2 Copyright (C) 2003, 2004, 2005, 2006, 2007, 2008, 2009, 2010, 2011, 2012
3 Free Software Foundation, Inc.
4 Contributed by Dorit Naishlos <dorit@il.ibm.com> and
5 Ira Rosen <irar@il.ibm.com>
7 This file is part of GCC.
9 GCC is free software; you can redistribute it and/or modify it under
10 the terms of the GNU General Public License as published by the Free
11 Software Foundation; either version 3, or (at your option) any later
14 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
15 WARRANTY; without even the implied warranty of MERCHANTABILITY or
16 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
19 You should have received a copy of the GNU General Public License
20 along with GCC; see the file COPYING3. If not see
21 <http://www.gnu.org/licenses/>. */
25 #include "coretypes.h"
29 #include "basic-block.h"
30 #include "tree-pretty-print.h"
31 #include "gimple-pretty-print.h"
32 #include "tree-flow.h"
33 #include "tree-dump.h"
35 #include "cfglayout.h"
40 #include "diagnostic-core.h"
41 #include "tree-chrec.h"
42 #include "tree-scalar-evolution.h"
43 #include "tree-vectorizer.h"
46 /* Loop Vectorization Pass.
48 This pass tries to vectorize loops.
50 For example, the vectorizer transforms the following simple loop:
52 short a[N]; short b[N]; short c[N]; int i;
58 as if it was manually vectorized by rewriting the source code into:
60 typedef int __attribute__((mode(V8HI))) v8hi;
61 short a[N]; short b[N]; short c[N]; int i;
62 v8hi *pa = (v8hi*)a, *pb = (v8hi*)b, *pc = (v8hi*)c;
65 for (i=0; i<N/8; i++){
72 The main entry to this pass is vectorize_loops(), in which
73 the vectorizer applies a set of analyses on a given set of loops,
74 followed by the actual vectorization transformation for the loops that
75 had successfully passed the analysis phase.
76 Throughout this pass we make a distinction between two types of
77 data: scalars (which are represented by SSA_NAMES), and memory references
78 ("data-refs"). These two types of data require different handling both
79 during analysis and transformation. The types of data-refs that the
80 vectorizer currently supports are ARRAY_REFS which base is an array DECL
81 (not a pointer), and INDIRECT_REFS through pointers; both array and pointer
82 accesses are required to have a simple (consecutive) access pattern.
86 The driver for the analysis phase is vect_analyze_loop().
87 It applies a set of analyses, some of which rely on the scalar evolution
88 analyzer (scev) developed by Sebastian Pop.
90 During the analysis phase the vectorizer records some information
91 per stmt in a "stmt_vec_info" struct which is attached to each stmt in the
92 loop, as well as general information about the loop as a whole, which is
93 recorded in a "loop_vec_info" struct attached to each loop.
97 The loop transformation phase scans all the stmts in the loop, and
98 creates a vector stmt (or a sequence of stmts) for each scalar stmt S in
99 the loop that needs to be vectorized. It inserts the vector code sequence
100 just before the scalar stmt S, and records a pointer to the vector code
101 in STMT_VINFO_VEC_STMT (stmt_info) (stmt_info is the stmt_vec_info struct
102 attached to S). This pointer will be used for the vectorization of following
103 stmts which use the def of stmt S. Stmt S is removed if it writes to memory;
104 otherwise, we rely on dead code elimination for removing it.
106 For example, say stmt S1 was vectorized into stmt VS1:
109 S1: b = x[i]; STMT_VINFO_VEC_STMT (stmt_info (S1)) = VS1
112 To vectorize stmt S2, the vectorizer first finds the stmt that defines
113 the operand 'b' (S1), and gets the relevant vector def 'vb' from the
114 vector stmt VS1 pointed to by STMT_VINFO_VEC_STMT (stmt_info (S1)). The
115 resulting sequence would be:
118 S1: b = x[i]; STMT_VINFO_VEC_STMT (stmt_info (S1)) = VS1
120 S2: a = b; STMT_VINFO_VEC_STMT (stmt_info (S2)) = VS2
122 Operands that are not SSA_NAMEs, are data-refs that appear in
123 load/store operations (like 'x[i]' in S1), and are handled differently.
127 Currently the only target specific information that is used is the
128 size of the vector (in bytes) - "TARGET_VECTORIZE_UNITS_PER_SIMD_WORD".
129 Targets that can support different sizes of vectors, for now will need
130 to specify one value for "TARGET_VECTORIZE_UNITS_PER_SIMD_WORD". More
131 flexibility will be added in the future.
133 Since we only vectorize operations which vector form can be
134 expressed using existing tree codes, to verify that an operation is
135 supported, the vectorizer checks the relevant optab at the relevant
136 machine_mode (e.g, optab_handler (add_optab, V8HImode)). If
137 the value found is CODE_FOR_nothing, then there's no target support, and
138 we can't vectorize the stmt.
140 For additional information on this project see:
141 http://gcc.gnu.org/projects/tree-ssa/vectorization.html
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_start (NULL);
187 bool analyze_pattern_stmt = false;
189 if (vect_print_dump_info (REPORT_DETAILS))
190 fprintf (vect_dump, "=== vect_determine_vectorization_factor ===");
192 for (i = 0; i < nbbs; i++)
194 basic_block bb = bbs[i];
196 for (si = gsi_start_phis (bb); !gsi_end_p (si); gsi_next (&si))
199 stmt_info = vinfo_for_stmt (phi);
200 if (vect_print_dump_info (REPORT_DETAILS))
202 fprintf (vect_dump, "==> examining phi: ");
203 print_gimple_stmt (vect_dump, phi, 0, TDF_SLIM);
206 gcc_assert (stmt_info);
208 if (STMT_VINFO_RELEVANT_P (stmt_info))
210 gcc_assert (!STMT_VINFO_VECTYPE (stmt_info));
211 scalar_type = TREE_TYPE (PHI_RESULT (phi));
213 if (vect_print_dump_info (REPORT_DETAILS))
215 fprintf (vect_dump, "get vectype for scalar type: ");
216 print_generic_expr (vect_dump, scalar_type, TDF_SLIM);
219 vectype = get_vectype_for_scalar_type (scalar_type);
222 if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS))
225 "not vectorized: unsupported data-type ");
226 print_generic_expr (vect_dump, scalar_type, TDF_SLIM);
230 STMT_VINFO_VECTYPE (stmt_info) = vectype;
232 if (vect_print_dump_info (REPORT_DETAILS))
234 fprintf (vect_dump, "vectype: ");
235 print_generic_expr (vect_dump, vectype, TDF_SLIM);
238 nunits = TYPE_VECTOR_SUBPARTS (vectype);
239 if (vect_print_dump_info (REPORT_DETAILS))
240 fprintf (vect_dump, "nunits = %d", nunits);
242 if (!vectorization_factor
243 || (nunits > vectorization_factor))
244 vectorization_factor = nunits;
248 for (si = gsi_start_bb (bb); !gsi_end_p (si) || analyze_pattern_stmt;)
252 if (analyze_pattern_stmt)
255 stmt = gsi_stmt (si);
257 stmt_info = vinfo_for_stmt (stmt);
259 if (vect_print_dump_info (REPORT_DETAILS))
261 fprintf (vect_dump, "==> examining statement: ");
262 print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM);
265 gcc_assert (stmt_info);
267 /* Skip stmts which do not need to be vectorized. */
268 if (!STMT_VINFO_RELEVANT_P (stmt_info)
269 && !STMT_VINFO_LIVE_P (stmt_info))
271 if (STMT_VINFO_IN_PATTERN_P (stmt_info)
272 && (pattern_stmt = STMT_VINFO_RELATED_STMT (stmt_info))
273 && (STMT_VINFO_RELEVANT_P (vinfo_for_stmt (pattern_stmt))
274 || STMT_VINFO_LIVE_P (vinfo_for_stmt (pattern_stmt))))
277 stmt_info = vinfo_for_stmt (pattern_stmt);
278 if (vect_print_dump_info (REPORT_DETAILS))
280 fprintf (vect_dump, "==> examining pattern statement: ");
281 print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM);
286 if (vect_print_dump_info (REPORT_DETAILS))
287 fprintf (vect_dump, "skip.");
292 else if (STMT_VINFO_IN_PATTERN_P (stmt_info)
293 && (pattern_stmt = STMT_VINFO_RELATED_STMT (stmt_info))
294 && (STMT_VINFO_RELEVANT_P (vinfo_for_stmt (pattern_stmt))
295 || STMT_VINFO_LIVE_P (vinfo_for_stmt (pattern_stmt))))
296 analyze_pattern_stmt = true;
298 /* If a pattern statement has def stmts, analyze them too. */
299 if (is_pattern_stmt_p (stmt_info))
301 if (pattern_def_seq == NULL)
303 pattern_def_seq = STMT_VINFO_PATTERN_DEF_SEQ (stmt_info);
304 pattern_def_si = gsi_start (pattern_def_seq);
306 else if (!gsi_end_p (pattern_def_si))
307 gsi_next (&pattern_def_si);
308 if (pattern_def_seq != NULL)
310 gimple pattern_def_stmt = NULL;
311 stmt_vec_info pattern_def_stmt_info = NULL;
313 while (!gsi_end_p (pattern_def_si))
315 pattern_def_stmt = gsi_stmt (pattern_def_si);
316 pattern_def_stmt_info
317 = vinfo_for_stmt (pattern_def_stmt);
318 if (STMT_VINFO_RELEVANT_P (pattern_def_stmt_info)
319 || STMT_VINFO_LIVE_P (pattern_def_stmt_info))
321 gsi_next (&pattern_def_si);
324 if (!gsi_end_p (pattern_def_si))
326 if (vect_print_dump_info (REPORT_DETAILS))
329 "==> examining pattern def stmt: ");
330 print_gimple_stmt (vect_dump, pattern_def_stmt, 0,
334 stmt = pattern_def_stmt;
335 stmt_info = pattern_def_stmt_info;
339 pattern_def_si = gsi_start (NULL);
340 analyze_pattern_stmt = false;
344 analyze_pattern_stmt = false;
347 if (gimple_get_lhs (stmt) == NULL_TREE)
349 if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS))
351 fprintf (vect_dump, "not vectorized: irregular stmt.");
352 print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM);
357 if (VECTOR_MODE_P (TYPE_MODE (gimple_expr_type (stmt))))
359 if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS))
361 fprintf (vect_dump, "not vectorized: vector stmt in loop:");
362 print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM);
367 if (STMT_VINFO_VECTYPE (stmt_info))
369 /* The only case when a vectype had been already set is for stmts
370 that contain a dataref, or for "pattern-stmts" (stmts
371 generated by the vectorizer to represent/replace a certain
373 gcc_assert (STMT_VINFO_DATA_REF (stmt_info)
374 || is_pattern_stmt_p (stmt_info)
375 || !gsi_end_p (pattern_def_si));
376 vectype = STMT_VINFO_VECTYPE (stmt_info);
380 gcc_assert (!STMT_VINFO_DATA_REF (stmt_info));
381 scalar_type = TREE_TYPE (gimple_get_lhs (stmt));
382 if (vect_print_dump_info (REPORT_DETAILS))
384 fprintf (vect_dump, "get vectype for scalar type: ");
385 print_generic_expr (vect_dump, scalar_type, TDF_SLIM);
387 vectype = get_vectype_for_scalar_type (scalar_type);
390 if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS))
393 "not vectorized: unsupported data-type ");
394 print_generic_expr (vect_dump, scalar_type, TDF_SLIM);
399 STMT_VINFO_VECTYPE (stmt_info) = vectype;
402 /* The vectorization factor is according to the smallest
403 scalar type (or the largest vector size, but we only
404 support one vector size per loop). */
405 scalar_type = vect_get_smallest_scalar_type (stmt, &dummy,
407 if (vect_print_dump_info (REPORT_DETAILS))
409 fprintf (vect_dump, "get vectype for scalar type: ");
410 print_generic_expr (vect_dump, scalar_type, TDF_SLIM);
412 vf_vectype = get_vectype_for_scalar_type (scalar_type);
415 if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS))
418 "not vectorized: unsupported data-type ");
419 print_generic_expr (vect_dump, scalar_type, TDF_SLIM);
424 if ((GET_MODE_SIZE (TYPE_MODE (vectype))
425 != GET_MODE_SIZE (TYPE_MODE (vf_vectype))))
427 if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS))
430 "not vectorized: different sized vector "
431 "types in statement, ");
432 print_generic_expr (vect_dump, vectype, TDF_SLIM);
433 fprintf (vect_dump, " and ");
434 print_generic_expr (vect_dump, vf_vectype, TDF_SLIM);
439 if (vect_print_dump_info (REPORT_DETAILS))
441 fprintf (vect_dump, "vectype: ");
442 print_generic_expr (vect_dump, vf_vectype, TDF_SLIM);
445 nunits = TYPE_VECTOR_SUBPARTS (vf_vectype);
446 if (vect_print_dump_info (REPORT_DETAILS))
447 fprintf (vect_dump, "nunits = %d", nunits);
449 if (!vectorization_factor
450 || (nunits > vectorization_factor))
451 vectorization_factor = nunits;
453 if (!analyze_pattern_stmt && gsi_end_p (pattern_def_si))
455 pattern_def_seq = NULL;
461 /* TODO: Analyze cost. Decide if worth while to vectorize. */
462 if (vect_print_dump_info (REPORT_DETAILS))
463 fprintf (vect_dump, "vectorization factor = %d", vectorization_factor);
464 if (vectorization_factor <= 1)
466 if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS))
467 fprintf (vect_dump, "not vectorized: unsupported data-type");
470 LOOP_VINFO_VECT_FACTOR (loop_vinfo) = vectorization_factor;
476 /* Function vect_is_simple_iv_evolution.
478 FORNOW: A simple evolution of an induction variables in the loop is
479 considered a polynomial evolution with constant step. */
482 vect_is_simple_iv_evolution (unsigned loop_nb, tree access_fn, tree * init,
487 tree evolution_part = evolution_part_in_loop_num (access_fn, loop_nb);
489 /* When there is no evolution in this loop, the evolution function
491 if (evolution_part == NULL_TREE)
494 /* When the evolution is a polynomial of degree >= 2
495 the evolution function is not "simple". */
496 if (tree_is_chrec (evolution_part))
499 step_expr = evolution_part;
500 init_expr = unshare_expr (initial_condition_in_loop_num (access_fn, loop_nb));
502 if (vect_print_dump_info (REPORT_DETAILS))
504 fprintf (vect_dump, "step: ");
505 print_generic_expr (vect_dump, step_expr, TDF_SLIM);
506 fprintf (vect_dump, ", init: ");
507 print_generic_expr (vect_dump, init_expr, TDF_SLIM);
513 if (TREE_CODE (step_expr) != INTEGER_CST)
515 if (vect_print_dump_info (REPORT_DETAILS))
516 fprintf (vect_dump, "step unknown.");
523 /* Function vect_analyze_scalar_cycles_1.
525 Examine the cross iteration def-use cycles of scalar variables
526 in LOOP. LOOP_VINFO represents the loop that is now being
527 considered for vectorization (can be LOOP, or an outer-loop
531 vect_analyze_scalar_cycles_1 (loop_vec_info loop_vinfo, struct loop *loop)
533 basic_block bb = loop->header;
535 VEC(gimple,heap) *worklist = VEC_alloc (gimple, heap, 64);
536 gimple_stmt_iterator gsi;
539 if (vect_print_dump_info (REPORT_DETAILS))
540 fprintf (vect_dump, "=== vect_analyze_scalar_cycles ===");
542 /* First - identify all inductions. Reduction detection assumes that all the
543 inductions have been identified, therefore, this order must not be
545 for (gsi = gsi_start_phis (bb); !gsi_end_p (gsi); gsi_next (&gsi))
547 gimple phi = gsi_stmt (gsi);
548 tree access_fn = NULL;
549 tree def = PHI_RESULT (phi);
550 stmt_vec_info stmt_vinfo = vinfo_for_stmt (phi);
552 if (vect_print_dump_info (REPORT_DETAILS))
554 fprintf (vect_dump, "Analyze phi: ");
555 print_gimple_stmt (vect_dump, phi, 0, TDF_SLIM);
558 /* Skip virtual phi's. The data dependences that are associated with
559 virtual defs/uses (i.e., memory accesses) are analyzed elsewhere. */
560 if (!is_gimple_reg (SSA_NAME_VAR (def)))
563 STMT_VINFO_DEF_TYPE (stmt_vinfo) = vect_unknown_def_type;
565 /* Analyze the evolution function. */
566 access_fn = analyze_scalar_evolution (loop, def);
569 STRIP_NOPS (access_fn);
570 if (vect_print_dump_info (REPORT_DETAILS))
572 fprintf (vect_dump, "Access function of PHI: ");
573 print_generic_expr (vect_dump, access_fn, TDF_SLIM);
575 STMT_VINFO_LOOP_PHI_EVOLUTION_PART (stmt_vinfo)
576 = evolution_part_in_loop_num (access_fn, loop->num);
580 || !vect_is_simple_iv_evolution (loop->num, access_fn, &dumy, &dumy))
582 VEC_safe_push (gimple, heap, worklist, phi);
586 gcc_assert (STMT_VINFO_LOOP_PHI_EVOLUTION_PART (stmt_vinfo) != NULL_TREE);
588 if (vect_print_dump_info (REPORT_DETAILS))
589 fprintf (vect_dump, "Detected induction.");
590 STMT_VINFO_DEF_TYPE (stmt_vinfo) = vect_induction_def;
594 /* Second - identify all reductions and nested cycles. */
595 while (VEC_length (gimple, worklist) > 0)
597 gimple phi = VEC_pop (gimple, worklist);
598 tree def = PHI_RESULT (phi);
599 stmt_vec_info stmt_vinfo = vinfo_for_stmt (phi);
603 if (vect_print_dump_info (REPORT_DETAILS))
605 fprintf (vect_dump, "Analyze phi: ");
606 print_gimple_stmt (vect_dump, phi, 0, TDF_SLIM);
609 gcc_assert (is_gimple_reg (SSA_NAME_VAR (def)));
610 gcc_assert (STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_unknown_def_type);
612 nested_cycle = (loop != LOOP_VINFO_LOOP (loop_vinfo));
613 reduc_stmt = vect_force_simple_reduction (loop_vinfo, phi, !nested_cycle,
619 if (vect_print_dump_info (REPORT_DETAILS))
620 fprintf (vect_dump, "Detected double reduction.");
622 STMT_VINFO_DEF_TYPE (stmt_vinfo) = vect_double_reduction_def;
623 STMT_VINFO_DEF_TYPE (vinfo_for_stmt (reduc_stmt)) =
624 vect_double_reduction_def;
630 if (vect_print_dump_info (REPORT_DETAILS))
631 fprintf (vect_dump, "Detected vectorizable nested cycle.");
633 STMT_VINFO_DEF_TYPE (stmt_vinfo) = vect_nested_cycle;
634 STMT_VINFO_DEF_TYPE (vinfo_for_stmt (reduc_stmt)) =
639 if (vect_print_dump_info (REPORT_DETAILS))
640 fprintf (vect_dump, "Detected reduction.");
642 STMT_VINFO_DEF_TYPE (stmt_vinfo) = vect_reduction_def;
643 STMT_VINFO_DEF_TYPE (vinfo_for_stmt (reduc_stmt)) =
645 /* Store the reduction cycles for possible vectorization in
647 VEC_safe_push (gimple, heap,
648 LOOP_VINFO_REDUCTIONS (loop_vinfo),
654 if (vect_print_dump_info (REPORT_DETAILS))
655 fprintf (vect_dump, "Unknown def-use cycle pattern.");
658 VEC_free (gimple, heap, worklist);
662 /* Function vect_analyze_scalar_cycles.
664 Examine the cross iteration def-use cycles of scalar variables, by
665 analyzing the loop-header PHIs of scalar variables. Classify each
666 cycle as one of the following: invariant, induction, reduction, unknown.
667 We do that for the loop represented by LOOP_VINFO, and also to its
668 inner-loop, if exists.
669 Examples for scalar cycles:
684 vect_analyze_scalar_cycles (loop_vec_info loop_vinfo)
686 struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
688 vect_analyze_scalar_cycles_1 (loop_vinfo, loop);
690 /* When vectorizing an outer-loop, the inner-loop is executed sequentially.
691 Reductions in such inner-loop therefore have different properties than
692 the reductions in the nest that gets vectorized:
693 1. When vectorized, they are executed in the same order as in the original
694 scalar loop, so we can't change the order of computation when
696 2. FIXME: Inner-loop reductions can be used in the inner-loop, so the
697 current checks are too strict. */
700 vect_analyze_scalar_cycles_1 (loop_vinfo, loop->inner);
703 /* Function vect_get_loop_niters.
705 Determine how many iterations the loop is executed.
706 If an expression that represents the number of iterations
707 can be constructed, place it in NUMBER_OF_ITERATIONS.
708 Return the loop exit condition. */
711 vect_get_loop_niters (struct loop *loop, tree *number_of_iterations)
715 if (vect_print_dump_info (REPORT_DETAILS))
716 fprintf (vect_dump, "=== get_loop_niters ===");
718 niters = number_of_exit_cond_executions (loop);
720 if (niters != NULL_TREE
721 && niters != chrec_dont_know)
723 *number_of_iterations = niters;
725 if (vect_print_dump_info (REPORT_DETAILS))
727 fprintf (vect_dump, "==> get_loop_niters:" );
728 print_generic_expr (vect_dump, *number_of_iterations, TDF_SLIM);
732 return get_loop_exit_condition (loop);
736 /* Function bb_in_loop_p
738 Used as predicate for dfs order traversal of the loop bbs. */
741 bb_in_loop_p (const_basic_block bb, const void *data)
743 const struct loop *const loop = (const struct loop *)data;
744 if (flow_bb_inside_loop_p (loop, bb))
750 /* Function new_loop_vec_info.
752 Create and initialize a new loop_vec_info struct for LOOP, as well as
753 stmt_vec_info structs for all the stmts in LOOP. */
756 new_loop_vec_info (struct loop *loop)
760 gimple_stmt_iterator si;
761 unsigned int i, nbbs;
763 res = (loop_vec_info) xcalloc (1, sizeof (struct _loop_vec_info));
764 LOOP_VINFO_LOOP (res) = loop;
766 bbs = get_loop_body (loop);
768 /* Create/Update stmt_info for all stmts in the loop. */
769 for (i = 0; i < loop->num_nodes; i++)
771 basic_block bb = bbs[i];
773 /* BBs in a nested inner-loop will have been already processed (because
774 we will have called vect_analyze_loop_form for any nested inner-loop).
775 Therefore, for stmts in an inner-loop we just want to update the
776 STMT_VINFO_LOOP_VINFO field of their stmt_info to point to the new
777 loop_info of the outer-loop we are currently considering to vectorize
778 (instead of the loop_info of the inner-loop).
779 For stmts in other BBs we need to create a stmt_info from scratch. */
780 if (bb->loop_father != loop)
783 gcc_assert (loop->inner && bb->loop_father == loop->inner);
784 for (si = gsi_start_phis (bb); !gsi_end_p (si); gsi_next (&si))
786 gimple phi = gsi_stmt (si);
787 stmt_vec_info stmt_info = vinfo_for_stmt (phi);
788 loop_vec_info inner_loop_vinfo =
789 STMT_VINFO_LOOP_VINFO (stmt_info);
790 gcc_assert (loop->inner == LOOP_VINFO_LOOP (inner_loop_vinfo));
791 STMT_VINFO_LOOP_VINFO (stmt_info) = res;
793 for (si = gsi_start_bb (bb); !gsi_end_p (si); gsi_next (&si))
795 gimple stmt = gsi_stmt (si);
796 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
797 loop_vec_info inner_loop_vinfo =
798 STMT_VINFO_LOOP_VINFO (stmt_info);
799 gcc_assert (loop->inner == LOOP_VINFO_LOOP (inner_loop_vinfo));
800 STMT_VINFO_LOOP_VINFO (stmt_info) = res;
805 /* bb in current nest. */
806 for (si = gsi_start_phis (bb); !gsi_end_p (si); gsi_next (&si))
808 gimple phi = gsi_stmt (si);
809 gimple_set_uid (phi, 0);
810 set_vinfo_for_stmt (phi, new_stmt_vec_info (phi, res, NULL));
813 for (si = gsi_start_bb (bb); !gsi_end_p (si); gsi_next (&si))
815 gimple stmt = gsi_stmt (si);
816 gimple_set_uid (stmt, 0);
817 set_vinfo_for_stmt (stmt, new_stmt_vec_info (stmt, res, NULL));
822 /* CHECKME: We want to visit all BBs before their successors (except for
823 latch blocks, for which this assertion wouldn't hold). In the simple
824 case of the loop forms we allow, a dfs order of the BBs would the same
825 as reversed postorder traversal, so we are safe. */
828 bbs = XCNEWVEC (basic_block, loop->num_nodes);
829 nbbs = dfs_enumerate_from (loop->header, 0, bb_in_loop_p,
830 bbs, loop->num_nodes, loop);
831 gcc_assert (nbbs == loop->num_nodes);
833 LOOP_VINFO_BBS (res) = bbs;
834 LOOP_VINFO_NITERS (res) = NULL;
835 LOOP_VINFO_NITERS_UNCHANGED (res) = NULL;
836 LOOP_VINFO_COST_MODEL_MIN_ITERS (res) = 0;
837 LOOP_VINFO_VECTORIZABLE_P (res) = 0;
838 LOOP_PEELING_FOR_ALIGNMENT (res) = 0;
839 LOOP_VINFO_VECT_FACTOR (res) = 0;
840 LOOP_VINFO_LOOP_NEST (res) = VEC_alloc (loop_p, heap, 3);
841 LOOP_VINFO_DATAREFS (res) = VEC_alloc (data_reference_p, heap, 10);
842 LOOP_VINFO_DDRS (res) = VEC_alloc (ddr_p, heap, 10 * 10);
843 LOOP_VINFO_UNALIGNED_DR (res) = NULL;
844 LOOP_VINFO_MAY_MISALIGN_STMTS (res) =
845 VEC_alloc (gimple, heap,
846 PARAM_VALUE (PARAM_VECT_MAX_VERSION_FOR_ALIGNMENT_CHECKS));
847 LOOP_VINFO_MAY_ALIAS_DDRS (res) =
848 VEC_alloc (ddr_p, heap,
849 PARAM_VALUE (PARAM_VECT_MAX_VERSION_FOR_ALIAS_CHECKS));
850 LOOP_VINFO_STRIDED_STORES (res) = VEC_alloc (gimple, heap, 10);
851 LOOP_VINFO_REDUCTIONS (res) = VEC_alloc (gimple, heap, 10);
852 LOOP_VINFO_REDUCTION_CHAINS (res) = VEC_alloc (gimple, heap, 10);
853 LOOP_VINFO_SLP_INSTANCES (res) = VEC_alloc (slp_instance, heap, 10);
854 LOOP_VINFO_SLP_UNROLLING_FACTOR (res) = 1;
855 LOOP_VINFO_PEELING_HTAB (res) = NULL;
856 LOOP_VINFO_PEELING_FOR_GAPS (res) = false;
862 /* Function destroy_loop_vec_info.
864 Free LOOP_VINFO struct, as well as all the stmt_vec_info structs of all the
865 stmts in the loop. */
868 destroy_loop_vec_info (loop_vec_info loop_vinfo, bool clean_stmts)
873 gimple_stmt_iterator si;
875 VEC (slp_instance, heap) *slp_instances;
876 slp_instance instance;
881 loop = LOOP_VINFO_LOOP (loop_vinfo);
883 bbs = LOOP_VINFO_BBS (loop_vinfo);
884 nbbs = loop->num_nodes;
888 free (LOOP_VINFO_BBS (loop_vinfo));
889 free_data_refs (LOOP_VINFO_DATAREFS (loop_vinfo));
890 free_dependence_relations (LOOP_VINFO_DDRS (loop_vinfo));
891 VEC_free (loop_p, heap, LOOP_VINFO_LOOP_NEST (loop_vinfo));
892 VEC_free (gimple, heap, LOOP_VINFO_MAY_MISALIGN_STMTS (loop_vinfo));
893 VEC_free (ddr_p, heap, LOOP_VINFO_MAY_ALIAS_DDRS (loop_vinfo));
900 for (j = 0; j < nbbs; j++)
902 basic_block bb = bbs[j];
903 for (si = gsi_start_phis (bb); !gsi_end_p (si); gsi_next (&si))
904 free_stmt_vec_info (gsi_stmt (si));
906 for (si = gsi_start_bb (bb); !gsi_end_p (si); )
908 gimple stmt = gsi_stmt (si);
909 /* Free stmt_vec_info. */
910 free_stmt_vec_info (stmt);
915 free (LOOP_VINFO_BBS (loop_vinfo));
916 free_data_refs (LOOP_VINFO_DATAREFS (loop_vinfo));
917 free_dependence_relations (LOOP_VINFO_DDRS (loop_vinfo));
918 VEC_free (loop_p, heap, LOOP_VINFO_LOOP_NEST (loop_vinfo));
919 VEC_free (gimple, heap, LOOP_VINFO_MAY_MISALIGN_STMTS (loop_vinfo));
920 VEC_free (ddr_p, heap, LOOP_VINFO_MAY_ALIAS_DDRS (loop_vinfo));
921 slp_instances = LOOP_VINFO_SLP_INSTANCES (loop_vinfo);
922 FOR_EACH_VEC_ELT (slp_instance, slp_instances, j, instance)
923 vect_free_slp_instance (instance);
925 VEC_free (slp_instance, heap, LOOP_VINFO_SLP_INSTANCES (loop_vinfo));
926 VEC_free (gimple, heap, LOOP_VINFO_STRIDED_STORES (loop_vinfo));
927 VEC_free (gimple, heap, LOOP_VINFO_REDUCTIONS (loop_vinfo));
928 VEC_free (gimple, heap, LOOP_VINFO_REDUCTION_CHAINS (loop_vinfo));
930 if (LOOP_VINFO_PEELING_HTAB (loop_vinfo))
931 htab_delete (LOOP_VINFO_PEELING_HTAB (loop_vinfo));
938 /* Function vect_analyze_loop_1.
940 Apply a set of analyses on LOOP, and create a loop_vec_info struct
941 for it. The different analyses will record information in the
942 loop_vec_info struct. This is a subset of the analyses applied in
943 vect_analyze_loop, to be applied on an inner-loop nested in the loop
944 that is now considered for (outer-loop) vectorization. */
947 vect_analyze_loop_1 (struct loop *loop)
949 loop_vec_info loop_vinfo;
951 if (vect_print_dump_info (REPORT_DETAILS))
952 fprintf (vect_dump, "===== analyze_loop_nest_1 =====");
954 /* Check the CFG characteristics of the loop (nesting, entry/exit, etc. */
956 loop_vinfo = vect_analyze_loop_form (loop);
959 if (vect_print_dump_info (REPORT_DETAILS))
960 fprintf (vect_dump, "bad inner-loop form.");
968 /* Function vect_analyze_loop_form.
970 Verify that certain CFG restrictions hold, including:
971 - the loop has a pre-header
972 - the loop has a single entry and exit
973 - the loop exit condition is simple enough, and the number of iterations
974 can be analyzed (a countable loop). */
977 vect_analyze_loop_form (struct loop *loop)
979 loop_vec_info loop_vinfo;
981 tree number_of_iterations = NULL;
982 loop_vec_info inner_loop_vinfo = NULL;
984 if (vect_print_dump_info (REPORT_DETAILS))
985 fprintf (vect_dump, "=== vect_analyze_loop_form ===");
987 /* Different restrictions apply when we are considering an inner-most loop,
988 vs. an outer (nested) loop.
989 (FORNOW. May want to relax some of these restrictions in the future). */
993 /* Inner-most loop. We currently require that the number of BBs is
994 exactly 2 (the header and latch). Vectorizable inner-most loops
1005 if (loop->num_nodes != 2)
1007 if (vect_print_dump_info (REPORT_BAD_FORM_LOOPS))
1008 fprintf (vect_dump, "not vectorized: control flow in loop.");
1012 if (empty_block_p (loop->header))
1014 if (vect_print_dump_info (REPORT_BAD_FORM_LOOPS))
1015 fprintf (vect_dump, "not vectorized: empty loop.");
1021 struct loop *innerloop = loop->inner;
1024 /* Nested loop. We currently require that the loop is doubly-nested,
1025 contains a single inner loop, and the number of BBs is exactly 5.
1026 Vectorizable outer-loops look like this:
1038 The inner-loop has the properties expected of inner-most loops
1039 as described above. */
1041 if ((loop->inner)->inner || (loop->inner)->next)
1043 if (vect_print_dump_info (REPORT_BAD_FORM_LOOPS))
1044 fprintf (vect_dump, "not vectorized: multiple nested loops.");
1048 /* Analyze the inner-loop. */
1049 inner_loop_vinfo = vect_analyze_loop_1 (loop->inner);
1050 if (!inner_loop_vinfo)
1052 if (vect_print_dump_info (REPORT_BAD_FORM_LOOPS))
1053 fprintf (vect_dump, "not vectorized: Bad inner loop.");
1057 if (!expr_invariant_in_loop_p (loop,
1058 LOOP_VINFO_NITERS (inner_loop_vinfo)))
1060 if (vect_print_dump_info (REPORT_BAD_FORM_LOOPS))
1062 "not vectorized: inner-loop count not invariant.");
1063 destroy_loop_vec_info (inner_loop_vinfo, true);
1067 if (loop->num_nodes != 5)
1069 if (vect_print_dump_info (REPORT_BAD_FORM_LOOPS))
1070 fprintf (vect_dump, "not vectorized: control flow in loop.");
1071 destroy_loop_vec_info (inner_loop_vinfo, true);
1075 gcc_assert (EDGE_COUNT (innerloop->header->preds) == 2);
1076 entryedge = EDGE_PRED (innerloop->header, 0);
1077 if (EDGE_PRED (innerloop->header, 0)->src == innerloop->latch)
1078 entryedge = EDGE_PRED (innerloop->header, 1);
1080 if (entryedge->src != loop->header
1081 || !single_exit (innerloop)
1082 || single_exit (innerloop)->dest != EDGE_PRED (loop->latch, 0)->src)
1084 if (vect_print_dump_info (REPORT_BAD_FORM_LOOPS))
1085 fprintf (vect_dump, "not vectorized: unsupported outerloop form.");
1086 destroy_loop_vec_info (inner_loop_vinfo, true);
1090 if (vect_print_dump_info (REPORT_DETAILS))
1091 fprintf (vect_dump, "Considering outer-loop vectorization.");
1094 if (!single_exit (loop)
1095 || EDGE_COUNT (loop->header->preds) != 2)
1097 if (vect_print_dump_info (REPORT_BAD_FORM_LOOPS))
1099 if (!single_exit (loop))
1100 fprintf (vect_dump, "not vectorized: multiple exits.");
1101 else if (EDGE_COUNT (loop->header->preds) != 2)
1102 fprintf (vect_dump, "not vectorized: too many incoming edges.");
1104 if (inner_loop_vinfo)
1105 destroy_loop_vec_info (inner_loop_vinfo, true);
1109 /* We assume that the loop exit condition is at the end of the loop. i.e,
1110 that the loop is represented as a do-while (with a proper if-guard
1111 before the loop if needed), where the loop header contains all the
1112 executable statements, and the latch is empty. */
1113 if (!empty_block_p (loop->latch)
1114 || !gimple_seq_empty_p (phi_nodes (loop->latch)))
1116 if (vect_print_dump_info (REPORT_BAD_FORM_LOOPS))
1117 fprintf (vect_dump, "not vectorized: unexpected loop form.");
1118 if (inner_loop_vinfo)
1119 destroy_loop_vec_info (inner_loop_vinfo, true);
1123 /* Make sure there exists a single-predecessor exit bb: */
1124 if (!single_pred_p (single_exit (loop)->dest))
1126 edge e = single_exit (loop);
1127 if (!(e->flags & EDGE_ABNORMAL))
1129 split_loop_exit_edge (e);
1130 if (vect_print_dump_info (REPORT_DETAILS))
1131 fprintf (vect_dump, "split exit edge.");
1135 if (vect_print_dump_info (REPORT_BAD_FORM_LOOPS))
1136 fprintf (vect_dump, "not vectorized: abnormal loop exit edge.");
1137 if (inner_loop_vinfo)
1138 destroy_loop_vec_info (inner_loop_vinfo, true);
1143 loop_cond = vect_get_loop_niters (loop, &number_of_iterations);
1146 if (vect_print_dump_info (REPORT_BAD_FORM_LOOPS))
1147 fprintf (vect_dump, "not vectorized: complicated exit condition.");
1148 if (inner_loop_vinfo)
1149 destroy_loop_vec_info (inner_loop_vinfo, true);
1153 if (!number_of_iterations)
1155 if (vect_print_dump_info (REPORT_BAD_FORM_LOOPS))
1157 "not vectorized: number of iterations cannot be computed.");
1158 if (inner_loop_vinfo)
1159 destroy_loop_vec_info (inner_loop_vinfo, true);
1163 if (chrec_contains_undetermined (number_of_iterations))
1165 if (vect_print_dump_info (REPORT_BAD_FORM_LOOPS))
1166 fprintf (vect_dump, "Infinite number of iterations.");
1167 if (inner_loop_vinfo)
1168 destroy_loop_vec_info (inner_loop_vinfo, true);
1172 if (!NITERS_KNOWN_P (number_of_iterations))
1174 if (vect_print_dump_info (REPORT_DETAILS))
1176 fprintf (vect_dump, "Symbolic number of iterations is ");
1177 print_generic_expr (vect_dump, number_of_iterations, TDF_DETAILS);
1180 else if (TREE_INT_CST_LOW (number_of_iterations) == 0)
1182 if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS))
1183 fprintf (vect_dump, "not vectorized: number of iterations = 0.");
1184 if (inner_loop_vinfo)
1185 destroy_loop_vec_info (inner_loop_vinfo, false);
1189 loop_vinfo = new_loop_vec_info (loop);
1190 LOOP_VINFO_NITERS (loop_vinfo) = number_of_iterations;
1191 LOOP_VINFO_NITERS_UNCHANGED (loop_vinfo) = number_of_iterations;
1193 STMT_VINFO_TYPE (vinfo_for_stmt (loop_cond)) = loop_exit_ctrl_vec_info_type;
1195 /* CHECKME: May want to keep it around it in the future. */
1196 if (inner_loop_vinfo)
1197 destroy_loop_vec_info (inner_loop_vinfo, false);
1199 gcc_assert (!loop->aux);
1200 loop->aux = loop_vinfo;
1205 /* Get cost by calling cost target builtin. */
1208 vect_get_cost (enum vect_cost_for_stmt type_of_cost)
1210 tree dummy_type = NULL;
1213 return targetm.vectorize.builtin_vectorization_cost (type_of_cost,
1218 /* Function vect_analyze_loop_operations.
1220 Scan the loop stmts and make sure they are all vectorizable. */
1223 vect_analyze_loop_operations (loop_vec_info loop_vinfo, bool slp)
1225 struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
1226 basic_block *bbs = LOOP_VINFO_BBS (loop_vinfo);
1227 int nbbs = loop->num_nodes;
1228 gimple_stmt_iterator si;
1229 unsigned int vectorization_factor = 0;
1232 stmt_vec_info stmt_info;
1233 bool need_to_vectorize = false;
1234 int min_profitable_iters;
1235 int min_scalar_loop_bound;
1237 bool only_slp_in_loop = true, ok;
1239 if (vect_print_dump_info (REPORT_DETAILS))
1240 fprintf (vect_dump, "=== vect_analyze_loop_operations ===");
1242 gcc_assert (LOOP_VINFO_VECT_FACTOR (loop_vinfo));
1243 vectorization_factor = LOOP_VINFO_VECT_FACTOR (loop_vinfo);
1246 /* If all the stmts in the loop can be SLPed, we perform only SLP, and
1247 vectorization factor of the loop is the unrolling factor required by
1248 the SLP instances. If that unrolling factor is 1, we say, that we
1249 perform pure SLP on loop - cross iteration parallelism is not
1251 for (i = 0; i < nbbs; i++)
1253 basic_block bb = bbs[i];
1254 for (si = gsi_start_bb (bb); !gsi_end_p (si); gsi_next (&si))
1256 gimple stmt = gsi_stmt (si);
1257 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
1258 gcc_assert (stmt_info);
1259 if ((STMT_VINFO_RELEVANT_P (stmt_info)
1260 || VECTORIZABLE_CYCLE_DEF (STMT_VINFO_DEF_TYPE (stmt_info)))
1261 && !PURE_SLP_STMT (stmt_info))
1262 /* STMT needs both SLP and loop-based vectorization. */
1263 only_slp_in_loop = false;
1267 if (only_slp_in_loop)
1268 vectorization_factor = LOOP_VINFO_SLP_UNROLLING_FACTOR (loop_vinfo);
1270 vectorization_factor = least_common_multiple (vectorization_factor,
1271 LOOP_VINFO_SLP_UNROLLING_FACTOR (loop_vinfo));
1273 LOOP_VINFO_VECT_FACTOR (loop_vinfo) = vectorization_factor;
1274 if (vect_print_dump_info (REPORT_DETAILS))
1275 fprintf (vect_dump, "Updating vectorization factor to %d ",
1276 vectorization_factor);
1279 for (i = 0; i < nbbs; i++)
1281 basic_block bb = bbs[i];
1283 for (si = gsi_start_phis (bb); !gsi_end_p (si); gsi_next (&si))
1285 phi = gsi_stmt (si);
1288 stmt_info = vinfo_for_stmt (phi);
1289 if (vect_print_dump_info (REPORT_DETAILS))
1291 fprintf (vect_dump, "examining phi: ");
1292 print_gimple_stmt (vect_dump, phi, 0, TDF_SLIM);
1295 /* Inner-loop loop-closed exit phi in outer-loop vectorization
1296 (i.e., a phi in the tail of the outer-loop). */
1297 if (! is_loop_header_bb_p (bb))
1299 /* FORNOW: we currently don't support the case that these phis
1300 are not used in the outerloop (unless it is double reduction,
1301 i.e., this phi is vect_reduction_def), cause this case
1302 requires to actually do something here. */
1303 if ((!STMT_VINFO_RELEVANT_P (stmt_info)
1304 || STMT_VINFO_LIVE_P (stmt_info))
1305 && STMT_VINFO_DEF_TYPE (stmt_info)
1306 != vect_double_reduction_def)
1308 if (vect_print_dump_info (REPORT_DETAILS))
1310 "Unsupported loop-closed phi in outer-loop.");
1314 /* If PHI is used in the outer loop, we check that its operand
1315 is defined in the inner loop. */
1316 if (STMT_VINFO_RELEVANT_P (stmt_info))
1321 if (gimple_phi_num_args (phi) != 1)
1324 phi_op = PHI_ARG_DEF (phi, 0);
1325 if (TREE_CODE (phi_op) != SSA_NAME)
1328 op_def_stmt = SSA_NAME_DEF_STMT (phi_op);
1330 || !flow_bb_inside_loop_p (loop, gimple_bb (op_def_stmt))
1331 || !vinfo_for_stmt (op_def_stmt))
1334 if (STMT_VINFO_RELEVANT (vinfo_for_stmt (op_def_stmt))
1335 != vect_used_in_outer
1336 && STMT_VINFO_RELEVANT (vinfo_for_stmt (op_def_stmt))
1337 != vect_used_in_outer_by_reduction)
1344 gcc_assert (stmt_info);
1346 if (STMT_VINFO_LIVE_P (stmt_info))
1348 /* FORNOW: not yet supported. */
1349 if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS))
1350 fprintf (vect_dump, "not vectorized: value used after loop.");
1354 if (STMT_VINFO_RELEVANT (stmt_info) == vect_used_in_scope
1355 && STMT_VINFO_DEF_TYPE (stmt_info) != vect_induction_def)
1357 /* A scalar-dependence cycle that we don't support. */
1358 if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS))
1359 fprintf (vect_dump, "not vectorized: scalar dependence cycle.");
1363 if (STMT_VINFO_RELEVANT_P (stmt_info))
1365 need_to_vectorize = true;
1366 if (STMT_VINFO_DEF_TYPE (stmt_info) == vect_induction_def)
1367 ok = vectorizable_induction (phi, NULL, NULL);
1372 if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS))
1375 "not vectorized: relevant phi not supported: ");
1376 print_gimple_stmt (vect_dump, phi, 0, TDF_SLIM);
1382 for (si = gsi_start_bb (bb); !gsi_end_p (si); gsi_next (&si))
1384 gimple stmt = gsi_stmt (si);
1385 if (!vect_analyze_stmt (stmt, &need_to_vectorize, NULL))
1390 /* All operations in the loop are either irrelevant (deal with loop
1391 control, or dead), or only used outside the loop and can be moved
1392 out of the loop (e.g. invariants, inductions). The loop can be
1393 optimized away by scalar optimizations. We're better off not
1394 touching this loop. */
1395 if (!need_to_vectorize)
1397 if (vect_print_dump_info (REPORT_DETAILS))
1399 "All the computation can be taken out of the loop.");
1400 if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS))
1402 "not vectorized: redundant loop. no profit to vectorize.");
1406 if (LOOP_VINFO_NITERS_KNOWN_P (loop_vinfo)
1407 && vect_print_dump_info (REPORT_DETAILS))
1409 "vectorization_factor = %d, niters = " HOST_WIDE_INT_PRINT_DEC,
1410 vectorization_factor, LOOP_VINFO_INT_NITERS (loop_vinfo));
1412 if (LOOP_VINFO_NITERS_KNOWN_P (loop_vinfo)
1413 && (LOOP_VINFO_INT_NITERS (loop_vinfo) < vectorization_factor))
1415 if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS))
1416 fprintf (vect_dump, "not vectorized: iteration count too small.");
1417 if (vect_print_dump_info (REPORT_DETAILS))
1418 fprintf (vect_dump,"not vectorized: iteration count smaller than "
1419 "vectorization factor.");
1423 /* Analyze cost. Decide if worth while to vectorize. */
1425 /* Once VF is set, SLP costs should be updated since the number of created
1426 vector stmts depends on VF. */
1427 vect_update_slp_costs_according_to_vf (loop_vinfo);
1429 min_profitable_iters = vect_estimate_min_profitable_iters (loop_vinfo);
1430 LOOP_VINFO_COST_MODEL_MIN_ITERS (loop_vinfo) = min_profitable_iters;
1432 if (min_profitable_iters < 0)
1434 if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS))
1435 fprintf (vect_dump, "not vectorized: vectorization not profitable.");
1436 if (vect_print_dump_info (REPORT_DETAILS))
1437 fprintf (vect_dump, "not vectorized: vector version will never be "
1442 min_scalar_loop_bound = ((PARAM_VALUE (PARAM_MIN_VECT_LOOP_BOUND)
1443 * vectorization_factor) - 1);
1445 /* Use the cost model only if it is more conservative than user specified
1448 th = (unsigned) min_scalar_loop_bound;
1449 if (min_profitable_iters
1450 && (!min_scalar_loop_bound
1451 || min_profitable_iters > min_scalar_loop_bound))
1452 th = (unsigned) min_profitable_iters;
1454 if (LOOP_VINFO_NITERS_KNOWN_P (loop_vinfo)
1455 && LOOP_VINFO_INT_NITERS (loop_vinfo) <= th)
1457 if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS))
1458 fprintf (vect_dump, "not vectorized: vectorization not "
1460 if (vect_print_dump_info (REPORT_DETAILS))
1461 fprintf (vect_dump, "not vectorized: iteration count smaller than "
1462 "user specified loop bound parameter or minimum "
1463 "profitable iterations (whichever is more conservative).");
1467 if (!LOOP_VINFO_NITERS_KNOWN_P (loop_vinfo)
1468 || LOOP_VINFO_INT_NITERS (loop_vinfo) % vectorization_factor != 0
1469 || LOOP_PEELING_FOR_ALIGNMENT (loop_vinfo))
1471 if (vect_print_dump_info (REPORT_DETAILS))
1472 fprintf (vect_dump, "epilog loop required.");
1473 if (!vect_can_advance_ivs_p (loop_vinfo))
1475 if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS))
1477 "not vectorized: can't create epilog loop 1.");
1480 if (!slpeel_can_duplicate_loop_p (loop, single_exit (loop)))
1482 if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS))
1484 "not vectorized: can't create epilog loop 2.");
1493 /* Function vect_analyze_loop_2.
1495 Apply a set of analyses on LOOP, and create a loop_vec_info struct
1496 for it. The different analyses will record information in the
1497 loop_vec_info struct. */
1499 vect_analyze_loop_2 (loop_vec_info loop_vinfo)
1501 bool ok, slp = false;
1502 int max_vf = MAX_VECTORIZATION_FACTOR;
1505 /* Find all data references in the loop (which correspond to vdefs/vuses)
1506 and analyze their evolution in the loop. Also adjust the minimal
1507 vectorization factor according to the loads and stores.
1509 FORNOW: Handle only simple, array references, which
1510 alignment can be forced, and aligned pointer-references. */
1512 ok = vect_analyze_data_refs (loop_vinfo, NULL, &min_vf);
1515 if (vect_print_dump_info (REPORT_DETAILS))
1516 fprintf (vect_dump, "bad data references.");
1520 /* Classify all cross-iteration scalar data-flow cycles.
1521 Cross-iteration cycles caused by virtual phis are analyzed separately. */
1523 vect_analyze_scalar_cycles (loop_vinfo);
1525 vect_pattern_recog (loop_vinfo);
1527 /* Data-flow analysis to detect stmts that do not need to be vectorized. */
1529 ok = vect_mark_stmts_to_be_vectorized (loop_vinfo);
1532 if (vect_print_dump_info (REPORT_DETAILS))
1533 fprintf (vect_dump, "unexpected pattern.");
1537 /* Analyze data dependences between the data-refs in the loop
1538 and adjust the maximum vectorization factor according to
1540 FORNOW: fail at the first data dependence that we encounter. */
1542 ok = vect_analyze_data_ref_dependences (loop_vinfo, NULL, &max_vf);
1546 if (vect_print_dump_info (REPORT_DETAILS))
1547 fprintf (vect_dump, "bad data dependence.");
1551 ok = vect_determine_vectorization_factor (loop_vinfo);
1554 if (vect_print_dump_info (REPORT_DETAILS))
1555 fprintf (vect_dump, "can't determine vectorization factor.");
1558 if (max_vf < LOOP_VINFO_VECT_FACTOR (loop_vinfo))
1560 if (vect_print_dump_info (REPORT_DETAILS))
1561 fprintf (vect_dump, "bad data dependence.");
1565 /* Analyze the alignment of the data-refs in the loop.
1566 Fail if a data reference is found that cannot be vectorized. */
1568 ok = vect_analyze_data_refs_alignment (loop_vinfo, NULL);
1571 if (vect_print_dump_info (REPORT_DETAILS))
1572 fprintf (vect_dump, "bad data alignment.");
1576 /* Analyze the access patterns of the data-refs in the loop (consecutive,
1577 complex, etc.). FORNOW: Only handle consecutive access pattern. */
1579 ok = vect_analyze_data_ref_accesses (loop_vinfo, NULL);
1582 if (vect_print_dump_info (REPORT_DETAILS))
1583 fprintf (vect_dump, "bad data access.");
1587 /* Prune the list of ddrs to be tested at run-time by versioning for alias.
1588 It is important to call pruning after vect_analyze_data_ref_accesses,
1589 since we use grouping information gathered by interleaving analysis. */
1590 ok = vect_prune_runtime_alias_test_list (loop_vinfo);
1593 if (vect_print_dump_info (REPORT_DETAILS))
1594 fprintf (vect_dump, "too long list of versioning for alias "
1599 /* This pass will decide on using loop versioning and/or loop peeling in
1600 order to enhance the alignment of data references in the loop. */
1602 ok = vect_enhance_data_refs_alignment (loop_vinfo);
1605 if (vect_print_dump_info (REPORT_DETAILS))
1606 fprintf (vect_dump, "bad data alignment.");
1610 /* Check the SLP opportunities in the loop, analyze and build SLP trees. */
1611 ok = vect_analyze_slp (loop_vinfo, NULL);
1614 /* Decide which possible SLP instances to SLP. */
1615 slp = vect_make_slp_decision (loop_vinfo);
1617 /* Find stmts that need to be both vectorized and SLPed. */
1618 vect_detect_hybrid_slp (loop_vinfo);
1623 /* Scan all the operations in the loop and make sure they are
1626 ok = vect_analyze_loop_operations (loop_vinfo, slp);
1629 if (vect_print_dump_info (REPORT_DETAILS))
1630 fprintf (vect_dump, "bad operation or unsupported loop bound.");
1637 /* Function vect_analyze_loop.
1639 Apply a set of analyses on LOOP, and create a loop_vec_info struct
1640 for it. The different analyses will record information in the
1641 loop_vec_info struct. */
1643 vect_analyze_loop (struct loop *loop)
1645 loop_vec_info loop_vinfo;
1646 unsigned int vector_sizes;
1648 /* Autodetect first vector size we try. */
1649 current_vector_size = 0;
1650 vector_sizes = targetm.vectorize.autovectorize_vector_sizes ();
1652 if (vect_print_dump_info (REPORT_DETAILS))
1653 fprintf (vect_dump, "===== analyze_loop_nest =====");
1655 if (loop_outer (loop)
1656 && loop_vec_info_for_loop (loop_outer (loop))
1657 && LOOP_VINFO_VECTORIZABLE_P (loop_vec_info_for_loop (loop_outer (loop))))
1659 if (vect_print_dump_info (REPORT_DETAILS))
1660 fprintf (vect_dump, "outer-loop already vectorized.");
1666 /* Check the CFG characteristics of the loop (nesting, entry/exit). */
1667 loop_vinfo = vect_analyze_loop_form (loop);
1670 if (vect_print_dump_info (REPORT_DETAILS))
1671 fprintf (vect_dump, "bad loop form.");
1675 if (vect_analyze_loop_2 (loop_vinfo))
1677 LOOP_VINFO_VECTORIZABLE_P (loop_vinfo) = 1;
1682 destroy_loop_vec_info (loop_vinfo, true);
1684 vector_sizes &= ~current_vector_size;
1685 if (vector_sizes == 0
1686 || current_vector_size == 0)
1689 /* Try the next biggest vector size. */
1690 current_vector_size = 1 << floor_log2 (vector_sizes);
1691 if (vect_print_dump_info (REPORT_DETAILS))
1692 fprintf (vect_dump, "***** Re-trying analysis with "
1693 "vector size %d\n", current_vector_size);
1698 /* Function reduction_code_for_scalar_code
1701 CODE - tree_code of a reduction operations.
1704 REDUC_CODE - the corresponding tree-code to be used to reduce the
1705 vector of partial results into a single scalar result (which
1706 will also reside in a vector) or ERROR_MARK if the operation is
1707 a supported reduction operation, but does not have such tree-code.
1709 Return FALSE if CODE currently cannot be vectorized as reduction. */
1712 reduction_code_for_scalar_code (enum tree_code code,
1713 enum tree_code *reduc_code)
1718 *reduc_code = REDUC_MAX_EXPR;
1722 *reduc_code = REDUC_MIN_EXPR;
1726 *reduc_code = REDUC_PLUS_EXPR;
1734 *reduc_code = ERROR_MARK;
1743 /* Error reporting helper for vect_is_simple_reduction below. GIMPLE statement
1744 STMT is printed with a message MSG. */
1747 report_vect_op (gimple stmt, const char *msg)
1749 fprintf (vect_dump, "%s", msg);
1750 print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM);
1754 /* Detect SLP reduction of the form:
1764 PHI is the reduction phi node (#a1 = phi <a5, a0> above)
1765 FIRST_STMT is the first reduction stmt in the chain
1766 (a2 = operation (a1)).
1768 Return TRUE if a reduction chain was detected. */
1771 vect_is_slp_reduction (loop_vec_info loop_info, gimple phi, gimple first_stmt)
1773 struct loop *loop = (gimple_bb (phi))->loop_father;
1774 struct loop *vect_loop = LOOP_VINFO_LOOP (loop_info);
1775 enum tree_code code;
1776 gimple current_stmt = NULL, loop_use_stmt = NULL, first, next_stmt;
1777 stmt_vec_info use_stmt_info, current_stmt_info;
1779 imm_use_iterator imm_iter;
1780 use_operand_p use_p;
1781 int nloop_uses, size = 0, n_out_of_loop_uses;
1784 if (loop != vect_loop)
1787 lhs = PHI_RESULT (phi);
1788 code = gimple_assign_rhs_code (first_stmt);
1792 n_out_of_loop_uses = 0;
1793 FOR_EACH_IMM_USE_FAST (use_p, imm_iter, lhs)
1795 gimple use_stmt = USE_STMT (use_p);
1796 if (is_gimple_debug (use_stmt))
1799 use_stmt = USE_STMT (use_p);
1801 /* Check if we got back to the reduction phi. */
1802 if (use_stmt == phi)
1804 loop_use_stmt = use_stmt;
1809 if (flow_bb_inside_loop_p (loop, gimple_bb (use_stmt)))
1811 if (vinfo_for_stmt (use_stmt)
1812 && !STMT_VINFO_IN_PATTERN_P (vinfo_for_stmt (use_stmt)))
1814 loop_use_stmt = use_stmt;
1819 n_out_of_loop_uses++;
1821 /* There are can be either a single use in the loop or two uses in
1823 if (nloop_uses > 1 || (n_out_of_loop_uses && nloop_uses))
1830 /* We reached a statement with no loop uses. */
1831 if (nloop_uses == 0)
1834 /* This is a loop exit phi, and we haven't reached the reduction phi. */
1835 if (gimple_code (loop_use_stmt) == GIMPLE_PHI)
1838 if (!is_gimple_assign (loop_use_stmt)
1839 || code != gimple_assign_rhs_code (loop_use_stmt)
1840 || !flow_bb_inside_loop_p (loop, gimple_bb (loop_use_stmt)))
1843 /* Insert USE_STMT into reduction chain. */
1844 use_stmt_info = vinfo_for_stmt (loop_use_stmt);
1847 current_stmt_info = vinfo_for_stmt (current_stmt);
1848 GROUP_NEXT_ELEMENT (current_stmt_info) = loop_use_stmt;
1849 GROUP_FIRST_ELEMENT (use_stmt_info)
1850 = GROUP_FIRST_ELEMENT (current_stmt_info);
1853 GROUP_FIRST_ELEMENT (use_stmt_info) = loop_use_stmt;
1855 lhs = gimple_assign_lhs (loop_use_stmt);
1856 current_stmt = loop_use_stmt;
1860 if (!found || loop_use_stmt != phi || size < 2)
1863 /* Swap the operands, if needed, to make the reduction operand be the second
1865 lhs = PHI_RESULT (phi);
1866 next_stmt = GROUP_FIRST_ELEMENT (vinfo_for_stmt (current_stmt));
1869 if (gimple_assign_rhs2 (next_stmt) == lhs)
1871 tree op = gimple_assign_rhs1 (next_stmt);
1872 gimple def_stmt = NULL;
1874 if (TREE_CODE (op) == SSA_NAME)
1875 def_stmt = SSA_NAME_DEF_STMT (op);
1877 /* Check that the other def is either defined in the loop
1878 ("vect_internal_def"), or it's an induction (defined by a
1879 loop-header phi-node). */
1881 && gimple_bb (def_stmt)
1882 && flow_bb_inside_loop_p (loop, gimple_bb (def_stmt))
1883 && (is_gimple_assign (def_stmt)
1884 || is_gimple_call (def_stmt)
1885 || STMT_VINFO_DEF_TYPE (vinfo_for_stmt (def_stmt))
1886 == vect_induction_def
1887 || (gimple_code (def_stmt) == GIMPLE_PHI
1888 && STMT_VINFO_DEF_TYPE (vinfo_for_stmt (def_stmt))
1889 == vect_internal_def
1890 && !is_loop_header_bb_p (gimple_bb (def_stmt)))))
1892 lhs = gimple_assign_lhs (next_stmt);
1893 next_stmt = GROUP_NEXT_ELEMENT (vinfo_for_stmt (next_stmt));
1901 tree op = gimple_assign_rhs2 (next_stmt);
1902 gimple def_stmt = NULL;
1904 if (TREE_CODE (op) == SSA_NAME)
1905 def_stmt = SSA_NAME_DEF_STMT (op);
1907 /* Check that the other def is either defined in the loop
1908 ("vect_internal_def"), or it's an induction (defined by a
1909 loop-header phi-node). */
1911 && gimple_bb (def_stmt)
1912 && flow_bb_inside_loop_p (loop, gimple_bb (def_stmt))
1913 && (is_gimple_assign (def_stmt)
1914 || is_gimple_call (def_stmt)
1915 || STMT_VINFO_DEF_TYPE (vinfo_for_stmt (def_stmt))
1916 == vect_induction_def
1917 || (gimple_code (def_stmt) == GIMPLE_PHI
1918 && STMT_VINFO_DEF_TYPE (vinfo_for_stmt (def_stmt))
1919 == vect_internal_def
1920 && !is_loop_header_bb_p (gimple_bb (def_stmt)))))
1922 if (vect_print_dump_info (REPORT_DETAILS))
1924 fprintf (vect_dump, "swapping oprnds: ");
1925 print_gimple_stmt (vect_dump, next_stmt, 0, TDF_SLIM);
1928 swap_tree_operands (next_stmt,
1929 gimple_assign_rhs1_ptr (next_stmt),
1930 gimple_assign_rhs2_ptr (next_stmt));
1931 mark_symbols_for_renaming (next_stmt);
1937 lhs = gimple_assign_lhs (next_stmt);
1938 next_stmt = GROUP_NEXT_ELEMENT (vinfo_for_stmt (next_stmt));
1941 /* Save the chain for further analysis in SLP detection. */
1942 first = GROUP_FIRST_ELEMENT (vinfo_for_stmt (current_stmt));
1943 VEC_safe_push (gimple, heap, LOOP_VINFO_REDUCTION_CHAINS (loop_info), first);
1944 GROUP_SIZE (vinfo_for_stmt (first)) = size;
1950 /* Function vect_is_simple_reduction_1
1952 (1) Detect a cross-iteration def-use cycle that represents a simple
1953 reduction computation. We look for the following pattern:
1958 a2 = operation (a3, a1)
1961 1. operation is commutative and associative and it is safe to
1962 change the order of the computation (if CHECK_REDUCTION is true)
1963 2. no uses for a2 in the loop (a2 is used out of the loop)
1964 3. no uses of a1 in the loop besides the reduction operation
1965 4. no uses of a1 outside the loop.
1967 Conditions 1,4 are tested here.
1968 Conditions 2,3 are tested in vect_mark_stmts_to_be_vectorized.
1970 (2) Detect a cross-iteration def-use cycle in nested loops, i.e.,
1971 nested cycles, if CHECK_REDUCTION is false.
1973 (3) Detect cycles of phi nodes in outer-loop vectorization, i.e., double
1977 inner loop (def of a3)
1980 If MODIFY is true it tries also to rework the code in-place to enable
1981 detection of more reduction patterns. For the time being we rewrite
1982 "res -= RHS" into "rhs += -RHS" when it seems worthwhile.
1986 vect_is_simple_reduction_1 (loop_vec_info loop_info, gimple phi,
1987 bool check_reduction, bool *double_reduc,
1990 struct loop *loop = (gimple_bb (phi))->loop_father;
1991 struct loop *vect_loop = LOOP_VINFO_LOOP (loop_info);
1992 edge latch_e = loop_latch_edge (loop);
1993 tree loop_arg = PHI_ARG_DEF_FROM_EDGE (phi, latch_e);
1994 gimple def_stmt, def1 = NULL, def2 = NULL;
1995 enum tree_code orig_code, code;
1996 tree op1, op2, op3 = NULL_TREE, op4 = NULL_TREE;
2000 imm_use_iterator imm_iter;
2001 use_operand_p use_p;
2004 *double_reduc = false;
2006 /* If CHECK_REDUCTION is true, we assume inner-most loop vectorization,
2007 otherwise, we assume outer loop vectorization. */
2008 gcc_assert ((check_reduction && loop == vect_loop)
2009 || (!check_reduction && flow_loop_nested_p (vect_loop, loop)));
2011 name = PHI_RESULT (phi);
2013 FOR_EACH_IMM_USE_FAST (use_p, imm_iter, name)
2015 gimple use_stmt = USE_STMT (use_p);
2016 if (is_gimple_debug (use_stmt))
2019 if (!flow_bb_inside_loop_p (loop, gimple_bb (use_stmt)))
2021 if (vect_print_dump_info (REPORT_DETAILS))
2022 fprintf (vect_dump, "intermediate value used outside loop.");
2027 if (vinfo_for_stmt (use_stmt)
2028 && !is_pattern_stmt_p (vinfo_for_stmt (use_stmt)))
2032 if (vect_print_dump_info (REPORT_DETAILS))
2033 fprintf (vect_dump, "reduction used in loop.");
2038 if (TREE_CODE (loop_arg) != SSA_NAME)
2040 if (vect_print_dump_info (REPORT_DETAILS))
2042 fprintf (vect_dump, "reduction: not ssa_name: ");
2043 print_generic_expr (vect_dump, loop_arg, TDF_SLIM);
2048 def_stmt = SSA_NAME_DEF_STMT (loop_arg);
2051 if (vect_print_dump_info (REPORT_DETAILS))
2052 fprintf (vect_dump, "reduction: no def_stmt.");
2056 if (!is_gimple_assign (def_stmt) && gimple_code (def_stmt) != GIMPLE_PHI)
2058 if (vect_print_dump_info (REPORT_DETAILS))
2059 print_gimple_stmt (vect_dump, def_stmt, 0, TDF_SLIM);
2063 if (is_gimple_assign (def_stmt))
2065 name = gimple_assign_lhs (def_stmt);
2070 name = PHI_RESULT (def_stmt);
2075 FOR_EACH_IMM_USE_FAST (use_p, imm_iter, name)
2077 gimple use_stmt = USE_STMT (use_p);
2078 if (is_gimple_debug (use_stmt))
2080 if (flow_bb_inside_loop_p (loop, gimple_bb (use_stmt))
2081 && vinfo_for_stmt (use_stmt)
2082 && !is_pattern_stmt_p (vinfo_for_stmt (use_stmt)))
2086 if (vect_print_dump_info (REPORT_DETAILS))
2087 fprintf (vect_dump, "reduction used in loop.");
2092 /* If DEF_STMT is a phi node itself, we expect it to have a single argument
2093 defined in the inner loop. */
2096 op1 = PHI_ARG_DEF (def_stmt, 0);
2098 if (gimple_phi_num_args (def_stmt) != 1
2099 || TREE_CODE (op1) != SSA_NAME)
2101 if (vect_print_dump_info (REPORT_DETAILS))
2102 fprintf (vect_dump, "unsupported phi node definition.");
2107 def1 = SSA_NAME_DEF_STMT (op1);
2108 if (flow_bb_inside_loop_p (loop, gimple_bb (def_stmt))
2110 && flow_bb_inside_loop_p (loop->inner, gimple_bb (def1))
2111 && is_gimple_assign (def1))
2113 if (vect_print_dump_info (REPORT_DETAILS))
2114 report_vect_op (def_stmt, "detected double reduction: ");
2116 *double_reduc = true;
2123 code = orig_code = gimple_assign_rhs_code (def_stmt);
2125 /* We can handle "res -= x[i]", which is non-associative by
2126 simply rewriting this into "res += -x[i]". Avoid changing
2127 gimple instruction for the first simple tests and only do this
2128 if we're allowed to change code at all. */
2129 if (code == MINUS_EXPR
2131 && (op1 = gimple_assign_rhs1 (def_stmt))
2132 && TREE_CODE (op1) == SSA_NAME
2133 && SSA_NAME_DEF_STMT (op1) == phi)
2137 && (!commutative_tree_code (code) || !associative_tree_code (code)))
2139 if (vect_print_dump_info (REPORT_DETAILS))
2140 report_vect_op (def_stmt, "reduction: not commutative/associative: ");
2144 if (get_gimple_rhs_class (code) != GIMPLE_BINARY_RHS)
2146 if (code != COND_EXPR)
2148 if (vect_print_dump_info (REPORT_DETAILS))
2149 report_vect_op (def_stmt, "reduction: not binary operation: ");
2154 op3 = gimple_assign_rhs1 (def_stmt);
2155 if (COMPARISON_CLASS_P (op3))
2157 op4 = TREE_OPERAND (op3, 1);
2158 op3 = TREE_OPERAND (op3, 0);
2161 op1 = gimple_assign_rhs2 (def_stmt);
2162 op2 = gimple_assign_rhs3 (def_stmt);
2164 if (TREE_CODE (op1) != SSA_NAME && TREE_CODE (op2) != SSA_NAME)
2166 if (vect_print_dump_info (REPORT_DETAILS))
2167 report_vect_op (def_stmt, "reduction: uses not ssa_names: ");
2174 op1 = gimple_assign_rhs1 (def_stmt);
2175 op2 = gimple_assign_rhs2 (def_stmt);
2177 if (TREE_CODE (op1) != SSA_NAME && TREE_CODE (op2) != SSA_NAME)
2179 if (vect_print_dump_info (REPORT_DETAILS))
2180 report_vect_op (def_stmt, "reduction: uses not ssa_names: ");
2186 type = TREE_TYPE (gimple_assign_lhs (def_stmt));
2187 if ((TREE_CODE (op1) == SSA_NAME
2188 && !types_compatible_p (type,TREE_TYPE (op1)))
2189 || (TREE_CODE (op2) == SSA_NAME
2190 && !types_compatible_p (type, TREE_TYPE (op2)))
2191 || (op3 && TREE_CODE (op3) == SSA_NAME
2192 && !types_compatible_p (type, TREE_TYPE (op3)))
2193 || (op4 && TREE_CODE (op4) == SSA_NAME
2194 && !types_compatible_p (type, TREE_TYPE (op4))))
2196 if (vect_print_dump_info (REPORT_DETAILS))
2198 fprintf (vect_dump, "reduction: multiple types: operation type: ");
2199 print_generic_expr (vect_dump, type, TDF_SLIM);
2200 fprintf (vect_dump, ", operands types: ");
2201 print_generic_expr (vect_dump, TREE_TYPE (op1), TDF_SLIM);
2202 fprintf (vect_dump, ",");
2203 print_generic_expr (vect_dump, TREE_TYPE (op2), TDF_SLIM);
2206 fprintf (vect_dump, ",");
2207 print_generic_expr (vect_dump, TREE_TYPE (op3), TDF_SLIM);
2212 fprintf (vect_dump, ",");
2213 print_generic_expr (vect_dump, TREE_TYPE (op4), TDF_SLIM);
2220 /* Check that it's ok to change the order of the computation.
2221 Generally, when vectorizing a reduction we change the order of the
2222 computation. This may change the behavior of the program in some
2223 cases, so we need to check that this is ok. One exception is when
2224 vectorizing an outer-loop: the inner-loop is executed sequentially,
2225 and therefore vectorizing reductions in the inner-loop during
2226 outer-loop vectorization is safe. */
2228 /* CHECKME: check for !flag_finite_math_only too? */
2229 if (SCALAR_FLOAT_TYPE_P (type) && !flag_associative_math
2232 /* Changing the order of operations changes the semantics. */
2233 if (vect_print_dump_info (REPORT_DETAILS))
2234 report_vect_op (def_stmt, "reduction: unsafe fp math optimization: ");
2237 else if (INTEGRAL_TYPE_P (type) && TYPE_OVERFLOW_TRAPS (type)
2240 /* Changing the order of operations changes the semantics. */
2241 if (vect_print_dump_info (REPORT_DETAILS))
2242 report_vect_op (def_stmt, "reduction: unsafe int math optimization: ");
2245 else if (SAT_FIXED_POINT_TYPE_P (type) && check_reduction)
2247 /* Changing the order of operations changes the semantics. */
2248 if (vect_print_dump_info (REPORT_DETAILS))
2249 report_vect_op (def_stmt,
2250 "reduction: unsafe fixed-point math optimization: ");
2254 /* If we detected "res -= x[i]" earlier, rewrite it into
2255 "res += -x[i]" now. If this turns out to be useless reassoc
2256 will clean it up again. */
2257 if (orig_code == MINUS_EXPR)
2259 tree rhs = gimple_assign_rhs2 (def_stmt);
2260 tree var = TREE_CODE (rhs) == SSA_NAME
2261 ? SSA_NAME_VAR (rhs)
2262 : create_tmp_reg (TREE_TYPE (rhs), NULL);
2263 tree negrhs = make_ssa_name (var, NULL);
2264 gimple negate_stmt = gimple_build_assign_with_ops (NEGATE_EXPR, negrhs,
2266 gimple_stmt_iterator gsi = gsi_for_stmt (def_stmt);
2267 set_vinfo_for_stmt (negate_stmt, new_stmt_vec_info (negate_stmt,
2269 gsi_insert_before (&gsi, negate_stmt, GSI_NEW_STMT);
2270 gimple_assign_set_rhs2 (def_stmt, negrhs);
2271 gimple_assign_set_rhs_code (def_stmt, PLUS_EXPR);
2272 update_stmt (def_stmt);
2275 /* Reduction is safe. We're dealing with one of the following:
2276 1) integer arithmetic and no trapv
2277 2) floating point arithmetic, and special flags permit this optimization
2278 3) nested cycle (i.e., outer loop vectorization). */
2279 if (TREE_CODE (op1) == SSA_NAME)
2280 def1 = SSA_NAME_DEF_STMT (op1);
2282 if (TREE_CODE (op2) == SSA_NAME)
2283 def2 = SSA_NAME_DEF_STMT (op2);
2285 if (code != COND_EXPR
2286 && ((!def1 || gimple_nop_p (def1)) && (!def2 || gimple_nop_p (def2))))
2288 if (vect_print_dump_info (REPORT_DETAILS))
2289 report_vect_op (def_stmt, "reduction: no defs for operands: ");
2293 /* Check that one def is the reduction def, defined by PHI,
2294 the other def is either defined in the loop ("vect_internal_def"),
2295 or it's an induction (defined by a loop-header phi-node). */
2297 if (def2 && def2 == phi
2298 && (code == COND_EXPR
2299 || !def1 || gimple_nop_p (def1)
2300 || (def1 && flow_bb_inside_loop_p (loop, gimple_bb (def1))
2301 && (is_gimple_assign (def1)
2302 || is_gimple_call (def1)
2303 || STMT_VINFO_DEF_TYPE (vinfo_for_stmt (def1))
2304 == vect_induction_def
2305 || (gimple_code (def1) == GIMPLE_PHI
2306 && STMT_VINFO_DEF_TYPE (vinfo_for_stmt (def1))
2307 == vect_internal_def
2308 && !is_loop_header_bb_p (gimple_bb (def1)))))))
2310 if (vect_print_dump_info (REPORT_DETAILS))
2311 report_vect_op (def_stmt, "detected reduction: ");
2315 if (def1 && def1 == phi
2316 && (code == COND_EXPR
2317 || !def2 || gimple_nop_p (def2)
2318 || (def2 && flow_bb_inside_loop_p (loop, gimple_bb (def2))
2319 && (is_gimple_assign (def2)
2320 || is_gimple_call (def2)
2321 || STMT_VINFO_DEF_TYPE (vinfo_for_stmt (def2))
2322 == vect_induction_def
2323 || (gimple_code (def2) == GIMPLE_PHI
2324 && STMT_VINFO_DEF_TYPE (vinfo_for_stmt (def2))
2325 == vect_internal_def
2326 && !is_loop_header_bb_p (gimple_bb (def2)))))))
2328 if (check_reduction)
2330 /* Swap operands (just for simplicity - so that the rest of the code
2331 can assume that the reduction variable is always the last (second)
2333 if (vect_print_dump_info (REPORT_DETAILS))
2334 report_vect_op (def_stmt,
2335 "detected reduction: need to swap operands: ");
2337 swap_tree_operands (def_stmt, gimple_assign_rhs1_ptr (def_stmt),
2338 gimple_assign_rhs2_ptr (def_stmt));
2342 if (vect_print_dump_info (REPORT_DETAILS))
2343 report_vect_op (def_stmt, "detected reduction: ");
2349 /* Try to find SLP reduction chain. */
2350 if (check_reduction && vect_is_slp_reduction (loop_info, phi, def_stmt))
2352 if (vect_print_dump_info (REPORT_DETAILS))
2353 report_vect_op (def_stmt, "reduction: detected reduction chain: ");
2358 if (vect_print_dump_info (REPORT_DETAILS))
2359 report_vect_op (def_stmt, "reduction: unknown pattern: ");
2364 /* Wrapper around vect_is_simple_reduction_1, that won't modify code
2365 in-place. Arguments as there. */
2368 vect_is_simple_reduction (loop_vec_info loop_info, gimple phi,
2369 bool check_reduction, bool *double_reduc)
2371 return vect_is_simple_reduction_1 (loop_info, phi, check_reduction,
2372 double_reduc, false);
2375 /* Wrapper around vect_is_simple_reduction_1, which will modify code
2376 in-place if it enables detection of more reductions. Arguments
2380 vect_force_simple_reduction (loop_vec_info loop_info, gimple phi,
2381 bool check_reduction, bool *double_reduc)
2383 return vect_is_simple_reduction_1 (loop_info, phi, check_reduction,
2384 double_reduc, true);
2387 /* Calculate the cost of one scalar iteration of the loop. */
2389 vect_get_single_scalar_iteration_cost (loop_vec_info loop_vinfo)
2391 struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
2392 basic_block *bbs = LOOP_VINFO_BBS (loop_vinfo);
2393 int nbbs = loop->num_nodes, factor, scalar_single_iter_cost = 0;
2394 int innerloop_iters, i, stmt_cost;
2396 /* Count statements in scalar loop. Using this as scalar cost for a single
2399 TODO: Add outer loop support.
2401 TODO: Consider assigning different costs to different scalar
2405 innerloop_iters = 1;
2407 innerloop_iters = 50; /* FIXME */
2409 for (i = 0; i < nbbs; i++)
2411 gimple_stmt_iterator si;
2412 basic_block bb = bbs[i];
2414 if (bb->loop_father == loop->inner)
2415 factor = innerloop_iters;
2419 for (si = gsi_start_bb (bb); !gsi_end_p (si); gsi_next (&si))
2421 gimple stmt = gsi_stmt (si);
2422 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
2424 if (!is_gimple_assign (stmt) && !is_gimple_call (stmt))
2427 /* Skip stmts that are not vectorized inside the loop. */
2429 && !STMT_VINFO_RELEVANT_P (stmt_info)
2430 && (!STMT_VINFO_LIVE_P (stmt_info)
2431 || !VECTORIZABLE_CYCLE_DEF (STMT_VINFO_DEF_TYPE (stmt_info)))
2432 && !STMT_VINFO_IN_PATTERN_P (stmt_info))
2435 if (STMT_VINFO_DATA_REF (vinfo_for_stmt (stmt)))
2437 if (DR_IS_READ (STMT_VINFO_DATA_REF (vinfo_for_stmt (stmt))))
2438 stmt_cost = vect_get_cost (scalar_load);
2440 stmt_cost = vect_get_cost (scalar_store);
2443 stmt_cost = vect_get_cost (scalar_stmt);
2445 scalar_single_iter_cost += stmt_cost * factor;
2448 return scalar_single_iter_cost;
2451 /* Calculate cost of peeling the loop PEEL_ITERS_PROLOGUE times. */
2453 vect_get_known_peeling_cost (loop_vec_info loop_vinfo, int peel_iters_prologue,
2454 int *peel_iters_epilogue,
2455 int scalar_single_iter_cost)
2457 int peel_guard_costs = 0;
2458 int vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo);
2460 if (!LOOP_VINFO_NITERS_KNOWN_P (loop_vinfo))
2462 *peel_iters_epilogue = vf/2;
2463 if (vect_print_dump_info (REPORT_COST))
2464 fprintf (vect_dump, "cost model: "
2465 "epilogue peel iters set to vf/2 because "
2466 "loop iterations are unknown .");
2468 /* If peeled iterations are known but number of scalar loop
2469 iterations are unknown, count a taken branch per peeled loop. */
2470 peel_guard_costs = 2 * vect_get_cost (cond_branch_taken);
2474 int niters = LOOP_VINFO_INT_NITERS (loop_vinfo);
2475 peel_iters_prologue = niters < peel_iters_prologue ?
2476 niters : peel_iters_prologue;
2477 *peel_iters_epilogue = (niters - peel_iters_prologue) % vf;
2478 /* If we need to peel for gaps, but no peeling is required, we have to
2479 peel VF iterations. */
2480 if (LOOP_VINFO_PEELING_FOR_GAPS (loop_vinfo) && !*peel_iters_epilogue)
2481 *peel_iters_epilogue = vf;
2484 return (peel_iters_prologue * scalar_single_iter_cost)
2485 + (*peel_iters_epilogue * scalar_single_iter_cost)
2489 /* Function vect_estimate_min_profitable_iters
2491 Return the number of iterations required for the vector version of the
2492 loop to be profitable relative to the cost of the scalar version of the
2495 TODO: Take profile info into account before making vectorization
2496 decisions, if available. */
2499 vect_estimate_min_profitable_iters (loop_vec_info loop_vinfo)
2502 int min_profitable_iters;
2503 int peel_iters_prologue;
2504 int peel_iters_epilogue;
2505 int vec_inside_cost = 0;
2506 int vec_outside_cost = 0;
2507 int scalar_single_iter_cost = 0;
2508 int scalar_outside_cost = 0;
2509 int vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo);
2510 struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
2511 basic_block *bbs = LOOP_VINFO_BBS (loop_vinfo);
2512 int nbbs = loop->num_nodes;
2513 int npeel = LOOP_PEELING_FOR_ALIGNMENT (loop_vinfo);
2514 int peel_guard_costs = 0;
2515 int innerloop_iters = 0, factor;
2516 VEC (slp_instance, heap) *slp_instances;
2517 slp_instance instance;
2519 /* Cost model disabled. */
2520 if (!flag_vect_cost_model)
2522 if (vect_print_dump_info (REPORT_COST))
2523 fprintf (vect_dump, "cost model disabled.");
2527 /* Requires loop versioning tests to handle misalignment. */
2528 if (LOOP_REQUIRES_VERSIONING_FOR_ALIGNMENT (loop_vinfo))
2530 /* FIXME: Make cost depend on complexity of individual check. */
2532 VEC_length (gimple, LOOP_VINFO_MAY_MISALIGN_STMTS (loop_vinfo));
2533 if (vect_print_dump_info (REPORT_COST))
2534 fprintf (vect_dump, "cost model: Adding cost of checks for loop "
2535 "versioning to treat misalignment.\n");
2538 /* Requires loop versioning with alias checks. */
2539 if (LOOP_REQUIRES_VERSIONING_FOR_ALIAS (loop_vinfo))
2541 /* FIXME: Make cost depend on complexity of individual check. */
2543 VEC_length (ddr_p, LOOP_VINFO_MAY_ALIAS_DDRS (loop_vinfo));
2544 if (vect_print_dump_info (REPORT_COST))
2545 fprintf (vect_dump, "cost model: Adding cost of checks for loop "
2546 "versioning aliasing.\n");
2549 if (LOOP_REQUIRES_VERSIONING_FOR_ALIGNMENT (loop_vinfo)
2550 || LOOP_REQUIRES_VERSIONING_FOR_ALIAS (loop_vinfo))
2551 vec_outside_cost += vect_get_cost (cond_branch_taken);
2553 /* Count statements in scalar loop. Using this as scalar cost for a single
2556 TODO: Add outer loop support.
2558 TODO: Consider assigning different costs to different scalar
2563 innerloop_iters = 50; /* FIXME */
2565 for (i = 0; i < nbbs; i++)
2567 gimple_stmt_iterator si;
2568 basic_block bb = bbs[i];
2570 if (bb->loop_father == loop->inner)
2571 factor = innerloop_iters;
2575 for (si = gsi_start_bb (bb); !gsi_end_p (si); gsi_next (&si))
2577 gimple stmt = gsi_stmt (si);
2578 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
2580 if (STMT_VINFO_IN_PATTERN_P (stmt_info))
2582 stmt = STMT_VINFO_RELATED_STMT (stmt_info);
2583 stmt_info = vinfo_for_stmt (stmt);
2586 /* Skip stmts that are not vectorized inside the loop. */
2587 if (!STMT_VINFO_RELEVANT_P (stmt_info)
2588 && (!STMT_VINFO_LIVE_P (stmt_info)
2589 || !VECTORIZABLE_CYCLE_DEF (STMT_VINFO_DEF_TYPE (stmt_info))))
2592 vec_inside_cost += STMT_VINFO_INSIDE_OF_LOOP_COST (stmt_info) * factor;
2593 /* FIXME: for stmts in the inner-loop in outer-loop vectorization,
2594 some of the "outside" costs are generated inside the outer-loop. */
2595 vec_outside_cost += STMT_VINFO_OUTSIDE_OF_LOOP_COST (stmt_info);
2596 if (is_pattern_stmt_p (stmt_info)
2597 && STMT_VINFO_PATTERN_DEF_SEQ (stmt_info))
2599 gimple_stmt_iterator gsi;
2601 for (gsi = gsi_start (STMT_VINFO_PATTERN_DEF_SEQ (stmt_info));
2602 !gsi_end_p (gsi); gsi_next (&gsi))
2604 gimple pattern_def_stmt = gsi_stmt (gsi);
2605 stmt_vec_info pattern_def_stmt_info
2606 = vinfo_for_stmt (pattern_def_stmt);
2607 if (STMT_VINFO_RELEVANT_P (pattern_def_stmt_info)
2608 || STMT_VINFO_LIVE_P (pattern_def_stmt_info))
2611 += STMT_VINFO_INSIDE_OF_LOOP_COST
2612 (pattern_def_stmt_info) * factor;
2614 += STMT_VINFO_OUTSIDE_OF_LOOP_COST
2615 (pattern_def_stmt_info);
2622 scalar_single_iter_cost = vect_get_single_scalar_iteration_cost (loop_vinfo);
2624 /* Add additional cost for the peeled instructions in prologue and epilogue
2627 FORNOW: If we don't know the value of peel_iters for prologue or epilogue
2628 at compile-time - we assume it's vf/2 (the worst would be vf-1).
2630 TODO: Build an expression that represents peel_iters for prologue and
2631 epilogue to be used in a run-time test. */
2635 peel_iters_prologue = vf/2;
2636 if (vect_print_dump_info (REPORT_COST))
2637 fprintf (vect_dump, "cost model: "
2638 "prologue peel iters set to vf/2.");
2640 /* If peeling for alignment is unknown, loop bound of main loop becomes
2642 peel_iters_epilogue = vf/2;
2643 if (vect_print_dump_info (REPORT_COST))
2644 fprintf (vect_dump, "cost model: "
2645 "epilogue peel iters set to vf/2 because "
2646 "peeling for alignment is unknown .");
2648 /* If peeled iterations are unknown, count a taken branch and a not taken
2649 branch per peeled loop. Even if scalar loop iterations are known,
2650 vector iterations are not known since peeled prologue iterations are
2651 not known. Hence guards remain the same. */
2652 peel_guard_costs += 2 * (vect_get_cost (cond_branch_taken)
2653 + vect_get_cost (cond_branch_not_taken));
2654 vec_outside_cost += (peel_iters_prologue * scalar_single_iter_cost)
2655 + (peel_iters_epilogue * scalar_single_iter_cost)
2660 peel_iters_prologue = npeel;
2661 vec_outside_cost += vect_get_known_peeling_cost (loop_vinfo,
2662 peel_iters_prologue, &peel_iters_epilogue,
2663 scalar_single_iter_cost);
2666 /* FORNOW: The scalar outside cost is incremented in one of the
2669 1. The vectorizer checks for alignment and aliasing and generates
2670 a condition that allows dynamic vectorization. A cost model
2671 check is ANDED with the versioning condition. Hence scalar code
2672 path now has the added cost of the versioning check.
2674 if (cost > th & versioning_check)
2677 Hence run-time scalar is incremented by not-taken branch cost.
2679 2. The vectorizer then checks if a prologue is required. If the
2680 cost model check was not done before during versioning, it has to
2681 be done before the prologue check.
2684 prologue = scalar_iters
2689 if (prologue == num_iters)
2692 Hence the run-time scalar cost is incremented by a taken branch,
2693 plus a not-taken branch, plus a taken branch cost.
2695 3. The vectorizer then checks if an epilogue is required. If the
2696 cost model check was not done before during prologue check, it
2697 has to be done with the epilogue check.
2703 if (prologue == num_iters)
2706 if ((cost <= th) | (scalar_iters-prologue-epilogue == 0))
2709 Hence the run-time scalar cost should be incremented by 2 taken
2712 TODO: The back end may reorder the BBS's differently and reverse
2713 conditions/branch directions. Change the estimates below to
2714 something more reasonable. */
2716 /* If the number of iterations is known and we do not do versioning, we can
2717 decide whether to vectorize at compile time. Hence the scalar version
2718 do not carry cost model guard costs. */
2719 if (!LOOP_VINFO_NITERS_KNOWN_P (loop_vinfo)
2720 || LOOP_REQUIRES_VERSIONING_FOR_ALIGNMENT (loop_vinfo)
2721 || LOOP_REQUIRES_VERSIONING_FOR_ALIAS (loop_vinfo))
2723 /* Cost model check occurs at versioning. */
2724 if (LOOP_REQUIRES_VERSIONING_FOR_ALIGNMENT (loop_vinfo)
2725 || LOOP_REQUIRES_VERSIONING_FOR_ALIAS (loop_vinfo))
2726 scalar_outside_cost += vect_get_cost (cond_branch_not_taken);
2729 /* Cost model check occurs at prologue generation. */
2730 if (LOOP_PEELING_FOR_ALIGNMENT (loop_vinfo) < 0)
2731 scalar_outside_cost += 2 * vect_get_cost (cond_branch_taken)
2732 + vect_get_cost (cond_branch_not_taken);
2733 /* Cost model check occurs at epilogue generation. */
2735 scalar_outside_cost += 2 * vect_get_cost (cond_branch_taken);
2739 /* Add SLP costs. */
2740 slp_instances = LOOP_VINFO_SLP_INSTANCES (loop_vinfo);
2741 FOR_EACH_VEC_ELT (slp_instance, slp_instances, i, instance)
2743 vec_outside_cost += SLP_INSTANCE_OUTSIDE_OF_LOOP_COST (instance);
2744 vec_inside_cost += SLP_INSTANCE_INSIDE_OF_LOOP_COST (instance);
2747 /* Calculate number of iterations required to make the vector version
2748 profitable, relative to the loop bodies only. The following condition
2750 SIC * niters + SOC > VIC * ((niters-PL_ITERS-EP_ITERS)/VF) + VOC
2752 SIC = scalar iteration cost, VIC = vector iteration cost,
2753 VOC = vector outside cost, VF = vectorization factor,
2754 PL_ITERS = prologue iterations, EP_ITERS= epilogue iterations
2755 SOC = scalar outside cost for run time cost model check. */
2757 if ((scalar_single_iter_cost * vf) > vec_inside_cost)
2759 if (vec_outside_cost <= 0)
2760 min_profitable_iters = 1;
2763 min_profitable_iters = ((vec_outside_cost - scalar_outside_cost) * vf
2764 - vec_inside_cost * peel_iters_prologue
2765 - vec_inside_cost * peel_iters_epilogue)
2766 / ((scalar_single_iter_cost * vf)
2769 if ((scalar_single_iter_cost * vf * min_profitable_iters)
2770 <= ((vec_inside_cost * min_profitable_iters)
2771 + ((vec_outside_cost - scalar_outside_cost) * vf)))
2772 min_profitable_iters++;
2775 /* vector version will never be profitable. */
2778 if (vect_print_dump_info (REPORT_COST))
2779 fprintf (vect_dump, "cost model: the vector iteration cost = %d "
2780 "divided by the scalar iteration cost = %d "
2781 "is greater or equal to the vectorization factor = %d.",
2782 vec_inside_cost, scalar_single_iter_cost, vf);
2786 if (vect_print_dump_info (REPORT_COST))
2788 fprintf (vect_dump, "Cost model analysis: \n");
2789 fprintf (vect_dump, " Vector inside of loop cost: %d\n",
2791 fprintf (vect_dump, " Vector outside of loop cost: %d\n",
2793 fprintf (vect_dump, " Scalar iteration cost: %d\n",
2794 scalar_single_iter_cost);
2795 fprintf (vect_dump, " Scalar outside cost: %d\n", scalar_outside_cost);
2796 fprintf (vect_dump, " prologue iterations: %d\n",
2797 peel_iters_prologue);
2798 fprintf (vect_dump, " epilogue iterations: %d\n",
2799 peel_iters_epilogue);
2800 fprintf (vect_dump, " Calculated minimum iters for profitability: %d\n",
2801 min_profitable_iters);
2804 min_profitable_iters =
2805 min_profitable_iters < vf ? vf : min_profitable_iters;
2807 /* Because the condition we create is:
2808 if (niters <= min_profitable_iters)
2809 then skip the vectorized loop. */
2810 min_profitable_iters--;
2812 if (vect_print_dump_info (REPORT_COST))
2813 fprintf (vect_dump, " Profitability threshold = %d\n",
2814 min_profitable_iters);
2816 return min_profitable_iters;
2820 /* TODO: Close dependency between vect_model_*_cost and vectorizable_*
2821 functions. Design better to avoid maintenance issues. */
2823 /* Function vect_model_reduction_cost.
2825 Models cost for a reduction operation, including the vector ops
2826 generated within the strip-mine loop, the initial definition before
2827 the loop, and the epilogue code that must be generated. */
2830 vect_model_reduction_cost (stmt_vec_info stmt_info, enum tree_code reduc_code,
2834 enum tree_code code;
2837 gimple stmt, orig_stmt;
2839 enum machine_mode mode;
2840 loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info);
2841 struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
2844 /* Cost of reduction op inside loop. */
2845 STMT_VINFO_INSIDE_OF_LOOP_COST (stmt_info)
2846 += ncopies * vect_get_cost (vector_stmt);
2848 stmt = STMT_VINFO_STMT (stmt_info);
2850 switch (get_gimple_rhs_class (gimple_assign_rhs_code (stmt)))
2852 case GIMPLE_SINGLE_RHS:
2853 gcc_assert (TREE_OPERAND_LENGTH (gimple_assign_rhs1 (stmt)) == ternary_op);
2854 reduction_op = TREE_OPERAND (gimple_assign_rhs1 (stmt), 2);
2856 case GIMPLE_UNARY_RHS:
2857 reduction_op = gimple_assign_rhs1 (stmt);
2859 case GIMPLE_BINARY_RHS:
2860 reduction_op = gimple_assign_rhs2 (stmt);
2862 case GIMPLE_TERNARY_RHS:
2863 reduction_op = gimple_assign_rhs3 (stmt);
2869 vectype = get_vectype_for_scalar_type (TREE_TYPE (reduction_op));
2872 if (vect_print_dump_info (REPORT_COST))
2874 fprintf (vect_dump, "unsupported data-type ");
2875 print_generic_expr (vect_dump, TREE_TYPE (reduction_op), TDF_SLIM);
2880 mode = TYPE_MODE (vectype);
2881 orig_stmt = STMT_VINFO_RELATED_STMT (stmt_info);
2884 orig_stmt = STMT_VINFO_STMT (stmt_info);
2886 code = gimple_assign_rhs_code (orig_stmt);
2888 /* Add in cost for initial definition. */
2889 outer_cost += vect_get_cost (scalar_to_vec);
2891 /* Determine cost of epilogue code.
2893 We have a reduction operator that will reduce the vector in one statement.
2894 Also requires scalar extract. */
2896 if (!nested_in_vect_loop_p (loop, orig_stmt))
2898 if (reduc_code != ERROR_MARK)
2899 outer_cost += vect_get_cost (vector_stmt)
2900 + vect_get_cost (vec_to_scalar);
2903 int vec_size_in_bits = tree_low_cst (TYPE_SIZE (vectype), 1);
2905 TYPE_SIZE (TREE_TYPE (gimple_assign_lhs (orig_stmt)));
2906 int element_bitsize = tree_low_cst (bitsize, 1);
2907 int nelements = vec_size_in_bits / element_bitsize;
2909 optab = optab_for_tree_code (code, vectype, optab_default);
2911 /* We have a whole vector shift available. */
2912 if (VECTOR_MODE_P (mode)
2913 && optab_handler (optab, mode) != CODE_FOR_nothing
2914 && optab_handler (vec_shr_optab, mode) != CODE_FOR_nothing)
2915 /* Final reduction via vector shifts and the reduction operator. Also
2916 requires scalar extract. */
2917 outer_cost += ((exact_log2(nelements) * 2)
2918 * vect_get_cost (vector_stmt)
2919 + vect_get_cost (vec_to_scalar));
2921 /* Use extracts and reduction op for final reduction. For N elements,
2922 we have N extracts and N-1 reduction ops. */
2923 outer_cost += ((nelements + nelements - 1)
2924 * vect_get_cost (vector_stmt));
2928 STMT_VINFO_OUTSIDE_OF_LOOP_COST (stmt_info) = outer_cost;
2930 if (vect_print_dump_info (REPORT_COST))
2931 fprintf (vect_dump, "vect_model_reduction_cost: inside_cost = %d, "
2932 "outside_cost = %d .", STMT_VINFO_INSIDE_OF_LOOP_COST (stmt_info),
2933 STMT_VINFO_OUTSIDE_OF_LOOP_COST (stmt_info));
2939 /* Function vect_model_induction_cost.
2941 Models cost for induction operations. */
2944 vect_model_induction_cost (stmt_vec_info stmt_info, int ncopies)
2946 /* loop cost for vec_loop. */
2947 STMT_VINFO_INSIDE_OF_LOOP_COST (stmt_info)
2948 = ncopies * vect_get_cost (vector_stmt);
2949 /* prologue cost for vec_init and vec_step. */
2950 STMT_VINFO_OUTSIDE_OF_LOOP_COST (stmt_info)
2951 = 2 * vect_get_cost (scalar_to_vec);
2953 if (vect_print_dump_info (REPORT_COST))
2954 fprintf (vect_dump, "vect_model_induction_cost: inside_cost = %d, "
2955 "outside_cost = %d .", STMT_VINFO_INSIDE_OF_LOOP_COST (stmt_info),
2956 STMT_VINFO_OUTSIDE_OF_LOOP_COST (stmt_info));
2960 /* Function get_initial_def_for_induction
2963 STMT - a stmt that performs an induction operation in the loop.
2964 IV_PHI - the initial value of the induction variable
2967 Return a vector variable, initialized with the first VF values of
2968 the induction variable. E.g., for an iv with IV_PHI='X' and
2969 evolution S, for a vector of 4 units, we want to return:
2970 [X, X + S, X + 2*S, X + 3*S]. */
2973 get_initial_def_for_induction (gimple iv_phi)
2975 stmt_vec_info stmt_vinfo = vinfo_for_stmt (iv_phi);
2976 loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_vinfo);
2977 struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
2981 edge pe = loop_preheader_edge (loop);
2982 struct loop *iv_loop;
2984 tree vec, vec_init, vec_step, t;
2988 gimple init_stmt, induction_phi, new_stmt;
2989 tree induc_def, vec_def, vec_dest;
2990 tree init_expr, step_expr;
2991 int vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo);
2996 stmt_vec_info phi_info = vinfo_for_stmt (iv_phi);
2997 bool nested_in_vect_loop = false;
2998 gimple_seq stmts = NULL;
2999 imm_use_iterator imm_iter;
3000 use_operand_p use_p;
3004 gimple_stmt_iterator si;
3005 basic_block bb = gimple_bb (iv_phi);
3009 /* Is phi in an inner-loop, while vectorizing an enclosing outer-loop? */
3010 if (nested_in_vect_loop_p (loop, iv_phi))
3012 nested_in_vect_loop = true;
3013 iv_loop = loop->inner;
3017 gcc_assert (iv_loop == (gimple_bb (iv_phi))->loop_father);
3019 latch_e = loop_latch_edge (iv_loop);
3020 loop_arg = PHI_ARG_DEF_FROM_EDGE (iv_phi, latch_e);
3022 access_fn = analyze_scalar_evolution (iv_loop, PHI_RESULT (iv_phi));
3023 gcc_assert (access_fn);
3024 STRIP_NOPS (access_fn);
3025 ok = vect_is_simple_iv_evolution (iv_loop->num, access_fn,
3026 &init_expr, &step_expr);
3028 pe = loop_preheader_edge (iv_loop);
3030 scalar_type = TREE_TYPE (init_expr);
3031 vectype = get_vectype_for_scalar_type (scalar_type);
3032 resvectype = get_vectype_for_scalar_type (TREE_TYPE (PHI_RESULT (iv_phi)));
3033 gcc_assert (vectype);
3034 nunits = TYPE_VECTOR_SUBPARTS (vectype);
3035 ncopies = vf / nunits;
3037 gcc_assert (phi_info);
3038 gcc_assert (ncopies >= 1);
3040 /* Find the first insertion point in the BB. */
3041 si = gsi_after_labels (bb);
3043 /* Create the vector that holds the initial_value of the induction. */
3044 if (nested_in_vect_loop)
3046 /* iv_loop is nested in the loop to be vectorized. init_expr had already
3047 been created during vectorization of previous stmts. We obtain it
3048 from the STMT_VINFO_VEC_STMT of the defining stmt. */
3049 tree iv_def = PHI_ARG_DEF_FROM_EDGE (iv_phi,
3050 loop_preheader_edge (iv_loop));
3051 vec_init = vect_get_vec_def_for_operand (iv_def, iv_phi, NULL);
3055 /* iv_loop is the loop to be vectorized. Create:
3056 vec_init = [X, X+S, X+2*S, X+3*S] (S = step_expr, X = init_expr) */
3057 new_var = vect_get_new_vect_var (scalar_type, vect_scalar_var, "var_");
3058 add_referenced_var (new_var);
3060 new_name = force_gimple_operand (init_expr, &stmts, false, new_var);
3063 new_bb = gsi_insert_seq_on_edge_immediate (pe, stmts);
3064 gcc_assert (!new_bb);
3068 t = tree_cons (NULL_TREE, new_name, t);
3069 for (i = 1; i < nunits; i++)
3071 /* Create: new_name_i = new_name + step_expr */
3072 enum tree_code code = POINTER_TYPE_P (scalar_type)
3073 ? POINTER_PLUS_EXPR : PLUS_EXPR;
3074 init_stmt = gimple_build_assign_with_ops (code, new_var,
3075 new_name, step_expr);
3076 new_name = make_ssa_name (new_var, init_stmt);
3077 gimple_assign_set_lhs (init_stmt, new_name);
3079 new_bb = gsi_insert_on_edge_immediate (pe, init_stmt);
3080 gcc_assert (!new_bb);
3082 if (vect_print_dump_info (REPORT_DETAILS))
3084 fprintf (vect_dump, "created new init_stmt: ");
3085 print_gimple_stmt (vect_dump, init_stmt, 0, TDF_SLIM);
3087 t = tree_cons (NULL_TREE, new_name, t);
3089 /* Create a vector from [new_name_0, new_name_1, ..., new_name_nunits-1] */
3090 vec = build_constructor_from_list (vectype, nreverse (t));
3091 vec_init = vect_init_vector (iv_phi, vec, vectype, NULL);
3095 /* Create the vector that holds the step of the induction. */
3096 if (nested_in_vect_loop)
3097 /* iv_loop is nested in the loop to be vectorized. Generate:
3098 vec_step = [S, S, S, S] */
3099 new_name = step_expr;
3102 /* iv_loop is the loop to be vectorized. Generate:
3103 vec_step = [VF*S, VF*S, VF*S, VF*S] */
3104 expr = build_int_cst (TREE_TYPE (step_expr), vf);
3105 new_name = fold_build2 (MULT_EXPR, TREE_TYPE (step_expr),
3109 t = unshare_expr (new_name);
3110 gcc_assert (CONSTANT_CLASS_P (new_name));
3111 stepvectype = get_vectype_for_scalar_type (TREE_TYPE (new_name));
3112 gcc_assert (stepvectype);
3113 vec = build_vector_from_val (stepvectype, t);
3114 vec_step = vect_init_vector (iv_phi, vec, stepvectype, NULL);
3117 /* Create the following def-use cycle:
3122 vec_iv = PHI <vec_init, vec_loop>
3126 vec_loop = vec_iv + vec_step; */
3128 /* Create the induction-phi that defines the induction-operand. */
3129 vec_dest = vect_get_new_vect_var (vectype, vect_simple_var, "vec_iv_");
3130 add_referenced_var (vec_dest);
3131 induction_phi = create_phi_node (vec_dest, iv_loop->header);
3132 set_vinfo_for_stmt (induction_phi,
3133 new_stmt_vec_info (induction_phi, loop_vinfo, NULL));
3134 induc_def = PHI_RESULT (induction_phi);
3136 /* Create the iv update inside the loop */
3137 new_stmt = gimple_build_assign_with_ops (PLUS_EXPR, vec_dest,
3138 induc_def, vec_step);
3139 vec_def = make_ssa_name (vec_dest, new_stmt);
3140 gimple_assign_set_lhs (new_stmt, vec_def);
3141 gsi_insert_before (&si, new_stmt, GSI_SAME_STMT);
3142 set_vinfo_for_stmt (new_stmt, new_stmt_vec_info (new_stmt, loop_vinfo,
3145 /* Set the arguments of the phi node: */
3146 add_phi_arg (induction_phi, vec_init, pe, UNKNOWN_LOCATION);
3147 add_phi_arg (induction_phi, vec_def, loop_latch_edge (iv_loop),
3151 /* In case that vectorization factor (VF) is bigger than the number
3152 of elements that we can fit in a vectype (nunits), we have to generate
3153 more than one vector stmt - i.e - we need to "unroll" the
3154 vector stmt by a factor VF/nunits. For more details see documentation
3155 in vectorizable_operation. */
3159 stmt_vec_info prev_stmt_vinfo;
3160 /* FORNOW. This restriction should be relaxed. */
3161 gcc_assert (!nested_in_vect_loop);
3163 /* Create the vector that holds the step of the induction. */
3164 expr = build_int_cst (TREE_TYPE (step_expr), nunits);
3165 new_name = fold_build2 (MULT_EXPR, TREE_TYPE (step_expr),
3167 t = unshare_expr (new_name);
3168 gcc_assert (CONSTANT_CLASS_P (new_name));
3169 vec = build_vector_from_val (stepvectype, t);
3170 vec_step = vect_init_vector (iv_phi, vec, stepvectype, NULL);
3172 vec_def = induc_def;
3173 prev_stmt_vinfo = vinfo_for_stmt (induction_phi);
3174 for (i = 1; i < ncopies; i++)
3176 /* vec_i = vec_prev + vec_step */
3177 new_stmt = gimple_build_assign_with_ops (PLUS_EXPR, vec_dest,
3179 vec_def = make_ssa_name (vec_dest, new_stmt);
3180 gimple_assign_set_lhs (new_stmt, vec_def);
3182 gsi_insert_before (&si, new_stmt, GSI_SAME_STMT);
3183 if (!useless_type_conversion_p (resvectype, vectype))
3185 new_stmt = gimple_build_assign_with_ops
3187 vect_get_new_vect_var (resvectype, vect_simple_var,
3189 build1 (VIEW_CONVERT_EXPR, resvectype,
3190 gimple_assign_lhs (new_stmt)), NULL_TREE);
3191 gimple_assign_set_lhs (new_stmt,
3193 (gimple_assign_lhs (new_stmt), new_stmt));
3194 gsi_insert_before (&si, new_stmt, GSI_SAME_STMT);
3196 set_vinfo_for_stmt (new_stmt,
3197 new_stmt_vec_info (new_stmt, loop_vinfo, NULL));
3198 STMT_VINFO_RELATED_STMT (prev_stmt_vinfo) = new_stmt;
3199 prev_stmt_vinfo = vinfo_for_stmt (new_stmt);
3203 if (nested_in_vect_loop)
3205 /* Find the loop-closed exit-phi of the induction, and record
3206 the final vector of induction results: */
3208 FOR_EACH_IMM_USE_FAST (use_p, imm_iter, loop_arg)
3210 if (!flow_bb_inside_loop_p (iv_loop, gimple_bb (USE_STMT (use_p))))
3212 exit_phi = USE_STMT (use_p);
3218 stmt_vec_info stmt_vinfo = vinfo_for_stmt (exit_phi);
3219 /* FORNOW. Currently not supporting the case that an inner-loop induction
3220 is not used in the outer-loop (i.e. only outside the outer-loop). */
3221 gcc_assert (STMT_VINFO_RELEVANT_P (stmt_vinfo)
3222 && !STMT_VINFO_LIVE_P (stmt_vinfo));
3224 STMT_VINFO_VEC_STMT (stmt_vinfo) = new_stmt;
3225 if (vect_print_dump_info (REPORT_DETAILS))
3227 fprintf (vect_dump, "vector of inductions after inner-loop:");
3228 print_gimple_stmt (vect_dump, new_stmt, 0, TDF_SLIM);
3234 if (vect_print_dump_info (REPORT_DETAILS))
3236 fprintf (vect_dump, "transform induction: created def-use cycle: ");
3237 print_gimple_stmt (vect_dump, induction_phi, 0, TDF_SLIM);
3238 fprintf (vect_dump, "\n");
3239 print_gimple_stmt (vect_dump, SSA_NAME_DEF_STMT (vec_def), 0, TDF_SLIM);
3242 STMT_VINFO_VEC_STMT (phi_info) = induction_phi;
3243 if (!useless_type_conversion_p (resvectype, vectype))
3245 new_stmt = gimple_build_assign_with_ops
3247 vect_get_new_vect_var (resvectype, vect_simple_var, "vec_iv_"),
3248 build1 (VIEW_CONVERT_EXPR, resvectype, induc_def), NULL_TREE);
3249 induc_def = make_ssa_name (gimple_assign_lhs (new_stmt), new_stmt);
3250 gimple_assign_set_lhs (new_stmt, induc_def);
3251 si = gsi_start_bb (bb);
3252 gsi_insert_before (&si, new_stmt, GSI_SAME_STMT);
3253 set_vinfo_for_stmt (new_stmt,
3254 new_stmt_vec_info (new_stmt, loop_vinfo, NULL));
3255 STMT_VINFO_RELATED_STMT (vinfo_for_stmt (new_stmt))
3256 = STMT_VINFO_RELATED_STMT (vinfo_for_stmt (induction_phi));
3263 /* Function get_initial_def_for_reduction
3266 STMT - a stmt that performs a reduction operation in the loop.
3267 INIT_VAL - the initial value of the reduction variable
3270 ADJUSTMENT_DEF - a tree that holds a value to be added to the final result
3271 of the reduction (used for adjusting the epilog - see below).
3272 Return a vector variable, initialized according to the operation that STMT
3273 performs. This vector will be used as the initial value of the
3274 vector of partial results.
3276 Option1 (adjust in epilog): Initialize the vector as follows:
3277 add/bit or/xor: [0,0,...,0,0]
3278 mult/bit and: [1,1,...,1,1]
3279 min/max/cond_expr: [init_val,init_val,..,init_val,init_val]
3280 and when necessary (e.g. add/mult case) let the caller know
3281 that it needs to adjust the result by init_val.
3283 Option2: Initialize the vector as follows:
3284 add/bit or/xor: [init_val,0,0,...,0]
3285 mult/bit and: [init_val,1,1,...,1]
3286 min/max/cond_expr: [init_val,init_val,...,init_val]
3287 and no adjustments are needed.
3289 For example, for the following code:
3295 STMT is 's = s + a[i]', and the reduction variable is 's'.
3296 For a vector of 4 units, we want to return either [0,0,0,init_val],
3297 or [0,0,0,0] and let the caller know that it needs to adjust
3298 the result at the end by 'init_val'.
3300 FORNOW, we are using the 'adjust in epilog' scheme, because this way the
3301 initialization vector is simpler (same element in all entries), if
3302 ADJUSTMENT_DEF is not NULL, and Option2 otherwise.
3304 A cost model should help decide between these two schemes. */
3307 get_initial_def_for_reduction (gimple stmt, tree init_val,
3308 tree *adjustment_def)
3310 stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt);
3311 loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_vinfo);
3312 struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
3313 tree scalar_type = TREE_TYPE (init_val);
3314 tree vectype = get_vectype_for_scalar_type (scalar_type);
3316 enum tree_code code = gimple_assign_rhs_code (stmt);
3321 bool nested_in_vect_loop = false;
3323 REAL_VALUE_TYPE real_init_val = dconst0;
3324 int int_init_val = 0;
3325 gimple def_stmt = NULL;
3327 gcc_assert (vectype);
3328 nunits = TYPE_VECTOR_SUBPARTS (vectype);
3330 gcc_assert (POINTER_TYPE_P (scalar_type) || INTEGRAL_TYPE_P (scalar_type)
3331 || SCALAR_FLOAT_TYPE_P (scalar_type));
3333 if (nested_in_vect_loop_p (loop, stmt))
3334 nested_in_vect_loop = true;
3336 gcc_assert (loop == (gimple_bb (stmt))->loop_father);
3338 /* In case of double reduction we only create a vector variable to be put
3339 in the reduction phi node. The actual statement creation is done in
3340 vect_create_epilog_for_reduction. */
3341 if (adjustment_def && nested_in_vect_loop
3342 && TREE_CODE (init_val) == SSA_NAME
3343 && (def_stmt = SSA_NAME_DEF_STMT (init_val))
3344 && gimple_code (def_stmt) == GIMPLE_PHI
3345 && flow_bb_inside_loop_p (loop, gimple_bb (def_stmt))
3346 && vinfo_for_stmt (def_stmt)
3347 && STMT_VINFO_DEF_TYPE (vinfo_for_stmt (def_stmt))
3348 == vect_double_reduction_def)
3350 *adjustment_def = NULL;
3351 return vect_create_destination_var (init_val, vectype);
3354 if (TREE_CONSTANT (init_val))
3356 if (SCALAR_FLOAT_TYPE_P (scalar_type))
3357 init_value = build_real (scalar_type, TREE_REAL_CST (init_val));
3359 init_value = build_int_cst (scalar_type, TREE_INT_CST_LOW (init_val));
3362 init_value = init_val;
3366 case WIDEN_SUM_EXPR:
3374 /* ADJUSMENT_DEF is NULL when called from
3375 vect_create_epilog_for_reduction to vectorize double reduction. */
3378 if (nested_in_vect_loop)
3379 *adjustment_def = vect_get_vec_def_for_operand (init_val, stmt,
3382 *adjustment_def = init_val;
3385 if (code == MULT_EXPR)
3387 real_init_val = dconst1;
3391 if (code == BIT_AND_EXPR)
3394 if (SCALAR_FLOAT_TYPE_P (scalar_type))
3395 def_for_init = build_real (scalar_type, real_init_val);
3397 def_for_init = build_int_cst (scalar_type, int_init_val);
3399 /* Create a vector of '0' or '1' except the first element. */
3400 for (i = nunits - 2; i >= 0; --i)
3401 t = tree_cons (NULL_TREE, def_for_init, t);
3403 /* Option1: the first element is '0' or '1' as well. */
3406 t = tree_cons (NULL_TREE, def_for_init, t);
3407 init_def = build_vector (vectype, t);
3411 /* Option2: the first element is INIT_VAL. */
3412 t = tree_cons (NULL_TREE, init_value, t);
3413 if (TREE_CONSTANT (init_val))
3414 init_def = build_vector (vectype, t);
3416 init_def = build_constructor_from_list (vectype, t);
3425 *adjustment_def = NULL_TREE;
3426 init_def = vect_get_vec_def_for_operand (init_val, stmt, NULL);
3430 init_def = build_vector_from_val (vectype, init_value);
3441 /* Function vect_create_epilog_for_reduction
3443 Create code at the loop-epilog to finalize the result of a reduction
3446 VECT_DEFS is list of vector of partial results, i.e., the lhs's of vector
3447 reduction statements.
3448 STMT is the scalar reduction stmt that is being vectorized.
3449 NCOPIES is > 1 in case the vectorization factor (VF) is bigger than the
3450 number of elements that we can fit in a vectype (nunits). In this case
3451 we have to generate more than one vector stmt - i.e - we need to "unroll"
3452 the vector stmt by a factor VF/nunits. For more details see documentation
3453 in vectorizable_operation.
3454 REDUC_CODE is the tree-code for the epilog reduction.
3455 REDUCTION_PHIS is a list of the phi-nodes that carry the reduction
3457 REDUC_INDEX is the index of the operand in the right hand side of the
3458 statement that is defined by REDUCTION_PHI.
3459 DOUBLE_REDUC is TRUE if double reduction phi nodes should be handled.
3460 SLP_NODE is an SLP node containing a group of reduction statements. The
3461 first one in this group is STMT.
3464 1. Creates the reduction def-use cycles: sets the arguments for
3466 The loop-entry argument is the vectorized initial-value of the reduction.
3467 The loop-latch argument is taken from VECT_DEFS - the vector of partial
3469 2. "Reduces" each vector of partial results VECT_DEFS into a single result,
3470 by applying the operation specified by REDUC_CODE if available, or by
3471 other means (whole-vector shifts or a scalar loop).
3472 The function also creates a new phi node at the loop exit to preserve
3473 loop-closed form, as illustrated below.
3475 The flow at the entry to this function:
3478 vec_def = phi <null, null> # REDUCTION_PHI
3479 VECT_DEF = vector_stmt # vectorized form of STMT
3480 s_loop = scalar_stmt # (scalar) STMT
3482 s_out0 = phi <s_loop> # (scalar) EXIT_PHI
3486 The above is transformed by this function into:
3489 vec_def = phi <vec_init, VECT_DEF> # REDUCTION_PHI
3490 VECT_DEF = vector_stmt # vectorized form of STMT
3491 s_loop = scalar_stmt # (scalar) STMT
3493 s_out0 = phi <s_loop> # (scalar) EXIT_PHI
3494 v_out1 = phi <VECT_DEF> # NEW_EXIT_PHI
3495 v_out2 = reduce <v_out1>
3496 s_out3 = extract_field <v_out2, 0>
3497 s_out4 = adjust_result <s_out3>
3503 vect_create_epilog_for_reduction (VEC (tree, heap) *vect_defs, gimple stmt,
3504 int ncopies, enum tree_code reduc_code,
3505 VEC (gimple, heap) *reduction_phis,
3506 int reduc_index, bool double_reduc,
3509 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
3510 stmt_vec_info prev_phi_info;
3512 enum machine_mode mode;
3513 loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info);
3514 struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo), *outer_loop = NULL;
3515 basic_block exit_bb;
3518 gimple new_phi = NULL, phi;
3519 gimple_stmt_iterator exit_gsi;
3521 tree new_temp = NULL_TREE, new_dest, new_name, new_scalar_dest;
3522 gimple epilog_stmt = NULL;
3523 enum tree_code code = gimple_assign_rhs_code (stmt);
3525 tree bitsize, bitpos;
3526 tree adjustment_def = NULL;
3527 tree vec_initial_def = NULL;
3528 tree reduction_op, expr, def;
3529 tree orig_name, scalar_result;
3530 imm_use_iterator imm_iter, phi_imm_iter;
3531 use_operand_p use_p, phi_use_p;
3532 bool extract_scalar_result = false;
3533 gimple use_stmt, orig_stmt, reduction_phi = NULL;
3534 bool nested_in_vect_loop = false;
3535 VEC (gimple, heap) *new_phis = NULL;
3536 VEC (gimple, heap) *inner_phis = NULL;
3537 enum vect_def_type dt = vect_unknown_def_type;
3539 VEC (tree, heap) *scalar_results = NULL;
3540 unsigned int group_size = 1, k, ratio;
3541 VEC (tree, heap) *vec_initial_defs = NULL;
3542 VEC (gimple, heap) *phis;
3543 bool slp_reduc = false;
3544 tree new_phi_result;
3545 gimple inner_phi = NULL;
3548 group_size = VEC_length (gimple, SLP_TREE_SCALAR_STMTS (slp_node));
3550 if (nested_in_vect_loop_p (loop, stmt))
3554 nested_in_vect_loop = true;
3555 gcc_assert (!slp_node);
3558 switch (get_gimple_rhs_class (gimple_assign_rhs_code (stmt)))
3560 case GIMPLE_SINGLE_RHS:
3561 gcc_assert (TREE_OPERAND_LENGTH (gimple_assign_rhs1 (stmt))
3563 reduction_op = TREE_OPERAND (gimple_assign_rhs1 (stmt), reduc_index);
3565 case GIMPLE_UNARY_RHS:
3566 reduction_op = gimple_assign_rhs1 (stmt);
3568 case GIMPLE_BINARY_RHS:
3569 reduction_op = reduc_index ?
3570 gimple_assign_rhs2 (stmt) : gimple_assign_rhs1 (stmt);
3572 case GIMPLE_TERNARY_RHS:
3573 reduction_op = gimple_op (stmt, reduc_index + 1);
3579 vectype = get_vectype_for_scalar_type (TREE_TYPE (reduction_op));
3580 gcc_assert (vectype);
3581 mode = TYPE_MODE (vectype);
3583 /* 1. Create the reduction def-use cycle:
3584 Set the arguments of REDUCTION_PHIS, i.e., transform
3587 vec_def = phi <null, null> # REDUCTION_PHI
3588 VECT_DEF = vector_stmt # vectorized form of STMT
3594 vec_def = phi <vec_init, VECT_DEF> # REDUCTION_PHI
3595 VECT_DEF = vector_stmt # vectorized form of STMT
3598 (in case of SLP, do it for all the phis). */
3600 /* Get the loop-entry arguments. */
3602 vect_get_vec_defs (reduction_op, NULL_TREE, stmt, &vec_initial_defs,
3603 NULL, slp_node, reduc_index);
3606 vec_initial_defs = VEC_alloc (tree, heap, 1);
3607 /* For the case of reduction, vect_get_vec_def_for_operand returns
3608 the scalar def before the loop, that defines the initial value
3609 of the reduction variable. */
3610 vec_initial_def = vect_get_vec_def_for_operand (reduction_op, stmt,
3612 VEC_quick_push (tree, vec_initial_defs, vec_initial_def);
3615 /* Set phi nodes arguments. */
3616 FOR_EACH_VEC_ELT (gimple, reduction_phis, i, phi)
3618 tree vec_init_def = VEC_index (tree, vec_initial_defs, i);
3619 tree def = VEC_index (tree, vect_defs, i);
3620 for (j = 0; j < ncopies; j++)
3622 /* Set the loop-entry arg of the reduction-phi. */
3623 add_phi_arg (phi, vec_init_def, loop_preheader_edge (loop),
3626 /* Set the loop-latch arg for the reduction-phi. */
3628 def = vect_get_vec_def_for_stmt_copy (vect_unknown_def_type, def);
3630 add_phi_arg (phi, def, loop_latch_edge (loop), UNKNOWN_LOCATION);
3632 if (vect_print_dump_info (REPORT_DETAILS))
3634 fprintf (vect_dump, "transform reduction: created def-use"
3636 print_gimple_stmt (vect_dump, phi, 0, TDF_SLIM);
3637 fprintf (vect_dump, "\n");
3638 print_gimple_stmt (vect_dump, SSA_NAME_DEF_STMT (def), 0,
3642 phi = STMT_VINFO_RELATED_STMT (vinfo_for_stmt (phi));
3646 VEC_free (tree, heap, vec_initial_defs);
3648 /* 2. Create epilog code.
3649 The reduction epilog code operates across the elements of the vector
3650 of partial results computed by the vectorized loop.
3651 The reduction epilog code consists of:
3653 step 1: compute the scalar result in a vector (v_out2)
3654 step 2: extract the scalar result (s_out3) from the vector (v_out2)
3655 step 3: adjust the scalar result (s_out3) if needed.
3657 Step 1 can be accomplished using one the following three schemes:
3658 (scheme 1) using reduc_code, if available.
3659 (scheme 2) using whole-vector shifts, if available.
3660 (scheme 3) using a scalar loop. In this case steps 1+2 above are
3663 The overall epilog code looks like this:
3665 s_out0 = phi <s_loop> # original EXIT_PHI
3666 v_out1 = phi <VECT_DEF> # NEW_EXIT_PHI
3667 v_out2 = reduce <v_out1> # step 1
3668 s_out3 = extract_field <v_out2, 0> # step 2
3669 s_out4 = adjust_result <s_out3> # step 3
3671 (step 3 is optional, and steps 1 and 2 may be combined).
3672 Lastly, the uses of s_out0 are replaced by s_out4. */
3675 /* 2.1 Create new loop-exit-phis to preserve loop-closed form:
3676 v_out1 = phi <VECT_DEF>
3677 Store them in NEW_PHIS. */
3679 exit_bb = single_exit (loop)->dest;
3680 prev_phi_info = NULL;
3681 new_phis = VEC_alloc (gimple, heap, VEC_length (tree, vect_defs));
3682 FOR_EACH_VEC_ELT (tree, vect_defs, i, def)
3684 for (j = 0; j < ncopies; j++)
3686 phi = create_phi_node (SSA_NAME_VAR (def), exit_bb);
3687 set_vinfo_for_stmt (phi, new_stmt_vec_info (phi, loop_vinfo, NULL));
3689 VEC_quick_push (gimple, new_phis, phi);
3692 def = vect_get_vec_def_for_stmt_copy (dt, def);
3693 STMT_VINFO_RELATED_STMT (prev_phi_info) = phi;
3696 SET_PHI_ARG_DEF (phi, single_exit (loop)->dest_idx, def);
3697 prev_phi_info = vinfo_for_stmt (phi);
3701 /* The epilogue is created for the outer-loop, i.e., for the loop being
3702 vectorized. Create exit phis for the outer loop. */
3706 exit_bb = single_exit (loop)->dest;
3707 inner_phis = VEC_alloc (gimple, heap, VEC_length (tree, vect_defs));
3708 FOR_EACH_VEC_ELT (gimple, new_phis, i, phi)
3710 gimple outer_phi = create_phi_node (SSA_NAME_VAR (PHI_RESULT (phi)),
3712 SET_PHI_ARG_DEF (outer_phi, single_exit (loop)->dest_idx,
3714 set_vinfo_for_stmt (outer_phi, new_stmt_vec_info (outer_phi,
3716 VEC_quick_push (gimple, inner_phis, phi);
3717 VEC_replace (gimple, new_phis, i, outer_phi);
3718 prev_phi_info = vinfo_for_stmt (outer_phi);
3719 while (STMT_VINFO_RELATED_STMT (vinfo_for_stmt (phi)))
3721 phi = STMT_VINFO_RELATED_STMT (vinfo_for_stmt (phi));
3722 outer_phi = create_phi_node (SSA_NAME_VAR (PHI_RESULT (phi)),
3724 SET_PHI_ARG_DEF (outer_phi, single_exit (loop)->dest_idx,
3726 set_vinfo_for_stmt (outer_phi, new_stmt_vec_info (outer_phi,
3728 STMT_VINFO_RELATED_STMT (prev_phi_info) = outer_phi;
3729 prev_phi_info = vinfo_for_stmt (outer_phi);
3734 exit_gsi = gsi_after_labels (exit_bb);
3736 /* 2.2 Get the relevant tree-code to use in the epilog for schemes 2,3
3737 (i.e. when reduc_code is not available) and in the final adjustment
3738 code (if needed). Also get the original scalar reduction variable as
3739 defined in the loop. In case STMT is a "pattern-stmt" (i.e. - it
3740 represents a reduction pattern), the tree-code and scalar-def are
3741 taken from the original stmt that the pattern-stmt (STMT) replaces.
3742 Otherwise (it is a regular reduction) - the tree-code and scalar-def
3743 are taken from STMT. */
3745 orig_stmt = STMT_VINFO_RELATED_STMT (stmt_info);
3748 /* Regular reduction */
3753 /* Reduction pattern */
3754 stmt_vec_info stmt_vinfo = vinfo_for_stmt (orig_stmt);
3755 gcc_assert (STMT_VINFO_IN_PATTERN_P (stmt_vinfo));
3756 gcc_assert (STMT_VINFO_RELATED_STMT (stmt_vinfo) == stmt);
3759 code = gimple_assign_rhs_code (orig_stmt);
3760 /* For MINUS_EXPR the initial vector is [init_val,0,...,0], therefore,
3761 partial results are added and not subtracted. */
3762 if (code == MINUS_EXPR)
3765 scalar_dest = gimple_assign_lhs (orig_stmt);
3766 scalar_type = TREE_TYPE (scalar_dest);
3767 scalar_results = VEC_alloc (tree, heap, group_size);
3768 new_scalar_dest = vect_create_destination_var (scalar_dest, NULL);
3769 bitsize = TYPE_SIZE (scalar_type);
3771 /* In case this is a reduction in an inner-loop while vectorizing an outer
3772 loop - we don't need to extract a single scalar result at the end of the
3773 inner-loop (unless it is double reduction, i.e., the use of reduction is
3774 outside the outer-loop). The final vector of partial results will be used
3775 in the vectorized outer-loop, or reduced to a scalar result at the end of
3777 if (nested_in_vect_loop && !double_reduc)
3778 goto vect_finalize_reduction;
3780 /* SLP reduction without reduction chain, e.g.,
3784 b2 = operation (b1) */
3785 slp_reduc = (slp_node && !GROUP_FIRST_ELEMENT (vinfo_for_stmt (stmt)));
3787 /* In case of reduction chain, e.g.,
3790 a3 = operation (a2),
3792 we may end up with more than one vector result. Here we reduce them to
3794 if (GROUP_FIRST_ELEMENT (vinfo_for_stmt (stmt)))
3796 tree first_vect = PHI_RESULT (VEC_index (gimple, new_phis, 0));
3798 gimple new_vec_stmt = NULL;
3800 vec_dest = vect_create_destination_var (scalar_dest, vectype);
3801 for (k = 1; k < VEC_length (gimple, new_phis); k++)
3803 gimple next_phi = VEC_index (gimple, new_phis, k);
3804 tree second_vect = PHI_RESULT (next_phi);
3806 tmp = build2 (code, vectype, first_vect, second_vect);
3807 new_vec_stmt = gimple_build_assign (vec_dest, tmp);
3808 first_vect = make_ssa_name (vec_dest, new_vec_stmt);
3809 gimple_assign_set_lhs (new_vec_stmt, first_vect);
3810 gsi_insert_before (&exit_gsi, new_vec_stmt, GSI_SAME_STMT);
3813 new_phi_result = first_vect;
3816 VEC_truncate (gimple, new_phis, 0);
3817 VEC_safe_push (gimple, heap, new_phis, new_vec_stmt);
3821 new_phi_result = PHI_RESULT (VEC_index (gimple, new_phis, 0));
3823 /* 2.3 Create the reduction code, using one of the three schemes described
3824 above. In SLP we simply need to extract all the elements from the
3825 vector (without reducing them), so we use scalar shifts. */
3826 if (reduc_code != ERROR_MARK && !slp_reduc)
3830 /*** Case 1: Create:
3831 v_out2 = reduc_expr <v_out1> */
3833 if (vect_print_dump_info (REPORT_DETAILS))
3834 fprintf (vect_dump, "Reduce using direct vector reduction.");
3836 vec_dest = vect_create_destination_var (scalar_dest, vectype);
3837 tmp = build1 (reduc_code, vectype, new_phi_result);
3838 epilog_stmt = gimple_build_assign (vec_dest, tmp);
3839 new_temp = make_ssa_name (vec_dest, epilog_stmt);
3840 gimple_assign_set_lhs (epilog_stmt, new_temp);
3841 gsi_insert_before (&exit_gsi, epilog_stmt, GSI_SAME_STMT);
3843 extract_scalar_result = true;
3847 enum tree_code shift_code = ERROR_MARK;
3848 bool have_whole_vector_shift = true;
3850 int element_bitsize = tree_low_cst (bitsize, 1);
3851 int vec_size_in_bits = tree_low_cst (TYPE_SIZE (vectype), 1);
3854 if (optab_handler (vec_shr_optab, mode) != CODE_FOR_nothing)
3855 shift_code = VEC_RSHIFT_EXPR;
3857 have_whole_vector_shift = false;
3859 /* Regardless of whether we have a whole vector shift, if we're
3860 emulating the operation via tree-vect-generic, we don't want
3861 to use it. Only the first round of the reduction is likely
3862 to still be profitable via emulation. */
3863 /* ??? It might be better to emit a reduction tree code here, so that
3864 tree-vect-generic can expand the first round via bit tricks. */
3865 if (!VECTOR_MODE_P (mode))
3866 have_whole_vector_shift = false;
3869 optab optab = optab_for_tree_code (code, vectype, optab_default);
3870 if (optab_handler (optab, mode) == CODE_FOR_nothing)
3871 have_whole_vector_shift = false;
3874 if (have_whole_vector_shift && !slp_reduc)
3876 /*** Case 2: Create:
3877 for (offset = VS/2; offset >= element_size; offset/=2)
3879 Create: va' = vec_shift <va, offset>
3880 Create: va = vop <va, va'>
3883 if (vect_print_dump_info (REPORT_DETAILS))
3884 fprintf (vect_dump, "Reduce using vector shifts");
3886 vec_dest = vect_create_destination_var (scalar_dest, vectype);
3887 new_temp = new_phi_result;
3888 for (bit_offset = vec_size_in_bits/2;
3889 bit_offset >= element_bitsize;
3892 tree bitpos = size_int (bit_offset);
3894 epilog_stmt = gimple_build_assign_with_ops (shift_code,
3895 vec_dest, new_temp, bitpos);
3896 new_name = make_ssa_name (vec_dest, epilog_stmt);
3897 gimple_assign_set_lhs (epilog_stmt, new_name);
3898 gsi_insert_before (&exit_gsi, epilog_stmt, GSI_SAME_STMT);
3900 epilog_stmt = gimple_build_assign_with_ops (code, vec_dest,
3901 new_name, new_temp);
3902 new_temp = make_ssa_name (vec_dest, epilog_stmt);
3903 gimple_assign_set_lhs (epilog_stmt, new_temp);
3904 gsi_insert_before (&exit_gsi, epilog_stmt, GSI_SAME_STMT);
3907 extract_scalar_result = true;
3913 /*** Case 3: Create:
3914 s = extract_field <v_out2, 0>
3915 for (offset = element_size;
3916 offset < vector_size;
3917 offset += element_size;)
3919 Create: s' = extract_field <v_out2, offset>
3920 Create: s = op <s, s'> // For non SLP cases
3923 if (vect_print_dump_info (REPORT_DETAILS))
3924 fprintf (vect_dump, "Reduce using scalar code. ");
3926 vec_size_in_bits = tree_low_cst (TYPE_SIZE (vectype), 1);
3927 FOR_EACH_VEC_ELT (gimple, new_phis, i, new_phi)
3929 if (gimple_code (new_phi) == GIMPLE_PHI)
3930 vec_temp = PHI_RESULT (new_phi);
3932 vec_temp = gimple_assign_lhs (new_phi);
3933 rhs = build3 (BIT_FIELD_REF, scalar_type, vec_temp, bitsize,
3935 epilog_stmt = gimple_build_assign (new_scalar_dest, rhs);
3936 new_temp = make_ssa_name (new_scalar_dest, epilog_stmt);
3937 gimple_assign_set_lhs (epilog_stmt, new_temp);
3938 gsi_insert_before (&exit_gsi, epilog_stmt, GSI_SAME_STMT);
3940 /* In SLP we don't need to apply reduction operation, so we just
3941 collect s' values in SCALAR_RESULTS. */
3943 VEC_safe_push (tree, heap, scalar_results, new_temp);
3945 for (bit_offset = element_bitsize;
3946 bit_offset < vec_size_in_bits;
3947 bit_offset += element_bitsize)
3949 tree bitpos = bitsize_int (bit_offset);
3950 tree rhs = build3 (BIT_FIELD_REF, scalar_type, vec_temp,
3953 epilog_stmt = gimple_build_assign (new_scalar_dest, rhs);
3954 new_name = make_ssa_name (new_scalar_dest, epilog_stmt);
3955 gimple_assign_set_lhs (epilog_stmt, new_name);
3956 gsi_insert_before (&exit_gsi, epilog_stmt, GSI_SAME_STMT);
3960 /* In SLP we don't need to apply reduction operation, so
3961 we just collect s' values in SCALAR_RESULTS. */
3962 new_temp = new_name;
3963 VEC_safe_push (tree, heap, scalar_results, new_name);
3967 epilog_stmt = gimple_build_assign_with_ops (code,
3968 new_scalar_dest, new_name, new_temp);
3969 new_temp = make_ssa_name (new_scalar_dest, epilog_stmt);
3970 gimple_assign_set_lhs (epilog_stmt, new_temp);
3971 gsi_insert_before (&exit_gsi, epilog_stmt, GSI_SAME_STMT);
3976 /* The only case where we need to reduce scalar results in SLP, is
3977 unrolling. If the size of SCALAR_RESULTS is greater than
3978 GROUP_SIZE, we reduce them combining elements modulo
3982 tree res, first_res, new_res;
3985 /* Reduce multiple scalar results in case of SLP unrolling. */
3986 for (j = group_size; VEC_iterate (tree, scalar_results, j, res);
3989 first_res = VEC_index (tree, scalar_results, j % group_size);
3990 new_stmt = gimple_build_assign_with_ops (code,
3991 new_scalar_dest, first_res, res);
3992 new_res = make_ssa_name (new_scalar_dest, new_stmt);
3993 gimple_assign_set_lhs (new_stmt, new_res);
3994 gsi_insert_before (&exit_gsi, new_stmt, GSI_SAME_STMT);
3995 VEC_replace (tree, scalar_results, j % group_size, new_res);
3999 /* Not SLP - we have one scalar to keep in SCALAR_RESULTS. */
4000 VEC_safe_push (tree, heap, scalar_results, new_temp);
4002 extract_scalar_result = false;
4006 /* 2.4 Extract the final scalar result. Create:
4007 s_out3 = extract_field <v_out2, bitpos> */
4009 if (extract_scalar_result)
4013 if (vect_print_dump_info (REPORT_DETAILS))
4014 fprintf (vect_dump, "extract scalar result");
4016 if (BYTES_BIG_ENDIAN)
4017 bitpos = size_binop (MULT_EXPR,
4018 bitsize_int (TYPE_VECTOR_SUBPARTS (vectype) - 1),
4019 TYPE_SIZE (scalar_type));
4021 bitpos = bitsize_zero_node;
4023 rhs = build3 (BIT_FIELD_REF, scalar_type, new_temp, bitsize, bitpos);
4024 epilog_stmt = gimple_build_assign (new_scalar_dest, rhs);
4025 new_temp = make_ssa_name (new_scalar_dest, epilog_stmt);
4026 gimple_assign_set_lhs (epilog_stmt, new_temp);
4027 gsi_insert_before (&exit_gsi, epilog_stmt, GSI_SAME_STMT);
4028 VEC_safe_push (tree, heap, scalar_results, new_temp);
4031 vect_finalize_reduction:
4036 /* 2.5 Adjust the final result by the initial value of the reduction
4037 variable. (When such adjustment is not needed, then
4038 'adjustment_def' is zero). For example, if code is PLUS we create:
4039 new_temp = loop_exit_def + adjustment_def */
4043 gcc_assert (!slp_reduc);
4044 if (nested_in_vect_loop)
4046 new_phi = VEC_index (gimple, new_phis, 0);
4047 gcc_assert (TREE_CODE (TREE_TYPE (adjustment_def)) == VECTOR_TYPE);
4048 expr = build2 (code, vectype, PHI_RESULT (new_phi), adjustment_def);
4049 new_dest = vect_create_destination_var (scalar_dest, vectype);
4053 new_temp = VEC_index (tree, scalar_results, 0);
4054 gcc_assert (TREE_CODE (TREE_TYPE (adjustment_def)) != VECTOR_TYPE);
4055 expr = build2 (code, scalar_type, new_temp, adjustment_def);
4056 new_dest = vect_create_destination_var (scalar_dest, scalar_type);
4059 epilog_stmt = gimple_build_assign (new_dest, expr);
4060 new_temp = make_ssa_name (new_dest, epilog_stmt);
4061 gimple_assign_set_lhs (epilog_stmt, new_temp);
4062 SSA_NAME_DEF_STMT (new_temp) = epilog_stmt;
4063 gsi_insert_before (&exit_gsi, epilog_stmt, GSI_SAME_STMT);
4064 if (nested_in_vect_loop)
4066 set_vinfo_for_stmt (epilog_stmt,
4067 new_stmt_vec_info (epilog_stmt, loop_vinfo,
4069 STMT_VINFO_RELATED_STMT (vinfo_for_stmt (epilog_stmt)) =
4070 STMT_VINFO_RELATED_STMT (vinfo_for_stmt (new_phi));
4073 VEC_quick_push (tree, scalar_results, new_temp);
4075 VEC_replace (tree, scalar_results, 0, new_temp);
4078 VEC_replace (tree, scalar_results, 0, new_temp);
4080 VEC_replace (gimple, new_phis, 0, epilog_stmt);
4083 /* 2.6 Handle the loop-exit phis. Replace the uses of scalar loop-exit
4084 phis with new adjusted scalar results, i.e., replace use <s_out0>
4089 s_out0 = phi <s_loop> # (scalar) EXIT_PHI
4090 v_out1 = phi <VECT_DEF> # NEW_EXIT_PHI
4091 v_out2 = reduce <v_out1>
4092 s_out3 = extract_field <v_out2, 0>
4093 s_out4 = adjust_result <s_out3>
4100 s_out0 = phi <s_loop> # (scalar) EXIT_PHI
4101 v_out1 = phi <VECT_DEF> # NEW_EXIT_PHI
4102 v_out2 = reduce <v_out1>
4103 s_out3 = extract_field <v_out2, 0>
4104 s_out4 = adjust_result <s_out3>
4109 /* In SLP reduction chain we reduce vector results into one vector if
4110 necessary, hence we set here GROUP_SIZE to 1. SCALAR_DEST is the LHS of
4111 the last stmt in the reduction chain, since we are looking for the loop
4113 if (GROUP_FIRST_ELEMENT (vinfo_for_stmt (stmt)))
4115 scalar_dest = gimple_assign_lhs (VEC_index (gimple,
4116 SLP_TREE_SCALAR_STMTS (slp_node),
4121 /* In SLP we may have several statements in NEW_PHIS and REDUCTION_PHIS (in
4122 case that GROUP_SIZE is greater than vectorization factor). Therefore, we
4123 need to match SCALAR_RESULTS with corresponding statements. The first
4124 (GROUP_SIZE / number of new vector stmts) scalar results correspond to
4125 the first vector stmt, etc.
4126 (RATIO is equal to (GROUP_SIZE / number of new vector stmts)). */
4127 if (group_size > VEC_length (gimple, new_phis))
4129 ratio = group_size / VEC_length (gimple, new_phis);
4130 gcc_assert (!(group_size % VEC_length (gimple, new_phis)));
4135 for (k = 0; k < group_size; k++)
4139 epilog_stmt = VEC_index (gimple, new_phis, k / ratio);
4140 reduction_phi = VEC_index (gimple, reduction_phis, k / ratio);
4142 inner_phi = VEC_index (gimple, inner_phis, k / ratio);
4147 gimple current_stmt = VEC_index (gimple,
4148 SLP_TREE_SCALAR_STMTS (slp_node), k);
4150 orig_stmt = STMT_VINFO_RELATED_STMT (vinfo_for_stmt (current_stmt));
4151 /* SLP statements can't participate in patterns. */
4152 gcc_assert (!orig_stmt);
4153 scalar_dest = gimple_assign_lhs (current_stmt);
4156 phis = VEC_alloc (gimple, heap, 3);
4157 /* Find the loop-closed-use at the loop exit of the original scalar
4158 result. (The reduction result is expected to have two immediate uses -
4159 one at the latch block, and one at the loop exit). */
4160 FOR_EACH_IMM_USE_FAST (use_p, imm_iter, scalar_dest)
4161 if (!flow_bb_inside_loop_p (loop, gimple_bb (USE_STMT (use_p))))
4162 VEC_safe_push (gimple, heap, phis, USE_STMT (use_p));
4164 /* We expect to have found an exit_phi because of loop-closed-ssa
4166 gcc_assert (!VEC_empty (gimple, phis));
4168 FOR_EACH_VEC_ELT (gimple, phis, i, exit_phi)
4172 stmt_vec_info exit_phi_vinfo = vinfo_for_stmt (exit_phi);
4175 /* FORNOW. Currently not supporting the case that an inner-loop
4176 reduction is not used in the outer-loop (but only outside the
4177 outer-loop), unless it is double reduction. */
4178 gcc_assert ((STMT_VINFO_RELEVANT_P (exit_phi_vinfo)
4179 && !STMT_VINFO_LIVE_P (exit_phi_vinfo))
4182 STMT_VINFO_VEC_STMT (exit_phi_vinfo) = epilog_stmt;
4184 || STMT_VINFO_DEF_TYPE (exit_phi_vinfo)
4185 != vect_double_reduction_def)
4188 /* Handle double reduction:
4190 stmt1: s1 = phi <s0, s2> - double reduction phi (outer loop)
4191 stmt2: s3 = phi <s1, s4> - (regular) reduc phi (inner loop)
4192 stmt3: s4 = use (s3) - (regular) reduc stmt (inner loop)
4193 stmt4: s2 = phi <s4> - double reduction stmt (outer loop)
4195 At that point the regular reduction (stmt2 and stmt3) is
4196 already vectorized, as well as the exit phi node, stmt4.
4197 Here we vectorize the phi node of double reduction, stmt1, and
4198 update all relevant statements. */
4200 /* Go through all the uses of s2 to find double reduction phi
4201 node, i.e., stmt1 above. */
4202 orig_name = PHI_RESULT (exit_phi);
4203 FOR_EACH_IMM_USE_STMT (use_stmt, imm_iter, orig_name)
4205 stmt_vec_info use_stmt_vinfo = vinfo_for_stmt (use_stmt);
4206 stmt_vec_info new_phi_vinfo;
4207 tree vect_phi_init, preheader_arg, vect_phi_res, init_def;
4208 basic_block bb = gimple_bb (use_stmt);
4211 /* Check that USE_STMT is really double reduction phi
4213 if (gimple_code (use_stmt) != GIMPLE_PHI
4214 || gimple_phi_num_args (use_stmt) != 2
4216 || STMT_VINFO_DEF_TYPE (use_stmt_vinfo)
4217 != vect_double_reduction_def
4218 || bb->loop_father != outer_loop)
4221 /* Create vector phi node for double reduction:
4222 vs1 = phi <vs0, vs2>
4223 vs1 was created previously in this function by a call to
4224 vect_get_vec_def_for_operand and is stored in
4226 vs2 is defined by INNER_PHI, the vectorized EXIT_PHI;
4227 vs0 is created here. */
4229 /* Create vector phi node. */
4230 vect_phi = create_phi_node (vec_initial_def, bb);
4231 new_phi_vinfo = new_stmt_vec_info (vect_phi,
4232 loop_vec_info_for_loop (outer_loop), NULL);
4233 set_vinfo_for_stmt (vect_phi, new_phi_vinfo);
4235 /* Create vs0 - initial def of the double reduction phi. */
4236 preheader_arg = PHI_ARG_DEF_FROM_EDGE (use_stmt,
4237 loop_preheader_edge (outer_loop));
4238 init_def = get_initial_def_for_reduction (stmt,
4239 preheader_arg, NULL);
4240 vect_phi_init = vect_init_vector (use_stmt, init_def,
4243 /* Update phi node arguments with vs0 and vs2. */
4244 add_phi_arg (vect_phi, vect_phi_init,
4245 loop_preheader_edge (outer_loop),
4247 add_phi_arg (vect_phi, PHI_RESULT (inner_phi),
4248 loop_latch_edge (outer_loop), UNKNOWN_LOCATION);
4249 if (vect_print_dump_info (REPORT_DETAILS))
4251 fprintf (vect_dump, "created double reduction phi "
4253 print_gimple_stmt (vect_dump, vect_phi, 0, TDF_SLIM);
4256 vect_phi_res = PHI_RESULT (vect_phi);
4258 /* Replace the use, i.e., set the correct vs1 in the regular
4259 reduction phi node. FORNOW, NCOPIES is always 1, so the
4260 loop is redundant. */
4261 use = reduction_phi;
4262 for (j = 0; j < ncopies; j++)
4264 edge pr_edge = loop_preheader_edge (loop);
4265 SET_PHI_ARG_DEF (use, pr_edge->dest_idx, vect_phi_res);
4266 use = STMT_VINFO_RELATED_STMT (vinfo_for_stmt (use));
4272 VEC_free (gimple, heap, phis);
4273 if (nested_in_vect_loop)
4281 phis = VEC_alloc (gimple, heap, 3);
4282 /* Find the loop-closed-use at the loop exit of the original scalar
4283 result. (The reduction result is expected to have two immediate uses,
4284 one at the latch block, and one at the loop exit). For double
4285 reductions we are looking for exit phis of the outer loop. */
4286 FOR_EACH_IMM_USE_FAST (use_p, imm_iter, scalar_dest)
4288 if (!flow_bb_inside_loop_p (loop, gimple_bb (USE_STMT (use_p))))
4289 VEC_safe_push (gimple, heap, phis, USE_STMT (use_p));
4292 if (double_reduc && gimple_code (USE_STMT (use_p)) == GIMPLE_PHI)
4294 tree phi_res = PHI_RESULT (USE_STMT (use_p));
4296 FOR_EACH_IMM_USE_FAST (phi_use_p, phi_imm_iter, phi_res)
4298 if (!flow_bb_inside_loop_p (loop,
4299 gimple_bb (USE_STMT (phi_use_p))))
4300 VEC_safe_push (gimple, heap, phis,
4301 USE_STMT (phi_use_p));
4307 FOR_EACH_VEC_ELT (gimple, phis, i, exit_phi)
4309 /* Replace the uses: */
4310 orig_name = PHI_RESULT (exit_phi);
4311 scalar_result = VEC_index (tree, scalar_results, k);
4312 FOR_EACH_IMM_USE_STMT (use_stmt, imm_iter, orig_name)
4313 FOR_EACH_IMM_USE_ON_STMT (use_p, imm_iter)
4314 SET_USE (use_p, scalar_result);
4317 VEC_free (gimple, heap, phis);
4320 VEC_free (tree, heap, scalar_results);
4321 VEC_free (gimple, heap, new_phis);
4325 /* Function vectorizable_reduction.
4327 Check if STMT performs a reduction operation that can be vectorized.
4328 If VEC_STMT is also passed, vectorize the STMT: create a vectorized
4329 stmt to replace it, put it in VEC_STMT, and insert it at GSI.
4330 Return FALSE if not a vectorizable STMT, TRUE otherwise.
4332 This function also handles reduction idioms (patterns) that have been
4333 recognized in advance during vect_pattern_recog. In this case, STMT may be
4335 X = pattern_expr (arg0, arg1, ..., X)
4336 and it's STMT_VINFO_RELATED_STMT points to the last stmt in the original
4337 sequence that had been detected and replaced by the pattern-stmt (STMT).
4339 In some cases of reduction patterns, the type of the reduction variable X is
4340 different than the type of the other arguments of STMT.
4341 In such cases, the vectype that is used when transforming STMT into a vector
4342 stmt is different than the vectype that is used to determine the
4343 vectorization factor, because it consists of a different number of elements
4344 than the actual number of elements that are being operated upon in parallel.
4346 For example, consider an accumulation of shorts into an int accumulator.
4347 On some targets it's possible to vectorize this pattern operating on 8
4348 shorts at a time (hence, the vectype for purposes of determining the
4349 vectorization factor should be V8HI); on the other hand, the vectype that
4350 is used to create the vector form is actually V4SI (the type of the result).
4352 Upon entry to this function, STMT_VINFO_VECTYPE records the vectype that
4353 indicates what is the actual level of parallelism (V8HI in the example), so
4354 that the right vectorization factor would be derived. This vectype
4355 corresponds to the type of arguments to the reduction stmt, and should *NOT*
4356 be used to create the vectorized stmt. The right vectype for the vectorized
4357 stmt is obtained from the type of the result X:
4358 get_vectype_for_scalar_type (TREE_TYPE (X))
4360 This means that, contrary to "regular" reductions (or "regular" stmts in
4361 general), the following equation:
4362 STMT_VINFO_VECTYPE == get_vectype_for_scalar_type (TREE_TYPE (X))
4363 does *NOT* necessarily hold for reduction patterns. */
4366 vectorizable_reduction (gimple stmt, gimple_stmt_iterator *gsi,
4367 gimple *vec_stmt, slp_tree slp_node)
4371 tree loop_vec_def0 = NULL_TREE, loop_vec_def1 = NULL_TREE;
4372 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
4373 tree vectype_out = STMT_VINFO_VECTYPE (stmt_info);
4374 tree vectype_in = NULL_TREE;
4375 loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info);
4376 struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
4377 enum tree_code code, orig_code, epilog_reduc_code;
4378 enum machine_mode vec_mode;
4380 optab optab, reduc_optab;
4381 tree new_temp = NULL_TREE;
4384 enum vect_def_type dt;
4385 gimple new_phi = NULL;
4389 stmt_vec_info orig_stmt_info;
4390 tree expr = NULL_TREE;
4394 stmt_vec_info prev_stmt_info, prev_phi_info;
4395 bool single_defuse_cycle = false;
4396 tree reduc_def = NULL_TREE;
4397 gimple new_stmt = NULL;
4400 bool nested_cycle = false, found_nested_cycle_def = false;
4401 gimple reduc_def_stmt = NULL;
4402 /* The default is that the reduction variable is the last in statement. */
4403 int reduc_index = 2;
4404 bool double_reduc = false, dummy;
4406 struct loop * def_stmt_loop, *outer_loop = NULL;
4408 gimple def_arg_stmt;
4409 VEC (tree, heap) *vec_oprnds0 = NULL, *vec_oprnds1 = NULL, *vect_defs = NULL;
4410 VEC (gimple, heap) *phis = NULL;
4412 tree def0, def1, tem, op0, op1 = NULL_TREE;
4414 /* In case of reduction chain we switch to the first stmt in the chain, but
4415 we don't update STMT_INFO, since only the last stmt is marked as reduction
4416 and has reduction properties. */
4417 if (GROUP_FIRST_ELEMENT (vinfo_for_stmt (stmt)))
4418 stmt = GROUP_FIRST_ELEMENT (stmt_info);
4420 if (nested_in_vect_loop_p (loop, stmt))
4424 nested_cycle = true;
4427 /* 1. Is vectorizable reduction? */
4428 /* Not supportable if the reduction variable is used in the loop, unless
4429 it's a reduction chain. */
4430 if (STMT_VINFO_RELEVANT (stmt_info) > vect_used_in_outer
4431 && !GROUP_FIRST_ELEMENT (stmt_info))
4434 /* Reductions that are not used even in an enclosing outer-loop,
4435 are expected to be "live" (used out of the loop). */
4436 if (STMT_VINFO_RELEVANT (stmt_info) == vect_unused_in_scope
4437 && !STMT_VINFO_LIVE_P (stmt_info))
4440 /* Make sure it was already recognized as a reduction computation. */
4441 if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_reduction_def
4442 && STMT_VINFO_DEF_TYPE (stmt_info) != vect_nested_cycle)
4445 /* 2. Has this been recognized as a reduction pattern?
4447 Check if STMT represents a pattern that has been recognized
4448 in earlier analysis stages. For stmts that represent a pattern,
4449 the STMT_VINFO_RELATED_STMT field records the last stmt in
4450 the original sequence that constitutes the pattern. */
4452 orig_stmt = STMT_VINFO_RELATED_STMT (stmt_info);
4455 orig_stmt_info = vinfo_for_stmt (orig_stmt);
4456 gcc_assert (STMT_VINFO_IN_PATTERN_P (orig_stmt_info));
4457 gcc_assert (!STMT_VINFO_IN_PATTERN_P (stmt_info));
4460 /* 3. Check the operands of the operation. The first operands are defined
4461 inside the loop body. The last operand is the reduction variable,
4462 which is defined by the loop-header-phi. */
4464 gcc_assert (is_gimple_assign (stmt));
4467 switch (get_gimple_rhs_class (gimple_assign_rhs_code (stmt)))
4469 case GIMPLE_SINGLE_RHS:
4470 op_type = TREE_OPERAND_LENGTH (gimple_assign_rhs1 (stmt));
4471 if (op_type == ternary_op)
4473 tree rhs = gimple_assign_rhs1 (stmt);
4474 ops[0] = TREE_OPERAND (rhs, 0);
4475 ops[1] = TREE_OPERAND (rhs, 1);
4476 ops[2] = TREE_OPERAND (rhs, 2);
4477 code = TREE_CODE (rhs);
4483 case GIMPLE_BINARY_RHS:
4484 code = gimple_assign_rhs_code (stmt);
4485 op_type = TREE_CODE_LENGTH (code);
4486 gcc_assert (op_type == binary_op);
4487 ops[0] = gimple_assign_rhs1 (stmt);
4488 ops[1] = gimple_assign_rhs2 (stmt);
4491 case GIMPLE_TERNARY_RHS:
4492 code = gimple_assign_rhs_code (stmt);
4493 op_type = TREE_CODE_LENGTH (code);
4494 gcc_assert (op_type == ternary_op);
4495 ops[0] = gimple_assign_rhs1 (stmt);
4496 ops[1] = gimple_assign_rhs2 (stmt);
4497 ops[2] = gimple_assign_rhs3 (stmt);
4500 case GIMPLE_UNARY_RHS:
4507 if (code == COND_EXPR && slp_node)
4510 scalar_dest = gimple_assign_lhs (stmt);
4511 scalar_type = TREE_TYPE (scalar_dest);
4512 if (!POINTER_TYPE_P (scalar_type) && !INTEGRAL_TYPE_P (scalar_type)
4513 && !SCALAR_FLOAT_TYPE_P (scalar_type))
4516 /* Do not try to vectorize bit-precision reductions. */
4517 if ((TYPE_PRECISION (scalar_type)
4518 != GET_MODE_PRECISION (TYPE_MODE (scalar_type))))
4521 /* All uses but the last are expected to be defined in the loop.
4522 The last use is the reduction variable. In case of nested cycle this
4523 assumption is not true: we use reduc_index to record the index of the
4524 reduction variable. */
4525 for (i = 0; i < op_type - 1; i++)
4527 /* The condition of COND_EXPR is checked in vectorizable_condition(). */
4528 if (i == 0 && code == COND_EXPR)
4531 is_simple_use = vect_is_simple_use_1 (ops[i], stmt, loop_vinfo, NULL,
4532 &def_stmt, &def, &dt, &tem);
4535 gcc_assert (is_simple_use);
4537 if (dt != vect_internal_def
4538 && dt != vect_external_def
4539 && dt != vect_constant_def
4540 && dt != vect_induction_def
4541 && !(dt == vect_nested_cycle && nested_cycle))
4544 if (dt == vect_nested_cycle)
4546 found_nested_cycle_def = true;
4547 reduc_def_stmt = def_stmt;
4552 is_simple_use = vect_is_simple_use_1 (ops[i], stmt, loop_vinfo, NULL,
4553 &def_stmt, &def, &dt, &tem);
4556 gcc_assert (is_simple_use);
4557 if (!(dt == vect_reduction_def
4558 || dt == vect_nested_cycle
4559 || ((dt == vect_internal_def || dt == vect_external_def
4560 || dt == vect_constant_def || dt == vect_induction_def)
4561 && nested_cycle && found_nested_cycle_def)))
4563 /* For pattern recognized stmts, orig_stmt might be a reduction,
4564 but some helper statements for the pattern might not, or
4565 might be COND_EXPRs with reduction uses in the condition. */
4566 gcc_assert (orig_stmt);
4569 if (!found_nested_cycle_def)
4570 reduc_def_stmt = def_stmt;
4572 gcc_assert (gimple_code (reduc_def_stmt) == GIMPLE_PHI);
4574 gcc_assert (orig_stmt == vect_is_simple_reduction (loop_vinfo,
4580 gimple tmp = vect_is_simple_reduction (loop_vinfo, reduc_def_stmt,
4581 !nested_cycle, &dummy);
4582 /* We changed STMT to be the first stmt in reduction chain, hence we
4583 check that in this case the first element in the chain is STMT. */
4584 gcc_assert (stmt == tmp
4585 || GROUP_FIRST_ELEMENT (vinfo_for_stmt (tmp)) == stmt);
4588 if (STMT_VINFO_LIVE_P (vinfo_for_stmt (reduc_def_stmt)))
4591 if (slp_node || PURE_SLP_STMT (stmt_info))
4594 ncopies = (LOOP_VINFO_VECT_FACTOR (loop_vinfo)
4595 / TYPE_VECTOR_SUBPARTS (vectype_in));
4597 gcc_assert (ncopies >= 1);
4599 vec_mode = TYPE_MODE (vectype_in);
4601 if (code == COND_EXPR)
4603 if (!vectorizable_condition (stmt, gsi, NULL, ops[reduc_index], 0, NULL))
4605 if (vect_print_dump_info (REPORT_DETAILS))
4606 fprintf (vect_dump, "unsupported condition in reduction");
4613 /* 4. Supportable by target? */
4615 /* 4.1. check support for the operation in the loop */
4616 optab = optab_for_tree_code (code, vectype_in, optab_default);
4619 if (vect_print_dump_info (REPORT_DETAILS))
4620 fprintf (vect_dump, "no optab.");
4625 if (optab_handler (optab, vec_mode) == CODE_FOR_nothing)
4627 if (vect_print_dump_info (REPORT_DETAILS))
4628 fprintf (vect_dump, "op not supported by target.");
4630 if (GET_MODE_SIZE (vec_mode) != UNITS_PER_WORD
4631 || LOOP_VINFO_VECT_FACTOR (loop_vinfo)
4632 < vect_min_worthwhile_factor (code))
4635 if (vect_print_dump_info (REPORT_DETAILS))
4636 fprintf (vect_dump, "proceeding using word mode.");
4639 /* Worthwhile without SIMD support? */
4640 if (!VECTOR_MODE_P (TYPE_MODE (vectype_in))
4641 && LOOP_VINFO_VECT_FACTOR (loop_vinfo)
4642 < vect_min_worthwhile_factor (code))
4644 if (vect_print_dump_info (REPORT_DETAILS))
4645 fprintf (vect_dump, "not worthwhile without SIMD support.");
4651 /* 4.2. Check support for the epilog operation.
4653 If STMT represents a reduction pattern, then the type of the
4654 reduction variable may be different than the type of the rest
4655 of the arguments. For example, consider the case of accumulation
4656 of shorts into an int accumulator; The original code:
4657 S1: int_a = (int) short_a;
4658 orig_stmt-> S2: int_acc = plus <int_a ,int_acc>;
4661 STMT: int_acc = widen_sum <short_a, int_acc>
4664 1. The tree-code that is used to create the vector operation in the
4665 epilog code (that reduces the partial results) is not the
4666 tree-code of STMT, but is rather the tree-code of the original
4667 stmt from the pattern that STMT is replacing. I.e, in the example
4668 above we want to use 'widen_sum' in the loop, but 'plus' in the
4670 2. The type (mode) we use to check available target support
4671 for the vector operation to be created in the *epilog*, is
4672 determined by the type of the reduction variable (in the example
4673 above we'd check this: optab_handler (plus_optab, vect_int_mode])).
4674 However the type (mode) we use to check available target support
4675 for the vector operation to be created *inside the loop*, is
4676 determined by the type of the other arguments to STMT (in the
4677 example we'd check this: optab_handler (widen_sum_optab,
4680 This is contrary to "regular" reductions, in which the types of all
4681 the arguments are the same as the type of the reduction variable.
4682 For "regular" reductions we can therefore use the same vector type
4683 (and also the same tree-code) when generating the epilog code and
4684 when generating the code inside the loop. */
4688 /* This is a reduction pattern: get the vectype from the type of the
4689 reduction variable, and get the tree-code from orig_stmt. */
4690 orig_code = gimple_assign_rhs_code (orig_stmt);
4691 gcc_assert (vectype_out);
4692 vec_mode = TYPE_MODE (vectype_out);
4696 /* Regular reduction: use the same vectype and tree-code as used for
4697 the vector code inside the loop can be used for the epilog code. */
4703 def_bb = gimple_bb (reduc_def_stmt);
4704 def_stmt_loop = def_bb->loop_father;
4705 def_arg = PHI_ARG_DEF_FROM_EDGE (reduc_def_stmt,
4706 loop_preheader_edge (def_stmt_loop));
4707 if (TREE_CODE (def_arg) == SSA_NAME
4708 && (def_arg_stmt = SSA_NAME_DEF_STMT (def_arg))
4709 && gimple_code (def_arg_stmt) == GIMPLE_PHI
4710 && flow_bb_inside_loop_p (outer_loop, gimple_bb (def_arg_stmt))
4711 && vinfo_for_stmt (def_arg_stmt)
4712 && STMT_VINFO_DEF_TYPE (vinfo_for_stmt (def_arg_stmt))
4713 == vect_double_reduction_def)
4714 double_reduc = true;
4717 epilog_reduc_code = ERROR_MARK;
4718 if (reduction_code_for_scalar_code (orig_code, &epilog_reduc_code))
4720 reduc_optab = optab_for_tree_code (epilog_reduc_code, vectype_out,
4724 if (vect_print_dump_info (REPORT_DETAILS))
4725 fprintf (vect_dump, "no optab for reduction.");
4727 epilog_reduc_code = ERROR_MARK;
4731 && optab_handler (reduc_optab, vec_mode) == CODE_FOR_nothing)
4733 if (vect_print_dump_info (REPORT_DETAILS))
4734 fprintf (vect_dump, "reduc op not supported by target.");
4736 epilog_reduc_code = ERROR_MARK;
4741 if (!nested_cycle || double_reduc)
4743 if (vect_print_dump_info (REPORT_DETAILS))
4744 fprintf (vect_dump, "no reduc code for scalar code.");
4750 if (double_reduc && ncopies > 1)
4752 if (vect_print_dump_info (REPORT_DETAILS))
4753 fprintf (vect_dump, "multiple types in double reduction");
4758 /* In case of widenning multiplication by a constant, we update the type
4759 of the constant to be the type of the other operand. We check that the
4760 constant fits the type in the pattern recognition pass. */
4761 if (code == DOT_PROD_EXPR
4762 && !types_compatible_p (TREE_TYPE (ops[0]), TREE_TYPE (ops[1])))
4764 if (TREE_CODE (ops[0]) == INTEGER_CST)
4765 ops[0] = fold_convert (TREE_TYPE (ops[1]), ops[0]);
4766 else if (TREE_CODE (ops[1]) == INTEGER_CST)
4767 ops[1] = fold_convert (TREE_TYPE (ops[0]), ops[1]);
4770 if (vect_print_dump_info (REPORT_DETAILS))
4771 fprintf (vect_dump, "invalid types in dot-prod");
4777 if (!vec_stmt) /* transformation not required. */
4779 if (!vect_model_reduction_cost (stmt_info, epilog_reduc_code, ncopies))
4781 STMT_VINFO_TYPE (stmt_info) = reduc_vec_info_type;
4787 if (vect_print_dump_info (REPORT_DETAILS))
4788 fprintf (vect_dump, "transform reduction.");
4790 /* FORNOW: Multiple types are not supported for condition. */
4791 if (code == COND_EXPR)
4792 gcc_assert (ncopies == 1);
4794 /* Create the destination vector */
4795 vec_dest = vect_create_destination_var (scalar_dest, vectype_out);
4797 /* In case the vectorization factor (VF) is bigger than the number
4798 of elements that we can fit in a vectype (nunits), we have to generate
4799 more than one vector stmt - i.e - we need to "unroll" the
4800 vector stmt by a factor VF/nunits. For more details see documentation
4801 in vectorizable_operation. */
4803 /* If the reduction is used in an outer loop we need to generate
4804 VF intermediate results, like so (e.g. for ncopies=2):
4809 (i.e. we generate VF results in 2 registers).
4810 In this case we have a separate def-use cycle for each copy, and therefore
4811 for each copy we get the vector def for the reduction variable from the
4812 respective phi node created for this copy.
4814 Otherwise (the reduction is unused in the loop nest), we can combine
4815 together intermediate results, like so (e.g. for ncopies=2):
4819 (i.e. we generate VF/2 results in a single register).
4820 In this case for each copy we get the vector def for the reduction variable
4821 from the vectorized reduction operation generated in the previous iteration.
4824 if (STMT_VINFO_RELEVANT (stmt_info) == vect_unused_in_scope)
4826 single_defuse_cycle = true;
4830 epilog_copies = ncopies;
4832 prev_stmt_info = NULL;
4833 prev_phi_info = NULL;
4836 vec_num = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node);
4837 gcc_assert (TYPE_VECTOR_SUBPARTS (vectype_out)
4838 == TYPE_VECTOR_SUBPARTS (vectype_in));
4843 vec_oprnds0 = VEC_alloc (tree, heap, 1);
4844 if (op_type == ternary_op)
4845 vec_oprnds1 = VEC_alloc (tree, heap, 1);
4848 phis = VEC_alloc (gimple, heap, vec_num);
4849 vect_defs = VEC_alloc (tree, heap, vec_num);
4851 VEC_quick_push (tree, vect_defs, NULL_TREE);
4853 for (j = 0; j < ncopies; j++)
4855 if (j == 0 || !single_defuse_cycle)
4857 for (i = 0; i < vec_num; i++)
4859 /* Create the reduction-phi that defines the reduction
4861 new_phi = create_phi_node (vec_dest, loop->header);
4862 set_vinfo_for_stmt (new_phi,
4863 new_stmt_vec_info (new_phi, loop_vinfo,
4865 if (j == 0 || slp_node)
4866 VEC_quick_push (gimple, phis, new_phi);
4870 if (code == COND_EXPR)
4872 gcc_assert (!slp_node);
4873 vectorizable_condition (stmt, gsi, vec_stmt,
4874 PHI_RESULT (VEC_index (gimple, phis, 0)),
4876 /* Multiple types are not supported for condition. */
4883 op0 = ops[!reduc_index];
4884 if (op_type == ternary_op)
4886 if (reduc_index == 0)
4893 vect_get_vec_defs (op0, op1, stmt, &vec_oprnds0, &vec_oprnds1,
4897 loop_vec_def0 = vect_get_vec_def_for_operand (ops[!reduc_index],
4899 VEC_quick_push (tree, vec_oprnds0, loop_vec_def0);
4900 if (op_type == ternary_op)
4902 loop_vec_def1 = vect_get_vec_def_for_operand (op1, stmt,
4904 VEC_quick_push (tree, vec_oprnds1, loop_vec_def1);
4912 enum vect_def_type dt;
4916 vect_is_simple_use (ops[!reduc_index], stmt, loop_vinfo, NULL,
4917 &dummy_stmt, &dummy, &dt);
4918 loop_vec_def0 = vect_get_vec_def_for_stmt_copy (dt,
4920 VEC_replace (tree, vec_oprnds0, 0, loop_vec_def0);
4921 if (op_type == ternary_op)
4923 vect_is_simple_use (op1, stmt, loop_vinfo, NULL, &dummy_stmt,
4925 loop_vec_def1 = vect_get_vec_def_for_stmt_copy (dt,
4927 VEC_replace (tree, vec_oprnds1, 0, loop_vec_def1);
4931 if (single_defuse_cycle)
4932 reduc_def = gimple_assign_lhs (new_stmt);
4934 STMT_VINFO_RELATED_STMT (prev_phi_info) = new_phi;
4937 FOR_EACH_VEC_ELT (tree, vec_oprnds0, i, def0)
4940 reduc_def = PHI_RESULT (VEC_index (gimple, phis, i));
4943 if (!single_defuse_cycle || j == 0)
4944 reduc_def = PHI_RESULT (new_phi);
4947 def1 = ((op_type == ternary_op)
4948 ? VEC_index (tree, vec_oprnds1, i) : NULL);
4949 if (op_type == binary_op)
4951 if (reduc_index == 0)
4952 expr = build2 (code, vectype_out, reduc_def, def0);
4954 expr = build2 (code, vectype_out, def0, reduc_def);
4958 if (reduc_index == 0)
4959 expr = build3 (code, vectype_out, reduc_def, def0, def1);
4962 if (reduc_index == 1)
4963 expr = build3 (code, vectype_out, def0, reduc_def, def1);
4965 expr = build3 (code, vectype_out, def0, def1, reduc_def);
4969 new_stmt = gimple_build_assign (vec_dest, expr);
4970 new_temp = make_ssa_name (vec_dest, new_stmt);
4971 gimple_assign_set_lhs (new_stmt, new_temp);
4972 vect_finish_stmt_generation (stmt, new_stmt, gsi);
4976 VEC_quick_push (gimple, SLP_TREE_VEC_STMTS (slp_node), new_stmt);
4977 VEC_quick_push (tree, vect_defs, new_temp);
4980 VEC_replace (tree, vect_defs, 0, new_temp);
4987 STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt;
4989 STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt;
4991 prev_stmt_info = vinfo_for_stmt (new_stmt);
4992 prev_phi_info = vinfo_for_stmt (new_phi);
4995 /* Finalize the reduction-phi (set its arguments) and create the
4996 epilog reduction code. */
4997 if ((!single_defuse_cycle || code == COND_EXPR) && !slp_node)
4999 new_temp = gimple_assign_lhs (*vec_stmt);
5000 VEC_replace (tree, vect_defs, 0, new_temp);
5003 vect_create_epilog_for_reduction (vect_defs, stmt, epilog_copies,
5004 epilog_reduc_code, phis, reduc_index,
5005 double_reduc, slp_node);
5007 VEC_free (gimple, heap, phis);
5008 VEC_free (tree, heap, vec_oprnds0);
5010 VEC_free (tree, heap, vec_oprnds1);
5015 /* Function vect_min_worthwhile_factor.
5017 For a loop where we could vectorize the operation indicated by CODE,
5018 return the minimum vectorization factor that makes it worthwhile
5019 to use generic vectors. */
5021 vect_min_worthwhile_factor (enum tree_code code)
5042 /* Function vectorizable_induction
5044 Check if PHI performs an induction computation that can be vectorized.
5045 If VEC_STMT is also passed, vectorize the induction PHI: create a vectorized
5046 phi to replace it, put it in VEC_STMT, and add it to the same basic block.
5047 Return FALSE if not a vectorizable STMT, TRUE otherwise. */
5050 vectorizable_induction (gimple phi, gimple_stmt_iterator *gsi ATTRIBUTE_UNUSED,
5053 stmt_vec_info stmt_info = vinfo_for_stmt (phi);
5054 tree vectype = STMT_VINFO_VECTYPE (stmt_info);
5055 loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info);
5056 struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
5057 int nunits = TYPE_VECTOR_SUBPARTS (vectype);
5058 int ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits;
5061 gcc_assert (ncopies >= 1);
5062 /* FORNOW. These restrictions should be relaxed. */
5063 if (nested_in_vect_loop_p (loop, phi))
5065 imm_use_iterator imm_iter;
5066 use_operand_p use_p;
5073 if (vect_print_dump_info (REPORT_DETAILS))
5074 fprintf (vect_dump, "multiple types in nested loop.");
5079 latch_e = loop_latch_edge (loop->inner);
5080 loop_arg = PHI_ARG_DEF_FROM_EDGE (phi, latch_e);
5081 FOR_EACH_IMM_USE_FAST (use_p, imm_iter, loop_arg)
5083 if (!flow_bb_inside_loop_p (loop->inner,
5084 gimple_bb (USE_STMT (use_p))))
5086 exit_phi = USE_STMT (use_p);
5092 stmt_vec_info exit_phi_vinfo = vinfo_for_stmt (exit_phi);
5093 if (!(STMT_VINFO_RELEVANT_P (exit_phi_vinfo)
5094 && !STMT_VINFO_LIVE_P (exit_phi_vinfo)))
5096 if (vect_print_dump_info (REPORT_DETAILS))
5097 fprintf (vect_dump, "inner-loop induction only used outside "
5098 "of the outer vectorized loop.");
5104 if (!STMT_VINFO_RELEVANT_P (stmt_info))
5107 /* FORNOW: SLP not supported. */
5108 if (STMT_SLP_TYPE (stmt_info))
5111 gcc_assert (STMT_VINFO_DEF_TYPE (stmt_info) == vect_induction_def);
5113 if (gimple_code (phi) != GIMPLE_PHI)
5116 if (!vec_stmt) /* transformation not required. */
5118 STMT_VINFO_TYPE (stmt_info) = induc_vec_info_type;
5119 if (vect_print_dump_info (REPORT_DETAILS))
5120 fprintf (vect_dump, "=== vectorizable_induction ===");
5121 vect_model_induction_cost (stmt_info, ncopies);
5127 if (vect_print_dump_info (REPORT_DETAILS))
5128 fprintf (vect_dump, "transform induction phi.");
5130 vec_def = get_initial_def_for_induction (phi);
5131 *vec_stmt = SSA_NAME_DEF_STMT (vec_def);
5135 /* Function vectorizable_live_operation.
5137 STMT computes a value that is used outside the loop. Check if
5138 it can be supported. */
5141 vectorizable_live_operation (gimple stmt,
5142 gimple_stmt_iterator *gsi ATTRIBUTE_UNUSED,
5143 gimple *vec_stmt ATTRIBUTE_UNUSED)
5145 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
5146 loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info);
5147 struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
5153 enum vect_def_type dt;
5154 enum tree_code code;
5155 enum gimple_rhs_class rhs_class;
5157 gcc_assert (STMT_VINFO_LIVE_P (stmt_info));
5159 if (STMT_VINFO_DEF_TYPE (stmt_info) == vect_reduction_def)
5162 if (!is_gimple_assign (stmt))
5165 if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME)
5168 /* FORNOW. CHECKME. */
5169 if (nested_in_vect_loop_p (loop, stmt))
5172 code = gimple_assign_rhs_code (stmt);
5173 op_type = TREE_CODE_LENGTH (code);
5174 rhs_class = get_gimple_rhs_class (code);
5175 gcc_assert (rhs_class != GIMPLE_UNARY_RHS || op_type == unary_op);
5176 gcc_assert (rhs_class != GIMPLE_BINARY_RHS || op_type == binary_op);
5178 /* FORNOW: support only if all uses are invariant. This means
5179 that the scalar operations can remain in place, unvectorized.
5180 The original last scalar value that they compute will be used. */
5182 for (i = 0; i < op_type; i++)
5184 if (rhs_class == GIMPLE_SINGLE_RHS)
5185 op = TREE_OPERAND (gimple_op (stmt, 1), i);
5187 op = gimple_op (stmt, i + 1);
5189 && !vect_is_simple_use (op, stmt, loop_vinfo, NULL, &def_stmt, &def,
5192 if (vect_print_dump_info (REPORT_DETAILS))
5193 fprintf (vect_dump, "use not simple.");
5197 if (dt != vect_external_def && dt != vect_constant_def)
5201 /* No transformation is required for the cases we currently support. */
5205 /* Kill any debug uses outside LOOP of SSA names defined in STMT. */
5208 vect_loop_kill_debug_uses (struct loop *loop, gimple stmt)
5210 ssa_op_iter op_iter;
5211 imm_use_iterator imm_iter;
5212 def_operand_p def_p;
5215 FOR_EACH_PHI_OR_STMT_DEF (def_p, stmt, op_iter, SSA_OP_DEF)
5217 FOR_EACH_IMM_USE_STMT (ustmt, imm_iter, DEF_FROM_PTR (def_p))
5221 if (!is_gimple_debug (ustmt))
5224 bb = gimple_bb (ustmt);
5226 if (!flow_bb_inside_loop_p (loop, bb))
5228 if (gimple_debug_bind_p (ustmt))
5230 if (vect_print_dump_info (REPORT_DETAILS))
5231 fprintf (vect_dump, "killing debug use");
5233 gimple_debug_bind_reset_value (ustmt);
5234 update_stmt (ustmt);
5243 /* Function vect_transform_loop.
5245 The analysis phase has determined that the loop is vectorizable.
5246 Vectorize the loop - created vectorized stmts to replace the scalar
5247 stmts in the loop, and update the loop exit condition. */
5250 vect_transform_loop (loop_vec_info loop_vinfo)
5252 struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
5253 basic_block *bbs = LOOP_VINFO_BBS (loop_vinfo);
5254 int nbbs = loop->num_nodes;
5255 gimple_stmt_iterator si;
5258 int vectorization_factor = LOOP_VINFO_VECT_FACTOR (loop_vinfo);
5260 bool slp_scheduled = false;
5261 unsigned int nunits;
5262 tree cond_expr = NULL_TREE;
5263 gimple_seq cond_expr_stmt_list = NULL;
5264 bool do_peeling_for_loop_bound;
5265 gimple stmt, pattern_stmt;
5266 gimple_seq pattern_def_seq = NULL;
5267 gimple_stmt_iterator pattern_def_si = gsi_start (NULL);
5268 bool transform_pattern_stmt = false;
5270 if (vect_print_dump_info (REPORT_DETAILS))
5271 fprintf (vect_dump, "=== vec_transform_loop ===");
5273 /* Peel the loop if there are data refs with unknown alignment.
5274 Only one data ref with unknown store is allowed. */
5276 if (LOOP_PEELING_FOR_ALIGNMENT (loop_vinfo))
5277 vect_do_peeling_for_alignment (loop_vinfo);
5279 do_peeling_for_loop_bound
5280 = (!LOOP_VINFO_NITERS_KNOWN_P (loop_vinfo)
5281 || (LOOP_VINFO_NITERS_KNOWN_P (loop_vinfo)
5282 && LOOP_VINFO_INT_NITERS (loop_vinfo) % vectorization_factor != 0)
5283 || LOOP_VINFO_PEELING_FOR_GAPS (loop_vinfo));
5285 if (LOOP_REQUIRES_VERSIONING_FOR_ALIGNMENT (loop_vinfo)
5286 || LOOP_REQUIRES_VERSIONING_FOR_ALIAS (loop_vinfo))
5287 vect_loop_versioning (loop_vinfo,
5288 !do_peeling_for_loop_bound,
5289 &cond_expr, &cond_expr_stmt_list);
5291 /* If the loop has a symbolic number of iterations 'n' (i.e. it's not a
5292 compile time constant), or it is a constant that doesn't divide by the
5293 vectorization factor, then an epilog loop needs to be created.
5294 We therefore duplicate the loop: the original loop will be vectorized,
5295 and will compute the first (n/VF) iterations. The second copy of the loop
5296 will remain scalar and will compute the remaining (n%VF) iterations.
5297 (VF is the vectorization factor). */
5299 if (do_peeling_for_loop_bound)
5300 vect_do_peeling_for_loop_bound (loop_vinfo, &ratio,
5301 cond_expr, cond_expr_stmt_list);
5303 ratio = build_int_cst (TREE_TYPE (LOOP_VINFO_NITERS (loop_vinfo)),
5304 LOOP_VINFO_INT_NITERS (loop_vinfo) / vectorization_factor);
5306 /* 1) Make sure the loop header has exactly two entries
5307 2) Make sure we have a preheader basic block. */
5309 gcc_assert (EDGE_COUNT (loop->header->preds) == 2);
5311 split_edge (loop_preheader_edge (loop));
5313 /* FORNOW: the vectorizer supports only loops which body consist
5314 of one basic block (header + empty latch). When the vectorizer will
5315 support more involved loop forms, the order by which the BBs are
5316 traversed need to be reconsidered. */
5318 for (i = 0; i < nbbs; i++)
5320 basic_block bb = bbs[i];
5321 stmt_vec_info stmt_info;
5324 for (si = gsi_start_phis (bb); !gsi_end_p (si); gsi_next (&si))
5326 phi = gsi_stmt (si);
5327 if (vect_print_dump_info (REPORT_DETAILS))
5329 fprintf (vect_dump, "------>vectorizing phi: ");
5330 print_gimple_stmt (vect_dump, phi, 0, TDF_SLIM);
5332 stmt_info = vinfo_for_stmt (phi);
5336 if (MAY_HAVE_DEBUG_STMTS && !STMT_VINFO_LIVE_P (stmt_info))
5337 vect_loop_kill_debug_uses (loop, phi);
5339 if (!STMT_VINFO_RELEVANT_P (stmt_info)
5340 && !STMT_VINFO_LIVE_P (stmt_info))
5343 if ((TYPE_VECTOR_SUBPARTS (STMT_VINFO_VECTYPE (stmt_info))
5344 != (unsigned HOST_WIDE_INT) vectorization_factor)
5345 && vect_print_dump_info (REPORT_DETAILS))
5346 fprintf (vect_dump, "multiple-types.");
5348 if (STMT_VINFO_DEF_TYPE (stmt_info) == vect_induction_def)
5350 if (vect_print_dump_info (REPORT_DETAILS))
5351 fprintf (vect_dump, "transform phi.");
5352 vect_transform_stmt (phi, NULL, NULL, NULL, NULL);
5356 pattern_stmt = NULL;
5357 for (si = gsi_start_bb (bb); !gsi_end_p (si) || transform_pattern_stmt;)
5361 if (transform_pattern_stmt)
5362 stmt = pattern_stmt;
5364 stmt = gsi_stmt (si);
5366 if (vect_print_dump_info (REPORT_DETAILS))
5368 fprintf (vect_dump, "------>vectorizing statement: ");
5369 print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM);
5372 stmt_info = vinfo_for_stmt (stmt);
5374 /* vector stmts created in the outer-loop during vectorization of
5375 stmts in an inner-loop may not have a stmt_info, and do not
5376 need to be vectorized. */
5383 if (MAY_HAVE_DEBUG_STMTS && !STMT_VINFO_LIVE_P (stmt_info))
5384 vect_loop_kill_debug_uses (loop, stmt);
5386 if (!STMT_VINFO_RELEVANT_P (stmt_info)
5387 && !STMT_VINFO_LIVE_P (stmt_info))
5389 if (STMT_VINFO_IN_PATTERN_P (stmt_info)
5390 && (pattern_stmt = STMT_VINFO_RELATED_STMT (stmt_info))
5391 && (STMT_VINFO_RELEVANT_P (vinfo_for_stmt (pattern_stmt))
5392 || STMT_VINFO_LIVE_P (vinfo_for_stmt (pattern_stmt))))
5394 stmt = pattern_stmt;
5395 stmt_info = vinfo_for_stmt (stmt);
5403 else if (STMT_VINFO_IN_PATTERN_P (stmt_info)
5404 && (pattern_stmt = STMT_VINFO_RELATED_STMT (stmt_info))
5405 && (STMT_VINFO_RELEVANT_P (vinfo_for_stmt (pattern_stmt))
5406 || STMT_VINFO_LIVE_P (vinfo_for_stmt (pattern_stmt))))
5407 transform_pattern_stmt = true;
5409 /* If pattern statement has def stmts, vectorize them too. */
5410 if (is_pattern_stmt_p (stmt_info))
5412 if (pattern_def_seq == NULL)
5414 pattern_def_seq = STMT_VINFO_PATTERN_DEF_SEQ (stmt_info);
5415 pattern_def_si = gsi_start (pattern_def_seq);
5417 else if (!gsi_end_p (pattern_def_si))
5418 gsi_next (&pattern_def_si);
5419 if (pattern_def_seq != NULL)
5421 gimple pattern_def_stmt = NULL;
5422 stmt_vec_info pattern_def_stmt_info = NULL;
5424 while (!gsi_end_p (pattern_def_si))
5426 pattern_def_stmt = gsi_stmt (pattern_def_si);
5427 pattern_def_stmt_info
5428 = vinfo_for_stmt (pattern_def_stmt);
5429 if (STMT_VINFO_RELEVANT_P (pattern_def_stmt_info)
5430 || STMT_VINFO_LIVE_P (pattern_def_stmt_info))
5432 gsi_next (&pattern_def_si);
5435 if (!gsi_end_p (pattern_def_si))
5437 if (vect_print_dump_info (REPORT_DETAILS))
5439 fprintf (vect_dump, "==> vectorizing pattern def"
5441 print_gimple_stmt (vect_dump, pattern_def_stmt, 0,
5445 stmt = pattern_def_stmt;
5446 stmt_info = pattern_def_stmt_info;
5450 pattern_def_si = gsi_start (NULL);
5451 transform_pattern_stmt = false;
5455 transform_pattern_stmt = false;
5458 gcc_assert (STMT_VINFO_VECTYPE (stmt_info));
5459 nunits = (unsigned int) TYPE_VECTOR_SUBPARTS (
5460 STMT_VINFO_VECTYPE (stmt_info));
5461 if (!STMT_SLP_TYPE (stmt_info)
5462 && nunits != (unsigned int) vectorization_factor
5463 && vect_print_dump_info (REPORT_DETAILS))
5464 /* For SLP VF is set according to unrolling factor, and not to
5465 vector size, hence for SLP this print is not valid. */
5466 fprintf (vect_dump, "multiple-types.");
5468 /* SLP. Schedule all the SLP instances when the first SLP stmt is
5470 if (STMT_SLP_TYPE (stmt_info))
5474 slp_scheduled = true;
5476 if (vect_print_dump_info (REPORT_DETAILS))
5477 fprintf (vect_dump, "=== scheduling SLP instances ===");
5479 vect_schedule_slp (loop_vinfo, NULL);
5482 /* Hybrid SLP stmts must be vectorized in addition to SLP. */
5483 if (!vinfo_for_stmt (stmt) || PURE_SLP_STMT (stmt_info))
5485 if (!transform_pattern_stmt && gsi_end_p (pattern_def_si))
5487 pattern_def_seq = NULL;
5494 /* -------- vectorize statement ------------ */
5495 if (vect_print_dump_info (REPORT_DETAILS))
5496 fprintf (vect_dump, "transform statement.");
5498 strided_store = false;
5499 is_store = vect_transform_stmt (stmt, &si, &strided_store, NULL, NULL);
5502 if (STMT_VINFO_STRIDED_ACCESS (stmt_info))
5504 /* Interleaving. If IS_STORE is TRUE, the vectorization of the
5505 interleaving chain was completed - free all the stores in
5508 vect_remove_stores (GROUP_FIRST_ELEMENT (stmt_info));
5513 /* Free the attached stmt_vec_info and remove the stmt. */
5514 free_stmt_vec_info (gsi_stmt (si));
5515 gsi_remove (&si, true);
5520 if (!transform_pattern_stmt && gsi_end_p (pattern_def_si))
5522 pattern_def_seq = NULL;
5528 slpeel_make_loop_iterate_ntimes (loop, ratio);
5530 /* The memory tags and pointers in vectorized statements need to
5531 have their SSA forms updated. FIXME, why can't this be delayed
5532 until all the loops have been transformed? */
5533 update_ssa (TODO_update_ssa);
5535 if (vect_print_dump_info (REPORT_VECTORIZED_LOCATIONS))
5536 fprintf (vect_dump, "LOOP VECTORIZED.");
5537 if (loop->inner && vect_print_dump_info (REPORT_VECTORIZED_LOCATIONS))
5538 fprintf (vect_dump, "OUTER LOOP VECTORIZED.");