1 /* Transformation Utilities for Loop Vectorization.
2 Copyright (C) 2003, 2004, 2005, 2006, 2007, 2008 Free Software Foundation, Inc.
3 Contributed by Dorit Naishlos <dorit@il.ibm.com>
5 This file is part of GCC.
7 GCC is free software; you can redistribute it and/or modify it under
8 the terms of the GNU General Public License as published by the Free
9 Software Foundation; either version 3, or (at your option) any later
12 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
13 WARRANTY; without even the implied warranty of MERCHANTABILITY or
14 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
17 You should have received a copy of the GNU General Public License
18 along with GCC; see the file COPYING3. If not see
19 <http://www.gnu.org/licenses/>. */
23 #include "coretypes.h"
29 #include "basic-block.h"
30 #include "diagnostic.h"
31 #include "tree-flow.h"
32 #include "tree-dump.h"
39 #include "tree-data-ref.h"
40 #include "tree-chrec.h"
41 #include "tree-scalar-evolution.h"
42 #include "tree-vectorizer.h"
43 #include "langhooks.h"
44 #include "tree-pass.h"
48 /* Utility functions for the code transformation. */
49 static bool vect_transform_stmt (tree, block_stmt_iterator *, bool *, slp_tree);
50 static tree vect_create_destination_var (tree, tree);
51 static tree vect_create_data_ref_ptr
52 (tree, struct loop*, tree, tree *, tree *, bool, tree, bool *);
53 static tree vect_create_addr_base_for_vector_ref
54 (tree, tree *, tree, struct loop *);
55 static tree vect_get_new_vect_var (tree, enum vect_var_kind, const char *);
56 static tree vect_get_vec_def_for_operand (tree, tree, tree *);
57 static tree vect_init_vector (tree, tree, tree, block_stmt_iterator *);
58 static void vect_finish_stmt_generation
59 (tree stmt, tree vec_stmt, block_stmt_iterator *);
60 static bool vect_is_simple_cond (tree, loop_vec_info);
61 static void vect_create_epilog_for_reduction (tree, tree, enum tree_code, tree);
62 static tree get_initial_def_for_reduction (tree, tree, tree *);
64 /* Utility function dealing with loop peeling (not peeling itself). */
65 static void vect_generate_tmps_on_preheader
66 (loop_vec_info, tree *, tree *, tree *);
67 static tree vect_build_loop_niters (loop_vec_info);
68 static void vect_update_ivs_after_vectorizer (loop_vec_info, tree, edge);
69 static tree vect_gen_niters_for_prolog_loop (loop_vec_info, tree);
70 static void vect_update_init_of_dr (struct data_reference *, tree niters);
71 static void vect_update_inits_of_drs (loop_vec_info, tree);
72 static int vect_min_worthwhile_factor (enum tree_code);
76 cost_for_stmt (tree stmt)
78 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
80 switch (STMT_VINFO_TYPE (stmt_info))
82 case load_vec_info_type:
83 return TARG_SCALAR_LOAD_COST;
84 case store_vec_info_type:
85 return TARG_SCALAR_STORE_COST;
86 case op_vec_info_type:
87 case condition_vec_info_type:
88 case assignment_vec_info_type:
89 case reduc_vec_info_type:
90 case induc_vec_info_type:
91 case type_promotion_vec_info_type:
92 case type_demotion_vec_info_type:
93 case type_conversion_vec_info_type:
94 case call_vec_info_type:
95 return TARG_SCALAR_STMT_COST;
96 case undef_vec_info_type:
103 /* Function vect_estimate_min_profitable_iters
105 Return the number of iterations required for the vector version of the
106 loop to be profitable relative to the cost of the scalar version of the
109 TODO: Take profile info into account before making vectorization
110 decisions, if available. */
113 vect_estimate_min_profitable_iters (loop_vec_info loop_vinfo)
116 int min_profitable_iters;
117 int peel_iters_prologue;
118 int peel_iters_epilogue;
119 int vec_inside_cost = 0;
120 int vec_outside_cost = 0;
121 int scalar_single_iter_cost = 0;
122 int scalar_outside_cost = 0;
123 bool runtime_test = false;
124 int vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo);
125 struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
126 basic_block *bbs = LOOP_VINFO_BBS (loop_vinfo);
127 int nbbs = loop->num_nodes;
128 int byte_misalign = LOOP_PEELING_FOR_ALIGNMENT (loop_vinfo);
129 int peel_guard_costs = 0;
130 int innerloop_iters = 0, factor;
131 VEC (slp_instance, heap) *slp_instances;
132 slp_instance instance;
134 /* Cost model disabled. */
135 if (!flag_vect_cost_model)
137 if (vect_print_dump_info (REPORT_COST))
138 fprintf (vect_dump, "cost model disabled.");
142 /* If the number of iterations is unknown, or the
143 peeling-for-misalignment amount is unknown, we will have to generate
144 a runtime test to test the loop count against the threshold. */
145 if (!LOOP_VINFO_NITERS_KNOWN_P (loop_vinfo)
146 || (byte_misalign < 0))
149 /* Requires loop versioning tests to handle misalignment. */
151 if (VEC_length (tree, LOOP_VINFO_MAY_MISALIGN_STMTS (loop_vinfo)))
153 /* FIXME: Make cost depend on complexity of individual check. */
155 VEC_length (tree, LOOP_VINFO_MAY_MISALIGN_STMTS (loop_vinfo));
156 if (vect_print_dump_info (REPORT_COST))
157 fprintf (vect_dump, "cost model: Adding cost of checks for loop "
158 "versioning to treat misalignment.\n");
161 if (VEC_length (ddr_p, LOOP_VINFO_MAY_ALIAS_DDRS (loop_vinfo)))
163 /* FIXME: Make cost depend on complexity of individual check. */
165 VEC_length (ddr_p, LOOP_VINFO_MAY_ALIAS_DDRS (loop_vinfo));
166 if (vect_print_dump_info (REPORT_COST))
167 fprintf (vect_dump, "cost model: Adding cost of checks for loop "
168 "versioning aliasing.\n");
171 if (VEC_length (tree, LOOP_VINFO_MAY_MISALIGN_STMTS (loop_vinfo))
172 || VEC_length (ddr_p, LOOP_VINFO_MAY_ALIAS_DDRS (loop_vinfo)))
174 vec_outside_cost += TARG_COND_TAKEN_BRANCH_COST;
177 /* Count statements in scalar loop. Using this as scalar cost for a single
180 TODO: Add outer loop support.
182 TODO: Consider assigning different costs to different scalar
187 innerloop_iters = 50; /* FIXME */
189 for (i = 0; i < nbbs; i++)
191 block_stmt_iterator si;
192 basic_block bb = bbs[i];
194 if (bb->loop_father == loop->inner)
195 factor = innerloop_iters;
199 for (si = bsi_start (bb); !bsi_end_p (si); bsi_next (&si))
201 tree stmt = bsi_stmt (si);
202 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
203 /* Skip stmts that are not vectorized inside the loop. */
204 if (!STMT_VINFO_RELEVANT_P (stmt_info)
205 && (!STMT_VINFO_LIVE_P (stmt_info)
206 || STMT_VINFO_DEF_TYPE (stmt_info) != vect_reduction_def))
208 scalar_single_iter_cost += cost_for_stmt (stmt) * factor;
209 vec_inside_cost += STMT_VINFO_INSIDE_OF_LOOP_COST (stmt_info) * factor;
210 /* FIXME: for stmts in the inner-loop in outer-loop vectorization,
211 some of the "outside" costs are generated inside the outer-loop. */
212 vec_outside_cost += STMT_VINFO_OUTSIDE_OF_LOOP_COST (stmt_info);
216 /* Add additional cost for the peeled instructions in prologue and epilogue
219 FORNOW: If we dont know the value of peel_iters for prologue or epilogue
220 at compile-time - we assume it's vf/2 (the worst would be vf-1).
222 TODO: Build an expression that represents peel_iters for prologue and
223 epilogue to be used in a run-time test. */
225 if (byte_misalign < 0)
227 peel_iters_prologue = vf/2;
228 if (vect_print_dump_info (REPORT_COST))
229 fprintf (vect_dump, "cost model: "
230 "prologue peel iters set to vf/2.");
232 /* If peeling for alignment is unknown, loop bound of main loop becomes
234 peel_iters_epilogue = vf/2;
235 if (vect_print_dump_info (REPORT_COST))
236 fprintf (vect_dump, "cost model: "
237 "epilogue peel iters set to vf/2 because "
238 "peeling for alignment is unknown .");
240 /* If peeled iterations are unknown, count a taken branch and a not taken
241 branch per peeled loop. Even if scalar loop iterations are known,
242 vector iterations are not known since peeled prologue iterations are
243 not known. Hence guards remain the same. */
244 peel_guard_costs += 2 * (TARG_COND_TAKEN_BRANCH_COST
245 + TARG_COND_NOT_TAKEN_BRANCH_COST);
252 struct data_reference *dr = LOOP_VINFO_UNALIGNED_DR (loop_vinfo);
253 int element_size = GET_MODE_SIZE (TYPE_MODE (TREE_TYPE (DR_REF (dr))));
254 tree vectype = STMT_VINFO_VECTYPE (vinfo_for_stmt (DR_STMT (dr)));
255 int nelements = TYPE_VECTOR_SUBPARTS (vectype);
257 peel_iters_prologue = nelements - (byte_misalign / element_size);
260 peel_iters_prologue = 0;
262 if (!LOOP_VINFO_NITERS_KNOWN_P (loop_vinfo))
264 peel_iters_epilogue = vf/2;
265 if (vect_print_dump_info (REPORT_COST))
266 fprintf (vect_dump, "cost model: "
267 "epilogue peel iters set to vf/2 because "
268 "loop iterations are unknown .");
270 /* If peeled iterations are known but number of scalar loop
271 iterations are unknown, count a taken branch per peeled loop. */
272 peel_guard_costs += 2 * TARG_COND_TAKEN_BRANCH_COST;
277 int niters = LOOP_VINFO_INT_NITERS (loop_vinfo);
278 peel_iters_prologue = niters < peel_iters_prologue ?
279 niters : peel_iters_prologue;
280 peel_iters_epilogue = (niters - peel_iters_prologue) % vf;
284 vec_outside_cost += (peel_iters_prologue * scalar_single_iter_cost)
285 + (peel_iters_epilogue * scalar_single_iter_cost)
288 /* FORNOW: The scalar outside cost is incremented in one of the
291 1. The vectorizer checks for alignment and aliasing and generates
292 a condition that allows dynamic vectorization. A cost model
293 check is ANDED with the versioning condition. Hence scalar code
294 path now has the added cost of the versioning check.
296 if (cost > th & versioning_check)
299 Hence run-time scalar is incremented by not-taken branch cost.
301 2. The vectorizer then checks if a prologue is required. If the
302 cost model check was not done before during versioning, it has to
303 be done before the prologue check.
306 prologue = scalar_iters
311 if (prologue == num_iters)
314 Hence the run-time scalar cost is incremented by a taken branch,
315 plus a not-taken branch, plus a taken branch cost.
317 3. The vectorizer then checks if an epilogue is required. If the
318 cost model check was not done before during prologue check, it
319 has to be done with the epilogue check.
325 if (prologue == num_iters)
328 if ((cost <= th) | (scalar_iters-prologue-epilogue == 0))
331 Hence the run-time scalar cost should be incremented by 2 taken
334 TODO: The back end may reorder the BBS's differently and reverse
335 conditions/branch directions. Change the stimates below to
336 something more reasonable. */
340 /* Cost model check occurs at versioning. */
341 if (VEC_length (tree, LOOP_VINFO_MAY_MISALIGN_STMTS (loop_vinfo))
342 || VEC_length (ddr_p, LOOP_VINFO_MAY_ALIAS_DDRS (loop_vinfo)))
343 scalar_outside_cost += TARG_COND_NOT_TAKEN_BRANCH_COST;
346 /* Cost model occurs at prologue generation. */
347 if (LOOP_VINFO_NITERS_KNOWN_P (loop_vinfo))
348 scalar_outside_cost += 2 * TARG_COND_TAKEN_BRANCH_COST
349 + TARG_COND_NOT_TAKEN_BRANCH_COST;
350 /* Cost model check occurs at epilogue generation. */
352 scalar_outside_cost += 2 * TARG_COND_TAKEN_BRANCH_COST;
357 slp_instances = LOOP_VINFO_SLP_INSTANCES (loop_vinfo);
358 for (i = 0; VEC_iterate (slp_instance, slp_instances, i, instance); i++)
360 vec_outside_cost += SLP_INSTANCE_OUTSIDE_OF_LOOP_COST (instance);
361 vec_inside_cost += SLP_INSTANCE_INSIDE_OF_LOOP_COST (instance);
364 /* Calculate number of iterations required to make the vector version
365 profitable, relative to the loop bodies only. The following condition
367 SIC * niters + SOC > VIC * ((niters-PL_ITERS-EP_ITERS)/VF) + VOC
369 SIC = scalar iteration cost, VIC = vector iteration cost,
370 VOC = vector outside cost, VF = vectorization factor,
371 PL_ITERS = prologue iterations, EP_ITERS= epilogue iterations
372 SOC = scalar outside cost for run time cost model check. */
374 if ((scalar_single_iter_cost * vf) > vec_inside_cost)
376 if (vec_outside_cost <= 0)
377 min_profitable_iters = 1;
380 min_profitable_iters = ((vec_outside_cost - scalar_outside_cost) * vf
381 - vec_inside_cost * peel_iters_prologue
382 - vec_inside_cost * peel_iters_epilogue)
383 / ((scalar_single_iter_cost * vf)
386 if ((scalar_single_iter_cost * vf * min_profitable_iters)
387 <= ((vec_inside_cost * min_profitable_iters)
388 + ((vec_outside_cost - scalar_outside_cost) * vf)))
389 min_profitable_iters++;
392 /* vector version will never be profitable. */
395 if (vect_print_dump_info (REPORT_COST))
396 fprintf (vect_dump, "cost model: vector iteration cost = %d "
397 "is divisible by scalar iteration cost = %d by a factor "
398 "greater than or equal to the vectorization factor = %d .",
399 vec_inside_cost, scalar_single_iter_cost, vf);
403 if (vect_print_dump_info (REPORT_COST))
405 fprintf (vect_dump, "Cost model analysis: \n");
406 fprintf (vect_dump, " Vector inside of loop cost: %d\n",
408 fprintf (vect_dump, " Vector outside of loop cost: %d\n",
410 fprintf (vect_dump, " Scalar iteration cost: %d\n",
411 scalar_single_iter_cost);
412 fprintf (vect_dump, " Scalar outside cost: %d\n", scalar_outside_cost);
413 fprintf (vect_dump, " prologue iterations: %d\n",
414 peel_iters_prologue);
415 fprintf (vect_dump, " epilogue iterations: %d\n",
416 peel_iters_epilogue);
417 fprintf (vect_dump, " Calculated minimum iters for profitability: %d\n",
418 min_profitable_iters);
421 min_profitable_iters =
422 min_profitable_iters < vf ? vf : min_profitable_iters;
424 /* Because the condition we create is:
425 if (niters <= min_profitable_iters)
426 then skip the vectorized loop. */
427 min_profitable_iters--;
429 if (vect_print_dump_info (REPORT_COST))
430 fprintf (vect_dump, " Profitability threshold = %d\n",
431 min_profitable_iters);
433 return min_profitable_iters;
437 /* TODO: Close dependency between vect_model_*_cost and vectorizable_*
438 functions. Design better to avoid maintenance issues. */
440 /* Function vect_model_reduction_cost.
442 Models cost for a reduction operation, including the vector ops
443 generated within the strip-mine loop, the initial definition before
444 the loop, and the epilogue code that must be generated. */
447 vect_model_reduction_cost (stmt_vec_info stmt_info, enum tree_code reduc_code,
456 enum machine_mode mode;
457 tree operation = GIMPLE_STMT_OPERAND (STMT_VINFO_STMT (stmt_info), 1);
458 int op_type = TREE_CODE_LENGTH (TREE_CODE (operation));
459 loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info);
460 struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
462 /* Cost of reduction op inside loop. */
463 STMT_VINFO_INSIDE_OF_LOOP_COST (stmt_info) += ncopies * TARG_VEC_STMT_COST;
465 reduction_op = TREE_OPERAND (operation, op_type-1);
466 vectype = get_vectype_for_scalar_type (TREE_TYPE (reduction_op));
469 if (vect_print_dump_info (REPORT_COST))
471 fprintf (vect_dump, "unsupported data-type ");
472 print_generic_expr (vect_dump, TREE_TYPE (reduction_op), TDF_SLIM);
477 mode = TYPE_MODE (vectype);
478 orig_stmt = STMT_VINFO_RELATED_STMT (stmt_info);
481 orig_stmt = STMT_VINFO_STMT (stmt_info);
483 code = TREE_CODE (GIMPLE_STMT_OPERAND (orig_stmt, 1));
485 /* Add in cost for initial definition. */
486 outer_cost += TARG_SCALAR_TO_VEC_COST;
488 /* Determine cost of epilogue code.
490 We have a reduction operator that will reduce the vector in one statement.
491 Also requires scalar extract. */
493 if (!nested_in_vect_loop_p (loop, orig_stmt))
495 if (reduc_code < NUM_TREE_CODES)
496 outer_cost += TARG_VEC_STMT_COST + TARG_VEC_TO_SCALAR_COST;
499 int vec_size_in_bits = tree_low_cst (TYPE_SIZE (vectype), 1);
501 TYPE_SIZE (TREE_TYPE ( GIMPLE_STMT_OPERAND (orig_stmt, 0)));
502 int element_bitsize = tree_low_cst (bitsize, 1);
503 int nelements = vec_size_in_bits / element_bitsize;
505 optab = optab_for_tree_code (code, vectype);
507 /* We have a whole vector shift available. */
508 if (VECTOR_MODE_P (mode)
509 && optab_handler (optab, mode)->insn_code != CODE_FOR_nothing
510 && optab_handler (vec_shr_optab, mode)->insn_code != CODE_FOR_nothing)
511 /* Final reduction via vector shifts and the reduction operator. Also
512 requires scalar extract. */
513 outer_cost += ((exact_log2(nelements) * 2) * TARG_VEC_STMT_COST
514 + TARG_VEC_TO_SCALAR_COST);
516 /* Use extracts and reduction op for final reduction. For N elements,
517 we have N extracts and N-1 reduction ops. */
518 outer_cost += ((nelements + nelements - 1) * TARG_VEC_STMT_COST);
522 STMT_VINFO_OUTSIDE_OF_LOOP_COST (stmt_info) = outer_cost;
524 if (vect_print_dump_info (REPORT_COST))
525 fprintf (vect_dump, "vect_model_reduction_cost: inside_cost = %d, "
526 "outside_cost = %d .", STMT_VINFO_INSIDE_OF_LOOP_COST (stmt_info),
527 STMT_VINFO_OUTSIDE_OF_LOOP_COST (stmt_info));
533 /* Function vect_model_induction_cost.
535 Models cost for induction operations. */
538 vect_model_induction_cost (stmt_vec_info stmt_info, int ncopies)
540 /* loop cost for vec_loop. */
541 STMT_VINFO_INSIDE_OF_LOOP_COST (stmt_info) = ncopies * TARG_VEC_STMT_COST;
542 /* prologue cost for vec_init and vec_step. */
543 STMT_VINFO_OUTSIDE_OF_LOOP_COST (stmt_info) = 2 * TARG_SCALAR_TO_VEC_COST;
545 if (vect_print_dump_info (REPORT_COST))
546 fprintf (vect_dump, "vect_model_induction_cost: inside_cost = %d, "
547 "outside_cost = %d .", STMT_VINFO_INSIDE_OF_LOOP_COST (stmt_info),
548 STMT_VINFO_OUTSIDE_OF_LOOP_COST (stmt_info));
552 /* Function vect_model_simple_cost.
554 Models cost for simple operations, i.e. those that only emit ncopies of a
555 single op. Right now, this does not account for multiple insns that could
556 be generated for the single vector op. We will handle that shortly. */
559 vect_model_simple_cost (stmt_vec_info stmt_info, int ncopies,
560 enum vect_def_type *dt, slp_tree slp_node)
563 int inside_cost = 0, outside_cost = 0;
565 inside_cost = ncopies * TARG_VEC_STMT_COST;
567 /* FORNOW: Assuming maximum 2 args per stmts. */
568 for (i = 0; i < 2; i++)
570 if (dt[i] == vect_constant_def || dt[i] == vect_invariant_def)
571 outside_cost += TARG_SCALAR_TO_VEC_COST;
574 if (vect_print_dump_info (REPORT_COST))
575 fprintf (vect_dump, "vect_model_simple_cost: inside_cost = %d, "
576 "outside_cost = %d .", inside_cost, outside_cost);
578 /* Set the costs either in STMT_INFO or SLP_NODE (if exists). */
579 stmt_vinfo_set_inside_of_loop_cost (stmt_info, slp_node, inside_cost);
580 stmt_vinfo_set_outside_of_loop_cost (stmt_info, slp_node, outside_cost);
584 /* Function vect_cost_strided_group_size
586 For strided load or store, return the group_size only if it is the first
587 load or store of a group, else return 1. This ensures that group size is
588 only returned once per group. */
591 vect_cost_strided_group_size (stmt_vec_info stmt_info)
593 tree first_stmt = DR_GROUP_FIRST_DR (stmt_info);
595 if (first_stmt == STMT_VINFO_STMT (stmt_info))
596 return DR_GROUP_SIZE (stmt_info);
602 /* Function vect_model_store_cost
604 Models cost for stores. In the case of strided accesses, one access
605 has the overhead of the strided access attributed to it. */
608 vect_model_store_cost (stmt_vec_info stmt_info, int ncopies,
609 enum vect_def_type dt, slp_tree slp_node)
612 int inside_cost = 0, outside_cost = 0;
614 if (dt == vect_constant_def || dt == vect_invariant_def)
615 outside_cost = TARG_SCALAR_TO_VEC_COST;
617 /* Strided access? */
618 if (DR_GROUP_FIRST_DR (stmt_info))
619 group_size = vect_cost_strided_group_size (stmt_info);
620 /* Not a strided access. */
624 /* Is this an access in a group of stores, which provide strided access?
625 If so, add in the cost of the permutes. */
628 /* Uses a high and low interleave operation for each needed permute. */
629 inside_cost = ncopies * exact_log2(group_size) * group_size
630 * TARG_VEC_STMT_COST;
632 if (vect_print_dump_info (REPORT_COST))
633 fprintf (vect_dump, "vect_model_store_cost: strided group_size = %d .",
638 /* Costs of the stores. */
639 inside_cost += ncopies * TARG_VEC_STORE_COST;
641 if (vect_print_dump_info (REPORT_COST))
642 fprintf (vect_dump, "vect_model_store_cost: inside_cost = %d, "
643 "outside_cost = %d .", inside_cost, outside_cost);
645 /* Set the costs either in STMT_INFO or SLP_NODE (if exists). */
646 stmt_vinfo_set_inside_of_loop_cost (stmt_info, slp_node, inside_cost);
647 stmt_vinfo_set_outside_of_loop_cost (stmt_info, slp_node, outside_cost);
651 /* Function vect_model_load_cost
653 Models cost for loads. In the case of strided accesses, the last access
654 has the overhead of the strided access attributed to it. Since unaligned
655 accesses are supported for loads, we also account for the costs of the
656 access scheme chosen. */
659 vect_model_load_cost (stmt_vec_info stmt_info, int ncopies, slp_tree slp_node)
663 int alignment_support_cheme;
665 struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info), *first_dr;
666 int inside_cost = 0, outside_cost = 0;
668 /* Strided accesses? */
669 first_stmt = DR_GROUP_FIRST_DR (stmt_info);
670 if (first_stmt && !slp_node)
672 group_size = vect_cost_strided_group_size (stmt_info);
673 first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt));
675 /* Not a strided access. */
682 alignment_support_cheme = vect_supportable_dr_alignment (first_dr);
684 /* Is this an access in a group of loads providing strided access?
685 If so, add in the cost of the permutes. */
688 /* Uses an even and odd extract operations for each needed permute. */
689 inside_cost = ncopies * exact_log2(group_size) * group_size
690 * TARG_VEC_STMT_COST;
692 if (vect_print_dump_info (REPORT_COST))
693 fprintf (vect_dump, "vect_model_load_cost: strided group_size = %d .",
698 /* The loads themselves. */
699 switch (alignment_support_cheme)
703 inside_cost += ncopies * TARG_VEC_LOAD_COST;
705 if (vect_print_dump_info (REPORT_COST))
706 fprintf (vect_dump, "vect_model_load_cost: aligned.");
710 case dr_unaligned_supported:
712 /* Here, we assign an additional cost for the unaligned load. */
713 inside_cost += ncopies * TARG_VEC_UNALIGNED_LOAD_COST;
715 if (vect_print_dump_info (REPORT_COST))
716 fprintf (vect_dump, "vect_model_load_cost: unaligned supported by "
721 case dr_explicit_realign:
723 inside_cost += ncopies * (2*TARG_VEC_LOAD_COST + TARG_VEC_STMT_COST);
725 /* FIXME: If the misalignment remains fixed across the iterations of
726 the containing loop, the following cost should be added to the
728 if (targetm.vectorize.builtin_mask_for_load)
729 inside_cost += TARG_VEC_STMT_COST;
733 case dr_explicit_realign_optimized:
735 if (vect_print_dump_info (REPORT_COST))
736 fprintf (vect_dump, "vect_model_load_cost: unaligned software "
739 /* Unaligned software pipeline has a load of an address, an initial
740 load, and possibly a mask operation to "prime" the loop. However,
741 if this is an access in a group of loads, which provide strided
742 access, then the above cost should only be considered for one
743 access in the group. Inside the loop, there is a load op
744 and a realignment op. */
746 if ((!DR_GROUP_FIRST_DR (stmt_info)) || group_size > 1 || slp_node)
748 outside_cost = 2*TARG_VEC_STMT_COST;
749 if (targetm.vectorize.builtin_mask_for_load)
750 outside_cost += TARG_VEC_STMT_COST;
753 inside_cost += ncopies * (TARG_VEC_LOAD_COST + TARG_VEC_STMT_COST);
762 if (vect_print_dump_info (REPORT_COST))
763 fprintf (vect_dump, "vect_model_load_cost: inside_cost = %d, "
764 "outside_cost = %d .", inside_cost, outside_cost);
766 /* Set the costs either in STMT_INFO or SLP_NODE (if exists). */
767 stmt_vinfo_set_inside_of_loop_cost (stmt_info, slp_node, inside_cost);
768 stmt_vinfo_set_outside_of_loop_cost (stmt_info, slp_node, outside_cost);
772 /* Function vect_get_new_vect_var.
774 Returns a name for a new variable. The current naming scheme appends the
775 prefix "vect_" or "vect_p" (depending on the value of VAR_KIND) to
776 the name of vectorizer generated variables, and appends that to NAME if
780 vect_get_new_vect_var (tree type, enum vect_var_kind var_kind, const char *name)
787 case vect_simple_var:
790 case vect_scalar_var:
793 case vect_pointer_var:
802 char* tmp = concat (prefix, name, NULL);
803 new_vect_var = create_tmp_var (type, tmp);
807 new_vect_var = create_tmp_var (type, prefix);
809 /* Mark vector typed variable as a gimple register variable. */
810 if (TREE_CODE (type) == VECTOR_TYPE)
811 DECL_GIMPLE_REG_P (new_vect_var) = true;
817 /* Function vect_create_addr_base_for_vector_ref.
819 Create an expression that computes the address of the first memory location
820 that will be accessed for a data reference.
823 STMT: The statement containing the data reference.
824 NEW_STMT_LIST: Must be initialized to NULL_TREE or a statement list.
825 OFFSET: Optional. If supplied, it is be added to the initial address.
826 LOOP: Specify relative to which loop-nest should the address be computed.
827 For example, when the dataref is in an inner-loop nested in an
828 outer-loop that is now being vectorized, LOOP can be either the
829 outer-loop, or the inner-loop. The first memory location accessed
830 by the following dataref ('in' points to short):
837 if LOOP=i_loop: &in (relative to i_loop)
838 if LOOP=j_loop: &in+i*2B (relative to j_loop)
841 1. Return an SSA_NAME whose value is the address of the memory location of
842 the first vector of the data reference.
843 2. If new_stmt_list is not NULL_TREE after return then the caller must insert
844 these statement(s) which define the returned SSA_NAME.
846 FORNOW: We are only handling array accesses with step 1. */
849 vect_create_addr_base_for_vector_ref (tree stmt,
854 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
855 struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info);
856 struct loop *containing_loop = (bb_for_stmt (stmt))->loop_father;
857 tree data_ref_base = unshare_expr (DR_BASE_ADDRESS (dr));
859 tree data_ref_base_var;
862 tree addr_base, addr_expr;
864 tree base_offset = unshare_expr (DR_OFFSET (dr));
865 tree init = unshare_expr (DR_INIT (dr));
866 tree vect_ptr_type, addr_expr2;
867 tree step = TYPE_SIZE_UNIT (TREE_TYPE (DR_REF (dr)));
870 if (loop != containing_loop)
872 loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info);
873 struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
875 gcc_assert (nested_in_vect_loop_p (loop, stmt));
877 data_ref_base = unshare_expr (STMT_VINFO_DR_BASE_ADDRESS (stmt_info));
878 base_offset = unshare_expr (STMT_VINFO_DR_OFFSET (stmt_info));
879 init = unshare_expr (STMT_VINFO_DR_INIT (stmt_info));
882 /* Create data_ref_base */
883 base_name = build_fold_indirect_ref (data_ref_base);
884 data_ref_base_var = create_tmp_var (TREE_TYPE (data_ref_base), "batmp");
885 add_referenced_var (data_ref_base_var);
886 data_ref_base = force_gimple_operand (data_ref_base, &new_base_stmt,
887 true, data_ref_base_var);
888 append_to_statement_list_force(new_base_stmt, new_stmt_list);
890 /* Create base_offset */
891 base_offset = size_binop (PLUS_EXPR, base_offset, init);
892 base_offset = fold_convert (sizetype, base_offset);
893 dest = create_tmp_var (TREE_TYPE (base_offset), "base_off");
894 add_referenced_var (dest);
895 base_offset = force_gimple_operand (base_offset, &new_stmt, true, dest);
896 append_to_statement_list_force (new_stmt, new_stmt_list);
900 tree tmp = create_tmp_var (sizetype, "offset");
902 add_referenced_var (tmp);
903 offset = fold_build2 (MULT_EXPR, TREE_TYPE (offset), offset, step);
904 base_offset = fold_build2 (PLUS_EXPR, TREE_TYPE (base_offset),
905 base_offset, offset);
906 base_offset = force_gimple_operand (base_offset, &new_stmt, false, tmp);
907 append_to_statement_list_force (new_stmt, new_stmt_list);
910 /* base + base_offset */
911 addr_base = fold_build2 (POINTER_PLUS_EXPR, TREE_TYPE (data_ref_base),
912 data_ref_base, base_offset);
914 vect_ptr_type = build_pointer_type (STMT_VINFO_VECTYPE (stmt_info));
916 /* addr_expr = addr_base */
917 addr_expr = vect_get_new_vect_var (vect_ptr_type, vect_pointer_var,
918 get_name (base_name));
919 add_referenced_var (addr_expr);
920 vec_stmt = fold_convert (vect_ptr_type, addr_base);
921 addr_expr2 = vect_get_new_vect_var (vect_ptr_type, vect_pointer_var,
922 get_name (base_name));
923 add_referenced_var (addr_expr2);
924 vec_stmt = force_gimple_operand (vec_stmt, &new_stmt, false, addr_expr2);
925 append_to_statement_list_force (new_stmt, new_stmt_list);
927 if (vect_print_dump_info (REPORT_DETAILS))
929 fprintf (vect_dump, "created ");
930 print_generic_expr (vect_dump, vec_stmt, TDF_SLIM);
936 /* Function vect_create_data_ref_ptr.
938 Create a new pointer to vector type (vp), that points to the first location
939 accessed in the loop by STMT, along with the def-use update chain to
940 appropriately advance the pointer through the loop iterations. Also set
941 aliasing information for the pointer. This vector pointer is used by the
942 callers to this function to create a memory reference expression for vector
946 1. STMT: a stmt that references memory. Expected to be of the form
947 GIMPLE_MODIFY_STMT <name, data-ref> or
948 GIMPLE_MODIFY_STMT <data-ref, name>.
949 2. AT_LOOP: the loop where the vector memref is to be created.
950 3. OFFSET (optional): an offset to be added to the initial address accessed
951 by the data-ref in STMT.
952 4. ONLY_INIT: indicate if vp is to be updated in the loop, or remain
953 pointing to the initial address.
954 5. TYPE: if not NULL indicates the required type of the data-ref
957 1. Declare a new ptr to vector_type, and have it point to the base of the
958 data reference (initial addressed accessed by the data reference).
959 For example, for vector of type V8HI, the following code is generated:
962 vp = (v8hi *)initial_address;
964 if OFFSET is not supplied:
965 initial_address = &a[init];
966 if OFFSET is supplied:
967 initial_address = &a[init + OFFSET];
969 Return the initial_address in INITIAL_ADDRESS.
971 2. If ONLY_INIT is true, just return the initial pointer. Otherwise, also
972 update the pointer in each iteration of the loop.
974 Return the increment stmt that updates the pointer in PTR_INCR.
976 3. Set INV_P to true if the access pattern of the data reference in the
977 vectorized loop is invariant. Set it to false otherwise.
979 4. Return the pointer. */
982 vect_create_data_ref_ptr (tree stmt, struct loop *at_loop,
983 tree offset, tree *initial_address, tree *ptr_incr,
984 bool only_init, tree type, bool *inv_p)
987 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
988 loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info);
989 struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
990 bool nested_in_vect_loop = nested_in_vect_loop_p (loop, stmt);
991 struct loop *containing_loop = (bb_for_stmt (stmt))->loop_father;
992 tree vectype = STMT_VINFO_VECTYPE (stmt_info);
998 tree new_stmt_list = NULL_TREE;
1002 struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info);
1004 block_stmt_iterator incr_bsi;
1006 tree indx_before_incr, indx_after_incr;
1010 /* Check the step (evolution) of the load in LOOP, and record
1011 whether it's invariant. */
1012 if (nested_in_vect_loop)
1013 step = STMT_VINFO_DR_STEP (stmt_info);
1015 step = DR_STEP (STMT_VINFO_DATA_REF (stmt_info));
1017 if (tree_int_cst_compare (step, size_zero_node) == 0)
1022 /* Create an expression for the first address accessed by this load
1024 base_name = build_fold_indirect_ref (unshare_expr (DR_BASE_ADDRESS (dr)));
1026 if (vect_print_dump_info (REPORT_DETAILS))
1028 tree data_ref_base = base_name;
1029 fprintf (vect_dump, "create vector-pointer variable to type: ");
1030 print_generic_expr (vect_dump, vectype, TDF_SLIM);
1031 if (TREE_CODE (data_ref_base) == VAR_DECL)
1032 fprintf (vect_dump, " vectorizing a one dimensional array ref: ");
1033 else if (TREE_CODE (data_ref_base) == ARRAY_REF)
1034 fprintf (vect_dump, " vectorizing a multidimensional array ref: ");
1035 else if (TREE_CODE (data_ref_base) == COMPONENT_REF)
1036 fprintf (vect_dump, " vectorizing a record based array ref: ");
1037 else if (TREE_CODE (data_ref_base) == SSA_NAME)
1038 fprintf (vect_dump, " vectorizing a pointer ref: ");
1039 print_generic_expr (vect_dump, base_name, TDF_SLIM);
1042 /** (1) Create the new vector-pointer variable: **/
1044 vect_ptr_type = build_pointer_type (type);
1046 vect_ptr_type = build_pointer_type (vectype);
1047 vect_ptr = vect_get_new_vect_var (vect_ptr_type, vect_pointer_var,
1048 get_name (base_name));
1049 add_referenced_var (vect_ptr);
1051 /** (2) Add aliasing information to the new vector-pointer:
1052 (The points-to info (DR_PTR_INFO) may be defined later.) **/
1054 tag = DR_SYMBOL_TAG (dr);
1057 /* If tag is a variable (and NOT_A_TAG) than a new symbol memory
1058 tag must be created with tag added to its may alias list. */
1060 new_type_alias (vect_ptr, tag, DR_REF (dr));
1062 set_symbol_mem_tag (vect_ptr, tag);
1064 var_ann (vect_ptr)->subvars = DR_SUBVARS (dr);
1066 /** Note: If the dataref is in an inner-loop nested in LOOP, and we are
1067 vectorizing LOOP (i.e. outer-loop vectorization), we need to create two
1068 def-use update cycles for the pointer: One relative to the outer-loop
1069 (LOOP), which is what steps (3) and (4) below do. The other is relative
1070 to the inner-loop (which is the inner-most loop containing the dataref),
1071 and this is done be step (5) below.
1073 When vectorizing inner-most loops, the vectorized loop (LOOP) is also the
1074 inner-most loop, and so steps (3),(4) work the same, and step (5) is
1075 redundant. Steps (3),(4) create the following:
1078 LOOP: vp1 = phi(vp0,vp2)
1084 If there is an inner-loop nested in loop, then step (5) will also be
1085 applied, and an additional update in the inner-loop will be created:
1088 LOOP: vp1 = phi(vp0,vp2)
1090 inner: vp3 = phi(vp1,vp4)
1091 vp4 = vp3 + inner_step
1097 /** (3) Calculate the initial address the vector-pointer, and set
1098 the vector-pointer to point to it before the loop: **/
1100 /* Create: (&(base[init_val+offset]) in the loop preheader. */
1102 new_temp = vect_create_addr_base_for_vector_ref (stmt, &new_stmt_list,
1104 pe = loop_preheader_edge (loop);
1105 new_bb = bsi_insert_on_edge_immediate (pe, new_stmt_list);
1106 gcc_assert (!new_bb);
1107 *initial_address = new_temp;
1109 /* Create: p = (vectype *) initial_base */
1110 vec_stmt = fold_convert (vect_ptr_type, new_temp);
1111 vec_stmt = build_gimple_modify_stmt (vect_ptr, vec_stmt);
1112 vect_ptr_init = make_ssa_name (vect_ptr, vec_stmt);
1113 GIMPLE_STMT_OPERAND (vec_stmt, 0) = vect_ptr_init;
1114 new_bb = bsi_insert_on_edge_immediate (pe, vec_stmt);
1115 gcc_assert (!new_bb);
1118 /** (4) Handle the updating of the vector-pointer inside the loop.
1119 This is needed when ONLY_INIT is false, and also when AT_LOOP
1120 is the inner-loop nested in LOOP (during outer-loop vectorization).
1123 if (only_init && at_loop == loop) /* No update in loop is required. */
1125 /* Copy the points-to information if it exists. */
1126 if (DR_PTR_INFO (dr))
1127 duplicate_ssa_name_ptr_info (vect_ptr_init, DR_PTR_INFO (dr));
1128 vptr = vect_ptr_init;
1132 /* The step of the vector pointer is the Vector Size. */
1133 tree step = TYPE_SIZE_UNIT (vectype);
1134 /* One exception to the above is when the scalar step of the load in
1135 LOOP is zero. In this case the step here is also zero. */
1137 step = size_zero_node;
1139 standard_iv_increment_position (loop, &incr_bsi, &insert_after);
1141 create_iv (vect_ptr_init,
1142 fold_convert (vect_ptr_type, step),
1143 NULL_TREE, loop, &incr_bsi, insert_after,
1144 &indx_before_incr, &indx_after_incr);
1145 incr = bsi_stmt (incr_bsi);
1146 set_stmt_info (stmt_ann (incr),
1147 new_stmt_vec_info (incr, loop_vinfo));
1149 /* Copy the points-to information if it exists. */
1150 if (DR_PTR_INFO (dr))
1152 duplicate_ssa_name_ptr_info (indx_before_incr, DR_PTR_INFO (dr));
1153 duplicate_ssa_name_ptr_info (indx_after_incr, DR_PTR_INFO (dr));
1155 merge_alias_info (vect_ptr_init, indx_before_incr);
1156 merge_alias_info (vect_ptr_init, indx_after_incr);
1160 vptr = indx_before_incr;
1163 if (!nested_in_vect_loop || only_init)
1167 /** (5) Handle the updating of the vector-pointer inside the inner-loop
1168 nested in LOOP, if exists: **/
1170 gcc_assert (nested_in_vect_loop);
1173 standard_iv_increment_position (containing_loop, &incr_bsi,
1175 create_iv (vptr, fold_convert (vect_ptr_type, DR_STEP (dr)), NULL_TREE,
1176 containing_loop, &incr_bsi, insert_after, &indx_before_incr,
1178 incr = bsi_stmt (incr_bsi);
1179 set_stmt_info (stmt_ann (incr), new_stmt_vec_info (incr, loop_vinfo));
1181 /* Copy the points-to information if it exists. */
1182 if (DR_PTR_INFO (dr))
1184 duplicate_ssa_name_ptr_info (indx_before_incr, DR_PTR_INFO (dr));
1185 duplicate_ssa_name_ptr_info (indx_after_incr, DR_PTR_INFO (dr));
1187 merge_alias_info (vect_ptr_init, indx_before_incr);
1188 merge_alias_info (vect_ptr_init, indx_after_incr);
1192 return indx_before_incr;
1199 /* Function bump_vector_ptr
1201 Increment a pointer (to a vector type) by vector-size. If requested,
1202 i.e. if PTR-INCR is given, then also connect the new increment stmt
1203 to the existing def-use update-chain of the pointer, by modifying
1204 the PTR_INCR as illustrated below:
1206 The pointer def-use update-chain before this function:
1207 DATAREF_PTR = phi (p_0, p_2)
1209 PTR_INCR: p_2 = DATAREF_PTR + step
1211 The pointer def-use update-chain after this function:
1212 DATAREF_PTR = phi (p_0, p_2)
1214 NEW_DATAREF_PTR = DATAREF_PTR + BUMP
1216 PTR_INCR: p_2 = NEW_DATAREF_PTR + step
1219 DATAREF_PTR - ssa_name of a pointer (to vector type) that is being updated
1221 PTR_INCR - optional. The stmt that updates the pointer in each iteration of
1222 the loop. The increment amount across iterations is expected
1224 BSI - location where the new update stmt is to be placed.
1225 STMT - the original scalar memory-access stmt that is being vectorized.
1226 BUMP - optional. The offset by which to bump the pointer. If not given,
1227 the offset is assumed to be vector_size.
1229 Output: Return NEW_DATAREF_PTR as illustrated above.
1234 bump_vector_ptr (tree dataref_ptr, tree ptr_incr, block_stmt_iterator *bsi,
1235 tree stmt, tree bump)
1237 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
1238 struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info);
1239 tree vectype = STMT_VINFO_VECTYPE (stmt_info);
1240 tree vptr_type = TREE_TYPE (dataref_ptr);
1241 tree ptr_var = SSA_NAME_VAR (dataref_ptr);
1242 tree update = TYPE_SIZE_UNIT (vectype);
1245 use_operand_p use_p;
1246 tree new_dataref_ptr;
1251 incr_stmt = build_gimple_modify_stmt (ptr_var,
1252 build2 (POINTER_PLUS_EXPR, vptr_type,
1253 dataref_ptr, update));
1254 new_dataref_ptr = make_ssa_name (ptr_var, incr_stmt);
1255 GIMPLE_STMT_OPERAND (incr_stmt, 0) = new_dataref_ptr;
1256 vect_finish_stmt_generation (stmt, incr_stmt, bsi);
1258 /* Copy the points-to information if it exists. */
1259 if (DR_PTR_INFO (dr))
1260 duplicate_ssa_name_ptr_info (new_dataref_ptr, DR_PTR_INFO (dr));
1261 merge_alias_info (new_dataref_ptr, dataref_ptr);
1264 return new_dataref_ptr;
1266 /* Update the vector-pointer's cross-iteration increment. */
1267 FOR_EACH_SSA_USE_OPERAND (use_p, ptr_incr, iter, SSA_OP_USE)
1269 tree use = USE_FROM_PTR (use_p);
1271 if (use == dataref_ptr)
1272 SET_USE (use_p, new_dataref_ptr);
1274 gcc_assert (tree_int_cst_compare (use, update) == 0);
1277 return new_dataref_ptr;
1281 /* Function vect_create_destination_var.
1283 Create a new temporary of type VECTYPE. */
1286 vect_create_destination_var (tree scalar_dest, tree vectype)
1289 const char *new_name;
1291 enum vect_var_kind kind;
1293 kind = vectype ? vect_simple_var : vect_scalar_var;
1294 type = vectype ? vectype : TREE_TYPE (scalar_dest);
1296 gcc_assert (TREE_CODE (scalar_dest) == SSA_NAME);
1298 new_name = get_name (scalar_dest);
1301 vec_dest = vect_get_new_vect_var (type, kind, new_name);
1302 add_referenced_var (vec_dest);
1308 /* Function vect_init_vector.
1310 Insert a new stmt (INIT_STMT) that initializes a new vector variable with
1311 the vector elements of VECTOR_VAR. Place the initialization at BSI if it
1312 is not NULL. Otherwise, place the initialization at the loop preheader.
1313 Return the DEF of INIT_STMT.
1314 It will be used in the vectorization of STMT. */
1317 vect_init_vector (tree stmt, tree vector_var, tree vector_type,
1318 block_stmt_iterator *bsi)
1320 stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt);
1328 new_var = vect_get_new_vect_var (vector_type, vect_simple_var, "cst_");
1329 add_referenced_var (new_var);
1330 init_stmt = build_gimple_modify_stmt (new_var, vector_var);
1331 new_temp = make_ssa_name (new_var, init_stmt);
1332 GIMPLE_STMT_OPERAND (init_stmt, 0) = new_temp;
1335 vect_finish_stmt_generation (stmt, init_stmt, bsi);
1338 loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_vinfo);
1339 struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
1341 if (nested_in_vect_loop_p (loop, stmt))
1343 pe = loop_preheader_edge (loop);
1344 new_bb = bsi_insert_on_edge_immediate (pe, init_stmt);
1345 gcc_assert (!new_bb);
1348 if (vect_print_dump_info (REPORT_DETAILS))
1350 fprintf (vect_dump, "created new init_stmt: ");
1351 print_generic_expr (vect_dump, init_stmt, TDF_SLIM);
1354 vec_oprnd = GIMPLE_STMT_OPERAND (init_stmt, 0);
1359 /* For constant and loop invariant defs of SLP_NODE this function returns
1360 (vector) defs (VEC_OPRNDS) that will be used in the vectorized stmts.
1361 OP_NUM determines if we gather defs for operand 0 or operand 1 of the scalar
1365 vect_get_constant_vectors (slp_tree slp_node, VEC(tree,heap) **vec_oprnds,
1366 unsigned int op_num)
1368 VEC (tree, heap) *stmts = SLP_TREE_SCALAR_STMTS (slp_node);
1369 tree stmt = VEC_index (tree, stmts, 0);
1370 stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt);
1371 tree vectype = STMT_VINFO_VECTYPE (stmt_vinfo);
1372 int nunits = TYPE_VECTOR_SUBPARTS (vectype);
1375 int j, number_of_places_left_in_vector;
1377 tree op, vop, operation;
1378 int group_size = VEC_length (tree, stmts);
1379 unsigned int vec_num, i;
1380 int number_of_copies = 1;
1381 bool is_store = false;
1382 unsigned int number_of_vectors = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node);
1383 VEC (tree, heap) *voprnds = VEC_alloc (tree, heap, number_of_vectors);
1386 if (STMT_VINFO_DATA_REF (stmt_vinfo))
1389 /* NUMBER_OF_COPIES is the number of times we need to use the same values in
1390 created vectors. It is greater than 1 if unrolling is performed.
1392 For example, we have two scalar operands, s1 and s2 (e.g., group of
1393 strided accesses of size two), while NUINTS is four (i.e., four scalars
1394 of this type can be packed in a vector). The output vector will contain
1395 two copies of each scalar operand: {s1, s2, s1, s2}. (NUMBER_OF_COPIES
1398 If GROUP_SIZE > NUNITS, the scalars will be split into several vectors
1399 containing the operands.
1401 For example, NUINTS is four as before, and the group size is 8
1402 (s1, s2, ..., s8). We will create two vectors {s1, s2, s3, s4} and
1403 {s5, s6, s7, s8}. */
1405 number_of_copies = least_common_multiple (nunits, group_size) / group_size;
1407 number_of_places_left_in_vector = nunits;
1409 for (j = 0; j < number_of_copies; j++)
1411 for (i = group_size - 1; VEC_iterate (tree, stmts, i, stmt); i--)
1413 operation = GIMPLE_STMT_OPERAND (stmt, 1);
1417 op = TREE_OPERAND (operation, op_num);
1418 if (!CONSTANT_CLASS_P (op))
1421 /* Create 'vect_ = {op0,op1,...,opn}'. */
1422 t = tree_cons (NULL_TREE, op, t);
1424 number_of_places_left_in_vector--;
1426 if (number_of_places_left_in_vector == 0)
1428 number_of_places_left_in_vector = nunits;
1430 vector_type = get_vectype_for_scalar_type (TREE_TYPE (op));
1431 gcc_assert (vector_type);
1433 vec_cst = build_vector (vector_type, t);
1435 vec_cst = build_constructor_from_list (vector_type, t);
1437 VEC_quick_push (tree, voprnds,
1438 vect_init_vector (stmt, vec_cst, vector_type,
1445 /* Since the vectors are created in the reverse order, we should invert
1447 vec_num = VEC_length (tree, voprnds);
1448 for (j = vec_num - 1; j >= 0; j--)
1450 vop = VEC_index (tree, voprnds, j);
1451 VEC_quick_push (tree, *vec_oprnds, vop);
1454 VEC_free (tree, heap, voprnds);
1456 /* In case that VF is greater than the unrolling factor needed for the SLP
1457 group of stmts, NUMBER_OF_VECTORS to be created is greater than
1458 NUMBER_OF_SCALARS/NUNITS or NUNITS/NUMBER_OF_SCALARS, and hence we have
1459 to replicate the vectors. */
1460 while (number_of_vectors > VEC_length (tree, *vec_oprnds))
1462 for (i = 0; VEC_iterate (tree, *vec_oprnds, i, vop) && i < vec_num; i++)
1463 VEC_quick_push (tree, *vec_oprnds, vop);
1468 /* Get vectorized definitions from SLP_NODE that contains corresponding
1469 vectorized def-stmts. */
1472 vect_get_slp_vect_defs (slp_tree slp_node, VEC (tree,heap) **vec_oprnds)
1478 gcc_assert (SLP_TREE_VEC_STMTS (slp_node));
1481 VEC_iterate (tree, SLP_TREE_VEC_STMTS (slp_node), i, vec_def_stmt);
1484 gcc_assert (vec_def_stmt);
1485 vec_oprnd = GIMPLE_STMT_OPERAND (vec_def_stmt, 0);
1486 VEC_quick_push (tree, *vec_oprnds, vec_oprnd);
1491 /* Get vectorized definitions for SLP_NODE.
1492 If the scalar definitions are loop invariants or constants, collect them and
1493 call vect_get_constant_vectors() to create vector stmts.
1494 Otherwise, the def-stmts must be already vectorized and the vectorized stmts
1495 must be stored in the LEFT/RIGHT node of SLP_NODE, and we call
1496 vect_get_slp_vect_defs() to retrieve them.
1497 If VEC_OPRNDS1 is NULL, don't get vector defs for the second operand (from
1498 the right node. This is used when the second operand must remain scalar. */
1501 vect_get_slp_defs (slp_tree slp_node, VEC (tree,heap) **vec_oprnds0,
1502 VEC (tree,heap) **vec_oprnds1)
1504 tree operation, first_stmt;
1506 /* Allocate memory for vectorized defs. */
1507 *vec_oprnds0 = VEC_alloc (tree, heap,
1508 SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node));
1510 /* SLP_NODE corresponds either to a group of stores or to a group of
1511 unary/binary operations. We don't call this function for loads. */
1512 if (SLP_TREE_LEFT (slp_node))
1513 /* The defs are already vectorized. */
1514 vect_get_slp_vect_defs (SLP_TREE_LEFT (slp_node), vec_oprnds0);
1516 /* Build vectors from scalar defs. */
1517 vect_get_constant_vectors (slp_node, vec_oprnds0, 0);
1519 first_stmt = VEC_index (tree, SLP_TREE_SCALAR_STMTS (slp_node), 0);
1520 if (STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt)))
1521 /* Since we don't call this function with loads, this is a group of
1525 operation = GIMPLE_STMT_OPERAND (first_stmt, 1);
1526 if (TREE_OPERAND_LENGTH (operation) == unary_op || !vec_oprnds1)
1529 *vec_oprnds1 = VEC_alloc (tree, heap,
1530 SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node));
1532 if (SLP_TREE_RIGHT (slp_node))
1533 /* The defs are already vectorized. */
1534 vect_get_slp_vect_defs (SLP_TREE_RIGHT (slp_node), vec_oprnds1);
1536 /* Build vectors from scalar defs. */
1537 vect_get_constant_vectors (slp_node, vec_oprnds1, 1);
1541 /* Function get_initial_def_for_induction
1544 STMT - a stmt that performs an induction operation in the loop.
1545 IV_PHI - the initial value of the induction variable
1548 Return a vector variable, initialized with the first VF values of
1549 the induction variable. E.g., for an iv with IV_PHI='X' and
1550 evolution S, for a vector of 4 units, we want to return:
1551 [X, X + S, X + 2*S, X + 3*S]. */
1554 get_initial_def_for_induction (tree iv_phi)
1556 stmt_vec_info stmt_vinfo = vinfo_for_stmt (iv_phi);
1557 loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_vinfo);
1558 struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
1559 tree scalar_type = TREE_TYPE (PHI_RESULT_TREE (iv_phi));
1562 edge pe = loop_preheader_edge (loop);
1563 struct loop *iv_loop;
1565 tree vec, vec_init, vec_step, t;
1570 tree induction_phi, induc_def, new_stmt, vec_def, vec_dest;
1571 tree init_expr, step_expr;
1572 int vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo);
1577 stmt_vec_info phi_info = vinfo_for_stmt (iv_phi);
1578 bool nested_in_vect_loop = false;
1580 imm_use_iterator imm_iter;
1581 use_operand_p use_p;
1585 block_stmt_iterator si;
1586 basic_block bb = bb_for_stmt (iv_phi);
1588 vectype = get_vectype_for_scalar_type (scalar_type);
1589 gcc_assert (vectype);
1590 nunits = TYPE_VECTOR_SUBPARTS (vectype);
1591 ncopies = vf / nunits;
1593 gcc_assert (phi_info);
1594 gcc_assert (ncopies >= 1);
1596 /* Find the first insertion point in the BB. */
1597 si = bsi_after_labels (bb);
1599 if (INTEGRAL_TYPE_P (scalar_type))
1600 step_expr = build_int_cst (scalar_type, 0);
1602 step_expr = build_real (scalar_type, dconst0);
1604 /* Is phi in an inner-loop, while vectorizing an enclosing outer-loop? */
1605 if (nested_in_vect_loop_p (loop, iv_phi))
1607 nested_in_vect_loop = true;
1608 iv_loop = loop->inner;
1612 gcc_assert (iv_loop == (bb_for_stmt (iv_phi))->loop_father);
1614 latch_e = loop_latch_edge (iv_loop);
1615 loop_arg = PHI_ARG_DEF_FROM_EDGE (iv_phi, latch_e);
1617 access_fn = analyze_scalar_evolution (iv_loop, PHI_RESULT (iv_phi));
1618 gcc_assert (access_fn);
1619 ok = vect_is_simple_iv_evolution (iv_loop->num, access_fn,
1620 &init_expr, &step_expr);
1622 pe = loop_preheader_edge (iv_loop);
1624 /* Create the vector that holds the initial_value of the induction. */
1625 if (nested_in_vect_loop)
1627 /* iv_loop is nested in the loop to be vectorized. init_expr had already
1628 been created during vectorization of previous stmts; We obtain it from
1629 the STMT_VINFO_VEC_STMT of the defining stmt. */
1630 tree iv_def = PHI_ARG_DEF_FROM_EDGE (iv_phi, loop_preheader_edge (iv_loop));
1631 vec_init = vect_get_vec_def_for_operand (iv_def, iv_phi, NULL);
1635 /* iv_loop is the loop to be vectorized. Create:
1636 vec_init = [X, X+S, X+2*S, X+3*S] (S = step_expr, X = init_expr) */
1637 new_var = vect_get_new_vect_var (scalar_type, vect_scalar_var, "var_");
1638 add_referenced_var (new_var);
1640 new_name = force_gimple_operand (init_expr, &stmts, false, new_var);
1643 new_bb = bsi_insert_on_edge_immediate (pe, stmts);
1644 gcc_assert (!new_bb);
1648 t = tree_cons (NULL_TREE, init_expr, t);
1649 for (i = 1; i < nunits; i++)
1653 /* Create: new_name_i = new_name + step_expr */
1654 tmp = fold_build2 (PLUS_EXPR, scalar_type, new_name, step_expr);
1655 init_stmt = build_gimple_modify_stmt (new_var, tmp);
1656 new_name = make_ssa_name (new_var, init_stmt);
1657 GIMPLE_STMT_OPERAND (init_stmt, 0) = new_name;
1659 new_bb = bsi_insert_on_edge_immediate (pe, init_stmt);
1660 gcc_assert (!new_bb);
1662 if (vect_print_dump_info (REPORT_DETAILS))
1664 fprintf (vect_dump, "created new init_stmt: ");
1665 print_generic_expr (vect_dump, init_stmt, TDF_SLIM);
1667 t = tree_cons (NULL_TREE, new_name, t);
1669 /* Create a vector from [new_name_0, new_name_1, ..., new_name_nunits-1] */
1670 vec = build_constructor_from_list (vectype, nreverse (t));
1671 vec_init = vect_init_vector (iv_phi, vec, vectype, NULL);
1675 /* Create the vector that holds the step of the induction. */
1676 if (nested_in_vect_loop)
1677 /* iv_loop is nested in the loop to be vectorized. Generate:
1678 vec_step = [S, S, S, S] */
1679 new_name = step_expr;
1682 /* iv_loop is the loop to be vectorized. Generate:
1683 vec_step = [VF*S, VF*S, VF*S, VF*S] */
1684 expr = build_int_cst (scalar_type, vf);
1685 new_name = fold_build2 (MULT_EXPR, scalar_type, expr, step_expr);
1689 for (i = 0; i < nunits; i++)
1690 t = tree_cons (NULL_TREE, unshare_expr (new_name), t);
1691 gcc_assert (CONSTANT_CLASS_P (new_name));
1692 vec = build_vector (vectype, t);
1693 vec_step = vect_init_vector (iv_phi, vec, vectype, NULL);
1696 /* Create the following def-use cycle:
1701 vec_iv = PHI <vec_init, vec_loop>
1705 vec_loop = vec_iv + vec_step; */
1707 /* Create the induction-phi that defines the induction-operand. */
1708 vec_dest = vect_get_new_vect_var (vectype, vect_simple_var, "vec_iv_");
1709 add_referenced_var (vec_dest);
1710 induction_phi = create_phi_node (vec_dest, iv_loop->header);
1711 set_stmt_info (get_stmt_ann (induction_phi),
1712 new_stmt_vec_info (induction_phi, loop_vinfo));
1713 induc_def = PHI_RESULT (induction_phi);
1715 /* Create the iv update inside the loop */
1716 new_stmt = build_gimple_modify_stmt (NULL_TREE,
1717 build2 (PLUS_EXPR, vectype,
1718 induc_def, vec_step));
1719 vec_def = make_ssa_name (vec_dest, new_stmt);
1720 GIMPLE_STMT_OPERAND (new_stmt, 0) = vec_def;
1721 bsi_insert_before (&si, new_stmt, BSI_SAME_STMT);
1722 set_stmt_info (get_stmt_ann (new_stmt),
1723 new_stmt_vec_info (new_stmt, loop_vinfo));
1725 /* Set the arguments of the phi node: */
1726 add_phi_arg (induction_phi, vec_init, pe);
1727 add_phi_arg (induction_phi, vec_def, loop_latch_edge (iv_loop));
1730 /* In case that vectorization factor (VF) is bigger than the number
1731 of elements that we can fit in a vectype (nunits), we have to generate
1732 more than one vector stmt - i.e - we need to "unroll" the
1733 vector stmt by a factor VF/nunits. For more details see documentation
1734 in vectorizable_operation. */
1738 stmt_vec_info prev_stmt_vinfo;
1739 /* FORNOW. This restriction should be relaxed. */
1740 gcc_assert (!nested_in_vect_loop);
1742 /* Create the vector that holds the step of the induction. */
1743 expr = build_int_cst (scalar_type, nunits);
1744 new_name = fold_build2 (MULT_EXPR, scalar_type, expr, step_expr);
1746 for (i = 0; i < nunits; i++)
1747 t = tree_cons (NULL_TREE, unshare_expr (new_name), t);
1748 gcc_assert (CONSTANT_CLASS_P (new_name));
1749 vec = build_vector (vectype, t);
1750 vec_step = vect_init_vector (iv_phi, vec, vectype, NULL);
1752 vec_def = induc_def;
1753 prev_stmt_vinfo = vinfo_for_stmt (induction_phi);
1754 for (i = 1; i < ncopies; i++)
1758 /* vec_i = vec_prev + vec_step */
1759 tmp = build2 (PLUS_EXPR, vectype, vec_def, vec_step);
1760 new_stmt = build_gimple_modify_stmt (NULL_TREE, tmp);
1761 vec_def = make_ssa_name (vec_dest, new_stmt);
1762 GIMPLE_STMT_OPERAND (new_stmt, 0) = vec_def;
1763 bsi_insert_before (&si, new_stmt, BSI_SAME_STMT);
1764 set_stmt_info (get_stmt_ann (new_stmt),
1765 new_stmt_vec_info (new_stmt, loop_vinfo));
1766 STMT_VINFO_RELATED_STMT (prev_stmt_vinfo) = new_stmt;
1767 prev_stmt_vinfo = vinfo_for_stmt (new_stmt);
1771 if (nested_in_vect_loop)
1773 /* Find the loop-closed exit-phi of the induction, and record
1774 the final vector of induction results: */
1776 FOR_EACH_IMM_USE_FAST (use_p, imm_iter, loop_arg)
1778 if (!flow_bb_inside_loop_p (iv_loop, bb_for_stmt (USE_STMT (use_p))))
1780 exit_phi = USE_STMT (use_p);
1786 stmt_vec_info stmt_vinfo = vinfo_for_stmt (exit_phi);
1787 /* FORNOW. Currently not supporting the case that an inner-loop induction
1788 is not used in the outer-loop (i.e. only outside the outer-loop). */
1789 gcc_assert (STMT_VINFO_RELEVANT_P (stmt_vinfo)
1790 && !STMT_VINFO_LIVE_P (stmt_vinfo));
1792 STMT_VINFO_VEC_STMT (stmt_vinfo) = new_stmt;
1793 if (vect_print_dump_info (REPORT_DETAILS))
1795 fprintf (vect_dump, "vector of inductions after inner-loop:");
1796 print_generic_expr (vect_dump, new_stmt, TDF_SLIM);
1802 if (vect_print_dump_info (REPORT_DETAILS))
1804 fprintf (vect_dump, "transform induction: created def-use cycle:");
1805 print_generic_expr (vect_dump, induction_phi, TDF_SLIM);
1806 fprintf (vect_dump, "\n");
1807 print_generic_expr (vect_dump, SSA_NAME_DEF_STMT (vec_def), TDF_SLIM);
1810 STMT_VINFO_VEC_STMT (phi_info) = induction_phi;
1815 /* Function vect_get_vec_def_for_operand.
1817 OP is an operand in STMT. This function returns a (vector) def that will be
1818 used in the vectorized stmt for STMT.
1820 In the case that OP is an SSA_NAME which is defined in the loop, then
1821 STMT_VINFO_VEC_STMT of the defining stmt holds the relevant def.
1823 In case OP is an invariant or constant, a new stmt that creates a vector def
1824 needs to be introduced. */
1827 vect_get_vec_def_for_operand (tree op, tree stmt, tree *scalar_def)
1832 stmt_vec_info def_stmt_info = NULL;
1833 stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt);
1834 tree vectype = STMT_VINFO_VECTYPE (stmt_vinfo);
1835 int nunits = TYPE_VECTOR_SUBPARTS (vectype);
1836 loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_vinfo);
1842 enum vect_def_type dt;
1846 if (vect_print_dump_info (REPORT_DETAILS))
1848 fprintf (vect_dump, "vect_get_vec_def_for_operand: ");
1849 print_generic_expr (vect_dump, op, TDF_SLIM);
1852 is_simple_use = vect_is_simple_use (op, loop_vinfo, &def_stmt, &def, &dt);
1853 gcc_assert (is_simple_use);
1854 if (vect_print_dump_info (REPORT_DETAILS))
1858 fprintf (vect_dump, "def = ");
1859 print_generic_expr (vect_dump, def, TDF_SLIM);
1863 fprintf (vect_dump, " def_stmt = ");
1864 print_generic_expr (vect_dump, def_stmt, TDF_SLIM);
1870 /* Case 1: operand is a constant. */
1871 case vect_constant_def:
1876 /* Create 'vect_cst_ = {cst,cst,...,cst}' */
1877 if (vect_print_dump_info (REPORT_DETAILS))
1878 fprintf (vect_dump, "Create vector_cst. nunits = %d", nunits);
1880 for (i = nunits - 1; i >= 0; --i)
1882 t = tree_cons (NULL_TREE, op, t);
1884 vector_type = get_vectype_for_scalar_type (TREE_TYPE (op));
1885 gcc_assert (vector_type);
1886 vec_cst = build_vector (vector_type, t);
1888 return vect_init_vector (stmt, vec_cst, vector_type, NULL);
1891 /* Case 2: operand is defined outside the loop - loop invariant. */
1892 case vect_invariant_def:
1897 /* Create 'vec_inv = {inv,inv,..,inv}' */
1898 if (vect_print_dump_info (REPORT_DETAILS))
1899 fprintf (vect_dump, "Create vector_inv.");
1901 for (i = nunits - 1; i >= 0; --i)
1903 t = tree_cons (NULL_TREE, def, t);
1906 /* FIXME: use build_constructor directly. */
1907 vector_type = get_vectype_for_scalar_type (TREE_TYPE (def));
1908 gcc_assert (vector_type);
1909 vec_inv = build_constructor_from_list (vector_type, t);
1910 return vect_init_vector (stmt, vec_inv, vector_type, NULL);
1913 /* Case 3: operand is defined inside the loop. */
1917 *scalar_def = def_stmt;
1919 /* Get the def from the vectorized stmt. */
1920 def_stmt_info = vinfo_for_stmt (def_stmt);
1921 vec_stmt = STMT_VINFO_VEC_STMT (def_stmt_info);
1922 gcc_assert (vec_stmt);
1923 if (TREE_CODE (vec_stmt) == PHI_NODE)
1924 vec_oprnd = PHI_RESULT (vec_stmt);
1926 vec_oprnd = GIMPLE_STMT_OPERAND (vec_stmt, 0);
1930 /* Case 4: operand is defined by a loop header phi - reduction */
1931 case vect_reduction_def:
1935 gcc_assert (TREE_CODE (def_stmt) == PHI_NODE);
1936 loop = (bb_for_stmt (def_stmt))->loop_father;
1938 /* Get the def before the loop */
1939 op = PHI_ARG_DEF_FROM_EDGE (def_stmt, loop_preheader_edge (loop));
1940 return get_initial_def_for_reduction (stmt, op, scalar_def);
1943 /* Case 5: operand is defined by loop-header phi - induction. */
1944 case vect_induction_def:
1946 gcc_assert (TREE_CODE (def_stmt) == PHI_NODE);
1948 /* Get the def from the vectorized stmt. */
1949 def_stmt_info = vinfo_for_stmt (def_stmt);
1950 vec_stmt = STMT_VINFO_VEC_STMT (def_stmt_info);
1951 gcc_assert (vec_stmt && (TREE_CODE (vec_stmt) == PHI_NODE));
1952 vec_oprnd = PHI_RESULT (vec_stmt);
1962 /* Function vect_get_vec_def_for_stmt_copy
1964 Return a vector-def for an operand. This function is used when the
1965 vectorized stmt to be created (by the caller to this function) is a "copy"
1966 created in case the vectorized result cannot fit in one vector, and several
1967 copies of the vector-stmt are required. In this case the vector-def is
1968 retrieved from the vector stmt recorded in the STMT_VINFO_RELATED_STMT field
1969 of the stmt that defines VEC_OPRND.
1970 DT is the type of the vector def VEC_OPRND.
1973 In case the vectorization factor (VF) is bigger than the number
1974 of elements that can fit in a vectype (nunits), we have to generate
1975 more than one vector stmt to vectorize the scalar stmt. This situation
1976 arises when there are multiple data-types operated upon in the loop; the
1977 smallest data-type determines the VF, and as a result, when vectorizing
1978 stmts operating on wider types we need to create 'VF/nunits' "copies" of the
1979 vector stmt (each computing a vector of 'nunits' results, and together
1980 computing 'VF' results in each iteration). This function is called when
1981 vectorizing such a stmt (e.g. vectorizing S2 in the illustration below, in
1982 which VF=16 and nunits=4, so the number of copies required is 4):
1984 scalar stmt: vectorized into: STMT_VINFO_RELATED_STMT
1986 S1: x = load VS1.0: vx.0 = memref0 VS1.1
1987 VS1.1: vx.1 = memref1 VS1.2
1988 VS1.2: vx.2 = memref2 VS1.3
1989 VS1.3: vx.3 = memref3
1991 S2: z = x + ... VSnew.0: vz0 = vx.0 + ... VSnew.1
1992 VSnew.1: vz1 = vx.1 + ... VSnew.2
1993 VSnew.2: vz2 = vx.2 + ... VSnew.3
1994 VSnew.3: vz3 = vx.3 + ...
1996 The vectorization of S1 is explained in vectorizable_load.
1997 The vectorization of S2:
1998 To create the first vector-stmt out of the 4 copies - VSnew.0 -
1999 the function 'vect_get_vec_def_for_operand' is called to
2000 get the relevant vector-def for each operand of S2. For operand x it
2001 returns the vector-def 'vx.0'.
2003 To create the remaining copies of the vector-stmt (VSnew.j), this
2004 function is called to get the relevant vector-def for each operand. It is
2005 obtained from the respective VS1.j stmt, which is recorded in the
2006 STMT_VINFO_RELATED_STMT field of the stmt that defines VEC_OPRND.
2008 For example, to obtain the vector-def 'vx.1' in order to create the
2009 vector stmt 'VSnew.1', this function is called with VEC_OPRND='vx.0'.
2010 Given 'vx0' we obtain the stmt that defines it ('VS1.0'); from the
2011 STMT_VINFO_RELATED_STMT field of 'VS1.0' we obtain the next copy - 'VS1.1',
2012 and return its def ('vx.1').
2013 Overall, to create the above sequence this function will be called 3 times:
2014 vx.1 = vect_get_vec_def_for_stmt_copy (dt, vx.0);
2015 vx.2 = vect_get_vec_def_for_stmt_copy (dt, vx.1);
2016 vx.3 = vect_get_vec_def_for_stmt_copy (dt, vx.2); */
2019 vect_get_vec_def_for_stmt_copy (enum vect_def_type dt, tree vec_oprnd)
2021 tree vec_stmt_for_operand;
2022 stmt_vec_info def_stmt_info;
2024 /* Do nothing; can reuse same def. */
2025 if (dt == vect_invariant_def || dt == vect_constant_def )
2028 vec_stmt_for_operand = SSA_NAME_DEF_STMT (vec_oprnd);
2029 def_stmt_info = vinfo_for_stmt (vec_stmt_for_operand);
2030 gcc_assert (def_stmt_info);
2031 vec_stmt_for_operand = STMT_VINFO_RELATED_STMT (def_stmt_info);
2032 gcc_assert (vec_stmt_for_operand);
2033 vec_oprnd = GIMPLE_STMT_OPERAND (vec_stmt_for_operand, 0);
2038 /* Get vectorized definitions for the operands to create a copy of an original
2039 stmt. See vect_get_vec_def_for_stmt_copy() for details. */
2042 vect_get_vec_defs_for_stmt_copy (enum vect_def_type *dt,
2043 VEC(tree,heap) **vec_oprnds0,
2044 VEC(tree,heap) **vec_oprnds1)
2046 tree vec_oprnd = VEC_pop (tree, *vec_oprnds0);
2048 vec_oprnd = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd);
2049 VEC_quick_push (tree, *vec_oprnds0, vec_oprnd);
2051 if (vec_oprnds1 && *vec_oprnds1)
2053 vec_oprnd = VEC_pop (tree, *vec_oprnds1);
2054 vec_oprnd = vect_get_vec_def_for_stmt_copy (dt[1], vec_oprnd);
2055 VEC_quick_push (tree, *vec_oprnds1, vec_oprnd);
2060 /* Get vectorized definitions for OP0 and OP1, or SLP_NODE if it is not NULL. */
2063 vect_get_vec_defs (tree op0, tree op1, tree stmt, VEC(tree,heap) **vec_oprnds0,
2064 VEC(tree,heap) **vec_oprnds1, slp_tree slp_node)
2067 vect_get_slp_defs (slp_node, vec_oprnds0, vec_oprnds1);
2072 *vec_oprnds0 = VEC_alloc (tree, heap, 1);
2073 vec_oprnd = vect_get_vec_def_for_operand (op0, stmt, NULL);
2074 VEC_quick_push (tree, *vec_oprnds0, vec_oprnd);
2078 *vec_oprnds1 = VEC_alloc (tree, heap, 1);
2079 vec_oprnd = vect_get_vec_def_for_operand (op1, stmt, NULL);
2080 VEC_quick_push (tree, *vec_oprnds1, vec_oprnd);
2086 /* Function vect_finish_stmt_generation.
2088 Insert a new stmt. */
2091 vect_finish_stmt_generation (tree stmt, tree vec_stmt,
2092 block_stmt_iterator *bsi)
2094 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
2095 loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info);
2097 gcc_assert (stmt == bsi_stmt (*bsi));
2098 gcc_assert (TREE_CODE (stmt) != LABEL_EXPR);
2100 bsi_insert_before (bsi, vec_stmt, BSI_SAME_STMT);
2102 set_stmt_info (get_stmt_ann (vec_stmt),
2103 new_stmt_vec_info (vec_stmt, loop_vinfo));
2105 if (vect_print_dump_info (REPORT_DETAILS))
2107 fprintf (vect_dump, "add new stmt: ");
2108 print_generic_expr (vect_dump, vec_stmt, TDF_SLIM);
2111 /* Make sure bsi points to the stmt that is being vectorized. */
2112 gcc_assert (stmt == bsi_stmt (*bsi));
2114 SET_EXPR_LOCATION (vec_stmt, EXPR_LOCATION (stmt));
2118 /* Function get_initial_def_for_reduction
2121 STMT - a stmt that performs a reduction operation in the loop.
2122 INIT_VAL - the initial value of the reduction variable
2125 ADJUSTMENT_DEF - a tree that holds a value to be added to the final result
2126 of the reduction (used for adjusting the epilog - see below).
2127 Return a vector variable, initialized according to the operation that STMT
2128 performs. This vector will be used as the initial value of the
2129 vector of partial results.
2131 Option1 (adjust in epilog): Initialize the vector as follows:
2134 min/max: [init_val,init_val,..,init_val,init_val]
2135 bit and/or: [init_val,init_val,..,init_val,init_val]
2136 and when necessary (e.g. add/mult case) let the caller know
2137 that it needs to adjust the result by init_val.
2139 Option2: Initialize the vector as follows:
2140 add: [0,0,...,0,init_val]
2141 mult: [1,1,...,1,init_val]
2142 min/max: [init_val,init_val,...,init_val]
2143 bit and/or: [init_val,init_val,...,init_val]
2144 and no adjustments are needed.
2146 For example, for the following code:
2152 STMT is 's = s + a[i]', and the reduction variable is 's'.
2153 For a vector of 4 units, we want to return either [0,0,0,init_val],
2154 or [0,0,0,0] and let the caller know that it needs to adjust
2155 the result at the end by 'init_val'.
2157 FORNOW, we are using the 'adjust in epilog' scheme, because this way the
2158 initialization vector is simpler (same element in all entries).
2159 A cost model should help decide between these two schemes. */
2162 get_initial_def_for_reduction (tree stmt, tree init_val, tree *adjustment_def)
2164 stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt);
2165 loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_vinfo);
2166 struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
2167 tree vectype = STMT_VINFO_VECTYPE (stmt_vinfo);
2168 int nunits = TYPE_VECTOR_SUBPARTS (vectype);
2169 enum tree_code code = TREE_CODE (GIMPLE_STMT_OPERAND (stmt, 1));
2170 tree type = TREE_TYPE (init_val);
2177 bool nested_in_vect_loop = false;
2179 gcc_assert (POINTER_TYPE_P (type) || INTEGRAL_TYPE_P (type) || SCALAR_FLOAT_TYPE_P (type));
2180 if (nested_in_vect_loop_p (loop, stmt))
2181 nested_in_vect_loop = true;
2183 gcc_assert (loop == (bb_for_stmt (stmt))->loop_father);
2185 vecdef = vect_get_vec_def_for_operand (init_val, stmt, NULL);
2189 case WIDEN_SUM_EXPR:
2192 if (nested_in_vect_loop)
2193 *adjustment_def = vecdef;
2195 *adjustment_def = init_val;
2196 /* Create a vector of zeros for init_def. */
2197 if (SCALAR_FLOAT_TYPE_P (type))
2198 def_for_init = build_real (type, dconst0);
2200 def_for_init = build_int_cst (type, 0);
2201 for (i = nunits - 1; i >= 0; --i)
2202 t = tree_cons (NULL_TREE, def_for_init, t);
2203 vector_type = get_vectype_for_scalar_type (TREE_TYPE (def_for_init));
2204 gcc_assert (vector_type);
2205 init_def = build_vector (vector_type, t);
2210 *adjustment_def = NULL_TREE;
2222 /* Function vect_create_epilog_for_reduction
2224 Create code at the loop-epilog to finalize the result of a reduction
2227 VECT_DEF is a vector of partial results.
2228 REDUC_CODE is the tree-code for the epilog reduction.
2229 STMT is the scalar reduction stmt that is being vectorized.
2230 REDUCTION_PHI is the phi-node that carries the reduction computation.
2233 1. Creates the reduction def-use cycle: sets the arguments for
2235 The loop-entry argument is the vectorized initial-value of the reduction.
2236 The loop-latch argument is VECT_DEF - the vector of partial sums.
2237 2. "Reduces" the vector of partial results VECT_DEF into a single result,
2238 by applying the operation specified by REDUC_CODE if available, or by
2239 other means (whole-vector shifts or a scalar loop).
2240 The function also creates a new phi node at the loop exit to preserve
2241 loop-closed form, as illustrated below.
2243 The flow at the entry to this function:
2246 vec_def = phi <null, null> # REDUCTION_PHI
2247 VECT_DEF = vector_stmt # vectorized form of STMT
2248 s_loop = scalar_stmt # (scalar) STMT
2250 s_out0 = phi <s_loop> # (scalar) EXIT_PHI
2254 The above is transformed by this function into:
2257 vec_def = phi <vec_init, VECT_DEF> # REDUCTION_PHI
2258 VECT_DEF = vector_stmt # vectorized form of STMT
2259 s_loop = scalar_stmt # (scalar) STMT
2261 s_out0 = phi <s_loop> # (scalar) EXIT_PHI
2262 v_out1 = phi <VECT_DEF> # NEW_EXIT_PHI
2263 v_out2 = reduce <v_out1>
2264 s_out3 = extract_field <v_out2, 0>
2265 s_out4 = adjust_result <s_out3>
2271 vect_create_epilog_for_reduction (tree vect_def, tree stmt,
2272 enum tree_code reduc_code, tree reduction_phi)
2274 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
2276 enum machine_mode mode;
2277 loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info);
2278 struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
2279 basic_block exit_bb;
2283 block_stmt_iterator exit_bsi;
2285 tree new_temp = NULL_TREE;
2287 tree epilog_stmt = NULL_TREE;
2288 tree new_scalar_dest, exit_phi, new_dest;
2289 tree bitsize, bitpos, bytesize;
2290 enum tree_code code = TREE_CODE (GIMPLE_STMT_OPERAND (stmt, 1));
2291 tree adjustment_def;
2292 tree vec_initial_def;
2294 imm_use_iterator imm_iter;
2295 use_operand_p use_p;
2296 bool extract_scalar_result = false;
2297 tree reduction_op, expr;
2300 tree operation = GIMPLE_STMT_OPERAND (stmt, 1);
2301 bool nested_in_vect_loop = false;
2303 VEC(tree,heap) *phis = NULL;
2306 if (nested_in_vect_loop_p (loop, stmt))
2309 nested_in_vect_loop = true;
2312 op_type = TREE_OPERAND_LENGTH (operation);
2313 reduction_op = TREE_OPERAND (operation, op_type-1);
2314 vectype = get_vectype_for_scalar_type (TREE_TYPE (reduction_op));
2315 gcc_assert (vectype);
2316 mode = TYPE_MODE (vectype);
2318 /*** 1. Create the reduction def-use cycle ***/
2320 /* 1.1 set the loop-entry arg of the reduction-phi: */
2321 /* For the case of reduction, vect_get_vec_def_for_operand returns
2322 the scalar def before the loop, that defines the initial value
2323 of the reduction variable. */
2324 vec_initial_def = vect_get_vec_def_for_operand (reduction_op, stmt,
2326 add_phi_arg (reduction_phi, vec_initial_def, loop_preheader_edge (loop));
2328 /* 1.2 set the loop-latch arg for the reduction-phi: */
2329 add_phi_arg (reduction_phi, vect_def, loop_latch_edge (loop));
2331 if (vect_print_dump_info (REPORT_DETAILS))
2333 fprintf (vect_dump, "transform reduction: created def-use cycle:");
2334 print_generic_expr (vect_dump, reduction_phi, TDF_SLIM);
2335 fprintf (vect_dump, "\n");
2336 print_generic_expr (vect_dump, SSA_NAME_DEF_STMT (vect_def), TDF_SLIM);
2340 /*** 2. Create epilog code
2341 The reduction epilog code operates across the elements of the vector
2342 of partial results computed by the vectorized loop.
2343 The reduction epilog code consists of:
2344 step 1: compute the scalar result in a vector (v_out2)
2345 step 2: extract the scalar result (s_out3) from the vector (v_out2)
2346 step 3: adjust the scalar result (s_out3) if needed.
2348 Step 1 can be accomplished using one the following three schemes:
2349 (scheme 1) using reduc_code, if available.
2350 (scheme 2) using whole-vector shifts, if available.
2351 (scheme 3) using a scalar loop. In this case steps 1+2 above are
2354 The overall epilog code looks like this:
2356 s_out0 = phi <s_loop> # original EXIT_PHI
2357 v_out1 = phi <VECT_DEF> # NEW_EXIT_PHI
2358 v_out2 = reduce <v_out1> # step 1
2359 s_out3 = extract_field <v_out2, 0> # step 2
2360 s_out4 = adjust_result <s_out3> # step 3
2362 (step 3 is optional, and step2 1 and 2 may be combined).
2363 Lastly, the uses of s_out0 are replaced by s_out4.
2367 /* 2.1 Create new loop-exit-phi to preserve loop-closed form:
2368 v_out1 = phi <v_loop> */
2370 exit_bb = single_exit (loop)->dest;
2371 new_phi = create_phi_node (SSA_NAME_VAR (vect_def), exit_bb);
2372 SET_PHI_ARG_DEF (new_phi, single_exit (loop)->dest_idx, vect_def);
2373 exit_bsi = bsi_after_labels (exit_bb);
2375 /* 2.2 Get the relevant tree-code to use in the epilog for schemes 2,3
2376 (i.e. when reduc_code is not available) and in the final adjustment
2377 code (if needed). Also get the original scalar reduction variable as
2378 defined in the loop. In case STMT is a "pattern-stmt" (i.e. - it
2379 represents a reduction pattern), the tree-code and scalar-def are
2380 taken from the original stmt that the pattern-stmt (STMT) replaces.
2381 Otherwise (it is a regular reduction) - the tree-code and scalar-def
2382 are taken from STMT. */
2384 orig_stmt = STMT_VINFO_RELATED_STMT (stmt_info);
2387 /* Regular reduction */
2392 /* Reduction pattern */
2393 stmt_vec_info stmt_vinfo = vinfo_for_stmt (orig_stmt);
2394 gcc_assert (STMT_VINFO_IN_PATTERN_P (stmt_vinfo));
2395 gcc_assert (STMT_VINFO_RELATED_STMT (stmt_vinfo) == stmt);
2397 code = TREE_CODE (GIMPLE_STMT_OPERAND (orig_stmt, 1));
2398 scalar_dest = GIMPLE_STMT_OPERAND (orig_stmt, 0);
2399 scalar_type = TREE_TYPE (scalar_dest);
2400 new_scalar_dest = vect_create_destination_var (scalar_dest, NULL);
2401 bitsize = TYPE_SIZE (scalar_type);
2402 bytesize = TYPE_SIZE_UNIT (scalar_type);
2405 /* In case this is a reduction in an inner-loop while vectorizing an outer
2406 loop - we don't need to extract a single scalar result at the end of the
2407 inner-loop. The final vector of partial results will be used in the
2408 vectorized outer-loop, or reduced to a scalar result at the end of the
2410 if (nested_in_vect_loop)
2411 goto vect_finalize_reduction;
2413 /* 2.3 Create the reduction code, using one of the three schemes described
2416 if (reduc_code < NUM_TREE_CODES)
2420 /*** Case 1: Create:
2421 v_out2 = reduc_expr <v_out1> */
2423 if (vect_print_dump_info (REPORT_DETAILS))
2424 fprintf (vect_dump, "Reduce using direct vector reduction.");
2426 vec_dest = vect_create_destination_var (scalar_dest, vectype);
2427 tmp = build1 (reduc_code, vectype, PHI_RESULT (new_phi));
2428 epilog_stmt = build_gimple_modify_stmt (vec_dest, tmp);
2429 new_temp = make_ssa_name (vec_dest, epilog_stmt);
2430 GIMPLE_STMT_OPERAND (epilog_stmt, 0) = new_temp;
2431 bsi_insert_before (&exit_bsi, epilog_stmt, BSI_SAME_STMT);
2433 extract_scalar_result = true;
2437 enum tree_code shift_code = 0;
2438 bool have_whole_vector_shift = true;
2440 int element_bitsize = tree_low_cst (bitsize, 1);
2441 int vec_size_in_bits = tree_low_cst (TYPE_SIZE (vectype), 1);
2444 if (optab_handler (vec_shr_optab, mode)->insn_code != CODE_FOR_nothing)
2445 shift_code = VEC_RSHIFT_EXPR;
2447 have_whole_vector_shift = false;
2449 /* Regardless of whether we have a whole vector shift, if we're
2450 emulating the operation via tree-vect-generic, we don't want
2451 to use it. Only the first round of the reduction is likely
2452 to still be profitable via emulation. */
2453 /* ??? It might be better to emit a reduction tree code here, so that
2454 tree-vect-generic can expand the first round via bit tricks. */
2455 if (!VECTOR_MODE_P (mode))
2456 have_whole_vector_shift = false;
2459 optab optab = optab_for_tree_code (code, vectype);
2460 if (optab_handler (optab, mode)->insn_code == CODE_FOR_nothing)
2461 have_whole_vector_shift = false;
2464 if (have_whole_vector_shift)
2466 /*** Case 2: Create:
2467 for (offset = VS/2; offset >= element_size; offset/=2)
2469 Create: va' = vec_shift <va, offset>
2470 Create: va = vop <va, va'>
2473 if (vect_print_dump_info (REPORT_DETAILS))
2474 fprintf (vect_dump, "Reduce using vector shifts");
2476 vec_dest = vect_create_destination_var (scalar_dest, vectype);
2477 new_temp = PHI_RESULT (new_phi);
2479 for (bit_offset = vec_size_in_bits/2;
2480 bit_offset >= element_bitsize;
2483 tree bitpos = size_int (bit_offset);
2484 tree tmp = build2 (shift_code, vectype, new_temp, bitpos);
2485 epilog_stmt = build_gimple_modify_stmt (vec_dest, tmp);
2486 new_name = make_ssa_name (vec_dest, epilog_stmt);
2487 GIMPLE_STMT_OPERAND (epilog_stmt, 0) = new_name;
2488 bsi_insert_before (&exit_bsi, epilog_stmt, BSI_SAME_STMT);
2490 tmp = build2 (code, vectype, new_name, new_temp);
2491 epilog_stmt = build_gimple_modify_stmt (vec_dest, tmp);
2492 new_temp = make_ssa_name (vec_dest, epilog_stmt);
2493 GIMPLE_STMT_OPERAND (epilog_stmt, 0) = new_temp;
2494 bsi_insert_before (&exit_bsi, epilog_stmt, BSI_SAME_STMT);
2497 extract_scalar_result = true;
2503 /*** Case 3: Create:
2504 s = extract_field <v_out2, 0>
2505 for (offset = element_size;
2506 offset < vector_size;
2507 offset += element_size;)
2509 Create: s' = extract_field <v_out2, offset>
2510 Create: s = op <s, s'>
2513 if (vect_print_dump_info (REPORT_DETAILS))
2514 fprintf (vect_dump, "Reduce using scalar code. ");
2516 vec_temp = PHI_RESULT (new_phi);
2517 vec_size_in_bits = tree_low_cst (TYPE_SIZE (vectype), 1);
2518 rhs = build3 (BIT_FIELD_REF, scalar_type, vec_temp, bitsize,
2520 epilog_stmt = build_gimple_modify_stmt (new_scalar_dest, rhs);
2521 new_temp = make_ssa_name (new_scalar_dest, epilog_stmt);
2522 GIMPLE_STMT_OPERAND (epilog_stmt, 0) = new_temp;
2523 bsi_insert_before (&exit_bsi, epilog_stmt, BSI_SAME_STMT);
2525 for (bit_offset = element_bitsize;
2526 bit_offset < vec_size_in_bits;
2527 bit_offset += element_bitsize)
2530 tree bitpos = bitsize_int (bit_offset);
2531 tree rhs = build3 (BIT_FIELD_REF, scalar_type, vec_temp, bitsize,
2534 epilog_stmt = build_gimple_modify_stmt (new_scalar_dest, rhs);
2535 new_name = make_ssa_name (new_scalar_dest, epilog_stmt);
2536 GIMPLE_STMT_OPERAND (epilog_stmt, 0) = new_name;
2537 bsi_insert_before (&exit_bsi, epilog_stmt, BSI_SAME_STMT);
2539 tmp = build2 (code, scalar_type, new_name, new_temp);
2540 epilog_stmt = build_gimple_modify_stmt (new_scalar_dest, tmp);
2541 new_temp = make_ssa_name (new_scalar_dest, epilog_stmt);
2542 GIMPLE_STMT_OPERAND (epilog_stmt, 0) = new_temp;
2543 bsi_insert_before (&exit_bsi, epilog_stmt, BSI_SAME_STMT);
2546 extract_scalar_result = false;
2550 /* 2.4 Extract the final scalar result. Create:
2551 s_out3 = extract_field <v_out2, bitpos> */
2553 if (extract_scalar_result)
2557 gcc_assert (!nested_in_vect_loop);
2558 if (vect_print_dump_info (REPORT_DETAILS))
2559 fprintf (vect_dump, "extract scalar result");
2561 if (BYTES_BIG_ENDIAN)
2562 bitpos = size_binop (MULT_EXPR,
2563 bitsize_int (TYPE_VECTOR_SUBPARTS (vectype) - 1),
2564 TYPE_SIZE (scalar_type));
2566 bitpos = bitsize_zero_node;
2568 rhs = build3 (BIT_FIELD_REF, scalar_type, new_temp, bitsize, bitpos);
2569 epilog_stmt = build_gimple_modify_stmt (new_scalar_dest, rhs);
2570 new_temp = make_ssa_name (new_scalar_dest, epilog_stmt);
2571 GIMPLE_STMT_OPERAND (epilog_stmt, 0) = new_temp;
2572 bsi_insert_before (&exit_bsi, epilog_stmt, BSI_SAME_STMT);
2575 vect_finalize_reduction:
2577 /* 2.5 Adjust the final result by the initial value of the reduction
2578 variable. (When such adjustment is not needed, then
2579 'adjustment_def' is zero). For example, if code is PLUS we create:
2580 new_temp = loop_exit_def + adjustment_def */
2584 if (nested_in_vect_loop)
2586 gcc_assert (TREE_CODE (TREE_TYPE (adjustment_def)) == VECTOR_TYPE);
2587 expr = build2 (code, vectype, PHI_RESULT (new_phi), adjustment_def);
2588 new_dest = vect_create_destination_var (scalar_dest, vectype);
2592 gcc_assert (TREE_CODE (TREE_TYPE (adjustment_def)) != VECTOR_TYPE);
2593 expr = build2 (code, scalar_type, new_temp, adjustment_def);
2594 new_dest = vect_create_destination_var (scalar_dest, scalar_type);
2596 epilog_stmt = build_gimple_modify_stmt (new_dest, expr);
2597 new_temp = make_ssa_name (new_dest, epilog_stmt);
2598 GIMPLE_STMT_OPERAND (epilog_stmt, 0) = new_temp;
2599 bsi_insert_before (&exit_bsi, epilog_stmt, BSI_SAME_STMT);
2603 /* 2.6 Handle the loop-exit phi */
2605 /* Replace uses of s_out0 with uses of s_out3:
2606 Find the loop-closed-use at the loop exit of the original scalar result.
2607 (The reduction result is expected to have two immediate uses - one at the
2608 latch block, and one at the loop exit). */
2609 phis = VEC_alloc (tree, heap, 10);
2610 FOR_EACH_IMM_USE_FAST (use_p, imm_iter, scalar_dest)
2612 if (!flow_bb_inside_loop_p (loop, bb_for_stmt (USE_STMT (use_p))))
2614 exit_phi = USE_STMT (use_p);
2615 VEC_quick_push (tree, phis, exit_phi);
2618 /* We expect to have found an exit_phi because of loop-closed-ssa form. */
2619 gcc_assert (!VEC_empty (tree, phis));
2621 for (i = 0; VEC_iterate (tree, phis, i, exit_phi); i++)
2623 if (nested_in_vect_loop)
2625 stmt_vec_info stmt_vinfo = vinfo_for_stmt (exit_phi);
2627 /* FORNOW. Currently not supporting the case that an inner-loop reduction
2628 is not used in the outer-loop (but only outside the outer-loop). */
2629 gcc_assert (STMT_VINFO_RELEVANT_P (stmt_vinfo)
2630 && !STMT_VINFO_LIVE_P (stmt_vinfo));
2632 epilog_stmt = adjustment_def ? epilog_stmt : new_phi;
2633 STMT_VINFO_VEC_STMT (stmt_vinfo) = epilog_stmt;
2634 set_stmt_info (get_stmt_ann (epilog_stmt),
2635 new_stmt_vec_info (epilog_stmt, loop_vinfo));
2639 /* Replace the uses: */
2640 orig_name = PHI_RESULT (exit_phi);
2641 FOR_EACH_IMM_USE_STMT (use_stmt, imm_iter, orig_name)
2642 FOR_EACH_IMM_USE_ON_STMT (use_p, imm_iter)
2643 SET_USE (use_p, new_temp);
2645 VEC_free (tree, heap, phis);
2649 /* Function vectorizable_reduction.
2651 Check if STMT performs a reduction operation that can be vectorized.
2652 If VEC_STMT is also passed, vectorize the STMT: create a vectorized
2653 stmt to replace it, put it in VEC_STMT, and insert it at BSI.
2654 Return FALSE if not a vectorizable STMT, TRUE otherwise.
2656 This function also handles reduction idioms (patterns) that have been
2657 recognized in advance during vect_pattern_recog. In this case, STMT may be
2659 X = pattern_expr (arg0, arg1, ..., X)
2660 and it's STMT_VINFO_RELATED_STMT points to the last stmt in the original
2661 sequence that had been detected and replaced by the pattern-stmt (STMT).
2663 In some cases of reduction patterns, the type of the reduction variable X is
2664 different than the type of the other arguments of STMT.
2665 In such cases, the vectype that is used when transforming STMT into a vector
2666 stmt is different than the vectype that is used to determine the
2667 vectorization factor, because it consists of a different number of elements
2668 than the actual number of elements that are being operated upon in parallel.
2670 For example, consider an accumulation of shorts into an int accumulator.
2671 On some targets it's possible to vectorize this pattern operating on 8
2672 shorts at a time (hence, the vectype for purposes of determining the
2673 vectorization factor should be V8HI); on the other hand, the vectype that
2674 is used to create the vector form is actually V4SI (the type of the result).
2676 Upon entry to this function, STMT_VINFO_VECTYPE records the vectype that
2677 indicates what is the actual level of parallelism (V8HI in the example), so
2678 that the right vectorization factor would be derived. This vectype
2679 corresponds to the type of arguments to the reduction stmt, and should *NOT*
2680 be used to create the vectorized stmt. The right vectype for the vectorized
2681 stmt is obtained from the type of the result X:
2682 get_vectype_for_scalar_type (TREE_TYPE (X))
2684 This means that, contrary to "regular" reductions (or "regular" stmts in
2685 general), the following equation:
2686 STMT_VINFO_VECTYPE == get_vectype_for_scalar_type (TREE_TYPE (X))
2687 does *NOT* necessarily hold for reduction patterns. */
2690 vectorizable_reduction (tree stmt, block_stmt_iterator *bsi, tree *vec_stmt)
2695 tree loop_vec_def0 = NULL_TREE, loop_vec_def1 = NULL_TREE;
2696 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
2697 tree vectype = STMT_VINFO_VECTYPE (stmt_info);
2698 loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info);
2699 struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
2701 enum tree_code code, orig_code, epilog_reduc_code = 0;
2702 enum machine_mode vec_mode;
2704 optab optab, reduc_optab;
2705 tree new_temp = NULL_TREE;
2707 enum vect_def_type dt;
2712 stmt_vec_info orig_stmt_info;
2713 tree expr = NULL_TREE;
2715 int nunits = TYPE_VECTOR_SUBPARTS (vectype);
2716 int ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits;
2717 stmt_vec_info prev_stmt_info;
2719 tree new_stmt = NULL_TREE;
2722 if (nested_in_vect_loop_p (loop, stmt))
2725 /* FORNOW. This restriction should be relaxed. */
2728 if (vect_print_dump_info (REPORT_DETAILS))
2729 fprintf (vect_dump, "multiple types in nested loop.");
2734 gcc_assert (ncopies >= 1);
2736 /* FORNOW: SLP not supported. */
2737 if (STMT_SLP_TYPE (stmt_info))
2740 /* 1. Is vectorizable reduction? */
2742 /* Not supportable if the reduction variable is used in the loop. */
2743 if (STMT_VINFO_RELEVANT (stmt_info) > vect_used_in_outer)
2746 /* Reductions that are not used even in an enclosing outer-loop,
2747 are expected to be "live" (used out of the loop). */
2748 if (STMT_VINFO_RELEVANT (stmt_info) == vect_unused_in_loop
2749 && !STMT_VINFO_LIVE_P (stmt_info))
2752 /* Make sure it was already recognized as a reduction computation. */
2753 if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_reduction_def)
2756 /* 2. Has this been recognized as a reduction pattern?
2758 Check if STMT represents a pattern that has been recognized
2759 in earlier analysis stages. For stmts that represent a pattern,
2760 the STMT_VINFO_RELATED_STMT field records the last stmt in
2761 the original sequence that constitutes the pattern. */
2763 orig_stmt = STMT_VINFO_RELATED_STMT (stmt_info);
2766 orig_stmt_info = vinfo_for_stmt (orig_stmt);
2767 gcc_assert (STMT_VINFO_RELATED_STMT (orig_stmt_info) == stmt);
2768 gcc_assert (STMT_VINFO_IN_PATTERN_P (orig_stmt_info));
2769 gcc_assert (!STMT_VINFO_IN_PATTERN_P (stmt_info));
2772 /* 3. Check the operands of the operation. The first operands are defined
2773 inside the loop body. The last operand is the reduction variable,
2774 which is defined by the loop-header-phi. */
2776 gcc_assert (TREE_CODE (stmt) == GIMPLE_MODIFY_STMT);
2778 operation = GIMPLE_STMT_OPERAND (stmt, 1);
2779 code = TREE_CODE (operation);
2780 op_type = TREE_OPERAND_LENGTH (operation);
2781 if (op_type != binary_op && op_type != ternary_op)
2783 scalar_dest = GIMPLE_STMT_OPERAND (stmt, 0);
2784 scalar_type = TREE_TYPE (scalar_dest);
2785 if (!POINTER_TYPE_P (scalar_type) && !INTEGRAL_TYPE_P (scalar_type)
2786 && !SCALAR_FLOAT_TYPE_P (scalar_type))
2789 /* All uses but the last are expected to be defined in the loop.
2790 The last use is the reduction variable. */
2791 for (i = 0; i < op_type-1; i++)
2793 op = TREE_OPERAND (operation, i);
2794 is_simple_use = vect_is_simple_use (op, loop_vinfo, &def_stmt, &def, &dt);
2795 gcc_assert (is_simple_use);
2796 if (dt != vect_loop_def
2797 && dt != vect_invariant_def
2798 && dt != vect_constant_def
2799 && dt != vect_induction_def)
2803 op = TREE_OPERAND (operation, i);
2804 is_simple_use = vect_is_simple_use (op, loop_vinfo, &def_stmt, &def, &dt);
2805 gcc_assert (is_simple_use);
2806 gcc_assert (dt == vect_reduction_def);
2807 gcc_assert (TREE_CODE (def_stmt) == PHI_NODE);
2809 gcc_assert (orig_stmt == vect_is_simple_reduction (loop_vinfo, def_stmt));
2811 gcc_assert (stmt == vect_is_simple_reduction (loop_vinfo, def_stmt));
2813 if (STMT_VINFO_LIVE_P (vinfo_for_stmt (def_stmt)))
2816 /* 4. Supportable by target? */
2818 /* 4.1. check support for the operation in the loop */
2819 optab = optab_for_tree_code (code, vectype);
2822 if (vect_print_dump_info (REPORT_DETAILS))
2823 fprintf (vect_dump, "no optab.");
2826 vec_mode = TYPE_MODE (vectype);
2827 if (optab_handler (optab, vec_mode)->insn_code == CODE_FOR_nothing)
2829 if (vect_print_dump_info (REPORT_DETAILS))
2830 fprintf (vect_dump, "op not supported by target.");
2831 if (GET_MODE_SIZE (vec_mode) != UNITS_PER_WORD
2832 || LOOP_VINFO_VECT_FACTOR (loop_vinfo)
2833 < vect_min_worthwhile_factor (code))
2835 if (vect_print_dump_info (REPORT_DETAILS))
2836 fprintf (vect_dump, "proceeding using word mode.");
2839 /* Worthwhile without SIMD support? */
2840 if (!VECTOR_MODE_P (TYPE_MODE (vectype))
2841 && LOOP_VINFO_VECT_FACTOR (loop_vinfo)
2842 < vect_min_worthwhile_factor (code))
2844 if (vect_print_dump_info (REPORT_DETAILS))
2845 fprintf (vect_dump, "not worthwhile without SIMD support.");
2849 /* 4.2. Check support for the epilog operation.
2851 If STMT represents a reduction pattern, then the type of the
2852 reduction variable may be different than the type of the rest
2853 of the arguments. For example, consider the case of accumulation
2854 of shorts into an int accumulator; The original code:
2855 S1: int_a = (int) short_a;
2856 orig_stmt-> S2: int_acc = plus <int_a ,int_acc>;
2859 STMT: int_acc = widen_sum <short_a, int_acc>
2862 1. The tree-code that is used to create the vector operation in the
2863 epilog code (that reduces the partial results) is not the
2864 tree-code of STMT, but is rather the tree-code of the original
2865 stmt from the pattern that STMT is replacing. I.e, in the example
2866 above we want to use 'widen_sum' in the loop, but 'plus' in the
2868 2. The type (mode) we use to check available target support
2869 for the vector operation to be created in the *epilog*, is
2870 determined by the type of the reduction variable (in the example
2871 above we'd check this: plus_optab[vect_int_mode]).
2872 However the type (mode) we use to check available target support
2873 for the vector operation to be created *inside the loop*, is
2874 determined by the type of the other arguments to STMT (in the
2875 example we'd check this: widen_sum_optab[vect_short_mode]).
2877 This is contrary to "regular" reductions, in which the types of all
2878 the arguments are the same as the type of the reduction variable.
2879 For "regular" reductions we can therefore use the same vector type
2880 (and also the same tree-code) when generating the epilog code and
2881 when generating the code inside the loop. */
2885 /* This is a reduction pattern: get the vectype from the type of the
2886 reduction variable, and get the tree-code from orig_stmt. */
2887 orig_code = TREE_CODE (GIMPLE_STMT_OPERAND (orig_stmt, 1));
2888 vectype = get_vectype_for_scalar_type (TREE_TYPE (def));
2891 if (vect_print_dump_info (REPORT_DETAILS))
2893 fprintf (vect_dump, "unsupported data-type ");
2894 print_generic_expr (vect_dump, TREE_TYPE (def), TDF_SLIM);
2899 vec_mode = TYPE_MODE (vectype);
2903 /* Regular reduction: use the same vectype and tree-code as used for
2904 the vector code inside the loop can be used for the epilog code. */
2908 if (!reduction_code_for_scalar_code (orig_code, &epilog_reduc_code))
2910 reduc_optab = optab_for_tree_code (epilog_reduc_code, vectype);
2913 if (vect_print_dump_info (REPORT_DETAILS))
2914 fprintf (vect_dump, "no optab for reduction.");
2915 epilog_reduc_code = NUM_TREE_CODES;
2917 if (optab_handler (reduc_optab, vec_mode)->insn_code == CODE_FOR_nothing)
2919 if (vect_print_dump_info (REPORT_DETAILS))
2920 fprintf (vect_dump, "reduc op not supported by target.");
2921 epilog_reduc_code = NUM_TREE_CODES;
2924 if (!vec_stmt) /* transformation not required. */
2926 STMT_VINFO_TYPE (stmt_info) = reduc_vec_info_type;
2927 if (!vect_model_reduction_cost (stmt_info, epilog_reduc_code, ncopies))
2934 if (vect_print_dump_info (REPORT_DETAILS))
2935 fprintf (vect_dump, "transform reduction.");
2937 /* Create the destination vector */
2938 vec_dest = vect_create_destination_var (scalar_dest, vectype);
2940 /* Create the reduction-phi that defines the reduction-operand. */
2941 new_phi = create_phi_node (vec_dest, loop->header);
2943 /* In case the vectorization factor (VF) is bigger than the number
2944 of elements that we can fit in a vectype (nunits), we have to generate
2945 more than one vector stmt - i.e - we need to "unroll" the
2946 vector stmt by a factor VF/nunits. For more details see documentation
2947 in vectorizable_operation. */
2949 prev_stmt_info = NULL;
2950 for (j = 0; j < ncopies; j++)
2955 op = TREE_OPERAND (operation, 0);
2956 loop_vec_def0 = vect_get_vec_def_for_operand (op, stmt, NULL);
2957 if (op_type == ternary_op)
2959 op = TREE_OPERAND (operation, 1);
2960 loop_vec_def1 = vect_get_vec_def_for_operand (op, stmt, NULL);
2963 /* Get the vector def for the reduction variable from the phi node */
2964 reduc_def = PHI_RESULT (new_phi);
2968 enum vect_def_type dt = vect_unknown_def_type; /* Dummy */
2969 loop_vec_def0 = vect_get_vec_def_for_stmt_copy (dt, loop_vec_def0);
2970 if (op_type == ternary_op)
2971 loop_vec_def1 = vect_get_vec_def_for_stmt_copy (dt, loop_vec_def1);
2973 /* Get the vector def for the reduction variable from the vectorized
2974 reduction operation generated in the previous iteration (j-1) */
2975 reduc_def = GIMPLE_STMT_OPERAND (new_stmt ,0);
2978 /* Arguments are ready. create the new vector stmt. */
2979 if (op_type == binary_op)
2980 expr = build2 (code, vectype, loop_vec_def0, reduc_def);
2982 expr = build3 (code, vectype, loop_vec_def0, loop_vec_def1,
2984 new_stmt = build_gimple_modify_stmt (vec_dest, expr);
2985 new_temp = make_ssa_name (vec_dest, new_stmt);
2986 GIMPLE_STMT_OPERAND (new_stmt, 0) = new_temp;
2987 vect_finish_stmt_generation (stmt, new_stmt, bsi);
2990 STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt;
2992 STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt;
2993 prev_stmt_info = vinfo_for_stmt (new_stmt);
2996 /* Finalize the reduction-phi (set it's arguments) and create the
2997 epilog reduction code. */
2998 vect_create_epilog_for_reduction (new_temp, stmt, epilog_reduc_code, new_phi);
3002 /* Checks if CALL can be vectorized in type VECTYPE. Returns
3003 a function declaration if the target has a vectorized version
3004 of the function, or NULL_TREE if the function cannot be vectorized. */
3007 vectorizable_function (tree call, tree vectype_out, tree vectype_in)
3009 tree fndecl = get_callee_fndecl (call);
3010 enum built_in_function code;
3012 /* We only handle functions that do not read or clobber memory -- i.e.
3013 const or novops ones. */
3014 if (!(call_expr_flags (call) & (ECF_CONST | ECF_NOVOPS)))
3018 || TREE_CODE (fndecl) != FUNCTION_DECL
3019 || !DECL_BUILT_IN (fndecl))
3022 code = DECL_FUNCTION_CODE (fndecl);
3023 return targetm.vectorize.builtin_vectorized_function (code, vectype_out,
3027 /* Function vectorizable_call.
3029 Check if STMT performs a function call that can be vectorized.
3030 If VEC_STMT is also passed, vectorize the STMT: create a vectorized
3031 stmt to replace it, put it in VEC_STMT, and insert it at BSI.
3032 Return FALSE if not a vectorizable STMT, TRUE otherwise. */
3035 vectorizable_call (tree stmt, block_stmt_iterator *bsi, tree *vec_stmt)
3041 tree vec_oprnd0 = NULL_TREE, vec_oprnd1 = NULL_TREE;
3042 stmt_vec_info stmt_info = vinfo_for_stmt (stmt), prev_stmt_info;
3043 tree vectype_out, vectype_in;
3046 loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info);
3047 struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
3048 tree fndecl, rhs, new_temp, def, def_stmt, rhs_type, lhs_type;
3049 enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type};
3051 int ncopies, j, nargs;
3052 call_expr_arg_iterator iter;
3054 enum { NARROW, NONE, WIDEN } modifier;
3056 if (!STMT_VINFO_RELEVANT_P (stmt_info))
3059 if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_loop_def)
3062 /* FORNOW: SLP not supported. */
3063 if (STMT_SLP_TYPE (stmt_info))
3066 /* Is STMT a vectorizable call? */
3067 if (TREE_CODE (stmt) != GIMPLE_MODIFY_STMT)
3070 if (TREE_CODE (GIMPLE_STMT_OPERAND (stmt, 0)) != SSA_NAME)
3073 operation = GIMPLE_STMT_OPERAND (stmt, 1);
3074 if (TREE_CODE (operation) != CALL_EXPR)
3077 /* Process function arguments. */
3078 rhs_type = NULL_TREE;
3080 FOR_EACH_CALL_EXPR_ARG (op, iter, operation)
3082 /* Bail out if the function has more than two arguments, we
3083 do not have interesting builtin functions to vectorize with
3084 more than two arguments. */
3088 /* We can only handle calls with arguments of the same type. */
3090 && rhs_type != TREE_TYPE (op))
3092 if (vect_print_dump_info (REPORT_DETAILS))
3093 fprintf (vect_dump, "argument types differ.");
3096 rhs_type = TREE_TYPE (op);
3098 if (!vect_is_simple_use (op, loop_vinfo, &def_stmt, &def, &dt[nargs]))
3100 if (vect_print_dump_info (REPORT_DETAILS))
3101 fprintf (vect_dump, "use not simple.");
3108 /* No arguments is also not good. */
3112 vectype_in = get_vectype_for_scalar_type (rhs_type);
3115 nunits_in = TYPE_VECTOR_SUBPARTS (vectype_in);
3117 lhs_type = TREE_TYPE (GIMPLE_STMT_OPERAND (stmt, 0));
3118 vectype_out = get_vectype_for_scalar_type (lhs_type);
3121 nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out);
3124 if (nunits_in == nunits_out / 2)
3126 else if (nunits_out == nunits_in)
3128 else if (nunits_out == nunits_in / 2)
3133 /* For now, we only vectorize functions if a target specific builtin
3134 is available. TODO -- in some cases, it might be profitable to
3135 insert the calls for pieces of the vector, in order to be able
3136 to vectorize other operations in the loop. */
3137 fndecl = vectorizable_function (operation, vectype_out, vectype_in);
3138 if (fndecl == NULL_TREE)
3140 if (vect_print_dump_info (REPORT_DETAILS))
3141 fprintf (vect_dump, "function is not vectorizable.");
3146 gcc_assert (ZERO_SSA_OPERANDS (stmt, SSA_OP_ALL_VIRTUALS));
3148 if (modifier == NARROW)
3149 ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_out;
3151 ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in;
3153 /* Sanity check: make sure that at least one copy of the vectorized stmt
3154 needs to be generated. */
3155 gcc_assert (ncopies >= 1);
3157 /* FORNOW. This restriction should be relaxed. */
3158 if (nested_in_vect_loop_p (loop, stmt) && ncopies > 1)
3160 if (vect_print_dump_info (REPORT_DETAILS))
3161 fprintf (vect_dump, "multiple types in nested loop.");
3165 if (!vec_stmt) /* transformation not required. */
3167 STMT_VINFO_TYPE (stmt_info) = call_vec_info_type;
3168 if (vect_print_dump_info (REPORT_DETAILS))
3169 fprintf (vect_dump, "=== vectorizable_call ===");
3170 vect_model_simple_cost (stmt_info, ncopies, dt, NULL);
3176 if (vect_print_dump_info (REPORT_DETAILS))
3177 fprintf (vect_dump, "transform operation.");
3179 /* FORNOW. This restriction should be relaxed. */
3180 if (nested_in_vect_loop_p (loop, stmt) && ncopies > 1)
3182 if (vect_print_dump_info (REPORT_DETAILS))
3183 fprintf (vect_dump, "multiple types in nested loop.");
3188 scalar_dest = GIMPLE_STMT_OPERAND (stmt, 0);
3189 vec_dest = vect_create_destination_var (scalar_dest, vectype_out);
3191 prev_stmt_info = NULL;
3195 for (j = 0; j < ncopies; ++j)
3197 /* Build argument list for the vectorized call. */
3198 /* FIXME: Rewrite this so that it doesn't
3199 construct a temporary list. */
3202 FOR_EACH_CALL_EXPR_ARG (op, iter, operation)
3206 = vect_get_vec_def_for_operand (op, stmt, NULL);
3209 = vect_get_vec_def_for_stmt_copy (dt[nargs], vec_oprnd0);
3211 vargs = tree_cons (NULL_TREE, vec_oprnd0, vargs);
3215 vargs = nreverse (vargs);
3217 rhs = build_function_call_expr (fndecl, vargs);
3218 new_stmt = build_gimple_modify_stmt (vec_dest, rhs);
3219 new_temp = make_ssa_name (vec_dest, new_stmt);
3220 GIMPLE_STMT_OPERAND (new_stmt, 0) = new_temp;
3222 vect_finish_stmt_generation (stmt, new_stmt, bsi);
3225 STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt;
3227 STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt;
3229 prev_stmt_info = vinfo_for_stmt (new_stmt);
3235 for (j = 0; j < ncopies; ++j)
3237 /* Build argument list for the vectorized call. */
3238 /* FIXME: Rewrite this so that it doesn't
3239 construct a temporary list. */
3242 FOR_EACH_CALL_EXPR_ARG (op, iter, operation)
3247 = vect_get_vec_def_for_operand (op, stmt, NULL);
3249 = vect_get_vec_def_for_stmt_copy (dt[nargs], vec_oprnd0);
3254 = vect_get_vec_def_for_stmt_copy (dt[nargs], vec_oprnd1);
3256 = vect_get_vec_def_for_stmt_copy (dt[nargs], vec_oprnd0);
3259 vargs = tree_cons (NULL_TREE, vec_oprnd0, vargs);
3260 vargs = tree_cons (NULL_TREE, vec_oprnd1, vargs);
3264 vargs = nreverse (vargs);
3266 rhs = build_function_call_expr (fndecl, vargs);
3267 new_stmt = build_gimple_modify_stmt (vec_dest, rhs);
3268 new_temp = make_ssa_name (vec_dest, new_stmt);
3269 GIMPLE_STMT_OPERAND (new_stmt, 0) = new_temp;
3271 vect_finish_stmt_generation (stmt, new_stmt, bsi);
3274 STMT_VINFO_VEC_STMT (stmt_info) = new_stmt;
3276 STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt;
3278 prev_stmt_info = vinfo_for_stmt (new_stmt);
3281 *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info);
3286 /* No current target implements this case. */
3290 /* The call in STMT might prevent it from being removed in dce.
3291 We however cannot remove it here, due to the way the ssa name
3292 it defines is mapped to the new definition. So just replace
3293 rhs of the statement with something harmless. */
3294 type = TREE_TYPE (scalar_dest);
3295 GIMPLE_STMT_OPERAND (stmt, 1) = fold_convert (type, integer_zero_node);
3302 /* Function vect_gen_widened_results_half
3304 Create a vector stmt whose code, type, number of arguments, and result
3305 variable are CODE, VECTYPE, OP_TYPE, and VEC_DEST, and its arguments are
3306 VEC_OPRND0 and VEC_OPRND1. The new vector stmt is to be inserted at BSI.
3307 In the case that CODE is a CALL_EXPR, this means that a call to DECL
3308 needs to be created (DECL is a function-decl of a target-builtin).
3309 STMT is the original scalar stmt that we are vectorizing. */
3312 vect_gen_widened_results_half (enum tree_code code, tree vectype, tree decl,
3313 tree vec_oprnd0, tree vec_oprnd1, int op_type,
3314 tree vec_dest, block_stmt_iterator *bsi,
3323 /* Generate half of the widened result: */
3324 if (code == CALL_EXPR)
3326 /* Target specific support */
3327 if (op_type == binary_op)
3328 expr = build_call_expr (decl, 2, vec_oprnd0, vec_oprnd1);
3330 expr = build_call_expr (decl, 1, vec_oprnd0);
3334 /* Generic support */
3335 gcc_assert (op_type == TREE_CODE_LENGTH (code));
3336 if (op_type == binary_op)
3337 expr = build2 (code, vectype, vec_oprnd0, vec_oprnd1);
3339 expr = build1 (code, vectype, vec_oprnd0);
3341 new_stmt = build_gimple_modify_stmt (vec_dest, expr);
3342 new_temp = make_ssa_name (vec_dest, new_stmt);
3343 GIMPLE_STMT_OPERAND (new_stmt, 0) = new_temp;
3344 vect_finish_stmt_generation (stmt, new_stmt, bsi);
3346 if (code == CALL_EXPR)
3348 FOR_EACH_SSA_TREE_OPERAND (sym, new_stmt, iter, SSA_OP_ALL_VIRTUALS)
3350 if (TREE_CODE (sym) == SSA_NAME)
3351 sym = SSA_NAME_VAR (sym);
3352 mark_sym_for_renaming (sym);
3360 /* Check if STMT performs a conversion operation, that can be vectorized.
3361 If VEC_STMT is also passed, vectorize the STMT: create a vectorized
3362 stmt to replace it, put it in VEC_STMT, and insert it at BSI.
3363 Return FALSE if not a vectorizable STMT, TRUE otherwise. */
3366 vectorizable_conversion (tree stmt, block_stmt_iterator *bsi,
3367 tree *vec_stmt, slp_tree slp_node)
3373 tree vec_oprnd0 = NULL_TREE, vec_oprnd1 = NULL_TREE;
3374 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
3375 loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info);
3376 struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
3377 enum tree_code code, code1 = ERROR_MARK, code2 = ERROR_MARK;
3378 tree decl1 = NULL_TREE, decl2 = NULL_TREE;
3381 enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type};
3382 tree new_stmt = NULL_TREE;
3383 stmt_vec_info prev_stmt_info;
3386 tree vectype_out, vectype_in;
3389 tree rhs_type, lhs_type;
3391 enum { NARROW, NONE, WIDEN } modifier;
3393 VEC(tree,heap) *vec_oprnds0 = NULL;
3396 /* Is STMT a vectorizable conversion? */
3398 if (!STMT_VINFO_RELEVANT_P (stmt_info))
3401 if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_loop_def)
3404 if (TREE_CODE (stmt) != GIMPLE_MODIFY_STMT)
3407 if (TREE_CODE (GIMPLE_STMT_OPERAND (stmt, 0)) != SSA_NAME)
3410 operation = GIMPLE_STMT_OPERAND (stmt, 1);
3411 code = TREE_CODE (operation);
3412 if (code != FIX_TRUNC_EXPR && code != FLOAT_EXPR)
3415 /* Check types of lhs and rhs. */
3416 op0 = TREE_OPERAND (operation, 0);
3417 rhs_type = TREE_TYPE (op0);
3418 vectype_in = get_vectype_for_scalar_type (rhs_type);
3421 nunits_in = TYPE_VECTOR_SUBPARTS (vectype_in);
3423 scalar_dest = GIMPLE_STMT_OPERAND (stmt, 0);
3424 lhs_type = TREE_TYPE (scalar_dest);
3425 vectype_out = get_vectype_for_scalar_type (lhs_type);
3428 nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out);
3431 if (nunits_in == nunits_out / 2)
3433 else if (nunits_out == nunits_in)
3435 else if (nunits_out == nunits_in / 2)
3440 if (modifier == NONE)
3441 gcc_assert (STMT_VINFO_VECTYPE (stmt_info) == vectype_out);
3443 /* Bail out if the types are both integral or non-integral. */
3444 if ((INTEGRAL_TYPE_P (rhs_type) && INTEGRAL_TYPE_P (lhs_type))
3445 || (!INTEGRAL_TYPE_P (rhs_type) && !INTEGRAL_TYPE_P (lhs_type)))
3448 if (modifier == NARROW)
3449 ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_out;
3451 ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in;
3453 /* FORNOW: SLP with multiple types is not supported. The SLP analysis verifies
3454 this, so we can safely override NCOPIES with 1 here. */
3458 /* Sanity check: make sure that at least one copy of the vectorized stmt
3459 needs to be generated. */
3460 gcc_assert (ncopies >= 1);
3462 /* FORNOW. This restriction should be relaxed. */
3463 if (nested_in_vect_loop_p (loop, stmt) && ncopies > 1)
3465 if (vect_print_dump_info (REPORT_DETAILS))
3466 fprintf (vect_dump, "multiple types in nested loop.");
3470 /* Check the operands of the operation. */
3471 if (!vect_is_simple_use (op0, loop_vinfo, &def_stmt, &def, &dt[0]))
3473 if (vect_print_dump_info (REPORT_DETAILS))
3474 fprintf (vect_dump, "use not simple.");
3478 /* Supportable by target? */
3479 if ((modifier == NONE
3480 && !targetm.vectorize.builtin_conversion (code, vectype_in))
3481 || (modifier == WIDEN
3482 && !supportable_widening_operation (code, stmt, vectype_in,
3485 || (modifier == NARROW
3486 && !supportable_narrowing_operation (code, stmt, vectype_in,
3489 if (vect_print_dump_info (REPORT_DETAILS))
3490 fprintf (vect_dump, "op not supported by target.");
3494 if (modifier != NONE)
3496 STMT_VINFO_VECTYPE (stmt_info) = vectype_in;
3497 /* FORNOW: SLP not supported. */
3498 if (STMT_SLP_TYPE (stmt_info))
3502 if (!vec_stmt) /* transformation not required. */
3504 STMT_VINFO_TYPE (stmt_info) = type_conversion_vec_info_type;
3509 if (vect_print_dump_info (REPORT_DETAILS))
3510 fprintf (vect_dump, "transform conversion.");
3513 vec_dest = vect_create_destination_var (scalar_dest, vectype_out);
3515 if (modifier == NONE && !slp_node)
3516 vec_oprnds0 = VEC_alloc (tree, heap, 1);
3518 prev_stmt_info = NULL;
3522 for (j = 0; j < ncopies; j++)
3528 vect_get_vec_defs (op0, NULL, stmt, &vec_oprnds0, NULL, slp_node);
3530 vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds0, NULL);
3533 targetm.vectorize.builtin_conversion (code, vectype_in);
3534 for (i = 0; VEC_iterate (tree, vec_oprnds0, i, vop0); i++)
3536 new_stmt = build_call_expr (builtin_decl, 1, vop0);
3538 /* Arguments are ready. create the new vector stmt. */
3539 new_stmt = build_gimple_modify_stmt (vec_dest, new_stmt);
3540 new_temp = make_ssa_name (vec_dest, new_stmt);
3541 GIMPLE_STMT_OPERAND (new_stmt, 0) = new_temp;
3542 vect_finish_stmt_generation (stmt, new_stmt, bsi);
3543 FOR_EACH_SSA_TREE_OPERAND (sym, new_stmt, iter,
3544 SSA_OP_ALL_VIRTUALS)
3546 if (TREE_CODE (sym) == SSA_NAME)
3547 sym = SSA_NAME_VAR (sym);
3548 mark_sym_for_renaming (sym);
3551 VEC_quick_push (tree, SLP_TREE_VEC_STMTS (slp_node), new_stmt);
3555 STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt;
3557 STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt;
3558 prev_stmt_info = vinfo_for_stmt (new_stmt);
3563 /* In case the vectorization factor (VF) is bigger than the number
3564 of elements that we can fit in a vectype (nunits), we have to
3565 generate more than one vector stmt - i.e - we need to "unroll"
3566 the vector stmt by a factor VF/nunits. */
3567 for (j = 0; j < ncopies; j++)
3570 vec_oprnd0 = vect_get_vec_def_for_operand (op0, stmt, NULL);
3572 vec_oprnd0 = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd0);
3574 STMT_VINFO_VECTYPE (stmt_info) = vectype_in;
3576 /* Generate first half of the widened result: */
3578 = vect_gen_widened_results_half (code1, vectype_out, decl1,
3579 vec_oprnd0, vec_oprnd1,
3580 unary_op, vec_dest, bsi, stmt);
3582 STMT_VINFO_VEC_STMT (stmt_info) = new_stmt;
3584 STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt;
3585 prev_stmt_info = vinfo_for_stmt (new_stmt);
3587 /* Generate second half of the widened result: */
3589 = vect_gen_widened_results_half (code2, vectype_out, decl2,
3590 vec_oprnd0, vec_oprnd1,
3591 unary_op, vec_dest, bsi, stmt);
3592 STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt;
3593 prev_stmt_info = vinfo_for_stmt (new_stmt);
3598 /* In case the vectorization factor (VF) is bigger than the number
3599 of elements that we can fit in a vectype (nunits), we have to
3600 generate more than one vector stmt - i.e - we need to "unroll"
3601 the vector stmt by a factor VF/nunits. */
3602 for (j = 0; j < ncopies; j++)
3607 vec_oprnd0 = vect_get_vec_def_for_operand (op0, stmt, NULL);
3608 vec_oprnd1 = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd0);
3612 vec_oprnd0 = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd1);
3613 vec_oprnd1 = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd0);
3616 /* Arguments are ready. Create the new vector stmt. */
3617 expr = build2 (code1, vectype_out, vec_oprnd0, vec_oprnd1);
3618 new_stmt = build_gimple_modify_stmt (vec_dest, expr);
3619 new_temp = make_ssa_name (vec_dest, new_stmt);
3620 GIMPLE_STMT_OPERAND (new_stmt, 0) = new_temp;
3621 vect_finish_stmt_generation (stmt, new_stmt, bsi);
3624 STMT_VINFO_VEC_STMT (stmt_info) = new_stmt;
3626 STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt;
3628 prev_stmt_info = vinfo_for_stmt (new_stmt);
3631 *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info);
3638 /* Function vectorizable_assignment.
3640 Check if STMT performs an assignment (copy) that can be vectorized.
3641 If VEC_STMT is also passed, vectorize the STMT: create a vectorized
3642 stmt to replace it, put it in VEC_STMT, and insert it at BSI.
3643 Return FALSE if not a vectorizable STMT, TRUE otherwise. */
3646 vectorizable_assignment (tree stmt, block_stmt_iterator *bsi, tree *vec_stmt,
3652 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
3653 tree vectype = STMT_VINFO_VECTYPE (stmt_info);
3654 loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info);
3657 enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type};
3658 int nunits = TYPE_VECTOR_SUBPARTS (vectype);
3659 int ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits;
3661 VEC(tree,heap) *vec_oprnds = NULL;
3664 gcc_assert (ncopies >= 1);
3666 return false; /* FORNOW */
3668 if (!STMT_VINFO_RELEVANT_P (stmt_info))
3671 if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_loop_def)
3674 /* Is vectorizable assignment? */
3675 if (TREE_CODE (stmt) != GIMPLE_MODIFY_STMT)
3678 scalar_dest = GIMPLE_STMT_OPERAND (stmt, 0);
3679 if (TREE_CODE (scalar_dest) != SSA_NAME)
3682 op = GIMPLE_STMT_OPERAND (stmt, 1);
3683 if (!vect_is_simple_use (op, loop_vinfo, &def_stmt, &def, &dt[0]))
3685 if (vect_print_dump_info (REPORT_DETAILS))
3686 fprintf (vect_dump, "use not simple.");
3690 if (!vec_stmt) /* transformation not required. */
3692 STMT_VINFO_TYPE (stmt_info) = assignment_vec_info_type;
3693 if (vect_print_dump_info (REPORT_DETAILS))
3694 fprintf (vect_dump, "=== vectorizable_assignment ===");
3695 vect_model_simple_cost (stmt_info, ncopies, dt, NULL);
3700 if (vect_print_dump_info (REPORT_DETAILS))
3701 fprintf (vect_dump, "transform assignment.");
3704 vec_dest = vect_create_destination_var (scalar_dest, vectype);
3707 vect_get_vec_defs (op, NULL, stmt, &vec_oprnds, NULL, slp_node);
3709 /* Arguments are ready. create the new vector stmt. */
3710 for (i = 0; VEC_iterate (tree, vec_oprnds, i, vop); i++)
3712 *vec_stmt = build_gimple_modify_stmt (vec_dest, vop);
3713 new_temp = make_ssa_name (vec_dest, *vec_stmt);
3714 GIMPLE_STMT_OPERAND (*vec_stmt, 0) = new_temp;
3715 vect_finish_stmt_generation (stmt, *vec_stmt, bsi);
3716 STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt;
3719 VEC_quick_push (tree, SLP_TREE_VEC_STMTS (slp_node), *vec_stmt);
3722 VEC_free (tree, heap, vec_oprnds);
3727 /* Function vect_min_worthwhile_factor.
3729 For a loop where we could vectorize the operation indicated by CODE,
3730 return the minimum vectorization factor that makes it worthwhile
3731 to use generic vectors. */
3733 vect_min_worthwhile_factor (enum tree_code code)
3754 /* Function vectorizable_induction
3756 Check if PHI performs an induction computation that can be vectorized.
3757 If VEC_STMT is also passed, vectorize the induction PHI: create a vectorized
3758 phi to replace it, put it in VEC_STMT, and add it to the same basic block.
3759 Return FALSE if not a vectorizable STMT, TRUE otherwise. */
3762 vectorizable_induction (tree phi, block_stmt_iterator *bsi ATTRIBUTE_UNUSED,
3765 stmt_vec_info stmt_info = vinfo_for_stmt (phi);
3766 tree vectype = STMT_VINFO_VECTYPE (stmt_info);
3767 loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info);
3768 int nunits = TYPE_VECTOR_SUBPARTS (vectype);
3769 int ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits;
3772 gcc_assert (ncopies >= 1);
3774 if (!STMT_VINFO_RELEVANT_P (stmt_info))
3777 /* FORNOW: SLP not supported. */
3778 if (STMT_SLP_TYPE (stmt_info))
3781 gcc_assert (STMT_VINFO_DEF_TYPE (stmt_info) == vect_induction_def);
3783 if (TREE_CODE (phi) != PHI_NODE)
3786 if (!vec_stmt) /* transformation not required. */
3788 STMT_VINFO_TYPE (stmt_info) = induc_vec_info_type;
3789 if (vect_print_dump_info (REPORT_DETAILS))
3790 fprintf (vect_dump, "=== vectorizable_induction ===");
3791 vect_model_induction_cost (stmt_info, ncopies);
3797 if (vect_print_dump_info (REPORT_DETAILS))
3798 fprintf (vect_dump, "transform induction phi.");
3800 vec_def = get_initial_def_for_induction (phi);
3801 *vec_stmt = SSA_NAME_DEF_STMT (vec_def);
3806 /* Function vectorizable_operation.
3808 Check if STMT performs a binary or unary operation that can be vectorized.
3809 If VEC_STMT is also passed, vectorize the STMT: create a vectorized
3810 stmt to replace it, put it in VEC_STMT, and insert it at BSI.
3811 Return FALSE if not a vectorizable STMT, TRUE otherwise. */
3814 vectorizable_operation (tree stmt, block_stmt_iterator *bsi, tree *vec_stmt,
3820 tree op0, op1 = NULL;
3821 tree vec_oprnd1 = NULL_TREE;
3822 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
3823 tree vectype = STMT_VINFO_VECTYPE (stmt_info);
3824 loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info);
3825 struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
3826 enum tree_code code;
3827 enum machine_mode vec_mode;
3832 enum machine_mode optab_op2_mode;
3834 enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type};
3835 tree new_stmt = NULL_TREE;
3836 stmt_vec_info prev_stmt_info;
3837 int nunits_in = TYPE_VECTOR_SUBPARTS (vectype);
3840 int ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in;
3842 VEC(tree,heap) *vec_oprnds0 = NULL, *vec_oprnds1 = NULL;
3845 bool scalar_shift_arg = false;
3847 /* FORNOW: SLP with multiple types is not supported. The SLP analysis verifies
3848 this, so we can safely override NCOPIES with 1 here. */
3851 gcc_assert (ncopies >= 1);
3852 /* FORNOW. This restriction should be relaxed. */
3853 if (nested_in_vect_loop_p (loop, stmt) && ncopies > 1)
3855 if (vect_print_dump_info (REPORT_DETAILS))
3856 fprintf (vect_dump, "multiple types in nested loop.");
3860 if (!STMT_VINFO_RELEVANT_P (stmt_info))
3863 if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_loop_def)
3866 /* Is STMT a vectorizable binary/unary operation? */
3867 if (TREE_CODE (stmt) != GIMPLE_MODIFY_STMT)
3870 if (TREE_CODE (GIMPLE_STMT_OPERAND (stmt, 0)) != SSA_NAME)
3873 scalar_dest = GIMPLE_STMT_OPERAND (stmt, 0);
3874 vectype_out = get_vectype_for_scalar_type (TREE_TYPE (scalar_dest));
3877 nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out);
3878 if (nunits_out != nunits_in)
3881 operation = GIMPLE_STMT_OPERAND (stmt, 1);
3882 code = TREE_CODE (operation);
3884 /* For pointer addition, we should use the normal plus for
3885 the vector addition. */
3886 if (code == POINTER_PLUS_EXPR)
3889 optab = optab_for_tree_code (code, vectype);
3891 /* Support only unary or binary operations. */
3892 op_type = TREE_OPERAND_LENGTH (operation);
3893 if (op_type != unary_op && op_type != binary_op)
3895 if (vect_print_dump_info (REPORT_DETAILS))
3896 fprintf (vect_dump, "num. args = %d (not unary/binary op).", op_type);
3900 op0 = TREE_OPERAND (operation, 0);
3901 if (!vect_is_simple_use (op0, loop_vinfo, &def_stmt, &def, &dt[0]))
3903 if (vect_print_dump_info (REPORT_DETAILS))
3904 fprintf (vect_dump, "use not simple.");
3908 if (op_type == binary_op)
3910 op1 = TREE_OPERAND (operation, 1);
3911 if (!vect_is_simple_use (op1, loop_vinfo, &def_stmt, &def, &dt[1]))
3913 if (vect_print_dump_info (REPORT_DETAILS))
3914 fprintf (vect_dump, "use not simple.");
3919 /* Supportable by target? */
3922 if (vect_print_dump_info (REPORT_DETAILS))
3923 fprintf (vect_dump, "no optab.");
3926 vec_mode = TYPE_MODE (vectype);
3927 icode = (int) optab_handler (optab, vec_mode)->insn_code;
3928 if (icode == CODE_FOR_nothing)
3930 if (vect_print_dump_info (REPORT_DETAILS))
3931 fprintf (vect_dump, "op not supported by target.");
3932 /* Check only during analysis. */
3933 if (GET_MODE_SIZE (vec_mode) != UNITS_PER_WORD
3934 || (LOOP_VINFO_VECT_FACTOR (loop_vinfo)
3935 < vect_min_worthwhile_factor (code)
3938 if (vect_print_dump_info (REPORT_DETAILS))
3939 fprintf (vect_dump, "proceeding using word mode.");
3942 /* Worthwhile without SIMD support? Check only during analysis. */
3943 if (!VECTOR_MODE_P (TYPE_MODE (vectype))
3944 && LOOP_VINFO_VECT_FACTOR (loop_vinfo)
3945 < vect_min_worthwhile_factor (code)
3948 if (vect_print_dump_info (REPORT_DETAILS))
3949 fprintf (vect_dump, "not worthwhile without SIMD support.");
3953 if (code == LSHIFT_EXPR || code == RSHIFT_EXPR)
3955 /* FORNOW: not yet supported. */
3956 if (!VECTOR_MODE_P (vec_mode))
3959 /* Invariant argument is needed for a vector shift
3960 by a scalar shift operand. */
3961 optab_op2_mode = insn_data[icode].operand[2].mode;
3962 if (!VECTOR_MODE_P (optab_op2_mode))
3964 if (dt[1] != vect_constant_def && dt[1] != vect_invariant_def)
3966 if (vect_print_dump_info (REPORT_DETAILS))
3967 fprintf (vect_dump, "operand mode requires invariant"
3972 scalar_shift_arg = true;
3976 if (!vec_stmt) /* transformation not required. */
3978 STMT_VINFO_TYPE (stmt_info) = op_vec_info_type;
3979 if (vect_print_dump_info (REPORT_DETAILS))
3980 fprintf (vect_dump, "=== vectorizable_operation ===");
3981 vect_model_simple_cost (stmt_info, ncopies, dt, NULL);
3987 if (vect_print_dump_info (REPORT_DETAILS))
3988 fprintf (vect_dump, "transform binary/unary operation.");
3991 vec_dest = vect_create_destination_var (scalar_dest, vectype);
3993 /* Allocate VECs for vector operands. In case of SLP, vector operands are
3994 created in the previous stages of the recursion, so no allocation is
3995 needed, except for the case of shift with scalar shift argument. In that
3996 case we store the scalar operand in VEC_OPRNDS1 for every vector stmt to
3997 be created to vectorize the SLP group, i.e., SLP_NODE->VEC_STMTS_SIZE.
3998 In case of loop-based vectorization we allocate VECs of size 1. We
3999 allocate VEC_OPRNDS1 only in case of binary operation. */
4002 vec_oprnds0 = VEC_alloc (tree, heap, 1);
4003 if (op_type == binary_op)
4004 vec_oprnds1 = VEC_alloc (tree, heap, 1);
4006 else if (scalar_shift_arg)
4007 vec_oprnds1 = VEC_alloc (tree, heap, slp_node->vec_stmts_size);
4009 /* In case the vectorization factor (VF) is bigger than the number
4010 of elements that we can fit in a vectype (nunits), we have to generate
4011 more than one vector stmt - i.e - we need to "unroll" the
4012 vector stmt by a factor VF/nunits. In doing so, we record a pointer
4013 from one copy of the vector stmt to the next, in the field
4014 STMT_VINFO_RELATED_STMT. This is necessary in order to allow following
4015 stages to find the correct vector defs to be used when vectorizing
4016 stmts that use the defs of the current stmt. The example below illustrates
4017 the vectorization process when VF=16 and nunits=4 (i.e - we need to create
4018 4 vectorized stmts):
4020 before vectorization:
4021 RELATED_STMT VEC_STMT
4025 step 1: vectorize stmt S1 (done in vectorizable_load. See more details
4027 RELATED_STMT VEC_STMT
4028 VS1_0: vx0 = memref0 VS1_1 -
4029 VS1_1: vx1 = memref1 VS1_2 -
4030 VS1_2: vx2 = memref2 VS1_3 -
4031 VS1_3: vx3 = memref3 - -
4032 S1: x = load - VS1_0
4035 step2: vectorize stmt S2 (done here):
4036 To vectorize stmt S2 we first need to find the relevant vector
4037 def for the first operand 'x'. This is, as usual, obtained from
4038 the vector stmt recorded in the STMT_VINFO_VEC_STMT of the stmt
4039 that defines 'x' (S1). This way we find the stmt VS1_0, and the
4040 relevant vector def 'vx0'. Having found 'vx0' we can generate
4041 the vector stmt VS2_0, and as usual, record it in the
4042 STMT_VINFO_VEC_STMT of stmt S2.
4043 When creating the second copy (VS2_1), we obtain the relevant vector
4044 def from the vector stmt recorded in the STMT_VINFO_RELATED_STMT of
4045 stmt VS1_0. This way we find the stmt VS1_1 and the relevant
4046 vector def 'vx1'. Using 'vx1' we create stmt VS2_1 and record a
4047 pointer to it in the STMT_VINFO_RELATED_STMT of the vector stmt VS2_0.
4048 Similarly when creating stmts VS2_2 and VS2_3. This is the resulting
4049 chain of stmts and pointers:
4050 RELATED_STMT VEC_STMT
4051 VS1_0: vx0 = memref0 VS1_1 -
4052 VS1_1: vx1 = memref1 VS1_2 -
4053 VS1_2: vx2 = memref2 VS1_3 -
4054 VS1_3: vx3 = memref3 - -
4055 S1: x = load - VS1_0
4056 VS2_0: vz0 = vx0 + v1 VS2_1 -
4057 VS2_1: vz1 = vx1 + v1 VS2_2 -
4058 VS2_2: vz2 = vx2 + v1 VS2_3 -
4059 VS2_3: vz3 = vx3 + v1 - -
4060 S2: z = x + 1 - VS2_0 */
4062 prev_stmt_info = NULL;
4063 for (j = 0; j < ncopies; j++)
4068 if (op_type == binary_op
4069 && (code == LSHIFT_EXPR || code == RSHIFT_EXPR))
4071 /* Vector shl and shr insn patterns can be defined with scalar
4072 operand 2 (shift operand). In this case, use constant or loop
4073 invariant op1 directly, without extending it to vector mode
4075 optab_op2_mode = insn_data[icode].operand[2].mode;
4076 if (!VECTOR_MODE_P (optab_op2_mode))
4078 if (vect_print_dump_info (REPORT_DETAILS))
4079 fprintf (vect_dump, "operand 1 using scalar mode.");
4081 VEC_quick_push (tree, vec_oprnds1, vec_oprnd1);
4084 /* Store vec_oprnd1 for every vector stmt to be created
4085 for SLP_NODE. We check during the analysis that all the
4086 shift arguments are the same.
4087 TODO: Allow different constants for different vector
4088 stmts generated for an SLP instance. */
4089 for (k = 0; k < slp_node->vec_stmts_size - 1; k++)
4090 VEC_quick_push (tree, vec_oprnds1, vec_oprnd1);
4095 /* vec_oprnd1 is available if operand 1 should be of a scalar-type
4096 (a special case for certain kind of vector shifts); otherwise,
4097 operand 1 should be of a vector type (the usual case). */
4098 if (op_type == binary_op && !vec_oprnd1)
4099 vect_get_vec_defs (op0, op1, stmt, &vec_oprnds0, &vec_oprnds1,
4102 vect_get_vec_defs (op0, NULL_TREE, stmt, &vec_oprnds0, NULL,
4106 vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds0, &vec_oprnds1);
4108 /* Arguments are ready. Create the new vector stmt. */
4109 for (i = 0; VEC_iterate (tree, vec_oprnds0, i, vop0); i++)
4111 if (op_type == binary_op)
4113 vop1 = VEC_index (tree, vec_oprnds1, i);
4114 new_stmt = build_gimple_modify_stmt (vec_dest,
4115 build2 (code, vectype, vop0, vop1));
4118 new_stmt = build_gimple_modify_stmt (vec_dest,
4119 build1 (code, vectype, vop0));
4121 new_temp = make_ssa_name (vec_dest, new_stmt);
4122 GIMPLE_STMT_OPERAND (new_stmt, 0) = new_temp;
4123 vect_finish_stmt_generation (stmt, new_stmt, bsi);
4125 VEC_quick_push (tree, SLP_TREE_VEC_STMTS (slp_node), new_stmt);
4129 STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt;
4131 STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt;
4132 prev_stmt_info = vinfo_for_stmt (new_stmt);
4135 VEC_free (tree, heap, vec_oprnds0);
4137 VEC_free (tree, heap, vec_oprnds1);
4143 /* Function vectorizable_type_demotion
4145 Check if STMT performs a binary or unary operation that involves
4146 type demotion, and if it can be vectorized.
4147 If VEC_STMT is also passed, vectorize the STMT: create a vectorized
4148 stmt to replace it, put it in VEC_STMT, and insert it at BSI.
4149 Return FALSE if not a vectorizable STMT, TRUE otherwise. */
4152 vectorizable_type_demotion (tree stmt, block_stmt_iterator *bsi,
4159 tree vec_oprnd0=NULL, vec_oprnd1=NULL;
4160 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
4161 loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info);
4162 struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
4163 enum tree_code code, code1 = ERROR_MARK;
4166 enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type};
4168 stmt_vec_info prev_stmt_info;
4177 if (!STMT_VINFO_RELEVANT_P (stmt_info))
4180 if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_loop_def)
4183 /* Is STMT a vectorizable type-demotion operation? */
4184 if (TREE_CODE (stmt) != GIMPLE_MODIFY_STMT)
4187 if (TREE_CODE (GIMPLE_STMT_OPERAND (stmt, 0)) != SSA_NAME)
4190 operation = GIMPLE_STMT_OPERAND (stmt, 1);
4191 code = TREE_CODE (operation);
4192 if (code != NOP_EXPR && code != CONVERT_EXPR)
4195 op0 = TREE_OPERAND (operation, 0);
4196 vectype_in = get_vectype_for_scalar_type (TREE_TYPE (op0));
4199 nunits_in = TYPE_VECTOR_SUBPARTS (vectype_in);
4201 scalar_dest = GIMPLE_STMT_OPERAND (stmt, 0);
4202 vectype_out = get_vectype_for_scalar_type (TREE_TYPE (scalar_dest));
4205 nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out);
4206 if (nunits_in != nunits_out / 2) /* FORNOW */
4209 ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_out;
4210 gcc_assert (ncopies >= 1);
4211 /* FORNOW. This restriction should be relaxed. */
4212 if (nested_in_vect_loop_p (loop, stmt) && ncopies > 1)
4214 if (vect_print_dump_info (REPORT_DETAILS))
4215 fprintf (vect_dump, "multiple types in nested loop.");
4219 if (! ((INTEGRAL_TYPE_P (TREE_TYPE (scalar_dest))
4220 && INTEGRAL_TYPE_P (TREE_TYPE (op0)))
4221 || (SCALAR_FLOAT_TYPE_P (TREE_TYPE (scalar_dest))
4222 && SCALAR_FLOAT_TYPE_P (TREE_TYPE (op0))
4223 && (code == NOP_EXPR || code == CONVERT_EXPR))))
4226 /* Check the operands of the operation. */
4227 if (!vect_is_simple_use (op0, loop_vinfo, &def_stmt, &def, &dt[0]))
4229 if (vect_print_dump_info (REPORT_DETAILS))
4230 fprintf (vect_dump, "use not simple.");
4234 /* Supportable by target? */
4235 if (!supportable_narrowing_operation (code, stmt, vectype_in, &code1))
4238 STMT_VINFO_VECTYPE (stmt_info) = vectype_in;
4240 if (!vec_stmt) /* transformation not required. */
4242 STMT_VINFO_TYPE (stmt_info) = type_demotion_vec_info_type;
4243 if (vect_print_dump_info (REPORT_DETAILS))
4244 fprintf (vect_dump, "=== vectorizable_demotion ===");
4245 vect_model_simple_cost (stmt_info, ncopies, dt, NULL);
4250 if (vect_print_dump_info (REPORT_DETAILS))
4251 fprintf (vect_dump, "transform type demotion operation. ncopies = %d.",
4255 vec_dest = vect_create_destination_var (scalar_dest, vectype_out);
4257 /* In case the vectorization factor (VF) is bigger than the number
4258 of elements that we can fit in a vectype (nunits), we have to generate
4259 more than one vector stmt - i.e - we need to "unroll" the
4260 vector stmt by a factor VF/nunits. */
4261 prev_stmt_info = NULL;
4262 for (j = 0; j < ncopies; j++)
4267 vec_oprnd0 = vect_get_vec_def_for_operand (op0, stmt, NULL);
4268 vec_oprnd1 = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd0);
4272 vec_oprnd0 = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd1);
4273 vec_oprnd1 = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd0);
4276 /* Arguments are ready. Create the new vector stmt. */
4277 expr = build2 (code1, vectype_out, vec_oprnd0, vec_oprnd1);
4278 new_stmt = build_gimple_modify_stmt (vec_dest, expr);
4279 new_temp = make_ssa_name (vec_dest, new_stmt);
4280 GIMPLE_STMT_OPERAND (new_stmt, 0) = new_temp;
4281 vect_finish_stmt_generation (stmt, new_stmt, bsi);
4284 STMT_VINFO_VEC_STMT (stmt_info) = new_stmt;
4286 STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt;
4288 prev_stmt_info = vinfo_for_stmt (new_stmt);
4291 *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info);
4296 /* Function vectorizable_type_promotion
4298 Check if STMT performs a binary or unary operation that involves
4299 type promotion, and if it can be vectorized.
4300 If VEC_STMT is also passed, vectorize the STMT: create a vectorized
4301 stmt to replace it, put it in VEC_STMT, and insert it at BSI.
4302 Return FALSE if not a vectorizable STMT, TRUE otherwise. */
4305 vectorizable_type_promotion (tree stmt, block_stmt_iterator *bsi,
4311 tree op0, op1 = NULL;
4312 tree vec_oprnd0=NULL, vec_oprnd1=NULL;
4313 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
4314 loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info);
4315 struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
4316 enum tree_code code, code1 = ERROR_MARK, code2 = ERROR_MARK;
4317 tree decl1 = NULL_TREE, decl2 = NULL_TREE;
4320 enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type};
4322 stmt_vec_info prev_stmt_info;
4330 if (!STMT_VINFO_RELEVANT_P (stmt_info))
4333 if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_loop_def)
4336 /* Is STMT a vectorizable type-promotion operation? */
4337 if (TREE_CODE (stmt) != GIMPLE_MODIFY_STMT)
4340 if (TREE_CODE (GIMPLE_STMT_OPERAND (stmt, 0)) != SSA_NAME)
4343 operation = GIMPLE_STMT_OPERAND (stmt, 1);
4344 code = TREE_CODE (operation);
4345 if (code != NOP_EXPR && code != CONVERT_EXPR
4346 && code != WIDEN_MULT_EXPR)
4349 op0 = TREE_OPERAND (operation, 0);
4350 vectype_in = get_vectype_for_scalar_type (TREE_TYPE (op0));
4353 nunits_in = TYPE_VECTOR_SUBPARTS (vectype_in);
4355 scalar_dest = GIMPLE_STMT_OPERAND (stmt, 0);
4356 vectype_out = get_vectype_for_scalar_type (TREE_TYPE (scalar_dest));
4359 nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out);
4360 if (nunits_out != nunits_in / 2) /* FORNOW */
4363 ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in;
4364 gcc_assert (ncopies >= 1);
4365 /* FORNOW. This restriction should be relaxed. */
4366 if (nested_in_vect_loop_p (loop, stmt) && ncopies > 1)
4368 if (vect_print_dump_info (REPORT_DETAILS))
4369 fprintf (vect_dump, "multiple types in nested loop.");
4373 if (! ((INTEGRAL_TYPE_P (TREE_TYPE (scalar_dest))
4374 && INTEGRAL_TYPE_P (TREE_TYPE (op0)))
4375 || (SCALAR_FLOAT_TYPE_P (TREE_TYPE (scalar_dest))
4376 && SCALAR_FLOAT_TYPE_P (TREE_TYPE (op0))
4377 && (code == CONVERT_EXPR || code == NOP_EXPR))))
4380 /* Check the operands of the operation. */
4381 if (!vect_is_simple_use (op0, loop_vinfo, &def_stmt, &def, &dt[0]))
4383 if (vect_print_dump_info (REPORT_DETAILS))
4384 fprintf (vect_dump, "use not simple.");
4388 op_type = TREE_CODE_LENGTH (code);
4389 if (op_type == binary_op)
4391 op1 = TREE_OPERAND (operation, 1);
4392 if (!vect_is_simple_use (op1, loop_vinfo, &def_stmt, &def, &dt[1]))
4394 if (vect_print_dump_info (REPORT_DETAILS))
4395 fprintf (vect_dump, "use not simple.");
4400 /* Supportable by target? */
4401 if (!supportable_widening_operation (code, stmt, vectype_in,
4402 &decl1, &decl2, &code1, &code2))
4405 STMT_VINFO_VECTYPE (stmt_info) = vectype_in;
4407 if (!vec_stmt) /* transformation not required. */
4409 STMT_VINFO_TYPE (stmt_info) = type_promotion_vec_info_type;
4410 if (vect_print_dump_info (REPORT_DETAILS))
4411 fprintf (vect_dump, "=== vectorizable_promotion ===");
4412 vect_model_simple_cost (stmt_info, 2*ncopies, dt, NULL);
4418 if (vect_print_dump_info (REPORT_DETAILS))
4419 fprintf (vect_dump, "transform type promotion operation. ncopies = %d.",
4423 vec_dest = vect_create_destination_var (scalar_dest, vectype_out);
4425 /* In case the vectorization factor (VF) is bigger than the number
4426 of elements that we can fit in a vectype (nunits), we have to generate
4427 more than one vector stmt - i.e - we need to "unroll" the
4428 vector stmt by a factor VF/nunits. */
4430 prev_stmt_info = NULL;
4431 for (j = 0; j < ncopies; j++)
4436 vec_oprnd0 = vect_get_vec_def_for_operand (op0, stmt, NULL);
4437 if (op_type == binary_op)
4438 vec_oprnd1 = vect_get_vec_def_for_operand (op1, stmt, NULL);
4442 vec_oprnd0 = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd0);
4443 if (op_type == binary_op)
4444 vec_oprnd1 = vect_get_vec_def_for_stmt_copy (dt[1], vec_oprnd1);
4447 /* Arguments are ready. Create the new vector stmt. We are creating
4448 two vector defs because the widened result does not fit in one vector.
4449 The vectorized stmt can be expressed as a call to a taregt builtin,
4450 or a using a tree-code. */
4451 /* Generate first half of the widened result: */
4452 new_stmt = vect_gen_widened_results_half (code1, vectype_out, decl1,
4453 vec_oprnd0, vec_oprnd1, op_type, vec_dest, bsi, stmt);
4455 STMT_VINFO_VEC_STMT (stmt_info) = new_stmt;
4457 STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt;
4458 prev_stmt_info = vinfo_for_stmt (new_stmt);
4460 /* Generate second half of the widened result: */
4461 new_stmt = vect_gen_widened_results_half (code2, vectype_out, decl2,
4462 vec_oprnd0, vec_oprnd1, op_type, vec_dest, bsi, stmt);
4463 STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt;
4464 prev_stmt_info = vinfo_for_stmt (new_stmt);
4468 *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info);
4473 /* Function vect_strided_store_supported.
4475 Returns TRUE is INTERLEAVE_HIGH and INTERLEAVE_LOW operations are supported,
4476 and FALSE otherwise. */
4479 vect_strided_store_supported (tree vectype)
4481 optab interleave_high_optab, interleave_low_optab;
4484 mode = (int) TYPE_MODE (vectype);
4486 /* Check that the operation is supported. */
4487 interleave_high_optab = optab_for_tree_code (VEC_INTERLEAVE_HIGH_EXPR,
4489 interleave_low_optab = optab_for_tree_code (VEC_INTERLEAVE_LOW_EXPR,
4491 if (!interleave_high_optab || !interleave_low_optab)
4493 if (vect_print_dump_info (REPORT_DETAILS))
4494 fprintf (vect_dump, "no optab for interleave.");
4498 if (optab_handler (interleave_high_optab, mode)->insn_code
4500 || optab_handler (interleave_low_optab, mode)->insn_code
4501 == CODE_FOR_nothing)
4503 if (vect_print_dump_info (REPORT_DETAILS))
4504 fprintf (vect_dump, "interleave op not supported by target.");
4512 /* Function vect_permute_store_chain.
4514 Given a chain of interleaved stores in DR_CHAIN of LENGTH that must be
4515 a power of 2, generate interleave_high/low stmts to reorder the data
4516 correctly for the stores. Return the final references for stores in
4519 E.g., LENGTH is 4 and the scalar type is short, i.e., VF is 8.
4520 The input is 4 vectors each containing 8 elements. We assign a number to each
4521 element, the input sequence is:
4523 1st vec: 0 1 2 3 4 5 6 7
4524 2nd vec: 8 9 10 11 12 13 14 15
4525 3rd vec: 16 17 18 19 20 21 22 23
4526 4th vec: 24 25 26 27 28 29 30 31
4528 The output sequence should be:
4530 1st vec: 0 8 16 24 1 9 17 25
4531 2nd vec: 2 10 18 26 3 11 19 27
4532 3rd vec: 4 12 20 28 5 13 21 30
4533 4th vec: 6 14 22 30 7 15 23 31
4535 i.e., we interleave the contents of the four vectors in their order.
4537 We use interleave_high/low instructions to create such output. The input of
4538 each interleave_high/low operation is two vectors:
4541 the even elements of the result vector are obtained left-to-right from the
4542 high/low elements of the first vector. The odd elements of the result are
4543 obtained left-to-right from the high/low elements of the second vector.
4544 The output of interleave_high will be: 0 4 1 5
4545 and of interleave_low: 2 6 3 7
4548 The permutation is done in log LENGTH stages. In each stage interleave_high
4549 and interleave_low stmts are created for each pair of vectors in DR_CHAIN,
4550 where the first argument is taken from the first half of DR_CHAIN and the
4551 second argument from it's second half.
4554 I1: interleave_high (1st vec, 3rd vec)
4555 I2: interleave_low (1st vec, 3rd vec)
4556 I3: interleave_high (2nd vec, 4th vec)
4557 I4: interleave_low (2nd vec, 4th vec)
4559 The output for the first stage is:
4561 I1: 0 16 1 17 2 18 3 19
4562 I2: 4 20 5 21 6 22 7 23
4563 I3: 8 24 9 25 10 26 11 27
4564 I4: 12 28 13 29 14 30 15 31
4566 The output of the second stage, i.e. the final result is:
4568 I1: 0 8 16 24 1 9 17 25
4569 I2: 2 10 18 26 3 11 19 27
4570 I3: 4 12 20 28 5 13 21 30
4571 I4: 6 14 22 30 7 15 23 31. */
4574 vect_permute_store_chain (VEC(tree,heap) *dr_chain,
4575 unsigned int length,
4577 block_stmt_iterator *bsi,
4578 VEC(tree,heap) **result_chain)
4580 tree perm_dest, perm_stmt, vect1, vect2, high, low;
4581 tree vectype = STMT_VINFO_VECTYPE (vinfo_for_stmt (stmt));
4582 tree scalar_dest, tmp;
4585 VEC(tree,heap) *first, *second;
4587 scalar_dest = GIMPLE_STMT_OPERAND (stmt, 0);
4588 first = VEC_alloc (tree, heap, length/2);
4589 second = VEC_alloc (tree, heap, length/2);
4591 /* Check that the operation is supported. */
4592 if (!vect_strided_store_supported (vectype))
4595 *result_chain = VEC_copy (tree, heap, dr_chain);
4597 for (i = 0; i < exact_log2 (length); i++)
4599 for (j = 0; j < length/2; j++)
4601 vect1 = VEC_index (tree, dr_chain, j);
4602 vect2 = VEC_index (tree, dr_chain, j+length/2);
4604 /* Create interleaving stmt:
4605 in the case of big endian:
4606 high = interleave_high (vect1, vect2)
4607 and in the case of little endian:
4608 high = interleave_low (vect1, vect2). */
4609 perm_dest = create_tmp_var (vectype, "vect_inter_high");
4610 DECL_GIMPLE_REG_P (perm_dest) = 1;
4611 add_referenced_var (perm_dest);
4612 if (BYTES_BIG_ENDIAN)
4613 tmp = build2 (VEC_INTERLEAVE_HIGH_EXPR, vectype, vect1, vect2);
4615 tmp = build2 (VEC_INTERLEAVE_LOW_EXPR, vectype, vect1, vect2);
4616 perm_stmt = build_gimple_modify_stmt (perm_dest, tmp);
4617 high = make_ssa_name (perm_dest, perm_stmt);
4618 GIMPLE_STMT_OPERAND (perm_stmt, 0) = high;
4619 vect_finish_stmt_generation (stmt, perm_stmt, bsi);
4620 VEC_replace (tree, *result_chain, 2*j, high);
4622 /* Create interleaving stmt:
4623 in the case of big endian:
4624 low = interleave_low (vect1, vect2)
4625 and in the case of little endian:
4626 low = interleave_high (vect1, vect2). */
4627 perm_dest = create_tmp_var (vectype, "vect_inter_low");
4628 DECL_GIMPLE_REG_P (perm_dest) = 1;
4629 add_referenced_var (perm_dest);
4630 if (BYTES_BIG_ENDIAN)
4631 tmp = build2 (VEC_INTERLEAVE_LOW_EXPR, vectype, vect1, vect2);
4633 tmp = build2 (VEC_INTERLEAVE_HIGH_EXPR, vectype, vect1, vect2);
4634 perm_stmt = build_gimple_modify_stmt (perm_dest, tmp);
4635 low = make_ssa_name (perm_dest, perm_stmt);
4636 GIMPLE_STMT_OPERAND (perm_stmt, 0) = low;
4637 vect_finish_stmt_generation (stmt, perm_stmt, bsi);
4638 VEC_replace (tree, *result_chain, 2*j+1, low);
4640 dr_chain = VEC_copy (tree, heap, *result_chain);
4646 /* Function vectorizable_store.
4648 Check if STMT defines a non scalar data-ref (array/pointer/structure) that
4650 If VEC_STMT is also passed, vectorize the STMT: create a vectorized
4651 stmt to replace it, put it in VEC_STMT, and insert it at BSI.
4652 Return FALSE if not a vectorizable STMT, TRUE otherwise. */
4655 vectorizable_store (tree stmt, block_stmt_iterator *bsi, tree *vec_stmt,
4661 tree vec_oprnd = NULL_TREE;
4662 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
4663 struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info), *first_dr = NULL;
4664 tree vectype = STMT_VINFO_VECTYPE (stmt_info);
4665 loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info);
4666 struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
4667 enum machine_mode vec_mode;
4669 enum dr_alignment_support alignment_support_scheme;
4671 enum vect_def_type dt;
4672 stmt_vec_info prev_stmt_info = NULL;
4673 tree dataref_ptr = NULL_TREE;
4674 int nunits = TYPE_VECTOR_SUBPARTS (vectype);
4675 int ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits;
4677 tree next_stmt, first_stmt = NULL_TREE;
4678 bool strided_store = false;
4679 unsigned int group_size, i;
4680 VEC(tree,heap) *dr_chain = NULL, *oprnds = NULL, *result_chain = NULL;
4682 VEC(tree,heap) *vec_oprnds = NULL;
4683 bool slp = (slp_node != NULL);
4684 stmt_vec_info first_stmt_vinfo;
4685 unsigned int vec_num;
4687 /* FORNOW: SLP with multiple types is not supported. The SLP analysis verifies
4688 this, so we can safely override NCOPIES with 1 here. */
4692 gcc_assert (ncopies >= 1);
4694 /* FORNOW. This restriction should be relaxed. */
4695 if (nested_in_vect_loop_p (loop, stmt) && ncopies > 1)
4697 if (vect_print_dump_info (REPORT_DETAILS))
4698 fprintf (vect_dump, "multiple types in nested loop.");
4702 if (!STMT_VINFO_RELEVANT_P (stmt_info))
4705 if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_loop_def)
4708 /* Is vectorizable store? */
4710 if (TREE_CODE (stmt) != GIMPLE_MODIFY_STMT)
4713 scalar_dest = GIMPLE_STMT_OPERAND (stmt, 0);
4714 if (TREE_CODE (scalar_dest) != ARRAY_REF
4715 && TREE_CODE (scalar_dest) != INDIRECT_REF
4716 && !STMT_VINFO_STRIDED_ACCESS (stmt_info))
4719 op = GIMPLE_STMT_OPERAND (stmt, 1);
4720 if (!vect_is_simple_use (op, loop_vinfo, &def_stmt, &def, &dt))
4722 if (vect_print_dump_info (REPORT_DETAILS))
4723 fprintf (vect_dump, "use not simple.");
4727 vec_mode = TYPE_MODE (vectype);
4728 /* FORNOW. In some cases can vectorize even if data-type not supported
4729 (e.g. - array initialization with 0). */
4730 if (optab_handler (mov_optab, (int)vec_mode)->insn_code == CODE_FOR_nothing)
4733 if (!STMT_VINFO_DATA_REF (stmt_info))
4736 if (STMT_VINFO_STRIDED_ACCESS (stmt_info))
4738 strided_store = true;
4739 first_stmt = DR_GROUP_FIRST_DR (stmt_info);
4740 if (!vect_strided_store_supported (vectype)
4741 && !PURE_SLP_STMT (stmt_info) && !slp)
4744 if (first_stmt == stmt)
4746 /* STMT is the leader of the group. Check the operands of all the
4747 stmts of the group. */
4748 next_stmt = DR_GROUP_NEXT_DR (stmt_info);
4751 op = GIMPLE_STMT_OPERAND (next_stmt, 1);
4752 if (!vect_is_simple_use (op, loop_vinfo, &def_stmt, &def, &dt))
4754 if (vect_print_dump_info (REPORT_DETAILS))
4755 fprintf (vect_dump, "use not simple.");
4758 next_stmt = DR_GROUP_NEXT_DR (vinfo_for_stmt (next_stmt));
4763 if (!vec_stmt) /* transformation not required. */
4765 STMT_VINFO_TYPE (stmt_info) = store_vec_info_type;
4766 if (!PURE_SLP_STMT (stmt_info))
4767 vect_model_store_cost (stmt_info, ncopies, dt, NULL);
4775 first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt));
4776 group_size = DR_GROUP_SIZE (vinfo_for_stmt (first_stmt));
4778 DR_GROUP_STORE_COUNT (vinfo_for_stmt (first_stmt))++;
4781 gcc_assert (!nested_in_vect_loop_p (loop, stmt));
4783 /* We vectorize all the stmts of the interleaving group when we
4784 reach the last stmt in the group. */
4785 if (DR_GROUP_STORE_COUNT (vinfo_for_stmt (first_stmt))
4786 < DR_GROUP_SIZE (vinfo_for_stmt (first_stmt))
4789 *vec_stmt = NULL_TREE;
4794 strided_store = false;
4796 /* VEC_NUM is the number of vect stmts to be created for this group. */
4797 if (slp && SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node) < group_size)
4798 vec_num = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node);
4800 vec_num = group_size;
4806 group_size = vec_num = 1;
4807 first_stmt_vinfo = stmt_info;
4810 if (vect_print_dump_info (REPORT_DETAILS))
4811 fprintf (vect_dump, "transform store. ncopies = %d",ncopies);
4813 dr_chain = VEC_alloc (tree, heap, group_size);
4814 oprnds = VEC_alloc (tree, heap, group_size);
4816 alignment_support_scheme = vect_supportable_dr_alignment (first_dr);
4817 gcc_assert (alignment_support_scheme);
4818 gcc_assert (alignment_support_scheme == dr_aligned); /* FORNOW */
4820 /* In case the vectorization factor (VF) is bigger than the number
4821 of elements that we can fit in a vectype (nunits), we have to generate
4822 more than one vector stmt - i.e - we need to "unroll" the
4823 vector stmt by a factor VF/nunits. For more details see documentation in
4824 vect_get_vec_def_for_copy_stmt. */
4826 /* In case of interleaving (non-unit strided access):
4833 We create vectorized stores starting from base address (the access of the
4834 first stmt in the chain (S2 in the above example), when the last store stmt
4835 of the chain (S4) is reached:
4838 VS2: &base + vec_size*1 = vx0
4839 VS3: &base + vec_size*2 = vx1
4840 VS4: &base + vec_size*3 = vx3
4842 Then permutation statements are generated:
4844 VS5: vx5 = VEC_INTERLEAVE_HIGH_EXPR < vx0, vx3 >
4845 VS6: vx6 = VEC_INTERLEAVE_LOW_EXPR < vx0, vx3 >
4848 And they are put in STMT_VINFO_VEC_STMT of the corresponding scalar stmts
4849 (the order of the data-refs in the output of vect_permute_store_chain
4850 corresponds to the order of scalar stmts in the interleaving chain - see
4851 the documentation of vect_permute_store_chain()).
4853 In case of both multiple types and interleaving, above vector stores and
4854 permutation stmts are created for every copy. The result vector stmts are
4855 put in STMT_VINFO_VEC_STMT for the first copy and in the corresponding
4856 STMT_VINFO_RELATED_STMT for the next copies.
4859 prev_stmt_info = NULL;
4860 for (j = 0; j < ncopies; j++)
4869 /* Get vectorized arguments for SLP_NODE. */
4870 vect_get_slp_defs (slp_node, &vec_oprnds, NULL);
4872 vec_oprnd = VEC_index (tree, vec_oprnds, 0);
4876 /* For interleaved stores we collect vectorized defs for all the
4877 stores in the group in DR_CHAIN and OPRNDS. DR_CHAIN is then
4878 used as an input to vect_permute_store_chain(), and OPRNDS as
4879 an input to vect_get_vec_def_for_stmt_copy() for the next copy.
4881 If the store is not strided, GROUP_SIZE is 1, and DR_CHAIN and
4882 OPRNDS are of size 1. */
4883 next_stmt = first_stmt;
4884 for (i = 0; i < group_size; i++)
4886 /* Since gaps are not supported for interleaved stores,
4887 GROUP_SIZE is the exact number of stmts in the chain.
4888 Therefore, NEXT_STMT can't be NULL_TREE. In case that
4889 there is no interleaving, GROUP_SIZE is 1, and only one
4890 iteration of the loop will be executed. */
4891 gcc_assert (next_stmt);
4892 op = GIMPLE_STMT_OPERAND (next_stmt, 1);
4894 vec_oprnd = vect_get_vec_def_for_operand (op, next_stmt,
4896 VEC_quick_push(tree, dr_chain, vec_oprnd);
4897 VEC_quick_push(tree, oprnds, vec_oprnd);
4898 next_stmt = DR_GROUP_NEXT_DR (vinfo_for_stmt (next_stmt));
4901 dataref_ptr = vect_create_data_ref_ptr (first_stmt, NULL, NULL_TREE,
4902 &dummy, &ptr_incr, false,
4903 TREE_TYPE (vec_oprnd), &inv_p);
4904 gcc_assert (!inv_p);
4908 /* FORNOW SLP doesn't work for multiple types. */
4911 /* For interleaved stores we created vectorized defs for all the
4912 defs stored in OPRNDS in the previous iteration (previous copy).
4913 DR_CHAIN is then used as an input to vect_permute_store_chain(),
4914 and OPRNDS as an input to vect_get_vec_def_for_stmt_copy() for the
4916 If the store is not strided, GROUP_SIZE is 1, and DR_CHAIN and
4917 OPRNDS are of size 1. */
4918 for (i = 0; i < group_size; i++)
4920 op = VEC_index (tree, oprnds, i);
4921 vect_is_simple_use (op, loop_vinfo, &def_stmt, &def, &dt);
4922 vec_oprnd = vect_get_vec_def_for_stmt_copy (dt, op);
4923 VEC_replace(tree, dr_chain, i, vec_oprnd);
4924 VEC_replace(tree, oprnds, i, vec_oprnd);
4927 bump_vector_ptr (dataref_ptr, ptr_incr, bsi, stmt, NULL_TREE);
4932 result_chain = VEC_alloc (tree, heap, group_size);
4934 if (!vect_permute_store_chain (dr_chain, group_size, stmt, bsi,
4939 next_stmt = first_stmt;
4940 for (i = 0; i < vec_num; i++)
4943 /* Bump the vector pointer. */
4944 dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, bsi, stmt,
4948 vec_oprnd = VEC_index (tree, vec_oprnds, i);
4949 else if (strided_store)
4950 /* For strided stores vectorized defs are interleaved in
4951 vect_permute_store_chain(). */
4952 vec_oprnd = VEC_index (tree, result_chain, i);
4954 data_ref = build_fold_indirect_ref (dataref_ptr);
4955 /* Arguments are ready. Create the new vector stmt. */
4956 new_stmt = build_gimple_modify_stmt (data_ref, vec_oprnd);
4957 vect_finish_stmt_generation (stmt, new_stmt, bsi);
4958 mark_symbols_for_renaming (new_stmt);
4961 STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt;
4963 STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt;
4965 prev_stmt_info = vinfo_for_stmt (new_stmt);
4966 next_stmt = DR_GROUP_NEXT_DR (vinfo_for_stmt (next_stmt));
4976 /* Function vect_setup_realignment
4978 This function is called when vectorizing an unaligned load using
4979 the dr_explicit_realign[_optimized] scheme.
4980 This function generates the following code at the loop prolog:
4983 x msq_init = *(floor(p)); # prolog load
4984 realignment_token = call target_builtin;
4986 x msq = phi (msq_init, ---)
4988 The stmts marked with x are generated only for the case of
4989 dr_explicit_realign_optimized.
4991 The code above sets up a new (vector) pointer, pointing to the first
4992 location accessed by STMT, and a "floor-aligned" load using that pointer.
4993 It also generates code to compute the "realignment-token" (if the relevant
4994 target hook was defined), and creates a phi-node at the loop-header bb
4995 whose arguments are the result of the prolog-load (created by this
4996 function) and the result of a load that takes place in the loop (to be
4997 created by the caller to this function).
4999 For the case of dr_explicit_realign_optimized:
5000 The caller to this function uses the phi-result (msq) to create the
5001 realignment code inside the loop, and sets up the missing phi argument,
5004 msq = phi (msq_init, lsq)
5005 lsq = *(floor(p')); # load in loop
5006 result = realign_load (msq, lsq, realignment_token);
5008 For the case of dr_explicit_realign:
5010 msq = *(floor(p)); # load in loop
5012 lsq = *(floor(p')); # load in loop
5013 result = realign_load (msq, lsq, realignment_token);
5016 STMT - (scalar) load stmt to be vectorized. This load accesses
5017 a memory location that may be unaligned.
5018 BSI - place where new code is to be inserted.
5019 ALIGNMENT_SUPPORT_SCHEME - which of the two misalignment handling schemes
5023 REALIGNMENT_TOKEN - the result of a call to the builtin_mask_for_load
5024 target hook, if defined.
5025 Return value - the result of the loop-header phi node. */
5028 vect_setup_realignment (tree stmt, block_stmt_iterator *bsi,
5029 tree *realignment_token,
5030 enum dr_alignment_support alignment_support_scheme,
5032 struct loop **at_loop)
5034 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
5035 tree vectype = STMT_VINFO_VECTYPE (stmt_info);
5036 loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info);
5037 struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
5039 tree scalar_dest = GIMPLE_STMT_OPERAND (stmt, 0);
5046 tree msq_init = NULL_TREE;
5049 tree msq = NULL_TREE;
5050 tree stmts = NULL_TREE;
5052 bool compute_in_loop = false;
5053 bool nested_in_vect_loop = nested_in_vect_loop_p (loop, stmt);
5054 struct loop *containing_loop = (bb_for_stmt (stmt))->loop_father;
5055 struct loop *loop_for_initial_load;
5057 gcc_assert (alignment_support_scheme == dr_explicit_realign
5058 || alignment_support_scheme == dr_explicit_realign_optimized);
5060 /* We need to generate three things:
5061 1. the misalignment computation
5062 2. the extra vector load (for the optimized realignment scheme).
5063 3. the phi node for the two vectors from which the realignment is
5064 done (for the optimized realignment scheme).
5067 /* 1. Determine where to generate the misalignment computation.
5069 If INIT_ADDR is NULL_TREE, this indicates that the misalignment
5070 calculation will be generated by this function, outside the loop (in the
5071 preheader). Otherwise, INIT_ADDR had already been computed for us by the
5072 caller, inside the loop.
5074 Background: If the misalignment remains fixed throughout the iterations of
5075 the loop, then both realignment schemes are applicable, and also the
5076 misalignment computation can be done outside LOOP. This is because we are
5077 vectorizing LOOP, and so the memory accesses in LOOP advance in steps that
5078 are a multiple of VS (the Vector Size), and therefore the misalignment in
5079 different vectorized LOOP iterations is always the same.
5080 The problem arises only if the memory access is in an inner-loop nested
5081 inside LOOP, which is now being vectorized using outer-loop vectorization.
5082 This is the only case when the misalignment of the memory access may not
5083 remain fixed throughout the iterations of the inner-loop (as explained in
5084 detail in vect_supportable_dr_alignment). In this case, not only is the
5085 optimized realignment scheme not applicable, but also the misalignment
5086 computation (and generation of the realignment token that is passed to
5087 REALIGN_LOAD) have to be done inside the loop.
5089 In short, INIT_ADDR indicates whether we are in a COMPUTE_IN_LOOP mode
5090 or not, which in turn determines if the misalignment is computed inside
5091 the inner-loop, or outside LOOP. */
5093 if (init_addr != NULL_TREE)
5095 compute_in_loop = true;
5096 gcc_assert (alignment_support_scheme == dr_explicit_realign);
5100 /* 2. Determine where to generate the extra vector load.
5102 For the optimized realignment scheme, instead of generating two vector
5103 loads in each iteration, we generate a single extra vector load in the
5104 preheader of the loop, and in each iteration reuse the result of the
5105 vector load from the previous iteration. In case the memory access is in
5106 an inner-loop nested inside LOOP, which is now being vectorized using
5107 outer-loop vectorization, we need to determine whether this initial vector
5108 load should be generated at the preheader of the inner-loop, or can be
5109 generated at the preheader of LOOP. If the memory access has no evolution
5110 in LOOP, it can be generated in the preheader of LOOP. Otherwise, it has
5111 to be generated inside LOOP (in the preheader of the inner-loop). */
5113 if (nested_in_vect_loop)
5115 tree outerloop_step = STMT_VINFO_DR_STEP (stmt_info);
5116 bool invariant_in_outerloop =
5117 (tree_int_cst_compare (outerloop_step, size_zero_node) == 0);
5118 loop_for_initial_load = (invariant_in_outerloop ? loop : loop->inner);
5121 loop_for_initial_load = loop;
5123 *at_loop = loop_for_initial_load;
5125 /* 3. For the case of the optimized realignment, create the first vector
5126 load at the loop preheader. */
5128 if (alignment_support_scheme == dr_explicit_realign_optimized)
5130 /* Create msq_init = *(floor(p1)) in the loop preheader */
5132 gcc_assert (!compute_in_loop);
5133 pe = loop_preheader_edge (loop_for_initial_load);
5134 vec_dest = vect_create_destination_var (scalar_dest, vectype);
5135 ptr = vect_create_data_ref_ptr (stmt, loop_for_initial_load, NULL_TREE,
5136 &init_addr, &inc, true, NULL_TREE, &inv_p);
5137 data_ref = build1 (ALIGN_INDIRECT_REF, vectype, ptr);
5138 new_stmt = build_gimple_modify_stmt (vec_dest, data_ref);
5139 new_temp = make_ssa_name (vec_dest, new_stmt);
5140 GIMPLE_STMT_OPERAND (new_stmt, 0) = new_temp;
5141 mark_symbols_for_renaming (new_stmt);
5142 new_bb = bsi_insert_on_edge_immediate (pe, new_stmt);
5143 gcc_assert (!new_bb);
5144 msq_init = GIMPLE_STMT_OPERAND (new_stmt, 0);
5147 /* 4. Create realignment token using a target builtin, if available.
5148 It is done either inside the containing loop, or before LOOP (as
5149 determined above). */
5151 if (targetm.vectorize.builtin_mask_for_load)
5155 /* Compute INIT_ADDR - the initial addressed accessed by this memref. */
5156 if (compute_in_loop)
5157 gcc_assert (init_addr); /* already computed by the caller. */
5160 /* Generate the INIT_ADDR computation outside LOOP. */
5161 init_addr = vect_create_addr_base_for_vector_ref (stmt, &stmts,
5163 pe = loop_preheader_edge (loop);
5164 new_bb = bsi_insert_on_edge_immediate (pe, stmts);
5165 gcc_assert (!new_bb);
5168 builtin_decl = targetm.vectorize.builtin_mask_for_load ();
5169 new_stmt = build_call_expr (builtin_decl, 1, init_addr);
5170 vec_dest = vect_create_destination_var (scalar_dest,
5171 TREE_TYPE (new_stmt));
5172 new_stmt = build_gimple_modify_stmt (vec_dest, new_stmt);
5173 new_temp = make_ssa_name (vec_dest, new_stmt);
5174 GIMPLE_STMT_OPERAND (new_stmt, 0) = new_temp;
5176 if (compute_in_loop)
5177 bsi_insert_before (bsi, new_stmt, BSI_SAME_STMT);
5180 /* Generate the misalignment computation outside LOOP. */
5181 pe = loop_preheader_edge (loop);
5182 new_bb = bsi_insert_on_edge_immediate (pe, new_stmt);
5183 gcc_assert (!new_bb);
5186 *realignment_token = GIMPLE_STMT_OPERAND (new_stmt, 0);
5188 /* The result of the CALL_EXPR to this builtin is determined from
5189 the value of the parameter and no global variables are touched
5190 which makes the builtin a "const" function. Requiring the
5191 builtin to have the "const" attribute makes it unnecessary
5192 to call mark_call_clobbered. */
5193 gcc_assert (TREE_READONLY (builtin_decl));
5196 if (alignment_support_scheme == dr_explicit_realign)
5199 gcc_assert (!compute_in_loop);
5200 gcc_assert (alignment_support_scheme == dr_explicit_realign_optimized);
5203 /* 5. Create msq = phi <msq_init, lsq> in loop */
5205 pe = loop_preheader_edge (containing_loop);
5206 vec_dest = vect_create_destination_var (scalar_dest, vectype);
5207 msq = make_ssa_name (vec_dest, NULL_TREE);
5208 phi_stmt = create_phi_node (msq, containing_loop->header);
5209 SSA_NAME_DEF_STMT (msq) = phi_stmt;
5210 add_phi_arg (phi_stmt, msq_init, pe);
5216 /* Function vect_strided_load_supported.
5218 Returns TRUE is EXTRACT_EVEN and EXTRACT_ODD operations are supported,
5219 and FALSE otherwise. */
5222 vect_strided_load_supported (tree vectype)
5224 optab perm_even_optab, perm_odd_optab;
5227 mode = (int) TYPE_MODE (vectype);
5229 perm_even_optab = optab_for_tree_code (VEC_EXTRACT_EVEN_EXPR, vectype);
5230 if (!perm_even_optab)
5232 if (vect_print_dump_info (REPORT_DETAILS))
5233 fprintf (vect_dump, "no optab for perm_even.");
5237 if (optab_handler (perm_even_optab, mode)->insn_code == CODE_FOR_nothing)
5239 if (vect_print_dump_info (REPORT_DETAILS))
5240 fprintf (vect_dump, "perm_even op not supported by target.");
5244 perm_odd_optab = optab_for_tree_code (VEC_EXTRACT_ODD_EXPR, vectype);
5245 if (!perm_odd_optab)
5247 if (vect_print_dump_info (REPORT_DETAILS))
5248 fprintf (vect_dump, "no optab for perm_odd.");
5252 if (optab_handler (perm_odd_optab, mode)->insn_code == CODE_FOR_nothing)
5254 if (vect_print_dump_info (REPORT_DETAILS))
5255 fprintf (vect_dump, "perm_odd op not supported by target.");
5262 /* Function vect_permute_load_chain.
5264 Given a chain of interleaved loads in DR_CHAIN of LENGTH that must be
5265 a power of 2, generate extract_even/odd stmts to reorder the input data
5266 correctly. Return the final references for loads in RESULT_CHAIN.
5268 E.g., LENGTH is 4 and the scalar type is short, i.e., VF is 8.
5269 The input is 4 vectors each containing 8 elements. We assign a number to each
5270 element, the input sequence is:
5272 1st vec: 0 1 2 3 4 5 6 7
5273 2nd vec: 8 9 10 11 12 13 14 15
5274 3rd vec: 16 17 18 19 20 21 22 23
5275 4th vec: 24 25 26 27 28 29 30 31
5277 The output sequence should be:
5279 1st vec: 0 4 8 12 16 20 24 28
5280 2nd vec: 1 5 9 13 17 21 25 29
5281 3rd vec: 2 6 10 14 18 22 26 30
5282 4th vec: 3 7 11 15 19 23 27 31
5284 i.e., the first output vector should contain the first elements of each
5285 interleaving group, etc.
5287 We use extract_even/odd instructions to create such output. The input of each
5288 extract_even/odd operation is two vectors
5292 and the output is the vector of extracted even/odd elements. The output of
5293 extract_even will be: 0 2 4 6
5294 and of extract_odd: 1 3 5 7
5297 The permutation is done in log LENGTH stages. In each stage extract_even and
5298 extract_odd stmts are created for each pair of vectors in DR_CHAIN in their
5299 order. In our example,
5301 E1: extract_even (1st vec, 2nd vec)
5302 E2: extract_odd (1st vec, 2nd vec)
5303 E3: extract_even (3rd vec, 4th vec)
5304 E4: extract_odd (3rd vec, 4th vec)
5306 The output for the first stage will be:
5308 E1: 0 2 4 6 8 10 12 14
5309 E2: 1 3 5 7 9 11 13 15
5310 E3: 16 18 20 22 24 26 28 30
5311 E4: 17 19 21 23 25 27 29 31
5313 In order to proceed and create the correct sequence for the next stage (or
5314 for the correct output, if the second stage is the last one, as in our
5315 example), we first put the output of extract_even operation and then the
5316 output of extract_odd in RESULT_CHAIN (which is then copied to DR_CHAIN).
5317 The input for the second stage is:
5319 1st vec (E1): 0 2 4 6 8 10 12 14
5320 2nd vec (E3): 16 18 20 22 24 26 28 30
5321 3rd vec (E2): 1 3 5 7 9 11 13 15
5322 4th vec (E4): 17 19 21 23 25 27 29 31
5324 The output of the second stage:
5326 E1: 0 4 8 12 16 20 24 28
5327 E2: 2 6 10 14 18 22 26 30
5328 E3: 1 5 9 13 17 21 25 29
5329 E4: 3 7 11 15 19 23 27 31
5331 And RESULT_CHAIN after reordering:
5333 1st vec (E1): 0 4 8 12 16 20 24 28
5334 2nd vec (E3): 1 5 9 13 17 21 25 29
5335 3rd vec (E2): 2 6 10 14 18 22 26 30
5336 4th vec (E4): 3 7 11 15 19 23 27 31. */
5339 vect_permute_load_chain (VEC(tree,heap) *dr_chain,
5340 unsigned int length,
5342 block_stmt_iterator *bsi,
5343 VEC(tree,heap) **result_chain)
5345 tree perm_dest, perm_stmt, data_ref, first_vect, second_vect;
5346 tree vectype = STMT_VINFO_VECTYPE (vinfo_for_stmt (stmt));
5351 /* Check that the operation is supported. */
5352 if (!vect_strided_load_supported (vectype))
5355 *result_chain = VEC_copy (tree, heap, dr_chain);
5356 for (i = 0; i < exact_log2 (length); i++)
5358 for (j = 0; j < length; j +=2)
5360 first_vect = VEC_index (tree, dr_chain, j);
5361 second_vect = VEC_index (tree, dr_chain, j+1);
5363 /* data_ref = permute_even (first_data_ref, second_data_ref); */
5364 perm_dest = create_tmp_var (vectype, "vect_perm_even");
5365 DECL_GIMPLE_REG_P (perm_dest) = 1;
5366 add_referenced_var (perm_dest);
5368 tmp = build2 (VEC_EXTRACT_EVEN_EXPR, vectype,
5369 first_vect, second_vect);
5370 perm_stmt = build_gimple_modify_stmt (perm_dest, tmp);
5372 data_ref = make_ssa_name (perm_dest, perm_stmt);
5373 GIMPLE_STMT_OPERAND (perm_stmt, 0) = data_ref;
5374 vect_finish_stmt_generation (stmt, perm_stmt, bsi);
5375 mark_symbols_for_renaming (perm_stmt);
5377 VEC_replace (tree, *result_chain, j/2, data_ref);
5379 /* data_ref = permute_odd (first_data_ref, second_data_ref); */
5380 perm_dest = create_tmp_var (vectype, "vect_perm_odd");
5381 DECL_GIMPLE_REG_P (perm_dest) = 1;
5382 add_referenced_var (perm_dest);
5384 tmp = build2 (VEC_EXTRACT_ODD_EXPR, vectype,
5385 first_vect, second_vect);
5386 perm_stmt = build_gimple_modify_stmt (perm_dest, tmp);
5387 data_ref = make_ssa_name (perm_dest, perm_stmt);
5388 GIMPLE_STMT_OPERAND (perm_stmt, 0) = data_ref;
5389 vect_finish_stmt_generation (stmt, perm_stmt, bsi);
5390 mark_symbols_for_renaming (perm_stmt);
5392 VEC_replace (tree, *result_chain, j/2+length/2, data_ref);
5394 dr_chain = VEC_copy (tree, heap, *result_chain);
5400 /* Function vect_transform_strided_load.
5402 Given a chain of input interleaved data-refs (in DR_CHAIN), build statements
5403 to perform their permutation and ascribe the result vectorized statements to
5404 the scalar statements.
5408 vect_transform_strided_load (tree stmt, VEC(tree,heap) *dr_chain, int size,
5409 block_stmt_iterator *bsi)
5411 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
5412 tree first_stmt = DR_GROUP_FIRST_DR (stmt_info);
5413 tree next_stmt, new_stmt;
5414 VEC(tree,heap) *result_chain = NULL;
5415 unsigned int i, gap_count;
5418 /* DR_CHAIN contains input data-refs that are a part of the interleaving.
5419 RESULT_CHAIN is the output of vect_permute_load_chain, it contains permuted
5420 vectors, that are ready for vector computation. */
5421 result_chain = VEC_alloc (tree, heap, size);
5423 if (!vect_permute_load_chain (dr_chain, size, stmt, bsi, &result_chain))
5426 /* Put a permuted data-ref in the VECTORIZED_STMT field.
5427 Since we scan the chain starting from it's first node, their order
5428 corresponds the order of data-refs in RESULT_CHAIN. */
5429 next_stmt = first_stmt;
5431 for (i = 0; VEC_iterate (tree, result_chain, i, tmp_data_ref); i++)
5436 /* Skip the gaps. Loads created for the gaps will be removed by dead
5437 code elimination pass later.
5438 DR_GROUP_GAP is the number of steps in elements from the previous
5439 access (if there is no gap DR_GROUP_GAP is 1). We skip loads that
5440 correspond to the gaps.
5442 if (gap_count < DR_GROUP_GAP (vinfo_for_stmt (next_stmt)))
5450 new_stmt = SSA_NAME_DEF_STMT (tmp_data_ref);
5451 /* We assume that if VEC_STMT is not NULL, this is a case of multiple
5452 copies, and we put the new vector statement in the first available
5454 if (!STMT_VINFO_VEC_STMT (vinfo_for_stmt (next_stmt)))
5455 STMT_VINFO_VEC_STMT (vinfo_for_stmt (next_stmt)) = new_stmt;
5458 tree prev_stmt = STMT_VINFO_VEC_STMT (vinfo_for_stmt (next_stmt));
5459 tree rel_stmt = STMT_VINFO_RELATED_STMT (
5460 vinfo_for_stmt (prev_stmt));
5463 prev_stmt = rel_stmt;
5464 rel_stmt = STMT_VINFO_RELATED_STMT (vinfo_for_stmt (rel_stmt));
5466 STMT_VINFO_RELATED_STMT (vinfo_for_stmt (prev_stmt)) = new_stmt;
5468 next_stmt = DR_GROUP_NEXT_DR (vinfo_for_stmt (next_stmt));
5470 /* If NEXT_STMT accesses the same DR as the previous statement,
5471 put the same TMP_DATA_REF as its vectorized statement; otherwise
5472 get the next data-ref from RESULT_CHAIN. */
5473 if (!next_stmt || !DR_GROUP_SAME_DR_STMT (vinfo_for_stmt (next_stmt)))
5481 /* vectorizable_load.
5483 Check if STMT reads a non scalar data-ref (array/pointer/structure) that
5485 If VEC_STMT is also passed, vectorize the STMT: create a vectorized
5486 stmt to replace it, put it in VEC_STMT, and insert it at BSI.
5487 Return FALSE if not a vectorizable STMT, TRUE otherwise. */
5490 vectorizable_load (tree stmt, block_stmt_iterator *bsi, tree *vec_stmt,
5494 tree vec_dest = NULL;
5495 tree data_ref = NULL;
5497 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
5498 stmt_vec_info prev_stmt_info;
5499 loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info);
5500 struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
5501 struct loop *containing_loop = (bb_for_stmt (stmt))->loop_father;
5502 bool nested_in_vect_loop = nested_in_vect_loop_p (loop, stmt);
5503 struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info), *first_dr;
5504 tree vectype = STMT_VINFO_VECTYPE (stmt_info);
5507 tree new_stmt = NULL_TREE;
5509 enum dr_alignment_support alignment_support_scheme;
5510 tree dataref_ptr = NULL_TREE;
5512 int nunits = TYPE_VECTOR_SUBPARTS (vectype);
5513 int ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits;
5514 int i, j, group_size;
5515 tree msq = NULL_TREE, lsq;
5516 tree offset = NULL_TREE;
5517 tree realignment_token = NULL_TREE;
5518 tree phi = NULL_TREE;
5519 VEC(tree,heap) *dr_chain = NULL;
5520 bool strided_load = false;
5524 bool compute_in_loop = false;
5525 struct loop *at_loop;
5527 bool slp = (slp_node != NULL);
5529 /* FORNOW: SLP with multiple types is not supported. The SLP analysis verifies
5530 this, so we can safely override NCOPIES with 1 here. */
5534 gcc_assert (ncopies >= 1);
5536 /* FORNOW. This restriction should be relaxed. */
5537 if (nested_in_vect_loop && ncopies > 1)
5539 if (vect_print_dump_info (REPORT_DETAILS))
5540 fprintf (vect_dump, "multiple types in nested loop.");
5544 if (!STMT_VINFO_RELEVANT_P (stmt_info))
5547 if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_loop_def)
5550 /* Is vectorizable load? */
5551 if (TREE_CODE (stmt) != GIMPLE_MODIFY_STMT)
5554 scalar_dest = GIMPLE_STMT_OPERAND (stmt, 0);
5555 if (TREE_CODE (scalar_dest) != SSA_NAME)
5558 op = GIMPLE_STMT_OPERAND (stmt, 1);
5559 if (TREE_CODE (op) != ARRAY_REF
5560 && TREE_CODE (op) != INDIRECT_REF
5561 && !STMT_VINFO_STRIDED_ACCESS (stmt_info))
5564 if (!STMT_VINFO_DATA_REF (stmt_info))
5567 scalar_type = TREE_TYPE (DR_REF (dr));
5568 mode = (int) TYPE_MODE (vectype);
5570 /* FORNOW. In some cases can vectorize even if data-type not supported
5571 (e.g. - data copies). */
5572 if (optab_handler (mov_optab, mode)->insn_code == CODE_FOR_nothing)
5574 if (vect_print_dump_info (REPORT_DETAILS))
5575 fprintf (vect_dump, "Aligned load, but unsupported type.");
5579 /* Check if the load is a part of an interleaving chain. */
5580 if (STMT_VINFO_STRIDED_ACCESS (stmt_info))
5582 strided_load = true;
5584 gcc_assert (! nested_in_vect_loop);
5586 /* Check if interleaving is supported. */
5587 if (!vect_strided_load_supported (vectype)
5588 && !PURE_SLP_STMT (stmt_info) && !slp)
5592 if (!vec_stmt) /* transformation not required. */
5594 STMT_VINFO_TYPE (stmt_info) = load_vec_info_type;
5595 vect_model_load_cost (stmt_info, ncopies, NULL);
5599 if (vect_print_dump_info (REPORT_DETAILS))
5600 fprintf (vect_dump, "transform load.");
5606 first_stmt = DR_GROUP_FIRST_DR (stmt_info);
5607 /* Check if the chain of loads is already vectorized. */
5608 if (STMT_VINFO_VEC_STMT (vinfo_for_stmt (first_stmt)))
5610 *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info);
5613 first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt));
5614 group_size = DR_GROUP_SIZE (vinfo_for_stmt (first_stmt));
5615 dr_chain = VEC_alloc (tree, heap, group_size);
5617 /* VEC_NUM is the number of vect stmts to be created for this group. */
5620 strided_load = false;
5621 vec_num = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node);
5624 vec_num = group_size;
5630 group_size = vec_num = 1;
5633 alignment_support_scheme = vect_supportable_dr_alignment (first_dr);
5634 gcc_assert (alignment_support_scheme);
5636 /* In case the vectorization factor (VF) is bigger than the number
5637 of elements that we can fit in a vectype (nunits), we have to generate
5638 more than one vector stmt - i.e - we need to "unroll" the
5639 vector stmt by a factor VF/nunits. In doing so, we record a pointer
5640 from one copy of the vector stmt to the next, in the field
5641 STMT_VINFO_RELATED_STMT. This is necessary in order to allow following
5642 stages to find the correct vector defs to be used when vectorizing
5643 stmts that use the defs of the current stmt. The example below illustrates
5644 the vectorization process when VF=16 and nunits=4 (i.e - we need to create
5645 4 vectorized stmts):
5647 before vectorization:
5648 RELATED_STMT VEC_STMT
5652 step 1: vectorize stmt S1:
5653 We first create the vector stmt VS1_0, and, as usual, record a
5654 pointer to it in the STMT_VINFO_VEC_STMT of the scalar stmt S1.
5655 Next, we create the vector stmt VS1_1, and record a pointer to
5656 it in the STMT_VINFO_RELATED_STMT of the vector stmt VS1_0.
5657 Similarly, for VS1_2 and VS1_3. This is the resulting chain of
5659 RELATED_STMT VEC_STMT
5660 VS1_0: vx0 = memref0 VS1_1 -
5661 VS1_1: vx1 = memref1 VS1_2 -
5662 VS1_2: vx2 = memref2 VS1_3 -
5663 VS1_3: vx3 = memref3 - -
5664 S1: x = load - VS1_0
5667 See in documentation in vect_get_vec_def_for_stmt_copy for how the
5668 information we recorded in RELATED_STMT field is used to vectorize
5671 /* In case of interleaving (non-unit strided access):
5678 Vectorized loads are created in the order of memory accesses
5679 starting from the access of the first stmt of the chain:
5682 VS2: vx1 = &base + vec_size*1
5683 VS3: vx3 = &base + vec_size*2
5684 VS4: vx4 = &base + vec_size*3
5686 Then permutation statements are generated:
5688 VS5: vx5 = VEC_EXTRACT_EVEN_EXPR < vx0, vx1 >
5689 VS6: vx6 = VEC_EXTRACT_ODD_EXPR < vx0, vx1 >
5692 And they are put in STMT_VINFO_VEC_STMT of the corresponding scalar stmts
5693 (the order of the data-refs in the output of vect_permute_load_chain
5694 corresponds to the order of scalar stmts in the interleaving chain - see
5695 the documentation of vect_permute_load_chain()).
5696 The generation of permutation stmts and recording them in
5697 STMT_VINFO_VEC_STMT is done in vect_transform_strided_load().
5699 In case of both multiple types and interleaving, the vector loads and
5700 permutation stmts above are created for every copy. The result vector stmts
5701 are put in STMT_VINFO_VEC_STMT for the first copy and in the corresponding
5702 STMT_VINFO_RELATED_STMT for the next copies. */
5704 /* If the data reference is aligned (dr_aligned) or potentially unaligned
5705 on a target that supports unaligned accesses (dr_unaligned_supported)
5706 we generate the following code:
5710 p = p + indx * vectype_size;
5715 Otherwise, the data reference is potentially unaligned on a target that
5716 does not support unaligned accesses (dr_explicit_realign_optimized) -
5717 then generate the following code, in which the data in each iteration is
5718 obtained by two vector loads, one from the previous iteration, and one
5719 from the current iteration:
5721 msq_init = *(floor(p1))
5722 p2 = initial_addr + VS - 1;
5723 realignment_token = call target_builtin;
5726 p2 = p2 + indx * vectype_size
5728 vec_dest = realign_load (msq, lsq, realignment_token)
5733 /* If the misalignment remains the same throughout the execution of the
5734 loop, we can create the init_addr and permutation mask at the loop
5735 preheader. Otherwise, it needs to be created inside the loop.
5736 This can only occur when vectorizing memory accesses in the inner-loop
5737 nested within an outer-loop that is being vectorized. */
5739 if (nested_in_vect_loop_p (loop, stmt)
5740 && (TREE_INT_CST_LOW (DR_STEP (dr)) % UNITS_PER_SIMD_WORD != 0))
5742 gcc_assert (alignment_support_scheme != dr_explicit_realign_optimized);
5743 compute_in_loop = true;
5746 if ((alignment_support_scheme == dr_explicit_realign_optimized
5747 || alignment_support_scheme == dr_explicit_realign)
5748 && !compute_in_loop)
5750 msq = vect_setup_realignment (first_stmt, bsi, &realignment_token,
5751 alignment_support_scheme, NULL_TREE,
5753 if (alignment_support_scheme == dr_explicit_realign_optimized)
5755 phi = SSA_NAME_DEF_STMT (msq);
5756 offset = size_int (TYPE_VECTOR_SUBPARTS (vectype) - 1);
5762 prev_stmt_info = NULL;
5763 for (j = 0; j < ncopies; j++)
5765 /* 1. Create the vector pointer update chain. */
5767 dataref_ptr = vect_create_data_ref_ptr (first_stmt,
5769 &dummy, &ptr_incr, false,
5773 bump_vector_ptr (dataref_ptr, ptr_incr, bsi, stmt, NULL_TREE);
5775 for (i = 0; i < vec_num; i++)
5778 dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, bsi, stmt,
5781 /* 2. Create the vector-load in the loop. */
5782 switch (alignment_support_scheme)
5785 gcc_assert (aligned_access_p (first_dr));
5786 data_ref = build_fold_indirect_ref (dataref_ptr);
5788 case dr_unaligned_supported:
5790 int mis = DR_MISALIGNMENT (first_dr);
5791 tree tmis = (mis == -1 ? size_zero_node : size_int (mis));
5793 tmis = size_binop (MULT_EXPR, tmis, size_int(BITS_PER_UNIT));
5795 build2 (MISALIGNED_INDIRECT_REF, vectype, dataref_ptr, tmis);
5798 case dr_explicit_realign:
5801 tree vs_minus_1 = size_int (TYPE_VECTOR_SUBPARTS (vectype) - 1);
5803 if (compute_in_loop)
5804 msq = vect_setup_realignment (first_stmt, bsi,
5806 dr_explicit_realign,
5809 data_ref = build1 (ALIGN_INDIRECT_REF, vectype, dataref_ptr);
5810 vec_dest = vect_create_destination_var (scalar_dest, vectype);
5811 new_stmt = build_gimple_modify_stmt (vec_dest, data_ref);
5812 new_temp = make_ssa_name (vec_dest, new_stmt);
5813 GIMPLE_STMT_OPERAND (new_stmt, 0) = new_temp;
5814 vect_finish_stmt_generation (stmt, new_stmt, bsi);
5815 copy_virtual_operands (new_stmt, stmt);
5816 mark_symbols_for_renaming (new_stmt);
5819 bump = size_binop (MULT_EXPR, vs_minus_1,
5820 TYPE_SIZE_UNIT (scalar_type));
5821 ptr = bump_vector_ptr (dataref_ptr, NULL_TREE, bsi, stmt, bump);
5822 data_ref = build1 (ALIGN_INDIRECT_REF, vectype, ptr);
5825 case dr_explicit_realign_optimized:
5826 data_ref = build1 (ALIGN_INDIRECT_REF, vectype, dataref_ptr);
5831 vec_dest = vect_create_destination_var (scalar_dest, vectype);
5832 new_stmt = build_gimple_modify_stmt (vec_dest, data_ref);
5833 new_temp = make_ssa_name (vec_dest, new_stmt);
5834 GIMPLE_STMT_OPERAND (new_stmt, 0) = new_temp;
5835 vect_finish_stmt_generation (stmt, new_stmt, bsi);
5836 mark_symbols_for_renaming (new_stmt);
5838 /* 3. Handle explicit realignment if necessary/supported. Create in
5839 loop: vec_dest = realign_load (msq, lsq, realignment_token) */
5840 if (alignment_support_scheme == dr_explicit_realign_optimized
5841 || alignment_support_scheme == dr_explicit_realign)
5843 lsq = GIMPLE_STMT_OPERAND (new_stmt, 0);
5844 if (!realignment_token)
5845 realignment_token = dataref_ptr;
5846 vec_dest = vect_create_destination_var (scalar_dest, vectype);
5847 new_stmt = build3 (REALIGN_LOAD_EXPR, vectype, msq, lsq,
5849 new_stmt = build_gimple_modify_stmt (vec_dest, new_stmt);
5850 new_temp = make_ssa_name (vec_dest, new_stmt);
5851 GIMPLE_STMT_OPERAND (new_stmt, 0) = new_temp;
5852 vect_finish_stmt_generation (stmt, new_stmt, bsi);
5854 if (alignment_support_scheme == dr_explicit_realign_optimized)
5856 if (i == vec_num - 1 && j == ncopies - 1)
5857 add_phi_arg (phi, lsq, loop_latch_edge (containing_loop));
5862 /* 4. Handle invariant-load. */
5865 gcc_assert (!strided_load);
5866 gcc_assert (nested_in_vect_loop_p (loop, stmt));
5871 tree vec_inv, bitpos, bitsize = TYPE_SIZE (scalar_type);
5873 /* CHECKME: bitpos depends on endianess? */
5874 bitpos = bitsize_zero_node;
5875 vec_inv = build3 (BIT_FIELD_REF, scalar_type, new_temp,
5878 vect_create_destination_var (scalar_dest, NULL_TREE);
5879 new_stmt = build_gimple_modify_stmt (vec_dest, vec_inv);
5880 new_temp = make_ssa_name (vec_dest, new_stmt);
5881 GIMPLE_STMT_OPERAND (new_stmt, 0) = new_temp;
5882 vect_finish_stmt_generation (stmt, new_stmt, bsi);
5884 for (k = nunits - 1; k >= 0; --k)
5885 t = tree_cons (NULL_TREE, new_temp, t);
5886 /* FIXME: use build_constructor directly. */
5887 vec_inv = build_constructor_from_list (vectype, t);
5888 new_temp = vect_init_vector (stmt, vec_inv, vectype, bsi);
5889 new_stmt = SSA_NAME_DEF_STMT (new_temp);
5892 gcc_unreachable (); /* FORNOW. */
5895 /* Collect vector loads and later create their permutation in
5896 vect_transform_strided_load (). */
5898 VEC_quick_push (tree, dr_chain, new_temp);
5900 /* Store vector loads in the corresponding SLP_NODE. */
5902 VEC_quick_push (tree, SLP_TREE_VEC_STMTS (slp_node), new_stmt);
5905 /* FORNOW: SLP with multiple types is unsupported. */
5911 if (!vect_transform_strided_load (stmt, dr_chain, group_size, bsi))
5913 *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info);
5914 dr_chain = VEC_alloc (tree, heap, group_size);
5919 STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt;
5921 STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt;
5922 prev_stmt_info = vinfo_for_stmt (new_stmt);
5930 /* Function vectorizable_live_operation.
5932 STMT computes a value that is used outside the loop. Check if
5933 it can be supported. */
5936 vectorizable_live_operation (tree stmt,
5937 block_stmt_iterator *bsi ATTRIBUTE_UNUSED,
5938 tree *vec_stmt ATTRIBUTE_UNUSED)
5941 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
5942 loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info);
5943 struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
5948 enum vect_def_type dt;
5950 gcc_assert (STMT_VINFO_LIVE_P (stmt_info));
5952 if (STMT_VINFO_DEF_TYPE (stmt_info) == vect_reduction_def)
5955 if (TREE_CODE (stmt) != GIMPLE_MODIFY_STMT)
5958 if (TREE_CODE (GIMPLE_STMT_OPERAND (stmt, 0)) != SSA_NAME)
5961 /* FORNOW. CHECKME. */
5962 if (nested_in_vect_loop_p (loop, stmt))
5965 operation = GIMPLE_STMT_OPERAND (stmt, 1);
5966 op_type = TREE_OPERAND_LENGTH (operation);
5968 /* FORNOW: support only if all uses are invariant. This means
5969 that the scalar operations can remain in place, unvectorized.
5970 The original last scalar value that they compute will be used. */
5972 for (i = 0; i < op_type; i++)
5974 op = TREE_OPERAND (operation, i);
5975 if (op && !vect_is_simple_use (op, loop_vinfo, &def_stmt, &def, &dt))
5977 if (vect_print_dump_info (REPORT_DETAILS))
5978 fprintf (vect_dump, "use not simple.");
5982 if (dt != vect_invariant_def && dt != vect_constant_def)
5986 /* No transformation is required for the cases we currently support. */
5991 /* Function vect_is_simple_cond.
5994 LOOP - the loop that is being vectorized.
5995 COND - Condition that is checked for simple use.
5997 Returns whether a COND can be vectorized. Checks whether
5998 condition operands are supportable using vec_is_simple_use. */
6001 vect_is_simple_cond (tree cond, loop_vec_info loop_vinfo)
6005 enum vect_def_type dt;
6007 if (!COMPARISON_CLASS_P (cond))
6010 lhs = TREE_OPERAND (cond, 0);
6011 rhs = TREE_OPERAND (cond, 1);
6013 if (TREE_CODE (lhs) == SSA_NAME)
6015 tree lhs_def_stmt = SSA_NAME_DEF_STMT (lhs);
6016 if (!vect_is_simple_use (lhs, loop_vinfo, &lhs_def_stmt, &def, &dt))
6019 else if (TREE_CODE (lhs) != INTEGER_CST && TREE_CODE (lhs) != REAL_CST
6020 && TREE_CODE (lhs) != FIXED_CST)
6023 if (TREE_CODE (rhs) == SSA_NAME)
6025 tree rhs_def_stmt = SSA_NAME_DEF_STMT (rhs);
6026 if (!vect_is_simple_use (rhs, loop_vinfo, &rhs_def_stmt, &def, &dt))
6029 else if (TREE_CODE (rhs) != INTEGER_CST && TREE_CODE (rhs) != REAL_CST
6030 && TREE_CODE (rhs) != FIXED_CST)
6036 /* vectorizable_condition.
6038 Check if STMT is conditional modify expression that can be vectorized.
6039 If VEC_STMT is also passed, vectorize the STMT: create a vectorized
6040 stmt using VEC_COND_EXPR to replace it, put it in VEC_STMT, and insert it
6043 Return FALSE if not a vectorizable STMT, TRUE otherwise. */
6046 vectorizable_condition (tree stmt, block_stmt_iterator *bsi, tree *vec_stmt)
6048 tree scalar_dest = NULL_TREE;
6049 tree vec_dest = NULL_TREE;
6050 tree op = NULL_TREE;
6051 tree cond_expr, then_clause, else_clause;
6052 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
6053 tree vectype = STMT_VINFO_VECTYPE (stmt_info);
6054 tree vec_cond_lhs, vec_cond_rhs, vec_then_clause, vec_else_clause;
6055 tree vec_compare, vec_cond_expr;
6057 loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info);
6058 enum machine_mode vec_mode;
6060 enum vect_def_type dt;
6061 int nunits = TYPE_VECTOR_SUBPARTS (vectype);
6062 int ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits;
6064 gcc_assert (ncopies >= 1);
6066 return false; /* FORNOW */
6068 if (!STMT_VINFO_RELEVANT_P (stmt_info))
6071 if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_loop_def)
6074 /* FORNOW: SLP not supported. */
6075 if (STMT_SLP_TYPE (stmt_info))
6078 /* FORNOW: not yet supported. */
6079 if (STMT_VINFO_LIVE_P (stmt_info))
6081 if (vect_print_dump_info (REPORT_DETAILS))
6082 fprintf (vect_dump, "value used after loop.");
6086 /* Is vectorizable conditional operation? */
6087 if (TREE_CODE (stmt) != GIMPLE_MODIFY_STMT)
6090 op = GIMPLE_STMT_OPERAND (stmt, 1);
6092 if (TREE_CODE (op) != COND_EXPR)
6095 cond_expr = TREE_OPERAND (op, 0);
6096 then_clause = TREE_OPERAND (op, 1);
6097 else_clause = TREE_OPERAND (op, 2);
6099 if (!vect_is_simple_cond (cond_expr, loop_vinfo))
6102 /* We do not handle two different vector types for the condition
6104 if (TREE_TYPE (TREE_OPERAND (cond_expr, 0)) != TREE_TYPE (vectype))
6107 if (TREE_CODE (then_clause) == SSA_NAME)
6109 tree then_def_stmt = SSA_NAME_DEF_STMT (then_clause);
6110 if (!vect_is_simple_use (then_clause, loop_vinfo,
6111 &then_def_stmt, &def, &dt))
6114 else if (TREE_CODE (then_clause) != INTEGER_CST
6115 && TREE_CODE (then_clause) != REAL_CST
6116 && TREE_CODE (then_clause) != FIXED_CST)
6119 if (TREE_CODE (else_clause) == SSA_NAME)
6121 tree else_def_stmt = SSA_NAME_DEF_STMT (else_clause);
6122 if (!vect_is_simple_use (else_clause, loop_vinfo,
6123 &else_def_stmt, &def, &dt))
6126 else if (TREE_CODE (else_clause) != INTEGER_CST
6127 && TREE_CODE (else_clause) != REAL_CST
6128 && TREE_CODE (else_clause) != FIXED_CST)
6132 vec_mode = TYPE_MODE (vectype);
6136 STMT_VINFO_TYPE (stmt_info) = condition_vec_info_type;
6137 return expand_vec_cond_expr_p (op, vec_mode);
6143 scalar_dest = GIMPLE_STMT_OPERAND (stmt, 0);
6144 vec_dest = vect_create_destination_var (scalar_dest, vectype);
6146 /* Handle cond expr. */
6148 vect_get_vec_def_for_operand (TREE_OPERAND (cond_expr, 0), stmt, NULL);
6150 vect_get_vec_def_for_operand (TREE_OPERAND (cond_expr, 1), stmt, NULL);
6151 vec_then_clause = vect_get_vec_def_for_operand (then_clause, stmt, NULL);
6152 vec_else_clause = vect_get_vec_def_for_operand (else_clause, stmt, NULL);
6154 /* Arguments are ready. create the new vector stmt. */
6155 vec_compare = build2 (TREE_CODE (cond_expr), vectype,
6156 vec_cond_lhs, vec_cond_rhs);
6157 vec_cond_expr = build3 (VEC_COND_EXPR, vectype,
6158 vec_compare, vec_then_clause, vec_else_clause);
6160 *vec_stmt = build_gimple_modify_stmt (vec_dest, vec_cond_expr);
6161 new_temp = make_ssa_name (vec_dest, *vec_stmt);
6162 GIMPLE_STMT_OPERAND (*vec_stmt, 0) = new_temp;
6163 vect_finish_stmt_generation (stmt, *vec_stmt, bsi);
6169 /* Function vect_transform_stmt.
6171 Create a vectorized stmt to replace STMT, and insert it at BSI. */
6174 vect_transform_stmt (tree stmt, block_stmt_iterator *bsi, bool *strided_store,
6177 bool is_store = false;
6178 tree vec_stmt = NULL_TREE;
6179 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
6180 tree orig_stmt_in_pattern;
6183 switch (STMT_VINFO_TYPE (stmt_info))
6185 case type_demotion_vec_info_type:
6186 gcc_assert (!slp_node);
6187 done = vectorizable_type_demotion (stmt, bsi, &vec_stmt);
6191 case type_promotion_vec_info_type:
6192 gcc_assert (!slp_node);
6193 done = vectorizable_type_promotion (stmt, bsi, &vec_stmt);
6197 case type_conversion_vec_info_type:
6198 done = vectorizable_conversion (stmt, bsi, &vec_stmt, slp_node);
6202 case induc_vec_info_type:
6203 gcc_assert (!slp_node);
6204 done = vectorizable_induction (stmt, bsi, &vec_stmt);
6208 case op_vec_info_type:
6209 done = vectorizable_operation (stmt, bsi, &vec_stmt, slp_node);
6213 case assignment_vec_info_type:
6214 done = vectorizable_assignment (stmt, bsi, &vec_stmt, slp_node);
6218 case load_vec_info_type:
6219 done = vectorizable_load (stmt, bsi, &vec_stmt, slp_node);
6223 case store_vec_info_type:
6224 done = vectorizable_store (stmt, bsi, &vec_stmt, slp_node);
6226 if (STMT_VINFO_STRIDED_ACCESS (stmt_info))
6228 /* In case of interleaving, the whole chain is vectorized when the
6229 last store in the chain is reached. Store stmts before the last
6230 one are skipped, and there vec_stmt_info shouldn't be freed
6232 *strided_store = true;
6233 if (STMT_VINFO_VEC_STMT (stmt_info))
6240 case condition_vec_info_type:
6241 gcc_assert (!slp_node);
6242 done = vectorizable_condition (stmt, bsi, &vec_stmt);
6246 case call_vec_info_type:
6247 gcc_assert (!slp_node);
6248 done = vectorizable_call (stmt, bsi, &vec_stmt);
6251 case reduc_vec_info_type:
6252 gcc_assert (!slp_node);
6253 done = vectorizable_reduction (stmt, bsi, &vec_stmt);
6258 if (!STMT_VINFO_LIVE_P (stmt_info))
6260 if (vect_print_dump_info (REPORT_DETAILS))
6261 fprintf (vect_dump, "stmt not supported.");
6266 if (STMT_VINFO_LIVE_P (stmt_info)
6267 && STMT_VINFO_TYPE (stmt_info) != reduc_vec_info_type)
6269 done = vectorizable_live_operation (stmt, bsi, &vec_stmt);
6275 STMT_VINFO_VEC_STMT (stmt_info) = vec_stmt;
6276 orig_stmt_in_pattern = STMT_VINFO_RELATED_STMT (stmt_info);
6277 if (orig_stmt_in_pattern)
6279 stmt_vec_info stmt_vinfo = vinfo_for_stmt (orig_stmt_in_pattern);
6280 /* STMT was inserted by the vectorizer to replace a computation idiom.
6281 ORIG_STMT_IN_PATTERN is a stmt in the original sequence that
6282 computed this idiom. We need to record a pointer to VEC_STMT in
6283 the stmt_info of ORIG_STMT_IN_PATTERN. See more details in the
6284 documentation of vect_pattern_recog. */
6285 if (STMT_VINFO_IN_PATTERN_P (stmt_vinfo))
6287 gcc_assert (STMT_VINFO_RELATED_STMT (stmt_vinfo) == stmt);
6288 STMT_VINFO_VEC_STMT (stmt_vinfo) = vec_stmt;
6297 /* This function builds ni_name = number of iterations loop executes
6298 on the loop preheader. */
6301 vect_build_loop_niters (loop_vec_info loop_vinfo)
6303 tree ni_name, stmt, var;
6305 struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
6306 tree ni = unshare_expr (LOOP_VINFO_NITERS (loop_vinfo));
6308 var = create_tmp_var (TREE_TYPE (ni), "niters");
6309 add_referenced_var (var);
6310 ni_name = force_gimple_operand (ni, &stmt, false, var);
6312 pe = loop_preheader_edge (loop);
6315 basic_block new_bb = bsi_insert_on_edge_immediate (pe, stmt);
6316 gcc_assert (!new_bb);
6323 /* This function generates the following statements:
6325 ni_name = number of iterations loop executes
6326 ratio = ni_name / vf
6327 ratio_mult_vf_name = ratio * vf
6329 and places them at the loop preheader edge. */
6332 vect_generate_tmps_on_preheader (loop_vec_info loop_vinfo,
6334 tree *ratio_mult_vf_name_ptr,
6335 tree *ratio_name_ptr)
6343 tree ratio_mult_vf_name;
6344 struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
6345 tree ni = LOOP_VINFO_NITERS (loop_vinfo);
6346 int vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo);
6349 pe = loop_preheader_edge (loop);
6351 /* Generate temporary variable that contains
6352 number of iterations loop executes. */
6354 ni_name = vect_build_loop_niters (loop_vinfo);
6355 log_vf = build_int_cst (TREE_TYPE (ni), exact_log2 (vf));
6357 /* Create: ratio = ni >> log2(vf) */
6359 ratio_name = fold_build2 (RSHIFT_EXPR, TREE_TYPE (ni_name), ni_name, log_vf);
6360 if (!is_gimple_val (ratio_name))
6362 var = create_tmp_var (TREE_TYPE (ni), "bnd");
6363 add_referenced_var (var);
6365 ratio_name = force_gimple_operand (ratio_name, &stmt, true, var);
6366 pe = loop_preheader_edge (loop);
6367 new_bb = bsi_insert_on_edge_immediate (pe, stmt);
6368 gcc_assert (!new_bb);
6371 /* Create: ratio_mult_vf = ratio << log2 (vf). */
6373 ratio_mult_vf_name = fold_build2 (LSHIFT_EXPR, TREE_TYPE (ratio_name),
6374 ratio_name, log_vf);
6375 if (!is_gimple_val (ratio_mult_vf_name))
6377 var = create_tmp_var (TREE_TYPE (ni), "ratio_mult_vf");
6378 add_referenced_var (var);
6380 ratio_mult_vf_name = force_gimple_operand (ratio_mult_vf_name, &stmt,
6382 pe = loop_preheader_edge (loop);
6383 new_bb = bsi_insert_on_edge_immediate (pe, stmt);
6384 gcc_assert (!new_bb);
6387 *ni_name_ptr = ni_name;
6388 *ratio_mult_vf_name_ptr = ratio_mult_vf_name;
6389 *ratio_name_ptr = ratio_name;
6395 /* Function vect_update_ivs_after_vectorizer.
6397 "Advance" the induction variables of LOOP to the value they should take
6398 after the execution of LOOP. This is currently necessary because the
6399 vectorizer does not handle induction variables that are used after the
6400 loop. Such a situation occurs when the last iterations of LOOP are
6402 1. We introduced new uses after LOOP for IVs that were not originally used
6403 after LOOP: the IVs of LOOP are now used by an epilog loop.
6404 2. LOOP is going to be vectorized; this means that it will iterate N/VF
6405 times, whereas the loop IVs should be bumped N times.
6408 - LOOP - a loop that is going to be vectorized. The last few iterations
6409 of LOOP were peeled.
6410 - NITERS - the number of iterations that LOOP executes (before it is
6411 vectorized). i.e, the number of times the ivs should be bumped.
6412 - UPDATE_E - a successor edge of LOOP->exit that is on the (only) path
6413 coming out from LOOP on which there are uses of the LOOP ivs
6414 (this is the path from LOOP->exit to epilog_loop->preheader).
6416 The new definitions of the ivs are placed in LOOP->exit.
6417 The phi args associated with the edge UPDATE_E in the bb
6418 UPDATE_E->dest are updated accordingly.
6420 Assumption 1: Like the rest of the vectorizer, this function assumes
6421 a single loop exit that has a single predecessor.
6423 Assumption 2: The phi nodes in the LOOP header and in update_bb are
6424 organized in the same order.
6426 Assumption 3: The access function of the ivs is simple enough (see
6427 vect_can_advance_ivs_p). This assumption will be relaxed in the future.
6429 Assumption 4: Exactly one of the successors of LOOP exit-bb is on a path
6430 coming out of LOOP on which the ivs of LOOP are used (this is the path
6431 that leads to the epilog loop; other paths skip the epilog loop). This
6432 path starts with the edge UPDATE_E, and its destination (denoted update_bb)
6433 needs to have its phis updated.
6437 vect_update_ivs_after_vectorizer (loop_vec_info loop_vinfo, tree niters,
6440 struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
6441 basic_block exit_bb = single_exit (loop)->dest;
6443 basic_block update_bb = update_e->dest;
6445 /* gcc_assert (vect_can_advance_ivs_p (loop_vinfo)); */
6447 /* Make sure there exists a single-predecessor exit bb: */
6448 gcc_assert (single_pred_p (exit_bb));
6450 for (phi = phi_nodes (loop->header), phi1 = phi_nodes (update_bb);
6452 phi = PHI_CHAIN (phi), phi1 = PHI_CHAIN (phi1))
6454 tree access_fn = NULL;
6455 tree evolution_part;
6458 tree var, ni, ni_name;
6459 block_stmt_iterator last_bsi;
6461 if (vect_print_dump_info (REPORT_DETAILS))
6463 fprintf (vect_dump, "vect_update_ivs_after_vectorizer: phi: ");
6464 print_generic_expr (vect_dump, phi, TDF_SLIM);
6467 /* Skip virtual phi's. */
6468 if (!is_gimple_reg (SSA_NAME_VAR (PHI_RESULT (phi))))
6470 if (vect_print_dump_info (REPORT_DETAILS))
6471 fprintf (vect_dump, "virtual phi. skip.");
6475 /* Skip reduction phis. */
6476 if (STMT_VINFO_DEF_TYPE (vinfo_for_stmt (phi)) == vect_reduction_def)
6478 if (vect_print_dump_info (REPORT_DETAILS))
6479 fprintf (vect_dump, "reduc phi. skip.");
6483 access_fn = analyze_scalar_evolution (loop, PHI_RESULT (phi));
6484 gcc_assert (access_fn);
6486 unshare_expr (evolution_part_in_loop_num (access_fn, loop->num));
6487 gcc_assert (evolution_part != NULL_TREE);
6489 /* FORNOW: We do not support IVs whose evolution function is a polynomial
6490 of degree >= 2 or exponential. */
6491 gcc_assert (!tree_is_chrec (evolution_part));
6493 step_expr = evolution_part;
6494 init_expr = unshare_expr (initial_condition_in_loop_num (access_fn,
6497 if (POINTER_TYPE_P (TREE_TYPE (init_expr)))
6498 ni = fold_build2 (POINTER_PLUS_EXPR, TREE_TYPE (init_expr),
6500 fold_convert (sizetype,
6501 fold_build2 (MULT_EXPR, TREE_TYPE (niters),
6502 niters, step_expr)));
6504 ni = fold_build2 (PLUS_EXPR, TREE_TYPE (init_expr),
6505 fold_build2 (MULT_EXPR, TREE_TYPE (init_expr),
6506 fold_convert (TREE_TYPE (init_expr),
6513 var = create_tmp_var (TREE_TYPE (init_expr), "tmp");
6514 add_referenced_var (var);
6516 last_bsi = bsi_last (exit_bb);
6517 ni_name = force_gimple_operand_bsi (&last_bsi, ni, false, var,
6518 true, BSI_SAME_STMT);
6520 /* Fix phi expressions in the successor bb. */
6521 SET_PHI_ARG_DEF (phi1, update_e->dest_idx, ni_name);
6525 /* Return the more conservative threshold between the
6526 min_profitable_iters returned by the cost model and the user
6527 specified threshold, if provided. */
6530 conservative_cost_threshold (loop_vec_info loop_vinfo,
6531 int min_profitable_iters)
6534 int min_scalar_loop_bound;
6536 min_scalar_loop_bound = ((PARAM_VALUE (PARAM_MIN_VECT_LOOP_BOUND)
6537 * LOOP_VINFO_VECT_FACTOR (loop_vinfo)) - 1);
6539 /* Use the cost model only if it is more conservative than user specified
6541 th = (unsigned) min_scalar_loop_bound;
6542 if (min_profitable_iters
6543 && (!min_scalar_loop_bound
6544 || min_profitable_iters > min_scalar_loop_bound))
6545 th = (unsigned) min_profitable_iters;
6547 if (th && vect_print_dump_info (REPORT_COST))
6548 fprintf (vect_dump, "Vectorization may not be profitable.");
6553 /* Function vect_do_peeling_for_loop_bound
6555 Peel the last iterations of the loop represented by LOOP_VINFO.
6556 The peeled iterations form a new epilog loop. Given that the loop now
6557 iterates NITERS times, the new epilog loop iterates
6558 NITERS % VECTORIZATION_FACTOR times.
6560 The original loop will later be made to iterate
6561 NITERS / VECTORIZATION_FACTOR times (this value is placed into RATIO). */
6564 vect_do_peeling_for_loop_bound (loop_vec_info loop_vinfo, tree *ratio)
6566 tree ni_name, ratio_mult_vf_name;
6567 struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
6568 struct loop *new_loop;
6570 basic_block preheader;
6572 bool check_profitability = false;
6573 unsigned int th = 0;
6574 int min_profitable_iters;
6576 if (vect_print_dump_info (REPORT_DETAILS))
6577 fprintf (vect_dump, "=== vect_do_peeling_for_loop_bound ===");
6579 initialize_original_copy_tables ();
6581 /* Generate the following variables on the preheader of original loop:
6583 ni_name = number of iteration the original loop executes
6584 ratio = ni_name / vf
6585 ratio_mult_vf_name = ratio * vf */
6586 vect_generate_tmps_on_preheader (loop_vinfo, &ni_name,
6587 &ratio_mult_vf_name, ratio);
6589 loop_num = loop->num;
6591 /* If cost model check not done during versioning and
6592 peeling for alignment. */
6593 if (!VEC_length (tree, LOOP_VINFO_MAY_MISALIGN_STMTS (loop_vinfo))
6594 && !VEC_length (ddr_p, LOOP_VINFO_MAY_ALIAS_DDRS (loop_vinfo))
6595 && !LOOP_PEELING_FOR_ALIGNMENT (loop_vinfo))
6597 check_profitability = true;
6599 /* Get profitability threshold for vectorized loop. */
6600 min_profitable_iters = LOOP_VINFO_COST_MODEL_MIN_ITERS (loop_vinfo);
6602 th = conservative_cost_threshold (loop_vinfo,
6603 min_profitable_iters);
6606 new_loop = slpeel_tree_peel_loop_to_edge (loop, single_exit (loop),
6607 ratio_mult_vf_name, ni_name, false,
6608 th, check_profitability);
6609 gcc_assert (new_loop);
6610 gcc_assert (loop_num == loop->num);
6611 #ifdef ENABLE_CHECKING
6612 slpeel_verify_cfg_after_peeling (loop, new_loop);
6615 /* A guard that controls whether the new_loop is to be executed or skipped
6616 is placed in LOOP->exit. LOOP->exit therefore has two successors - one
6617 is the preheader of NEW_LOOP, where the IVs from LOOP are used. The other
6618 is a bb after NEW_LOOP, where these IVs are not used. Find the edge that
6619 is on the path where the LOOP IVs are used and need to be updated. */
6621 preheader = loop_preheader_edge (new_loop)->src;
6622 if (EDGE_PRED (preheader, 0)->src == single_exit (loop)->dest)
6623 update_e = EDGE_PRED (preheader, 0);
6625 update_e = EDGE_PRED (preheader, 1);
6627 /* Update IVs of original loop as if they were advanced
6628 by ratio_mult_vf_name steps. */
6629 vect_update_ivs_after_vectorizer (loop_vinfo, ratio_mult_vf_name, update_e);
6631 /* After peeling we have to reset scalar evolution analyzer. */
6634 free_original_copy_tables ();
6638 /* Function vect_gen_niters_for_prolog_loop
6640 Set the number of iterations for the loop represented by LOOP_VINFO
6641 to the minimum between LOOP_NITERS (the original iteration count of the loop)
6642 and the misalignment of DR - the data reference recorded in
6643 LOOP_VINFO_UNALIGNED_DR (LOOP_VINFO). As a result, after the execution of
6644 this loop, the data reference DR will refer to an aligned location.
6646 The following computation is generated:
6648 If the misalignment of DR is known at compile time:
6649 addr_mis = int mis = DR_MISALIGNMENT (dr);
6650 Else, compute address misalignment in bytes:
6651 addr_mis = addr & (vectype_size - 1)
6653 prolog_niters = min ( LOOP_NITERS , (VF - addr_mis/elem_size)&(VF-1) )
6655 (elem_size = element type size; an element is the scalar element
6656 whose type is the inner type of the vectype)
6660 prolog_niters = min ( LOOP_NITERS ,
6661 (VF/group_size - addr_mis/elem_size)&(VF/group_size-1) )
6662 where group_size is the size of the interleaved group.
6664 The above formulas assume that VF == number of elements in the vector. This
6665 may not hold when there are multiple-types in the loop.
6666 In this case, for some data-references in the loop the VF does not represent
6667 the number of elements that fit in the vector. Therefore, instead of VF we
6668 use TYPE_VECTOR_SUBPARTS. */
6671 vect_gen_niters_for_prolog_loop (loop_vec_info loop_vinfo, tree loop_niters)
6673 struct data_reference *dr = LOOP_VINFO_UNALIGNED_DR (loop_vinfo);
6674 struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
6676 tree iters, iters_name;
6679 tree dr_stmt = DR_STMT (dr);
6680 stmt_vec_info stmt_info = vinfo_for_stmt (dr_stmt);
6681 tree vectype = STMT_VINFO_VECTYPE (stmt_info);
6682 int vectype_align = TYPE_ALIGN (vectype) / BITS_PER_UNIT;
6683 tree niters_type = TREE_TYPE (loop_niters);
6685 int element_size = GET_MODE_SIZE (TYPE_MODE (TREE_TYPE (DR_REF (dr))));
6686 int nelements = TYPE_VECTOR_SUBPARTS (vectype);
6688 if (STMT_VINFO_STRIDED_ACCESS (stmt_info))
6690 /* For interleaved access element size must be multiplied by the size of
6691 the interleaved group. */
6692 group_size = DR_GROUP_SIZE (vinfo_for_stmt (
6693 DR_GROUP_FIRST_DR (stmt_info)));
6694 element_size *= group_size;
6697 pe = loop_preheader_edge (loop);
6699 if (LOOP_PEELING_FOR_ALIGNMENT (loop_vinfo) > 0)
6701 int byte_misalign = LOOP_PEELING_FOR_ALIGNMENT (loop_vinfo);
6702 int elem_misalign = byte_misalign / element_size;
6704 if (vect_print_dump_info (REPORT_DETAILS))
6705 fprintf (vect_dump, "known alignment = %d.", byte_misalign);
6706 iters = build_int_cst (niters_type,
6707 (nelements - elem_misalign)&(nelements/group_size-1));
6711 tree new_stmts = NULL_TREE;
6712 tree start_addr = vect_create_addr_base_for_vector_ref (dr_stmt,
6713 &new_stmts, NULL_TREE, loop);
6714 tree ptr_type = TREE_TYPE (start_addr);
6715 tree size = TYPE_SIZE (ptr_type);
6716 tree type = lang_hooks.types.type_for_size (tree_low_cst (size, 1), 1);
6717 tree vectype_size_minus_1 = build_int_cst (type, vectype_align - 1);
6718 tree elem_size_log =
6719 build_int_cst (type, exact_log2 (vectype_align/nelements));
6720 tree nelements_minus_1 = build_int_cst (type, nelements - 1);
6721 tree nelements_tree = build_int_cst (type, nelements);
6725 new_bb = bsi_insert_on_edge_immediate (pe, new_stmts);
6726 gcc_assert (!new_bb);
6728 /* Create: byte_misalign = addr & (vectype_size - 1) */
6730 fold_build2 (BIT_AND_EXPR, type, fold_convert (type, start_addr), vectype_size_minus_1);
6732 /* Create: elem_misalign = byte_misalign / element_size */
6734 fold_build2 (RSHIFT_EXPR, type, byte_misalign, elem_size_log);
6736 /* Create: (niters_type) (nelements - elem_misalign)&(nelements - 1) */
6737 iters = fold_build2 (MINUS_EXPR, type, nelements_tree, elem_misalign);
6738 iters = fold_build2 (BIT_AND_EXPR, type, iters, nelements_minus_1);
6739 iters = fold_convert (niters_type, iters);
6742 /* Create: prolog_loop_niters = min (iters, loop_niters) */
6743 /* If the loop bound is known at compile time we already verified that it is
6744 greater than vf; since the misalignment ('iters') is at most vf, there's
6745 no need to generate the MIN_EXPR in this case. */
6746 if (TREE_CODE (loop_niters) != INTEGER_CST)
6747 iters = fold_build2 (MIN_EXPR, niters_type, iters, loop_niters);
6749 if (vect_print_dump_info (REPORT_DETAILS))
6751 fprintf (vect_dump, "niters for prolog loop: ");
6752 print_generic_expr (vect_dump, iters, TDF_SLIM);
6755 var = create_tmp_var (niters_type, "prolog_loop_niters");
6756 add_referenced_var (var);
6757 iters_name = force_gimple_operand (iters, &stmt, false, var);
6759 /* Insert stmt on loop preheader edge. */
6762 basic_block new_bb = bsi_insert_on_edge_immediate (pe, stmt);
6763 gcc_assert (!new_bb);
6770 /* Function vect_update_init_of_dr
6772 NITERS iterations were peeled from LOOP. DR represents a data reference
6773 in LOOP. This function updates the information recorded in DR to
6774 account for the fact that the first NITERS iterations had already been
6775 executed. Specifically, it updates the OFFSET field of DR. */
6778 vect_update_init_of_dr (struct data_reference *dr, tree niters)
6780 tree offset = DR_OFFSET (dr);
6782 niters = fold_build2 (MULT_EXPR, TREE_TYPE (niters), niters, DR_STEP (dr));
6783 offset = fold_build2 (PLUS_EXPR, TREE_TYPE (offset), offset, niters);
6784 DR_OFFSET (dr) = offset;
6788 /* Function vect_update_inits_of_drs
6790 NITERS iterations were peeled from the loop represented by LOOP_VINFO.
6791 This function updates the information recorded for the data references in
6792 the loop to account for the fact that the first NITERS iterations had
6793 already been executed. Specifically, it updates the initial_condition of
6794 the access_function of all the data_references in the loop. */
6797 vect_update_inits_of_drs (loop_vec_info loop_vinfo, tree niters)
6800 VEC (data_reference_p, heap) *datarefs = LOOP_VINFO_DATAREFS (loop_vinfo);
6801 struct data_reference *dr;
6803 if (vect_print_dump_info (REPORT_DETAILS))
6804 fprintf (vect_dump, "=== vect_update_inits_of_dr ===");
6806 for (i = 0; VEC_iterate (data_reference_p, datarefs, i, dr); i++)
6807 vect_update_init_of_dr (dr, niters);
6811 /* Function vect_do_peeling_for_alignment
6813 Peel the first 'niters' iterations of the loop represented by LOOP_VINFO.
6814 'niters' is set to the misalignment of one of the data references in the
6815 loop, thereby forcing it to refer to an aligned location at the beginning
6816 of the execution of this loop. The data reference for which we are
6817 peeling is recorded in LOOP_VINFO_UNALIGNED_DR. */
6820 vect_do_peeling_for_alignment (loop_vec_info loop_vinfo)
6822 struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
6823 tree niters_of_prolog_loop, ni_name;
6825 struct loop *new_loop;
6826 bool check_profitability = false;
6827 unsigned int th = 0;
6828 int min_profitable_iters;
6830 if (vect_print_dump_info (REPORT_DETAILS))
6831 fprintf (vect_dump, "=== vect_do_peeling_for_alignment ===");
6833 initialize_original_copy_tables ();
6835 ni_name = vect_build_loop_niters (loop_vinfo);
6836 niters_of_prolog_loop = vect_gen_niters_for_prolog_loop (loop_vinfo, ni_name);
6839 /* If cost model check not done during versioning. */
6840 if (!VEC_length (tree, LOOP_VINFO_MAY_MISALIGN_STMTS (loop_vinfo))
6841 && !VEC_length (ddr_p, LOOP_VINFO_MAY_ALIAS_DDRS (loop_vinfo)))
6843 check_profitability = true;
6845 /* Get profitability threshold for vectorized loop. */
6846 min_profitable_iters = LOOP_VINFO_COST_MODEL_MIN_ITERS (loop_vinfo);
6848 th = conservative_cost_threshold (loop_vinfo,
6849 min_profitable_iters);
6852 /* Peel the prolog loop and iterate it niters_of_prolog_loop. */
6854 slpeel_tree_peel_loop_to_edge (loop, loop_preheader_edge (loop),
6855 niters_of_prolog_loop, ni_name, true,
6856 th, check_profitability);
6858 gcc_assert (new_loop);
6859 #ifdef ENABLE_CHECKING
6860 slpeel_verify_cfg_after_peeling (new_loop, loop);
6863 /* Update number of times loop executes. */
6864 n_iters = LOOP_VINFO_NITERS (loop_vinfo);
6865 LOOP_VINFO_NITERS (loop_vinfo) = fold_build2 (MINUS_EXPR,
6866 TREE_TYPE (n_iters), n_iters, niters_of_prolog_loop);
6868 /* Update the init conditions of the access functions of all data refs. */
6869 vect_update_inits_of_drs (loop_vinfo, niters_of_prolog_loop);
6871 /* After peeling we have to reset scalar evolution analyzer. */
6874 free_original_copy_tables ();
6878 /* Function vect_create_cond_for_align_checks.
6880 Create a conditional expression that represents the alignment checks for
6881 all of data references (array element references) whose alignment must be
6885 COND_EXPR - input conditional expression. New conditions will be chained
6886 with logical AND operation.
6887 LOOP_VINFO - two fields of the loop information are used.
6888 LOOP_VINFO_PTR_MASK is the mask used to check the alignment.
6889 LOOP_VINFO_MAY_MISALIGN_STMTS contains the refs to be checked.
6892 COND_EXPR_STMT_LIST - statements needed to construct the conditional
6894 The returned value is the conditional expression to be used in the if
6895 statement that controls which version of the loop gets executed at runtime.
6897 The algorithm makes two assumptions:
6898 1) The number of bytes "n" in a vector is a power of 2.
6899 2) An address "a" is aligned if a%n is zero and that this
6900 test can be done as a&(n-1) == 0. For example, for 16
6901 byte vectors the test is a&0xf == 0. */
6904 vect_create_cond_for_align_checks (loop_vec_info loop_vinfo,
6906 tree *cond_expr_stmt_list)
6908 struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
6909 VEC(tree,heap) *may_misalign_stmts
6910 = LOOP_VINFO_MAY_MISALIGN_STMTS (loop_vinfo);
6912 int mask = LOOP_VINFO_PTR_MASK (loop_vinfo);
6916 tree int_ptrsize_type;
6918 tree or_tmp_name = NULL_TREE;
6919 tree and_tmp, and_tmp_name, and_stmt;
6921 tree part_cond_expr;
6923 /* Check that mask is one less than a power of 2, i.e., mask is
6924 all zeros followed by all ones. */
6925 gcc_assert ((mask != 0) && ((mask & (mask+1)) == 0));
6927 /* CHECKME: what is the best integer or unsigned type to use to hold a
6928 cast from a pointer value? */
6929 psize = TYPE_SIZE (ptr_type_node);
6931 = lang_hooks.types.type_for_size (tree_low_cst (psize, 1), 0);
6933 /* Create expression (mask & (dr_1 || ... || dr_n)) where dr_i is the address
6934 of the first vector of the i'th data reference. */
6936 for (i = 0; VEC_iterate (tree, may_misalign_stmts, i, ref_stmt); i++)
6938 tree new_stmt_list = NULL_TREE;
6940 tree addr_tmp, addr_tmp_name, addr_stmt;
6941 tree or_tmp, new_or_tmp_name, or_stmt;
6943 /* create: addr_tmp = (int)(address_of_first_vector) */
6944 addr_base = vect_create_addr_base_for_vector_ref (ref_stmt,
6945 &new_stmt_list, NULL_TREE, loop);
6947 if (new_stmt_list != NULL_TREE)
6948 append_to_statement_list_force (new_stmt_list, cond_expr_stmt_list);
6950 sprintf (tmp_name, "%s%d", "addr2int", i);
6951 addr_tmp = create_tmp_var (int_ptrsize_type, tmp_name);
6952 add_referenced_var (addr_tmp);
6953 addr_tmp_name = make_ssa_name (addr_tmp, NULL_TREE);
6954 addr_stmt = fold_convert (int_ptrsize_type, addr_base);
6955 addr_stmt = build_gimple_modify_stmt (addr_tmp_name, addr_stmt);
6956 SSA_NAME_DEF_STMT (addr_tmp_name) = addr_stmt;
6957 append_to_statement_list_force (addr_stmt, cond_expr_stmt_list);
6959 /* The addresses are OR together. */
6961 if (or_tmp_name != NULL_TREE)
6963 /* create: or_tmp = or_tmp | addr_tmp */
6964 sprintf (tmp_name, "%s%d", "orptrs", i);
6965 or_tmp = create_tmp_var (int_ptrsize_type, tmp_name);
6966 add_referenced_var (or_tmp);
6967 new_or_tmp_name = make_ssa_name (or_tmp, NULL_TREE);
6968 tmp = build2 (BIT_IOR_EXPR, int_ptrsize_type,
6969 or_tmp_name, addr_tmp_name);
6970 or_stmt = build_gimple_modify_stmt (new_or_tmp_name, tmp);
6971 SSA_NAME_DEF_STMT (new_or_tmp_name) = or_stmt;
6972 append_to_statement_list_force (or_stmt, cond_expr_stmt_list);
6973 or_tmp_name = new_or_tmp_name;
6976 or_tmp_name = addr_tmp_name;
6980 mask_cst = build_int_cst (int_ptrsize_type, mask);
6982 /* create: and_tmp = or_tmp & mask */
6983 and_tmp = create_tmp_var (int_ptrsize_type, "andmask" );
6984 add_referenced_var (and_tmp);
6985 and_tmp_name = make_ssa_name (and_tmp, NULL_TREE);
6987 tmp = build2 (BIT_AND_EXPR, int_ptrsize_type, or_tmp_name, mask_cst);
6988 and_stmt = build_gimple_modify_stmt (and_tmp_name, tmp);
6989 SSA_NAME_DEF_STMT (and_tmp_name) = and_stmt;
6990 append_to_statement_list_force (and_stmt, cond_expr_stmt_list);
6992 /* Make and_tmp the left operand of the conditional test against zero.
6993 if and_tmp has a nonzero bit then some address is unaligned. */
6994 ptrsize_zero = build_int_cst (int_ptrsize_type, 0);
6995 part_cond_expr = fold_build2 (EQ_EXPR, boolean_type_node,
6996 and_tmp_name, ptrsize_zero);
6998 *cond_expr = fold_build2 (TRUTH_AND_EXPR, boolean_type_node,
6999 *cond_expr, part_cond_expr);
7001 *cond_expr = part_cond_expr;
7004 /* Function vect_vfa_segment_size.
7006 Create an expression that computes the size of segment
7007 that will be accessed for a data reference. The functions takes into
7008 account that realignment loads may access one more vector.
7011 DR: The data reference.
7012 VECT_FACTOR: vectorization factor.
7014 Return an expression whose value is the size of segment which will be
7018 vect_vfa_segment_size (struct data_reference *dr, tree vect_factor)
7020 tree segment_length = fold_build2 (MULT_EXPR, integer_type_node,
7021 DR_STEP (dr), vect_factor);
7023 if (vect_supportable_dr_alignment (dr) == dr_explicit_realign_optimized)
7025 tree vector_size = TYPE_SIZE_UNIT
7026 (STMT_VINFO_VECTYPE (vinfo_for_stmt (DR_STMT (dr))));
7028 segment_length = fold_build2 (PLUS_EXPR, integer_type_node,
7029 segment_length, vector_size);
7031 return fold_convert (sizetype, segment_length);
7034 /* Function vect_create_cond_for_alias_checks.
7036 Create a conditional expression that represents the run-time checks for
7037 overlapping of address ranges represented by a list of data references
7038 relations passed as input.
7041 COND_EXPR - input conditional expression. New conditions will be chained
7042 with logical AND operation.
7043 LOOP_VINFO - field LOOP_VINFO_MAY_ALIAS_STMTS contains the list of ddrs
7047 COND_EXPR - conditional expression.
7048 COND_EXPR_STMT_LIST - statements needed to construct the conditional
7052 The returned value is the conditional expression to be used in the if
7053 statement that controls which version of the loop gets executed at runtime.
7057 vect_create_cond_for_alias_checks (loop_vec_info loop_vinfo,
7059 tree * cond_expr_stmt_list)
7061 struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
7062 VEC (ddr_p, heap) * may_alias_ddrs =
7063 LOOP_VINFO_MAY_ALIAS_DDRS (loop_vinfo);
7065 build_int_cst (integer_type_node, LOOP_VINFO_VECT_FACTOR (loop_vinfo));
7069 tree part_cond_expr;
7071 /* Create expression
7072 ((store_ptr_0 + store_segment_length_0) < load_ptr_0)
7073 || (load_ptr_0 + load_segment_length_0) < store_ptr_0))
7077 ((store_ptr_n + store_segment_length_n) < load_ptr_n)
7078 || (load_ptr_n + load_segment_length_n) < store_ptr_n)) */
7080 if (VEC_empty (ddr_p, may_alias_ddrs))
7083 for (i = 0; VEC_iterate (ddr_p, may_alias_ddrs, i, ddr); i++)
7085 struct data_reference *dr_a, *dr_b;
7086 tree dr_group_first_a, dr_group_first_b;
7087 tree addr_base_a, addr_base_b;
7088 tree segment_length_a, segment_length_b;
7089 tree stmt_a, stmt_b;
7092 stmt_a = DR_STMT (DDR_A (ddr));
7093 dr_group_first_a = DR_GROUP_FIRST_DR (vinfo_for_stmt (stmt_a));
7094 if (dr_group_first_a)
7096 stmt_a = dr_group_first_a;
7097 dr_a = STMT_VINFO_DATA_REF (vinfo_for_stmt (stmt_a));
7101 stmt_b = DR_STMT (DDR_B (ddr));
7102 dr_group_first_b = DR_GROUP_FIRST_DR (vinfo_for_stmt (stmt_b));
7103 if (dr_group_first_b)
7105 stmt_b = dr_group_first_b;
7106 dr_b = STMT_VINFO_DATA_REF (vinfo_for_stmt (stmt_b));
7110 vect_create_addr_base_for_vector_ref (stmt_a, cond_expr_stmt_list,
7113 vect_create_addr_base_for_vector_ref (stmt_b, cond_expr_stmt_list,
7116 segment_length_a = vect_vfa_segment_size (dr_a, vect_factor);
7117 segment_length_b = vect_vfa_segment_size (dr_b, vect_factor);
7119 if (vect_print_dump_info (REPORT_DR_DETAILS))
7122 "create runtime check for data references ");
7123 print_generic_expr (vect_dump, DR_REF (dr_a), TDF_SLIM);
7124 fprintf (vect_dump, " and ");
7125 print_generic_expr (vect_dump, DR_REF (dr_b), TDF_SLIM);
7130 fold_build2 (TRUTH_OR_EXPR, boolean_type_node,
7131 fold_build2 (LT_EXPR, boolean_type_node,
7132 fold_build2 (POINTER_PLUS_EXPR, TREE_TYPE (addr_base_a),
7136 fold_build2 (LT_EXPR, boolean_type_node,
7137 fold_build2 (POINTER_PLUS_EXPR, TREE_TYPE (addr_base_b),
7143 *cond_expr = fold_build2 (TRUTH_AND_EXPR, boolean_type_node,
7144 *cond_expr, part_cond_expr);
7146 *cond_expr = part_cond_expr;
7148 if (vect_print_dump_info (REPORT_VECTORIZED_LOOPS))
7149 fprintf (vect_dump, "created %u versioning for alias checks.\n",
7150 VEC_length (ddr_p, may_alias_ddrs));
7154 /* Function vect_loop_versioning.
7156 If the loop has data references that may or may not be aligned or/and
7157 has data reference relations whose independence was not proven then
7158 two versions of the loop need to be generated, one which is vectorized
7159 and one which isn't. A test is then generated to control which of the
7160 loops is executed. The test checks for the alignment of all of the
7161 data references that may or may not be aligned. An additional
7162 sequence of runtime tests is generated for each pairs of DDRs whose
7163 independence was not proven. The vectorized version of loop is
7164 executed only if both alias and alignment tests are passed.
7166 The test generated to check which version of loop is executed
7167 is modified to also check for profitability as indicated by the
7168 cost model initially. */
7171 vect_loop_versioning (loop_vec_info loop_vinfo)
7173 struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
7175 tree cond_expr = NULL_TREE;
7176 tree cond_expr_stmt_list = NULL_TREE;
7177 basic_block condition_bb;
7178 block_stmt_iterator cond_exp_bsi;
7179 basic_block merge_bb;
7180 basic_block new_exit_bb;
7182 tree orig_phi, new_phi, arg;
7183 unsigned prob = 4 * REG_BR_PROB_BASE / 5;
7184 tree gimplify_stmt_list;
7185 tree scalar_loop_iters = LOOP_VINFO_NITERS (loop_vinfo);
7186 int min_profitable_iters = 0;
7189 /* Get profitability threshold for vectorized loop. */
7190 min_profitable_iters = LOOP_VINFO_COST_MODEL_MIN_ITERS (loop_vinfo);
7192 th = conservative_cost_threshold (loop_vinfo,
7193 min_profitable_iters);
7196 build2 (GT_EXPR, boolean_type_node, scalar_loop_iters,
7197 build_int_cst (TREE_TYPE (scalar_loop_iters), th));
7199 cond_expr = force_gimple_operand (cond_expr, &cond_expr_stmt_list,
7202 if (VEC_length (tree, LOOP_VINFO_MAY_MISALIGN_STMTS (loop_vinfo)))
7203 vect_create_cond_for_align_checks (loop_vinfo, &cond_expr,
7204 &cond_expr_stmt_list);
7206 if (VEC_length (ddr_p, LOOP_VINFO_MAY_ALIAS_DDRS (loop_vinfo)))
7207 vect_create_cond_for_alias_checks (loop_vinfo, &cond_expr,
7208 &cond_expr_stmt_list);
7211 fold_build2 (NE_EXPR, boolean_type_node, cond_expr, integer_zero_node);
7213 force_gimple_operand (cond_expr, &gimplify_stmt_list, true,
7215 append_to_statement_list (gimplify_stmt_list, &cond_expr_stmt_list);
7217 initialize_original_copy_tables ();
7218 nloop = loop_version (loop, cond_expr, &condition_bb,
7219 prob, prob, REG_BR_PROB_BASE - prob, true);
7220 free_original_copy_tables();
7222 /* Loop versioning violates an assumption we try to maintain during
7223 vectorization - that the loop exit block has a single predecessor.
7224 After versioning, the exit block of both loop versions is the same
7225 basic block (i.e. it has two predecessors). Just in order to simplify
7226 following transformations in the vectorizer, we fix this situation
7227 here by adding a new (empty) block on the exit-edge of the loop,
7228 with the proper loop-exit phis to maintain loop-closed-form. */
7230 merge_bb = single_exit (loop)->dest;
7231 gcc_assert (EDGE_COUNT (merge_bb->preds) == 2);
7232 new_exit_bb = split_edge (single_exit (loop));
7233 new_exit_e = single_exit (loop);
7234 e = EDGE_SUCC (new_exit_bb, 0);
7236 for (orig_phi = phi_nodes (merge_bb); orig_phi;
7237 orig_phi = PHI_CHAIN (orig_phi))
7239 new_phi = create_phi_node (SSA_NAME_VAR (PHI_RESULT (orig_phi)),
7241 arg = PHI_ARG_DEF_FROM_EDGE (orig_phi, e);
7242 add_phi_arg (new_phi, arg, new_exit_e);
7243 SET_PHI_ARG_DEF (orig_phi, e->dest_idx, PHI_RESULT (new_phi));
7246 /* End loop-exit-fixes after versioning. */
7248 update_ssa (TODO_update_ssa);
7249 if (cond_expr_stmt_list)
7251 cond_exp_bsi = bsi_last (condition_bb);
7252 bsi_insert_before (&cond_exp_bsi, cond_expr_stmt_list, BSI_SAME_STMT);
7256 /* Remove a group of stores (for SLP or interleaving), free their
7260 vect_remove_stores (tree first_stmt)
7263 tree next = first_stmt;
7265 stmt_vec_info next_stmt_info;
7266 block_stmt_iterator next_si;
7270 /* Free the attached stmt_vec_info and remove the stmt. */
7271 next_si = bsi_for_stmt (next);
7272 bsi_remove (&next_si, true);
7273 next_stmt_info = vinfo_for_stmt (next);
7274 ann = stmt_ann (next);
7275 tmp = DR_GROUP_NEXT_DR (next_stmt_info);
7276 free (next_stmt_info);
7277 set_stmt_info (ann, NULL);
7283 /* Vectorize SLP instance tree in postorder. */
7286 vect_schedule_slp_instance (slp_tree node, unsigned int vec_stmts_size)
7289 bool strided_store, is_store;
7290 block_stmt_iterator si;
7291 stmt_vec_info stmt_info;
7296 vect_schedule_slp_instance (SLP_TREE_LEFT (node), vec_stmts_size);
7297 vect_schedule_slp_instance (SLP_TREE_RIGHT (node), vec_stmts_size);
7299 stmt = VEC_index(tree, SLP_TREE_SCALAR_STMTS (node), 0);
7300 stmt_info = vinfo_for_stmt (stmt);
7301 SLP_TREE_VEC_STMTS (node) = VEC_alloc (tree, heap, vec_stmts_size);
7302 SLP_TREE_NUMBER_OF_VEC_STMTS (node) = vec_stmts_size;
7304 if (vect_print_dump_info (REPORT_DETAILS))
7306 fprintf (vect_dump, "------>vectorizing SLP node starting from: ");
7307 print_generic_expr (vect_dump, stmt, TDF_SLIM);
7310 si = bsi_for_stmt (stmt);
7311 is_store = vect_transform_stmt (stmt, &si, &strided_store, node);
7314 if (DR_GROUP_FIRST_DR (stmt_info))
7315 /* If IS_STORE is TRUE, the vectorization of the
7316 interleaving chain was completed - free all the stores in
7318 vect_remove_stores (DR_GROUP_FIRST_DR (stmt_info));
7320 /* FORNOW: SLP originates only from strided stores. */
7326 /* FORNOW: SLP originates only from strided stores. */
7332 vect_schedule_slp (loop_vec_info loop_vinfo, unsigned int nunits)
7334 VEC (slp_instance, heap) *slp_instances =
7335 LOOP_VINFO_SLP_INSTANCES (loop_vinfo);
7336 slp_instance instance;
7337 unsigned int vec_stmts_size;
7338 unsigned int group_size, i;
7339 unsigned int vectorization_factor = LOOP_VINFO_VECT_FACTOR (loop_vinfo);
7340 bool is_store = false;
7342 for (i = 0; VEC_iterate (slp_instance, slp_instances, i, instance); i++)
7344 group_size = SLP_INSTANCE_GROUP_SIZE (instance);
7345 /* For each SLP instance calculate number of vector stmts to be created
7346 for the scalar stmts in each node of the SLP tree. Number of vector
7347 elements in one vector iteration is the number of scalar elements in
7348 one scalar iteration (GROUP_SIZE) multiplied by VF divided by vector
7350 vec_stmts_size = vectorization_factor * group_size / nunits;
7352 /* Schedule the tree of INSTANCE. */
7353 is_store = vect_schedule_slp_instance (SLP_INSTANCE_TREE (instance),
7356 if (vect_print_dump_info (REPORT_VECTORIZED_LOOPS)
7357 || vect_print_dump_info (REPORT_UNVECTORIZED_LOOPS))
7358 fprintf (vect_dump, "vectorizing stmts using SLP.");
7364 /* Function vect_transform_loop.
7366 The analysis phase has determined that the loop is vectorizable.
7367 Vectorize the loop - created vectorized stmts to replace the scalar
7368 stmts in the loop, and update the loop exit condition. */
7371 vect_transform_loop (loop_vec_info loop_vinfo)
7373 struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
7374 basic_block *bbs = LOOP_VINFO_BBS (loop_vinfo);
7375 int nbbs = loop->num_nodes;
7376 block_stmt_iterator si, next_si;
7379 int vectorization_factor = LOOP_VINFO_VECT_FACTOR (loop_vinfo);
7381 bool slp_scheduled = false;
7382 unsigned int nunits;
7384 if (vect_print_dump_info (REPORT_DETAILS))
7385 fprintf (vect_dump, "=== vec_transform_loop ===");
7387 if (VEC_length (tree, LOOP_VINFO_MAY_MISALIGN_STMTS (loop_vinfo))
7388 || VEC_length (ddr_p, LOOP_VINFO_MAY_ALIAS_DDRS (loop_vinfo)))
7389 vect_loop_versioning (loop_vinfo);
7391 /* CHECKME: we wouldn't need this if we called update_ssa once
7393 bitmap_zero (vect_memsyms_to_rename);
7395 /* Peel the loop if there are data refs with unknown alignment.
7396 Only one data ref with unknown store is allowed. */
7398 if (LOOP_PEELING_FOR_ALIGNMENT (loop_vinfo))
7399 vect_do_peeling_for_alignment (loop_vinfo);
7401 /* If the loop has a symbolic number of iterations 'n' (i.e. it's not a
7402 compile time constant), or it is a constant that doesn't divide by the
7403 vectorization factor, then an epilog loop needs to be created.
7404 We therefore duplicate the loop: the original loop will be vectorized,
7405 and will compute the first (n/VF) iterations. The second copy of the loop
7406 will remain scalar and will compute the remaining (n%VF) iterations.
7407 (VF is the vectorization factor). */
7409 if (!LOOP_VINFO_NITERS_KNOWN_P (loop_vinfo)
7410 || (LOOP_VINFO_NITERS_KNOWN_P (loop_vinfo)
7411 && LOOP_VINFO_INT_NITERS (loop_vinfo) % vectorization_factor != 0))
7412 vect_do_peeling_for_loop_bound (loop_vinfo, &ratio);
7414 ratio = build_int_cst (TREE_TYPE (LOOP_VINFO_NITERS (loop_vinfo)),
7415 LOOP_VINFO_INT_NITERS (loop_vinfo) / vectorization_factor);
7417 /* 1) Make sure the loop header has exactly two entries
7418 2) Make sure we have a preheader basic block. */
7420 gcc_assert (EDGE_COUNT (loop->header->preds) == 2);
7422 split_edge (loop_preheader_edge (loop));
7424 /* FORNOW: the vectorizer supports only loops which body consist
7425 of one basic block (header + empty latch). When the vectorizer will
7426 support more involved loop forms, the order by which the BBs are
7427 traversed need to be reconsidered. */
7429 for (i = 0; i < nbbs; i++)
7431 basic_block bb = bbs[i];
7432 stmt_vec_info stmt_info;
7435 for (phi = phi_nodes (bb); phi; phi = PHI_CHAIN (phi))
7437 if (vect_print_dump_info (REPORT_DETAILS))
7439 fprintf (vect_dump, "------>vectorizing phi: ");
7440 print_generic_expr (vect_dump, phi, TDF_SLIM);
7442 stmt_info = vinfo_for_stmt (phi);
7446 if (!STMT_VINFO_RELEVANT_P (stmt_info)
7447 && !STMT_VINFO_LIVE_P (stmt_info))
7450 if ((TYPE_VECTOR_SUBPARTS (STMT_VINFO_VECTYPE (stmt_info))
7451 != (unsigned HOST_WIDE_INT) vectorization_factor)
7452 && vect_print_dump_info (REPORT_DETAILS))
7453 fprintf (vect_dump, "multiple-types.");
7455 if (STMT_VINFO_DEF_TYPE (stmt_info) == vect_induction_def)
7457 if (vect_print_dump_info (REPORT_DETAILS))
7458 fprintf (vect_dump, "transform phi.");
7459 vect_transform_stmt (phi, NULL, NULL, NULL);
7463 for (si = bsi_start (bb); !bsi_end_p (si);)
7465 tree stmt = bsi_stmt (si);
7468 if (vect_print_dump_info (REPORT_DETAILS))
7470 fprintf (vect_dump, "------>vectorizing statement: ");
7471 print_generic_expr (vect_dump, stmt, TDF_SLIM);
7474 stmt_info = vinfo_for_stmt (stmt);
7476 /* vector stmts created in the outer-loop during vectorization of
7477 stmts in an inner-loop may not have a stmt_info, and do not
7478 need to be vectorized. */
7485 if (!STMT_VINFO_RELEVANT_P (stmt_info)
7486 && !STMT_VINFO_LIVE_P (stmt_info))
7492 gcc_assert (STMT_VINFO_VECTYPE (stmt_info));
7494 (unsigned int) TYPE_VECTOR_SUBPARTS (STMT_VINFO_VECTYPE (stmt_info));
7495 if (!STMT_SLP_TYPE (stmt_info)
7496 && nunits != (unsigned int) vectorization_factor
7497 && vect_print_dump_info (REPORT_DETAILS))
7498 /* For SLP VF is set according to unrolling factor, and not to
7499 vector size, hence for SLP this print is not valid. */
7500 fprintf (vect_dump, "multiple-types.");
7502 /* SLP. Schedule all the SLP instances when the first SLP stmt is
7504 if (STMT_SLP_TYPE (stmt_info))
7508 slp_scheduled = true;
7510 if (vect_print_dump_info (REPORT_DETAILS))
7511 fprintf (vect_dump, "=== scheduling SLP instances ===");
7513 is_store = vect_schedule_slp (loop_vinfo, nunits);
7515 /* IS_STORE is true if STMT is a store. Stores cannot be of
7516 hybrid SLP type. They are removed in
7517 vect_schedule_slp_instance and their vinfo is destroyed. */
7525 /* Hybrid SLP stmts must be vectorized in addition to SLP. */
7526 if (PURE_SLP_STMT (stmt_info))
7533 /* -------- vectorize statement ------------ */
7534 if (vect_print_dump_info (REPORT_DETAILS))
7535 fprintf (vect_dump, "transform statement.");
7537 strided_store = false;
7538 is_store = vect_transform_stmt (stmt, &si, &strided_store, NULL);
7542 if (STMT_VINFO_STRIDED_ACCESS (stmt_info))
7544 /* Interleaving. If IS_STORE is TRUE, the vectorization of the
7545 interleaving chain was completed - free all the stores in
7547 tree next = DR_GROUP_FIRST_DR (stmt_info);
7549 stmt_vec_info next_stmt_info;
7553 next_si = bsi_for_stmt (next);
7554 next_stmt_info = vinfo_for_stmt (next);
7555 /* Free the attached stmt_vec_info and remove the stmt. */
7556 ann = stmt_ann (next);
7557 tmp = DR_GROUP_NEXT_DR (next_stmt_info);
7558 free (next_stmt_info);
7559 set_stmt_info (ann, NULL);
7560 bsi_remove (&next_si, true);
7563 bsi_remove (&si, true);
7568 /* Free the attached stmt_vec_info and remove the stmt. */
7569 ann = stmt_ann (stmt);
7571 set_stmt_info (ann, NULL);
7572 bsi_remove (&si, true);
7580 slpeel_make_loop_iterate_ntimes (loop, ratio);
7582 mark_set_for_renaming (vect_memsyms_to_rename);
7584 /* The memory tags and pointers in vectorized statements need to
7585 have their SSA forms updated. FIXME, why can't this be delayed
7586 until all the loops have been transformed? */
7587 update_ssa (TODO_update_ssa);
7589 if (vect_print_dump_info (REPORT_VECTORIZED_LOOPS))
7590 fprintf (vect_dump, "LOOP VECTORIZED.");
7591 if (loop->inner && vect_print_dump_info (REPORT_VECTORIZED_LOOPS))
7592 fprintf (vect_dump, "OUTER LOOP VECTORIZED.");