| blob | 4b88bdf7f4389aece69ac311790e509e2d04d816 |
1 /* Transformation Utilities for Loop Vectorization.
2 Copyright (C) 2003, 2004, 2005, 2006, 2007 Free Software Foundation, Inc.
3 Contributed by Dorit Naishlos <dorit@il.ibm.com>
5 This file is part of GCC.
7 GCC is free software; you can redistribute it and/or modify it under
8 the terms of the GNU General Public License as published by the Free
9 Software Foundation; either version 3, or (at your option) any later
10 version.
12 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
13 WARRANTY; without even the implied warranty of MERCHANTABILITY or
14 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
15 for more details.
17 You should have received a copy of the GNU General Public License
18 along with GCC; see the file COPYING3. If not see
19 <http://www.gnu.org/licenses/>. */
48 /* Utility functions for the code transformation. */
51 static tree vect_create_data_ref_ptr
53 static tree vect_create_addr_base_for_vector_ref
58 static void vect_finish_stmt_generation
64 /* Utility function dealing with loop peeling (not peeling itself). */
65 static void vect_generate_tmps_on_preheader
75 static int
77 {
81 {
99 }
100 }
103 /* Function vect_estimate_min_profitable_iters
105 Return the number of iterations required for the vector version of the
106 loop to be profitable relative to the cost of the scalar version of the
107 loop.
109 TODO: Take profile info into account before making vectorization
110 decisions, if available. */
112 int
114 {
130 slp_instance instance;
132 /* Cost model disabled. */
134 {
138 }
140 /* Requires loop versioning tests to handle misalignment. */
143 {
144 /* FIXME: Make cost depend on complexity of individual check. */
145 vec_outside_cost +=
150 }
153 {
154 /* FIXME: Make cost depend on complexity of individual check. */
155 vec_outside_cost +=
160 }
164 {
166 }
168 /* Count statements in scalar loop. Using this as scalar cost for a single
169 iteration for now.
171 TODO: Add outer loop support.
173 TODO: Consider assigning different costs to different scalar
174 statements. */
176 /* FORNOW. */
181 {
182 block_stmt_iterator si;
187 else
191 {
199 /* FIXME: for stmts in the inner-loop in outer-loop vectorization,
200 some of the "outside" costs are generated inside the outer-loop. */
202 }
203 }
205 /* Add additional cost for the peeled instructions in prologue and epilogue
206 loop.
208 FORNOW: If we dont know the value of peel_iters for prologue or epilogue
209 at compile-time - we assume it's vf/2 (the worst would be vf-1).
211 TODO: Build an expression that represents peel_iters for prologue and
212 epilogue to be used in a run-time test. */
217 {
223 /* If peeling for alignment is unknown, loop bound of main loop becomes
224 unknown. */
228 "epilogue peel iters set to vf/2 because "
231 /* If peeled iterations are unknown, count a taken branch and a not taken
232 branch per peeled loop. Even if scalar loop iterations are known,
233 vector iterations are not known since peeled prologue iterations are
234 not known. Hence guards remain the same. */
238 }
239 else
240 {
242 {
249 }
250 else
254 {
258 "epilogue peel iters set to vf/2 because "
261 /* If peeled iterations are known but number of scalar loop
262 iterations are unknown, count a taken branch per peeled loop. */
265 }
266 else
267 {
272 }
273 }
279 /* Allow targets add additional (outside-of-loop) costs. FORNOW, the only
280 information we provide for the target is whether testing against the
281 threshold involves a runtime test. */
283 {
286 /* If the number of iterations is unknown, or the
287 peeling-for-misalignment amount is unknown, we eill have to generate
288 a runtime test to test the loop count against the threshold. */
292 vec_outside_cost +=
297 }
299 /* Add SLP costs. */
302 {
305 }
307 /* Calculate number of iterations required to make the vector version
308 profitable, relative to the loop bodies only. The following condition
309 must hold true: ((SIC*VF)-VIC)*niters > VOC*VF, where
310 SIC = scalar iteration cost, VIC = vector iteration cost,
311 VOC = vector outside cost and VF = vectorization factor. */
314 {
317 else
318 {
328 min_profitable_iters++;
329 }
330 }
331 /* vector version will never be profitable. */
332 else
333 {
336 "is divisible by scalar iteration cost = %d by a factor "
340 }
343 {
346 vec_inside_cost);
348 vec_outside_cost);
351 peel_iters_prologue);
353 peel_iters_epilogue);
355 min_profitable_iters);
356 }
358 min_profitable_iters =
361 /* Because the condition we create is:
362 if (niters <= min_profitable_iters)
363 then skip the vectorized loop. */
364 min_profitable_iters--;
368 min_profitable_iters);
371 }
374 /* TODO: Close dependency between vect_model_*_cost and vectorizable_*
375 functions. Design better to avoid maintenance issues. */
377 /* Function vect_model_reduction_cost.
379 Models cost for a reduction operation, including the vector ops
380 generated within the strip-mine loop, the initial definition before
381 the loop, and the epilogue code that must be generated. */
383 static bool
386 {
389 optab optab;
390 tree vectype;
391 tree orig_stmt;
392 tree reduction_op;
399 /* Cost of reduction op inside loop. */
405 {
407 {
410 }
412 }
422 /* Add in cost for initial definition. */
425 /* Determine cost of epilogue code.
427 We have a reduction operator that will reduce the vector in one statement.
428 Also requires scalar extract. */
431 {
434 else
435 {
437 tree bitsize =
444 /* We have a whole vector shift available. */
448 /* Final reduction via vector shifts and the reduction operator. Also
449 requires scalar extract. */
452 else
453 /* Use extracts and reduction op for final reduction. For N elements,
454 we have N extracts and N-1 reduction ops. */
456 }
457 }
467 }
470 /* Function vect_model_induction_cost.
472 Models cost for induction operations. */
474 static void
476 {
477 /* loop cost for vec_loop. */
479 /* prologue cost for vec_init and vec_step. */
486 }
489 /* Function vect_model_simple_cost.
491 Models cost for simple operations, i.e. those that only emit ncopies of a
492 single op. Right now, this does not account for multiple insns that could
493 be generated for the single vector op. We will handle that shortly. */
495 void
498 {
504 /* FORNOW: Assuming maximum 2 args per stmts. */
506 {
509 }
515 /* Set the costs either in STMT_INFO or SLP_NODE (if exists). */
518 }
521 /* Function vect_cost_strided_group_size
523 For strided load or store, return the group_size only if it is the first
524 load or store of a group, else return 1. This ensures that group size is
525 only returned once per group. */
527 static int
529 {
536 }
539 /* Function vect_model_store_cost
541 Models cost for stores. In the case of strided accesses, one access
542 has the overhead of the strided access attributed to it. */
544 void
547 {
554 /* Strided access? */
557 /* Not a strided access. */
558 else
561 /* Is this an access in a group of stores, which provide strided access?
562 If so, add in the cost of the permutes. */
564 {
565 /* Uses a high and low interleave operation for each needed permute. */
571 group_size);
573 }
575 /* Costs of the stores. */
582 /* Set the costs either in STMT_INFO or SLP_NODE (if exists). */
585 }
588 /* Function vect_model_load_cost
590 Models cost for loads. In the case of strided accesses, the last access
591 has the overhead of the strided access attributed to it. Since unaligned
592 accesses are supported for loads, we also account for the costs of the
593 access scheme chosen. */
595 void
598 {
601 tree first_stmt;
605 /* Strided accesses? */
608 {
611 }
612 /* Not a strided access. */
613 else
614 {
617 }
621 /* Is this an access in a group of loads providing strided access?
622 If so, add in the cost of the permutes. */
624 {
625 /* Uses an even and odd extract operations for each needed permute. */
631 group_size);
633 }
635 /* The loads themselves. */
637 {
639 {
646 }
648 {
649 /* Here, we assign an additional cost for the unaligned load. */
657 }
659 {
662 /* FIXME: If the misalignment remains fixed across the iterations of
663 the containing loop, the following cost should be added to the
664 outside costs. */
669 }
671 {
676 /* Unaligned software pipeline has a load of an address, an initial
677 load, and possibly a mask operation to "prime" the loop. However,
678 if this is an access in a group of loads, which provide strided
679 access, then the above cost should only be considered for one
680 access in the group. Inside the loop, there is a load op
681 and a realignment op. */
684 {
688 }
693 }
697 }
703 /* Set the costs either in STMT_INFO or SLP_NODE (if exists). */
706 }
709 /* Function vect_get_new_vect_var.
711 Returns a name for a new variable. The current naming scheme appends the
712 prefix "vect_" or "vect_p" (depending on the value of VAR_KIND) to
713 the name of vectorizer generated variables, and appends that to NAME if
714 provided. */
716 static tree
718 {
720 tree new_vect_var;
723 {
735 }
738 {
742 }
743 else
746 /* Mark vector typed variable as a gimple register variable. */
751 }
754 /* Function vect_create_addr_base_for_vector_ref.
756 Create an expression that computes the address of the first memory location
757 that will be accessed for a data reference.
759 Input:
760 STMT: The statement containing the data reference.
761 NEW_STMT_LIST: Must be initialized to NULL_TREE or a statement list.
762 OFFSET: Optional. If supplied, it is be added to the initial address.
763 LOOP: Specify relative to which loop-nest should the address be computed.
764 For example, when the dataref is in an inner-loop nested in an
765 outer-loop that is now being vectorized, LOOP can be either the
766 outer-loop, or the inner-loop. The first memory location accessed
767 by the following dataref ('in' points to short):
769 for (i=0; i<N; i++)
770 for (j=0; j<M; j++)
771 s += in[i+j]
773 is as follows:
774 if LOOP=i_loop: &in (relative to i_loop)
775 if LOOP=j_loop: &in+i*2B (relative to j_loop)
777 Output:
778 1. Return an SSA_NAME whose value is the address of the memory location of
779 the first vector of the data reference.
780 2. If new_stmt_list is not NULL_TREE after return then the caller must insert
781 these statement(s) which define the returned SSA_NAME.
783 FORNOW: We are only handling array accesses with step 1. */
785 static tree
788 tree offset,
790 {
795 tree base_name;
796 tree data_ref_base_var;
797 tree new_base_stmt;
798 tree vec_stmt;
808 {
817 }
819 /* Create data_ref_base */
827 /* Create base_offset */
836 {
845 }
847 /* base + base_offset */
853 /* addr_expr = addr_base */
865 {
868 }
870 }
873 /* Function vect_create_data_ref_ptr.
875 Create a new pointer to vector type (vp), that points to the first location
876 accessed in the loop by STMT, along with the def-use update chain to
877 appropriately advance the pointer through the loop iterations. Also set
878 aliasing information for the pointer. This vector pointer is used by the
879 callers to this function to create a memory reference expression for vector
880 load/store access.
882 Input:
883 1. STMT: a stmt that references memory. Expected to be of the form
884 GIMPLE_MODIFY_STMT <name, data-ref> or
885 GIMPLE_MODIFY_STMT <data-ref, name>.
886 2. AT_LOOP: the loop where the vector memref is to be created.
887 3. OFFSET (optional): an offset to be added to the initial address accessed
888 by the data-ref in STMT.
889 4. ONLY_INIT: indicate if vp is to be updated in the loop, or remain
890 pointing to the initial address.
891 5. TYPE: if not NULL indicates the required type of the data-ref
893 Output:
894 1. Declare a new ptr to vector_type, and have it point to the base of the
895 data reference (initial addressed accessed by the data reference).
896 For example, for vector of type V8HI, the following code is generated:
898 v8hi *vp;
899 vp = (v8hi *)initial_address;
901 if OFFSET is not supplied:
902 initial_address = &a[init];
903 if OFFSET is supplied:
904 initial_address = &a[init + OFFSET];
906 Return the initial_address in INITIAL_ADDRESS.
908 2. If ONLY_INIT is true, just return the initial pointer. Otherwise, also
909 update the pointer in each iteration of the loop.
911 Return the increment stmt that updates the pointer in PTR_INCR.
913 3. Set INV_P to true if the access pattern of the data reference in the
914 vectorized loop is invariant. Set it to false otherwise.
916 4. Return the pointer. */
918 static tree
922 {
923 tree base_name;
930 tree vect_ptr_type;
931 tree vect_ptr;
932 tree tag;
933 tree new_temp;
934 tree vec_stmt;
936 edge pe;
937 basic_block new_bb;
938 tree vect_ptr_init;
940 tree vptr;
941 block_stmt_iterator incr_bsi;
944 tree incr;
945 tree step;
947 /* Check the step (evolution) of the load in LOOP, and record
948 whether it's invariant. */
951 else
956 else
959 /* Create an expression for the first address accessed by this load
960 in LOOP. */
964 {
977 }
979 /** (1) Create the new vector-pointer variable: **/
982 else
988 /** (2) Add aliasing information to the new vector-pointer:
989 (The points-to info (DR_PTR_INFO) may be defined later.) **/
994 /* If tag is a variable (and NOT_A_TAG) than a new symbol memory
995 tag must be created with tag added to its may alias list. */
998 else
1003 /** Note: If the dataref is in an inner-loop nested in LOOP, and we are
1004 vectorizing LOOP (i.e. outer-loop vectorization), we need to create two
1005 def-use update cycles for the pointer: One relative to the outer-loop
1006 (LOOP), which is what steps (3) and (4) below do. The other is relative
1007 to the inner-loop (which is the inner-most loop containing the dataref),
1008 and this is done be step (5) below.
1010 When vectorizing inner-most loops, the vectorized loop (LOOP) is also the
1011 inner-most loop, and so steps (3),(4) work the same, and step (5) is
1012 redundant. Steps (3),(4) create the following:
1014 vp0 = &base_addr;
1015 LOOP: vp1 = phi(vp0,vp2)
1016 ...
1017 ...
1018 vp2 = vp1 + step
1019 goto LOOP
1021 If there is an inner-loop nested in loop, then step (5) will also be
1022 applied, and an additional update in the inner-loop will be created:
1024 vp0 = &base_addr;
1025 LOOP: vp1 = phi(vp0,vp2)
1026 ...
1027 inner: vp3 = phi(vp1,vp4)
1028 vp4 = vp3 + inner_step
1029 if () goto inner
1030 ...
1031 vp2 = vp1 + step
1032 if () goto LOOP */
1034 /** (3) Calculate the initial address the vector-pointer, and set
1035 the vector-pointer to point to it before the loop: **/
1037 /* Create: (&(base[init_val+offset]) in the loop preheader. */
1046 /* Create: p = (vectype *) initial_base */
1055 /** (4) Handle the updating of the vector-pointer inside the loop.
1056 This is needed when ONLY_INIT is false, and also when AT_LOOP
1057 is the inner-loop nested in LOOP (during outer-loop vectorization).
1058 **/
1061 {
1062 /* Copy the points-to information if it exists. */
1066 }
1067 else
1068 {
1069 /* The step of the vector pointer is the Vector Size. */
1071 /* One exception to the above is when the scalar step of the load in
1072 LOOP is zero. In this case the step here is also zero. */
1086 /* Copy the points-to information if it exists. */
1088 {
1091 }
1098 }
1104 /** (5) Handle the updating of the vector-pointer inside the inner-loop
1105 nested in LOOP, if exists: **/
1109 {
1118 /* Copy the points-to information if it exists. */
1120 {
1123 }
1130 }
1131 else
1133 }
1136 /* Function bump_vector_ptr
1138 Increment a pointer (to a vector type) by vector-size. If requested,
1139 i.e. if PTR-INCR is given, then also connect the new increment stmt
1140 to the existing def-use update-chain of the pointer, by modifying
1141 the PTR_INCR as illustrated below:
1143 The pointer def-use update-chain before this function:
1144 DATAREF_PTR = phi (p_0, p_2)
1145 ....
1146 PTR_INCR: p_2 = DATAREF_PTR + step
1148 The pointer def-use update-chain after this function:
1149 DATAREF_PTR = phi (p_0, p_2)
1150 ....
1151 NEW_DATAREF_PTR = DATAREF_PTR + BUMP
1152 ....
1153 PTR_INCR: p_2 = NEW_DATAREF_PTR + step
1155 Input:
1156 DATAREF_PTR - ssa_name of a pointer (to vector type) that is being updated
1157 in the loop.
1158 PTR_INCR - optional. The stmt that updates the pointer in each iteration of
1159 the loop. The increment amount across iterations is expected
1160 to be vector_size.
1161 BSI - location where the new update stmt is to be placed.
1162 STMT - the original scalar memory-access stmt that is being vectorized.
1163 BUMP - optional. The offset by which to bump the pointer. If not given,
1164 the offset is assumed to be vector_size.
1166 Output: Return NEW_DATAREF_PTR as illustrated above.
1168 */
1170 static tree
1173 {
1180 tree incr_stmt;
1181 ssa_op_iter iter;
1182 use_operand_p use_p;
1183 tree new_dataref_ptr;
1195 /* Copy the points-to information if it exists. */
1203 /* Update the vector-pointer's cross-iteration increment. */
1205 {
1210 else
1212 }
1215 }
1218 /* Function vect_create_destination_var.
1220 Create a new temporary of type VECTYPE. */
1222 static tree
1224 {
1225 tree vec_dest;
1227 tree type;
1242 }
1245 /* Function vect_init_vector.
1247 Insert a new stmt (INIT_STMT) that initializes a new vector variable with
1248 the vector elements of VECTOR_VAR. Place the initialization at BSI if it
1249 is not NULL. Otherwise, place the initialization at the loop preheader.
1250 Return the DEF of INIT_STMT.
1251 It will be used in the vectorization of STMT. */
1253 static tree
1256 {
1258 tree new_var;
1259 tree init_stmt;
1260 tree vec_oprnd;
1261 edge pe;
1262 tree new_temp;
1263 basic_block new_bb;
1273 else
1274 {
1283 }
1286 {
1289 }
1293 }
1296 /* For constant and loop invariant defs of SLP_NODE this function returns
1297 (vector) defs (VEC_OPRNDS) that will be used in the vectorized stmts.
1298 OP_NUM determines if we gather defs for operand 0 or operand 1 of the scalar
1299 stmts. */
1301 static void
1304 {
1310 tree vec_cst;
1313 tree vector_type;
1325 /* NUMBER_OF_COPIES is the number of times we need to use the same values in
1326 created vectors. It is greater than 1 if unrolling is performed.
1328 For example, we have two scalar operands, s1 and s2 (e.g., group of
1329 strided accesses of size two), while NUINTS is four (i.e., four scalars
1330 of this type can be packed in a vector). The output vector will contain
1331 two copies of each scalar operand: {s1, s2, s1, s2}. (NUMBER_OF_COPIES
1332 will be 2).
1334 If GROUP_SIZE > NUNITS, the scalars will be split into several vectors
1335 containing the operands.
1337 For example, NUINTS is four as before, and the group size is 8
1338 (s1, s2, ..., s8). We will create two vectors {s1, s2, s3, s4} and
1339 {s5, s6, s7, s8}. */
1345 {
1347 {
1351 else
1354 /* Create 'vect_ = {op0,op1,...,opn}'. */
1357 number_of_places_left_in_vector--;
1360 {
1368 NULL));
1370 }
1371 }
1372 }
1374 /* Since the vectors are created in the reverse order, we should invert
1375 them. */
1378 {
1381 }
1385 /* In case that VF is greater than the unrolling factor needed for the SLP
1386 group of stmts, NUMBER_OF_VECTORS to be created is greater than
1387 NUMBER_OF_SCALARS/NUNITS or NUNITS/NUMBER_OF_SCALARS, and hence we have
1388 to replicate the vectors. */
1390 {
1393 }
1394 }
1397 /* Get vectorized definitions from SLP_NODE that contains corresponding
1398 vectorized def-stmts. */
1400 static void
1402 {
1403 tree vec_oprnd;
1404 tree vec_def_stmt;
1411 i++)
1412 {
1416 }
1417 }
1420 /* Get vectorized definitions for SLP_NODE.
1421 If the scalar definitions are loop invariants or constants, collect them and
1422 call vect_get_constant_vectors() to create vector stmts.
1423 Otherwise, the def-stmts must be already vectorized and the vectorized stmts
1424 must be stored in the LEFT/RIGHT node of SLP_NODE, and we call
1425 vect_get_slp_vect_defs() to retrieve them.
1426 If VEC_OPRNDS1 is NULL, don't get vector defs for the second operand (from
1427 the right node. This is used when the second operand must remain scalar. */
1429 static void
1432 {
1435 /* Allocate memory for vectorized defs. */
1439 /* SLP_NODE corresponds either to a group of stores or to a group of
1440 unary/binary operations. We don't call this function for loads. */
1442 /* The defs are already vectorized. */
1444 else
1445 /* Build vectors from scalar defs. */
1450 /* Since we don't call this function with loads, this is a group of
1451 stores. */
1462 /* The defs are already vectorized. */
1464 else
1465 /* Build vectors from scalar defs. */
1467 }
1470 /* Function get_initial_def_for_induction
1472 Input:
1473 STMT - a stmt that performs an induction operation in the loop.
1474 IV_PHI - the initial value of the induction variable
1476 Output:
1477 Return a vector variable, initialized with the first VF values of
1478 the induction variable. E.g., for an iv with IV_PHI='X' and
1479 evolution S, for a vector of 4 units, we want to return:
1480 [X, X + S, X + 2*S, X + 3*S]. */
1482 static tree
1484 {
1489 tree vectype;
1493 basic_block new_bb;
1495 tree access_fn;
1496 tree new_var;
1497 tree new_name;
1498 tree init_stmt;
1505 tree expr;
1508 tree stmts;
1509 imm_use_iterator imm_iter;
1510 use_operand_p use_p;
1511 tree exit_phi;
1512 edge latch_e;
1513 tree loop_arg;
1514 block_stmt_iterator si;
1525 /* Find the first insertion point in the BB. */
1530 else
1533 /* Is phi in an inner-loop, while vectorizing an enclosing outer-loop? */
1535 {
1538 }
1539 else
1553 /* Create the vector that holds the initial_value of the induction. */
1555 {
1556 /* iv_loop is nested in the loop to be vectorized. init_expr had already
1557 been created during vectorization of previous stmts; We obtain it from
1558 the STMT_VINFO_VEC_STMT of the defining stmt. */
1561 }
1562 else
1563 {
1564 /* iv_loop is the loop to be vectorized. Create:
1565 vec_init = [X, X+S, X+2*S, X+3*S] (S = step_expr, X = init_expr) */
1571 {
1574 }
1579 {
1580 tree tmp;
1582 /* Create: new_name_i = new_name + step_expr */
1592 {
1595 }
1597 }
1598 /* Create a vector from [new_name_0, new_name_1, ..., new_name_nunits-1] */
1601 }
1604 /* Create the vector that holds the step of the induction. */
1606 /* iv_loop is nested in the loop to be vectorized. Generate:
1607 vec_step = [S, S, S, S] */
1609 else
1610 {
1611 /* iv_loop is the loop to be vectorized. Generate:
1612 vec_step = [VF*S, VF*S, VF*S, VF*S] */
1615 }
1624 /* Create the following def-use cycle:
1625 loop prolog:
1626 vec_init = ...
1627 vec_step = ...
1628 loop:
1629 vec_iv = PHI <vec_init, vec_loop>
1630 ...
1631 STMT
1632 ...
1633 vec_loop = vec_iv + vec_step; */
1635 /* Create the induction-phi that defines the induction-operand. */
1643 /* Create the iv update inside the loop */
1653 /* Set the arguments of the phi node: */
1658 /* In case that vectorization factor (VF) is bigger than the number
1659 of elements that we can fit in a vectype (nunits), we have to generate
1660 more than one vector stmt - i.e - we need to "unroll" the
1661 vector stmt by a factor VF/nunits. For more details see documentation
1662 in vectorizable_operation. */
1665 {
1666 stmt_vec_info prev_stmt_vinfo;
1667 /* FORNOW. This restriction should be relaxed. */
1670 /* Create the vector that holds the step of the induction. */
1682 {
1683 tree tmp;
1685 /* vec_i = vec_prev + vec_step */
1695 }
1696 }
1699 {
1700 /* Find the loop-closed exit-phi of the induction, and record
1701 the final vector of induction results: */
1704 {
1706 {
1709 }
1710 }
1712 {
1714 /* FORNOW. Currently not supporting the case that an inner-loop induction
1715 is not used in the outer-loop (i.e. only outside the outer-loop). */
1721 {
1724 }
1725 }
1726 }
1730 {
1735 }
1739 }
1742 /* Function vect_get_vec_def_for_operand.
1744 OP is an operand in STMT. This function returns a (vector) def that will be
1745 used in the vectorized stmt for STMT.
1747 In the case that OP is an SSA_NAME which is defined in the loop, then
1748 STMT_VINFO_VEC_STMT of the defining stmt holds the relevant def.
1750 In case OP is an invariant or constant, a new stmt that creates a vector def
1751 needs to be introduced. */
1753 static tree
1755 {
1756 tree vec_oprnd;
1757 tree vec_stmt;
1758 tree def_stmt;
1764 tree vec_inv;
1765 tree vec_cst;
1767 tree def;
1771 tree vector_type;
1774 {
1777 }
1782 {
1784 {
1787 }
1789 {
1792 }
1793 }
1796 {
1797 /* Case 1: operand is a constant. */
1799 {
1803 /* Create 'vect_cst_ = {cst,cst,...,cst}' */
1808 {
1810 }
1816 }
1818 /* Case 2: operand is defined outside the loop - loop invariant. */
1820 {
1824 /* Create 'vec_inv = {inv,inv,..,inv}' */
1829 {
1831 }
1833 /* FIXME: use build_constructor directly. */
1838 }
1840 /* Case 3: operand is defined inside the loop. */
1842 {
1846 /* Get the def from the vectorized stmt. */
1852 else
1855 }
1857 /* Case 4: operand is defined by a loop header phi - reduction */
1859 {
1865 /* Get the def before the loop */
1868 }
1870 /* Case 5: operand is defined by loop-header phi - induction. */
1872 {
1875 /* Get the def from the vectorized stmt. */
1881 }
1885 }
1886 }
1889 /* Function vect_get_vec_def_for_stmt_copy
1891 Return a vector-def for an operand. This function is used when the
1892 vectorized stmt to be created (by the caller to this function) is a "copy"
1893 created in case the vectorized result cannot fit in one vector, and several
1894 copies of the vector-stmt are required. In this case the vector-def is
1895 retrieved from the vector stmt recorded in the STMT_VINFO_RELATED_STMT field
1896 of the stmt that defines VEC_OPRND.
1897 DT is the type of the vector def VEC_OPRND.
1899 Context:
1900 In case the vectorization factor (VF) is bigger than the number
1901 of elements that can fit in a vectype (nunits), we have to generate
1902 more than one vector stmt to vectorize the scalar stmt. This situation
1903 arises when there are multiple data-types operated upon in the loop; the
1904 smallest data-type determines the VF, and as a result, when vectorizing
1905 stmts operating on wider types we need to create 'VF/nunits' "copies" of the
1906 vector stmt (each computing a vector of 'nunits' results, and together
1907 computing 'VF' results in each iteration). This function is called when
1908 vectorizing such a stmt (e.g. vectorizing S2 in the illustration below, in
1909 which VF=16 and nunits=4, so the number of copies required is 4):
1911 scalar stmt: vectorized into: STMT_VINFO_RELATED_STMT
1913 S1: x = load VS1.0: vx.0 = memref0 VS1.1
1914 VS1.1: vx.1 = memref1 VS1.2
1915 VS1.2: vx.2 = memref2 VS1.3
1916 VS1.3: vx.3 = memref3
1918 S2: z = x + ... VSnew.0: vz0 = vx.0 + ... VSnew.1
1919 VSnew.1: vz1 = vx.1 + ... VSnew.2
1920 VSnew.2: vz2 = vx.2 + ... VSnew.3
1921 VSnew.3: vz3 = vx.3 + ...
1923 The vectorization of S1 is explained in vectorizable_load.
1924 The vectorization of S2:
1925 To create the first vector-stmt out of the 4 copies - VSnew.0 -
1926 the function 'vect_get_vec_def_for_operand' is called to
1927 get the relevant vector-def for each operand of S2. For operand x it
1928 returns the vector-def 'vx.0'.
1930 To create the remaining copies of the vector-stmt (VSnew.j), this
1931 function is called to get the relevant vector-def for each operand. It is
1932 obtained from the respective VS1.j stmt, which is recorded in the
1933 STMT_VINFO_RELATED_STMT field of the stmt that defines VEC_OPRND.
1935 For example, to obtain the vector-def 'vx.1' in order to create the
1936 vector stmt 'VSnew.1', this function is called with VEC_OPRND='vx.0'.
1937 Given 'vx0' we obtain the stmt that defines it ('VS1.0'); from the
1938 STMT_VINFO_RELATED_STMT field of 'VS1.0' we obtain the next copy - 'VS1.1',
1939 and return its def ('vx.1').
1940 Overall, to create the above sequence this function will be called 3 times:
1941 vx.1 = vect_get_vec_def_for_stmt_copy (dt, vx.0);
1942 vx.2 = vect_get_vec_def_for_stmt_copy (dt, vx.1);
1943 vx.3 = vect_get_vec_def_for_stmt_copy (dt, vx.2); */
1945 static tree
1947 {
1948 tree vec_stmt_for_operand;
1949 stmt_vec_info def_stmt_info;
1951 /* Do nothing; can reuse same def. */
1962 }
1965 /* Get vectorized definitions for the operands to create a copy of an original
1966 stmt. See vect_get_vec_def_for_stmt_copy() for details. */
1968 static void
1972 {
1979 {
1983 }
1984 }
1987 /* Get vectorized definitions for OP0 and OP1, or SLP_NODE if it is not NULL. */
1989 static void
1992 {
1995 else
1996 {
1997 tree vec_oprnd;
2004 {
2008 }
2009 }
2010 }
2013 /* Function vect_finish_stmt_generation.
2015 Insert a new stmt. */
2017 static void
2020 {
2033 {
2036 }
2038 /* Make sure bsi points to the stmt that is being vectorized. */
2041 #ifdef USE_MAPPED_LOCATION
2043 #else
2045 #endif
2046 }
2049 /* Function get_initial_def_for_reduction
2051 Input:
2052 STMT - a stmt that performs a reduction operation in the loop.
2053 INIT_VAL - the initial value of the reduction variable
2055 Output:
2056 ADJUSTMENT_DEF - a tree that holds a value to be added to the final result
2057 of the reduction (used for adjusting the epilog - see below).
2058 Return a vector variable, initialized according to the operation that STMT
2059 performs. This vector will be used as the initial value of the
2060 vector of partial results.
2062 Option1 (adjust in epilog): Initialize the vector as follows:
2063 add: [0,0,...,0,0]
2064 mult: [1,1,...,1,1]
2065 min/max: [init_val,init_val,..,init_val,init_val]
2066 bit and/or: [init_val,init_val,..,init_val,init_val]
2067 and when necessary (e.g. add/mult case) let the caller know
2068 that it needs to adjust the result by init_val.
2070 Option2: Initialize the vector as follows:
2071 add: [0,0,...,0,init_val]
2072 mult: [1,1,...,1,init_val]
2073 min/max: [init_val,init_val,...,init_val]
2074 bit and/or: [init_val,init_val,...,init_val]
2075 and no adjustments are needed.
2077 For example, for the following code:
2079 s = init_val;
2080 for (i=0;i<n;i++)
2081 s = s + a[i];
2083 STMT is 's = s + a[i]', and the reduction variable is 's'.
2084 For a vector of 4 units, we want to return either [0,0,0,init_val],
2085 or [0,0,0,0] and let the caller know that it needs to adjust
2086 the result at the end by 'init_val'.
2088 FORNOW, we are using the 'adjust in epilog' scheme, because this way the
2089 initialization vector is simpler (same element in all entries).
2090 A cost model should help decide between these two schemes. */
2092 static tree
2094 {
2102 tree vecdef;
2103 tree def_for_init;
2104 tree init_def;
2107 tree vector_type;
2113 else
2119 {
2125 else
2127 /* Create a vector of zeros for init_def. */
2130 else
2147 }
2150 }
2153 /* Function vect_create_epilog_for_reduction
2155 Create code at the loop-epilog to finalize the result of a reduction
2156 computation.
2158 VECT_DEF is a vector of partial results.
2159 REDUC_CODE is the tree-code for the epilog reduction.
2160 STMT is the scalar reduction stmt that is being vectorized.
2161 REDUCTION_PHI is the phi-node that carries the reduction computation.
2163 This function:
2164 1. Creates the reduction def-use cycle: sets the arguments for
2165 REDUCTION_PHI:
2166 The loop-entry argument is the vectorized initial-value of the reduction.
2167 The loop-latch argument is VECT_DEF - the vector of partial sums.
2168 2. "Reduces" the vector of partial results VECT_DEF into a single result,
2169 by applying the operation specified by REDUC_CODE if available, or by
2170 other means (whole-vector shifts or a scalar loop).
2171 The function also creates a new phi node at the loop exit to preserve
2172 loop-closed form, as illustrated below.
2174 The flow at the entry to this function:
2176 loop:
2177 vec_def = phi <null, null> # REDUCTION_PHI
2178 VECT_DEF = vector_stmt # vectorized form of STMT
2179 s_loop = scalar_stmt # (scalar) STMT
2180 loop_exit:
2181 s_out0 = phi <s_loop> # (scalar) EXIT_PHI
2182 use <s_out0>
2183 use <s_out0>
2185 The above is transformed by this function into:
2187 loop:
2188 vec_def = phi <vec_init, VECT_DEF> # REDUCTION_PHI
2189 VECT_DEF = vector_stmt # vectorized form of STMT
2190 s_loop = scalar_stmt # (scalar) STMT
2191 loop_exit:
2192 s_out0 = phi <s_loop> # (scalar) EXIT_PHI
2193 v_out1 = phi <VECT_DEF> # NEW_EXIT_PHI
2194 v_out2 = reduce <v_out1>
2195 s_out3 = extract_field <v_out2, 0>
2196 s_out4 = adjust_result <s_out3>
2197 use <s_out4>
2198 use <s_out4>
2199 */
2201 static void
2204 {
2206 tree vectype;
2210 basic_block exit_bb;
2211 tree scalar_dest;
2212 tree scalar_type;
2213 tree new_phi;
2214 block_stmt_iterator exit_bsi;
2215 tree vec_dest;
2217 tree new_name;
2222 tree adjustment_def;
2223 tree vec_initial_def;
2224 tree orig_name;
2225 imm_use_iterator imm_iter;
2226 use_operand_p use_p;
2229 tree orig_stmt;
2230 tree use_stmt;
2238 {
2241 }
2249 /*** 1. Create the reduction def-use cycle ***/
2251 /* 1.1 set the loop-entry arg of the reduction-phi: */
2252 /* For the case of reduction, vect_get_vec_def_for_operand returns
2253 the scalar def before the loop, that defines the initial value
2254 of the reduction variable. */
2259 /* 1.2 set the loop-latch arg for the reduction-phi: */
2263 {
2268 }
2271 /*** 2. Create epilog code
2272 The reduction epilog code operates across the elements of the vector
2273 of partial results computed by the vectorized loop.
2274 The reduction epilog code consists of:
2275 step 1: compute the scalar result in a vector (v_out2)
2276 step 2: extract the scalar result (s_out3) from the vector (v_out2)
2277 step 3: adjust the scalar result (s_out3) if needed.
2279 Step 1 can be accomplished using one the following three schemes:
2280 (scheme 1) using reduc_code, if available.
2281 (scheme 2) using whole-vector shifts, if available.
2282 (scheme 3) using a scalar loop. In this case steps 1+2 above are
2283 combined.
2285 The overall epilog code looks like this:
2287 s_out0 = phi <s_loop> # original EXIT_PHI
2288 v_out1 = phi <VECT_DEF> # NEW_EXIT_PHI
2289 v_out2 = reduce <v_out1> # step 1
2290 s_out3 = extract_field <v_out2, 0> # step 2
2291 s_out4 = adjust_result <s_out3> # step 3
2293 (step 3 is optional, and step2 1 and 2 may be combined).
2294 Lastly, the uses of s_out0 are replaced by s_out4.
2296 ***/
2298 /* 2.1 Create new loop-exit-phi to preserve loop-closed form:
2299 v_out1 = phi <v_loop> */
2306 /* 2.2 Get the relevant tree-code to use in the epilog for schemes 2,3
2307 (i.e. when reduc_code is not available) and in the final adjustment
2308 code (if needed). Also get the original scalar reduction variable as
2309 defined in the loop. In case STMT is a "pattern-stmt" (i.e. - it
2310 represents a reduction pattern), the tree-code and scalar-def are
2311 taken from the original stmt that the pattern-stmt (STMT) replaces.
2312 Otherwise (it is a regular reduction) - the tree-code and scalar-def
2313 are taken from STMT. */
2317 {
2318 /* Regular reduction */
2320 }
2321 else
2322 {
2323 /* Reduction pattern */
2327 }
2336 /* In case this is a reduction in an inner-loop while vectorizing an outer
2337 loop - we don't need to extract a single scalar result at the end of the
2338 inner-loop. The final vector of partial results will be used in the
2339 vectorized outer-loop, or reduced to a scalar result at the end of the
2340 outer-loop. */
2344 /* 2.3 Create the reduction code, using one of the three schemes described
2345 above. */
2348 {
2349 tree tmp;
2351 /*** Case 1: Create:
2352 v_out2 = reduc_expr <v_out1> */
2365 }
2366 else
2367 {
2373 tree vec_temp;
2377 else
2380 /* Regardless of whether we have a whole vector shift, if we're
2381 emulating the operation via tree-vect-generic, we don't want
2382 to use it. Only the first round of the reduction is likely
2383 to still be profitable via emulation. */
2384 /* ??? It might be better to emit a reduction tree code here, so that
2385 tree-vect-generic can expand the first round via bit tricks. */
2388 else
2389 {
2393 }
2396 {
2397 /*** Case 2: Create:
2398 for (offset = VS/2; offset >= element_size; offset/=2)
2399 {
2400 Create: va' = vec_shift <va, offset>
2401 Create: va = vop <va, va'>
2402 } */
2413 {
2426 }
2429 }
2430 else
2431 {
2432 tree rhs;
2434 /*** Case 3: Create:
2435 s = extract_field <v_out2, 0>
2436 for (offset = element_size;
2437 offset < vector_size;
2438 offset += element_size;)
2439 {
2440 Create: s' = extract_field <v_out2, offset>
2441 Create: s = op <s, s'>
2442 } */
2450 bitsize_zero_node);
2460 {
2461 tree tmp;
2464 bitpos);
2477 }
2480 }
2481 }
2483 /* 2.4 Extract the final scalar result. Create:
2484 s_out3 = extract_field <v_out2, bitpos> */
2487 {
2488 tree rhs;
2498 else
2507 }
2509 vect_finalize_reduction:
2511 /* 2.5 Adjust the final result by the initial value of the reduction
2512 variable. (When such adjustment is not needed, then
2513 'adjustment_def' is zero). For example, if code is PLUS we create:
2514 new_temp = loop_exit_def + adjustment_def */
2517 {
2519 {
2523 }
2524 else
2525 {
2529 }
2534 }
2537 /* 2.6 Handle the loop-exit phi */
2539 /* Replace uses of s_out0 with uses of s_out3:
2540 Find the loop-closed-use at the loop exit of the original scalar result.
2541 (The reduction result is expected to have two immediate uses - one at the
2542 latch block, and one at the loop exit). */
2545 {
2547 {
2550 }
2551 }
2552 /* We expect to have found an exit_phi because of loop-closed-ssa form. */
2556 {
2558 {
2561 /* FORNOW. Currently not supporting the case that an inner-loop reduction
2562 is not used in the outer-loop (but only outside the outer-loop). */
2571 }
2573 /* Replace the uses: */
2578 }
2580 }
2583 /* Function vectorizable_reduction.
2585 Check if STMT performs a reduction operation that can be vectorized.
2586 If VEC_STMT is also passed, vectorize the STMT: create a vectorized
2587 stmt to replace it, put it in VEC_STMT, and insert it at BSI.
2588 Return FALSE if not a vectorizable STMT, TRUE otherwise.
2590 This function also handles reduction idioms (patterns) that have been
2591 recognized in advance during vect_pattern_recog. In this case, STMT may be
2592 of this form:
2593 X = pattern_expr (arg0, arg1, ..., X)
2594 and it's STMT_VINFO_RELATED_STMT points to the last stmt in the original
2595 sequence that had been detected and replaced by the pattern-stmt (STMT).
2597 In some cases of reduction patterns, the type of the reduction variable X is
2598 different than the type of the other arguments of STMT.
2599 In such cases, the vectype that is used when transforming STMT into a vector
2600 stmt is different than the vectype that is used to determine the
2601 vectorization factor, because it consists of a different number of elements
2602 than the actual number of elements that are being operated upon in parallel.
2604 For example, consider an accumulation of shorts into an int accumulator.
2605 On some targets it's possible to vectorize this pattern operating on 8
2606 shorts at a time (hence, the vectype for purposes of determining the
2607 vectorization factor should be V8HI); on the other hand, the vectype that
2608 is used to create the vector form is actually V4SI (the type of the result).
2610 Upon entry to this function, STMT_VINFO_VECTYPE records the vectype that
2611 indicates what is the actual level of parallelism (V8HI in the example), so
2612 that the right vectorization factor would be derived. This vectype
2613 corresponds to the type of arguments to the reduction stmt, and should *NOT*
2614 be used to create the vectorized stmt. The right vectype for the vectorized
2615 stmt is obtained from the type of the result X:
2616 get_vectype_for_scalar_type (TREE_TYPE (X))
2618 This means that, contrary to "regular" reductions (or "regular" stmts in
2619 general), the following equation:
2620 STMT_VINFO_VECTYPE == get_vectype_for_scalar_type (TREE_TYPE (X))
2621 does *NOT* necessarily hold for reduction patterns. */
2623 bool
2625 {
2626 tree vec_dest;
2627 tree scalar_dest;
2628 tree op;
2634 tree operation;
2642 tree new_phi;
2643 tree scalar_type;
2645 tree orig_stmt;
2646 stmt_vec_info orig_stmt_info;
2651 stmt_vec_info prev_stmt_info;
2652 tree reduc_def;
2657 {
2659 /* FORNOW. This restriction should be relaxed. */
2661 {
2665 }
2666 }
2670 /* FORNOW: SLP not supported. */
2674 /* 1. Is vectorizable reduction? */
2676 /* Not supportable if the reduction variable is used in the loop. */
2680 /* Reductions that are not used even in an enclosing outer-loop,
2681 are expected to be "live" (used out of the loop). */
2686 /* Make sure it was already recognized as a reduction computation. */
2690 /* 2. Has this been recognized as a reduction pattern?
2692 Check if STMT represents a pattern that has been recognized
2693 in earlier analysis stages. For stmts that represent a pattern,
2694 the STMT_VINFO_RELATED_STMT field records the last stmt in
2695 the original sequence that constitutes the pattern. */
2699 {
2704 }
2706 /* 3. Check the operands of the operation. The first operands are defined
2707 inside the loop body. The last operand is the reduction variable,
2708 which is defined by the loop-header-phi. */
2720 /* All uses but the last are expected to be defined in the loop.
2721 The last use is the reduction variable. */
2723 {
2732 }
2741 else
2747 /* 4. Supportable by target? */
2749 /* 4.1. check support for the operation in the loop */
2752 {
2756 }
2759 {
2768 }
2770 /* Worthwhile without SIMD support? */
2774 {
2778 }
2780 /* 4.2. Check support for the epilog operation.
2782 If STMT represents a reduction pattern, then the type of the
2783 reduction variable may be different than the type of the rest
2784 of the arguments. For example, consider the case of accumulation
2785 of shorts into an int accumulator; The original code:
2786 S1: int_a = (int) short_a;
2787 orig_stmt-> S2: int_acc = plus <int_a ,int_acc>;
2789 was replaced with:
2790 STMT: int_acc = widen_sum <short_a, int_acc>
2792 This means that:
2793 1. The tree-code that is used to create the vector operation in the
2794 epilog code (that reduces the partial results) is not the
2795 tree-code of STMT, but is rather the tree-code of the original
2796 stmt from the pattern that STMT is replacing. I.e, in the example
2797 above we want to use 'widen_sum' in the loop, but 'plus' in the
2798 epilog.
2799 2. The type (mode) we use to check available target support
2800 for the vector operation to be created in the *epilog*, is
2801 determined by the type of the reduction variable (in the example
2802 above we'd check this: plus_optab[vect_int_mode]).
2803 However the type (mode) we use to check available target support
2804 for the vector operation to be created *inside the loop*, is
2805 determined by the type of the other arguments to STMT (in the
2806 example we'd check this: widen_sum_optab[vect_short_mode]).
2808 This is contrary to "regular" reductions, in which the types of all
2809 the arguments are the same as the type of the reduction variable.
2810 For "regular" reductions we can therefore use the same vector type
2811 (and also the same tree-code) when generating the epilog code and
2812 when generating the code inside the loop. */
2815 {
2816 /* This is a reduction pattern: get the vectype from the type of the
2817 reduction variable, and get the tree-code from orig_stmt. */
2821 {
2823 {
2826 }
2828 }
2831 }
2832 else
2833 {
2834 /* Regular reduction: use the same vectype and tree-code as used for
2835 the vector code inside the loop can be used for the epilog code. */
2837 }
2843 {
2847 }
2849 {
2853 }
2856 {
2861 }
2863 /** Transform. **/
2868 /* Create the destination vector */
2871 /* Create the reduction-phi that defines the reduction-operand. */
2874 /* In case the vectorization factor (VF) is bigger than the number
2875 of elements that we can fit in a vectype (nunits), we have to generate
2876 more than one vector stmt - i.e - we need to "unroll" the
2877 vector stmt by a factor VF/nunits. For more details see documentation
2878 in vectorizable_operation. */
2882 {
2883 /* Handle uses. */
2885 {
2889 {
2892 }
2894 /* Get the vector def for the reduction variable from the phi node */
2896 }
2897 else
2898 {
2904 /* Get the vector def for the reduction variable from the vectorized
2905 reduction operation generated in the previous iteration (j-1) */
2907 }
2909 /* Arguments are ready. create the new vector stmt. */
2912 else
2914 reduc_def);
2922 else
2925 }
2927 /* Finalize the reduction-phi (set it's arguments) and create the
2928 epilog reduction code. */
2931 }
2933 /* Checks if CALL can be vectorized in type VECTYPE. Returns
2934 a function declaration if the target has a vectorized version
2935 of the function, or NULL_TREE if the function cannot be vectorized. */
2937 tree
2939 {
2943 /* We only handle functions that do not read or clobber memory -- i.e.
2944 const or novops ones. */
2955 vectype_in);
2956 }
2958 /* Function vectorizable_call.
2960 Check if STMT performs a function call that can be vectorized.
2961 If VEC_STMT is also passed, vectorize the STMT: create a vectorized
2962 stmt to replace it, put it in VEC_STMT, and insert it at BSI.
2963 Return FALSE if not a vectorizable STMT, TRUE otherwise. */
2965 bool
2967 {
2968 tree vec_dest;
2969 tree scalar_dest;
2970 tree operation;
2981 tree new_stmt;
2983 call_expr_arg_iterator iter;
2984 tree vargs;
2993 /* FORNOW: SLP not supported. */
2997 /* Is STMT a vectorizable call? */
3008 /* Process function arguments. */
3012 {
3013 /* Bail out if the function has more than two arguments, we
3014 do not have interesting builtin functions to vectorize with
3015 more than two arguments. */
3019 /* We can only handle calls with arguments of the same type. */
3022 {
3026 }
3030 {
3034 }
3037 }
3039 /* No arguments is also not good. */
3054 /* FORNOW */
3061 else
3064 /* For now, we only vectorize functions if a target specific builtin
3065 is available. TODO -- in some cases, it might be profitable to
3066 insert the calls for pieces of the vector, in order to be able
3067 to vectorize other operations in the loop. */
3070 {
3075 }
3081 else
3084 /* Sanity check: make sure that at least one copy of the vectorized stmt
3085 needs to be generated. */
3088 /* FORNOW. This restriction should be relaxed. */
3090 {
3094 }
3097 {
3103 }
3105 /** Transform. **/
3110 /* FORNOW. This restriction should be relaxed. */
3112 {
3116 }
3118 /* Handle def. */
3124 {
3127 {
3128 /* Build argument list for the vectorized call. */
3129 /* FIXME: Rewrite this so that it doesn't
3130 construct a temporary list. */
3134 {
3136 vec_oprnd0
3138 else
3139 vec_oprnd0
3145 }
3157 else
3161 }
3167 {
3168 /* Build argument list for the vectorized call. */
3169 /* FIXME: Rewrite this so that it doesn't
3170 construct a temporary list. */
3174 {
3176 {
3177 vec_oprnd0
3179 vec_oprnd1
3181 }
3182 else
3183 {
3184 vec_oprnd0
3186 vec_oprnd1
3188 }
3194 }
3206 else
3210 }
3217 /* No current target implements this case. */
3219 }
3221 /* The call in STMT might prevent it from being removed in dce.
3222 We however cannot remove it here, due to the way the ssa name
3223 it defines is mapped to the new definition. So just replace
3224 rhs of the statement with something harmless. */
3230 }
3233 /* Function vect_gen_widened_results_half
3235 Create a vector stmt whose code, type, number of arguments, and result
3236 variable are CODE, VECTYPE, OP_TYPE, and VEC_DEST, and its arguments are
3237 VEC_OPRND0 and VEC_OPRND1. The new vector stmt is to be inserted at BSI.
3238 In the case that CODE is a CALL_EXPR, this means that a call to DECL
3239 needs to be created (DECL is a function-decl of a target-builtin).
3240 STMT is the original scalar stmt that we are vectorizing. */
3242 static tree
3246 tree stmt)
3247 {
3248 tree expr;
3249 tree new_stmt;
3250 tree new_temp;
3251 tree sym;
3252 ssa_op_iter iter;
3254 /* Generate half of the widened result: */
3256 {
3257 /* Target specific support */
3260 else
3262 }
3263 else
3264 {
3265 /* Generic support */
3269 else
3271 }
3278 {
3280 {
3284 }
3285 }
3288 }
3291 /* Check if STMT performs a conversion operation, that can be vectorized.
3292 If VEC_STMT is also passed, vectorize the STMT: create a vectorized
3293 stmt to replace it, put it in VEC_STMT, and insert it at BSI.
3294 Return FALSE if not a vectorizable STMT, TRUE otherwise. */
3296 bool
3299 {
3300 tree vec_dest;
3301 tree scalar_dest;
3302 tree operation;
3303 tree op0;
3310 tree new_temp;
3314 stmt_vec_info prev_stmt_info;
3319 tree expr;
3321 tree builtin_decl;
3325 tree vop0;
3327 /* Is STMT a vectorizable conversion? */
3346 /* Check types of lhs and rhs. */
3361 /* FORNOW */
3368 else
3374 /* Bail out if the types are both integral or non-integral. */
3381 else
3384 /* FORNOW: SLP with multiple types is not supported. The SLP analysis verifies
3385 this, so we can safely override NCOPIES with 1 here. */
3389 /* Sanity check: make sure that at least one copy of the vectorized stmt
3390 needs to be generated. */
3393 /* FORNOW. This restriction should be relaxed. */
3395 {
3399 }
3401 /* Check the operands of the operation. */
3403 {
3407 }
3409 /* Supportable by target? */
3419 {
3423 }
3426 {
3428 /* FORNOW: SLP not supported. */
3431 }
3434 {
3437 }
3439 /** Transform. **/
3443 /* Handle def. */
3451 {
3454 {
3455 tree sym;
3456 ssa_op_iter iter;
3460 else
3463 builtin_decl =
3466 {
3469 /* Arguments are ready. create the new vector stmt. */
3475 SSA_OP_ALL_VIRTUALS)
3476 {
3480 }
3483 }
3487 else
3490 }
3494 /* In case the vectorization factor (VF) is bigger than the number
3495 of elements that we can fit in a vectype (nunits), we have to
3496 generate more than one vector stmt - i.e - we need to "unroll"
3497 the vector stmt by a factor VF/nunits. */
3499 {
3502 else
3507 /* Generate first half of the widened result: */
3508 new_stmt
3514 else
3518 /* Generate second half of the widened result: */
3519 new_stmt
3525 }
3529 /* In case the vectorization factor (VF) is bigger than the number
3530 of elements that we can fit in a vectype (nunits), we have to
3531 generate more than one vector stmt - i.e - we need to "unroll"
3532 the vector stmt by a factor VF/nunits. */
3534 {
3535 /* Handle uses. */
3537 {
3540 }
3541 else
3542 {
3545 }
3547 /* Arguments are ready. Create the new vector stmt. */
3556 else
3560 }
3563 }
3566 }
3569 /* Function vectorizable_assignment.
3571 Check if STMT performs an assignment (copy) that can be vectorized.
3572 If VEC_STMT is also passed, vectorize the STMT: create a vectorized
3573 stmt to replace it, put it in VEC_STMT, and insert it at BSI.
3574 Return FALSE if not a vectorizable STMT, TRUE otherwise. */
3576 bool
3578 slp_tree slp_node)
3579 {
3580 tree vec_dest;
3581 tree scalar_dest;
3582 tree op;
3586 tree new_temp;
3593 tree vop;
3605 /* Is vectorizable assignment? */
3615 {
3619 }
3622 {
3628 }
3630 /** Transform. **/
3634 /* Handle def. */
3637 /* Handle use. */
3640 /* Arguments are ready. create the new vector stmt. */
3642 {
3651 }
3655 }
3658 /* Function vect_min_worthwhile_factor.
3660 For a loop where we could vectorize the operation indicated by CODE,
3661 return the minimum vectorization factor that makes it worthwhile
3662 to use generic vectors. */
3663 static int
3665 {
3667 {
3681 }
3682 }
3685 /* Function vectorizable_induction
3687 Check if PHI performs an induction computation that can be vectorized.
3688 If VEC_STMT is also passed, vectorize the induction PHI: create a vectorized
3689 phi to replace it, put it in VEC_STMT, and add it to the same basic block.
3690 Return FALSE if not a vectorizable STMT, TRUE otherwise. */
3692 bool
3695 {
3701 tree vec_def;
3708 /* FORNOW: SLP not supported. */
3718 {
3724 }
3726 /** Transform. **/
3734 }
3737 /* Function vectorizable_operation.
3739 Check if STMT performs a binary or unary operation that can be vectorized.
3740 If VEC_STMT is also passed, vectorize the STMT: create a vectorized
3741 stmt to replace it, put it in VEC_STMT, and insert it at BSI.
3742 Return FALSE if not a vectorizable STMT, TRUE otherwise. */
3744 bool
3746 slp_tree slp_node)
3747 {
3748 tree vec_dest;
3749 tree scalar_dest;
3750 tree operation;
3759 tree new_temp;
3761 optab optab;
3767 stmt_vec_info prev_stmt_info;
3770 tree vectype_out;
3776 /* FORNOW: SLP with multiple types is not supported. The SLP analysis verifies
3777 this, so we can safely override NCOPIES with 1 here. */
3781 /* FORNOW. This restriction should be relaxed. */
3783 {
3787 }
3795 /* Is STMT a vectorizable binary/unary operation? */
3813 /* For pointer addition, we should use the normal plus for
3814 the vector addition. */
3820 /* Support only unary or binary operations. */
3823 {
3827 }
3831 {
3835 }
3838 {
3841 {
3845 }
3846 }
3848 /* Supportable by target? */
3850 {
3854 }
3858 {
3861 /* Check only during analysis. */
3869 }
3871 /* Worthwhile without SIMD support? Check only during analysis. */
3876 {
3880 }
3883 {
3884 /* FORNOW: not yet supported. */
3888 /* Invariant argument is needed for a vector shift
3889 by a scalar shift operand. */
3894 {
3898 }
3899 }
3902 {
3908 }
3910 /** Transform. **/
3915 /* Handle def. */
3923 /* In case the vectorization factor (VF) is bigger than the number
3924 of elements that we can fit in a vectype (nunits), we have to generate
3925 more than one vector stmt - i.e - we need to "unroll" the
3926 vector stmt by a factor VF/nunits. In doing so, we record a pointer
3927 from one copy of the vector stmt to the next, in the field
3928 STMT_VINFO_RELATED_STMT. This is necessary in order to allow following
3929 stages to find the correct vector defs to be used when vectorizing
3930 stmts that use the defs of the current stmt. The example below illustrates
3931 the vectorization process when VF=16 and nunits=4 (i.e - we need to create
3932 4 vectorized stmts):
3934 before vectorization:
3935 RELATED_STMT VEC_STMT
3936 S1: x = memref - -
3937 S2: z = x + 1 - -
3939 step 1: vectorize stmt S1 (done in vectorizable_load. See more details
3940 there):
3941 RELATED_STMT VEC_STMT
3942 VS1_0: vx0 = memref0 VS1_1 -
3943 VS1_1: vx1 = memref1 VS1_2 -
3944 VS1_2: vx2 = memref2 VS1_3 -
3945 VS1_3: vx3 = memref3 - -
3946 S1: x = load - VS1_0
3947 S2: z = x + 1 - -
3949 step2: vectorize stmt S2 (done here):
3950 To vectorize stmt S2 we first need to find the relevant vector
3951 def for the first operand 'x'. This is, as usual, obtained from
3952 the vector stmt recorded in the STMT_VINFO_VEC_STMT of the stmt
3953 that defines 'x' (S1). This way we find the stmt VS1_0, and the
3954 relevant vector def 'vx0'. Having found 'vx0' we can generate
3955 the vector stmt VS2_0, and as usual, record it in the
3956 STMT_VINFO_VEC_STMT of stmt S2.
3957 When creating the second copy (VS2_1), we obtain the relevant vector
3958 def from the vector stmt recorded in the STMT_VINFO_RELATED_STMT of
3959 stmt VS1_0. This way we find the stmt VS1_1 and the relevant
3960 vector def 'vx1'. Using 'vx1' we create stmt VS2_1 and record a
3961 pointer to it in the STMT_VINFO_RELATED_STMT of the vector stmt VS2_0.
3962 Similarly when creating stmts VS2_2 and VS2_3. This is the resulting
3963 chain of stmts and pointers:
3964 RELATED_STMT VEC_STMT
3965 VS1_0: vx0 = memref0 VS1_1 -
3966 VS1_1: vx1 = memref1 VS1_2 -
3967 VS1_2: vx2 = memref2 VS1_3 -
3968 VS1_3: vx3 = memref3 - -
3969 S1: x = load - VS1_0
3970 VS2_0: vz0 = vx0 + v1 VS2_1 -
3971 VS2_1: vz1 = vx1 + v1 VS2_2 -
3972 VS2_2: vz2 = vx2 + v1 VS2_3 -
3973 VS2_3: vz3 = vx3 + v1 - -
3974 S2: z = x + 1 - VS2_0 */
3978 {
3979 /* Handle uses. */
3981 {
3984 {
3985 /* Vector shl and shr insn patterns can be defined with scalar
3986 operand 2 (shift operand). In this case, use constant or loop
3987 invariant op1 directly, without extending it to vector mode
3988 first. */
3991 {
3996 }
3997 }
3999 /* vec_oprnd is available if operand 1 should be of a scalar-type
4000 (a special case for certain kind of vector shifts); otherwise,
4001 operand 1 should be of a vector type (the usual case). */
4004 slp_node);
4005 else
4007 slp_node);
4008 }
4009 else
4012 /* Arguments are ready. Create the new vector stmt. */
4014 {
4016 {
4020 }
4021 else
4030 }
4034 else
4037 }
4044 }
4047 /* Function vectorizable_type_demotion
4049 Check if STMT performs a binary or unary operation that involves
4050 type demotion, and if it can be vectorized.
4051 If VEC_STMT is also passed, vectorize the STMT: create a vectorized
4052 stmt to replace it, put it in VEC_STMT, and insert it at BSI.
4053 Return FALSE if not a vectorizable STMT, TRUE otherwise. */
4055 bool
4058 {
4059 tree vec_dest;
4060 tree scalar_dest;
4061 tree operation;
4062 tree op0;
4068 tree new_temp;
4071 tree new_stmt;
4072 stmt_vec_info prev_stmt_info;
4075 tree vectype_out;
4078 tree expr;
4079 tree vectype_in;
4087 /* Is STMT a vectorizable type-demotion operation? */
4115 /* FORNOW. This restriction should be relaxed. */
4117 {
4121 }
4130 /* Check the operands of the operation. */
4132 {
4136 }
4138 /* Supportable by target? */
4145 {
4151 }
4153 /** Transform. **/
4156 ncopies);
4158 /* Handle def. */
4161 /* In case the vectorization factor (VF) is bigger than the number
4162 of elements that we can fit in a vectype (nunits), we have to generate
4163 more than one vector stmt - i.e - we need to "unroll" the
4164 vector stmt by a factor VF/nunits. */
4167 {
4168 /* Handle uses. */
4170 {
4173 }
4174 else
4175 {
4178 }
4180 /* Arguments are ready. Create the new vector stmt. */
4189 else
4193 }
4197 }
4200 /* Function vectorizable_type_promotion
4202 Check if STMT performs a binary or unary operation that involves
4203 type promotion, and if it can be vectorized.
4204 If VEC_STMT is also passed, vectorize the STMT: create a vectorized
4205 stmt to replace it, put it in VEC_STMT, and insert it at BSI.
4206 Return FALSE if not a vectorizable STMT, TRUE otherwise. */
4208 bool
4211 {
4212 tree vec_dest;
4213 tree scalar_dest;
4214 tree operation;
4225 tree new_stmt;
4226 stmt_vec_info prev_stmt_info;
4229 tree vectype_out;
4232 tree vectype_in;
4240 /* Is STMT a vectorizable type-promotion operation? */
4269 /* FORNOW. This restriction should be relaxed. */
4271 {
4275 }
4284 /* Check the operands of the operation. */
4286 {
4290 }
4294 {
4297 {
4301 }
4302 }
4304 /* Supportable by target? */
4312 {
4318 }
4320 /** Transform. **/
4324 ncopies);
4326 /* Handle def. */
4329 /* In case the vectorization factor (VF) is bigger than the number
4330 of elements that we can fit in a vectype (nunits), we have to generate
4331 more than one vector stmt - i.e - we need to "unroll" the
4332 vector stmt by a factor VF/nunits. */
4336 {
4337 /* Handle uses. */
4339 {
4343 }
4344 else
4345 {
4349 }
4351 /* Arguments are ready. Create the new vector stmt. We are creating
4352 two vector defs because the widened result does not fit in one vector.
4353 The vectorized stmt can be expressed as a call to a taregt builtin,
4354 or a using a tree-code. */
4355 /* Generate first half of the widened result: */
4360 else
4364 /* Generate second half of the widened result: */
4370 }
4374 }
4377 /* Function vect_strided_store_supported.
4379 Returns TRUE is INTERLEAVE_HIGH and INTERLEAVE_LOW operations are supported,
4380 and FALSE otherwise. */
4382 static bool
4384 {
4390 /* Check that the operation is supported. */
4392 vectype);
4394 vectype);
4396 {
4400 }
4403 == CODE_FOR_nothing
4406 {
4410 }
4413 }
4416 /* Function vect_permute_store_chain.
4418 Given a chain of interleaved stores in DR_CHAIN of LENGTH that must be
4419 a power of 2, generate interleave_high/low stmts to reorder the data
4420 correctly for the stores. Return the final references for stores in
4421 RESULT_CHAIN.
4423 E.g., LENGTH is 4 and the scalar type is short, i.e., VF is 8.
4424 The input is 4 vectors each containing 8 elements. We assign a number to each
4425 element, the input sequence is:
4427 1st vec: 0 1 2 3 4 5 6 7
4428 2nd vec: 8 9 10 11 12 13 14 15
4429 3rd vec: 16 17 18 19 20 21 22 23
4430 4th vec: 24 25 26 27 28 29 30 31
4432 The output sequence should be:
4434 1st vec: 0 8 16 24 1 9 17 25
4435 2nd vec: 2 10 18 26 3 11 19 27
4436 3rd vec: 4 12 20 28 5 13 21 30
4437 4th vec: 6 14 22 30 7 15 23 31
4439 i.e., we interleave the contents of the four vectors in their order.
4441 We use interleave_high/low instructions to create such output. The input of
4442 each interleave_high/low operation is two vectors:
4443 1st vec 2nd vec
4444 0 1 2 3 4 5 6 7
4445 the even elements of the result vector are obtained left-to-right from the
4446 high/low elements of the first vector. The odd elements of the result are
4447 obtained left-to-right from the high/low elements of the second vector.
4448 The output of interleave_high will be: 0 4 1 5
4449 and of interleave_low: 2 6 3 7
4452 The permutation is done in log LENGTH stages. In each stage interleave_high
4453 and interleave_low stmts are created for each pair of vectors in DR_CHAIN,
4454 where the first argument is taken from the first half of DR_CHAIN and the
4455 second argument from it's second half.
4456 In our example,
4458 I1: interleave_high (1st vec, 3rd vec)
4459 I2: interleave_low (1st vec, 3rd vec)
4460 I3: interleave_high (2nd vec, 4th vec)
4461 I4: interleave_low (2nd vec, 4th vec)
4463 The output for the first stage is:
4465 I1: 0 16 1 17 2 18 3 19
4466 I2: 4 20 5 21 6 22 7 23
4467 I3: 8 24 9 25 10 26 11 27
4468 I4: 12 28 13 29 14 30 15 31
4470 The output of the second stage, i.e. the final result is:
4472 I1: 0 8 16 24 1 9 17 25
4473 I2: 2 10 18 26 3 11 19 27
4474 I3: 4 12 20 28 5 13 21 30
4475 I4: 6 14 22 30 7 15 23 31. */
4477 static bool
4480 tree stmt,
4483 {
4495 /* Check that the operation is supported. */
4502 {
4504 {
4508 /* Create interleaving stmt:
4509 in the case of big endian:
4510 high = interleave_high (vect1, vect2)
4511 and in the case of little endian:
4512 high = interleave_low (vect1, vect2). */
4518 else
4526 /* Create interleaving stmt:
4527 in the case of big endian:
4528 low = interleave_low (vect1, vect2)
4529 and in the case of little endian:
4530 low = interleave_high (vect1, vect2). */
4536 else
4543 }
4545 }
4547 }
4550 /* Function vectorizable_store.
4552 Check if STMT defines a non scalar data-ref (array/pointer/structure) that
4553 can be vectorized.
4554 If VEC_STMT is also passed, vectorize the STMT: create a vectorized
4555 stmt to replace it, put it in VEC_STMT, and insert it at BSI.
4556 Return FALSE if not a vectorizable STMT, TRUE otherwise. */
4558 bool
4560 slp_tree slp_node)
4561 {
4562 tree scalar_dest;
4563 tree data_ref;
4564 tree op;
4572 tree dummy;
4588 stmt_vec_info first_stmt_vinfo;
4591 /* FORNOW: SLP with multiple types is not supported. The SLP analysis verifies
4592 this, so we can safely override NCOPIES with 1 here. */
4598 /* FORNOW. This restriction should be relaxed. */
4600 {
4604 }
4612 /* Is vectorizable store? */
4625 {
4629 }
4632 /* FORNOW. In some cases can vectorize even if data-type not supported
4633 (e.g. - array initialization with 0). */
4641 {
4649 {
4650 /* STMT is the leader of the group. Check the operands of all the
4651 stmts of the group. */
4654 {
4657 {
4661 }
4663 }
4664 }
4665 }
4668 {
4673 }
4675 /** Transform. **/
4678 {
4684 /* FORNOW */
4687 /* We vectorize all the stmts of the interleaving group when we
4688 reach the last stmt in the group. */
4692 {
4695 }
4700 /* VEC_NUM is the number of vect stmts to be created for this group. */
4703 else
4705 }
4706 else
4707 {
4712 }
4724 /* In case the vectorization factor (VF) is bigger than the number
4725 of elements that we can fit in a vectype (nunits), we have to generate
4726 more than one vector stmt - i.e - we need to "unroll" the
4727 vector stmt by a factor VF/nunits. For more details see documentation in
4728 vect_get_vec_def_for_copy_stmt. */
4730 /* In case of interleaving (non-unit strided access):
4732 S1: &base + 2 = x2
4733 S2: &base = x0
4734 S3: &base + 1 = x1
4735 S4: &base + 3 = x3
4737 We create vectorized stores starting from base address (the access of the
4738 first stmt in the chain (S2 in the above example), when the last store stmt
4739 of the chain (S4) is reached:
4741 VS1: &base = vx2
4742 VS2: &base + vec_size*1 = vx0
4743 VS3: &base + vec_size*2 = vx1
4744 VS4: &base + vec_size*3 = vx3
4746 Then permutation statements are generated:
4748 VS5: vx5 = VEC_INTERLEAVE_HIGH_EXPR < vx0, vx3 >
4749 VS6: vx6 = VEC_INTERLEAVE_LOW_EXPR < vx0, vx3 >
4750 ...
4752 And they are put in STMT_VINFO_VEC_STMT of the corresponding scalar stmts
4753 (the order of the data-refs in the output of vect_permute_store_chain
4754 corresponds to the order of scalar stmts in the interleaving chain - see
4755 the documentation of vect_permute_store_chain()).
4757 In case of both multiple types and interleaving, above vector stores and
4758 permutation stmts are created for every copy. The result vector stmts are
4759 put in STMT_VINFO_VEC_STMT for the first copy and in the corresponding
4760 STMT_VINFO_RELATED_STMT for the next copies.
4761 */
4765 {
4766 tree new_stmt;
4767 tree ptr_incr;
4770 {
4772 {
4773 /* Get vectorized arguments for SLP_NODE. */
4777 }
4778 else
4779 {
4780 /* For interleaved stores we collect vectorized defs for all the
4781 stores in the group in DR_CHAIN and OPRNDS. DR_CHAIN is then
4782 used as an input to vect_permute_store_chain(), and OPRNDS as
4783 an input to vect_get_vec_def_for_stmt_copy() for the next copy.
4785 If the store is not strided, GROUP_SIZE is 1, and DR_CHAIN and
4786 OPRNDS are of size 1. */
4789 {
4790 /* Since gaps are not supported for interleaved stores,
4791 GROUP_SIZE is the exact number of stmts in the chain.
4792 Therefore, NEXT_STMT can't be NULL_TREE. In case that
4793 there is no interleaving, GROUP_SIZE is 1, and only one
4794 iteration of the loop will be executed. */
4799 NULL);
4803 }
4804 }
4809 }
4810 else
4811 {
4812 /* FORNOW SLP doesn't work for multiple types. */
4815 /* For interleaved stores we created vectorized defs for all the
4816 defs stored in OPRNDS in the previous iteration (previous copy).
4817 DR_CHAIN is then used as an input to vect_permute_store_chain(),
4818 and OPRNDS as an input to vect_get_vec_def_for_stmt_copy() for the
4819 next copy.
4820 If the store is not strided, GROUP_SIZE is 1, and DR_CHAIN and
4821 OPRNDS are of size 1. */
4823 {
4829 }
4830 dataref_ptr =
4832 }
4835 {
4837 /* Permute. */
4841 }
4845 {
4847 /* Bump the vector pointer. */
4849 NULL_TREE);
4854 /* For strided stores vectorized defs are interleaved in
4855 vect_permute_store_chain(). */
4859 /* Arguments are ready. Create the new vector stmt. */
4866 else
4873 }
4874 }
4877 }
4880 /* Function vect_setup_realignment
4882 This function is called when vectorizing an unaligned load using
4883 the dr_explicit_realign[_optimized] scheme.
4884 This function generates the following code at the loop prolog:
4886 p = initial_addr;
4887 x msq_init = *(floor(p)); # prolog load
4888 realignment_token = call target_builtin;
4889 loop:
4890 x msq = phi (msq_init, ---)
4892 The stmts marked with x are generated only for the case of
4893 dr_explicit_realign_optimized.
4895 The code above sets up a new (vector) pointer, pointing to the first
4896 location accessed by STMT, and a "floor-aligned" load using that pointer.
4897 It also generates code to compute the "realignment-token" (if the relevant
4898 target hook was defined), and creates a phi-node at the loop-header bb
4899 whose arguments are the result of the prolog-load (created by this
4900 function) and the result of a load that takes place in the loop (to be
4901 created by the caller to this function).
4903 For the case of dr_explicit_realign_optimized:
4904 The caller to this function uses the phi-result (msq) to create the
4905 realignment code inside the loop, and sets up the missing phi argument,
4906 as follows:
4907 loop:
4908 msq = phi (msq_init, lsq)
4909 lsq = *(floor(p')); # load in loop
4910 result = realign_load (msq, lsq, realignment_token);
4912 For the case of dr_explicit_realign:
4913 loop:
4914 msq = *(floor(p)); # load in loop
4915 p' = p + (VS-1);
4916 lsq = *(floor(p')); # load in loop
4917 result = realign_load (msq, lsq, realignment_token);
4919 Input:
4920 STMT - (scalar) load stmt to be vectorized. This load accesses
4921 a memory location that may be unaligned.
4922 BSI - place where new code is to be inserted.
4923 ALIGNMENT_SUPPORT_SCHEME - which of the two misalignment handling schemes
4924 is used.
4926 Output:
4927 REALIGNMENT_TOKEN - the result of a call to the builtin_mask_for_load
4928 target hook, if defined.
4929 Return value - the result of the loop-header phi node. */
4931 static tree
4935 tree init_addr,
4937 {
4942 edge pe;
4944 tree vec_dest;
4945 tree inc;
4946 tree ptr;
4947 tree data_ref;
4948 tree new_stmt;
4949 basic_block new_bb;
4951 tree new_temp;
4952 tree phi_stmt;
4964 /* We need to generate three things:
4965 1. the misalignment computation
4966 2. the extra vector load (for the optimized realignment scheme).
4967 3. the phi node for the two vectors from which the realignment is
4968 done (for the optimized realignment scheme).
4969 */
4971 /* 1. Determine where to generate the misalignment computation.
4973 If INIT_ADDR is NULL_TREE, this indicates that the misalignment
4974 calculation will be generated by this function, outside the loop (in the
4975 preheader). Otherwise, INIT_ADDR had already been computed for us by the
4976 caller, inside the loop.
4978 Background: If the misalignment remains fixed throughout the iterations of
4979 the loop, then both realignment schemes are applicable, and also the
4980 misalignment computation can be done outside LOOP. This is because we are
4981 vectorizing LOOP, and so the memory accesses in LOOP advance in steps that
4982 are a multiple of VS (the Vector Size), and therefore the misalignment in
4983 different vectorized LOOP iterations is always the same.
4984 The problem arises only if the memory access is in an inner-loop nested
4985 inside LOOP, which is now being vectorized using outer-loop vectorization.
4986 This is the only case when the misalignment of the memory access may not
4987 remain fixed throughout the iterations of the inner-loop (as explained in
4988 detail in vect_supportable_dr_alignment). In this case, not only is the
4989 optimized realignment scheme not applicable, but also the misalignment
4990 computation (and generation of the realignment token that is passed to
4991 REALIGN_LOAD) have to be done inside the loop.
4993 In short, INIT_ADDR indicates whether we are in a COMPUTE_IN_LOOP mode
4994 or not, which in turn determines if the misalignment is computed inside
4995 the inner-loop, or outside LOOP. */
4998 {
5001 }
5004 /* 2. Determine where to generate the extra vector load.
5006 For the optimized realignment scheme, instead of generating two vector
5007 loads in each iteration, we generate a single extra vector load in the
5008 preheader of the loop, and in each iteration reuse the result of the
5009 vector load from the previous iteration. In case the memory access is in
5010 an inner-loop nested inside LOOP, which is now being vectorized using
5011 outer-loop vectorization, we need to determine whether this initial vector
5012 load should be generated at the preheader of the inner-loop, or can be
5013 generated at the preheader of LOOP. If the memory access has no evolution
5014 in LOOP, it can be generated in the preheader of LOOP. Otherwise, it has
5015 to be generated inside LOOP (in the preheader of the inner-loop). */
5018 {
5023 }
5024 else
5029 /* 3. For the case of the optimized realignment, create the first vector
5030 load at the loop preheader. */
5033 {
5034 /* Create msq_init = *(floor(p1)) in the loop preheader */
5048 }
5050 /* 4. Create realignment token using a target builtin, if available.
5051 It is done either inside the containing loop, or before LOOP (as
5052 determined above). */
5055 {
5056 tree builtin_decl;
5058 /* Compute INIT_ADDR - the initial addressed accessed by this memref. */
5061 else
5062 {
5063 /* Generate the INIT_ADDR computation outside LOOP. */
5069 }
5081 else
5082 {
5083 /* Generate the misalignment computation outside LOOP. */
5087 }
5091 /* The result of the CALL_EXPR to this builtin is determined from
5092 the value of the parameter and no global variables are touched
5093 which makes the builtin a "const" function. Requiring the
5094 builtin to have the "const" attribute makes it unnecessary
5095 to call mark_call_clobbered. */
5097 }
5106 /* 5. Create msq = phi <msq_init, lsq> in loop */
5116 }
5119 /* Function vect_strided_load_supported.
5121 Returns TRUE is EXTRACT_EVEN and EXTRACT_ODD operations are supported,
5122 and FALSE otherwise. */
5124 static bool
5126 {
5134 {
5138 }
5141 {
5145 }
5149 {
5153 }
5156 {
5160 }
5162 }
5165 /* Function vect_permute_load_chain.
5167 Given a chain of interleaved loads in DR_CHAIN of LENGTH that must be
5168 a power of 2, generate extract_even/odd stmts to reorder the input data
5169 correctly. Return the final references for loads in RESULT_CHAIN.
5171 E.g., LENGTH is 4 and the scalar type is short, i.e., VF is 8.
5172 The input is 4 vectors each containing 8 elements. We assign a number to each
5173 element, the input sequence is:
5175 1st vec: 0 1 2 3 4 5 6 7
5176 2nd vec: 8 9 10 11 12 13 14 15
5177 3rd vec: 16 17 18 19 20 21 22 23
5178 4th vec: 24 25 26 27 28 29 30 31
5180 The output sequence should be:
5182 1st vec: 0 4 8 12 16 20 24 28
5183 2nd vec: 1 5 9 13 17 21 25 29
5184 3rd vec: 2 6 10 14 18 22 26 30
5185 4th vec: 3 7 11 15 19 23 27 31
5187 i.e., the first output vector should contain the first elements of each
5188 interleaving group, etc.
5190 We use extract_even/odd instructions to create such output. The input of each
5191 extract_even/odd operation is two vectors
5192 1st vec 2nd vec
5193 0 1 2 3 4 5 6 7
5195 and the output is the vector of extracted even/odd elements. The output of
5196 extract_even will be: 0 2 4 6
5197 and of extract_odd: 1 3 5 7
5200 The permutation is done in log LENGTH stages. In each stage extract_even and
5201 extract_odd stmts are created for each pair of vectors in DR_CHAIN in their
5202 order. In our example,
5204 E1: extract_even (1st vec, 2nd vec)
5205 E2: extract_odd (1st vec, 2nd vec)
5206 E3: extract_even (3rd vec, 4th vec)
5207 E4: extract_odd (3rd vec, 4th vec)
5209 The output for the first stage will be:
5211 E1: 0 2 4 6 8 10 12 14
5212 E2: 1 3 5 7 9 11 13 15
5213 E3: 16 18 20 22 24 26 28 30
5214 E4: 17 19 21 23 25 27 29 31
5216 In order to proceed and create the correct sequence for the next stage (or
5217 for the correct output, if the second stage is the last one, as in our
5218 example), we first put the output of extract_even operation and then the
5219 output of extract_odd in RESULT_CHAIN (which is then copied to DR_CHAIN).
5220 The input for the second stage is:
5222 1st vec (E1): 0 2 4 6 8 10 12 14
5223 2nd vec (E3): 16 18 20 22 24 26 28 30
5224 3rd vec (E2): 1 3 5 7 9 11 13 15
5225 4th vec (E4): 17 19 21 23 25 27 29 31
5227 The output of the second stage:
5229 E1: 0 4 8 12 16 20 24 28
5230 E2: 2 6 10 14 18 22 26 30
5231 E3: 1 5 9 13 17 21 25 29
5232 E4: 3 7 11 15 19 23 27 31
5234 And RESULT_CHAIN after reordering:
5236 1st vec (E1): 0 4 8 12 16 20 24 28
5237 2nd vec (E3): 1 5 9 13 17 21 25 29
5238 3rd vec (E2): 2 6 10 14 18 22 26 30
5239 4th vec (E4): 3 7 11 15 19 23 27 31. */
5241 static bool
5244 tree stmt,
5247 {
5250 tree tmp;
5254 /* Check that the operation is supported. */
5260 {
5262 {
5266 /* data_ref = permute_even (first_data_ref, second_data_ref); */
5282 /* data_ref = permute_odd (first_data_ref, second_data_ref); */
5296 }
5298 }
5300 }
5303 /* Function vect_transform_strided_load.
5305 Given a chain of input interleaved data-refs (in DR_CHAIN), build statements
5306 to perform their permutation and ascribe the result vectorized statements to
5307 the scalar statements.
5308 */
5310 static bool
5313 {
5319 tree tmp_data_ref;
5321 /* DR_CHAIN contains input data-refs that are a part of the interleaving.
5322 RESULT_CHAIN is the output of vect_permute_load_chain, it contains permuted
5323 vectors, that are ready for vector computation. */
5325 /* Permute. */
5329 /* Put a permuted data-ref in the VECTORIZED_STMT field.
5330 Since we scan the chain starting from it's first node, their order
5331 corresponds the order of data-refs in RESULT_CHAIN. */
5335 {
5339 /* Skip the gaps. Loads created for the gaps will be removed by dead
5340 code elimination pass later.
5341 DR_GROUP_GAP is the number of steps in elements from the previous
5342 access (if there is no gap DR_GROUP_GAP is 1). We skip loads that
5343 correspond to the gaps.
5344 */
5346 {
5347 gap_count++;
5349 }
5352 {
5354 /* We assume that if VEC_STMT is not NULL, this is a case of multiple
5355 copies, and we put the new vector statement in the first available
5356 RELATED_STMT. */
5359 else
5360 {
5365 {
5368 }
5370 }
5373 /* If NEXT_STMT accesses the same DR as the previous statement,
5374 put the same TMP_DATA_REF as its vectorized statement; otherwise
5375 get the next data-ref from RESULT_CHAIN. */
5378 }
5379 }
5381 }
5384 /* vectorizable_load.
5386 Check if STMT reads a non scalar data-ref (array/pointer/structure) that
5387 can be vectorized.
5388 If VEC_STMT is also passed, vectorize the STMT: create a vectorized
5389 stmt to replace it, put it in VEC_STMT, and insert it at BSI.
5390 Return FALSE if not a vectorizable STMT, TRUE otherwise. */
5392 bool
5394 slp_tree slp_node)
5395 {
5396 tree scalar_dest;
5399 tree op;
5401 stmt_vec_info prev_stmt_info;
5408 tree new_temp;
5411 tree dummy;
5414 tree ptr_incr;
5424 tree first_stmt;
5425 tree scalar_type;
5432 /* FORNOW: SLP with multiple types is not supported. The SLP analysis verifies
5433 this, so we can safely override NCOPIES with 1 here. */
5439 /* FORNOW. This restriction should be relaxed. */
5441 {
5445 }
5453 /* Is vectorizable load? */
5473 /* FORNOW. In some cases can vectorize even if data-type not supported
5474 (e.g. - data copies). */
5476 {
5480 }
5482 /* Check if the load is a part of an interleaving chain. */
5484 {
5486 /* FORNOW */
5489 /* Check if interleaving is supported. */
5493 }
5496 {
5500 }
5505 /** Transform. **/
5508 {
5510 /* Check if the chain of loads is already vectorized. */
5512 {
5515 }
5520 /* VEC_NUM is the number of vect stmts to be created for this group. */
5522 {
5525 }
5526 else
5528 }
5529 else
5530 {
5534 }
5539 /* In case the vectorization factor (VF) is bigger than the number
5540 of elements that we can fit in a vectype (nunits), we have to generate
5541 more than one vector stmt - i.e - we need to "unroll" the
5542 vector stmt by a factor VF/nunits. In doing so, we record a pointer
5543 from one copy of the vector stmt to the next, in the field
5544 STMT_VINFO_RELATED_STMT. This is necessary in order to allow following
5545 stages to find the correct vector defs to be used when vectorizing
5546 stmts that use the defs of the current stmt. The example below illustrates
5547 the vectorization process when VF=16 and nunits=4 (i.e - we need to create
5548 4 vectorized stmts):
5550 before vectorization:
5551 RELATED_STMT VEC_STMT
5552 S1: x = memref - -
5553 S2: z = x + 1 - -
5555 step 1: vectorize stmt S1:
5556 We first create the vector stmt VS1_0, and, as usual, record a
5557 pointer to it in the STMT_VINFO_VEC_STMT of the scalar stmt S1.
5558 Next, we create the vector stmt VS1_1, and record a pointer to
5559 it in the STMT_VINFO_RELATED_STMT of the vector stmt VS1_0.
5560 Similarly, for VS1_2 and VS1_3. This is the resulting chain of
5561 stmts and pointers:
5562 RELATED_STMT VEC_STMT
5563 VS1_0: vx0 = memref0 VS1_1 -
5564 VS1_1: vx1 = memref1 VS1_2 -
5565 VS1_2: vx2 = memref2 VS1_3 -
5566 VS1_3: vx3 = memref3 - -
5567 S1: x = load - VS1_0
5568 S2: z = x + 1 - -
5570 See in documentation in vect_get_vec_def_for_stmt_copy for how the
5571 information we recorded in RELATED_STMT field is used to vectorize
5572 stmt S2. */
5574 /* In case of interleaving (non-unit strided access):
5576 S1: x2 = &base + 2
5577 S2: x0 = &base
5578 S3: x1 = &base + 1
5579 S4: x3 = &base + 3
5581 Vectorized loads are created in the order of memory accesses
5582 starting from the access of the first stmt of the chain:
5584 VS1: vx0 = &base
5585 VS2: vx1 = &base + vec_size*1
5586 VS3: vx3 = &base + vec_size*2
5587 VS4: vx4 = &base + vec_size*3
5589 Then permutation statements are generated:
5591 VS5: vx5 = VEC_EXTRACT_EVEN_EXPR < vx0, vx1 >
5592 VS6: vx6 = VEC_EXTRACT_ODD_EXPR < vx0, vx1 >
5593 ...
5595 And they are put in STMT_VINFO_VEC_STMT of the corresponding scalar stmts
5596 (the order of the data-refs in the output of vect_permute_load_chain
5597 corresponds to the order of scalar stmts in the interleaving chain - see
5598 the documentation of vect_permute_load_chain()).
5599 The generation of permutation stmts and recording them in
5600 STMT_VINFO_VEC_STMT is done in vect_transform_strided_load().
5602 In case of both multiple types and interleaving, the vector loads and
5603 permutation stmts above are created for every copy. The result vector stmts
5604 are put in STMT_VINFO_VEC_STMT for the first copy and in the corresponding
5605 STMT_VINFO_RELATED_STMT for the next copies. */
5607 /* If the data reference is aligned (dr_aligned) or potentially unaligned
5608 on a target that supports unaligned accesses (dr_unaligned_supported)
5609 we generate the following code:
5610 p = initial_addr;
5611 indx = 0;
5612 loop {
5613 p = p + indx * vectype_size;
5614 vec_dest = *(p);
5615 indx = indx + 1;
5616 }
5618 Otherwise, the data reference is potentially unaligned on a target that
5619 does not support unaligned accesses (dr_explicit_realign_optimized) -
5620 then generate the following code, in which the data in each iteration is
5621 obtained by two vector loads, one from the previous iteration, and one
5622 from the current iteration:
5623 p1 = initial_addr;
5624 msq_init = *(floor(p1))
5625 p2 = initial_addr + VS - 1;
5626 realignment_token = call target_builtin;
5627 indx = 0;
5628 loop {
5629 p2 = p2 + indx * vectype_size
5630 lsq = *(floor(p2))
5631 vec_dest = realign_load (msq, lsq, realignment_token)
5632 indx = indx + 1;
5633 msq = lsq;
5634 } */
5636 /* If the misalignment remains the same throughout the execution of the
5637 loop, we can create the init_addr and permutation mask at the loop
5638 preheader. Otherwise, it needs to be created inside the loop.
5639 This can only occur when vectorizing memory accesses in the inner-loop
5640 nested within an outer-loop that is being vectorized. */
5644 {
5647 }
5652 {
5657 {
5660 }
5661 }
5662 else
5667 {
5668 /* 1. Create the vector pointer update chain. */
5674 else
5675 dataref_ptr =
5679 {
5682 NULL_TREE);
5684 /* 2. Create the vector-load in the loop. */
5686 {
5692 {
5697 data_ref =
5700 }
5702 {
5709 dr_explicit_realign,
5727 }
5733 }
5741 /* 3. Handle explicit realignment if necessary/supported. Create in
5742 loop: vec_dest = realign_load (msq, lsq, realignment_token) */
5745 {
5751 realignment_token);
5758 {
5762 }
5763 }
5765 /* 4. Handle invariant-load. */
5767 {
5771 {
5776 /* CHECKME: bitpos depends on endianess? */
5782 vec_dest =
5791 /* FIXME: use build_constructor directly. */
5795 }
5796 else
5798 }
5800 /* Collect vector loads and later create their permutation in
5801 vect_transform_strided_load (). */
5805 /* Store vector loads in the corresponding SLP_NODE. */
5808 }
5810 /* FORNOW: SLP with multiple types is unsupported. */
5815 {
5820 }
5821 else
5822 {
5825 else
5828 }
5829 }
5832 }
5835 /* Function vectorizable_live_operation.
5837 STMT computes a value that is used outside the loop. Check if
5838 it can be supported. */
5840 bool
5844 {
5845 tree operation;
5851 tree op;
5866 /* FORNOW. CHECKME. */
5873 /* FORNOW: support only if all uses are invariant. This means
5874 that the scalar operations can remain in place, unvectorized.
5875 The original last scalar value that they compute will be used. */
5878 {
5881 {
5885 }
5889 }
5891 /* No transformation is required for the cases we currently support. */
5893 }
5896 /* Function vect_is_simple_cond.
5898 Input:
5899 LOOP - the loop that is being vectorized.
5900 COND - Condition that is checked for simple use.
5902 Returns whether a COND can be vectorized. Checks whether
5903 condition operands are supportable using vec_is_simple_use. */
5905 static bool
5907 {
5909 tree def;
5919 {
5923 }
5929 {
5933 }
5939 }
5941 /* vectorizable_condition.
5943 Check if STMT is conditional modify expression that can be vectorized.
5944 If VEC_STMT is also passed, vectorize the STMT: create a vectorized
5945 stmt using VEC_COND_EXPR to replace it, put it in VEC_STMT, and insert it
5946 at BSI.
5948 Return FALSE if not a vectorizable STMT, TRUE otherwise. */
5950 bool
5952 {
5961 tree new_temp;
5964 tree def;
5979 /* FORNOW: SLP not supported. */
5983 /* FORNOW: not yet supported. */
5985 {
5989 }
5991 /* Is vectorizable conditional operation? */
6007 /* We do not handle two different vector types for the condition
6008 and the values. */
6013 {
6018 }
6025 {
6030 }
6040 {
6043 }
6045 /* Transform */
6047 /* Handle def. */
6051 /* Handle cond expr. */
6052 vec_cond_lhs =
6054 vec_cond_rhs =
6059 /* Arguments are ready. create the new vector stmt. */
6071 }
6074 /* Function vect_transform_stmt.
6076 Create a vectorized stmt to replace STMT, and insert it at BSI. */
6078 static bool
6080 slp_tree slp_node)
6081 {
6085 tree orig_stmt_in_pattern;
6089 {
6132 {
6133 /* In case of interleaving, the whole chain is vectorized when the
6134 last store in the chain is reached. Store stmts before the last
6135 one are skipped, and there vec_stmt_info shouldn't be freed
6136 meanwhile. */
6140 }
6141 else
6164 {
6168 }
6169 }
6173 {
6176 }
6179 {
6183 {
6185 /* STMT was inserted by the vectorizer to replace a computation idiom.
6186 ORIG_STMT_IN_PATTERN is a stmt in the original sequence that
6187 computed this idiom. We need to record a pointer to VEC_STMT in
6188 the stmt_info of ORIG_STMT_IN_PATTERN. See more details in the
6189 documentation of vect_pattern_recog. */
6191 {
6194 }
6195 }
6196 }
6199 }
6202 /* This function builds ni_name = number of iterations loop executes
6203 on the loop preheader. */
6205 static tree
6207 {
6209 edge pe;
6219 {
6222 }
6225 }
6228 /* This function generates the following statements:
6230 ni_name = number of iterations loop executes
6231 ratio = ni_name / vf
6232 ratio_mult_vf_name = ratio * vf
6234 and places them at the loop preheader edge. */
6236 static void
6241 {
6243 edge pe;
6244 basic_block new_bb;
6246 tree var;
6247 tree ratio_name;
6248 tree ratio_mult_vf_name;
6252 tree log_vf;
6256 /* Generate temporary variable that contains
6257 number of iterations loop executes. */
6262 /* Create: ratio = ni >> log2(vf) */
6266 {
6274 }
6276 /* Create: ratio_mult_vf = ratio << log2 (vf). */
6281 {
6290 }
6297 }
6300 /* Function vect_update_ivs_after_vectorizer.
6302 "Advance" the induction variables of LOOP to the value they should take
6303 after the execution of LOOP. This is currently necessary because the
6304 vectorizer does not handle induction variables that are used after the
6305 loop. Such a situation occurs when the last iterations of LOOP are
6306 peeled, because:
6307 1. We introduced new uses after LOOP for IVs that were not originally used
6308 after LOOP: the IVs of LOOP are now used by an epilog loop.
6309 2. LOOP is going to be vectorized; this means that it will iterate N/VF
6310 times, whereas the loop IVs should be bumped N times.
6312 Input:
6313 - LOOP - a loop that is going to be vectorized. The last few iterations
6314 of LOOP were peeled.
6315 - NITERS - the number of iterations that LOOP executes (before it is
6316 vectorized). i.e, the number of times the ivs should be bumped.
6317 - UPDATE_E - a successor edge of LOOP->exit that is on the (only) path
6318 coming out from LOOP on which there are uses of the LOOP ivs
6319 (this is the path from LOOP->exit to epilog_loop->preheader).
6321 The new definitions of the ivs are placed in LOOP->exit.
6322 The phi args associated with the edge UPDATE_E in the bb
6323 UPDATE_E->dest are updated accordingly.
6325 Assumption 1: Like the rest of the vectorizer, this function assumes
6326 a single loop exit that has a single predecessor.
6328 Assumption 2: The phi nodes in the LOOP header and in update_bb are
6329 organized in the same order.
6331 Assumption 3: The access function of the ivs is simple enough (see
6332 vect_can_advance_ivs_p). This assumption will be relaxed in the future.
6334 Assumption 4: Exactly one of the successors of LOOP exit-bb is on a path
6335 coming out of LOOP on which the ivs of LOOP are used (this is the path
6336 that leads to the epilog loop; other paths skip the epilog loop). This
6337 path starts with the edge UPDATE_E, and its destination (denoted update_bb)
6338 needs to have its phis updated.
6339 */
6341 static void
6343 edge update_e)
6344 {
6350 /* gcc_assert (vect_can_advance_ivs_p (loop_vinfo)); */
6352 /* Make sure there exists a single-predecessor exit bb: */
6358 {
6360 tree evolution_part;
6361 tree init_expr;
6362 tree step_expr;
6364 block_stmt_iterator last_bsi;
6367 {
6370 }
6372 /* Skip virtual phi's. */
6374 {
6378 }
6380 /* Skip reduction phis. */
6382 {
6386 }
6390 evolution_part =
6394 /* FORNOW: We do not support IVs whose evolution function is a polynomial
6395 of degree >= 2 or exponential. */
6404 init_expr,
6408 else
6412 niters),
6413 step_expr),
6414 init_expr);
6425 /* Fix phi expressions in the successor bb. */
6427 }
6428 }
6431 /* Function vect_do_peeling_for_loop_bound
6433 Peel the last iterations of the loop represented by LOOP_VINFO.
6434 The peeled iterations form a new epilog loop. Given that the loop now
6435 iterates NITERS times, the new epilog loop iterates
6436 NITERS % VECTORIZATION_FACTOR times.
6438 The original loop will later be made to iterate
6439 NITERS / VECTORIZATION_FACTOR times (this value is placed into RATIO). */
6441 static void
6443 {
6447 edge update_e;
6448 basic_block preheader;
6459 /* Generate the following variables on the preheader of original loop:
6461 ni_name = number of iteration the original loop executes
6462 ratio = ni_name / vf
6463 ratio_mult_vf_name = ratio * vf */
6469 /* Analyze cost to set threshhold for vectorized loop. */
6474 /* Use the cost model only if it is more conservative than user specified
6475 threshold. */
6479 && (!min_scalar_loop_bound
6490 th);
6493 #ifdef ENABLE_CHECKING
6495 #endif
6497 /* A guard that controls whether the new_loop is to be executed or skipped
6498 is placed in LOOP->exit. LOOP->exit therefore has two successors - one
6499 is the preheader of NEW_LOOP, where the IVs from LOOP are used. The other
6500 is a bb after NEW_LOOP, where these IVs are not used. Find the edge that
6501 is on the path where the LOOP IVs are used and need to be updated. */
6506 else
6509 /* Update IVs of original loop as if they were advanced
6510 by ratio_mult_vf_name steps. */
6513 /* After peeling we have to reset scalar evolution analyzer. */
6517 }
6520 /* Function vect_gen_niters_for_prolog_loop
6522 Set the number of iterations for the loop represented by LOOP_VINFO
6523 to the minimum between LOOP_NITERS (the original iteration count of the loop)
6524 and the misalignment of DR - the data reference recorded in
6525 LOOP_VINFO_UNALIGNED_DR (LOOP_VINFO). As a result, after the execution of
6526 this loop, the data reference DR will refer to an aligned location.
6528 The following computation is generated:
6530 If the misalignment of DR is known at compile time:
6531 addr_mis = int mis = DR_MISALIGNMENT (dr);
6532 Else, compute address misalignment in bytes:
6533 addr_mis = addr & (vectype_size - 1)
6535 prolog_niters = min ( LOOP_NITERS , (VF - addr_mis/elem_size)&(VF-1) )
6537 (elem_size = element type size; an element is the scalar element
6538 whose type is the inner type of the vectype)
6540 For interleaving,
6542 prolog_niters = min ( LOOP_NITERS ,
6543 (VF/group_size - addr_mis/elem_size)&(VF/group_size-1) )
6544 where group_size is the size of the interleaved group.
6546 The above formulas assume that VF == number of elements in the vector. This
6547 may not hold when there are multiple-types in the loop.
6548 In this case, for some data-references in the loop the VF does not represent
6549 the number of elements that fit in the vector. Therefore, instead of VF we
6550 use TYPE_VECTOR_SUBPARTS. */
6552 static tree
6554 {
6559 edge pe;
6560 basic_block new_bb;
6571 {
6572 /* For interleaved access element size must be multiplied by the size of
6573 the interleaved group. */
6577 }
6582 {
6590 }
6591 else
6592 {
6600 tree elem_size_log =
6604 tree byte_misalign;
6605 tree elem_misalign;
6610 /* Create: byte_misalign = addr & (vectype_size - 1) */
6611 byte_misalign =
6614 /* Create: elem_misalign = byte_misalign / element_size */
6615 elem_misalign =
6618 /* Create: (niters_type) (nelements - elem_misalign)&(nelements - 1) */
6622 }
6624 /* Create: prolog_loop_niters = min (iters, loop_niters) */
6625 /* If the loop bound is known at compile time we already verified that it is
6626 greater than vf; since the misalignment ('iters') is at most vf, there's
6627 no need to generate the MIN_EXPR in this case. */
6632 {
6635 }
6641 /* Insert stmt on loop preheader edge. */
6643 {
6646 }
6649 }
6652 /* Function vect_update_init_of_dr
6654 NITERS iterations were peeled from LOOP. DR represents a data reference
6655 in LOOP. This function updates the information recorded in DR to
6656 account for the fact that the first NITERS iterations had already been
6657 executed. Specifically, it updates the OFFSET field of DR. */
6659 static void
6661 {
6667 }
6670 /* Function vect_update_inits_of_drs
6672 NITERS iterations were peeled from the loop represented by LOOP_VINFO.
6673 This function updates the information recorded for the data references in
6674 the loop to account for the fact that the first NITERS iterations had
6675 already been executed. Specifically, it updates the initial_condition of
6676 the access_function of all the data_references in the loop. */
6678 static void
6680 {
6690 }
6693 /* Function vect_do_peeling_for_alignment
6695 Peel the first 'niters' iterations of the loop represented by LOOP_VINFO.
6696 'niters' is set to the misalignment of one of the data references in the
6697 loop, thereby forcing it to refer to an aligned location at the beginning
6698 of the execution of this loop. The data reference for which we are
6699 peeling is recorded in LOOP_VINFO_UNALIGNED_DR. */
6701 static void
6703 {
6706 tree n_iters;
6717 /* Peel the prolog loop and iterate it niters_of_prolog_loop. */
6718 new_loop =
6722 #ifdef ENABLE_CHECKING
6724 #endif
6726 /* Update number of times loop executes. */
6731 /* Update the init conditions of the access functions of all data refs. */
6734 /* After peeling we have to reset scalar evolution analyzer. */
6738 }
6741 /* Function vect_create_cond_for_align_checks.
6743 Create a conditional expression that represents the alignment checks for
6744 all of data references (array element references) whose alignment must be
6745 checked at runtime.
6747 Input:
6748 LOOP_VINFO - two fields of the loop information are used.
6749 LOOP_VINFO_PTR_MASK is the mask used to check the alignment.
6750 LOOP_VINFO_MAY_MISALIGN_STMTS contains the refs to be checked.
6752 Output:
6753 COND_EXPR_STMT_LIST - statements needed to construct the conditional
6754 expression.
6755 The returned value is the conditional expression to be used in the if
6756 statement that controls which version of the loop gets executed at runtime.
6758 The algorithm makes two assumptions:
6759 1) The number of bytes "n" in a vector is a power of 2.
6760 2) An address "a" is aligned if a%n is zero and that this
6761 test can be done as a&(n-1) == 0. For example, for 16
6762 byte vectors the test is a&0xf == 0. */
6764 static tree
6767 {
6773 tree mask_cst;
6775 tree psize;
6776 tree int_ptrsize_type;
6780 tree ptrsize_zero;
6782 /* Check that mask is one less than a power of 2, i.e., mask is
6783 all zeros followed by all ones. */
6786 /* CHECKME: what is the best integer or unsigned type to use to hold a
6787 cast from a pointer value? */
6789 int_ptrsize_type
6792 /* Create expression (mask & (dr_1 || ... || dr_n)) where dr_i is the address
6793 of the first vector of the i'th data reference. */
6796 {
6798 tree addr_base;
6802 /* create: addr_tmp = (int)(address_of_first_vector) */
6818 /* The addresses are OR together. */
6821 {
6822 /* create: or_tmp = or_tmp | addr_tmp */
6833 }
6834 else
6841 /* create: and_tmp = or_tmp & mask */
6851 /* Make and_tmp the left operand of the conditional test against zero.
6852 if and_tmp has a nonzero bit then some address is unaligned. */
6856 }
6858 /* Function vect_vfa_segment_size.
6860 Create an expression that computes the size of segment
6861 that will be accessed for a data reference. The functions takes into
6862 account that realignment loads may access one more vector.
6864 Input:
6865 DR: The data reference.
6866 VECT_FACTOR: vectorization factor.
6868 Return an expression whose value is the size of segment which will be
6869 accessed by DR. */
6871 static tree
6873 {
6878 {
6879 tree vector_size = TYPE_SIZE_UNIT
6884 }
6886 }
6888 /* Function vect_create_cond_for_alias_checks.
6890 Create a conditional expression that represents the run-time checks for
6891 overlapping of address ranges represented by a list of data references
6892 relations passed as input.
6894 Input:
6895 COND_EXPR - input conditional expression. New conditions will be chained
6896 with logical and operation.
6897 LOOP_VINFO - field LOOP_VINFO_MAY_ALIAS_STMTS contains the list of ddrs
6898 to be checked.
6900 Output:
6901 COND_EXPR - conditional expression.
6902 COND_EXPR_STMT_LIST - statements needed to construct the conditional
6903 expression.
6906 The returned value is the conditional expression to be used in the if
6907 statement that controls which version of the loop gets executed at runtime.
6908 */
6910 static void
6914 {
6918 tree vect_factor =
6921 ddr_p ddr;
6923 tree part_cond_expr;
6925 /* Create expression
6926 ((store_ptr_0 + store_segment_length_0) < load_ptr_0)
6927 || (load_ptr_0 + load_segment_length_0) < store_ptr_0))
6928 &&
6929 ...
6930 &&
6931 ((store_ptr_n + store_segment_length_n) < load_ptr_n)
6932 || (load_ptr_n + load_segment_length_n) < store_ptr_n)) */
6938 {
6949 {
6952 }
6958 {
6961 }
6963 addr_base_a =
6966 addr_base_b =
6974 {
6980 }
6983 part_cond_expr =
6987 addr_base_a,
6988 segment_length_a),
6989 addr_base_b),
6992 addr_base_b,
6993 segment_length_b),
6994 addr_base_a));
6999 else
7001 }
7006 }
7008 /* Function vect_loop_versioning.
7010 If the loop has data references that may or may not be aligned or/and
7011 has data reference relations whose independence was not proven then
7012 two versions of the loop need to be generated, one which is vectorized
7013 and one which isn't. A test is then generated to control which of the
7014 loops is executed. The test checks for the alignment of all of the
7015 data references that may or may not be aligned. An additional
7016 sequence of runtime tests is generated for each pairs of DDRs whose
7017 independence was not proven. The vectorized version of loop is
7018 executed only if both alias and alignment tests are passed. */
7020 static void
7022 {
7027 basic_block condition_bb;
7028 block_stmt_iterator cond_exp_bsi;
7029 basic_block merge_bb;
7030 basic_block new_exit_bb;
7034 tree gimplify_stmt_list;
7041 cond_expr =
7047 cond_expr =
7049 cond_expr =
7051 NULL_TREE);
7059 /* Loop versioning violates an assumption we try to maintain during
7060 vectorization - that the loop exit block has a single predecessor.
7061 After versioning, the exit block of both loop versions is the same
7062 basic block (i.e. it has two predecessors). Just in order to simplify
7063 following transformations in the vectorizer, we fix this situation
7064 here by adding a new (empty) block on the exit-edge of the loop,
7065 with the proper loop-exit phis to maintain loop-closed-form. */
7075 {
7077 new_exit_bb);
7081 }
7083 /* End loop-exit-fixes after versioning. */
7087 {
7090 }
7091 }
7093 /* Remove a group of stores (for SLP or interleaving), free their
7094 stmt_vec_info. */
7096 static void
7098 {
7099 stmt_ann_t ann;
7101 tree tmp;
7102 stmt_vec_info next_stmt_info;
7103 block_stmt_iterator next_si;
7106 {
7107 /* Free the attached stmt_vec_info and remove the stmt. */
7116 }
7117 }
7120 /* Vectorize SLP instance tree in postorder. */
7122 static bool
7124 {
7125 tree stmt;
7127 block_stmt_iterator si;
7128 stmt_vec_info stmt_info;
7142 {
7145 }
7150 {
7152 /* If IS_STORE is TRUE, the vectorization of the
7153 interleaving chain was completed - free all the stores in
7154 the chain. */
7156 else
7157 /* FORNOW: SLP originates only from strided stores. */
7161 }
7163 /* FORNOW: SLP originates only from strided stores. */
7165 }
7168 static bool
7170 {
7173 slp_instance instance;
7180 {
7182 /* For each SLP instance calculate number of vector stmts to be created
7183 for the scalar stmts in each node of the SLP tree. Number of vector
7184 elements in one vector iteration is the number of scalar elements in
7185 one scalar iteration (GROUP_SIZE) multiplied by VF divided by vector
7186 size. */
7189 /* Schedule the tree of INSTANCE. */
7191 vec_stmts_size);
7196 }
7199 }
7201 /* Function vect_transform_loop.
7203 The analysis phase has determined that the loop is vectorizable.
7204 Vectorize the loop - created vectorized stmts to replace the scalar
7205 stmts in the loop, and update the loop exit condition. */
7207 void
7209 {
7225 /* CHECKME: we wouldn't need this if we called update_ssa once
7226 for all loops. */
7229 /* Peel the loop if there are data refs with unknown alignment.
7230 Only one data ref with unknown store is allowed. */
7235 /* If the loop has a symbolic number of iterations 'n' (i.e. it's not a
7236 compile time constant), or it is a constant that doesn't divide by the
7237 vectorization factor, then an epilog loop needs to be created.
7238 We therefore duplicate the loop: the original loop will be vectorized,
7239 and will compute the first (n/VF) iterations. The second copy of the loop
7240 will remain scalar and will compute the remaining (n%VF) iterations.
7241 (VF is the vectorization factor). */
7247 else
7251 /* 1) Make sure the loop header has exactly two entries
7252 2) Make sure we have a preheader basic block. */
7258 /* FORNOW: the vectorizer supports only loops which body consist
7259 of one basic block (header + empty latch). When the vectorizer will
7260 support more involved loop forms, the order by which the BBs are
7261 traversed need to be reconsidered. */
7264 {
7266 stmt_vec_info stmt_info;
7267 tree phi;
7270 {
7272 {
7275 }
7290 {
7294 }
7295 }
7298 {
7303 {
7306 }
7310 /* vector stmts created in the outer-loop during vectorization of
7311 stmts in an inner-loop may not have a stmt_info, and do not
7312 need to be vectorized. */
7314 {
7317 }
7321 {
7324 }
7327 nunits =
7332 /* For SLP VF is set according to unrolling factor, and not to
7333 vector size, hence for SLP this print is not valid. */
7336 /* SLP. Schedule all the SLP instances when the first SLP stmt is
7337 reached. */
7339 {
7341 {
7349 /* IS_STORE is true if STMT is a store. Stores cannot be of
7350 hybrid SLP type. They are removed in
7351 vect_schedule_slp_instance and their vinfo is destroyed. */
7353 {
7356 }
7357 }
7359 /* Hybrid SLP stmts must be vectorized in addition to SLP. */
7361 {
7364 }
7365 }
7367 /* -------- vectorize statement ------------ */
7374 {
7375 stmt_ann_t ann;
7377 {
7378 /* Interleaving. If IS_STORE is TRUE, the vectorization of the
7379 interleaving chain was completed - free all the stores in
7380 the chain. */
7382 tree tmp;
7383 stmt_vec_info next_stmt_info;
7386 {
7389 /* Free the attached stmt_vec_info and remove the stmt. */
7396 }
7399 }
7400 else
7401 {
7402 /* Free the attached stmt_vec_info and remove the stmt. */
7408 }
7409 }
7418 /* The memory tags and pointers in vectorized statements need to
7419 have their SSA forms updated. FIXME, why can't this be delayed
7420 until all the loops have been transformed? */
7427 }