect_loop
= true;
3167 iv_loop
= loop
->inner
;
3171 gcc_assert (iv_loop
== (gimple_bb (iv_phi
))->loop_father
);
3173 latch_e
= loop_latch_edge (iv_loop
);
3174 loop_arg
= PHI_ARG_DEF_FROM_EDGE (iv_phi
, latch_e
);
3176 access_fn
= analyze_scalar_evolution (iv_loop
, PHI_RESULT (iv_phi
));
3177 gcc_assert (access_fn
);
3178 STRIP_NOPS (access_fn
);
3179 ok
= vect_is_simple_iv_evolution (iv_loop
->num
, access_fn
,
3180 &init_expr
, &step_expr
);
3182 pe
= loop_preheader_edge (iv_loop
);
3184 scalar_type
= TREE_TYPE (init_expr
);
3185 vectype
= get_vectype_for_scalar_type (scalar_type
);
3186 resvectype
= get_vectype_for_scalar_type (TREE_TYPE (PHI_RESULT (iv_phi
)));
3187 gcc_assert (vectype
);
3188 nunits
= TYPE_VECTOR_SUBPARTS (vectype
);
3189 ncopies
= vf
/ nunits
;
3191 gcc_assert (phi_info
);
3192 gcc_assert (ncopies
>= 1);
3194 /* Find the first insertion point in the BB. */
3195 si
= gsi_after_labels (bb
);
3197 /* Create the vector that holds the initial_value of the induction. */
3198 if (nested_in_vect_loop
)
3200 /* iv_loop is nested in the loop to be vectorized. init_expr had already
3201 been created during vectorization of previous stmts. We obtain it
3202 from the STMT_VINFO_VEC_STMT of the defining stmt. */
3203 tree iv_def
= PHI_ARG_DEF_FROM_EDGE (iv_phi
,
3204 loop_preheader_edge (iv_loop
));
3205 vec_init
= vect_get_vec_def_for_operand (iv_def
, iv_phi
, NULL
);
3209 vec
<constructor_elt
, va_gc
> *v
;
3211 /* iv_loop is the loop to be vectorized. Create:
3212 vec_init = [X, X+S, X+2*S, X+3*S] (S = step_expr, X = init_expr) */
3213 new_var
= vect_get_new_vect_var (scalar_type
, vect_scalar_var
, "var_");
3214 new_name
= force_gimple_operand (init_expr
, &stmts
, false, new_var
);
3217 new_bb
= gsi_insert_seq_on_edge_immediate (pe
, stmts
);
3218 gcc_assert (!new_bb
);
3221 vec_alloc (v
, nunits
);
3222 CONSTRUCTOR_APPEND_ELT (v
, NULL_TREE
, new_name
);
3223 for (i
= 1; i
< nunits
; i
++)
3225 /* Create: new_name_i = new_name + step_expr */
3226 enum tree_code code
= POINTER_TYPE_P (scalar_type
)
3227 ? POINTER_PLUS_EXPR
: PLUS_EXPR
;
3228 init_stmt
= gimple_build_assign_with_ops (code
, new_var
,
3229 new_name
, step_expr
);
3230 new_name
= make_ssa_name (new_var
, init_stmt
);
3231 gimple_assign_set_lhs (init_stmt
, new_name
);
3233 new_bb
= gsi_insert_on_edge_immediate (pe
, init_stmt
);
3234 gcc_assert (!new_bb
);
3236 if (dump_enabled_p ())
3238 dump_printf_loc (MSG_NOTE
, vect_location
,
3239 "created new init_stmt: ");
3240 dump_gimple_stmt (MSG_NOTE
, TDF_SLIM
, init_stmt
, 0);
3242 CONSTRUCTOR_APPEND_ELT (v
, NULL_TREE
, new_name
);
3244 /* Create a vector from [new_name_0, new_name_1, ..., new_name_nunits-1] */
3245 new_vec
= build_constructor (vectype
, v
);
3246 vec_init
= vect_init_vector (iv_phi
, new_vec
, vectype
, NULL
);
3250 /* Create the vector that holds the step of the induction. */
3251 if (nested_in_vect_loop
)
3252 /* iv_loop is nested in the loop to be vectorized. Generate:
3253 vec_step = [S, S, S, S] */
3254 new_name
= step_expr
;
3257 /* iv_loop is the loop to be vectorized. Generate:
3258 vec_step = [VF*S, VF*S, VF*S, VF*S] */
3259 expr
= build_int_cst (TREE_TYPE (step_expr
), vf
);
3260 new_name
= fold_build2 (MULT_EXPR
, TREE_TYPE (step_expr
),
3264 t
= unshare_expr (new_name
);
3265 gcc_assert (CONSTANT_CLASS_P (new_name
));
3266 stepvectype
= get_vectype_for_scalar_type (TREE_TYPE (new_name
));
3267 gcc_assert (stepvectype
);
3268 new_vec
= build_vector_from_val (stepvectype
, t
);
3269 vec_step
= vect_init_vector (iv_phi
, new_vec
, stepvectype
, NULL
);
3272 /* Create the following def-use cycle:
3277 vec_iv = PHI <vec_init, vec_loop>
3281 vec_loop = vec_iv + vec_step; */
3283 /* Create the induction-phi that defines the induction-operand. */
3284 vec_dest
= vect_get_new_vect_var (vectype
, vect_simple_var
, "vec_iv_");
3285 induction_phi
= create_phi_node (vec_dest
, iv_loop
->header
);
3286 set_vinfo_for_stmt (induction_phi
,
3287 new_stmt_vec_info (induction_phi
, loop_vinfo
, NULL
));
3288 induc_def
= PHI_RESULT (induction_phi
);
3290 /* Create the iv update inside the loop */
3291 new_stmt
= gimple_build_assign_with_ops (PLUS_EXPR
, vec_dest
,
3292 induc_def
, vec_step
);
3293 vec_def
= make_ssa_name (vec_dest
, new_stmt
);
3294 gimple_assign_set_lhs (new_stmt
, vec_def
);
3295 gsi_insert_before (&si
, new_stmt
, GSI_SAME_STMT
);
3296 set_vinfo_for_stmt (new_stmt
, new_stmt_vec_info (new_stmt
, loop_vinfo
,
3299 /* Set the arguments of the phi node: */
3300 add_phi_arg (induction_phi
, vec_init
, pe
, UNKNOWN_LOCATION
);
3301 add_phi_arg (induction_phi
, vec_def
, loop_latch_edge (iv_loop
),
3305 /* In case that vectorization factor (VF) is bigger than the number
3306 of elements that we can fit in a vectype (nunits), we have to generate
3307 more than one vector stmt - i.e - we need to "unroll" the
3308 vector stmt by a factor VF/nunits. For more details see documentation
3309 in vectorizable_operation. */
3313 stmt_vec_info prev_stmt_vinfo
;
3314 /* FORNOW. This restriction should be relaxed. */
3315 gcc_assert (!nested_in_vect_loop
);
3317 /* Create the vector that holds the step of the induction. */
3318 expr
= build_int_cst (TREE_TYPE (step_expr
), nunits
);
3319 new_name
= fold_build2 (MULT_EXPR
, TREE_TYPE (step_expr
),
3321 t
= unshare_expr (new_name
);
3322 gcc_assert (CONSTANT_CLASS_P (new_name
));
3323 new_vec
= build_vector_from_val (stepvectype
, t
);
3324 vec_step
= vect_init_vector (iv_phi
, new_vec
, stepvectype
, NULL
);
3326 vec_def
= induc_def
;
3327 prev_stmt_vinfo
= vinfo_for_stmt (induction_phi
);
3328 for (i
= 1; i
< ncopies
; i
++)
3330 /* vec_i = vec_prev + vec_step */
3331 new_stmt
= gimple_build_assign_with_ops (PLUS_EXPR
, vec_dest
,
3333 vec_def
= make_ssa_name (vec_dest
, new_stmt
);
3334 gimple_assign_set_lhs (new_stmt
, vec_def
);
3336 gsi_insert_before (&si
, new_stmt
, GSI_SAME_STMT
);
3337 if (!useless_type_conversion_p (resvectype
, vectype
))
3339 new_stmt
= gimple_build_assign_with_ops
3341 vect_get_new_vect_var (resvectype
, vect_simple_var
,
3343 build1 (VIEW_CONVERT_EXPR
, resvectype
,
3344 gimple_assign_lhs (new_stmt
)), NULL_TREE
);
3345 gimple_assign_set_lhs (new_stmt
,
3347 (gimple_assign_lhs (new_stmt
), new_stmt
));
3348 gsi_insert_before (&si
, new_stmt
, GSI_SAME_STMT
);
3350 set_vinfo_for_stmt (new_stmt
,
3351 new_stmt_vec_info (new_stmt
, loop_vinfo
, NULL
));
3352 STMT_VINFO_RELATED_STMT (prev_stmt_vinfo
) = new_stmt
;
3353 prev_stmt_vinfo
= vinfo_for_stmt (new_stmt
);
3357 if (nested_in_vect_loop
)
3359 /* Find the loop-closed exit-phi of the induction, and record
3360 the final vector of induction results: */
3362 FOR_EACH_IMM_USE_FAST (use_p
, imm_iter
, loop_arg
)
3364 if (!flow_bb_inside_loop_p (iv_loop
, gimple_bb (USE_STMT (use_p
))))
3366 exit_phi
= USE_STMT (use_p
);
3372 stmt_vec_info stmt_vinfo
= vinfo_for_stmt (exit_phi
);
3373 /* FORNOW. Currently not supporting the case that an inner-loop induction
3374 is not used in the outer-loop (i.e. only outside the outer-loop). */
3375 gcc_assert (STMT_VINFO_RELEVANT_P (stmt_vinfo
)
3376 && !STMT_VINFO_LIVE_P (stmt_vinfo
));
3378 STMT_VINFO_VEC_STMT (stmt_vinfo
) = new_stmt
;
3379 if (dump_enabled_p ())
3381 dump_printf_loc (MSG_NOTE
, vect_location
,
3382 "vector of inductions after inner-loop:");
3383 dump_gimple_stmt (MSG_NOTE
, TDF_SLIM
, new_stmt
, 0);
3389 if (dump_enabled_p ())
3391 dump_printf_loc (MSG_NOTE
, vect_location
,
3392 "transform induction: created def-use cycle: ");
3393 dump_gimple_stmt (MSG_NOTE
, TDF_SLIM
, induction_phi
, 0);
3394 dump_printf (MSG_NOTE
, "\n");
3395 dump_gimple_stmt (MSG_NOTE
, TDF_SLIM
,
3396 SSA_NAME_DEF_STMT (vec_def
), 0);
3399 STMT_VINFO_VEC_STMT (phi_info
) = induction_phi
;
3400 if (!useless_type_conversion_p (resvectype
, vectype
))
3402 new_stmt
= gimple_build_assign_with_ops
3404 vect_get_new_vect_var (resvectype
, vect_simple_var
, "vec_iv_"),
3405 build1 (VIEW_CONVERT_EXPR
, resvectype
, induc_def
), NULL_TREE
);
3406 induc_def
= make_ssa_name (gimple_assign_lhs (new_stmt
), new_stmt
);
3407 gimple_assign_set_lhs (new_stmt
, induc_def
);
3408 si
= gsi_after_labels (bb
);
3409 gsi_insert_before (&si
, new_stmt
, GSI_SAME_STMT
);
3410 set_vinfo_for_stmt (new_stmt
,
3411 new_stmt_vec_info (new_stmt
, loop_vinfo
, NULL
));
3412 STMT_VINFO_RELATED_STMT (vinfo_for_stmt (new_stmt
))
3413 = STMT_VINFO_RELATED_STMT (vinfo_for_stmt (induction_phi
));
3420 /* Function get_initial_def_for_reduction
3423 STMT - a stmt that performs a reduction operation in the loop.
3424 INIT_VAL - the initial value of the reduction variable
3427 ADJUSTMENT_DEF - a tree that holds a value to be added to the final result
3428 of the reduction (used for adjusting the epilog - see below).
3429 Return a vector variable, initialized according to the operation that STMT
3430 performs. This vector will be used as the initial value of the
3431 vector of partial results.
3433 Option1 (adjust in epilog): Initialize the vector as follows:
3434 add/bit or/xor: [0,0,...,0,0]
3435 mult/bit and: [1,1,...,1,1]
3436 min/max/cond_expr: [init_val,init_val,..,init_val,init_val]
3437 and when necessary (e.g. add/mult case) let the caller know
3438 that it needs to adjust the result by init_val.
3440 Option2: Initialize the vector as follows:
3441 add/bit or/xor: [init_val,0,0,...,0]
3442 mult/bit and: [init_val,1,1,...,1]
3443 min/max/cond_expr: [init_val,init_val,...,init_val]
3444 and no adjustments are needed.
3446 For example, for the following code:
3452 STMT is 's = s + a[i]', and the reduction variable is 's'.
3453 For a vector of 4 units, we want to return either [0,0,0,init_val],
3454 or [0,0,0,0] and let the caller know that it needs to adjust
3455 the result at the end by 'init_val'.
3457 FORNOW, we are using the 'adjust in epilog' scheme, because this way the
3458 initialization vector is simpler (same element in all entries), if
3459 ADJUSTMENT_DEF is not NULL, and Option2 otherwise.
3461 A cost model should help decide between these two schemes. */
3464 get_initial_def_for_reduction (gimple stmt
, tree init_val
,
3465 tree
*adjustment_def
)
3467 stmt_vec_info stmt_vinfo
= vinfo_for_stmt (stmt
);
3468 loop_vec_info loop_vinfo
= STMT_VINFO_LOOP_VINFO (stmt_vinfo
);
3469 struct loop
*loop
= LOOP_VINFO_LOOP (loop_vinfo
);
3470 tree scalar_type
= TREE_TYPE (init_val
);
3471 tree vectype
= get_vectype_for_scalar_type (scalar_type
);
3473 enum tree_code code
= gimple_assign_rhs_code (stmt
);
3478 bool nested_in_vect_loop
= false;
3480 REAL_VALUE_TYPE real_init_val
= dconst0
;
3481 int int_init_val
= 0;
3482 gimple def_stmt
= NULL
;
3484 gcc_assert (vectype
);
3485 nunits
= TYPE_VECTOR_SUBPARTS (vectype
);
3487 gcc_assert (POINTER_TYPE_P (scalar_type
) || INTEGRAL_TYPE_P (scalar_type
)
3488 || SCALAR_FLOAT_TYPE_P (scalar_type
));
3490 if (nested_in_vect_loop_p (loop
, stmt
))
3491 nested_in_vect_loop
= true;
3493 gcc_assert (loop
== (gimple_bb (stmt
))->loop_father
);
3495 /* In case of double reduction we only create a vector variable to be put
3496 in the reduction phi node. The actual statement creation is done in
3497 vect_create_epilog_for_reduction. */
3498 if (adjustment_def
&& nested_in_vect_loop
3499 && TREE_CODE (init_val
) == SSA_NAME
3500 && (def_stmt
= SSA_NAME_DEF_STMT (init_val
))
3501 && gimple_code (def_stmt
) == GIMPLE_PHI
3502 && flow_bb_inside_loop_p (loop
, gimple_bb (def_stmt
))
3503 && vinfo_for_stmt (def_stmt
)
3504 && STMT_VINFO_DEF_TYPE (vinfo_for_stmt (def_stmt
))
3505 == vect_double_reduction_def
)
3507 *adjustment_def
= NULL
;
3508 return vect_create_destination_var (init_val
, vectype
);
3511 if (TREE_CONSTANT (init_val
))
3513 if (SCALAR_FLOAT_TYPE_P (scalar_type
))
3514 init_value
= build_real (scalar_type
, TREE_REAL_CST (init_val
));
3516 init_value
= build_int_cst (scalar_type
, TREE_INT_CST_LOW (init_val
));
3519 init_value
= init_val
;
3523 case WIDEN_SUM_EXPR
:
3531 /* ADJUSMENT_DEF is NULL when called from
3532 vect_create_epilog_for_reduction to vectorize double reduction. */
3535 if (nested_in_vect_loop
)
3536 *adjustment_def
= vect_get_vec_def_for_operand (init_val
, stmt
,
3539 *adjustment_def
= init_val
;
3542 if (code
== MULT_EXPR
)
3544 real_init_val
= dconst1
;
3548 if (code
== BIT_AND_EXPR
)
3551 if (SCALAR_FLOAT_TYPE_P (scalar_type
))
3552 def_for_init
= build_real (scalar_type
, real_init_val
);
3554 def_for_init
= build_int_cst (scalar_type
, int_init_val
);
3556 /* Create a vector of '0' or '1' except the first element. */
3557 elts
= XALLOCAVEC (tree
, nunits
);
3558 for (i
= nunits
- 2; i
>= 0; --i
)
3559 elts
[i
+ 1] = def_for_init
;
3561 /* Option1: the first element is '0' or '1' as well. */
3564 elts
[0] = def_for_init
;
3565 init_def
= build_vector (vectype
, elts
);
3569 /* Option2: the first element is INIT_VAL. */
3571 if (TREE_CONSTANT (init_val
))
3572 init_def
= build_vector (vectype
, elts
);
3575 vec
<constructor_elt
, va_gc
> *v
;
3576 vec_alloc (v
, nunits
);
3577 CONSTRUCTOR_APPEND_ELT (v
, NULL_TREE
, init_val
);
3578 for (i
= 1; i
< nunits
; ++i
)
3579 CONSTRUCTOR_APPEND_ELT (v
, NULL_TREE
, elts
[i
]);
3580 init_def
= build_constructor (vectype
, v
);
3590 *adjustment_def
= NULL_TREE
;
3591 init_def
= vect_get_vec_def_for_operand (init_val
, stmt
, NULL
);
3595 init_def
= build_vector_from_val (vectype
, init_value
);
3606 /* Function vect_create_epilog_for_reduction
3608 Create code at the loop-epilog to finalize the result of a reduction
3611 VECT_DEFS is list of vector of partial results, i.e., the lhs's of vector
3612 reduction statements.
3613 STMT is the scalar reduction stmt that is being vectorized.
3614 NCOPIES is > 1 in case the vectorization factor (VF) is bigger than the
3615 number of elements that we can fit in a vectype (nunits). In this case
3616 we have to generate more than one vector stmt - i.e - we need to "unroll"
3617 the vector stmt by a factor VF/nunits. For more details see documentation
3618 in vectorizable_operation.
3619 REDUC_CODE is the tree-code for the epilog reduction.
3620 REDUCTION_PHIS is a list of the phi-nodes that carry the reduction
3622 REDUC_INDEX is the index of the operand in the right hand side of the
3623 statement that is defined by REDUCTION_PHI.
3624 DOUBLE_REDUC is TRUE if double reduction phi nodes should be handled.
3625 SLP_NODE is an SLP node containing a group of reduction statements. The
3626 first one in this group is STMT.
3629 1. Creates the reduction def-use cycles: sets the arguments for
3631 The loop-entry argument is the vectorized initial-value of the reduction.
3632 The loop-latch argument is taken from VECT_DEFS - the vector of partial
3634 2. "Reduces" each vector of partial results VECT_DEFS into a single result,
3635 by applying the operation specified by REDUC_CODE if available, or by
3636 other means (whole-vector shifts or a scalar loop).
3637 The function also creates a new phi node at the loop exit to preserve
3638 loop-closed form, as illustrated below.
3640 The flow at the entry to this function:
3643 vec_def = phi <null, null> # REDUCTION_PHI
3644 VECT_DEF = vector_stmt # vectorized form of STMT
3645 s_loop = scalar_stmt # (scalar) STMT
3647 s_out0 = phi <s_loop> # (scalar) EXIT_PHI
3651 The above is transformed by this function into:
3654 vec_def = phi <vec_init, VECT_DEF> # REDUCTION_PHI
3655 VECT_DEF = vector_stmt # vectorized form of STMT
3656 s_loop = scalar_stmt # (scalar) STMT
3658 s_out0 = phi <s_loop> # (scalar) EXIT_PHI
3659 v_out1 = phi <VECT_DEF> # NEW_EXIT_PHI
3660 v_out2 = reduce <v_out1>
3661 s_out3 = extract_field <v_out2, 0>
3662 s_out4 = adjust_result <s_out3>
3668 vect_create_epilog_for_reduction (vec
<tree
> vect_defs
, gimple stmt
,
3669 int ncopies
, enum tree_code reduc_code
,
3670 vec
<gimple
> reduction_phis
,
3671 int reduc_index
, bool double_reduc
,
3674 stmt_vec_info stmt_info
= vinfo_for_stmt (stmt
);
3675 stmt_vec_info prev_phi_info
;
3677 enum machine_mode mode
;
3678 loop_vec_info loop_vinfo
= STMT_VINFO_LOOP_VINFO (stmt_info
);
3679 struct loop
*loop
= LOOP_VINFO_LOOP (loop_vinfo
), *outer_loop
= NULL
;
3680 basic_block exit_bb
;
3683 gimple new_phi
= NULL
, phi
;
3684 gimple_stmt_iterator exit_gsi
;
3686 tree new_temp
= NULL_TREE
, new_dest
, new_name
, new_scalar_dest
;
3687 gimple epilog_stmt
= NULL
;
3688 enum tree_code code
= gimple_assign_rhs_code (stmt
);
3690 tree bitsize
, bitpos
;
3691 tree adjustment_def
= NULL
;
3692 tree vec_initial_def
= NULL
;
3693 tree reduction_op
, expr
, def
;
3694 tree orig_name
, scalar_result
;
3695 imm_use_iterator imm_iter
, phi_imm_iter
;
3696 use_operand_p use_p
, phi_use_p
;
3697 bool extract_scalar_result
= false;
3698 gimple use_stmt
, orig_stmt
, reduction_phi
= NULL
;
3699 bool nested_in_vect_loop
= false;
3700 vec
<gimple
> new_phis
= vNULL
;
3701 vec
<gimple
> inner_phis
= vNULL
;
3702 enum vect_def_type dt
= vect_unknown_def_type
;
3704 vec
<tree
> scalar_results
= vNULL
;
3705 unsigned int group_size
= 1, k
, ratio
;
3706 vec
<tree
> vec_initial_defs
= vNULL
;
3708 bool slp_reduc
= false;
3709 tree new_phi_result
;
3710 gimple inner_phi
= NULL
;
3713 group_size
= SLP_TREE_SCALAR_STMTS (slp_node
).length ();
3715 if (nested_in_vect_loop_p (loop
, stmt
))
3719 nested_in_vect_loop
= true;
3720 gcc_assert (!slp_node
);
3723 switch (get_gimple_rhs_class (gimple_assign_rhs_code (stmt
)))
3725 case GIMPLE_SINGLE_RHS
:
3726 gcc_assert (TREE_OPERAND_LENGTH (gimple_assign_rhs1 (stmt
))
3728 reduction_op
= TREE_OPERAND (gimple_assign_rhs1 (stmt
), reduc_index
);
3730 case GIMPLE_UNARY_RHS
:
3731 reduction_op
= gimple_assign_rhs1 (stmt
);
3733 case GIMPLE_BINARY_RHS
:
3734 reduction_op
= reduc_index
?
3735 gimple_assign_rhs2 (stmt
) : gimple_assign_rhs1 (stmt
);
3737 case GIMPLE_TERNARY_RHS
:
3738 reduction_op
= gimple_op (stmt
, reduc_index
+ 1);
3744 vectype
= get_vectype_for_scalar_type (TREE_TYPE (reduction_op
));
3745 gcc_assert (vectype
);
3746 mode
= TYPE_MODE (vectype
);
3748 /* 1. Create the reduction def-use cycle:
3749 Set the arguments of REDUCTION_PHIS, i.e., transform
3752 vec_def = phi <null, null> # REDUCTION_PHI
3753 VECT_DEF = vector_stmt # vectorized form of STMT
3759 vec_def = phi <vec_init, VECT_DEF> # REDUCTION_PHI
3760 VECT_DEF = vector_stmt # vectorized form of STMT
3763 (in case of SLP, do it for all the phis). */
3765 /* Get the loop-entry arguments. */
3767 vect_get_vec_defs (reduction_op
, NULL_TREE
, stmt
, &vec_initial_defs
,
3768 NULL
, slp_node
, reduc_index
);
3771 vec_initial_defs
.create (1);
3772 /* For the case of reduction, vect_get_vec_def_for_operand returns
3773 the scalar def before the loop, that defines the initial value
3774 of the reduction variable. */
3775 vec_initial_def
= vect_get_vec_def_for_operand (reduction_op
, stmt
,
3777 vec_initial_defs
.quick_push (vec_initial_def
);
3780 /* Set phi nodes arguments. */
3781 FOR_EACH_VEC_ELT (reduction_phis
, i
, phi
)
3783 tree vec_init_def
= vec_initial_defs
[i
];
3784 tree def
= vect_defs
[i
];
3785 for (j
= 0; j
< ncopies
; j
++)
3787 /* Set the loop-entry arg of the reduction-phi. */
3788 add_phi_arg (phi
, vec_init_def
, loop_preheader_edge (loop
),
3791 /* Set the loop-latch arg for the reduction-phi. */
3793 def
= vect_get_vec_def_for_stmt_copy (vect_unknown_def_type
, def
);
3795 add_phi_arg (phi
, def
, loop_latch_edge (loop
), UNKNOWN_LOCATION
);
3797 if (dump_enabled_p ())
3799 dump_printf_loc (MSG_NOTE
, vect_location
,
3800 "transform reduction: created def-use cycle: ");
3801 dump_gimple_stmt (MSG_NOTE
, TDF_SLIM
, phi
, 0);
3802 dump_printf (MSG_NOTE
, "\n");
3803 dump_gimple_stmt (MSG_NOTE
, TDF_SLIM
, SSA_NAME_DEF_STMT (def
), 0);
3806 phi
= STMT_VINFO_RELATED_STMT (vinfo_for_stmt (phi
));
3810 vec_initial_defs
.release ();
3812 /* 2. Create epilog code.
3813 The reduction epilog code operates across the elements of the vector
3814 of partial results computed by the vectorized loop.
3815 The reduction epilog code consists of:
3817 step 1: compute the scalar result in a vector (v_out2)
3818 step 2: extract the scalar result (s_out3) from the vector (v_out2)
3819 step 3: adjust the scalar result (s_out3) if needed.
3821 Step 1 can be accomplished using one the following three schemes:
3822 (scheme 1) using reduc_code, if available.
3823 (scheme 2) using whole-vector shifts, if available.
3824 (scheme 3) using a scalar loop. In this case steps 1+2 above are
3827 The overall epilog code looks like this:
3829 s_out0 = phi <s_loop> # original EXIT_PHI
3830 v_out1 = phi <VECT_DEF> # NEW_EXIT_PHI
3831 v_out2 = reduce <v_out1> # step 1
3832 s_out3 = extract_field <v_out2, 0> # step 2
3833 s_out4 = adjust_result <s_out3> # step 3
3835 (step 3 is optional, and steps 1 and 2 may be combined).
3836 Lastly, the uses of s_out0 are replaced by s_out4. */
3839 /* 2.1 Create new loop-exit-phis to preserve loop-closed form:
3840 v_out1 = phi <VECT_DEF>
3841 Store them in NEW_PHIS. */
3843 exit_bb
= single_exit (loop
)->dest
;
3844 prev_phi_info
= NULL
;
3845 new_phis
.create (vect_defs
.length ());
3846 FOR_EACH_VEC_ELT (vect_defs
, i
, def
)
3848 for (j
= 0; j
< ncopies
; j
++)
3850 tree new_def
= copy_ssa_name (def
, NULL
);
3851 phi
= create_phi_node (new_def
, exit_bb
);
3852 set_vinfo_for_stmt (phi
, new_stmt_vec_info (phi
, loop_vinfo
, NULL
));
3854 new_phis
.quick_push (phi
);
3857 def
= vect_get_vec_def_for_stmt_copy (dt
, def
);
3858 STMT_VINFO_RELATED_STMT (prev_phi_info
) = phi
;
3861 SET_PHI_ARG_DEF (phi
, single_exit (loop
)->dest_idx
, def
);
3862 prev_phi_info
= vinfo_for_stmt (phi
);
3866 /* The epilogue is created for the outer-loop, i.e., for the loop being
3867 vectorized. Create exit phis for the outer loop. */
3871 exit_bb
= single_exit (loop
)->dest
;
3872 inner_phis
.create (vect_defs
.length ());
3873 FOR_EACH_VEC_ELT (new_phis
, i
, phi
)
3875 tree new_result
= copy_ssa_name (PHI_RESULT (phi
), NULL
);
3876 gimple outer_phi
= create_phi_node (new_result
, exit_bb
);
3877 SET_PHI_ARG_DEF (outer_phi
, single_exit (loop
)->dest_idx
,
3879 set_vinfo_for_stmt (outer_phi
, new_stmt_vec_info (outer_phi
,
3881 inner_phis
.quick_push (phi
);
3882 new_phis
[i
] = outer_phi
;
3883 prev_phi_info
= vinfo_for_stmt (outer_phi
);
3884 while (STMT_VINFO_RELATED_STMT (vinfo_for_stmt (phi
)))
3886 phi
= STMT_VINFO_RELATED_STMT (vinfo_for_stmt (phi
));
3887 new_result
= copy_ssa_name (PHI_RESULT (phi
), NULL
);
3888 outer_phi
= create_phi_node (new_result
, exit_bb
);
3889 SET_PHI_ARG_DEF (outer_phi
, single_exit (loop
)->dest_idx
,
3891 set_vinfo_for_stmt (outer_phi
, new_stmt_vec_info (outer_phi
,
3893 STMT_VINFO_RELATED_STMT (prev_phi_info
) = outer_phi
;
3894 prev_phi_info
= vinfo_for_stmt (outer_phi
);
3899 exit_gsi
= gsi_after_labels (exit_bb
);
3901 /* 2.2 Get the relevant tree-code to use in the epilog for schemes 2,3
3902 (i.e. when reduc_code is not available) and in the final adjustment
3903 code (if needed). Also get the original scalar reduction variable as
3904 defined in the loop. In case STMT is a "pattern-stmt" (i.e. - it
3905 represents a reduction pattern), the tree-code and scalar-def are
3906 taken from the original stmt that the pattern-stmt (STMT) replaces.
3907 Otherwise (it is a regular reduction) - the tree-code and scalar-def
3908 are taken from STMT. */
3910 orig_stmt
= STMT_VINFO_RELATED_STMT (stmt_info
);
3913 /* Regular reduction */
3918 /* Reduction pattern */
3919 stmt_vec_info stmt_vinfo
= vinfo_for_stmt (orig_stmt
);
3920 gcc_assert (STMT_VINFO_IN_PATTERN_P (stmt_vinfo
));
3921 gcc_assert (STMT_VINFO_RELATED_STMT (stmt_vinfo
) == stmt
);
3924 code
= gimple_assign_rhs_code (orig_stmt
);
3925 /* For MINUS_EXPR the initial vector is [init_val,0,...,0], therefore,
3926 partial results are added and not subtracted. */
3927 if (code
== MINUS_EXPR
)
3930 scalar_dest
= gimple_assign_lhs (orig_stmt
);
3931 scalar_type
= TREE_TYPE (scalar_dest
);
3932 scalar_results
.create (group_size
);
3933 new_scalar_dest
= vect_create_destination_var (scalar_dest
, NULL
);
3934 bitsize
= TYPE_SIZE (scalar_type
);
3936 /* In case this is a reduction in an inner-loop while vectorizing an outer
3937 loop - we don't need to extract a single scalar result at the end of the
3938 inner-loop (unless it is double reduction, i.e., the use of reduction is
3939 outside the outer-loop). The final vector of partial results will be used
3940 in the vectorized outer-loop, or reduced to a scalar result at the end of
3942 if (nested_in_vect_loop
&& !double_reduc
)
3943 goto vect_finalize_reduction
;
3945 /* SLP reduction without reduction chain, e.g.,
3949 b2 = operation (b1) */
3950 slp_reduc
= (slp_node
&& !GROUP_FIRST_ELEMENT (vinfo_for_stmt (stmt
)));
3952 /* In case of reduction chain, e.g.,
3955 a3 = operation (a2),
3957 we may end up with more than one vector result. Here we reduce them to
3959 if (GROUP_FIRST_ELEMENT (vinfo_for_stmt (stmt
)))
3961 tree first_vect
= PHI_RESULT (new_phis
[0]);
3963 gimple new_vec_stmt
= NULL
;
3965 vec_dest
= vect_create_destination_var (scalar_dest
, vectype
);
3966 for (k
= 1; k
< new_phis
.length (); k
++)
3968 gimple next_phi
= new_phis
[k
];
3969 tree second_vect
= PHI_RESULT (next_phi
);
3971 tmp
= build2 (code
, vectype
, first_vect
, second_vect
);
3972 new_vec_stmt
= gimple_build_assign (vec_dest
, tmp
);
3973 first_vect
= make_ssa_name (vec_dest
, new_vec_stmt
);
3974 gimple_assign_set_lhs (new_vec_stmt
, first_vect
);
3975 gsi_insert_before (&exit_gsi
, new_vec_stmt
, GSI_SAME_STMT
);
3978 new_phi_result
= first_vect
;
3981 new_phis
.truncate (0);
3982 new_phis
.safe_push (new_vec_stmt
);
3986 new_phi_result
= PHI_RESULT (new_phis
[0]);
3988 /* 2.3 Create the reduction code, using one of the three schemes described
3989 above. In SLP we simply need to extract all the elements from the
3990 vector (without reducing them), so we use scalar shifts. */
3991 if (reduc_code
!= ERROR_MARK
&& !slp_reduc
)
3995 /*** Case 1: Create:
3996 v_out2 = reduc_expr <v_out1> */
3998 if (dump_enabled_p ())
3999 dump_printf_loc (MSG_NOTE
, vect_location
,
4000 "Reduce using direct vector reduction.");
4002 vec_dest
= vect_create_destination_var (scalar_dest
, vectype
);
4003 tmp
= build1 (reduc_code
, vectype
, new_phi_result
);
4004 epilog_stmt
= gimple_build_assign (vec_dest
, tmp
);
4005 new_temp
= make_ssa_name (vec_dest
, epilog_stmt
);
4006 gimple_assign_set_lhs (epilog_stmt
, new_temp
);
4007 gsi_insert_before (&exit_gsi
, epilog_stmt
, GSI_SAME_STMT
);
4009 extract_scalar_result
= true;
4013 enum tree_code shift_code
= ERROR_MARK
;
4014 bool have_whole_vector_shift
= true;
4016 int element_bitsize
= tree_low_cst (bitsize
, 1);
4017 int vec_size_in_bits
= tree_low_cst (TYPE_SIZE (vectype
), 1);
4020 if (optab_handler (vec_shr_optab
, mode
) != CODE_FOR_nothing
)
4021 shift_code
= VEC_RSHIFT_EXPR
;
4023 have_whole_vector_shift
= false;
4025 /* Regardless of whether we have a whole vector shift, if we're
4026 emulating the operation via tree-vect-generic, we don't want
4027 to use it. Only the first round of the reduction is likely
4028 to still be profitable via emulation. */
4029 /* ??? It might be better to emit a reduction tree code here, so that
4030 tree-vect-generic can expand the first round via bit tricks. */
4031 if (!VECTOR_MODE_P (mode
))
4032 have_whole_vector_shift
= false;
4035 optab optab
= optab_for_tree_code (code
, vectype
, optab_default
);
4036 if (optab_handler (optab
, mode
) == CODE_FOR_nothing
)
4037 have_whole_vector_shift
= false;
4040 if (have_whole_vector_shift
&& !slp_reduc
)
4042 /*** Case 2: Create:
4043 for (offset = VS/2; offset >= element_size; offset/=2)
4045 Create: va' = vec_shift <va, offset>
4046 Create: va = vop <va, va'>
4049 if (dump_enabled_p ())
4050 dump_printf_loc (MSG_NOTE
, vect_location
,
4051 "Reduce using vector shifts");
4053 vec_dest
= vect_create_destination_var (scalar_dest
, vectype
);
4054 new_temp
= new_phi_result
;
4055 for (bit_offset
= vec_size_in_bits
/2;
4056 bit_offset
>= element_bitsize
;
4059 tree bitpos
= size_int (bit_offset
);
4061 epilog_stmt
= gimple_build_assign_with_ops (shift_code
,
4062 vec_dest
, new_temp
, bitpos
);
4063 new_name
= make_ssa_name (vec_dest
, epilog_stmt
);
4064 gimple_assign_set_lhs (epilog_stmt
, new_name
);
4065 gsi_insert_before (&exit_gsi
, epilog_stmt
, GSI_SAME_STMT
);
4067 epilog_stmt
= gimple_build_assign_with_ops (code
, vec_dest
,
4068 new_name
, new_temp
);
4069 new_temp
= make_ssa_name (vec_dest
, epilog_stmt
);
4070 gimple_assign_set_lhs (epilog_stmt
, new_temp
);
4071 gsi_insert_before (&exit_gsi
, epilog_stmt
, GSI_SAME_STMT
);
4074 extract_scalar_result
= true;
4080 /*** Case 3: Create:
4081 s = extract_field <v_out2, 0>
4082 for (offset = element_size;
4083 offset < vector_size;
4084 offset += element_size;)
4086 Create: s' = extract_field <v_out2, offset>
4087 Create: s = op <s, s'> // For non SLP cases
4090 if (dump_enabled_p ())
4091 dump_printf_loc (MSG_NOTE
, vect_location
,
4092 "Reduce using scalar code. ");
4094 vec_size_in_bits
= tree_low_cst (TYPE_SIZE (vectype
), 1);
4095 FOR_EACH_VEC_ELT (new_phis
, i
, new_phi
)
4097 if (gimple_code (new_phi
) == GIMPLE_PHI
)
4098 vec_temp
= PHI_RESULT (new_phi
);
4100 vec_temp
= gimple_assign_lhs (new_phi
);
4101 rhs
= build3 (BIT_FIELD_REF
, scalar_type
, vec_temp
, bitsize
,
4103 epilog_stmt
= gimple_build_assign (new_scalar_dest
, rhs
);
4104 new_temp
= make_ssa_name (new_scalar_dest
, epilog_stmt
);
4105 gimple_assign_set_lhs (epilog_stmt
, new_temp
);
4106 gsi_insert_before (&exit_gsi
, epilog_stmt
, GSI_SAME_STMT
);
4108 /* In SLP we don't need to apply reduction operation, so we just
4109 collect s' values in SCALAR_RESULTS. */
4111 scalar_results
.safe_push (new_temp
);
4113 for (bit_offset
= element_bitsize
;
4114 bit_offset
< vec_size_in_bits
;
4115 bit_offset
+= element_bitsize
)
4117 tree bitpos
= bitsize_int (bit_offset
);
4118 tree rhs
= build3 (BIT_FIELD_REF
, scalar_type
, vec_temp
,
4121 epilog_stmt
= gimple_build_assign (new_scalar_dest
, rhs
);
4122 new_name
= make_ssa_name (new_scalar_dest
, epilog_stmt
);
4123 gimple_assign_set_lhs (epilog_stmt
, new_name
);
4124 gsi_insert_before (&exit_gsi
, epilog_stmt
, GSI_SAME_STMT
);
4128 /* In SLP we don't need to apply reduction operation, so
4129 we just collect s' values in SCALAR_RESULTS. */
4130 new_temp
= new_name
;
4131 scalar_results
.safe_push (new_name
);
4135 epilog_stmt
= gimple_build_assign_with_ops (code
,
4136 new_scalar_dest
, new_name
, new_temp
);
4137 new_temp
= make_ssa_name (new_scalar_dest
, epilog_stmt
);
4138 gimple_assign_set_lhs (epilog_stmt
, new_temp
);
4139 gsi_insert_before (&exit_gsi
, epilog_stmt
, GSI_SAME_STMT
);
4144 /* The only case where we need to reduce scalar results in SLP, is
4145 unrolling. If the size of SCALAR_RESULTS is greater than
4146 GROUP_SIZE, we reduce them combining elements modulo
4150 tree res
, first_res
, new_res
;
4153 /* Reduce multiple scalar results in case of SLP unrolling. */
4154 for (j
= group_size
; scalar_results
.iterate (j
, &res
);
4157 first_res
= scalar_results
[j
% group_size
];
4158 new_stmt
= gimple_build_assign_with_ops (code
,
4159 new_scalar_dest
, first_res
, res
);
4160 new_res
= make_ssa_name (new_scalar_dest
, new_stmt
);
4161 gimple_assign_set_lhs (new_stmt
, new_res
);
4162 gsi_insert_before (&exit_gsi
, new_stmt
, GSI_SAME_STMT
);
4163 scalar_results
[j
% group_size
] = new_res
;
4167 /* Not SLP - we have one scalar to keep in SCALAR_RESULTS. */
4168 scalar_results
.safe_push (new_temp
);
4170 extract_scalar_result
= false;
4174 /* 2.4 Extract the final scalar result. Create:
4175 s_out3 = extract_field <v_out2, bitpos> */
4177 if (extract_scalar_result
)
4181 if (dump_enabled_p ())
4182 dump_printf_loc (MSG_NOTE
, vect_location
,
4183 "extract scalar result");
4185 if (BYTES_BIG_ENDIAN
)
4186 bitpos
= size_binop (MULT_EXPR
,
4187 bitsize_int (TYPE_VECTOR_SUBPARTS (vectype
) - 1),
4188 TYPE_SIZE (scalar_type
));
4190 bitpos
= bitsize_zero_node
;
4192 rhs
= build3 (BIT_FIELD_REF
, scalar_type
, new_temp
, bitsize
, bitpos
);
4193 epilog_stmt
= gimple_build_assign (new_scalar_dest
, rhs
);
4194 new_temp
= make_ssa_name (new_scalar_dest
, epilog_stmt
);
4195 gimple_assign_set_lhs (epilog_stmt
, new_temp
);
4196 gsi_insert_before (&exit_gsi
, epilog_stmt
, GSI_SAME_STMT
);
4197 scalar_results
.safe_push (new_temp
);
4200 vect_finalize_reduction
:
4205 /* 2.5 Adjust the final result by the initial value of the reduction
4206 variable. (When such adjustment is not needed, then
4207 'adjustment_def' is zero). For example, if code is PLUS we create:
4208 new_temp = loop_exit_def + adjustment_def */
4212 gcc_assert (!slp_reduc
);
4213 if (nested_in_vect_loop
)
4215 new_phi
= new_phis
[0];
4216 gcc_assert (TREE_CODE (TREE_TYPE (adjustment_def
)) == VECTOR_TYPE
);
4217 expr
= build2 (code
, vectype
, PHI_RESULT (new_phi
), adjustment_def
);
4218 new_dest
= vect_create_destination_var (scalar_dest
, vectype
);
4222 new_temp
= scalar_results
[0];
4223 gcc_assert (TREE_CODE (TREE_TYPE (adjustment_def
)) != VECTOR_TYPE
);
4224 expr
= build2 (code
, scalar_type
, new_temp
, adjustment_def
);
4225 new_dest
= vect_create_destination_var (scalar_dest
, scalar_type
);
4228 epilog_stmt
= gimple_build_assign (new_dest
, expr
);
4229 new_temp
= make_ssa_name (new_dest
, epilog_stmt
);
4230 gimple_assign_set_lhs (epilog_stmt
, new_temp
);
4231 SSA_NAME_DEF_STMT (new_temp
) = epilog_stmt
;
4232 gsi_insert_before (&exit_gsi
, epilog_stmt
, GSI_SAME_STMT
);
4233 if (nested_in_vect_loop
)
4235 set_vinfo_for_stmt (epilog_stmt
,
4236 new_stmt_vec_info (epilog_stmt
, loop_vinfo
,
4238 STMT_VINFO_RELATED_STMT (vinfo_for_stmt (epilog_stmt
)) =
4239 STMT_VINFO_RELATED_STMT (vinfo_for_stmt (new_phi
));
4242 scalar_results
.quick_push (new_temp
);
4244 scalar_results
[0] = new_temp
;
4247 scalar_results
[0] = new_temp
;
4249 new_phis
[0] = epilog_stmt
;
4252 /* 2.6 Handle the loop-exit phis. Replace the uses of scalar loop-exit
4253 phis with new adjusted scalar results, i.e., replace use <s_out0>
4258 s_out0 = phi <s_loop> # (scalar) EXIT_PHI
4259 v_out1 = phi <VECT_DEF> # NEW_EXIT_PHI
4260 v_out2 = reduce <v_out1>
4261 s_out3 = extract_field <v_out2, 0>
4262 s_out4 = adjust_result <s_out3>
4269 s_out0 = phi <s_loop> # (scalar) EXIT_PHI
4270 v_out1 = phi <VECT_DEF> # NEW_EXIT_PHI
4271 v_out2 = reduce <v_out1>
4272 s_out3 = extract_field <v_out2, 0>
4273 s_out4 = adjust_result <s_out3>
4278 /* In SLP reduction chain we reduce vector results into one vector if
4279 necessary, hence we set here GROUP_SIZE to 1. SCALAR_DEST is the LHS of
4280 the last stmt in the reduction chain, since we are looking for the loop
4282 if (GROUP_FIRST_ELEMENT (vinfo_for_stmt (stmt
)))
4284 scalar_dest
= gimple_assign_lhs (
4285 SLP_TREE_SCALAR_STMTS (slp_node
)[group_size
- 1]);
4289 /* In SLP we may have several statements in NEW_PHIS and REDUCTION_PHIS (in
4290 case that GROUP_SIZE is greater than vectorization factor). Therefore, we
4291 need to match SCALAR_RESULTS with corresponding statements. The first
4292 (GROUP_SIZE / number of new vector stmts) scalar results correspond to
4293 the first vector stmt, etc.
4294 (RATIO is equal to (GROUP_SIZE / number of new vector stmts)). */
4295 if (group_size
> new_phis
.length ())
4297 ratio
= group_size
/ new_phis
.length ();
4298 gcc_assert (!(group_size
% new_phis
.length ()));
4303 for (k
= 0; k
< group_size
; k
++)
4307 epilog_stmt
= new_phis
[k
/ ratio
];
4308 reduction_phi
= reduction_phis
[k
/ ratio
];
4310 inner_phi
= inner_phis
[k
/ ratio
];
4315 gimple current_stmt
= SLP_TREE_SCALAR_STMTS (slp_node
)[k
];
4317 orig_stmt
= STMT_VINFO_RELATED_STMT (vinfo_for_stmt (current_stmt
));
4318 /* SLP statements can't participate in patterns. */
4319 gcc_assert (!orig_stmt
);
4320 scalar_dest
= gimple_assign_lhs (current_stmt
);
4324 /* Find the loop-closed-use at the loop exit of the original scalar
4325 result. (The reduction result is expected to have two immediate uses -
4326 one at the latch block, and one at the loop exit). */
4327 FOR_EACH_IMM_USE_FAST (use_p
, imm_iter
, scalar_dest
)
4328 if (!flow_bb_inside_loop_p (loop
, gimple_bb (USE_STMT (use_p
))))
4329 phis
.safe_push (USE_STMT (use_p
));
4331 /* We expect to have found an exit_phi because of loop-closed-ssa
4333 gcc_assert (!phis
.is_empty ());
4335 FOR_EACH_VEC_ELT (phis
, i
, exit_phi
)
4339 stmt_vec_info exit_phi_vinfo
= vinfo_for_stmt (exit_phi
);
4342 /* FORNOW. Currently not supporting the case that an inner-loop
4343 reduction is not used in the outer-loop (but only outside the
4344 outer-loop), unless it is double reduction. */
4345 gcc_assert ((STMT_VINFO_RELEVANT_P (exit_phi_vinfo
)
4346 && !STMT_VINFO_LIVE_P (exit_phi_vinfo
))
4349 STMT_VINFO_VEC_STMT (exit_phi_vinfo
) = epilog_stmt
;
4351 || STMT_VINFO_DEF_TYPE (exit_phi_vinfo
)
4352 != vect_double_reduction_def
)
4355 /* Handle double reduction:
4357 stmt1: s1 = phi <s0, s2> - double reduction phi (outer loop)
4358 stmt2: s3 = phi <s1, s4> - (regular) reduc phi (inner loop)
4359 stmt3: s4 = use (s3) - (regular) reduc stmt (inner loop)
4360 stmt4: s2 = phi <s4> - double reduction stmt (outer loop)
4362 At that point the regular reduction (stmt2 and stmt3) is
4363 already vectorized, as well as the exit phi node, stmt4.
4364 Here we vectorize the phi node of double reduction, stmt1, and
4365 update all relevant statements. */
4367 /* Go through all the uses of s2 to find double reduction phi
4368 node, i.e., stmt1 above. */
4369 orig_name
= PHI_RESULT (exit_phi
);
4370 FOR_EACH_IMM_USE_STMT (use_stmt
, imm_iter
, orig_name
)
4372 stmt_vec_info use_stmt_vinfo
;
4373 stmt_vec_info new_phi_vinfo
;
4374 tree vect_phi_init
, preheader_arg
, vect_phi_res
, init_def
;
4375 basic_block bb
= gimple_bb (use_stmt
);
4378 /* Check that USE_STMT is really double reduction phi
4380 if (gimple_code (use_stmt
) != GIMPLE_PHI
4381 || gimple_phi_num_args (use_stmt
) != 2
4382 || bb
->loop_father
!= outer_loop
)
4384 use_stmt_vinfo
= vinfo_for_stmt (use_stmt
);
4386 || STMT_VINFO_DEF_TYPE (use_stmt_vinfo
)
4387 != vect_double_reduction_def
)
4390 /* Create vector phi node for double reduction:
4391 vs1 = phi <vs0, vs2>
4392 vs1 was created previously in this function by a call to
4393 vect_get_vec_def_for_operand and is stored in
4395 vs2 is defined by INNER_PHI, the vectorized EXIT_PHI;
4396 vs0 is created here. */
4398 /* Create vector phi node. */
4399 vect_phi
= create_phi_node (vec_initial_def
, bb
);
4400 new_phi_vinfo
= new_stmt_vec_info (vect_phi
,
4401 loop_vec_info_for_loop (outer_loop
), NULL
);
4402 set_vinfo_for_stmt (vect_phi
, new_phi_vinfo
);
4404 /* Create vs0 - initial def of the double reduction phi. */
4405 preheader_arg
= PHI_ARG_DEF_FROM_EDGE (use_stmt
,
4406 loop_preheader_edge (outer_loop
));
4407 init_def
= get_initial_def_for_reduction (stmt
,
4408 preheader_arg
, NULL
);
4409 vect_phi_init
= vect_init_vector (use_stmt
, init_def
,
4412 /* Update phi node arguments with vs0 and vs2. */
4413 add_phi_arg (vect_phi
, vect_phi_init
,
4414 loop_preheader_edge (outer_loop
),
4416 add_phi_arg (vect_phi
, PHI_RESULT (inner_phi
),
4417 loop_latch_edge (outer_loop
), UNKNOWN_LOCATION
);
4418 if (dump_enabled_p ())
4420 dump_printf_loc (MSG_NOTE
, vect_location
,
4421 "created double reduction phi node: ");
4422 dump_gimple_stmt (MSG_NOTE
, TDF_SLIM
, vect_phi
, 0);
4425 vect_phi_res
= PHI_RESULT (vect_phi
);
4427 /* Replace the use, i.e., set the correct vs1 in the regular
4428 reduction phi node. FORNOW, NCOPIES is always 1, so the
4429 loop is redundant. */
4430 use
= reduction_phi
;
4431 for (j
= 0; j
< ncopies
; j
++)
4433 edge pr_edge
= loop_preheader_edge (loop
);
4434 SET_PHI_ARG_DEF (use
, pr_edge
->dest_idx
, vect_phi_res
);
4435 use
= STMT_VINFO_RELATED_STMT (vinfo_for_stmt (use
));
4442 if (nested_in_vect_loop
)
4451 /* Find the loop-closed-use at the loop exit of the original scalar
4452 result. (The reduction result is expected to have two immediate uses,
4453 one at the latch block, and one at the loop exit). For double
4454 reductions we are looking for exit phis of the outer loop. */
4455 FOR_EACH_IMM_USE_FAST (use_p
, imm_iter
, scalar_dest
)
4457 if (!flow_bb_inside_loop_p (loop
, gimple_bb (USE_STMT (use_p
))))
4458 phis
.safe_push (USE_STMT (use_p
));
4461 if (double_reduc
&& gimple_code (USE_STMT (use_p
)) == GIMPLE_PHI
)
4463 tree phi_res
= PHI_RESULT (USE_STMT (use_p
));
4465 FOR_EACH_IMM_USE_FAST (phi_use_p
, phi_imm_iter
, phi_res
)
4467 if (!flow_bb_inside_loop_p (loop
,
4468 gimple_bb (USE_STMT (phi_use_p
))))
4469 phis
.safe_push (USE_STMT (phi_use_p
));
4475 FOR_EACH_VEC_ELT (phis
, i
, exit_phi
)
4477 /* Replace the uses: */
4478 orig_name
= PHI_RESULT (exit_phi
);
4479 scalar_result
= scalar_results
[k
];
4480 FOR_EACH_IMM_USE_STMT (use_stmt
, imm_iter
, orig_name
)
4481 FOR_EACH_IMM_USE_ON_STMT (use_p
, imm_iter
)
4482 SET_USE (use_p
, scalar_result
);
4488 scalar_results
.release ();
4489 new_phis
.release ();
4493 /* Function vectorizable_reduction.
4495 Check if STMT performs a reduction operation that can be vectorized.
4496 If VEC_STMT is also passed, vectorize the STMT: create a vectorized
4497 stmt to replace it, put it in VEC_STMT, and insert it at GSI.
4498 Return FALSE if not a vectorizable STMT, TRUE otherwise.
4500 This function also handles reduction idioms (patterns) that have been
4501 recognized in advance during vect_pattern_recog. In this case, STMT may be
4503 X = pattern_expr (arg0, arg1, ..., X)
4504 and it's STMT_VINFO_RELATED_STMT points to the last stmt in the original
4505 sequence that had been detected and replaced by the pattern-stmt (STMT).
4507 In some cases of reduction patterns, the type of the reduction variable X is
4508 different than the type of the other arguments of STMT.
4509 In such cases, the vectype that is used when transforming STMT into a vector
4510 stmt is different than the vectype that is used to determine the
4511 vectorization factor, because it consists of a different number of elements
4512 than the actual number of elements that are being operated upon in parallel.
4514 For example, consider an accumulation of shorts into an int accumulator.
4515 On some targets it's possible to vectorize this pattern operating on 8
4516 shorts at a time (hence, the vectype for purposes of determining the
4517 vectorization factor should be V8HI); on the other hand, the vectype that
4518 is used to create the vector form is actually V4SI (the type of the result).
4520 Upon entry to this function, STMT_VINFO_VECTYPE records the vectype that
4521 indicates what is the actual level of parallelism (V8HI in the example), so
4522 that the right vectorization factor would be derived. This vectype
4523 corresponds to the type of arguments to the reduction stmt, and should *NOT*
4524 be used to create the vectorized stmt. The right vectype for the vectorized
4525 stmt is obtained from the type of the result X:
4526 get_vectype_for_scalar_type (TREE_TYPE (X))
4528 This means that, contrary to "regular" reductions (or "regular" stmts in
4529 general), the following equation:
4530 STMT_VINFO_VECTYPE == get_vectype_for_scalar_type (TREE_TYPE (X))
4531 does *NOT* necessarily hold for reduction patterns. */
4534 vectorizable_reduction (gimple stmt
, gimple_stmt_iterator
*gsi
,
4535 gimple
*vec_stmt
, slp_tree slp_node
)
4539 tree loop_vec_def0
= NULL_TREE
, loop_vec_def1
= NULL_TREE
;
4540 stmt_vec_info stmt_info
= vinfo_for_stmt (stmt
);
4541 tree vectype_out
= STMT_VINFO_VECTYPE (stmt_info
);
4542 tree vectype_in
= NULL_TREE
;
4543 loop_vec_info loop_vinfo
= STMT_VINFO_LOOP_VINFO (stmt_info
);
4544 struct loop
*loop
= LOOP_VINFO_LOOP (loop_vinfo
);
4545 enum tree_code code
, orig_code
, epilog_reduc_code
;
4546 enum machine_mode vec_mode
;
4548 optab optab
, reduc_optab
;
4549 tree new_temp
= NULL_TREE
;
4552 enum vect_def_type dt
;
4553 gimple new_phi
= NULL
;
4557 stmt_vec_info orig_stmt_info
;
4558 tree expr
= NULL_TREE
;
4562 stmt_vec_info prev_stmt_info
, prev_phi_info
;
4563 bool single_defuse_cycle
= false;
4564 tree reduc_def
= NULL_TREE
;
4565 gimple new_stmt
= NULL
;
4568 bool nested_cycle
= false, found_nested_cycle_def
= false;
4569 gimple reduc_def_stmt
= NULL
;
4570 /* The default is that the reduction variable is the last in statement. */
4571 int reduc_index
= 2;
4572 bool double_reduc
= false, dummy
;
4574 struct loop
* def_stmt_loop
, *outer_loop
= NULL
;
4576 gimple def_arg_stmt
;
4577 vec
<tree
> vec_oprnds0
= vNULL
;
4578 vec
<tree
> vec_oprnds1
= vNULL
;
4579 vec
<tree
> vect_defs
= vNULL
;
4580 vec
<gimple
> phis
= vNULL
;
4582 tree def0
, def1
, tem
, op0
, op1
= NULL_TREE
;
4584 /* In case of reduction chain we switch to the first stmt in the chain, but
4585 we don't update STMT_INFO, since only the last stmt is marked as reduction
4586 and has reduction properties. */
4587 if (GROUP_FIRST_ELEMENT (vinfo_for_stmt (stmt
)))
4588 stmt
= GROUP_FIRST_ELEMENT (stmt_info
);
4590 if (nested_in_vect_loop_p (loop
, stmt
))
4594 nested_cycle
= true;
4597 /* 1. Is vectorizable reduction? */
4598 /* Not supportable if the reduction variable is used in the loop, unless
4599 it's a reduction chain. */
4600 if (STMT_VINFO_RELEVANT (stmt_info
) > vect_used_in_outer
4601 && !GROUP_FIRST_ELEMENT (stmt_info
))
4604 /* Reductions that are not used even in an enclosing outer-loop,
4605 are expected to be "live" (used out of the loop). */
4606 if (STMT_VINFO_RELEVANT (stmt_info
) == vect_unused_in_scope
4607 && !STMT_VINFO_LIVE_P (stmt_info
))
4610 /* Make sure it was already recognized as a reduction computation. */
4611 if (STMT_VINFO_DEF_TYPE (stmt_info
) != vect_reduction_def
4612 && STMT_VINFO_DEF_TYPE (stmt_info
) != vect_nested_cycle
)
4615 /* 2. Has this been recognized as a reduction pattern?
4617 Check if STMT represents a pattern that has been recognized
4618 in earlier analysis stages. For stmts that represent a pattern,
4619 the STMT_VINFO_RELATED_STMT field records the last stmt in
4620 the original sequence that constitutes the pattern. */
4622 orig_stmt
= STMT_VINFO_RELATED_STMT (stmt_info
);
4625 orig_stmt_info
= vinfo_for_stmt (orig_stmt
);
4626 gcc_assert (STMT_VINFO_IN_PATTERN_P (orig_stmt_info
));
4627 gcc_assert (!STMT_VINFO_IN_PATTERN_P (stmt_info
));
4630 /* 3. Check the operands of the operation. The first operands are defined
4631 inside the loop body. The last operand is the reduction variable,
4632 which is defined by the loop-header-phi. */
4634 gcc_assert (is_gimple_assign (stmt
));
4637 switch (get_gimple_rhs_class (gimple_assign_rhs_code (stmt
)))
4639 case GIMPLE_SINGLE_RHS
:
4640 op_type
= TREE_OPERAND_LENGTH (gimple_assign_rhs1 (stmt
));
4641 if (op_type
== ternary_op
)
4643 tree rhs
= gimple_assign_rhs1 (stmt
);
4644 ops
[0] = TREE_OPERAND (rhs
, 0);
4645 ops
[1] = TREE_OPERAND (rhs
, 1);
4646 ops
[2] = TREE_OPERAND (rhs
, 2);
4647 code
= TREE_CODE (rhs
);
4653 case GIMPLE_BINARY_RHS
:
4654 code
= gimple_assign_rhs_code (stmt
);
4655 op_type
= TREE_CODE_LENGTH (code
);
4656 gcc_assert (op_type
== binary_op
);
4657 ops
[0] = gimple_assign_rhs1 (stmt
);
4658 ops
[1] = gimple_assign_rhs2 (stmt
);
4661 case GIMPLE_TERNARY_RHS
:
4662 code
= gimple_assign_rhs_code (stmt
);
4663 op_type
= TREE_CODE_LENGTH (code
);
4664 gcc_assert (op_type
== ternary_op
);
4665 ops
[0] = gimple_assign_rhs1 (stmt
);
4666 ops
[1] = gimple_assign_rhs2 (stmt
);
4667 ops
[2] = gimple_assign_rhs3 (stmt
);
4670 case GIMPLE_UNARY_RHS
:
4677 if (code
== COND_EXPR
&& slp_node
)
4680 scalar_dest
= gimple_assign_lhs (stmt
);
4681 scalar_type
= TREE_TYPE (scalar_dest
);
4682 if (!POINTER_TYPE_P (scalar_type
) && !INTEGRAL_TYPE_P (scalar_type
)
4683 && !SCALAR_FLOAT_TYPE_P (scalar_type
))
4686 /* Do not try to vectorize bit-precision reductions. */
4687 if ((TYPE_PRECISION (scalar_type
)
4688 != GET_MODE_PRECISION (TYPE_MODE (scalar_type
))))
4691 /* All uses but the last are expected to be defined in the loop.
4692 The last use is the reduction variable. In case of nested cycle this
4693 assumption is not true: we use reduc_index to record the index of the
4694 reduction variable. */
4695 for (i
= 0; i
< op_type
-1; i
++)
4697 /* The condition of COND_EXPR is checked in vectorizable_condition(). */
4698 if (i
== 0 && code
== COND_EXPR
)
4701 is_simple_use
= vect_is_simple_use_1 (ops
[i
], stmt
, loop_vinfo
, NULL
,
4702 &def_stmt
, &def
, &dt
, &tem
);
4705 gcc_assert (is_simple_use
);
4707 if (dt
!= vect_internal_def
4708 && dt
!= vect_external_def
4709 && dt
!= vect_constant_def
4710 && dt
!= vect_induction_def
4711 && !(dt
== vect_nested_cycle
&& nested_cycle
))
4714 if (dt
== vect_nested_cycle
)
4716 found_nested_cycle_def
= true;
4717 reduc_def_stmt
= def_stmt
;
4722 is_simple_use
= vect_is_simple_use_1 (ops
[i
], stmt
, loop_vinfo
, NULL
,
4723 &def_stmt
, &def
, &dt
, &tem
);
4726 gcc_assert (is_simple_use
);
4727 gcc_assert (dt
== vect_reduction_def
4728 || dt
== vect_nested_cycle
4729 || ((dt
== vect_internal_def
|| dt
== vect_external_def
4730 || dt
== vect_constant_def
|| dt
== vect_induction_def
)
4731 && nested_cycle
&& found_nested_cycle_def
));
4732 if (!found_nested_cycle_def
)
4733 reduc_def_stmt
= def_stmt
;
4735 gcc_assert (gimple_code (reduc_def_stmt
) == GIMPLE_PHI
);
4737 gcc_assert (orig_stmt
== vect_is_simple_reduction (loop_vinfo
,
4743 gimple tmp
= vect_is_simple_reduction (loop_vinfo
, reduc_def_stmt
,
4744 !nested_cycle
, &dummy
);
4745 /* We changed STMT to be the first stmt in reduction chain, hence we
4746 check that in this case the first element in the chain is STMT. */
4747 gcc_assert (stmt
== tmp
4748 || GROUP_FIRST_ELEMENT (vinfo_for_stmt (tmp
)) == stmt
);
4751 if (STMT_VINFO_LIVE_P (vinfo_for_stmt (reduc_def_stmt
)))
4754 if (slp_node
|| PURE_SLP_STMT (stmt_info
))
4757 ncopies
= (LOOP_VINFO_VECT_FACTOR (loop_vinfo
)
4758 / TYPE_VECTOR_SUBPARTS (vectype_in
));
4760 gcc_assert (ncopies
>= 1);
4762 vec_mode
= TYPE_MODE (vectype_in
);
4764 if (code
== COND_EXPR
)
4766 if (!vectorizable_condition (stmt
, gsi
, NULL
, ops
[reduc_index
], 0, NULL
))
4768 if (dump_enabled_p ())
4769 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
4770 "unsupported condition in reduction");
4777 /* 4. Supportable by target? */
4779 /* 4.1. check support for the operation in the loop */
4780 optab
= optab_for_tree_code (code
, vectype_in
, optab_default
);
4783 if (dump_enabled_p ())
4784 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
4790 if (optab_handler (optab
, vec_mode
) == CODE_FOR_nothing
)
4792 if (dump_enabled_p ())
4793 dump_printf (MSG_NOTE
, "op not supported by target.");
4795 if (GET_MODE_SIZE (vec_mode
) != UNITS_PER_WORD
4796 || LOOP_VINFO_VECT_FACTOR (loop_vinfo
)
4797 < vect_min_worthwhile_factor (code
))
4800 if (dump_enabled_p ())
4801 dump_printf (MSG_NOTE
, "proceeding using word mode.");
4804 /* Worthwhile without SIMD support? */
4805 if (!VECTOR_MODE_P (TYPE_MODE (vectype_in
))
4806 && LOOP_VINFO_VECT_FACTOR (loop_vinfo
)
4807 < vect_min_worthwhile_factor (code
))
4809 if (dump_enabled_p ())
4810 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
4811 "not worthwhile without SIMD support.");
4817 /* 4.2. Check support for the epilog operation.
4819 If STMT represents a reduction pattern, then the type of the
4820 reduction variable may be different than the type of the rest
4821 of the arguments. For example, consider the case of accumulation
4822 of shorts into an int accumulator; The original code:
4823 S1: int_a = (int) short_a;
4824 orig_stmt-> S2: int_acc = plus <int_a ,int_acc>;
4827 STMT: int_acc = widen_sum <short_a, int_acc>
4830 1. The tree-code that is used to create the vector operation in the
4831 epilog code (that reduces the partial results) is not the
4832 tree-code of STMT, but is rather the tree-code of the original
4833 stmt from the pattern that STMT is replacing. I.e, in the example
4834 above we want to use 'widen_sum' in the loop, but 'plus' in the
4836 2. The type (mode) we use to check available target support
4837 for the vector operation to be created in the *epilog*, is
4838 determined by the type of the reduction variable (in the example
4839 above we'd check this: optab_handler (plus_optab, vect_int_mode])).
4840 However the type (mode) we use to check available target support
4841 for the vector operation to be created *inside the loop*, is
4842 determined by the type of the other arguments to STMT (in the
4843 example we'd check this: optab_handler (widen_sum_optab,
4846 This is contrary to "regular" reductions, in which the types of all
4847 the arguments are the same as the type of the reduction variable.
4848 For "regular" reductions we can therefore use the same vector type
4849 (and also the same tree-code) when generating the epilog code and
4850 when generating the code inside the loop. */
4854 /* This is a reduction pattern: get the vectype from the type of the
4855 reduction variable, and get the tree-code from orig_stmt. */
4856 orig_code
= gimple_assign_rhs_code (orig_stmt
);
4857 gcc_assert (vectype_out
);
4858 vec_mode
= TYPE_MODE (vectype_out
);
4862 /* Regular reduction: use the same vectype and tree-code as used for
4863 the vector code inside the loop can be used for the epilog code. */
4869 def_bb
= gimple_bb (reduc_def_stmt
);
4870 def_stmt_loop
= def_bb
->loop_father
;
4871 def_arg
= PHI_ARG_DEF_FROM_EDGE (reduc_def_stmt
,
4872 loop_preheader_edge (def_stmt_loop
));
4873 if (TREE_CODE (def_arg
) == SSA_NAME
4874 && (def_arg_stmt
= SSA_NAME_DEF_STMT (def_arg
))
4875 && gimple_code (def_arg_stmt
) == GIMPLE_PHI
4876 && flow_bb_inside_loop_p (outer_loop
, gimple_bb (def_arg_stmt
))
4877 && vinfo_for_stmt (def_arg_stmt
)
4878 && STMT_VINFO_DEF_TYPE (vinfo_for_stmt (def_arg_stmt
))
4879 == vect_double_reduction_def
)
4880 double_reduc
= true;
4883 epilog_reduc_code
= ERROR_MARK
;
4884 if (reduction_code_for_scalar_code (orig_code
, &epilog_reduc_code
))
4886 reduc_optab
= optab_for_tree_code (epilog_reduc_code
, vectype_out
,
4890 if (dump_enabled_p ())
4891 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
4892 "no optab for reduction.");
4894 epilog_reduc_code
= ERROR_MARK
;
4898 && optab_handler (reduc_optab
, vec_mode
) == CODE_FOR_nothing
)
4900 if (dump_enabled_p ())
4901 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
4902 "reduc op not supported by target.");
4904 epilog_reduc_code
= ERROR_MARK
;
4909 if (!nested_cycle
|| double_reduc
)
4911 if (dump_enabled_p ())
4912 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
4913 "no reduc code for scalar code.");
4919 if (double_reduc
&& ncopies
> 1)
4921 if (dump_enabled_p ())
4922 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
4923 "multiple types in double reduction");
4928 /* In case of widenning multiplication by a constant, we update the type
4929 of the constant to be the type of the other operand. We check that the
4930 constant fits the type in the pattern recognition pass. */
4931 if (code
== DOT_PROD_EXPR
4932 && !types_compatible_p (TREE_TYPE (ops
[0]), TREE_TYPE (ops
[1])))
4934 if (TREE_CODE (ops
[0]) == INTEGER_CST
)
4935 ops
[0] = fold_convert (TREE_TYPE (ops
[1]), ops
[0]);
4936 else if (TREE_CODE (ops
[1]) == INTEGER_CST
)
4937 ops
[1] = fold_convert (TREE_TYPE (ops
[0]), ops
[1]);
4940 if (dump_enabled_p ())
4941 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
4942 "invalid types in dot-prod");
4948 if (!vec_stmt
) /* transformation not required. */
4950 if (!vect_model_reduction_cost (stmt_info
, epilog_reduc_code
, ncopies
))
4952 STMT_VINFO_TYPE (stmt_info
) = reduc_vec_info_type
;
4958 if (dump_enabled_p ())
4959 dump_printf_loc (MSG_NOTE
, vect_location
, "transform reduction.");
4961 /* FORNOW: Multiple types are not supported for condition. */
4962 if (code
== COND_EXPR
)
4963 gcc_assert (ncopies
== 1);
4965 /* Create the destination vector */
4966 vec_dest
= vect_create_destination_var (scalar_dest
, vectype_out
);
4968 /* In case the vectorization factor (VF) is bigger than the number
4969 of elements that we can fit in a vectype (nunits), we have to generate
4970 more than one vector stmt - i.e - we need to "unroll" the
4971 vector stmt by a factor VF/nunits. For more details see documentation
4972 in vectorizable_operation. */
4974 /* If the reduction is used in an outer loop we need to generate
4975 VF intermediate results, like so (e.g. for ncopies=2):
4980 (i.e. we generate VF results in 2 registers).
4981 In this case we have a separate def-use cycle for each copy, and therefore
4982 for each copy we get the vector def for the reduction variable from the
4983 respective phi node created for this copy.
4985 Otherwise (the reduction is unused in the loop nest), we can combine
4986 together intermediate results, like so (e.g. for ncopies=2):
4990 (i.e. we generate VF/2 results in a single register).
4991 In this case for each copy we get the vector def for the reduction variable
4992 from the vectorized reduction operation generated in the previous iteration.
4995 if (STMT_VINFO_RELEVANT (stmt_info
) == vect_unused_in_scope
)
4997 single_defuse_cycle
= true;
5001 epilog_copies
= ncopies
;
5003 prev_stmt_info
= NULL
;
5004 prev_phi_info
= NULL
;
5007 vec_num
= SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node
);
5008 gcc_assert (TYPE_VECTOR_SUBPARTS (vectype_out
)
5009 == TYPE_VECTOR_SUBPARTS (vectype_in
));
5014 vec_oprnds0
.create (1);
5015 if (op_type
== ternary_op
)
5016 vec_oprnds1
.create (1);
5019 phis
.create (vec_num
);
5020 vect_defs
.create (vec_num
);
5022 vect_defs
.quick_push (NULL_TREE
);
5024 for (j
= 0; j
< ncopies
; j
++)
5026 if (j
== 0 || !single_defuse_cycle
)
5028 for (i
= 0; i
< vec_num
; i
++)
5030 /* Create the reduction-phi that defines the reduction
5032 new_phi
= create_phi_node (vec_dest
, loop
->header
);
5033 set_vinfo_for_stmt (new_phi
,
5034 new_stmt_vec_info (new_phi
, loop_vinfo
,
5036 if (j
== 0 || slp_node
)
5037 phis
.quick_push (new_phi
);
5041 if (code
== COND_EXPR
)
5043 gcc_assert (!slp_node
);
5044 vectorizable_condition (stmt
, gsi
, vec_stmt
,
5045 PHI_RESULT (phis
[0]),
5047 /* Multiple types are not supported for condition. */
5054 op0
= ops
[!reduc_index
];
5055 if (op_type
== ternary_op
)
5057 if (reduc_index
== 0)
5064 vect_get_vec_defs (op0
, op1
, stmt
, &vec_oprnds0
, &vec_oprnds1
,
5068 loop_vec_def0
= vect_get_vec_def_for_operand (ops
[!reduc_index
],
5070 vec_oprnds0
.quick_push (loop_vec_def0
);
5071 if (op_type
== ternary_op
)
5073 loop_vec_def1
= vect_get_vec_def_for_operand (op1
, stmt
,
5075 vec_oprnds1
.quick_push (loop_vec_def1
);
5083 enum vect_def_type dt
;
5087 vect_is_simple_use (ops
[!reduc_index
], stmt
, loop_vinfo
, NULL
,
5088 &dummy_stmt
, &dummy
, &dt
);
5089 loop_vec_def0
= vect_get_vec_def_for_stmt_copy (dt
,
5091 vec_oprnds0
[0] = loop_vec_def0
;
5092 if (op_type
== ternary_op
)
5094 vect_is_simple_use (op1
, stmt
, loop_vinfo
, NULL
, &dummy_stmt
,
5096 loop_vec_def1
= vect_get_vec_def_for_stmt_copy (dt
,
5098 vec_oprnds1
[0] = loop_vec_def1
;
5102 if (single_defuse_cycle
)
5103 reduc_def
= gimple_assign_lhs (new_stmt
);
5105 STMT_VINFO_RELATED_STMT (prev_phi_info
) = new_phi
;
5108 FOR_EACH_VEC_ELT (vec_oprnds0
, i
, def0
)
5111 reduc_def
= PHI_RESULT (phis
[i
]);
5114 if (!single_defuse_cycle
|| j
== 0)
5115 reduc_def
= PHI_RESULT (new_phi
);
5118 def1
= ((op_type
== ternary_op
)
5119 ? vec_oprnds1
[i
] : NULL
);
5120 if (op_type
== binary_op
)
5122 if (reduc_index
== 0)
5123 expr
= build2 (code
, vectype_out
, reduc_def
, def0
);
5125 expr
= build2 (code
, vectype_out
, def0
, reduc_def
);
5129 if (reduc_index
== 0)
5130 expr
= build3 (code
, vectype_out
, reduc_def
, def0
, def1
);
5133 if (reduc_index
== 1)
5134 expr
= build3 (code
, vectype_out
, def0
, reduc_def
, def1
);
5136 expr
= build3 (code
, vectype_out
, def0
, def1
, reduc_def
);
5140 new_stmt
= gimple_build_assign (vec_dest
, expr
);
5141 new_temp
= make_ssa_name (vec_dest
, new_stmt
);
5142 gimple_assign_set_lhs (new_stmt
, new_temp
);
5143 vect_finish_stmt_generation (stmt
, new_stmt
, gsi
);
5147 SLP_TREE_VEC_STMTS (slp_node
).quick_push (new_stmt
);
5148 vect_defs
.quick_push (new_temp
);
5151 vect_defs
[0] = new_temp
;
5158 STMT_VINFO_VEC_STMT (stmt_info
) = *vec_stmt
= new_stmt
;
5160 STMT_VINFO_RELATED_STMT (prev_stmt_info
) = new_stmt
;
5162 prev_stmt_info
= vinfo_for_stmt (new_stmt
);
5163 prev_phi_info
= vinfo_for_stmt (new_phi
);
5166 /* Finalize the reduction-phi (set its arguments) and create the
5167 epilog reduction code. */
5168 if ((!single_defuse_cycle
|| code
== COND_EXPR
) && !slp_node
)
5170 new_temp
= gimple_assign_lhs (*vec_stmt
);
5171 vect_defs
[0] = new_temp
;
5174 vect_create_epilog_for_reduction (vect_defs
, stmt
, epilog_copies
,
5175 epilog_reduc_code
, phis
, reduc_index
,
5176 double_reduc
, slp_node
);
5179 vec_oprnds0
.release ();
5180 vec_oprnds1
.release ();
5185 /* Function vect_min_worthwhile_factor.
5187 For a loop where we could vectorize the operation indicated by CODE,
5188 return the minimum vectorization factor that makes it worthwhile
5189 to use generic vectors. */
5191 vect_min_worthwhile_factor (enum tree_code code
)
5212 /* Function vectorizable_induction
5214 Check if PHI performs an induction computation that can be vectorized.
5215 If VEC_STMT is also passed, vectorize the induction PHI: create a vectorized
5216 phi to replace it, put it in VEC_STMT, and add it to the same basic block.
5217 Return FALSE if not a vectorizable STMT, TRUE otherwise. */
5220 vectorizable_induction (gimple phi
, gimple_stmt_iterator
*gsi ATTRIBUTE_UNUSED
,
5223 stmt_vec_info stmt_info
= vinfo_for_stmt (phi
);
5224 tree vectype
= STMT_VINFO_VECTYPE (stmt_info
);
5225 loop_vec_info loop_vinfo
= STMT_VINFO_LOOP_VINFO (stmt_info
);
5226 struct loop
*loop
= LOOP_VINFO_LOOP (loop_vinfo
);
5227 int nunits
= TYPE_VECTOR_SUBPARTS (vectype
);
5228 int ncopies
= LOOP_VINFO_VECT_FACTOR (loop_vinfo
) / nunits
;
5231 gcc_assert (ncopies
>= 1);
5232 /* FORNOW. These restrictions should be relaxed. */
5233 if (nested_in_vect_loop_p (loop
, phi
))
5235 imm_use_iterator imm_iter
;
5236 use_operand_p use_p
;
5243 if (dump_enabled_p ())
5244 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
5245 "multiple types in nested loop.");
5250 latch_e
= loop_latch_edge (loop
->inner
);
5251 loop_arg
= PHI_ARG_DEF_FROM_EDGE (phi
, latch_e
);
5252 FOR_EACH_IMM_USE_FAST (use_p
, imm_iter
, loop_arg
)
5254 if (!flow_bb_inside_loop_p (loop
->inner
,
5255 gimple_bb (USE_STMT (use_p
))))
5257 exit_phi
= USE_STMT (use_p
);
5263 stmt_vec_info exit_phi_vinfo
= vinfo_for_stmt (exit_phi
);
5264 if (!(STMT_VINFO_RELEVANT_P (exit_phi_vinfo
)
5265 && !STMT_VINFO_LIVE_P (exit_phi_vinfo
)))
5267 if (dump_enabled_p ())
5268 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
5269 "inner-loop induction only used outside "
5270 "of the outer vectorized loop.");
5276 if (!STMT_VINFO_RELEVANT_P (stmt_info
))
5279 /* FORNOW: SLP not supported. */
5280 if (STMT_SLP_TYPE (stmt_info
))
5283 gcc_assert (STMT_VINFO_DEF_TYPE (stmt_info
) == vect_induction_def
);
5285 if (gimple_code (phi
) != GIMPLE_PHI
)
5288 if (!vec_stmt
) /* transformation not required. */
5290 STMT_VINFO_TYPE (stmt_info
) = induc_vec_info_type
;
5291 if (dump_enabled_p ())
5292 dump_printf_loc (MSG_NOTE
, vect_location
,
5293 "=== vectorizable_induction ===");
5294 vect_model_induction_cost (stmt_info
, ncopies
);
5300 if (dump_enabled_p ())
5301 dump_printf_loc (MSG_NOTE
, vect_location
, "transform induction phi.");
5303 vec_def
= get_initial_def_for_induction (phi
);
5304 *vec_stmt
= SSA_NAME_DEF_STMT (vec_def
);
5308 /* Function vectorizable_live_operation.
5310 STMT computes a value that is used outside the loop. Check if
5311 it can be supported. */
5314 vectorizable_live_operation (gimple stmt
,
5315 gimple_stmt_iterator
*gsi ATTRIBUTE_UNUSED
,
5316 gimple
*vec_stmt ATTRIBUTE_UNUSED
)
5318 stmt_vec_info stmt_info
= vinfo_for_stmt (stmt
);
5319 loop_vec_info loop_vinfo
= STMT_VINFO_LOOP_VINFO (stmt_info
);
5320 struct loop
*loop
= LOOP_VINFO_LOOP (loop_vinfo
);
5326 enum vect_def_type dt
;
5327 enum tree_code code
;
5328 enum gimple_rhs_class rhs_class
;
5330 gcc_assert (STMT_VINFO_LIVE_P (stmt_info
));
5332 if (STMT_VINFO_DEF_TYPE (stmt_info
) == vect_reduction_def
)
5335 if (!is_gimple_assign (stmt
))
5338 if (TREE_CODE (gimple_assign_lhs (stmt
)) != SSA_NAME
)
5341 /* FORNOW. CHECKME. */
5342 if (nested_in_vect_loop_p (loop
, stmt
))
5345 code
= gimple_assign_rhs_code (stmt
);
5346 op_type
= TREE_CODE_LENGTH (code
);
5347 rhs_class
= get_gimple_rhs_class (code
);
5348 gcc_assert (rhs_class
!= GIMPLE_UNARY_RHS
|| op_type
== unary_op
);
5349 gcc_assert (rhs_class
!= GIMPLE_BINARY_RHS
|| op_type
== binary_op
);
5351 /* FORNOW: support only if all uses are invariant. This means
5352 that the scalar operations can remain in place, unvectorized.
5353 The original last scalar value that they compute will be used. */
5355 for (i
= 0; i
< op_type
; i
++)
5357 if (rhs_class
== GIMPLE_SINGLE_RHS
)
5358 op
= TREE_OPERAND (gimple_op (stmt
, 1), i
);
5360 op
= gimple_op (stmt
, i
+ 1);
5362 && !vect_is_simple_use (op
, stmt
, loop_vinfo
, NULL
, &def_stmt
, &def
,
5365 if (dump_enabled_p ())
5366 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
5371 if (dt
!= vect_external_def
&& dt
!= vect_constant_def
)
5375 /* No transformation is required for the cases we currently support. */
5379 /* Kill any debug uses outside LOOP of SSA names defined in STMT. */
5382 vect_loop_kill_debug_uses (struct loop
*loop
, gimple stmt
)
5384 ssa_op_iter op_iter
;
5385 imm_use_iterator imm_iter
;
5386 def_operand_p def_p
;
5389 FOR_EACH_PHI_OR_STMT_DEF (def_p
, stmt
, op_iter
, SSA_OP_DEF
)
5391 FOR_EACH_IMM_USE_STMT (ustmt
, imm_iter
, DEF_FROM_PTR (def_p
))
5395 if (!is_gimple_debug (ustmt
))
5398 bb
= gimple_bb (ustmt
);
5400 if (!flow_bb_inside_loop_p (loop
, bb
))
5402 if (gimple_debug_bind_p (ustmt
))
5404 if (dump_enabled_p ())
5405 dump_printf_loc (MSG_NOTE
, vect_location
,
5406 "killing debug use");
5408 gimple_debug_bind_reset_value (ustmt
);
5409 update_stmt (ustmt
);
5418 /* Function vect_transform_loop.
5420 The analysis phase has determined that the loop is vectorizable.
5421 Vectorize the loop - created vectorized stmts to replace the scalar
5422 stmts in the loop, and update the loop exit condition. */
5425 vect_transform_loop (loop_vec_info loop_vinfo
)
5427 struct loop
*loop
= LOOP_VINFO_LOOP (loop_vinfo
);
5428 basic_block
*bbs
= LOOP_VINFO_BBS (loop_vinfo
);
5429 int nbbs
= loop
->num_nodes
;
5430 gimple_stmt_iterator si
;
5433 int vectorization_factor
= LOOP_VINFO_VECT_FACTOR (loop_vinfo
);
5435 bool slp_scheduled
= false;
5436 unsigned int nunits
;
5437 gimple stmt
, pattern_stmt
;
5438 gimple_seq pattern_def_seq
= NULL
;
5439 gimple_stmt_iterator pattern_def_si
= gsi_none ();
5440 bool transform_pattern_stmt
= false;
5441 bool check_profitability
= false;
5443 /* Record number of iterations before we started tampering with the profile. */
5444 gcov_type expected_iterations
= expected_loop_iterations_unbounded (loop
);
5446 if (dump_enabled_p ())
5447 dump_printf_loc (MSG_NOTE
, vect_location
, "=== vec_transform_loop ===");
5449 /* If profile is inprecise, we have chance to fix it up. */
5450 if (LOOP_VINFO_NITERS_KNOWN_P (loop_vinfo
))
5451 expected_iterations
= LOOP_VINFO_INT_NITERS (loop_vinfo
);
5453 /* Use the more conservative vectorization threshold. If the number
5454 of iterations is constant assume the cost check has been performed
5455 by our caller. If the threshold makes all loops profitable that
5456 run at least the vectorization factor number of times checking
5457 is pointless, too. */
5458 th
= ((PARAM_VALUE (PARAM_MIN_VECT_LOOP_BOUND
)
5459 * LOOP_VINFO_VECT_FACTOR (loop_vinfo
)) - 1);
5460 th
= MAX (th
, LOOP_VINFO_COST_MODEL_MIN_ITERS (loop_vinfo
));
5461 if (th
>= LOOP_VINFO_VECT_FACTOR (loop_vinfo
) - 1
5462 && !LOOP_VINFO_NITERS_KNOWN_P (loop_vinfo
))
5464 if (dump_enabled_p ())
5465 dump_printf_loc (MSG_NOTE
, vect_location
,
5466 "Profitability threshold is %d loop iterations.", th
);
5467 check_profitability
= true;
5470 /* Peel the loop if there are data refs with unknown alignment.
5471 Only one data ref with unknown store is allowed. */
5473 if (LOOP_PEELING_FOR_ALIGNMENT (loop_vinfo
))
5475 vect_do_peeling_for_alignment (loop_vinfo
, th
, check_profitability
);
5476 check_profitability
= false;
5479 if (LOOP_REQUIRES_VERSIONING_FOR_ALIGNMENT (loop_vinfo
)
5480 || LOOP_REQUIRES_VERSIONING_FOR_ALIAS (loop_vinfo
))
5482 vect_loop_versioning (loop_vinfo
, th
, check_profitability
);
5483 check_profitability
= false;
5486 /* If the loop has a symbolic number of iterations 'n' (i.e. it's not a
5487 compile time constant), or it is a constant that doesn't divide by the
5488 vectorization factor, then an epilog loop needs to be created.
5489 We therefore duplicate the loop: the original loop will be vectorized,
5490 and will compute the first (n/VF) iterations. The second copy of the loop
5491 will remain scalar and will compute the remaining (n%VF) iterations.
5492 (VF is the vectorization factor). */
5494 if (!LOOP_VINFO_NITERS_KNOWN_P (loop_vinfo
)
5495 || (LOOP_VINFO_NITERS_KNOWN_P (loop_vinfo
)
5496 && LOOP_VINFO_INT_NITERS (loop_vinfo
) % vectorization_factor
!= 0)
5497 || LOOP_VINFO_PEELING_FOR_GAPS (loop_vinfo
))
5498 vect_do_peeling_for_loop_bound (loop_vinfo
, &ratio
,
5499 th
, check_profitability
);
5501 ratio
= build_int_cst (TREE_TYPE (LOOP_VINFO_NITERS (loop_vinfo
)),
5502 LOOP_VINFO_INT_NITERS (loop_vinfo
) / vectorization_factor
);
5504 /* 1) Make sure the loop header has exactly two entries
5505 2) Make sure we have a preheader basic block. */
5507 gcc_assert (EDGE_COUNT (loop
->header
->preds
) == 2);
5509 split_edge (loop_preheader_edge (loop
));
5511 /* FORNOW: the vectorizer supports only loops which body consist
5512 of one basic block (header + empty latch). When the vectorizer will
5513 support more involved loop forms, the order by which the BBs are
5514 traversed need to be reconsidered. */
5516 for (i
= 0; i
< nbbs
; i
++)
5518 basic_block bb
= bbs
[i
];
5519 stmt_vec_info stmt_info
;
5522 for (si
= gsi_start_phis (bb
); !gsi_end_p (si
); gsi_next (&si
))
5524 phi
= gsi_stmt (si
);
5525 if (dump_enabled_p ())
5527 dump_printf_loc (MSG_NOTE
, vect_location
,
5528 "------>vectorizing phi: ");
5529 dump_gimple_stmt (MSG_NOTE
, TDF_SLIM
, phi
, 0);
5531 stmt_info
= vinfo_for_stmt (phi
);
5535 if (MAY_HAVE_DEBUG_STMTS
&& !STMT_VINFO_LIVE_P (stmt_info
))
5536 vect_loop_kill_debug_uses (loop
, phi
);
5538 if (!STMT_VINFO_RELEVANT_P (stmt_info
)
5539 && !STMT_VINFO_LIVE_P (stmt_info
))
5542 if ((TYPE_VECTOR_SUBPARTS (STMT_VINFO_VECTYPE (stmt_info
))
5543 != (unsigned HOST_WIDE_INT
) vectorization_factor
)
5544 && dump_enabled_p ())
5545 dump_printf_loc (MSG_NOTE
, vect_location
, "multiple-types.");
5547 if (STMT_VINFO_DEF_TYPE (stmt_info
) == vect_induction_def
)
5549 if (dump_enabled_p ())
5550 dump_printf_loc (MSG_NOTE
, vect_location
, "transform phi.");
5551 vect_transform_stmt (phi
, NULL
, NULL
, NULL
, NULL
);
5555 pattern_stmt
= NULL
;
5556 for (si
= gsi_start_bb (bb
); !gsi_end_p (si
) || transform_pattern_stmt
;)
5560 if (transform_pattern_stmt
)
5561 stmt
= pattern_stmt
;
5563 stmt
= gsi_stmt (si
);
5565 if (dump_enabled_p ())
5567 dump_printf_loc (MSG_NOTE
, vect_location
,
5568 "------>vectorizing statement: ");
5569 dump_gimple_stmt (MSG_NOTE
, TDF_SLIM
, stmt
, 0);
5572 stmt_info
= vinfo_for_stmt (stmt
);
5574 /* vector stmts created in the outer-loop during vectorization of
5575 stmts in an inner-loop may not have a stmt_info, and do not
5576 need to be vectorized. */
5583 if (MAY_HAVE_DEBUG_STMTS
&& !STMT_VINFO_LIVE_P (stmt_info
))
5584 vect_loop_kill_debug_uses (loop
, stmt
);
5586 if (!STMT_VINFO_RELEVANT_P (stmt_info
)
5587 && !STMT_VINFO_LIVE_P (stmt_info
))
5589 if (STMT_VINFO_IN_PATTERN_P (stmt_info
)
5590 && (pattern_stmt
= STMT_VINFO_RELATED_STMT (stmt_info
))
5591 && (STMT_VINFO_RELEVANT_P (vinfo_for_stmt (pattern_stmt
))
5592 || STMT_VINFO_LIVE_P (vinfo_for_stmt (pattern_stmt
))))
5594 stmt
= pattern_stmt
;
5595 stmt_info
= vinfo_for_stmt (stmt
);
5603 else if (STMT_VINFO_IN_PATTERN_P (stmt_info
)
5604 && (pattern_stmt
= STMT_VINFO_RELATED_STMT (stmt_info
))
5605 && (STMT_VINFO_RELEVANT_P (vinfo_for_stmt (pattern_stmt
))
5606 || STMT_VINFO_LIVE_P (vinfo_for_stmt (pattern_stmt
))))
5607 transform_pattern_stmt
= true;
5609 /* If pattern statement has def stmts, vectorize them too. */
5610 if (is_pattern_stmt_p (stmt_info
))
5612 if (pattern_def_seq
== NULL
)
5614 pattern_def_seq
= STMT_VINFO_PATTERN_DEF_SEQ (stmt_info
);
5615 pattern_def_si
= gsi_start (pattern_def_seq
);
5617 else if (!gsi_end_p (pattern_def_si
))
5618 gsi_next (&pattern_def_si
);
5619 if (pattern_def_seq
!= NULL
)
5621 gimple pattern_def_stmt
= NULL
;
5622 stmt_vec_info pattern_def_stmt_info
= NULL
;
5624 while (!gsi_end_p (pattern_def_si
))
5626 pattern_def_stmt
= gsi_stmt (pattern_def_si
);
5627 pattern_def_stmt_info
5628 = vinfo_for_stmt (pattern_def_stmt
);
5629 if (STMT_VINFO_RELEVANT_P (pattern_def_stmt_info
)
5630 || STMT_VINFO_LIVE_P (pattern_def_stmt_info
))
5632 gsi_next (&pattern_def_si
);
5635 if (!gsi_end_p (pattern_def_si
))
5637 if (dump_enabled_p ())
5639 dump_printf_loc (MSG_NOTE
, vect_location
,
5640 "==> vectorizing pattern def "
5642 dump_gimple_stmt (MSG_NOTE
, TDF_SLIM
,
5643 pattern_def_stmt
, 0);
5646 stmt
= pattern_def_stmt
;
5647 stmt_info
= pattern_def_stmt_info
;
5651 pattern_def_si
= gsi_none ();
5652 transform_pattern_stmt
= false;
5656 transform_pattern_stmt
= false;
5659 gcc_assert (STMT_VINFO_VECTYPE (stmt_info
));
5660 nunits
= (unsigned int) TYPE_VECTOR_SUBPARTS (
5661 STMT_VINFO_VECTYPE (stmt_info
));
5662 if (!STMT_SLP_TYPE (stmt_info
)
5663 && nunits
!= (unsigned int) vectorization_factor
5664 && dump_enabled_p ())
5665 /* For SLP VF is set according to unrolling factor, and not to
5666 vector size, hence for SLP this print is not valid. */
5667 dump_printf_loc (MSG_NOTE
, vect_location
,
5670 /* SLP. Schedule all the SLP instances when the first SLP stmt is
5672 if (STMT_SLP_TYPE (stmt_info
))
5676 slp_scheduled
= true;
5678 if (dump_enabled_p ())
5679 dump_printf_loc (MSG_NOTE
, vect_location
,
5680 "=== scheduling SLP instances ===");
5682 vect_schedule_slp (loop_vinfo
, NULL
);
5685 /* Hybrid SLP stmts must be vectorized in addition to SLP. */
5686 if (!vinfo_for_stmt (stmt
) || PURE_SLP_STMT (stmt_info
))
5688 if (!transform_pattern_stmt
&& gsi_end_p (pattern_def_si
))
5690 pattern_def_seq
= NULL
;
5697 /* -------- vectorize statement ------------ */
5698 if (dump_enabled_p ())
5699 dump_printf_loc (MSG_NOTE
, vect_location
, "transform statement.");
5701 grouped_store
= false;
5702 is_store
= vect_transform_stmt (stmt
, &si
, &grouped_store
, NULL
, NULL
);
5705 if (STMT_VINFO_GROUPED_ACCESS (stmt_info
))
5707 /* Interleaving. If IS_STORE is TRUE, the vectorization of the
5708 interleaving chain was completed - free all the stores in
5711 vect_remove_stores (GROUP_FIRST_ELEMENT (stmt_info
));
5716 /* Free the attached stmt_vec_info and remove the stmt. */
5717 gimple store
= gsi_stmt (si
);
5718 free_stmt_vec_info (store
);
5719 unlink_stmt_vdef (store
);
5720 gsi_remove (&si
, true);
5721 release_defs (store
);
5726 if (!transform_pattern_stmt
&& gsi_end_p (pattern_def_si
))
5728 pattern_def_seq
= NULL
;
5734 slpeel_make_loop_iterate_ntimes (loop
, ratio
);
5736 /* Reduce loop iterations by the vectorization factor. */
5737 scale_loop_profile (loop
, RDIV (REG_BR_PROB_BASE
, vectorization_factor
),
5738 expected_iterations
/ vectorization_factor
);
5739 loop
->nb_iterations_upper_bound
5740 = loop
->nb_iterations_upper_bound
.udiv (double_int::from_uhwi (vectorization_factor
),
5742 if (LOOP_VINFO_PEELING_FOR_GAPS (loop_vinfo
)
5743 && loop
->nb_iterations_upper_bound
!= double_int_zero
)
5744 loop
->nb_iterations_upper_bound
= loop
->nb_iterations_upper_bound
- double_int_one
;
5745 if (loop
->any_estimate
)
5747 loop
->nb_iterations_estimate
5748 = loop
->nb_iterations_estimate
.udiv (double_int::from_uhwi (vectorization_factor
),
5750 if (LOOP_VINFO_PEELING_FOR_GAPS (loop_vinfo
)
5751 && loop
->nb_iterations_estimate
!= double_int_zero
)
5752 loop
->nb_iterations_estimate
= loop
->nb_iterations_estimate
- double_int_one
;
5755 /* The memory tags and pointers in vectorized statements need to
5756 have their SSA forms updated. FIXME, why can't this be delayed
5757 until all the loops have been transformed? */
5758 update_ssa (TODO_update_ssa
);
5760 if (dump_enabled_p ())
5761 dump_printf_loc (MSG_OPTIMIZED_LOCATIONS
, vect_location
, "LOOP VECTORIZED.");
5762 if (loop
->inner
&& dump_enabled_p ())
5763 dump_printf_loc (MSG_OPTIMIZED_LOCATIONS
, vect_location
,
5764 "OUTER LOOP VECTORIZED.");