2 Copyright (C) 2003, 2004, 2005, 2006, 2007, 2008, 2009 Free Software
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 "diagnostic.h"
31 #include "tree-flow.h"
32 #include "tree-dump.h"
34 #include "cfglayout.h"
40 #include "tree-chrec.h"
41 #include "tree-scalar-evolution.h"
42 #include "tree-vectorizer.h"
44 /* Loop Vectorization Pass.
46 This pass tries to vectorize loops.
48 For example, the vectorizer transforms the following simple loop:
50 short a[N]; short b[N]; short c[N]; int i;
56 as if it was manually vectorized by rewriting the source code into:
58 typedef int __attribute__((mode(V8HI))) v8hi;
59 short a[N]; short b[N]; short c[N]; int i;
60 v8hi *pa = (v8hi*)a, *pb = (v8hi*)b, *pc = (v8hi*)c;
63 for (i=0; i<N/8; i++){
70 The main entry to this pass is vectorize_loops(), in which
71 the vectorizer applies a set of analyses on a given set of loops,
72 followed by the actual vectorization transformation for the loops that
73 had successfully passed the analysis phase.
74 Throughout this pass we make a distinction between two types of
75 data: scalars (which are represented by SSA_NAMES), and memory references
76 ("data-refs"). These two types of data require different handling both
77 during analysis and transformation. The types of data-refs that the
78 vectorizer currently supports are ARRAY_REFS which base is an array DECL
79 (not a pointer), and INDIRECT_REFS through pointers; both array and pointer
80 accesses are required to have a simple (consecutive) access pattern.
84 The driver for the analysis phase is vect_analyze_loop().
85 It applies a set of analyses, some of which rely on the scalar evolution
86 analyzer (scev) developed by Sebastian Pop.
88 During the analysis phase the vectorizer records some information
89 per stmt in a "stmt_vec_info" struct which is attached to each stmt in the
90 loop, as well as general information about the loop as a whole, which is
91 recorded in a "loop_vec_info" struct attached to each loop.
95 The loop transformation phase scans all the stmts in the loop, and
96 creates a vector stmt (or a sequence of stmts) for each scalar stmt S in
97 the loop that needs to be vectorized. It inserts the vector code sequence
98 just before the scalar stmt S, and records a pointer to the vector code
99 in STMT_VINFO_VEC_STMT (stmt_info) (stmt_info is the stmt_vec_info struct
100 attached to S). This pointer will be used for the vectorization of following
101 stmts which use the def of stmt S. Stmt S is removed if it writes to memory;
102 otherwise, we rely on dead code elimination for removing it.
104 For example, say stmt S1 was vectorized into stmt VS1:
107 S1: b = x[i]; STMT_VINFO_VEC_STMT (stmt_info (S1)) = VS1
110 To vectorize stmt S2, the vectorizer first finds the stmt that defines
111 the operand 'b' (S1), and gets the relevant vector def 'vb' from the
112 vector stmt VS1 pointed to by STMT_VINFO_VEC_STMT (stmt_info (S1)). The
113 resulting sequence would be:
116 S1: b = x[i]; STMT_VINFO_VEC_STMT (stmt_info (S1)) = VS1
118 S2: a = b; STMT_VINFO_VEC_STMT (stmt_info (S2)) = VS2
120 Operands that are not SSA_NAMEs, are data-refs that appear in
121 load/store operations (like 'x[i]' in S1), and are handled differently.
125 Currently the only target specific information that is used is the
126 size of the vector (in bytes) - "UNITS_PER_SIMD_WORD". Targets that can
127 support different sizes of vectors, for now will need to specify one value
128 for "UNITS_PER_SIMD_WORD". More flexibility will be added in the future.
130 Since we only vectorize operations which vector form can be
131 expressed using existing tree codes, to verify that an operation is
132 supported, the vectorizer checks the relevant optab at the relevant
133 machine_mode (e.g, optab_handler (add_optab, V8HImode)->insn_code). If
134 the value found is CODE_FOR_nothing, then there's no target support, and
135 we can't vectorize the stmt.
137 For additional information on this project see:
138 http://gcc.gnu.org/projects/tree-ssa/vectorization.html
141 /* Function vect_determine_vectorization_factor
143 Determine the vectorization factor (VF). VF is the number of data elements
144 that are operated upon in parallel in a single iteration of the vectorized
145 loop. For example, when vectorizing a loop that operates on 4byte elements,
146 on a target with vector size (VS) 16byte, the VF is set to 4, since 4
147 elements can fit in a single vector register.
149 We currently support vectorization of loops in which all types operated upon
150 are of the same size. Therefore this function currently sets VF according to
151 the size of the types operated upon, and fails if there are multiple sizes
154 VF is also the factor by which the loop iterations are strip-mined, e.g.:
161 for (i=0; i<N; i+=VF){
162 a[i:VF] = b[i:VF] + c[i:VF];
167 vect_determine_vectorization_factor (loop_vec_info loop_vinfo
)
169 struct loop
*loop
= LOOP_VINFO_LOOP (loop_vinfo
);
170 basic_block
*bbs
= LOOP_VINFO_BBS (loop_vinfo
);
171 int nbbs
= loop
->num_nodes
;
172 gimple_stmt_iterator si
;
173 unsigned int vectorization_factor
= 0;
178 stmt_vec_info stmt_info
;
182 if (vect_print_dump_info (REPORT_DETAILS
))
183 fprintf (vect_dump
, "=== vect_determine_vectorization_factor ===");
185 for (i
= 0; i
< nbbs
; i
++)
187 basic_block bb
= bbs
[i
];
189 for (si
= gsi_start_phis (bb
); !gsi_end_p (si
); gsi_next (&si
))
192 stmt_info
= vinfo_for_stmt (phi
);
193 if (vect_print_dump_info (REPORT_DETAILS
))
195 fprintf (vect_dump
, "==> examining phi: ");
196 print_gimple_stmt (vect_dump
, phi
, 0, TDF_SLIM
);
199 gcc_assert (stmt_info
);
201 if (STMT_VINFO_RELEVANT_P (stmt_info
))
203 gcc_assert (!STMT_VINFO_VECTYPE (stmt_info
));
204 scalar_type
= TREE_TYPE (PHI_RESULT (phi
));
206 if (vect_print_dump_info (REPORT_DETAILS
))
208 fprintf (vect_dump
, "get vectype for scalar type: ");
209 print_generic_expr (vect_dump
, scalar_type
, TDF_SLIM
);
212 vectype
= get_vectype_for_scalar_type (scalar_type
);
215 if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS
))
218 "not vectorized: unsupported data-type ");
219 print_generic_expr (vect_dump
, scalar_type
, TDF_SLIM
);
223 STMT_VINFO_VECTYPE (stmt_info
) = vectype
;
225 if (vect_print_dump_info (REPORT_DETAILS
))
227 fprintf (vect_dump
, "vectype: ");
228 print_generic_expr (vect_dump
, vectype
, TDF_SLIM
);
231 nunits
= TYPE_VECTOR_SUBPARTS (vectype
);
232 if (vect_print_dump_info (REPORT_DETAILS
))
233 fprintf (vect_dump
, "nunits = %d", nunits
);
235 if (!vectorization_factor
236 || (nunits
> vectorization_factor
))
237 vectorization_factor
= nunits
;
241 for (si
= gsi_start_bb (bb
); !gsi_end_p (si
); gsi_next (&si
))
243 gimple stmt
= gsi_stmt (si
);
244 stmt_info
= vinfo_for_stmt (stmt
);
246 if (vect_print_dump_info (REPORT_DETAILS
))
248 fprintf (vect_dump
, "==> examining statement: ");
249 print_gimple_stmt (vect_dump
, stmt
, 0, TDF_SLIM
);
252 gcc_assert (stmt_info
);
254 /* skip stmts which do not need to be vectorized. */
255 if (!STMT_VINFO_RELEVANT_P (stmt_info
)
256 && !STMT_VINFO_LIVE_P (stmt_info
))
258 if (vect_print_dump_info (REPORT_DETAILS
))
259 fprintf (vect_dump
, "skip.");
263 if (gimple_get_lhs (stmt
) == NULL_TREE
)
265 if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS
))
267 fprintf (vect_dump
, "not vectorized: irregular stmt.");
268 print_gimple_stmt (vect_dump
, stmt
, 0, TDF_SLIM
);
273 if (VECTOR_MODE_P (TYPE_MODE (gimple_expr_type (stmt
))))
275 if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS
))
277 fprintf (vect_dump
, "not vectorized: vector stmt in loop:");
278 print_gimple_stmt (vect_dump
, stmt
, 0, TDF_SLIM
);
283 if (STMT_VINFO_VECTYPE (stmt_info
))
285 /* The only case when a vectype had been already set is for stmts
286 that contain a dataref, or for "pattern-stmts" (stmts generated
287 by the vectorizer to represent/replace a certain idiom). */
288 gcc_assert (STMT_VINFO_DATA_REF (stmt_info
)
289 || is_pattern_stmt_p (stmt_info
));
290 vectype
= STMT_VINFO_VECTYPE (stmt_info
);
294 gcc_assert (!STMT_VINFO_DATA_REF (stmt_info
)
295 && !is_pattern_stmt_p (stmt_info
));
297 scalar_type
= vect_get_smallest_scalar_type (stmt
, &dummy
,
299 if (vect_print_dump_info (REPORT_DETAILS
))
301 fprintf (vect_dump
, "get vectype for scalar type: ");
302 print_generic_expr (vect_dump
, scalar_type
, TDF_SLIM
);
305 vectype
= get_vectype_for_scalar_type (scalar_type
);
308 if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS
))
311 "not vectorized: unsupported data-type ");
312 print_generic_expr (vect_dump
, scalar_type
, TDF_SLIM
);
316 STMT_VINFO_VECTYPE (stmt_info
) = vectype
;
319 if (vect_print_dump_info (REPORT_DETAILS
))
321 fprintf (vect_dump
, "vectype: ");
322 print_generic_expr (vect_dump
, vectype
, TDF_SLIM
);
325 nunits
= TYPE_VECTOR_SUBPARTS (vectype
);
326 if (vect_print_dump_info (REPORT_DETAILS
))
327 fprintf (vect_dump
, "nunits = %d", nunits
);
329 if (!vectorization_factor
330 || (nunits
> vectorization_factor
))
331 vectorization_factor
= nunits
;
336 /* TODO: Analyze cost. Decide if worth while to vectorize. */
337 if (vect_print_dump_info (REPORT_DETAILS
))
338 fprintf (vect_dump
, "vectorization factor = %d", vectorization_factor
);
339 if (vectorization_factor
<= 1)
341 if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS
))
342 fprintf (vect_dump
, "not vectorized: unsupported data-type");
345 LOOP_VINFO_VECT_FACTOR (loop_vinfo
) = vectorization_factor
;
351 /* Function vect_is_simple_iv_evolution.
353 FORNOW: A simple evolution of an induction variables in the loop is
354 considered a polynomial evolution with constant step. */
357 vect_is_simple_iv_evolution (unsigned loop_nb
, tree access_fn
, tree
* init
,
362 tree evolution_part
= evolution_part_in_loop_num (access_fn
, loop_nb
);
364 /* When there is no evolution in this loop, the evolution function
366 if (evolution_part
== NULL_TREE
)
369 /* When the evolution is a polynomial of degree >= 2
370 the evolution function is not "simple". */
371 if (tree_is_chrec (evolution_part
))
374 step_expr
= evolution_part
;
375 init_expr
= unshare_expr (initial_condition_in_loop_num (access_fn
, loop_nb
));
377 if (vect_print_dump_info (REPORT_DETAILS
))
379 fprintf (vect_dump
, "step: ");
380 print_generic_expr (vect_dump
, step_expr
, TDF_SLIM
);
381 fprintf (vect_dump
, ", init: ");
382 print_generic_expr (vect_dump
, init_expr
, TDF_SLIM
);
388 if (TREE_CODE (step_expr
) != INTEGER_CST
)
390 if (vect_print_dump_info (REPORT_DETAILS
))
391 fprintf (vect_dump
, "step unknown.");
398 /* Function vect_analyze_scalar_cycles_1.
400 Examine the cross iteration def-use cycles of scalar variables
401 in LOOP. LOOP_VINFO represents the loop that is now being
402 considered for vectorization (can be LOOP, or an outer-loop
406 vect_analyze_scalar_cycles_1 (loop_vec_info loop_vinfo
, struct loop
*loop
)
408 basic_block bb
= loop
->header
;
410 VEC(gimple
,heap
) *worklist
= VEC_alloc (gimple
, heap
, 64);
411 gimple_stmt_iterator gsi
;
414 if (vect_print_dump_info (REPORT_DETAILS
))
415 fprintf (vect_dump
, "=== vect_analyze_scalar_cycles ===");
417 /* First - identify all inductions. Reduction detection assumes that all the
418 inductions have been identified, therefore, this order must not be
420 for (gsi
= gsi_start_phis (bb
); !gsi_end_p (gsi
); gsi_next (&gsi
))
422 gimple phi
= gsi_stmt (gsi
);
423 tree access_fn
= NULL
;
424 tree def
= PHI_RESULT (phi
);
425 stmt_vec_info stmt_vinfo
= vinfo_for_stmt (phi
);
427 if (vect_print_dump_info (REPORT_DETAILS
))
429 fprintf (vect_dump
, "Analyze phi: ");
430 print_gimple_stmt (vect_dump
, phi
, 0, TDF_SLIM
);
433 /* Skip virtual phi's. The data dependences that are associated with
434 virtual defs/uses (i.e., memory accesses) are analyzed elsewhere. */
435 if (!is_gimple_reg (SSA_NAME_VAR (def
)))
438 STMT_VINFO_DEF_TYPE (stmt_vinfo
) = vect_unknown_def_type
;
440 /* Analyze the evolution function. */
441 access_fn
= analyze_scalar_evolution (loop
, def
);
442 if (access_fn
&& vect_print_dump_info (REPORT_DETAILS
))
444 fprintf (vect_dump
, "Access function of PHI: ");
445 print_generic_expr (vect_dump
, access_fn
, TDF_SLIM
);
449 || !vect_is_simple_iv_evolution (loop
->num
, access_fn
, &dumy
, &dumy
))
451 VEC_safe_push (gimple
, heap
, worklist
, phi
);
455 if (vect_print_dump_info (REPORT_DETAILS
))
456 fprintf (vect_dump
, "Detected induction.");
457 STMT_VINFO_DEF_TYPE (stmt_vinfo
) = vect_induction_def
;
461 /* Second - identify all reductions and nested cycles. */
462 while (VEC_length (gimple
, worklist
) > 0)
464 gimple phi
= VEC_pop (gimple
, worklist
);
465 tree def
= PHI_RESULT (phi
);
466 stmt_vec_info stmt_vinfo
= vinfo_for_stmt (phi
);
470 if (vect_print_dump_info (REPORT_DETAILS
))
472 fprintf (vect_dump
, "Analyze phi: ");
473 print_gimple_stmt (vect_dump
, phi
, 0, TDF_SLIM
);
476 gcc_assert (is_gimple_reg (SSA_NAME_VAR (def
)));
477 gcc_assert (STMT_VINFO_DEF_TYPE (stmt_vinfo
) == vect_unknown_def_type
);
479 nested_cycle
= (loop
!= LOOP_VINFO_LOOP (loop_vinfo
));
480 reduc_stmt
= vect_is_simple_reduction (loop_vinfo
, phi
, !nested_cycle
,
486 if (vect_print_dump_info (REPORT_DETAILS
))
487 fprintf (vect_dump
, "Detected double reduction.");
489 STMT_VINFO_DEF_TYPE (stmt_vinfo
) = vect_double_reduction_def
;
490 STMT_VINFO_DEF_TYPE (vinfo_for_stmt (reduc_stmt
)) =
491 vect_double_reduction_def
;
497 if (vect_print_dump_info (REPORT_DETAILS
))
498 fprintf (vect_dump
, "Detected vectorizable nested cycle.");
500 STMT_VINFO_DEF_TYPE (stmt_vinfo
) = vect_nested_cycle
;
501 STMT_VINFO_DEF_TYPE (vinfo_for_stmt (reduc_stmt
)) =
506 if (vect_print_dump_info (REPORT_DETAILS
))
507 fprintf (vect_dump
, "Detected reduction.");
509 STMT_VINFO_DEF_TYPE (stmt_vinfo
) = vect_reduction_def
;
510 STMT_VINFO_DEF_TYPE (vinfo_for_stmt (reduc_stmt
)) =
516 if (vect_print_dump_info (REPORT_DETAILS
))
517 fprintf (vect_dump
, "Unknown def-use cycle pattern.");
520 VEC_free (gimple
, heap
, worklist
);
524 /* Function vect_analyze_scalar_cycles.
526 Examine the cross iteration def-use cycles of scalar variables, by
527 analyzing the loop-header PHIs of scalar variables; Classify each
528 cycle as one of the following: invariant, induction, reduction, unknown.
529 We do that for the loop represented by LOOP_VINFO, and also to its
530 inner-loop, if exists.
531 Examples for scalar cycles:
546 vect_analyze_scalar_cycles (loop_vec_info loop_vinfo
)
548 struct loop
*loop
= LOOP_VINFO_LOOP (loop_vinfo
);
550 vect_analyze_scalar_cycles_1 (loop_vinfo
, loop
);
552 /* When vectorizing an outer-loop, the inner-loop is executed sequentially.
553 Reductions in such inner-loop therefore have different properties than
554 the reductions in the nest that gets vectorized:
555 1. When vectorized, they are executed in the same order as in the original
556 scalar loop, so we can't change the order of computation when
558 2. FIXME: Inner-loop reductions can be used in the inner-loop, so the
559 current checks are too strict. */
562 vect_analyze_scalar_cycles_1 (loop_vinfo
, loop
->inner
);
565 /* Function vect_get_loop_niters.
567 Determine how many iterations the loop is executed.
568 If an expression that represents the number of iterations
569 can be constructed, place it in NUMBER_OF_ITERATIONS.
570 Return the loop exit condition. */
573 vect_get_loop_niters (struct loop
*loop
, tree
*number_of_iterations
)
577 if (vect_print_dump_info (REPORT_DETAILS
))
578 fprintf (vect_dump
, "=== get_loop_niters ===");
580 niters
= number_of_exit_cond_executions (loop
);
582 if (niters
!= NULL_TREE
583 && niters
!= chrec_dont_know
)
585 *number_of_iterations
= niters
;
587 if (vect_print_dump_info (REPORT_DETAILS
))
589 fprintf (vect_dump
, "==> get_loop_niters:" );
590 print_generic_expr (vect_dump
, *number_of_iterations
, TDF_SLIM
);
594 return get_loop_exit_condition (loop
);
598 /* Function bb_in_loop_p
600 Used as predicate for dfs order traversal of the loop bbs. */
603 bb_in_loop_p (const_basic_block bb
, const void *data
)
605 const struct loop
*const loop
= (const struct loop
*)data
;
606 if (flow_bb_inside_loop_p (loop
, bb
))
612 /* Function new_loop_vec_info.
614 Create and initialize a new loop_vec_info struct for LOOP, as well as
615 stmt_vec_info structs for all the stmts in LOOP. */
618 new_loop_vec_info (struct loop
*loop
)
622 gimple_stmt_iterator si
;
623 unsigned int i
, nbbs
;
625 res
= (loop_vec_info
) xcalloc (1, sizeof (struct _loop_vec_info
));
626 LOOP_VINFO_LOOP (res
) = loop
;
628 bbs
= get_loop_body (loop
);
630 /* Create/Update stmt_info for all stmts in the loop. */
631 for (i
= 0; i
< loop
->num_nodes
; i
++)
633 basic_block bb
= bbs
[i
];
635 /* BBs in a nested inner-loop will have been already processed (because
636 we will have called vect_analyze_loop_form for any nested inner-loop).
637 Therefore, for stmts in an inner-loop we just want to update the
638 STMT_VINFO_LOOP_VINFO field of their stmt_info to point to the new
639 loop_info of the outer-loop we are currently considering to vectorize
640 (instead of the loop_info of the inner-loop).
641 For stmts in other BBs we need to create a stmt_info from scratch. */
642 if (bb
->loop_father
!= loop
)
645 gcc_assert (loop
->inner
&& bb
->loop_father
== loop
->inner
);
646 for (si
= gsi_start_phis (bb
); !gsi_end_p (si
); gsi_next (&si
))
648 gimple phi
= gsi_stmt (si
);
649 stmt_vec_info stmt_info
= vinfo_for_stmt (phi
);
650 loop_vec_info inner_loop_vinfo
=
651 STMT_VINFO_LOOP_VINFO (stmt_info
);
652 gcc_assert (loop
->inner
== LOOP_VINFO_LOOP (inner_loop_vinfo
));
653 STMT_VINFO_LOOP_VINFO (stmt_info
) = res
;
655 for (si
= gsi_start_bb (bb
); !gsi_end_p (si
); gsi_next (&si
))
657 gimple stmt
= gsi_stmt (si
);
658 stmt_vec_info stmt_info
= vinfo_for_stmt (stmt
);
659 loop_vec_info inner_loop_vinfo
=
660 STMT_VINFO_LOOP_VINFO (stmt_info
);
661 gcc_assert (loop
->inner
== LOOP_VINFO_LOOP (inner_loop_vinfo
));
662 STMT_VINFO_LOOP_VINFO (stmt_info
) = res
;
667 /* bb in current nest. */
668 for (si
= gsi_start_phis (bb
); !gsi_end_p (si
); gsi_next (&si
))
670 gimple phi
= gsi_stmt (si
);
671 gimple_set_uid (phi
, 0);
672 set_vinfo_for_stmt (phi
, new_stmt_vec_info (phi
, res
, NULL
));
675 for (si
= gsi_start_bb (bb
); !gsi_end_p (si
); gsi_next (&si
))
677 gimple stmt
= gsi_stmt (si
);
678 gimple_set_uid (stmt
, 0);
679 set_vinfo_for_stmt (stmt
, new_stmt_vec_info (stmt
, res
, NULL
));
684 /* CHECKME: We want to visit all BBs before their successors (except for
685 latch blocks, for which this assertion wouldn't hold). In the simple
686 case of the loop forms we allow, a dfs order of the BBs would the same
687 as reversed postorder traversal, so we are safe. */
690 bbs
= XCNEWVEC (basic_block
, loop
->num_nodes
);
691 nbbs
= dfs_enumerate_from (loop
->header
, 0, bb_in_loop_p
,
692 bbs
, loop
->num_nodes
, loop
);
693 gcc_assert (nbbs
== loop
->num_nodes
);
695 LOOP_VINFO_BBS (res
) = bbs
;
696 LOOP_VINFO_NITERS (res
) = NULL
;
697 LOOP_VINFO_NITERS_UNCHANGED (res
) = NULL
;
698 LOOP_VINFO_COST_MODEL_MIN_ITERS (res
) = 0;
699 LOOP_VINFO_VECTORIZABLE_P (res
) = 0;
700 LOOP_PEELING_FOR_ALIGNMENT (res
) = 0;
701 LOOP_VINFO_VECT_FACTOR (res
) = 0;
702 LOOP_VINFO_DATAREFS (res
) = VEC_alloc (data_reference_p
, heap
, 10);
703 LOOP_VINFO_DDRS (res
) = VEC_alloc (ddr_p
, heap
, 10 * 10);
704 LOOP_VINFO_UNALIGNED_DR (res
) = NULL
;
705 LOOP_VINFO_MAY_MISALIGN_STMTS (res
) =
706 VEC_alloc (gimple
, heap
,
707 PARAM_VALUE (PARAM_VECT_MAX_VERSION_FOR_ALIGNMENT_CHECKS
));
708 LOOP_VINFO_MAY_ALIAS_DDRS (res
) =
709 VEC_alloc (ddr_p
, heap
,
710 PARAM_VALUE (PARAM_VECT_MAX_VERSION_FOR_ALIAS_CHECKS
));
711 LOOP_VINFO_STRIDED_STORES (res
) = VEC_alloc (gimple
, heap
, 10);
712 LOOP_VINFO_SLP_INSTANCES (res
) = VEC_alloc (slp_instance
, heap
, 10);
713 LOOP_VINFO_SLP_UNROLLING_FACTOR (res
) = 1;
719 /* Function destroy_loop_vec_info.
721 Free LOOP_VINFO struct, as well as all the stmt_vec_info structs of all the
722 stmts in the loop. */
725 destroy_loop_vec_info (loop_vec_info loop_vinfo
, bool clean_stmts
)
730 gimple_stmt_iterator si
;
732 VEC (slp_instance
, heap
) *slp_instances
;
733 slp_instance instance
;
738 loop
= LOOP_VINFO_LOOP (loop_vinfo
);
740 bbs
= LOOP_VINFO_BBS (loop_vinfo
);
741 nbbs
= loop
->num_nodes
;
745 free (LOOP_VINFO_BBS (loop_vinfo
));
746 free_data_refs (LOOP_VINFO_DATAREFS (loop_vinfo
));
747 free_dependence_relations (LOOP_VINFO_DDRS (loop_vinfo
));
748 VEC_free (gimple
, heap
, LOOP_VINFO_MAY_MISALIGN_STMTS (loop_vinfo
));
755 for (j
= 0; j
< nbbs
; j
++)
757 basic_block bb
= bbs
[j
];
758 for (si
= gsi_start_phis (bb
); !gsi_end_p (si
); gsi_next (&si
))
759 free_stmt_vec_info (gsi_stmt (si
));
761 for (si
= gsi_start_bb (bb
); !gsi_end_p (si
); )
763 gimple stmt
= gsi_stmt (si
);
764 stmt_vec_info stmt_info
= vinfo_for_stmt (stmt
);
768 /* Check if this is a "pattern stmt" (introduced by the
769 vectorizer during the pattern recognition pass). */
770 bool remove_stmt_p
= false;
771 gimple orig_stmt
= STMT_VINFO_RELATED_STMT (stmt_info
);
774 stmt_vec_info orig_stmt_info
= vinfo_for_stmt (orig_stmt
);
776 && STMT_VINFO_IN_PATTERN_P (orig_stmt_info
))
777 remove_stmt_p
= true;
780 /* Free stmt_vec_info. */
781 free_stmt_vec_info (stmt
);
783 /* Remove dead "pattern stmts". */
785 gsi_remove (&si
, true);
791 free (LOOP_VINFO_BBS (loop_vinfo
));
792 free_data_refs (LOOP_VINFO_DATAREFS (loop_vinfo
));
793 free_dependence_relations (LOOP_VINFO_DDRS (loop_vinfo
));
794 VEC_free (gimple
, heap
, LOOP_VINFO_MAY_MISALIGN_STMTS (loop_vinfo
));
795 VEC_free (ddr_p
, heap
, LOOP_VINFO_MAY_ALIAS_DDRS (loop_vinfo
));
796 slp_instances
= LOOP_VINFO_SLP_INSTANCES (loop_vinfo
);
797 for (j
= 0; VEC_iterate (slp_instance
, slp_instances
, j
, instance
); j
++)
798 vect_free_slp_instance (instance
);
800 VEC_free (slp_instance
, heap
, LOOP_VINFO_SLP_INSTANCES (loop_vinfo
));
801 VEC_free (gimple
, heap
, LOOP_VINFO_STRIDED_STORES (loop_vinfo
));
808 /* Function vect_analyze_loop_1.
810 Apply a set of analyses on LOOP, and create a loop_vec_info struct
811 for it. The different analyses will record information in the
812 loop_vec_info struct. This is a subset of the analyses applied in
813 vect_analyze_loop, to be applied on an inner-loop nested in the loop
814 that is now considered for (outer-loop) vectorization. */
817 vect_analyze_loop_1 (struct loop
*loop
)
819 loop_vec_info loop_vinfo
;
821 if (vect_print_dump_info (REPORT_DETAILS
))
822 fprintf (vect_dump
, "===== analyze_loop_nest_1 =====");
824 /* Check the CFG characteristics of the loop (nesting, entry/exit, etc. */
826 loop_vinfo
= vect_analyze_loop_form (loop
);
829 if (vect_print_dump_info (REPORT_DETAILS
))
830 fprintf (vect_dump
, "bad inner-loop form.");
838 /* Function vect_analyze_loop_form.
840 Verify that certain CFG restrictions hold, including:
841 - the loop has a pre-header
842 - the loop has a single entry and exit
843 - the loop exit condition is simple enough, and the number of iterations
844 can be analyzed (a countable loop). */
847 vect_analyze_loop_form (struct loop
*loop
)
849 loop_vec_info loop_vinfo
;
851 tree number_of_iterations
= NULL
;
852 loop_vec_info inner_loop_vinfo
= NULL
;
854 if (vect_print_dump_info (REPORT_DETAILS
))
855 fprintf (vect_dump
, "=== vect_analyze_loop_form ===");
857 /* Different restrictions apply when we are considering an inner-most loop,
858 vs. an outer (nested) loop.
859 (FORNOW. May want to relax some of these restrictions in the future). */
863 /* Inner-most loop. We currently require that the number of BBs is
864 exactly 2 (the header and latch). Vectorizable inner-most loops
875 if (loop
->num_nodes
!= 2)
877 if (vect_print_dump_info (REPORT_BAD_FORM_LOOPS
))
878 fprintf (vect_dump
, "not vectorized: control flow in loop.");
882 if (empty_block_p (loop
->header
))
884 if (vect_print_dump_info (REPORT_BAD_FORM_LOOPS
))
885 fprintf (vect_dump
, "not vectorized: empty loop.");
891 struct loop
*innerloop
= loop
->inner
;
892 edge backedge
, entryedge
;
894 /* Nested loop. We currently require that the loop is doubly-nested,
895 contains a single inner loop, and the number of BBs is exactly 5.
896 Vectorizable outer-loops look like this:
908 The inner-loop has the properties expected of inner-most loops
909 as described above. */
911 if ((loop
->inner
)->inner
|| (loop
->inner
)->next
)
913 if (vect_print_dump_info (REPORT_BAD_FORM_LOOPS
))
914 fprintf (vect_dump
, "not vectorized: multiple nested loops.");
918 /* Analyze the inner-loop. */
919 inner_loop_vinfo
= vect_analyze_loop_1 (loop
->inner
);
920 if (!inner_loop_vinfo
)
922 if (vect_print_dump_info (REPORT_BAD_FORM_LOOPS
))
923 fprintf (vect_dump
, "not vectorized: Bad inner loop.");
927 if (!expr_invariant_in_loop_p (loop
,
928 LOOP_VINFO_NITERS (inner_loop_vinfo
)))
930 if (vect_print_dump_info (REPORT_BAD_FORM_LOOPS
))
932 "not vectorized: inner-loop count not invariant.");
933 destroy_loop_vec_info (inner_loop_vinfo
, true);
937 if (loop
->num_nodes
!= 5)
939 if (vect_print_dump_info (REPORT_BAD_FORM_LOOPS
))
940 fprintf (vect_dump
, "not vectorized: control flow in loop.");
941 destroy_loop_vec_info (inner_loop_vinfo
, true);
945 gcc_assert (EDGE_COUNT (innerloop
->header
->preds
) == 2);
946 backedge
= EDGE_PRED (innerloop
->header
, 1);
947 entryedge
= EDGE_PRED (innerloop
->header
, 0);
948 if (EDGE_PRED (innerloop
->header
, 0)->src
== innerloop
->latch
)
950 backedge
= EDGE_PRED (innerloop
->header
, 0);
951 entryedge
= EDGE_PRED (innerloop
->header
, 1);
954 if (entryedge
->src
!= loop
->header
955 || !single_exit (innerloop
)
956 || single_exit (innerloop
)->dest
!= EDGE_PRED (loop
->latch
, 0)->src
)
958 if (vect_print_dump_info (REPORT_BAD_FORM_LOOPS
))
959 fprintf (vect_dump
, "not vectorized: unsupported outerloop form.");
960 destroy_loop_vec_info (inner_loop_vinfo
, true);
964 if (vect_print_dump_info (REPORT_DETAILS
))
965 fprintf (vect_dump
, "Considering outer-loop vectorization.");
968 if (!single_exit (loop
)
969 || EDGE_COUNT (loop
->header
->preds
) != 2)
971 if (vect_print_dump_info (REPORT_BAD_FORM_LOOPS
))
973 if (!single_exit (loop
))
974 fprintf (vect_dump
, "not vectorized: multiple exits.");
975 else if (EDGE_COUNT (loop
->header
->preds
) != 2)
976 fprintf (vect_dump
, "not vectorized: too many incoming edges.");
978 if (inner_loop_vinfo
)
979 destroy_loop_vec_info (inner_loop_vinfo
, true);
983 /* We assume that the loop exit condition is at the end of the loop. i.e,
984 that the loop is represented as a do-while (with a proper if-guard
985 before the loop if needed), where the loop header contains all the
986 executable statements, and the latch is empty. */
987 if (!empty_block_p (loop
->latch
)
988 || phi_nodes (loop
->latch
))
990 if (vect_print_dump_info (REPORT_BAD_FORM_LOOPS
))
991 fprintf (vect_dump
, "not vectorized: unexpected loop form.");
992 if (inner_loop_vinfo
)
993 destroy_loop_vec_info (inner_loop_vinfo
, true);
997 /* Make sure there exists a single-predecessor exit bb: */
998 if (!single_pred_p (single_exit (loop
)->dest
))
1000 edge e
= single_exit (loop
);
1001 if (!(e
->flags
& EDGE_ABNORMAL
))
1003 split_loop_exit_edge (e
);
1004 if (vect_print_dump_info (REPORT_DETAILS
))
1005 fprintf (vect_dump
, "split exit edge.");
1009 if (vect_print_dump_info (REPORT_BAD_FORM_LOOPS
))
1010 fprintf (vect_dump
, "not vectorized: abnormal loop exit edge.");
1011 if (inner_loop_vinfo
)
1012 destroy_loop_vec_info (inner_loop_vinfo
, true);
1017 loop_cond
= vect_get_loop_niters (loop
, &number_of_iterations
);
1020 if (vect_print_dump_info (REPORT_BAD_FORM_LOOPS
))
1021 fprintf (vect_dump
, "not vectorized: complicated exit condition.");
1022 if (inner_loop_vinfo
)
1023 destroy_loop_vec_info (inner_loop_vinfo
, true);
1027 if (!number_of_iterations
)
1029 if (vect_print_dump_info (REPORT_BAD_FORM_LOOPS
))
1031 "not vectorized: number of iterations cannot be computed.");
1032 if (inner_loop_vinfo
)
1033 destroy_loop_vec_info (inner_loop_vinfo
, true);
1037 if (chrec_contains_undetermined (number_of_iterations
))
1039 if (vect_print_dump_info (REPORT_BAD_FORM_LOOPS
))
1040 fprintf (vect_dump
, "Infinite number of iterations.");
1041 if (inner_loop_vinfo
)
1042 destroy_loop_vec_info (inner_loop_vinfo
, true);
1046 if (!NITERS_KNOWN_P (number_of_iterations
))
1048 if (vect_print_dump_info (REPORT_DETAILS
))
1050 fprintf (vect_dump
, "Symbolic number of iterations is ");
1051 print_generic_expr (vect_dump
, number_of_iterations
, TDF_DETAILS
);
1054 else if (TREE_INT_CST_LOW (number_of_iterations
) == 0)
1056 if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS
))
1057 fprintf (vect_dump
, "not vectorized: number of iterations = 0.");
1058 if (inner_loop_vinfo
)
1059 destroy_loop_vec_info (inner_loop_vinfo
, false);
1063 loop_vinfo
= new_loop_vec_info (loop
);
1064 LOOP_VINFO_NITERS (loop_vinfo
) = number_of_iterations
;
1065 LOOP_VINFO_NITERS_UNCHANGED (loop_vinfo
) = number_of_iterations
;
1067 STMT_VINFO_TYPE (vinfo_for_stmt (loop_cond
)) = loop_exit_ctrl_vec_info_type
;
1069 /* CHECKME: May want to keep it around it in the future. */
1070 if (inner_loop_vinfo
)
1071 destroy_loop_vec_info (inner_loop_vinfo
, false);
1073 gcc_assert (!loop
->aux
);
1074 loop
->aux
= loop_vinfo
;
1079 /* Function vect_analyze_loop_operations.
1081 Scan the loop stmts and make sure they are all vectorizable. */
1084 vect_analyze_loop_operations (loop_vec_info loop_vinfo
)
1086 struct loop
*loop
= LOOP_VINFO_LOOP (loop_vinfo
);
1087 basic_block
*bbs
= LOOP_VINFO_BBS (loop_vinfo
);
1088 int nbbs
= loop
->num_nodes
;
1089 gimple_stmt_iterator si
;
1090 unsigned int vectorization_factor
= 0;
1093 stmt_vec_info stmt_info
;
1094 bool need_to_vectorize
= false;
1095 int min_profitable_iters
;
1096 int min_scalar_loop_bound
;
1098 bool only_slp_in_loop
= true, ok
;
1100 if (vect_print_dump_info (REPORT_DETAILS
))
1101 fprintf (vect_dump
, "=== vect_analyze_loop_operations ===");
1103 gcc_assert (LOOP_VINFO_VECT_FACTOR (loop_vinfo
));
1104 vectorization_factor
= LOOP_VINFO_VECT_FACTOR (loop_vinfo
);
1106 for (i
= 0; i
< nbbs
; i
++)
1108 basic_block bb
= bbs
[i
];
1110 for (si
= gsi_start_phis (bb
); !gsi_end_p (si
); gsi_next (&si
))
1112 phi
= gsi_stmt (si
);
1115 stmt_info
= vinfo_for_stmt (phi
);
1116 if (vect_print_dump_info (REPORT_DETAILS
))
1118 fprintf (vect_dump
, "examining phi: ");
1119 print_gimple_stmt (vect_dump
, phi
, 0, TDF_SLIM
);
1122 if (! is_loop_header_bb_p (bb
))
1124 /* inner-loop loop-closed exit phi in outer-loop vectorization
1125 (i.e. a phi in the tail of the outer-loop).
1126 FORNOW: we currently don't support the case that these phis
1127 are not used in the outerloop (unless it is double reduction,
1128 i.e., this phi is vect_reduction_def), cause this case
1129 requires to actually do something here. */
1130 if ((!STMT_VINFO_RELEVANT_P (stmt_info
)
1131 || STMT_VINFO_LIVE_P (stmt_info
))
1132 && STMT_VINFO_DEF_TYPE (stmt_info
)
1133 != vect_double_reduction_def
)
1135 if (vect_print_dump_info (REPORT_DETAILS
))
1137 "Unsupported loop-closed phi in outer-loop.");
1143 gcc_assert (stmt_info
);
1145 if (STMT_VINFO_LIVE_P (stmt_info
))
1147 /* FORNOW: not yet supported. */
1148 if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS
))
1149 fprintf (vect_dump
, "not vectorized: value used after loop.");
1153 if (STMT_VINFO_RELEVANT (stmt_info
) == vect_used_in_scope
1154 && STMT_VINFO_DEF_TYPE (stmt_info
) != vect_induction_def
)
1156 /* A scalar-dependence cycle that we don't support. */
1157 if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS
))
1158 fprintf (vect_dump
, "not vectorized: scalar dependence cycle.");
1162 if (STMT_VINFO_RELEVANT_P (stmt_info
))
1164 need_to_vectorize
= true;
1165 if (STMT_VINFO_DEF_TYPE (stmt_info
) == vect_induction_def
)
1166 ok
= vectorizable_induction (phi
, NULL
, NULL
);
1171 if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS
))
1174 "not vectorized: relevant phi not supported: ");
1175 print_gimple_stmt (vect_dump
, phi
, 0, TDF_SLIM
);
1181 for (si
= gsi_start_bb (bb
); !gsi_end_p (si
); gsi_next (&si
))
1183 gimple stmt
= gsi_stmt (si
);
1184 stmt_vec_info stmt_info
= vinfo_for_stmt (stmt
);
1186 gcc_assert (stmt_info
);
1188 if (!vect_analyze_stmt (stmt
, &need_to_vectorize
, NULL
))
1191 if (STMT_VINFO_RELEVANT_P (stmt_info
) && !PURE_SLP_STMT (stmt_info
))
1192 /* STMT needs both SLP and loop-based vectorization. */
1193 only_slp_in_loop
= false;
1197 /* All operations in the loop are either irrelevant (deal with loop
1198 control, or dead), or only used outside the loop and can be moved
1199 out of the loop (e.g. invariants, inductions). The loop can be
1200 optimized away by scalar optimizations. We're better off not
1201 touching this loop. */
1202 if (!need_to_vectorize
)
1204 if (vect_print_dump_info (REPORT_DETAILS
))
1206 "All the computation can be taken out of the loop.");
1207 if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS
))
1209 "not vectorized: redundant loop. no profit to vectorize.");
1213 /* If all the stmts in the loop can be SLPed, we perform only SLP, and
1214 vectorization factor of the loop is the unrolling factor required by the
1215 SLP instances. If that unrolling factor is 1, we say, that we perform
1216 pure SLP on loop - cross iteration parallelism is not exploited. */
1217 if (only_slp_in_loop
)
1218 vectorization_factor
= LOOP_VINFO_SLP_UNROLLING_FACTOR (loop_vinfo
);
1220 vectorization_factor
= least_common_multiple (vectorization_factor
,
1221 LOOP_VINFO_SLP_UNROLLING_FACTOR (loop_vinfo
));
1223 LOOP_VINFO_VECT_FACTOR (loop_vinfo
) = vectorization_factor
;
1225 if (LOOP_VINFO_NITERS_KNOWN_P (loop_vinfo
)
1226 && vect_print_dump_info (REPORT_DETAILS
))
1228 "vectorization_factor = %d, niters = " HOST_WIDE_INT_PRINT_DEC
,
1229 vectorization_factor
, LOOP_VINFO_INT_NITERS (loop_vinfo
));
1231 if (LOOP_VINFO_NITERS_KNOWN_P (loop_vinfo
)
1232 && (LOOP_VINFO_INT_NITERS (loop_vinfo
) < vectorization_factor
))
1234 if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS
))
1235 fprintf (vect_dump
, "not vectorized: iteration count too small.");
1236 if (vect_print_dump_info (REPORT_DETAILS
))
1237 fprintf (vect_dump
,"not vectorized: iteration count smaller than "
1238 "vectorization factor.");
1242 /* Analyze cost. Decide if worth while to vectorize. */
1244 /* Once VF is set, SLP costs should be updated since the number of created
1245 vector stmts depends on VF. */
1246 vect_update_slp_costs_according_to_vf (loop_vinfo
);
1248 min_profitable_iters
= vect_estimate_min_profitable_iters (loop_vinfo
);
1249 LOOP_VINFO_COST_MODEL_MIN_ITERS (loop_vinfo
) = min_profitable_iters
;
1251 if (min_profitable_iters
< 0)
1253 if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS
))
1254 fprintf (vect_dump
, "not vectorized: vectorization not profitable.");
1255 if (vect_print_dump_info (REPORT_DETAILS
))
1256 fprintf (vect_dump
, "not vectorized: vector version will never be "
1261 min_scalar_loop_bound
= ((PARAM_VALUE (PARAM_MIN_VECT_LOOP_BOUND
)
1262 * vectorization_factor
) - 1);
1264 /* Use the cost model only if it is more conservative than user specified
1267 th
= (unsigned) min_scalar_loop_bound
;
1268 if (min_profitable_iters
1269 && (!min_scalar_loop_bound
1270 || min_profitable_iters
> min_scalar_loop_bound
))
1271 th
= (unsigned) min_profitable_iters
;
1273 if (LOOP_VINFO_NITERS_KNOWN_P (loop_vinfo
)
1274 && LOOP_VINFO_INT_NITERS (loop_vinfo
) <= th
)
1276 if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS
))
1277 fprintf (vect_dump
, "not vectorized: vectorization not "
1279 if (vect_print_dump_info (REPORT_DETAILS
))
1280 fprintf (vect_dump
, "not vectorized: iteration count smaller than "
1281 "user specified loop bound parameter or minimum "
1282 "profitable iterations (whichever is more conservative).");
1286 if (!LOOP_VINFO_NITERS_KNOWN_P (loop_vinfo
)
1287 || LOOP_VINFO_INT_NITERS (loop_vinfo
) % vectorization_factor
!= 0
1288 || LOOP_PEELING_FOR_ALIGNMENT (loop_vinfo
))
1290 if (vect_print_dump_info (REPORT_DETAILS
))
1291 fprintf (vect_dump
, "epilog loop required.");
1292 if (!vect_can_advance_ivs_p (loop_vinfo
))
1294 if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS
))
1296 "not vectorized: can't create epilog loop 1.");
1299 if (!slpeel_can_duplicate_loop_p (loop
, single_exit (loop
)))
1301 if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS
))
1303 "not vectorized: can't create epilog loop 2.");
1312 /* Function vect_analyze_loop.
1314 Apply a set of analyses on LOOP, and create a loop_vec_info struct
1315 for it. The different analyses will record information in the
1316 loop_vec_info struct. */
1318 vect_analyze_loop (struct loop
*loop
)
1321 loop_vec_info loop_vinfo
;
1323 if (vect_print_dump_info (REPORT_DETAILS
))
1324 fprintf (vect_dump
, "===== analyze_loop_nest =====");
1326 if (loop_outer (loop
)
1327 && loop_vec_info_for_loop (loop_outer (loop
))
1328 && LOOP_VINFO_VECTORIZABLE_P (loop_vec_info_for_loop (loop_outer (loop
))))
1330 if (vect_print_dump_info (REPORT_DETAILS
))
1331 fprintf (vect_dump
, "outer-loop already vectorized.");
1335 /* Check the CFG characteristics of the loop (nesting, entry/exit, etc. */
1337 loop_vinfo
= vect_analyze_loop_form (loop
);
1340 if (vect_print_dump_info (REPORT_DETAILS
))
1341 fprintf (vect_dump
, "bad loop form.");
1345 /* Find all data references in the loop (which correspond to vdefs/vuses)
1346 and analyze their evolution in the loop.
1348 FORNOW: Handle only simple, array references, which
1349 alignment can be forced, and aligned pointer-references. */
1351 ok
= vect_analyze_data_refs (loop_vinfo
, NULL
);
1354 if (vect_print_dump_info (REPORT_DETAILS
))
1355 fprintf (vect_dump
, "bad data references.");
1356 destroy_loop_vec_info (loop_vinfo
, true);
1360 /* Classify all cross-iteration scalar data-flow cycles.
1361 Cross-iteration cycles caused by virtual phis are analyzed separately. */
1363 vect_analyze_scalar_cycles (loop_vinfo
);
1365 vect_pattern_recog (loop_vinfo
);
1367 /* Data-flow analysis to detect stmts that do not need to be vectorized. */
1369 ok
= vect_mark_stmts_to_be_vectorized (loop_vinfo
);
1372 if (vect_print_dump_info (REPORT_DETAILS
))
1373 fprintf (vect_dump
, "unexpected pattern.");
1374 destroy_loop_vec_info (loop_vinfo
, true);
1378 /* Analyze the alignment of the data-refs in the loop.
1379 Fail if a data reference is found that cannot be vectorized. */
1381 ok
= vect_analyze_data_refs_alignment (loop_vinfo
, NULL
);
1384 if (vect_print_dump_info (REPORT_DETAILS
))
1385 fprintf (vect_dump
, "bad data alignment.");
1386 destroy_loop_vec_info (loop_vinfo
, true);
1390 ok
= vect_determine_vectorization_factor (loop_vinfo
);
1393 if (vect_print_dump_info (REPORT_DETAILS
))
1394 fprintf (vect_dump
, "can't determine vectorization factor.");
1395 destroy_loop_vec_info (loop_vinfo
, true);
1399 /* Analyze data dependences between the data-refs in the loop.
1400 FORNOW: fail at the first data dependence that we encounter. */
1402 ok
= vect_analyze_data_ref_dependences (loop_vinfo
, NULL
);
1405 if (vect_print_dump_info (REPORT_DETAILS
))
1406 fprintf (vect_dump
, "bad data dependence.");
1407 destroy_loop_vec_info (loop_vinfo
, true);
1411 /* Analyze the access patterns of the data-refs in the loop (consecutive,
1412 complex, etc.). FORNOW: Only handle consecutive access pattern. */
1414 ok
= vect_analyze_data_ref_accesses (loop_vinfo
, NULL
);
1417 if (vect_print_dump_info (REPORT_DETAILS
))
1418 fprintf (vect_dump
, "bad data access.");
1419 destroy_loop_vec_info (loop_vinfo
, true);
1423 /* Prune the list of ddrs to be tested at run-time by versioning for alias.
1424 It is important to call pruning after vect_analyze_data_ref_accesses,
1425 since we use grouping information gathered by interleaving analysis. */
1426 ok
= vect_prune_runtime_alias_test_list (loop_vinfo
);
1429 if (vect_print_dump_info (REPORT_DETAILS
))
1430 fprintf (vect_dump
, "too long list of versioning for alias "
1432 destroy_loop_vec_info (loop_vinfo
, true);
1436 /* Check the SLP opportunities in the loop, analyze and build SLP trees. */
1437 ok
= vect_analyze_slp (loop_vinfo
, NULL
);
1440 /* Decide which possible SLP instances to SLP. */
1441 vect_make_slp_decision (loop_vinfo
);
1443 /* Find stmts that need to be both vectorized and SLPed. */
1444 vect_detect_hybrid_slp (loop_vinfo
);
1447 /* This pass will decide on using loop versioning and/or loop peeling in
1448 order to enhance the alignment of data references in the loop. */
1450 ok
= vect_enhance_data_refs_alignment (loop_vinfo
);
1453 if (vect_print_dump_info (REPORT_DETAILS
))
1454 fprintf (vect_dump
, "bad data alignment.");
1455 destroy_loop_vec_info (loop_vinfo
, true);
1459 /* Scan all the operations in the loop and make sure they are
1462 ok
= vect_analyze_loop_operations (loop_vinfo
);
1465 if (vect_print_dump_info (REPORT_DETAILS
))
1466 fprintf (vect_dump
, "bad operation or unsupported loop bound.");
1467 destroy_loop_vec_info (loop_vinfo
, true);
1471 LOOP_VINFO_VECTORIZABLE_P (loop_vinfo
) = 1;
1477 /* Function reduction_code_for_scalar_code
1480 CODE - tree_code of a reduction operations.
1483 REDUC_CODE - the corresponding tree-code to be used to reduce the
1484 vector of partial results into a single scalar result (which
1485 will also reside in a vector) or ERROR_MARK if the operation is
1486 a supported reduction operation, but does not have such tree-code.
1488 Return FALSE if CODE currently cannot be vectorized as reduction. */
1491 reduction_code_for_scalar_code (enum tree_code code
,
1492 enum tree_code
*reduc_code
)
1497 *reduc_code
= REDUC_MAX_EXPR
;
1501 *reduc_code
= REDUC_MIN_EXPR
;
1505 *reduc_code
= REDUC_PLUS_EXPR
;
1513 *reduc_code
= ERROR_MARK
;
1522 /* Error reporting helper for vect_is_simple_reduction below. GIMPLE statement
1523 STMT is printed with a message MSG. */
1526 report_vect_op (gimple stmt
, const char *msg
)
1528 fprintf (vect_dump
, "%s", msg
);
1529 print_gimple_stmt (vect_dump
, stmt
, 0, TDF_SLIM
);
1533 /* Function vect_is_simple_reduction
1535 (1) Detect a cross-iteration def-use cycle that represents a simple
1536 reduction computation. We look for the following pattern:
1541 a2 = operation (a3, a1)
1544 1. operation is commutative and associative and it is safe to
1545 change the order of the computation (if CHECK_REDUCTION is true)
1546 2. no uses for a2 in the loop (a2 is used out of the loop)
1547 3. no uses of a1 in the loop besides the reduction operation.
1549 Condition 1 is tested here.
1550 Conditions 2,3 are tested in vect_mark_stmts_to_be_vectorized.
1552 (2) Detect a cross-iteration def-use cycle in nested loops, i.e.,
1553 nested cycles, if CHECK_REDUCTION is false.
1555 (3) Detect cycles of phi nodes in outer-loop vectorization, i.e., double
1559 inner loop (def of a3)
1564 vect_is_simple_reduction (loop_vec_info loop_info
, gimple phi
,
1565 bool check_reduction
, bool *double_reduc
)
1567 struct loop
*loop
= (gimple_bb (phi
))->loop_father
;
1568 struct loop
*vect_loop
= LOOP_VINFO_LOOP (loop_info
);
1569 edge latch_e
= loop_latch_edge (loop
);
1570 tree loop_arg
= PHI_ARG_DEF_FROM_EDGE (phi
, latch_e
);
1571 gimple def_stmt
, def1
, def2
;
1572 enum tree_code code
;
1577 imm_use_iterator imm_iter
;
1578 use_operand_p use_p
;
1581 *double_reduc
= false;
1583 /* If CHECK_REDUCTION is true, we assume inner-most loop vectorization,
1584 otherwise, we assume outer loop vectorization. */
1585 gcc_assert ((check_reduction
&& loop
== vect_loop
)
1586 || (!check_reduction
&& flow_loop_nested_p (vect_loop
, loop
)));
1588 name
= PHI_RESULT (phi
);
1590 FOR_EACH_IMM_USE_FAST (use_p
, imm_iter
, name
)
1592 gimple use_stmt
= USE_STMT (use_p
);
1593 if (flow_bb_inside_loop_p (loop
, gimple_bb (use_stmt
))
1594 && vinfo_for_stmt (use_stmt
)
1595 && !is_pattern_stmt_p (vinfo_for_stmt (use_stmt
)))
1599 if (vect_print_dump_info (REPORT_DETAILS
))
1600 fprintf (vect_dump
, "reduction used in loop.");
1605 if (TREE_CODE (loop_arg
) != SSA_NAME
)
1607 if (vect_print_dump_info (REPORT_DETAILS
))
1609 fprintf (vect_dump
, "reduction: not ssa_name: ");
1610 print_generic_expr (vect_dump
, loop_arg
, TDF_SLIM
);
1615 def_stmt
= SSA_NAME_DEF_STMT (loop_arg
);
1618 if (vect_print_dump_info (REPORT_DETAILS
))
1619 fprintf (vect_dump
, "reduction: no def_stmt.");
1623 if (!is_gimple_assign (def_stmt
) && gimple_code (def_stmt
) != GIMPLE_PHI
)
1625 if (vect_print_dump_info (REPORT_DETAILS
))
1626 print_gimple_stmt (vect_dump
, def_stmt
, 0, TDF_SLIM
);
1630 if (is_gimple_assign (def_stmt
))
1632 name
= gimple_assign_lhs (def_stmt
);
1637 name
= PHI_RESULT (def_stmt
);
1642 FOR_EACH_IMM_USE_FAST (use_p
, imm_iter
, name
)
1644 gimple use_stmt
= USE_STMT (use_p
);
1645 if (flow_bb_inside_loop_p (loop
, gimple_bb (use_stmt
))
1646 && vinfo_for_stmt (use_stmt
)
1647 && !is_pattern_stmt_p (vinfo_for_stmt (use_stmt
)))
1651 if (vect_print_dump_info (REPORT_DETAILS
))
1652 fprintf (vect_dump
, "reduction used in loop.");
1657 /* If DEF_STMT is a phi node itself, we expect it to have a single argument
1658 defined in the inner loop. */
1661 op1
= PHI_ARG_DEF (def_stmt
, 0);
1663 if (gimple_phi_num_args (def_stmt
) != 1
1664 || TREE_CODE (op1
) != SSA_NAME
)
1666 if (vect_print_dump_info (REPORT_DETAILS
))
1667 fprintf (vect_dump
, "unsupported phi node definition.");
1672 def1
= SSA_NAME_DEF_STMT (op1
);
1673 if (flow_bb_inside_loop_p (loop
, gimple_bb (def_stmt
))
1675 && flow_bb_inside_loop_p (loop
->inner
, gimple_bb (def1
))
1676 && is_gimple_assign (def1
))
1678 if (vect_print_dump_info (REPORT_DETAILS
))
1679 report_vect_op (def_stmt
, "detected double reduction: ");
1681 *double_reduc
= true;
1688 code
= gimple_assign_rhs_code (def_stmt
);
1691 && (!commutative_tree_code (code
) || !associative_tree_code (code
)))
1693 if (vect_print_dump_info (REPORT_DETAILS
))
1694 report_vect_op (def_stmt
, "reduction: not commutative/associative: ");
1698 if (get_gimple_rhs_class (code
) != GIMPLE_BINARY_RHS
)
1700 if (vect_print_dump_info (REPORT_DETAILS
))
1701 report_vect_op (def_stmt
, "reduction: not binary operation: ");
1705 op1
= gimple_assign_rhs1 (def_stmt
);
1706 op2
= gimple_assign_rhs2 (def_stmt
);
1707 if (TREE_CODE (op1
) != SSA_NAME
|| TREE_CODE (op2
) != SSA_NAME
)
1709 if (vect_print_dump_info (REPORT_DETAILS
))
1710 report_vect_op (def_stmt
, "reduction: uses not ssa_names: ");
1714 type
= TREE_TYPE (gimple_assign_lhs (def_stmt
));
1715 if (TYPE_MAIN_VARIANT (type
) != TYPE_MAIN_VARIANT (TREE_TYPE (op1
))
1716 || TYPE_MAIN_VARIANT (type
) != TYPE_MAIN_VARIANT (TREE_TYPE (op2
)))
1718 if (vect_print_dump_info (REPORT_DETAILS
))
1720 fprintf (vect_dump
, "reduction: multiple types: operation type: ");
1721 print_generic_expr (vect_dump
, type
, TDF_SLIM
);
1722 fprintf (vect_dump
, ", operands types: ");
1723 print_generic_expr (vect_dump
, TREE_TYPE (op1
), TDF_SLIM
);
1724 fprintf (vect_dump
, ",");
1725 print_generic_expr (vect_dump
, TREE_TYPE (op2
), TDF_SLIM
);
1730 /* Check that it's ok to change the order of the computation.
1731 Generally, when vectorizing a reduction we change the order of the
1732 computation. This may change the behavior of the program in some
1733 cases, so we need to check that this is ok. One exception is when
1734 vectorizing an outer-loop: the inner-loop is executed sequentially,
1735 and therefore vectorizing reductions in the inner-loop during
1736 outer-loop vectorization is safe. */
1738 /* CHECKME: check for !flag_finite_math_only too? */
1739 if (SCALAR_FLOAT_TYPE_P (type
) && !flag_associative_math
1742 /* Changing the order of operations changes the semantics. */
1743 if (vect_print_dump_info (REPORT_DETAILS
))
1744 report_vect_op (def_stmt
, "reduction: unsafe fp math optimization: ");
1747 else if (INTEGRAL_TYPE_P (type
) && TYPE_OVERFLOW_TRAPS (type
)
1750 /* Changing the order of operations changes the semantics. */
1751 if (vect_print_dump_info (REPORT_DETAILS
))
1752 report_vect_op (def_stmt
, "reduction: unsafe int math optimization: ");
1755 else if (SAT_FIXED_POINT_TYPE_P (type
) && check_reduction
)
1757 /* Changing the order of operations changes the semantics. */
1758 if (vect_print_dump_info (REPORT_DETAILS
))
1759 report_vect_op (def_stmt
,
1760 "reduction: unsafe fixed-point math optimization: ");
1764 /* Reduction is safe. We're dealing with one of the following:
1765 1) integer arithmetic and no trapv
1766 2) floating point arithmetic, and special flags permit this optimization
1767 3) nested cycle (i.e., outer loop vectorization). */
1768 def1
= SSA_NAME_DEF_STMT (op1
);
1769 def2
= SSA_NAME_DEF_STMT (op2
);
1770 if (!def1
|| !def2
|| gimple_nop_p (def1
) || gimple_nop_p (def2
))
1772 if (vect_print_dump_info (REPORT_DETAILS
))
1773 report_vect_op (def_stmt
, "reduction: no defs for operands: ");
1777 /* Check that one def is the reduction def, defined by PHI,
1778 the other def is either defined in the loop ("vect_internal_def"),
1779 or it's an induction (defined by a loop-header phi-node). */
1782 && flow_bb_inside_loop_p (loop
, gimple_bb (def1
))
1783 && (is_gimple_assign (def1
)
1784 || STMT_VINFO_DEF_TYPE (vinfo_for_stmt (def1
)) == vect_induction_def
1785 || (gimple_code (def1
) == GIMPLE_PHI
1786 && STMT_VINFO_DEF_TYPE (vinfo_for_stmt (def1
))
1787 == vect_internal_def
1788 && !is_loop_header_bb_p (gimple_bb (def1
)))))
1790 if (vect_print_dump_info (REPORT_DETAILS
))
1791 report_vect_op (def_stmt
, "detected reduction: ");
1794 else if (def1
== phi
1795 && flow_bb_inside_loop_p (loop
, gimple_bb (def2
))
1796 && (is_gimple_assign (def2
)
1797 || STMT_VINFO_DEF_TYPE (vinfo_for_stmt (def2
))
1798 == vect_induction_def
1799 || (gimple_code (def2
) == GIMPLE_PHI
1800 && STMT_VINFO_DEF_TYPE (vinfo_for_stmt (def2
))
1801 == vect_internal_def
1802 && !is_loop_header_bb_p (gimple_bb (def2
)))))
1804 if (check_reduction
)
1806 /* Swap operands (just for simplicity - so that the rest of the code
1807 can assume that the reduction variable is always the last (second)
1809 if (vect_print_dump_info (REPORT_DETAILS
))
1810 report_vect_op (def_stmt
,
1811 "detected reduction: need to swap operands: ");
1813 swap_tree_operands (def_stmt
, gimple_assign_rhs1_ptr (def_stmt
),
1814 gimple_assign_rhs2_ptr (def_stmt
));
1818 if (vect_print_dump_info (REPORT_DETAILS
))
1819 report_vect_op (def_stmt
, "detected reduction: ");
1826 if (vect_print_dump_info (REPORT_DETAILS
))
1827 report_vect_op (def_stmt
, "reduction: unknown pattern: ");
1834 /* Function vect_estimate_min_profitable_iters
1836 Return the number of iterations required for the vector version of the
1837 loop to be profitable relative to the cost of the scalar version of the
1840 TODO: Take profile info into account before making vectorization
1841 decisions, if available. */
1844 vect_estimate_min_profitable_iters (loop_vec_info loop_vinfo
)
1847 int min_profitable_iters
;
1848 int peel_iters_prologue
;
1849 int peel_iters_epilogue
;
1850 int vec_inside_cost
= 0;
1851 int vec_outside_cost
= 0;
1852 int scalar_single_iter_cost
= 0;
1853 int scalar_outside_cost
= 0;
1854 int vf
= LOOP_VINFO_VECT_FACTOR (loop_vinfo
);
1855 struct loop
*loop
= LOOP_VINFO_LOOP (loop_vinfo
);
1856 basic_block
*bbs
= LOOP_VINFO_BBS (loop_vinfo
);
1857 int nbbs
= loop
->num_nodes
;
1858 int byte_misalign
= LOOP_PEELING_FOR_ALIGNMENT (loop_vinfo
);
1859 int peel_guard_costs
= 0;
1860 int innerloop_iters
= 0, factor
;
1861 VEC (slp_instance
, heap
) *slp_instances
;
1862 slp_instance instance
;
1864 /* Cost model disabled. */
1865 if (!flag_vect_cost_model
)
1867 if (vect_print_dump_info (REPORT_COST
))
1868 fprintf (vect_dump
, "cost model disabled.");
1872 /* Requires loop versioning tests to handle misalignment. */
1873 if (LOOP_REQUIRES_VERSIONING_FOR_ALIGNMENT (loop_vinfo
))
1875 /* FIXME: Make cost depend on complexity of individual check. */
1877 VEC_length (gimple
, LOOP_VINFO_MAY_MISALIGN_STMTS (loop_vinfo
));
1878 if (vect_print_dump_info (REPORT_COST
))
1879 fprintf (vect_dump
, "cost model: Adding cost of checks for loop "
1880 "versioning to treat misalignment.\n");
1883 /* Requires loop versioning with alias checks. */
1884 if (LOOP_REQUIRES_VERSIONING_FOR_ALIAS (loop_vinfo
))
1886 /* FIXME: Make cost depend on complexity of individual check. */
1888 VEC_length (ddr_p
, LOOP_VINFO_MAY_ALIAS_DDRS (loop_vinfo
));
1889 if (vect_print_dump_info (REPORT_COST
))
1890 fprintf (vect_dump
, "cost model: Adding cost of checks for loop "
1891 "versioning aliasing.\n");
1894 if (LOOP_REQUIRES_VERSIONING_FOR_ALIGNMENT (loop_vinfo
)
1895 || LOOP_REQUIRES_VERSIONING_FOR_ALIAS (loop_vinfo
))
1896 vec_outside_cost
+= TARG_COND_TAKEN_BRANCH_COST
;
1898 /* Count statements in scalar loop. Using this as scalar cost for a single
1901 TODO: Add outer loop support.
1903 TODO: Consider assigning different costs to different scalar
1908 innerloop_iters
= 50; /* FIXME */
1910 for (i
= 0; i
< nbbs
; i
++)
1912 gimple_stmt_iterator si
;
1913 basic_block bb
= bbs
[i
];
1915 if (bb
->loop_father
== loop
->inner
)
1916 factor
= innerloop_iters
;
1920 for (si
= gsi_start_bb (bb
); !gsi_end_p (si
); gsi_next (&si
))
1922 gimple stmt
= gsi_stmt (si
);
1923 stmt_vec_info stmt_info
= vinfo_for_stmt (stmt
);
1924 /* Skip stmts that are not vectorized inside the loop. */
1925 if (!STMT_VINFO_RELEVANT_P (stmt_info
)
1926 && (!STMT_VINFO_LIVE_P (stmt_info
)
1927 || STMT_VINFO_DEF_TYPE (stmt_info
) != vect_reduction_def
))
1929 scalar_single_iter_cost
+= cost_for_stmt (stmt
) * factor
;
1930 vec_inside_cost
+= STMT_VINFO_INSIDE_OF_LOOP_COST (stmt_info
) * factor
;
1931 /* FIXME: for stmts in the inner-loop in outer-loop vectorization,
1932 some of the "outside" costs are generated inside the outer-loop. */
1933 vec_outside_cost
+= STMT_VINFO_OUTSIDE_OF_LOOP_COST (stmt_info
);
1937 /* Add additional cost for the peeled instructions in prologue and epilogue
1940 FORNOW: If we don't know the value of peel_iters for prologue or epilogue
1941 at compile-time - we assume it's vf/2 (the worst would be vf-1).
1943 TODO: Build an expression that represents peel_iters for prologue and
1944 epilogue to be used in a run-time test. */
1946 if (byte_misalign
< 0)
1948 peel_iters_prologue
= vf
/2;
1949 if (vect_print_dump_info (REPORT_COST
))
1950 fprintf (vect_dump
, "cost model: "
1951 "prologue peel iters set to vf/2.");
1953 /* If peeling for alignment is unknown, loop bound of main loop becomes
1955 peel_iters_epilogue
= vf
/2;
1956 if (vect_print_dump_info (REPORT_COST
))
1957 fprintf (vect_dump
, "cost model: "
1958 "epilogue peel iters set to vf/2 because "
1959 "peeling for alignment is unknown .");
1961 /* If peeled iterations are unknown, count a taken branch and a not taken
1962 branch per peeled loop. Even if scalar loop iterations are known,
1963 vector iterations are not known since peeled prologue iterations are
1964 not known. Hence guards remain the same. */
1965 peel_guard_costs
+= 2 * (TARG_COND_TAKEN_BRANCH_COST
1966 + TARG_COND_NOT_TAKEN_BRANCH_COST
);
1972 struct data_reference
*dr
= LOOP_VINFO_UNALIGNED_DR (loop_vinfo
);
1973 int element_size
= GET_MODE_SIZE (TYPE_MODE (TREE_TYPE (DR_REF (dr
))));
1974 tree vectype
= STMT_VINFO_VECTYPE (vinfo_for_stmt (DR_STMT (dr
)));
1975 int nelements
= TYPE_VECTOR_SUBPARTS (vectype
);
1977 peel_iters_prologue
= nelements
- (byte_misalign
/ element_size
);
1980 peel_iters_prologue
= 0;
1982 if (!LOOP_VINFO_NITERS_KNOWN_P (loop_vinfo
))
1984 peel_iters_epilogue
= vf
/2;
1985 if (vect_print_dump_info (REPORT_COST
))
1986 fprintf (vect_dump
, "cost model: "
1987 "epilogue peel iters set to vf/2 because "
1988 "loop iterations are unknown .");
1990 /* If peeled iterations are known but number of scalar loop
1991 iterations are unknown, count a taken branch per peeled loop. */
1992 peel_guard_costs
+= 2 * TARG_COND_TAKEN_BRANCH_COST
;
1997 int niters
= LOOP_VINFO_INT_NITERS (loop_vinfo
);
1998 peel_iters_prologue
= niters
< peel_iters_prologue
?
1999 niters
: peel_iters_prologue
;
2000 peel_iters_epilogue
= (niters
- peel_iters_prologue
) % vf
;
2004 vec_outside_cost
+= (peel_iters_prologue
* scalar_single_iter_cost
)
2005 + (peel_iters_epilogue
* scalar_single_iter_cost
)
2008 /* FORNOW: The scalar outside cost is incremented in one of the
2011 1. The vectorizer checks for alignment and aliasing and generates
2012 a condition that allows dynamic vectorization. A cost model
2013 check is ANDED with the versioning condition. Hence scalar code
2014 path now has the added cost of the versioning check.
2016 if (cost > th & versioning_check)
2019 Hence run-time scalar is incremented by not-taken branch cost.
2021 2. The vectorizer then checks if a prologue is required. If the
2022 cost model check was not done before during versioning, it has to
2023 be done before the prologue check.
2026 prologue = scalar_iters
2031 if (prologue == num_iters)
2034 Hence the run-time scalar cost is incremented by a taken branch,
2035 plus a not-taken branch, plus a taken branch cost.
2037 3. The vectorizer then checks if an epilogue is required. If the
2038 cost model check was not done before during prologue check, it
2039 has to be done with the epilogue check.
2045 if (prologue == num_iters)
2048 if ((cost <= th) | (scalar_iters-prologue-epilogue == 0))
2051 Hence the run-time scalar cost should be incremented by 2 taken
2054 TODO: The back end may reorder the BBS's differently and reverse
2055 conditions/branch directions. Change the estimates below to
2056 something more reasonable. */
2058 /* If the number of iterations is known and we do not do versioning, we can
2059 decide whether to vectorize at compile time. Hence the scalar version
2060 do not carry cost model guard costs. */
2061 if (!LOOP_VINFO_NITERS_KNOWN_P (loop_vinfo
)
2062 || LOOP_REQUIRES_VERSIONING_FOR_ALIGNMENT (loop_vinfo
)
2063 || LOOP_REQUIRES_VERSIONING_FOR_ALIAS (loop_vinfo
))
2065 /* Cost model check occurs at versioning. */
2066 if (LOOP_REQUIRES_VERSIONING_FOR_ALIGNMENT (loop_vinfo
)
2067 || LOOP_REQUIRES_VERSIONING_FOR_ALIAS (loop_vinfo
))
2068 scalar_outside_cost
+= TARG_COND_NOT_TAKEN_BRANCH_COST
;
2071 /* Cost model check occurs at prologue generation. */
2072 if (LOOP_PEELING_FOR_ALIGNMENT (loop_vinfo
) < 0)
2073 scalar_outside_cost
+= 2 * TARG_COND_TAKEN_BRANCH_COST
2074 + TARG_COND_NOT_TAKEN_BRANCH_COST
;
2075 /* Cost model check occurs at epilogue generation. */
2077 scalar_outside_cost
+= 2 * TARG_COND_TAKEN_BRANCH_COST
;
2081 /* Add SLP costs. */
2082 slp_instances
= LOOP_VINFO_SLP_INSTANCES (loop_vinfo
);
2083 for (i
= 0; VEC_iterate (slp_instance
, slp_instances
, i
, instance
); i
++)
2085 vec_outside_cost
+= SLP_INSTANCE_OUTSIDE_OF_LOOP_COST (instance
);
2086 vec_inside_cost
+= SLP_INSTANCE_INSIDE_OF_LOOP_COST (instance
);
2089 /* Calculate number of iterations required to make the vector version
2090 profitable, relative to the loop bodies only. The following condition
2092 SIC * niters + SOC > VIC * ((niters-PL_ITERS-EP_ITERS)/VF) + VOC
2094 SIC = scalar iteration cost, VIC = vector iteration cost,
2095 VOC = vector outside cost, VF = vectorization factor,
2096 PL_ITERS = prologue iterations, EP_ITERS= epilogue iterations
2097 SOC = scalar outside cost for run time cost model check. */
2099 if ((scalar_single_iter_cost
* vf
) > vec_inside_cost
)
2101 if (vec_outside_cost
<= 0)
2102 min_profitable_iters
= 1;
2105 min_profitable_iters
= ((vec_outside_cost
- scalar_outside_cost
) * vf
2106 - vec_inside_cost
* peel_iters_prologue
2107 - vec_inside_cost
* peel_iters_epilogue
)
2108 / ((scalar_single_iter_cost
* vf
)
2111 if ((scalar_single_iter_cost
* vf
* min_profitable_iters
)
2112 <= ((vec_inside_cost
* min_profitable_iters
)
2113 + ((vec_outside_cost
- scalar_outside_cost
) * vf
)))
2114 min_profitable_iters
++;
2117 /* vector version will never be profitable. */
2120 if (vect_print_dump_info (REPORT_COST
))
2121 fprintf (vect_dump
, "cost model: vector iteration cost = %d "
2122 "is divisible by scalar iteration cost = %d by a factor "
2123 "greater than or equal to the vectorization factor = %d .",
2124 vec_inside_cost
, scalar_single_iter_cost
, vf
);
2128 if (vect_print_dump_info (REPORT_COST
))
2130 fprintf (vect_dump
, "Cost model analysis: \n");
2131 fprintf (vect_dump
, " Vector inside of loop cost: %d\n",
2133 fprintf (vect_dump
, " Vector outside of loop cost: %d\n",
2135 fprintf (vect_dump
, " Scalar iteration cost: %d\n",
2136 scalar_single_iter_cost
);
2137 fprintf (vect_dump
, " Scalar outside cost: %d\n", scalar_outside_cost
);
2138 fprintf (vect_dump
, " prologue iterations: %d\n",
2139 peel_iters_prologue
);
2140 fprintf (vect_dump
, " epilogue iterations: %d\n",
2141 peel_iters_epilogue
);
2142 fprintf (vect_dump
, " Calculated minimum iters for profitability: %d\n",
2143 min_profitable_iters
);
2146 min_profitable_iters
=
2147 min_profitable_iters
< vf
? vf
: min_profitable_iters
;
2149 /* Because the condition we create is:
2150 if (niters <= min_profitable_iters)
2151 then skip the vectorized loop. */
2152 min_profitable_iters
--;
2154 if (vect_print_dump_info (REPORT_COST
))
2155 fprintf (vect_dump
, " Profitability threshold = %d\n",
2156 min_profitable_iters
);
2158 return min_profitable_iters
;
2162 /* TODO: Close dependency between vect_model_*_cost and vectorizable_*
2163 functions. Design better to avoid maintenance issues. */
2165 /* Function vect_model_reduction_cost.
2167 Models cost for a reduction operation, including the vector ops
2168 generated within the strip-mine loop, the initial definition before
2169 the loop, and the epilogue code that must be generated. */
2172 vect_model_reduction_cost (stmt_vec_info stmt_info
, enum tree_code reduc_code
,
2176 enum tree_code code
;
2179 gimple stmt
, orig_stmt
;
2181 enum machine_mode mode
;
2182 loop_vec_info loop_vinfo
= STMT_VINFO_LOOP_VINFO (stmt_info
);
2183 struct loop
*loop
= LOOP_VINFO_LOOP (loop_vinfo
);
2186 /* Cost of reduction op inside loop. */
2187 STMT_VINFO_INSIDE_OF_LOOP_COST (stmt_info
) += ncopies
* TARG_VEC_STMT_COST
;
2189 stmt
= STMT_VINFO_STMT (stmt_info
);
2191 switch (get_gimple_rhs_class (gimple_assign_rhs_code (stmt
)))
2193 case GIMPLE_SINGLE_RHS
:
2194 gcc_assert (TREE_OPERAND_LENGTH (gimple_assign_rhs1 (stmt
)) == ternary_op
);
2195 reduction_op
= TREE_OPERAND (gimple_assign_rhs1 (stmt
), 2);
2197 case GIMPLE_UNARY_RHS
:
2198 reduction_op
= gimple_assign_rhs1 (stmt
);
2200 case GIMPLE_BINARY_RHS
:
2201 reduction_op
= gimple_assign_rhs2 (stmt
);
2207 vectype
= get_vectype_for_scalar_type (TREE_TYPE (reduction_op
));
2210 if (vect_print_dump_info (REPORT_COST
))
2212 fprintf (vect_dump
, "unsupported data-type ");
2213 print_generic_expr (vect_dump
, TREE_TYPE (reduction_op
), TDF_SLIM
);
2218 mode
= TYPE_MODE (vectype
);
2219 orig_stmt
= STMT_VINFO_RELATED_STMT (stmt_info
);
2222 orig_stmt
= STMT_VINFO_STMT (stmt_info
);
2224 code
= gimple_assign_rhs_code (orig_stmt
);
2226 /* Add in cost for initial definition. */
2227 outer_cost
+= TARG_SCALAR_TO_VEC_COST
;
2229 /* Determine cost of epilogue code.
2231 We have a reduction operator that will reduce the vector in one statement.
2232 Also requires scalar extract. */
2234 if (!nested_in_vect_loop_p (loop
, orig_stmt
))
2236 if (reduc_code
!= ERROR_MARK
)
2237 outer_cost
+= TARG_VEC_STMT_COST
+ TARG_VEC_TO_SCALAR_COST
;
2240 int vec_size_in_bits
= tree_low_cst (TYPE_SIZE (vectype
), 1);
2242 TYPE_SIZE (TREE_TYPE (gimple_assign_lhs (orig_stmt
)));
2243 int element_bitsize
= tree_low_cst (bitsize
, 1);
2244 int nelements
= vec_size_in_bits
/ element_bitsize
;
2246 optab
= optab_for_tree_code (code
, vectype
, optab_default
);
2248 /* We have a whole vector shift available. */
2249 if (VECTOR_MODE_P (mode
)
2250 && optab_handler (optab
, mode
)->insn_code
!= CODE_FOR_nothing
2251 && optab_handler (vec_shr_optab
, mode
)->insn_code
!= CODE_FOR_nothing
)
2252 /* Final reduction via vector shifts and the reduction operator. Also
2253 requires scalar extract. */
2254 outer_cost
+= ((exact_log2(nelements
) * 2) * TARG_VEC_STMT_COST
2255 + TARG_VEC_TO_SCALAR_COST
);
2257 /* Use extracts and reduction op for final reduction. For N elements,
2258 we have N extracts and N-1 reduction ops. */
2259 outer_cost
+= ((nelements
+ nelements
- 1) * TARG_VEC_STMT_COST
);
2263 STMT_VINFO_OUTSIDE_OF_LOOP_COST (stmt_info
) = outer_cost
;
2265 if (vect_print_dump_info (REPORT_COST
))
2266 fprintf (vect_dump
, "vect_model_reduction_cost: inside_cost = %d, "
2267 "outside_cost = %d .", STMT_VINFO_INSIDE_OF_LOOP_COST (stmt_info
),
2268 STMT_VINFO_OUTSIDE_OF_LOOP_COST (stmt_info
));
2274 /* Function vect_model_induction_cost.
2276 Models cost for induction operations. */
2279 vect_model_induction_cost (stmt_vec_info stmt_info
, int ncopies
)
2281 /* loop cost for vec_loop. */
2282 STMT_VINFO_INSIDE_OF_LOOP_COST (stmt_info
) = ncopies
* TARG_VEC_STMT_COST
;
2283 /* prologue cost for vec_init and vec_step. */
2284 STMT_VINFO_OUTSIDE_OF_LOOP_COST (stmt_info
) = 2 * TARG_SCALAR_TO_VEC_COST
;
2286 if (vect_print_dump_info (REPORT_COST
))
2287 fprintf (vect_dump
, "vect_model_induction_cost: inside_cost = %d, "
2288 "outside_cost = %d .", STMT_VINFO_INSIDE_OF_LOOP_COST (stmt_info
),
2289 STMT_VINFO_OUTSIDE_OF_LOOP_COST (stmt_info
));
2293 /* Function get_initial_def_for_induction
2296 STMT - a stmt that performs an induction operation in the loop.
2297 IV_PHI - the initial value of the induction variable
2300 Return a vector variable, initialized with the first VF values of
2301 the induction variable. E.g., for an iv with IV_PHI='X' and
2302 evolution S, for a vector of 4 units, we want to return:
2303 [X, X + S, X + 2*S, X + 3*S]. */
2306 get_initial_def_for_induction (gimple iv_phi
)
2308 stmt_vec_info stmt_vinfo
= vinfo_for_stmt (iv_phi
);
2309 loop_vec_info loop_vinfo
= STMT_VINFO_LOOP_VINFO (stmt_vinfo
);
2310 struct loop
*loop
= LOOP_VINFO_LOOP (loop_vinfo
);
2311 tree scalar_type
= TREE_TYPE (gimple_phi_result (iv_phi
));
2314 edge pe
= loop_preheader_edge (loop
);
2315 struct loop
*iv_loop
;
2317 tree vec
, vec_init
, vec_step
, t
;
2321 gimple init_stmt
, induction_phi
, new_stmt
;
2322 tree induc_def
, vec_def
, vec_dest
;
2323 tree init_expr
, step_expr
;
2324 int vf
= LOOP_VINFO_VECT_FACTOR (loop_vinfo
);
2329 stmt_vec_info phi_info
= vinfo_for_stmt (iv_phi
);
2330 bool nested_in_vect_loop
= false;
2331 gimple_seq stmts
= NULL
;
2332 imm_use_iterator imm_iter
;
2333 use_operand_p use_p
;
2337 gimple_stmt_iterator si
;
2338 basic_block bb
= gimple_bb (iv_phi
);
2341 vectype
= get_vectype_for_scalar_type (scalar_type
);
2342 gcc_assert (vectype
);
2343 nunits
= TYPE_VECTOR_SUBPARTS (vectype
);
2344 ncopies
= vf
/ nunits
;
2346 gcc_assert (phi_info
);
2347 gcc_assert (ncopies
>= 1);
2349 /* Find the first insertion point in the BB. */
2350 si
= gsi_after_labels (bb
);
2352 if (INTEGRAL_TYPE_P (scalar_type
))
2353 step_expr
= build_int_cst (scalar_type
, 0);
2354 else if (POINTER_TYPE_P (scalar_type
))
2355 step_expr
= build_int_cst (sizetype
, 0);
2357 step_expr
= build_real (scalar_type
, dconst0
);
2359 /* Is phi in an inner-loop, while vectorizing an enclosing outer-loop? */
2360 if (nested_in_vect_loop_p (loop
, iv_phi
))
2362 nested_in_vect_loop
= true;
2363 iv_loop
= loop
->inner
;
2367 gcc_assert (iv_loop
== (gimple_bb (iv_phi
))->loop_father
);
2369 latch_e
= loop_latch_edge (iv_loop
);
2370 loop_arg
= PHI_ARG_DEF_FROM_EDGE (iv_phi
, latch_e
);
2372 access_fn
= analyze_scalar_evolution (iv_loop
, PHI_RESULT (iv_phi
));
2373 gcc_assert (access_fn
);
2374 ok
= vect_is_simple_iv_evolution (iv_loop
->num
, access_fn
,
2375 &init_expr
, &step_expr
);
2377 pe
= loop_preheader_edge (iv_loop
);
2379 /* Create the vector that holds the initial_value of the induction. */
2380 if (nested_in_vect_loop
)
2382 /* iv_loop is nested in the loop to be vectorized. init_expr had already
2383 been created during vectorization of previous stmts; We obtain it from
2384 the STMT_VINFO_VEC_STMT of the defining stmt. */
2385 tree iv_def
= PHI_ARG_DEF_FROM_EDGE (iv_phi
,
2386 loop_preheader_edge (iv_loop
));
2387 vec_init
= vect_get_vec_def_for_operand (iv_def
, iv_phi
, NULL
);
2391 /* iv_loop is the loop to be vectorized. Create:
2392 vec_init = [X, X+S, X+2*S, X+3*S] (S = step_expr, X = init_expr) */
2393 new_var
= vect_get_new_vect_var (scalar_type
, vect_scalar_var
, "var_");
2394 add_referenced_var (new_var
);
2396 new_name
= force_gimple_operand (init_expr
, &stmts
, false, new_var
);
2399 new_bb
= gsi_insert_seq_on_edge_immediate (pe
, stmts
);
2400 gcc_assert (!new_bb
);
2404 t
= tree_cons (NULL_TREE
, init_expr
, t
);
2405 for (i
= 1; i
< nunits
; i
++)
2407 /* Create: new_name_i = new_name + step_expr */
2408 enum tree_code code
= POINTER_TYPE_P (scalar_type
)
2409 ? POINTER_PLUS_EXPR
: PLUS_EXPR
;
2410 init_stmt
= gimple_build_assign_with_ops (code
, new_var
,
2411 new_name
, step_expr
);
2412 new_name
= make_ssa_name (new_var
, init_stmt
);
2413 gimple_assign_set_lhs (init_stmt
, new_name
);
2415 new_bb
= gsi_insert_on_edge_immediate (pe
, init_stmt
);
2416 gcc_assert (!new_bb
);
2418 if (vect_print_dump_info (REPORT_DETAILS
))
2420 fprintf (vect_dump
, "created new init_stmt: ");
2421 print_gimple_stmt (vect_dump
, init_stmt
, 0, TDF_SLIM
);
2423 t
= tree_cons (NULL_TREE
, new_name
, t
);
2425 /* Create a vector from [new_name_0, new_name_1, ..., new_name_nunits-1] */
2426 vec
= build_constructor_from_list (vectype
, nreverse (t
));
2427 vec_init
= vect_init_vector (iv_phi
, vec
, vectype
, NULL
);
2431 /* Create the vector that holds the step of the induction. */
2432 if (nested_in_vect_loop
)
2433 /* iv_loop is nested in the loop to be vectorized. Generate:
2434 vec_step = [S, S, S, S] */
2435 new_name
= step_expr
;
2438 /* iv_loop is the loop to be vectorized. Generate:
2439 vec_step = [VF*S, VF*S, VF*S, VF*S] */
2440 expr
= build_int_cst (TREE_TYPE (step_expr
), vf
);
2441 new_name
= fold_build2 (MULT_EXPR
, TREE_TYPE (step_expr
),
2446 for (i
= 0; i
< nunits
; i
++)
2447 t
= tree_cons (NULL_TREE
, unshare_expr (new_name
), t
);
2448 gcc_assert (CONSTANT_CLASS_P (new_name
));
2449 stepvectype
= get_vectype_for_scalar_type (TREE_TYPE (new_name
));
2450 gcc_assert (stepvectype
);
2451 vec
= build_vector (stepvectype
, t
);
2452 vec_step
= vect_init_vector (iv_phi
, vec
, stepvectype
, NULL
);
2455 /* Create the following def-use cycle:
2460 vec_iv = PHI <vec_init, vec_loop>
2464 vec_loop = vec_iv + vec_step; */
2466 /* Create the induction-phi that defines the induction-operand. */
2467 vec_dest
= vect_get_new_vect_var (vectype
, vect_simple_var
, "vec_iv_");
2468 add_referenced_var (vec_dest
);
2469 induction_phi
= create_phi_node (vec_dest
, iv_loop
->header
);
2470 set_vinfo_for_stmt (induction_phi
,
2471 new_stmt_vec_info (induction_phi
, loop_vinfo
, NULL
));
2472 induc_def
= PHI_RESULT (induction_phi
);
2474 /* Create the iv update inside the loop */
2475 new_stmt
= gimple_build_assign_with_ops (PLUS_EXPR
, vec_dest
,
2476 induc_def
, vec_step
);
2477 vec_def
= make_ssa_name (vec_dest
, new_stmt
);
2478 gimple_assign_set_lhs (new_stmt
, vec_def
);
2479 gsi_insert_before (&si
, new_stmt
, GSI_SAME_STMT
);
2480 set_vinfo_for_stmt (new_stmt
, new_stmt_vec_info (new_stmt
, loop_vinfo
,
2483 /* Set the arguments of the phi node: */
2484 add_phi_arg (induction_phi
, vec_init
, pe
);
2485 add_phi_arg (induction_phi
, vec_def
, loop_latch_edge (iv_loop
));
2488 /* In case that vectorization factor (VF) is bigger than the number
2489 of elements that we can fit in a vectype (nunits), we have to generate
2490 more than one vector stmt - i.e - we need to "unroll" the
2491 vector stmt by a factor VF/nunits. For more details see documentation
2492 in vectorizable_operation. */
2496 stmt_vec_info prev_stmt_vinfo
;
2497 /* FORNOW. This restriction should be relaxed. */
2498 gcc_assert (!nested_in_vect_loop
);
2500 /* Create the vector that holds the step of the induction. */
2501 expr
= build_int_cst (TREE_TYPE (step_expr
), nunits
);
2502 new_name
= fold_build2 (MULT_EXPR
, TREE_TYPE (step_expr
),
2505 for (i
= 0; i
< nunits
; i
++)
2506 t
= tree_cons (NULL_TREE
, unshare_expr (new_name
), t
);
2507 gcc_assert (CONSTANT_CLASS_P (new_name
));
2508 vec
= build_vector (stepvectype
, t
);
2509 vec_step
= vect_init_vector (iv_phi
, vec
, stepvectype
, NULL
);
2511 vec_def
= induc_def
;
2512 prev_stmt_vinfo
= vinfo_for_stmt (induction_phi
);
2513 for (i
= 1; i
< ncopies
; i
++)
2515 /* vec_i = vec_prev + vec_step */
2516 new_stmt
= gimple_build_assign_with_ops (PLUS_EXPR
, vec_dest
,
2518 vec_def
= make_ssa_name (vec_dest
, new_stmt
);
2519 gimple_assign_set_lhs (new_stmt
, vec_def
);
2521 gsi_insert_before (&si
, new_stmt
, GSI_SAME_STMT
);
2522 set_vinfo_for_stmt (new_stmt
,
2523 new_stmt_vec_info (new_stmt
, loop_vinfo
, NULL
));
2524 STMT_VINFO_RELATED_STMT (prev_stmt_vinfo
) = new_stmt
;
2525 prev_stmt_vinfo
= vinfo_for_stmt (new_stmt
);
2529 if (nested_in_vect_loop
)
2531 /* Find the loop-closed exit-phi of the induction, and record
2532 the final vector of induction results: */
2534 FOR_EACH_IMM_USE_FAST (use_p
, imm_iter
, loop_arg
)
2536 if (!flow_bb_inside_loop_p (iv_loop
, gimple_bb (USE_STMT (use_p
))))
2538 exit_phi
= USE_STMT (use_p
);
2544 stmt_vec_info stmt_vinfo
= vinfo_for_stmt (exit_phi
);
2545 /* FORNOW. Currently not supporting the case that an inner-loop induction
2546 is not used in the outer-loop (i.e. only outside the outer-loop). */
2547 gcc_assert (STMT_VINFO_RELEVANT_P (stmt_vinfo
)
2548 && !STMT_VINFO_LIVE_P (stmt_vinfo
));
2550 STMT_VINFO_VEC_STMT (stmt_vinfo
) = new_stmt
;
2551 if (vect_print_dump_info (REPORT_DETAILS
))
2553 fprintf (vect_dump
, "vector of inductions after inner-loop:");
2554 print_gimple_stmt (vect_dump
, new_stmt
, 0, TDF_SLIM
);
2560 if (vect_print_dump_info (REPORT_DETAILS
))
2562 fprintf (vect_dump
, "transform induction: created def-use cycle: ");
2563 print_gimple_stmt (vect_dump
, induction_phi
, 0, TDF_SLIM
);
2564 fprintf (vect_dump
, "\n");
2565 print_gimple_stmt (vect_dump
, SSA_NAME_DEF_STMT (vec_def
), 0, TDF_SLIM
);
2568 STMT_VINFO_VEC_STMT (phi_info
) = induction_phi
;
2573 /* Function get_initial_def_for_reduction
2576 STMT - a stmt that performs a reduction operation in the loop.
2577 INIT_VAL - the initial value of the reduction variable
2580 ADJUSTMENT_DEF - a tree that holds a value to be added to the final result
2581 of the reduction (used for adjusting the epilog - see below).
2582 Return a vector variable, initialized according to the operation that STMT
2583 performs. This vector will be used as the initial value of the
2584 vector of partial results.
2586 Option1 (adjust in epilog): Initialize the vector as follows:
2587 add/bit or/xor: [0,0,...,0,0]
2588 mult/bit and: [1,1,...,1,1]
2589 min/max: [init_val,init_val,..,init_val,init_val]
2590 and when necessary (e.g. add/mult case) let the caller know
2591 that it needs to adjust the result by init_val.
2593 Option2: Initialize the vector as follows:
2594 add/bit or/xor: [init_val,0,0,...,0]
2595 mult/bit and: [init_val,1,1,...,1]
2596 min/max: [init_val,init_val,...,init_val]
2597 and no adjustments are needed.
2599 For example, for the following code:
2605 STMT is 's = s + a[i]', and the reduction variable is 's'.
2606 For a vector of 4 units, we want to return either [0,0,0,init_val],
2607 or [0,0,0,0] and let the caller know that it needs to adjust
2608 the result at the end by 'init_val'.
2610 FORNOW, we are using the 'adjust in epilog' scheme, because this way the
2611 initialization vector is simpler (same element in all entries), if
2612 ADJUSTMENT_DEF is not NULL, and Option2 otherwise.
2614 A cost model should help decide between these two schemes. */
2617 get_initial_def_for_reduction (gimple stmt
, tree init_val
,
2618 tree
*adjustment_def
)
2620 stmt_vec_info stmt_vinfo
= vinfo_for_stmt (stmt
);
2621 loop_vec_info loop_vinfo
= STMT_VINFO_LOOP_VINFO (stmt_vinfo
);
2622 struct loop
*loop
= LOOP_VINFO_LOOP (loop_vinfo
);
2623 tree scalar_type
= TREE_TYPE (init_val
);
2624 tree vectype
= get_vectype_for_scalar_type (scalar_type
);
2626 enum tree_code code
= gimple_assign_rhs_code (stmt
);
2631 bool nested_in_vect_loop
= false;
2633 REAL_VALUE_TYPE real_init_val
= dconst0
;
2634 int int_init_val
= 0;
2636 gcc_assert (vectype
);
2637 nunits
= TYPE_VECTOR_SUBPARTS (vectype
);
2639 gcc_assert (POINTER_TYPE_P (scalar_type
) || INTEGRAL_TYPE_P (scalar_type
)
2640 || SCALAR_FLOAT_TYPE_P (scalar_type
));
2642 if (nested_in_vect_loop_p (loop
, stmt
))
2643 nested_in_vect_loop
= true;
2645 gcc_assert (loop
== (gimple_bb (stmt
))->loop_father
);
2647 /* In case of double reduction we only create a vector variable to be put
2648 in the reduction phi node. The actual statement creation is done in
2649 vect_create_epilog_for_reduction. */
2650 if (TREE_CODE (init_val
) == SSA_NAME
2651 && vinfo_for_stmt (SSA_NAME_DEF_STMT (init_val
))
2652 && STMT_VINFO_DEF_TYPE (vinfo_for_stmt (SSA_NAME_DEF_STMT (init_val
)))
2653 == vect_double_reduction_def
)
2655 *adjustment_def
= NULL
;
2656 return vect_create_destination_var (init_val
, vectype
);
2659 if (TREE_CONSTANT (init_val
))
2661 if (SCALAR_FLOAT_TYPE_P (scalar_type
))
2662 init_value
= build_real (scalar_type
, TREE_REAL_CST (init_val
));
2664 init_value
= build_int_cst (scalar_type
, TREE_INT_CST_LOW (init_val
));
2667 init_value
= init_val
;
2671 case WIDEN_SUM_EXPR
:
2679 /* ADJUSMENT_DEF is NULL when called from
2680 vect_create_epilog_for_reduction to vectorize double reduction. */
2683 if (nested_in_vect_loop
)
2684 *adjustment_def
= vect_get_vec_def_for_operand (init_val
, stmt
,
2687 *adjustment_def
= init_val
;
2690 if (code
== MULT_EXPR
|| code
== BIT_AND_EXPR
)
2692 real_init_val
= dconst1
;
2696 if (SCALAR_FLOAT_TYPE_P (scalar_type
))
2697 def_for_init
= build_real (scalar_type
, real_init_val
);
2699 def_for_init
= build_int_cst (scalar_type
, int_init_val
);
2701 /* Create a vector of '0' or '1' except the first element. */
2702 for (i
= nunits
- 2; i
>= 0; --i
)
2703 t
= tree_cons (NULL_TREE
, def_for_init
, t
);
2705 /* Option1: the first element is '0' or '1' as well. */
2708 t
= tree_cons (NULL_TREE
, def_for_init
, t
);
2709 init_def
= build_vector (vectype
, t
);
2713 /* Option2: the first element is INIT_VAL. */
2714 t
= tree_cons (NULL_TREE
, init_value
, t
);
2715 if (TREE_CONSTANT (init_val
))
2716 init_def
= build_vector (vectype
, t
);
2718 init_def
= build_constructor_from_list (vectype
, t
);
2726 *adjustment_def
= NULL_TREE
;
2727 init_def
= vect_get_vec_def_for_operand (init_val
, stmt
, NULL
);
2731 for (i
= nunits
- 1; i
>= 0; --i
)
2732 t
= tree_cons (NULL_TREE
, init_value
, t
);
2734 if (TREE_CONSTANT (init_val
))
2735 init_def
= build_vector (vectype
, t
);
2737 init_def
= build_constructor_from_list (vectype
, t
);
2749 /* Function vect_create_epilog_for_reduction
2751 Create code at the loop-epilog to finalize the result of a reduction
2754 VECT_DEF is a vector of partial results.
2755 REDUC_CODE is the tree-code for the epilog reduction.
2756 NCOPIES is > 1 in case the vectorization factor (VF) is bigger than the
2757 number of elements that we can fit in a vectype (nunits). In this case
2758 we have to generate more than one vector stmt - i.e - we need to "unroll"
2759 the vector stmt by a factor VF/nunits. For more details see documentation
2760 in vectorizable_operation.
2761 STMT is the scalar reduction stmt that is being vectorized.
2762 REDUCTION_PHI is the phi-node that carries the reduction computation.
2763 REDUC_INDEX is the index of the operand in the right hand side of the
2764 statement that is defined by REDUCTION_PHI.
2765 DOUBLE_REDUC is TRUE if double reduction phi nodes should be handled.
2768 1. Creates the reduction def-use cycle: sets the arguments for
2770 The loop-entry argument is the vectorized initial-value of the reduction.
2771 The loop-latch argument is VECT_DEF - the vector of partial sums.
2772 2. "Reduces" the vector of partial results VECT_DEF into a single result,
2773 by applying the operation specified by REDUC_CODE if available, or by
2774 other means (whole-vector shifts or a scalar loop).
2775 The function also creates a new phi node at the loop exit to preserve
2776 loop-closed form, as illustrated below.
2778 The flow at the entry to this function:
2781 vec_def = phi <null, null> # REDUCTION_PHI
2782 VECT_DEF = vector_stmt # vectorized form of STMT
2783 s_loop = scalar_stmt # (scalar) STMT
2785 s_out0 = phi <s_loop> # (scalar) EXIT_PHI
2789 The above is transformed by this function into:
2792 vec_def = phi <vec_init, VECT_DEF> # REDUCTION_PHI
2793 VECT_DEF = vector_stmt # vectorized form of STMT
2794 s_loop = scalar_stmt # (scalar) STMT
2796 s_out0 = phi <s_loop> # (scalar) EXIT_PHI
2797 v_out1 = phi <VECT_DEF> # NEW_EXIT_PHI
2798 v_out2 = reduce <v_out1>
2799 s_out3 = extract_field <v_out2, 0>
2800 s_out4 = adjust_result <s_out3>
2806 vect_create_epilog_for_reduction (tree vect_def
, gimple stmt
,
2808 enum tree_code reduc_code
,
2809 gimple reduction_phi
,
2813 stmt_vec_info stmt_info
= vinfo_for_stmt (stmt
);
2814 stmt_vec_info prev_phi_info
;
2816 enum machine_mode mode
;
2817 loop_vec_info loop_vinfo
= STMT_VINFO_LOOP_VINFO (stmt_info
);
2818 struct loop
*loop
= LOOP_VINFO_LOOP (loop_vinfo
), *outer_loop
= NULL
;
2819 basic_block exit_bb
;
2822 gimple new_phi
= NULL
, phi
;
2823 gimple_stmt_iterator exit_gsi
;
2825 tree new_temp
= NULL_TREE
;
2827 gimple epilog_stmt
= NULL
;
2828 tree new_scalar_dest
, new_dest
;
2830 tree bitsize
, bitpos
, bytesize
;
2831 enum tree_code code
= gimple_assign_rhs_code (stmt
);
2832 tree adjustment_def
;
2833 tree vec_initial_def
, def
;
2835 imm_use_iterator imm_iter
;
2836 use_operand_p use_p
;
2837 bool extract_scalar_result
= false;
2838 tree reduction_op
, expr
;
2841 bool nested_in_vect_loop
= false;
2842 VEC(gimple
,heap
) *phis
= NULL
;
2843 enum vect_def_type dt
= vect_unknown_def_type
;
2846 if (nested_in_vect_loop_p (loop
, stmt
))
2850 nested_in_vect_loop
= true;
2853 switch (get_gimple_rhs_class (gimple_assign_rhs_code (stmt
)))
2855 case GIMPLE_SINGLE_RHS
:
2856 gcc_assert (TREE_OPERAND_LENGTH (gimple_assign_rhs1 (stmt
))
2858 reduction_op
= TREE_OPERAND (gimple_assign_rhs1 (stmt
), reduc_index
);
2860 case GIMPLE_UNARY_RHS
:
2861 reduction_op
= gimple_assign_rhs1 (stmt
);
2863 case GIMPLE_BINARY_RHS
:
2864 reduction_op
= reduc_index
?
2865 gimple_assign_rhs2 (stmt
) : gimple_assign_rhs1 (stmt
);
2871 vectype
= get_vectype_for_scalar_type (TREE_TYPE (reduction_op
));
2872 gcc_assert (vectype
);
2873 mode
= TYPE_MODE (vectype
);
2875 /*** 1. Create the reduction def-use cycle ***/
2877 /* For the case of reduction, vect_get_vec_def_for_operand returns
2878 the scalar def before the loop, that defines the initial value
2879 of the reduction variable. */
2880 vec_initial_def
= vect_get_vec_def_for_operand (reduction_op
, stmt
,
2883 phi
= reduction_phi
;
2885 for (j
= 0; j
< ncopies
; j
++)
2887 /* 1.1 set the loop-entry arg of the reduction-phi: */
2888 add_phi_arg (phi
, vec_initial_def
, loop_preheader_edge (loop
));
2890 /* 1.2 set the loop-latch arg for the reduction-phi: */
2892 def
= vect_get_vec_def_for_stmt_copy (dt
, def
);
2893 add_phi_arg (phi
, def
, loop_latch_edge (loop
));
2895 if (vect_print_dump_info (REPORT_DETAILS
))
2897 fprintf (vect_dump
, "transform reduction: created def-use cycle: ");
2898 print_gimple_stmt (vect_dump
, phi
, 0, TDF_SLIM
);
2899 fprintf (vect_dump
, "\n");
2900 print_gimple_stmt (vect_dump
, SSA_NAME_DEF_STMT (def
), 0, TDF_SLIM
);
2903 phi
= STMT_VINFO_RELATED_STMT (vinfo_for_stmt (phi
));
2906 /*** 2. Create epilog code
2907 The reduction epilog code operates across the elements of the vector
2908 of partial results computed by the vectorized loop.
2909 The reduction epilog code consists of:
2910 step 1: compute the scalar result in a vector (v_out2)
2911 step 2: extract the scalar result (s_out3) from the vector (v_out2)
2912 step 3: adjust the scalar result (s_out3) if needed.
2914 Step 1 can be accomplished using one the following three schemes:
2915 (scheme 1) using reduc_code, if available.
2916 (scheme 2) using whole-vector shifts, if available.
2917 (scheme 3) using a scalar loop. In this case steps 1+2 above are
2920 The overall epilog code looks like this:
2922 s_out0 = phi <s_loop> # original EXIT_PHI
2923 v_out1 = phi <VECT_DEF> # NEW_EXIT_PHI
2924 v_out2 = reduce <v_out1> # step 1
2925 s_out3 = extract_field <v_out2, 0> # step 2
2926 s_out4 = adjust_result <s_out3> # step 3
2928 (step 3 is optional, and steps 1 and 2 may be combined).
2929 Lastly, the uses of s_out0 are replaced by s_out4.
2933 /* 2.1 Create new loop-exit-phi to preserve loop-closed form:
2934 v_out1 = phi <v_loop> */
2936 exit_bb
= single_exit (loop
)->dest
;
2938 prev_phi_info
= NULL
;
2939 for (j
= 0; j
< ncopies
; j
++)
2941 phi
= create_phi_node (SSA_NAME_VAR (vect_def
), exit_bb
);
2942 set_vinfo_for_stmt (phi
, new_stmt_vec_info (phi
, loop_vinfo
, NULL
));
2947 def
= vect_get_vec_def_for_stmt_copy (dt
, def
);
2948 STMT_VINFO_RELATED_STMT (prev_phi_info
) = phi
;
2950 SET_PHI_ARG_DEF (phi
, single_exit (loop
)->dest_idx
, def
);
2951 prev_phi_info
= vinfo_for_stmt (phi
);
2954 exit_gsi
= gsi_after_labels (exit_bb
);
2956 /* 2.2 Get the relevant tree-code to use in the epilog for schemes 2,3
2957 (i.e. when reduc_code is not available) and in the final adjustment
2958 code (if needed). Also get the original scalar reduction variable as
2959 defined in the loop. In case STMT is a "pattern-stmt" (i.e. - it
2960 represents a reduction pattern), the tree-code and scalar-def are
2961 taken from the original stmt that the pattern-stmt (STMT) replaces.
2962 Otherwise (it is a regular reduction) - the tree-code and scalar-def
2963 are taken from STMT. */
2965 orig_stmt
= STMT_VINFO_RELATED_STMT (stmt_info
);
2968 /* Regular reduction */
2973 /* Reduction pattern */
2974 stmt_vec_info stmt_vinfo
= vinfo_for_stmt (orig_stmt
);
2975 gcc_assert (STMT_VINFO_IN_PATTERN_P (stmt_vinfo
));
2976 gcc_assert (STMT_VINFO_RELATED_STMT (stmt_vinfo
) == stmt
);
2979 code
= gimple_assign_rhs_code (orig_stmt
);
2980 scalar_dest
= gimple_assign_lhs (orig_stmt
);
2981 scalar_type
= TREE_TYPE (scalar_dest
);
2982 new_scalar_dest
= vect_create_destination_var (scalar_dest
, NULL
);
2983 bitsize
= TYPE_SIZE (scalar_type
);
2984 bytesize
= TYPE_SIZE_UNIT (scalar_type
);
2986 /* For MINUS_EXPR the initial vector is [init_val,0,...,0], therefore,
2987 partial results are added and not subtracted. */
2988 if (code
== MINUS_EXPR
)
2991 /* In case this is a reduction in an inner-loop while vectorizing an outer
2992 loop - we don't need to extract a single scalar result at the end of the
2993 inner-loop (unless it is double reduction, i.e., the use of reduction is
2994 outside the outer-loop). The final vector of partial results will be used
2995 in the vectorized outer-loop, or reduced to a scalar result at the end of
2997 if (nested_in_vect_loop
&& !double_reduc
)
2998 goto vect_finalize_reduction
;
3000 /* The epilogue is created for the outer-loop, i.e., for the loop being
3006 gcc_assert (ncopies
== 1);
3008 /* 2.3 Create the reduction code, using one of the three schemes described
3011 if (reduc_code
!= ERROR_MARK
)
3015 /*** Case 1: Create:
3016 v_out2 = reduc_expr <v_out1> */
3018 if (vect_print_dump_info (REPORT_DETAILS
))
3019 fprintf (vect_dump
, "Reduce using direct vector reduction.");
3021 vec_dest
= vect_create_destination_var (scalar_dest
, vectype
);
3022 tmp
= build1 (reduc_code
, vectype
, PHI_RESULT (new_phi
));
3023 epilog_stmt
= gimple_build_assign (vec_dest
, tmp
);
3024 new_temp
= make_ssa_name (vec_dest
, epilog_stmt
);
3025 gimple_assign_set_lhs (epilog_stmt
, new_temp
);
3026 gsi_insert_before (&exit_gsi
, epilog_stmt
, GSI_SAME_STMT
);
3028 extract_scalar_result
= true;
3032 enum tree_code shift_code
= ERROR_MARK
;
3033 bool have_whole_vector_shift
= true;
3035 int element_bitsize
= tree_low_cst (bitsize
, 1);
3036 int vec_size_in_bits
= tree_low_cst (TYPE_SIZE (vectype
), 1);
3039 if (optab_handler (vec_shr_optab
, mode
)->insn_code
!= CODE_FOR_nothing
)
3040 shift_code
= VEC_RSHIFT_EXPR
;
3042 have_whole_vector_shift
= false;
3044 /* Regardless of whether we have a whole vector shift, if we're
3045 emulating the operation via tree-vect-generic, we don't want
3046 to use it. Only the first round of the reduction is likely
3047 to still be profitable via emulation. */
3048 /* ??? It might be better to emit a reduction tree code here, so that
3049 tree-vect-generic can expand the first round via bit tricks. */
3050 if (!VECTOR_MODE_P (mode
))
3051 have_whole_vector_shift
= false;
3054 optab optab
= optab_for_tree_code (code
, vectype
, optab_default
);
3055 if (optab_handler (optab
, mode
)->insn_code
== CODE_FOR_nothing
)
3056 have_whole_vector_shift
= false;
3059 if (have_whole_vector_shift
)
3061 /*** Case 2: Create:
3062 for (offset = VS/2; offset >= element_size; offset/=2)
3064 Create: va' = vec_shift <va, offset>
3065 Create: va = vop <va, va'>
3068 if (vect_print_dump_info (REPORT_DETAILS
))
3069 fprintf (vect_dump
, "Reduce using vector shifts");
3071 vec_dest
= vect_create_destination_var (scalar_dest
, vectype
);
3072 new_temp
= PHI_RESULT (new_phi
);
3074 for (bit_offset
= vec_size_in_bits
/2;
3075 bit_offset
>= element_bitsize
;
3078 tree bitpos
= size_int (bit_offset
);
3080 epilog_stmt
= gimple_build_assign_with_ops (shift_code
, vec_dest
,
3082 new_name
= make_ssa_name (vec_dest
, epilog_stmt
);
3083 gimple_assign_set_lhs (epilog_stmt
, new_name
);
3084 gsi_insert_before (&exit_gsi
, epilog_stmt
, GSI_SAME_STMT
);
3086 epilog_stmt
= gimple_build_assign_with_ops (code
, vec_dest
,
3087 new_name
, new_temp
);
3088 new_temp
= make_ssa_name (vec_dest
, epilog_stmt
);
3089 gimple_assign_set_lhs (epilog_stmt
, new_temp
);
3090 gsi_insert_before (&exit_gsi
, epilog_stmt
, GSI_SAME_STMT
);
3093 extract_scalar_result
= true;
3099 /*** Case 3: Create:
3100 s = extract_field <v_out2, 0>
3101 for (offset = element_size;
3102 offset < vector_size;
3103 offset += element_size;)
3105 Create: s' = extract_field <v_out2, offset>
3106 Create: s = op <s, s'>
3109 if (vect_print_dump_info (REPORT_DETAILS
))
3110 fprintf (vect_dump
, "Reduce using scalar code. ");
3112 vec_temp
= PHI_RESULT (new_phi
);
3113 vec_size_in_bits
= tree_low_cst (TYPE_SIZE (vectype
), 1);
3114 rhs
= build3 (BIT_FIELD_REF
, scalar_type
, vec_temp
, bitsize
,
3116 epilog_stmt
= gimple_build_assign (new_scalar_dest
, rhs
);
3117 new_temp
= make_ssa_name (new_scalar_dest
, epilog_stmt
);
3118 gimple_assign_set_lhs (epilog_stmt
, new_temp
);
3119 gsi_insert_before (&exit_gsi
, epilog_stmt
, GSI_SAME_STMT
);
3121 for (bit_offset
= element_bitsize
;
3122 bit_offset
< vec_size_in_bits
;
3123 bit_offset
+= element_bitsize
)
3125 tree bitpos
= bitsize_int (bit_offset
);
3126 tree rhs
= build3 (BIT_FIELD_REF
, scalar_type
, vec_temp
, bitsize
,
3129 epilog_stmt
= gimple_build_assign (new_scalar_dest
, rhs
);
3130 new_name
= make_ssa_name (new_scalar_dest
, epilog_stmt
);
3131 gimple_assign_set_lhs (epilog_stmt
, new_name
);
3132 gsi_insert_before (&exit_gsi
, epilog_stmt
, GSI_SAME_STMT
);
3134 epilog_stmt
= gimple_build_assign_with_ops (code
,
3136 new_name
, new_temp
);
3137 new_temp
= make_ssa_name (new_scalar_dest
, epilog_stmt
);
3138 gimple_assign_set_lhs (epilog_stmt
, new_temp
);
3139 gsi_insert_before (&exit_gsi
, epilog_stmt
, GSI_SAME_STMT
);
3142 extract_scalar_result
= false;
3146 /* 2.4 Extract the final scalar result. Create:
3147 s_out3 = extract_field <v_out2, bitpos> */
3149 if (extract_scalar_result
)
3153 gcc_assert (!nested_in_vect_loop
|| double_reduc
);
3154 if (vect_print_dump_info (REPORT_DETAILS
))
3155 fprintf (vect_dump
, "extract scalar result");
3157 if (BYTES_BIG_ENDIAN
)
3158 bitpos
= size_binop (MULT_EXPR
,
3159 bitsize_int (TYPE_VECTOR_SUBPARTS (vectype
) - 1),
3160 TYPE_SIZE (scalar_type
));
3162 bitpos
= bitsize_zero_node
;
3164 rhs
= build3 (BIT_FIELD_REF
, scalar_type
, new_temp
, bitsize
, bitpos
);
3165 epilog_stmt
= gimple_build_assign (new_scalar_dest
, rhs
);
3166 new_temp
= make_ssa_name (new_scalar_dest
, epilog_stmt
);
3167 gimple_assign_set_lhs (epilog_stmt
, new_temp
);
3168 gsi_insert_before (&exit_gsi
, epilog_stmt
, GSI_SAME_STMT
);
3171 vect_finalize_reduction
:
3176 /* 2.5 Adjust the final result by the initial value of the reduction
3177 variable. (When such adjustment is not needed, then
3178 'adjustment_def' is zero). For example, if code is PLUS we create:
3179 new_temp = loop_exit_def + adjustment_def */
3183 if (nested_in_vect_loop
)
3185 gcc_assert (TREE_CODE (TREE_TYPE (adjustment_def
)) == VECTOR_TYPE
);
3186 expr
= build2 (code
, vectype
, PHI_RESULT (new_phi
), adjustment_def
);
3187 new_dest
= vect_create_destination_var (scalar_dest
, vectype
);
3191 gcc_assert (TREE_CODE (TREE_TYPE (adjustment_def
)) != VECTOR_TYPE
);
3192 expr
= build2 (code
, scalar_type
, new_temp
, adjustment_def
);
3193 new_dest
= vect_create_destination_var (scalar_dest
, scalar_type
);
3196 epilog_stmt
= gimple_build_assign (new_dest
, expr
);
3197 new_temp
= make_ssa_name (new_dest
, epilog_stmt
);
3198 gimple_assign_set_lhs (epilog_stmt
, new_temp
);
3199 SSA_NAME_DEF_STMT (new_temp
) = epilog_stmt
;
3200 gsi_insert_before (&exit_gsi
, epilog_stmt
, GSI_SAME_STMT
);
3204 /* 2.6 Handle the loop-exit phi */
3206 /* Replace uses of s_out0 with uses of s_out3:
3207 Find the loop-closed-use at the loop exit of the original scalar result.
3208 (The reduction result is expected to have two immediate uses - one at the
3209 latch block, and one at the loop exit). */
3210 phis
= VEC_alloc (gimple
, heap
, 10);
3211 FOR_EACH_IMM_USE_FAST (use_p
, imm_iter
, scalar_dest
)
3213 if (!flow_bb_inside_loop_p (loop
, gimple_bb (USE_STMT (use_p
))))
3215 exit_phi
= USE_STMT (use_p
);
3216 VEC_quick_push (gimple
, phis
, exit_phi
);
3220 /* We expect to have found an exit_phi because of loop-closed-ssa form. */
3221 gcc_assert (!VEC_empty (gimple
, phis
));
3223 for (i
= 0; VEC_iterate (gimple
, phis
, i
, exit_phi
); i
++)
3225 if (nested_in_vect_loop
)
3227 stmt_vec_info stmt_vinfo
= vinfo_for_stmt (exit_phi
);
3230 /* FORNOW. Currently not supporting the case that an inner-loop
3231 reduction is not used in the outer-loop (but only outside the
3232 outer-loop), unless it is double reduction. */
3233 gcc_assert ((STMT_VINFO_RELEVANT_P (stmt_vinfo
)
3234 && !STMT_VINFO_LIVE_P (stmt_vinfo
)) || double_reduc
);
3236 epilog_stmt
= adjustment_def
? epilog_stmt
: new_phi
;
3237 STMT_VINFO_VEC_STMT (stmt_vinfo
) = epilog_stmt
;
3238 set_vinfo_for_stmt (epilog_stmt
,
3239 new_stmt_vec_info (epilog_stmt
, loop_vinfo
,
3242 STMT_VINFO_RELATED_STMT (vinfo_for_stmt (epilog_stmt
)) =
3243 STMT_VINFO_RELATED_STMT (vinfo_for_stmt (new_phi
));
3246 || STMT_VINFO_DEF_TYPE (stmt_vinfo
) != vect_double_reduction_def
)
3249 /* Handle double reduction:
3251 stmt1: s1 = phi <s0, s2> - double reduction phi (outer loop)
3252 stmt2: s3 = phi <s1, s4> - (regular) reduction phi (inner loop)
3253 stmt3: s4 = use (s3) - (regular) reduction stmt (inner loop)
3254 stmt4: s2 = phi <s4> - double reduction stmt (outer loop)
3256 At that point the regular reduction (stmt2 and stmt3) is already
3257 vectorized, as well as the exit phi node, stmt4.
3258 Here we vectorize the phi node of double reduction, stmt1, and
3259 update all relevant statements. */
3261 /* Go through all the uses of s2 to find double reduction phi node,
3262 i.e., stmt1 above. */
3263 orig_name
= PHI_RESULT (exit_phi
);
3264 FOR_EACH_IMM_USE_STMT (use_stmt
, imm_iter
, orig_name
)
3266 stmt_vec_info use_stmt_vinfo
= vinfo_for_stmt (use_stmt
);
3267 stmt_vec_info new_phi_vinfo
;
3268 tree vect_phi_init
, preheader_arg
, vect_phi_res
, init_def
;
3269 basic_block bb
= gimple_bb (use_stmt
);
3272 /* Check that USE_STMT is really double reduction phi node. */
3273 if (gimple_code (use_stmt
) != GIMPLE_PHI
3274 || gimple_phi_num_args (use_stmt
) != 2
3276 || STMT_VINFO_DEF_TYPE (use_stmt_vinfo
)
3277 != vect_double_reduction_def
3278 || bb
->loop_father
!= outer_loop
)
3281 /* Create vector phi node for double reduction:
3282 vs1 = phi <vs0, vs2>
3283 vs1 was created previously in this function by a call to
3284 vect_get_vec_def_for_operand and is stored in vec_initial_def;
3285 vs2 is defined by EPILOG_STMT, the vectorized EXIT_PHI;
3286 vs0 is created here. */
3288 /* Create vector phi node. */
3289 vect_phi
= create_phi_node (vec_initial_def
, bb
);
3290 new_phi_vinfo
= new_stmt_vec_info (vect_phi
,
3291 loop_vec_info_for_loop (outer_loop
), NULL
);
3292 set_vinfo_for_stmt (vect_phi
, new_phi_vinfo
);
3294 /* Create vs0 - initial def of the double reduction phi. */
3295 preheader_arg
= PHI_ARG_DEF_FROM_EDGE (use_stmt
,
3296 loop_preheader_edge (outer_loop
));
3297 init_def
= get_initial_def_for_reduction (stmt
, preheader_arg
,
3299 vect_phi_init
= vect_init_vector (use_stmt
, init_def
, vectype
,
3302 /* Update phi node arguments with vs0 and vs2. */
3303 add_phi_arg (vect_phi
, vect_phi_init
,
3304 loop_preheader_edge (outer_loop
));
3305 add_phi_arg (vect_phi
, PHI_RESULT (epilog_stmt
),
3306 loop_latch_edge (outer_loop
));
3307 if (vect_print_dump_info (REPORT_DETAILS
))
3309 fprintf (vect_dump
, "created double reduction phi node: ");
3310 print_gimple_stmt (vect_dump
, vect_phi
, 0, TDF_SLIM
);
3313 vect_phi_res
= PHI_RESULT (vect_phi
);
3315 /* Replace the use, i.e., set the correct vs1 in the regular
3316 reduction phi node. FORNOW, NCOPIES is always 1, so the loop
3318 use
= reduction_phi
;
3319 for (j
= 0; j
< ncopies
; j
++)
3321 edge pr_edge
= loop_preheader_edge (loop
);
3322 SET_PHI_ARG_DEF (use
, pr_edge
->dest_idx
, vect_phi_res
);
3323 use
= STMT_VINFO_RELATED_STMT (vinfo_for_stmt (use
));
3328 /* Replace the uses: */
3329 orig_name
= PHI_RESULT (exit_phi
);
3330 FOR_EACH_IMM_USE_STMT (use_stmt
, imm_iter
, orig_name
)
3331 FOR_EACH_IMM_USE_ON_STMT (use_p
, imm_iter
)
3332 SET_USE (use_p
, new_temp
);
3335 VEC_free (gimple
, heap
, phis
);
3339 /* Function vectorizable_reduction.
3341 Check if STMT performs a reduction operation that can be vectorized.
3342 If VEC_STMT is also passed, vectorize the STMT: create a vectorized
3343 stmt to replace it, put it in VEC_STMT, and insert it at GSI.
3344 Return FALSE if not a vectorizable STMT, TRUE otherwise.
3346 This function also handles reduction idioms (patterns) that have been
3347 recognized in advance during vect_pattern_recog. In this case, STMT may be
3349 X = pattern_expr (arg0, arg1, ..., X)
3350 and it's STMT_VINFO_RELATED_STMT points to the last stmt in the original
3351 sequence that had been detected and replaced by the pattern-stmt (STMT).
3353 In some cases of reduction patterns, the type of the reduction variable X is
3354 different than the type of the other arguments of STMT.
3355 In such cases, the vectype that is used when transforming STMT into a vector
3356 stmt is different than the vectype that is used to determine the
3357 vectorization factor, because it consists of a different number of elements
3358 than the actual number of elements that are being operated upon in parallel.
3360 For example, consider an accumulation of shorts into an int accumulator.
3361 On some targets it's possible to vectorize this pattern operating on 8
3362 shorts at a time (hence, the vectype for purposes of determining the
3363 vectorization factor should be V8HI); on the other hand, the vectype that
3364 is used to create the vector form is actually V4SI (the type of the result).
3366 Upon entry to this function, STMT_VINFO_VECTYPE records the vectype that
3367 indicates what is the actual level of parallelism (V8HI in the example), so
3368 that the right vectorization factor would be derived. This vectype
3369 corresponds to the type of arguments to the reduction stmt, and should *NOT*
3370 be used to create the vectorized stmt. The right vectype for the vectorized
3371 stmt is obtained from the type of the result X:
3372 get_vectype_for_scalar_type (TREE_TYPE (X))
3374 This means that, contrary to "regular" reductions (or "regular" stmts in
3375 general), the following equation:
3376 STMT_VINFO_VECTYPE == get_vectype_for_scalar_type (TREE_TYPE (X))
3377 does *NOT* necessarily hold for reduction patterns. */
3380 vectorizable_reduction (gimple stmt
, gimple_stmt_iterator
*gsi
,
3385 tree loop_vec_def0
= NULL_TREE
, loop_vec_def1
= NULL_TREE
;
3386 stmt_vec_info stmt_info
= vinfo_for_stmt (stmt
);
3387 tree vectype
= STMT_VINFO_VECTYPE (stmt_info
);
3388 loop_vec_info loop_vinfo
= STMT_VINFO_LOOP_VINFO (stmt_info
);
3389 struct loop
*loop
= LOOP_VINFO_LOOP (loop_vinfo
);
3390 enum tree_code code
, orig_code
, epilog_reduc_code
;
3391 enum machine_mode vec_mode
;
3393 optab optab
, reduc_optab
;
3394 tree new_temp
= NULL_TREE
;
3397 enum vect_def_type dt
;
3398 gimple new_phi
= NULL
;
3402 stmt_vec_info orig_stmt_info
;
3403 tree expr
= NULL_TREE
;
3405 int nunits
= TYPE_VECTOR_SUBPARTS (vectype
);
3406 int ncopies
= LOOP_VINFO_VECT_FACTOR (loop_vinfo
) / nunits
;
3408 stmt_vec_info prev_stmt_info
, prev_phi_info
;
3409 gimple first_phi
= NULL
;
3410 bool single_defuse_cycle
= false;
3412 gimple new_stmt
= NULL
;
3415 bool nested_cycle
= false, found_nested_cycle_def
= false;
3416 gimple reduc_def_stmt
= NULL
;
3417 /* The default is that the reduction variable is the last in statement. */
3418 int reduc_index
= 2;
3419 bool double_reduc
= false, dummy
;
3421 struct loop
* def_stmt_loop
;
3424 if (nested_in_vect_loop_p (loop
, stmt
))
3427 nested_cycle
= true;
3430 gcc_assert (ncopies
>= 1);
3432 /* FORNOW: SLP not supported. */
3433 if (STMT_SLP_TYPE (stmt_info
))
3436 /* 1. Is vectorizable reduction? */
3437 /* Not supportable if the reduction variable is used in the loop. */
3438 if (STMT_VINFO_RELEVANT (stmt_info
) > vect_used_in_outer
)
3441 /* Reductions that are not used even in an enclosing outer-loop,
3442 are expected to be "live" (used out of the loop). */
3443 if (STMT_VINFO_RELEVANT (stmt_info
) == vect_unused_in_scope
3444 && !STMT_VINFO_LIVE_P (stmt_info
))
3447 /* Make sure it was already recognized as a reduction computation. */
3448 if (STMT_VINFO_DEF_TYPE (stmt_info
) != vect_reduction_def
3449 && STMT_VINFO_DEF_TYPE (stmt_info
) != vect_nested_cycle
)
3452 /* 2. Has this been recognized as a reduction pattern?
3454 Check if STMT represents a pattern that has been recognized
3455 in earlier analysis stages. For stmts that represent a pattern,
3456 the STMT_VINFO_RELATED_STMT field records the last stmt in
3457 the original sequence that constitutes the pattern. */
3459 orig_stmt
= STMT_VINFO_RELATED_STMT (stmt_info
);
3462 orig_stmt_info
= vinfo_for_stmt (orig_stmt
);
3463 gcc_assert (STMT_VINFO_RELATED_STMT (orig_stmt_info
) == stmt
);
3464 gcc_assert (STMT_VINFO_IN_PATTERN_P (orig_stmt_info
));
3465 gcc_assert (!STMT_VINFO_IN_PATTERN_P (stmt_info
));
3468 /* 3. Check the operands of the operation. The first operands are defined
3469 inside the loop body. The last operand is the reduction variable,
3470 which is defined by the loop-header-phi. */
3472 gcc_assert (is_gimple_assign (stmt
));
3475 switch (get_gimple_rhs_class (gimple_assign_rhs_code (stmt
)))
3477 case GIMPLE_SINGLE_RHS
:
3478 op_type
= TREE_OPERAND_LENGTH (gimple_assign_rhs1 (stmt
));
3479 if (op_type
== ternary_op
)
3481 tree rhs
= gimple_assign_rhs1 (stmt
);
3482 ops
[0] = TREE_OPERAND (rhs
, 0);
3483 ops
[1] = TREE_OPERAND (rhs
, 1);
3484 ops
[2] = TREE_OPERAND (rhs
, 2);
3485 code
= TREE_CODE (rhs
);
3491 case GIMPLE_BINARY_RHS
:
3492 code
= gimple_assign_rhs_code (stmt
);
3493 op_type
= TREE_CODE_LENGTH (code
);
3494 gcc_assert (op_type
== binary_op
);
3495 ops
[0] = gimple_assign_rhs1 (stmt
);
3496 ops
[1] = gimple_assign_rhs2 (stmt
);
3499 case GIMPLE_UNARY_RHS
:
3506 scalar_dest
= gimple_assign_lhs (stmt
);
3507 scalar_type
= TREE_TYPE (scalar_dest
);
3508 if (!POINTER_TYPE_P (scalar_type
) && !INTEGRAL_TYPE_P (scalar_type
)
3509 && !SCALAR_FLOAT_TYPE_P (scalar_type
))
3512 /* All uses but the last are expected to be defined in the loop.
3513 The last use is the reduction variable. In case of nested cycle this
3514 assumption is not true: we use reduc_index to record the index of the
3515 reduction variable. */
3516 for (i
= 0; i
< op_type
-1; i
++)
3518 is_simple_use
= vect_is_simple_use (ops
[i
], loop_vinfo
, NULL
, &def_stmt
,
3520 gcc_assert (is_simple_use
);
3521 if (dt
!= vect_internal_def
3522 && dt
!= vect_external_def
3523 && dt
!= vect_constant_def
3524 && dt
!= vect_induction_def
3525 && dt
!= vect_nested_cycle
)
3528 if (dt
== vect_nested_cycle
)
3530 found_nested_cycle_def
= true;
3531 reduc_def_stmt
= def_stmt
;
3536 is_simple_use
= vect_is_simple_use (ops
[i
], loop_vinfo
, NULL
, &def_stmt
,
3538 gcc_assert (is_simple_use
);
3539 gcc_assert (dt
== vect_reduction_def
3540 || dt
== vect_nested_cycle
3541 || ((dt
== vect_internal_def
|| dt
== vect_external_def
3542 || dt
== vect_constant_def
|| dt
== vect_induction_def
)
3543 && nested_cycle
&& found_nested_cycle_def
));
3544 if (!found_nested_cycle_def
)
3545 reduc_def_stmt
= def_stmt
;
3547 gcc_assert (gimple_code (reduc_def_stmt
) == GIMPLE_PHI
);
3549 gcc_assert (orig_stmt
== vect_is_simple_reduction (loop_vinfo
,
3554 gcc_assert (stmt
== vect_is_simple_reduction (loop_vinfo
, reduc_def_stmt
,
3555 !nested_cycle
, &dummy
));
3557 if (STMT_VINFO_LIVE_P (vinfo_for_stmt (reduc_def_stmt
)))
3560 /* 4. Supportable by target? */
3562 /* 4.1. check support for the operation in the loop */
3563 optab
= optab_for_tree_code (code
, vectype
, optab_default
);
3566 if (vect_print_dump_info (REPORT_DETAILS
))
3567 fprintf (vect_dump
, "no optab.");
3570 vec_mode
= TYPE_MODE (vectype
);
3571 if (optab_handler (optab
, vec_mode
)->insn_code
== CODE_FOR_nothing
)
3573 if (vect_print_dump_info (REPORT_DETAILS
))
3574 fprintf (vect_dump
, "op not supported by target.");
3575 if (GET_MODE_SIZE (vec_mode
) != UNITS_PER_WORD
3576 || LOOP_VINFO_VECT_FACTOR (loop_vinfo
)
3577 < vect_min_worthwhile_factor (code
))
3579 if (vect_print_dump_info (REPORT_DETAILS
))
3580 fprintf (vect_dump
, "proceeding using word mode.");
3583 /* Worthwhile without SIMD support? */
3584 if (!VECTOR_MODE_P (TYPE_MODE (vectype
))
3585 && LOOP_VINFO_VECT_FACTOR (loop_vinfo
)
3586 < vect_min_worthwhile_factor (code
))
3588 if (vect_print_dump_info (REPORT_DETAILS
))
3589 fprintf (vect_dump
, "not worthwhile without SIMD support.");
3593 /* 4.2. Check support for the epilog operation.
3595 If STMT represents a reduction pattern, then the type of the
3596 reduction variable may be different than the type of the rest
3597 of the arguments. For example, consider the case of accumulation
3598 of shorts into an int accumulator; The original code:
3599 S1: int_a = (int) short_a;
3600 orig_stmt-> S2: int_acc = plus <int_a ,int_acc>;
3603 STMT: int_acc = widen_sum <short_a, int_acc>
3606 1. The tree-code that is used to create the vector operation in the
3607 epilog code (that reduces the partial results) is not the
3608 tree-code of STMT, but is rather the tree-code of the original
3609 stmt from the pattern that STMT is replacing. I.e, in the example
3610 above we want to use 'widen_sum' in the loop, but 'plus' in the
3612 2. The type (mode) we use to check available target support
3613 for the vector operation to be created in the *epilog*, is
3614 determined by the type of the reduction variable (in the example
3615 above we'd check this: plus_optab[vect_int_mode]).
3616 However the type (mode) we use to check available target support
3617 for the vector operation to be created *inside the loop*, is
3618 determined by the type of the other arguments to STMT (in the
3619 example we'd check this: widen_sum_optab[vect_short_mode]).
3621 This is contrary to "regular" reductions, in which the types of all
3622 the arguments are the same as the type of the reduction variable.
3623 For "regular" reductions we can therefore use the same vector type
3624 (and also the same tree-code) when generating the epilog code and
3625 when generating the code inside the loop. */
3629 /* This is a reduction pattern: get the vectype from the type of the
3630 reduction variable, and get the tree-code from orig_stmt. */
3631 orig_code
= gimple_assign_rhs_code (orig_stmt
);
3632 vectype
= get_vectype_for_scalar_type (TREE_TYPE (def
));
3635 if (vect_print_dump_info (REPORT_DETAILS
))
3637 fprintf (vect_dump
, "unsupported data-type ");
3638 print_generic_expr (vect_dump
, TREE_TYPE (def
), TDF_SLIM
);
3643 vec_mode
= TYPE_MODE (vectype
);
3647 /* Regular reduction: use the same vectype and tree-code as used for
3648 the vector code inside the loop can be used for the epilog code. */
3652 if (!reduction_code_for_scalar_code (orig_code
, &epilog_reduc_code
))
3655 reduc_optab
= optab_for_tree_code (epilog_reduc_code
, vectype
,
3659 if (vect_print_dump_info (REPORT_DETAILS
))
3660 fprintf (vect_dump
, "no optab for reduction.");
3661 epilog_reduc_code
= ERROR_MARK
;
3665 && optab_handler (reduc_optab
, vec_mode
)->insn_code
== CODE_FOR_nothing
)
3667 if (vect_print_dump_info (REPORT_DETAILS
))
3668 fprintf (vect_dump
, "reduc op not supported by target.");
3669 epilog_reduc_code
= ERROR_MARK
;
3672 def_bb
= gimple_bb (reduc_def_stmt
);
3673 def_stmt_loop
= def_bb
->loop_father
;
3674 def_arg
= PHI_ARG_DEF_FROM_EDGE (reduc_def_stmt
,
3675 loop_preheader_edge (def_stmt_loop
));
3676 if (TREE_CODE (def_arg
) == SSA_NAME
3677 && vinfo_for_stmt (SSA_NAME_DEF_STMT (def_arg
))
3678 && STMT_VINFO_DEF_TYPE (vinfo_for_stmt (SSA_NAME_DEF_STMT (def_arg
)))
3679 == vect_double_reduction_def
)
3680 double_reduc
= true;
3682 if (double_reduc
&& ncopies
> 1)
3684 if (vect_print_dump_info (REPORT_DETAILS
))
3685 fprintf (vect_dump
, "multiple types in double reduction");
3690 if (!vec_stmt
) /* transformation not required. */
3692 STMT_VINFO_TYPE (stmt_info
) = reduc_vec_info_type
;
3693 if (!vect_model_reduction_cost (stmt_info
, epilog_reduc_code
, ncopies
))
3700 if (vect_print_dump_info (REPORT_DETAILS
))
3701 fprintf (vect_dump
, "transform reduction.");
3703 /* Create the destination vector */
3704 vec_dest
= vect_create_destination_var (scalar_dest
, vectype
);
3706 /* In case the vectorization factor (VF) is bigger than the number
3707 of elements that we can fit in a vectype (nunits), we have to generate
3708 more than one vector stmt - i.e - we need to "unroll" the
3709 vector stmt by a factor VF/nunits. For more details see documentation
3710 in vectorizable_operation. */
3712 /* If the reduction is used in an outer loop we need to generate
3713 VF intermediate results, like so (e.g. for ncopies=2):
3718 (i.e. we generate VF results in 2 registers).
3719 In this case we have a separate def-use cycle for each copy, and therefore
3720 for each copy we get the vector def for the reduction variable from the
3721 respective phi node created for this copy.
3723 Otherwise (the reduction is unused in the loop nest), we can combine
3724 together intermediate results, like so (e.g. for ncopies=2):
3728 (i.e. we generate VF/2 results in a single register).
3729 In this case for each copy we get the vector def for the reduction variable
3730 from the vectorized reduction operation generated in the previous iteration.
3733 if (STMT_VINFO_RELEVANT (stmt_info
) == vect_unused_in_scope
)
3735 single_defuse_cycle
= true;
3739 epilog_copies
= ncopies
;
3741 prev_stmt_info
= NULL
;
3742 prev_phi_info
= NULL
;
3743 for (j
= 0; j
< ncopies
; j
++)
3745 if (j
== 0 || !single_defuse_cycle
)
3747 /* Create the reduction-phi that defines the reduction-operand. */
3748 new_phi
= create_phi_node (vec_dest
, loop
->header
);
3749 set_vinfo_for_stmt (new_phi
, new_stmt_vec_info (new_phi
, loop_vinfo
,
3756 loop_vec_def0
= vect_get_vec_def_for_operand (ops
[!reduc_index
],
3758 if (op_type
== ternary_op
)
3760 if (reduc_index
== 0)
3761 loop_vec_def1
= vect_get_vec_def_for_operand (ops
[2], stmt
,
3764 loop_vec_def1
= vect_get_vec_def_for_operand (ops
[1], stmt
,
3768 /* Get the vector def for the reduction variable from the phi
3770 reduc_def
= PHI_RESULT (new_phi
);
3771 first_phi
= new_phi
;
3775 enum vect_def_type dt
= vect_unknown_def_type
; /* Dummy */
3776 loop_vec_def0
= vect_get_vec_def_for_stmt_copy (dt
, loop_vec_def0
);
3777 if (op_type
== ternary_op
)
3778 loop_vec_def1
= vect_get_vec_def_for_stmt_copy (dt
, loop_vec_def1
);
3780 if (single_defuse_cycle
)
3781 reduc_def
= gimple_assign_lhs (new_stmt
);
3783 reduc_def
= PHI_RESULT (new_phi
);
3785 STMT_VINFO_RELATED_STMT (prev_phi_info
) = new_phi
;
3789 /* Arguments are ready. create the new vector stmt. */
3790 if (op_type
== binary_op
)
3792 if (reduc_index
== 0)
3793 expr
= build2 (code
, vectype
, reduc_def
, loop_vec_def0
);
3795 expr
= build2 (code
, vectype
, loop_vec_def0
, reduc_def
);
3799 if (reduc_index
== 0)
3800 expr
= build3 (code
, vectype
, reduc_def
, loop_vec_def0
,
3804 if (reduc_index
== 1)
3805 expr
= build3 (code
, vectype
, loop_vec_def0
, reduc_def
,
3808 expr
= build3 (code
, vectype
, loop_vec_def0
, loop_vec_def1
,
3813 new_stmt
= gimple_build_assign (vec_dest
, expr
);
3814 new_temp
= make_ssa_name (vec_dest
, new_stmt
);
3815 gimple_assign_set_lhs (new_stmt
, new_temp
);
3816 vect_finish_stmt_generation (stmt
, new_stmt
, gsi
);
3819 STMT_VINFO_VEC_STMT (stmt_info
) = *vec_stmt
= new_stmt
;
3821 STMT_VINFO_RELATED_STMT (prev_stmt_info
) = new_stmt
;
3822 prev_stmt_info
= vinfo_for_stmt (new_stmt
);
3823 prev_phi_info
= vinfo_for_stmt (new_phi
);
3826 /* Finalize the reduction-phi (set its arguments) and create the
3827 epilog reduction code. */
3828 if (!single_defuse_cycle
)
3829 new_temp
= gimple_assign_lhs (*vec_stmt
);
3831 vect_create_epilog_for_reduction (new_temp
, stmt
, epilog_copies
,
3832 epilog_reduc_code
, first_phi
, reduc_index
,
3837 /* Function vect_min_worthwhile_factor.
3839 For a loop where we could vectorize the operation indicated by CODE,
3840 return the minimum vectorization factor that makes it worthwhile
3841 to use generic vectors. */
3843 vect_min_worthwhile_factor (enum tree_code code
)
3864 /* Function vectorizable_induction
3866 Check if PHI performs an induction computation that can be vectorized.
3867 If VEC_STMT is also passed, vectorize the induction PHI: create a vectorized
3868 phi to replace it, put it in VEC_STMT, and add it to the same basic block.
3869 Return FALSE if not a vectorizable STMT, TRUE otherwise. */
3872 vectorizable_induction (gimple phi
, gimple_stmt_iterator
*gsi ATTRIBUTE_UNUSED
,
3875 stmt_vec_info stmt_info
= vinfo_for_stmt (phi
);
3876 tree vectype
= STMT_VINFO_VECTYPE (stmt_info
);
3877 loop_vec_info loop_vinfo
= STMT_VINFO_LOOP_VINFO (stmt_info
);
3878 struct loop
*loop
= LOOP_VINFO_LOOP (loop_vinfo
);
3879 int nunits
= TYPE_VECTOR_SUBPARTS (vectype
);
3880 int ncopies
= LOOP_VINFO_VECT_FACTOR (loop_vinfo
) / nunits
;
3883 gcc_assert (ncopies
>= 1);
3884 /* FORNOW. This restriction should be relaxed. */
3885 if (nested_in_vect_loop_p (loop
, phi
) && ncopies
> 1)
3887 if (vect_print_dump_info (REPORT_DETAILS
))
3888 fprintf (vect_dump
, "multiple types in nested loop.");
3892 if (!STMT_VINFO_RELEVANT_P (stmt_info
))
3895 /* FORNOW: SLP not supported. */
3896 if (STMT_SLP_TYPE (stmt_info
))
3899 gcc_assert (STMT_VINFO_DEF_TYPE (stmt_info
) == vect_induction_def
);
3901 if (gimple_code (phi
) != GIMPLE_PHI
)
3904 if (!vec_stmt
) /* transformation not required. */
3906 STMT_VINFO_TYPE (stmt_info
) = induc_vec_info_type
;
3907 if (vect_print_dump_info (REPORT_DETAILS
))
3908 fprintf (vect_dump
, "=== vectorizable_induction ===");
3909 vect_model_induction_cost (stmt_info
, ncopies
);
3915 if (vect_print_dump_info (REPORT_DETAILS
))
3916 fprintf (vect_dump
, "transform induction phi.");
3918 vec_def
= get_initial_def_for_induction (phi
);
3919 *vec_stmt
= SSA_NAME_DEF_STMT (vec_def
);
3923 /* Function vectorizable_live_operation.
3925 STMT computes a value that is used outside the loop. Check if
3926 it can be supported. */
3929 vectorizable_live_operation (gimple stmt
,
3930 gimple_stmt_iterator
*gsi ATTRIBUTE_UNUSED
,
3931 gimple
*vec_stmt ATTRIBUTE_UNUSED
)
3933 stmt_vec_info stmt_info
= vinfo_for_stmt (stmt
);
3934 loop_vec_info loop_vinfo
= STMT_VINFO_LOOP_VINFO (stmt_info
);
3935 struct loop
*loop
= LOOP_VINFO_LOOP (loop_vinfo
);
3941 enum vect_def_type dt
;
3942 enum tree_code code
;
3943 enum gimple_rhs_class rhs_class
;
3945 gcc_assert (STMT_VINFO_LIVE_P (stmt_info
));
3947 if (STMT_VINFO_DEF_TYPE (stmt_info
) == vect_reduction_def
)
3950 if (!is_gimple_assign (stmt
))
3953 if (TREE_CODE (gimple_assign_lhs (stmt
)) != SSA_NAME
)
3956 /* FORNOW. CHECKME. */
3957 if (nested_in_vect_loop_p (loop
, stmt
))
3960 code
= gimple_assign_rhs_code (stmt
);
3961 op_type
= TREE_CODE_LENGTH (code
);
3962 rhs_class
= get_gimple_rhs_class (code
);
3963 gcc_assert (rhs_class
!= GIMPLE_UNARY_RHS
|| op_type
== unary_op
);
3964 gcc_assert (rhs_class
!= GIMPLE_BINARY_RHS
|| op_type
== binary_op
);
3966 /* FORNOW: support only if all uses are invariant. This means
3967 that the scalar operations can remain in place, unvectorized.
3968 The original last scalar value that they compute will be used. */
3970 for (i
= 0; i
< op_type
; i
++)
3972 if (rhs_class
== GIMPLE_SINGLE_RHS
)
3973 op
= TREE_OPERAND (gimple_op (stmt
, 1), i
);
3975 op
= gimple_op (stmt
, i
+ 1);
3977 && !vect_is_simple_use (op
, loop_vinfo
, NULL
, &def_stmt
, &def
, &dt
))
3979 if (vect_print_dump_info (REPORT_DETAILS
))
3980 fprintf (vect_dump
, "use not simple.");
3984 if (dt
!= vect_external_def
&& dt
!= vect_constant_def
)
3988 /* No transformation is required for the cases we currently support. */
3992 /* Function vect_transform_loop.
3994 The analysis phase has determined that the loop is vectorizable.
3995 Vectorize the loop - created vectorized stmts to replace the scalar
3996 stmts in the loop, and update the loop exit condition. */
3999 vect_transform_loop (loop_vec_info loop_vinfo
)
4001 struct loop
*loop
= LOOP_VINFO_LOOP (loop_vinfo
);
4002 basic_block
*bbs
= LOOP_VINFO_BBS (loop_vinfo
);
4003 int nbbs
= loop
->num_nodes
;
4004 gimple_stmt_iterator si
;
4007 int vectorization_factor
= LOOP_VINFO_VECT_FACTOR (loop_vinfo
);
4009 bool slp_scheduled
= false;
4010 unsigned int nunits
;
4011 tree cond_expr
= NULL_TREE
;
4012 gimple_seq cond_expr_stmt_list
= NULL
;
4013 bool do_peeling_for_loop_bound
;
4015 if (vect_print_dump_info (REPORT_DETAILS
))
4016 fprintf (vect_dump
, "=== vec_transform_loop ===");
4018 /* Peel the loop if there are data refs with unknown alignment.
4019 Only one data ref with unknown store is allowed. */
4021 if (LOOP_PEELING_FOR_ALIGNMENT (loop_vinfo
))
4022 vect_do_peeling_for_alignment (loop_vinfo
);
4024 do_peeling_for_loop_bound
4025 = (!LOOP_VINFO_NITERS_KNOWN_P (loop_vinfo
)
4026 || (LOOP_VINFO_NITERS_KNOWN_P (loop_vinfo
)
4027 && LOOP_VINFO_INT_NITERS (loop_vinfo
) % vectorization_factor
!= 0));
4029 if (LOOP_REQUIRES_VERSIONING_FOR_ALIGNMENT (loop_vinfo
)
4030 || LOOP_REQUIRES_VERSIONING_FOR_ALIAS (loop_vinfo
))
4031 vect_loop_versioning (loop_vinfo
,
4032 !do_peeling_for_loop_bound
,
4033 &cond_expr
, &cond_expr_stmt_list
);
4035 /* If the loop has a symbolic number of iterations 'n' (i.e. it's not a
4036 compile time constant), or it is a constant that doesn't divide by the
4037 vectorization factor, then an epilog loop needs to be created.
4038 We therefore duplicate the loop: the original loop will be vectorized,
4039 and will compute the first (n/VF) iterations. The second copy of the loop
4040 will remain scalar and will compute the remaining (n%VF) iterations.
4041 (VF is the vectorization factor). */
4043 if (do_peeling_for_loop_bound
)
4044 vect_do_peeling_for_loop_bound (loop_vinfo
, &ratio
,
4045 cond_expr
, cond_expr_stmt_list
);
4047 ratio
= build_int_cst (TREE_TYPE (LOOP_VINFO_NITERS (loop_vinfo
)),
4048 LOOP_VINFO_INT_NITERS (loop_vinfo
) / vectorization_factor
);
4050 /* 1) Make sure the loop header has exactly two entries
4051 2) Make sure we have a preheader basic block. */
4053 gcc_assert (EDGE_COUNT (loop
->header
->preds
) == 2);
4055 split_edge (loop_preheader_edge (loop
));
4057 /* FORNOW: the vectorizer supports only loops which body consist
4058 of one basic block (header + empty latch). When the vectorizer will
4059 support more involved loop forms, the order by which the BBs are
4060 traversed need to be reconsidered. */
4062 for (i
= 0; i
< nbbs
; i
++)
4064 basic_block bb
= bbs
[i
];
4065 stmt_vec_info stmt_info
;
4068 for (si
= gsi_start_phis (bb
); !gsi_end_p (si
); gsi_next (&si
))
4070 phi
= gsi_stmt (si
);
4071 if (vect_print_dump_info (REPORT_DETAILS
))
4073 fprintf (vect_dump
, "------>vectorizing phi: ");
4074 print_gimple_stmt (vect_dump
, phi
, 0, TDF_SLIM
);
4076 stmt_info
= vinfo_for_stmt (phi
);
4080 if (!STMT_VINFO_RELEVANT_P (stmt_info
)
4081 && !STMT_VINFO_LIVE_P (stmt_info
))
4084 if ((TYPE_VECTOR_SUBPARTS (STMT_VINFO_VECTYPE (stmt_info
))
4085 != (unsigned HOST_WIDE_INT
) vectorization_factor
)
4086 && vect_print_dump_info (REPORT_DETAILS
))
4087 fprintf (vect_dump
, "multiple-types.");
4089 if (STMT_VINFO_DEF_TYPE (stmt_info
) == vect_induction_def
)
4091 if (vect_print_dump_info (REPORT_DETAILS
))
4092 fprintf (vect_dump
, "transform phi.");
4093 vect_transform_stmt (phi
, NULL
, NULL
, NULL
, NULL
);
4097 for (si
= gsi_start_bb (bb
); !gsi_end_p (si
);)
4099 gimple stmt
= gsi_stmt (si
);
4102 if (vect_print_dump_info (REPORT_DETAILS
))
4104 fprintf (vect_dump
, "------>vectorizing statement: ");
4105 print_gimple_stmt (vect_dump
, stmt
, 0, TDF_SLIM
);
4108 stmt_info
= vinfo_for_stmt (stmt
);
4110 /* vector stmts created in the outer-loop during vectorization of
4111 stmts in an inner-loop may not have a stmt_info, and do not
4112 need to be vectorized. */
4119 if (!STMT_VINFO_RELEVANT_P (stmt_info
)
4120 && !STMT_VINFO_LIVE_P (stmt_info
))
4126 gcc_assert (STMT_VINFO_VECTYPE (stmt_info
));
4128 (unsigned int) TYPE_VECTOR_SUBPARTS (STMT_VINFO_VECTYPE (stmt_info
));
4129 if (!STMT_SLP_TYPE (stmt_info
)
4130 && nunits
!= (unsigned int) vectorization_factor
4131 && vect_print_dump_info (REPORT_DETAILS
))
4132 /* For SLP VF is set according to unrolling factor, and not to
4133 vector size, hence for SLP this print is not valid. */
4134 fprintf (vect_dump
, "multiple-types.");
4136 /* SLP. Schedule all the SLP instances when the first SLP stmt is
4138 if (STMT_SLP_TYPE (stmt_info
))
4142 slp_scheduled
= true;
4144 if (vect_print_dump_info (REPORT_DETAILS
))
4145 fprintf (vect_dump
, "=== scheduling SLP instances ===");
4147 vect_schedule_slp (loop_vinfo
, NULL
);
4150 /* Hybrid SLP stmts must be vectorized in addition to SLP. */
4151 if (!vinfo_for_stmt (stmt
) || PURE_SLP_STMT (stmt_info
))
4158 /* -------- vectorize statement ------------ */
4159 if (vect_print_dump_info (REPORT_DETAILS
))
4160 fprintf (vect_dump
, "transform statement.");
4162 strided_store
= false;
4163 is_store
= vect_transform_stmt (stmt
, &si
, &strided_store
, NULL
, NULL
);
4166 if (STMT_VINFO_STRIDED_ACCESS (stmt_info
))
4168 /* Interleaving. If IS_STORE is TRUE, the vectorization of the
4169 interleaving chain was completed - free all the stores in
4171 vect_remove_stores (DR_GROUP_FIRST_DR (stmt_info
));
4172 gsi_remove (&si
, true);
4177 /* Free the attached stmt_vec_info and remove the stmt. */
4178 free_stmt_vec_info (stmt
);
4179 gsi_remove (&si
, true);
4187 slpeel_make_loop_iterate_ntimes (loop
, ratio
);
4189 /* The memory tags and pointers in vectorized statements need to
4190 have their SSA forms updated. FIXME, why can't this be delayed
4191 until all the loops have been transformed? */
4192 update_ssa (TODO_update_ssa
);
4194 if (vect_print_dump_info (REPORT_VECTORIZED_LOCATIONS
))
4195 fprintf (vect_dump
, "LOOP VECTORIZED.");
4196 if (loop
->inner
&& vect_print_dump_info (REPORT_VECTORIZED_LOCATIONS
))
4197 fprintf (vect_dump
, "OUTER LOOP VECTORIZED.");