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"
61 #include "hard-reg-set.h"
65 #include "tree-pass.h"
68 #include "tree-pretty-print.h"
71 #include "fold-const.h"
72 #include "stor-layout.h"
73 #include "internal-fn.h"
74 #include "gimple-iterator.h"
75 #include "gimple-walk.h"
76 #include "tree-ssa-loop-manip.h"
79 #include "tree-vectorizer.h"
80 #include "tree-ssa-propagate.h"
82 #include "gimple-fold.h"
83 #include "tree-scalar-evolution.h"
86 /* Loop or bb location. */
87 source_location vect_location
;
89 /* Vector mapping GIMPLE stmt to stmt_vec_info. */
90 vec
<stmt_vec_info
> stmt_vec_info_vec
;
92 /* For mapping simduid to vectorization factor. */
94 struct simduid_to_vf
: free_ptr_hash
<simduid_to_vf
>
99 /* hash_table support. */
100 static inline hashval_t
hash (const simduid_to_vf
*);
101 static inline int equal (const simduid_to_vf
*, const simduid_to_vf
*);
105 simduid_to_vf::hash (const simduid_to_vf
*p
)
111 simduid_to_vf::equal (const simduid_to_vf
*p1
, const simduid_to_vf
*p2
)
113 return p1
->simduid
== p2
->simduid
;
116 /* This hash maps the OMP simd array to the corresponding simduid used
117 to index into it. Like thus,
119 _7 = GOMP_SIMD_LANE (simduid.0)
125 This hash maps from the OMP simd array (D.1737[]) to DECL_UID of
128 struct simd_array_to_simduid
: free_ptr_hash
<simd_array_to_simduid
>
131 unsigned int simduid
;
133 /* hash_table support. */
134 static inline hashval_t
hash (const simd_array_to_simduid
*);
135 static inline int equal (const simd_array_to_simduid
*,
136 const simd_array_to_simduid
*);
140 simd_array_to_simduid::hash (const simd_array_to_simduid
*p
)
142 return DECL_UID (p
->decl
);
146 simd_array_to_simduid::equal (const simd_array_to_simduid
*p1
,
147 const simd_array_to_simduid
*p2
)
149 return p1
->decl
== p2
->decl
;
152 /* Fold IFN_GOMP_SIMD_LANE, IFN_GOMP_SIMD_VF, IFN_GOMP_SIMD_LAST_LANE,
153 into their corresponding constants and remove
154 IFN_GOMP_SIMD_ORDERED_{START,END}. */
157 adjust_simduid_builtins (hash_table
<simduid_to_vf
> *htab
)
161 FOR_EACH_BB_FN (bb
, cfun
)
163 gimple_stmt_iterator i
;
165 for (i
= gsi_start_bb (bb
); !gsi_end_p (i
); )
168 enum internal_fn ifn
;
169 gimple
*stmt
= gsi_stmt (i
);
171 if (!is_gimple_call (stmt
)
172 || !gimple_call_internal_p (stmt
))
177 ifn
= gimple_call_internal_fn (stmt
);
180 case IFN_GOMP_SIMD_LANE
:
181 case IFN_GOMP_SIMD_VF
:
182 case IFN_GOMP_SIMD_LAST_LANE
:
184 case IFN_GOMP_SIMD_ORDERED_START
:
185 case IFN_GOMP_SIMD_ORDERED_END
:
186 gsi_remove (&i
, true);
187 unlink_stmt_vdef (stmt
);
193 tree arg
= gimple_call_arg (stmt
, 0);
194 gcc_assert (arg
!= NULL_TREE
);
195 gcc_assert (TREE_CODE (arg
) == SSA_NAME
);
196 simduid_to_vf
*p
= NULL
, data
;
197 data
.simduid
= DECL_UID (SSA_NAME_VAR (arg
));
200 p
= htab
->find (&data
);
206 case IFN_GOMP_SIMD_VF
:
207 t
= build_int_cst (unsigned_type_node
, vf
);
209 case IFN_GOMP_SIMD_LANE
:
210 t
= build_int_cst (unsigned_type_node
, 0);
212 case IFN_GOMP_SIMD_LAST_LANE
:
213 t
= gimple_call_arg (stmt
, 1);
218 update_call_from_tree (&i
, t
);
224 /* Helper structure for note_simd_array_uses. */
226 struct note_simd_array_uses_struct
228 hash_table
<simd_array_to_simduid
> **htab
;
229 unsigned int simduid
;
232 /* Callback for note_simd_array_uses, called through walk_gimple_op. */
235 note_simd_array_uses_cb (tree
*tp
, int *walk_subtrees
, void *data
)
237 struct walk_stmt_info
*wi
= (struct walk_stmt_info
*) data
;
238 struct note_simd_array_uses_struct
*ns
239 = (struct note_simd_array_uses_struct
*) wi
->info
;
244 && lookup_attribute ("omp simd array", DECL_ATTRIBUTES (*tp
))
245 && DECL_CONTEXT (*tp
) == current_function_decl
)
247 simd_array_to_simduid data
;
249 *ns
->htab
= new hash_table
<simd_array_to_simduid
> (15);
251 data
.simduid
= ns
->simduid
;
252 simd_array_to_simduid
**slot
= (*ns
->htab
)->find_slot (&data
, INSERT
);
255 simd_array_to_simduid
*p
= XNEW (simd_array_to_simduid
);
259 else if ((*slot
)->simduid
!= ns
->simduid
)
260 (*slot
)->simduid
= -1U;
266 /* Find "omp simd array" temporaries and map them to corresponding
270 note_simd_array_uses (hash_table
<simd_array_to_simduid
> **htab
)
273 gimple_stmt_iterator gsi
;
274 struct walk_stmt_info wi
;
275 struct note_simd_array_uses_struct ns
;
277 memset (&wi
, 0, sizeof (wi
));
281 FOR_EACH_BB_FN (bb
, cfun
)
282 for (gsi
= gsi_start_bb (bb
); !gsi_end_p (gsi
); gsi_next (&gsi
))
284 gimple
*stmt
= gsi_stmt (gsi
);
285 if (!is_gimple_call (stmt
) || !gimple_call_internal_p (stmt
))
287 switch (gimple_call_internal_fn (stmt
))
289 case IFN_GOMP_SIMD_LANE
:
290 case IFN_GOMP_SIMD_VF
:
291 case IFN_GOMP_SIMD_LAST_LANE
:
296 tree lhs
= gimple_call_lhs (stmt
);
297 if (lhs
== NULL_TREE
)
299 imm_use_iterator use_iter
;
301 ns
.simduid
= DECL_UID (SSA_NAME_VAR (gimple_call_arg (stmt
, 0)));
302 FOR_EACH_IMM_USE_STMT (use_stmt
, use_iter
, lhs
)
303 if (!is_gimple_debug (use_stmt
))
304 walk_gimple_op (use_stmt
, note_simd_array_uses_cb
, &wi
);
308 /* Shrink arrays with "omp simd array" attribute to the corresponding
309 vectorization factor. */
313 (hash_table
<simd_array_to_simduid
> *simd_array_to_simduid_htab
,
314 hash_table
<simduid_to_vf
> *simduid_to_vf_htab
)
316 for (hash_table
<simd_array_to_simduid
>::iterator iter
317 = simd_array_to_simduid_htab
->begin ();
318 iter
!= simd_array_to_simduid_htab
->end (); ++iter
)
319 if ((*iter
)->simduid
!= -1U)
321 tree decl
= (*iter
)->decl
;
323 if (simduid_to_vf_htab
)
325 simduid_to_vf
*p
= NULL
, data
;
326 data
.simduid
= (*iter
)->simduid
;
327 p
= simduid_to_vf_htab
->find (&data
);
332 = build_array_type_nelts (TREE_TYPE (TREE_TYPE (decl
)), vf
);
333 TREE_TYPE (decl
) = atype
;
334 relayout_decl (decl
);
337 delete simd_array_to_simduid_htab
;
340 /* A helper function to free data refs. */
343 vect_destroy_datarefs (vec_info
*vinfo
)
345 struct data_reference
*dr
;
348 FOR_EACH_VEC_ELT (vinfo
->datarefs
, i
, dr
)
355 free_data_refs (vinfo
->datarefs
);
359 /* If LOOP has been versioned during ifcvt, return the internal call
363 vect_loop_vectorized_call (struct loop
*loop
)
365 basic_block bb
= loop_preheader_edge (loop
)->src
;
372 if (!single_pred_p (bb
))
374 bb
= single_pred (bb
);
377 if (g
&& gimple_code (g
) == GIMPLE_COND
)
379 gimple_stmt_iterator gsi
= gsi_for_stmt (g
);
381 if (!gsi_end_p (gsi
))
384 if (is_gimple_call (g
)
385 && gimple_call_internal_p (g
)
386 && gimple_call_internal_fn (g
) == IFN_LOOP_VECTORIZED
387 && (tree_to_shwi (gimple_call_arg (g
, 0)) == loop
->num
388 || tree_to_shwi (gimple_call_arg (g
, 1)) == loop
->num
))
395 /* Fold LOOP_VECTORIZED internal call G to VALUE and
396 update any immediate uses of it's LHS. */
399 fold_loop_vectorized_call (gimple
*g
, tree value
)
401 tree lhs
= gimple_call_lhs (g
);
403 imm_use_iterator iter
;
405 gimple_stmt_iterator gsi
= gsi_for_stmt (g
);
407 update_call_from_tree (&gsi
, value
);
408 FOR_EACH_IMM_USE_STMT (use_stmt
, iter
, lhs
)
410 FOR_EACH_IMM_USE_ON_STMT (use_p
, iter
)
411 SET_USE (use_p
, value
);
412 update_stmt (use_stmt
);
415 /* Set the uids of all the statements in basic blocks inside loop
416 represented by LOOP_VINFO. LOOP_VECTORIZED_CALL is the internal
417 call guarding the loop which has been if converted. */
419 set_uid_loop_bbs (loop_vec_info loop_vinfo
, gimple
*loop_vectorized_call
)
421 tree arg
= gimple_call_arg (loop_vectorized_call
, 1);
424 struct loop
*scalar_loop
= get_loop (cfun
, tree_to_shwi (arg
));
426 LOOP_VINFO_SCALAR_LOOP (loop_vinfo
) = scalar_loop
;
427 gcc_checking_assert (vect_loop_vectorized_call
428 (LOOP_VINFO_SCALAR_LOOP (loop_vinfo
))
429 == loop_vectorized_call
);
430 bbs
= get_loop_body (scalar_loop
);
431 for (i
= 0; i
< scalar_loop
->num_nodes
; i
++)
433 basic_block bb
= bbs
[i
];
434 gimple_stmt_iterator gsi
;
435 for (gsi
= gsi_start_phis (bb
); !gsi_end_p (gsi
); gsi_next (&gsi
))
437 gimple
*phi
= gsi_stmt (gsi
);
438 gimple_set_uid (phi
, 0);
440 for (gsi
= gsi_start_bb (bb
); !gsi_end_p (gsi
); gsi_next (&gsi
))
442 gimple
*stmt
= gsi_stmt (gsi
);
443 gimple_set_uid (stmt
, 0);
449 /* Function vectorize_loops.
451 Entry point to loop vectorization phase. */
454 vectorize_loops (void)
457 unsigned int num_vectorized_loops
= 0;
458 unsigned int vect_loops_num
;
460 hash_table
<simduid_to_vf
> *simduid_to_vf_htab
= NULL
;
461 hash_table
<simd_array_to_simduid
> *simd_array_to_simduid_htab
= NULL
;
462 bool any_ifcvt_loops
= false;
465 vect_loops_num
= number_of_loops (cfun
);
467 /* Bail out if there are no loops. */
468 if (vect_loops_num
<= 1)
471 if (cfun
->has_simduid_loops
)
472 note_simd_array_uses (&simd_array_to_simduid_htab
);
474 init_stmt_vec_info_vec ();
476 /* ----------- Analyze loops. ----------- */
478 /* If some loop was duplicated, it gets bigger number
479 than all previously defined loops. This fact allows us to run
480 only over initial loops skipping newly generated ones. */
481 FOR_EACH_LOOP (loop
, 0)
482 if (loop
->dont_vectorize
)
483 any_ifcvt_loops
= true;
484 else if ((flag_tree_loop_vectorize
485 && optimize_loop_nest_for_speed_p (loop
))
486 || loop
->force_vectorize
)
488 loop_vec_info loop_vinfo
;
489 vect_location
= find_loop_location (loop
);
490 if (LOCATION_LOCUS (vect_location
) != UNKNOWN_LOCATION
491 && dump_enabled_p ())
492 dump_printf (MSG_NOTE
, "\nAnalyzing loop at %s:%d\n",
493 LOCATION_FILE (vect_location
),
494 LOCATION_LINE (vect_location
));
496 loop_vinfo
= vect_analyze_loop (loop
);
497 loop
->aux
= loop_vinfo
;
499 if (!loop_vinfo
|| !LOOP_VINFO_VECTORIZABLE_P (loop_vinfo
))
502 if (!dbg_cnt (vect_loop
))
505 gimple
*loop_vectorized_call
= vect_loop_vectorized_call (loop
);
506 if (loop_vectorized_call
)
507 set_uid_loop_bbs (loop_vinfo
, loop_vectorized_call
);
508 if (LOCATION_LOCUS (vect_location
) != UNKNOWN_LOCATION
509 && dump_enabled_p ())
510 dump_printf_loc (MSG_OPTIMIZED_LOCATIONS
, vect_location
,
511 "loop vectorized\n");
512 vect_transform_loop (loop_vinfo
);
513 num_vectorized_loops
++;
514 /* Now that the loop has been vectorized, allow it to be unrolled
516 loop
->force_vectorize
= false;
520 simduid_to_vf
*simduid_to_vf_data
= XNEW (simduid_to_vf
);
521 if (!simduid_to_vf_htab
)
522 simduid_to_vf_htab
= new hash_table
<simduid_to_vf
> (15);
523 simduid_to_vf_data
->simduid
= DECL_UID (loop
->simduid
);
524 simduid_to_vf_data
->vf
= loop_vinfo
->vectorization_factor
;
525 *simduid_to_vf_htab
->find_slot (simduid_to_vf_data
, INSERT
)
526 = simduid_to_vf_data
;
529 if (loop_vectorized_call
)
531 fold_loop_vectorized_call (loop_vectorized_call
, boolean_true_node
);
532 ret
|= TODO_cleanup_cfg
;
536 vect_location
= UNKNOWN_LOCATION
;
538 statistics_counter_event (cfun
, "Vectorized loops", num_vectorized_loops
);
539 if (dump_enabled_p ()
540 || (num_vectorized_loops
> 0 && dump_enabled_p ()))
541 dump_printf_loc (MSG_NOTE
, vect_location
,
542 "vectorized %u loops in function.\n",
543 num_vectorized_loops
);
545 /* ----------- Finalize. ----------- */
548 for (i
= 1; i
< vect_loops_num
; i
++)
550 loop
= get_loop (cfun
, i
);
551 if (loop
&& loop
->dont_vectorize
)
553 gimple
*g
= vect_loop_vectorized_call (loop
);
556 fold_loop_vectorized_call (g
, boolean_false_node
);
557 ret
|= TODO_cleanup_cfg
;
562 for (i
= 1; i
< vect_loops_num
; i
++)
564 loop_vec_info loop_vinfo
;
566 loop
= get_loop (cfun
, i
);
569 loop_vinfo
= (loop_vec_info
) loop
->aux
;
570 destroy_loop_vec_info (loop_vinfo
, true);
574 free_stmt_vec_info_vec ();
576 /* Fold IFN_GOMP_SIMD_{VF,LANE,LAST_LANE,ORDERED_{START,END}} builtins. */
577 if (cfun
->has_simduid_loops
)
578 adjust_simduid_builtins (simduid_to_vf_htab
);
580 /* Shrink any "omp array simd" temporary arrays to the
581 actual vectorization factors. */
582 if (simd_array_to_simduid_htab
)
583 shrink_simd_arrays (simd_array_to_simduid_htab
, simduid_to_vf_htab
);
584 delete simduid_to_vf_htab
;
585 cfun
->has_simduid_loops
= false;
587 if (num_vectorized_loops
> 0)
589 /* If we vectorized any loop only virtual SSA form needs to be updated.
590 ??? Also while we try hard to update loop-closed SSA form we fail
591 to properly do this in some corner-cases (see PR56286). */
592 rewrite_into_loop_closed_ssa (NULL
, TODO_update_ssa_only_virtuals
);
593 return TODO_cleanup_cfg
;
600 /* Entry point to the simduid cleanup pass. */
604 const pass_data pass_data_simduid_cleanup
=
606 GIMPLE_PASS
, /* type */
607 "simduid", /* name */
608 OPTGROUP_NONE
, /* optinfo_flags */
610 ( PROP_ssa
| PROP_cfg
), /* properties_required */
611 0, /* properties_provided */
612 0, /* properties_destroyed */
613 0, /* todo_flags_start */
614 0, /* todo_flags_finish */
617 class pass_simduid_cleanup
: public gimple_opt_pass
620 pass_simduid_cleanup (gcc::context
*ctxt
)
621 : gimple_opt_pass (pass_data_simduid_cleanup
, ctxt
)
624 /* opt_pass methods: */
625 opt_pass
* clone () { return new pass_simduid_cleanup (m_ctxt
); }
626 virtual bool gate (function
*fun
) { return fun
->has_simduid_loops
; }
627 virtual unsigned int execute (function
*);
629 }; // class pass_simduid_cleanup
632 pass_simduid_cleanup::execute (function
*fun
)
634 hash_table
<simd_array_to_simduid
> *simd_array_to_simduid_htab
= NULL
;
636 note_simd_array_uses (&simd_array_to_simduid_htab
);
638 /* Fold IFN_GOMP_SIMD_{VF,LANE,LAST_LANE,ORDERED_{START,END}} builtins. */
639 adjust_simduid_builtins (NULL
);
641 /* Shrink any "omp array simd" temporary arrays to the
642 actual vectorization factors. */
643 if (simd_array_to_simduid_htab
)
644 shrink_simd_arrays (simd_array_to_simduid_htab
, NULL
);
645 fun
->has_simduid_loops
= false;
652 make_pass_simduid_cleanup (gcc::context
*ctxt
)
654 return new pass_simduid_cleanup (ctxt
);
658 /* Entry point to basic block SLP phase. */
662 const pass_data pass_data_slp_vectorize
=
664 GIMPLE_PASS
, /* type */
666 OPTGROUP_LOOP
| OPTGROUP_VEC
, /* optinfo_flags */
667 TV_TREE_SLP_VECTORIZATION
, /* tv_id */
668 ( PROP_ssa
| PROP_cfg
), /* properties_required */
669 0, /* properties_provided */
670 0, /* properties_destroyed */
671 0, /* todo_flags_start */
672 TODO_update_ssa
, /* todo_flags_finish */
675 class pass_slp_vectorize
: public gimple_opt_pass
678 pass_slp_vectorize (gcc::context
*ctxt
)
679 : gimple_opt_pass (pass_data_slp_vectorize
, ctxt
)
682 /* opt_pass methods: */
683 opt_pass
* clone () { return new pass_slp_vectorize (m_ctxt
); }
684 virtual bool gate (function
*) { return flag_tree_slp_vectorize
!= 0; }
685 virtual unsigned int execute (function
*);
687 }; // class pass_slp_vectorize
690 pass_slp_vectorize::execute (function
*fun
)
694 bool in_loop_pipeline
= scev_initialized_p ();
695 if (!in_loop_pipeline
)
697 loop_optimizer_init (LOOPS_NORMAL
);
701 init_stmt_vec_info_vec ();
703 FOR_EACH_BB_FN (bb
, fun
)
705 vect_location
= find_bb_location (bb
);
707 if (vect_slp_analyze_bb (bb
))
709 if (!dbg_cnt (vect_slp
))
712 vect_slp_transform_bb (bb
);
713 if (dump_enabled_p ())
714 dump_printf_loc (MSG_OPTIMIZED_LOCATIONS
, vect_location
,
715 "basic block vectorized\n");
719 free_stmt_vec_info_vec ();
721 if (!in_loop_pipeline
)
724 loop_optimizer_finalize ();
733 make_pass_slp_vectorize (gcc::context
*ctxt
)
735 return new pass_slp_vectorize (ctxt
);
739 /* Increase alignment of global arrays to improve vectorization potential.
741 - Consider also structs that have an array field.
742 - Use ipa analysis to prune arrays that can't be vectorized?
743 This should involve global alignment analysis and in the future also
747 increase_alignment (void)
751 vect_location
= UNKNOWN_LOCATION
;
753 /* Increase the alignment of all global arrays for vectorization. */
754 FOR_EACH_DEFINED_VARIABLE (vnode
)
756 tree vectype
, decl
= vnode
->decl
;
758 unsigned int alignment
;
760 t
= TREE_TYPE (decl
);
761 if (TREE_CODE (t
) != ARRAY_TYPE
)
763 vectype
= get_vectype_for_scalar_type (strip_array_types (t
));
766 alignment
= TYPE_ALIGN (vectype
);
767 if (DECL_ALIGN (decl
) >= alignment
)
770 if (vect_can_force_dr_alignment_p (decl
, alignment
))
772 vnode
->increase_alignment (TYPE_ALIGN (vectype
));
773 dump_printf (MSG_NOTE
, "Increasing alignment of decl: ");
774 dump_generic_expr (MSG_NOTE
, TDF_SLIM
, decl
);
775 dump_printf (MSG_NOTE
, "\n");
784 const pass_data pass_data_ipa_increase_alignment
=
786 SIMPLE_IPA_PASS
, /* type */
787 "increase_alignment", /* name */
788 OPTGROUP_LOOP
| OPTGROUP_VEC
, /* optinfo_flags */
789 TV_IPA_OPT
, /* tv_id */
790 0, /* properties_required */
791 0, /* properties_provided */
792 0, /* properties_destroyed */
793 0, /* todo_flags_start */
794 0, /* todo_flags_finish */
797 class pass_ipa_increase_alignment
: public simple_ipa_opt_pass
800 pass_ipa_increase_alignment (gcc::context
*ctxt
)
801 : simple_ipa_opt_pass (pass_data_ipa_increase_alignment
, ctxt
)
804 /* opt_pass methods: */
805 virtual bool gate (function
*)
807 return flag_section_anchors
&& flag_tree_loop_vectorize
;
810 virtual unsigned int execute (function
*) { return increase_alignment (); }
812 }; // class pass_ipa_increase_alignment
816 simple_ipa_opt_pass
*
817 make_pass_ipa_increase_alignment (gcc::context
*ctxt
)
819 return new pass_ipa_increase_alignment (ctxt
);