| blob | 59fa6d5de8078ffeb6a5d7fd72b451d9d2cb5b4f |
1 /* Transformation Utilities for Loop Vectorization.
2 Copyright (C) 2003, 2004, 2005, 2006, 2007, 2008 Free Software Foundation, Inc.
3 Contributed by Dorit Naishlos <dorit@il.ibm.com>
5 This file is part of GCC.
7 GCC is free software; you can redistribute it and/or modify it under
8 the terms of the GNU General Public License as published by the Free
9 Software Foundation; either version 3, or (at your option) any later
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 {
132 slp_instance instance;
134 /* Cost model disabled. */
136 {
140 }
142 /* If the number of iterations is unknown, or the
143 peeling-for-misalignment amount is unknown, we will have to generate
144 a runtime test to test the loop count against the threshold. */
149 /* Requires loop versioning tests to handle misalignment. */
152 {
153 /* FIXME: Make cost depend on complexity of individual check. */
154 vec_outside_cost +=
159 }
162 {
163 /* FIXME: Make cost depend on complexity of individual check. */
164 vec_outside_cost +=
169 }
173 {
175 }
177 /* Count statements in scalar loop. Using this as scalar cost for a single
178 iteration for now.
180 TODO: Add outer loop support.
182 TODO: Consider assigning different costs to different scalar
183 statements. */
185 /* FORNOW. */
190 {
191 block_stmt_iterator si;
196 else
200 {
203 /* Skip stmts that are not vectorized inside the loop. */
210 /* FIXME: for stmts in the inner-loop in outer-loop vectorization,
211 some of the "outside" costs are generated inside the outer-loop. */
213 }
214 }
216 /* Add additional cost for the peeled instructions in prologue and epilogue
217 loop.
219 FORNOW: If we dont know the value of peel_iters for prologue or epilogue
220 at compile-time - we assume it's vf/2 (the worst would be vf-1).
222 TODO: Build an expression that represents peel_iters for prologue and
223 epilogue to be used in a run-time test. */
226 {
232 /* If peeling for alignment is unknown, loop bound of main loop becomes
233 unknown. */
237 "epilogue peel iters set to vf/2 because "
240 /* If peeled iterations are unknown, count a taken branch and a not taken
241 branch per peeled loop. Even if scalar loop iterations are known,
242 vector iterations are not known since peeled prologue iterations are
243 not known. Hence guards remain the same. */
247 }
248 else
249 {
251 {
258 }
259 else
263 {
267 "epilogue peel iters set to vf/2 because "
270 /* If peeled iterations are known but number of scalar loop
271 iterations are unknown, count a taken branch per peeled loop. */
274 }
275 else
276 {
281 }
282 }
288 /* FORNOW: The scalar outside cost is incremented in one of the
289 following ways:
291 1. The vectorizer checks for alignment and aliasing and generates
292 a condition that allows dynamic vectorization. A cost model
293 check is ANDED with the versioning condition. Hence scalar code
294 path now has the added cost of the versioning check.
296 if (cost > th & versioning_check)
297 jmp to vector code
299 Hence run-time scalar is incremented by not-taken branch cost.
301 2. The vectorizer then checks if a prologue is required. If the
302 cost model check was not done before during versioning, it has to
303 be done before the prologue check.
305 if (cost <= th)
306 prologue = scalar_iters
307 if (prologue == 0)
308 jmp to vector code
309 else
310 execute prologue
311 if (prologue == num_iters)
312 go to exit
314 Hence the run-time scalar cost is incremented by a taken branch,
315 plus a not-taken branch, plus a taken branch cost.
317 3. The vectorizer then checks if an epilogue is required. If the
318 cost model check was not done before during prologue check, it
319 has to be done with the epilogue check.
321 if (prologue == 0)
322 jmp to vector code
323 else
324 execute prologue
325 if (prologue == num_iters)
326 go to exit
327 vector code:
328 if ((cost <= th) | (scalar_iters-prologue-epilogue == 0))
329 jmp to epilogue
331 Hence the run-time scalar cost should be incremented by 2 taken
332 branches.
334 TODO: The back end may reorder the BBS's differently and reverse
335 conditions/branch directions. Change the stimates below to
336 something more reasonable. */
339 {
340 /* Cost model check occurs at versioning. */
344 else
345 {
346 /* Cost model occurs at prologue generation. */
350 /* Cost model check occurs at epilogue generation. */
351 else
353 }
354 }
356 /* Add SLP costs. */
359 {
362 }
364 /* Calculate number of iterations required to make the vector version
365 profitable, relative to the loop bodies only. The following condition
366 must hold true:
367 SIC * niters + SOC > VIC * ((niters-PL_ITERS-EP_ITERS)/VF) + VOC
368 where
369 SIC = scalar iteration cost, VIC = vector iteration cost,
370 VOC = vector outside cost, VF = vectorization factor,
371 PL_ITERS = prologue iterations, EP_ITERS= epilogue iterations
372 SOC = scalar outside cost for run time cost model check. */
375 {
378 else
379 {
389 min_profitable_iters++;
390 }
391 }
392 /* vector version will never be profitable. */
393 else
394 {
397 "is divisible by scalar iteration cost = %d by a factor "
401 }
404 {
407 vec_inside_cost);
409 vec_outside_cost);
411 scalar_single_iter_cost);
414 peel_iters_prologue);
416 peel_iters_epilogue);
418 min_profitable_iters);
419 }
421 min_profitable_iters =
424 /* Because the condition we create is:
425 if (niters <= min_profitable_iters)
426 then skip the vectorized loop. */
427 min_profitable_iters--;
431 min_profitable_iters);
434 }
437 /* TODO: Close dependency between vect_model_*_cost and vectorizable_*
438 functions. Design better to avoid maintenance issues. */
440 /* Function vect_model_reduction_cost.
442 Models cost for a reduction operation, including the vector ops
443 generated within the strip-mine loop, the initial definition before
444 the loop, and the epilogue code that must be generated. */
446 static bool
449 {
452 optab optab;
453 tree vectype;
454 tree orig_stmt;
455 tree reduction_op;
462 /* Cost of reduction op inside loop. */
468 {
470 {
473 }
475 }
485 /* Add in cost for initial definition. */
488 /* Determine cost of epilogue code.
490 We have a reduction operator that will reduce the vector in one statement.
491 Also requires scalar extract. */
494 {
497 else
498 {
500 tree bitsize =
507 /* We have a whole vector shift available. */
511 /* Final reduction via vector shifts and the reduction operator. Also
512 requires scalar extract. */
515 else
516 /* Use extracts and reduction op for final reduction. For N elements,
517 we have N extracts and N-1 reduction ops. */
519 }
520 }
530 }
533 /* Function vect_model_induction_cost.
535 Models cost for induction operations. */
537 static void
539 {
540 /* loop cost for vec_loop. */
542 /* prologue cost for vec_init and vec_step. */
549 }
552 /* Function vect_model_simple_cost.
554 Models cost for simple operations, i.e. those that only emit ncopies of a
555 single op. Right now, this does not account for multiple insns that could
556 be generated for the single vector op. We will handle that shortly. */
558 void
561 {
567 /* FORNOW: Assuming maximum 2 args per stmts. */
569 {
572 }
578 /* Set the costs either in STMT_INFO or SLP_NODE (if exists). */
581 }
584 /* Function vect_cost_strided_group_size
586 For strided load or store, return the group_size only if it is the first
587 load or store of a group, else return 1. This ensures that group size is
588 only returned once per group. */
590 static int
592 {
599 }
602 /* Function vect_model_store_cost
604 Models cost for stores. In the case of strided accesses, one access
605 has the overhead of the strided access attributed to it. */
607 void
610 {
617 /* Strided access? */
620 /* Not a strided access. */
621 else
624 /* Is this an access in a group of stores, which provide strided access?
625 If so, add in the cost of the permutes. */
627 {
628 /* Uses a high and low interleave operation for each needed permute. */
634 group_size);
636 }
638 /* Costs of the stores. */
645 /* Set the costs either in STMT_INFO or SLP_NODE (if exists). */
648 }
651 /* Function vect_model_load_cost
653 Models cost for loads. In the case of strided accesses, the last access
654 has the overhead of the strided access attributed to it. Since unaligned
655 accesses are supported for loads, we also account for the costs of the
656 access scheme chosen. */
658 void
661 {
664 tree first_stmt;
668 /* Strided accesses? */
671 {
674 }
675 /* Not a strided access. */
676 else
677 {
680 }
684 /* Is this an access in a group of loads providing strided access?
685 If so, add in the cost of the permutes. */
687 {
688 /* Uses an even and odd extract operations for each needed permute. */
694 group_size);
696 }
698 /* The loads themselves. */
700 {
702 {
709 }
711 {
712 /* Here, we assign an additional cost for the unaligned load. */
720 }
722 {
725 /* FIXME: If the misalignment remains fixed across the iterations of
726 the containing loop, the following cost should be added to the
727 outside costs. */
732 }
734 {
739 /* Unaligned software pipeline has a load of an address, an initial
740 load, and possibly a mask operation to "prime" the loop. However,
741 if this is an access in a group of loads, which provide strided
742 access, then the above cost should only be considered for one
743 access in the group. Inside the loop, there is a load op
744 and a realignment op. */
747 {
751 }
756 }
760 }
766 /* Set the costs either in STMT_INFO or SLP_NODE (if exists). */
769 }
772 /* Function vect_get_new_vect_var.
774 Returns a name for a new variable. The current naming scheme appends the
775 prefix "vect_" or "vect_p" (depending on the value of VAR_KIND) to
776 the name of vectorizer generated variables, and appends that to NAME if
777 provided. */
779 static tree
781 {
783 tree new_vect_var;
786 {
798 }
801 {
805 }
806 else
809 /* Mark vector typed variable as a gimple register variable. */
814 }
817 /* Function vect_create_addr_base_for_vector_ref.
819 Create an expression that computes the address of the first memory location
820 that will be accessed for a data reference.
822 Input:
823 STMT: The statement containing the data reference.
824 NEW_STMT_LIST: Must be initialized to NULL_TREE or a statement list.
825 OFFSET: Optional. If supplied, it is be added to the initial address.
826 LOOP: Specify relative to which loop-nest should the address be computed.
827 For example, when the dataref is in an inner-loop nested in an
828 outer-loop that is now being vectorized, LOOP can be either the
829 outer-loop, or the inner-loop. The first memory location accessed
830 by the following dataref ('in' points to short):
832 for (i=0; i<N; i++)
833 for (j=0; j<M; j++)
834 s += in[i+j]
836 is as follows:
837 if LOOP=i_loop: &in (relative to i_loop)
838 if LOOP=j_loop: &in+i*2B (relative to j_loop)
840 Output:
841 1. Return an SSA_NAME whose value is the address of the memory location of
842 the first vector of the data reference.
843 2. If new_stmt_list is not NULL_TREE after return then the caller must insert
844 these statement(s) which define the returned SSA_NAME.
846 FORNOW: We are only handling array accesses with step 1. */
848 static tree
851 tree offset,
853 {
858 tree base_name;
859 tree data_ref_base_var;
860 tree new_base_stmt;
861 tree vec_stmt;
871 {
880 }
882 /* Create data_ref_base */
890 /* Create base_offset */
899 {
908 }
910 /* base + base_offset */
916 /* addr_expr = addr_base */
928 {
931 }
933 }
936 /* Function vect_create_data_ref_ptr.
938 Create a new pointer to vector type (vp), that points to the first location
939 accessed in the loop by STMT, along with the def-use update chain to
940 appropriately advance the pointer through the loop iterations. Also set
941 aliasing information for the pointer. This vector pointer is used by the
942 callers to this function to create a memory reference expression for vector
943 load/store access.
945 Input:
946 1. STMT: a stmt that references memory. Expected to be of the form
947 GIMPLE_MODIFY_STMT <name, data-ref> or
948 GIMPLE_MODIFY_STMT <data-ref, name>.
949 2. AT_LOOP: the loop where the vector memref is to be created.
950 3. OFFSET (optional): an offset to be added to the initial address accessed
951 by the data-ref in STMT.
952 4. ONLY_INIT: indicate if vp is to be updated in the loop, or remain
953 pointing to the initial address.
954 5. TYPE: if not NULL indicates the required type of the data-ref
956 Output:
957 1. Declare a new ptr to vector_type, and have it point to the base of the
958 data reference (initial addressed accessed by the data reference).
959 For example, for vector of type V8HI, the following code is generated:
961 v8hi *vp;
962 vp = (v8hi *)initial_address;
964 if OFFSET is not supplied:
965 initial_address = &a[init];
966 if OFFSET is supplied:
967 initial_address = &a[init + OFFSET];
969 Return the initial_address in INITIAL_ADDRESS.
971 2. If ONLY_INIT is true, just return the initial pointer. Otherwise, also
972 update the pointer in each iteration of the loop.
974 Return the increment stmt that updates the pointer in PTR_INCR.
976 3. Set INV_P to true if the access pattern of the data reference in the
977 vectorized loop is invariant. Set it to false otherwise.
979 4. Return the pointer. */
981 static tree
985 {
986 tree base_name;
993 tree vect_ptr_type;
994 tree vect_ptr;
995 tree tag;
996 tree new_temp;
997 tree vec_stmt;
999 edge pe;
1000 basic_block new_bb;
1001 tree vect_ptr_init;
1003 tree vptr;
1004 block_stmt_iterator incr_bsi;
1007 tree incr;
1008 tree step;
1010 /* Check the step (evolution) of the load in LOOP, and record
1011 whether it's invariant. */
1014 else
1019 else
1022 /* Create an expression for the first address accessed by this load
1023 in LOOP. */
1027 {
1040 }
1042 /** (1) Create the new vector-pointer variable: **/
1045 else
1051 /** (2) Add aliasing information to the new vector-pointer:
1052 (The points-to info (DR_PTR_INFO) may be defined later.) **/
1057 /* If tag is a variable (and NOT_A_TAG) than a new symbol memory
1058 tag must be created with tag added to its may alias list. */
1061 else
1066 /** Note: If the dataref is in an inner-loop nested in LOOP, and we are
1067 vectorizing LOOP (i.e. outer-loop vectorization), we need to create two
1068 def-use update cycles for the pointer: One relative to the outer-loop
1069 (LOOP), which is what steps (3) and (4) below do. The other is relative
1070 to the inner-loop (which is the inner-most loop containing the dataref),
1071 and this is done be step (5) below.
1073 When vectorizing inner-most loops, the vectorized loop (LOOP) is also the
1074 inner-most loop, and so steps (3),(4) work the same, and step (5) is
1075 redundant. Steps (3),(4) create the following:
1077 vp0 = &base_addr;
1078 LOOP: vp1 = phi(vp0,vp2)
1079 ...
1080 ...
1081 vp2 = vp1 + step
1082 goto LOOP
1084 If there is an inner-loop nested in loop, then step (5) will also be
1085 applied, and an additional update in the inner-loop will be created:
1087 vp0 = &base_addr;
1088 LOOP: vp1 = phi(vp0,vp2)
1089 ...
1090 inner: vp3 = phi(vp1,vp4)
1091 vp4 = vp3 + inner_step
1092 if () goto inner
1093 ...
1094 vp2 = vp1 + step
1095 if () goto LOOP */
1097 /** (3) Calculate the initial address the vector-pointer, and set
1098 the vector-pointer to point to it before the loop: **/
1100 /* Create: (&(base[init_val+offset]) in the loop preheader. */
1109 /* Create: p = (vectype *) initial_base */
1118 /** (4) Handle the updating of the vector-pointer inside the loop.
1119 This is needed when ONLY_INIT is false, and also when AT_LOOP
1120 is the inner-loop nested in LOOP (during outer-loop vectorization).
1121 **/
1124 {
1125 /* Copy the points-to information if it exists. */
1129 }
1130 else
1131 {
1132 /* The step of the vector pointer is the Vector Size. */
1134 /* One exception to the above is when the scalar step of the load in
1135 LOOP is zero. In this case the step here is also zero. */
1149 /* Copy the points-to information if it exists. */
1151 {
1154 }
1161 }
1167 /** (5) Handle the updating of the vector-pointer inside the inner-loop
1168 nested in LOOP, if exists: **/
1172 {
1181 /* Copy the points-to information if it exists. */
1183 {
1186 }
1193 }
1194 else
1196 }
1199 /* Function bump_vector_ptr
1201 Increment a pointer (to a vector type) by vector-size. If requested,
1202 i.e. if PTR-INCR is given, then also connect the new increment stmt
1203 to the existing def-use update-chain of the pointer, by modifying
1204 the PTR_INCR as illustrated below:
1206 The pointer def-use update-chain before this function:
1207 DATAREF_PTR = phi (p_0, p_2)
1208 ....
1209 PTR_INCR: p_2 = DATAREF_PTR + step
1211 The pointer def-use update-chain after this function:
1212 DATAREF_PTR = phi (p_0, p_2)
1213 ....
1214 NEW_DATAREF_PTR = DATAREF_PTR + BUMP
1215 ....
1216 PTR_INCR: p_2 = NEW_DATAREF_PTR + step
1218 Input:
1219 DATAREF_PTR - ssa_name of a pointer (to vector type) that is being updated
1220 in the loop.
1221 PTR_INCR - optional. The stmt that updates the pointer in each iteration of
1222 the loop. The increment amount across iterations is expected
1223 to be vector_size.
1224 BSI - location where the new update stmt is to be placed.
1225 STMT - the original scalar memory-access stmt that is being vectorized.
1226 BUMP - optional. The offset by which to bump the pointer. If not given,
1227 the offset is assumed to be vector_size.
1229 Output: Return NEW_DATAREF_PTR as illustrated above.
1231 */
1233 static tree
1236 {
1243 tree incr_stmt;
1244 ssa_op_iter iter;
1245 use_operand_p use_p;
1246 tree new_dataref_ptr;
1258 /* Copy the points-to information if it exists. */
1266 /* Update the vector-pointer's cross-iteration increment. */
1268 {
1273 else
1275 }
1278 }
1281 /* Function vect_create_destination_var.
1283 Create a new temporary of type VECTYPE. */
1285 static tree
1287 {
1288 tree vec_dest;
1290 tree type;
1305 }
1308 /* Function vect_init_vector.
1310 Insert a new stmt (INIT_STMT) that initializes a new vector variable with
1311 the vector elements of VECTOR_VAR. Place the initialization at BSI if it
1312 is not NULL. Otherwise, place the initialization at the loop preheader.
1313 Return the DEF of INIT_STMT.
1314 It will be used in the vectorization of STMT. */
1316 static tree
1319 {
1321 tree new_var;
1322 tree init_stmt;
1323 tree vec_oprnd;
1324 edge pe;
1325 tree new_temp;
1326 basic_block new_bb;
1336 else
1337 {
1346 }
1349 {
1352 }
1356 }
1359 /* For constant and loop invariant defs of SLP_NODE this function returns
1360 (vector) defs (VEC_OPRNDS) that will be used in the vectorized stmts.
1361 OP_NUM determines if we gather defs for operand 0 or operand 1 of the scalar
1362 stmts. */
1364 static void
1367 {
1373 tree vec_cst;
1376 tree vector_type;
1389 /* NUMBER_OF_COPIES is the number of times we need to use the same values in
1390 created vectors. It is greater than 1 if unrolling is performed.
1392 For example, we have two scalar operands, s1 and s2 (e.g., group of
1393 strided accesses of size two), while NUINTS is four (i.e., four scalars
1394 of this type can be packed in a vector). The output vector will contain
1395 two copies of each scalar operand: {s1, s2, s1, s2}. (NUMBER_OF_COPIES
1396 will be 2).
1398 If GROUP_SIZE > NUNITS, the scalars will be split into several vectors
1399 containing the operands.
1401 For example, NUINTS is four as before, and the group size is 8
1402 (s1, s2, ..., s8). We will create two vectors {s1, s2, s3, s4} and
1403 {s5, s6, s7, s8}. */
1410 {
1412 {
1416 else
1421 /* Create 'vect_ = {op0,op1,...,opn}'. */
1424 number_of_places_left_in_vector--;
1427 {
1434 else
1439 NULL));
1441 }
1442 }
1443 }
1445 /* Since the vectors are created in the reverse order, we should invert
1446 them. */
1449 {
1452 }
1456 /* In case that VF is greater than the unrolling factor needed for the SLP
1457 group of stmts, NUMBER_OF_VECTORS to be created is greater than
1458 NUMBER_OF_SCALARS/NUNITS or NUNITS/NUMBER_OF_SCALARS, and hence we have
1459 to replicate the vectors. */
1461 {
1464 }
1465 }
1468 /* Get vectorized definitions from SLP_NODE that contains corresponding
1469 vectorized def-stmts. */
1471 static void
1473 {
1474 tree vec_oprnd;
1475 tree vec_def_stmt;
1482 i++)
1483 {
1487 }
1488 }
1491 /* Get vectorized definitions for SLP_NODE.
1492 If the scalar definitions are loop invariants or constants, collect them and
1493 call vect_get_constant_vectors() to create vector stmts.
1494 Otherwise, the def-stmts must be already vectorized and the vectorized stmts
1495 must be stored in the LEFT/RIGHT node of SLP_NODE, and we call
1496 vect_get_slp_vect_defs() to retrieve them.
1497 If VEC_OPRNDS1 is NULL, don't get vector defs for the second operand (from
1498 the right node. This is used when the second operand must remain scalar. */
1500 static void
1503 {
1506 /* Allocate memory for vectorized defs. */
1510 /* SLP_NODE corresponds either to a group of stores or to a group of
1511 unary/binary operations. We don't call this function for loads. */
1513 /* The defs are already vectorized. */
1515 else
1516 /* Build vectors from scalar defs. */
1521 /* Since we don't call this function with loads, this is a group of
1522 stores. */
1533 /* The defs are already vectorized. */
1535 else
1536 /* Build vectors from scalar defs. */
1538 }
1541 /* Function get_initial_def_for_induction
1543 Input:
1544 STMT - a stmt that performs an induction operation in the loop.
1545 IV_PHI - the initial value of the induction variable
1547 Output:
1548 Return a vector variable, initialized with the first VF values of
1549 the induction variable. E.g., for an iv with IV_PHI='X' and
1550 evolution S, for a vector of 4 units, we want to return:
1551 [X, X + S, X + 2*S, X + 3*S]. */
1553 static tree
1555 {
1560 tree vectype;
1564 basic_block new_bb;
1566 tree access_fn;
1567 tree new_var;
1568 tree new_name;
1569 tree init_stmt;
1576 tree expr;
1579 tree stmts;
1580 imm_use_iterator imm_iter;
1581 use_operand_p use_p;
1582 tree exit_phi;
1583 edge latch_e;
1584 tree loop_arg;
1585 block_stmt_iterator si;
1596 /* Find the first insertion point in the BB. */
1601 else
1604 /* Is phi in an inner-loop, while vectorizing an enclosing outer-loop? */
1606 {
1609 }
1610 else
1624 /* Create the vector that holds the initial_value of the induction. */
1626 {
1627 /* iv_loop is nested in the loop to be vectorized. init_expr had already
1628 been created during vectorization of previous stmts; We obtain it from
1629 the STMT_VINFO_VEC_STMT of the defining stmt. */
1632 }
1633 else
1634 {
1635 /* iv_loop is the loop to be vectorized. Create:
1636 vec_init = [X, X+S, X+2*S, X+3*S] (S = step_expr, X = init_expr) */
1642 {
1645 }
1650 {
1651 tree tmp;
1653 /* Create: new_name_i = new_name + step_expr */
1663 {
1666 }
1668 }
1669 /* Create a vector from [new_name_0, new_name_1, ..., new_name_nunits-1] */
1672 }
1675 /* Create the vector that holds the step of the induction. */
1677 /* iv_loop is nested in the loop to be vectorized. Generate:
1678 vec_step = [S, S, S, S] */
1680 else
1681 {
1682 /* iv_loop is the loop to be vectorized. Generate:
1683 vec_step = [VF*S, VF*S, VF*S, VF*S] */
1686 }
1696 /* Create the following def-use cycle:
1697 loop prolog:
1698 vec_init = ...
1699 vec_step = ...
1700 loop:
1701 vec_iv = PHI <vec_init, vec_loop>
1702 ...
1703 STMT
1704 ...
1705 vec_loop = vec_iv + vec_step; */
1707 /* Create the induction-phi that defines the induction-operand. */
1715 /* Create the iv update inside the loop */
1725 /* Set the arguments of the phi node: */
1730 /* In case that vectorization factor (VF) is bigger than the number
1731 of elements that we can fit in a vectype (nunits), we have to generate
1732 more than one vector stmt - i.e - we need to "unroll" the
1733 vector stmt by a factor VF/nunits. For more details see documentation
1734 in vectorizable_operation. */
1737 {
1738 stmt_vec_info prev_stmt_vinfo;
1739 /* FORNOW. This restriction should be relaxed. */
1742 /* Create the vector that holds the step of the induction. */
1755 {
1756 tree tmp;
1758 /* vec_i = vec_prev + vec_step */
1768 }
1769 }
1772 {
1773 /* Find the loop-closed exit-phi of the induction, and record
1774 the final vector of induction results: */
1777 {
1779 {
1782 }
1783 }
1785 {
1787 /* FORNOW. Currently not supporting the case that an inner-loop induction
1788 is not used in the outer-loop (i.e. only outside the outer-loop). */
1794 {
1797 }
1798 }
1799 }
1803 {
1808 }
1812 }
1815 /* Function vect_get_vec_def_for_operand.
1817 OP is an operand in STMT. This function returns a (vector) def that will be
1818 used in the vectorized stmt for STMT.
1820 In the case that OP is an SSA_NAME which is defined in the loop, then
1821 STMT_VINFO_VEC_STMT of the defining stmt holds the relevant def.
1823 In case OP is an invariant or constant, a new stmt that creates a vector def
1824 needs to be introduced. */
1826 static tree
1828 {
1829 tree vec_oprnd;
1830 tree vec_stmt;
1831 tree def_stmt;
1837 tree vec_inv;
1838 tree vec_cst;
1840 tree def;
1844 tree vector_type;
1847 {
1850 }
1855 {
1857 {
1860 }
1862 {
1865 }
1866 }
1869 {
1870 /* Case 1: operand is a constant. */
1872 {
1876 /* Create 'vect_cst_ = {cst,cst,...,cst}' */
1881 {
1883 }
1889 }
1891 /* Case 2: operand is defined outside the loop - loop invariant. */
1893 {
1897 /* Create 'vec_inv = {inv,inv,..,inv}' */
1902 {
1904 }
1906 /* FIXME: use build_constructor directly. */
1911 }
1913 /* Case 3: operand is defined inside the loop. */
1915 {
1919 /* Get the def from the vectorized stmt. */
1925 else
1928 }
1930 /* Case 4: operand is defined by a loop header phi - reduction */
1932 {
1938 /* Get the def before the loop */
1941 }
1943 /* Case 5: operand is defined by loop-header phi - induction. */
1945 {
1948 /* Get the def from the vectorized stmt. */
1954 }
1958 }
1959 }
1962 /* Function vect_get_vec_def_for_stmt_copy
1964 Return a vector-def for an operand. This function is used when the
1965 vectorized stmt to be created (by the caller to this function) is a "copy"
1966 created in case the vectorized result cannot fit in one vector, and several
1967 copies of the vector-stmt are required. In this case the vector-def is
1968 retrieved from the vector stmt recorded in the STMT_VINFO_RELATED_STMT field
1969 of the stmt that defines VEC_OPRND.
1970 DT is the type of the vector def VEC_OPRND.
1972 Context:
1973 In case the vectorization factor (VF) is bigger than the number
1974 of elements that can fit in a vectype (nunits), we have to generate
1975 more than one vector stmt to vectorize the scalar stmt. This situation
1976 arises when there are multiple data-types operated upon in the loop; the
1977 smallest data-type determines the VF, and as a result, when vectorizing
1978 stmts operating on wider types we need to create 'VF/nunits' "copies" of the
1979 vector stmt (each computing a vector of 'nunits' results, and together
1980 computing 'VF' results in each iteration). This function is called when
1981 vectorizing such a stmt (e.g. vectorizing S2 in the illustration below, in
1982 which VF=16 and nunits=4, so the number of copies required is 4):
1984 scalar stmt: vectorized into: STMT_VINFO_RELATED_STMT
1986 S1: x = load VS1.0: vx.0 = memref0 VS1.1
1987 VS1.1: vx.1 = memref1 VS1.2
1988 VS1.2: vx.2 = memref2 VS1.3
1989 VS1.3: vx.3 = memref3
1991 S2: z = x + ... VSnew.0: vz0 = vx.0 + ... VSnew.1
1992 VSnew.1: vz1 = vx.1 + ... VSnew.2
1993 VSnew.2: vz2 = vx.2 + ... VSnew.3
1994 VSnew.3: vz3 = vx.3 + ...
1996 The vectorization of S1 is explained in vectorizable_load.
1997 The vectorization of S2:
1998 To create the first vector-stmt out of the 4 copies - VSnew.0 -
1999 the function 'vect_get_vec_def_for_operand' is called to
2000 get the relevant vector-def for each operand of S2. For operand x it
2001 returns the vector-def 'vx.0'.
2003 To create the remaining copies of the vector-stmt (VSnew.j), this
2004 function is called to get the relevant vector-def for each operand. It is
2005 obtained from the respective VS1.j stmt, which is recorded in the
2006 STMT_VINFO_RELATED_STMT field of the stmt that defines VEC_OPRND.
2008 For example, to obtain the vector-def 'vx.1' in order to create the
2009 vector stmt 'VSnew.1', this function is called with VEC_OPRND='vx.0'.
2010 Given 'vx0' we obtain the stmt that defines it ('VS1.0'); from the
2011 STMT_VINFO_RELATED_STMT field of 'VS1.0' we obtain the next copy - 'VS1.1',
2012 and return its def ('vx.1').
2013 Overall, to create the above sequence this function will be called 3 times:
2014 vx.1 = vect_get_vec_def_for_stmt_copy (dt, vx.0);
2015 vx.2 = vect_get_vec_def_for_stmt_copy (dt, vx.1);
2016 vx.3 = vect_get_vec_def_for_stmt_copy (dt, vx.2); */
2018 static tree
2020 {
2021 tree vec_stmt_for_operand;
2022 stmt_vec_info def_stmt_info;
2024 /* Do nothing; can reuse same def. */
2035 }
2038 /* Get vectorized definitions for the operands to create a copy of an original
2039 stmt. See vect_get_vec_def_for_stmt_copy() for details. */
2041 static void
2045 {
2052 {
2056 }
2057 }
2060 /* Get vectorized definitions for OP0 and OP1, or SLP_NODE if it is not NULL. */
2062 static void
2065 {
2068 else
2069 {
2070 tree vec_oprnd;
2077 {
2081 }
2082 }
2083 }
2086 /* Function vect_finish_stmt_generation.
2088 Insert a new stmt. */
2090 static void
2093 {
2106 {
2109 }
2111 /* Make sure bsi points to the stmt that is being vectorized. */
2115 }
2118 /* Function get_initial_def_for_reduction
2120 Input:
2121 STMT - a stmt that performs a reduction operation in the loop.
2122 INIT_VAL - the initial value of the reduction variable
2124 Output:
2125 ADJUSTMENT_DEF - a tree that holds a value to be added to the final result
2126 of the reduction (used for adjusting the epilog - see below).
2127 Return a vector variable, initialized according to the operation that STMT
2128 performs. This vector will be used as the initial value of the
2129 vector of partial results.
2131 Option1 (adjust in epilog): Initialize the vector as follows:
2132 add: [0,0,...,0,0]
2133 mult: [1,1,...,1,1]
2134 min/max: [init_val,init_val,..,init_val,init_val]
2135 bit and/or: [init_val,init_val,..,init_val,init_val]
2136 and when necessary (e.g. add/mult case) let the caller know
2137 that it needs to adjust the result by init_val.
2139 Option2: Initialize the vector as follows:
2140 add: [0,0,...,0,init_val]
2141 mult: [1,1,...,1,init_val]
2142 min/max: [init_val,init_val,...,init_val]
2143 bit and/or: [init_val,init_val,...,init_val]
2144 and no adjustments are needed.
2146 For example, for the following code:
2148 s = init_val;
2149 for (i=0;i<n;i++)
2150 s = s + a[i];
2152 STMT is 's = s + a[i]', and the reduction variable is 's'.
2153 For a vector of 4 units, we want to return either [0,0,0,init_val],
2154 or [0,0,0,0] and let the caller know that it needs to adjust
2155 the result at the end by 'init_val'.
2157 FORNOW, we are using the 'adjust in epilog' scheme, because this way the
2158 initialization vector is simpler (same element in all entries).
2159 A cost model should help decide between these two schemes. */
2161 static tree
2163 {
2171 tree vecdef;
2172 tree def_for_init;
2173 tree init_def;
2176 tree vector_type;
2182 else
2188 {
2194 else
2196 /* Create a vector of zeros for init_def. */
2199 else
2216 }
2219 }
2222 /* Function vect_create_epilog_for_reduction
2224 Create code at the loop-epilog to finalize the result of a reduction
2225 computation.
2227 VECT_DEF is a vector of partial results.
2228 REDUC_CODE is the tree-code for the epilog reduction.
2229 STMT is the scalar reduction stmt that is being vectorized.
2230 REDUCTION_PHI is the phi-node that carries the reduction computation.
2232 This function:
2233 1. Creates the reduction def-use cycle: sets the arguments for
2234 REDUCTION_PHI:
2235 The loop-entry argument is the vectorized initial-value of the reduction.
2236 The loop-latch argument is VECT_DEF - the vector of partial sums.
2237 2. "Reduces" the vector of partial results VECT_DEF into a single result,
2238 by applying the operation specified by REDUC_CODE if available, or by
2239 other means (whole-vector shifts or a scalar loop).
2240 The function also creates a new phi node at the loop exit to preserve
2241 loop-closed form, as illustrated below.
2243 The flow at the entry to this function:
2245 loop:
2246 vec_def = phi <null, null> # REDUCTION_PHI
2247 VECT_DEF = vector_stmt # vectorized form of STMT
2248 s_loop = scalar_stmt # (scalar) STMT
2249 loop_exit:
2250 s_out0 = phi <s_loop> # (scalar) EXIT_PHI
2251 use <s_out0>
2252 use <s_out0>
2254 The above is transformed by this function into:
2256 loop:
2257 vec_def = phi <vec_init, VECT_DEF> # REDUCTION_PHI
2258 VECT_DEF = vector_stmt # vectorized form of STMT
2259 s_loop = scalar_stmt # (scalar) STMT
2260 loop_exit:
2261 s_out0 = phi <s_loop> # (scalar) EXIT_PHI
2262 v_out1 = phi <VECT_DEF> # NEW_EXIT_PHI
2263 v_out2 = reduce <v_out1>
2264 s_out3 = extract_field <v_out2, 0>
2265 s_out4 = adjust_result <s_out3>
2266 use <s_out4>
2267 use <s_out4>
2268 */
2270 static void
2273 {
2275 tree vectype;
2279 basic_block exit_bb;
2280 tree scalar_dest;
2281 tree scalar_type;
2282 tree new_phi;
2283 block_stmt_iterator exit_bsi;
2284 tree vec_dest;
2286 tree new_name;
2291 tree adjustment_def;
2292 tree vec_initial_def;
2293 tree orig_name;
2294 imm_use_iterator imm_iter;
2295 use_operand_p use_p;
2298 tree orig_stmt;
2299 tree use_stmt;
2307 {
2310 }
2318 /*** 1. Create the reduction def-use cycle ***/
2320 /* 1.1 set the loop-entry arg of the reduction-phi: */
2321 /* For the case of reduction, vect_get_vec_def_for_operand returns
2322 the scalar def before the loop, that defines the initial value
2323 of the reduction variable. */
2328 /* 1.2 set the loop-latch arg for the reduction-phi: */
2332 {
2337 }
2340 /*** 2. Create epilog code
2341 The reduction epilog code operates across the elements of the vector
2342 of partial results computed by the vectorized loop.
2343 The reduction epilog code consists of:
2344 step 1: compute the scalar result in a vector (v_out2)
2345 step 2: extract the scalar result (s_out3) from the vector (v_out2)
2346 step 3: adjust the scalar result (s_out3) if needed.
2348 Step 1 can be accomplished using one the following three schemes:
2349 (scheme 1) using reduc_code, if available.
2350 (scheme 2) using whole-vector shifts, if available.
2351 (scheme 3) using a scalar loop. In this case steps 1+2 above are
2352 combined.
2354 The overall epilog code looks like this:
2356 s_out0 = phi <s_loop> # original EXIT_PHI
2357 v_out1 = phi <VECT_DEF> # NEW_EXIT_PHI
2358 v_out2 = reduce <v_out1> # step 1
2359 s_out3 = extract_field <v_out2, 0> # step 2
2360 s_out4 = adjust_result <s_out3> # step 3
2362 (step 3 is optional, and step2 1 and 2 may be combined).
2363 Lastly, the uses of s_out0 are replaced by s_out4.
2365 ***/
2367 /* 2.1 Create new loop-exit-phi to preserve loop-closed form:
2368 v_out1 = phi <v_loop> */
2375 /* 2.2 Get the relevant tree-code to use in the epilog for schemes 2,3
2376 (i.e. when reduc_code is not available) and in the final adjustment
2377 code (if needed). Also get the original scalar reduction variable as
2378 defined in the loop. In case STMT is a "pattern-stmt" (i.e. - it
2379 represents a reduction pattern), the tree-code and scalar-def are
2380 taken from the original stmt that the pattern-stmt (STMT) replaces.
2381 Otherwise (it is a regular reduction) - the tree-code and scalar-def
2382 are taken from STMT. */
2386 {
2387 /* Regular reduction */
2389 }
2390 else
2391 {
2392 /* Reduction pattern */
2396 }
2405 /* In case this is a reduction in an inner-loop while vectorizing an outer
2406 loop - we don't need to extract a single scalar result at the end of the
2407 inner-loop. The final vector of partial results will be used in the
2408 vectorized outer-loop, or reduced to a scalar result at the end of the
2409 outer-loop. */
2413 /* 2.3 Create the reduction code, using one of the three schemes described
2414 above. */
2417 {
2418 tree tmp;
2420 /*** Case 1: Create:
2421 v_out2 = reduc_expr <v_out1> */
2434 }
2435 else
2436 {
2442 tree vec_temp;
2446 else
2449 /* Regardless of whether we have a whole vector shift, if we're
2450 emulating the operation via tree-vect-generic, we don't want
2451 to use it. Only the first round of the reduction is likely
2452 to still be profitable via emulation. */
2453 /* ??? It might be better to emit a reduction tree code here, so that
2454 tree-vect-generic can expand the first round via bit tricks. */
2457 else
2458 {
2462 }
2465 {
2466 /*** Case 2: Create:
2467 for (offset = VS/2; offset >= element_size; offset/=2)
2468 {
2469 Create: va' = vec_shift <va, offset>
2470 Create: va = vop <va, va'>
2471 } */
2482 {
2495 }
2498 }
2499 else
2500 {
2501 tree rhs;
2503 /*** Case 3: Create:
2504 s = extract_field <v_out2, 0>
2505 for (offset = element_size;
2506 offset < vector_size;
2507 offset += element_size;)
2508 {
2509 Create: s' = extract_field <v_out2, offset>
2510 Create: s = op <s, s'>
2511 } */
2519 bitsize_zero_node);
2528 {
2529 tree tmp;
2532 bitpos);
2544 }
2547 }
2548 }
2550 /* 2.4 Extract the final scalar result. Create:
2551 s_out3 = extract_field <v_out2, bitpos> */
2554 {
2555 tree rhs;
2565 else
2573 }
2575 vect_finalize_reduction:
2577 /* 2.5 Adjust the final result by the initial value of the reduction
2578 variable. (When such adjustment is not needed, then
2579 'adjustment_def' is zero). For example, if code is PLUS we create:
2580 new_temp = loop_exit_def + adjustment_def */
2583 {
2585 {
2589 }
2590 else
2591 {
2595 }
2600 }
2603 /* 2.6 Handle the loop-exit phi */
2605 /* Replace uses of s_out0 with uses of s_out3:
2606 Find the loop-closed-use at the loop exit of the original scalar result.
2607 (The reduction result is expected to have two immediate uses - one at the
2608 latch block, and one at the loop exit). */
2611 {
2613 {
2616 }
2617 }
2618 /* We expect to have found an exit_phi because of loop-closed-ssa form. */
2622 {
2624 {
2627 /* FORNOW. Currently not supporting the case that an inner-loop reduction
2628 is not used in the outer-loop (but only outside the outer-loop). */
2637 }
2639 /* Replace the uses: */
2644 }
2646 }
2649 /* Function vectorizable_reduction.
2651 Check if STMT performs a reduction operation that can be vectorized.
2652 If VEC_STMT is also passed, vectorize the STMT: create a vectorized
2653 stmt to replace it, put it in VEC_STMT, and insert it at BSI.
2654 Return FALSE if not a vectorizable STMT, TRUE otherwise.
2656 This function also handles reduction idioms (patterns) that have been
2657 recognized in advance during vect_pattern_recog. In this case, STMT may be
2658 of this form:
2659 X = pattern_expr (arg0, arg1, ..., X)
2660 and it's STMT_VINFO_RELATED_STMT points to the last stmt in the original
2661 sequence that had been detected and replaced by the pattern-stmt (STMT).
2663 In some cases of reduction patterns, the type of the reduction variable X is
2664 different than the type of the other arguments of STMT.
2665 In such cases, the vectype that is used when transforming STMT into a vector
2666 stmt is different than the vectype that is used to determine the
2667 vectorization factor, because it consists of a different number of elements
2668 than the actual number of elements that are being operated upon in parallel.
2670 For example, consider an accumulation of shorts into an int accumulator.
2671 On some targets it's possible to vectorize this pattern operating on 8
2672 shorts at a time (hence, the vectype for purposes of determining the
2673 vectorization factor should be V8HI); on the other hand, the vectype that
2674 is used to create the vector form is actually V4SI (the type of the result).
2676 Upon entry to this function, STMT_VINFO_VECTYPE records the vectype that
2677 indicates what is the actual level of parallelism (V8HI in the example), so
2678 that the right vectorization factor would be derived. This vectype
2679 corresponds to the type of arguments to the reduction stmt, and should *NOT*
2680 be used to create the vectorized stmt. The right vectype for the vectorized
2681 stmt is obtained from the type of the result X:
2682 get_vectype_for_scalar_type (TREE_TYPE (X))
2684 This means that, contrary to "regular" reductions (or "regular" stmts in
2685 general), the following equation:
2686 STMT_VINFO_VECTYPE == get_vectype_for_scalar_type (TREE_TYPE (X))
2687 does *NOT* necessarily hold for reduction patterns. */
2689 bool
2691 {
2692 tree vec_dest;
2693 tree scalar_dest;
2694 tree op;
2700 tree operation;
2708 tree new_phi;
2709 tree scalar_type;
2711 tree orig_stmt;
2712 stmt_vec_info orig_stmt_info;
2717 stmt_vec_info prev_stmt_info;
2718 tree reduc_def;
2723 {
2725 /* FORNOW. This restriction should be relaxed. */
2727 {
2731 }
2732 }
2736 /* FORNOW: SLP not supported. */
2740 /* 1. Is vectorizable reduction? */
2742 /* Not supportable if the reduction variable is used in the loop. */
2746 /* Reductions that are not used even in an enclosing outer-loop,
2747 are expected to be "live" (used out of the loop). */
2752 /* Make sure it was already recognized as a reduction computation. */
2756 /* 2. Has this been recognized as a reduction pattern?
2758 Check if STMT represents a pattern that has been recognized
2759 in earlier analysis stages. For stmts that represent a pattern,
2760 the STMT_VINFO_RELATED_STMT field records the last stmt in
2761 the original sequence that constitutes the pattern. */
2765 {
2770 }
2772 /* 3. Check the operands of the operation. The first operands are defined
2773 inside the loop body. The last operand is the reduction variable,
2774 which is defined by the loop-header-phi. */
2789 /* All uses but the last are expected to be defined in the loop.
2790 The last use is the reduction variable. */
2792 {
2801 }
2810 else
2816 /* 4. Supportable by target? */
2818 /* 4.1. check support for the operation in the loop */
2821 {
2825 }
2828 {
2837 }
2839 /* Worthwhile without SIMD support? */
2843 {
2847 }
2849 /* 4.2. Check support for the epilog operation.
2851 If STMT represents a reduction pattern, then the type of the
2852 reduction variable may be different than the type of the rest
2853 of the arguments. For example, consider the case of accumulation
2854 of shorts into an int accumulator; The original code:
2855 S1: int_a = (int) short_a;
2856 orig_stmt-> S2: int_acc = plus <int_a ,int_acc>;
2858 was replaced with:
2859 STMT: int_acc = widen_sum <short_a, int_acc>
2861 This means that:
2862 1. The tree-code that is used to create the vector operation in the
2863 epilog code (that reduces the partial results) is not the
2864 tree-code of STMT, but is rather the tree-code of the original
2865 stmt from the pattern that STMT is replacing. I.e, in the example
2866 above we want to use 'widen_sum' in the loop, but 'plus' in the
2867 epilog.
2868 2. The type (mode) we use to check available target support
2869 for the vector operation to be created in the *epilog*, is
2870 determined by the type of the reduction variable (in the example
2871 above we'd check this: plus_optab[vect_int_mode]).
2872 However the type (mode) we use to check available target support
2873 for the vector operation to be created *inside the loop*, is
2874 determined by the type of the other arguments to STMT (in the
2875 example we'd check this: widen_sum_optab[vect_short_mode]).
2877 This is contrary to "regular" reductions, in which the types of all
2878 the arguments are the same as the type of the reduction variable.
2879 For "regular" reductions we can therefore use the same vector type
2880 (and also the same tree-code) when generating the epilog code and
2881 when generating the code inside the loop. */
2884 {
2885 /* This is a reduction pattern: get the vectype from the type of the
2886 reduction variable, and get the tree-code from orig_stmt. */
2890 {
2892 {
2895 }
2897 }
2900 }
2901 else
2902 {
2903 /* Regular reduction: use the same vectype and tree-code as used for
2904 the vector code inside the loop can be used for the epilog code. */
2906 }
2912 {
2916 }
2918 {
2922 }
2925 {
2930 }
2932 /** Transform. **/
2937 /* Create the destination vector */
2940 /* Create the reduction-phi that defines the reduction-operand. */
2943 /* In case the vectorization factor (VF) is bigger than the number
2944 of elements that we can fit in a vectype (nunits), we have to generate
2945 more than one vector stmt - i.e - we need to "unroll" the
2946 vector stmt by a factor VF/nunits. For more details see documentation
2947 in vectorizable_operation. */
2951 {
2952 /* Handle uses. */
2954 {
2958 {
2961 }
2963 /* Get the vector def for the reduction variable from the phi node */
2965 }
2966 else
2967 {
2973 /* Get the vector def for the reduction variable from the vectorized
2974 reduction operation generated in the previous iteration (j-1) */
2976 }
2978 /* Arguments are ready. create the new vector stmt. */
2981 else
2983 reduc_def);
2991 else
2994 }
2996 /* Finalize the reduction-phi (set it's arguments) and create the
2997 epilog reduction code. */
3000 }
3002 /* Checks if CALL can be vectorized in type VECTYPE. Returns
3003 a function declaration if the target has a vectorized version
3004 of the function, or NULL_TREE if the function cannot be vectorized. */
3006 tree
3008 {
3012 /* We only handle functions that do not read or clobber memory -- i.e.
3013 const or novops ones. */
3024 vectype_in);
3025 }
3027 /* Function vectorizable_call.
3029 Check if STMT performs a function call that can be vectorized.
3030 If VEC_STMT is also passed, vectorize the STMT: create a vectorized
3031 stmt to replace it, put it in VEC_STMT, and insert it at BSI.
3032 Return FALSE if not a vectorizable STMT, TRUE otherwise. */
3034 bool
3036 {
3037 tree vec_dest;
3038 tree scalar_dest;
3039 tree operation;
3050 tree new_stmt;
3052 call_expr_arg_iterator iter;
3053 tree vargs;
3062 /* FORNOW: SLP not supported. */
3066 /* Is STMT a vectorizable call? */
3077 /* Process function arguments. */
3081 {
3082 /* Bail out if the function has more than two arguments, we
3083 do not have interesting builtin functions to vectorize with
3084 more than two arguments. */
3088 /* We can only handle calls with arguments of the same type. */
3091 {
3095 }
3099 {
3103 }
3106 }
3108 /* No arguments is also not good. */
3123 /* FORNOW */
3130 else
3133 /* For now, we only vectorize functions if a target specific builtin
3134 is available. TODO -- in some cases, it might be profitable to
3135 insert the calls for pieces of the vector, in order to be able
3136 to vectorize other operations in the loop. */
3139 {
3144 }
3150 else
3153 /* Sanity check: make sure that at least one copy of the vectorized stmt
3154 needs to be generated. */
3157 /* FORNOW. This restriction should be relaxed. */
3159 {
3163 }
3166 {
3172 }
3174 /** Transform. **/
3179 /* FORNOW. This restriction should be relaxed. */
3181 {
3185 }
3187 /* Handle def. */
3193 {
3196 {
3197 /* Build argument list for the vectorized call. */
3198 /* FIXME: Rewrite this so that it doesn't
3199 construct a temporary list. */
3203 {
3205 vec_oprnd0
3207 else
3208 vec_oprnd0
3214 }
3226 else
3230 }
3236 {
3237 /* Build argument list for the vectorized call. */
3238 /* FIXME: Rewrite this so that it doesn't
3239 construct a temporary list. */
3243 {
3245 {
3246 vec_oprnd0
3248 vec_oprnd1
3250 }
3251 else
3252 {
3253 vec_oprnd0
3255 vec_oprnd1
3257 }
3263 }
3275 else
3279 }
3286 /* No current target implements this case. */
3288 }
3290 /* The call in STMT might prevent it from being removed in dce.
3291 We however cannot remove it here, due to the way the ssa name
3292 it defines is mapped to the new definition. So just replace
3293 rhs of the statement with something harmless. */
3299 }
3302 /* Function vect_gen_widened_results_half
3304 Create a vector stmt whose code, type, number of arguments, and result
3305 variable are CODE, VECTYPE, OP_TYPE, and VEC_DEST, and its arguments are
3306 VEC_OPRND0 and VEC_OPRND1. The new vector stmt is to be inserted at BSI.
3307 In the case that CODE is a CALL_EXPR, this means that a call to DECL
3308 needs to be created (DECL is a function-decl of a target-builtin).
3309 STMT is the original scalar stmt that we are vectorizing. */
3311 static tree
3315 tree stmt)
3316 {
3317 tree expr;
3318 tree new_stmt;
3319 tree new_temp;
3320 tree sym;
3321 ssa_op_iter iter;
3323 /* Generate half of the widened result: */
3325 {
3326 /* Target specific support */
3329 else
3331 }
3332 else
3333 {
3334 /* Generic support */
3338 else
3340 }
3347 {
3349 {
3353 }
3354 }
3357 }
3360 /* Check if STMT performs a conversion operation, that can be vectorized.
3361 If VEC_STMT is also passed, vectorize the STMT: create a vectorized
3362 stmt to replace it, put it in VEC_STMT, and insert it at BSI.
3363 Return FALSE if not a vectorizable STMT, TRUE otherwise. */
3365 bool
3368 {
3369 tree vec_dest;
3370 tree scalar_dest;
3371 tree operation;
3372 tree op0;
3379 tree new_temp;
3383 stmt_vec_info prev_stmt_info;
3388 tree expr;
3390 tree builtin_decl;
3394 tree vop0;
3396 /* Is STMT a vectorizable conversion? */
3415 /* Check types of lhs and rhs. */
3430 /* FORNOW */
3437 else
3443 /* Bail out if the types are both integral or non-integral. */
3450 else
3453 /* FORNOW: SLP with multiple types is not supported. The SLP analysis verifies
3454 this, so we can safely override NCOPIES with 1 here. */
3458 /* Sanity check: make sure that at least one copy of the vectorized stmt
3459 needs to be generated. */
3462 /* FORNOW. This restriction should be relaxed. */
3464 {
3468 }
3470 /* Check the operands of the operation. */
3472 {
3476 }
3478 /* Supportable by target? */
3488 {
3492 }
3495 {
3497 /* FORNOW: SLP not supported. */
3500 }
3503 {
3506 }
3508 /** Transform. **/
3512 /* Handle def. */
3520 {
3523 {
3524 tree sym;
3525 ssa_op_iter iter;
3529 else
3532 builtin_decl =
3535 {
3538 /* Arguments are ready. create the new vector stmt. */
3544 SSA_OP_ALL_VIRTUALS)
3545 {
3549 }
3552 }
3556 else
3559 }
3563 /* In case the vectorization factor (VF) is bigger than the number
3564 of elements that we can fit in a vectype (nunits), we have to
3565 generate more than one vector stmt - i.e - we need to "unroll"
3566 the vector stmt by a factor VF/nunits. */
3568 {
3571 else
3576 /* Generate first half of the widened result: */
3577 new_stmt
3583 else
3587 /* Generate second half of the widened result: */
3588 new_stmt
3594 }
3598 /* In case the vectorization factor (VF) is bigger than the number
3599 of elements that we can fit in a vectype (nunits), we have to
3600 generate more than one vector stmt - i.e - we need to "unroll"
3601 the vector stmt by a factor VF/nunits. */
3603 {
3604 /* Handle uses. */
3606 {
3609 }
3610 else
3611 {
3614 }
3616 /* Arguments are ready. Create the new vector stmt. */
3625 else
3629 }
3632 }
3638 }
3641 /* Function vectorizable_assignment.
3643 Check if STMT performs an assignment (copy) that can be vectorized.
3644 If VEC_STMT is also passed, vectorize the STMT: create a vectorized
3645 stmt to replace it, put it in VEC_STMT, and insert it at BSI.
3646 Return FALSE if not a vectorizable STMT, TRUE otherwise. */
3648 bool
3650 slp_tree slp_node)
3651 {
3652 tree vec_dest;
3653 tree scalar_dest;
3654 tree op;
3658 tree new_temp;
3665 tree vop;
3677 /* Is vectorizable assignment? */
3687 {
3691 }
3694 {
3700 }
3702 /** Transform. **/
3706 /* Handle def. */
3709 /* Handle use. */
3712 /* Arguments are ready. create the new vector stmt. */
3714 {
3723 }
3727 }
3730 /* Function vect_min_worthwhile_factor.
3732 For a loop where we could vectorize the operation indicated by CODE,
3733 return the minimum vectorization factor that makes it worthwhile
3734 to use generic vectors. */
3735 static int
3737 {
3739 {
3753 }
3754 }
3757 /* Function vectorizable_induction
3759 Check if PHI performs an induction computation that can be vectorized.
3760 If VEC_STMT is also passed, vectorize the induction PHI: create a vectorized
3761 phi to replace it, put it in VEC_STMT, and add it to the same basic block.
3762 Return FALSE if not a vectorizable STMT, TRUE otherwise. */
3764 bool
3767 {
3773 tree vec_def;
3780 /* FORNOW: SLP not supported. */
3790 {
3796 }
3798 /** Transform. **/
3806 }
3809 /* Function vectorizable_operation.
3811 Check if STMT performs a binary or unary operation that can be vectorized.
3812 If VEC_STMT is also passed, vectorize the STMT: create a vectorized
3813 stmt to replace it, put it in VEC_STMT, and insert it at BSI.
3814 Return FALSE if not a vectorizable STMT, TRUE otherwise. */
3816 bool
3818 slp_tree slp_node)
3819 {
3820 tree vec_dest;
3821 tree scalar_dest;
3822 tree operation;
3831 tree new_temp;
3833 optab optab;
3839 stmt_vec_info prev_stmt_info;
3842 tree vectype_out;
3850 /* FORNOW: SLP with multiple types is not supported. The SLP analysis verifies
3851 this, so we can safely override NCOPIES with 1 here. */
3855 /* FORNOW. This restriction should be relaxed. */
3857 {
3861 }
3869 /* Is STMT a vectorizable binary/unary operation? */
3887 /* For pointer addition, we should use the normal plus for
3888 the vector addition. */
3894 /* Support only unary or binary operations. */
3897 {
3901 }
3905 {
3909 }
3912 {
3915 {
3919 }
3920 }
3922 /* Supportable by target? */
3924 {
3928 }
3932 {
3935 /* Check only during analysis. */
3943 }
3945 /* Worthwhile without SIMD support? Check only during analysis. */
3950 {
3954 }
3957 {
3958 /* FORNOW: not yet supported. */
3962 /* Invariant argument is needed for a vector shift
3963 by a scalar shift operand. */
3966 {
3968 {
3973 }
3976 }
3977 }
3980 {
3986 }
3988 /** Transform. **/
3993 /* Handle def. */
3996 /* Allocate VECs for vector operands. In case of SLP, vector operands are
3997 created in the previous stages of the recursion, so no allocation is
3998 needed, except for the case of shift with scalar shift argument. In that
3999 case we store the scalar operand in VEC_OPRNDS1 for every vector stmt to
4000 be created to vectorize the SLP group, i.e., SLP_NODE->VEC_STMTS_SIZE.
4001 In case of loop-based vectorization we allocate VECs of size 1. We
4002 allocate VEC_OPRNDS1 only in case of binary operation. */
4004 {
4008 }
4012 /* In case the vectorization factor (VF) is bigger than the number
4013 of elements that we can fit in a vectype (nunits), we have to generate
4014 more than one vector stmt - i.e - we need to "unroll" the
4015 vector stmt by a factor VF/nunits. In doing so, we record a pointer
4016 from one copy of the vector stmt to the next, in the field
4017 STMT_VINFO_RELATED_STMT. This is necessary in order to allow following
4018 stages to find the correct vector defs to be used when vectorizing
4019 stmts that use the defs of the current stmt. The example below illustrates
4020 the vectorization process when VF=16 and nunits=4 (i.e - we need to create
4021 4 vectorized stmts):
4023 before vectorization:
4024 RELATED_STMT VEC_STMT
4025 S1: x = memref - -
4026 S2: z = x + 1 - -
4028 step 1: vectorize stmt S1 (done in vectorizable_load. See more details
4029 there):
4030 RELATED_STMT VEC_STMT
4031 VS1_0: vx0 = memref0 VS1_1 -
4032 VS1_1: vx1 = memref1 VS1_2 -
4033 VS1_2: vx2 = memref2 VS1_3 -
4034 VS1_3: vx3 = memref3 - -
4035 S1: x = load - VS1_0
4036 S2: z = x + 1 - -
4038 step2: vectorize stmt S2 (done here):
4039 To vectorize stmt S2 we first need to find the relevant vector
4040 def for the first operand 'x'. This is, as usual, obtained from
4041 the vector stmt recorded in the STMT_VINFO_VEC_STMT of the stmt
4042 that defines 'x' (S1). This way we find the stmt VS1_0, and the
4043 relevant vector def 'vx0'. Having found 'vx0' we can generate
4044 the vector stmt VS2_0, and as usual, record it in the
4045 STMT_VINFO_VEC_STMT of stmt S2.
4046 When creating the second copy (VS2_1), we obtain the relevant vector
4047 def from the vector stmt recorded in the STMT_VINFO_RELATED_STMT of
4048 stmt VS1_0. This way we find the stmt VS1_1 and the relevant
4049 vector def 'vx1'. Using 'vx1' we create stmt VS2_1 and record a
4050 pointer to it in the STMT_VINFO_RELATED_STMT of the vector stmt VS2_0.
4051 Similarly when creating stmts VS2_2 and VS2_3. This is the resulting
4052 chain of stmts and pointers:
4053 RELATED_STMT VEC_STMT
4054 VS1_0: vx0 = memref0 VS1_1 -
4055 VS1_1: vx1 = memref1 VS1_2 -
4056 VS1_2: vx2 = memref2 VS1_3 -
4057 VS1_3: vx3 = memref3 - -
4058 S1: x = load - VS1_0
4059 VS2_0: vz0 = vx0 + v1 VS2_1 -
4060 VS2_1: vz1 = vx1 + v1 VS2_2 -
4061 VS2_2: vz2 = vx2 + v1 VS2_3 -
4062 VS2_3: vz3 = vx3 + v1 - -
4063 S2: z = x + 1 - VS2_0 */
4067 {
4068 /* Handle uses. */
4070 {
4073 {
4074 /* Vector shl and shr insn patterns can be defined with scalar
4075 operand 2 (shift operand). In this case, use constant or loop
4076 invariant op1 directly, without extending it to vector mode
4077 first. */
4080 {
4086 {
4087 /* Store vec_oprnd1 for every vector stmt to be created
4088 for SLP_NODE. We check during the analysis that all the
4089 shift arguments are the same.
4090 TODO: Allow different constants for different vector
4091 stmts generated for an SLP instance. */
4094 }
4095 }
4096 }
4098 /* vec_oprnd1 is available if operand 1 should be of a scalar-type
4099 (a special case for certain kind of vector shifts); otherwise,
4100 operand 1 should be of a vector type (the usual case). */
4103 slp_node);
4104 else
4106 slp_node);
4107 }
4108 else
4111 /* Arguments are ready. Create the new vector stmt. */
4113 {
4115 {
4119 }
4120 else
4129 }
4133 else
4136 }
4143 }
4146 /* Function vectorizable_type_demotion
4148 Check if STMT performs a binary or unary operation that involves
4149 type demotion, and if it can be vectorized.
4150 If VEC_STMT is also passed, vectorize the STMT: create a vectorized
4151 stmt to replace it, put it in VEC_STMT, and insert it at BSI.
4152 Return FALSE if not a vectorizable STMT, TRUE otherwise. */
4154 bool
4157 {
4158 tree vec_dest;
4159 tree scalar_dest;
4160 tree operation;
4161 tree op0;
4167 tree new_temp;
4170 tree new_stmt;
4171 stmt_vec_info prev_stmt_info;
4174 tree vectype_out;
4177 tree expr;
4178 tree vectype_in;
4186 /* Is STMT a vectorizable type-demotion operation? */
4214 /* FORNOW. This restriction should be relaxed. */
4216 {
4220 }
4229 /* Check the operands of the operation. */
4231 {
4235 }
4237 /* Supportable by target? */
4244 {
4250 }
4252 /** Transform. **/
4255 ncopies);
4257 /* Handle def. */
4260 /* In case the vectorization factor (VF) is bigger than the number
4261 of elements that we can fit in a vectype (nunits), we have to generate
4262 more than one vector stmt - i.e - we need to "unroll" the
4263 vector stmt by a factor VF/nunits. */
4266 {
4267 /* Handle uses. */
4269 {
4272 }
4273 else
4274 {
4277 }
4279 /* Arguments are ready. Create the new vector stmt. */
4288 else
4292 }
4296 }
4299 /* Function vectorizable_type_promotion
4301 Check if STMT performs a binary or unary operation that involves
4302 type promotion, and if it can be vectorized.
4303 If VEC_STMT is also passed, vectorize the STMT: create a vectorized
4304 stmt to replace it, put it in VEC_STMT, and insert it at BSI.
4305 Return FALSE if not a vectorizable STMT, TRUE otherwise. */
4307 bool
4310 {
4311 tree vec_dest;
4312 tree scalar_dest;
4313 tree operation;
4324 tree new_stmt;
4325 stmt_vec_info prev_stmt_info;
4328 tree vectype_out;
4331 tree vectype_in;
4339 /* Is STMT a vectorizable type-promotion operation? */
4368 /* FORNOW. This restriction should be relaxed. */
4370 {
4374 }
4383 /* Check the operands of the operation. */
4385 {
4389 }
4393 {
4396 {
4400 }
4401 }
4403 /* Supportable by target? */
4411 {
4417 }
4419 /** Transform. **/
4423 ncopies);
4425 /* Handle def. */
4428 /* In case the vectorization factor (VF) is bigger than the number
4429 of elements that we can fit in a vectype (nunits), we have to generate
4430 more than one vector stmt - i.e - we need to "unroll" the
4431 vector stmt by a factor VF/nunits. */
4435 {
4436 /* Handle uses. */
4438 {
4442 }
4443 else
4444 {
4448 }
4450 /* Arguments are ready. Create the new vector stmt. We are creating
4451 two vector defs because the widened result does not fit in one vector.
4452 The vectorized stmt can be expressed as a call to a taregt builtin,
4453 or a using a tree-code. */
4454 /* Generate first half of the widened result: */
4459 else
4463 /* Generate second half of the widened result: */
4469 }
4473 }
4476 /* Function vect_strided_store_supported.
4478 Returns TRUE is INTERLEAVE_HIGH and INTERLEAVE_LOW operations are supported,
4479 and FALSE otherwise. */
4481 static bool
4483 {
4489 /* Check that the operation is supported. */
4491 vectype);
4493 vectype);
4495 {
4499 }
4502 == CODE_FOR_nothing
4505 {
4509 }
4512 }
4515 /* Function vect_permute_store_chain.
4517 Given a chain of interleaved stores in DR_CHAIN of LENGTH that must be
4518 a power of 2, generate interleave_high/low stmts to reorder the data
4519 correctly for the stores. Return the final references for stores in
4520 RESULT_CHAIN.
4522 E.g., LENGTH is 4 and the scalar type is short, i.e., VF is 8.
4523 The input is 4 vectors each containing 8 elements. We assign a number to each
4524 element, the input sequence is:
4526 1st vec: 0 1 2 3 4 5 6 7
4527 2nd vec: 8 9 10 11 12 13 14 15
4528 3rd vec: 16 17 18 19 20 21 22 23
4529 4th vec: 24 25 26 27 28 29 30 31
4531 The output sequence should be:
4533 1st vec: 0 8 16 24 1 9 17 25
4534 2nd vec: 2 10 18 26 3 11 19 27
4535 3rd vec: 4 12 20 28 5 13 21 30
4536 4th vec: 6 14 22 30 7 15 23 31
4538 i.e., we interleave the contents of the four vectors in their order.
4540 We use interleave_high/low instructions to create such output. The input of
4541 each interleave_high/low operation is two vectors:
4542 1st vec 2nd vec
4543 0 1 2 3 4 5 6 7
4544 the even elements of the result vector are obtained left-to-right from the
4545 high/low elements of the first vector. The odd elements of the result are
4546 obtained left-to-right from the high/low elements of the second vector.
4547 The output of interleave_high will be: 0 4 1 5
4548 and of interleave_low: 2 6 3 7
4551 The permutation is done in log LENGTH stages. In each stage interleave_high
4552 and interleave_low stmts are created for each pair of vectors in DR_CHAIN,
4553 where the first argument is taken from the first half of DR_CHAIN and the
4554 second argument from it's second half.
4555 In our example,
4557 I1: interleave_high (1st vec, 3rd vec)
4558 I2: interleave_low (1st vec, 3rd vec)
4559 I3: interleave_high (2nd vec, 4th vec)
4560 I4: interleave_low (2nd vec, 4th vec)
4562 The output for the first stage is:
4564 I1: 0 16 1 17 2 18 3 19
4565 I2: 4 20 5 21 6 22 7 23
4566 I3: 8 24 9 25 10 26 11 27
4567 I4: 12 28 13 29 14 30 15 31
4569 The output of the second stage, i.e. the final result is:
4571 I1: 0 8 16 24 1 9 17 25
4572 I2: 2 10 18 26 3 11 19 27
4573 I3: 4 12 20 28 5 13 21 30
4574 I4: 6 14 22 30 7 15 23 31. */
4576 static bool
4579 tree stmt,
4582 {
4591 /* Check that the operation is supported. */
4598 {
4600 {
4604 /* Create interleaving stmt:
4605 in the case of big endian:
4606 high = interleave_high (vect1, vect2)
4607 and in the case of little endian:
4608 high = interleave_low (vect1, vect2). */
4614 else
4622 /* Create interleaving stmt:
4623 in the case of big endian:
4624 low = interleave_low (vect1, vect2)
4625 and in the case of little endian:
4626 low = interleave_high (vect1, vect2). */
4632 else
4639 }
4641 }
4643 }
4646 /* Function vectorizable_store.
4648 Check if STMT defines a non scalar data-ref (array/pointer/structure) that
4649 can be vectorized.
4650 If VEC_STMT is also passed, vectorize the STMT: create a vectorized
4651 stmt to replace it, put it in VEC_STMT, and insert it at BSI.
4652 Return FALSE if not a vectorizable STMT, TRUE otherwise. */
4654 bool
4656 slp_tree slp_node)
4657 {
4658 tree scalar_dest;
4659 tree data_ref;
4660 tree op;
4668 tree dummy;
4684 stmt_vec_info first_stmt_vinfo;
4687 /* FORNOW: SLP with multiple types is not supported. The SLP analysis verifies
4688 this, so we can safely override NCOPIES with 1 here. */
4694 /* FORNOW. This restriction should be relaxed. */
4696 {
4700 }
4708 /* Is vectorizable store? */
4721 {
4725 }
4728 /* FORNOW. In some cases can vectorize even if data-type not supported
4729 (e.g. - array initialization with 0). */
4737 {
4745 {
4746 /* STMT is the leader of the group. Check the operands of all the
4747 stmts of the group. */
4750 {
4753 {
4757 }
4759 }
4760 }
4761 }
4764 {
4769 }
4771 /** Transform. **/
4774 {
4780 /* FORNOW */
4783 /* We vectorize all the stmts of the interleaving group when we
4784 reach the last stmt in the group. */
4788 {
4791 }
4796 /* VEC_NUM is the number of vect stmts to be created for this group. */
4799 else
4801 }
4802 else
4803 {
4808 }
4820 /* In case the vectorization factor (VF) is bigger than the number
4821 of elements that we can fit in a vectype (nunits), we have to generate
4822 more than one vector stmt - i.e - we need to "unroll" the
4823 vector stmt by a factor VF/nunits. For more details see documentation in
4824 vect_get_vec_def_for_copy_stmt. */
4826 /* In case of interleaving (non-unit strided access):
4828 S1: &base + 2 = x2
4829 S2: &base = x0
4830 S3: &base + 1 = x1
4831 S4: &base + 3 = x3
4833 We create vectorized stores starting from base address (the access of the
4834 first stmt in the chain (S2 in the above example), when the last store stmt
4835 of the chain (S4) is reached:
4837 VS1: &base = vx2
4838 VS2: &base + vec_size*1 = vx0
4839 VS3: &base + vec_size*2 = vx1
4840 VS4: &base + vec_size*3 = vx3
4842 Then permutation statements are generated:
4844 VS5: vx5 = VEC_INTERLEAVE_HIGH_EXPR < vx0, vx3 >
4845 VS6: vx6 = VEC_INTERLEAVE_LOW_EXPR < vx0, vx3 >
4846 ...
4848 And they are put in STMT_VINFO_VEC_STMT of the corresponding scalar stmts
4849 (the order of the data-refs in the output of vect_permute_store_chain
4850 corresponds to the order of scalar stmts in the interleaving chain - see
4851 the documentation of vect_permute_store_chain()).
4853 In case of both multiple types and interleaving, above vector stores and
4854 permutation stmts are created for every copy. The result vector stmts are
4855 put in STMT_VINFO_VEC_STMT for the first copy and in the corresponding
4856 STMT_VINFO_RELATED_STMT for the next copies.
4857 */
4861 {
4862 tree new_stmt;
4863 tree ptr_incr;
4866 {
4868 {
4869 /* Get vectorized arguments for SLP_NODE. */
4873 }
4874 else
4875 {
4876 /* For interleaved stores we collect vectorized defs for all the
4877 stores in the group in DR_CHAIN and OPRNDS. DR_CHAIN is then
4878 used as an input to vect_permute_store_chain(), and OPRNDS as
4879 an input to vect_get_vec_def_for_stmt_copy() for the next copy.
4881 If the store is not strided, GROUP_SIZE is 1, and DR_CHAIN and
4882 OPRNDS are of size 1. */
4885 {
4886 /* Since gaps are not supported for interleaved stores,
4887 GROUP_SIZE is the exact number of stmts in the chain.
4888 Therefore, NEXT_STMT can't be NULL_TREE. In case that
4889 there is no interleaving, GROUP_SIZE is 1, and only one
4890 iteration of the loop will be executed. */
4895 NULL);
4899 }
4900 }
4905 }
4906 else
4907 {
4908 /* FORNOW SLP doesn't work for multiple types. */
4911 /* For interleaved stores we created vectorized defs for all the
4912 defs stored in OPRNDS in the previous iteration (previous copy).
4913 DR_CHAIN is then used as an input to vect_permute_store_chain(),
4914 and OPRNDS as an input to vect_get_vec_def_for_stmt_copy() for the
4915 next copy.
4916 If the store is not strided, GROUP_SIZE is 1, and DR_CHAIN and
4917 OPRNDS are of size 1. */
4919 {
4925 }
4926 dataref_ptr =
4928 }
4931 {
4933 /* Permute. */
4937 }
4941 {
4943 /* Bump the vector pointer. */
4945 NULL_TREE);
4950 /* For strided stores vectorized defs are interleaved in
4951 vect_permute_store_chain(). */
4955 /* Arguments are ready. Create the new vector stmt. */
4962 else
4969 }
4970 }
4978 }
4981 /* Function vect_setup_realignment
4983 This function is called when vectorizing an unaligned load using
4984 the dr_explicit_realign[_optimized] scheme.
4985 This function generates the following code at the loop prolog:
4987 p = initial_addr;
4988 x msq_init = *(floor(p)); # prolog load
4989 realignment_token = call target_builtin;
4990 loop:
4991 x msq = phi (msq_init, ---)
4993 The stmts marked with x are generated only for the case of
4994 dr_explicit_realign_optimized.
4996 The code above sets up a new (vector) pointer, pointing to the first
4997 location accessed by STMT, and a "floor-aligned" load using that pointer.
4998 It also generates code to compute the "realignment-token" (if the relevant
4999 target hook was defined), and creates a phi-node at the loop-header bb
5000 whose arguments are the result of the prolog-load (created by this
5001 function) and the result of a load that takes place in the loop (to be
5002 created by the caller to this function).
5004 For the case of dr_explicit_realign_optimized:
5005 The caller to this function uses the phi-result (msq) to create the
5006 realignment code inside the loop, and sets up the missing phi argument,
5007 as follows:
5008 loop:
5009 msq = phi (msq_init, lsq)
5010 lsq = *(floor(p')); # load in loop
5011 result = realign_load (msq, lsq, realignment_token);
5013 For the case of dr_explicit_realign:
5014 loop:
5015 msq = *(floor(p)); # load in loop
5016 p' = p + (VS-1);
5017 lsq = *(floor(p')); # load in loop
5018 result = realign_load (msq, lsq, realignment_token);
5020 Input:
5021 STMT - (scalar) load stmt to be vectorized. This load accesses
5022 a memory location that may be unaligned.
5023 BSI - place where new code is to be inserted.
5024 ALIGNMENT_SUPPORT_SCHEME - which of the two misalignment handling schemes
5025 is used.
5027 Output:
5028 REALIGNMENT_TOKEN - the result of a call to the builtin_mask_for_load
5029 target hook, if defined.
5030 Return value - the result of the loop-header phi node. */
5032 static tree
5036 tree init_addr,
5038 {
5043 edge pe;
5045 tree vec_dest;
5046 tree inc;
5047 tree ptr;
5048 tree data_ref;
5049 tree new_stmt;
5050 basic_block new_bb;
5052 tree new_temp;
5053 tree phi_stmt;
5065 /* We need to generate three things:
5066 1. the misalignment computation
5067 2. the extra vector load (for the optimized realignment scheme).
5068 3. the phi node for the two vectors from which the realignment is
5069 done (for the optimized realignment scheme).
5070 */
5072 /* 1. Determine where to generate the misalignment computation.
5074 If INIT_ADDR is NULL_TREE, this indicates that the misalignment
5075 calculation will be generated by this function, outside the loop (in the
5076 preheader). Otherwise, INIT_ADDR had already been computed for us by the
5077 caller, inside the loop.
5079 Background: If the misalignment remains fixed throughout the iterations of
5080 the loop, then both realignment schemes are applicable, and also the
5081 misalignment computation can be done outside LOOP. This is because we are
5082 vectorizing LOOP, and so the memory accesses in LOOP advance in steps that
5083 are a multiple of VS (the Vector Size), and therefore the misalignment in
5084 different vectorized LOOP iterations is always the same.
5085 The problem arises only if the memory access is in an inner-loop nested
5086 inside LOOP, which is now being vectorized using outer-loop vectorization.
5087 This is the only case when the misalignment of the memory access may not
5088 remain fixed throughout the iterations of the inner-loop (as explained in
5089 detail in vect_supportable_dr_alignment). In this case, not only is the
5090 optimized realignment scheme not applicable, but also the misalignment
5091 computation (and generation of the realignment token that is passed to
5092 REALIGN_LOAD) have to be done inside the loop.
5094 In short, INIT_ADDR indicates whether we are in a COMPUTE_IN_LOOP mode
5095 or not, which in turn determines if the misalignment is computed inside
5096 the inner-loop, or outside LOOP. */
5099 {
5102 }
5105 /* 2. Determine where to generate the extra vector load.
5107 For the optimized realignment scheme, instead of generating two vector
5108 loads in each iteration, we generate a single extra vector load in the
5109 preheader of the loop, and in each iteration reuse the result of the
5110 vector load from the previous iteration. In case the memory access is in
5111 an inner-loop nested inside LOOP, which is now being vectorized using
5112 outer-loop vectorization, we need to determine whether this initial vector
5113 load should be generated at the preheader of the inner-loop, or can be
5114 generated at the preheader of LOOP. If the memory access has no evolution
5115 in LOOP, it can be generated in the preheader of LOOP. Otherwise, it has
5116 to be generated inside LOOP (in the preheader of the inner-loop). */
5119 {
5124 }
5125 else
5130 /* 3. For the case of the optimized realignment, create the first vector
5131 load at the loop preheader. */
5134 {
5135 /* Create msq_init = *(floor(p1)) in the loop preheader */
5150 }
5152 /* 4. Create realignment token using a target builtin, if available.
5153 It is done either inside the containing loop, or before LOOP (as
5154 determined above). */
5157 {
5158 tree builtin_decl;
5160 /* Compute INIT_ADDR - the initial addressed accessed by this memref. */
5163 else
5164 {
5165 /* Generate the INIT_ADDR computation outside LOOP. */
5171 }
5183 else
5184 {
5185 /* Generate the misalignment computation outside LOOP. */
5189 }
5193 /* The result of the CALL_EXPR to this builtin is determined from
5194 the value of the parameter and no global variables are touched
5195 which makes the builtin a "const" function. Requiring the
5196 builtin to have the "const" attribute makes it unnecessary
5197 to call mark_call_clobbered. */
5199 }
5208 /* 5. Create msq = phi <msq_init, lsq> in loop */
5218 }
5221 /* Function vect_strided_load_supported.
5223 Returns TRUE is EXTRACT_EVEN and EXTRACT_ODD operations are supported,
5224 and FALSE otherwise. */
5226 static bool
5228 {
5236 {
5240 }
5243 {
5247 }
5251 {
5255 }
5258 {
5262 }
5264 }
5267 /* Function vect_permute_load_chain.
5269 Given a chain of interleaved loads in DR_CHAIN of LENGTH that must be
5270 a power of 2, generate extract_even/odd stmts to reorder the input data
5271 correctly. Return the final references for loads in RESULT_CHAIN.
5273 E.g., LENGTH is 4 and the scalar type is short, i.e., VF is 8.
5274 The input is 4 vectors each containing 8 elements. We assign a number to each
5275 element, the input sequence is:
5277 1st vec: 0 1 2 3 4 5 6 7
5278 2nd vec: 8 9 10 11 12 13 14 15
5279 3rd vec: 16 17 18 19 20 21 22 23
5280 4th vec: 24 25 26 27 28 29 30 31
5282 The output sequence should be:
5284 1st vec: 0 4 8 12 16 20 24 28
5285 2nd vec: 1 5 9 13 17 21 25 29
5286 3rd vec: 2 6 10 14 18 22 26 30
5287 4th vec: 3 7 11 15 19 23 27 31
5289 i.e., the first output vector should contain the first elements of each
5290 interleaving group, etc.
5292 We use extract_even/odd instructions to create such output. The input of each
5293 extract_even/odd operation is two vectors
5294 1st vec 2nd vec
5295 0 1 2 3 4 5 6 7
5297 and the output is the vector of extracted even/odd elements. The output of
5298 extract_even will be: 0 2 4 6
5299 and of extract_odd: 1 3 5 7
5302 The permutation is done in log LENGTH stages. In each stage extract_even and
5303 extract_odd stmts are created for each pair of vectors in DR_CHAIN in their
5304 order. In our example,
5306 E1: extract_even (1st vec, 2nd vec)
5307 E2: extract_odd (1st vec, 2nd vec)
5308 E3: extract_even (3rd vec, 4th vec)
5309 E4: extract_odd (3rd vec, 4th vec)
5311 The output for the first stage will be:
5313 E1: 0 2 4 6 8 10 12 14
5314 E2: 1 3 5 7 9 11 13 15
5315 E3: 16 18 20 22 24 26 28 30
5316 E4: 17 19 21 23 25 27 29 31
5318 In order to proceed and create the correct sequence for the next stage (or
5319 for the correct output, if the second stage is the last one, as in our
5320 example), we first put the output of extract_even operation and then the
5321 output of extract_odd in RESULT_CHAIN (which is then copied to DR_CHAIN).
5322 The input for the second stage is:
5324 1st vec (E1): 0 2 4 6 8 10 12 14
5325 2nd vec (E3): 16 18 20 22 24 26 28 30
5326 3rd vec (E2): 1 3 5 7 9 11 13 15
5327 4th vec (E4): 17 19 21 23 25 27 29 31
5329 The output of the second stage:
5331 E1: 0 4 8 12 16 20 24 28
5332 E2: 2 6 10 14 18 22 26 30
5333 E3: 1 5 9 13 17 21 25 29
5334 E4: 3 7 11 15 19 23 27 31
5336 And RESULT_CHAIN after reordering:
5338 1st vec (E1): 0 4 8 12 16 20 24 28
5339 2nd vec (E3): 1 5 9 13 17 21 25 29
5340 3rd vec (E2): 2 6 10 14 18 22 26 30
5341 4th vec (E4): 3 7 11 15 19 23 27 31. */
5343 static bool
5346 tree stmt,
5349 {
5352 tree tmp;
5356 /* Check that the operation is supported. */
5362 {
5364 {
5368 /* data_ref = permute_even (first_data_ref, second_data_ref); */
5384 /* data_ref = permute_odd (first_data_ref, second_data_ref); */
5398 }
5400 }
5402 }
5405 /* Function vect_transform_strided_load.
5407 Given a chain of input interleaved data-refs (in DR_CHAIN), build statements
5408 to perform their permutation and ascribe the result vectorized statements to
5409 the scalar statements.
5410 */
5412 static bool
5415 {
5421 tree tmp_data_ref;
5423 /* DR_CHAIN contains input data-refs that are a part of the interleaving.
5424 RESULT_CHAIN is the output of vect_permute_load_chain, it contains permuted
5425 vectors, that are ready for vector computation. */
5427 /* Permute. */
5431 /* Put a permuted data-ref in the VECTORIZED_STMT field.
5432 Since we scan the chain starting from it's first node, their order
5433 corresponds the order of data-refs in RESULT_CHAIN. */
5437 {
5441 /* Skip the gaps. Loads created for the gaps will be removed by dead
5442 code elimination pass later.
5443 DR_GROUP_GAP is the number of steps in elements from the previous
5444 access (if there is no gap DR_GROUP_GAP is 1). We skip loads that
5445 correspond to the gaps.
5446 */
5448 {
5449 gap_count++;
5451 }
5454 {
5456 /* We assume that if VEC_STMT is not NULL, this is a case of multiple
5457 copies, and we put the new vector statement in the first available
5458 RELATED_STMT. */
5461 else
5462 {
5467 {
5470 }
5472 }
5475 /* If NEXT_STMT accesses the same DR as the previous statement,
5476 put the same TMP_DATA_REF as its vectorized statement; otherwise
5477 get the next data-ref from RESULT_CHAIN. */
5480 }
5481 }
5485 }
5488 /* vectorizable_load.
5490 Check if STMT reads a non scalar data-ref (array/pointer/structure) that
5491 can be vectorized.
5492 If VEC_STMT is also passed, vectorize the STMT: create a vectorized
5493 stmt to replace it, put it in VEC_STMT, and insert it at BSI.
5494 Return FALSE if not a vectorizable STMT, TRUE otherwise. */
5496 bool
5498 slp_tree slp_node)
5499 {
5500 tree scalar_dest;
5503 tree op;
5505 stmt_vec_info prev_stmt_info;
5512 tree new_temp;
5515 tree dummy;
5518 tree ptr_incr;
5528 tree first_stmt;
5529 tree scalar_type;
5536 /* FORNOW: SLP with multiple types is not supported. The SLP analysis verifies
5537 this, so we can safely override NCOPIES with 1 here. */
5543 /* FORNOW. This restriction should be relaxed. */
5545 {
5549 }
5557 /* Is vectorizable load? */
5577 /* FORNOW. In some cases can vectorize even if data-type not supported
5578 (e.g. - data copies). */
5580 {
5584 }
5586 /* Check if the load is a part of an interleaving chain. */
5588 {
5590 /* FORNOW */
5593 /* Check if interleaving is supported. */
5597 }
5600 {
5604 }
5609 /** Transform. **/
5612 {
5614 /* Check if the chain of loads is already vectorized. */
5616 {
5619 }
5624 /* VEC_NUM is the number of vect stmts to be created for this group. */
5626 {
5629 }
5630 else
5632 }
5633 else
5634 {
5638 }
5643 /* In case the vectorization factor (VF) is bigger than the number
5644 of elements that we can fit in a vectype (nunits), we have to generate
5645 more than one vector stmt - i.e - we need to "unroll" the
5646 vector stmt by a factor VF/nunits. In doing so, we record a pointer
5647 from one copy of the vector stmt to the next, in the field
5648 STMT_VINFO_RELATED_STMT. This is necessary in order to allow following
5649 stages to find the correct vector defs to be used when vectorizing
5650 stmts that use the defs of the current stmt. The example below illustrates
5651 the vectorization process when VF=16 and nunits=4 (i.e - we need to create
5652 4 vectorized stmts):
5654 before vectorization:
5655 RELATED_STMT VEC_STMT
5656 S1: x = memref - -
5657 S2: z = x + 1 - -
5659 step 1: vectorize stmt S1:
5660 We first create the vector stmt VS1_0, and, as usual, record a
5661 pointer to it in the STMT_VINFO_VEC_STMT of the scalar stmt S1.
5662 Next, we create the vector stmt VS1_1, and record a pointer to
5663 it in the STMT_VINFO_RELATED_STMT of the vector stmt VS1_0.
5664 Similarly, for VS1_2 and VS1_3. This is the resulting chain of
5665 stmts and pointers:
5666 RELATED_STMT VEC_STMT
5667 VS1_0: vx0 = memref0 VS1_1 -
5668 VS1_1: vx1 = memref1 VS1_2 -
5669 VS1_2: vx2 = memref2 VS1_3 -
5670 VS1_3: vx3 = memref3 - -
5671 S1: x = load - VS1_0
5672 S2: z = x + 1 - -
5674 See in documentation in vect_get_vec_def_for_stmt_copy for how the
5675 information we recorded in RELATED_STMT field is used to vectorize
5676 stmt S2. */
5678 /* In case of interleaving (non-unit strided access):
5680 S1: x2 = &base + 2
5681 S2: x0 = &base
5682 S3: x1 = &base + 1
5683 S4: x3 = &base + 3
5685 Vectorized loads are created in the order of memory accesses
5686 starting from the access of the first stmt of the chain:
5688 VS1: vx0 = &base
5689 VS2: vx1 = &base + vec_size*1
5690 VS3: vx3 = &base + vec_size*2
5691 VS4: vx4 = &base + vec_size*3
5693 Then permutation statements are generated:
5695 VS5: vx5 = VEC_EXTRACT_EVEN_EXPR < vx0, vx1 >
5696 VS6: vx6 = VEC_EXTRACT_ODD_EXPR < vx0, vx1 >
5697 ...
5699 And they are put in STMT_VINFO_VEC_STMT of the corresponding scalar stmts
5700 (the order of the data-refs in the output of vect_permute_load_chain
5701 corresponds to the order of scalar stmts in the interleaving chain - see
5702 the documentation of vect_permute_load_chain()).
5703 The generation of permutation stmts and recording them in
5704 STMT_VINFO_VEC_STMT is done in vect_transform_strided_load().
5706 In case of both multiple types and interleaving, the vector loads and
5707 permutation stmts above are created for every copy. The result vector stmts
5708 are put in STMT_VINFO_VEC_STMT for the first copy and in the corresponding
5709 STMT_VINFO_RELATED_STMT for the next copies. */
5711 /* If the data reference is aligned (dr_aligned) or potentially unaligned
5712 on a target that supports unaligned accesses (dr_unaligned_supported)
5713 we generate the following code:
5714 p = initial_addr;
5715 indx = 0;
5716 loop {
5717 p = p + indx * vectype_size;
5718 vec_dest = *(p);
5719 indx = indx + 1;
5720 }
5722 Otherwise, the data reference is potentially unaligned on a target that
5723 does not support unaligned accesses (dr_explicit_realign_optimized) -
5724 then generate the following code, in which the data in each iteration is
5725 obtained by two vector loads, one from the previous iteration, and one
5726 from the current iteration:
5727 p1 = initial_addr;
5728 msq_init = *(floor(p1))
5729 p2 = initial_addr + VS - 1;
5730 realignment_token = call target_builtin;
5731 indx = 0;
5732 loop {
5733 p2 = p2 + indx * vectype_size
5734 lsq = *(floor(p2))
5735 vec_dest = realign_load (msq, lsq, realignment_token)
5736 indx = indx + 1;
5737 msq = lsq;
5738 } */
5740 /* If the misalignment remains the same throughout the execution of the
5741 loop, we can create the init_addr and permutation mask at the loop
5742 preheader. Otherwise, it needs to be created inside the loop.
5743 This can only occur when vectorizing memory accesses in the inner-loop
5744 nested within an outer-loop that is being vectorized. */
5748 {
5751 }
5756 {
5761 {
5764 }
5765 }
5766 else
5771 {
5772 /* 1. Create the vector pointer update chain. */
5778 else
5779 dataref_ptr =
5783 {
5786 NULL_TREE);
5788 /* 2. Create the vector-load in the loop. */
5790 {
5796 {
5801 data_ref =
5804 }
5806 {
5813 dr_explicit_realign,
5831 }
5837 }
5845 /* 3. Handle explicit realignment if necessary/supported. Create in
5846 loop: vec_dest = realign_load (msq, lsq, realignment_token) */
5849 {
5855 realignment_token);
5862 {
5866 }
5867 }
5869 /* 4. Handle invariant-load. */
5871 {
5875 {
5880 /* CHECKME: bitpos depends on endianess? */
5884 vec_dest =
5893 /* FIXME: use build_constructor directly. */
5897 }
5898 else
5900 }
5902 /* Collect vector loads and later create their permutation in
5903 vect_transform_strided_load (). */
5907 /* Store vector loads in the corresponding SLP_NODE. */
5910 }
5912 /* FORNOW: SLP with multiple types is unsupported. */
5917 {
5923 }
5924 else
5925 {
5928 else
5931 }
5932 }
5938 }
5941 /* Function vectorizable_live_operation.
5943 STMT computes a value that is used outside the loop. Check if
5944 it can be supported. */
5946 bool
5950 {
5951 tree operation;
5957 tree op;
5972 /* FORNOW. CHECKME. */
5979 /* FORNOW: support only if all uses are invariant. This means
5980 that the scalar operations can remain in place, unvectorized.
5981 The original last scalar value that they compute will be used. */
5984 {
5987 {
5991 }
5995 }
5997 /* No transformation is required for the cases we currently support. */
5999 }
6002 /* Function vect_is_simple_cond.
6004 Input:
6005 LOOP - the loop that is being vectorized.
6006 COND - Condition that is checked for simple use.
6008 Returns whether a COND can be vectorized. Checks whether
6009 condition operands are supportable using vec_is_simple_use. */
6011 static bool
6013 {
6015 tree def;
6025 {
6029 }
6035 {
6039 }
6045 }
6047 /* vectorizable_condition.
6049 Check if STMT is conditional modify expression that can be vectorized.
6050 If VEC_STMT is also passed, vectorize the STMT: create a vectorized
6051 stmt using VEC_COND_EXPR to replace it, put it in VEC_STMT, and insert it
6052 at BSI.
6054 Return FALSE if not a vectorizable STMT, TRUE otherwise. */
6056 bool
6058 {
6067 tree new_temp;
6070 tree def;
6085 /* FORNOW: SLP not supported. */
6089 /* FORNOW: not yet supported. */
6091 {
6095 }
6097 /* Is vectorizable conditional operation? */
6113 /* We do not handle two different vector types for the condition
6114 and the values. */
6119 {
6124 }
6131 {
6136 }
6146 {
6149 }
6151 /* Transform */
6153 /* Handle def. */
6157 /* Handle cond expr. */
6158 vec_cond_lhs =
6160 vec_cond_rhs =
6165 /* Arguments are ready. create the new vector stmt. */
6177 }
6180 /* Function vect_transform_stmt.
6182 Create a vectorized stmt to replace STMT, and insert it at BSI. */
6184 static bool
6186 slp_tree slp_node)
6187 {
6191 tree orig_stmt_in_pattern;
6195 {
6238 {
6239 /* In case of interleaving, the whole chain is vectorized when the
6240 last store in the chain is reached. Store stmts before the last
6241 one are skipped, and there vec_stmt_info shouldn't be freed
6242 meanwhile. */
6246 }
6247 else
6270 {
6274 }
6275 }
6279 {
6282 }
6285 {
6289 {
6291 /* STMT was inserted by the vectorizer to replace a computation idiom.
6292 ORIG_STMT_IN_PATTERN is a stmt in the original sequence that
6293 computed this idiom. We need to record a pointer to VEC_STMT in
6294 the stmt_info of ORIG_STMT_IN_PATTERN. See more details in the
6295 documentation of vect_pattern_recog. */
6297 {
6300 }
6301 }
6302 }
6305 }
6308 /* This function builds ni_name = number of iterations loop executes
6309 on the loop preheader. */
6311 static tree
6313 {
6315 edge pe;
6325 {
6328 }
6331 }
6334 /* This function generates the following statements:
6336 ni_name = number of iterations loop executes
6337 ratio = ni_name / vf
6338 ratio_mult_vf_name = ratio * vf
6340 and places them at the loop preheader edge. */
6342 static void
6347 {
6349 edge pe;
6350 basic_block new_bb;
6352 tree var;
6353 tree ratio_name;
6354 tree ratio_mult_vf_name;
6358 tree log_vf;
6362 /* Generate temporary variable that contains
6363 number of iterations loop executes. */
6368 /* Create: ratio = ni >> log2(vf) */
6372 {
6380 }
6382 /* Create: ratio_mult_vf = ratio << log2 (vf). */
6387 {
6396 }
6403 }
6406 /* Function vect_update_ivs_after_vectorizer.
6408 "Advance" the induction variables of LOOP to the value they should take
6409 after the execution of LOOP. This is currently necessary because the
6410 vectorizer does not handle induction variables that are used after the
6411 loop. Such a situation occurs when the last iterations of LOOP are
6412 peeled, because:
6413 1. We introduced new uses after LOOP for IVs that were not originally used
6414 after LOOP: the IVs of LOOP are now used by an epilog loop.
6415 2. LOOP is going to be vectorized; this means that it will iterate N/VF
6416 times, whereas the loop IVs should be bumped N times.
6418 Input:
6419 - LOOP - a loop that is going to be vectorized. The last few iterations
6420 of LOOP were peeled.
6421 - NITERS - the number of iterations that LOOP executes (before it is
6422 vectorized). i.e, the number of times the ivs should be bumped.
6423 - UPDATE_E - a successor edge of LOOP->exit that is on the (only) path
6424 coming out from LOOP on which there are uses of the LOOP ivs
6425 (this is the path from LOOP->exit to epilog_loop->preheader).
6427 The new definitions of the ivs are placed in LOOP->exit.
6428 The phi args associated with the edge UPDATE_E in the bb
6429 UPDATE_E->dest are updated accordingly.
6431 Assumption 1: Like the rest of the vectorizer, this function assumes
6432 a single loop exit that has a single predecessor.
6434 Assumption 2: The phi nodes in the LOOP header and in update_bb are
6435 organized in the same order.
6437 Assumption 3: The access function of the ivs is simple enough (see
6438 vect_can_advance_ivs_p). This assumption will be relaxed in the future.
6440 Assumption 4: Exactly one of the successors of LOOP exit-bb is on a path
6441 coming out of LOOP on which the ivs of LOOP are used (this is the path
6442 that leads to the epilog loop; other paths skip the epilog loop). This
6443 path starts with the edge UPDATE_E, and its destination (denoted update_bb)
6444 needs to have its phis updated.
6445 */
6447 static void
6449 edge update_e)
6450 {
6456 /* gcc_assert (vect_can_advance_ivs_p (loop_vinfo)); */
6458 /* Make sure there exists a single-predecessor exit bb: */
6464 {
6466 tree evolution_part;
6467 tree init_expr;
6468 tree step_expr;
6470 block_stmt_iterator last_bsi;
6473 {
6476 }
6478 /* Skip virtual phi's. */
6480 {
6484 }
6486 /* Skip reduction phis. */
6488 {
6492 }
6496 evolution_part =
6500 /* FORNOW: We do not support IVs whose evolution function is a polynomial
6501 of degree >= 2 or exponential. */
6510 init_expr,
6514 else
6518 niters),
6519 step_expr),
6520 init_expr);
6531 /* Fix phi expressions in the successor bb. */
6533 }
6534 }
6536 /* Return the more conservative threshold between the
6537 min_profitable_iters returned by the cost model and the user
6538 specified threshold, if provided. */
6540 static unsigned int
6543 {
6550 /* Use the cost model only if it is more conservative than user specified
6551 threshold. */
6554 && (!min_scalar_loop_bound
6562 }
6564 /* Function vect_do_peeling_for_loop_bound
6566 Peel the last iterations of the loop represented by LOOP_VINFO.
6567 The peeled iterations form a new epilog loop. Given that the loop now
6568 iterates NITERS times, the new epilog loop iterates
6569 NITERS % VECTORIZATION_FACTOR times.
6571 The original loop will later be made to iterate
6572 NITERS / VECTORIZATION_FACTOR times (this value is placed into RATIO). */
6574 static void
6576 {
6580 edge update_e;
6581 basic_block preheader;
6592 /* Generate the following variables on the preheader of original loop:
6594 ni_name = number of iteration the original loop executes
6595 ratio = ni_name / vf
6596 ratio_mult_vf_name = ratio * vf */
6602 /* If cost model check not done during versioning and
6603 peeling for alignment. */
6607 {
6610 /* Get profitability threshold for vectorized loop. */
6614 min_profitable_iters);
6615 }
6622 #ifdef ENABLE_CHECKING
6624 #endif
6626 /* A guard that controls whether the new_loop is to be executed or skipped
6627 is placed in LOOP->exit. LOOP->exit therefore has two successors - one
6628 is the preheader of NEW_LOOP, where the IVs from LOOP are used. The other
6629 is a bb after NEW_LOOP, where these IVs are not used. Find the edge that
6630 is on the path where the LOOP IVs are used and need to be updated. */
6635 else
6638 /* Update IVs of original loop as if they were advanced
6639 by ratio_mult_vf_name steps. */
6642 /* After peeling we have to reset scalar evolution analyzer. */
6646 }
6649 /* Function vect_gen_niters_for_prolog_loop
6651 Set the number of iterations for the loop represented by LOOP_VINFO
6652 to the minimum between LOOP_NITERS (the original iteration count of the loop)
6653 and the misalignment of DR - the data reference recorded in
6654 LOOP_VINFO_UNALIGNED_DR (LOOP_VINFO). As a result, after the execution of
6655 this loop, the data reference DR will refer to an aligned location.
6657 The following computation is generated:
6659 If the misalignment of DR is known at compile time:
6660 addr_mis = int mis = DR_MISALIGNMENT (dr);
6661 Else, compute address misalignment in bytes:
6662 addr_mis = addr & (vectype_size - 1)
6664 prolog_niters = min ( LOOP_NITERS , (VF - addr_mis/elem_size)&(VF-1) )
6666 (elem_size = element type size; an element is the scalar element
6667 whose type is the inner type of the vectype)
6669 For interleaving,
6671 prolog_niters = min ( LOOP_NITERS ,
6672 (VF/group_size - addr_mis/elem_size)&(VF/group_size-1) )
6673 where group_size is the size of the interleaved group.
6675 The above formulas assume that VF == number of elements in the vector. This
6676 may not hold when there are multiple-types in the loop.
6677 In this case, for some data-references in the loop the VF does not represent
6678 the number of elements that fit in the vector. Therefore, instead of VF we
6679 use TYPE_VECTOR_SUBPARTS. */
6681 static tree
6683 {
6688 edge pe;
6689 basic_block new_bb;
6700 {
6701 /* For interleaved access element size must be multiplied by the size of
6702 the interleaved group. */
6706 }
6711 {
6719 }
6720 else
6721 {
6729 tree elem_size_log =
6733 tree byte_misalign;
6734 tree elem_misalign;
6739 /* Create: byte_misalign = addr & (vectype_size - 1) */
6740 byte_misalign =
6743 /* Create: elem_misalign = byte_misalign / element_size */
6744 elem_misalign =
6747 /* Create: (niters_type) (nelements - elem_misalign)&(nelements - 1) */
6751 }
6753 /* Create: prolog_loop_niters = min (iters, loop_niters) */
6754 /* If the loop bound is known at compile time we already verified that it is
6755 greater than vf; since the misalignment ('iters') is at most vf, there's
6756 no need to generate the MIN_EXPR in this case. */
6761 {
6764 }
6770 /* Insert stmt on loop preheader edge. */
6772 {
6775 }
6778 }
6781 /* Function vect_update_init_of_dr
6783 NITERS iterations were peeled from LOOP. DR represents a data reference
6784 in LOOP. This function updates the information recorded in DR to
6785 account for the fact that the first NITERS iterations had already been
6786 executed. Specifically, it updates the OFFSET field of DR. */
6788 static void
6790 {
6796 }
6799 /* Function vect_update_inits_of_drs
6801 NITERS iterations were peeled from the loop represented by LOOP_VINFO.
6802 This function updates the information recorded for the data references in
6803 the loop to account for the fact that the first NITERS iterations had
6804 already been executed. Specifically, it updates the initial_condition of
6805 the access_function of all the data_references in the loop. */
6807 static void
6809 {
6819 }
6822 /* Function vect_do_peeling_for_alignment
6824 Peel the first 'niters' iterations of the loop represented by LOOP_VINFO.
6825 'niters' is set to the misalignment of one of the data references in the
6826 loop, thereby forcing it to refer to an aligned location at the beginning
6827 of the execution of this loop. The data reference for which we are
6828 peeling is recorded in LOOP_VINFO_UNALIGNED_DR. */
6830 static void
6832 {
6835 tree n_iters;
6850 /* If cost model check not done during versioning. */
6853 {
6856 /* Get profitability threshold for vectorized loop. */
6860 min_profitable_iters);
6861 }
6863 /* Peel the prolog loop and iterate it niters_of_prolog_loop. */
6864 new_loop =
6870 #ifdef ENABLE_CHECKING
6872 #endif
6874 /* Update number of times loop executes. */
6879 /* Update the init conditions of the access functions of all data refs. */
6882 /* After peeling we have to reset scalar evolution analyzer. */
6886 }
6889 /* Function vect_create_cond_for_align_checks.
6891 Create a conditional expression that represents the alignment checks for
6892 all of data references (array element references) whose alignment must be
6893 checked at runtime.
6895 Input:
6896 COND_EXPR - input conditional expression. New conditions will be chained
6897 with logical AND operation.
6898 LOOP_VINFO - two fields of the loop information are used.
6899 LOOP_VINFO_PTR_MASK is the mask used to check the alignment.
6900 LOOP_VINFO_MAY_MISALIGN_STMTS contains the refs to be checked.
6902 Output:
6903 COND_EXPR_STMT_LIST - statements needed to construct the conditional
6904 expression.
6905 The returned value is the conditional expression to be used in the if
6906 statement that controls which version of the loop gets executed at runtime.
6908 The algorithm makes two assumptions:
6909 1) The number of bytes "n" in a vector is a power of 2.
6910 2) An address "a" is aligned if a%n is zero and that this
6911 test can be done as a&(n-1) == 0. For example, for 16
6912 byte vectors the test is a&0xf == 0. */
6914 static void
6918 {
6924 tree mask_cst;
6926 tree psize;
6927 tree int_ptrsize_type;
6931 tree ptrsize_zero;
6932 tree part_cond_expr;
6934 /* Check that mask is one less than a power of 2, i.e., mask is
6935 all zeros followed by all ones. */
6938 /* CHECKME: what is the best integer or unsigned type to use to hold a
6939 cast from a pointer value? */
6941 int_ptrsize_type
6944 /* Create expression (mask & (dr_1 || ... || dr_n)) where dr_i is the address
6945 of the first vector of the i'th data reference. */
6948 {
6950 tree addr_base;
6954 /* create: addr_tmp = (int)(address_of_first_vector) */
6970 /* The addresses are OR together. */
6973 {
6974 /* create: or_tmp = or_tmp | addr_tmp */
6985 }
6986 else
6993 /* create: and_tmp = or_tmp & mask */
7003 /* Make and_tmp the left operand of the conditional test against zero.
7004 if and_tmp has a nonzero bit then some address is unaligned. */
7011 else
7013 }
7015 /* Function vect_vfa_segment_size.
7017 Create an expression that computes the size of segment
7018 that will be accessed for a data reference. The functions takes into
7019 account that realignment loads may access one more vector.
7021 Input:
7022 DR: The data reference.
7023 VECT_FACTOR: vectorization factor.
7025 Return an expression whose value is the size of segment which will be
7026 accessed by DR. */
7028 static tree
7030 {
7035 {
7036 tree vector_size = TYPE_SIZE_UNIT
7041 }
7043 }
7045 /* Function vect_create_cond_for_alias_checks.
7047 Create a conditional expression that represents the run-time checks for
7048 overlapping of address ranges represented by a list of data references
7049 relations passed as input.
7051 Input:
7052 COND_EXPR - input conditional expression. New conditions will be chained
7053 with logical AND operation.
7054 LOOP_VINFO - field LOOP_VINFO_MAY_ALIAS_STMTS contains the list of ddrs
7055 to be checked.
7057 Output:
7058 COND_EXPR - conditional expression.
7059 COND_EXPR_STMT_LIST - statements needed to construct the conditional
7060 expression.
7063 The returned value is the conditional expression to be used in the if
7064 statement that controls which version of the loop gets executed at runtime.
7065 */
7067 static void
7071 {
7075 tree vect_factor =
7078 ddr_p ddr;
7080 tree part_cond_expr;
7082 /* Create expression
7083 ((store_ptr_0 + store_segment_length_0) < load_ptr_0)
7084 || (load_ptr_0 + load_segment_length_0) < store_ptr_0))
7085 &&
7086 ...
7087 &&
7088 ((store_ptr_n + store_segment_length_n) < load_ptr_n)
7089 || (load_ptr_n + load_segment_length_n) < store_ptr_n)) */
7095 {
7106 {
7109 }
7115 {
7118 }
7120 addr_base_a =
7123 addr_base_b =
7131 {
7137 }
7140 part_cond_expr =
7144 addr_base_a,
7145 segment_length_a),
7146 addr_base_b),
7149 addr_base_b,
7150 segment_length_b),
7151 addr_base_a));
7156 else
7158 }
7163 }
7165 /* Function vect_loop_versioning.
7167 If the loop has data references that may or may not be aligned or/and
7168 has data reference relations whose independence was not proven then
7169 two versions of the loop need to be generated, one which is vectorized
7170 and one which isn't. A test is then generated to control which of the
7171 loops is executed. The test checks for the alignment of all of the
7172 data references that may or may not be aligned. An additional
7173 sequence of runtime tests is generated for each pairs of DDRs whose
7174 independence was not proven. The vectorized version of loop is
7175 executed only if both alias and alignment tests are passed.
7177 The test generated to check which version of loop is executed
7178 is modified to also check for profitability as indicated by the
7179 cost model initially. */
7181 static void
7183 {
7188 basic_block condition_bb;
7189 block_stmt_iterator cond_exp_bsi;
7190 basic_block merge_bb;
7191 basic_block new_exit_bb;
7195 tree gimplify_stmt_list;
7200 /* Get profitability threshold for vectorized loop. */
7204 min_profitable_iters);
7206 cond_expr =
7221 cond_expr =
7223 cond_expr =
7225 NULL_TREE);
7233 /* Loop versioning violates an assumption we try to maintain during
7234 vectorization - that the loop exit block has a single predecessor.
7235 After versioning, the exit block of both loop versions is the same
7236 basic block (i.e. it has two predecessors). Just in order to simplify
7237 following transformations in the vectorizer, we fix this situation
7238 here by adding a new (empty) block on the exit-edge of the loop,
7239 with the proper loop-exit phis to maintain loop-closed-form. */
7249 {
7251 new_exit_bb);
7255 }
7257 /* End loop-exit-fixes after versioning. */
7261 {
7264 }
7265 }
7267 /* Remove a group of stores (for SLP or interleaving), free their
7268 stmt_vec_info. */
7270 static void
7272 {
7274 tree tmp;
7275 block_stmt_iterator next_si;
7278 {
7279 /* Free the attached stmt_vec_info and remove the stmt. */
7285 }
7286 }
7289 /* Vectorize SLP instance tree in postorder. */
7291 static bool
7293 {
7294 tree stmt;
7296 block_stmt_iterator si;
7297 stmt_vec_info stmt_info;
7311 {
7314 }
7319 {
7321 /* If IS_STORE is TRUE, the vectorization of the
7322 interleaving chain was completed - free all the stores in
7323 the chain. */
7325 else
7326 /* FORNOW: SLP originates only from strided stores. */
7330 }
7332 /* FORNOW: SLP originates only from strided stores. */
7334 }
7337 static bool
7339 {
7342 slp_instance instance;
7349 {
7351 /* For each SLP instance calculate number of vector stmts to be created
7352 for the scalar stmts in each node of the SLP tree. Number of vector
7353 elements in one vector iteration is the number of scalar elements in
7354 one scalar iteration (GROUP_SIZE) multiplied by VF divided by vector
7355 size. */
7358 /* Schedule the tree of INSTANCE. */
7360 vec_stmts_size);
7365 }
7368 }
7370 /* Function vect_transform_loop.
7372 The analysis phase has determined that the loop is vectorizable.
7373 Vectorize the loop - created vectorized stmts to replace the scalar
7374 stmts in the loop, and update the loop exit condition. */
7376 void
7378 {
7382 block_stmt_iterator si;
7397 /* CHECKME: we wouldn't need this if we called update_ssa once
7398 for all loops. */
7401 /* Peel the loop if there are data refs with unknown alignment.
7402 Only one data ref with unknown store is allowed. */
7407 /* If the loop has a symbolic number of iterations 'n' (i.e. it's not a
7408 compile time constant), or it is a constant that doesn't divide by the
7409 vectorization factor, then an epilog loop needs to be created.
7410 We therefore duplicate the loop: the original loop will be vectorized,
7411 and will compute the first (n/VF) iterations. The second copy of the loop
7412 will remain scalar and will compute the remaining (n%VF) iterations.
7413 (VF is the vectorization factor). */
7419 else
7423 /* 1) Make sure the loop header has exactly two entries
7424 2) Make sure we have a preheader basic block. */
7430 /* FORNOW: the vectorizer supports only loops which body consist
7431 of one basic block (header + empty latch). When the vectorizer will
7432 support more involved loop forms, the order by which the BBs are
7433 traversed need to be reconsidered. */
7436 {
7438 stmt_vec_info stmt_info;
7439 tree phi;
7442 {
7444 {
7447 }
7462 {
7466 }
7467 }
7470 {
7475 {
7478 }
7482 /* vector stmts created in the outer-loop during vectorization of
7483 stmts in an inner-loop may not have a stmt_info, and do not
7484 need to be vectorized. */
7486 {
7489 }
7493 {
7496 }
7499 nunits =
7504 /* For SLP VF is set according to unrolling factor, and not to
7505 vector size, hence for SLP this print is not valid. */
7508 /* SLP. Schedule all the SLP instances when the first SLP stmt is
7509 reached. */
7511 {
7513 {
7521 /* IS_STORE is true if STMT is a store. Stores cannot be of
7522 hybrid SLP type. They are removed in
7523 vect_schedule_slp_instance and their vinfo is destroyed. */
7525 {
7528 }
7529 }
7531 /* Hybrid SLP stmts must be vectorized in addition to SLP. */
7533 {
7536 }
7537 }
7539 /* -------- vectorize statement ------------ */
7546 {
7548 {
7549 /* Interleaving. If IS_STORE is TRUE, the vectorization of the
7550 interleaving chain was completed - free all the stores in
7551 the chain. */
7555 }
7556 else
7557 {
7558 /* Free the attached stmt_vec_info and remove the stmt. */
7562 }
7563 }
7572 /* The memory tags and pointers in vectorized statements need to
7573 have their SSA forms updated. FIXME, why can't this be delayed
7574 until all the loops have been transformed? */
7581 }