]>
Commit | Line | Data |
---|---|---|
b8698a0f | 1 | /* Data References Analysis and Manipulation Utilities for Vectorization. |
ebfd146a IR |
2 | Copyright (C) 2003, 2004, 2005, 2006, 2007, 2008, 2009 Free Software |
3 | Foundation, Inc. | |
b8698a0f | 4 | Contributed by Dorit Naishlos <dorit@il.ibm.com> |
ebfd146a IR |
5 | and Ira Rosen <irar@il.ibm.com> |
6 | ||
7 | This file is part of GCC. | |
8 | ||
9 | GCC is free software; you can redistribute it and/or modify it under | |
10 | the terms of the GNU General Public License as published by the Free | |
11 | Software Foundation; either version 3, or (at your option) any later | |
12 | version. | |
13 | ||
14 | GCC is distributed in the hope that it will be useful, but WITHOUT ANY | |
15 | WARRANTY; without even the implied warranty of MERCHANTABILITY or | |
16 | FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License | |
17 | for more details. | |
18 | ||
19 | You should have received a copy of the GNU General Public License | |
20 | along with GCC; see the file COPYING3. If not see | |
21 | <http://www.gnu.org/licenses/>. */ | |
22 | ||
23 | #include "config.h" | |
24 | #include "system.h" | |
25 | #include "coretypes.h" | |
26 | #include "tm.h" | |
27 | #include "ggc.h" | |
28 | #include "tree.h" | |
29 | #include "target.h" | |
30 | #include "basic-block.h" | |
31 | #include "diagnostic.h" | |
32 | #include "tree-flow.h" | |
33 | #include "tree-dump.h" | |
34 | #include "cfgloop.h" | |
35 | #include "expr.h" | |
36 | #include "optabs.h" | |
37 | #include "tree-chrec.h" | |
38 | #include "tree-scalar-evolution.h" | |
39 | #include "tree-vectorizer.h" | |
40 | #include "toplev.h" | |
41 | ||
42 | ||
43 | /* Return the smallest scalar part of STMT. | |
b8698a0f L |
44 | This is used to determine the vectype of the stmt. We generally set the |
45 | vectype according to the type of the result (lhs). For stmts whose | |
ebfd146a | 46 | result-type is different than the type of the arguments (e.g., demotion, |
b8698a0f | 47 | promotion), vectype will be reset appropriately (later). Note that we have |
ebfd146a | 48 | to visit the smallest datatype in this function, because that determines the |
b8698a0f | 49 | VF. If the smallest datatype in the loop is present only as the rhs of a |
ebfd146a IR |
50 | promotion operation - we'd miss it. |
51 | Such a case, where a variable of this datatype does not appear in the lhs | |
52 | anywhere in the loop, can only occur if it's an invariant: e.g.: | |
b8698a0f | 53 | 'int_x = (int) short_inv', which we'd expect to have been optimized away by |
ebfd146a IR |
54 | invariant motion. However, we cannot rely on invariant motion to always take |
55 | invariants out of the loop, and so in the case of promotion we also have to | |
b8698a0f | 56 | check the rhs. |
ebfd146a IR |
57 | LHS_SIZE_UNIT and RHS_SIZE_UNIT contain the sizes of the corresponding |
58 | types. */ | |
59 | ||
60 | tree | |
61 | vect_get_smallest_scalar_type (gimple stmt, HOST_WIDE_INT *lhs_size_unit, | |
62 | HOST_WIDE_INT *rhs_size_unit) | |
63 | { | |
64 | tree scalar_type = gimple_expr_type (stmt); | |
65 | HOST_WIDE_INT lhs, rhs; | |
66 | ||
67 | lhs = rhs = TREE_INT_CST_LOW (TYPE_SIZE_UNIT (scalar_type)); | |
68 | ||
69 | if (is_gimple_assign (stmt) | |
70 | && (gimple_assign_cast_p (stmt) | |
71 | || gimple_assign_rhs_code (stmt) == WIDEN_MULT_EXPR | |
72 | || gimple_assign_rhs_code (stmt) == FLOAT_EXPR)) | |
73 | { | |
74 | tree rhs_type = TREE_TYPE (gimple_assign_rhs1 (stmt)); | |
75 | ||
76 | rhs = TREE_INT_CST_LOW (TYPE_SIZE_UNIT (rhs_type)); | |
77 | if (rhs < lhs) | |
78 | scalar_type = rhs_type; | |
79 | } | |
b8698a0f L |
80 | |
81 | *lhs_size_unit = lhs; | |
ebfd146a IR |
82 | *rhs_size_unit = rhs; |
83 | return scalar_type; | |
84 | } | |
85 | ||
86 | ||
87 | /* Find the place of the data-ref in STMT in the interleaving chain that starts | |
88 | from FIRST_STMT. Return -1 if the data-ref is not a part of the chain. */ | |
89 | ||
b8698a0f | 90 | int |
ebfd146a IR |
91 | vect_get_place_in_interleaving_chain (gimple stmt, gimple first_stmt) |
92 | { | |
93 | gimple next_stmt = first_stmt; | |
94 | int result = 0; | |
95 | ||
96 | if (first_stmt != DR_GROUP_FIRST_DR (vinfo_for_stmt (stmt))) | |
97 | return -1; | |
98 | ||
99 | while (next_stmt && next_stmt != stmt) | |
100 | { | |
101 | result++; | |
102 | next_stmt = DR_GROUP_NEXT_DR (vinfo_for_stmt (next_stmt)); | |
103 | } | |
104 | ||
105 | if (next_stmt) | |
106 | return result; | |
107 | else | |
108 | return -1; | |
109 | } | |
110 | ||
111 | ||
112 | /* Function vect_insert_into_interleaving_chain. | |
113 | ||
114 | Insert DRA into the interleaving chain of DRB according to DRA's INIT. */ | |
115 | ||
116 | static void | |
117 | vect_insert_into_interleaving_chain (struct data_reference *dra, | |
118 | struct data_reference *drb) | |
119 | { | |
120 | gimple prev, next; | |
121 | tree next_init; | |
b8698a0f | 122 | stmt_vec_info stmtinfo_a = vinfo_for_stmt (DR_STMT (dra)); |
ebfd146a IR |
123 | stmt_vec_info stmtinfo_b = vinfo_for_stmt (DR_STMT (drb)); |
124 | ||
125 | prev = DR_GROUP_FIRST_DR (stmtinfo_b); | |
b8698a0f | 126 | next = DR_GROUP_NEXT_DR (vinfo_for_stmt (prev)); |
ebfd146a IR |
127 | while (next) |
128 | { | |
129 | next_init = DR_INIT (STMT_VINFO_DATA_REF (vinfo_for_stmt (next))); | |
130 | if (tree_int_cst_compare (next_init, DR_INIT (dra)) > 0) | |
131 | { | |
132 | /* Insert here. */ | |
133 | DR_GROUP_NEXT_DR (vinfo_for_stmt (prev)) = DR_STMT (dra); | |
134 | DR_GROUP_NEXT_DR (stmtinfo_a) = next; | |
135 | return; | |
136 | } | |
137 | prev = next; | |
138 | next = DR_GROUP_NEXT_DR (vinfo_for_stmt (prev)); | |
139 | } | |
140 | ||
141 | /* We got to the end of the list. Insert here. */ | |
142 | DR_GROUP_NEXT_DR (vinfo_for_stmt (prev)) = DR_STMT (dra); | |
143 | DR_GROUP_NEXT_DR (stmtinfo_a) = NULL; | |
144 | } | |
145 | ||
146 | ||
147 | /* Function vect_update_interleaving_chain. | |
b8698a0f L |
148 | |
149 | For two data-refs DRA and DRB that are a part of a chain interleaved data | |
ebfd146a IR |
150 | accesses, update the interleaving chain. DRB's INIT is smaller than DRA's. |
151 | ||
152 | There are four possible cases: | |
153 | 1. New stmts - both DRA and DRB are not a part of any chain: | |
154 | FIRST_DR = DRB | |
155 | NEXT_DR (DRB) = DRA | |
156 | 2. DRB is a part of a chain and DRA is not: | |
157 | no need to update FIRST_DR | |
158 | no need to insert DRB | |
159 | insert DRA according to init | |
160 | 3. DRA is a part of a chain and DRB is not: | |
161 | if (init of FIRST_DR > init of DRB) | |
162 | FIRST_DR = DRB | |
163 | NEXT(FIRST_DR) = previous FIRST_DR | |
164 | else | |
165 | insert DRB according to its init | |
166 | 4. both DRA and DRB are in some interleaving chains: | |
167 | choose the chain with the smallest init of FIRST_DR | |
168 | insert the nodes of the second chain into the first one. */ | |
169 | ||
170 | static void | |
171 | vect_update_interleaving_chain (struct data_reference *drb, | |
172 | struct data_reference *dra) | |
173 | { | |
b8698a0f | 174 | stmt_vec_info stmtinfo_a = vinfo_for_stmt (DR_STMT (dra)); |
ebfd146a IR |
175 | stmt_vec_info stmtinfo_b = vinfo_for_stmt (DR_STMT (drb)); |
176 | tree next_init, init_dra_chain, init_drb_chain; | |
177 | gimple first_a, first_b; | |
178 | tree node_init; | |
179 | gimple node, prev, next, first_stmt; | |
180 | ||
181 | /* 1. New stmts - both DRA and DRB are not a part of any chain. */ | |
182 | if (!DR_GROUP_FIRST_DR (stmtinfo_a) && !DR_GROUP_FIRST_DR (stmtinfo_b)) | |
183 | { | |
184 | DR_GROUP_FIRST_DR (stmtinfo_a) = DR_STMT (drb); | |
185 | DR_GROUP_FIRST_DR (stmtinfo_b) = DR_STMT (drb); | |
186 | DR_GROUP_NEXT_DR (stmtinfo_b) = DR_STMT (dra); | |
187 | return; | |
188 | } | |
189 | ||
190 | /* 2. DRB is a part of a chain and DRA is not. */ | |
191 | if (!DR_GROUP_FIRST_DR (stmtinfo_a) && DR_GROUP_FIRST_DR (stmtinfo_b)) | |
192 | { | |
193 | DR_GROUP_FIRST_DR (stmtinfo_a) = DR_GROUP_FIRST_DR (stmtinfo_b); | |
194 | /* Insert DRA into the chain of DRB. */ | |
195 | vect_insert_into_interleaving_chain (dra, drb); | |
196 | return; | |
197 | } | |
198 | ||
b8698a0f | 199 | /* 3. DRA is a part of a chain and DRB is not. */ |
ebfd146a IR |
200 | if (DR_GROUP_FIRST_DR (stmtinfo_a) && !DR_GROUP_FIRST_DR (stmtinfo_b)) |
201 | { | |
202 | gimple old_first_stmt = DR_GROUP_FIRST_DR (stmtinfo_a); | |
203 | tree init_old = DR_INIT (STMT_VINFO_DATA_REF (vinfo_for_stmt ( | |
204 | old_first_stmt))); | |
205 | gimple tmp; | |
206 | ||
207 | if (tree_int_cst_compare (init_old, DR_INIT (drb)) > 0) | |
208 | { | |
b8698a0f L |
209 | /* DRB's init is smaller than the init of the stmt previously marked |
210 | as the first stmt of the interleaving chain of DRA. Therefore, we | |
ebfd146a IR |
211 | update FIRST_STMT and put DRB in the head of the list. */ |
212 | DR_GROUP_FIRST_DR (stmtinfo_b) = DR_STMT (drb); | |
213 | DR_GROUP_NEXT_DR (stmtinfo_b) = old_first_stmt; | |
b8698a0f | 214 | |
ebfd146a IR |
215 | /* Update all the stmts in the list to point to the new FIRST_STMT. */ |
216 | tmp = old_first_stmt; | |
217 | while (tmp) | |
218 | { | |
219 | DR_GROUP_FIRST_DR (vinfo_for_stmt (tmp)) = DR_STMT (drb); | |
220 | tmp = DR_GROUP_NEXT_DR (vinfo_for_stmt (tmp)); | |
221 | } | |
222 | } | |
223 | else | |
224 | { | |
225 | /* Insert DRB in the list of DRA. */ | |
226 | vect_insert_into_interleaving_chain (drb, dra); | |
b8698a0f | 227 | DR_GROUP_FIRST_DR (stmtinfo_b) = DR_GROUP_FIRST_DR (stmtinfo_a); |
ebfd146a IR |
228 | } |
229 | return; | |
230 | } | |
b8698a0f | 231 | |
ebfd146a IR |
232 | /* 4. both DRA and DRB are in some interleaving chains. */ |
233 | first_a = DR_GROUP_FIRST_DR (stmtinfo_a); | |
234 | first_b = DR_GROUP_FIRST_DR (stmtinfo_b); | |
235 | if (first_a == first_b) | |
236 | return; | |
237 | init_dra_chain = DR_INIT (STMT_VINFO_DATA_REF (vinfo_for_stmt (first_a))); | |
238 | init_drb_chain = DR_INIT (STMT_VINFO_DATA_REF (vinfo_for_stmt (first_b))); | |
239 | ||
240 | if (tree_int_cst_compare (init_dra_chain, init_drb_chain) > 0) | |
241 | { | |
b8698a0f | 242 | /* Insert the nodes of DRA chain into the DRB chain. |
ebfd146a IR |
243 | After inserting a node, continue from this node of the DRB chain (don't |
244 | start from the beginning. */ | |
245 | node = DR_GROUP_FIRST_DR (stmtinfo_a); | |
b8698a0f | 246 | prev = DR_GROUP_FIRST_DR (stmtinfo_b); |
ebfd146a IR |
247 | first_stmt = first_b; |
248 | } | |
249 | else | |
250 | { | |
b8698a0f | 251 | /* Insert the nodes of DRB chain into the DRA chain. |
ebfd146a IR |
252 | After inserting a node, continue from this node of the DRA chain (don't |
253 | start from the beginning. */ | |
254 | node = DR_GROUP_FIRST_DR (stmtinfo_b); | |
b8698a0f | 255 | prev = DR_GROUP_FIRST_DR (stmtinfo_a); |
ebfd146a IR |
256 | first_stmt = first_a; |
257 | } | |
b8698a0f | 258 | |
ebfd146a IR |
259 | while (node) |
260 | { | |
261 | node_init = DR_INIT (STMT_VINFO_DATA_REF (vinfo_for_stmt (node))); | |
b8698a0f | 262 | next = DR_GROUP_NEXT_DR (vinfo_for_stmt (prev)); |
ebfd146a | 263 | while (next) |
b8698a0f | 264 | { |
ebfd146a IR |
265 | next_init = DR_INIT (STMT_VINFO_DATA_REF (vinfo_for_stmt (next))); |
266 | if (tree_int_cst_compare (next_init, node_init) > 0) | |
267 | { | |
268 | /* Insert here. */ | |
269 | DR_GROUP_NEXT_DR (vinfo_for_stmt (prev)) = node; | |
270 | DR_GROUP_NEXT_DR (vinfo_for_stmt (node)) = next; | |
271 | prev = node; | |
272 | break; | |
273 | } | |
274 | prev = next; | |
275 | next = DR_GROUP_NEXT_DR (vinfo_for_stmt (prev)); | |
276 | } | |
277 | if (!next) | |
278 | { | |
279 | /* We got to the end of the list. Insert here. */ | |
280 | DR_GROUP_NEXT_DR (vinfo_for_stmt (prev)) = node; | |
281 | DR_GROUP_NEXT_DR (vinfo_for_stmt (node)) = NULL; | |
282 | prev = node; | |
b8698a0f | 283 | } |
ebfd146a | 284 | DR_GROUP_FIRST_DR (vinfo_for_stmt (node)) = first_stmt; |
b8698a0f | 285 | node = DR_GROUP_NEXT_DR (vinfo_for_stmt (node)); |
ebfd146a IR |
286 | } |
287 | } | |
288 | ||
289 | ||
290 | /* Function vect_equal_offsets. | |
291 | ||
292 | Check if OFFSET1 and OFFSET2 are identical expressions. */ | |
293 | ||
294 | static bool | |
295 | vect_equal_offsets (tree offset1, tree offset2) | |
296 | { | |
297 | bool res0, res1; | |
298 | ||
299 | STRIP_NOPS (offset1); | |
300 | STRIP_NOPS (offset2); | |
301 | ||
302 | if (offset1 == offset2) | |
303 | return true; | |
304 | ||
305 | if (TREE_CODE (offset1) != TREE_CODE (offset2) | |
306 | || !BINARY_CLASS_P (offset1) | |
b8698a0f | 307 | || !BINARY_CLASS_P (offset2)) |
ebfd146a | 308 | return false; |
b8698a0f L |
309 | |
310 | res0 = vect_equal_offsets (TREE_OPERAND (offset1, 0), | |
ebfd146a | 311 | TREE_OPERAND (offset2, 0)); |
b8698a0f | 312 | res1 = vect_equal_offsets (TREE_OPERAND (offset1, 1), |
ebfd146a IR |
313 | TREE_OPERAND (offset2, 1)); |
314 | ||
315 | return (res0 && res1); | |
316 | } | |
317 | ||
318 | ||
319 | /* Function vect_check_interleaving. | |
320 | ||
321 | Check if DRA and DRB are a part of interleaving. In case they are, insert | |
322 | DRA and DRB in an interleaving chain. */ | |
323 | ||
b8698a0f | 324 | static bool |
ebfd146a IR |
325 | vect_check_interleaving (struct data_reference *dra, |
326 | struct data_reference *drb) | |
327 | { | |
328 | HOST_WIDE_INT type_size_a, type_size_b, diff_mod_size, step, init_a, init_b; | |
329 | ||
330 | /* Check that the data-refs have same first location (except init) and they | |
331 | are both either store or load (not load and store). */ | |
332 | if ((DR_BASE_ADDRESS (dra) != DR_BASE_ADDRESS (drb) | |
b8698a0f | 333 | && (TREE_CODE (DR_BASE_ADDRESS (dra)) != ADDR_EXPR |
ebfd146a | 334 | || TREE_CODE (DR_BASE_ADDRESS (drb)) != ADDR_EXPR |
b8698a0f | 335 | || TREE_OPERAND (DR_BASE_ADDRESS (dra), 0) |
ebfd146a IR |
336 | != TREE_OPERAND (DR_BASE_ADDRESS (drb),0))) |
337 | || !vect_equal_offsets (DR_OFFSET (dra), DR_OFFSET (drb)) | |
b8698a0f | 338 | || !tree_int_cst_compare (DR_INIT (dra), DR_INIT (drb)) |
ebfd146a | 339 | || DR_IS_READ (dra) != DR_IS_READ (drb)) |
8644a673 | 340 | return false; |
ebfd146a IR |
341 | |
342 | /* Check: | |
343 | 1. data-refs are of the same type | |
344 | 2. their steps are equal | |
b8698a0f | 345 | 3. the step (if greater than zero) is greater than the difference between |
8644a673 | 346 | data-refs' inits. */ |
ebfd146a IR |
347 | type_size_a = TREE_INT_CST_LOW (TYPE_SIZE_UNIT (TREE_TYPE (DR_REF (dra)))); |
348 | type_size_b = TREE_INT_CST_LOW (TYPE_SIZE_UNIT (TREE_TYPE (DR_REF (drb)))); | |
349 | ||
350 | if (type_size_a != type_size_b | |
351 | || tree_int_cst_compare (DR_STEP (dra), DR_STEP (drb)) | |
b8698a0f | 352 | || !types_compatible_p (TREE_TYPE (DR_REF (dra)), |
ebfd146a | 353 | TREE_TYPE (DR_REF (drb)))) |
8644a673 | 354 | return false; |
ebfd146a IR |
355 | |
356 | init_a = TREE_INT_CST_LOW (DR_INIT (dra)); | |
357 | init_b = TREE_INT_CST_LOW (DR_INIT (drb)); | |
358 | step = TREE_INT_CST_LOW (DR_STEP (dra)); | |
359 | ||
360 | if (init_a > init_b) | |
361 | { | |
b8698a0f | 362 | /* If init_a == init_b + the size of the type * k, we have an interleaving, |
ebfd146a IR |
363 | and DRB is accessed before DRA. */ |
364 | diff_mod_size = (init_a - init_b) % type_size_a; | |
365 | ||
a70d6342 | 366 | if (step && (init_a - init_b) > step) |
b8698a0f | 367 | return false; |
ebfd146a IR |
368 | |
369 | if (diff_mod_size == 0) | |
370 | { | |
b8698a0f | 371 | vect_update_interleaving_chain (drb, dra); |
ebfd146a IR |
372 | if (vect_print_dump_info (REPORT_DR_DETAILS)) |
373 | { | |
374 | fprintf (vect_dump, "Detected interleaving "); | |
375 | print_generic_expr (vect_dump, DR_REF (dra), TDF_SLIM); | |
376 | fprintf (vect_dump, " and "); | |
377 | print_generic_expr (vect_dump, DR_REF (drb), TDF_SLIM); | |
378 | } | |
8644a673 | 379 | return true; |
b8698a0f | 380 | } |
ebfd146a | 381 | } |
b8698a0f | 382 | else |
ebfd146a | 383 | { |
b8698a0f | 384 | /* If init_b == init_a + the size of the type * k, we have an |
ebfd146a IR |
385 | interleaving, and DRA is accessed before DRB. */ |
386 | diff_mod_size = (init_b - init_a) % type_size_a; | |
387 | ||
a70d6342 | 388 | if (step && (init_b - init_a) > step) |
8644a673 | 389 | return false; |
ebfd146a IR |
390 | |
391 | if (diff_mod_size == 0) | |
392 | { | |
b8698a0f | 393 | vect_update_interleaving_chain (dra, drb); |
ebfd146a IR |
394 | if (vect_print_dump_info (REPORT_DR_DETAILS)) |
395 | { | |
396 | fprintf (vect_dump, "Detected interleaving "); | |
397 | print_generic_expr (vect_dump, DR_REF (dra), TDF_SLIM); | |
398 | fprintf (vect_dump, " and "); | |
399 | print_generic_expr (vect_dump, DR_REF (drb), TDF_SLIM); | |
400 | } | |
8644a673 | 401 | return true; |
b8698a0f | 402 | } |
ebfd146a | 403 | } |
b8698a0f | 404 | |
8644a673 | 405 | return false; |
ebfd146a IR |
406 | } |
407 | ||
408 | /* Check if data references pointed by DR_I and DR_J are same or | |
409 | belong to same interleaving group. Return FALSE if drs are | |
410 | different, otherwise return TRUE. */ | |
411 | ||
412 | static bool | |
413 | vect_same_range_drs (data_reference_p dr_i, data_reference_p dr_j) | |
414 | { | |
415 | gimple stmt_i = DR_STMT (dr_i); | |
416 | gimple stmt_j = DR_STMT (dr_j); | |
417 | ||
418 | if (operand_equal_p (DR_REF (dr_i), DR_REF (dr_j), 0) | |
419 | || (DR_GROUP_FIRST_DR (vinfo_for_stmt (stmt_i)) | |
420 | && DR_GROUP_FIRST_DR (vinfo_for_stmt (stmt_j)) | |
421 | && (DR_GROUP_FIRST_DR (vinfo_for_stmt (stmt_i)) | |
422 | == DR_GROUP_FIRST_DR (vinfo_for_stmt (stmt_j))))) | |
423 | return true; | |
424 | else | |
425 | return false; | |
426 | } | |
427 | ||
428 | /* If address ranges represented by DDR_I and DDR_J are equal, | |
429 | return TRUE, otherwise return FALSE. */ | |
430 | ||
431 | static bool | |
432 | vect_vfa_range_equal (ddr_p ddr_i, ddr_p ddr_j) | |
433 | { | |
434 | if ((vect_same_range_drs (DDR_A (ddr_i), DDR_A (ddr_j)) | |
435 | && vect_same_range_drs (DDR_B (ddr_i), DDR_B (ddr_j))) | |
436 | || (vect_same_range_drs (DDR_A (ddr_i), DDR_B (ddr_j)) | |
437 | && vect_same_range_drs (DDR_B (ddr_i), DDR_A (ddr_j)))) | |
438 | return true; | |
439 | else | |
440 | return false; | |
441 | } | |
442 | ||
443 | /* Insert DDR into LOOP_VINFO list of ddrs that may alias and need to be | |
444 | tested at run-time. Return TRUE if DDR was successfully inserted. | |
445 | Return false if versioning is not supported. */ | |
446 | ||
447 | static bool | |
448 | vect_mark_for_runtime_alias_test (ddr_p ddr, loop_vec_info loop_vinfo) | |
449 | { | |
450 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
451 | ||
452 | if ((unsigned) PARAM_VALUE (PARAM_VECT_MAX_VERSION_FOR_ALIAS_CHECKS) == 0) | |
453 | return false; | |
454 | ||
455 | if (vect_print_dump_info (REPORT_DR_DETAILS)) | |
456 | { | |
457 | fprintf (vect_dump, "mark for run-time aliasing test between "); | |
458 | print_generic_expr (vect_dump, DR_REF (DDR_A (ddr)), TDF_SLIM); | |
459 | fprintf (vect_dump, " and "); | |
460 | print_generic_expr (vect_dump, DR_REF (DDR_B (ddr)), TDF_SLIM); | |
461 | } | |
462 | ||
463 | if (optimize_loop_nest_for_size_p (loop)) | |
464 | { | |
465 | if (vect_print_dump_info (REPORT_DR_DETAILS)) | |
466 | fprintf (vect_dump, "versioning not supported when optimizing for size."); | |
467 | return false; | |
468 | } | |
469 | ||
470 | /* FORNOW: We don't support versioning with outer-loop vectorization. */ | |
471 | if (loop->inner) | |
472 | { | |
473 | if (vect_print_dump_info (REPORT_DR_DETAILS)) | |
474 | fprintf (vect_dump, "versioning not yet supported for outer-loops."); | |
475 | return false; | |
476 | } | |
477 | ||
478 | VEC_safe_push (ddr_p, heap, LOOP_VINFO_MAY_ALIAS_DDRS (loop_vinfo), ddr); | |
479 | return true; | |
480 | } | |
481 | ||
a70d6342 | 482 | |
ebfd146a IR |
483 | /* Function vect_analyze_data_ref_dependence. |
484 | ||
485 | Return TRUE if there (might) exist a dependence between a memory-reference | |
486 | DRA and a memory-reference DRB. When versioning for alias may check a | |
487 | dependence at run-time, return FALSE. */ | |
b8698a0f | 488 | |
ebfd146a IR |
489 | static bool |
490 | vect_analyze_data_ref_dependence (struct data_dependence_relation *ddr, | |
491 | loop_vec_info loop_vinfo) | |
492 | { | |
493 | unsigned int i; | |
a70d6342 IR |
494 | struct loop *loop = NULL; |
495 | int vectorization_factor = 0; | |
ebfd146a IR |
496 | struct data_reference *dra = DDR_A (ddr); |
497 | struct data_reference *drb = DDR_B (ddr); | |
b8698a0f | 498 | stmt_vec_info stmtinfo_a = vinfo_for_stmt (DR_STMT (dra)); |
ebfd146a IR |
499 | stmt_vec_info stmtinfo_b = vinfo_for_stmt (DR_STMT (drb)); |
500 | int dra_size = GET_MODE_SIZE (TYPE_MODE (TREE_TYPE (DR_REF (dra)))); | |
501 | int drb_size = GET_MODE_SIZE (TYPE_MODE (TREE_TYPE (DR_REF (drb)))); | |
502 | lambda_vector dist_v; | |
503 | unsigned int loop_depth; | |
b8698a0f | 504 | |
ebfd146a IR |
505 | if (DDR_ARE_DEPENDENT (ddr) == chrec_known) |
506 | { | |
507 | /* Independent data accesses. */ | |
508 | vect_check_interleaving (dra, drb); | |
509 | return false; | |
510 | } | |
511 | ||
a70d6342 IR |
512 | if (loop_vinfo) |
513 | { | |
514 | loop = LOOP_VINFO_LOOP (loop_vinfo); | |
515 | vectorization_factor = LOOP_VINFO_VECT_FACTOR (loop_vinfo); | |
516 | } | |
517 | ||
518 | if ((DR_IS_READ (dra) && DR_IS_READ (drb) && loop_vinfo) || dra == drb) | |
ebfd146a | 519 | return false; |
b8698a0f | 520 | |
ebfd146a IR |
521 | if (DDR_ARE_DEPENDENT (ddr) == chrec_dont_know) |
522 | { | |
b8698a0f | 523 | if (loop_vinfo) |
a70d6342 IR |
524 | { |
525 | if (vect_print_dump_info (REPORT_DR_DETAILS)) | |
526 | { | |
527 | fprintf (vect_dump, "versioning for alias required: " | |
528 | "can't determine dependence between "); | |
529 | print_generic_expr (vect_dump, DR_REF (dra), TDF_SLIM); | |
530 | fprintf (vect_dump, " and "); | |
531 | print_generic_expr (vect_dump, DR_REF (drb), TDF_SLIM); | |
532 | } | |
b8698a0f | 533 | |
a70d6342 IR |
534 | /* Add to list of ddrs that need to be tested at run-time. */ |
535 | return !vect_mark_for_runtime_alias_test (ddr, loop_vinfo); | |
536 | } | |
537 | ||
538 | /* When vectorizing a basic block unknown depnedence can still mean | |
539 | strided access. */ | |
540 | if (vect_check_interleaving (dra, drb)) | |
541 | return false; | |
542 | ||
ebfd146a IR |
543 | if (vect_print_dump_info (REPORT_DR_DETAILS)) |
544 | { | |
a70d6342 | 545 | fprintf (vect_dump, "can't determine dependence between "); |
ebfd146a IR |
546 | print_generic_expr (vect_dump, DR_REF (dra), TDF_SLIM); |
547 | fprintf (vect_dump, " and "); | |
548 | print_generic_expr (vect_dump, DR_REF (drb), TDF_SLIM); | |
549 | } | |
a70d6342 IR |
550 | |
551 | return true; | |
ebfd146a IR |
552 | } |
553 | ||
a70d6342 | 554 | /* Versioning for alias is not yet supported for basic block SLP, and |
b8698a0f | 555 | dependence distance is unapplicable, hence, in case of known data |
a70d6342 IR |
556 | dependence, basic block vectorization is impossible for now. */ |
557 | if (!loop_vinfo) | |
558 | { | |
559 | if (dra != drb && vect_check_interleaving (dra, drb)) | |
560 | return false; | |
b8698a0f | 561 | |
a70d6342 IR |
562 | if (vect_print_dump_info (REPORT_DR_DETAILS)) |
563 | { | |
564 | fprintf (vect_dump, "determined dependence between "); | |
565 | print_generic_expr (vect_dump, DR_REF (dra), TDF_SLIM); | |
566 | fprintf (vect_dump, " and "); | |
567 | print_generic_expr (vect_dump, DR_REF (drb), TDF_SLIM); | |
568 | } | |
569 | ||
b8698a0f | 570 | return true; |
a70d6342 IR |
571 | } |
572 | ||
573 | /* Loop-based vectorization and known data dependence. */ | |
ebfd146a IR |
574 | if (DDR_NUM_DIST_VECTS (ddr) == 0) |
575 | { | |
576 | if (vect_print_dump_info (REPORT_DR_DETAILS)) | |
577 | { | |
578 | fprintf (vect_dump, "versioning for alias required: bad dist vector for "); | |
579 | print_generic_expr (vect_dump, DR_REF (dra), TDF_SLIM); | |
580 | fprintf (vect_dump, " and "); | |
581 | print_generic_expr (vect_dump, DR_REF (drb), TDF_SLIM); | |
582 | } | |
583 | /* Add to list of ddrs that need to be tested at run-time. */ | |
584 | return !vect_mark_for_runtime_alias_test (ddr, loop_vinfo); | |
b8698a0f | 585 | } |
ebfd146a IR |
586 | |
587 | loop_depth = index_in_loop_nest (loop->num, DDR_LOOP_NEST (ddr)); | |
588 | for (i = 0; VEC_iterate (lambda_vector, DDR_DIST_VECTS (ddr), i, dist_v); i++) | |
589 | { | |
590 | int dist = dist_v[loop_depth]; | |
591 | ||
592 | if (vect_print_dump_info (REPORT_DR_DETAILS)) | |
593 | fprintf (vect_dump, "dependence distance = %d.", dist); | |
594 | ||
595 | /* Same loop iteration. */ | |
596 | if (dist % vectorization_factor == 0 && dra_size == drb_size) | |
597 | { | |
598 | /* Two references with distance zero have the same alignment. */ | |
599 | VEC_safe_push (dr_p, heap, STMT_VINFO_SAME_ALIGN_REFS (stmtinfo_a), drb); | |
600 | VEC_safe_push (dr_p, heap, STMT_VINFO_SAME_ALIGN_REFS (stmtinfo_b), dra); | |
601 | if (vect_print_dump_info (REPORT_ALIGNMENT)) | |
602 | fprintf (vect_dump, "accesses have the same alignment."); | |
603 | if (vect_print_dump_info (REPORT_DR_DETAILS)) | |
604 | { | |
605 | fprintf (vect_dump, "dependence distance modulo vf == 0 between "); | |
606 | print_generic_expr (vect_dump, DR_REF (dra), TDF_SLIM); | |
607 | fprintf (vect_dump, " and "); | |
608 | print_generic_expr (vect_dump, DR_REF (drb), TDF_SLIM); | |
609 | } | |
610 | ||
611 | /* For interleaving, mark that there is a read-write dependency if | |
b8698a0f | 612 | necessary. We check before that one of the data-refs is store. */ |
ebfd146a IR |
613 | if (DR_IS_READ (dra)) |
614 | DR_GROUP_READ_WRITE_DEPENDENCE (stmtinfo_a) = true; | |
615 | else | |
616 | { | |
617 | if (DR_IS_READ (drb)) | |
618 | DR_GROUP_READ_WRITE_DEPENDENCE (stmtinfo_b) = true; | |
619 | } | |
b8698a0f | 620 | |
ebfd146a IR |
621 | continue; |
622 | } | |
623 | ||
b8698a0f | 624 | if (abs (dist) >= vectorization_factor |
ebfd146a IR |
625 | || (dist > 0 && DDR_REVERSED_P (ddr))) |
626 | { | |
b8698a0f L |
627 | /* Dependence distance does not create dependence, as far as |
628 | vectorization is concerned, in this case. If DDR_REVERSED_P the | |
ebfd146a IR |
629 | order of the data-refs in DDR was reversed (to make distance |
630 | vector positive), and the actual distance is negative. */ | |
631 | if (vect_print_dump_info (REPORT_DR_DETAILS)) | |
632 | fprintf (vect_dump, "dependence distance >= VF or negative."); | |
633 | continue; | |
634 | } | |
635 | ||
8644a673 | 636 | if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS)) |
ebfd146a | 637 | { |
a70d6342 IR |
638 | fprintf (vect_dump, "not vectorized, possible dependence " |
639 | "between data-refs "); | |
ebfd146a IR |
640 | print_generic_expr (vect_dump, DR_REF (dra), TDF_SLIM); |
641 | fprintf (vect_dump, " and "); | |
642 | print_generic_expr (vect_dump, DR_REF (drb), TDF_SLIM); | |
643 | } | |
644 | ||
645 | return true; | |
646 | } | |
647 | ||
648 | return false; | |
649 | } | |
650 | ||
651 | /* Function vect_analyze_data_ref_dependences. | |
b8698a0f | 652 | |
ebfd146a IR |
653 | Examine all the data references in the loop, and make sure there do not |
654 | exist any data dependences between them. */ | |
b8698a0f | 655 | |
ebfd146a | 656 | bool |
b8698a0f | 657 | vect_analyze_data_ref_dependences (loop_vec_info loop_vinfo, |
a70d6342 | 658 | bb_vec_info bb_vinfo) |
ebfd146a IR |
659 | { |
660 | unsigned int i; | |
a70d6342 | 661 | VEC (ddr_p, heap) *ddrs = NULL; |
ebfd146a IR |
662 | struct data_dependence_relation *ddr; |
663 | ||
b8698a0f | 664 | if (vect_print_dump_info (REPORT_DETAILS)) |
ebfd146a | 665 | fprintf (vect_dump, "=== vect_analyze_dependences ==="); |
b8698a0f | 666 | |
a70d6342 IR |
667 | if (loop_vinfo) |
668 | ddrs = LOOP_VINFO_DDRS (loop_vinfo); | |
669 | else | |
670 | ddrs = BB_VINFO_DDRS (bb_vinfo); | |
b8698a0f | 671 | |
ebfd146a IR |
672 | for (i = 0; VEC_iterate (ddr_p, ddrs, i, ddr); i++) |
673 | if (vect_analyze_data_ref_dependence (ddr, loop_vinfo)) | |
674 | return false; | |
675 | ||
676 | return true; | |
677 | } | |
678 | ||
679 | ||
680 | /* Function vect_compute_data_ref_alignment | |
681 | ||
682 | Compute the misalignment of the data reference DR. | |
683 | ||
684 | Output: | |
685 | 1. If during the misalignment computation it is found that the data reference | |
686 | cannot be vectorized then false is returned. | |
687 | 2. DR_MISALIGNMENT (DR) is defined. | |
688 | ||
689 | FOR NOW: No analysis is actually performed. Misalignment is calculated | |
690 | only for trivial cases. TODO. */ | |
691 | ||
692 | static bool | |
693 | vect_compute_data_ref_alignment (struct data_reference *dr) | |
694 | { | |
695 | gimple stmt = DR_STMT (dr); | |
b8698a0f | 696 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); |
ebfd146a | 697 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
a70d6342 | 698 | struct loop *loop = NULL; |
ebfd146a IR |
699 | tree ref = DR_REF (dr); |
700 | tree vectype; | |
701 | tree base, base_addr; | |
702 | bool base_aligned; | |
703 | tree misalign; | |
704 | tree aligned_to, alignment; | |
b8698a0f | 705 | |
ebfd146a IR |
706 | if (vect_print_dump_info (REPORT_DETAILS)) |
707 | fprintf (vect_dump, "vect_compute_data_ref_alignment:"); | |
708 | ||
a70d6342 IR |
709 | if (loop_vinfo) |
710 | loop = LOOP_VINFO_LOOP (loop_vinfo); | |
b8698a0f | 711 | |
ebfd146a IR |
712 | /* Initialize misalignment to unknown. */ |
713 | SET_DR_MISALIGNMENT (dr, -1); | |
714 | ||
715 | misalign = DR_INIT (dr); | |
716 | aligned_to = DR_ALIGNED_TO (dr); | |
717 | base_addr = DR_BASE_ADDRESS (dr); | |
718 | vectype = STMT_VINFO_VECTYPE (stmt_info); | |
719 | ||
720 | /* In case the dataref is in an inner-loop of the loop that is being | |
721 | vectorized (LOOP), we use the base and misalignment information | |
722 | relative to the outer-loop (LOOP). This is ok only if the misalignment | |
723 | stays the same throughout the execution of the inner-loop, which is why | |
724 | we have to check that the stride of the dataref in the inner-loop evenly | |
725 | divides by the vector size. */ | |
a70d6342 | 726 | if (loop && nested_in_vect_loop_p (loop, stmt)) |
ebfd146a IR |
727 | { |
728 | tree step = DR_STEP (dr); | |
729 | HOST_WIDE_INT dr_step = TREE_INT_CST_LOW (step); | |
b8698a0f | 730 | |
ebfd146a IR |
731 | if (dr_step % GET_MODE_SIZE (TYPE_MODE (vectype)) == 0) |
732 | { | |
733 | if (vect_print_dump_info (REPORT_ALIGNMENT)) | |
734 | fprintf (vect_dump, "inner step divides the vector-size."); | |
735 | misalign = STMT_VINFO_DR_INIT (stmt_info); | |
736 | aligned_to = STMT_VINFO_DR_ALIGNED_TO (stmt_info); | |
737 | base_addr = STMT_VINFO_DR_BASE_ADDRESS (stmt_info); | |
738 | } | |
739 | else | |
740 | { | |
741 | if (vect_print_dump_info (REPORT_ALIGNMENT)) | |
742 | fprintf (vect_dump, "inner step doesn't divide the vector-size."); | |
743 | misalign = NULL_TREE; | |
744 | } | |
745 | } | |
746 | ||
747 | base = build_fold_indirect_ref (base_addr); | |
748 | alignment = ssize_int (TYPE_ALIGN (vectype)/BITS_PER_UNIT); | |
749 | ||
750 | if ((aligned_to && tree_int_cst_compare (aligned_to, alignment) < 0) | |
751 | || !misalign) | |
752 | { | |
753 | if (vect_print_dump_info (REPORT_ALIGNMENT)) | |
754 | { | |
755 | fprintf (vect_dump, "Unknown alignment for access: "); | |
756 | print_generic_expr (vect_dump, base, TDF_SLIM); | |
757 | } | |
758 | return true; | |
759 | } | |
760 | ||
b8698a0f | 761 | if ((DECL_P (base) |
ebfd146a IR |
762 | && tree_int_cst_compare (ssize_int (DECL_ALIGN_UNIT (base)), |
763 | alignment) >= 0) | |
764 | || (TREE_CODE (base_addr) == SSA_NAME | |
765 | && tree_int_cst_compare (ssize_int (TYPE_ALIGN_UNIT (TREE_TYPE ( | |
766 | TREE_TYPE (base_addr)))), | |
767 | alignment) >= 0)) | |
768 | base_aligned = true; | |
769 | else | |
b8698a0f | 770 | base_aligned = false; |
ebfd146a | 771 | |
b8698a0f | 772 | if (!base_aligned) |
ebfd146a | 773 | { |
b8698a0f | 774 | /* Do not change the alignment of global variables if |
ebfd146a IR |
775 | flag_section_anchors is enabled. */ |
776 | if (!vect_can_force_dr_alignment_p (base, TYPE_ALIGN (vectype)) | |
777 | || (TREE_STATIC (base) && flag_section_anchors)) | |
778 | { | |
779 | if (vect_print_dump_info (REPORT_DETAILS)) | |
780 | { | |
781 | fprintf (vect_dump, "can't force alignment of ref: "); | |
782 | print_generic_expr (vect_dump, ref, TDF_SLIM); | |
783 | } | |
784 | return true; | |
785 | } | |
b8698a0f | 786 | |
ebfd146a IR |
787 | /* Force the alignment of the decl. |
788 | NOTE: This is the only change to the code we make during | |
789 | the analysis phase, before deciding to vectorize the loop. */ | |
790 | if (vect_print_dump_info (REPORT_DETAILS)) | |
791 | fprintf (vect_dump, "force alignment"); | |
792 | DECL_ALIGN (base) = TYPE_ALIGN (vectype); | |
793 | DECL_USER_ALIGN (base) = 1; | |
794 | } | |
795 | ||
796 | /* At this point we assume that the base is aligned. */ | |
797 | gcc_assert (base_aligned | |
b8698a0f | 798 | || (TREE_CODE (base) == VAR_DECL |
ebfd146a IR |
799 | && DECL_ALIGN (base) >= TYPE_ALIGN (vectype))); |
800 | ||
801 | /* Modulo alignment. */ | |
802 | misalign = size_binop (FLOOR_MOD_EXPR, misalign, alignment); | |
803 | ||
804 | if (!host_integerp (misalign, 1)) | |
805 | { | |
806 | /* Negative or overflowed misalignment value. */ | |
807 | if (vect_print_dump_info (REPORT_DETAILS)) | |
808 | fprintf (vect_dump, "unexpected misalign value"); | |
809 | return false; | |
810 | } | |
811 | ||
812 | SET_DR_MISALIGNMENT (dr, TREE_INT_CST_LOW (misalign)); | |
813 | ||
814 | if (vect_print_dump_info (REPORT_DETAILS)) | |
815 | { | |
816 | fprintf (vect_dump, "misalign = %d bytes of ref ", DR_MISALIGNMENT (dr)); | |
817 | print_generic_expr (vect_dump, ref, TDF_SLIM); | |
818 | } | |
819 | ||
820 | return true; | |
821 | } | |
822 | ||
823 | ||
824 | /* Function vect_compute_data_refs_alignment | |
825 | ||
826 | Compute the misalignment of data references in the loop. | |
827 | Return FALSE if a data reference is found that cannot be vectorized. */ | |
828 | ||
829 | static bool | |
b8698a0f | 830 | vect_compute_data_refs_alignment (loop_vec_info loop_vinfo, |
a70d6342 | 831 | bb_vec_info bb_vinfo) |
ebfd146a | 832 | { |
a70d6342 | 833 | VEC (data_reference_p, heap) *datarefs; |
ebfd146a IR |
834 | struct data_reference *dr; |
835 | unsigned int i; | |
836 | ||
a70d6342 IR |
837 | if (loop_vinfo) |
838 | datarefs = LOOP_VINFO_DATAREFS (loop_vinfo); | |
839 | else | |
840 | datarefs = BB_VINFO_DATAREFS (bb_vinfo); | |
b8698a0f | 841 | |
ebfd146a IR |
842 | for (i = 0; VEC_iterate (data_reference_p, datarefs, i, dr); i++) |
843 | if (!vect_compute_data_ref_alignment (dr)) | |
844 | return false; | |
845 | ||
846 | return true; | |
847 | } | |
848 | ||
849 | ||
850 | /* Function vect_update_misalignment_for_peel | |
851 | ||
852 | DR - the data reference whose misalignment is to be adjusted. | |
853 | DR_PEEL - the data reference whose misalignment is being made | |
854 | zero in the vector loop by the peel. | |
855 | NPEEL - the number of iterations in the peel loop if the misalignment | |
856 | of DR_PEEL is known at compile time. */ | |
857 | ||
858 | static void | |
859 | vect_update_misalignment_for_peel (struct data_reference *dr, | |
860 | struct data_reference *dr_peel, int npeel) | |
861 | { | |
862 | unsigned int i; | |
863 | VEC(dr_p,heap) *same_align_drs; | |
864 | struct data_reference *current_dr; | |
865 | int dr_size = GET_MODE_SIZE (TYPE_MODE (TREE_TYPE (DR_REF (dr)))); | |
866 | int dr_peel_size = GET_MODE_SIZE (TYPE_MODE (TREE_TYPE (DR_REF (dr_peel)))); | |
867 | stmt_vec_info stmt_info = vinfo_for_stmt (DR_STMT (dr)); | |
868 | stmt_vec_info peel_stmt_info = vinfo_for_stmt (DR_STMT (dr_peel)); | |
869 | ||
870 | /* For interleaved data accesses the step in the loop must be multiplied by | |
871 | the size of the interleaving group. */ | |
872 | if (STMT_VINFO_STRIDED_ACCESS (stmt_info)) | |
873 | dr_size *= DR_GROUP_SIZE (vinfo_for_stmt (DR_GROUP_FIRST_DR (stmt_info))); | |
874 | if (STMT_VINFO_STRIDED_ACCESS (peel_stmt_info)) | |
875 | dr_peel_size *= DR_GROUP_SIZE (peel_stmt_info); | |
876 | ||
877 | /* It can be assumed that the data refs with the same alignment as dr_peel | |
878 | are aligned in the vector loop. */ | |
879 | same_align_drs | |
880 | = STMT_VINFO_SAME_ALIGN_REFS (vinfo_for_stmt (DR_STMT (dr_peel))); | |
881 | for (i = 0; VEC_iterate (dr_p, same_align_drs, i, current_dr); i++) | |
882 | { | |
883 | if (current_dr != dr) | |
884 | continue; | |
885 | gcc_assert (DR_MISALIGNMENT (dr) / dr_size == | |
886 | DR_MISALIGNMENT (dr_peel) / dr_peel_size); | |
887 | SET_DR_MISALIGNMENT (dr, 0); | |
888 | return; | |
889 | } | |
890 | ||
891 | if (known_alignment_for_access_p (dr) | |
892 | && known_alignment_for_access_p (dr_peel)) | |
893 | { | |
894 | int misal = DR_MISALIGNMENT (dr); | |
895 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
896 | misal += npeel * dr_size; | |
897 | misal %= GET_MODE_SIZE (TYPE_MODE (vectype)); | |
898 | SET_DR_MISALIGNMENT (dr, misal); | |
899 | return; | |
900 | } | |
901 | ||
902 | if (vect_print_dump_info (REPORT_DETAILS)) | |
903 | fprintf (vect_dump, "Setting misalignment to -1."); | |
904 | SET_DR_MISALIGNMENT (dr, -1); | |
905 | } | |
906 | ||
907 | ||
908 | /* Function vect_verify_datarefs_alignment | |
909 | ||
910 | Return TRUE if all data references in the loop can be | |
911 | handled with respect to alignment. */ | |
912 | ||
a70d6342 IR |
913 | bool |
914 | vect_verify_datarefs_alignment (loop_vec_info loop_vinfo, bb_vec_info bb_vinfo) | |
ebfd146a | 915 | { |
a70d6342 | 916 | VEC (data_reference_p, heap) *datarefs; |
ebfd146a IR |
917 | struct data_reference *dr; |
918 | enum dr_alignment_support supportable_dr_alignment; | |
919 | unsigned int i; | |
920 | ||
a70d6342 IR |
921 | if (loop_vinfo) |
922 | datarefs = LOOP_VINFO_DATAREFS (loop_vinfo); | |
923 | else | |
924 | datarefs = BB_VINFO_DATAREFS (bb_vinfo); | |
925 | ||
ebfd146a IR |
926 | for (i = 0; VEC_iterate (data_reference_p, datarefs, i, dr); i++) |
927 | { | |
928 | gimple stmt = DR_STMT (dr); | |
929 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
930 | ||
931 | /* For interleaving, only the alignment of the first access matters. */ | |
932 | if (STMT_VINFO_STRIDED_ACCESS (stmt_info) | |
933 | && DR_GROUP_FIRST_DR (stmt_info) != stmt) | |
934 | continue; | |
935 | ||
936 | supportable_dr_alignment = vect_supportable_dr_alignment (dr); | |
937 | if (!supportable_dr_alignment) | |
938 | { | |
8644a673 | 939 | if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS)) |
ebfd146a IR |
940 | { |
941 | if (DR_IS_READ (dr)) | |
b8698a0f | 942 | fprintf (vect_dump, |
ebfd146a IR |
943 | "not vectorized: unsupported unaligned load."); |
944 | else | |
b8698a0f | 945 | fprintf (vect_dump, |
ebfd146a IR |
946 | "not vectorized: unsupported unaligned store."); |
947 | } | |
948 | return false; | |
949 | } | |
950 | if (supportable_dr_alignment != dr_aligned | |
951 | && vect_print_dump_info (REPORT_ALIGNMENT)) | |
952 | fprintf (vect_dump, "Vectorizing an unaligned access."); | |
953 | } | |
954 | return true; | |
955 | } | |
956 | ||
957 | ||
958 | /* Function vector_alignment_reachable_p | |
959 | ||
960 | Return true if vector alignment for DR is reachable by peeling | |
961 | a few loop iterations. Return false otherwise. */ | |
962 | ||
963 | static bool | |
964 | vector_alignment_reachable_p (struct data_reference *dr) | |
965 | { | |
966 | gimple stmt = DR_STMT (dr); | |
967 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
968 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
969 | ||
970 | if (STMT_VINFO_STRIDED_ACCESS (stmt_info)) | |
971 | { | |
972 | /* For interleaved access we peel only if number of iterations in | |
973 | the prolog loop ({VF - misalignment}), is a multiple of the | |
974 | number of the interleaved accesses. */ | |
975 | int elem_size, mis_in_elements; | |
976 | int nelements = TYPE_VECTOR_SUBPARTS (vectype); | |
977 | ||
978 | /* FORNOW: handle only known alignment. */ | |
979 | if (!known_alignment_for_access_p (dr)) | |
980 | return false; | |
981 | ||
982 | elem_size = GET_MODE_SIZE (TYPE_MODE (vectype)) / nelements; | |
983 | mis_in_elements = DR_MISALIGNMENT (dr) / elem_size; | |
984 | ||
985 | if ((nelements - mis_in_elements) % DR_GROUP_SIZE (stmt_info)) | |
986 | return false; | |
987 | } | |
988 | ||
989 | /* If misalignment is known at the compile time then allow peeling | |
990 | only if natural alignment is reachable through peeling. */ | |
991 | if (known_alignment_for_access_p (dr) && !aligned_access_p (dr)) | |
992 | { | |
b8698a0f | 993 | HOST_WIDE_INT elmsize = |
ebfd146a IR |
994 | int_cst_value (TYPE_SIZE_UNIT (TREE_TYPE (vectype))); |
995 | if (vect_print_dump_info (REPORT_DETAILS)) | |
996 | { | |
997 | fprintf (vect_dump, "data size =" HOST_WIDE_INT_PRINT_DEC, elmsize); | |
998 | fprintf (vect_dump, ". misalignment = %d. ", DR_MISALIGNMENT (dr)); | |
999 | } | |
1000 | if (DR_MISALIGNMENT (dr) % elmsize) | |
1001 | { | |
1002 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1003 | fprintf (vect_dump, "data size does not divide the misalignment.\n"); | |
1004 | return false; | |
1005 | } | |
1006 | } | |
1007 | ||
1008 | if (!known_alignment_for_access_p (dr)) | |
1009 | { | |
1010 | tree type = (TREE_TYPE (DR_REF (dr))); | |
1011 | tree ba = DR_BASE_OBJECT (dr); | |
1012 | bool is_packed = false; | |
1013 | ||
1014 | if (ba) | |
1015 | is_packed = contains_packed_reference (ba); | |
1016 | ||
1017 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1018 | fprintf (vect_dump, "Unknown misalignment, is_packed = %d",is_packed); | |
1019 | if (targetm.vectorize.vector_alignment_reachable (type, is_packed)) | |
1020 | return true; | |
1021 | else | |
1022 | return false; | |
1023 | } | |
1024 | ||
1025 | return true; | |
1026 | } | |
1027 | ||
1028 | /* Function vect_enhance_data_refs_alignment | |
1029 | ||
1030 | This pass will use loop versioning and loop peeling in order to enhance | |
1031 | the alignment of data references in the loop. | |
1032 | ||
1033 | FOR NOW: we assume that whatever versioning/peeling takes place, only the | |
1034 | original loop is to be vectorized; Any other loops that are created by | |
1035 | the transformations performed in this pass - are not supposed to be | |
1036 | vectorized. This restriction will be relaxed. | |
1037 | ||
1038 | This pass will require a cost model to guide it whether to apply peeling | |
1039 | or versioning or a combination of the two. For example, the scheme that | |
1040 | intel uses when given a loop with several memory accesses, is as follows: | |
1041 | choose one memory access ('p') which alignment you want to force by doing | |
1042 | peeling. Then, either (1) generate a loop in which 'p' is aligned and all | |
1043 | other accesses are not necessarily aligned, or (2) use loop versioning to | |
1044 | generate one loop in which all accesses are aligned, and another loop in | |
1045 | which only 'p' is necessarily aligned. | |
1046 | ||
1047 | ("Automatic Intra-Register Vectorization for the Intel Architecture", | |
1048 | Aart J.C. Bik, Milind Girkar, Paul M. Grey and Ximmin Tian, International | |
1049 | Journal of Parallel Programming, Vol. 30, No. 2, April 2002.) | |
1050 | ||
1051 | Devising a cost model is the most critical aspect of this work. It will | |
1052 | guide us on which access to peel for, whether to use loop versioning, how | |
1053 | many versions to create, etc. The cost model will probably consist of | |
1054 | generic considerations as well as target specific considerations (on | |
1055 | powerpc for example, misaligned stores are more painful than misaligned | |
1056 | loads). | |
1057 | ||
1058 | Here are the general steps involved in alignment enhancements: | |
1059 | ||
1060 | -- original loop, before alignment analysis: | |
1061 | for (i=0; i<N; i++){ | |
1062 | x = q[i]; # DR_MISALIGNMENT(q) = unknown | |
1063 | p[i] = y; # DR_MISALIGNMENT(p) = unknown | |
1064 | } | |
1065 | ||
1066 | -- After vect_compute_data_refs_alignment: | |
1067 | for (i=0; i<N; i++){ | |
1068 | x = q[i]; # DR_MISALIGNMENT(q) = 3 | |
1069 | p[i] = y; # DR_MISALIGNMENT(p) = unknown | |
1070 | } | |
1071 | ||
1072 | -- Possibility 1: we do loop versioning: | |
1073 | if (p is aligned) { | |
1074 | for (i=0; i<N; i++){ # loop 1A | |
1075 | x = q[i]; # DR_MISALIGNMENT(q) = 3 | |
1076 | p[i] = y; # DR_MISALIGNMENT(p) = 0 | |
1077 | } | |
1078 | } | |
1079 | else { | |
1080 | for (i=0; i<N; i++){ # loop 1B | |
1081 | x = q[i]; # DR_MISALIGNMENT(q) = 3 | |
1082 | p[i] = y; # DR_MISALIGNMENT(p) = unaligned | |
1083 | } | |
1084 | } | |
1085 | ||
1086 | -- Possibility 2: we do loop peeling: | |
1087 | for (i = 0; i < 3; i++){ # (scalar loop, not to be vectorized). | |
1088 | x = q[i]; | |
1089 | p[i] = y; | |
1090 | } | |
1091 | for (i = 3; i < N; i++){ # loop 2A | |
1092 | x = q[i]; # DR_MISALIGNMENT(q) = 0 | |
1093 | p[i] = y; # DR_MISALIGNMENT(p) = unknown | |
1094 | } | |
1095 | ||
1096 | -- Possibility 3: combination of loop peeling and versioning: | |
1097 | for (i = 0; i < 3; i++){ # (scalar loop, not to be vectorized). | |
1098 | x = q[i]; | |
1099 | p[i] = y; | |
1100 | } | |
1101 | if (p is aligned) { | |
1102 | for (i = 3; i<N; i++){ # loop 3A | |
1103 | x = q[i]; # DR_MISALIGNMENT(q) = 0 | |
1104 | p[i] = y; # DR_MISALIGNMENT(p) = 0 | |
1105 | } | |
1106 | } | |
1107 | else { | |
1108 | for (i = 3; i<N; i++){ # loop 3B | |
1109 | x = q[i]; # DR_MISALIGNMENT(q) = 0 | |
1110 | p[i] = y; # DR_MISALIGNMENT(p) = unaligned | |
1111 | } | |
1112 | } | |
1113 | ||
1114 | These loops are later passed to loop_transform to be vectorized. The | |
1115 | vectorizer will use the alignment information to guide the transformation | |
1116 | (whether to generate regular loads/stores, or with special handling for | |
1117 | misalignment). */ | |
1118 | ||
1119 | bool | |
1120 | vect_enhance_data_refs_alignment (loop_vec_info loop_vinfo) | |
1121 | { | |
1122 | VEC (data_reference_p, heap) *datarefs = LOOP_VINFO_DATAREFS (loop_vinfo); | |
1123 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
1124 | enum dr_alignment_support supportable_dr_alignment; | |
1125 | struct data_reference *dr0 = NULL; | |
1126 | struct data_reference *dr; | |
1127 | unsigned int i; | |
1128 | bool do_peeling = false; | |
1129 | bool do_versioning = false; | |
1130 | bool stat; | |
1131 | gimple stmt; | |
1132 | stmt_vec_info stmt_info; | |
1133 | int vect_versioning_for_alias_required; | |
1134 | ||
1135 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1136 | fprintf (vect_dump, "=== vect_enhance_data_refs_alignment ==="); | |
1137 | ||
1138 | /* While cost model enhancements are expected in the future, the high level | |
1139 | view of the code at this time is as follows: | |
1140 | ||
673beced RE |
1141 | A) If there is a misaligned access then see if peeling to align |
1142 | this access can make all data references satisfy | |
8f439681 RE |
1143 | vect_supportable_dr_alignment. If so, update data structures |
1144 | as needed and return true. | |
ebfd146a IR |
1145 | |
1146 | B) If peeling wasn't possible and there is a data reference with an | |
1147 | unknown misalignment that does not satisfy vect_supportable_dr_alignment | |
1148 | then see if loop versioning checks can be used to make all data | |
1149 | references satisfy vect_supportable_dr_alignment. If so, update | |
1150 | data structures as needed and return true. | |
1151 | ||
1152 | C) If neither peeling nor versioning were successful then return false if | |
1153 | any data reference does not satisfy vect_supportable_dr_alignment. | |
1154 | ||
1155 | D) Return true (all data references satisfy vect_supportable_dr_alignment). | |
1156 | ||
1157 | Note, Possibility 3 above (which is peeling and versioning together) is not | |
1158 | being done at this time. */ | |
1159 | ||
1160 | /* (1) Peeling to force alignment. */ | |
1161 | ||
1162 | /* (1.1) Decide whether to perform peeling, and how many iterations to peel: | |
1163 | Considerations: | |
1164 | + How many accesses will become aligned due to the peeling | |
1165 | - How many accesses will become unaligned due to the peeling, | |
1166 | and the cost of misaligned accesses. | |
b8698a0f | 1167 | - The cost of peeling (the extra runtime checks, the increase |
ebfd146a IR |
1168 | in code size). |
1169 | ||
1170 | The scheme we use FORNOW: peel to force the alignment of the first | |
8f439681 | 1171 | unsupported misaligned access in the loop. |
ebfd146a IR |
1172 | |
1173 | TODO: Use a cost model. */ | |
1174 | ||
1175 | for (i = 0; VEC_iterate (data_reference_p, datarefs, i, dr); i++) | |
1176 | { | |
1177 | stmt = DR_STMT (dr); | |
1178 | stmt_info = vinfo_for_stmt (stmt); | |
1179 | ||
1180 | /* For interleaving, only the alignment of the first access | |
1181 | matters. */ | |
1182 | if (STMT_VINFO_STRIDED_ACCESS (stmt_info) | |
1183 | && DR_GROUP_FIRST_DR (stmt_info) != stmt) | |
1184 | continue; | |
1185 | ||
0cf7986c | 1186 | if (!DR_IS_READ (dr) && !aligned_access_p (dr)) |
ebfd146a IR |
1187 | { |
1188 | do_peeling = vector_alignment_reachable_p (dr); | |
1189 | if (do_peeling) | |
1190 | dr0 = dr; | |
1191 | if (!do_peeling && vect_print_dump_info (REPORT_DETAILS)) | |
1192 | fprintf (vect_dump, "vector alignment may not be reachable"); | |
1193 | break; | |
1194 | } | |
1195 | } | |
1196 | ||
b8698a0f | 1197 | vect_versioning_for_alias_required |
e9dbe7bb | 1198 | = LOOP_REQUIRES_VERSIONING_FOR_ALIAS (loop_vinfo); |
ebfd146a IR |
1199 | |
1200 | /* Temporarily, if versioning for alias is required, we disable peeling | |
1201 | until we support peeling and versioning. Often peeling for alignment | |
1202 | will require peeling for loop-bound, which in turn requires that we | |
1203 | know how to adjust the loop ivs after the loop. */ | |
1204 | if (vect_versioning_for_alias_required | |
e9dbe7bb | 1205 | || !vect_can_advance_ivs_p (loop_vinfo) |
ebfd146a IR |
1206 | || !slpeel_can_duplicate_loop_p (loop, single_exit (loop))) |
1207 | do_peeling = false; | |
1208 | ||
1209 | if (do_peeling) | |
1210 | { | |
1211 | int mis; | |
1212 | int npeel = 0; | |
1213 | gimple stmt = DR_STMT (dr0); | |
1214 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
1215 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
1216 | int nelements = TYPE_VECTOR_SUBPARTS (vectype); | |
1217 | ||
1218 | if (known_alignment_for_access_p (dr0)) | |
1219 | { | |
1220 | /* Since it's known at compile time, compute the number of iterations | |
1221 | in the peeled loop (the peeling factor) for use in updating | |
1222 | DR_MISALIGNMENT values. The peeling factor is the vectorization | |
1223 | factor minus the misalignment as an element count. */ | |
1224 | mis = DR_MISALIGNMENT (dr0); | |
1225 | mis /= GET_MODE_SIZE (TYPE_MODE (TREE_TYPE (DR_REF (dr0)))); | |
1226 | npeel = nelements - mis; | |
1227 | ||
b8698a0f | 1228 | /* For interleaved data access every iteration accesses all the |
ebfd146a IR |
1229 | members of the group, therefore we divide the number of iterations |
1230 | by the group size. */ | |
b8698a0f | 1231 | stmt_info = vinfo_for_stmt (DR_STMT (dr0)); |
ebfd146a IR |
1232 | if (STMT_VINFO_STRIDED_ACCESS (stmt_info)) |
1233 | npeel /= DR_GROUP_SIZE (stmt_info); | |
1234 | ||
1235 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1236 | fprintf (vect_dump, "Try peeling by %d", npeel); | |
1237 | } | |
1238 | ||
1239 | /* Ensure that all data refs can be vectorized after the peel. */ | |
1240 | for (i = 0; VEC_iterate (data_reference_p, datarefs, i, dr); i++) | |
1241 | { | |
1242 | int save_misalignment; | |
1243 | ||
1244 | if (dr == dr0) | |
1245 | continue; | |
1246 | ||
1247 | stmt = DR_STMT (dr); | |
1248 | stmt_info = vinfo_for_stmt (stmt); | |
1249 | /* For interleaving, only the alignment of the first access | |
1250 | matters. */ | |
1251 | if (STMT_VINFO_STRIDED_ACCESS (stmt_info) | |
1252 | && DR_GROUP_FIRST_DR (stmt_info) != stmt) | |
1253 | continue; | |
1254 | ||
1255 | save_misalignment = DR_MISALIGNMENT (dr); | |
1256 | vect_update_misalignment_for_peel (dr, dr0, npeel); | |
1257 | supportable_dr_alignment = vect_supportable_dr_alignment (dr); | |
1258 | SET_DR_MISALIGNMENT (dr, save_misalignment); | |
b8698a0f | 1259 | |
ebfd146a IR |
1260 | if (!supportable_dr_alignment) |
1261 | { | |
1262 | do_peeling = false; | |
1263 | break; | |
1264 | } | |
1265 | } | |
1266 | ||
1267 | if (do_peeling) | |
1268 | { | |
1269 | /* (1.2) Update the DR_MISALIGNMENT of each data reference DR_i. | |
1270 | If the misalignment of DR_i is identical to that of dr0 then set | |
1271 | DR_MISALIGNMENT (DR_i) to zero. If the misalignment of DR_i and | |
1272 | dr0 are known at compile time then increment DR_MISALIGNMENT (DR_i) | |
1273 | by the peeling factor times the element size of DR_i (MOD the | |
1274 | vectorization factor times the size). Otherwise, the | |
1275 | misalignment of DR_i must be set to unknown. */ | |
1276 | for (i = 0; VEC_iterate (data_reference_p, datarefs, i, dr); i++) | |
1277 | if (dr != dr0) | |
1278 | vect_update_misalignment_for_peel (dr, dr0, npeel); | |
1279 | ||
1280 | LOOP_VINFO_UNALIGNED_DR (loop_vinfo) = dr0; | |
1281 | LOOP_PEELING_FOR_ALIGNMENT (loop_vinfo) = DR_MISALIGNMENT (dr0); | |
1282 | SET_DR_MISALIGNMENT (dr0, 0); | |
1283 | if (vect_print_dump_info (REPORT_ALIGNMENT)) | |
1284 | fprintf (vect_dump, "Alignment of access forced using peeling."); | |
1285 | ||
1286 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1287 | fprintf (vect_dump, "Peeling for alignment will be applied."); | |
1288 | ||
a70d6342 | 1289 | stat = vect_verify_datarefs_alignment (loop_vinfo, NULL); |
ebfd146a IR |
1290 | gcc_assert (stat); |
1291 | return stat; | |
1292 | } | |
1293 | } | |
1294 | ||
1295 | ||
1296 | /* (2) Versioning to force alignment. */ | |
1297 | ||
1298 | /* Try versioning if: | |
1299 | 1) flag_tree_vect_loop_version is TRUE | |
1300 | 2) optimize loop for speed | |
1301 | 3) there is at least one unsupported misaligned data ref with an unknown | |
1302 | misalignment, and | |
1303 | 4) all misaligned data refs with a known misalignment are supported, and | |
1304 | 5) the number of runtime alignment checks is within reason. */ | |
1305 | ||
b8698a0f L |
1306 | do_versioning = |
1307 | flag_tree_vect_loop_version | |
ebfd146a IR |
1308 | && optimize_loop_nest_for_speed_p (loop) |
1309 | && (!loop->inner); /* FORNOW */ | |
1310 | ||
1311 | if (do_versioning) | |
1312 | { | |
1313 | for (i = 0; VEC_iterate (data_reference_p, datarefs, i, dr); i++) | |
1314 | { | |
1315 | stmt = DR_STMT (dr); | |
1316 | stmt_info = vinfo_for_stmt (stmt); | |
1317 | ||
1318 | /* For interleaving, only the alignment of the first access | |
1319 | matters. */ | |
1320 | if (aligned_access_p (dr) | |
1321 | || (STMT_VINFO_STRIDED_ACCESS (stmt_info) | |
1322 | && DR_GROUP_FIRST_DR (stmt_info) != stmt)) | |
1323 | continue; | |
1324 | ||
1325 | supportable_dr_alignment = vect_supportable_dr_alignment (dr); | |
1326 | ||
1327 | if (!supportable_dr_alignment) | |
1328 | { | |
1329 | gimple stmt; | |
1330 | int mask; | |
1331 | tree vectype; | |
1332 | ||
1333 | if (known_alignment_for_access_p (dr) | |
1334 | || VEC_length (gimple, | |
1335 | LOOP_VINFO_MAY_MISALIGN_STMTS (loop_vinfo)) | |
1336 | >= (unsigned) PARAM_VALUE (PARAM_VECT_MAX_VERSION_FOR_ALIGNMENT_CHECKS)) | |
1337 | { | |
1338 | do_versioning = false; | |
1339 | break; | |
1340 | } | |
1341 | ||
1342 | stmt = DR_STMT (dr); | |
1343 | vectype = STMT_VINFO_VECTYPE (vinfo_for_stmt (stmt)); | |
1344 | gcc_assert (vectype); | |
b8698a0f | 1345 | |
ebfd146a IR |
1346 | /* The rightmost bits of an aligned address must be zeros. |
1347 | Construct the mask needed for this test. For example, | |
1348 | GET_MODE_SIZE for the vector mode V4SI is 16 bytes so the | |
1349 | mask must be 15 = 0xf. */ | |
1350 | mask = GET_MODE_SIZE (TYPE_MODE (vectype)) - 1; | |
1351 | ||
1352 | /* FORNOW: use the same mask to test all potentially unaligned | |
1353 | references in the loop. The vectorizer currently supports | |
1354 | a single vector size, see the reference to | |
1355 | GET_MODE_NUNITS (TYPE_MODE (vectype)) where the | |
1356 | vectorization factor is computed. */ | |
1357 | gcc_assert (!LOOP_VINFO_PTR_MASK (loop_vinfo) | |
1358 | || LOOP_VINFO_PTR_MASK (loop_vinfo) == mask); | |
1359 | LOOP_VINFO_PTR_MASK (loop_vinfo) = mask; | |
1360 | VEC_safe_push (gimple, heap, | |
1361 | LOOP_VINFO_MAY_MISALIGN_STMTS (loop_vinfo), | |
1362 | DR_STMT (dr)); | |
1363 | } | |
1364 | } | |
b8698a0f | 1365 | |
ebfd146a | 1366 | /* Versioning requires at least one misaligned data reference. */ |
e9dbe7bb | 1367 | if (!LOOP_REQUIRES_VERSIONING_FOR_ALIGNMENT (loop_vinfo)) |
ebfd146a IR |
1368 | do_versioning = false; |
1369 | else if (!do_versioning) | |
1370 | VEC_truncate (gimple, LOOP_VINFO_MAY_MISALIGN_STMTS (loop_vinfo), 0); | |
1371 | } | |
1372 | ||
1373 | if (do_versioning) | |
1374 | { | |
1375 | VEC(gimple,heap) *may_misalign_stmts | |
1376 | = LOOP_VINFO_MAY_MISALIGN_STMTS (loop_vinfo); | |
1377 | gimple stmt; | |
1378 | ||
1379 | /* It can now be assumed that the data references in the statements | |
1380 | in LOOP_VINFO_MAY_MISALIGN_STMTS will be aligned in the version | |
1381 | of the loop being vectorized. */ | |
1382 | for (i = 0; VEC_iterate (gimple, may_misalign_stmts, i, stmt); i++) | |
1383 | { | |
1384 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
1385 | dr = STMT_VINFO_DATA_REF (stmt_info); | |
1386 | SET_DR_MISALIGNMENT (dr, 0); | |
1387 | if (vect_print_dump_info (REPORT_ALIGNMENT)) | |
1388 | fprintf (vect_dump, "Alignment of access forced using versioning."); | |
1389 | } | |
1390 | ||
1391 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1392 | fprintf (vect_dump, "Versioning for alignment will be applied."); | |
1393 | ||
1394 | /* Peeling and versioning can't be done together at this time. */ | |
1395 | gcc_assert (! (do_peeling && do_versioning)); | |
1396 | ||
a70d6342 | 1397 | stat = vect_verify_datarefs_alignment (loop_vinfo, NULL); |
ebfd146a IR |
1398 | gcc_assert (stat); |
1399 | return stat; | |
1400 | } | |
1401 | ||
1402 | /* This point is reached if neither peeling nor versioning is being done. */ | |
1403 | gcc_assert (! (do_peeling || do_versioning)); | |
1404 | ||
a70d6342 | 1405 | stat = vect_verify_datarefs_alignment (loop_vinfo, NULL); |
ebfd146a IR |
1406 | return stat; |
1407 | } | |
1408 | ||
1409 | ||
1410 | /* Function vect_analyze_data_refs_alignment | |
1411 | ||
1412 | Analyze the alignment of the data-references in the loop. | |
1413 | Return FALSE if a data reference is found that cannot be vectorized. */ | |
1414 | ||
1415 | bool | |
b8698a0f | 1416 | vect_analyze_data_refs_alignment (loop_vec_info loop_vinfo, |
a70d6342 | 1417 | bb_vec_info bb_vinfo) |
ebfd146a IR |
1418 | { |
1419 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1420 | fprintf (vect_dump, "=== vect_analyze_data_refs_alignment ==="); | |
1421 | ||
a70d6342 | 1422 | if (!vect_compute_data_refs_alignment (loop_vinfo, bb_vinfo)) |
ebfd146a | 1423 | { |
8644a673 | 1424 | if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS)) |
b8698a0f | 1425 | fprintf (vect_dump, |
ebfd146a IR |
1426 | "not vectorized: can't calculate alignment for data ref."); |
1427 | return false; | |
1428 | } | |
1429 | ||
1430 | return true; | |
1431 | } | |
1432 | ||
1433 | ||
1434 | /* Analyze groups of strided accesses: check that DR belongs to a group of | |
1435 | strided accesses of legal size, step, etc. Detect gaps, single element | |
1436 | interleaving, and other special cases. Set strided access info. | |
1437 | Collect groups of strided stores for further use in SLP analysis. */ | |
1438 | ||
1439 | static bool | |
1440 | vect_analyze_group_access (struct data_reference *dr) | |
1441 | { | |
1442 | tree step = DR_STEP (dr); | |
1443 | tree scalar_type = TREE_TYPE (DR_REF (dr)); | |
1444 | HOST_WIDE_INT type_size = TREE_INT_CST_LOW (TYPE_SIZE_UNIT (scalar_type)); | |
1445 | gimple stmt = DR_STMT (dr); | |
1446 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
1447 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
a70d6342 | 1448 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
ebfd146a IR |
1449 | HOST_WIDE_INT dr_step = TREE_INT_CST_LOW (step); |
1450 | HOST_WIDE_INT stride; | |
1451 | bool slp_impossible = false; | |
1452 | ||
b8698a0f | 1453 | /* For interleaving, STRIDE is STEP counted in elements, i.e., the size of the |
ebfd146a | 1454 | interleaving group (including gaps). */ |
b8698a0f | 1455 | stride = dr_step / type_size; |
ebfd146a IR |
1456 | |
1457 | /* Not consecutive access is possible only if it is a part of interleaving. */ | |
1458 | if (!DR_GROUP_FIRST_DR (vinfo_for_stmt (stmt))) | |
1459 | { | |
1460 | /* Check if it this DR is a part of interleaving, and is a single | |
1461 | element of the group that is accessed in the loop. */ | |
b8698a0f | 1462 | |
ebfd146a IR |
1463 | /* Gaps are supported only for loads. STEP must be a multiple of the type |
1464 | size. The size of the group must be a power of 2. */ | |
1465 | if (DR_IS_READ (dr) | |
1466 | && (dr_step % type_size) == 0 | |
1467 | && stride > 0 | |
1468 | && exact_log2 (stride) != -1) | |
1469 | { | |
1470 | DR_GROUP_FIRST_DR (vinfo_for_stmt (stmt)) = stmt; | |
1471 | DR_GROUP_SIZE (vinfo_for_stmt (stmt)) = stride; | |
1472 | if (vect_print_dump_info (REPORT_DR_DETAILS)) | |
1473 | { | |
a70d6342 | 1474 | fprintf (vect_dump, "Detected single element interleaving "); |
ebfd146a IR |
1475 | print_generic_expr (vect_dump, DR_REF (dr), TDF_SLIM); |
1476 | fprintf (vect_dump, " step "); | |
1477 | print_generic_expr (vect_dump, step, TDF_SLIM); | |
1478 | } | |
1479 | return true; | |
1480 | } | |
1481 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1482 | fprintf (vect_dump, "not consecutive access"); | |
1483 | return false; | |
1484 | } | |
1485 | ||
1486 | if (DR_GROUP_FIRST_DR (vinfo_for_stmt (stmt)) == stmt) | |
1487 | { | |
1488 | /* First stmt in the interleaving chain. Check the chain. */ | |
1489 | gimple next = DR_GROUP_NEXT_DR (vinfo_for_stmt (stmt)); | |
1490 | struct data_reference *data_ref = dr; | |
df398a37 | 1491 | unsigned int count = 1; |
ebfd146a IR |
1492 | tree next_step; |
1493 | tree prev_init = DR_INIT (data_ref); | |
1494 | gimple prev = stmt; | |
df398a37 | 1495 | HOST_WIDE_INT diff, count_in_bytes, gaps = 0; |
ebfd146a IR |
1496 | |
1497 | while (next) | |
1498 | { | |
1499 | /* Skip same data-refs. In case that two or more stmts share data-ref | |
1500 | (supported only for loads), we vectorize only the first stmt, and | |
1501 | the rest get their vectorized loads from the first one. */ | |
1502 | if (!tree_int_cst_compare (DR_INIT (data_ref), | |
1503 | DR_INIT (STMT_VINFO_DATA_REF ( | |
1504 | vinfo_for_stmt (next))))) | |
1505 | { | |
1506 | if (!DR_IS_READ (data_ref)) | |
1507 | { | |
1508 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1509 | fprintf (vect_dump, "Two store stmts share the same dr."); | |
1510 | return false; | |
1511 | } | |
1512 | ||
1513 | /* Check that there is no load-store dependencies for this loads | |
1514 | to prevent a case of load-store-load to the same location. */ | |
1515 | if (DR_GROUP_READ_WRITE_DEPENDENCE (vinfo_for_stmt (next)) | |
1516 | || DR_GROUP_READ_WRITE_DEPENDENCE (vinfo_for_stmt (prev))) | |
1517 | { | |
1518 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1519 | fprintf (vect_dump, | |
1520 | "READ_WRITE dependence in interleaving."); | |
1521 | return false; | |
1522 | } | |
1523 | ||
1524 | /* For load use the same data-ref load. */ | |
1525 | DR_GROUP_SAME_DR_STMT (vinfo_for_stmt (next)) = prev; | |
1526 | ||
1527 | prev = next; | |
1528 | next = DR_GROUP_NEXT_DR (vinfo_for_stmt (next)); | |
1529 | continue; | |
1530 | } | |
1531 | prev = next; | |
1532 | ||
1533 | /* Check that all the accesses have the same STEP. */ | |
1534 | next_step = DR_STEP (STMT_VINFO_DATA_REF (vinfo_for_stmt (next))); | |
1535 | if (tree_int_cst_compare (step, next_step)) | |
1536 | { | |
1537 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1538 | fprintf (vect_dump, "not consecutive access in interleaving"); | |
1539 | return false; | |
1540 | } | |
1541 | ||
1542 | data_ref = STMT_VINFO_DATA_REF (vinfo_for_stmt (next)); | |
1543 | /* Check that the distance between two accesses is equal to the type | |
1544 | size. Otherwise, we have gaps. */ | |
1545 | diff = (TREE_INT_CST_LOW (DR_INIT (data_ref)) | |
1546 | - TREE_INT_CST_LOW (prev_init)) / type_size; | |
1547 | if (diff != 1) | |
1548 | { | |
1549 | /* FORNOW: SLP of accesses with gaps is not supported. */ | |
1550 | slp_impossible = true; | |
1551 | if (!DR_IS_READ (data_ref)) | |
1552 | { | |
1553 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1554 | fprintf (vect_dump, "interleaved store with gaps"); | |
1555 | return false; | |
1556 | } | |
4da39468 IR |
1557 | |
1558 | gaps += diff - 1; | |
ebfd146a IR |
1559 | } |
1560 | ||
1561 | /* Store the gap from the previous member of the group. If there is no | |
1562 | gap in the access, DR_GROUP_GAP is always 1. */ | |
1563 | DR_GROUP_GAP (vinfo_for_stmt (next)) = diff; | |
1564 | ||
1565 | prev_init = DR_INIT (data_ref); | |
1566 | next = DR_GROUP_NEXT_DR (vinfo_for_stmt (next)); | |
1567 | /* Count the number of data-refs in the chain. */ | |
1568 | count++; | |
1569 | } | |
1570 | ||
1571 | /* COUNT is the number of accesses found, we multiply it by the size of | |
1572 | the type to get COUNT_IN_BYTES. */ | |
1573 | count_in_bytes = type_size * count; | |
1574 | ||
b8698a0f | 1575 | /* Check that the size of the interleaving (including gaps) is not |
a70d6342 | 1576 | greater than STEP. */ |
4da39468 | 1577 | if (dr_step && dr_step < count_in_bytes + gaps * type_size) |
ebfd146a IR |
1578 | { |
1579 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1580 | { | |
1581 | fprintf (vect_dump, "interleaving size is greater than step for "); | |
1582 | print_generic_expr (vect_dump, DR_REF (dr), TDF_SLIM); | |
1583 | } | |
1584 | return false; | |
1585 | } | |
1586 | ||
1587 | /* Check that the size of the interleaving is equal to STEP for stores, | |
1588 | i.e., that there are no gaps. */ | |
a70d6342 | 1589 | if (dr_step && dr_step != count_in_bytes) |
ebfd146a IR |
1590 | { |
1591 | if (DR_IS_READ (dr)) | |
1592 | { | |
1593 | slp_impossible = true; | |
1594 | /* There is a gap after the last load in the group. This gap is a | |
b8698a0f L |
1595 | difference between the stride and the number of elements. When |
1596 | there is no gap, this difference should be 0. */ | |
1597 | DR_GROUP_GAP (vinfo_for_stmt (stmt)) = stride - count; | |
ebfd146a IR |
1598 | } |
1599 | else | |
1600 | { | |
1601 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1602 | fprintf (vect_dump, "interleaved store with gaps"); | |
1603 | return false; | |
1604 | } | |
1605 | } | |
1606 | ||
1607 | /* Check that STEP is a multiple of type size. */ | |
a70d6342 | 1608 | if (dr_step && (dr_step % type_size) != 0) |
ebfd146a IR |
1609 | { |
1610 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1611 | { | |
1612 | fprintf (vect_dump, "step is not a multiple of type size: step "); | |
1613 | print_generic_expr (vect_dump, step, TDF_SLIM); | |
1614 | fprintf (vect_dump, " size "); | |
1615 | print_generic_expr (vect_dump, TYPE_SIZE_UNIT (scalar_type), | |
1616 | TDF_SLIM); | |
1617 | } | |
1618 | return false; | |
1619 | } | |
1620 | ||
b8698a0f | 1621 | /* FORNOW: we handle only interleaving that is a power of 2. |
ebfd146a IR |
1622 | We don't fail here if it may be still possible to vectorize the |
1623 | group using SLP. If not, the size of the group will be checked in | |
1624 | vect_analyze_operations, and the vectorization will fail. */ | |
1625 | if (exact_log2 (stride) == -1) | |
1626 | { | |
1627 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1628 | fprintf (vect_dump, "interleaving is not a power of 2"); | |
1629 | ||
1630 | if (slp_impossible) | |
1631 | return false; | |
1632 | } | |
a70d6342 IR |
1633 | |
1634 | if (stride == 0) | |
1635 | stride = count; | |
b8698a0f | 1636 | |
ebfd146a IR |
1637 | DR_GROUP_SIZE (vinfo_for_stmt (stmt)) = stride; |
1638 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1639 | fprintf (vect_dump, "Detected interleaving of size %d", (int)stride); | |
1640 | ||
b8698a0f | 1641 | /* SLP: create an SLP data structure for every interleaving group of |
ebfd146a IR |
1642 | stores for further analysis in vect_analyse_slp. */ |
1643 | if (!DR_IS_READ (dr) && !slp_impossible) | |
a70d6342 IR |
1644 | { |
1645 | if (loop_vinfo) | |
1646 | VEC_safe_push (gimple, heap, LOOP_VINFO_STRIDED_STORES (loop_vinfo), | |
1647 | stmt); | |
1648 | if (bb_vinfo) | |
b8698a0f | 1649 | VEC_safe_push (gimple, heap, BB_VINFO_STRIDED_STORES (bb_vinfo), |
a70d6342 IR |
1650 | stmt); |
1651 | } | |
ebfd146a IR |
1652 | } |
1653 | ||
1654 | return true; | |
1655 | } | |
1656 | ||
1657 | ||
1658 | /* Analyze the access pattern of the data-reference DR. | |
1659 | In case of non-consecutive accesses call vect_analyze_group_access() to | |
1660 | analyze groups of strided accesses. */ | |
1661 | ||
1662 | static bool | |
1663 | vect_analyze_data_ref_access (struct data_reference *dr) | |
1664 | { | |
1665 | tree step = DR_STEP (dr); | |
1666 | tree scalar_type = TREE_TYPE (DR_REF (dr)); | |
1667 | gimple stmt = DR_STMT (dr); | |
1668 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
1669 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
a70d6342 | 1670 | struct loop *loop = NULL; |
ebfd146a IR |
1671 | HOST_WIDE_INT dr_step = TREE_INT_CST_LOW (step); |
1672 | ||
a70d6342 IR |
1673 | if (loop_vinfo) |
1674 | loop = LOOP_VINFO_LOOP (loop_vinfo); | |
b8698a0f | 1675 | |
a70d6342 | 1676 | if (loop_vinfo && !step) |
ebfd146a IR |
1677 | { |
1678 | if (vect_print_dump_info (REPORT_DETAILS)) | |
a70d6342 | 1679 | fprintf (vect_dump, "bad data-ref access in loop"); |
ebfd146a IR |
1680 | return false; |
1681 | } | |
1682 | ||
a70d6342 IR |
1683 | /* Don't allow invariant accesses in loops. */ |
1684 | if (loop_vinfo && dr_step == 0) | |
b8698a0f | 1685 | return false; |
ebfd146a | 1686 | |
a70d6342 | 1687 | if (loop && nested_in_vect_loop_p (loop, stmt)) |
ebfd146a IR |
1688 | { |
1689 | /* Interleaved accesses are not yet supported within outer-loop | |
1690 | vectorization for references in the inner-loop. */ | |
1691 | DR_GROUP_FIRST_DR (vinfo_for_stmt (stmt)) = NULL; | |
1692 | ||
1693 | /* For the rest of the analysis we use the outer-loop step. */ | |
1694 | step = STMT_VINFO_DR_STEP (stmt_info); | |
1695 | dr_step = TREE_INT_CST_LOW (step); | |
b8698a0f | 1696 | |
ebfd146a IR |
1697 | if (dr_step == 0) |
1698 | { | |
1699 | if (vect_print_dump_info (REPORT_ALIGNMENT)) | |
1700 | fprintf (vect_dump, "zero step in outer loop."); | |
1701 | if (DR_IS_READ (dr)) | |
b8698a0f | 1702 | return true; |
ebfd146a IR |
1703 | else |
1704 | return false; | |
1705 | } | |
1706 | } | |
1707 | ||
1708 | /* Consecutive? */ | |
1709 | if (!tree_int_cst_compare (step, TYPE_SIZE_UNIT (scalar_type))) | |
1710 | { | |
1711 | /* Mark that it is not interleaving. */ | |
1712 | DR_GROUP_FIRST_DR (vinfo_for_stmt (stmt)) = NULL; | |
1713 | return true; | |
1714 | } | |
1715 | ||
a70d6342 | 1716 | if (loop && nested_in_vect_loop_p (loop, stmt)) |
ebfd146a IR |
1717 | { |
1718 | if (vect_print_dump_info (REPORT_ALIGNMENT)) | |
1719 | fprintf (vect_dump, "strided access in outer loop."); | |
1720 | return false; | |
1721 | } | |
1722 | ||
1723 | /* Not consecutive access - check if it's a part of interleaving group. */ | |
1724 | return vect_analyze_group_access (dr); | |
1725 | } | |
1726 | ||
1727 | ||
1728 | /* Function vect_analyze_data_ref_accesses. | |
1729 | ||
1730 | Analyze the access pattern of all the data references in the loop. | |
1731 | ||
1732 | FORNOW: the only access pattern that is considered vectorizable is a | |
1733 | simple step 1 (consecutive) access. | |
1734 | ||
1735 | FORNOW: handle only arrays and pointer accesses. */ | |
1736 | ||
1737 | bool | |
a70d6342 | 1738 | vect_analyze_data_ref_accesses (loop_vec_info loop_vinfo, bb_vec_info bb_vinfo) |
ebfd146a IR |
1739 | { |
1740 | unsigned int i; | |
a70d6342 | 1741 | VEC (data_reference_p, heap) *datarefs; |
ebfd146a IR |
1742 | struct data_reference *dr; |
1743 | ||
1744 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1745 | fprintf (vect_dump, "=== vect_analyze_data_ref_accesses ==="); | |
1746 | ||
a70d6342 IR |
1747 | if (loop_vinfo) |
1748 | datarefs = LOOP_VINFO_DATAREFS (loop_vinfo); | |
1749 | else | |
1750 | datarefs = BB_VINFO_DATAREFS (bb_vinfo); | |
1751 | ||
ebfd146a IR |
1752 | for (i = 0; VEC_iterate (data_reference_p, datarefs, i, dr); i++) |
1753 | if (!vect_analyze_data_ref_access (dr)) | |
1754 | { | |
8644a673 | 1755 | if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS)) |
ebfd146a IR |
1756 | fprintf (vect_dump, "not vectorized: complicated access pattern."); |
1757 | return false; | |
1758 | } | |
1759 | ||
1760 | return true; | |
1761 | } | |
1762 | ||
1763 | /* Function vect_prune_runtime_alias_test_list. | |
1764 | ||
1765 | Prune a list of ddrs to be tested at run-time by versioning for alias. | |
1766 | Return FALSE if resulting list of ddrs is longer then allowed by | |
1767 | PARAM_VECT_MAX_VERSION_FOR_ALIAS_CHECKS, otherwise return TRUE. */ | |
1768 | ||
1769 | bool | |
1770 | vect_prune_runtime_alias_test_list (loop_vec_info loop_vinfo) | |
1771 | { | |
1772 | VEC (ddr_p, heap) * ddrs = | |
1773 | LOOP_VINFO_MAY_ALIAS_DDRS (loop_vinfo); | |
1774 | unsigned i, j; | |
1775 | ||
1776 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1777 | fprintf (vect_dump, "=== vect_prune_runtime_alias_test_list ==="); | |
1778 | ||
1779 | for (i = 0; i < VEC_length (ddr_p, ddrs); ) | |
1780 | { | |
1781 | bool found; | |
1782 | ddr_p ddr_i; | |
1783 | ||
1784 | ddr_i = VEC_index (ddr_p, ddrs, i); | |
1785 | found = false; | |
1786 | ||
1787 | for (j = 0; j < i; j++) | |
1788 | { | |
1789 | ddr_p ddr_j = VEC_index (ddr_p, ddrs, j); | |
1790 | ||
1791 | if (vect_vfa_range_equal (ddr_i, ddr_j)) | |
1792 | { | |
1793 | if (vect_print_dump_info (REPORT_DR_DETAILS)) | |
1794 | { | |
1795 | fprintf (vect_dump, "found equal ranges "); | |
1796 | print_generic_expr (vect_dump, DR_REF (DDR_A (ddr_i)), TDF_SLIM); | |
1797 | fprintf (vect_dump, ", "); | |
1798 | print_generic_expr (vect_dump, DR_REF (DDR_B (ddr_i)), TDF_SLIM); | |
1799 | fprintf (vect_dump, " and "); | |
1800 | print_generic_expr (vect_dump, DR_REF (DDR_A (ddr_j)), TDF_SLIM); | |
1801 | fprintf (vect_dump, ", "); | |
1802 | print_generic_expr (vect_dump, DR_REF (DDR_B (ddr_j)), TDF_SLIM); | |
1803 | } | |
1804 | found = true; | |
1805 | break; | |
1806 | } | |
1807 | } | |
b8698a0f | 1808 | |
ebfd146a IR |
1809 | if (found) |
1810 | { | |
1811 | VEC_ordered_remove (ddr_p, ddrs, i); | |
1812 | continue; | |
1813 | } | |
1814 | i++; | |
1815 | } | |
1816 | ||
1817 | if (VEC_length (ddr_p, ddrs) > | |
1818 | (unsigned) PARAM_VALUE (PARAM_VECT_MAX_VERSION_FOR_ALIAS_CHECKS)) | |
1819 | { | |
1820 | if (vect_print_dump_info (REPORT_DR_DETAILS)) | |
1821 | { | |
1822 | fprintf (vect_dump, | |
1823 | "disable versioning for alias - max number of generated " | |
1824 | "checks exceeded."); | |
1825 | } | |
1826 | ||
1827 | VEC_truncate (ddr_p, LOOP_VINFO_MAY_ALIAS_DDRS (loop_vinfo), 0); | |
1828 | ||
1829 | return false; | |
1830 | } | |
1831 | ||
1832 | return true; | |
1833 | } | |
1834 | ||
1835 | ||
1836 | /* Function vect_analyze_data_refs. | |
1837 | ||
a70d6342 | 1838 | Find all the data references in the loop or basic block. |
ebfd146a IR |
1839 | |
1840 | The general structure of the analysis of data refs in the vectorizer is as | |
1841 | follows: | |
b8698a0f | 1842 | 1- vect_analyze_data_refs(loop/bb): call |
a70d6342 IR |
1843 | compute_data_dependences_for_loop/bb to find and analyze all data-refs |
1844 | in the loop/bb and their dependences. | |
ebfd146a IR |
1845 | 2- vect_analyze_dependences(): apply dependence testing using ddrs. |
1846 | 3- vect_analyze_drs_alignment(): check that ref_stmt.alignment is ok. | |
1847 | 4- vect_analyze_drs_access(): check that ref_stmt.step is ok. | |
1848 | ||
1849 | */ | |
1850 | ||
1851 | bool | |
b8698a0f | 1852 | vect_analyze_data_refs (loop_vec_info loop_vinfo, bb_vec_info bb_vinfo) |
ebfd146a | 1853 | { |
a70d6342 IR |
1854 | struct loop *loop = NULL; |
1855 | basic_block bb = NULL; | |
ebfd146a IR |
1856 | unsigned int i; |
1857 | VEC (data_reference_p, heap) *datarefs; | |
1858 | struct data_reference *dr; | |
1859 | tree scalar_type; | |
1860 | ||
1861 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1862 | fprintf (vect_dump, "=== vect_analyze_data_refs ===\n"); | |
b8698a0f | 1863 | |
a70d6342 IR |
1864 | if (loop_vinfo) |
1865 | { | |
1866 | loop = LOOP_VINFO_LOOP (loop_vinfo); | |
1867 | compute_data_dependences_for_loop (loop, true, | |
1868 | &LOOP_VINFO_DATAREFS (loop_vinfo), | |
1869 | &LOOP_VINFO_DDRS (loop_vinfo)); | |
1870 | datarefs = LOOP_VINFO_DATAREFS (loop_vinfo); | |
1871 | } | |
1872 | else | |
1873 | { | |
1874 | bb = BB_VINFO_BB (bb_vinfo); | |
1875 | compute_data_dependences_for_bb (bb, true, | |
1876 | &BB_VINFO_DATAREFS (bb_vinfo), | |
1877 | &BB_VINFO_DDRS (bb_vinfo)); | |
1878 | datarefs = BB_VINFO_DATAREFS (bb_vinfo); | |
1879 | } | |
ebfd146a IR |
1880 | |
1881 | /* Go through the data-refs, check that the analysis succeeded. Update pointer | |
1882 | from stmt_vec_info struct to DR and vectype. */ | |
ebfd146a IR |
1883 | |
1884 | for (i = 0; VEC_iterate (data_reference_p, datarefs, i, dr); i++) | |
1885 | { | |
1886 | gimple stmt; | |
1887 | stmt_vec_info stmt_info; | |
b8698a0f L |
1888 | tree base, offset, init; |
1889 | ||
ebfd146a IR |
1890 | if (!dr || !DR_REF (dr)) |
1891 | { | |
8644a673 | 1892 | if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS)) |
ebfd146a IR |
1893 | fprintf (vect_dump, "not vectorized: unhandled data-ref "); |
1894 | return false; | |
1895 | } | |
1896 | ||
1897 | stmt = DR_STMT (dr); | |
1898 | stmt_info = vinfo_for_stmt (stmt); | |
1899 | ||
1900 | /* Check that analysis of the data-ref succeeded. */ | |
1901 | if (!DR_BASE_ADDRESS (dr) || !DR_OFFSET (dr) || !DR_INIT (dr) | |
1902 | || !DR_STEP (dr)) | |
1903 | { | |
8644a673 | 1904 | if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS)) |
ebfd146a IR |
1905 | { |
1906 | fprintf (vect_dump, "not vectorized: data ref analysis failed "); | |
1907 | print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM); | |
1908 | } | |
1909 | return false; | |
1910 | } | |
1911 | ||
1912 | if (TREE_CODE (DR_BASE_ADDRESS (dr)) == INTEGER_CST) | |
1913 | { | |
8644a673 | 1914 | if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS)) |
ebfd146a IR |
1915 | fprintf (vect_dump, "not vectorized: base addr of dr is a " |
1916 | "constant"); | |
1917 | return false; | |
1918 | } | |
1919 | ||
ebfd146a IR |
1920 | base = unshare_expr (DR_BASE_ADDRESS (dr)); |
1921 | offset = unshare_expr (DR_OFFSET (dr)); | |
1922 | init = unshare_expr (DR_INIT (dr)); | |
b8698a0f | 1923 | |
ebfd146a | 1924 | /* Update DR field in stmt_vec_info struct. */ |
ebfd146a IR |
1925 | |
1926 | /* If the dataref is in an inner-loop of the loop that is considered for | |
1927 | for vectorization, we also want to analyze the access relative to | |
b8698a0f | 1928 | the outer-loop (DR contains information only relative to the |
ebfd146a IR |
1929 | inner-most enclosing loop). We do that by building a reference to the |
1930 | first location accessed by the inner-loop, and analyze it relative to | |
b8698a0f L |
1931 | the outer-loop. */ |
1932 | if (loop && nested_in_vect_loop_p (loop, stmt)) | |
ebfd146a IR |
1933 | { |
1934 | tree outer_step, outer_base, outer_init; | |
1935 | HOST_WIDE_INT pbitsize, pbitpos; | |
1936 | tree poffset; | |
1937 | enum machine_mode pmode; | |
1938 | int punsignedp, pvolatilep; | |
1939 | affine_iv base_iv, offset_iv; | |
1940 | tree dinit; | |
1941 | ||
b8698a0f | 1942 | /* Build a reference to the first location accessed by the |
ebfd146a IR |
1943 | inner-loop: *(BASE+INIT). (The first location is actually |
1944 | BASE+INIT+OFFSET, but we add OFFSET separately later). */ | |
1945 | tree inner_base = build_fold_indirect_ref | |
1946 | (fold_build2 (POINTER_PLUS_EXPR, | |
b8698a0f | 1947 | TREE_TYPE (base), base, |
ebfd146a IR |
1948 | fold_convert (sizetype, init))); |
1949 | ||
1950 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1951 | { | |
1952 | fprintf (vect_dump, "analyze in outer-loop: "); | |
1953 | print_generic_expr (vect_dump, inner_base, TDF_SLIM); | |
1954 | } | |
1955 | ||
b8698a0f | 1956 | outer_base = get_inner_reference (inner_base, &pbitsize, &pbitpos, |
ebfd146a IR |
1957 | &poffset, &pmode, &punsignedp, &pvolatilep, false); |
1958 | gcc_assert (outer_base != NULL_TREE); | |
1959 | ||
1960 | if (pbitpos % BITS_PER_UNIT != 0) | |
1961 | { | |
1962 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1963 | fprintf (vect_dump, "failed: bit offset alignment.\n"); | |
1964 | return false; | |
1965 | } | |
1966 | ||
1967 | outer_base = build_fold_addr_expr (outer_base); | |
b8698a0f | 1968 | if (!simple_iv (loop, loop_containing_stmt (stmt), outer_base, |
ebfd146a IR |
1969 | &base_iv, false)) |
1970 | { | |
1971 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1972 | fprintf (vect_dump, "failed: evolution of base is not affine.\n"); | |
1973 | return false; | |
1974 | } | |
1975 | ||
1976 | if (offset) | |
1977 | { | |
1978 | if (poffset) | |
b8698a0f | 1979 | poffset = fold_build2 (PLUS_EXPR, TREE_TYPE (offset), offset, |
ebfd146a IR |
1980 | poffset); |
1981 | else | |
1982 | poffset = offset; | |
1983 | } | |
1984 | ||
1985 | if (!poffset) | |
1986 | { | |
1987 | offset_iv.base = ssize_int (0); | |
1988 | offset_iv.step = ssize_int (0); | |
1989 | } | |
b8698a0f | 1990 | else if (!simple_iv (loop, loop_containing_stmt (stmt), poffset, |
ebfd146a IR |
1991 | &offset_iv, false)) |
1992 | { | |
1993 | if (vect_print_dump_info (REPORT_DETAILS)) | |
1994 | fprintf (vect_dump, "evolution of offset is not affine.\n"); | |
1995 | return false; | |
1996 | } | |
1997 | ||
1998 | outer_init = ssize_int (pbitpos / BITS_PER_UNIT); | |
1999 | split_constant_offset (base_iv.base, &base_iv.base, &dinit); | |
2000 | outer_init = size_binop (PLUS_EXPR, outer_init, dinit); | |
2001 | split_constant_offset (offset_iv.base, &offset_iv.base, &dinit); | |
2002 | outer_init = size_binop (PLUS_EXPR, outer_init, dinit); | |
2003 | ||
2004 | outer_step = size_binop (PLUS_EXPR, | |
2005 | fold_convert (ssizetype, base_iv.step), | |
2006 | fold_convert (ssizetype, offset_iv.step)); | |
2007 | ||
2008 | STMT_VINFO_DR_STEP (stmt_info) = outer_step; | |
2009 | /* FIXME: Use canonicalize_base_object_address (base_iv.base); */ | |
b8698a0f | 2010 | STMT_VINFO_DR_BASE_ADDRESS (stmt_info) = base_iv.base; |
ebfd146a | 2011 | STMT_VINFO_DR_INIT (stmt_info) = outer_init; |
b8698a0f | 2012 | STMT_VINFO_DR_OFFSET (stmt_info) = |
ebfd146a | 2013 | fold_convert (ssizetype, offset_iv.base); |
b8698a0f | 2014 | STMT_VINFO_DR_ALIGNED_TO (stmt_info) = |
ebfd146a IR |
2015 | size_int (highest_pow2_factor (offset_iv.base)); |
2016 | ||
2017 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2018 | { | |
2019 | fprintf (vect_dump, "\touter base_address: "); | |
2020 | print_generic_expr (vect_dump, STMT_VINFO_DR_BASE_ADDRESS (stmt_info), TDF_SLIM); | |
2021 | fprintf (vect_dump, "\n\touter offset from base address: "); | |
2022 | print_generic_expr (vect_dump, STMT_VINFO_DR_OFFSET (stmt_info), TDF_SLIM); | |
2023 | fprintf (vect_dump, "\n\touter constant offset from base address: "); | |
2024 | print_generic_expr (vect_dump, STMT_VINFO_DR_INIT (stmt_info), TDF_SLIM); | |
2025 | fprintf (vect_dump, "\n\touter step: "); | |
2026 | print_generic_expr (vect_dump, STMT_VINFO_DR_STEP (stmt_info), TDF_SLIM); | |
2027 | fprintf (vect_dump, "\n\touter aligned to: "); | |
2028 | print_generic_expr (vect_dump, STMT_VINFO_DR_ALIGNED_TO (stmt_info), TDF_SLIM); | |
2029 | } | |
2030 | } | |
2031 | ||
2032 | if (STMT_VINFO_DATA_REF (stmt_info)) | |
2033 | { | |
8644a673 | 2034 | if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS)) |
ebfd146a IR |
2035 | { |
2036 | fprintf (vect_dump, | |
2037 | "not vectorized: more than one data ref in stmt: "); | |
2038 | print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM); | |
2039 | } | |
2040 | return false; | |
2041 | } | |
8644a673 | 2042 | |
ebfd146a | 2043 | STMT_VINFO_DATA_REF (stmt_info) = dr; |
b8698a0f | 2044 | |
ebfd146a IR |
2045 | /* Set vectype for STMT. */ |
2046 | scalar_type = TREE_TYPE (DR_REF (dr)); | |
2047 | STMT_VINFO_VECTYPE (stmt_info) = | |
2048 | get_vectype_for_scalar_type (scalar_type); | |
b8698a0f | 2049 | if (!STMT_VINFO_VECTYPE (stmt_info)) |
ebfd146a | 2050 | { |
8644a673 | 2051 | if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS)) |
ebfd146a IR |
2052 | { |
2053 | fprintf (vect_dump, | |
2054 | "not vectorized: no vectype for stmt: "); | |
2055 | print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM); | |
2056 | fprintf (vect_dump, " scalar_type: "); | |
2057 | print_generic_expr (vect_dump, scalar_type, TDF_DETAILS); | |
2058 | } | |
2059 | return false; | |
2060 | } | |
2061 | } | |
b8698a0f | 2062 | |
ebfd146a IR |
2063 | return true; |
2064 | } | |
2065 | ||
2066 | ||
2067 | /* Function vect_get_new_vect_var. | |
2068 | ||
b8698a0f L |
2069 | Returns a name for a new variable. The current naming scheme appends the |
2070 | prefix "vect_" or "vect_p" (depending on the value of VAR_KIND) to | |
2071 | the name of vectorizer generated variables, and appends that to NAME if | |
ebfd146a IR |
2072 | provided. */ |
2073 | ||
2074 | tree | |
2075 | vect_get_new_vect_var (tree type, enum vect_var_kind var_kind, const char *name) | |
2076 | { | |
2077 | const char *prefix; | |
2078 | tree new_vect_var; | |
2079 | ||
2080 | switch (var_kind) | |
2081 | { | |
2082 | case vect_simple_var: | |
2083 | prefix = "vect_"; | |
2084 | break; | |
2085 | case vect_scalar_var: | |
2086 | prefix = "stmp_"; | |
2087 | break; | |
2088 | case vect_pointer_var: | |
2089 | prefix = "vect_p"; | |
2090 | break; | |
2091 | default: | |
2092 | gcc_unreachable (); | |
2093 | } | |
2094 | ||
2095 | if (name) | |
2096 | { | |
2097 | char* tmp = concat (prefix, name, NULL); | |
2098 | new_vect_var = create_tmp_var (type, tmp); | |
2099 | free (tmp); | |
2100 | } | |
2101 | else | |
2102 | new_vect_var = create_tmp_var (type, prefix); | |
2103 | ||
2104 | /* Mark vector typed variable as a gimple register variable. */ | |
2105 | if (TREE_CODE (type) == VECTOR_TYPE) | |
2106 | DECL_GIMPLE_REG_P (new_vect_var) = true; | |
2107 | ||
2108 | return new_vect_var; | |
2109 | } | |
2110 | ||
2111 | ||
2112 | /* Function vect_create_addr_base_for_vector_ref. | |
2113 | ||
2114 | Create an expression that computes the address of the first memory location | |
2115 | that will be accessed for a data reference. | |
2116 | ||
2117 | Input: | |
2118 | STMT: The statement containing the data reference. | |
2119 | NEW_STMT_LIST: Must be initialized to NULL_TREE or a statement list. | |
2120 | OFFSET: Optional. If supplied, it is be added to the initial address. | |
2121 | LOOP: Specify relative to which loop-nest should the address be computed. | |
2122 | For example, when the dataref is in an inner-loop nested in an | |
2123 | outer-loop that is now being vectorized, LOOP can be either the | |
2124 | outer-loop, or the inner-loop. The first memory location accessed | |
2125 | by the following dataref ('in' points to short): | |
2126 | ||
2127 | for (i=0; i<N; i++) | |
2128 | for (j=0; j<M; j++) | |
2129 | s += in[i+j] | |
2130 | ||
2131 | is as follows: | |
2132 | if LOOP=i_loop: &in (relative to i_loop) | |
2133 | if LOOP=j_loop: &in+i*2B (relative to j_loop) | |
2134 | ||
2135 | Output: | |
b8698a0f | 2136 | 1. Return an SSA_NAME whose value is the address of the memory location of |
ebfd146a IR |
2137 | the first vector of the data reference. |
2138 | 2. If new_stmt_list is not NULL_TREE after return then the caller must insert | |
2139 | these statement(s) which define the returned SSA_NAME. | |
2140 | ||
2141 | FORNOW: We are only handling array accesses with step 1. */ | |
2142 | ||
2143 | tree | |
2144 | vect_create_addr_base_for_vector_ref (gimple stmt, | |
2145 | gimple_seq *new_stmt_list, | |
2146 | tree offset, | |
2147 | struct loop *loop) | |
2148 | { | |
2149 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
2150 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info); | |
ebfd146a IR |
2151 | tree data_ref_base = unshare_expr (DR_BASE_ADDRESS (dr)); |
2152 | tree base_name; | |
2153 | tree data_ref_base_var; | |
2154 | tree vec_stmt; | |
2155 | tree addr_base, addr_expr; | |
2156 | tree dest; | |
2157 | gimple_seq seq = NULL; | |
2158 | tree base_offset = unshare_expr (DR_OFFSET (dr)); | |
2159 | tree init = unshare_expr (DR_INIT (dr)); | |
8644a673 | 2160 | tree vect_ptr_type; |
ebfd146a | 2161 | tree step = TYPE_SIZE_UNIT (TREE_TYPE (DR_REF (dr))); |
a70d6342 | 2162 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
ebfd146a | 2163 | |
a70d6342 | 2164 | if (loop_vinfo && loop && loop != (gimple_bb (stmt))->loop_father) |
ebfd146a | 2165 | { |
a70d6342 | 2166 | struct loop *outer_loop = LOOP_VINFO_LOOP (loop_vinfo); |
ebfd146a | 2167 | |
a70d6342 | 2168 | gcc_assert (nested_in_vect_loop_p (outer_loop, stmt)); |
ebfd146a IR |
2169 | |
2170 | data_ref_base = unshare_expr (STMT_VINFO_DR_BASE_ADDRESS (stmt_info)); | |
2171 | base_offset = unshare_expr (STMT_VINFO_DR_OFFSET (stmt_info)); | |
2172 | init = unshare_expr (STMT_VINFO_DR_INIT (stmt_info)); | |
2173 | } | |
2174 | ||
a70d6342 IR |
2175 | if (loop_vinfo) |
2176 | base_name = build_fold_indirect_ref (data_ref_base); | |
2177 | else | |
2178 | { | |
2179 | base_offset = ssize_int (0); | |
2180 | init = ssize_int (0); | |
2181 | base_name = build_fold_indirect_ref (unshare_expr (DR_REF (dr))); | |
b8698a0f | 2182 | } |
a70d6342 | 2183 | |
ebfd146a IR |
2184 | data_ref_base_var = create_tmp_var (TREE_TYPE (data_ref_base), "batmp"); |
2185 | add_referenced_var (data_ref_base_var); | |
2186 | data_ref_base = force_gimple_operand (data_ref_base, &seq, true, | |
2187 | data_ref_base_var); | |
2188 | gimple_seq_add_seq (new_stmt_list, seq); | |
2189 | ||
2190 | /* Create base_offset */ | |
2191 | base_offset = size_binop (PLUS_EXPR, | |
2192 | fold_convert (sizetype, base_offset), | |
2193 | fold_convert (sizetype, init)); | |
2194 | dest = create_tmp_var (sizetype, "base_off"); | |
2195 | add_referenced_var (dest); | |
2196 | base_offset = force_gimple_operand (base_offset, &seq, true, dest); | |
2197 | gimple_seq_add_seq (new_stmt_list, seq); | |
2198 | ||
2199 | if (offset) | |
2200 | { | |
2201 | tree tmp = create_tmp_var (sizetype, "offset"); | |
2202 | ||
2203 | add_referenced_var (tmp); | |
2204 | offset = fold_build2 (MULT_EXPR, sizetype, | |
2205 | fold_convert (sizetype, offset), step); | |
2206 | base_offset = fold_build2 (PLUS_EXPR, sizetype, | |
2207 | base_offset, offset); | |
2208 | base_offset = force_gimple_operand (base_offset, &seq, false, tmp); | |
2209 | gimple_seq_add_seq (new_stmt_list, seq); | |
2210 | } | |
2211 | ||
2212 | /* base + base_offset */ | |
a70d6342 IR |
2213 | if (loop_vinfo) |
2214 | addr_base = fold_build2 (POINTER_PLUS_EXPR, TREE_TYPE (data_ref_base), | |
2215 | data_ref_base, base_offset); | |
2216 | else | |
2217 | { | |
2218 | if (TREE_CODE (DR_REF (dr)) == INDIRECT_REF) | |
2219 | addr_base = unshare_expr (TREE_OPERAND (DR_REF (dr), 0)); | |
2220 | else | |
b8698a0f | 2221 | addr_base = build1 (ADDR_EXPR, |
a70d6342 IR |
2222 | build_pointer_type (TREE_TYPE (DR_REF (dr))), |
2223 | unshare_expr (DR_REF (dr))); | |
2224 | } | |
b8698a0f | 2225 | |
ebfd146a IR |
2226 | vect_ptr_type = build_pointer_type (STMT_VINFO_VECTYPE (stmt_info)); |
2227 | ||
8644a673 | 2228 | vec_stmt = fold_convert (vect_ptr_type, addr_base); |
ebfd146a IR |
2229 | addr_expr = vect_get_new_vect_var (vect_ptr_type, vect_pointer_var, |
2230 | get_name (base_name)); | |
2231 | add_referenced_var (addr_expr); | |
8644a673 | 2232 | vec_stmt = force_gimple_operand (vec_stmt, &seq, false, addr_expr); |
ebfd146a IR |
2233 | gimple_seq_add_seq (new_stmt_list, seq); |
2234 | ||
2235 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2236 | { | |
2237 | fprintf (vect_dump, "created "); | |
2238 | print_generic_expr (vect_dump, vec_stmt, TDF_SLIM); | |
2239 | } | |
8644a673 | 2240 | |
ebfd146a IR |
2241 | return vec_stmt; |
2242 | } | |
2243 | ||
2244 | ||
2245 | /* Function vect_create_data_ref_ptr. | |
2246 | ||
2247 | Create a new pointer to vector type (vp), that points to the first location | |
b8698a0f | 2248 | accessed in the loop by STMT, along with the def-use update chain to |
ebfd146a IR |
2249 | appropriately advance the pointer through the loop iterations. Also set |
2250 | aliasing information for the pointer. This vector pointer is used by the | |
2251 | callers to this function to create a memory reference expression for vector | |
2252 | load/store access. | |
2253 | ||
2254 | Input: | |
2255 | 1. STMT: a stmt that references memory. Expected to be of the form | |
2256 | GIMPLE_ASSIGN <name, data-ref> or | |
2257 | GIMPLE_ASSIGN <data-ref, name>. | |
2258 | 2. AT_LOOP: the loop where the vector memref is to be created. | |
2259 | 3. OFFSET (optional): an offset to be added to the initial address accessed | |
2260 | by the data-ref in STMT. | |
2261 | 4. ONLY_INIT: indicate if vp is to be updated in the loop, or remain | |
2262 | pointing to the initial address. | |
2263 | 5. TYPE: if not NULL indicates the required type of the data-ref. | |
2264 | ||
2265 | Output: | |
2266 | 1. Declare a new ptr to vector_type, and have it point to the base of the | |
2267 | data reference (initial addressed accessed by the data reference). | |
2268 | For example, for vector of type V8HI, the following code is generated: | |
2269 | ||
2270 | v8hi *vp; | |
2271 | vp = (v8hi *)initial_address; | |
2272 | ||
2273 | if OFFSET is not supplied: | |
2274 | initial_address = &a[init]; | |
2275 | if OFFSET is supplied: | |
2276 | initial_address = &a[init + OFFSET]; | |
2277 | ||
2278 | Return the initial_address in INITIAL_ADDRESS. | |
2279 | ||
2280 | 2. If ONLY_INIT is true, just return the initial pointer. Otherwise, also | |
b8698a0f | 2281 | update the pointer in each iteration of the loop. |
ebfd146a IR |
2282 | |
2283 | Return the increment stmt that updates the pointer in PTR_INCR. | |
2284 | ||
b8698a0f | 2285 | 3. Set INV_P to true if the access pattern of the data reference in the |
ebfd146a IR |
2286 | vectorized loop is invariant. Set it to false otherwise. |
2287 | ||
2288 | 4. Return the pointer. */ | |
2289 | ||
2290 | tree | |
2291 | vect_create_data_ref_ptr (gimple stmt, struct loop *at_loop, | |
2292 | tree offset, tree *initial_address, gimple *ptr_incr, | |
5006671f | 2293 | bool only_init, bool *inv_p) |
ebfd146a IR |
2294 | { |
2295 | tree base_name; | |
2296 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
2297 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
a70d6342 IR |
2298 | struct loop *loop = NULL; |
2299 | bool nested_in_vect_loop = false; | |
2300 | struct loop *containing_loop = NULL; | |
ebfd146a IR |
2301 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); |
2302 | tree vect_ptr_type; | |
2303 | tree vect_ptr; | |
ebfd146a IR |
2304 | tree new_temp; |
2305 | gimple vec_stmt; | |
2306 | gimple_seq new_stmt_list = NULL; | |
a70d6342 | 2307 | edge pe = NULL; |
ebfd146a IR |
2308 | basic_block new_bb; |
2309 | tree vect_ptr_init; | |
2310 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info); | |
2311 | tree vptr; | |
2312 | gimple_stmt_iterator incr_gsi; | |
2313 | bool insert_after; | |
2314 | tree indx_before_incr, indx_after_incr; | |
2315 | gimple incr; | |
2316 | tree step; | |
a70d6342 IR |
2317 | bb_vec_info bb_vinfo = STMT_VINFO_BB_VINFO (stmt_info); |
2318 | gimple_stmt_iterator gsi = gsi_for_stmt (stmt); | |
b8698a0f | 2319 | |
a70d6342 IR |
2320 | if (loop_vinfo) |
2321 | { | |
2322 | loop = LOOP_VINFO_LOOP (loop_vinfo); | |
2323 | nested_in_vect_loop = nested_in_vect_loop_p (loop, stmt); | |
2324 | containing_loop = (gimple_bb (stmt))->loop_father; | |
2325 | pe = loop_preheader_edge (loop); | |
2326 | } | |
2327 | else | |
2328 | { | |
2329 | gcc_assert (bb_vinfo); | |
2330 | only_init = true; | |
2331 | *ptr_incr = NULL; | |
2332 | } | |
b8698a0f | 2333 | |
ebfd146a IR |
2334 | /* Check the step (evolution) of the load in LOOP, and record |
2335 | whether it's invariant. */ | |
2336 | if (nested_in_vect_loop) | |
2337 | step = STMT_VINFO_DR_STEP (stmt_info); | |
2338 | else | |
2339 | step = DR_STEP (STMT_VINFO_DATA_REF (stmt_info)); | |
b8698a0f | 2340 | |
ebfd146a IR |
2341 | if (tree_int_cst_compare (step, size_zero_node) == 0) |
2342 | *inv_p = true; | |
2343 | else | |
2344 | *inv_p = false; | |
2345 | ||
2346 | /* Create an expression for the first address accessed by this load | |
b8698a0f | 2347 | in LOOP. */ |
ebfd146a IR |
2348 | base_name = build_fold_indirect_ref (unshare_expr (DR_BASE_ADDRESS (dr))); |
2349 | ||
2350 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2351 | { | |
2352 | tree data_ref_base = base_name; | |
2353 | fprintf (vect_dump, "create vector-pointer variable to type: "); | |
2354 | print_generic_expr (vect_dump, vectype, TDF_SLIM); | |
b8698a0f | 2355 | if (TREE_CODE (data_ref_base) == VAR_DECL |
e9dbe7bb IR |
2356 | || TREE_CODE (data_ref_base) == ARRAY_REF) |
2357 | fprintf (vect_dump, " vectorizing an array ref: "); | |
ebfd146a IR |
2358 | else if (TREE_CODE (data_ref_base) == COMPONENT_REF) |
2359 | fprintf (vect_dump, " vectorizing a record based array ref: "); | |
2360 | else if (TREE_CODE (data_ref_base) == SSA_NAME) | |
2361 | fprintf (vect_dump, " vectorizing a pointer ref: "); | |
2362 | print_generic_expr (vect_dump, base_name, TDF_SLIM); | |
2363 | } | |
2364 | ||
2365 | /** (1) Create the new vector-pointer variable: **/ | |
5006671f | 2366 | vect_ptr_type = build_pointer_type (vectype); |
ebfd146a IR |
2367 | vect_ptr = vect_get_new_vect_var (vect_ptr_type, vect_pointer_var, |
2368 | get_name (base_name)); | |
3f49ba3f EB |
2369 | |
2370 | /* Vector types inherit the alias set of their component type by default so | |
2371 | we need to use a ref-all pointer if the data reference does not conflict | |
2372 | with the created vector data reference because it is not addressable. */ | |
2373 | if (!alias_sets_conflict_p (get_deref_alias_set (vect_ptr), | |
2374 | get_alias_set (DR_REF (dr)))) | |
2375 | { | |
d4ebfa65 BE |
2376 | vect_ptr_type |
2377 | = build_pointer_type_for_mode (vectype, | |
2378 | TYPE_MODE (vect_ptr_type), true); | |
3f49ba3f EB |
2379 | vect_ptr = vect_get_new_vect_var (vect_ptr_type, vect_pointer_var, |
2380 | get_name (base_name)); | |
2381 | } | |
2382 | ||
2383 | /* Likewise for any of the data references in the stmt group. */ | |
2384 | else if (STMT_VINFO_DR_GROUP_SIZE (stmt_info) > 1) | |
ebfd146a | 2385 | { |
5006671f RG |
2386 | gimple orig_stmt = STMT_VINFO_DR_GROUP_FIRST_DR (stmt_info); |
2387 | do | |
2388 | { | |
2389 | tree lhs = gimple_assign_lhs (orig_stmt); | |
2390 | if (!alias_sets_conflict_p (get_deref_alias_set (vect_ptr), | |
2391 | get_alias_set (lhs))) | |
2392 | { | |
3f49ba3f | 2393 | vect_ptr_type |
d4ebfa65 BE |
2394 | = build_pointer_type_for_mode (vectype, |
2395 | TYPE_MODE (vect_ptr_type), true); | |
3f49ba3f EB |
2396 | vect_ptr |
2397 | = vect_get_new_vect_var (vect_ptr_type, vect_pointer_var, | |
2398 | get_name (base_name)); | |
5006671f RG |
2399 | break; |
2400 | } | |
2401 | ||
2402 | orig_stmt = STMT_VINFO_DR_GROUP_NEXT_DR (vinfo_for_stmt (orig_stmt)); | |
2403 | } | |
2404 | while (orig_stmt); | |
ebfd146a IR |
2405 | } |
2406 | ||
2407 | add_referenced_var (vect_ptr); | |
2408 | ||
ebfd146a IR |
2409 | /** Note: If the dataref is in an inner-loop nested in LOOP, and we are |
2410 | vectorizing LOOP (i.e. outer-loop vectorization), we need to create two | |
2411 | def-use update cycles for the pointer: One relative to the outer-loop | |
2412 | (LOOP), which is what steps (3) and (4) below do. The other is relative | |
2413 | to the inner-loop (which is the inner-most loop containing the dataref), | |
b8698a0f | 2414 | and this is done be step (5) below. |
ebfd146a IR |
2415 | |
2416 | When vectorizing inner-most loops, the vectorized loop (LOOP) is also the | |
2417 | inner-most loop, and so steps (3),(4) work the same, and step (5) is | |
2418 | redundant. Steps (3),(4) create the following: | |
2419 | ||
2420 | vp0 = &base_addr; | |
2421 | LOOP: vp1 = phi(vp0,vp2) | |
b8698a0f | 2422 | ... |
ebfd146a IR |
2423 | ... |
2424 | vp2 = vp1 + step | |
2425 | goto LOOP | |
b8698a0f | 2426 | |
ebfd146a IR |
2427 | If there is an inner-loop nested in loop, then step (5) will also be |
2428 | applied, and an additional update in the inner-loop will be created: | |
2429 | ||
2430 | vp0 = &base_addr; | |
2431 | LOOP: vp1 = phi(vp0,vp2) | |
2432 | ... | |
2433 | inner: vp3 = phi(vp1,vp4) | |
2434 | vp4 = vp3 + inner_step | |
2435 | if () goto inner | |
2436 | ... | |
2437 | vp2 = vp1 + step | |
2438 | if () goto LOOP */ | |
2439 | ||
2440 | /** (3) Calculate the initial address the vector-pointer, and set | |
2441 | the vector-pointer to point to it before the loop: **/ | |
2442 | ||
2443 | /* Create: (&(base[init_val+offset]) in the loop preheader. */ | |
2444 | ||
2445 | new_temp = vect_create_addr_base_for_vector_ref (stmt, &new_stmt_list, | |
2446 | offset, loop); | |
ebfd146a IR |
2447 | if (new_stmt_list) |
2448 | { | |
a70d6342 IR |
2449 | if (pe) |
2450 | { | |
2451 | new_bb = gsi_insert_seq_on_edge_immediate (pe, new_stmt_list); | |
2452 | gcc_assert (!new_bb); | |
2453 | } | |
2454 | else | |
2455 | gsi_insert_seq_before (&gsi, new_stmt_list, GSI_SAME_STMT); | |
ebfd146a IR |
2456 | } |
2457 | ||
2458 | *initial_address = new_temp; | |
2459 | ||
2460 | /* Create: p = (vectype *) initial_base */ | |
2461 | vec_stmt = gimple_build_assign (vect_ptr, | |
2462 | fold_convert (vect_ptr_type, new_temp)); | |
2463 | vect_ptr_init = make_ssa_name (vect_ptr, vec_stmt); | |
2464 | gimple_assign_set_lhs (vec_stmt, vect_ptr_init); | |
a70d6342 IR |
2465 | if (pe) |
2466 | { | |
2467 | new_bb = gsi_insert_on_edge_immediate (pe, vec_stmt); | |
2468 | gcc_assert (!new_bb); | |
2469 | } | |
2470 | else | |
2471 | gsi_insert_before (&gsi, vec_stmt, GSI_SAME_STMT); | |
ebfd146a IR |
2472 | |
2473 | /** (4) Handle the updating of the vector-pointer inside the loop. | |
2474 | This is needed when ONLY_INIT is false, and also when AT_LOOP | |
2475 | is the inner-loop nested in LOOP (during outer-loop vectorization). | |
2476 | **/ | |
2477 | ||
a70d6342 | 2478 | /* No update in loop is required. */ |
b8698a0f | 2479 | if (only_init && (!loop_vinfo || at_loop == loop)) |
ebfd146a IR |
2480 | { |
2481 | /* Copy the points-to information if it exists. */ | |
2482 | if (DR_PTR_INFO (dr)) | |
2483 | duplicate_ssa_name_ptr_info (vect_ptr_init, DR_PTR_INFO (dr)); | |
2484 | vptr = vect_ptr_init; | |
2485 | } | |
2486 | else | |
2487 | { | |
2488 | /* The step of the vector pointer is the Vector Size. */ | |
2489 | tree step = TYPE_SIZE_UNIT (vectype); | |
b8698a0f | 2490 | /* One exception to the above is when the scalar step of the load in |
ebfd146a IR |
2491 | LOOP is zero. In this case the step here is also zero. */ |
2492 | if (*inv_p) | |
2493 | step = size_zero_node; | |
2494 | ||
2495 | standard_iv_increment_position (loop, &incr_gsi, &insert_after); | |
2496 | ||
2497 | create_iv (vect_ptr_init, | |
2498 | fold_convert (vect_ptr_type, step), | |
2499 | vect_ptr, loop, &incr_gsi, insert_after, | |
2500 | &indx_before_incr, &indx_after_incr); | |
2501 | incr = gsi_stmt (incr_gsi); | |
a70d6342 | 2502 | set_vinfo_for_stmt (incr, new_stmt_vec_info (incr, loop_vinfo, NULL)); |
ebfd146a IR |
2503 | |
2504 | /* Copy the points-to information if it exists. */ | |
2505 | if (DR_PTR_INFO (dr)) | |
2506 | { | |
2507 | duplicate_ssa_name_ptr_info (indx_before_incr, DR_PTR_INFO (dr)); | |
2508 | duplicate_ssa_name_ptr_info (indx_after_incr, DR_PTR_INFO (dr)); | |
2509 | } | |
ebfd146a IR |
2510 | if (ptr_incr) |
2511 | *ptr_incr = incr; | |
2512 | ||
2513 | vptr = indx_before_incr; | |
2514 | } | |
2515 | ||
2516 | if (!nested_in_vect_loop || only_init) | |
2517 | return vptr; | |
2518 | ||
2519 | ||
2520 | /** (5) Handle the updating of the vector-pointer inside the inner-loop | |
2521 | nested in LOOP, if exists: **/ | |
2522 | ||
2523 | gcc_assert (nested_in_vect_loop); | |
2524 | if (!only_init) | |
2525 | { | |
2526 | standard_iv_increment_position (containing_loop, &incr_gsi, | |
2527 | &insert_after); | |
b8698a0f | 2528 | create_iv (vptr, fold_convert (vect_ptr_type, DR_STEP (dr)), vect_ptr, |
ebfd146a IR |
2529 | containing_loop, &incr_gsi, insert_after, &indx_before_incr, |
2530 | &indx_after_incr); | |
2531 | incr = gsi_stmt (incr_gsi); | |
a70d6342 | 2532 | set_vinfo_for_stmt (incr, new_stmt_vec_info (incr, loop_vinfo, NULL)); |
ebfd146a IR |
2533 | |
2534 | /* Copy the points-to information if it exists. */ | |
2535 | if (DR_PTR_INFO (dr)) | |
2536 | { | |
2537 | duplicate_ssa_name_ptr_info (indx_before_incr, DR_PTR_INFO (dr)); | |
2538 | duplicate_ssa_name_ptr_info (indx_after_incr, DR_PTR_INFO (dr)); | |
2539 | } | |
ebfd146a IR |
2540 | if (ptr_incr) |
2541 | *ptr_incr = incr; | |
2542 | ||
b8698a0f | 2543 | return indx_before_incr; |
ebfd146a IR |
2544 | } |
2545 | else | |
2546 | gcc_unreachable (); | |
2547 | } | |
2548 | ||
2549 | ||
2550 | /* Function bump_vector_ptr | |
2551 | ||
2552 | Increment a pointer (to a vector type) by vector-size. If requested, | |
b8698a0f | 2553 | i.e. if PTR-INCR is given, then also connect the new increment stmt |
ebfd146a IR |
2554 | to the existing def-use update-chain of the pointer, by modifying |
2555 | the PTR_INCR as illustrated below: | |
2556 | ||
2557 | The pointer def-use update-chain before this function: | |
2558 | DATAREF_PTR = phi (p_0, p_2) | |
2559 | .... | |
b8698a0f | 2560 | PTR_INCR: p_2 = DATAREF_PTR + step |
ebfd146a IR |
2561 | |
2562 | The pointer def-use update-chain after this function: | |
2563 | DATAREF_PTR = phi (p_0, p_2) | |
2564 | .... | |
2565 | NEW_DATAREF_PTR = DATAREF_PTR + BUMP | |
2566 | .... | |
2567 | PTR_INCR: p_2 = NEW_DATAREF_PTR + step | |
2568 | ||
2569 | Input: | |
b8698a0f | 2570 | DATAREF_PTR - ssa_name of a pointer (to vector type) that is being updated |
ebfd146a | 2571 | in the loop. |
b8698a0f | 2572 | PTR_INCR - optional. The stmt that updates the pointer in each iteration of |
ebfd146a | 2573 | the loop. The increment amount across iterations is expected |
b8698a0f | 2574 | to be vector_size. |
ebfd146a IR |
2575 | BSI - location where the new update stmt is to be placed. |
2576 | STMT - the original scalar memory-access stmt that is being vectorized. | |
2577 | BUMP - optional. The offset by which to bump the pointer. If not given, | |
2578 | the offset is assumed to be vector_size. | |
2579 | ||
2580 | Output: Return NEW_DATAREF_PTR as illustrated above. | |
b8698a0f | 2581 | |
ebfd146a IR |
2582 | */ |
2583 | ||
2584 | tree | |
2585 | bump_vector_ptr (tree dataref_ptr, gimple ptr_incr, gimple_stmt_iterator *gsi, | |
2586 | gimple stmt, tree bump) | |
2587 | { | |
2588 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
2589 | struct data_reference *dr = STMT_VINFO_DATA_REF (stmt_info); | |
2590 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
2591 | tree ptr_var = SSA_NAME_VAR (dataref_ptr); | |
2592 | tree update = TYPE_SIZE_UNIT (vectype); | |
2593 | gimple incr_stmt; | |
2594 | ssa_op_iter iter; | |
2595 | use_operand_p use_p; | |
2596 | tree new_dataref_ptr; | |
2597 | ||
2598 | if (bump) | |
2599 | update = bump; | |
b8698a0f | 2600 | |
ebfd146a IR |
2601 | incr_stmt = gimple_build_assign_with_ops (POINTER_PLUS_EXPR, ptr_var, |
2602 | dataref_ptr, update); | |
2603 | new_dataref_ptr = make_ssa_name (ptr_var, incr_stmt); | |
2604 | gimple_assign_set_lhs (incr_stmt, new_dataref_ptr); | |
2605 | vect_finish_stmt_generation (stmt, incr_stmt, gsi); | |
2606 | ||
2607 | /* Copy the points-to information if it exists. */ | |
2608 | if (DR_PTR_INFO (dr)) | |
2609 | duplicate_ssa_name_ptr_info (new_dataref_ptr, DR_PTR_INFO (dr)); | |
ebfd146a IR |
2610 | |
2611 | if (!ptr_incr) | |
2612 | return new_dataref_ptr; | |
2613 | ||
2614 | /* Update the vector-pointer's cross-iteration increment. */ | |
2615 | FOR_EACH_SSA_USE_OPERAND (use_p, ptr_incr, iter, SSA_OP_USE) | |
2616 | { | |
2617 | tree use = USE_FROM_PTR (use_p); | |
2618 | ||
2619 | if (use == dataref_ptr) | |
2620 | SET_USE (use_p, new_dataref_ptr); | |
2621 | else | |
2622 | gcc_assert (tree_int_cst_compare (use, update) == 0); | |
2623 | } | |
2624 | ||
2625 | return new_dataref_ptr; | |
2626 | } | |
2627 | ||
2628 | ||
2629 | /* Function vect_create_destination_var. | |
2630 | ||
2631 | Create a new temporary of type VECTYPE. */ | |
2632 | ||
2633 | tree | |
2634 | vect_create_destination_var (tree scalar_dest, tree vectype) | |
2635 | { | |
2636 | tree vec_dest; | |
2637 | const char *new_name; | |
2638 | tree type; | |
2639 | enum vect_var_kind kind; | |
2640 | ||
2641 | kind = vectype ? vect_simple_var : vect_scalar_var; | |
2642 | type = vectype ? vectype : TREE_TYPE (scalar_dest); | |
2643 | ||
2644 | gcc_assert (TREE_CODE (scalar_dest) == SSA_NAME); | |
2645 | ||
2646 | new_name = get_name (scalar_dest); | |
2647 | if (!new_name) | |
2648 | new_name = "var_"; | |
2649 | vec_dest = vect_get_new_vect_var (type, kind, new_name); | |
2650 | add_referenced_var (vec_dest); | |
2651 | ||
2652 | return vec_dest; | |
2653 | } | |
2654 | ||
2655 | /* Function vect_strided_store_supported. | |
2656 | ||
2657 | Returns TRUE is INTERLEAVE_HIGH and INTERLEAVE_LOW operations are supported, | |
2658 | and FALSE otherwise. */ | |
2659 | ||
2660 | bool | |
2661 | vect_strided_store_supported (tree vectype) | |
2662 | { | |
2663 | optab interleave_high_optab, interleave_low_optab; | |
2664 | int mode; | |
2665 | ||
2666 | mode = (int) TYPE_MODE (vectype); | |
b8698a0f | 2667 | |
ebfd146a | 2668 | /* Check that the operation is supported. */ |
b8698a0f | 2669 | interleave_high_optab = optab_for_tree_code (VEC_INTERLEAVE_HIGH_EXPR, |
ebfd146a | 2670 | vectype, optab_default); |
b8698a0f | 2671 | interleave_low_optab = optab_for_tree_code (VEC_INTERLEAVE_LOW_EXPR, |
ebfd146a IR |
2672 | vectype, optab_default); |
2673 | if (!interleave_high_optab || !interleave_low_optab) | |
2674 | { | |
2675 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2676 | fprintf (vect_dump, "no optab for interleave."); | |
2677 | return false; | |
2678 | } | |
2679 | ||
b8698a0f | 2680 | if (optab_handler (interleave_high_optab, mode)->insn_code |
ebfd146a | 2681 | == CODE_FOR_nothing |
b8698a0f | 2682 | || optab_handler (interleave_low_optab, mode)->insn_code |
ebfd146a IR |
2683 | == CODE_FOR_nothing) |
2684 | { | |
2685 | if (vect_print_dump_info (REPORT_DETAILS)) | |
2686 | fprintf (vect_dump, "interleave op not supported by target."); | |
2687 | return false; | |
2688 | } | |
2689 | ||
2690 | return true; | |
2691 | } | |
2692 | ||
2693 | ||
2694 | /* Function vect_permute_store_chain. | |
2695 | ||
2696 | Given a chain of interleaved stores in DR_CHAIN of LENGTH that must be | |
b8698a0f | 2697 | a power of 2, generate interleave_high/low stmts to reorder the data |
ebfd146a IR |
2698 | correctly for the stores. Return the final references for stores in |
2699 | RESULT_CHAIN. | |
2700 | ||
2701 | E.g., LENGTH is 4 and the scalar type is short, i.e., VF is 8. | |
2702 | The input is 4 vectors each containing 8 elements. We assign a number to each | |
2703 | element, the input sequence is: | |
2704 | ||
2705 | 1st vec: 0 1 2 3 4 5 6 7 | |
2706 | 2nd vec: 8 9 10 11 12 13 14 15 | |
b8698a0f | 2707 | 3rd vec: 16 17 18 19 20 21 22 23 |
ebfd146a IR |
2708 | 4th vec: 24 25 26 27 28 29 30 31 |
2709 | ||
2710 | The output sequence should be: | |
2711 | ||
2712 | 1st vec: 0 8 16 24 1 9 17 25 | |
2713 | 2nd vec: 2 10 18 26 3 11 19 27 | |
2714 | 3rd vec: 4 12 20 28 5 13 21 30 | |
2715 | 4th vec: 6 14 22 30 7 15 23 31 | |
2716 | ||
2717 | i.e., we interleave the contents of the four vectors in their order. | |
2718 | ||
b8698a0f | 2719 | We use interleave_high/low instructions to create such output. The input of |
ebfd146a | 2720 | each interleave_high/low operation is two vectors: |
b8698a0f L |
2721 | 1st vec 2nd vec |
2722 | 0 1 2 3 4 5 6 7 | |
2723 | the even elements of the result vector are obtained left-to-right from the | |
2724 | high/low elements of the first vector. The odd elements of the result are | |
ebfd146a IR |
2725 | obtained left-to-right from the high/low elements of the second vector. |
2726 | The output of interleave_high will be: 0 4 1 5 | |
2727 | and of interleave_low: 2 6 3 7 | |
2728 | ||
b8698a0f | 2729 | |
ebfd146a | 2730 | The permutation is done in log LENGTH stages. In each stage interleave_high |
b8698a0f L |
2731 | and interleave_low stmts are created for each pair of vectors in DR_CHAIN, |
2732 | where the first argument is taken from the first half of DR_CHAIN and the | |
2733 | second argument from it's second half. | |
2734 | In our example, | |
ebfd146a IR |
2735 | |
2736 | I1: interleave_high (1st vec, 3rd vec) | |
2737 | I2: interleave_low (1st vec, 3rd vec) | |
2738 | I3: interleave_high (2nd vec, 4th vec) | |
2739 | I4: interleave_low (2nd vec, 4th vec) | |
2740 | ||
2741 | The output for the first stage is: | |
2742 | ||
2743 | I1: 0 16 1 17 2 18 3 19 | |
2744 | I2: 4 20 5 21 6 22 7 23 | |
2745 | I3: 8 24 9 25 10 26 11 27 | |
2746 | I4: 12 28 13 29 14 30 15 31 | |
2747 | ||
2748 | The output of the second stage, i.e. the final result is: | |
2749 | ||
2750 | I1: 0 8 16 24 1 9 17 25 | |
2751 | I2: 2 10 18 26 3 11 19 27 | |
2752 | I3: 4 12 20 28 5 13 21 30 | |
2753 | I4: 6 14 22 30 7 15 23 31. */ | |
b8698a0f | 2754 | |
ebfd146a | 2755 | bool |
b8698a0f L |
2756 | vect_permute_store_chain (VEC(tree,heap) *dr_chain, |
2757 | unsigned int length, | |
ebfd146a IR |
2758 | gimple stmt, |
2759 | gimple_stmt_iterator *gsi, | |
2760 | VEC(tree,heap) **result_chain) | |
2761 | { | |
2762 | tree perm_dest, vect1, vect2, high, low; | |
2763 | gimple perm_stmt; | |
2764 | tree vectype = STMT_VINFO_VECTYPE (vinfo_for_stmt (stmt)); | |
ebfd146a IR |
2765 | int i; |
2766 | unsigned int j; | |
2767 | enum tree_code high_code, low_code; | |
b8698a0f | 2768 | |
ebfd146a IR |
2769 | /* Check that the operation is supported. */ |
2770 | if (!vect_strided_store_supported (vectype)) | |
2771 | return false; | |
2772 | ||
2773 | *result_chain = VEC_copy (tree, heap, dr_chain); | |
2774 | ||
2775 | for (i = 0; i < exact_log2 (length); i++) | |
2776 | { | |
2777 | for (j = 0; j < length/2; j++) | |
2778 | { | |
2779 | vect1 = VEC_index (tree, dr_chain, j); | |
2780 | vect2 = VEC_index (tree, dr_chain, j+length/2); | |
2781 | ||
2782 | /* Create interleaving stmt: | |
b8698a0f L |
2783 | in the case of big endian: |
2784 | high = interleave_high (vect1, vect2) | |
2785 | and in the case of little endian: | |
ebfd146a IR |
2786 | high = interleave_low (vect1, vect2). */ |
2787 | perm_dest = create_tmp_var (vectype, "vect_inter_high"); | |
2788 | DECL_GIMPLE_REG_P (perm_dest) = 1; | |
2789 | add_referenced_var (perm_dest); | |
2790 | if (BYTES_BIG_ENDIAN) | |
2791 | { | |
2792 | high_code = VEC_INTERLEAVE_HIGH_EXPR; | |
2793 | low_code = VEC_INTERLEAVE_LOW_EXPR; | |
2794 | } | |
2795 | else | |
2796 | { | |
2797 | low_code = VEC_INTERLEAVE_HIGH_EXPR; | |
2798 | high_code = VEC_INTERLEAVE_LOW_EXPR; | |
2799 | } | |
2800 | perm_stmt = gimple_build_assign_with_ops (high_code, perm_dest, | |
2801 | vect1, vect2); | |
2802 | high = make_ssa_name (perm_dest, perm_stmt); | |
2803 | gimple_assign_set_lhs (perm_stmt, high); | |
2804 | vect_finish_stmt_generation (stmt, perm_stmt, gsi); | |
2805 | VEC_replace (tree, *result_chain, 2*j, high); | |
2806 | ||
2807 | /* Create interleaving stmt: | |
2808 | in the case of big endian: | |
b8698a0f | 2809 | low = interleave_low (vect1, vect2) |
ebfd146a | 2810 | and in the case of little endian: |
b8698a0f | 2811 | low = interleave_high (vect1, vect2). */ |
ebfd146a IR |
2812 | perm_dest = create_tmp_var (vectype, "vect_inter_low"); |
2813 | DECL_GIMPLE_REG_P (perm_dest) = 1; | |
2814 | add_referenced_var (perm_dest); | |
2815 | perm_stmt = gimple_build_assign_with_ops (low_code, perm_dest, | |
2816 | vect1, vect2); | |
2817 | low = make_ssa_name (perm_dest, perm_stmt); | |
2818 | gimple_assign_set_lhs (perm_stmt, low); | |
2819 | vect_finish_stmt_generation (stmt, perm_stmt, gsi); | |
2820 | VEC_replace (tree, *result_chain, 2*j+1, low); | |
2821 | } | |
2822 | dr_chain = VEC_copy (tree, heap, *result_chain); | |
2823 | } | |
2824 | return true; | |
2825 | } | |
2826 | ||
2827 | /* Function vect_setup_realignment | |
b8698a0f | 2828 | |
ebfd146a IR |
2829 | This function is called when vectorizing an unaligned load using |
2830 | the dr_explicit_realign[_optimized] scheme. | |
2831 | This function generates the following code at the loop prolog: | |
2832 | ||
2833 | p = initial_addr; | |
2834 | x msq_init = *(floor(p)); # prolog load | |
b8698a0f | 2835 | realignment_token = call target_builtin; |
ebfd146a IR |
2836 | loop: |
2837 | x msq = phi (msq_init, ---) | |
2838 | ||
b8698a0f | 2839 | The stmts marked with x are generated only for the case of |
ebfd146a IR |
2840 | dr_explicit_realign_optimized. |
2841 | ||
b8698a0f | 2842 | The code above sets up a new (vector) pointer, pointing to the first |
ebfd146a IR |
2843 | location accessed by STMT, and a "floor-aligned" load using that pointer. |
2844 | It also generates code to compute the "realignment-token" (if the relevant | |
2845 | target hook was defined), and creates a phi-node at the loop-header bb | |
2846 | whose arguments are the result of the prolog-load (created by this | |
2847 | function) and the result of a load that takes place in the loop (to be | |
2848 | created by the caller to this function). | |
2849 | ||
2850 | For the case of dr_explicit_realign_optimized: | |
b8698a0f | 2851 | The caller to this function uses the phi-result (msq) to create the |
ebfd146a IR |
2852 | realignment code inside the loop, and sets up the missing phi argument, |
2853 | as follows: | |
b8698a0f | 2854 | loop: |
ebfd146a IR |
2855 | msq = phi (msq_init, lsq) |
2856 | lsq = *(floor(p')); # load in loop | |
2857 | result = realign_load (msq, lsq, realignment_token); | |
2858 | ||
2859 | For the case of dr_explicit_realign: | |
2860 | loop: | |
2861 | msq = *(floor(p)); # load in loop | |
2862 | p' = p + (VS-1); | |
2863 | lsq = *(floor(p')); # load in loop | |
2864 | result = realign_load (msq, lsq, realignment_token); | |
2865 | ||
2866 | Input: | |
2867 | STMT - (scalar) load stmt to be vectorized. This load accesses | |
2868 | a memory location that may be unaligned. | |
2869 | BSI - place where new code is to be inserted. | |
2870 | ALIGNMENT_SUPPORT_SCHEME - which of the two misalignment handling schemes | |
b8698a0f L |
2871 | is used. |
2872 | ||
ebfd146a IR |
2873 | Output: |
2874 | REALIGNMENT_TOKEN - the result of a call to the builtin_mask_for_load | |
2875 | target hook, if defined. | |
2876 | Return value - the result of the loop-header phi node. */ | |
2877 | ||
2878 | tree | |
2879 | vect_setup_realignment (gimple stmt, gimple_stmt_iterator *gsi, | |
2880 | tree *realignment_token, | |
2881 | enum dr_alignment_support alignment_support_scheme, | |
2882 | tree init_addr, | |
2883 | struct loop **at_loop) | |
2884 | { | |
2885 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
2886 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
2887 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); | |
2888 | struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); | |
2889 | edge pe; | |
2890 | tree scalar_dest = gimple_assign_lhs (stmt); | |
2891 | tree vec_dest; | |
2892 | gimple inc; | |
2893 | tree ptr; | |
2894 | tree data_ref; | |
2895 | gimple new_stmt; | |
2896 | basic_block new_bb; | |
2897 | tree msq_init = NULL_TREE; | |
2898 | tree new_temp; | |
2899 | gimple phi_stmt; | |
2900 | tree msq = NULL_TREE; | |
2901 | gimple_seq stmts = NULL; | |
2902 | bool inv_p; | |
2903 | bool compute_in_loop = false; | |
2904 | bool nested_in_vect_loop = nested_in_vect_loop_p (loop, stmt); | |
2905 | struct loop *containing_loop = (gimple_bb (stmt))->loop_father; | |
2906 | struct loop *loop_for_initial_load; | |
2907 | ||
2908 | gcc_assert (alignment_support_scheme == dr_explicit_realign | |
2909 | || alignment_support_scheme == dr_explicit_realign_optimized); | |
2910 | ||
2911 | /* We need to generate three things: | |
2912 | 1. the misalignment computation | |
2913 | 2. the extra vector load (for the optimized realignment scheme). | |
2914 | 3. the phi node for the two vectors from which the realignment is | |
2915 | done (for the optimized realignment scheme). | |
2916 | */ | |
2917 | ||
2918 | /* 1. Determine where to generate the misalignment computation. | |
2919 | ||
2920 | If INIT_ADDR is NULL_TREE, this indicates that the misalignment | |
2921 | calculation will be generated by this function, outside the loop (in the | |
2922 | preheader). Otherwise, INIT_ADDR had already been computed for us by the | |
2923 | caller, inside the loop. | |
2924 | ||
2925 | Background: If the misalignment remains fixed throughout the iterations of | |
2926 | the loop, then both realignment schemes are applicable, and also the | |
2927 | misalignment computation can be done outside LOOP. This is because we are | |
2928 | vectorizing LOOP, and so the memory accesses in LOOP advance in steps that | |
2929 | are a multiple of VS (the Vector Size), and therefore the misalignment in | |
2930 | different vectorized LOOP iterations is always the same. | |
2931 | The problem arises only if the memory access is in an inner-loop nested | |
2932 | inside LOOP, which is now being vectorized using outer-loop vectorization. | |
2933 | This is the only case when the misalignment of the memory access may not | |
2934 | remain fixed throughout the iterations of the inner-loop (as explained in | |
2935 | detail in vect_supportable_dr_alignment). In this case, not only is the | |
2936 | optimized realignment scheme not applicable, but also the misalignment | |
2937 | computation (and generation of the realignment token that is passed to | |
2938 | REALIGN_LOAD) have to be done inside the loop. | |
2939 | ||
2940 | In short, INIT_ADDR indicates whether we are in a COMPUTE_IN_LOOP mode | |
2941 | or not, which in turn determines if the misalignment is computed inside | |
2942 | the inner-loop, or outside LOOP. */ | |
2943 | ||
2944 | if (init_addr != NULL_TREE) | |
2945 | { | |
2946 | compute_in_loop = true; | |
2947 | gcc_assert (alignment_support_scheme == dr_explicit_realign); | |
2948 | } | |
2949 | ||
2950 | ||
2951 | /* 2. Determine where to generate the extra vector load. | |
2952 | ||
2953 | For the optimized realignment scheme, instead of generating two vector | |
2954 | loads in each iteration, we generate a single extra vector load in the | |
2955 | preheader of the loop, and in each iteration reuse the result of the | |
2956 | vector load from the previous iteration. In case the memory access is in | |
2957 | an inner-loop nested inside LOOP, which is now being vectorized using | |
2958 | outer-loop vectorization, we need to determine whether this initial vector | |
2959 | load should be generated at the preheader of the inner-loop, or can be | |
2960 | generated at the preheader of LOOP. If the memory access has no evolution | |
2961 | in LOOP, it can be generated in the preheader of LOOP. Otherwise, it has | |
2962 | to be generated inside LOOP (in the preheader of the inner-loop). */ | |
2963 | ||
2964 | if (nested_in_vect_loop) | |
2965 | { | |
2966 | tree outerloop_step = STMT_VINFO_DR_STEP (stmt_info); | |
2967 | bool invariant_in_outerloop = | |
2968 | (tree_int_cst_compare (outerloop_step, size_zero_node) == 0); | |
2969 | loop_for_initial_load = (invariant_in_outerloop ? loop : loop->inner); | |
2970 | } | |
2971 | else | |
2972 | loop_for_initial_load = loop; | |
2973 | if (at_loop) | |
2974 | *at_loop = loop_for_initial_load; | |
2975 | ||
2976 | /* 3. For the case of the optimized realignment, create the first vector | |
2977 | load at the loop preheader. */ | |
2978 | ||
2979 | if (alignment_support_scheme == dr_explicit_realign_optimized) | |
2980 | { | |
2981 | /* Create msq_init = *(floor(p1)) in the loop preheader */ | |
2982 | ||
2983 | gcc_assert (!compute_in_loop); | |
2984 | pe = loop_preheader_edge (loop_for_initial_load); | |
2985 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
2986 | ptr = vect_create_data_ref_ptr (stmt, loop_for_initial_load, NULL_TREE, | |
5006671f | 2987 | &init_addr, &inc, true, &inv_p); |
ebfd146a IR |
2988 | data_ref = build1 (ALIGN_INDIRECT_REF, vectype, ptr); |
2989 | new_stmt = gimple_build_assign (vec_dest, data_ref); | |
2990 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
2991 | gimple_assign_set_lhs (new_stmt, new_temp); | |
2992 | mark_symbols_for_renaming (new_stmt); | |
2993 | new_bb = gsi_insert_on_edge_immediate (pe, new_stmt); | |
2994 | gcc_assert (!new_bb); | |
2995 | msq_init = gimple_assign_lhs (new_stmt); | |
2996 | } | |
2997 | ||
2998 | /* 4. Create realignment token using a target builtin, if available. | |
2999 | It is done either inside the containing loop, or before LOOP (as | |
3000 | determined above). */ | |
3001 | ||
3002 | if (targetm.vectorize.builtin_mask_for_load) | |
3003 | { | |
3004 | tree builtin_decl; | |
3005 | ||
3006 | /* Compute INIT_ADDR - the initial addressed accessed by this memref. */ | |
3007 | if (compute_in_loop) | |
3008 | gcc_assert (init_addr); /* already computed by the caller. */ | |
3009 | else | |
3010 | { | |
3011 | /* Generate the INIT_ADDR computation outside LOOP. */ | |
3012 | init_addr = vect_create_addr_base_for_vector_ref (stmt, &stmts, | |
3013 | NULL_TREE, loop); | |
3014 | pe = loop_preheader_edge (loop); | |
3015 | new_bb = gsi_insert_seq_on_edge_immediate (pe, stmts); | |
3016 | gcc_assert (!new_bb); | |
3017 | } | |
3018 | ||
3019 | builtin_decl = targetm.vectorize.builtin_mask_for_load (); | |
3020 | new_stmt = gimple_build_call (builtin_decl, 1, init_addr); | |
3021 | vec_dest = | |
3022 | vect_create_destination_var (scalar_dest, | |
3023 | gimple_call_return_type (new_stmt)); | |
3024 | new_temp = make_ssa_name (vec_dest, new_stmt); | |
3025 | gimple_call_set_lhs (new_stmt, new_temp); | |
3026 | ||
3027 | if (compute_in_loop) | |
3028 | gsi_insert_before (gsi, new_stmt, GSI_SAME_STMT); | |
3029 | else | |
3030 | { | |
3031 | /* Generate the misalignment computation outside LOOP. */ | |
3032 | pe = loop_preheader_edge (loop); | |
3033 | new_bb = gsi_insert_on_edge_immediate (pe, new_stmt); | |
3034 | gcc_assert (!new_bb); | |
3035 | } | |
3036 | ||
3037 | *realignment_token = gimple_call_lhs (new_stmt); | |
3038 | ||
3039 | /* The result of the CALL_EXPR to this builtin is determined from | |
3040 | the value of the parameter and no global variables are touched | |
3041 | which makes the builtin a "const" function. Requiring the | |
3042 | builtin to have the "const" attribute makes it unnecessary | |
3043 | to call mark_call_clobbered. */ | |
3044 | gcc_assert (TREE_READONLY (builtin_decl)); | |
3045 | } | |
3046 | ||
3047 | if (alignment_support_scheme == dr_explicit_realign) | |
3048 | return msq; | |
3049 | ||
3050 | gcc_assert (!compute_in_loop); | |
3051 | gcc_assert (alignment_support_scheme == dr_explicit_realign_optimized); | |
3052 | ||
3053 | ||
3054 | /* 5. Create msq = phi <msq_init, lsq> in loop */ | |
3055 | ||
3056 | pe = loop_preheader_edge (containing_loop); | |
3057 | vec_dest = vect_create_destination_var (scalar_dest, vectype); | |
3058 | msq = make_ssa_name (vec_dest, NULL); | |
3059 | phi_stmt = create_phi_node (msq, containing_loop->header); | |
3060 | SSA_NAME_DEF_STMT (msq) = phi_stmt; | |
f5045c96 | 3061 | add_phi_arg (phi_stmt, msq_init, pe, UNKNOWN_LOCATION); |
ebfd146a IR |
3062 | |
3063 | return msq; | |
3064 | } | |
3065 | ||
3066 | ||
3067 | /* Function vect_strided_load_supported. | |
3068 | ||
3069 | Returns TRUE is EXTRACT_EVEN and EXTRACT_ODD operations are supported, | |
3070 | and FALSE otherwise. */ | |
3071 | ||
3072 | bool | |
3073 | vect_strided_load_supported (tree vectype) | |
3074 | { | |
3075 | optab perm_even_optab, perm_odd_optab; | |
3076 | int mode; | |
3077 | ||
3078 | mode = (int) TYPE_MODE (vectype); | |
3079 | ||
3080 | perm_even_optab = optab_for_tree_code (VEC_EXTRACT_EVEN_EXPR, vectype, | |
3081 | optab_default); | |
3082 | if (!perm_even_optab) | |
3083 | { | |
3084 | if (vect_print_dump_info (REPORT_DETAILS)) | |
3085 | fprintf (vect_dump, "no optab for perm_even."); | |
3086 | return false; | |
3087 | } | |
3088 | ||
3089 | if (optab_handler (perm_even_optab, mode)->insn_code == CODE_FOR_nothing) | |
3090 | { | |
3091 | if (vect_print_dump_info (REPORT_DETAILS)) | |
3092 | fprintf (vect_dump, "perm_even op not supported by target."); | |
3093 | return false; | |
3094 | } | |
3095 | ||
3096 | perm_odd_optab = optab_for_tree_code (VEC_EXTRACT_ODD_EXPR, vectype, | |
3097 | optab_default); | |
3098 | if (!perm_odd_optab) | |
3099 | { | |
3100 | if (vect_print_dump_info (REPORT_DETAILS)) | |
3101 | fprintf (vect_dump, "no optab for perm_odd."); | |
3102 | return false; | |
3103 | } | |
3104 | ||
3105 | if (optab_handler (perm_odd_optab, mode)->insn_code == CODE_FOR_nothing) | |
3106 | { | |
3107 | if (vect_print_dump_info (REPORT_DETAILS)) | |
3108 | fprintf (vect_dump, "perm_odd op not supported by target."); | |
3109 | return false; | |
3110 | } | |
3111 | return true; | |
3112 | } | |
3113 | ||
3114 | ||
3115 | /* Function vect_permute_load_chain. | |
3116 | ||
3117 | Given a chain of interleaved loads in DR_CHAIN of LENGTH that must be | |
b8698a0f | 3118 | a power of 2, generate extract_even/odd stmts to reorder the input data |
ebfd146a IR |
3119 | correctly. Return the final references for loads in RESULT_CHAIN. |
3120 | ||
3121 | E.g., LENGTH is 4 and the scalar type is short, i.e., VF is 8. | |
3122 | The input is 4 vectors each containing 8 elements. We assign a number to each | |
3123 | element, the input sequence is: | |
3124 | ||
3125 | 1st vec: 0 1 2 3 4 5 6 7 | |
3126 | 2nd vec: 8 9 10 11 12 13 14 15 | |
b8698a0f | 3127 | 3rd vec: 16 17 18 19 20 21 22 23 |
ebfd146a IR |
3128 | 4th vec: 24 25 26 27 28 29 30 31 |
3129 | ||
3130 | The output sequence should be: | |
3131 | ||
3132 | 1st vec: 0 4 8 12 16 20 24 28 | |
3133 | 2nd vec: 1 5 9 13 17 21 25 29 | |
b8698a0f | 3134 | 3rd vec: 2 6 10 14 18 22 26 30 |
ebfd146a IR |
3135 | 4th vec: 3 7 11 15 19 23 27 31 |
3136 | ||
3137 | i.e., the first output vector should contain the first elements of each | |
3138 | interleaving group, etc. | |
3139 | ||
3140 | We use extract_even/odd instructions to create such output. The input of each | |
3141 | extract_even/odd operation is two vectors | |
b8698a0f L |
3142 | 1st vec 2nd vec |
3143 | 0 1 2 3 4 5 6 7 | |
ebfd146a | 3144 | |
b8698a0f | 3145 | and the output is the vector of extracted even/odd elements. The output of |
ebfd146a IR |
3146 | extract_even will be: 0 2 4 6 |
3147 | and of extract_odd: 1 3 5 7 | |
3148 | ||
b8698a0f | 3149 | |
ebfd146a | 3150 | The permutation is done in log LENGTH stages. In each stage extract_even and |
b8698a0f L |
3151 | extract_odd stmts are created for each pair of vectors in DR_CHAIN in their |
3152 | order. In our example, | |
ebfd146a IR |
3153 | |
3154 | E1: extract_even (1st vec, 2nd vec) | |
3155 | E2: extract_odd (1st vec, 2nd vec) | |
3156 | E3: extract_even (3rd vec, 4th vec) | |
3157 | E4: extract_odd (3rd vec, 4th vec) | |
3158 | ||
3159 | The output for the first stage will be: | |
3160 | ||
3161 | E1: 0 2 4 6 8 10 12 14 | |
3162 | E2: 1 3 5 7 9 11 13 15 | |
b8698a0f | 3163 | E3: 16 18 20 22 24 26 28 30 |
ebfd146a IR |
3164 | E4: 17 19 21 23 25 27 29 31 |
3165 | ||
3166 | In order to proceed and create the correct sequence for the next stage (or | |
b8698a0f L |
3167 | for the correct output, if the second stage is the last one, as in our |
3168 | example), we first put the output of extract_even operation and then the | |
ebfd146a IR |
3169 | output of extract_odd in RESULT_CHAIN (which is then copied to DR_CHAIN). |
3170 | The input for the second stage is: | |
3171 | ||
3172 | 1st vec (E1): 0 2 4 6 8 10 12 14 | |
b8698a0f L |
3173 | 2nd vec (E3): 16 18 20 22 24 26 28 30 |
3174 | 3rd vec (E2): 1 3 5 7 9 11 13 15 | |
ebfd146a IR |
3175 | 4th vec (E4): 17 19 21 23 25 27 29 31 |
3176 | ||
3177 | The output of the second stage: | |
3178 | ||
3179 | E1: 0 4 8 12 16 20 24 28 | |
3180 | E2: 2 6 10 14 18 22 26 30 | |
3181 | E3: 1 5 9 13 17 21 25 29 | |
3182 | E4: 3 7 11 15 19 23 27 31 | |
3183 | ||
3184 | And RESULT_CHAIN after reordering: | |
3185 | ||
3186 | 1st vec (E1): 0 4 8 12 16 20 24 28 | |
3187 | 2nd vec (E3): 1 5 9 13 17 21 25 29 | |
b8698a0f | 3188 | 3rd vec (E2): 2 6 10 14 18 22 26 30 |
ebfd146a IR |
3189 | 4th vec (E4): 3 7 11 15 19 23 27 31. */ |
3190 | ||
3191 | bool | |
b8698a0f L |
3192 | vect_permute_load_chain (VEC(tree,heap) *dr_chain, |
3193 | unsigned int length, | |
ebfd146a IR |
3194 | gimple stmt, |
3195 | gimple_stmt_iterator *gsi, | |
3196 | VEC(tree,heap) **result_chain) | |
3197 | { | |
3198 | tree perm_dest, data_ref, first_vect, second_vect; | |
3199 | gimple perm_stmt; | |
3200 | tree vectype = STMT_VINFO_VECTYPE (vinfo_for_stmt (stmt)); | |
3201 | int i; | |
3202 | unsigned int j; | |
3203 | ||
3204 | /* Check that the operation is supported. */ | |
3205 | if (!vect_strided_load_supported (vectype)) | |
3206 | return false; | |
3207 | ||
3208 | *result_chain = VEC_copy (tree, heap, dr_chain); | |
3209 | for (i = 0; i < exact_log2 (length); i++) | |
3210 | { | |
3211 | for (j = 0; j < length; j +=2) | |
3212 | { | |
3213 | first_vect = VEC_index (tree, dr_chain, j); | |
3214 | second_vect = VEC_index (tree, dr_chain, j+1); | |
3215 | ||
3216 | /* data_ref = permute_even (first_data_ref, second_data_ref); */ | |
3217 | perm_dest = create_tmp_var (vectype, "vect_perm_even"); | |
3218 | DECL_GIMPLE_REG_P (perm_dest) = 1; | |
3219 | add_referenced_var (perm_dest); | |
3220 | ||
3221 | perm_stmt = gimple_build_assign_with_ops (VEC_EXTRACT_EVEN_EXPR, | |
3222 | perm_dest, first_vect, | |
3223 | second_vect); | |
3224 | ||
3225 | data_ref = make_ssa_name (perm_dest, perm_stmt); | |
3226 | gimple_assign_set_lhs (perm_stmt, data_ref); | |
3227 | vect_finish_stmt_generation (stmt, perm_stmt, gsi); | |
3228 | mark_symbols_for_renaming (perm_stmt); | |
3229 | ||
b8698a0f L |
3230 | VEC_replace (tree, *result_chain, j/2, data_ref); |
3231 | ||
ebfd146a IR |
3232 | /* data_ref = permute_odd (first_data_ref, second_data_ref); */ |
3233 | perm_dest = create_tmp_var (vectype, "vect_perm_odd"); | |
3234 | DECL_GIMPLE_REG_P (perm_dest) = 1; | |
3235 | add_referenced_var (perm_dest); | |
3236 | ||
3237 | perm_stmt = gimple_build_assign_with_ops (VEC_EXTRACT_ODD_EXPR, | |
3238 | perm_dest, first_vect, | |
3239 | second_vect); | |
3240 | data_ref = make_ssa_name (perm_dest, perm_stmt); | |
3241 | gimple_assign_set_lhs (perm_stmt, data_ref); | |
3242 | vect_finish_stmt_generation (stmt, perm_stmt, gsi); | |
3243 | mark_symbols_for_renaming (perm_stmt); | |
3244 | ||
3245 | VEC_replace (tree, *result_chain, j/2+length/2, data_ref); | |
3246 | } | |
3247 | dr_chain = VEC_copy (tree, heap, *result_chain); | |
3248 | } | |
3249 | return true; | |
3250 | } | |
3251 | ||
3252 | ||
3253 | /* Function vect_transform_strided_load. | |
3254 | ||
3255 | Given a chain of input interleaved data-refs (in DR_CHAIN), build statements | |
3256 | to perform their permutation and ascribe the result vectorized statements to | |
3257 | the scalar statements. | |
3258 | */ | |
3259 | ||
3260 | bool | |
3261 | vect_transform_strided_load (gimple stmt, VEC(tree,heap) *dr_chain, int size, | |
3262 | gimple_stmt_iterator *gsi) | |
3263 | { | |
3264 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
3265 | gimple first_stmt = DR_GROUP_FIRST_DR (stmt_info); | |
3266 | gimple next_stmt, new_stmt; | |
3267 | VEC(tree,heap) *result_chain = NULL; | |
3268 | unsigned int i, gap_count; | |
3269 | tree tmp_data_ref; | |
3270 | ||
b8698a0f L |
3271 | /* DR_CHAIN contains input data-refs that are a part of the interleaving. |
3272 | RESULT_CHAIN is the output of vect_permute_load_chain, it contains permuted | |
ebfd146a IR |
3273 | vectors, that are ready for vector computation. */ |
3274 | result_chain = VEC_alloc (tree, heap, size); | |
3275 | /* Permute. */ | |
3276 | if (!vect_permute_load_chain (dr_chain, size, stmt, gsi, &result_chain)) | |
3277 | return false; | |
3278 | ||
b8698a0f L |
3279 | /* Put a permuted data-ref in the VECTORIZED_STMT field. |
3280 | Since we scan the chain starting from it's first node, their order | |
ebfd146a IR |
3281 | corresponds the order of data-refs in RESULT_CHAIN. */ |
3282 | next_stmt = first_stmt; | |
3283 | gap_count = 1; | |
3284 | for (i = 0; VEC_iterate (tree, result_chain, i, tmp_data_ref); i++) | |
3285 | { | |
3286 | if (!next_stmt) | |
3287 | break; | |
3288 | ||
3289 | /* Skip the gaps. Loads created for the gaps will be removed by dead | |
3290 | code elimination pass later. No need to check for the first stmt in | |
3291 | the group, since it always exists. | |
3292 | DR_GROUP_GAP is the number of steps in elements from the previous | |
3293 | access (if there is no gap DR_GROUP_GAP is 1). We skip loads that | |
3294 | correspond to the gaps. | |
3295 | */ | |
b8698a0f | 3296 | if (next_stmt != first_stmt |
ebfd146a IR |
3297 | && gap_count < DR_GROUP_GAP (vinfo_for_stmt (next_stmt))) |
3298 | { | |
3299 | gap_count++; | |
3300 | continue; | |
3301 | } | |
3302 | ||
3303 | while (next_stmt) | |
3304 | { | |
3305 | new_stmt = SSA_NAME_DEF_STMT (tmp_data_ref); | |
3306 | /* We assume that if VEC_STMT is not NULL, this is a case of multiple | |
3307 | copies, and we put the new vector statement in the first available | |
3308 | RELATED_STMT. */ | |
3309 | if (!STMT_VINFO_VEC_STMT (vinfo_for_stmt (next_stmt))) | |
3310 | STMT_VINFO_VEC_STMT (vinfo_for_stmt (next_stmt)) = new_stmt; | |
3311 | else | |
3312 | { | |
3313 | if (!DR_GROUP_SAME_DR_STMT (vinfo_for_stmt (next_stmt))) | |
3314 | { | |
3315 | gimple prev_stmt = | |
3316 | STMT_VINFO_VEC_STMT (vinfo_for_stmt (next_stmt)); | |
3317 | gimple rel_stmt = | |
3318 | STMT_VINFO_RELATED_STMT (vinfo_for_stmt (prev_stmt)); | |
3319 | while (rel_stmt) | |
3320 | { | |
3321 | prev_stmt = rel_stmt; | |
b8698a0f | 3322 | rel_stmt = |
ebfd146a IR |
3323 | STMT_VINFO_RELATED_STMT (vinfo_for_stmt (rel_stmt)); |
3324 | } | |
3325 | ||
b8698a0f | 3326 | STMT_VINFO_RELATED_STMT (vinfo_for_stmt (prev_stmt)) = |
ebfd146a IR |
3327 | new_stmt; |
3328 | } | |
3329 | } | |
3330 | ||
3331 | next_stmt = DR_GROUP_NEXT_DR (vinfo_for_stmt (next_stmt)); | |
3332 | gap_count = 1; | |
3333 | /* If NEXT_STMT accesses the same DR as the previous statement, | |
3334 | put the same TMP_DATA_REF as its vectorized statement; otherwise | |
3335 | get the next data-ref from RESULT_CHAIN. */ | |
3336 | if (!next_stmt || !DR_GROUP_SAME_DR_STMT (vinfo_for_stmt (next_stmt))) | |
3337 | break; | |
3338 | } | |
3339 | } | |
3340 | ||
3341 | VEC_free (tree, heap, result_chain); | |
3342 | return true; | |
3343 | } | |
3344 | ||
3345 | /* Function vect_force_dr_alignment_p. | |
3346 | ||
3347 | Returns whether the alignment of a DECL can be forced to be aligned | |
3348 | on ALIGNMENT bit boundary. */ | |
3349 | ||
b8698a0f | 3350 | bool |
ebfd146a IR |
3351 | vect_can_force_dr_alignment_p (const_tree decl, unsigned int alignment) |
3352 | { | |
3353 | if (TREE_CODE (decl) != VAR_DECL) | |
3354 | return false; | |
3355 | ||
3356 | if (DECL_EXTERNAL (decl)) | |
3357 | return false; | |
3358 | ||
3359 | if (TREE_ASM_WRITTEN (decl)) | |
3360 | return false; | |
3361 | ||
3362 | if (TREE_STATIC (decl)) | |
3363 | return (alignment <= MAX_OFILE_ALIGNMENT); | |
3364 | else | |
3365 | return (alignment <= MAX_STACK_ALIGNMENT); | |
3366 | } | |
3367 | ||
3368 | /* Function vect_supportable_dr_alignment | |
3369 | ||
3370 | Return whether the data reference DR is supported with respect to its | |
3371 | alignment. */ | |
3372 | ||
3373 | enum dr_alignment_support | |
3374 | vect_supportable_dr_alignment (struct data_reference *dr) | |
3375 | { | |
3376 | gimple stmt = DR_STMT (dr); | |
3377 | stmt_vec_info stmt_info = vinfo_for_stmt (stmt); | |
3378 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
81f40b79 | 3379 | enum machine_mode mode = TYPE_MODE (vectype); |
a70d6342 IR |
3380 | loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); |
3381 | struct loop *vect_loop = NULL; | |
3382 | bool nested_in_vect_loop = false; | |
ebfd146a IR |
3383 | |
3384 | if (aligned_access_p (dr)) | |
3385 | return dr_aligned; | |
3386 | ||
a70d6342 IR |
3387 | if (!loop_vinfo) |
3388 | /* FORNOW: Misaligned accesses are supported only in loops. */ | |
3389 | return dr_unaligned_unsupported; | |
3390 | ||
3391 | vect_loop = LOOP_VINFO_LOOP (loop_vinfo); | |
3392 | nested_in_vect_loop = nested_in_vect_loop_p (vect_loop, stmt); | |
3393 | ||
ebfd146a IR |
3394 | /* Possibly unaligned access. */ |
3395 | ||
3396 | /* We can choose between using the implicit realignment scheme (generating | |
3397 | a misaligned_move stmt) and the explicit realignment scheme (generating | |
3398 | aligned loads with a REALIGN_LOAD). There are two variants to the explicit | |
3399 | realignment scheme: optimized, and unoptimized. | |
3400 | We can optimize the realignment only if the step between consecutive | |
3401 | vector loads is equal to the vector size. Since the vector memory | |
3402 | accesses advance in steps of VS (Vector Size) in the vectorized loop, it | |
3403 | is guaranteed that the misalignment amount remains the same throughout the | |
3404 | execution of the vectorized loop. Therefore, we can create the | |
3405 | "realignment token" (the permutation mask that is passed to REALIGN_LOAD) | |
3406 | at the loop preheader. | |
3407 | ||
3408 | However, in the case of outer-loop vectorization, when vectorizing a | |
3409 | memory access in the inner-loop nested within the LOOP that is now being | |
3410 | vectorized, while it is guaranteed that the misalignment of the | |
3411 | vectorized memory access will remain the same in different outer-loop | |
3412 | iterations, it is *not* guaranteed that is will remain the same throughout | |
3413 | the execution of the inner-loop. This is because the inner-loop advances | |
3414 | with the original scalar step (and not in steps of VS). If the inner-loop | |
3415 | step happens to be a multiple of VS, then the misalignment remains fixed | |
3416 | and we can use the optimized realignment scheme. For example: | |
3417 | ||
3418 | for (i=0; i<N; i++) | |
3419 | for (j=0; j<M; j++) | |
3420 | s += a[i+j]; | |
3421 | ||
3422 | When vectorizing the i-loop in the above example, the step between | |
3423 | consecutive vector loads is 1, and so the misalignment does not remain | |
3424 | fixed across the execution of the inner-loop, and the realignment cannot | |
3425 | be optimized (as illustrated in the following pseudo vectorized loop): | |
3426 | ||
3427 | for (i=0; i<N; i+=4) | |
3428 | for (j=0; j<M; j++){ | |
3429 | vs += vp[i+j]; // misalignment of &vp[i+j] is {0,1,2,3,0,1,2,3,...} | |
3430 | // when j is {0,1,2,3,4,5,6,7,...} respectively. | |
3431 | // (assuming that we start from an aligned address). | |
3432 | } | |
3433 | ||
3434 | We therefore have to use the unoptimized realignment scheme: | |
3435 | ||
3436 | for (i=0; i<N; i+=4) | |
3437 | for (j=k; j<M; j+=4) | |
3438 | vs += vp[i+j]; // misalignment of &vp[i+j] is always k (assuming | |
3439 | // that the misalignment of the initial address is | |
3440 | // 0). | |
3441 | ||
3442 | The loop can then be vectorized as follows: | |
3443 | ||
3444 | for (k=0; k<4; k++){ | |
3445 | rt = get_realignment_token (&vp[k]); | |
3446 | for (i=0; i<N; i+=4){ | |
3447 | v1 = vp[i+k]; | |
3448 | for (j=k; j<M; j+=4){ | |
3449 | v2 = vp[i+j+VS-1]; | |
3450 | va = REALIGN_LOAD <v1,v2,rt>; | |
3451 | vs += va; | |
3452 | v1 = v2; | |
3453 | } | |
3454 | } | |
3455 | } */ | |
3456 | ||
3457 | if (DR_IS_READ (dr)) | |
3458 | { | |
0601d0cf RE |
3459 | bool is_packed = false; |
3460 | tree type = (TREE_TYPE (DR_REF (dr))); | |
3461 | ||
b8698a0f | 3462 | if (optab_handler (vec_realign_load_optab, mode)->insn_code != |
ebfd146a IR |
3463 | CODE_FOR_nothing |
3464 | && (!targetm.vectorize.builtin_mask_for_load | |
3465 | || targetm.vectorize.builtin_mask_for_load ())) | |
3466 | { | |
3467 | tree vectype = STMT_VINFO_VECTYPE (stmt_info); | |
3468 | if (nested_in_vect_loop | |
3469 | && (TREE_INT_CST_LOW (DR_STEP (dr)) | |
3470 | != GET_MODE_SIZE (TYPE_MODE (vectype)))) | |
3471 | return dr_explicit_realign; | |
3472 | else | |
3473 | return dr_explicit_realign_optimized; | |
3474 | } | |
0601d0cf RE |
3475 | if (!known_alignment_for_access_p (dr)) |
3476 | { | |
3477 | tree ba = DR_BASE_OBJECT (dr); | |
b8698a0f | 3478 | |
0601d0cf RE |
3479 | if (ba) |
3480 | is_packed = contains_packed_reference (ba); | |
3481 | } | |
b8698a0f | 3482 | |
0601d0cf RE |
3483 | if (targetm.vectorize. |
3484 | builtin_support_vector_misalignment (mode, type, | |
3485 | DR_MISALIGNMENT (dr), is_packed)) | |
ebfd146a IR |
3486 | /* Can't software pipeline the loads, but can at least do them. */ |
3487 | return dr_unaligned_supported; | |
3488 | } | |
0601d0cf RE |
3489 | else |
3490 | { | |
3491 | bool is_packed = false; | |
3492 | tree type = (TREE_TYPE (DR_REF (dr))); | |
ebfd146a | 3493 | |
0601d0cf RE |
3494 | if (!known_alignment_for_access_p (dr)) |
3495 | { | |
3496 | tree ba = DR_BASE_OBJECT (dr); | |
b8698a0f | 3497 | |
0601d0cf RE |
3498 | if (ba) |
3499 | is_packed = contains_packed_reference (ba); | |
3500 | } | |
b8698a0f | 3501 | |
0601d0cf | 3502 | if (targetm.vectorize. |
b8698a0f | 3503 | builtin_support_vector_misalignment (mode, type, |
0601d0cf RE |
3504 | DR_MISALIGNMENT (dr), is_packed)) |
3505 | return dr_unaligned_supported; | |
3506 | } | |
b8698a0f | 3507 | |
ebfd146a IR |
3508 | /* Unsupported. */ |
3509 | return dr_unaligned_unsupported; | |
3510 | } |