2 Copyright (C) 2003-2013 Free Software Foundation, Inc.
3 Contributed by Dorit Naishlos <dorit@il.ibm.com> and
4 Ira Rosen <irar@il.ibm.com>
6 This file is part of GCC.
8 GCC is free software; you can redistribute it and/or modify it under
9 the terms of the GNU General Public License as published by the Free
10 Software Foundation; either version 3, or (at your option) any later
13 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
14 WARRANTY; without even the implied warranty of MERCHANTABILITY or
15 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
18 You should have received a copy of the GNU General Public License
19 along with GCC; see the file COPYING3. If not see
20 <http://www.gnu.org/licenses/>. */
24 #include "coretypes.h"
29 #include "basic-block.h"
30 #include "gimple-pretty-print.h"
33 #include "gimple-iterator.h"
34 #include "gimplify-me.h"
35 #include "gimple-ssa.h"
36 #include "tree-phinodes.h"
37 #include "ssa-iterators.h"
38 #include "tree-ssanames.h"
39 #include "tree-ssa-loop-ivopts.h"
40 #include "tree-ssa-loop-manip.h"
41 #include "tree-ssa-loop-niter.h"
42 #include "tree-pass.h"
48 #include "diagnostic-core.h"
49 #include "tree-chrec.h"
50 #include "tree-scalar-evolution.h"
51 #include "tree-vectorizer.h"
54 /* Loop Vectorization Pass.
56 This pass tries to vectorize loops.
58 For example, the vectorizer transforms the following simple loop:
60 short a[N]; short b[N]; short c[N]; int i;
66 as if it was manually vectorized by rewriting the source code into:
68 typedef int __attribute__((mode(V8HI))) v8hi;
69 short a[N]; short b[N]; short c[N]; int i;
70 v8hi *pa = (v8hi*)a, *pb = (v8hi*)b, *pc = (v8hi*)c;
73 for (i=0; i<N/8; i++){
80 The main entry to this pass is vectorize_loops(), in which
81 the vectorizer applies a set of analyses on a given set of loops,
82 followed by the actual vectorization transformation for the loops that
83 had successfully passed the analysis phase.
84 Throughout this pass we make a distinction between two types of
85 data: scalars (which are represented by SSA_NAMES), and memory references
86 ("data-refs"). These two types of data require different handling both
87 during analysis and transformation. The types of data-refs that the
88 vectorizer currently supports are ARRAY_REFS which base is an array DECL
89 (not a pointer), and INDIRECT_REFS through pointers; both array and pointer
90 accesses are required to have a simple (consecutive) access pattern.
94 The driver for the analysis phase is vect_analyze_loop().
95 It applies a set of analyses, some of which rely on the scalar evolution
96 analyzer (scev) developed by Sebastian Pop.
98 During the analysis phase the vectorizer records some information
99 per stmt in a "stmt_vec_info" struct which is attached to each stmt in the
100 loop, as well as general information about the loop as a whole, which is
101 recorded in a "loop_vec_info" struct attached to each loop.
103 Transformation phase:
104 =====================
105 The loop transformation phase scans all the stmts in the loop, and
106 creates a vector stmt (or a sequence of stmts) for each scalar stmt S in
107 the loop that needs to be vectorized. It inserts the vector code sequence
108 just before the scalar stmt S, and records a pointer to the vector code
109 in STMT_VINFO_VEC_STMT (stmt_info) (stmt_info is the stmt_vec_info struct
110 attached to S). This pointer will be used for the vectorization of following
111 stmts which use the def of stmt S. Stmt S is removed if it writes to memory;
112 otherwise, we rely on dead code elimination for removing it.
114 For example, say stmt S1 was vectorized into stmt VS1:
117 S1: b = x[i]; STMT_VINFO_VEC_STMT (stmt_info (S1)) = VS1
120 To vectorize stmt S2, the vectorizer first finds the stmt that defines
121 the operand 'b' (S1), and gets the relevant vector def 'vb' from the
122 vector stmt VS1 pointed to by STMT_VINFO_VEC_STMT (stmt_info (S1)). The
123 resulting sequence would be:
126 S1: b = x[i]; STMT_VINFO_VEC_STMT (stmt_info (S1)) = VS1
128 S2: a = b; STMT_VINFO_VEC_STMT (stmt_info (S2)) = VS2
130 Operands that are not SSA_NAMEs, are data-refs that appear in
131 load/store operations (like 'x[i]' in S1), and are handled differently.
135 Currently the only target specific information that is used is the
136 size of the vector (in bytes) - "TARGET_VECTORIZE_UNITS_PER_SIMD_WORD".
137 Targets that can support different sizes of vectors, for now will need
138 to specify one value for "TARGET_VECTORIZE_UNITS_PER_SIMD_WORD". More
139 flexibility will be added in the future.
141 Since we only vectorize operations which vector form can be
142 expressed using existing tree codes, to verify that an operation is
143 supported, the vectorizer checks the relevant optab at the relevant
144 machine_mode (e.g, optab_handler (add_optab, V8HImode)). If
145 the value found is CODE_FOR_nothing, then there's no target support, and
146 we can't vectorize the stmt.
148 For additional information on this project see:
149 http://gcc.gnu.org/projects/tree-ssa/vectorization.html
152 static void vect_estimate_min_profitable_iters (loop_vec_info
, int *, int *);
154 /* Function vect_determine_vectorization_factor
156 Determine the vectorization factor (VF). VF is the number of data elements
157 that are operated upon in parallel in a single iteration of the vectorized
158 loop. For example, when vectorizing a loop that operates on 4byte elements,
159 on a target with vector size (VS) 16byte, the VF is set to 4, since 4
160 elements can fit in a single vector register.
162 We currently support vectorization of loops in which all types operated upon
163 are of the same size. Therefore this function currently sets VF according to
164 the size of the types operated upon, and fails if there are multiple sizes
167 VF is also the factor by which the loop iterations are strip-mined, e.g.:
174 for (i=0; i<N; i+=VF){
175 a[i:VF] = b[i:VF] + c[i:VF];
180 vect_determine_vectorization_factor (loop_vec_info loop_vinfo
)
182 struct loop
*loop
= LOOP_VINFO_LOOP (loop_vinfo
);
183 basic_block
*bbs
= LOOP_VINFO_BBS (loop_vinfo
);
184 int nbbs
= loop
->num_nodes
;
185 gimple_stmt_iterator si
;
186 unsigned int vectorization_factor
= 0;
191 stmt_vec_info stmt_info
;
194 gimple stmt
, pattern_stmt
= NULL
;
195 gimple_seq pattern_def_seq
= NULL
;
196 gimple_stmt_iterator pattern_def_si
= gsi_none ();
197 bool analyze_pattern_stmt
= false;
199 if (dump_enabled_p ())
200 dump_printf_loc (MSG_NOTE
, vect_location
,
201 "=== vect_determine_vectorization_factor ===\n");
203 for (i
= 0; i
< nbbs
; i
++)
205 basic_block bb
= bbs
[i
];
207 for (si
= gsi_start_phis (bb
); !gsi_end_p (si
); gsi_next (&si
))
210 stmt_info
= vinfo_for_stmt (phi
);
211 if (dump_enabled_p ())
213 dump_printf_loc (MSG_NOTE
, vect_location
, "==> examining phi: ");
214 dump_gimple_stmt (MSG_NOTE
, TDF_SLIM
, phi
, 0);
215 dump_printf (MSG_NOTE
, "\n");
218 gcc_assert (stmt_info
);
220 if (STMT_VINFO_RELEVANT_P (stmt_info
))
222 gcc_assert (!STMT_VINFO_VECTYPE (stmt_info
));
223 scalar_type
= TREE_TYPE (PHI_RESULT (phi
));
225 if (dump_enabled_p ())
227 dump_printf_loc (MSG_NOTE
, vect_location
,
228 "get vectype for scalar type: ");
229 dump_generic_expr (MSG_NOTE
, TDF_SLIM
, scalar_type
);
230 dump_printf (MSG_NOTE
, "\n");
233 vectype
= get_vectype_for_scalar_type (scalar_type
);
236 if (dump_enabled_p ())
238 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
239 "not vectorized: unsupported "
241 dump_generic_expr (MSG_MISSED_OPTIMIZATION
, TDF_SLIM
,
243 dump_printf (MSG_MISSED_OPTIMIZATION
, "\n");
247 STMT_VINFO_VECTYPE (stmt_info
) = vectype
;
249 if (dump_enabled_p ())
251 dump_printf_loc (MSG_NOTE
, vect_location
, "vectype: ");
252 dump_generic_expr (MSG_NOTE
, TDF_SLIM
, vectype
);
253 dump_printf (MSG_NOTE
, "\n");
256 nunits
= TYPE_VECTOR_SUBPARTS (vectype
);
257 if (dump_enabled_p ())
258 dump_printf_loc (MSG_NOTE
, vect_location
, "nunits = %d\n",
261 if (!vectorization_factor
262 || (nunits
> vectorization_factor
))
263 vectorization_factor
= nunits
;
267 for (si
= gsi_start_bb (bb
); !gsi_end_p (si
) || analyze_pattern_stmt
;)
271 if (analyze_pattern_stmt
)
274 stmt
= gsi_stmt (si
);
276 stmt_info
= vinfo_for_stmt (stmt
);
278 if (dump_enabled_p ())
280 dump_printf_loc (MSG_NOTE
, vect_location
,
281 "==> examining statement: ");
282 dump_gimple_stmt (MSG_NOTE
, TDF_SLIM
, stmt
, 0);
283 dump_printf (MSG_NOTE
, "\n");
286 gcc_assert (stmt_info
);
288 /* Skip stmts which do not need to be vectorized. */
289 if ((!STMT_VINFO_RELEVANT_P (stmt_info
)
290 && !STMT_VINFO_LIVE_P (stmt_info
))
291 || gimple_clobber_p (stmt
))
293 if (STMT_VINFO_IN_PATTERN_P (stmt_info
)
294 && (pattern_stmt
= STMT_VINFO_RELATED_STMT (stmt_info
))
295 && (STMT_VINFO_RELEVANT_P (vinfo_for_stmt (pattern_stmt
))
296 || STMT_VINFO_LIVE_P (vinfo_for_stmt (pattern_stmt
))))
299 stmt_info
= vinfo_for_stmt (pattern_stmt
);
300 if (dump_enabled_p ())
302 dump_printf_loc (MSG_NOTE
, vect_location
,
303 "==> examining pattern statement: ");
304 dump_gimple_stmt (MSG_NOTE
, TDF_SLIM
, stmt
, 0);
305 dump_printf (MSG_NOTE
, "\n");
310 if (dump_enabled_p ())
311 dump_printf_loc (MSG_NOTE
, vect_location
, "skip.\n");
316 else if (STMT_VINFO_IN_PATTERN_P (stmt_info
)
317 && (pattern_stmt
= STMT_VINFO_RELATED_STMT (stmt_info
))
318 && (STMT_VINFO_RELEVANT_P (vinfo_for_stmt (pattern_stmt
))
319 || STMT_VINFO_LIVE_P (vinfo_for_stmt (pattern_stmt
))))
320 analyze_pattern_stmt
= true;
322 /* If a pattern statement has def stmts, analyze them too. */
323 if (is_pattern_stmt_p (stmt_info
))
325 if (pattern_def_seq
== NULL
)
327 pattern_def_seq
= STMT_VINFO_PATTERN_DEF_SEQ (stmt_info
);
328 pattern_def_si
= gsi_start (pattern_def_seq
);
330 else if (!gsi_end_p (pattern_def_si
))
331 gsi_next (&pattern_def_si
);
332 if (pattern_def_seq
!= NULL
)
334 gimple pattern_def_stmt
= NULL
;
335 stmt_vec_info pattern_def_stmt_info
= NULL
;
337 while (!gsi_end_p (pattern_def_si
))
339 pattern_def_stmt
= gsi_stmt (pattern_def_si
);
340 pattern_def_stmt_info
341 = vinfo_for_stmt (pattern_def_stmt
);
342 if (STMT_VINFO_RELEVANT_P (pattern_def_stmt_info
)
343 || STMT_VINFO_LIVE_P (pattern_def_stmt_info
))
345 gsi_next (&pattern_def_si
);
348 if (!gsi_end_p (pattern_def_si
))
350 if (dump_enabled_p ())
352 dump_printf_loc (MSG_NOTE
, vect_location
,
353 "==> examining pattern def stmt: ");
354 dump_gimple_stmt (MSG_NOTE
, TDF_SLIM
,
355 pattern_def_stmt
, 0);
356 dump_printf (MSG_NOTE
, "\n");
359 stmt
= pattern_def_stmt
;
360 stmt_info
= pattern_def_stmt_info
;
364 pattern_def_si
= gsi_none ();
365 analyze_pattern_stmt
= false;
369 analyze_pattern_stmt
= false;
372 if (gimple_get_lhs (stmt
) == NULL_TREE
)
374 if (dump_enabled_p ())
376 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
377 "not vectorized: irregular stmt.");
378 dump_gimple_stmt (MSG_MISSED_OPTIMIZATION
, TDF_SLIM
, stmt
,
380 dump_printf (MSG_MISSED_OPTIMIZATION
, "\n");
385 if (VECTOR_MODE_P (TYPE_MODE (gimple_expr_type (stmt
))))
387 if (dump_enabled_p ())
389 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
390 "not vectorized: vector stmt in loop:");
391 dump_gimple_stmt (MSG_MISSED_OPTIMIZATION
, TDF_SLIM
, stmt
, 0);
392 dump_printf (MSG_MISSED_OPTIMIZATION
, "\n");
397 if (STMT_VINFO_VECTYPE (stmt_info
))
399 /* The only case when a vectype had been already set is for stmts
400 that contain a dataref, or for "pattern-stmts" (stmts
401 generated by the vectorizer to represent/replace a certain
403 gcc_assert (STMT_VINFO_DATA_REF (stmt_info
)
404 || is_pattern_stmt_p (stmt_info
)
405 || !gsi_end_p (pattern_def_si
));
406 vectype
= STMT_VINFO_VECTYPE (stmt_info
);
410 gcc_assert (!STMT_VINFO_DATA_REF (stmt_info
));
411 scalar_type
= TREE_TYPE (gimple_get_lhs (stmt
));
412 if (dump_enabled_p ())
414 dump_printf_loc (MSG_NOTE
, vect_location
,
415 "get vectype for scalar type: ");
416 dump_generic_expr (MSG_NOTE
, TDF_SLIM
, scalar_type
);
417 dump_printf (MSG_NOTE
, "\n");
419 vectype
= get_vectype_for_scalar_type (scalar_type
);
422 if (dump_enabled_p ())
424 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
425 "not vectorized: unsupported "
427 dump_generic_expr (MSG_MISSED_OPTIMIZATION
, TDF_SLIM
,
429 dump_printf (MSG_MISSED_OPTIMIZATION
, "\n");
434 STMT_VINFO_VECTYPE (stmt_info
) = vectype
;
436 if (dump_enabled_p ())
438 dump_printf_loc (MSG_NOTE
, vect_location
, "vectype: ");
439 dump_generic_expr (MSG_NOTE
, TDF_SLIM
, vectype
);
440 dump_printf (MSG_NOTE
, "\n");
444 /* The vectorization factor is according to the smallest
445 scalar type (or the largest vector size, but we only
446 support one vector size per loop). */
447 scalar_type
= vect_get_smallest_scalar_type (stmt
, &dummy
,
449 if (dump_enabled_p ())
451 dump_printf_loc (MSG_NOTE
, vect_location
,
452 "get vectype for scalar type: ");
453 dump_generic_expr (MSG_NOTE
, TDF_SLIM
, scalar_type
);
454 dump_printf (MSG_NOTE
, "\n");
456 vf_vectype
= get_vectype_for_scalar_type (scalar_type
);
459 if (dump_enabled_p ())
461 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
462 "not vectorized: unsupported data-type ");
463 dump_generic_expr (MSG_MISSED_OPTIMIZATION
, TDF_SLIM
,
465 dump_printf (MSG_MISSED_OPTIMIZATION
, "\n");
470 if ((GET_MODE_SIZE (TYPE_MODE (vectype
))
471 != GET_MODE_SIZE (TYPE_MODE (vf_vectype
))))
473 if (dump_enabled_p ())
475 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
476 "not vectorized: different sized vector "
477 "types in statement, ");
478 dump_generic_expr (MSG_MISSED_OPTIMIZATION
, TDF_SLIM
,
480 dump_printf (MSG_MISSED_OPTIMIZATION
, " and ");
481 dump_generic_expr (MSG_MISSED_OPTIMIZATION
, TDF_SLIM
,
483 dump_printf (MSG_MISSED_OPTIMIZATION
, "\n");
488 if (dump_enabled_p ())
490 dump_printf_loc (MSG_NOTE
, vect_location
, "vectype: ");
491 dump_generic_expr (MSG_NOTE
, TDF_SLIM
, vf_vectype
);
492 dump_printf (MSG_NOTE
, "\n");
495 nunits
= TYPE_VECTOR_SUBPARTS (vf_vectype
);
496 if (dump_enabled_p ())
497 dump_printf_loc (MSG_NOTE
, vect_location
, "nunits = %d\n", nunits
);
498 if (!vectorization_factor
499 || (nunits
> vectorization_factor
))
500 vectorization_factor
= nunits
;
502 if (!analyze_pattern_stmt
&& gsi_end_p (pattern_def_si
))
504 pattern_def_seq
= NULL
;
510 /* TODO: Analyze cost. Decide if worth while to vectorize. */
511 if (dump_enabled_p ())
512 dump_printf_loc (MSG_NOTE
, vect_location
, "vectorization factor = %d\n",
513 vectorization_factor
);
514 if (vectorization_factor
<= 1)
516 if (dump_enabled_p ())
517 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
518 "not vectorized: unsupported data-type\n");
521 LOOP_VINFO_VECT_FACTOR (loop_vinfo
) = vectorization_factor
;
527 /* Function vect_is_simple_iv_evolution.
529 FORNOW: A simple evolution of an induction variables in the loop is
530 considered a polynomial evolution. */
533 vect_is_simple_iv_evolution (unsigned loop_nb
, tree access_fn
, tree
* init
,
538 tree evolution_part
= evolution_part_in_loop_num (access_fn
, loop_nb
);
541 /* When there is no evolution in this loop, the evolution function
543 if (evolution_part
== NULL_TREE
)
546 /* When the evolution is a polynomial of degree >= 2
547 the evolution function is not "simple". */
548 if (tree_is_chrec (evolution_part
))
551 step_expr
= evolution_part
;
552 init_expr
= unshare_expr (initial_condition_in_loop_num (access_fn
, loop_nb
));
554 if (dump_enabled_p ())
556 dump_printf_loc (MSG_NOTE
, vect_location
, "step: ");
557 dump_generic_expr (MSG_NOTE
, TDF_SLIM
, step_expr
);
558 dump_printf (MSG_NOTE
, ", init: ");
559 dump_generic_expr (MSG_NOTE
, TDF_SLIM
, init_expr
);
560 dump_printf (MSG_NOTE
, "\n");
566 if (TREE_CODE (step_expr
) != INTEGER_CST
567 && (TREE_CODE (step_expr
) != SSA_NAME
568 || ((bb
= gimple_bb (SSA_NAME_DEF_STMT (step_expr
)))
569 && flow_bb_inside_loop_p (get_loop (cfun
, loop_nb
), bb
))
570 || (!INTEGRAL_TYPE_P (TREE_TYPE (step_expr
))
571 && (!SCALAR_FLOAT_TYPE_P (TREE_TYPE (step_expr
))
572 || !flag_associative_math
)))
573 && (TREE_CODE (step_expr
) != REAL_CST
574 || !flag_associative_math
))
576 if (dump_enabled_p ())
577 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
585 /* Function vect_analyze_scalar_cycles_1.
587 Examine the cross iteration def-use cycles of scalar variables
588 in LOOP. LOOP_VINFO represents the loop that is now being
589 considered for vectorization (can be LOOP, or an outer-loop
593 vect_analyze_scalar_cycles_1 (loop_vec_info loop_vinfo
, struct loop
*loop
)
595 basic_block bb
= loop
->header
;
597 stack_vec
<gimple
, 64> worklist
;
598 gimple_stmt_iterator gsi
;
601 if (dump_enabled_p ())
602 dump_printf_loc (MSG_NOTE
, vect_location
,
603 "=== vect_analyze_scalar_cycles ===\n");
605 /* First - identify all inductions. Reduction detection assumes that all the
606 inductions have been identified, therefore, this order must not be
608 for (gsi
= gsi_start_phis (bb
); !gsi_end_p (gsi
); gsi_next (&gsi
))
610 gimple phi
= gsi_stmt (gsi
);
611 tree access_fn
= NULL
;
612 tree def
= PHI_RESULT (phi
);
613 stmt_vec_info stmt_vinfo
= vinfo_for_stmt (phi
);
615 if (dump_enabled_p ())
617 dump_printf_loc (MSG_NOTE
, vect_location
, "Analyze phi: ");
618 dump_gimple_stmt (MSG_NOTE
, TDF_SLIM
, phi
, 0);
619 dump_printf (MSG_NOTE
, "\n");
622 /* Skip virtual phi's. The data dependences that are associated with
623 virtual defs/uses (i.e., memory accesses) are analyzed elsewhere. */
624 if (virtual_operand_p (def
))
627 STMT_VINFO_DEF_TYPE (stmt_vinfo
) = vect_unknown_def_type
;
629 /* Analyze the evolution function. */
630 access_fn
= analyze_scalar_evolution (loop
, def
);
633 STRIP_NOPS (access_fn
);
634 if (dump_enabled_p ())
636 dump_printf_loc (MSG_NOTE
, vect_location
,
637 "Access function of PHI: ");
638 dump_generic_expr (MSG_NOTE
, TDF_SLIM
, access_fn
);
639 dump_printf (MSG_NOTE
, "\n");
641 STMT_VINFO_LOOP_PHI_EVOLUTION_PART (stmt_vinfo
)
642 = evolution_part_in_loop_num (access_fn
, loop
->num
);
646 || !vect_is_simple_iv_evolution (loop
->num
, access_fn
, &init
, &step
)
647 || (LOOP_VINFO_LOOP (loop_vinfo
) != loop
648 && TREE_CODE (step
) != INTEGER_CST
))
650 worklist
.safe_push (phi
);
654 gcc_assert (STMT_VINFO_LOOP_PHI_EVOLUTION_PART (stmt_vinfo
) != NULL_TREE
);
656 if (dump_enabled_p ())
657 dump_printf_loc (MSG_NOTE
, vect_location
, "Detected induction.\n");
658 STMT_VINFO_DEF_TYPE (stmt_vinfo
) = vect_induction_def
;
662 /* Second - identify all reductions and nested cycles. */
663 while (worklist
.length () > 0)
665 gimple phi
= worklist
.pop ();
666 tree def
= PHI_RESULT (phi
);
667 stmt_vec_info stmt_vinfo
= vinfo_for_stmt (phi
);
671 if (dump_enabled_p ())
673 dump_printf_loc (MSG_NOTE
, vect_location
, "Analyze phi: ");
674 dump_gimple_stmt (MSG_NOTE
, TDF_SLIM
, phi
, 0);
675 dump_printf (MSG_NOTE
, "\n");
678 gcc_assert (!virtual_operand_p (def
)
679 && STMT_VINFO_DEF_TYPE (stmt_vinfo
) == vect_unknown_def_type
);
681 nested_cycle
= (loop
!= LOOP_VINFO_LOOP (loop_vinfo
));
682 reduc_stmt
= vect_force_simple_reduction (loop_vinfo
, phi
, !nested_cycle
,
688 if (dump_enabled_p ())
689 dump_printf_loc (MSG_NOTE
, vect_location
,
690 "Detected double reduction.\n");
692 STMT_VINFO_DEF_TYPE (stmt_vinfo
) = vect_double_reduction_def
;
693 STMT_VINFO_DEF_TYPE (vinfo_for_stmt (reduc_stmt
)) =
694 vect_double_reduction_def
;
700 if (dump_enabled_p ())
701 dump_printf_loc (MSG_NOTE
, vect_location
,
702 "Detected vectorizable nested cycle.\n");
704 STMT_VINFO_DEF_TYPE (stmt_vinfo
) = vect_nested_cycle
;
705 STMT_VINFO_DEF_TYPE (vinfo_for_stmt (reduc_stmt
)) =
710 if (dump_enabled_p ())
711 dump_printf_loc (MSG_NOTE
, vect_location
,
712 "Detected reduction.\n");
714 STMT_VINFO_DEF_TYPE (stmt_vinfo
) = vect_reduction_def
;
715 STMT_VINFO_DEF_TYPE (vinfo_for_stmt (reduc_stmt
)) =
717 /* Store the reduction cycles for possible vectorization in
719 LOOP_VINFO_REDUCTIONS (loop_vinfo
).safe_push (reduc_stmt
);
724 if (dump_enabled_p ())
725 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
726 "Unknown def-use cycle pattern.\n");
731 /* Function vect_analyze_scalar_cycles.
733 Examine the cross iteration def-use cycles of scalar variables, by
734 analyzing the loop-header PHIs of scalar variables. Classify each
735 cycle as one of the following: invariant, induction, reduction, unknown.
736 We do that for the loop represented by LOOP_VINFO, and also to its
737 inner-loop, if exists.
738 Examples for scalar cycles:
753 vect_analyze_scalar_cycles (loop_vec_info loop_vinfo
)
755 struct loop
*loop
= LOOP_VINFO_LOOP (loop_vinfo
);
757 vect_analyze_scalar_cycles_1 (loop_vinfo
, loop
);
759 /* When vectorizing an outer-loop, the inner-loop is executed sequentially.
760 Reductions in such inner-loop therefore have different properties than
761 the reductions in the nest that gets vectorized:
762 1. When vectorized, they are executed in the same order as in the original
763 scalar loop, so we can't change the order of computation when
765 2. FIXME: Inner-loop reductions can be used in the inner-loop, so the
766 current checks are too strict. */
769 vect_analyze_scalar_cycles_1 (loop_vinfo
, loop
->inner
);
772 /* Function vect_get_loop_niters.
774 Determine how many iterations the loop is executed.
775 If an expression that represents the number of iterations
776 can be constructed, place it in NUMBER_OF_ITERATIONS.
777 Return the loop exit condition. */
780 vect_get_loop_niters (struct loop
*loop
, tree
*number_of_iterations
)
784 if (dump_enabled_p ())
785 dump_printf_loc (MSG_NOTE
, vect_location
,
786 "=== get_loop_niters ===\n");
787 niters
= number_of_exit_cond_executions (loop
);
789 if (niters
!= NULL_TREE
790 && niters
!= chrec_dont_know
)
792 *number_of_iterations
= niters
;
794 if (dump_enabled_p ())
796 dump_printf_loc (MSG_NOTE
, vect_location
, "==> get_loop_niters:");
797 dump_generic_expr (MSG_NOTE
, TDF_SLIM
, *number_of_iterations
);
798 dump_printf (MSG_NOTE
, "\n");
802 return get_loop_exit_condition (loop
);
806 /* Function bb_in_loop_p
808 Used as predicate for dfs order traversal of the loop bbs. */
811 bb_in_loop_p (const_basic_block bb
, const void *data
)
813 const struct loop
*const loop
= (const struct loop
*)data
;
814 if (flow_bb_inside_loop_p (loop
, bb
))
820 /* Function new_loop_vec_info.
822 Create and initialize a new loop_vec_info struct for LOOP, as well as
823 stmt_vec_info structs for all the stmts in LOOP. */
826 new_loop_vec_info (struct loop
*loop
)
830 gimple_stmt_iterator si
;
831 unsigned int i
, nbbs
;
833 res
= (loop_vec_info
) xcalloc (1, sizeof (struct _loop_vec_info
));
834 LOOP_VINFO_LOOP (res
) = loop
;
836 bbs
= get_loop_body (loop
);
838 /* Create/Update stmt_info for all stmts in the loop. */
839 for (i
= 0; i
< loop
->num_nodes
; i
++)
841 basic_block bb
= bbs
[i
];
843 /* BBs in a nested inner-loop will have been already processed (because
844 we will have called vect_analyze_loop_form for any nested inner-loop).
845 Therefore, for stmts in an inner-loop we just want to update the
846 STMT_VINFO_LOOP_VINFO field of their stmt_info to point to the new
847 loop_info of the outer-loop we are currently considering to vectorize
848 (instead of the loop_info of the inner-loop).
849 For stmts in other BBs we need to create a stmt_info from scratch. */
850 if (bb
->loop_father
!= loop
)
853 gcc_assert (loop
->inner
&& bb
->loop_father
== loop
->inner
);
854 for (si
= gsi_start_phis (bb
); !gsi_end_p (si
); gsi_next (&si
))
856 gimple phi
= gsi_stmt (si
);
857 stmt_vec_info stmt_info
= vinfo_for_stmt (phi
);
858 loop_vec_info inner_loop_vinfo
=
859 STMT_VINFO_LOOP_VINFO (stmt_info
);
860 gcc_assert (loop
->inner
== LOOP_VINFO_LOOP (inner_loop_vinfo
));
861 STMT_VINFO_LOOP_VINFO (stmt_info
) = res
;
863 for (si
= gsi_start_bb (bb
); !gsi_end_p (si
); gsi_next (&si
))
865 gimple stmt
= gsi_stmt (si
);
866 stmt_vec_info stmt_info
= vinfo_for_stmt (stmt
);
867 loop_vec_info inner_loop_vinfo
=
868 STMT_VINFO_LOOP_VINFO (stmt_info
);
869 gcc_assert (loop
->inner
== LOOP_VINFO_LOOP (inner_loop_vinfo
));
870 STMT_VINFO_LOOP_VINFO (stmt_info
) = res
;
875 /* bb in current nest. */
876 for (si
= gsi_start_phis (bb
); !gsi_end_p (si
); gsi_next (&si
))
878 gimple phi
= gsi_stmt (si
);
879 gimple_set_uid (phi
, 0);
880 set_vinfo_for_stmt (phi
, new_stmt_vec_info (phi
, res
, NULL
));
883 for (si
= gsi_start_bb (bb
); !gsi_end_p (si
); gsi_next (&si
))
885 gimple stmt
= gsi_stmt (si
);
886 gimple_set_uid (stmt
, 0);
887 set_vinfo_for_stmt (stmt
, new_stmt_vec_info (stmt
, res
, NULL
));
892 /* CHECKME: We want to visit all BBs before their successors (except for
893 latch blocks, for which this assertion wouldn't hold). In the simple
894 case of the loop forms we allow, a dfs order of the BBs would the same
895 as reversed postorder traversal, so we are safe. */
898 bbs
= XCNEWVEC (basic_block
, loop
->num_nodes
);
899 nbbs
= dfs_enumerate_from (loop
->header
, 0, bb_in_loop_p
,
900 bbs
, loop
->num_nodes
, loop
);
901 gcc_assert (nbbs
== loop
->num_nodes
);
903 LOOP_VINFO_BBS (res
) = bbs
;
904 LOOP_VINFO_NITERS (res
) = NULL
;
905 LOOP_VINFO_NITERS_UNCHANGED (res
) = NULL
;
906 LOOP_VINFO_COST_MODEL_MIN_ITERS (res
) = 0;
907 LOOP_VINFO_VECTORIZABLE_P (res
) = 0;
908 LOOP_PEELING_FOR_ALIGNMENT (res
) = 0;
909 LOOP_VINFO_VECT_FACTOR (res
) = 0;
910 LOOP_VINFO_LOOP_NEST (res
).create (3);
911 LOOP_VINFO_DATAREFS (res
).create (10);
912 LOOP_VINFO_DDRS (res
).create (10 * 10);
913 LOOP_VINFO_UNALIGNED_DR (res
) = NULL
;
914 LOOP_VINFO_MAY_MISALIGN_STMTS (res
).create (
915 PARAM_VALUE (PARAM_VECT_MAX_VERSION_FOR_ALIGNMENT_CHECKS
));
916 LOOP_VINFO_MAY_ALIAS_DDRS (res
).create (
917 PARAM_VALUE (PARAM_VECT_MAX_VERSION_FOR_ALIAS_CHECKS
));
918 LOOP_VINFO_GROUPED_STORES (res
).create (10);
919 LOOP_VINFO_REDUCTIONS (res
).create (10);
920 LOOP_VINFO_REDUCTION_CHAINS (res
).create (10);
921 LOOP_VINFO_SLP_INSTANCES (res
).create (10);
922 LOOP_VINFO_SLP_UNROLLING_FACTOR (res
) = 1;
923 LOOP_VINFO_TARGET_COST_DATA (res
) = init_cost (loop
);
924 LOOP_VINFO_PEELING_FOR_GAPS (res
) = false;
925 LOOP_VINFO_OPERANDS_SWAPPED (res
) = false;
931 /* Function destroy_loop_vec_info.
933 Free LOOP_VINFO struct, as well as all the stmt_vec_info structs of all the
934 stmts in the loop. */
937 destroy_loop_vec_info (loop_vec_info loop_vinfo
, bool clean_stmts
)
942 gimple_stmt_iterator si
;
944 vec
<slp_instance
> slp_instances
;
945 slp_instance instance
;
951 loop
= LOOP_VINFO_LOOP (loop_vinfo
);
953 bbs
= LOOP_VINFO_BBS (loop_vinfo
);
954 nbbs
= clean_stmts
? loop
->num_nodes
: 0;
955 swapped
= LOOP_VINFO_OPERANDS_SWAPPED (loop_vinfo
);
957 for (j
= 0; j
< nbbs
; j
++)
959 basic_block bb
= bbs
[j
];
960 for (si
= gsi_start_phis (bb
); !gsi_end_p (si
); gsi_next (&si
))
961 free_stmt_vec_info (gsi_stmt (si
));
963 for (si
= gsi_start_bb (bb
); !gsi_end_p (si
); )
965 gimple stmt
= gsi_stmt (si
);
967 /* We may have broken canonical form by moving a constant
968 into RHS1 of a commutative op. Fix such occurrences. */
969 if (swapped
&& is_gimple_assign (stmt
))
971 enum tree_code code
= gimple_assign_rhs_code (stmt
);
973 if ((code
== PLUS_EXPR
974 || code
== POINTER_PLUS_EXPR
975 || code
== MULT_EXPR
)
976 && CONSTANT_CLASS_P (gimple_assign_rhs1 (stmt
)))
977 swap_ssa_operands (stmt
,
978 gimple_assign_rhs1_ptr (stmt
),
979 gimple_assign_rhs2_ptr (stmt
));
982 /* Free stmt_vec_info. */
983 free_stmt_vec_info (stmt
);
988 free (LOOP_VINFO_BBS (loop_vinfo
));
989 vect_destroy_datarefs (loop_vinfo
, NULL
);
990 free_dependence_relations (LOOP_VINFO_DDRS (loop_vinfo
));
991 LOOP_VINFO_LOOP_NEST (loop_vinfo
).release ();
992 LOOP_VINFO_MAY_MISALIGN_STMTS (loop_vinfo
).release ();
993 LOOP_VINFO_MAY_ALIAS_DDRS (loop_vinfo
).release ();
994 slp_instances
= LOOP_VINFO_SLP_INSTANCES (loop_vinfo
);
995 FOR_EACH_VEC_ELT (slp_instances
, j
, instance
)
996 vect_free_slp_instance (instance
);
998 LOOP_VINFO_SLP_INSTANCES (loop_vinfo
).release ();
999 LOOP_VINFO_GROUPED_STORES (loop_vinfo
).release ();
1000 LOOP_VINFO_REDUCTIONS (loop_vinfo
).release ();
1001 LOOP_VINFO_REDUCTION_CHAINS (loop_vinfo
).release ();
1003 if (LOOP_VINFO_PEELING_HTAB (loop_vinfo
).is_created ())
1004 LOOP_VINFO_PEELING_HTAB (loop_vinfo
).dispose ();
1006 destroy_cost_data (LOOP_VINFO_TARGET_COST_DATA (loop_vinfo
));
1013 /* Function vect_analyze_loop_1.
1015 Apply a set of analyses on LOOP, and create a loop_vec_info struct
1016 for it. The different analyses will record information in the
1017 loop_vec_info struct. This is a subset of the analyses applied in
1018 vect_analyze_loop, to be applied on an inner-loop nested in the loop
1019 that is now considered for (outer-loop) vectorization. */
1021 static loop_vec_info
1022 vect_analyze_loop_1 (struct loop
*loop
)
1024 loop_vec_info loop_vinfo
;
1026 if (dump_enabled_p ())
1027 dump_printf_loc (MSG_NOTE
, vect_location
,
1028 "===== analyze_loop_nest_1 =====\n");
1030 /* Check the CFG characteristics of the loop (nesting, entry/exit, etc. */
1032 loop_vinfo
= vect_analyze_loop_form (loop
);
1035 if (dump_enabled_p ())
1036 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
1037 "bad inner-loop form.\n");
1045 /* Function vect_analyze_loop_form.
1047 Verify that certain CFG restrictions hold, including:
1048 - the loop has a pre-header
1049 - the loop has a single entry and exit
1050 - the loop exit condition is simple enough, and the number of iterations
1051 can be analyzed (a countable loop). */
1054 vect_analyze_loop_form (struct loop
*loop
)
1056 loop_vec_info loop_vinfo
;
1058 tree number_of_iterations
= NULL
;
1059 loop_vec_info inner_loop_vinfo
= NULL
;
1061 if (dump_enabled_p ())
1062 dump_printf_loc (MSG_NOTE
, vect_location
,
1063 "=== vect_analyze_loop_form ===\n");
1065 /* Different restrictions apply when we are considering an inner-most loop,
1066 vs. an outer (nested) loop.
1067 (FORNOW. May want to relax some of these restrictions in the future). */
1071 /* Inner-most loop. We currently require that the number of BBs is
1072 exactly 2 (the header and latch). Vectorizable inner-most loops
1083 if (loop
->num_nodes
!= 2)
1085 if (dump_enabled_p ())
1086 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
1087 "not vectorized: control flow in loop.\n");
1091 if (empty_block_p (loop
->header
))
1093 if (dump_enabled_p ())
1094 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
1095 "not vectorized: empty loop.\n");
1101 struct loop
*innerloop
= loop
->inner
;
1104 /* Nested loop. We currently require that the loop is doubly-nested,
1105 contains a single inner loop, and the number of BBs is exactly 5.
1106 Vectorizable outer-loops look like this:
1118 The inner-loop has the properties expected of inner-most loops
1119 as described above. */
1121 if ((loop
->inner
)->inner
|| (loop
->inner
)->next
)
1123 if (dump_enabled_p ())
1124 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
1125 "not vectorized: multiple nested loops.\n");
1129 /* Analyze the inner-loop. */
1130 inner_loop_vinfo
= vect_analyze_loop_1 (loop
->inner
);
1131 if (!inner_loop_vinfo
)
1133 if (dump_enabled_p ())
1134 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
1135 "not vectorized: Bad inner loop.\n");
1139 if (!expr_invariant_in_loop_p (loop
,
1140 LOOP_VINFO_NITERS (inner_loop_vinfo
)))
1142 if (dump_enabled_p ())
1143 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
1144 "not vectorized: inner-loop count not"
1146 destroy_loop_vec_info (inner_loop_vinfo
, true);
1150 if (loop
->num_nodes
!= 5)
1152 if (dump_enabled_p ())
1153 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
1154 "not vectorized: control flow in loop.\n");
1155 destroy_loop_vec_info (inner_loop_vinfo
, true);
1159 gcc_assert (EDGE_COUNT (innerloop
->header
->preds
) == 2);
1160 entryedge
= EDGE_PRED (innerloop
->header
, 0);
1161 if (EDGE_PRED (innerloop
->header
, 0)->src
== innerloop
->latch
)
1162 entryedge
= EDGE_PRED (innerloop
->header
, 1);
1164 if (entryedge
->src
!= loop
->header
1165 || !single_exit (innerloop
)
1166 || single_exit (innerloop
)->dest
!= EDGE_PRED (loop
->latch
, 0)->src
)
1168 if (dump_enabled_p ())
1169 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
1170 "not vectorized: unsupported outerloop form.\n");
1171 destroy_loop_vec_info (inner_loop_vinfo
, true);
1175 if (dump_enabled_p ())
1176 dump_printf_loc (MSG_NOTE
, vect_location
,
1177 "Considering outer-loop vectorization.\n");
1180 if (!single_exit (loop
)
1181 || EDGE_COUNT (loop
->header
->preds
) != 2)
1183 if (dump_enabled_p ())
1185 if (!single_exit (loop
))
1186 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
1187 "not vectorized: multiple exits.\n");
1188 else if (EDGE_COUNT (loop
->header
->preds
) != 2)
1189 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
1190 "not vectorized: too many incoming edges.\n");
1192 if (inner_loop_vinfo
)
1193 destroy_loop_vec_info (inner_loop_vinfo
, true);
1197 /* We assume that the loop exit condition is at the end of the loop. i.e,
1198 that the loop is represented as a do-while (with a proper if-guard
1199 before the loop if needed), where the loop header contains all the
1200 executable statements, and the latch is empty. */
1201 if (!empty_block_p (loop
->latch
)
1202 || !gimple_seq_empty_p (phi_nodes (loop
->latch
)))
1204 if (dump_enabled_p ())
1205 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
1206 "not vectorized: latch block not empty.\n");
1207 if (inner_loop_vinfo
)
1208 destroy_loop_vec_info (inner_loop_vinfo
, true);
1212 /* Make sure there exists a single-predecessor exit bb: */
1213 if (!single_pred_p (single_exit (loop
)->dest
))
1215 edge e
= single_exit (loop
);
1216 if (!(e
->flags
& EDGE_ABNORMAL
))
1218 split_loop_exit_edge (e
);
1219 if (dump_enabled_p ())
1220 dump_printf (MSG_NOTE
, "split exit edge.\n");
1224 if (dump_enabled_p ())
1225 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
1226 "not vectorized: abnormal loop exit edge.\n");
1227 if (inner_loop_vinfo
)
1228 destroy_loop_vec_info (inner_loop_vinfo
, true);
1233 loop_cond
= vect_get_loop_niters (loop
, &number_of_iterations
);
1236 if (dump_enabled_p ())
1237 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
1238 "not vectorized: complicated exit condition.\n");
1239 if (inner_loop_vinfo
)
1240 destroy_loop_vec_info (inner_loop_vinfo
, true);
1244 if (!number_of_iterations
)
1246 if (dump_enabled_p ())
1247 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
1248 "not vectorized: number of iterations cannot be "
1250 if (inner_loop_vinfo
)
1251 destroy_loop_vec_info (inner_loop_vinfo
, true);
1255 if (chrec_contains_undetermined (number_of_iterations
))
1257 if (dump_enabled_p ())
1258 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
1259 "Infinite number of iterations.\n");
1260 if (inner_loop_vinfo
)
1261 destroy_loop_vec_info (inner_loop_vinfo
, true);
1265 if (!NITERS_KNOWN_P (number_of_iterations
))
1267 if (dump_enabled_p ())
1269 dump_printf_loc (MSG_NOTE
, vect_location
,
1270 "Symbolic number of iterations is ");
1271 dump_generic_expr (MSG_NOTE
, TDF_DETAILS
, number_of_iterations
);
1272 dump_printf (MSG_NOTE
, "\n");
1275 else if (TREE_INT_CST_LOW (number_of_iterations
) == 0)
1277 if (dump_enabled_p ())
1278 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
1279 "not vectorized: number of iterations = 0.\n");
1280 if (inner_loop_vinfo
)
1281 destroy_loop_vec_info (inner_loop_vinfo
, true);
1285 loop_vinfo
= new_loop_vec_info (loop
);
1286 LOOP_VINFO_NITERS (loop_vinfo
) = number_of_iterations
;
1287 LOOP_VINFO_NITERS_UNCHANGED (loop_vinfo
) = number_of_iterations
;
1289 STMT_VINFO_TYPE (vinfo_for_stmt (loop_cond
)) = loop_exit_ctrl_vec_info_type
;
1291 /* CHECKME: May want to keep it around it in the future. */
1292 if (inner_loop_vinfo
)
1293 destroy_loop_vec_info (inner_loop_vinfo
, false);
1295 gcc_assert (!loop
->aux
);
1296 loop
->aux
= loop_vinfo
;
1301 /* Function vect_analyze_loop_operations.
1303 Scan the loop stmts and make sure they are all vectorizable. */
1306 vect_analyze_loop_operations (loop_vec_info loop_vinfo
, bool slp
)
1308 struct loop
*loop
= LOOP_VINFO_LOOP (loop_vinfo
);
1309 basic_block
*bbs
= LOOP_VINFO_BBS (loop_vinfo
);
1310 int nbbs
= loop
->num_nodes
;
1311 gimple_stmt_iterator si
;
1312 unsigned int vectorization_factor
= 0;
1315 stmt_vec_info stmt_info
;
1316 bool need_to_vectorize
= false;
1317 int min_profitable_iters
;
1318 int min_scalar_loop_bound
;
1320 bool only_slp_in_loop
= true, ok
;
1321 HOST_WIDE_INT max_niter
;
1322 HOST_WIDE_INT estimated_niter
;
1323 int min_profitable_estimate
;
1325 if (dump_enabled_p ())
1326 dump_printf_loc (MSG_NOTE
, vect_location
,
1327 "=== vect_analyze_loop_operations ===\n");
1329 gcc_assert (LOOP_VINFO_VECT_FACTOR (loop_vinfo
));
1330 vectorization_factor
= LOOP_VINFO_VECT_FACTOR (loop_vinfo
);
1333 /* If all the stmts in the loop can be SLPed, we perform only SLP, and
1334 vectorization factor of the loop is the unrolling factor required by
1335 the SLP instances. If that unrolling factor is 1, we say, that we
1336 perform pure SLP on loop - cross iteration parallelism is not
1338 for (i
= 0; i
< nbbs
; i
++)
1340 basic_block bb
= bbs
[i
];
1341 for (si
= gsi_start_bb (bb
); !gsi_end_p (si
); gsi_next (&si
))
1343 gimple stmt
= gsi_stmt (si
);
1344 stmt_vec_info stmt_info
= vinfo_for_stmt (stmt
);
1345 gcc_assert (stmt_info
);
1346 if ((STMT_VINFO_RELEVANT_P (stmt_info
)
1347 || VECTORIZABLE_CYCLE_DEF (STMT_VINFO_DEF_TYPE (stmt_info
)))
1348 && !PURE_SLP_STMT (stmt_info
))
1349 /* STMT needs both SLP and loop-based vectorization. */
1350 only_slp_in_loop
= false;
1354 if (only_slp_in_loop
)
1355 vectorization_factor
= LOOP_VINFO_SLP_UNROLLING_FACTOR (loop_vinfo
);
1357 vectorization_factor
= least_common_multiple (vectorization_factor
,
1358 LOOP_VINFO_SLP_UNROLLING_FACTOR (loop_vinfo
));
1360 LOOP_VINFO_VECT_FACTOR (loop_vinfo
) = vectorization_factor
;
1361 if (dump_enabled_p ())
1362 dump_printf_loc (MSG_NOTE
, vect_location
,
1363 "Updating vectorization factor to %d\n",
1364 vectorization_factor
);
1367 for (i
= 0; i
< nbbs
; i
++)
1369 basic_block bb
= bbs
[i
];
1371 for (si
= gsi_start_phis (bb
); !gsi_end_p (si
); gsi_next (&si
))
1373 phi
= gsi_stmt (si
);
1376 stmt_info
= vinfo_for_stmt (phi
);
1377 if (dump_enabled_p ())
1379 dump_printf_loc (MSG_NOTE
, vect_location
, "examining phi: ");
1380 dump_gimple_stmt (MSG_NOTE
, TDF_SLIM
, phi
, 0);
1381 dump_printf (MSG_NOTE
, "\n");
1384 /* Inner-loop loop-closed exit phi in outer-loop vectorization
1385 (i.e., a phi in the tail of the outer-loop). */
1386 if (! is_loop_header_bb_p (bb
))
1388 /* FORNOW: we currently don't support the case that these phis
1389 are not used in the outerloop (unless it is double reduction,
1390 i.e., this phi is vect_reduction_def), cause this case
1391 requires to actually do something here. */
1392 if ((!STMT_VINFO_RELEVANT_P (stmt_info
)
1393 || STMT_VINFO_LIVE_P (stmt_info
))
1394 && STMT_VINFO_DEF_TYPE (stmt_info
)
1395 != vect_double_reduction_def
)
1397 if (dump_enabled_p ())
1398 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
1399 "Unsupported loop-closed phi in "
1404 /* If PHI is used in the outer loop, we check that its operand
1405 is defined in the inner loop. */
1406 if (STMT_VINFO_RELEVANT_P (stmt_info
))
1411 if (gimple_phi_num_args (phi
) != 1)
1414 phi_op
= PHI_ARG_DEF (phi
, 0);
1415 if (TREE_CODE (phi_op
) != SSA_NAME
)
1418 op_def_stmt
= SSA_NAME_DEF_STMT (phi_op
);
1419 if (gimple_nop_p (op_def_stmt
)
1420 || !flow_bb_inside_loop_p (loop
, gimple_bb (op_def_stmt
))
1421 || !vinfo_for_stmt (op_def_stmt
))
1424 if (STMT_VINFO_RELEVANT (vinfo_for_stmt (op_def_stmt
))
1425 != vect_used_in_outer
1426 && STMT_VINFO_RELEVANT (vinfo_for_stmt (op_def_stmt
))
1427 != vect_used_in_outer_by_reduction
)
1434 gcc_assert (stmt_info
);
1436 if (STMT_VINFO_LIVE_P (stmt_info
))
1438 /* FORNOW: not yet supported. */
1439 if (dump_enabled_p ())
1440 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
1441 "not vectorized: value used after loop.\n");
1445 if (STMT_VINFO_RELEVANT (stmt_info
) == vect_used_in_scope
1446 && STMT_VINFO_DEF_TYPE (stmt_info
) != vect_induction_def
)
1448 /* A scalar-dependence cycle that we don't support. */
1449 if (dump_enabled_p ())
1450 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
1451 "not vectorized: scalar dependence cycle.\n");
1455 if (STMT_VINFO_RELEVANT_P (stmt_info
))
1457 need_to_vectorize
= true;
1458 if (STMT_VINFO_DEF_TYPE (stmt_info
) == vect_induction_def
)
1459 ok
= vectorizable_induction (phi
, NULL
, NULL
);
1464 if (dump_enabled_p ())
1466 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
1467 "not vectorized: relevant phi not "
1469 dump_gimple_stmt (MSG_MISSED_OPTIMIZATION
, TDF_SLIM
, phi
, 0);
1470 dump_printf (MSG_MISSED_OPTIMIZATION
, "\n");
1476 for (si
= gsi_start_bb (bb
); !gsi_end_p (si
); gsi_next (&si
))
1478 gimple stmt
= gsi_stmt (si
);
1479 if (!gimple_clobber_p (stmt
)
1480 && !vect_analyze_stmt (stmt
, &need_to_vectorize
, NULL
))
1485 /* All operations in the loop are either irrelevant (deal with loop
1486 control, or dead), or only used outside the loop and can be moved
1487 out of the loop (e.g. invariants, inductions). The loop can be
1488 optimized away by scalar optimizations. We're better off not
1489 touching this loop. */
1490 if (!need_to_vectorize
)
1492 if (dump_enabled_p ())
1493 dump_printf_loc (MSG_NOTE
, vect_location
,
1494 "All the computation can be taken out of the loop.\n");
1495 if (dump_enabled_p ())
1496 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
1497 "not vectorized: redundant loop. no profit to "
1502 if (LOOP_VINFO_NITERS_KNOWN_P (loop_vinfo
) && dump_enabled_p ())
1503 dump_printf_loc (MSG_NOTE
, vect_location
,
1504 "vectorization_factor = %d, niters = "
1505 HOST_WIDE_INT_PRINT_DEC
"\n", vectorization_factor
,
1506 LOOP_VINFO_INT_NITERS (loop_vinfo
));
1508 if ((LOOP_VINFO_NITERS_KNOWN_P (loop_vinfo
)
1509 && (LOOP_VINFO_INT_NITERS (loop_vinfo
) < vectorization_factor
))
1510 || ((max_niter
= max_stmt_executions_int (loop
)) != -1
1511 && (unsigned HOST_WIDE_INT
) max_niter
< vectorization_factor
))
1513 if (dump_enabled_p ())
1514 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
1515 "not vectorized: iteration count too small.\n");
1516 if (dump_enabled_p ())
1517 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
1518 "not vectorized: iteration count smaller than "
1519 "vectorization factor.\n");
1523 /* Analyze cost. Decide if worth while to vectorize. */
1525 /* Once VF is set, SLP costs should be updated since the number of created
1526 vector stmts depends on VF. */
1527 vect_update_slp_costs_according_to_vf (loop_vinfo
);
1529 vect_estimate_min_profitable_iters (loop_vinfo
, &min_profitable_iters
,
1530 &min_profitable_estimate
);
1531 LOOP_VINFO_COST_MODEL_MIN_ITERS (loop_vinfo
) = min_profitable_iters
;
1533 if (min_profitable_iters
< 0)
1535 if (dump_enabled_p ())
1536 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
1537 "not vectorized: vectorization not profitable.\n");
1538 if (dump_enabled_p ())
1539 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
1540 "not vectorized: vector version will never be "
1545 min_scalar_loop_bound
= ((PARAM_VALUE (PARAM_MIN_VECT_LOOP_BOUND
)
1546 * vectorization_factor
) - 1);
1549 /* Use the cost model only if it is more conservative than user specified
1552 th
= (unsigned) min_scalar_loop_bound
;
1553 if (min_profitable_iters
1554 && (!min_scalar_loop_bound
1555 || min_profitable_iters
> min_scalar_loop_bound
))
1556 th
= (unsigned) min_profitable_iters
;
1558 if (LOOP_VINFO_NITERS_KNOWN_P (loop_vinfo
)
1559 && LOOP_VINFO_INT_NITERS (loop_vinfo
) <= th
)
1561 if (dump_enabled_p ())
1562 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
1563 "not vectorized: vectorization not profitable.\n");
1564 if (dump_enabled_p ())
1565 dump_printf_loc (MSG_NOTE
, vect_location
,
1566 "not vectorized: iteration count smaller than user "
1567 "specified loop bound parameter or minimum profitable "
1568 "iterations (whichever is more conservative).\n");
1572 if ((estimated_niter
= estimated_stmt_executions_int (loop
)) != -1
1573 && ((unsigned HOST_WIDE_INT
) estimated_niter
1574 <= MAX (th
, (unsigned)min_profitable_estimate
)))
1576 if (dump_enabled_p ())
1577 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
1578 "not vectorized: estimated iteration count too "
1580 if (dump_enabled_p ())
1581 dump_printf_loc (MSG_NOTE
, vect_location
,
1582 "not vectorized: estimated iteration count smaller "
1583 "than specified loop bound parameter or minimum "
1584 "profitable iterations (whichever is more "
1585 "conservative).\n");
1589 if (LOOP_PEELING_FOR_ALIGNMENT (loop_vinfo
)
1590 || ((int) tree_ctz (LOOP_VINFO_NITERS (loop_vinfo
))
1591 < exact_log2 (vectorization_factor
)))
1593 if (dump_enabled_p ())
1594 dump_printf_loc (MSG_NOTE
, vect_location
, "epilog loop required.\n");
1595 if (!vect_can_advance_ivs_p (loop_vinfo
))
1597 if (dump_enabled_p ())
1598 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
1599 "not vectorized: can't create epilog loop 1.\n");
1602 if (!slpeel_can_duplicate_loop_p (loop
, single_exit (loop
)))
1604 if (dump_enabled_p ())
1605 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
1606 "not vectorized: can't create epilog loop 2.\n");
1615 /* Function vect_analyze_loop_2.
1617 Apply a set of analyses on LOOP, and create a loop_vec_info struct
1618 for it. The different analyses will record information in the
1619 loop_vec_info struct. */
1621 vect_analyze_loop_2 (loop_vec_info loop_vinfo
)
1623 bool ok
, slp
= false;
1624 int max_vf
= MAX_VECTORIZATION_FACTOR
;
1627 /* Find all data references in the loop (which correspond to vdefs/vuses)
1628 and analyze their evolution in the loop. Also adjust the minimal
1629 vectorization factor according to the loads and stores.
1631 FORNOW: Handle only simple, array references, which
1632 alignment can be forced, and aligned pointer-references. */
1634 ok
= vect_analyze_data_refs (loop_vinfo
, NULL
, &min_vf
);
1637 if (dump_enabled_p ())
1638 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
1639 "bad data references.\n");
1643 /* Analyze the access patterns of the data-refs in the loop (consecutive,
1644 complex, etc.). FORNOW: Only handle consecutive access pattern. */
1646 ok
= vect_analyze_data_ref_accesses (loop_vinfo
, NULL
);
1649 if (dump_enabled_p ())
1650 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
1651 "bad data access.\n");
1655 /* Classify all cross-iteration scalar data-flow cycles.
1656 Cross-iteration cycles caused by virtual phis are analyzed separately. */
1658 vect_analyze_scalar_cycles (loop_vinfo
);
1660 vect_pattern_recog (loop_vinfo
, NULL
);
1662 /* Data-flow analysis to detect stmts that do not need to be vectorized. */
1664 ok
= vect_mark_stmts_to_be_vectorized (loop_vinfo
);
1667 if (dump_enabled_p ())
1668 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
1669 "unexpected pattern.\n");
1673 /* Analyze data dependences between the data-refs in the loop
1674 and adjust the maximum vectorization factor according to
1676 FORNOW: fail at the first data dependence that we encounter. */
1678 ok
= vect_analyze_data_ref_dependences (loop_vinfo
, &max_vf
);
1682 if (dump_enabled_p ())
1683 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
1684 "bad data dependence.\n");
1688 ok
= vect_determine_vectorization_factor (loop_vinfo
);
1691 if (dump_enabled_p ())
1692 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
1693 "can't determine vectorization factor.\n");
1696 if (max_vf
< LOOP_VINFO_VECT_FACTOR (loop_vinfo
))
1698 if (dump_enabled_p ())
1699 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
1700 "bad data dependence.\n");
1704 /* Analyze the alignment of the data-refs in the loop.
1705 Fail if a data reference is found that cannot be vectorized. */
1707 ok
= vect_analyze_data_refs_alignment (loop_vinfo
, NULL
);
1710 if (dump_enabled_p ())
1711 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
1712 "bad data alignment.\n");
1716 /* Prune the list of ddrs to be tested at run-time by versioning for alias.
1717 It is important to call pruning after vect_analyze_data_ref_accesses,
1718 since we use grouping information gathered by interleaving analysis. */
1719 ok
= vect_prune_runtime_alias_test_list (loop_vinfo
);
1722 if (dump_enabled_p ())
1723 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
1724 "too long list of versioning for alias "
1725 "run-time tests.\n");
1729 /* This pass will decide on using loop versioning and/or loop peeling in
1730 order to enhance the alignment of data references in the loop. */
1732 ok
= vect_enhance_data_refs_alignment (loop_vinfo
);
1735 if (dump_enabled_p ())
1736 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
1737 "bad data alignment.\n");
1741 /* Check the SLP opportunities in the loop, analyze and build SLP trees. */
1742 ok
= vect_analyze_slp (loop_vinfo
, NULL
);
1745 /* Decide which possible SLP instances to SLP. */
1746 slp
= vect_make_slp_decision (loop_vinfo
);
1748 /* Find stmts that need to be both vectorized and SLPed. */
1749 vect_detect_hybrid_slp (loop_vinfo
);
1754 /* Scan all the operations in the loop and make sure they are
1757 ok
= vect_analyze_loop_operations (loop_vinfo
, slp
);
1760 if (dump_enabled_p ())
1761 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
1762 "bad operation or unsupported loop bound.\n");
1769 /* Function vect_analyze_loop.
1771 Apply a set of analyses on LOOP, and create a loop_vec_info struct
1772 for it. The different analyses will record information in the
1773 loop_vec_info struct. */
1775 vect_analyze_loop (struct loop
*loop
)
1777 loop_vec_info loop_vinfo
;
1778 unsigned int vector_sizes
;
1780 /* Autodetect first vector size we try. */
1781 current_vector_size
= 0;
1782 vector_sizes
= targetm
.vectorize
.autovectorize_vector_sizes ();
1784 if (dump_enabled_p ())
1785 dump_printf_loc (MSG_NOTE
, vect_location
,
1786 "===== analyze_loop_nest =====\n");
1788 if (loop_outer (loop
)
1789 && loop_vec_info_for_loop (loop_outer (loop
))
1790 && LOOP_VINFO_VECTORIZABLE_P (loop_vec_info_for_loop (loop_outer (loop
))))
1792 if (dump_enabled_p ())
1793 dump_printf_loc (MSG_NOTE
, vect_location
,
1794 "outer-loop already vectorized.\n");
1800 /* Check the CFG characteristics of the loop (nesting, entry/exit). */
1801 loop_vinfo
= vect_analyze_loop_form (loop
);
1804 if (dump_enabled_p ())
1805 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
1806 "bad loop form.\n");
1810 if (vect_analyze_loop_2 (loop_vinfo
))
1812 LOOP_VINFO_VECTORIZABLE_P (loop_vinfo
) = 1;
1817 destroy_loop_vec_info (loop_vinfo
, true);
1819 vector_sizes
&= ~current_vector_size
;
1820 if (vector_sizes
== 0
1821 || current_vector_size
== 0)
1824 /* Try the next biggest vector size. */
1825 current_vector_size
= 1 << floor_log2 (vector_sizes
);
1826 if (dump_enabled_p ())
1827 dump_printf_loc (MSG_NOTE
, vect_location
,
1828 "***** Re-trying analysis with "
1829 "vector size %d\n", current_vector_size
);
1834 /* Function reduction_code_for_scalar_code
1837 CODE - tree_code of a reduction operations.
1840 REDUC_CODE - the corresponding tree-code to be used to reduce the
1841 vector of partial results into a single scalar result (which
1842 will also reside in a vector) or ERROR_MARK if the operation is
1843 a supported reduction operation, but does not have such tree-code.
1845 Return FALSE if CODE currently cannot be vectorized as reduction. */
1848 reduction_code_for_scalar_code (enum tree_code code
,
1849 enum tree_code
*reduc_code
)
1854 *reduc_code
= REDUC_MAX_EXPR
;
1858 *reduc_code
= REDUC_MIN_EXPR
;
1862 *reduc_code
= REDUC_PLUS_EXPR
;
1870 *reduc_code
= ERROR_MARK
;
1879 /* Error reporting helper for vect_is_simple_reduction below. GIMPLE statement
1880 STMT is printed with a message MSG. */
1883 report_vect_op (int msg_type
, gimple stmt
, const char *msg
)
1885 dump_printf_loc (msg_type
, vect_location
, "%s", msg
);
1886 dump_gimple_stmt (msg_type
, TDF_SLIM
, stmt
, 0);
1887 dump_printf (msg_type
, "\n");
1891 /* Detect SLP reduction of the form:
1901 PHI is the reduction phi node (#a1 = phi <a5, a0> above)
1902 FIRST_STMT is the first reduction stmt in the chain
1903 (a2 = operation (a1)).
1905 Return TRUE if a reduction chain was detected. */
1908 vect_is_slp_reduction (loop_vec_info loop_info
, gimple phi
, gimple first_stmt
)
1910 struct loop
*loop
= (gimple_bb (phi
))->loop_father
;
1911 struct loop
*vect_loop
= LOOP_VINFO_LOOP (loop_info
);
1912 enum tree_code code
;
1913 gimple current_stmt
= NULL
, loop_use_stmt
= NULL
, first
, next_stmt
;
1914 stmt_vec_info use_stmt_info
, current_stmt_info
;
1916 imm_use_iterator imm_iter
;
1917 use_operand_p use_p
;
1918 int nloop_uses
, size
= 0, n_out_of_loop_uses
;
1921 if (loop
!= vect_loop
)
1924 lhs
= PHI_RESULT (phi
);
1925 code
= gimple_assign_rhs_code (first_stmt
);
1929 n_out_of_loop_uses
= 0;
1930 FOR_EACH_IMM_USE_FAST (use_p
, imm_iter
, lhs
)
1932 gimple use_stmt
= USE_STMT (use_p
);
1933 if (is_gimple_debug (use_stmt
))
1936 use_stmt
= USE_STMT (use_p
);
1938 /* Check if we got back to the reduction phi. */
1939 if (use_stmt
== phi
)
1941 loop_use_stmt
= use_stmt
;
1946 if (flow_bb_inside_loop_p (loop
, gimple_bb (use_stmt
)))
1948 if (vinfo_for_stmt (use_stmt
)
1949 && !STMT_VINFO_IN_PATTERN_P (vinfo_for_stmt (use_stmt
)))
1951 loop_use_stmt
= use_stmt
;
1956 n_out_of_loop_uses
++;
1958 /* There are can be either a single use in the loop or two uses in
1960 if (nloop_uses
> 1 || (n_out_of_loop_uses
&& nloop_uses
))
1967 /* We reached a statement with no loop uses. */
1968 if (nloop_uses
== 0)
1971 /* This is a loop exit phi, and we haven't reached the reduction phi. */
1972 if (gimple_code (loop_use_stmt
) == GIMPLE_PHI
)
1975 if (!is_gimple_assign (loop_use_stmt
)
1976 || code
!= gimple_assign_rhs_code (loop_use_stmt
)
1977 || !flow_bb_inside_loop_p (loop
, gimple_bb (loop_use_stmt
)))
1980 /* Insert USE_STMT into reduction chain. */
1981 use_stmt_info
= vinfo_for_stmt (loop_use_stmt
);
1984 current_stmt_info
= vinfo_for_stmt (current_stmt
);
1985 GROUP_NEXT_ELEMENT (current_stmt_info
) = loop_use_stmt
;
1986 GROUP_FIRST_ELEMENT (use_stmt_info
)
1987 = GROUP_FIRST_ELEMENT (current_stmt_info
);
1990 GROUP_FIRST_ELEMENT (use_stmt_info
) = loop_use_stmt
;
1992 lhs
= gimple_assign_lhs (loop_use_stmt
);
1993 current_stmt
= loop_use_stmt
;
1997 if (!found
|| loop_use_stmt
!= phi
|| size
< 2)
2000 /* Swap the operands, if needed, to make the reduction operand be the second
2002 lhs
= PHI_RESULT (phi
);
2003 next_stmt
= GROUP_FIRST_ELEMENT (vinfo_for_stmt (current_stmt
));
2006 if (gimple_assign_rhs2 (next_stmt
) == lhs
)
2008 tree op
= gimple_assign_rhs1 (next_stmt
);
2009 gimple def_stmt
= NULL
;
2011 if (TREE_CODE (op
) == SSA_NAME
)
2012 def_stmt
= SSA_NAME_DEF_STMT (op
);
2014 /* Check that the other def is either defined in the loop
2015 ("vect_internal_def"), or it's an induction (defined by a
2016 loop-header phi-node). */
2018 && gimple_bb (def_stmt
)
2019 && flow_bb_inside_loop_p (loop
, gimple_bb (def_stmt
))
2020 && (is_gimple_assign (def_stmt
)
2021 || is_gimple_call (def_stmt
)
2022 || STMT_VINFO_DEF_TYPE (vinfo_for_stmt (def_stmt
))
2023 == vect_induction_def
2024 || (gimple_code (def_stmt
) == GIMPLE_PHI
2025 && STMT_VINFO_DEF_TYPE (vinfo_for_stmt (def_stmt
))
2026 == vect_internal_def
2027 && !is_loop_header_bb_p (gimple_bb (def_stmt
)))))
2029 lhs
= gimple_assign_lhs (next_stmt
);
2030 next_stmt
= GROUP_NEXT_ELEMENT (vinfo_for_stmt (next_stmt
));
2038 tree op
= gimple_assign_rhs2 (next_stmt
);
2039 gimple def_stmt
= NULL
;
2041 if (TREE_CODE (op
) == SSA_NAME
)
2042 def_stmt
= SSA_NAME_DEF_STMT (op
);
2044 /* Check that the other def is either defined in the loop
2045 ("vect_internal_def"), or it's an induction (defined by a
2046 loop-header phi-node). */
2048 && gimple_bb (def_stmt
)
2049 && flow_bb_inside_loop_p (loop
, gimple_bb (def_stmt
))
2050 && (is_gimple_assign (def_stmt
)
2051 || is_gimple_call (def_stmt
)
2052 || STMT_VINFO_DEF_TYPE (vinfo_for_stmt (def_stmt
))
2053 == vect_induction_def
2054 || (gimple_code (def_stmt
) == GIMPLE_PHI
2055 && STMT_VINFO_DEF_TYPE (vinfo_for_stmt (def_stmt
))
2056 == vect_internal_def
2057 && !is_loop_header_bb_p (gimple_bb (def_stmt
)))))
2059 if (dump_enabled_p ())
2061 dump_printf_loc (MSG_NOTE
, vect_location
, "swapping oprnds: ");
2062 dump_gimple_stmt (MSG_NOTE
, TDF_SLIM
, next_stmt
, 0);
2063 dump_printf (MSG_NOTE
, "\n");
2066 swap_ssa_operands (next_stmt
,
2067 gimple_assign_rhs1_ptr (next_stmt
),
2068 gimple_assign_rhs2_ptr (next_stmt
));
2069 update_stmt (next_stmt
);
2071 if (CONSTANT_CLASS_P (gimple_assign_rhs1 (next_stmt
)))
2072 LOOP_VINFO_OPERANDS_SWAPPED (loop_info
) = true;
2078 lhs
= gimple_assign_lhs (next_stmt
);
2079 next_stmt
= GROUP_NEXT_ELEMENT (vinfo_for_stmt (next_stmt
));
2082 /* Save the chain for further analysis in SLP detection. */
2083 first
= GROUP_FIRST_ELEMENT (vinfo_for_stmt (current_stmt
));
2084 LOOP_VINFO_REDUCTION_CHAINS (loop_info
).safe_push (first
);
2085 GROUP_SIZE (vinfo_for_stmt (first
)) = size
;
2091 /* Function vect_is_simple_reduction_1
2093 (1) Detect a cross-iteration def-use cycle that represents a simple
2094 reduction computation. We look for the following pattern:
2099 a2 = operation (a3, a1)
2106 a2 = operation (a3, a1)
2109 1. operation is commutative and associative and it is safe to
2110 change the order of the computation (if CHECK_REDUCTION is true)
2111 2. no uses for a2 in the loop (a2 is used out of the loop)
2112 3. no uses of a1 in the loop besides the reduction operation
2113 4. no uses of a1 outside the loop.
2115 Conditions 1,4 are tested here.
2116 Conditions 2,3 are tested in vect_mark_stmts_to_be_vectorized.
2118 (2) Detect a cross-iteration def-use cycle in nested loops, i.e.,
2119 nested cycles, if CHECK_REDUCTION is false.
2121 (3) Detect cycles of phi nodes in outer-loop vectorization, i.e., double
2125 inner loop (def of a3)
2128 If MODIFY is true it tries also to rework the code in-place to enable
2129 detection of more reduction patterns. For the time being we rewrite
2130 "res -= RHS" into "rhs += -RHS" when it seems worthwhile.
2134 vect_is_simple_reduction_1 (loop_vec_info loop_info
, gimple phi
,
2135 bool check_reduction
, bool *double_reduc
,
2138 struct loop
*loop
= (gimple_bb (phi
))->loop_father
;
2139 struct loop
*vect_loop
= LOOP_VINFO_LOOP (loop_info
);
2140 edge latch_e
= loop_latch_edge (loop
);
2141 tree loop_arg
= PHI_ARG_DEF_FROM_EDGE (phi
, latch_e
);
2142 gimple def_stmt
, def1
= NULL
, def2
= NULL
;
2143 enum tree_code orig_code
, code
;
2144 tree op1
, op2
, op3
= NULL_TREE
, op4
= NULL_TREE
;
2148 imm_use_iterator imm_iter
;
2149 use_operand_p use_p
;
2152 *double_reduc
= false;
2154 /* If CHECK_REDUCTION is true, we assume inner-most loop vectorization,
2155 otherwise, we assume outer loop vectorization. */
2156 gcc_assert ((check_reduction
&& loop
== vect_loop
)
2157 || (!check_reduction
&& flow_loop_nested_p (vect_loop
, loop
)));
2159 name
= PHI_RESULT (phi
);
2161 FOR_EACH_IMM_USE_FAST (use_p
, imm_iter
, name
)
2163 gimple use_stmt
= USE_STMT (use_p
);
2164 if (is_gimple_debug (use_stmt
))
2167 if (!flow_bb_inside_loop_p (loop
, gimple_bb (use_stmt
)))
2169 if (dump_enabled_p ())
2170 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
2171 "intermediate value used outside loop.\n");
2176 if (vinfo_for_stmt (use_stmt
)
2177 && !is_pattern_stmt_p (vinfo_for_stmt (use_stmt
)))
2181 if (dump_enabled_p ())
2182 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
2183 "reduction used in loop.\n");
2188 if (TREE_CODE (loop_arg
) != SSA_NAME
)
2190 if (dump_enabled_p ())
2192 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
2193 "reduction: not ssa_name: ");
2194 dump_generic_expr (MSG_MISSED_OPTIMIZATION
, TDF_SLIM
, loop_arg
);
2195 dump_printf (MSG_MISSED_OPTIMIZATION
, "\n");
2200 def_stmt
= SSA_NAME_DEF_STMT (loop_arg
);
2203 if (dump_enabled_p ())
2204 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
2205 "reduction: no def_stmt.\n");
2209 if (!is_gimple_assign (def_stmt
) && gimple_code (def_stmt
) != GIMPLE_PHI
)
2211 if (dump_enabled_p ())
2213 dump_gimple_stmt (MSG_NOTE
, TDF_SLIM
, def_stmt
, 0);
2214 dump_printf (MSG_NOTE
, "\n");
2219 if (is_gimple_assign (def_stmt
))
2221 name
= gimple_assign_lhs (def_stmt
);
2226 name
= PHI_RESULT (def_stmt
);
2231 FOR_EACH_IMM_USE_FAST (use_p
, imm_iter
, name
)
2233 gimple use_stmt
= USE_STMT (use_p
);
2234 if (is_gimple_debug (use_stmt
))
2236 if (flow_bb_inside_loop_p (loop
, gimple_bb (use_stmt
))
2237 && vinfo_for_stmt (use_stmt
)
2238 && !is_pattern_stmt_p (vinfo_for_stmt (use_stmt
)))
2242 if (dump_enabled_p ())
2243 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
2244 "reduction used in loop.\n");
2249 /* If DEF_STMT is a phi node itself, we expect it to have a single argument
2250 defined in the inner loop. */
2253 op1
= PHI_ARG_DEF (def_stmt
, 0);
2255 if (gimple_phi_num_args (def_stmt
) != 1
2256 || TREE_CODE (op1
) != SSA_NAME
)
2258 if (dump_enabled_p ())
2259 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
2260 "unsupported phi node definition.\n");
2265 def1
= SSA_NAME_DEF_STMT (op1
);
2266 if (flow_bb_inside_loop_p (loop
, gimple_bb (def_stmt
))
2268 && flow_bb_inside_loop_p (loop
->inner
, gimple_bb (def1
))
2269 && is_gimple_assign (def1
))
2271 if (dump_enabled_p ())
2272 report_vect_op (MSG_NOTE
, def_stmt
,
2273 "detected double reduction: ");
2275 *double_reduc
= true;
2282 code
= orig_code
= gimple_assign_rhs_code (def_stmt
);
2284 /* We can handle "res -= x[i]", which is non-associative by
2285 simply rewriting this into "res += -x[i]". Avoid changing
2286 gimple instruction for the first simple tests and only do this
2287 if we're allowed to change code at all. */
2288 if (code
== MINUS_EXPR
2290 && (op1
= gimple_assign_rhs1 (def_stmt
))
2291 && TREE_CODE (op1
) == SSA_NAME
2292 && SSA_NAME_DEF_STMT (op1
) == phi
)
2296 && (!commutative_tree_code (code
) || !associative_tree_code (code
)))
2298 if (dump_enabled_p ())
2299 report_vect_op (MSG_MISSED_OPTIMIZATION
, def_stmt
,
2300 "reduction: not commutative/associative: ");
2304 if (get_gimple_rhs_class (code
) != GIMPLE_BINARY_RHS
)
2306 if (code
!= COND_EXPR
)
2308 if (dump_enabled_p ())
2309 report_vect_op (MSG_MISSED_OPTIMIZATION
, def_stmt
,
2310 "reduction: not binary operation: ");
2315 op3
= gimple_assign_rhs1 (def_stmt
);
2316 if (COMPARISON_CLASS_P (op3
))
2318 op4
= TREE_OPERAND (op3
, 1);
2319 op3
= TREE_OPERAND (op3
, 0);
2322 op1
= gimple_assign_rhs2 (def_stmt
);
2323 op2
= gimple_assign_rhs3 (def_stmt
);
2325 if (TREE_CODE (op1
) != SSA_NAME
&& TREE_CODE (op2
) != SSA_NAME
)
2327 if (dump_enabled_p ())
2328 report_vect_op (MSG_MISSED_OPTIMIZATION
, def_stmt
,
2329 "reduction: uses not ssa_names: ");
2336 op1
= gimple_assign_rhs1 (def_stmt
);
2337 op2
= gimple_assign_rhs2 (def_stmt
);
2339 if (TREE_CODE (op1
) != SSA_NAME
&& TREE_CODE (op2
) != SSA_NAME
)
2341 if (dump_enabled_p ())
2342 report_vect_op (MSG_MISSED_OPTIMIZATION
, def_stmt
,
2343 "reduction: uses not ssa_names: ");
2349 type
= TREE_TYPE (gimple_assign_lhs (def_stmt
));
2350 if ((TREE_CODE (op1
) == SSA_NAME
2351 && !types_compatible_p (type
,TREE_TYPE (op1
)))
2352 || (TREE_CODE (op2
) == SSA_NAME
2353 && !types_compatible_p (type
, TREE_TYPE (op2
)))
2354 || (op3
&& TREE_CODE (op3
) == SSA_NAME
2355 && !types_compatible_p (type
, TREE_TYPE (op3
)))
2356 || (op4
&& TREE_CODE (op4
) == SSA_NAME
2357 && !types_compatible_p (type
, TREE_TYPE (op4
))))
2359 if (dump_enabled_p ())
2361 dump_printf_loc (MSG_NOTE
, vect_location
,
2362 "reduction: multiple types: operation type: ");
2363 dump_generic_expr (MSG_NOTE
, TDF_SLIM
, type
);
2364 dump_printf (MSG_NOTE
, ", operands types: ");
2365 dump_generic_expr (MSG_NOTE
, TDF_SLIM
,
2367 dump_printf (MSG_NOTE
, ",");
2368 dump_generic_expr (MSG_NOTE
, TDF_SLIM
,
2372 dump_printf (MSG_NOTE
, ",");
2373 dump_generic_expr (MSG_NOTE
, TDF_SLIM
,
2379 dump_printf (MSG_NOTE
, ",");
2380 dump_generic_expr (MSG_NOTE
, TDF_SLIM
,
2383 dump_printf (MSG_NOTE
, "\n");
2389 /* Check that it's ok to change the order of the computation.
2390 Generally, when vectorizing a reduction we change the order of the
2391 computation. This may change the behavior of the program in some
2392 cases, so we need to check that this is ok. One exception is when
2393 vectorizing an outer-loop: the inner-loop is executed sequentially,
2394 and therefore vectorizing reductions in the inner-loop during
2395 outer-loop vectorization is safe. */
2397 /* CHECKME: check for !flag_finite_math_only too? */
2398 if (SCALAR_FLOAT_TYPE_P (type
) && !flag_associative_math
2401 /* Changing the order of operations changes the semantics. */
2402 if (dump_enabled_p ())
2403 report_vect_op (MSG_MISSED_OPTIMIZATION
, def_stmt
,
2404 "reduction: unsafe fp math optimization: ");
2407 else if (INTEGRAL_TYPE_P (type
) && TYPE_OVERFLOW_TRAPS (type
)
2410 /* Changing the order of operations changes the semantics. */
2411 if (dump_enabled_p ())
2412 report_vect_op (MSG_MISSED_OPTIMIZATION
, def_stmt
,
2413 "reduction: unsafe int math optimization: ");
2416 else if (SAT_FIXED_POINT_TYPE_P (type
) && check_reduction
)
2418 /* Changing the order of operations changes the semantics. */
2419 if (dump_enabled_p ())
2420 report_vect_op (MSG_MISSED_OPTIMIZATION
, def_stmt
,
2421 "reduction: unsafe fixed-point math optimization: ");
2425 /* If we detected "res -= x[i]" earlier, rewrite it into
2426 "res += -x[i]" now. If this turns out to be useless reassoc
2427 will clean it up again. */
2428 if (orig_code
== MINUS_EXPR
)
2430 tree rhs
= gimple_assign_rhs2 (def_stmt
);
2431 tree negrhs
= make_ssa_name (TREE_TYPE (rhs
), NULL
);
2432 gimple negate_stmt
= gimple_build_assign_with_ops (NEGATE_EXPR
, negrhs
,
2434 gimple_stmt_iterator gsi
= gsi_for_stmt (def_stmt
);
2435 set_vinfo_for_stmt (negate_stmt
, new_stmt_vec_info (negate_stmt
,
2437 gsi_insert_before (&gsi
, negate_stmt
, GSI_NEW_STMT
);
2438 gimple_assign_set_rhs2 (def_stmt
, negrhs
);
2439 gimple_assign_set_rhs_code (def_stmt
, PLUS_EXPR
);
2440 update_stmt (def_stmt
);
2443 /* Reduction is safe. We're dealing with one of the following:
2444 1) integer arithmetic and no trapv
2445 2) floating point arithmetic, and special flags permit this optimization
2446 3) nested cycle (i.e., outer loop vectorization). */
2447 if (TREE_CODE (op1
) == SSA_NAME
)
2448 def1
= SSA_NAME_DEF_STMT (op1
);
2450 if (TREE_CODE (op2
) == SSA_NAME
)
2451 def2
= SSA_NAME_DEF_STMT (op2
);
2453 if (code
!= COND_EXPR
2454 && ((!def1
|| gimple_nop_p (def1
)) && (!def2
|| gimple_nop_p (def2
))))
2456 if (dump_enabled_p ())
2457 report_vect_op (MSG_NOTE
, def_stmt
, "reduction: no defs for operands: ");
2461 /* Check that one def is the reduction def, defined by PHI,
2462 the other def is either defined in the loop ("vect_internal_def"),
2463 or it's an induction (defined by a loop-header phi-node). */
2465 if (def2
&& def2
== phi
2466 && (code
== COND_EXPR
2467 || !def1
|| gimple_nop_p (def1
)
2468 || !flow_bb_inside_loop_p (loop
, gimple_bb (def1
))
2469 || (def1
&& flow_bb_inside_loop_p (loop
, gimple_bb (def1
))
2470 && (is_gimple_assign (def1
)
2471 || is_gimple_call (def1
)
2472 || STMT_VINFO_DEF_TYPE (vinfo_for_stmt (def1
))
2473 == vect_induction_def
2474 || (gimple_code (def1
) == GIMPLE_PHI
2475 && STMT_VINFO_DEF_TYPE (vinfo_for_stmt (def1
))
2476 == vect_internal_def
2477 && !is_loop_header_bb_p (gimple_bb (def1
)))))))
2479 if (dump_enabled_p ())
2480 report_vect_op (MSG_NOTE
, def_stmt
, "detected reduction: ");
2484 if (def1
&& def1
== phi
2485 && (code
== COND_EXPR
2486 || !def2
|| gimple_nop_p (def2
)
2487 || !flow_bb_inside_loop_p (loop
, gimple_bb (def2
))
2488 || (def2
&& flow_bb_inside_loop_p (loop
, gimple_bb (def2
))
2489 && (is_gimple_assign (def2
)
2490 || is_gimple_call (def2
)
2491 || STMT_VINFO_DEF_TYPE (vinfo_for_stmt (def2
))
2492 == vect_induction_def
2493 || (gimple_code (def2
) == GIMPLE_PHI
2494 && STMT_VINFO_DEF_TYPE (vinfo_for_stmt (def2
))
2495 == vect_internal_def
2496 && !is_loop_header_bb_p (gimple_bb (def2
)))))))
2498 if (check_reduction
)
2500 /* Swap operands (just for simplicity - so that the rest of the code
2501 can assume that the reduction variable is always the last (second)
2503 if (dump_enabled_p ())
2504 report_vect_op (MSG_NOTE
, def_stmt
,
2505 "detected reduction: need to swap operands: ");
2507 swap_ssa_operands (def_stmt
, gimple_assign_rhs1_ptr (def_stmt
),
2508 gimple_assign_rhs2_ptr (def_stmt
));
2510 if (CONSTANT_CLASS_P (gimple_assign_rhs1 (def_stmt
)))
2511 LOOP_VINFO_OPERANDS_SWAPPED (loop_info
) = true;
2515 if (dump_enabled_p ())
2516 report_vect_op (MSG_NOTE
, def_stmt
, "detected reduction: ");
2522 /* Try to find SLP reduction chain. */
2523 if (check_reduction
&& vect_is_slp_reduction (loop_info
, phi
, def_stmt
))
2525 if (dump_enabled_p ())
2526 report_vect_op (MSG_NOTE
, def_stmt
,
2527 "reduction: detected reduction chain: ");
2532 if (dump_enabled_p ())
2533 report_vect_op (MSG_MISSED_OPTIMIZATION
, def_stmt
,
2534 "reduction: unknown pattern: ");
2539 /* Wrapper around vect_is_simple_reduction_1, that won't modify code
2540 in-place. Arguments as there. */
2543 vect_is_simple_reduction (loop_vec_info loop_info
, gimple phi
,
2544 bool check_reduction
, bool *double_reduc
)
2546 return vect_is_simple_reduction_1 (loop_info
, phi
, check_reduction
,
2547 double_reduc
, false);
2550 /* Wrapper around vect_is_simple_reduction_1, which will modify code
2551 in-place if it enables detection of more reductions. Arguments
2555 vect_force_simple_reduction (loop_vec_info loop_info
, gimple phi
,
2556 bool check_reduction
, bool *double_reduc
)
2558 return vect_is_simple_reduction_1 (loop_info
, phi
, check_reduction
,
2559 double_reduc
, true);
2562 /* Calculate the cost of one scalar iteration of the loop. */
2564 vect_get_single_scalar_iteration_cost (loop_vec_info loop_vinfo
)
2566 struct loop
*loop
= LOOP_VINFO_LOOP (loop_vinfo
);
2567 basic_block
*bbs
= LOOP_VINFO_BBS (loop_vinfo
);
2568 int nbbs
= loop
->num_nodes
, factor
, scalar_single_iter_cost
= 0;
2569 int innerloop_iters
, i
, stmt_cost
;
2571 /* Count statements in scalar loop. Using this as scalar cost for a single
2574 TODO: Add outer loop support.
2576 TODO: Consider assigning different costs to different scalar
2580 innerloop_iters
= 1;
2582 innerloop_iters
= 50; /* FIXME */
2584 for (i
= 0; i
< nbbs
; i
++)
2586 gimple_stmt_iterator si
;
2587 basic_block bb
= bbs
[i
];
2589 if (bb
->loop_father
== loop
->inner
)
2590 factor
= innerloop_iters
;
2594 for (si
= gsi_start_bb (bb
); !gsi_end_p (si
); gsi_next (&si
))
2596 gimple stmt
= gsi_stmt (si
);
2597 stmt_vec_info stmt_info
= vinfo_for_stmt (stmt
);
2599 if (!is_gimple_assign (stmt
) && !is_gimple_call (stmt
))
2602 /* Skip stmts that are not vectorized inside the loop. */
2604 && !STMT_VINFO_RELEVANT_P (stmt_info
)
2605 && (!STMT_VINFO_LIVE_P (stmt_info
)
2606 || !VECTORIZABLE_CYCLE_DEF (STMT_VINFO_DEF_TYPE (stmt_info
)))
2607 && !STMT_VINFO_IN_PATTERN_P (stmt_info
))
2610 if (STMT_VINFO_DATA_REF (vinfo_for_stmt (stmt
)))
2612 if (DR_IS_READ (STMT_VINFO_DATA_REF (vinfo_for_stmt (stmt
))))
2613 stmt_cost
= vect_get_stmt_cost (scalar_load
);
2615 stmt_cost
= vect_get_stmt_cost (scalar_store
);
2618 stmt_cost
= vect_get_stmt_cost (scalar_stmt
);
2620 scalar_single_iter_cost
+= stmt_cost
* factor
;
2623 return scalar_single_iter_cost
;
2626 /* Calculate cost of peeling the loop PEEL_ITERS_PROLOGUE times. */
2628 vect_get_known_peeling_cost (loop_vec_info loop_vinfo
, int peel_iters_prologue
,
2629 int *peel_iters_epilogue
,
2630 int scalar_single_iter_cost
,
2631 stmt_vector_for_cost
*prologue_cost_vec
,
2632 stmt_vector_for_cost
*epilogue_cost_vec
)
2635 int vf
= LOOP_VINFO_VECT_FACTOR (loop_vinfo
);
2637 if (!LOOP_VINFO_NITERS_KNOWN_P (loop_vinfo
))
2639 *peel_iters_epilogue
= vf
/2;
2640 if (dump_enabled_p ())
2641 dump_printf_loc (MSG_NOTE
, vect_location
,
2642 "cost model: epilogue peel iters set to vf/2 "
2643 "because loop iterations are unknown .\n");
2645 /* If peeled iterations are known but number of scalar loop
2646 iterations are unknown, count a taken branch per peeled loop. */
2647 retval
= record_stmt_cost (prologue_cost_vec
, 2, cond_branch_taken
,
2648 NULL
, 0, vect_prologue
);
2652 int niters
= LOOP_VINFO_INT_NITERS (loop_vinfo
);
2653 peel_iters_prologue
= niters
< peel_iters_prologue
?
2654 niters
: peel_iters_prologue
;
2655 *peel_iters_epilogue
= (niters
- peel_iters_prologue
) % vf
;
2656 /* If we need to peel for gaps, but no peeling is required, we have to
2657 peel VF iterations. */
2658 if (LOOP_VINFO_PEELING_FOR_GAPS (loop_vinfo
) && !*peel_iters_epilogue
)
2659 *peel_iters_epilogue
= vf
;
2662 if (peel_iters_prologue
)
2663 retval
+= record_stmt_cost (prologue_cost_vec
,
2664 peel_iters_prologue
* scalar_single_iter_cost
,
2665 scalar_stmt
, NULL
, 0, vect_prologue
);
2666 if (*peel_iters_epilogue
)
2667 retval
+= record_stmt_cost (epilogue_cost_vec
,
2668 *peel_iters_epilogue
* scalar_single_iter_cost
,
2669 scalar_stmt
, NULL
, 0, vect_epilogue
);
2673 /* Function vect_estimate_min_profitable_iters
2675 Return the number of iterations required for the vector version of the
2676 loop to be profitable relative to the cost of the scalar version of the
2680 vect_estimate_min_profitable_iters (loop_vec_info loop_vinfo
,
2681 int *ret_min_profitable_niters
,
2682 int *ret_min_profitable_estimate
)
2684 int min_profitable_iters
;
2685 int min_profitable_estimate
;
2686 int peel_iters_prologue
;
2687 int peel_iters_epilogue
;
2688 unsigned vec_inside_cost
= 0;
2689 int vec_outside_cost
= 0;
2690 unsigned vec_prologue_cost
= 0;
2691 unsigned vec_epilogue_cost
= 0;
2692 int scalar_single_iter_cost
= 0;
2693 int scalar_outside_cost
= 0;
2694 int vf
= LOOP_VINFO_VECT_FACTOR (loop_vinfo
);
2695 int npeel
= LOOP_PEELING_FOR_ALIGNMENT (loop_vinfo
);
2696 void *target_cost_data
= LOOP_VINFO_TARGET_COST_DATA (loop_vinfo
);
2698 /* Cost model disabled. */
2699 if (unlimited_cost_model ())
2701 dump_printf_loc (MSG_NOTE
, vect_location
, "cost model disabled.\n");
2702 *ret_min_profitable_niters
= 0;
2703 *ret_min_profitable_estimate
= 0;
2707 /* Requires loop versioning tests to handle misalignment. */
2708 if (LOOP_REQUIRES_VERSIONING_FOR_ALIGNMENT (loop_vinfo
))
2710 /* FIXME: Make cost depend on complexity of individual check. */
2711 unsigned len
= LOOP_VINFO_MAY_MISALIGN_STMTS (loop_vinfo
).length ();
2712 (void) add_stmt_cost (target_cost_data
, len
, vector_stmt
, NULL
, 0,
2714 dump_printf (MSG_NOTE
,
2715 "cost model: Adding cost of checks for loop "
2716 "versioning to treat misalignment.\n");
2719 /* Requires loop versioning with alias checks. */
2720 if (LOOP_REQUIRES_VERSIONING_FOR_ALIAS (loop_vinfo
))
2722 /* FIXME: Make cost depend on complexity of individual check. */
2723 unsigned len
= LOOP_VINFO_MAY_ALIAS_DDRS (loop_vinfo
).length ();
2724 (void) add_stmt_cost (target_cost_data
, len
, vector_stmt
, NULL
, 0,
2726 dump_printf (MSG_NOTE
,
2727 "cost model: Adding cost of checks for loop "
2728 "versioning aliasing.\n");
2731 if (LOOP_REQUIRES_VERSIONING_FOR_ALIGNMENT (loop_vinfo
)
2732 || LOOP_REQUIRES_VERSIONING_FOR_ALIAS (loop_vinfo
))
2733 (void) add_stmt_cost (target_cost_data
, 1, cond_branch_taken
, NULL
, 0,
2736 /* Count statements in scalar loop. Using this as scalar cost for a single
2739 TODO: Add outer loop support.
2741 TODO: Consider assigning different costs to different scalar
2744 scalar_single_iter_cost
= vect_get_single_scalar_iteration_cost (loop_vinfo
);
2746 /* Add additional cost for the peeled instructions in prologue and epilogue
2749 FORNOW: If we don't know the value of peel_iters for prologue or epilogue
2750 at compile-time - we assume it's vf/2 (the worst would be vf-1).
2752 TODO: Build an expression that represents peel_iters for prologue and
2753 epilogue to be used in a run-time test. */
2757 peel_iters_prologue
= vf
/2;
2758 dump_printf (MSG_NOTE
, "cost model: "
2759 "prologue peel iters set to vf/2.\n");
2761 /* If peeling for alignment is unknown, loop bound of main loop becomes
2763 peel_iters_epilogue
= vf
/2;
2764 dump_printf (MSG_NOTE
, "cost model: "
2765 "epilogue peel iters set to vf/2 because "
2766 "peeling for alignment is unknown.\n");
2768 /* If peeled iterations are unknown, count a taken branch and a not taken
2769 branch per peeled loop. Even if scalar loop iterations are known,
2770 vector iterations are not known since peeled prologue iterations are
2771 not known. Hence guards remain the same. */
2772 (void) add_stmt_cost (target_cost_data
, 2, cond_branch_taken
,
2773 NULL
, 0, vect_prologue
);
2774 (void) add_stmt_cost (target_cost_data
, 2, cond_branch_not_taken
,
2775 NULL
, 0, vect_prologue
);
2776 /* FORNOW: Don't attempt to pass individual scalar instructions to
2777 the model; just assume linear cost for scalar iterations. */
2778 (void) add_stmt_cost (target_cost_data
,
2779 peel_iters_prologue
* scalar_single_iter_cost
,
2780 scalar_stmt
, NULL
, 0, vect_prologue
);
2781 (void) add_stmt_cost (target_cost_data
,
2782 peel_iters_epilogue
* scalar_single_iter_cost
,
2783 scalar_stmt
, NULL
, 0, vect_epilogue
);
2787 stmt_vector_for_cost prologue_cost_vec
, epilogue_cost_vec
;
2788 stmt_info_for_cost
*si
;
2790 void *data
= LOOP_VINFO_TARGET_COST_DATA (loop_vinfo
);
2792 prologue_cost_vec
.create (2);
2793 epilogue_cost_vec
.create (2);
2794 peel_iters_prologue
= npeel
;
2796 (void) vect_get_known_peeling_cost (loop_vinfo
, peel_iters_prologue
,
2797 &peel_iters_epilogue
,
2798 scalar_single_iter_cost
,
2800 &epilogue_cost_vec
);
2802 FOR_EACH_VEC_ELT (prologue_cost_vec
, j
, si
)
2804 struct _stmt_vec_info
*stmt_info
2805 = si
->stmt
? vinfo_for_stmt (si
->stmt
) : NULL
;
2806 (void) add_stmt_cost (data
, si
->count
, si
->kind
, stmt_info
,
2807 si
->misalign
, vect_prologue
);
2810 FOR_EACH_VEC_ELT (epilogue_cost_vec
, j
, si
)
2812 struct _stmt_vec_info
*stmt_info
2813 = si
->stmt
? vinfo_for_stmt (si
->stmt
) : NULL
;
2814 (void) add_stmt_cost (data
, si
->count
, si
->kind
, stmt_info
,
2815 si
->misalign
, vect_epilogue
);
2818 prologue_cost_vec
.release ();
2819 epilogue_cost_vec
.release ();
2822 /* FORNOW: The scalar outside cost is incremented in one of the
2825 1. The vectorizer checks for alignment and aliasing and generates
2826 a condition that allows dynamic vectorization. A cost model
2827 check is ANDED with the versioning condition. Hence scalar code
2828 path now has the added cost of the versioning check.
2830 if (cost > th & versioning_check)
2833 Hence run-time scalar is incremented by not-taken branch cost.
2835 2. The vectorizer then checks if a prologue is required. If the
2836 cost model check was not done before during versioning, it has to
2837 be done before the prologue check.
2840 prologue = scalar_iters
2845 if (prologue == num_iters)
2848 Hence the run-time scalar cost is incremented by a taken branch,
2849 plus a not-taken branch, plus a taken branch cost.
2851 3. The vectorizer then checks if an epilogue is required. If the
2852 cost model check was not done before during prologue check, it
2853 has to be done with the epilogue check.
2859 if (prologue == num_iters)
2862 if ((cost <= th) | (scalar_iters-prologue-epilogue == 0))
2865 Hence the run-time scalar cost should be incremented by 2 taken
2868 TODO: The back end may reorder the BBS's differently and reverse
2869 conditions/branch directions. Change the estimates below to
2870 something more reasonable. */
2872 /* If the number of iterations is known and we do not do versioning, we can
2873 decide whether to vectorize at compile time. Hence the scalar version
2874 do not carry cost model guard costs. */
2875 if (!LOOP_VINFO_NITERS_KNOWN_P (loop_vinfo
)
2876 || LOOP_REQUIRES_VERSIONING_FOR_ALIGNMENT (loop_vinfo
)
2877 || LOOP_REQUIRES_VERSIONING_FOR_ALIAS (loop_vinfo
))
2879 /* Cost model check occurs at versioning. */
2880 if (LOOP_REQUIRES_VERSIONING_FOR_ALIGNMENT (loop_vinfo
)
2881 || LOOP_REQUIRES_VERSIONING_FOR_ALIAS (loop_vinfo
))
2882 scalar_outside_cost
+= vect_get_stmt_cost (cond_branch_not_taken
);
2885 /* Cost model check occurs at prologue generation. */
2886 if (LOOP_PEELING_FOR_ALIGNMENT (loop_vinfo
) < 0)
2887 scalar_outside_cost
+= 2 * vect_get_stmt_cost (cond_branch_taken
)
2888 + vect_get_stmt_cost (cond_branch_not_taken
);
2889 /* Cost model check occurs at epilogue generation. */
2891 scalar_outside_cost
+= 2 * vect_get_stmt_cost (cond_branch_taken
);
2895 /* Complete the target-specific cost calculations. */
2896 finish_cost (LOOP_VINFO_TARGET_COST_DATA (loop_vinfo
), &vec_prologue_cost
,
2897 &vec_inside_cost
, &vec_epilogue_cost
);
2899 vec_outside_cost
= (int)(vec_prologue_cost
+ vec_epilogue_cost
);
2901 /* Calculate number of iterations required to make the vector version
2902 profitable, relative to the loop bodies only. The following condition
2904 SIC * niters + SOC > VIC * ((niters-PL_ITERS-EP_ITERS)/VF) + VOC
2906 SIC = scalar iteration cost, VIC = vector iteration cost,
2907 VOC = vector outside cost, VF = vectorization factor,
2908 PL_ITERS = prologue iterations, EP_ITERS= epilogue iterations
2909 SOC = scalar outside cost for run time cost model check. */
2911 if ((scalar_single_iter_cost
* vf
) > (int) vec_inside_cost
)
2913 if (vec_outside_cost
<= 0)
2914 min_profitable_iters
= 1;
2917 min_profitable_iters
= ((vec_outside_cost
- scalar_outside_cost
) * vf
2918 - vec_inside_cost
* peel_iters_prologue
2919 - vec_inside_cost
* peel_iters_epilogue
)
2920 / ((scalar_single_iter_cost
* vf
)
2923 if ((scalar_single_iter_cost
* vf
* min_profitable_iters
)
2924 <= (((int) vec_inside_cost
* min_profitable_iters
)
2925 + (((int) vec_outside_cost
- scalar_outside_cost
) * vf
)))
2926 min_profitable_iters
++;
2929 /* vector version will never be profitable. */
2932 if (dump_enabled_p ())
2933 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
2934 "cost model: the vector iteration cost = %d "
2935 "divided by the scalar iteration cost = %d "
2936 "is greater or equal to the vectorization factor = %d"
2938 vec_inside_cost
, scalar_single_iter_cost
, vf
);
2939 *ret_min_profitable_niters
= -1;
2940 *ret_min_profitable_estimate
= -1;
2944 if (dump_enabled_p ())
2946 dump_printf_loc (MSG_NOTE
, vect_location
, "Cost model analysis: \n");
2947 dump_printf (MSG_NOTE
, " Vector inside of loop cost: %d\n",
2949 dump_printf (MSG_NOTE
, " Vector prologue cost: %d\n",
2951 dump_printf (MSG_NOTE
, " Vector epilogue cost: %d\n",
2953 dump_printf (MSG_NOTE
, " Scalar iteration cost: %d\n",
2954 scalar_single_iter_cost
);
2955 dump_printf (MSG_NOTE
, " Scalar outside cost: %d\n",
2956 scalar_outside_cost
);
2957 dump_printf (MSG_NOTE
, " Vector outside cost: %d\n",
2959 dump_printf (MSG_NOTE
, " prologue iterations: %d\n",
2960 peel_iters_prologue
);
2961 dump_printf (MSG_NOTE
, " epilogue iterations: %d\n",
2962 peel_iters_epilogue
);
2963 dump_printf (MSG_NOTE
,
2964 " Calculated minimum iters for profitability: %d\n",
2965 min_profitable_iters
);
2966 dump_printf (MSG_NOTE
, "\n");
2969 min_profitable_iters
=
2970 min_profitable_iters
< vf
? vf
: min_profitable_iters
;
2972 /* Because the condition we create is:
2973 if (niters <= min_profitable_iters)
2974 then skip the vectorized loop. */
2975 min_profitable_iters
--;
2977 if (dump_enabled_p ())
2978 dump_printf_loc (MSG_NOTE
, vect_location
,
2979 " Runtime profitability threshold = %d\n",
2980 min_profitable_iters
);
2982 *ret_min_profitable_niters
= min_profitable_iters
;
2984 /* Calculate number of iterations required to make the vector version
2985 profitable, relative to the loop bodies only.
2987 Non-vectorized variant is SIC * niters and it must win over vector
2988 variant on the expected loop trip count. The following condition must hold true:
2989 SIC * niters > VIC * ((niters-PL_ITERS-EP_ITERS)/VF) + VOC + SOC */
2991 if (vec_outside_cost
<= 0)
2992 min_profitable_estimate
= 1;
2995 min_profitable_estimate
= ((vec_outside_cost
+ scalar_outside_cost
) * vf
2996 - vec_inside_cost
* peel_iters_prologue
2997 - vec_inside_cost
* peel_iters_epilogue
)
2998 / ((scalar_single_iter_cost
* vf
)
3001 min_profitable_estimate
--;
3002 min_profitable_estimate
= MAX (min_profitable_estimate
, min_profitable_iters
);
3003 if (dump_enabled_p ())
3004 dump_printf_loc (MSG_NOTE
, vect_location
,
3005 " Static estimate profitability threshold = %d\n",
3006 min_profitable_iters
);
3008 *ret_min_profitable_estimate
= min_profitable_estimate
;
3012 /* TODO: Close dependency between vect_model_*_cost and vectorizable_*
3013 functions. Design better to avoid maintenance issues. */
3015 /* Function vect_model_reduction_cost.
3017 Models cost for a reduction operation, including the vector ops
3018 generated within the strip-mine loop, the initial definition before
3019 the loop, and the epilogue code that must be generated. */
3022 vect_model_reduction_cost (stmt_vec_info stmt_info
, enum tree_code reduc_code
,
3025 int prologue_cost
= 0, epilogue_cost
= 0;
3026 enum tree_code code
;
3029 gimple stmt
, orig_stmt
;
3031 enum machine_mode mode
;
3032 loop_vec_info loop_vinfo
= STMT_VINFO_LOOP_VINFO (stmt_info
);
3033 struct loop
*loop
= LOOP_VINFO_LOOP (loop_vinfo
);
3034 void *target_cost_data
= LOOP_VINFO_TARGET_COST_DATA (loop_vinfo
);
3036 /* Cost of reduction op inside loop. */
3037 unsigned inside_cost
= add_stmt_cost (target_cost_data
, ncopies
, vector_stmt
,
3038 stmt_info
, 0, vect_body
);
3039 stmt
= STMT_VINFO_STMT (stmt_info
);
3041 switch (get_gimple_rhs_class (gimple_assign_rhs_code (stmt
)))
3043 case GIMPLE_SINGLE_RHS
:
3044 gcc_assert (TREE_OPERAND_LENGTH (gimple_assign_rhs1 (stmt
)) == ternary_op
);
3045 reduction_op
= TREE_OPERAND (gimple_assign_rhs1 (stmt
), 2);
3047 case GIMPLE_UNARY_RHS
:
3048 reduction_op
= gimple_assign_rhs1 (stmt
);
3050 case GIMPLE_BINARY_RHS
:
3051 reduction_op
= gimple_assign_rhs2 (stmt
);
3053 case GIMPLE_TERNARY_RHS
:
3054 reduction_op
= gimple_assign_rhs3 (stmt
);
3060 vectype
= get_vectype_for_scalar_type (TREE_TYPE (reduction_op
));
3063 if (dump_enabled_p ())
3065 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
3066 "unsupported data-type ");
3067 dump_generic_expr (MSG_MISSED_OPTIMIZATION
, TDF_SLIM
,
3068 TREE_TYPE (reduction_op
));
3069 dump_printf (MSG_MISSED_OPTIMIZATION
, "\n");
3074 mode
= TYPE_MODE (vectype
);
3075 orig_stmt
= STMT_VINFO_RELATED_STMT (stmt_info
);
3078 orig_stmt
= STMT_VINFO_STMT (stmt_info
);
3080 code
= gimple_assign_rhs_code (orig_stmt
);
3082 /* Add in cost for initial definition. */
3083 prologue_cost
+= add_stmt_cost (target_cost_data
, 1, scalar_to_vec
,
3084 stmt_info
, 0, vect_prologue
);
3086 /* Determine cost of epilogue code.
3088 We have a reduction operator that will reduce the vector in one statement.
3089 Also requires scalar extract. */
3091 if (!nested_in_vect_loop_p (loop
, orig_stmt
))
3093 if (reduc_code
!= ERROR_MARK
)
3095 epilogue_cost
+= add_stmt_cost (target_cost_data
, 1, vector_stmt
,
3096 stmt_info
, 0, vect_epilogue
);
3097 epilogue_cost
+= add_stmt_cost (target_cost_data
, 1, vec_to_scalar
,
3098 stmt_info
, 0, vect_epilogue
);
3102 int vec_size_in_bits
= tree_low_cst (TYPE_SIZE (vectype
), 1);
3104 TYPE_SIZE (TREE_TYPE (gimple_assign_lhs (orig_stmt
)));
3105 int element_bitsize
= tree_low_cst (bitsize
, 1);
3106 int nelements
= vec_size_in_bits
/ element_bitsize
;
3108 optab
= optab_for_tree_code (code
, vectype
, optab_default
);
3110 /* We have a whole vector shift available. */
3111 if (VECTOR_MODE_P (mode
)
3112 && optab_handler (optab
, mode
) != CODE_FOR_nothing
3113 && optab_handler (vec_shr_optab
, mode
) != CODE_FOR_nothing
)
3115 /* Final reduction via vector shifts and the reduction operator.
3116 Also requires scalar extract. */
3117 epilogue_cost
+= add_stmt_cost (target_cost_data
,
3118 exact_log2 (nelements
) * 2,
3119 vector_stmt
, stmt_info
, 0,
3121 epilogue_cost
+= add_stmt_cost (target_cost_data
, 1,
3122 vec_to_scalar
, stmt_info
, 0,
3126 /* Use extracts and reduction op for final reduction. For N
3127 elements, we have N extracts and N-1 reduction ops. */
3128 epilogue_cost
+= add_stmt_cost (target_cost_data
,
3129 nelements
+ nelements
- 1,
3130 vector_stmt
, stmt_info
, 0,
3135 if (dump_enabled_p ())
3136 dump_printf (MSG_NOTE
,
3137 "vect_model_reduction_cost: inside_cost = %d, "
3138 "prologue_cost = %d, epilogue_cost = %d .\n", inside_cost
,
3139 prologue_cost
, epilogue_cost
);
3145 /* Function vect_model_induction_cost.
3147 Models cost for induction operations. */
3150 vect_model_induction_cost (stmt_vec_info stmt_info
, int ncopies
)
3152 loop_vec_info loop_vinfo
= STMT_VINFO_LOOP_VINFO (stmt_info
);
3153 void *target_cost_data
= LOOP_VINFO_TARGET_COST_DATA (loop_vinfo
);
3154 unsigned inside_cost
, prologue_cost
;
3156 /* loop cost for vec_loop. */
3157 inside_cost
= add_stmt_cost (target_cost_data
, ncopies
, vector_stmt
,
3158 stmt_info
, 0, vect_body
);
3160 /* prologue cost for vec_init and vec_step. */
3161 prologue_cost
= add_stmt_cost (target_cost_data
, 2, scalar_to_vec
,
3162 stmt_info
, 0, vect_prologue
);
3164 if (dump_enabled_p ())
3165 dump_printf_loc (MSG_NOTE
, vect_location
,
3166 "vect_model_induction_cost: inside_cost = %d, "
3167 "prologue_cost = %d .\n", inside_cost
, prologue_cost
);
3171 /* Function get_initial_def_for_induction
3174 STMT - a stmt that performs an induction operation in the loop.
3175 IV_PHI - the initial value of the induction variable
3178 Return a vector variable, initialized with the first VF values of
3179 the induction variable. E.g., for an iv with IV_PHI='X' and
3180 evolution S, for a vector of 4 units, we want to return:
3181 [X, X + S, X + 2*S, X + 3*S]. */
3184 get_initial_def_for_induction (gimple iv_phi
)
3186 stmt_vec_info stmt_vinfo
= vinfo_for_stmt (iv_phi
);
3187 loop_vec_info loop_vinfo
= STMT_VINFO_LOOP_VINFO (stmt_vinfo
);
3188 struct loop
*loop
= LOOP_VINFO_LOOP (loop_vinfo
);
3191 edge pe
= loop_preheader_edge (loop
);
3192 struct loop
*iv_loop
;
3194 tree new_vec
, vec_init
, vec_step
, t
;
3198 gimple init_stmt
, induction_phi
, new_stmt
;
3199 tree induc_def
, vec_def
, vec_dest
;
3200 tree init_expr
, step_expr
;
3201 int vf
= LOOP_VINFO_VECT_FACTOR (loop_vinfo
);
3206 stmt_vec_info phi_info
= vinfo_for_stmt (iv_phi
);
3207 bool nested_in_vect_loop
= false;
3208 gimple_seq stmts
= NULL
;
3209 imm_use_iterator imm_iter
;
3210 use_operand_p use_p
;
3214 gimple_stmt_iterator si
;
3215 basic_block bb
= gimple_bb (iv_phi
);
3219 /* Is phi in an inner-loop, while vectorizing an enclosing outer-loop? */
3220 if (nested_in_vect_loop_p (loop
, iv_phi
))
3222 nested_in_vect_loop
= true;
3223 iv_loop
= loop
->inner
;
3227 gcc_assert (iv_loop
== (gimple_bb (iv_phi
))->loop_father
);
3229 latch_e
= loop_latch_edge (iv_loop
);
3230 loop_arg
= PHI_ARG_DEF_FROM_EDGE (iv_phi
, latch_e
);
3232 access_fn
= analyze_scalar_evolution (iv_loop
, PHI_RESULT (iv_phi
));
3233 gcc_assert (access_fn
);
3234 STRIP_NOPS (access_fn
);
3235 ok
= vect_is_simple_iv_evolution (iv_loop
->num
, access_fn
,
3236 &init_expr
, &step_expr
);
3238 pe
= loop_preheader_edge (iv_loop
);
3240 vectype
= get_vectype_for_scalar_type (TREE_TYPE (init_expr
));
3241 resvectype
= get_vectype_for_scalar_type (TREE_TYPE (PHI_RESULT (iv_phi
)));
3242 gcc_assert (vectype
);
3243 nunits
= TYPE_VECTOR_SUBPARTS (vectype
);
3244 ncopies
= vf
/ nunits
;
3246 gcc_assert (phi_info
);
3247 gcc_assert (ncopies
>= 1);
3249 /* Find the first insertion point in the BB. */
3250 si
= gsi_after_labels (bb
);
3252 /* Create the vector that holds the initial_value of the induction. */
3253 if (nested_in_vect_loop
)
3255 /* iv_loop is nested in the loop to be vectorized. init_expr had already
3256 been created during vectorization of previous stmts. We obtain it
3257 from the STMT_VINFO_VEC_STMT of the defining stmt. */
3258 tree iv_def
= PHI_ARG_DEF_FROM_EDGE (iv_phi
,
3259 loop_preheader_edge (iv_loop
));
3260 vec_init
= vect_get_vec_def_for_operand (iv_def
, iv_phi
, NULL
);
3261 /* If the initial value is not of proper type, convert it. */
3262 if (!useless_type_conversion_p (vectype
, TREE_TYPE (vec_init
)))
3264 new_stmt
= gimple_build_assign_with_ops
3266 vect_get_new_vect_var (vectype
, vect_simple_var
, "vec_iv_"),
3267 build1 (VIEW_CONVERT_EXPR
, vectype
, vec_init
), NULL_TREE
);
3268 vec_init
= make_ssa_name (gimple_assign_lhs (new_stmt
), new_stmt
);
3269 gimple_assign_set_lhs (new_stmt
, vec_init
);
3270 new_bb
= gsi_insert_on_edge_immediate (loop_preheader_edge (iv_loop
),
3272 gcc_assert (!new_bb
);
3273 set_vinfo_for_stmt (new_stmt
,
3274 new_stmt_vec_info (new_stmt
, loop_vinfo
, NULL
));
3279 vec
<constructor_elt
, va_gc
> *v
;
3281 /* iv_loop is the loop to be vectorized. Create:
3282 vec_init = [X, X+S, X+2*S, X+3*S] (S = step_expr, X = init_expr) */
3283 new_var
= vect_get_new_vect_var (TREE_TYPE (vectype
),
3284 vect_scalar_var
, "var_");
3285 new_name
= force_gimple_operand (fold_convert (TREE_TYPE (vectype
),
3287 &stmts
, false, new_var
);
3290 new_bb
= gsi_insert_seq_on_edge_immediate (pe
, stmts
);
3291 gcc_assert (!new_bb
);
3294 vec_alloc (v
, nunits
);
3295 bool constant_p
= is_gimple_min_invariant (new_name
);
3296 CONSTRUCTOR_APPEND_ELT (v
, NULL_TREE
, new_name
);
3297 for (i
= 1; i
< nunits
; i
++)
3299 /* Create: new_name_i = new_name + step_expr */
3300 new_name
= fold_build2 (PLUS_EXPR
, TREE_TYPE (new_name
),
3301 new_name
, step_expr
);
3302 if (!is_gimple_min_invariant (new_name
))
3304 init_stmt
= gimple_build_assign (new_var
, new_name
);
3305 new_name
= make_ssa_name (new_var
, init_stmt
);
3306 gimple_assign_set_lhs (init_stmt
, new_name
);
3307 new_bb
= gsi_insert_on_edge_immediate (pe
, init_stmt
);
3308 gcc_assert (!new_bb
);
3309 if (dump_enabled_p ())
3311 dump_printf_loc (MSG_NOTE
, vect_location
,
3312 "created new init_stmt: ");
3313 dump_gimple_stmt (MSG_NOTE
, TDF_SLIM
, init_stmt
, 0);
3314 dump_printf (MSG_NOTE
, "\n");
3318 CONSTRUCTOR_APPEND_ELT (v
, NULL_TREE
, new_name
);
3320 /* Create a vector from [new_name_0, new_name_1, ..., new_name_nunits-1] */
3322 new_vec
= build_vector_from_ctor (vectype
, v
);
3324 new_vec
= build_constructor (vectype
, v
);
3325 vec_init
= vect_init_vector (iv_phi
, new_vec
, vectype
, NULL
);
3329 /* Create the vector that holds the step of the induction. */
3330 if (nested_in_vect_loop
)
3331 /* iv_loop is nested in the loop to be vectorized. Generate:
3332 vec_step = [S, S, S, S] */
3333 new_name
= step_expr
;
3336 /* iv_loop is the loop to be vectorized. Generate:
3337 vec_step = [VF*S, VF*S, VF*S, VF*S] */
3338 if (SCALAR_FLOAT_TYPE_P (TREE_TYPE (step_expr
)))
3340 expr
= build_int_cst (integer_type_node
, vf
);
3341 expr
= fold_convert (TREE_TYPE (step_expr
), expr
);
3344 expr
= build_int_cst (TREE_TYPE (step_expr
), vf
);
3345 new_name
= fold_build2 (MULT_EXPR
, TREE_TYPE (step_expr
),
3347 if (TREE_CODE (step_expr
) == SSA_NAME
)
3348 new_name
= vect_init_vector (iv_phi
, new_name
,
3349 TREE_TYPE (step_expr
), NULL
);
3352 t
= unshare_expr (new_name
);
3353 gcc_assert (CONSTANT_CLASS_P (new_name
)
3354 || TREE_CODE (new_name
) == SSA_NAME
);
3355 stepvectype
= get_vectype_for_scalar_type (TREE_TYPE (new_name
));
3356 gcc_assert (stepvectype
);
3357 new_vec
= build_vector_from_val (stepvectype
, t
);
3358 vec_step
= vect_init_vector (iv_phi
, new_vec
, stepvectype
, NULL
);
3361 /* Create the following def-use cycle:
3366 vec_iv = PHI <vec_init, vec_loop>
3370 vec_loop = vec_iv + vec_step; */
3372 /* Create the induction-phi that defines the induction-operand. */
3373 vec_dest
= vect_get_new_vect_var (vectype
, vect_simple_var
, "vec_iv_");
3374 induction_phi
= create_phi_node (vec_dest
, iv_loop
->header
);
3375 set_vinfo_for_stmt (induction_phi
,
3376 new_stmt_vec_info (induction_phi
, loop_vinfo
, NULL
));
3377 induc_def
= PHI_RESULT (induction_phi
);
3379 /* Create the iv update inside the loop */
3380 new_stmt
= gimple_build_assign_with_ops (PLUS_EXPR
, vec_dest
,
3381 induc_def
, vec_step
);
3382 vec_def
= make_ssa_name (vec_dest
, new_stmt
);
3383 gimple_assign_set_lhs (new_stmt
, vec_def
);
3384 gsi_insert_before (&si
, new_stmt
, GSI_SAME_STMT
);
3385 set_vinfo_for_stmt (new_stmt
, new_stmt_vec_info (new_stmt
, loop_vinfo
,
3388 /* Set the arguments of the phi node: */
3389 add_phi_arg (induction_phi
, vec_init
, pe
, UNKNOWN_LOCATION
);
3390 add_phi_arg (induction_phi
, vec_def
, loop_latch_edge (iv_loop
),
3394 /* In case that vectorization factor (VF) is bigger than the number
3395 of elements that we can fit in a vectype (nunits), we have to generate
3396 more than one vector stmt - i.e - we need to "unroll" the
3397 vector stmt by a factor VF/nunits. For more details see documentation
3398 in vectorizable_operation. */
3402 stmt_vec_info prev_stmt_vinfo
;
3403 /* FORNOW. This restriction should be relaxed. */
3404 gcc_assert (!nested_in_vect_loop
);
3406 /* Create the vector that holds the step of the induction. */
3407 if (SCALAR_FLOAT_TYPE_P (TREE_TYPE (step_expr
)))
3409 expr
= build_int_cst (integer_type_node
, nunits
);
3410 expr
= fold_convert (TREE_TYPE (step_expr
), expr
);
3413 expr
= build_int_cst (TREE_TYPE (step_expr
), nunits
);
3414 new_name
= fold_build2 (MULT_EXPR
, TREE_TYPE (step_expr
),
3416 if (TREE_CODE (step_expr
) == SSA_NAME
)
3417 new_name
= vect_init_vector (iv_phi
, new_name
,
3418 TREE_TYPE (step_expr
), NULL
);
3419 t
= unshare_expr (new_name
);
3420 gcc_assert (CONSTANT_CLASS_P (new_name
)
3421 || TREE_CODE (new_name
) == SSA_NAME
);
3422 new_vec
= build_vector_from_val (stepvectype
, t
);
3423 vec_step
= vect_init_vector (iv_phi
, new_vec
, stepvectype
, NULL
);
3425 vec_def
= induc_def
;
3426 prev_stmt_vinfo
= vinfo_for_stmt (induction_phi
);
3427 for (i
= 1; i
< ncopies
; i
++)
3429 /* vec_i = vec_prev + vec_step */
3430 new_stmt
= gimple_build_assign_with_ops (PLUS_EXPR
, vec_dest
,
3432 vec_def
= make_ssa_name (vec_dest
, new_stmt
);
3433 gimple_assign_set_lhs (new_stmt
, vec_def
);
3435 gsi_insert_before (&si
, new_stmt
, GSI_SAME_STMT
);
3436 if (!useless_type_conversion_p (resvectype
, vectype
))
3438 new_stmt
= gimple_build_assign_with_ops
3440 vect_get_new_vect_var (resvectype
, vect_simple_var
,
3442 build1 (VIEW_CONVERT_EXPR
, resvectype
,
3443 gimple_assign_lhs (new_stmt
)), NULL_TREE
);
3444 gimple_assign_set_lhs (new_stmt
,
3446 (gimple_assign_lhs (new_stmt
), new_stmt
));
3447 gsi_insert_before (&si
, new_stmt
, GSI_SAME_STMT
);
3449 set_vinfo_for_stmt (new_stmt
,
3450 new_stmt_vec_info (new_stmt
, loop_vinfo
, NULL
));
3451 STMT_VINFO_RELATED_STMT (prev_stmt_vinfo
) = new_stmt
;
3452 prev_stmt_vinfo
= vinfo_for_stmt (new_stmt
);
3456 if (nested_in_vect_loop
)
3458 /* Find the loop-closed exit-phi of the induction, and record
3459 the final vector of induction results: */
3461 FOR_EACH_IMM_USE_FAST (use_p
, imm_iter
, loop_arg
)
3463 if (!flow_bb_inside_loop_p (iv_loop
, gimple_bb (USE_STMT (use_p
))))
3465 exit_phi
= USE_STMT (use_p
);
3471 stmt_vec_info stmt_vinfo
= vinfo_for_stmt (exit_phi
);
3472 /* FORNOW. Currently not supporting the case that an inner-loop induction
3473 is not used in the outer-loop (i.e. only outside the outer-loop). */
3474 gcc_assert (STMT_VINFO_RELEVANT_P (stmt_vinfo
)
3475 && !STMT_VINFO_LIVE_P (stmt_vinfo
));
3477 STMT_VINFO_VEC_STMT (stmt_vinfo
) = new_stmt
;
3478 if (dump_enabled_p ())
3480 dump_printf_loc (MSG_NOTE
, vect_location
,
3481 "vector of inductions after inner-loop:");
3482 dump_gimple_stmt (MSG_NOTE
, TDF_SLIM
, new_stmt
, 0);
3483 dump_printf (MSG_NOTE
, "\n");
3489 if (dump_enabled_p ())
3491 dump_printf_loc (MSG_NOTE
, vect_location
,
3492 "transform induction: created def-use cycle: ");
3493 dump_gimple_stmt (MSG_NOTE
, TDF_SLIM
, induction_phi
, 0);
3494 dump_printf (MSG_NOTE
, "\n");
3495 dump_gimple_stmt (MSG_NOTE
, TDF_SLIM
,
3496 SSA_NAME_DEF_STMT (vec_def
), 0);
3497 dump_printf (MSG_NOTE
, "\n");
3500 STMT_VINFO_VEC_STMT (phi_info
) = induction_phi
;
3501 if (!useless_type_conversion_p (resvectype
, vectype
))
3503 new_stmt
= gimple_build_assign_with_ops
3505 vect_get_new_vect_var (resvectype
, vect_simple_var
, "vec_iv_"),
3506 build1 (VIEW_CONVERT_EXPR
, resvectype
, induc_def
), NULL_TREE
);
3507 induc_def
= make_ssa_name (gimple_assign_lhs (new_stmt
), new_stmt
);
3508 gimple_assign_set_lhs (new_stmt
, induc_def
);
3509 si
= gsi_after_labels (bb
);
3510 gsi_insert_before (&si
, new_stmt
, GSI_SAME_STMT
);
3511 set_vinfo_for_stmt (new_stmt
,
3512 new_stmt_vec_info (new_stmt
, loop_vinfo
, NULL
));
3513 STMT_VINFO_RELATED_STMT (vinfo_for_stmt (new_stmt
))
3514 = STMT_VINFO_RELATED_STMT (vinfo_for_stmt (induction_phi
));
3521 /* Function get_initial_def_for_reduction
3524 STMT - a stmt that performs a reduction operation in the loop.
3525 INIT_VAL - the initial value of the reduction variable
3528 ADJUSTMENT_DEF - a tree that holds a value to be added to the final result
3529 of the reduction (used for adjusting the epilog - see below).
3530 Return a vector variable, initialized according to the operation that STMT
3531 performs. This vector will be used as the initial value of the
3532 vector of partial results.
3534 Option1 (adjust in epilog): Initialize the vector as follows:
3535 add/bit or/xor: [0,0,...,0,0]
3536 mult/bit and: [1,1,...,1,1]
3537 min/max/cond_expr: [init_val,init_val,..,init_val,init_val]
3538 and when necessary (e.g. add/mult case) let the caller know
3539 that it needs to adjust the result by init_val.
3541 Option2: Initialize the vector as follows:
3542 add/bit or/xor: [init_val,0,0,...,0]
3543 mult/bit and: [init_val,1,1,...,1]
3544 min/max/cond_expr: [init_val,init_val,...,init_val]
3545 and no adjustments are needed.
3547 For example, for the following code:
3553 STMT is 's = s + a[i]', and the reduction variable is 's'.
3554 For a vector of 4 units, we want to return either [0,0,0,init_val],
3555 or [0,0,0,0] and let the caller know that it needs to adjust
3556 the result at the end by 'init_val'.
3558 FORNOW, we are using the 'adjust in epilog' scheme, because this way the
3559 initialization vector is simpler (same element in all entries), if
3560 ADJUSTMENT_DEF is not NULL, and Option2 otherwise.
3562 A cost model should help decide between these two schemes. */
3565 get_initial_def_for_reduction (gimple stmt
, tree init_val
,
3566 tree
*adjustment_def
)
3568 stmt_vec_info stmt_vinfo
= vinfo_for_stmt (stmt
);
3569 loop_vec_info loop_vinfo
= STMT_VINFO_LOOP_VINFO (stmt_vinfo
);
3570 struct loop
*loop
= LOOP_VINFO_LOOP (loop_vinfo
);
3571 tree scalar_type
= TREE_TYPE (init_val
);
3572 tree vectype
= get_vectype_for_scalar_type (scalar_type
);
3574 enum tree_code code
= gimple_assign_rhs_code (stmt
);
3579 bool nested_in_vect_loop
= false;
3581 REAL_VALUE_TYPE real_init_val
= dconst0
;
3582 int int_init_val
= 0;
3583 gimple def_stmt
= NULL
;
3585 gcc_assert (vectype
);
3586 nunits
= TYPE_VECTOR_SUBPARTS (vectype
);
3588 gcc_assert (POINTER_TYPE_P (scalar_type
) || INTEGRAL_TYPE_P (scalar_type
)
3589 || SCALAR_FLOAT_TYPE_P (scalar_type
));
3591 if (nested_in_vect_loop_p (loop
, stmt
))
3592 nested_in_vect_loop
= true;
3594 gcc_assert (loop
== (gimple_bb (stmt
))->loop_father
);
3596 /* In case of double reduction we only create a vector variable to be put
3597 in the reduction phi node. The actual statement creation is done in
3598 vect_create_epilog_for_reduction. */
3599 if (adjustment_def
&& nested_in_vect_loop
3600 && TREE_CODE (init_val
) == SSA_NAME
3601 && (def_stmt
= SSA_NAME_DEF_STMT (init_val
))
3602 && gimple_code (def_stmt
) == GIMPLE_PHI
3603 && flow_bb_inside_loop_p (loop
, gimple_bb (def_stmt
))
3604 && vinfo_for_stmt (def_stmt
)
3605 && STMT_VINFO_DEF_TYPE (vinfo_for_stmt (def_stmt
))
3606 == vect_double_reduction_def
)
3608 *adjustment_def
= NULL
;
3609 return vect_create_destination_var (init_val
, vectype
);
3612 if (TREE_CONSTANT (init_val
))
3614 if (SCALAR_FLOAT_TYPE_P (scalar_type
))
3615 init_value
= build_real (scalar_type
, TREE_REAL_CST (init_val
));
3617 init_value
= build_int_cst (scalar_type
, TREE_INT_CST_LOW (init_val
));
3620 init_value
= init_val
;
3624 case WIDEN_SUM_EXPR
:
3632 /* ADJUSMENT_DEF is NULL when called from
3633 vect_create_epilog_for_reduction to vectorize double reduction. */
3636 if (nested_in_vect_loop
)
3637 *adjustment_def
= vect_get_vec_def_for_operand (init_val
, stmt
,
3640 *adjustment_def
= init_val
;
3643 if (code
== MULT_EXPR
)
3645 real_init_val
= dconst1
;
3649 if (code
== BIT_AND_EXPR
)
3652 if (SCALAR_FLOAT_TYPE_P (scalar_type
))
3653 def_for_init
= build_real (scalar_type
, real_init_val
);
3655 def_for_init
= build_int_cst (scalar_type
, int_init_val
);
3657 /* Create a vector of '0' or '1' except the first element. */
3658 elts
= XALLOCAVEC (tree
, nunits
);
3659 for (i
= nunits
- 2; i
>= 0; --i
)
3660 elts
[i
+ 1] = def_for_init
;
3662 /* Option1: the first element is '0' or '1' as well. */
3665 elts
[0] = def_for_init
;
3666 init_def
= build_vector (vectype
, elts
);
3670 /* Option2: the first element is INIT_VAL. */
3672 if (TREE_CONSTANT (init_val
))
3673 init_def
= build_vector (vectype
, elts
);
3676 vec
<constructor_elt
, va_gc
> *v
;
3677 vec_alloc (v
, nunits
);
3678 CONSTRUCTOR_APPEND_ELT (v
, NULL_TREE
, init_val
);
3679 for (i
= 1; i
< nunits
; ++i
)
3680 CONSTRUCTOR_APPEND_ELT (v
, NULL_TREE
, elts
[i
]);
3681 init_def
= build_constructor (vectype
, v
);
3691 *adjustment_def
= NULL_TREE
;
3692 init_def
= vect_get_vec_def_for_operand (init_val
, stmt
, NULL
);
3696 init_def
= build_vector_from_val (vectype
, init_value
);
3707 /* Function vect_create_epilog_for_reduction
3709 Create code at the loop-epilog to finalize the result of a reduction
3712 VECT_DEFS is list of vector of partial results, i.e., the lhs's of vector
3713 reduction statements.
3714 STMT is the scalar reduction stmt that is being vectorized.
3715 NCOPIES is > 1 in case the vectorization factor (VF) is bigger than the
3716 number of elements that we can fit in a vectype (nunits). In this case
3717 we have to generate more than one vector stmt - i.e - we need to "unroll"
3718 the vector stmt by a factor VF/nunits. For more details see documentation
3719 in vectorizable_operation.
3720 REDUC_CODE is the tree-code for the epilog reduction.
3721 REDUCTION_PHIS is a list of the phi-nodes that carry the reduction
3723 REDUC_INDEX is the index of the operand in the right hand side of the
3724 statement that is defined by REDUCTION_PHI.
3725 DOUBLE_REDUC is TRUE if double reduction phi nodes should be handled.
3726 SLP_NODE is an SLP node containing a group of reduction statements. The
3727 first one in this group is STMT.
3730 1. Creates the reduction def-use cycles: sets the arguments for
3732 The loop-entry argument is the vectorized initial-value of the reduction.
3733 The loop-latch argument is taken from VECT_DEFS - the vector of partial
3735 2. "Reduces" each vector of partial results VECT_DEFS into a single result,
3736 by applying the operation specified by REDUC_CODE if available, or by
3737 other means (whole-vector shifts or a scalar loop).
3738 The function also creates a new phi node at the loop exit to preserve
3739 loop-closed form, as illustrated below.
3741 The flow at the entry to this function:
3744 vec_def = phi <null, null> # REDUCTION_PHI
3745 VECT_DEF = vector_stmt # vectorized form of STMT
3746 s_loop = scalar_stmt # (scalar) STMT
3748 s_out0 = phi <s_loop> # (scalar) EXIT_PHI
3752 The above is transformed by this function into:
3755 vec_def = phi <vec_init, VECT_DEF> # REDUCTION_PHI
3756 VECT_DEF = vector_stmt # vectorized form of STMT
3757 s_loop = scalar_stmt # (scalar) STMT
3759 s_out0 = phi <s_loop> # (scalar) EXIT_PHI
3760 v_out1 = phi <VECT_DEF> # NEW_EXIT_PHI
3761 v_out2 = reduce <v_out1>
3762 s_out3 = extract_field <v_out2, 0>
3763 s_out4 = adjust_result <s_out3>
3769 vect_create_epilog_for_reduction (vec
<tree
> vect_defs
, gimple stmt
,
3770 int ncopies
, enum tree_code reduc_code
,
3771 vec
<gimple
> reduction_phis
,
3772 int reduc_index
, bool double_reduc
,
3775 stmt_vec_info stmt_info
= vinfo_for_stmt (stmt
);
3776 stmt_vec_info prev_phi_info
;
3778 enum machine_mode mode
;
3779 loop_vec_info loop_vinfo
= STMT_VINFO_LOOP_VINFO (stmt_info
);
3780 struct loop
*loop
= LOOP_VINFO_LOOP (loop_vinfo
), *outer_loop
= NULL
;
3781 basic_block exit_bb
;
3784 gimple new_phi
= NULL
, phi
;
3785 gimple_stmt_iterator exit_gsi
;
3787 tree new_temp
= NULL_TREE
, new_dest
, new_name
, new_scalar_dest
;
3788 gimple epilog_stmt
= NULL
;
3789 enum tree_code code
= gimple_assign_rhs_code (stmt
);
3791 tree bitsize
, bitpos
;
3792 tree adjustment_def
= NULL
;
3793 tree vec_initial_def
= NULL
;
3794 tree reduction_op
, expr
, def
;
3795 tree orig_name
, scalar_result
;
3796 imm_use_iterator imm_iter
, phi_imm_iter
;
3797 use_operand_p use_p
, phi_use_p
;
3798 bool extract_scalar_result
= false;
3799 gimple use_stmt
, orig_stmt
, reduction_phi
= NULL
;
3800 bool nested_in_vect_loop
= false;
3801 vec
<gimple
> new_phis
= vNULL
;
3802 vec
<gimple
> inner_phis
= vNULL
;
3803 enum vect_def_type dt
= vect_unknown_def_type
;
3805 vec
<tree
> scalar_results
= vNULL
;
3806 unsigned int group_size
= 1, k
, ratio
;
3807 vec
<tree
> vec_initial_defs
= vNULL
;
3809 bool slp_reduc
= false;
3810 tree new_phi_result
;
3811 gimple inner_phi
= NULL
;
3814 group_size
= SLP_TREE_SCALAR_STMTS (slp_node
).length ();
3816 if (nested_in_vect_loop_p (loop
, stmt
))
3820 nested_in_vect_loop
= true;
3821 gcc_assert (!slp_node
);
3824 switch (get_gimple_rhs_class (gimple_assign_rhs_code (stmt
)))
3826 case GIMPLE_SINGLE_RHS
:
3827 gcc_assert (TREE_OPERAND_LENGTH (gimple_assign_rhs1 (stmt
))
3829 reduction_op
= TREE_OPERAND (gimple_assign_rhs1 (stmt
), reduc_index
);
3831 case GIMPLE_UNARY_RHS
:
3832 reduction_op
= gimple_assign_rhs1 (stmt
);
3834 case GIMPLE_BINARY_RHS
:
3835 reduction_op
= reduc_index
?
3836 gimple_assign_rhs2 (stmt
) : gimple_assign_rhs1 (stmt
);
3838 case GIMPLE_TERNARY_RHS
:
3839 reduction_op
= gimple_op (stmt
, reduc_index
+ 1);
3845 vectype
= get_vectype_for_scalar_type (TREE_TYPE (reduction_op
));
3846 gcc_assert (vectype
);
3847 mode
= TYPE_MODE (vectype
);
3849 /* 1. Create the reduction def-use cycle:
3850 Set the arguments of REDUCTION_PHIS, i.e., transform
3853 vec_def = phi <null, null> # REDUCTION_PHI
3854 VECT_DEF = vector_stmt # vectorized form of STMT
3860 vec_def = phi <vec_init, VECT_DEF> # REDUCTION_PHI
3861 VECT_DEF = vector_stmt # vectorized form of STMT
3864 (in case of SLP, do it for all the phis). */
3866 /* Get the loop-entry arguments. */
3868 vect_get_vec_defs (reduction_op
, NULL_TREE
, stmt
, &vec_initial_defs
,
3869 NULL
, slp_node
, reduc_index
);
3872 vec_initial_defs
.create (1);
3873 /* For the case of reduction, vect_get_vec_def_for_operand returns
3874 the scalar def before the loop, that defines the initial value
3875 of the reduction variable. */
3876 vec_initial_def
= vect_get_vec_def_for_operand (reduction_op
, stmt
,
3878 vec_initial_defs
.quick_push (vec_initial_def
);
3881 /* Set phi nodes arguments. */
3882 FOR_EACH_VEC_ELT (reduction_phis
, i
, phi
)
3884 tree vec_init_def
= vec_initial_defs
[i
];
3885 tree def
= vect_defs
[i
];
3886 for (j
= 0; j
< ncopies
; j
++)
3888 /* Set the loop-entry arg of the reduction-phi. */
3889 add_phi_arg (phi
, vec_init_def
, loop_preheader_edge (loop
),
3892 /* Set the loop-latch arg for the reduction-phi. */
3894 def
= vect_get_vec_def_for_stmt_copy (vect_unknown_def_type
, def
);
3896 add_phi_arg (phi
, def
, loop_latch_edge (loop
), UNKNOWN_LOCATION
);
3898 if (dump_enabled_p ())
3900 dump_printf_loc (MSG_NOTE
, vect_location
,
3901 "transform reduction: created def-use cycle: ");
3902 dump_gimple_stmt (MSG_NOTE
, TDF_SLIM
, phi
, 0);
3903 dump_printf (MSG_NOTE
, "\n");
3904 dump_gimple_stmt (MSG_NOTE
, TDF_SLIM
, SSA_NAME_DEF_STMT (def
), 0);
3905 dump_printf (MSG_NOTE
, "\n");
3908 phi
= STMT_VINFO_RELATED_STMT (vinfo_for_stmt (phi
));
3912 vec_initial_defs
.release ();
3914 /* 2. Create epilog code.
3915 The reduction epilog code operates across the elements of the vector
3916 of partial results computed by the vectorized loop.
3917 The reduction epilog code consists of:
3919 step 1: compute the scalar result in a vector (v_out2)
3920 step 2: extract the scalar result (s_out3) from the vector (v_out2)
3921 step 3: adjust the scalar result (s_out3) if needed.
3923 Step 1 can be accomplished using one the following three schemes:
3924 (scheme 1) using reduc_code, if available.
3925 (scheme 2) using whole-vector shifts, if available.
3926 (scheme 3) using a scalar loop. In this case steps 1+2 above are
3929 The overall epilog code looks like this:
3931 s_out0 = phi <s_loop> # original EXIT_PHI
3932 v_out1 = phi <VECT_DEF> # NEW_EXIT_PHI
3933 v_out2 = reduce <v_out1> # step 1
3934 s_out3 = extract_field <v_out2, 0> # step 2
3935 s_out4 = adjust_result <s_out3> # step 3
3937 (step 3 is optional, and steps 1 and 2 may be combined).
3938 Lastly, the uses of s_out0 are replaced by s_out4. */
3941 /* 2.1 Create new loop-exit-phis to preserve loop-closed form:
3942 v_out1 = phi <VECT_DEF>
3943 Store them in NEW_PHIS. */
3945 exit_bb
= single_exit (loop
)->dest
;
3946 prev_phi_info
= NULL
;
3947 new_phis
.create (vect_defs
.length ());
3948 FOR_EACH_VEC_ELT (vect_defs
, i
, def
)
3950 for (j
= 0; j
< ncopies
; j
++)
3952 tree new_def
= copy_ssa_name (def
, NULL
);
3953 phi
= create_phi_node (new_def
, exit_bb
);
3954 set_vinfo_for_stmt (phi
, new_stmt_vec_info (phi
, loop_vinfo
, NULL
));
3956 new_phis
.quick_push (phi
);
3959 def
= vect_get_vec_def_for_stmt_copy (dt
, def
);
3960 STMT_VINFO_RELATED_STMT (prev_phi_info
) = phi
;
3963 SET_PHI_ARG_DEF (phi
, single_exit (loop
)->dest_idx
, def
);
3964 prev_phi_info
= vinfo_for_stmt (phi
);
3968 /* The epilogue is created for the outer-loop, i.e., for the loop being
3969 vectorized. Create exit phis for the outer loop. */
3973 exit_bb
= single_exit (loop
)->dest
;
3974 inner_phis
.create (vect_defs
.length ());
3975 FOR_EACH_VEC_ELT (new_phis
, i
, phi
)
3977 tree new_result
= copy_ssa_name (PHI_RESULT (phi
), NULL
);
3978 gimple outer_phi
= create_phi_node (new_result
, exit_bb
);
3979 SET_PHI_ARG_DEF (outer_phi
, single_exit (loop
)->dest_idx
,
3981 set_vinfo_for_stmt (outer_phi
, new_stmt_vec_info (outer_phi
,
3983 inner_phis
.quick_push (phi
);
3984 new_phis
[i
] = outer_phi
;
3985 prev_phi_info
= vinfo_for_stmt (outer_phi
);
3986 while (STMT_VINFO_RELATED_STMT (vinfo_for_stmt (phi
)))
3988 phi
= STMT_VINFO_RELATED_STMT (vinfo_for_stmt (phi
));
3989 new_result
= copy_ssa_name (PHI_RESULT (phi
), NULL
);
3990 outer_phi
= create_phi_node (new_result
, exit_bb
);
3991 SET_PHI_ARG_DEF (outer_phi
, single_exit (loop
)->dest_idx
,
3993 set_vinfo_for_stmt (outer_phi
, new_stmt_vec_info (outer_phi
,
3995 STMT_VINFO_RELATED_STMT (prev_phi_info
) = outer_phi
;
3996 prev_phi_info
= vinfo_for_stmt (outer_phi
);
4001 exit_gsi
= gsi_after_labels (exit_bb
);
4003 /* 2.2 Get the relevant tree-code to use in the epilog for schemes 2,3
4004 (i.e. when reduc_code is not available) and in the final adjustment
4005 code (if needed). Also get the original scalar reduction variable as
4006 defined in the loop. In case STMT is a "pattern-stmt" (i.e. - it
4007 represents a reduction pattern), the tree-code and scalar-def are
4008 taken from the original stmt that the pattern-stmt (STMT) replaces.
4009 Otherwise (it is a regular reduction) - the tree-code and scalar-def
4010 are taken from STMT. */
4012 orig_stmt
= STMT_VINFO_RELATED_STMT (stmt_info
);
4015 /* Regular reduction */
4020 /* Reduction pattern */
4021 stmt_vec_info stmt_vinfo
= vinfo_for_stmt (orig_stmt
);
4022 gcc_assert (STMT_VINFO_IN_PATTERN_P (stmt_vinfo
));
4023 gcc_assert (STMT_VINFO_RELATED_STMT (stmt_vinfo
) == stmt
);
4026 code
= gimple_assign_rhs_code (orig_stmt
);
4027 /* For MINUS_EXPR the initial vector is [init_val,0,...,0], therefore,
4028 partial results are added and not subtracted. */
4029 if (code
== MINUS_EXPR
)
4032 scalar_dest
= gimple_assign_lhs (orig_stmt
);
4033 scalar_type
= TREE_TYPE (scalar_dest
);
4034 scalar_results
.create (group_size
);
4035 new_scalar_dest
= vect_create_destination_var (scalar_dest
, NULL
);
4036 bitsize
= TYPE_SIZE (scalar_type
);
4038 /* In case this is a reduction in an inner-loop while vectorizing an outer
4039 loop - we don't need to extract a single scalar result at the end of the
4040 inner-loop (unless it is double reduction, i.e., the use of reduction is
4041 outside the outer-loop). The final vector of partial results will be used
4042 in the vectorized outer-loop, or reduced to a scalar result at the end of
4044 if (nested_in_vect_loop
&& !double_reduc
)
4045 goto vect_finalize_reduction
;
4047 /* SLP reduction without reduction chain, e.g.,
4051 b2 = operation (b1) */
4052 slp_reduc
= (slp_node
&& !GROUP_FIRST_ELEMENT (vinfo_for_stmt (stmt
)));
4054 /* In case of reduction chain, e.g.,
4057 a3 = operation (a2),
4059 we may end up with more than one vector result. Here we reduce them to
4061 if (GROUP_FIRST_ELEMENT (vinfo_for_stmt (stmt
)))
4063 tree first_vect
= PHI_RESULT (new_phis
[0]);
4065 gimple new_vec_stmt
= NULL
;
4067 vec_dest
= vect_create_destination_var (scalar_dest
, vectype
);
4068 for (k
= 1; k
< new_phis
.length (); k
++)
4070 gimple next_phi
= new_phis
[k
];
4071 tree second_vect
= PHI_RESULT (next_phi
);
4073 tmp
= build2 (code
, vectype
, first_vect
, second_vect
);
4074 new_vec_stmt
= gimple_build_assign (vec_dest
, tmp
);
4075 first_vect
= make_ssa_name (vec_dest
, new_vec_stmt
);
4076 gimple_assign_set_lhs (new_vec_stmt
, first_vect
);
4077 gsi_insert_before (&exit_gsi
, new_vec_stmt
, GSI_SAME_STMT
);
4080 new_phi_result
= first_vect
;
4083 new_phis
.truncate (0);
4084 new_phis
.safe_push (new_vec_stmt
);
4088 new_phi_result
= PHI_RESULT (new_phis
[0]);
4090 /* 2.3 Create the reduction code, using one of the three schemes described
4091 above. In SLP we simply need to extract all the elements from the
4092 vector (without reducing them), so we use scalar shifts. */
4093 if (reduc_code
!= ERROR_MARK
&& !slp_reduc
)
4097 /*** Case 1: Create:
4098 v_out2 = reduc_expr <v_out1> */
4100 if (dump_enabled_p ())
4101 dump_printf_loc (MSG_NOTE
, vect_location
,
4102 "Reduce using direct vector reduction.\n");
4104 vec_dest
= vect_create_destination_var (scalar_dest
, vectype
);
4105 tmp
= build1 (reduc_code
, vectype
, new_phi_result
);
4106 epilog_stmt
= gimple_build_assign (vec_dest
, tmp
);
4107 new_temp
= make_ssa_name (vec_dest
, epilog_stmt
);
4108 gimple_assign_set_lhs (epilog_stmt
, new_temp
);
4109 gsi_insert_before (&exit_gsi
, epilog_stmt
, GSI_SAME_STMT
);
4111 extract_scalar_result
= true;
4115 enum tree_code shift_code
= ERROR_MARK
;
4116 bool have_whole_vector_shift
= true;
4118 int element_bitsize
= tree_low_cst (bitsize
, 1);
4119 int vec_size_in_bits
= tree_low_cst (TYPE_SIZE (vectype
), 1);
4122 if (optab_handler (vec_shr_optab
, mode
) != CODE_FOR_nothing
)
4123 shift_code
= VEC_RSHIFT_EXPR
;
4125 have_whole_vector_shift
= false;
4127 /* Regardless of whether we have a whole vector shift, if we're
4128 emulating the operation via tree-vect-generic, we don't want
4129 to use it. Only the first round of the reduction is likely
4130 to still be profitable via emulation. */
4131 /* ??? It might be better to emit a reduction tree code here, so that
4132 tree-vect-generic can expand the first round via bit tricks. */
4133 if (!VECTOR_MODE_P (mode
))
4134 have_whole_vector_shift
= false;
4137 optab optab
= optab_for_tree_code (code
, vectype
, optab_default
);
4138 if (optab_handler (optab
, mode
) == CODE_FOR_nothing
)
4139 have_whole_vector_shift
= false;
4142 if (have_whole_vector_shift
&& !slp_reduc
)
4144 /*** Case 2: Create:
4145 for (offset = VS/2; offset >= element_size; offset/=2)
4147 Create: va' = vec_shift <va, offset>
4148 Create: va = vop <va, va'>
4151 if (dump_enabled_p ())
4152 dump_printf_loc (MSG_NOTE
, vect_location
,
4153 "Reduce using vector shifts\n");
4155 vec_dest
= vect_create_destination_var (scalar_dest
, vectype
);
4156 new_temp
= new_phi_result
;
4157 for (bit_offset
= vec_size_in_bits
/2;
4158 bit_offset
>= element_bitsize
;
4161 tree bitpos
= size_int (bit_offset
);
4163 epilog_stmt
= gimple_build_assign_with_ops (shift_code
,
4164 vec_dest
, new_temp
, bitpos
);
4165 new_name
= make_ssa_name (vec_dest
, epilog_stmt
);
4166 gimple_assign_set_lhs (epilog_stmt
, new_name
);
4167 gsi_insert_before (&exit_gsi
, epilog_stmt
, GSI_SAME_STMT
);
4169 epilog_stmt
= gimple_build_assign_with_ops (code
, vec_dest
,
4170 new_name
, new_temp
);
4171 new_temp
= make_ssa_name (vec_dest
, epilog_stmt
);
4172 gimple_assign_set_lhs (epilog_stmt
, new_temp
);
4173 gsi_insert_before (&exit_gsi
, epilog_stmt
, GSI_SAME_STMT
);
4176 extract_scalar_result
= true;
4182 /*** Case 3: Create:
4183 s = extract_field <v_out2, 0>
4184 for (offset = element_size;
4185 offset < vector_size;
4186 offset += element_size;)
4188 Create: s' = extract_field <v_out2, offset>
4189 Create: s = op <s, s'> // For non SLP cases
4192 if (dump_enabled_p ())
4193 dump_printf_loc (MSG_NOTE
, vect_location
,
4194 "Reduce using scalar code.\n");
4196 vec_size_in_bits
= tree_low_cst (TYPE_SIZE (vectype
), 1);
4197 FOR_EACH_VEC_ELT (new_phis
, i
, new_phi
)
4199 if (gimple_code (new_phi
) == GIMPLE_PHI
)
4200 vec_temp
= PHI_RESULT (new_phi
);
4202 vec_temp
= gimple_assign_lhs (new_phi
);
4203 rhs
= build3 (BIT_FIELD_REF
, scalar_type
, vec_temp
, bitsize
,
4205 epilog_stmt
= gimple_build_assign (new_scalar_dest
, rhs
);
4206 new_temp
= make_ssa_name (new_scalar_dest
, epilog_stmt
);
4207 gimple_assign_set_lhs (epilog_stmt
, new_temp
);
4208 gsi_insert_before (&exit_gsi
, epilog_stmt
, GSI_SAME_STMT
);
4210 /* In SLP we don't need to apply reduction operation, so we just
4211 collect s' values in SCALAR_RESULTS. */
4213 scalar_results
.safe_push (new_temp
);
4215 for (bit_offset
= element_bitsize
;
4216 bit_offset
< vec_size_in_bits
;
4217 bit_offset
+= element_bitsize
)
4219 tree bitpos
= bitsize_int (bit_offset
);
4220 tree rhs
= build3 (BIT_FIELD_REF
, scalar_type
, vec_temp
,
4223 epilog_stmt
= gimple_build_assign (new_scalar_dest
, rhs
);
4224 new_name
= make_ssa_name (new_scalar_dest
, epilog_stmt
);
4225 gimple_assign_set_lhs (epilog_stmt
, new_name
);
4226 gsi_insert_before (&exit_gsi
, epilog_stmt
, GSI_SAME_STMT
);
4230 /* In SLP we don't need to apply reduction operation, so
4231 we just collect s' values in SCALAR_RESULTS. */
4232 new_temp
= new_name
;
4233 scalar_results
.safe_push (new_name
);
4237 epilog_stmt
= gimple_build_assign_with_ops (code
,
4238 new_scalar_dest
, new_name
, new_temp
);
4239 new_temp
= make_ssa_name (new_scalar_dest
, epilog_stmt
);
4240 gimple_assign_set_lhs (epilog_stmt
, new_temp
);
4241 gsi_insert_before (&exit_gsi
, epilog_stmt
, GSI_SAME_STMT
);
4246 /* The only case where we need to reduce scalar results in SLP, is
4247 unrolling. If the size of SCALAR_RESULTS is greater than
4248 GROUP_SIZE, we reduce them combining elements modulo
4252 tree res
, first_res
, new_res
;
4255 /* Reduce multiple scalar results in case of SLP unrolling. */
4256 for (j
= group_size
; scalar_results
.iterate (j
, &res
);
4259 first_res
= scalar_results
[j
% group_size
];
4260 new_stmt
= gimple_build_assign_with_ops (code
,
4261 new_scalar_dest
, first_res
, res
);
4262 new_res
= make_ssa_name (new_scalar_dest
, new_stmt
);
4263 gimple_assign_set_lhs (new_stmt
, new_res
);
4264 gsi_insert_before (&exit_gsi
, new_stmt
, GSI_SAME_STMT
);
4265 scalar_results
[j
% group_size
] = new_res
;
4269 /* Not SLP - we have one scalar to keep in SCALAR_RESULTS. */
4270 scalar_results
.safe_push (new_temp
);
4272 extract_scalar_result
= false;
4276 /* 2.4 Extract the final scalar result. Create:
4277 s_out3 = extract_field <v_out2, bitpos> */
4279 if (extract_scalar_result
)
4283 if (dump_enabled_p ())
4284 dump_printf_loc (MSG_NOTE
, vect_location
,
4285 "extract scalar result\n");
4287 if (BYTES_BIG_ENDIAN
)
4288 bitpos
= size_binop (MULT_EXPR
,
4289 bitsize_int (TYPE_VECTOR_SUBPARTS (vectype
) - 1),
4290 TYPE_SIZE (scalar_type
));
4292 bitpos
= bitsize_zero_node
;
4294 rhs
= build3 (BIT_FIELD_REF
, scalar_type
, new_temp
, bitsize
, bitpos
);
4295 epilog_stmt
= gimple_build_assign (new_scalar_dest
, rhs
);
4296 new_temp
= make_ssa_name (new_scalar_dest
, epilog_stmt
);
4297 gimple_assign_set_lhs (epilog_stmt
, new_temp
);
4298 gsi_insert_before (&exit_gsi
, epilog_stmt
, GSI_SAME_STMT
);
4299 scalar_results
.safe_push (new_temp
);
4302 vect_finalize_reduction
:
4307 /* 2.5 Adjust the final result by the initial value of the reduction
4308 variable. (When such adjustment is not needed, then
4309 'adjustment_def' is zero). For example, if code is PLUS we create:
4310 new_temp = loop_exit_def + adjustment_def */
4314 gcc_assert (!slp_reduc
);
4315 if (nested_in_vect_loop
)
4317 new_phi
= new_phis
[0];
4318 gcc_assert (TREE_CODE (TREE_TYPE (adjustment_def
)) == VECTOR_TYPE
);
4319 expr
= build2 (code
, vectype
, PHI_RESULT (new_phi
), adjustment_def
);
4320 new_dest
= vect_create_destination_var (scalar_dest
, vectype
);
4324 new_temp
= scalar_results
[0];
4325 gcc_assert (TREE_CODE (TREE_TYPE (adjustment_def
)) != VECTOR_TYPE
);
4326 expr
= build2 (code
, scalar_type
, new_temp
, adjustment_def
);
4327 new_dest
= vect_create_destination_var (scalar_dest
, scalar_type
);
4330 epilog_stmt
= gimple_build_assign (new_dest
, expr
);
4331 new_temp
= make_ssa_name (new_dest
, epilog_stmt
);
4332 gimple_assign_set_lhs (epilog_stmt
, new_temp
);
4333 gsi_insert_before (&exit_gsi
, epilog_stmt
, GSI_SAME_STMT
);
4334 if (nested_in_vect_loop
)
4336 set_vinfo_for_stmt (epilog_stmt
,
4337 new_stmt_vec_info (epilog_stmt
, loop_vinfo
,
4339 STMT_VINFO_RELATED_STMT (vinfo_for_stmt (epilog_stmt
)) =
4340 STMT_VINFO_RELATED_STMT (vinfo_for_stmt (new_phi
));
4343 scalar_results
.quick_push (new_temp
);
4345 scalar_results
[0] = new_temp
;
4348 scalar_results
[0] = new_temp
;
4350 new_phis
[0] = epilog_stmt
;
4353 /* 2.6 Handle the loop-exit phis. Replace the uses of scalar loop-exit
4354 phis with new adjusted scalar results, i.e., replace use <s_out0>
4359 s_out0 = phi <s_loop> # (scalar) EXIT_PHI
4360 v_out1 = phi <VECT_DEF> # NEW_EXIT_PHI
4361 v_out2 = reduce <v_out1>
4362 s_out3 = extract_field <v_out2, 0>
4363 s_out4 = adjust_result <s_out3>
4370 s_out0 = phi <s_loop> # (scalar) EXIT_PHI
4371 v_out1 = phi <VECT_DEF> # NEW_EXIT_PHI
4372 v_out2 = reduce <v_out1>
4373 s_out3 = extract_field <v_out2, 0>
4374 s_out4 = adjust_result <s_out3>
4379 /* In SLP reduction chain we reduce vector results into one vector if
4380 necessary, hence we set here GROUP_SIZE to 1. SCALAR_DEST is the LHS of
4381 the last stmt in the reduction chain, since we are looking for the loop
4383 if (GROUP_FIRST_ELEMENT (vinfo_for_stmt (stmt
)))
4385 scalar_dest
= gimple_assign_lhs (
4386 SLP_TREE_SCALAR_STMTS (slp_node
)[group_size
- 1]);
4390 /* In SLP we may have several statements in NEW_PHIS and REDUCTION_PHIS (in
4391 case that GROUP_SIZE is greater than vectorization factor). Therefore, we
4392 need to match SCALAR_RESULTS with corresponding statements. The first
4393 (GROUP_SIZE / number of new vector stmts) scalar results correspond to
4394 the first vector stmt, etc.
4395 (RATIO is equal to (GROUP_SIZE / number of new vector stmts)). */
4396 if (group_size
> new_phis
.length ())
4398 ratio
= group_size
/ new_phis
.length ();
4399 gcc_assert (!(group_size
% new_phis
.length ()));
4404 for (k
= 0; k
< group_size
; k
++)
4408 epilog_stmt
= new_phis
[k
/ ratio
];
4409 reduction_phi
= reduction_phis
[k
/ ratio
];
4411 inner_phi
= inner_phis
[k
/ ratio
];
4416 gimple current_stmt
= SLP_TREE_SCALAR_STMTS (slp_node
)[k
];
4418 orig_stmt
= STMT_VINFO_RELATED_STMT (vinfo_for_stmt (current_stmt
));
4419 /* SLP statements can't participate in patterns. */
4420 gcc_assert (!orig_stmt
);
4421 scalar_dest
= gimple_assign_lhs (current_stmt
);
4425 /* Find the loop-closed-use at the loop exit of the original scalar
4426 result. (The reduction result is expected to have two immediate uses -
4427 one at the latch block, and one at the loop exit). */
4428 FOR_EACH_IMM_USE_FAST (use_p
, imm_iter
, scalar_dest
)
4429 if (!flow_bb_inside_loop_p (loop
, gimple_bb (USE_STMT (use_p
)))
4430 && !is_gimple_debug (USE_STMT (use_p
)))
4431 phis
.safe_push (USE_STMT (use_p
));
4433 /* While we expect to have found an exit_phi because of loop-closed-ssa
4434 form we can end up without one if the scalar cycle is dead. */
4436 FOR_EACH_VEC_ELT (phis
, i
, exit_phi
)
4440 stmt_vec_info exit_phi_vinfo
= vinfo_for_stmt (exit_phi
);
4443 /* FORNOW. Currently not supporting the case that an inner-loop
4444 reduction is not used in the outer-loop (but only outside the
4445 outer-loop), unless it is double reduction. */
4446 gcc_assert ((STMT_VINFO_RELEVANT_P (exit_phi_vinfo
)
4447 && !STMT_VINFO_LIVE_P (exit_phi_vinfo
))
4450 STMT_VINFO_VEC_STMT (exit_phi_vinfo
) = epilog_stmt
;
4452 || STMT_VINFO_DEF_TYPE (exit_phi_vinfo
)
4453 != vect_double_reduction_def
)
4456 /* Handle double reduction:
4458 stmt1: s1 = phi <s0, s2> - double reduction phi (outer loop)
4459 stmt2: s3 = phi <s1, s4> - (regular) reduc phi (inner loop)
4460 stmt3: s4 = use (s3) - (regular) reduc stmt (inner loop)
4461 stmt4: s2 = phi <s4> - double reduction stmt (outer loop)
4463 At that point the regular reduction (stmt2 and stmt3) is
4464 already vectorized, as well as the exit phi node, stmt4.
4465 Here we vectorize the phi node of double reduction, stmt1, and
4466 update all relevant statements. */
4468 /* Go through all the uses of s2 to find double reduction phi
4469 node, i.e., stmt1 above. */
4470 orig_name
= PHI_RESULT (exit_phi
);
4471 FOR_EACH_IMM_USE_STMT (use_stmt
, imm_iter
, orig_name
)
4473 stmt_vec_info use_stmt_vinfo
;
4474 stmt_vec_info new_phi_vinfo
;
4475 tree vect_phi_init
, preheader_arg
, vect_phi_res
, init_def
;
4476 basic_block bb
= gimple_bb (use_stmt
);
4479 /* Check that USE_STMT is really double reduction phi
4481 if (gimple_code (use_stmt
) != GIMPLE_PHI
4482 || gimple_phi_num_args (use_stmt
) != 2
4483 || bb
->loop_father
!= outer_loop
)
4485 use_stmt_vinfo
= vinfo_for_stmt (use_stmt
);
4487 || STMT_VINFO_DEF_TYPE (use_stmt_vinfo
)
4488 != vect_double_reduction_def
)
4491 /* Create vector phi node for double reduction:
4492 vs1 = phi <vs0, vs2>
4493 vs1 was created previously in this function by a call to
4494 vect_get_vec_def_for_operand and is stored in
4496 vs2 is defined by INNER_PHI, the vectorized EXIT_PHI;
4497 vs0 is created here. */
4499 /* Create vector phi node. */
4500 vect_phi
= create_phi_node (vec_initial_def
, bb
);
4501 new_phi_vinfo
= new_stmt_vec_info (vect_phi
,
4502 loop_vec_info_for_loop (outer_loop
), NULL
);
4503 set_vinfo_for_stmt (vect_phi
, new_phi_vinfo
);
4505 /* Create vs0 - initial def of the double reduction phi. */
4506 preheader_arg
= PHI_ARG_DEF_FROM_EDGE (use_stmt
,
4507 loop_preheader_edge (outer_loop
));
4508 init_def
= get_initial_def_for_reduction (stmt
,
4509 preheader_arg
, NULL
);
4510 vect_phi_init
= vect_init_vector (use_stmt
, init_def
,
4513 /* Update phi node arguments with vs0 and vs2. */
4514 add_phi_arg (vect_phi
, vect_phi_init
,
4515 loop_preheader_edge (outer_loop
),
4517 add_phi_arg (vect_phi
, PHI_RESULT (inner_phi
),
4518 loop_latch_edge (outer_loop
), UNKNOWN_LOCATION
);
4519 if (dump_enabled_p ())
4521 dump_printf_loc (MSG_NOTE
, vect_location
,
4522 "created double reduction phi node: ");
4523 dump_gimple_stmt (MSG_NOTE
, TDF_SLIM
, vect_phi
, 0);
4524 dump_printf (MSG_NOTE
, "\n");
4527 vect_phi_res
= PHI_RESULT (vect_phi
);
4529 /* Replace the use, i.e., set the correct vs1 in the regular
4530 reduction phi node. FORNOW, NCOPIES is always 1, so the
4531 loop is redundant. */
4532 use
= reduction_phi
;
4533 for (j
= 0; j
< ncopies
; j
++)
4535 edge pr_edge
= loop_preheader_edge (loop
);
4536 SET_PHI_ARG_DEF (use
, pr_edge
->dest_idx
, vect_phi_res
);
4537 use
= STMT_VINFO_RELATED_STMT (vinfo_for_stmt (use
));
4544 if (nested_in_vect_loop
)
4553 /* Find the loop-closed-use at the loop exit of the original scalar
4554 result. (The reduction result is expected to have two immediate uses,
4555 one at the latch block, and one at the loop exit). For double
4556 reductions we are looking for exit phis of the outer loop. */
4557 FOR_EACH_IMM_USE_FAST (use_p
, imm_iter
, scalar_dest
)
4559 if (!flow_bb_inside_loop_p (loop
, gimple_bb (USE_STMT (use_p
))))
4561 if (!is_gimple_debug (USE_STMT (use_p
)))
4562 phis
.safe_push (USE_STMT (use_p
));
4566 if (double_reduc
&& gimple_code (USE_STMT (use_p
)) == GIMPLE_PHI
)
4568 tree phi_res
= PHI_RESULT (USE_STMT (use_p
));
4570 FOR_EACH_IMM_USE_FAST (phi_use_p
, phi_imm_iter
, phi_res
)
4572 if (!flow_bb_inside_loop_p (loop
,
4573 gimple_bb (USE_STMT (phi_use_p
)))
4574 && !is_gimple_debug (USE_STMT (phi_use_p
)))
4575 phis
.safe_push (USE_STMT (phi_use_p
));
4581 FOR_EACH_VEC_ELT (phis
, i
, exit_phi
)
4583 /* Replace the uses: */
4584 orig_name
= PHI_RESULT (exit_phi
);
4585 scalar_result
= scalar_results
[k
];
4586 FOR_EACH_IMM_USE_STMT (use_stmt
, imm_iter
, orig_name
)
4587 FOR_EACH_IMM_USE_ON_STMT (use_p
, imm_iter
)
4588 SET_USE (use_p
, scalar_result
);
4594 scalar_results
.release ();
4595 inner_phis
.release ();
4596 new_phis
.release ();
4600 /* Function vectorizable_reduction.
4602 Check if STMT performs a reduction operation that can be vectorized.
4603 If VEC_STMT is also passed, vectorize the STMT: create a vectorized
4604 stmt to replace it, put it in VEC_STMT, and insert it at GSI.
4605 Return FALSE if not a vectorizable STMT, TRUE otherwise.
4607 This function also handles reduction idioms (patterns) that have been
4608 recognized in advance during vect_pattern_recog. In this case, STMT may be
4610 X = pattern_expr (arg0, arg1, ..., X)
4611 and it's STMT_VINFO_RELATED_STMT points to the last stmt in the original
4612 sequence that had been detected and replaced by the pattern-stmt (STMT).
4614 In some cases of reduction patterns, the type of the reduction variable X is
4615 different than the type of the other arguments of STMT.
4616 In such cases, the vectype that is used when transforming STMT into a vector
4617 stmt is different than the vectype that is used to determine the
4618 vectorization factor, because it consists of a different number of elements
4619 than the actual number of elements that are being operated upon in parallel.
4621 For example, consider an accumulation of shorts into an int accumulator.
4622 On some targets it's possible to vectorize this pattern operating on 8
4623 shorts at a time (hence, the vectype for purposes of determining the
4624 vectorization factor should be V8HI); on the other hand, the vectype that
4625 is used to create the vector form is actually V4SI (the type of the result).
4627 Upon entry to this function, STMT_VINFO_VECTYPE records the vectype that
4628 indicates what is the actual level of parallelism (V8HI in the example), so
4629 that the right vectorization factor would be derived. This vectype
4630 corresponds to the type of arguments to the reduction stmt, and should *NOT*
4631 be used to create the vectorized stmt. The right vectype for the vectorized
4632 stmt is obtained from the type of the result X:
4633 get_vectype_for_scalar_type (TREE_TYPE (X))
4635 This means that, contrary to "regular" reductions (or "regular" stmts in
4636 general), the following equation:
4637 STMT_VINFO_VECTYPE == get_vectype_for_scalar_type (TREE_TYPE (X))
4638 does *NOT* necessarily hold for reduction patterns. */
4641 vectorizable_reduction (gimple stmt
, gimple_stmt_iterator
*gsi
,
4642 gimple
*vec_stmt
, slp_tree slp_node
)
4646 tree loop_vec_def0
= NULL_TREE
, loop_vec_def1
= NULL_TREE
;
4647 stmt_vec_info stmt_info
= vinfo_for_stmt (stmt
);
4648 tree vectype_out
= STMT_VINFO_VECTYPE (stmt_info
);
4649 tree vectype_in
= NULL_TREE
;
4650 loop_vec_info loop_vinfo
= STMT_VINFO_LOOP_VINFO (stmt_info
);
4651 struct loop
*loop
= LOOP_VINFO_LOOP (loop_vinfo
);
4652 enum tree_code code
, orig_code
, epilog_reduc_code
;
4653 enum machine_mode vec_mode
;
4655 optab optab
, reduc_optab
;
4656 tree new_temp
= NULL_TREE
;
4659 enum vect_def_type dt
;
4660 gimple new_phi
= NULL
;
4664 stmt_vec_info orig_stmt_info
;
4665 tree expr
= NULL_TREE
;
4669 stmt_vec_info prev_stmt_info
, prev_phi_info
;
4670 bool single_defuse_cycle
= false;
4671 tree reduc_def
= NULL_TREE
;
4672 gimple new_stmt
= NULL
;
4675 bool nested_cycle
= false, found_nested_cycle_def
= false;
4676 gimple reduc_def_stmt
= NULL
;
4677 /* The default is that the reduction variable is the last in statement. */
4678 int reduc_index
= 2;
4679 bool double_reduc
= false, dummy
;
4681 struct loop
* def_stmt_loop
, *outer_loop
= NULL
;
4683 gimple def_arg_stmt
;
4684 vec
<tree
> vec_oprnds0
= vNULL
;
4685 vec
<tree
> vec_oprnds1
= vNULL
;
4686 vec
<tree
> vect_defs
= vNULL
;
4687 vec
<gimple
> phis
= vNULL
;
4689 tree def0
, def1
, tem
, op0
, op1
= NULL_TREE
;
4691 /* In case of reduction chain we switch to the first stmt in the chain, but
4692 we don't update STMT_INFO, since only the last stmt is marked as reduction
4693 and has reduction properties. */
4694 if (GROUP_FIRST_ELEMENT (vinfo_for_stmt (stmt
)))
4695 stmt
= GROUP_FIRST_ELEMENT (stmt_info
);
4697 if (nested_in_vect_loop_p (loop
, stmt
))
4701 nested_cycle
= true;
4704 /* 1. Is vectorizable reduction? */
4705 /* Not supportable if the reduction variable is used in the loop, unless
4706 it's a reduction chain. */
4707 if (STMT_VINFO_RELEVANT (stmt_info
) > vect_used_in_outer
4708 && !GROUP_FIRST_ELEMENT (stmt_info
))
4711 /* Reductions that are not used even in an enclosing outer-loop,
4712 are expected to be "live" (used out of the loop). */
4713 if (STMT_VINFO_RELEVANT (stmt_info
) == vect_unused_in_scope
4714 && !STMT_VINFO_LIVE_P (stmt_info
))
4717 /* Make sure it was already recognized as a reduction computation. */
4718 if (STMT_VINFO_DEF_TYPE (stmt_info
) != vect_reduction_def
4719 && STMT_VINFO_DEF_TYPE (stmt_info
) != vect_nested_cycle
)
4722 /* 2. Has this been recognized as a reduction pattern?
4724 Check if STMT represents a pattern that has been recognized
4725 in earlier analysis stages. For stmts that represent a pattern,
4726 the STMT_VINFO_RELATED_STMT field records the last stmt in
4727 the original sequence that constitutes the pattern. */
4729 orig_stmt
= STMT_VINFO_RELATED_STMT (stmt_info
);
4732 orig_stmt_info
= vinfo_for_stmt (orig_stmt
);
4733 gcc_assert (STMT_VINFO_IN_PATTERN_P (orig_stmt_info
));
4734 gcc_assert (!STMT_VINFO_IN_PATTERN_P (stmt_info
));
4737 /* 3. Check the operands of the operation. The first operands are defined
4738 inside the loop body. The last operand is the reduction variable,
4739 which is defined by the loop-header-phi. */
4741 gcc_assert (is_gimple_assign (stmt
));
4744 switch (get_gimple_rhs_class (gimple_assign_rhs_code (stmt
)))
4746 case GIMPLE_SINGLE_RHS
:
4747 op_type
= TREE_OPERAND_LENGTH (gimple_assign_rhs1 (stmt
));
4748 if (op_type
== ternary_op
)
4750 tree rhs
= gimple_assign_rhs1 (stmt
);
4751 ops
[0] = TREE_OPERAND (rhs
, 0);
4752 ops
[1] = TREE_OPERAND (rhs
, 1);
4753 ops
[2] = TREE_OPERAND (rhs
, 2);
4754 code
= TREE_CODE (rhs
);
4760 case GIMPLE_BINARY_RHS
:
4761 code
= gimple_assign_rhs_code (stmt
);
4762 op_type
= TREE_CODE_LENGTH (code
);
4763 gcc_assert (op_type
== binary_op
);
4764 ops
[0] = gimple_assign_rhs1 (stmt
);
4765 ops
[1] = gimple_assign_rhs2 (stmt
);
4768 case GIMPLE_TERNARY_RHS
:
4769 code
= gimple_assign_rhs_code (stmt
);
4770 op_type
= TREE_CODE_LENGTH (code
);
4771 gcc_assert (op_type
== ternary_op
);
4772 ops
[0] = gimple_assign_rhs1 (stmt
);
4773 ops
[1] = gimple_assign_rhs2 (stmt
);
4774 ops
[2] = gimple_assign_rhs3 (stmt
);
4777 case GIMPLE_UNARY_RHS
:
4784 if (code
== COND_EXPR
&& slp_node
)
4787 scalar_dest
= gimple_assign_lhs (stmt
);
4788 scalar_type
= TREE_TYPE (scalar_dest
);
4789 if (!POINTER_TYPE_P (scalar_type
) && !INTEGRAL_TYPE_P (scalar_type
)
4790 && !SCALAR_FLOAT_TYPE_P (scalar_type
))
4793 /* Do not try to vectorize bit-precision reductions. */
4794 if ((TYPE_PRECISION (scalar_type
)
4795 != GET_MODE_PRECISION (TYPE_MODE (scalar_type
))))
4798 /* All uses but the last are expected to be defined in the loop.
4799 The last use is the reduction variable. In case of nested cycle this
4800 assumption is not true: we use reduc_index to record the index of the
4801 reduction variable. */
4802 for (i
= 0; i
< op_type
- 1; i
++)
4804 /* The condition of COND_EXPR is checked in vectorizable_condition(). */
4805 if (i
== 0 && code
== COND_EXPR
)
4808 is_simple_use
= vect_is_simple_use_1 (ops
[i
], stmt
, loop_vinfo
, NULL
,
4809 &def_stmt
, &def
, &dt
, &tem
);
4812 gcc_assert (is_simple_use
);
4814 if (dt
!= vect_internal_def
4815 && dt
!= vect_external_def
4816 && dt
!= vect_constant_def
4817 && dt
!= vect_induction_def
4818 && !(dt
== vect_nested_cycle
&& nested_cycle
))
4821 if (dt
== vect_nested_cycle
)
4823 found_nested_cycle_def
= true;
4824 reduc_def_stmt
= def_stmt
;
4829 is_simple_use
= vect_is_simple_use_1 (ops
[i
], stmt
, loop_vinfo
, NULL
,
4830 &def_stmt
, &def
, &dt
, &tem
);
4833 gcc_assert (is_simple_use
);
4834 if (!(dt
== vect_reduction_def
4835 || dt
== vect_nested_cycle
4836 || ((dt
== vect_internal_def
|| dt
== vect_external_def
4837 || dt
== vect_constant_def
|| dt
== vect_induction_def
)
4838 && nested_cycle
&& found_nested_cycle_def
)))
4840 /* For pattern recognized stmts, orig_stmt might be a reduction,
4841 but some helper statements for the pattern might not, or
4842 might be COND_EXPRs with reduction uses in the condition. */
4843 gcc_assert (orig_stmt
);
4846 if (!found_nested_cycle_def
)
4847 reduc_def_stmt
= def_stmt
;
4849 gcc_assert (gimple_code (reduc_def_stmt
) == GIMPLE_PHI
);
4851 gcc_assert (orig_stmt
== vect_is_simple_reduction (loop_vinfo
,
4857 gimple tmp
= vect_is_simple_reduction (loop_vinfo
, reduc_def_stmt
,
4858 !nested_cycle
, &dummy
);
4859 /* We changed STMT to be the first stmt in reduction chain, hence we
4860 check that in this case the first element in the chain is STMT. */
4861 gcc_assert (stmt
== tmp
4862 || GROUP_FIRST_ELEMENT (vinfo_for_stmt (tmp
)) == stmt
);
4865 if (STMT_VINFO_LIVE_P (vinfo_for_stmt (reduc_def_stmt
)))
4868 if (slp_node
|| PURE_SLP_STMT (stmt_info
))
4871 ncopies
= (LOOP_VINFO_VECT_FACTOR (loop_vinfo
)
4872 / TYPE_VECTOR_SUBPARTS (vectype_in
));
4874 gcc_assert (ncopies
>= 1);
4876 vec_mode
= TYPE_MODE (vectype_in
);
4878 if (code
== COND_EXPR
)
4880 if (!vectorizable_condition (stmt
, gsi
, NULL
, ops
[reduc_index
], 0, NULL
))
4882 if (dump_enabled_p ())
4883 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
4884 "unsupported condition in reduction\n");
4891 /* 4. Supportable by target? */
4893 if (code
== LSHIFT_EXPR
|| code
== RSHIFT_EXPR
4894 || code
== LROTATE_EXPR
|| code
== RROTATE_EXPR
)
4896 /* Shifts and rotates are only supported by vectorizable_shifts,
4897 not vectorizable_reduction. */
4898 if (dump_enabled_p ())
4899 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
4900 "unsupported shift or rotation.\n");
4904 /* 4.1. check support for the operation in the loop */
4905 optab
= optab_for_tree_code (code
, vectype_in
, optab_default
);
4908 if (dump_enabled_p ())
4909 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
4915 if (optab_handler (optab
, vec_mode
) == CODE_FOR_nothing
)
4917 if (dump_enabled_p ())
4918 dump_printf (MSG_NOTE
, "op not supported by target.\n");
4920 if (GET_MODE_SIZE (vec_mode
) != UNITS_PER_WORD
4921 || LOOP_VINFO_VECT_FACTOR (loop_vinfo
)
4922 < vect_min_worthwhile_factor (code
))
4925 if (dump_enabled_p ())
4926 dump_printf (MSG_NOTE
, "proceeding using word mode.\n");
4929 /* Worthwhile without SIMD support? */
4930 if (!VECTOR_MODE_P (TYPE_MODE (vectype_in
))
4931 && LOOP_VINFO_VECT_FACTOR (loop_vinfo
)
4932 < vect_min_worthwhile_factor (code
))
4934 if (dump_enabled_p ())
4935 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
4936 "not worthwhile without SIMD support.\n");
4942 /* 4.2. Check support for the epilog operation.
4944 If STMT represents a reduction pattern, then the type of the
4945 reduction variable may be different than the type of the rest
4946 of the arguments. For example, consider the case of accumulation
4947 of shorts into an int accumulator; The original code:
4948 S1: int_a = (int) short_a;
4949 orig_stmt-> S2: int_acc = plus <int_a ,int_acc>;
4952 STMT: int_acc = widen_sum <short_a, int_acc>
4955 1. The tree-code that is used to create the vector operation in the
4956 epilog code (that reduces the partial results) is not the
4957 tree-code of STMT, but is rather the tree-code of the original
4958 stmt from the pattern that STMT is replacing. I.e, in the example
4959 above we want to use 'widen_sum' in the loop, but 'plus' in the
4961 2. The type (mode) we use to check available target support
4962 for the vector operation to be created in the *epilog*, is
4963 determined by the type of the reduction variable (in the example
4964 above we'd check this: optab_handler (plus_optab, vect_int_mode])).
4965 However the type (mode) we use to check available target support
4966 for the vector operation to be created *inside the loop*, is
4967 determined by the type of the other arguments to STMT (in the
4968 example we'd check this: optab_handler (widen_sum_optab,
4971 This is contrary to "regular" reductions, in which the types of all
4972 the arguments are the same as the type of the reduction variable.
4973 For "regular" reductions we can therefore use the same vector type
4974 (and also the same tree-code) when generating the epilog code and
4975 when generating the code inside the loop. */
4979 /* This is a reduction pattern: get the vectype from the type of the
4980 reduction variable, and get the tree-code from orig_stmt. */
4981 orig_code
= gimple_assign_rhs_code (orig_stmt
);
4982 gcc_assert (vectype_out
);
4983 vec_mode
= TYPE_MODE (vectype_out
);
4987 /* Regular reduction: use the same vectype and tree-code as used for
4988 the vector code inside the loop can be used for the epilog code. */
4994 def_bb
= gimple_bb (reduc_def_stmt
);
4995 def_stmt_loop
= def_bb
->loop_father
;
4996 def_arg
= PHI_ARG_DEF_FROM_EDGE (reduc_def_stmt
,
4997 loop_preheader_edge (def_stmt_loop
));
4998 if (TREE_CODE (def_arg
) == SSA_NAME
4999 && (def_arg_stmt
= SSA_NAME_DEF_STMT (def_arg
))
5000 && gimple_code (def_arg_stmt
) == GIMPLE_PHI
5001 && flow_bb_inside_loop_p (outer_loop
, gimple_bb (def_arg_stmt
))
5002 && vinfo_for_stmt (def_arg_stmt
)
5003 && STMT_VINFO_DEF_TYPE (vinfo_for_stmt (def_arg_stmt
))
5004 == vect_double_reduction_def
)
5005 double_reduc
= true;
5008 epilog_reduc_code
= ERROR_MARK
;
5009 if (reduction_code_for_scalar_code (orig_code
, &epilog_reduc_code
))
5011 reduc_optab
= optab_for_tree_code (epilog_reduc_code
, vectype_out
,
5015 if (dump_enabled_p ())
5016 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
5017 "no optab for reduction.\n");
5019 epilog_reduc_code
= ERROR_MARK
;
5023 && optab_handler (reduc_optab
, vec_mode
) == CODE_FOR_nothing
)
5025 if (dump_enabled_p ())
5026 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
5027 "reduc op not supported by target.\n");
5029 epilog_reduc_code
= ERROR_MARK
;
5034 if (!nested_cycle
|| double_reduc
)
5036 if (dump_enabled_p ())
5037 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
5038 "no reduc code for scalar code.\n");
5044 if (double_reduc
&& ncopies
> 1)
5046 if (dump_enabled_p ())
5047 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
5048 "multiple types in double reduction\n");
5053 /* In case of widenning multiplication by a constant, we update the type
5054 of the constant to be the type of the other operand. We check that the
5055 constant fits the type in the pattern recognition pass. */
5056 if (code
== DOT_PROD_EXPR
5057 && !types_compatible_p (TREE_TYPE (ops
[0]), TREE_TYPE (ops
[1])))
5059 if (TREE_CODE (ops
[0]) == INTEGER_CST
)
5060 ops
[0] = fold_convert (TREE_TYPE (ops
[1]), ops
[0]);
5061 else if (TREE_CODE (ops
[1]) == INTEGER_CST
)
5062 ops
[1] = fold_convert (TREE_TYPE (ops
[0]), ops
[1]);
5065 if (dump_enabled_p ())
5066 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
5067 "invalid types in dot-prod\n");
5073 if (!vec_stmt
) /* transformation not required. */
5075 if (!vect_model_reduction_cost (stmt_info
, epilog_reduc_code
, ncopies
))
5077 STMT_VINFO_TYPE (stmt_info
) = reduc_vec_info_type
;
5083 if (dump_enabled_p ())
5084 dump_printf_loc (MSG_NOTE
, vect_location
, "transform reduction.\n");
5086 /* FORNOW: Multiple types are not supported for condition. */
5087 if (code
== COND_EXPR
)
5088 gcc_assert (ncopies
== 1);
5090 /* Create the destination vector */
5091 vec_dest
= vect_create_destination_var (scalar_dest
, vectype_out
);
5093 /* In case the vectorization factor (VF) is bigger than the number
5094 of elements that we can fit in a vectype (nunits), we have to generate
5095 more than one vector stmt - i.e - we need to "unroll" the
5096 vector stmt by a factor VF/nunits. For more details see documentation
5097 in vectorizable_operation. */
5099 /* If the reduction is used in an outer loop we need to generate
5100 VF intermediate results, like so (e.g. for ncopies=2):
5105 (i.e. we generate VF results in 2 registers).
5106 In this case we have a separate def-use cycle for each copy, and therefore
5107 for each copy we get the vector def for the reduction variable from the
5108 respective phi node created for this copy.
5110 Otherwise (the reduction is unused in the loop nest), we can combine
5111 together intermediate results, like so (e.g. for ncopies=2):
5115 (i.e. we generate VF/2 results in a single register).
5116 In this case for each copy we get the vector def for the reduction variable
5117 from the vectorized reduction operation generated in the previous iteration.
5120 if (STMT_VINFO_RELEVANT (stmt_info
) == vect_unused_in_scope
)
5122 single_defuse_cycle
= true;
5126 epilog_copies
= ncopies
;
5128 prev_stmt_info
= NULL
;
5129 prev_phi_info
= NULL
;
5132 vec_num
= SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node
);
5133 gcc_assert (TYPE_VECTOR_SUBPARTS (vectype_out
)
5134 == TYPE_VECTOR_SUBPARTS (vectype_in
));
5139 vec_oprnds0
.create (1);
5140 if (op_type
== ternary_op
)
5141 vec_oprnds1
.create (1);
5144 phis
.create (vec_num
);
5145 vect_defs
.create (vec_num
);
5147 vect_defs
.quick_push (NULL_TREE
);
5149 for (j
= 0; j
< ncopies
; j
++)
5151 if (j
== 0 || !single_defuse_cycle
)
5153 for (i
= 0; i
< vec_num
; i
++)
5155 /* Create the reduction-phi that defines the reduction
5157 new_phi
= create_phi_node (vec_dest
, loop
->header
);
5158 set_vinfo_for_stmt (new_phi
,
5159 new_stmt_vec_info (new_phi
, loop_vinfo
,
5161 if (j
== 0 || slp_node
)
5162 phis
.quick_push (new_phi
);
5166 if (code
== COND_EXPR
)
5168 gcc_assert (!slp_node
);
5169 vectorizable_condition (stmt
, gsi
, vec_stmt
,
5170 PHI_RESULT (phis
[0]),
5172 /* Multiple types are not supported for condition. */
5179 op0
= ops
[!reduc_index
];
5180 if (op_type
== ternary_op
)
5182 if (reduc_index
== 0)
5189 vect_get_vec_defs (op0
, op1
, stmt
, &vec_oprnds0
, &vec_oprnds1
,
5193 loop_vec_def0
= vect_get_vec_def_for_operand (ops
[!reduc_index
],
5195 vec_oprnds0
.quick_push (loop_vec_def0
);
5196 if (op_type
== ternary_op
)
5198 loop_vec_def1
= vect_get_vec_def_for_operand (op1
, stmt
,
5200 vec_oprnds1
.quick_push (loop_vec_def1
);
5208 enum vect_def_type dt
;
5212 vect_is_simple_use (ops
[!reduc_index
], stmt
, loop_vinfo
, NULL
,
5213 &dummy_stmt
, &dummy
, &dt
);
5214 loop_vec_def0
= vect_get_vec_def_for_stmt_copy (dt
,
5216 vec_oprnds0
[0] = loop_vec_def0
;
5217 if (op_type
== ternary_op
)
5219 vect_is_simple_use (op1
, stmt
, loop_vinfo
, NULL
, &dummy_stmt
,
5221 loop_vec_def1
= vect_get_vec_def_for_stmt_copy (dt
,
5223 vec_oprnds1
[0] = loop_vec_def1
;
5227 if (single_defuse_cycle
)
5228 reduc_def
= gimple_assign_lhs (new_stmt
);
5230 STMT_VINFO_RELATED_STMT (prev_phi_info
) = new_phi
;
5233 FOR_EACH_VEC_ELT (vec_oprnds0
, i
, def0
)
5236 reduc_def
= PHI_RESULT (phis
[i
]);
5239 if (!single_defuse_cycle
|| j
== 0)
5240 reduc_def
= PHI_RESULT (new_phi
);
5243 def1
= ((op_type
== ternary_op
)
5244 ? vec_oprnds1
[i
] : NULL
);
5245 if (op_type
== binary_op
)
5247 if (reduc_index
== 0)
5248 expr
= build2 (code
, vectype_out
, reduc_def
, def0
);
5250 expr
= build2 (code
, vectype_out
, def0
, reduc_def
);
5254 if (reduc_index
== 0)
5255 expr
= build3 (code
, vectype_out
, reduc_def
, def0
, def1
);
5258 if (reduc_index
== 1)
5259 expr
= build3 (code
, vectype_out
, def0
, reduc_def
, def1
);
5261 expr
= build3 (code
, vectype_out
, def0
, def1
, reduc_def
);
5265 new_stmt
= gimple_build_assign (vec_dest
, expr
);
5266 new_temp
= make_ssa_name (vec_dest
, new_stmt
);
5267 gimple_assign_set_lhs (new_stmt
, new_temp
);
5268 vect_finish_stmt_generation (stmt
, new_stmt
, gsi
);
5272 SLP_TREE_VEC_STMTS (slp_node
).quick_push (new_stmt
);
5273 vect_defs
.quick_push (new_temp
);
5276 vect_defs
[0] = new_temp
;
5283 STMT_VINFO_VEC_STMT (stmt_info
) = *vec_stmt
= new_stmt
;
5285 STMT_VINFO_RELATED_STMT (prev_stmt_info
) = new_stmt
;
5287 prev_stmt_info
= vinfo_for_stmt (new_stmt
);
5288 prev_phi_info
= vinfo_for_stmt (new_phi
);
5291 /* Finalize the reduction-phi (set its arguments) and create the
5292 epilog reduction code. */
5293 if ((!single_defuse_cycle
|| code
== COND_EXPR
) && !slp_node
)
5295 new_temp
= gimple_assign_lhs (*vec_stmt
);
5296 vect_defs
[0] = new_temp
;
5299 vect_create_epilog_for_reduction (vect_defs
, stmt
, epilog_copies
,
5300 epilog_reduc_code
, phis
, reduc_index
,
5301 double_reduc
, slp_node
);
5304 vect_defs
.release ();
5305 vec_oprnds0
.release ();
5306 vec_oprnds1
.release ();
5311 /* Function vect_min_worthwhile_factor.
5313 For a loop where we could vectorize the operation indicated by CODE,
5314 return the minimum vectorization factor that makes it worthwhile
5315 to use generic vectors. */
5317 vect_min_worthwhile_factor (enum tree_code code
)
5338 /* Function vectorizable_induction
5340 Check if PHI performs an induction computation that can be vectorized.
5341 If VEC_STMT is also passed, vectorize the induction PHI: create a vectorized
5342 phi to replace it, put it in VEC_STMT, and add it to the same basic block.
5343 Return FALSE if not a vectorizable STMT, TRUE otherwise. */
5346 vectorizable_induction (gimple phi
, gimple_stmt_iterator
*gsi ATTRIBUTE_UNUSED
,
5349 stmt_vec_info stmt_info
= vinfo_for_stmt (phi
);
5350 tree vectype
= STMT_VINFO_VECTYPE (stmt_info
);
5351 loop_vec_info loop_vinfo
= STMT_VINFO_LOOP_VINFO (stmt_info
);
5352 struct loop
*loop
= LOOP_VINFO_LOOP (loop_vinfo
);
5353 int nunits
= TYPE_VECTOR_SUBPARTS (vectype
);
5354 int ncopies
= LOOP_VINFO_VECT_FACTOR (loop_vinfo
) / nunits
;
5357 gcc_assert (ncopies
>= 1);
5358 /* FORNOW. These restrictions should be relaxed. */
5359 if (nested_in_vect_loop_p (loop
, phi
))
5361 imm_use_iterator imm_iter
;
5362 use_operand_p use_p
;
5369 if (dump_enabled_p ())
5370 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
5371 "multiple types in nested loop.\n");
5376 latch_e
= loop_latch_edge (loop
->inner
);
5377 loop_arg
= PHI_ARG_DEF_FROM_EDGE (phi
, latch_e
);
5378 FOR_EACH_IMM_USE_FAST (use_p
, imm_iter
, loop_arg
)
5380 if (!flow_bb_inside_loop_p (loop
->inner
,
5381 gimple_bb (USE_STMT (use_p
))))
5383 exit_phi
= USE_STMT (use_p
);
5389 stmt_vec_info exit_phi_vinfo
= vinfo_for_stmt (exit_phi
);
5390 if (!(STMT_VINFO_RELEVANT_P (exit_phi_vinfo
)
5391 && !STMT_VINFO_LIVE_P (exit_phi_vinfo
)))
5393 if (dump_enabled_p ())
5394 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
5395 "inner-loop induction only used outside "
5396 "of the outer vectorized loop.\n");
5402 if (!STMT_VINFO_RELEVANT_P (stmt_info
))
5405 /* FORNOW: SLP not supported. */
5406 if (STMT_SLP_TYPE (stmt_info
))
5409 gcc_assert (STMT_VINFO_DEF_TYPE (stmt_info
) == vect_induction_def
);
5411 if (gimple_code (phi
) != GIMPLE_PHI
)
5414 if (!vec_stmt
) /* transformation not required. */
5416 STMT_VINFO_TYPE (stmt_info
) = induc_vec_info_type
;
5417 if (dump_enabled_p ())
5418 dump_printf_loc (MSG_NOTE
, vect_location
,
5419 "=== vectorizable_induction ===\n");
5420 vect_model_induction_cost (stmt_info
, ncopies
);
5426 if (dump_enabled_p ())
5427 dump_printf_loc (MSG_NOTE
, vect_location
, "transform induction phi.\n");
5429 vec_def
= get_initial_def_for_induction (phi
);
5430 *vec_stmt
= SSA_NAME_DEF_STMT (vec_def
);
5434 /* Function vectorizable_live_operation.
5436 STMT computes a value that is used outside the loop. Check if
5437 it can be supported. */
5440 vectorizable_live_operation (gimple stmt
,
5441 gimple_stmt_iterator
*gsi ATTRIBUTE_UNUSED
,
5444 stmt_vec_info stmt_info
= vinfo_for_stmt (stmt
);
5445 loop_vec_info loop_vinfo
= STMT_VINFO_LOOP_VINFO (stmt_info
);
5446 struct loop
*loop
= LOOP_VINFO_LOOP (loop_vinfo
);
5452 enum vect_def_type dt
;
5453 enum tree_code code
;
5454 enum gimple_rhs_class rhs_class
;
5456 gcc_assert (STMT_VINFO_LIVE_P (stmt_info
));
5458 if (STMT_VINFO_DEF_TYPE (stmt_info
) == vect_reduction_def
)
5461 if (!is_gimple_assign (stmt
))
5463 if (gimple_call_internal_p (stmt
)
5464 && gimple_call_internal_fn (stmt
) == IFN_GOMP_SIMD_LANE
5465 && gimple_call_lhs (stmt
)
5467 && TREE_CODE (gimple_call_arg (stmt
, 0)) == SSA_NAME
5469 == SSA_NAME_VAR (gimple_call_arg (stmt
, 0)))
5471 edge e
= single_exit (loop
);
5472 basic_block merge_bb
= e
->dest
;
5473 imm_use_iterator imm_iter
;
5474 use_operand_p use_p
;
5475 tree lhs
= gimple_call_lhs (stmt
);
5477 FOR_EACH_IMM_USE_FAST (use_p
, imm_iter
, lhs
)
5479 gimple use_stmt
= USE_STMT (use_p
);
5480 if (gimple_code (use_stmt
) == GIMPLE_PHI
5481 || gimple_bb (use_stmt
) == merge_bb
)
5486 = build_int_cst (unsigned_type_node
,
5487 loop_vinfo
->vectorization_factor
- 1);
5488 SET_PHI_ARG_DEF (use_stmt
, e
->dest_idx
, vfm1
);
5498 if (TREE_CODE (gimple_assign_lhs (stmt
)) != SSA_NAME
)
5501 /* FORNOW. CHECKME. */
5502 if (nested_in_vect_loop_p (loop
, stmt
))
5505 code
= gimple_assign_rhs_code (stmt
);
5506 op_type
= TREE_CODE_LENGTH (code
);
5507 rhs_class
= get_gimple_rhs_class (code
);
5508 gcc_assert (rhs_class
!= GIMPLE_UNARY_RHS
|| op_type
== unary_op
);
5509 gcc_assert (rhs_class
!= GIMPLE_BINARY_RHS
|| op_type
== binary_op
);
5511 /* FORNOW: support only if all uses are invariant. This means
5512 that the scalar operations can remain in place, unvectorized.
5513 The original last scalar value that they compute will be used. */
5515 for (i
= 0; i
< op_type
; i
++)
5517 if (rhs_class
== GIMPLE_SINGLE_RHS
)
5518 op
= TREE_OPERAND (gimple_op (stmt
, 1), i
);
5520 op
= gimple_op (stmt
, i
+ 1);
5522 && !vect_is_simple_use (op
, stmt
, loop_vinfo
, NULL
, &def_stmt
, &def
,
5525 if (dump_enabled_p ())
5526 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
5527 "use not simple.\n");
5531 if (dt
!= vect_external_def
&& dt
!= vect_constant_def
)
5535 /* No transformation is required for the cases we currently support. */
5539 /* Kill any debug uses outside LOOP of SSA names defined in STMT. */
5542 vect_loop_kill_debug_uses (struct loop
*loop
, gimple stmt
)
5544 ssa_op_iter op_iter
;
5545 imm_use_iterator imm_iter
;
5546 def_operand_p def_p
;
5549 FOR_EACH_PHI_OR_STMT_DEF (def_p
, stmt
, op_iter
, SSA_OP_DEF
)
5551 FOR_EACH_IMM_USE_STMT (ustmt
, imm_iter
, DEF_FROM_PTR (def_p
))
5555 if (!is_gimple_debug (ustmt
))
5558 bb
= gimple_bb (ustmt
);
5560 if (!flow_bb_inside_loop_p (loop
, bb
))
5562 if (gimple_debug_bind_p (ustmt
))
5564 if (dump_enabled_p ())
5565 dump_printf_loc (MSG_NOTE
, vect_location
,
5566 "killing debug use\n");
5568 gimple_debug_bind_reset_value (ustmt
);
5569 update_stmt (ustmt
);
5578 /* Function vect_transform_loop.
5580 The analysis phase has determined that the loop is vectorizable.
5581 Vectorize the loop - created vectorized stmts to replace the scalar
5582 stmts in the loop, and update the loop exit condition. */
5585 vect_transform_loop (loop_vec_info loop_vinfo
)
5587 struct loop
*loop
= LOOP_VINFO_LOOP (loop_vinfo
);
5588 basic_block
*bbs
= LOOP_VINFO_BBS (loop_vinfo
);
5589 int nbbs
= loop
->num_nodes
;
5590 gimple_stmt_iterator si
;
5593 int vectorization_factor
= LOOP_VINFO_VECT_FACTOR (loop_vinfo
);
5595 bool slp_scheduled
= false;
5596 unsigned int nunits
;
5597 gimple stmt
, pattern_stmt
;
5598 gimple_seq pattern_def_seq
= NULL
;
5599 gimple_stmt_iterator pattern_def_si
= gsi_none ();
5600 bool transform_pattern_stmt
= false;
5601 bool check_profitability
= false;
5603 /* Record number of iterations before we started tampering with the profile. */
5604 gcov_type expected_iterations
= expected_loop_iterations_unbounded (loop
);
5606 if (dump_enabled_p ())
5607 dump_printf_loc (MSG_NOTE
, vect_location
, "=== vec_transform_loop ===\n");
5609 /* If profile is inprecise, we have chance to fix it up. */
5610 if (LOOP_VINFO_NITERS_KNOWN_P (loop_vinfo
))
5611 expected_iterations
= LOOP_VINFO_INT_NITERS (loop_vinfo
);
5613 /* Use the more conservative vectorization threshold. If the number
5614 of iterations is constant assume the cost check has been performed
5615 by our caller. If the threshold makes all loops profitable that
5616 run at least the vectorization factor number of times checking
5617 is pointless, too. */
5618 th
= ((PARAM_VALUE (PARAM_MIN_VECT_LOOP_BOUND
)
5619 * LOOP_VINFO_VECT_FACTOR (loop_vinfo
)) - 1);
5620 th
= MAX (th
, LOOP_VINFO_COST_MODEL_MIN_ITERS (loop_vinfo
));
5621 if (th
>= LOOP_VINFO_VECT_FACTOR (loop_vinfo
) - 1
5622 && !LOOP_VINFO_NITERS_KNOWN_P (loop_vinfo
))
5624 if (dump_enabled_p ())
5625 dump_printf_loc (MSG_NOTE
, vect_location
,
5626 "Profitability threshold is %d loop iterations.\n",
5628 check_profitability
= true;
5631 /* Version the loop first, if required, so the profitability check
5634 if (LOOP_REQUIRES_VERSIONING_FOR_ALIGNMENT (loop_vinfo
)
5635 || LOOP_REQUIRES_VERSIONING_FOR_ALIAS (loop_vinfo
))
5637 vect_loop_versioning (loop_vinfo
, th
, check_profitability
);
5638 check_profitability
= false;
5641 /* Peel the loop if there are data refs with unknown alignment.
5642 Only one data ref with unknown store is allowed. */
5644 if (LOOP_PEELING_FOR_ALIGNMENT (loop_vinfo
))
5646 vect_do_peeling_for_alignment (loop_vinfo
, th
, check_profitability
);
5647 check_profitability
= false;
5650 /* If the loop has a symbolic number of iterations 'n' (i.e. it's not a
5651 compile time constant), or it is a constant that doesn't divide by the
5652 vectorization factor, then an epilog loop needs to be created.
5653 We therefore duplicate the loop: the original loop will be vectorized,
5654 and will compute the first (n/VF) iterations. The second copy of the loop
5655 will remain scalar and will compute the remaining (n%VF) iterations.
5656 (VF is the vectorization factor). */
5658 if ((int) tree_ctz (LOOP_VINFO_NITERS (loop_vinfo
))
5659 < exact_log2 (vectorization_factor
)
5660 || LOOP_VINFO_PEELING_FOR_GAPS (loop_vinfo
))
5661 vect_do_peeling_for_loop_bound (loop_vinfo
, &ratio
,
5662 th
, check_profitability
);
5663 else if (LOOP_VINFO_NITERS_KNOWN_P (loop_vinfo
))
5664 ratio
= build_int_cst (TREE_TYPE (LOOP_VINFO_NITERS (loop_vinfo
)),
5665 LOOP_VINFO_INT_NITERS (loop_vinfo
) / vectorization_factor
);
5668 tree ni_name
, ratio_mult_vf
;
5669 vect_generate_tmps_on_preheader (loop_vinfo
, &ni_name
, &ratio_mult_vf
,
5673 /* 1) Make sure the loop header has exactly two entries
5674 2) Make sure we have a preheader basic block. */
5676 gcc_assert (EDGE_COUNT (loop
->header
->preds
) == 2);
5678 split_edge (loop_preheader_edge (loop
));
5680 /* FORNOW: the vectorizer supports only loops which body consist
5681 of one basic block (header + empty latch). When the vectorizer will
5682 support more involved loop forms, the order by which the BBs are
5683 traversed need to be reconsidered. */
5685 for (i
= 0; i
< nbbs
; i
++)
5687 basic_block bb
= bbs
[i
];
5688 stmt_vec_info stmt_info
;
5691 for (si
= gsi_start_phis (bb
); !gsi_end_p (si
); gsi_next (&si
))
5693 phi
= gsi_stmt (si
);
5694 if (dump_enabled_p ())
5696 dump_printf_loc (MSG_NOTE
, vect_location
,
5697 "------>vectorizing phi: ");
5698 dump_gimple_stmt (MSG_NOTE
, TDF_SLIM
, phi
, 0);
5699 dump_printf (MSG_NOTE
, "\n");
5701 stmt_info
= vinfo_for_stmt (phi
);
5705 if (MAY_HAVE_DEBUG_STMTS
&& !STMT_VINFO_LIVE_P (stmt_info
))
5706 vect_loop_kill_debug_uses (loop
, phi
);
5708 if (!STMT_VINFO_RELEVANT_P (stmt_info
)
5709 && !STMT_VINFO_LIVE_P (stmt_info
))
5712 if ((TYPE_VECTOR_SUBPARTS (STMT_VINFO_VECTYPE (stmt_info
))
5713 != (unsigned HOST_WIDE_INT
) vectorization_factor
)
5714 && dump_enabled_p ())
5715 dump_printf_loc (MSG_NOTE
, vect_location
, "multiple-types.\n");
5717 if (STMT_VINFO_DEF_TYPE (stmt_info
) == vect_induction_def
)
5719 if (dump_enabled_p ())
5720 dump_printf_loc (MSG_NOTE
, vect_location
, "transform phi.\n");
5721 vect_transform_stmt (phi
, NULL
, NULL
, NULL
, NULL
);
5725 pattern_stmt
= NULL
;
5726 for (si
= gsi_start_bb (bb
); !gsi_end_p (si
) || transform_pattern_stmt
;)
5730 if (transform_pattern_stmt
)
5731 stmt
= pattern_stmt
;
5734 stmt
= gsi_stmt (si
);
5735 /* During vectorization remove existing clobber stmts. */
5736 if (gimple_clobber_p (stmt
))
5738 unlink_stmt_vdef (stmt
);
5739 gsi_remove (&si
, true);
5740 release_defs (stmt
);
5745 if (dump_enabled_p ())
5747 dump_printf_loc (MSG_NOTE
, vect_location
,
5748 "------>vectorizing statement: ");
5749 dump_gimple_stmt (MSG_NOTE
, TDF_SLIM
, stmt
, 0);
5750 dump_printf (MSG_NOTE
, "\n");
5753 stmt_info
= vinfo_for_stmt (stmt
);
5755 /* vector stmts created in the outer-loop during vectorization of
5756 stmts in an inner-loop may not have a stmt_info, and do not
5757 need to be vectorized. */
5764 if (MAY_HAVE_DEBUG_STMTS
&& !STMT_VINFO_LIVE_P (stmt_info
))
5765 vect_loop_kill_debug_uses (loop
, stmt
);
5767 if (!STMT_VINFO_RELEVANT_P (stmt_info
)
5768 && !STMT_VINFO_LIVE_P (stmt_info
))
5770 if (STMT_VINFO_IN_PATTERN_P (stmt_info
)
5771 && (pattern_stmt
= STMT_VINFO_RELATED_STMT (stmt_info
))
5772 && (STMT_VINFO_RELEVANT_P (vinfo_for_stmt (pattern_stmt
))
5773 || STMT_VINFO_LIVE_P (vinfo_for_stmt (pattern_stmt
))))
5775 stmt
= pattern_stmt
;
5776 stmt_info
= vinfo_for_stmt (stmt
);
5784 else if (STMT_VINFO_IN_PATTERN_P (stmt_info
)
5785 && (pattern_stmt
= STMT_VINFO_RELATED_STMT (stmt_info
))
5786 && (STMT_VINFO_RELEVANT_P (vinfo_for_stmt (pattern_stmt
))
5787 || STMT_VINFO_LIVE_P (vinfo_for_stmt (pattern_stmt
))))
5788 transform_pattern_stmt
= true;
5790 /* If pattern statement has def stmts, vectorize them too. */
5791 if (is_pattern_stmt_p (stmt_info
))
5793 if (pattern_def_seq
== NULL
)
5795 pattern_def_seq
= STMT_VINFO_PATTERN_DEF_SEQ (stmt_info
);
5796 pattern_def_si
= gsi_start (pattern_def_seq
);
5798 else if (!gsi_end_p (pattern_def_si
))
5799 gsi_next (&pattern_def_si
);
5800 if (pattern_def_seq
!= NULL
)
5802 gimple pattern_def_stmt
= NULL
;
5803 stmt_vec_info pattern_def_stmt_info
= NULL
;
5805 while (!gsi_end_p (pattern_def_si
))
5807 pattern_def_stmt
= gsi_stmt (pattern_def_si
);
5808 pattern_def_stmt_info
5809 = vinfo_for_stmt (pattern_def_stmt
);
5810 if (STMT_VINFO_RELEVANT_P (pattern_def_stmt_info
)
5811 || STMT_VINFO_LIVE_P (pattern_def_stmt_info
))
5813 gsi_next (&pattern_def_si
);
5816 if (!gsi_end_p (pattern_def_si
))
5818 if (dump_enabled_p ())
5820 dump_printf_loc (MSG_NOTE
, vect_location
,
5821 "==> vectorizing pattern def "
5823 dump_gimple_stmt (MSG_NOTE
, TDF_SLIM
,
5824 pattern_def_stmt
, 0);
5825 dump_printf (MSG_NOTE
, "\n");
5828 stmt
= pattern_def_stmt
;
5829 stmt_info
= pattern_def_stmt_info
;
5833 pattern_def_si
= gsi_none ();
5834 transform_pattern_stmt
= false;
5838 transform_pattern_stmt
= false;
5841 gcc_assert (STMT_VINFO_VECTYPE (stmt_info
));
5842 nunits
= (unsigned int) TYPE_VECTOR_SUBPARTS (
5843 STMT_VINFO_VECTYPE (stmt_info
));
5844 if (!STMT_SLP_TYPE (stmt_info
)
5845 && nunits
!= (unsigned int) vectorization_factor
5846 && dump_enabled_p ())
5847 /* For SLP VF is set according to unrolling factor, and not to
5848 vector size, hence for SLP this print is not valid. */
5849 dump_printf_loc (MSG_NOTE
, vect_location
,
5850 "multiple-types.\n");
5852 /* SLP. Schedule all the SLP instances when the first SLP stmt is
5854 if (STMT_SLP_TYPE (stmt_info
))
5858 slp_scheduled
= true;
5860 if (dump_enabled_p ())
5861 dump_printf_loc (MSG_NOTE
, vect_location
,
5862 "=== scheduling SLP instances ===\n");
5864 vect_schedule_slp (loop_vinfo
, NULL
);
5867 /* Hybrid SLP stmts must be vectorized in addition to SLP. */
5868 if (!vinfo_for_stmt (stmt
) || PURE_SLP_STMT (stmt_info
))
5870 if (!transform_pattern_stmt
&& gsi_end_p (pattern_def_si
))
5872 pattern_def_seq
= NULL
;
5879 /* -------- vectorize statement ------------ */
5880 if (dump_enabled_p ())
5881 dump_printf_loc (MSG_NOTE
, vect_location
, "transform statement.\n");
5883 grouped_store
= false;
5884 is_store
= vect_transform_stmt (stmt
, &si
, &grouped_store
, NULL
, NULL
);
5887 if (STMT_VINFO_GROUPED_ACCESS (stmt_info
))
5889 /* Interleaving. If IS_STORE is TRUE, the vectorization of the
5890 interleaving chain was completed - free all the stores in
5893 vect_remove_stores (GROUP_FIRST_ELEMENT (stmt_info
));
5898 /* Free the attached stmt_vec_info and remove the stmt. */
5899 gimple store
= gsi_stmt (si
);
5900 free_stmt_vec_info (store
);
5901 unlink_stmt_vdef (store
);
5902 gsi_remove (&si
, true);
5903 release_defs (store
);
5908 if (!transform_pattern_stmt
&& gsi_end_p (pattern_def_si
))
5910 pattern_def_seq
= NULL
;
5916 slpeel_make_loop_iterate_ntimes (loop
, ratio
);
5918 /* Reduce loop iterations by the vectorization factor. */
5919 scale_loop_profile (loop
, GCOV_COMPUTE_SCALE (1, vectorization_factor
),
5920 expected_iterations
/ vectorization_factor
);
5921 loop
->nb_iterations_upper_bound
5922 = loop
->nb_iterations_upper_bound
.udiv (double_int::from_uhwi (vectorization_factor
),
5924 if (LOOP_VINFO_PEELING_FOR_GAPS (loop_vinfo
)
5925 && loop
->nb_iterations_upper_bound
!= double_int_zero
)
5926 loop
->nb_iterations_upper_bound
= loop
->nb_iterations_upper_bound
- double_int_one
;
5927 if (loop
->any_estimate
)
5929 loop
->nb_iterations_estimate
5930 = loop
->nb_iterations_estimate
.udiv (double_int::from_uhwi (vectorization_factor
),
5932 if (LOOP_VINFO_PEELING_FOR_GAPS (loop_vinfo
)
5933 && loop
->nb_iterations_estimate
!= double_int_zero
)
5934 loop
->nb_iterations_estimate
= loop
->nb_iterations_estimate
- double_int_one
;
5937 if (dump_enabled_p ())
5939 dump_printf_loc (MSG_NOTE
, vect_location
,
5940 "LOOP VECTORIZED\n");
5942 dump_printf_loc (MSG_NOTE
, vect_location
,
5943 "OUTER LOOP VECTORIZED\n");
5944 dump_printf (MSG_NOTE
, "\n");