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 BIT_FIELD_REF_UNSIGNED (rhs) = TYPE_UNSIGNED (scalar_type);
2521 epilog_stmt = build_gimple_modify_stmt (new_scalar_dest, rhs);
2522 new_temp = make_ssa_name (new_scalar_dest, epilog_stmt);
2523 GIMPLE_STMT_OPERAND (epilog_stmt, 0) = new_temp;
2524 bsi_insert_before (&exit_bsi, epilog_stmt, BSI_SAME_STMT);
2526 for (bit_offset = element_bitsize;
2527 bit_offset < vec_size_in_bits;
2528 bit_offset += element_bitsize)
2531 tree bitpos = bitsize_int (bit_offset);
2532 tree rhs = build3 (BIT_FIELD_REF, scalar_type, vec_temp, bitsize,
2535 BIT_FIELD_REF_UNSIGNED (rhs) = TYPE_UNSIGNED (scalar_type);
2536 epilog_stmt = build_gimple_modify_stmt (new_scalar_dest, rhs);
2537 new_name = make_ssa_name (new_scalar_dest, epilog_stmt);
2538 GIMPLE_STMT_OPERAND (epilog_stmt, 0) = new_name;
2539 bsi_insert_before (&exit_bsi, epilog_stmt, BSI_SAME_STMT);
2541 tmp = build2 (code, scalar_type, new_name, new_temp);
2542 epilog_stmt = build_gimple_modify_stmt (new_scalar_dest, tmp);
2543 new_temp = make_ssa_name (new_scalar_dest, epilog_stmt);
2544 GIMPLE_STMT_OPERAND (epilog_stmt, 0) = new_temp;
2545 bsi_insert_before (&exit_bsi, epilog_stmt, BSI_SAME_STMT);
2548 extract_scalar_result = false;
2552 /* 2.4 Extract the final scalar result. Create:
2553 s_out3 = extract_field <v_out2, bitpos> */
2555 if (extract_scalar_result)
2559 gcc_assert (!nested_in_vect_loop);
2560 if (vect_print_dump_info (REPORT_DETAILS))
2561 fprintf (vect_dump, "extract scalar result");
2563 if (BYTES_BIG_ENDIAN)
2564 bitpos = size_binop (MULT_EXPR,
2565 bitsize_int (TYPE_VECTOR_SUBPARTS (vectype) - 1),
2566 TYPE_SIZE (scalar_type));
2568 bitpos = bitsize_zero_node;
2570 rhs = build3 (BIT_FIELD_REF, scalar_type, new_temp, bitsize, bitpos);
2571 BIT_FIELD_REF_UNSIGNED (rhs) = TYPE_UNSIGNED (scalar_type);
2572 epilog_stmt = build_gimple_modify_stmt (new_scalar_dest, rhs);
2573 new_temp = make_ssa_name (new_scalar_dest, epilog_stmt);
2574 GIMPLE_STMT_OPERAND (epilog_stmt, 0) = new_temp;
2575 bsi_insert_before (&exit_bsi, epilog_stmt, BSI_SAME_STMT);
2578 vect_finalize_reduction:
2580 /* 2.5 Adjust the final result by the initial value of the reduction
2581 variable. (When such adjustment is not needed, then
2582 'adjustment_def' is zero). For example, if code is PLUS we create:
2583 new_temp = loop_exit_def + adjustment_def */
2587 if (nested_in_vect_loop)
2589 gcc_assert (TREE_CODE (TREE_TYPE (adjustment_def)) == VECTOR_TYPE);
2590 expr = build2 (code, vectype, PHI_RESULT (new_phi), adjustment_def);
2591 new_dest = vect_create_destination_var (scalar_dest, vectype);
2595 gcc_assert (TREE_CODE (TREE_TYPE (adjustment_def)) != VECTOR_TYPE);
2596 expr = build2 (code, scalar_type, new_temp, adjustment_def);
2597 new_dest = vect_create_destination_var (scalar_dest, scalar_type);
2599 epilog_stmt = build_gimple_modify_stmt (new_dest, expr);
2600 new_temp = make_ssa_name (new_dest, epilog_stmt);
2601 GIMPLE_STMT_OPERAND (epilog_stmt, 0) = new_temp;
2602 bsi_insert_before (&exit_bsi, epilog_stmt, BSI_SAME_STMT);
2606 /* 2.6 Handle the loop-exit phi */
2608 /* Replace uses of s_out0 with uses of s_out3:
2609 Find the loop-closed-use at the loop exit of the original scalar result.
2610 (The reduction result is expected to have two immediate uses - one at the
2611 latch block, and one at the loop exit). */
2612 phis = VEC_alloc (tree, heap, 10);
2613 FOR_EACH_IMM_USE_FAST (use_p, imm_iter, scalar_dest)
2615 if (!flow_bb_inside_loop_p (loop, bb_for_stmt (USE_STMT (use_p))))
2617 exit_phi = USE_STMT (use_p);
2618 VEC_quick_push (tree, phis, exit_phi);
2621 /* We expect to have found an exit_phi because of loop-closed-ssa form. */
2622 gcc_assert (!VEC_empty (tree, phis));
2624 for (i = 0; VEC_iterate (tree, phis, i, exit_phi); i++)
2626 if (nested_in_vect_loop)
2628 stmt_vec_info stmt_vinfo = vinfo_for_stmt (exit_phi);
2630 /* FORNOW. Currently not supporting the case that an inner-loop reduction
2631 is not used in the outer-loop (but only outside the outer-loop). */
2632 gcc_assert (STMT_VINFO_RELEVANT_P (stmt_vinfo)
2633 && !STMT_VINFO_LIVE_P (stmt_vinfo));
2635 epilog_stmt = adjustment_def ? epilog_stmt : new_phi;
2636 STMT_VINFO_VEC_STMT (stmt_vinfo) = epilog_stmt;
2637 set_stmt_info (get_stmt_ann (epilog_stmt),
2638 new_stmt_vec_info (epilog_stmt, loop_vinfo));
2642 /* Replace the uses: */
2643 orig_name = PHI_RESULT (exit_phi);
2644 FOR_EACH_IMM_USE_STMT (use_stmt, imm_iter, orig_name)
2645 FOR_EACH_IMM_USE_ON_STMT (use_p, imm_iter)
2646 SET_USE (use_p, new_temp);
2648 VEC_free (tree, heap, phis);
2652 /* Function vectorizable_reduction.
2654 Check if STMT performs a reduction operation that can be vectorized.
2655 If VEC_STMT is also passed, vectorize the STMT: create a vectorized
2656 stmt to replace it, put it in VEC_STMT, and insert it at BSI.
2657 Return FALSE if not a vectorizable STMT, TRUE otherwise.
2659 This function also handles reduction idioms (patterns) that have been
2660 recognized in advance during vect_pattern_recog. In this case, STMT may be
2662 X = pattern_expr (arg0, arg1, ..., X)
2663 and it's STMT_VINFO_RELATED_STMT points to the last stmt in the original
2664 sequence that had been detected and replaced by the pattern-stmt (STMT).
2666 In some cases of reduction patterns, the type of the reduction variable X is
2667 different than the type of the other arguments of STMT.
2668 In such cases, the vectype that is used when transforming STMT into a vector
2669 stmt is different than the vectype that is used to determine the
2670 vectorization factor, because it consists of a different number of elements
2671 than the actual number of elements that are being operated upon in parallel.
2673 For example, consider an accumulation of shorts into an int accumulator.
2674 On some targets it's possible to vectorize this pattern operating on 8
2675 shorts at a time (hence, the vectype for purposes of determining the
2676 vectorization factor should be V8HI); on the other hand, the vectype that
2677 is used to create the vector form is actually V4SI (the type of the result).
2679 Upon entry to this function, STMT_VINFO_VECTYPE records the vectype that
2680 indicates what is the actual level of parallelism (V8HI in the example), so
2681 that the right vectorization factor would be derived. This vectype
2682 corresponds to the type of arguments to the reduction stmt, and should *NOT*
2683 be used to create the vectorized stmt. The right vectype for the vectorized
2684 stmt is obtained from the type of the result X:
2685 get_vectype_for_scalar_type (TREE_TYPE (X))
2687 This means that, contrary to "regular" reductions (or "regular" stmts in
2688 general), the following equation:
2689 STMT_VINFO_VECTYPE == get_vectype_for_scalar_type (TREE_TYPE (X))
2690 does *NOT* necessarily hold for reduction patterns. */
2693 vectorizable_reduction (tree stmt, block_stmt_iterator *bsi, tree *vec_stmt)
2698 tree loop_vec_def0 = NULL_TREE, loop_vec_def1 = NULL_TREE;
2699 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
2700 tree vectype = STMT_VINFO_VECTYPE (stmt_info);
2701 loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info);
2702 struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
2704 enum tree_code code, orig_code, epilog_reduc_code = 0;
2705 enum machine_mode vec_mode;
2707 optab optab, reduc_optab;
2708 tree new_temp = NULL_TREE;
2710 enum vect_def_type dt;
2715 stmt_vec_info orig_stmt_info;
2716 tree expr = NULL_TREE;
2718 int nunits = TYPE_VECTOR_SUBPARTS (vectype);
2719 int ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits;
2720 stmt_vec_info prev_stmt_info;
2722 tree new_stmt = NULL_TREE;
2725 if (nested_in_vect_loop_p (loop, stmt))
2728 /* FORNOW. This restriction should be relaxed. */
2731 if (vect_print_dump_info (REPORT_DETAILS))
2732 fprintf (vect_dump, "multiple types in nested loop.");
2737 gcc_assert (ncopies >= 1);
2739 /* FORNOW: SLP not supported. */
2740 if (STMT_SLP_TYPE (stmt_info))
2743 /* 1. Is vectorizable reduction? */
2745 /* Not supportable if the reduction variable is used in the loop. */
2746 if (STMT_VINFO_RELEVANT (stmt_info) > vect_used_in_outer)
2749 /* Reductions that are not used even in an enclosing outer-loop,
2750 are expected to be "live" (used out of the loop). */
2751 if (STMT_VINFO_RELEVANT (stmt_info) == vect_unused_in_loop
2752 && !STMT_VINFO_LIVE_P (stmt_info))
2755 /* Make sure it was already recognized as a reduction computation. */
2756 if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_reduction_def)
2759 /* 2. Has this been recognized as a reduction pattern?
2761 Check if STMT represents a pattern that has been recognized
2762 in earlier analysis stages. For stmts that represent a pattern,
2763 the STMT_VINFO_RELATED_STMT field records the last stmt in
2764 the original sequence that constitutes the pattern. */
2766 orig_stmt = STMT_VINFO_RELATED_STMT (stmt_info);
2769 orig_stmt_info = vinfo_for_stmt (orig_stmt);
2770 gcc_assert (STMT_VINFO_RELATED_STMT (orig_stmt_info) == stmt);
2771 gcc_assert (STMT_VINFO_IN_PATTERN_P (orig_stmt_info));
2772 gcc_assert (!STMT_VINFO_IN_PATTERN_P (stmt_info));
2775 /* 3. Check the operands of the operation. The first operands are defined
2776 inside the loop body. The last operand is the reduction variable,
2777 which is defined by the loop-header-phi. */
2779 gcc_assert (TREE_CODE (stmt) == GIMPLE_MODIFY_STMT);
2781 operation = GIMPLE_STMT_OPERAND (stmt, 1);
2782 code = TREE_CODE (operation);
2783 op_type = TREE_OPERAND_LENGTH (operation);
2784 if (op_type != binary_op && op_type != ternary_op)
2786 scalar_dest = GIMPLE_STMT_OPERAND (stmt, 0);
2787 scalar_type = TREE_TYPE (scalar_dest);
2788 if (!POINTER_TYPE_P (scalar_type) && !INTEGRAL_TYPE_P (scalar_type)
2789 && !SCALAR_FLOAT_TYPE_P (scalar_type))
2792 /* All uses but the last are expected to be defined in the loop.
2793 The last use is the reduction variable. */
2794 for (i = 0; i < op_type-1; i++)
2796 op = TREE_OPERAND (operation, i);
2797 is_simple_use = vect_is_simple_use (op, loop_vinfo, &def_stmt, &def, &dt);
2798 gcc_assert (is_simple_use);
2799 if (dt != vect_loop_def
2800 && dt != vect_invariant_def
2801 && dt != vect_constant_def
2802 && dt != vect_induction_def)
2806 op = TREE_OPERAND (operation, i);
2807 is_simple_use = vect_is_simple_use (op, loop_vinfo, &def_stmt, &def, &dt);
2808 gcc_assert (is_simple_use);
2809 gcc_assert (dt == vect_reduction_def);
2810 gcc_assert (TREE_CODE (def_stmt) == PHI_NODE);
2812 gcc_assert (orig_stmt == vect_is_simple_reduction (loop_vinfo, def_stmt));
2814 gcc_assert (stmt == vect_is_simple_reduction (loop_vinfo, def_stmt));
2816 if (STMT_VINFO_LIVE_P (vinfo_for_stmt (def_stmt)))
2819 /* 4. Supportable by target? */
2821 /* 4.1. check support for the operation in the loop */
2822 optab = optab_for_tree_code (code, vectype);
2825 if (vect_print_dump_info (REPORT_DETAILS))
2826 fprintf (vect_dump, "no optab.");
2829 vec_mode = TYPE_MODE (vectype);
2830 if (optab_handler (optab, vec_mode)->insn_code == CODE_FOR_nothing)
2832 if (vect_print_dump_info (REPORT_DETAILS))
2833 fprintf (vect_dump, "op not supported by target.");
2834 if (GET_MODE_SIZE (vec_mode) != UNITS_PER_WORD
2835 || LOOP_VINFO_VECT_FACTOR (loop_vinfo)
2836 < vect_min_worthwhile_factor (code))
2838 if (vect_print_dump_info (REPORT_DETAILS))
2839 fprintf (vect_dump, "proceeding using word mode.");
2842 /* Worthwhile without SIMD support? */
2843 if (!VECTOR_MODE_P (TYPE_MODE (vectype))
2844 && LOOP_VINFO_VECT_FACTOR (loop_vinfo)
2845 < vect_min_worthwhile_factor (code))
2847 if (vect_print_dump_info (REPORT_DETAILS))
2848 fprintf (vect_dump, "not worthwhile without SIMD support.");
2852 /* 4.2. Check support for the epilog operation.
2854 If STMT represents a reduction pattern, then the type of the
2855 reduction variable may be different than the type of the rest
2856 of the arguments. For example, consider the case of accumulation
2857 of shorts into an int accumulator; The original code:
2858 S1: int_a = (int) short_a;
2859 orig_stmt-> S2: int_acc = plus <int_a ,int_acc>;
2862 STMT: int_acc = widen_sum <short_a, int_acc>
2865 1. The tree-code that is used to create the vector operation in the
2866 epilog code (that reduces the partial results) is not the
2867 tree-code of STMT, but is rather the tree-code of the original
2868 stmt from the pattern that STMT is replacing. I.e, in the example
2869 above we want to use 'widen_sum' in the loop, but 'plus' in the
2871 2. The type (mode) we use to check available target support
2872 for the vector operation to be created in the *epilog*, is
2873 determined by the type of the reduction variable (in the example
2874 above we'd check this: plus_optab[vect_int_mode]).
2875 However the type (mode) we use to check available target support
2876 for the vector operation to be created *inside the loop*, is
2877 determined by the type of the other arguments to STMT (in the
2878 example we'd check this: widen_sum_optab[vect_short_mode]).
2880 This is contrary to "regular" reductions, in which the types of all
2881 the arguments are the same as the type of the reduction variable.
2882 For "regular" reductions we can therefore use the same vector type
2883 (and also the same tree-code) when generating the epilog code and
2884 when generating the code inside the loop. */
2888 /* This is a reduction pattern: get the vectype from the type of the
2889 reduction variable, and get the tree-code from orig_stmt. */
2890 orig_code = TREE_CODE (GIMPLE_STMT_OPERAND (orig_stmt, 1));
2891 vectype = get_vectype_for_scalar_type (TREE_TYPE (def));
2894 if (vect_print_dump_info (REPORT_DETAILS))
2896 fprintf (vect_dump, "unsupported data-type ");
2897 print_generic_expr (vect_dump, TREE_TYPE (def), TDF_SLIM);
2902 vec_mode = TYPE_MODE (vectype);
2906 /* Regular reduction: use the same vectype and tree-code as used for
2907 the vector code inside the loop can be used for the epilog code. */
2911 if (!reduction_code_for_scalar_code (orig_code, &epilog_reduc_code))
2913 reduc_optab = optab_for_tree_code (epilog_reduc_code, vectype);
2916 if (vect_print_dump_info (REPORT_DETAILS))
2917 fprintf (vect_dump, "no optab for reduction.");
2918 epilog_reduc_code = NUM_TREE_CODES;
2920 if (optab_handler (reduc_optab, vec_mode)->insn_code == CODE_FOR_nothing)
2922 if (vect_print_dump_info (REPORT_DETAILS))
2923 fprintf (vect_dump, "reduc op not supported by target.");
2924 epilog_reduc_code = NUM_TREE_CODES;
2927 if (!vec_stmt) /* transformation not required. */
2929 STMT_VINFO_TYPE (stmt_info) = reduc_vec_info_type;
2930 if (!vect_model_reduction_cost (stmt_info, epilog_reduc_code, ncopies))
2937 if (vect_print_dump_info (REPORT_DETAILS))
2938 fprintf (vect_dump, "transform reduction.");
2940 /* Create the destination vector */
2941 vec_dest = vect_create_destination_var (scalar_dest, vectype);
2943 /* Create the reduction-phi that defines the reduction-operand. */
2944 new_phi = create_phi_node (vec_dest, loop->header);
2946 /* In case the vectorization factor (VF) is bigger than the number
2947 of elements that we can fit in a vectype (nunits), we have to generate
2948 more than one vector stmt - i.e - we need to "unroll" the
2949 vector stmt by a factor VF/nunits. For more details see documentation
2950 in vectorizable_operation. */
2952 prev_stmt_info = NULL;
2953 for (j = 0; j < ncopies; j++)
2958 op = TREE_OPERAND (operation, 0);
2959 loop_vec_def0 = vect_get_vec_def_for_operand (op, stmt, NULL);
2960 if (op_type == ternary_op)
2962 op = TREE_OPERAND (operation, 1);
2963 loop_vec_def1 = vect_get_vec_def_for_operand (op, stmt, NULL);
2966 /* Get the vector def for the reduction variable from the phi node */
2967 reduc_def = PHI_RESULT (new_phi);
2971 enum vect_def_type dt = vect_unknown_def_type; /* Dummy */
2972 loop_vec_def0 = vect_get_vec_def_for_stmt_copy (dt, loop_vec_def0);
2973 if (op_type == ternary_op)
2974 loop_vec_def1 = vect_get_vec_def_for_stmt_copy (dt, loop_vec_def1);
2976 /* Get the vector def for the reduction variable from the vectorized
2977 reduction operation generated in the previous iteration (j-1) */
2978 reduc_def = GIMPLE_STMT_OPERAND (new_stmt ,0);
2981 /* Arguments are ready. create the new vector stmt. */
2982 if (op_type == binary_op)
2983 expr = build2 (code, vectype, loop_vec_def0, reduc_def);
2985 expr = build3 (code, vectype, loop_vec_def0, loop_vec_def1,
2987 new_stmt = build_gimple_modify_stmt (vec_dest, expr);
2988 new_temp = make_ssa_name (vec_dest, new_stmt);
2989 GIMPLE_STMT_OPERAND (new_stmt, 0) = new_temp;
2990 vect_finish_stmt_generation (stmt, new_stmt, bsi);
2993 STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt;
2995 STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt;
2996 prev_stmt_info = vinfo_for_stmt (new_stmt);
2999 /* Finalize the reduction-phi (set it's arguments) and create the
3000 epilog reduction code. */
3001 vect_create_epilog_for_reduction (new_temp, stmt, epilog_reduc_code, new_phi);
3005 /* Checks if CALL can be vectorized in type VECTYPE. Returns
3006 a function declaration if the target has a vectorized version
3007 of the function, or NULL_TREE if the function cannot be vectorized. */
3010 vectorizable_function (tree call, tree vectype_out, tree vectype_in)
3012 tree fndecl = get_callee_fndecl (call);
3013 enum built_in_function code;
3015 /* We only handle functions that do not read or clobber memory -- i.e.
3016 const or novops ones. */
3017 if (!(call_expr_flags (call) & (ECF_CONST | ECF_NOVOPS)))
3021 || TREE_CODE (fndecl) != FUNCTION_DECL
3022 || !DECL_BUILT_IN (fndecl))
3025 code = DECL_FUNCTION_CODE (fndecl);
3026 return targetm.vectorize.builtin_vectorized_function (code, vectype_out,
3030 /* Function vectorizable_call.
3032 Check if STMT performs a function call that can be vectorized.
3033 If VEC_STMT is also passed, vectorize the STMT: create a vectorized
3034 stmt to replace it, put it in VEC_STMT, and insert it at BSI.
3035 Return FALSE if not a vectorizable STMT, TRUE otherwise. */
3038 vectorizable_call (tree stmt, block_stmt_iterator *bsi, tree *vec_stmt)
3044 tree vec_oprnd0 = NULL_TREE, vec_oprnd1 = NULL_TREE;
3045 stmt_vec_info stmt_info = vinfo_for_stmt (stmt), prev_stmt_info;
3046 tree vectype_out, vectype_in;
3049 loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info);
3050 struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
3051 tree fndecl, rhs, new_temp, def, def_stmt, rhs_type, lhs_type;
3052 enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type};
3054 int ncopies, j, nargs;
3055 call_expr_arg_iterator iter;
3057 enum { NARROW, NONE, WIDEN } modifier;
3059 if (!STMT_VINFO_RELEVANT_P (stmt_info))
3062 if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_loop_def)
3065 /* FORNOW: SLP not supported. */
3066 if (STMT_SLP_TYPE (stmt_info))
3069 /* Is STMT a vectorizable call? */
3070 if (TREE_CODE (stmt) != GIMPLE_MODIFY_STMT)
3073 if (TREE_CODE (GIMPLE_STMT_OPERAND (stmt, 0)) != SSA_NAME)
3076 operation = GIMPLE_STMT_OPERAND (stmt, 1);
3077 if (TREE_CODE (operation) != CALL_EXPR)
3080 /* Process function arguments. */
3081 rhs_type = NULL_TREE;
3083 FOR_EACH_CALL_EXPR_ARG (op, iter, operation)
3085 /* Bail out if the function has more than two arguments, we
3086 do not have interesting builtin functions to vectorize with
3087 more than two arguments. */
3091 /* We can only handle calls with arguments of the same type. */
3093 && rhs_type != TREE_TYPE (op))
3095 if (vect_print_dump_info (REPORT_DETAILS))
3096 fprintf (vect_dump, "argument types differ.");
3099 rhs_type = TREE_TYPE (op);
3101 if (!vect_is_simple_use (op, loop_vinfo, &def_stmt, &def, &dt[nargs]))
3103 if (vect_print_dump_info (REPORT_DETAILS))
3104 fprintf (vect_dump, "use not simple.");
3111 /* No arguments is also not good. */
3115 vectype_in = get_vectype_for_scalar_type (rhs_type);
3118 nunits_in = TYPE_VECTOR_SUBPARTS (vectype_in);
3120 lhs_type = TREE_TYPE (GIMPLE_STMT_OPERAND (stmt, 0));
3121 vectype_out = get_vectype_for_scalar_type (lhs_type);
3124 nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out);
3127 if (nunits_in == nunits_out / 2)
3129 else if (nunits_out == nunits_in)
3131 else if (nunits_out == nunits_in / 2)
3136 /* For now, we only vectorize functions if a target specific builtin
3137 is available. TODO -- in some cases, it might be profitable to
3138 insert the calls for pieces of the vector, in order to be able
3139 to vectorize other operations in the loop. */
3140 fndecl = vectorizable_function (operation, vectype_out, vectype_in);
3141 if (fndecl == NULL_TREE)
3143 if (vect_print_dump_info (REPORT_DETAILS))
3144 fprintf (vect_dump, "function is not vectorizable.");
3149 gcc_assert (ZERO_SSA_OPERANDS (stmt, SSA_OP_ALL_VIRTUALS));
3151 if (modifier == NARROW)
3152 ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_out;
3154 ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in;
3156 /* Sanity check: make sure that at least one copy of the vectorized stmt
3157 needs to be generated. */
3158 gcc_assert (ncopies >= 1);
3160 /* FORNOW. This restriction should be relaxed. */
3161 if (nested_in_vect_loop_p (loop, stmt) && ncopies > 1)
3163 if (vect_print_dump_info (REPORT_DETAILS))
3164 fprintf (vect_dump, "multiple types in nested loop.");
3168 if (!vec_stmt) /* transformation not required. */
3170 STMT_VINFO_TYPE (stmt_info) = call_vec_info_type;
3171 if (vect_print_dump_info (REPORT_DETAILS))
3172 fprintf (vect_dump, "=== vectorizable_call ===");
3173 vect_model_simple_cost (stmt_info, ncopies, dt, NULL);
3179 if (vect_print_dump_info (REPORT_DETAILS))
3180 fprintf (vect_dump, "transform operation.");
3182 /* FORNOW. This restriction should be relaxed. */
3183 if (nested_in_vect_loop_p (loop, stmt) && ncopies > 1)
3185 if (vect_print_dump_info (REPORT_DETAILS))
3186 fprintf (vect_dump, "multiple types in nested loop.");
3191 scalar_dest = GIMPLE_STMT_OPERAND (stmt, 0);
3192 vec_dest = vect_create_destination_var (scalar_dest, vectype_out);
3194 prev_stmt_info = NULL;
3198 for (j = 0; j < ncopies; ++j)
3200 /* Build argument list for the vectorized call. */
3201 /* FIXME: Rewrite this so that it doesn't
3202 construct a temporary list. */
3205 FOR_EACH_CALL_EXPR_ARG (op, iter, operation)
3209 = vect_get_vec_def_for_operand (op, stmt, NULL);
3212 = vect_get_vec_def_for_stmt_copy (dt[nargs], vec_oprnd0);
3214 vargs = tree_cons (NULL_TREE, vec_oprnd0, vargs);
3218 vargs = nreverse (vargs);
3220 rhs = build_function_call_expr (fndecl, vargs);
3221 new_stmt = build_gimple_modify_stmt (vec_dest, rhs);
3222 new_temp = make_ssa_name (vec_dest, new_stmt);
3223 GIMPLE_STMT_OPERAND (new_stmt, 0) = new_temp;
3225 vect_finish_stmt_generation (stmt, new_stmt, bsi);
3228 STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt;
3230 STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt;
3232 prev_stmt_info = vinfo_for_stmt (new_stmt);
3238 for (j = 0; j < ncopies; ++j)
3240 /* Build argument list for the vectorized call. */
3241 /* FIXME: Rewrite this so that it doesn't
3242 construct a temporary list. */
3245 FOR_EACH_CALL_EXPR_ARG (op, iter, operation)
3250 = vect_get_vec_def_for_operand (op, stmt, NULL);
3252 = vect_get_vec_def_for_stmt_copy (dt[nargs], vec_oprnd0);
3257 = vect_get_vec_def_for_stmt_copy (dt[nargs], vec_oprnd1);
3259 = vect_get_vec_def_for_stmt_copy (dt[nargs], vec_oprnd0);
3262 vargs = tree_cons (NULL_TREE, vec_oprnd0, vargs);
3263 vargs = tree_cons (NULL_TREE, vec_oprnd1, vargs);
3267 vargs = nreverse (vargs);
3269 rhs = build_function_call_expr (fndecl, vargs);
3270 new_stmt = build_gimple_modify_stmt (vec_dest, rhs);
3271 new_temp = make_ssa_name (vec_dest, new_stmt);
3272 GIMPLE_STMT_OPERAND (new_stmt, 0) = new_temp;
3274 vect_finish_stmt_generation (stmt, new_stmt, bsi);
3277 STMT_VINFO_VEC_STMT (stmt_info) = new_stmt;
3279 STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt;
3281 prev_stmt_info = vinfo_for_stmt (new_stmt);
3284 *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info);
3289 /* No current target implements this case. */
3293 /* The call in STMT might prevent it from being removed in dce.
3294 We however cannot remove it here, due to the way the ssa name
3295 it defines is mapped to the new definition. So just replace
3296 rhs of the statement with something harmless. */
3297 type = TREE_TYPE (scalar_dest);
3298 GIMPLE_STMT_OPERAND (stmt, 1) = fold_convert (type, integer_zero_node);
3305 /* Function vect_gen_widened_results_half
3307 Create a vector stmt whose code, type, number of arguments, and result
3308 variable are CODE, VECTYPE, OP_TYPE, and VEC_DEST, and its arguments are
3309 VEC_OPRND0 and VEC_OPRND1. The new vector stmt is to be inserted at BSI.
3310 In the case that CODE is a CALL_EXPR, this means that a call to DECL
3311 needs to be created (DECL is a function-decl of a target-builtin).
3312 STMT is the original scalar stmt that we are vectorizing. */
3315 vect_gen_widened_results_half (enum tree_code code, tree vectype, tree decl,
3316 tree vec_oprnd0, tree vec_oprnd1, int op_type,
3317 tree vec_dest, block_stmt_iterator *bsi,
3326 /* Generate half of the widened result: */
3327 if (code == CALL_EXPR)
3329 /* Target specific support */
3330 if (op_type == binary_op)
3331 expr = build_call_expr (decl, 2, vec_oprnd0, vec_oprnd1);
3333 expr = build_call_expr (decl, 1, vec_oprnd0);
3337 /* Generic support */
3338 gcc_assert (op_type == TREE_CODE_LENGTH (code));
3339 if (op_type == binary_op)
3340 expr = build2 (code, vectype, vec_oprnd0, vec_oprnd1);
3342 expr = build1 (code, vectype, vec_oprnd0);
3344 new_stmt = build_gimple_modify_stmt (vec_dest, expr);
3345 new_temp = make_ssa_name (vec_dest, new_stmt);
3346 GIMPLE_STMT_OPERAND (new_stmt, 0) = new_temp;
3347 vect_finish_stmt_generation (stmt, new_stmt, bsi);
3349 if (code == CALL_EXPR)
3351 FOR_EACH_SSA_TREE_OPERAND (sym, new_stmt, iter, SSA_OP_ALL_VIRTUALS)
3353 if (TREE_CODE (sym) == SSA_NAME)
3354 sym = SSA_NAME_VAR (sym);
3355 mark_sym_for_renaming (sym);
3363 /* Check if STMT performs a conversion operation, that can be vectorized.
3364 If VEC_STMT is also passed, vectorize the STMT: create a vectorized
3365 stmt to replace it, put it in VEC_STMT, and insert it at BSI.
3366 Return FALSE if not a vectorizable STMT, TRUE otherwise. */
3369 vectorizable_conversion (tree stmt, block_stmt_iterator *bsi,
3370 tree *vec_stmt, slp_tree slp_node)
3376 tree vec_oprnd0 = NULL_TREE, vec_oprnd1 = NULL_TREE;
3377 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
3378 loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info);
3379 struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
3380 enum tree_code code, code1 = ERROR_MARK, code2 = ERROR_MARK;
3381 tree decl1 = NULL_TREE, decl2 = NULL_TREE;
3384 enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type};
3385 tree new_stmt = NULL_TREE;
3386 stmt_vec_info prev_stmt_info;
3389 tree vectype_out, vectype_in;
3392 tree rhs_type, lhs_type;
3394 enum { NARROW, NONE, WIDEN } modifier;
3396 VEC(tree,heap) *vec_oprnds0 = NULL;
3399 /* Is STMT a vectorizable conversion? */
3401 if (!STMT_VINFO_RELEVANT_P (stmt_info))
3404 if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_loop_def)
3407 if (TREE_CODE (stmt) != GIMPLE_MODIFY_STMT)
3410 if (TREE_CODE (GIMPLE_STMT_OPERAND (stmt, 0)) != SSA_NAME)
3413 operation = GIMPLE_STMT_OPERAND (stmt, 1);
3414 code = TREE_CODE (operation);
3415 if (code != FIX_TRUNC_EXPR && code != FLOAT_EXPR)
3418 /* Check types of lhs and rhs. */
3419 op0 = TREE_OPERAND (operation, 0);
3420 rhs_type = TREE_TYPE (op0);
3421 vectype_in = get_vectype_for_scalar_type (rhs_type);
3424 nunits_in = TYPE_VECTOR_SUBPARTS (vectype_in);
3426 scalar_dest = GIMPLE_STMT_OPERAND (stmt, 0);
3427 lhs_type = TREE_TYPE (scalar_dest);
3428 vectype_out = get_vectype_for_scalar_type (lhs_type);
3431 nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out);
3434 if (nunits_in == nunits_out / 2)
3436 else if (nunits_out == nunits_in)
3438 else if (nunits_out == nunits_in / 2)
3443 if (modifier == NONE)
3444 gcc_assert (STMT_VINFO_VECTYPE (stmt_info) == vectype_out);
3446 /* Bail out if the types are both integral or non-integral. */
3447 if ((INTEGRAL_TYPE_P (rhs_type) && INTEGRAL_TYPE_P (lhs_type))
3448 || (!INTEGRAL_TYPE_P (rhs_type) && !INTEGRAL_TYPE_P (lhs_type)))
3451 if (modifier == NARROW)
3452 ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_out;
3454 ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in;
3456 /* FORNOW: SLP with multiple types is not supported. The SLP analysis verifies
3457 this, so we can safely override NCOPIES with 1 here. */
3461 /* Sanity check: make sure that at least one copy of the vectorized stmt
3462 needs to be generated. */
3463 gcc_assert (ncopies >= 1);
3465 /* FORNOW. This restriction should be relaxed. */
3466 if (nested_in_vect_loop_p (loop, stmt) && ncopies > 1)
3468 if (vect_print_dump_info (REPORT_DETAILS))
3469 fprintf (vect_dump, "multiple types in nested loop.");
3473 /* Check the operands of the operation. */
3474 if (!vect_is_simple_use (op0, loop_vinfo, &def_stmt, &def, &dt[0]))
3476 if (vect_print_dump_info (REPORT_DETAILS))
3477 fprintf (vect_dump, "use not simple.");
3481 /* Supportable by target? */
3482 if ((modifier == NONE
3483 && !targetm.vectorize.builtin_conversion (code, vectype_in))
3484 || (modifier == WIDEN
3485 && !supportable_widening_operation (code, stmt, vectype_in,
3488 || (modifier == NARROW
3489 && !supportable_narrowing_operation (code, stmt, vectype_in,
3492 if (vect_print_dump_info (REPORT_DETAILS))
3493 fprintf (vect_dump, "op not supported by target.");
3497 if (modifier != NONE)
3499 STMT_VINFO_VECTYPE (stmt_info) = vectype_in;
3500 /* FORNOW: SLP not supported. */
3501 if (STMT_SLP_TYPE (stmt_info))
3505 if (!vec_stmt) /* transformation not required. */
3507 STMT_VINFO_TYPE (stmt_info) = type_conversion_vec_info_type;
3512 if (vect_print_dump_info (REPORT_DETAILS))
3513 fprintf (vect_dump, "transform conversion.");
3516 vec_dest = vect_create_destination_var (scalar_dest, vectype_out);
3518 if (modifier == NONE && !slp_node)
3519 vec_oprnds0 = VEC_alloc (tree, heap, 1);
3521 prev_stmt_info = NULL;
3525 for (j = 0; j < ncopies; j++)
3531 vect_get_vec_defs (op0, NULL, stmt, &vec_oprnds0, NULL, slp_node);
3533 vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds0, NULL);
3536 targetm.vectorize.builtin_conversion (code, vectype_in);
3537 for (i = 0; VEC_iterate (tree, vec_oprnds0, i, vop0); i++)
3539 new_stmt = build_call_expr (builtin_decl, 1, vop0);
3541 /* Arguments are ready. create the new vector stmt. */
3542 new_stmt = build_gimple_modify_stmt (vec_dest, new_stmt);
3543 new_temp = make_ssa_name (vec_dest, new_stmt);
3544 GIMPLE_STMT_OPERAND (new_stmt, 0) = new_temp;
3545 vect_finish_stmt_generation (stmt, new_stmt, bsi);
3546 FOR_EACH_SSA_TREE_OPERAND (sym, new_stmt, iter,
3547 SSA_OP_ALL_VIRTUALS)
3549 if (TREE_CODE (sym) == SSA_NAME)
3550 sym = SSA_NAME_VAR (sym);
3551 mark_sym_for_renaming (sym);
3554 VEC_quick_push (tree, SLP_TREE_VEC_STMTS (slp_node), new_stmt);
3558 STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt;
3560 STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt;
3561 prev_stmt_info = vinfo_for_stmt (new_stmt);
3566 /* In case the vectorization factor (VF) is bigger than the number
3567 of elements that we can fit in a vectype (nunits), we have to
3568 generate more than one vector stmt - i.e - we need to "unroll"
3569 the vector stmt by a factor VF/nunits. */
3570 for (j = 0; j < ncopies; j++)
3573 vec_oprnd0 = vect_get_vec_def_for_operand (op0, stmt, NULL);
3575 vec_oprnd0 = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd0);
3577 STMT_VINFO_VECTYPE (stmt_info) = vectype_in;
3579 /* Generate first half of the widened result: */
3581 = vect_gen_widened_results_half (code1, vectype_out, decl1,
3582 vec_oprnd0, vec_oprnd1,
3583 unary_op, vec_dest, bsi, stmt);
3585 STMT_VINFO_VEC_STMT (stmt_info) = new_stmt;
3587 STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt;
3588 prev_stmt_info = vinfo_for_stmt (new_stmt);
3590 /* Generate second half of the widened result: */
3592 = vect_gen_widened_results_half (code2, vectype_out, decl2,
3593 vec_oprnd0, vec_oprnd1,
3594 unary_op, vec_dest, bsi, stmt);
3595 STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt;
3596 prev_stmt_info = vinfo_for_stmt (new_stmt);
3601 /* In case the vectorization factor (VF) is bigger than the number
3602 of elements that we can fit in a vectype (nunits), we have to
3603 generate more than one vector stmt - i.e - we need to "unroll"
3604 the vector stmt by a factor VF/nunits. */
3605 for (j = 0; j < ncopies; j++)
3610 vec_oprnd0 = vect_get_vec_def_for_operand (op0, stmt, NULL);
3611 vec_oprnd1 = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd0);
3615 vec_oprnd0 = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd1);
3616 vec_oprnd1 = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd0);
3619 /* Arguments are ready. Create the new vector stmt. */
3620 expr = build2 (code1, vectype_out, vec_oprnd0, vec_oprnd1);
3621 new_stmt = build_gimple_modify_stmt (vec_dest, expr);
3622 new_temp = make_ssa_name (vec_dest, new_stmt);
3623 GIMPLE_STMT_OPERAND (new_stmt, 0) = new_temp;
3624 vect_finish_stmt_generation (stmt, new_stmt, bsi);
3627 STMT_VINFO_VEC_STMT (stmt_info) = new_stmt;
3629 STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt;
3631 prev_stmt_info = vinfo_for_stmt (new_stmt);
3634 *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info);
3641 /* Function vectorizable_assignment.
3643 Check if STMT performs an assignment (copy) that can be vectorized.
3644 If VEC_STMT is also passed, vectorize the STMT: create a vectorized
3645 stmt to replace it, put it in VEC_STMT, and insert it at BSI.
3646 Return FALSE if not a vectorizable STMT, TRUE otherwise. */
3649 vectorizable_assignment (tree stmt, block_stmt_iterator *bsi, tree *vec_stmt,
3655 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
3656 tree vectype = STMT_VINFO_VECTYPE (stmt_info);
3657 loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info);
3660 enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type};
3661 int nunits = TYPE_VECTOR_SUBPARTS (vectype);
3662 int ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits;
3664 VEC(tree,heap) *vec_oprnds = NULL;
3667 gcc_assert (ncopies >= 1);
3669 return false; /* FORNOW */
3671 if (!STMT_VINFO_RELEVANT_P (stmt_info))
3674 if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_loop_def)
3677 /* Is vectorizable assignment? */
3678 if (TREE_CODE (stmt) != GIMPLE_MODIFY_STMT)
3681 scalar_dest = GIMPLE_STMT_OPERAND (stmt, 0);
3682 if (TREE_CODE (scalar_dest) != SSA_NAME)
3685 op = GIMPLE_STMT_OPERAND (stmt, 1);
3686 if (!vect_is_simple_use (op, loop_vinfo, &def_stmt, &def, &dt[0]))
3688 if (vect_print_dump_info (REPORT_DETAILS))
3689 fprintf (vect_dump, "use not simple.");
3693 if (!vec_stmt) /* transformation not required. */
3695 STMT_VINFO_TYPE (stmt_info) = assignment_vec_info_type;
3696 if (vect_print_dump_info (REPORT_DETAILS))
3697 fprintf (vect_dump, "=== vectorizable_assignment ===");
3698 vect_model_simple_cost (stmt_info, ncopies, dt, NULL);
3703 if (vect_print_dump_info (REPORT_DETAILS))
3704 fprintf (vect_dump, "transform assignment.");
3707 vec_dest = vect_create_destination_var (scalar_dest, vectype);
3710 vect_get_vec_defs (op, NULL, stmt, &vec_oprnds, NULL, slp_node);
3712 /* Arguments are ready. create the new vector stmt. */
3713 for (i = 0; VEC_iterate (tree, vec_oprnds, i, vop); i++)
3715 *vec_stmt = build_gimple_modify_stmt (vec_dest, vop);
3716 new_temp = make_ssa_name (vec_dest, *vec_stmt);
3717 GIMPLE_STMT_OPERAND (*vec_stmt, 0) = new_temp;
3718 vect_finish_stmt_generation (stmt, *vec_stmt, bsi);
3719 STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt;
3722 VEC_quick_push (tree, SLP_TREE_VEC_STMTS (slp_node), *vec_stmt);
3725 VEC_free (tree, heap, vec_oprnds);
3730 /* Function vect_min_worthwhile_factor.
3732 For a loop where we could vectorize the operation indicated by CODE,
3733 return the minimum vectorization factor that makes it worthwhile
3734 to use generic vectors. */
3736 vect_min_worthwhile_factor (enum tree_code code)
3757 /* Function vectorizable_induction
3759 Check if PHI performs an induction computation that can be vectorized.
3760 If VEC_STMT is also passed, vectorize the induction PHI: create a vectorized
3761 phi to replace it, put it in VEC_STMT, and add it to the same basic block.
3762 Return FALSE if not a vectorizable STMT, TRUE otherwise. */
3765 vectorizable_induction (tree phi, block_stmt_iterator *bsi ATTRIBUTE_UNUSED,
3768 stmt_vec_info stmt_info = vinfo_for_stmt (phi);
3769 tree vectype = STMT_VINFO_VECTYPE (stmt_info);
3770 loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info);
3771 int nunits = TYPE_VECTOR_SUBPARTS (vectype);
3772 int ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits;
3775 gcc_assert (ncopies >= 1);
3777 if (!STMT_VINFO_RELEVANT_P (stmt_info))
3780 /* FORNOW: SLP not supported. */
3781 if (STMT_SLP_TYPE (stmt_info))
3784 gcc_assert (STMT_VINFO_DEF_TYPE (stmt_info) == vect_induction_def);
3786 if (TREE_CODE (phi) != PHI_NODE)
3789 if (!vec_stmt) /* transformation not required. */
3791 STMT_VINFO_TYPE (stmt_info) = induc_vec_info_type;
3792 if (vect_print_dump_info (REPORT_DETAILS))
3793 fprintf (vect_dump, "=== vectorizable_induction ===");
3794 vect_model_induction_cost (stmt_info, ncopies);
3800 if (vect_print_dump_info (REPORT_DETAILS))
3801 fprintf (vect_dump, "transform induction phi.");
3803 vec_def = get_initial_def_for_induction (phi);
3804 *vec_stmt = SSA_NAME_DEF_STMT (vec_def);
3809 /* Function vectorizable_operation.
3811 Check if STMT performs a binary or unary operation that can be vectorized.
3812 If VEC_STMT is also passed, vectorize the STMT: create a vectorized
3813 stmt to replace it, put it in VEC_STMT, and insert it at BSI.
3814 Return FALSE if not a vectorizable STMT, TRUE otherwise. */
3817 vectorizable_operation (tree stmt, block_stmt_iterator *bsi, tree *vec_stmt,
3823 tree op0, op1 = NULL;
3824 tree vec_oprnd1 = NULL_TREE;
3825 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
3826 tree vectype = STMT_VINFO_VECTYPE (stmt_info);
3827 loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info);
3828 struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
3829 enum tree_code code;
3830 enum machine_mode vec_mode;
3835 enum machine_mode optab_op2_mode;
3837 enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type};
3838 tree new_stmt = NULL_TREE;
3839 stmt_vec_info prev_stmt_info;
3840 int nunits_in = TYPE_VECTOR_SUBPARTS (vectype);
3843 int ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in;
3845 VEC(tree,heap) *vec_oprnds0 = NULL, *vec_oprnds1 = NULL;
3848 bool scalar_shift_arg = false;
3850 /* FORNOW: SLP with multiple types is not supported. The SLP analysis verifies
3851 this, so we can safely override NCOPIES with 1 here. */
3854 gcc_assert (ncopies >= 1);
3855 /* FORNOW. This restriction should be relaxed. */
3856 if (nested_in_vect_loop_p (loop, stmt) && ncopies > 1)
3858 if (vect_print_dump_info (REPORT_DETAILS))
3859 fprintf (vect_dump, "multiple types in nested loop.");
3863 if (!STMT_VINFO_RELEVANT_P (stmt_info))
3866 if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_loop_def)
3869 /* Is STMT a vectorizable binary/unary operation? */
3870 if (TREE_CODE (stmt) != GIMPLE_MODIFY_STMT)
3873 if (TREE_CODE (GIMPLE_STMT_OPERAND (stmt, 0)) != SSA_NAME)
3876 scalar_dest = GIMPLE_STMT_OPERAND (stmt, 0);
3877 vectype_out = get_vectype_for_scalar_type (TREE_TYPE (scalar_dest));
3880 nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out);
3881 if (nunits_out != nunits_in)
3884 operation = GIMPLE_STMT_OPERAND (stmt, 1);
3885 code = TREE_CODE (operation);
3887 /* For pointer addition, we should use the normal plus for
3888 the vector addition. */
3889 if (code == POINTER_PLUS_EXPR)
3892 optab = optab_for_tree_code (code, vectype);
3894 /* Support only unary or binary operations. */
3895 op_type = TREE_OPERAND_LENGTH (operation);
3896 if (op_type != unary_op && op_type != binary_op)
3898 if (vect_print_dump_info (REPORT_DETAILS))
3899 fprintf (vect_dump, "num. args = %d (not unary/binary op).", op_type);
3903 op0 = TREE_OPERAND (operation, 0);
3904 if (!vect_is_simple_use (op0, loop_vinfo, &def_stmt, &def, &dt[0]))
3906 if (vect_print_dump_info (REPORT_DETAILS))
3907 fprintf (vect_dump, "use not simple.");
3911 if (op_type == binary_op)
3913 op1 = TREE_OPERAND (operation, 1);
3914 if (!vect_is_simple_use (op1, loop_vinfo, &def_stmt, &def, &dt[1]))
3916 if (vect_print_dump_info (REPORT_DETAILS))
3917 fprintf (vect_dump, "use not simple.");
3922 /* Supportable by target? */
3925 if (vect_print_dump_info (REPORT_DETAILS))
3926 fprintf (vect_dump, "no optab.");
3929 vec_mode = TYPE_MODE (vectype);
3930 icode = (int) optab_handler (optab, vec_mode)->insn_code;
3931 if (icode == CODE_FOR_nothing)
3933 if (vect_print_dump_info (REPORT_DETAILS))
3934 fprintf (vect_dump, "op not supported by target.");
3935 /* Check only during analysis. */
3936 if (GET_MODE_SIZE (vec_mode) != UNITS_PER_WORD
3937 || (LOOP_VINFO_VECT_FACTOR (loop_vinfo)
3938 < vect_min_worthwhile_factor (code)
3941 if (vect_print_dump_info (REPORT_DETAILS))
3942 fprintf (vect_dump, "proceeding using word mode.");
3945 /* Worthwhile without SIMD support? Check only during analysis. */
3946 if (!VECTOR_MODE_P (TYPE_MODE (vectype))
3947 && LOOP_VINFO_VECT_FACTOR (loop_vinfo)
3948 < vect_min_worthwhile_factor (code)
3951 if (vect_print_dump_info (REPORT_DETAILS))
3952 fprintf (vect_dump, "not worthwhile without SIMD support.");
3956 if (code == LSHIFT_EXPR || code == RSHIFT_EXPR)
3958 /* FORNOW: not yet supported. */
3959 if (!VECTOR_MODE_P (vec_mode))
3962 /* Invariant argument is needed for a vector shift
3963 by a scalar shift operand. */
3964 optab_op2_mode = insn_data[icode].operand[2].mode;
3965 if (!VECTOR_MODE_P (optab_op2_mode))
3967 if (dt[1] != vect_constant_def && dt[1] != vect_invariant_def)
3969 if (vect_print_dump_info (REPORT_DETAILS))
3970 fprintf (vect_dump, "operand mode requires invariant"
3975 scalar_shift_arg = true;
3979 if (!vec_stmt) /* transformation not required. */
3981 STMT_VINFO_TYPE (stmt_info) = op_vec_info_type;
3982 if (vect_print_dump_info (REPORT_DETAILS))
3983 fprintf (vect_dump, "=== vectorizable_operation ===");
3984 vect_model_simple_cost (stmt_info, ncopies, dt, NULL);
3990 if (vect_print_dump_info (REPORT_DETAILS))
3991 fprintf (vect_dump, "transform binary/unary operation.");
3994 vec_dest = vect_create_destination_var (scalar_dest, vectype);
3996 /* Allocate VECs for vector operands. In case of SLP, vector operands are
3997 created in the previous stages of the recursion, so no allocation is
3998 needed, except for the case of shift with scalar shift argument. In that
3999 case we store the scalar operand in VEC_OPRNDS1 for every vector stmt to
4000 be created to vectorize the SLP group, i.e., SLP_NODE->VEC_STMTS_SIZE.
4001 In case of loop-based vectorization we allocate VECs of size 1. We
4002 allocate VEC_OPRNDS1 only in case of binary operation. */
4005 vec_oprnds0 = VEC_alloc (tree, heap, 1);
4006 if (op_type == binary_op)
4007 vec_oprnds1 = VEC_alloc (tree, heap, 1);
4009 else if (scalar_shift_arg)
4010 vec_oprnds1 = VEC_alloc (tree, heap, slp_node->vec_stmts_size);
4012 /* In case the vectorization factor (VF) is bigger than the number
4013 of elements that we can fit in a vectype (nunits), we have to generate
4014 more than one vector stmt - i.e - we need to "unroll" the
4015 vector stmt by a factor VF/nunits. In doing so, we record a pointer
4016 from one copy of the vector stmt to the next, in the field
4017 STMT_VINFO_RELATED_STMT. This is necessary in order to allow following
4018 stages to find the correct vector defs to be used when vectorizing
4019 stmts that use the defs of the current stmt. The example below illustrates
4020 the vectorization process when VF=16 and nunits=4 (i.e - we need to create
4021 4 vectorized stmts):
4023 before vectorization:
4024 RELATED_STMT VEC_STMT
4028 step 1: vectorize stmt S1 (done in vectorizable_load. See more details
4030 RELATED_STMT VEC_STMT
4031 VS1_0: vx0 = memref0 VS1_1 -
4032 VS1_1: vx1 = memref1 VS1_2 -
4033 VS1_2: vx2 = memref2 VS1_3 -
4034 VS1_3: vx3 = memref3 - -
4035 S1: x = load - VS1_0
4038 step2: vectorize stmt S2 (done here):
4039 To vectorize stmt S2 we first need to find the relevant vector
4040 def for the first operand 'x'. This is, as usual, obtained from
4041 the vector stmt recorded in the STMT_VINFO_VEC_STMT of the stmt
4042 that defines 'x' (S1). This way we find the stmt VS1_0, and the
4043 relevant vector def 'vx0'. Having found 'vx0' we can generate
4044 the vector stmt VS2_0, and as usual, record it in the
4045 STMT_VINFO_VEC_STMT of stmt S2.
4046 When creating the second copy (VS2_1), we obtain the relevant vector
4047 def from the vector stmt recorded in the STMT_VINFO_RELATED_STMT of
4048 stmt VS1_0. This way we find the stmt VS1_1 and the relevant
4049 vector def 'vx1'. Using 'vx1' we create stmt VS2_1 and record a
4050 pointer to it in the STMT_VINFO_RELATED_STMT of the vector stmt VS2_0.
4051 Similarly when creating stmts VS2_2 and VS2_3. This is the resulting
4052 chain of stmts and pointers:
4053 RELATED_STMT VEC_STMT
4054 VS1_0: vx0 = memref0 VS1_1 -
4055 VS1_1: vx1 = memref1 VS1_2 -
4056 VS1_2: vx2 = memref2 VS1_3 -
4057 VS1_3: vx3 = memref3 - -
4058 S1: x = load - VS1_0
4059 VS2_0: vz0 = vx0 + v1 VS2_1 -
4060 VS2_1: vz1 = vx1 + v1 VS2_2 -
4061 VS2_2: vz2 = vx2 + v1 VS2_3 -
4062 VS2_3: vz3 = vx3 + v1 - -
4063 S2: z = x + 1 - VS2_0 */
4065 prev_stmt_info = NULL;
4066 for (j = 0; j < ncopies; j++)
4071 if (op_type == binary_op
4072 && (code == LSHIFT_EXPR || code == RSHIFT_EXPR))
4074 /* Vector shl and shr insn patterns can be defined with scalar
4075 operand 2 (shift operand). In this case, use constant or loop
4076 invariant op1 directly, without extending it to vector mode
4078 optab_op2_mode = insn_data[icode].operand[2].mode;
4079 if (!VECTOR_MODE_P (optab_op2_mode))
4081 if (vect_print_dump_info (REPORT_DETAILS))
4082 fprintf (vect_dump, "operand 1 using scalar mode.");
4084 VEC_quick_push (tree, vec_oprnds1, vec_oprnd1);
4087 /* Store vec_oprnd1 for every vector stmt to be created
4088 for SLP_NODE. We check during the analysis that all the
4089 shift arguments are the same.
4090 TODO: Allow different constants for different vector
4091 stmts generated for an SLP instance. */
4092 for (k = 0; k < slp_node->vec_stmts_size - 1; k++)
4093 VEC_quick_push (tree, vec_oprnds1, vec_oprnd1);
4098 /* vec_oprnd1 is available if operand 1 should be of a scalar-type
4099 (a special case for certain kind of vector shifts); otherwise,
4100 operand 1 should be of a vector type (the usual case). */
4101 if (op_type == binary_op && !vec_oprnd1)
4102 vect_get_vec_defs (op0, op1, stmt, &vec_oprnds0, &vec_oprnds1,
4105 vect_get_vec_defs (op0, NULL_TREE, stmt, &vec_oprnds0, NULL,
4109 vect_get_vec_defs_for_stmt_copy (dt, &vec_oprnds0, &vec_oprnds1);
4111 /* Arguments are ready. Create the new vector stmt. */
4112 for (i = 0; VEC_iterate (tree, vec_oprnds0, i, vop0); i++)
4114 if (op_type == binary_op)
4116 vop1 = VEC_index (tree, vec_oprnds1, i);
4117 new_stmt = build_gimple_modify_stmt (vec_dest,
4118 build2 (code, vectype, vop0, vop1));
4121 new_stmt = build_gimple_modify_stmt (vec_dest,
4122 build1 (code, vectype, vop0));
4124 new_temp = make_ssa_name (vec_dest, new_stmt);
4125 GIMPLE_STMT_OPERAND (new_stmt, 0) = new_temp;
4126 vect_finish_stmt_generation (stmt, new_stmt, bsi);
4128 VEC_quick_push (tree, SLP_TREE_VEC_STMTS (slp_node), new_stmt);
4132 STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt;
4134 STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt;
4135 prev_stmt_info = vinfo_for_stmt (new_stmt);
4138 VEC_free (tree, heap, vec_oprnds0);
4140 VEC_free (tree, heap, vec_oprnds1);
4146 /* Function vectorizable_type_demotion
4148 Check if STMT performs a binary or unary operation that involves
4149 type demotion, and if it can be vectorized.
4150 If VEC_STMT is also passed, vectorize the STMT: create a vectorized
4151 stmt to replace it, put it in VEC_STMT, and insert it at BSI.
4152 Return FALSE if not a vectorizable STMT, TRUE otherwise. */
4155 vectorizable_type_demotion (tree stmt, block_stmt_iterator *bsi,
4162 tree vec_oprnd0=NULL, vec_oprnd1=NULL;
4163 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
4164 loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info);
4165 struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
4166 enum tree_code code, code1 = ERROR_MARK;
4169 enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type};
4171 stmt_vec_info prev_stmt_info;
4180 if (!STMT_VINFO_RELEVANT_P (stmt_info))
4183 if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_loop_def)
4186 /* Is STMT a vectorizable type-demotion operation? */
4187 if (TREE_CODE (stmt) != GIMPLE_MODIFY_STMT)
4190 if (TREE_CODE (GIMPLE_STMT_OPERAND (stmt, 0)) != SSA_NAME)
4193 operation = GIMPLE_STMT_OPERAND (stmt, 1);
4194 code = TREE_CODE (operation);
4195 if (code != NOP_EXPR && code != CONVERT_EXPR)
4198 op0 = TREE_OPERAND (operation, 0);
4199 vectype_in = get_vectype_for_scalar_type (TREE_TYPE (op0));
4202 nunits_in = TYPE_VECTOR_SUBPARTS (vectype_in);
4204 scalar_dest = GIMPLE_STMT_OPERAND (stmt, 0);
4205 vectype_out = get_vectype_for_scalar_type (TREE_TYPE (scalar_dest));
4208 nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out);
4209 if (nunits_in != nunits_out / 2) /* FORNOW */
4212 ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_out;
4213 gcc_assert (ncopies >= 1);
4214 /* FORNOW. This restriction should be relaxed. */
4215 if (nested_in_vect_loop_p (loop, stmt) && ncopies > 1)
4217 if (vect_print_dump_info (REPORT_DETAILS))
4218 fprintf (vect_dump, "multiple types in nested loop.");
4222 if (! ((INTEGRAL_TYPE_P (TREE_TYPE (scalar_dest))
4223 && INTEGRAL_TYPE_P (TREE_TYPE (op0)))
4224 || (SCALAR_FLOAT_TYPE_P (TREE_TYPE (scalar_dest))
4225 && SCALAR_FLOAT_TYPE_P (TREE_TYPE (op0))
4226 && (code == NOP_EXPR || code == CONVERT_EXPR))))
4229 /* Check the operands of the operation. */
4230 if (!vect_is_simple_use (op0, loop_vinfo, &def_stmt, &def, &dt[0]))
4232 if (vect_print_dump_info (REPORT_DETAILS))
4233 fprintf (vect_dump, "use not simple.");
4237 /* Supportable by target? */
4238 if (!supportable_narrowing_operation (code, stmt, vectype_in, &code1))
4241 STMT_VINFO_VECTYPE (stmt_info) = vectype_in;
4243 if (!vec_stmt) /* transformation not required. */
4245 STMT_VINFO_TYPE (stmt_info) = type_demotion_vec_info_type;
4246 if (vect_print_dump_info (REPORT_DETAILS))
4247 fprintf (vect_dump, "=== vectorizable_demotion ===");
4248 vect_model_simple_cost (stmt_info, ncopies, dt, NULL);
4253 if (vect_print_dump_info (REPORT_DETAILS))
4254 fprintf (vect_dump, "transform type demotion operation. ncopies = %d.",
4258 vec_dest = vect_create_destination_var (scalar_dest, vectype_out);
4260 /* In case the vectorization factor (VF) is bigger than the number
4261 of elements that we can fit in a vectype (nunits), we have to generate
4262 more than one vector stmt - i.e - we need to "unroll" the
4263 vector stmt by a factor VF/nunits. */
4264 prev_stmt_info = NULL;
4265 for (j = 0; j < ncopies; j++)
4270 vec_oprnd0 = vect_get_vec_def_for_operand (op0, stmt, NULL);
4271 vec_oprnd1 = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd0);
4275 vec_oprnd0 = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd1);
4276 vec_oprnd1 = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd0);
4279 /* Arguments are ready. Create the new vector stmt. */
4280 expr = build2 (code1, vectype_out, vec_oprnd0, vec_oprnd1);
4281 new_stmt = build_gimple_modify_stmt (vec_dest, expr);
4282 new_temp = make_ssa_name (vec_dest, new_stmt);
4283 GIMPLE_STMT_OPERAND (new_stmt, 0) = new_temp;
4284 vect_finish_stmt_generation (stmt, new_stmt, bsi);
4287 STMT_VINFO_VEC_STMT (stmt_info) = new_stmt;
4289 STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt;
4291 prev_stmt_info = vinfo_for_stmt (new_stmt);
4294 *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info);
4299 /* Function vectorizable_type_promotion
4301 Check if STMT performs a binary or unary operation that involves
4302 type promotion, and if it can be vectorized.
4303 If VEC_STMT is also passed, vectorize the STMT: create a vectorized
4304 stmt to replace it, put it in VEC_STMT, and insert it at BSI.
4305 Return FALSE if not a vectorizable STMT, TRUE otherwise. */
4308 vectorizable_type_promotion (tree stmt, block_stmt_iterator *bsi,
4314 tree op0, op1 = NULL;
4315 tree vec_oprnd0=NULL, vec_oprnd1=NULL;
4316 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
4317 loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info);
4318 struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
4319 enum tree_code code, code1 = ERROR_MARK, code2 = ERROR_MARK;
4320 tree decl1 = NULL_TREE, decl2 = NULL_TREE;
4323 enum vect_def_type dt[2] = {vect_unknown_def_type, vect_unknown_def_type};
4325 stmt_vec_info prev_stmt_info;
4333 if (!STMT_VINFO_RELEVANT_P (stmt_info))
4336 if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_loop_def)
4339 /* Is STMT a vectorizable type-promotion operation? */
4340 if (TREE_CODE (stmt) != GIMPLE_MODIFY_STMT)
4343 if (TREE_CODE (GIMPLE_STMT_OPERAND (stmt, 0)) != SSA_NAME)
4346 operation = GIMPLE_STMT_OPERAND (stmt, 1);
4347 code = TREE_CODE (operation);
4348 if (code != NOP_EXPR && code != CONVERT_EXPR
4349 && code != WIDEN_MULT_EXPR)
4352 op0 = TREE_OPERAND (operation, 0);
4353 vectype_in = get_vectype_for_scalar_type (TREE_TYPE (op0));
4356 nunits_in = TYPE_VECTOR_SUBPARTS (vectype_in);
4358 scalar_dest = GIMPLE_STMT_OPERAND (stmt, 0);
4359 vectype_out = get_vectype_for_scalar_type (TREE_TYPE (scalar_dest));
4362 nunits_out = TYPE_VECTOR_SUBPARTS (vectype_out);
4363 if (nunits_out != nunits_in / 2) /* FORNOW */
4366 ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits_in;
4367 gcc_assert (ncopies >= 1);
4368 /* FORNOW. This restriction should be relaxed. */
4369 if (nested_in_vect_loop_p (loop, stmt) && ncopies > 1)
4371 if (vect_print_dump_info (REPORT_DETAILS))
4372 fprintf (vect_dump, "multiple types in nested loop.");
4376 if (! ((INTEGRAL_TYPE_P (TREE_TYPE (scalar_dest))
4377 && INTEGRAL_TYPE_P (TREE_TYPE (op0)))
4378 || (SCALAR_FLOAT_TYPE_P (TREE_TYPE (scalar_dest))
4379 && SCALAR_FLOAT_TYPE_P (TREE_TYPE (op0))
4380 && (code == CONVERT_EXPR || code == NOP_EXPR))))
4383 /* Check the operands of the operation. */
4384 if (!vect_is_simple_use (op0, loop_vinfo, &def_stmt, &def, &dt[0]))
4386 if (vect_print_dump_info (REPORT_DETAILS))
4387 fprintf (vect_dump, "use not simple.");
4391 op_type = TREE_CODE_LENGTH (code);
4392 if (op_type == binary_op)
4394 op1 = TREE_OPERAND (operation, 1);
4395 if (!vect_is_simple_use (op1, loop_vinfo, &def_stmt, &def, &dt[1]))
4397 if (vect_print_dump_info (REPORT_DETAILS))
4398 fprintf (vect_dump, "use not simple.");
4403 /* Supportable by target? */
4404 if (!supportable_widening_operation (code, stmt, vectype_in,
4405 &decl1, &decl2, &code1, &code2))
4408 STMT_VINFO_VECTYPE (stmt_info) = vectype_in;
4410 if (!vec_stmt) /* transformation not required. */
4412 STMT_VINFO_TYPE (stmt_info) = type_promotion_vec_info_type;
4413 if (vect_print_dump_info (REPORT_DETAILS))
4414 fprintf (vect_dump, "=== vectorizable_promotion ===");
4415 vect_model_simple_cost (stmt_info, 2*ncopies, dt, NULL);
4421 if (vect_print_dump_info (REPORT_DETAILS))
4422 fprintf (vect_dump, "transform type promotion operation. ncopies = %d.",
4426 vec_dest = vect_create_destination_var (scalar_dest, vectype_out);
4428 /* In case the vectorization factor (VF) is bigger than the number
4429 of elements that we can fit in a vectype (nunits), we have to generate
4430 more than one vector stmt - i.e - we need to "unroll" the
4431 vector stmt by a factor VF/nunits. */
4433 prev_stmt_info = NULL;
4434 for (j = 0; j < ncopies; j++)
4439 vec_oprnd0 = vect_get_vec_def_for_operand (op0, stmt, NULL);
4440 if (op_type == binary_op)
4441 vec_oprnd1 = vect_get_vec_def_for_operand (op1, stmt, NULL);
4445 vec_oprnd0 = vect_get_vec_def_for_stmt_copy (dt[0], vec_oprnd0);
4446 if (op_type == binary_op)
4447 vec_oprnd1 = vect_get_vec_def_for_stmt_copy (dt[1], vec_oprnd1);
4450 /* Arguments are ready. Create the new vector stmt. We are creating
4451 two vector defs because the widened result does not fit in one vector.
4452 The vectorized stmt can be expressed as a call to a taregt builtin,
4453 or a using a tree-code. */
4454 /* Generate first half of the widened result: */
4455 new_stmt = vect_gen_widened_results_half (code1, vectype_out, decl1,
4456 vec_oprnd0, vec_oprnd1, op_type, vec_dest, bsi, stmt);
4458 STMT_VINFO_VEC_STMT (stmt_info) = new_stmt;
4460 STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt;
4461 prev_stmt_info = vinfo_for_stmt (new_stmt);
4463 /* Generate second half of the widened result: */
4464 new_stmt = vect_gen_widened_results_half (code2, vectype_out, decl2,
4465 vec_oprnd0, vec_oprnd1, op_type, vec_dest, bsi, stmt);
4466 STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt;
4467 prev_stmt_info = vinfo_for_stmt (new_stmt);
4471 *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info);
4476 /* Function vect_strided_store_supported.
4478 Returns TRUE is INTERLEAVE_HIGH and INTERLEAVE_LOW operations are supported,
4479 and FALSE otherwise. */
4482 vect_strided_store_supported (tree vectype)
4484 optab interleave_high_optab, interleave_low_optab;
4487 mode = (int) TYPE_MODE (vectype);
4489 /* Check that the operation is supported. */
4490 interleave_high_optab = optab_for_tree_code (VEC_INTERLEAVE_HIGH_EXPR,
4492 interleave_low_optab = optab_for_tree_code (VEC_INTERLEAVE_LOW_EXPR,
4494 if (!interleave_high_optab || !interleave_low_optab)
4496 if (vect_print_dump_info (REPORT_DETAILS))
4497 fprintf (vect_dump, "no optab for interleave.");
4501 if (optab_handler (interleave_high_optab, mode)->insn_code
4503 || optab_handler (interleave_low_optab, mode)->insn_code
4504 == CODE_FOR_nothing)
4506 if (vect_print_dump_info (REPORT_DETAILS))
4507 fprintf (vect_dump, "interleave op not supported by target.");
4515 /* Function vect_permute_store_chain.
4517 Given a chain of interleaved stores in DR_CHAIN of LENGTH that must be
4518 a power of 2, generate interleave_high/low stmts to reorder the data
4519 correctly for the stores. Return the final references for stores in
4522 E.g., LENGTH is 4 and the scalar type is short, i.e., VF is 8.
4523 The input is 4 vectors each containing 8 elements. We assign a number to each
4524 element, the input sequence is:
4526 1st vec: 0 1 2 3 4 5 6 7
4527 2nd vec: 8 9 10 11 12 13 14 15
4528 3rd vec: 16 17 18 19 20 21 22 23
4529 4th vec: 24 25 26 27 28 29 30 31
4531 The output sequence should be:
4533 1st vec: 0 8 16 24 1 9 17 25
4534 2nd vec: 2 10 18 26 3 11 19 27
4535 3rd vec: 4 12 20 28 5 13 21 30
4536 4th vec: 6 14 22 30 7 15 23 31
4538 i.e., we interleave the contents of the four vectors in their order.
4540 We use interleave_high/low instructions to create such output. The input of
4541 each interleave_high/low operation is two vectors:
4544 the even elements of the result vector are obtained left-to-right from the
4545 high/low elements of the first vector. The odd elements of the result are
4546 obtained left-to-right from the high/low elements of the second vector.
4547 The output of interleave_high will be: 0 4 1 5
4548 and of interleave_low: 2 6 3 7
4551 The permutation is done in log LENGTH stages. In each stage interleave_high
4552 and interleave_low stmts are created for each pair of vectors in DR_CHAIN,
4553 where the first argument is taken from the first half of DR_CHAIN and the
4554 second argument from it's second half.
4557 I1: interleave_high (1st vec, 3rd vec)
4558 I2: interleave_low (1st vec, 3rd vec)
4559 I3: interleave_high (2nd vec, 4th vec)
4560 I4: interleave_low (2nd vec, 4th vec)
4562 The output for the first stage is:
4564 I1: 0 16 1 17 2 18 3 19
4565 I2: 4 20 5 21 6 22 7 23
4566 I3: 8 24 9 25 10 26 11 27
4567 I4: 12 28 13 29 14 30 15 31
4569 The output of the second stage, i.e. the final result is:
4571 I1: 0 8 16 24 1 9 17 25
4572 I2: 2 10 18 26 3 11 19 27
4573 I3: 4 12 20 28 5 13 21 30
4574 I4: 6 14 22 30 7 15 23 31. */
4577 vect_permute_store_chain (VEC(tree,heap) *dr_chain,
4578 unsigned int length,
4580 block_stmt_iterator *bsi,
4581 VEC(tree,heap) **result_chain)
4583 tree perm_dest, perm_stmt, vect1, vect2, high, low;
4584 tree vectype = STMT_VINFO_VECTYPE (vinfo_for_stmt (stmt));
4585 tree scalar_dest, tmp;
4588 VEC(tree,heap) *first, *second;
4590 scalar_dest = GIMPLE_STMT_OPERAND (stmt, 0);
4591 first = VEC_alloc (tree, heap, length/2);
4592 second = VEC_alloc (tree, heap, length/2);
4594 /* Check that the operation is supported. */
4595 if (!vect_strided_store_supported (vectype))
4598 *result_chain = VEC_copy (tree, heap, dr_chain);
4600 for (i = 0; i < exact_log2 (length); i++)
4602 for (j = 0; j < length/2; j++)
4604 vect1 = VEC_index (tree, dr_chain, j);
4605 vect2 = VEC_index (tree, dr_chain, j+length/2);
4607 /* Create interleaving stmt:
4608 in the case of big endian:
4609 high = interleave_high (vect1, vect2)
4610 and in the case of little endian:
4611 high = interleave_low (vect1, vect2). */
4612 perm_dest = create_tmp_var (vectype, "vect_inter_high");
4613 DECL_GIMPLE_REG_P (perm_dest) = 1;
4614 add_referenced_var (perm_dest);
4615 if (BYTES_BIG_ENDIAN)
4616 tmp = build2 (VEC_INTERLEAVE_HIGH_EXPR, vectype, vect1, vect2);
4618 tmp = build2 (VEC_INTERLEAVE_LOW_EXPR, vectype, vect1, vect2);
4619 perm_stmt = build_gimple_modify_stmt (perm_dest, tmp);
4620 high = make_ssa_name (perm_dest, perm_stmt);
4621 GIMPLE_STMT_OPERAND (perm_stmt, 0) = high;
4622 vect_finish_stmt_generation (stmt, perm_stmt, bsi);
4623 VEC_replace (tree, *result_chain, 2*j, high);
4625 /* Create interleaving stmt:
4626 in the case of big endian:
4627 low = interleave_low (vect1, vect2)
4628 and in the case of little endian:
4629 low = interleave_high (vect1, vect2). */
4630 perm_dest = create_tmp_var (vectype, "vect_inter_low");
4631 DECL_GIMPLE_REG_P (perm_dest) = 1;
4632 add_referenced_var (perm_dest);
4633 if (BYTES_BIG_ENDIAN)
4634 tmp = build2 (VEC_INTERLEAVE_LOW_EXPR, vectype, vect1, vect2);
4636 tmp = build2 (VEC_INTERLEAVE_HIGH_EXPR, vectype, vect1, vect2);
4637 perm_stmt = build_gimple_modify_stmt (perm_dest, tmp);
4638 low = make_ssa_name (perm_dest, perm_stmt);
4639 GIMPLE_STMT_OPERAND (perm_stmt, 0) = low;
4640 vect_finish_stmt_generation (stmt, perm_stmt, bsi);
4641 VEC_replace (tree, *result_chain, 2*j+1, low);
4643 dr_chain = VEC_copy (tree, heap, *result_chain);
4649 /* Function vectorizable_store.
4651 Check if STMT defines a non scalar data-ref (array/pointer/structure) that
4653 If VEC_STMT is also passed, vectorize the STMT: create a vectorized
4654 stmt to replace it, put it in VEC_STMT, and insert it at BSI.
4655 Return FALSE if not a vectorizable STMT, TRUE otherwise. */
4658 vectorizable_store (tree stmt, block_stmt_iterator *bsi, tree *vec_stmt,
4664 tree vec_oprnd = NULL_TREE;
4665 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
4666 struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info), *first_dr = NULL;
4667 tree vectype = STMT_VINFO_VECTYPE (stmt_info);
4668 loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info);
4669 struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
4670 enum machine_mode vec_mode;
4672 enum dr_alignment_support alignment_support_scheme;
4674 enum vect_def_type dt;
4675 stmt_vec_info prev_stmt_info = NULL;
4676 tree dataref_ptr = NULL_TREE;
4677 int nunits = TYPE_VECTOR_SUBPARTS (vectype);
4678 int ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits;
4680 tree next_stmt, first_stmt = NULL_TREE;
4681 bool strided_store = false;
4682 unsigned int group_size, i;
4683 VEC(tree,heap) *dr_chain = NULL, *oprnds = NULL, *result_chain = NULL;
4685 VEC(tree,heap) *vec_oprnds = NULL;
4686 bool slp = (slp_node != NULL);
4687 stmt_vec_info first_stmt_vinfo;
4688 unsigned int vec_num;
4690 /* FORNOW: SLP with multiple types is not supported. The SLP analysis verifies
4691 this, so we can safely override NCOPIES with 1 here. */
4695 gcc_assert (ncopies >= 1);
4697 /* FORNOW. This restriction should be relaxed. */
4698 if (nested_in_vect_loop_p (loop, stmt) && ncopies > 1)
4700 if (vect_print_dump_info (REPORT_DETAILS))
4701 fprintf (vect_dump, "multiple types in nested loop.");
4705 if (!STMT_VINFO_RELEVANT_P (stmt_info))
4708 if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_loop_def)
4711 /* Is vectorizable store? */
4713 if (TREE_CODE (stmt) != GIMPLE_MODIFY_STMT)
4716 scalar_dest = GIMPLE_STMT_OPERAND (stmt, 0);
4717 if (TREE_CODE (scalar_dest) != ARRAY_REF
4718 && TREE_CODE (scalar_dest) != INDIRECT_REF
4719 && !STMT_VINFO_STRIDED_ACCESS (stmt_info))
4722 op = GIMPLE_STMT_OPERAND (stmt, 1);
4723 if (!vect_is_simple_use (op, loop_vinfo, &def_stmt, &def, &dt))
4725 if (vect_print_dump_info (REPORT_DETAILS))
4726 fprintf (vect_dump, "use not simple.");
4730 vec_mode = TYPE_MODE (vectype);
4731 /* FORNOW. In some cases can vectorize even if data-type not supported
4732 (e.g. - array initialization with 0). */
4733 if (optab_handler (mov_optab, (int)vec_mode)->insn_code == CODE_FOR_nothing)
4736 if (!STMT_VINFO_DATA_REF (stmt_info))
4739 if (STMT_VINFO_STRIDED_ACCESS (stmt_info))
4741 strided_store = true;
4742 first_stmt = DR_GROUP_FIRST_DR (stmt_info);
4743 if (!vect_strided_store_supported (vectype)
4744 && !PURE_SLP_STMT (stmt_info) && !slp)
4747 if (first_stmt == stmt)
4749 /* STMT is the leader of the group. Check the operands of all the
4750 stmts of the group. */
4751 next_stmt = DR_GROUP_NEXT_DR (stmt_info);
4754 op = GIMPLE_STMT_OPERAND (next_stmt, 1);
4755 if (!vect_is_simple_use (op, loop_vinfo, &def_stmt, &def, &dt))
4757 if (vect_print_dump_info (REPORT_DETAILS))
4758 fprintf (vect_dump, "use not simple.");
4761 next_stmt = DR_GROUP_NEXT_DR (vinfo_for_stmt (next_stmt));
4766 if (!vec_stmt) /* transformation not required. */
4768 STMT_VINFO_TYPE (stmt_info) = store_vec_info_type;
4769 if (!PURE_SLP_STMT (stmt_info))
4770 vect_model_store_cost (stmt_info, ncopies, dt, NULL);
4778 first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt));
4779 group_size = DR_GROUP_SIZE (vinfo_for_stmt (first_stmt));
4781 DR_GROUP_STORE_COUNT (vinfo_for_stmt (first_stmt))++;
4784 gcc_assert (!nested_in_vect_loop_p (loop, stmt));
4786 /* We vectorize all the stmts of the interleaving group when we
4787 reach the last stmt in the group. */
4788 if (DR_GROUP_STORE_COUNT (vinfo_for_stmt (first_stmt))
4789 < DR_GROUP_SIZE (vinfo_for_stmt (first_stmt))
4792 *vec_stmt = NULL_TREE;
4797 strided_store = false;
4799 /* VEC_NUM is the number of vect stmts to be created for this group. */
4800 if (slp && SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node) < group_size)
4801 vec_num = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node);
4803 vec_num = group_size;
4809 group_size = vec_num = 1;
4810 first_stmt_vinfo = stmt_info;
4813 if (vect_print_dump_info (REPORT_DETAILS))
4814 fprintf (vect_dump, "transform store. ncopies = %d",ncopies);
4816 dr_chain = VEC_alloc (tree, heap, group_size);
4817 oprnds = VEC_alloc (tree, heap, group_size);
4819 alignment_support_scheme = vect_supportable_dr_alignment (first_dr);
4820 gcc_assert (alignment_support_scheme);
4821 gcc_assert (alignment_support_scheme == dr_aligned); /* FORNOW */
4823 /* In case the vectorization factor (VF) is bigger than the number
4824 of elements that we can fit in a vectype (nunits), we have to generate
4825 more than one vector stmt - i.e - we need to "unroll" the
4826 vector stmt by a factor VF/nunits. For more details see documentation in
4827 vect_get_vec_def_for_copy_stmt. */
4829 /* In case of interleaving (non-unit strided access):
4836 We create vectorized stores starting from base address (the access of the
4837 first stmt in the chain (S2 in the above example), when the last store stmt
4838 of the chain (S4) is reached:
4841 VS2: &base + vec_size*1 = vx0
4842 VS3: &base + vec_size*2 = vx1
4843 VS4: &base + vec_size*3 = vx3
4845 Then permutation statements are generated:
4847 VS5: vx5 = VEC_INTERLEAVE_HIGH_EXPR < vx0, vx3 >
4848 VS6: vx6 = VEC_INTERLEAVE_LOW_EXPR < vx0, vx3 >
4851 And they are put in STMT_VINFO_VEC_STMT of the corresponding scalar stmts
4852 (the order of the data-refs in the output of vect_permute_store_chain
4853 corresponds to the order of scalar stmts in the interleaving chain - see
4854 the documentation of vect_permute_store_chain()).
4856 In case of both multiple types and interleaving, above vector stores and
4857 permutation stmts are created for every copy. The result vector stmts are
4858 put in STMT_VINFO_VEC_STMT for the first copy and in the corresponding
4859 STMT_VINFO_RELATED_STMT for the next copies.
4862 prev_stmt_info = NULL;
4863 for (j = 0; j < ncopies; j++)
4872 /* Get vectorized arguments for SLP_NODE. */
4873 vect_get_slp_defs (slp_node, &vec_oprnds, NULL);
4875 vec_oprnd = VEC_index (tree, vec_oprnds, 0);
4879 /* For interleaved stores we collect vectorized defs for all the
4880 stores in the group in DR_CHAIN and OPRNDS. DR_CHAIN is then
4881 used as an input to vect_permute_store_chain(), and OPRNDS as
4882 an input to vect_get_vec_def_for_stmt_copy() for the next copy.
4884 If the store is not strided, GROUP_SIZE is 1, and DR_CHAIN and
4885 OPRNDS are of size 1. */
4886 next_stmt = first_stmt;
4887 for (i = 0; i < group_size; i++)
4889 /* Since gaps are not supported for interleaved stores,
4890 GROUP_SIZE is the exact number of stmts in the chain.
4891 Therefore, NEXT_STMT can't be NULL_TREE. In case that
4892 there is no interleaving, GROUP_SIZE is 1, and only one
4893 iteration of the loop will be executed. */
4894 gcc_assert (next_stmt);
4895 op = GIMPLE_STMT_OPERAND (next_stmt, 1);
4897 vec_oprnd = vect_get_vec_def_for_operand (op, next_stmt,
4899 VEC_quick_push(tree, dr_chain, vec_oprnd);
4900 VEC_quick_push(tree, oprnds, vec_oprnd);
4901 next_stmt = DR_GROUP_NEXT_DR (vinfo_for_stmt (next_stmt));
4904 dataref_ptr = vect_create_data_ref_ptr (first_stmt, NULL, NULL_TREE,
4905 &dummy, &ptr_incr, false,
4906 TREE_TYPE (vec_oprnd), &inv_p);
4907 gcc_assert (!inv_p);
4911 /* FORNOW SLP doesn't work for multiple types. */
4914 /* For interleaved stores we created vectorized defs for all the
4915 defs stored in OPRNDS in the previous iteration (previous copy).
4916 DR_CHAIN is then used as an input to vect_permute_store_chain(),
4917 and OPRNDS as an input to vect_get_vec_def_for_stmt_copy() for the
4919 If the store is not strided, GROUP_SIZE is 1, and DR_CHAIN and
4920 OPRNDS are of size 1. */
4921 for (i = 0; i < group_size; i++)
4923 op = VEC_index (tree, oprnds, i);
4924 vect_is_simple_use (op, loop_vinfo, &def_stmt, &def, &dt);
4925 vec_oprnd = vect_get_vec_def_for_stmt_copy (dt, op);
4926 VEC_replace(tree, dr_chain, i, vec_oprnd);
4927 VEC_replace(tree, oprnds, i, vec_oprnd);
4930 bump_vector_ptr (dataref_ptr, ptr_incr, bsi, stmt, NULL_TREE);
4935 result_chain = VEC_alloc (tree, heap, group_size);
4937 if (!vect_permute_store_chain (dr_chain, group_size, stmt, bsi,
4942 next_stmt = first_stmt;
4943 for (i = 0; i < vec_num; i++)
4946 /* Bump the vector pointer. */
4947 dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, bsi, stmt,
4951 vec_oprnd = VEC_index (tree, vec_oprnds, i);
4952 else if (strided_store)
4953 /* For strided stores vectorized defs are interleaved in
4954 vect_permute_store_chain(). */
4955 vec_oprnd = VEC_index (tree, result_chain, i);
4957 data_ref = build_fold_indirect_ref (dataref_ptr);
4958 /* Arguments are ready. Create the new vector stmt. */
4959 new_stmt = build_gimple_modify_stmt (data_ref, vec_oprnd);
4960 vect_finish_stmt_generation (stmt, new_stmt, bsi);
4961 mark_symbols_for_renaming (new_stmt);
4964 STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt;
4966 STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt;
4968 prev_stmt_info = vinfo_for_stmt (new_stmt);
4969 next_stmt = DR_GROUP_NEXT_DR (vinfo_for_stmt (next_stmt));
4979 /* Function vect_setup_realignment
4981 This function is called when vectorizing an unaligned load using
4982 the dr_explicit_realign[_optimized] scheme.
4983 This function generates the following code at the loop prolog:
4986 x msq_init = *(floor(p)); # prolog load
4987 realignment_token = call target_builtin;
4989 x msq = phi (msq_init, ---)
4991 The stmts marked with x are generated only for the case of
4992 dr_explicit_realign_optimized.
4994 The code above sets up a new (vector) pointer, pointing to the first
4995 location accessed by STMT, and a "floor-aligned" load using that pointer.
4996 It also generates code to compute the "realignment-token" (if the relevant
4997 target hook was defined), and creates a phi-node at the loop-header bb
4998 whose arguments are the result of the prolog-load (created by this
4999 function) and the result of a load that takes place in the loop (to be
5000 created by the caller to this function).
5002 For the case of dr_explicit_realign_optimized:
5003 The caller to this function uses the phi-result (msq) to create the
5004 realignment code inside the loop, and sets up the missing phi argument,
5007 msq = phi (msq_init, lsq)
5008 lsq = *(floor(p')); # load in loop
5009 result = realign_load (msq, lsq, realignment_token);
5011 For the case of dr_explicit_realign:
5013 msq = *(floor(p)); # load in loop
5015 lsq = *(floor(p')); # load in loop
5016 result = realign_load (msq, lsq, realignment_token);
5019 STMT - (scalar) load stmt to be vectorized. This load accesses
5020 a memory location that may be unaligned.
5021 BSI - place where new code is to be inserted.
5022 ALIGNMENT_SUPPORT_SCHEME - which of the two misalignment handling schemes
5026 REALIGNMENT_TOKEN - the result of a call to the builtin_mask_for_load
5027 target hook, if defined.
5028 Return value - the result of the loop-header phi node. */
5031 vect_setup_realignment (tree stmt, block_stmt_iterator *bsi,
5032 tree *realignment_token,
5033 enum dr_alignment_support alignment_support_scheme,
5035 struct loop **at_loop)
5037 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
5038 tree vectype = STMT_VINFO_VECTYPE (stmt_info);
5039 loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info);
5040 struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
5042 tree scalar_dest = GIMPLE_STMT_OPERAND (stmt, 0);
5049 tree msq_init = NULL_TREE;
5052 tree msq = NULL_TREE;
5053 tree stmts = NULL_TREE;
5055 bool compute_in_loop = false;
5056 bool nested_in_vect_loop = nested_in_vect_loop_p (loop, stmt);
5057 struct loop *containing_loop = (bb_for_stmt (stmt))->loop_father;
5058 struct loop *loop_for_initial_load;
5060 gcc_assert (alignment_support_scheme == dr_explicit_realign
5061 || alignment_support_scheme == dr_explicit_realign_optimized);
5063 /* We need to generate three things:
5064 1. the misalignment computation
5065 2. the extra vector load (for the optimized realignment scheme).
5066 3. the phi node for the two vectors from which the realignment is
5067 done (for the optimized realignment scheme).
5070 /* 1. Determine where to generate the misalignment computation.
5072 If INIT_ADDR is NULL_TREE, this indicates that the misalignment
5073 calculation will be generated by this function, outside the loop (in the
5074 preheader). Otherwise, INIT_ADDR had already been computed for us by the
5075 caller, inside the loop.
5077 Background: If the misalignment remains fixed throughout the iterations of
5078 the loop, then both realignment schemes are applicable, and also the
5079 misalignment computation can be done outside LOOP. This is because we are
5080 vectorizing LOOP, and so the memory accesses in LOOP advance in steps that
5081 are a multiple of VS (the Vector Size), and therefore the misalignment in
5082 different vectorized LOOP iterations is always the same.
5083 The problem arises only if the memory access is in an inner-loop nested
5084 inside LOOP, which is now being vectorized using outer-loop vectorization.
5085 This is the only case when the misalignment of the memory access may not
5086 remain fixed throughout the iterations of the inner-loop (as explained in
5087 detail in vect_supportable_dr_alignment). In this case, not only is the
5088 optimized realignment scheme not applicable, but also the misalignment
5089 computation (and generation of the realignment token that is passed to
5090 REALIGN_LOAD) have to be done inside the loop.
5092 In short, INIT_ADDR indicates whether we are in a COMPUTE_IN_LOOP mode
5093 or not, which in turn determines if the misalignment is computed inside
5094 the inner-loop, or outside LOOP. */
5096 if (init_addr != NULL_TREE)
5098 compute_in_loop = true;
5099 gcc_assert (alignment_support_scheme == dr_explicit_realign);
5103 /* 2. Determine where to generate the extra vector load.
5105 For the optimized realignment scheme, instead of generating two vector
5106 loads in each iteration, we generate a single extra vector load in the
5107 preheader of the loop, and in each iteration reuse the result of the
5108 vector load from the previous iteration. In case the memory access is in
5109 an inner-loop nested inside LOOP, which is now being vectorized using
5110 outer-loop vectorization, we need to determine whether this initial vector
5111 load should be generated at the preheader of the inner-loop, or can be
5112 generated at the preheader of LOOP. If the memory access has no evolution
5113 in LOOP, it can be generated in the preheader of LOOP. Otherwise, it has
5114 to be generated inside LOOP (in the preheader of the inner-loop). */
5116 if (nested_in_vect_loop)
5118 tree outerloop_step = STMT_VINFO_DR_STEP (stmt_info);
5119 bool invariant_in_outerloop =
5120 (tree_int_cst_compare (outerloop_step, size_zero_node) == 0);
5121 loop_for_initial_load = (invariant_in_outerloop ? loop : loop->inner);
5124 loop_for_initial_load = loop;
5126 *at_loop = loop_for_initial_load;
5128 /* 3. For the case of the optimized realignment, create the first vector
5129 load at the loop preheader. */
5131 if (alignment_support_scheme == dr_explicit_realign_optimized)
5133 /* Create msq_init = *(floor(p1)) in the loop preheader */
5135 gcc_assert (!compute_in_loop);
5136 pe = loop_preheader_edge (loop_for_initial_load);
5137 vec_dest = vect_create_destination_var (scalar_dest, vectype);
5138 ptr = vect_create_data_ref_ptr (stmt, loop_for_initial_load, NULL_TREE,
5139 &init_addr, &inc, true, NULL_TREE, &inv_p);
5140 data_ref = build1 (ALIGN_INDIRECT_REF, vectype, ptr);
5141 new_stmt = build_gimple_modify_stmt (vec_dest, data_ref);
5142 new_temp = make_ssa_name (vec_dest, new_stmt);
5143 GIMPLE_STMT_OPERAND (new_stmt, 0) = new_temp;
5144 mark_symbols_for_renaming (new_stmt);
5145 new_bb = bsi_insert_on_edge_immediate (pe, new_stmt);
5146 gcc_assert (!new_bb);
5147 msq_init = GIMPLE_STMT_OPERAND (new_stmt, 0);
5150 /* 4. Create realignment token using a target builtin, if available.
5151 It is done either inside the containing loop, or before LOOP (as
5152 determined above). */
5154 if (targetm.vectorize.builtin_mask_for_load)
5158 /* Compute INIT_ADDR - the initial addressed accessed by this memref. */
5159 if (compute_in_loop)
5160 gcc_assert (init_addr); /* already computed by the caller. */
5163 /* Generate the INIT_ADDR computation outside LOOP. */
5164 init_addr = vect_create_addr_base_for_vector_ref (stmt, &stmts,
5166 pe = loop_preheader_edge (loop);
5167 new_bb = bsi_insert_on_edge_immediate (pe, stmts);
5168 gcc_assert (!new_bb);
5171 builtin_decl = targetm.vectorize.builtin_mask_for_load ();
5172 new_stmt = build_call_expr (builtin_decl, 1, init_addr);
5173 vec_dest = vect_create_destination_var (scalar_dest,
5174 TREE_TYPE (new_stmt));
5175 new_stmt = build_gimple_modify_stmt (vec_dest, new_stmt);
5176 new_temp = make_ssa_name (vec_dest, new_stmt);
5177 GIMPLE_STMT_OPERAND (new_stmt, 0) = new_temp;
5179 if (compute_in_loop)
5180 bsi_insert_before (bsi, new_stmt, BSI_SAME_STMT);
5183 /* Generate the misalignment computation outside LOOP. */
5184 pe = loop_preheader_edge (loop);
5185 new_bb = bsi_insert_on_edge_immediate (pe, new_stmt);
5186 gcc_assert (!new_bb);
5189 *realignment_token = GIMPLE_STMT_OPERAND (new_stmt, 0);
5191 /* The result of the CALL_EXPR to this builtin is determined from
5192 the value of the parameter and no global variables are touched
5193 which makes the builtin a "const" function. Requiring the
5194 builtin to have the "const" attribute makes it unnecessary
5195 to call mark_call_clobbered. */
5196 gcc_assert (TREE_READONLY (builtin_decl));
5199 if (alignment_support_scheme == dr_explicit_realign)
5202 gcc_assert (!compute_in_loop);
5203 gcc_assert (alignment_support_scheme == dr_explicit_realign_optimized);
5206 /* 5. Create msq = phi <msq_init, lsq> in loop */
5208 pe = loop_preheader_edge (containing_loop);
5209 vec_dest = vect_create_destination_var (scalar_dest, vectype);
5210 msq = make_ssa_name (vec_dest, NULL_TREE);
5211 phi_stmt = create_phi_node (msq, containing_loop->header);
5212 SSA_NAME_DEF_STMT (msq) = phi_stmt;
5213 add_phi_arg (phi_stmt, msq_init, pe);
5219 /* Function vect_strided_load_supported.
5221 Returns TRUE is EXTRACT_EVEN and EXTRACT_ODD operations are supported,
5222 and FALSE otherwise. */
5225 vect_strided_load_supported (tree vectype)
5227 optab perm_even_optab, perm_odd_optab;
5230 mode = (int) TYPE_MODE (vectype);
5232 perm_even_optab = optab_for_tree_code (VEC_EXTRACT_EVEN_EXPR, vectype);
5233 if (!perm_even_optab)
5235 if (vect_print_dump_info (REPORT_DETAILS))
5236 fprintf (vect_dump, "no optab for perm_even.");
5240 if (optab_handler (perm_even_optab, mode)->insn_code == CODE_FOR_nothing)
5242 if (vect_print_dump_info (REPORT_DETAILS))
5243 fprintf (vect_dump, "perm_even op not supported by target.");
5247 perm_odd_optab = optab_for_tree_code (VEC_EXTRACT_ODD_EXPR, vectype);
5248 if (!perm_odd_optab)
5250 if (vect_print_dump_info (REPORT_DETAILS))
5251 fprintf (vect_dump, "no optab for perm_odd.");
5255 if (optab_handler (perm_odd_optab, mode)->insn_code == CODE_FOR_nothing)
5257 if (vect_print_dump_info (REPORT_DETAILS))
5258 fprintf (vect_dump, "perm_odd op not supported by target.");
5265 /* Function vect_permute_load_chain.
5267 Given a chain of interleaved loads in DR_CHAIN of LENGTH that must be
5268 a power of 2, generate extract_even/odd stmts to reorder the input data
5269 correctly. Return the final references for loads in RESULT_CHAIN.
5271 E.g., LENGTH is 4 and the scalar type is short, i.e., VF is 8.
5272 The input is 4 vectors each containing 8 elements. We assign a number to each
5273 element, the input sequence is:
5275 1st vec: 0 1 2 3 4 5 6 7
5276 2nd vec: 8 9 10 11 12 13 14 15
5277 3rd vec: 16 17 18 19 20 21 22 23
5278 4th vec: 24 25 26 27 28 29 30 31
5280 The output sequence should be:
5282 1st vec: 0 4 8 12 16 20 24 28
5283 2nd vec: 1 5 9 13 17 21 25 29
5284 3rd vec: 2 6 10 14 18 22 26 30
5285 4th vec: 3 7 11 15 19 23 27 31
5287 i.e., the first output vector should contain the first elements of each
5288 interleaving group, etc.
5290 We use extract_even/odd instructions to create such output. The input of each
5291 extract_even/odd operation is two vectors
5295 and the output is the vector of extracted even/odd elements. The output of
5296 extract_even will be: 0 2 4 6
5297 and of extract_odd: 1 3 5 7
5300 The permutation is done in log LENGTH stages. In each stage extract_even and
5301 extract_odd stmts are created for each pair of vectors in DR_CHAIN in their
5302 order. In our example,
5304 E1: extract_even (1st vec, 2nd vec)
5305 E2: extract_odd (1st vec, 2nd vec)
5306 E3: extract_even (3rd vec, 4th vec)
5307 E4: extract_odd (3rd vec, 4th vec)
5309 The output for the first stage will be:
5311 E1: 0 2 4 6 8 10 12 14
5312 E2: 1 3 5 7 9 11 13 15
5313 E3: 16 18 20 22 24 26 28 30
5314 E4: 17 19 21 23 25 27 29 31
5316 In order to proceed and create the correct sequence for the next stage (or
5317 for the correct output, if the second stage is the last one, as in our
5318 example), we first put the output of extract_even operation and then the
5319 output of extract_odd in RESULT_CHAIN (which is then copied to DR_CHAIN).
5320 The input for the second stage is:
5322 1st vec (E1): 0 2 4 6 8 10 12 14
5323 2nd vec (E3): 16 18 20 22 24 26 28 30
5324 3rd vec (E2): 1 3 5 7 9 11 13 15
5325 4th vec (E4): 17 19 21 23 25 27 29 31
5327 The output of the second stage:
5329 E1: 0 4 8 12 16 20 24 28
5330 E2: 2 6 10 14 18 22 26 30
5331 E3: 1 5 9 13 17 21 25 29
5332 E4: 3 7 11 15 19 23 27 31
5334 And RESULT_CHAIN after reordering:
5336 1st vec (E1): 0 4 8 12 16 20 24 28
5337 2nd vec (E3): 1 5 9 13 17 21 25 29
5338 3rd vec (E2): 2 6 10 14 18 22 26 30
5339 4th vec (E4): 3 7 11 15 19 23 27 31. */
5342 vect_permute_load_chain (VEC(tree,heap) *dr_chain,
5343 unsigned int length,
5345 block_stmt_iterator *bsi,
5346 VEC(tree,heap) **result_chain)
5348 tree perm_dest, perm_stmt, data_ref, first_vect, second_vect;
5349 tree vectype = STMT_VINFO_VECTYPE (vinfo_for_stmt (stmt));
5354 /* Check that the operation is supported. */
5355 if (!vect_strided_load_supported (vectype))
5358 *result_chain = VEC_copy (tree, heap, dr_chain);
5359 for (i = 0; i < exact_log2 (length); i++)
5361 for (j = 0; j < length; j +=2)
5363 first_vect = VEC_index (tree, dr_chain, j);
5364 second_vect = VEC_index (tree, dr_chain, j+1);
5366 /* data_ref = permute_even (first_data_ref, second_data_ref); */
5367 perm_dest = create_tmp_var (vectype, "vect_perm_even");
5368 DECL_GIMPLE_REG_P (perm_dest) = 1;
5369 add_referenced_var (perm_dest);
5371 tmp = build2 (VEC_EXTRACT_EVEN_EXPR, vectype,
5372 first_vect, second_vect);
5373 perm_stmt = build_gimple_modify_stmt (perm_dest, tmp);
5375 data_ref = make_ssa_name (perm_dest, perm_stmt);
5376 GIMPLE_STMT_OPERAND (perm_stmt, 0) = data_ref;
5377 vect_finish_stmt_generation (stmt, perm_stmt, bsi);
5378 mark_symbols_for_renaming (perm_stmt);
5380 VEC_replace (tree, *result_chain, j/2, data_ref);
5382 /* data_ref = permute_odd (first_data_ref, second_data_ref); */
5383 perm_dest = create_tmp_var (vectype, "vect_perm_odd");
5384 DECL_GIMPLE_REG_P (perm_dest) = 1;
5385 add_referenced_var (perm_dest);
5387 tmp = build2 (VEC_EXTRACT_ODD_EXPR, vectype,
5388 first_vect, second_vect);
5389 perm_stmt = build_gimple_modify_stmt (perm_dest, tmp);
5390 data_ref = make_ssa_name (perm_dest, perm_stmt);
5391 GIMPLE_STMT_OPERAND (perm_stmt, 0) = data_ref;
5392 vect_finish_stmt_generation (stmt, perm_stmt, bsi);
5393 mark_symbols_for_renaming (perm_stmt);
5395 VEC_replace (tree, *result_chain, j/2+length/2, data_ref);
5397 dr_chain = VEC_copy (tree, heap, *result_chain);
5403 /* Function vect_transform_strided_load.
5405 Given a chain of input interleaved data-refs (in DR_CHAIN), build statements
5406 to perform their permutation and ascribe the result vectorized statements to
5407 the scalar statements.
5411 vect_transform_strided_load (tree stmt, VEC(tree,heap) *dr_chain, int size,
5412 block_stmt_iterator *bsi)
5414 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
5415 tree first_stmt = DR_GROUP_FIRST_DR (stmt_info);
5416 tree next_stmt, new_stmt;
5417 VEC(tree,heap) *result_chain = NULL;
5418 unsigned int i, gap_count;
5421 /* DR_CHAIN contains input data-refs that are a part of the interleaving.
5422 RESULT_CHAIN is the output of vect_permute_load_chain, it contains permuted
5423 vectors, that are ready for vector computation. */
5424 result_chain = VEC_alloc (tree, heap, size);
5426 if (!vect_permute_load_chain (dr_chain, size, stmt, bsi, &result_chain))
5429 /* Put a permuted data-ref in the VECTORIZED_STMT field.
5430 Since we scan the chain starting from it's first node, their order
5431 corresponds the order of data-refs in RESULT_CHAIN. */
5432 next_stmt = first_stmt;
5434 for (i = 0; VEC_iterate (tree, result_chain, i, tmp_data_ref); i++)
5439 /* Skip the gaps. Loads created for the gaps will be removed by dead
5440 code elimination pass later.
5441 DR_GROUP_GAP is the number of steps in elements from the previous
5442 access (if there is no gap DR_GROUP_GAP is 1). We skip loads that
5443 correspond to the gaps.
5445 if (gap_count < DR_GROUP_GAP (vinfo_for_stmt (next_stmt)))
5453 new_stmt = SSA_NAME_DEF_STMT (tmp_data_ref);
5454 /* We assume that if VEC_STMT is not NULL, this is a case of multiple
5455 copies, and we put the new vector statement in the first available
5457 if (!STMT_VINFO_VEC_STMT (vinfo_for_stmt (next_stmt)))
5458 STMT_VINFO_VEC_STMT (vinfo_for_stmt (next_stmt)) = new_stmt;
5461 tree prev_stmt = STMT_VINFO_VEC_STMT (vinfo_for_stmt (next_stmt));
5462 tree rel_stmt = STMT_VINFO_RELATED_STMT (
5463 vinfo_for_stmt (prev_stmt));
5466 prev_stmt = rel_stmt;
5467 rel_stmt = STMT_VINFO_RELATED_STMT (vinfo_for_stmt (rel_stmt));
5469 STMT_VINFO_RELATED_STMT (vinfo_for_stmt (prev_stmt)) = new_stmt;
5471 next_stmt = DR_GROUP_NEXT_DR (vinfo_for_stmt (next_stmt));
5473 /* If NEXT_STMT accesses the same DR as the previous statement,
5474 put the same TMP_DATA_REF as its vectorized statement; otherwise
5475 get the next data-ref from RESULT_CHAIN. */
5476 if (!next_stmt || !DR_GROUP_SAME_DR_STMT (vinfo_for_stmt (next_stmt)))
5484 /* vectorizable_load.
5486 Check if STMT reads a non scalar data-ref (array/pointer/structure) that
5488 If VEC_STMT is also passed, vectorize the STMT: create a vectorized
5489 stmt to replace it, put it in VEC_STMT, and insert it at BSI.
5490 Return FALSE if not a vectorizable STMT, TRUE otherwise. */
5493 vectorizable_load (tree stmt, block_stmt_iterator *bsi, tree *vec_stmt,
5497 tree vec_dest = NULL;
5498 tree data_ref = NULL;
5500 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
5501 stmt_vec_info prev_stmt_info;
5502 loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info);
5503 struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
5504 struct loop *containing_loop = (bb_for_stmt (stmt))->loop_father;
5505 bool nested_in_vect_loop = nested_in_vect_loop_p (loop, stmt);
5506 struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info), *first_dr;
5507 tree vectype = STMT_VINFO_VECTYPE (stmt_info);
5510 tree new_stmt = NULL_TREE;
5512 enum dr_alignment_support alignment_support_scheme;
5513 tree dataref_ptr = NULL_TREE;
5515 int nunits = TYPE_VECTOR_SUBPARTS (vectype);
5516 int ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits;
5517 int i, j, group_size;
5518 tree msq = NULL_TREE, lsq;
5519 tree offset = NULL_TREE;
5520 tree realignment_token = NULL_TREE;
5521 tree phi = NULL_TREE;
5522 VEC(tree,heap) *dr_chain = NULL;
5523 bool strided_load = false;
5527 bool compute_in_loop = false;
5528 struct loop *at_loop;
5530 bool slp = (slp_node != NULL);
5532 /* FORNOW: SLP with multiple types is not supported. The SLP analysis verifies
5533 this, so we can safely override NCOPIES with 1 here. */
5537 gcc_assert (ncopies >= 1);
5539 /* FORNOW. This restriction should be relaxed. */
5540 if (nested_in_vect_loop && ncopies > 1)
5542 if (vect_print_dump_info (REPORT_DETAILS))
5543 fprintf (vect_dump, "multiple types in nested loop.");
5547 if (!STMT_VINFO_RELEVANT_P (stmt_info))
5550 if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_loop_def)
5553 /* Is vectorizable load? */
5554 if (TREE_CODE (stmt) != GIMPLE_MODIFY_STMT)
5557 scalar_dest = GIMPLE_STMT_OPERAND (stmt, 0);
5558 if (TREE_CODE (scalar_dest) != SSA_NAME)
5561 op = GIMPLE_STMT_OPERAND (stmt, 1);
5562 if (TREE_CODE (op) != ARRAY_REF
5563 && TREE_CODE (op) != INDIRECT_REF
5564 && !STMT_VINFO_STRIDED_ACCESS (stmt_info))
5567 if (!STMT_VINFO_DATA_REF (stmt_info))
5570 scalar_type = TREE_TYPE (DR_REF (dr));
5571 mode = (int) TYPE_MODE (vectype);
5573 /* FORNOW. In some cases can vectorize even if data-type not supported
5574 (e.g. - data copies). */
5575 if (optab_handler (mov_optab, mode)->insn_code == CODE_FOR_nothing)
5577 if (vect_print_dump_info (REPORT_DETAILS))
5578 fprintf (vect_dump, "Aligned load, but unsupported type.");
5582 /* Check if the load is a part of an interleaving chain. */
5583 if (STMT_VINFO_STRIDED_ACCESS (stmt_info))
5585 strided_load = true;
5587 gcc_assert (! nested_in_vect_loop);
5589 /* Check if interleaving is supported. */
5590 if (!vect_strided_load_supported (vectype)
5591 && !PURE_SLP_STMT (stmt_info) && !slp)
5595 if (!vec_stmt) /* transformation not required. */
5597 STMT_VINFO_TYPE (stmt_info) = load_vec_info_type;
5598 vect_model_load_cost (stmt_info, ncopies, NULL);
5602 if (vect_print_dump_info (REPORT_DETAILS))
5603 fprintf (vect_dump, "transform load.");
5609 first_stmt = DR_GROUP_FIRST_DR (stmt_info);
5610 /* Check if the chain of loads is already vectorized. */
5611 if (STMT_VINFO_VEC_STMT (vinfo_for_stmt (first_stmt)))
5613 *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info);
5616 first_dr = STMT_VINFO_DATA_REF (vinfo_for_stmt (first_stmt));
5617 group_size = DR_GROUP_SIZE (vinfo_for_stmt (first_stmt));
5618 dr_chain = VEC_alloc (tree, heap, group_size);
5620 /* VEC_NUM is the number of vect stmts to be created for this group. */
5623 strided_load = false;
5624 vec_num = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node);
5627 vec_num = group_size;
5633 group_size = vec_num = 1;
5636 alignment_support_scheme = vect_supportable_dr_alignment (first_dr);
5637 gcc_assert (alignment_support_scheme);
5639 /* In case the vectorization factor (VF) is bigger than the number
5640 of elements that we can fit in a vectype (nunits), we have to generate
5641 more than one vector stmt - i.e - we need to "unroll" the
5642 vector stmt by a factor VF/nunits. In doing so, we record a pointer
5643 from one copy of the vector stmt to the next, in the field
5644 STMT_VINFO_RELATED_STMT. This is necessary in order to allow following
5645 stages to find the correct vector defs to be used when vectorizing
5646 stmts that use the defs of the current stmt. The example below illustrates
5647 the vectorization process when VF=16 and nunits=4 (i.e - we need to create
5648 4 vectorized stmts):
5650 before vectorization:
5651 RELATED_STMT VEC_STMT
5655 step 1: vectorize stmt S1:
5656 We first create the vector stmt VS1_0, and, as usual, record a
5657 pointer to it in the STMT_VINFO_VEC_STMT of the scalar stmt S1.
5658 Next, we create the vector stmt VS1_1, and record a pointer to
5659 it in the STMT_VINFO_RELATED_STMT of the vector stmt VS1_0.
5660 Similarly, for VS1_2 and VS1_3. This is the resulting chain of
5662 RELATED_STMT VEC_STMT
5663 VS1_0: vx0 = memref0 VS1_1 -
5664 VS1_1: vx1 = memref1 VS1_2 -
5665 VS1_2: vx2 = memref2 VS1_3 -
5666 VS1_3: vx3 = memref3 - -
5667 S1: x = load - VS1_0
5670 See in documentation in vect_get_vec_def_for_stmt_copy for how the
5671 information we recorded in RELATED_STMT field is used to vectorize
5674 /* In case of interleaving (non-unit strided access):
5681 Vectorized loads are created in the order of memory accesses
5682 starting from the access of the first stmt of the chain:
5685 VS2: vx1 = &base + vec_size*1
5686 VS3: vx3 = &base + vec_size*2
5687 VS4: vx4 = &base + vec_size*3
5689 Then permutation statements are generated:
5691 VS5: vx5 = VEC_EXTRACT_EVEN_EXPR < vx0, vx1 >
5692 VS6: vx6 = VEC_EXTRACT_ODD_EXPR < vx0, vx1 >
5695 And they are put in STMT_VINFO_VEC_STMT of the corresponding scalar stmts
5696 (the order of the data-refs in the output of vect_permute_load_chain
5697 corresponds to the order of scalar stmts in the interleaving chain - see
5698 the documentation of vect_permute_load_chain()).
5699 The generation of permutation stmts and recording them in
5700 STMT_VINFO_VEC_STMT is done in vect_transform_strided_load().
5702 In case of both multiple types and interleaving, the vector loads and
5703 permutation stmts above are created for every copy. The result vector stmts
5704 are put in STMT_VINFO_VEC_STMT for the first copy and in the corresponding
5705 STMT_VINFO_RELATED_STMT for the next copies. */
5707 /* If the data reference is aligned (dr_aligned) or potentially unaligned
5708 on a target that supports unaligned accesses (dr_unaligned_supported)
5709 we generate the following code:
5713 p = p + indx * vectype_size;
5718 Otherwise, the data reference is potentially unaligned on a target that
5719 does not support unaligned accesses (dr_explicit_realign_optimized) -
5720 then generate the following code, in which the data in each iteration is
5721 obtained by two vector loads, one from the previous iteration, and one
5722 from the current iteration:
5724 msq_init = *(floor(p1))
5725 p2 = initial_addr + VS - 1;
5726 realignment_token = call target_builtin;
5729 p2 = p2 + indx * vectype_size
5731 vec_dest = realign_load (msq, lsq, realignment_token)
5736 /* If the misalignment remains the same throughout the execution of the
5737 loop, we can create the init_addr and permutation mask at the loop
5738 preheader. Otherwise, it needs to be created inside the loop.
5739 This can only occur when vectorizing memory accesses in the inner-loop
5740 nested within an outer-loop that is being vectorized. */
5742 if (nested_in_vect_loop_p (loop, stmt)
5743 && (TREE_INT_CST_LOW (DR_STEP (dr)) % UNITS_PER_SIMD_WORD != 0))
5745 gcc_assert (alignment_support_scheme != dr_explicit_realign_optimized);
5746 compute_in_loop = true;
5749 if ((alignment_support_scheme == dr_explicit_realign_optimized
5750 || alignment_support_scheme == dr_explicit_realign)
5751 && !compute_in_loop)
5753 msq = vect_setup_realignment (first_stmt, bsi, &realignment_token,
5754 alignment_support_scheme, NULL_TREE,
5756 if (alignment_support_scheme == dr_explicit_realign_optimized)
5758 phi = SSA_NAME_DEF_STMT (msq);
5759 offset = size_int (TYPE_VECTOR_SUBPARTS (vectype) - 1);
5765 prev_stmt_info = NULL;
5766 for (j = 0; j < ncopies; j++)
5768 /* 1. Create the vector pointer update chain. */
5770 dataref_ptr = vect_create_data_ref_ptr (first_stmt,
5772 &dummy, &ptr_incr, false,
5776 bump_vector_ptr (dataref_ptr, ptr_incr, bsi, stmt, NULL_TREE);
5778 for (i = 0; i < vec_num; i++)
5781 dataref_ptr = bump_vector_ptr (dataref_ptr, ptr_incr, bsi, stmt,
5784 /* 2. Create the vector-load in the loop. */
5785 switch (alignment_support_scheme)
5788 gcc_assert (aligned_access_p (first_dr));
5789 data_ref = build_fold_indirect_ref (dataref_ptr);
5791 case dr_unaligned_supported:
5793 int mis = DR_MISALIGNMENT (first_dr);
5794 tree tmis = (mis == -1 ? size_zero_node : size_int (mis));
5796 tmis = size_binop (MULT_EXPR, tmis, size_int(BITS_PER_UNIT));
5798 build2 (MISALIGNED_INDIRECT_REF, vectype, dataref_ptr, tmis);
5801 case dr_explicit_realign:
5804 tree vs_minus_1 = size_int (TYPE_VECTOR_SUBPARTS (vectype) - 1);
5806 if (compute_in_loop)
5807 msq = vect_setup_realignment (first_stmt, bsi,
5809 dr_explicit_realign,
5812 data_ref = build1 (ALIGN_INDIRECT_REF, vectype, dataref_ptr);
5813 vec_dest = vect_create_destination_var (scalar_dest, vectype);
5814 new_stmt = build_gimple_modify_stmt (vec_dest, data_ref);
5815 new_temp = make_ssa_name (vec_dest, new_stmt);
5816 GIMPLE_STMT_OPERAND (new_stmt, 0) = new_temp;
5817 vect_finish_stmt_generation (stmt, new_stmt, bsi);
5818 copy_virtual_operands (new_stmt, stmt);
5819 mark_symbols_for_renaming (new_stmt);
5822 bump = size_binop (MULT_EXPR, vs_minus_1,
5823 TYPE_SIZE_UNIT (scalar_type));
5824 ptr = bump_vector_ptr (dataref_ptr, NULL_TREE, bsi, stmt, bump);
5825 data_ref = build1 (ALIGN_INDIRECT_REF, vectype, ptr);
5828 case dr_explicit_realign_optimized:
5829 data_ref = build1 (ALIGN_INDIRECT_REF, vectype, dataref_ptr);
5834 vec_dest = vect_create_destination_var (scalar_dest, vectype);
5835 new_stmt = build_gimple_modify_stmt (vec_dest, data_ref);
5836 new_temp = make_ssa_name (vec_dest, new_stmt);
5837 GIMPLE_STMT_OPERAND (new_stmt, 0) = new_temp;
5838 vect_finish_stmt_generation (stmt, new_stmt, bsi);
5839 mark_symbols_for_renaming (new_stmt);
5841 /* 3. Handle explicit realignment if necessary/supported. Create in
5842 loop: vec_dest = realign_load (msq, lsq, realignment_token) */
5843 if (alignment_support_scheme == dr_explicit_realign_optimized
5844 || alignment_support_scheme == dr_explicit_realign)
5846 lsq = GIMPLE_STMT_OPERAND (new_stmt, 0);
5847 if (!realignment_token)
5848 realignment_token = dataref_ptr;
5849 vec_dest = vect_create_destination_var (scalar_dest, vectype);
5850 new_stmt = build3 (REALIGN_LOAD_EXPR, vectype, msq, lsq,
5852 new_stmt = build_gimple_modify_stmt (vec_dest, new_stmt);
5853 new_temp = make_ssa_name (vec_dest, new_stmt);
5854 GIMPLE_STMT_OPERAND (new_stmt, 0) = new_temp;
5855 vect_finish_stmt_generation (stmt, new_stmt, bsi);
5857 if (alignment_support_scheme == dr_explicit_realign_optimized)
5859 if (i == vec_num - 1 && j == ncopies - 1)
5860 add_phi_arg (phi, lsq, loop_latch_edge (containing_loop));
5865 /* 4. Handle invariant-load. */
5868 gcc_assert (!strided_load);
5869 gcc_assert (nested_in_vect_loop_p (loop, stmt));
5874 tree vec_inv, bitpos, bitsize = TYPE_SIZE (scalar_type);
5876 /* CHECKME: bitpos depends on endianess? */
5877 bitpos = bitsize_zero_node;
5878 vec_inv = build3 (BIT_FIELD_REF, scalar_type, new_temp,
5880 BIT_FIELD_REF_UNSIGNED (vec_inv) =
5881 TYPE_UNSIGNED (scalar_type);
5883 vect_create_destination_var (scalar_dest, NULL_TREE);
5884 new_stmt = build_gimple_modify_stmt (vec_dest, vec_inv);
5885 new_temp = make_ssa_name (vec_dest, new_stmt);
5886 GIMPLE_STMT_OPERAND (new_stmt, 0) = new_temp;
5887 vect_finish_stmt_generation (stmt, new_stmt, bsi);
5889 for (k = nunits - 1; k >= 0; --k)
5890 t = tree_cons (NULL_TREE, new_temp, t);
5891 /* FIXME: use build_constructor directly. */
5892 vec_inv = build_constructor_from_list (vectype, t);
5893 new_temp = vect_init_vector (stmt, vec_inv, vectype, bsi);
5894 new_stmt = SSA_NAME_DEF_STMT (new_temp);
5897 gcc_unreachable (); /* FORNOW. */
5900 /* Collect vector loads and later create their permutation in
5901 vect_transform_strided_load (). */
5903 VEC_quick_push (tree, dr_chain, new_temp);
5905 /* Store vector loads in the corresponding SLP_NODE. */
5907 VEC_quick_push (tree, SLP_TREE_VEC_STMTS (slp_node), new_stmt);
5910 /* FORNOW: SLP with multiple types is unsupported. */
5916 if (!vect_transform_strided_load (stmt, dr_chain, group_size, bsi))
5918 *vec_stmt = STMT_VINFO_VEC_STMT (stmt_info);
5919 dr_chain = VEC_alloc (tree, heap, group_size);
5924 STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt;
5926 STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt;
5927 prev_stmt_info = vinfo_for_stmt (new_stmt);
5935 /* Function vectorizable_live_operation.
5937 STMT computes a value that is used outside the loop. Check if
5938 it can be supported. */
5941 vectorizable_live_operation (tree stmt,
5942 block_stmt_iterator *bsi ATTRIBUTE_UNUSED,
5943 tree *vec_stmt ATTRIBUTE_UNUSED)
5946 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
5947 loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info);
5948 struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
5953 enum vect_def_type dt;
5955 gcc_assert (STMT_VINFO_LIVE_P (stmt_info));
5957 if (STMT_VINFO_DEF_TYPE (stmt_info) == vect_reduction_def)
5960 if (TREE_CODE (stmt) != GIMPLE_MODIFY_STMT)
5963 if (TREE_CODE (GIMPLE_STMT_OPERAND (stmt, 0)) != SSA_NAME)
5966 /* FORNOW. CHECKME. */
5967 if (nested_in_vect_loop_p (loop, stmt))
5970 operation = GIMPLE_STMT_OPERAND (stmt, 1);
5971 op_type = TREE_OPERAND_LENGTH (operation);
5973 /* FORNOW: support only if all uses are invariant. This means
5974 that the scalar operations can remain in place, unvectorized.
5975 The original last scalar value that they compute will be used. */
5977 for (i = 0; i < op_type; i++)
5979 op = TREE_OPERAND (operation, i);
5980 if (op && !vect_is_simple_use (op, loop_vinfo, &def_stmt, &def, &dt))
5982 if (vect_print_dump_info (REPORT_DETAILS))
5983 fprintf (vect_dump, "use not simple.");
5987 if (dt != vect_invariant_def && dt != vect_constant_def)
5991 /* No transformation is required for the cases we currently support. */
5996 /* Function vect_is_simple_cond.
5999 LOOP - the loop that is being vectorized.
6000 COND - Condition that is checked for simple use.
6002 Returns whether a COND can be vectorized. Checks whether
6003 condition operands are supportable using vec_is_simple_use. */
6006 vect_is_simple_cond (tree cond, loop_vec_info loop_vinfo)
6010 enum vect_def_type dt;
6012 if (!COMPARISON_CLASS_P (cond))
6015 lhs = TREE_OPERAND (cond, 0);
6016 rhs = TREE_OPERAND (cond, 1);
6018 if (TREE_CODE (lhs) == SSA_NAME)
6020 tree lhs_def_stmt = SSA_NAME_DEF_STMT (lhs);
6021 if (!vect_is_simple_use (lhs, loop_vinfo, &lhs_def_stmt, &def, &dt))
6024 else if (TREE_CODE (lhs) != INTEGER_CST && TREE_CODE (lhs) != REAL_CST
6025 && TREE_CODE (lhs) != FIXED_CST)
6028 if (TREE_CODE (rhs) == SSA_NAME)
6030 tree rhs_def_stmt = SSA_NAME_DEF_STMT (rhs);
6031 if (!vect_is_simple_use (rhs, loop_vinfo, &rhs_def_stmt, &def, &dt))
6034 else if (TREE_CODE (rhs) != INTEGER_CST && TREE_CODE (rhs) != REAL_CST
6035 && TREE_CODE (rhs) != FIXED_CST)
6041 /* vectorizable_condition.
6043 Check if STMT is conditional modify expression that can be vectorized.
6044 If VEC_STMT is also passed, vectorize the STMT: create a vectorized
6045 stmt using VEC_COND_EXPR to replace it, put it in VEC_STMT, and insert it
6048 Return FALSE if not a vectorizable STMT, TRUE otherwise. */
6051 vectorizable_condition (tree stmt, block_stmt_iterator *bsi, tree *vec_stmt)
6053 tree scalar_dest = NULL_TREE;
6054 tree vec_dest = NULL_TREE;
6055 tree op = NULL_TREE;
6056 tree cond_expr, then_clause, else_clause;
6057 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
6058 tree vectype = STMT_VINFO_VECTYPE (stmt_info);
6059 tree vec_cond_lhs, vec_cond_rhs, vec_then_clause, vec_else_clause;
6060 tree vec_compare, vec_cond_expr;
6062 loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info);
6063 enum machine_mode vec_mode;
6065 enum vect_def_type dt;
6066 int nunits = TYPE_VECTOR_SUBPARTS (vectype);
6067 int ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits;
6069 gcc_assert (ncopies >= 1);
6071 return false; /* FORNOW */
6073 if (!STMT_VINFO_RELEVANT_P (stmt_info))
6076 if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_loop_def)
6079 /* FORNOW: SLP not supported. */
6080 if (STMT_SLP_TYPE (stmt_info))
6083 /* FORNOW: not yet supported. */
6084 if (STMT_VINFO_LIVE_P (stmt_info))
6086 if (vect_print_dump_info (REPORT_DETAILS))
6087 fprintf (vect_dump, "value used after loop.");
6091 /* Is vectorizable conditional operation? */
6092 if (TREE_CODE (stmt) != GIMPLE_MODIFY_STMT)
6095 op = GIMPLE_STMT_OPERAND (stmt, 1);
6097 if (TREE_CODE (op) != COND_EXPR)
6100 cond_expr = TREE_OPERAND (op, 0);
6101 then_clause = TREE_OPERAND (op, 1);
6102 else_clause = TREE_OPERAND (op, 2);
6104 if (!vect_is_simple_cond (cond_expr, loop_vinfo))
6107 /* We do not handle two different vector types for the condition
6109 if (TREE_TYPE (TREE_OPERAND (cond_expr, 0)) != TREE_TYPE (vectype))
6112 if (TREE_CODE (then_clause) == SSA_NAME)
6114 tree then_def_stmt = SSA_NAME_DEF_STMT (then_clause);
6115 if (!vect_is_simple_use (then_clause, loop_vinfo,
6116 &then_def_stmt, &def, &dt))
6119 else if (TREE_CODE (then_clause) != INTEGER_CST
6120 && TREE_CODE (then_clause) != REAL_CST
6121 && TREE_CODE (then_clause) != FIXED_CST)
6124 if (TREE_CODE (else_clause) == SSA_NAME)
6126 tree else_def_stmt = SSA_NAME_DEF_STMT (else_clause);
6127 if (!vect_is_simple_use (else_clause, loop_vinfo,
6128 &else_def_stmt, &def, &dt))
6131 else if (TREE_CODE (else_clause) != INTEGER_CST
6132 && TREE_CODE (else_clause) != REAL_CST
6133 && TREE_CODE (else_clause) != FIXED_CST)
6137 vec_mode = TYPE_MODE (vectype);
6141 STMT_VINFO_TYPE (stmt_info) = condition_vec_info_type;
6142 return expand_vec_cond_expr_p (op, vec_mode);
6148 scalar_dest = GIMPLE_STMT_OPERAND (stmt, 0);
6149 vec_dest = vect_create_destination_var (scalar_dest, vectype);
6151 /* Handle cond expr. */
6153 vect_get_vec_def_for_operand (TREE_OPERAND (cond_expr, 0), stmt, NULL);
6155 vect_get_vec_def_for_operand (TREE_OPERAND (cond_expr, 1), stmt, NULL);
6156 vec_then_clause = vect_get_vec_def_for_operand (then_clause, stmt, NULL);
6157 vec_else_clause = vect_get_vec_def_for_operand (else_clause, stmt, NULL);
6159 /* Arguments are ready. create the new vector stmt. */
6160 vec_compare = build2 (TREE_CODE (cond_expr), vectype,
6161 vec_cond_lhs, vec_cond_rhs);
6162 vec_cond_expr = build3 (VEC_COND_EXPR, vectype,
6163 vec_compare, vec_then_clause, vec_else_clause);
6165 *vec_stmt = build_gimple_modify_stmt (vec_dest, vec_cond_expr);
6166 new_temp = make_ssa_name (vec_dest, *vec_stmt);
6167 GIMPLE_STMT_OPERAND (*vec_stmt, 0) = new_temp;
6168 vect_finish_stmt_generation (stmt, *vec_stmt, bsi);
6174 /* Function vect_transform_stmt.
6176 Create a vectorized stmt to replace STMT, and insert it at BSI. */
6179 vect_transform_stmt (tree stmt, block_stmt_iterator *bsi, bool *strided_store,
6182 bool is_store = false;
6183 tree vec_stmt = NULL_TREE;
6184 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
6185 tree orig_stmt_in_pattern;
6188 switch (STMT_VINFO_TYPE (stmt_info))
6190 case type_demotion_vec_info_type:
6191 gcc_assert (!slp_node);
6192 done = vectorizable_type_demotion (stmt, bsi, &vec_stmt);
6196 case type_promotion_vec_info_type:
6197 gcc_assert (!slp_node);
6198 done = vectorizable_type_promotion (stmt, bsi, &vec_stmt);
6202 case type_conversion_vec_info_type:
6203 done = vectorizable_conversion (stmt, bsi, &vec_stmt, slp_node);
6207 case induc_vec_info_type:
6208 gcc_assert (!slp_node);
6209 done = vectorizable_induction (stmt, bsi, &vec_stmt);
6213 case op_vec_info_type:
6214 done = vectorizable_operation (stmt, bsi, &vec_stmt, slp_node);
6218 case assignment_vec_info_type:
6219 done = vectorizable_assignment (stmt, bsi, &vec_stmt, slp_node);
6223 case load_vec_info_type:
6224 done = vectorizable_load (stmt, bsi, &vec_stmt, slp_node);
6228 case store_vec_info_type:
6229 done = vectorizable_store (stmt, bsi, &vec_stmt, slp_node);
6231 if (STMT_VINFO_STRIDED_ACCESS (stmt_info))
6233 /* In case of interleaving, the whole chain is vectorized when the
6234 last store in the chain is reached. Store stmts before the last
6235 one are skipped, and there vec_stmt_info shouldn't be freed
6237 *strided_store = true;
6238 if (STMT_VINFO_VEC_STMT (stmt_info))
6245 case condition_vec_info_type:
6246 gcc_assert (!slp_node);
6247 done = vectorizable_condition (stmt, bsi, &vec_stmt);
6251 case call_vec_info_type:
6252 gcc_assert (!slp_node);
6253 done = vectorizable_call (stmt, bsi, &vec_stmt);
6256 case reduc_vec_info_type:
6257 gcc_assert (!slp_node);
6258 done = vectorizable_reduction (stmt, bsi, &vec_stmt);
6263 if (!STMT_VINFO_LIVE_P (stmt_info))
6265 if (vect_print_dump_info (REPORT_DETAILS))
6266 fprintf (vect_dump, "stmt not supported.");
6271 if (STMT_VINFO_LIVE_P (stmt_info)
6272 && STMT_VINFO_TYPE (stmt_info) != reduc_vec_info_type)
6274 done = vectorizable_live_operation (stmt, bsi, &vec_stmt);
6280 STMT_VINFO_VEC_STMT (stmt_info) = vec_stmt;
6281 orig_stmt_in_pattern = STMT_VINFO_RELATED_STMT (stmt_info);
6282 if (orig_stmt_in_pattern)
6284 stmt_vec_info stmt_vinfo = vinfo_for_stmt (orig_stmt_in_pattern);
6285 /* STMT was inserted by the vectorizer to replace a computation idiom.
6286 ORIG_STMT_IN_PATTERN is a stmt in the original sequence that
6287 computed this idiom. We need to record a pointer to VEC_STMT in
6288 the stmt_info of ORIG_STMT_IN_PATTERN. See more details in the
6289 documentation of vect_pattern_recog. */
6290 if (STMT_VINFO_IN_PATTERN_P (stmt_vinfo))
6292 gcc_assert (STMT_VINFO_RELATED_STMT (stmt_vinfo) == stmt);
6293 STMT_VINFO_VEC_STMT (stmt_vinfo) = vec_stmt;
6302 /* This function builds ni_name = number of iterations loop executes
6303 on the loop preheader. */
6306 vect_build_loop_niters (loop_vec_info loop_vinfo)
6308 tree ni_name, stmt, var;
6310 struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
6311 tree ni = unshare_expr (LOOP_VINFO_NITERS (loop_vinfo));
6313 var = create_tmp_var (TREE_TYPE (ni), "niters");
6314 add_referenced_var (var);
6315 ni_name = force_gimple_operand (ni, &stmt, false, var);
6317 pe = loop_preheader_edge (loop);
6320 basic_block new_bb = bsi_insert_on_edge_immediate (pe, stmt);
6321 gcc_assert (!new_bb);
6328 /* This function generates the following statements:
6330 ni_name = number of iterations loop executes
6331 ratio = ni_name / vf
6332 ratio_mult_vf_name = ratio * vf
6334 and places them at the loop preheader edge. */
6337 vect_generate_tmps_on_preheader (loop_vec_info loop_vinfo,
6339 tree *ratio_mult_vf_name_ptr,
6340 tree *ratio_name_ptr)
6348 tree ratio_mult_vf_name;
6349 struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
6350 tree ni = LOOP_VINFO_NITERS (loop_vinfo);
6351 int vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo);
6354 pe = loop_preheader_edge (loop);
6356 /* Generate temporary variable that contains
6357 number of iterations loop executes. */
6359 ni_name = vect_build_loop_niters (loop_vinfo);
6360 log_vf = build_int_cst (TREE_TYPE (ni), exact_log2 (vf));
6362 /* Create: ratio = ni >> log2(vf) */
6364 ratio_name = fold_build2 (RSHIFT_EXPR, TREE_TYPE (ni_name), ni_name, log_vf);
6365 if (!is_gimple_val (ratio_name))
6367 var = create_tmp_var (TREE_TYPE (ni), "bnd");
6368 add_referenced_var (var);
6370 ratio_name = force_gimple_operand (ratio_name, &stmt, true, var);
6371 pe = loop_preheader_edge (loop);
6372 new_bb = bsi_insert_on_edge_immediate (pe, stmt);
6373 gcc_assert (!new_bb);
6376 /* Create: ratio_mult_vf = ratio << log2 (vf). */
6378 ratio_mult_vf_name = fold_build2 (LSHIFT_EXPR, TREE_TYPE (ratio_name),
6379 ratio_name, log_vf);
6380 if (!is_gimple_val (ratio_mult_vf_name))
6382 var = create_tmp_var (TREE_TYPE (ni), "ratio_mult_vf");
6383 add_referenced_var (var);
6385 ratio_mult_vf_name = force_gimple_operand (ratio_mult_vf_name, &stmt,
6387 pe = loop_preheader_edge (loop);
6388 new_bb = bsi_insert_on_edge_immediate (pe, stmt);
6389 gcc_assert (!new_bb);
6392 *ni_name_ptr = ni_name;
6393 *ratio_mult_vf_name_ptr = ratio_mult_vf_name;
6394 *ratio_name_ptr = ratio_name;
6400 /* Function vect_update_ivs_after_vectorizer.
6402 "Advance" the induction variables of LOOP to the value they should take
6403 after the execution of LOOP. This is currently necessary because the
6404 vectorizer does not handle induction variables that are used after the
6405 loop. Such a situation occurs when the last iterations of LOOP are
6407 1. We introduced new uses after LOOP for IVs that were not originally used
6408 after LOOP: the IVs of LOOP are now used by an epilog loop.
6409 2. LOOP is going to be vectorized; this means that it will iterate N/VF
6410 times, whereas the loop IVs should be bumped N times.
6413 - LOOP - a loop that is going to be vectorized. The last few iterations
6414 of LOOP were peeled.
6415 - NITERS - the number of iterations that LOOP executes (before it is
6416 vectorized). i.e, the number of times the ivs should be bumped.
6417 - UPDATE_E - a successor edge of LOOP->exit that is on the (only) path
6418 coming out from LOOP on which there are uses of the LOOP ivs
6419 (this is the path from LOOP->exit to epilog_loop->preheader).
6421 The new definitions of the ivs are placed in LOOP->exit.
6422 The phi args associated with the edge UPDATE_E in the bb
6423 UPDATE_E->dest are updated accordingly.
6425 Assumption 1: Like the rest of the vectorizer, this function assumes
6426 a single loop exit that has a single predecessor.
6428 Assumption 2: The phi nodes in the LOOP header and in update_bb are
6429 organized in the same order.
6431 Assumption 3: The access function of the ivs is simple enough (see
6432 vect_can_advance_ivs_p). This assumption will be relaxed in the future.
6434 Assumption 4: Exactly one of the successors of LOOP exit-bb is on a path
6435 coming out of LOOP on which the ivs of LOOP are used (this is the path
6436 that leads to the epilog loop; other paths skip the epilog loop). This
6437 path starts with the edge UPDATE_E, and its destination (denoted update_bb)
6438 needs to have its phis updated.
6442 vect_update_ivs_after_vectorizer (loop_vec_info loop_vinfo, tree niters,
6445 struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
6446 basic_block exit_bb = single_exit (loop)->dest;
6448 basic_block update_bb = update_e->dest;
6450 /* gcc_assert (vect_can_advance_ivs_p (loop_vinfo)); */
6452 /* Make sure there exists a single-predecessor exit bb: */
6453 gcc_assert (single_pred_p (exit_bb));
6455 for (phi = phi_nodes (loop->header), phi1 = phi_nodes (update_bb);
6457 phi = PHI_CHAIN (phi), phi1 = PHI_CHAIN (phi1))
6459 tree access_fn = NULL;
6460 tree evolution_part;
6463 tree var, ni, ni_name;
6464 block_stmt_iterator last_bsi;
6466 if (vect_print_dump_info (REPORT_DETAILS))
6468 fprintf (vect_dump, "vect_update_ivs_after_vectorizer: phi: ");
6469 print_generic_expr (vect_dump, phi, TDF_SLIM);
6472 /* Skip virtual phi's. */
6473 if (!is_gimple_reg (SSA_NAME_VAR (PHI_RESULT (phi))))
6475 if (vect_print_dump_info (REPORT_DETAILS))
6476 fprintf (vect_dump, "virtual phi. skip.");
6480 /* Skip reduction phis. */
6481 if (STMT_VINFO_DEF_TYPE (vinfo_for_stmt (phi)) == vect_reduction_def)
6483 if (vect_print_dump_info (REPORT_DETAILS))
6484 fprintf (vect_dump, "reduc phi. skip.");
6488 access_fn = analyze_scalar_evolution (loop, PHI_RESULT (phi));
6489 gcc_assert (access_fn);
6491 unshare_expr (evolution_part_in_loop_num (access_fn, loop->num));
6492 gcc_assert (evolution_part != NULL_TREE);
6494 /* FORNOW: We do not support IVs whose evolution function is a polynomial
6495 of degree >= 2 or exponential. */
6496 gcc_assert (!tree_is_chrec (evolution_part));
6498 step_expr = evolution_part;
6499 init_expr = unshare_expr (initial_condition_in_loop_num (access_fn,
6502 if (POINTER_TYPE_P (TREE_TYPE (init_expr)))
6503 ni = fold_build2 (POINTER_PLUS_EXPR, TREE_TYPE (init_expr),
6505 fold_convert (sizetype,
6506 fold_build2 (MULT_EXPR, TREE_TYPE (niters),
6507 niters, step_expr)));
6509 ni = fold_build2 (PLUS_EXPR, TREE_TYPE (init_expr),
6510 fold_build2 (MULT_EXPR, TREE_TYPE (init_expr),
6511 fold_convert (TREE_TYPE (init_expr),
6518 var = create_tmp_var (TREE_TYPE (init_expr), "tmp");
6519 add_referenced_var (var);
6521 last_bsi = bsi_last (exit_bb);
6522 ni_name = force_gimple_operand_bsi (&last_bsi, ni, false, var,
6523 true, BSI_SAME_STMT);
6525 /* Fix phi expressions in the successor bb. */
6526 SET_PHI_ARG_DEF (phi1, update_e->dest_idx, ni_name);
6530 /* Return the more conservative threshold between the
6531 min_profitable_iters returned by the cost model and the user
6532 specified threshold, if provided. */
6535 conservative_cost_threshold (loop_vec_info loop_vinfo,
6536 int min_profitable_iters)
6539 int min_scalar_loop_bound;
6541 min_scalar_loop_bound = ((PARAM_VALUE (PARAM_MIN_VECT_LOOP_BOUND)
6542 * LOOP_VINFO_VECT_FACTOR (loop_vinfo)) - 1);
6544 /* Use the cost model only if it is more conservative than user specified
6546 th = (unsigned) min_scalar_loop_bound;
6547 if (min_profitable_iters
6548 && (!min_scalar_loop_bound
6549 || min_profitable_iters > min_scalar_loop_bound))
6550 th = (unsigned) min_profitable_iters;
6552 if (th && vect_print_dump_info (REPORT_COST))
6553 fprintf (vect_dump, "Vectorization may not be profitable.");
6558 /* Function vect_do_peeling_for_loop_bound
6560 Peel the last iterations of the loop represented by LOOP_VINFO.
6561 The peeled iterations form a new epilog loop. Given that the loop now
6562 iterates NITERS times, the new epilog loop iterates
6563 NITERS % VECTORIZATION_FACTOR times.
6565 The original loop will later be made to iterate
6566 NITERS / VECTORIZATION_FACTOR times (this value is placed into RATIO). */
6569 vect_do_peeling_for_loop_bound (loop_vec_info loop_vinfo, tree *ratio)
6571 tree ni_name, ratio_mult_vf_name;
6572 struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
6573 struct loop *new_loop;
6575 basic_block preheader;
6577 bool check_profitability = false;
6578 unsigned int th = 0;
6579 int min_profitable_iters;
6581 if (vect_print_dump_info (REPORT_DETAILS))
6582 fprintf (vect_dump, "=== vect_do_peeling_for_loop_bound ===");
6584 initialize_original_copy_tables ();
6586 /* Generate the following variables on the preheader of original loop:
6588 ni_name = number of iteration the original loop executes
6589 ratio = ni_name / vf
6590 ratio_mult_vf_name = ratio * vf */
6591 vect_generate_tmps_on_preheader (loop_vinfo, &ni_name,
6592 &ratio_mult_vf_name, ratio);
6594 loop_num = loop->num;
6596 /* If cost model check not done during versioning and
6597 peeling for alignment. */
6598 if (!VEC_length (tree, LOOP_VINFO_MAY_MISALIGN_STMTS (loop_vinfo))
6599 && !VEC_length (ddr_p, LOOP_VINFO_MAY_ALIAS_DDRS (loop_vinfo))
6600 && !LOOP_PEELING_FOR_ALIGNMENT (loop_vinfo))
6602 check_profitability = true;
6604 /* Get profitability threshold for vectorized loop. */
6605 min_profitable_iters = LOOP_VINFO_COST_MODEL_MIN_ITERS (loop_vinfo);
6607 th = conservative_cost_threshold (loop_vinfo,
6608 min_profitable_iters);
6611 new_loop = slpeel_tree_peel_loop_to_edge (loop, single_exit (loop),
6612 ratio_mult_vf_name, ni_name, false,
6613 th, check_profitability);
6614 gcc_assert (new_loop);
6615 gcc_assert (loop_num == loop->num);
6616 #ifdef ENABLE_CHECKING
6617 slpeel_verify_cfg_after_peeling (loop, new_loop);
6620 /* A guard that controls whether the new_loop is to be executed or skipped
6621 is placed in LOOP->exit. LOOP->exit therefore has two successors - one
6622 is the preheader of NEW_LOOP, where the IVs from LOOP are used. The other
6623 is a bb after NEW_LOOP, where these IVs are not used. Find the edge that
6624 is on the path where the LOOP IVs are used and need to be updated. */
6626 preheader = loop_preheader_edge (new_loop)->src;
6627 if (EDGE_PRED (preheader, 0)->src == single_exit (loop)->dest)
6628 update_e = EDGE_PRED (preheader, 0);
6630 update_e = EDGE_PRED (preheader, 1);
6632 /* Update IVs of original loop as if they were advanced
6633 by ratio_mult_vf_name steps. */
6634 vect_update_ivs_after_vectorizer (loop_vinfo, ratio_mult_vf_name, update_e);
6636 /* After peeling we have to reset scalar evolution analyzer. */
6639 free_original_copy_tables ();
6643 /* Function vect_gen_niters_for_prolog_loop
6645 Set the number of iterations for the loop represented by LOOP_VINFO
6646 to the minimum between LOOP_NITERS (the original iteration count of the loop)
6647 and the misalignment of DR - the data reference recorded in
6648 LOOP_VINFO_UNALIGNED_DR (LOOP_VINFO). As a result, after the execution of
6649 this loop, the data reference DR will refer to an aligned location.
6651 The following computation is generated:
6653 If the misalignment of DR is known at compile time:
6654 addr_mis = int mis = DR_MISALIGNMENT (dr);
6655 Else, compute address misalignment in bytes:
6656 addr_mis = addr & (vectype_size - 1)
6658 prolog_niters = min ( LOOP_NITERS , (VF - addr_mis/elem_size)&(VF-1) )
6660 (elem_size = element type size; an element is the scalar element
6661 whose type is the inner type of the vectype)
6665 prolog_niters = min ( LOOP_NITERS ,
6666 (VF/group_size - addr_mis/elem_size)&(VF/group_size-1) )
6667 where group_size is the size of the interleaved group.
6669 The above formulas assume that VF == number of elements in the vector. This
6670 may not hold when there are multiple-types in the loop.
6671 In this case, for some data-references in the loop the VF does not represent
6672 the number of elements that fit in the vector. Therefore, instead of VF we
6673 use TYPE_VECTOR_SUBPARTS. */
6676 vect_gen_niters_for_prolog_loop (loop_vec_info loop_vinfo, tree loop_niters)
6678 struct data_reference *dr = LOOP_VINFO_UNALIGNED_DR (loop_vinfo);
6679 struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
6681 tree iters, iters_name;
6684 tree dr_stmt = DR_STMT (dr);
6685 stmt_vec_info stmt_info = vinfo_for_stmt (dr_stmt);
6686 tree vectype = STMT_VINFO_VECTYPE (stmt_info);
6687 int vectype_align = TYPE_ALIGN (vectype) / BITS_PER_UNIT;
6688 tree niters_type = TREE_TYPE (loop_niters);
6690 int element_size = GET_MODE_SIZE (TYPE_MODE (TREE_TYPE (DR_REF (dr))));
6691 int nelements = TYPE_VECTOR_SUBPARTS (vectype);
6693 if (STMT_VINFO_STRIDED_ACCESS (stmt_info))
6695 /* For interleaved access element size must be multiplied by the size of
6696 the interleaved group. */
6697 group_size = DR_GROUP_SIZE (vinfo_for_stmt (
6698 DR_GROUP_FIRST_DR (stmt_info)));
6699 element_size *= group_size;
6702 pe = loop_preheader_edge (loop);
6704 if (LOOP_PEELING_FOR_ALIGNMENT (loop_vinfo) > 0)
6706 int byte_misalign = LOOP_PEELING_FOR_ALIGNMENT (loop_vinfo);
6707 int elem_misalign = byte_misalign / element_size;
6709 if (vect_print_dump_info (REPORT_DETAILS))
6710 fprintf (vect_dump, "known alignment = %d.", byte_misalign);
6711 iters = build_int_cst (niters_type,
6712 (nelements - elem_misalign)&(nelements/group_size-1));
6716 tree new_stmts = NULL_TREE;
6717 tree start_addr = vect_create_addr_base_for_vector_ref (dr_stmt,
6718 &new_stmts, NULL_TREE, loop);
6719 tree ptr_type = TREE_TYPE (start_addr);
6720 tree size = TYPE_SIZE (ptr_type);
6721 tree type = lang_hooks.types.type_for_size (tree_low_cst (size, 1), 1);
6722 tree vectype_size_minus_1 = build_int_cst (type, vectype_align - 1);
6723 tree elem_size_log =
6724 build_int_cst (type, exact_log2 (vectype_align/nelements));
6725 tree nelements_minus_1 = build_int_cst (type, nelements - 1);
6726 tree nelements_tree = build_int_cst (type, nelements);
6730 new_bb = bsi_insert_on_edge_immediate (pe, new_stmts);
6731 gcc_assert (!new_bb);
6733 /* Create: byte_misalign = addr & (vectype_size - 1) */
6735 fold_build2 (BIT_AND_EXPR, type, fold_convert (type, start_addr), vectype_size_minus_1);
6737 /* Create: elem_misalign = byte_misalign / element_size */
6739 fold_build2 (RSHIFT_EXPR, type, byte_misalign, elem_size_log);
6741 /* Create: (niters_type) (nelements - elem_misalign)&(nelements - 1) */
6742 iters = fold_build2 (MINUS_EXPR, type, nelements_tree, elem_misalign);
6743 iters = fold_build2 (BIT_AND_EXPR, type, iters, nelements_minus_1);
6744 iters = fold_convert (niters_type, iters);
6747 /* Create: prolog_loop_niters = min (iters, loop_niters) */
6748 /* If the loop bound is known at compile time we already verified that it is
6749 greater than vf; since the misalignment ('iters') is at most vf, there's
6750 no need to generate the MIN_EXPR in this case. */
6751 if (TREE_CODE (loop_niters) != INTEGER_CST)
6752 iters = fold_build2 (MIN_EXPR, niters_type, iters, loop_niters);
6754 if (vect_print_dump_info (REPORT_DETAILS))
6756 fprintf (vect_dump, "niters for prolog loop: ");
6757 print_generic_expr (vect_dump, iters, TDF_SLIM);
6760 var = create_tmp_var (niters_type, "prolog_loop_niters");
6761 add_referenced_var (var);
6762 iters_name = force_gimple_operand (iters, &stmt, false, var);
6764 /* Insert stmt on loop preheader edge. */
6767 basic_block new_bb = bsi_insert_on_edge_immediate (pe, stmt);
6768 gcc_assert (!new_bb);
6775 /* Function vect_update_init_of_dr
6777 NITERS iterations were peeled from LOOP. DR represents a data reference
6778 in LOOP. This function updates the information recorded in DR to
6779 account for the fact that the first NITERS iterations had already been
6780 executed. Specifically, it updates the OFFSET field of DR. */
6783 vect_update_init_of_dr (struct data_reference *dr, tree niters)
6785 tree offset = DR_OFFSET (dr);
6787 niters = fold_build2 (MULT_EXPR, TREE_TYPE (niters), niters, DR_STEP (dr));
6788 offset = fold_build2 (PLUS_EXPR, TREE_TYPE (offset), offset, niters);
6789 DR_OFFSET (dr) = offset;
6793 /* Function vect_update_inits_of_drs
6795 NITERS iterations were peeled from the loop represented by LOOP_VINFO.
6796 This function updates the information recorded for the data references in
6797 the loop to account for the fact that the first NITERS iterations had
6798 already been executed. Specifically, it updates the initial_condition of
6799 the access_function of all the data_references in the loop. */
6802 vect_update_inits_of_drs (loop_vec_info loop_vinfo, tree niters)
6805 VEC (data_reference_p, heap) *datarefs = LOOP_VINFO_DATAREFS (loop_vinfo);
6806 struct data_reference *dr;
6808 if (vect_print_dump_info (REPORT_DETAILS))
6809 fprintf (vect_dump, "=== vect_update_inits_of_dr ===");
6811 for (i = 0; VEC_iterate (data_reference_p, datarefs, i, dr); i++)
6812 vect_update_init_of_dr (dr, niters);
6816 /* Function vect_do_peeling_for_alignment
6818 Peel the first 'niters' iterations of the loop represented by LOOP_VINFO.
6819 'niters' is set to the misalignment of one of the data references in the
6820 loop, thereby forcing it to refer to an aligned location at the beginning
6821 of the execution of this loop. The data reference for which we are
6822 peeling is recorded in LOOP_VINFO_UNALIGNED_DR. */
6825 vect_do_peeling_for_alignment (loop_vec_info loop_vinfo)
6827 struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
6828 tree niters_of_prolog_loop, ni_name;
6830 struct loop *new_loop;
6831 bool check_profitability = false;
6832 unsigned int th = 0;
6833 int min_profitable_iters;
6835 if (vect_print_dump_info (REPORT_DETAILS))
6836 fprintf (vect_dump, "=== vect_do_peeling_for_alignment ===");
6838 initialize_original_copy_tables ();
6840 ni_name = vect_build_loop_niters (loop_vinfo);
6841 niters_of_prolog_loop = vect_gen_niters_for_prolog_loop (loop_vinfo, ni_name);
6844 /* If cost model check not done during versioning. */
6845 if (!VEC_length (tree, LOOP_VINFO_MAY_MISALIGN_STMTS (loop_vinfo))
6846 && !VEC_length (ddr_p, LOOP_VINFO_MAY_ALIAS_DDRS (loop_vinfo)))
6848 check_profitability = true;
6850 /* Get profitability threshold for vectorized loop. */
6851 min_profitable_iters = LOOP_VINFO_COST_MODEL_MIN_ITERS (loop_vinfo);
6853 th = conservative_cost_threshold (loop_vinfo,
6854 min_profitable_iters);
6857 /* Peel the prolog loop and iterate it niters_of_prolog_loop. */
6859 slpeel_tree_peel_loop_to_edge (loop, loop_preheader_edge (loop),
6860 niters_of_prolog_loop, ni_name, true,
6861 th, check_profitability);
6863 gcc_assert (new_loop);
6864 #ifdef ENABLE_CHECKING
6865 slpeel_verify_cfg_after_peeling (new_loop, loop);
6868 /* Update number of times loop executes. */
6869 n_iters = LOOP_VINFO_NITERS (loop_vinfo);
6870 LOOP_VINFO_NITERS (loop_vinfo) = fold_build2 (MINUS_EXPR,
6871 TREE_TYPE (n_iters), n_iters, niters_of_prolog_loop);
6873 /* Update the init conditions of the access functions of all data refs. */
6874 vect_update_inits_of_drs (loop_vinfo, niters_of_prolog_loop);
6876 /* After peeling we have to reset scalar evolution analyzer. */
6879 free_original_copy_tables ();
6883 /* Function vect_create_cond_for_align_checks.
6885 Create a conditional expression that represents the alignment checks for
6886 all of data references (array element references) whose alignment must be
6890 COND_EXPR - input conditional expression. New conditions will be chained
6891 with logical AND operation.
6892 LOOP_VINFO - two fields of the loop information are used.
6893 LOOP_VINFO_PTR_MASK is the mask used to check the alignment.
6894 LOOP_VINFO_MAY_MISALIGN_STMTS contains the refs to be checked.
6897 COND_EXPR_STMT_LIST - statements needed to construct the conditional
6899 The returned value is the conditional expression to be used in the if
6900 statement that controls which version of the loop gets executed at runtime.
6902 The algorithm makes two assumptions:
6903 1) The number of bytes "n" in a vector is a power of 2.
6904 2) An address "a" is aligned if a%n is zero and that this
6905 test can be done as a&(n-1) == 0. For example, for 16
6906 byte vectors the test is a&0xf == 0. */
6909 vect_create_cond_for_align_checks (loop_vec_info loop_vinfo,
6911 tree *cond_expr_stmt_list)
6913 struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
6914 VEC(tree,heap) *may_misalign_stmts
6915 = LOOP_VINFO_MAY_MISALIGN_STMTS (loop_vinfo);
6917 int mask = LOOP_VINFO_PTR_MASK (loop_vinfo);
6921 tree int_ptrsize_type;
6923 tree or_tmp_name = NULL_TREE;
6924 tree and_tmp, and_tmp_name, and_stmt;
6926 tree part_cond_expr;
6928 /* Check that mask is one less than a power of 2, i.e., mask is
6929 all zeros followed by all ones. */
6930 gcc_assert ((mask != 0) && ((mask & (mask+1)) == 0));
6932 /* CHECKME: what is the best integer or unsigned type to use to hold a
6933 cast from a pointer value? */
6934 psize = TYPE_SIZE (ptr_type_node);
6936 = lang_hooks.types.type_for_size (tree_low_cst (psize, 1), 0);
6938 /* Create expression (mask & (dr_1 || ... || dr_n)) where dr_i is the address
6939 of the first vector of the i'th data reference. */
6941 for (i = 0; VEC_iterate (tree, may_misalign_stmts, i, ref_stmt); i++)
6943 tree new_stmt_list = NULL_TREE;
6945 tree addr_tmp, addr_tmp_name, addr_stmt;
6946 tree or_tmp, new_or_tmp_name, or_stmt;
6948 /* create: addr_tmp = (int)(address_of_first_vector) */
6949 addr_base = vect_create_addr_base_for_vector_ref (ref_stmt,
6950 &new_stmt_list, NULL_TREE, loop);
6952 if (new_stmt_list != NULL_TREE)
6953 append_to_statement_list_force (new_stmt_list, cond_expr_stmt_list);
6955 sprintf (tmp_name, "%s%d", "addr2int", i);
6956 addr_tmp = create_tmp_var (int_ptrsize_type, tmp_name);
6957 add_referenced_var (addr_tmp);
6958 addr_tmp_name = make_ssa_name (addr_tmp, NULL_TREE);
6959 addr_stmt = fold_convert (int_ptrsize_type, addr_base);
6960 addr_stmt = build_gimple_modify_stmt (addr_tmp_name, addr_stmt);
6961 SSA_NAME_DEF_STMT (addr_tmp_name) = addr_stmt;
6962 append_to_statement_list_force (addr_stmt, cond_expr_stmt_list);
6964 /* The addresses are OR together. */
6966 if (or_tmp_name != NULL_TREE)
6968 /* create: or_tmp = or_tmp | addr_tmp */
6969 sprintf (tmp_name, "%s%d", "orptrs", i);
6970 or_tmp = create_tmp_var (int_ptrsize_type, tmp_name);
6971 add_referenced_var (or_tmp);
6972 new_or_tmp_name = make_ssa_name (or_tmp, NULL_TREE);
6973 tmp = build2 (BIT_IOR_EXPR, int_ptrsize_type,
6974 or_tmp_name, addr_tmp_name);
6975 or_stmt = build_gimple_modify_stmt (new_or_tmp_name, tmp);
6976 SSA_NAME_DEF_STMT (new_or_tmp_name) = or_stmt;
6977 append_to_statement_list_force (or_stmt, cond_expr_stmt_list);
6978 or_tmp_name = new_or_tmp_name;
6981 or_tmp_name = addr_tmp_name;
6985 mask_cst = build_int_cst (int_ptrsize_type, mask);
6987 /* create: and_tmp = or_tmp & mask */
6988 and_tmp = create_tmp_var (int_ptrsize_type, "andmask" );
6989 add_referenced_var (and_tmp);
6990 and_tmp_name = make_ssa_name (and_tmp, NULL_TREE);
6992 tmp = build2 (BIT_AND_EXPR, int_ptrsize_type, or_tmp_name, mask_cst);
6993 and_stmt = build_gimple_modify_stmt (and_tmp_name, tmp);
6994 SSA_NAME_DEF_STMT (and_tmp_name) = and_stmt;
6995 append_to_statement_list_force (and_stmt, cond_expr_stmt_list);
6997 /* Make and_tmp the left operand of the conditional test against zero.
6998 if and_tmp has a nonzero bit then some address is unaligned. */
6999 ptrsize_zero = build_int_cst (int_ptrsize_type, 0);
7000 part_cond_expr = fold_build2 (EQ_EXPR, boolean_type_node,
7001 and_tmp_name, ptrsize_zero);
7003 *cond_expr = fold_build2 (TRUTH_AND_EXPR, boolean_type_node,
7004 *cond_expr, part_cond_expr);
7006 *cond_expr = part_cond_expr;
7009 /* Function vect_vfa_segment_size.
7011 Create an expression that computes the size of segment
7012 that will be accessed for a data reference. The functions takes into
7013 account that realignment loads may access one more vector.
7016 DR: The data reference.
7017 VECT_FACTOR: vectorization factor.
7019 Return an expression whose value is the size of segment which will be
7023 vect_vfa_segment_size (struct data_reference *dr, tree vect_factor)
7025 tree segment_length = fold_build2 (MULT_EXPR, integer_type_node,
7026 DR_STEP (dr), vect_factor);
7028 if (vect_supportable_dr_alignment (dr) == dr_explicit_realign_optimized)
7030 tree vector_size = TYPE_SIZE_UNIT
7031 (STMT_VINFO_VECTYPE (vinfo_for_stmt (DR_STMT (dr))));
7033 segment_length = fold_build2 (PLUS_EXPR, integer_type_node,
7034 segment_length, vector_size);
7036 return fold_convert (sizetype, segment_length);
7039 /* Function vect_create_cond_for_alias_checks.
7041 Create a conditional expression that represents the run-time checks for
7042 overlapping of address ranges represented by a list of data references
7043 relations passed as input.
7046 COND_EXPR - input conditional expression. New conditions will be chained
7047 with logical AND operation.
7048 LOOP_VINFO - field LOOP_VINFO_MAY_ALIAS_STMTS contains the list of ddrs
7052 COND_EXPR - conditional expression.
7053 COND_EXPR_STMT_LIST - statements needed to construct the conditional
7057 The returned value is the conditional expression to be used in the if
7058 statement that controls which version of the loop gets executed at runtime.
7062 vect_create_cond_for_alias_checks (loop_vec_info loop_vinfo,
7064 tree * cond_expr_stmt_list)
7066 struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
7067 VEC (ddr_p, heap) * may_alias_ddrs =
7068 LOOP_VINFO_MAY_ALIAS_DDRS (loop_vinfo);
7070 build_int_cst (integer_type_node, LOOP_VINFO_VECT_FACTOR (loop_vinfo));
7074 tree part_cond_expr;
7076 /* Create expression
7077 ((store_ptr_0 + store_segment_length_0) < load_ptr_0)
7078 || (load_ptr_0 + load_segment_length_0) < store_ptr_0))
7082 ((store_ptr_n + store_segment_length_n) < load_ptr_n)
7083 || (load_ptr_n + load_segment_length_n) < store_ptr_n)) */
7085 if (VEC_empty (ddr_p, may_alias_ddrs))
7088 for (i = 0; VEC_iterate (ddr_p, may_alias_ddrs, i, ddr); i++)
7090 struct data_reference *dr_a, *dr_b;
7091 tree dr_group_first_a, dr_group_first_b;
7092 tree addr_base_a, addr_base_b;
7093 tree segment_length_a, segment_length_b;
7094 tree stmt_a, stmt_b;
7097 stmt_a = DR_STMT (DDR_A (ddr));
7098 dr_group_first_a = DR_GROUP_FIRST_DR (vinfo_for_stmt (stmt_a));
7099 if (dr_group_first_a)
7101 stmt_a = dr_group_first_a;
7102 dr_a = STMT_VINFO_DATA_REF (vinfo_for_stmt (stmt_a));
7106 stmt_b = DR_STMT (DDR_B (ddr));
7107 dr_group_first_b = DR_GROUP_FIRST_DR (vinfo_for_stmt (stmt_b));
7108 if (dr_group_first_b)
7110 stmt_b = dr_group_first_b;
7111 dr_b = STMT_VINFO_DATA_REF (vinfo_for_stmt (stmt_b));
7115 vect_create_addr_base_for_vector_ref (stmt_a, cond_expr_stmt_list,
7118 vect_create_addr_base_for_vector_ref (stmt_b, cond_expr_stmt_list,
7121 segment_length_a = vect_vfa_segment_size (dr_a, vect_factor);
7122 segment_length_b = vect_vfa_segment_size (dr_b, vect_factor);
7124 if (vect_print_dump_info (REPORT_DR_DETAILS))
7127 "create runtime check for data references ");
7128 print_generic_expr (vect_dump, DR_REF (dr_a), TDF_SLIM);
7129 fprintf (vect_dump, " and ");
7130 print_generic_expr (vect_dump, DR_REF (dr_b), TDF_SLIM);
7135 fold_build2 (TRUTH_OR_EXPR, boolean_type_node,
7136 fold_build2 (LT_EXPR, boolean_type_node,
7137 fold_build2 (POINTER_PLUS_EXPR, TREE_TYPE (addr_base_a),
7141 fold_build2 (LT_EXPR, boolean_type_node,
7142 fold_build2 (POINTER_PLUS_EXPR, TREE_TYPE (addr_base_b),
7148 *cond_expr = fold_build2 (TRUTH_AND_EXPR, boolean_type_node,
7149 *cond_expr, part_cond_expr);
7151 *cond_expr = part_cond_expr;
7153 if (vect_print_dump_info (REPORT_VECTORIZED_LOOPS))
7154 fprintf (vect_dump, "created %u versioning for alias checks.\n",
7155 VEC_length (ddr_p, may_alias_ddrs));
7159 /* Function vect_loop_versioning.
7161 If the loop has data references that may or may not be aligned or/and
7162 has data reference relations whose independence was not proven then
7163 two versions of the loop need to be generated, one which is vectorized
7164 and one which isn't. A test is then generated to control which of the
7165 loops is executed. The test checks for the alignment of all of the
7166 data references that may or may not be aligned. An additional
7167 sequence of runtime tests is generated for each pairs of DDRs whose
7168 independence was not proven. The vectorized version of loop is
7169 executed only if both alias and alignment tests are passed.
7171 The test generated to check which version of loop is executed
7172 is modified to also check for profitability as indicated by the
7173 cost model initially. */
7176 vect_loop_versioning (loop_vec_info loop_vinfo)
7178 struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
7180 tree cond_expr = NULL_TREE;
7181 tree cond_expr_stmt_list = NULL_TREE;
7182 basic_block condition_bb;
7183 block_stmt_iterator cond_exp_bsi;
7184 basic_block merge_bb;
7185 basic_block new_exit_bb;
7187 tree orig_phi, new_phi, arg;
7188 unsigned prob = 4 * REG_BR_PROB_BASE / 5;
7189 tree gimplify_stmt_list;
7190 tree scalar_loop_iters = LOOP_VINFO_NITERS (loop_vinfo);
7191 int min_profitable_iters = 0;
7194 /* Get profitability threshold for vectorized loop. */
7195 min_profitable_iters = LOOP_VINFO_COST_MODEL_MIN_ITERS (loop_vinfo);
7197 th = conservative_cost_threshold (loop_vinfo,
7198 min_profitable_iters);
7201 build2 (GT_EXPR, boolean_type_node, scalar_loop_iters,
7202 build_int_cst (TREE_TYPE (scalar_loop_iters), th));
7204 cond_expr = force_gimple_operand (cond_expr, &cond_expr_stmt_list,
7207 if (VEC_length (tree, LOOP_VINFO_MAY_MISALIGN_STMTS (loop_vinfo)))
7208 vect_create_cond_for_align_checks (loop_vinfo, &cond_expr,
7209 &cond_expr_stmt_list);
7211 if (VEC_length (ddr_p, LOOP_VINFO_MAY_ALIAS_DDRS (loop_vinfo)))
7212 vect_create_cond_for_alias_checks (loop_vinfo, &cond_expr,
7213 &cond_expr_stmt_list);
7216 fold_build2 (NE_EXPR, boolean_type_node, cond_expr, integer_zero_node);
7218 force_gimple_operand (cond_expr, &gimplify_stmt_list, true,
7220 append_to_statement_list (gimplify_stmt_list, &cond_expr_stmt_list);
7222 initialize_original_copy_tables ();
7223 nloop = loop_version (loop, cond_expr, &condition_bb,
7224 prob, prob, REG_BR_PROB_BASE - prob, true);
7225 free_original_copy_tables();
7227 /* Loop versioning violates an assumption we try to maintain during
7228 vectorization - that the loop exit block has a single predecessor.
7229 After versioning, the exit block of both loop versions is the same
7230 basic block (i.e. it has two predecessors). Just in order to simplify
7231 following transformations in the vectorizer, we fix this situation
7232 here by adding a new (empty) block on the exit-edge of the loop,
7233 with the proper loop-exit phis to maintain loop-closed-form. */
7235 merge_bb = single_exit (loop)->dest;
7236 gcc_assert (EDGE_COUNT (merge_bb->preds) == 2);
7237 new_exit_bb = split_edge (single_exit (loop));
7238 new_exit_e = single_exit (loop);
7239 e = EDGE_SUCC (new_exit_bb, 0);
7241 for (orig_phi = phi_nodes (merge_bb); orig_phi;
7242 orig_phi = PHI_CHAIN (orig_phi))
7244 new_phi = create_phi_node (SSA_NAME_VAR (PHI_RESULT (orig_phi)),
7246 arg = PHI_ARG_DEF_FROM_EDGE (orig_phi, e);
7247 add_phi_arg (new_phi, arg, new_exit_e);
7248 SET_PHI_ARG_DEF (orig_phi, e->dest_idx, PHI_RESULT (new_phi));
7251 /* End loop-exit-fixes after versioning. */
7253 update_ssa (TODO_update_ssa);
7254 if (cond_expr_stmt_list)
7256 cond_exp_bsi = bsi_last (condition_bb);
7257 bsi_insert_before (&cond_exp_bsi, cond_expr_stmt_list, BSI_SAME_STMT);
7261 /* Remove a group of stores (for SLP or interleaving), free their
7265 vect_remove_stores (tree first_stmt)
7268 tree next = first_stmt;
7270 stmt_vec_info next_stmt_info;
7271 block_stmt_iterator next_si;
7275 /* Free the attached stmt_vec_info and remove the stmt. */
7276 next_si = bsi_for_stmt (next);
7277 bsi_remove (&next_si, true);
7278 next_stmt_info = vinfo_for_stmt (next);
7279 ann = stmt_ann (next);
7280 tmp = DR_GROUP_NEXT_DR (next_stmt_info);
7281 free (next_stmt_info);
7282 set_stmt_info (ann, NULL);
7288 /* Vectorize SLP instance tree in postorder. */
7291 vect_schedule_slp_instance (slp_tree node, unsigned int vec_stmts_size)
7294 bool strided_store, is_store;
7295 block_stmt_iterator si;
7296 stmt_vec_info stmt_info;
7301 vect_schedule_slp_instance (SLP_TREE_LEFT (node), vec_stmts_size);
7302 vect_schedule_slp_instance (SLP_TREE_RIGHT (node), vec_stmts_size);
7304 stmt = VEC_index(tree, SLP_TREE_SCALAR_STMTS (node), 0);
7305 stmt_info = vinfo_for_stmt (stmt);
7306 SLP_TREE_VEC_STMTS (node) = VEC_alloc (tree, heap, vec_stmts_size);
7307 SLP_TREE_NUMBER_OF_VEC_STMTS (node) = vec_stmts_size;
7309 if (vect_print_dump_info (REPORT_DETAILS))
7311 fprintf (vect_dump, "------>vectorizing SLP node starting from: ");
7312 print_generic_expr (vect_dump, stmt, TDF_SLIM);
7315 si = bsi_for_stmt (stmt);
7316 is_store = vect_transform_stmt (stmt, &si, &strided_store, node);
7319 if (DR_GROUP_FIRST_DR (stmt_info))
7320 /* If IS_STORE is TRUE, the vectorization of the
7321 interleaving chain was completed - free all the stores in
7323 vect_remove_stores (DR_GROUP_FIRST_DR (stmt_info));
7325 /* FORNOW: SLP originates only from strided stores. */
7331 /* FORNOW: SLP originates only from strided stores. */
7337 vect_schedule_slp (loop_vec_info loop_vinfo, unsigned int nunits)
7339 VEC (slp_instance, heap) *slp_instances =
7340 LOOP_VINFO_SLP_INSTANCES (loop_vinfo);
7341 slp_instance instance;
7342 unsigned int vec_stmts_size;
7343 unsigned int group_size, i;
7344 unsigned int vectorization_factor = LOOP_VINFO_VECT_FACTOR (loop_vinfo);
7345 bool is_store = false;
7347 for (i = 0; VEC_iterate (slp_instance, slp_instances, i, instance); i++)
7349 group_size = SLP_INSTANCE_GROUP_SIZE (instance);
7350 /* For each SLP instance calculate number of vector stmts to be created
7351 for the scalar stmts in each node of the SLP tree. Number of vector
7352 elements in one vector iteration is the number of scalar elements in
7353 one scalar iteration (GROUP_SIZE) multiplied by VF divided by vector
7355 vec_stmts_size = vectorization_factor * group_size / nunits;
7357 /* Schedule the tree of INSTANCE. */
7358 is_store = vect_schedule_slp_instance (SLP_INSTANCE_TREE (instance),
7361 if (vect_print_dump_info (REPORT_VECTORIZED_LOOPS)
7362 || vect_print_dump_info (REPORT_UNVECTORIZED_LOOPS))
7363 fprintf (vect_dump, "vectorizing stmts using SLP.");
7369 /* Function vect_transform_loop.
7371 The analysis phase has determined that the loop is vectorizable.
7372 Vectorize the loop - created vectorized stmts to replace the scalar
7373 stmts in the loop, and update the loop exit condition. */
7376 vect_transform_loop (loop_vec_info loop_vinfo)
7378 struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
7379 basic_block *bbs = LOOP_VINFO_BBS (loop_vinfo);
7380 int nbbs = loop->num_nodes;
7381 block_stmt_iterator si, next_si;
7384 int vectorization_factor = LOOP_VINFO_VECT_FACTOR (loop_vinfo);
7386 bool slp_scheduled = false;
7387 unsigned int nunits;
7389 if (vect_print_dump_info (REPORT_DETAILS))
7390 fprintf (vect_dump, "=== vec_transform_loop ===");
7392 if (VEC_length (tree, LOOP_VINFO_MAY_MISALIGN_STMTS (loop_vinfo))
7393 || VEC_length (ddr_p, LOOP_VINFO_MAY_ALIAS_DDRS (loop_vinfo)))
7394 vect_loop_versioning (loop_vinfo);
7396 /* CHECKME: we wouldn't need this if we called update_ssa once
7398 bitmap_zero (vect_memsyms_to_rename);
7400 /* Peel the loop if there are data refs with unknown alignment.
7401 Only one data ref with unknown store is allowed. */
7403 if (LOOP_PEELING_FOR_ALIGNMENT (loop_vinfo))
7404 vect_do_peeling_for_alignment (loop_vinfo);
7406 /* If the loop has a symbolic number of iterations 'n' (i.e. it's not a
7407 compile time constant), or it is a constant that doesn't divide by the
7408 vectorization factor, then an epilog loop needs to be created.
7409 We therefore duplicate the loop: the original loop will be vectorized,
7410 and will compute the first (n/VF) iterations. The second copy of the loop
7411 will remain scalar and will compute the remaining (n%VF) iterations.
7412 (VF is the vectorization factor). */
7414 if (!LOOP_VINFO_NITERS_KNOWN_P (loop_vinfo)
7415 || (LOOP_VINFO_NITERS_KNOWN_P (loop_vinfo)
7416 && LOOP_VINFO_INT_NITERS (loop_vinfo) % vectorization_factor != 0))
7417 vect_do_peeling_for_loop_bound (loop_vinfo, &ratio);
7419 ratio = build_int_cst (TREE_TYPE (LOOP_VINFO_NITERS (loop_vinfo)),
7420 LOOP_VINFO_INT_NITERS (loop_vinfo) / vectorization_factor);
7422 /* 1) Make sure the loop header has exactly two entries
7423 2) Make sure we have a preheader basic block. */
7425 gcc_assert (EDGE_COUNT (loop->header->preds) == 2);
7427 split_edge (loop_preheader_edge (loop));
7429 /* FORNOW: the vectorizer supports only loops which body consist
7430 of one basic block (header + empty latch). When the vectorizer will
7431 support more involved loop forms, the order by which the BBs are
7432 traversed need to be reconsidered. */
7434 for (i = 0; i < nbbs; i++)
7436 basic_block bb = bbs[i];
7437 stmt_vec_info stmt_info;
7440 for (phi = phi_nodes (bb); phi; phi = PHI_CHAIN (phi))
7442 if (vect_print_dump_info (REPORT_DETAILS))
7444 fprintf (vect_dump, "------>vectorizing phi: ");
7445 print_generic_expr (vect_dump, phi, TDF_SLIM);
7447 stmt_info = vinfo_for_stmt (phi);
7451 if (!STMT_VINFO_RELEVANT_P (stmt_info)
7452 && !STMT_VINFO_LIVE_P (stmt_info))
7455 if ((TYPE_VECTOR_SUBPARTS (STMT_VINFO_VECTYPE (stmt_info))
7456 != (unsigned HOST_WIDE_INT) vectorization_factor)
7457 && vect_print_dump_info (REPORT_DETAILS))
7458 fprintf (vect_dump, "multiple-types.");
7460 if (STMT_VINFO_DEF_TYPE (stmt_info) == vect_induction_def)
7462 if (vect_print_dump_info (REPORT_DETAILS))
7463 fprintf (vect_dump, "transform phi.");
7464 vect_transform_stmt (phi, NULL, NULL, NULL);
7468 for (si = bsi_start (bb); !bsi_end_p (si);)
7470 tree stmt = bsi_stmt (si);
7473 if (vect_print_dump_info (REPORT_DETAILS))
7475 fprintf (vect_dump, "------>vectorizing statement: ");
7476 print_generic_expr (vect_dump, stmt, TDF_SLIM);
7479 stmt_info = vinfo_for_stmt (stmt);
7481 /* vector stmts created in the outer-loop during vectorization of
7482 stmts in an inner-loop may not have a stmt_info, and do not
7483 need to be vectorized. */
7490 if (!STMT_VINFO_RELEVANT_P (stmt_info)
7491 && !STMT_VINFO_LIVE_P (stmt_info))
7497 gcc_assert (STMT_VINFO_VECTYPE (stmt_info));
7499 (unsigned int) TYPE_VECTOR_SUBPARTS (STMT_VINFO_VECTYPE (stmt_info));
7500 if (!STMT_SLP_TYPE (stmt_info)
7501 && nunits != (unsigned int) vectorization_factor
7502 && vect_print_dump_info (REPORT_DETAILS))
7503 /* For SLP VF is set according to unrolling factor, and not to
7504 vector size, hence for SLP this print is not valid. */
7505 fprintf (vect_dump, "multiple-types.");
7507 /* SLP. Schedule all the SLP instances when the first SLP stmt is
7509 if (STMT_SLP_TYPE (stmt_info))
7513 slp_scheduled = true;
7515 if (vect_print_dump_info (REPORT_DETAILS))
7516 fprintf (vect_dump, "=== scheduling SLP instances ===");
7518 is_store = vect_schedule_slp (loop_vinfo, nunits);
7520 /* IS_STORE is true if STMT is a store. Stores cannot be of
7521 hybrid SLP type. They are removed in
7522 vect_schedule_slp_instance and their vinfo is destroyed. */
7530 /* Hybrid SLP stmts must be vectorized in addition to SLP. */
7531 if (PURE_SLP_STMT (stmt_info))
7538 /* -------- vectorize statement ------------ */
7539 if (vect_print_dump_info (REPORT_DETAILS))
7540 fprintf (vect_dump, "transform statement.");
7542 strided_store = false;
7543 is_store = vect_transform_stmt (stmt, &si, &strided_store, NULL);
7547 if (STMT_VINFO_STRIDED_ACCESS (stmt_info))
7549 /* Interleaving. If IS_STORE is TRUE, the vectorization of the
7550 interleaving chain was completed - free all the stores in
7552 tree next = DR_GROUP_FIRST_DR (stmt_info);
7554 stmt_vec_info next_stmt_info;
7558 next_si = bsi_for_stmt (next);
7559 next_stmt_info = vinfo_for_stmt (next);
7560 /* Free the attached stmt_vec_info and remove the stmt. */
7561 ann = stmt_ann (next);
7562 tmp = DR_GROUP_NEXT_DR (next_stmt_info);
7563 free (next_stmt_info);
7564 set_stmt_info (ann, NULL);
7565 bsi_remove (&next_si, true);
7568 bsi_remove (&si, true);
7573 /* Free the attached stmt_vec_info and remove the stmt. */
7574 ann = stmt_ann (stmt);
7576 set_stmt_info (ann, NULL);
7577 bsi_remove (&si, true);
7585 slpeel_make_loop_iterate_ntimes (loop, ratio);
7587 mark_set_for_renaming (vect_memsyms_to_rename);
7589 /* The memory tags and pointers in vectorized statements need to
7590 have their SSA forms updated. FIXME, why can't this be delayed
7591 until all the loops have been transformed? */
7592 update_ssa (TODO_update_ssa);
7594 if (vect_print_dump_info (REPORT_VECTORIZED_LOOPS))
7595 fprintf (vect_dump, "LOOP VECTORIZED.");
7596 if (loop->inner && vect_print_dump_info (REPORT_VECTORIZED_LOOPS))
7597 fprintf (vect_dump, "OUTER LOOP VECTORIZED.");