| blob | 82a4c84893a80ee3f459f669d75e0268fdb86ae1 |
1 /* Transformation Utilities for Loop Vectorization.
2 Copyright (C) 2003, 2004, 2005, 2006, 2007, 2008 Free Software Foundation, Inc.
3 Contributed by Dorit Naishlos <dorit@il.ibm.com>
5 This file is part of GCC.
7 GCC is free software; you can redistribute it and/or modify it under
8 the terms of the GNU General Public License as published by the Free
9 Software Foundation; either version 3, or (at your option) any later
10 version.
12 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
13 WARRANTY; without even the implied warranty of MERCHANTABILITY or
14 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
15 for more details.
17 You should have received a copy of the GNU General Public License
18 along with GCC; see the file COPYING3. If not see
19 <http://www.gnu.org/licenses/>. */
48 /* Utility functions for the code transformation. */
51 static tree vect_create_data_ref_ptr
53 static tree vect_create_addr_base_for_vector_ref
58 static void vect_finish_stmt_generation
64 /* Utility function dealing with loop peeling (not peeling itself). */
65 static void vect_generate_tmps_on_preheader
75 static int
77 {
81 {
99 }
100 }
103 /* Function vect_estimate_min_profitable_iters
105 Return the number of iterations required for the vector version of the
106 loop to be profitable relative to the cost of the scalar version of the
107 loop.
109 TODO: Take profile info into account before making vectorization
110 decisions, if available. */
112 int
114 {
132 slp_instance instance;
134 /* Cost model disabled. */
136 {
140 }
142 /* If the number of iterations is unknown, or the
143 peeling-for-misalignment amount is unknown, we will have to generate
144 a runtime test to test the loop count against the threshold. */
149 /* Requires loop versioning tests to handle misalignment. */
152 {
153 /* FIXME: Make cost depend on complexity of individual check. */
154 vec_outside_cost +=
159 }
162 {
163 /* FIXME: Make cost depend on complexity of individual check. */
164 vec_outside_cost +=
169 }
173 {
175 }
177 /* Count statements in scalar loop. Using this as scalar cost for a single
178 iteration for now.
180 TODO: Add outer loop support.
182 TODO: Consider assigning different costs to different scalar
183 statements. */
185 /* FORNOW. */
190 {
191 block_stmt_iterator si;
196 else
200 {
203 /* Skip stmts that are not vectorized inside the loop. */
210 /* FIXME: for stmts in the inner-loop in outer-loop vectorization,
211 some of the "outside" costs are generated inside the outer-loop. */
213 }
214 }
216 /* Add additional cost for the peeled instructions in prologue and epilogue
217 loop.
219 FORNOW: If we dont know the value of peel_iters for prologue or epilogue
220 at compile-time - we assume it's vf/2 (the worst would be vf-1).
222 TODO: Build an expression that represents peel_iters for prologue and
223 epilogue to be used in a run-time test. */
226 {
232 /* If peeling for alignment is unknown, loop bound of main loop becomes
233 unknown. */
237 "epilogue peel iters set to vf/2 because "
240 /* If peeled iterations are unknown, count a taken branch and a not taken
241 branch per peeled loop. Even if scalar loop iterations are known,
242 vector iterations are not known since peeled prologue iterations are
243 not known. Hence guards remain the same. */
247 }
248 else
249 {
251 {
258 }
259 else
263 {
267 "epilogue peel iters set to vf/2 because "
270 /* If peeled iterations are known but number of scalar loop
271 iterations are unknown, count a taken branch per peeled loop. */
274 }
275 else
276 {
281 }
282 }
288 /* FORNOW: The scalar outside cost is incremented in one of the
289 following ways:
291 1. The vectorizer checks for alignment and aliasing and generates
292 a condition that allows dynamic vectorization. A cost model
293 check is ANDED with the versioning condition. Hence scalar code
294 path now has the added cost of the versioning check.
296 if (cost > th & versioning_check)
297 jmp to vector code
299 Hence run-time scalar is incremented by not-taken branch cost.
301 2. The vectorizer then checks if a prologue is required. If the
302 cost model check was not done before during versioning, it has to
303 be done before the prologue check.
305 if (cost <= th)
306 prologue = scalar_iters
307 if (prologue == 0)
308 jmp to vector code
309 else
310 execute prologue
311 if (prologue == num_iters)
312 go to exit
314 Hence the run-time scalar cost is incremented by a taken branch,
315 plus a not-taken branch, plus a taken branch cost.
317 3. The vectorizer then checks if an epilogue is required. If the
318 cost model check was not done before during prologue check, it
319 has to be done with the epilogue check.
321 if (prologue == 0)
322 jmp to vector code
323 else
324 execute prologue
325 if (prologue == num_iters)
326 go to exit
327 vector code:
328 if ((cost <= th) | (scalar_iters-prologue-epilogue == 0))
329 jmp to epilogue
331 Hence the run-time scalar cost should be incremented by 2 taken
332 branches.
334 TODO: The back end may reorder the BBS's differently and reverse
335 conditions/branch directions. Change the stimates below to
336 something more reasonable. */
339 {
340 /* Cost model check occurs at versioning. */
344 else
345 {
346 /* Cost model occurs at prologue generation. */
350 /* Cost model check occurs at epilogue generation. */
351 else
353 }
354 }
356 /* Add SLP costs. */
359 {
362 }
364 /* Calculate number of iterations required to make the vector version
365 profitable, relative to the loop bodies only. The following condition
366 must hold true:
367 SIC * niters + SOC > VIC * ((niters-PL_ITERS-EP_ITERS)/VF) + VOC
368 where
369 SIC = scalar iteration cost, VIC = vector iteration cost,
370 VOC = vector outside cost, VF = vectorization factor,
371 PL_ITERS = prologue iterations, EP_ITERS= epilogue iterations
372 SOC = scalar outside cost for run time cost model check. */
375 {
378 else
379 {
389 min_profitable_iters++;
390 }
391 }
392 /* vector version will never be profitable. */
393 else
394 {
397 "is divisible by scalar iteration cost = %d by a factor "
401 }
404 {
407 vec_inside_cost);
409 vec_outside_cost);
411 scalar_single_iter_cost);
414 peel_iters_prologue);
416 peel_iters_epilogue);
418 min_profitable_iters);
419 }
421 min_profitable_iters =
424 /* Because the condition we create is:
425 if (niters <= min_profitable_iters)
426 then skip the vectorized loop. */
427 min_profitable_iters--;
431 min_profitable_iters);
434 }
437 /* TODO: Close dependency between vect_model_*_cost and vectorizable_*
438 functions. Design better to avoid maintenance issues. */
440 /* Function vect_model_reduction_cost.
442 Models cost for a reduction operation, including the vector ops
443 generated within the strip-mine loop, the initial definition before
444 the loop, and the epilogue code that must be generated. */
446 static bool
449 {
452 optab optab;
453 tree vectype;
454 tree orig_stmt;
455 tree reduction_op;
462 /* Cost of reduction op inside loop. */
468 {
470 {
473 }
475 }
485 /* Add in cost for initial definition. */
488 /* Determine cost of epilogue code.
490 We have a reduction operator that will reduce the vector in one statement.
491 Also requires scalar extract. */
494 {
497 else
498 {
500 tree bitsize =
507 /* We have a whole vector shift available. */
511 /* Final reduction via vector shifts and the reduction operator. Also
512 requires scalar extract. */
515 else
516 /* Use extracts and reduction op for final reduction. For N elements,
517 we have N extracts and N-1 reduction ops. */
519 }
520 }
530 }
533 /* Function vect_model_induction_cost.
535 Models cost for induction operations. */
537 static void
539 {
540 /* loop cost for vec_loop. */
542 /* prologue cost for vec_init and vec_step. */
549 }
552 /* Function vect_model_simple_cost.
554 Models cost for simple operations, i.e. those that only emit ncopies of a
555 single op. Right now, this does not account for multiple insns that could
556 be generated for the single vector op. We will handle that shortly. */
558 void
561 {
567 /* FORNOW: Assuming maximum 2 args per stmts. */
569 {
572 }
578 /* Set the costs either in STMT_INFO or SLP_NODE (if exists). */
581 }
584 /* Function vect_cost_strided_group_size
586 For strided load or store, return the group_size only if it is the first
587 load or store of a group, else return 1. This ensures that group size is
588 only returned once per group. */
590 static int
592 {
599 }
602 /* Function vect_model_store_cost
604 Models cost for stores. In the case of strided accesses, one access
605 has the overhead of the strided access attributed to it. */
607 void
610 {
617 /* Strided access? */
620 /* Not a strided access. */
621 else
624 /* Is this an access in a group of stores, which provide strided access?
625 If so, add in the cost of the permutes. */
627 {
628 /* Uses a high and low interleave operation for each needed permute. */
634 group_size);
636 }
638 /* Costs of the stores. */
645 /* Set the costs either in STMT_INFO or SLP_NODE (if exists). */
648 }
651 /* Function vect_model_load_cost
653 Models cost for loads. In the case of strided accesses, the last access
654 has the overhead of the strided access attributed to it. Since unaligned
655 accesses are supported for loads, we also account for the costs of the
656 access scheme chosen. */
658 void
661 {
664 tree first_stmt;
668 /* Strided accesses? */
671 {
674 }
675 /* Not a strided access. */
676 else
677 {
680 }
684 /* Is this an access in a group of loads providing strided access?
685 If so, add in the cost of the permutes. */
687 {
688 /* Uses an even and odd extract operations for each needed permute. */
694 group_size);
696 }
698 /* The loads themselves. */
700 {
702 {
709 }
711 {
712 /* Here, we assign an additional cost for the unaligned load. */
720 }
722 {
725 /* FIXME: If the misalignment remains fixed across the iterations of
726 the containing loop, the following cost should be added to the
727 outside costs. */
732 }
734 {
739 /* Unaligned software pipeline has a load of an address, an initial
740 load, and possibly a mask operation to "prime" the loop. However,
741 if this is an access in a group of loads, which provide strided
742 access, then the above cost should only be considered for one
743 access in the group. Inside the loop, there is a load op
744 and a realignment op. */
747 {
751 }
756 }
760 }
766 /* Set the costs either in STMT_INFO or SLP_NODE (if exists). */
769 }
772 /* Function vect_get_new_vect_var.
774 Returns a name for a new variable. The current naming scheme appends the
775 prefix "vect_" or "vect_p" (depending on the value of VAR_KIND) to
776 the name of vectorizer generated variables, and appends that to NAME if
777 provided. */
779 static tree
781 {
783 tree new_vect_var;
786 {
798 }
801 {
805 }
806 else
809 /* Mark vector typed variable as a gimple register variable. */
814 }
817 /* Function vect_create_addr_base_for_vector_ref.
819 Create an expression that computes the address of the first memory location
820 that will be accessed for a data reference.
822 Input:
823 STMT: The statement containing the data reference.
824 NEW_STMT_LIST: Must be initialized to NULL_TREE or a statement list.
825 OFFSET: Optional. If supplied, it is be added to the initial address.
826 LOOP: Specify relative to which loop-nest should the address be computed.
827 For example, when the dataref is in an inner-loop nested in an
828 outer-loop that is now being vectorized, LOOP can be either the
829 outer-loop, or the inner-loop. The first memory location accessed
830 by the following dataref ('in' points to short):
832 for (i=0; i<N; i++)
833 for (j=0; j<M; j++)
834 s += in[i+j]
836 is as follows:
837 if LOOP=i_loop: &in (relative to i_loop)
838 if LOOP=j_loop: &in+i*2B (relative to j_loop)
840 Output:
841 1. Return an SSA_NAME whose value is the address of the memory location of
842 the first vector of the data reference.
843 2. If new_stmt_list is not NULL_TREE after return then the caller must insert
844 these statement(s) which define the returned SSA_NAME.
846 FORNOW: We are only handling array accesses with step 1. */
848 static tree
851 tree offset,
853 {
858 tree base_name;
859 tree data_ref_base_var;
860 tree new_base_stmt;
861 tree vec_stmt;
871 {
880 }
882 /* Create data_ref_base */
890 /* Create base_offset */
899 {
908 }
910 /* base + base_offset */
916 /* addr_expr = addr_base */
928 {
931 }
933 }
936 /* Function vect_create_data_ref_ptr.
938 Create a new pointer to vector type (vp), that points to the first location
939 accessed in the loop by STMT, along with the def-use update chain to
940 appropriately advance the pointer through the loop iterations. Also set
941 aliasing information for the pointer. This vector pointer is used by the
942 callers to this function to create a memory reference expression for vector
943 load/store access.
945 Input:
946 1. STMT: a stmt that references memory. Expected to be of the form
947 GIMPLE_MODIFY_STMT <name, data-ref> or
948 GIMPLE_MODIFY_STMT <data-ref, name>.
949 2. AT_LOOP: the loop where the vector memref is to be created.
950 3. OFFSET (optional): an offset to be added to the initial address accessed
951 by the data-ref in STMT.
952 4. ONLY_INIT: indicate if vp is to be updated in the loop, or remain
953 pointing to the initial address.
954 5. TYPE: if not NULL indicates the required type of the data-ref
956 Output:
957 1. Declare a new ptr to vector_type, and have it point to the base of the
958 data reference (initial addressed accessed by the data reference).
959 For example, for vector of type V8HI, the following code is generated:
961 v8hi *vp;
962 vp = (v8hi *)initial_address;
964 if OFFSET is not supplied:
965 initial_address = &a[init];
966 if OFFSET is supplied:
967 initial_address = &a[init + OFFSET];
969 Return the initial_address in INITIAL_ADDRESS.
971 2. If ONLY_INIT is true, just return the initial pointer. Otherwise, also
972 update the pointer in each iteration of the loop.
974 Return the increment stmt that updates the pointer in PTR_INCR.
976 3. Set INV_P to true if the access pattern of the data reference in the
977 vectorized loop is invariant. Set it to false otherwise.
979 4. Return the pointer. */
981 static tree
985 {
986 tree base_name;
993 tree vect_ptr_type;
994 tree vect_ptr;
995 tree tag;
996 tree new_temp;
997 tree vec_stmt;
999 edge pe;
1000 basic_block new_bb;
1001 tree vect_ptr_init;
1003 tree vptr;
1004 block_stmt_iterator incr_bsi;
1007 tree incr;
1008 tree step;
1010 /* Check the step (evolution) of the load in LOOP, and record
1011 whether it's invariant. */
1014 else
1019 else
1022 /* Create an expression for the first address accessed by this load
1023 in LOOP. */
1027 {
1040 }
1042 /** (1) Create the new vector-pointer variable: **/
1045 else
1051 /** (2) Add aliasing information to the new vector-pointer:
1052 (The points-to info (DR_PTR_INFO) may be defined later.) **/
1057 /* If tag is a variable (and NOT_A_TAG) than a new symbol memory
1058 tag must be created with tag added to its may alias list. */
1061 else
1064 /** Note: If the dataref is in an inner-loop nested in LOOP, and we are
1065 vectorizing LOOP (i.e. outer-loop vectorization), we need to create two
1066 def-use update cycles for the pointer: One relative to the outer-loop
1067 (LOOP), which is what steps (3) and (4) below do. The other is relative
1068 to the inner-loop (which is the inner-most loop containing the dataref),
1069 and this is done be step (5) below.
1071 When vectorizing inner-most loops, the vectorized loop (LOOP) is also the
1072 inner-most loop, and so steps (3),(4) work the same, and step (5) is
1073 redundant. Steps (3),(4) create the following:
1075 vp0 = &base_addr;
1076 LOOP: vp1 = phi(vp0,vp2)
1077 ...
1078 ...
1079 vp2 = vp1 + step
1080 goto LOOP
1082 If there is an inner-loop nested in loop, then step (5) will also be
1083 applied, and an additional update in the inner-loop will be created:
1085 vp0 = &base_addr;
1086 LOOP: vp1 = phi(vp0,vp2)
1087 ...
1088 inner: vp3 = phi(vp1,vp4)
1089 vp4 = vp3 + inner_step
1090 if () goto inner
1091 ...
1092 vp2 = vp1 + step
1093 if () goto LOOP */
1095 /** (3) Calculate the initial address the vector-pointer, and set
1096 the vector-pointer to point to it before the loop: **/
1098 /* Create: (&(base[init_val+offset]) in the loop preheader. */
1104 {
1107 }
1111 /* Create: p = (vectype *) initial_base */
1120 /** (4) Handle the updating of the vector-pointer inside the loop.
1121 This is needed when ONLY_INIT is false, and also when AT_LOOP
1122 is the inner-loop nested in LOOP (during outer-loop vectorization).
1123 **/
1126 {
1127 /* Copy the points-to information if it exists. */
1131 }
1132 else
1133 {
1134 /* The step of the vector pointer is the Vector Size. */
1136 /* One exception to the above is when the scalar step of the load in
1137 LOOP is zero. In this case the step here is also zero. */
1151 /* Copy the points-to information if it exists. */
1153 {
1156 }
1163 }
1169 /** (5) Handle the updating of the vector-pointer inside the inner-loop
1170 nested in LOOP, if exists: **/
1174 {
1183 /* Copy the points-to information if it exists. */
1185 {
1188 }
1195 }
1196 else
1198 }
1201 /* Function bump_vector_ptr
1203 Increment a pointer (to a vector type) by vector-size. If requested,
1204 i.e. if PTR-INCR is given, then also connect the new increment stmt
1205 to the existing def-use update-chain of the pointer, by modifying
1206 the PTR_INCR as illustrated below:
1208 The pointer def-use update-chain before this function:
1209 DATAREF_PTR = phi (p_0, p_2)
1210 ....
1211 PTR_INCR: p_2 = DATAREF_PTR + step
1213 The pointer def-use update-chain after this function:
1214 DATAREF_PTR = phi (p_0, p_2)
1215 ....
1216 NEW_DATAREF_PTR = DATAREF_PTR + BUMP
1217 ....
1218 PTR_INCR: p_2 = NEW_DATAREF_PTR + step
1220 Input:
1221 DATAREF_PTR - ssa_name of a pointer (to vector type) that is being updated
1222 in the loop.
1223 PTR_INCR - optional. The stmt that updates the pointer in each iteration of
1224 the loop. The increment amount across iterations is expected
1225 to be vector_size.
1226 BSI - location where the new update stmt is to be placed.
1227 STMT - the original scalar memory-access stmt that is being vectorized.
1228 BUMP - optional. The offset by which to bump the pointer. If not given,
1229 the offset is assumed to be vector_size.
1231 Output: Return NEW_DATAREF_PTR as illustrated above.
1233 */
1235 static tree
1238 {
1245 tree incr_stmt;
1246 ssa_op_iter iter;
1247 use_operand_p use_p;
1248 tree new_dataref_ptr;
1260 /* Copy the points-to information if it exists. */
1268 /* Update the vector-pointer's cross-iteration increment. */
1270 {
1275 else
1277 }
1280 }
1283 /* Function vect_create_destination_var.
1285 Create a new temporary of type VECTYPE. */
1287 static tree
1289 {
1290 tree vec_dest;
1292 tree type;
1307 }
1310 /* Function vect_init_vector.
1312 Insert a new stmt (INIT_STMT) that initializes a new vector variable with
1313 the vector elements of VECTOR_VAR. Place the initialization at BSI if it
1314 is not NULL. Otherwise, place the initialization at the loop preheader.
1315 Return the DEF of INIT_STMT.
1316 It will be used in the vectorization of STMT. */
1318 static tree
1321 {
1323 tree new_var;
1324 tree init_stmt;
1325 tree vec_oprnd;
1326 edge pe;
1327 tree new_temp;
1328 basic_block new_bb;
1338 else
1339 {
1348 }
1351 {
1354 }
1358 }
1361 /* For constant and loop invariant defs of SLP_NODE this function returns
1362 (vector) defs (VEC_OPRNDS) that will be used in the vectorized stmts.
1363 OP_NUM determines if we gather defs for operand 0 or operand 1 of the scalar
1364 stmts. */
1366 static void
1369 {
1375 tree vec_cst;
1378 tree vector_type;
1391 /* NUMBER_OF_COPIES is the number of times we need to use the same values in
1392 created vectors. It is greater than 1 if unrolling is performed.
1394 For example, we have two scalar operands, s1 and s2 (e.g., group of
1395 strided accesses of size two), while NUINTS is four (i.e., four scalars
1396 of this type can be packed in a vector). The output vector will contain
1397 two copies of each scalar operand: {s1, s2, s1, s2}. (NUMBER_OF_COPIES
1398 will be 2).
1400 If GROUP_SIZE > NUNITS, the scalars will be split into several vectors
1401 containing the operands.
1403 For example, NUINTS is four as before, and the group size is 8
1404 (s1, s2, ..., s8). We will create two vectors {s1, s2, s3, s4} and
1405 {s5, s6, s7, s8}. */
1412 {
1414 {
1418 else
1423 /* Create 'vect_ = {op0,op1,...,opn}'. */
1426 number_of_places_left_in_vector--;
1429 {
1436 else
1441 NULL));
1443 }
1444 }
1445 }
1447 /* Since the vectors are created in the reverse order, we should invert
1448 them. */
1451 {
1454 }
1458 /* In case that VF is greater than the unrolling factor needed for the SLP
1459 group of stmts, NUMBER_OF_VECTORS to be created is greater than
1460 NUMBER_OF_SCALARS/NUNITS or NUNITS/NUMBER_OF_SCALARS, and hence we have
1461 to replicate the vectors. */
1463 {
1466 }
1467 }
1470 /* Get vectorized definitions from SLP_NODE that contains corresponding
1471 vectorized def-stmts. */
1473 static void
1475 {
1476 tree vec_oprnd;
1477 tree vec_def_stmt;
1484 i++)
1485 {
1489 }
1490 }
1493 /* Get vectorized definitions for SLP_NODE.
1494 If the scalar definitions are loop invariants or constants, collect them and
1495 call vect_get_constant_vectors() to create vector stmts.
1496 Otherwise, the def-stmts must be already vectorized and the vectorized stmts
1497 must be stored in the LEFT/RIGHT node of SLP_NODE, and we call
1498 vect_get_slp_vect_defs() to retrieve them.
1499 If VEC_OPRNDS1 is NULL, don't get vector defs for the second operand (from
1500 the right node. This is used when the second operand must remain scalar. */
1502 static void
1505 {
1508 /* Allocate memory for vectorized defs. */
1512 /* SLP_NODE corresponds either to a group of stores or to a group of
1513 unary/binary operations. We don't call this function for loads. */
1515 /* The defs are already vectorized. */
1517 else
1518 /* Build vectors from scalar defs. */
1523 /* Since we don't call this function with loads, this is a group of
1524 stores. */
1535 /* The defs are already vectorized. */
1537 else
1538 /* Build vectors from scalar defs. */
1540 }
1543 /* Function get_initial_def_for_induction
1545 Input:
1546 STMT - a stmt that performs an induction operation in the loop.
1547 IV_PHI - the initial value of the induction variable
1549 Output:
1550 Return a vector variable, initialized with the first VF values of
1551 the induction variable. E.g., for an iv with IV_PHI='X' and
1552 evolution S, for a vector of 4 units, we want to return:
1553 [X, X + S, X + 2*S, X + 3*S]. */
1555 static tree
1557 {
1562 tree vectype;
1566 basic_block new_bb;
1568 tree access_fn;
1569 tree new_var;
1570 tree new_name;
1571 tree init_stmt;
1578 tree expr;
1581 tree stmts;
1582 imm_use_iterator imm_iter;
1583 use_operand_p use_p;
1584 tree exit_phi;
1585 edge latch_e;
1586 tree loop_arg;
1587 block_stmt_iterator si;
1598 /* Find the first insertion point in the BB. */
1603 else
1606 /* Is phi in an inner-loop, while vectorizing an enclosing outer-loop? */
1608 {
1611 }
1612 else
1626 /* Create the vector that holds the initial_value of the induction. */
1628 {
1629 /* iv_loop is nested in the loop to be vectorized. init_expr had already
1630 been created during vectorization of previous stmts; We obtain it from
1631 the STMT_VINFO_VEC_STMT of the defining stmt. */
1634 }
1635 else
1636 {
1637 /* iv_loop is the loop to be vectorized. Create:
1638 vec_init = [X, X+S, X+2*S, X+3*S] (S = step_expr, X = init_expr) */
1644 {
1647 }
1652 {
1653 tree tmp;
1655 /* Create: new_name_i = new_name + step_expr */
1665 {
1668 }
1670 }
1671 /* Create a vector from [new_name_0, new_name_1, ..., new_name_nunits-1] */
1674 }
1677 /* Create the vector that holds the step of the induction. */
1679 /* iv_loop is nested in the loop to be vectorized. Generate:
1680 vec_step = [S, S, S, S] */
1682 else
1683 {
1684 /* iv_loop is the loop to be vectorized. Generate:
1685 vec_step = [VF*S, VF*S, VF*S, VF*S] */
1688 }
1698 /* Create the following def-use cycle:
1699 loop prolog:
1700 vec_init = ...
1701 vec_step = ...
1702 loop:
1703 vec_iv = PHI <vec_init, vec_loop>
1704 ...
1705 STMT
1706 ...
1707 vec_loop = vec_iv + vec_step; */
1709 /* Create the induction-phi that defines the induction-operand. */
1717 /* Create the iv update inside the loop */
1727 /* Set the arguments of the phi node: */
1732 /* In case that vectorization factor (VF) is bigger than the number
1733 of elements that we can fit in a vectype (nunits), we have to generate
1734 more than one vector stmt - i.e - we need to "unroll" the
1735 vector stmt by a factor VF/nunits. For more details see documentation
1736 in vectorizable_operation. */
1739 {
1740 stmt_vec_info prev_stmt_vinfo;
1741 /* FORNOW. This restriction should be relaxed. */
1744 /* Create the vector that holds the step of the induction. */
1757 {
1758 tree tmp;
1760 /* vec_i = vec_prev + vec_step */
1770 }
1771 }
1774 {
1775 /* Find the loop-closed exit-phi of the induction, and record
1776 the final vector of induction results: */
1779 {
1781 {
1784 }
1785 }
1787 {
1789 /* FORNOW. Currently not supporting the case that an inner-loop induction
1790 is not used in the outer-loop (i.e. only outside the outer-loop). */
1796 {
1799 }
1800 }
1801 }
1805 {
1810 }
1814 }
1817 /* Function vect_get_vec_def_for_operand.
1819 OP is an operand in STMT. This function returns a (vector) def that will be
1820 used in the vectorized stmt for STMT.
1822 In the case that OP is an SSA_NAME which is defined in the loop, then
1823 STMT_VINFO_VEC_STMT of the defining stmt holds the relevant def.
1825 In case OP is an invariant or constant, a new stmt that creates a vector def
1826 needs to be introduced. */
1828 static tree
1830 {
1831 tree vec_oprnd;
1832 tree vec_stmt;
1833 tree def_stmt;
1839 tree vec_inv;
1840 tree vec_cst;
1842 tree def;
1846 tree vector_type;
1849 {
1852 }
1857 {
1859 {
1862 }
1864 {
1867 }
1868 }
1871 {
1872 /* Case 1: operand is a constant. */
1874 {
1878 /* Create 'vect_cst_ = {cst,cst,...,cst}' */
1883 {
1885 }
1891 }
1893 /* Case 2: operand is defined outside the loop - loop invariant. */
1895 {
1899 /* Create 'vec_inv = {inv,inv,..,inv}' */
1904 {
1906 }
1908 /* FIXME: use build_constructor directly. */
1913 }
1915 /* Case 3: operand is defined inside the loop. */
1917 {
1921 /* Get the def from the vectorized stmt. */
1927 else
1930 }
1932 /* Case 4: operand is defined by a loop header phi - reduction */
1934 {
1940 /* Get the def before the loop */
1943 }
1945 /* Case 5: operand is defined by loop-header phi - induction. */
1947 {
1950 /* Get the def from the vectorized stmt. */
1956 }
1960 }
1961 }
1964 /* Function vect_get_vec_def_for_stmt_copy
1966 Return a vector-def for an operand. This function is used when the
1967 vectorized stmt to be created (by the caller to this function) is a "copy"
1968 created in case the vectorized result cannot fit in one vector, and several
1969 copies of the vector-stmt are required. In this case the vector-def is
1970 retrieved from the vector stmt recorded in the STMT_VINFO_RELATED_STMT field
1971 of the stmt that defines VEC_OPRND.
1972 DT is the type of the vector def VEC_OPRND.
1974 Context:
1975 In case the vectorization factor (VF) is bigger than the number
1976 of elements that can fit in a vectype (nunits), we have to generate
1977 more than one vector stmt to vectorize the scalar stmt. This situation
1978 arises when there are multiple data-types operated upon in the loop; the
1979 smallest data-type determines the VF, and as a result, when vectorizing
1980 stmts operating on wider types we need to create 'VF/nunits' "copies" of the
1981 vector stmt (each computing a vector of 'nunits' results, and together
1982 computing 'VF' results in each iteration). This function is called when
1983 vectorizing such a stmt (e.g. vectorizing S2 in the illustration below, in
1984 which VF=16 and nunits=4, so the number of copies required is 4):
1986 scalar stmt: vectorized into: STMT_VINFO_RELATED_STMT
1988 S1: x = load VS1.0: vx.0 = memref0 VS1.1
1989 VS1.1: vx.1 = memref1 VS1.2
1990 VS1.2: vx.2 = memref2 VS1.3
1991 VS1.3: vx.3 = memref3
1993 S2: z = x + ... VSnew.0: vz0 = vx.0 + ... VSnew.1
1994 VSnew.1: vz1 = vx.1 + ... VSnew.2
1995 VSnew.2: vz2 = vx.2 + ... VSnew.3
1996 VSnew.3: vz3 = vx.3 + ...
1998 The vectorization of S1 is explained in vectorizable_load.
1999 The vectorization of S2:
2000 To create the first vector-stmt out of the 4 copies - VSnew.0 -
2001 the function 'vect_get_vec_def_for_operand' is called to
2002 get the relevant vector-def for each operand of S2. For operand x it
2003 returns the vector-def 'vx.0'.
2005 To create the remaining copies of the vector-stmt (VSnew.j), this
2006 function is called to get the relevant vector-def for each operand. It is
2007 obtained from the respective VS1.j stmt, which is recorded in the
2008 STMT_VINFO_RELATED_STMT field of the stmt that defines VEC_OPRND.
2010 For example, to obtain the vector-def 'vx.1' in order to create the
2011 vector stmt 'VSnew.1', this function is called with VEC_OPRND='vx.0'.
2012 Given 'vx0' we obtain the stmt that defines it ('VS1.0'); from the
2013 STMT_VINFO_RELATED_STMT field of 'VS1.0' we obtain the next copy - 'VS1.1',
2014 and return its def ('vx.1').
2015 Overall, to create the above sequence this function will be called 3 times:
2016 vx.1 = vect_get_vec_def_for_stmt_copy (dt, vx.0);
2017 vx.2 = vect_get_vec_def_for_stmt_copy (dt, vx.1);
2018 vx.3 = vect_get_vec_def_for_stmt_copy (dt, vx.2); */
2020 static tree
2022 {
2023 tree vec_stmt_for_operand;
2024 stmt_vec_info def_stmt_info;
2026 /* Do nothing; can reuse same def. */
2037 }
2040 /* Get vectorized definitions for the operands to create a copy of an original
2041 stmt. See vect_get_vec_def_for_stmt_copy() for details. */
2043 static void
2047 {
2054 {
2058 }
2059 }
2062 /* Get vectorized definitions for OP0 and OP1, or SLP_NODE if it is not NULL. */
2064 static void
2067 {
2070 else
2071 {
2072 tree vec_oprnd;
2079 {
2083 }
2084 }
2085 }
2088 /* Function vect_finish_stmt_generation.
2090 Insert a new stmt. */
2092 static void
2095 {
2108 {
2111 }
2113 /* Make sure bsi points to the stmt that is being vectorized. */
2117 }
2120 /* Function get_initial_def_for_reduction
2122 Input:
2123 STMT - a stmt that performs a reduction operation in the loop.
2124 INIT_VAL - the initial value of the reduction variable
2126 Output:
2127 ADJUSTMENT_DEF - a tree that holds a value to be added to the final result
2128 of the reduction (used for adjusting the epilog - see below).
2129 Return a vector variable, initialized according to the operation that STMT
2130 performs. This vector will be used as the initial value of the
2131 vector of partial results.
2133 Option1 (adjust in epilog): Initialize the vector as follows:
2134 add: [0,0,...,0,0]
2135 mult: [1,1,...,1,1]
2136 min/max: [init_val,init_val,..,init_val,init_val]
2137 bit and/or: [init_val,init_val,..,init_val,init_val]
2138 and when necessary (e.g. add/mult case) let the caller know
2139 that it needs to adjust the result by init_val.
2141 Option2: Initialize the vector as follows:
2142 add: [0,0,...,0,init_val]
2143 mult: [1,1,...,1,init_val]
2144 min/max: [init_val,init_val,...,init_val]
2145 bit and/or: [init_val,init_val,...,init_val]
2146 and no adjustments are needed.
2148 For example, for the following code:
2150 s = init_val;
2151 for (i=0;i<n;i++)
2152 s = s + a[i];
2154 STMT is 's = s + a[i]', and the reduction variable is 's'.
2155 For a vector of 4 units, we want to return either [0,0,0,init_val],
2156 or [0,0,0,0] and let the caller know that it needs to adjust
2157 the result at the end by 'init_val'.
2159 FORNOW, we are using the 'adjust in epilog' scheme, because this way the
2160 initialization vector is simpler (same element in all entries).
2161 A cost model should help decide between these two schemes. */
2163 static tree
2165 {
2173 tree vecdef;
2174 tree def_for_init;
2175 tree init_def;
2178 tree vector_type;
2184 else
2190 {
2196 else
2198 /* Create a vector of zeros for init_def. */
2201 else
2218 }
2221 }
2224 /* Function vect_create_epilog_for_reduction
2226 Create code at the loop-epilog to finalize the result of a reduction
2227 computation.
2229 VECT_DEF is a vector of partial results.
2230 REDUC_CODE is the tree-code for the epilog reduction.
2231 STMT is the scalar reduction stmt that is being vectorized.
2232 REDUCTION_PHI is the phi-node that carries the reduction computation.
2234 This function:
2235 1. Creates the reduction def-use cycle: sets the arguments for
2236 REDUCTION_PHI:
2237 The loop-entry argument is the vectorized initial-value of the reduction.
2238 The loop-latch argument is VECT_DEF - the vector of partial sums.
2239 2. "Reduces" the vector of partial results VECT_DEF into a single result,
2240 by applying the operation specified by REDUC_CODE if available, or by
2241 other means (whole-vector shifts or a scalar loop).
2242 The function also creates a new phi node at the loop exit to preserve
2243 loop-closed form, as illustrated below.
2245 The flow at the entry to this function:
2247 loop:
2248 vec_def = phi <null, null> # REDUCTION_PHI
2249 VECT_DEF = vector_stmt # vectorized form of STMT
2250 s_loop = scalar_stmt # (scalar) STMT
2251 loop_exit:
2252 s_out0 = phi <s_loop> # (scalar) EXIT_PHI
2253 use <s_out0>
2254 use <s_out0>
2256 The above is transformed by this function into:
2258 loop:
2259 vec_def = phi <vec_init, VECT_DEF> # REDUCTION_PHI
2260 VECT_DEF = vector_stmt # vectorized form of STMT
2261 s_loop = scalar_stmt # (scalar) STMT
2262 loop_exit:
2263 s_out0 = phi <s_loop> # (scalar) EXIT_PHI
2264 v_out1 = phi <VECT_DEF> # NEW_EXIT_PHI
2265 v_out2 = reduce <v_out1>
2266 s_out3 = extract_field <v_out2, 0>
2267 s_out4 = adjust_result <s_out3>
2268 use <s_out4>
2269 use <s_out4>
2270 */
2272 static void
2275 {
2277 tree vectype;
2281 basic_block exit_bb;
2282 tree scalar_dest;
2283 tree scalar_type;
2284 tree new_phi;
2285 block_stmt_iterator exit_bsi;
2286 tree vec_dest;
2288 tree new_name;
2293 tree adjustment_def;
2294 tree vec_initial_def;
2295 tree orig_name;
2296 imm_use_iterator imm_iter;
2297 use_operand_p use_p;
2300 tree orig_stmt;
2301 tree use_stmt;
2309 {
2312 }
2320 /*** 1. Create the reduction def-use cycle ***/
2322 /* 1.1 set the loop-entry arg of the reduction-phi: */
2323 /* For the case of reduction, vect_get_vec_def_for_operand returns
2324 the scalar def before the loop, that defines the initial value
2325 of the reduction variable. */
2330 /* 1.2 set the loop-latch arg for the reduction-phi: */
2334 {
2339 }
2342 /*** 2. Create epilog code
2343 The reduction epilog code operates across the elements of the vector
2344 of partial results computed by the vectorized loop.
2345 The reduction epilog code consists of:
2346 step 1: compute the scalar result in a vector (v_out2)
2347 step 2: extract the scalar result (s_out3) from the vector (v_out2)
2348 step 3: adjust the scalar result (s_out3) if needed.
2350 Step 1 can be accomplished using one the following three schemes:
2351 (scheme 1) using reduc_code, if available.
2352 (scheme 2) using whole-vector shifts, if available.
2353 (scheme 3) using a scalar loop. In this case steps 1+2 above are
2354 combined.
2356 The overall epilog code looks like this:
2358 s_out0 = phi <s_loop> # original EXIT_PHI
2359 v_out1 = phi <VECT_DEF> # NEW_EXIT_PHI
2360 v_out2 = reduce <v_out1> # step 1
2361 s_out3 = extract_field <v_out2, 0> # step 2
2362 s_out4 = adjust_result <s_out3> # step 3
2364 (step 3 is optional, and step2 1 and 2 may be combined).
2365 Lastly, the uses of s_out0 are replaced by s_out4.
2367 ***/
2369 /* 2.1 Create new loop-exit-phi to preserve loop-closed form:
2370 v_out1 = phi <v_loop> */
2377 /* 2.2 Get the relevant tree-code to use in the epilog for schemes 2,3
2378 (i.e. when reduc_code is not available) and in the final adjustment
2379 code (if needed). Also get the original scalar reduction variable as
2380 defined in the loop. In case STMT is a "pattern-stmt" (i.e. - it
2381 represents a reduction pattern), the tree-code and scalar-def are
2382 taken from the original stmt that the pattern-stmt (STMT) replaces.
2383 Otherwise (it is a regular reduction) - the tree-code and scalar-def
2384 are taken from STMT. */
2388 {
2389 /* Regular reduction */
2391 }
2392 else
2393 {
2394 /* Reduction pattern */
2398 }
2407 /* In case this is a reduction in an inner-loop while vectorizing an outer
2408 loop - we don't need to extract a single scalar result at the end of the
2409 inner-loop. The final vector of partial results will be used in the
2410 vectorized outer-loop, or reduced to a scalar result at the end of the
2411 outer-loop. */
2415 /* 2.3 Create the reduction code, using one of the three schemes described
2416 above. */
2419 {
2420 tree tmp;
2422 /*** Case 1: Create:
2423 v_out2 = reduc_expr <v_out1> */
2436 }
2437 else
2438 {
2444 tree vec_temp;
2448 else
2451 /* Regardless of whether we have a whole vector shift, if we're
2452 emulating the operation via tree-vect-generic, we don't want
2453 to use it. Only the first round of the reduction is likely
2454 to still be profitable via emulation. */
2455 /* ??? It might be better to emit a reduction tree code here, so that
2456 tree-vect-generic can expand the first round via bit tricks. */
2459 else
2460 {
2464 }
2467 {
2468 /*** Case 2: Create:
2469 for (offset = VS/2; offset >= element_size; offset/=2)
2470 {
2471 Create: va' = vec_shift <va, offset>
2472 Create: va = vop <va, va'>
2473 } */
2484 {
2497 }
2500 }
2501 else
2502 {
2503 tree rhs;
2505 /*** Case 3: Create:
2506 s = extract_field <v_out2, 0>
2507 for (offset = element_size;
2508 offset < vector_size;
2509 offset += element_size;)
2510 {
2511 Create: s' = extract_field <v_out2, offset>
2512 Create: s = op <s, s'>
2513 } */
2521 bitsize_zero_node);
2530 {
2531 tree tmp;
2534 bitpos);
2546 }
2549 }
2550 }
2552 /* 2.4 Extract the final scalar result. Create:
2553 s_out3 = extract_field <v_out2, bitpos> */
2556 {
2557 tree rhs;
2567 else
2575 }
2577 vect_finalize_reduction:
2579 /* 2.5 Adjust the final result by the initial value of the reduction
2580 variable. (When such adjustment is not needed, then
2581 'adjustment_def' is zero). For example, if code is PLUS we create:
2582 new_temp = loop_exit_def + adjustment_def */
2585 {
2587 {
2591 }
2592 else
2593 {
2597 }
2602 }
2605 /* 2.6 Handle the loop-exit phi */
2607 /* Replace uses of s_out0 with uses of s_out3:
2608 Find the loop-closed-use at the loop exit of the original scalar result.
2609 (The reduction result is expected to have two immediate uses - one at the
2610 latch block, and one at the loop exit). */
2613 {
2615 {
2618 }
2619 }
2620 /* We expect to have found an exit_phi because of loop-closed-ssa form. */
2624 {
2626 {
2629 /* FORNOW. Currently not supporting the case that an inner-loop reduction
2630 is not used in the outer-loop (but only outside the outer-loop). */
2639 }
2641 /* Replace the uses: */
2646 }
2648 }
2651 /* Function vectorizable_reduction.
2653 Check if STMT performs a reduction operation that can be vectorized.
2654 If VEC_STMT is also passed, vectorize the STMT: create a vectorized
2655 stmt to replace it, put it in VEC_STMT, and insert it at BSI.
2656 Return FALSE if not a vectorizable STMT, TRUE otherwise.
2658 This function also handles reduction idioms (patterns) that have been
2659 recognized in advance during vect_pattern_recog. In this case, STMT may be
2660 of this form:
2661 X = pattern_expr (arg0, arg1, ..., X)
2662 and it's STMT_VINFO_RELATED_STMT points to the last stmt in the original
2663 sequence that had been detected and replaced by the pattern-stmt (STMT).
2665 In some cases of reduction patterns, the type of the reduction variable X is
2666 different than the type of the other arguments of STMT.
2667 In such cases, the vectype that is used when transforming STMT into a vector
2668 stmt is different than the vectype that is used to determine the
2669 vectorization factor, because it consists of a different number of elements
2670 than the actual number of elements that are being operated upon in parallel.
2672 For example, consider an accumulation of shorts into an int accumulator.
2673 On some targets it's possible to vectorize this pattern operating on 8
2674 shorts at a time (hence, the vectype for purposes of determining the
2675 vectorization factor should be V8HI); on the other hand, the vectype that
2676 is used to create the vector form is actually V4SI (the type of the result).
2678 Upon entry to this function, STMT_VINFO_VECTYPE records the vectype that
2679 indicates what is the actual level of parallelism (V8HI in the example), so
2680 that the right vectorization factor would be derived. This vectype
2681 corresponds to the type of arguments to the reduction stmt, and should *NOT*
2682 be used to create the vectorized stmt. The right vectype for the vectorized
2683 stmt is obtained from the type of the result X:
2684 get_vectype_for_scalar_type (TREE_TYPE (X))
2686 This means that, contrary to "regular" reductions (or "regular" stmts in
2687 general), the following equation:
2688 STMT_VINFO_VECTYPE == get_vectype_for_scalar_type (TREE_TYPE (X))
2689 does *NOT* necessarily hold for reduction patterns. */
2691 bool
2693 {
2694 tree vec_dest;
2695 tree scalar_dest;
2696 tree op;
2702 tree operation;
2710 tree new_phi;
2711 tree scalar_type;
2713 tree orig_stmt;
2714 stmt_vec_info orig_stmt_info;
2719 stmt_vec_info prev_stmt_info;
2720 tree reduc_def;
2725 {
2727 /* FORNOW. This restriction should be relaxed. */
2729 {
2733 }
2734 }
2738 /* FORNOW: SLP not supported. */
2742 /* 1. Is vectorizable reduction? */
2744 /* Not supportable if the reduction variable is used in the loop. */
2748 /* Reductions that are not used even in an enclosing outer-loop,
2749 are expected to be "live" (used out of the loop). */
2754 /* Make sure it was already recognized as a reduction computation. */
2758 /* 2. Has this been recognized as a reduction pattern?
2760 Check if STMT represents a pattern that has been recognized
2761 in earlier analysis stages. For stmts that represent a pattern,
2762 the STMT_VINFO_RELATED_STMT field records the last stmt in
2763 the original sequence that constitutes the pattern. */
2767 {
2772 }
2774 /* 3. Check the operands of the operation. The first operands are defined
2775 inside the loop body. The last operand is the reduction variable,
2776 which is defined by the loop-header-phi. */
2791 /* All uses but the last are expected to be defined in the loop.
2792 The last use is the reduction variable. */
2794 {
2803 }
2812 else
2818 /* 4. Supportable by target? */
2820 /* 4.1. check support for the operation in the loop */
2823 {
2827 }
2830 {
2839 }
2841 /* Worthwhile without SIMD support? */
2845 {
2849 }
2851 /* 4.2. Check support for the epilog operation.
2853 If STMT represents a reduction pattern, then the type of the
2854 reduction variable may be different than the type of the rest
2855 of the arguments. For example, consider the case of accumulation
2856 of shorts into an int accumulator; The original code:
2857 S1: int_a = (int) short_a;
2858 orig_stmt-> S2: int_acc = plus <int_a ,int_acc>;
2860 was replaced with:
2861 STMT: int_acc = widen_sum <short_a, int_acc>
2863 This means that:
2864 1. The tree-code that is used to create the vector operation in the
2865 epilog code (that reduces the partial results) is not the
2866 tree-code of STMT, but is rather the tree-code of the original
2867 stmt from the pattern that STMT is replacing. I.e, in the example
2868 above we want to use 'widen_sum' in the loop, but 'plus' in the
2869 epilog.
2870 2. The type (mode) we use to check available target support
2871 for the vector operation to be created in the *epilog*, is
2872 determined by the type of the reduction variable (in the example
2873 above we'd check this: plus_optab[vect_int_mode]).
2874 However the type (mode) we use to check available target support
2875 for the vector operation to be created *inside the loop*, is
2876 determined by the type of the other arguments to STMT (in the
2877 example we'd check this: widen_sum_optab[vect_short_mode]).
2879 This is contrary to "regular" reductions, in which the types of all
2880 the arguments are the same as the type of the reduction variable.
2881 For "regular" reductions we can therefore use the same vector type
2882 (and also the same tree-code) when generating the epilog code and
2883 when generating the code inside the loop. */
2886 {
2887 /* This is a reduction pattern: get the vectype from the type of the
2888 reduction variable, and get the tree-code from orig_stmt. */
2892 {
2894 {
2897 }
2899 }
2902 }
2903 else
2904 {
2905 /* Regular reduction: use the same vectype and tree-code as used for
2906 the vector code inside the loop can be used for the epilog code. */
2908 }
2914 {
2918 }
2920 {
2924 }
2927 {
2932 }
2934 /** Transform. **/
2939 /* Create the destination vector */
2942 /* Create the reduction-phi that defines the reduction-operand. */
2945 /* In case the vectorization factor (VF) is bigger than the number
2946 of elements that we can fit in a vectype (nunits), we have to generate
2947 more than one vector stmt - i.e - we need to "unroll" the
2948 vector stmt by a factor VF/nunits. For more details see documentation
2949 in vectorizable_operation. */
2953 {
2954 /* Handle uses. */
2956 {
2960 {
2963 }
2965 /* Get the vector def for the reduction variable from the phi node */
2967 }
2968 else
2969 {
2975 /* Get the vector def for the reduction variable from the vectorized
2976 reduction operation generated in the previous iteration (j-1) */
2978 }
2980 /* Arguments are ready. create the new vector stmt. */
2983 else
2985 reduc_def);
2993 else
2996 }
2998 /* Finalize the reduction-phi (set it's arguments) and create the
2999 epilog reduction code. */
3002 }
3004 /* Checks if CALL can be vectorized in type VECTYPE. Returns
3005 a function declaration if the target has a vectorized version
3006 of the function, or NULL_TREE if the function cannot be vectorized. */
3008 tree
3010 {
3014 /* We only handle functions that do not read or clobber memory -- i.e.
3015 const or novops ones. */
3026 vectype_in);
3027 }
3029 /* Function vectorizable_call.
3031 Check if STMT performs a function call that can be vectorized.
3032 If VEC_STMT is also passed, vectorize the STMT: create a vectorized
3033 stmt to replace it, put it in VEC_STMT, and insert it at BSI.
3034 Return FALSE if not a vectorizable STMT, TRUE otherwise. */
3036 bool
3038 {
3039 tree vec_dest;
3040 tree scalar_dest;
3041 tree operation;
3052 tree new_stmt;
3054 call_expr_arg_iterator iter;
3055 tree vargs;
3064 /* FORNOW: SLP not supported. */
3068 /* Is STMT a vectorizable call? */
3079 /* Process function arguments. */
3083 {
3084 /* Bail out if the function has more than two arguments, we
3085 do not have interesting builtin functions to vectorize with
3086 more than two arguments. */
3090 /* We can only handle calls with arguments of the same type. */
3093 {
3097 }
3101 {
3105 }
3108 }
3110 /* No arguments is also not good. */
3125 /* FORNOW */
3132 else
3135 /* For now, we only vectorize functions if a target specific builtin
3136 is available. TODO -- in some cases, it might be profitable to
3137 insert the calls for pieces of the vector, in order to be able
3138 to vectorize other operations in the loop. */
3141 {
3146 }
3152 else
3155 /* Sanity check: make sure that at least one copy of the vectorized stmt
3156 needs to be generated. */
3159 /* FORNOW. This restriction should be relaxed. */
3161 {
3165 }
3168 {
3174 }
3176 /** Transform. **/
3181 /* FORNOW. This restriction should be relaxed. */
3183 {
3187 }
3189 /* Handle def. */
3195 {
3198 {
3199 /* Build argument list for the vectorized call. */
3200 /* FIXME: Rewrite this so that it doesn't
3201 construct a temporary list. */
3205 {
3207 vec_oprnd0
3209 else
3210 vec_oprnd0
3216 }
3228 else
3232 }
3238 {
3239 /* Build argument list for the vectorized call. */
3240 /* FIXME: Rewrite this so that it doesn't
3241 construct a temporary list. */
3245 {
3247 {
3248 vec_oprnd0
3250 vec_oprnd1
3252 }
3253 else
3254 {
3255 vec_oprnd0
3257 vec_oprnd1
3259 }
3265 }
3277 else
3281 }
3288 /* No current target implements this case. */
3290 }
3292 /* The call in STMT might prevent it from being removed in dce.
3293 We however cannot remove it here, due to the way the ssa name
3294 it defines is mapped to the new definition. So just replace
3295 rhs of the statement with something harmless. */
3301 }
3304 /* Function vect_gen_widened_results_half
3306 Create a vector stmt whose code, type, number of arguments, and result
3307 variable are CODE, VECTYPE, OP_TYPE, and VEC_DEST, and its arguments are
3308 VEC_OPRND0 and VEC_OPRND1. The new vector stmt is to be inserted at BSI.
3309 In the case that CODE is a CALL_EXPR, this means that a call to DECL
3310 needs to be created (DECL is a function-decl of a target-builtin).
3311 STMT is the original scalar stmt that we are vectorizing. */
3313 static tree
3317 tree stmt)
3318 {
3319 tree expr;
3320 tree new_stmt;
3321 tree new_temp;
3322 tree sym;
3323 ssa_op_iter iter;
3325 /* Generate half of the widened result: */
3327 {
3328 /* Target specific support */
3331 else
3333 }
3334 else
3335 {
3336 /* Generic support */
3340 else
3342 }
3349 {
3351 {
3355 }
3356 }
3359 }
3362 /* Check if STMT performs a conversion operation, that can be vectorized.
3363 If VEC_STMT is also passed, vectorize the STMT: create a vectorized
3364 stmt to replace it, put it in VEC_STMT, and insert it at BSI.
3365 Return FALSE if not a vectorizable STMT, TRUE otherwise. */
3367 bool
3370 {
3371 tree vec_dest;
3372 tree scalar_dest;
3373 tree operation;
3374 tree op0;
3381 tree new_temp;
3385 stmt_vec_info prev_stmt_info;
3390 tree expr;
3392 tree builtin_decl;
3396 tree vop0;
3398 /* Is STMT a vectorizable conversion? */
3417 /* Check types of lhs and rhs. */
3432 /* FORNOW */
3439 else
3445 /* Bail out if the types are both integral or non-integral. */
3452 else
3455 /* FORNOW: SLP with multiple types is not supported. The SLP analysis verifies
3456 this, so we can safely override NCOPIES with 1 here. */
3460 /* Sanity check: make sure that at least one copy of the vectorized stmt
3461 needs to be generated. */
3464 /* FORNOW. This restriction should be relaxed. */
3466 {
3470 }
3472 /* Check the operands of the operation. */
3474 {
3478 }
3480 /* Supportable by target? */
3490 {
3494 }
3497 {
3499 /* FORNOW: SLP not supported. */
3502 }
3505 {
3508 }
3510 /** Transform. **/
3514 /* Handle def. */
3522 {
3525 {
3526 tree sym;
3527 ssa_op_iter iter;
3531 else
3534 builtin_decl =
3537 {
3540 /* Arguments are ready. create the new vector stmt. */
3546 SSA_OP_ALL_VIRTUALS)
3547 {
3551 }
3554 }
3558 else
3561 }
3565 /* In case the vectorization factor (VF) is bigger than the number
3566 of elements that we can fit in a vectype (nunits), we have to
3567 generate more than one vector stmt - i.e - we need to "unroll"
3568 the vector stmt by a factor VF/nunits. */
3570 {
3573 else
3578 /* Generate first half of the widened result: */
3579 new_stmt
3585 else
3589 /* Generate second half of the widened result: */
3590 new_stmt
3596 }
3600 /* In case the vectorization factor (VF) is bigger than the number
3601 of elements that we can fit in a vectype (nunits), we have to
3602 generate more than one vector stmt - i.e - we need to "unroll"
3603 the vector stmt by a factor VF/nunits. */
3605 {
3606 /* Handle uses. */
3608 {
3611 }
3612 else
3613 {
3616 }
3618 /* Arguments are ready. Create the new vector stmt. */
3627 else
3631 }
3634 }
3640 }
3643 /* Function vectorizable_assignment.
3645 Check if STMT performs an assignment (copy) that can be vectorized.
3646 If VEC_STMT is also passed, vectorize the STMT: create a vectorized
3647 stmt to replace it, put it in VEC_STMT, and insert it at BSI.
3648 Return FALSE if not a vectorizable STMT, TRUE otherwise. */
3650 bool
3652 slp_tree slp_node)
3653 {
3654 tree vec_dest;
3655 tree scalar_dest;
3656 tree op;
3660 tree new_temp;
3667 tree vop;
3669 /* FORNOW: SLP with multiple types is not supported. The SLP analysis
3670 verifies this, so we can safely override NCOPIES with 1 here. */
3684 /* Is vectorizable assignment? */
3694 {
3698 }
3701 {
3707 }
3709 /** Transform. **/
3713 /* Handle def. */
3716 /* Handle use. */
3719 /* Arguments are ready. create the new vector stmt. */
3721 {
3730 }
3734 }
3737 /* Function vect_min_worthwhile_factor.
3739 For a loop where we could vectorize the operation indicated by CODE,
3740 return the minimum vectorization factor that makes it worthwhile
3741 to use generic vectors. */
3742 static int
3744 {
3746 {
3760 }
3761 }
3764 /* Function vectorizable_induction
3766 Check if PHI performs an induction computation that can be vectorized.
3767 If VEC_STMT is also passed, vectorize the induction PHI: create a vectorized
3768 phi to replace it, put it in VEC_STMT, and add it to the same basic block.
3769 Return FALSE if not a vectorizable STMT, TRUE otherwise. */
3771 bool
3774 {
3780 tree vec_def;
3787 /* FORNOW: SLP not supported. */
3797 {
3803 }
3805 /** Transform. **/
3813 }
3816 /* Function vectorizable_operation.
3818 Check if STMT performs a binary or unary operation that can be vectorized.
3819 If VEC_STMT is also passed, vectorize the STMT: create a vectorized
3820 stmt to replace it, put it in VEC_STMT, and insert it at BSI.
3821 Return FALSE if not a vectorizable STMT, TRUE otherwise. */
3823 bool
3825 slp_tree slp_node)
3826 {
3827 tree vec_dest;
3828 tree scalar_dest;
3829 tree operation;
3838 tree new_temp;
3840 optab optab;
3846 stmt_vec_info prev_stmt_info;
3849 tree vectype_out;
3858 /* FORNOW: SLP with multiple types is not supported. The SLP analysis verifies
3859 this, so we can safely override NCOPIES with 1 here. */
3863 /* FORNOW. This restriction should be relaxed. */
3865 {
3869 }
3877 /* Is STMT a vectorizable binary/unary operation? */
3895 /* For pointer addition, we should use the normal plus for
3896 the vector addition. */
3900 /* Support only unary or binary operations. */
3903 {
3907 }
3911 {
3915 }
3918 {
3921 {
3925 }
3926 }
3928 /* If this is a shift/rotate, determine whether the shift amount is a vector,
3929 or scalar. If the shift/rotate amount is a vector, use the vector/vector
3930 shift optabs. */
3933 {
3936 /* vector shifted by vector */
3938 {
3942 }
3944 /* See if the machine has a vector shifted by scalar insn and if not
3945 then see if it has a vector shifted by vector insn */
3947 {
3952 {
3956 }
3957 else
3958 {
3961 && optab
3965 }
3966 }
3968 else
3969 {
3973 }
3974 }
3975 else
3978 /* Supportable by target? */
3980 {
3984 }
3988 {
3991 /* Check only during analysis. */
3999 }
4001 /* Worthwhile without SIMD support? Check only during analysis. */
4006 {
4010 }
4013 {
4019 }
4021 /** Transform. **/
4026 /* Handle def. */
4029 /* Allocate VECs for vector operands. In case of SLP, vector operands are
4030 created in the previous stages of the recursion, so no allocation is
4031 needed, except for the case of shift with scalar shift argument. In that
4032 case we store the scalar operand in VEC_OPRNDS1 for every vector stmt to
4033 be created to vectorize the SLP group, i.e., SLP_NODE->VEC_STMTS_SIZE.
4034 In case of loop-based vectorization we allocate VECs of size 1. We
4035 allocate VEC_OPRNDS1 only in case of binary operation. */
4037 {
4041 }
4045 /* In case the vectorization factor (VF) is bigger than the number
4046 of elements that we can fit in a vectype (nunits), we have to generate
4047 more than one vector stmt - i.e - we need to "unroll" the
4048 vector stmt by a factor VF/nunits. In doing so, we record a pointer
4049 from one copy of the vector stmt to the next, in the field
4050 STMT_VINFO_RELATED_STMT. This is necessary in order to allow following
4051 stages to find the correct vector defs to be used when vectorizing
4052 stmts that use the defs of the current stmt. The example below illustrates
4053 the vectorization process when VF=16 and nunits=4 (i.e - we need to create
4054 4 vectorized stmts):
4056 before vectorization:
4057 RELATED_STMT VEC_STMT
4058 S1: x = memref - -
4059 S2: z = x + 1 - -
4061 step 1: vectorize stmt S1 (done in vectorizable_load. See more details
4062 there):
4063 RELATED_STMT VEC_STMT
4064 VS1_0: vx0 = memref0 VS1_1 -
4065 VS1_1: vx1 = memref1 VS1_2 -
4066 VS1_2: vx2 = memref2 VS1_3 -
4067 VS1_3: vx3 = memref3 - -
4068 S1: x = load - VS1_0
4069 S2: z = x + 1 - -
4071 step2: vectorize stmt S2 (done here):
4072 To vectorize stmt S2 we first need to find the relevant vector
4073 def for the first operand 'x'. This is, as usual, obtained from
4074 the vector stmt recorded in the STMT_VINFO_VEC_STMT of the stmt
4075 that defines 'x' (S1). This way we find the stmt VS1_0, and the
4076 relevant vector def 'vx0'. Having found 'vx0' we can generate
4077 the vector stmt VS2_0, and as usual, record it in the
4078 STMT_VINFO_VEC_STMT of stmt S2.
4079 When creating the second copy (VS2_1), we obtain the relevant vector
4080 def from the vector stmt recorded in the STMT_VINFO_RELATED_STMT of
4081 stmt VS1_0. This way we find the stmt VS1_1 and the relevant
4082 vector def 'vx1'. Using 'vx1' we create stmt VS2_1 and record a
4083 pointer to it in the STMT_VINFO_RELATED_STMT of the vector stmt VS2_0.
4084 Similarly when creating stmts VS2_2 and VS2_3. This is the resulting
4085 chain of stmts and pointers:
4086 RELATED_STMT VEC_STMT
4087 VS1_0: vx0 = memref0 VS1_1 -
4088 VS1_1: vx1 = memref1 VS1_2 -
4089 VS1_2: vx2 = memref2 VS1_3 -
4090 VS1_3: vx3 = memref3 - -
4091 S1: x = load - VS1_0
4092 VS2_0: vz0 = vx0 + v1 VS2_1 -
4093 VS2_1: vz1 = vx1 + v1 VS2_2 -
4094 VS2_2: vz2 = vx2 + v1 VS2_3 -
4095 VS2_3: vz3 = vx3 + v1 - -
4096 S2: z = x + 1 - VS2_0 */
4100 {
4101 /* Handle uses. */
4103 {
4105 {
4106 /* Vector shl and shr insn patterns can be defined with scalar
4107 operand 2 (shift operand). In this case, use constant or loop
4108 invariant op1 directly, without extending it to vector mode
4109 first. */
4112 {
4118 {
4119 /* Store vec_oprnd1 for every vector stmt to be created
4120 for SLP_NODE. We check during the analysis that all the
4121 shift arguments are the same.
4122 TODO: Allow different constants for different vector
4123 stmts generated for an SLP instance. */
4126 }
4127 }
4128 }
4130 /* vec_oprnd1 is available if operand 1 should be of a scalar-type
4131 (a special case for certain kind of vector shifts); otherwise,
4132 operand 1 should be of a vector type (the usual case). */
4135 slp_node);
4136 else
4138 slp_node);
4139 }
4140 else
4143 /* Arguments are ready. Create the new vector stmt. */
4145 {
4147 {
4151 }
4152 else
4161 }
4165 else
4168 }
4175 }
4178 /* Function vectorizable_type_demotion
4180 Check if STMT performs a binary or unary operation that involves
4181 type demotion, and if it can be vectorized.
4182 If VEC_STMT is also passed, vectorize the STMT: create a vectorized
4183 stmt to replace it, put it in VEC_STMT, and insert it at BSI.
4184 Return FALSE if not a vectorizable STMT, TRUE otherwise. */
4186 bool
4189 {
4190 tree vec_dest;
4191 tree scalar_dest;
4192 tree operation;
4193 tree op0;
4199 tree new_temp;
4202 tree new_stmt;
4203 stmt_vec_info prev_stmt_info;
4206 tree vectype_out;
4209 tree expr;
4210 tree vectype_in;
4218 /* Is STMT a vectorizable type-demotion operation? */
4246 /* FORNOW. This restriction should be relaxed. */
4248 {
4252 }
4261 /* Check the operands of the operation. */
4263 {
4267 }
4269 /* Supportable by target? */
4276 {
4282 }
4284 /** Transform. **/
4287 ncopies);
4289 /* Handle def. */
4292 /* In case the vectorization factor (VF) is bigger than the number
4293 of elements that we can fit in a vectype (nunits), we have to generate
4294 more than one vector stmt - i.e - we need to "unroll" the
4295 vector stmt by a factor VF/nunits. */
4298 {
4299 /* Handle uses. */
4301 {
4304 }
4305 else
4306 {
4309 }
4311 /* Arguments are ready. Create the new vector stmt. */
4320 else
4324 }
4328 }
4331 /* Function vectorizable_type_promotion
4333 Check if STMT performs a binary or unary operation that involves
4334 type promotion, and if it can be vectorized.
4335 If VEC_STMT is also passed, vectorize the STMT: create a vectorized
4336 stmt to replace it, put it in VEC_STMT, and insert it at BSI.
4337 Return FALSE if not a vectorizable STMT, TRUE otherwise. */
4339 bool
4342 {
4343 tree vec_dest;
4344 tree scalar_dest;
4345 tree operation;
4356 tree new_stmt;
4357 stmt_vec_info prev_stmt_info;
4360 tree vectype_out;
4363 tree vectype_in;
4371 /* Is STMT a vectorizable type-promotion operation? */
4400 /* FORNOW. This restriction should be relaxed. */
4402 {
4406 }
4415 /* Check the operands of the operation. */
4417 {
4421 }
4425 {
4428 {
4432 }
4433 }
4435 /* Supportable by target? */
4443 {
4449 }
4451 /** Transform. **/
4455 ncopies);
4457 /* Handle def. */
4460 /* In case the vectorization factor (VF) is bigger than the number
4461 of elements that we can fit in a vectype (nunits), we have to generate
4462 more than one vector stmt - i.e - we need to "unroll" the
4463 vector stmt by a factor VF/nunits. */
4467 {
4468 /* Handle uses. */
4470 {
4474 }
4475 else
4476 {
4480 }
4482 /* Arguments are ready. Create the new vector stmt. We are creating
4483 two vector defs because the widened result does not fit in one vector.
4484 The vectorized stmt can be expressed as a call to a taregt builtin,
4485 or a using a tree-code. */
4486 /* Generate first half of the widened result: */
4491 else
4495 /* Generate second half of the widened result: */
4501 }
4505 }
4508 /* Function vect_strided_store_supported.
4510 Returns TRUE is INTERLEAVE_HIGH and INTERLEAVE_LOW operations are supported,
4511 and FALSE otherwise. */
4513 static bool
4515 {
4521 /* Check that the operation is supported. */
4527 {
4531 }
4534 == CODE_FOR_nothing
4537 {
4541 }
4544 }
4547 /* Function vect_permute_store_chain.
4549 Given a chain of interleaved stores in DR_CHAIN of LENGTH that must be
4550 a power of 2, generate interleave_high/low stmts to reorder the data
4551 correctly for the stores. Return the final references for stores in
4552 RESULT_CHAIN.
4554 E.g., LENGTH is 4 and the scalar type is short, i.e., VF is 8.
4555 The input is 4 vectors each containing 8 elements. We assign a number to each
4556 element, the input sequence is:
4558 1st vec: 0 1 2 3 4 5 6 7
4559 2nd vec: 8 9 10 11 12 13 14 15
4560 3rd vec: 16 17 18 19 20 21 22 23
4561 4th vec: 24 25 26 27 28 29 30 31
4563 The output sequence should be:
4565 1st vec: 0 8 16 24 1 9 17 25
4566 2nd vec: 2 10 18 26 3 11 19 27
4567 3rd vec: 4 12 20 28 5 13 21 30
4568 4th vec: 6 14 22 30 7 15 23 31
4570 i.e., we interleave the contents of the four vectors in their order.
4572 We use interleave_high/low instructions to create such output. The input of
4573 each interleave_high/low operation is two vectors:
4574 1st vec 2nd vec
4575 0 1 2 3 4 5 6 7
4576 the even elements of the result vector are obtained left-to-right from the
4577 high/low elements of the first vector. The odd elements of the result are
4578 obtained left-to-right from the high/low elements of the second vector.
4579 The output of interleave_high will be: 0 4 1 5
4580 and of interleave_low: 2 6 3 7
4583 The permutation is done in log LENGTH stages. In each stage interleave_high
4584 and interleave_low stmts are created for each pair of vectors in DR_CHAIN,
4585 where the first argument is taken from the first half of DR_CHAIN and the
4586 second argument from it's second half.
4587 In our example,
4589 I1: interleave_high (1st vec, 3rd vec)
4590 I2: interleave_low (1st vec, 3rd vec)
4591 I3: interleave_high (2nd vec, 4th vec)
4592 I4: interleave_low (2nd vec, 4th vec)
4594 The output for the first stage is:
4596 I1: 0 16 1 17 2 18 3 19
4597 I2: 4 20 5 21 6 22 7 23
4598 I3: 8 24 9 25 10 26 11 27
4599 I4: 12 28 13 29 14 30 15 31
4601 The output of the second stage, i.e. the final result is:
4603 I1: 0 8 16 24 1 9 17 25
4604 I2: 2 10 18 26 3 11 19 27
4605 I3: 4 12 20 28 5 13 21 30
4606 I4: 6 14 22 30 7 15 23 31. */
4608 static bool
4611 tree stmt,
4614 {
4623 /* Check that the operation is supported. */
4630 {
4632 {
4636 /* Create interleaving stmt:
4637 in the case of big endian:
4638 high = interleave_high (vect1, vect2)
4639 and in the case of little endian:
4640 high = interleave_low (vect1, vect2). */
4646 else
4654 /* Create interleaving stmt:
4655 in the case of big endian:
4656 low = interleave_low (vect1, vect2)
4657 and in the case of little endian:
4658 low = interleave_high (vect1, vect2). */
4664 else
4671 }
4673 }
4675 }
4678 /* Function vectorizable_store.
4680 Check if STMT defines a non scalar data-ref (array/pointer/structure) that
4681 can be vectorized.
4682 If VEC_STMT is also passed, vectorize the STMT: create a vectorized
4683 stmt to replace it, put it in VEC_STMT, and insert it at BSI.
4684 Return FALSE if not a vectorizable STMT, TRUE otherwise. */
4686 bool
4688 slp_tree slp_node)
4689 {
4690 tree scalar_dest;
4691 tree data_ref;
4692 tree op;
4700 tree dummy;
4716 stmt_vec_info first_stmt_vinfo;
4719 /* FORNOW: SLP with multiple types is not supported. The SLP analysis verifies
4720 this, so we can safely override NCOPIES with 1 here. */
4726 /* FORNOW. This restriction should be relaxed. */
4728 {
4732 }
4740 /* Is vectorizable store? */
4753 {
4757 }
4760 /* FORNOW. In some cases can vectorize even if data-type not supported
4761 (e.g. - array initialization with 0). */
4769 {
4777 {
4778 /* STMT is the leader of the group. Check the operands of all the
4779 stmts of the group. */
4782 {
4785 {
4789 }
4791 }
4792 }
4793 }
4796 {
4801 }
4803 /** Transform. **/
4806 {
4812 /* FORNOW */
4815 /* We vectorize all the stmts of the interleaving group when we
4816 reach the last stmt in the group. */
4820 {
4823 }
4828 /* VEC_NUM is the number of vect stmts to be created for this group. */
4831 else
4833 }
4834 else
4835 {
4840 }
4852 /* In case the vectorization factor (VF) is bigger than the number
4853 of elements that we can fit in a vectype (nunits), we have to generate
4854 more than one vector stmt - i.e - we need to "unroll" the
4855 vector stmt by a factor VF/nunits. For more details see documentation in
4856 vect_get_vec_def_for_copy_stmt. */
4858 /* In case of interleaving (non-unit strided access):
4860 S1: &base + 2 = x2
4861 S2: &base = x0
4862 S3: &base + 1 = x1
4863 S4: &base + 3 = x3
4865 We create vectorized stores starting from base address (the access of the
4866 first stmt in the chain (S2 in the above example), when the last store stmt
4867 of the chain (S4) is reached:
4869 VS1: &base = vx2
4870 VS2: &base + vec_size*1 = vx0
4871 VS3: &base + vec_size*2 = vx1
4872 VS4: &base + vec_size*3 = vx3
4874 Then permutation statements are generated:
4876 VS5: vx5 = VEC_INTERLEAVE_HIGH_EXPR < vx0, vx3 >
4877 VS6: vx6 = VEC_INTERLEAVE_LOW_EXPR < vx0, vx3 >
4878 ...
4880 And they are put in STMT_VINFO_VEC_STMT of the corresponding scalar stmts
4881 (the order of the data-refs in the output of vect_permute_store_chain
4882 corresponds to the order of scalar stmts in the interleaving chain - see
4883 the documentation of vect_permute_store_chain()).
4885 In case of both multiple types and interleaving, above vector stores and
4886 permutation stmts are created for every copy. The result vector stmts are
4887 put in STMT_VINFO_VEC_STMT for the first copy and in the corresponding
4888 STMT_VINFO_RELATED_STMT for the next copies.
4889 */
4893 {
4894 tree new_stmt;
4895 tree ptr_incr;
4898 {
4900 {
4901 /* Get vectorized arguments for SLP_NODE. */
4905 }
4906 else
4907 {
4908 /* For interleaved stores we collect vectorized defs for all the
4909 stores in the group in DR_CHAIN and OPRNDS. DR_CHAIN is then
4910 used as an input to vect_permute_store_chain(), and OPRNDS as
4911 an input to vect_get_vec_def_for_stmt_copy() for the next copy.
4913 If the store is not strided, GROUP_SIZE is 1, and DR_CHAIN and
4914 OPRNDS are of size 1. */
4917 {
4918 /* Since gaps are not supported for interleaved stores,
4919 GROUP_SIZE is the exact number of stmts in the chain.
4920 Therefore, NEXT_STMT can't be NULL_TREE. In case that
4921 there is no interleaving, GROUP_SIZE is 1, and only one
4922 iteration of the loop will be executed. */
4927 NULL);
4931 }
4932 }
4937 }
4938 else
4939 {
4940 /* FORNOW SLP doesn't work for multiple types. */
4943 /* For interleaved stores we created vectorized defs for all the
4944 defs stored in OPRNDS in the previous iteration (previous copy).
4945 DR_CHAIN is then used as an input to vect_permute_store_chain(),
4946 and OPRNDS as an input to vect_get_vec_def_for_stmt_copy() for the
4947 next copy.
4948 If the store is not strided, GROUP_SIZE is 1, and DR_CHAIN and
4949 OPRNDS are of size 1. */
4951 {
4957 }
4958 dataref_ptr =
4960 }
4963 {
4965 /* Permute. */
4969 }
4973 {
4975 /* Bump the vector pointer. */
4977 NULL_TREE);
4982 /* For strided stores vectorized defs are interleaved in
4983 vect_permute_store_chain(). */
4987 /* Arguments are ready. Create the new vector stmt. */
4994 else
5001 }
5002 }
5010 }
5013 /* Function vect_setup_realignment
5015 This function is called when vectorizing an unaligned load using
5016 the dr_explicit_realign[_optimized] scheme.
5017 This function generates the following code at the loop prolog:
5019 p = initial_addr;
5020 x msq_init = *(floor(p)); # prolog load
5021 realignment_token = call target_builtin;
5022 loop:
5023 x msq = phi (msq_init, ---)
5025 The stmts marked with x are generated only for the case of
5026 dr_explicit_realign_optimized.
5028 The code above sets up a new (vector) pointer, pointing to the first
5029 location accessed by STMT, and a "floor-aligned" load using that pointer.
5030 It also generates code to compute the "realignment-token" (if the relevant
5031 target hook was defined), and creates a phi-node at the loop-header bb
5032 whose arguments are the result of the prolog-load (created by this
5033 function) and the result of a load that takes place in the loop (to be
5034 created by the caller to this function).
5036 For the case of dr_explicit_realign_optimized:
5037 The caller to this function uses the phi-result (msq) to create the
5038 realignment code inside the loop, and sets up the missing phi argument,
5039 as follows:
5040 loop:
5041 msq = phi (msq_init, lsq)
5042 lsq = *(floor(p')); # load in loop
5043 result = realign_load (msq, lsq, realignment_token);
5045 For the case of dr_explicit_realign:
5046 loop:
5047 msq = *(floor(p)); # load in loop
5048 p' = p + (VS-1);
5049 lsq = *(floor(p')); # load in loop
5050 result = realign_load (msq, lsq, realignment_token);
5052 Input:
5053 STMT - (scalar) load stmt to be vectorized. This load accesses
5054 a memory location that may be unaligned.
5055 BSI - place where new code is to be inserted.
5056 ALIGNMENT_SUPPORT_SCHEME - which of the two misalignment handling schemes
5057 is used.
5059 Output:
5060 REALIGNMENT_TOKEN - the result of a call to the builtin_mask_for_load
5061 target hook, if defined.
5062 Return value - the result of the loop-header phi node. */
5064 static tree
5068 tree init_addr,
5070 {
5075 edge pe;
5077 tree vec_dest;
5078 tree inc;
5079 tree ptr;
5080 tree data_ref;
5081 tree new_stmt;
5082 basic_block new_bb;
5084 tree new_temp;
5085 tree phi_stmt;
5097 /* We need to generate three things:
5098 1. the misalignment computation
5099 2. the extra vector load (for the optimized realignment scheme).
5100 3. the phi node for the two vectors from which the realignment is
5101 done (for the optimized realignment scheme).
5102 */
5104 /* 1. Determine where to generate the misalignment computation.
5106 If INIT_ADDR is NULL_TREE, this indicates that the misalignment
5107 calculation will be generated by this function, outside the loop (in the
5108 preheader). Otherwise, INIT_ADDR had already been computed for us by the
5109 caller, inside the loop.
5111 Background: If the misalignment remains fixed throughout the iterations of
5112 the loop, then both realignment schemes are applicable, and also the
5113 misalignment computation can be done outside LOOP. This is because we are
5114 vectorizing LOOP, and so the memory accesses in LOOP advance in steps that
5115 are a multiple of VS (the Vector Size), and therefore the misalignment in
5116 different vectorized LOOP iterations is always the same.
5117 The problem arises only if the memory access is in an inner-loop nested
5118 inside LOOP, which is now being vectorized using outer-loop vectorization.
5119 This is the only case when the misalignment of the memory access may not
5120 remain fixed throughout the iterations of the inner-loop (as explained in
5121 detail in vect_supportable_dr_alignment). In this case, not only is the
5122 optimized realignment scheme not applicable, but also the misalignment
5123 computation (and generation of the realignment token that is passed to
5124 REALIGN_LOAD) have to be done inside the loop.
5126 In short, INIT_ADDR indicates whether we are in a COMPUTE_IN_LOOP mode
5127 or not, which in turn determines if the misalignment is computed inside
5128 the inner-loop, or outside LOOP. */
5131 {
5134 }
5137 /* 2. Determine where to generate the extra vector load.
5139 For the optimized realignment scheme, instead of generating two vector
5140 loads in each iteration, we generate a single extra vector load in the
5141 preheader of the loop, and in each iteration reuse the result of the
5142 vector load from the previous iteration. In case the memory access is in
5143 an inner-loop nested inside LOOP, which is now being vectorized using
5144 outer-loop vectorization, we need to determine whether this initial vector
5145 load should be generated at the preheader of the inner-loop, or can be
5146 generated at the preheader of LOOP. If the memory access has no evolution
5147 in LOOP, it can be generated in the preheader of LOOP. Otherwise, it has
5148 to be generated inside LOOP (in the preheader of the inner-loop). */
5151 {
5156 }
5157 else
5162 /* 3. For the case of the optimized realignment, create the first vector
5163 load at the loop preheader. */
5166 {
5167 /* Create msq_init = *(floor(p1)) in the loop preheader */
5182 }
5184 /* 4. Create realignment token using a target builtin, if available.
5185 It is done either inside the containing loop, or before LOOP (as
5186 determined above). */
5189 {
5190 tree builtin_decl;
5192 /* Compute INIT_ADDR - the initial addressed accessed by this memref. */
5195 else
5196 {
5197 /* Generate the INIT_ADDR computation outside LOOP. */
5203 }
5215 else
5216 {
5217 /* Generate the misalignment computation outside LOOP. */
5221 }
5225 /* The result of the CALL_EXPR to this builtin is determined from
5226 the value of the parameter and no global variables are touched
5227 which makes the builtin a "const" function. Requiring the
5228 builtin to have the "const" attribute makes it unnecessary
5229 to call mark_call_clobbered. */
5231 }
5240 /* 5. Create msq = phi <msq_init, lsq> in loop */
5250 }
5253 /* Function vect_strided_load_supported.
5255 Returns TRUE is EXTRACT_EVEN and EXTRACT_ODD operations are supported,
5256 and FALSE otherwise. */
5258 static bool
5260 {
5267 optab_default);
5269 {
5273 }
5276 {
5280 }
5283 optab_default);
5285 {
5289 }
5292 {
5296 }
5298 }
5301 /* Function vect_permute_load_chain.
5303 Given a chain of interleaved loads in DR_CHAIN of LENGTH that must be
5304 a power of 2, generate extract_even/odd stmts to reorder the input data
5305 correctly. Return the final references for loads in RESULT_CHAIN.
5307 E.g., LENGTH is 4 and the scalar type is short, i.e., VF is 8.
5308 The input is 4 vectors each containing 8 elements. We assign a number to each
5309 element, the input sequence is:
5311 1st vec: 0 1 2 3 4 5 6 7
5312 2nd vec: 8 9 10 11 12 13 14 15
5313 3rd vec: 16 17 18 19 20 21 22 23
5314 4th vec: 24 25 26 27 28 29 30 31
5316 The output sequence should be:
5318 1st vec: 0 4 8 12 16 20 24 28
5319 2nd vec: 1 5 9 13 17 21 25 29
5320 3rd vec: 2 6 10 14 18 22 26 30
5321 4th vec: 3 7 11 15 19 23 27 31
5323 i.e., the first output vector should contain the first elements of each
5324 interleaving group, etc.
5326 We use extract_even/odd instructions to create such output. The input of each
5327 extract_even/odd operation is two vectors
5328 1st vec 2nd vec
5329 0 1 2 3 4 5 6 7
5331 and the output is the vector of extracted even/odd elements. The output of
5332 extract_even will be: 0 2 4 6
5333 and of extract_odd: 1 3 5 7
5336 The permutation is done in log LENGTH stages. In each stage extract_even and
5337 extract_odd stmts are created for each pair of vectors in DR_CHAIN in their
5338 order. In our example,
5340 E1: extract_even (1st vec, 2nd vec)
5341 E2: extract_odd (1st vec, 2nd vec)
5342 E3: extract_even (3rd vec, 4th vec)
5343 E4: extract_odd (3rd vec, 4th vec)
5345 The output for the first stage will be:
5347 E1: 0 2 4 6 8 10 12 14
5348 E2: 1 3 5 7 9 11 13 15
5349 E3: 16 18 20 22 24 26 28 30
5350 E4: 17 19 21 23 25 27 29 31
5352 In order to proceed and create the correct sequence for the next stage (or
5353 for the correct output, if the second stage is the last one, as in our
5354 example), we first put the output of extract_even operation and then the
5355 output of extract_odd in RESULT_CHAIN (which is then copied to DR_CHAIN).
5356 The input for the second stage is:
5358 1st vec (E1): 0 2 4 6 8 10 12 14
5359 2nd vec (E3): 16 18 20 22 24 26 28 30
5360 3rd vec (E2): 1 3 5 7 9 11 13 15
5361 4th vec (E4): 17 19 21 23 25 27 29 31
5363 The output of the second stage:
5365 E1: 0 4 8 12 16 20 24 28
5366 E2: 2 6 10 14 18 22 26 30
5367 E3: 1 5 9 13 17 21 25 29
5368 E4: 3 7 11 15 19 23 27 31
5370 And RESULT_CHAIN after reordering:
5372 1st vec (E1): 0 4 8 12 16 20 24 28
5373 2nd vec (E3): 1 5 9 13 17 21 25 29
5374 3rd vec (E2): 2 6 10 14 18 22 26 30
5375 4th vec (E4): 3 7 11 15 19 23 27 31. */
5377 static bool
5380 tree stmt,
5383 {
5386 tree tmp;
5390 /* Check that the operation is supported. */
5396 {
5398 {
5402 /* data_ref = permute_even (first_data_ref, second_data_ref); */
5418 /* data_ref = permute_odd (first_data_ref, second_data_ref); */
5432 }
5434 }
5436 }
5439 /* Function vect_transform_strided_load.
5441 Given a chain of input interleaved data-refs (in DR_CHAIN), build statements
5442 to perform their permutation and ascribe the result vectorized statements to
5443 the scalar statements.
5444 */
5446 static bool
5449 {
5455 tree tmp_data_ref;
5457 /* DR_CHAIN contains input data-refs that are a part of the interleaving.
5458 RESULT_CHAIN is the output of vect_permute_load_chain, it contains permuted
5459 vectors, that are ready for vector computation. */
5461 /* Permute. */
5465 /* Put a permuted data-ref in the VECTORIZED_STMT field.
5466 Since we scan the chain starting from it's first node, their order
5467 corresponds the order of data-refs in RESULT_CHAIN. */
5471 {
5475 /* Skip the gaps. Loads created for the gaps will be removed by dead
5476 code elimination pass later.
5477 DR_GROUP_GAP is the number of steps in elements from the previous
5478 access (if there is no gap DR_GROUP_GAP is 1). We skip loads that
5479 correspond to the gaps.
5480 */
5482 {
5483 gap_count++;
5485 }
5488 {
5490 /* We assume that if VEC_STMT is not NULL, this is a case of multiple
5491 copies, and we put the new vector statement in the first available
5492 RELATED_STMT. */
5495 else
5496 {
5501 {
5504 }
5506 }
5509 /* If NEXT_STMT accesses the same DR as the previous statement,
5510 put the same TMP_DATA_REF as its vectorized statement; otherwise
5511 get the next data-ref from RESULT_CHAIN. */
5514 }
5515 }
5519 }
5522 /* vectorizable_load.
5524 Check if STMT reads a non scalar data-ref (array/pointer/structure) that
5525 can be vectorized.
5526 If VEC_STMT is also passed, vectorize the STMT: create a vectorized
5527 stmt to replace it, put it in VEC_STMT, and insert it at BSI.
5528 Return FALSE if not a vectorizable STMT, TRUE otherwise. */
5530 bool
5532 slp_tree slp_node)
5533 {
5534 tree scalar_dest;
5537 tree op;
5539 stmt_vec_info prev_stmt_info;
5546 tree new_temp;
5549 tree dummy;
5552 tree ptr_incr;
5562 tree first_stmt;
5563 tree scalar_type;
5570 /* FORNOW: SLP with multiple types is not supported. The SLP analysis verifies
5571 this, so we can safely override NCOPIES with 1 here. */
5577 /* FORNOW. This restriction should be relaxed. */
5579 {
5583 }
5591 /* Is vectorizable load? */
5611 /* FORNOW. In some cases can vectorize even if data-type not supported
5612 (e.g. - data copies). */
5614 {
5618 }
5620 /* Check if the load is a part of an interleaving chain. */
5622 {
5624 /* FORNOW */
5627 /* Check if interleaving is supported. */
5631 }
5634 {
5638 }
5643 /** Transform. **/
5646 {
5648 /* Check if the chain of loads is already vectorized. */
5650 {
5653 }
5658 /* VEC_NUM is the number of vect stmts to be created for this group. */
5660 {
5663 }
5664 else
5666 }
5667 else
5668 {
5672 }
5677 /* In case the vectorization factor (VF) is bigger than the number
5678 of elements that we can fit in a vectype (nunits), we have to generate
5679 more than one vector stmt - i.e - we need to "unroll" the
5680 vector stmt by a factor VF/nunits. In doing so, we record a pointer
5681 from one copy of the vector stmt to the next, in the field
5682 STMT_VINFO_RELATED_STMT. This is necessary in order to allow following
5683 stages to find the correct vector defs to be used when vectorizing
5684 stmts that use the defs of the current stmt. The example below illustrates
5685 the vectorization process when VF=16 and nunits=4 (i.e - we need to create
5686 4 vectorized stmts):
5688 before vectorization:
5689 RELATED_STMT VEC_STMT
5690 S1: x = memref - -
5691 S2: z = x + 1 - -
5693 step 1: vectorize stmt S1:
5694 We first create the vector stmt VS1_0, and, as usual, record a
5695 pointer to it in the STMT_VINFO_VEC_STMT of the scalar stmt S1.
5696 Next, we create the vector stmt VS1_1, and record a pointer to
5697 it in the STMT_VINFO_RELATED_STMT of the vector stmt VS1_0.
5698 Similarly, for VS1_2 and VS1_3. This is the resulting chain of
5699 stmts and pointers:
5700 RELATED_STMT VEC_STMT
5701 VS1_0: vx0 = memref0 VS1_1 -
5702 VS1_1: vx1 = memref1 VS1_2 -
5703 VS1_2: vx2 = memref2 VS1_3 -
5704 VS1_3: vx3 = memref3 - -
5705 S1: x = load - VS1_0
5706 S2: z = x + 1 - -
5708 See in documentation in vect_get_vec_def_for_stmt_copy for how the
5709 information we recorded in RELATED_STMT field is used to vectorize
5710 stmt S2. */
5712 /* In case of interleaving (non-unit strided access):
5714 S1: x2 = &base + 2
5715 S2: x0 = &base
5716 S3: x1 = &base + 1
5717 S4: x3 = &base + 3
5719 Vectorized loads are created in the order of memory accesses
5720 starting from the access of the first stmt of the chain:
5722 VS1: vx0 = &base
5723 VS2: vx1 = &base + vec_size*1
5724 VS3: vx3 = &base + vec_size*2
5725 VS4: vx4 = &base + vec_size*3
5727 Then permutation statements are generated:
5729 VS5: vx5 = VEC_EXTRACT_EVEN_EXPR < vx0, vx1 >
5730 VS6: vx6 = VEC_EXTRACT_ODD_EXPR < vx0, vx1 >
5731 ...
5733 And they are put in STMT_VINFO_VEC_STMT of the corresponding scalar stmts
5734 (the order of the data-refs in the output of vect_permute_load_chain
5735 corresponds to the order of scalar stmts in the interleaving chain - see
5736 the documentation of vect_permute_load_chain()).
5737 The generation of permutation stmts and recording them in
5738 STMT_VINFO_VEC_STMT is done in vect_transform_strided_load().
5740 In case of both multiple types and interleaving, the vector loads and
5741 permutation stmts above are created for every copy. The result vector stmts
5742 are put in STMT_VINFO_VEC_STMT for the first copy and in the corresponding
5743 STMT_VINFO_RELATED_STMT for the next copies. */
5745 /* If the data reference is aligned (dr_aligned) or potentially unaligned
5746 on a target that supports unaligned accesses (dr_unaligned_supported)
5747 we generate the following code:
5748 p = initial_addr;
5749 indx = 0;
5750 loop {
5751 p = p + indx * vectype_size;
5752 vec_dest = *(p);
5753 indx = indx + 1;
5754 }
5756 Otherwise, the data reference is potentially unaligned on a target that
5757 does not support unaligned accesses (dr_explicit_realign_optimized) -
5758 then generate the following code, in which the data in each iteration is
5759 obtained by two vector loads, one from the previous iteration, and one
5760 from the current iteration:
5761 p1 = initial_addr;
5762 msq_init = *(floor(p1))
5763 p2 = initial_addr + VS - 1;
5764 realignment_token = call target_builtin;
5765 indx = 0;
5766 loop {
5767 p2 = p2 + indx * vectype_size
5768 lsq = *(floor(p2))
5769 vec_dest = realign_load (msq, lsq, realignment_token)
5770 indx = indx + 1;
5771 msq = lsq;
5772 } */
5774 /* If the misalignment remains the same throughout the execution of the
5775 loop, we can create the init_addr and permutation mask at the loop
5776 preheader. Otherwise, it needs to be created inside the loop.
5777 This can only occur when vectorizing memory accesses in the inner-loop
5778 nested within an outer-loop that is being vectorized. */
5782 {
5785 }
5790 {
5795 {
5798 }
5799 }
5800 else
5805 {
5806 /* 1. Create the vector pointer update chain. */
5812 else
5813 dataref_ptr =
5817 {
5820 NULL_TREE);
5822 /* 2. Create the vector-load in the loop. */
5824 {
5830 {
5835 data_ref =
5838 }
5840 {
5847 dr_explicit_realign,
5865 }
5871 }
5879 /* 3. Handle explicit realignment if necessary/supported. Create in
5880 loop: vec_dest = realign_load (msq, lsq, realignment_token) */
5883 {
5889 realignment_token);
5896 {
5900 }
5901 }
5903 /* 4. Handle invariant-load. */
5905 {
5909 {
5914 /* CHECKME: bitpos depends on endianess? */
5918 vec_dest =
5927 /* FIXME: use build_constructor directly. */
5931 }
5932 else
5934 }
5936 /* Collect vector loads and later create their permutation in
5937 vect_transform_strided_load (). */
5941 /* Store vector loads in the corresponding SLP_NODE. */
5944 }
5946 /* FORNOW: SLP with multiple types is unsupported. */
5951 {
5957 }
5958 else
5959 {
5962 else
5965 }
5966 }
5972 }
5975 /* Function vectorizable_live_operation.
5977 STMT computes a value that is used outside the loop. Check if
5978 it can be supported. */
5980 bool
5984 {
5985 tree operation;
5991 tree op;
6006 /* FORNOW. CHECKME. */
6013 /* FORNOW: support only if all uses are invariant. This means
6014 that the scalar operations can remain in place, unvectorized.
6015 The original last scalar value that they compute will be used. */
6018 {
6021 {
6025 }
6029 }
6031 /* No transformation is required for the cases we currently support. */
6033 }
6036 /* Function vect_is_simple_cond.
6038 Input:
6039 LOOP - the loop that is being vectorized.
6040 COND - Condition that is checked for simple use.
6042 Returns whether a COND can be vectorized. Checks whether
6043 condition operands are supportable using vec_is_simple_use. */
6045 static bool
6047 {
6049 tree def;
6059 {
6063 }
6069 {
6073 }
6079 }
6081 /* vectorizable_condition.
6083 Check if STMT is conditional modify expression that can be vectorized.
6084 If VEC_STMT is also passed, vectorize the STMT: create a vectorized
6085 stmt using VEC_COND_EXPR to replace it, put it in VEC_STMT, and insert it
6086 at BSI.
6088 Return FALSE if not a vectorizable STMT, TRUE otherwise. */
6090 bool
6092 {
6101 tree new_temp;
6104 tree def;
6119 /* FORNOW: SLP not supported. */
6123 /* FORNOW: not yet supported. */
6125 {
6129 }
6131 /* Is vectorizable conditional operation? */
6147 /* We do not handle two different vector types for the condition
6148 and the values. */
6153 {
6158 }
6165 {
6170 }
6180 {
6183 }
6185 /* Transform */
6187 /* Handle def. */
6191 /* Handle cond expr. */
6192 vec_cond_lhs =
6194 vec_cond_rhs =
6199 /* Arguments are ready. create the new vector stmt. */
6211 }
6214 /* Function vect_transform_stmt.
6216 Create a vectorized stmt to replace STMT, and insert it at BSI. */
6218 static bool
6220 slp_tree slp_node)
6221 {
6225 tree orig_stmt_in_pattern;
6229 {
6272 {
6273 /* In case of interleaving, the whole chain is vectorized when the
6274 last store in the chain is reached. Store stmts before the last
6275 one are skipped, and there vec_stmt_info shouldn't be freed
6276 meanwhile. */
6280 }
6281 else
6304 {
6308 }
6309 }
6313 {
6316 }
6319 {
6323 {
6325 /* STMT was inserted by the vectorizer to replace a computation idiom.
6326 ORIG_STMT_IN_PATTERN is a stmt in the original sequence that
6327 computed this idiom. We need to record a pointer to VEC_STMT in
6328 the stmt_info of ORIG_STMT_IN_PATTERN. See more details in the
6329 documentation of vect_pattern_recog. */
6331 {
6334 }
6335 }
6336 }
6339 }
6342 /* This function builds ni_name = number of iterations loop executes
6343 on the loop preheader. */
6345 static tree
6347 {
6349 edge pe;
6359 {
6362 }
6365 }
6368 /* This function generates the following statements:
6370 ni_name = number of iterations loop executes
6371 ratio = ni_name / vf
6372 ratio_mult_vf_name = ratio * vf
6374 and places them at the loop preheader edge. */
6376 static void
6381 {
6383 edge pe;
6384 basic_block new_bb;
6386 tree var;
6387 tree ratio_name;
6388 tree ratio_mult_vf_name;
6392 tree log_vf;
6396 /* Generate temporary variable that contains
6397 number of iterations loop executes. */
6402 /* Create: ratio = ni >> log2(vf) */
6406 {
6414 }
6416 /* Create: ratio_mult_vf = ratio << log2 (vf). */
6421 {
6430 }
6437 }
6440 /* Function vect_update_ivs_after_vectorizer.
6442 "Advance" the induction variables of LOOP to the value they should take
6443 after the execution of LOOP. This is currently necessary because the
6444 vectorizer does not handle induction variables that are used after the
6445 loop. Such a situation occurs when the last iterations of LOOP are
6446 peeled, because:
6447 1. We introduced new uses after LOOP for IVs that were not originally used
6448 after LOOP: the IVs of LOOP are now used by an epilog loop.
6449 2. LOOP is going to be vectorized; this means that it will iterate N/VF
6450 times, whereas the loop IVs should be bumped N times.
6452 Input:
6453 - LOOP - a loop that is going to be vectorized. The last few iterations
6454 of LOOP were peeled.
6455 - NITERS - the number of iterations that LOOP executes (before it is
6456 vectorized). i.e, the number of times the ivs should be bumped.
6457 - UPDATE_E - a successor edge of LOOP->exit that is on the (only) path
6458 coming out from LOOP on which there are uses of the LOOP ivs
6459 (this is the path from LOOP->exit to epilog_loop->preheader).
6461 The new definitions of the ivs are placed in LOOP->exit.
6462 The phi args associated with the edge UPDATE_E in the bb
6463 UPDATE_E->dest are updated accordingly.
6465 Assumption 1: Like the rest of the vectorizer, this function assumes
6466 a single loop exit that has a single predecessor.
6468 Assumption 2: The phi nodes in the LOOP header and in update_bb are
6469 organized in the same order.
6471 Assumption 3: The access function of the ivs is simple enough (see
6472 vect_can_advance_ivs_p). This assumption will be relaxed in the future.
6474 Assumption 4: Exactly one of the successors of LOOP exit-bb is on a path
6475 coming out of LOOP on which the ivs of LOOP are used (this is the path
6476 that leads to the epilog loop; other paths skip the epilog loop). This
6477 path starts with the edge UPDATE_E, and its destination (denoted update_bb)
6478 needs to have its phis updated.
6479 */
6481 static void
6483 edge update_e)
6484 {
6490 /* gcc_assert (vect_can_advance_ivs_p (loop_vinfo)); */
6492 /* Make sure there exists a single-predecessor exit bb: */
6498 {
6500 tree evolution_part;
6501 tree init_expr;
6502 tree step_expr;
6504 block_stmt_iterator last_bsi;
6507 {
6510 }
6512 /* Skip virtual phi's. */
6514 {
6518 }
6520 /* Skip reduction phis. */
6522 {
6526 }
6530 evolution_part =
6534 /* FORNOW: We do not support IVs whose evolution function is a polynomial
6535 of degree >= 2 or exponential. */
6544 init_expr,
6548 else
6552 niters),
6553 step_expr),
6554 init_expr);
6565 /* Fix phi expressions in the successor bb. */
6567 }
6568 }
6570 /* Return the more conservative threshold between the
6571 min_profitable_iters returned by the cost model and the user
6572 specified threshold, if provided. */
6574 static unsigned int
6577 {
6584 /* Use the cost model only if it is more conservative than user specified
6585 threshold. */
6588 && (!min_scalar_loop_bound
6596 }
6598 /* Function vect_do_peeling_for_loop_bound
6600 Peel the last iterations of the loop represented by LOOP_VINFO.
6601 The peeled iterations form a new epilog loop. Given that the loop now
6602 iterates NITERS times, the new epilog loop iterates
6603 NITERS % VECTORIZATION_FACTOR times.
6605 The original loop will later be made to iterate
6606 NITERS / VECTORIZATION_FACTOR times (this value is placed into RATIO). */
6608 static void
6610 {
6614 edge update_e;
6615 basic_block preheader;
6626 /* Generate the following variables on the preheader of original loop:
6628 ni_name = number of iteration the original loop executes
6629 ratio = ni_name / vf
6630 ratio_mult_vf_name = ratio * vf */
6636 /* If cost model check not done during versioning and
6637 peeling for alignment. */
6641 {
6644 /* Get profitability threshold for vectorized loop. */
6648 min_profitable_iters);
6649 }
6656 #ifdef ENABLE_CHECKING
6658 #endif
6660 /* A guard that controls whether the new_loop is to be executed or skipped
6661 is placed in LOOP->exit. LOOP->exit therefore has two successors - one
6662 is the preheader of NEW_LOOP, where the IVs from LOOP are used. The other
6663 is a bb after NEW_LOOP, where these IVs are not used. Find the edge that
6664 is on the path where the LOOP IVs are used and need to be updated. */
6669 else
6672 /* Update IVs of original loop as if they were advanced
6673 by ratio_mult_vf_name steps. */
6676 /* After peeling we have to reset scalar evolution analyzer. */
6680 }
6683 /* Function vect_gen_niters_for_prolog_loop
6685 Set the number of iterations for the loop represented by LOOP_VINFO
6686 to the minimum between LOOP_NITERS (the original iteration count of the loop)
6687 and the misalignment of DR - the data reference recorded in
6688 LOOP_VINFO_UNALIGNED_DR (LOOP_VINFO). As a result, after the execution of
6689 this loop, the data reference DR will refer to an aligned location.
6691 The following computation is generated:
6693 If the misalignment of DR is known at compile time:
6694 addr_mis = int mis = DR_MISALIGNMENT (dr);
6695 Else, compute address misalignment in bytes:
6696 addr_mis = addr & (vectype_size - 1)
6698 prolog_niters = min ( LOOP_NITERS , (VF - addr_mis/elem_size)&(VF-1) )
6700 (elem_size = element type size; an element is the scalar element
6701 whose type is the inner type of the vectype)
6703 For interleaving,
6705 prolog_niters = min ( LOOP_NITERS ,
6706 (VF/group_size - addr_mis/elem_size)&(VF/group_size-1) )
6707 where group_size is the size of the interleaved group.
6709 The above formulas assume that VF == number of elements in the vector. This
6710 may not hold when there are multiple-types in the loop.
6711 In this case, for some data-references in the loop the VF does not represent
6712 the number of elements that fit in the vector. Therefore, instead of VF we
6713 use TYPE_VECTOR_SUBPARTS. */
6715 static tree
6717 {
6722 edge pe;
6723 basic_block new_bb;
6734 {
6735 /* For interleaved access element size must be multiplied by the size of
6736 the interleaved group. */
6740 }
6745 {
6753 }
6754 else
6755 {
6763 tree elem_size_log =
6767 tree byte_misalign;
6768 tree elem_misalign;
6773 /* Create: byte_misalign = addr & (vectype_size - 1) */
6774 byte_misalign =
6777 /* Create: elem_misalign = byte_misalign / element_size */
6778 elem_misalign =
6781 /* Create: (niters_type) (nelements - elem_misalign)&(nelements - 1) */
6785 }
6787 /* Create: prolog_loop_niters = min (iters, loop_niters) */
6788 /* If the loop bound is known at compile time we already verified that it is
6789 greater than vf; since the misalignment ('iters') is at most vf, there's
6790 no need to generate the MIN_EXPR in this case. */
6795 {
6798 }
6804 /* Insert stmt on loop preheader edge. */
6806 {
6809 }
6812 }
6815 /* Function vect_update_init_of_dr
6817 NITERS iterations were peeled from LOOP. DR represents a data reference
6818 in LOOP. This function updates the information recorded in DR to
6819 account for the fact that the first NITERS iterations had already been
6820 executed. Specifically, it updates the OFFSET field of DR. */
6822 static void
6824 {
6830 }
6833 /* Function vect_update_inits_of_drs
6835 NITERS iterations were peeled from the loop represented by LOOP_VINFO.
6836 This function updates the information recorded for the data references in
6837 the loop to account for the fact that the first NITERS iterations had
6838 already been executed. Specifically, it updates the initial_condition of
6839 the access_function of all the data_references in the loop. */
6841 static void
6843 {
6853 }
6856 /* Function vect_do_peeling_for_alignment
6858 Peel the first 'niters' iterations of the loop represented by LOOP_VINFO.
6859 'niters' is set to the misalignment of one of the data references in the
6860 loop, thereby forcing it to refer to an aligned location at the beginning
6861 of the execution of this loop. The data reference for which we are
6862 peeling is recorded in LOOP_VINFO_UNALIGNED_DR. */
6864 static void
6866 {
6869 tree n_iters;
6884 /* If cost model check not done during versioning. */
6887 {
6890 /* Get profitability threshold for vectorized loop. */
6894 min_profitable_iters);
6895 }
6897 /* Peel the prolog loop and iterate it niters_of_prolog_loop. */
6898 new_loop =
6904 #ifdef ENABLE_CHECKING
6906 #endif
6908 /* Update number of times loop executes. */
6913 /* Update the init conditions of the access functions of all data refs. */
6916 /* After peeling we have to reset scalar evolution analyzer. */
6920 }
6923 /* Function vect_create_cond_for_align_checks.
6925 Create a conditional expression that represents the alignment checks for
6926 all of data references (array element references) whose alignment must be
6927 checked at runtime.
6929 Input:
6930 COND_EXPR - input conditional expression. New conditions will be chained
6931 with logical AND operation.
6932 LOOP_VINFO - two fields of the loop information are used.
6933 LOOP_VINFO_PTR_MASK is the mask used to check the alignment.
6934 LOOP_VINFO_MAY_MISALIGN_STMTS contains the refs to be checked.
6936 Output:
6937 COND_EXPR_STMT_LIST - statements needed to construct the conditional
6938 expression.
6939 The returned value is the conditional expression to be used in the if
6940 statement that controls which version of the loop gets executed at runtime.
6942 The algorithm makes two assumptions:
6943 1) The number of bytes "n" in a vector is a power of 2.
6944 2) An address "a" is aligned if a%n is zero and that this
6945 test can be done as a&(n-1) == 0. For example, for 16
6946 byte vectors the test is a&0xf == 0. */
6948 static void
6952 {
6958 tree mask_cst;
6960 tree psize;
6961 tree int_ptrsize_type;
6965 tree ptrsize_zero;
6966 tree part_cond_expr;
6968 /* Check that mask is one less than a power of 2, i.e., mask is
6969 all zeros followed by all ones. */
6972 /* CHECKME: what is the best integer or unsigned type to use to hold a
6973 cast from a pointer value? */
6975 int_ptrsize_type
6978 /* Create expression (mask & (dr_1 || ... || dr_n)) where dr_i is the address
6979 of the first vector of the i'th data reference. */
6982 {
6984 tree addr_base;
6988 /* create: addr_tmp = (int)(address_of_first_vector) */
7004 /* The addresses are OR together. */
7007 {
7008 /* create: or_tmp = or_tmp | addr_tmp */
7019 }
7020 else
7027 /* create: and_tmp = or_tmp & mask */
7037 /* Make and_tmp the left operand of the conditional test against zero.
7038 if and_tmp has a nonzero bit then some address is unaligned. */
7045 else
7047 }
7049 /* Function vect_vfa_segment_size.
7051 Create an expression that computes the size of segment
7052 that will be accessed for a data reference. The functions takes into
7053 account that realignment loads may access one more vector.
7055 Input:
7056 DR: The data reference.
7057 VECT_FACTOR: vectorization factor.
7059 Return an expression whose value is the size of segment which will be
7060 accessed by DR. */
7062 static tree
7064 {
7069 {
7070 tree vector_size = TYPE_SIZE_UNIT
7075 }
7077 }
7079 /* Function vect_create_cond_for_alias_checks.
7081 Create a conditional expression that represents the run-time checks for
7082 overlapping of address ranges represented by a list of data references
7083 relations passed as input.
7085 Input:
7086 COND_EXPR - input conditional expression. New conditions will be chained
7087 with logical AND operation.
7088 LOOP_VINFO - field LOOP_VINFO_MAY_ALIAS_STMTS contains the list of ddrs
7089 to be checked.
7091 Output:
7092 COND_EXPR - conditional expression.
7093 COND_EXPR_STMT_LIST - statements needed to construct the conditional
7094 expression.
7097 The returned value is the conditional expression to be used in the if
7098 statement that controls which version of the loop gets executed at runtime.
7099 */
7101 static void
7105 {
7109 tree vect_factor =
7112 ddr_p ddr;
7114 tree part_cond_expr;
7116 /* Create expression
7117 ((store_ptr_0 + store_segment_length_0) < load_ptr_0)
7118 || (load_ptr_0 + load_segment_length_0) < store_ptr_0))
7119 &&
7120 ...
7121 &&
7122 ((store_ptr_n + store_segment_length_n) < load_ptr_n)
7123 || (load_ptr_n + load_segment_length_n) < store_ptr_n)) */
7129 {
7140 {
7143 }
7149 {
7152 }
7154 addr_base_a =
7157 addr_base_b =
7165 {
7171 }
7174 part_cond_expr =
7178 addr_base_a,
7179 segment_length_a),
7180 addr_base_b),
7183 addr_base_b,
7184 segment_length_b),
7185 addr_base_a));
7190 else
7192 }
7197 }
7199 /* Function vect_loop_versioning.
7201 If the loop has data references that may or may not be aligned or/and
7202 has data reference relations whose independence was not proven then
7203 two versions of the loop need to be generated, one which is vectorized
7204 and one which isn't. A test is then generated to control which of the
7205 loops is executed. The test checks for the alignment of all of the
7206 data references that may or may not be aligned. An additional
7207 sequence of runtime tests is generated for each pairs of DDRs whose
7208 independence was not proven. The vectorized version of loop is
7209 executed only if both alias and alignment tests are passed.
7211 The test generated to check which version of loop is executed
7212 is modified to also check for profitability as indicated by the
7213 cost model initially. */
7215 static void
7217 {
7222 basic_block condition_bb;
7223 block_stmt_iterator cond_exp_bsi;
7224 basic_block merge_bb;
7225 basic_block new_exit_bb;
7229 tree gimplify_stmt_list;
7234 /* Get profitability threshold for vectorized loop. */
7238 min_profitable_iters);
7240 cond_expr =
7255 cond_expr =
7257 cond_expr =
7259 NULL_TREE);
7267 /* Loop versioning violates an assumption we try to maintain during
7268 vectorization - that the loop exit block has a single predecessor.
7269 After versioning, the exit block of both loop versions is the same
7270 basic block (i.e. it has two predecessors). Just in order to simplify
7271 following transformations in the vectorizer, we fix this situation
7272 here by adding a new (empty) block on the exit-edge of the loop,
7273 with the proper loop-exit phis to maintain loop-closed-form. */
7283 {
7285 new_exit_bb);
7289 }
7291 /* End loop-exit-fixes after versioning. */
7295 {
7298 }
7299 }
7301 /* Remove a group of stores (for SLP or interleaving), free their
7302 stmt_vec_info. */
7304 static void
7306 {
7308 tree tmp;
7309 block_stmt_iterator next_si;
7312 {
7313 /* Free the attached stmt_vec_info and remove the stmt. */
7319 }
7320 }
7323 /* Vectorize SLP instance tree in postorder. */
7325 static bool
7327 {
7328 tree stmt;
7330 block_stmt_iterator si;
7331 stmt_vec_info stmt_info;
7345 {
7348 }
7353 {
7355 /* If IS_STORE is TRUE, the vectorization of the
7356 interleaving chain was completed - free all the stores in
7357 the chain. */
7359 else
7360 /* FORNOW: SLP originates only from strided stores. */
7364 }
7366 /* FORNOW: SLP originates only from strided stores. */
7368 }
7371 static bool
7373 {
7376 slp_instance instance;
7383 {
7385 /* For each SLP instance calculate number of vector stmts to be created
7386 for the scalar stmts in each node of the SLP tree. Number of vector
7387 elements in one vector iteration is the number of scalar elements in
7388 one scalar iteration (GROUP_SIZE) multiplied by VF divided by vector
7389 size. */
7392 /* Schedule the tree of INSTANCE. */
7394 vec_stmts_size);
7399 }
7402 }
7404 /* Function vect_transform_loop.
7406 The analysis phase has determined that the loop is vectorizable.
7407 Vectorize the loop - created vectorized stmts to replace the scalar
7408 stmts in the loop, and update the loop exit condition. */
7410 void
7412 {
7416 block_stmt_iterator si;
7431 /* CHECKME: we wouldn't need this if we called update_ssa once
7432 for all loops. */
7435 /* Peel the loop if there are data refs with unknown alignment.
7436 Only one data ref with unknown store is allowed. */
7441 /* If the loop has a symbolic number of iterations 'n' (i.e. it's not a
7442 compile time constant), or it is a constant that doesn't divide by the
7443 vectorization factor, then an epilog loop needs to be created.
7444 We therefore duplicate the loop: the original loop will be vectorized,
7445 and will compute the first (n/VF) iterations. The second copy of the loop
7446 will remain scalar and will compute the remaining (n%VF) iterations.
7447 (VF is the vectorization factor). */
7453 else
7457 /* 1) Make sure the loop header has exactly two entries
7458 2) Make sure we have a preheader basic block. */
7464 /* FORNOW: the vectorizer supports only loops which body consist
7465 of one basic block (header + empty latch). When the vectorizer will
7466 support more involved loop forms, the order by which the BBs are
7467 traversed need to be reconsidered. */
7470 {
7472 stmt_vec_info stmt_info;
7473 tree phi;
7476 {
7478 {
7481 }
7496 {
7500 }
7501 }
7504 {
7509 {
7512 }
7516 /* vector stmts created in the outer-loop during vectorization of
7517 stmts in an inner-loop may not have a stmt_info, and do not
7518 need to be vectorized. */
7520 {
7523 }
7527 {
7530 }
7533 nunits =
7538 /* For SLP VF is set according to unrolling factor, and not to
7539 vector size, hence for SLP this print is not valid. */
7542 /* SLP. Schedule all the SLP instances when the first SLP stmt is
7543 reached. */
7545 {
7547 {
7555 /* IS_STORE is true if STMT is a store. Stores cannot be of
7556 hybrid SLP type. They are removed in
7557 vect_schedule_slp_instance and their vinfo is destroyed. */
7559 {
7562 }
7563 }
7565 /* Hybrid SLP stmts must be vectorized in addition to SLP. */
7567 {
7570 }
7571 }
7573 /* -------- vectorize statement ------------ */
7580 {
7582 {
7583 /* Interleaving. If IS_STORE is TRUE, the vectorization of the
7584 interleaving chain was completed - free all the stores in
7585 the chain. */
7589 }
7590 else
7591 {
7592 /* Free the attached stmt_vec_info and remove the stmt. */
7596 }
7597 }
7606 /* The memory tags and pointers in vectorized statements need to
7607 have their SSA forms updated. FIXME, why can't this be delayed
7608 until all the loops have been transformed? */
7615 }