| blob | 846d52bf90cac03642898a1b4ce1ef1f8dd02c98 |
1 /* Transformation Utilities for Loop Vectorization.
2 Copyright (C) 2003,2004,2005,2006 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 2, 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 COPYING. If not, write to the Free
19 Software Foundation, 51 Franklin Street, Fifth Floor, Boston, MA
20 02110-1301, USA. */
49 /* Utility functions for the code transformation. */
52 static tree vect_create_data_ref_ptr
59 static void vect_finish_stmt_generation
66 /* Utility function dealing with loop peeling (not peeling itself). */
67 static void vect_generate_tmps_on_preheader
77 /* Function vect_get_new_vect_var.
79 Returns a name for a new variable. The current naming scheme appends the
80 prefix "vect_" or "vect_p" (depending on the value of VAR_KIND) to
81 the name of vectorizer generated variables, and appends that to NAME if
82 provided. */
84 static tree
86 {
88 tree new_vect_var;
91 {
103 }
107 else
110 /* Mark vector typed variable as a gimple register variable. */
115 }
118 /* Function vect_create_addr_base_for_vector_ref.
120 Create an expression that computes the address of the first memory location
121 that will be accessed for a data reference.
123 Input:
124 STMT: The statement containing the data reference.
125 NEW_STMT_LIST: Must be initialized to NULL_TREE or a statement list.
126 OFFSET: Optional. If supplied, it is be added to the initial address.
128 Output:
129 1. Return an SSA_NAME whose value is the address of the memory location of
130 the first vector of the data reference.
131 2. If new_stmt_list is not NULL_TREE after return then the caller must insert
132 these statement(s) which define the returned SSA_NAME.
134 FORNOW: We are only handling array accesses with step 1. */
136 static tree
139 tree offset)
140 {
145 tree vec_stmt;
152 /* Create base_offset */
160 {
162 tree step;
164 /* For interleaved access step we divide STEP by the size of the
165 interleaving group. */
170 else
179 }
181 /* base + base_offset */
183 base_offset);
187 /* addr_expr = addr_base */
199 {
202 }
204 }
207 /* Function vect_create_data_ref_ptr.
209 Create a new pointer to vector type (vp), that points to the first location
210 accessed in the loop by STMT, along with the def-use update chain to
211 appropriately advance the pointer through the loop iterations. Also set
212 aliasing information for the pointer. This vector pointer is used by the
213 callers to this function to create a memory reference expression for vector
214 load/store access.
216 Input:
217 1. STMT: a stmt that references memory. Expected to be of the form
218 GIMPLE_MODIFY_STMT <name, data-ref> or
219 GIMPLE_MODIFY_STMT <data-ref, name>.
220 2. BSI: block_stmt_iterator where new stmts can be added.
221 3. OFFSET (optional): an offset to be added to the initial address accessed
222 by the data-ref in STMT.
223 4. ONLY_INIT: indicate if vp is to be updated in the loop, or remain
224 pointing to the initial address.
225 5. TYPE: if not NULL indicates the required type of the data-ref
227 Output:
228 1. Declare a new ptr to vector_type, and have it point to the base of the
229 data reference (initial addressed accessed by the data reference).
230 For example, for vector of type V8HI, the following code is generated:
232 v8hi *vp;
233 vp = (v8hi *)initial_address;
235 if OFFSET is not supplied:
236 initial_address = &a[init];
237 if OFFSET is supplied:
238 initial_address = &a[init + OFFSET];
240 Return the initial_address in INITIAL_ADDRESS.
242 2. If ONLY_INIT is true, just return the initial pointer. Otherwise, also
243 update the pointer in each iteration of the loop.
245 Return the increment stmt that updates the pointer in PTR_INCR.
247 3. Return the pointer. */
249 static tree
254 {
255 tree base_name;
260 tree vect_ptr_type;
261 tree vect_ptr;
262 tree tag;
263 tree new_temp;
264 tree vec_stmt;
267 basic_block new_bb;
268 tree vect_ptr_init;
274 {
287 }
289 /** (1) Create the new vector-pointer variable: **/
292 else
298 /** (2) Add aliasing information to the new vector-pointer:
299 (The points-to info (DR_PTR_INFO) may be defined later.) **/
304 /* If tag is a variable (and NOT_A_TAG) than a new symbol memory
305 tag must be created with tag added to its may alias list. */
308 else
313 /** (3) Calculate the initial address the vector-pointer, and set
314 the vector-pointer to point to it before the loop: **/
316 /* Create: (&(base[init_val+offset]) in the loop preheader. */
318 offset);
324 /* Create: p = (vectype *) initial_base */
333 /** (4) Handle the updating of the vector-pointer inside the loop: **/
336 {
337 /* Copy the points-to information if it exists. */
341 }
342 else
343 {
344 block_stmt_iterator incr_bsi;
347 tree incr;
358 /* Copy the points-to information if it exists. */
360 {
363 }
370 }
371 }
374 /* Function bump_vector_ptr
376 Increment a pointer (to a vector type) by vector-size. Connect the new
377 increment stmt to the existing def-use update-chain of the pointer.
379 The pointer def-use update-chain before this function:
380 DATAREF_PTR = phi (p_0, p_2)
381 ....
382 PTR_INCR: p_2 = DATAREF_PTR + step
384 The pointer def-use update-chain after this function:
385 DATAREF_PTR = phi (p_0, p_2)
386 ....
387 NEW_DATAREF_PTR = DATAREF_PTR + vector_size
388 ....
389 PTR_INCR: p_2 = NEW_DATAREF_PTR + step
391 Input:
392 DATAREF_PTR - ssa_name of a pointer (to vector type) that is being updated
393 in the loop.
394 PTR_INCR - the stmt that updates the pointer in each iteration of the loop.
395 The increment amount across iterations is also expected to be
396 vector_size.
397 BSI - location where the new update stmt is to be placed.
398 STMT - the original scalar memory-access stmt that is being vectorized.
400 Output: Return NEW_DATAREF_PTR as illustrated above.
402 */
404 static tree
406 tree stmt)
407 {
414 tree incr_stmt;
415 ssa_op_iter iter;
416 use_operand_p use_p;
417 tree new_dataref_ptr;
425 /* Update the vector-pointer's cross-iteration increment. */
427 {
432 else
434 }
436 /* Copy the points-to information if it exists. */
442 }
445 /* Function vect_create_destination_var.
447 Create a new temporary of type VECTYPE. */
449 static tree
451 {
452 tree vec_dest;
454 tree type;
469 }
472 /* Function vect_init_vector.
474 Insert a new stmt (INIT_STMT) that initializes a new vector variable with
475 the vector elements of VECTOR_VAR. Return the DEF of INIT_STMT. It will be
476 used in the vectorization of STMT. */
478 static tree
480 {
484 tree new_var;
485 tree init_stmt;
486 tree vec_oprnd;
487 edge pe;
488 tree new_temp;
489 basic_block new_bb;
503 {
506 }
510 }
513 /* Function vect_get_vec_def_for_operand.
515 OP is an operand in STMT. This function returns a (vector) def that will be
516 used in the vectorized stmt for STMT.
518 In the case that OP is an SSA_NAME which is defined in the loop, then
519 STMT_VINFO_VEC_STMT of the defining stmt holds the relevant def.
521 In case OP is an invariant or constant, a new stmt that creates a vector def
522 needs to be introduced. */
524 static tree
526 {
527 tree vec_oprnd;
528 tree vec_stmt;
529 tree def_stmt;
536 tree vec_inv;
537 tree vec_cst;
539 tree def;
543 tree vector_type;
546 {
549 }
554 {
556 {
559 }
561 {
564 }
565 }
568 {
569 /* Case 1: operand is a constant. */
571 {
575 /* Create 'vect_cst_ = {cst,cst,...,cst}' */
580 {
582 }
587 }
589 /* Case 2: operand is defined outside the loop - loop invariant. */
591 {
595 /* Create 'vec_inv = {inv,inv,..,inv}' */
600 {
602 }
604 /* FIXME: use build_constructor directly. */
608 }
610 /* Case 3: operand is defined inside the loop. */
612 {
616 /* Get the def from the vectorized stmt. */
622 }
624 /* Case 4: operand is defined by a loop header phi - reduction */
626 {
629 /* Get the def before the loop */
632 }
634 /* Case 5: operand is defined by loop-header phi - induction. */
636 {
640 }
644 }
645 }
648 /* Function vect_get_vec_def_for_stmt_copy
650 Return a vector-def for an operand. This function is used when the
651 vectorized stmt to be created (by the caller to this function) is a "copy"
652 created in case the vectorized result cannot fit in one vector, and several
653 copies of the vector-stmt are required. In this case the vector-def is
654 retrieved from the vector stmt recorded in the STMT_VINFO_RELATED_STMT field
655 of the stmt that defines VEC_OPRND.
656 DT is the type of the vector def VEC_OPRND.
658 Context:
659 In case the vectorization factor (VF) is bigger than the number
660 of elements that can fit in a vectype (nunits), we have to generate
661 more than one vector stmt to vectorize the scalar stmt. This situation
662 arises when there are multiple data-types operated upon in the loop; the
663 smallest data-type determines the VF, and as a result, when vectorizing
664 stmts operating on wider types we need to create 'VF/nunits' "copies" of the
665 vector stmt (each computing a vector of 'nunits' results, and together
666 computing 'VF' results in each iteration). This function is called when
667 vectorizing such a stmt (e.g. vectorizing S2 in the illustration below, in
668 which VF=16 and nunits=4, so the number of copies required is 4):
670 scalar stmt: vectorized into: STMT_VINFO_RELATED_STMT
672 S1: x = load VS1.0: vx.0 = memref0 VS1.1
673 VS1.1: vx.1 = memref1 VS1.2
674 VS1.2: vx.2 = memref2 VS1.3
675 VS1.3: vx.3 = memref3
677 S2: z = x + ... VSnew.0: vz0 = vx.0 + ... VSnew.1
678 VSnew.1: vz1 = vx.1 + ... VSnew.2
679 VSnew.2: vz2 = vx.2 + ... VSnew.3
680 VSnew.3: vz3 = vx.3 + ...
682 The vectorization of S1 is explained in vectorizable_load.
683 The vectorization of S2:
684 To create the first vector-stmt out of the 4 copies - VSnew.0 -
685 the function 'vect_get_vec_def_for_operand' is called to
686 get the relevant vector-def for each operand of S2. For operand x it
687 returns the vector-def 'vx.0'.
689 To create the remaining copies of the vector-stmt (VSnew.j), this
690 function is called to get the relevant vector-def for each operand. It is
691 obtained from the respective VS1.j stmt, which is recorded in the
692 STMT_VINFO_RELATED_STMT field of the stmt that defines VEC_OPRND.
694 For example, to obtain the vector-def 'vx.1' in order to create the
695 vector stmt 'VSnew.1', this function is called with VEC_OPRND='vx.0'.
696 Given 'vx0' we obtain the stmt that defines it ('VS1.0'); from the
697 STMT_VINFO_RELATED_STMT field of 'VS1.0' we obtain the next copy - 'VS1.1',
698 and return its def ('vx.1').
699 Overall, to create the above sequence this function will be called 3 times:
700 vx.1 = vect_get_vec_def_for_stmt_copy (dt, vx.0);
701 vx.2 = vect_get_vec_def_for_stmt_copy (dt, vx.1);
702 vx.3 = vect_get_vec_def_for_stmt_copy (dt, vx.2); */
704 static tree
706 {
707 tree vec_stmt_for_operand;
708 stmt_vec_info def_stmt_info;
711 {
714 }
724 }
727 /* Function vect_finish_stmt_generation.
729 Insert a new stmt. */
731 static void
734 {
743 {
746 }
748 /* Make sure bsi points to the stmt that is being vectorized. */
751 #ifdef USE_MAPPED_LOCATION
753 #else
755 #endif
756 }
759 #define ADJUST_IN_EPILOG 1
761 /* Function get_initial_def_for_reduction
763 Input:
764 STMT - a stmt that performs a reduction operation in the loop.
765 INIT_VAL - the initial value of the reduction variable
767 Output:
768 SCALAR_DEF - a tree that holds a value to be added to the final result
769 of the reduction (used for "ADJUST_IN_EPILOG" - see below).
770 Return a vector variable, initialized according to the operation that STMT
771 performs. This vector will be used as the initial value of the
772 vector of partial results.
774 Option1 ("ADJUST_IN_EPILOG"): Initialize the vector as follows:
775 add: [0,0,...,0,0]
776 mult: [1,1,...,1,1]
777 min/max: [init_val,init_val,..,init_val,init_val]
778 bit and/or: [init_val,init_val,..,init_val,init_val]
779 and when necessary (e.g. add/mult case) let the caller know
780 that it needs to adjust the result by init_val.
782 Option2: Initialize the vector as follows:
783 add: [0,0,...,0,init_val]
784 mult: [1,1,...,1,init_val]
785 min/max: [init_val,init_val,...,init_val]
786 bit and/or: [init_val,init_val,...,init_val]
787 and no adjustments are needed.
789 For example, for the following code:
791 s = init_val;
792 for (i=0;i<n;i++)
793 s = s + a[i];
795 STMT is 's = s + a[i]', and the reduction variable is 's'.
796 For a vector of 4 units, we want to return either [0,0,0,init_val],
797 or [0,0,0,0] and let the caller know that it needs to adjust
798 the result at the end by 'init_val'.
800 FORNOW: We use the "ADJUST_IN_EPILOG" scheme.
801 TODO: Use some cost-model to estimate which scheme is more profitable.
802 */
804 static tree
806 {
813 tree def;
817 tree vector_type;
822 {
828 else
831 #ifdef ADJUST_IN_EPILOG
832 /* All the 'nunits' elements are set to 0. The final result will be
833 adjusted by 'init_val' at the loop epilog. */
836 #else
837 /* 'nunits - 1' elements are set to 0; The last element is set to
838 'init_val'. No further adjustments at the epilog are needed. */
841 #endif
853 }
859 {
860 /* Set the last element of the vector. */
863 }
869 else
874 else
878 }
881 /* Function vect_create_epilog_for_reduction
883 Create code at the loop-epilog to finalize the result of a reduction
884 computation.
886 VECT_DEF is a vector of partial results.
887 REDUC_CODE is the tree-code for the epilog reduction.
888 STMT is the scalar reduction stmt that is being vectorized.
889 REDUCTION_PHI is the phi-node that carries the reduction computation.
891 This function:
892 1. Creates the reduction def-use cycle: sets the the arguments for
893 REDUCTION_PHI:
894 The loop-entry argument is the vectorized initial-value of the reduction.
895 The loop-latch argument is VECT_DEF - the vector of partial sums.
896 2. "Reduces" the vector of partial results VECT_DEF into a single result,
897 by applying the operation specified by REDUC_CODE if available, or by
898 other means (whole-vector shifts or a scalar loop).
899 The function also creates a new phi node at the loop exit to preserve
900 loop-closed form, as illustrated below.
902 The flow at the entry to this function:
904 loop:
905 vec_def = phi <null, null> # REDUCTION_PHI
906 VECT_DEF = vector_stmt # vectorized form of STMT
907 s_loop = scalar_stmt # (scalar) STMT
908 loop_exit:
909 s_out0 = phi <s_loop> # (scalar) EXIT_PHI
910 use <s_out0>
911 use <s_out0>
913 The above is transformed by this function into:
915 loop:
916 vec_def = phi <vec_init, VECT_DEF> # REDUCTION_PHI
917 VECT_DEF = vector_stmt # vectorized form of STMT
918 s_loop = scalar_stmt # (scalar) STMT
919 loop_exit:
920 s_out0 = phi <s_loop> # (scalar) EXIT_PHI
921 v_out1 = phi <VECT_DEF> # NEW_EXIT_PHI
922 v_out2 = reduce <v_out1>
923 s_out3 = extract_field <v_out2, 0>
924 s_out4 = adjust_result <s_out3>
925 use <s_out4>
926 use <s_out4>
927 */
929 static void
932 {
934 tree vectype;
938 basic_block exit_bb;
939 tree scalar_dest;
940 tree scalar_type;
941 tree new_phi;
942 block_stmt_iterator exit_bsi;
943 tree vec_dest;
944 tree new_temp;
945 tree new_name;
946 tree epilog_stmt;
950 tree scalar_initial_def;
951 tree vec_initial_def;
952 tree orig_name;
953 imm_use_iterator imm_iter;
954 use_operand_p use_p;
956 tree reduction_op;
957 tree orig_stmt;
958 tree use_stmt;
967 /*** 1. Create the reduction def-use cycle ***/
969 /* 1.1 set the loop-entry arg of the reduction-phi: */
970 /* For the case of reduction, vect_get_vec_def_for_operand returns
971 the scalar def before the loop, that defines the initial value
972 of the reduction variable. */
977 /* 1.2 set the loop-latch arg for the reduction-phi: */
981 {
986 }
989 /*** 2. Create epilog code
990 The reduction epilog code operates across the elements of the vector
991 of partial results computed by the vectorized loop.
992 The reduction epilog code consists of:
993 step 1: compute the scalar result in a vector (v_out2)
994 step 2: extract the scalar result (s_out3) from the vector (v_out2)
995 step 3: adjust the scalar result (s_out3) if needed.
997 Step 1 can be accomplished using one the following three schemes:
998 (scheme 1) using reduc_code, if available.
999 (scheme 2) using whole-vector shifts, if available.
1000 (scheme 3) using a scalar loop. In this case steps 1+2 above are
1001 combined.
1003 The overall epilog code looks like this:
1005 s_out0 = phi <s_loop> # original EXIT_PHI
1006 v_out1 = phi <VECT_DEF> # NEW_EXIT_PHI
1007 v_out2 = reduce <v_out1> # step 1
1008 s_out3 = extract_field <v_out2, 0> # step 2
1009 s_out4 = adjust_result <s_out3> # step 3
1011 (step 3 is optional, and step2 1 and 2 may be combined).
1012 Lastly, the uses of s_out0 are replaced by s_out4.
1014 ***/
1016 /* 2.1 Create new loop-exit-phi to preserve loop-closed form:
1017 v_out1 = phi <v_loop> */
1024 /* 2.2 Get the relevant tree-code to use in the epilog for schemes 2,3
1025 (i.e. when reduc_code is not available) and in the final adjustment code
1026 (if needed). Also get the original scalar reduction variable as
1027 defined in the loop. In case STMT is a "pattern-stmt" (i.e. - it
1028 represents a reduction pattern), the tree-code and scalar-def are
1029 taken from the original stmt that the pattern-stmt (STMT) replaces.
1030 Otherwise (it is a regular reduction) - the tree-code and scalar-def
1031 are taken from STMT. */
1035 {
1036 /* Regular reduction */
1038 }
1039 else
1040 {
1041 /* Reduction pattern */
1045 }
1053 /* 2.3 Create the reduction code, using one of the three schemes described
1054 above. */
1057 {
1058 /*** Case 1: Create:
1059 v_out2 = reduc_expr <v_out1> */
1072 }
1073 else
1074 {
1080 tree vec_temp;
1084 else
1087 /* Regardless of whether we have a whole vector shift, if we're
1088 emulating the operation via tree-vect-generic, we don't want
1089 to use it. Only the first round of the reduction is likely
1090 to still be profitable via emulation. */
1091 /* ??? It might be better to emit a reduction tree code here, so that
1092 tree-vect-generic can expand the first round via bit tricks. */
1095 else
1096 {
1100 }
1103 {
1104 /*** Case 2: Create:
1105 for (offset = VS/2; offset >= element_size; offset/=2)
1106 {
1107 Create: va' = vec_shift <va, offset>
1108 Create: va = vop <va, va'>
1109 } */
1120 {
1124 vec_dest,
1132 vec_dest,
1138 }
1141 }
1142 else
1143 {
1144 tree rhs;
1146 /*** Case 3: Create:
1147 s = extract_field <v_out2, 0>
1148 for (offset = element_size;
1149 offset < vector_size;
1150 offset += element_size;)
1151 {
1152 Create: s' = extract_field <v_out2, offset>
1153 Create: s = op <s, s'>
1154 } */
1162 bitsize_zero_node);
1173 {
1176 bitpos);
1186 new_scalar_dest,
1191 }
1194 }
1195 }
1197 /* 2.4 Extract the final scalar result. Create:
1198 s_out3 = extract_field <v_out2, bitpos> */
1201 {
1202 tree rhs;
1211 else
1221 }
1223 /* 2.4 Adjust the final result by the initial value of the reduction
1224 variable. (When such adjustment is not needed, then
1225 'scalar_initial_def' is zero).
1227 Create:
1228 s_out4 = scalar_expr <s_out3, scalar_initial_def> */
1231 {
1233 new_scalar_dest,
1238 }
1240 /* 2.6 Replace uses of s_out0 with uses of s_out3 */
1242 /* Find the loop-closed-use at the loop exit of the original scalar result.
1243 (The reduction result is expected to have two immediate uses - one at the
1244 latch block, and one at the loop exit). */
1247 {
1249 {
1252 }
1253 }
1254 /* We expect to have found an exit_phi because of loop-closed-ssa form. */
1256 /* Replace the uses: */
1261 }
1264 /* Function vectorizable_reduction.
1266 Check if STMT performs a reduction operation that can be vectorized.
1267 If VEC_STMT is also passed, vectorize the STMT: create a vectorized
1268 stmt to replace it, put it in VEC_STMT, and insert it at BSI.
1269 Return FALSE if not a vectorizable STMT, TRUE otherwise.
1271 This function also handles reduction idioms (patterns) that have been
1272 recognized in advance during vect_pattern_recog. In this case, STMT may be
1273 of this form:
1274 X = pattern_expr (arg0, arg1, ..., X)
1275 and it's STMT_VINFO_RELATED_STMT points to the last stmt in the original
1276 sequence that had been detected and replaced by the pattern-stmt (STMT).
1278 In some cases of reduction patterns, the type of the reduction variable X is
1279 different than the type of the other arguments of STMT.
1280 In such cases, the vectype that is used when transforming STMT into a vector
1281 stmt is different than the vectype that is used to determine the
1282 vectorization factor, because it consists of a different number of elements
1283 than the actual number of elements that are being operated upon in parallel.
1285 For example, consider an accumulation of shorts into an int accumulator.
1286 On some targets it's possible to vectorize this pattern operating on 8
1287 shorts at a time (hence, the vectype for purposes of determining the
1288 vectorization factor should be V8HI); on the other hand, the vectype that
1289 is used to create the vector form is actually V4SI (the type of the result).
1291 Upon entry to this function, STMT_VINFO_VECTYPE records the vectype that
1292 indicates what is the actual level of parallelism (V8HI in the example), so
1293 that the right vectorization factor would be derived. This vectype
1294 corresponds to the type of arguments to the reduction stmt, and should *NOT*
1295 be used to create the vectorized stmt. The right vectype for the vectorized
1296 stmt is obtained from the type of the result X:
1297 get_vectype_for_scalar_type (TREE_TYPE (X))
1299 This means that, contrary to "regular" reductions (or "regular" stmts in
1300 general), the following equation:
1301 STMT_VINFO_VECTYPE == get_vectype_for_scalar_type (TREE_TYPE (X))
1302 does *NOT* necessarily hold for reduction patterns. */
1304 bool
1306 {
1307 tree vec_dest;
1308 tree scalar_dest;
1309 tree op;
1315 tree operation;
1323 tree new_phi;
1324 tree scalar_type;
1326 tree orig_stmt;
1327 stmt_vec_info orig_stmt_info;
1332 stmt_vec_info prev_stmt_info;
1333 tree reduc_def;
1339 /* 1. Is vectorizable reduction? */
1341 /* Not supportable if the reduction variable is used in the loop. */
1348 /* Make sure it was already recognized as a reduction computation. */
1352 /* 2. Has this been recognized as a reduction pattern?
1354 Check if STMT represents a pattern that has been recognized
1355 in earlier analysis stages. For stmts that represent a pattern,
1356 the STMT_VINFO_RELATED_STMT field records the last stmt in
1357 the original sequence that constitutes the pattern. */
1361 {
1366 }
1368 /* 3. Check the operands of the operation. The first operands are defined
1369 inside the loop body. The last operand is the reduction variable,
1370 which is defined by the loop-header-phi. */
1382 /* All uses but the last are expected to be defined in the loop.
1383 The last use is the reduction variable. */
1385 {
1391 }
1400 else
1406 /* 4. Supportable by target? */
1408 /* 4.1. check support for the operation in the loop */
1411 {
1415 }
1418 {
1427 }
1429 /* Worthwhile without SIMD support? */
1433 {
1437 }
1439 /* 4.2. Check support for the epilog operation.
1441 If STMT represents a reduction pattern, then the type of the
1442 reduction variable may be different than the type of the rest
1443 of the arguments. For example, consider the case of accumulation
1444 of shorts into an int accumulator; The original code:
1445 S1: int_a = (int) short_a;
1446 orig_stmt-> S2: int_acc = plus <int_a ,int_acc>;
1448 was replaced with:
1449 STMT: int_acc = widen_sum <short_a, int_acc>
1451 This means that:
1452 1. The tree-code that is used to create the vector operation in the
1453 epilog code (that reduces the partial results) is not the
1454 tree-code of STMT, but is rather the tree-code of the original
1455 stmt from the pattern that STMT is replacing. I.e, in the example
1456 above we want to use 'widen_sum' in the loop, but 'plus' in the
1457 epilog.
1458 2. The type (mode) we use to check available target support
1459 for the vector operation to be created in the *epilog*, is
1460 determined by the type of the reduction variable (in the example
1461 above we'd check this: plus_optab[vect_int_mode]).
1462 However the type (mode) we use to check available target support
1463 for the vector operation to be created *inside the loop*, is
1464 determined by the type of the other arguments to STMT (in the
1465 example we'd check this: widen_sum_optab[vect_short_mode]).
1467 This is contrary to "regular" reductions, in which the types of all
1468 the arguments are the same as the type of the reduction variable.
1469 For "regular" reductions we can therefore use the same vector type
1470 (and also the same tree-code) when generating the epilog code and
1471 when generating the code inside the loop. */
1474 {
1475 /* This is a reduction pattern: get the vectype from the type of the
1476 reduction variable, and get the tree-code from orig_stmt. */
1480 }
1481 else
1482 {
1483 /* Regular reduction: use the same vectype and tree-code as used for
1484 the vector code inside the loop can be used for the epilog code. */
1486 }
1492 {
1496 }
1498 {
1502 }
1505 {
1508 }
1510 /** Transform. **/
1515 /* Create the destination vector */
1518 /* Create the reduction-phi that defines the reduction-operand. */
1521 /* In case the vectorization factor (VF) is bigger than the number
1522 of elements that we can fit in a vectype (nunits), we have to generate
1523 more than one vector stmt - i.e - we need to "unroll" the
1524 vector stmt by a factor VF/nunits. For more details see documentation
1525 in vectorizable_operation. */
1529 {
1530 /* Handle uses. */
1532 {
1536 {
1539 }
1541 /* Get the vector def for the reduction variable from the phi node */
1543 }
1544 else
1545 {
1551 /* Get the vector def for the reduction variable from the vectorized
1552 reduction operation generated in the previous iteration (j-1) */
1554 }
1556 /* Arguments are ready. create the new vector stmt. */
1560 else
1562 reduc_def);
1570 else
1573 }
1575 /* Finalize the reduction-phi (set it's arguments) and create the
1576 epilog reduction code. */
1579 }
1581 /* Checks if CALL can be vectorized in type VECTYPE. Returns
1582 true if the target has a vectorized version of the function,
1583 or false if the function cannot be vectorized. */
1585 bool
1587 {
1590 /* We only handle functions that do not read or clobber memory -- i.e.
1591 const or novops ones. */
1604 }
1606 /* Returns an expression that performs a call to vectorized version
1607 of FNDECL in type VECTYPE, with the arguments given by ARGS.
1608 If extra statements need to be generated, they are inserted
1609 before BSI. */
1611 static tree
1614 {
1615 tree vfndecl;
1618 /* The target specific builtin should be available. */
1623 }
1625 /* Function vectorizable_call.
1627 Check if STMT performs a function call that can be vectorized.
1628 If VEC_STMT is also passed, vectorize the STMT: create a vectorized
1629 stmt to replace it, put it in VEC_STMT, and insert it at BSI.
1630 Return FALSE if not a vectorizable STMT, TRUE otherwise. */
1632 bool
1634 {
1635 tree vec_dest;
1636 tree scalar_dest;
1637 tree operation;
1646 /* Is STMT a vectorizable call? */
1657 /* For now, we only vectorize functions if a target specific builtin
1658 is available. TODO -- in some cases, it might be profitable to
1659 insert the calls for pieces of the vector, in order to be able
1660 to vectorize other operations in the loop. */
1662 {
1667 }
1671 {
1675 {
1679 }
1680 }
1683 {
1686 }
1688 /** Transform. **/
1693 /* Handle def. */
1697 /* Handle uses. */
1701 {
1707 }
1717 /* The call in STMT might prevent it from being removed in dce. We however
1718 cannot remove it here, due to the way the ssa name it defines is mapped
1719 to the new definition. So just replace rhs of the statement with something
1720 harmless. */
1725 }
1728 /* Function vectorizable_assignment.
1730 Check if STMT performs an assignment (copy) that can be vectorized.
1731 If VEC_STMT is also passed, vectorize the STMT: create a vectorized
1732 stmt to replace it, put it in VEC_STMT, and insert it at BSI.
1733 Return FALSE if not a vectorizable STMT, TRUE otherwise. */
1735 bool
1737 {
1738 tree vec_dest;
1739 tree scalar_dest;
1740 tree op;
1741 tree vec_oprnd;
1745 tree new_temp;
1755 /* Is vectorizable assignment? */
1770 {
1774 }
1777 {
1780 }
1782 /** Transform. **/
1786 /* Handle def. */
1789 /* Handle use. */
1793 /* Arguments are ready. create the new vector stmt. */
1800 }
1803 /* Function vect_min_worthwhile_factor.
1805 For a loop where we could vectorize the operation indicated by CODE,
1806 return the minimum vectorization factor that makes it worthwhile
1807 to use generic vectors. */
1808 static int
1810 {
1812 {
1826 }
1827 }
1830 /* Function vectorizable_operation.
1832 Check if STMT performs a binary or unary operation that can be vectorized.
1833 If VEC_STMT is also passed, vectorize the STMT: create a vectorized
1834 stmt to replace it, put it in VEC_STMT, and insert it at BSI.
1835 Return FALSE if not a vectorizable STMT, TRUE otherwise. */
1837 bool
1839 {
1840 tree vec_dest;
1841 tree scalar_dest;
1842 tree operation;
1850 tree new_temp;
1852 optab optab;
1857 tree new_stmt;
1858 stmt_vec_info prev_stmt_info;
1861 tree vectype_out;
1867 /* Is STMT a vectorizable binary/unary operation? */
1874 {
1875 /* FORNOW: not yet supported. */
1879 }
1897 /* Support only unary or binary operations. */
1900 {
1904 }
1908 {
1912 }
1915 {
1918 {
1922 }
1923 }
1925 /* Supportable by target? */
1927 {
1931 }
1935 {
1944 }
1946 /* Worthwhile without SIMD support? */
1950 {
1954 }
1957 {
1958 /* FORNOW: not yet supported. */
1962 /* Invariant argument is needed for a vector shift
1963 by a scalar shift operand. */
1968 {
1972 }
1973 }
1976 {
1979 }
1981 /** Transform. **/
1986 /* Handle def. */
1989 /* In case the vectorization factor (VF) is bigger than the number
1990 of elements that we can fit in a vectype (nunits), we have to generate
1991 more than one vector stmt - i.e - we need to "unroll" the
1992 vector stmt by a factor VF/nunits. In doing so, we record a pointer
1993 from one copy of the vector stmt to the next, in the field
1994 STMT_VINFO_RELATED_STMT. This is necessary in order to allow following
1995 stages to find the correct vector defs to be used when vectorizing
1996 stmts that use the defs of the current stmt. The example below illustrates
1997 the vectorization process when VF=16 and nunits=4 (i.e - we need to create
1998 4 vectorized stmts):
2000 before vectorization:
2001 RELATED_STMT VEC_STMT
2002 S1: x = memref - -
2003 S2: z = x + 1 - -
2005 step 1: vectorize stmt S1 (done in vectorizable_load. See more details
2006 there):
2007 RELATED_STMT VEC_STMT
2008 VS1_0: vx0 = memref0 VS1_1 -
2009 VS1_1: vx1 = memref1 VS1_2 -
2010 VS1_2: vx2 = memref2 VS1_3 -
2011 VS1_3: vx3 = memref3 - -
2012 S1: x = load - VS1_0
2013 S2: z = x + 1 - -
2015 step2: vectorize stmt S2 (done here):
2016 To vectorize stmt S2 we first need to find the relevant vector
2017 def for the first operand 'x'. This is, as usual, obtained from
2018 the vector stmt recorded in the STMT_VINFO_VEC_STMT of the stmt
2019 that defines 'x' (S1). This way we find the stmt VS1_0, and the
2020 relevant vector def 'vx0'. Having found 'vx0' we can generate
2021 the vector stmt VS2_0, and as usual, record it in the
2022 STMT_VINFO_VEC_STMT of stmt S2.
2023 When creating the second copy (VS2_1), we obtain the relevant vector
2024 def from the vector stmt recorded in the STMT_VINFO_RELATED_STMT of
2025 stmt VS1_0. This way we find the stmt VS1_1 and the relevant
2026 vector def 'vx1'. Using 'vx1' we create stmt VS2_1 and record a
2027 pointer to it in the STMT_VINFO_RELATED_STMT of the vector stmt VS2_0.
2028 Similarly when creating stmts VS2_2 and VS2_3. This is the resulting
2029 chain of stmts and pointers:
2030 RELATED_STMT VEC_STMT
2031 VS1_0: vx0 = memref0 VS1_1 -
2032 VS1_1: vx1 = memref1 VS1_2 -
2033 VS1_2: vx2 = memref2 VS1_3 -
2034 VS1_3: vx3 = memref3 - -
2035 S1: x = load - VS1_0
2036 VS2_0: vz0 = vx0 + v1 VS2_1 -
2037 VS2_1: vz1 = vx1 + v1 VS2_2 -
2038 VS2_2: vz2 = vx2 + v1 VS2_3 -
2039 VS2_3: vz3 = vx3 + v1 - -
2040 S2: z = x + 1 - VS2_0 */
2044 {
2045 /* Handle uses. */
2047 {
2050 {
2052 {
2053 /* Vector shl and shr insn patterns can be defined with
2054 scalar operand 2 (shift operand). In this case, use
2055 constant or loop invariant op1 directly, without
2056 extending it to vector mode first. */
2059 {
2063 }
2064 }
2067 }
2068 }
2069 else
2070 {
2074 }
2076 /* Arguments are ready. create the new vector stmt. */
2081 else
2090 else
2093 }
2096 }
2099 /* Function vectorizable_type_demotion
2101 Check if STMT performs a binary or unary operation that involves
2102 type demotion, and if it can be vectorized.
2103 If VEC_STMT is also passed, vectorize the STMT: create a vectorized
2104 stmt to replace it, put it in VEC_STMT, and insert it at BSI.
2105 Return FALSE if not a vectorizable STMT, TRUE otherwise. */
2107 bool
2110 {
2111 tree vec_dest;
2112 tree scalar_dest;
2113 tree operation;
2114 tree op0;
2119 tree new_temp;
2122 tree new_stmt;
2123 stmt_vec_info prev_stmt_info;
2126 tree vectype_out;
2129 tree expr;
2130 tree vectype_in;
2131 tree scalar_type;
2132 optab optab;
2135 /* Is STMT a vectorizable type-demotion operation? */
2143 {
2144 /* FORNOW: not yet supported. */
2148 }
2179 /* Check the operands of the operation. */
2181 {
2185 }
2187 /* Supportable by target? */
2200 {
2203 }
2205 /** Transform. **/
2209 ncopies);
2211 /* Handle def. */
2214 /* In case the vectorization factor (VF) is bigger than the number
2215 of elements that we can fit in a vectype (nunits), we have to generate
2216 more than one vector stmt - i.e - we need to "unroll" the
2217 vector stmt by a factor VF/nunits. */
2220 {
2221 /* Handle uses. */
2223 {
2227 }
2228 else
2229 {
2232 }
2234 /* Arguments are ready. Create the new vector stmt. */
2243 else
2247 }
2251 }
2254 /* Function vect_gen_widened_results_half
2256 Create a vector stmt whose code, type, number of arguments, and result
2257 variable are CODE, VECTYPE, OP_TYPE, and VEC_DEST, and its arguments are
2258 VEC_OPRND0 and VEC_OPRND1. The new vector stmt is to be inserted at BSI.
2259 In the case that CODE is a CALL_EXPR, this means that a call to DECL
2260 needs to be created (DECL is a function-decl of a target-builtin).
2261 STMT is the original scalar stmt that we are vectorizing. */
2263 static tree
2267 tree stmt)
2268 {
2269 tree vec_params;
2270 tree expr;
2271 tree new_stmt;
2272 tree new_temp;
2273 tree sym;
2274 ssa_op_iter iter;
2276 /* Generate half of the widened result: */
2278 {
2279 /* Target specific support */
2284 }
2285 else
2286 {
2287 /* Generic support */
2291 else
2293 }
2300 {
2302 {
2306 }
2307 }
2310 }
2313 /* Function vectorizable_type_promotion
2315 Check if STMT performs a binary or unary operation that involves
2316 type promotion, and if it can be vectorized.
2317 If VEC_STMT is also passed, vectorize the STMT: create a vectorized
2318 stmt to replace it, put it in VEC_STMT, and insert it at BSI.
2319 Return FALSE if not a vectorizable STMT, TRUE otherwise. */
2321 bool
2324 {
2325 tree vec_dest;
2326 tree scalar_dest;
2327 tree operation;
2337 tree new_stmt;
2338 stmt_vec_info prev_stmt_info;
2341 tree vectype_out;
2344 tree vectype_in;
2346 /* Is STMT a vectorizable type-promotion operation? */
2354 {
2355 /* FORNOW: not yet supported. */
2359 }
2388 /* Check the operands of the operation. */
2390 {
2394 }
2398 {
2401 {
2405 }
2406 }
2408 /* Supportable by target? */
2416 {
2419 }
2421 /** Transform. **/
2425 ncopies);
2427 /* Handle def. */
2430 /* In case the vectorization factor (VF) is bigger than the number
2431 of elements that we can fit in a vectype (nunits), we have to generate
2432 more than one vector stmt - i.e - we need to "unroll" the
2433 vector stmt by a factor VF/nunits. */
2437 {
2438 /* Handle uses. */
2440 {
2444 }
2445 else
2446 {
2450 }
2452 /* Arguments are ready. Create the new vector stmt. We are creating
2453 two vector defs because the widened result does not fit in one vector.
2454 The vectorized stmt can be expressed as a call to a taregt builtin,
2455 or a using a tree-code. */
2456 /* Generate first half of the widened result: */
2461 else
2465 /* Generate second half of the widened result: */
2471 }
2475 }
2478 /* Function vect_strided_store_supported.
2480 Returns TRUE is INTERLEAVE_HIGH and INTERLEAVE_LOW operations are supported,
2481 and FALSE otherwise. */
2483 static bool
2485 {
2491 /* Check that the operation is supported. */
2493 vectype);
2495 vectype);
2497 {
2501 }
2504 == CODE_FOR_nothing
2507 {
2511 }
2513 }
2516 /* Function vect_permute_store_chain.
2518 Given a chain of interleaved stores in DR_CHAIN of LENGTH that must be
2519 a power of 2, generate interleave_high/low stmts to reorder the data
2520 correctly for the stores. Return the final references for stores in
2521 RESULT_CHAIN.
2523 E.g., LENGTH is 4 and the scalar type is short, i.e., VF is 8.
2524 The input is 4 vectors each containing 8 elements. We assign a number to each
2525 element, the input sequence is:
2527 1st vec: 0 1 2 3 4 5 6 7
2528 2nd vec: 8 9 10 11 12 13 14 15
2529 3rd vec: 16 17 18 19 20 21 22 23
2530 4th vec: 24 25 26 27 28 29 30 31
2532 The output sequence should be:
2534 1st vec: 0 8 16 24 1 9 17 25
2535 2nd vec: 2 10 18 26 3 11 19 27
2536 3rd vec: 4 12 20 28 5 13 21 30
2537 4th vec: 6 14 22 30 7 15 23 31
2539 i.e., we interleave the contents of the four vectors in their order.
2541 We use interleave_high/low instructions to create such output. The input of
2542 each interleave_high/low operation is two vectors:
2543 1st vec 2nd vec
2544 0 1 2 3 4 5 6 7
2545 the even elements of the result vector are obtained left-to-right from the
2546 high/low elements of the first vector. The odd elements of the result are
2547 obtained left-to-right from the high/low elements of the second vector.
2548 The output of interleave_high will be: 0 4 1 5
2549 and of interleave_low: 2 6 3 7
2552 The permutation is done in log LENGTH stages. In each stage interleave_high
2553 and interleave_low stmts are created for each pair of vectors in DR_CHAIN,
2554 where the first argument is taken from the first half of DR_CHAIN and the
2555 second argument from it's second half.
2556 In our example,
2558 I1: interleave_high (1st vec, 3rd vec)
2559 I2: interleave_low (1st vec, 3rd vec)
2560 I3: interleave_high (2nd vec, 4th vec)
2561 I4: interleave_low (2nd vec, 4th vec)
2563 The output for the first stage is:
2565 I1: 0 16 1 17 2 18 3 19
2566 I2: 4 20 5 21 6 22 7 23
2567 I3: 8 24 9 25 10 26 11 27
2568 I4: 12 28 13 29 14 30 15 31
2570 The output of the second stage, i.e. the final result is:
2572 I1: 0 8 16 24 1 9 17 25
2573 I2: 2 10 18 26 3 11 19 27
2574 I3: 4 12 20 28 5 13 21 30
2575 I4: 6 14 22 30 7 15 23 31. */
2577 static bool
2580 tree stmt,
2583 {
2586 tree scalar_dest;
2595 /* Check that the operation is supported. */
2602 {
2604 {
2608 /* Create interleaving stmt:
2609 in the case of big endian:
2610 high = interleave_high (vect1, vect2)
2611 and in the case of little endian:
2612 high = interleave_low (vect1, vect2). */
2620 else
2629 /* Create interleaving stmt:
2630 in the case of big endian:
2631 low = interleave_low (vect1, vect2)
2632 and in the case of little endian:
2633 low = interleave_high (vect1, vect2). */
2641 else
2649 }
2651 }
2653 }
2656 /* Function vectorizable_store.
2658 Check if STMT defines a non scalar data-ref (array/pointer/structure) that
2659 can be vectorized.
2660 If VEC_STMT is also passed, vectorize the STMT: create a vectorized
2661 stmt to replace it, put it in VEC_STMT, and insert it at BSI.
2662 Return FALSE if not a vectorizable STMT, TRUE otherwise. */
2664 bool
2666 {
2667 tree scalar_dest;
2668 tree data_ref;
2669 tree op;
2676 tree dummy;
2678 ssa_op_iter iter;
2679 def_operand_p def_p;
2693 /* Is vectorizable store? */
2706 {
2710 }
2713 /* FORNOW. In some cases can vectorize even if data-type not supported
2714 (e.g. - array initialization with 0). */
2722 {
2726 }
2729 {
2732 }
2734 /** Transform. **/
2740 {
2747 /* We vectorize all the stmts of the interleaving group when we
2748 reach the last stmt in the group. */
2751 {
2754 }
2755 }
2756 else
2757 {
2761 }
2770 /* In case the vectorization factor (VF) is bigger than the number
2771 of elements that we can fit in a vectype (nunits), we have to generate
2772 more than one vector stmt - i.e - we need to "unroll" the
2773 vector stmt by a factor VF/nunits. For more details see documentation in
2774 vect_get_vec_def_for_copy_stmt. */
2776 /* In case of interleaving (non-unit strided access):
2778 S1: &base + 2 = x2
2779 S2: &base = x0
2780 S3: &base + 1 = x1
2781 S4: &base + 3 = x3
2783 We create vectorized stores starting from base address (the access of the
2784 first stmt in the chain (S2 in the above example), when the last store stmt
2785 of the chain (S4) is reached:
2787 VS1: &base = vx2
2788 VS2: &base + vec_size*1 = vx0
2789 VS3: &base + vec_size*2 = vx1
2790 VS4: &base + vec_size*3 = vx3
2792 Then permutation statements are generated:
2794 VS5: vx5 = VEC_INTERLEAVE_HIGH_EXPR < vx0, vx3 >
2795 VS6: vx6 = VEC_INTERLEAVE_LOW_EXPR < vx0, vx3 >
2796 ...
2798 And they are put in STMT_VINFO_VEC_STMT of the corresponding scalar stmts
2799 (the order of the data-refs in the output of vect_permute_store_chain
2800 corresponds to the order of scalar stmts in the interleaving chain - see
2801 the documentation of vect_permute_store_chain()).
2803 In case of both multiple types and interleaving, above vector stores and
2804 permutation stmts are created for every copy. The result vector stmts are
2805 put in STMT_VINFO_VEC_STMT for the first copy and in the corresponding
2806 STMT_VINFO_RELATED_STMT for the next copies.
2807 */
2811 {
2812 tree new_stmt;
2813 tree ptr_incr;
2816 {
2817 /* For interleaved stores we collect vectorized defs for all the
2818 stores in the group in DR_CHAIN and OPRNDS. DR_CHAIN is then used
2819 as an input to vect_permute_store_chain(), and OPRNDS as an input
2820 to vect_get_vec_def_for_stmt_copy() for the next copy.
2821 If the store is not strided, GROUP_SIZE is 1, and DR_CHAIN and
2822 OPRNDS are of size 1. */
2825 {
2826 /* Since gaps are not supported for interleaved stores, GROUP_SIZE
2827 is the exact number of stmts in the chain. Therefore, NEXT_STMT
2828 can't be NULL_TREE. In case that there is no interleaving,
2829 GROUP_SIZE is 1, and only one iteration of the loop will be
2830 executed. */
2837 }
2841 }
2842 else
2843 {
2844 /* For interleaved stores we created vectorized defs for all the
2845 defs stored in OPRNDS in the previous iteration (previous copy).
2846 DR_CHAIN is then used as an input to vect_permute_store_chain(),
2847 and OPRNDS as an input to vect_get_vec_def_for_stmt_copy() for the
2848 next copy.
2849 If the store is not strided, GROUP_SIZE is 1, and DR_CHAIN and
2850 OPRNDS are of size 1. */
2852 {
2857 }
2859 }
2862 {
2864 /* Permute. */
2868 }
2872 {
2873 /* For strided stores vectorized defs are interleaved in
2874 vect_permute_store_chain(). */
2879 /* Arguments are ready. Create the new vector stmt. */
2881 vec_oprnd);
2884 /* Set the VDEFs for the vector pointer. If this virtual def
2885 has a use outside the loop and a loop peel is performed
2886 then the def may be renamed by the peel. Mark it for
2887 renaming so the later use will also be renamed. */
2890 {
2891 /* The original store is deleted so the same SSA_NAMEs
2892 can be used. */
2894 {
2897 }
2900 }
2901 else
2902 {
2903 /* Create new names for all the definitions created by COPY and
2904 add replacement mappings for each new name. */
2906 {
2909 }
2912 }
2918 /* Bump the vector pointer. */
2920 }
2921 }
2924 }
2927 /* Function vect_setup_realignment
2929 This function is called when vectorizing an unaligned load using
2930 the dr_unaligned_software_pipeline scheme.
2931 This function generates the following code at the loop prolog:
2933 p = initial_addr;
2934 msq_init = *(floor(p)); # prolog load
2935 realignment_token = call target_builtin;
2936 loop:
2937 msq = phi (msq_init, ---)
2939 The code above sets up a new (vector) pointer, pointing to the first
2940 location accessed by STMT, and a "floor-aligned" load using that pointer.
2941 It also generates code to compute the "realignment-token" (if the relevant
2942 target hook was defined), and creates a phi-node at the loop-header bb
2943 whose arguments are the result of the prolog-load (created by this
2944 function) and the result of a load that takes place in the loop (to be
2945 created by the caller to this function).
2946 The caller to this function uses the phi-result (msq) to create the
2947 realignment code inside the loop, and sets up the missing phi argument,
2948 as follows:
2950 loop:
2951 msq = phi (msq_init, lsq)
2952 lsq = *(floor(p')); # load in loop
2953 result = realign_load (msq, lsq, realignment_token);
2955 Input:
2956 STMT - (scalar) load stmt to be vectorized. This load accesses
2957 a memory location that may be unaligned.
2958 BSI - place where new code is to be inserted.
2960 Output:
2961 REALIGNMENT_TOKEN - the result of a call to the builtin_mask_for_load
2962 target hook, if defined.
2963 Return value - the result of the loop-header phi node. */
2965 static tree
2968 {
2975 tree vec_dest;
2976 tree init_addr;
2977 tree inc;
2978 tree ptr;
2979 tree data_ref;
2980 tree new_stmt;
2981 basic_block new_bb;
2982 tree msq_init;
2983 tree new_temp;
2984 tree phi_stmt;
2985 tree msq;
2987 /* 1. Create msq_init = *(floor(p1)) in the loop preheader */
2990 NULL_TREE);
3001 /* 2. Create permutation mask, if required, in loop preheader. */
3003 {
3004 tree builtin_decl;
3012 new_stmt);
3019 /* The result of the CALL_EXPR to this builtin is determined from
3020 the value of the parameter and no global variables are touched
3021 which makes the builtin a "const" function. Requiring the
3022 builtin to have the "const" attribute makes it unnecessary
3023 to call mark_call_clobbered. */
3025 }
3027 /* 3. Create msq = phi <msq_init, lsq> in loop */
3035 }
3038 /* Function vect_strided_load_supported.
3040 Returns TRUE is EXTRACT_EVEN and EXTRACT_ODD operations are supported,
3041 and FALSE otherwise. */
3043 static bool
3045 {
3053 {
3057 }
3060 {
3064 }
3068 {
3072 }
3075 {
3079 }
3081 }
3084 /* Function vect_permute_load_chain.
3086 Given a chain of interleaved loads in DR_CHAIN of LENGTH that must be
3087 a power of 2, generate extract_even/odd stmts to reorder the input data
3088 correctly. Return the final references for loads in RESULT_CHAIN.
3090 E.g., LENGTH is 4 and the scalar type is short, i.e., VF is 8.
3091 The input is 4 vectors each containing 8 elements. We assign a number to each
3092 element, the input sequence is:
3094 1st vec: 0 1 2 3 4 5 6 7
3095 2nd vec: 8 9 10 11 12 13 14 15
3096 3rd vec: 16 17 18 19 20 21 22 23
3097 4th vec: 24 25 26 27 28 29 30 31
3099 The output sequence should be:
3101 1st vec: 0 4 8 12 16 20 24 28
3102 2nd vec: 1 5 9 13 17 21 25 29
3103 3rd vec: 2 6 10 14 18 22 26 30
3104 4th vec: 3 7 11 15 19 23 27 31
3106 i.e., the first output vector should contain the first elements of each
3107 interleaving group, etc.
3109 We use extract_even/odd instructions to create such output. The input of each
3110 extract_even/odd operation is two vectors
3111 1st vec 2nd vec
3112 0 1 2 3 4 5 6 7
3114 and the output is the vector of extracted even/odd elements. The output of
3115 extract_even will be: 0 2 4 6
3116 and of extract_odd: 1 3 5 7
3119 The permutation is done in log LENGTH stages. In each stage extract_even and
3120 extract_odd stmts are created for each pair of vectors in DR_CHAIN in their
3121 order. In our example,
3123 E1: extract_even (1st vec, 2nd vec)
3124 E2: extract_odd (1st vec, 2nd vec)
3125 E3: extract_even (3rd vec, 4th vec)
3126 E4: extract_odd (3rd vec, 4th vec)
3128 The output for the first stage will be:
3130 E1: 0 2 4 6 8 10 12 14
3131 E2: 1 3 5 7 9 11 13 15
3132 E3: 16 18 20 22 24 26 28 30
3133 E4: 17 19 21 23 25 27 29 31
3135 In order to proceed and create the correct sequence for the next stage (or
3136 for the correct output, if the second stage is the last one, as in our
3137 example), we first put the output of extract_even operation and then the
3138 output of extract_odd in RESULT_CHAIN (which is then copied to DR_CHAIN).
3139 The input for the second stage is:
3141 1st vec (E1): 0 2 4 6 8 10 12 14
3142 2nd vec (E3): 16 18 20 22 24 26 28 30
3143 3rd vec (E2): 1 3 5 7 9 11 13 15
3144 4th vec (E4): 17 19 21 23 25 27 29 31
3146 The output of the second stage:
3148 E1: 0 4 8 12 16 20 24 28
3149 E2: 2 6 10 14 18 22 26 30
3150 E3: 1 5 9 13 17 21 25 29
3151 E4: 3 7 11 15 19 23 27 31
3153 And RESULT_CHAIN after reordering:
3155 1st vec (E1): 0 4 8 12 16 20 24 28
3156 2nd vec (E3): 1 5 9 13 17 21 25 29
3157 3rd vec (E2): 2 6 10 14 18 22 26 30
3158 4th vec (E4): 3 7 11 15 19 23 27 31. */
3160 static bool
3163 tree stmt,
3166 {
3172 /* Check that the operation is supported. */
3178 {
3180 {
3184 /* data_ref = permute_even (first_data_ref, second_data_ref); */
3200 /* data_ref = permute_odd (first_data_ref, second_data_ref); */
3214 }
3216 }
3218 }
3221 /* Function vect_transform_strided_load.
3223 Given a chain of input interleaved data-refs (in DR_CHAIN), build statements
3224 to perform their permutation and ascribe the result vectorized statements to
3225 the scalar statements.
3226 */
3228 static bool
3231 {
3237 tree tmp_data_ref;
3239 /* DR_CHAIN contains input data-refs that are a part of the interleaving.
3240 RESULT_CHAIN is the output of vect_permute_load_chain, it contains permuted
3241 vectors, that are ready for vector computation. */
3243 /* Permute. */
3247 /* Put a permuted data-ref in the VECTORIZED_STMT field.
3248 Since we scan the chain starting from it's first node, their order
3249 corresponds the order of data-refs in RESULT_CHAIN. */
3253 {
3257 /* Skip the gaps. Loads created for the gaps will be removed by dead
3258 code elimination pass later.
3259 DR_GROUP_GAP is the number of steps in elements from the previous
3260 access (if there is no gap DR_GROUP_GAP is 1). We skip loads that
3261 correspond to the gaps.
3262 */
3264 {
3265 gap_count++;
3267 }
3270 {
3272 /* We assume that if VEC_STMT is not NULL, this is a case of multiple
3273 copies, and we put the new vector statement in the first available
3274 RELATED_STMT. */
3277 else
3278 {
3283 {
3286 }
3288 }
3291 /* If NEXT_STMT accesses the same DR as the previous statement,
3292 put the same TMP_DATA_REF as its vectorized statement; otherwise
3293 get the next data-ref from RESULT_CHAIN. */
3296 }
3297 }
3299 }
3302 /* vectorizable_load.
3304 Check if STMT reads a non scalar data-ref (array/pointer/structure) that
3305 can be vectorized.
3306 If VEC_STMT is also passed, vectorize the STMT: create a vectorized
3307 stmt to replace it, put it in VEC_STMT, and insert it at BSI.
3308 Return FALSE if not a vectorizable STMT, TRUE otherwise. */
3310 bool
3312 {
3313 tree scalar_dest;
3316 tree op;
3318 stmt_vec_info prev_stmt_info;
3323 tree new_temp;
3326 tree dummy;
3329 tree ptr_incr;
3339 tree first_stmt;
3341 /* Is vectorizable load? */
3348 {
3349 /* FORNOW: not yet supported. */
3353 }
3373 /* FORNOW. In some cases can vectorize even if data-type not supported
3374 (e.g. - data copies). */
3376 {
3380 }
3382 /* Check if the load is a part of an interleaving chain. */
3384 {
3387 /* Check if interleaving is supported. */
3390 }
3393 {
3396 }
3398 /** Transform. **/
3404 {
3406 /* Check if the chain of loads is already vectorized. */
3408 {
3411 }
3415 }
3416 else
3417 {
3421 }
3427 /* In case the vectorization factor (VF) is bigger than the number
3428 of elements that we can fit in a vectype (nunits), we have to generate
3429 more than one vector stmt - i.e - we need to "unroll" the
3430 vector stmt by a factor VF/nunits. In doing so, we record a pointer
3431 from one copy of the vector stmt to the next, in the field
3432 STMT_VINFO_RELATED_STMT. This is necessary in order to allow following
3433 stages to find the correct vector defs to be used when vectorizing
3434 stmts that use the defs of the current stmt. The example below illustrates
3435 the vectorization process when VF=16 and nunits=4 (i.e - we need to create
3436 4 vectorized stmts):
3438 before vectorization:
3439 RELATED_STMT VEC_STMT
3440 S1: x = memref - -
3441 S2: z = x + 1 - -
3443 step 1: vectorize stmt S1:
3444 We first create the vector stmt VS1_0, and, as usual, record a
3445 pointer to it in the STMT_VINFO_VEC_STMT of the scalar stmt S1.
3446 Next, we create the vector stmt VS1_1, and record a pointer to
3447 it in the STMT_VINFO_RELATED_STMT of the vector stmt VS1_0.
3448 Similarly, for VS1_2 and VS1_3. This is the resulting chain of
3449 stmts and pointers:
3450 RELATED_STMT VEC_STMT
3451 VS1_0: vx0 = memref0 VS1_1 -
3452 VS1_1: vx1 = memref1 VS1_2 -
3453 VS1_2: vx2 = memref2 VS1_3 -
3454 VS1_3: vx3 = memref3 - -
3455 S1: x = load - VS1_0
3456 S2: z = x + 1 - -
3458 See in documentation in vect_get_vec_def_for_stmt_copy for how the
3459 information we recorded in RELATED_STMT field is used to vectorize
3460 stmt S2. */
3462 /* In case of interleaving (non-unit strided access):
3464 S1: x2 = &base + 2
3465 S2: x0 = &base
3466 S3: x1 = &base + 1
3467 S4: x3 = &base + 3
3469 Vectorized loads are created in the order of memory accesses
3470 starting from the access of the first stmt of the chain:
3472 VS1: vx0 = &base
3473 VS2: vx1 = &base + vec_size*1
3474 VS3: vx3 = &base + vec_size*2
3475 VS4: vx4 = &base + vec_size*3
3477 Then permutation statements are generated:
3479 VS5: vx5 = VEC_EXTRACT_EVEN_EXPR < vx0, vx1 >
3480 VS6: vx6 = VEC_EXTRACT_ODD_EXPR < vx0, vx1 >
3481 ...
3483 And they are put in STMT_VINFO_VEC_STMT of the corresponding scalar stmts
3484 (the order of the data-refs in the output of vect_permute_load_chain
3485 corresponds to the order of scalar stmts in the interleaving chain - see
3486 the documentation of vect_permute_load_chain()).
3487 The generation of permutation stmts and recording them in
3488 STMT_VINFO_VEC_STMT is done in vect_transform_strided_load().
3490 In case of both multiple types and interleaving, the vector loads and
3491 permutation stmts above are created for every copy. The result vector stmts
3492 are put in STMT_VINFO_VEC_STMT for the first copy and in the corresponding
3493 STMT_VINFO_RELATED_STMT for the next copies. */
3495 /* If the data reference is aligned (dr_aligned) or potentially unaligned
3496 on a target that supports unaligned accesses (dr_unaligned_supported)
3497 we generate the following code:
3498 p = initial_addr;
3499 indx = 0;
3500 loop {
3501 p = p + indx * vectype_size;
3502 vec_dest = *(p);
3503 indx = indx + 1;
3504 }
3506 Otherwise, the data reference is potentially unaligned on a target that
3507 does not support unaligned accesses (dr_unaligned_software_pipeline) -
3508 then generate the following code, in which the data in each iteration is
3509 obtained by two vector loads, one from the previous iteration, and one
3510 from the current iteration:
3511 p1 = initial_addr;
3512 msq_init = *(floor(p1))
3513 p2 = initial_addr + VS - 1;
3514 realignment_token = call target_builtin;
3515 indx = 0;
3516 loop {
3517 p2 = p2 + indx * vectype_size
3518 lsq = *(floor(p2))
3519 vec_dest = realign_load (msq, lsq, realignment_token)
3520 indx = indx + 1;
3521 msq = lsq;
3522 } */
3525 {
3529 }
3533 {
3534 /* 1. Create the vector pointer update chain. */
3538 else
3542 {
3543 /* 2. Create the vector-load in the loop. */
3545 {
3551 {
3557 data_ref =
3560 }
3567 }
3570 data_ref);
3577 /* 3. Handle explicit realignment if necessary/supported. */
3579 {
3580 /* Create in loop:
3581 <vec_dest = realign_load (msq, lsq, realignment_token)> */
3586 new_stmt =
3589 new_stmt);
3596 }
3601 }
3604 {
3609 }
3610 else
3611 {
3614 else
3617 }
3618 }
3621 }
3624 /* Function vectorizable_live_operation.
3626 STMT computes a value that is used outside the loop. Check if
3627 it can be supported. */
3629 bool
3633 {
3634 tree operation;
3640 tree op;
3658 /* FORNOW: support only if all uses are invariant. This means
3659 that the scalar operations can remain in place, unvectorized.
3660 The original last scalar value that they compute will be used. */
3663 {
3666 {
3670 }
3674 }
3676 /* No transformation is required for the cases we currently support. */
3678 }
3681 /* Function vect_is_simple_cond.
3683 Input:
3684 LOOP - the loop that is being vectorized.
3685 COND - Condition that is checked for simple use.
3687 Returns whether a COND can be vectorized. Checks whether
3688 condition operands are supportable using vec_is_simple_use. */
3690 static bool
3692 {
3694 tree def;
3704 {
3708 }
3713 {
3717 }
3722 }
3724 /* vectorizable_condition.
3726 Check if STMT is conditional modify expression that can be vectorized.
3727 If VEC_STMT is also passed, vectorize the STMT: create a vectorized
3728 stmt using VEC_COND_EXPR to replace it, put it in VEC_STMT, and insert it
3729 at BSI.
3731 Return FALSE if not a vectorizable STMT, TRUE otherwise. */
3733 bool
3735 {
3744 tree new_temp;
3747 tree def;
3762 {
3763 /* FORNOW: not yet supported. */
3767 }
3784 /* We do not handle two different vector types for the condition
3785 and the values. */
3790 {
3795 }
3801 {
3806 }
3815 {
3818 }
3820 /* Transform */
3822 /* Handle def. */
3826 /* Handle cond expr. */
3827 vec_cond_lhs =
3829 vec_cond_rhs =
3834 /* Arguments are ready. create the new vector stmt. */
3841 vec_cond_expr);
3847 }
3849 /* Function vect_transform_stmt.
3851 Create a vectorized stmt to replace STMT, and insert it at BSI. */
3853 bool
3855 {
3859 tree orig_stmt_in_pattern;
3863 {
3865 {
3895 {
3896 /* In case of interleaving, the whole chain is vectorized when the
3897 last store in the chain is reached. Store stmts before the last
3898 one are skipped, and there vec_stmt_info shouldn't be freed
3899 meanwhile. */
3903 }
3904 else
3921 }
3925 {
3929 {
3932 {
3935 /* STMT was inserted by the vectorizer to replace a
3936 computation idiom. ORIG_STMT_IN_PATTERN is a stmt in the
3937 original sequence that computed this idiom. We need to
3938 record a pointer to VEC_STMT in the stmt_info of
3939 ORIG_STMT_IN_PATTERN. See more details in the
3940 documentation of vect_pattern_recog. */
3943 }
3944 }
3945 }
3946 }
3949 {
3951 {
3960 }
3961 }
3964 }
3967 /* This function builds ni_name = number of iterations loop executes
3968 on the loop preheader. */
3970 static tree
3972 {
3974 edge pe;
3984 {
3987 }
3990 }
3993 /* This function generates the following statements:
3995 ni_name = number of iterations loop executes
3996 ratio = ni_name / vf
3997 ratio_mult_vf_name = ratio * vf
3999 and places them at the loop preheader edge. */
4001 static void
4006 {
4008 edge pe;
4009 basic_block new_bb;
4011 tree var;
4012 tree ratio_name;
4013 tree ratio_mult_vf_name;
4017 tree log_vf;
4021 /* Generate temporary variable that contains
4022 number of iterations loop executes. */
4027 /* Create: ratio = ni >> log2(vf) */
4031 {
4039 }
4041 /* Create: ratio_mult_vf = ratio << log2 (vf). */
4046 {
4055 }
4062 }
4065 /* Function update_vuses_to_preheader.
4067 Input:
4068 STMT - a statement with potential VUSEs.
4069 LOOP - the loop whose preheader will contain STMT.
4071 It's possible to vectorize a loop even though an SSA_NAME from a VUSE
4072 appears to be defined in a VDEF in another statement in a loop.
4073 One such case is when the VUSE is at the dereference of a __restricted__
4074 pointer in a load and the VDEF is at the dereference of a different
4075 __restricted__ pointer in a store. Vectorization may result in
4076 copy_virtual_uses being called to copy the problematic VUSE to a new
4077 statement that is being inserted in the loop preheader. This procedure
4078 is called to change the SSA_NAME in the new statement's VUSE from the
4079 SSA_NAME updated in the loop to the related SSA_NAME available on the
4080 path entering the loop.
4082 When this function is called, we have the following situation:
4084 # vuse <name1>
4085 S1: vload
4086 do {
4087 # name1 = phi < name0 , name2>
4089 # vuse <name1>
4090 S2: vload
4092 # name2 = vdef <name1>
4093 S3: vstore
4095 }while...
4097 Stmt S1 was created in the loop preheader block as part of misaligned-load
4098 handling. This function fixes the name of the vuse of S1 from 'name1' to
4099 'name0'. */
4101 static void
4103 {
4106 ssa_op_iter iter;
4107 use_operand_p use_p;
4110 {
4116 /* For a use before any definitions, def_stmt is a NOP_EXPR. */
4119 {
4120 /* If the block containing the statement defining the SSA_NAME
4121 is in the loop then it's necessary to find the definition
4122 outside the loop using the PHI nodes of the header. */
4123 tree phi;
4127 {
4129 {
4133 }
4134 }
4136 }
4137 }
4138 }
4141 /* Function vect_update_ivs_after_vectorizer.
4143 "Advance" the induction variables of LOOP to the value they should take
4144 after the execution of LOOP. This is currently necessary because the
4145 vectorizer does not handle induction variables that are used after the
4146 loop. Such a situation occurs when the last iterations of LOOP are
4147 peeled, because:
4148 1. We introduced new uses after LOOP for IVs that were not originally used
4149 after LOOP: the IVs of LOOP are now used by an epilog loop.
4150 2. LOOP is going to be vectorized; this means that it will iterate N/VF
4151 times, whereas the loop IVs should be bumped N times.
4153 Input:
4154 - LOOP - a loop that is going to be vectorized. The last few iterations
4155 of LOOP were peeled.
4156 - NITERS - the number of iterations that LOOP executes (before it is
4157 vectorized). i.e, the number of times the ivs should be bumped.
4158 - UPDATE_E - a successor edge of LOOP->exit that is on the (only) path
4159 coming out from LOOP on which there are uses of the LOOP ivs
4160 (this is the path from LOOP->exit to epilog_loop->preheader).
4162 The new definitions of the ivs are placed in LOOP->exit.
4163 The phi args associated with the edge UPDATE_E in the bb
4164 UPDATE_E->dest are updated accordingly.
4166 Assumption 1: Like the rest of the vectorizer, this function assumes
4167 a single loop exit that has a single predecessor.
4169 Assumption 2: The phi nodes in the LOOP header and in update_bb are
4170 organized in the same order.
4172 Assumption 3: The access function of the ivs is simple enough (see
4173 vect_can_advance_ivs_p). This assumption will be relaxed in the future.
4175 Assumption 4: Exactly one of the successors of LOOP exit-bb is on a path
4176 coming out of LOOP on which the ivs of LOOP are used (this is the path
4177 that leads to the epilog loop; other paths skip the epilog loop). This
4178 path starts with the edge UPDATE_E, and its destination (denoted update_bb)
4179 needs to have its phis updated.
4180 */
4182 static void
4184 edge update_e)
4185 {
4191 /* gcc_assert (vect_can_advance_ivs_p (loop_vinfo)); */
4193 /* Make sure there exists a single-predecessor exit bb: */
4199 {
4201 tree evolution_part;
4202 tree init_expr;
4203 tree step_expr;
4205 block_stmt_iterator last_bsi;
4208 {
4211 }
4213 /* Skip virtual phi's. */
4215 {
4219 }
4221 /* Skip reduction phis. */
4223 {
4227 }
4231 evolution_part =
4235 /* FORNOW: We do not support IVs whose evolution function is a polynomial
4236 of degree >= 2 or exponential. */
4246 niters),
4247 step_expr),
4248 init_expr);
4255 /* Insert stmt into exit_bb. */
4260 /* Fix phi expressions in the successor bb. */
4262 }
4263 }
4266 /* Function vect_do_peeling_for_loop_bound
4268 Peel the last iterations of the loop represented by LOOP_VINFO.
4269 The peeled iterations form a new epilog loop. Given that the loop now
4270 iterates NITERS times, the new epilog loop iterates
4271 NITERS % VECTORIZATION_FACTOR times.
4273 The original loop will later be made to iterate
4274 NITERS / VECTORIZATION_FACTOR times (this value is placed into RATIO). */
4276 static void
4278 {
4282 edge update_e;
4283 basic_block preheader;
4292 /* Generate the following variables on the preheader of original loop:
4294 ni_name = number of iteration the original loop executes
4295 ratio = ni_name / vf
4296 ratio_mult_vf_name = ratio * vf */
4301 /* Threshold for vectorized loop. */
4308 #ifdef ENABLE_CHECKING
4310 #endif
4312 /* A guard that controls whether the new_loop is to be executed or skipped
4313 is placed in LOOP->exit. LOOP->exit therefore has two successors - one
4314 is the preheader of NEW_LOOP, where the IVs from LOOP are used. The other
4315 is a bb after NEW_LOOP, where these IVs are not used. Find the edge that
4316 is on the path where the LOOP IVs are used and need to be updated. */
4321 else
4324 /* Update IVs of original loop as if they were advanced
4325 by ratio_mult_vf_name steps. */
4328 /* After peeling we have to reset scalar evolution analyzer. */
4332 }
4335 /* Function vect_gen_niters_for_prolog_loop
4337 Set the number of iterations for the loop represented by LOOP_VINFO
4338 to the minimum between LOOP_NITERS (the original iteration count of the loop)
4339 and the misalignment of DR - the data reference recorded in
4340 LOOP_VINFO_UNALIGNED_DR (LOOP_VINFO). As a result, after the execution of
4341 this loop, the data reference DR will refer to an aligned location.
4343 The following computation is generated:
4345 If the misalignment of DR is known at compile time:
4346 addr_mis = int mis = DR_MISALIGNMENT (dr);
4347 Else, compute address misalignment in bytes:
4348 addr_mis = addr & (vectype_size - 1)
4350 prolog_niters = min ( LOOP_NITERS , (VF - addr_mis/elem_size)&(VF-1) )
4352 (elem_size = element type size; an element is the scalar element
4353 whose type is the inner type of the vectype)
4355 For interleaving,
4357 prolog_niters = min ( LOOP_NITERS ,
4358 (VF/group_size - addr_mis/elem_size)&(VF/group_size-1) )
4359 where group_size is the size of the interleaved group.
4360 */
4362 static tree
4364 {
4370 edge pe;
4371 basic_block new_bb;
4381 {
4382 /* For interleaved access element size must be multiplied by the size of
4383 the interleaved group. */
4387 }
4392 {
4400 }
4401 else
4402 {
4404 tree start_addr =
4410 tree elem_size_log =
4414 tree byte_misalign;
4415 tree elem_misalign;
4420 /* Create: byte_misalign = addr & (vectype_size - 1) */
4421 byte_misalign =
4424 /* Create: elem_misalign = byte_misalign / element_size */
4425 elem_misalign =
4428 /* Create: (niters_type) (VF - elem_misalign)&(VF - 1) */
4432 }
4434 /* Create: prolog_loop_niters = min (iters, loop_niters) */
4435 /* If the loop bound is known at compile time we already verified that it is
4436 greater than vf; since the misalignment ('iters') is at most vf, there's
4437 no need to generate the MIN_EXPR in this case. */
4442 {
4445 }
4451 /* Insert stmt on loop preheader edge. */
4453 {
4456 }
4459 }
4462 /* Function vect_update_init_of_dr
4464 NITERS iterations were peeled from LOOP. DR represents a data reference
4465 in LOOP. This function updates the information recorded in DR to
4466 account for the fact that the first NITERS iterations had already been
4467 executed. Specifically, it updates the OFFSET field of DR. */
4469 static void
4471 {
4477 }
4480 /* Function vect_update_inits_of_drs
4482 NITERS iterations were peeled from the loop represented by LOOP_VINFO.
4483 This function updates the information recorded for the data references in
4484 the loop to account for the fact that the first NITERS iterations had
4485 already been executed. Specifically, it updates the initial_condition of the
4486 access_function of all the data_references in the loop. */
4488 static void
4490 {
4500 }
4503 /* Function vect_do_peeling_for_alignment
4505 Peel the first 'niters' iterations of the loop represented by LOOP_VINFO.
4506 'niters' is set to the misalignment of one of the data references in the
4507 loop, thereby forcing it to refer to an aligned location at the beginning
4508 of the execution of this loop. The data reference for which we are
4509 peeling is recorded in LOOP_VINFO_UNALIGNED_DR. */
4511 static void
4513 {
4516 tree n_iters;
4527 /* Peel the prolog loop and iterate it niters_of_prolog_loop. */
4528 new_loop =
4532 #ifdef ENABLE_CHECKING
4534 #endif
4536 /* Update number of times loop executes. */
4541 /* Update the init conditions of the access functions of all data refs. */
4544 /* After peeling we have to reset scalar evolution analyzer. */
4548 }
4551 /* Function vect_create_cond_for_align_checks.
4553 Create a conditional expression that represents the alignment checks for
4554 all of data references (array element references) whose alignment must be
4555 checked at runtime.
4557 Input:
4558 LOOP_VINFO - two fields of the loop information are used.
4559 LOOP_VINFO_PTR_MASK is the mask used to check the alignment.
4560 LOOP_VINFO_MAY_MISALIGN_STMTS contains the refs to be checked.
4562 Output:
4563 COND_EXPR_STMT_LIST - statements needed to construct the conditional
4564 expression.
4565 The returned value is the conditional expression to be used in the if
4566 statement that controls which version of the loop gets executed at runtime.
4568 The algorithm makes two assumptions:
4569 1) The number of bytes "n" in a vector is a power of 2.
4570 2) An address "a" is aligned if a%n is zero and that this
4571 test can be done as a&(n-1) == 0. For example, for 16
4572 byte vectors the test is a&0xf == 0. */
4574 static tree
4577 {
4580 tree ref_stmt;
4582 tree mask_cst;
4584 tree psize;
4585 tree int_ptrsize_type;
4589 tree ptrsize_zero;
4591 /* Check that mask is one less than a power of 2, i.e., mask is
4592 all zeros followed by all ones. */
4595 /* CHECKME: what is the best integer or unsigned type to use to hold a
4596 cast from a pointer value? */
4598 int_ptrsize_type
4601 /* Create expression (mask & (dr_1 || ... || dr_n)) where dr_i is the address
4602 of the first vector of the i'th data reference. */
4605 {
4607 tree addr_base;
4611 /* create: addr_tmp = (int)(address_of_first_vector) */
4614 NULL_TREE);
4629 /* The addresses are OR together. */
4632 {
4633 /* create: or_tmp = or_tmp | addr_tmp */
4639 new_or_tmp_name,
4641 or_tmp_name,
4642 addr_tmp_name));
4646 }
4647 else
4654 /* create: and_tmp = or_tmp & mask */
4660 and_tmp_name,
4666 /* Make and_tmp the left operand of the conditional test against zero.
4667 if and_tmp has a nonzero bit then some address is unaligned. */
4671 }
4674 /* Function vect_transform_loop.
4676 The analysis phase has determined that the loop is vectorizable.
4677 Vectorize the loop - created vectorized stmts to replace the scalar
4678 stmts in the loop, and update the loop exit condition. */
4680 void
4682 {
4686 block_stmt_iterator si;
4695 /* If the loop has data references that may or may not be aligned then
4696 two versions of the loop need to be generated, one which is vectorized
4697 and one which isn't. A test is then generated to control which of the
4698 loops is executed. The test checks for the alignment of all of the
4699 data references that may or may not be aligned. */
4702 {
4704 tree cond_expr;
4706 basic_block condition_bb;
4707 block_stmt_iterator cond_exp_bsi;
4708 basic_block merge_bb;
4709 basic_block new_exit_bb;
4721 /** Loop versioning violates an assumption we try to maintain during
4722 vectorization - that the loop exit block has a single predecessor.
4723 After versioning, the exit block of both loop versions is the same
4724 basic block (i.e. it has two predecessors). Just in order to simplify
4725 following transformations in the vectorizer, we fix this situation
4726 here by adding a new (empty) block on the exit-edge of the loop,
4727 with the proper loop-exit phis to maintain loop-closed-form. **/
4737 {
4739 new_exit_bb);
4743 }
4745 /** end loop-exit-fixes after versioning **/
4750 }
4752 /* CHECKME: we wouldn't need this if we called update_ssa once
4753 for all loops. */
4756 /* Peel the loop if there are data refs with unknown alignment.
4757 Only one data ref with unknown store is allowed. */
4762 /* If the loop has a symbolic number of iterations 'n' (i.e. it's not a
4763 compile time constant), or it is a constant that doesn't divide by the
4764 vectorization factor, then an epilog loop needs to be created.
4765 We therefore duplicate the loop: the original loop will be vectorized,
4766 and will compute the first (n/VF) iterations. The second copy of the loop
4767 will remain scalar and will compute the remaining (n%VF) iterations.
4768 (VF is the vectorization factor). */
4774 else
4778 /* 1) Make sure the loop header has exactly two entries
4779 2) Make sure we have a preheader basic block. */
4785 /* FORNOW: the vectorizer supports only loops which body consist
4786 of one basic block (header + empty latch). When the vectorizer will
4787 support more involved loop forms, the order by which the BBs are
4788 traversed need to be reconsidered. */
4791 {
4795 {
4797 stmt_vec_info stmt_info;
4801 {
4804 }
4809 {
4812 }
4819 /* -------- vectorize statement ------------ */
4826 {
4827 stmt_ann_t ann;
4829 {
4830 /* Interleaving. If IS_STORE is TRUE, the vectorization of the
4831 interleaving chain was completed - free all the stores in
4832 the chain. */
4834 tree tmp;
4835 stmt_vec_info next_stmt_info;
4838 {
4840 /* Free the attached stmt_vec_info and remove the stmt. */
4846 }
4849 }
4850 else
4851 {
4852 /* Free the attached stmt_vec_info and remove the stmt. */
4858 }
4859 }
4860 else
4861 {
4863 {
4864 /* This is case of skipped interleaved store. We don't free
4865 its stmt_vec_info. */
4868 }
4869 }
4878 /* The memory tags and pointers in vectorized statements need to
4879 have their SSA forms updated. FIXME, why can't this be delayed
4880 until all the loops have been transformed? */
4885 }