| blob | bf8c99779aee4b92f57ec63ec610eed85d87468e |
1 /* Statement Analysis and Transformation for Vectorization
2 Copyright (C) 2003-2023 Free Software Foundation, Inc.
3 Contributed by Dorit Naishlos <dorit@il.ibm.com>
4 and Ira Rosen <irar@il.ibm.com>
6 This file is part of GCC.
8 GCC is free software; you can redistribute it and/or modify it under
9 the terms of the GNU General Public License as published by the Free
10 Software Foundation; either version 3, or (at your option) any later
11 version.
13 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
14 WARRANTY; without even the implied warranty of MERCHANTABILITY or
15 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
16 for more details.
18 You should have received a copy of the GNU General Public License
19 along with GCC; see the file COPYING3. If not see
20 <http://www.gnu.org/licenses/>. */
61 /* For lang_hooks.types.type_for_mode. */
64 /* Return the vectorized type for the given statement. */
66 tree
68 {
70 }
72 /* Return TRUE iff the given statement is in an inner loop relative to
73 the loop being vectorized. */
74 bool
76 {
88 }
90 /* Record the cost of a statement, either by directly informing the
91 target model or by saving it in a vector for later processing.
92 Return a preliminary estimate of the statement's cost. */
94 static unsigned
100 {
108 stmt_info_for_cost si
114 }
116 unsigned
121 {
124 }
126 unsigned
131 {
134 }
136 unsigned
140 {
145 }
147 /* Return a variable of type ELEM_TYPE[NELEMS]. */
149 static tree
151 {
154 }
156 /* ARRAY is an array of vectors created by create_vector_array.
157 Return an SSA_NAME for the vector in index N. The reference
158 is part of the vectorization of STMT_INFO and the vector is associated
159 with scalar destination SCALAR_DEST. */
161 static tree
165 {
182 }
184 /* ARRAY is an array of vectors created by create_vector_array.
185 Emit code to store SSA_NAME VECT in index N of the array.
186 The store is part of the vectorization of STMT_INFO. */
188 static void
192 {
193 tree array_ref;
202 }
204 /* PTR is a pointer to an array of type TYPE. Return a representation
205 of *PTR. The memory reference replaces those in FIRST_DR
206 (and its group). */
208 static tree
210 {
211 tree mem_ref;
214 /* Arrays have the same alignment as their type. */
217 }
219 /* Add a clobber of variable VAR to the vectorization of STMT_INFO.
220 Emit the clobber before *GSI. */
222 static void
225 {
229 }
231 /* Utility functions used by vect_mark_stmts_to_be_vectorized. */
233 /* Function vect_mark_relevant.
235 Mark STMT_INFO as "relevant for vectorization" and add it to WORKLIST. */
237 static void
240 {
249 /* If this stmt is an original stmt in a pattern, we might need to mark its
250 related pattern stmt instead of the original stmt. However, such stmts
251 may have their own uses that are not in any pattern, in such cases the
252 stmt itself should be marked. */
254 {
255 /* This is the last stmt in a sequence that was detected as a
256 pattern that can potentially be vectorized. Don't mark the stmt
257 as relevant/live because it's not going to be vectorized.
258 Instead mark the pattern-stmt that replaces it. */
262 "last stmt in pattern. don't mark"
272 {
275 "vec_stmt_relevant_p: forcing live pattern stmt "
278 }
284 }
292 {
297 }
300 }
303 /* Function is_simple_and_all_uses_invariant
305 Return true if STMT_INFO is simple and all uses of it are invariant. */
307 bool
309 loop_vec_info loop_vinfo)
310 {
311 tree op;
312 ssa_op_iter iter;
319 {
323 {
328 }
332 }
334 }
336 /* Function vect_stmt_relevant_p.
338 Return true if STMT_INFO, in the loop that is represented by LOOP_VINFO,
339 is "relevant for vectorization".
341 A stmt is considered "relevant for vectorization" if:
342 - it has uses outside the loop.
343 - it has vdefs (it alters memory).
344 - control stmts in the loop (except for the exit condition).
346 CHECKME: what other side effects would the vectorizer allow? */
348 static bool
351 {
353 ssa_op_iter op_iter;
354 imm_use_iterator imm_iter;
355 use_operand_p use_p;
356 def_operand_p def_p;
361 /* cond stmt other than loop exit cond. */
366 /* changing memory. */
370 {
375 }
377 /* uses outside the loop. */
379 {
381 {
384 {
392 /* We expect all such uses to be in the loop exit phis
393 (because of loop closed form) */
398 }
399 }
400 }
404 {
409 }
412 }
415 /* Function exist_non_indexing_operands_for_use_p
417 USE is one of the uses attached to STMT_INFO. Check if USE is
418 used in STMT_INFO for anything other than indexing an array. */
420 static bool
422 {
423 tree operand;
425 /* USE corresponds to some operand in STMT. If there is no data
426 reference in STMT, then any operand that corresponds to USE
427 is not indexing an array. */
431 /* STMT has a data_ref. FORNOW this means that its of one of
432 the following forms:
433 -1- ARRAY_REF = var
434 -2- var = ARRAY_REF
435 (This should have been verified in analyze_data_refs).
437 'var' in the second case corresponds to a def, not a use,
438 so USE cannot correspond to any operands that are not used
439 for array indexing.
441 Therefore, all we need to check is if STMT falls into the
442 first case, and whether var corresponds to USE. */
446 {
449 {
462 }
464 }
476 }
479 /*
480 Function process_use.
482 Inputs:
483 - a USE in STMT_VINFO in a loop represented by LOOP_VINFO
484 - RELEVANT - enum value to be set in the STMT_VINFO of the stmt
485 that defined USE. This is done by calling mark_relevant and passing it
486 the WORKLIST (to add DEF_STMT to the WORKLIST in case it is relevant).
487 - FORCE is true if exist_non_indexing_operands_for_use_p check shouldn't
488 be performed.
490 Outputs:
491 Generally, LIVE_P and RELEVANT are used to define the liveness and
492 relevance info of the DEF_STMT of this USE:
493 STMT_VINFO_LIVE_P (DEF_stmt_vinfo) <-- live_p
494 STMT_VINFO_RELEVANT (DEF_stmt_vinfo) <-- relevant
495 Exceptions:
496 - case 1: If USE is used only for address computations (e.g. array indexing),
497 which does not need to be directly vectorized, then the liveness/relevance
498 of the respective DEF_STMT is left unchanged.
499 - case 2: If STMT_VINFO is a reduction phi and DEF_STMT is a reduction stmt,
500 we skip DEF_STMT cause it had already been processed.
501 - case 3: If DEF_STMT and STMT_VINFO are in different nests, then
502 "relevant" will be modified accordingly.
504 Return true if everything is as expected. Return false otherwise. */
506 static opt_result
510 {
511 stmt_vec_info dstmt_vinfo;
514 /* case 1: we are only interested in uses that need to be vectorized. Uses
515 that are used for address computation are not considered relevant. */
521 "not vectorized:"
530 /* case 2: A reduction phi (STMT) defined by a reduction stmt (DSTMT_VINFO).
531 We have to force the stmt live since the epilogue loop needs it to
532 continue computing the reduction. */
538 {
544 }
546 /* case 3a: outer-loop stmt defining an inner-loop stmt:
547 outer-loop-header-bb:
548 d = dstmt_vinfo
549 inner-loop:
550 stmt # use (d)
551 outer-loop-tail-bb:
552 ... */
554 {
560 {
581 }
582 }
584 /* case 3b: inner-loop stmt defining an outer-loop stmt:
585 outer-loop-header-bb:
586 ...
587 inner-loop:
588 d = dstmt_vinfo
589 outer-loop-tail-bb (or outer-loop-exit-bb in double reduction):
590 stmt # use (d) */
592 {
598 {
616 }
617 }
618 /* We are also not interested in uses on loop PHI backedges that are
619 inductions. Otherwise we'll needlessly vectorize the IV increment
620 and cause hybrid SLP for SLP inductions. Unless the PHI is live
621 of course. */
628 {
633 }
638 }
641 /* Function vect_mark_stmts_to_be_vectorized.
643 Not all stmts in the loop need to be vectorized. For example:
645 for i...
646 for j...
647 1. T0 = i + j
648 2. T1 = a[T0]
650 3. j = j + 1
652 Stmt 1 and 3 do not need to be vectorized, because loop control and
653 addressing of vectorized data-refs are handled differently.
655 This pass detects such stmts. */
657 opt_result
659 {
663 gimple_stmt_iterator si;
665 basic_block bb;
673 /* 1. Init worklist. */
675 {
678 {
686 }
688 {
698 }
699 }
701 /* 2. Process_worklist */
703 {
704 use_operand_p use_p;
705 ssa_op_iter iter;
712 /* Examine the USEs of STMT. For each USE, mark the stmt that defines it
713 (DEF_STMT) as relevant/irrelevant according to the relevance property
714 of STMT. */
717 /* Generally, the relevance property of STMT (in STMT_VINFO_RELEVANT) is
718 propagated as is to the DEF_STMTs of its USEs.
720 One exception is when STMT has been identified as defining a reduction
721 variable; in this case we set the relevance to vect_used_by_reduction.
722 This is because we distinguish between two kinds of relevant stmts -
723 those that are used by a reduction computation, and those that are
724 (also) used by a regular computation. This allows us later on to
725 identify stmts that are used solely by a reduction, and therefore the
726 order of the results that they produce does not have to be kept. */
729 {
758 }
761 {
762 /* Pattern statements are not inserted into the code, so
763 FOR_EACH_PHI_OR_STMT_USE optimizes their operands out, and we
764 have to scan the RHS or function arguments instead. */
766 {
772 {
773 opt_result res
783 }
785 {
788 {
789 opt_result res
794 }
795 }
796 }
798 {
800 {
802 opt_result res
807 }
808 }
809 }
810 else
812 {
814 opt_result res
819 }
822 {
823 gather_scatter_info gs_info;
826 opt_result res
830 {
834 }
835 }
839 }
841 /* Function vect_model_simple_cost.
843 Models cost for simple operations, i.e. those that only emit ncopies of a
844 single op. Right now, this does not account for multiple insns that could
845 be generated for the single vector op. We will handle that shortly. */
847 static void
852 slp_tree node,
855 {
860 /* ??? Somehow we need to fix this at the callers. */
865 /* Cost the "broadcast" of a scalar operand in to a vector operand.
866 Use scalar_to_vec to cost the broadcast, as elsewhere in the vector
867 cost model. */
873 /* Pass the inside-of-loop statements to the target-specific cost model. */
879 "vect_model_simple_cost: inside_cost = %d, "
881 }
884 /* Model cost for type demotion and promotion operations. PWR is
885 normally zero for single-step promotions and demotions. It will be
886 one if two-step promotion/demotion is required, and so on. NCOPIES
887 is the number of vector results (and thus number of instructions)
888 for the narrowest end of the operation chain. Each additional
889 step doubles the number of instructions required. If WIDEN_ARITH
890 is true the stmt is doing widening arithmetic. */
892 static void
898 {
903 {
905 widen_arith
909 }
911 /* FORNOW: Assuming maximum 2 args per stmts. */
919 "vect_model_promotion_demotion_cost: inside_cost = %d, "
921 }
923 /* Returns true if the current function returns DECL. */
925 static bool
927 {
928 edge_iterator ei;
929 edge e;
931 {
937 /* We often end up with an aggregate copy to the result decl,
938 handle that case as well. First skip intermediate clobbers
939 though. */
941 do
942 {
944 }
950 }
952 }
954 /* Calculate cost of DR's memory access. */
955 void
957 dr_alignment_support alignment_support_scheme,
961 {
963 {
965 {
968 vect_body);
974 }
977 {
978 /* Here, we assign an additional cost for the unaligned store. */
984 "vect_model_store_cost: unaligned supported by "
987 }
990 {
997 }
1001 }
1002 }
1004 /* Calculate cost of DR's memory access. */
1005 void
1007 dr_alignment_support alignment_support_scheme,
1014 {
1016 {
1018 {
1027 }
1029 {
1030 /* Here, we assign an additional cost for the unaligned load. */
1037 "vect_model_load_cost: unaligned supported by "
1041 }
1043 {
1049 /* FIXME: If the misalignment remains fixed across the iterations of
1050 the containing loop, the following cost should be added to the
1051 prologue costs. */
1061 }
1063 {
1066 "vect_model_load_cost: unaligned software "
1069 /* Unaligned software pipeline has a load of an address, an initial
1070 load, and possibly a mask operation to "prime" the loop. However,
1071 if this is an access in a group of loads, which provide grouped
1072 access, then the above cost should only be considered for one
1073 access in the group. Inside the loop, there is a load op
1074 and a realignment op. */
1077 {
1085 }
1094 "vect_model_load_cost: explicit realign optimized"
1098 }
1101 {
1108 }
1112 }
1113 }
1115 /* Insert the new stmt NEW_STMT at *GSI or at the appropriate place in
1116 the loop preheader for the vectorized stmt STMT_VINFO. */
1118 static void
1121 {
1124 else
1130 }
1132 /* Function vect_init_vector.
1134 Insert a new stmt (INIT_STMT) that initializes a new variable of type
1135 TYPE with the value VAL. If TYPE is a vector type and VAL does not have
1136 vector type a vector with all elements equal to VAL is created first.
1137 Place the initialization at GSI if it is not NULL. Otherwise, place the
1138 initialization at the loop preheader.
1139 Return the DEF of INIT_STMT.
1140 It will be used in the vectorization of STMT_INFO. */
1142 tree
1145 {
1147 tree new_temp;
1149 /* We abuse this function to push sth to a SSA name with initial 'val'. */
1151 {
1154 {
1155 /* Scalar boolean value should be transformed into
1156 all zeros or all ones value before building a vector. */
1158 {
1164 else
1165 {
1171 }
1172 }
1173 else
1174 {
1179 else
1180 /* ??? Condition vectorization expects us to do
1181 promotion of invariant/external defs. */
1185 {
1189 }
1190 }
1191 }
1193 }
1199 }
1202 /* Function vect_get_vec_defs_for_operand.
1204 OP is an operand in STMT_VINFO. This function returns a vector of
1205 NCOPIES defs that will be used in the vectorized stmts for STMT_VINFO.
1207 In the case that OP is an SSA_NAME which is defined in the loop, then
1208 STMT_VINFO_VEC_STMTS of the defining stmt holds the relevant defs.
1210 In case OP is an invariant or constant, a new stmt that creates a vector def
1211 needs to be introduced. VECTYPE may be used to specify a required type for
1212 vector invariant. */
1214 void
1218 {
1228 stmt_vec_info def_stmt_info;
1237 {
1239 tree vector_type;
1246 else
1253 }
1254 else
1255 {
1261 }
1262 }
1265 /* Get vectorized definitions for OP0 and OP1. */
1267 void
1274 {
1276 {
1285 }
1286 else
1287 {
1300 }
1301 }
1303 void
1310 {
1316 }
1318 /* Helper function called by vect_finish_replace_stmt and
1319 vect_finish_stmt_generation. Set the location of the new
1320 statement and create and return a stmt_vec_info for it. */
1322 static void
1325 {
1330 {
1333 /* While EH edges will generally prevent vectorization, stmt might
1334 e.g. be in a must-not-throw region. Ensure newly created stmts
1335 that could throw are part of the same region. */
1339 }
1340 else
1342 }
1344 /* Replace the scalar statement STMT_INFO with a new vector statement VEC_STMT,
1345 which sets the same scalar result as STMT_INFO did. Create and return a
1346 stmt_vec_info for VEC_STMT. */
1348 void
1351 {
1359 }
1361 /* Add VEC_STMT to the vectorized implementation of STMT_INFO and insert it
1362 before *GSI. Create and return a stmt_vec_info for VEC_STMT. */
1364 void
1368 {
1373 {
1377 {
1381 /* If we have an SSA vuse and insert a store, update virtual
1382 SSA form to avoid triggering the renamer. Do so only
1383 if we can easily see all uses - which is what almost always
1384 happens with the way vectorized stmts are inserted. */
1393 {
1397 }
1398 }
1399 }
1402 }
1404 /* We want to vectorize a call to combined function CFN with function
1405 decl FNDECL, using VECTYPE_OUT as the type of the output and VECTYPE_IN
1406 as the types of all inputs. Check whether this is possible using
1407 an internal function, returning its code if so or IFN_LAST if not. */
1409 static internal_fn
1412 {
1413 internal_fn ifn;
1416 else
1419 {
1422 {
1427 /* The type size of both the vectype_in and vectype_out should be
1428 exactly the same when vectype_out isn't participating the optab.
1429 While there is no restriction for type size when vectype_out
1430 is part of the optab query. */
1435 OPTIMIZE_FOR_SPEED))
1437 }
1438 }
1440 }
1444 gimple_stmt_iterator *);
1446 /* Check whether a load or store statement in the loop described by
1447 LOOP_VINFO is possible in a loop using partial vectors. This is
1448 testing whether the vectorizer pass has the appropriate support,
1449 as well as whether the target does.
1451 VLS_TYPE says whether the statement is a load or store and VECTYPE
1452 is the type of the vector being loaded or stored. SLP_NODE is the SLP
1453 node that contains the statement, or null if none. MEMORY_ACCESS_TYPE
1454 says how the load or store is going to be implemented and GROUP_SIZE
1455 is the number of load or store statements in the containing group.
1456 If the access is a gather load or scatter store, GS_INFO describes
1457 its arguments. If the load or store is conditional, SCALAR_MASK is the
1458 condition under which it occurs.
1460 Clear LOOP_VINFO_CAN_USE_PARTIAL_VECTORS_P if a loop using partial
1461 vectors is not supported, otherwise record the required rgroup control
1462 types. */
1464 static void
1466 slp_tree slp_node,
1467 vec_load_store_type vls_type,
1469 vect_memory_access_type
1470 memory_access_type,
1472 tree scalar_mask)
1473 {
1474 /* Invariant loads need no special support. */
1481 else
1489 {
1490 internal_fn ifn
1497 scalar_mask);
1498 else
1499 {
1502 "can't operate on partial vectors because"
1503 " the target doesn't have an appropriate"
1506 }
1508 }
1511 {
1512 internal_fn ifn = (is_load
1513 ? IFN_MASK_GATHER_LOAD
1515 internal_fn len_ifn = (is_load
1516 ? IFN_MASK_LEN_GATHER_LOAD
1528 scalar_mask);
1529 else
1530 {
1533 "can't operate on partial vectors because"
1534 " the target doesn't have an appropriate"
1537 }
1539 }
1543 {
1544 /* Element X of the data must come from iteration i * VF + X of the
1545 scalar loop. We need more work to support other mappings. */
1548 "can't operate on partial vectors because an"
1552 }
1555 {
1558 "can't operate on partial vectors when emulating"
1562 }
1564 /* We might load more scalars than we need for permuting SLP loads.
1565 We checked in get_group_load_store_type that the extra elements
1566 don't leak into a new vector. */
1568 {
1573 };
1577 machine_mode mask_mode;
1578 machine_mode vmode;
1581 {
1586 }
1589 {
1593 }
1596 {
1599 "can't operate on partial vectors because the"
1600 " target doesn't have the appropriate partial"
1603 }
1604 }
1606 /* Return the mask input to a masked load or store. VEC_MASK is the vectorized
1607 form of the scalar mask condition and LOOP_MASK, if nonnull, is the mask
1608 that needs to be applied to all loads and stores in a vectorized loop.
1609 Return VEC_MASK if LOOP_MASK is null or if VEC_MASK is already masked,
1610 otherwise return VEC_MASK & LOOP_MASK.
1612 MASK_TYPE is the type of both masks. If new statements are needed,
1613 insert them before GSI. */
1615 static tree
1618 {
1634 }
1636 /* Determine whether we can use a gather load or scatter store to vectorize
1637 strided load or store STMT_INFO by truncating the current offset to a
1638 smaller width. We need to be able to construct an offset vector:
1640 { 0, X, X*2, X*3, ... }
1642 without loss of precision, where X is STMT_INFO's DR_STEP.
1644 Return true if this is possible, describing the gather load or scatter
1645 store in GS_INFO. MASKED_P is true if the load or store is conditional. */
1647 static bool
1651 {
1656 {
1657 /* ??? Perhaps we could use range information here? */
1662 }
1664 /* Get the number of bits in an element. */
1669 /* Set COUNT to the upper limit on the number of elements - 1.
1670 Start with the maximum vectorization factor. */
1673 /* Try lowering COUNT to the number of scalar latch iterations. */
1675 widest_int max_iters;
1680 /* Try scales of 1 and the element size. */
1684 {
1686 widest_int factor;
1690 /* Determine the minimum precision of (COUNT - 1) * STEP / SCALE. */
1697 /* Find the narrowest viable offset type. */
1702 /* See whether the target supports the operation with an offset
1703 no narrower than OFFSET_TYPE. */
1712 /* Logically the sum of DR_BASE_ADDRESS, DR_INIT and DR_OFFSET,
1713 but we don't need to store that here. */
1721 }
1725 "truncating gather/scatter offset to %d bits"
1729 }
1731 /* Return true if we can use gather/scatter internal functions to
1732 vectorize STMT_INFO, which is a grouped or strided load or store.
1733 MASKED_P is true if load or store is conditional. When returning
1734 true, fill in GS_INFO with the information required to perform the
1735 operation. */
1737 static bool
1741 {
1756 "using gather/scatter for strided/grouped access,"
1760 }
1762 /* STMT_INFO is a non-strided load or store, meaning that it accesses
1763 elements with a known constant step. Return -1 if that step
1764 is negative, 0 if it is zero, and 1 if it is greater than zero. */
1766 static int
1768 {
1771 size_zero_node);
1772 }
1774 /* If the target supports a permute mask that reverses the elements in
1775 a vector of type VECTYPE, return that mask, otherwise return null. */
1777 static tree
1779 {
1782 /* The encoding has a single stepped pattern. */
1789 indices))
1792 }
1794 /* A subroutine of get_load_store_type, with a subset of the same
1795 arguments. Handle the case where STMT_INFO is a load or store that
1796 accesses consecutive elements with a negative step. Sets *POFFSET
1797 to the offset to be applied to the DR for the first access. */
1799 static vect_memory_access_type
1802 vec_load_store_type vls_type,
1804 {
1806 dr_alignment_support alignment_support_scheme;
1809 {
1814 }
1816 /* For backward running DRs the first access in vectype actually is
1817 N-1 elements before the address of the DR. */
1822 alignment_support_scheme
1826 {
1832 }
1835 {
1838 "negative step with invariant source;"
1841 }
1844 {
1850 }
1853 }
1855 /* STMT_INFO is either a masked or unconditional store. Return the value
1856 being stored. */
1858 tree
1860 {
1862 {
1865 }
1867 {
1872 }
1874 }
1876 /* Function VECTOR_VECTOR_COMPOSITION_TYPE
1878 This function returns a vector type which can be composed with NETLS pieces,
1879 whose type is recorded in PTYPE. VTYPE should be a vector type, and has the
1880 same vector size as the return vector. It checks target whether supports
1881 pieces-size vector mode for construction firstly, if target fails to, check
1882 pieces-size scalar mode for construction further. It returns NULL_TREE if
1883 fails to find the available composition.
1885 For example, for (vtype=V16QI, nelts=4), we can probably get:
1886 - V16QI with PTYPE V4QI.
1887 - V4SI with PTYPE SI.
1888 - NULL_TREE. */
1890 static tree
1892 {
1900 /* When we are asked to compose the vector from its components let
1901 that happen directly. */
1903 {
1906 }
1911 {
1912 /* First check if vec_init optab supports construction from
1913 vector pieces directly. */
1916 machine_mode rmode;
1920 {
1923 }
1925 /* Otherwise check if exists an integer type of the same piece size and
1926 if vec_init optab supports construction from it directly. */
1931 {
1934 }
1935 }
1938 }
1940 /* A subroutine of get_load_store_type, with a subset of the same
1941 arguments. Handle the case where STMT_INFO is part of a grouped load
1942 or store.
1944 For stores, the statements in the group are all consecutive
1945 and there is no gap at the end. For loads, the statements in the
1946 group might not be consecutive; there can be gaps between statements
1947 as well as at the end. */
1949 static bool
1959 {
1962 stmt_vec_info first_stmt_info;
1966 {
1970 }
1971 else
1972 {
1976 }
1982 /* True if the vectorized statements would access beyond the last
1983 statement in the group. */
1986 /* True if we can cope with such overrun by peeling for gaps, so that
1987 there is at least one final scalar iteration after the vector loop. */
1990 && loop_vinfo
1993 /* There can only be a gap at the end of the group if the stride is
1994 known at compile time. */
1997 /* Stores can't yet have gaps. */
2001 {
2002 /* For SLP vectorization we directly vectorize a subchain
2003 without permutation. */
2005 first_dr_info
2008 {
2009 /* Try to use consecutive accesses of DR_GROUP_SIZE elements,
2010 separated by the stride, until we have a complete vector.
2011 Fall back to scalar accesses if that isn't possible. */
2014 else
2016 }
2017 else
2018 {
2021 {
2023 "Grouped store with gaps requires"
2026 }
2027 /* An overrun is fine if the trailing elements are smaller
2028 than the alignment boundary B. Every vector access will
2029 be a multiple of B and so we are guaranteed to access a
2030 non-gap element in the same B-sized block. */
2033 vectype)
2037 /* If the gap splits the vector in half and the target
2038 can do half-vector operations avoid the epilogue peeling
2039 by simply loading half of the vector only. Usually
2040 the construction with an upper zero half will be elided. */
2041 dr_alignment_support alss;
2043 tree half_vtype;
2045 && !masked_p
2048 == dr_aligned
2057 {
2062 }
2065 {
2067 /* ??? The VMAT_CONTIGUOUS_REVERSE code generation is
2068 only correct for single element "interleaving" SLP. */
2071 else
2072 {
2073 /* Try to use consecutive accesses of DR_GROUP_SIZE elements,
2074 separated by the stride, until we have a complete vector.
2075 Fall back to scalar accesses if that isn't possible. */
2078 else
2080 }
2081 }
2083 {
2086 /* Invariant accesses perform only component accesses, alignment
2087 is irrelevant for them. */
2089 }
2090 else
2093 /* When we have a contiguous access across loop iterations
2094 but the access in the loop doesn't cover the full vector
2095 we can end up with no gap recorded but still excess
2096 elements accessed, see PR103116. Make sure we peel for
2097 gaps if necessary and sufficient and give up if not.
2099 If there is a combination of the access not covering the full
2100 vector and a gap recorded then we may need to peel twice. */
2105 nunits))
2106 {
2111 /* Peeling for gaps assumes that a single scalar iteration
2112 is enough to make sure the last vector iteration doesn't
2113 access excess elements.
2114 ??? Enhancements include peeling multiple iterations
2115 or using masked loads with a static mask. */
2117 {
2120 "peeling for gaps insufficient for "
2123 }
2125 }
2126 }
2127 }
2128 else
2129 {
2130 /* We can always handle this case using elementwise accesses,
2131 but see if something more efficient is available. */
2134 /* If there is a gap at the end of the group then these optimizations
2135 would access excess elements in the last iteration. */
2137 /* An overrun is fine if the trailing elements are smaller than the
2138 alignment boundary B. Every vector access will be a multiple of B
2139 and so we are guaranteed to access a non-gap element in the
2140 same B-sized block. */
2142 && !masked_p
2150 {
2151 /* First cope with the degenerate case of a single-element
2152 vector. */
2154 ;
2156 else
2157 {
2158 /* Otherwise try using LOAD/STORE_LANES. */
2159 *lanes_ifn
2163 masked_p);
2165 {
2168 }
2170 /* If that fails, try using permuting loads. */
2173 single_element_p,
2174 group_size)
2176 {
2179 }
2180 }
2181 }
2183 /* As a last resort, trying using a gather load or scatter store.
2185 ??? Although the code can handle all group sizes correctly,
2186 it probably isn't a win to use separate strided accesses based
2187 on nearby locations. Or, even if it's a win over scalar code,
2188 it might not be a win over vectorizing at a lower VF, if that
2189 allows us to use contiguous accesses. */
2191 && single_element_p
2192 && loop_vinfo
2196 }
2200 {
2203 }
2204 else
2205 {
2207 *alignment_support_scheme
2210 }
2213 {
2214 /* STMT is the leader of the group. Check the operands of all the
2215 stmts of the group. */
2218 {
2222 {
2227 }
2229 }
2230 }
2233 {
2237 "Data access with gaps requires scalar "
2240 }
2243 }
2245 /* Analyze load or store statement STMT_INFO of type VLS_TYPE. Return true
2246 if there is a memory access type that the vectorized form can use,
2247 storing it in *MEMORY_ACCESS_TYPE if so. If we decide to use gathers
2248 or scatters, fill in GS_INFO accordingly. In addition
2249 *ALIGNMENT_SUPPORT_SCHEME is filled out and false is returned if
2250 the target does not support the alignment scheme. *MISALIGNMENT
2251 is set according to the alignment of the access (including
2252 DR_MISALIGNMENT_UNKNOWN when it is unknown).
2254 SLP says whether we're performing SLP rather than loop vectorization.
2255 MASKED_P is true if the statement is conditional on a vectorized mask.
2256 VECTYPE is the vector type that the vectorized statements will use.
2257 NCOPIES is the number of vector statements that will be needed. */
2259 static bool
2270 {
2276 {
2280 /* When using internal functions, we rely on pattern recognition
2281 to convert the type of the offset to the type that the target
2282 requires, with the result being a call to an internal function.
2283 If that failed for some reason (e.g. because another pattern
2284 took priority), just handle cases in which the offset already
2285 has the right type. */
2290 {
2296 }
2300 {
2306 }
2308 {
2314 {
2317 "unsupported vector types for emulated "
2320 }
2321 }
2322 /* Gather-scatter accesses perform only component accesses, alignment
2323 is irrelevant for them. */
2325 }
2327 {
2329 masked_p,
2331 alignment_support_scheme,
2334 }
2336 {
2342 else
2344 /* Alignment is irrelevant here. */
2346 }
2347 else
2348 {
2351 {
2354 /* Invariant accesses perform only component accesses, alignment
2355 is irrelevant for them. */
2357 }
2358 else
2359 {
2363 else
2367 *alignment_support_scheme
2371 }
2372 }
2377 {
2380 "Not using elementwise accesses due to variable "
2383 }
2386 {
2391 }
2393 /* FIXME: At the moment the cost model seems to underestimate the
2394 cost of using elementwise accesses. This check preserves the
2395 traditional behavior until that can be fixed. */
2404 {
2409 }
2411 }
2413 /* Return true if boolean argument at MASK_INDEX is suitable for vectorizing
2414 conditional operation STMT_INFO. When returning true, store the mask
2415 in *MASK, the type of its definition in *MASK_DT_OUT, the type of the
2416 vectorized mask in *MASK_VECTYPE_OUT and the SLP node corresponding
2417 to the mask in *MASK_NODE if MASK_NODE is not NULL. */
2419 static bool
2424 {
2426 tree mask_vectype;
2427 slp_tree mask_node_1;
2430 {
2435 }
2438 {
2443 }
2445 /* If the caller is not prepared for adjusting an external/constant
2446 SLP mask vector type fail. */
2448 && !mask_node
2450 {
2455 }
2460 mask_node_1);
2463 {
2468 }
2472 {
2475 "vector mask type %T"
2480 }
2487 }
2489 /* Return true if stored value is suitable for vectorizing store
2490 statement STMT_INFO. When returning true, store the scalar stored
2491 in *RHS and *RHS_NODE, the type of the definition in *RHS_DT_OUT,
2492 the type of the vectorized store value in
2493 *RHS_VECTYPE_OUT and the type of the store in *VLS_TYPE_OUT. */
2495 static bool
2500 {
2503 {
2507 }
2509 op_no = vect_slp_child_index_for_operand
2513 tree rhs_vectype;
2516 {
2521 }
2523 /* In the case this is a store from a constant make sure
2524 native_encode_expr can handle it. */
2526 {
2531 }
2535 {
2540 }
2546 else
2549 }
2551 /* Build an all-ones vector mask of type MASKTYPE while vectorizing STMT_INFO.
2552 Note that we support masks with floating-point type, in which case the
2553 floats are interpreted as a bitmask. */
2555 static tree
2558 {
2563 {
2567 }
2569 {
2570 REAL_VALUE_TYPE r;
2578 }
2580 }
2582 /* Build an all-zero merge value of type VECTYPE while vectorizing
2583 STMT_INFO as a gather load. */
2585 static tree
2588 {
2589 tree merge;
2593 {
2594 REAL_VALUE_TYPE r;
2600 }
2601 else
2605 }
2607 /* Build a gather load call while vectorizing STMT_INFO. Insert new
2608 instructions before GSI and add them to VEC_STMT. GS_INFO describes
2609 the gather load operation. If the load is conditional, MASK is the
2610 vectorized condition, otherwise MASK is null. PTR is the base
2611 pointer and OFFSET is the vectorized offset. */
2618 {
2627 tree var;
2629 && (!mask
2635 {
2643 }
2648 {
2650 {
2655 else
2659 gassign *new_stmt
2664 {
2671 }
2674 }
2675 else
2676 {
2679 }
2680 }
2681 else
2682 {
2685 }
2692 {
2700 }
2703 }
2705 /* Build a scatter store call while vectorizing STMT_INFO. Insert new
2706 instructions before GSI. GS_INFO describes the scatter store operation.
2707 PTR is the base pointer, OFFSET the vectorized offsets and OPRND the
2708 vectorized data to store.
2709 If the store is conditional, MASK is the vectorized condition, otherwise
2710 MASK is null. */
2717 {
2730 {
2733 tree utype;
2736 else
2740 gassign *new_stmt
2745 {
2751 }
2752 }
2753 else
2754 {
2757 }
2761 {
2769 }
2773 {
2781 }
2784 gcall *new_stmt
2787 }
2789 /* Prepare the base and offset in GS_INFO for vectorization.
2790 Set *DATAREF_PTR to the loop-invariant base address and *VEC_OFFSET
2791 to the vectorized offset argument for the first copy of STMT_INFO.
2792 STMT_INFO is the statement described by GS_INFO and LOOP is the
2793 containing loop. */
2795 static void
2800 {
2804 {
2805 basic_block new_bb;
2809 }
2812 else
2813 {
2814 unsigned ncopies
2819 }
2820 }
2822 /* Prepare to implement a grouped or strided load or store using
2823 the gather load or scatter store operation described by GS_INFO.
2824 STMT_INFO is the load or store statement.
2826 Set *DATAREF_BUMP to the amount that should be added to the base
2827 address after each copy of the vectorized statement. Set *VEC_OFFSET
2828 to an invariant offset vector in which element I has the value
2829 I * DR_STEP / SCALE. */
2831 static void
2833 loop_vec_info loop_vinfo,
2838 {
2843 {
2844 /* _31 = .SELECT_VL (ivtmp_29, POLY_INT_CST [4, 4]);
2845 ivtmp_8 = _31 * 16 (step in bytes);
2846 .MASK_LEN_SCATTER_STORE (vectp_a.9_7, ... );
2847 vectp_a.9_26 = vectp_a.9_7 + ivtmp_8; */
2848 tree loop_len
2850 tree tmp
2853 loop_len);
2855 GSI_SAME_STMT);
2856 }
2857 else
2858 {
2859 tree bump
2864 }
2866 /* The offset given in GS_INFO can have pointer type, so use the element
2867 type of the vector instead. */
2870 /* Calculate X = DR_STEP / SCALE and convert it to the appropriate type. */
2875 /* Create {0, X, X*2, X*3, ...}. */
2879 }
2881 /* Prepare the pointer IVs which needs to be updated by a variable amount.
2882 Such variable amount is the outcome of .SELECT_VL. In this case, we can
2883 allow each iteration process the flexible number of elements as long as
2884 the number <= vf elments.
2886 Return data reference according to SELECT_VL.
2887 If new statements are needed, insert them before GSI. */
2889 static tree
2893 vect_memory_access_type memory_access_type)
2894 {
2898 /* gather/scatter never reach here. */
2901 /* When we support SELECT_VL pattern, we dynamic adjust
2902 the memory address by .SELECT_VL result.
2904 The result of .SELECT_VL is the number of elements to
2905 be processed of each iteration. So the memory address
2906 adjustment operation should be:
2908 addr = addr + .SELECT_VL (ARG..) * step;
2909 */
2910 tree loop_len
2913 /* Since the outcome of .SELECT_VL is element size, we should adjust
2914 it into bytesize so that it can be used in address pointer variable
2915 amount IVs adjustment. */
2922 }
2924 /* Return the amount that should be added to a vector pointer to move
2925 to the next or previous copy of AGGR_TYPE. DR_INFO is the data reference
2926 being vectorized and MEMORY_ACCESS_TYPE describes the type of
2927 vectorization. */
2929 static tree
2932 vect_memory_access_type memory_access_type,
2934 {
2942 memory_access_type);
2949 }
2951 /* Check and perform vectorization of BUILT_IN_BSWAP{16,32,64,128}. */
2953 static bool
2959 {
2969 /* Multiple types in SLP are handled by creating the appropriate number of
2970 vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in
2971 case of SLP. */
2974 else
2988 /* The encoding uses one stepped pattern for each byte in the word. */
3000 {
3003 {
3008 }
3015 slp_node
3019 }
3023 /* Transform. */
3027 /* Arguments are ready. create the new vector stmt. */
3029 tree vop;
3031 {
3047 else
3049 }
3056 }
3058 /* Return true if vector types VECTYPE_IN and VECTYPE_OUT have
3059 integer elements and if we can narrow VECTYPE_IN to VECTYPE_OUT
3060 in a single step. On success, store the binary pack code in
3061 *CONVERT_CODE. */
3063 static bool
3066 {
3071 code_helper code;
3081 }
3083 /* Function vectorizable_call.
3085 Check if STMT_INFO performs a function call that can be vectorized.
3086 If VEC_STMT is also passed, vectorize STMT_INFO: create a vectorized
3087 stmt to replace it, put it in VEC_STMT, and insert it at GSI.
3088 Return true if STMT_INFO is vectorizable in this way. */
3090 static bool
3095 {
3097 tree vec_dest;
3098 tree scalar_dest;
3099 tree op;
3102 poly_uint64 nunits_in;
3103 poly_uint64 nunits_out;
3109 vect_unknown_def_type };
3117 tree lhs;
3127 /* Is STMT_INFO a vectorizable call? */
3135 /* Handled by vectorizable_load and vectorizable_store. */
3146 /* Process function arguments. */
3151 /* Bail out if the function has more than four arguments, we do not have
3152 interesting builtin functions to vectorize with more than two arguments
3153 except for fma. No arguments is also not good. */
3157 /* Ignore the arguments of IFN_GOMP_SIMD_LANE, they are magic. */
3160 {
3163 }
3164 /* Similarly pretend IFN_CLZ and IFN_CTZ only has one argument, the second
3165 argument just says whether it is well-defined at zero or not and what
3166 value should be returned for it. */
3168 {
3171 }
3178 {
3180 {
3185 }
3189 {
3194 }
3196 /* We can only handle calls with arguments of the same type. */
3199 {
3204 }
3212 {
3217 }
3218 }
3219 /* If all arguments are external or constant defs, infer the vector type
3220 from the scalar type. */
3226 {
3232 }
3236 {
3241 }
3244 {
3249 }
3251 /* FORNOW */
3260 else
3263 /* We only handle functions that do not read or clobber memory. */
3265 {
3270 }
3272 /* For now, we only vectorize functions if a target specific builtin
3273 is available. TODO -- in some cases, it might be profitable to
3274 insert the calls for pieces of the vector, in order to be able
3275 to vectorize other operations in the loop. */
3280 /* First try using an internal function. */
3288 vectype_in);
3290 /* If that fails, try asking for a target-specific built-in function. */
3292 {
3299 }
3302 {
3304 && !slp_node
3305 && loop_vinfo
3310 {
3311 /* We can handle IFN_GOMP_SIMD_LANE by returning a
3312 { 0, 1, 2, ... vf - 1 } vector. */
3314 }
3322 else
3323 {
3328 }
3329 }
3335 else
3338 /* Sanity check: make sure that at least one copy of the vectorized stmt
3339 needs to be generated. */
3349 {
3355 {
3360 }
3372 {
3376 OPTIMIZE_FOR_SPEED))
3379 OPTIMIZE_FOR_SPEED)))
3380 {
3383 "can't use a fully-masked loop because no"
3386 }
3387 else
3388 {
3389 unsigned int nvectors
3390 = (slp_node
3398 OPTIMIZE_FOR_SPEED))
3401 else
3403 scalar_mask);
3404 }
3405 }
3407 }
3409 /* Transform. */
3414 /* Handle def. */
3422 {
3424 {
3426 /* COND_* -> COND_LEN_* takes 2 extra arguments:LEN,BIAS. */
3428 }
3431 }
3433 {
3436 }
3441 {
3446 {
3447 /* Build argument list for the vectorized call. */
3449 {
3455 /* Arguments are ready. Create the new vector stmt. */
3457 {
3460 {
3462 /* Always true for SLP. */
3467 }
3470 {
3473 }
3481 {
3482 /* We don't define any narrowing conditional functions
3483 at present. */
3486 gcall *call
3492 {
3495 }
3501 }
3502 else
3503 {
3505 {
3507 /* Always true for SLP. */
3509 tree len
3512 signed char biasval
3517 }
3519 {
3521 /* Always true for SLP. */
3529 }
3534 else
3541 }
3543 }
3545 }
3552 {
3555 {
3560 }
3562 }
3569 {
3572 signed char biasval
3577 }
3579 {
3585 }
3589 {
3591 tree new_var
3598 }
3600 {
3601 /* We don't define any narrowing conditional functions at
3602 present. */
3610 {
3613 }
3616 half_res);
3618 }
3619 else
3620 {
3624 else
3631 }
3636 }
3638 {
3641 }
3642 }
3644 {
3646 /* We don't define any narrowing conditional functions at present. */
3649 {
3650 /* Build argument list for the vectorized call. */
3653 else
3657 {
3663 /* Arguments are ready. Create the new vector stmt. */
3665 {
3669 {
3673 }
3677 else
3684 }
3686 }
3689 {
3692 {
3696 }
3702 }
3710 }
3716 {
3719 }
3720 }
3721 else
3722 /* No current target implements this case. */
3727 /* The call in STMT might prevent it from being removed in dce.
3728 We however cannot remove it here, due to the way the ssa name
3729 it defines is mapped to the new definition. So just replace
3730 rhs of the statement with something harmless. */
3738 gassign *new_stmt
3743 }
3746 struct simd_call_arg_info
3747 {
3748 tree vectype;
3749 tree op;
3750 HOST_WIDE_INT linear_step;
3754 };
3756 /* Helper function of vectorizable_simd_clone_call. If OP, an SSA_NAME,
3757 is linear within simd lane (but not within whole loop), note it in
3758 *ARGINFO. */
3760 static void
3763 {
3775 {
3776 tree t;
3780 {
3795 CASE_CONVERT:
3807 }
3813 {
3820 }
3821 }
3822 }
3824 /* Function vectorizable_simd_clone_call.
3826 Check if STMT_INFO performs a function call that can be vectorized
3827 by calling a simd clone of the function.
3828 If VEC_STMT is also passed, vectorize STMT_INFO: create a vectorized
3829 stmt to replace it, put it in VEC_STMT, and insert it at GSI.
3830 Return true if STMT_INFO is vectorizable in this way. */
3832 static bool
3836 stmt_vector_for_cost *)
3837 {
3838 tree vec_dest;
3839 tree scalar_dest;
3842 tree vectype;
3843 poly_uint64 nunits;
3856 /* Is STMT a vectorizable call? */
3864 {
3870 }
3896 /* Process function arguments. */
3899 /* Bail out if the function has zero arguments. */
3910 {
3911 simd_call_arg_info thisarginfo;
3912 affine_iv iv;
3926 {
3931 }
3935 {
3936 /* With SLP we determine the vector type of constants/externals
3937 at analysis time, handling conflicts via
3938 vect_maybe_update_slp_op_vectype. At transform time
3939 we have a vector type recorded for SLP. */
3941 || !slp_node
3946 slp_node);
3947 }
3948 else
3951 /* For linear arguments, the analyze phase should have saved
3952 the base and step in {STMT_VINFO,SLP_TREE}_SIMD_CLONE_INFO. */
3955 {
3959 thisarginfo.simd_lane_linear
3961 /* If loop has been peeled for alignment, we need to adjust it. */
3965 {
3971 thisarginfo.op
3975 }
3976 }
3980 && loop_vinfo
3985 {
3988 }
3993 /* Addresses of array elements indexed by GOMP_SIMD_LANE are
3994 linear too. */
3997 && !vec_stmt
4000 && loop_vinfo
4005 }
4013 else
4016 {
4019 /* The number of arguments in the call and the number of parameters in
4020 the simdclone should match. However, when the simdclone is
4021 'inbranch', it could have one more paramater than nargs when using
4022 an inbranch simdclone to call a non-inbranch call, either in a
4023 non-masked loop using a all true constant mask, or inside a masked
4024 loop using it's mask. */
4027 simd_nargs--;
4042 {
4044 {
4075 /* FORNOW */
4079 /* While we can create a traditional data vector from
4080 an incoming integer mode mask we have no good way to
4081 force generate an integer mode mask from a traditional
4082 boolean vector input. */
4090 }
4094 {
4097 }
4101 }
4107 {
4109 SIMD_CLONE_ARG_TYPE_MASK);
4110 /* Penalize using a masked SIMD clone in a non-masked loop, that is
4111 not in a branch, as we'd have to construct an all-true mask. */
4114 }
4116 {
4119 }
4120 }
4129 num_mask_args++;
4132 {
4136 {
4140 slp_node);
4145 }
4149 {
4154 }
4157 {
4160 {
4163 {
4164 /* FORNOW we only have partial support for vector-type masks
4165 that can't hold all of simdlen. */
4168 vect_location,
4169 "in-branch vector clones are not yet"
4172 }
4173 }
4175 {
4178 num_mask_args),
4180 {
4181 /* FORNOW we only have partial support for integer-type masks
4182 that represent the same number of lanes as the
4183 vectorized mask inputs. */
4186 vect_location,
4187 "in-branch vector clones are not yet "
4190 }
4191 }
4192 else
4193 {
4196 vect_location,
4197 "in-branch vector clones not supported"
4200 }
4201 }
4202 }
4208 else
4211 /* If the function isn't const, only allow it in simd loops where user
4212 has asserted that at least nunits consecutive iterations can be
4213 performed using SIMD instructions. */
4218 /* Sanity check: make sure that at least one copy of the vectorized stmt
4219 needs to be generated. */
4223 {
4227 {
4232 }
4233 /* When the original call is pure or const but the SIMD ABI dictates
4234 an aggregate return we will have to use a virtual definition and
4235 in a loop eventually even need to add a virtual PHI. That's
4236 not straight-forward so allow to fix this up via renaming. */
4241 /* ??? For SLP code-gen we end up inserting after the last
4242 vector argument def rather than at the original call position
4243 so automagic virtual operand updating doesn't work. */
4248 {
4250 {
4255 {
4265 }
4275 }
4276 }
4279 {
4283 "can't use a fully-masked loop because a"
4286 }
4290 /* vect_model_simple_cost (vinfo, stmt_info, ncopies,
4291 dt, slp_node, cost_vec); */
4293 }
4295 /* Transform. */
4300 /* Handle def. */
4306 {
4310 {
4313 }
4314 }
4320 {
4323 }
4324 else
4327 {
4328 poly_uint64 callee_nelements;
4329 poly_uint64 caller_nelements;
4330 /* Build argument list for the vectorized call. */
4333 else
4337 {
4339 tree atype;
4342 {
4349 {
4351 {
4359 {
4366 }
4367 else
4368 {
4372 }
4374 vec_oprnd0
4378 gassign *new_stmt
4380 vec_oprnd0);
4384 }
4385 else
4386 {
4394 else
4397 {
4399 {
4403 op,
4407 }
4408 else
4414 vec_oprnd0);
4415 }
4418 atype))
4419 {
4421 vec_oprnd0);
4422 gassign *new_stmt
4424 vec_oprnd0);
4428 }
4429 else
4431 else
4432 {
4434 gassign *new_stmt
4436 vec_oprnd0);
4440 }
4441 }
4442 }
4446 {
4455 {
4457 {
4458 /* The mask type has fewer elements than simdlen. */
4460 /* FORNOW */
4462 }
4464 {
4465 /* The SIMD clone function has the same number of
4466 elements as the current function. */
4468 {
4472 op,
4475 }
4479 {
4480 vec_loop_masks *loop_masks
4482 tree loop_mask
4486 vec_oprnd0
4490 gsi);
4492 loop_mask });
4494 }
4495 vec_oprnd0
4499 gassign *new_stmt
4501 vec_oprnd0);
4505 }
4506 else
4507 {
4508 /* The mask type has more elements than simdlen. */
4510 /* FORNOW */
4512 }
4513 }
4514 }
4516 {
4518 /* Guess the number of lanes represented by atype. */
4519 poly_uint64 atype_subparts
4521 num_mask_args);
4524 {
4526 {
4530 op,
4533 }
4536 {
4537 /* The mask argument has fewer elements than the
4538 input vector. */
4539 /* FORNOW */
4541 }
4544 {
4545 /* The vector mask argument matches the input
4546 in the number of lanes, but not necessarily
4547 in the mode. */
4552 vec_oprnd0);
4553 gassign *new_stmt
4555 vec_oprnd0);
4559 {
4560 new_stmt
4562 NOP_EXPR,
4563 gimple_assign_lhs
4567 }
4569 }
4570 else
4571 {
4572 /* The mask argument has more elements than the
4573 input vector. */
4574 /* FORNOW */
4576 }
4577 }
4578 }
4579 else
4588 {
4589 gimple_seq stmts;
4594 {
4595 basic_block new_bb;
4599 }
4601 {
4604 }
4609 enum tree_code code
4614 poly_widest_int cst
4619 gassign *new_stmt
4624 UNKNOWN_LOCATION);
4627 }
4628 else
4629 {
4630 enum tree_code code
4635 poly_widest_int cst
4640 gassign *new_stmt
4645 }
4655 }
4656 }
4661 {
4664 tree mask;
4666 SIMD_CLONE_ARG_TYPE_MASK);
4672 {
4674 {
4678 }
4679 else
4684 {
4685 /* This means we are dealing with integer mask modes.
4686 First convert to an integer type with the same size as
4687 the current vector type. */
4688 unsigned HOST_WIDE_INT intermediate_size
4690 tree mid_int_type =
4693 new_stmt
4695 mask);
4697 /* Then zero-extend to the mask mode. */
4700 }
4702 {
4704 integer_one_node);
4706 integer_zero_node);
4710 }
4711 else
4719 }
4720 }
4724 {
4731 else
4734 }
4739 {
4741 {
4749 {
4750 tree t;
4752 {
4756 }
4757 else
4768 else
4770 }
4775 }
4777 {
4786 {
4791 {
4795 tem);
4800 }
4802 }
4803 else
4808 new_stmt
4817 else
4820 }
4822 {
4829 }
4831 {
4833 new_stmt
4836 }
4837 }
4843 else
4845 }
4848 {
4851 }
4854 /* Mark the clone as no longer being a candidate for GC. */
4857 /* The call in STMT might prevent it from being removed in dce.
4858 We however cannot remove it here, due to the way the ssa name
4859 it defines is mapped to the new definition. So just replace
4860 rhs of the statement with something harmless. */
4867 {
4871 }
4872 else
4878 }
4881 /* Function vect_gen_widened_results_half
4883 Create a vector stmt whose code, type, number of arguments, and result
4884 variable are CODE, OP_TYPE, and VEC_DEST, and its arguments are
4885 VEC_OPRND0 and VEC_OPRND1. The new vector stmt is to be inserted at GSI.
4886 In the case that CODE is a CALL_EXPR, this means that a call to DECL
4887 needs to be created (DECL is a function-decl of a target-builtin).
4888 STMT_INFO is the original scalar stmt that we are vectorizing. */
4894 stmt_vec_info stmt_info)
4895 {
4897 tree new_temp;
4899 /* Generate half of the widened result: */
4908 }
4911 /* Create vectorized demotion statements for vector operands from VEC_OPRNDS.
4912 For multi-step conversions store the resulting vectors and call the function
4913 recursively. When NARROW_SRC_P is true, there's still a conversion after
4914 narrowing, don't store the vectors in the SLP_NODE or in vector info of
4915 the scalar statement(or in STMT_VINFO_RELATED_STMT chain). */
4917 static void
4920 stmt_vec_info stmt_info,
4925 {
4932 {
4933 /* Create demotion operation. */
4941 /* Store the resulting vector for next recursive call,
4942 or return the resulting vector_tmp for NARROW FLOAT_EXPR. */
4944 else
4945 {
4946 /* This is the last step of the conversion sequence. Store the
4947 vectors in SLP_NODE or in vector info of the scalar statement
4948 (or in STMT_VINFO_RELATED_STMT chain). */
4951 else
4953 }
4954 }
4956 /* For multi-step demotion operations we first generate demotion operations
4957 from the source type to the intermediate types, and then combine the
4958 results (stored in VEC_OPRNDS) in demotion operation to the destination
4959 type. */
4961 {
4962 /* At each level of recursion we have half of the operands we had at the
4963 previous level. */
4969 narrow_src_p);
4970 }
4973 }
4976 /* Create vectorized promotion statements for vector operands from VEC_OPRNDS0
4977 and VEC_OPRNDS1, for a binary operation associated with scalar statement
4978 STMT_INFO. For multi-step conversions store the resulting vectors and
4979 call the function recursively. */
4981 static void
4987 code_helper ch1,
4989 {
4997 {
5000 else
5003 /* Generate the two halves of promotion operation. */
5006 stmt_info);
5009 stmt_info);
5011 {
5014 }
5015 else
5016 {
5019 }
5021 /* Store the results for the next step. */
5024 }
5028 }
5030 /* Create vectorized promotion stmts for widening stmts using only half the
5031 potential vector size for input. */
5032 static void
5038 code_helper code1,
5040 {
5050 {
5056 /* Widen the first vector input. */
5062 {
5063 /* Widen the second vector input. */
5067 /* Perform the operation. With both vector inputs widened. */
5069 }
5070 else
5071 {
5072 /* Perform the operation. With the single vector input widened. */
5074 }
5080 /* Store the results for the next step. */
5082 }
5086 }
5089 /* Check if STMT_INFO performs a conversion operation that can be vectorized.
5090 If VEC_STMT is also passed, vectorize STMT_INFO: create a vectorized
5091 stmt to replace it, put it in VEC_STMT, and insert it at GSI.
5092 Return true if STMT_INFO is vectorizable in this way. */
5094 static bool
5099 {
5101 tree scalar_dest;
5107 tree new_temp;
5110 poly_uint64 nunits_in;
5111 poly_uint64 nunits_out;
5115 /* For conversions between floating point and integer, there're 2 NARROW
5116 cases. NARROW_SRC is for FLOAT_EXPR, means
5117 integer --DEMOTION--> integer --FLOAT_EXPR--> floating point.
5118 This is safe when the range of the source integer can fit into the lower
5119 precision. NARROW_DST is for FIX_TRUNC_EXPR, means
5120 floating point --FIX_TRUNC_EXPR--> integer --DEMOTION--> INTEGER.
5121 For other conversions, when there's narrowing, NARROW_DST is used as
5122 default. */
5126 tree vop0;
5134 /* Is STMT a vectorizable conversion? */
5155 {
5158 }
5160 {
5163 }
5164 else
5177 /* Check types of lhs and rhs. */
5182 /* Check the operands of the operation. */
5186 {
5191 }
5206 {
5209 "type conversion to/from bit-precision unsupported."
5212 }
5215 {
5222 tree vectype1_in;
5225 {
5230 }
5231 /* For WIDEN_MULT_EXPR, if OP0 is a constant, use the type of
5232 OP1. */
5235 }
5237 /* If op0 is an external or constant def, infer the vector type
5238 from the scalar type. */
5244 {
5250 }
5254 {
5257 "can't convert between boolean and non "
5261 }
5268 else
5272 else
5273 {
5276 }
5278 /* Multiple types in SLP are handled by creating the appropriate number of
5279 vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in
5280 case of SLP. */
5285 else
5288 /* Sanity check: make sure that at least one copy of the vectorized stmt
5289 needs to be generated. */
5295 opt_scalar_mode rhs_mode_iter;
5297 /* Supportable by target? */
5299 {
5308 {
5311 }
5313 /* For conversions between float and integer types try whether
5314 we can use intermediate signed integer types to support the
5315 conversion. */
5319 {
5323 scalar_mode intermediate_mode;
5325 {
5328 }
5329 else
5330 {
5335 }
5338 opt_scalar_mode mode_iter;
5340 {
5346 scalar_mode cvt_mode;
5351 cvt_type = build_nonstandard_integer_type
5354 /* Check if the intermediate type can hold OP0's range.
5355 When converting from float to integer this is not necessary
5356 because values that do not fit the (smaller) target type are
5357 unspecified anyway. */
5359 {
5370 }
5373 /* This should only happened for SLP as long as loop vectorizer
5374 only supports same-sized vector. */
5378 vectype_out,
5381 cvt_type,
5387 }
5390 {
5391 multi_step_cvt++;
5397 }
5398 }
5399 /* FALLTHRU */
5400 unsupported:
5408 {
5417 }
5422 {
5423 /* Binary widening operation can only be supported directly by the
5424 architecture. */
5427 }
5435 {
5440 cvt_type
5447 {
5454 }
5461 else
5465 cvt_type,
5469 {
5472 }
5473 }
5480 else
5481 {
5482 multi_step_cvt++;
5485 }
5499 {
5500 cvt_type
5508 else
5514 }
5515 /* If op0 can be represented with low precision integer,
5516 truncate it to cvt_type and the do FLOAT_EXPR. */
5518 {
5523 cvt_type
5541 {
5545 }
5546 }
5552 }
5555 {
5559 {
5564 }
5567 {
5572 }
5574 {
5576 /* The final packing step produces one vector result per copy. */
5577 unsigned int nvectors
5581 widen_arith);
5582 }
5583 else
5584 {
5586 /* The initial unpacking step produces two vector results
5587 per copy. MULTI_STEP_CVT is 0 for a single conversion,
5588 so >> MULTI_STEP_CVT divides by 2^(number of steps - 1). */
5589 unsigned int nvectors
5590 = (slp_node
5595 widen_arith);
5596 }
5599 }
5601 /* Transform. */
5607 {
5612 }
5614 /* In case of multi-step conversion, we first generate conversion operations
5615 to the intermediate types, and then from that types to the final one.
5616 We create vector destinations for the intermediate type (TYPES) received
5617 from supportable_*_operation, and store them in the correct order
5618 for future use in vect_create_vectorized_*_stmts (). */
5620 bool widen_or_narrow_float_p
5623 widen_or_narrow_float_p
5628 {
5631 {
5633 intermediate_type);
5635 }
5636 }
5640 widen_or_narrow_float_p
5645 {
5647 ;
5649 {
5653 }
5654 }
5657 {
5661 /* vec_dest is intermediate type operand when multi_step_cvt. */
5663 {
5666 }
5669 {
5670 /* Arguments are ready, create the new vector stmt. */
5673 {
5680 }
5688 else
5690 }
5694 /* In case the vectorization factor (VF) is bigger than the number
5695 of elements that we can fit in a vectype (nunits), we have to
5696 generate more than one vector stmt - i.e - we need to "unroll"
5697 the vector stmt by a factor VF/nunits. */
5703 {
5708 }
5709 /* Arguments are ready. Create the new vector stmts. */
5711 {
5715 {
5718 }
5722 op_type);
5723 else
5728 }
5731 {
5734 {
5738 }
5739 else
5744 else
5746 }
5751 /* In case the vectorization factor (VF) is bigger than the number
5752 of elements that we can fit in a vectype (nunits), we have to
5753 generate more than one vector stmt - i.e - we need to "unroll"
5754 the vector stmt by a factor VF/nunits. */
5757 /* Arguments are ready. Create the new vector stmts. */
5760 {
5765 }
5768 multi_step_cvt,
5772 /* After demoting op0 to cvt_type, convert it to dest. */
5774 {
5776 {
5777 /* Arguments are ready, create the new vector stmt. */
5779 gimple *new_stmt
5785 /* This is the last step of the conversion sequence. Store the
5786 vectors in SLP_NODE or in vector info of the scalar statement
5787 (or in STMT_VINFO_RELATED_STMT chain). */
5790 else
5792 }
5793 }
5795 }
5804 }
5806 /* Return true if we can assume from the scalar form of STMT_INFO that
5807 neither the scalar nor the vector forms will generate code. STMT_INFO
5808 is known not to involve a data reference. */
5810 bool
5812 {
5828 }
5830 /* Function vectorizable_assignment.
5832 Check if STMT_INFO performs an assignment (copy) that can be vectorized.
5833 If VEC_STMT is also passed, vectorize the STMT_INFO: create a vectorized
5834 stmt to replace it, put it in VEC_STMT, and insert it at GSI.
5835 Return true if STMT_INFO is vectorizable in this way. */
5837 static bool
5842 {
5843 tree vec_dest;
5844 tree scalar_dest;
5845 tree op;
5847 tree new_temp;
5853 tree vop;
5856 tree vectype_in;
5865 /* Is vectorizable assignment? */
5886 /* Multiple types in SLP are handled by creating the appropriate number of
5887 vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in
5888 case of SLP. */
5891 else
5896 slp_tree slp_op;
5899 {
5904 }
5908 /* We can handle NOP_EXPR conversions that do not change the number
5909 of elements or the vector size. */
5912 && (!vectype_in
5919 {
5922 "can't convert between boolean and non "
5926 }
5928 /* We do not handle bit-precision changes. */
5935 /* But a conversion that does not change the bit-pattern is ok. */
5943 {
5946 "type conversion to/from bit-precision "
5949 }
5952 {
5955 {
5960 }
5965 cost_vec);
5967 }
5969 /* Transform. */
5973 /* Handle def. */
5976 /* Handle use. */
5979 /* Arguments are ready. create the new vector stmt. */
5981 {
5991 else
5993 }
5999 }
6002 /* Return TRUE if CODE (a shift operation) is supported for SCALAR_TYPE
6003 either as shift by a scalar or by a vector. */
6005 bool
6007 {
6009 machine_mode vec_mode;
6010 optab optab;
6012 tree vectype;
6021 {
6027 }
6035 }
6038 /* Function vectorizable_shift.
6040 Check if STMT_INFO performs a shift operation that can be vectorized.
6041 If VEC_STMT is also passed, vectorize the STMT_INFO: create a vectorized
6042 stmt to replace it, put it in VEC_STMT, and insert it at GSI.
6043 Return true if STMT_INFO is vectorizable in this way. */
6045 static bool
6050 {
6051 tree vec_dest;
6052 tree scalar_dest;
6055 tree vectype;
6058 machine_mode vec_mode;
6059 tree new_temp;
6060 optab optab;
6062 machine_mode optab_op2_mode;
6065 poly_uint64 nunits_in;
6066 poly_uint64 nunits_out;
6067 tree vectype_out;
6068 tree op1_vectype;
6087 /* Is STMT a vectorizable binary/unary operation? */
6104 {
6109 }
6111 slp_tree slp_op0;
6114 {
6119 }
6120 /* If op0 is an external or constant def, infer the vector type
6121 from the scalar type. */
6127 {
6132 }
6139 stmt_vec_info op1_def_stmt_info;
6140 slp_tree slp_op1;
6143 {
6148 }
6150 /* Multiple types in SLP are handled by creating the appropriate number of
6151 vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in
6152 case of SLP. */
6155 else
6160 /* Determine whether the shift amount is a vector, or scalar. If the
6161 shift/rotate amount is a vector, use the vector/vector shift optabs. */
6171 {
6172 /* In SLP, need to check whether the shift count is the same,
6173 in loops if it is a constant or invariant, it is always
6174 a scalar shift. */
6176 {
6178 stmt_vec_info slpstmt_info;
6181 {
6185 }
6187 /* For internal SLP defs we have to make sure we see scalar stmts
6188 for all vector elements.
6189 ??? For different vectors we could resort to a different
6190 scalar shift operand but code-generation below simply always
6191 takes the first. */
6196 }
6198 /* If the shift amount is computed by a pattern stmt we cannot
6199 use the scalar amount directly thus give up and use a vector
6200 shift. */
6203 }
6204 else
6205 {
6210 }
6212 /* Vector shifted by vector. */
6215 {
6223 slp_op1);
6224 incompatible_op1_vectype_p
6230 && (!slp_node
6233 {
6236 "unusable type for last operand in"
6239 }
6240 }
6241 /* See if the machine has a vector shifted by scalar insn and if not
6242 then see if it has a vector shifted by vector insn. */
6243 else
6244 {
6248 {
6252 }
6253 else
6254 {
6259 {
6269 slp_op1);
6271 /* Unlike the other binary operators, shifts/rotates have
6272 the rhs being int, instead of the same type as the lhs,
6273 so make sure the scalar is the right type if we are
6274 dealing with vectors of long long/long/short/char. */
6275 incompatible_op1_vectype_p
6276 = (!op1_vectype
6281 {
6284 "unusable type for last operand in"
6287 }
6288 }
6289 }
6290 }
6292 /* Supportable by target? */
6294 {
6299 }
6303 {
6308 }
6309 /* vector lowering cannot optimize vector shifts using word arithmetic. */
6314 {
6318 && (!incompatible_op1_vectype_p
6320 && !vect_maybe_update_slp_op_vectype
6323 {
6328 }
6329 /* Now adjust the constant shift amount in place. */
6331 && incompatible_op1_vectype_p
6333 {
6336 {
6342 }
6343 }
6349 }
6351 /* Transform. */
6358 {
6364 }
6366 /* Handle def. */
6370 {
6371 /* Vector shl and shr insn patterns can be defined with scalar
6372 operand 2 (shift operand). In this case, use constant or loop
6373 invariant op1 directly, without extending it to vector mode
6374 first. */
6377 {
6384 /* Store vec_oprnd1 for every vector stmt to be created.
6385 We check during the analysis that all the shift arguments
6386 are the same.
6387 TODO: Allow different constants for different vector
6388 stmts generated for an SLP instance. */
6392 }
6393 }
6395 {
6397 {
6398 /* If the argument was the same in all lanes create
6399 the correctly typed vector shift amount directly. */
6408 }
6410 /* The constant shift amount has been adjusted in place. */
6411 ;
6412 else
6414 }
6416 /* vec_oprnd1 is available if operand 1 should be of a scalar-type
6417 (a special case for certain kind of vector shifts); otherwise,
6418 operand 1 should be of a vector type (the usual case). */
6423 /* Arguments are ready. Create the new vector stmt. */
6425 {
6426 /* For internal defs where we need to use a scalar shift arg
6427 extract the first lane. */
6429 {
6432 gassign *new_stmt
6435 vop1,
6437 bitsize_zero_node));
6440 }
6441 else
6449 else
6451 }
6460 }
6462 /* Function vectorizable_operation.
6464 Check if STMT_INFO performs a binary, unary or ternary operation that can
6465 be vectorized.
6466 If VEC_STMT is also passed, vectorize STMT_INFO: create a vectorized
6467 stmt to replace it, put it in VEC_STMT, and insert it at GSI.
6468 Return true if STMT_INFO is vectorizable in this way. */
6470 static bool
6475 {
6476 tree vec_dest;
6477 tree scalar_dest;
6479 tree vectype;
6482 machine_mode vec_mode;
6483 tree new_temp;
6485 optab optab;
6490 poly_uint64 nunits_in;
6491 poly_uint64 nunits_out;
6492 tree vectype_out;
6508 /* Is STMT a vectorizable binary/unary operation? */
6513 /* Loads and stores are handled in vectorizable_{load,store}. */
6519 /* Shifts are handled in vectorizable_shift. */
6526 /* Comparisons are handled in vectorizable_comparison. */
6530 /* Conditions are handled in vectorizable_condition. */
6534 /* For pointer addition and subtraction, we should use the normal
6535 plus and minus for the vector operation. */
6541 /* Support only unary or binary operations. */
6544 {
6548 op_type);
6550 }
6555 /* Most operations cannot handle bit-precision types without extra
6556 truncations. */
6560 /* Exception are bitwise binary operations. */
6564 {
6569 }
6571 slp_tree slp_op0;
6574 {
6579 }
6582 /* If op0 is an external or constant def, infer the vector type
6583 from the scalar type. */
6585 {
6586 /* For boolean type we cannot determine vectype by
6587 invariant value (don't know whether it is a vector
6588 of booleans or vector of integers). We use output
6589 vectype because operations on boolean don't change
6590 type. */
6592 {
6594 {
6599 }
6601 }
6602 else
6604 slp_node);
6605 }
6609 {
6616 }
6626 {
6629 {
6634 }
6640 }
6642 {
6645 {
6650 }
6656 }
6658 /* Multiple types in SLP are handled by creating the appropriate number of
6659 vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in
6660 case of SLP. */
6662 {
6665 }
6666 else
6667 {
6670 }
6674 /* Reject attempts to combine mask types with nonmask types, e.g. if
6675 we have an AND between a (nonmask) boolean loaded from memory and
6676 a (mask) boolean result of a comparison.
6678 TODO: We could easily fix these cases up using pattern statements. */
6682 {
6687 }
6689 /* Supportable by target? */
6694 else
6695 {
6698 {
6703 }
6706 }
6710 {
6714 /* When vec_mode is not a vector mode and we verified ops we
6715 do not have to lower like AND are natively supported let
6716 those through even when the mode isn't word_mode. For
6717 ops we have to lower the lowering code assumes we are
6718 dealing with word_mode. */
6722 /* Check only during analysis. */
6724 {
6728 }
6733 }
6741 /* If operating on inactive elements could generate spurious traps,
6742 we need to restrict the operation to active lanes. Note that this
6743 specifically doesn't apply to unhoisted invariants, since they
6744 operate on the same value for every lane.
6746 Similarly, if this operation is part of a reduction, a fully-masked
6747 loop should only change the active lanes of the reduction chain,
6748 keeping the inactive lanes as-is. */
6753 {
6757 {
6760 OPTIMIZE_FOR_SPEED))
6765 OPTIMIZE_FOR_SPEED))
6768 else
6769 {
6772 "can't use a fully-masked loop because no"
6775 }
6776 }
6778 /* Put types on constant and invariant SLP children. */
6783 {
6788 }
6795 {
6796 /* The above vect_model_simple_cost call handles constants
6797 in the prologue and (mis-)costs one of the stmts as
6798 vector stmt. See below for the actual lowering that will
6799 be applied. */
6800 unsigned n
6803 {
6814 /* Bit operations do not have extra cost and are accounted
6815 as vector stmt by vect_model_simple_cost. */
6818 }
6820 {
6821 /* We also need to materialize two large constants. */
6826 }
6827 }
6829 }
6831 /* Transform. */
6840 /* POINTER_DIFF_EXPR has pointer arguments which are vectorized as
6841 vectors with unsigned elements, but the result is signed. So, we
6842 need to compute the MINUS_EXPR into vectype temporary and
6843 VIEW_CONVERT_EXPR it into the final vectype_out result. */
6846 {
6849 }
6850 /* Handle def. */
6851 else
6854 /* In case the vectorization factor (VF) is bigger than the number
6855 of elements that we can fit in a vectype (nunits), we have to generate
6856 more than one vector stmt - i.e - we need to "unroll" the
6857 vector stmt by a factor VF/nunits. In doing so, we record a pointer
6858 from one copy of the vector stmt to the next, in the field
6859 STMT_VINFO_RELATED_STMT. This is necessary in order to allow following
6860 stages to find the correct vector defs to be used when vectorizing
6861 stmts that use the defs of the current stmt. The example below
6862 illustrates the vectorization process when VF=16 and nunits=4 (i.e.,
6863 we need to create 4 vectorized stmts):
6865 before vectorization:
6866 RELATED_STMT VEC_STMT
6867 S1: x = memref - -
6868 S2: z = x + 1 - -
6870 step 1: vectorize stmt S1 (done in vectorizable_load. See more details
6871 there):
6872 RELATED_STMT VEC_STMT
6873 VS1_0: vx0 = memref0 VS1_1 -
6874 VS1_1: vx1 = memref1 VS1_2 -
6875 VS1_2: vx2 = memref2 VS1_3 -
6876 VS1_3: vx3 = memref3 - -
6877 S1: x = load - VS1_0
6878 S2: z = x + 1 - -
6880 step2: vectorize stmt S2 (done here):
6881 To vectorize stmt S2 we first need to find the relevant vector
6882 def for the first operand 'x'. This is, as usual, obtained from
6883 the vector stmt recorded in the STMT_VINFO_VEC_STMT of the stmt
6884 that defines 'x' (S1). This way we find the stmt VS1_0, and the
6885 relevant vector def 'vx0'. Having found 'vx0' we can generate
6886 the vector stmt VS2_0, and as usual, record it in the
6887 STMT_VINFO_VEC_STMT of stmt S2.
6888 When creating the second copy (VS2_1), we obtain the relevant vector
6889 def from the vector stmt recorded in the STMT_VINFO_RELATED_STMT of
6890 stmt VS1_0. This way we find the stmt VS1_1 and the relevant
6891 vector def 'vx1'. Using 'vx1' we create stmt VS2_1 and record a
6892 pointer to it in the STMT_VINFO_RELATED_STMT of the vector stmt VS2_0.
6893 Similarly when creating stmts VS2_2 and VS2_3. This is the resulting
6894 chain of stmts and pointers:
6895 RELATED_STMT VEC_STMT
6896 VS1_0: vx0 = memref0 VS1_1 -
6897 VS1_1: vx1 = memref1 VS1_2 -
6898 VS1_2: vx2 = memref2 VS1_3 -
6899 VS1_3: vx3 = memref3 - -
6900 S1: x = load - VS1_0
6901 VS2_0: vz0 = vx0 + v1 VS2_1 -
6902 VS2_1: vz1 = vx1 + v1 VS2_2 -
6903 VS2_2: vz2 = vx2 + v1 VS2_3 -
6904 VS2_3: vz3 = vx3 + v1 - -
6905 S2: z = x + 1 - VS2_0 */
6909 /* Arguments are ready. Create the new vector stmt. */
6911 {
6918 {
6919 /* Lower the operation. This follows vector lowering. */
6922 tree word_type
6926 tree high_bits
6935 {
6953 else
6958 {
6964 }
6973 }
6974 else
6975 {
6992 }
6994 signs);
7004 }
7006 {
7007 tree mask;
7011 else
7012 /* Dummy mask. */
7022 {
7023 /* Perform the operation on active elements only and take
7024 inactive elements from the reduction chain input. */
7027 }
7028 else
7029 {
7033 }
7035 {
7038 signed char biasval
7043 }
7044 gcall *call
7047 vops);
7053 }
7054 else
7055 {
7057 /* When combining two masks check if either of them is elsewhere
7058 combined with a loop mask, if that's the case we can mark that the
7059 new combined mask doesn't need to be combined with a loop mask. */
7063 {
7065 ncopies}))
7066 {
7072 }
7075 ncopies }))
7076 {
7082 }
7083 }
7092 /* Enter the combined value into the vector cond hash so we don't
7093 AND it with a loop mask again. */
7096 }
7099 {
7102 new_temp);
7107 }
7111 else
7113 }
7123 }
7125 /* A helper function to ensure data reference DR_INFO's base alignment. */
7127 static void
7129 {
7130 /* Alignment is only analyzed for the first element of a DR group,
7131 use that to look at base alignment we need to enforce. */
7138 {
7141 // We should only be able to increase the alignment of a base object if
7142 // we know what its new alignment should be at compile time.
7149 {
7152 }
7154 }
7155 }
7158 /* Function get_group_alias_ptr_type.
7160 Return the alias type for the group starting at FIRST_STMT_INFO. */
7162 static tree
7164 {
7170 {
7174 {
7179 }
7181 }
7183 }
7186 /* Function scan_operand_equal_p.
7188 Helper function for check_scan_store. Compare two references
7189 with .GOMP_SIMD_LANE bases. */
7191 static bool
7193 {
7198 {
7199 machine_mode mode;
7209 {
7215 {
7220 }
7221 }
7222 }
7229 {
7233 {
7236 {
7239 {
7245 {
7248 }
7249 }
7251 {
7254 }
7257 {
7261 }
7270 }
7274 }
7275 }
7277 }
7281 /* Normal permutation. */
7282 scan_store_kind_perm,
7284 /* Whole vector left shift permutation with zero init. */
7285 scan_store_kind_lshift_zero,
7287 /* Whole vector left shift permutation and VEC_COND_EXPR. */
7288 scan_store_kind_lshift_cond
7289 };
7291 /* Function check_scan_store.
7293 Verify if we can perform the needed permutations or whole vector shifts.
7294 Return -1 on failure, otherwise exact log2 of vectype's nunits.
7295 USE_WHOLE_VECTOR is a vector of enum scan_store_kind which operation
7296 to do at each step. */
7298 static int
7301 {
7313 {
7319 {
7322 }
7323 else
7324 {
7329 }
7332 {
7337 {
7341 /* Whole vector shifts shift in zeros, so if init is all zero
7342 constant, there is no need to do anything further. */
7346 {
7351 }
7352 }
7354 }
7356 {
7361 }
7362 }
7365 }
7368 /* Function check_scan_store.
7370 Check magic stores for #pragma omp scan {in,ex}clusive reductions. */
7372 static bool
7375 vect_memory_access_type memory_access_type)
7376 {
7379 tree ref_type;
7383 || mask
7395 {
7400 }
7402 /* We need to pattern match code built by OpenMP lowering and simplified
7403 by following optimizations into something we can handle.
7404 #pragma omp simd reduction(inscan,+:r)
7405 for (...)
7406 {
7407 r += something ();
7408 #pragma omp scan inclusive (r)
7409 use (r);
7410 }
7411 shall have body with:
7412 // Initialization for input phase, store the reduction initializer:
7413 _20 = .GOMP_SIMD_LANE (simduid.3_14(D), 0);
7414 _21 = .GOMP_SIMD_LANE (simduid.3_14(D), 1);
7415 D.2042[_21] = 0;
7416 // Actual input phase:
7417 ...
7418 r.0_5 = D.2042[_20];
7419 _6 = _4 + r.0_5;
7420 D.2042[_20] = _6;
7421 // Initialization for scan phase:
7422 _25 = .GOMP_SIMD_LANE (simduid.3_14(D), 2);
7423 _26 = D.2043[_25];
7424 _27 = D.2042[_25];
7425 _28 = _26 + _27;
7426 D.2043[_25] = _28;
7427 D.2042[_25] = _28;
7428 // Actual scan phase:
7429 ...
7430 r.1_8 = D.2042[_20];
7431 ...
7432 The "omp simd array" variable D.2042 holds the privatized copy used
7433 inside of the loop and D.2043 is another one that holds copies of
7434 the current original list item. The separate GOMP_SIMD_LANE ifn
7435 kinds are there in order to allow optimizing the initializer store
7436 and combiner sequence, e.g. if it is originally some C++ish user
7437 defined reduction, but allow the vectorizer to pattern recognize it
7438 and turn into the appropriate vectorized scan.
7440 For exclusive scan, this is slightly different:
7441 #pragma omp simd reduction(inscan,+:r)
7442 for (...)
7443 {
7444 use (r);
7445 #pragma omp scan exclusive (r)
7446 r += something ();
7447 }
7448 shall have body with:
7449 // Initialization for input phase, store the reduction initializer:
7450 _20 = .GOMP_SIMD_LANE (simduid.3_14(D), 0);
7451 _21 = .GOMP_SIMD_LANE (simduid.3_14(D), 1);
7452 D.2042[_21] = 0;
7453 // Actual input phase:
7454 ...
7455 r.0_5 = D.2042[_20];
7456 _6 = _4 + r.0_5;
7457 D.2042[_20] = _6;
7458 // Initialization for scan phase:
7459 _25 = .GOMP_SIMD_LANE (simduid.3_14(D), 3);
7460 _26 = D.2043[_25];
7461 D.2044[_25] = _26;
7462 _27 = D.2042[_25];
7463 _28 = _26 + _27;
7464 D.2043[_25] = _28;
7465 // Actual scan phase:
7466 ...
7467 r.1_8 = D.2044[_20];
7468 ... */
7471 {
7472 /* Match the D.2042[_21] = 0; store above. Just require that
7473 it is a constant or external definition store. */
7475 {
7476 fail_init:
7481 }
7491 /* These stores can be vectorized normally. */
7493 }
7496 {
7497 fail:
7502 }
7511 bool inscan_var_store
7515 {
7517 {
7518 use_operand_p use_p;
7519 imm_use_iterator iter;
7521 {
7528 || other_store_stmt
7532 }
7538 }
7539 }
7541 {
7542 use_operand_p use_p;
7543 imm_use_iterator iter;
7545 {
7552 }
7553 }
7554 else
7564 /* For pointer addition, we should use the normal plus for the vector
7565 operation. */
7567 {
7575 }
7603 {
7609 tree lrhs;
7612 else
7614 use_operand_p use_p;
7615 imm_use_iterator iter;
7617 {
7624 }
7625 }
7633 stmt_vec_info other_store_stmt_info
7646 {
7650 }
7661 {
7672 }
7691 {
7696 }
7704 /* The IL is as expected, now check if we can actually vectorize it.
7705 Inclusive scan:
7706 _26 = D.2043[_25];
7707 _27 = D.2042[_25];
7708 _28 = _26 + _27;
7709 D.2043[_25] = _28;
7710 D.2042[_25] = _28;
7711 should be vectorized as (where _40 is the vectorized rhs
7712 from the D.2042[_21] = 0; store):
7713 _30 = MEM <vector(8) int> [(int *)&D.2043];
7714 _31 = MEM <vector(8) int> [(int *)&D.2042];
7715 _32 = VEC_PERM_EXPR <_40, _31, { 0, 8, 9, 10, 11, 12, 13, 14 }>;
7716 _33 = _31 + _32;
7717 // _33 = { _31[0], _31[0]+_31[1], _31[1]+_31[2], ..., _31[6]+_31[7] };
7718 _34 = VEC_PERM_EXPR <_40, _33, { 0, 1, 8, 9, 10, 11, 12, 13 }>;
7719 _35 = _33 + _34;
7720 // _35 = { _31[0], _31[0]+_31[1], _31[0]+.._31[2], _31[0]+.._31[3],
7721 // _31[1]+.._31[4], ... _31[4]+.._31[7] };
7722 _36 = VEC_PERM_EXPR <_40, _35, { 0, 1, 2, 3, 8, 9, 10, 11 }>;
7723 _37 = _35 + _36;
7724 // _37 = { _31[0], _31[0]+_31[1], _31[0]+.._31[2], _31[0]+.._31[3],
7725 // _31[0]+.._31[4], ... _31[0]+.._31[7] };
7726 _38 = _30 + _37;
7727 _39 = VEC_PERM_EXPR <_38, _38, { 7, 7, 7, 7, 7, 7, 7, 7 }>;
7728 MEM <vector(8) int> [(int *)&D.2043] = _39;
7729 MEM <vector(8) int> [(int *)&D.2042] = _38;
7730 Exclusive scan:
7731 _26 = D.2043[_25];
7732 D.2044[_25] = _26;
7733 _27 = D.2042[_25];
7734 _28 = _26 + _27;
7735 D.2043[_25] = _28;
7736 should be vectorized as (where _40 is the vectorized rhs
7737 from the D.2042[_21] = 0; store):
7738 _30 = MEM <vector(8) int> [(int *)&D.2043];
7739 _31 = MEM <vector(8) int> [(int *)&D.2042];
7740 _32 = VEC_PERM_EXPR <_40, _31, { 0, 8, 9, 10, 11, 12, 13, 14 }>;
7741 _33 = VEC_PERM_EXPR <_40, _32, { 0, 8, 9, 10, 11, 12, 13, 14 }>;
7742 _34 = _32 + _33;
7743 // _34 = { 0, _31[0], _31[0]+_31[1], _31[1]+_31[2], _31[2]+_31[3],
7744 // _31[3]+_31[4], ... _31[5]+.._31[6] };
7745 _35 = VEC_PERM_EXPR <_40, _34, { 0, 1, 8, 9, 10, 11, 12, 13 }>;
7746 _36 = _34 + _35;
7747 // _36 = { 0, _31[0], _31[0]+_31[1], _31[0]+.._31[2], _31[0]+.._31[3],
7748 // _31[1]+.._31[4], ... _31[3]+.._31[6] };
7749 _37 = VEC_PERM_EXPR <_40, _36, { 0, 1, 2, 3, 8, 9, 10, 11 }>;
7750 _38 = _36 + _37;
7751 // _38 = { 0, _31[0], _31[0]+_31[1], _31[0]+.._31[2], _31[0]+.._31[3],
7752 // _31[0]+.._31[4], ... _31[0]+.._31[6] };
7753 _39 = _30 + _38;
7754 _50 = _31 + _39;
7755 _51 = VEC_PERM_EXPR <_50, _50, { 7, 7, 7, 7, 7, 7, 7, 7 }>;
7756 MEM <vector(8) int> [(int *)&D.2044] = _39;
7757 MEM <vector(8) int> [(int *)&D.2042] = _51; */
7768 }
7771 /* Function vectorizable_scan_store.
7773 Helper of vectorizable_score, arguments like on vectorizable_store.
7774 Handle only the transformation, checking is done in check_scan_store. */
7776 static bool
7780 {
7795 bool inscan_var_store
7799 {
7800 use_operand_p use_p;
7801 imm_use_iterator iter;
7803 {
7809 }
7810 }
7831 {
7835 }
7850 {
7857 else
7858 {
7863 }
7867 {
7874 }
7875 else
7877 }
7898 {
7909 {
7913 dataref_offset);
7919 }
7923 {
7926 (zero_vec
7936 {
7937 /* Whole vector shift shifted in zero bits, but if *init
7938 is not initializer_zerop, we need to replace those elements
7939 with elements from vec_oprnd1. */
7952 }
7954 /* For exclusive scan, perform the perms[i] permutation once
7955 more. */
7959 {
7963 }
7971 }
7979 /* For exclusive scan, new_temp computed above is the exclusive scan
7980 prefix sum. Turn it into inclusive prefix sum for the broadcast
7981 of the last element into orig. */
7983 {
7988 }
7997 {
8000 dataref_offset);
8005 }
8006 }
8010 {
8016 dataref_offset);
8021 }
8023 }
8026 /* Function vectorizable_store.
8028 Check if STMT_INFO defines a non scalar data-ref (array/pointer/structure)
8029 that can be vectorized.
8030 If VEC_STMT is also passed, vectorize STMT_INFO: create a vectorized
8031 stmt to replace it, put it in VEC_STMT, and insert it at GSI.
8032 Return true if STMT_INFO is vectorizable in this way. */
8034 static bool
8039 {
8040 tree data_ref;
8042 tree elem_type;
8045 machine_mode vec_mode;
8046 tree dummy;
8054 stmt_vec_info first_stmt_info;
8060 tree aggr_type;
8061 gather_scatter_info gs_info;
8062 poly_uint64 vf;
8063 vec_load_store_type vls_type;
8064 tree ref_type;
8073 /* Is vectorizable store? */
8078 {
8091 }
8092 else
8093 {
8104 mask_index = vect_slp_child_index_for_operand
8111 }
8113 /* Cannot have hybrid store SLP -- that would mean storing to the
8114 same location twice. */
8121 {
8124 }
8125 else
8128 /* Multiple types in SLP are handled by creating the appropriate number of
8129 vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in
8130 case of SLP. */
8133 else
8138 /* FORNOW. This restriction should be relaxed. */
8140 {
8145 }
8147 tree op;
8148 slp_tree op_node;
8159 vect_memory_access_type memory_access_type;
8162 poly_int64 poffset;
8163 internal_fn lanes_ifn;
8171 {
8173 {
8178 }
8182 {
8187 }
8191 {
8196 }
8197 }
8198 else
8199 {
8200 /* FORNOW. In some cases can vectorize even if data-type not supported
8201 (e.g. - array initialization with 0). */
8204 }
8211 {
8215 }
8216 else
8217 {
8221 }
8224 {
8226 memory_access_type))
8228 }
8232 {
8240 mask);
8244 || (mask
8246 mask_vectype))))
8247 {
8252 }
8263 /* As function vect_transform_stmt shows, for interleaving stores
8264 the whole chain is vectorized when the last store in the chain
8265 is reached, the other stores in the group are skipped. So we
8266 want to only cost the last one here, but it's not trivial to
8267 get the last, as it's equivalent to use the first one for
8268 costing, use the first one instead. */
8270 && !slp
8273 }
8276 /* Transform. */
8281 {
8285 {
8296 "vect_model_store_cost: inside_cost = %d, "
8301 }
8303 }
8306 {
8307 /* FORNOW */
8311 {
8313 /* VEC_NUM is the number of vect stmts to be created for this
8314 group. */
8321 }
8322 else
8323 /* VEC_NUM is the number of vect stmts to be created for this
8324 group. */
8328 }
8329 else
8334 ncopies);
8336 /* Check if we need to update prologue cost for invariant,
8337 and update it accordingly if so. If it's not for
8338 interleaving store, we can just check vls_type; but if
8339 it's for interleaving store, need to check the def_type
8340 of the stored value since the current vls_type is just
8341 for first_stmt_info. */
8343 {
8348 {
8354 }
8359 };
8363 {
8365 gimple_stmt_iterator incr_gsi;
8368 tree offvar;
8369 tree ivstep;
8370 tree running_off;
8373 tree dr_offset;
8375 /* Checked by get_load_store_type. */
8382 stride_base
8383 = fold_build_pointer_plus
8390 /* For a store with loop-invariant (but other than power-of-2)
8391 stride (i.e. not a grouped access) like so:
8393 for (i = 0; i < n; i += stride)
8394 array[i] = ...;
8396 we generate a new induction variable and new stores from
8397 the components of the (vectorized) rhs:
8399 for (j = 0; ; j += VF*stride)
8400 vectemp = ...;
8401 tmp1 = vectemp[0];
8402 array[j] = tmp1;
8403 tmp2 = vectemp[1];
8404 array[j + stride] = tmp2;
8405 ...
8406 */
8413 {
8416 {
8422 /* First check if vec_extract optab doesn't support extraction
8423 of vector elts directly. */
8425 machine_mode vmode;
8432 {
8433 /* Try to avoid emitting an extract of vector elements
8434 by performing the extracts using an integer type of the
8435 same size, extracting from a vector of those and then
8436 re-interpreting it as the original vector type if
8437 supported. */
8438 unsigned lsize
8441 /* If we can't construct such a vector fall back to
8442 element extracts from the original vector type and
8443 element size stores. */
8451 {
8456 }
8457 /* Else fall back to vector extraction anyway.
8458 Fewer stores are more important than avoiding spilling
8459 of the vector we extract from. Compared to the
8460 construction case in vectorizable_load no store-forwarding
8461 issue exists here for reasonable archs. */
8462 }
8463 }
8466 {
8468 dr_alignment_support dr_align
8470 mis_align);
8473 {
8480 }
8481 }
8484 }
8487 {
8501 }
8506 /* For costing some adjacent vector stores, we'd like to cost with
8507 the total number of them once instead of cost each one by one. */
8510 {
8513 {
8515 {
8517 tree pos
8524 }
8525 }
8531 else
8534 unsigned HOST_WIDE_INT
8537 {
8539 {
8541 /* Pun the vector to extract from if necessary. */
8543 {
8545 tree cvt
8550 }
8551 }
8553 {
8555 {
8556 /* Only need vector extracting when there are more
8557 than one stores. */
8559 inside_cost
8562 /* Take a single lane vector type store as scalar
8563 store to avoid ICE like 110776. */
8566 n_adjacent_stores++;
8567 else
8568 inside_cost
8572 }
8576 /* Extract the i'th component. */
8584 GSI_SAME_STMT);
8592 /* And store it to *running_off. */
8599 {
8607 }
8610 {
8614 }
8615 }
8616 }
8621 }
8624 {
8631 "vect_model_store_cost: inside_cost = %d, "
8634 }
8637 }
8640 vec_loop_masks *loop_masks
8644 vec_loop_lens *loop_lens
8649 /* Shouldn't go with length-based approach if fully masked. */
8652 /* Targets with store-lane instructions must not require explicit
8653 realignment. vect_supportable_dr_alignment always returns either
8654 dr_aligned or dr_unaligned_supported for masked operations. */
8656 && !mask
8665 tree bump;
8668 {
8671 }
8673 {
8678 }
8679 else
8680 {
8683 else
8687 }
8692 /* In case the vectorization factor (VF) is bigger than the number
8693 of elements that we can fit in a vectype (nunits), we have to generate
8694 more than one vector stmt - i.e - we need to "unroll" the
8695 vector stmt by a factor VF/nunits. */
8697 /* In case of interleaving (non-unit grouped access):
8699 S1: &base + 2 = x2
8700 S2: &base = x0
8701 S3: &base + 1 = x1
8702 S4: &base + 3 = x3
8704 We create vectorized stores starting from base address (the access of the
8705 first stmt in the chain (S2 in the above example), when the last store stmt
8706 of the chain (S4) is reached:
8708 VS1: &base = vx2
8709 VS2: &base + vec_size*1 = vx0
8710 VS3: &base + vec_size*2 = vx1
8711 VS4: &base + vec_size*3 = vx3
8713 Then permutation statements are generated:
8715 VS5: vx5 = VEC_PERM_EXPR < vx0, vx3, {0, 8, 1, 9, 2, 10, 3, 11} >
8716 VS6: vx6 = VEC_PERM_EXPR < vx0, vx3, {4, 12, 5, 13, 6, 14, 7, 15} >
8717 ...
8719 And they are put in STMT_VINFO_VEC_STMT of the corresponding scalar stmts
8720 (the order of the data-refs in the output of vect_permute_store_chain
8721 corresponds to the order of scalar stmts in the interleaving chain - see
8722 the documentation of vect_permute_store_chain()).
8724 In case of both multiple types and interleaving, above vector stores and
8725 permutation stmts are created for every copy. The result vector stmts are
8726 put in STMT_VINFO_VEC_STMT for the first copy and in the corresponding
8727 STMT_VINFO_RELATED_STMT for the next copies.
8728 */
8738 {
8741 /* For costing some adjacent vector stores, we'd like to cost with
8742 the total number of them once instead of cost each one by one. */
8745 {
8748 {
8749 /* For interleaved stores we collect vectorized defs for all
8750 the stores in the group in DR_CHAIN. DR_CHAIN is then used
8751 as an input to vect_permute_store_chain(). */
8754 {
8755 /* Since gaps are not supported for interleaved stores,
8756 DR_GROUP_SIZE is the exact number of stmts in the
8757 chain. Therefore, NEXT_STMT_INFO can't be NULL_TREE. */
8761 else
8762 {
8768 }
8770 }
8773 {
8775 {
8778 mask_vectype);
8780 }
8782 /* We should have catched mismatched types earlier. */
8785 dataref_ptr
8789 }
8790 }
8792 {
8794 /* DR_CHAIN is then used as an input to
8795 vect_permute_store_chain(). */
8797 {
8800 }
8805 }
8808 {
8811 }
8813 /* Get an array into which we can store the individual vectors. */
8816 /* Invalidate the current contents of VEC_ARRAY. This should
8817 become an RTL clobber too, which prevents the vector registers
8818 from being upward-exposed. */
8821 /* Store the individual vectors into the array. */
8823 {
8826 i);
8827 }
8840 {
8844 else
8846 signed char biasval
8850 {
8853 }
8854 }
8858 {
8859 /* Emit:
8860 MASK_LEN_STORE_LANES (DATAREF_PTR, ALIAS_PTR, VEC_MASK,
8861 LEN, BIAS, VEC_ARRAY). */
8867 vec_array);
8868 }
8870 {
8871 /* Emit:
8872 MASK_STORE_LANES (DATAREF_PTR, ALIAS_PTR, VEC_MASK,
8873 VEC_ARRAY). */
8879 }
8880 else
8881 {
8882 /* Emit:
8883 MEM_REF[...all elements...] = STORE_LANES (VEC_ARRAY). */
8887 }
8892 /* Record that VEC_ARRAY is now dead. */
8897 }
8900 {
8907 "vect_model_store_cost: inside_cost = %d, "
8910 }
8913 }
8916 {
8921 {
8924 {
8929 {
8930 /* Since the store is not grouped, DR_GROUP_SIZE is 1, and
8931 DR_CHAIN is of size 1. */
8935 else
8939 {
8942 else
8944 ncopies,
8946 mask_vectype);
8947 }
8953 else
8954 dataref_ptr
8958 bump);
8959 }
8960 }
8962 {
8967 }
8971 {
8973 {
8977 /* We should have catched mismatched types earlier. */
8980 }
8986 {
8994 }
8997 {
8999 {
9001 inside_cost
9005 }
9012 {
9017 else
9019 signed char biasval
9023 {
9026 }
9027 }
9031 call = gimple_build_call_internal
9036 call = gimple_build_call_internal
9039 else
9046 }
9048 {
9049 /* The builtin decls path for scatter is legacy, x86 only. */
9051 && (!final_mask
9052 || SCALAR_INT_MODE_P
9055 {
9057 inside_cost
9061 }
9062 poly_uint64 offset_nunits
9065 {
9066 new_stmt = vect_build_one_scatter_store_call
9072 }
9074 {
9075 /* We have a offset vector with half the number of
9076 lanes but the builtins will store full vectype
9077 data from the lower lanes. */
9078 new_stmt = vect_build_one_scatter_store_call
9080 dataref_ptr,
9091 tree perm_mask
9095 perm_mask);
9101 {
9103 VEC_UNPACK_HI_EXPR,
9104 final_mask);
9105 final_mask = make_ssa_name
9110 }
9111 new_stmt = vect_build_one_scatter_store_call
9113 dataref_ptr,
9118 }
9120 {
9121 /* We have a offset vector with double the number of
9122 lanes. Select the low/high part accordingly. */
9125 {
9132 tree perm_mask = vect_gen_perm_mask_checked
9135 VEC_PERM_EXPR,
9136 vec_offset,
9137 vec_offset,
9138 perm_mask);
9143 }
9144 new_stmt = vect_build_one_scatter_store_call
9150 }
9151 else
9153 }
9154 else
9155 {
9156 /* Emulated scatter. */
9159 {
9161 /* For emulated scatter N offset vector element extracts
9162 (we assume the scalar scaling and ptr + offset add is
9163 consumed by the load). */
9164 inside_cost
9167 /* N scalar stores plus extracting the elements. */
9168 inside_cost
9171 inside_cost
9175 }
9178 unsigned HOST_WIDE_INT const_offset_nunits
9184 unsigned HOST_WIDE_INT elt_size
9186 /* We support offset vectors with more elements
9187 than the data vector for now. */
9188 unsigned HOST_WIDE_INT factor
9191 unsigned elt_offset
9198 {
9199 /* Compute the offsetted pointer. */
9202 tree idx
9208 tree ptr
9213 /* Extract the element to be stored. */
9214 tree elt
9221 tree ref
9227 }
9230 }
9231 }
9234 }
9241 "vect_model_store_cost: inside_cost = %d, "
9246 }
9254 /* For costing some adjacent vector stores, we'd like to cost with
9255 the total number of them once instead of cost each one by one. */
9260 {
9263 {
9265 {
9266 /* Get vectorized arguments for SLP_NODE. */
9272 }
9273 else
9274 {
9275 /* For interleaved stores we collect vectorized defs for all the
9276 stores in the group in DR_CHAIN. DR_CHAIN is then used as an
9277 input to vect_permute_store_chain().
9279 If the store is not grouped, DR_GROUP_SIZE is 1, and DR_CHAIN
9280 is of size 1. */
9283 {
9284 /* Since gaps are not supported for interleaved stores,
9285 DR_GROUP_SIZE is the exact number of stmts in the chain.
9286 Therefore, NEXT_STMT_INFO can't be NULL_TREE. In case
9287 that there is no interleaving, DR_GROUP_SIZE is 1,
9288 and only one iteration of the loop will be executed. */
9292 else
9293 {
9299 }
9301 }
9303 {
9306 mask_vectype);
9308 }
9309 }
9311 /* We should have catched mismatched types earlier. */
9315 bool simd_lane_access_p
9318 && simd_lane_access_p
9319 && !loop_masks
9326 {
9329 }
9331 dataref_ptr
9336 }
9338 {
9340 /* DR_CHAIN is then used as an input to vect_permute_store_chain().
9341 If the store is not grouped, DR_GROUP_SIZE is 1, and DR_CHAIN is
9342 of size 1. */
9344 {
9347 }
9352 else
9355 }
9359 {
9360 /* Permute. */
9363 {
9370 "vect_model_store_cost: "
9372 group_size);
9373 }
9374 else
9377 }
9381 {
9383 {
9387 /* For grouped stores vectorized defs are interleaved in
9388 vect_permute_store_chain(). */
9390 }
9393 {
9397 else
9398 {
9404 /* Generate the permute statement. */
9405 gimple *perm_stmt
9409 gsi);
9413 }
9414 }
9417 {
9418 n_adjacent_stores++;
9421 {
9425 }
9428 }
9444 /* Bump the vector pointer. */
9454 {
9457 }
9458 else
9463 misalign);
9466 /* Compute IFN when LOOP_LENS or final_mask valid. */
9471 {
9472 opt_machine_mode new_ovmode
9475 unsigned factor
9480 }
9482 {
9487 }
9490 {
9492 {
9493 /* Pass VF value to 'len' argument of
9494 MASK_LEN_STORE if LOOP_LENS is invalid. */
9496 }
9498 {
9499 /* Pass all ones value to 'mask' argument of
9500 MASK_LEN_STORE if final_mask is invalid. */
9503 }
9504 }
9506 {
9507 signed char biasval
9511 }
9513 /* Arguments are ready. Create the new vector stmt. */
9515 {
9518 /* Need conversion if it's wrapped with VnQI. */
9520 {
9521 tree new_vtype
9523 new_vmode);
9526 gassign *new_stmt
9530 }
9536 else
9543 }
9545 {
9547 gcall *call
9553 }
9554 else
9555 {
9556 data_ref
9558 dataref_offset ? dataref_offset
9561 ;
9562 else
9569 }
9577 }
9579 {
9583 }
9584 }
9587 {
9593 /* When vectorizing a store into the function result assign
9594 a penalty if the function returns in a multi-register location.
9595 In this case we assume we'll end up with having to spill the
9596 vector result and do piecewise loads as a conservative estimate. */
9602 {
9604 /* ??? Handle PARALLEL in some way. */
9606 {
9608 /* Assume that a single reg-reg move is possible and cheap,
9609 do not account for vector to gp register move cost. */
9611 {
9612 /* Spill. */
9613 prologue_cost
9616 /* Loads. */
9617 prologue_cost
9620 }
9621 }
9622 }
9625 "vect_model_store_cost: inside_cost = %d, "
9628 }
9631 }
9633 /* Given a vector type VECTYPE, turns permutation SEL into the equivalent
9634 VECTOR_CST mask. No checks are made that the target platform supports the
9635 mask, so callers may wish to test can_vec_perm_const_p separately, or use
9636 vect_gen_perm_mask_checked. */
9638 tree
9640 {
9641 tree mask_type;
9648 }
9650 /* Checked version of vect_gen_perm_mask_any. Asserts can_vec_perm_const_p,
9651 i.e. that the target supports the pattern _for arbitrary input vectors_. */
9653 tree
9655 {
9659 }
9661 /* Given a vector variable X and Y, that was generated for the scalar
9662 STMT_INFO, generate instructions to permute the vector elements of X and Y
9663 using permutation mask MASK_VEC, insert them at *GSI and return the
9664 permuted vector variable. */
9666 static tree
9670 {
9678 else
9682 /* Generate the permute statement. */
9687 }
9689 /* Hoist the definitions of all SSA uses on STMT_INFO out of the loop LOOP,
9690 inserting them on the loops preheader edge. Returns true if we
9691 were successful in doing so (and thus STMT_INFO can be moved then),
9692 otherwise returns false. HOIST_P indicates if we want to hoist the
9693 definitions of all SSA uses, it would be false when we are costing. */
9695 static bool
9697 {
9698 ssa_op_iter i;
9699 tree op;
9703 {
9707 {
9708 /* Make sure we don't need to recurse. While we could do
9709 so in simple cases when there are more complex use webs
9710 we don't have an easy way to preserve stmt order to fulfil
9711 dependencies within them. */
9712 tree op2;
9713 ssa_op_iter i2;
9717 {
9722 }
9724 }
9725 }
9734 {
9738 {
9742 }
9743 }
9746 }
9748 /* vectorizable_load.
9750 Check if STMT_INFO reads a non scalar data-ref (array/pointer/structure)
9751 that can be vectorized.
9752 If VEC_STMT is also passed, vectorize STMT_INFO: create a vectorized
9753 stmt to replace it, put it in VEC_STMT, and insert it at GSI.
9754 Return true if STMT_INFO is vectorizable in this way. */
9756 static bool
9761 {
9762 tree scalar_dest;
9769 tree elem_type;
9770 /* Avoid false positive uninitialized warning, see PR110652. */
9772 machine_mode mode;
9773 tree dummy;
9780 poly_uint64 group_gap_adj;
9786 stmt_vec_info first_stmt_info;
9794 poly_uint64 vf;
9795 tree aggr_type;
9796 gather_scatter_info gs_info;
9797 tree ref_type;
9814 {
9829 }
9830 else
9831 {
9846 mask_index = vect_slp_child_index_for_operand
9852 }
9858 {
9862 }
9863 else
9866 /* Multiple types in SLP are handled by creating the appropriate number of
9867 vectorized stmts for each SLP node. Hence, NCOPIES is always 1 in
9868 case of SLP. */
9871 else
9876 /* FORNOW. This restriction should be relaxed. */
9878 {
9883 }
9885 /* Invalidate assumptions made by dependence analysis when vectorization
9886 on the unrolled body effectively re-orders stmts. */
9891 {
9894 "cannot perform implicit CSE when unrolling "
9897 }
9902 /* FORNOW. In some cases can vectorize even if data-type not supported
9903 (e.g. - data copies). */
9905 {
9910 }
9912 /* Check if the load is a part of an interleaving chain. */
9914 {
9916 /* FORNOW */
9923 /* Refuse non-SLP vectorization of SLP-only groups. */
9925 {
9930 }
9932 /* Invalidate assumptions made by dependence analysis when vectorization
9933 on the unrolled body effectively re-orders stmts. */
9938 {
9941 "cannot perform implicit CSE when performing "
9944 }
9945 }
9946 else
9950 {
9954 {
9955 /* In BB vectorization we may not actually use a loaded vector
9956 accessing elements in excess of DR_GROUP_SIZE. */
9967 {
9970 "BB vectorization with gaps at the end of "
9973 }
9974 }
9980 {
9983 vect_location,
9986 }
9987 }
9989 vect_memory_access_type memory_access_type;
9992 poly_int64 poffset;
9993 internal_fn lanes_ifn;
10001 {
10003 {
10009 }
10012 {
10017 }
10021 {
10026 }
10027 }
10032 {
10034 && mask
10036 mask_vectype))
10037 {
10042 }
10052 mask);
10065 }
10075 /* Transform. */
10081 {
10083 /* If we have versioned for aliasing or the loop doesn't
10084 have any data dependencies that would preclude this,
10085 then we are sure this is a loop invariant load and
10086 thus we can insert it on the preheader edge.
10087 TODO: hoist_defs_of_uses should ideally be computed
10088 once at analysis time, remembered and used in the
10089 transform time. */
10091 && !nested_in_vect_loop
10094 {
10095 enum vect_cost_model_location cost_loc
10100 cost_loc);
10105 "vect_model_load_cost: inside_cost = %d, "
10109 }
10111 {
10121 tree vuse;
10124 else
10129 }
10130 /* These copies are all equivalent. */
10134 else
10135 {
10140 }
10145 else
10146 {
10150 }
10152 }
10156 {
10157 gimple_stmt_iterator incr_gsi;
10159 tree offvar;
10160 tree ivstep;
10161 tree running_off;
10164 /* Checked by get_load_store_type. */
10167 tree dr_offset;
10174 {
10177 }
10178 else
10179 {
10182 }
10185 {
10188 }
10189 else
10190 {
10192 cst_offset
10195 first_stmt_info));
10198 }
10201 {
10209 /* For a load with loop-invariant (but other than power-of-2)
10210 stride (i.e. not a grouped access) like so:
10212 for (i = 0; i < n; i += stride)
10213 ... = array[i];
10215 we generate a new induction variable and new accesses to
10216 form a new vector (or vectors, depending on ncopies):
10218 for (j = 0; ; j += VF*stride)
10219 tmp1 = array[j];
10220 tmp2 = array[j + stride];
10221 ...
10222 vectemp = {tmp1, tmp2, ...}
10223 */
10237 }
10247 {
10249 {
10250 /* First check if vec_init optab supports construction from vector
10251 elts directly. Otherwise avoid emitting a constructor of
10252 vector elements by performing the loads using an integer type
10253 of the same size, constructing a vector of those and then
10254 re-interpreting it as the original vector type. This avoids a
10255 huge runtime penalty due to the general inability to perform
10256 store forwarding from smaller stores to a larger load. */
10257 tree ptype;
10258 tree vtype
10263 {
10268 }
10269 }
10270 else
10271 {
10275 }
10277 }
10278 /* Load vector(1) scalar_type if it's 1 element-wise vectype. */
10283 {
10284 /* For SLP permutation support we need to load the whole group,
10285 not only the number of vector stmts the permutation result
10286 fits in. */
10288 {
10289 /* We don't yet generate SLP_TREE_LOAD_PERMUTATIONs for
10290 variable VF. */
10294 }
10295 else
10297 }
10299 unsigned HOST_WIDE_INT
10302 /* For costing some adjacent vector loads, we'd like to cost with
10303 the total number of them once instead of cost each one by one. */
10306 {
10311 {
10313 {
10314 /* For VMAT_ELEMENTWISE, just cost it as scalar_load to
10315 avoid ICE, see PR110776. */
10318 n_adjacent_loads++;
10319 else
10323 }
10337 {
10338 n_groups++;
10339 /* When doing SLP make sure to not load elements from
10340 the next vector iteration, those will not be accessed
10341 so just use the last element again. See PR107451. */
10343 {
10345 gimple *incr
10350 }
10352 }
10353 }
10356 {
10360 else
10361 {
10367 {
10368 new_stmt
10370 VIEW_CONVERT_EXPR,
10374 gsi);
10375 }
10376 }
10377 }
10380 {
10382 {
10385 else
10387 }
10388 else
10389 {
10393 }
10394 }
10395 }
10397 {
10400 {
10406 }
10407 else
10410 }
10413 {
10421 "vect_model_load_cost: inside_cost = %u, "
10423 inside_cost);
10424 }
10427 }
10435 {
10437 {
10440 }
10441 else
10442 {
10445 }
10446 /* For SLP vectorization we directly vectorize a subchain
10447 without permutation. */
10450 /* For BB vectorization always use the first stmt to base
10451 the data ref pointer on. */
10453 first_stmt_info_for_drptr
10456 /* Check if the chain of loads is already vectorized. */
10458 /* For SLP we would need to copy over SLP_TREE_VEC_DEFS.
10459 ??? But we can only do so if there is exactly one
10460 as we have no way to get at the rest. Leave the CSE
10461 opportunity alone.
10462 ??? With the group load eventually participating
10463 in multiple different permutations (having multiple
10464 slp nodes which refer to the same group) the CSE
10465 is even wrong code. See PR56270. */
10467 {
10470 }
10474 /* VEC_NUM is the number of vect stmts to be created for this group. */
10476 {
10478 /* If an SLP permutation is from N elements to N elements,
10479 and if one vector holds a whole number of N, we can load
10480 the inputs to the permutation in the same way as an
10481 unpermuted sequence. In other cases we need to load the
10482 whole group, not only the number of vector stmts the
10483 permutation result fits in. */
10488 {
10489 /* We don't yet generate such SLP_TREE_LOAD_PERMUTATIONs for
10490 variable VF; see vect_transform_slp_perm_load. */
10495 }
10496 else
10497 {
10499 group_gap_adj
10501 }
10502 }
10503 else
10507 }
10508 else
10509 {
10517 }
10520 vec_loop_masks *loop_masks
10524 vec_loop_lens *loop_lens
10529 /* Shouldn't go with length-based approach if fully masked. */
10532 /* Targets with store-lane instructions must not require explicit
10533 realignment. vect_supportable_dr_alignment always returns either
10534 dr_aligned or dr_unaligned_supported for masked operations. */
10536 && !mask
10541 /* In case the vectorization factor (VF) is bigger than the number
10542 of elements that we can fit in a vectype (nunits), we have to generate
10543 more than one vector stmt - i.e - we need to "unroll" the
10544 vector stmt by a factor VF/nunits. In doing so, we record a pointer
10545 from one copy of the vector stmt to the next, in the field
10546 STMT_VINFO_RELATED_STMT. This is necessary in order to allow following
10547 stages to find the correct vector defs to be used when vectorizing
10548 stmts that use the defs of the current stmt. The example below
10549 illustrates the vectorization process when VF=16 and nunits=4 (i.e., we
10550 need to create 4 vectorized stmts):
10552 before vectorization:
10553 RELATED_STMT VEC_STMT
10554 S1: x = memref - -
10555 S2: z = x + 1 - -
10557 step 1: vectorize stmt S1:
10558 We first create the vector stmt VS1_0, and, as usual, record a
10559 pointer to it in the STMT_VINFO_VEC_STMT of the scalar stmt S1.
10560 Next, we create the vector stmt VS1_1, and record a pointer to
10561 it in the STMT_VINFO_RELATED_STMT of the vector stmt VS1_0.
10562 Similarly, for VS1_2 and VS1_3. This is the resulting chain of
10563 stmts and pointers:
10564 RELATED_STMT VEC_STMT
10565 VS1_0: vx0 = memref0 VS1_1 -
10566 VS1_1: vx1 = memref1 VS1_2 -
10567 VS1_2: vx2 = memref2 VS1_3 -
10568 VS1_3: vx3 = memref3 - -
10569 S1: x = load - VS1_0
10570 S2: z = x + 1 - -
10571 */
10573 /* In case of interleaving (non-unit grouped access):
10575 S1: x2 = &base + 2
10576 S2: x0 = &base
10577 S3: x1 = &base + 1
10578 S4: x3 = &base + 3
10580 Vectorized loads are created in the order of memory accesses
10581 starting from the access of the first stmt of the chain:
10583 VS1: vx0 = &base
10584 VS2: vx1 = &base + vec_size*1
10585 VS3: vx3 = &base + vec_size*2
10586 VS4: vx4 = &base + vec_size*3
10588 Then permutation statements are generated:
10590 VS5: vx5 = VEC_PERM_EXPR < vx0, vx1, { 0, 2, ..., i*2 } >
10591 VS6: vx6 = VEC_PERM_EXPR < vx0, vx1, { 1, 3, ..., i*2+1 } >
10592 ...
10594 And they are put in STMT_VINFO_VEC_STMT of the corresponding scalar stmts
10595 (the order of the data-refs in the output of vect_permute_load_chain
10596 corresponds to the order of scalar stmts in the interleaving chain - see
10597 the documentation of vect_permute_load_chain()).
10598 The generation of permutation stmts and recording them in
10599 STMT_VINFO_VEC_STMT is done in vect_transform_grouped_load().
10601 In case of both multiple types and interleaving, the vector loads and
10602 permutation stmts above are created for every copy. The result vector
10603 stmts are put in STMT_VINFO_VEC_STMT for the first copy and in the
10604 corresponding STMT_VINFO_RELATED_STMT for the next copies. */
10606 /* If the data reference is aligned (dr_aligned) or potentially unaligned
10607 on a target that supports unaligned accesses (dr_unaligned_supported)
10608 we generate the following code:
10609 p = initial_addr;
10610 indx = 0;
10611 loop {
10612 p = p + indx * vectype_size;
10613 vec_dest = *(p);
10614 indx = indx + 1;
10615 }
10617 Otherwise, the data reference is potentially unaligned on a target that
10618 does not support unaligned accesses (dr_explicit_realign_optimized) -
10619 then generate the following code, in which the data in each iteration is
10620 obtained by two vector loads, one from the previous iteration, and one
10621 from the current iteration:
10622 p1 = initial_addr;
10623 msq_init = *(floor(p1))
10624 p2 = initial_addr + VS - 1;
10625 realignment_token = call target_builtin;
10626 indx = 0;
10627 loop {
10628 p2 = p2 + indx * vectype_size
10629 lsq = *(floor(p2))
10630 vec_dest = realign_load (msq, lsq, realignment_token)
10631 indx = indx + 1;
10632 msq = lsq;
10633 } */
10635 /* If the misalignment remains the same throughout the execution of the
10636 loop, we can create the init_addr and permutation mask at the loop
10637 preheader. Otherwise, it needs to be created inside the loop.
10638 This can only occur when vectorizing memory accesses in the inner-loop
10639 nested within an outer-loop that is being vectorized. */
10644 {
10647 }
10649 bool diff_first_stmt_info
10656 {
10657 /* If we have different first_stmt_info, we can't set up realignment
10658 here, since we can't guarantee first_stmt_info DR has been
10659 initialized yet, use first_stmt_info_for_drptr DR by bumping the
10660 distance from first_stmt_info DR instead as below. */
10662 {
10669 {
10672 size_one_node);
10674 }
10675 }
10676 }
10677 else
10681 offset = (offset
10685 tree bump;
10688 {
10691 }
10693 {
10698 }
10699 else
10700 {
10703 else
10707 }
10712 {
10716 else
10719 }
10723 {
10729 /* For costing some adjacent vector loads, we'd like to cost with
10730 the total number of them once instead of cost each one by one. */
10733 {
10735 {
10736 /* An IFN_LOAD_LANES will load all its vector results,
10737 regardless of which ones we actually need. Account
10738 for the cost of unused results. */
10740 {
10743 do
10744 {
10747 }
10750 {
10753 "vect_model_load_cost: %d "
10755 gaps);
10757 alignment_support_scheme,
10761 }
10762 }
10763 n_adjacent_loads++;
10765 }
10767 /* 1. Create the vector or array pointer update chain. */
10769 dataref_ptr
10773 else
10774 {
10778 }
10795 {
10799 else
10801 signed char biasval
10805 {
10808 }
10809 }
10813 {
10814 /* Emit:
10815 VEC_ARRAY = MASK_LEN_LOAD_LANES (DATAREF_PTR, ALIAS_PTR,
10816 VEC_MASK, LEN, BIAS). */
10822 }
10824 {
10825 /* Emit:
10826 VEC_ARRAY = MASK_LOAD_LANES (DATAREF_PTR, ALIAS_PTR,
10827 VEC_MASK). */
10832 final_mask);
10833 }
10834 else
10835 {
10836 /* Emit:
10837 VEC_ARRAY = LOAD_LANES (MEM_REF[...all elements...]). */
10840 }
10846 /* Extract each vector into an SSA_NAME. */
10848 {
10852 }
10854 /* Record the mapping between SSA_NAMEs and statements. */
10857 /* Record that VEC_ARRAY is now dead. */
10863 }
10866 {
10874 "vect_model_load_cost: inside_cost = %u, "
10877 }
10880 }
10883 {
10890 {
10891 /* 1. Create the vector or array pointer update chain. */
10893 {
10898 else
10899 dataref_ptr
10903 }
10905 {
10910 }
10914 {
10919 {
10923 final_mask
10934 }
10936 /* 2. Create the vector-load in the loop. */
10939 {
10941 {
10943 inside_cost
10947 }
10954 {
10956 final_len
10960 else
10961 final_len
10964 signed char biasval
10968 {
10971 }
10972 }
10976 call
10985 else
10992 }
10994 {
10995 /* The builtin decls path for gather is legacy, x86 only. */
10998 {
11000 inside_cost
11004 }
11005 poly_uint64 offset_nunits
11008 {
11009 new_stmt = vect_build_one_gather_load_call
11012 final_mask);
11014 }
11016 {
11017 /* We have a offset vector with half the number of
11018 lanes but the builtins will produce full vectype
11019 data with just the lower lanes filled. */
11020 new_stmt = vect_build_one_gather_load_call
11023 final_mask);
11029 /* now put upper half of final_mask in final_mask low. */
11031 && !SCALAR_INT_MODE_P
11033 {
11040 tree perm_mask = vect_gen_perm_mask_checked
11043 VEC_PERM_EXPR,
11044 final_mask,
11045 final_mask,
11046 perm_mask);
11051 }
11053 {
11055 VEC_UNPACK_HI_EXPR,
11056 final_mask);
11057 final_mask = make_ssa_name
11062 }
11064 new_stmt = vect_build_one_gather_load_call
11066 dataref_ptr,
11068 final_mask);
11074 /* compose low + high. */
11081 tree perm_mask
11084 VEC_PERM_EXPR,
11087 }
11089 {
11090 /* We have a offset vector with double the number of
11091 lanes. Select the low/high part accordingly. */
11094 {
11101 tree perm_mask = vect_gen_perm_mask_checked
11104 VEC_PERM_EXPR,
11105 vec_offset,
11106 vec_offset,
11107 perm_mask);
11112 }
11113 new_stmt = vect_build_one_gather_load_call
11117 }
11118 else
11120 }
11121 else
11122 {
11123 /* Emulated gather-scatter. */
11127 {
11128 /* For emulated gathers N offset vector element
11129 offset add is consumed by the load). */
11133 /* N scalar loads plus gathering them into a
11134 vector. */
11135 inside_cost
11138 inside_cost
11142 }
11143 unsigned HOST_WIDE_INT const_offset_nunits
11149 /* We support offset vectors with more elements
11150 than the data vector for now. */
11151 unsigned HOST_WIDE_INT factor
11154 unsigned elt_offset
11161 {
11164 tree idx
11169 scale);
11181 }
11186 }
11189 /* DATA_REF is null if we've already built the statement. */
11191 {
11194 }
11199 /* Store vector loads in the corresponding SLP_NODE. */
11202 }
11206 }
11213 "vect_model_load_cost: inside_cost = %u, "
11217 }
11221 /* For costing some adjacent vector loads, we'd like to cost with
11222 the total number of them once instead of cost each one by one. */
11225 {
11226 /* 1. Create the vector or array pointer update chain. */
11228 {
11229 bool simd_lane_access_p
11240 {
11243 }
11245 {
11246 dataref_ptr
11250 bump);
11251 /* Adjust the pointer by the difference to first_stmt. */
11252 data_reference_p ptrdr
11254 tree diff
11262 {
11266 alignment_support_scheme,
11269 }
11270 }
11271 else
11272 dataref_ptr
11274 at_loop,
11277 }
11279 {
11283 bump);
11284 else
11287 }
11294 {
11299 {
11313 }
11315 /* 2. Create the vector-load in the loop. */
11317 {
11320 {
11330 {
11331 align
11334 }
11335 else
11340 misalign);
11343 /* Compute IFN when LOOP_LENS or final_mask valid. */
11348 {
11349 opt_machine_mode new_ovmode
11352 unsigned factor
11357 }
11359 {
11364 }
11367 {
11369 {
11370 /* Pass VF value to 'len' argument of
11371 MASK_LEN_LOAD if LOOP_LENS is invalid. */
11373 }
11375 {
11376 /* Pass all ones value to 'mask' argument of
11377 MASK_LEN_LOAD if final_mask is invalid. */
11380 }
11381 }
11383 {
11384 signed char biasval
11388 }
11391 {
11398 bias);
11399 else
11407 /* Need conversion if it's wrapped with VnQI. */
11409 {
11412 tree var
11416 gsi);
11420 }
11421 }
11423 {
11427 final_mask);
11431 }
11432 else
11433 {
11437 unsigned int vect_align
11439 unsigned int scalar_dr_size
11441 /* If there's no peeling for gaps but we have a gap
11442 with slp loads then load the lower half of the
11443 vector only. See get_group_load_store_type for
11444 when we apply this optimization. */
11446 && loop_vinfo
11451 {
11452 tree half_vtype;
11453 new_vtype
11458 }
11459 tree offset
11464 {
11465 unsigned HOST_WIDE_INT gap_offset
11469 }
11470 data_ref
11473 ;
11474 else
11479 {
11485 gsi);
11490 {
11494 }
11495 else
11496 {
11500 }
11505 else
11506 {
11513 vec_dest,
11515 }
11516 }
11517 }
11519 }
11521 {
11531 dr_explicit_realign,
11536 else
11538 // For explicit realign the target alignment should be
11539 // known at compile time.
11540 unsigned HOST_WIDE_INT align
11547 data_ref
11561 bump);
11567 else
11571 data_ref
11574 }
11576 {
11581 else
11583 // We should only be doing this if we know the target
11584 // alignment at compile time.
11585 unsigned HOST_WIDE_INT align
11595 }
11598 }
11600 /* One common place to cost the above vect load for different
11601 alignment support schemes. */
11603 {
11604 /* For VMAT_CONTIGUOUS_PERMUTE if it's grouped load, we
11605 only need to take care of the first stmt, whose
11606 stmt_info is first_stmt_info, vec_num iterating on it
11607 will cover the cost for the remaining, it's consistent
11608 with transforming. For the prologue cost for realign,
11609 we only need to count it once for the whole group. */
11616 {
11617 /* Leave realign cases alone to keep them simple. */
11625 else
11626 n_adjacent_loads++;
11627 }
11628 }
11629 else
11630 {
11632 /* DATA_REF is null if we've already built the statement. */
11634 {
11637 }
11641 }
11643 /* 3. Handle explicit realignment if necessary/supported.
11644 Create in loop:
11645 vec_dest = realign_load (msq, lsq, realignment_token) */
11649 {
11661 {
11665 UNKNOWN_LOCATION);
11667 }
11668 }
11671 {
11675 else
11676 {
11681 }
11682 }
11684 /* Collect vector loads and later create their permutation in
11685 vect_transform_grouped_load (). */
11689 /* Store vector loads in the corresponding SLP_NODE. */
11693 /* With SLP permutation we load the gaps as well, without
11694 we need to skip the gaps after we manage to fully load
11695 all elements. group_gap_adj is DR_GROUP_SIZE here. */
11699 && !slp_perm
11701 {
11702 poly_wide_int bump_val
11711 }
11712 }
11713 /* Bump the vector pointer to account for a gap or for excess
11714 elements loaded for a permuted SLP load. */
11718 {
11719 poly_wide_int bump_val
11726 }
11732 {
11734 /* For SLP we know we've seen all possible uses of dr_chain so
11735 direct vect_transform_slp_perm_load to DCE the unused parts.
11736 ??? This is a hack to prevent compile-time issues as seen
11737 in PR101120 and friends. */
11739 {
11744 }
11745 else
11746 {
11751 }
11752 }
11753 else
11754 {
11756 {
11758 /* We assume that the cost of a single load-lanes instruction
11759 is equivalent to the cost of DR_GROUP_SIZE separate loads.
11760 If a grouped access is instead being provided by a
11761 load-and-permute operation, include the cost of the
11762 permutes. */
11764 {
11765 /* Uses an even and odd extract operations or shuffle
11766 operations for each needed permute. */
11774 "vect_model_load_cost:"
11776 group_size);
11777 }
11779 {
11783 }
11784 }
11787 }
11789 }
11794 {
11805 "vect_model_load_cost: inside_cost = %u, "
11808 }
11811 }
11813 /* Function vect_is_simple_cond.
11815 Input:
11816 LOOP - the loop that is being vectorized.
11817 COND - Condition that is checked for simple use.
11819 Output:
11820 *COMP_VECTYPE - the vector type for the comparison.
11821 *DTS - The def types for the arguments of the comparison
11823 Returns whether a COND can be vectorized. Checks whether
11824 condition operands are supportable using vec_is_simple_use. */
11826 static bool
11830 {
11833 slp_tree slp_op;
11835 /* Mask case. */
11838 {
11841 || !*comp_vectype
11845 }
11854 {
11858 }
11862 else
11866 {
11870 }
11874 else
11883 /* Invariant comparison. */
11885 {
11889 else
11890 {
11891 /* If we can widen the comparison to match vectype do so. */
11893 && !slp_node
11896 scalar_type = build_nonstandard_integer_type
11899 slp_node);
11900 }
11901 }
11904 }
11906 /* vectorizable_condition.
11908 Check if STMT_INFO is conditional modify expression that can be vectorized.
11909 If VEC_STMT is also passed, vectorize STMT_INFO: create a vectorized
11910 stmt using VEC_COND_EXPR to replace it, put it in VEC_STMT, and insert it
11911 at GSI.
11913 When STMT_INFO is vectorized as a nested cycle, for_reduction is true.
11915 Return true if STMT_INFO is vectorizable in this way. */
11917 static bool
11922 {
11930 tree vec_compare;
11931 tree new_temp;
11946 tree vec_cmp_type;
11952 /* Is vectorizable conditional operation? */
11964 bool for_reduction
11967 {
11975 }
11976 else
11977 {
11980 }
11986 {
11989 }
11990 else
11991 {
11994 }
12030 {
12033 }
12035 /* For conditional reductions, the "then" value needs to be the candidate
12036 value calculated by this iteration while the "else" value needs to be
12037 the result carried over from previous iterations. If the COND_EXPR
12038 is the other way around, we need to swap it. */
12041 {
12044 else
12045 {
12050 else
12051 {
12053 /* Make sure we don't accidentally use the old condition. */
12055 }
12056 }
12058 }
12061 {
12062 /* Boolean values may have another representation in vectors
12063 and therefore we prefer bit operations over comparison for
12064 them (which also works for scalar masks). We store opcodes
12065 to use in bitop1 and bitop2. Statement is vectorized as
12066 BITOP2 (rhs1 BITOP1 rhs2) or rhs1 BITOP2 (BITOP1 rhs2)
12067 depending on bitop1 and bitop2 arity. */
12069 {
12097 }
12099 }
12104 {
12109 }
12112 {
12114 {
12116 optab optab;
12123 {
12125 optab_default);
12128 }
12129 }
12133 /* Count one reduction-like operation per vector. */
12136 && (masked
12138 cond_code)
12140 ERROR_MARK))))
12144 && (!vect_maybe_update_slp_op_vectype
12147 && !vect_maybe_update_slp_op_vectype
12151 {
12156 }
12160 {
12162 {
12168 else
12172 }
12173 /* Extra inactive lanes should be safe for vect_nested_cycle. */
12175 {
12178 "conditional reduction prevents the use"
12181 }
12182 }
12188 }
12190 /* Transform. */
12192 /* Handle def. */
12199 /* See whether another part of the vectorized code applies a loop
12200 mask to the condition, or to its inverse. */
12205 {
12208 }
12210 {
12213 else
12214 {
12218 else
12219 {
12224 {
12228 }
12229 else
12230 {
12231 /* Try the inverse of the current mask. We check if the
12232 inverse mask is live and if so we generate a negate of
12233 the current mask such that we still honor NaNs. */
12237 {
12242 }
12243 }
12244 }
12245 }
12246 }
12248 /* Handle cond expr. */
12253 reduction_type != EXTRACT_LAST_REDUCTION
12255 else
12260 reduction_type != EXTRACT_LAST_REDUCTION
12263 /* Arguments are ready. Create the new vector stmt. */
12265 {
12275 else
12276 {
12279 {
12284 }
12285 else
12286 {
12291 vec_cond_rhs);
12292 else
12293 new_stmt
12295 vec_cond_rhs);
12301 {
12302 /* Instead of doing ~x ? y : z do x ? z : y. */
12305 }
12306 else
12307 {
12310 new_stmt
12312 else
12313 new_stmt
12318 }
12319 }
12320 }
12322 /* If we decided to apply a loop mask to the result of the vector
12323 comparison, AND the comparison with the mask now. Later passes
12324 should then be able to reuse the AND results between mulitple
12325 vector statements.
12327 For example:
12328 for (int i = 0; i < 100; ++i)
12329 x[i] = y[i] ? z[i] : 10;
12331 results in following optimized GIMPLE:
12333 mask__35.8_43 = vect__4.7_41 != { 0, ... };
12334 vec_mask_and_46 = loop_mask_40 & mask__35.8_43;
12335 _19 = &MEM[base: z_12(D), index: ivtmp_56, step: 4, offset: 0B];
12336 vect_iftmp.11_47 = .MASK_LOAD (_19, 4B, vec_mask_and_46);
12337 vect_iftmp.12_52 = VEC_COND_EXPR <vec_mask_and_46,
12338 vect_iftmp.11_47, { 10, ... }>;
12340 instead of using a masked and unmasked forms of
12341 vec != { 0, ... } (masked in the MASK_LOAD,
12342 unmasked in the VEC_COND_EXPR). */
12344 /* Force vec_compare to be an SSA_NAME rather than a comparison,
12345 in cases where that's necessary. */
12349 {
12351 {
12354 vec_compare);
12357 }
12360 {
12363 BIT_NOT_EXPR,
12364 vec_compare);
12367 }
12371 {
12373 {
12376 signed char biasval
12379 }
12380 else
12381 {
12384 }
12385 }
12387 {
12388 tree loop_mask
12392 gassign *g
12394 loop_mask);
12397 }
12398 }
12402 {
12406 new_stmt = gimple_build_call_internal
12409 else
12410 new_stmt = gimple_build_call_internal
12412 vec_then_clause);
12417 else
12418 {
12419 /* In this case we're moving the definition to later in the
12420 block. That doesn't matter because the only uses of the
12421 lhs are in phi statements. */
12425 }
12426 }
12427 else
12428 {
12433 }
12436 else
12438 }
12449 }
12451 /* Helper of vectorizable_comparison.
12453 Check if STMT_INFO is comparison expression CODE that can be vectorized.
12454 If VEC_STMT is also passed, vectorize STMT_INFO: create a vectorized
12455 comparison, put it in VEC_STMT, and insert it at GSI.
12457 Return true if STMT_INFO is vectorizable in this way. */
12459 static bool
12464 {
12468 tree new_temp;
12472 poly_uint64 nunits;
12479 tree mask_type;
12480 tree mask;
12493 else
12517 /* Invariant comparison. */
12519 {
12522 else
12524 slp_node);
12527 }
12531 /* Can't compare mask and non-mask types. */
12536 /* Boolean values may have another representation in vectors
12537 and therefore we prefer bit operations over comparison for
12538 them (which also works for scalar masks). We store opcodes
12539 to use in bitop1 and bitop2. Statement is vectorized as
12540 BITOP2 (rhs1 BITOP1 rhs2) or
12541 rhs1 BITOP2 (BITOP1 rhs2)
12542 depending on bitop1 and bitop2 arity. */
12545 {
12547 {
12550 }
12552 {
12555 }
12557 {
12561 }
12563 {
12567 }
12568 else
12569 {
12573 }
12574 }
12577 {
12579 {
12582 }
12583 else
12584 {
12586 optab optab;
12593 {
12597 }
12598 }
12600 /* Put types on constant and invariant SLP children. */
12604 {
12609 }
12615 }
12617 /* Transform. */
12619 /* Handle def. */
12629 /* Arguments are ready. Create the new vector stmt. */
12631 {
12637 {
12641 }
12642 else
12643 {
12646 else
12648 vec_rhs2);
12651 {
12655 else
12657 new_temp);
12659 }
12660 }
12663 else
12665 }
12674 }
12676 /* vectorizable_comparison.
12678 Check if STMT_INFO is comparison expression that can be vectorized.
12679 If VEC_STMT is also passed, vectorize STMT_INFO: create a vectorized
12680 comparison, put it in VEC_STMT, and insert it at GSI.
12682 Return true if STMT_INFO is vectorizable in this way. */
12684 static bool
12689 {
12712 }
12714 /* If SLP_NODE is nonnull, return true if vectorizable_live_operation
12715 can handle all live statements in the node. Otherwise return true
12716 if STMT_INFO is not live or if vectorizable_live_operation can handle it.
12717 VEC_STMT_P is as for vectorizable_live_operation. */
12719 static bool
12724 {
12726 {
12727 stmt_vec_info slp_stmt_info;
12730 {
12736 }
12737 }
12745 }
12747 /* Make sure the statement is vectorizable. */
12749 opt_result
12754 {
12758 gimple_seq pattern_def_seq;
12766 "not vectorized:"
12773 {
12774 gimple_stmt_iterator si;
12777 {
12778 stmt_vec_info pattern_def_stmt_info
12782 {
12783 /* Analyze def stmt of STMT if it's a pattern stmt. */
12789 opt_result res
12792 cost_vec);
12795 }
12796 }
12797 }
12799 /* Skip stmts that do not need to be vectorized. In loops this is expected
12800 to include:
12801 - the COND_EXPR which is the loop exit condition
12802 - any LABEL_EXPRs in the loop
12803 - computations that are used only for array indexing or loop control.
12804 In basic blocks we only analyze statements that are a part of some SLP
12805 instance, therefore, all the statements are relevant.
12807 Pattern statement needs to be analyzed instead of the original statement
12808 if the original statement is not relevant. Otherwise, we analyze both
12809 statements. In basic blocks we are called from some SLP instance
12810 traversal, don't analyze pattern stmts instead, the pattern stmts
12811 already will be part of SLP instance. */
12816 {
12818 && pattern_stmt_info
12821 {
12822 /* Analyze PATTERN_STMT instead of the original stmt. */
12828 }
12829 else
12830 {
12835 }
12836 }
12839 && pattern_stmt_info
12842 {
12843 /* Analyze PATTERN_STMT too. */
12849 opt_result res
12854 }
12857 {
12881 }
12888 {
12893 }
12896 {
12901 }
12907 /* Prefer vectorizable_call over vectorizable_simd_clone_call so
12908 -mveclibabi= takes preference over library functions with
12909 the simd attribute. */
12912 cost_vec)
12929 cost_vec)
12934 else
12935 {
12941 cost_vec)
12947 cost_vec)
12955 cost_vec)
12957 }
12964 "not vectorized:"
12968 /* Stmts that are (also) "live" (i.e. - that are used out of the loop)
12969 need extra handling, except for vectorizable reductions. */
12977 "not vectorized:"
12982 }
12985 /* Function vect_transform_stmt.
12987 Create a vectorized stmt to replace STMT_INFO, and insert it at GSI. */
12989 bool
12993 {
13005 {
13017 NULL);
13029 NULL);
13041 NULL);
13047 && !slp_node
13050 /* In case of interleaving, the whole chain is vectorized when the
13051 last store in the chain is reached. Store stmts before the last
13052 one are skipped, and there vec_stmt_info shouldn't be freed
13053 meanwhile. */
13054 ;
13055 else
13056 {
13061 }
13117 {
13122 }
13124 }
13130 {
13131 /* Handle stmts whose DEF is used outside the loop-nest that is
13132 being vectorized. */
13136 }
13142 }
13145 /* Remove a group of stores (for SLP or interleaving), free their
13146 stmt_vec_info. */
13148 void
13150 {
13154 {
13157 /* Free the attached stmt_vec_info and remove the stmt. */
13160 }
13161 }
13163 /* If NUNITS is nonzero, return a vector type that contains NUNITS
13164 elements of type SCALAR_TYPE, or null if the target doesn't support
13165 such a type.
13167 If NUNITS is zero, return a vector type that contains elements of
13168 type SCALAR_TYPE, choosing whichever vector size the target prefers.
13170 If PREVAILING_MODE is VOIDmode, we have not yet chosen a vector mode
13171 for this vectorization region and want to "autodetect" the best choice.
13172 Otherwise, PREVAILING_MODE is a previously-chosen vector TYPE_MODE
13173 and we want the new type to be interoperable with it. PREVAILING_MODE
13174 in this case can be a scalar integer mode or a vector mode; when it
13175 is a vector mode, the function acts like a tree-level version of
13176 related_vector_mode. */
13178 tree
13181 {
13183 scalar_mode inner_mode;
13184 machine_mode simd_mode;
13185 tree vectype;
13196 /* Interoperability between modes requires one to be a constant multiple
13197 of the other, so that the number of vectors required for each operation
13198 is a compile-time constant. */
13206 /* For vector types of elements whose mode precision doesn't
13207 match their types precision we use a element type of mode
13208 precision. The vectorization routines will have to make sure
13209 they support the proper result truncation/extension.
13210 We also make sure to build vector types with INTEGER_TYPE
13211 component type only. */
13218 /* We shouldn't end up building VECTOR_TYPEs of non-scalar components.
13219 When the component mode passes the above test simply use a type
13220 corresponding to that mode. The theory is that any use that
13221 would cause problems with this will disable vectorization anyway. */
13226 /* We can't build a vector type of elements with alignment bigger than
13227 their size. */
13232 /* If we felt back to using the mode fail if there was
13233 no scalar type for it. */
13237 /* If no prevailing mode was supplied, use the mode the target prefers.
13238 Otherwise lookup a vector mode based on the prevailing mode. */
13240 {
13244 {
13245 /* Traditional behavior is not to take the integer mode
13246 literally, but simply to use it as a way of determining
13247 the vector size. It is up to mode_for_vector to decide
13248 what the TYPE_MODE should be.
13250 Note that nunits == 1 is allowed in order to support single
13251 element vector types. */
13255 }
13256 }
13260 {
13261 /* Fall back to using mode_for_vector, mostly in the hope of being
13262 able to use an integer mode. */
13269 }
13273 /* In cases where the mode was chosen by mode_for_vector, check that
13274 the target actually supports the chosen mode, or that it at least
13275 allows the vector mode to be replaced by a like-sized integer. */
13280 /* Re-attach the address-space qualifier if we canonicalized the scalar
13281 type. */
13283 return build_qualified_type
13287 }
13289 /* Function get_vectype_for_scalar_type.
13291 Returns the vector type corresponding to SCALAR_TYPE as supported
13292 by the target. If GROUP_SIZE is nonzero and we're performing BB
13293 vectorization, make sure that the number of elements in the vector
13294 is no bigger than GROUP_SIZE. */
13296 tree
13299 {
13300 /* For BB vectorization, we should always have a group size once we've
13301 constructed the SLP tree; the only valid uses of zero GROUP_SIZEs
13302 are tentative requests during things like early data reference
13303 analysis and pattern recognition. */
13306 else
13310 scalar_type);
13314 /* Register the natural choice of vector type, before the group size
13315 has been applied. */
13319 /* If the natural choice of vector type doesn't satisfy GROUP_SIZE,
13320 try again with an explicit number of elements. */
13322 && group_size
13324 {
13325 /* Start with the biggest number of units that fits within
13326 GROUP_SIZE and halve it until we find a valid vector type.
13327 Usually either the first attempt will succeed or all will
13328 fail (in the latter case because GROUP_SIZE is too small
13329 for the target), but it's possible that a target could have
13330 a hole between supported vector types.
13332 If GROUP_SIZE is not a power of 2, this has the effect of
13333 trying the largest power of 2 that fits within the group,
13334 even though the group is not a multiple of that vector size.
13335 The BB vectorizer will then try to carve up the group into
13336 smaller pieces. */
13338 do
13339 {
13343 }
13345 }
13348 }
13350 /* Return the vector type corresponding to SCALAR_TYPE as supported
13351 by the target. NODE, if nonnull, is the SLP tree node that will
13352 use the returned vector type. */
13354 tree
13356 {
13361 }
13363 /* Function get_mask_type_for_scalar_type.
13365 Returns the mask type corresponding to a result of comparison
13366 of vectors of specified SCALAR_TYPE as supported by target.
13367 If GROUP_SIZE is nonzero and we're performing BB vectorization,
13368 make sure that the number of elements in the vector is no bigger
13369 than GROUP_SIZE. */
13371 tree
13374 {
13381 }
13383 /* Function get_mask_type_for_scalar_type.
13385 Returns the mask type corresponding to a result of comparison
13386 of vectors of specified SCALAR_TYPE as supported by target.
13387 NODE, if nonnull, is the SLP tree node that will use the returned
13388 vector type. */
13390 tree
13392 slp_tree node)
13393 {
13400 }
13402 /* Function get_same_sized_vectype
13404 Returns a vector type corresponding to SCALAR_TYPE of size
13405 VECTOR_TYPE if supported by the target. */
13407 tree
13409 {
13413 poly_uint64 nunits;
13420 }
13422 /* Return true if replacing LOOP_VINFO->vector_mode with VECTOR_MODE
13423 would not change the chosen vector modes. */
13425 bool
13427 {
13434 }
13436 /* Function vect_is_simple_use.
13438 Input:
13439 VINFO - the vect info of the loop or basic block that is being vectorized.
13440 OPERAND - operand in the loop or bb.
13441 Output:
13442 DEF_STMT_INFO_OUT (optional) - information about the defining stmt in
13443 case OPERAND is an SSA_NAME that is defined in the vectorizable region
13444 DEF_STMT_OUT (optional) - the defining stmt in case OPERAND is an SSA_NAME;
13445 the definition could be anywhere in the function
13446 DT - the type of definition
13448 Returns whether a stmt with OPERAND can be vectorized.
13449 For loops, supportable operands are constants, loop invariants, and operands
13450 that are defined by the current iteration of the loop. Unsupportable
13451 operands are those that are defined by a previous iteration of the loop (as
13452 is the case in reduction/induction computations).
13453 For basic blocks, supportable operands are constants and bb invariants.
13454 For now, operands defined outside the basic block are not supported. */
13456 bool
13459 {
13467 {
13473 else
13475 }
13485 else
13486 {
13491 else
13492 {
13498 }
13501 }
13504 {
13507 {
13538 }
13539 }
13542 {
13547 }
13550 }
13552 /* Function vect_is_simple_use.
13554 Same as vect_is_simple_use but also determines the vector operand
13555 type of OPERAND and stores it to *VECTYPE. If the definition of
13556 OPERAND is vect_uninitialized_def, vect_constant_def or
13557 vect_external_def *VECTYPE will be set to NULL_TREE and the caller
13558 is responsible to compute the best suited vector type for the
13559 scalar operand. */
13561 bool
13565 {
13566 stmt_vec_info def_stmt_info;
13576 /* Now get a vector type if the def is internal, otherwise supply
13577 NULL_TREE and leave it up to the caller to figure out a proper
13578 type for the use stmt. */
13585 {
13591 }
13596 else
13600 }
13602 /* Function vect_is_simple_use.
13604 Same as vect_is_simple_use but determines the operand by operand
13605 position OPERAND from either STMT or SLP_NODE, filling in *OP
13606 and *SLP_DEF (when SLP_NODE is not NULL). */
13608 bool
13613 {
13615 {
13620 {
13623 }
13624 else
13625 {
13631 }
13632 }
13633 else
13634 {
13637 {
13640 {
13643 else
13645 }
13648 else
13650 }
13653 else
13656 }
13657 }
13659 /* If OP is not NULL and is external or constant update its vector
13660 type with VECTYPE. Returns true if successful or false if not,
13661 for example when conflicting vector types are present. */
13663 bool
13665 {
13670 /* For external defs refuse to produce VECTOR_BOOLEAN_TYPE_P, those
13671 should be handled by patters. Allow vect_constant_def for now. */
13677 }
13679 /* Function supportable_widening_operation
13681 Check whether an operation represented by the code CODE is a
13682 widening operation that is supported by the target platform in
13683 vector form (i.e., when operating on arguments of type VECTYPE_IN
13684 producing a result of type VECTYPE_OUT).
13686 Widening operations we currently support are NOP (CONVERT), FLOAT,
13687 FIX_TRUNC and WIDEN_MULT. This function checks if these operations
13688 are supported by the target platform either directly (via vector
13689 tree-codes), or via target builtins.
13691 Output:
13692 - CODE1 and CODE2 are codes of vector operations to be used when
13693 vectorizing the operation, if available.
13694 - MULTI_STEP_CVT determines the number of required intermediate steps in
13695 case of multi-step conversion (like char->short->int - in that case
13696 MULTI_STEP_CVT will be 1).
13697 - INTERM_TYPES contains the intermediate type required to perform the
13698 widening operation (short in the above example). */
13700 bool
13702 code_helper code,
13703 stmt_vec_info stmt_info,
13709 {
13712 machine_mode vec_mode;
13728 {
13730 /* Don't set c1 and c2 if code is not a tree_code. */
13734 /* The result of a vectorized widening operation usually requires
13735 two vectors (because the widened results do not fit into one vector).
13736 The generated vector results would normally be expected to be
13737 generated in the same order as in the original scalar computation,
13738 i.e. if 8 results are generated in each vector iteration, they are
13739 to be organized as follows:
13740 vect1: [res1,res2,res3,res4],
13741 vect2: [res5,res6,res7,res8].
13743 However, in the special case that the result of the widening
13744 operation is used in a reduction computation only, the order doesn't
13745 matter (because when vectorizing a reduction we change the order of
13746 the computation). Some targets can take advantage of this and
13747 generate more efficient code. For example, targets like Altivec,
13748 that support widen_mult using a sequence of {mult_even,mult_odd}
13749 generate the following vectors:
13750 vect1: [res1,res3,res5,res7],
13751 vect2: [res2,res4,res6,res8].
13753 When vectorizing outer-loops, we execute the inner-loop sequentially
13754 (each vectorized inner-loop iteration contributes to VF outer-loop
13755 iterations in parallel). We therefore don't allow to change the
13756 order of the computation in the inner-loop during outer-loop
13757 vectorization. */
13758 /* TODO: Another case in which order doesn't *really* matter is when we
13759 widen and then contract again, e.g. (short)((int)x * y >> 8).
13760 Normally, pack_trunc performs an even/odd permute, whereas the
13761 repack from an even/odd expansion would be an interleave, which
13762 would be significantly simpler for e.g. AVX2. */
13763 /* In any case, in order to avoid duplicating the code below, recurse
13764 on VEC_WIDEN_MULT_EVEN_EXPR. If it succeeds, all the return values
13765 are properly set up for the caller. If we fail, we'll continue with
13766 a VEC_WIDEN_MULT_LO/HI_EXPR check. */
13774 interm_types))
13775 {
13776 /* Elements in a vector with vect_used_by_reduction property cannot
13777 be reordered if the use chain with this property does not have the
13778 same operation. One such an example is s += a * b, where elements
13779 in a and b cannot be reordered. Here we check if the vector defined
13780 by STMT is only directly used in the reduction statement. */
13786 }
13802 /* Support the recursion induced just above. */
13812 CASE_CONVERT:
13829 }
13835 {
13836 /* The signedness is determined from output operand. */
13839 }
13845 {
13846 /* If the input and result modes are the same, a different optab
13847 is needed where we pass in the number of units in vectype. */
13850 }
13854 {
13855 /* If this is an internal fn then we must check whether the target
13856 supports either a low-high split or an even-odd split. */
13866 /* If we don't support low-high, then check for even-odd. */
13869 || !optab2
13871 {
13877 }
13878 }
13880 {
13882 {
13883 /* The signedness is determined from output operand. */
13886 }
13892 {
13893 /* If the input and result modes are the same, a different optab
13894 is needed where we pass in the number of units in vectype. */
13897 }
13898 else
13899 {
13902 }
13905 }
13917 {
13920 /* For scalar masks we may have different boolean
13921 vector types having the same QImode. Thus we
13922 add additional check for elements number. */
13926 }
13928 /* Check if it's a multi-step conversion that can be done using intermediate
13929 types. */
13937 /* We assume here that there will not be more than MAX_INTERM_CVT_STEPS
13938 intermediate steps in promotion sequence. We try
13939 MAX_INTERM_CVT_STEPS to get to NARROW_VECTYPE, and fail if we do
13940 not. */
13943 {
13946 intermediate_type
13949 {
13950 tree intermediate_element_type
13953 intermediate_type
13955 intermediate_mode);
13956 }
13957 else
13958 intermediate_type
13966 {
13967 /* If the input and result modes are the same, a different optab
13968 is needed where we pass in the number of units in vectype. */
13971 }
13972 else
13973 {
13976 }
13994 {
14000 }
14004 }
14008 }
14011 /* Function supportable_narrowing_operation
14013 Check whether an operation represented by the code CODE is a
14014 narrowing operation that is supported by the target platform in
14015 vector form (i.e., when operating on arguments of type VECTYPE_IN
14016 and producing a result of type VECTYPE_OUT).
14018 Narrowing operations we currently support are NOP (CONVERT), FIX_TRUNC
14019 and FLOAT. This function checks if these operations are supported by
14020 the target platform directly via vector tree-codes.
14022 Output:
14023 - CODE1 is the code of a vector operation to be used when
14024 vectorizing the operation, if available.
14025 - MULTI_STEP_CVT determines the number of required intermediate steps in
14026 case of multi-step conversion (like int->short->char - in that case
14027 MULTI_STEP_CVT will be 1).
14028 - INTERM_TYPES contains the intermediate type required to perform the
14029 narrowing operation (short in the above example). */
14031 bool
14036 {
14037 machine_mode vec_mode;
14054 {
14055 CASE_CONVERT:
14063 else
14069 /* The signedness is determined from output operand. */
14080 }
14092 {
14095 /* For scalar masks we may have different boolean
14096 vector types having the same QImode. Thus we
14097 add additional check for elements number. */
14101 }
14106 /* Check if it's a multi-step conversion that can be done using intermediate
14107 types. */
14112 else
14115 /* For multi-step FIX_TRUNC_EXPR prefer signed floating to integer
14116 conversion over unsigned, as unsigned FIX_TRUNC_EXPR is often more
14117 costly than signed. */
14119 {
14122 intermediate_type
14124 interm_optab
14130 {
14134 }
14135 }
14137 /* We assume here that there will not be more than MAX_INTERM_CVT_STEPS
14138 intermediate steps in promotion sequence. We try
14139 MAX_INTERM_CVT_STEPS to get to NARROW_VECTYPE, and fail if we do not. */
14142 {
14145 intermediate_type
14147 else
14148 intermediate_type
14156 else
14157 interm_optab
14159 optab_default);
14171 {
14177 }
14182 }
14186 }
14188 /* Generate and return a vector mask of MASK_TYPE such that
14189 mask[I] is true iff J + START_INDEX < END_INDEX for all J <= I.
14190 Add the statements to SEQ. */
14192 tree
14195 {
14199 OPTIMIZE_FOR_SPEED));
14203 tree tmp;
14206 else
14211 }
14213 /* Generate a vector mask of type MASK_TYPE for which index I is false iff
14214 J + START_INDEX < END_INDEX for all J <= I. Add the statements to SEQ. */
14216 tree
14218 tree end_index)
14219 {
14222 }
14224 /* Try to compute the vector types required to vectorize STMT_INFO,
14225 returning true on success and false if vectorization isn't possible.
14226 If GROUP_SIZE is nonzero and we're performing BB vectorization,
14227 take sure that the number of elements in the vectors is no bigger
14228 than GROUP_SIZE.
14230 On success:
14232 - Set *STMT_VECTYPE_OUT to:
14233 - NULL_TREE if the statement doesn't need to be vectorized;
14234 - the equivalent of STMT_VINFO_VECTYPE otherwise.
14236 - Set *NUNITS_VECTYPE_OUT to the vector type that contains the maximum
14237 number of units needed to vectorize STMT_INFO, or NULL_TREE if the
14238 statement does not help to determine the overall number of units. */
14240 opt_result
14245 {
14248 /* For BB vectorization, we should always have a group size once we've
14249 constructed the SLP tree; the only valid uses of zero GROUP_SIZEs
14250 are tentative requests during things like early data reference
14251 analysis and pattern recognition. */
14254 else
14261 /* MASK_STORE has no lhs, but is ok. */
14263 {
14265 {
14266 /* Ignore calls with no lhs. These must be calls to
14267 #pragma omp simd functions, and what vectorization factor
14268 it really needs can't be determined until
14269 vectorizable_simd_clone_call. */
14274 }
14278 }
14280 tree vectype;
14283 {
14288 }
14290 {
14299 }
14300 else
14301 {
14306 else
14310 {
14313 "get vectype for scalar type (group size %d):"
14315 else
14318 }
14322 "not vectorized:"
14324 scalar_type);
14328 }
14333 stmt);
14337 /* Don't try to compute scalar types if the stmt produces a boolean
14338 vector; use the existing vector type instead. */
14341 {
14342 /* The number of units is set according to the smallest scalar
14343 type (or the largest vector size, but we only support one
14344 vector size per vectorization). */
14348 {
14352 scalar_type);
14354 group_size);
14358 scalar_type);
14361 nunits_vectype);
14362 }
14363 }
14368 "Not vectorized: Incompatible number "
14373 {
14377 }
14381 }
14383 /* Generate and return statement sequence that sets vector length LEN that is:
14385 min_of_start_and_end = min (START_INDEX, END_INDEX);
14386 left_len = END_INDEX - min_of_start_and_end;
14387 rhs = min (left_len, LEN_LIMIT);
14388 LEN = rhs;
14390 Note: the cost of the code generated by this function is modeled
14391 by vect_estimate_min_profitable_iters, so changes here may need
14392 corresponding changes there. */
14394 gimple_seq
14396 {
14408 }