2 Copyright (C) 2003-2015 Free Software Foundation, Inc.
3 Contributed by Dorit Naishlos <dorit@il.ibm.com>
5 This file is part of GCC.
7 GCC is free software; you can redistribute it and/or modify it under
8 the terms of the GNU General Public License as published by the Free
9 Software Foundation; either version 3, or (at your option) any later
12 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
13 WARRANTY; without even the implied warranty of MERCHANTABILITY or
14 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
17 You should have received a copy of the GNU General Public License
18 along with GCC; see the file COPYING3. If not see
19 <http://www.gnu.org/licenses/>. */
21 /* Loop and basic block vectorizer.
23 This file contains drivers for the three vectorizers:
24 (1) loop vectorizer (inter-iteration parallelism),
25 (2) loop-aware SLP (intra-iteration parallelism) (invoked by the loop
27 (3) BB vectorizer (out-of-loops), aka SLP
29 The rest of the vectorizer's code is organized as follows:
30 - tree-vect-loop.c - loop specific parts such as reductions, etc. These are
31 used by drivers (1) and (2).
32 - tree-vect-loop-manip.c - vectorizer's loop control-flow utilities, used by
34 - tree-vect-slp.c - BB vectorization specific analysis and transformation,
35 used by drivers (2) and (3).
36 - tree-vect-stmts.c - statements analysis and transformation (used by all).
37 - tree-vect-data-refs.c - vectorizer specific data-refs analysis and
38 manipulations (used by all).
39 - tree-vect-patterns.c - vectorizable code patterns detector (used by all)
41 Here's a poor attempt at illustrating that:
44 loop_vect() loop_aware_slp() slp_vect()
47 tree-vect-loop.c tree-vect-slp.c
52 tree-vect-stmts.c tree-vect-data-refs.c
59 #include "coretypes.h"
65 #include "double-int.h"
72 #include "fold-const.h"
73 #include "stor-layout.h"
74 #include "tree-pretty-print.h"
76 #include "hard-reg-set.h"
79 #include "dominance.h"
81 #include "basic-block.h"
82 #include "tree-ssa-alias.h"
83 #include "internal-fn.h"
84 #include "gimple-expr.h"
87 #include "gimple-iterator.h"
88 #include "gimple-walk.h"
89 #include "gimple-ssa.h"
91 #include "plugin-api.h"
94 #include "tree-phinodes.h"
95 #include "ssa-iterators.h"
96 #include "tree-ssa-loop-manip.h"
99 #include "tree-vectorizer.h"
100 #include "tree-pass.h"
101 #include "tree-ssa-propagate.h"
103 #include "gimple-fold.h"
104 #include "tree-scalar-evolution.h"
107 /* Loop or bb location. */
108 source_location vect_location
;
110 /* Vector mapping GIMPLE stmt to stmt_vec_info. */
111 vec
<vec_void_p
> stmt_vec_info_vec
;
113 /* For mapping simduid to vectorization factor. */
115 struct simduid_to_vf
: typed_free_remove
<simduid_to_vf
>
117 unsigned int simduid
;
120 /* hash_table support. */
121 typedef simduid_to_vf value_type
;
122 typedef simduid_to_vf compare_type
;
123 static inline hashval_t
hash (const value_type
*);
124 static inline int equal (const value_type
*, const compare_type
*);
128 simduid_to_vf::hash (const value_type
*p
)
134 simduid_to_vf::equal (const value_type
*p1
, const value_type
*p2
)
136 return p1
->simduid
== p2
->simduid
;
139 /* This hash maps the OMP simd array to the corresponding simduid used
140 to index into it. Like thus,
142 _7 = GOMP_SIMD_LANE (simduid.0)
148 This hash maps from the OMP simd array (D.1737[]) to DECL_UID of
151 struct simd_array_to_simduid
: typed_free_remove
<simd_array_to_simduid
>
154 unsigned int simduid
;
156 /* hash_table support. */
157 typedef simd_array_to_simduid value_type
;
158 typedef simd_array_to_simduid compare_type
;
159 static inline hashval_t
hash (const value_type
*);
160 static inline int equal (const value_type
*, const compare_type
*);
164 simd_array_to_simduid::hash (const value_type
*p
)
166 return DECL_UID (p
->decl
);
170 simd_array_to_simduid::equal (const value_type
*p1
, const value_type
*p2
)
172 return p1
->decl
== p2
->decl
;
175 /* Fold IFN_GOMP_SIMD_LANE, IFN_GOMP_SIMD_VF and IFN_GOMP_SIMD_LAST_LANE
176 into their corresponding constants. */
179 adjust_simduid_builtins (hash_table
<simduid_to_vf
> **htab
)
183 FOR_EACH_BB_FN (bb
, cfun
)
185 gimple_stmt_iterator i
;
187 for (i
= gsi_start_bb (bb
); !gsi_end_p (i
); gsi_next (&i
))
190 enum internal_fn ifn
;
191 gimple stmt
= gsi_stmt (i
);
193 if (!is_gimple_call (stmt
)
194 || !gimple_call_internal_p (stmt
))
196 ifn
= gimple_call_internal_fn (stmt
);
199 case IFN_GOMP_SIMD_LANE
:
200 case IFN_GOMP_SIMD_VF
:
201 case IFN_GOMP_SIMD_LAST_LANE
:
206 tree arg
= gimple_call_arg (stmt
, 0);
207 gcc_assert (arg
!= NULL_TREE
);
208 gcc_assert (TREE_CODE (arg
) == SSA_NAME
);
209 simduid_to_vf
*p
= NULL
, data
;
210 data
.simduid
= DECL_UID (SSA_NAME_VAR (arg
));
212 p
= (*htab
)->find (&data
);
217 case IFN_GOMP_SIMD_VF
:
218 t
= build_int_cst (unsigned_type_node
, vf
);
220 case IFN_GOMP_SIMD_LANE
:
221 t
= build_int_cst (unsigned_type_node
, 0);
223 case IFN_GOMP_SIMD_LAST_LANE
:
224 t
= gimple_call_arg (stmt
, 1);
229 update_call_from_tree (&i
, t
);
234 /* Helper structure for note_simd_array_uses. */
236 struct note_simd_array_uses_struct
238 hash_table
<simd_array_to_simduid
> **htab
;
239 unsigned int simduid
;
242 /* Callback for note_simd_array_uses, called through walk_gimple_op. */
245 note_simd_array_uses_cb (tree
*tp
, int *walk_subtrees
, void *data
)
247 struct walk_stmt_info
*wi
= (struct walk_stmt_info
*) data
;
248 struct note_simd_array_uses_struct
*ns
249 = (struct note_simd_array_uses_struct
*) wi
->info
;
254 && lookup_attribute ("omp simd array", DECL_ATTRIBUTES (*tp
))
255 && DECL_CONTEXT (*tp
) == current_function_decl
)
257 simd_array_to_simduid data
;
259 *ns
->htab
= new hash_table
<simd_array_to_simduid
> (15);
261 data
.simduid
= ns
->simduid
;
262 simd_array_to_simduid
**slot
= (*ns
->htab
)->find_slot (&data
, INSERT
);
265 simd_array_to_simduid
*p
= XNEW (simd_array_to_simduid
);
269 else if ((*slot
)->simduid
!= ns
->simduid
)
270 (*slot
)->simduid
= -1U;
276 /* Find "omp simd array" temporaries and map them to corresponding
280 note_simd_array_uses (hash_table
<simd_array_to_simduid
> **htab
)
283 gimple_stmt_iterator gsi
;
284 struct walk_stmt_info wi
;
285 struct note_simd_array_uses_struct ns
;
287 memset (&wi
, 0, sizeof (wi
));
291 FOR_EACH_BB_FN (bb
, cfun
)
292 for (gsi
= gsi_start_bb (bb
); !gsi_end_p (gsi
); gsi_next (&gsi
))
294 gimple stmt
= gsi_stmt (gsi
);
295 if (!is_gimple_call (stmt
) || !gimple_call_internal_p (stmt
))
297 switch (gimple_call_internal_fn (stmt
))
299 case IFN_GOMP_SIMD_LANE
:
300 case IFN_GOMP_SIMD_VF
:
301 case IFN_GOMP_SIMD_LAST_LANE
:
306 tree lhs
= gimple_call_lhs (stmt
);
307 if (lhs
== NULL_TREE
)
309 imm_use_iterator use_iter
;
311 ns
.simduid
= DECL_UID (SSA_NAME_VAR (gimple_call_arg (stmt
, 0)));
312 FOR_EACH_IMM_USE_STMT (use_stmt
, use_iter
, lhs
)
313 if (!is_gimple_debug (use_stmt
))
314 walk_gimple_op (use_stmt
, note_simd_array_uses_cb
, &wi
);
318 /* A helper function to free data refs. */
321 vect_destroy_datarefs (loop_vec_info loop_vinfo
, bb_vec_info bb_vinfo
)
323 vec
<data_reference_p
> datarefs
;
324 struct data_reference
*dr
;
328 datarefs
= LOOP_VINFO_DATAREFS (loop_vinfo
);
330 datarefs
= BB_VINFO_DATAREFS (bb_vinfo
);
332 FOR_EACH_VEC_ELT (datarefs
, i
, dr
)
339 free_data_refs (datarefs
);
343 /* If LOOP has been versioned during ifcvt, return the internal call
347 vect_loop_vectorized_call (struct loop
*loop
)
349 basic_block bb
= loop_preheader_edge (loop
)->src
;
356 if (!single_pred_p (bb
))
358 bb
= single_pred (bb
);
361 if (g
&& gimple_code (g
) == GIMPLE_COND
)
363 gimple_stmt_iterator gsi
= gsi_for_stmt (g
);
365 if (!gsi_end_p (gsi
))
368 if (is_gimple_call (g
)
369 && gimple_call_internal_p (g
)
370 && gimple_call_internal_fn (g
) == IFN_LOOP_VECTORIZED
371 && (tree_to_shwi (gimple_call_arg (g
, 0)) == loop
->num
372 || tree_to_shwi (gimple_call_arg (g
, 1)) == loop
->num
))
379 /* Fold LOOP_VECTORIZED internal call G to VALUE and
380 update any immediate uses of it's LHS. */
383 fold_loop_vectorized_call (gimple g
, tree value
)
385 tree lhs
= gimple_call_lhs (g
);
387 imm_use_iterator iter
;
389 gimple_stmt_iterator gsi
= gsi_for_stmt (g
);
391 update_call_from_tree (&gsi
, value
);
392 FOR_EACH_IMM_USE_STMT (use_stmt
, iter
, lhs
)
394 FOR_EACH_IMM_USE_ON_STMT (use_p
, iter
)
395 SET_USE (use_p
, value
);
396 update_stmt (use_stmt
);
400 /* Function vectorize_loops.
402 Entry point to loop vectorization phase. */
405 vectorize_loops (void)
408 unsigned int num_vectorized_loops
= 0;
409 unsigned int vect_loops_num
;
411 hash_table
<simduid_to_vf
> *simduid_to_vf_htab
= NULL
;
412 hash_table
<simd_array_to_simduid
> *simd_array_to_simduid_htab
= NULL
;
413 bool any_ifcvt_loops
= false;
416 vect_loops_num
= number_of_loops (cfun
);
418 /* Bail out if there are no loops. */
419 if (vect_loops_num
<= 1)
421 if (cfun
->has_simduid_loops
)
422 adjust_simduid_builtins (&simduid_to_vf_htab
);
426 if (cfun
->has_simduid_loops
)
427 note_simd_array_uses (&simd_array_to_simduid_htab
);
429 init_stmt_vec_info_vec ();
431 /* ----------- Analyze loops. ----------- */
433 /* If some loop was duplicated, it gets bigger number
434 than all previously defined loops. This fact allows us to run
435 only over initial loops skipping newly generated ones. */
436 FOR_EACH_LOOP (loop
, 0)
437 if (loop
->dont_vectorize
)
438 any_ifcvt_loops
= true;
439 else if ((flag_tree_loop_vectorize
440 && optimize_loop_nest_for_speed_p (loop
))
441 || loop
->force_vectorize
)
443 loop_vec_info loop_vinfo
;
444 vect_location
= find_loop_location (loop
);
445 if (LOCATION_LOCUS (vect_location
) != UNKNOWN_LOCATION
446 && dump_enabled_p ())
447 dump_printf (MSG_NOTE
, "\nAnalyzing loop at %s:%d\n",
448 LOCATION_FILE (vect_location
),
449 LOCATION_LINE (vect_location
));
451 loop_vinfo
= vect_analyze_loop (loop
);
452 loop
->aux
= loop_vinfo
;
454 if (!loop_vinfo
|| !LOOP_VINFO_VECTORIZABLE_P (loop_vinfo
))
457 if (!dbg_cnt (vect_loop
))
460 gimple loop_vectorized_call
= vect_loop_vectorized_call (loop
);
461 if (loop_vectorized_call
)
463 tree arg
= gimple_call_arg (loop_vectorized_call
, 1);
466 struct loop
*scalar_loop
= get_loop (cfun
, tree_to_shwi (arg
));
468 LOOP_VINFO_SCALAR_LOOP (loop_vinfo
) = scalar_loop
;
469 gcc_checking_assert (vect_loop_vectorized_call
470 (LOOP_VINFO_SCALAR_LOOP (loop_vinfo
))
471 == loop_vectorized_call
);
472 bbs
= get_loop_body (scalar_loop
);
473 for (i
= 0; i
< scalar_loop
->num_nodes
; i
++)
475 basic_block bb
= bbs
[i
];
476 gimple_stmt_iterator gsi
;
477 for (gsi
= gsi_start_phis (bb
); !gsi_end_p (gsi
);
480 gimple phi
= gsi_stmt (gsi
);
481 gimple_set_uid (phi
, 0);
483 for (gsi
= gsi_start_bb (bb
); !gsi_end_p (gsi
);
486 gimple stmt
= gsi_stmt (gsi
);
487 gimple_set_uid (stmt
, 0);
493 if (LOCATION_LOCUS (vect_location
) != UNKNOWN_LOCATION
494 && dump_enabled_p ())
495 dump_printf_loc (MSG_OPTIMIZED_LOCATIONS
, vect_location
,
496 "loop vectorized\n");
497 vect_transform_loop (loop_vinfo
);
498 num_vectorized_loops
++;
499 /* Now that the loop has been vectorized, allow it to be unrolled
501 loop
->force_vectorize
= false;
505 simduid_to_vf
*simduid_to_vf_data
= XNEW (simduid_to_vf
);
506 if (!simduid_to_vf_htab
)
507 simduid_to_vf_htab
= new hash_table
<simduid_to_vf
> (15);
508 simduid_to_vf_data
->simduid
= DECL_UID (loop
->simduid
);
509 simduid_to_vf_data
->vf
= loop_vinfo
->vectorization_factor
;
510 *simduid_to_vf_htab
->find_slot (simduid_to_vf_data
, INSERT
)
511 = simduid_to_vf_data
;
514 if (loop_vectorized_call
)
516 fold_loop_vectorized_call (loop_vectorized_call
, boolean_true_node
);
517 ret
|= TODO_cleanup_cfg
;
521 vect_location
= UNKNOWN_LOCATION
;
523 statistics_counter_event (cfun
, "Vectorized loops", num_vectorized_loops
);
524 if (dump_enabled_p ()
525 || (num_vectorized_loops
> 0 && dump_enabled_p ()))
526 dump_printf_loc (MSG_NOTE
, vect_location
,
527 "vectorized %u loops in function.\n",
528 num_vectorized_loops
);
530 /* ----------- Finalize. ----------- */
533 for (i
= 1; i
< vect_loops_num
; i
++)
535 loop
= get_loop (cfun
, i
);
536 if (loop
&& loop
->dont_vectorize
)
538 gimple g
= vect_loop_vectorized_call (loop
);
541 fold_loop_vectorized_call (g
, boolean_false_node
);
542 ret
|= TODO_cleanup_cfg
;
547 for (i
= 1; i
< vect_loops_num
; i
++)
549 loop_vec_info loop_vinfo
;
551 loop
= get_loop (cfun
, i
);
554 loop_vinfo
= (loop_vec_info
) loop
->aux
;
555 destroy_loop_vec_info (loop_vinfo
, true);
559 free_stmt_vec_info_vec ();
561 /* Fold IFN_GOMP_SIMD_{VF,LANE,LAST_LANE} builtins. */
562 if (cfun
->has_simduid_loops
)
563 adjust_simduid_builtins (&simduid_to_vf_htab
);
565 /* Shrink any "omp array simd" temporary arrays to the
566 actual vectorization factors. */
567 if (simd_array_to_simduid_htab
)
569 for (hash_table
<simd_array_to_simduid
>::iterator iter
570 = simd_array_to_simduid_htab
->begin ();
571 iter
!= simd_array_to_simduid_htab
->end (); ++iter
)
572 if ((*iter
)->simduid
!= -1U)
574 tree decl
= (*iter
)->decl
;
576 if (simduid_to_vf_htab
)
578 simduid_to_vf
*p
= NULL
, data
;
579 data
.simduid
= (*iter
)->simduid
;
580 p
= simduid_to_vf_htab
->find (&data
);
585 = build_array_type_nelts (TREE_TYPE (TREE_TYPE (decl
)), vf
);
586 TREE_TYPE (decl
) = atype
;
587 relayout_decl (decl
);
590 delete simd_array_to_simduid_htab
;
592 delete simduid_to_vf_htab
;
593 simduid_to_vf_htab
= NULL
;
595 if (num_vectorized_loops
> 0)
597 /* If we vectorized any loop only virtual SSA form needs to be updated.
598 ??? Also while we try hard to update loop-closed SSA form we fail
599 to properly do this in some corner-cases (see PR56286). */
600 rewrite_into_loop_closed_ssa (NULL
, TODO_update_ssa_only_virtuals
);
601 return TODO_cleanup_cfg
;
608 /* Entry point to basic block SLP phase. */
612 const pass_data pass_data_slp_vectorize
=
614 GIMPLE_PASS
, /* type */
616 OPTGROUP_LOOP
| OPTGROUP_VEC
, /* optinfo_flags */
617 TV_TREE_SLP_VECTORIZATION
, /* tv_id */
618 ( PROP_ssa
| PROP_cfg
), /* properties_required */
619 0, /* properties_provided */
620 0, /* properties_destroyed */
621 0, /* todo_flags_start */
622 TODO_update_ssa
, /* todo_flags_finish */
625 class pass_slp_vectorize
: public gimple_opt_pass
628 pass_slp_vectorize (gcc::context
*ctxt
)
629 : gimple_opt_pass (pass_data_slp_vectorize
, ctxt
)
632 /* opt_pass methods: */
633 opt_pass
* clone () { return new pass_slp_vectorize (m_ctxt
); }
634 virtual bool gate (function
*) { return flag_tree_slp_vectorize
!= 0; }
635 virtual unsigned int execute (function
*);
637 }; // class pass_slp_vectorize
640 pass_slp_vectorize::execute (function
*fun
)
644 bool in_loop_pipeline
= scev_initialized_p ();
645 if (!in_loop_pipeline
)
647 loop_optimizer_init (LOOPS_NORMAL
);
651 init_stmt_vec_info_vec ();
653 FOR_EACH_BB_FN (bb
, fun
)
655 vect_location
= find_bb_location (bb
);
657 if (vect_slp_analyze_bb (bb
))
659 if (!dbg_cnt (vect_slp
))
662 vect_slp_transform_bb (bb
);
663 if (dump_enabled_p ())
664 dump_printf_loc (MSG_OPTIMIZED_LOCATIONS
, vect_location
,
665 "basic block vectorized\n");
669 free_stmt_vec_info_vec ();
671 if (!in_loop_pipeline
)
674 loop_optimizer_finalize ();
683 make_pass_slp_vectorize (gcc::context
*ctxt
)
685 return new pass_slp_vectorize (ctxt
);
689 /* Increase alignment of global arrays to improve vectorization potential.
691 - Consider also structs that have an array field.
692 - Use ipa analysis to prune arrays that can't be vectorized?
693 This should involve global alignment analysis and in the future also
697 increase_alignment (void)
701 vect_location
= UNKNOWN_LOCATION
;
703 /* Increase the alignment of all global arrays for vectorization. */
704 FOR_EACH_DEFINED_VARIABLE (vnode
)
706 tree vectype
, decl
= vnode
->decl
;
708 unsigned int alignment
;
710 t
= TREE_TYPE (decl
);
711 if (TREE_CODE (t
) != ARRAY_TYPE
)
713 vectype
= get_vectype_for_scalar_type (strip_array_types (t
));
716 alignment
= TYPE_ALIGN (vectype
);
717 if (DECL_ALIGN (decl
) >= alignment
)
720 if (vect_can_force_dr_alignment_p (decl
, alignment
))
722 DECL_ALIGN (decl
) = TYPE_ALIGN (vectype
);
723 DECL_USER_ALIGN (decl
) = 1;
724 if (TREE_STATIC (decl
))
726 tree target
= symtab_node::get (decl
)->ultimate_alias_target ()->decl
;
727 DECL_ALIGN (target
) = TYPE_ALIGN (vectype
);
728 DECL_USER_ALIGN (target
) = 1;
730 dump_printf (MSG_NOTE
, "Increasing alignment of decl: ");
731 dump_generic_expr (MSG_NOTE
, TDF_SLIM
, decl
);
732 dump_printf (MSG_NOTE
, "\n");
741 const pass_data pass_data_ipa_increase_alignment
=
743 SIMPLE_IPA_PASS
, /* type */
744 "increase_alignment", /* name */
745 OPTGROUP_LOOP
| OPTGROUP_VEC
, /* optinfo_flags */
746 TV_IPA_OPT
, /* tv_id */
747 0, /* properties_required */
748 0, /* properties_provided */
749 0, /* properties_destroyed */
750 0, /* todo_flags_start */
751 0, /* todo_flags_finish */
754 class pass_ipa_increase_alignment
: public simple_ipa_opt_pass
757 pass_ipa_increase_alignment (gcc::context
*ctxt
)
758 : simple_ipa_opt_pass (pass_data_ipa_increase_alignment
, ctxt
)
761 /* opt_pass methods: */
762 virtual bool gate (function
*)
764 return flag_section_anchors
&& flag_tree_loop_vectorize
;
767 virtual unsigned int execute (function
*) { return increase_alignment (); }
769 }; // class pass_ipa_increase_alignment
773 simple_ipa_opt_pass
*
774 make_pass_ipa_increase_alignment (gcc::context
*ctxt
)
776 return new pass_ipa_increase_alignment (ctxt
);