1 /* Transformation Utilities for Loop Vectorization.
2 Copyright (C) 2003, 2004, 2005, 2006, 2007 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 2, 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 COPYING. If not, write to the Free
19 Software Foundation, 51 Franklin Street, Fifth Floor, Boston, MA
24 #include "coretypes.h"
30 #include "basic-block.h"
31 #include "diagnostic.h"
32 #include "tree-flow.h"
33 #include "tree-dump.h"
40 #include "tree-data-ref.h"
41 #include "tree-chrec.h"
42 #include "tree-scalar-evolution.h"
43 #include "tree-vectorizer.h"
44 #include "langhooks.h"
45 #include "tree-pass.h"
49 /* Utility functions for the code transformation. */
50 static bool vect_transform_stmt (tree
, block_stmt_iterator
*, bool *);
51 static tree
vect_create_destination_var (tree
, tree
);
52 static tree vect_create_data_ref_ptr
53 (tree
, block_stmt_iterator
*, tree
, tree
*, tree
*, bool, tree
);
54 static tree
vect_create_addr_base_for_vector_ref (tree
, tree
*, tree
);
55 static tree
vect_setup_realignment (tree
, block_stmt_iterator
*, tree
*);
56 static tree
vect_get_new_vect_var (tree
, enum vect_var_kind
, const char *);
57 static tree
vect_get_vec_def_for_operand (tree
, tree
, tree
*);
58 static tree
vect_init_vector (tree
, tree
, tree
);
59 static void vect_finish_stmt_generation
60 (tree stmt
, tree vec_stmt
, block_stmt_iterator
*bsi
);
61 static bool vect_is_simple_cond (tree
, loop_vec_info
);
62 static void update_vuses_to_preheader (tree
, struct loop
*);
63 static void vect_create_epilog_for_reduction (tree
, tree
, enum tree_code
, tree
);
64 static tree
get_initial_def_for_reduction (tree
, tree
, tree
*);
66 /* Utility function dealing with loop peeling (not peeling itself). */
67 static void vect_generate_tmps_on_preheader
68 (loop_vec_info
, tree
*, tree
*, tree
*);
69 static tree
vect_build_loop_niters (loop_vec_info
);
70 static void vect_update_ivs_after_vectorizer (loop_vec_info
, tree
, edge
);
71 static tree
vect_gen_niters_for_prolog_loop (loop_vec_info
, tree
);
72 static void vect_update_init_of_dr (struct data_reference
*, tree niters
);
73 static void vect_update_inits_of_drs (loop_vec_info
, tree
);
74 static int vect_min_worthwhile_factor (enum tree_code
);
77 /* Function vect_estimate_min_profitable_iters
79 Return the number of iterations required for the vector version of the
80 loop to be profitable relative to the cost of the scalar version of the
83 TODO: Take profile info into account before making vectorization
84 decisions, if available. */
87 vect_estimate_min_profitable_iters (loop_vec_info loop_vinfo
)
90 int min_profitable_iters
;
91 int peel_iters_prologue
;
92 int peel_iters_epilogue
;
93 int vec_inside_cost
= 0;
94 int vec_outside_cost
= 0;
95 int scalar_single_iter_cost
= 0;
96 int vf
= LOOP_VINFO_VECT_FACTOR (loop_vinfo
);
97 struct loop
*loop
= LOOP_VINFO_LOOP (loop_vinfo
);
98 basic_block
*bbs
= LOOP_VINFO_BBS (loop_vinfo
);
99 int nbbs
= loop
->num_nodes
;
101 /* Cost model disabled. */
102 if (!flag_vect_cost_model
)
104 if (vect_print_dump_info (REPORT_DETAILS
))
105 fprintf (vect_dump
, "cost model disabled.");
109 /* Requires loop versioning tests to handle misalignment.
110 FIXME: Make cost depend on number of stmts in may_misalign list. */
112 if (LOOP_VINFO_MAY_MISALIGN_STMTS (loop_vinfo
))
114 vec_outside_cost
+= TARG_COND_BRANCH_COST
;
115 if (vect_print_dump_info (REPORT_DETAILS
))
116 fprintf (vect_dump
, "cost model: Adding cost of checks for loop "
120 /* Requires a prologue loop when peeling to handle misalignment. Add cost of
121 two guards, one for the peeled loop and one for the vector loop. */
123 peel_iters_prologue
= LOOP_PEELING_FOR_ALIGNMENT (loop_vinfo
);
124 if (peel_iters_prologue
)
126 vec_outside_cost
+= 2 * TARG_COND_BRANCH_COST
;
127 if (vect_print_dump_info (REPORT_DETAILS
))
128 fprintf (vect_dump
, "cost model: Adding cost of checks for "
132 /* Requires an epilogue loop to finish up remaining iterations after vector
133 loop. Add cost of two guards, one for the peeled loop and one for the
136 if ((peel_iters_prologue
< 0)
137 || !LOOP_VINFO_NITERS_KNOWN_P (loop_vinfo
)
138 || LOOP_VINFO_INT_NITERS (loop_vinfo
) % vf
)
140 vec_outside_cost
+= 2 * TARG_COND_BRANCH_COST
;
141 if (vect_print_dump_info (REPORT_DETAILS
))
142 fprintf (vect_dump
, "cost model : Adding cost of checks for "
146 /* Count statements in scalar loop. Using this as scalar cost for a single
149 TODO: Add outer loop support.
151 TODO: Consider assigning different costs to different scalar
154 for (i
= 0; i
< nbbs
; i
++)
156 block_stmt_iterator si
;
157 basic_block bb
= bbs
[i
];
159 for (si
= bsi_start (bb
); !bsi_end_p (si
); bsi_next (&si
))
161 tree stmt
= bsi_stmt (si
);
162 stmt_vec_info stmt_info
= vinfo_for_stmt (stmt
);
163 if (!STMT_VINFO_RELEVANT_P (stmt_info
)
164 && !STMT_VINFO_LIVE_P (stmt_info
))
166 scalar_single_iter_cost
++;
167 vec_inside_cost
+= STMT_VINFO_INSIDE_OF_LOOP_COST (stmt_info
);
168 vec_outside_cost
+= STMT_VINFO_OUTSIDE_OF_LOOP_COST (stmt_info
);
172 /* Add additional cost for the peeled instructions in prologue and epilogue
175 FORNOW: If we dont know the value of peel_iters for prologue or epilogue
176 at compile-time - we assume the worst.
178 TODO: Build an expression that represents peel_iters for prologue and
179 epilogue to be used in a run-time test. */
181 peel_iters_prologue
= LOOP_PEELING_FOR_ALIGNMENT (loop_vinfo
);
183 if (peel_iters_prologue
< 0)
185 peel_iters_prologue
= vf
- 1;
186 if (vect_print_dump_info (REPORT_DETAILS
))
187 fprintf (vect_dump
, "cost model: "
188 "prologue peel iters set conservatively.");
190 /* If peeling for alignment is unknown, loop bound of main loop becomes
192 peel_iters_epilogue
= vf
- 1;
193 if (vect_print_dump_info (REPORT_DETAILS
))
194 fprintf (vect_dump
, "cost model: "
195 "epilogue peel iters set conservatively because "
196 "peeling for alignment is unknown .");
200 if (!LOOP_VINFO_NITERS_KNOWN_P (loop_vinfo
))
202 peel_iters_epilogue
= vf
- 1;
203 if (vect_print_dump_info (REPORT_DETAILS
))
204 fprintf (vect_dump
, "cost model: "
205 "epilogue peel iters set conservatively because "
206 "loop iterations are unknown .");
209 peel_iters_epilogue
=
210 (LOOP_VINFO_INT_NITERS (loop_vinfo
) - peel_iters_prologue
)
214 vec_outside_cost
+= (peel_iters_prologue
* scalar_single_iter_cost
)
215 + (peel_iters_epilogue
* scalar_single_iter_cost
);
217 /* Calculate number of iterations required to make the vector version
218 profitable, relative to the loop bodies only. The following condition
219 must hold true: ((SIC*VF)-VIC)*niters > VOC*VF, where
220 SIC = scalar iteration cost, VIC = vector iteration cost,
221 VOC = vector outside cost and VF = vectorization factor. */
223 if ((scalar_single_iter_cost
* vf
) > vec_inside_cost
)
225 if (vec_outside_cost
== 0)
226 min_profitable_iters
= 1;
229 min_profitable_iters
= (vec_outside_cost
* vf
)
230 / ((scalar_single_iter_cost
* vf
)
233 if ((scalar_single_iter_cost
* vf
* min_profitable_iters
)
234 <= ((vec_inside_cost
* min_profitable_iters
)
235 + (vec_outside_cost
* vf
)))
236 min_profitable_iters
++;
239 /* vector version will never be profitable. */
242 if (vect_print_dump_info (REPORT_DETAILS
))
243 fprintf (vect_dump
, "cost model: vector iteration cost = %d "
244 "is divisible by scalar iteration cost = %d by a factor "
245 "greater than or equal to the vectorization factor = %d .",
246 vec_inside_cost
, scalar_single_iter_cost
, vf
);
250 if (vect_print_dump_info (REPORT_DETAILS
))
252 fprintf (vect_dump
, "Cost model analysis: \n");
253 fprintf (vect_dump
, " Vector inside of loop cost: %d\n",
255 fprintf (vect_dump
, " Vector outside of loop cost: %d\n",
257 fprintf (vect_dump
, " Scalar cost: %d\n", scalar_single_iter_cost
);
258 fprintf (vect_dump
, " prologue iterations: %d\n",
259 peel_iters_prologue
);
260 fprintf (vect_dump
, " epilogue iterations: %d\n",
261 peel_iters_epilogue
);
262 fprintf (vect_dump
, " Calculated minimum iters for profitability: %d\n",
263 min_profitable_iters
);
264 fprintf (vect_dump
, " Actual minimum iters for profitability: %d\n",
265 min_profitable_iters
< vf
? vf
: min_profitable_iters
);
268 return min_profitable_iters
< vf
? vf
: min_profitable_iters
;
272 /* TODO: Close dependency between vect_model_*_cost and vectorizable_*
273 functions. Design better to avoid maintenance issues. */
275 /* Function vect_model_reduction_cost.
277 Models cost for a reduction operation, including the vector ops
278 generated within the strip-mine loop, the initial definition before
279 the loop, and the epilogue code that must be generated. */
282 vect_model_reduction_cost (stmt_vec_info stmt_info
, enum tree_code reduc_code
,
291 enum machine_mode mode
;
292 tree operation
= GIMPLE_STMT_OPERAND (STMT_VINFO_STMT (stmt_info
), 1);
293 int op_type
= TREE_CODE_LENGTH (TREE_CODE (operation
));
295 /* Cost of reduction op inside loop. */
296 STMT_VINFO_INSIDE_OF_LOOP_COST (stmt_info
) += ncopies
* TARG_VEC_STMT_COST
;
298 reduction_op
= TREE_OPERAND (operation
, op_type
-1);
299 vectype
= get_vectype_for_scalar_type (TREE_TYPE (reduction_op
));
300 mode
= TYPE_MODE (vectype
);
301 orig_stmt
= STMT_VINFO_RELATED_STMT (stmt_info
);
304 orig_stmt
= STMT_VINFO_STMT (stmt_info
);
306 code
= TREE_CODE (GIMPLE_STMT_OPERAND (orig_stmt
, 1));
308 /* Add in cost for initial definition. */
309 outer_cost
+= TARG_VEC_STMT_COST
;
311 /* Determine cost of epilogue code.
313 We have a reduction operator that will reduce the vector in one statement.
314 Also requires scalar extract. */
316 if (reduc_code
< NUM_TREE_CODES
)
317 outer_cost
+= TARG_VEC_STMT_COST
+ TARG_VEC_TO_SCALAR_COST
;
320 int vec_size_in_bits
= tree_low_cst (TYPE_SIZE (vectype
), 1);
322 TYPE_SIZE (TREE_TYPE ( GIMPLE_STMT_OPERAND (orig_stmt
, 0)));
323 int element_bitsize
= tree_low_cst (bitsize
, 1);
324 int nelements
= vec_size_in_bits
/ element_bitsize
;
326 optab
= optab_for_tree_code (code
, vectype
);
328 /* We have a whole vector shift available. */
329 if (!VECTOR_MODE_P (mode
)
330 || optab
->handlers
[mode
].insn_code
== CODE_FOR_nothing
)
331 /* Final reduction via vector shifts and the reduction operator. Also
332 requires scalar extract. */
333 outer_cost
+= ((exact_log2(nelements
) * 2 + 1) * TARG_VEC_STMT_COST
);
335 /* Use extracts and reduction op for final reduction. For N elements,
336 we have N extracts and N-1 reduction ops. */
337 outer_cost
+= ((nelements
+ nelements
- 1) * TARG_VEC_STMT_COST
);
340 STMT_VINFO_OUTSIDE_OF_LOOP_COST (stmt_info
) = outer_cost
;
342 if (vect_print_dump_info (REPORT_DETAILS
))
343 fprintf (vect_dump
, "vect_model_reduction_cost: inside_cost = %d, "
344 "outside_cost = %d .", STMT_VINFO_INSIDE_OF_LOOP_COST (stmt_info
),
345 STMT_VINFO_OUTSIDE_OF_LOOP_COST (stmt_info
));
349 /* Function vect_model_induction_cost.
351 Models cost for induction operations. */
354 vect_model_induction_cost (stmt_vec_info stmt_info
, int ncopies
)
356 /* loop cost for vec_loop. */
357 STMT_VINFO_INSIDE_OF_LOOP_COST (stmt_info
) = ncopies
* TARG_VEC_STMT_COST
;
358 /* prologue cost for vec_init and vec_step. */
359 STMT_VINFO_OUTSIDE_OF_LOOP_COST (stmt_info
) = 2 * TARG_VEC_STMT_COST
;
361 if (vect_print_dump_info (REPORT_DETAILS
))
362 fprintf (vect_dump
, "vect_model_induction_cost: inside_cost = %d, "
363 "outside_cost = %d .", STMT_VINFO_INSIDE_OF_LOOP_COST (stmt_info
),
364 STMT_VINFO_OUTSIDE_OF_LOOP_COST (stmt_info
));
368 /* Function vect_model_simple_cost.
370 Models cost for simple operations, i.e. those that only emit ncopies of a
371 single op. Right now, this does not account for multiple insns that could
372 be generated for the single vector op. We will handle that shortly. */
375 vect_model_simple_cost (stmt_vec_info stmt_info
, int ncopies
)
377 STMT_VINFO_INSIDE_OF_LOOP_COST (stmt_info
) = ncopies
* TARG_VEC_STMT_COST
;
379 if (vect_print_dump_info (REPORT_DETAILS
))
380 fprintf (vect_dump
, "vect_model_simple_cost: inside_cost = %d, "
381 "outside_cost = %d .", STMT_VINFO_INSIDE_OF_LOOP_COST (stmt_info
),
382 STMT_VINFO_OUTSIDE_OF_LOOP_COST (stmt_info
));
386 /* Function vect_cost_strided_group_size
388 For strided load or store, return the group_size only if it is the first
389 load or store of a group, else return 1. This ensures that group size is
390 only returned once per group. */
393 vect_cost_strided_group_size (stmt_vec_info stmt_info
)
395 tree first_stmt
= DR_GROUP_FIRST_DR (stmt_info
);
397 if (first_stmt
== STMT_VINFO_STMT (stmt_info
))
398 return DR_GROUP_SIZE (stmt_info
);
404 /* Function vect_model_store_cost
406 Models cost for stores. In the case of strided accesses, one access
407 has the overhead of the strided access attributed to it. */
410 vect_model_store_cost (stmt_vec_info stmt_info
, int ncopies
)
415 /* Strided access? */
416 if (DR_GROUP_FIRST_DR (stmt_info
))
417 group_size
= vect_cost_strided_group_size (stmt_info
);
418 /* Not a strided access. */
422 /* Is this an access in a group of stores, which provide strided access?
423 If so, add in the cost of the permutes. */
426 /* Uses a high and low interleave operation for each needed permute. */
427 cost
= ncopies
* exact_log2(group_size
) * group_size
428 * TARG_VEC_STMT_COST
;
430 if (vect_print_dump_info (REPORT_DETAILS
))
431 fprintf (vect_dump
, "vect_model_store_cost: strided group_size = %d .",
436 /* Costs of the stores. */
437 cost
+= ncopies
* TARG_VEC_STORE_COST
;
439 STMT_VINFO_INSIDE_OF_LOOP_COST (stmt_info
) = cost
;
441 if (vect_print_dump_info (REPORT_DETAILS
))
442 fprintf (vect_dump
, "vect_model_store_cost: inside_cost = %d, "
443 "outside_cost = %d .", STMT_VINFO_INSIDE_OF_LOOP_COST (stmt_info
),
444 STMT_VINFO_OUTSIDE_OF_LOOP_COST (stmt_info
));
448 /* Function vect_model_load_cost
450 Models cost for loads. In the case of strided accesses, the last access
451 has the overhead of the strided access attributed to it. Since unaligned
452 accesses are supported for loads, we also account for the costs of the
453 access scheme chosen. */
456 vect_model_load_cost (stmt_vec_info stmt_info
, int ncopies
)
461 int alignment_support_cheme
;
463 struct data_reference
*dr
= STMT_VINFO_DATA_REF (stmt_info
), *first_dr
;
465 /* Strided accesses? */
466 first_stmt
= DR_GROUP_FIRST_DR (stmt_info
);
469 group_size
= vect_cost_strided_group_size (stmt_info
);
470 first_dr
= STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt
));
472 /* Not a strided access. */
479 alignment_support_cheme
= vect_supportable_dr_alignment (first_dr
);
481 /* Is this an access in a group of loads providing strided access?
482 If so, add in the cost of the permutes. */
485 /* Uses an even and odd extract operations for each needed permute. */
486 inner_cost
= ncopies
* exact_log2(group_size
) * group_size
487 * TARG_VEC_STMT_COST
;
489 if (vect_print_dump_info (REPORT_DETAILS
))
490 fprintf (vect_dump
, "vect_model_load_cost: strided group_size = %d .",
495 /* The loads themselves. */
496 switch (alignment_support_cheme
)
500 inner_cost
+= ncopies
* TARG_VEC_LOAD_COST
;
502 if (vect_print_dump_info (REPORT_DETAILS
))
503 fprintf (vect_dump
, "vect_model_load_cost: aligned.");
507 case dr_unaligned_supported
:
509 /* Here, we assign an additional cost for the unaligned load. */
510 inner_cost
+= ncopies
* TARG_VEC_UNALIGNED_LOAD_COST
;
512 if (vect_print_dump_info (REPORT_DETAILS
))
513 fprintf (vect_dump
, "vect_model_load_cost: unaligned supported by "
518 case dr_unaligned_software_pipeline
:
522 if (vect_print_dump_info (REPORT_DETAILS
))
523 fprintf (vect_dump
, "vect_model_load_cost: unaligned software "
526 /* Unaligned software pipeline has a load of an address, an initial
527 load, and possibly a mask operation to "prime" the loop. However,
528 if this is an access in a group of loads, which provide strided
529 access, then the above cost should only be considered for one
530 access in the group. Inside the loop, there is a load op
531 and a realignment op. */
533 if ((!DR_GROUP_FIRST_DR (stmt_info
)) || group_size
> 1)
535 outer_cost
= 2*TARG_VEC_STMT_COST
;
536 if (targetm
.vectorize
.builtin_mask_for_load
)
537 outer_cost
+= TARG_VEC_STMT_COST
;
540 STMT_VINFO_OUTSIDE_OF_LOOP_COST (stmt_info
) = outer_cost
;
542 inner_cost
+= ncopies
* (TARG_VEC_LOAD_COST
+ TARG_VEC_STMT_COST
);
551 STMT_VINFO_INSIDE_OF_LOOP_COST (stmt_info
) = inner_cost
;
553 if (vect_print_dump_info (REPORT_DETAILS
))
554 fprintf (vect_dump
, "vect_model_load_cost: inside_cost = %d, "
555 "outside_cost = %d .", STMT_VINFO_INSIDE_OF_LOOP_COST (stmt_info
),
556 STMT_VINFO_OUTSIDE_OF_LOOP_COST (stmt_info
));
561 /* Function vect_get_new_vect_var.
563 Returns a name for a new variable. The current naming scheme appends the
564 prefix "vect_" or "vect_p" (depending on the value of VAR_KIND) to
565 the name of vectorizer generated variables, and appends that to NAME if
569 vect_get_new_vect_var (tree type
, enum vect_var_kind var_kind
, const char *name
)
576 case vect_simple_var
:
579 case vect_scalar_var
:
582 case vect_pointer_var
:
590 new_vect_var
= create_tmp_var (type
, concat (prefix
, name
, NULL
));
592 new_vect_var
= create_tmp_var (type
, prefix
);
594 /* Mark vector typed variable as a gimple register variable. */
595 if (TREE_CODE (type
) == VECTOR_TYPE
)
596 DECL_GIMPLE_REG_P (new_vect_var
) = true;
602 /* Function vect_create_addr_base_for_vector_ref.
604 Create an expression that computes the address of the first memory location
605 that will be accessed for a data reference.
608 STMT: The statement containing the data reference.
609 NEW_STMT_LIST: Must be initialized to NULL_TREE or a statement list.
610 OFFSET: Optional. If supplied, it is be added to the initial address.
613 1. Return an SSA_NAME whose value is the address of the memory location of
614 the first vector of the data reference.
615 2. If new_stmt_list is not NULL_TREE after return then the caller must insert
616 these statement(s) which define the returned SSA_NAME.
618 FORNOW: We are only handling array accesses with step 1. */
621 vect_create_addr_base_for_vector_ref (tree stmt
,
625 stmt_vec_info stmt_info
= vinfo_for_stmt (stmt
);
626 struct data_reference
*dr
= STMT_VINFO_DATA_REF (stmt_info
);
627 tree data_ref_base
= unshare_expr (DR_BASE_ADDRESS (dr
));
628 tree base_name
= build_fold_indirect_ref (data_ref_base
);
630 tree addr_base
, addr_expr
;
632 tree base_offset
= unshare_expr (DR_OFFSET (dr
));
633 tree init
= unshare_expr (DR_INIT (dr
));
634 tree vect_ptr_type
, addr_expr2
;
636 /* Create base_offset */
637 base_offset
= size_binop (PLUS_EXPR
, base_offset
, init
);
638 base_offset
= fold_convert (sizetype
, base_offset
);
639 dest
= create_tmp_var (TREE_TYPE (base_offset
), "base_off");
640 add_referenced_var (dest
);
641 base_offset
= force_gimple_operand (base_offset
, &new_stmt
, false, dest
);
642 append_to_statement_list_force (new_stmt
, new_stmt_list
);
646 tree tmp
= create_tmp_var (sizetype
, "offset");
649 /* For interleaved access step we divide STEP by the size of the
650 interleaving group. */
651 if (DR_GROUP_SIZE (stmt_info
))
652 step
= fold_build2 (TRUNC_DIV_EXPR
, TREE_TYPE (offset
), DR_STEP (dr
),
653 build_int_cst (TREE_TYPE (offset
),
654 DR_GROUP_SIZE (stmt_info
)));
658 add_referenced_var (tmp
);
659 offset
= fold_build2 (MULT_EXPR
, TREE_TYPE (offset
), offset
, step
);
660 base_offset
= fold_build2 (PLUS_EXPR
, TREE_TYPE (base_offset
),
661 base_offset
, offset
);
662 base_offset
= force_gimple_operand (base_offset
, &new_stmt
, false, tmp
);
663 append_to_statement_list_force (new_stmt
, new_stmt_list
);
666 /* base + base_offset */
667 addr_base
= fold_build2 (POINTER_PLUS_EXPR
, TREE_TYPE (data_ref_base
), data_ref_base
,
670 vect_ptr_type
= build_pointer_type (STMT_VINFO_VECTYPE (stmt_info
));
672 /* addr_expr = addr_base */
673 addr_expr
= vect_get_new_vect_var (vect_ptr_type
, vect_pointer_var
,
674 get_name (base_name
));
675 add_referenced_var (addr_expr
);
676 vec_stmt
= fold_convert (vect_ptr_type
, addr_base
);
677 addr_expr2
= vect_get_new_vect_var (vect_ptr_type
, vect_pointer_var
,
678 get_name (base_name
));
679 add_referenced_var (addr_expr2
);
680 vec_stmt
= force_gimple_operand (vec_stmt
, &new_stmt
, false, addr_expr2
);
681 append_to_statement_list_force (new_stmt
, new_stmt_list
);
683 if (vect_print_dump_info (REPORT_DETAILS
))
685 fprintf (vect_dump
, "created ");
686 print_generic_expr (vect_dump
, vec_stmt
, TDF_SLIM
);
692 /* Function vect_create_data_ref_ptr.
694 Create a new pointer to vector type (vp), that points to the first location
695 accessed in the loop by STMT, along with the def-use update chain to
696 appropriately advance the pointer through the loop iterations. Also set
697 aliasing information for the pointer. This vector pointer is used by the
698 callers to this function to create a memory reference expression for vector
702 1. STMT: a stmt that references memory. Expected to be of the form
703 GIMPLE_MODIFY_STMT <name, data-ref> or
704 GIMPLE_MODIFY_STMT <data-ref, name>.
705 2. BSI: block_stmt_iterator where new stmts can be added.
706 3. OFFSET (optional): an offset to be added to the initial address accessed
707 by the data-ref in STMT.
708 4. ONLY_INIT: indicate if vp is to be updated in the loop, or remain
709 pointing to the initial address.
710 5. TYPE: if not NULL indicates the required type of the data-ref
713 1. Declare a new ptr to vector_type, and have it point to the base of the
714 data reference (initial addressed accessed by the data reference).
715 For example, for vector of type V8HI, the following code is generated:
718 vp = (v8hi *)initial_address;
720 if OFFSET is not supplied:
721 initial_address = &a[init];
722 if OFFSET is supplied:
723 initial_address = &a[init + OFFSET];
725 Return the initial_address in INITIAL_ADDRESS.
727 2. If ONLY_INIT is true, just return the initial pointer. Otherwise, also
728 update the pointer in each iteration of the loop.
730 Return the increment stmt that updates the pointer in PTR_INCR.
732 3. Return the pointer. */
735 vect_create_data_ref_ptr (tree stmt
,
736 block_stmt_iterator
*bsi ATTRIBUTE_UNUSED
,
737 tree offset
, tree
*initial_address
, tree
*ptr_incr
,
738 bool only_init
, tree type
)
741 stmt_vec_info stmt_info
= vinfo_for_stmt (stmt
);
742 loop_vec_info loop_vinfo
= STMT_VINFO_LOOP_VINFO (stmt_info
);
743 struct loop
*loop
= LOOP_VINFO_LOOP (loop_vinfo
);
744 tree vectype
= STMT_VINFO_VECTYPE (stmt_info
);
750 tree new_stmt_list
= NULL_TREE
;
751 edge pe
= loop_preheader_edge (loop
);
754 struct data_reference
*dr
= STMT_VINFO_DATA_REF (stmt_info
);
756 base_name
= build_fold_indirect_ref (unshare_expr (DR_BASE_ADDRESS (dr
)));
758 if (vect_print_dump_info (REPORT_DETAILS
))
760 tree data_ref_base
= base_name
;
761 fprintf (vect_dump
, "create vector-pointer variable to type: ");
762 print_generic_expr (vect_dump
, vectype
, TDF_SLIM
);
763 if (TREE_CODE (data_ref_base
) == VAR_DECL
)
764 fprintf (vect_dump
, " vectorizing a one dimensional array ref: ");
765 else if (TREE_CODE (data_ref_base
) == ARRAY_REF
)
766 fprintf (vect_dump
, " vectorizing a multidimensional array ref: ");
767 else if (TREE_CODE (data_ref_base
) == COMPONENT_REF
)
768 fprintf (vect_dump
, " vectorizing a record based array ref: ");
769 else if (TREE_CODE (data_ref_base
) == SSA_NAME
)
770 fprintf (vect_dump
, " vectorizing a pointer ref: ");
771 print_generic_expr (vect_dump
, base_name
, TDF_SLIM
);
774 /** (1) Create the new vector-pointer variable: **/
776 vect_ptr_type
= build_pointer_type (type
);
778 vect_ptr_type
= build_pointer_type (vectype
);
779 vect_ptr
= vect_get_new_vect_var (vect_ptr_type
, vect_pointer_var
,
780 get_name (base_name
));
781 add_referenced_var (vect_ptr
);
783 /** (2) Add aliasing information to the new vector-pointer:
784 (The points-to info (DR_PTR_INFO) may be defined later.) **/
786 tag
= DR_SYMBOL_TAG (dr
);
789 /* If tag is a variable (and NOT_A_TAG) than a new symbol memory
790 tag must be created with tag added to its may alias list. */
792 new_type_alias (vect_ptr
, tag
, DR_REF (dr
));
794 set_symbol_mem_tag (vect_ptr
, tag
);
796 var_ann (vect_ptr
)->subvars
= DR_SUBVARS (dr
);
798 /** (3) Calculate the initial address the vector-pointer, and set
799 the vector-pointer to point to it before the loop: **/
801 /* Create: (&(base[init_val+offset]) in the loop preheader. */
802 new_temp
= vect_create_addr_base_for_vector_ref (stmt
, &new_stmt_list
,
804 pe
= loop_preheader_edge (loop
);
805 new_bb
= bsi_insert_on_edge_immediate (pe
, new_stmt_list
);
806 gcc_assert (!new_bb
);
807 *initial_address
= new_temp
;
809 /* Create: p = (vectype *) initial_base */
810 vec_stmt
= fold_convert (vect_ptr_type
, new_temp
);
811 vec_stmt
= build_gimple_modify_stmt (vect_ptr
, vec_stmt
);
812 vect_ptr_init
= make_ssa_name (vect_ptr
, vec_stmt
);
813 GIMPLE_STMT_OPERAND (vec_stmt
, 0) = vect_ptr_init
;
814 new_bb
= bsi_insert_on_edge_immediate (pe
, vec_stmt
);
815 gcc_assert (!new_bb
);
818 /** (4) Handle the updating of the vector-pointer inside the loop: **/
820 if (only_init
) /* No update in loop is required. */
822 /* Copy the points-to information if it exists. */
823 if (DR_PTR_INFO (dr
))
824 duplicate_ssa_name_ptr_info (vect_ptr_init
, DR_PTR_INFO (dr
));
825 return vect_ptr_init
;
829 block_stmt_iterator incr_bsi
;
831 tree indx_before_incr
, indx_after_incr
;
834 standard_iv_increment_position (loop
, &incr_bsi
, &insert_after
);
835 create_iv (vect_ptr_init
,
836 fold_convert (vect_ptr_type
, TYPE_SIZE_UNIT (vectype
)),
837 NULL_TREE
, loop
, &incr_bsi
, insert_after
,
838 &indx_before_incr
, &indx_after_incr
);
839 incr
= bsi_stmt (incr_bsi
);
840 set_stmt_info (stmt_ann (incr
),
841 new_stmt_vec_info (incr
, loop_vinfo
));
843 /* Copy the points-to information if it exists. */
844 if (DR_PTR_INFO (dr
))
846 duplicate_ssa_name_ptr_info (indx_before_incr
, DR_PTR_INFO (dr
));
847 duplicate_ssa_name_ptr_info (indx_after_incr
, DR_PTR_INFO (dr
));
849 merge_alias_info (vect_ptr_init
, indx_before_incr
);
850 merge_alias_info (vect_ptr_init
, indx_after_incr
);
854 return indx_before_incr
;
859 /* Function bump_vector_ptr
861 Increment a pointer (to a vector type) by vector-size. Connect the new
862 increment stmt to the existing def-use update-chain of the pointer.
864 The pointer def-use update-chain before this function:
865 DATAREF_PTR = phi (p_0, p_2)
867 PTR_INCR: p_2 = DATAREF_PTR + step
869 The pointer def-use update-chain after this function:
870 DATAREF_PTR = phi (p_0, p_2)
872 NEW_DATAREF_PTR = DATAREF_PTR + vector_size
874 PTR_INCR: p_2 = NEW_DATAREF_PTR + step
877 DATAREF_PTR - ssa_name of a pointer (to vector type) that is being updated
879 PTR_INCR - the stmt that updates the pointer in each iteration of the loop.
880 The increment amount across iterations is also expected to be
882 BSI - location where the new update stmt is to be placed.
883 STMT - the original scalar memory-access stmt that is being vectorized.
885 Output: Return NEW_DATAREF_PTR as illustrated above.
890 bump_vector_ptr (tree dataref_ptr
, tree ptr_incr
, block_stmt_iterator
*bsi
,
893 stmt_vec_info stmt_info
= vinfo_for_stmt (stmt
);
894 struct data_reference
*dr
= STMT_VINFO_DATA_REF (stmt_info
);
895 tree vectype
= STMT_VINFO_VECTYPE (stmt_info
);
896 tree vptr_type
= TREE_TYPE (dataref_ptr
);
897 tree ptr_var
= SSA_NAME_VAR (dataref_ptr
);
898 tree update
= TYPE_SIZE_UNIT (vectype
);
902 tree new_dataref_ptr
;
904 incr_stmt
= build_gimple_modify_stmt (ptr_var
,
905 build2 (POINTER_PLUS_EXPR
, vptr_type
,
906 dataref_ptr
, update
));
907 new_dataref_ptr
= make_ssa_name (ptr_var
, incr_stmt
);
908 GIMPLE_STMT_OPERAND (incr_stmt
, 0) = new_dataref_ptr
;
909 vect_finish_stmt_generation (stmt
, incr_stmt
, bsi
);
911 /* Update the vector-pointer's cross-iteration increment. */
912 FOR_EACH_SSA_USE_OPERAND (use_p
, ptr_incr
, iter
, SSA_OP_USE
)
914 tree use
= USE_FROM_PTR (use_p
);
916 if (use
== dataref_ptr
)
917 SET_USE (use_p
, new_dataref_ptr
);
919 gcc_assert (tree_int_cst_compare (use
, update
) == 0);
922 /* Copy the points-to information if it exists. */
923 if (DR_PTR_INFO (dr
))
924 duplicate_ssa_name_ptr_info (new_dataref_ptr
, DR_PTR_INFO (dr
));
925 merge_alias_info (new_dataref_ptr
, dataref_ptr
);
927 return new_dataref_ptr
;
931 /* Function vect_create_destination_var.
933 Create a new temporary of type VECTYPE. */
936 vect_create_destination_var (tree scalar_dest
, tree vectype
)
939 const char *new_name
;
941 enum vect_var_kind kind
;
943 kind
= vectype
? vect_simple_var
: vect_scalar_var
;
944 type
= vectype
? vectype
: TREE_TYPE (scalar_dest
);
946 gcc_assert (TREE_CODE (scalar_dest
) == SSA_NAME
);
948 new_name
= get_name (scalar_dest
);
951 vec_dest
= vect_get_new_vect_var (type
, kind
, new_name
);
952 add_referenced_var (vec_dest
);
958 /* Function vect_init_vector.
960 Insert a new stmt (INIT_STMT) that initializes a new vector variable with
961 the vector elements of VECTOR_VAR. Return the DEF of INIT_STMT. It will be
962 used in the vectorization of STMT. */
965 vect_init_vector (tree stmt
, tree vector_var
, tree vector_type
)
967 stmt_vec_info stmt_vinfo
= vinfo_for_stmt (stmt
);
968 loop_vec_info loop_vinfo
= STMT_VINFO_LOOP_VINFO (stmt_vinfo
);
969 struct loop
*loop
= LOOP_VINFO_LOOP (loop_vinfo
);
977 new_var
= vect_get_new_vect_var (vector_type
, vect_simple_var
, "cst_");
978 add_referenced_var (new_var
);
980 init_stmt
= build_gimple_modify_stmt (new_var
, vector_var
);
981 new_temp
= make_ssa_name (new_var
, init_stmt
);
982 GIMPLE_STMT_OPERAND (init_stmt
, 0) = new_temp
;
984 pe
= loop_preheader_edge (loop
);
985 new_bb
= bsi_insert_on_edge_immediate (pe
, init_stmt
);
986 gcc_assert (!new_bb
);
988 if (vect_print_dump_info (REPORT_DETAILS
))
990 fprintf (vect_dump
, "created new init_stmt: ");
991 print_generic_expr (vect_dump
, init_stmt
, TDF_SLIM
);
994 vec_oprnd
= GIMPLE_STMT_OPERAND (init_stmt
, 0);
999 /* Function get_initial_def_for_induction
1002 IV_PHI - the initial value of the induction variable
1005 Return a vector variable, initialized with the first VF values of
1006 the induction variable. E.g., for an iv with IV_PHI='X' and
1007 evolution S, for a vector of 4 units, we want to return:
1008 [X, X + S, X + 2*S, X + 3*S]. */
1011 get_initial_def_for_induction (tree iv_phi
)
1013 stmt_vec_info stmt_vinfo
= vinfo_for_stmt (iv_phi
);
1014 loop_vec_info loop_vinfo
= STMT_VINFO_LOOP_VINFO (stmt_vinfo
);
1015 struct loop
*loop
= LOOP_VINFO_LOOP (loop_vinfo
);
1016 tree scalar_type
= TREE_TYPE (PHI_RESULT_TREE (iv_phi
));
1017 tree vectype
= get_vectype_for_scalar_type (scalar_type
);
1018 int nunits
= TYPE_VECTOR_SUBPARTS (vectype
);
1019 edge pe
= loop_preheader_edge (loop
);
1021 block_stmt_iterator bsi
;
1022 tree vec
, vec_init
, vec_step
, t
;
1027 tree induction_phi
, induc_def
, new_stmt
, vec_def
, vec_dest
;
1028 tree init_expr
, step_expr
;
1029 int vf
= LOOP_VINFO_VECT_FACTOR (loop_vinfo
);
1032 int ncopies
= vf
/ nunits
;
1034 stmt_vec_info phi_info
= vinfo_for_stmt (iv_phi
);
1036 tree stmt
= NULL_TREE
;
1037 block_stmt_iterator si
;
1038 basic_block bb
= bb_for_stmt (iv_phi
);
1040 gcc_assert (phi_info
);
1041 gcc_assert (ncopies
>= 1);
1043 /* Find the first insertion point in the BB. */
1044 si
= bsi_after_labels (bb
);
1045 stmt
= bsi_stmt (si
);
1047 access_fn
= analyze_scalar_evolution (loop
, PHI_RESULT (iv_phi
));
1048 gcc_assert (access_fn
);
1049 ok
= vect_is_simple_iv_evolution (loop
->num
, access_fn
,
1050 &init_expr
, &step_expr
);
1053 /* Create the vector that holds the initial_value of the induction. */
1054 new_var
= vect_get_new_vect_var (scalar_type
, vect_scalar_var
, "var_");
1055 add_referenced_var (new_var
);
1057 new_name
= force_gimple_operand (init_expr
, &stmts
, false, new_var
);
1060 new_bb
= bsi_insert_on_edge_immediate (pe
, stmts
);
1061 gcc_assert (!new_bb
);
1065 t
= tree_cons (NULL_TREE
, new_name
, t
);
1066 for (i
= 1; i
< nunits
; i
++)
1070 /* Create: new_name = new_name + step_expr */
1071 tmp
= fold_build2 (PLUS_EXPR
, scalar_type
, new_name
, step_expr
);
1072 init_stmt
= build_gimple_modify_stmt (new_var
, tmp
);
1073 new_name
= make_ssa_name (new_var
, init_stmt
);
1074 GIMPLE_STMT_OPERAND (init_stmt
, 0) = new_name
;
1076 new_bb
= bsi_insert_on_edge_immediate (pe
, init_stmt
);
1077 gcc_assert (!new_bb
);
1079 if (vect_print_dump_info (REPORT_DETAILS
))
1081 fprintf (vect_dump
, "created new init_stmt: ");
1082 print_generic_expr (vect_dump
, init_stmt
, TDF_SLIM
);
1084 t
= tree_cons (NULL_TREE
, new_name
, t
);
1086 vec
= build_constructor_from_list (vectype
, nreverse (t
));
1087 vec_init
= vect_init_vector (stmt
, vec
, vectype
);
1090 /* Create the vector that holds the step of the induction. */
1091 expr
= build_int_cst (scalar_type
, vf
);
1092 new_name
= fold_build2 (MULT_EXPR
, scalar_type
, expr
, step_expr
);
1094 for (i
= 0; i
< nunits
; i
++)
1095 t
= tree_cons (NULL_TREE
, unshare_expr (new_name
), t
);
1096 vec
= build_constructor_from_list (vectype
, t
);
1097 vec_step
= vect_init_vector (stmt
, vec
, vectype
);
1100 /* Create the following def-use cycle:
1102 vec_init = [X, X+S, X+2*S, X+3*S]
1103 vec_step = [VF*S, VF*S, VF*S, VF*S]
1105 vec_iv = PHI <vec_init, vec_loop>
1109 vec_loop = vec_iv + vec_step; */
1111 /* Create the induction-phi that defines the induction-operand. */
1112 vec_dest
= vect_get_new_vect_var (vectype
, vect_simple_var
, "vec_iv_");
1113 add_referenced_var (vec_dest
);
1114 induction_phi
= create_phi_node (vec_dest
, loop
->header
);
1115 set_stmt_info (get_stmt_ann (induction_phi
),
1116 new_stmt_vec_info (induction_phi
, loop_vinfo
));
1117 induc_def
= PHI_RESULT (induction_phi
);
1119 /* Create the iv update inside the loop */
1120 new_stmt
= build_gimple_modify_stmt (NULL_TREE
,
1121 build2 (PLUS_EXPR
, vectype
,
1122 induc_def
, vec_step
));
1123 vec_def
= make_ssa_name (vec_dest
, new_stmt
);
1124 GIMPLE_STMT_OPERAND (new_stmt
, 0) = vec_def
;
1125 bsi
= bsi_for_stmt (stmt
);
1126 vect_finish_stmt_generation (stmt
, new_stmt
, &bsi
);
1128 /* Set the arguments of the phi node: */
1129 add_phi_arg (induction_phi
, vec_init
, loop_preheader_edge (loop
));
1130 add_phi_arg (induction_phi
, vec_def
, loop_latch_edge (loop
));
1133 /* In case the vectorization factor (VF) is bigger than the number
1134 of elements that we can fit in a vectype (nunits), we have to generate
1135 more than one vector stmt - i.e - we need to "unroll" the
1136 vector stmt by a factor VF/nunits. For more details see documentation
1137 in vectorizable_operation. */
1141 stmt_vec_info prev_stmt_vinfo
;
1143 /* Create the vector that holds the step of the induction. */
1144 expr
= build_int_cst (scalar_type
, nunits
);
1145 new_name
= fold_build2 (MULT_EXPR
, scalar_type
, expr
, step_expr
);
1147 for (i
= 0; i
< nunits
; i
++)
1148 t
= tree_cons (NULL_TREE
, unshare_expr (new_name
), t
);
1149 vec
= build_constructor_from_list (vectype
, t
);
1150 vec_step
= vect_init_vector (stmt
, vec
, vectype
);
1152 vec_def
= induc_def
;
1153 prev_stmt_vinfo
= vinfo_for_stmt (induction_phi
);
1154 for (i
= 1; i
< ncopies
; i
++)
1158 /* vec_i = vec_prev + vec_{step*nunits} */
1159 tmp
= build2 (PLUS_EXPR
, vectype
, vec_def
, vec_step
);
1160 new_stmt
= build_gimple_modify_stmt (NULL_TREE
, tmp
);
1161 vec_def
= make_ssa_name (vec_dest
, new_stmt
);
1162 GIMPLE_STMT_OPERAND (new_stmt
, 0) = vec_def
;
1163 bsi
= bsi_for_stmt (stmt
);
1164 vect_finish_stmt_generation (stmt
, new_stmt
, &bsi
);
1166 STMT_VINFO_RELATED_STMT (prev_stmt_vinfo
) = new_stmt
;
1167 prev_stmt_vinfo
= vinfo_for_stmt (new_stmt
);
1171 if (vect_print_dump_info (REPORT_DETAILS
))
1173 fprintf (vect_dump
, "transform induction: created def-use cycle:");
1174 print_generic_expr (vect_dump
, induction_phi
, TDF_SLIM
);
1175 fprintf (vect_dump
, "\n");
1176 print_generic_expr (vect_dump
, SSA_NAME_DEF_STMT (vec_def
), TDF_SLIM
);
1179 STMT_VINFO_VEC_STMT (phi_info
) = induction_phi
;
1184 /* Function vect_get_vec_def_for_operand.
1186 OP is an operand in STMT. This function returns a (vector) def that will be
1187 used in the vectorized stmt for STMT.
1189 In the case that OP is an SSA_NAME which is defined in the loop, then
1190 STMT_VINFO_VEC_STMT of the defining stmt holds the relevant def.
1192 In case OP is an invariant or constant, a new stmt that creates a vector def
1193 needs to be introduced. */
1196 vect_get_vec_def_for_operand (tree op
, tree stmt
, tree
*scalar_def
)
1201 stmt_vec_info def_stmt_info
= NULL
;
1202 stmt_vec_info stmt_vinfo
= vinfo_for_stmt (stmt
);
1203 tree vectype
= STMT_VINFO_VECTYPE (stmt_vinfo
);
1204 int nunits
= TYPE_VECTOR_SUBPARTS (vectype
);
1205 loop_vec_info loop_vinfo
= STMT_VINFO_LOOP_VINFO (stmt_vinfo
);
1206 struct loop
*loop
= LOOP_VINFO_LOOP (loop_vinfo
);
1212 enum vect_def_type dt
;
1216 if (vect_print_dump_info (REPORT_DETAILS
))
1218 fprintf (vect_dump
, "vect_get_vec_def_for_operand: ");
1219 print_generic_expr (vect_dump
, op
, TDF_SLIM
);
1222 is_simple_use
= vect_is_simple_use (op
, loop_vinfo
, &def_stmt
, &def
, &dt
);
1223 gcc_assert (is_simple_use
);
1224 if (vect_print_dump_info (REPORT_DETAILS
))
1228 fprintf (vect_dump
, "def = ");
1229 print_generic_expr (vect_dump
, def
, TDF_SLIM
);
1233 fprintf (vect_dump
, " def_stmt = ");
1234 print_generic_expr (vect_dump
, def_stmt
, TDF_SLIM
);
1240 /* Case 1: operand is a constant. */
1241 case vect_constant_def
:
1246 /* Create 'vect_cst_ = {cst,cst,...,cst}' */
1247 if (vect_print_dump_info (REPORT_DETAILS
))
1248 fprintf (vect_dump
, "Create vector_cst. nunits = %d", nunits
);
1250 for (i
= nunits
- 1; i
>= 0; --i
)
1252 t
= tree_cons (NULL_TREE
, op
, t
);
1254 vector_type
= get_vectype_for_scalar_type (TREE_TYPE (op
));
1255 vec_cst
= build_vector (vector_type
, t
);
1257 return vect_init_vector (stmt
, vec_cst
, vector_type
);
1260 /* Case 2: operand is defined outside the loop - loop invariant. */
1261 case vect_invariant_def
:
1266 /* Create 'vec_inv = {inv,inv,..,inv}' */
1267 if (vect_print_dump_info (REPORT_DETAILS
))
1268 fprintf (vect_dump
, "Create vector_inv.");
1270 for (i
= nunits
- 1; i
>= 0; --i
)
1272 t
= tree_cons (NULL_TREE
, def
, t
);
1275 /* FIXME: use build_constructor directly. */
1276 vector_type
= get_vectype_for_scalar_type (TREE_TYPE (def
));
1277 vec_inv
= build_constructor_from_list (vector_type
, t
);
1278 return vect_init_vector (stmt
, vec_inv
, vector_type
);
1281 /* Case 3: operand is defined inside the loop. */
1285 *scalar_def
= def_stmt
;
1287 /* Get the def from the vectorized stmt. */
1288 def_stmt_info
= vinfo_for_stmt (def_stmt
);
1289 vec_stmt
= STMT_VINFO_VEC_STMT (def_stmt_info
);
1290 gcc_assert (vec_stmt
);
1291 vec_oprnd
= GIMPLE_STMT_OPERAND (vec_stmt
, 0);
1295 /* Case 4: operand is defined by a loop header phi - reduction */
1296 case vect_reduction_def
:
1298 gcc_assert (TREE_CODE (def_stmt
) == PHI_NODE
);
1300 /* Get the def before the loop */
1301 op
= PHI_ARG_DEF_FROM_EDGE (def_stmt
, loop_preheader_edge (loop
));
1302 return get_initial_def_for_reduction (stmt
, op
, scalar_def
);
1305 /* Case 5: operand is defined by loop-header phi - induction. */
1306 case vect_induction_def
:
1308 gcc_assert (TREE_CODE (def_stmt
) == PHI_NODE
);
1310 /* Get the def before the loop */
1311 return get_initial_def_for_induction (def_stmt
);
1320 /* Function vect_get_vec_def_for_stmt_copy
1322 Return a vector-def for an operand. This function is used when the
1323 vectorized stmt to be created (by the caller to this function) is a "copy"
1324 created in case the vectorized result cannot fit in one vector, and several
1325 copies of the vector-stmt are required. In this case the vector-def is
1326 retrieved from the vector stmt recorded in the STMT_VINFO_RELATED_STMT field
1327 of the stmt that defines VEC_OPRND.
1328 DT is the type of the vector def VEC_OPRND.
1331 In case the vectorization factor (VF) is bigger than the number
1332 of elements that can fit in a vectype (nunits), we have to generate
1333 more than one vector stmt to vectorize the scalar stmt. This situation
1334 arises when there are multiple data-types operated upon in the loop; the
1335 smallest data-type determines the VF, and as a result, when vectorizing
1336 stmts operating on wider types we need to create 'VF/nunits' "copies" of the
1337 vector stmt (each computing a vector of 'nunits' results, and together
1338 computing 'VF' results in each iteration). This function is called when
1339 vectorizing such a stmt (e.g. vectorizing S2 in the illustration below, in
1340 which VF=16 and nunits=4, so the number of copies required is 4):
1342 scalar stmt: vectorized into: STMT_VINFO_RELATED_STMT
1344 S1: x = load VS1.0: vx.0 = memref0 VS1.1
1345 VS1.1: vx.1 = memref1 VS1.2
1346 VS1.2: vx.2 = memref2 VS1.3
1347 VS1.3: vx.3 = memref3
1349 S2: z = x + ... VSnew.0: vz0 = vx.0 + ... VSnew.1
1350 VSnew.1: vz1 = vx.1 + ... VSnew.2
1351 VSnew.2: vz2 = vx.2 + ... VSnew.3
1352 VSnew.3: vz3 = vx.3 + ...
1354 The vectorization of S1 is explained in vectorizable_load.
1355 The vectorization of S2:
1356 To create the first vector-stmt out of the 4 copies - VSnew.0 -
1357 the function 'vect_get_vec_def_for_operand' is called to
1358 get the relevant vector-def for each operand of S2. For operand x it
1359 returns the vector-def 'vx.0'.
1361 To create the remaining copies of the vector-stmt (VSnew.j), this
1362 function is called to get the relevant vector-def for each operand. It is
1363 obtained from the respective VS1.j stmt, which is recorded in the
1364 STMT_VINFO_RELATED_STMT field of the stmt that defines VEC_OPRND.
1366 For example, to obtain the vector-def 'vx.1' in order to create the
1367 vector stmt 'VSnew.1', this function is called with VEC_OPRND='vx.0'.
1368 Given 'vx0' we obtain the stmt that defines it ('VS1.0'); from the
1369 STMT_VINFO_RELATED_STMT field of 'VS1.0' we obtain the next copy - 'VS1.1',
1370 and return its def ('vx.1').
1371 Overall, to create the above sequence this function will be called 3 times:
1372 vx.1 = vect_get_vec_def_for_stmt_copy (dt, vx.0);
1373 vx.2 = vect_get_vec_def_for_stmt_copy (dt, vx.1);
1374 vx.3 = vect_get_vec_def_for_stmt_copy (dt, vx.2); */
1377 vect_get_vec_def_for_stmt_copy (enum vect_def_type dt
, tree vec_oprnd
)
1379 tree vec_stmt_for_operand
;
1380 stmt_vec_info def_stmt_info
;
1382 /* Do nothing; can reuse same def. */
1383 if (dt
== vect_invariant_def
|| dt
== vect_constant_def
)
1386 vec_stmt_for_operand
= SSA_NAME_DEF_STMT (vec_oprnd
);
1387 def_stmt_info
= vinfo_for_stmt (vec_stmt_for_operand
);
1388 gcc_assert (def_stmt_info
);
1389 vec_stmt_for_operand
= STMT_VINFO_RELATED_STMT (def_stmt_info
);
1390 gcc_assert (vec_stmt_for_operand
);
1391 vec_oprnd
= GIMPLE_STMT_OPERAND (vec_stmt_for_operand
, 0);
1397 /* Function vect_finish_stmt_generation.
1399 Insert a new stmt. */
1402 vect_finish_stmt_generation (tree stmt
, tree vec_stmt
,
1403 block_stmt_iterator
*bsi
)
1405 stmt_vec_info stmt_info
= vinfo_for_stmt (stmt
);
1406 loop_vec_info loop_vinfo
= STMT_VINFO_LOOP_VINFO (stmt_info
);
1408 bsi_insert_before (bsi
, vec_stmt
, BSI_SAME_STMT
);
1409 set_stmt_info (get_stmt_ann (vec_stmt
),
1410 new_stmt_vec_info (vec_stmt
, loop_vinfo
));
1412 if (vect_print_dump_info (REPORT_DETAILS
))
1414 fprintf (vect_dump
, "add new stmt: ");
1415 print_generic_expr (vect_dump
, vec_stmt
, TDF_SLIM
);
1418 /* Make sure bsi points to the stmt that is being vectorized. */
1419 gcc_assert (stmt
== bsi_stmt (*bsi
));
1421 #ifdef USE_MAPPED_LOCATION
1422 SET_EXPR_LOCATION (vec_stmt
, EXPR_LOCATION (stmt
));
1424 SET_EXPR_LOCUS (vec_stmt
, EXPR_LOCUS (stmt
));
1429 /* Function get_initial_def_for_reduction
1432 STMT - a stmt that performs a reduction operation in the loop.
1433 INIT_VAL - the initial value of the reduction variable
1436 ADJUSTMENT_DEF - a tree that holds a value to be added to the final result
1437 of the reduction (used for adjusting the epilog - see below).
1438 Return a vector variable, initialized according to the operation that STMT
1439 performs. This vector will be used as the initial value of the
1440 vector of partial results.
1442 Option1 (adjust in epilog): Initialize the vector as follows:
1445 min/max: [init_val,init_val,..,init_val,init_val]
1446 bit and/or: [init_val,init_val,..,init_val,init_val]
1447 and when necessary (e.g. add/mult case) let the caller know
1448 that it needs to adjust the result by init_val.
1450 Option2: Initialize the vector as follows:
1451 add: [0,0,...,0,init_val]
1452 mult: [1,1,...,1,init_val]
1453 min/max: [init_val,init_val,...,init_val]
1454 bit and/or: [init_val,init_val,...,init_val]
1455 and no adjustments are needed.
1457 For example, for the following code:
1463 STMT is 's = s + a[i]', and the reduction variable is 's'.
1464 For a vector of 4 units, we want to return either [0,0,0,init_val],
1465 or [0,0,0,0] and let the caller know that it needs to adjust
1466 the result at the end by 'init_val'.
1468 FORNOW, we are using the 'adjust in epilog' scheme, because this way the
1469 initialization vector is simpler (same element in all entries).
1470 A cost model should help decide between these two schemes. */
1473 get_initial_def_for_reduction (tree stmt
, tree init_val
, tree
*adjustment_def
)
1475 stmt_vec_info stmt_vinfo
= vinfo_for_stmt (stmt
);
1476 tree vectype
= STMT_VINFO_VECTYPE (stmt_vinfo
);
1477 int nunits
= TYPE_VECTOR_SUBPARTS (vectype
);
1478 enum tree_code code
= TREE_CODE (GIMPLE_STMT_OPERAND (stmt
, 1));
1479 tree type
= TREE_TYPE (init_val
);
1487 gcc_assert (INTEGRAL_TYPE_P (type
) || SCALAR_FLOAT_TYPE_P (type
));
1488 vecdef
= vect_get_vec_def_for_operand (init_val
, stmt
, NULL
);
1492 case WIDEN_SUM_EXPR
:
1495 *adjustment_def
= init_val
;
1496 /* Create a vector of zeros for init_def. */
1497 if (INTEGRAL_TYPE_P (type
))
1498 def_for_init
= build_int_cst (type
, 0);
1500 def_for_init
= build_real (type
, dconst0
);
1501 for (i
= nunits
- 1; i
>= 0; --i
)
1502 t
= tree_cons (NULL_TREE
, def_for_init
, t
);
1503 vector_type
= get_vectype_for_scalar_type (TREE_TYPE (def_for_init
));
1504 init_def
= build_vector (vector_type
, t
);
1509 *adjustment_def
= NULL_TREE
;
1521 /* Function vect_create_epilog_for_reduction
1523 Create code at the loop-epilog to finalize the result of a reduction
1526 VECT_DEF is a vector of partial results.
1527 REDUC_CODE is the tree-code for the epilog reduction.
1528 STMT is the scalar reduction stmt that is being vectorized.
1529 REDUCTION_PHI is the phi-node that carries the reduction computation.
1532 1. Creates the reduction def-use cycle: sets the arguments for
1534 The loop-entry argument is the vectorized initial-value of the reduction.
1535 The loop-latch argument is VECT_DEF - the vector of partial sums.
1536 2. "Reduces" the vector of partial results VECT_DEF into a single result,
1537 by applying the operation specified by REDUC_CODE if available, or by
1538 other means (whole-vector shifts or a scalar loop).
1539 The function also creates a new phi node at the loop exit to preserve
1540 loop-closed form, as illustrated below.
1542 The flow at the entry to this function:
1545 vec_def = phi <null, null> # REDUCTION_PHI
1546 VECT_DEF = vector_stmt # vectorized form of STMT
1547 s_loop = scalar_stmt # (scalar) STMT
1549 s_out0 = phi <s_loop> # (scalar) EXIT_PHI
1553 The above is transformed by this function into:
1556 vec_def = phi <vec_init, VECT_DEF> # REDUCTION_PHI
1557 VECT_DEF = vector_stmt # vectorized form of STMT
1558 s_loop = scalar_stmt # (scalar) STMT
1560 s_out0 = phi <s_loop> # (scalar) EXIT_PHI
1561 v_out1 = phi <VECT_DEF> # NEW_EXIT_PHI
1562 v_out2 = reduce <v_out1>
1563 s_out3 = extract_field <v_out2, 0>
1564 s_out4 = adjust_result <s_out3>
1570 vect_create_epilog_for_reduction (tree vect_def
, tree stmt
,
1571 enum tree_code reduc_code
, tree reduction_phi
)
1573 stmt_vec_info stmt_info
= vinfo_for_stmt (stmt
);
1575 enum machine_mode mode
;
1576 loop_vec_info loop_vinfo
= STMT_VINFO_LOOP_VINFO (stmt_info
);
1577 struct loop
*loop
= LOOP_VINFO_LOOP (loop_vinfo
);
1578 basic_block exit_bb
;
1582 block_stmt_iterator exit_bsi
;
1587 tree new_scalar_dest
, exit_phi
;
1588 tree bitsize
, bitpos
, bytesize
;
1589 enum tree_code code
= TREE_CODE (GIMPLE_STMT_OPERAND (stmt
, 1));
1590 tree scalar_initial_def
;
1591 tree vec_initial_def
;
1593 imm_use_iterator imm_iter
;
1594 use_operand_p use_p
;
1595 bool extract_scalar_result
;
1599 tree operation
= GIMPLE_STMT_OPERAND (stmt
, 1);
1602 op_type
= TREE_OPERAND_LENGTH (operation
);
1603 reduction_op
= TREE_OPERAND (operation
, op_type
-1);
1604 vectype
= get_vectype_for_scalar_type (TREE_TYPE (reduction_op
));
1605 mode
= TYPE_MODE (vectype
);
1607 /*** 1. Create the reduction def-use cycle ***/
1609 /* 1.1 set the loop-entry arg of the reduction-phi: */
1610 /* For the case of reduction, vect_get_vec_def_for_operand returns
1611 the scalar def before the loop, that defines the initial value
1612 of the reduction variable. */
1613 vec_initial_def
= vect_get_vec_def_for_operand (reduction_op
, stmt
,
1614 &scalar_initial_def
);
1615 add_phi_arg (reduction_phi
, vec_initial_def
, loop_preheader_edge (loop
));
1617 /* 1.2 set the loop-latch arg for the reduction-phi: */
1618 add_phi_arg (reduction_phi
, vect_def
, loop_latch_edge (loop
));
1620 if (vect_print_dump_info (REPORT_DETAILS
))
1622 fprintf (vect_dump
, "transform reduction: created def-use cycle:");
1623 print_generic_expr (vect_dump
, reduction_phi
, TDF_SLIM
);
1624 fprintf (vect_dump
, "\n");
1625 print_generic_expr (vect_dump
, SSA_NAME_DEF_STMT (vect_def
), TDF_SLIM
);
1629 /*** 2. Create epilog code
1630 The reduction epilog code operates across the elements of the vector
1631 of partial results computed by the vectorized loop.
1632 The reduction epilog code consists of:
1633 step 1: compute the scalar result in a vector (v_out2)
1634 step 2: extract the scalar result (s_out3) from the vector (v_out2)
1635 step 3: adjust the scalar result (s_out3) if needed.
1637 Step 1 can be accomplished using one the following three schemes:
1638 (scheme 1) using reduc_code, if available.
1639 (scheme 2) using whole-vector shifts, if available.
1640 (scheme 3) using a scalar loop. In this case steps 1+2 above are
1643 The overall epilog code looks like this:
1645 s_out0 = phi <s_loop> # original EXIT_PHI
1646 v_out1 = phi <VECT_DEF> # NEW_EXIT_PHI
1647 v_out2 = reduce <v_out1> # step 1
1648 s_out3 = extract_field <v_out2, 0> # step 2
1649 s_out4 = adjust_result <s_out3> # step 3
1651 (step 3 is optional, and step2 1 and 2 may be combined).
1652 Lastly, the uses of s_out0 are replaced by s_out4.
1656 /* 2.1 Create new loop-exit-phi to preserve loop-closed form:
1657 v_out1 = phi <v_loop> */
1659 exit_bb
= single_exit (loop
)->dest
;
1660 new_phi
= create_phi_node (SSA_NAME_VAR (vect_def
), exit_bb
);
1661 SET_PHI_ARG_DEF (new_phi
, single_exit (loop
)->dest_idx
, vect_def
);
1662 exit_bsi
= bsi_after_labels (exit_bb
);
1664 /* 2.2 Get the relevant tree-code to use in the epilog for schemes 2,3
1665 (i.e. when reduc_code is not available) and in the final adjustment
1666 code (if needed). Also get the original scalar reduction variable as
1667 defined in the loop. In case STMT is a "pattern-stmt" (i.e. - it
1668 represents a reduction pattern), the tree-code and scalar-def are
1669 taken from the original stmt that the pattern-stmt (STMT) replaces.
1670 Otherwise (it is a regular reduction) - the tree-code and scalar-def
1671 are taken from STMT. */
1673 orig_stmt
= STMT_VINFO_RELATED_STMT (stmt_info
);
1676 /* Regular reduction */
1681 /* Reduction pattern */
1682 stmt_vec_info stmt_vinfo
= vinfo_for_stmt (orig_stmt
);
1683 gcc_assert (STMT_VINFO_IN_PATTERN_P (stmt_vinfo
));
1684 gcc_assert (STMT_VINFO_RELATED_STMT (stmt_vinfo
) == stmt
);
1686 code
= TREE_CODE (GIMPLE_STMT_OPERAND (orig_stmt
, 1));
1687 scalar_dest
= GIMPLE_STMT_OPERAND (orig_stmt
, 0);
1688 scalar_type
= TREE_TYPE (scalar_dest
);
1689 new_scalar_dest
= vect_create_destination_var (scalar_dest
, NULL
);
1690 bitsize
= TYPE_SIZE (scalar_type
);
1691 bytesize
= TYPE_SIZE_UNIT (scalar_type
);
1693 /* 2.3 Create the reduction code, using one of the three schemes described
1696 if (reduc_code
< NUM_TREE_CODES
)
1700 /*** Case 1: Create:
1701 v_out2 = reduc_expr <v_out1> */
1703 if (vect_print_dump_info (REPORT_DETAILS
))
1704 fprintf (vect_dump
, "Reduce using direct vector reduction.");
1706 vec_dest
= vect_create_destination_var (scalar_dest
, vectype
);
1707 tmp
= build1 (reduc_code
, vectype
, PHI_RESULT (new_phi
));
1708 epilog_stmt
= build_gimple_modify_stmt (vec_dest
, tmp
);
1709 new_temp
= make_ssa_name (vec_dest
, epilog_stmt
);
1710 GIMPLE_STMT_OPERAND (epilog_stmt
, 0) = new_temp
;
1711 bsi_insert_before (&exit_bsi
, epilog_stmt
, BSI_SAME_STMT
);
1713 extract_scalar_result
= true;
1717 enum tree_code shift_code
= 0;
1718 bool have_whole_vector_shift
= true;
1720 int element_bitsize
= tree_low_cst (bitsize
, 1);
1721 int vec_size_in_bits
= tree_low_cst (TYPE_SIZE (vectype
), 1);
1724 if (vec_shr_optab
->handlers
[mode
].insn_code
!= CODE_FOR_nothing
)
1725 shift_code
= VEC_RSHIFT_EXPR
;
1727 have_whole_vector_shift
= false;
1729 /* Regardless of whether we have a whole vector shift, if we're
1730 emulating the operation via tree-vect-generic, we don't want
1731 to use it. Only the first round of the reduction is likely
1732 to still be profitable via emulation. */
1733 /* ??? It might be better to emit a reduction tree code here, so that
1734 tree-vect-generic can expand the first round via bit tricks. */
1735 if (!VECTOR_MODE_P (mode
))
1736 have_whole_vector_shift
= false;
1739 optab optab
= optab_for_tree_code (code
, vectype
);
1740 if (optab
->handlers
[mode
].insn_code
== CODE_FOR_nothing
)
1741 have_whole_vector_shift
= false;
1744 if (have_whole_vector_shift
)
1746 /*** Case 2: Create:
1747 for (offset = VS/2; offset >= element_size; offset/=2)
1749 Create: va' = vec_shift <va, offset>
1750 Create: va = vop <va, va'>
1753 if (vect_print_dump_info (REPORT_DETAILS
))
1754 fprintf (vect_dump
, "Reduce using vector shifts");
1756 vec_dest
= vect_create_destination_var (scalar_dest
, vectype
);
1757 new_temp
= PHI_RESULT (new_phi
);
1759 for (bit_offset
= vec_size_in_bits
/2;
1760 bit_offset
>= element_bitsize
;
1763 tree bitpos
= size_int (bit_offset
);
1764 tree tmp
= build2 (shift_code
, vectype
, new_temp
, bitpos
);
1765 epilog_stmt
= build_gimple_modify_stmt (vec_dest
, tmp
);
1766 new_name
= make_ssa_name (vec_dest
, epilog_stmt
);
1767 GIMPLE_STMT_OPERAND (epilog_stmt
, 0) = new_name
;
1768 bsi_insert_before (&exit_bsi
, epilog_stmt
, BSI_SAME_STMT
);
1770 tmp
= build2 (code
, vectype
, new_name
, new_temp
);
1771 epilog_stmt
= build_gimple_modify_stmt (vec_dest
, tmp
);
1772 new_temp
= make_ssa_name (vec_dest
, epilog_stmt
);
1773 GIMPLE_STMT_OPERAND (epilog_stmt
, 0) = new_temp
;
1774 bsi_insert_before (&exit_bsi
, epilog_stmt
, BSI_SAME_STMT
);
1777 extract_scalar_result
= true;
1783 /*** Case 3: Create:
1784 s = extract_field <v_out2, 0>
1785 for (offset = element_size;
1786 offset < vector_size;
1787 offset += element_size;)
1789 Create: s' = extract_field <v_out2, offset>
1790 Create: s = op <s, s'>
1793 if (vect_print_dump_info (REPORT_DETAILS
))
1794 fprintf (vect_dump
, "Reduce using scalar code. ");
1796 vec_temp
= PHI_RESULT (new_phi
);
1797 vec_size_in_bits
= tree_low_cst (TYPE_SIZE (vectype
), 1);
1798 rhs
= build3 (BIT_FIELD_REF
, scalar_type
, vec_temp
, bitsize
,
1800 BIT_FIELD_REF_UNSIGNED (rhs
) = TYPE_UNSIGNED (scalar_type
);
1801 epilog_stmt
= build_gimple_modify_stmt (new_scalar_dest
, rhs
);
1802 new_temp
= make_ssa_name (new_scalar_dest
, epilog_stmt
);
1803 GIMPLE_STMT_OPERAND (epilog_stmt
, 0) = new_temp
;
1804 bsi_insert_before (&exit_bsi
, epilog_stmt
, BSI_SAME_STMT
);
1806 for (bit_offset
= element_bitsize
;
1807 bit_offset
< vec_size_in_bits
;
1808 bit_offset
+= element_bitsize
)
1811 tree bitpos
= bitsize_int (bit_offset
);
1812 tree rhs
= build3 (BIT_FIELD_REF
, scalar_type
, vec_temp
, bitsize
,
1815 BIT_FIELD_REF_UNSIGNED (rhs
) = TYPE_UNSIGNED (scalar_type
);
1816 epilog_stmt
= build_gimple_modify_stmt (new_scalar_dest
, rhs
);
1817 new_name
= make_ssa_name (new_scalar_dest
, epilog_stmt
);
1818 GIMPLE_STMT_OPERAND (epilog_stmt
, 0) = new_name
;
1819 bsi_insert_before (&exit_bsi
, epilog_stmt
, BSI_SAME_STMT
);
1821 tmp
= build2 (code
, scalar_type
, new_name
, new_temp
);
1822 epilog_stmt
= build_gimple_modify_stmt (new_scalar_dest
, tmp
);
1823 new_temp
= make_ssa_name (new_scalar_dest
, epilog_stmt
);
1824 GIMPLE_STMT_OPERAND (epilog_stmt
, 0) = new_temp
;
1825 bsi_insert_before (&exit_bsi
, epilog_stmt
, BSI_SAME_STMT
);
1828 extract_scalar_result
= false;
1832 /* 2.4 Extract the final scalar result. Create:
1833 s_out3 = extract_field <v_out2, bitpos> */
1835 if (extract_scalar_result
)
1839 if (vect_print_dump_info (REPORT_DETAILS
))
1840 fprintf (vect_dump
, "extract scalar result");
1842 if (BYTES_BIG_ENDIAN
)
1843 bitpos
= size_binop (MULT_EXPR
,
1844 bitsize_int (TYPE_VECTOR_SUBPARTS (vectype
) - 1),
1845 TYPE_SIZE (scalar_type
));
1847 bitpos
= bitsize_zero_node
;
1849 rhs
= build3 (BIT_FIELD_REF
, scalar_type
, new_temp
, bitsize
, bitpos
);
1850 BIT_FIELD_REF_UNSIGNED (rhs
) = TYPE_UNSIGNED (scalar_type
);
1851 epilog_stmt
= build_gimple_modify_stmt (new_scalar_dest
, rhs
);
1852 new_temp
= make_ssa_name (new_scalar_dest
, epilog_stmt
);
1853 GIMPLE_STMT_OPERAND (epilog_stmt
, 0) = new_temp
;
1854 bsi_insert_before (&exit_bsi
, epilog_stmt
, BSI_SAME_STMT
);
1857 /* 2.4 Adjust the final result by the initial value of the reduction
1858 variable. (When such adjustment is not needed, then
1859 'scalar_initial_def' is zero).
1862 s_out4 = scalar_expr <s_out3, scalar_initial_def> */
1864 if (scalar_initial_def
)
1866 tree tmp
= build2 (code
, scalar_type
, new_temp
, scalar_initial_def
);
1867 epilog_stmt
= build_gimple_modify_stmt (new_scalar_dest
, tmp
);
1868 new_temp
= make_ssa_name (new_scalar_dest
, epilog_stmt
);
1869 GIMPLE_STMT_OPERAND (epilog_stmt
, 0) = new_temp
;
1870 bsi_insert_before (&exit_bsi
, epilog_stmt
, BSI_SAME_STMT
);
1873 /* 2.6 Replace uses of s_out0 with uses of s_out3 */
1875 /* Find the loop-closed-use at the loop exit of the original scalar result.
1876 (The reduction result is expected to have two immediate uses - one at the
1877 latch block, and one at the loop exit). */
1879 FOR_EACH_IMM_USE_FAST (use_p
, imm_iter
, scalar_dest
)
1881 if (!flow_bb_inside_loop_p (loop
, bb_for_stmt (USE_STMT (use_p
))))
1883 exit_phi
= USE_STMT (use_p
);
1887 /* We expect to have found an exit_phi because of loop-closed-ssa form. */
1888 gcc_assert (exit_phi
);
1889 /* Replace the uses: */
1890 orig_name
= PHI_RESULT (exit_phi
);
1891 FOR_EACH_IMM_USE_STMT (use_stmt
, imm_iter
, orig_name
)
1892 FOR_EACH_IMM_USE_ON_STMT (use_p
, imm_iter
)
1893 SET_USE (use_p
, new_temp
);
1897 /* Function vectorizable_reduction.
1899 Check if STMT performs a reduction operation that can be vectorized.
1900 If VEC_STMT is also passed, vectorize the STMT: create a vectorized
1901 stmt to replace it, put it in VEC_STMT, and insert it at BSI.
1902 Return FALSE if not a vectorizable STMT, TRUE otherwise.
1904 This function also handles reduction idioms (patterns) that have been
1905 recognized in advance during vect_pattern_recog. In this case, STMT may be
1907 X = pattern_expr (arg0, arg1, ..., X)
1908 and it's STMT_VINFO_RELATED_STMT points to the last stmt in the original
1909 sequence that had been detected and replaced by the pattern-stmt (STMT).
1911 In some cases of reduction patterns, the type of the reduction variable X is
1912 different than the type of the other arguments of STMT.
1913 In such cases, the vectype that is used when transforming STMT into a vector
1914 stmt is different than the vectype that is used to determine the
1915 vectorization factor, because it consists of a different number of elements
1916 than the actual number of elements that are being operated upon in parallel.
1918 For example, consider an accumulation of shorts into an int accumulator.
1919 On some targets it's possible to vectorize this pattern operating on 8
1920 shorts at a time (hence, the vectype for purposes of determining the
1921 vectorization factor should be V8HI); on the other hand, the vectype that
1922 is used to create the vector form is actually V4SI (the type of the result).
1924 Upon entry to this function, STMT_VINFO_VECTYPE records the vectype that
1925 indicates what is the actual level of parallelism (V8HI in the example), so
1926 that the right vectorization factor would be derived. This vectype
1927 corresponds to the type of arguments to the reduction stmt, and should *NOT*
1928 be used to create the vectorized stmt. The right vectype for the vectorized
1929 stmt is obtained from the type of the result X:
1930 get_vectype_for_scalar_type (TREE_TYPE (X))
1932 This means that, contrary to "regular" reductions (or "regular" stmts in
1933 general), the following equation:
1934 STMT_VINFO_VECTYPE == get_vectype_for_scalar_type (TREE_TYPE (X))
1935 does *NOT* necessarily hold for reduction patterns. */
1938 vectorizable_reduction (tree stmt
, block_stmt_iterator
*bsi
, tree
*vec_stmt
)
1943 tree loop_vec_def0
= NULL_TREE
, loop_vec_def1
= NULL_TREE
;
1944 stmt_vec_info stmt_info
= vinfo_for_stmt (stmt
);
1945 tree vectype
= STMT_VINFO_VECTYPE (stmt_info
);
1946 loop_vec_info loop_vinfo
= STMT_VINFO_LOOP_VINFO (stmt_info
);
1947 struct loop
*loop
= LOOP_VINFO_LOOP (loop_vinfo
);
1949 enum tree_code code
, orig_code
, epilog_reduc_code
= 0;
1950 enum machine_mode vec_mode
;
1952 optab optab
, reduc_optab
;
1953 tree new_temp
= NULL_TREE
;
1955 enum vect_def_type dt
;
1960 stmt_vec_info orig_stmt_info
;
1961 tree expr
= NULL_TREE
;
1963 int nunits
= TYPE_VECTOR_SUBPARTS (vectype
);
1964 int ncopies
= LOOP_VINFO_VECT_FACTOR (loop_vinfo
) / nunits
;
1965 stmt_vec_info prev_stmt_info
;
1967 tree new_stmt
= NULL_TREE
;
1970 gcc_assert (ncopies
>= 1);
1972 /* 1. Is vectorizable reduction? */
1974 /* Not supportable if the reduction variable is used in the loop. */
1975 if (STMT_VINFO_RELEVANT_P (stmt_info
))
1978 if (!STMT_VINFO_LIVE_P (stmt_info
))
1981 /* Make sure it was already recognized as a reduction computation. */
1982 if (STMT_VINFO_DEF_TYPE (stmt_info
) != vect_reduction_def
)
1985 /* 2. Has this been recognized as a reduction pattern?
1987 Check if STMT represents a pattern that has been recognized
1988 in earlier analysis stages. For stmts that represent a pattern,
1989 the STMT_VINFO_RELATED_STMT field records the last stmt in
1990 the original sequence that constitutes the pattern. */
1992 orig_stmt
= STMT_VINFO_RELATED_STMT (stmt_info
);
1995 orig_stmt_info
= vinfo_for_stmt (orig_stmt
);
1996 gcc_assert (STMT_VINFO_RELATED_STMT (orig_stmt_info
) == stmt
);
1997 gcc_assert (STMT_VINFO_IN_PATTERN_P (orig_stmt_info
));
1998 gcc_assert (!STMT_VINFO_IN_PATTERN_P (stmt_info
));
2001 /* 3. Check the operands of the operation. The first operands are defined
2002 inside the loop body. The last operand is the reduction variable,
2003 which is defined by the loop-header-phi. */
2005 gcc_assert (TREE_CODE (stmt
) == GIMPLE_MODIFY_STMT
);
2007 operation
= GIMPLE_STMT_OPERAND (stmt
, 1);
2008 code
= TREE_CODE (operation
);
2009 op_type
= TREE_OPERAND_LENGTH (operation
);
2010 if (op_type
!= binary_op
&& op_type
!= ternary_op
)
2012 scalar_dest
= GIMPLE_STMT_OPERAND (stmt
, 0);
2013 scalar_type
= TREE_TYPE (scalar_dest
);
2015 /* All uses but the last are expected to be defined in the loop.
2016 The last use is the reduction variable. */
2017 for (i
= 0; i
< op_type
-1; i
++)
2019 op
= TREE_OPERAND (operation
, i
);
2020 is_simple_use
= vect_is_simple_use (op
, loop_vinfo
, &def_stmt
, &def
, &dt
);
2021 gcc_assert (is_simple_use
);
2022 if (dt
!= vect_loop_def
2023 && dt
!= vect_invariant_def
2024 && dt
!= vect_constant_def
2025 && dt
!= vect_induction_def
)
2029 op
= TREE_OPERAND (operation
, i
);
2030 is_simple_use
= vect_is_simple_use (op
, loop_vinfo
, &def_stmt
, &def
, &dt
);
2031 gcc_assert (is_simple_use
);
2032 gcc_assert (dt
== vect_reduction_def
);
2033 gcc_assert (TREE_CODE (def_stmt
) == PHI_NODE
);
2035 gcc_assert (orig_stmt
== vect_is_simple_reduction (loop
, def_stmt
));
2037 gcc_assert (stmt
== vect_is_simple_reduction (loop
, def_stmt
));
2039 if (STMT_VINFO_LIVE_P (vinfo_for_stmt (def_stmt
)))
2042 /* 4. Supportable by target? */
2044 /* 4.1. check support for the operation in the loop */
2045 optab
= optab_for_tree_code (code
, vectype
);
2048 if (vect_print_dump_info (REPORT_DETAILS
))
2049 fprintf (vect_dump
, "no optab.");
2052 vec_mode
= TYPE_MODE (vectype
);
2053 if (optab
->handlers
[(int) vec_mode
].insn_code
== CODE_FOR_nothing
)
2055 if (vect_print_dump_info (REPORT_DETAILS
))
2056 fprintf (vect_dump
, "op not supported by target.");
2057 if (GET_MODE_SIZE (vec_mode
) != UNITS_PER_WORD
2058 || LOOP_VINFO_VECT_FACTOR (loop_vinfo
)
2059 < vect_min_worthwhile_factor (code
))
2061 if (vect_print_dump_info (REPORT_DETAILS
))
2062 fprintf (vect_dump
, "proceeding using word mode.");
2065 /* Worthwhile without SIMD support? */
2066 if (!VECTOR_MODE_P (TYPE_MODE (vectype
))
2067 && LOOP_VINFO_VECT_FACTOR (loop_vinfo
)
2068 < vect_min_worthwhile_factor (code
))
2070 if (vect_print_dump_info (REPORT_DETAILS
))
2071 fprintf (vect_dump
, "not worthwhile without SIMD support.");
2075 /* 4.2. Check support for the epilog operation.
2077 If STMT represents a reduction pattern, then the type of the
2078 reduction variable may be different than the type of the rest
2079 of the arguments. For example, consider the case of accumulation
2080 of shorts into an int accumulator; The original code:
2081 S1: int_a = (int) short_a;
2082 orig_stmt-> S2: int_acc = plus <int_a ,int_acc>;
2085 STMT: int_acc = widen_sum <short_a, int_acc>
2088 1. The tree-code that is used to create the vector operation in the
2089 epilog code (that reduces the partial results) is not the
2090 tree-code of STMT, but is rather the tree-code of the original
2091 stmt from the pattern that STMT is replacing. I.e, in the example
2092 above we want to use 'widen_sum' in the loop, but 'plus' in the
2094 2. The type (mode) we use to check available target support
2095 for the vector operation to be created in the *epilog*, is
2096 determined by the type of the reduction variable (in the example
2097 above we'd check this: plus_optab[vect_int_mode]).
2098 However the type (mode) we use to check available target support
2099 for the vector operation to be created *inside the loop*, is
2100 determined by the type of the other arguments to STMT (in the
2101 example we'd check this: widen_sum_optab[vect_short_mode]).
2103 This is contrary to "regular" reductions, in which the types of all
2104 the arguments are the same as the type of the reduction variable.
2105 For "regular" reductions we can therefore use the same vector type
2106 (and also the same tree-code) when generating the epilog code and
2107 when generating the code inside the loop. */
2111 /* This is a reduction pattern: get the vectype from the type of the
2112 reduction variable, and get the tree-code from orig_stmt. */
2113 orig_code
= TREE_CODE (GIMPLE_STMT_OPERAND (orig_stmt
, 1));
2114 vectype
= get_vectype_for_scalar_type (TREE_TYPE (def
));
2115 vec_mode
= TYPE_MODE (vectype
);
2119 /* Regular reduction: use the same vectype and tree-code as used for
2120 the vector code inside the loop can be used for the epilog code. */
2124 if (!reduction_code_for_scalar_code (orig_code
, &epilog_reduc_code
))
2126 reduc_optab
= optab_for_tree_code (epilog_reduc_code
, vectype
);
2129 if (vect_print_dump_info (REPORT_DETAILS
))
2130 fprintf (vect_dump
, "no optab for reduction.");
2131 epilog_reduc_code
= NUM_TREE_CODES
;
2133 if (reduc_optab
->handlers
[(int) vec_mode
].insn_code
== CODE_FOR_nothing
)
2135 if (vect_print_dump_info (REPORT_DETAILS
))
2136 fprintf (vect_dump
, "reduc op not supported by target.");
2137 epilog_reduc_code
= NUM_TREE_CODES
;
2140 if (!vec_stmt
) /* transformation not required. */
2142 STMT_VINFO_TYPE (stmt_info
) = reduc_vec_info_type
;
2143 vect_model_reduction_cost (stmt_info
, epilog_reduc_code
, ncopies
);
2149 if (vect_print_dump_info (REPORT_DETAILS
))
2150 fprintf (vect_dump
, "transform reduction.");
2152 /* Create the destination vector */
2153 vec_dest
= vect_create_destination_var (scalar_dest
, vectype
);
2155 /* Create the reduction-phi that defines the reduction-operand. */
2156 new_phi
= create_phi_node (vec_dest
, loop
->header
);
2158 /* In case the vectorization factor (VF) is bigger than the number
2159 of elements that we can fit in a vectype (nunits), we have to generate
2160 more than one vector stmt - i.e - we need to "unroll" the
2161 vector stmt by a factor VF/nunits. For more details see documentation
2162 in vectorizable_operation. */
2164 prev_stmt_info
= NULL
;
2165 for (j
= 0; j
< ncopies
; j
++)
2170 op
= TREE_OPERAND (operation
, 0);
2171 loop_vec_def0
= vect_get_vec_def_for_operand (op
, stmt
, NULL
);
2172 if (op_type
== ternary_op
)
2174 op
= TREE_OPERAND (operation
, 1);
2175 loop_vec_def1
= vect_get_vec_def_for_operand (op
, stmt
, NULL
);
2178 /* Get the vector def for the reduction variable from the phi node */
2179 reduc_def
= PHI_RESULT (new_phi
);
2183 enum vect_def_type dt
= vect_unknown_def_type
; /* Dummy */
2184 loop_vec_def0
= vect_get_vec_def_for_stmt_copy (dt
, loop_vec_def0
);
2185 if (op_type
== ternary_op
)
2186 loop_vec_def1
= vect_get_vec_def_for_stmt_copy (dt
, loop_vec_def1
);
2188 /* Get the vector def for the reduction variable from the vectorized
2189 reduction operation generated in the previous iteration (j-1) */
2190 reduc_def
= GIMPLE_STMT_OPERAND (new_stmt
,0);
2193 /* Arguments are ready. create the new vector stmt. */
2194 if (op_type
== binary_op
)
2195 expr
= build2 (code
, vectype
, loop_vec_def0
, reduc_def
);
2197 expr
= build3 (code
, vectype
, loop_vec_def0
, loop_vec_def1
,
2199 new_stmt
= build_gimple_modify_stmt (vec_dest
, expr
);
2200 new_temp
= make_ssa_name (vec_dest
, new_stmt
);
2201 GIMPLE_STMT_OPERAND (new_stmt
, 0) = new_temp
;
2202 vect_finish_stmt_generation (stmt
, new_stmt
, bsi
);
2205 STMT_VINFO_VEC_STMT (stmt_info
) = *vec_stmt
= new_stmt
;
2207 STMT_VINFO_RELATED_STMT (prev_stmt_info
) = new_stmt
;
2208 prev_stmt_info
= vinfo_for_stmt (new_stmt
);
2211 /* Finalize the reduction-phi (set it's arguments) and create the
2212 epilog reduction code. */
2213 vect_create_epilog_for_reduction (new_temp
, stmt
, epilog_reduc_code
, new_phi
);
2217 /* Checks if CALL can be vectorized in type VECTYPE. Returns
2218 a function declaration if the target has a vectorized version
2219 of the function, or NULL_TREE if the function cannot be vectorized. */
2222 vectorizable_function (tree call
, tree vectype_out
, tree vectype_in
)
2224 tree fndecl
= get_callee_fndecl (call
);
2225 enum built_in_function code
;
2227 /* We only handle functions that do not read or clobber memory -- i.e.
2228 const or novops ones. */
2229 if (!(call_expr_flags (call
) & (ECF_CONST
| ECF_NOVOPS
)))
2233 || TREE_CODE (fndecl
) != FUNCTION_DECL
2234 || !DECL_BUILT_IN (fndecl
))
2237 code
= DECL_FUNCTION_CODE (fndecl
);
2238 return targetm
.vectorize
.builtin_vectorized_function (code
, vectype_out
,
2242 /* Function vectorizable_call.
2244 Check if STMT performs a function call that can be vectorized.
2245 If VEC_STMT is also passed, vectorize the STMT: create a vectorized
2246 stmt to replace it, put it in VEC_STMT, and insert it at BSI.
2247 Return FALSE if not a vectorizable STMT, TRUE otherwise. */
2250 vectorizable_call (tree stmt
, block_stmt_iterator
*bsi
, tree
*vec_stmt
)
2256 tree vec_oprnd0
= NULL_TREE
, vec_oprnd1
= NULL_TREE
;
2257 stmt_vec_info stmt_info
= vinfo_for_stmt (stmt
), prev_stmt_info
;
2258 tree vectype_out
, vectype_in
;
2261 loop_vec_info loop_vinfo
= STMT_VINFO_LOOP_VINFO (stmt_info
);
2262 tree fndecl
, rhs
, new_temp
, def
, def_stmt
, rhs_type
, lhs_type
;
2263 enum vect_def_type dt
[2];
2265 int ncopies
, j
, nargs
;
2266 call_expr_arg_iterator iter
;
2268 enum { NARROW
, NONE
, WIDEN
} modifier
;
2270 if (!STMT_VINFO_RELEVANT_P (stmt_info
))
2273 if (STMT_VINFO_DEF_TYPE (stmt_info
) != vect_loop_def
)
2276 /* FORNOW: not yet supported. */
2277 if (STMT_VINFO_LIVE_P (stmt_info
))
2279 if (vect_print_dump_info (REPORT_DETAILS
))
2280 fprintf (vect_dump
, "value used after loop.");
2284 /* Is STMT a vectorizable call? */
2285 if (TREE_CODE (stmt
) != GIMPLE_MODIFY_STMT
)
2288 if (TREE_CODE (GIMPLE_STMT_OPERAND (stmt
, 0)) != SSA_NAME
)
2291 operation
= GIMPLE_STMT_OPERAND (stmt
, 1);
2292 if (TREE_CODE (operation
) != CALL_EXPR
)
2295 /* Process function arguments. */
2296 rhs_type
= NULL_TREE
;
2298 FOR_EACH_CALL_EXPR_ARG (op
, iter
, operation
)
2300 /* Bail out if the function has more than two arguments, we
2301 do not have interesting builtin functions to vectorize with
2302 more than two arguments. */
2306 /* We can only handle calls with arguments of the same type. */
2308 && rhs_type
!= TREE_TYPE (op
))
2310 if (vect_print_dump_info (REPORT_DETAILS
))
2311 fprintf (vect_dump
, "argument types differ.");
2314 rhs_type
= TREE_TYPE (op
);
2316 if (!vect_is_simple_use (op
, loop_vinfo
, &def_stmt
, &def
, &dt
[nargs
]))
2318 if (vect_print_dump_info (REPORT_DETAILS
))
2319 fprintf (vect_dump
, "use not simple.");
2326 /* No arguments is also not good. */
2330 vectype_in
= get_vectype_for_scalar_type (rhs_type
);
2331 nunits_in
= TYPE_VECTOR_SUBPARTS (vectype_in
);
2333 lhs_type
= TREE_TYPE (GIMPLE_STMT_OPERAND (stmt
, 0));
2334 vectype_out
= get_vectype_for_scalar_type (lhs_type
);
2335 nunits_out
= TYPE_VECTOR_SUBPARTS (vectype_out
);
2338 if (nunits_in
== nunits_out
/ 2)
2340 else if (nunits_out
== nunits_in
)
2342 else if (nunits_out
== nunits_in
/ 2)
2347 /* For now, we only vectorize functions if a target specific builtin
2348 is available. TODO -- in some cases, it might be profitable to
2349 insert the calls for pieces of the vector, in order to be able
2350 to vectorize other operations in the loop. */
2351 fndecl
= vectorizable_function (operation
, vectype_out
, vectype_in
);
2352 if (fndecl
== NULL_TREE
)
2354 if (vect_print_dump_info (REPORT_DETAILS
))
2355 fprintf (vect_dump
, "function is not vectorizable.");
2360 gcc_assert (ZERO_SSA_OPERANDS (stmt
, SSA_OP_ALL_VIRTUALS
));
2362 if (modifier
== NARROW
)
2363 ncopies
= LOOP_VINFO_VECT_FACTOR (loop_vinfo
) / nunits_out
;
2365 ncopies
= LOOP_VINFO_VECT_FACTOR (loop_vinfo
) / nunits_in
;
2367 /* Sanity check: make sure that at least one copy of the vectorized stmt
2368 needs to be generated. */
2369 gcc_assert (ncopies
>= 1);
2371 if (!vec_stmt
) /* transformation not required. */
2373 STMT_VINFO_TYPE (stmt_info
) = call_vec_info_type
;
2374 if (vect_print_dump_info (REPORT_DETAILS
))
2375 fprintf (vect_dump
, "=== vectorizable_call ===");
2376 vect_model_simple_cost (stmt_info
, ncopies
);
2382 if (vect_print_dump_info (REPORT_DETAILS
))
2383 fprintf (vect_dump
, "transform operation.");
2386 scalar_dest
= GIMPLE_STMT_OPERAND (stmt
, 0);
2387 vec_dest
= vect_create_destination_var (scalar_dest
, vectype_out
);
2389 prev_stmt_info
= NULL
;
2393 for (j
= 0; j
< ncopies
; ++j
)
2395 /* Build argument list for the vectorized call. */
2396 /* FIXME: Rewrite this so that it doesn't
2397 construct a temporary list. */
2400 FOR_EACH_CALL_EXPR_ARG (op
, iter
, operation
)
2404 = vect_get_vec_def_for_operand (op
, stmt
, NULL
);
2407 = vect_get_vec_def_for_stmt_copy (dt
[nargs
], vec_oprnd0
);
2409 vargs
= tree_cons (NULL_TREE
, vec_oprnd0
, vargs
);
2413 vargs
= nreverse (vargs
);
2415 rhs
= build_function_call_expr (fndecl
, vargs
);
2416 new_stmt
= build_gimple_modify_stmt (vec_dest
, rhs
);
2417 new_temp
= make_ssa_name (vec_dest
, new_stmt
);
2418 GIMPLE_STMT_OPERAND (new_stmt
, 0) = new_temp
;
2420 vect_finish_stmt_generation (stmt
, new_stmt
, bsi
);
2423 STMT_VINFO_VEC_STMT (stmt_info
) = *vec_stmt
= new_stmt
;
2425 STMT_VINFO_RELATED_STMT (prev_stmt_info
) = new_stmt
;
2427 prev_stmt_info
= vinfo_for_stmt (new_stmt
);
2433 for (j
= 0; j
< ncopies
; ++j
)
2435 /* Build argument list for the vectorized call. */
2436 /* FIXME: Rewrite this so that it doesn't
2437 construct a temporary list. */
2440 FOR_EACH_CALL_EXPR_ARG (op
, iter
, operation
)
2445 = vect_get_vec_def_for_operand (op
, stmt
, NULL
);
2447 = vect_get_vec_def_for_stmt_copy (dt
[nargs
], vec_oprnd0
);
2452 = vect_get_vec_def_for_stmt_copy (dt
[nargs
], vec_oprnd1
);
2454 = vect_get_vec_def_for_stmt_copy (dt
[nargs
], vec_oprnd0
);
2457 vargs
= tree_cons (NULL_TREE
, vec_oprnd0
, vargs
);
2458 vargs
= tree_cons (NULL_TREE
, vec_oprnd1
, vargs
);
2462 vargs
= nreverse (vargs
);
2464 rhs
= build_function_call_expr (fndecl
, vargs
);
2465 new_stmt
= build_gimple_modify_stmt (vec_dest
, rhs
);
2466 new_temp
= make_ssa_name (vec_dest
, new_stmt
);
2467 GIMPLE_STMT_OPERAND (new_stmt
, 0) = new_temp
;
2469 vect_finish_stmt_generation (stmt
, new_stmt
, bsi
);
2472 STMT_VINFO_VEC_STMT (stmt_info
) = new_stmt
;
2474 STMT_VINFO_RELATED_STMT (prev_stmt_info
) = new_stmt
;
2476 prev_stmt_info
= vinfo_for_stmt (new_stmt
);
2479 *vec_stmt
= STMT_VINFO_VEC_STMT (stmt_info
);
2484 /* No current target implements this case. */
2488 /* The call in STMT might prevent it from being removed in dce.
2489 We however cannot remove it here, due to the way the ssa name
2490 it defines is mapped to the new definition. So just replace
2491 rhs of the statement with something harmless. */
2492 type
= TREE_TYPE (scalar_dest
);
2493 GIMPLE_STMT_OPERAND (stmt
, 1) = fold_convert (type
, integer_zero_node
);
2500 /* Function vect_gen_widened_results_half
2502 Create a vector stmt whose code, type, number of arguments, and result
2503 variable are CODE, VECTYPE, OP_TYPE, and VEC_DEST, and its arguments are
2504 VEC_OPRND0 and VEC_OPRND1. The new vector stmt is to be inserted at BSI.
2505 In the case that CODE is a CALL_EXPR, this means that a call to DECL
2506 needs to be created (DECL is a function-decl of a target-builtin).
2507 STMT is the original scalar stmt that we are vectorizing. */
2510 vect_gen_widened_results_half (enum tree_code code
, tree vectype
, tree decl
,
2511 tree vec_oprnd0
, tree vec_oprnd1
, int op_type
,
2512 tree vec_dest
, block_stmt_iterator
*bsi
,
2521 /* Generate half of the widened result: */
2522 if (code
== CALL_EXPR
)
2524 /* Target specific support */
2525 if (op_type
== binary_op
)
2526 expr
= build_call_expr (decl
, 2, vec_oprnd0
, vec_oprnd1
);
2528 expr
= build_call_expr (decl
, 1, vec_oprnd0
);
2532 /* Generic support */
2533 gcc_assert (op_type
== TREE_CODE_LENGTH (code
));
2534 if (op_type
== binary_op
)
2535 expr
= build2 (code
, vectype
, vec_oprnd0
, vec_oprnd1
);
2537 expr
= build1 (code
, vectype
, vec_oprnd0
);
2539 new_stmt
= build_gimple_modify_stmt (vec_dest
, expr
);
2540 new_temp
= make_ssa_name (vec_dest
, new_stmt
);
2541 GIMPLE_STMT_OPERAND (new_stmt
, 0) = new_temp
;
2542 vect_finish_stmt_generation (stmt
, new_stmt
, bsi
);
2544 if (code
== CALL_EXPR
)
2546 FOR_EACH_SSA_TREE_OPERAND (sym
, new_stmt
, iter
, SSA_OP_ALL_VIRTUALS
)
2548 if (TREE_CODE (sym
) == SSA_NAME
)
2549 sym
= SSA_NAME_VAR (sym
);
2550 mark_sym_for_renaming (sym
);
2558 /* Function vectorizable_conversion.
2560 Check if STMT performs a conversion operation, that can be vectorized.
2561 If VEC_STMT is also passed, vectorize the STMT: create a vectorized
2562 stmt to replace it, put it in VEC_STMT, and insert it at BSI.
2563 Return FALSE if not a vectorizable STMT, TRUE otherwise. */
2566 vectorizable_conversion (tree stmt
, block_stmt_iterator
* bsi
,
2573 tree vec_oprnd0
= NULL_TREE
, vec_oprnd1
= NULL_TREE
;
2574 stmt_vec_info stmt_info
= vinfo_for_stmt (stmt
);
2575 loop_vec_info loop_vinfo
= STMT_VINFO_LOOP_VINFO (stmt_info
);
2576 enum tree_code code
, code1
= ERROR_MARK
, code2
= ERROR_MARK
;
2577 tree decl1
= NULL_TREE
, decl2
= NULL_TREE
;
2580 enum vect_def_type dt0
;
2582 stmt_vec_info prev_stmt_info
;
2585 tree vectype_out
, vectype_in
;
2588 tree rhs_type
, lhs_type
;
2590 enum { NARROW
, NONE
, WIDEN
} modifier
;
2592 /* Is STMT a vectorizable conversion? */
2594 if (!STMT_VINFO_RELEVANT_P (stmt_info
))
2597 if (STMT_VINFO_DEF_TYPE (stmt_info
) != vect_loop_def
)
2600 if (STMT_VINFO_LIVE_P (stmt_info
))
2602 /* FORNOW: not yet supported. */
2603 if (vect_print_dump_info (REPORT_DETAILS
))
2604 fprintf (vect_dump
, "value used after loop.");
2608 if (TREE_CODE (stmt
) != GIMPLE_MODIFY_STMT
)
2611 if (TREE_CODE (GIMPLE_STMT_OPERAND (stmt
, 0)) != SSA_NAME
)
2614 operation
= GIMPLE_STMT_OPERAND (stmt
, 1);
2615 code
= TREE_CODE (operation
);
2616 if (code
!= FIX_TRUNC_EXPR
&& code
!= FLOAT_EXPR
)
2619 /* Check types of lhs and rhs */
2620 op0
= TREE_OPERAND (operation
, 0);
2621 rhs_type
= TREE_TYPE (op0
);
2622 vectype_in
= get_vectype_for_scalar_type (rhs_type
);
2623 nunits_in
= TYPE_VECTOR_SUBPARTS (vectype_in
);
2625 scalar_dest
= GIMPLE_STMT_OPERAND (stmt
, 0);
2626 lhs_type
= TREE_TYPE (scalar_dest
);
2627 vectype_out
= get_vectype_for_scalar_type (lhs_type
);
2628 nunits_out
= TYPE_VECTOR_SUBPARTS (vectype_out
);
2631 if (nunits_in
== nunits_out
/ 2)
2633 else if (nunits_out
== nunits_in
)
2635 else if (nunits_out
== nunits_in
/ 2)
2640 if (modifier
== NONE
)
2641 gcc_assert (STMT_VINFO_VECTYPE (stmt_info
) == vectype_out
);
2643 /* Bail out if the types are both integral or non-integral */
2644 if ((INTEGRAL_TYPE_P (rhs_type
) && INTEGRAL_TYPE_P (lhs_type
))
2645 || (!INTEGRAL_TYPE_P (rhs_type
) && !INTEGRAL_TYPE_P (lhs_type
)))
2648 if (modifier
== NARROW
)
2649 ncopies
= LOOP_VINFO_VECT_FACTOR (loop_vinfo
) / nunits_out
;
2651 ncopies
= LOOP_VINFO_VECT_FACTOR (loop_vinfo
) / nunits_in
;
2653 /* Sanity check: make sure that at least one copy of the vectorized stmt
2654 needs to be generated. */
2655 gcc_assert (ncopies
>= 1);
2657 /* Check the operands of the operation. */
2658 if (!vect_is_simple_use (op0
, loop_vinfo
, &def_stmt
, &def
, &dt0
))
2660 if (vect_print_dump_info (REPORT_DETAILS
))
2661 fprintf (vect_dump
, "use not simple.");
2665 /* Supportable by target? */
2666 if ((modifier
== NONE
2667 && !targetm
.vectorize
.builtin_conversion (code
, vectype_in
))
2668 || (modifier
== WIDEN
2669 && !supportable_widening_operation (code
, stmt
, vectype_in
,
2672 || (modifier
== NARROW
2673 && !supportable_narrowing_operation (code
, stmt
, vectype_in
,
2676 if (vect_print_dump_info (REPORT_DETAILS
))
2677 fprintf (vect_dump
, "op not supported by target.");
2681 if (modifier
!= NONE
)
2682 STMT_VINFO_VECTYPE (stmt_info
) = vectype_in
;
2684 if (!vec_stmt
) /* transformation not required. */
2686 STMT_VINFO_TYPE (stmt_info
) = type_conversion_vec_info_type
;
2691 if (vect_print_dump_info (REPORT_DETAILS
))
2692 fprintf (vect_dump
, "transform conversion.");
2695 vec_dest
= vect_create_destination_var (scalar_dest
, vectype_out
);
2697 prev_stmt_info
= NULL
;
2701 for (j
= 0; j
< ncopies
; j
++)
2707 vec_oprnd0
= vect_get_vec_def_for_operand (op0
, stmt
, NULL
);
2709 vec_oprnd0
= vect_get_vec_def_for_stmt_copy (dt0
, vec_oprnd0
);
2712 targetm
.vectorize
.builtin_conversion (code
, vectype_in
);
2713 new_stmt
= build_call_expr (builtin_decl
, 1, vec_oprnd0
);
2715 /* Arguments are ready. create the new vector stmt. */
2716 new_stmt
= build_gimple_modify_stmt (vec_dest
, new_stmt
);
2717 new_temp
= make_ssa_name (vec_dest
, new_stmt
);
2718 GIMPLE_STMT_OPERAND (new_stmt
, 0) = new_temp
;
2719 vect_finish_stmt_generation (stmt
, new_stmt
, bsi
);
2720 FOR_EACH_SSA_TREE_OPERAND (sym
, new_stmt
, iter
, SSA_OP_ALL_VIRTUALS
)
2722 if (TREE_CODE (sym
) == SSA_NAME
)
2723 sym
= SSA_NAME_VAR (sym
);
2724 mark_sym_for_renaming (sym
);
2728 STMT_VINFO_VEC_STMT (stmt_info
) = *vec_stmt
= new_stmt
;
2730 STMT_VINFO_RELATED_STMT (prev_stmt_info
) = new_stmt
;
2731 prev_stmt_info
= vinfo_for_stmt (new_stmt
);
2736 /* In case the vectorization factor (VF) is bigger than the number
2737 of elements that we can fit in a vectype (nunits), we have to
2738 generate more than one vector stmt - i.e - we need to "unroll"
2739 the vector stmt by a factor VF/nunits. */
2740 for (j
= 0; j
< ncopies
; j
++)
2743 vec_oprnd0
= vect_get_vec_def_for_operand (op0
, stmt
, NULL
);
2745 vec_oprnd0
= vect_get_vec_def_for_stmt_copy (dt0
, vec_oprnd0
);
2747 STMT_VINFO_VECTYPE (stmt_info
) = vectype_in
;
2749 /* Generate first half of the widened result: */
2751 = vect_gen_widened_results_half (code1
, vectype_out
, decl1
,
2752 vec_oprnd0
, vec_oprnd1
,
2753 unary_op
, vec_dest
, bsi
, stmt
);
2755 STMT_VINFO_VEC_STMT (stmt_info
) = new_stmt
;
2757 STMT_VINFO_RELATED_STMT (prev_stmt_info
) = new_stmt
;
2758 prev_stmt_info
= vinfo_for_stmt (new_stmt
);
2760 /* Generate second half of the widened result: */
2762 = vect_gen_widened_results_half (code2
, vectype_out
, decl2
,
2763 vec_oprnd0
, vec_oprnd1
,
2764 unary_op
, vec_dest
, bsi
, stmt
);
2765 STMT_VINFO_RELATED_STMT (prev_stmt_info
) = new_stmt
;
2766 prev_stmt_info
= vinfo_for_stmt (new_stmt
);
2771 /* In case the vectorization factor (VF) is bigger than the number
2772 of elements that we can fit in a vectype (nunits), we have to
2773 generate more than one vector stmt - i.e - we need to "unroll"
2774 the vector stmt by a factor VF/nunits. */
2775 for (j
= 0; j
< ncopies
; j
++)
2780 vec_oprnd0
= vect_get_vec_def_for_operand (op0
, stmt
, NULL
);
2781 vec_oprnd1
= vect_get_vec_def_for_stmt_copy (dt0
, vec_oprnd0
);
2785 vec_oprnd0
= vect_get_vec_def_for_stmt_copy (dt0
, vec_oprnd1
);
2786 vec_oprnd1
= vect_get_vec_def_for_stmt_copy (dt0
, vec_oprnd0
);
2789 /* Arguments are ready. Create the new vector stmt. */
2790 expr
= build2 (code1
, vectype_out
, vec_oprnd0
, vec_oprnd1
);
2791 new_stmt
= build_gimple_modify_stmt (vec_dest
, expr
);
2792 new_temp
= make_ssa_name (vec_dest
, new_stmt
);
2793 GIMPLE_STMT_OPERAND (new_stmt
, 0) = new_temp
;
2794 vect_finish_stmt_generation (stmt
, new_stmt
, bsi
);
2797 STMT_VINFO_VEC_STMT (stmt_info
) = new_stmt
;
2799 STMT_VINFO_RELATED_STMT (prev_stmt_info
) = new_stmt
;
2801 prev_stmt_info
= vinfo_for_stmt (new_stmt
);
2804 *vec_stmt
= STMT_VINFO_VEC_STMT (stmt_info
);
2810 /* Function vectorizable_assignment.
2812 Check if STMT performs an assignment (copy) that can be vectorized.
2813 If VEC_STMT is also passed, vectorize the STMT: create a vectorized
2814 stmt to replace it, put it in VEC_STMT, and insert it at BSI.
2815 Return FALSE if not a vectorizable STMT, TRUE otherwise. */
2818 vectorizable_assignment (tree stmt
, block_stmt_iterator
*bsi
, tree
*vec_stmt
)
2824 stmt_vec_info stmt_info
= vinfo_for_stmt (stmt
);
2825 tree vectype
= STMT_VINFO_VECTYPE (stmt_info
);
2826 loop_vec_info loop_vinfo
= STMT_VINFO_LOOP_VINFO (stmt_info
);
2829 enum vect_def_type dt
;
2830 int nunits
= TYPE_VECTOR_SUBPARTS (vectype
);
2831 int ncopies
= LOOP_VINFO_VECT_FACTOR (loop_vinfo
) / nunits
;
2833 gcc_assert (ncopies
>= 1);
2835 return false; /* FORNOW */
2837 if (!STMT_VINFO_RELEVANT_P (stmt_info
))
2840 if (STMT_VINFO_DEF_TYPE (stmt_info
) != vect_loop_def
)
2843 /* FORNOW: not yet supported. */
2844 if (STMT_VINFO_LIVE_P (stmt_info
))
2846 if (vect_print_dump_info (REPORT_DETAILS
))
2847 fprintf (vect_dump
, "value used after loop.");
2851 /* Is vectorizable assignment? */
2852 if (TREE_CODE (stmt
) != GIMPLE_MODIFY_STMT
)
2855 scalar_dest
= GIMPLE_STMT_OPERAND (stmt
, 0);
2856 if (TREE_CODE (scalar_dest
) != SSA_NAME
)
2859 op
= GIMPLE_STMT_OPERAND (stmt
, 1);
2860 if (!vect_is_simple_use (op
, loop_vinfo
, &def_stmt
, &def
, &dt
))
2862 if (vect_print_dump_info (REPORT_DETAILS
))
2863 fprintf (vect_dump
, "use not simple.");
2867 if (!vec_stmt
) /* transformation not required. */
2869 STMT_VINFO_TYPE (stmt_info
) = assignment_vec_info_type
;
2870 if (vect_print_dump_info (REPORT_DETAILS
))
2871 fprintf (vect_dump
, "=== vectorizable_assignment ===");
2872 vect_model_simple_cost (stmt_info
, ncopies
);
2877 if (vect_print_dump_info (REPORT_DETAILS
))
2878 fprintf (vect_dump
, "transform assignment.");
2881 vec_dest
= vect_create_destination_var (scalar_dest
, vectype
);
2884 op
= GIMPLE_STMT_OPERAND (stmt
, 1);
2885 vec_oprnd
= vect_get_vec_def_for_operand (op
, stmt
, NULL
);
2887 /* Arguments are ready. create the new vector stmt. */
2888 *vec_stmt
= build_gimple_modify_stmt (vec_dest
, vec_oprnd
);
2889 new_temp
= make_ssa_name (vec_dest
, *vec_stmt
);
2890 GIMPLE_STMT_OPERAND (*vec_stmt
, 0) = new_temp
;
2891 vect_finish_stmt_generation (stmt
, *vec_stmt
, bsi
);
2897 /* Function vect_min_worthwhile_factor.
2899 For a loop where we could vectorize the operation indicated by CODE,
2900 return the minimum vectorization factor that makes it worthwhile
2901 to use generic vectors. */
2903 vect_min_worthwhile_factor (enum tree_code code
)
2924 /* Function vectorizable_induction
2926 Check if PHI performs an induction computation that can be vectorized.
2927 If VEC_STMT is also passed, vectorize the induction PHI: create a vectorized
2928 phi to replace it, put it in VEC_STMT, and add it to the same basic block.
2929 Return FALSE if not a vectorizable STMT, TRUE otherwise. */
2932 vectorizable_induction (tree phi
, block_stmt_iterator
*bsi ATTRIBUTE_UNUSED
,
2935 stmt_vec_info stmt_info
= vinfo_for_stmt (phi
);
2936 tree vectype
= STMT_VINFO_VECTYPE (stmt_info
);
2937 loop_vec_info loop_vinfo
= STMT_VINFO_LOOP_VINFO (stmt_info
);
2938 int nunits
= TYPE_VECTOR_SUBPARTS (vectype
);
2939 int ncopies
= LOOP_VINFO_VECT_FACTOR (loop_vinfo
) / nunits
;
2942 gcc_assert (ncopies
>= 1);
2944 if (!STMT_VINFO_RELEVANT_P (stmt_info
))
2947 gcc_assert (STMT_VINFO_DEF_TYPE (stmt_info
) == vect_induction_def
);
2949 if (STMT_VINFO_LIVE_P (stmt_info
))
2951 /* FORNOW: not yet supported. */
2952 if (vect_print_dump_info (REPORT_DETAILS
))
2953 fprintf (vect_dump
, "value used after loop.");
2957 if (TREE_CODE (phi
) != PHI_NODE
)
2960 if (!vec_stmt
) /* transformation not required. */
2962 STMT_VINFO_TYPE (stmt_info
) = induc_vec_info_type
;
2963 if (vect_print_dump_info (REPORT_DETAILS
))
2964 fprintf (vect_dump
, "=== vectorizable_induction ===");
2965 vect_model_induction_cost (stmt_info
, ncopies
);
2971 if (vect_print_dump_info (REPORT_DETAILS
))
2972 fprintf (vect_dump
, "transform induction phi.");
2974 vec_def
= get_initial_def_for_induction (phi
);
2975 *vec_stmt
= SSA_NAME_DEF_STMT (vec_def
);
2980 /* Function vectorizable_operation.
2982 Check if STMT performs a binary or unary operation that can be vectorized.
2983 If VEC_STMT is also passed, vectorize the STMT: create a vectorized
2984 stmt to replace it, put it in VEC_STMT, and insert it at BSI.
2985 Return FALSE if not a vectorizable STMT, TRUE otherwise. */
2988 vectorizable_operation (tree stmt
, block_stmt_iterator
*bsi
, tree
*vec_stmt
)
2993 tree op0
, op1
= NULL
;
2994 tree vec_oprnd0
= NULL_TREE
, vec_oprnd1
= NULL_TREE
;
2995 stmt_vec_info stmt_info
= vinfo_for_stmt (stmt
);
2996 tree vectype
= STMT_VINFO_VECTYPE (stmt_info
);
2997 loop_vec_info loop_vinfo
= STMT_VINFO_LOOP_VINFO (stmt_info
);
2998 enum tree_code code
;
2999 enum machine_mode vec_mode
;
3004 enum machine_mode optab_op2_mode
;
3006 enum vect_def_type dt0
, dt1
;
3008 stmt_vec_info prev_stmt_info
;
3009 int nunits_in
= TYPE_VECTOR_SUBPARTS (vectype
);
3012 int ncopies
= LOOP_VINFO_VECT_FACTOR (loop_vinfo
) / nunits_in
;
3015 gcc_assert (ncopies
>= 1);
3017 if (!STMT_VINFO_RELEVANT_P (stmt_info
))
3020 if (STMT_VINFO_DEF_TYPE (stmt_info
) != vect_loop_def
)
3023 /* FORNOW: not yet supported. */
3024 if (STMT_VINFO_LIVE_P (stmt_info
))
3026 if (vect_print_dump_info (REPORT_DETAILS
))
3027 fprintf (vect_dump
, "value used after loop.");
3031 /* Is STMT a vectorizable binary/unary operation? */
3032 if (TREE_CODE (stmt
) != GIMPLE_MODIFY_STMT
)
3035 if (TREE_CODE (GIMPLE_STMT_OPERAND (stmt
, 0)) != SSA_NAME
)
3038 scalar_dest
= GIMPLE_STMT_OPERAND (stmt
, 0);
3039 vectype_out
= get_vectype_for_scalar_type (TREE_TYPE (scalar_dest
));
3040 nunits_out
= TYPE_VECTOR_SUBPARTS (vectype_out
);
3041 if (nunits_out
!= nunits_in
)
3044 operation
= GIMPLE_STMT_OPERAND (stmt
, 1);
3045 code
= TREE_CODE (operation
);
3047 /* For pointer addition, we should use the normal plus for
3048 the vector addition. */
3049 if (code
== POINTER_PLUS_EXPR
)
3052 optab
= optab_for_tree_code (code
, vectype
);
3054 /* Support only unary or binary operations. */
3055 op_type
= TREE_OPERAND_LENGTH (operation
);
3056 if (op_type
!= unary_op
&& op_type
!= binary_op
)
3058 if (vect_print_dump_info (REPORT_DETAILS
))
3059 fprintf (vect_dump
, "num. args = %d (not unary/binary op).", op_type
);
3063 op0
= TREE_OPERAND (operation
, 0);
3064 if (!vect_is_simple_use (op0
, loop_vinfo
, &def_stmt
, &def
, &dt0
))
3066 if (vect_print_dump_info (REPORT_DETAILS
))
3067 fprintf (vect_dump
, "use not simple.");
3071 if (op_type
== binary_op
)
3073 op1
= TREE_OPERAND (operation
, 1);
3074 if (!vect_is_simple_use (op1
, loop_vinfo
, &def_stmt
, &def
, &dt1
))
3076 if (vect_print_dump_info (REPORT_DETAILS
))
3077 fprintf (vect_dump
, "use not simple.");
3082 /* Supportable by target? */
3085 if (vect_print_dump_info (REPORT_DETAILS
))
3086 fprintf (vect_dump
, "no optab.");
3089 vec_mode
= TYPE_MODE (vectype
);
3090 icode
= (int) optab
->handlers
[(int) vec_mode
].insn_code
;
3091 if (icode
== CODE_FOR_nothing
)
3093 if (vect_print_dump_info (REPORT_DETAILS
))
3094 fprintf (vect_dump
, "op not supported by target.");
3095 if (GET_MODE_SIZE (vec_mode
) != UNITS_PER_WORD
3096 || LOOP_VINFO_VECT_FACTOR (loop_vinfo
)
3097 < vect_min_worthwhile_factor (code
))
3099 if (vect_print_dump_info (REPORT_DETAILS
))
3100 fprintf (vect_dump
, "proceeding using word mode.");
3103 /* Worthwhile without SIMD support? */
3104 if (!VECTOR_MODE_P (TYPE_MODE (vectype
))
3105 && LOOP_VINFO_VECT_FACTOR (loop_vinfo
)
3106 < vect_min_worthwhile_factor (code
))
3108 if (vect_print_dump_info (REPORT_DETAILS
))
3109 fprintf (vect_dump
, "not worthwhile without SIMD support.");
3113 if (code
== LSHIFT_EXPR
|| code
== RSHIFT_EXPR
)
3115 /* FORNOW: not yet supported. */
3116 if (!VECTOR_MODE_P (vec_mode
))
3119 /* Invariant argument is needed for a vector shift
3120 by a scalar shift operand. */
3121 optab_op2_mode
= insn_data
[icode
].operand
[2].mode
;
3122 if (! (VECTOR_MODE_P (optab_op2_mode
)
3123 || dt1
== vect_constant_def
3124 || dt1
== vect_invariant_def
))
3126 if (vect_print_dump_info (REPORT_DETAILS
))
3127 fprintf (vect_dump
, "operand mode requires invariant argument.");
3132 if (!vec_stmt
) /* transformation not required. */
3134 STMT_VINFO_TYPE (stmt_info
) = op_vec_info_type
;
3135 if (vect_print_dump_info (REPORT_DETAILS
))
3136 fprintf (vect_dump
, "=== vectorizable_operation ===");
3137 vect_model_simple_cost (stmt_info
, ncopies
);
3143 if (vect_print_dump_info (REPORT_DETAILS
))
3144 fprintf (vect_dump
, "transform binary/unary operation.");
3147 vec_dest
= vect_create_destination_var (scalar_dest
, vectype
);
3149 /* In case the vectorization factor (VF) is bigger than the number
3150 of elements that we can fit in a vectype (nunits), we have to generate
3151 more than one vector stmt - i.e - we need to "unroll" the
3152 vector stmt by a factor VF/nunits. In doing so, we record a pointer
3153 from one copy of the vector stmt to the next, in the field
3154 STMT_VINFO_RELATED_STMT. This is necessary in order to allow following
3155 stages to find the correct vector defs to be used when vectorizing
3156 stmts that use the defs of the current stmt. The example below illustrates
3157 the vectorization process when VF=16 and nunits=4 (i.e - we need to create
3158 4 vectorized stmts):
3160 before vectorization:
3161 RELATED_STMT VEC_STMT
3165 step 1: vectorize stmt S1 (done in vectorizable_load. See more details
3167 RELATED_STMT VEC_STMT
3168 VS1_0: vx0 = memref0 VS1_1 -
3169 VS1_1: vx1 = memref1 VS1_2 -
3170 VS1_2: vx2 = memref2 VS1_3 -
3171 VS1_3: vx3 = memref3 - -
3172 S1: x = load - VS1_0
3175 step2: vectorize stmt S2 (done here):
3176 To vectorize stmt S2 we first need to find the relevant vector
3177 def for the first operand 'x'. This is, as usual, obtained from
3178 the vector stmt recorded in the STMT_VINFO_VEC_STMT of the stmt
3179 that defines 'x' (S1). This way we find the stmt VS1_0, and the
3180 relevant vector def 'vx0'. Having found 'vx0' we can generate
3181 the vector stmt VS2_0, and as usual, record it in the
3182 STMT_VINFO_VEC_STMT of stmt S2.
3183 When creating the second copy (VS2_1), we obtain the relevant vector
3184 def from the vector stmt recorded in the STMT_VINFO_RELATED_STMT of
3185 stmt VS1_0. This way we find the stmt VS1_1 and the relevant
3186 vector def 'vx1'. Using 'vx1' we create stmt VS2_1 and record a
3187 pointer to it in the STMT_VINFO_RELATED_STMT of the vector stmt VS2_0.
3188 Similarly when creating stmts VS2_2 and VS2_3. This is the resulting
3189 chain of stmts and pointers:
3190 RELATED_STMT VEC_STMT
3191 VS1_0: vx0 = memref0 VS1_1 -
3192 VS1_1: vx1 = memref1 VS1_2 -
3193 VS1_2: vx2 = memref2 VS1_3 -
3194 VS1_3: vx3 = memref3 - -
3195 S1: x = load - VS1_0
3196 VS2_0: vz0 = vx0 + v1 VS2_1 -
3197 VS2_1: vz1 = vx1 + v1 VS2_2 -
3198 VS2_2: vz2 = vx2 + v1 VS2_3 -
3199 VS2_3: vz3 = vx3 + v1 - -
3200 S2: z = x + 1 - VS2_0 */
3202 prev_stmt_info
= NULL
;
3203 for (j
= 0; j
< ncopies
; j
++)
3208 vec_oprnd0
= vect_get_vec_def_for_operand (op0
, stmt
, NULL
);
3209 if (op_type
== binary_op
)
3211 if (code
== LSHIFT_EXPR
|| code
== RSHIFT_EXPR
)
3213 /* Vector shl and shr insn patterns can be defined with
3214 scalar operand 2 (shift operand). In this case, use
3215 constant or loop invariant op1 directly, without
3216 extending it to vector mode first. */
3217 optab_op2_mode
= insn_data
[icode
].operand
[2].mode
;
3218 if (!VECTOR_MODE_P (optab_op2_mode
))
3220 if (vect_print_dump_info (REPORT_DETAILS
))
3221 fprintf (vect_dump
, "operand 1 using scalar mode.");
3226 vec_oprnd1
= vect_get_vec_def_for_operand (op1
, stmt
, NULL
);
3231 vec_oprnd0
= vect_get_vec_def_for_stmt_copy (dt0
, vec_oprnd0
);
3232 if (op_type
== binary_op
)
3233 vec_oprnd1
= vect_get_vec_def_for_stmt_copy (dt1
, vec_oprnd1
);
3236 /* Arguments are ready. create the new vector stmt. */
3238 if (op_type
== binary_op
)
3239 new_stmt
= build_gimple_modify_stmt (vec_dest
,
3240 build2 (code
, vectype
, vec_oprnd0
, vec_oprnd1
));
3242 new_stmt
= build_gimple_modify_stmt (vec_dest
,
3243 build1 (code
, vectype
, vec_oprnd0
));
3244 new_temp
= make_ssa_name (vec_dest
, new_stmt
);
3245 GIMPLE_STMT_OPERAND (new_stmt
, 0) = new_temp
;
3246 vect_finish_stmt_generation (stmt
, new_stmt
, bsi
);
3249 STMT_VINFO_VEC_STMT (stmt_info
) = *vec_stmt
= new_stmt
;
3251 STMT_VINFO_RELATED_STMT (prev_stmt_info
) = new_stmt
;
3252 prev_stmt_info
= vinfo_for_stmt (new_stmt
);
3259 /* Function vectorizable_type_demotion
3261 Check if STMT performs a binary or unary operation that involves
3262 type demotion, and if it can be vectorized.
3263 If VEC_STMT is also passed, vectorize the STMT: create a vectorized
3264 stmt to replace it, put it in VEC_STMT, and insert it at BSI.
3265 Return FALSE if not a vectorizable STMT, TRUE otherwise. */
3268 vectorizable_type_demotion (tree stmt
, block_stmt_iterator
*bsi
,
3275 tree vec_oprnd0
=NULL
, vec_oprnd1
=NULL
;
3276 stmt_vec_info stmt_info
= vinfo_for_stmt (stmt
);
3277 loop_vec_info loop_vinfo
= STMT_VINFO_LOOP_VINFO (stmt_info
);
3278 enum tree_code code
, code1
= ERROR_MARK
;
3281 enum vect_def_type dt0
;
3283 stmt_vec_info prev_stmt_info
;
3292 if (!STMT_VINFO_RELEVANT_P (stmt_info
))
3295 if (STMT_VINFO_DEF_TYPE (stmt_info
) != vect_loop_def
)
3298 /* FORNOW: not yet supported. */
3299 if (STMT_VINFO_LIVE_P (stmt_info
))
3301 if (vect_print_dump_info (REPORT_DETAILS
))
3302 fprintf (vect_dump
, "value used after loop.");
3306 /* Is STMT a vectorizable type-demotion operation? */
3307 if (TREE_CODE (stmt
) != GIMPLE_MODIFY_STMT
)
3310 if (TREE_CODE (GIMPLE_STMT_OPERAND (stmt
, 0)) != SSA_NAME
)
3313 operation
= GIMPLE_STMT_OPERAND (stmt
, 1);
3314 code
= TREE_CODE (operation
);
3315 if (code
!= NOP_EXPR
&& code
!= CONVERT_EXPR
)
3318 op0
= TREE_OPERAND (operation
, 0);
3319 vectype_in
= get_vectype_for_scalar_type (TREE_TYPE (op0
));
3320 nunits_in
= TYPE_VECTOR_SUBPARTS (vectype_in
);
3322 scalar_dest
= GIMPLE_STMT_OPERAND (stmt
, 0);
3323 vectype_out
= get_vectype_for_scalar_type (TREE_TYPE (scalar_dest
));
3324 nunits_out
= TYPE_VECTOR_SUBPARTS (vectype_out
);
3325 if (nunits_in
!= nunits_out
/ 2) /* FORNOW */
3328 ncopies
= LOOP_VINFO_VECT_FACTOR (loop_vinfo
) / nunits_out
;
3329 gcc_assert (ncopies
>= 1);
3331 if (! ((INTEGRAL_TYPE_P (TREE_TYPE (scalar_dest
))
3332 && INTEGRAL_TYPE_P (TREE_TYPE (op0
)))
3333 || (SCALAR_FLOAT_TYPE_P (TREE_TYPE (scalar_dest
))
3334 && SCALAR_FLOAT_TYPE_P (TREE_TYPE (op0
))
3335 && (code
== NOP_EXPR
|| code
== CONVERT_EXPR
))))
3338 /* Check the operands of the operation. */
3339 if (!vect_is_simple_use (op0
, loop_vinfo
, &def_stmt
, &def
, &dt0
))
3341 if (vect_print_dump_info (REPORT_DETAILS
))
3342 fprintf (vect_dump
, "use not simple.");
3346 /* Supportable by target? */
3347 if (!supportable_narrowing_operation (code
, stmt
, vectype_in
, &code1
))
3350 STMT_VINFO_VECTYPE (stmt_info
) = vectype_in
;
3352 if (!vec_stmt
) /* transformation not required. */
3354 STMT_VINFO_TYPE (stmt_info
) = type_demotion_vec_info_type
;
3355 if (vect_print_dump_info (REPORT_DETAILS
))
3356 fprintf (vect_dump
, "=== vectorizable_demotion ===");
3357 vect_model_simple_cost (stmt_info
, ncopies
);
3362 if (vect_print_dump_info (REPORT_DETAILS
))
3363 fprintf (vect_dump
, "transform type demotion operation. ncopies = %d.",
3367 vec_dest
= vect_create_destination_var (scalar_dest
, vectype_out
);
3369 /* In case the vectorization factor (VF) is bigger than the number
3370 of elements that we can fit in a vectype (nunits), we have to generate
3371 more than one vector stmt - i.e - we need to "unroll" the
3372 vector stmt by a factor VF/nunits. */
3373 prev_stmt_info
= NULL
;
3374 for (j
= 0; j
< ncopies
; j
++)
3379 vec_oprnd0
= vect_get_vec_def_for_operand (op0
, stmt
, NULL
);
3380 vec_oprnd1
= vect_get_vec_def_for_stmt_copy (dt0
, vec_oprnd0
);
3384 vec_oprnd0
= vect_get_vec_def_for_stmt_copy (dt0
, vec_oprnd1
);
3385 vec_oprnd1
= vect_get_vec_def_for_stmt_copy (dt0
, vec_oprnd0
);
3388 /* Arguments are ready. Create the new vector stmt. */
3389 expr
= build2 (code1
, vectype_out
, vec_oprnd0
, vec_oprnd1
);
3390 new_stmt
= build_gimple_modify_stmt (vec_dest
, expr
);
3391 new_temp
= make_ssa_name (vec_dest
, new_stmt
);
3392 GIMPLE_STMT_OPERAND (new_stmt
, 0) = new_temp
;
3393 vect_finish_stmt_generation (stmt
, new_stmt
, bsi
);
3396 STMT_VINFO_VEC_STMT (stmt_info
) = new_stmt
;
3398 STMT_VINFO_RELATED_STMT (prev_stmt_info
) = new_stmt
;
3400 prev_stmt_info
= vinfo_for_stmt (new_stmt
);
3403 *vec_stmt
= STMT_VINFO_VEC_STMT (stmt_info
);
3408 /* Function vectorizable_type_promotion
3410 Check if STMT performs a binary or unary operation that involves
3411 type promotion, and if it can be vectorized.
3412 If VEC_STMT is also passed, vectorize the STMT: create a vectorized
3413 stmt to replace it, put it in VEC_STMT, and insert it at BSI.
3414 Return FALSE if not a vectorizable STMT, TRUE otherwise. */
3417 vectorizable_type_promotion (tree stmt
, block_stmt_iterator
*bsi
,
3423 tree op0
, op1
= NULL
;
3424 tree vec_oprnd0
=NULL
, vec_oprnd1
=NULL
;
3425 stmt_vec_info stmt_info
= vinfo_for_stmt (stmt
);
3426 loop_vec_info loop_vinfo
= STMT_VINFO_LOOP_VINFO (stmt_info
);
3427 enum tree_code code
, code1
= ERROR_MARK
, code2
= ERROR_MARK
;
3428 tree decl1
= NULL_TREE
, decl2
= NULL_TREE
;
3431 enum vect_def_type dt0
, dt1
;
3433 stmt_vec_info prev_stmt_info
;
3441 if (!STMT_VINFO_RELEVANT_P (stmt_info
))
3444 if (STMT_VINFO_DEF_TYPE (stmt_info
) != vect_loop_def
)
3447 /* FORNOW: not yet supported. */
3448 if (STMT_VINFO_LIVE_P (stmt_info
))
3450 if (vect_print_dump_info (REPORT_DETAILS
))
3451 fprintf (vect_dump
, "value used after loop.");
3455 /* Is STMT a vectorizable type-promotion operation? */
3456 if (TREE_CODE (stmt
) != GIMPLE_MODIFY_STMT
)
3459 if (TREE_CODE (GIMPLE_STMT_OPERAND (stmt
, 0)) != SSA_NAME
)
3462 operation
= GIMPLE_STMT_OPERAND (stmt
, 1);
3463 code
= TREE_CODE (operation
);
3464 if (code
!= NOP_EXPR
&& code
!= CONVERT_EXPR
3465 && code
!= WIDEN_MULT_EXPR
)
3468 op0
= TREE_OPERAND (operation
, 0);
3469 vectype_in
= get_vectype_for_scalar_type (TREE_TYPE (op0
));
3470 nunits_in
= TYPE_VECTOR_SUBPARTS (vectype_in
);
3472 scalar_dest
= GIMPLE_STMT_OPERAND (stmt
, 0);
3473 vectype_out
= get_vectype_for_scalar_type (TREE_TYPE (scalar_dest
));
3474 nunits_out
= TYPE_VECTOR_SUBPARTS (vectype_out
);
3475 if (nunits_out
!= nunits_in
/ 2) /* FORNOW */
3478 ncopies
= LOOP_VINFO_VECT_FACTOR (loop_vinfo
) / nunits_in
;
3479 gcc_assert (ncopies
>= 1);
3481 if (! ((INTEGRAL_TYPE_P (TREE_TYPE (scalar_dest
))
3482 && INTEGRAL_TYPE_P (TREE_TYPE (op0
)))
3483 || (SCALAR_FLOAT_TYPE_P (TREE_TYPE (scalar_dest
))
3484 && SCALAR_FLOAT_TYPE_P (TREE_TYPE (op0
))
3485 && (code
== CONVERT_EXPR
|| code
== NOP_EXPR
))))
3488 /* Check the operands of the operation. */
3489 if (!vect_is_simple_use (op0
, loop_vinfo
, &def_stmt
, &def
, &dt0
))
3491 if (vect_print_dump_info (REPORT_DETAILS
))
3492 fprintf (vect_dump
, "use not simple.");
3496 op_type
= TREE_CODE_LENGTH (code
);
3497 if (op_type
== binary_op
)
3499 op1
= TREE_OPERAND (operation
, 1);
3500 if (!vect_is_simple_use (op1
, loop_vinfo
, &def_stmt
, &def
, &dt1
))
3502 if (vect_print_dump_info (REPORT_DETAILS
))
3503 fprintf (vect_dump
, "use not simple.");
3508 /* Supportable by target? */
3509 if (!supportable_widening_operation (code
, stmt
, vectype_in
,
3510 &decl1
, &decl2
, &code1
, &code2
))
3513 STMT_VINFO_VECTYPE (stmt_info
) = vectype_in
;
3515 if (!vec_stmt
) /* transformation not required. */
3517 STMT_VINFO_TYPE (stmt_info
) = type_promotion_vec_info_type
;
3518 if (vect_print_dump_info (REPORT_DETAILS
))
3519 fprintf (vect_dump
, "=== vectorizable_promotion ===");
3520 vect_model_simple_cost (stmt_info
, 2*ncopies
);
3526 if (vect_print_dump_info (REPORT_DETAILS
))
3527 fprintf (vect_dump
, "transform type promotion operation. ncopies = %d.",
3531 vec_dest
= vect_create_destination_var (scalar_dest
, vectype_out
);
3533 /* In case the vectorization factor (VF) is bigger than the number
3534 of elements that we can fit in a vectype (nunits), we have to generate
3535 more than one vector stmt - i.e - we need to "unroll" the
3536 vector stmt by a factor VF/nunits. */
3538 prev_stmt_info
= NULL
;
3539 for (j
= 0; j
< ncopies
; j
++)
3544 vec_oprnd0
= vect_get_vec_def_for_operand (op0
, stmt
, NULL
);
3545 if (op_type
== binary_op
)
3546 vec_oprnd1
= vect_get_vec_def_for_operand (op1
, stmt
, NULL
);
3550 vec_oprnd0
= vect_get_vec_def_for_stmt_copy (dt0
, vec_oprnd0
);
3551 if (op_type
== binary_op
)
3552 vec_oprnd1
= vect_get_vec_def_for_stmt_copy (dt1
, vec_oprnd1
);
3555 /* Arguments are ready. Create the new vector stmt. We are creating
3556 two vector defs because the widened result does not fit in one vector.
3557 The vectorized stmt can be expressed as a call to a taregt builtin,
3558 or a using a tree-code. */
3559 /* Generate first half of the widened result: */
3560 new_stmt
= vect_gen_widened_results_half (code1
, vectype_out
, decl1
,
3561 vec_oprnd0
, vec_oprnd1
, op_type
, vec_dest
, bsi
, stmt
);
3563 STMT_VINFO_VEC_STMT (stmt_info
) = new_stmt
;
3565 STMT_VINFO_RELATED_STMT (prev_stmt_info
) = new_stmt
;
3566 prev_stmt_info
= vinfo_for_stmt (new_stmt
);
3568 /* Generate second half of the widened result: */
3569 new_stmt
= vect_gen_widened_results_half (code2
, vectype_out
, decl2
,
3570 vec_oprnd0
, vec_oprnd1
, op_type
, vec_dest
, bsi
, stmt
);
3571 STMT_VINFO_RELATED_STMT (prev_stmt_info
) = new_stmt
;
3572 prev_stmt_info
= vinfo_for_stmt (new_stmt
);
3576 *vec_stmt
= STMT_VINFO_VEC_STMT (stmt_info
);
3581 /* Function vect_strided_store_supported.
3583 Returns TRUE is INTERLEAVE_HIGH and INTERLEAVE_LOW operations are supported,
3584 and FALSE otherwise. */
3587 vect_strided_store_supported (tree vectype
)
3589 optab interleave_high_optab
, interleave_low_optab
;
3592 mode
= (int) TYPE_MODE (vectype
);
3594 /* Check that the operation is supported. */
3595 interleave_high_optab
= optab_for_tree_code (VEC_INTERLEAVE_HIGH_EXPR
,
3597 interleave_low_optab
= optab_for_tree_code (VEC_INTERLEAVE_LOW_EXPR
,
3599 if (!interleave_high_optab
|| !interleave_low_optab
)
3601 if (vect_print_dump_info (REPORT_DETAILS
))
3602 fprintf (vect_dump
, "no optab for interleave.");
3606 if (interleave_high_optab
->handlers
[(int) mode
].insn_code
3608 || interleave_low_optab
->handlers
[(int) mode
].insn_code
3609 == CODE_FOR_nothing
)
3611 if (vect_print_dump_info (REPORT_DETAILS
))
3612 fprintf (vect_dump
, "interleave op not supported by target.");
3619 /* Function vect_permute_store_chain.
3621 Given a chain of interleaved stores in DR_CHAIN of LENGTH that must be
3622 a power of 2, generate interleave_high/low stmts to reorder the data
3623 correctly for the stores. Return the final references for stores in
3626 E.g., LENGTH is 4 and the scalar type is short, i.e., VF is 8.
3627 The input is 4 vectors each containing 8 elements. We assign a number to each
3628 element, the input sequence is:
3630 1st vec: 0 1 2 3 4 5 6 7
3631 2nd vec: 8 9 10 11 12 13 14 15
3632 3rd vec: 16 17 18 19 20 21 22 23
3633 4th vec: 24 25 26 27 28 29 30 31
3635 The output sequence should be:
3637 1st vec: 0 8 16 24 1 9 17 25
3638 2nd vec: 2 10 18 26 3 11 19 27
3639 3rd vec: 4 12 20 28 5 13 21 30
3640 4th vec: 6 14 22 30 7 15 23 31
3642 i.e., we interleave the contents of the four vectors in their order.
3644 We use interleave_high/low instructions to create such output. The input of
3645 each interleave_high/low operation is two vectors:
3648 the even elements of the result vector are obtained left-to-right from the
3649 high/low elements of the first vector. The odd elements of the result are
3650 obtained left-to-right from the high/low elements of the second vector.
3651 The output of interleave_high will be: 0 4 1 5
3652 and of interleave_low: 2 6 3 7
3655 The permutation is done in log LENGTH stages. In each stage interleave_high
3656 and interleave_low stmts are created for each pair of vectors in DR_CHAIN,
3657 where the first argument is taken from the first half of DR_CHAIN and the
3658 second argument from it's second half.
3661 I1: interleave_high (1st vec, 3rd vec)
3662 I2: interleave_low (1st vec, 3rd vec)
3663 I3: interleave_high (2nd vec, 4th vec)
3664 I4: interleave_low (2nd vec, 4th vec)
3666 The output for the first stage is:
3668 I1: 0 16 1 17 2 18 3 19
3669 I2: 4 20 5 21 6 22 7 23
3670 I3: 8 24 9 25 10 26 11 27
3671 I4: 12 28 13 29 14 30 15 31
3673 The output of the second stage, i.e. the final result is:
3675 I1: 0 8 16 24 1 9 17 25
3676 I2: 2 10 18 26 3 11 19 27
3677 I3: 4 12 20 28 5 13 21 30
3678 I4: 6 14 22 30 7 15 23 31. */
3681 vect_permute_store_chain (VEC(tree
,heap
) *dr_chain
,
3682 unsigned int length
,
3684 block_stmt_iterator
*bsi
,
3685 VEC(tree
,heap
) **result_chain
)
3687 tree perm_dest
, perm_stmt
, vect1
, vect2
, high
, low
;
3688 tree vectype
= STMT_VINFO_VECTYPE (vinfo_for_stmt (stmt
));
3689 tree scalar_dest
, tmp
;
3692 VEC(tree
,heap
) *first
, *second
;
3694 scalar_dest
= GIMPLE_STMT_OPERAND (stmt
, 0);
3695 first
= VEC_alloc (tree
, heap
, length
/2);
3696 second
= VEC_alloc (tree
, heap
, length
/2);
3698 /* Check that the operation is supported. */
3699 if (!vect_strided_store_supported (vectype
))
3702 *result_chain
= VEC_copy (tree
, heap
, dr_chain
);
3704 for (i
= 0; i
< exact_log2 (length
); i
++)
3706 for (j
= 0; j
< length
/2; j
++)
3708 vect1
= VEC_index (tree
, dr_chain
, j
);
3709 vect2
= VEC_index (tree
, dr_chain
, j
+length
/2);
3711 /* Create interleaving stmt:
3712 in the case of big endian:
3713 high = interleave_high (vect1, vect2)
3714 and in the case of little endian:
3715 high = interleave_low (vect1, vect2). */
3716 perm_dest
= create_tmp_var (vectype
, "vect_inter_high");
3717 DECL_GIMPLE_REG_P (perm_dest
) = 1;
3718 add_referenced_var (perm_dest
);
3719 if (BYTES_BIG_ENDIAN
)
3720 tmp
= build2 (VEC_INTERLEAVE_HIGH_EXPR
, vectype
, vect1
, vect2
);
3722 tmp
= build2 (VEC_INTERLEAVE_LOW_EXPR
, vectype
, vect1
, vect2
);
3723 perm_stmt
= build_gimple_modify_stmt (perm_dest
, tmp
);
3724 high
= make_ssa_name (perm_dest
, perm_stmt
);
3725 GIMPLE_STMT_OPERAND (perm_stmt
, 0) = high
;
3726 vect_finish_stmt_generation (stmt
, perm_stmt
, bsi
);
3727 VEC_replace (tree
, *result_chain
, 2*j
, high
);
3729 /* Create interleaving stmt:
3730 in the case of big endian:
3731 low = interleave_low (vect1, vect2)
3732 and in the case of little endian:
3733 low = interleave_high (vect1, vect2). */
3734 perm_dest
= create_tmp_var (vectype
, "vect_inter_low");
3735 DECL_GIMPLE_REG_P (perm_dest
) = 1;
3736 add_referenced_var (perm_dest
);
3737 if (BYTES_BIG_ENDIAN
)
3738 tmp
= build2 (VEC_INTERLEAVE_LOW_EXPR
, vectype
, vect1
, vect2
);
3740 tmp
= build2 (VEC_INTERLEAVE_HIGH_EXPR
, vectype
, vect1
, vect2
);
3741 perm_stmt
= build_gimple_modify_stmt (perm_dest
, tmp
);
3742 low
= make_ssa_name (perm_dest
, perm_stmt
);
3743 GIMPLE_STMT_OPERAND (perm_stmt
, 0) = low
;
3744 vect_finish_stmt_generation (stmt
, perm_stmt
, bsi
);
3745 VEC_replace (tree
, *result_chain
, 2*j
+1, low
);
3747 dr_chain
= VEC_copy (tree
, heap
, *result_chain
);
3753 /* Function vectorizable_store.
3755 Check if STMT defines a non scalar data-ref (array/pointer/structure) that
3757 If VEC_STMT is also passed, vectorize the STMT: create a vectorized
3758 stmt to replace it, put it in VEC_STMT, and insert it at BSI.
3759 Return FALSE if not a vectorizable STMT, TRUE otherwise. */
3762 vectorizable_store (tree stmt
, block_stmt_iterator
*bsi
, tree
*vec_stmt
)
3767 tree vec_oprnd
= NULL_TREE
;
3768 stmt_vec_info stmt_info
= vinfo_for_stmt (stmt
);
3769 struct data_reference
*dr
= STMT_VINFO_DATA_REF (stmt_info
), *first_dr
= NULL
;
3770 tree vectype
= STMT_VINFO_VECTYPE (stmt_info
);
3771 loop_vec_info loop_vinfo
= STMT_VINFO_LOOP_VINFO (stmt_info
);
3772 enum machine_mode vec_mode
;
3774 enum dr_alignment_support alignment_support_cheme
;
3776 def_operand_p def_p
;
3778 enum vect_def_type dt
;
3779 stmt_vec_info prev_stmt_info
= NULL
;
3780 tree dataref_ptr
= NULL_TREE
;
3781 int nunits
= TYPE_VECTOR_SUBPARTS (vectype
);
3782 int ncopies
= LOOP_VINFO_VECT_FACTOR (loop_vinfo
) / nunits
;
3784 tree next_stmt
, first_stmt
;
3785 bool strided_store
= false;
3786 unsigned int group_size
, i
;
3787 VEC(tree
,heap
) *dr_chain
= NULL
, *oprnds
= NULL
, *result_chain
= NULL
;
3788 gcc_assert (ncopies
>= 1);
3790 if (!STMT_VINFO_RELEVANT_P (stmt_info
))
3793 if (STMT_VINFO_DEF_TYPE (stmt_info
) != vect_loop_def
)
3796 if (STMT_VINFO_LIVE_P (stmt_info
))
3798 if (vect_print_dump_info (REPORT_DETAILS
))
3799 fprintf (vect_dump
, "value used after loop.");
3803 /* Is vectorizable store? */
3805 if (TREE_CODE (stmt
) != GIMPLE_MODIFY_STMT
)
3808 scalar_dest
= GIMPLE_STMT_OPERAND (stmt
, 0);
3809 if (TREE_CODE (scalar_dest
) != ARRAY_REF
3810 && TREE_CODE (scalar_dest
) != INDIRECT_REF
3811 && !DR_GROUP_FIRST_DR (stmt_info
))
3814 op
= GIMPLE_STMT_OPERAND (stmt
, 1);
3815 if (!vect_is_simple_use (op
, loop_vinfo
, &def_stmt
, &def
, &dt
))
3817 if (vect_print_dump_info (REPORT_DETAILS
))
3818 fprintf (vect_dump
, "use not simple.");
3822 vec_mode
= TYPE_MODE (vectype
);
3823 /* FORNOW. In some cases can vectorize even if data-type not supported
3824 (e.g. - array initialization with 0). */
3825 if (mov_optab
->handlers
[(int)vec_mode
].insn_code
== CODE_FOR_nothing
)
3828 if (!STMT_VINFO_DATA_REF (stmt_info
))
3831 if (DR_GROUP_FIRST_DR (stmt_info
))
3833 strided_store
= true;
3834 if (!vect_strided_store_supported (vectype
))
3838 if (!vec_stmt
) /* transformation not required. */
3840 STMT_VINFO_TYPE (stmt_info
) = store_vec_info_type
;
3841 vect_model_store_cost (stmt_info
, ncopies
);
3849 first_stmt
= DR_GROUP_FIRST_DR (stmt_info
);
3850 first_dr
= STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt
));
3851 group_size
= DR_GROUP_SIZE (vinfo_for_stmt (first_stmt
));
3853 DR_GROUP_STORE_COUNT (vinfo_for_stmt (first_stmt
))++;
3855 /* We vectorize all the stmts of the interleaving group when we
3856 reach the last stmt in the group. */
3857 if (DR_GROUP_STORE_COUNT (vinfo_for_stmt (first_stmt
))
3858 < DR_GROUP_SIZE (vinfo_for_stmt (first_stmt
)))
3860 *vec_stmt
= NULL_TREE
;
3871 if (vect_print_dump_info (REPORT_DETAILS
))
3872 fprintf (vect_dump
, "transform store. ncopies = %d",ncopies
);
3874 dr_chain
= VEC_alloc (tree
, heap
, group_size
);
3875 oprnds
= VEC_alloc (tree
, heap
, group_size
);
3877 alignment_support_cheme
= vect_supportable_dr_alignment (first_dr
);
3878 gcc_assert (alignment_support_cheme
);
3879 gcc_assert (alignment_support_cheme
== dr_aligned
); /* FORNOW */
3881 /* In case the vectorization factor (VF) is bigger than the number
3882 of elements that we can fit in a vectype (nunits), we have to generate
3883 more than one vector stmt - i.e - we need to "unroll" the
3884 vector stmt by a factor VF/nunits. For more details see documentation in
3885 vect_get_vec_def_for_copy_stmt. */
3887 /* In case of interleaving (non-unit strided access):
3894 We create vectorized stores starting from base address (the access of the
3895 first stmt in the chain (S2 in the above example), when the last store stmt
3896 of the chain (S4) is reached:
3899 VS2: &base + vec_size*1 = vx0
3900 VS3: &base + vec_size*2 = vx1
3901 VS4: &base + vec_size*3 = vx3
3903 Then permutation statements are generated:
3905 VS5: vx5 = VEC_INTERLEAVE_HIGH_EXPR < vx0, vx3 >
3906 VS6: vx6 = VEC_INTERLEAVE_LOW_EXPR < vx0, vx3 >
3909 And they are put in STMT_VINFO_VEC_STMT of the corresponding scalar stmts
3910 (the order of the data-refs in the output of vect_permute_store_chain
3911 corresponds to the order of scalar stmts in the interleaving chain - see
3912 the documentation of vect_permute_store_chain()).
3914 In case of both multiple types and interleaving, above vector stores and
3915 permutation stmts are created for every copy. The result vector stmts are
3916 put in STMT_VINFO_VEC_STMT for the first copy and in the corresponding
3917 STMT_VINFO_RELATED_STMT for the next copies.
3920 prev_stmt_info
= NULL
;
3921 for (j
= 0; j
< ncopies
; j
++)
3928 /* For interleaved stores we collect vectorized defs for all the
3929 stores in the group in DR_CHAIN and OPRNDS. DR_CHAIN is then used
3930 as an input to vect_permute_store_chain(), and OPRNDS as an input
3931 to vect_get_vec_def_for_stmt_copy() for the next copy.
3932 If the store is not strided, GROUP_SIZE is 1, and DR_CHAIN and
3933 OPRNDS are of size 1. */
3934 next_stmt
= first_stmt
;
3935 for (i
= 0; i
< group_size
; i
++)
3937 /* Since gaps are not supported for interleaved stores, GROUP_SIZE
3938 is the exact number of stmts in the chain. Therefore, NEXT_STMT
3939 can't be NULL_TREE. In case that there is no interleaving,
3940 GROUP_SIZE is 1, and only one iteration of the loop will be
3942 gcc_assert (next_stmt
);
3943 op
= GIMPLE_STMT_OPERAND (next_stmt
, 1);
3944 vec_oprnd
= vect_get_vec_def_for_operand (op
, next_stmt
, NULL
);
3945 VEC_quick_push(tree
, dr_chain
, vec_oprnd
);
3946 VEC_quick_push(tree
, oprnds
, vec_oprnd
);
3947 next_stmt
= DR_GROUP_NEXT_DR (vinfo_for_stmt (next_stmt
));
3949 dataref_ptr
= vect_create_data_ref_ptr (first_stmt
, bsi
, NULL_TREE
,
3950 &dummy
, &ptr_incr
, false,
3951 TREE_TYPE (vec_oprnd
));
3955 /* For interleaved stores we created vectorized defs for all the
3956 defs stored in OPRNDS in the previous iteration (previous copy).
3957 DR_CHAIN is then used as an input to vect_permute_store_chain(),
3958 and OPRNDS as an input to vect_get_vec_def_for_stmt_copy() for the
3960 If the store is not strided, GROUP_SIZE is 1, and DR_CHAIN and
3961 OPRNDS are of size 1. */
3962 for (i
= 0; i
< group_size
; i
++)
3964 vec_oprnd
= vect_get_vec_def_for_stmt_copy (dt
,
3965 VEC_index (tree
, oprnds
, i
));
3966 VEC_replace(tree
, dr_chain
, i
, vec_oprnd
);
3967 VEC_replace(tree
, oprnds
, i
, vec_oprnd
);
3969 dataref_ptr
= bump_vector_ptr (dataref_ptr
, ptr_incr
, bsi
, stmt
);
3974 result_chain
= VEC_alloc (tree
, heap
, group_size
);
3976 if (!vect_permute_store_chain (dr_chain
, group_size
, stmt
, bsi
,
3981 next_stmt
= first_stmt
;
3982 for (i
= 0; i
< group_size
; i
++)
3984 /* For strided stores vectorized defs are interleaved in
3985 vect_permute_store_chain(). */
3987 vec_oprnd
= VEC_index(tree
, result_chain
, i
);
3989 data_ref
= build_fold_indirect_ref (dataref_ptr
);
3990 /* Arguments are ready. Create the new vector stmt. */
3991 new_stmt
= build_gimple_modify_stmt (data_ref
, vec_oprnd
);
3992 vect_finish_stmt_generation (stmt
, new_stmt
, bsi
);
3994 /* Set the VDEFs for the vector pointer. If this virtual def
3995 has a use outside the loop and a loop peel is performed
3996 then the def may be renamed by the peel. Mark it for
3997 renaming so the later use will also be renamed. */
3998 copy_virtual_operands (new_stmt
, next_stmt
);
4001 /* The original store is deleted so the same SSA_NAMEs
4003 FOR_EACH_SSA_TREE_OPERAND (def
, next_stmt
, iter
, SSA_OP_VDEF
)
4005 SSA_NAME_DEF_STMT (def
) = new_stmt
;
4006 mark_sym_for_renaming (SSA_NAME_VAR (def
));
4009 STMT_VINFO_VEC_STMT (stmt_info
) = *vec_stmt
= new_stmt
;
4013 /* Create new names for all the definitions created by COPY and
4014 add replacement mappings for each new name. */
4015 FOR_EACH_SSA_DEF_OPERAND (def_p
, new_stmt
, iter
, SSA_OP_VDEF
)
4017 create_new_def_for (DEF_FROM_PTR (def_p
), new_stmt
, def_p
);
4018 mark_sym_for_renaming (SSA_NAME_VAR (DEF_FROM_PTR (def_p
)));
4021 STMT_VINFO_RELATED_STMT (prev_stmt_info
) = new_stmt
;
4024 prev_stmt_info
= vinfo_for_stmt (new_stmt
);
4025 next_stmt
= DR_GROUP_NEXT_DR (vinfo_for_stmt (next_stmt
));
4028 /* Bump the vector pointer. */
4029 dataref_ptr
= bump_vector_ptr (dataref_ptr
, ptr_incr
, bsi
, stmt
);
4037 /* Function vect_setup_realignment
4039 This function is called when vectorizing an unaligned load using
4040 the dr_unaligned_software_pipeline scheme.
4041 This function generates the following code at the loop prolog:
4044 msq_init = *(floor(p)); # prolog load
4045 realignment_token = call target_builtin;
4047 msq = phi (msq_init, ---)
4049 The code above sets up a new (vector) pointer, pointing to the first
4050 location accessed by STMT, and a "floor-aligned" load using that pointer.
4051 It also generates code to compute the "realignment-token" (if the relevant
4052 target hook was defined), and creates a phi-node at the loop-header bb
4053 whose arguments are the result of the prolog-load (created by this
4054 function) and the result of a load that takes place in the loop (to be
4055 created by the caller to this function).
4056 The caller to this function uses the phi-result (msq) to create the
4057 realignment code inside the loop, and sets up the missing phi argument,
4061 msq = phi (msq_init, lsq)
4062 lsq = *(floor(p')); # load in loop
4063 result = realign_load (msq, lsq, realignment_token);
4066 STMT - (scalar) load stmt to be vectorized. This load accesses
4067 a memory location that may be unaligned.
4068 BSI - place where new code is to be inserted.
4071 REALIGNMENT_TOKEN - the result of a call to the builtin_mask_for_load
4072 target hook, if defined.
4073 Return value - the result of the loop-header phi node. */
4076 vect_setup_realignment (tree stmt
, block_stmt_iterator
*bsi
,
4077 tree
*realignment_token
)
4079 stmt_vec_info stmt_info
= vinfo_for_stmt (stmt
);
4080 tree vectype
= STMT_VINFO_VECTYPE (stmt_info
);
4081 loop_vec_info loop_vinfo
= STMT_VINFO_LOOP_VINFO (stmt_info
);
4082 struct loop
*loop
= LOOP_VINFO_LOOP (loop_vinfo
);
4083 edge pe
= loop_preheader_edge (loop
);
4084 tree scalar_dest
= GIMPLE_STMT_OPERAND (stmt
, 0);
4097 /* 1. Create msq_init = *(floor(p1)) in the loop preheader */
4098 vec_dest
= vect_create_destination_var (scalar_dest
, vectype
);
4099 ptr
= vect_create_data_ref_ptr (stmt
, bsi
, NULL_TREE
, &init_addr
, &inc
, true,
4101 data_ref
= build1 (ALIGN_INDIRECT_REF
, vectype
, ptr
);
4102 new_stmt
= build_gimple_modify_stmt (vec_dest
, data_ref
);
4103 new_temp
= make_ssa_name (vec_dest
, new_stmt
);
4104 GIMPLE_STMT_OPERAND (new_stmt
, 0) = new_temp
;
4105 new_bb
= bsi_insert_on_edge_immediate (pe
, new_stmt
);
4106 gcc_assert (!new_bb
);
4107 msq_init
= GIMPLE_STMT_OPERAND (new_stmt
, 0);
4108 copy_virtual_operands (new_stmt
, stmt
);
4109 update_vuses_to_preheader (new_stmt
, loop
);
4111 /* 2. Create permutation mask, if required, in loop preheader. */
4112 if (targetm
.vectorize
.builtin_mask_for_load
)
4116 builtin_decl
= targetm
.vectorize
.builtin_mask_for_load ();
4117 new_stmt
= build_call_expr (builtin_decl
, 1, init_addr
);
4118 vec_dest
= vect_create_destination_var (scalar_dest
,
4119 TREE_TYPE (new_stmt
));
4120 new_stmt
= build_gimple_modify_stmt (vec_dest
, new_stmt
);
4121 new_temp
= make_ssa_name (vec_dest
, new_stmt
);
4122 GIMPLE_STMT_OPERAND (new_stmt
, 0) = new_temp
;
4123 new_bb
= bsi_insert_on_edge_immediate (pe
, new_stmt
);
4124 gcc_assert (!new_bb
);
4125 *realignment_token
= GIMPLE_STMT_OPERAND (new_stmt
, 0);
4127 /* The result of the CALL_EXPR to this builtin is determined from
4128 the value of the parameter and no global variables are touched
4129 which makes the builtin a "const" function. Requiring the
4130 builtin to have the "const" attribute makes it unnecessary
4131 to call mark_call_clobbered. */
4132 gcc_assert (TREE_READONLY (builtin_decl
));
4135 /* 3. Create msq = phi <msq_init, lsq> in loop */
4136 vec_dest
= vect_create_destination_var (scalar_dest
, vectype
);
4137 msq
= make_ssa_name (vec_dest
, NULL_TREE
);
4138 phi_stmt
= create_phi_node (msq
, loop
->header
);
4139 SSA_NAME_DEF_STMT (msq
) = phi_stmt
;
4140 add_phi_arg (phi_stmt
, msq_init
, loop_preheader_edge (loop
));
4146 /* Function vect_strided_load_supported.
4148 Returns TRUE is EXTRACT_EVEN and EXTRACT_ODD operations are supported,
4149 and FALSE otherwise. */
4152 vect_strided_load_supported (tree vectype
)
4154 optab perm_even_optab
, perm_odd_optab
;
4157 mode
= (int) TYPE_MODE (vectype
);
4159 perm_even_optab
= optab_for_tree_code (VEC_EXTRACT_EVEN_EXPR
, vectype
);
4160 if (!perm_even_optab
)
4162 if (vect_print_dump_info (REPORT_DETAILS
))
4163 fprintf (vect_dump
, "no optab for perm_even.");
4167 if (perm_even_optab
->handlers
[mode
].insn_code
== CODE_FOR_nothing
)
4169 if (vect_print_dump_info (REPORT_DETAILS
))
4170 fprintf (vect_dump
, "perm_even op not supported by target.");
4174 perm_odd_optab
= optab_for_tree_code (VEC_EXTRACT_ODD_EXPR
, vectype
);
4175 if (!perm_odd_optab
)
4177 if (vect_print_dump_info (REPORT_DETAILS
))
4178 fprintf (vect_dump
, "no optab for perm_odd.");
4182 if (perm_odd_optab
->handlers
[mode
].insn_code
== CODE_FOR_nothing
)
4184 if (vect_print_dump_info (REPORT_DETAILS
))
4185 fprintf (vect_dump
, "perm_odd op not supported by target.");
4192 /* Function vect_permute_load_chain.
4194 Given a chain of interleaved loads in DR_CHAIN of LENGTH that must be
4195 a power of 2, generate extract_even/odd stmts to reorder the input data
4196 correctly. Return the final references for loads in RESULT_CHAIN.
4198 E.g., LENGTH is 4 and the scalar type is short, i.e., VF is 8.
4199 The input is 4 vectors each containing 8 elements. We assign a number to each
4200 element, the input sequence is:
4202 1st vec: 0 1 2 3 4 5 6 7
4203 2nd vec: 8 9 10 11 12 13 14 15
4204 3rd vec: 16 17 18 19 20 21 22 23
4205 4th vec: 24 25 26 27 28 29 30 31
4207 The output sequence should be:
4209 1st vec: 0 4 8 12 16 20 24 28
4210 2nd vec: 1 5 9 13 17 21 25 29
4211 3rd vec: 2 6 10 14 18 22 26 30
4212 4th vec: 3 7 11 15 19 23 27 31
4214 i.e., the first output vector should contain the first elements of each
4215 interleaving group, etc.
4217 We use extract_even/odd instructions to create such output. The input of each
4218 extract_even/odd operation is two vectors
4222 and the output is the vector of extracted even/odd elements. The output of
4223 extract_even will be: 0 2 4 6
4224 and of extract_odd: 1 3 5 7
4227 The permutation is done in log LENGTH stages. In each stage extract_even and
4228 extract_odd stmts are created for each pair of vectors in DR_CHAIN in their
4229 order. In our example,
4231 E1: extract_even (1st vec, 2nd vec)
4232 E2: extract_odd (1st vec, 2nd vec)
4233 E3: extract_even (3rd vec, 4th vec)
4234 E4: extract_odd (3rd vec, 4th vec)
4236 The output for the first stage will be:
4238 E1: 0 2 4 6 8 10 12 14
4239 E2: 1 3 5 7 9 11 13 15
4240 E3: 16 18 20 22 24 26 28 30
4241 E4: 17 19 21 23 25 27 29 31
4243 In order to proceed and create the correct sequence for the next stage (or
4244 for the correct output, if the second stage is the last one, as in our
4245 example), we first put the output of extract_even operation and then the
4246 output of extract_odd in RESULT_CHAIN (which is then copied to DR_CHAIN).
4247 The input for the second stage is:
4249 1st vec (E1): 0 2 4 6 8 10 12 14
4250 2nd vec (E3): 16 18 20 22 24 26 28 30
4251 3rd vec (E2): 1 3 5 7 9 11 13 15
4252 4th vec (E4): 17 19 21 23 25 27 29 31
4254 The output of the second stage:
4256 E1: 0 4 8 12 16 20 24 28
4257 E2: 2 6 10 14 18 22 26 30
4258 E3: 1 5 9 13 17 21 25 29
4259 E4: 3 7 11 15 19 23 27 31
4261 And RESULT_CHAIN after reordering:
4263 1st vec (E1): 0 4 8 12 16 20 24 28
4264 2nd vec (E3): 1 5 9 13 17 21 25 29
4265 3rd vec (E2): 2 6 10 14 18 22 26 30
4266 4th vec (E4): 3 7 11 15 19 23 27 31. */
4269 vect_permute_load_chain (VEC(tree
,heap
) *dr_chain
,
4270 unsigned int length
,
4272 block_stmt_iterator
*bsi
,
4273 VEC(tree
,heap
) **result_chain
)
4275 tree perm_dest
, perm_stmt
, data_ref
, first_vect
, second_vect
;
4276 tree vectype
= STMT_VINFO_VECTYPE (vinfo_for_stmt (stmt
));
4281 /* Check that the operation is supported. */
4282 if (!vect_strided_load_supported (vectype
))
4285 *result_chain
= VEC_copy (tree
, heap
, dr_chain
);
4286 for (i
= 0; i
< exact_log2 (length
); i
++)
4288 for (j
= 0; j
< length
; j
+=2)
4290 first_vect
= VEC_index (tree
, dr_chain
, j
);
4291 second_vect
= VEC_index (tree
, dr_chain
, j
+1);
4293 /* data_ref = permute_even (first_data_ref, second_data_ref); */
4294 perm_dest
= create_tmp_var (vectype
, "vect_perm_even");
4295 DECL_GIMPLE_REG_P (perm_dest
) = 1;
4296 add_referenced_var (perm_dest
);
4298 tmp
= build2 (VEC_EXTRACT_EVEN_EXPR
, vectype
,
4299 first_vect
, second_vect
);
4300 perm_stmt
= build_gimple_modify_stmt (perm_dest
, tmp
);
4302 data_ref
= make_ssa_name (perm_dest
, perm_stmt
);
4303 GIMPLE_STMT_OPERAND (perm_stmt
, 0) = data_ref
;
4304 vect_finish_stmt_generation (stmt
, perm_stmt
, bsi
);
4305 mark_symbols_for_renaming (perm_stmt
);
4307 VEC_replace (tree
, *result_chain
, j
/2, data_ref
);
4309 /* data_ref = permute_odd (first_data_ref, second_data_ref); */
4310 perm_dest
= create_tmp_var (vectype
, "vect_perm_odd");
4311 DECL_GIMPLE_REG_P (perm_dest
) = 1;
4312 add_referenced_var (perm_dest
);
4314 tmp
= build2 (VEC_EXTRACT_ODD_EXPR
, vectype
,
4315 first_vect
, second_vect
);
4316 perm_stmt
= build_gimple_modify_stmt (perm_dest
, tmp
);
4317 data_ref
= make_ssa_name (perm_dest
, perm_stmt
);
4318 GIMPLE_STMT_OPERAND (perm_stmt
, 0) = data_ref
;
4319 vect_finish_stmt_generation (stmt
, perm_stmt
, bsi
);
4320 mark_symbols_for_renaming (perm_stmt
);
4322 VEC_replace (tree
, *result_chain
, j
/2+length
/2, data_ref
);
4324 dr_chain
= VEC_copy (tree
, heap
, *result_chain
);
4330 /* Function vect_transform_strided_load.
4332 Given a chain of input interleaved data-refs (in DR_CHAIN), build statements
4333 to perform their permutation and ascribe the result vectorized statements to
4334 the scalar statements.
4338 vect_transform_strided_load (tree stmt
, VEC(tree
,heap
) *dr_chain
, int size
,
4339 block_stmt_iterator
*bsi
)
4341 stmt_vec_info stmt_info
= vinfo_for_stmt (stmt
);
4342 tree first_stmt
= DR_GROUP_FIRST_DR (stmt_info
);
4343 tree next_stmt
, new_stmt
;
4344 VEC(tree
,heap
) *result_chain
= NULL
;
4345 unsigned int i
, gap_count
;
4348 /* DR_CHAIN contains input data-refs that are a part of the interleaving.
4349 RESULT_CHAIN is the output of vect_permute_load_chain, it contains permuted
4350 vectors, that are ready for vector computation. */
4351 result_chain
= VEC_alloc (tree
, heap
, size
);
4353 if (!vect_permute_load_chain (dr_chain
, size
, stmt
, bsi
, &result_chain
))
4356 /* Put a permuted data-ref in the VECTORIZED_STMT field.
4357 Since we scan the chain starting from it's first node, their order
4358 corresponds the order of data-refs in RESULT_CHAIN. */
4359 next_stmt
= first_stmt
;
4361 for (i
= 0; VEC_iterate(tree
, result_chain
, i
, tmp_data_ref
); i
++)
4366 /* Skip the gaps. Loads created for the gaps will be removed by dead
4367 code elimination pass later.
4368 DR_GROUP_GAP is the number of steps in elements from the previous
4369 access (if there is no gap DR_GROUP_GAP is 1). We skip loads that
4370 correspond to the gaps.
4372 if (gap_count
< DR_GROUP_GAP (vinfo_for_stmt (next_stmt
)))
4380 new_stmt
= SSA_NAME_DEF_STMT (tmp_data_ref
);
4381 /* We assume that if VEC_STMT is not NULL, this is a case of multiple
4382 copies, and we put the new vector statement in the first available
4384 if (!STMT_VINFO_VEC_STMT (vinfo_for_stmt (next_stmt
)))
4385 STMT_VINFO_VEC_STMT (vinfo_for_stmt (next_stmt
)) = new_stmt
;
4388 tree prev_stmt
= STMT_VINFO_VEC_STMT (vinfo_for_stmt (next_stmt
));
4389 tree rel_stmt
= STMT_VINFO_RELATED_STMT (
4390 vinfo_for_stmt (prev_stmt
));
4393 prev_stmt
= rel_stmt
;
4394 rel_stmt
= STMT_VINFO_RELATED_STMT (vinfo_for_stmt (rel_stmt
));
4396 STMT_VINFO_RELATED_STMT (vinfo_for_stmt (prev_stmt
)) = new_stmt
;
4398 next_stmt
= DR_GROUP_NEXT_DR (vinfo_for_stmt (next_stmt
));
4400 /* If NEXT_STMT accesses the same DR as the previous statement,
4401 put the same TMP_DATA_REF as its vectorized statement; otherwise
4402 get the next data-ref from RESULT_CHAIN. */
4403 if (!next_stmt
|| !DR_GROUP_SAME_DR_STMT (vinfo_for_stmt (next_stmt
)))
4411 /* vectorizable_load.
4413 Check if STMT reads a non scalar data-ref (array/pointer/structure) that
4415 If VEC_STMT is also passed, vectorize the STMT: create a vectorized
4416 stmt to replace it, put it in VEC_STMT, and insert it at BSI.
4417 Return FALSE if not a vectorizable STMT, TRUE otherwise. */
4420 vectorizable_load (tree stmt
, block_stmt_iterator
*bsi
, tree
*vec_stmt
)
4423 tree vec_dest
= NULL
;
4424 tree data_ref
= NULL
;
4426 stmt_vec_info stmt_info
= vinfo_for_stmt (stmt
);
4427 stmt_vec_info prev_stmt_info
;
4428 loop_vec_info loop_vinfo
= STMT_VINFO_LOOP_VINFO (stmt_info
);
4429 struct loop
*loop
= LOOP_VINFO_LOOP (loop_vinfo
);
4430 struct data_reference
*dr
= STMT_VINFO_DATA_REF (stmt_info
), *first_dr
;
4431 tree vectype
= STMT_VINFO_VECTYPE (stmt_info
);
4434 tree new_stmt
= NULL_TREE
;
4436 enum dr_alignment_support alignment_support_cheme
;
4437 tree dataref_ptr
= NULL_TREE
;
4439 int nunits
= TYPE_VECTOR_SUBPARTS (vectype
);
4440 int ncopies
= LOOP_VINFO_VECT_FACTOR (loop_vinfo
) / nunits
;
4441 int i
, j
, group_size
;
4442 tree msq
= NULL_TREE
, lsq
;
4443 tree offset
= NULL_TREE
;
4444 tree realignment_token
= NULL_TREE
;
4445 tree phi_stmt
= NULL_TREE
;
4446 VEC(tree
,heap
) *dr_chain
= NULL
;
4447 bool strided_load
= false;
4450 if (!STMT_VINFO_RELEVANT_P (stmt_info
))
4453 if (STMT_VINFO_DEF_TYPE (stmt_info
) != vect_loop_def
)
4456 /* FORNOW: not yet supported. */
4457 if (STMT_VINFO_LIVE_P (stmt_info
))
4459 if (vect_print_dump_info (REPORT_DETAILS
))
4460 fprintf (vect_dump
, "value used after loop.");
4464 /* Is vectorizable load? */
4465 if (TREE_CODE (stmt
) != GIMPLE_MODIFY_STMT
)
4468 scalar_dest
= GIMPLE_STMT_OPERAND (stmt
, 0);
4469 if (TREE_CODE (scalar_dest
) != SSA_NAME
)
4472 op
= GIMPLE_STMT_OPERAND (stmt
, 1);
4473 if (TREE_CODE (op
) != ARRAY_REF
4474 && TREE_CODE (op
) != INDIRECT_REF
4475 && !DR_GROUP_FIRST_DR (stmt_info
))
4478 if (!STMT_VINFO_DATA_REF (stmt_info
))
4481 mode
= (int) TYPE_MODE (vectype
);
4483 /* FORNOW. In some cases can vectorize even if data-type not supported
4484 (e.g. - data copies). */
4485 if (mov_optab
->handlers
[mode
].insn_code
== CODE_FOR_nothing
)
4487 if (vect_print_dump_info (REPORT_DETAILS
))
4488 fprintf (vect_dump
, "Aligned load, but unsupported type.");
4492 /* Check if the load is a part of an interleaving chain. */
4493 if (DR_GROUP_FIRST_DR (stmt_info
))
4495 strided_load
= true;
4497 /* Check if interleaving is supported. */
4498 if (!vect_strided_load_supported (vectype
))
4502 if (!vec_stmt
) /* transformation not required. */
4504 STMT_VINFO_TYPE (stmt_info
) = load_vec_info_type
;
4505 vect_model_load_cost (stmt_info
, ncopies
);
4509 if (vect_print_dump_info (REPORT_DETAILS
))
4510 fprintf (vect_dump
, "transform load.");
4516 first_stmt
= DR_GROUP_FIRST_DR (stmt_info
);
4517 /* Check if the chain of loads is already vectorized. */
4518 if (STMT_VINFO_VEC_STMT (vinfo_for_stmt (first_stmt
)))
4520 *vec_stmt
= STMT_VINFO_VEC_STMT (stmt_info
);
4523 first_dr
= STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt
));
4524 group_size
= DR_GROUP_SIZE (vinfo_for_stmt (first_stmt
));
4525 dr_chain
= VEC_alloc (tree
, heap
, group_size
);
4534 alignment_support_cheme
= vect_supportable_dr_alignment (first_dr
);
4535 gcc_assert (alignment_support_cheme
);
4538 /* In case the vectorization factor (VF) is bigger than the number
4539 of elements that we can fit in a vectype (nunits), we have to generate
4540 more than one vector stmt - i.e - we need to "unroll" the
4541 vector stmt by a factor VF/nunits. In doing so, we record a pointer
4542 from one copy of the vector stmt to the next, in the field
4543 STMT_VINFO_RELATED_STMT. This is necessary in order to allow following
4544 stages to find the correct vector defs to be used when vectorizing
4545 stmts that use the defs of the current stmt. The example below illustrates
4546 the vectorization process when VF=16 and nunits=4 (i.e - we need to create
4547 4 vectorized stmts):
4549 before vectorization:
4550 RELATED_STMT VEC_STMT
4554 step 1: vectorize stmt S1:
4555 We first create the vector stmt VS1_0, and, as usual, record a
4556 pointer to it in the STMT_VINFO_VEC_STMT of the scalar stmt S1.
4557 Next, we create the vector stmt VS1_1, and record a pointer to
4558 it in the STMT_VINFO_RELATED_STMT of the vector stmt VS1_0.
4559 Similarly, for VS1_2 and VS1_3. This is the resulting chain of
4561 RELATED_STMT VEC_STMT
4562 VS1_0: vx0 = memref0 VS1_1 -
4563 VS1_1: vx1 = memref1 VS1_2 -
4564 VS1_2: vx2 = memref2 VS1_3 -
4565 VS1_3: vx3 = memref3 - -
4566 S1: x = load - VS1_0
4569 See in documentation in vect_get_vec_def_for_stmt_copy for how the
4570 information we recorded in RELATED_STMT field is used to vectorize
4573 /* In case of interleaving (non-unit strided access):
4580 Vectorized loads are created in the order of memory accesses
4581 starting from the access of the first stmt of the chain:
4584 VS2: vx1 = &base + vec_size*1
4585 VS3: vx3 = &base + vec_size*2
4586 VS4: vx4 = &base + vec_size*3
4588 Then permutation statements are generated:
4590 VS5: vx5 = VEC_EXTRACT_EVEN_EXPR < vx0, vx1 >
4591 VS6: vx6 = VEC_EXTRACT_ODD_EXPR < vx0, vx1 >
4594 And they are put in STMT_VINFO_VEC_STMT of the corresponding scalar stmts
4595 (the order of the data-refs in the output of vect_permute_load_chain
4596 corresponds to the order of scalar stmts in the interleaving chain - see
4597 the documentation of vect_permute_load_chain()).
4598 The generation of permutation stmts and recording them in
4599 STMT_VINFO_VEC_STMT is done in vect_transform_strided_load().
4601 In case of both multiple types and interleaving, the vector loads and
4602 permutation stmts above are created for every copy. The result vector stmts
4603 are put in STMT_VINFO_VEC_STMT for the first copy and in the corresponding
4604 STMT_VINFO_RELATED_STMT for the next copies. */
4606 /* If the data reference is aligned (dr_aligned) or potentially unaligned
4607 on a target that supports unaligned accesses (dr_unaligned_supported)
4608 we generate the following code:
4612 p = p + indx * vectype_size;
4617 Otherwise, the data reference is potentially unaligned on a target that
4618 does not support unaligned accesses (dr_unaligned_software_pipeline) -
4619 then generate the following code, in which the data in each iteration is
4620 obtained by two vector loads, one from the previous iteration, and one
4621 from the current iteration:
4623 msq_init = *(floor(p1))
4624 p2 = initial_addr + VS - 1;
4625 realignment_token = call target_builtin;
4628 p2 = p2 + indx * vectype_size
4630 vec_dest = realign_load (msq, lsq, realignment_token)
4635 if (alignment_support_cheme
== dr_unaligned_software_pipeline
)
4637 msq
= vect_setup_realignment (first_stmt
, bsi
, &realignment_token
);
4638 phi_stmt
= SSA_NAME_DEF_STMT (msq
);
4639 offset
= size_int (TYPE_VECTOR_SUBPARTS (vectype
) - 1);
4642 prev_stmt_info
= NULL
;
4643 for (j
= 0; j
< ncopies
; j
++)
4645 /* 1. Create the vector pointer update chain. */
4647 dataref_ptr
= vect_create_data_ref_ptr (first_stmt
, bsi
, offset
, &dummy
,
4648 &ptr_incr
, false, NULL_TREE
);
4650 dataref_ptr
= bump_vector_ptr (dataref_ptr
, ptr_incr
, bsi
, stmt
);
4652 for (i
= 0; i
< group_size
; i
++)
4654 /* 2. Create the vector-load in the loop. */
4655 switch (alignment_support_cheme
)
4658 gcc_assert (aligned_access_p (first_dr
));
4659 data_ref
= build_fold_indirect_ref (dataref_ptr
);
4661 case dr_unaligned_supported
:
4663 int mis
= DR_MISALIGNMENT (first_dr
);
4664 tree tmis
= (mis
== -1 ? size_zero_node
: size_int (mis
));
4666 gcc_assert (!aligned_access_p (first_dr
));
4667 tmis
= size_binop (MULT_EXPR
, tmis
, size_int(BITS_PER_UNIT
));
4669 build2 (MISALIGNED_INDIRECT_REF
, vectype
, dataref_ptr
, tmis
);
4672 case dr_unaligned_software_pipeline
:
4673 gcc_assert (!aligned_access_p (first_dr
));
4674 data_ref
= build1 (ALIGN_INDIRECT_REF
, vectype
, dataref_ptr
);
4679 vec_dest
= vect_create_destination_var (scalar_dest
, vectype
);
4680 new_stmt
= build_gimple_modify_stmt (vec_dest
, data_ref
);
4681 new_temp
= make_ssa_name (vec_dest
, new_stmt
);
4682 GIMPLE_STMT_OPERAND (new_stmt
, 0) = new_temp
;
4683 vect_finish_stmt_generation (stmt
, new_stmt
, bsi
);
4684 copy_virtual_operands (new_stmt
, stmt
);
4685 mark_symbols_for_renaming (new_stmt
);
4687 /* 3. Handle explicit realignment if necessary/supported. */
4688 if (alignment_support_cheme
== dr_unaligned_software_pipeline
)
4691 <vec_dest = realign_load (msq, lsq, realignment_token)> */
4692 lsq
= GIMPLE_STMT_OPERAND (new_stmt
, 0);
4693 if (!realignment_token
)
4694 realignment_token
= dataref_ptr
;
4695 vec_dest
= vect_create_destination_var (scalar_dest
, vectype
);
4697 build3 (REALIGN_LOAD_EXPR
, vectype
, msq
, lsq
, realignment_token
);
4698 new_stmt
= build_gimple_modify_stmt (vec_dest
, new_stmt
);
4699 new_temp
= make_ssa_name (vec_dest
, new_stmt
);
4700 GIMPLE_STMT_OPERAND (new_stmt
, 0) = new_temp
;
4701 vect_finish_stmt_generation (stmt
, new_stmt
, bsi
);
4702 if (i
== group_size
- 1 && j
== ncopies
- 1)
4703 add_phi_arg (phi_stmt
, lsq
, loop_latch_edge (loop
));
4707 VEC_quick_push (tree
, dr_chain
, new_temp
);
4708 if (i
< group_size
- 1)
4709 dataref_ptr
= bump_vector_ptr (dataref_ptr
, ptr_incr
, bsi
, stmt
);
4714 if (!vect_transform_strided_load (stmt
, dr_chain
, group_size
, bsi
))
4716 *vec_stmt
= STMT_VINFO_VEC_STMT (stmt_info
);
4717 dr_chain
= VEC_alloc (tree
, heap
, group_size
);
4722 STMT_VINFO_VEC_STMT (stmt_info
) = *vec_stmt
= new_stmt
;
4724 STMT_VINFO_RELATED_STMT (prev_stmt_info
) = new_stmt
;
4725 prev_stmt_info
= vinfo_for_stmt (new_stmt
);
4733 /* Function vectorizable_live_operation.
4735 STMT computes a value that is used outside the loop. Check if
4736 it can be supported. */
4739 vectorizable_live_operation (tree stmt
,
4740 block_stmt_iterator
*bsi ATTRIBUTE_UNUSED
,
4741 tree
*vec_stmt ATTRIBUTE_UNUSED
)
4744 stmt_vec_info stmt_info
= vinfo_for_stmt (stmt
);
4745 loop_vec_info loop_vinfo
= STMT_VINFO_LOOP_VINFO (stmt_info
);
4750 enum vect_def_type dt
;
4752 gcc_assert (STMT_VINFO_LIVE_P (stmt_info
));
4754 if (STMT_VINFO_DEF_TYPE (stmt_info
) == vect_reduction_def
)
4757 if (TREE_CODE (stmt
) != GIMPLE_MODIFY_STMT
)
4760 if (TREE_CODE (GIMPLE_STMT_OPERAND (stmt
, 0)) != SSA_NAME
)
4763 operation
= GIMPLE_STMT_OPERAND (stmt
, 1);
4764 op_type
= TREE_OPERAND_LENGTH (operation
);
4766 /* FORNOW: support only if all uses are invariant. This means
4767 that the scalar operations can remain in place, unvectorized.
4768 The original last scalar value that they compute will be used. */
4770 for (i
= 0; i
< op_type
; i
++)
4772 op
= TREE_OPERAND (operation
, i
);
4773 if (!vect_is_simple_use (op
, loop_vinfo
, &def_stmt
, &def
, &dt
))
4775 if (vect_print_dump_info (REPORT_DETAILS
))
4776 fprintf (vect_dump
, "use not simple.");
4780 if (dt
!= vect_invariant_def
&& dt
!= vect_constant_def
)
4784 /* No transformation is required for the cases we currently support. */
4789 /* Function vect_is_simple_cond.
4792 LOOP - the loop that is being vectorized.
4793 COND - Condition that is checked for simple use.
4795 Returns whether a COND can be vectorized. Checks whether
4796 condition operands are supportable using vec_is_simple_use. */
4799 vect_is_simple_cond (tree cond
, loop_vec_info loop_vinfo
)
4803 enum vect_def_type dt
;
4805 if (!COMPARISON_CLASS_P (cond
))
4808 lhs
= TREE_OPERAND (cond
, 0);
4809 rhs
= TREE_OPERAND (cond
, 1);
4811 if (TREE_CODE (lhs
) == SSA_NAME
)
4813 tree lhs_def_stmt
= SSA_NAME_DEF_STMT (lhs
);
4814 if (!vect_is_simple_use (lhs
, loop_vinfo
, &lhs_def_stmt
, &def
, &dt
))
4817 else if (TREE_CODE (lhs
) != INTEGER_CST
&& TREE_CODE (lhs
) != REAL_CST
)
4820 if (TREE_CODE (rhs
) == SSA_NAME
)
4822 tree rhs_def_stmt
= SSA_NAME_DEF_STMT (rhs
);
4823 if (!vect_is_simple_use (rhs
, loop_vinfo
, &rhs_def_stmt
, &def
, &dt
))
4826 else if (TREE_CODE (rhs
) != INTEGER_CST
&& TREE_CODE (rhs
) != REAL_CST
)
4832 /* vectorizable_condition.
4834 Check if STMT is conditional modify expression that can be vectorized.
4835 If VEC_STMT is also passed, vectorize the STMT: create a vectorized
4836 stmt using VEC_COND_EXPR to replace it, put it in VEC_STMT, and insert it
4839 Return FALSE if not a vectorizable STMT, TRUE otherwise. */
4842 vectorizable_condition (tree stmt
, block_stmt_iterator
*bsi
, tree
*vec_stmt
)
4844 tree scalar_dest
= NULL_TREE
;
4845 tree vec_dest
= NULL_TREE
;
4846 tree op
= NULL_TREE
;
4847 tree cond_expr
, then_clause
, else_clause
;
4848 stmt_vec_info stmt_info
= vinfo_for_stmt (stmt
);
4849 tree vectype
= STMT_VINFO_VECTYPE (stmt_info
);
4850 tree vec_cond_lhs
, vec_cond_rhs
, vec_then_clause
, vec_else_clause
;
4851 tree vec_compare
, vec_cond_expr
;
4853 loop_vec_info loop_vinfo
= STMT_VINFO_LOOP_VINFO (stmt_info
);
4854 enum machine_mode vec_mode
;
4856 enum vect_def_type dt
;
4857 int nunits
= TYPE_VECTOR_SUBPARTS (vectype
);
4858 int ncopies
= LOOP_VINFO_VECT_FACTOR (loop_vinfo
) / nunits
;
4860 gcc_assert (ncopies
>= 1);
4862 return false; /* FORNOW */
4864 if (!STMT_VINFO_RELEVANT_P (stmt_info
))
4867 if (STMT_VINFO_DEF_TYPE (stmt_info
) != vect_loop_def
)
4870 /* FORNOW: not yet supported. */
4871 if (STMT_VINFO_LIVE_P (stmt_info
))
4873 if (vect_print_dump_info (REPORT_DETAILS
))
4874 fprintf (vect_dump
, "value used after loop.");
4878 /* Is vectorizable conditional operation? */
4879 if (TREE_CODE (stmt
) != GIMPLE_MODIFY_STMT
)
4882 op
= GIMPLE_STMT_OPERAND (stmt
, 1);
4884 if (TREE_CODE (op
) != COND_EXPR
)
4887 cond_expr
= TREE_OPERAND (op
, 0);
4888 then_clause
= TREE_OPERAND (op
, 1);
4889 else_clause
= TREE_OPERAND (op
, 2);
4891 if (!vect_is_simple_cond (cond_expr
, loop_vinfo
))
4894 /* We do not handle two different vector types for the condition
4896 if (TREE_TYPE (TREE_OPERAND (cond_expr
, 0)) != TREE_TYPE (vectype
))
4899 if (TREE_CODE (then_clause
) == SSA_NAME
)
4901 tree then_def_stmt
= SSA_NAME_DEF_STMT (then_clause
);
4902 if (!vect_is_simple_use (then_clause
, loop_vinfo
,
4903 &then_def_stmt
, &def
, &dt
))
4906 else if (TREE_CODE (then_clause
) != INTEGER_CST
4907 && TREE_CODE (then_clause
) != REAL_CST
)
4910 if (TREE_CODE (else_clause
) == SSA_NAME
)
4912 tree else_def_stmt
= SSA_NAME_DEF_STMT (else_clause
);
4913 if (!vect_is_simple_use (else_clause
, loop_vinfo
,
4914 &else_def_stmt
, &def
, &dt
))
4917 else if (TREE_CODE (else_clause
) != INTEGER_CST
4918 && TREE_CODE (else_clause
) != REAL_CST
)
4922 vec_mode
= TYPE_MODE (vectype
);
4926 STMT_VINFO_TYPE (stmt_info
) = condition_vec_info_type
;
4927 return expand_vec_cond_expr_p (op
, vec_mode
);
4933 scalar_dest
= GIMPLE_STMT_OPERAND (stmt
, 0);
4934 vec_dest
= vect_create_destination_var (scalar_dest
, vectype
);
4936 /* Handle cond expr. */
4938 vect_get_vec_def_for_operand (TREE_OPERAND (cond_expr
, 0), stmt
, NULL
);
4940 vect_get_vec_def_for_operand (TREE_OPERAND (cond_expr
, 1), stmt
, NULL
);
4941 vec_then_clause
= vect_get_vec_def_for_operand (then_clause
, stmt
, NULL
);
4942 vec_else_clause
= vect_get_vec_def_for_operand (else_clause
, stmt
, NULL
);
4944 /* Arguments are ready. create the new vector stmt. */
4945 vec_compare
= build2 (TREE_CODE (cond_expr
), vectype
,
4946 vec_cond_lhs
, vec_cond_rhs
);
4947 vec_cond_expr
= build3 (VEC_COND_EXPR
, vectype
,
4948 vec_compare
, vec_then_clause
, vec_else_clause
);
4950 *vec_stmt
= build_gimple_modify_stmt (vec_dest
, vec_cond_expr
);
4951 new_temp
= make_ssa_name (vec_dest
, *vec_stmt
);
4952 GIMPLE_STMT_OPERAND (*vec_stmt
, 0) = new_temp
;
4953 vect_finish_stmt_generation (stmt
, *vec_stmt
, bsi
);
4958 /* Function vect_transform_stmt.
4960 Create a vectorized stmt to replace STMT, and insert it at BSI. */
4963 vect_transform_stmt (tree stmt
, block_stmt_iterator
*bsi
, bool *strided_store
)
4965 bool is_store
= false;
4966 tree vec_stmt
= NULL_TREE
;
4967 stmt_vec_info stmt_info
= vinfo_for_stmt (stmt
);
4968 tree orig_stmt_in_pattern
;
4971 switch (STMT_VINFO_TYPE (stmt_info
))
4973 case type_demotion_vec_info_type
:
4974 done
= vectorizable_type_demotion (stmt
, bsi
, &vec_stmt
);
4978 case type_promotion_vec_info_type
:
4979 done
= vectorizable_type_promotion (stmt
, bsi
, &vec_stmt
);
4983 case type_conversion_vec_info_type
:
4984 done
= vectorizable_conversion (stmt
, bsi
, &vec_stmt
);
4988 case induc_vec_info_type
:
4989 done
= vectorizable_induction (stmt
, bsi
, &vec_stmt
);
4993 case op_vec_info_type
:
4994 done
= vectorizable_operation (stmt
, bsi
, &vec_stmt
);
4998 case assignment_vec_info_type
:
4999 done
= vectorizable_assignment (stmt
, bsi
, &vec_stmt
);
5003 case load_vec_info_type
:
5004 done
= vectorizable_load (stmt
, bsi
, &vec_stmt
);
5008 case store_vec_info_type
:
5009 done
= vectorizable_store (stmt
, bsi
, &vec_stmt
);
5011 if (DR_GROUP_FIRST_DR (stmt_info
))
5013 /* In case of interleaving, the whole chain is vectorized when the
5014 last store in the chain is reached. Store stmts before the last
5015 one are skipped, and there vec_stmt_info shouldn't be freed
5017 *strided_store
= true;
5018 if (STMT_VINFO_VEC_STMT (stmt_info
))
5025 case condition_vec_info_type
:
5026 done
= vectorizable_condition (stmt
, bsi
, &vec_stmt
);
5030 case call_vec_info_type
:
5031 done
= vectorizable_call (stmt
, bsi
, &vec_stmt
);
5034 case reduc_vec_info_type
:
5035 done
= vectorizable_reduction (stmt
, bsi
, &vec_stmt
);
5040 if (!STMT_VINFO_LIVE_P (stmt_info
))
5042 if (vect_print_dump_info (REPORT_DETAILS
))
5043 fprintf (vect_dump
, "stmt not supported.");
5048 if (STMT_VINFO_LIVE_P (stmt_info
)
5049 && STMT_VINFO_TYPE (stmt_info
) != reduc_vec_info_type
)
5051 done
= vectorizable_live_operation (stmt
, bsi
, &vec_stmt
);
5057 STMT_VINFO_VEC_STMT (stmt_info
) = vec_stmt
;
5058 orig_stmt_in_pattern
= STMT_VINFO_RELATED_STMT (stmt_info
);
5059 if (orig_stmt_in_pattern
)
5061 stmt_vec_info stmt_vinfo
= vinfo_for_stmt (orig_stmt_in_pattern
);
5062 /* STMT was inserted by the vectorizer to replace a computation idiom.
5063 ORIG_STMT_IN_PATTERN is a stmt in the original sequence that
5064 computed this idiom. We need to record a pointer to VEC_STMT in
5065 the stmt_info of ORIG_STMT_IN_PATTERN. See more details in the
5066 documentation of vect_pattern_recog. */
5067 if (STMT_VINFO_IN_PATTERN_P (stmt_vinfo
))
5069 gcc_assert (STMT_VINFO_RELATED_STMT (stmt_vinfo
) == stmt
);
5070 STMT_VINFO_VEC_STMT (stmt_vinfo
) = vec_stmt
;
5079 /* This function builds ni_name = number of iterations loop executes
5080 on the loop preheader. */
5083 vect_build_loop_niters (loop_vec_info loop_vinfo
)
5085 tree ni_name
, stmt
, var
;
5087 struct loop
*loop
= LOOP_VINFO_LOOP (loop_vinfo
);
5088 tree ni
= unshare_expr (LOOP_VINFO_NITERS (loop_vinfo
));
5090 var
= create_tmp_var (TREE_TYPE (ni
), "niters");
5091 add_referenced_var (var
);
5092 ni_name
= force_gimple_operand (ni
, &stmt
, false, var
);
5094 pe
= loop_preheader_edge (loop
);
5097 basic_block new_bb
= bsi_insert_on_edge_immediate (pe
, stmt
);
5098 gcc_assert (!new_bb
);
5105 /* This function generates the following statements:
5107 ni_name = number of iterations loop executes
5108 ratio = ni_name / vf
5109 ratio_mult_vf_name = ratio * vf
5111 and places them at the loop preheader edge. */
5114 vect_generate_tmps_on_preheader (loop_vec_info loop_vinfo
,
5116 tree
*ratio_mult_vf_name_ptr
,
5117 tree
*ratio_name_ptr
)
5125 tree ratio_mult_vf_name
;
5126 struct loop
*loop
= LOOP_VINFO_LOOP (loop_vinfo
);
5127 tree ni
= LOOP_VINFO_NITERS (loop_vinfo
);
5128 int vf
= LOOP_VINFO_VECT_FACTOR (loop_vinfo
);
5131 pe
= loop_preheader_edge (loop
);
5133 /* Generate temporary variable that contains
5134 number of iterations loop executes. */
5136 ni_name
= vect_build_loop_niters (loop_vinfo
);
5137 log_vf
= build_int_cst (TREE_TYPE (ni
), exact_log2 (vf
));
5139 /* Create: ratio = ni >> log2(vf) */
5141 ratio_name
= fold_build2 (RSHIFT_EXPR
, TREE_TYPE (ni_name
), ni_name
, log_vf
);
5142 if (!is_gimple_val (ratio_name
))
5144 var
= create_tmp_var (TREE_TYPE (ni
), "bnd");
5145 add_referenced_var (var
);
5147 ratio_name
= force_gimple_operand (ratio_name
, &stmt
, true, var
);
5148 pe
= loop_preheader_edge (loop
);
5149 new_bb
= bsi_insert_on_edge_immediate (pe
, stmt
);
5150 gcc_assert (!new_bb
);
5153 /* Create: ratio_mult_vf = ratio << log2 (vf). */
5155 ratio_mult_vf_name
= fold_build2 (LSHIFT_EXPR
, TREE_TYPE (ratio_name
),
5156 ratio_name
, log_vf
);
5157 if (!is_gimple_val (ratio_mult_vf_name
))
5159 var
= create_tmp_var (TREE_TYPE (ni
), "ratio_mult_vf");
5160 add_referenced_var (var
);
5162 ratio_mult_vf_name
= force_gimple_operand (ratio_mult_vf_name
, &stmt
,
5164 pe
= loop_preheader_edge (loop
);
5165 new_bb
= bsi_insert_on_edge_immediate (pe
, stmt
);
5166 gcc_assert (!new_bb
);
5169 *ni_name_ptr
= ni_name
;
5170 *ratio_mult_vf_name_ptr
= ratio_mult_vf_name
;
5171 *ratio_name_ptr
= ratio_name
;
5177 /* Function update_vuses_to_preheader.
5180 STMT - a statement with potential VUSEs.
5181 LOOP - the loop whose preheader will contain STMT.
5183 It's possible to vectorize a loop even though an SSA_NAME from a VUSE
5184 appears to be defined in a VDEF in another statement in a loop.
5185 One such case is when the VUSE is at the dereference of a __restricted__
5186 pointer in a load and the VDEF is at the dereference of a different
5187 __restricted__ pointer in a store. Vectorization may result in
5188 copy_virtual_uses being called to copy the problematic VUSE to a new
5189 statement that is being inserted in the loop preheader. This procedure
5190 is called to change the SSA_NAME in the new statement's VUSE from the
5191 SSA_NAME updated in the loop to the related SSA_NAME available on the
5192 path entering the loop.
5194 When this function is called, we have the following situation:
5199 # name1 = phi < name0 , name2>
5204 # name2 = vdef <name1>
5209 Stmt S1 was created in the loop preheader block as part of misaligned-load
5210 handling. This function fixes the name of the vuse of S1 from 'name1' to
5214 update_vuses_to_preheader (tree stmt
, struct loop
*loop
)
5216 basic_block header_bb
= loop
->header
;
5217 edge preheader_e
= loop_preheader_edge (loop
);
5219 use_operand_p use_p
;
5221 FOR_EACH_SSA_USE_OPERAND (use_p
, stmt
, iter
, SSA_OP_VUSE
)
5223 tree ssa_name
= USE_FROM_PTR (use_p
);
5224 tree def_stmt
= SSA_NAME_DEF_STMT (ssa_name
);
5225 tree name_var
= SSA_NAME_VAR (ssa_name
);
5226 basic_block bb
= bb_for_stmt (def_stmt
);
5228 /* For a use before any definitions, def_stmt is a NOP_EXPR. */
5229 if (!IS_EMPTY_STMT (def_stmt
)
5230 && flow_bb_inside_loop_p (loop
, bb
))
5232 /* If the block containing the statement defining the SSA_NAME
5233 is in the loop then it's necessary to find the definition
5234 outside the loop using the PHI nodes of the header. */
5236 bool updated
= false;
5238 for (phi
= phi_nodes (header_bb
); phi
; phi
= PHI_CHAIN (phi
))
5240 if (SSA_NAME_VAR (PHI_RESULT (phi
)) == name_var
)
5242 SET_USE (use_p
, PHI_ARG_DEF (phi
, preheader_e
->dest_idx
));
5247 gcc_assert (updated
);
5253 /* Function vect_update_ivs_after_vectorizer.
5255 "Advance" the induction variables of LOOP to the value they should take
5256 after the execution of LOOP. This is currently necessary because the
5257 vectorizer does not handle induction variables that are used after the
5258 loop. Such a situation occurs when the last iterations of LOOP are
5260 1. We introduced new uses after LOOP for IVs that were not originally used
5261 after LOOP: the IVs of LOOP are now used by an epilog loop.
5262 2. LOOP is going to be vectorized; this means that it will iterate N/VF
5263 times, whereas the loop IVs should be bumped N times.
5266 - LOOP - a loop that is going to be vectorized. The last few iterations
5267 of LOOP were peeled.
5268 - NITERS - the number of iterations that LOOP executes (before it is
5269 vectorized). i.e, the number of times the ivs should be bumped.
5270 - UPDATE_E - a successor edge of LOOP->exit that is on the (only) path
5271 coming out from LOOP on which there are uses of the LOOP ivs
5272 (this is the path from LOOP->exit to epilog_loop->preheader).
5274 The new definitions of the ivs are placed in LOOP->exit.
5275 The phi args associated with the edge UPDATE_E in the bb
5276 UPDATE_E->dest are updated accordingly.
5278 Assumption 1: Like the rest of the vectorizer, this function assumes
5279 a single loop exit that has a single predecessor.
5281 Assumption 2: The phi nodes in the LOOP header and in update_bb are
5282 organized in the same order.
5284 Assumption 3: The access function of the ivs is simple enough (see
5285 vect_can_advance_ivs_p). This assumption will be relaxed in the future.
5287 Assumption 4: Exactly one of the successors of LOOP exit-bb is on a path
5288 coming out of LOOP on which the ivs of LOOP are used (this is the path
5289 that leads to the epilog loop; other paths skip the epilog loop). This
5290 path starts with the edge UPDATE_E, and its destination (denoted update_bb)
5291 needs to have its phis updated.
5295 vect_update_ivs_after_vectorizer (loop_vec_info loop_vinfo
, tree niters
,
5298 struct loop
*loop
= LOOP_VINFO_LOOP (loop_vinfo
);
5299 basic_block exit_bb
= single_exit (loop
)->dest
;
5301 basic_block update_bb
= update_e
->dest
;
5303 /* gcc_assert (vect_can_advance_ivs_p (loop_vinfo)); */
5305 /* Make sure there exists a single-predecessor exit bb: */
5306 gcc_assert (single_pred_p (exit_bb
));
5308 for (phi
= phi_nodes (loop
->header
), phi1
= phi_nodes (update_bb
);
5310 phi
= PHI_CHAIN (phi
), phi1
= PHI_CHAIN (phi1
))
5312 tree access_fn
= NULL
;
5313 tree evolution_part
;
5316 tree var
, stmt
, ni
, ni_name
;
5317 block_stmt_iterator last_bsi
;
5319 if (vect_print_dump_info (REPORT_DETAILS
))
5321 fprintf (vect_dump
, "vect_update_ivs_after_vectorizer: phi: ");
5322 print_generic_expr (vect_dump
, phi
, TDF_SLIM
);
5325 /* Skip virtual phi's. */
5326 if (!is_gimple_reg (SSA_NAME_VAR (PHI_RESULT (phi
))))
5328 if (vect_print_dump_info (REPORT_DETAILS
))
5329 fprintf (vect_dump
, "virtual phi. skip.");
5333 /* Skip reduction phis. */
5334 if (STMT_VINFO_DEF_TYPE (vinfo_for_stmt (phi
)) == vect_reduction_def
)
5336 if (vect_print_dump_info (REPORT_DETAILS
))
5337 fprintf (vect_dump
, "reduc phi. skip.");
5341 access_fn
= analyze_scalar_evolution (loop
, PHI_RESULT (phi
));
5342 gcc_assert (access_fn
);
5344 unshare_expr (evolution_part_in_loop_num (access_fn
, loop
->num
));
5345 gcc_assert (evolution_part
!= NULL_TREE
);
5347 /* FORNOW: We do not support IVs whose evolution function is a polynomial
5348 of degree >= 2 or exponential. */
5349 gcc_assert (!tree_is_chrec (evolution_part
));
5351 step_expr
= evolution_part
;
5352 init_expr
= unshare_expr (initial_condition_in_loop_num (access_fn
,
5355 if (POINTER_TYPE_P (TREE_TYPE (init_expr
)))
5356 ni
= fold_build2 (POINTER_PLUS_EXPR
, TREE_TYPE (init_expr
),
5358 fold_convert (sizetype
,
5359 fold_build2 (MULT_EXPR
, TREE_TYPE (niters
),
5360 niters
, step_expr
)));
5362 ni
= fold_build2 (PLUS_EXPR
, TREE_TYPE (init_expr
),
5363 fold_build2 (MULT_EXPR
, TREE_TYPE (init_expr
),
5364 fold_convert (TREE_TYPE (init_expr
),
5371 var
= create_tmp_var (TREE_TYPE (init_expr
), "tmp");
5372 add_referenced_var (var
);
5374 ni_name
= force_gimple_operand (ni
, &stmt
, false, var
);
5376 /* Insert stmt into exit_bb. */
5377 last_bsi
= bsi_last (exit_bb
);
5379 bsi_insert_before (&last_bsi
, stmt
, BSI_SAME_STMT
);
5381 /* Fix phi expressions in the successor bb. */
5382 SET_PHI_ARG_DEF (phi1
, update_e
->dest_idx
, ni_name
);
5387 /* Function vect_do_peeling_for_loop_bound
5389 Peel the last iterations of the loop represented by LOOP_VINFO.
5390 The peeled iterations form a new epilog loop. Given that the loop now
5391 iterates NITERS times, the new epilog loop iterates
5392 NITERS % VECTORIZATION_FACTOR times.
5394 The original loop will later be made to iterate
5395 NITERS / VECTORIZATION_FACTOR times (this value is placed into RATIO). */
5398 vect_do_peeling_for_loop_bound (loop_vec_info loop_vinfo
, tree
*ratio
)
5400 tree ni_name
, ratio_mult_vf_name
;
5401 struct loop
*loop
= LOOP_VINFO_LOOP (loop_vinfo
);
5402 struct loop
*new_loop
;
5404 basic_block preheader
;
5407 int min_scalar_loop_bound
;
5408 int min_profitable_iters
;
5410 if (vect_print_dump_info (REPORT_DETAILS
))
5411 fprintf (vect_dump
, "=== vect_do_peeling_for_loop_bound ===");
5413 initialize_original_copy_tables ();
5415 /* Generate the following variables on the preheader of original loop:
5417 ni_name = number of iteration the original loop executes
5418 ratio = ni_name / vf
5419 ratio_mult_vf_name = ratio * vf */
5420 vect_generate_tmps_on_preheader (loop_vinfo
, &ni_name
,
5421 &ratio_mult_vf_name
, ratio
);
5423 loop_num
= loop
->num
;
5425 /* Analyze cost to set threshhold for vectorized loop. */
5426 min_profitable_iters
= vect_estimate_min_profitable_iters (loop_vinfo
);
5428 min_scalar_loop_bound
= (PARAM_VALUE (PARAM_MIN_VECT_LOOP_BOUND
))
5429 * LOOP_VINFO_VECT_FACTOR (loop_vinfo
);
5431 /* Use the cost model only if it is more conservative than user specified
5434 th
= (unsigned) min_scalar_loop_bound
;
5435 if (min_profitable_iters
5436 && (!min_scalar_loop_bound
5437 || min_profitable_iters
> min_scalar_loop_bound
))
5438 th
= (unsigned) min_profitable_iters
;
5440 if (vect_print_dump_info (REPORT_DETAILS
))
5441 fprintf (vect_dump
, "vectorization may not be profitable.");
5443 new_loop
= slpeel_tree_peel_loop_to_edge (loop
, single_exit (loop
),
5444 ratio_mult_vf_name
, ni_name
, false,
5446 gcc_assert (new_loop
);
5447 gcc_assert (loop_num
== loop
->num
);
5448 #ifdef ENABLE_CHECKING
5449 slpeel_verify_cfg_after_peeling (loop
, new_loop
);
5452 /* A guard that controls whether the new_loop is to be executed or skipped
5453 is placed in LOOP->exit. LOOP->exit therefore has two successors - one
5454 is the preheader of NEW_LOOP, where the IVs from LOOP are used. The other
5455 is a bb after NEW_LOOP, where these IVs are not used. Find the edge that
5456 is on the path where the LOOP IVs are used and need to be updated. */
5458 preheader
= loop_preheader_edge (new_loop
)->src
;
5459 if (EDGE_PRED (preheader
, 0)->src
== single_exit (loop
)->dest
)
5460 update_e
= EDGE_PRED (preheader
, 0);
5462 update_e
= EDGE_PRED (preheader
, 1);
5464 /* Update IVs of original loop as if they were advanced
5465 by ratio_mult_vf_name steps. */
5466 vect_update_ivs_after_vectorizer (loop_vinfo
, ratio_mult_vf_name
, update_e
);
5468 /* After peeling we have to reset scalar evolution analyzer. */
5471 free_original_copy_tables ();
5475 /* Function vect_gen_niters_for_prolog_loop
5477 Set the number of iterations for the loop represented by LOOP_VINFO
5478 to the minimum between LOOP_NITERS (the original iteration count of the loop)
5479 and the misalignment of DR - the data reference recorded in
5480 LOOP_VINFO_UNALIGNED_DR (LOOP_VINFO). As a result, after the execution of
5481 this loop, the data reference DR will refer to an aligned location.
5483 The following computation is generated:
5485 If the misalignment of DR is known at compile time:
5486 addr_mis = int mis = DR_MISALIGNMENT (dr);
5487 Else, compute address misalignment in bytes:
5488 addr_mis = addr & (vectype_size - 1)
5490 prolog_niters = min ( LOOP_NITERS , (VF - addr_mis/elem_size)&(VF-1) )
5492 (elem_size = element type size; an element is the scalar element
5493 whose type is the inner type of the vectype)
5497 prolog_niters = min ( LOOP_NITERS ,
5498 (VF/group_size - addr_mis/elem_size)&(VF/group_size-1) )
5499 where group_size is the size of the interleaved group.
5501 The above formulas assume that VF == number of elements in the vector. This
5502 may not hold when there are multiple-types in the loop.
5503 In this case, for some data-references in the loop the VF does not represent
5504 the number of elements that fit in the vector. Therefore, instead of VF we
5505 use TYPE_VECTOR_SUBPARTS. */
5508 vect_gen_niters_for_prolog_loop (loop_vec_info loop_vinfo
, tree loop_niters
)
5510 struct data_reference
*dr
= LOOP_VINFO_UNALIGNED_DR (loop_vinfo
);
5511 struct loop
*loop
= LOOP_VINFO_LOOP (loop_vinfo
);
5513 tree iters
, iters_name
;
5516 tree dr_stmt
= DR_STMT (dr
);
5517 stmt_vec_info stmt_info
= vinfo_for_stmt (dr_stmt
);
5518 tree vectype
= STMT_VINFO_VECTYPE (stmt_info
);
5519 int vectype_align
= TYPE_ALIGN (vectype
) / BITS_PER_UNIT
;
5520 tree niters_type
= TREE_TYPE (loop_niters
);
5522 int element_size
= GET_MODE_SIZE (TYPE_MODE (TREE_TYPE (DR_REF (dr
))));
5523 int nelements
= TYPE_VECTOR_SUBPARTS (vectype
);
5525 if (DR_GROUP_FIRST_DR (stmt_info
))
5527 /* For interleaved access element size must be multiplied by the size of
5528 the interleaved group. */
5529 group_size
= DR_GROUP_SIZE (vinfo_for_stmt (
5530 DR_GROUP_FIRST_DR (stmt_info
)));
5531 element_size
*= group_size
;
5534 pe
= loop_preheader_edge (loop
);
5536 if (LOOP_PEELING_FOR_ALIGNMENT (loop_vinfo
) > 0)
5538 int byte_misalign
= LOOP_PEELING_FOR_ALIGNMENT (loop_vinfo
);
5539 int elem_misalign
= byte_misalign
/ element_size
;
5541 if (vect_print_dump_info (REPORT_DETAILS
))
5542 fprintf (vect_dump
, "known alignment = %d.", byte_misalign
);
5543 iters
= build_int_cst (niters_type
,
5544 (nelements
- elem_misalign
)&(nelements
/group_size
-1));
5548 tree new_stmts
= NULL_TREE
;
5550 vect_create_addr_base_for_vector_ref (dr_stmt
, &new_stmts
, NULL_TREE
);
5551 tree ptr_type
= TREE_TYPE (start_addr
);
5552 tree size
= TYPE_SIZE (ptr_type
);
5553 tree type
= lang_hooks
.types
.type_for_size (tree_low_cst (size
, 1), 1);
5554 tree vectype_size_minus_1
= build_int_cst (type
, vectype_align
- 1);
5555 tree elem_size_log
=
5556 build_int_cst (type
, exact_log2 (vectype_align
/nelements
));
5557 tree nelements_minus_1
= build_int_cst (type
, nelements
- 1);
5558 tree nelements_tree
= build_int_cst (type
, nelements
);
5562 new_bb
= bsi_insert_on_edge_immediate (pe
, new_stmts
);
5563 gcc_assert (!new_bb
);
5565 /* Create: byte_misalign = addr & (vectype_size - 1) */
5567 fold_build2 (BIT_AND_EXPR
, type
, fold_convert (type
, start_addr
), vectype_size_minus_1
);
5569 /* Create: elem_misalign = byte_misalign / element_size */
5571 fold_build2 (RSHIFT_EXPR
, type
, byte_misalign
, elem_size_log
);
5573 /* Create: (niters_type) (nelements - elem_misalign)&(nelements - 1) */
5574 iters
= fold_build2 (MINUS_EXPR
, type
, nelements_tree
, elem_misalign
);
5575 iters
= fold_build2 (BIT_AND_EXPR
, type
, iters
, nelements_minus_1
);
5576 iters
= fold_convert (niters_type
, iters
);
5579 /* Create: prolog_loop_niters = min (iters, loop_niters) */
5580 /* If the loop bound is known at compile time we already verified that it is
5581 greater than vf; since the misalignment ('iters') is at most vf, there's
5582 no need to generate the MIN_EXPR in this case. */
5583 if (TREE_CODE (loop_niters
) != INTEGER_CST
)
5584 iters
= fold_build2 (MIN_EXPR
, niters_type
, iters
, loop_niters
);
5586 if (vect_print_dump_info (REPORT_DETAILS
))
5588 fprintf (vect_dump
, "niters for prolog loop: ");
5589 print_generic_expr (vect_dump
, iters
, TDF_SLIM
);
5592 var
= create_tmp_var (niters_type
, "prolog_loop_niters");
5593 add_referenced_var (var
);
5594 iters_name
= force_gimple_operand (iters
, &stmt
, false, var
);
5596 /* Insert stmt on loop preheader edge. */
5599 basic_block new_bb
= bsi_insert_on_edge_immediate (pe
, stmt
);
5600 gcc_assert (!new_bb
);
5607 /* Function vect_update_init_of_dr
5609 NITERS iterations were peeled from LOOP. DR represents a data reference
5610 in LOOP. This function updates the information recorded in DR to
5611 account for the fact that the first NITERS iterations had already been
5612 executed. Specifically, it updates the OFFSET field of DR. */
5615 vect_update_init_of_dr (struct data_reference
*dr
, tree niters
)
5617 tree offset
= DR_OFFSET (dr
);
5619 niters
= fold_build2 (MULT_EXPR
, TREE_TYPE (niters
), niters
, DR_STEP (dr
));
5620 offset
= fold_build2 (PLUS_EXPR
, TREE_TYPE (offset
), offset
, niters
);
5621 DR_OFFSET (dr
) = offset
;
5625 /* Function vect_update_inits_of_drs
5627 NITERS iterations were peeled from the loop represented by LOOP_VINFO.
5628 This function updates the information recorded for the data references in
5629 the loop to account for the fact that the first NITERS iterations had
5630 already been executed. Specifically, it updates the initial_condition of
5631 the access_function of all the data_references in the loop. */
5634 vect_update_inits_of_drs (loop_vec_info loop_vinfo
, tree niters
)
5637 VEC (data_reference_p
, heap
) *datarefs
= LOOP_VINFO_DATAREFS (loop_vinfo
);
5638 struct data_reference
*dr
;
5640 if (vect_print_dump_info (REPORT_DETAILS
))
5641 fprintf (vect_dump
, "=== vect_update_inits_of_dr ===");
5643 for (i
= 0; VEC_iterate (data_reference_p
, datarefs
, i
, dr
); i
++)
5644 vect_update_init_of_dr (dr
, niters
);
5648 /* Function vect_do_peeling_for_alignment
5650 Peel the first 'niters' iterations of the loop represented by LOOP_VINFO.
5651 'niters' is set to the misalignment of one of the data references in the
5652 loop, thereby forcing it to refer to an aligned location at the beginning
5653 of the execution of this loop. The data reference for which we are
5654 peeling is recorded in LOOP_VINFO_UNALIGNED_DR. */
5657 vect_do_peeling_for_alignment (loop_vec_info loop_vinfo
)
5659 struct loop
*loop
= LOOP_VINFO_LOOP (loop_vinfo
);
5660 tree niters_of_prolog_loop
, ni_name
;
5662 struct loop
*new_loop
;
5664 if (vect_print_dump_info (REPORT_DETAILS
))
5665 fprintf (vect_dump
, "=== vect_do_peeling_for_alignment ===");
5667 initialize_original_copy_tables ();
5669 ni_name
= vect_build_loop_niters (loop_vinfo
);
5670 niters_of_prolog_loop
= vect_gen_niters_for_prolog_loop (loop_vinfo
, ni_name
);
5672 /* Peel the prolog loop and iterate it niters_of_prolog_loop. */
5674 slpeel_tree_peel_loop_to_edge (loop
, loop_preheader_edge (loop
),
5675 niters_of_prolog_loop
, ni_name
, true, 0);
5676 gcc_assert (new_loop
);
5677 #ifdef ENABLE_CHECKING
5678 slpeel_verify_cfg_after_peeling (new_loop
, loop
);
5681 /* Update number of times loop executes. */
5682 n_iters
= LOOP_VINFO_NITERS (loop_vinfo
);
5683 LOOP_VINFO_NITERS (loop_vinfo
) = fold_build2 (MINUS_EXPR
,
5684 TREE_TYPE (n_iters
), n_iters
, niters_of_prolog_loop
);
5686 /* Update the init conditions of the access functions of all data refs. */
5687 vect_update_inits_of_drs (loop_vinfo
, niters_of_prolog_loop
);
5689 /* After peeling we have to reset scalar evolution analyzer. */
5692 free_original_copy_tables ();
5696 /* Function vect_create_cond_for_align_checks.
5698 Create a conditional expression that represents the alignment checks for
5699 all of data references (array element references) whose alignment must be
5703 LOOP_VINFO - two fields of the loop information are used.
5704 LOOP_VINFO_PTR_MASK is the mask used to check the alignment.
5705 LOOP_VINFO_MAY_MISALIGN_STMTS contains the refs to be checked.
5708 COND_EXPR_STMT_LIST - statements needed to construct the conditional
5710 The returned value is the conditional expression to be used in the if
5711 statement that controls which version of the loop gets executed at runtime.
5713 The algorithm makes two assumptions:
5714 1) The number of bytes "n" in a vector is a power of 2.
5715 2) An address "a" is aligned if a%n is zero and that this
5716 test can be done as a&(n-1) == 0. For example, for 16
5717 byte vectors the test is a&0xf == 0. */
5720 vect_create_cond_for_align_checks (loop_vec_info loop_vinfo
,
5721 tree
*cond_expr_stmt_list
)
5723 VEC(tree
,heap
) *may_misalign_stmts
5724 = LOOP_VINFO_MAY_MISALIGN_STMTS (loop_vinfo
);
5726 int mask
= LOOP_VINFO_PTR_MASK (loop_vinfo
);
5730 tree int_ptrsize_type
;
5732 tree or_tmp_name
= NULL_TREE
;
5733 tree and_tmp
, and_tmp_name
, and_stmt
;
5736 /* Check that mask is one less than a power of 2, i.e., mask is
5737 all zeros followed by all ones. */
5738 gcc_assert ((mask
!= 0) && ((mask
& (mask
+1)) == 0));
5740 /* CHECKME: what is the best integer or unsigned type to use to hold a
5741 cast from a pointer value? */
5742 psize
= TYPE_SIZE (ptr_type_node
);
5744 = lang_hooks
.types
.type_for_size (tree_low_cst (psize
, 1), 0);
5746 /* Create expression (mask & (dr_1 || ... || dr_n)) where dr_i is the address
5747 of the first vector of the i'th data reference. */
5749 for (i
= 0; VEC_iterate (tree
, may_misalign_stmts
, i
, ref_stmt
); i
++)
5751 tree new_stmt_list
= NULL_TREE
;
5753 tree addr_tmp
, addr_tmp_name
, addr_stmt
;
5754 tree or_tmp
, new_or_tmp_name
, or_stmt
;
5756 /* create: addr_tmp = (int)(address_of_first_vector) */
5757 addr_base
= vect_create_addr_base_for_vector_ref (ref_stmt
,
5761 if (new_stmt_list
!= NULL_TREE
)
5762 append_to_statement_list_force (new_stmt_list
, cond_expr_stmt_list
);
5764 sprintf (tmp_name
, "%s%d", "addr2int", i
);
5765 addr_tmp
= create_tmp_var (int_ptrsize_type
, tmp_name
);
5766 add_referenced_var (addr_tmp
);
5767 addr_tmp_name
= make_ssa_name (addr_tmp
, NULL_TREE
);
5768 addr_stmt
= fold_convert (int_ptrsize_type
, addr_base
);
5769 addr_stmt
= build_gimple_modify_stmt (addr_tmp_name
, addr_stmt
);
5770 SSA_NAME_DEF_STMT (addr_tmp_name
) = addr_stmt
;
5771 append_to_statement_list_force (addr_stmt
, cond_expr_stmt_list
);
5773 /* The addresses are OR together. */
5775 if (or_tmp_name
!= NULL_TREE
)
5777 /* create: or_tmp = or_tmp | addr_tmp */
5778 sprintf (tmp_name
, "%s%d", "orptrs", i
);
5779 or_tmp
= create_tmp_var (int_ptrsize_type
, tmp_name
);
5780 add_referenced_var (or_tmp
);
5781 new_or_tmp_name
= make_ssa_name (or_tmp
, NULL_TREE
);
5782 tmp
= build2 (BIT_IOR_EXPR
, int_ptrsize_type
,
5783 or_tmp_name
, addr_tmp_name
);
5784 or_stmt
= build_gimple_modify_stmt (new_or_tmp_name
, tmp
);
5785 SSA_NAME_DEF_STMT (new_or_tmp_name
) = or_stmt
;
5786 append_to_statement_list_force (or_stmt
, cond_expr_stmt_list
);
5787 or_tmp_name
= new_or_tmp_name
;
5790 or_tmp_name
= addr_tmp_name
;
5794 mask_cst
= build_int_cst (int_ptrsize_type
, mask
);
5796 /* create: and_tmp = or_tmp & mask */
5797 and_tmp
= create_tmp_var (int_ptrsize_type
, "andmask" );
5798 add_referenced_var (and_tmp
);
5799 and_tmp_name
= make_ssa_name (and_tmp
, NULL_TREE
);
5801 tmp
= build2 (BIT_AND_EXPR
, int_ptrsize_type
, or_tmp_name
, mask_cst
);
5802 and_stmt
= build_gimple_modify_stmt (and_tmp_name
, tmp
);
5803 SSA_NAME_DEF_STMT (and_tmp_name
) = and_stmt
;
5804 append_to_statement_list_force (and_stmt
, cond_expr_stmt_list
);
5806 /* Make and_tmp the left operand of the conditional test against zero.
5807 if and_tmp has a nonzero bit then some address is unaligned. */
5808 ptrsize_zero
= build_int_cst (int_ptrsize_type
, 0);
5809 return build2 (EQ_EXPR
, boolean_type_node
,
5810 and_tmp_name
, ptrsize_zero
);
5814 /* Function vect_transform_loop.
5816 The analysis phase has determined that the loop is vectorizable.
5817 Vectorize the loop - created vectorized stmts to replace the scalar
5818 stmts in the loop, and update the loop exit condition. */
5821 vect_transform_loop (loop_vec_info loop_vinfo
)
5823 struct loop
*loop
= LOOP_VINFO_LOOP (loop_vinfo
);
5824 basic_block
*bbs
= LOOP_VINFO_BBS (loop_vinfo
);
5825 int nbbs
= loop
->num_nodes
;
5826 block_stmt_iterator si
, next_si
;
5829 int vectorization_factor
= LOOP_VINFO_VECT_FACTOR (loop_vinfo
);
5832 if (vect_print_dump_info (REPORT_DETAILS
))
5833 fprintf (vect_dump
, "=== vec_transform_loop ===");
5835 /* If the loop has data references that may or may not be aligned then
5836 two versions of the loop need to be generated, one which is vectorized
5837 and one which isn't. A test is then generated to control which of the
5838 loops is executed. The test checks for the alignment of all of the
5839 data references that may or may not be aligned. */
5841 if (VEC_length (tree
, LOOP_VINFO_MAY_MISALIGN_STMTS (loop_vinfo
)))
5845 tree cond_expr_stmt_list
= NULL_TREE
;
5846 basic_block condition_bb
;
5847 block_stmt_iterator cond_exp_bsi
;
5848 basic_block merge_bb
;
5849 basic_block new_exit_bb
;
5851 tree orig_phi
, new_phi
, arg
;
5852 unsigned prob
= 4 * REG_BR_PROB_BASE
/ 5;
5854 cond_expr
= vect_create_cond_for_align_checks (loop_vinfo
,
5855 &cond_expr_stmt_list
);
5856 initialize_original_copy_tables ();
5857 nloop
= loop_version (loop
, cond_expr
, &condition_bb
,
5858 prob
, prob
, REG_BR_PROB_BASE
- prob
, true);
5859 free_original_copy_tables();
5861 /** Loop versioning violates an assumption we try to maintain during
5862 vectorization - that the loop exit block has a single predecessor.
5863 After versioning, the exit block of both loop versions is the same
5864 basic block (i.e. it has two predecessors). Just in order to simplify
5865 following transformations in the vectorizer, we fix this situation
5866 here by adding a new (empty) block on the exit-edge of the loop,
5867 with the proper loop-exit phis to maintain loop-closed-form. **/
5869 merge_bb
= single_exit (loop
)->dest
;
5870 gcc_assert (EDGE_COUNT (merge_bb
->preds
) == 2);
5871 new_exit_bb
= split_edge (single_exit (loop
));
5872 new_exit_e
= single_exit (loop
);
5873 e
= EDGE_SUCC (new_exit_bb
, 0);
5875 for (orig_phi
= phi_nodes (merge_bb
); orig_phi
;
5876 orig_phi
= PHI_CHAIN (orig_phi
))
5878 new_phi
= create_phi_node (SSA_NAME_VAR (PHI_RESULT (orig_phi
)),
5880 arg
= PHI_ARG_DEF_FROM_EDGE (orig_phi
, e
);
5881 add_phi_arg (new_phi
, arg
, new_exit_e
);
5882 SET_PHI_ARG_DEF (orig_phi
, e
->dest_idx
, PHI_RESULT (new_phi
));
5885 /** end loop-exit-fixes after versioning **/
5887 update_ssa (TODO_update_ssa
);
5888 cond_exp_bsi
= bsi_last (condition_bb
);
5889 bsi_insert_before (&cond_exp_bsi
, cond_expr_stmt_list
, BSI_SAME_STMT
);
5892 /* CHECKME: we wouldn't need this if we called update_ssa once
5894 bitmap_zero (vect_memsyms_to_rename
);
5896 /* Peel the loop if there are data refs with unknown alignment.
5897 Only one data ref with unknown store is allowed. */
5899 if (LOOP_PEELING_FOR_ALIGNMENT (loop_vinfo
))
5900 vect_do_peeling_for_alignment (loop_vinfo
);
5902 /* If the loop has a symbolic number of iterations 'n' (i.e. it's not a
5903 compile time constant), or it is a constant that doesn't divide by the
5904 vectorization factor, then an epilog loop needs to be created.
5905 We therefore duplicate the loop: the original loop will be vectorized,
5906 and will compute the first (n/VF) iterations. The second copy of the loop
5907 will remain scalar and will compute the remaining (n%VF) iterations.
5908 (VF is the vectorization factor). */
5910 if (!LOOP_VINFO_NITERS_KNOWN_P (loop_vinfo
)
5911 || (LOOP_VINFO_NITERS_KNOWN_P (loop_vinfo
)
5912 && LOOP_VINFO_INT_NITERS (loop_vinfo
) % vectorization_factor
!= 0))
5913 vect_do_peeling_for_loop_bound (loop_vinfo
, &ratio
);
5915 ratio
= build_int_cst (TREE_TYPE (LOOP_VINFO_NITERS (loop_vinfo
)),
5916 LOOP_VINFO_INT_NITERS (loop_vinfo
) / vectorization_factor
);
5918 /* 1) Make sure the loop header has exactly two entries
5919 2) Make sure we have a preheader basic block. */
5921 gcc_assert (EDGE_COUNT (loop
->header
->preds
) == 2);
5923 split_edge (loop_preheader_edge (loop
));
5925 /* FORNOW: the vectorizer supports only loops which body consist
5926 of one basic block (header + empty latch). When the vectorizer will
5927 support more involved loop forms, the order by which the BBs are
5928 traversed need to be reconsidered. */
5930 for (i
= 0; i
< nbbs
; i
++)
5932 basic_block bb
= bbs
[i
];
5933 stmt_vec_info stmt_info
;
5936 for (phi
= phi_nodes (bb
); phi
; phi
= PHI_CHAIN (phi
))
5938 if (vect_print_dump_info (REPORT_DETAILS
))
5940 fprintf (vect_dump
, "------>vectorizing phi: ");
5941 print_generic_expr (vect_dump
, phi
, TDF_SLIM
);
5943 stmt_info
= vinfo_for_stmt (phi
);
5946 if (!STMT_VINFO_RELEVANT_P (stmt_info
)
5947 && !STMT_VINFO_LIVE_P (stmt_info
))
5950 if ((TYPE_VECTOR_SUBPARTS (STMT_VINFO_VECTYPE (stmt_info
))
5951 != (unsigned HOST_WIDE_INT
) vectorization_factor
)
5952 && vect_print_dump_info (REPORT_DETAILS
))
5953 fprintf (vect_dump
, "multiple-types.");
5955 if (STMT_VINFO_DEF_TYPE (stmt_info
) == vect_induction_def
)
5957 if (vect_print_dump_info (REPORT_DETAILS
))
5958 fprintf (vect_dump
, "transform phi.");
5959 vect_transform_stmt (phi
, NULL
, NULL
);
5963 for (si
= bsi_start (bb
); !bsi_end_p (si
);)
5965 tree stmt
= bsi_stmt (si
);
5968 if (vect_print_dump_info (REPORT_DETAILS
))
5970 fprintf (vect_dump
, "------>vectorizing statement: ");
5971 print_generic_expr (vect_dump
, stmt
, TDF_SLIM
);
5973 stmt_info
= vinfo_for_stmt (stmt
);
5974 gcc_assert (stmt_info
);
5975 if (!STMT_VINFO_RELEVANT_P (stmt_info
)
5976 && !STMT_VINFO_LIVE_P (stmt_info
))
5982 gcc_assert (STMT_VINFO_VECTYPE (stmt_info
));
5983 if ((TYPE_VECTOR_SUBPARTS (STMT_VINFO_VECTYPE (stmt_info
))
5984 != (unsigned HOST_WIDE_INT
) vectorization_factor
)
5985 && vect_print_dump_info (REPORT_DETAILS
))
5986 fprintf (vect_dump
, "multiple-types.");
5988 /* -------- vectorize statement ------------ */
5989 if (vect_print_dump_info (REPORT_DETAILS
))
5990 fprintf (vect_dump
, "transform statement.");
5992 strided_store
= false;
5993 is_store
= vect_transform_stmt (stmt
, &si
, &strided_store
);
5997 if (DR_GROUP_FIRST_DR (stmt_info
))
5999 /* Interleaving. If IS_STORE is TRUE, the vectorization of the
6000 interleaving chain was completed - free all the stores in
6002 tree next
= DR_GROUP_FIRST_DR (stmt_info
);
6004 stmt_vec_info next_stmt_info
;
6008 next_si
= bsi_for_stmt (next
);
6009 next_stmt_info
= vinfo_for_stmt (next
);
6010 /* Free the attached stmt_vec_info and remove the stmt. */
6011 ann
= stmt_ann (next
);
6012 tmp
= DR_GROUP_NEXT_DR (next_stmt_info
);
6013 free (next_stmt_info
);
6014 set_stmt_info (ann
, NULL
);
6015 bsi_remove (&next_si
, true);
6018 bsi_remove (&si
, true);
6023 /* Free the attached stmt_vec_info and remove the stmt. */
6024 ann
= stmt_ann (stmt
);
6026 set_stmt_info (ann
, NULL
);
6027 bsi_remove (&si
, true);
6035 slpeel_make_loop_iterate_ntimes (loop
, ratio
);
6037 mark_set_for_renaming (vect_memsyms_to_rename
);
6039 /* The memory tags and pointers in vectorized statements need to
6040 have their SSA forms updated. FIXME, why can't this be delayed
6041 until all the loops have been transformed? */
6042 update_ssa (TODO_update_ssa
);
6044 if (vect_print_dump_info (REPORT_VECTORIZED_LOOPS
))
6045 fprintf (vect_dump
, "LOOP VECTORIZED.");