1 /* Data References Analysis and Manipulation Utilities for Vectorization.
2 Copyright (C) 2003, 2004, 2005, 2006, 2007, 2008, 2009, 2010, 2011
3 Free Software Foundation, Inc.
4 Contributed by Dorit Naishlos <dorit@il.ibm.com>
5 and Ira Rosen <irar@il.ibm.com>
7 This file is part of GCC.
9 GCC is free software; you can redistribute it and/or modify it under
10 the terms of the GNU General Public License as published by the Free
11 Software Foundation; either version 3, or (at your option) any later
14 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
15 WARRANTY; without even the implied warranty of MERCHANTABILITY or
16 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
19 You should have received a copy of the GNU General Public License
20 along with GCC; see the file COPYING3. If not see
21 <http://www.gnu.org/licenses/>. */
25 #include "coretypes.h"
31 #include "basic-block.h"
32 #include "tree-pretty-print.h"
33 #include "gimple-pretty-print.h"
34 #include "tree-flow.h"
35 #include "tree-dump.h"
37 #include "tree-chrec.h"
38 #include "tree-scalar-evolution.h"
39 #include "tree-vectorizer.h"
40 #include "diagnostic-core.h"
42 /* Need to include rtl.h, expr.h, etc. for optabs. */
46 /* Return true if load- or store-lanes optab OPTAB is implemented for
47 COUNT vectors of type VECTYPE. NAME is the name of OPTAB. */
50 vect_lanes_optab_supported_p (const char *name, convert_optab optab,
51 tree vectype, unsigned HOST_WIDE_INT count)
53 enum machine_mode mode, array_mode;
56 mode = TYPE_MODE (vectype);
57 limit_p = !targetm.array_mode_supported_p (mode, count);
58 array_mode = mode_for_size (count * GET_MODE_BITSIZE (mode),
61 if (array_mode == BLKmode)
63 if (vect_print_dump_info (REPORT_DETAILS))
64 fprintf (vect_dump, "no array mode for %s[" HOST_WIDE_INT_PRINT_DEC "]",
65 GET_MODE_NAME (mode), count);
69 if (convert_optab_handler (optab, array_mode, mode) == CODE_FOR_nothing)
71 if (vect_print_dump_info (REPORT_DETAILS))
72 fprintf (vect_dump, "cannot use %s<%s><%s>",
73 name, GET_MODE_NAME (array_mode), GET_MODE_NAME (mode));
77 if (vect_print_dump_info (REPORT_DETAILS))
78 fprintf (vect_dump, "can use %s<%s><%s>",
79 name, GET_MODE_NAME (array_mode), GET_MODE_NAME (mode));
85 /* Return the smallest scalar part of STMT.
86 This is used to determine the vectype of the stmt. We generally set the
87 vectype according to the type of the result (lhs). For stmts whose
88 result-type is different than the type of the arguments (e.g., demotion,
89 promotion), vectype will be reset appropriately (later). Note that we have
90 to visit the smallest datatype in this function, because that determines the
91 VF. If the smallest datatype in the loop is present only as the rhs of a
92 promotion operation - we'd miss it.
93 Such a case, where a variable of this datatype does not appear in the lhs
94 anywhere in the loop, can only occur if it's an invariant: e.g.:
95 'int_x = (int) short_inv', which we'd expect to have been optimized away by
96 invariant motion. However, we cannot rely on invariant motion to always
97 take invariants out of the loop, and so in the case of promotion we also
98 have to check the rhs.
99 LHS_SIZE_UNIT and RHS_SIZE_UNIT contain the sizes of the corresponding
103 vect_get_smallest_scalar_type (gimple stmt, HOST_WIDE_INT *lhs_size_unit,
104 HOST_WIDE_INT *rhs_size_unit)
106 tree scalar_type = gimple_expr_type (stmt);
107 HOST_WIDE_INT lhs, rhs;
109 lhs = rhs = TREE_INT_CST_LOW (TYPE_SIZE_UNIT (scalar_type));
111 if (is_gimple_assign (stmt)
112 && (gimple_assign_cast_p (stmt)
113 || gimple_assign_rhs_code (stmt) == WIDEN_MULT_EXPR
114 || gimple_assign_rhs_code (stmt) == FLOAT_EXPR))
116 tree rhs_type = TREE_TYPE (gimple_assign_rhs1 (stmt));
118 rhs = TREE_INT_CST_LOW (TYPE_SIZE_UNIT (rhs_type));
120 scalar_type = rhs_type;
123 *lhs_size_unit = lhs;
124 *rhs_size_unit = rhs;
129 /* Find the place of the data-ref in STMT in the interleaving chain that starts
130 from FIRST_STMT. Return -1 if the data-ref is not a part of the chain. */
133 vect_get_place_in_interleaving_chain (gimple stmt, gimple first_stmt)
135 gimple next_stmt = first_stmt;
138 if (first_stmt != GROUP_FIRST_ELEMENT (vinfo_for_stmt (stmt)))
141 while (next_stmt && next_stmt != stmt)
144 next_stmt = GROUP_NEXT_ELEMENT (vinfo_for_stmt (next_stmt));
154 /* Function vect_insert_into_interleaving_chain.
156 Insert DRA into the interleaving chain of DRB according to DRA's INIT. */
159 vect_insert_into_interleaving_chain (struct data_reference *dra,
160 struct data_reference *drb)
164 stmt_vec_info stmtinfo_a = vinfo_for_stmt (DR_STMT (dra));
165 stmt_vec_info stmtinfo_b = vinfo_for_stmt (DR_STMT (drb));
167 prev = GROUP_FIRST_ELEMENT (stmtinfo_b);
168 next = GROUP_NEXT_ELEMENT (vinfo_for_stmt (prev));
171 next_init = DR_INIT (STMT_VINFO_DATA_REF (vinfo_for_stmt (next)));
172 if (tree_int_cst_compare (next_init, DR_INIT (dra)) > 0)
175 GROUP_NEXT_ELEMENT (vinfo_for_stmt (prev)) = DR_STMT (dra);
176 GROUP_NEXT_ELEMENT (stmtinfo_a) = next;
180 next = GROUP_NEXT_ELEMENT (vinfo_for_stmt (prev));
183 /* We got to the end of the list. Insert here. */
184 GROUP_NEXT_ELEMENT (vinfo_for_stmt (prev)) = DR_STMT (dra);
185 GROUP_NEXT_ELEMENT (stmtinfo_a) = NULL;
189 /* Function vect_update_interleaving_chain.
191 For two data-refs DRA and DRB that are a part of a chain interleaved data
192 accesses, update the interleaving chain. DRB's INIT is smaller than DRA's.
194 There are four possible cases:
195 1. New stmts - both DRA and DRB are not a part of any chain:
198 2. DRB is a part of a chain and DRA is not:
199 no need to update FIRST_DR
200 no need to insert DRB
201 insert DRA according to init
202 3. DRA is a part of a chain and DRB is not:
203 if (init of FIRST_DR > init of DRB)
205 NEXT(FIRST_DR) = previous FIRST_DR
207 insert DRB according to its init
208 4. both DRA and DRB are in some interleaving chains:
209 choose the chain with the smallest init of FIRST_DR
210 insert the nodes of the second chain into the first one. */
213 vect_update_interleaving_chain (struct data_reference *drb,
214 struct data_reference *dra)
216 stmt_vec_info stmtinfo_a = vinfo_for_stmt (DR_STMT (dra));
217 stmt_vec_info stmtinfo_b = vinfo_for_stmt (DR_STMT (drb));
218 tree next_init, init_dra_chain, init_drb_chain;
219 gimple first_a, first_b;
221 gimple node, prev, next, first_stmt;
223 /* 1. New stmts - both DRA and DRB are not a part of any chain. */
224 if (!GROUP_FIRST_ELEMENT (stmtinfo_a) && !GROUP_FIRST_ELEMENT (stmtinfo_b))
226 GROUP_FIRST_ELEMENT (stmtinfo_a) = DR_STMT (drb);
227 GROUP_FIRST_ELEMENT (stmtinfo_b) = DR_STMT (drb);
228 GROUP_NEXT_ELEMENT (stmtinfo_b) = DR_STMT (dra);
232 /* 2. DRB is a part of a chain and DRA is not. */
233 if (!GROUP_FIRST_ELEMENT (stmtinfo_a) && GROUP_FIRST_ELEMENT (stmtinfo_b))
235 GROUP_FIRST_ELEMENT (stmtinfo_a) = GROUP_FIRST_ELEMENT (stmtinfo_b);
236 /* Insert DRA into the chain of DRB. */
237 vect_insert_into_interleaving_chain (dra, drb);
241 /* 3. DRA is a part of a chain and DRB is not. */
242 if (GROUP_FIRST_ELEMENT (stmtinfo_a) && !GROUP_FIRST_ELEMENT (stmtinfo_b))
244 gimple old_first_stmt = GROUP_FIRST_ELEMENT (stmtinfo_a);
245 tree init_old = DR_INIT (STMT_VINFO_DATA_REF (vinfo_for_stmt (
249 if (tree_int_cst_compare (init_old, DR_INIT (drb)) > 0)
251 /* DRB's init is smaller than the init of the stmt previously marked
252 as the first stmt of the interleaving chain of DRA. Therefore, we
253 update FIRST_STMT and put DRB in the head of the list. */
254 GROUP_FIRST_ELEMENT (stmtinfo_b) = DR_STMT (drb);
255 GROUP_NEXT_ELEMENT (stmtinfo_b) = old_first_stmt;
257 /* Update all the stmts in the list to point to the new FIRST_STMT. */
258 tmp = old_first_stmt;
261 GROUP_FIRST_ELEMENT (vinfo_for_stmt (tmp)) = DR_STMT (drb);
262 tmp = GROUP_NEXT_ELEMENT (vinfo_for_stmt (tmp));
267 /* Insert DRB in the list of DRA. */
268 vect_insert_into_interleaving_chain (drb, dra);
269 GROUP_FIRST_ELEMENT (stmtinfo_b) = GROUP_FIRST_ELEMENT (stmtinfo_a);
274 /* 4. both DRA and DRB are in some interleaving chains. */
275 first_a = GROUP_FIRST_ELEMENT (stmtinfo_a);
276 first_b = GROUP_FIRST_ELEMENT (stmtinfo_b);
277 if (first_a == first_b)
279 init_dra_chain = DR_INIT (STMT_VINFO_DATA_REF (vinfo_for_stmt (first_a)));
280 init_drb_chain = DR_INIT (STMT_VINFO_DATA_REF (vinfo_for_stmt (first_b)));
282 if (tree_int_cst_compare (init_dra_chain, init_drb_chain) > 0)
284 /* Insert the nodes of DRA chain into the DRB chain.
285 After inserting a node, continue from this node of the DRB chain (don't
286 start from the beginning. */
287 node = GROUP_FIRST_ELEMENT (stmtinfo_a);
288 prev = GROUP_FIRST_ELEMENT (stmtinfo_b);
289 first_stmt = first_b;
293 /* Insert the nodes of DRB chain into the DRA chain.
294 After inserting a node, continue from this node of the DRA chain (don't
295 start from the beginning. */
296 node = GROUP_FIRST_ELEMENT (stmtinfo_b);
297 prev = GROUP_FIRST_ELEMENT (stmtinfo_a);
298 first_stmt = first_a;
303 node_init = DR_INIT (STMT_VINFO_DATA_REF (vinfo_for_stmt (node)));
304 next = GROUP_NEXT_ELEMENT (vinfo_for_stmt (prev));
307 next_init = DR_INIT (STMT_VINFO_DATA_REF (vinfo_for_stmt (next)));
308 if (tree_int_cst_compare (next_init, node_init) > 0)
311 GROUP_NEXT_ELEMENT (vinfo_for_stmt (prev)) = node;
312 GROUP_NEXT_ELEMENT (vinfo_for_stmt (node)) = next;
317 next = GROUP_NEXT_ELEMENT (vinfo_for_stmt (prev));
321 /* We got to the end of the list. Insert here. */
322 GROUP_NEXT_ELEMENT (vinfo_for_stmt (prev)) = node;
323 GROUP_NEXT_ELEMENT (vinfo_for_stmt (node)) = NULL;
326 GROUP_FIRST_ELEMENT (vinfo_for_stmt (node)) = first_stmt;
327 node = GROUP_NEXT_ELEMENT (vinfo_for_stmt (node));
331 /* Check dependence between DRA and DRB for basic block vectorization.
332 If the accesses share same bases and offsets, we can compare their initial
333 constant offsets to decide whether they differ or not. In case of a read-
334 write dependence we check that the load is before the store to ensure that
335 vectorization will not change the order of the accesses. */
338 vect_drs_dependent_in_basic_block (struct data_reference *dra,
339 struct data_reference *drb)
341 HOST_WIDE_INT type_size_a, type_size_b, init_a, init_b;
344 /* We only call this function for pairs of loads and stores, but we verify
346 if (DR_IS_READ (dra) == DR_IS_READ (drb))
348 if (DR_IS_READ (dra))
354 /* Check that the data-refs have same bases and offsets. If not, we can't
355 determine if they are dependent. */
356 if (!operand_equal_p (DR_BASE_ADDRESS (dra), DR_BASE_ADDRESS (drb), 0)
357 || !dr_equal_offsets_p (dra, drb))
360 /* Check the types. */
361 type_size_a = TREE_INT_CST_LOW (TYPE_SIZE_UNIT (TREE_TYPE (DR_REF (dra))));
362 type_size_b = TREE_INT_CST_LOW (TYPE_SIZE_UNIT (TREE_TYPE (DR_REF (drb))));
364 if (type_size_a != type_size_b
365 || !types_compatible_p (TREE_TYPE (DR_REF (dra)),
366 TREE_TYPE (DR_REF (drb))))
369 init_a = TREE_INT_CST_LOW (DR_INIT (dra));
370 init_b = TREE_INT_CST_LOW (DR_INIT (drb));
372 /* Two different locations - no dependence. */
373 if (init_a != init_b)
376 /* We have a read-write dependence. Check that the load is before the store.
377 When we vectorize basic blocks, vector load can be only before
378 corresponding scalar load, and vector store can be only after its
379 corresponding scalar store. So the order of the acceses is preserved in
380 case the load is before the store. */
381 earlier_stmt = get_earlier_stmt (DR_STMT (dra), DR_STMT (drb));
382 if (DR_IS_READ (STMT_VINFO_DATA_REF (vinfo_for_stmt (earlier_stmt))))
389 /* Function vect_check_interleaving.
391 Check if DRA and DRB are a part of interleaving. In case they are, insert
392 DRA and DRB in an interleaving chain. */
395 vect_check_interleaving (struct data_reference *dra,
396 struct data_reference *drb)
398 HOST_WIDE_INT type_size_a, type_size_b, diff_mod_size, step, init_a, init_b;
400 /* Check that the data-refs have same first location (except init) and they
401 are both either store or load (not load and store). */
402 if (!operand_equal_p (DR_BASE_ADDRESS (dra), DR_BASE_ADDRESS (drb), 0)
403 || !dr_equal_offsets_p (dra, drb)
404 || !tree_int_cst_compare (DR_INIT (dra), DR_INIT (drb))
405 || DR_IS_READ (dra) != DR_IS_READ (drb))
409 1. data-refs are of the same type
410 2. their steps are equal
411 3. the step (if greater than zero) is greater than the difference between
413 type_size_a = TREE_INT_CST_LOW (TYPE_SIZE_UNIT (TREE_TYPE (DR_REF (dra))));
414 type_size_b = TREE_INT_CST_LOW (TYPE_SIZE_UNIT (TREE_TYPE (DR_REF (drb))));
416 if (type_size_a != type_size_b
417 || tree_int_cst_compare (DR_STEP (dra), DR_STEP (drb))
418 || !types_compatible_p (TREE_TYPE (DR_REF (dra)),
419 TREE_TYPE (DR_REF (drb))))
422 init_a = TREE_INT_CST_LOW (DR_INIT (dra));
423 init_b = TREE_INT_CST_LOW (DR_INIT (drb));
424 step = TREE_INT_CST_LOW (DR_STEP (dra));
428 /* If init_a == init_b + the size of the type * k, we have an interleaving,
429 and DRB is accessed before DRA. */
430 diff_mod_size = (init_a - init_b) % type_size_a;
432 if (step && (init_a - init_b) > step)
435 if (diff_mod_size == 0)
437 vect_update_interleaving_chain (drb, dra);
438 if (vect_print_dump_info (REPORT_DR_DETAILS))
440 fprintf (vect_dump, "Detected interleaving ");
441 print_generic_expr (vect_dump, DR_REF (dra), TDF_SLIM);
442 fprintf (vect_dump, " and ");
443 print_generic_expr (vect_dump, DR_REF (drb), TDF_SLIM);
450 /* If init_b == init_a + the size of the type * k, we have an
451 interleaving, and DRA is accessed before DRB. */
452 diff_mod_size = (init_b - init_a) % type_size_a;
454 if (step && (init_b - init_a) > step)
457 if (diff_mod_size == 0)
459 vect_update_interleaving_chain (dra, drb);
460 if (vect_print_dump_info (REPORT_DR_DETAILS))
462 fprintf (vect_dump, "Detected interleaving ");
463 print_generic_expr (vect_dump, DR_REF (dra), TDF_SLIM);
464 fprintf (vect_dump, " and ");
465 print_generic_expr (vect_dump, DR_REF (drb), TDF_SLIM);
474 /* Check if data references pointed by DR_I and DR_J are same or
475 belong to same interleaving group. Return FALSE if drs are
476 different, otherwise return TRUE. */
479 vect_same_range_drs (data_reference_p dr_i, data_reference_p dr_j)
481 gimple stmt_i = DR_STMT (dr_i);
482 gimple stmt_j = DR_STMT (dr_j);
484 if (operand_equal_p (DR_REF (dr_i), DR_REF (dr_j), 0)
485 || (GROUP_FIRST_ELEMENT (vinfo_for_stmt (stmt_i))
486 && GROUP_FIRST_ELEMENT (vinfo_for_stmt (stmt_j))
487 && (GROUP_FIRST_ELEMENT (vinfo_for_stmt (stmt_i))
488 == GROUP_FIRST_ELEMENT (vinfo_for_stmt (stmt_j)))))
494 /* If address ranges represented by DDR_I and DDR_J are equal,
495 return TRUE, otherwise return FALSE. */
498 vect_vfa_range_equal (ddr_p ddr_i, ddr_p ddr_j)
500 if ((vect_same_range_drs (DDR_A (ddr_i), DDR_A (ddr_j))
501 && vect_same_range_drs (DDR_B (ddr_i), DDR_B (ddr_j)))
502 || (vect_same_range_drs (DDR_A (ddr_i), DDR_B (ddr_j))
503 && vect_same_range_drs (DDR_B (ddr_i), DDR_A (ddr_j))))
509 /* Insert DDR into LOOP_VINFO list of ddrs that may alias and need to be
510 tested at run-time. Return TRUE if DDR was successfully inserted.
511 Return false if versioning is not supported. */
514 vect_mark_for_runtime_alias_test (ddr_p ddr, loop_vec_info loop_vinfo)
516 struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
518 if ((unsigned) PARAM_VALUE (PARAM_VECT_MAX_VERSION_FOR_ALIAS_CHECKS) == 0)
521 if (vect_print_dump_info (REPORT_DR_DETAILS))
523 fprintf (vect_dump, "mark for run-time aliasing test between ");
524 print_generic_expr (vect_dump, DR_REF (DDR_A (ddr)), TDF_SLIM);
525 fprintf (vect_dump, " and ");
526 print_generic_expr (vect_dump, DR_REF (DDR_B (ddr)), TDF_SLIM);
529 if (optimize_loop_nest_for_size_p (loop))
531 if (vect_print_dump_info (REPORT_DR_DETAILS))
532 fprintf (vect_dump, "versioning not supported when optimizing for size.");
536 /* FORNOW: We don't support versioning with outer-loop vectorization. */
539 if (vect_print_dump_info (REPORT_DR_DETAILS))
540 fprintf (vect_dump, "versioning not yet supported for outer-loops.");
544 VEC_safe_push (ddr_p, heap, LOOP_VINFO_MAY_ALIAS_DDRS (loop_vinfo), ddr);
549 /* Function vect_analyze_data_ref_dependence.
551 Return TRUE if there (might) exist a dependence between a memory-reference
552 DRA and a memory-reference DRB. When versioning for alias may check a
553 dependence at run-time, return FALSE. Adjust *MAX_VF according to
554 the data dependence. */
557 vect_analyze_data_ref_dependence (struct data_dependence_relation *ddr,
558 loop_vec_info loop_vinfo, int *max_vf,
559 bool *data_dependence_in_bb)
562 struct loop *loop = NULL;
563 struct data_reference *dra = DDR_A (ddr);
564 struct data_reference *drb = DDR_B (ddr);
565 stmt_vec_info stmtinfo_a = vinfo_for_stmt (DR_STMT (dra));
566 stmt_vec_info stmtinfo_b = vinfo_for_stmt (DR_STMT (drb));
567 lambda_vector dist_v;
568 unsigned int loop_depth;
570 /* Don't bother to analyze statements marked as unvectorizable. */
571 if (!STMT_VINFO_VECTORIZABLE (stmtinfo_a)
572 || !STMT_VINFO_VECTORIZABLE (stmtinfo_b))
575 if (DDR_ARE_DEPENDENT (ddr) == chrec_known)
577 /* Independent data accesses. */
578 vect_check_interleaving (dra, drb);
583 loop = LOOP_VINFO_LOOP (loop_vinfo);
585 if ((DR_IS_READ (dra) && DR_IS_READ (drb) && loop_vinfo) || dra == drb)
588 if (DDR_ARE_DEPENDENT (ddr) == chrec_dont_know)
592 if (vect_print_dump_info (REPORT_DR_DETAILS))
594 fprintf (vect_dump, "versioning for alias required: "
595 "can't determine dependence between ");
596 print_generic_expr (vect_dump, DR_REF (dra), TDF_SLIM);
597 fprintf (vect_dump, " and ");
598 print_generic_expr (vect_dump, DR_REF (drb), TDF_SLIM);
601 /* Add to list of ddrs that need to be tested at run-time. */
602 return !vect_mark_for_runtime_alias_test (ddr, loop_vinfo);
605 /* When vectorizing a basic block unknown depnedence can still mean
607 if (vect_check_interleaving (dra, drb))
610 /* Read-read is OK (we need this check here, after checking for
612 if (DR_IS_READ (dra) && DR_IS_READ (drb))
615 if (vect_print_dump_info (REPORT_DR_DETAILS))
617 fprintf (vect_dump, "can't determine dependence between ");
618 print_generic_expr (vect_dump, DR_REF (dra), TDF_SLIM);
619 fprintf (vect_dump, " and ");
620 print_generic_expr (vect_dump, DR_REF (drb), TDF_SLIM);
623 /* We do not vectorize basic blocks with write-write dependencies. */
624 if (DR_IS_WRITE (dra) && DR_IS_WRITE (drb))
627 /* We deal with read-write dependencies in basic blocks later (by
628 verifying that all the loads in the basic block are before all the
630 *data_dependence_in_bb = true;
634 /* Versioning for alias is not yet supported for basic block SLP, and
635 dependence distance is unapplicable, hence, in case of known data
636 dependence, basic block vectorization is impossible for now. */
639 if (dra != drb && vect_check_interleaving (dra, drb))
642 if (vect_print_dump_info (REPORT_DR_DETAILS))
644 fprintf (vect_dump, "determined dependence between ");
645 print_generic_expr (vect_dump, DR_REF (dra), TDF_SLIM);
646 fprintf (vect_dump, " and ");
647 print_generic_expr (vect_dump, DR_REF (drb), TDF_SLIM);
650 /* Do not vectorize basic blcoks with write-write dependences. */
651 if (DR_IS_WRITE (dra) && DR_IS_WRITE (drb))
654 /* Check if this dependence is allowed in basic block vectorization. */
655 return vect_drs_dependent_in_basic_block (dra, drb);
658 /* Loop-based vectorization and known data dependence. */
659 if (DDR_NUM_DIST_VECTS (ddr) == 0)
661 if (vect_print_dump_info (REPORT_DR_DETAILS))
663 fprintf (vect_dump, "versioning for alias required: bad dist vector for ");
664 print_generic_expr (vect_dump, DR_REF (dra), TDF_SLIM);
665 fprintf (vect_dump, " and ");
666 print_generic_expr (vect_dump, DR_REF (drb), TDF_SLIM);
668 /* Add to list of ddrs that need to be tested at run-time. */
669 return !vect_mark_for_runtime_alias_test (ddr, loop_vinfo);
672 loop_depth = index_in_loop_nest (loop->num, DDR_LOOP_NEST (ddr));
673 FOR_EACH_VEC_ELT (lambda_vector, DDR_DIST_VECTS (ddr), i, dist_v)
675 int dist = dist_v[loop_depth];
677 if (vect_print_dump_info (REPORT_DR_DETAILS))
678 fprintf (vect_dump, "dependence distance = %d.", dist);
682 if (vect_print_dump_info (REPORT_DR_DETAILS))
684 fprintf (vect_dump, "dependence distance == 0 between ");
685 print_generic_expr (vect_dump, DR_REF (dra), TDF_SLIM);
686 fprintf (vect_dump, " and ");
687 print_generic_expr (vect_dump, DR_REF (drb), TDF_SLIM);
690 /* For interleaving, mark that there is a read-write dependency if
691 necessary. We check before that one of the data-refs is store. */
692 if (DR_IS_READ (dra))
693 GROUP_READ_WRITE_DEPENDENCE (stmtinfo_a) = true;
696 if (DR_IS_READ (drb))
697 GROUP_READ_WRITE_DEPENDENCE (stmtinfo_b) = true;
703 if (dist > 0 && DDR_REVERSED_P (ddr))
705 /* If DDR_REVERSED_P the order of the data-refs in DDR was
706 reversed (to make distance vector positive), and the actual
707 distance is negative. */
708 if (vect_print_dump_info (REPORT_DR_DETAILS))
709 fprintf (vect_dump, "dependence distance negative.");
714 && abs (dist) < *max_vf)
716 /* The dependence distance requires reduction of the maximal
717 vectorization factor. */
718 *max_vf = abs (dist);
719 if (vect_print_dump_info (REPORT_DR_DETAILS))
720 fprintf (vect_dump, "adjusting maximal vectorization factor to %i",
724 if (abs (dist) >= *max_vf)
726 /* Dependence distance does not create dependence, as far as
727 vectorization is concerned, in this case. */
728 if (vect_print_dump_info (REPORT_DR_DETAILS))
729 fprintf (vect_dump, "dependence distance >= VF.");
733 if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS))
735 fprintf (vect_dump, "not vectorized, possible dependence "
736 "between data-refs ");
737 print_generic_expr (vect_dump, DR_REF (dra), TDF_SLIM);
738 fprintf (vect_dump, " and ");
739 print_generic_expr (vect_dump, DR_REF (drb), TDF_SLIM);
748 /* Function vect_analyze_data_ref_dependences.
750 Examine all the data references in the loop, and make sure there do not
751 exist any data dependences between them. Set *MAX_VF according to
752 the maximum vectorization factor the data dependences allow. */
755 vect_analyze_data_ref_dependences (loop_vec_info loop_vinfo,
756 bb_vec_info bb_vinfo, int *max_vf,
757 bool *data_dependence_in_bb)
760 VEC (ddr_p, heap) *ddrs = NULL;
761 struct data_dependence_relation *ddr;
763 if (vect_print_dump_info (REPORT_DETAILS))
764 fprintf (vect_dump, "=== vect_analyze_dependences ===");
767 ddrs = LOOP_VINFO_DDRS (loop_vinfo);
769 ddrs = BB_VINFO_DDRS (bb_vinfo);
771 FOR_EACH_VEC_ELT (ddr_p, ddrs, i, ddr)
772 if (vect_analyze_data_ref_dependence (ddr, loop_vinfo, max_vf,
773 data_dependence_in_bb))
780 /* Function vect_compute_data_ref_alignment
782 Compute the misalignment of the data reference DR.
785 1. If during the misalignment computation it is found that the data reference
786 cannot be vectorized then false is returned.
787 2. DR_MISALIGNMENT (DR) is defined.
789 FOR NOW: No analysis is actually performed. Misalignment is calculated
790 only for trivial cases. TODO. */
793 vect_compute_data_ref_alignment (struct data_reference *dr)
795 gimple stmt = DR_STMT (dr);
796 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
797 loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info);
798 struct loop *loop = NULL;
799 tree ref = DR_REF (dr);
801 tree base, base_addr;
804 tree aligned_to, alignment;
806 if (vect_print_dump_info (REPORT_DETAILS))
807 fprintf (vect_dump, "vect_compute_data_ref_alignment:");
810 loop = LOOP_VINFO_LOOP (loop_vinfo);
812 /* Initialize misalignment to unknown. */
813 SET_DR_MISALIGNMENT (dr, -1);
815 misalign = DR_INIT (dr);
816 aligned_to = DR_ALIGNED_TO (dr);
817 base_addr = DR_BASE_ADDRESS (dr);
818 vectype = STMT_VINFO_VECTYPE (stmt_info);
820 /* In case the dataref is in an inner-loop of the loop that is being
821 vectorized (LOOP), we use the base and misalignment information
822 relative to the outer-loop (LOOP). This is ok only if the misalignment
823 stays the same throughout the execution of the inner-loop, which is why
824 we have to check that the stride of the dataref in the inner-loop evenly
825 divides by the vector size. */
826 if (loop && nested_in_vect_loop_p (loop, stmt))
828 tree step = DR_STEP (dr);
829 HOST_WIDE_INT dr_step = TREE_INT_CST_LOW (step);
831 if (dr_step % GET_MODE_SIZE (TYPE_MODE (vectype)) == 0)
833 if (vect_print_dump_info (REPORT_ALIGNMENT))
834 fprintf (vect_dump, "inner step divides the vector-size.");
835 misalign = STMT_VINFO_DR_INIT (stmt_info);
836 aligned_to = STMT_VINFO_DR_ALIGNED_TO (stmt_info);
837 base_addr = STMT_VINFO_DR_BASE_ADDRESS (stmt_info);
841 if (vect_print_dump_info (REPORT_ALIGNMENT))
842 fprintf (vect_dump, "inner step doesn't divide the vector-size.");
843 misalign = NULL_TREE;
847 base = build_fold_indirect_ref (base_addr);
848 alignment = ssize_int (TYPE_ALIGN (vectype)/BITS_PER_UNIT);
850 if ((aligned_to && tree_int_cst_compare (aligned_to, alignment) < 0)
853 if (vect_print_dump_info (REPORT_ALIGNMENT))
855 fprintf (vect_dump, "Unknown alignment for access: ");
856 print_generic_expr (vect_dump, base, TDF_SLIM);
862 && tree_int_cst_compare (ssize_int (DECL_ALIGN_UNIT (base)),
864 || (TREE_CODE (base_addr) == SSA_NAME
865 && tree_int_cst_compare (ssize_int (TYPE_ALIGN_UNIT (TREE_TYPE (
866 TREE_TYPE (base_addr)))),
868 || (get_pointer_alignment (base_addr) >= TYPE_ALIGN (vectype)))
871 base_aligned = false;
875 /* Do not change the alignment of global variables if
876 flag_section_anchors is enabled. */
877 if (!vect_can_force_dr_alignment_p (base, TYPE_ALIGN (vectype))
878 || (TREE_STATIC (base) && flag_section_anchors))
880 if (vect_print_dump_info (REPORT_DETAILS))
882 fprintf (vect_dump, "can't force alignment of ref: ");
883 print_generic_expr (vect_dump, ref, TDF_SLIM);
888 /* Force the alignment of the decl.
889 NOTE: This is the only change to the code we make during
890 the analysis phase, before deciding to vectorize the loop. */
891 if (vect_print_dump_info (REPORT_DETAILS))
893 fprintf (vect_dump, "force alignment of ");
894 print_generic_expr (vect_dump, ref, TDF_SLIM);
897 DECL_ALIGN (base) = TYPE_ALIGN (vectype);
898 DECL_USER_ALIGN (base) = 1;
901 /* At this point we assume that the base is aligned. */
902 gcc_assert (base_aligned
903 || (TREE_CODE (base) == VAR_DECL
904 && DECL_ALIGN (base) >= TYPE_ALIGN (vectype)));
906 /* If this is a backward running DR then first access in the larger
907 vectype actually is N-1 elements before the address in the DR.
908 Adjust misalign accordingly. */
909 if (tree_int_cst_compare (DR_STEP (dr), size_zero_node) < 0)
911 tree offset = ssize_int (TYPE_VECTOR_SUBPARTS (vectype) - 1);
912 /* DR_STEP(dr) is the same as -TYPE_SIZE of the scalar type,
913 otherwise we wouldn't be here. */
914 offset = fold_build2 (MULT_EXPR, ssizetype, offset, DR_STEP (dr));
915 /* PLUS because DR_STEP was negative. */
916 misalign = size_binop (PLUS_EXPR, misalign, offset);
919 /* Modulo alignment. */
920 misalign = size_binop (FLOOR_MOD_EXPR, misalign, alignment);
922 if (!host_integerp (misalign, 1))
924 /* Negative or overflowed misalignment value. */
925 if (vect_print_dump_info (REPORT_DETAILS))
926 fprintf (vect_dump, "unexpected misalign value");
930 SET_DR_MISALIGNMENT (dr, TREE_INT_CST_LOW (misalign));
932 if (vect_print_dump_info (REPORT_DETAILS))
934 fprintf (vect_dump, "misalign = %d bytes of ref ", DR_MISALIGNMENT (dr));
935 print_generic_expr (vect_dump, ref, TDF_SLIM);
942 /* Function vect_compute_data_refs_alignment
944 Compute the misalignment of data references in the loop.
945 Return FALSE if a data reference is found that cannot be vectorized. */
948 vect_compute_data_refs_alignment (loop_vec_info loop_vinfo,
949 bb_vec_info bb_vinfo)
951 VEC (data_reference_p, heap) *datarefs;
952 struct data_reference *dr;
956 datarefs = LOOP_VINFO_DATAREFS (loop_vinfo);
958 datarefs = BB_VINFO_DATAREFS (bb_vinfo);
960 FOR_EACH_VEC_ELT (data_reference_p, datarefs, i, dr)
961 if (STMT_VINFO_VECTORIZABLE (vinfo_for_stmt (DR_STMT (dr)))
962 && !vect_compute_data_ref_alignment (dr))
966 /* Mark unsupported statement as unvectorizable. */
967 STMT_VINFO_VECTORIZABLE (vinfo_for_stmt (DR_STMT (dr))) = false;
978 /* Function vect_update_misalignment_for_peel
980 DR - the data reference whose misalignment is to be adjusted.
981 DR_PEEL - the data reference whose misalignment is being made
982 zero in the vector loop by the peel.
983 NPEEL - the number of iterations in the peel loop if the misalignment
984 of DR_PEEL is known at compile time. */
987 vect_update_misalignment_for_peel (struct data_reference *dr,
988 struct data_reference *dr_peel, int npeel)
991 VEC(dr_p,heap) *same_align_drs;
992 struct data_reference *current_dr;
993 int dr_size = GET_MODE_SIZE (TYPE_MODE (TREE_TYPE (DR_REF (dr))));
994 int dr_peel_size = GET_MODE_SIZE (TYPE_MODE (TREE_TYPE (DR_REF (dr_peel))));
995 stmt_vec_info stmt_info = vinfo_for_stmt (DR_STMT (dr));
996 stmt_vec_info peel_stmt_info = vinfo_for_stmt (DR_STMT (dr_peel));
998 /* For interleaved data accesses the step in the loop must be multiplied by
999 the size of the interleaving group. */
1000 if (STMT_VINFO_STRIDED_ACCESS (stmt_info))
1001 dr_size *= GROUP_SIZE (vinfo_for_stmt (GROUP_FIRST_ELEMENT (stmt_info)));
1002 if (STMT_VINFO_STRIDED_ACCESS (peel_stmt_info))
1003 dr_peel_size *= GROUP_SIZE (peel_stmt_info);
1005 /* It can be assumed that the data refs with the same alignment as dr_peel
1006 are aligned in the vector loop. */
1008 = STMT_VINFO_SAME_ALIGN_REFS (vinfo_for_stmt (DR_STMT (dr_peel)));
1009 FOR_EACH_VEC_ELT (dr_p, same_align_drs, i, current_dr)
1011 if (current_dr != dr)
1013 gcc_assert (DR_MISALIGNMENT (dr) / dr_size ==
1014 DR_MISALIGNMENT (dr_peel) / dr_peel_size);
1015 SET_DR_MISALIGNMENT (dr, 0);
1019 if (known_alignment_for_access_p (dr)
1020 && known_alignment_for_access_p (dr_peel))
1022 bool negative = tree_int_cst_compare (DR_STEP (dr), size_zero_node) < 0;
1023 int misal = DR_MISALIGNMENT (dr);
1024 tree vectype = STMT_VINFO_VECTYPE (stmt_info);
1025 misal += negative ? -npeel * dr_size : npeel * dr_size;
1026 misal &= GET_MODE_SIZE (TYPE_MODE (vectype)) - 1;
1027 SET_DR_MISALIGNMENT (dr, misal);
1031 if (vect_print_dump_info (REPORT_DETAILS))
1032 fprintf (vect_dump, "Setting misalignment to -1.");
1033 SET_DR_MISALIGNMENT (dr, -1);
1037 /* Function vect_verify_datarefs_alignment
1039 Return TRUE if all data references in the loop can be
1040 handled with respect to alignment. */
1043 vect_verify_datarefs_alignment (loop_vec_info loop_vinfo, bb_vec_info bb_vinfo)
1045 VEC (data_reference_p, heap) *datarefs;
1046 struct data_reference *dr;
1047 enum dr_alignment_support supportable_dr_alignment;
1051 datarefs = LOOP_VINFO_DATAREFS (loop_vinfo);
1053 datarefs = BB_VINFO_DATAREFS (bb_vinfo);
1055 FOR_EACH_VEC_ELT (data_reference_p, datarefs, i, dr)
1057 gimple stmt = DR_STMT (dr);
1058 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
1060 /* For interleaving, only the alignment of the first access matters.
1061 Skip statements marked as not vectorizable. */
1062 if ((STMT_VINFO_STRIDED_ACCESS (stmt_info)
1063 && GROUP_FIRST_ELEMENT (stmt_info) != stmt)
1064 || !STMT_VINFO_VECTORIZABLE (stmt_info))
1067 supportable_dr_alignment = vect_supportable_dr_alignment (dr, false);
1068 if (!supportable_dr_alignment)
1070 if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS))
1072 if (DR_IS_READ (dr))
1074 "not vectorized: unsupported unaligned load.");
1077 "not vectorized: unsupported unaligned store.");
1079 print_generic_expr (vect_dump, DR_REF (dr), TDF_SLIM);
1083 if (supportable_dr_alignment != dr_aligned
1084 && vect_print_dump_info (REPORT_ALIGNMENT))
1085 fprintf (vect_dump, "Vectorizing an unaligned access.");
1091 /* Function vector_alignment_reachable_p
1093 Return true if vector alignment for DR is reachable by peeling
1094 a few loop iterations. Return false otherwise. */
1097 vector_alignment_reachable_p (struct data_reference *dr)
1099 gimple stmt = DR_STMT (dr);
1100 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
1101 tree vectype = STMT_VINFO_VECTYPE (stmt_info);
1103 if (STMT_VINFO_STRIDED_ACCESS (stmt_info))
1105 /* For interleaved access we peel only if number of iterations in
1106 the prolog loop ({VF - misalignment}), is a multiple of the
1107 number of the interleaved accesses. */
1108 int elem_size, mis_in_elements;
1109 int nelements = TYPE_VECTOR_SUBPARTS (vectype);
1111 /* FORNOW: handle only known alignment. */
1112 if (!known_alignment_for_access_p (dr))
1115 elem_size = GET_MODE_SIZE (TYPE_MODE (vectype)) / nelements;
1116 mis_in_elements = DR_MISALIGNMENT (dr) / elem_size;
1118 if ((nelements - mis_in_elements) % GROUP_SIZE (stmt_info))
1122 /* If misalignment is known at the compile time then allow peeling
1123 only if natural alignment is reachable through peeling. */
1124 if (known_alignment_for_access_p (dr) && !aligned_access_p (dr))
1126 HOST_WIDE_INT elmsize =
1127 int_cst_value (TYPE_SIZE_UNIT (TREE_TYPE (vectype)));
1128 if (vect_print_dump_info (REPORT_DETAILS))
1130 fprintf (vect_dump, "data size =" HOST_WIDE_INT_PRINT_DEC, elmsize);
1131 fprintf (vect_dump, ". misalignment = %d. ", DR_MISALIGNMENT (dr));
1133 if (DR_MISALIGNMENT (dr) % elmsize)
1135 if (vect_print_dump_info (REPORT_DETAILS))
1136 fprintf (vect_dump, "data size does not divide the misalignment.\n");
1141 if (!known_alignment_for_access_p (dr))
1143 tree type = (TREE_TYPE (DR_REF (dr)));
1144 tree ba = DR_BASE_OBJECT (dr);
1145 bool is_packed = false;
1148 is_packed = contains_packed_reference (ba);
1150 if (compare_tree_int (TYPE_SIZE (type), TYPE_ALIGN (type)) > 0)
1153 if (vect_print_dump_info (REPORT_DETAILS))
1154 fprintf (vect_dump, "Unknown misalignment, is_packed = %d",is_packed);
1155 if (targetm.vectorize.vector_alignment_reachable (type, is_packed))
1165 /* Calculate the cost of the memory access represented by DR. */
1168 vect_get_data_access_cost (struct data_reference *dr,
1169 unsigned int *inside_cost,
1170 unsigned int *outside_cost)
1172 gimple stmt = DR_STMT (dr);
1173 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
1174 int nunits = TYPE_VECTOR_SUBPARTS (STMT_VINFO_VECTYPE (stmt_info));
1175 loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info);
1176 int vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo);
1177 int ncopies = vf / nunits;
1178 bool supportable_dr_alignment = vect_supportable_dr_alignment (dr, true);
1180 if (!supportable_dr_alignment)
1181 *inside_cost = VECT_MAX_COST;
1184 if (DR_IS_READ (dr))
1185 vect_get_load_cost (dr, ncopies, true, inside_cost, outside_cost);
1187 vect_get_store_cost (dr, ncopies, inside_cost);
1190 if (vect_print_dump_info (REPORT_COST))
1191 fprintf (vect_dump, "vect_get_data_access_cost: inside_cost = %d, "
1192 "outside_cost = %d.", *inside_cost, *outside_cost);
1197 vect_peeling_hash (const void *elem)
1199 const struct _vect_peel_info *peel_info;
1201 peel_info = (const struct _vect_peel_info *) elem;
1202 return (hashval_t) peel_info->npeel;
1207 vect_peeling_hash_eq (const void *elem1, const void *elem2)
1209 const struct _vect_peel_info *a, *b;
1211 a = (const struct _vect_peel_info *) elem1;
1212 b = (const struct _vect_peel_info *) elem2;
1213 return (a->npeel == b->npeel);
1217 /* Insert DR into peeling hash table with NPEEL as key. */
1220 vect_peeling_hash_insert (loop_vec_info loop_vinfo, struct data_reference *dr,
1223 struct _vect_peel_info elem, *slot;
1225 bool supportable_dr_alignment = vect_supportable_dr_alignment (dr, true);
1228 slot = (vect_peel_info) htab_find (LOOP_VINFO_PEELING_HTAB (loop_vinfo),
1234 slot = XNEW (struct _vect_peel_info);
1235 slot->npeel = npeel;
1238 new_slot = htab_find_slot (LOOP_VINFO_PEELING_HTAB (loop_vinfo), slot,
1243 if (!supportable_dr_alignment && !flag_vect_cost_model)
1244 slot->count += VECT_MAX_COST;
1248 /* Traverse peeling hash table to find peeling option that aligns maximum
1249 number of data accesses. */
1252 vect_peeling_hash_get_most_frequent (void **slot, void *data)
1254 vect_peel_info elem = (vect_peel_info) *slot;
1255 vect_peel_extended_info max = (vect_peel_extended_info) data;
1257 if (elem->count > max->peel_info.count
1258 || (elem->count == max->peel_info.count
1259 && max->peel_info.npeel > elem->npeel))
1261 max->peel_info.npeel = elem->npeel;
1262 max->peel_info.count = elem->count;
1263 max->peel_info.dr = elem->dr;
1270 /* Traverse peeling hash table and calculate cost for each peeling option.
1271 Find the one with the lowest cost. */
1274 vect_peeling_hash_get_lowest_cost (void **slot, void *data)
1276 vect_peel_info elem = (vect_peel_info) *slot;
1277 vect_peel_extended_info min = (vect_peel_extended_info) data;
1278 int save_misalignment, dummy;
1279 unsigned int inside_cost = 0, outside_cost = 0, i;
1280 gimple stmt = DR_STMT (elem->dr);
1281 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
1282 loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info);
1283 VEC (data_reference_p, heap) *datarefs = LOOP_VINFO_DATAREFS (loop_vinfo);
1284 struct data_reference *dr;
1286 FOR_EACH_VEC_ELT (data_reference_p, datarefs, i, dr)
1288 stmt = DR_STMT (dr);
1289 stmt_info = vinfo_for_stmt (stmt);
1290 /* For interleaving, only the alignment of the first access
1292 if (STMT_VINFO_STRIDED_ACCESS (stmt_info)
1293 && GROUP_FIRST_ELEMENT (stmt_info) != stmt)
1296 save_misalignment = DR_MISALIGNMENT (dr);
1297 vect_update_misalignment_for_peel (dr, elem->dr, elem->npeel);
1298 vect_get_data_access_cost (dr, &inside_cost, &outside_cost);
1299 SET_DR_MISALIGNMENT (dr, save_misalignment);
1302 outside_cost += vect_get_known_peeling_cost (loop_vinfo, elem->npeel, &dummy,
1303 vect_get_single_scalar_iteraion_cost (loop_vinfo));
1305 if (inside_cost < min->inside_cost
1306 || (inside_cost == min->inside_cost && outside_cost < min->outside_cost))
1308 min->inside_cost = inside_cost;
1309 min->outside_cost = outside_cost;
1310 min->peel_info.dr = elem->dr;
1311 min->peel_info.npeel = elem->npeel;
1318 /* Choose best peeling option by traversing peeling hash table and either
1319 choosing an option with the lowest cost (if cost model is enabled) or the
1320 option that aligns as many accesses as possible. */
1322 static struct data_reference *
1323 vect_peeling_hash_choose_best_peeling (loop_vec_info loop_vinfo,
1324 unsigned int *npeel)
1326 struct _vect_peel_extended_info res;
1328 res.peel_info.dr = NULL;
1330 if (flag_vect_cost_model)
1332 res.inside_cost = INT_MAX;
1333 res.outside_cost = INT_MAX;
1334 htab_traverse (LOOP_VINFO_PEELING_HTAB (loop_vinfo),
1335 vect_peeling_hash_get_lowest_cost, &res);
1339 res.peel_info.count = 0;
1340 htab_traverse (LOOP_VINFO_PEELING_HTAB (loop_vinfo),
1341 vect_peeling_hash_get_most_frequent, &res);
1344 *npeel = res.peel_info.npeel;
1345 return res.peel_info.dr;
1349 /* Function vect_enhance_data_refs_alignment
1351 This pass will use loop versioning and loop peeling in order to enhance
1352 the alignment of data references in the loop.
1354 FOR NOW: we assume that whatever versioning/peeling takes place, only the
1355 original loop is to be vectorized. Any other loops that are created by
1356 the transformations performed in this pass - are not supposed to be
1357 vectorized. This restriction will be relaxed.
1359 This pass will require a cost model to guide it whether to apply peeling
1360 or versioning or a combination of the two. For example, the scheme that
1361 intel uses when given a loop with several memory accesses, is as follows:
1362 choose one memory access ('p') which alignment you want to force by doing
1363 peeling. Then, either (1) generate a loop in which 'p' is aligned and all
1364 other accesses are not necessarily aligned, or (2) use loop versioning to
1365 generate one loop in which all accesses are aligned, and another loop in
1366 which only 'p' is necessarily aligned.
1368 ("Automatic Intra-Register Vectorization for the Intel Architecture",
1369 Aart J.C. Bik, Milind Girkar, Paul M. Grey and Ximmin Tian, International
1370 Journal of Parallel Programming, Vol. 30, No. 2, April 2002.)
1372 Devising a cost model is the most critical aspect of this work. It will
1373 guide us on which access to peel for, whether to use loop versioning, how
1374 many versions to create, etc. The cost model will probably consist of
1375 generic considerations as well as target specific considerations (on
1376 powerpc for example, misaligned stores are more painful than misaligned
1379 Here are the general steps involved in alignment enhancements:
1381 -- original loop, before alignment analysis:
1382 for (i=0; i<N; i++){
1383 x = q[i]; # DR_MISALIGNMENT(q) = unknown
1384 p[i] = y; # DR_MISALIGNMENT(p) = unknown
1387 -- After vect_compute_data_refs_alignment:
1388 for (i=0; i<N; i++){
1389 x = q[i]; # DR_MISALIGNMENT(q) = 3
1390 p[i] = y; # DR_MISALIGNMENT(p) = unknown
1393 -- Possibility 1: we do loop versioning:
1395 for (i=0; i<N; i++){ # loop 1A
1396 x = q[i]; # DR_MISALIGNMENT(q) = 3
1397 p[i] = y; # DR_MISALIGNMENT(p) = 0
1401 for (i=0; i<N; i++){ # loop 1B
1402 x = q[i]; # DR_MISALIGNMENT(q) = 3
1403 p[i] = y; # DR_MISALIGNMENT(p) = unaligned
1407 -- Possibility 2: we do loop peeling:
1408 for (i = 0; i < 3; i++){ # (scalar loop, not to be vectorized).
1412 for (i = 3; i < N; i++){ # loop 2A
1413 x = q[i]; # DR_MISALIGNMENT(q) = 0
1414 p[i] = y; # DR_MISALIGNMENT(p) = unknown
1417 -- Possibility 3: combination of loop peeling and versioning:
1418 for (i = 0; i < 3; i++){ # (scalar loop, not to be vectorized).
1423 for (i = 3; i<N; i++){ # loop 3A
1424 x = q[i]; # DR_MISALIGNMENT(q) = 0
1425 p[i] = y; # DR_MISALIGNMENT(p) = 0
1429 for (i = 3; i<N; i++){ # loop 3B
1430 x = q[i]; # DR_MISALIGNMENT(q) = 0
1431 p[i] = y; # DR_MISALIGNMENT(p) = unaligned
1435 These loops are later passed to loop_transform to be vectorized. The
1436 vectorizer will use the alignment information to guide the transformation
1437 (whether to generate regular loads/stores, or with special handling for
1441 vect_enhance_data_refs_alignment (loop_vec_info loop_vinfo)
1443 VEC (data_reference_p, heap) *datarefs = LOOP_VINFO_DATAREFS (loop_vinfo);
1444 struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
1445 enum dr_alignment_support supportable_dr_alignment;
1446 struct data_reference *dr0 = NULL, *first_store = NULL;
1447 struct data_reference *dr;
1449 bool do_peeling = false;
1450 bool do_versioning = false;
1453 stmt_vec_info stmt_info;
1454 int vect_versioning_for_alias_required;
1455 unsigned int npeel = 0;
1456 bool all_misalignments_unknown = true;
1457 unsigned int vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo);
1458 unsigned possible_npeel_number = 1;
1460 unsigned int nelements, mis, same_align_drs_max = 0;
1462 if (vect_print_dump_info (REPORT_DETAILS))
1463 fprintf (vect_dump, "=== vect_enhance_data_refs_alignment ===");
1465 /* While cost model enhancements are expected in the future, the high level
1466 view of the code at this time is as follows:
1468 A) If there is a misaligned access then see if peeling to align
1469 this access can make all data references satisfy
1470 vect_supportable_dr_alignment. If so, update data structures
1471 as needed and return true.
1473 B) If peeling wasn't possible and there is a data reference with an
1474 unknown misalignment that does not satisfy vect_supportable_dr_alignment
1475 then see if loop versioning checks can be used to make all data
1476 references satisfy vect_supportable_dr_alignment. If so, update
1477 data structures as needed and return true.
1479 C) If neither peeling nor versioning were successful then return false if
1480 any data reference does not satisfy vect_supportable_dr_alignment.
1482 D) Return true (all data references satisfy vect_supportable_dr_alignment).
1484 Note, Possibility 3 above (which is peeling and versioning together) is not
1485 being done at this time. */
1487 /* (1) Peeling to force alignment. */
1489 /* (1.1) Decide whether to perform peeling, and how many iterations to peel:
1491 + How many accesses will become aligned due to the peeling
1492 - How many accesses will become unaligned due to the peeling,
1493 and the cost of misaligned accesses.
1494 - The cost of peeling (the extra runtime checks, the increase
1497 FOR_EACH_VEC_ELT (data_reference_p, datarefs, i, dr)
1499 stmt = DR_STMT (dr);
1500 stmt_info = vinfo_for_stmt (stmt);
1502 if (!STMT_VINFO_RELEVANT (stmt_info))
1505 /* For interleaving, only the alignment of the first access
1507 if (STMT_VINFO_STRIDED_ACCESS (stmt_info)
1508 && GROUP_FIRST_ELEMENT (stmt_info) != stmt)
1511 /* For invariant accesses there is nothing to enhance. */
1512 if (integer_zerop (DR_STEP (dr)))
1515 supportable_dr_alignment = vect_supportable_dr_alignment (dr, true);
1516 do_peeling = vector_alignment_reachable_p (dr);
1519 if (known_alignment_for_access_p (dr))
1521 unsigned int npeel_tmp;
1522 bool negative = tree_int_cst_compare (DR_STEP (dr),
1523 size_zero_node) < 0;
1525 /* Save info about DR in the hash table. */
1526 if (!LOOP_VINFO_PEELING_HTAB (loop_vinfo))
1527 LOOP_VINFO_PEELING_HTAB (loop_vinfo) =
1528 htab_create (1, vect_peeling_hash,
1529 vect_peeling_hash_eq, free);
1531 vectype = STMT_VINFO_VECTYPE (stmt_info);
1532 nelements = TYPE_VECTOR_SUBPARTS (vectype);
1533 mis = DR_MISALIGNMENT (dr) / GET_MODE_SIZE (TYPE_MODE (
1534 TREE_TYPE (DR_REF (dr))));
1535 npeel_tmp = (negative
1536 ? (mis - nelements) : (nelements - mis))
1539 /* For multiple types, it is possible that the bigger type access
1540 will have more than one peeling option. E.g., a loop with two
1541 types: one of size (vector size / 4), and the other one of
1542 size (vector size / 8). Vectorization factor will 8. If both
1543 access are misaligned by 3, the first one needs one scalar
1544 iteration to be aligned, and the second one needs 5. But the
1545 the first one will be aligned also by peeling 5 scalar
1546 iterations, and in that case both accesses will be aligned.
1547 Hence, except for the immediate peeling amount, we also want
1548 to try to add full vector size, while we don't exceed
1549 vectorization factor.
1550 We do this automtically for cost model, since we calculate cost
1551 for every peeling option. */
1552 if (!flag_vect_cost_model)
1553 possible_npeel_number = vf /nelements;
1555 /* Handle the aligned case. We may decide to align some other
1556 access, making DR unaligned. */
1557 if (DR_MISALIGNMENT (dr) == 0)
1560 if (!flag_vect_cost_model)
1561 possible_npeel_number++;
1564 for (j = 0; j < possible_npeel_number; j++)
1566 gcc_assert (npeel_tmp <= vf);
1567 vect_peeling_hash_insert (loop_vinfo, dr, npeel_tmp);
1568 npeel_tmp += nelements;
1571 all_misalignments_unknown = false;
1572 /* Data-ref that was chosen for the case that all the
1573 misalignments are unknown is not relevant anymore, since we
1574 have a data-ref with known alignment. */
1579 /* If we don't know all the misalignment values, we prefer
1580 peeling for data-ref that has maximum number of data-refs
1581 with the same alignment, unless the target prefers to align
1582 stores over load. */
1583 if (all_misalignments_unknown)
1585 if (same_align_drs_max < VEC_length (dr_p,
1586 STMT_VINFO_SAME_ALIGN_REFS (stmt_info))
1589 same_align_drs_max = VEC_length (dr_p,
1590 STMT_VINFO_SAME_ALIGN_REFS (stmt_info));
1594 if (!first_store && DR_IS_WRITE (dr))
1598 /* If there are both known and unknown misaligned accesses in the
1599 loop, we choose peeling amount according to the known
1603 if (!supportable_dr_alignment)
1606 if (!first_store && DR_IS_WRITE (dr))
1613 if (!aligned_access_p (dr))
1615 if (vect_print_dump_info (REPORT_DETAILS))
1616 fprintf (vect_dump, "vector alignment may not be reachable");
1623 vect_versioning_for_alias_required
1624 = LOOP_REQUIRES_VERSIONING_FOR_ALIAS (loop_vinfo);
1626 /* Temporarily, if versioning for alias is required, we disable peeling
1627 until we support peeling and versioning. Often peeling for alignment
1628 will require peeling for loop-bound, which in turn requires that we
1629 know how to adjust the loop ivs after the loop. */
1630 if (vect_versioning_for_alias_required
1631 || !vect_can_advance_ivs_p (loop_vinfo)
1632 || !slpeel_can_duplicate_loop_p (loop, single_exit (loop)))
1635 if (do_peeling && all_misalignments_unknown
1636 && vect_supportable_dr_alignment (dr0, false))
1639 /* Check if the target requires to prefer stores over loads, i.e., if
1640 misaligned stores are more expensive than misaligned loads (taking
1641 drs with same alignment into account). */
1642 if (first_store && DR_IS_READ (dr0))
1644 unsigned int load_inside_cost = 0, load_outside_cost = 0;
1645 unsigned int store_inside_cost = 0, store_outside_cost = 0;
1646 unsigned int load_inside_penalty = 0, load_outside_penalty = 0;
1647 unsigned int store_inside_penalty = 0, store_outside_penalty = 0;
1649 vect_get_data_access_cost (dr0, &load_inside_cost,
1650 &load_outside_cost);
1651 vect_get_data_access_cost (first_store, &store_inside_cost,
1652 &store_outside_cost);
1654 /* Calculate the penalty for leaving FIRST_STORE unaligned (by
1655 aligning the load DR0). */
1656 load_inside_penalty = store_inside_cost;
1657 load_outside_penalty = store_outside_cost;
1658 for (i = 0; VEC_iterate (dr_p, STMT_VINFO_SAME_ALIGN_REFS
1659 (vinfo_for_stmt (DR_STMT (first_store))),
1662 if (DR_IS_READ (dr))
1664 load_inside_penalty += load_inside_cost;
1665 load_outside_penalty += load_outside_cost;
1669 load_inside_penalty += store_inside_cost;
1670 load_outside_penalty += store_outside_cost;
1673 /* Calculate the penalty for leaving DR0 unaligned (by
1674 aligning the FIRST_STORE). */
1675 store_inside_penalty = load_inside_cost;
1676 store_outside_penalty = load_outside_cost;
1677 for (i = 0; VEC_iterate (dr_p, STMT_VINFO_SAME_ALIGN_REFS
1678 (vinfo_for_stmt (DR_STMT (dr0))),
1681 if (DR_IS_READ (dr))
1683 store_inside_penalty += load_inside_cost;
1684 store_outside_penalty += load_outside_cost;
1688 store_inside_penalty += store_inside_cost;
1689 store_outside_penalty += store_outside_cost;
1692 if (load_inside_penalty > store_inside_penalty
1693 || (load_inside_penalty == store_inside_penalty
1694 && load_outside_penalty > store_outside_penalty))
1698 /* In case there are only loads with different unknown misalignments, use
1699 peeling only if it may help to align other accesses in the loop. */
1700 if (!first_store && !VEC_length (dr_p, STMT_VINFO_SAME_ALIGN_REFS
1701 (vinfo_for_stmt (DR_STMT (dr0))))
1702 && vect_supportable_dr_alignment (dr0, false)
1703 != dr_unaligned_supported)
1707 if (do_peeling && !dr0)
1709 /* Peeling is possible, but there is no data access that is not supported
1710 unless aligned. So we try to choose the best possible peeling. */
1712 /* We should get here only if there are drs with known misalignment. */
1713 gcc_assert (!all_misalignments_unknown);
1715 /* Choose the best peeling from the hash table. */
1716 dr0 = vect_peeling_hash_choose_best_peeling (loop_vinfo, &npeel);
1723 stmt = DR_STMT (dr0);
1724 stmt_info = vinfo_for_stmt (stmt);
1725 vectype = STMT_VINFO_VECTYPE (stmt_info);
1726 nelements = TYPE_VECTOR_SUBPARTS (vectype);
1728 if (known_alignment_for_access_p (dr0))
1730 bool negative = tree_int_cst_compare (DR_STEP (dr0),
1731 size_zero_node) < 0;
1734 /* Since it's known at compile time, compute the number of
1735 iterations in the peeled loop (the peeling factor) for use in
1736 updating DR_MISALIGNMENT values. The peeling factor is the
1737 vectorization factor minus the misalignment as an element
1739 mis = DR_MISALIGNMENT (dr0);
1740 mis /= GET_MODE_SIZE (TYPE_MODE (TREE_TYPE (DR_REF (dr0))));
1741 npeel = ((negative ? mis - nelements : nelements - mis)
1745 /* For interleaved data access every iteration accesses all the
1746 members of the group, therefore we divide the number of iterations
1747 by the group size. */
1748 stmt_info = vinfo_for_stmt (DR_STMT (dr0));
1749 if (STMT_VINFO_STRIDED_ACCESS (stmt_info))
1750 npeel /= GROUP_SIZE (stmt_info);
1752 if (vect_print_dump_info (REPORT_DETAILS))
1753 fprintf (vect_dump, "Try peeling by %d", npeel);
1756 /* Ensure that all data refs can be vectorized after the peel. */
1757 FOR_EACH_VEC_ELT (data_reference_p, datarefs, i, dr)
1759 int save_misalignment;
1764 stmt = DR_STMT (dr);
1765 stmt_info = vinfo_for_stmt (stmt);
1766 /* For interleaving, only the alignment of the first access
1768 if (STMT_VINFO_STRIDED_ACCESS (stmt_info)
1769 && GROUP_FIRST_ELEMENT (stmt_info) != stmt)
1772 save_misalignment = DR_MISALIGNMENT (dr);
1773 vect_update_misalignment_for_peel (dr, dr0, npeel);
1774 supportable_dr_alignment = vect_supportable_dr_alignment (dr, false);
1775 SET_DR_MISALIGNMENT (dr, save_misalignment);
1777 if (!supportable_dr_alignment)
1784 if (do_peeling && known_alignment_for_access_p (dr0) && npeel == 0)
1786 stat = vect_verify_datarefs_alignment (loop_vinfo, NULL);
1795 /* (1.2) Update the DR_MISALIGNMENT of each data reference DR_i.
1796 If the misalignment of DR_i is identical to that of dr0 then set
1797 DR_MISALIGNMENT (DR_i) to zero. If the misalignment of DR_i and
1798 dr0 are known at compile time then increment DR_MISALIGNMENT (DR_i)
1799 by the peeling factor times the element size of DR_i (MOD the
1800 vectorization factor times the size). Otherwise, the
1801 misalignment of DR_i must be set to unknown. */
1802 FOR_EACH_VEC_ELT (data_reference_p, datarefs, i, dr)
1804 vect_update_misalignment_for_peel (dr, dr0, npeel);
1806 LOOP_VINFO_UNALIGNED_DR (loop_vinfo) = dr0;
1808 LOOP_PEELING_FOR_ALIGNMENT (loop_vinfo) = npeel;
1810 LOOP_PEELING_FOR_ALIGNMENT (loop_vinfo) = DR_MISALIGNMENT (dr0);
1811 SET_DR_MISALIGNMENT (dr0, 0);
1812 if (vect_print_dump_info (REPORT_ALIGNMENT))
1813 fprintf (vect_dump, "Alignment of access forced using peeling.");
1815 if (vect_print_dump_info (REPORT_DETAILS))
1816 fprintf (vect_dump, "Peeling for alignment will be applied.");
1818 stat = vect_verify_datarefs_alignment (loop_vinfo, NULL);
1825 /* (2) Versioning to force alignment. */
1827 /* Try versioning if:
1828 1) flag_tree_vect_loop_version is TRUE
1829 2) optimize loop for speed
1830 3) there is at least one unsupported misaligned data ref with an unknown
1832 4) all misaligned data refs with a known misalignment are supported, and
1833 5) the number of runtime alignment checks is within reason. */
1836 flag_tree_vect_loop_version
1837 && optimize_loop_nest_for_speed_p (loop)
1838 && (!loop->inner); /* FORNOW */
1842 FOR_EACH_VEC_ELT (data_reference_p, datarefs, i, dr)
1844 stmt = DR_STMT (dr);
1845 stmt_info = vinfo_for_stmt (stmt);
1847 /* For interleaving, only the alignment of the first access
1849 if (aligned_access_p (dr)
1850 || (STMT_VINFO_STRIDED_ACCESS (stmt_info)
1851 && GROUP_FIRST_ELEMENT (stmt_info) != stmt))
1854 supportable_dr_alignment = vect_supportable_dr_alignment (dr, false);
1856 if (!supportable_dr_alignment)
1862 if (known_alignment_for_access_p (dr)
1863 || VEC_length (gimple,
1864 LOOP_VINFO_MAY_MISALIGN_STMTS (loop_vinfo))
1865 >= (unsigned) PARAM_VALUE (PARAM_VECT_MAX_VERSION_FOR_ALIGNMENT_CHECKS))
1867 do_versioning = false;
1871 stmt = DR_STMT (dr);
1872 vectype = STMT_VINFO_VECTYPE (vinfo_for_stmt (stmt));
1873 gcc_assert (vectype);
1875 /* The rightmost bits of an aligned address must be zeros.
1876 Construct the mask needed for this test. For example,
1877 GET_MODE_SIZE for the vector mode V4SI is 16 bytes so the
1878 mask must be 15 = 0xf. */
1879 mask = GET_MODE_SIZE (TYPE_MODE (vectype)) - 1;
1881 /* FORNOW: use the same mask to test all potentially unaligned
1882 references in the loop. The vectorizer currently supports
1883 a single vector size, see the reference to
1884 GET_MODE_NUNITS (TYPE_MODE (vectype)) where the
1885 vectorization factor is computed. */
1886 gcc_assert (!LOOP_VINFO_PTR_MASK (loop_vinfo)
1887 || LOOP_VINFO_PTR_MASK (loop_vinfo) == mask);
1888 LOOP_VINFO_PTR_MASK (loop_vinfo) = mask;
1889 VEC_safe_push (gimple, heap,
1890 LOOP_VINFO_MAY_MISALIGN_STMTS (loop_vinfo),
1895 /* Versioning requires at least one misaligned data reference. */
1896 if (!LOOP_REQUIRES_VERSIONING_FOR_ALIGNMENT (loop_vinfo))
1897 do_versioning = false;
1898 else if (!do_versioning)
1899 VEC_truncate (gimple, LOOP_VINFO_MAY_MISALIGN_STMTS (loop_vinfo), 0);
1904 VEC(gimple,heap) *may_misalign_stmts
1905 = LOOP_VINFO_MAY_MISALIGN_STMTS (loop_vinfo);
1908 /* It can now be assumed that the data references in the statements
1909 in LOOP_VINFO_MAY_MISALIGN_STMTS will be aligned in the version
1910 of the loop being vectorized. */
1911 FOR_EACH_VEC_ELT (gimple, may_misalign_stmts, i, stmt)
1913 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
1914 dr = STMT_VINFO_DATA_REF (stmt_info);
1915 SET_DR_MISALIGNMENT (dr, 0);
1916 if (vect_print_dump_info (REPORT_ALIGNMENT))
1917 fprintf (vect_dump, "Alignment of access forced using versioning.");
1920 if (vect_print_dump_info (REPORT_DETAILS))
1921 fprintf (vect_dump, "Versioning for alignment will be applied.");
1923 /* Peeling and versioning can't be done together at this time. */
1924 gcc_assert (! (do_peeling && do_versioning));
1926 stat = vect_verify_datarefs_alignment (loop_vinfo, NULL);
1931 /* This point is reached if neither peeling nor versioning is being done. */
1932 gcc_assert (! (do_peeling || do_versioning));
1934 stat = vect_verify_datarefs_alignment (loop_vinfo, NULL);
1939 /* Function vect_find_same_alignment_drs.
1941 Update group and alignment relations according to the chosen
1942 vectorization factor. */
1945 vect_find_same_alignment_drs (struct data_dependence_relation *ddr,
1946 loop_vec_info loop_vinfo)
1949 struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
1950 int vectorization_factor = LOOP_VINFO_VECT_FACTOR (loop_vinfo);
1951 struct data_reference *dra = DDR_A (ddr);
1952 struct data_reference *drb = DDR_B (ddr);
1953 stmt_vec_info stmtinfo_a = vinfo_for_stmt (DR_STMT (dra));
1954 stmt_vec_info stmtinfo_b = vinfo_for_stmt (DR_STMT (drb));
1955 int dra_size = GET_MODE_SIZE (TYPE_MODE (TREE_TYPE (DR_REF (dra))));
1956 int drb_size = GET_MODE_SIZE (TYPE_MODE (TREE_TYPE (DR_REF (drb))));
1957 lambda_vector dist_v;
1958 unsigned int loop_depth;
1960 if (DDR_ARE_DEPENDENT (ddr) == chrec_known)
1966 if (DDR_ARE_DEPENDENT (ddr) == chrec_dont_know)
1969 /* Loop-based vectorization and known data dependence. */
1970 if (DDR_NUM_DIST_VECTS (ddr) == 0)
1973 /* Data-dependence analysis reports a distance vector of zero
1974 for data-references that overlap only in the first iteration
1975 but have different sign step (see PR45764).
1976 So as a sanity check require equal DR_STEP. */
1977 if (!operand_equal_p (DR_STEP (dra), DR_STEP (drb), 0))
1980 loop_depth = index_in_loop_nest (loop->num, DDR_LOOP_NEST (ddr));
1981 FOR_EACH_VEC_ELT (lambda_vector, DDR_DIST_VECTS (ddr), i, dist_v)
1983 int dist = dist_v[loop_depth];
1985 if (vect_print_dump_info (REPORT_DR_DETAILS))
1986 fprintf (vect_dump, "dependence distance = %d.", dist);
1988 /* Same loop iteration. */
1990 || (dist % vectorization_factor == 0 && dra_size == drb_size))
1992 /* Two references with distance zero have the same alignment. */
1993 VEC_safe_push (dr_p, heap, STMT_VINFO_SAME_ALIGN_REFS (stmtinfo_a), drb);
1994 VEC_safe_push (dr_p, heap, STMT_VINFO_SAME_ALIGN_REFS (stmtinfo_b), dra);
1995 if (vect_print_dump_info (REPORT_ALIGNMENT))
1996 fprintf (vect_dump, "accesses have the same alignment.");
1997 if (vect_print_dump_info (REPORT_DR_DETAILS))
1999 fprintf (vect_dump, "dependence distance modulo vf == 0 between ");
2000 print_generic_expr (vect_dump, DR_REF (dra), TDF_SLIM);
2001 fprintf (vect_dump, " and ");
2002 print_generic_expr (vect_dump, DR_REF (drb), TDF_SLIM);
2009 /* Function vect_analyze_data_refs_alignment
2011 Analyze the alignment of the data-references in the loop.
2012 Return FALSE if a data reference is found that cannot be vectorized. */
2015 vect_analyze_data_refs_alignment (loop_vec_info loop_vinfo,
2016 bb_vec_info bb_vinfo)
2018 if (vect_print_dump_info (REPORT_DETAILS))
2019 fprintf (vect_dump, "=== vect_analyze_data_refs_alignment ===");
2021 /* Mark groups of data references with same alignment using
2022 data dependence information. */
2025 VEC (ddr_p, heap) *ddrs = LOOP_VINFO_DDRS (loop_vinfo);
2026 struct data_dependence_relation *ddr;
2029 FOR_EACH_VEC_ELT (ddr_p, ddrs, i, ddr)
2030 vect_find_same_alignment_drs (ddr, loop_vinfo);
2033 if (!vect_compute_data_refs_alignment (loop_vinfo, bb_vinfo))
2035 if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS))
2037 "not vectorized: can't calculate alignment for data ref.");
2045 /* Analyze groups of strided accesses: check that DR belongs to a group of
2046 strided accesses of legal size, step, etc. Detect gaps, single element
2047 interleaving, and other special cases. Set strided access info.
2048 Collect groups of strided stores for further use in SLP analysis. */
2051 vect_analyze_group_access (struct data_reference *dr)
2053 tree step = DR_STEP (dr);
2054 tree scalar_type = TREE_TYPE (DR_REF (dr));
2055 HOST_WIDE_INT type_size = TREE_INT_CST_LOW (TYPE_SIZE_UNIT (scalar_type));
2056 gimple stmt = DR_STMT (dr);
2057 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
2058 loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info);
2059 bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info);
2060 HOST_WIDE_INT dr_step = TREE_INT_CST_LOW (step);
2061 HOST_WIDE_INT stride, last_accessed_element = 1;
2062 bool slp_impossible = false;
2063 struct loop *loop = NULL;
2066 loop = LOOP_VINFO_LOOP (loop_vinfo);
2068 /* For interleaving, STRIDE is STEP counted in elements, i.e., the size of the
2069 interleaving group (including gaps). */
2070 stride = dr_step / type_size;
2072 /* Not consecutive access is possible only if it is a part of interleaving. */
2073 if (!GROUP_FIRST_ELEMENT (vinfo_for_stmt (stmt)))
2075 /* Check if it this DR is a part of interleaving, and is a single
2076 element of the group that is accessed in the loop. */
2078 /* Gaps are supported only for loads. STEP must be a multiple of the type
2079 size. The size of the group must be a power of 2. */
2081 && (dr_step % type_size) == 0
2083 && exact_log2 (stride) != -1)
2085 GROUP_FIRST_ELEMENT (vinfo_for_stmt (stmt)) = stmt;
2086 GROUP_SIZE (vinfo_for_stmt (stmt)) = stride;
2087 if (vect_print_dump_info (REPORT_DR_DETAILS))
2089 fprintf (vect_dump, "Detected single element interleaving ");
2090 print_generic_expr (vect_dump, DR_REF (dr), TDF_SLIM);
2091 fprintf (vect_dump, " step ");
2092 print_generic_expr (vect_dump, step, TDF_SLIM);
2097 if (vect_print_dump_info (REPORT_DETAILS))
2098 fprintf (vect_dump, "Data access with gaps requires scalar "
2102 if (vect_print_dump_info (REPORT_DETAILS))
2103 fprintf (vect_dump, "Peeling for outer loop is not"
2108 LOOP_VINFO_PEELING_FOR_GAPS (loop_vinfo) = true;
2114 if (vect_print_dump_info (REPORT_DETAILS))
2116 fprintf (vect_dump, "not consecutive access ");
2117 print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM);
2122 /* Mark the statement as unvectorizable. */
2123 STMT_VINFO_VECTORIZABLE (vinfo_for_stmt (DR_STMT (dr))) = false;
2130 if (GROUP_FIRST_ELEMENT (vinfo_for_stmt (stmt)) == stmt)
2132 /* First stmt in the interleaving chain. Check the chain. */
2133 gimple next = GROUP_NEXT_ELEMENT (vinfo_for_stmt (stmt));
2134 struct data_reference *data_ref = dr;
2135 unsigned int count = 1;
2137 tree prev_init = DR_INIT (data_ref);
2139 HOST_WIDE_INT diff, count_in_bytes, gaps = 0;
2143 /* Skip same data-refs. In case that two or more stmts share
2144 data-ref (supported only for loads), we vectorize only the first
2145 stmt, and the rest get their vectorized loads from the first
2147 if (!tree_int_cst_compare (DR_INIT (data_ref),
2148 DR_INIT (STMT_VINFO_DATA_REF (
2149 vinfo_for_stmt (next)))))
2151 if (DR_IS_WRITE (data_ref))
2153 if (vect_print_dump_info (REPORT_DETAILS))
2154 fprintf (vect_dump, "Two store stmts share the same dr.");
2158 /* Check that there is no load-store dependencies for this loads
2159 to prevent a case of load-store-load to the same location. */
2160 if (GROUP_READ_WRITE_DEPENDENCE (vinfo_for_stmt (next))
2161 || GROUP_READ_WRITE_DEPENDENCE (vinfo_for_stmt (prev)))
2163 if (vect_print_dump_info (REPORT_DETAILS))
2165 "READ_WRITE dependence in interleaving.");
2169 /* For load use the same data-ref load. */
2170 GROUP_SAME_DR_STMT (vinfo_for_stmt (next)) = prev;
2173 next = GROUP_NEXT_ELEMENT (vinfo_for_stmt (next));
2179 /* Check that all the accesses have the same STEP. */
2180 next_step = DR_STEP (STMT_VINFO_DATA_REF (vinfo_for_stmt (next)));
2181 if (tree_int_cst_compare (step, next_step))
2183 if (vect_print_dump_info (REPORT_DETAILS))
2184 fprintf (vect_dump, "not consecutive access in interleaving");
2188 data_ref = STMT_VINFO_DATA_REF (vinfo_for_stmt (next));
2189 /* Check that the distance between two accesses is equal to the type
2190 size. Otherwise, we have gaps. */
2191 diff = (TREE_INT_CST_LOW (DR_INIT (data_ref))
2192 - TREE_INT_CST_LOW (prev_init)) / type_size;
2195 /* FORNOW: SLP of accesses with gaps is not supported. */
2196 slp_impossible = true;
2197 if (DR_IS_WRITE (data_ref))
2199 if (vect_print_dump_info (REPORT_DETAILS))
2200 fprintf (vect_dump, "interleaved store with gaps");
2207 last_accessed_element += diff;
2209 /* Store the gap from the previous member of the group. If there is no
2210 gap in the access, GROUP_GAP is always 1. */
2211 GROUP_GAP (vinfo_for_stmt (next)) = diff;
2213 prev_init = DR_INIT (data_ref);
2214 next = GROUP_NEXT_ELEMENT (vinfo_for_stmt (next));
2215 /* Count the number of data-refs in the chain. */
2219 /* COUNT is the number of accesses found, we multiply it by the size of
2220 the type to get COUNT_IN_BYTES. */
2221 count_in_bytes = type_size * count;
2223 /* Check that the size of the interleaving (including gaps) is not
2224 greater than STEP. */
2225 if (dr_step && dr_step < count_in_bytes + gaps * type_size)
2227 if (vect_print_dump_info (REPORT_DETAILS))
2229 fprintf (vect_dump, "interleaving size is greater than step for ");
2230 print_generic_expr (vect_dump, DR_REF (dr), TDF_SLIM);
2235 /* Check that the size of the interleaving is equal to STEP for stores,
2236 i.e., that there are no gaps. */
2237 if (dr_step && dr_step != count_in_bytes)
2239 if (DR_IS_READ (dr))
2241 slp_impossible = true;
2242 /* There is a gap after the last load in the group. This gap is a
2243 difference between the stride and the number of elements. When
2244 there is no gap, this difference should be 0. */
2245 GROUP_GAP (vinfo_for_stmt (stmt)) = stride - count;
2249 if (vect_print_dump_info (REPORT_DETAILS))
2250 fprintf (vect_dump, "interleaved store with gaps");
2255 /* Check that STEP is a multiple of type size. */
2256 if (dr_step && (dr_step % type_size) != 0)
2258 if (vect_print_dump_info (REPORT_DETAILS))
2260 fprintf (vect_dump, "step is not a multiple of type size: step ");
2261 print_generic_expr (vect_dump, step, TDF_SLIM);
2262 fprintf (vect_dump, " size ");
2263 print_generic_expr (vect_dump, TYPE_SIZE_UNIT (scalar_type),
2272 GROUP_SIZE (vinfo_for_stmt (stmt)) = stride;
2273 if (vect_print_dump_info (REPORT_DETAILS))
2274 fprintf (vect_dump, "Detected interleaving of size %d", (int)stride);
2276 /* SLP: create an SLP data structure for every interleaving group of
2277 stores for further analysis in vect_analyse_slp. */
2278 if (DR_IS_WRITE (dr) && !slp_impossible)
2281 VEC_safe_push (gimple, heap, LOOP_VINFO_STRIDED_STORES (loop_vinfo),
2284 VEC_safe_push (gimple, heap, BB_VINFO_STRIDED_STORES (bb_vinfo),
2288 /* There is a gap in the end of the group. */
2289 if (stride - last_accessed_element > 0 && loop_vinfo)
2291 if (vect_print_dump_info (REPORT_DETAILS))
2292 fprintf (vect_dump, "Data access with gaps requires scalar "
2296 if (vect_print_dump_info (REPORT_DETAILS))
2297 fprintf (vect_dump, "Peeling for outer loop is not supported");
2301 LOOP_VINFO_PEELING_FOR_GAPS (loop_vinfo) = true;
2309 /* Analyze the access pattern of the data-reference DR.
2310 In case of non-consecutive accesses call vect_analyze_group_access() to
2311 analyze groups of strided accesses. */
2314 vect_analyze_data_ref_access (struct data_reference *dr)
2316 tree step = DR_STEP (dr);
2317 tree scalar_type = TREE_TYPE (DR_REF (dr));
2318 gimple stmt = DR_STMT (dr);
2319 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
2320 loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info);
2321 struct loop *loop = NULL;
2322 HOST_WIDE_INT dr_step = TREE_INT_CST_LOW (step);
2325 loop = LOOP_VINFO_LOOP (loop_vinfo);
2327 if (loop_vinfo && !step)
2329 if (vect_print_dump_info (REPORT_DETAILS))
2330 fprintf (vect_dump, "bad data-ref access in loop");
2334 /* Allow invariant loads in loops. */
2335 if (loop_vinfo && dr_step == 0)
2337 GROUP_FIRST_ELEMENT (vinfo_for_stmt (stmt)) = NULL;
2338 return DR_IS_READ (dr);
2341 if (loop && nested_in_vect_loop_p (loop, stmt))
2343 /* Interleaved accesses are not yet supported within outer-loop
2344 vectorization for references in the inner-loop. */
2345 GROUP_FIRST_ELEMENT (vinfo_for_stmt (stmt)) = NULL;
2347 /* For the rest of the analysis we use the outer-loop step. */
2348 step = STMT_VINFO_DR_STEP (stmt_info);
2349 dr_step = TREE_INT_CST_LOW (step);
2353 if (vect_print_dump_info (REPORT_ALIGNMENT))
2354 fprintf (vect_dump, "zero step in outer loop.");
2355 if (DR_IS_READ (dr))
2363 if (!tree_int_cst_compare (step, TYPE_SIZE_UNIT (scalar_type))
2365 && !compare_tree_int (TYPE_SIZE_UNIT (scalar_type), -dr_step)))
2367 /* Mark that it is not interleaving. */
2368 GROUP_FIRST_ELEMENT (vinfo_for_stmt (stmt)) = NULL;
2372 if (loop && nested_in_vect_loop_p (loop, stmt))
2374 if (vect_print_dump_info (REPORT_ALIGNMENT))
2375 fprintf (vect_dump, "strided access in outer loop.");
2379 /* Not consecutive access - check if it's a part of interleaving group. */
2380 return vect_analyze_group_access (dr);
2384 /* Function vect_analyze_data_ref_accesses.
2386 Analyze the access pattern of all the data references in the loop.
2388 FORNOW: the only access pattern that is considered vectorizable is a
2389 simple step 1 (consecutive) access.
2391 FORNOW: handle only arrays and pointer accesses. */
2394 vect_analyze_data_ref_accesses (loop_vec_info loop_vinfo, bb_vec_info bb_vinfo)
2397 VEC (data_reference_p, heap) *datarefs;
2398 struct data_reference *dr;
2400 if (vect_print_dump_info (REPORT_DETAILS))
2401 fprintf (vect_dump, "=== vect_analyze_data_ref_accesses ===");
2404 datarefs = LOOP_VINFO_DATAREFS (loop_vinfo);
2406 datarefs = BB_VINFO_DATAREFS (bb_vinfo);
2408 FOR_EACH_VEC_ELT (data_reference_p, datarefs, i, dr)
2409 if (STMT_VINFO_VECTORIZABLE (vinfo_for_stmt (DR_STMT (dr)))
2410 && !vect_analyze_data_ref_access (dr))
2412 if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS))
2413 fprintf (vect_dump, "not vectorized: complicated access pattern.");
2417 /* Mark the statement as not vectorizable. */
2418 STMT_VINFO_VECTORIZABLE (vinfo_for_stmt (DR_STMT (dr))) = false;
2428 /* Function vect_prune_runtime_alias_test_list.
2430 Prune a list of ddrs to be tested at run-time by versioning for alias.
2431 Return FALSE if resulting list of ddrs is longer then allowed by
2432 PARAM_VECT_MAX_VERSION_FOR_ALIAS_CHECKS, otherwise return TRUE. */
2435 vect_prune_runtime_alias_test_list (loop_vec_info loop_vinfo)
2437 VEC (ddr_p, heap) * ddrs =
2438 LOOP_VINFO_MAY_ALIAS_DDRS (loop_vinfo);
2441 if (vect_print_dump_info (REPORT_DETAILS))
2442 fprintf (vect_dump, "=== vect_prune_runtime_alias_test_list ===");
2444 for (i = 0; i < VEC_length (ddr_p, ddrs); )
2449 ddr_i = VEC_index (ddr_p, ddrs, i);
2452 for (j = 0; j < i; j++)
2454 ddr_p ddr_j = VEC_index (ddr_p, ddrs, j);
2456 if (vect_vfa_range_equal (ddr_i, ddr_j))
2458 if (vect_print_dump_info (REPORT_DR_DETAILS))
2460 fprintf (vect_dump, "found equal ranges ");
2461 print_generic_expr (vect_dump, DR_REF (DDR_A (ddr_i)), TDF_SLIM);
2462 fprintf (vect_dump, ", ");
2463 print_generic_expr (vect_dump, DR_REF (DDR_B (ddr_i)), TDF_SLIM);
2464 fprintf (vect_dump, " and ");
2465 print_generic_expr (vect_dump, DR_REF (DDR_A (ddr_j)), TDF_SLIM);
2466 fprintf (vect_dump, ", ");
2467 print_generic_expr (vect_dump, DR_REF (DDR_B (ddr_j)), TDF_SLIM);
2476 VEC_ordered_remove (ddr_p, ddrs, i);
2482 if (VEC_length (ddr_p, ddrs) >
2483 (unsigned) PARAM_VALUE (PARAM_VECT_MAX_VERSION_FOR_ALIAS_CHECKS))
2485 if (vect_print_dump_info (REPORT_DR_DETAILS))
2488 "disable versioning for alias - max number of generated "
2489 "checks exceeded.");
2492 VEC_truncate (ddr_p, LOOP_VINFO_MAY_ALIAS_DDRS (loop_vinfo), 0);
2501 /* Function vect_analyze_data_refs.
2503 Find all the data references in the loop or basic block.
2505 The general structure of the analysis of data refs in the vectorizer is as
2507 1- vect_analyze_data_refs(loop/bb): call
2508 compute_data_dependences_for_loop/bb to find and analyze all data-refs
2509 in the loop/bb and their dependences.
2510 2- vect_analyze_dependences(): apply dependence testing using ddrs.
2511 3- vect_analyze_drs_alignment(): check that ref_stmt.alignment is ok.
2512 4- vect_analyze_drs_access(): check that ref_stmt.step is ok.
2517 vect_analyze_data_refs (loop_vec_info loop_vinfo,
2518 bb_vec_info bb_vinfo,
2521 struct loop *loop = NULL;
2522 basic_block bb = NULL;
2524 VEC (data_reference_p, heap) *datarefs;
2525 struct data_reference *dr;
2529 if (vect_print_dump_info (REPORT_DETAILS))
2530 fprintf (vect_dump, "=== vect_analyze_data_refs ===\n");
2534 loop = LOOP_VINFO_LOOP (loop_vinfo);
2535 res = compute_data_dependences_for_loop
2537 &LOOP_VINFO_LOOP_NEST (loop_vinfo),
2538 &LOOP_VINFO_DATAREFS (loop_vinfo),
2539 &LOOP_VINFO_DDRS (loop_vinfo));
2543 if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS))
2544 fprintf (vect_dump, "not vectorized: loop contains function calls"
2545 " or data references that cannot be analyzed");
2549 datarefs = LOOP_VINFO_DATAREFS (loop_vinfo);
2553 bb = BB_VINFO_BB (bb_vinfo);
2554 res = compute_data_dependences_for_bb (bb, true,
2555 &BB_VINFO_DATAREFS (bb_vinfo),
2556 &BB_VINFO_DDRS (bb_vinfo));
2559 if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS))
2560 fprintf (vect_dump, "not vectorized: basic block contains function"
2561 " calls or data references that cannot be analyzed");
2565 datarefs = BB_VINFO_DATAREFS (bb_vinfo);
2568 /* Go through the data-refs, check that the analysis succeeded. Update
2569 pointer from stmt_vec_info struct to DR and vectype. */
2571 FOR_EACH_VEC_ELT (data_reference_p, datarefs, i, dr)
2574 stmt_vec_info stmt_info;
2575 tree base, offset, init;
2578 if (!dr || !DR_REF (dr))
2580 if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS))
2581 fprintf (vect_dump, "not vectorized: unhandled data-ref ");
2585 stmt = DR_STMT (dr);
2586 stmt_info = vinfo_for_stmt (stmt);
2588 /* Check that analysis of the data-ref succeeded. */
2589 if (!DR_BASE_ADDRESS (dr) || !DR_OFFSET (dr) || !DR_INIT (dr)
2592 if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS))
2594 fprintf (vect_dump, "not vectorized: data ref analysis failed ");
2595 print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM);
2601 if (TREE_CODE (DR_BASE_ADDRESS (dr)) == INTEGER_CST)
2603 if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS))
2604 fprintf (vect_dump, "not vectorized: base addr of dr is a "
2609 if (TREE_THIS_VOLATILE (DR_REF (dr)))
2611 if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS))
2613 fprintf (vect_dump, "not vectorized: volatile type ");
2614 print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM);
2619 base = unshare_expr (DR_BASE_ADDRESS (dr));
2620 offset = unshare_expr (DR_OFFSET (dr));
2621 init = unshare_expr (DR_INIT (dr));
2623 if (stmt_can_throw_internal (stmt))
2625 if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS))
2627 fprintf (vect_dump, "not vectorized: statement can throw an "
2629 print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM);
2634 /* Update DR field in stmt_vec_info struct. */
2636 /* If the dataref is in an inner-loop of the loop that is considered for
2637 for vectorization, we also want to analyze the access relative to
2638 the outer-loop (DR contains information only relative to the
2639 inner-most enclosing loop). We do that by building a reference to the
2640 first location accessed by the inner-loop, and analyze it relative to
2642 if (loop && nested_in_vect_loop_p (loop, stmt))
2644 tree outer_step, outer_base, outer_init;
2645 HOST_WIDE_INT pbitsize, pbitpos;
2647 enum machine_mode pmode;
2648 int punsignedp, pvolatilep;
2649 affine_iv base_iv, offset_iv;
2652 /* Build a reference to the first location accessed by the
2653 inner-loop: *(BASE+INIT). (The first location is actually
2654 BASE+INIT+OFFSET, but we add OFFSET separately later). */
2655 tree inner_base = build_fold_indirect_ref
2656 (fold_build_pointer_plus (base, init));
2658 if (vect_print_dump_info (REPORT_DETAILS))
2660 fprintf (vect_dump, "analyze in outer-loop: ");
2661 print_generic_expr (vect_dump, inner_base, TDF_SLIM);
2664 outer_base = get_inner_reference (inner_base, &pbitsize, &pbitpos,
2665 &poffset, &pmode, &punsignedp, &pvolatilep, false);
2666 gcc_assert (outer_base != NULL_TREE);
2668 if (pbitpos % BITS_PER_UNIT != 0)
2670 if (vect_print_dump_info (REPORT_DETAILS))
2671 fprintf (vect_dump, "failed: bit offset alignment.\n");
2675 outer_base = build_fold_addr_expr (outer_base);
2676 if (!simple_iv (loop, loop_containing_stmt (stmt), outer_base,
2679 if (vect_print_dump_info (REPORT_DETAILS))
2680 fprintf (vect_dump, "failed: evolution of base is not affine.\n");
2687 poffset = fold_build2 (PLUS_EXPR, TREE_TYPE (offset), offset,
2695 offset_iv.base = ssize_int (0);
2696 offset_iv.step = ssize_int (0);
2698 else if (!simple_iv (loop, loop_containing_stmt (stmt), poffset,
2701 if (vect_print_dump_info (REPORT_DETAILS))
2702 fprintf (vect_dump, "evolution of offset is not affine.\n");
2706 outer_init = ssize_int (pbitpos / BITS_PER_UNIT);
2707 split_constant_offset (base_iv.base, &base_iv.base, &dinit);
2708 outer_init = size_binop (PLUS_EXPR, outer_init, dinit);
2709 split_constant_offset (offset_iv.base, &offset_iv.base, &dinit);
2710 outer_init = size_binop (PLUS_EXPR, outer_init, dinit);
2712 outer_step = size_binop (PLUS_EXPR,
2713 fold_convert (ssizetype, base_iv.step),
2714 fold_convert (ssizetype, offset_iv.step));
2716 STMT_VINFO_DR_STEP (stmt_info) = outer_step;
2717 /* FIXME: Use canonicalize_base_object_address (base_iv.base); */
2718 STMT_VINFO_DR_BASE_ADDRESS (stmt_info) = base_iv.base;
2719 STMT_VINFO_DR_INIT (stmt_info) = outer_init;
2720 STMT_VINFO_DR_OFFSET (stmt_info) =
2721 fold_convert (ssizetype, offset_iv.base);
2722 STMT_VINFO_DR_ALIGNED_TO (stmt_info) =
2723 size_int (highest_pow2_factor (offset_iv.base));
2725 if (vect_print_dump_info (REPORT_DETAILS))
2727 fprintf (vect_dump, "\touter base_address: ");
2728 print_generic_expr (vect_dump, STMT_VINFO_DR_BASE_ADDRESS (stmt_info), TDF_SLIM);
2729 fprintf (vect_dump, "\n\touter offset from base address: ");
2730 print_generic_expr (vect_dump, STMT_VINFO_DR_OFFSET (stmt_info), TDF_SLIM);
2731 fprintf (vect_dump, "\n\touter constant offset from base address: ");
2732 print_generic_expr (vect_dump, STMT_VINFO_DR_INIT (stmt_info), TDF_SLIM);
2733 fprintf (vect_dump, "\n\touter step: ");
2734 print_generic_expr (vect_dump, STMT_VINFO_DR_STEP (stmt_info), TDF_SLIM);
2735 fprintf (vect_dump, "\n\touter aligned to: ");
2736 print_generic_expr (vect_dump, STMT_VINFO_DR_ALIGNED_TO (stmt_info), TDF_SLIM);
2740 if (STMT_VINFO_DATA_REF (stmt_info))
2742 if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS))
2745 "not vectorized: more than one data ref in stmt: ");
2746 print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM);
2751 STMT_VINFO_DATA_REF (stmt_info) = dr;
2753 /* Set vectype for STMT. */
2754 scalar_type = TREE_TYPE (DR_REF (dr));
2755 STMT_VINFO_VECTYPE (stmt_info) =
2756 get_vectype_for_scalar_type (scalar_type);
2757 if (!STMT_VINFO_VECTYPE (stmt_info))
2759 if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS))
2762 "not vectorized: no vectype for stmt: ");
2763 print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM);
2764 fprintf (vect_dump, " scalar_type: ");
2765 print_generic_expr (vect_dump, scalar_type, TDF_DETAILS);
2770 /* Mark the statement as not vectorizable. */
2771 STMT_VINFO_VECTORIZABLE (stmt_info) = false;
2778 /* Adjust the minimal vectorization factor according to the
2780 vf = TYPE_VECTOR_SUBPARTS (STMT_VINFO_VECTYPE (stmt_info));
2789 /* Function vect_get_new_vect_var.
2791 Returns a name for a new variable. The current naming scheme appends the
2792 prefix "vect_" or "vect_p" (depending on the value of VAR_KIND) to
2793 the name of vectorizer generated variables, and appends that to NAME if
2797 vect_get_new_vect_var (tree type, enum vect_var_kind var_kind, const char *name)
2804 case vect_simple_var:
2807 case vect_scalar_var:
2810 case vect_pointer_var:
2819 char* tmp = concat (prefix, name, NULL);
2820 new_vect_var = create_tmp_var (type, tmp);
2824 new_vect_var = create_tmp_var (type, prefix);
2826 /* Mark vector typed variable as a gimple register variable. */
2827 if (TREE_CODE (type) == VECTOR_TYPE)
2828 DECL_GIMPLE_REG_P (new_vect_var) = true;
2830 return new_vect_var;
2834 /* Function vect_create_addr_base_for_vector_ref.
2836 Create an expression that computes the address of the first memory location
2837 that will be accessed for a data reference.
2840 STMT: The statement containing the data reference.
2841 NEW_STMT_LIST: Must be initialized to NULL_TREE or a statement list.
2842 OFFSET: Optional. If supplied, it is be added to the initial address.
2843 LOOP: Specify relative to which loop-nest should the address be computed.
2844 For example, when the dataref is in an inner-loop nested in an
2845 outer-loop that is now being vectorized, LOOP can be either the
2846 outer-loop, or the inner-loop. The first memory location accessed
2847 by the following dataref ('in' points to short):
2854 if LOOP=i_loop: &in (relative to i_loop)
2855 if LOOP=j_loop: &in+i*2B (relative to j_loop)
2858 1. Return an SSA_NAME whose value is the address of the memory location of
2859 the first vector of the data reference.
2860 2. If new_stmt_list is not NULL_TREE after return then the caller must insert
2861 these statement(s) which define the returned SSA_NAME.
2863 FORNOW: We are only handling array accesses with step 1. */
2866 vect_create_addr_base_for_vector_ref (gimple stmt,
2867 gimple_seq *new_stmt_list,
2871 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
2872 struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info);
2873 tree data_ref_base = unshare_expr (DR_BASE_ADDRESS (dr));
2875 tree data_ref_base_var;
2877 tree addr_base, addr_expr;
2879 gimple_seq seq = NULL;
2880 tree base_offset = unshare_expr (DR_OFFSET (dr));
2881 tree init = unshare_expr (DR_INIT (dr));
2883 tree step = TYPE_SIZE_UNIT (TREE_TYPE (DR_REF (dr)));
2884 loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info);
2887 if (loop_vinfo && loop && loop != (gimple_bb (stmt))->loop_father)
2889 struct loop *outer_loop = LOOP_VINFO_LOOP (loop_vinfo);
2891 gcc_assert (nested_in_vect_loop_p (outer_loop, stmt));
2893 data_ref_base = unshare_expr (STMT_VINFO_DR_BASE_ADDRESS (stmt_info));
2894 base_offset = unshare_expr (STMT_VINFO_DR_OFFSET (stmt_info));
2895 init = unshare_expr (STMT_VINFO_DR_INIT (stmt_info));
2899 base_name = build_fold_indirect_ref (data_ref_base);
2902 base_offset = ssize_int (0);
2903 init = ssize_int (0);
2904 base_name = build_fold_indirect_ref (unshare_expr (DR_REF (dr)));
2907 data_ref_base_var = create_tmp_var (TREE_TYPE (data_ref_base), "batmp");
2908 add_referenced_var (data_ref_base_var);
2909 data_ref_base = force_gimple_operand (data_ref_base, &seq, true,
2911 gimple_seq_add_seq (new_stmt_list, seq);
2913 /* Create base_offset */
2914 base_offset = size_binop (PLUS_EXPR,
2915 fold_convert (sizetype, base_offset),
2916 fold_convert (sizetype, init));
2917 dest = create_tmp_var (sizetype, "base_off");
2918 add_referenced_var (dest);
2919 base_offset = force_gimple_operand (base_offset, &seq, true, dest);
2920 gimple_seq_add_seq (new_stmt_list, seq);
2924 tree tmp = create_tmp_var (sizetype, "offset");
2926 add_referenced_var (tmp);
2927 offset = fold_build2 (MULT_EXPR, sizetype,
2928 fold_convert (sizetype, offset), step);
2929 base_offset = fold_build2 (PLUS_EXPR, sizetype,
2930 base_offset, offset);
2931 base_offset = force_gimple_operand (base_offset, &seq, false, tmp);
2932 gimple_seq_add_seq (new_stmt_list, seq);
2935 /* base + base_offset */
2937 addr_base = fold_build_pointer_plus (data_ref_base, base_offset);
2940 addr_base = build1 (ADDR_EXPR,
2941 build_pointer_type (TREE_TYPE (DR_REF (dr))),
2942 unshare_expr (DR_REF (dr)));
2945 vect_ptr_type = build_pointer_type (STMT_VINFO_VECTYPE (stmt_info));
2946 base = get_base_address (DR_REF (dr));
2948 && TREE_CODE (base) == MEM_REF)
2950 = build_qualified_type (vect_ptr_type,
2951 TYPE_QUALS (TREE_TYPE (TREE_OPERAND (base, 0))));
2953 vec_stmt = fold_convert (vect_ptr_type, addr_base);
2954 addr_expr = vect_get_new_vect_var (vect_ptr_type, vect_pointer_var,
2955 get_name (base_name));
2956 add_referenced_var (addr_expr);
2957 vec_stmt = force_gimple_operand (vec_stmt, &seq, false, addr_expr);
2958 gimple_seq_add_seq (new_stmt_list, seq);
2960 if (DR_PTR_INFO (dr)
2961 && TREE_CODE (vec_stmt) == SSA_NAME)
2963 duplicate_ssa_name_ptr_info (vec_stmt, DR_PTR_INFO (dr));
2966 SSA_NAME_PTR_INFO (vec_stmt)->align = 1;
2967 SSA_NAME_PTR_INFO (vec_stmt)->misalign = 0;
2971 if (vect_print_dump_info (REPORT_DETAILS))
2973 fprintf (vect_dump, "created ");
2974 print_generic_expr (vect_dump, vec_stmt, TDF_SLIM);
2981 /* Function vect_create_data_ref_ptr.
2983 Create a new pointer-to-AGGR_TYPE variable (ap), that points to the first
2984 location accessed in the loop by STMT, along with the def-use update
2985 chain to appropriately advance the pointer through the loop iterations.
2986 Also set aliasing information for the pointer. This pointer is used by
2987 the callers to this function to create a memory reference expression for
2988 vector load/store access.
2991 1. STMT: a stmt that references memory. Expected to be of the form
2992 GIMPLE_ASSIGN <name, data-ref> or
2993 GIMPLE_ASSIGN <data-ref, name>.
2994 2. AGGR_TYPE: the type of the reference, which should be either a vector
2996 3. AT_LOOP: the loop where the vector memref is to be created.
2997 4. OFFSET (optional): an offset to be added to the initial address accessed
2998 by the data-ref in STMT.
2999 5. BSI: location where the new stmts are to be placed if there is no loop
3000 6. ONLY_INIT: indicate if ap is to be updated in the loop, or remain
3001 pointing to the initial address.
3004 1. Declare a new ptr to vector_type, and have it point to the base of the
3005 data reference (initial addressed accessed by the data reference).
3006 For example, for vector of type V8HI, the following code is generated:
3009 ap = (v8hi *)initial_address;
3011 if OFFSET is not supplied:
3012 initial_address = &a[init];
3013 if OFFSET is supplied:
3014 initial_address = &a[init + OFFSET];
3016 Return the initial_address in INITIAL_ADDRESS.
3018 2. If ONLY_INIT is true, just return the initial pointer. Otherwise, also
3019 update the pointer in each iteration of the loop.
3021 Return the increment stmt that updates the pointer in PTR_INCR.
3023 3. Set INV_P to true if the access pattern of the data reference in the
3024 vectorized loop is invariant. Set it to false otherwise.
3026 4. Return the pointer. */
3029 vect_create_data_ref_ptr (gimple stmt, tree aggr_type, struct loop *at_loop,
3030 tree offset, tree *initial_address,
3031 gimple_stmt_iterator *gsi, gimple *ptr_incr,
3032 bool only_init, bool *inv_p)
3035 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
3036 loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info);
3037 struct loop *loop = NULL;
3038 bool nested_in_vect_loop = false;
3039 struct loop *containing_loop = NULL;
3044 gimple_seq new_stmt_list = NULL;
3048 struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info);
3050 gimple_stmt_iterator incr_gsi;
3053 tree indx_before_incr, indx_after_incr;
3056 bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info);
3059 gcc_assert (TREE_CODE (aggr_type) == ARRAY_TYPE
3060 || TREE_CODE (aggr_type) == VECTOR_TYPE);
3064 loop = LOOP_VINFO_LOOP (loop_vinfo);
3065 nested_in_vect_loop = nested_in_vect_loop_p (loop, stmt);
3066 containing_loop = (gimple_bb (stmt))->loop_father;
3067 pe = loop_preheader_edge (loop);
3071 gcc_assert (bb_vinfo);
3076 /* Check the step (evolution) of the load in LOOP, and record
3077 whether it's invariant. */
3078 if (nested_in_vect_loop)
3079 step = STMT_VINFO_DR_STEP (stmt_info);
3081 step = DR_STEP (STMT_VINFO_DATA_REF (stmt_info));
3083 if (tree_int_cst_compare (step, size_zero_node) == 0)
3087 negative = tree_int_cst_compare (step, size_zero_node) < 0;
3089 /* Create an expression for the first address accessed by this load
3091 base_name = build_fold_indirect_ref (unshare_expr (DR_BASE_ADDRESS (dr)));
3093 if (vect_print_dump_info (REPORT_DETAILS))
3095 tree data_ref_base = base_name;
3096 fprintf (vect_dump, "create %s-pointer variable to type: ",
3097 tree_code_name[(int) TREE_CODE (aggr_type)]);
3098 print_generic_expr (vect_dump, aggr_type, TDF_SLIM);
3099 if (TREE_CODE (data_ref_base) == VAR_DECL
3100 || TREE_CODE (data_ref_base) == ARRAY_REF)
3101 fprintf (vect_dump, " vectorizing an array ref: ");
3102 else if (TREE_CODE (data_ref_base) == COMPONENT_REF)
3103 fprintf (vect_dump, " vectorizing a record based array ref: ");
3104 else if (TREE_CODE (data_ref_base) == SSA_NAME)
3105 fprintf (vect_dump, " vectorizing a pointer ref: ");
3106 print_generic_expr (vect_dump, base_name, TDF_SLIM);
3109 /* (1) Create the new aggregate-pointer variable. */
3110 aggr_ptr_type = build_pointer_type (aggr_type);
3111 base = get_base_address (DR_REF (dr));
3113 && TREE_CODE (base) == MEM_REF)
3115 = build_qualified_type (aggr_ptr_type,
3116 TYPE_QUALS (TREE_TYPE (TREE_OPERAND (base, 0))));
3117 aggr_ptr = vect_get_new_vect_var (aggr_ptr_type, vect_pointer_var,
3118 get_name (base_name));
3120 /* Vector and array types inherit the alias set of their component
3121 type by default so we need to use a ref-all pointer if the data
3122 reference does not conflict with the created aggregated data
3123 reference because it is not addressable. */
3124 if (!alias_sets_conflict_p (get_deref_alias_set (aggr_ptr),
3125 get_alias_set (DR_REF (dr))))
3128 = build_pointer_type_for_mode (aggr_type,
3129 TYPE_MODE (aggr_ptr_type), true);
3130 aggr_ptr = vect_get_new_vect_var (aggr_ptr_type, vect_pointer_var,
3131 get_name (base_name));
3134 /* Likewise for any of the data references in the stmt group. */
3135 else if (STMT_VINFO_GROUP_SIZE (stmt_info) > 1)
3137 gimple orig_stmt = STMT_VINFO_GROUP_FIRST_ELEMENT (stmt_info);
3140 tree lhs = gimple_assign_lhs (orig_stmt);
3141 if (!alias_sets_conflict_p (get_deref_alias_set (aggr_ptr),
3142 get_alias_set (lhs)))
3145 = build_pointer_type_for_mode (aggr_type,
3146 TYPE_MODE (aggr_ptr_type), true);
3148 = vect_get_new_vect_var (aggr_ptr_type, vect_pointer_var,
3149 get_name (base_name));
3153 orig_stmt = STMT_VINFO_GROUP_NEXT_ELEMENT (vinfo_for_stmt (orig_stmt));
3158 add_referenced_var (aggr_ptr);
3160 /* Note: If the dataref is in an inner-loop nested in LOOP, and we are
3161 vectorizing LOOP (i.e., outer-loop vectorization), we need to create two
3162 def-use update cycles for the pointer: one relative to the outer-loop
3163 (LOOP), which is what steps (3) and (4) below do. The other is relative
3164 to the inner-loop (which is the inner-most loop containing the dataref),
3165 and this is done be step (5) below.
3167 When vectorizing inner-most loops, the vectorized loop (LOOP) is also the
3168 inner-most loop, and so steps (3),(4) work the same, and step (5) is
3169 redundant. Steps (3),(4) create the following:
3172 LOOP: vp1 = phi(vp0,vp2)
3178 If there is an inner-loop nested in loop, then step (5) will also be
3179 applied, and an additional update in the inner-loop will be created:
3182 LOOP: vp1 = phi(vp0,vp2)
3184 inner: vp3 = phi(vp1,vp4)
3185 vp4 = vp3 + inner_step
3191 /* (2) Calculate the initial address of the aggregate-pointer, and set
3192 the aggregate-pointer to point to it before the loop. */
3194 /* Create: (&(base[init_val+offset]) in the loop preheader. */
3196 new_temp = vect_create_addr_base_for_vector_ref (stmt, &new_stmt_list,
3202 new_bb = gsi_insert_seq_on_edge_immediate (pe, new_stmt_list);
3203 gcc_assert (!new_bb);
3206 gsi_insert_seq_before (gsi, new_stmt_list, GSI_SAME_STMT);
3209 *initial_address = new_temp;
3211 /* Create: p = (aggr_type *) initial_base */
3212 if (TREE_CODE (new_temp) != SSA_NAME
3213 || !useless_type_conversion_p (aggr_ptr_type, TREE_TYPE (new_temp)))
3215 vec_stmt = gimple_build_assign (aggr_ptr,
3216 fold_convert (aggr_ptr_type, new_temp));
3217 aggr_ptr_init = make_ssa_name (aggr_ptr, vec_stmt);
3218 /* Copy the points-to information if it exists. */
3219 if (DR_PTR_INFO (dr))
3220 duplicate_ssa_name_ptr_info (aggr_ptr_init, DR_PTR_INFO (dr));
3221 gimple_assign_set_lhs (vec_stmt, aggr_ptr_init);
3224 new_bb = gsi_insert_on_edge_immediate (pe, vec_stmt);
3225 gcc_assert (!new_bb);
3228 gsi_insert_before (gsi, vec_stmt, GSI_SAME_STMT);
3231 aggr_ptr_init = new_temp;
3233 /* (3) Handle the updating of the aggregate-pointer inside the loop.
3234 This is needed when ONLY_INIT is false, and also when AT_LOOP is the
3235 inner-loop nested in LOOP (during outer-loop vectorization). */
3237 /* No update in loop is required. */
3238 if (only_init && (!loop_vinfo || at_loop == loop))
3239 aptr = aggr_ptr_init;
3242 /* The step of the aggregate pointer is the type size. */
3243 tree step = TYPE_SIZE_UNIT (aggr_type);
3244 /* One exception to the above is when the scalar step of the load in
3245 LOOP is zero. In this case the step here is also zero. */
3247 step = size_zero_node;
3249 step = fold_build1 (NEGATE_EXPR, TREE_TYPE (step), step);
3251 standard_iv_increment_position (loop, &incr_gsi, &insert_after);
3253 create_iv (aggr_ptr_init,
3254 fold_convert (aggr_ptr_type, step),
3255 aggr_ptr, loop, &incr_gsi, insert_after,
3256 &indx_before_incr, &indx_after_incr);
3257 incr = gsi_stmt (incr_gsi);
3258 set_vinfo_for_stmt (incr, new_stmt_vec_info (incr, loop_vinfo, NULL));
3260 /* Copy the points-to information if it exists. */
3261 if (DR_PTR_INFO (dr))
3263 duplicate_ssa_name_ptr_info (indx_before_incr, DR_PTR_INFO (dr));
3264 duplicate_ssa_name_ptr_info (indx_after_incr, DR_PTR_INFO (dr));
3269 aptr = indx_before_incr;
3272 if (!nested_in_vect_loop || only_init)
3276 /* (4) Handle the updating of the aggregate-pointer inside the inner-loop
3277 nested in LOOP, if exists. */
3279 gcc_assert (nested_in_vect_loop);
3282 standard_iv_increment_position (containing_loop, &incr_gsi,
3284 create_iv (aptr, fold_convert (aggr_ptr_type, DR_STEP (dr)), aggr_ptr,
3285 containing_loop, &incr_gsi, insert_after, &indx_before_incr,
3287 incr = gsi_stmt (incr_gsi);
3288 set_vinfo_for_stmt (incr, new_stmt_vec_info (incr, loop_vinfo, NULL));
3290 /* Copy the points-to information if it exists. */
3291 if (DR_PTR_INFO (dr))
3293 duplicate_ssa_name_ptr_info (indx_before_incr, DR_PTR_INFO (dr));
3294 duplicate_ssa_name_ptr_info (indx_after_incr, DR_PTR_INFO (dr));
3299 return indx_before_incr;
3306 /* Function bump_vector_ptr
3308 Increment a pointer (to a vector type) by vector-size. If requested,
3309 i.e. if PTR-INCR is given, then also connect the new increment stmt
3310 to the existing def-use update-chain of the pointer, by modifying
3311 the PTR_INCR as illustrated below:
3313 The pointer def-use update-chain before this function:
3314 DATAREF_PTR = phi (p_0, p_2)
3316 PTR_INCR: p_2 = DATAREF_PTR + step
3318 The pointer def-use update-chain after this function:
3319 DATAREF_PTR = phi (p_0, p_2)
3321 NEW_DATAREF_PTR = DATAREF_PTR + BUMP
3323 PTR_INCR: p_2 = NEW_DATAREF_PTR + step
3326 DATAREF_PTR - ssa_name of a pointer (to vector type) that is being updated
3328 PTR_INCR - optional. The stmt that updates the pointer in each iteration of
3329 the loop. The increment amount across iterations is expected
3331 BSI - location where the new update stmt is to be placed.
3332 STMT - the original scalar memory-access stmt that is being vectorized.
3333 BUMP - optional. The offset by which to bump the pointer. If not given,
3334 the offset is assumed to be vector_size.
3336 Output: Return NEW_DATAREF_PTR as illustrated above.
3341 bump_vector_ptr (tree dataref_ptr, gimple ptr_incr, gimple_stmt_iterator *gsi,
3342 gimple stmt, tree bump)
3344 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
3345 struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info);
3346 tree vectype = STMT_VINFO_VECTYPE (stmt_info);
3347 tree ptr_var = SSA_NAME_VAR (dataref_ptr);
3348 tree update = TYPE_SIZE_UNIT (vectype);
3351 use_operand_p use_p;
3352 tree new_dataref_ptr;
3357 incr_stmt = gimple_build_assign_with_ops (POINTER_PLUS_EXPR, ptr_var,
3358 dataref_ptr, update);
3359 new_dataref_ptr = make_ssa_name (ptr_var, incr_stmt);
3360 gimple_assign_set_lhs (incr_stmt, new_dataref_ptr);
3361 vect_finish_stmt_generation (stmt, incr_stmt, gsi);
3363 /* Copy the points-to information if it exists. */
3364 if (DR_PTR_INFO (dr))
3366 duplicate_ssa_name_ptr_info (new_dataref_ptr, DR_PTR_INFO (dr));
3367 SSA_NAME_PTR_INFO (new_dataref_ptr)->align = 1;
3368 SSA_NAME_PTR_INFO (new_dataref_ptr)->misalign = 0;
3372 return new_dataref_ptr;
3374 /* Update the vector-pointer's cross-iteration increment. */
3375 FOR_EACH_SSA_USE_OPERAND (use_p, ptr_incr, iter, SSA_OP_USE)
3377 tree use = USE_FROM_PTR (use_p);
3379 if (use == dataref_ptr)
3380 SET_USE (use_p, new_dataref_ptr);
3382 gcc_assert (tree_int_cst_compare (use, update) == 0);
3385 return new_dataref_ptr;
3389 /* Function vect_create_destination_var.
3391 Create a new temporary of type VECTYPE. */
3394 vect_create_destination_var (tree scalar_dest, tree vectype)
3397 const char *new_name;
3399 enum vect_var_kind kind;
3401 kind = vectype ? vect_simple_var : vect_scalar_var;
3402 type = vectype ? vectype : TREE_TYPE (scalar_dest);
3404 gcc_assert (TREE_CODE (scalar_dest) == SSA_NAME);
3406 new_name = get_name (scalar_dest);
3409 vec_dest = vect_get_new_vect_var (type, kind, new_name);
3410 add_referenced_var (vec_dest);
3415 /* Function vect_strided_store_supported.
3417 Returns TRUE is INTERLEAVE_HIGH and INTERLEAVE_LOW operations are supported,
3418 and FALSE otherwise. */
3421 vect_strided_store_supported (tree vectype, unsigned HOST_WIDE_INT count)
3423 optab interleave_high_optab, interleave_low_optab;
3424 enum machine_mode mode;
3426 mode = TYPE_MODE (vectype);
3428 /* vect_permute_store_chain requires the group size to be a power of two. */
3429 if (exact_log2 (count) == -1)
3431 if (vect_print_dump_info (REPORT_DETAILS))
3432 fprintf (vect_dump, "the size of the group of strided accesses"
3433 " is not a power of 2");
3437 /* Check that the operation is supported. */
3438 interleave_high_optab = optab_for_tree_code (VEC_INTERLEAVE_HIGH_EXPR,
3439 vectype, optab_default);
3440 interleave_low_optab = optab_for_tree_code (VEC_INTERLEAVE_LOW_EXPR,
3441 vectype, optab_default);
3442 if (!interleave_high_optab || !interleave_low_optab)
3444 if (vect_print_dump_info (REPORT_DETAILS))
3445 fprintf (vect_dump, "no optab for interleave.");
3449 if (optab_handler (interleave_high_optab, mode) == CODE_FOR_nothing
3450 || optab_handler (interleave_low_optab, mode) == CODE_FOR_nothing)
3452 if (vect_print_dump_info (REPORT_DETAILS))
3453 fprintf (vect_dump, "interleave op not supported by target.");
3461 /* Return TRUE if vec_store_lanes is available for COUNT vectors of
3465 vect_store_lanes_supported (tree vectype, unsigned HOST_WIDE_INT count)
3467 return vect_lanes_optab_supported_p ("vec_store_lanes",
3468 vec_store_lanes_optab,
3473 /* Function vect_permute_store_chain.
3475 Given a chain of interleaved stores in DR_CHAIN of LENGTH that must be
3476 a power of 2, generate interleave_high/low stmts to reorder the data
3477 correctly for the stores. Return the final references for stores in
3480 E.g., LENGTH is 4 and the scalar type is short, i.e., VF is 8.
3481 The input is 4 vectors each containing 8 elements. We assign a number to
3482 each element, the input sequence is:
3484 1st vec: 0 1 2 3 4 5 6 7
3485 2nd vec: 8 9 10 11 12 13 14 15
3486 3rd vec: 16 17 18 19 20 21 22 23
3487 4th vec: 24 25 26 27 28 29 30 31
3489 The output sequence should be:
3491 1st vec: 0 8 16 24 1 9 17 25
3492 2nd vec: 2 10 18 26 3 11 19 27
3493 3rd vec: 4 12 20 28 5 13 21 30
3494 4th vec: 6 14 22 30 7 15 23 31
3496 i.e., we interleave the contents of the four vectors in their order.
3498 We use interleave_high/low instructions to create such output. The input of
3499 each interleave_high/low operation is two vectors:
3502 the even elements of the result vector are obtained left-to-right from the
3503 high/low elements of the first vector. The odd elements of the result are
3504 obtained left-to-right from the high/low elements of the second vector.
3505 The output of interleave_high will be: 0 4 1 5
3506 and of interleave_low: 2 6 3 7
3509 The permutation is done in log LENGTH stages. In each stage interleave_high
3510 and interleave_low stmts are created for each pair of vectors in DR_CHAIN,
3511 where the first argument is taken from the first half of DR_CHAIN and the
3512 second argument from it's second half.
3515 I1: interleave_high (1st vec, 3rd vec)
3516 I2: interleave_low (1st vec, 3rd vec)
3517 I3: interleave_high (2nd vec, 4th vec)
3518 I4: interleave_low (2nd vec, 4th vec)
3520 The output for the first stage is:
3522 I1: 0 16 1 17 2 18 3 19
3523 I2: 4 20 5 21 6 22 7 23
3524 I3: 8 24 9 25 10 26 11 27
3525 I4: 12 28 13 29 14 30 15 31
3527 The output of the second stage, i.e. the final result is:
3529 I1: 0 8 16 24 1 9 17 25
3530 I2: 2 10 18 26 3 11 19 27
3531 I3: 4 12 20 28 5 13 21 30
3532 I4: 6 14 22 30 7 15 23 31. */
3535 vect_permute_store_chain (VEC(tree,heap) *dr_chain,
3536 unsigned int length,
3538 gimple_stmt_iterator *gsi,
3539 VEC(tree,heap) **result_chain)
3541 tree perm_dest, vect1, vect2, high, low;
3543 tree vectype = STMT_VINFO_VECTYPE (vinfo_for_stmt (stmt));
3546 enum tree_code high_code, low_code;
3548 gcc_assert (vect_strided_store_supported (vectype, length));
3550 *result_chain = VEC_copy (tree, heap, dr_chain);
3552 for (i = 0; i < exact_log2 (length); i++)
3554 for (j = 0; j < length/2; j++)
3556 vect1 = VEC_index (tree, dr_chain, j);
3557 vect2 = VEC_index (tree, dr_chain, j+length/2);
3559 /* Create interleaving stmt:
3560 in the case of big endian:
3561 high = interleave_high (vect1, vect2)
3562 and in the case of little endian:
3563 high = interleave_low (vect1, vect2). */
3564 perm_dest = create_tmp_var (vectype, "vect_inter_high");
3565 DECL_GIMPLE_REG_P (perm_dest) = 1;
3566 add_referenced_var (perm_dest);
3567 if (BYTES_BIG_ENDIAN)
3569 high_code = VEC_INTERLEAVE_HIGH_EXPR;
3570 low_code = VEC_INTERLEAVE_LOW_EXPR;
3574 low_code = VEC_INTERLEAVE_HIGH_EXPR;
3575 high_code = VEC_INTERLEAVE_LOW_EXPR;
3577 perm_stmt = gimple_build_assign_with_ops (high_code, perm_dest,
3579 high = make_ssa_name (perm_dest, perm_stmt);
3580 gimple_assign_set_lhs (perm_stmt, high);
3581 vect_finish_stmt_generation (stmt, perm_stmt, gsi);
3582 VEC_replace (tree, *result_chain, 2*j, high);
3584 /* Create interleaving stmt:
3585 in the case of big endian:
3586 low = interleave_low (vect1, vect2)
3587 and in the case of little endian:
3588 low = interleave_high (vect1, vect2). */
3589 perm_dest = create_tmp_var (vectype, "vect_inter_low");
3590 DECL_GIMPLE_REG_P (perm_dest) = 1;
3591 add_referenced_var (perm_dest);
3592 perm_stmt = gimple_build_assign_with_ops (low_code, perm_dest,
3594 low = make_ssa_name (perm_dest, perm_stmt);
3595 gimple_assign_set_lhs (perm_stmt, low);
3596 vect_finish_stmt_generation (stmt, perm_stmt, gsi);
3597 VEC_replace (tree, *result_chain, 2*j+1, low);
3599 dr_chain = VEC_copy (tree, heap, *result_chain);
3603 /* Function vect_setup_realignment
3605 This function is called when vectorizing an unaligned load using
3606 the dr_explicit_realign[_optimized] scheme.
3607 This function generates the following code at the loop prolog:
3610 x msq_init = *(floor(p)); # prolog load
3611 realignment_token = call target_builtin;
3613 x msq = phi (msq_init, ---)
3615 The stmts marked with x are generated only for the case of
3616 dr_explicit_realign_optimized.
3618 The code above sets up a new (vector) pointer, pointing to the first
3619 location accessed by STMT, and a "floor-aligned" load using that pointer.
3620 It also generates code to compute the "realignment-token" (if the relevant
3621 target hook was defined), and creates a phi-node at the loop-header bb
3622 whose arguments are the result of the prolog-load (created by this
3623 function) and the result of a load that takes place in the loop (to be
3624 created by the caller to this function).
3626 For the case of dr_explicit_realign_optimized:
3627 The caller to this function uses the phi-result (msq) to create the
3628 realignment code inside the loop, and sets up the missing phi argument,
3631 msq = phi (msq_init, lsq)
3632 lsq = *(floor(p')); # load in loop
3633 result = realign_load (msq, lsq, realignment_token);
3635 For the case of dr_explicit_realign:
3637 msq = *(floor(p)); # load in loop
3639 lsq = *(floor(p')); # load in loop
3640 result = realign_load (msq, lsq, realignment_token);
3643 STMT - (scalar) load stmt to be vectorized. This load accesses
3644 a memory location that may be unaligned.
3645 BSI - place where new code is to be inserted.
3646 ALIGNMENT_SUPPORT_SCHEME - which of the two misalignment handling schemes
3650 REALIGNMENT_TOKEN - the result of a call to the builtin_mask_for_load
3651 target hook, if defined.
3652 Return value - the result of the loop-header phi node. */
3655 vect_setup_realignment (gimple stmt, gimple_stmt_iterator *gsi,
3656 tree *realignment_token,
3657 enum dr_alignment_support alignment_support_scheme,
3659 struct loop **at_loop)
3661 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
3662 tree vectype = STMT_VINFO_VECTYPE (stmt_info);
3663 loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info);
3664 struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info);
3665 struct loop *loop = NULL;
3667 tree scalar_dest = gimple_assign_lhs (stmt);
3674 tree msq_init = NULL_TREE;
3677 tree msq = NULL_TREE;
3678 gimple_seq stmts = NULL;
3680 bool compute_in_loop = false;
3681 bool nested_in_vect_loop = false;
3682 struct loop *containing_loop = (gimple_bb (stmt))->loop_father;
3683 struct loop *loop_for_initial_load = NULL;
3687 loop = LOOP_VINFO_LOOP (loop_vinfo);
3688 nested_in_vect_loop = nested_in_vect_loop_p (loop, stmt);
3691 gcc_assert (alignment_support_scheme == dr_explicit_realign
3692 || alignment_support_scheme == dr_explicit_realign_optimized);
3694 /* We need to generate three things:
3695 1. the misalignment computation
3696 2. the extra vector load (for the optimized realignment scheme).
3697 3. the phi node for the two vectors from which the realignment is
3698 done (for the optimized realignment scheme). */
3700 /* 1. Determine where to generate the misalignment computation.
3702 If INIT_ADDR is NULL_TREE, this indicates that the misalignment
3703 calculation will be generated by this function, outside the loop (in the
3704 preheader). Otherwise, INIT_ADDR had already been computed for us by the
3705 caller, inside the loop.
3707 Background: If the misalignment remains fixed throughout the iterations of
3708 the loop, then both realignment schemes are applicable, and also the
3709 misalignment computation can be done outside LOOP. This is because we are
3710 vectorizing LOOP, and so the memory accesses in LOOP advance in steps that
3711 are a multiple of VS (the Vector Size), and therefore the misalignment in
3712 different vectorized LOOP iterations is always the same.
3713 The problem arises only if the memory access is in an inner-loop nested
3714 inside LOOP, which is now being vectorized using outer-loop vectorization.
3715 This is the only case when the misalignment of the memory access may not
3716 remain fixed throughout the iterations of the inner-loop (as explained in
3717 detail in vect_supportable_dr_alignment). In this case, not only is the
3718 optimized realignment scheme not applicable, but also the misalignment
3719 computation (and generation of the realignment token that is passed to
3720 REALIGN_LOAD) have to be done inside the loop.
3722 In short, INIT_ADDR indicates whether we are in a COMPUTE_IN_LOOP mode
3723 or not, which in turn determines if the misalignment is computed inside
3724 the inner-loop, or outside LOOP. */
3726 if (init_addr != NULL_TREE || !loop_vinfo)
3728 compute_in_loop = true;
3729 gcc_assert (alignment_support_scheme == dr_explicit_realign);
3733 /* 2. Determine where to generate the extra vector load.
3735 For the optimized realignment scheme, instead of generating two vector
3736 loads in each iteration, we generate a single extra vector load in the
3737 preheader of the loop, and in each iteration reuse the result of the
3738 vector load from the previous iteration. In case the memory access is in
3739 an inner-loop nested inside LOOP, which is now being vectorized using
3740 outer-loop vectorization, we need to determine whether this initial vector
3741 load should be generated at the preheader of the inner-loop, or can be
3742 generated at the preheader of LOOP. If the memory access has no evolution
3743 in LOOP, it can be generated in the preheader of LOOP. Otherwise, it has
3744 to be generated inside LOOP (in the preheader of the inner-loop). */
3746 if (nested_in_vect_loop)
3748 tree outerloop_step = STMT_VINFO_DR_STEP (stmt_info);
3749 bool invariant_in_outerloop =
3750 (tree_int_cst_compare (outerloop_step, size_zero_node) == 0);
3751 loop_for_initial_load = (invariant_in_outerloop ? loop : loop->inner);
3754 loop_for_initial_load = loop;
3756 *at_loop = loop_for_initial_load;
3758 if (loop_for_initial_load)
3759 pe = loop_preheader_edge (loop_for_initial_load);
3761 /* 3. For the case of the optimized realignment, create the first vector
3762 load at the loop preheader. */
3764 if (alignment_support_scheme == dr_explicit_realign_optimized)
3766 /* Create msq_init = *(floor(p1)) in the loop preheader */
3768 gcc_assert (!compute_in_loop);
3769 vec_dest = vect_create_destination_var (scalar_dest, vectype);
3770 ptr = vect_create_data_ref_ptr (stmt, vectype, loop_for_initial_load,
3771 NULL_TREE, &init_addr, NULL, &inc,
3773 new_stmt = gimple_build_assign_with_ops
3774 (BIT_AND_EXPR, NULL_TREE, ptr,
3775 build_int_cst (TREE_TYPE (ptr),
3776 -(HOST_WIDE_INT)TYPE_ALIGN_UNIT (vectype)));
3777 new_temp = make_ssa_name (SSA_NAME_VAR (ptr), new_stmt);
3778 gimple_assign_set_lhs (new_stmt, new_temp);
3779 new_bb = gsi_insert_on_edge_immediate (pe, new_stmt);
3780 gcc_assert (!new_bb);
3782 = build2 (MEM_REF, TREE_TYPE (vec_dest), new_temp,
3783 build_int_cst (reference_alias_ptr_type (DR_REF (dr)), 0));
3784 new_stmt = gimple_build_assign (vec_dest, data_ref);
3785 new_temp = make_ssa_name (vec_dest, new_stmt);
3786 gimple_assign_set_lhs (new_stmt, new_temp);
3787 mark_symbols_for_renaming (new_stmt);
3790 new_bb = gsi_insert_on_edge_immediate (pe, new_stmt);
3791 gcc_assert (!new_bb);
3794 gsi_insert_before (gsi, new_stmt, GSI_SAME_STMT);
3796 msq_init = gimple_assign_lhs (new_stmt);
3799 /* 4. Create realignment token using a target builtin, if available.
3800 It is done either inside the containing loop, or before LOOP (as
3801 determined above). */
3803 if (targetm.vectorize.builtin_mask_for_load)
3807 /* Compute INIT_ADDR - the initial addressed accessed by this memref. */
3810 /* Generate the INIT_ADDR computation outside LOOP. */
3811 init_addr = vect_create_addr_base_for_vector_ref (stmt, &stmts,
3815 pe = loop_preheader_edge (loop);
3816 new_bb = gsi_insert_seq_on_edge_immediate (pe, stmts);
3817 gcc_assert (!new_bb);
3820 gsi_insert_seq_before (gsi, stmts, GSI_SAME_STMT);
3823 builtin_decl = targetm.vectorize.builtin_mask_for_load ();
3824 new_stmt = gimple_build_call (builtin_decl, 1, init_addr);
3826 vect_create_destination_var (scalar_dest,
3827 gimple_call_return_type (new_stmt));
3828 new_temp = make_ssa_name (vec_dest, new_stmt);
3829 gimple_call_set_lhs (new_stmt, new_temp);
3831 if (compute_in_loop)
3832 gsi_insert_before (gsi, new_stmt, GSI_SAME_STMT);
3835 /* Generate the misalignment computation outside LOOP. */
3836 pe = loop_preheader_edge (loop);
3837 new_bb = gsi_insert_on_edge_immediate (pe, new_stmt);
3838 gcc_assert (!new_bb);
3841 *realignment_token = gimple_call_lhs (new_stmt);
3843 /* The result of the CALL_EXPR to this builtin is determined from
3844 the value of the parameter and no global variables are touched
3845 which makes the builtin a "const" function. Requiring the
3846 builtin to have the "const" attribute makes it unnecessary
3847 to call mark_call_clobbered. */
3848 gcc_assert (TREE_READONLY (builtin_decl));
3851 if (alignment_support_scheme == dr_explicit_realign)
3854 gcc_assert (!compute_in_loop);
3855 gcc_assert (alignment_support_scheme == dr_explicit_realign_optimized);
3858 /* 5. Create msq = phi <msq_init, lsq> in loop */
3860 pe = loop_preheader_edge (containing_loop);
3861 vec_dest = vect_create_destination_var (scalar_dest, vectype);
3862 msq = make_ssa_name (vec_dest, NULL);
3863 phi_stmt = create_phi_node (msq, containing_loop->header);
3864 SSA_NAME_DEF_STMT (msq) = phi_stmt;
3865 add_phi_arg (phi_stmt, msq_init, pe, UNKNOWN_LOCATION);
3871 /* Function vect_strided_load_supported.
3873 Returns TRUE is EXTRACT_EVEN and EXTRACT_ODD operations are supported,
3874 and FALSE otherwise. */
3877 vect_strided_load_supported (tree vectype, unsigned HOST_WIDE_INT count)
3879 optab perm_even_optab, perm_odd_optab;
3880 enum machine_mode mode;
3882 mode = TYPE_MODE (vectype);
3884 /* vect_permute_load_chain requires the group size to be a power of two. */
3885 if (exact_log2 (count) == -1)
3887 if (vect_print_dump_info (REPORT_DETAILS))
3888 fprintf (vect_dump, "the size of the group of strided accesses"
3889 " is not a power of 2");
3893 perm_even_optab = optab_for_tree_code (VEC_EXTRACT_EVEN_EXPR, vectype,
3895 if (!perm_even_optab)
3897 if (vect_print_dump_info (REPORT_DETAILS))
3898 fprintf (vect_dump, "no optab for perm_even.");
3902 if (optab_handler (perm_even_optab, mode) == CODE_FOR_nothing)
3904 if (vect_print_dump_info (REPORT_DETAILS))
3905 fprintf (vect_dump, "perm_even op not supported by target.");
3909 perm_odd_optab = optab_for_tree_code (VEC_EXTRACT_ODD_EXPR, vectype,
3911 if (!perm_odd_optab)
3913 if (vect_print_dump_info (REPORT_DETAILS))
3914 fprintf (vect_dump, "no optab for perm_odd.");
3918 if (optab_handler (perm_odd_optab, mode) == CODE_FOR_nothing)
3920 if (vect_print_dump_info (REPORT_DETAILS))
3921 fprintf (vect_dump, "perm_odd op not supported by target.");
3927 /* Return TRUE if vec_load_lanes is available for COUNT vectors of
3931 vect_load_lanes_supported (tree vectype, unsigned HOST_WIDE_INT count)
3933 return vect_lanes_optab_supported_p ("vec_load_lanes",
3934 vec_load_lanes_optab,
3938 /* Function vect_permute_load_chain.
3940 Given a chain of interleaved loads in DR_CHAIN of LENGTH that must be
3941 a power of 2, generate extract_even/odd stmts to reorder the input data
3942 correctly. Return the final references for loads in RESULT_CHAIN.
3944 E.g., LENGTH is 4 and the scalar type is short, i.e., VF is 8.
3945 The input is 4 vectors each containing 8 elements. We assign a number to each
3946 element, the input sequence is:
3948 1st vec: 0 1 2 3 4 5 6 7
3949 2nd vec: 8 9 10 11 12 13 14 15
3950 3rd vec: 16 17 18 19 20 21 22 23
3951 4th vec: 24 25 26 27 28 29 30 31
3953 The output sequence should be:
3955 1st vec: 0 4 8 12 16 20 24 28
3956 2nd vec: 1 5 9 13 17 21 25 29
3957 3rd vec: 2 6 10 14 18 22 26 30
3958 4th vec: 3 7 11 15 19 23 27 31
3960 i.e., the first output vector should contain the first elements of each
3961 interleaving group, etc.
3963 We use extract_even/odd instructions to create such output. The input of
3964 each extract_even/odd operation is two vectors
3968 and the output is the vector of extracted even/odd elements. The output of
3969 extract_even will be: 0 2 4 6
3970 and of extract_odd: 1 3 5 7
3973 The permutation is done in log LENGTH stages. In each stage extract_even
3974 and extract_odd stmts are created for each pair of vectors in DR_CHAIN in
3975 their order. In our example,
3977 E1: extract_even (1st vec, 2nd vec)
3978 E2: extract_odd (1st vec, 2nd vec)
3979 E3: extract_even (3rd vec, 4th vec)
3980 E4: extract_odd (3rd vec, 4th vec)
3982 The output for the first stage will be:
3984 E1: 0 2 4 6 8 10 12 14
3985 E2: 1 3 5 7 9 11 13 15
3986 E3: 16 18 20 22 24 26 28 30
3987 E4: 17 19 21 23 25 27 29 31
3989 In order to proceed and create the correct sequence for the next stage (or
3990 for the correct output, if the second stage is the last one, as in our
3991 example), we first put the output of extract_even operation and then the
3992 output of extract_odd in RESULT_CHAIN (which is then copied to DR_CHAIN).
3993 The input for the second stage is:
3995 1st vec (E1): 0 2 4 6 8 10 12 14
3996 2nd vec (E3): 16 18 20 22 24 26 28 30
3997 3rd vec (E2): 1 3 5 7 9 11 13 15
3998 4th vec (E4): 17 19 21 23 25 27 29 31
4000 The output of the second stage:
4002 E1: 0 4 8 12 16 20 24 28
4003 E2: 2 6 10 14 18 22 26 30
4004 E3: 1 5 9 13 17 21 25 29
4005 E4: 3 7 11 15 19 23 27 31
4007 And RESULT_CHAIN after reordering:
4009 1st vec (E1): 0 4 8 12 16 20 24 28
4010 2nd vec (E3): 1 5 9 13 17 21 25 29
4011 3rd vec (E2): 2 6 10 14 18 22 26 30
4012 4th vec (E4): 3 7 11 15 19 23 27 31. */
4015 vect_permute_load_chain (VEC(tree,heap) *dr_chain,
4016 unsigned int length,
4018 gimple_stmt_iterator *gsi,
4019 VEC(tree,heap) **result_chain)
4021 tree perm_dest, data_ref, first_vect, second_vect;
4023 tree vectype = STMT_VINFO_VECTYPE (vinfo_for_stmt (stmt));
4027 gcc_assert (vect_strided_load_supported (vectype, length));
4029 *result_chain = VEC_copy (tree, heap, dr_chain);
4030 for (i = 0; i < exact_log2 (length); i++)
4032 for (j = 0; j < length; j +=2)
4034 first_vect = VEC_index (tree, dr_chain, j);
4035 second_vect = VEC_index (tree, dr_chain, j+1);
4037 /* data_ref = permute_even (first_data_ref, second_data_ref); */
4038 perm_dest = create_tmp_var (vectype, "vect_perm_even");
4039 DECL_GIMPLE_REG_P (perm_dest) = 1;
4040 add_referenced_var (perm_dest);
4042 perm_stmt = gimple_build_assign_with_ops (VEC_EXTRACT_EVEN_EXPR,
4043 perm_dest, first_vect,
4046 data_ref = make_ssa_name (perm_dest, perm_stmt);
4047 gimple_assign_set_lhs (perm_stmt, data_ref);
4048 vect_finish_stmt_generation (stmt, perm_stmt, gsi);
4049 mark_symbols_for_renaming (perm_stmt);
4051 VEC_replace (tree, *result_chain, j/2, data_ref);
4053 /* data_ref = permute_odd (first_data_ref, second_data_ref); */
4054 perm_dest = create_tmp_var (vectype, "vect_perm_odd");
4055 DECL_GIMPLE_REG_P (perm_dest) = 1;
4056 add_referenced_var (perm_dest);
4058 perm_stmt = gimple_build_assign_with_ops (VEC_EXTRACT_ODD_EXPR,
4059 perm_dest, first_vect,
4061 data_ref = make_ssa_name (perm_dest, perm_stmt);
4062 gimple_assign_set_lhs (perm_stmt, data_ref);
4063 vect_finish_stmt_generation (stmt, perm_stmt, gsi);
4064 mark_symbols_for_renaming (perm_stmt);
4066 VEC_replace (tree, *result_chain, j/2+length/2, data_ref);
4068 dr_chain = VEC_copy (tree, heap, *result_chain);
4073 /* Function vect_transform_strided_load.
4075 Given a chain of input interleaved data-refs (in DR_CHAIN), build statements
4076 to perform their permutation and ascribe the result vectorized statements to
4077 the scalar statements.
4081 vect_transform_strided_load (gimple stmt, VEC(tree,heap) *dr_chain, int size,
4082 gimple_stmt_iterator *gsi)
4084 VEC(tree,heap) *result_chain = NULL;
4086 /* DR_CHAIN contains input data-refs that are a part of the interleaving.
4087 RESULT_CHAIN is the output of vect_permute_load_chain, it contains permuted
4088 vectors, that are ready for vector computation. */
4089 result_chain = VEC_alloc (tree, heap, size);
4090 vect_permute_load_chain (dr_chain, size, stmt, gsi, &result_chain);
4091 vect_record_strided_load_vectors (stmt, result_chain);
4092 VEC_free (tree, heap, result_chain);
4095 /* RESULT_CHAIN contains the output of a group of strided loads that were
4096 generated as part of the vectorization of STMT. Assign the statement
4097 for each vector to the associated scalar statement. */
4100 vect_record_strided_load_vectors (gimple stmt, VEC(tree,heap) *result_chain)
4102 gimple first_stmt = GROUP_FIRST_ELEMENT (vinfo_for_stmt (stmt));
4103 gimple next_stmt, new_stmt;
4104 unsigned int i, gap_count;
4107 /* Put a permuted data-ref in the VECTORIZED_STMT field.
4108 Since we scan the chain starting from it's first node, their order
4109 corresponds the order of data-refs in RESULT_CHAIN. */
4110 next_stmt = first_stmt;
4112 FOR_EACH_VEC_ELT (tree, result_chain, i, tmp_data_ref)
4117 /* Skip the gaps. Loads created for the gaps will be removed by dead
4118 code elimination pass later. No need to check for the first stmt in
4119 the group, since it always exists.
4120 GROUP_GAP is the number of steps in elements from the previous
4121 access (if there is no gap GROUP_GAP is 1). We skip loads that
4122 correspond to the gaps. */
4123 if (next_stmt != first_stmt
4124 && gap_count < GROUP_GAP (vinfo_for_stmt (next_stmt)))
4132 new_stmt = SSA_NAME_DEF_STMT (tmp_data_ref);
4133 /* We assume that if VEC_STMT is not NULL, this is a case of multiple
4134 copies, and we put the new vector statement in the first available
4136 if (!STMT_VINFO_VEC_STMT (vinfo_for_stmt (next_stmt)))
4137 STMT_VINFO_VEC_STMT (vinfo_for_stmt (next_stmt)) = new_stmt;
4140 if (!GROUP_SAME_DR_STMT (vinfo_for_stmt (next_stmt)))
4143 STMT_VINFO_VEC_STMT (vinfo_for_stmt (next_stmt));
4145 STMT_VINFO_RELATED_STMT (vinfo_for_stmt (prev_stmt));
4148 prev_stmt = rel_stmt;
4150 STMT_VINFO_RELATED_STMT (vinfo_for_stmt (rel_stmt));
4153 STMT_VINFO_RELATED_STMT (vinfo_for_stmt (prev_stmt)) =
4158 next_stmt = GROUP_NEXT_ELEMENT (vinfo_for_stmt (next_stmt));
4160 /* If NEXT_STMT accesses the same DR as the previous statement,
4161 put the same TMP_DATA_REF as its vectorized statement; otherwise
4162 get the next data-ref from RESULT_CHAIN. */
4163 if (!next_stmt || !GROUP_SAME_DR_STMT (vinfo_for_stmt (next_stmt)))
4169 /* Function vect_force_dr_alignment_p.
4171 Returns whether the alignment of a DECL can be forced to be aligned
4172 on ALIGNMENT bit boundary. */
4175 vect_can_force_dr_alignment_p (const_tree decl, unsigned int alignment)
4177 if (TREE_CODE (decl) != VAR_DECL)
4180 if (DECL_EXTERNAL (decl))
4183 if (TREE_ASM_WRITTEN (decl))
4186 if (TREE_STATIC (decl))
4187 return (alignment <= MAX_OFILE_ALIGNMENT);
4189 return (alignment <= MAX_STACK_ALIGNMENT);
4193 /* Return whether the data reference DR is supported with respect to its
4195 If CHECK_ALIGNED_ACCESSES is TRUE, check if the access is supported even
4196 it is aligned, i.e., check if it is possible to vectorize it with different
4199 enum dr_alignment_support
4200 vect_supportable_dr_alignment (struct data_reference *dr,
4201 bool check_aligned_accesses)
4203 gimple stmt = DR_STMT (dr);
4204 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
4205 tree vectype = STMT_VINFO_VECTYPE (stmt_info);
4206 enum machine_mode mode = TYPE_MODE (vectype);
4207 loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info);
4208 struct loop *vect_loop = NULL;
4209 bool nested_in_vect_loop = false;
4211 if (aligned_access_p (dr) && !check_aligned_accesses)
4216 vect_loop = LOOP_VINFO_LOOP (loop_vinfo);
4217 nested_in_vect_loop = nested_in_vect_loop_p (vect_loop, stmt);
4220 /* Possibly unaligned access. */
4222 /* We can choose between using the implicit realignment scheme (generating
4223 a misaligned_move stmt) and the explicit realignment scheme (generating
4224 aligned loads with a REALIGN_LOAD). There are two variants to the
4225 explicit realignment scheme: optimized, and unoptimized.
4226 We can optimize the realignment only if the step between consecutive
4227 vector loads is equal to the vector size. Since the vector memory
4228 accesses advance in steps of VS (Vector Size) in the vectorized loop, it
4229 is guaranteed that the misalignment amount remains the same throughout the
4230 execution of the vectorized loop. Therefore, we can create the
4231 "realignment token" (the permutation mask that is passed to REALIGN_LOAD)
4232 at the loop preheader.
4234 However, in the case of outer-loop vectorization, when vectorizing a
4235 memory access in the inner-loop nested within the LOOP that is now being
4236 vectorized, while it is guaranteed that the misalignment of the
4237 vectorized memory access will remain the same in different outer-loop
4238 iterations, it is *not* guaranteed that is will remain the same throughout
4239 the execution of the inner-loop. This is because the inner-loop advances
4240 with the original scalar step (and not in steps of VS). If the inner-loop
4241 step happens to be a multiple of VS, then the misalignment remains fixed
4242 and we can use the optimized realignment scheme. For example:
4248 When vectorizing the i-loop in the above example, the step between
4249 consecutive vector loads is 1, and so the misalignment does not remain
4250 fixed across the execution of the inner-loop, and the realignment cannot
4251 be optimized (as illustrated in the following pseudo vectorized loop):
4253 for (i=0; i<N; i+=4)
4254 for (j=0; j<M; j++){
4255 vs += vp[i+j]; // misalignment of &vp[i+j] is {0,1,2,3,0,1,2,3,...}
4256 // when j is {0,1,2,3,4,5,6,7,...} respectively.
4257 // (assuming that we start from an aligned address).
4260 We therefore have to use the unoptimized realignment scheme:
4262 for (i=0; i<N; i+=4)
4263 for (j=k; j<M; j+=4)
4264 vs += vp[i+j]; // misalignment of &vp[i+j] is always k (assuming
4265 // that the misalignment of the initial address is
4268 The loop can then be vectorized as follows:
4270 for (k=0; k<4; k++){
4271 rt = get_realignment_token (&vp[k]);
4272 for (i=0; i<N; i+=4){
4274 for (j=k; j<M; j+=4){
4276 va = REALIGN_LOAD <v1,v2,rt>;
4283 if (DR_IS_READ (dr))
4285 bool is_packed = false;
4286 tree type = (TREE_TYPE (DR_REF (dr)));
4288 if (optab_handler (vec_realign_load_optab, mode) != CODE_FOR_nothing
4289 && (!targetm.vectorize.builtin_mask_for_load
4290 || targetm.vectorize.builtin_mask_for_load ()))
4292 tree vectype = STMT_VINFO_VECTYPE (stmt_info);
4293 if ((nested_in_vect_loop
4294 && (TREE_INT_CST_LOW (DR_STEP (dr))
4295 != GET_MODE_SIZE (TYPE_MODE (vectype))))
4297 return dr_explicit_realign;
4299 return dr_explicit_realign_optimized;
4301 if (!known_alignment_for_access_p (dr))
4303 tree ba = DR_BASE_OBJECT (dr);
4306 is_packed = contains_packed_reference (ba);
4309 if (targetm.vectorize.
4310 support_vector_misalignment (mode, type,
4311 DR_MISALIGNMENT (dr), is_packed))
4312 /* Can't software pipeline the loads, but can at least do them. */
4313 return dr_unaligned_supported;
4317 bool is_packed = false;
4318 tree type = (TREE_TYPE (DR_REF (dr)));
4320 if (!known_alignment_for_access_p (dr))
4322 tree ba = DR_BASE_OBJECT (dr);
4325 is_packed = contains_packed_reference (ba);
4328 if (targetm.vectorize.
4329 support_vector_misalignment (mode, type,
4330 DR_MISALIGNMENT (dr), is_packed))
4331 return dr_unaligned_supported;
4335 return dr_unaligned_unsupported;