| blob | 352939f03b24179ead2bc79f4711fcf11d93e8d7 |
1 /* Transformation Utilities for Loop Vectorization.
2 Copyright (C) 2003, 2004, 2005, 2006, 2007 Free Software Foundation, Inc.
3 Contributed by Dorit Naishlos <dorit@il.ibm.com>
5 This file is part of GCC.
7 GCC is free software; you can redistribute it and/or modify it under
8 the terms of the GNU General Public License as published by the Free
9 Software Foundation; either version 3, or (at your option) any later
10 version.
12 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
13 WARRANTY; without even the implied warranty of MERCHANTABILITY or
14 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
15 for more details.
17 You should have received a copy of the GNU General Public License
18 along with GCC; see the file COPYING3. If not see
19 <http://www.gnu.org/licenses/>. */
48 /* Utility functions for the code transformation. */
51 static tree vect_create_data_ref_ptr
53 static tree vect_create_addr_base_for_vector_ref
58 static void vect_finish_stmt_generation
64 /* Utility function dealing with loop peeling (not peeling itself). */
65 static void vect_generate_tmps_on_preheader
75 static int
77 {
81 {
99 }
100 }
103 /* Function vect_estimate_min_profitable_iters
105 Return the number of iterations required for the vector version of the
106 loop to be profitable relative to the cost of the scalar version of the
107 loop.
109 TODO: Take profile info into account before making vectorization
110 decisions, if available. */
112 int
114 {
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. */
2114 #ifdef USE_MAPPED_LOCATION
2116 #else
2118 #endif
2119 }
2122 /* Function get_initial_def_for_reduction
2124 Input:
2125 STMT - a stmt that performs a reduction operation in the loop.
2126 INIT_VAL - the initial value of the reduction variable
2128 Output:
2129 ADJUSTMENT_DEF - a tree that holds a value to be added to the final result
2130 of the reduction (used for adjusting the epilog - see below).
2131 Return a vector variable, initialized according to the operation that STMT
2132 performs. This vector will be used as the initial value of the
2133 vector of partial results.
2135 Option1 (adjust in epilog): Initialize the vector as follows:
2136 add: [0,0,...,0,0]
2137 mult: [1,1,...,1,1]
2138 min/max: [init_val,init_val,..,init_val,init_val]
2139 bit and/or: [init_val,init_val,..,init_val,init_val]
2140 and when necessary (e.g. add/mult case) let the caller know
2141 that it needs to adjust the result by init_val.
2143 Option2: Initialize the vector as follows:
2144 add: [0,0,...,0,init_val]
2145 mult: [1,1,...,1,init_val]
2146 min/max: [init_val,init_val,...,init_val]
2147 bit and/or: [init_val,init_val,...,init_val]
2148 and no adjustments are needed.
2150 For example, for the following code:
2152 s = init_val;
2153 for (i=0;i<n;i++)
2154 s = s + a[i];
2156 STMT is 's = s + a[i]', and the reduction variable is 's'.
2157 For a vector of 4 units, we want to return either [0,0,0,init_val],
2158 or [0,0,0,0] and let the caller know that it needs to adjust
2159 the result at the end by 'init_val'.
2161 FORNOW, we are using the 'adjust in epilog' scheme, because this way the
2162 initialization vector is simpler (same element in all entries).
2163 A cost model should help decide between these two schemes. */
2165 static tree
2167 {
2175 tree vecdef;
2176 tree def_for_init;
2177 tree init_def;
2180 tree vector_type;
2186 else
2192 {
2198 else
2200 /* Create a vector of zeros for init_def. */
2203 else
2220 }
2223 }
2226 /* Function vect_create_epilog_for_reduction
2228 Create code at the loop-epilog to finalize the result of a reduction
2229 computation.
2231 VECT_DEF is a vector of partial results.
2232 REDUC_CODE is the tree-code for the epilog reduction.
2233 STMT is the scalar reduction stmt that is being vectorized.
2234 REDUCTION_PHI is the phi-node that carries the reduction computation.
2236 This function:
2237 1. Creates the reduction def-use cycle: sets the arguments for
2238 REDUCTION_PHI:
2239 The loop-entry argument is the vectorized initial-value of the reduction.
2240 The loop-latch argument is VECT_DEF - the vector of partial sums.
2241 2. "Reduces" the vector of partial results VECT_DEF into a single result,
2242 by applying the operation specified by REDUC_CODE if available, or by
2243 other means (whole-vector shifts or a scalar loop).
2244 The function also creates a new phi node at the loop exit to preserve
2245 loop-closed form, as illustrated below.
2247 The flow at the entry to this function:
2249 loop:
2250 vec_def = phi <null, null> # REDUCTION_PHI
2251 VECT_DEF = vector_stmt # vectorized form of STMT
2252 s_loop = scalar_stmt # (scalar) STMT
2253 loop_exit:
2254 s_out0 = phi <s_loop> # (scalar) EXIT_PHI
2255 use <s_out0>
2256 use <s_out0>
2258 The above is transformed by this function into:
2260 loop:
2261 vec_def = phi <vec_init, VECT_DEF> # REDUCTION_PHI
2262 VECT_DEF = vector_stmt # vectorized form of STMT
2263 s_loop = scalar_stmt # (scalar) STMT
2264 loop_exit:
2265 s_out0 = phi <s_loop> # (scalar) EXIT_PHI
2266 v_out1 = phi <VECT_DEF> # NEW_EXIT_PHI
2267 v_out2 = reduce <v_out1>
2268 s_out3 = extract_field <v_out2, 0>
2269 s_out4 = adjust_result <s_out3>
2270 use <s_out4>
2271 use <s_out4>
2272 */
2274 static void
2277 {
2279 tree vectype;
2283 basic_block exit_bb;
2284 tree scalar_dest;
2285 tree scalar_type;
2286 tree new_phi;
2287 block_stmt_iterator exit_bsi;
2288 tree vec_dest;
2290 tree new_name;
2295 tree adjustment_def;
2296 tree vec_initial_def;
2297 tree orig_name;
2298 imm_use_iterator imm_iter;
2299 use_operand_p use_p;
2302 tree orig_stmt;
2303 tree use_stmt;
2311 {
2314 }
2322 /*** 1. Create the reduction def-use cycle ***/
2324 /* 1.1 set the loop-entry arg of the reduction-phi: */
2325 /* For the case of reduction, vect_get_vec_def_for_operand returns
2326 the scalar def before the loop, that defines the initial value
2327 of the reduction variable. */
2332 /* 1.2 set the loop-latch arg for the reduction-phi: */
2336 {
2341 }
2344 /*** 2. Create epilog code
2345 The reduction epilog code operates across the elements of the vector
2346 of partial results computed by the vectorized loop.
2347 The reduction epilog code consists of:
2348 step 1: compute the scalar result in a vector (v_out2)
2349 step 2: extract the scalar result (s_out3) from the vector (v_out2)
2350 step 3: adjust the scalar result (s_out3) if needed.
2352 Step 1 can be accomplished using one the following three schemes:
2353 (scheme 1) using reduc_code, if available.
2354 (scheme 2) using whole-vector shifts, if available.
2355 (scheme 3) using a scalar loop. In this case steps 1+2 above are
2356 combined.
2358 The overall epilog code looks like this:
2360 s_out0 = phi <s_loop> # original EXIT_PHI
2361 v_out1 = phi <VECT_DEF> # NEW_EXIT_PHI
2362 v_out2 = reduce <v_out1> # step 1
2363 s_out3 = extract_field <v_out2, 0> # step 2
2364 s_out4 = adjust_result <s_out3> # step 3
2366 (step 3 is optional, and step2 1 and 2 may be combined).
2367 Lastly, the uses of s_out0 are replaced by s_out4.
2369 ***/
2371 /* 2.1 Create new loop-exit-phi to preserve loop-closed form:
2372 v_out1 = phi <v_loop> */
2379 /* 2.2 Get the relevant tree-code to use in the epilog for schemes 2,3
2380 (i.e. when reduc_code is not available) and in the final adjustment
2381 code (if needed). Also get the original scalar reduction variable as
2382 defined in the loop. In case STMT is a "pattern-stmt" (i.e. - it
2383 represents a reduction pattern), the tree-code and scalar-def are
2384 taken from the original stmt that the pattern-stmt (STMT) replaces.
2385 Otherwise (it is a regular reduction) - the tree-code and scalar-def
2386 are taken from STMT. */
2390 {
2391 /* Regular reduction */
2393 }
2394 else
2395 {
2396 /* Reduction pattern */
2400 }
2409 /* In case this is a reduction in an inner-loop while vectorizing an outer
2410 loop - we don't need to extract a single scalar result at the end of the
2411 inner-loop. The final vector of partial results will be used in the
2412 vectorized outer-loop, or reduced to a scalar result at the end of the
2413 outer-loop. */
2417 /* 2.3 Create the reduction code, using one of the three schemes described
2418 above. */
2421 {
2422 tree tmp;
2424 /*** Case 1: Create:
2425 v_out2 = reduc_expr <v_out1> */
2438 }
2439 else
2440 {
2446 tree vec_temp;
2450 else
2453 /* Regardless of whether we have a whole vector shift, if we're
2454 emulating the operation via tree-vect-generic, we don't want
2455 to use it. Only the first round of the reduction is likely
2456 to still be profitable via emulation. */
2457 /* ??? It might be better to emit a reduction tree code here, so that
2458 tree-vect-generic can expand the first round via bit tricks. */
2461 else
2462 {
2466 }
2469 {
2470 /*** Case 2: Create:
2471 for (offset = VS/2; offset >= element_size; offset/=2)
2472 {
2473 Create: va' = vec_shift <va, offset>
2474 Create: va = vop <va, va'>
2475 } */
2486 {
2499 }
2502 }
2503 else
2504 {
2505 tree rhs;
2507 /*** Case 3: Create:
2508 s = extract_field <v_out2, 0>
2509 for (offset = element_size;
2510 offset < vector_size;
2511 offset += element_size;)
2512 {
2513 Create: s' = extract_field <v_out2, offset>
2514 Create: s = op <s, s'>
2515 } */
2523 bitsize_zero_node);
2533 {
2534 tree tmp;
2537 bitpos);
2550 }
2553 }
2554 }
2556 /* 2.4 Extract the final scalar result. Create:
2557 s_out3 = extract_field <v_out2, bitpos> */
2560 {
2561 tree rhs;
2571 else
2580 }
2582 vect_finalize_reduction:
2584 /* 2.5 Adjust the final result by the initial value of the reduction
2585 variable. (When such adjustment is not needed, then
2586 'adjustment_def' is zero). For example, if code is PLUS we create:
2587 new_temp = loop_exit_def + adjustment_def */
2590 {
2592 {
2596 }
2597 else
2598 {
2602 }
2607 }
2610 /* 2.6 Handle the loop-exit phi */
2612 /* Replace uses of s_out0 with uses of s_out3:
2613 Find the loop-closed-use at the loop exit of the original scalar result.
2614 (The reduction result is expected to have two immediate uses - one at the
2615 latch block, and one at the loop exit). */
2618 {
2620 {
2623 }
2624 }
2625 /* We expect to have found an exit_phi because of loop-closed-ssa form. */
2629 {
2631 {
2634 /* FORNOW. Currently not supporting the case that an inner-loop reduction
2635 is not used in the outer-loop (but only outside the outer-loop). */
2644 }
2646 /* Replace the uses: */
2651 }
2653 }
2656 /* Function vectorizable_reduction.
2658 Check if STMT performs a reduction operation that can be vectorized.
2659 If VEC_STMT is also passed, vectorize the STMT: create a vectorized
2660 stmt to replace it, put it in VEC_STMT, and insert it at BSI.
2661 Return FALSE if not a vectorizable STMT, TRUE otherwise.
2663 This function also handles reduction idioms (patterns) that have been
2664 recognized in advance during vect_pattern_recog. In this case, STMT may be
2665 of this form:
2666 X = pattern_expr (arg0, arg1, ..., X)
2667 and it's STMT_VINFO_RELATED_STMT points to the last stmt in the original
2668 sequence that had been detected and replaced by the pattern-stmt (STMT).
2670 In some cases of reduction patterns, the type of the reduction variable X is
2671 different than the type of the other arguments of STMT.
2672 In such cases, the vectype that is used when transforming STMT into a vector
2673 stmt is different than the vectype that is used to determine the
2674 vectorization factor, because it consists of a different number of elements
2675 than the actual number of elements that are being operated upon in parallel.
2677 For example, consider an accumulation of shorts into an int accumulator.
2678 On some targets it's possible to vectorize this pattern operating on 8
2679 shorts at a time (hence, the vectype for purposes of determining the
2680 vectorization factor should be V8HI); on the other hand, the vectype that
2681 is used to create the vector form is actually V4SI (the type of the result).
2683 Upon entry to this function, STMT_VINFO_VECTYPE records the vectype that
2684 indicates what is the actual level of parallelism (V8HI in the example), so
2685 that the right vectorization factor would be derived. This vectype
2686 corresponds to the type of arguments to the reduction stmt, and should *NOT*
2687 be used to create the vectorized stmt. The right vectype for the vectorized
2688 stmt is obtained from the type of the result X:
2689 get_vectype_for_scalar_type (TREE_TYPE (X))
2691 This means that, contrary to "regular" reductions (or "regular" stmts in
2692 general), the following equation:
2693 STMT_VINFO_VECTYPE == get_vectype_for_scalar_type (TREE_TYPE (X))
2694 does *NOT* necessarily hold for reduction patterns. */
2696 bool
2698 {
2699 tree vec_dest;
2700 tree scalar_dest;
2701 tree op;
2707 tree operation;
2715 tree new_phi;
2716 tree scalar_type;
2718 tree orig_stmt;
2719 stmt_vec_info orig_stmt_info;
2724 stmt_vec_info prev_stmt_info;
2725 tree reduc_def;
2730 {
2732 /* FORNOW. This restriction should be relaxed. */
2734 {
2738 }
2739 }
2743 /* FORNOW: SLP not supported. */
2747 /* 1. Is vectorizable reduction? */
2749 /* Not supportable if the reduction variable is used in the loop. */
2753 /* Reductions that are not used even in an enclosing outer-loop,
2754 are expected to be "live" (used out of the loop). */
2759 /* Make sure it was already recognized as a reduction computation. */
2763 /* 2. Has this been recognized as a reduction pattern?
2765 Check if STMT represents a pattern that has been recognized
2766 in earlier analysis stages. For stmts that represent a pattern,
2767 the STMT_VINFO_RELATED_STMT field records the last stmt in
2768 the original sequence that constitutes the pattern. */
2772 {
2777 }
2779 /* 3. Check the operands of the operation. The first operands are defined
2780 inside the loop body. The last operand is the reduction variable,
2781 which is defined by the loop-header-phi. */
2796 /* All uses but the last are expected to be defined in the loop.
2797 The last use is the reduction variable. */
2799 {
2808 }
2817 else
2823 /* 4. Supportable by target? */
2825 /* 4.1. check support for the operation in the loop */
2828 {
2832 }
2835 {
2844 }
2846 /* Worthwhile without SIMD support? */
2850 {
2854 }
2856 /* 4.2. Check support for the epilog operation.
2858 If STMT represents a reduction pattern, then the type of the
2859 reduction variable may be different than the type of the rest
2860 of the arguments. For example, consider the case of accumulation
2861 of shorts into an int accumulator; The original code:
2862 S1: int_a = (int) short_a;
2863 orig_stmt-> S2: int_acc = plus <int_a ,int_acc>;
2865 was replaced with:
2866 STMT: int_acc = widen_sum <short_a, int_acc>
2868 This means that:
2869 1. The tree-code that is used to create the vector operation in the
2870 epilog code (that reduces the partial results) is not the
2871 tree-code of STMT, but is rather the tree-code of the original
2872 stmt from the pattern that STMT is replacing. I.e, in the example
2873 above we want to use 'widen_sum' in the loop, but 'plus' in the
2874 epilog.
2875 2. The type (mode) we use to check available target support
2876 for the vector operation to be created in the *epilog*, is
2877 determined by the type of the reduction variable (in the example
2878 above we'd check this: plus_optab[vect_int_mode]).
2879 However the type (mode) we use to check available target support
2880 for the vector operation to be created *inside the loop*, is
2881 determined by the type of the other arguments to STMT (in the
2882 example we'd check this: widen_sum_optab[vect_short_mode]).
2884 This is contrary to "regular" reductions, in which the types of all
2885 the arguments are the same as the type of the reduction variable.
2886 For "regular" reductions we can therefore use the same vector type
2887 (and also the same tree-code) when generating the epilog code and
2888 when generating the code inside the loop. */
2891 {
2892 /* This is a reduction pattern: get the vectype from the type of the
2893 reduction variable, and get the tree-code from orig_stmt. */
2897 {
2899 {
2902 }
2904 }
2907 }
2908 else
2909 {
2910 /* Regular reduction: use the same vectype and tree-code as used for
2911 the vector code inside the loop can be used for the epilog code. */
2913 }
2919 {
2923 }
2925 {
2929 }
2932 {
2937 }
2939 /** Transform. **/
2944 /* Create the destination vector */
2947 /* Create the reduction-phi that defines the reduction-operand. */
2950 /* In case the vectorization factor (VF) is bigger than the number
2951 of elements that we can fit in a vectype (nunits), we have to generate
2952 more than one vector stmt - i.e - we need to "unroll" the
2953 vector stmt by a factor VF/nunits. For more details see documentation
2954 in vectorizable_operation. */
2958 {
2959 /* Handle uses. */
2961 {
2965 {
2968 }
2970 /* Get the vector def for the reduction variable from the phi node */
2972 }
2973 else
2974 {
2980 /* Get the vector def for the reduction variable from the vectorized
2981 reduction operation generated in the previous iteration (j-1) */
2983 }
2985 /* Arguments are ready. create the new vector stmt. */
2988 else
2990 reduc_def);
2998 else
3001 }
3003 /* Finalize the reduction-phi (set it's arguments) and create the
3004 epilog reduction code. */
3007 }
3009 /* Checks if CALL can be vectorized in type VECTYPE. Returns
3010 a function declaration if the target has a vectorized version
3011 of the function, or NULL_TREE if the function cannot be vectorized. */
3013 tree
3015 {
3019 /* We only handle functions that do not read or clobber memory -- i.e.
3020 const or novops ones. */
3031 vectype_in);
3032 }
3034 /* Function vectorizable_call.
3036 Check if STMT performs a function call that can be vectorized.
3037 If VEC_STMT is also passed, vectorize the STMT: create a vectorized
3038 stmt to replace it, put it in VEC_STMT, and insert it at BSI.
3039 Return FALSE if not a vectorizable STMT, TRUE otherwise. */
3041 bool
3043 {
3044 tree vec_dest;
3045 tree scalar_dest;
3046 tree operation;
3057 tree new_stmt;
3059 call_expr_arg_iterator iter;
3060 tree vargs;
3069 /* FORNOW: SLP not supported. */
3073 /* Is STMT a vectorizable call? */
3084 /* Process function arguments. */
3088 {
3089 /* Bail out if the function has more than two arguments, we
3090 do not have interesting builtin functions to vectorize with
3091 more than two arguments. */
3095 /* We can only handle calls with arguments of the same type. */
3098 {
3102 }
3106 {
3110 }
3113 }
3115 /* No arguments is also not good. */
3130 /* FORNOW */
3137 else
3140 /* For now, we only vectorize functions if a target specific builtin
3141 is available. TODO -- in some cases, it might be profitable to
3142 insert the calls for pieces of the vector, in order to be able
3143 to vectorize other operations in the loop. */
3146 {
3151 }
3157 else
3160 /* Sanity check: make sure that at least one copy of the vectorized stmt
3161 needs to be generated. */
3164 /* FORNOW. This restriction should be relaxed. */
3166 {
3170 }
3173 {
3179 }
3181 /** Transform. **/
3186 /* FORNOW. This restriction should be relaxed. */
3188 {
3192 }
3194 /* Handle def. */
3200 {
3203 {
3204 /* Build argument list for the vectorized call. */
3205 /* FIXME: Rewrite this so that it doesn't
3206 construct a temporary list. */
3210 {
3212 vec_oprnd0
3214 else
3215 vec_oprnd0
3221 }
3233 else
3237 }
3243 {
3244 /* Build argument list for the vectorized call. */
3245 /* FIXME: Rewrite this so that it doesn't
3246 construct a temporary list. */
3250 {
3252 {
3253 vec_oprnd0
3255 vec_oprnd1
3257 }
3258 else
3259 {
3260 vec_oprnd0
3262 vec_oprnd1
3264 }
3270 }
3282 else
3286 }
3293 /* No current target implements this case. */
3295 }
3297 /* The call in STMT might prevent it from being removed in dce.
3298 We however cannot remove it here, due to the way the ssa name
3299 it defines is mapped to the new definition. So just replace
3300 rhs of the statement with something harmless. */
3306 }
3309 /* Function vect_gen_widened_results_half
3311 Create a vector stmt whose code, type, number of arguments, and result
3312 variable are CODE, VECTYPE, OP_TYPE, and VEC_DEST, and its arguments are
3313 VEC_OPRND0 and VEC_OPRND1. The new vector stmt is to be inserted at BSI.
3314 In the case that CODE is a CALL_EXPR, this means that a call to DECL
3315 needs to be created (DECL is a function-decl of a target-builtin).
3316 STMT is the original scalar stmt that we are vectorizing. */
3318 static tree
3322 tree stmt)
3323 {
3324 tree expr;
3325 tree new_stmt;
3326 tree new_temp;
3327 tree sym;
3328 ssa_op_iter iter;
3330 /* Generate half of the widened result: */
3332 {
3333 /* Target specific support */
3336 else
3338 }
3339 else
3340 {
3341 /* Generic support */
3345 else
3347 }
3354 {
3356 {
3360 }
3361 }
3364 }
3367 /* Check if STMT performs a conversion operation, that can be vectorized.
3368 If VEC_STMT is also passed, vectorize the STMT: create a vectorized
3369 stmt to replace it, put it in VEC_STMT, and insert it at BSI.
3370 Return FALSE if not a vectorizable STMT, TRUE otherwise. */
3372 bool
3375 {
3376 tree vec_dest;
3377 tree scalar_dest;
3378 tree operation;
3379 tree op0;
3386 tree new_temp;
3390 stmt_vec_info prev_stmt_info;
3395 tree expr;
3397 tree builtin_decl;
3401 tree vop0;
3403 /* Is STMT a vectorizable conversion? */
3422 /* Check types of lhs and rhs. */
3437 /* FORNOW */
3444 else
3450 /* Bail out if the types are both integral or non-integral. */
3457 else
3460 /* FORNOW: SLP with multiple types is not supported. The SLP analysis verifies
3461 this, so we can safely override NCOPIES with 1 here. */
3465 /* Sanity check: make sure that at least one copy of the vectorized stmt
3466 needs to be generated. */
3469 /* FORNOW. This restriction should be relaxed. */
3471 {
3475 }
3477 /* Check the operands of the operation. */
3479 {
3483 }
3485 /* Supportable by target? */
3495 {
3499 }
3502 {
3504 /* FORNOW: SLP not supported. */
3507 }
3510 {
3513 }
3515 /** Transform. **/
3519 /* Handle def. */
3527 {
3530 {
3531 tree sym;
3532 ssa_op_iter iter;
3536 else
3539 builtin_decl =
3542 {
3545 /* Arguments are ready. create the new vector stmt. */
3551 SSA_OP_ALL_VIRTUALS)
3552 {
3556 }
3559 }
3563 else
3566 }
3570 /* In case the vectorization factor (VF) is bigger than the number
3571 of elements that we can fit in a vectype (nunits), we have to
3572 generate more than one vector stmt - i.e - we need to "unroll"
3573 the vector stmt by a factor VF/nunits. */
3575 {
3578 else
3583 /* Generate first half of the widened result: */
3584 new_stmt
3590 else
3594 /* Generate second half of the widened result: */
3595 new_stmt
3601 }
3605 /* In case the vectorization factor (VF) is bigger than the number
3606 of elements that we can fit in a vectype (nunits), we have to
3607 generate more than one vector stmt - i.e - we need to "unroll"
3608 the vector stmt by a factor VF/nunits. */
3610 {
3611 /* Handle uses. */
3613 {
3616 }
3617 else
3618 {
3621 }
3623 /* Arguments are ready. Create the new vector stmt. */
3632 else
3636 }
3639 }
3642 }
3645 /* Function vectorizable_assignment.
3647 Check if STMT performs an assignment (copy) that can be vectorized.
3648 If VEC_STMT is also passed, vectorize the STMT: create a vectorized
3649 stmt to replace it, put it in VEC_STMT, and insert it at BSI.
3650 Return FALSE if not a vectorizable STMT, TRUE otherwise. */
3652 bool
3654 slp_tree slp_node)
3655 {
3656 tree vec_dest;
3657 tree scalar_dest;
3658 tree op;
3662 tree new_temp;
3669 tree vop;
3681 /* Is vectorizable assignment? */
3691 {
3695 }
3698 {
3704 }
3706 /** Transform. **/
3710 /* Handle def. */
3713 /* Handle use. */
3716 /* Arguments are ready. create the new vector stmt. */
3718 {
3727 }
3731 }
3734 /* Function vect_min_worthwhile_factor.
3736 For a loop where we could vectorize the operation indicated by CODE,
3737 return the minimum vectorization factor that makes it worthwhile
3738 to use generic vectors. */
3739 static int
3741 {
3743 {
3757 }
3758 }
3761 /* Function vectorizable_induction
3763 Check if PHI performs an induction computation that can be vectorized.
3764 If VEC_STMT is also passed, vectorize the induction PHI: create a vectorized
3765 phi to replace it, put it in VEC_STMT, and add it to the same basic block.
3766 Return FALSE if not a vectorizable STMT, TRUE otherwise. */
3768 bool
3771 {
3777 tree vec_def;
3784 /* FORNOW: SLP not supported. */
3794 {
3800 }
3802 /** Transform. **/
3810 }
3813 /* Function vectorizable_operation.
3815 Check if STMT performs a binary or unary operation that can be vectorized.
3816 If VEC_STMT is also passed, vectorize the STMT: create a vectorized
3817 stmt to replace it, put it in VEC_STMT, and insert it at BSI.
3818 Return FALSE if not a vectorizable STMT, TRUE otherwise. */
3820 bool
3822 slp_tree slp_node)
3823 {
3824 tree vec_dest;
3825 tree scalar_dest;
3826 tree operation;
3835 tree new_temp;
3837 optab optab;
3843 stmt_vec_info prev_stmt_info;
3846 tree vectype_out;
3854 /* FORNOW: SLP with multiple types is not supported. The SLP analysis verifies
3855 this, so we can safely override NCOPIES with 1 here. */
3859 /* FORNOW. This restriction should be relaxed. */
3861 {
3865 }
3873 /* Is STMT a vectorizable binary/unary operation? */
3891 /* For pointer addition, we should use the normal plus for
3892 the vector addition. */
3898 /* Support only unary or binary operations. */
3901 {
3905 }
3909 {
3913 }
3916 {
3919 {
3923 }
3924 }
3926 /* Supportable by target? */
3928 {
3932 }
3936 {
3939 /* Check only during analysis. */
3947 }
3949 /* Worthwhile without SIMD support? Check only during analysis. */
3954 {
3958 }
3961 {
3962 /* FORNOW: not yet supported. */
3966 /* Invariant argument is needed for a vector shift
3967 by a scalar shift operand. */
3970 {
3972 {
3977 }
3980 }
3981 }
3984 {
3990 }
3992 /** Transform. **/
3997 /* Handle def. */
4000 /* Allocate VECs for vector operands. In case of SLP, vector operands are
4001 created in the previous stages of the recursion, so no allocation is
4002 needed, except for the case of shift with scalar shift argument. In that
4003 case we store the scalar operand in VEC_OPRNDS1 for every vector stmt to
4004 be created to vectorize the SLP group, i.e., SLP_NODE->VEC_STMTS_SIZE.
4005 In case of loop-based vectorization we allocate VECs of size 1. We
4006 allocate VEC_OPRNDS1 only in case of binary operation. */
4008 {
4012 }
4016 /* In case the vectorization factor (VF) is bigger than the number
4017 of elements that we can fit in a vectype (nunits), we have to generate
4018 more than one vector stmt - i.e - we need to "unroll" the
4019 vector stmt by a factor VF/nunits. In doing so, we record a pointer
4020 from one copy of the vector stmt to the next, in the field
4021 STMT_VINFO_RELATED_STMT. This is necessary in order to allow following
4022 stages to find the correct vector defs to be used when vectorizing
4023 stmts that use the defs of the current stmt. The example below illustrates
4024 the vectorization process when VF=16 and nunits=4 (i.e - we need to create
4025 4 vectorized stmts):
4027 before vectorization:
4028 RELATED_STMT VEC_STMT
4029 S1: x = memref - -
4030 S2: z = x + 1 - -
4032 step 1: vectorize stmt S1 (done in vectorizable_load. See more details
4033 there):
4034 RELATED_STMT VEC_STMT
4035 VS1_0: vx0 = memref0 VS1_1 -
4036 VS1_1: vx1 = memref1 VS1_2 -
4037 VS1_2: vx2 = memref2 VS1_3 -
4038 VS1_3: vx3 = memref3 - -
4039 S1: x = load - VS1_0
4040 S2: z = x + 1 - -
4042 step2: vectorize stmt S2 (done here):
4043 To vectorize stmt S2 we first need to find the relevant vector
4044 def for the first operand 'x'. This is, as usual, obtained from
4045 the vector stmt recorded in the STMT_VINFO_VEC_STMT of the stmt
4046 that defines 'x' (S1). This way we find the stmt VS1_0, and the
4047 relevant vector def 'vx0'. Having found 'vx0' we can generate
4048 the vector stmt VS2_0, and as usual, record it in the
4049 STMT_VINFO_VEC_STMT of stmt S2.
4050 When creating the second copy (VS2_1), we obtain the relevant vector
4051 def from the vector stmt recorded in the STMT_VINFO_RELATED_STMT of
4052 stmt VS1_0. This way we find the stmt VS1_1 and the relevant
4053 vector def 'vx1'. Using 'vx1' we create stmt VS2_1 and record a
4054 pointer to it in the STMT_VINFO_RELATED_STMT of the vector stmt VS2_0.
4055 Similarly when creating stmts VS2_2 and VS2_3. This is the resulting
4056 chain of stmts and pointers:
4057 RELATED_STMT VEC_STMT
4058 VS1_0: vx0 = memref0 VS1_1 -
4059 VS1_1: vx1 = memref1 VS1_2 -
4060 VS1_2: vx2 = memref2 VS1_3 -
4061 VS1_3: vx3 = memref3 - -
4062 S1: x = load - VS1_0
4063 VS2_0: vz0 = vx0 + v1 VS2_1 -
4064 VS2_1: vz1 = vx1 + v1 VS2_2 -
4065 VS2_2: vz2 = vx2 + v1 VS2_3 -
4066 VS2_3: vz3 = vx3 + v1 - -
4067 S2: z = x + 1 - VS2_0 */
4071 {
4072 /* Handle uses. */
4074 {
4077 {
4078 /* Vector shl and shr insn patterns can be defined with scalar
4079 operand 2 (shift operand). In this case, use constant or loop
4080 invariant op1 directly, without extending it to vector mode
4081 first. */
4084 {
4090 {
4091 /* Store vec_oprnd1 for every vector stmt to be created
4092 for SLP_NODE. We check during the analysis that all the
4093 shift arguments are the same.
4094 TODO: Allow different constants for different vector
4095 stmts generated for an SLP instance. */
4098 }
4099 }
4100 }
4102 /* vec_oprnd1 is available if operand 1 should be of a scalar-type
4103 (a special case for certain kind of vector shifts); otherwise,
4104 operand 1 should be of a vector type (the usual case). */
4107 slp_node);
4108 else
4110 slp_node);
4111 }
4112 else
4115 /* Arguments are ready. Create the new vector stmt. */
4117 {
4119 {
4123 }
4124 else
4133 }
4137 else
4140 }
4147 }
4150 /* Function vectorizable_type_demotion
4152 Check if STMT performs a binary or unary operation that involves
4153 type demotion, and if it can be vectorized.
4154 If VEC_STMT is also passed, vectorize the STMT: create a vectorized
4155 stmt to replace it, put it in VEC_STMT, and insert it at BSI.
4156 Return FALSE if not a vectorizable STMT, TRUE otherwise. */
4158 bool
4161 {
4162 tree vec_dest;
4163 tree scalar_dest;
4164 tree operation;
4165 tree op0;
4171 tree new_temp;
4174 tree new_stmt;
4175 stmt_vec_info prev_stmt_info;
4178 tree vectype_out;
4181 tree expr;
4182 tree vectype_in;
4190 /* Is STMT a vectorizable type-demotion operation? */
4218 /* FORNOW. This restriction should be relaxed. */
4220 {
4224 }
4233 /* Check the operands of the operation. */
4235 {
4239 }
4241 /* Supportable by target? */
4248 {
4254 }
4256 /** Transform. **/
4259 ncopies);
4261 /* Handle def. */
4264 /* In case the vectorization factor (VF) is bigger than the number
4265 of elements that we can fit in a vectype (nunits), we have to generate
4266 more than one vector stmt - i.e - we need to "unroll" the
4267 vector stmt by a factor VF/nunits. */
4270 {
4271 /* Handle uses. */
4273 {
4276 }
4277 else
4278 {
4281 }
4283 /* Arguments are ready. Create the new vector stmt. */
4292 else
4296 }
4300 }
4303 /* Function vectorizable_type_promotion
4305 Check if STMT performs a binary or unary operation that involves
4306 type promotion, and if it can be vectorized.
4307 If VEC_STMT is also passed, vectorize the STMT: create a vectorized
4308 stmt to replace it, put it in VEC_STMT, and insert it at BSI.
4309 Return FALSE if not a vectorizable STMT, TRUE otherwise. */
4311 bool
4314 {
4315 tree vec_dest;
4316 tree scalar_dest;
4317 tree operation;
4328 tree new_stmt;
4329 stmt_vec_info prev_stmt_info;
4332 tree vectype_out;
4335 tree vectype_in;
4343 /* Is STMT a vectorizable type-promotion operation? */
4372 /* FORNOW. This restriction should be relaxed. */
4374 {
4378 }
4387 /* Check the operands of the operation. */
4389 {
4393 }
4397 {
4400 {
4404 }
4405 }
4407 /* Supportable by target? */
4415 {
4421 }
4423 /** Transform. **/
4427 ncopies);
4429 /* Handle def. */
4432 /* In case the vectorization factor (VF) is bigger than the number
4433 of elements that we can fit in a vectype (nunits), we have to generate
4434 more than one vector stmt - i.e - we need to "unroll" the
4435 vector stmt by a factor VF/nunits. */
4439 {
4440 /* Handle uses. */
4442 {
4446 }
4447 else
4448 {
4452 }
4454 /* Arguments are ready. Create the new vector stmt. We are creating
4455 two vector defs because the widened result does not fit in one vector.
4456 The vectorized stmt can be expressed as a call to a taregt builtin,
4457 or a using a tree-code. */
4458 /* Generate first half of the widened result: */
4463 else
4467 /* Generate second half of the widened result: */
4473 }
4477 }
4480 /* Function vect_strided_store_supported.
4482 Returns TRUE is INTERLEAVE_HIGH and INTERLEAVE_LOW operations are supported,
4483 and FALSE otherwise. */
4485 static bool
4487 {
4493 /* Check that the operation is supported. */
4495 vectype);
4497 vectype);
4499 {
4503 }
4506 == CODE_FOR_nothing
4509 {
4513 }
4516 }
4519 /* Function vect_permute_store_chain.
4521 Given a chain of interleaved stores in DR_CHAIN of LENGTH that must be
4522 a power of 2, generate interleave_high/low stmts to reorder the data
4523 correctly for the stores. Return the final references for stores in
4524 RESULT_CHAIN.
4526 E.g., LENGTH is 4 and the scalar type is short, i.e., VF is 8.
4527 The input is 4 vectors each containing 8 elements. We assign a number to each
4528 element, the input sequence is:
4530 1st vec: 0 1 2 3 4 5 6 7
4531 2nd vec: 8 9 10 11 12 13 14 15
4532 3rd vec: 16 17 18 19 20 21 22 23
4533 4th vec: 24 25 26 27 28 29 30 31
4535 The output sequence should be:
4537 1st vec: 0 8 16 24 1 9 17 25
4538 2nd vec: 2 10 18 26 3 11 19 27
4539 3rd vec: 4 12 20 28 5 13 21 30
4540 4th vec: 6 14 22 30 7 15 23 31
4542 i.e., we interleave the contents of the four vectors in their order.
4544 We use interleave_high/low instructions to create such output. The input of
4545 each interleave_high/low operation is two vectors:
4546 1st vec 2nd vec
4547 0 1 2 3 4 5 6 7
4548 the even elements of the result vector are obtained left-to-right from the
4549 high/low elements of the first vector. The odd elements of the result are
4550 obtained left-to-right from the high/low elements of the second vector.
4551 The output of interleave_high will be: 0 4 1 5
4552 and of interleave_low: 2 6 3 7
4555 The permutation is done in log LENGTH stages. In each stage interleave_high
4556 and interleave_low stmts are created for each pair of vectors in DR_CHAIN,
4557 where the first argument is taken from the first half of DR_CHAIN and the
4558 second argument from it's second half.
4559 In our example,
4561 I1: interleave_high (1st vec, 3rd vec)
4562 I2: interleave_low (1st vec, 3rd vec)
4563 I3: interleave_high (2nd vec, 4th vec)
4564 I4: interleave_low (2nd vec, 4th vec)
4566 The output for the first stage is:
4568 I1: 0 16 1 17 2 18 3 19
4569 I2: 4 20 5 21 6 22 7 23
4570 I3: 8 24 9 25 10 26 11 27
4571 I4: 12 28 13 29 14 30 15 31
4573 The output of the second stage, i.e. the final result is:
4575 I1: 0 8 16 24 1 9 17 25
4576 I2: 2 10 18 26 3 11 19 27
4577 I3: 4 12 20 28 5 13 21 30
4578 I4: 6 14 22 30 7 15 23 31. */
4580 static bool
4583 tree stmt,
4586 {
4598 /* Check that the operation is supported. */
4605 {
4607 {
4611 /* Create interleaving stmt:
4612 in the case of big endian:
4613 high = interleave_high (vect1, vect2)
4614 and in the case of little endian:
4615 high = interleave_low (vect1, vect2). */
4621 else
4629 /* Create interleaving stmt:
4630 in the case of big endian:
4631 low = interleave_low (vect1, vect2)
4632 and in the case of little endian:
4633 low = interleave_high (vect1, vect2). */
4639 else
4646 }
4648 }
4650 }
4653 /* Function vectorizable_store.
4655 Check if STMT defines a non scalar data-ref (array/pointer/structure) that
4656 can be vectorized.
4657 If VEC_STMT is also passed, vectorize the STMT: create a vectorized
4658 stmt to replace it, put it in VEC_STMT, and insert it at BSI.
4659 Return FALSE if not a vectorizable STMT, TRUE otherwise. */
4661 bool
4663 slp_tree slp_node)
4664 {
4665 tree scalar_dest;
4666 tree data_ref;
4667 tree op;
4675 tree dummy;
4691 stmt_vec_info first_stmt_vinfo;
4694 /* FORNOW: SLP with multiple types is not supported. The SLP analysis verifies
4695 this, so we can safely override NCOPIES with 1 here. */
4701 /* FORNOW. This restriction should be relaxed. */
4703 {
4707 }
4715 /* Is vectorizable store? */
4728 {
4732 }
4735 /* FORNOW. In some cases can vectorize even if data-type not supported
4736 (e.g. - array initialization with 0). */
4744 {
4752 {
4753 /* STMT is the leader of the group. Check the operands of all the
4754 stmts of the group. */
4757 {
4760 {
4764 }
4766 }
4767 }
4768 }
4771 {
4776 }
4778 /** Transform. **/
4781 {
4787 /* FORNOW */
4790 /* We vectorize all the stmts of the interleaving group when we
4791 reach the last stmt in the group. */
4795 {
4798 }
4803 /* VEC_NUM is the number of vect stmts to be created for this group. */
4806 else
4808 }
4809 else
4810 {
4815 }
4827 /* In case the vectorization factor (VF) is bigger than the number
4828 of elements that we can fit in a vectype (nunits), we have to generate
4829 more than one vector stmt - i.e - we need to "unroll" the
4830 vector stmt by a factor VF/nunits. For more details see documentation in
4831 vect_get_vec_def_for_copy_stmt. */
4833 /* In case of interleaving (non-unit strided access):
4835 S1: &base + 2 = x2
4836 S2: &base = x0
4837 S3: &base + 1 = x1
4838 S4: &base + 3 = x3
4840 We create vectorized stores starting from base address (the access of the
4841 first stmt in the chain (S2 in the above example), when the last store stmt
4842 of the chain (S4) is reached:
4844 VS1: &base = vx2
4845 VS2: &base + vec_size*1 = vx0
4846 VS3: &base + vec_size*2 = vx1
4847 VS4: &base + vec_size*3 = vx3
4849 Then permutation statements are generated:
4851 VS5: vx5 = VEC_INTERLEAVE_HIGH_EXPR < vx0, vx3 >
4852 VS6: vx6 = VEC_INTERLEAVE_LOW_EXPR < vx0, vx3 >
4853 ...
4855 And they are put in STMT_VINFO_VEC_STMT of the corresponding scalar stmts
4856 (the order of the data-refs in the output of vect_permute_store_chain
4857 corresponds to the order of scalar stmts in the interleaving chain - see
4858 the documentation of vect_permute_store_chain()).
4860 In case of both multiple types and interleaving, above vector stores and
4861 permutation stmts are created for every copy. The result vector stmts are
4862 put in STMT_VINFO_VEC_STMT for the first copy and in the corresponding
4863 STMT_VINFO_RELATED_STMT for the next copies.
4864 */
4868 {
4869 tree new_stmt;
4870 tree ptr_incr;
4873 {
4875 {
4876 /* Get vectorized arguments for SLP_NODE. */
4880 }
4881 else
4882 {
4883 /* For interleaved stores we collect vectorized defs for all the
4884 stores in the group in DR_CHAIN and OPRNDS. DR_CHAIN is then
4885 used as an input to vect_permute_store_chain(), and OPRNDS as
4886 an input to vect_get_vec_def_for_stmt_copy() for the next copy.
4888 If the store is not strided, GROUP_SIZE is 1, and DR_CHAIN and
4889 OPRNDS are of size 1. */
4892 {
4893 /* Since gaps are not supported for interleaved stores,
4894 GROUP_SIZE is the exact number of stmts in the chain.
4895 Therefore, NEXT_STMT can't be NULL_TREE. In case that
4896 there is no interleaving, GROUP_SIZE is 1, and only one
4897 iteration of the loop will be executed. */
4902 NULL);
4906 }
4907 }
4912 }
4913 else
4914 {
4915 /* FORNOW SLP doesn't work for multiple types. */
4918 /* For interleaved stores we created vectorized defs for all the
4919 defs stored in OPRNDS in the previous iteration (previous copy).
4920 DR_CHAIN is then used as an input to vect_permute_store_chain(),
4921 and OPRNDS as an input to vect_get_vec_def_for_stmt_copy() for the
4922 next copy.
4923 If the store is not strided, GROUP_SIZE is 1, and DR_CHAIN and
4924 OPRNDS are of size 1. */
4926 {
4932 }
4933 dataref_ptr =
4935 }
4938 {
4940 /* Permute. */
4944 }
4948 {
4950 /* Bump the vector pointer. */
4952 NULL_TREE);
4957 /* For strided stores vectorized defs are interleaved in
4958 vect_permute_store_chain(). */
4962 /* Arguments are ready. Create the new vector stmt. */
4969 else
4976 }
4977 }
4980 }
4983 /* Function vect_setup_realignment
4985 This function is called when vectorizing an unaligned load using
4986 the dr_explicit_realign[_optimized] scheme.
4987 This function generates the following code at the loop prolog:
4989 p = initial_addr;
4990 x msq_init = *(floor(p)); # prolog load
4991 realignment_token = call target_builtin;
4992 loop:
4993 x msq = phi (msq_init, ---)
4995 The stmts marked with x are generated only for the case of
4996 dr_explicit_realign_optimized.
4998 The code above sets up a new (vector) pointer, pointing to the first
4999 location accessed by STMT, and a "floor-aligned" load using that pointer.
5000 It also generates code to compute the "realignment-token" (if the relevant
5001 target hook was defined), and creates a phi-node at the loop-header bb
5002 whose arguments are the result of the prolog-load (created by this
5003 function) and the result of a load that takes place in the loop (to be
5004 created by the caller to this function).
5006 For the case of dr_explicit_realign_optimized:
5007 The caller to this function uses the phi-result (msq) to create the
5008 realignment code inside the loop, and sets up the missing phi argument,
5009 as follows:
5010 loop:
5011 msq = phi (msq_init, lsq)
5012 lsq = *(floor(p')); # load in loop
5013 result = realign_load (msq, lsq, realignment_token);
5015 For the case of dr_explicit_realign:
5016 loop:
5017 msq = *(floor(p)); # load in loop
5018 p' = p + (VS-1);
5019 lsq = *(floor(p')); # load in loop
5020 result = realign_load (msq, lsq, realignment_token);
5022 Input:
5023 STMT - (scalar) load stmt to be vectorized. This load accesses
5024 a memory location that may be unaligned.
5025 BSI - place where new code is to be inserted.
5026 ALIGNMENT_SUPPORT_SCHEME - which of the two misalignment handling schemes
5027 is used.
5029 Output:
5030 REALIGNMENT_TOKEN - the result of a call to the builtin_mask_for_load
5031 target hook, if defined.
5032 Return value - the result of the loop-header phi node. */
5034 static tree
5038 tree init_addr,
5040 {
5045 edge pe;
5047 tree vec_dest;
5048 tree inc;
5049 tree ptr;
5050 tree data_ref;
5051 tree new_stmt;
5052 basic_block new_bb;
5054 tree new_temp;
5055 tree phi_stmt;
5067 /* We need to generate three things:
5068 1. the misalignment computation
5069 2. the extra vector load (for the optimized realignment scheme).
5070 3. the phi node for the two vectors from which the realignment is
5071 done (for the optimized realignment scheme).
5072 */
5074 /* 1. Determine where to generate the misalignment computation.
5076 If INIT_ADDR is NULL_TREE, this indicates that the misalignment
5077 calculation will be generated by this function, outside the loop (in the
5078 preheader). Otherwise, INIT_ADDR had already been computed for us by the
5079 caller, inside the loop.
5081 Background: If the misalignment remains fixed throughout the iterations of
5082 the loop, then both realignment schemes are applicable, and also the
5083 misalignment computation can be done outside LOOP. This is because we are
5084 vectorizing LOOP, and so the memory accesses in LOOP advance in steps that
5085 are a multiple of VS (the Vector Size), and therefore the misalignment in
5086 different vectorized LOOP iterations is always the same.
5087 The problem arises only if the memory access is in an inner-loop nested
5088 inside LOOP, which is now being vectorized using outer-loop vectorization.
5089 This is the only case when the misalignment of the memory access may not
5090 remain fixed throughout the iterations of the inner-loop (as explained in
5091 detail in vect_supportable_dr_alignment). In this case, not only is the
5092 optimized realignment scheme not applicable, but also the misalignment
5093 computation (and generation of the realignment token that is passed to
5094 REALIGN_LOAD) have to be done inside the loop.
5096 In short, INIT_ADDR indicates whether we are in a COMPUTE_IN_LOOP mode
5097 or not, which in turn determines if the misalignment is computed inside
5098 the inner-loop, or outside LOOP. */
5101 {
5104 }
5107 /* 2. Determine where to generate the extra vector load.
5109 For the optimized realignment scheme, instead of generating two vector
5110 loads in each iteration, we generate a single extra vector load in the
5111 preheader of the loop, and in each iteration reuse the result of the
5112 vector load from the previous iteration. In case the memory access is in
5113 an inner-loop nested inside LOOP, which is now being vectorized using
5114 outer-loop vectorization, we need to determine whether this initial vector
5115 load should be generated at the preheader of the inner-loop, or can be
5116 generated at the preheader of LOOP. If the memory access has no evolution
5117 in LOOP, it can be generated in the preheader of LOOP. Otherwise, it has
5118 to be generated inside LOOP (in the preheader of the inner-loop). */
5121 {
5126 }
5127 else
5132 /* 3. For the case of the optimized realignment, create the first vector
5133 load at the loop preheader. */
5136 {
5137 /* Create msq_init = *(floor(p1)) in the loop preheader */
5152 }
5154 /* 4. Create realignment token using a target builtin, if available.
5155 It is done either inside the containing loop, or before LOOP (as
5156 determined above). */
5159 {
5160 tree builtin_decl;
5162 /* Compute INIT_ADDR - the initial addressed accessed by this memref. */
5165 else
5166 {
5167 /* Generate the INIT_ADDR computation outside LOOP. */
5173 }
5185 else
5186 {
5187 /* Generate the misalignment computation outside LOOP. */
5191 }
5195 /* The result of the CALL_EXPR to this builtin is determined from
5196 the value of the parameter and no global variables are touched
5197 which makes the builtin a "const" function. Requiring the
5198 builtin to have the "const" attribute makes it unnecessary
5199 to call mark_call_clobbered. */
5201 }
5210 /* 5. Create msq = phi <msq_init, lsq> in loop */
5220 }
5223 /* Function vect_strided_load_supported.
5225 Returns TRUE is EXTRACT_EVEN and EXTRACT_ODD operations are supported,
5226 and FALSE otherwise. */
5228 static bool
5230 {
5238 {
5242 }
5245 {
5249 }
5253 {
5257 }
5260 {
5264 }
5266 }
5269 /* Function vect_permute_load_chain.
5271 Given a chain of interleaved loads in DR_CHAIN of LENGTH that must be
5272 a power of 2, generate extract_even/odd stmts to reorder the input data
5273 correctly. Return the final references for loads in RESULT_CHAIN.
5275 E.g., LENGTH is 4 and the scalar type is short, i.e., VF is 8.
5276 The input is 4 vectors each containing 8 elements. We assign a number to each
5277 element, the input sequence is:
5279 1st vec: 0 1 2 3 4 5 6 7
5280 2nd vec: 8 9 10 11 12 13 14 15
5281 3rd vec: 16 17 18 19 20 21 22 23
5282 4th vec: 24 25 26 27 28 29 30 31
5284 The output sequence should be:
5286 1st vec: 0 4 8 12 16 20 24 28
5287 2nd vec: 1 5 9 13 17 21 25 29
5288 3rd vec: 2 6 10 14 18 22 26 30
5289 4th vec: 3 7 11 15 19 23 27 31
5291 i.e., the first output vector should contain the first elements of each
5292 interleaving group, etc.
5294 We use extract_even/odd instructions to create such output. The input of each
5295 extract_even/odd operation is two vectors
5296 1st vec 2nd vec
5297 0 1 2 3 4 5 6 7
5299 and the output is the vector of extracted even/odd elements. The output of
5300 extract_even will be: 0 2 4 6
5301 and of extract_odd: 1 3 5 7
5304 The permutation is done in log LENGTH stages. In each stage extract_even and
5305 extract_odd stmts are created for each pair of vectors in DR_CHAIN in their
5306 order. In our example,
5308 E1: extract_even (1st vec, 2nd vec)
5309 E2: extract_odd (1st vec, 2nd vec)
5310 E3: extract_even (3rd vec, 4th vec)
5311 E4: extract_odd (3rd vec, 4th vec)
5313 The output for the first stage will be:
5315 E1: 0 2 4 6 8 10 12 14
5316 E2: 1 3 5 7 9 11 13 15
5317 E3: 16 18 20 22 24 26 28 30
5318 E4: 17 19 21 23 25 27 29 31
5320 In order to proceed and create the correct sequence for the next stage (or
5321 for the correct output, if the second stage is the last one, as in our
5322 example), we first put the output of extract_even operation and then the
5323 output of extract_odd in RESULT_CHAIN (which is then copied to DR_CHAIN).
5324 The input for the second stage is:
5326 1st vec (E1): 0 2 4 6 8 10 12 14
5327 2nd vec (E3): 16 18 20 22 24 26 28 30
5328 3rd vec (E2): 1 3 5 7 9 11 13 15
5329 4th vec (E4): 17 19 21 23 25 27 29 31
5331 The output of the second stage:
5333 E1: 0 4 8 12 16 20 24 28
5334 E2: 2 6 10 14 18 22 26 30
5335 E3: 1 5 9 13 17 21 25 29
5336 E4: 3 7 11 15 19 23 27 31
5338 And RESULT_CHAIN after reordering:
5340 1st vec (E1): 0 4 8 12 16 20 24 28
5341 2nd vec (E3): 1 5 9 13 17 21 25 29
5342 3rd vec (E2): 2 6 10 14 18 22 26 30
5343 4th vec (E4): 3 7 11 15 19 23 27 31. */
5345 static bool
5348 tree stmt,
5351 {
5354 tree tmp;
5358 /* Check that the operation is supported. */
5364 {
5366 {
5370 /* data_ref = permute_even (first_data_ref, second_data_ref); */
5386 /* data_ref = permute_odd (first_data_ref, second_data_ref); */
5400 }
5402 }
5404 }
5407 /* Function vect_transform_strided_load.
5409 Given a chain of input interleaved data-refs (in DR_CHAIN), build statements
5410 to perform their permutation and ascribe the result vectorized statements to
5411 the scalar statements.
5412 */
5414 static bool
5417 {
5423 tree tmp_data_ref;
5425 /* DR_CHAIN contains input data-refs that are a part of the interleaving.
5426 RESULT_CHAIN is the output of vect_permute_load_chain, it contains permuted
5427 vectors, that are ready for vector computation. */
5429 /* Permute. */
5433 /* Put a permuted data-ref in the VECTORIZED_STMT field.
5434 Since we scan the chain starting from it's first node, their order
5435 corresponds the order of data-refs in RESULT_CHAIN. */
5439 {
5443 /* Skip the gaps. Loads created for the gaps will be removed by dead
5444 code elimination pass later.
5445 DR_GROUP_GAP is the number of steps in elements from the previous
5446 access (if there is no gap DR_GROUP_GAP is 1). We skip loads that
5447 correspond to the gaps.
5448 */
5450 {
5451 gap_count++;
5453 }
5456 {
5458 /* We assume that if VEC_STMT is not NULL, this is a case of multiple
5459 copies, and we put the new vector statement in the first available
5460 RELATED_STMT. */
5463 else
5464 {
5469 {
5472 }
5474 }
5477 /* If NEXT_STMT accesses the same DR as the previous statement,
5478 put the same TMP_DATA_REF as its vectorized statement; otherwise
5479 get the next data-ref from RESULT_CHAIN. */
5482 }
5483 }
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? */
5886 vec_dest =
5895 /* FIXME: use build_constructor directly. */
5899 }
5900 else
5902 }
5904 /* Collect vector loads and later create their permutation in
5905 vect_transform_strided_load (). */
5909 /* Store vector loads in the corresponding SLP_NODE. */
5912 }
5914 /* FORNOW: SLP with multiple types is unsupported. */
5919 {
5924 }
5925 else
5926 {
5929 else
5932 }
5933 }
5936 }
5939 /* Function vectorizable_live_operation.
5941 STMT computes a value that is used outside the loop. Check if
5942 it can be supported. */
5944 bool
5948 {
5949 tree operation;
5955 tree op;
5970 /* FORNOW. CHECKME. */
5977 /* FORNOW: support only if all uses are invariant. This means
5978 that the scalar operations can remain in place, unvectorized.
5979 The original last scalar value that they compute will be used. */
5982 {
5985 {
5989 }
5993 }
5995 /* No transformation is required for the cases we currently support. */
5997 }
6000 /* Function vect_is_simple_cond.
6002 Input:
6003 LOOP - the loop that is being vectorized.
6004 COND - Condition that is checked for simple use.
6006 Returns whether a COND can be vectorized. Checks whether
6007 condition operands are supportable using vec_is_simple_use. */
6009 static bool
6011 {
6013 tree def;
6023 {
6027 }
6033 {
6037 }
6043 }
6045 /* vectorizable_condition.
6047 Check if STMT is conditional modify expression that can be vectorized.
6048 If VEC_STMT is also passed, vectorize the STMT: create a vectorized
6049 stmt using VEC_COND_EXPR to replace it, put it in VEC_STMT, and insert it
6050 at BSI.
6052 Return FALSE if not a vectorizable STMT, TRUE otherwise. */
6054 bool
6056 {
6065 tree new_temp;
6068 tree def;
6083 /* FORNOW: SLP not supported. */
6087 /* FORNOW: not yet supported. */
6089 {
6093 }
6095 /* Is vectorizable conditional operation? */
6111 /* We do not handle two different vector types for the condition
6112 and the values. */
6117 {
6122 }
6129 {
6134 }
6144 {
6147 }
6149 /* Transform */
6151 /* Handle def. */
6155 /* Handle cond expr. */
6156 vec_cond_lhs =
6158 vec_cond_rhs =
6163 /* Arguments are ready. create the new vector stmt. */
6175 }
6178 /* Function vect_transform_stmt.
6180 Create a vectorized stmt to replace STMT, and insert it at BSI. */
6182 static bool
6184 slp_tree slp_node)
6185 {
6189 tree orig_stmt_in_pattern;
6193 {
6236 {
6237 /* In case of interleaving, the whole chain is vectorized when the
6238 last store in the chain is reached. Store stmts before the last
6239 one are skipped, and there vec_stmt_info shouldn't be freed
6240 meanwhile. */
6244 }
6245 else
6268 {
6272 }
6273 }
6277 {
6280 }
6283 {
6287 {
6289 /* STMT was inserted by the vectorizer to replace a computation idiom.
6290 ORIG_STMT_IN_PATTERN is a stmt in the original sequence that
6291 computed this idiom. We need to record a pointer to VEC_STMT in
6292 the stmt_info of ORIG_STMT_IN_PATTERN. See more details in the
6293 documentation of vect_pattern_recog. */
6295 {
6298 }
6299 }
6300 }
6303 }
6306 /* This function builds ni_name = number of iterations loop executes
6307 on the loop preheader. */
6309 static tree
6311 {
6313 edge pe;
6323 {
6326 }
6329 }
6332 /* This function generates the following statements:
6334 ni_name = number of iterations loop executes
6335 ratio = ni_name / vf
6336 ratio_mult_vf_name = ratio * vf
6338 and places them at the loop preheader edge. */
6340 static void
6345 {
6347 edge pe;
6348 basic_block new_bb;
6350 tree var;
6351 tree ratio_name;
6352 tree ratio_mult_vf_name;
6356 tree log_vf;
6360 /* Generate temporary variable that contains
6361 number of iterations loop executes. */
6366 /* Create: ratio = ni >> log2(vf) */
6370 {
6378 }
6380 /* Create: ratio_mult_vf = ratio << log2 (vf). */
6385 {
6394 }
6401 }
6404 /* Function vect_update_ivs_after_vectorizer.
6406 "Advance" the induction variables of LOOP to the value they should take
6407 after the execution of LOOP. This is currently necessary because the
6408 vectorizer does not handle induction variables that are used after the
6409 loop. Such a situation occurs when the last iterations of LOOP are
6410 peeled, because:
6411 1. We introduced new uses after LOOP for IVs that were not originally used
6412 after LOOP: the IVs of LOOP are now used by an epilog loop.
6413 2. LOOP is going to be vectorized; this means that it will iterate N/VF
6414 times, whereas the loop IVs should be bumped N times.
6416 Input:
6417 - LOOP - a loop that is going to be vectorized. The last few iterations
6418 of LOOP were peeled.
6419 - NITERS - the number of iterations that LOOP executes (before it is
6420 vectorized). i.e, the number of times the ivs should be bumped.
6421 - UPDATE_E - a successor edge of LOOP->exit that is on the (only) path
6422 coming out from LOOP on which there are uses of the LOOP ivs
6423 (this is the path from LOOP->exit to epilog_loop->preheader).
6425 The new definitions of the ivs are placed in LOOP->exit.
6426 The phi args associated with the edge UPDATE_E in the bb
6427 UPDATE_E->dest are updated accordingly.
6429 Assumption 1: Like the rest of the vectorizer, this function assumes
6430 a single loop exit that has a single predecessor.
6432 Assumption 2: The phi nodes in the LOOP header and in update_bb are
6433 organized in the same order.
6435 Assumption 3: The access function of the ivs is simple enough (see
6436 vect_can_advance_ivs_p). This assumption will be relaxed in the future.
6438 Assumption 4: Exactly one of the successors of LOOP exit-bb is on a path
6439 coming out of LOOP on which the ivs of LOOP are used (this is the path
6440 that leads to the epilog loop; other paths skip the epilog loop). This
6441 path starts with the edge UPDATE_E, and its destination (denoted update_bb)
6442 needs to have its phis updated.
6443 */
6445 static void
6447 edge update_e)
6448 {
6454 /* gcc_assert (vect_can_advance_ivs_p (loop_vinfo)); */
6456 /* Make sure there exists a single-predecessor exit bb: */
6462 {
6464 tree evolution_part;
6465 tree init_expr;
6466 tree step_expr;
6468 block_stmt_iterator last_bsi;
6471 {
6474 }
6476 /* Skip virtual phi's. */
6478 {
6482 }
6484 /* Skip reduction phis. */
6486 {
6490 }
6494 evolution_part =
6498 /* FORNOW: We do not support IVs whose evolution function is a polynomial
6499 of degree >= 2 or exponential. */
6508 init_expr,
6512 else
6516 niters),
6517 step_expr),
6518 init_expr);
6529 /* Fix phi expressions in the successor bb. */
6531 }
6532 }
6534 /* Return the more conservative threshold between the
6535 min_profitable_iters returned by the cost model and the user
6536 specified threshold, if provided. */
6538 static unsigned int
6541 {
6548 /* Use the cost model only if it is more conservative than user specified
6549 threshold. */
6552 && (!min_scalar_loop_bound
6560 }
6562 /* Function vect_do_peeling_for_loop_bound
6564 Peel the last iterations of the loop represented by LOOP_VINFO.
6565 The peeled iterations form a new epilog loop. Given that the loop now
6566 iterates NITERS times, the new epilog loop iterates
6567 NITERS % VECTORIZATION_FACTOR times.
6569 The original loop will later be made to iterate
6570 NITERS / VECTORIZATION_FACTOR times (this value is placed into RATIO). */
6572 static void
6574 {
6578 edge update_e;
6579 basic_block preheader;
6590 /* Generate the following variables on the preheader of original loop:
6592 ni_name = number of iteration the original loop executes
6593 ratio = ni_name / vf
6594 ratio_mult_vf_name = ratio * vf */
6600 /* If cost model check not done during versioning and
6601 peeling for alignment. */
6605 {
6608 /* Get profitability threshold for vectorized loop. */
6612 min_profitable_iters);
6613 }
6620 #ifdef ENABLE_CHECKING
6622 #endif
6624 /* A guard that controls whether the new_loop is to be executed or skipped
6625 is placed in LOOP->exit. LOOP->exit therefore has two successors - one
6626 is the preheader of NEW_LOOP, where the IVs from LOOP are used. The other
6627 is a bb after NEW_LOOP, where these IVs are not used. Find the edge that
6628 is on the path where the LOOP IVs are used and need to be updated. */
6633 else
6636 /* Update IVs of original loop as if they were advanced
6637 by ratio_mult_vf_name steps. */
6640 /* After peeling we have to reset scalar evolution analyzer. */
6644 }
6647 /* Function vect_gen_niters_for_prolog_loop
6649 Set the number of iterations for the loop represented by LOOP_VINFO
6650 to the minimum between LOOP_NITERS (the original iteration count of the loop)
6651 and the misalignment of DR - the data reference recorded in
6652 LOOP_VINFO_UNALIGNED_DR (LOOP_VINFO). As a result, after the execution of
6653 this loop, the data reference DR will refer to an aligned location.
6655 The following computation is generated:
6657 If the misalignment of DR is known at compile time:
6658 addr_mis = int mis = DR_MISALIGNMENT (dr);
6659 Else, compute address misalignment in bytes:
6660 addr_mis = addr & (vectype_size - 1)
6662 prolog_niters = min ( LOOP_NITERS , (VF - addr_mis/elem_size)&(VF-1) )
6664 (elem_size = element type size; an element is the scalar element
6665 whose type is the inner type of the vectype)
6667 For interleaving,
6669 prolog_niters = min ( LOOP_NITERS ,
6670 (VF/group_size - addr_mis/elem_size)&(VF/group_size-1) )
6671 where group_size is the size of the interleaved group.
6673 The above formulas assume that VF == number of elements in the vector. This
6674 may not hold when there are multiple-types in the loop.
6675 In this case, for some data-references in the loop the VF does not represent
6676 the number of elements that fit in the vector. Therefore, instead of VF we
6677 use TYPE_VECTOR_SUBPARTS. */
6679 static tree
6681 {
6686 edge pe;
6687 basic_block new_bb;
6698 {
6699 /* For interleaved access element size must be multiplied by the size of
6700 the interleaved group. */
6704 }
6709 {
6717 }
6718 else
6719 {
6727 tree elem_size_log =
6731 tree byte_misalign;
6732 tree elem_misalign;
6737 /* Create: byte_misalign = addr & (vectype_size - 1) */
6738 byte_misalign =
6741 /* Create: elem_misalign = byte_misalign / element_size */
6742 elem_misalign =
6745 /* Create: (niters_type) (nelements - elem_misalign)&(nelements - 1) */
6749 }
6751 /* Create: prolog_loop_niters = min (iters, loop_niters) */
6752 /* If the loop bound is known at compile time we already verified that it is
6753 greater than vf; since the misalignment ('iters') is at most vf, there's
6754 no need to generate the MIN_EXPR in this case. */
6759 {
6762 }
6768 /* Insert stmt on loop preheader edge. */
6770 {
6773 }
6776 }
6779 /* Function vect_update_init_of_dr
6781 NITERS iterations were peeled from LOOP. DR represents a data reference
6782 in LOOP. This function updates the information recorded in DR to
6783 account for the fact that the first NITERS iterations had already been
6784 executed. Specifically, it updates the OFFSET field of DR. */
6786 static void
6788 {
6794 }
6797 /* Function vect_update_inits_of_drs
6799 NITERS iterations were peeled from the loop represented by LOOP_VINFO.
6800 This function updates the information recorded for the data references in
6801 the loop to account for the fact that the first NITERS iterations had
6802 already been executed. Specifically, it updates the initial_condition of
6803 the access_function of all the data_references in the loop. */
6805 static void
6807 {
6817 }
6820 /* Function vect_do_peeling_for_alignment
6822 Peel the first 'niters' iterations of the loop represented by LOOP_VINFO.
6823 'niters' is set to the misalignment of one of the data references in the
6824 loop, thereby forcing it to refer to an aligned location at the beginning
6825 of the execution of this loop. The data reference for which we are
6826 peeling is recorded in LOOP_VINFO_UNALIGNED_DR. */
6828 static void
6830 {
6833 tree n_iters;
6848 /* If cost model check not done during versioning. */
6851 {
6854 /* Get profitability threshold for vectorized loop. */
6858 min_profitable_iters);
6859 }
6861 /* Peel the prolog loop and iterate it niters_of_prolog_loop. */
6862 new_loop =
6868 #ifdef ENABLE_CHECKING
6870 #endif
6872 /* Update number of times loop executes. */
6877 /* Update the init conditions of the access functions of all data refs. */
6880 /* After peeling we have to reset scalar evolution analyzer. */
6884 }
6887 /* Function vect_create_cond_for_align_checks.
6889 Create a conditional expression that represents the alignment checks for
6890 all of data references (array element references) whose alignment must be
6891 checked at runtime.
6893 Input:
6894 COND_EXPR - input conditional expression. New conditions will be chained
6895 with logical AND operation.
6896 LOOP_VINFO - two fields of the loop information are used.
6897 LOOP_VINFO_PTR_MASK is the mask used to check the alignment.
6898 LOOP_VINFO_MAY_MISALIGN_STMTS contains the refs to be checked.
6900 Output:
6901 COND_EXPR_STMT_LIST - statements needed to construct the conditional
6902 expression.
6903 The returned value is the conditional expression to be used in the if
6904 statement that controls which version of the loop gets executed at runtime.
6906 The algorithm makes two assumptions:
6907 1) The number of bytes "n" in a vector is a power of 2.
6908 2) An address "a" is aligned if a%n is zero and that this
6909 test can be done as a&(n-1) == 0. For example, for 16
6910 byte vectors the test is a&0xf == 0. */
6912 static void
6916 {
6922 tree mask_cst;
6924 tree psize;
6925 tree int_ptrsize_type;
6929 tree ptrsize_zero;
6930 tree part_cond_expr;
6932 /* Check that mask is one less than a power of 2, i.e., mask is
6933 all zeros followed by all ones. */
6936 /* CHECKME: what is the best integer or unsigned type to use to hold a
6937 cast from a pointer value? */
6939 int_ptrsize_type
6942 /* Create expression (mask & (dr_1 || ... || dr_n)) where dr_i is the address
6943 of the first vector of the i'th data reference. */
6946 {
6948 tree addr_base;
6952 /* create: addr_tmp = (int)(address_of_first_vector) */
6968 /* The addresses are OR together. */
6971 {
6972 /* create: or_tmp = or_tmp | addr_tmp */
6983 }
6984 else
6991 /* create: and_tmp = or_tmp & mask */
7001 /* Make and_tmp the left operand of the conditional test against zero.
7002 if and_tmp has a nonzero bit then some address is unaligned. */
7009 else
7011 }
7013 /* Function vect_vfa_segment_size.
7015 Create an expression that computes the size of segment
7016 that will be accessed for a data reference. The functions takes into
7017 account that realignment loads may access one more vector.
7019 Input:
7020 DR: The data reference.
7021 VECT_FACTOR: vectorization factor.
7023 Return an expression whose value is the size of segment which will be
7024 accessed by DR. */
7026 static tree
7028 {
7033 {
7034 tree vector_size = TYPE_SIZE_UNIT
7039 }
7041 }
7043 /* Function vect_create_cond_for_alias_checks.
7045 Create a conditional expression that represents the run-time checks for
7046 overlapping of address ranges represented by a list of data references
7047 relations passed as input.
7049 Input:
7050 COND_EXPR - input conditional expression. New conditions will be chained
7051 with logical AND operation.
7052 LOOP_VINFO - field LOOP_VINFO_MAY_ALIAS_STMTS contains the list of ddrs
7053 to be checked.
7055 Output:
7056 COND_EXPR - conditional expression.
7057 COND_EXPR_STMT_LIST - statements needed to construct the conditional
7058 expression.
7061 The returned value is the conditional expression to be used in the if
7062 statement that controls which version of the loop gets executed at runtime.
7063 */
7065 static void
7069 {
7073 tree vect_factor =
7076 ddr_p ddr;
7078 tree part_cond_expr;
7080 /* Create expression
7081 ((store_ptr_0 + store_segment_length_0) < load_ptr_0)
7082 || (load_ptr_0 + load_segment_length_0) < store_ptr_0))
7083 &&
7084 ...
7085 &&
7086 ((store_ptr_n + store_segment_length_n) < load_ptr_n)
7087 || (load_ptr_n + load_segment_length_n) < store_ptr_n)) */
7093 {
7104 {
7107 }
7113 {
7116 }
7118 addr_base_a =
7121 addr_base_b =
7129 {
7135 }
7138 part_cond_expr =
7142 addr_base_a,
7143 segment_length_a),
7144 addr_base_b),
7147 addr_base_b,
7148 segment_length_b),
7149 addr_base_a));
7154 else
7156 }
7161 }
7163 /* Function vect_loop_versioning.
7165 If the loop has data references that may or may not be aligned or/and
7166 has data reference relations whose independence was not proven then
7167 two versions of the loop need to be generated, one which is vectorized
7168 and one which isn't. A test is then generated to control which of the
7169 loops is executed. The test checks for the alignment of all of the
7170 data references that may or may not be aligned. An additional
7171 sequence of runtime tests is generated for each pairs of DDRs whose
7172 independence was not proven. The vectorized version of loop is
7173 executed only if both alias and alignment tests are passed.
7175 The test generated to check which version of loop is executed
7176 is modified to also check for profitability as indicated by the
7177 cost model initially. */
7179 static void
7181 {
7186 basic_block condition_bb;
7187 block_stmt_iterator cond_exp_bsi;
7188 basic_block merge_bb;
7189 basic_block new_exit_bb;
7193 tree gimplify_stmt_list;
7198 /* Get profitability threshold for vectorized loop. */
7202 min_profitable_iters);
7204 cond_expr =
7219 cond_expr =
7221 cond_expr =
7223 NULL_TREE);
7231 /* Loop versioning violates an assumption we try to maintain during
7232 vectorization - that the loop exit block has a single predecessor.
7233 After versioning, the exit block of both loop versions is the same
7234 basic block (i.e. it has two predecessors). Just in order to simplify
7235 following transformations in the vectorizer, we fix this situation
7236 here by adding a new (empty) block on the exit-edge of the loop,
7237 with the proper loop-exit phis to maintain loop-closed-form. */
7247 {
7249 new_exit_bb);
7253 }
7255 /* End loop-exit-fixes after versioning. */
7259 {
7262 }
7263 }
7265 /* Remove a group of stores (for SLP or interleaving), free their
7266 stmt_vec_info. */
7268 static void
7270 {
7271 stmt_ann_t ann;
7273 tree tmp;
7274 stmt_vec_info next_stmt_info;
7275 block_stmt_iterator next_si;
7278 {
7279 /* Free the attached stmt_vec_info and remove the stmt. */
7288 }
7289 }
7292 /* Vectorize SLP instance tree in postorder. */
7294 static bool
7296 {
7297 tree stmt;
7299 block_stmt_iterator si;
7300 stmt_vec_info stmt_info;
7314 {
7317 }
7322 {
7324 /* If IS_STORE is TRUE, the vectorization of the
7325 interleaving chain was completed - free all the stores in
7326 the chain. */
7328 else
7329 /* FORNOW: SLP originates only from strided stores. */
7333 }
7335 /* FORNOW: SLP originates only from strided stores. */
7337 }
7340 static bool
7342 {
7345 slp_instance instance;
7352 {
7354 /* For each SLP instance calculate number of vector stmts to be created
7355 for the scalar stmts in each node of the SLP tree. Number of vector
7356 elements in one vector iteration is the number of scalar elements in
7357 one scalar iteration (GROUP_SIZE) multiplied by VF divided by vector
7358 size. */
7361 /* Schedule the tree of INSTANCE. */
7363 vec_stmts_size);
7368 }
7371 }
7373 /* Function vect_transform_loop.
7375 The analysis phase has determined that the loop is vectorizable.
7376 Vectorize the loop - created vectorized stmts to replace the scalar
7377 stmts in the loop, and update the loop exit condition. */
7379 void
7381 {
7400 /* CHECKME: we wouldn't need this if we called update_ssa once
7401 for all loops. */
7404 /* Peel the loop if there are data refs with unknown alignment.
7405 Only one data ref with unknown store is allowed. */
7410 /* If the loop has a symbolic number of iterations 'n' (i.e. it's not a
7411 compile time constant), or it is a constant that doesn't divide by the
7412 vectorization factor, then an epilog loop needs to be created.
7413 We therefore duplicate the loop: the original loop will be vectorized,
7414 and will compute the first (n/VF) iterations. The second copy of the loop
7415 will remain scalar and will compute the remaining (n%VF) iterations.
7416 (VF is the vectorization factor). */
7422 else
7426 /* 1) Make sure the loop header has exactly two entries
7427 2) Make sure we have a preheader basic block. */
7433 /* FORNOW: the vectorizer supports only loops which body consist
7434 of one basic block (header + empty latch). When the vectorizer will
7435 support more involved loop forms, the order by which the BBs are
7436 traversed need to be reconsidered. */
7439 {
7441 stmt_vec_info stmt_info;
7442 tree phi;
7445 {
7447 {
7450 }
7465 {
7469 }
7470 }
7473 {
7478 {
7481 }
7485 /* vector stmts created in the outer-loop during vectorization of
7486 stmts in an inner-loop may not have a stmt_info, and do not
7487 need to be vectorized. */
7489 {
7492 }
7496 {
7499 }
7502 nunits =
7507 /* For SLP VF is set according to unrolling factor, and not to
7508 vector size, hence for SLP this print is not valid. */
7511 /* SLP. Schedule all the SLP instances when the first SLP stmt is
7512 reached. */
7514 {
7516 {
7524 /* IS_STORE is true if STMT is a store. Stores cannot be of
7525 hybrid SLP type. They are removed in
7526 vect_schedule_slp_instance and their vinfo is destroyed. */
7528 {
7531 }
7532 }
7534 /* Hybrid SLP stmts must be vectorized in addition to SLP. */
7536 {
7539 }
7540 }
7542 /* -------- vectorize statement ------------ */
7549 {
7550 stmt_ann_t ann;
7552 {
7553 /* Interleaving. If IS_STORE is TRUE, the vectorization of the
7554 interleaving chain was completed - free all the stores in
7555 the chain. */
7557 tree tmp;
7558 stmt_vec_info next_stmt_info;
7561 {
7564 /* Free the attached stmt_vec_info and remove the stmt. */
7571 }
7574 }
7575 else
7576 {
7577 /* Free the attached stmt_vec_info and remove the stmt. */
7583 }
7584 }
7593 /* The memory tags and pointers in vectorized statements need to
7594 have their SSA forms updated. FIXME, why can't this be delayed
7595 until all the loops have been transformed? */
7602 }