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);
568 STRIP_NOPS (access_fn);
569 if (access_fn && vect_print_dump_info (REPORT_DETAILS))
571 fprintf (vect_dump, "Access function of PHI: ");
572 print_generic_expr (vect_dump, access_fn, TDF_SLIM);
576 || !vect_is_simple_iv_evolution (loop->num, access_fn, &dumy, &dumy))
578 VEC_safe_push (gimple, heap, worklist, phi);
582 if (vect_print_dump_info (REPORT_DETAILS))
583 fprintf (vect_dump, "Detected induction.");
584 STMT_VINFO_DEF_TYPE (stmt_vinfo) = vect_induction_def;
588 /* Second - identify all reductions and nested cycles. */
589 while (VEC_length (gimple, worklist) > 0)
591 gimple phi = VEC_pop (gimple, worklist);
592 tree def = PHI_RESULT (phi);
593 stmt_vec_info stmt_vinfo = vinfo_for_stmt (phi);
597 if (vect_print_dump_info (REPORT_DETAILS))
599 fprintf (vect_dump, "Analyze phi: ");
600 print_gimple_stmt (vect_dump, phi, 0, TDF_SLIM);
603 gcc_assert (is_gimple_reg (SSA_NAME_VAR (def)));
604 gcc_assert (STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_unknown_def_type);
606 nested_cycle = (loop != LOOP_VINFO_LOOP (loop_vinfo));
607 reduc_stmt = vect_force_simple_reduction (loop_vinfo, phi, !nested_cycle,
613 if (vect_print_dump_info (REPORT_DETAILS))
614 fprintf (vect_dump, "Detected double reduction.");
616 STMT_VINFO_DEF_TYPE (stmt_vinfo) = vect_double_reduction_def;
617 STMT_VINFO_DEF_TYPE (vinfo_for_stmt (reduc_stmt)) =
618 vect_double_reduction_def;
624 if (vect_print_dump_info (REPORT_DETAILS))
625 fprintf (vect_dump, "Detected vectorizable nested cycle.");
627 STMT_VINFO_DEF_TYPE (stmt_vinfo) = vect_nested_cycle;
628 STMT_VINFO_DEF_TYPE (vinfo_for_stmt (reduc_stmt)) =
633 if (vect_print_dump_info (REPORT_DETAILS))
634 fprintf (vect_dump, "Detected reduction.");
636 STMT_VINFO_DEF_TYPE (stmt_vinfo) = vect_reduction_def;
637 STMT_VINFO_DEF_TYPE (vinfo_for_stmt (reduc_stmt)) =
639 /* Store the reduction cycles for possible vectorization in
641 VEC_safe_push (gimple, heap,
642 LOOP_VINFO_REDUCTIONS (loop_vinfo),
648 if (vect_print_dump_info (REPORT_DETAILS))
649 fprintf (vect_dump, "Unknown def-use cycle pattern.");
652 VEC_free (gimple, heap, worklist);
656 /* Function vect_analyze_scalar_cycles.
658 Examine the cross iteration def-use cycles of scalar variables, by
659 analyzing the loop-header PHIs of scalar variables. Classify each
660 cycle as one of the following: invariant, induction, reduction, unknown.
661 We do that for the loop represented by LOOP_VINFO, and also to its
662 inner-loop, if exists.
663 Examples for scalar cycles:
678 vect_analyze_scalar_cycles (loop_vec_info loop_vinfo)
680 struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
682 vect_analyze_scalar_cycles_1 (loop_vinfo, loop);
684 /* When vectorizing an outer-loop, the inner-loop is executed sequentially.
685 Reductions in such inner-loop therefore have different properties than
686 the reductions in the nest that gets vectorized:
687 1. When vectorized, they are executed in the same order as in the original
688 scalar loop, so we can't change the order of computation when
690 2. FIXME: Inner-loop reductions can be used in the inner-loop, so the
691 current checks are too strict. */
694 vect_analyze_scalar_cycles_1 (loop_vinfo, loop->inner);
697 /* Function vect_get_loop_niters.
699 Determine how many iterations the loop is executed.
700 If an expression that represents the number of iterations
701 can be constructed, place it in NUMBER_OF_ITERATIONS.
702 Return the loop exit condition. */
705 vect_get_loop_niters (struct loop *loop, tree *number_of_iterations)
709 if (vect_print_dump_info (REPORT_DETAILS))
710 fprintf (vect_dump, "=== get_loop_niters ===");
712 niters = number_of_exit_cond_executions (loop);
714 if (niters != NULL_TREE
715 && niters != chrec_dont_know)
717 *number_of_iterations = niters;
719 if (vect_print_dump_info (REPORT_DETAILS))
721 fprintf (vect_dump, "==> get_loop_niters:" );
722 print_generic_expr (vect_dump, *number_of_iterations, TDF_SLIM);
726 return get_loop_exit_condition (loop);
730 /* Function bb_in_loop_p
732 Used as predicate for dfs order traversal of the loop bbs. */
735 bb_in_loop_p (const_basic_block bb, const void *data)
737 const struct loop *const loop = (const struct loop *)data;
738 if (flow_bb_inside_loop_p (loop, bb))
744 /* Function new_loop_vec_info.
746 Create and initialize a new loop_vec_info struct for LOOP, as well as
747 stmt_vec_info structs for all the stmts in LOOP. */
750 new_loop_vec_info (struct loop *loop)
754 gimple_stmt_iterator si;
755 unsigned int i, nbbs;
757 res = (loop_vec_info) xcalloc (1, sizeof (struct _loop_vec_info));
758 LOOP_VINFO_LOOP (res) = loop;
760 bbs = get_loop_body (loop);
762 /* Create/Update stmt_info for all stmts in the loop. */
763 for (i = 0; i < loop->num_nodes; i++)
765 basic_block bb = bbs[i];
767 /* BBs in a nested inner-loop will have been already processed (because
768 we will have called vect_analyze_loop_form for any nested inner-loop).
769 Therefore, for stmts in an inner-loop we just want to update the
770 STMT_VINFO_LOOP_VINFO field of their stmt_info to point to the new
771 loop_info of the outer-loop we are currently considering to vectorize
772 (instead of the loop_info of the inner-loop).
773 For stmts in other BBs we need to create a stmt_info from scratch. */
774 if (bb->loop_father != loop)
777 gcc_assert (loop->inner && bb->loop_father == loop->inner);
778 for (si = gsi_start_phis (bb); !gsi_end_p (si); gsi_next (&si))
780 gimple phi = gsi_stmt (si);
781 stmt_vec_info stmt_info = vinfo_for_stmt (phi);
782 loop_vec_info inner_loop_vinfo =
783 STMT_VINFO_LOOP_VINFO (stmt_info);
784 gcc_assert (loop->inner == LOOP_VINFO_LOOP (inner_loop_vinfo));
785 STMT_VINFO_LOOP_VINFO (stmt_info) = res;
787 for (si = gsi_start_bb (bb); !gsi_end_p (si); gsi_next (&si))
789 gimple stmt = gsi_stmt (si);
790 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
791 loop_vec_info inner_loop_vinfo =
792 STMT_VINFO_LOOP_VINFO (stmt_info);
793 gcc_assert (loop->inner == LOOP_VINFO_LOOP (inner_loop_vinfo));
794 STMT_VINFO_LOOP_VINFO (stmt_info) = res;
799 /* bb in current nest. */
800 for (si = gsi_start_phis (bb); !gsi_end_p (si); gsi_next (&si))
802 gimple phi = gsi_stmt (si);
803 gimple_set_uid (phi, 0);
804 set_vinfo_for_stmt (phi, new_stmt_vec_info (phi, res, NULL));
807 for (si = gsi_start_bb (bb); !gsi_end_p (si); gsi_next (&si))
809 gimple stmt = gsi_stmt (si);
810 gimple_set_uid (stmt, 0);
811 set_vinfo_for_stmt (stmt, new_stmt_vec_info (stmt, res, NULL));
816 /* CHECKME: We want to visit all BBs before their successors (except for
817 latch blocks, for which this assertion wouldn't hold). In the simple
818 case of the loop forms we allow, a dfs order of the BBs would the same
819 as reversed postorder traversal, so we are safe. */
822 bbs = XCNEWVEC (basic_block, loop->num_nodes);
823 nbbs = dfs_enumerate_from (loop->header, 0, bb_in_loop_p,
824 bbs, loop->num_nodes, loop);
825 gcc_assert (nbbs == loop->num_nodes);
827 LOOP_VINFO_BBS (res) = bbs;
828 LOOP_VINFO_NITERS (res) = NULL;
829 LOOP_VINFO_NITERS_UNCHANGED (res) = NULL;
830 LOOP_VINFO_COST_MODEL_MIN_ITERS (res) = 0;
831 LOOP_VINFO_VECTORIZABLE_P (res) = 0;
832 LOOP_PEELING_FOR_ALIGNMENT (res) = 0;
833 LOOP_VINFO_VECT_FACTOR (res) = 0;
834 LOOP_VINFO_LOOP_NEST (res) = VEC_alloc (loop_p, heap, 3);
835 LOOP_VINFO_DATAREFS (res) = VEC_alloc (data_reference_p, heap, 10);
836 LOOP_VINFO_DDRS (res) = VEC_alloc (ddr_p, heap, 10 * 10);
837 LOOP_VINFO_UNALIGNED_DR (res) = NULL;
838 LOOP_VINFO_MAY_MISALIGN_STMTS (res) =
839 VEC_alloc (gimple, heap,
840 PARAM_VALUE (PARAM_VECT_MAX_VERSION_FOR_ALIGNMENT_CHECKS));
841 LOOP_VINFO_MAY_ALIAS_DDRS (res) =
842 VEC_alloc (ddr_p, heap,
843 PARAM_VALUE (PARAM_VECT_MAX_VERSION_FOR_ALIAS_CHECKS));
844 LOOP_VINFO_STRIDED_STORES (res) = VEC_alloc (gimple, heap, 10);
845 LOOP_VINFO_REDUCTIONS (res) = VEC_alloc (gimple, heap, 10);
846 LOOP_VINFO_REDUCTION_CHAINS (res) = VEC_alloc (gimple, heap, 10);
847 LOOP_VINFO_SLP_INSTANCES (res) = VEC_alloc (slp_instance, heap, 10);
848 LOOP_VINFO_SLP_UNROLLING_FACTOR (res) = 1;
849 LOOP_VINFO_PEELING_HTAB (res) = NULL;
850 LOOP_VINFO_PEELING_FOR_GAPS (res) = false;
856 /* Function destroy_loop_vec_info.
858 Free LOOP_VINFO struct, as well as all the stmt_vec_info structs of all the
859 stmts in the loop. */
862 destroy_loop_vec_info (loop_vec_info loop_vinfo, bool clean_stmts)
867 gimple_stmt_iterator si;
869 VEC (slp_instance, heap) *slp_instances;
870 slp_instance instance;
875 loop = LOOP_VINFO_LOOP (loop_vinfo);
877 bbs = LOOP_VINFO_BBS (loop_vinfo);
878 nbbs = loop->num_nodes;
882 free (LOOP_VINFO_BBS (loop_vinfo));
883 free_data_refs (LOOP_VINFO_DATAREFS (loop_vinfo));
884 free_dependence_relations (LOOP_VINFO_DDRS (loop_vinfo));
885 VEC_free (loop_p, heap, LOOP_VINFO_LOOP_NEST (loop_vinfo));
886 VEC_free (gimple, heap, LOOP_VINFO_MAY_MISALIGN_STMTS (loop_vinfo));
887 VEC_free (ddr_p, heap, LOOP_VINFO_MAY_ALIAS_DDRS (loop_vinfo));
894 for (j = 0; j < nbbs; j++)
896 basic_block bb = bbs[j];
897 for (si = gsi_start_phis (bb); !gsi_end_p (si); gsi_next (&si))
898 free_stmt_vec_info (gsi_stmt (si));
900 for (si = gsi_start_bb (bb); !gsi_end_p (si); )
902 gimple stmt = gsi_stmt (si);
903 /* Free stmt_vec_info. */
904 free_stmt_vec_info (stmt);
909 free (LOOP_VINFO_BBS (loop_vinfo));
910 free_data_refs (LOOP_VINFO_DATAREFS (loop_vinfo));
911 free_dependence_relations (LOOP_VINFO_DDRS (loop_vinfo));
912 VEC_free (loop_p, heap, LOOP_VINFO_LOOP_NEST (loop_vinfo));
913 VEC_free (gimple, heap, LOOP_VINFO_MAY_MISALIGN_STMTS (loop_vinfo));
914 VEC_free (ddr_p, heap, LOOP_VINFO_MAY_ALIAS_DDRS (loop_vinfo));
915 slp_instances = LOOP_VINFO_SLP_INSTANCES (loop_vinfo);
916 FOR_EACH_VEC_ELT (slp_instance, slp_instances, j, instance)
917 vect_free_slp_instance (instance);
919 VEC_free (slp_instance, heap, LOOP_VINFO_SLP_INSTANCES (loop_vinfo));
920 VEC_free (gimple, heap, LOOP_VINFO_STRIDED_STORES (loop_vinfo));
921 VEC_free (gimple, heap, LOOP_VINFO_REDUCTIONS (loop_vinfo));
922 VEC_free (gimple, heap, LOOP_VINFO_REDUCTION_CHAINS (loop_vinfo));
924 if (LOOP_VINFO_PEELING_HTAB (loop_vinfo))
925 htab_delete (LOOP_VINFO_PEELING_HTAB (loop_vinfo));
932 /* Function vect_analyze_loop_1.
934 Apply a set of analyses on LOOP, and create a loop_vec_info struct
935 for it. The different analyses will record information in the
936 loop_vec_info struct. This is a subset of the analyses applied in
937 vect_analyze_loop, to be applied on an inner-loop nested in the loop
938 that is now considered for (outer-loop) vectorization. */
941 vect_analyze_loop_1 (struct loop *loop)
943 loop_vec_info loop_vinfo;
945 if (vect_print_dump_info (REPORT_DETAILS))
946 fprintf (vect_dump, "===== analyze_loop_nest_1 =====");
948 /* Check the CFG characteristics of the loop (nesting, entry/exit, etc. */
950 loop_vinfo = vect_analyze_loop_form (loop);
953 if (vect_print_dump_info (REPORT_DETAILS))
954 fprintf (vect_dump, "bad inner-loop form.");
962 /* Function vect_analyze_loop_form.
964 Verify that certain CFG restrictions hold, including:
965 - the loop has a pre-header
966 - the loop has a single entry and exit
967 - the loop exit condition is simple enough, and the number of iterations
968 can be analyzed (a countable loop). */
971 vect_analyze_loop_form (struct loop *loop)
973 loop_vec_info loop_vinfo;
975 tree number_of_iterations = NULL;
976 loop_vec_info inner_loop_vinfo = NULL;
978 if (vect_print_dump_info (REPORT_DETAILS))
979 fprintf (vect_dump, "=== vect_analyze_loop_form ===");
981 /* Different restrictions apply when we are considering an inner-most loop,
982 vs. an outer (nested) loop.
983 (FORNOW. May want to relax some of these restrictions in the future). */
987 /* Inner-most loop. We currently require that the number of BBs is
988 exactly 2 (the header and latch). Vectorizable inner-most loops
999 if (loop->num_nodes != 2)
1001 if (vect_print_dump_info (REPORT_BAD_FORM_LOOPS))
1002 fprintf (vect_dump, "not vectorized: control flow in loop.");
1006 if (empty_block_p (loop->header))
1008 if (vect_print_dump_info (REPORT_BAD_FORM_LOOPS))
1009 fprintf (vect_dump, "not vectorized: empty loop.");
1015 struct loop *innerloop = loop->inner;
1018 /* Nested loop. We currently require that the loop is doubly-nested,
1019 contains a single inner loop, and the number of BBs is exactly 5.
1020 Vectorizable outer-loops look like this:
1032 The inner-loop has the properties expected of inner-most loops
1033 as described above. */
1035 if ((loop->inner)->inner || (loop->inner)->next)
1037 if (vect_print_dump_info (REPORT_BAD_FORM_LOOPS))
1038 fprintf (vect_dump, "not vectorized: multiple nested loops.");
1042 /* Analyze the inner-loop. */
1043 inner_loop_vinfo = vect_analyze_loop_1 (loop->inner);
1044 if (!inner_loop_vinfo)
1046 if (vect_print_dump_info (REPORT_BAD_FORM_LOOPS))
1047 fprintf (vect_dump, "not vectorized: Bad inner loop.");
1051 if (!expr_invariant_in_loop_p (loop,
1052 LOOP_VINFO_NITERS (inner_loop_vinfo)))
1054 if (vect_print_dump_info (REPORT_BAD_FORM_LOOPS))
1056 "not vectorized: inner-loop count not invariant.");
1057 destroy_loop_vec_info (inner_loop_vinfo, true);
1061 if (loop->num_nodes != 5)
1063 if (vect_print_dump_info (REPORT_BAD_FORM_LOOPS))
1064 fprintf (vect_dump, "not vectorized: control flow in loop.");
1065 destroy_loop_vec_info (inner_loop_vinfo, true);
1069 gcc_assert (EDGE_COUNT (innerloop->header->preds) == 2);
1070 entryedge = EDGE_PRED (innerloop->header, 0);
1071 if (EDGE_PRED (innerloop->header, 0)->src == innerloop->latch)
1072 entryedge = EDGE_PRED (innerloop->header, 1);
1074 if (entryedge->src != loop->header
1075 || !single_exit (innerloop)
1076 || single_exit (innerloop)->dest != EDGE_PRED (loop->latch, 0)->src)
1078 if (vect_print_dump_info (REPORT_BAD_FORM_LOOPS))
1079 fprintf (vect_dump, "not vectorized: unsupported outerloop form.");
1080 destroy_loop_vec_info (inner_loop_vinfo, true);
1084 if (vect_print_dump_info (REPORT_DETAILS))
1085 fprintf (vect_dump, "Considering outer-loop vectorization.");
1088 if (!single_exit (loop)
1089 || EDGE_COUNT (loop->header->preds) != 2)
1091 if (vect_print_dump_info (REPORT_BAD_FORM_LOOPS))
1093 if (!single_exit (loop))
1094 fprintf (vect_dump, "not vectorized: multiple exits.");
1095 else if (EDGE_COUNT (loop->header->preds) != 2)
1096 fprintf (vect_dump, "not vectorized: too many incoming edges.");
1098 if (inner_loop_vinfo)
1099 destroy_loop_vec_info (inner_loop_vinfo, true);
1103 /* We assume that the loop exit condition is at the end of the loop. i.e,
1104 that the loop is represented as a do-while (with a proper if-guard
1105 before the loop if needed), where the loop header contains all the
1106 executable statements, and the latch is empty. */
1107 if (!empty_block_p (loop->latch)
1108 || !gimple_seq_empty_p (phi_nodes (loop->latch)))
1110 if (vect_print_dump_info (REPORT_BAD_FORM_LOOPS))
1111 fprintf (vect_dump, "not vectorized: unexpected loop form.");
1112 if (inner_loop_vinfo)
1113 destroy_loop_vec_info (inner_loop_vinfo, true);
1117 /* Make sure there exists a single-predecessor exit bb: */
1118 if (!single_pred_p (single_exit (loop)->dest))
1120 edge e = single_exit (loop);
1121 if (!(e->flags & EDGE_ABNORMAL))
1123 split_loop_exit_edge (e);
1124 if (vect_print_dump_info (REPORT_DETAILS))
1125 fprintf (vect_dump, "split exit edge.");
1129 if (vect_print_dump_info (REPORT_BAD_FORM_LOOPS))
1130 fprintf (vect_dump, "not vectorized: abnormal loop exit edge.");
1131 if (inner_loop_vinfo)
1132 destroy_loop_vec_info (inner_loop_vinfo, true);
1137 loop_cond = vect_get_loop_niters (loop, &number_of_iterations);
1140 if (vect_print_dump_info (REPORT_BAD_FORM_LOOPS))
1141 fprintf (vect_dump, "not vectorized: complicated exit condition.");
1142 if (inner_loop_vinfo)
1143 destroy_loop_vec_info (inner_loop_vinfo, true);
1147 if (!number_of_iterations)
1149 if (vect_print_dump_info (REPORT_BAD_FORM_LOOPS))
1151 "not vectorized: number of iterations cannot be computed.");
1152 if (inner_loop_vinfo)
1153 destroy_loop_vec_info (inner_loop_vinfo, true);
1157 if (chrec_contains_undetermined (number_of_iterations))
1159 if (vect_print_dump_info (REPORT_BAD_FORM_LOOPS))
1160 fprintf (vect_dump, "Infinite number of iterations.");
1161 if (inner_loop_vinfo)
1162 destroy_loop_vec_info (inner_loop_vinfo, true);
1166 if (!NITERS_KNOWN_P (number_of_iterations))
1168 if (vect_print_dump_info (REPORT_DETAILS))
1170 fprintf (vect_dump, "Symbolic number of iterations is ");
1171 print_generic_expr (vect_dump, number_of_iterations, TDF_DETAILS);
1174 else if (TREE_INT_CST_LOW (number_of_iterations) == 0)
1176 if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS))
1177 fprintf (vect_dump, "not vectorized: number of iterations = 0.");
1178 if (inner_loop_vinfo)
1179 destroy_loop_vec_info (inner_loop_vinfo, false);
1183 loop_vinfo = new_loop_vec_info (loop);
1184 LOOP_VINFO_NITERS (loop_vinfo) = number_of_iterations;
1185 LOOP_VINFO_NITERS_UNCHANGED (loop_vinfo) = number_of_iterations;
1187 STMT_VINFO_TYPE (vinfo_for_stmt (loop_cond)) = loop_exit_ctrl_vec_info_type;
1189 /* CHECKME: May want to keep it around it in the future. */
1190 if (inner_loop_vinfo)
1191 destroy_loop_vec_info (inner_loop_vinfo, false);
1193 gcc_assert (!loop->aux);
1194 loop->aux = loop_vinfo;
1199 /* Get cost by calling cost target builtin. */
1202 vect_get_cost (enum vect_cost_for_stmt type_of_cost)
1204 tree dummy_type = NULL;
1207 return targetm.vectorize.builtin_vectorization_cost (type_of_cost,
1212 /* Function vect_analyze_loop_operations.
1214 Scan the loop stmts and make sure they are all vectorizable. */
1217 vect_analyze_loop_operations (loop_vec_info loop_vinfo, bool slp)
1219 struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
1220 basic_block *bbs = LOOP_VINFO_BBS (loop_vinfo);
1221 int nbbs = loop->num_nodes;
1222 gimple_stmt_iterator si;
1223 unsigned int vectorization_factor = 0;
1226 stmt_vec_info stmt_info;
1227 bool need_to_vectorize = false;
1228 int min_profitable_iters;
1229 int min_scalar_loop_bound;
1231 bool only_slp_in_loop = true, ok;
1233 if (vect_print_dump_info (REPORT_DETAILS))
1234 fprintf (vect_dump, "=== vect_analyze_loop_operations ===");
1236 gcc_assert (LOOP_VINFO_VECT_FACTOR (loop_vinfo));
1237 vectorization_factor = LOOP_VINFO_VECT_FACTOR (loop_vinfo);
1240 /* If all the stmts in the loop can be SLPed, we perform only SLP, and
1241 vectorization factor of the loop is the unrolling factor required by
1242 the SLP instances. If that unrolling factor is 1, we say, that we
1243 perform pure SLP on loop - cross iteration parallelism is not
1245 for (i = 0; i < nbbs; i++)
1247 basic_block bb = bbs[i];
1248 for (si = gsi_start_bb (bb); !gsi_end_p (si); gsi_next (&si))
1250 gimple stmt = gsi_stmt (si);
1251 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
1252 gcc_assert (stmt_info);
1253 if ((STMT_VINFO_RELEVANT_P (stmt_info)
1254 || VECTORIZABLE_CYCLE_DEF (STMT_VINFO_DEF_TYPE (stmt_info)))
1255 && !PURE_SLP_STMT (stmt_info))
1256 /* STMT needs both SLP and loop-based vectorization. */
1257 only_slp_in_loop = false;
1261 if (only_slp_in_loop)
1262 vectorization_factor = LOOP_VINFO_SLP_UNROLLING_FACTOR (loop_vinfo);
1264 vectorization_factor = least_common_multiple (vectorization_factor,
1265 LOOP_VINFO_SLP_UNROLLING_FACTOR (loop_vinfo));
1267 LOOP_VINFO_VECT_FACTOR (loop_vinfo) = vectorization_factor;
1268 if (vect_print_dump_info (REPORT_DETAILS))
1269 fprintf (vect_dump, "Updating vectorization factor to %d ",
1270 vectorization_factor);
1273 for (i = 0; i < nbbs; i++)
1275 basic_block bb = bbs[i];
1277 for (si = gsi_start_phis (bb); !gsi_end_p (si); gsi_next (&si))
1279 phi = gsi_stmt (si);
1282 stmt_info = vinfo_for_stmt (phi);
1283 if (vect_print_dump_info (REPORT_DETAILS))
1285 fprintf (vect_dump, "examining phi: ");
1286 print_gimple_stmt (vect_dump, phi, 0, TDF_SLIM);
1289 /* Inner-loop loop-closed exit phi in outer-loop vectorization
1290 (i.e., a phi in the tail of the outer-loop). */
1291 if (! is_loop_header_bb_p (bb))
1293 /* FORNOW: we currently don't support the case that these phis
1294 are not used in the outerloop (unless it is double reduction,
1295 i.e., this phi is vect_reduction_def), cause this case
1296 requires to actually do something here. */
1297 if ((!STMT_VINFO_RELEVANT_P (stmt_info)
1298 || STMT_VINFO_LIVE_P (stmt_info))
1299 && STMT_VINFO_DEF_TYPE (stmt_info)
1300 != vect_double_reduction_def)
1302 if (vect_print_dump_info (REPORT_DETAILS))
1304 "Unsupported loop-closed phi in outer-loop.");
1308 /* If PHI is used in the outer loop, we check that its operand
1309 is defined in the inner loop. */
1310 if (STMT_VINFO_RELEVANT_P (stmt_info))
1315 if (gimple_phi_num_args (phi) != 1)
1318 phi_op = PHI_ARG_DEF (phi, 0);
1319 if (TREE_CODE (phi_op) != SSA_NAME)
1322 op_def_stmt = SSA_NAME_DEF_STMT (phi_op);
1323 if (!op_def_stmt || !vinfo_for_stmt (op_def_stmt))
1326 if (STMT_VINFO_RELEVANT (vinfo_for_stmt (op_def_stmt))
1327 != vect_used_in_outer
1328 && STMT_VINFO_RELEVANT (vinfo_for_stmt (op_def_stmt))
1329 != vect_used_in_outer_by_reduction)
1336 gcc_assert (stmt_info);
1338 if (STMT_VINFO_LIVE_P (stmt_info))
1340 /* FORNOW: not yet supported. */
1341 if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS))
1342 fprintf (vect_dump, "not vectorized: value used after loop.");
1346 if (STMT_VINFO_RELEVANT (stmt_info) == vect_used_in_scope
1347 && STMT_VINFO_DEF_TYPE (stmt_info) != vect_induction_def)
1349 /* A scalar-dependence cycle that we don't support. */
1350 if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS))
1351 fprintf (vect_dump, "not vectorized: scalar dependence cycle.");
1355 if (STMT_VINFO_RELEVANT_P (stmt_info))
1357 need_to_vectorize = true;
1358 if (STMT_VINFO_DEF_TYPE (stmt_info) == vect_induction_def)
1359 ok = vectorizable_induction (phi, NULL, NULL);
1364 if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS))
1367 "not vectorized: relevant phi not supported: ");
1368 print_gimple_stmt (vect_dump, phi, 0, TDF_SLIM);
1374 for (si = gsi_start_bb (bb); !gsi_end_p (si); gsi_next (&si))
1376 gimple stmt = gsi_stmt (si);
1377 if (!vect_analyze_stmt (stmt, &need_to_vectorize, NULL))
1382 /* All operations in the loop are either irrelevant (deal with loop
1383 control, or dead), or only used outside the loop and can be moved
1384 out of the loop (e.g. invariants, inductions). The loop can be
1385 optimized away by scalar optimizations. We're better off not
1386 touching this loop. */
1387 if (!need_to_vectorize)
1389 if (vect_print_dump_info (REPORT_DETAILS))
1391 "All the computation can be taken out of the loop.");
1392 if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS))
1394 "not vectorized: redundant loop. no profit to vectorize.");
1398 if (LOOP_VINFO_NITERS_KNOWN_P (loop_vinfo)
1399 && vect_print_dump_info (REPORT_DETAILS))
1401 "vectorization_factor = %d, niters = " HOST_WIDE_INT_PRINT_DEC,
1402 vectorization_factor, LOOP_VINFO_INT_NITERS (loop_vinfo));
1404 if (LOOP_VINFO_NITERS_KNOWN_P (loop_vinfo)
1405 && (LOOP_VINFO_INT_NITERS (loop_vinfo) < vectorization_factor))
1407 if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS))
1408 fprintf (vect_dump, "not vectorized: iteration count too small.");
1409 if (vect_print_dump_info (REPORT_DETAILS))
1410 fprintf (vect_dump,"not vectorized: iteration count smaller than "
1411 "vectorization factor.");
1415 /* Analyze cost. Decide if worth while to vectorize. */
1417 /* Once VF is set, SLP costs should be updated since the number of created
1418 vector stmts depends on VF. */
1419 vect_update_slp_costs_according_to_vf (loop_vinfo);
1421 min_profitable_iters = vect_estimate_min_profitable_iters (loop_vinfo);
1422 LOOP_VINFO_COST_MODEL_MIN_ITERS (loop_vinfo) = min_profitable_iters;
1424 if (min_profitable_iters < 0)
1426 if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS))
1427 fprintf (vect_dump, "not vectorized: vectorization not profitable.");
1428 if (vect_print_dump_info (REPORT_DETAILS))
1429 fprintf (vect_dump, "not vectorized: vector version will never be "
1434 min_scalar_loop_bound = ((PARAM_VALUE (PARAM_MIN_VECT_LOOP_BOUND)
1435 * vectorization_factor) - 1);
1437 /* Use the cost model only if it is more conservative than user specified
1440 th = (unsigned) min_scalar_loop_bound;
1441 if (min_profitable_iters
1442 && (!min_scalar_loop_bound
1443 || min_profitable_iters > min_scalar_loop_bound))
1444 th = (unsigned) min_profitable_iters;
1446 if (LOOP_VINFO_NITERS_KNOWN_P (loop_vinfo)
1447 && LOOP_VINFO_INT_NITERS (loop_vinfo) <= th)
1449 if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS))
1450 fprintf (vect_dump, "not vectorized: vectorization not "
1452 if (vect_print_dump_info (REPORT_DETAILS))
1453 fprintf (vect_dump, "not vectorized: iteration count smaller than "
1454 "user specified loop bound parameter or minimum "
1455 "profitable iterations (whichever is more conservative).");
1459 if (!LOOP_VINFO_NITERS_KNOWN_P (loop_vinfo)
1460 || LOOP_VINFO_INT_NITERS (loop_vinfo) % vectorization_factor != 0
1461 || LOOP_PEELING_FOR_ALIGNMENT (loop_vinfo))
1463 if (vect_print_dump_info (REPORT_DETAILS))
1464 fprintf (vect_dump, "epilog loop required.");
1465 if (!vect_can_advance_ivs_p (loop_vinfo))
1467 if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS))
1469 "not vectorized: can't create epilog loop 1.");
1472 if (!slpeel_can_duplicate_loop_p (loop, single_exit (loop)))
1474 if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS))
1476 "not vectorized: can't create epilog loop 2.");
1485 /* Function vect_analyze_loop_2.
1487 Apply a set of analyses on LOOP, and create a loop_vec_info struct
1488 for it. The different analyses will record information in the
1489 loop_vec_info struct. */
1491 vect_analyze_loop_2 (loop_vec_info loop_vinfo)
1493 bool ok, slp = false;
1494 int max_vf = MAX_VECTORIZATION_FACTOR;
1497 /* Find all data references in the loop (which correspond to vdefs/vuses)
1498 and analyze their evolution in the loop. Also adjust the minimal
1499 vectorization factor according to the loads and stores.
1501 FORNOW: Handle only simple, array references, which
1502 alignment can be forced, and aligned pointer-references. */
1504 ok = vect_analyze_data_refs (loop_vinfo, NULL, &min_vf);
1507 if (vect_print_dump_info (REPORT_DETAILS))
1508 fprintf (vect_dump, "bad data references.");
1512 /* Classify all cross-iteration scalar data-flow cycles.
1513 Cross-iteration cycles caused by virtual phis are analyzed separately. */
1515 vect_analyze_scalar_cycles (loop_vinfo);
1517 vect_pattern_recog (loop_vinfo, NULL);
1519 /* Data-flow analysis to detect stmts that do not need to be vectorized. */
1521 ok = vect_mark_stmts_to_be_vectorized (loop_vinfo);
1524 if (vect_print_dump_info (REPORT_DETAILS))
1525 fprintf (vect_dump, "unexpected pattern.");
1529 /* Analyze data dependences between the data-refs in the loop
1530 and adjust the maximum vectorization factor according to
1532 FORNOW: fail at the first data dependence that we encounter. */
1534 ok = vect_analyze_data_ref_dependences (loop_vinfo, NULL, &max_vf);
1538 if (vect_print_dump_info (REPORT_DETAILS))
1539 fprintf (vect_dump, "bad data dependence.");
1543 ok = vect_determine_vectorization_factor (loop_vinfo);
1546 if (vect_print_dump_info (REPORT_DETAILS))
1547 fprintf (vect_dump, "can't determine vectorization factor.");
1550 if (max_vf < LOOP_VINFO_VECT_FACTOR (loop_vinfo))
1552 if (vect_print_dump_info (REPORT_DETAILS))
1553 fprintf (vect_dump, "bad data dependence.");
1557 /* Analyze the alignment of the data-refs in the loop.
1558 Fail if a data reference is found that cannot be vectorized. */
1560 ok = vect_analyze_data_refs_alignment (loop_vinfo, NULL);
1563 if (vect_print_dump_info (REPORT_DETAILS))
1564 fprintf (vect_dump, "bad data alignment.");
1568 /* Analyze the access patterns of the data-refs in the loop (consecutive,
1569 complex, etc.). FORNOW: Only handle consecutive access pattern. */
1571 ok = vect_analyze_data_ref_accesses (loop_vinfo, NULL);
1574 if (vect_print_dump_info (REPORT_DETAILS))
1575 fprintf (vect_dump, "bad data access.");
1579 /* Prune the list of ddrs to be tested at run-time by versioning for alias.
1580 It is important to call pruning after vect_analyze_data_ref_accesses,
1581 since we use grouping information gathered by interleaving analysis. */
1582 ok = vect_prune_runtime_alias_test_list (loop_vinfo);
1585 if (vect_print_dump_info (REPORT_DETAILS))
1586 fprintf (vect_dump, "too long list of versioning for alias "
1591 /* This pass will decide on using loop versioning and/or loop peeling in
1592 order to enhance the alignment of data references in the loop. */
1594 ok = vect_enhance_data_refs_alignment (loop_vinfo);
1597 if (vect_print_dump_info (REPORT_DETAILS))
1598 fprintf (vect_dump, "bad data alignment.");
1602 /* Check the SLP opportunities in the loop, analyze and build SLP trees. */
1603 ok = vect_analyze_slp (loop_vinfo, NULL);
1606 /* Decide which possible SLP instances to SLP. */
1607 slp = vect_make_slp_decision (loop_vinfo);
1609 /* Find stmts that need to be both vectorized and SLPed. */
1610 vect_detect_hybrid_slp (loop_vinfo);
1615 /* Scan all the operations in the loop and make sure they are
1618 ok = vect_analyze_loop_operations (loop_vinfo, slp);
1621 if (vect_print_dump_info (REPORT_DETAILS))
1622 fprintf (vect_dump, "bad operation or unsupported loop bound.");
1629 /* Function vect_analyze_loop.
1631 Apply a set of analyses on LOOP, and create a loop_vec_info struct
1632 for it. The different analyses will record information in the
1633 loop_vec_info struct. */
1635 vect_analyze_loop (struct loop *loop)
1637 loop_vec_info loop_vinfo;
1638 unsigned int vector_sizes;
1640 /* Autodetect first vector size we try. */
1641 current_vector_size = 0;
1642 vector_sizes = targetm.vectorize.autovectorize_vector_sizes ();
1644 if (vect_print_dump_info (REPORT_DETAILS))
1645 fprintf (vect_dump, "===== analyze_loop_nest =====");
1647 if (loop_outer (loop)
1648 && loop_vec_info_for_loop (loop_outer (loop))
1649 && LOOP_VINFO_VECTORIZABLE_P (loop_vec_info_for_loop (loop_outer (loop))))
1651 if (vect_print_dump_info (REPORT_DETAILS))
1652 fprintf (vect_dump, "outer-loop already vectorized.");
1658 /* Check the CFG characteristics of the loop (nesting, entry/exit). */
1659 loop_vinfo = vect_analyze_loop_form (loop);
1662 if (vect_print_dump_info (REPORT_DETAILS))
1663 fprintf (vect_dump, "bad loop form.");
1667 if (vect_analyze_loop_2 (loop_vinfo))
1669 LOOP_VINFO_VECTORIZABLE_P (loop_vinfo) = 1;
1674 destroy_loop_vec_info (loop_vinfo, true);
1676 vector_sizes &= ~current_vector_size;
1677 if (vector_sizes == 0
1678 || current_vector_size == 0)
1681 /* Try the next biggest vector size. */
1682 current_vector_size = 1 << floor_log2 (vector_sizes);
1683 if (vect_print_dump_info (REPORT_DETAILS))
1684 fprintf (vect_dump, "***** Re-trying analysis with "
1685 "vector size %d\n", current_vector_size);
1690 /* Function reduction_code_for_scalar_code
1693 CODE - tree_code of a reduction operations.
1696 REDUC_CODE - the corresponding tree-code to be used to reduce the
1697 vector of partial results into a single scalar result (which
1698 will also reside in a vector) or ERROR_MARK if the operation is
1699 a supported reduction operation, but does not have such tree-code.
1701 Return FALSE if CODE currently cannot be vectorized as reduction. */
1704 reduction_code_for_scalar_code (enum tree_code code,
1705 enum tree_code *reduc_code)
1710 *reduc_code = REDUC_MAX_EXPR;
1714 *reduc_code = REDUC_MIN_EXPR;
1718 *reduc_code = REDUC_PLUS_EXPR;
1726 *reduc_code = ERROR_MARK;
1735 /* Error reporting helper for vect_is_simple_reduction below. GIMPLE statement
1736 STMT is printed with a message MSG. */
1739 report_vect_op (gimple stmt, const char *msg)
1741 fprintf (vect_dump, "%s", msg);
1742 print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM);
1746 /* Detect SLP reduction of the form:
1756 PHI is the reduction phi node (#a1 = phi <a5, a0> above)
1757 FIRST_STMT is the first reduction stmt in the chain
1758 (a2 = operation (a1)).
1760 Return TRUE if a reduction chain was detected. */
1763 vect_is_slp_reduction (loop_vec_info loop_info, gimple phi, gimple first_stmt)
1765 struct loop *loop = (gimple_bb (phi))->loop_father;
1766 struct loop *vect_loop = LOOP_VINFO_LOOP (loop_info);
1767 enum tree_code code;
1768 gimple current_stmt = NULL, loop_use_stmt = NULL, first, next_stmt;
1769 stmt_vec_info use_stmt_info, current_stmt_info;
1771 imm_use_iterator imm_iter;
1772 use_operand_p use_p;
1773 int nloop_uses, size = 0, n_out_of_loop_uses;
1776 if (loop != vect_loop)
1779 lhs = PHI_RESULT (phi);
1780 code = gimple_assign_rhs_code (first_stmt);
1784 n_out_of_loop_uses = 0;
1785 FOR_EACH_IMM_USE_FAST (use_p, imm_iter, lhs)
1787 gimple use_stmt = USE_STMT (use_p);
1788 if (is_gimple_debug (use_stmt))
1791 use_stmt = USE_STMT (use_p);
1793 /* Check if we got back to the reduction phi. */
1794 if (use_stmt == phi)
1796 loop_use_stmt = use_stmt;
1801 if (flow_bb_inside_loop_p (loop, gimple_bb (use_stmt)))
1803 if (vinfo_for_stmt (use_stmt)
1804 && !STMT_VINFO_IN_PATTERN_P (vinfo_for_stmt (use_stmt)))
1806 loop_use_stmt = use_stmt;
1811 n_out_of_loop_uses++;
1813 /* There are can be either a single use in the loop or two uses in
1815 if (nloop_uses > 1 || (n_out_of_loop_uses && nloop_uses))
1822 /* We reached a statement with no loop uses. */
1823 if (nloop_uses == 0)
1826 /* This is a loop exit phi, and we haven't reached the reduction phi. */
1827 if (gimple_code (loop_use_stmt) == GIMPLE_PHI)
1830 if (!is_gimple_assign (loop_use_stmt)
1831 || code != gimple_assign_rhs_code (loop_use_stmt)
1832 || !flow_bb_inside_loop_p (loop, gimple_bb (loop_use_stmt)))
1835 /* Insert USE_STMT into reduction chain. */
1836 use_stmt_info = vinfo_for_stmt (loop_use_stmt);
1839 current_stmt_info = vinfo_for_stmt (current_stmt);
1840 GROUP_NEXT_ELEMENT (current_stmt_info) = loop_use_stmt;
1841 GROUP_FIRST_ELEMENT (use_stmt_info)
1842 = GROUP_FIRST_ELEMENT (current_stmt_info);
1845 GROUP_FIRST_ELEMENT (use_stmt_info) = loop_use_stmt;
1847 lhs = gimple_assign_lhs (loop_use_stmt);
1848 current_stmt = loop_use_stmt;
1852 if (!found || loop_use_stmt != phi || size < 2)
1855 /* Swap the operands, if needed, to make the reduction operand be the second
1857 lhs = PHI_RESULT (phi);
1858 next_stmt = GROUP_FIRST_ELEMENT (vinfo_for_stmt (current_stmt));
1861 if (gimple_assign_rhs2 (next_stmt) == lhs)
1863 tree op = gimple_assign_rhs1 (next_stmt);
1864 gimple def_stmt = NULL;
1866 if (TREE_CODE (op) == SSA_NAME)
1867 def_stmt = SSA_NAME_DEF_STMT (op);
1869 /* Check that the other def is either defined in the loop
1870 ("vect_internal_def"), or it's an induction (defined by a
1871 loop-header phi-node). */
1873 && gimple_bb (def_stmt)
1874 && flow_bb_inside_loop_p (loop, gimple_bb (def_stmt))
1875 && (is_gimple_assign (def_stmt)
1876 || is_gimple_call (def_stmt)
1877 || STMT_VINFO_DEF_TYPE (vinfo_for_stmt (def_stmt))
1878 == vect_induction_def
1879 || (gimple_code (def_stmt) == GIMPLE_PHI
1880 && STMT_VINFO_DEF_TYPE (vinfo_for_stmt (def_stmt))
1881 == vect_internal_def
1882 && !is_loop_header_bb_p (gimple_bb (def_stmt)))))
1884 lhs = gimple_assign_lhs (next_stmt);
1885 next_stmt = GROUP_NEXT_ELEMENT (vinfo_for_stmt (next_stmt));
1893 tree op = gimple_assign_rhs2 (next_stmt);
1894 gimple def_stmt = NULL;
1896 if (TREE_CODE (op) == SSA_NAME)
1897 def_stmt = SSA_NAME_DEF_STMT (op);
1899 /* Check that the other def is either defined in the loop
1900 ("vect_internal_def"), or it's an induction (defined by a
1901 loop-header phi-node). */
1903 && gimple_bb (def_stmt)
1904 && flow_bb_inside_loop_p (loop, gimple_bb (def_stmt))
1905 && (is_gimple_assign (def_stmt)
1906 || is_gimple_call (def_stmt)
1907 || STMT_VINFO_DEF_TYPE (vinfo_for_stmt (def_stmt))
1908 == vect_induction_def
1909 || (gimple_code (def_stmt) == GIMPLE_PHI
1910 && STMT_VINFO_DEF_TYPE (vinfo_for_stmt (def_stmt))
1911 == vect_internal_def
1912 && !is_loop_header_bb_p (gimple_bb (def_stmt)))))
1914 if (vect_print_dump_info (REPORT_DETAILS))
1916 fprintf (vect_dump, "swapping oprnds: ");
1917 print_gimple_stmt (vect_dump, next_stmt, 0, TDF_SLIM);
1920 swap_tree_operands (next_stmt,
1921 gimple_assign_rhs1_ptr (next_stmt),
1922 gimple_assign_rhs2_ptr (next_stmt));
1923 mark_symbols_for_renaming (next_stmt);
1929 lhs = gimple_assign_lhs (next_stmt);
1930 next_stmt = GROUP_NEXT_ELEMENT (vinfo_for_stmt (next_stmt));
1933 /* Save the chain for further analysis in SLP detection. */
1934 first = GROUP_FIRST_ELEMENT (vinfo_for_stmt (current_stmt));
1935 VEC_safe_push (gimple, heap, LOOP_VINFO_REDUCTION_CHAINS (loop_info), first);
1936 GROUP_SIZE (vinfo_for_stmt (first)) = size;
1942 /* Function vect_is_simple_reduction_1
1944 (1) Detect a cross-iteration def-use cycle that represents a simple
1945 reduction computation. We look for the following pattern:
1950 a2 = operation (a3, a1)
1953 1. operation is commutative and associative and it is safe to
1954 change the order of the computation (if CHECK_REDUCTION is true)
1955 2. no uses for a2 in the loop (a2 is used out of the loop)
1956 3. no uses of a1 in the loop besides the reduction operation
1957 4. no uses of a1 outside the loop.
1959 Conditions 1,4 are tested here.
1960 Conditions 2,3 are tested in vect_mark_stmts_to_be_vectorized.
1962 (2) Detect a cross-iteration def-use cycle in nested loops, i.e.,
1963 nested cycles, if CHECK_REDUCTION is false.
1965 (3) Detect cycles of phi nodes in outer-loop vectorization, i.e., double
1969 inner loop (def of a3)
1972 If MODIFY is true it tries also to rework the code in-place to enable
1973 detection of more reduction patterns. For the time being we rewrite
1974 "res -= RHS" into "rhs += -RHS" when it seems worthwhile.
1978 vect_is_simple_reduction_1 (loop_vec_info loop_info, gimple phi,
1979 bool check_reduction, bool *double_reduc,
1982 struct loop *loop = (gimple_bb (phi))->loop_father;
1983 struct loop *vect_loop = LOOP_VINFO_LOOP (loop_info);
1984 edge latch_e = loop_latch_edge (loop);
1985 tree loop_arg = PHI_ARG_DEF_FROM_EDGE (phi, latch_e);
1986 gimple def_stmt, def1 = NULL, def2 = NULL;
1987 enum tree_code orig_code, code;
1988 tree op1, op2, op3 = NULL_TREE, op4 = NULL_TREE;
1992 imm_use_iterator imm_iter;
1993 use_operand_p use_p;
1996 *double_reduc = false;
1998 /* If CHECK_REDUCTION is true, we assume inner-most loop vectorization,
1999 otherwise, we assume outer loop vectorization. */
2000 gcc_assert ((check_reduction && loop == vect_loop)
2001 || (!check_reduction && flow_loop_nested_p (vect_loop, loop)));
2003 name = PHI_RESULT (phi);
2005 FOR_EACH_IMM_USE_FAST (use_p, imm_iter, name)
2007 gimple use_stmt = USE_STMT (use_p);
2008 if (is_gimple_debug (use_stmt))
2011 if (!flow_bb_inside_loop_p (loop, gimple_bb (use_stmt)))
2013 if (vect_print_dump_info (REPORT_DETAILS))
2014 fprintf (vect_dump, "intermediate value used outside loop.");
2019 if (vinfo_for_stmt (use_stmt)
2020 && !is_pattern_stmt_p (vinfo_for_stmt (use_stmt)))
2024 if (vect_print_dump_info (REPORT_DETAILS))
2025 fprintf (vect_dump, "reduction used in loop.");
2030 if (TREE_CODE (loop_arg) != SSA_NAME)
2032 if (vect_print_dump_info (REPORT_DETAILS))
2034 fprintf (vect_dump, "reduction: not ssa_name: ");
2035 print_generic_expr (vect_dump, loop_arg, TDF_SLIM);
2040 def_stmt = SSA_NAME_DEF_STMT (loop_arg);
2043 if (vect_print_dump_info (REPORT_DETAILS))
2044 fprintf (vect_dump, "reduction: no def_stmt.");
2048 if (!is_gimple_assign (def_stmt) && gimple_code (def_stmt) != GIMPLE_PHI)
2050 if (vect_print_dump_info (REPORT_DETAILS))
2051 print_gimple_stmt (vect_dump, def_stmt, 0, TDF_SLIM);
2055 if (is_gimple_assign (def_stmt))
2057 name = gimple_assign_lhs (def_stmt);
2062 name = PHI_RESULT (def_stmt);
2067 FOR_EACH_IMM_USE_FAST (use_p, imm_iter, name)
2069 gimple use_stmt = USE_STMT (use_p);
2070 if (is_gimple_debug (use_stmt))
2072 if (flow_bb_inside_loop_p (loop, gimple_bb (use_stmt))
2073 && vinfo_for_stmt (use_stmt)
2074 && !is_pattern_stmt_p (vinfo_for_stmt (use_stmt)))
2078 if (vect_print_dump_info (REPORT_DETAILS))
2079 fprintf (vect_dump, "reduction used in loop.");
2084 /* If DEF_STMT is a phi node itself, we expect it to have a single argument
2085 defined in the inner loop. */
2088 op1 = PHI_ARG_DEF (def_stmt, 0);
2090 if (gimple_phi_num_args (def_stmt) != 1
2091 || TREE_CODE (op1) != SSA_NAME)
2093 if (vect_print_dump_info (REPORT_DETAILS))
2094 fprintf (vect_dump, "unsupported phi node definition.");
2099 def1 = SSA_NAME_DEF_STMT (op1);
2100 if (flow_bb_inside_loop_p (loop, gimple_bb (def_stmt))
2102 && flow_bb_inside_loop_p (loop->inner, gimple_bb (def1))
2103 && is_gimple_assign (def1))
2105 if (vect_print_dump_info (REPORT_DETAILS))
2106 report_vect_op (def_stmt, "detected double reduction: ");
2108 *double_reduc = true;
2115 code = orig_code = gimple_assign_rhs_code (def_stmt);
2117 /* We can handle "res -= x[i]", which is non-associative by
2118 simply rewriting this into "res += -x[i]". Avoid changing
2119 gimple instruction for the first simple tests and only do this
2120 if we're allowed to change code at all. */
2121 if (code == MINUS_EXPR
2123 && (op1 = gimple_assign_rhs1 (def_stmt))
2124 && TREE_CODE (op1) == SSA_NAME
2125 && SSA_NAME_DEF_STMT (op1) == phi)
2129 && (!commutative_tree_code (code) || !associative_tree_code (code)))
2131 if (vect_print_dump_info (REPORT_DETAILS))
2132 report_vect_op (def_stmt, "reduction: not commutative/associative: ");
2136 if (get_gimple_rhs_class (code) != GIMPLE_BINARY_RHS)
2138 if (code != COND_EXPR)
2140 if (vect_print_dump_info (REPORT_DETAILS))
2141 report_vect_op (def_stmt, "reduction: not binary operation: ");
2146 op3 = gimple_assign_rhs1 (def_stmt);
2147 if (COMPARISON_CLASS_P (op3))
2149 op4 = TREE_OPERAND (op3, 1);
2150 op3 = TREE_OPERAND (op3, 0);
2153 op1 = gimple_assign_rhs2 (def_stmt);
2154 op2 = gimple_assign_rhs3 (def_stmt);
2156 if (TREE_CODE (op1) != SSA_NAME && TREE_CODE (op2) != SSA_NAME)
2158 if (vect_print_dump_info (REPORT_DETAILS))
2159 report_vect_op (def_stmt, "reduction: uses not ssa_names: ");
2166 op1 = gimple_assign_rhs1 (def_stmt);
2167 op2 = gimple_assign_rhs2 (def_stmt);
2169 if (TREE_CODE (op1) != SSA_NAME && TREE_CODE (op2) != SSA_NAME)
2171 if (vect_print_dump_info (REPORT_DETAILS))
2172 report_vect_op (def_stmt, "reduction: uses not ssa_names: ");
2178 type = TREE_TYPE (gimple_assign_lhs (def_stmt));
2179 if ((TREE_CODE (op1) == SSA_NAME
2180 && !types_compatible_p (type,TREE_TYPE (op1)))
2181 || (TREE_CODE (op2) == SSA_NAME
2182 && !types_compatible_p (type, TREE_TYPE (op2)))
2183 || (op3 && TREE_CODE (op3) == SSA_NAME
2184 && !types_compatible_p (type, TREE_TYPE (op3)))
2185 || (op4 && TREE_CODE (op4) == SSA_NAME
2186 && !types_compatible_p (type, TREE_TYPE (op4))))
2188 if (vect_print_dump_info (REPORT_DETAILS))
2190 fprintf (vect_dump, "reduction: multiple types: operation type: ");
2191 print_generic_expr (vect_dump, type, TDF_SLIM);
2192 fprintf (vect_dump, ", operands types: ");
2193 print_generic_expr (vect_dump, TREE_TYPE (op1), TDF_SLIM);
2194 fprintf (vect_dump, ",");
2195 print_generic_expr (vect_dump, TREE_TYPE (op2), TDF_SLIM);
2198 fprintf (vect_dump, ",");
2199 print_generic_expr (vect_dump, TREE_TYPE (op3), TDF_SLIM);
2204 fprintf (vect_dump, ",");
2205 print_generic_expr (vect_dump, TREE_TYPE (op4), TDF_SLIM);
2212 /* Check that it's ok to change the order of the computation.
2213 Generally, when vectorizing a reduction we change the order of the
2214 computation. This may change the behavior of the program in some
2215 cases, so we need to check that this is ok. One exception is when
2216 vectorizing an outer-loop: the inner-loop is executed sequentially,
2217 and therefore vectorizing reductions in the inner-loop during
2218 outer-loop vectorization is safe. */
2220 /* CHECKME: check for !flag_finite_math_only too? */
2221 if (SCALAR_FLOAT_TYPE_P (type) && !flag_associative_math
2224 /* Changing the order of operations changes the semantics. */
2225 if (vect_print_dump_info (REPORT_DETAILS))
2226 report_vect_op (def_stmt, "reduction: unsafe fp math optimization: ");
2229 else if (INTEGRAL_TYPE_P (type) && TYPE_OVERFLOW_TRAPS (type)
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 int math optimization: ");
2237 else if (SAT_FIXED_POINT_TYPE_P (type) && check_reduction)
2239 /* Changing the order of operations changes the semantics. */
2240 if (vect_print_dump_info (REPORT_DETAILS))
2241 report_vect_op (def_stmt,
2242 "reduction: unsafe fixed-point math optimization: ");
2246 /* If we detected "res -= x[i]" earlier, rewrite it into
2247 "res += -x[i]" now. If this turns out to be useless reassoc
2248 will clean it up again. */
2249 if (orig_code == MINUS_EXPR)
2251 tree rhs = gimple_assign_rhs2 (def_stmt);
2252 tree negrhs = make_ssa_name (SSA_NAME_VAR (rhs), NULL);
2253 gimple negate_stmt = gimple_build_assign_with_ops (NEGATE_EXPR, negrhs,
2255 gimple_stmt_iterator gsi = gsi_for_stmt (def_stmt);
2256 set_vinfo_for_stmt (negate_stmt, new_stmt_vec_info (negate_stmt,
2258 gsi_insert_before (&gsi, negate_stmt, GSI_NEW_STMT);
2259 gimple_assign_set_rhs2 (def_stmt, negrhs);
2260 gimple_assign_set_rhs_code (def_stmt, PLUS_EXPR);
2261 update_stmt (def_stmt);
2264 /* Reduction is safe. We're dealing with one of the following:
2265 1) integer arithmetic and no trapv
2266 2) floating point arithmetic, and special flags permit this optimization
2267 3) nested cycle (i.e., outer loop vectorization). */
2268 if (TREE_CODE (op1) == SSA_NAME)
2269 def1 = SSA_NAME_DEF_STMT (op1);
2271 if (TREE_CODE (op2) == SSA_NAME)
2272 def2 = SSA_NAME_DEF_STMT (op2);
2274 if (code != COND_EXPR
2275 && ((!def1 || gimple_nop_p (def1)) && (!def2 || gimple_nop_p (def2))))
2277 if (vect_print_dump_info (REPORT_DETAILS))
2278 report_vect_op (def_stmt, "reduction: no defs for operands: ");
2282 /* Check that one def is the reduction def, defined by PHI,
2283 the other def is either defined in the loop ("vect_internal_def"),
2284 or it's an induction (defined by a loop-header phi-node). */
2286 if (def2 && def2 == phi
2287 && (code == COND_EXPR
2288 || !def1 || gimple_nop_p (def1)
2289 || (def1 && flow_bb_inside_loop_p (loop, gimple_bb (def1))
2290 && (is_gimple_assign (def1)
2291 || is_gimple_call (def1)
2292 || STMT_VINFO_DEF_TYPE (vinfo_for_stmt (def1))
2293 == vect_induction_def
2294 || (gimple_code (def1) == GIMPLE_PHI
2295 && STMT_VINFO_DEF_TYPE (vinfo_for_stmt (def1))
2296 == vect_internal_def
2297 && !is_loop_header_bb_p (gimple_bb (def1)))))))
2299 if (vect_print_dump_info (REPORT_DETAILS))
2300 report_vect_op (def_stmt, "detected reduction: ");
2304 if (def1 && def1 == phi
2305 && (code == COND_EXPR
2306 || !def2 || gimple_nop_p (def2)
2307 || (def2 && flow_bb_inside_loop_p (loop, gimple_bb (def2))
2308 && (is_gimple_assign (def2)
2309 || is_gimple_call (def2)
2310 || STMT_VINFO_DEF_TYPE (vinfo_for_stmt (def2))
2311 == vect_induction_def
2312 || (gimple_code (def2) == GIMPLE_PHI
2313 && STMT_VINFO_DEF_TYPE (vinfo_for_stmt (def2))
2314 == vect_internal_def
2315 && !is_loop_header_bb_p (gimple_bb (def2)))))))
2317 if (check_reduction)
2319 /* Swap operands (just for simplicity - so that the rest of the code
2320 can assume that the reduction variable is always the last (second)
2322 if (vect_print_dump_info (REPORT_DETAILS))
2323 report_vect_op (def_stmt,
2324 "detected reduction: need to swap operands: ");
2326 swap_tree_operands (def_stmt, gimple_assign_rhs1_ptr (def_stmt),
2327 gimple_assign_rhs2_ptr (def_stmt));
2331 if (vect_print_dump_info (REPORT_DETAILS))
2332 report_vect_op (def_stmt, "detected reduction: ");
2338 /* Try to find SLP reduction chain. */
2339 if (check_reduction && vect_is_slp_reduction (loop_info, phi, def_stmt))
2341 if (vect_print_dump_info (REPORT_DETAILS))
2342 report_vect_op (def_stmt, "reduction: detected reduction chain: ");
2347 if (vect_print_dump_info (REPORT_DETAILS))
2348 report_vect_op (def_stmt, "reduction: unknown pattern: ");
2353 /* Wrapper around vect_is_simple_reduction_1, that won't modify code
2354 in-place. Arguments as there. */
2357 vect_is_simple_reduction (loop_vec_info loop_info, gimple phi,
2358 bool check_reduction, bool *double_reduc)
2360 return vect_is_simple_reduction_1 (loop_info, phi, check_reduction,
2361 double_reduc, false);
2364 /* Wrapper around vect_is_simple_reduction_1, which will modify code
2365 in-place if it enables detection of more reductions. Arguments
2369 vect_force_simple_reduction (loop_vec_info loop_info, gimple phi,
2370 bool check_reduction, bool *double_reduc)
2372 return vect_is_simple_reduction_1 (loop_info, phi, check_reduction,
2373 double_reduc, true);
2376 /* Calculate the cost of one scalar iteration of the loop. */
2378 vect_get_single_scalar_iteraion_cost (loop_vec_info loop_vinfo)
2380 struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
2381 basic_block *bbs = LOOP_VINFO_BBS (loop_vinfo);
2382 int nbbs = loop->num_nodes, factor, scalar_single_iter_cost = 0;
2383 int innerloop_iters, i, stmt_cost;
2385 /* Count statements in scalar loop. Using this as scalar cost for a single
2388 TODO: Add outer loop support.
2390 TODO: Consider assigning different costs to different scalar
2394 innerloop_iters = 1;
2396 innerloop_iters = 50; /* FIXME */
2398 for (i = 0; i < nbbs; i++)
2400 gimple_stmt_iterator si;
2401 basic_block bb = bbs[i];
2403 if (bb->loop_father == loop->inner)
2404 factor = innerloop_iters;
2408 for (si = gsi_start_bb (bb); !gsi_end_p (si); gsi_next (&si))
2410 gimple stmt = gsi_stmt (si);
2411 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
2413 if (!is_gimple_assign (stmt) && !is_gimple_call (stmt))
2416 /* Skip stmts that are not vectorized inside the loop. */
2418 && !STMT_VINFO_RELEVANT_P (stmt_info)
2419 && (!STMT_VINFO_LIVE_P (stmt_info)
2420 || !VECTORIZABLE_CYCLE_DEF (STMT_VINFO_DEF_TYPE (stmt_info)))
2421 && !STMT_VINFO_IN_PATTERN_P (stmt_info))
2424 if (STMT_VINFO_DATA_REF (vinfo_for_stmt (stmt)))
2426 if (DR_IS_READ (STMT_VINFO_DATA_REF (vinfo_for_stmt (stmt))))
2427 stmt_cost = vect_get_cost (scalar_load);
2429 stmt_cost = vect_get_cost (scalar_store);
2432 stmt_cost = vect_get_cost (scalar_stmt);
2434 scalar_single_iter_cost += stmt_cost * factor;
2437 return scalar_single_iter_cost;
2440 /* Calculate cost of peeling the loop PEEL_ITERS_PROLOGUE times. */
2442 vect_get_known_peeling_cost (loop_vec_info loop_vinfo, int peel_iters_prologue,
2443 int *peel_iters_epilogue,
2444 int scalar_single_iter_cost)
2446 int peel_guard_costs = 0;
2447 int vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo);
2449 if (!LOOP_VINFO_NITERS_KNOWN_P (loop_vinfo))
2451 *peel_iters_epilogue = vf/2;
2452 if (vect_print_dump_info (REPORT_COST))
2453 fprintf (vect_dump, "cost model: "
2454 "epilogue peel iters set to vf/2 because "
2455 "loop iterations are unknown .");
2457 /* If peeled iterations are known but number of scalar loop
2458 iterations are unknown, count a taken branch per peeled loop. */
2459 peel_guard_costs = 2 * vect_get_cost (cond_branch_taken);
2463 int niters = LOOP_VINFO_INT_NITERS (loop_vinfo);
2464 peel_iters_prologue = niters < peel_iters_prologue ?
2465 niters : peel_iters_prologue;
2466 *peel_iters_epilogue = (niters - peel_iters_prologue) % vf;
2467 /* If we need to peel for gaps, but no peeling is required, we have to
2468 peel VF iterations. */
2469 if (LOOP_VINFO_PEELING_FOR_GAPS (loop_vinfo) && !*peel_iters_epilogue)
2470 *peel_iters_epilogue = vf;
2473 return (peel_iters_prologue * scalar_single_iter_cost)
2474 + (*peel_iters_epilogue * scalar_single_iter_cost)
2478 /* Function vect_estimate_min_profitable_iters
2480 Return the number of iterations required for the vector version of the
2481 loop to be profitable relative to the cost of the scalar version of the
2484 TODO: Take profile info into account before making vectorization
2485 decisions, if available. */
2488 vect_estimate_min_profitable_iters (loop_vec_info loop_vinfo)
2491 int min_profitable_iters;
2492 int peel_iters_prologue;
2493 int peel_iters_epilogue;
2494 int vec_inside_cost = 0;
2495 int vec_outside_cost = 0;
2496 int scalar_single_iter_cost = 0;
2497 int scalar_outside_cost = 0;
2498 int vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo);
2499 struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
2500 basic_block *bbs = LOOP_VINFO_BBS (loop_vinfo);
2501 int nbbs = loop->num_nodes;
2502 int npeel = LOOP_PEELING_FOR_ALIGNMENT (loop_vinfo);
2503 int peel_guard_costs = 0;
2504 int innerloop_iters = 0, factor;
2505 VEC (slp_instance, heap) *slp_instances;
2506 slp_instance instance;
2508 /* Cost model disabled. */
2509 if (!flag_vect_cost_model)
2511 if (vect_print_dump_info (REPORT_COST))
2512 fprintf (vect_dump, "cost model disabled.");
2516 /* Requires loop versioning tests to handle misalignment. */
2517 if (LOOP_REQUIRES_VERSIONING_FOR_ALIGNMENT (loop_vinfo))
2519 /* FIXME: Make cost depend on complexity of individual check. */
2521 VEC_length (gimple, LOOP_VINFO_MAY_MISALIGN_STMTS (loop_vinfo));
2522 if (vect_print_dump_info (REPORT_COST))
2523 fprintf (vect_dump, "cost model: Adding cost of checks for loop "
2524 "versioning to treat misalignment.\n");
2527 /* Requires loop versioning with alias checks. */
2528 if (LOOP_REQUIRES_VERSIONING_FOR_ALIAS (loop_vinfo))
2530 /* FIXME: Make cost depend on complexity of individual check. */
2532 VEC_length (ddr_p, LOOP_VINFO_MAY_ALIAS_DDRS (loop_vinfo));
2533 if (vect_print_dump_info (REPORT_COST))
2534 fprintf (vect_dump, "cost model: Adding cost of checks for loop "
2535 "versioning aliasing.\n");
2538 if (LOOP_REQUIRES_VERSIONING_FOR_ALIGNMENT (loop_vinfo)
2539 || LOOP_REQUIRES_VERSIONING_FOR_ALIAS (loop_vinfo))
2540 vec_outside_cost += vect_get_cost (cond_branch_taken);
2542 /* Count statements in scalar loop. Using this as scalar cost for a single
2545 TODO: Add outer loop support.
2547 TODO: Consider assigning different costs to different scalar
2552 innerloop_iters = 50; /* FIXME */
2554 for (i = 0; i < nbbs; i++)
2556 gimple_stmt_iterator si;
2557 basic_block bb = bbs[i];
2559 if (bb->loop_father == loop->inner)
2560 factor = innerloop_iters;
2564 for (si = gsi_start_bb (bb); !gsi_end_p (si); gsi_next (&si))
2566 gimple stmt = gsi_stmt (si);
2567 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
2569 if (STMT_VINFO_IN_PATTERN_P (stmt_info))
2571 stmt = STMT_VINFO_RELATED_STMT (stmt_info);
2572 stmt_info = vinfo_for_stmt (stmt);
2575 /* Skip stmts that are not vectorized inside the loop. */
2576 if (!STMT_VINFO_RELEVANT_P (stmt_info)
2577 && (!STMT_VINFO_LIVE_P (stmt_info)
2578 || !VECTORIZABLE_CYCLE_DEF (STMT_VINFO_DEF_TYPE (stmt_info))))
2581 vec_inside_cost += STMT_VINFO_INSIDE_OF_LOOP_COST (stmt_info) * factor;
2582 /* FIXME: for stmts in the inner-loop in outer-loop vectorization,
2583 some of the "outside" costs are generated inside the outer-loop. */
2584 vec_outside_cost += STMT_VINFO_OUTSIDE_OF_LOOP_COST (stmt_info);
2585 if (is_pattern_stmt_p (stmt_info)
2586 && STMT_VINFO_PATTERN_DEF_SEQ (stmt_info))
2588 gimple_stmt_iterator gsi;
2590 for (gsi = gsi_start (STMT_VINFO_PATTERN_DEF_SEQ (stmt_info));
2591 !gsi_end_p (gsi); gsi_next (&gsi))
2593 gimple pattern_def_stmt = gsi_stmt (gsi);
2594 stmt_vec_info pattern_def_stmt_info
2595 = vinfo_for_stmt (pattern_def_stmt);
2596 if (STMT_VINFO_RELEVANT_P (pattern_def_stmt_info)
2597 || STMT_VINFO_LIVE_P (pattern_def_stmt_info))
2600 += STMT_VINFO_INSIDE_OF_LOOP_COST
2601 (pattern_def_stmt_info) * factor;
2603 += STMT_VINFO_OUTSIDE_OF_LOOP_COST
2604 (pattern_def_stmt_info);
2611 scalar_single_iter_cost = vect_get_single_scalar_iteraion_cost (loop_vinfo);
2613 /* Add additional cost for the peeled instructions in prologue and epilogue
2616 FORNOW: If we don't know the value of peel_iters for prologue or epilogue
2617 at compile-time - we assume it's vf/2 (the worst would be vf-1).
2619 TODO: Build an expression that represents peel_iters for prologue and
2620 epilogue to be used in a run-time test. */
2624 peel_iters_prologue = vf/2;
2625 if (vect_print_dump_info (REPORT_COST))
2626 fprintf (vect_dump, "cost model: "
2627 "prologue peel iters set to vf/2.");
2629 /* If peeling for alignment is unknown, loop bound of main loop becomes
2631 peel_iters_epilogue = vf/2;
2632 if (vect_print_dump_info (REPORT_COST))
2633 fprintf (vect_dump, "cost model: "
2634 "epilogue peel iters set to vf/2 because "
2635 "peeling for alignment is unknown .");
2637 /* If peeled iterations are unknown, count a taken branch and a not taken
2638 branch per peeled loop. Even if scalar loop iterations are known,
2639 vector iterations are not known since peeled prologue iterations are
2640 not known. Hence guards remain the same. */
2641 peel_guard_costs += 2 * (vect_get_cost (cond_branch_taken)
2642 + vect_get_cost (cond_branch_not_taken));
2643 vec_outside_cost += (peel_iters_prologue * scalar_single_iter_cost)
2644 + (peel_iters_epilogue * scalar_single_iter_cost)
2649 peel_iters_prologue = npeel;
2650 vec_outside_cost += vect_get_known_peeling_cost (loop_vinfo,
2651 peel_iters_prologue, &peel_iters_epilogue,
2652 scalar_single_iter_cost);
2655 /* FORNOW: The scalar outside cost is incremented in one of the
2658 1. The vectorizer checks for alignment and aliasing and generates
2659 a condition that allows dynamic vectorization. A cost model
2660 check is ANDED with the versioning condition. Hence scalar code
2661 path now has the added cost of the versioning check.
2663 if (cost > th & versioning_check)
2666 Hence run-time scalar is incremented by not-taken branch cost.
2668 2. The vectorizer then checks if a prologue is required. If the
2669 cost model check was not done before during versioning, it has to
2670 be done before the prologue check.
2673 prologue = scalar_iters
2678 if (prologue == num_iters)
2681 Hence the run-time scalar cost is incremented by a taken branch,
2682 plus a not-taken branch, plus a taken branch cost.
2684 3. The vectorizer then checks if an epilogue is required. If the
2685 cost model check was not done before during prologue check, it
2686 has to be done with the epilogue check.
2692 if (prologue == num_iters)
2695 if ((cost <= th) | (scalar_iters-prologue-epilogue == 0))
2698 Hence the run-time scalar cost should be incremented by 2 taken
2701 TODO: The back end may reorder the BBS's differently and reverse
2702 conditions/branch directions. Change the estimates below to
2703 something more reasonable. */
2705 /* If the number of iterations is known and we do not do versioning, we can
2706 decide whether to vectorize at compile time. Hence the scalar version
2707 do not carry cost model guard costs. */
2708 if (!LOOP_VINFO_NITERS_KNOWN_P (loop_vinfo)
2709 || LOOP_REQUIRES_VERSIONING_FOR_ALIGNMENT (loop_vinfo)
2710 || LOOP_REQUIRES_VERSIONING_FOR_ALIAS (loop_vinfo))
2712 /* Cost model check occurs at versioning. */
2713 if (LOOP_REQUIRES_VERSIONING_FOR_ALIGNMENT (loop_vinfo)
2714 || LOOP_REQUIRES_VERSIONING_FOR_ALIAS (loop_vinfo))
2715 scalar_outside_cost += vect_get_cost (cond_branch_not_taken);
2718 /* Cost model check occurs at prologue generation. */
2719 if (LOOP_PEELING_FOR_ALIGNMENT (loop_vinfo) < 0)
2720 scalar_outside_cost += 2 * vect_get_cost (cond_branch_taken)
2721 + vect_get_cost (cond_branch_not_taken);
2722 /* Cost model check occurs at epilogue generation. */
2724 scalar_outside_cost += 2 * vect_get_cost (cond_branch_taken);
2728 /* Add SLP costs. */
2729 slp_instances = LOOP_VINFO_SLP_INSTANCES (loop_vinfo);
2730 FOR_EACH_VEC_ELT (slp_instance, slp_instances, i, instance)
2732 vec_outside_cost += SLP_INSTANCE_OUTSIDE_OF_LOOP_COST (instance);
2733 vec_inside_cost += SLP_INSTANCE_INSIDE_OF_LOOP_COST (instance);
2736 /* Calculate number of iterations required to make the vector version
2737 profitable, relative to the loop bodies only. The following condition
2739 SIC * niters + SOC > VIC * ((niters-PL_ITERS-EP_ITERS)/VF) + VOC
2741 SIC = scalar iteration cost, VIC = vector iteration cost,
2742 VOC = vector outside cost, VF = vectorization factor,
2743 PL_ITERS = prologue iterations, EP_ITERS= epilogue iterations
2744 SOC = scalar outside cost for run time cost model check. */
2746 if ((scalar_single_iter_cost * vf) > vec_inside_cost)
2748 if (vec_outside_cost <= 0)
2749 min_profitable_iters = 1;
2752 min_profitable_iters = ((vec_outside_cost - scalar_outside_cost) * vf
2753 - vec_inside_cost * peel_iters_prologue
2754 - vec_inside_cost * peel_iters_epilogue)
2755 / ((scalar_single_iter_cost * vf)
2758 if ((scalar_single_iter_cost * vf * min_profitable_iters)
2759 <= ((vec_inside_cost * min_profitable_iters)
2760 + ((vec_outside_cost - scalar_outside_cost) * vf)))
2761 min_profitable_iters++;
2764 /* vector version will never be profitable. */
2767 if (vect_print_dump_info (REPORT_COST))
2768 fprintf (vect_dump, "cost model: the vector iteration cost = %d "
2769 "divided by the scalar iteration cost = %d "
2770 "is greater or equal to the vectorization factor = %d.",
2771 vec_inside_cost, scalar_single_iter_cost, vf);
2775 if (vect_print_dump_info (REPORT_COST))
2777 fprintf (vect_dump, "Cost model analysis: \n");
2778 fprintf (vect_dump, " Vector inside of loop cost: %d\n",
2780 fprintf (vect_dump, " Vector outside of loop cost: %d\n",
2782 fprintf (vect_dump, " Scalar iteration cost: %d\n",
2783 scalar_single_iter_cost);
2784 fprintf (vect_dump, " Scalar outside cost: %d\n", scalar_outside_cost);
2785 fprintf (vect_dump, " prologue iterations: %d\n",
2786 peel_iters_prologue);
2787 fprintf (vect_dump, " epilogue iterations: %d\n",
2788 peel_iters_epilogue);
2789 fprintf (vect_dump, " Calculated minimum iters for profitability: %d\n",
2790 min_profitable_iters);
2793 min_profitable_iters =
2794 min_profitable_iters < vf ? vf : min_profitable_iters;
2796 /* Because the condition we create is:
2797 if (niters <= min_profitable_iters)
2798 then skip the vectorized loop. */
2799 min_profitable_iters--;
2801 if (vect_print_dump_info (REPORT_COST))
2802 fprintf (vect_dump, " Profitability threshold = %d\n",
2803 min_profitable_iters);
2805 return min_profitable_iters;
2809 /* TODO: Close dependency between vect_model_*_cost and vectorizable_*
2810 functions. Design better to avoid maintenance issues. */
2812 /* Function vect_model_reduction_cost.
2814 Models cost for a reduction operation, including the vector ops
2815 generated within the strip-mine loop, the initial definition before
2816 the loop, and the epilogue code that must be generated. */
2819 vect_model_reduction_cost (stmt_vec_info stmt_info, enum tree_code reduc_code,
2823 enum tree_code code;
2826 gimple stmt, orig_stmt;
2828 enum machine_mode mode;
2829 loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info);
2830 struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
2833 /* Cost of reduction op inside loop. */
2834 STMT_VINFO_INSIDE_OF_LOOP_COST (stmt_info)
2835 += ncopies * vect_get_cost (vector_stmt);
2837 stmt = STMT_VINFO_STMT (stmt_info);
2839 switch (get_gimple_rhs_class (gimple_assign_rhs_code (stmt)))
2841 case GIMPLE_SINGLE_RHS:
2842 gcc_assert (TREE_OPERAND_LENGTH (gimple_assign_rhs1 (stmt)) == ternary_op);
2843 reduction_op = TREE_OPERAND (gimple_assign_rhs1 (stmt), 2);
2845 case GIMPLE_UNARY_RHS:
2846 reduction_op = gimple_assign_rhs1 (stmt);
2848 case GIMPLE_BINARY_RHS:
2849 reduction_op = gimple_assign_rhs2 (stmt);
2851 case GIMPLE_TERNARY_RHS:
2852 reduction_op = gimple_assign_rhs3 (stmt);
2858 vectype = get_vectype_for_scalar_type (TREE_TYPE (reduction_op));
2861 if (vect_print_dump_info (REPORT_COST))
2863 fprintf (vect_dump, "unsupported data-type ");
2864 print_generic_expr (vect_dump, TREE_TYPE (reduction_op), TDF_SLIM);
2869 mode = TYPE_MODE (vectype);
2870 orig_stmt = STMT_VINFO_RELATED_STMT (stmt_info);
2873 orig_stmt = STMT_VINFO_STMT (stmt_info);
2875 code = gimple_assign_rhs_code (orig_stmt);
2877 /* Add in cost for initial definition. */
2878 outer_cost += vect_get_cost (scalar_to_vec);
2880 /* Determine cost of epilogue code.
2882 We have a reduction operator that will reduce the vector in one statement.
2883 Also requires scalar extract. */
2885 if (!nested_in_vect_loop_p (loop, orig_stmt))
2887 if (reduc_code != ERROR_MARK)
2888 outer_cost += vect_get_cost (vector_stmt)
2889 + vect_get_cost (vec_to_scalar);
2892 int vec_size_in_bits = tree_low_cst (TYPE_SIZE (vectype), 1);
2894 TYPE_SIZE (TREE_TYPE (gimple_assign_lhs (orig_stmt)));
2895 int element_bitsize = tree_low_cst (bitsize, 1);
2896 int nelements = vec_size_in_bits / element_bitsize;
2898 optab = optab_for_tree_code (code, vectype, optab_default);
2900 /* We have a whole vector shift available. */
2901 if (VECTOR_MODE_P (mode)
2902 && optab_handler (optab, mode) != CODE_FOR_nothing
2903 && optab_handler (vec_shr_optab, mode) != CODE_FOR_nothing)
2904 /* Final reduction via vector shifts and the reduction operator. Also
2905 requires scalar extract. */
2906 outer_cost += ((exact_log2(nelements) * 2)
2907 * vect_get_cost (vector_stmt)
2908 + vect_get_cost (vec_to_scalar));
2910 /* Use extracts and reduction op for final reduction. For N elements,
2911 we have N extracts and N-1 reduction ops. */
2912 outer_cost += ((nelements + nelements - 1)
2913 * vect_get_cost (vector_stmt));
2917 STMT_VINFO_OUTSIDE_OF_LOOP_COST (stmt_info) = outer_cost;
2919 if (vect_print_dump_info (REPORT_COST))
2920 fprintf (vect_dump, "vect_model_reduction_cost: inside_cost = %d, "
2921 "outside_cost = %d .", STMT_VINFO_INSIDE_OF_LOOP_COST (stmt_info),
2922 STMT_VINFO_OUTSIDE_OF_LOOP_COST (stmt_info));
2928 /* Function vect_model_induction_cost.
2930 Models cost for induction operations. */
2933 vect_model_induction_cost (stmt_vec_info stmt_info, int ncopies)
2935 /* loop cost for vec_loop. */
2936 STMT_VINFO_INSIDE_OF_LOOP_COST (stmt_info)
2937 = ncopies * vect_get_cost (vector_stmt);
2938 /* prologue cost for vec_init and vec_step. */
2939 STMT_VINFO_OUTSIDE_OF_LOOP_COST (stmt_info)
2940 = 2 * vect_get_cost (scalar_to_vec);
2942 if (vect_print_dump_info (REPORT_COST))
2943 fprintf (vect_dump, "vect_model_induction_cost: inside_cost = %d, "
2944 "outside_cost = %d .", STMT_VINFO_INSIDE_OF_LOOP_COST (stmt_info),
2945 STMT_VINFO_OUTSIDE_OF_LOOP_COST (stmt_info));
2949 /* Function get_initial_def_for_induction
2952 STMT - a stmt that performs an induction operation in the loop.
2953 IV_PHI - the initial value of the induction variable
2956 Return a vector variable, initialized with the first VF values of
2957 the induction variable. E.g., for an iv with IV_PHI='X' and
2958 evolution S, for a vector of 4 units, we want to return:
2959 [X, X + S, X + 2*S, X + 3*S]. */
2962 get_initial_def_for_induction (gimple iv_phi)
2964 stmt_vec_info stmt_vinfo = vinfo_for_stmt (iv_phi);
2965 loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_vinfo);
2966 struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
2970 edge pe = loop_preheader_edge (loop);
2971 struct loop *iv_loop;
2973 tree vec, vec_init, vec_step, t;
2977 gimple init_stmt, induction_phi, new_stmt;
2978 tree induc_def, vec_def, vec_dest;
2979 tree init_expr, step_expr;
2980 int vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo);
2985 stmt_vec_info phi_info = vinfo_for_stmt (iv_phi);
2986 bool nested_in_vect_loop = false;
2987 gimple_seq stmts = NULL;
2988 imm_use_iterator imm_iter;
2989 use_operand_p use_p;
2993 gimple_stmt_iterator si;
2994 basic_block bb = gimple_bb (iv_phi);
2998 /* Is phi in an inner-loop, while vectorizing an enclosing outer-loop? */
2999 if (nested_in_vect_loop_p (loop, iv_phi))
3001 nested_in_vect_loop = true;
3002 iv_loop = loop->inner;
3006 gcc_assert (iv_loop == (gimple_bb (iv_phi))->loop_father);
3008 latch_e = loop_latch_edge (iv_loop);
3009 loop_arg = PHI_ARG_DEF_FROM_EDGE (iv_phi, latch_e);
3011 access_fn = analyze_scalar_evolution (iv_loop, PHI_RESULT (iv_phi));
3012 gcc_assert (access_fn);
3013 STRIP_NOPS (access_fn);
3014 ok = vect_is_simple_iv_evolution (iv_loop->num, access_fn,
3015 &init_expr, &step_expr);
3017 pe = loop_preheader_edge (iv_loop);
3019 scalar_type = TREE_TYPE (init_expr);
3020 vectype = get_vectype_for_scalar_type (scalar_type);
3021 resvectype = get_vectype_for_scalar_type (TREE_TYPE (PHI_RESULT (iv_phi)));
3022 gcc_assert (vectype);
3023 nunits = TYPE_VECTOR_SUBPARTS (vectype);
3024 ncopies = vf / nunits;
3026 gcc_assert (phi_info);
3027 gcc_assert (ncopies >= 1);
3029 /* Find the first insertion point in the BB. */
3030 si = gsi_after_labels (bb);
3032 /* Create the vector that holds the initial_value of the induction. */
3033 if (nested_in_vect_loop)
3035 /* iv_loop is nested in the loop to be vectorized. init_expr had already
3036 been created during vectorization of previous stmts. We obtain it
3037 from the STMT_VINFO_VEC_STMT of the defining stmt. */
3038 tree iv_def = PHI_ARG_DEF_FROM_EDGE (iv_phi,
3039 loop_preheader_edge (iv_loop));
3040 vec_init = vect_get_vec_def_for_operand (iv_def, iv_phi, NULL);
3044 VEC(constructor_elt,gc) *v;
3046 /* iv_loop is the loop to be vectorized. Create:
3047 vec_init = [X, X+S, X+2*S, X+3*S] (S = step_expr, X = init_expr) */
3048 new_var = vect_get_new_vect_var (scalar_type, vect_scalar_var, "var_");
3049 add_referenced_var (new_var);
3051 new_name = force_gimple_operand (init_expr, &stmts, false, new_var);
3054 new_bb = gsi_insert_seq_on_edge_immediate (pe, stmts);
3055 gcc_assert (!new_bb);
3058 v = VEC_alloc (constructor_elt, gc, nunits);
3059 CONSTRUCTOR_APPEND_ELT (v, NULL_TREE, new_name);
3060 for (i = 1; i < nunits; i++)
3062 /* Create: new_name_i = new_name + step_expr */
3063 enum tree_code code = POINTER_TYPE_P (scalar_type)
3064 ? POINTER_PLUS_EXPR : PLUS_EXPR;
3065 init_stmt = gimple_build_assign_with_ops (code, new_var,
3066 new_name, step_expr);
3067 new_name = make_ssa_name (new_var, init_stmt);
3068 gimple_assign_set_lhs (init_stmt, new_name);
3070 new_bb = gsi_insert_on_edge_immediate (pe, init_stmt);
3071 gcc_assert (!new_bb);
3073 if (vect_print_dump_info (REPORT_DETAILS))
3075 fprintf (vect_dump, "created new init_stmt: ");
3076 print_gimple_stmt (vect_dump, init_stmt, 0, TDF_SLIM);
3078 CONSTRUCTOR_APPEND_ELT (v, NULL_TREE, new_name);
3080 /* Create a vector from [new_name_0, new_name_1, ..., new_name_nunits-1] */
3081 vec = build_constructor (vectype, v);
3082 vec_init = vect_init_vector (iv_phi, vec, vectype, NULL);
3086 /* Create the vector that holds the step of the induction. */
3087 if (nested_in_vect_loop)
3088 /* iv_loop is nested in the loop to be vectorized. Generate:
3089 vec_step = [S, S, S, S] */
3090 new_name = step_expr;
3093 /* iv_loop is the loop to be vectorized. Generate:
3094 vec_step = [VF*S, VF*S, VF*S, VF*S] */
3095 expr = build_int_cst (TREE_TYPE (step_expr), vf);
3096 new_name = fold_build2 (MULT_EXPR, TREE_TYPE (step_expr),
3100 t = unshare_expr (new_name);
3101 gcc_assert (CONSTANT_CLASS_P (new_name));
3102 stepvectype = get_vectype_for_scalar_type (TREE_TYPE (new_name));
3103 gcc_assert (stepvectype);
3104 vec = build_vector_from_val (stepvectype, t);
3105 vec_step = vect_init_vector (iv_phi, vec, stepvectype, NULL);
3108 /* Create the following def-use cycle:
3113 vec_iv = PHI <vec_init, vec_loop>
3117 vec_loop = vec_iv + vec_step; */
3119 /* Create the induction-phi that defines the induction-operand. */
3120 vec_dest = vect_get_new_vect_var (vectype, vect_simple_var, "vec_iv_");
3121 add_referenced_var (vec_dest);
3122 induction_phi = create_phi_node (vec_dest, iv_loop->header);
3123 set_vinfo_for_stmt (induction_phi,
3124 new_stmt_vec_info (induction_phi, loop_vinfo, NULL));
3125 induc_def = PHI_RESULT (induction_phi);
3127 /* Create the iv update inside the loop */
3128 new_stmt = gimple_build_assign_with_ops (PLUS_EXPR, vec_dest,
3129 induc_def, vec_step);
3130 vec_def = make_ssa_name (vec_dest, new_stmt);
3131 gimple_assign_set_lhs (new_stmt, vec_def);
3132 gsi_insert_before (&si, new_stmt, GSI_SAME_STMT);
3133 set_vinfo_for_stmt (new_stmt, new_stmt_vec_info (new_stmt, loop_vinfo,
3136 /* Set the arguments of the phi node: */
3137 add_phi_arg (induction_phi, vec_init, pe, UNKNOWN_LOCATION);
3138 add_phi_arg (induction_phi, vec_def, loop_latch_edge (iv_loop),
3142 /* In case that vectorization factor (VF) is bigger than the number
3143 of elements that we can fit in a vectype (nunits), we have to generate
3144 more than one vector stmt - i.e - we need to "unroll" the
3145 vector stmt by a factor VF/nunits. For more details see documentation
3146 in vectorizable_operation. */
3150 stmt_vec_info prev_stmt_vinfo;
3151 /* FORNOW. This restriction should be relaxed. */
3152 gcc_assert (!nested_in_vect_loop);
3154 /* Create the vector that holds the step of the induction. */
3155 expr = build_int_cst (TREE_TYPE (step_expr), nunits);
3156 new_name = fold_build2 (MULT_EXPR, TREE_TYPE (step_expr),
3158 t = unshare_expr (new_name);
3159 gcc_assert (CONSTANT_CLASS_P (new_name));
3160 vec = build_vector_from_val (stepvectype, t);
3161 vec_step = vect_init_vector (iv_phi, vec, stepvectype, NULL);
3163 vec_def = induc_def;
3164 prev_stmt_vinfo = vinfo_for_stmt (induction_phi);
3165 for (i = 1; i < ncopies; i++)
3167 /* vec_i = vec_prev + vec_step */
3168 new_stmt = gimple_build_assign_with_ops (PLUS_EXPR, vec_dest,
3170 vec_def = make_ssa_name (vec_dest, new_stmt);
3171 gimple_assign_set_lhs (new_stmt, vec_def);
3173 gsi_insert_before (&si, new_stmt, GSI_SAME_STMT);
3174 if (!useless_type_conversion_p (resvectype, vectype))
3176 new_stmt = gimple_build_assign_with_ops
3178 vect_get_new_vect_var (resvectype, vect_simple_var,
3180 build1 (VIEW_CONVERT_EXPR, resvectype,
3181 gimple_assign_lhs (new_stmt)), NULL_TREE);
3182 gimple_assign_set_lhs (new_stmt,
3184 (gimple_assign_lhs (new_stmt), new_stmt));
3185 gsi_insert_before (&si, new_stmt, GSI_SAME_STMT);
3187 set_vinfo_for_stmt (new_stmt,
3188 new_stmt_vec_info (new_stmt, loop_vinfo, NULL));
3189 STMT_VINFO_RELATED_STMT (prev_stmt_vinfo) = new_stmt;
3190 prev_stmt_vinfo = vinfo_for_stmt (new_stmt);
3194 if (nested_in_vect_loop)
3196 /* Find the loop-closed exit-phi of the induction, and record
3197 the final vector of induction results: */
3199 FOR_EACH_IMM_USE_FAST (use_p, imm_iter, loop_arg)
3201 if (!flow_bb_inside_loop_p (iv_loop, gimple_bb (USE_STMT (use_p))))
3203 exit_phi = USE_STMT (use_p);
3209 stmt_vec_info stmt_vinfo = vinfo_for_stmt (exit_phi);
3210 /* FORNOW. Currently not supporting the case that an inner-loop induction
3211 is not used in the outer-loop (i.e. only outside the outer-loop). */
3212 gcc_assert (STMT_VINFO_RELEVANT_P (stmt_vinfo)
3213 && !STMT_VINFO_LIVE_P (stmt_vinfo));
3215 STMT_VINFO_VEC_STMT (stmt_vinfo) = new_stmt;
3216 if (vect_print_dump_info (REPORT_DETAILS))
3218 fprintf (vect_dump, "vector of inductions after inner-loop:");
3219 print_gimple_stmt (vect_dump, new_stmt, 0, TDF_SLIM);
3225 if (vect_print_dump_info (REPORT_DETAILS))
3227 fprintf (vect_dump, "transform induction: created def-use cycle: ");
3228 print_gimple_stmt (vect_dump, induction_phi, 0, TDF_SLIM);
3229 fprintf (vect_dump, "\n");
3230 print_gimple_stmt (vect_dump, SSA_NAME_DEF_STMT (vec_def), 0, TDF_SLIM);
3233 STMT_VINFO_VEC_STMT (phi_info) = induction_phi;
3234 if (!useless_type_conversion_p (resvectype, vectype))
3236 new_stmt = gimple_build_assign_with_ops
3238 vect_get_new_vect_var (resvectype, vect_simple_var, "vec_iv_"),
3239 build1 (VIEW_CONVERT_EXPR, resvectype, induc_def), NULL_TREE);
3240 induc_def = make_ssa_name (gimple_assign_lhs (new_stmt), new_stmt);
3241 gimple_assign_set_lhs (new_stmt, induc_def);
3242 si = gsi_start_bb (bb);
3243 gsi_insert_before (&si, new_stmt, GSI_SAME_STMT);
3244 set_vinfo_for_stmt (new_stmt,
3245 new_stmt_vec_info (new_stmt, loop_vinfo, NULL));
3246 STMT_VINFO_RELATED_STMT (vinfo_for_stmt (new_stmt))
3247 = STMT_VINFO_RELATED_STMT (vinfo_for_stmt (induction_phi));
3254 /* Function get_initial_def_for_reduction
3257 STMT - a stmt that performs a reduction operation in the loop.
3258 INIT_VAL - the initial value of the reduction variable
3261 ADJUSTMENT_DEF - a tree that holds a value to be added to the final result
3262 of the reduction (used for adjusting the epilog - see below).
3263 Return a vector variable, initialized according to the operation that STMT
3264 performs. This vector will be used as the initial value of the
3265 vector of partial results.
3267 Option1 (adjust in epilog): Initialize the vector as follows:
3268 add/bit or/xor: [0,0,...,0,0]
3269 mult/bit and: [1,1,...,1,1]
3270 min/max/cond_expr: [init_val,init_val,..,init_val,init_val]
3271 and when necessary (e.g. add/mult case) let the caller know
3272 that it needs to adjust the result by init_val.
3274 Option2: Initialize the vector as follows:
3275 add/bit or/xor: [init_val,0,0,...,0]
3276 mult/bit and: [init_val,1,1,...,1]
3277 min/max/cond_expr: [init_val,init_val,...,init_val]
3278 and no adjustments are needed.
3280 For example, for the following code:
3286 STMT is 's = s + a[i]', and the reduction variable is 's'.
3287 For a vector of 4 units, we want to return either [0,0,0,init_val],
3288 or [0,0,0,0] and let the caller know that it needs to adjust
3289 the result at the end by 'init_val'.
3291 FORNOW, we are using the 'adjust in epilog' scheme, because this way the
3292 initialization vector is simpler (same element in all entries), if
3293 ADJUSTMENT_DEF is not NULL, and Option2 otherwise.
3295 A cost model should help decide between these two schemes. */
3298 get_initial_def_for_reduction (gimple stmt, tree init_val,
3299 tree *adjustment_def)
3301 stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt);
3302 loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_vinfo);
3303 struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
3304 tree scalar_type = TREE_TYPE (init_val);
3305 tree vectype = get_vectype_for_scalar_type (scalar_type);
3307 enum tree_code code = gimple_assign_rhs_code (stmt);
3312 bool nested_in_vect_loop = false;
3314 REAL_VALUE_TYPE real_init_val = dconst0;
3315 int int_init_val = 0;
3316 gimple def_stmt = NULL;
3318 gcc_assert (vectype);
3319 nunits = TYPE_VECTOR_SUBPARTS (vectype);
3321 gcc_assert (POINTER_TYPE_P (scalar_type) || INTEGRAL_TYPE_P (scalar_type)
3322 || SCALAR_FLOAT_TYPE_P (scalar_type));
3324 if (nested_in_vect_loop_p (loop, stmt))
3325 nested_in_vect_loop = true;
3327 gcc_assert (loop == (gimple_bb (stmt))->loop_father);
3329 /* In case of double reduction we only create a vector variable to be put
3330 in the reduction phi node. The actual statement creation is done in
3331 vect_create_epilog_for_reduction. */
3332 if (adjustment_def && nested_in_vect_loop
3333 && TREE_CODE (init_val) == SSA_NAME
3334 && (def_stmt = SSA_NAME_DEF_STMT (init_val))
3335 && gimple_code (def_stmt) == GIMPLE_PHI
3336 && flow_bb_inside_loop_p (loop, gimple_bb (def_stmt))
3337 && vinfo_for_stmt (def_stmt)
3338 && STMT_VINFO_DEF_TYPE (vinfo_for_stmt (def_stmt))
3339 == vect_double_reduction_def)
3341 *adjustment_def = NULL;
3342 return vect_create_destination_var (init_val, vectype);
3345 if (TREE_CONSTANT (init_val))
3347 if (SCALAR_FLOAT_TYPE_P (scalar_type))
3348 init_value = build_real (scalar_type, TREE_REAL_CST (init_val));
3350 init_value = build_int_cst (scalar_type, TREE_INT_CST_LOW (init_val));
3353 init_value = init_val;
3357 case WIDEN_SUM_EXPR:
3365 /* ADJUSMENT_DEF is NULL when called from
3366 vect_create_epilog_for_reduction to vectorize double reduction. */
3369 if (nested_in_vect_loop)
3370 *adjustment_def = vect_get_vec_def_for_operand (init_val, stmt,
3373 *adjustment_def = init_val;
3376 if (code == MULT_EXPR)
3378 real_init_val = dconst1;
3382 if (code == BIT_AND_EXPR)
3385 if (SCALAR_FLOAT_TYPE_P (scalar_type))
3386 def_for_init = build_real (scalar_type, real_init_val);
3388 def_for_init = build_int_cst (scalar_type, int_init_val);
3390 /* Create a vector of '0' or '1' except the first element. */
3391 elts = XALLOCAVEC (tree, nunits);
3392 for (i = nunits - 2; i >= 0; --i)
3393 elts[i + 1] = def_for_init;
3395 /* Option1: the first element is '0' or '1' as well. */
3398 elts[0] = def_for_init;
3399 init_def = build_vector (vectype, elts);
3403 /* Option2: the first element is INIT_VAL. */
3405 if (TREE_CONSTANT (init_val))
3406 init_def = build_vector (vectype, elts);
3409 VEC(constructor_elt,gc) *v;
3410 v = VEC_alloc (constructor_elt, gc, nunits);
3411 CONSTRUCTOR_APPEND_ELT (v, NULL_TREE, init_val);
3412 for (i = 1; i < nunits; ++i)
3413 CONSTRUCTOR_APPEND_ELT (v, NULL_TREE, elts[i]);
3414 init_def = build_constructor (vectype, v);
3424 *adjustment_def = NULL_TREE;
3425 init_def = vect_get_vec_def_for_operand (init_val, stmt, NULL);
3429 init_def = build_vector_from_val (vectype, init_value);
3440 /* Function vect_create_epilog_for_reduction
3442 Create code at the loop-epilog to finalize the result of a reduction
3445 VECT_DEFS is list of vector of partial results, i.e., the lhs's of vector
3446 reduction statements.
3447 STMT is the scalar reduction stmt that is being vectorized.
3448 NCOPIES is > 1 in case the vectorization factor (VF) is bigger than the
3449 number of elements that we can fit in a vectype (nunits). In this case
3450 we have to generate more than one vector stmt - i.e - we need to "unroll"
3451 the vector stmt by a factor VF/nunits. For more details see documentation
3452 in vectorizable_operation.
3453 REDUC_CODE is the tree-code for the epilog reduction.
3454 REDUCTION_PHIS is a list of the phi-nodes that carry the reduction
3456 REDUC_INDEX is the index of the operand in the right hand side of the
3457 statement that is defined by REDUCTION_PHI.
3458 DOUBLE_REDUC is TRUE if double reduction phi nodes should be handled.
3459 SLP_NODE is an SLP node containing a group of reduction statements. The
3460 first one in this group is STMT.
3463 1. Creates the reduction def-use cycles: sets the arguments for
3465 The loop-entry argument is the vectorized initial-value of the reduction.
3466 The loop-latch argument is taken from VECT_DEFS - the vector of partial
3468 2. "Reduces" each vector of partial results VECT_DEFS into a single result,
3469 by applying the operation specified by REDUC_CODE if available, or by
3470 other means (whole-vector shifts or a scalar loop).
3471 The function also creates a new phi node at the loop exit to preserve
3472 loop-closed form, as illustrated below.
3474 The flow at the entry to this function:
3477 vec_def = phi <null, null> # REDUCTION_PHI
3478 VECT_DEF = vector_stmt # vectorized form of STMT
3479 s_loop = scalar_stmt # (scalar) STMT
3481 s_out0 = phi <s_loop> # (scalar) EXIT_PHI
3485 The above is transformed by this function into:
3488 vec_def = phi <vec_init, VECT_DEF> # REDUCTION_PHI
3489 VECT_DEF = vector_stmt # vectorized form of STMT
3490 s_loop = scalar_stmt # (scalar) STMT
3492 s_out0 = phi <s_loop> # (scalar) EXIT_PHI
3493 v_out1 = phi <VECT_DEF> # NEW_EXIT_PHI
3494 v_out2 = reduce <v_out1>
3495 s_out3 = extract_field <v_out2, 0>
3496 s_out4 = adjust_result <s_out3>
3502 vect_create_epilog_for_reduction (VEC (tree, heap) *vect_defs, gimple stmt,
3503 int ncopies, enum tree_code reduc_code,
3504 VEC (gimple, heap) *reduction_phis,
3505 int reduc_index, bool double_reduc,
3508 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
3509 stmt_vec_info prev_phi_info;
3511 enum machine_mode mode;
3512 loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info);
3513 struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo), *outer_loop = NULL;
3514 basic_block exit_bb;
3517 gimple new_phi = NULL, phi;
3518 gimple_stmt_iterator exit_gsi;
3520 tree new_temp = NULL_TREE, new_dest, new_name, new_scalar_dest;
3521 gimple epilog_stmt = NULL;
3522 enum tree_code code = gimple_assign_rhs_code (stmt);
3524 tree bitsize, bitpos;
3525 tree adjustment_def = NULL;
3526 tree vec_initial_def = NULL;
3527 tree reduction_op, expr, def;
3528 tree orig_name, scalar_result;
3529 imm_use_iterator imm_iter, phi_imm_iter;
3530 use_operand_p use_p, phi_use_p;
3531 bool extract_scalar_result = false;
3532 gimple use_stmt, orig_stmt, reduction_phi = NULL;
3533 bool nested_in_vect_loop = false;
3534 VEC (gimple, heap) *new_phis = NULL;
3535 VEC (gimple, heap) *inner_phis = NULL;
3536 enum vect_def_type dt = vect_unknown_def_type;
3538 VEC (tree, heap) *scalar_results = NULL;
3539 unsigned int group_size = 1, k, ratio;
3540 VEC (tree, heap) *vec_initial_defs = NULL;
3541 VEC (gimple, heap) *phis;
3542 bool slp_reduc = false;
3543 tree new_phi_result;
3544 gimple inner_phi = NULL;
3547 group_size = VEC_length (gimple, SLP_TREE_SCALAR_STMTS (slp_node));
3549 if (nested_in_vect_loop_p (loop, stmt))
3553 nested_in_vect_loop = true;
3554 gcc_assert (!slp_node);
3557 switch (get_gimple_rhs_class (gimple_assign_rhs_code (stmt)))
3559 case GIMPLE_SINGLE_RHS:
3560 gcc_assert (TREE_OPERAND_LENGTH (gimple_assign_rhs1 (stmt))
3562 reduction_op = TREE_OPERAND (gimple_assign_rhs1 (stmt), reduc_index);
3564 case GIMPLE_UNARY_RHS:
3565 reduction_op = gimple_assign_rhs1 (stmt);
3567 case GIMPLE_BINARY_RHS:
3568 reduction_op = reduc_index ?
3569 gimple_assign_rhs2 (stmt) : gimple_assign_rhs1 (stmt);
3571 case GIMPLE_TERNARY_RHS:
3572 reduction_op = gimple_op (stmt, reduc_index + 1);
3578 vectype = get_vectype_for_scalar_type (TREE_TYPE (reduction_op));
3579 gcc_assert (vectype);
3580 mode = TYPE_MODE (vectype);
3582 /* 1. Create the reduction def-use cycle:
3583 Set the arguments of REDUCTION_PHIS, i.e., transform
3586 vec_def = phi <null, null> # REDUCTION_PHI
3587 VECT_DEF = vector_stmt # vectorized form of STMT
3593 vec_def = phi <vec_init, VECT_DEF> # REDUCTION_PHI
3594 VECT_DEF = vector_stmt # vectorized form of STMT
3597 (in case of SLP, do it for all the phis). */
3599 /* Get the loop-entry arguments. */
3601 vect_get_vec_defs (reduction_op, NULL_TREE, stmt, &vec_initial_defs,
3602 NULL, slp_node, reduc_index);
3605 vec_initial_defs = VEC_alloc (tree, heap, 1);
3606 /* For the case of reduction, vect_get_vec_def_for_operand returns
3607 the scalar def before the loop, that defines the initial value
3608 of the reduction variable. */
3609 vec_initial_def = vect_get_vec_def_for_operand (reduction_op, stmt,
3611 VEC_quick_push (tree, vec_initial_defs, vec_initial_def);
3614 /* Set phi nodes arguments. */
3615 FOR_EACH_VEC_ELT (gimple, reduction_phis, i, phi)
3617 tree vec_init_def = VEC_index (tree, vec_initial_defs, i);
3618 tree def = VEC_index (tree, vect_defs, i);
3619 for (j = 0; j < ncopies; j++)
3621 /* Set the loop-entry arg of the reduction-phi. */
3622 add_phi_arg (phi, vec_init_def, loop_preheader_edge (loop),
3625 /* Set the loop-latch arg for the reduction-phi. */
3627 def = vect_get_vec_def_for_stmt_copy (vect_unknown_def_type, def);
3629 add_phi_arg (phi, def, loop_latch_edge (loop), UNKNOWN_LOCATION);
3631 if (vect_print_dump_info (REPORT_DETAILS))
3633 fprintf (vect_dump, "transform reduction: created def-use"
3635 print_gimple_stmt (vect_dump, phi, 0, TDF_SLIM);
3636 fprintf (vect_dump, "\n");
3637 print_gimple_stmt (vect_dump, SSA_NAME_DEF_STMT (def), 0,
3641 phi = STMT_VINFO_RELATED_STMT (vinfo_for_stmt (phi));
3645 VEC_free (tree, heap, vec_initial_defs);
3647 /* 2. Create epilog code.
3648 The reduction epilog code operates across the elements of the vector
3649 of partial results computed by the vectorized loop.
3650 The reduction epilog code consists of:
3652 step 1: compute the scalar result in a vector (v_out2)
3653 step 2: extract the scalar result (s_out3) from the vector (v_out2)
3654 step 3: adjust the scalar result (s_out3) if needed.
3656 Step 1 can be accomplished using one the following three schemes:
3657 (scheme 1) using reduc_code, if available.
3658 (scheme 2) using whole-vector shifts, if available.
3659 (scheme 3) using a scalar loop. In this case steps 1+2 above are
3662 The overall epilog code looks like this:
3664 s_out0 = phi <s_loop> # original EXIT_PHI
3665 v_out1 = phi <VECT_DEF> # NEW_EXIT_PHI
3666 v_out2 = reduce <v_out1> # step 1
3667 s_out3 = extract_field <v_out2, 0> # step 2
3668 s_out4 = adjust_result <s_out3> # step 3
3670 (step 3 is optional, and steps 1 and 2 may be combined).
3671 Lastly, the uses of s_out0 are replaced by s_out4. */
3674 /* 2.1 Create new loop-exit-phis to preserve loop-closed form:
3675 v_out1 = phi <VECT_DEF>
3676 Store them in NEW_PHIS. */
3678 exit_bb = single_exit (loop)->dest;
3679 prev_phi_info = NULL;
3680 new_phis = VEC_alloc (gimple, heap, VEC_length (tree, vect_defs));
3681 FOR_EACH_VEC_ELT (tree, vect_defs, i, def)
3683 for (j = 0; j < ncopies; j++)
3685 phi = create_phi_node (SSA_NAME_VAR (def), exit_bb);
3686 set_vinfo_for_stmt (phi, new_stmt_vec_info (phi, loop_vinfo, NULL));
3688 VEC_quick_push (gimple, new_phis, phi);
3691 def = vect_get_vec_def_for_stmt_copy (dt, def);
3692 STMT_VINFO_RELATED_STMT (prev_phi_info) = phi;
3695 SET_PHI_ARG_DEF (phi, single_exit (loop)->dest_idx, def);
3696 prev_phi_info = vinfo_for_stmt (phi);
3700 /* The epilogue is created for the outer-loop, i.e., for the loop being
3701 vectorized. Create exit phis for the outer loop. */
3705 exit_bb = single_exit (loop)->dest;
3706 inner_phis = VEC_alloc (gimple, heap, VEC_length (tree, vect_defs));
3707 FOR_EACH_VEC_ELT (gimple, new_phis, i, phi)
3709 gimple outer_phi = create_phi_node (SSA_NAME_VAR (PHI_RESULT (phi)),
3711 SET_PHI_ARG_DEF (outer_phi, single_exit (loop)->dest_idx,
3713 set_vinfo_for_stmt (outer_phi, new_stmt_vec_info (outer_phi,
3715 VEC_quick_push (gimple, inner_phis, phi);
3716 VEC_replace (gimple, new_phis, i, outer_phi);
3717 prev_phi_info = vinfo_for_stmt (outer_phi);
3718 while (STMT_VINFO_RELATED_STMT (vinfo_for_stmt (phi)))
3720 phi = STMT_VINFO_RELATED_STMT (vinfo_for_stmt (phi));
3721 outer_phi = create_phi_node (SSA_NAME_VAR (PHI_RESULT (phi)),
3723 SET_PHI_ARG_DEF (outer_phi, single_exit (loop)->dest_idx,
3725 set_vinfo_for_stmt (outer_phi, new_stmt_vec_info (outer_phi,
3727 STMT_VINFO_RELATED_STMT (prev_phi_info) = outer_phi;
3728 prev_phi_info = vinfo_for_stmt (outer_phi);
3733 exit_gsi = gsi_after_labels (exit_bb);
3735 /* 2.2 Get the relevant tree-code to use in the epilog for schemes 2,3
3736 (i.e. when reduc_code is not available) and in the final adjustment
3737 code (if needed). Also get the original scalar reduction variable as
3738 defined in the loop. In case STMT is a "pattern-stmt" (i.e. - it
3739 represents a reduction pattern), the tree-code and scalar-def are
3740 taken from the original stmt that the pattern-stmt (STMT) replaces.
3741 Otherwise (it is a regular reduction) - the tree-code and scalar-def
3742 are taken from STMT. */
3744 orig_stmt = STMT_VINFO_RELATED_STMT (stmt_info);
3747 /* Regular reduction */
3752 /* Reduction pattern */
3753 stmt_vec_info stmt_vinfo = vinfo_for_stmt (orig_stmt);
3754 gcc_assert (STMT_VINFO_IN_PATTERN_P (stmt_vinfo));
3755 gcc_assert (STMT_VINFO_RELATED_STMT (stmt_vinfo) == stmt);
3758 code = gimple_assign_rhs_code (orig_stmt);
3759 /* For MINUS_EXPR the initial vector is [init_val,0,...,0], therefore,
3760 partial results are added and not subtracted. */
3761 if (code == MINUS_EXPR)
3764 scalar_dest = gimple_assign_lhs (orig_stmt);
3765 scalar_type = TREE_TYPE (scalar_dest);
3766 scalar_results = VEC_alloc (tree, heap, group_size);
3767 new_scalar_dest = vect_create_destination_var (scalar_dest, NULL);
3768 bitsize = TYPE_SIZE (scalar_type);
3770 /* In case this is a reduction in an inner-loop while vectorizing an outer
3771 loop - we don't need to extract a single scalar result at the end of the
3772 inner-loop (unless it is double reduction, i.e., the use of reduction is
3773 outside the outer-loop). The final vector of partial results will be used
3774 in the vectorized outer-loop, or reduced to a scalar result at the end of
3776 if (nested_in_vect_loop && !double_reduc)
3777 goto vect_finalize_reduction;
3779 /* SLP reduction without reduction chain, e.g.,
3783 b2 = operation (b1) */
3784 slp_reduc = (slp_node && !GROUP_FIRST_ELEMENT (vinfo_for_stmt (stmt)));
3786 /* In case of reduction chain, e.g.,
3789 a3 = operation (a2),
3791 we may end up with more than one vector result. Here we reduce them to
3793 if (GROUP_FIRST_ELEMENT (vinfo_for_stmt (stmt)))
3795 tree first_vect = PHI_RESULT (VEC_index (gimple, new_phis, 0));
3797 gimple new_vec_stmt = NULL;
3799 vec_dest = vect_create_destination_var (scalar_dest, vectype);
3800 for (k = 1; k < VEC_length (gimple, new_phis); k++)
3802 gimple next_phi = VEC_index (gimple, new_phis, k);
3803 tree second_vect = PHI_RESULT (next_phi);
3805 tmp = build2 (code, vectype, first_vect, second_vect);
3806 new_vec_stmt = gimple_build_assign (vec_dest, tmp);
3807 first_vect = make_ssa_name (vec_dest, new_vec_stmt);
3808 gimple_assign_set_lhs (new_vec_stmt, first_vect);
3809 gsi_insert_before (&exit_gsi, new_vec_stmt, GSI_SAME_STMT);
3812 new_phi_result = first_vect;
3815 VEC_truncate (gimple, new_phis, 0);
3816 VEC_safe_push (gimple, heap, new_phis, new_vec_stmt);
3820 new_phi_result = PHI_RESULT (VEC_index (gimple, new_phis, 0));
3822 /* 2.3 Create the reduction code, using one of the three schemes described
3823 above. In SLP we simply need to extract all the elements from the
3824 vector (without reducing them), so we use scalar shifts. */
3825 if (reduc_code != ERROR_MARK && !slp_reduc)
3829 /*** Case 1: Create:
3830 v_out2 = reduc_expr <v_out1> */
3832 if (vect_print_dump_info (REPORT_DETAILS))
3833 fprintf (vect_dump, "Reduce using direct vector reduction.");
3835 vec_dest = vect_create_destination_var (scalar_dest, vectype);
3836 tmp = build1 (reduc_code, vectype, new_phi_result);
3837 epilog_stmt = gimple_build_assign (vec_dest, tmp);
3838 new_temp = make_ssa_name (vec_dest, epilog_stmt);
3839 gimple_assign_set_lhs (epilog_stmt, new_temp);
3840 gsi_insert_before (&exit_gsi, epilog_stmt, GSI_SAME_STMT);
3842 extract_scalar_result = true;
3846 enum tree_code shift_code = ERROR_MARK;
3847 bool have_whole_vector_shift = true;
3849 int element_bitsize = tree_low_cst (bitsize, 1);
3850 int vec_size_in_bits = tree_low_cst (TYPE_SIZE (vectype), 1);
3853 if (optab_handler (vec_shr_optab, mode) != CODE_FOR_nothing)
3854 shift_code = VEC_RSHIFT_EXPR;
3856 have_whole_vector_shift = false;
3858 /* Regardless of whether we have a whole vector shift, if we're
3859 emulating the operation via tree-vect-generic, we don't want
3860 to use it. Only the first round of the reduction is likely
3861 to still be profitable via emulation. */
3862 /* ??? It might be better to emit a reduction tree code here, so that
3863 tree-vect-generic can expand the first round via bit tricks. */
3864 if (!VECTOR_MODE_P (mode))
3865 have_whole_vector_shift = false;
3868 optab optab = optab_for_tree_code (code, vectype, optab_default);
3869 if (optab_handler (optab, mode) == CODE_FOR_nothing)
3870 have_whole_vector_shift = false;
3873 if (have_whole_vector_shift && !slp_reduc)
3875 /*** Case 2: Create:
3876 for (offset = VS/2; offset >= element_size; offset/=2)
3878 Create: va' = vec_shift <va, offset>
3879 Create: va = vop <va, va'>
3882 if (vect_print_dump_info (REPORT_DETAILS))
3883 fprintf (vect_dump, "Reduce using vector shifts");
3885 vec_dest = vect_create_destination_var (scalar_dest, vectype);
3886 new_temp = new_phi_result;
3887 for (bit_offset = vec_size_in_bits/2;
3888 bit_offset >= element_bitsize;
3891 tree bitpos = size_int (bit_offset);
3893 epilog_stmt = gimple_build_assign_with_ops (shift_code,
3894 vec_dest, new_temp, bitpos);
3895 new_name = make_ssa_name (vec_dest, epilog_stmt);
3896 gimple_assign_set_lhs (epilog_stmt, new_name);
3897 gsi_insert_before (&exit_gsi, epilog_stmt, GSI_SAME_STMT);
3899 epilog_stmt = gimple_build_assign_with_ops (code, vec_dest,
3900 new_name, new_temp);
3901 new_temp = make_ssa_name (vec_dest, epilog_stmt);
3902 gimple_assign_set_lhs (epilog_stmt, new_temp);
3903 gsi_insert_before (&exit_gsi, epilog_stmt, GSI_SAME_STMT);
3906 extract_scalar_result = true;
3912 /*** Case 3: Create:
3913 s = extract_field <v_out2, 0>
3914 for (offset = element_size;
3915 offset < vector_size;
3916 offset += element_size;)
3918 Create: s' = extract_field <v_out2, offset>
3919 Create: s = op <s, s'> // For non SLP cases
3922 if (vect_print_dump_info (REPORT_DETAILS))
3923 fprintf (vect_dump, "Reduce using scalar code. ");
3925 vec_size_in_bits = tree_low_cst (TYPE_SIZE (vectype), 1);
3926 FOR_EACH_VEC_ELT (gimple, new_phis, i, new_phi)
3928 if (gimple_code (new_phi) == GIMPLE_PHI)
3929 vec_temp = PHI_RESULT (new_phi);
3931 vec_temp = gimple_assign_lhs (new_phi);
3932 rhs = build3 (BIT_FIELD_REF, scalar_type, vec_temp, bitsize,
3934 epilog_stmt = gimple_build_assign (new_scalar_dest, rhs);
3935 new_temp = make_ssa_name (new_scalar_dest, epilog_stmt);
3936 gimple_assign_set_lhs (epilog_stmt, new_temp);
3937 gsi_insert_before (&exit_gsi, epilog_stmt, GSI_SAME_STMT);
3939 /* In SLP we don't need to apply reduction operation, so we just
3940 collect s' values in SCALAR_RESULTS. */
3942 VEC_safe_push (tree, heap, scalar_results, new_temp);
3944 for (bit_offset = element_bitsize;
3945 bit_offset < vec_size_in_bits;
3946 bit_offset += element_bitsize)
3948 tree bitpos = bitsize_int (bit_offset);
3949 tree rhs = build3 (BIT_FIELD_REF, scalar_type, vec_temp,
3952 epilog_stmt = gimple_build_assign (new_scalar_dest, rhs);
3953 new_name = make_ssa_name (new_scalar_dest, epilog_stmt);
3954 gimple_assign_set_lhs (epilog_stmt, new_name);
3955 gsi_insert_before (&exit_gsi, epilog_stmt, GSI_SAME_STMT);
3959 /* In SLP we don't need to apply reduction operation, so
3960 we just collect s' values in SCALAR_RESULTS. */
3961 new_temp = new_name;
3962 VEC_safe_push (tree, heap, scalar_results, new_name);
3966 epilog_stmt = gimple_build_assign_with_ops (code,
3967 new_scalar_dest, new_name, new_temp);
3968 new_temp = make_ssa_name (new_scalar_dest, epilog_stmt);
3969 gimple_assign_set_lhs (epilog_stmt, new_temp);
3970 gsi_insert_before (&exit_gsi, epilog_stmt, GSI_SAME_STMT);
3975 /* The only case where we need to reduce scalar results in SLP, is
3976 unrolling. If the size of SCALAR_RESULTS is greater than
3977 GROUP_SIZE, we reduce them combining elements modulo
3981 tree res, first_res, new_res;
3984 /* Reduce multiple scalar results in case of SLP unrolling. */
3985 for (j = group_size; VEC_iterate (tree, scalar_results, j, res);
3988 first_res = VEC_index (tree, scalar_results, j % group_size);
3989 new_stmt = gimple_build_assign_with_ops (code,
3990 new_scalar_dest, first_res, res);
3991 new_res = make_ssa_name (new_scalar_dest, new_stmt);
3992 gimple_assign_set_lhs (new_stmt, new_res);
3993 gsi_insert_before (&exit_gsi, new_stmt, GSI_SAME_STMT);
3994 VEC_replace (tree, scalar_results, j % group_size, new_res);
3998 /* Not SLP - we have one scalar to keep in SCALAR_RESULTS. */
3999 VEC_safe_push (tree, heap, scalar_results, new_temp);
4001 extract_scalar_result = false;
4005 /* 2.4 Extract the final scalar result. Create:
4006 s_out3 = extract_field <v_out2, bitpos> */
4008 if (extract_scalar_result)
4012 if (vect_print_dump_info (REPORT_DETAILS))
4013 fprintf (vect_dump, "extract scalar result");
4015 if (BYTES_BIG_ENDIAN)
4016 bitpos = size_binop (MULT_EXPR,
4017 bitsize_int (TYPE_VECTOR_SUBPARTS (vectype) - 1),
4018 TYPE_SIZE (scalar_type));
4020 bitpos = bitsize_zero_node;
4022 rhs = build3 (BIT_FIELD_REF, scalar_type, new_temp, bitsize, bitpos);
4023 epilog_stmt = gimple_build_assign (new_scalar_dest, rhs);
4024 new_temp = make_ssa_name (new_scalar_dest, epilog_stmt);
4025 gimple_assign_set_lhs (epilog_stmt, new_temp);
4026 gsi_insert_before (&exit_gsi, epilog_stmt, GSI_SAME_STMT);
4027 VEC_safe_push (tree, heap, scalar_results, new_temp);
4030 vect_finalize_reduction:
4035 /* 2.5 Adjust the final result by the initial value of the reduction
4036 variable. (When such adjustment is not needed, then
4037 'adjustment_def' is zero). For example, if code is PLUS we create:
4038 new_temp = loop_exit_def + adjustment_def */
4042 gcc_assert (!slp_reduc);
4043 if (nested_in_vect_loop)
4045 new_phi = VEC_index (gimple, new_phis, 0);
4046 gcc_assert (TREE_CODE (TREE_TYPE (adjustment_def)) == VECTOR_TYPE);
4047 expr = build2 (code, vectype, PHI_RESULT (new_phi), adjustment_def);
4048 new_dest = vect_create_destination_var (scalar_dest, vectype);
4052 new_temp = VEC_index (tree, scalar_results, 0);
4053 gcc_assert (TREE_CODE (TREE_TYPE (adjustment_def)) != VECTOR_TYPE);
4054 expr = build2 (code, scalar_type, new_temp, adjustment_def);
4055 new_dest = vect_create_destination_var (scalar_dest, scalar_type);
4058 epilog_stmt = gimple_build_assign (new_dest, expr);
4059 new_temp = make_ssa_name (new_dest, epilog_stmt);
4060 gimple_assign_set_lhs (epilog_stmt, new_temp);
4061 SSA_NAME_DEF_STMT (new_temp) = epilog_stmt;
4062 gsi_insert_before (&exit_gsi, epilog_stmt, GSI_SAME_STMT);
4063 if (nested_in_vect_loop)
4065 set_vinfo_for_stmt (epilog_stmt,
4066 new_stmt_vec_info (epilog_stmt, loop_vinfo,
4068 STMT_VINFO_RELATED_STMT (vinfo_for_stmt (epilog_stmt)) =
4069 STMT_VINFO_RELATED_STMT (vinfo_for_stmt (new_phi));
4072 VEC_quick_push (tree, scalar_results, new_temp);
4074 VEC_replace (tree, scalar_results, 0, new_temp);
4077 VEC_replace (tree, scalar_results, 0, new_temp);
4079 VEC_replace (gimple, new_phis, 0, epilog_stmt);
4082 /* 2.6 Handle the loop-exit phis. Replace the uses of scalar loop-exit
4083 phis with new adjusted scalar results, i.e., replace use <s_out0>
4088 s_out0 = phi <s_loop> # (scalar) EXIT_PHI
4089 v_out1 = phi <VECT_DEF> # NEW_EXIT_PHI
4090 v_out2 = reduce <v_out1>
4091 s_out3 = extract_field <v_out2, 0>
4092 s_out4 = adjust_result <s_out3>
4099 s_out0 = phi <s_loop> # (scalar) EXIT_PHI
4100 v_out1 = phi <VECT_DEF> # NEW_EXIT_PHI
4101 v_out2 = reduce <v_out1>
4102 s_out3 = extract_field <v_out2, 0>
4103 s_out4 = adjust_result <s_out3>
4108 /* In SLP reduction chain we reduce vector results into one vector if
4109 necessary, hence we set here GROUP_SIZE to 1. SCALAR_DEST is the LHS of
4110 the last stmt in the reduction chain, since we are looking for the loop
4112 if (GROUP_FIRST_ELEMENT (vinfo_for_stmt (stmt)))
4114 scalar_dest = gimple_assign_lhs (VEC_index (gimple,
4115 SLP_TREE_SCALAR_STMTS (slp_node),
4120 /* In SLP we may have several statements in NEW_PHIS and REDUCTION_PHIS (in
4121 case that GROUP_SIZE is greater than vectorization factor). Therefore, we
4122 need to match SCALAR_RESULTS with corresponding statements. The first
4123 (GROUP_SIZE / number of new vector stmts) scalar results correspond to
4124 the first vector stmt, etc.
4125 (RATIO is equal to (GROUP_SIZE / number of new vector stmts)). */
4126 if (group_size > VEC_length (gimple, new_phis))
4128 ratio = group_size / VEC_length (gimple, new_phis);
4129 gcc_assert (!(group_size % VEC_length (gimple, new_phis)));
4134 for (k = 0; k < group_size; k++)
4138 epilog_stmt = VEC_index (gimple, new_phis, k / ratio);
4139 reduction_phi = VEC_index (gimple, reduction_phis, k / ratio);
4141 inner_phi = VEC_index (gimple, inner_phis, k / ratio);
4146 gimple current_stmt = VEC_index (gimple,
4147 SLP_TREE_SCALAR_STMTS (slp_node), k);
4149 orig_stmt = STMT_VINFO_RELATED_STMT (vinfo_for_stmt (current_stmt));
4150 /* SLP statements can't participate in patterns. */
4151 gcc_assert (!orig_stmt);
4152 scalar_dest = gimple_assign_lhs (current_stmt);
4155 phis = VEC_alloc (gimple, heap, 3);
4156 /* Find the loop-closed-use at the loop exit of the original scalar
4157 result. (The reduction result is expected to have two immediate uses -
4158 one at the latch block, and one at the loop exit). */
4159 FOR_EACH_IMM_USE_FAST (use_p, imm_iter, scalar_dest)
4160 if (!flow_bb_inside_loop_p (loop, gimple_bb (USE_STMT (use_p))))
4161 VEC_safe_push (gimple, heap, phis, USE_STMT (use_p));
4163 /* We expect to have found an exit_phi because of loop-closed-ssa
4165 gcc_assert (!VEC_empty (gimple, phis));
4167 FOR_EACH_VEC_ELT (gimple, phis, i, exit_phi)
4171 stmt_vec_info exit_phi_vinfo = vinfo_for_stmt (exit_phi);
4174 /* FORNOW. Currently not supporting the case that an inner-loop
4175 reduction is not used in the outer-loop (but only outside the
4176 outer-loop), unless it is double reduction. */
4177 gcc_assert ((STMT_VINFO_RELEVANT_P (exit_phi_vinfo)
4178 && !STMT_VINFO_LIVE_P (exit_phi_vinfo))
4181 STMT_VINFO_VEC_STMT (exit_phi_vinfo) = epilog_stmt;
4183 || STMT_VINFO_DEF_TYPE (exit_phi_vinfo)
4184 != vect_double_reduction_def)
4187 /* Handle double reduction:
4189 stmt1: s1 = phi <s0, s2> - double reduction phi (outer loop)
4190 stmt2: s3 = phi <s1, s4> - (regular) reduc phi (inner loop)
4191 stmt3: s4 = use (s3) - (regular) reduc stmt (inner loop)
4192 stmt4: s2 = phi <s4> - double reduction stmt (outer loop)
4194 At that point the regular reduction (stmt2 and stmt3) is
4195 already vectorized, as well as the exit phi node, stmt4.
4196 Here we vectorize the phi node of double reduction, stmt1, and
4197 update all relevant statements. */
4199 /* Go through all the uses of s2 to find double reduction phi
4200 node, i.e., stmt1 above. */
4201 orig_name = PHI_RESULT (exit_phi);
4202 FOR_EACH_IMM_USE_STMT (use_stmt, imm_iter, orig_name)
4204 stmt_vec_info use_stmt_vinfo = vinfo_for_stmt (use_stmt);
4205 stmt_vec_info new_phi_vinfo;
4206 tree vect_phi_init, preheader_arg, vect_phi_res, init_def;
4207 basic_block bb = gimple_bb (use_stmt);
4210 /* Check that USE_STMT is really double reduction phi
4212 if (gimple_code (use_stmt) != GIMPLE_PHI
4213 || gimple_phi_num_args (use_stmt) != 2
4215 || STMT_VINFO_DEF_TYPE (use_stmt_vinfo)
4216 != vect_double_reduction_def
4217 || bb->loop_father != outer_loop)
4220 /* Create vector phi node for double reduction:
4221 vs1 = phi <vs0, vs2>
4222 vs1 was created previously in this function by a call to
4223 vect_get_vec_def_for_operand and is stored in
4225 vs2 is defined by INNER_PHI, the vectorized EXIT_PHI;
4226 vs0 is created here. */
4228 /* Create vector phi node. */
4229 vect_phi = create_phi_node (vec_initial_def, bb);
4230 new_phi_vinfo = new_stmt_vec_info (vect_phi,
4231 loop_vec_info_for_loop (outer_loop), NULL);
4232 set_vinfo_for_stmt (vect_phi, new_phi_vinfo);
4234 /* Create vs0 - initial def of the double reduction phi. */
4235 preheader_arg = PHI_ARG_DEF_FROM_EDGE (use_stmt,
4236 loop_preheader_edge (outer_loop));
4237 init_def = get_initial_def_for_reduction (stmt,
4238 preheader_arg, NULL);
4239 vect_phi_init = vect_init_vector (use_stmt, init_def,
4242 /* Update phi node arguments with vs0 and vs2. */
4243 add_phi_arg (vect_phi, vect_phi_init,
4244 loop_preheader_edge (outer_loop),
4246 add_phi_arg (vect_phi, PHI_RESULT (inner_phi),
4247 loop_latch_edge (outer_loop), UNKNOWN_LOCATION);
4248 if (vect_print_dump_info (REPORT_DETAILS))
4250 fprintf (vect_dump, "created double reduction phi "
4252 print_gimple_stmt (vect_dump, vect_phi, 0, TDF_SLIM);
4255 vect_phi_res = PHI_RESULT (vect_phi);
4257 /* Replace the use, i.e., set the correct vs1 in the regular
4258 reduction phi node. FORNOW, NCOPIES is always 1, so the
4259 loop is redundant. */
4260 use = reduction_phi;
4261 for (j = 0; j < ncopies; j++)
4263 edge pr_edge = loop_preheader_edge (loop);
4264 SET_PHI_ARG_DEF (use, pr_edge->dest_idx, vect_phi_res);
4265 use = STMT_VINFO_RELATED_STMT (vinfo_for_stmt (use));
4271 VEC_free (gimple, heap, phis);
4272 if (nested_in_vect_loop)
4280 phis = VEC_alloc (gimple, heap, 3);
4281 /* Find the loop-closed-use at the loop exit of the original scalar
4282 result. (The reduction result is expected to have two immediate uses,
4283 one at the latch block, and one at the loop exit). For double
4284 reductions we are looking for exit phis of the outer loop. */
4285 FOR_EACH_IMM_USE_FAST (use_p, imm_iter, scalar_dest)
4287 if (!flow_bb_inside_loop_p (loop, gimple_bb (USE_STMT (use_p))))
4288 VEC_safe_push (gimple, heap, phis, USE_STMT (use_p));
4291 if (double_reduc && gimple_code (USE_STMT (use_p)) == GIMPLE_PHI)
4293 tree phi_res = PHI_RESULT (USE_STMT (use_p));
4295 FOR_EACH_IMM_USE_FAST (phi_use_p, phi_imm_iter, phi_res)
4297 if (!flow_bb_inside_loop_p (loop,
4298 gimple_bb (USE_STMT (phi_use_p))))
4299 VEC_safe_push (gimple, heap, phis,
4300 USE_STMT (phi_use_p));
4306 FOR_EACH_VEC_ELT (gimple, phis, i, exit_phi)
4308 /* Replace the uses: */
4309 orig_name = PHI_RESULT (exit_phi);
4310 scalar_result = VEC_index (tree, scalar_results, k);
4311 FOR_EACH_IMM_USE_STMT (use_stmt, imm_iter, orig_name)
4312 FOR_EACH_IMM_USE_ON_STMT (use_p, imm_iter)
4313 SET_USE (use_p, scalar_result);
4316 VEC_free (gimple, heap, phis);
4319 VEC_free (tree, heap, scalar_results);
4320 VEC_free (gimple, heap, new_phis);
4324 /* Function vectorizable_reduction.
4326 Check if STMT performs a reduction operation that can be vectorized.
4327 If VEC_STMT is also passed, vectorize the STMT: create a vectorized
4328 stmt to replace it, put it in VEC_STMT, and insert it at GSI.
4329 Return FALSE if not a vectorizable STMT, TRUE otherwise.
4331 This function also handles reduction idioms (patterns) that have been
4332 recognized in advance during vect_pattern_recog. In this case, STMT may be
4334 X = pattern_expr (arg0, arg1, ..., X)
4335 and it's STMT_VINFO_RELATED_STMT points to the last stmt in the original
4336 sequence that had been detected and replaced by the pattern-stmt (STMT).
4338 In some cases of reduction patterns, the type of the reduction variable X is
4339 different than the type of the other arguments of STMT.
4340 In such cases, the vectype that is used when transforming STMT into a vector
4341 stmt is different than the vectype that is used to determine the
4342 vectorization factor, because it consists of a different number of elements
4343 than the actual number of elements that are being operated upon in parallel.
4345 For example, consider an accumulation of shorts into an int accumulator.
4346 On some targets it's possible to vectorize this pattern operating on 8
4347 shorts at a time (hence, the vectype for purposes of determining the
4348 vectorization factor should be V8HI); on the other hand, the vectype that
4349 is used to create the vector form is actually V4SI (the type of the result).
4351 Upon entry to this function, STMT_VINFO_VECTYPE records the vectype that
4352 indicates what is the actual level of parallelism (V8HI in the example), so
4353 that the right vectorization factor would be derived. This vectype
4354 corresponds to the type of arguments to the reduction stmt, and should *NOT*
4355 be used to create the vectorized stmt. The right vectype for the vectorized
4356 stmt is obtained from the type of the result X:
4357 get_vectype_for_scalar_type (TREE_TYPE (X))
4359 This means that, contrary to "regular" reductions (or "regular" stmts in
4360 general), the following equation:
4361 STMT_VINFO_VECTYPE == get_vectype_for_scalar_type (TREE_TYPE (X))
4362 does *NOT* necessarily hold for reduction patterns. */
4365 vectorizable_reduction (gimple stmt, gimple_stmt_iterator *gsi,
4366 gimple *vec_stmt, slp_tree slp_node)
4370 tree loop_vec_def0 = NULL_TREE, loop_vec_def1 = NULL_TREE;
4371 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
4372 tree vectype_out = STMT_VINFO_VECTYPE (stmt_info);
4373 tree vectype_in = NULL_TREE;
4374 loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info);
4375 struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
4376 enum tree_code code, orig_code, epilog_reduc_code;
4377 enum machine_mode vec_mode;
4379 optab optab, reduc_optab;
4380 tree new_temp = NULL_TREE;
4383 enum vect_def_type dt;
4384 gimple new_phi = NULL;
4388 stmt_vec_info orig_stmt_info;
4389 tree expr = NULL_TREE;
4393 stmt_vec_info prev_stmt_info, prev_phi_info;
4394 bool single_defuse_cycle = false;
4395 tree reduc_def = NULL_TREE;
4396 gimple new_stmt = NULL;
4399 bool nested_cycle = false, found_nested_cycle_def = false;
4400 gimple reduc_def_stmt = NULL;
4401 /* The default is that the reduction variable is the last in statement. */
4402 int reduc_index = 2;
4403 bool double_reduc = false, dummy;
4405 struct loop * def_stmt_loop, *outer_loop = NULL;
4407 gimple def_arg_stmt;
4408 VEC (tree, heap) *vec_oprnds0 = NULL, *vec_oprnds1 = NULL, *vect_defs = NULL;
4409 VEC (gimple, heap) *phis = NULL;
4411 tree def0, def1, tem, op0, op1 = NULL_TREE;
4413 /* In case of reduction chain we switch to the first stmt in the chain, but
4414 we don't update STMT_INFO, since only the last stmt is marked as reduction
4415 and has reduction properties. */
4416 if (GROUP_FIRST_ELEMENT (vinfo_for_stmt (stmt)))
4417 stmt = GROUP_FIRST_ELEMENT (stmt_info);
4419 if (nested_in_vect_loop_p (loop, stmt))
4423 nested_cycle = true;
4426 /* 1. Is vectorizable reduction? */
4427 /* Not supportable if the reduction variable is used in the loop, unless
4428 it's a reduction chain. */
4429 if (STMT_VINFO_RELEVANT (stmt_info) > vect_used_in_outer
4430 && !GROUP_FIRST_ELEMENT (stmt_info))
4433 /* Reductions that are not used even in an enclosing outer-loop,
4434 are expected to be "live" (used out of the loop). */
4435 if (STMT_VINFO_RELEVANT (stmt_info) == vect_unused_in_scope
4436 && !STMT_VINFO_LIVE_P (stmt_info))
4439 /* Make sure it was already recognized as a reduction computation. */
4440 if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_reduction_def
4441 && STMT_VINFO_DEF_TYPE (stmt_info) != vect_nested_cycle)
4444 /* 2. Has this been recognized as a reduction pattern?
4446 Check if STMT represents a pattern that has been recognized
4447 in earlier analysis stages. For stmts that represent a pattern,
4448 the STMT_VINFO_RELATED_STMT field records the last stmt in
4449 the original sequence that constitutes the pattern. */
4451 orig_stmt = STMT_VINFO_RELATED_STMT (stmt_info);
4454 orig_stmt_info = vinfo_for_stmt (orig_stmt);
4455 gcc_assert (STMT_VINFO_RELATED_STMT (orig_stmt_info) == 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 gcc_assert (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));
4562 if (!found_nested_cycle_def)
4563 reduc_def_stmt = def_stmt;
4565 gcc_assert (gimple_code (reduc_def_stmt) == GIMPLE_PHI);
4567 gcc_assert (orig_stmt == vect_is_simple_reduction (loop_vinfo,
4573 gimple tmp = vect_is_simple_reduction (loop_vinfo, reduc_def_stmt,
4574 !nested_cycle, &dummy);
4575 /* We changed STMT to be the first stmt in reduction chain, hence we
4576 check that in this case the first element in the chain is STMT. */
4577 gcc_assert (stmt == tmp
4578 || GROUP_FIRST_ELEMENT (vinfo_for_stmt (tmp)) == stmt);
4581 if (STMT_VINFO_LIVE_P (vinfo_for_stmt (reduc_def_stmt)))
4584 if (slp_node || PURE_SLP_STMT (stmt_info))
4587 ncopies = (LOOP_VINFO_VECT_FACTOR (loop_vinfo)
4588 / TYPE_VECTOR_SUBPARTS (vectype_in));
4590 gcc_assert (ncopies >= 1);
4592 vec_mode = TYPE_MODE (vectype_in);
4594 if (code == COND_EXPR)
4596 if (!vectorizable_condition (stmt, gsi, NULL, ops[reduc_index], 0, NULL))
4598 if (vect_print_dump_info (REPORT_DETAILS))
4599 fprintf (vect_dump, "unsupported condition in reduction");
4606 /* 4. Supportable by target? */
4608 /* 4.1. check support for the operation in the loop */
4609 optab = optab_for_tree_code (code, vectype_in, optab_default);
4612 if (vect_print_dump_info (REPORT_DETAILS))
4613 fprintf (vect_dump, "no optab.");
4618 if (optab_handler (optab, vec_mode) == CODE_FOR_nothing)
4620 if (vect_print_dump_info (REPORT_DETAILS))
4621 fprintf (vect_dump, "op not supported by target.");
4623 if (GET_MODE_SIZE (vec_mode) != UNITS_PER_WORD
4624 || LOOP_VINFO_VECT_FACTOR (loop_vinfo)
4625 < vect_min_worthwhile_factor (code))
4628 if (vect_print_dump_info (REPORT_DETAILS))
4629 fprintf (vect_dump, "proceeding using word mode.");
4632 /* Worthwhile without SIMD support? */
4633 if (!VECTOR_MODE_P (TYPE_MODE (vectype_in))
4634 && LOOP_VINFO_VECT_FACTOR (loop_vinfo)
4635 < vect_min_worthwhile_factor (code))
4637 if (vect_print_dump_info (REPORT_DETAILS))
4638 fprintf (vect_dump, "not worthwhile without SIMD support.");
4644 /* 4.2. Check support for the epilog operation.
4646 If STMT represents a reduction pattern, then the type of the
4647 reduction variable may be different than the type of the rest
4648 of the arguments. For example, consider the case of accumulation
4649 of shorts into an int accumulator; The original code:
4650 S1: int_a = (int) short_a;
4651 orig_stmt-> S2: int_acc = plus <int_a ,int_acc>;
4654 STMT: int_acc = widen_sum <short_a, int_acc>
4657 1. The tree-code that is used to create the vector operation in the
4658 epilog code (that reduces the partial results) is not the
4659 tree-code of STMT, but is rather the tree-code of the original
4660 stmt from the pattern that STMT is replacing. I.e, in the example
4661 above we want to use 'widen_sum' in the loop, but 'plus' in the
4663 2. The type (mode) we use to check available target support
4664 for the vector operation to be created in the *epilog*, is
4665 determined by the type of the reduction variable (in the example
4666 above we'd check this: optab_handler (plus_optab, vect_int_mode])).
4667 However the type (mode) we use to check available target support
4668 for the vector operation to be created *inside the loop*, is
4669 determined by the type of the other arguments to STMT (in the
4670 example we'd check this: optab_handler (widen_sum_optab,
4673 This is contrary to "regular" reductions, in which the types of all
4674 the arguments are the same as the type of the reduction variable.
4675 For "regular" reductions we can therefore use the same vector type
4676 (and also the same tree-code) when generating the epilog code and
4677 when generating the code inside the loop. */
4681 /* This is a reduction pattern: get the vectype from the type of the
4682 reduction variable, and get the tree-code from orig_stmt. */
4683 orig_code = gimple_assign_rhs_code (orig_stmt);
4684 gcc_assert (vectype_out);
4685 vec_mode = TYPE_MODE (vectype_out);
4689 /* Regular reduction: use the same vectype and tree-code as used for
4690 the vector code inside the loop can be used for the epilog code. */
4696 def_bb = gimple_bb (reduc_def_stmt);
4697 def_stmt_loop = def_bb->loop_father;
4698 def_arg = PHI_ARG_DEF_FROM_EDGE (reduc_def_stmt,
4699 loop_preheader_edge (def_stmt_loop));
4700 if (TREE_CODE (def_arg) == SSA_NAME
4701 && (def_arg_stmt = SSA_NAME_DEF_STMT (def_arg))
4702 && gimple_code (def_arg_stmt) == GIMPLE_PHI
4703 && flow_bb_inside_loop_p (outer_loop, gimple_bb (def_arg_stmt))
4704 && vinfo_for_stmt (def_arg_stmt)
4705 && STMT_VINFO_DEF_TYPE (vinfo_for_stmt (def_arg_stmt))
4706 == vect_double_reduction_def)
4707 double_reduc = true;
4710 epilog_reduc_code = ERROR_MARK;
4711 if (reduction_code_for_scalar_code (orig_code, &epilog_reduc_code))
4713 reduc_optab = optab_for_tree_code (epilog_reduc_code, vectype_out,
4717 if (vect_print_dump_info (REPORT_DETAILS))
4718 fprintf (vect_dump, "no optab for reduction.");
4720 epilog_reduc_code = ERROR_MARK;
4724 && optab_handler (reduc_optab, vec_mode) == CODE_FOR_nothing)
4726 if (vect_print_dump_info (REPORT_DETAILS))
4727 fprintf (vect_dump, "reduc op not supported by target.");
4729 epilog_reduc_code = ERROR_MARK;
4734 if (!nested_cycle || double_reduc)
4736 if (vect_print_dump_info (REPORT_DETAILS))
4737 fprintf (vect_dump, "no reduc code for scalar code.");
4743 if (double_reduc && ncopies > 1)
4745 if (vect_print_dump_info (REPORT_DETAILS))
4746 fprintf (vect_dump, "multiple types in double reduction");
4751 /* In case of widenning multiplication by a constant, we update the type
4752 of the constant to be the type of the other operand. We check that the
4753 constant fits the type in the pattern recognition pass. */
4754 if (code == DOT_PROD_EXPR
4755 && !types_compatible_p (TREE_TYPE (ops[0]), TREE_TYPE (ops[1])))
4757 if (TREE_CODE (ops[0]) == INTEGER_CST)
4758 ops[0] = fold_convert (TREE_TYPE (ops[1]), ops[0]);
4759 else if (TREE_CODE (ops[1]) == INTEGER_CST)
4760 ops[1] = fold_convert (TREE_TYPE (ops[0]), ops[1]);
4763 if (vect_print_dump_info (REPORT_DETAILS))
4764 fprintf (vect_dump, "invalid types in dot-prod");
4770 if (!vec_stmt) /* transformation not required. */
4772 if (!vect_model_reduction_cost (stmt_info, epilog_reduc_code, ncopies))
4774 STMT_VINFO_TYPE (stmt_info) = reduc_vec_info_type;
4780 if (vect_print_dump_info (REPORT_DETAILS))
4781 fprintf (vect_dump, "transform reduction.");
4783 /* FORNOW: Multiple types are not supported for condition. */
4784 if (code == COND_EXPR)
4785 gcc_assert (ncopies == 1);
4787 /* Create the destination vector */
4788 vec_dest = vect_create_destination_var (scalar_dest, vectype_out);
4790 /* In case the vectorization factor (VF) is bigger than the number
4791 of elements that we can fit in a vectype (nunits), we have to generate
4792 more than one vector stmt - i.e - we need to "unroll" the
4793 vector stmt by a factor VF/nunits. For more details see documentation
4794 in vectorizable_operation. */
4796 /* If the reduction is used in an outer loop we need to generate
4797 VF intermediate results, like so (e.g. for ncopies=2):
4802 (i.e. we generate VF results in 2 registers).
4803 In this case we have a separate def-use cycle for each copy, and therefore
4804 for each copy we get the vector def for the reduction variable from the
4805 respective phi node created for this copy.
4807 Otherwise (the reduction is unused in the loop nest), we can combine
4808 together intermediate results, like so (e.g. for ncopies=2):
4812 (i.e. we generate VF/2 results in a single register).
4813 In this case for each copy we get the vector def for the reduction variable
4814 from the vectorized reduction operation generated in the previous iteration.
4817 if (STMT_VINFO_RELEVANT (stmt_info) == vect_unused_in_scope)
4819 single_defuse_cycle = true;
4823 epilog_copies = ncopies;
4825 prev_stmt_info = NULL;
4826 prev_phi_info = NULL;
4829 vec_num = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node);
4830 gcc_assert (TYPE_VECTOR_SUBPARTS (vectype_out)
4831 == TYPE_VECTOR_SUBPARTS (vectype_in));
4836 vec_oprnds0 = VEC_alloc (tree, heap, 1);
4837 if (op_type == ternary_op)
4838 vec_oprnds1 = VEC_alloc (tree, heap, 1);
4841 phis = VEC_alloc (gimple, heap, vec_num);
4842 vect_defs = VEC_alloc (tree, heap, vec_num);
4844 VEC_quick_push (tree, vect_defs, NULL_TREE);
4846 for (j = 0; j < ncopies; j++)
4848 if (j == 0 || !single_defuse_cycle)
4850 for (i = 0; i < vec_num; i++)
4852 /* Create the reduction-phi that defines the reduction
4854 new_phi = create_phi_node (vec_dest, loop->header);
4855 set_vinfo_for_stmt (new_phi,
4856 new_stmt_vec_info (new_phi, loop_vinfo,
4858 if (j == 0 || slp_node)
4859 VEC_quick_push (gimple, phis, new_phi);
4863 if (code == COND_EXPR)
4865 gcc_assert (!slp_node);
4866 vectorizable_condition (stmt, gsi, vec_stmt,
4867 PHI_RESULT (VEC_index (gimple, phis, 0)),
4869 /* Multiple types are not supported for condition. */
4876 op0 = ops[!reduc_index];
4877 if (op_type == ternary_op)
4879 if (reduc_index == 0)
4886 vect_get_vec_defs (op0, op1, stmt, &vec_oprnds0, &vec_oprnds1,
4890 loop_vec_def0 = vect_get_vec_def_for_operand (ops[!reduc_index],
4892 VEC_quick_push (tree, vec_oprnds0, loop_vec_def0);
4893 if (op_type == ternary_op)
4895 loop_vec_def1 = vect_get_vec_def_for_operand (op1, stmt,
4897 VEC_quick_push (tree, vec_oprnds1, loop_vec_def1);
4905 enum vect_def_type dt;
4909 vect_is_simple_use (ops[!reduc_index], stmt, loop_vinfo, NULL,
4910 &dummy_stmt, &dummy, &dt);
4911 loop_vec_def0 = vect_get_vec_def_for_stmt_copy (dt,
4913 VEC_replace (tree, vec_oprnds0, 0, loop_vec_def0);
4914 if (op_type == ternary_op)
4916 vect_is_simple_use (op1, stmt, loop_vinfo, NULL, &dummy_stmt,
4918 loop_vec_def1 = vect_get_vec_def_for_stmt_copy (dt,
4920 VEC_replace (tree, vec_oprnds1, 0, loop_vec_def1);
4924 if (single_defuse_cycle)
4925 reduc_def = gimple_assign_lhs (new_stmt);
4927 STMT_VINFO_RELATED_STMT (prev_phi_info) = new_phi;
4930 FOR_EACH_VEC_ELT (tree, vec_oprnds0, i, def0)
4933 reduc_def = PHI_RESULT (VEC_index (gimple, phis, i));
4936 if (!single_defuse_cycle || j == 0)
4937 reduc_def = PHI_RESULT (new_phi);
4940 def1 = ((op_type == ternary_op)
4941 ? VEC_index (tree, vec_oprnds1, i) : NULL);
4942 if (op_type == binary_op)
4944 if (reduc_index == 0)
4945 expr = build2 (code, vectype_out, reduc_def, def0);
4947 expr = build2 (code, vectype_out, def0, reduc_def);
4951 if (reduc_index == 0)
4952 expr = build3 (code, vectype_out, reduc_def, def0, def1);
4955 if (reduc_index == 1)
4956 expr = build3 (code, vectype_out, def0, reduc_def, def1);
4958 expr = build3 (code, vectype_out, def0, def1, reduc_def);
4962 new_stmt = gimple_build_assign (vec_dest, expr);
4963 new_temp = make_ssa_name (vec_dest, new_stmt);
4964 gimple_assign_set_lhs (new_stmt, new_temp);
4965 vect_finish_stmt_generation (stmt, new_stmt, gsi);
4969 VEC_quick_push (gimple, SLP_TREE_VEC_STMTS (slp_node), new_stmt);
4970 VEC_quick_push (tree, vect_defs, new_temp);
4973 VEC_replace (tree, vect_defs, 0, new_temp);
4980 STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt;
4982 STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt;
4984 prev_stmt_info = vinfo_for_stmt (new_stmt);
4985 prev_phi_info = vinfo_for_stmt (new_phi);
4988 /* Finalize the reduction-phi (set its arguments) and create the
4989 epilog reduction code. */
4990 if ((!single_defuse_cycle || code == COND_EXPR) && !slp_node)
4992 new_temp = gimple_assign_lhs (*vec_stmt);
4993 VEC_replace (tree, vect_defs, 0, new_temp);
4996 vect_create_epilog_for_reduction (vect_defs, stmt, epilog_copies,
4997 epilog_reduc_code, phis, reduc_index,
4998 double_reduc, slp_node);
5000 VEC_free (gimple, heap, phis);
5001 VEC_free (tree, heap, vec_oprnds0);
5003 VEC_free (tree, heap, vec_oprnds1);
5008 /* Function vect_min_worthwhile_factor.
5010 For a loop where we could vectorize the operation indicated by CODE,
5011 return the minimum vectorization factor that makes it worthwhile
5012 to use generic vectors. */
5014 vect_min_worthwhile_factor (enum tree_code code)
5035 /* Function vectorizable_induction
5037 Check if PHI performs an induction computation that can be vectorized.
5038 If VEC_STMT is also passed, vectorize the induction PHI: create a vectorized
5039 phi to replace it, put it in VEC_STMT, and add it to the same basic block.
5040 Return FALSE if not a vectorizable STMT, TRUE otherwise. */
5043 vectorizable_induction (gimple phi, gimple_stmt_iterator *gsi ATTRIBUTE_UNUSED,
5046 stmt_vec_info stmt_info = vinfo_for_stmt (phi);
5047 tree vectype = STMT_VINFO_VECTYPE (stmt_info);
5048 loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info);
5049 struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
5050 int nunits = TYPE_VECTOR_SUBPARTS (vectype);
5051 int ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits;
5054 gcc_assert (ncopies >= 1);
5055 /* FORNOW. This restriction should be relaxed. */
5056 if (nested_in_vect_loop_p (loop, phi) && ncopies > 1)
5058 if (vect_print_dump_info (REPORT_DETAILS))
5059 fprintf (vect_dump, "multiple types in nested loop.");
5063 if (!STMT_VINFO_RELEVANT_P (stmt_info))
5066 /* FORNOW: SLP not supported. */
5067 if (STMT_SLP_TYPE (stmt_info))
5070 gcc_assert (STMT_VINFO_DEF_TYPE (stmt_info) == vect_induction_def);
5072 if (gimple_code (phi) != GIMPLE_PHI)
5075 if (!vec_stmt) /* transformation not required. */
5077 STMT_VINFO_TYPE (stmt_info) = induc_vec_info_type;
5078 if (vect_print_dump_info (REPORT_DETAILS))
5079 fprintf (vect_dump, "=== vectorizable_induction ===");
5080 vect_model_induction_cost (stmt_info, ncopies);
5086 if (vect_print_dump_info (REPORT_DETAILS))
5087 fprintf (vect_dump, "transform induction phi.");
5089 vec_def = get_initial_def_for_induction (phi);
5090 *vec_stmt = SSA_NAME_DEF_STMT (vec_def);
5094 /* Function vectorizable_live_operation.
5096 STMT computes a value that is used outside the loop. Check if
5097 it can be supported. */
5100 vectorizable_live_operation (gimple stmt,
5101 gimple_stmt_iterator *gsi ATTRIBUTE_UNUSED,
5102 gimple *vec_stmt ATTRIBUTE_UNUSED)
5104 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
5105 loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info);
5106 struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
5112 enum vect_def_type dt;
5113 enum tree_code code;
5114 enum gimple_rhs_class rhs_class;
5116 gcc_assert (STMT_VINFO_LIVE_P (stmt_info));
5118 if (STMT_VINFO_DEF_TYPE (stmt_info) == vect_reduction_def)
5121 if (!is_gimple_assign (stmt))
5124 if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME)
5127 /* FORNOW. CHECKME. */
5128 if (nested_in_vect_loop_p (loop, stmt))
5131 code = gimple_assign_rhs_code (stmt);
5132 op_type = TREE_CODE_LENGTH (code);
5133 rhs_class = get_gimple_rhs_class (code);
5134 gcc_assert (rhs_class != GIMPLE_UNARY_RHS || op_type == unary_op);
5135 gcc_assert (rhs_class != GIMPLE_BINARY_RHS || op_type == binary_op);
5137 /* FORNOW: support only if all uses are invariant. This means
5138 that the scalar operations can remain in place, unvectorized.
5139 The original last scalar value that they compute will be used. */
5141 for (i = 0; i < op_type; i++)
5143 if (rhs_class == GIMPLE_SINGLE_RHS)
5144 op = TREE_OPERAND (gimple_op (stmt, 1), i);
5146 op = gimple_op (stmt, i + 1);
5148 && !vect_is_simple_use (op, stmt, loop_vinfo, NULL, &def_stmt, &def,
5151 if (vect_print_dump_info (REPORT_DETAILS))
5152 fprintf (vect_dump, "use not simple.");
5156 if (dt != vect_external_def && dt != vect_constant_def)
5160 /* No transformation is required for the cases we currently support. */
5164 /* Kill any debug uses outside LOOP of SSA names defined in STMT. */
5167 vect_loop_kill_debug_uses (struct loop *loop, gimple stmt)
5169 ssa_op_iter op_iter;
5170 imm_use_iterator imm_iter;
5171 def_operand_p def_p;
5174 FOR_EACH_PHI_OR_STMT_DEF (def_p, stmt, op_iter, SSA_OP_DEF)
5176 FOR_EACH_IMM_USE_STMT (ustmt, imm_iter, DEF_FROM_PTR (def_p))
5180 if (!is_gimple_debug (ustmt))
5183 bb = gimple_bb (ustmt);
5185 if (!flow_bb_inside_loop_p (loop, bb))
5187 if (gimple_debug_bind_p (ustmt))
5189 if (vect_print_dump_info (REPORT_DETAILS))
5190 fprintf (vect_dump, "killing debug use");
5192 gimple_debug_bind_reset_value (ustmt);
5193 update_stmt (ustmt);
5202 /* Function vect_transform_loop.
5204 The analysis phase has determined that the loop is vectorizable.
5205 Vectorize the loop - created vectorized stmts to replace the scalar
5206 stmts in the loop, and update the loop exit condition. */
5209 vect_transform_loop (loop_vec_info loop_vinfo)
5211 struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
5212 basic_block *bbs = LOOP_VINFO_BBS (loop_vinfo);
5213 int nbbs = loop->num_nodes;
5214 gimple_stmt_iterator si;
5217 int vectorization_factor = LOOP_VINFO_VECT_FACTOR (loop_vinfo);
5219 bool slp_scheduled = false;
5220 unsigned int nunits;
5221 tree cond_expr = NULL_TREE;
5222 gimple_seq cond_expr_stmt_list = NULL;
5223 bool do_peeling_for_loop_bound;
5224 gimple stmt, pattern_stmt;
5225 gimple_seq pattern_def_seq = NULL;
5226 gimple_stmt_iterator pattern_def_si = gsi_start (NULL);
5227 bool transform_pattern_stmt = false;
5229 if (vect_print_dump_info (REPORT_DETAILS))
5230 fprintf (vect_dump, "=== vec_transform_loop ===");
5232 /* Peel the loop if there are data refs with unknown alignment.
5233 Only one data ref with unknown store is allowed. */
5235 if (LOOP_PEELING_FOR_ALIGNMENT (loop_vinfo))
5236 vect_do_peeling_for_alignment (loop_vinfo);
5238 do_peeling_for_loop_bound
5239 = (!LOOP_VINFO_NITERS_KNOWN_P (loop_vinfo)
5240 || (LOOP_VINFO_NITERS_KNOWN_P (loop_vinfo)
5241 && LOOP_VINFO_INT_NITERS (loop_vinfo) % vectorization_factor != 0)
5242 || LOOP_VINFO_PEELING_FOR_GAPS (loop_vinfo));
5244 if (LOOP_REQUIRES_VERSIONING_FOR_ALIGNMENT (loop_vinfo)
5245 || LOOP_REQUIRES_VERSIONING_FOR_ALIAS (loop_vinfo))
5246 vect_loop_versioning (loop_vinfo,
5247 !do_peeling_for_loop_bound,
5248 &cond_expr, &cond_expr_stmt_list);
5250 /* If the loop has a symbolic number of iterations 'n' (i.e. it's not a
5251 compile time constant), or it is a constant that doesn't divide by the
5252 vectorization factor, then an epilog loop needs to be created.
5253 We therefore duplicate the loop: the original loop will be vectorized,
5254 and will compute the first (n/VF) iterations. The second copy of the loop
5255 will remain scalar and will compute the remaining (n%VF) iterations.
5256 (VF is the vectorization factor). */
5258 if (do_peeling_for_loop_bound)
5259 vect_do_peeling_for_loop_bound (loop_vinfo, &ratio,
5260 cond_expr, cond_expr_stmt_list);
5262 ratio = build_int_cst (TREE_TYPE (LOOP_VINFO_NITERS (loop_vinfo)),
5263 LOOP_VINFO_INT_NITERS (loop_vinfo) / vectorization_factor);
5265 /* 1) Make sure the loop header has exactly two entries
5266 2) Make sure we have a preheader basic block. */
5268 gcc_assert (EDGE_COUNT (loop->header->preds) == 2);
5270 split_edge (loop_preheader_edge (loop));
5272 /* FORNOW: the vectorizer supports only loops which body consist
5273 of one basic block (header + empty latch). When the vectorizer will
5274 support more involved loop forms, the order by which the BBs are
5275 traversed need to be reconsidered. */
5277 for (i = 0; i < nbbs; i++)
5279 basic_block bb = bbs[i];
5280 stmt_vec_info stmt_info;
5283 for (si = gsi_start_phis (bb); !gsi_end_p (si); gsi_next (&si))
5285 phi = gsi_stmt (si);
5286 if (vect_print_dump_info (REPORT_DETAILS))
5288 fprintf (vect_dump, "------>vectorizing phi: ");
5289 print_gimple_stmt (vect_dump, phi, 0, TDF_SLIM);
5291 stmt_info = vinfo_for_stmt (phi);
5295 if (MAY_HAVE_DEBUG_STMTS && !STMT_VINFO_LIVE_P (stmt_info))
5296 vect_loop_kill_debug_uses (loop, phi);
5298 if (!STMT_VINFO_RELEVANT_P (stmt_info)
5299 && !STMT_VINFO_LIVE_P (stmt_info))
5302 if ((TYPE_VECTOR_SUBPARTS (STMT_VINFO_VECTYPE (stmt_info))
5303 != (unsigned HOST_WIDE_INT) vectorization_factor)
5304 && vect_print_dump_info (REPORT_DETAILS))
5305 fprintf (vect_dump, "multiple-types.");
5307 if (STMT_VINFO_DEF_TYPE (stmt_info) == vect_induction_def)
5309 if (vect_print_dump_info (REPORT_DETAILS))
5310 fprintf (vect_dump, "transform phi.");
5311 vect_transform_stmt (phi, NULL, NULL, NULL, NULL);
5315 pattern_stmt = NULL;
5316 for (si = gsi_start_bb (bb); !gsi_end_p (si) || transform_pattern_stmt;)
5320 if (transform_pattern_stmt)
5321 stmt = pattern_stmt;
5323 stmt = gsi_stmt (si);
5325 if (vect_print_dump_info (REPORT_DETAILS))
5327 fprintf (vect_dump, "------>vectorizing statement: ");
5328 print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM);
5331 stmt_info = vinfo_for_stmt (stmt);
5333 /* vector stmts created in the outer-loop during vectorization of
5334 stmts in an inner-loop may not have a stmt_info, and do not
5335 need to be vectorized. */
5342 if (MAY_HAVE_DEBUG_STMTS && !STMT_VINFO_LIVE_P (stmt_info))
5343 vect_loop_kill_debug_uses (loop, stmt);
5345 if (!STMT_VINFO_RELEVANT_P (stmt_info)
5346 && !STMT_VINFO_LIVE_P (stmt_info))
5348 if (STMT_VINFO_IN_PATTERN_P (stmt_info)
5349 && (pattern_stmt = STMT_VINFO_RELATED_STMT (stmt_info))
5350 && (STMT_VINFO_RELEVANT_P (vinfo_for_stmt (pattern_stmt))
5351 || STMT_VINFO_LIVE_P (vinfo_for_stmt (pattern_stmt))))
5353 stmt = pattern_stmt;
5354 stmt_info = vinfo_for_stmt (stmt);
5362 else if (STMT_VINFO_IN_PATTERN_P (stmt_info)
5363 && (pattern_stmt = STMT_VINFO_RELATED_STMT (stmt_info))
5364 && (STMT_VINFO_RELEVANT_P (vinfo_for_stmt (pattern_stmt))
5365 || STMT_VINFO_LIVE_P (vinfo_for_stmt (pattern_stmt))))
5366 transform_pattern_stmt = true;
5368 /* If pattern statement has def stmts, vectorize them too. */
5369 if (is_pattern_stmt_p (stmt_info))
5371 if (pattern_def_seq == NULL)
5373 pattern_def_seq = STMT_VINFO_PATTERN_DEF_SEQ (stmt_info);
5374 pattern_def_si = gsi_start (pattern_def_seq);
5376 else if (!gsi_end_p (pattern_def_si))
5377 gsi_next (&pattern_def_si);
5378 if (pattern_def_seq != NULL)
5380 gimple pattern_def_stmt = NULL;
5381 stmt_vec_info pattern_def_stmt_info = NULL;
5383 while (!gsi_end_p (pattern_def_si))
5385 pattern_def_stmt = gsi_stmt (pattern_def_si);
5386 pattern_def_stmt_info
5387 = vinfo_for_stmt (pattern_def_stmt);
5388 if (STMT_VINFO_RELEVANT_P (pattern_def_stmt_info)
5389 || STMT_VINFO_LIVE_P (pattern_def_stmt_info))
5391 gsi_next (&pattern_def_si);
5394 if (!gsi_end_p (pattern_def_si))
5396 if (vect_print_dump_info (REPORT_DETAILS))
5398 fprintf (vect_dump, "==> vectorizing pattern def"
5400 print_gimple_stmt (vect_dump, pattern_def_stmt, 0,
5404 stmt = pattern_def_stmt;
5405 stmt_info = pattern_def_stmt_info;
5409 pattern_def_si = gsi_start (NULL);
5410 transform_pattern_stmt = false;
5414 transform_pattern_stmt = false;
5417 gcc_assert (STMT_VINFO_VECTYPE (stmt_info));
5418 nunits = (unsigned int) TYPE_VECTOR_SUBPARTS (
5419 STMT_VINFO_VECTYPE (stmt_info));
5420 if (!STMT_SLP_TYPE (stmt_info)
5421 && nunits != (unsigned int) vectorization_factor
5422 && vect_print_dump_info (REPORT_DETAILS))
5423 /* For SLP VF is set according to unrolling factor, and not to
5424 vector size, hence for SLP this print is not valid. */
5425 fprintf (vect_dump, "multiple-types.");
5427 /* SLP. Schedule all the SLP instances when the first SLP stmt is
5429 if (STMT_SLP_TYPE (stmt_info))
5433 slp_scheduled = true;
5435 if (vect_print_dump_info (REPORT_DETAILS))
5436 fprintf (vect_dump, "=== scheduling SLP instances ===");
5438 vect_schedule_slp (loop_vinfo, NULL);
5441 /* Hybrid SLP stmts must be vectorized in addition to SLP. */
5442 if (!vinfo_for_stmt (stmt) || PURE_SLP_STMT (stmt_info))
5444 if (!transform_pattern_stmt && gsi_end_p (pattern_def_si))
5446 pattern_def_seq = NULL;
5453 /* -------- vectorize statement ------------ */
5454 if (vect_print_dump_info (REPORT_DETAILS))
5455 fprintf (vect_dump, "transform statement.");
5457 strided_store = false;
5458 is_store = vect_transform_stmt (stmt, &si, &strided_store, NULL, NULL);
5461 if (STMT_VINFO_STRIDED_ACCESS (stmt_info))
5463 /* Interleaving. If IS_STORE is TRUE, the vectorization of the
5464 interleaving chain was completed - free all the stores in
5467 vect_remove_stores (GROUP_FIRST_ELEMENT (stmt_info));
5472 /* Free the attached stmt_vec_info and remove the stmt. */
5473 free_stmt_vec_info (gsi_stmt (si));
5474 gsi_remove (&si, true);
5479 if (!transform_pattern_stmt && gsi_end_p (pattern_def_si))
5481 pattern_def_seq = NULL;
5487 slpeel_make_loop_iterate_ntimes (loop, ratio);
5489 /* The memory tags and pointers in vectorized statements need to
5490 have their SSA forms updated. FIXME, why can't this be delayed
5491 until all the loops have been transformed? */
5492 update_ssa (TODO_update_ssa);
5494 if (vect_print_dump_info (REPORT_VECTORIZED_LOCATIONS))
5495 fprintf (vect_dump, "LOOP VECTORIZED.");
5496 if (loop->inner && vect_print_dump_info (REPORT_VECTORIZED_LOCATIONS))
5497 fprintf (vect_dump, "OUTER LOOP VECTORIZED.");