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 "fold-const.h"
66 #include "stor-layout.h"
67 #include "tree-pretty-print.h"
69 #include "hard-reg-set.h"
71 #include "dominance.h"
73 #include "basic-block.h"
74 #include "tree-ssa-alias.h"
75 #include "internal-fn.h"
76 #include "gimple-expr.h"
78 #include "gimple-iterator.h"
79 #include "gimple-walk.h"
80 #include "gimple-ssa.h"
82 #include "tree-phinodes.h"
83 #include "ssa-iterators.h"
84 #include "tree-ssa-loop-manip.h"
87 #include "tree-vectorizer.h"
88 #include "tree-pass.h"
89 #include "tree-ssa-propagate.h"
91 #include "gimple-fold.h"
92 #include "tree-scalar-evolution.h"
95 /* Loop or bb location. */
96 source_location vect_location
;
98 /* Vector mapping GIMPLE stmt to stmt_vec_info. */
99 vec
<vec_void_p
> stmt_vec_info_vec
;
101 /* For mapping simduid to vectorization factor. */
103 struct simduid_to_vf
: free_ptr_hash
<simduid_to_vf
>
105 unsigned int simduid
;
108 /* hash_table support. */
109 static inline hashval_t
hash (const simduid_to_vf
*);
110 static inline int equal (const simduid_to_vf
*, const simduid_to_vf
*);
114 simduid_to_vf::hash (const simduid_to_vf
*p
)
120 simduid_to_vf::equal (const simduid_to_vf
*p1
, const simduid_to_vf
*p2
)
122 return p1
->simduid
== p2
->simduid
;
125 /* This hash maps the OMP simd array to the corresponding simduid used
126 to index into it. Like thus,
128 _7 = GOMP_SIMD_LANE (simduid.0)
134 This hash maps from the OMP simd array (D.1737[]) to DECL_UID of
137 struct simd_array_to_simduid
: free_ptr_hash
<simd_array_to_simduid
>
140 unsigned int simduid
;
142 /* hash_table support. */
143 static inline hashval_t
hash (const simd_array_to_simduid
*);
144 static inline int equal (const simd_array_to_simduid
*,
145 const simd_array_to_simduid
*);
149 simd_array_to_simduid::hash (const simd_array_to_simduid
*p
)
151 return DECL_UID (p
->decl
);
155 simd_array_to_simduid::equal (const simd_array_to_simduid
*p1
,
156 const simd_array_to_simduid
*p2
)
158 return p1
->decl
== p2
->decl
;
161 /* Fold IFN_GOMP_SIMD_LANE, IFN_GOMP_SIMD_VF and IFN_GOMP_SIMD_LAST_LANE
162 into their corresponding constants. */
165 adjust_simduid_builtins (hash_table
<simduid_to_vf
> *htab
)
169 FOR_EACH_BB_FN (bb
, cfun
)
171 gimple_stmt_iterator i
;
173 for (i
= gsi_start_bb (bb
); !gsi_end_p (i
); gsi_next (&i
))
176 enum internal_fn ifn
;
177 gimple stmt
= gsi_stmt (i
);
179 if (!is_gimple_call (stmt
)
180 || !gimple_call_internal_p (stmt
))
182 ifn
= gimple_call_internal_fn (stmt
);
185 case IFN_GOMP_SIMD_LANE
:
186 case IFN_GOMP_SIMD_VF
:
187 case IFN_GOMP_SIMD_LAST_LANE
:
192 tree arg
= gimple_call_arg (stmt
, 0);
193 gcc_assert (arg
!= NULL_TREE
);
194 gcc_assert (TREE_CODE (arg
) == SSA_NAME
);
195 simduid_to_vf
*p
= NULL
, data
;
196 data
.simduid
= DECL_UID (SSA_NAME_VAR (arg
));
199 p
= htab
->find (&data
);
205 case IFN_GOMP_SIMD_VF
:
206 t
= build_int_cst (unsigned_type_node
, vf
);
208 case IFN_GOMP_SIMD_LANE
:
209 t
= build_int_cst (unsigned_type_node
, 0);
211 case IFN_GOMP_SIMD_LAST_LANE
:
212 t
= gimple_call_arg (stmt
, 1);
217 update_call_from_tree (&i
, t
);
222 /* Helper structure for note_simd_array_uses. */
224 struct note_simd_array_uses_struct
226 hash_table
<simd_array_to_simduid
> **htab
;
227 unsigned int simduid
;
230 /* Callback for note_simd_array_uses, called through walk_gimple_op. */
233 note_simd_array_uses_cb (tree
*tp
, int *walk_subtrees
, void *data
)
235 struct walk_stmt_info
*wi
= (struct walk_stmt_info
*) data
;
236 struct note_simd_array_uses_struct
*ns
237 = (struct note_simd_array_uses_struct
*) wi
->info
;
242 && lookup_attribute ("omp simd array", DECL_ATTRIBUTES (*tp
))
243 && DECL_CONTEXT (*tp
) == current_function_decl
)
245 simd_array_to_simduid data
;
247 *ns
->htab
= new hash_table
<simd_array_to_simduid
> (15);
249 data
.simduid
= ns
->simduid
;
250 simd_array_to_simduid
**slot
= (*ns
->htab
)->find_slot (&data
, INSERT
);
253 simd_array_to_simduid
*p
= XNEW (simd_array_to_simduid
);
257 else if ((*slot
)->simduid
!= ns
->simduid
)
258 (*slot
)->simduid
= -1U;
264 /* Find "omp simd array" temporaries and map them to corresponding
268 note_simd_array_uses (hash_table
<simd_array_to_simduid
> **htab
)
271 gimple_stmt_iterator gsi
;
272 struct walk_stmt_info wi
;
273 struct note_simd_array_uses_struct ns
;
275 memset (&wi
, 0, sizeof (wi
));
279 FOR_EACH_BB_FN (bb
, cfun
)
280 for (gsi
= gsi_start_bb (bb
); !gsi_end_p (gsi
); gsi_next (&gsi
))
282 gimple stmt
= gsi_stmt (gsi
);
283 if (!is_gimple_call (stmt
) || !gimple_call_internal_p (stmt
))
285 switch (gimple_call_internal_fn (stmt
))
287 case IFN_GOMP_SIMD_LANE
:
288 case IFN_GOMP_SIMD_VF
:
289 case IFN_GOMP_SIMD_LAST_LANE
:
294 tree lhs
= gimple_call_lhs (stmt
);
295 if (lhs
== NULL_TREE
)
297 imm_use_iterator use_iter
;
299 ns
.simduid
= DECL_UID (SSA_NAME_VAR (gimple_call_arg (stmt
, 0)));
300 FOR_EACH_IMM_USE_STMT (use_stmt
, use_iter
, lhs
)
301 if (!is_gimple_debug (use_stmt
))
302 walk_gimple_op (use_stmt
, note_simd_array_uses_cb
, &wi
);
306 /* Shrink arrays with "omp simd array" attribute to the corresponding
307 vectorization factor. */
311 (hash_table
<simd_array_to_simduid
> *simd_array_to_simduid_htab
,
312 hash_table
<simduid_to_vf
> *simduid_to_vf_htab
)
314 for (hash_table
<simd_array_to_simduid
>::iterator iter
315 = simd_array_to_simduid_htab
->begin ();
316 iter
!= simd_array_to_simduid_htab
->end (); ++iter
)
317 if ((*iter
)->simduid
!= -1U)
319 tree decl
= (*iter
)->decl
;
321 if (simduid_to_vf_htab
)
323 simduid_to_vf
*p
= NULL
, data
;
324 data
.simduid
= (*iter
)->simduid
;
325 p
= simduid_to_vf_htab
->find (&data
);
330 = build_array_type_nelts (TREE_TYPE (TREE_TYPE (decl
)), vf
);
331 TREE_TYPE (decl
) = atype
;
332 relayout_decl (decl
);
335 delete simd_array_to_simduid_htab
;
338 /* A helper function to free data refs. */
341 vect_destroy_datarefs (loop_vec_info loop_vinfo
, bb_vec_info bb_vinfo
)
343 vec
<data_reference_p
> datarefs
;
344 struct data_reference
*dr
;
348 datarefs
= LOOP_VINFO_DATAREFS (loop_vinfo
);
350 datarefs
= BB_VINFO_DATAREFS (bb_vinfo
);
352 FOR_EACH_VEC_ELT (datarefs
, i
, dr
)
359 free_data_refs (datarefs
);
363 /* If LOOP has been versioned during ifcvt, return the internal call
367 vect_loop_vectorized_call (struct loop
*loop
)
369 basic_block bb
= loop_preheader_edge (loop
)->src
;
376 if (!single_pred_p (bb
))
378 bb
= single_pred (bb
);
381 if (g
&& gimple_code (g
) == GIMPLE_COND
)
383 gimple_stmt_iterator gsi
= gsi_for_stmt (g
);
385 if (!gsi_end_p (gsi
))
388 if (is_gimple_call (g
)
389 && gimple_call_internal_p (g
)
390 && gimple_call_internal_fn (g
) == IFN_LOOP_VECTORIZED
391 && (tree_to_shwi (gimple_call_arg (g
, 0)) == loop
->num
392 || tree_to_shwi (gimple_call_arg (g
, 1)) == loop
->num
))
399 /* Fold LOOP_VECTORIZED internal call G to VALUE and
400 update any immediate uses of it's LHS. */
403 fold_loop_vectorized_call (gimple g
, tree value
)
405 tree lhs
= gimple_call_lhs (g
);
407 imm_use_iterator iter
;
409 gimple_stmt_iterator gsi
= gsi_for_stmt (g
);
411 update_call_from_tree (&gsi
, value
);
412 FOR_EACH_IMM_USE_STMT (use_stmt
, iter
, lhs
)
414 FOR_EACH_IMM_USE_ON_STMT (use_p
, iter
)
415 SET_USE (use_p
, value
);
416 update_stmt (use_stmt
);
419 /* Set the uids of all the statements in basic blocks inside loop
420 represented by LOOP_VINFO. LOOP_VECTORIZED_CALL is the internal
421 call guarding the loop which has been if converted. */
423 set_uid_loop_bbs (loop_vec_info loop_vinfo
, gimple loop_vectorized_call
)
425 tree arg
= gimple_call_arg (loop_vectorized_call
, 1);
428 struct loop
*scalar_loop
= get_loop (cfun
, tree_to_shwi (arg
));
430 LOOP_VINFO_SCALAR_LOOP (loop_vinfo
) = scalar_loop
;
431 gcc_checking_assert (vect_loop_vectorized_call
432 (LOOP_VINFO_SCALAR_LOOP (loop_vinfo
))
433 == loop_vectorized_call
);
434 bbs
= get_loop_body (scalar_loop
);
435 for (i
= 0; i
< scalar_loop
->num_nodes
; i
++)
437 basic_block bb
= bbs
[i
];
438 gimple_stmt_iterator gsi
;
439 for (gsi
= gsi_start_phis (bb
); !gsi_end_p (gsi
); gsi_next (&gsi
))
441 gimple phi
= gsi_stmt (gsi
);
442 gimple_set_uid (phi
, 0);
444 for (gsi
= gsi_start_bb (bb
); !gsi_end_p (gsi
); gsi_next (&gsi
))
446 gimple stmt
= gsi_stmt (gsi
);
447 gimple_set_uid (stmt
, 0);
453 /* Function vectorize_loops.
455 Entry point to loop vectorization phase. */
458 vectorize_loops (void)
461 unsigned int num_vectorized_loops
= 0;
462 unsigned int vect_loops_num
;
464 hash_table
<simduid_to_vf
> *simduid_to_vf_htab
= NULL
;
465 hash_table
<simd_array_to_simduid
> *simd_array_to_simduid_htab
= NULL
;
466 bool any_ifcvt_loops
= false;
469 vect_loops_num
= number_of_loops (cfun
);
471 /* Bail out if there are no loops. */
472 if (vect_loops_num
<= 1)
475 if (cfun
->has_simduid_loops
)
476 note_simd_array_uses (&simd_array_to_simduid_htab
);
478 init_stmt_vec_info_vec ();
480 /* ----------- Analyze loops. ----------- */
482 /* If some loop was duplicated, it gets bigger number
483 than all previously defined loops. This fact allows us to run
484 only over initial loops skipping newly generated ones. */
485 FOR_EACH_LOOP (loop
, 0)
486 if (loop
->dont_vectorize
)
487 any_ifcvt_loops
= true;
488 else if ((flag_tree_loop_vectorize
489 && optimize_loop_nest_for_speed_p (loop
))
490 || loop
->force_vectorize
)
492 loop_vec_info loop_vinfo
;
493 vect_location
= find_loop_location (loop
);
494 if (LOCATION_LOCUS (vect_location
) != UNKNOWN_LOCATION
495 && dump_enabled_p ())
496 dump_printf (MSG_NOTE
, "\nAnalyzing loop at %s:%d\n",
497 LOCATION_FILE (vect_location
),
498 LOCATION_LINE (vect_location
));
500 loop_vinfo
= vect_analyze_loop (loop
);
501 loop
->aux
= loop_vinfo
;
503 if (!loop_vinfo
|| !LOOP_VINFO_VECTORIZABLE_P (loop_vinfo
))
506 if (!dbg_cnt (vect_loop
))
509 gimple loop_vectorized_call
= vect_loop_vectorized_call (loop
);
510 if (loop_vectorized_call
)
511 set_uid_loop_bbs (loop_vinfo
, loop_vectorized_call
);
512 if (LOCATION_LOCUS (vect_location
) != UNKNOWN_LOCATION
513 && dump_enabled_p ())
514 dump_printf_loc (MSG_OPTIMIZED_LOCATIONS
, vect_location
,
515 "loop vectorized\n");
516 vect_transform_loop (loop_vinfo
);
517 num_vectorized_loops
++;
518 /* Now that the loop has been vectorized, allow it to be unrolled
520 loop
->force_vectorize
= false;
524 simduid_to_vf
*simduid_to_vf_data
= XNEW (simduid_to_vf
);
525 if (!simduid_to_vf_htab
)
526 simduid_to_vf_htab
= new hash_table
<simduid_to_vf
> (15);
527 simduid_to_vf_data
->simduid
= DECL_UID (loop
->simduid
);
528 simduid_to_vf_data
->vf
= loop_vinfo
->vectorization_factor
;
529 *simduid_to_vf_htab
->find_slot (simduid_to_vf_data
, INSERT
)
530 = simduid_to_vf_data
;
533 if (loop_vectorized_call
)
535 fold_loop_vectorized_call (loop_vectorized_call
, boolean_true_node
);
536 ret
|= TODO_cleanup_cfg
;
540 vect_location
= UNKNOWN_LOCATION
;
542 statistics_counter_event (cfun
, "Vectorized loops", num_vectorized_loops
);
543 if (dump_enabled_p ()
544 || (num_vectorized_loops
> 0 && dump_enabled_p ()))
545 dump_printf_loc (MSG_NOTE
, vect_location
,
546 "vectorized %u loops in function.\n",
547 num_vectorized_loops
);
549 /* ----------- Finalize. ----------- */
552 for (i
= 1; i
< vect_loops_num
; i
++)
554 loop
= get_loop (cfun
, i
);
555 if (loop
&& loop
->dont_vectorize
)
557 gimple g
= vect_loop_vectorized_call (loop
);
560 fold_loop_vectorized_call (g
, boolean_false_node
);
561 ret
|= TODO_cleanup_cfg
;
566 for (i
= 1; i
< vect_loops_num
; i
++)
568 loop_vec_info loop_vinfo
;
570 loop
= get_loop (cfun
, i
);
573 loop_vinfo
= (loop_vec_info
) loop
->aux
;
574 destroy_loop_vec_info (loop_vinfo
, true);
578 free_stmt_vec_info_vec ();
580 /* Fold IFN_GOMP_SIMD_{VF,LANE,LAST_LANE} builtins. */
581 if (cfun
->has_simduid_loops
)
582 adjust_simduid_builtins (simduid_to_vf_htab
);
584 /* Shrink any "omp array simd" temporary arrays to the
585 actual vectorization factors. */
586 if (simd_array_to_simduid_htab
)
587 shrink_simd_arrays (simd_array_to_simduid_htab
, simduid_to_vf_htab
);
588 delete simduid_to_vf_htab
;
589 cfun
->has_simduid_loops
= false;
591 if (num_vectorized_loops
> 0)
593 /* If we vectorized any loop only virtual SSA form needs to be updated.
594 ??? Also while we try hard to update loop-closed SSA form we fail
595 to properly do this in some corner-cases (see PR56286). */
596 rewrite_into_loop_closed_ssa (NULL
, TODO_update_ssa_only_virtuals
);
597 return TODO_cleanup_cfg
;
604 /* Entry point to the simduid cleanup pass. */
608 const pass_data pass_data_simduid_cleanup
=
610 GIMPLE_PASS
, /* type */
611 "simduid", /* name */
612 OPTGROUP_NONE
, /* optinfo_flags */
614 ( PROP_ssa
| PROP_cfg
), /* properties_required */
615 0, /* properties_provided */
616 0, /* properties_destroyed */
617 0, /* todo_flags_start */
618 0, /* todo_flags_finish */
621 class pass_simduid_cleanup
: public gimple_opt_pass
624 pass_simduid_cleanup (gcc::context
*ctxt
)
625 : gimple_opt_pass (pass_data_simduid_cleanup
, ctxt
)
628 /* opt_pass methods: */
629 opt_pass
* clone () { return new pass_simduid_cleanup (m_ctxt
); }
630 virtual bool gate (function
*fun
) { return fun
->has_simduid_loops
; }
631 virtual unsigned int execute (function
*);
633 }; // class pass_simduid_cleanup
636 pass_simduid_cleanup::execute (function
*fun
)
638 hash_table
<simd_array_to_simduid
> *simd_array_to_simduid_htab
= NULL
;
640 note_simd_array_uses (&simd_array_to_simduid_htab
);
642 /* Fold IFN_GOMP_SIMD_{VF,LANE,LAST_LANE} builtins. */
643 adjust_simduid_builtins (NULL
);
645 /* Shrink any "omp array simd" temporary arrays to the
646 actual vectorization factors. */
647 if (simd_array_to_simduid_htab
)
648 shrink_simd_arrays (simd_array_to_simduid_htab
, NULL
);
649 fun
->has_simduid_loops
= false;
656 make_pass_simduid_cleanup (gcc::context
*ctxt
)
658 return new pass_simduid_cleanup (ctxt
);
662 /* Entry point to basic block SLP phase. */
666 const pass_data pass_data_slp_vectorize
=
668 GIMPLE_PASS
, /* type */
670 OPTGROUP_LOOP
| OPTGROUP_VEC
, /* optinfo_flags */
671 TV_TREE_SLP_VECTORIZATION
, /* tv_id */
672 ( PROP_ssa
| PROP_cfg
), /* properties_required */
673 0, /* properties_provided */
674 0, /* properties_destroyed */
675 0, /* todo_flags_start */
676 TODO_update_ssa
, /* todo_flags_finish */
679 class pass_slp_vectorize
: public gimple_opt_pass
682 pass_slp_vectorize (gcc::context
*ctxt
)
683 : gimple_opt_pass (pass_data_slp_vectorize
, ctxt
)
686 /* opt_pass methods: */
687 opt_pass
* clone () { return new pass_slp_vectorize (m_ctxt
); }
688 virtual bool gate (function
*) { return flag_tree_slp_vectorize
!= 0; }
689 virtual unsigned int execute (function
*);
691 }; // class pass_slp_vectorize
694 pass_slp_vectorize::execute (function
*fun
)
698 bool in_loop_pipeline
= scev_initialized_p ();
699 if (!in_loop_pipeline
)
701 loop_optimizer_init (LOOPS_NORMAL
);
705 init_stmt_vec_info_vec ();
707 FOR_EACH_BB_FN (bb
, fun
)
709 vect_location
= find_bb_location (bb
);
711 if (vect_slp_analyze_bb (bb
))
713 if (!dbg_cnt (vect_slp
))
716 vect_slp_transform_bb (bb
);
717 if (dump_enabled_p ())
718 dump_printf_loc (MSG_OPTIMIZED_LOCATIONS
, vect_location
,
719 "basic block vectorized\n");
723 free_stmt_vec_info_vec ();
725 if (!in_loop_pipeline
)
728 loop_optimizer_finalize ();
737 make_pass_slp_vectorize (gcc::context
*ctxt
)
739 return new pass_slp_vectorize (ctxt
);
743 /* Increase alignment of global arrays to improve vectorization potential.
745 - Consider also structs that have an array field.
746 - Use ipa analysis to prune arrays that can't be vectorized?
747 This should involve global alignment analysis and in the future also
751 increase_alignment (void)
755 vect_location
= UNKNOWN_LOCATION
;
757 /* Increase the alignment of all global arrays for vectorization. */
758 FOR_EACH_DEFINED_VARIABLE (vnode
)
760 tree vectype
, decl
= vnode
->decl
;
762 unsigned int alignment
;
764 t
= TREE_TYPE (decl
);
765 if (TREE_CODE (t
) != ARRAY_TYPE
)
767 vectype
= get_vectype_for_scalar_type (strip_array_types (t
));
770 alignment
= TYPE_ALIGN (vectype
);
771 if (DECL_ALIGN (decl
) >= alignment
)
774 if (vect_can_force_dr_alignment_p (decl
, alignment
))
776 vnode
->increase_alignment (TYPE_ALIGN (vectype
));
777 dump_printf (MSG_NOTE
, "Increasing alignment of decl: ");
778 dump_generic_expr (MSG_NOTE
, TDF_SLIM
, decl
);
779 dump_printf (MSG_NOTE
, "\n");
788 const pass_data pass_data_ipa_increase_alignment
=
790 SIMPLE_IPA_PASS
, /* type */
791 "increase_alignment", /* name */
792 OPTGROUP_LOOP
| OPTGROUP_VEC
, /* optinfo_flags */
793 TV_IPA_OPT
, /* tv_id */
794 0, /* properties_required */
795 0, /* properties_provided */
796 0, /* properties_destroyed */
797 0, /* todo_flags_start */
798 0, /* todo_flags_finish */
801 class pass_ipa_increase_alignment
: public simple_ipa_opt_pass
804 pass_ipa_increase_alignment (gcc::context
*ctxt
)
805 : simple_ipa_opt_pass (pass_data_ipa_increase_alignment
, ctxt
)
808 /* opt_pass methods: */
809 virtual bool gate (function
*)
811 return flag_section_anchors
&& flag_tree_loop_vectorize
;
814 virtual unsigned int execute (function
*) { return increase_alignment (); }
816 }; // class pass_ipa_increase_alignment
820 simple_ipa_opt_pass
*
821 make_pass_ipa_increase_alignment (gcc::context
*ctxt
)
823 return new pass_ipa_increase_alignment (ctxt
);