2 Copyright (C) 2003-2020 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"
64 #include "tree-pass.h"
67 #include "fold-const.h"
68 #include "stor-layout.h"
69 #include "gimple-iterator.h"
70 #include "gimple-walk.h"
71 #include "tree-ssa-loop-manip.h"
72 #include "tree-ssa-loop-niter.h"
75 #include "tree-vectorizer.h"
76 #include "tree-ssa-propagate.h"
78 #include "tree-scalar-evolution.h"
79 #include "stringpool.h"
81 #include "gimple-pretty-print.h"
82 #include "opt-problem.h"
83 #include "internal-fn.h"
86 /* Loop or bb location, with hotness information. */
87 dump_user_location_t vect_location
;
89 /* auto_purge_vect_location's dtor: reset the vect_location
90 global, to avoid stale location_t values that could reference
93 auto_purge_vect_location::~auto_purge_vect_location ()
95 vect_location
= dump_user_location_t ();
98 /* Dump a cost entry according to args to F. */
101 dump_stmt_cost (FILE *f
, void *data
, int count
, enum vect_cost_for_stmt kind
,
102 stmt_vec_info stmt_info
, tree
, int misalign
, unsigned cost
,
103 enum vect_cost_model_location where
)
105 fprintf (f
, "%p ", data
);
108 print_gimple_expr (f
, STMT_VINFO_STMT (stmt_info
), 0, TDF_SLIM
);
112 fprintf (f
, "<unknown> ");
113 fprintf (f
, "%d times ", count
);
114 const char *ks
= "unknown";
132 case vector_gather_load
:
133 ks
= "vector_gather_load";
136 ks
= "unaligned_load";
138 case unaligned_store
:
139 ks
= "unaligned_store";
144 case vector_scatter_store
:
145 ks
= "vector_scatter_store";
148 ks
= "vec_to_scalar";
151 ks
= "scalar_to_vec";
153 case cond_branch_not_taken
:
154 ks
= "cond_branch_not_taken";
156 case cond_branch_taken
:
157 ks
= "cond_branch_taken";
162 case vec_promote_demote
:
163 ks
= "vec_promote_demote";
166 ks
= "vec_construct";
169 fprintf (f
, "%s ", ks
);
170 if (kind
== unaligned_load
|| kind
== unaligned_store
)
171 fprintf (f
, "(misalign %d) ", misalign
);
172 fprintf (f
, "costs %u ", cost
);
173 const char *ws
= "unknown";
186 fprintf (f
, "in %s\n", ws
);
189 /* For mapping simduid to vectorization factor. */
191 class simduid_to_vf
: public free_ptr_hash
<simduid_to_vf
>
194 unsigned int simduid
;
197 /* hash_table support. */
198 static inline hashval_t
hash (const simduid_to_vf
*);
199 static inline int equal (const simduid_to_vf
*, const simduid_to_vf
*);
203 simduid_to_vf::hash (const simduid_to_vf
*p
)
209 simduid_to_vf::equal (const simduid_to_vf
*p1
, const simduid_to_vf
*p2
)
211 return p1
->simduid
== p2
->simduid
;
214 /* This hash maps the OMP simd array to the corresponding simduid used
215 to index into it. Like thus,
217 _7 = GOMP_SIMD_LANE (simduid.0)
223 This hash maps from the OMP simd array (D.1737[]) to DECL_UID of
226 struct simd_array_to_simduid
: free_ptr_hash
<simd_array_to_simduid
>
229 unsigned int simduid
;
231 /* hash_table support. */
232 static inline hashval_t
hash (const simd_array_to_simduid
*);
233 static inline int equal (const simd_array_to_simduid
*,
234 const simd_array_to_simduid
*);
238 simd_array_to_simduid::hash (const simd_array_to_simduid
*p
)
240 return DECL_UID (p
->decl
);
244 simd_array_to_simduid::equal (const simd_array_to_simduid
*p1
,
245 const simd_array_to_simduid
*p2
)
247 return p1
->decl
== p2
->decl
;
250 /* Fold IFN_GOMP_SIMD_LANE, IFN_GOMP_SIMD_VF, IFN_GOMP_SIMD_LAST_LANE,
251 into their corresponding constants and remove
252 IFN_GOMP_SIMD_ORDERED_{START,END}. */
255 adjust_simduid_builtins (hash_table
<simduid_to_vf
> *htab
)
259 FOR_EACH_BB_FN (bb
, cfun
)
261 gimple_stmt_iterator i
;
263 for (i
= gsi_start_bb (bb
); !gsi_end_p (i
); )
266 enum internal_fn ifn
;
267 gimple
*stmt
= gsi_stmt (i
);
269 if (!is_gimple_call (stmt
)
270 || !gimple_call_internal_p (stmt
))
275 ifn
= gimple_call_internal_fn (stmt
);
278 case IFN_GOMP_SIMD_LANE
:
279 case IFN_GOMP_SIMD_VF
:
280 case IFN_GOMP_SIMD_LAST_LANE
:
282 case IFN_GOMP_SIMD_ORDERED_START
:
283 case IFN_GOMP_SIMD_ORDERED_END
:
284 if (integer_onep (gimple_call_arg (stmt
, 0)))
286 enum built_in_function bcode
287 = (ifn
== IFN_GOMP_SIMD_ORDERED_START
288 ? BUILT_IN_GOMP_ORDERED_START
289 : BUILT_IN_GOMP_ORDERED_END
);
291 = gimple_build_call (builtin_decl_explicit (bcode
), 0);
292 gimple_move_vops (g
, stmt
);
293 gsi_replace (&i
, g
, true);
296 gsi_remove (&i
, true);
297 unlink_stmt_vdef (stmt
);
303 tree arg
= gimple_call_arg (stmt
, 0);
304 gcc_assert (arg
!= NULL_TREE
);
305 gcc_assert (TREE_CODE (arg
) == SSA_NAME
);
306 simduid_to_vf
*p
= NULL
, data
;
307 data
.simduid
= DECL_UID (SSA_NAME_VAR (arg
));
308 /* Need to nullify loop safelen field since it's value is not
309 valid after transformation. */
310 if (bb
->loop_father
&& bb
->loop_father
->safelen
> 0)
311 bb
->loop_father
->safelen
= 0;
314 p
= htab
->find (&data
);
320 case IFN_GOMP_SIMD_VF
:
321 t
= build_int_cst (unsigned_type_node
, vf
);
323 case IFN_GOMP_SIMD_LANE
:
324 t
= build_int_cst (unsigned_type_node
, 0);
326 case IFN_GOMP_SIMD_LAST_LANE
:
327 t
= gimple_call_arg (stmt
, 1);
332 tree lhs
= gimple_call_lhs (stmt
);
334 replace_uses_by (lhs
, t
);
336 gsi_remove (&i
, true);
341 /* Helper structure for note_simd_array_uses. */
343 struct note_simd_array_uses_struct
345 hash_table
<simd_array_to_simduid
> **htab
;
346 unsigned int simduid
;
349 /* Callback for note_simd_array_uses, called through walk_gimple_op. */
352 note_simd_array_uses_cb (tree
*tp
, int *walk_subtrees
, void *data
)
354 struct walk_stmt_info
*wi
= (struct walk_stmt_info
*) data
;
355 struct note_simd_array_uses_struct
*ns
356 = (struct note_simd_array_uses_struct
*) wi
->info
;
361 && lookup_attribute ("omp simd array", DECL_ATTRIBUTES (*tp
))
362 && DECL_CONTEXT (*tp
) == current_function_decl
)
364 simd_array_to_simduid data
;
366 *ns
->htab
= new hash_table
<simd_array_to_simduid
> (15);
368 data
.simduid
= ns
->simduid
;
369 simd_array_to_simduid
**slot
= (*ns
->htab
)->find_slot (&data
, INSERT
);
372 simd_array_to_simduid
*p
= XNEW (simd_array_to_simduid
);
376 else if ((*slot
)->simduid
!= ns
->simduid
)
377 (*slot
)->simduid
= -1U;
383 /* Find "omp simd array" temporaries and map them to corresponding
387 note_simd_array_uses (hash_table
<simd_array_to_simduid
> **htab
)
390 gimple_stmt_iterator gsi
;
391 struct walk_stmt_info wi
;
392 struct note_simd_array_uses_struct ns
;
394 memset (&wi
, 0, sizeof (wi
));
398 FOR_EACH_BB_FN (bb
, cfun
)
399 for (gsi
= gsi_start_bb (bb
); !gsi_end_p (gsi
); gsi_next (&gsi
))
401 gimple
*stmt
= gsi_stmt (gsi
);
402 if (!is_gimple_call (stmt
) || !gimple_call_internal_p (stmt
))
404 switch (gimple_call_internal_fn (stmt
))
406 case IFN_GOMP_SIMD_LANE
:
407 case IFN_GOMP_SIMD_VF
:
408 case IFN_GOMP_SIMD_LAST_LANE
:
413 tree lhs
= gimple_call_lhs (stmt
);
414 if (lhs
== NULL_TREE
)
416 imm_use_iterator use_iter
;
418 ns
.simduid
= DECL_UID (SSA_NAME_VAR (gimple_call_arg (stmt
, 0)));
419 FOR_EACH_IMM_USE_STMT (use_stmt
, use_iter
, lhs
)
420 if (!is_gimple_debug (use_stmt
))
421 walk_gimple_op (use_stmt
, note_simd_array_uses_cb
, &wi
);
425 /* Shrink arrays with "omp simd array" attribute to the corresponding
426 vectorization factor. */
430 (hash_table
<simd_array_to_simduid
> *simd_array_to_simduid_htab
,
431 hash_table
<simduid_to_vf
> *simduid_to_vf_htab
)
433 for (hash_table
<simd_array_to_simduid
>::iterator iter
434 = simd_array_to_simduid_htab
->begin ();
435 iter
!= simd_array_to_simduid_htab
->end (); ++iter
)
436 if ((*iter
)->simduid
!= -1U)
438 tree decl
= (*iter
)->decl
;
440 if (simduid_to_vf_htab
)
442 simduid_to_vf
*p
= NULL
, data
;
443 data
.simduid
= (*iter
)->simduid
;
444 p
= simduid_to_vf_htab
->find (&data
);
449 = build_array_type_nelts (TREE_TYPE (TREE_TYPE (decl
)), vf
);
450 TREE_TYPE (decl
) = atype
;
451 relayout_decl (decl
);
454 delete simd_array_to_simduid_htab
;
457 /* Initialize the vec_info with kind KIND_IN and target cost data
458 TARGET_COST_DATA_IN. */
460 vec_info::vec_info (vec_info::vec_kind kind_in
, void *target_cost_data_in
,
461 vec_info_shared
*shared_
)
464 target_cost_data (target_cost_data_in
)
466 stmt_vec_infos
.create (50);
469 vec_info::~vec_info ()
471 slp_instance instance
;
474 FOR_EACH_VEC_ELT (slp_instances
, i
, instance
)
475 vect_free_slp_instance (instance
, true);
477 destroy_cost_data (target_cost_data
);
478 free_stmt_vec_infos ();
481 vec_info_shared::vec_info_shared ()
483 datarefs_copy (vNULL
),
488 vec_info_shared::~vec_info_shared ()
490 free_data_refs (datarefs
);
491 free_dependence_relations (ddrs
);
492 datarefs_copy
.release ();
496 vec_info_shared::save_datarefs ()
500 datarefs_copy
.reserve_exact (datarefs
.length ());
501 for (unsigned i
= 0; i
< datarefs
.length (); ++i
)
502 datarefs_copy
.quick_push (*datarefs
[i
]);
506 vec_info_shared::check_datarefs ()
510 gcc_assert (datarefs
.length () == datarefs_copy
.length ());
511 for (unsigned i
= 0; i
< datarefs
.length (); ++i
)
512 if (memcmp (&datarefs_copy
[i
], datarefs
[i
], sizeof (data_reference
)) != 0)
516 /* Record that STMT belongs to the vectorizable region. Create and return
517 an associated stmt_vec_info. */
520 vec_info::add_stmt (gimple
*stmt
)
522 stmt_vec_info res
= new_stmt_vec_info (stmt
);
523 set_vinfo_for_stmt (stmt
, res
);
527 /* If STMT has an associated stmt_vec_info, return that vec_info, otherwise
528 return null. It is safe to call this function on any statement, even if
529 it might not be part of the vectorizable region. */
532 vec_info::lookup_stmt (gimple
*stmt
)
534 unsigned int uid
= gimple_uid (stmt
);
535 if (uid
> 0 && uid
- 1 < stmt_vec_infos
.length ())
537 stmt_vec_info res
= stmt_vec_infos
[uid
- 1];
538 if (res
&& res
->stmt
== stmt
)
544 /* If NAME is an SSA_NAME and its definition has an associated stmt_vec_info,
545 return that stmt_vec_info, otherwise return null. It is safe to call
546 this on arbitrary operands. */
549 vec_info::lookup_def (tree name
)
551 if (TREE_CODE (name
) == SSA_NAME
552 && !SSA_NAME_IS_DEFAULT_DEF (name
))
553 return lookup_stmt (SSA_NAME_DEF_STMT (name
));
557 /* See whether there is a single non-debug statement that uses LHS and
558 whether that statement has an associated stmt_vec_info. Return the
559 stmt_vec_info if so, otherwise return null. */
562 vec_info::lookup_single_use (tree lhs
)
566 if (single_imm_use (lhs
, &dummy
, &use_stmt
))
567 return lookup_stmt (use_stmt
);
571 /* Return vectorization information about DR. */
574 vec_info::lookup_dr (data_reference
*dr
)
576 stmt_vec_info stmt_info
= lookup_stmt (DR_STMT (dr
));
577 /* DR_STMT should never refer to a stmt in a pattern replacement. */
578 gcc_checking_assert (!is_pattern_stmt_p (stmt_info
));
579 return STMT_VINFO_DR_INFO (stmt_info
->dr_aux
.stmt
);
582 /* Record that NEW_STMT_INFO now implements the same data reference
586 vec_info::move_dr (stmt_vec_info new_stmt_info
, stmt_vec_info old_stmt_info
)
588 gcc_assert (!is_pattern_stmt_p (old_stmt_info
));
589 STMT_VINFO_DR_INFO (old_stmt_info
)->stmt
= new_stmt_info
;
590 new_stmt_info
->dr_aux
= old_stmt_info
->dr_aux
;
591 STMT_VINFO_DR_WRT_VEC_LOOP (new_stmt_info
)
592 = STMT_VINFO_DR_WRT_VEC_LOOP (old_stmt_info
);
593 STMT_VINFO_GATHER_SCATTER_P (new_stmt_info
)
594 = STMT_VINFO_GATHER_SCATTER_P (old_stmt_info
);
597 /* Permanently remove the statement described by STMT_INFO from the
601 vec_info::remove_stmt (stmt_vec_info stmt_info
)
603 gcc_assert (!stmt_info
->pattern_stmt_p
);
604 set_vinfo_for_stmt (stmt_info
->stmt
, NULL
);
605 gimple_stmt_iterator si
= gsi_for_stmt (stmt_info
->stmt
);
606 unlink_stmt_vdef (stmt_info
->stmt
);
607 gsi_remove (&si
, true);
608 release_defs (stmt_info
->stmt
);
609 free_stmt_vec_info (stmt_info
);
612 /* Replace the statement at GSI by NEW_STMT, both the vectorization
613 information and the function itself. STMT_INFO describes the statement
617 vec_info::replace_stmt (gimple_stmt_iterator
*gsi
, stmt_vec_info stmt_info
,
620 gimple
*old_stmt
= stmt_info
->stmt
;
621 gcc_assert (!stmt_info
->pattern_stmt_p
&& old_stmt
== gsi_stmt (*gsi
));
622 set_vinfo_for_stmt (old_stmt
, NULL
);
623 set_vinfo_for_stmt (new_stmt
, stmt_info
);
624 stmt_info
->stmt
= new_stmt
;
625 gsi_replace (gsi
, new_stmt
, true);
628 /* Create and initialize a new stmt_vec_info struct for STMT. */
631 vec_info::new_stmt_vec_info (gimple
*stmt
)
633 stmt_vec_info res
= XCNEW (class _stmt_vec_info
);
636 STMT_VINFO_TYPE (res
) = undef_vec_info_type
;
637 STMT_VINFO_RELEVANT (res
) = vect_unused_in_scope
;
638 STMT_VINFO_VECTORIZABLE (res
) = true;
639 STMT_VINFO_REDUC_TYPE (res
) = TREE_CODE_REDUCTION
;
640 STMT_VINFO_REDUC_CODE (res
) = ERROR_MARK
;
641 STMT_VINFO_REDUC_FN (res
) = IFN_LAST
;
642 STMT_VINFO_REDUC_IDX (res
) = -1;
643 STMT_VINFO_SLP_VECT_ONLY (res
) = false;
645 if (gimple_code (stmt
) == GIMPLE_PHI
646 && is_loop_header_bb_p (gimple_bb (stmt
)))
647 STMT_VINFO_DEF_TYPE (res
) = vect_unknown_def_type
;
649 STMT_VINFO_DEF_TYPE (res
) = vect_internal_def
;
651 STMT_VINFO_SAME_ALIGN_REFS (res
).create (0);
652 STMT_SLP_TYPE (res
) = loop_vect
;
654 /* This is really "uninitialized" until vect_compute_data_ref_alignment. */
655 res
->dr_aux
.misalignment
= DR_MISALIGNMENT_UNINITIALIZED
;
660 /* Associate STMT with INFO. */
663 vec_info::set_vinfo_for_stmt (gimple
*stmt
, stmt_vec_info info
)
665 unsigned int uid
= gimple_uid (stmt
);
668 gcc_checking_assert (info
);
669 uid
= stmt_vec_infos
.length () + 1;
670 gimple_set_uid (stmt
, uid
);
671 stmt_vec_infos
.safe_push (info
);
675 gcc_checking_assert (info
== NULL
);
676 stmt_vec_infos
[uid
- 1] = info
;
680 /* Free the contents of stmt_vec_infos. */
683 vec_info::free_stmt_vec_infos (void)
687 FOR_EACH_VEC_ELT (stmt_vec_infos
, i
, info
)
689 free_stmt_vec_info (info
);
690 stmt_vec_infos
.release ();
693 /* Free STMT_INFO. */
696 vec_info::free_stmt_vec_info (stmt_vec_info stmt_info
)
698 if (stmt_info
->pattern_stmt_p
)
700 gimple_set_bb (stmt_info
->stmt
, NULL
);
701 tree lhs
= gimple_get_lhs (stmt_info
->stmt
);
702 if (lhs
&& TREE_CODE (lhs
) == SSA_NAME
)
703 release_ssa_name (lhs
);
706 STMT_VINFO_SAME_ALIGN_REFS (stmt_info
).release ();
707 STMT_VINFO_SIMD_CLONE_INFO (stmt_info
).release ();
711 /* A helper function to free scev and LOOP niter information, as well as
712 clear loop constraint LOOP_C_FINITE. */
715 vect_free_loop_info_assumptions (class loop
*loop
)
718 /* We need to explicitly reset upper bound information since they are
719 used even after free_numbers_of_iterations_estimates. */
720 loop
->any_upper_bound
= false;
721 loop
->any_likely_upper_bound
= false;
722 free_numbers_of_iterations_estimates (loop
);
723 loop_constraint_clear (loop
, LOOP_C_FINITE
);
726 /* If LOOP has been versioned during ifcvt, return the internal call
730 vect_loop_vectorized_call (class loop
*loop
, gcond
**cond
)
732 basic_block bb
= loop_preheader_edge (loop
)->src
;
739 if (!single_pred_p (bb
))
741 bb
= single_pred (bb
);
744 if (g
&& gimple_code (g
) == GIMPLE_COND
)
747 *cond
= as_a
<gcond
*> (g
);
748 gimple_stmt_iterator gsi
= gsi_for_stmt (g
);
750 if (!gsi_end_p (gsi
))
753 if (gimple_call_internal_p (g
, IFN_LOOP_VECTORIZED
)
754 && (tree_to_shwi (gimple_call_arg (g
, 0)) == loop
->num
755 || tree_to_shwi (gimple_call_arg (g
, 1)) == loop
->num
))
762 /* If LOOP has been versioned during loop distribution, return the gurading
766 vect_loop_dist_alias_call (class loop
*loop
)
770 class loop
*outer
, *orig
;
771 gimple_stmt_iterator gsi
;
774 if (loop
->orig_loop_num
== 0)
777 orig
= get_loop (cfun
, loop
->orig_loop_num
);
780 /* The original loop is somehow destroyed. Clear the information. */
781 loop
->orig_loop_num
= 0;
786 bb
= nearest_common_dominator (CDI_DOMINATORS
, loop
->header
, orig
->header
);
788 bb
= loop_preheader_edge (loop
)->src
;
790 outer
= bb
->loop_father
;
791 entry
= ENTRY_BLOCK_PTR_FOR_FN (cfun
);
793 /* Look upward in dominance tree. */
794 for (; bb
!= entry
&& flow_bb_inside_loop_p (outer
, bb
);
795 bb
= get_immediate_dominator (CDI_DOMINATORS
, bb
))
798 if (g
== NULL
|| gimple_code (g
) != GIMPLE_COND
)
801 gsi
= gsi_for_stmt (g
);
807 /* The guarding internal function call must have the same distribution
809 if (gimple_call_internal_p (g
, IFN_LOOP_DIST_ALIAS
)
810 && (tree_to_shwi (gimple_call_arg (g
, 0)) == loop
->orig_loop_num
))
816 /* Set the uids of all the statements in basic blocks inside loop
817 represented by LOOP_VINFO. LOOP_VECTORIZED_CALL is the internal
818 call guarding the loop which has been if converted. */
820 set_uid_loop_bbs (loop_vec_info loop_vinfo
, gimple
*loop_vectorized_call
)
822 tree arg
= gimple_call_arg (loop_vectorized_call
, 1);
825 class loop
*scalar_loop
= get_loop (cfun
, tree_to_shwi (arg
));
827 LOOP_VINFO_SCALAR_LOOP (loop_vinfo
) = scalar_loop
;
828 gcc_checking_assert (vect_loop_vectorized_call (scalar_loop
)
829 == loop_vectorized_call
);
830 /* If we are going to vectorize outer loop, prevent vectorization
831 of the inner loop in the scalar loop - either the scalar loop is
832 thrown away, so it is a wasted work, or is used only for
834 if (scalar_loop
->inner
)
836 gimple
*g
= vect_loop_vectorized_call (scalar_loop
->inner
);
839 arg
= gimple_call_arg (g
, 0);
840 get_loop (cfun
, tree_to_shwi (arg
))->dont_vectorize
= true;
841 fold_loop_internal_call (g
, boolean_false_node
);
844 bbs
= get_loop_body (scalar_loop
);
845 for (i
= 0; i
< scalar_loop
->num_nodes
; i
++)
847 basic_block bb
= bbs
[i
];
848 gimple_stmt_iterator gsi
;
849 for (gsi
= gsi_start_phis (bb
); !gsi_end_p (gsi
); gsi_next (&gsi
))
851 gimple
*phi
= gsi_stmt (gsi
);
852 gimple_set_uid (phi
, 0);
854 for (gsi
= gsi_start_bb (bb
); !gsi_end_p (gsi
); gsi_next (&gsi
))
856 gimple
*stmt
= gsi_stmt (gsi
);
857 gimple_set_uid (stmt
, 0);
863 /* Try to vectorize LOOP. */
866 try_vectorize_loop_1 (hash_table
<simduid_to_vf
> *&simduid_to_vf_htab
,
867 unsigned *num_vectorized_loops
, loop_p loop
,
868 gimple
*loop_vectorized_call
,
869 gimple
*loop_dist_alias_call
)
872 vec_info_shared shared
;
873 auto_purge_vect_location sentinel
;
874 vect_location
= find_loop_location (loop
);
876 if (LOCATION_LOCUS (vect_location
.get_location_t ()) != UNKNOWN_LOCATION
877 && dump_enabled_p ())
878 dump_printf (MSG_NOTE
| MSG_PRIORITY_INTERNALS
,
879 "\nAnalyzing loop at %s:%d\n",
880 LOCATION_FILE (vect_location
.get_location_t ()),
881 LOCATION_LINE (vect_location
.get_location_t ()));
883 opt_loop_vec_info loop_vinfo
= opt_loop_vec_info::success (NULL
);
884 /* In the case of epilogue vectorization the loop already has its
885 loop_vec_info set, we do not require to analyze the loop in this case. */
886 if (loop_vec_info vinfo
= loop_vec_info_for_loop (loop
))
887 loop_vinfo
= opt_loop_vec_info::success (vinfo
);
890 /* Try to analyze the loop, retaining an opt_problem if dump_enabled_p. */
891 loop_vinfo
= vect_analyze_loop (loop
, &shared
);
892 loop
->aux
= loop_vinfo
;
896 if (dump_enabled_p ())
897 if (opt_problem
*problem
= loop_vinfo
.get_problem ())
899 dump_printf_loc (MSG_MISSED_OPTIMIZATION
, vect_location
,
900 "couldn't vectorize loop\n");
901 problem
->emit_and_clear ();
904 if (!loop_vinfo
|| !LOOP_VINFO_VECTORIZABLE_P (loop_vinfo
))
906 /* Free existing information if loop is analyzed with some
908 if (loop_constraint_set_p (loop
, LOOP_C_FINITE
))
909 vect_free_loop_info_assumptions (loop
);
911 /* If we applied if-conversion then try to vectorize the
912 BB of innermost loops.
913 ??? Ideally BB vectorization would learn to vectorize
914 control flow by applying if-conversion on-the-fly, the
915 following retains the if-converted loop body even when
916 only non-if-converted parts took part in BB vectorization. */
917 if (flag_tree_slp_vectorize
!= 0
918 && loop_vectorized_call
921 basic_block bb
= loop
->header
;
922 bool require_loop_vectorize
= false;
923 for (gimple_stmt_iterator gsi
= gsi_start_bb (bb
);
924 !gsi_end_p (gsi
); gsi_next (&gsi
))
926 gimple
*stmt
= gsi_stmt (gsi
);
927 gcall
*call
= dyn_cast
<gcall
*> (stmt
);
928 if (call
&& gimple_call_internal_p (call
))
930 internal_fn ifn
= gimple_call_internal_fn (call
);
931 if (ifn
== IFN_MASK_LOAD
|| ifn
== IFN_MASK_STORE
932 /* Don't keep the if-converted parts when the ifn with
933 specifc type is not supported by the backend. */
934 || (direct_internal_fn_p (ifn
)
935 && !direct_internal_fn_supported_p
936 (call
, OPTIMIZE_FOR_SPEED
)))
938 require_loop_vectorize
= true;
942 gimple_set_uid (stmt
, -1);
943 gimple_set_visited (stmt
, false);
945 if (!require_loop_vectorize
&& vect_slp_bb (bb
))
947 if (dump_enabled_p ())
948 dump_printf_loc (MSG_NOTE
, vect_location
,
949 "basic block vectorized\n");
950 fold_loop_internal_call (loop_vectorized_call
,
952 loop_vectorized_call
= NULL
;
953 ret
|= TODO_cleanup_cfg
| TODO_update_ssa_only_virtuals
;
956 /* If outer loop vectorization fails for LOOP_VECTORIZED guarded
957 loop, don't vectorize its inner loop; we'll attempt to
958 vectorize LOOP_VECTORIZED guarded inner loop of the scalar
960 if (loop_vectorized_call
&& loop
->inner
)
961 loop
->inner
->dont_vectorize
= true;
965 if (!dbg_cnt (vect_loop
))
967 /* Free existing information if loop is analyzed with some
969 if (loop_constraint_set_p (loop
, LOOP_C_FINITE
))
970 vect_free_loop_info_assumptions (loop
);
974 if (loop_vectorized_call
)
975 set_uid_loop_bbs (loop_vinfo
, loop_vectorized_call
);
977 unsigned HOST_WIDE_INT bytes
;
978 if (dump_enabled_p ())
980 if (GET_MODE_SIZE (loop_vinfo
->vector_mode
).is_constant (&bytes
))
981 dump_printf_loc (MSG_OPTIMIZED_LOCATIONS
, vect_location
,
982 "loop vectorized using %wu byte vectors\n", bytes
);
984 dump_printf_loc (MSG_OPTIMIZED_LOCATIONS
, vect_location
,
985 "loop vectorized using variable length vectors\n");
988 loop_p new_loop
= vect_transform_loop (loop_vinfo
,
989 loop_vectorized_call
);
990 (*num_vectorized_loops
)++;
991 /* Now that the loop has been vectorized, allow it to be unrolled
993 loop
->force_vectorize
= false;
997 simduid_to_vf
*simduid_to_vf_data
= XNEW (simduid_to_vf
);
998 if (!simduid_to_vf_htab
)
999 simduid_to_vf_htab
= new hash_table
<simduid_to_vf
> (15);
1000 simduid_to_vf_data
->simduid
= DECL_UID (loop
->simduid
);
1001 simduid_to_vf_data
->vf
= loop_vinfo
->vectorization_factor
;
1002 *simduid_to_vf_htab
->find_slot (simduid_to_vf_data
, INSERT
)
1003 = simduid_to_vf_data
;
1006 if (loop_vectorized_call
)
1008 fold_loop_internal_call (loop_vectorized_call
, boolean_true_node
);
1009 loop_vectorized_call
= NULL
;
1010 ret
|= TODO_cleanup_cfg
;
1012 if (loop_dist_alias_call
)
1014 tree value
= gimple_call_arg (loop_dist_alias_call
, 1);
1015 fold_loop_internal_call (loop_dist_alias_call
, value
);
1016 loop_dist_alias_call
= NULL
;
1017 ret
|= TODO_cleanup_cfg
;
1020 /* Epilogue of vectorized loop must be vectorized too. */
1023 /* Don't include vectorized epilogues in the "vectorized loops" count.
1025 unsigned dont_count
= *num_vectorized_loops
;
1026 ret
|= try_vectorize_loop_1 (simduid_to_vf_htab
, &dont_count
,
1027 new_loop
, NULL
, NULL
);
1033 /* Try to vectorize LOOP. */
1036 try_vectorize_loop (hash_table
<simduid_to_vf
> *&simduid_to_vf_htab
,
1037 unsigned *num_vectorized_loops
, loop_p loop
)
1039 if (!((flag_tree_loop_vectorize
1040 && optimize_loop_nest_for_speed_p (loop
))
1041 || loop
->force_vectorize
))
1044 return try_vectorize_loop_1 (simduid_to_vf_htab
, num_vectorized_loops
, loop
,
1045 vect_loop_vectorized_call (loop
),
1046 vect_loop_dist_alias_call (loop
));
1050 /* Function vectorize_loops.
1052 Entry point to loop vectorization phase. */
1055 vectorize_loops (void)
1058 unsigned int num_vectorized_loops
= 0;
1059 unsigned int vect_loops_num
;
1061 hash_table
<simduid_to_vf
> *simduid_to_vf_htab
= NULL
;
1062 hash_table
<simd_array_to_simduid
> *simd_array_to_simduid_htab
= NULL
;
1063 bool any_ifcvt_loops
= false;
1066 vect_loops_num
= number_of_loops (cfun
);
1068 /* Bail out if there are no loops. */
1069 if (vect_loops_num
<= 1)
1072 if (cfun
->has_simduid_loops
)
1073 note_simd_array_uses (&simd_array_to_simduid_htab
);
1075 /* ----------- Analyze loops. ----------- */
1077 /* If some loop was duplicated, it gets bigger number
1078 than all previously defined loops. This fact allows us to run
1079 only over initial loops skipping newly generated ones. */
1080 FOR_EACH_LOOP (loop
, 0)
1081 if (loop
->dont_vectorize
)
1083 any_ifcvt_loops
= true;
1084 /* If-conversion sometimes versions both the outer loop
1085 (for the case when outer loop vectorization might be
1086 desirable) as well as the inner loop in the scalar version
1087 of the loop. So we have:
1088 if (LOOP_VECTORIZED (1, 3))
1094 loop3 (copy of loop1)
1095 if (LOOP_VECTORIZED (4, 5))
1096 loop4 (copy of loop2)
1098 loop5 (copy of loop4)
1099 If FOR_EACH_LOOP gives us loop3 first (which has
1100 dont_vectorize set), make sure to process loop1 before loop4;
1101 so that we can prevent vectorization of loop4 if loop1
1102 is successfully vectorized. */
1105 gimple
*loop_vectorized_call
1106 = vect_loop_vectorized_call (loop
);
1107 if (loop_vectorized_call
1108 && vect_loop_vectorized_call (loop
->inner
))
1110 tree arg
= gimple_call_arg (loop_vectorized_call
, 0);
1111 class loop
*vector_loop
1112 = get_loop (cfun
, tree_to_shwi (arg
));
1113 if (vector_loop
&& vector_loop
!= loop
)
1115 /* Make sure we don't vectorize it twice. */
1116 vector_loop
->dont_vectorize
= true;
1117 ret
|= try_vectorize_loop (simduid_to_vf_htab
,
1118 &num_vectorized_loops
,
1125 ret
|= try_vectorize_loop (simduid_to_vf_htab
, &num_vectorized_loops
,
1128 vect_location
= dump_user_location_t ();
1130 statistics_counter_event (cfun
, "Vectorized loops", num_vectorized_loops
);
1131 if (dump_enabled_p ()
1132 || (num_vectorized_loops
> 0 && dump_enabled_p ()))
1133 dump_printf_loc (MSG_NOTE
, vect_location
,
1134 "vectorized %u loops in function.\n",
1135 num_vectorized_loops
);
1137 /* ----------- Finalize. ----------- */
1139 if (any_ifcvt_loops
)
1140 for (i
= 1; i
< number_of_loops (cfun
); i
++)
1142 loop
= get_loop (cfun
, i
);
1143 if (loop
&& loop
->dont_vectorize
)
1145 gimple
*g
= vect_loop_vectorized_call (loop
);
1148 fold_loop_internal_call (g
, boolean_false_node
);
1149 ret
|= TODO_cleanup_cfg
;
1153 g
= vect_loop_dist_alias_call (loop
);
1157 fold_loop_internal_call (g
, boolean_false_node
);
1158 ret
|= TODO_cleanup_cfg
;
1163 for (i
= 1; i
< number_of_loops (cfun
); i
++)
1165 loop_vec_info loop_vinfo
;
1166 bool has_mask_store
;
1168 loop
= get_loop (cfun
, i
);
1169 if (!loop
|| !loop
->aux
)
1171 loop_vinfo
= (loop_vec_info
) loop
->aux
;
1172 has_mask_store
= LOOP_VINFO_HAS_MASK_STORE (loop_vinfo
);
1175 && targetm
.vectorize
.empty_mask_is_expensive (IFN_MASK_STORE
))
1176 optimize_mask_stores (loop
);
1180 /* Fold IFN_GOMP_SIMD_{VF,LANE,LAST_LANE,ORDERED_{START,END}} builtins. */
1181 if (cfun
->has_simduid_loops
)
1182 adjust_simduid_builtins (simduid_to_vf_htab
);
1184 /* Shrink any "omp array simd" temporary arrays to the
1185 actual vectorization factors. */
1186 if (simd_array_to_simduid_htab
)
1187 shrink_simd_arrays (simd_array_to_simduid_htab
, simduid_to_vf_htab
);
1188 delete simduid_to_vf_htab
;
1189 cfun
->has_simduid_loops
= false;
1191 if (num_vectorized_loops
> 0)
1193 /* If we vectorized any loop only virtual SSA form needs to be updated.
1194 ??? Also while we try hard to update loop-closed SSA form we fail
1195 to properly do this in some corner-cases (see PR56286). */
1196 rewrite_into_loop_closed_ssa (NULL
, TODO_update_ssa_only_virtuals
);
1197 return TODO_cleanup_cfg
;
1204 /* Entry point to the simduid cleanup pass. */
1208 const pass_data pass_data_simduid_cleanup
=
1210 GIMPLE_PASS
, /* type */
1211 "simduid", /* name */
1212 OPTGROUP_NONE
, /* optinfo_flags */
1213 TV_NONE
, /* tv_id */
1214 ( PROP_ssa
| PROP_cfg
), /* properties_required */
1215 0, /* properties_provided */
1216 0, /* properties_destroyed */
1217 0, /* todo_flags_start */
1218 0, /* todo_flags_finish */
1221 class pass_simduid_cleanup
: public gimple_opt_pass
1224 pass_simduid_cleanup (gcc::context
*ctxt
)
1225 : gimple_opt_pass (pass_data_simduid_cleanup
, ctxt
)
1228 /* opt_pass methods: */
1229 opt_pass
* clone () { return new pass_simduid_cleanup (m_ctxt
); }
1230 virtual bool gate (function
*fun
) { return fun
->has_simduid_loops
; }
1231 virtual unsigned int execute (function
*);
1233 }; // class pass_simduid_cleanup
1236 pass_simduid_cleanup::execute (function
*fun
)
1238 hash_table
<simd_array_to_simduid
> *simd_array_to_simduid_htab
= NULL
;
1240 note_simd_array_uses (&simd_array_to_simduid_htab
);
1242 /* Fold IFN_GOMP_SIMD_{VF,LANE,LAST_LANE,ORDERED_{START,END}} builtins. */
1243 adjust_simduid_builtins (NULL
);
1245 /* Shrink any "omp array simd" temporary arrays to the
1246 actual vectorization factors. */
1247 if (simd_array_to_simduid_htab
)
1248 shrink_simd_arrays (simd_array_to_simduid_htab
, NULL
);
1249 fun
->has_simduid_loops
= false;
1256 make_pass_simduid_cleanup (gcc::context
*ctxt
)
1258 return new pass_simduid_cleanup (ctxt
);
1262 /* Entry point to basic block SLP phase. */
1266 const pass_data pass_data_slp_vectorize
=
1268 GIMPLE_PASS
, /* type */
1270 OPTGROUP_LOOP
| OPTGROUP_VEC
, /* optinfo_flags */
1271 TV_TREE_SLP_VECTORIZATION
, /* tv_id */
1272 ( PROP_ssa
| PROP_cfg
), /* properties_required */
1273 0, /* properties_provided */
1274 0, /* properties_destroyed */
1275 0, /* todo_flags_start */
1276 TODO_update_ssa
, /* todo_flags_finish */
1279 class pass_slp_vectorize
: public gimple_opt_pass
1282 pass_slp_vectorize (gcc::context
*ctxt
)
1283 : gimple_opt_pass (pass_data_slp_vectorize
, ctxt
)
1286 /* opt_pass methods: */
1287 opt_pass
* clone () { return new pass_slp_vectorize (m_ctxt
); }
1288 virtual bool gate (function
*) { return flag_tree_slp_vectorize
!= 0; }
1289 virtual unsigned int execute (function
*);
1291 }; // class pass_slp_vectorize
1294 pass_slp_vectorize::execute (function
*fun
)
1296 auto_purge_vect_location sentinel
;
1299 bool in_loop_pipeline
= scev_initialized_p ();
1300 if (!in_loop_pipeline
)
1302 loop_optimizer_init (LOOPS_NORMAL
);
1306 /* Mark all stmts as not belonging to the current region and unvisited. */
1307 FOR_EACH_BB_FN (bb
, fun
)
1309 for (gimple_stmt_iterator gsi
= gsi_start_bb (bb
); !gsi_end_p (gsi
);
1312 gimple
*stmt
= gsi_stmt (gsi
);
1313 gimple_set_uid (stmt
, -1);
1314 gimple_set_visited (stmt
, false);
1318 FOR_EACH_BB_FN (bb
, fun
)
1320 if (vect_slp_bb (bb
))
1321 if (dump_enabled_p ())
1322 dump_printf_loc (MSG_NOTE
, vect_location
, "basic block vectorized\n");
1325 if (!in_loop_pipeline
)
1328 loop_optimizer_finalize ();
1337 make_pass_slp_vectorize (gcc::context
*ctxt
)
1339 return new pass_slp_vectorize (ctxt
);
1343 /* Increase alignment of global arrays to improve vectorization potential.
1345 - Consider also structs that have an array field.
1346 - Use ipa analysis to prune arrays that can't be vectorized?
1347 This should involve global alignment analysis and in the future also
1350 static unsigned get_vec_alignment_for_type (tree
);
1351 static hash_map
<tree
, unsigned> *type_align_map
;
1353 /* Return alignment of array's vector type corresponding to scalar type.
1354 0 if no vector type exists. */
1356 get_vec_alignment_for_array_type (tree type
)
1358 gcc_assert (TREE_CODE (type
) == ARRAY_TYPE
);
1359 poly_uint64 array_size
, vector_size
;
1361 tree scalar_type
= strip_array_types (type
);
1362 tree vectype
= get_related_vectype_for_scalar_type (VOIDmode
, scalar_type
);
1364 || !poly_int_tree_p (TYPE_SIZE (type
), &array_size
)
1365 || !poly_int_tree_p (TYPE_SIZE (vectype
), &vector_size
)
1366 || maybe_lt (array_size
, vector_size
))
1369 return TYPE_ALIGN (vectype
);
1372 /* Return alignment of field having maximum alignment of vector type
1373 corresponding to it's scalar type. For now, we only consider fields whose
1374 offset is a multiple of it's vector alignment.
1375 0 if no suitable field is found. */
1377 get_vec_alignment_for_record_type (tree type
)
1379 gcc_assert (TREE_CODE (type
) == RECORD_TYPE
);
1381 unsigned max_align
= 0, alignment
;
1382 HOST_WIDE_INT offset
;
1385 if (TYPE_PACKED (type
))
1388 unsigned *slot
= type_align_map
->get (type
);
1392 for (tree field
= first_field (type
);
1394 field
= DECL_CHAIN (field
))
1396 /* Skip if not FIELD_DECL or if alignment is set by user. */
1397 if (TREE_CODE (field
) != FIELD_DECL
1398 || DECL_USER_ALIGN (field
)
1399 || DECL_ARTIFICIAL (field
))
1402 /* We don't need to process the type further if offset is variable,
1403 since the offsets of remaining members will also be variable. */
1404 if (TREE_CODE (DECL_FIELD_OFFSET (field
)) != INTEGER_CST
1405 || TREE_CODE (DECL_FIELD_BIT_OFFSET (field
)) != INTEGER_CST
)
1408 /* Similarly stop processing the type if offset_tree
1409 does not fit in unsigned HOST_WIDE_INT. */
1410 offset_tree
= bit_position (field
);
1411 if (!tree_fits_uhwi_p (offset_tree
))
1414 offset
= tree_to_uhwi (offset_tree
);
1415 alignment
= get_vec_alignment_for_type (TREE_TYPE (field
));
1417 /* Get maximum alignment of vectorized field/array among those members
1418 whose offset is multiple of the vector alignment. */
1420 && (offset
% alignment
== 0)
1421 && (alignment
> max_align
))
1422 max_align
= alignment
;
1425 type_align_map
->put (type
, max_align
);
1429 /* Return alignment of vector type corresponding to decl's scalar type
1430 or 0 if it doesn't exist or the vector alignment is lesser than
1431 decl's alignment. */
1433 get_vec_alignment_for_type (tree type
)
1435 if (type
== NULL_TREE
)
1438 gcc_assert (TYPE_P (type
));
1440 static unsigned alignment
= 0;
1441 switch (TREE_CODE (type
))
1444 alignment
= get_vec_alignment_for_array_type (type
);
1447 alignment
= get_vec_alignment_for_record_type (type
);
1454 return (alignment
> TYPE_ALIGN (type
)) ? alignment
: 0;
1457 /* Entry point to increase_alignment pass. */
1459 increase_alignment (void)
1461 varpool_node
*vnode
;
1463 vect_location
= dump_user_location_t ();
1464 type_align_map
= new hash_map
<tree
, unsigned>;
1466 /* Increase the alignment of all global arrays for vectorization. */
1467 FOR_EACH_DEFINED_VARIABLE (vnode
)
1469 tree decl
= vnode
->decl
;
1470 unsigned int alignment
;
1472 if ((decl_in_symtab_p (decl
)
1473 && !symtab_node::get (decl
)->can_increase_alignment_p ())
1474 || DECL_USER_ALIGN (decl
) || DECL_ARTIFICIAL (decl
))
1477 alignment
= get_vec_alignment_for_type (TREE_TYPE (decl
));
1478 if (alignment
&& vect_can_force_dr_alignment_p (decl
, alignment
))
1480 vnode
->increase_alignment (alignment
);
1481 if (dump_enabled_p ())
1482 dump_printf (MSG_NOTE
, "Increasing alignment of decl: %T\n", decl
);
1486 delete type_align_map
;
1493 const pass_data pass_data_ipa_increase_alignment
=
1495 SIMPLE_IPA_PASS
, /* type */
1496 "increase_alignment", /* name */
1497 OPTGROUP_LOOP
| OPTGROUP_VEC
, /* optinfo_flags */
1498 TV_IPA_OPT
, /* tv_id */
1499 0, /* properties_required */
1500 0, /* properties_provided */
1501 0, /* properties_destroyed */
1502 0, /* todo_flags_start */
1503 0, /* todo_flags_finish */
1506 class pass_ipa_increase_alignment
: public simple_ipa_opt_pass
1509 pass_ipa_increase_alignment (gcc::context
*ctxt
)
1510 : simple_ipa_opt_pass (pass_data_ipa_increase_alignment
, ctxt
)
1513 /* opt_pass methods: */
1514 virtual bool gate (function
*)
1516 return flag_section_anchors
&& flag_tree_loop_vectorize
;
1519 virtual unsigned int execute (function
*) { return increase_alignment (); }
1521 }; // class pass_ipa_increase_alignment
1525 simple_ipa_opt_pass
*
1526 make_pass_ipa_increase_alignment (gcc::context
*ctxt
)
1528 return new pass_ipa_increase_alignment (ctxt
);
1531 /* If the condition represented by T is a comparison or the SSA name
1532 result of a comparison, extract the comparison's operands. Represent
1533 T as NE_EXPR <T, 0> otherwise. */
1536 scalar_cond_masked_key::get_cond_ops_from_tree (tree t
)
1538 if (TREE_CODE_CLASS (TREE_CODE (t
)) == tcc_comparison
)
1540 this->code
= TREE_CODE (t
);
1541 this->op0
= TREE_OPERAND (t
, 0);
1542 this->op1
= TREE_OPERAND (t
, 1);
1546 if (TREE_CODE (t
) == SSA_NAME
)
1547 if (gassign
*stmt
= dyn_cast
<gassign
*> (SSA_NAME_DEF_STMT (t
)))
1549 tree_code code
= gimple_assign_rhs_code (stmt
);
1550 if (TREE_CODE_CLASS (code
) == tcc_comparison
)
1553 this->op0
= gimple_assign_rhs1 (stmt
);
1554 this->op1
= gimple_assign_rhs2 (stmt
);
1559 this->code
= NE_EXPR
;
1561 this->op1
= build_zero_cst (TREE_TYPE (t
));