2 Copyright (C) 2003-2018 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"
84 /* Loop or bb location, with hotness information. */
85 dump_user_location_t vect_location
;
87 /* Vector mapping GIMPLE stmt to stmt_vec_info. */
88 vec
<stmt_vec_info
> *stmt_vec_info_vec
;
90 /* Dump a cost entry according to args to F. */
93 dump_stmt_cost (FILE *f
, void *data
, int count
, enum vect_cost_for_stmt kind
,
94 stmt_vec_info stmt_info
, int misalign
,
95 enum vect_cost_model_location where
)
97 fprintf (f
, "%p ", data
);
100 print_gimple_expr (f
, STMT_VINFO_STMT (stmt_info
), 0, TDF_SLIM
);
104 fprintf (f
, "<unknown> ");
105 fprintf (f
, "%d times ", count
);
106 const char *ks
= "unknown";
124 case vector_gather_load
:
125 ks
= "vector_gather_load";
128 ks
= "unaligned_load";
130 case unaligned_store
:
131 ks
= "unaligned_store";
134 ks
= "unaligned_store";
136 case vector_scatter_store
:
137 ks
= "unaligned_store";
140 ks
= "unaligned_store";
143 ks
= "unaligned_store";
145 case cond_branch_not_taken
:
146 ks
= "unaligned_store";
148 case cond_branch_taken
:
149 ks
= "unaligned_store";
152 ks
= "unaligned_store";
154 case vec_promote_demote
:
155 ks
= "unaligned_store";
158 ks
= "unaligned_store";
161 fprintf (f
, "%s ", ks
);
162 if (kind
== unaligned_load
|| kind
== unaligned_store
)
163 fprintf (f
, "(misalign %d) ", misalign
);
164 const char *ws
= "unknown";
177 fprintf (f
, "in %s\n", ws
);
180 /* For mapping simduid to vectorization factor. */
182 struct simduid_to_vf
: free_ptr_hash
<simduid_to_vf
>
184 unsigned int simduid
;
187 /* hash_table support. */
188 static inline hashval_t
hash (const simduid_to_vf
*);
189 static inline int equal (const simduid_to_vf
*, const simduid_to_vf
*);
193 simduid_to_vf::hash (const simduid_to_vf
*p
)
199 simduid_to_vf::equal (const simduid_to_vf
*p1
, const simduid_to_vf
*p2
)
201 return p1
->simduid
== p2
->simduid
;
204 /* This hash maps the OMP simd array to the corresponding simduid used
205 to index into it. Like thus,
207 _7 = GOMP_SIMD_LANE (simduid.0)
213 This hash maps from the OMP simd array (D.1737[]) to DECL_UID of
216 struct simd_array_to_simduid
: free_ptr_hash
<simd_array_to_simduid
>
219 unsigned int simduid
;
221 /* hash_table support. */
222 static inline hashval_t
hash (const simd_array_to_simduid
*);
223 static inline int equal (const simd_array_to_simduid
*,
224 const simd_array_to_simduid
*);
228 simd_array_to_simduid::hash (const simd_array_to_simduid
*p
)
230 return DECL_UID (p
->decl
);
234 simd_array_to_simduid::equal (const simd_array_to_simduid
*p1
,
235 const simd_array_to_simduid
*p2
)
237 return p1
->decl
== p2
->decl
;
240 /* Fold IFN_GOMP_SIMD_LANE, IFN_GOMP_SIMD_VF, IFN_GOMP_SIMD_LAST_LANE,
241 into their corresponding constants and remove
242 IFN_GOMP_SIMD_ORDERED_{START,END}. */
245 adjust_simduid_builtins (hash_table
<simduid_to_vf
> *htab
)
249 FOR_EACH_BB_FN (bb
, cfun
)
251 gimple_stmt_iterator i
;
253 for (i
= gsi_start_bb (bb
); !gsi_end_p (i
); )
256 enum internal_fn ifn
;
257 gimple
*stmt
= gsi_stmt (i
);
259 if (!is_gimple_call (stmt
)
260 || !gimple_call_internal_p (stmt
))
265 ifn
= gimple_call_internal_fn (stmt
);
268 case IFN_GOMP_SIMD_LANE
:
269 case IFN_GOMP_SIMD_VF
:
270 case IFN_GOMP_SIMD_LAST_LANE
:
272 case IFN_GOMP_SIMD_ORDERED_START
:
273 case IFN_GOMP_SIMD_ORDERED_END
:
274 if (integer_onep (gimple_call_arg (stmt
, 0)))
276 enum built_in_function bcode
277 = (ifn
== IFN_GOMP_SIMD_ORDERED_START
278 ? BUILT_IN_GOMP_ORDERED_START
279 : BUILT_IN_GOMP_ORDERED_END
);
281 = gimple_build_call (builtin_decl_explicit (bcode
), 0);
282 tree vdef
= gimple_vdef (stmt
);
283 gimple_set_vdef (g
, vdef
);
284 SSA_NAME_DEF_STMT (vdef
) = g
;
285 gimple_set_vuse (g
, gimple_vuse (stmt
));
286 gsi_replace (&i
, g
, true);
289 gsi_remove (&i
, true);
290 unlink_stmt_vdef (stmt
);
296 tree arg
= gimple_call_arg (stmt
, 0);
297 gcc_assert (arg
!= NULL_TREE
);
298 gcc_assert (TREE_CODE (arg
) == SSA_NAME
);
299 simduid_to_vf
*p
= NULL
, data
;
300 data
.simduid
= DECL_UID (SSA_NAME_VAR (arg
));
301 /* Need to nullify loop safelen field since it's value is not
302 valid after transformation. */
303 if (bb
->loop_father
&& bb
->loop_father
->safelen
> 0)
304 bb
->loop_father
->safelen
= 0;
307 p
= htab
->find (&data
);
313 case IFN_GOMP_SIMD_VF
:
314 t
= build_int_cst (unsigned_type_node
, vf
);
316 case IFN_GOMP_SIMD_LANE
:
317 t
= build_int_cst (unsigned_type_node
, 0);
319 case IFN_GOMP_SIMD_LAST_LANE
:
320 t
= gimple_call_arg (stmt
, 1);
325 tree lhs
= gimple_call_lhs (stmt
);
327 replace_uses_by (lhs
, t
);
329 gsi_remove (&i
, true);
334 /* Helper structure for note_simd_array_uses. */
336 struct note_simd_array_uses_struct
338 hash_table
<simd_array_to_simduid
> **htab
;
339 unsigned int simduid
;
342 /* Callback for note_simd_array_uses, called through walk_gimple_op. */
345 note_simd_array_uses_cb (tree
*tp
, int *walk_subtrees
, void *data
)
347 struct walk_stmt_info
*wi
= (struct walk_stmt_info
*) data
;
348 struct note_simd_array_uses_struct
*ns
349 = (struct note_simd_array_uses_struct
*) wi
->info
;
354 && lookup_attribute ("omp simd array", DECL_ATTRIBUTES (*tp
))
355 && DECL_CONTEXT (*tp
) == current_function_decl
)
357 simd_array_to_simduid data
;
359 *ns
->htab
= new hash_table
<simd_array_to_simduid
> (15);
361 data
.simduid
= ns
->simduid
;
362 simd_array_to_simduid
**slot
= (*ns
->htab
)->find_slot (&data
, INSERT
);
365 simd_array_to_simduid
*p
= XNEW (simd_array_to_simduid
);
369 else if ((*slot
)->simduid
!= ns
->simduid
)
370 (*slot
)->simduid
= -1U;
376 /* Find "omp simd array" temporaries and map them to corresponding
380 note_simd_array_uses (hash_table
<simd_array_to_simduid
> **htab
)
383 gimple_stmt_iterator gsi
;
384 struct walk_stmt_info wi
;
385 struct note_simd_array_uses_struct ns
;
387 memset (&wi
, 0, sizeof (wi
));
391 FOR_EACH_BB_FN (bb
, cfun
)
392 for (gsi
= gsi_start_bb (bb
); !gsi_end_p (gsi
); gsi_next (&gsi
))
394 gimple
*stmt
= gsi_stmt (gsi
);
395 if (!is_gimple_call (stmt
) || !gimple_call_internal_p (stmt
))
397 switch (gimple_call_internal_fn (stmt
))
399 case IFN_GOMP_SIMD_LANE
:
400 case IFN_GOMP_SIMD_VF
:
401 case IFN_GOMP_SIMD_LAST_LANE
:
406 tree lhs
= gimple_call_lhs (stmt
);
407 if (lhs
== NULL_TREE
)
409 imm_use_iterator use_iter
;
411 ns
.simduid
= DECL_UID (SSA_NAME_VAR (gimple_call_arg (stmt
, 0)));
412 FOR_EACH_IMM_USE_STMT (use_stmt
, use_iter
, lhs
)
413 if (!is_gimple_debug (use_stmt
))
414 walk_gimple_op (use_stmt
, note_simd_array_uses_cb
, &wi
);
418 /* Shrink arrays with "omp simd array" attribute to the corresponding
419 vectorization factor. */
423 (hash_table
<simd_array_to_simduid
> *simd_array_to_simduid_htab
,
424 hash_table
<simduid_to_vf
> *simduid_to_vf_htab
)
426 for (hash_table
<simd_array_to_simduid
>::iterator iter
427 = simd_array_to_simduid_htab
->begin ();
428 iter
!= simd_array_to_simduid_htab
->end (); ++iter
)
429 if ((*iter
)->simduid
!= -1U)
431 tree decl
= (*iter
)->decl
;
433 if (simduid_to_vf_htab
)
435 simduid_to_vf
*p
= NULL
, data
;
436 data
.simduid
= (*iter
)->simduid
;
437 p
= simduid_to_vf_htab
->find (&data
);
442 = build_array_type_nelts (TREE_TYPE (TREE_TYPE (decl
)), vf
);
443 TREE_TYPE (decl
) = atype
;
444 relayout_decl (decl
);
447 delete simd_array_to_simduid_htab
;
450 /* Initialize the vec_info with kind KIND_IN and target cost data
451 TARGET_COST_DATA_IN. */
453 vec_info::vec_info (vec_info::vec_kind kind_in
, void *target_cost_data_in
,
454 vec_info_shared
*shared_
)
457 target_cost_data (target_cost_data_in
)
459 stmt_vec_infos
.create (50);
460 set_stmt_vec_info_vec (&stmt_vec_infos
);
463 vec_info::~vec_info ()
465 slp_instance instance
;
468 FOR_EACH_VEC_ELT (slp_instances
, i
, instance
)
469 vect_free_slp_instance (instance
, true);
471 destroy_cost_data (target_cost_data
);
472 free_stmt_vec_infos (&stmt_vec_infos
);
475 vec_info_shared::vec_info_shared ()
477 datarefs_copy (vNULL
),
482 vec_info_shared::~vec_info_shared ()
484 free_data_refs (datarefs
);
485 free_dependence_relations (ddrs
);
486 datarefs_copy
.release ();
490 vec_info_shared::save_datarefs ()
494 datarefs_copy
.reserve_exact (datarefs
.length ());
495 for (unsigned i
= 0; i
< datarefs
.length (); ++i
)
496 datarefs_copy
.quick_push (*datarefs
[i
]);
500 vec_info_shared::check_datarefs ()
504 gcc_assert (datarefs
.length () == datarefs_copy
.length ());
505 for (unsigned i
= 0; i
< datarefs
.length (); ++i
)
506 if (memcmp (&datarefs_copy
[i
], datarefs
[i
], sizeof (data_reference
)) != 0)
510 /* Record that STMT belongs to the vectorizable region. Create and return
511 an associated stmt_vec_info. */
514 vec_info::add_stmt (gimple
*stmt
)
516 stmt_vec_info res
= new_stmt_vec_info (stmt
, this);
517 set_vinfo_for_stmt (stmt
, res
);
521 /* If STMT has an associated stmt_vec_info, return that vec_info, otherwise
522 return null. It is safe to call this function on any statement, even if
523 it might not be part of the vectorizable region. */
526 vec_info::lookup_stmt (gimple
*stmt
)
528 unsigned int uid
= gimple_uid (stmt
);
529 if (uid
> 0 && uid
- 1 < stmt_vec_infos
.length ())
531 stmt_vec_info res
= stmt_vec_infos
[uid
- 1];
532 if (res
&& res
->stmt
== stmt
)
538 /* If NAME is an SSA_NAME and its definition has an associated stmt_vec_info,
539 return that stmt_vec_info, otherwise return null. It is safe to call
540 this on arbitrary operands. */
543 vec_info::lookup_def (tree name
)
545 if (TREE_CODE (name
) == SSA_NAME
546 && !SSA_NAME_IS_DEFAULT_DEF (name
))
547 return lookup_stmt (SSA_NAME_DEF_STMT (name
));
551 /* See whether there is a single non-debug statement that uses LHS and
552 whether that statement has an associated stmt_vec_info. Return the
553 stmt_vec_info if so, otherwise return null. */
556 vec_info::lookup_single_use (tree lhs
)
560 if (single_imm_use (lhs
, &dummy
, &use_stmt
))
561 return lookup_stmt (use_stmt
);
565 /* A helper function to free scev and LOOP niter information, as well as
566 clear loop constraint LOOP_C_FINITE. */
569 vect_free_loop_info_assumptions (struct loop
*loop
)
572 /* We need to explicitly reset upper bound information since they are
573 used even after free_numbers_of_iterations_estimates. */
574 loop
->any_upper_bound
= false;
575 loop
->any_likely_upper_bound
= false;
576 free_numbers_of_iterations_estimates (loop
);
577 loop_constraint_clear (loop
, LOOP_C_FINITE
);
580 /* Return whether STMT is inside the region we try to vectorize. */
583 vect_stmt_in_region_p (vec_info
*vinfo
, gimple
*stmt
)
585 if (!gimple_bb (stmt
))
588 if (loop_vec_info loop_vinfo
= dyn_cast
<loop_vec_info
> (vinfo
))
590 struct loop
*loop
= LOOP_VINFO_LOOP (loop_vinfo
);
591 if (!flow_bb_inside_loop_p (loop
, gimple_bb (stmt
)))
596 bb_vec_info bb_vinfo
= as_a
<bb_vec_info
> (vinfo
);
597 if (gimple_bb (stmt
) != BB_VINFO_BB (bb_vinfo
)
598 || gimple_uid (stmt
) == -1U
599 || gimple_code (stmt
) == GIMPLE_PHI
)
607 /* If LOOP has been versioned during ifcvt, return the internal call
611 vect_loop_vectorized_call (struct loop
*loop
)
613 basic_block bb
= loop_preheader_edge (loop
)->src
;
620 if (!single_pred_p (bb
))
622 bb
= single_pred (bb
);
625 if (g
&& gimple_code (g
) == GIMPLE_COND
)
627 gimple_stmt_iterator gsi
= gsi_for_stmt (g
);
629 if (!gsi_end_p (gsi
))
632 if (gimple_call_internal_p (g
, IFN_LOOP_VECTORIZED
)
633 && (tree_to_shwi (gimple_call_arg (g
, 0)) == loop
->num
634 || tree_to_shwi (gimple_call_arg (g
, 1)) == loop
->num
))
641 /* If LOOP has been versioned during loop distribution, return the gurading
645 vect_loop_dist_alias_call (struct loop
*loop
)
649 struct loop
*outer
, *orig
;
650 gimple_stmt_iterator gsi
;
653 if (loop
->orig_loop_num
== 0)
656 orig
= get_loop (cfun
, loop
->orig_loop_num
);
659 /* The original loop is somehow destroyed. Clear the information. */
660 loop
->orig_loop_num
= 0;
665 bb
= nearest_common_dominator (CDI_DOMINATORS
, loop
->header
, orig
->header
);
667 bb
= loop_preheader_edge (loop
)->src
;
669 outer
= bb
->loop_father
;
670 entry
= ENTRY_BLOCK_PTR_FOR_FN (cfun
);
672 /* Look upward in dominance tree. */
673 for (; bb
!= entry
&& flow_bb_inside_loop_p (outer
, bb
);
674 bb
= get_immediate_dominator (CDI_DOMINATORS
, bb
))
677 if (g
== NULL
|| gimple_code (g
) != GIMPLE_COND
)
680 gsi
= gsi_for_stmt (g
);
686 /* The guarding internal function call must have the same distribution
688 if (gimple_call_internal_p (g
, IFN_LOOP_DIST_ALIAS
)
689 && (tree_to_shwi (gimple_call_arg (g
, 0)) == loop
->orig_loop_num
))
695 /* Set the uids of all the statements in basic blocks inside loop
696 represented by LOOP_VINFO. LOOP_VECTORIZED_CALL is the internal
697 call guarding the loop which has been if converted. */
699 set_uid_loop_bbs (loop_vec_info loop_vinfo
, gimple
*loop_vectorized_call
)
701 tree arg
= gimple_call_arg (loop_vectorized_call
, 1);
704 struct loop
*scalar_loop
= get_loop (cfun
, tree_to_shwi (arg
));
706 LOOP_VINFO_SCALAR_LOOP (loop_vinfo
) = scalar_loop
;
707 gcc_checking_assert (vect_loop_vectorized_call (scalar_loop
)
708 == loop_vectorized_call
);
709 /* If we are going to vectorize outer loop, prevent vectorization
710 of the inner loop in the scalar loop - either the scalar loop is
711 thrown away, so it is a wasted work, or is used only for
713 if (scalar_loop
->inner
)
715 gimple
*g
= vect_loop_vectorized_call (scalar_loop
->inner
);
718 arg
= gimple_call_arg (g
, 0);
719 get_loop (cfun
, tree_to_shwi (arg
))->dont_vectorize
= true;
720 fold_loop_internal_call (g
, boolean_false_node
);
723 bbs
= get_loop_body (scalar_loop
);
724 for (i
= 0; i
< scalar_loop
->num_nodes
; i
++)
726 basic_block bb
= bbs
[i
];
727 gimple_stmt_iterator gsi
;
728 for (gsi
= gsi_start_phis (bb
); !gsi_end_p (gsi
); gsi_next (&gsi
))
730 gimple
*phi
= gsi_stmt (gsi
);
731 gimple_set_uid (phi
, 0);
733 for (gsi
= gsi_start_bb (bb
); !gsi_end_p (gsi
); gsi_next (&gsi
))
735 gimple
*stmt
= gsi_stmt (gsi
);
736 gimple_set_uid (stmt
, 0);
742 /* Try to vectorize LOOP. */
745 try_vectorize_loop_1 (hash_table
<simduid_to_vf
> *&simduid_to_vf_htab
,
746 unsigned *num_vectorized_loops
,
747 loop_p loop
, loop_vec_info orig_loop_vinfo
,
748 gimple
*loop_vectorized_call
,
749 gimple
*loop_dist_alias_call
)
752 vec_info_shared shared
;
753 vect_location
= find_loop_location (loop
);
754 if (LOCATION_LOCUS (vect_location
.get_location_t ()) != UNKNOWN_LOCATION
755 && dump_enabled_p ())
756 dump_printf (MSG_NOTE
, "\nAnalyzing loop at %s:%d\n",
757 LOCATION_FILE (vect_location
.get_location_t ()),
758 LOCATION_LINE (vect_location
.get_location_t ()));
760 loop_vec_info loop_vinfo
= vect_analyze_loop (loop
, orig_loop_vinfo
, &shared
);
761 loop
->aux
= loop_vinfo
;
763 if (!loop_vinfo
|| !LOOP_VINFO_VECTORIZABLE_P (loop_vinfo
))
765 /* Free existing information if loop is analyzed with some
767 if (loop_constraint_set_p (loop
, LOOP_C_FINITE
))
768 vect_free_loop_info_assumptions (loop
);
770 /* If we applied if-conversion then try to vectorize the
771 BB of innermost loops.
772 ??? Ideally BB vectorization would learn to vectorize
773 control flow by applying if-conversion on-the-fly, the
774 following retains the if-converted loop body even when
775 only non-if-converted parts took part in BB vectorization. */
776 if (flag_tree_slp_vectorize
!= 0
777 && loop_vectorized_call
780 basic_block bb
= loop
->header
;
781 bool has_mask_load_store
= false;
782 for (gimple_stmt_iterator gsi
= gsi_start_bb (bb
);
783 !gsi_end_p (gsi
); gsi_next (&gsi
))
785 gimple
*stmt
= gsi_stmt (gsi
);
786 if (is_gimple_call (stmt
)
787 && gimple_call_internal_p (stmt
)
788 && (gimple_call_internal_fn (stmt
) == IFN_MASK_LOAD
789 || gimple_call_internal_fn (stmt
) == IFN_MASK_STORE
))
791 has_mask_load_store
= true;
794 gimple_set_uid (stmt
, -1);
795 gimple_set_visited (stmt
, false);
797 if (! has_mask_load_store
&& vect_slp_bb (bb
))
799 dump_printf_loc (MSG_NOTE
, vect_location
,
800 "basic block vectorized\n");
801 fold_loop_internal_call (loop_vectorized_call
,
803 loop_vectorized_call
= NULL
;
804 ret
|= TODO_cleanup_cfg
;
807 /* If outer loop vectorization fails for LOOP_VECTORIZED guarded
808 loop, don't vectorize its inner loop; we'll attempt to
809 vectorize LOOP_VECTORIZED guarded inner loop of the scalar
811 if (loop_vectorized_call
&& loop
->inner
)
812 loop
->inner
->dont_vectorize
= true;
816 if (!dbg_cnt (vect_loop
))
818 /* Free existing information if loop is analyzed with some
820 if (loop_constraint_set_p (loop
, LOOP_C_FINITE
))
821 vect_free_loop_info_assumptions (loop
);
825 if (loop_vectorized_call
)
826 set_uid_loop_bbs (loop_vinfo
, loop_vectorized_call
);
828 unsigned HOST_WIDE_INT bytes
;
829 if (current_vector_size
.is_constant (&bytes
))
830 dump_printf_loc (MSG_OPTIMIZED_LOCATIONS
, vect_location
,
831 "loop vectorized vectorized using "
832 HOST_WIDE_INT_PRINT_UNSIGNED
" byte "
835 dump_printf_loc (MSG_OPTIMIZED_LOCATIONS
, vect_location
,
836 "loop vectorized using variable length vectors\n");
838 loop_p new_loop
= vect_transform_loop (loop_vinfo
);
839 (*num_vectorized_loops
)++;
840 /* Now that the loop has been vectorized, allow it to be unrolled
842 loop
->force_vectorize
= false;
846 simduid_to_vf
*simduid_to_vf_data
= XNEW (simduid_to_vf
);
847 if (!simduid_to_vf_htab
)
848 simduid_to_vf_htab
= new hash_table
<simduid_to_vf
> (15);
849 simduid_to_vf_data
->simduid
= DECL_UID (loop
->simduid
);
850 simduid_to_vf_data
->vf
= loop_vinfo
->vectorization_factor
;
851 *simduid_to_vf_htab
->find_slot (simduid_to_vf_data
, INSERT
)
852 = simduid_to_vf_data
;
855 if (loop_vectorized_call
)
857 fold_loop_internal_call (loop_vectorized_call
, boolean_true_node
);
858 loop_vectorized_call
= NULL
;
859 ret
|= TODO_cleanup_cfg
;
861 if (loop_dist_alias_call
)
863 tree value
= gimple_call_arg (loop_dist_alias_call
, 1);
864 fold_loop_internal_call (loop_dist_alias_call
, value
);
865 loop_dist_alias_call
= NULL
;
866 ret
|= TODO_cleanup_cfg
;
869 /* Epilogue of vectorized loop must be vectorized too. */
871 ret
|= try_vectorize_loop_1 (simduid_to_vf_htab
, num_vectorized_loops
,
872 new_loop
, loop_vinfo
, NULL
, NULL
);
877 /* Try to vectorize LOOP. */
880 try_vectorize_loop (hash_table
<simduid_to_vf
> *&simduid_to_vf_htab
,
881 unsigned *num_vectorized_loops
, loop_p loop
)
883 if (!((flag_tree_loop_vectorize
884 && optimize_loop_nest_for_speed_p (loop
))
885 || loop
->force_vectorize
))
888 return try_vectorize_loop_1 (simduid_to_vf_htab
, num_vectorized_loops
,
890 vect_loop_vectorized_call (loop
),
891 vect_loop_dist_alias_call (loop
));
895 /* Function vectorize_loops.
897 Entry point to loop vectorization phase. */
900 vectorize_loops (void)
903 unsigned int num_vectorized_loops
= 0;
904 unsigned int vect_loops_num
;
906 hash_table
<simduid_to_vf
> *simduid_to_vf_htab
= NULL
;
907 hash_table
<simd_array_to_simduid
> *simd_array_to_simduid_htab
= NULL
;
908 bool any_ifcvt_loops
= false;
911 vect_loops_num
= number_of_loops (cfun
);
913 /* Bail out if there are no loops. */
914 if (vect_loops_num
<= 1)
917 if (cfun
->has_simduid_loops
)
918 note_simd_array_uses (&simd_array_to_simduid_htab
);
920 set_stmt_vec_info_vec (NULL
);
922 /* ----------- Analyze loops. ----------- */
924 /* If some loop was duplicated, it gets bigger number
925 than all previously defined loops. This fact allows us to run
926 only over initial loops skipping newly generated ones. */
927 FOR_EACH_LOOP (loop
, 0)
928 if (loop
->dont_vectorize
)
930 any_ifcvt_loops
= true;
931 /* If-conversion sometimes versions both the outer loop
932 (for the case when outer loop vectorization might be
933 desirable) as well as the inner loop in the scalar version
934 of the loop. So we have:
935 if (LOOP_VECTORIZED (1, 3))
941 loop3 (copy of loop1)
942 if (LOOP_VECTORIZED (4, 5))
943 loop4 (copy of loop2)
945 loop5 (copy of loop4)
946 If FOR_EACH_LOOP gives us loop3 first (which has
947 dont_vectorize set), make sure to process loop1 before loop4;
948 so that we can prevent vectorization of loop4 if loop1
949 is successfully vectorized. */
952 gimple
*loop_vectorized_call
953 = vect_loop_vectorized_call (loop
);
954 if (loop_vectorized_call
955 && vect_loop_vectorized_call (loop
->inner
))
957 tree arg
= gimple_call_arg (loop_vectorized_call
, 0);
958 struct loop
*vector_loop
959 = get_loop (cfun
, tree_to_shwi (arg
));
960 if (vector_loop
&& vector_loop
!= loop
)
962 /* Make sure we don't vectorize it twice. */
963 vector_loop
->dont_vectorize
= true;
964 ret
|= try_vectorize_loop (simduid_to_vf_htab
,
965 &num_vectorized_loops
,
972 ret
|= try_vectorize_loop (simduid_to_vf_htab
, &num_vectorized_loops
,
975 vect_location
= dump_user_location_t ();
977 statistics_counter_event (cfun
, "Vectorized loops", num_vectorized_loops
);
978 if (dump_enabled_p ()
979 || (num_vectorized_loops
> 0 && dump_enabled_p ()))
980 dump_printf_loc (MSG_NOTE
, vect_location
,
981 "vectorized %u loops in function.\n",
982 num_vectorized_loops
);
984 /* ----------- Finalize. ----------- */
987 for (i
= 1; i
< number_of_loops (cfun
); i
++)
989 loop
= get_loop (cfun
, i
);
990 if (loop
&& loop
->dont_vectorize
)
992 gimple
*g
= vect_loop_vectorized_call (loop
);
995 fold_loop_internal_call (g
, boolean_false_node
);
996 ret
|= TODO_cleanup_cfg
;
1000 g
= vect_loop_dist_alias_call (loop
);
1004 fold_loop_internal_call (g
, boolean_false_node
);
1005 ret
|= TODO_cleanup_cfg
;
1010 for (i
= 1; i
< number_of_loops (cfun
); i
++)
1012 loop_vec_info loop_vinfo
;
1013 bool has_mask_store
;
1015 loop
= get_loop (cfun
, i
);
1016 if (!loop
|| !loop
->aux
)
1018 loop_vinfo
= (loop_vec_info
) loop
->aux
;
1019 has_mask_store
= LOOP_VINFO_HAS_MASK_STORE (loop_vinfo
);
1022 && targetm
.vectorize
.empty_mask_is_expensive (IFN_MASK_STORE
))
1023 optimize_mask_stores (loop
);
1027 /* Fold IFN_GOMP_SIMD_{VF,LANE,LAST_LANE,ORDERED_{START,END}} builtins. */
1028 if (cfun
->has_simduid_loops
)
1029 adjust_simduid_builtins (simduid_to_vf_htab
);
1031 /* Shrink any "omp array simd" temporary arrays to the
1032 actual vectorization factors. */
1033 if (simd_array_to_simduid_htab
)
1034 shrink_simd_arrays (simd_array_to_simduid_htab
, simduid_to_vf_htab
);
1035 delete simduid_to_vf_htab
;
1036 cfun
->has_simduid_loops
= false;
1038 if (num_vectorized_loops
> 0)
1040 /* If we vectorized any loop only virtual SSA form needs to be updated.
1041 ??? Also while we try hard to update loop-closed SSA form we fail
1042 to properly do this in some corner-cases (see PR56286). */
1043 rewrite_into_loop_closed_ssa (NULL
, TODO_update_ssa_only_virtuals
);
1044 return TODO_cleanup_cfg
;
1051 /* Entry point to the simduid cleanup pass. */
1055 const pass_data pass_data_simduid_cleanup
=
1057 GIMPLE_PASS
, /* type */
1058 "simduid", /* name */
1059 OPTGROUP_NONE
, /* optinfo_flags */
1060 TV_NONE
, /* tv_id */
1061 ( PROP_ssa
| PROP_cfg
), /* properties_required */
1062 0, /* properties_provided */
1063 0, /* properties_destroyed */
1064 0, /* todo_flags_start */
1065 0, /* todo_flags_finish */
1068 class pass_simduid_cleanup
: public gimple_opt_pass
1071 pass_simduid_cleanup (gcc::context
*ctxt
)
1072 : gimple_opt_pass (pass_data_simduid_cleanup
, ctxt
)
1075 /* opt_pass methods: */
1076 opt_pass
* clone () { return new pass_simduid_cleanup (m_ctxt
); }
1077 virtual bool gate (function
*fun
) { return fun
->has_simduid_loops
; }
1078 virtual unsigned int execute (function
*);
1080 }; // class pass_simduid_cleanup
1083 pass_simduid_cleanup::execute (function
*fun
)
1085 hash_table
<simd_array_to_simduid
> *simd_array_to_simduid_htab
= NULL
;
1087 note_simd_array_uses (&simd_array_to_simduid_htab
);
1089 /* Fold IFN_GOMP_SIMD_{VF,LANE,LAST_LANE,ORDERED_{START,END}} builtins. */
1090 adjust_simduid_builtins (NULL
);
1092 /* Shrink any "omp array simd" temporary arrays to the
1093 actual vectorization factors. */
1094 if (simd_array_to_simduid_htab
)
1095 shrink_simd_arrays (simd_array_to_simduid_htab
, NULL
);
1096 fun
->has_simduid_loops
= false;
1103 make_pass_simduid_cleanup (gcc::context
*ctxt
)
1105 return new pass_simduid_cleanup (ctxt
);
1109 /* Entry point to basic block SLP phase. */
1113 const pass_data pass_data_slp_vectorize
=
1115 GIMPLE_PASS
, /* type */
1117 OPTGROUP_LOOP
| OPTGROUP_VEC
, /* optinfo_flags */
1118 TV_TREE_SLP_VECTORIZATION
, /* tv_id */
1119 ( PROP_ssa
| PROP_cfg
), /* properties_required */
1120 0, /* properties_provided */
1121 0, /* properties_destroyed */
1122 0, /* todo_flags_start */
1123 TODO_update_ssa
, /* todo_flags_finish */
1126 class pass_slp_vectorize
: public gimple_opt_pass
1129 pass_slp_vectorize (gcc::context
*ctxt
)
1130 : gimple_opt_pass (pass_data_slp_vectorize
, ctxt
)
1133 /* opt_pass methods: */
1134 opt_pass
* clone () { return new pass_slp_vectorize (m_ctxt
); }
1135 virtual bool gate (function
*) { return flag_tree_slp_vectorize
!= 0; }
1136 virtual unsigned int execute (function
*);
1138 }; // class pass_slp_vectorize
1141 pass_slp_vectorize::execute (function
*fun
)
1145 bool in_loop_pipeline
= scev_initialized_p ();
1146 if (!in_loop_pipeline
)
1148 loop_optimizer_init (LOOPS_NORMAL
);
1152 /* Mark all stmts as not belonging to the current region and unvisited. */
1153 FOR_EACH_BB_FN (bb
, fun
)
1155 for (gimple_stmt_iterator gsi
= gsi_start_bb (bb
); !gsi_end_p (gsi
);
1158 gimple
*stmt
= gsi_stmt (gsi
);
1159 gimple_set_uid (stmt
, -1);
1160 gimple_set_visited (stmt
, false);
1164 FOR_EACH_BB_FN (bb
, fun
)
1166 if (vect_slp_bb (bb
))
1167 dump_printf_loc (MSG_NOTE
, vect_location
, "basic block vectorized\n");
1170 if (!in_loop_pipeline
)
1173 loop_optimizer_finalize ();
1182 make_pass_slp_vectorize (gcc::context
*ctxt
)
1184 return new pass_slp_vectorize (ctxt
);
1188 /* Increase alignment of global arrays to improve vectorization potential.
1190 - Consider also structs that have an array field.
1191 - Use ipa analysis to prune arrays that can't be vectorized?
1192 This should involve global alignment analysis and in the future also
1195 static unsigned get_vec_alignment_for_type (tree
);
1196 static hash_map
<tree
, unsigned> *type_align_map
;
1198 /* Return alignment of array's vector type corresponding to scalar type.
1199 0 if no vector type exists. */
1201 get_vec_alignment_for_array_type (tree type
)
1203 gcc_assert (TREE_CODE (type
) == ARRAY_TYPE
);
1204 poly_uint64 array_size
, vector_size
;
1206 tree vectype
= get_vectype_for_scalar_type (strip_array_types (type
));
1208 || !poly_int_tree_p (TYPE_SIZE (type
), &array_size
)
1209 || !poly_int_tree_p (TYPE_SIZE (vectype
), &vector_size
)
1210 || maybe_lt (array_size
, vector_size
))
1213 return TYPE_ALIGN (vectype
);
1216 /* Return alignment of field having maximum alignment of vector type
1217 corresponding to it's scalar type. For now, we only consider fields whose
1218 offset is a multiple of it's vector alignment.
1219 0 if no suitable field is found. */
1221 get_vec_alignment_for_record_type (tree type
)
1223 gcc_assert (TREE_CODE (type
) == RECORD_TYPE
);
1225 unsigned max_align
= 0, alignment
;
1226 HOST_WIDE_INT offset
;
1229 if (TYPE_PACKED (type
))
1232 unsigned *slot
= type_align_map
->get (type
);
1236 for (tree field
= first_field (type
);
1238 field
= DECL_CHAIN (field
))
1240 /* Skip if not FIELD_DECL or if alignment is set by user. */
1241 if (TREE_CODE (field
) != FIELD_DECL
1242 || DECL_USER_ALIGN (field
)
1243 || DECL_ARTIFICIAL (field
))
1246 /* We don't need to process the type further if offset is variable,
1247 since the offsets of remaining members will also be variable. */
1248 if (TREE_CODE (DECL_FIELD_OFFSET (field
)) != INTEGER_CST
1249 || TREE_CODE (DECL_FIELD_BIT_OFFSET (field
)) != INTEGER_CST
)
1252 /* Similarly stop processing the type if offset_tree
1253 does not fit in unsigned HOST_WIDE_INT. */
1254 offset_tree
= bit_position (field
);
1255 if (!tree_fits_uhwi_p (offset_tree
))
1258 offset
= tree_to_uhwi (offset_tree
);
1259 alignment
= get_vec_alignment_for_type (TREE_TYPE (field
));
1261 /* Get maximum alignment of vectorized field/array among those members
1262 whose offset is multiple of the vector alignment. */
1264 && (offset
% alignment
== 0)
1265 && (alignment
> max_align
))
1266 max_align
= alignment
;
1269 type_align_map
->put (type
, max_align
);
1273 /* Return alignment of vector type corresponding to decl's scalar type
1274 or 0 if it doesn't exist or the vector alignment is lesser than
1275 decl's alignment. */
1277 get_vec_alignment_for_type (tree type
)
1279 if (type
== NULL_TREE
)
1282 gcc_assert (TYPE_P (type
));
1284 static unsigned alignment
= 0;
1285 switch (TREE_CODE (type
))
1288 alignment
= get_vec_alignment_for_array_type (type
);
1291 alignment
= get_vec_alignment_for_record_type (type
);
1298 return (alignment
> TYPE_ALIGN (type
)) ? alignment
: 0;
1301 /* Entry point to increase_alignment pass. */
1303 increase_alignment (void)
1305 varpool_node
*vnode
;
1307 vect_location
= dump_user_location_t ();
1308 type_align_map
= new hash_map
<tree
, unsigned>;
1310 /* Increase the alignment of all global arrays for vectorization. */
1311 FOR_EACH_DEFINED_VARIABLE (vnode
)
1313 tree decl
= vnode
->decl
;
1314 unsigned int alignment
;
1316 if ((decl_in_symtab_p (decl
)
1317 && !symtab_node::get (decl
)->can_increase_alignment_p ())
1318 || DECL_USER_ALIGN (decl
) || DECL_ARTIFICIAL (decl
))
1321 alignment
= get_vec_alignment_for_type (TREE_TYPE (decl
));
1322 if (alignment
&& vect_can_force_dr_alignment_p (decl
, alignment
))
1324 vnode
->increase_alignment (alignment
);
1325 dump_printf (MSG_NOTE
, "Increasing alignment of decl: ");
1326 dump_generic_expr (MSG_NOTE
, TDF_SLIM
, decl
);
1327 dump_printf (MSG_NOTE
, "\n");
1331 delete type_align_map
;
1338 const pass_data pass_data_ipa_increase_alignment
=
1340 SIMPLE_IPA_PASS
, /* type */
1341 "increase_alignment", /* name */
1342 OPTGROUP_LOOP
| OPTGROUP_VEC
, /* optinfo_flags */
1343 TV_IPA_OPT
, /* tv_id */
1344 0, /* properties_required */
1345 0, /* properties_provided */
1346 0, /* properties_destroyed */
1347 0, /* todo_flags_start */
1348 0, /* todo_flags_finish */
1351 class pass_ipa_increase_alignment
: public simple_ipa_opt_pass
1354 pass_ipa_increase_alignment (gcc::context
*ctxt
)
1355 : simple_ipa_opt_pass (pass_data_ipa_increase_alignment
, ctxt
)
1358 /* opt_pass methods: */
1359 virtual bool gate (function
*)
1361 return flag_section_anchors
&& flag_tree_loop_vectorize
;
1364 virtual unsigned int execute (function
*) { return increase_alignment (); }
1366 }; // class pass_ipa_increase_alignment
1370 simple_ipa_opt_pass
*
1371 make_pass_ipa_increase_alignment (gcc::context
*ctxt
)
1373 return new pass_ipa_increase_alignment (ctxt
);