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
));
213 p
= htab
->find (&data
);
219 case IFN_GOMP_SIMD_VF
:
220 t
= build_int_cst (unsigned_type_node
, vf
);
222 case IFN_GOMP_SIMD_LANE
:
223 t
= build_int_cst (unsigned_type_node
, 0);
225 case IFN_GOMP_SIMD_LAST_LANE
:
226 t
= gimple_call_arg (stmt
, 1);
231 update_call_from_tree (&i
, t
);
236 /* Helper structure for note_simd_array_uses. */
238 struct note_simd_array_uses_struct
240 hash_table
<simd_array_to_simduid
> **htab
;
241 unsigned int simduid
;
244 /* Callback for note_simd_array_uses, called through walk_gimple_op. */
247 note_simd_array_uses_cb (tree
*tp
, int *walk_subtrees
, void *data
)
249 struct walk_stmt_info
*wi
= (struct walk_stmt_info
*) data
;
250 struct note_simd_array_uses_struct
*ns
251 = (struct note_simd_array_uses_struct
*) wi
->info
;
256 && lookup_attribute ("omp simd array", DECL_ATTRIBUTES (*tp
))
257 && DECL_CONTEXT (*tp
) == current_function_decl
)
259 simd_array_to_simduid data
;
261 *ns
->htab
= new hash_table
<simd_array_to_simduid
> (15);
263 data
.simduid
= ns
->simduid
;
264 simd_array_to_simduid
**slot
= (*ns
->htab
)->find_slot (&data
, INSERT
);
267 simd_array_to_simduid
*p
= XNEW (simd_array_to_simduid
);
271 else if ((*slot
)->simduid
!= ns
->simduid
)
272 (*slot
)->simduid
= -1U;
278 /* Find "omp simd array" temporaries and map them to corresponding
282 note_simd_array_uses (hash_table
<simd_array_to_simduid
> **htab
)
285 gimple_stmt_iterator gsi
;
286 struct walk_stmt_info wi
;
287 struct note_simd_array_uses_struct ns
;
289 memset (&wi
, 0, sizeof (wi
));
293 FOR_EACH_BB_FN (bb
, cfun
)
294 for (gsi
= gsi_start_bb (bb
); !gsi_end_p (gsi
); gsi_next (&gsi
))
296 gimple stmt
= gsi_stmt (gsi
);
297 if (!is_gimple_call (stmt
) || !gimple_call_internal_p (stmt
))
299 switch (gimple_call_internal_fn (stmt
))
301 case IFN_GOMP_SIMD_LANE
:
302 case IFN_GOMP_SIMD_VF
:
303 case IFN_GOMP_SIMD_LAST_LANE
:
308 tree lhs
= gimple_call_lhs (stmt
);
309 if (lhs
== NULL_TREE
)
311 imm_use_iterator use_iter
;
313 ns
.simduid
= DECL_UID (SSA_NAME_VAR (gimple_call_arg (stmt
, 0)));
314 FOR_EACH_IMM_USE_STMT (use_stmt
, use_iter
, lhs
)
315 if (!is_gimple_debug (use_stmt
))
316 walk_gimple_op (use_stmt
, note_simd_array_uses_cb
, &wi
);
320 /* Shrink arrays with "omp simd array" attribute to the corresponding
321 vectorization factor. */
325 (hash_table
<simd_array_to_simduid
> *simd_array_to_simduid_htab
,
326 hash_table
<simduid_to_vf
> *simduid_to_vf_htab
)
328 for (hash_table
<simd_array_to_simduid
>::iterator iter
329 = simd_array_to_simduid_htab
->begin ();
330 iter
!= simd_array_to_simduid_htab
->end (); ++iter
)
331 if ((*iter
)->simduid
!= -1U)
333 tree decl
= (*iter
)->decl
;
335 if (simduid_to_vf_htab
)
337 simduid_to_vf
*p
= NULL
, data
;
338 data
.simduid
= (*iter
)->simduid
;
339 p
= simduid_to_vf_htab
->find (&data
);
344 = build_array_type_nelts (TREE_TYPE (TREE_TYPE (decl
)), vf
);
345 TREE_TYPE (decl
) = atype
;
346 relayout_decl (decl
);
349 delete simd_array_to_simduid_htab
;
352 /* A helper function to free data refs. */
355 vect_destroy_datarefs (loop_vec_info loop_vinfo
, bb_vec_info bb_vinfo
)
357 vec
<data_reference_p
> datarefs
;
358 struct data_reference
*dr
;
362 datarefs
= LOOP_VINFO_DATAREFS (loop_vinfo
);
364 datarefs
= BB_VINFO_DATAREFS (bb_vinfo
);
366 FOR_EACH_VEC_ELT (datarefs
, i
, dr
)
373 free_data_refs (datarefs
);
377 /* If LOOP has been versioned during ifcvt, return the internal call
381 vect_loop_vectorized_call (struct loop
*loop
)
383 basic_block bb
= loop_preheader_edge (loop
)->src
;
390 if (!single_pred_p (bb
))
392 bb
= single_pred (bb
);
395 if (g
&& gimple_code (g
) == GIMPLE_COND
)
397 gimple_stmt_iterator gsi
= gsi_for_stmt (g
);
399 if (!gsi_end_p (gsi
))
402 if (is_gimple_call (g
)
403 && gimple_call_internal_p (g
)
404 && gimple_call_internal_fn (g
) == IFN_LOOP_VECTORIZED
405 && (tree_to_shwi (gimple_call_arg (g
, 0)) == loop
->num
406 || tree_to_shwi (gimple_call_arg (g
, 1)) == loop
->num
))
413 /* Fold LOOP_VECTORIZED internal call G to VALUE and
414 update any immediate uses of it's LHS. */
417 fold_loop_vectorized_call (gimple g
, tree value
)
419 tree lhs
= gimple_call_lhs (g
);
421 imm_use_iterator iter
;
423 gimple_stmt_iterator gsi
= gsi_for_stmt (g
);
425 update_call_from_tree (&gsi
, value
);
426 FOR_EACH_IMM_USE_STMT (use_stmt
, iter
, lhs
)
428 FOR_EACH_IMM_USE_ON_STMT (use_p
, iter
)
429 SET_USE (use_p
, value
);
430 update_stmt (use_stmt
);
434 /* Function vectorize_loops.
436 Entry point to loop vectorization phase. */
439 vectorize_loops (void)
442 unsigned int num_vectorized_loops
= 0;
443 unsigned int vect_loops_num
;
445 hash_table
<simduid_to_vf
> *simduid_to_vf_htab
= NULL
;
446 hash_table
<simd_array_to_simduid
> *simd_array_to_simduid_htab
= NULL
;
447 bool any_ifcvt_loops
= false;
450 vect_loops_num
= number_of_loops (cfun
);
452 /* Bail out if there are no loops. */
453 if (vect_loops_num
<= 1)
456 if (cfun
->has_simduid_loops
)
457 note_simd_array_uses (&simd_array_to_simduid_htab
);
459 init_stmt_vec_info_vec ();
461 /* ----------- Analyze loops. ----------- */
463 /* If some loop was duplicated, it gets bigger number
464 than all previously defined loops. This fact allows us to run
465 only over initial loops skipping newly generated ones. */
466 FOR_EACH_LOOP (loop
, 0)
467 if (loop
->dont_vectorize
)
468 any_ifcvt_loops
= true;
469 else if ((flag_tree_loop_vectorize
470 && optimize_loop_nest_for_speed_p (loop
))
471 || loop
->force_vectorize
)
473 loop_vec_info loop_vinfo
;
474 vect_location
= find_loop_location (loop
);
475 if (LOCATION_LOCUS (vect_location
) != UNKNOWN_LOCATION
476 && dump_enabled_p ())
477 dump_printf (MSG_NOTE
, "\nAnalyzing loop at %s:%d\n",
478 LOCATION_FILE (vect_location
),
479 LOCATION_LINE (vect_location
));
481 loop_vinfo
= vect_analyze_loop (loop
);
482 loop
->aux
= loop_vinfo
;
484 if (!loop_vinfo
|| !LOOP_VINFO_VECTORIZABLE_P (loop_vinfo
))
487 if (!dbg_cnt (vect_loop
))
490 gimple loop_vectorized_call
= vect_loop_vectorized_call (loop
);
491 if (loop_vectorized_call
)
493 tree arg
= gimple_call_arg (loop_vectorized_call
, 1);
496 struct loop
*scalar_loop
= get_loop (cfun
, tree_to_shwi (arg
));
498 LOOP_VINFO_SCALAR_LOOP (loop_vinfo
) = scalar_loop
;
499 gcc_checking_assert (vect_loop_vectorized_call
500 (LOOP_VINFO_SCALAR_LOOP (loop_vinfo
))
501 == loop_vectorized_call
);
502 bbs
= get_loop_body (scalar_loop
);
503 for (i
= 0; i
< scalar_loop
->num_nodes
; i
++)
505 basic_block bb
= bbs
[i
];
506 gimple_stmt_iterator gsi
;
507 for (gsi
= gsi_start_phis (bb
); !gsi_end_p (gsi
);
510 gimple phi
= gsi_stmt (gsi
);
511 gimple_set_uid (phi
, 0);
513 for (gsi
= gsi_start_bb (bb
); !gsi_end_p (gsi
);
516 gimple stmt
= gsi_stmt (gsi
);
517 gimple_set_uid (stmt
, 0);
523 if (LOCATION_LOCUS (vect_location
) != UNKNOWN_LOCATION
524 && dump_enabled_p ())
525 dump_printf_loc (MSG_OPTIMIZED_LOCATIONS
, vect_location
,
526 "loop vectorized\n");
527 vect_transform_loop (loop_vinfo
);
528 num_vectorized_loops
++;
529 /* Now that the loop has been vectorized, allow it to be unrolled
531 loop
->force_vectorize
= false;
535 simduid_to_vf
*simduid_to_vf_data
= XNEW (simduid_to_vf
);
536 if (!simduid_to_vf_htab
)
537 simduid_to_vf_htab
= new hash_table
<simduid_to_vf
> (15);
538 simduid_to_vf_data
->simduid
= DECL_UID (loop
->simduid
);
539 simduid_to_vf_data
->vf
= loop_vinfo
->vectorization_factor
;
540 *simduid_to_vf_htab
->find_slot (simduid_to_vf_data
, INSERT
)
541 = simduid_to_vf_data
;
544 if (loop_vectorized_call
)
546 fold_loop_vectorized_call (loop_vectorized_call
, boolean_true_node
);
547 ret
|= TODO_cleanup_cfg
;
551 vect_location
= UNKNOWN_LOCATION
;
553 statistics_counter_event (cfun
, "Vectorized loops", num_vectorized_loops
);
554 if (dump_enabled_p ()
555 || (num_vectorized_loops
> 0 && dump_enabled_p ()))
556 dump_printf_loc (MSG_NOTE
, vect_location
,
557 "vectorized %u loops in function.\n",
558 num_vectorized_loops
);
560 /* ----------- Finalize. ----------- */
563 for (i
= 1; i
< vect_loops_num
; i
++)
565 loop
= get_loop (cfun
, i
);
566 if (loop
&& loop
->dont_vectorize
)
568 gimple g
= vect_loop_vectorized_call (loop
);
571 fold_loop_vectorized_call (g
, boolean_false_node
);
572 ret
|= TODO_cleanup_cfg
;
577 for (i
= 1; i
< vect_loops_num
; i
++)
579 loop_vec_info loop_vinfo
;
581 loop
= get_loop (cfun
, i
);
584 loop_vinfo
= (loop_vec_info
) loop
->aux
;
585 destroy_loop_vec_info (loop_vinfo
, true);
589 free_stmt_vec_info_vec ();
591 /* Fold IFN_GOMP_SIMD_{VF,LANE,LAST_LANE} builtins. */
592 if (cfun
->has_simduid_loops
)
593 adjust_simduid_builtins (simduid_to_vf_htab
);
595 /* Shrink any "omp array simd" temporary arrays to the
596 actual vectorization factors. */
597 if (simd_array_to_simduid_htab
)
598 shrink_simd_arrays (simd_array_to_simduid_htab
, simduid_to_vf_htab
);
599 delete simduid_to_vf_htab
;
600 cfun
->has_simduid_loops
= false;
602 if (num_vectorized_loops
> 0)
604 /* If we vectorized any loop only virtual SSA form needs to be updated.
605 ??? Also while we try hard to update loop-closed SSA form we fail
606 to properly do this in some corner-cases (see PR56286). */
607 rewrite_into_loop_closed_ssa (NULL
, TODO_update_ssa_only_virtuals
);
608 return TODO_cleanup_cfg
;
615 /* Entry point to the simduid cleanup pass. */
619 const pass_data pass_data_simduid_cleanup
=
621 GIMPLE_PASS
, /* type */
622 "simduid", /* name */
623 OPTGROUP_NONE
, /* optinfo_flags */
625 ( PROP_ssa
| PROP_cfg
), /* properties_required */
626 0, /* properties_provided */
627 0, /* properties_destroyed */
628 0, /* todo_flags_start */
629 0, /* todo_flags_finish */
632 class pass_simduid_cleanup
: public gimple_opt_pass
635 pass_simduid_cleanup (gcc::context
*ctxt
)
636 : gimple_opt_pass (pass_data_simduid_cleanup
, ctxt
)
639 /* opt_pass methods: */
640 opt_pass
* clone () { return new pass_simduid_cleanup (m_ctxt
); }
641 virtual bool gate (function
*fun
) { return fun
->has_simduid_loops
; }
642 virtual unsigned int execute (function
*);
644 }; // class pass_simduid_cleanup
647 pass_simduid_cleanup::execute (function
*fun
)
649 hash_table
<simd_array_to_simduid
> *simd_array_to_simduid_htab
= NULL
;
651 note_simd_array_uses (&simd_array_to_simduid_htab
);
653 /* Fold IFN_GOMP_SIMD_{VF,LANE,LAST_LANE} builtins. */
654 adjust_simduid_builtins (NULL
);
656 /* Shrink any "omp array simd" temporary arrays to the
657 actual vectorization factors. */
658 if (simd_array_to_simduid_htab
)
659 shrink_simd_arrays (simd_array_to_simduid_htab
, NULL
);
660 fun
->has_simduid_loops
= false;
667 make_pass_simduid_cleanup (gcc::context
*ctxt
)
669 return new pass_simduid_cleanup (ctxt
);
673 /* Entry point to basic block SLP phase. */
677 const pass_data pass_data_slp_vectorize
=
679 GIMPLE_PASS
, /* type */
681 OPTGROUP_LOOP
| OPTGROUP_VEC
, /* optinfo_flags */
682 TV_TREE_SLP_VECTORIZATION
, /* tv_id */
683 ( PROP_ssa
| PROP_cfg
), /* properties_required */
684 0, /* properties_provided */
685 0, /* properties_destroyed */
686 0, /* todo_flags_start */
687 TODO_update_ssa
, /* todo_flags_finish */
690 class pass_slp_vectorize
: public gimple_opt_pass
693 pass_slp_vectorize (gcc::context
*ctxt
)
694 : gimple_opt_pass (pass_data_slp_vectorize
, ctxt
)
697 /* opt_pass methods: */
698 opt_pass
* clone () { return new pass_slp_vectorize (m_ctxt
); }
699 virtual bool gate (function
*) { return flag_tree_slp_vectorize
!= 0; }
700 virtual unsigned int execute (function
*);
702 }; // class pass_slp_vectorize
705 pass_slp_vectorize::execute (function
*fun
)
709 bool in_loop_pipeline
= scev_initialized_p ();
710 if (!in_loop_pipeline
)
712 loop_optimizer_init (LOOPS_NORMAL
);
716 init_stmt_vec_info_vec ();
718 FOR_EACH_BB_FN (bb
, fun
)
720 vect_location
= find_bb_location (bb
);
722 if (vect_slp_analyze_bb (bb
))
724 if (!dbg_cnt (vect_slp
))
727 vect_slp_transform_bb (bb
);
728 if (dump_enabled_p ())
729 dump_printf_loc (MSG_OPTIMIZED_LOCATIONS
, vect_location
,
730 "basic block vectorized\n");
734 free_stmt_vec_info_vec ();
736 if (!in_loop_pipeline
)
739 loop_optimizer_finalize ();
748 make_pass_slp_vectorize (gcc::context
*ctxt
)
750 return new pass_slp_vectorize (ctxt
);
754 /* Increase alignment of global arrays to improve vectorization potential.
756 - Consider also structs that have an array field.
757 - Use ipa analysis to prune arrays that can't be vectorized?
758 This should involve global alignment analysis and in the future also
762 increase_alignment (void)
766 vect_location
= UNKNOWN_LOCATION
;
768 /* Increase the alignment of all global arrays for vectorization. */
769 FOR_EACH_DEFINED_VARIABLE (vnode
)
771 tree vectype
, decl
= vnode
->decl
;
773 unsigned int alignment
;
775 t
= TREE_TYPE (decl
);
776 if (TREE_CODE (t
) != ARRAY_TYPE
)
778 vectype
= get_vectype_for_scalar_type (strip_array_types (t
));
781 alignment
= TYPE_ALIGN (vectype
);
782 if (DECL_ALIGN (decl
) >= alignment
)
785 if (vect_can_force_dr_alignment_p (decl
, alignment
))
787 vnode
->increase_alignment (TYPE_ALIGN (vectype
));
788 dump_printf (MSG_NOTE
, "Increasing alignment of decl: ");
789 dump_generic_expr (MSG_NOTE
, TDF_SLIM
, decl
);
790 dump_printf (MSG_NOTE
, "\n");
799 const pass_data pass_data_ipa_increase_alignment
=
801 SIMPLE_IPA_PASS
, /* type */
802 "increase_alignment", /* name */
803 OPTGROUP_LOOP
| OPTGROUP_VEC
, /* optinfo_flags */
804 TV_IPA_OPT
, /* tv_id */
805 0, /* properties_required */
806 0, /* properties_provided */
807 0, /* properties_destroyed */
808 0, /* todo_flags_start */
809 0, /* todo_flags_finish */
812 class pass_ipa_increase_alignment
: public simple_ipa_opt_pass
815 pass_ipa_increase_alignment (gcc::context
*ctxt
)
816 : simple_ipa_opt_pass (pass_data_ipa_increase_alignment
, ctxt
)
819 /* opt_pass methods: */
820 virtual bool gate (function
*)
822 return flag_section_anchors
&& flag_tree_loop_vectorize
;
825 virtual unsigned int execute (function
*) { return increase_alignment (); }
827 }; // class pass_ipa_increase_alignment
831 simple_ipa_opt_pass
*
832 make_pass_ipa_increase_alignment (gcc::context
*ctxt
)
834 return new pass_ipa_increase_alignment (ctxt
);