]>
Commit | Line | Data |
---|---|---|
2146e26d | 1 | /* Data references and dependences detectors. |
f1717362 | 2 | Copyright (C) 2003-2016 Free Software Foundation, Inc. |
6b421feb | 3 | Contributed by Sebastian Pop <pop@cri.ensmp.fr> |
2146e26d | 4 | |
5 | This file is part of GCC. | |
6 | ||
7 | GCC is free software; you can redistribute it and/or modify it under | |
8 | the terms of the GNU General Public License as published by the Free | |
8c4c00c1 | 9 | Software Foundation; either version 3, or (at your option) any later |
2146e26d | 10 | version. |
11 | ||
12 | GCC is distributed in the hope that it will be useful, but WITHOUT ANY | |
13 | WARRANTY; without even the implied warranty of MERCHANTABILITY or | |
14 | FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License | |
15 | for more details. | |
16 | ||
17 | You should have received a copy of the GNU General Public License | |
8c4c00c1 | 18 | along with GCC; see the file COPYING3. If not see |
19 | <http://www.gnu.org/licenses/>. */ | |
2146e26d | 20 | |
21 | /* This pass walks a given loop structure searching for array | |
22 | references. The information about the array accesses is recorded | |
48e1416a | 23 | in DATA_REFERENCE structures. |
24 | ||
25 | The basic test for determining the dependences is: | |
26 | given two access functions chrec1 and chrec2 to a same array, and | |
27 | x and y two vectors from the iteration domain, the same element of | |
2146e26d | 28 | the array is accessed twice at iterations x and y if and only if: |
29 | | chrec1 (x) == chrec2 (y). | |
48e1416a | 30 | |
2146e26d | 31 | The goals of this analysis are: |
48e1416a | 32 | |
2146e26d | 33 | - to determine the independence: the relation between two |
34 | independent accesses is qualified with the chrec_known (this | |
35 | information allows a loop parallelization), | |
48e1416a | 36 | |
2146e26d | 37 | - when two data references access the same data, to qualify the |
38 | dependence relation with classic dependence representations: | |
48e1416a | 39 | |
2146e26d | 40 | - distance vectors |
41 | - direction vectors | |
42 | - loop carried level dependence | |
43 | - polyhedron dependence | |
44 | or with the chains of recurrences based representation, | |
48e1416a | 45 | |
46 | - to define a knowledge base for storing the data dependence | |
2146e26d | 47 | information, |
48e1416a | 48 | |
2146e26d | 49 | - to define an interface to access this data. |
48e1416a | 50 | |
51 | ||
2146e26d | 52 | Definitions: |
48e1416a | 53 | |
2146e26d | 54 | - subscript: given two array accesses a subscript is the tuple |
55 | composed of the access functions for a given dimension. Example: | |
56 | Given A[f1][f2][f3] and B[g1][g2][g3], there are three subscripts: | |
57 | (f1, g1), (f2, g2), (f3, g3). | |
58 | ||
59 | - Diophantine equation: an equation whose coefficients and | |
48e1416a | 60 | solutions are integer constants, for example the equation |
2146e26d | 61 | | 3*x + 2*y = 1 |
62 | has an integer solution x = 1 and y = -1. | |
48e1416a | 63 | |
2146e26d | 64 | References: |
48e1416a | 65 | |
2146e26d | 66 | - "Advanced Compilation for High Performance Computing" by Randy |
67 | Allen and Ken Kennedy. | |
48e1416a | 68 | http://citeseer.ist.psu.edu/goff91practical.html |
69 | ||
70 | - "Loop Transformations for Restructuring Compilers - The Foundations" | |
2146e26d | 71 | by Utpal Banerjee. |
72 | ||
48e1416a | 73 | |
2146e26d | 74 | */ |
75 | ||
76 | #include "config.h" | |
77 | #include "system.h" | |
78 | #include "coretypes.h" | |
9ef16211 | 79 | #include "backend.h" |
7c29e30e | 80 | #include "rtl.h" |
0d9585ca | 81 | #include "tree.h" |
9ef16211 | 82 | #include "gimple.h" |
7c29e30e | 83 | #include "gimple-pretty-print.h" |
84 | #include "alias.h" | |
9ef16211 | 85 | #include "fold-const.h" |
d53441c8 | 86 | #include "expr.h" |
dcf1a1ec | 87 | #include "gimple-iterator.h" |
05d9c18a | 88 | #include "tree-ssa-loop-niter.h" |
073c1fd5 | 89 | #include "tree-ssa-loop.h" |
69ee5dbb | 90 | #include "tree-ssa.h" |
2146e26d | 91 | #include "cfgloop.h" |
2146e26d | 92 | #include "tree-data-ref.h" |
93 | #include "tree-scalar-evolution.h" | |
b9ed1410 | 94 | #include "dumpfile.h" |
5fc88ffd | 95 | #include "tree-affine.h" |
c7af8ae7 | 96 | #include "params.h" |
2146e26d | 97 | |
6b421feb | 98 | static struct datadep_stats |
99 | { | |
100 | int num_dependence_tests; | |
101 | int num_dependence_dependent; | |
102 | int num_dependence_independent; | |
103 | int num_dependence_undetermined; | |
104 | ||
105 | int num_subscript_tests; | |
106 | int num_subscript_undetermined; | |
107 | int num_same_subscript_function; | |
108 | ||
109 | int num_ziv; | |
110 | int num_ziv_independent; | |
111 | int num_ziv_dependent; | |
112 | int num_ziv_unimplemented; | |
113 | ||
114 | int num_siv; | |
115 | int num_siv_independent; | |
116 | int num_siv_dependent; | |
117 | int num_siv_unimplemented; | |
118 | ||
119 | int num_miv; | |
120 | int num_miv_independent; | |
121 | int num_miv_dependent; | |
122 | int num_miv_unimplemented; | |
123 | } dependence_stats; | |
124 | ||
b44d1046 | 125 | static bool subscript_dependence_tester_1 (struct data_dependence_relation *, |
126 | struct data_reference *, | |
8c7b1d8e | 127 | struct data_reference *, |
128 | struct loop *); | |
2146e26d | 129 | /* Returns true iff A divides B. */ |
130 | ||
48e1416a | 131 | static inline bool |
7ecb5bb2 | 132 | tree_fold_divides_p (const_tree a, const_tree b) |
2146e26d | 133 | { |
426a138f | 134 | gcc_assert (TREE_CODE (a) == INTEGER_CST); |
135 | gcc_assert (TREE_CODE (b) == INTEGER_CST); | |
317e2a67 | 136 | return integer_zerop (int_const_binop (TRUNC_MOD_EXPR, b, a)); |
2146e26d | 137 | } |
138 | ||
bc3c8ad4 | 139 | /* Returns true iff A divides B. */ |
140 | ||
48e1416a | 141 | static inline bool |
bc3c8ad4 | 142 | int_divides_p (int a, int b) |
143 | { | |
144 | return ((b % a) == 0); | |
2146e26d | 145 | } |
146 | ||
147 | \f | |
148 | ||
48e1416a | 149 | /* Dump into FILE all the data references from DATAREFS. */ |
2146e26d | 150 | |
d129ada0 | 151 | static void |
f1f41a6c | 152 | dump_data_references (FILE *file, vec<data_reference_p> datarefs) |
2146e26d | 153 | { |
154 | unsigned int i; | |
41c7a324 | 155 | struct data_reference *dr; |
156 | ||
f1f41a6c | 157 | FOR_EACH_VEC_ELT (datarefs, i, dr) |
41c7a324 | 158 | dump_data_reference (file, dr); |
2146e26d | 159 | } |
160 | ||
c7d89805 | 161 | /* Unified dump into FILE all the data references from DATAREFS. */ |
162 | ||
163 | DEBUG_FUNCTION void | |
164 | debug (vec<data_reference_p> &ref) | |
165 | { | |
166 | dump_data_references (stderr, ref); | |
167 | } | |
168 | ||
169 | DEBUG_FUNCTION void | |
170 | debug (vec<data_reference_p> *ptr) | |
171 | { | |
172 | if (ptr) | |
173 | debug (*ptr); | |
174 | else | |
175 | fprintf (stderr, "<nil>\n"); | |
176 | } | |
177 | ||
178 | ||
48e1416a | 179 | /* Dump into STDERR all the data references from DATAREFS. */ |
5df4cc8d | 180 | |
4b987fac | 181 | DEBUG_FUNCTION void |
f1f41a6c | 182 | debug_data_references (vec<data_reference_p> datarefs) |
5df4cc8d | 183 | { |
184 | dump_data_references (stderr, datarefs); | |
185 | } | |
186 | ||
5df4cc8d | 187 | /* Print to STDERR the data_reference DR. */ |
188 | ||
4b987fac | 189 | DEBUG_FUNCTION void |
5df4cc8d | 190 | debug_data_reference (struct data_reference *dr) |
191 | { | |
192 | dump_data_reference (stderr, dr); | |
193 | } | |
194 | ||
2146e26d | 195 | /* Dump function for a DATA_REFERENCE structure. */ |
196 | ||
48e1416a | 197 | void |
198 | dump_data_reference (FILE *outf, | |
2146e26d | 199 | struct data_reference *dr) |
200 | { | |
201 | unsigned int i; | |
48e1416a | 202 | |
b40e3432 | 203 | fprintf (outf, "#(Data Ref: \n"); |
204 | fprintf (outf, "# bb: %d \n", gimple_bb (DR_STMT (dr))->index); | |
205 | fprintf (outf, "# stmt: "); | |
75a70cf9 | 206 | print_gimple_stmt (outf, DR_STMT (dr), 0, 0); |
5dc5fe13 | 207 | fprintf (outf, "# ref: "); |
2146e26d | 208 | print_generic_stmt (outf, DR_REF (dr), 0); |
5dc5fe13 | 209 | fprintf (outf, "# base_object: "); |
516849c7 | 210 | print_generic_stmt (outf, DR_BASE_OBJECT (dr), 0); |
48e1416a | 211 | |
2146e26d | 212 | for (i = 0; i < DR_NUM_DIMENSIONS (dr); i++) |
213 | { | |
5dc5fe13 | 214 | fprintf (outf, "# Access function %d: ", i); |
2146e26d | 215 | print_generic_stmt (outf, DR_ACCESS_FN (dr, i), 0); |
216 | } | |
5dc5fe13 | 217 | fprintf (outf, "#)\n"); |
2146e26d | 218 | } |
219 | ||
c7d89805 | 220 | /* Unified dump function for a DATA_REFERENCE structure. */ |
221 | ||
222 | DEBUG_FUNCTION void | |
223 | debug (data_reference &ref) | |
224 | { | |
225 | dump_data_reference (stderr, &ref); | |
226 | } | |
227 | ||
228 | DEBUG_FUNCTION void | |
229 | debug (data_reference *ptr) | |
230 | { | |
231 | if (ptr) | |
232 | debug (*ptr); | |
233 | else | |
234 | fprintf (stderr, "<nil>\n"); | |
235 | } | |
236 | ||
237 | ||
87da4f2e | 238 | /* Dumps the affine function described by FN to the file OUTF. */ |
239 | ||
0d8001a7 | 240 | DEBUG_FUNCTION void |
87da4f2e | 241 | dump_affine_function (FILE *outf, affine_fn fn) |
242 | { | |
243 | unsigned i; | |
244 | tree coef; | |
245 | ||
f1f41a6c | 246 | print_generic_expr (outf, fn[0], TDF_SLIM); |
247 | for (i = 1; fn.iterate (i, &coef); i++) | |
87da4f2e | 248 | { |
249 | fprintf (outf, " + "); | |
250 | print_generic_expr (outf, coef, TDF_SLIM); | |
251 | fprintf (outf, " * x_%u", i); | |
252 | } | |
253 | } | |
254 | ||
255 | /* Dumps the conflict function CF to the file OUTF. */ | |
256 | ||
0d8001a7 | 257 | DEBUG_FUNCTION void |
87da4f2e | 258 | dump_conflict_function (FILE *outf, conflict_function *cf) |
259 | { | |
260 | unsigned i; | |
261 | ||
262 | if (cf->n == NO_DEPENDENCE) | |
984cebc1 | 263 | fprintf (outf, "no dependence"); |
87da4f2e | 264 | else if (cf->n == NOT_KNOWN) |
984cebc1 | 265 | fprintf (outf, "not known"); |
87da4f2e | 266 | else |
267 | { | |
268 | for (i = 0; i < cf->n; i++) | |
269 | { | |
984cebc1 | 270 | if (i != 0) |
271 | fprintf (outf, " "); | |
87da4f2e | 272 | fprintf (outf, "["); |
273 | dump_affine_function (outf, cf->fns[i]); | |
984cebc1 | 274 | fprintf (outf, "]"); |
87da4f2e | 275 | } |
276 | } | |
277 | } | |
278 | ||
bc3c8ad4 | 279 | /* Dump function for a SUBSCRIPT structure. */ |
280 | ||
0d8001a7 | 281 | DEBUG_FUNCTION void |
bc3c8ad4 | 282 | dump_subscript (FILE *outf, struct subscript *subscript) |
283 | { | |
87da4f2e | 284 | conflict_function *cf = SUB_CONFLICTS_IN_A (subscript); |
bc3c8ad4 | 285 | |
286 | fprintf (outf, "\n (subscript \n"); | |
287 | fprintf (outf, " iterations_that_access_an_element_twice_in_A: "); | |
87da4f2e | 288 | dump_conflict_function (outf, cf); |
289 | if (CF_NONTRIVIAL_P (cf)) | |
bc3c8ad4 | 290 | { |
291 | tree last_iteration = SUB_LAST_CONFLICT (subscript); | |
984cebc1 | 292 | fprintf (outf, "\n last_conflict: "); |
293 | print_generic_expr (outf, last_iteration, 0); | |
bc3c8ad4 | 294 | } |
48e1416a | 295 | |
87da4f2e | 296 | cf = SUB_CONFLICTS_IN_B (subscript); |
984cebc1 | 297 | fprintf (outf, "\n iterations_that_access_an_element_twice_in_B: "); |
87da4f2e | 298 | dump_conflict_function (outf, cf); |
299 | if (CF_NONTRIVIAL_P (cf)) | |
bc3c8ad4 | 300 | { |
301 | tree last_iteration = SUB_LAST_CONFLICT (subscript); | |
984cebc1 | 302 | fprintf (outf, "\n last_conflict: "); |
303 | print_generic_expr (outf, last_iteration, 0); | |
bc3c8ad4 | 304 | } |
305 | ||
984cebc1 | 306 | fprintf (outf, "\n (Subscript distance: "); |
307 | print_generic_expr (outf, SUB_DISTANCE (subscript), 0); | |
308 | fprintf (outf, " ))\n"); | |
bc3c8ad4 | 309 | } |
310 | ||
6b421feb | 311 | /* Print the classic direction vector DIRV to OUTF. */ |
312 | ||
0d8001a7 | 313 | DEBUG_FUNCTION void |
6b421feb | 314 | print_direction_vector (FILE *outf, |
315 | lambda_vector dirv, | |
316 | int length) | |
317 | { | |
318 | int eq; | |
319 | ||
320 | for (eq = 0; eq < length; eq++) | |
321 | { | |
bc620c5c | 322 | enum data_dependence_direction dir = ((enum data_dependence_direction) |
323 | dirv[eq]); | |
6b421feb | 324 | |
325 | switch (dir) | |
326 | { | |
327 | case dir_positive: | |
328 | fprintf (outf, " +"); | |
329 | break; | |
330 | case dir_negative: | |
331 | fprintf (outf, " -"); | |
332 | break; | |
333 | case dir_equal: | |
334 | fprintf (outf, " ="); | |
335 | break; | |
336 | case dir_positive_or_equal: | |
337 | fprintf (outf, " +="); | |
338 | break; | |
339 | case dir_positive_or_negative: | |
340 | fprintf (outf, " +-"); | |
341 | break; | |
342 | case dir_negative_or_equal: | |
343 | fprintf (outf, " -="); | |
344 | break; | |
345 | case dir_star: | |
346 | fprintf (outf, " *"); | |
347 | break; | |
348 | default: | |
349 | fprintf (outf, "indep"); | |
350 | break; | |
351 | } | |
352 | } | |
353 | fprintf (outf, "\n"); | |
354 | } | |
355 | ||
b44d1046 | 356 | /* Print a vector of direction vectors. */ |
357 | ||
0d8001a7 | 358 | DEBUG_FUNCTION void |
f1f41a6c | 359 | print_dir_vectors (FILE *outf, vec<lambda_vector> dir_vects, |
b44d1046 | 360 | int length) |
361 | { | |
362 | unsigned j; | |
363 | lambda_vector v; | |
364 | ||
f1f41a6c | 365 | FOR_EACH_VEC_ELT (dir_vects, j, v) |
b44d1046 | 366 | print_direction_vector (outf, v, length); |
367 | } | |
368 | ||
e01f9f1f | 369 | /* Print out a vector VEC of length N to OUTFILE. */ |
370 | ||
0d8001a7 | 371 | DEBUG_FUNCTION void |
e01f9f1f | 372 | print_lambda_vector (FILE * outfile, lambda_vector vector, int n) |
373 | { | |
374 | int i; | |
375 | ||
376 | for (i = 0; i < n; i++) | |
377 | fprintf (outfile, "%3d ", vector[i]); | |
378 | fprintf (outfile, "\n"); | |
379 | } | |
380 | ||
b44d1046 | 381 | /* Print a vector of distance vectors. */ |
382 | ||
0d8001a7 | 383 | DEBUG_FUNCTION void |
f1f41a6c | 384 | print_dist_vectors (FILE *outf, vec<lambda_vector> dist_vects, |
d129ada0 | 385 | int length) |
b44d1046 | 386 | { |
387 | unsigned j; | |
388 | lambda_vector v; | |
389 | ||
f1f41a6c | 390 | FOR_EACH_VEC_ELT (dist_vects, j, v) |
b44d1046 | 391 | print_lambda_vector (outf, v, length); |
392 | } | |
393 | ||
2146e26d | 394 | /* Dump function for a DATA_DEPENDENCE_RELATION structure. */ |
395 | ||
0d8001a7 | 396 | DEBUG_FUNCTION void |
48e1416a | 397 | dump_data_dependence_relation (FILE *outf, |
2146e26d | 398 | struct data_dependence_relation *ddr) |
399 | { | |
2146e26d | 400 | struct data_reference *dra, *drb; |
bc3c8ad4 | 401 | |
bc3c8ad4 | 402 | fprintf (outf, "(Data Dep: \n"); |
801c5610 | 403 | |
4cc34f6b | 404 | if (!ddr || DDR_ARE_DEPENDENT (ddr) == chrec_dont_know) |
405 | { | |
be2e5c02 | 406 | if (ddr) |
407 | { | |
408 | dra = DDR_A (ddr); | |
409 | drb = DDR_B (ddr); | |
410 | if (dra) | |
411 | dump_data_reference (outf, dra); | |
412 | else | |
413 | fprintf (outf, " (nil)\n"); | |
414 | if (drb) | |
415 | dump_data_reference (outf, drb); | |
416 | else | |
417 | fprintf (outf, " (nil)\n"); | |
418 | } | |
4cc34f6b | 419 | fprintf (outf, " (don't know)\n)\n"); |
420 | return; | |
421 | } | |
422 | ||
423 | dra = DDR_A (ddr); | |
424 | drb = DDR_B (ddr); | |
801c5610 | 425 | dump_data_reference (outf, dra); |
426 | dump_data_reference (outf, drb); | |
427 | ||
4cc34f6b | 428 | if (DDR_ARE_DEPENDENT (ddr) == chrec_known) |
2146e26d | 429 | fprintf (outf, " (no dependence)\n"); |
48e1416a | 430 | |
bc3c8ad4 | 431 | else if (DDR_ARE_DEPENDENT (ddr) == NULL_TREE) |
2146e26d | 432 | { |
bc3c8ad4 | 433 | unsigned int i; |
b44d1046 | 434 | struct loop *loopi; |
1532ec98 | 435 | |
2146e26d | 436 | for (i = 0; i < DDR_NUM_SUBSCRIPTS (ddr); i++) |
437 | { | |
2146e26d | 438 | fprintf (outf, " access_fn_A: "); |
439 | print_generic_stmt (outf, DR_ACCESS_FN (dra, i), 0); | |
440 | fprintf (outf, " access_fn_B: "); | |
441 | print_generic_stmt (outf, DR_ACCESS_FN (drb, i), 0); | |
bc3c8ad4 | 442 | dump_subscript (outf, DDR_SUBSCRIPT (ddr, i)); |
2146e26d | 443 | } |
1532ec98 | 444 | |
355572cc | 445 | fprintf (outf, " inner loop index: %d\n", DDR_INNER_LOOP (ddr)); |
b44d1046 | 446 | fprintf (outf, " loop nest: ("); |
f1f41a6c | 447 | FOR_EACH_VEC_ELT (DDR_LOOP_NEST (ddr), i, loopi) |
b44d1046 | 448 | fprintf (outf, "%d ", loopi->num); |
449 | fprintf (outf, ")\n"); | |
450 | ||
1532ec98 | 451 | for (i = 0; i < DDR_NUM_DIST_VECTS (ddr); i++) |
2146e26d | 452 | { |
1532ec98 | 453 | fprintf (outf, " distance_vector: "); |
454 | print_lambda_vector (outf, DDR_DIST_VECT (ddr, i), | |
b44d1046 | 455 | DDR_NB_LOOPS (ddr)); |
bc3c8ad4 | 456 | } |
1532ec98 | 457 | |
458 | for (i = 0; i < DDR_NUM_DIR_VECTS (ddr); i++) | |
bc3c8ad4 | 459 | { |
1532ec98 | 460 | fprintf (outf, " direction_vector: "); |
6b421feb | 461 | print_direction_vector (outf, DDR_DIR_VECT (ddr, i), |
b44d1046 | 462 | DDR_NB_LOOPS (ddr)); |
2146e26d | 463 | } |
2146e26d | 464 | } |
465 | ||
466 | fprintf (outf, ")\n"); | |
467 | } | |
468 | ||
d129ada0 | 469 | /* Debug version. */ |
2146e26d | 470 | |
d129ada0 | 471 | DEBUG_FUNCTION void |
472 | debug_data_dependence_relation (struct data_dependence_relation *ddr) | |
2146e26d | 473 | { |
d129ada0 | 474 | dump_data_dependence_relation (stderr, ddr); |
475 | } | |
48e1416a | 476 | |
d129ada0 | 477 | /* Dump into FILE all the dependence relations from DDRS. */ |
48e1416a | 478 | |
0d8001a7 | 479 | DEBUG_FUNCTION void |
d129ada0 | 480 | dump_data_dependence_relations (FILE *file, |
f1f41a6c | 481 | vec<ddr_p> ddrs) |
d129ada0 | 482 | { |
483 | unsigned int i; | |
484 | struct data_dependence_relation *ddr; | |
48e1416a | 485 | |
f1f41a6c | 486 | FOR_EACH_VEC_ELT (ddrs, i, ddr) |
d129ada0 | 487 | dump_data_dependence_relation (file, ddr); |
488 | } | |
48e1416a | 489 | |
c7d89805 | 490 | DEBUG_FUNCTION void |
491 | debug (vec<ddr_p> &ref) | |
492 | { | |
493 | dump_data_dependence_relations (stderr, ref); | |
494 | } | |
495 | ||
496 | DEBUG_FUNCTION void | |
497 | debug (vec<ddr_p> *ptr) | |
498 | { | |
499 | if (ptr) | |
500 | debug (*ptr); | |
501 | else | |
502 | fprintf (stderr, "<nil>\n"); | |
503 | } | |
504 | ||
505 | ||
d129ada0 | 506 | /* Dump to STDERR all the dependence relations from DDRS. */ |
48e1416a | 507 | |
d129ada0 | 508 | DEBUG_FUNCTION void |
f1f41a6c | 509 | debug_data_dependence_relations (vec<ddr_p> ddrs) |
d129ada0 | 510 | { |
511 | dump_data_dependence_relations (stderr, ddrs); | |
2146e26d | 512 | } |
513 | ||
bc3c8ad4 | 514 | /* Dumps the distance and direction vectors in FILE. DDRS contains |
515 | the dependence relations, and VECT_SIZE is the size of the | |
516 | dependence vectors, or in other words the number of loops in the | |
517 | considered nest. */ | |
518 | ||
0d8001a7 | 519 | DEBUG_FUNCTION void |
f1f41a6c | 520 | dump_dist_dir_vectors (FILE *file, vec<ddr_p> ddrs) |
bc3c8ad4 | 521 | { |
1532ec98 | 522 | unsigned int i, j; |
41c7a324 | 523 | struct data_dependence_relation *ddr; |
524 | lambda_vector v; | |
bc3c8ad4 | 525 | |
f1f41a6c | 526 | FOR_EACH_VEC_ELT (ddrs, i, ddr) |
41c7a324 | 527 | if (DDR_ARE_DEPENDENT (ddr) == NULL_TREE && DDR_AFFINE_P (ddr)) |
528 | { | |
f1f41a6c | 529 | FOR_EACH_VEC_ELT (DDR_DIST_VECTS (ddr), j, v) |
41c7a324 | 530 | { |
531 | fprintf (file, "DISTANCE_V ("); | |
532 | print_lambda_vector (file, v, DDR_NB_LOOPS (ddr)); | |
533 | fprintf (file, ")\n"); | |
534 | } | |
535 | ||
f1f41a6c | 536 | FOR_EACH_VEC_ELT (DDR_DIR_VECTS (ddr), j, v) |
41c7a324 | 537 | { |
538 | fprintf (file, "DIRECTION_V ("); | |
539 | print_direction_vector (file, v, DDR_NB_LOOPS (ddr)); | |
540 | fprintf (file, ")\n"); | |
541 | } | |
542 | } | |
1532ec98 | 543 | |
bc3c8ad4 | 544 | fprintf (file, "\n\n"); |
545 | } | |
546 | ||
547 | /* Dumps the data dependence relations DDRS in FILE. */ | |
548 | ||
0d8001a7 | 549 | DEBUG_FUNCTION void |
f1f41a6c | 550 | dump_ddrs (FILE *file, vec<ddr_p> ddrs) |
bc3c8ad4 | 551 | { |
552 | unsigned int i; | |
41c7a324 | 553 | struct data_dependence_relation *ddr; |
554 | ||
f1f41a6c | 555 | FOR_EACH_VEC_ELT (ddrs, i, ddr) |
41c7a324 | 556 | dump_data_dependence_relation (file, ddr); |
bc3c8ad4 | 557 | |
bc3c8ad4 | 558 | fprintf (file, "\n\n"); |
559 | } | |
560 | ||
d129ada0 | 561 | DEBUG_FUNCTION void |
f1f41a6c | 562 | debug_ddrs (vec<ddr_p> ddrs) |
d129ada0 | 563 | { |
564 | dump_ddrs (stderr, ddrs); | |
565 | } | |
566 | ||
75a70cf9 | 567 | /* Helper function for split_constant_offset. Expresses OP0 CODE OP1 |
568 | (the type of the result is TYPE) as VAR + OFF, where OFF is a nonzero | |
569 | constant of type ssizetype, and returns true. If we cannot do this | |
570 | with OFF nonzero, OFF and VAR are set to NULL_TREE instead and false | |
571 | is returned. */ | |
516849c7 | 572 | |
75a70cf9 | 573 | static bool |
574 | split_constant_offset_1 (tree type, tree op0, enum tree_code code, tree op1, | |
575 | tree *var, tree *off) | |
516849c7 | 576 | { |
80bb306a | 577 | tree var0, var1; |
578 | tree off0, off1; | |
75a70cf9 | 579 | enum tree_code ocode = code; |
516849c7 | 580 | |
75a70cf9 | 581 | *var = NULL_TREE; |
582 | *off = NULL_TREE; | |
516849c7 | 583 | |
0de36bdb | 584 | switch (code) |
516849c7 | 585 | { |
80bb306a | 586 | case INTEGER_CST: |
587 | *var = build_int_cst (type, 0); | |
75a70cf9 | 588 | *off = fold_convert (ssizetype, op0); |
589 | return true; | |
516849c7 | 590 | |
0de36bdb | 591 | case POINTER_PLUS_EXPR: |
75a70cf9 | 592 | ocode = PLUS_EXPR; |
0de36bdb | 593 | /* FALLTHROUGH */ |
80bb306a | 594 | case PLUS_EXPR: |
595 | case MINUS_EXPR: | |
75a70cf9 | 596 | split_constant_offset (op0, &var0, &off0); |
597 | split_constant_offset (op1, &var1, &off1); | |
598 | *var = fold_build2 (code, type, var0, var1); | |
599 | *off = size_binop (ocode, off0, off1); | |
600 | return true; | |
516849c7 | 601 | |
516849c7 | 602 | case MULT_EXPR: |
75a70cf9 | 603 | if (TREE_CODE (op1) != INTEGER_CST) |
604 | return false; | |
80bb306a | 605 | |
75a70cf9 | 606 | split_constant_offset (op0, &var0, &off0); |
607 | *var = fold_build2 (MULT_EXPR, type, var0, op1); | |
608 | *off = size_binop (MULT_EXPR, off0, fold_convert (ssizetype, op1)); | |
609 | return true; | |
516849c7 | 610 | |
80bb306a | 611 | case ADDR_EXPR: |
612 | { | |
75a70cf9 | 613 | tree base, poffset; |
80bb306a | 614 | HOST_WIDE_INT pbitsize, pbitpos; |
3754d046 | 615 | machine_mode pmode; |
292237f3 | 616 | int punsignedp, preversep, pvolatilep; |
516849c7 | 617 | |
97d819e5 | 618 | op0 = TREE_OPERAND (op0, 0); |
292237f3 | 619 | base |
620 | = get_inner_reference (op0, &pbitsize, &pbitpos, &poffset, &pmode, | |
621 | &punsignedp, &preversep, &pvolatilep, false); | |
516849c7 | 622 | |
80bb306a | 623 | if (pbitpos % BITS_PER_UNIT != 0) |
75a70cf9 | 624 | return false; |
80bb306a | 625 | base = build_fold_addr_expr (base); |
626 | off0 = ssize_int (pbitpos / BITS_PER_UNIT); | |
516849c7 | 627 | |
80bb306a | 628 | if (poffset) |
629 | { | |
630 | split_constant_offset (poffset, &poffset, &off1); | |
631 | off0 = size_binop (PLUS_EXPR, off0, off1); | |
55ed4df6 | 632 | if (POINTER_TYPE_P (TREE_TYPE (base))) |
2cc66f2a | 633 | base = fold_build_pointer_plus (base, poffset); |
55ed4df6 | 634 | else |
635 | base = fold_build2 (PLUS_EXPR, TREE_TYPE (base), base, | |
636 | fold_convert (TREE_TYPE (base), poffset)); | |
80bb306a | 637 | } |
638 | ||
69a83716 | 639 | var0 = fold_convert (type, base); |
640 | ||
641 | /* If variable length types are involved, punt, otherwise casts | |
642 | might be converted into ARRAY_REFs in gimplify_conversion. | |
643 | To compute that ARRAY_REF's element size TYPE_SIZE_UNIT, which | |
644 | possibly no longer appears in current GIMPLE, might resurface. | |
645 | This perhaps could run | |
d9659041 | 646 | if (CONVERT_EXPR_P (var0)) |
69a83716 | 647 | { |
648 | gimplify_conversion (&var0); | |
649 | // Attempt to fill in any within var0 found ARRAY_REF's | |
650 | // element size from corresponding op embedded ARRAY_REF, | |
651 | // if unsuccessful, just punt. | |
652 | } */ | |
653 | while (POINTER_TYPE_P (type)) | |
654 | type = TREE_TYPE (type); | |
655 | if (int_size_in_bytes (type) < 0) | |
75a70cf9 | 656 | return false; |
69a83716 | 657 | |
658 | *var = var0; | |
80bb306a | 659 | *off = off0; |
75a70cf9 | 660 | return true; |
80bb306a | 661 | } |
516849c7 | 662 | |
a091367a | 663 | case SSA_NAME: |
664 | { | |
520f63be | 665 | if (SSA_NAME_OCCURS_IN_ABNORMAL_PHI (op0)) |
666 | return false; | |
667 | ||
42acab1c | 668 | gimple *def_stmt = SSA_NAME_DEF_STMT (op0); |
75a70cf9 | 669 | enum tree_code subcode; |
a091367a | 670 | |
75a70cf9 | 671 | if (gimple_code (def_stmt) != GIMPLE_ASSIGN) |
672 | return false; | |
673 | ||
674 | var0 = gimple_assign_rhs1 (def_stmt); | |
675 | subcode = gimple_assign_rhs_code (def_stmt); | |
676 | var1 = gimple_assign_rhs2 (def_stmt); | |
677 | ||
678 | return split_constant_offset_1 (type, var0, subcode, var1, var, off); | |
a091367a | 679 | } |
be2e5c02 | 680 | CASE_CONVERT: |
681 | { | |
682 | /* We must not introduce undefined overflow, and we must not change the value. | |
683 | Hence we're okay if the inner type doesn't overflow to start with | |
684 | (pointer or signed), the outer type also is an integer or pointer | |
685 | and the outer precision is at least as large as the inner. */ | |
686 | tree itype = TREE_TYPE (op0); | |
687 | if ((POINTER_TYPE_P (itype) | |
688 | || (INTEGRAL_TYPE_P (itype) && TYPE_OVERFLOW_UNDEFINED (itype))) | |
689 | && TYPE_PRECISION (type) >= TYPE_PRECISION (itype) | |
690 | && (POINTER_TYPE_P (type) || INTEGRAL_TYPE_P (type))) | |
691 | { | |
692 | split_constant_offset (op0, &var0, off); | |
693 | *var = fold_convert (type, var0); | |
694 | return true; | |
695 | } | |
696 | return false; | |
697 | } | |
a091367a | 698 | |
516849c7 | 699 | default: |
75a70cf9 | 700 | return false; |
516849c7 | 701 | } |
75a70cf9 | 702 | } |
703 | ||
704 | /* Expresses EXP as VAR + OFF, where off is a constant. The type of OFF | |
705 | will be ssizetype. */ | |
706 | ||
707 | void | |
708 | split_constant_offset (tree exp, tree *var, tree *off) | |
709 | { | |
710 | tree type = TREE_TYPE (exp), otype, op0, op1, e, o; | |
711 | enum tree_code code; | |
516849c7 | 712 | |
75a70cf9 | 713 | *var = exp; |
80bb306a | 714 | *off = ssize_int (0); |
75a70cf9 | 715 | STRIP_NOPS (exp); |
716 | ||
32246fd4 | 717 | if (tree_is_chrec (exp) |
718 | || get_gimple_rhs_class (TREE_CODE (exp)) == GIMPLE_TERNARY_RHS) | |
75a70cf9 | 719 | return; |
720 | ||
721 | otype = TREE_TYPE (exp); | |
722 | code = TREE_CODE (exp); | |
723 | extract_ops_from_tree (exp, &code, &op0, &op1); | |
724 | if (split_constant_offset_1 (otype, op0, code, op1, &e, &o)) | |
725 | { | |
726 | *var = fold_convert (type, e); | |
727 | *off = o; | |
728 | } | |
516849c7 | 729 | } |
730 | ||
80bb306a | 731 | /* Returns the address ADDR of an object in a canonical shape (without nop |
732 | casts, and with type of pointer to the object). */ | |
516849c7 | 733 | |
734 | static tree | |
80bb306a | 735 | canonicalize_base_object_address (tree addr) |
516849c7 | 736 | { |
ad4a85ad | 737 | tree orig = addr; |
738 | ||
80bb306a | 739 | STRIP_NOPS (addr); |
516849c7 | 740 | |
ad4a85ad | 741 | /* The base address may be obtained by casting from integer, in that case |
742 | keep the cast. */ | |
743 | if (!POINTER_TYPE_P (TREE_TYPE (addr))) | |
744 | return orig; | |
745 | ||
80bb306a | 746 | if (TREE_CODE (addr) != ADDR_EXPR) |
747 | return addr; | |
516849c7 | 748 | |
80bb306a | 749 | return build_fold_addr_expr (TREE_OPERAND (addr, 0)); |
516849c7 | 750 | } |
751 | ||
48e1416a | 752 | /* Analyzes the behavior of the memory reference DR in the innermost loop or |
0c257e4c | 753 | basic block that contains it. Returns true if analysis succeed or false |
37545e54 | 754 | otherwise. */ |
516849c7 | 755 | |
880734c8 | 756 | bool |
0c257e4c | 757 | dr_analyze_innermost (struct data_reference *dr, struct loop *nest) |
516849c7 | 758 | { |
42acab1c | 759 | gimple *stmt = DR_STMT (dr); |
80bb306a | 760 | struct loop *loop = loop_containing_stmt (stmt); |
761 | tree ref = DR_REF (dr); | |
516849c7 | 762 | HOST_WIDE_INT pbitsize, pbitpos; |
80bb306a | 763 | tree base, poffset; |
3754d046 | 764 | machine_mode pmode; |
292237f3 | 765 | int punsignedp, preversep, pvolatilep; |
80bb306a | 766 | affine_iv base_iv, offset_iv; |
767 | tree init, dinit, step; | |
37545e54 | 768 | bool in_loop = (loop && loop->num); |
80bb306a | 769 | |
770 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
771 | fprintf (dump_file, "analyze_innermost: "); | |
516849c7 | 772 | |
292237f3 | 773 | base = get_inner_reference (ref, &pbitsize, &pbitpos, &poffset, &pmode, |
774 | &punsignedp, &preversep, &pvolatilep, false); | |
80bb306a | 775 | gcc_assert (base != NULL_TREE); |
516849c7 | 776 | |
80bb306a | 777 | if (pbitpos % BITS_PER_UNIT != 0) |
516849c7 | 778 | { |
80bb306a | 779 | if (dump_file && (dump_flags & TDF_DETAILS)) |
780 | fprintf (dump_file, "failed: bit offset alignment.\n"); | |
880734c8 | 781 | return false; |
80bb306a | 782 | } |
516849c7 | 783 | |
292237f3 | 784 | if (preversep) |
785 | { | |
786 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
787 | fprintf (dump_file, "failed: reverse storage order.\n"); | |
788 | return false; | |
789 | } | |
790 | ||
182cf5a9 | 791 | if (TREE_CODE (base) == MEM_REF) |
792 | { | |
793 | if (!integer_zerop (TREE_OPERAND (base, 1))) | |
794 | { | |
5de9d3ed | 795 | offset_int moff = mem_ref_offset (base); |
e913b5cd | 796 | tree mofft = wide_int_to_tree (sizetype, moff); |
182cf5a9 | 797 | if (!poffset) |
7667733b | 798 | poffset = mofft; |
182cf5a9 | 799 | else |
7667733b | 800 | poffset = size_binop (PLUS_EXPR, poffset, mofft); |
182cf5a9 | 801 | } |
802 | base = TREE_OPERAND (base, 0); | |
803 | } | |
804 | else | |
805 | base = build_fold_addr_expr (base); | |
0c257e4c | 806 | |
37545e54 | 807 | if (in_loop) |
80bb306a | 808 | { |
48e1416a | 809 | if (!simple_iv (loop, loop_containing_stmt (stmt), base, &base_iv, |
f634c3e9 | 810 | nest ? true : false)) |
37545e54 | 811 | { |
0c257e4c | 812 | if (nest) |
813 | { | |
814 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
815 | fprintf (dump_file, "failed: evolution of base is not" | |
816 | " affine.\n"); | |
817 | return false; | |
818 | } | |
819 | else | |
820 | { | |
821 | base_iv.base = base; | |
822 | base_iv.step = ssize_int (0); | |
823 | base_iv.no_overflow = true; | |
824 | } | |
37545e54 | 825 | } |
826 | } | |
827 | else | |
828 | { | |
829 | base_iv.base = base; | |
830 | base_iv.step = ssize_int (0); | |
831 | base_iv.no_overflow = true; | |
80bb306a | 832 | } |
37545e54 | 833 | |
dd2644cf | 834 | if (!poffset) |
80bb306a | 835 | { |
836 | offset_iv.base = ssize_int (0); | |
837 | offset_iv.step = ssize_int (0); | |
838 | } | |
dd2644cf | 839 | else |
80bb306a | 840 | { |
dd2644cf | 841 | if (!in_loop) |
842 | { | |
843 | offset_iv.base = poffset; | |
844 | offset_iv.step = ssize_int (0); | |
845 | } | |
846 | else if (!simple_iv (loop, loop_containing_stmt (stmt), | |
f634c3e9 | 847 | poffset, &offset_iv, |
848 | nest ? true : false)) | |
dd2644cf | 849 | { |
0c257e4c | 850 | if (nest) |
851 | { | |
852 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
853 | fprintf (dump_file, "failed: evolution of offset is not" | |
854 | " affine.\n"); | |
855 | return false; | |
856 | } | |
857 | else | |
858 | { | |
859 | offset_iv.base = poffset; | |
860 | offset_iv.step = ssize_int (0); | |
861 | } | |
dd2644cf | 862 | } |
80bb306a | 863 | } |
516849c7 | 864 | |
80bb306a | 865 | init = ssize_int (pbitpos / BITS_PER_UNIT); |
866 | split_constant_offset (base_iv.base, &base_iv.base, &dinit); | |
867 | init = size_binop (PLUS_EXPR, init, dinit); | |
868 | split_constant_offset (offset_iv.base, &offset_iv.base, &dinit); | |
869 | init = size_binop (PLUS_EXPR, init, dinit); | |
516849c7 | 870 | |
80bb306a | 871 | step = size_binop (PLUS_EXPR, |
872 | fold_convert (ssizetype, base_iv.step), | |
873 | fold_convert (ssizetype, offset_iv.step)); | |
516849c7 | 874 | |
80bb306a | 875 | DR_BASE_ADDRESS (dr) = canonicalize_base_object_address (base_iv.base); |
516849c7 | 876 | |
80bb306a | 877 | DR_OFFSET (dr) = fold_convert (ssizetype, offset_iv.base); |
878 | DR_INIT (dr) = init; | |
879 | DR_STEP (dr) = step; | |
516849c7 | 880 | |
80bb306a | 881 | DR_ALIGNED_TO (dr) = size_int (highest_pow2_factor (offset_iv.base)); |
516849c7 | 882 | |
80bb306a | 883 | if (dump_file && (dump_flags & TDF_DETAILS)) |
884 | fprintf (dump_file, "success.\n"); | |
880734c8 | 885 | |
886 | return true; | |
80bb306a | 887 | } |
516849c7 | 888 | |
80bb306a | 889 | /* Determines the base object and the list of indices of memory reference |
221a697e | 890 | DR, analyzed in LOOP and instantiated in loop nest NEST. */ |
516849c7 | 891 | |
80bb306a | 892 | static void |
221a697e | 893 | dr_analyze_indices (struct data_reference *dr, loop_p nest, loop_p loop) |
80bb306a | 894 | { |
1e094109 | 895 | vec<tree> access_fns = vNULL; |
95539e1d | 896 | tree ref, op; |
93fc59e8 | 897 | tree base, off, access_fn; |
898 | basic_block before_loop; | |
48e1416a | 899 | |
93fc59e8 | 900 | /* If analyzing a basic-block there are no indices to analyze |
901 | and thus no access functions. */ | |
5fc88ffd | 902 | if (!nest) |
903 | { | |
93fc59e8 | 904 | DR_BASE_OBJECT (dr) = DR_REF (dr); |
f1f41a6c | 905 | DR_ACCESS_FNS (dr).create (0); |
5fc88ffd | 906 | return; |
907 | } | |
908 | ||
95539e1d | 909 | ref = DR_REF (dr); |
5fc88ffd | 910 | before_loop = block_before_loop (nest); |
48e1416a | 911 | |
93fc59e8 | 912 | /* REALPART_EXPR and IMAGPART_EXPR can be handled like accesses |
913 | into a two element array with a constant index. The base is | |
914 | then just the immediate underlying object. */ | |
915 | if (TREE_CODE (ref) == REALPART_EXPR) | |
916 | { | |
917 | ref = TREE_OPERAND (ref, 0); | |
f1f41a6c | 918 | access_fns.safe_push (integer_zero_node); |
93fc59e8 | 919 | } |
920 | else if (TREE_CODE (ref) == IMAGPART_EXPR) | |
921 | { | |
922 | ref = TREE_OPERAND (ref, 0); | |
f1f41a6c | 923 | access_fns.safe_push (integer_one_node); |
93fc59e8 | 924 | } |
925 | ||
bc4113a9 | 926 | /* Analyze access functions of dimensions we know to be independent. */ |
95539e1d | 927 | while (handled_component_p (ref)) |
516849c7 | 928 | { |
95539e1d | 929 | if (TREE_CODE (ref) == ARRAY_REF) |
516849c7 | 930 | { |
95539e1d | 931 | op = TREE_OPERAND (ref, 1); |
5fc88ffd | 932 | access_fn = analyze_scalar_evolution (loop, op); |
933 | access_fn = instantiate_scev (before_loop, loop, access_fn); | |
f1f41a6c | 934 | access_fns.safe_push (access_fn); |
bc4113a9 | 935 | } |
95539e1d | 936 | else if (TREE_CODE (ref) == COMPONENT_REF |
937 | && TREE_CODE (TREE_TYPE (TREE_OPERAND (ref, 0))) == RECORD_TYPE) | |
938 | { | |
939 | /* For COMPONENT_REFs of records (but not unions!) use the | |
940 | FIELD_DECL offset as constant access function so we can | |
941 | disambiguate a[i].f1 and a[i].f2. */ | |
942 | tree off = component_ref_field_offset (ref); | |
943 | off = size_binop (PLUS_EXPR, | |
944 | size_binop (MULT_EXPR, | |
945 | fold_convert (bitsizetype, off), | |
946 | bitsize_int (BITS_PER_UNIT)), | |
947 | DECL_FIELD_BIT_OFFSET (TREE_OPERAND (ref, 1))); | |
f1f41a6c | 948 | access_fns.safe_push (off); |
95539e1d | 949 | } |
950 | else | |
951 | /* If we have an unhandled component we could not translate | |
952 | to an access function stop analyzing. We have determined | |
953 | our base object in this case. */ | |
954 | break; | |
48e1416a | 955 | |
95539e1d | 956 | ref = TREE_OPERAND (ref, 0); |
516849c7 | 957 | } |
958 | ||
0f00ec21 | 959 | /* If the address operand of a MEM_REF base has an evolution in the |
960 | analyzed nest, add it as an additional independent access-function. */ | |
95539e1d | 961 | if (TREE_CODE (ref) == MEM_REF) |
516849c7 | 962 | { |
95539e1d | 963 | op = TREE_OPERAND (ref, 0); |
80bb306a | 964 | access_fn = analyze_scalar_evolution (loop, op); |
089aa668 | 965 | access_fn = instantiate_scev (before_loop, loop, access_fn); |
0f00ec21 | 966 | if (TREE_CODE (access_fn) == POLYNOMIAL_CHREC) |
9488020c | 967 | { |
a927c5fa | 968 | tree orig_type; |
95539e1d | 969 | tree memoff = TREE_OPERAND (ref, 1); |
0f00ec21 | 970 | base = initial_condition (access_fn); |
a927c5fa | 971 | orig_type = TREE_TYPE (base); |
972 | STRIP_USELESS_TYPE_CONVERSION (base); | |
0f00ec21 | 973 | split_constant_offset (base, &base, &off); |
1e074e77 | 974 | STRIP_USELESS_TYPE_CONVERSION (base); |
0f00ec21 | 975 | /* Fold the MEM_REF offset into the evolutions initial |
976 | value to make more bases comparable. */ | |
95539e1d | 977 | if (!integer_zerop (memoff)) |
0f00ec21 | 978 | { |
979 | off = size_binop (PLUS_EXPR, off, | |
95539e1d | 980 | fold_convert (ssizetype, memoff)); |
981 | memoff = build_int_cst (TREE_TYPE (memoff), 0); | |
0f00ec21 | 982 | } |
43085be3 | 983 | /* Adjust the offset so it is a multiple of the access type |
984 | size and thus we separate bases that can possibly be used | |
985 | to produce partial overlaps (which the access_fn machinery | |
986 | cannot handle). */ | |
987 | wide_int rem; | |
988 | if (TYPE_SIZE_UNIT (TREE_TYPE (ref)) | |
989 | && TREE_CODE (TYPE_SIZE_UNIT (TREE_TYPE (ref))) == INTEGER_CST | |
990 | && !integer_zerop (TYPE_SIZE_UNIT (TREE_TYPE (ref)))) | |
991 | rem = wi::mod_trunc (off, TYPE_SIZE_UNIT (TREE_TYPE (ref)), SIGNED); | |
992 | else | |
993 | /* If we can't compute the remainder simply force the initial | |
994 | condition to zero. */ | |
995 | rem = off; | |
996 | off = wide_int_to_tree (ssizetype, wi::sub (off, rem)); | |
997 | memoff = wide_int_to_tree (TREE_TYPE (memoff), rem); | |
998 | /* And finally replace the initial condition. */ | |
0f00ec21 | 999 | access_fn = chrec_replace_initial_condition |
a927c5fa | 1000 | (access_fn, fold_convert (orig_type, off)); |
95539e1d | 1001 | /* ??? This is still not a suitable base object for |
1002 | dr_may_alias_p - the base object needs to be an | |
1003 | access that covers the object as whole. With | |
1004 | an evolution in the pointer this cannot be | |
1005 | guaranteed. | |
1006 | As a band-aid, mark the access so we can special-case | |
1007 | it in dr_may_alias_p. */ | |
62b0a610 | 1008 | tree old = ref; |
95539e1d | 1009 | ref = fold_build2_loc (EXPR_LOCATION (ref), |
1010 | MEM_REF, TREE_TYPE (ref), | |
1011 | base, memoff); | |
62b0a610 | 1012 | MR_DEPENDENCE_CLIQUE (ref) = MR_DEPENDENCE_CLIQUE (old); |
1013 | MR_DEPENDENCE_BASE (ref) = MR_DEPENDENCE_BASE (old); | |
f146c442 | 1014 | DR_UNCONSTRAINED_BASE (dr) = true; |
f1f41a6c | 1015 | access_fns.safe_push (access_fn); |
9488020c | 1016 | } |
516849c7 | 1017 | } |
95539e1d | 1018 | else if (DECL_P (ref)) |
1019 | { | |
1020 | /* Canonicalize DR_BASE_OBJECT to MEM_REF form. */ | |
1021 | ref = build2 (MEM_REF, TREE_TYPE (ref), | |
1022 | build_fold_addr_expr (ref), | |
1023 | build_int_cst (reference_alias_ptr_type (ref), 0)); | |
1024 | } | |
516849c7 | 1025 | |
80bb306a | 1026 | DR_BASE_OBJECT (dr) = ref; |
1027 | DR_ACCESS_FNS (dr) = access_fns; | |
516849c7 | 1028 | } |
1029 | ||
80bb306a | 1030 | /* Extracts the alias analysis information from the memory reference DR. */ |
516849c7 | 1031 | |
80bb306a | 1032 | static void |
1033 | dr_analyze_alias (struct data_reference *dr) | |
516849c7 | 1034 | { |
80bb306a | 1035 | tree ref = DR_REF (dr); |
dd277d48 | 1036 | tree base = get_base_address (ref), addr; |
1037 | ||
182cf5a9 | 1038 | if (INDIRECT_REF_P (base) |
1039 | || TREE_CODE (base) == MEM_REF) | |
80bb306a | 1040 | { |
1041 | addr = TREE_OPERAND (base, 0); | |
1042 | if (TREE_CODE (addr) == SSA_NAME) | |
dd277d48 | 1043 | DR_PTR_INFO (dr) = SSA_NAME_PTR_INFO (addr); |
80bb306a | 1044 | } |
80bb306a | 1045 | } |
516849c7 | 1046 | |
80bb306a | 1047 | /* Frees data reference DR. */ |
31e389a2 | 1048 | |
801c5610 | 1049 | void |
31e389a2 | 1050 | free_data_ref (data_reference_p dr) |
1051 | { | |
f1f41a6c | 1052 | DR_ACCESS_FNS (dr).release (); |
31e389a2 | 1053 | free (dr); |
1054 | } | |
516849c7 | 1055 | |
80bb306a | 1056 | /* Analyzes memory reference MEMREF accessed in STMT. The reference |
1057 | is read if IS_READ is true, write otherwise. Returns the | |
221a697e | 1058 | data_reference description of MEMREF. NEST is the outermost loop |
1059 | in which the reference should be instantiated, LOOP is the loop in | |
1060 | which the data reference should be analyzed. */ | |
516849c7 | 1061 | |
5c205353 | 1062 | struct data_reference * |
42acab1c | 1063 | create_data_ref (loop_p nest, loop_p loop, tree memref, gimple *stmt, |
221a697e | 1064 | bool is_read) |
516849c7 | 1065 | { |
80bb306a | 1066 | struct data_reference *dr; |
6b421feb | 1067 | |
80bb306a | 1068 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6b421feb | 1069 | { |
80bb306a | 1070 | fprintf (dump_file, "Creating dr for "); |
1071 | print_generic_expr (dump_file, memref, TDF_SLIM); | |
1072 | fprintf (dump_file, "\n"); | |
6b421feb | 1073 | } |
f84a688a | 1074 | |
80bb306a | 1075 | dr = XCNEW (struct data_reference); |
1076 | DR_STMT (dr) = stmt; | |
1077 | DR_REF (dr) = memref; | |
1078 | DR_IS_READ (dr) = is_read; | |
516849c7 | 1079 | |
0c257e4c | 1080 | dr_analyze_innermost (dr, nest); |
221a697e | 1081 | dr_analyze_indices (dr, nest, loop); |
80bb306a | 1082 | dr_analyze_alias (dr); |
516849c7 | 1083 | |
1084 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
1085 | { | |
bc4113a9 | 1086 | unsigned i; |
80bb306a | 1087 | fprintf (dump_file, "\tbase_address: "); |
516849c7 | 1088 | print_generic_expr (dump_file, DR_BASE_ADDRESS (dr), TDF_SLIM); |
1089 | fprintf (dump_file, "\n\toffset from base address: "); | |
1090 | print_generic_expr (dump_file, DR_OFFSET (dr), TDF_SLIM); | |
1091 | fprintf (dump_file, "\n\tconstant offset from base address: "); | |
1092 | print_generic_expr (dump_file, DR_INIT (dr), TDF_SLIM); | |
516849c7 | 1093 | fprintf (dump_file, "\n\tstep: "); |
1094 | print_generic_expr (dump_file, DR_STEP (dr), TDF_SLIM); | |
80bb306a | 1095 | fprintf (dump_file, "\n\taligned to: "); |
1096 | print_generic_expr (dump_file, DR_ALIGNED_TO (dr), TDF_SLIM); | |
1097 | fprintf (dump_file, "\n\tbase_object: "); | |
1098 | print_generic_expr (dump_file, DR_BASE_OBJECT (dr), TDF_SLIM); | |
516849c7 | 1099 | fprintf (dump_file, "\n"); |
bc4113a9 | 1100 | for (i = 0; i < DR_NUM_DIMENSIONS (dr); i++) |
1101 | { | |
1102 | fprintf (dump_file, "\tAccess function %d: ", i); | |
1103 | print_generic_stmt (dump_file, DR_ACCESS_FN (dr, i), TDF_SLIM); | |
1104 | } | |
80bb306a | 1105 | } |
1106 | ||
48e1416a | 1107 | return dr; |
516849c7 | 1108 | } |
1109 | ||
ec611e12 | 1110 | /* Check if OFFSET1 and OFFSET2 (DR_OFFSETs of some data-refs) are identical |
1111 | expressions. */ | |
1112 | static bool | |
1113 | dr_equal_offsets_p1 (tree offset1, tree offset2) | |
1114 | { | |
1115 | bool res; | |
1116 | ||
1117 | STRIP_NOPS (offset1); | |
1118 | STRIP_NOPS (offset2); | |
1119 | ||
1120 | if (offset1 == offset2) | |
1121 | return true; | |
1122 | ||
1123 | if (TREE_CODE (offset1) != TREE_CODE (offset2) | |
1124 | || (!BINARY_CLASS_P (offset1) && !UNARY_CLASS_P (offset1))) | |
1125 | return false; | |
1126 | ||
1127 | res = dr_equal_offsets_p1 (TREE_OPERAND (offset1, 0), | |
1128 | TREE_OPERAND (offset2, 0)); | |
1129 | ||
1130 | if (!res || !BINARY_CLASS_P (offset1)) | |
1131 | return res; | |
1132 | ||
1133 | res = dr_equal_offsets_p1 (TREE_OPERAND (offset1, 1), | |
1134 | TREE_OPERAND (offset2, 1)); | |
1135 | ||
1136 | return res; | |
1137 | } | |
1138 | ||
1139 | /* Check if DRA and DRB have equal offsets. */ | |
1140 | bool | |
1141 | dr_equal_offsets_p (struct data_reference *dra, | |
1142 | struct data_reference *drb) | |
1143 | { | |
1144 | tree offset1, offset2; | |
1145 | ||
1146 | offset1 = DR_OFFSET (dra); | |
1147 | offset2 = DR_OFFSET (drb); | |
1148 | ||
1149 | return dr_equal_offsets_p1 (offset1, offset2); | |
1150 | } | |
1151 | ||
87da4f2e | 1152 | /* Returns true if FNA == FNB. */ |
1153 | ||
1154 | static bool | |
1155 | affine_function_equal_p (affine_fn fna, affine_fn fnb) | |
1156 | { | |
f1f41a6c | 1157 | unsigned i, n = fna.length (); |
516849c7 | 1158 | |
f1f41a6c | 1159 | if (n != fnb.length ()) |
de33470d | 1160 | return false; |
bc3c8ad4 | 1161 | |
87da4f2e | 1162 | for (i = 0; i < n; i++) |
f1f41a6c | 1163 | if (!operand_equal_p (fna[i], fnb[i], 0)) |
87da4f2e | 1164 | return false; |
1165 | ||
1166 | return true; | |
1167 | } | |
1168 | ||
1169 | /* If all the functions in CF are the same, returns one of them, | |
1170 | otherwise returns NULL. */ | |
1171 | ||
1172 | static affine_fn | |
1173 | common_affine_function (conflict_function *cf) | |
bc3c8ad4 | 1174 | { |
87da4f2e | 1175 | unsigned i; |
1176 | affine_fn comm; | |
1177 | ||
1178 | if (!CF_NONTRIVIAL_P (cf)) | |
9af5ce0c | 1179 | return affine_fn (); |
87da4f2e | 1180 | |
1181 | comm = cf->fns[0]; | |
1182 | ||
1183 | for (i = 1; i < cf->n; i++) | |
1184 | if (!affine_function_equal_p (comm, cf->fns[i])) | |
9af5ce0c | 1185 | return affine_fn (); |
87da4f2e | 1186 | |
1187 | return comm; | |
1188 | } | |
bc3c8ad4 | 1189 | |
87da4f2e | 1190 | /* Returns the base of the affine function FN. */ |
1191 | ||
1192 | static tree | |
1193 | affine_function_base (affine_fn fn) | |
1194 | { | |
f1f41a6c | 1195 | return fn[0]; |
87da4f2e | 1196 | } |
1197 | ||
1198 | /* Returns true if FN is a constant. */ | |
1199 | ||
1200 | static bool | |
1201 | affine_function_constant_p (affine_fn fn) | |
1202 | { | |
1203 | unsigned i; | |
1204 | tree coef; | |
1205 | ||
f1f41a6c | 1206 | for (i = 1; fn.iterate (i, &coef); i++) |
87da4f2e | 1207 | if (!integer_zerop (coef)) |
f84a688a | 1208 | return false; |
1209 | ||
bc3c8ad4 | 1210 | return true; |
1211 | } | |
1212 | ||
2a21a5df | 1213 | /* Returns true if FN is the zero constant function. */ |
1214 | ||
1215 | static bool | |
1216 | affine_function_zero_p (affine_fn fn) | |
1217 | { | |
1218 | return (integer_zerop (affine_function_base (fn)) | |
1219 | && affine_function_constant_p (fn)); | |
1220 | } | |
1221 | ||
9a3dad1c | 1222 | /* Returns a signed integer type with the largest precision from TA |
1223 | and TB. */ | |
1224 | ||
1225 | static tree | |
1226 | signed_type_for_types (tree ta, tree tb) | |
1227 | { | |
1228 | if (TYPE_PRECISION (ta) > TYPE_PRECISION (tb)) | |
1229 | return signed_type_for (ta); | |
1230 | else | |
1231 | return signed_type_for (tb); | |
1232 | } | |
1233 | ||
87da4f2e | 1234 | /* Applies operation OP on affine functions FNA and FNB, and returns the |
1235 | result. */ | |
1236 | ||
1237 | static affine_fn | |
1238 | affine_fn_op (enum tree_code op, affine_fn fna, affine_fn fnb) | |
1239 | { | |
1240 | unsigned i, n, m; | |
1241 | affine_fn ret; | |
1242 | tree coef; | |
1243 | ||
f1f41a6c | 1244 | if (fnb.length () > fna.length ()) |
87da4f2e | 1245 | { |
f1f41a6c | 1246 | n = fna.length (); |
1247 | m = fnb.length (); | |
87da4f2e | 1248 | } |
1249 | else | |
1250 | { | |
f1f41a6c | 1251 | n = fnb.length (); |
1252 | m = fna.length (); | |
87da4f2e | 1253 | } |
1254 | ||
f1f41a6c | 1255 | ret.create (m); |
87da4f2e | 1256 | for (i = 0; i < n; i++) |
9a3dad1c | 1257 | { |
f1f41a6c | 1258 | tree type = signed_type_for_types (TREE_TYPE (fna[i]), |
1259 | TREE_TYPE (fnb[i])); | |
1260 | ret.quick_push (fold_build2 (op, type, fna[i], fnb[i])); | |
9a3dad1c | 1261 | } |
87da4f2e | 1262 | |
f1f41a6c | 1263 | for (; fna.iterate (i, &coef); i++) |
1264 | ret.quick_push (fold_build2 (op, signed_type_for (TREE_TYPE (coef)), | |
87da4f2e | 1265 | coef, integer_zero_node)); |
f1f41a6c | 1266 | for (; fnb.iterate (i, &coef); i++) |
1267 | ret.quick_push (fold_build2 (op, signed_type_for (TREE_TYPE (coef)), | |
87da4f2e | 1268 | integer_zero_node, coef)); |
1269 | ||
1270 | return ret; | |
1271 | } | |
1272 | ||
1273 | /* Returns the sum of affine functions FNA and FNB. */ | |
1274 | ||
1275 | static affine_fn | |
1276 | affine_fn_plus (affine_fn fna, affine_fn fnb) | |
1277 | { | |
1278 | return affine_fn_op (PLUS_EXPR, fna, fnb); | |
1279 | } | |
1280 | ||
1281 | /* Returns the difference of affine functions FNA and FNB. */ | |
1282 | ||
1283 | static affine_fn | |
1284 | affine_fn_minus (affine_fn fna, affine_fn fnb) | |
1285 | { | |
1286 | return affine_fn_op (MINUS_EXPR, fna, fnb); | |
1287 | } | |
1288 | ||
1289 | /* Frees affine function FN. */ | |
1290 | ||
1291 | static void | |
1292 | affine_fn_free (affine_fn fn) | |
1293 | { | |
f1f41a6c | 1294 | fn.release (); |
87da4f2e | 1295 | } |
1296 | ||
bc3c8ad4 | 1297 | /* Determine for each subscript in the data dependence relation DDR |
1298 | the distance. */ | |
2146e26d | 1299 | |
6b421feb | 1300 | static void |
bc3c8ad4 | 1301 | compute_subscript_distance (struct data_dependence_relation *ddr) |
2146e26d | 1302 | { |
87da4f2e | 1303 | conflict_function *cf_a, *cf_b; |
1304 | affine_fn fn_a, fn_b, diff; | |
1305 | ||
2146e26d | 1306 | if (DDR_ARE_DEPENDENT (ddr) == NULL_TREE) |
1307 | { | |
1308 | unsigned int i; | |
48e1416a | 1309 | |
2146e26d | 1310 | for (i = 0; i < DDR_NUM_SUBSCRIPTS (ddr); i++) |
1311 | { | |
2146e26d | 1312 | struct subscript *subscript; |
48e1416a | 1313 | |
2146e26d | 1314 | subscript = DDR_SUBSCRIPT (ddr, i); |
87da4f2e | 1315 | cf_a = SUB_CONFLICTS_IN_A (subscript); |
1316 | cf_b = SUB_CONFLICTS_IN_B (subscript); | |
bc3c8ad4 | 1317 | |
87da4f2e | 1318 | fn_a = common_affine_function (cf_a); |
1319 | fn_b = common_affine_function (cf_b); | |
f1f41a6c | 1320 | if (!fn_a.exists () || !fn_b.exists ()) |
bc3c8ad4 | 1321 | { |
87da4f2e | 1322 | SUB_DISTANCE (subscript) = chrec_dont_know; |
1323 | return; | |
bc3c8ad4 | 1324 | } |
87da4f2e | 1325 | diff = affine_fn_minus (fn_a, fn_b); |
48e1416a | 1326 | |
87da4f2e | 1327 | if (affine_function_constant_p (diff)) |
1328 | SUB_DISTANCE (subscript) = affine_function_base (diff); | |
2146e26d | 1329 | else |
1330 | SUB_DISTANCE (subscript) = chrec_dont_know; | |
87da4f2e | 1331 | |
1332 | affine_fn_free (diff); | |
2146e26d | 1333 | } |
1334 | } | |
1335 | } | |
1336 | ||
87da4f2e | 1337 | /* Returns the conflict function for "unknown". */ |
1338 | ||
1339 | static conflict_function * | |
1340 | conflict_fn_not_known (void) | |
1341 | { | |
1342 | conflict_function *fn = XCNEW (conflict_function); | |
1343 | fn->n = NOT_KNOWN; | |
1344 | ||
1345 | return fn; | |
1346 | } | |
1347 | ||
1348 | /* Returns the conflict function for "independent". */ | |
1349 | ||
1350 | static conflict_function * | |
1351 | conflict_fn_no_dependence (void) | |
1352 | { | |
1353 | conflict_function *fn = XCNEW (conflict_function); | |
1354 | fn->n = NO_DEPENDENCE; | |
1355 | ||
1356 | return fn; | |
1357 | } | |
1358 | ||
80bb306a | 1359 | /* Returns true if the address of OBJ is invariant in LOOP. */ |
1360 | ||
1361 | static bool | |
7ecb5bb2 | 1362 | object_address_invariant_in_loop_p (const struct loop *loop, const_tree obj) |
80bb306a | 1363 | { |
1364 | while (handled_component_p (obj)) | |
1365 | { | |
1366 | if (TREE_CODE (obj) == ARRAY_REF) | |
1367 | { | |
1368 | /* Index of the ARRAY_REF was zeroed in analyze_indices, thus we only | |
1369 | need to check the stride and the lower bound of the reference. */ | |
1370 | if (chrec_contains_symbols_defined_in_loop (TREE_OPERAND (obj, 2), | |
1371 | loop->num) | |
1372 | || chrec_contains_symbols_defined_in_loop (TREE_OPERAND (obj, 3), | |
1373 | loop->num)) | |
1374 | return false; | |
1375 | } | |
1376 | else if (TREE_CODE (obj) == COMPONENT_REF) | |
1377 | { | |
1378 | if (chrec_contains_symbols_defined_in_loop (TREE_OPERAND (obj, 2), | |
1379 | loop->num)) | |
1380 | return false; | |
1381 | } | |
1382 | obj = TREE_OPERAND (obj, 0); | |
1383 | } | |
1384 | ||
182cf5a9 | 1385 | if (!INDIRECT_REF_P (obj) |
1386 | && TREE_CODE (obj) != MEM_REF) | |
80bb306a | 1387 | return true; |
1388 | ||
1389 | return !chrec_contains_symbols_defined_in_loop (TREE_OPERAND (obj, 0), | |
1390 | loop->num); | |
1391 | } | |
1392 | ||
80bb306a | 1393 | /* Returns false if we can prove that data references A and B do not alias, |
5fc88ffd | 1394 | true otherwise. If LOOP_NEST is false no cross-iteration aliases are |
1395 | considered. */ | |
80bb306a | 1396 | |
255b6be7 | 1397 | bool |
5fc88ffd | 1398 | dr_may_alias_p (const struct data_reference *a, const struct data_reference *b, |
1399 | bool loop_nest) | |
80bb306a | 1400 | { |
2760c0e3 | 1401 | tree addr_a = DR_BASE_OBJECT (a); |
1402 | tree addr_b = DR_BASE_OBJECT (b); | |
80bb306a | 1403 | |
5fc88ffd | 1404 | /* If we are not processing a loop nest but scalar code we |
1405 | do not need to care about possible cross-iteration dependences | |
1406 | and thus can process the full original reference. Do so, | |
1407 | similar to how loop invariant motion applies extra offset-based | |
1408 | disambiguation. */ | |
1409 | if (!loop_nest) | |
1410 | { | |
1411 | aff_tree off1, off2; | |
5de9d3ed | 1412 | widest_int size1, size2; |
5fc88ffd | 1413 | get_inner_reference_aff (DR_REF (a), &off1, &size1); |
1414 | get_inner_reference_aff (DR_REF (b), &off2, &size2); | |
e913b5cd | 1415 | aff_combination_scale (&off1, -1); |
5fc88ffd | 1416 | aff_combination_add (&off2, &off1); |
1417 | if (aff_comb_cannot_overlap_p (&off2, size1, size2)) | |
1418 | return false; | |
1419 | } | |
1420 | ||
62b0a610 | 1421 | if ((TREE_CODE (addr_a) == MEM_REF || TREE_CODE (addr_a) == TARGET_MEM_REF) |
1422 | && (TREE_CODE (addr_b) == MEM_REF || TREE_CODE (addr_b) == TARGET_MEM_REF) | |
1423 | && MR_DEPENDENCE_CLIQUE (addr_a) == MR_DEPENDENCE_CLIQUE (addr_b) | |
1424 | && MR_DEPENDENCE_BASE (addr_a) != MR_DEPENDENCE_BASE (addr_b)) | |
1425 | return false; | |
1426 | ||
c8efcccb | 1427 | /* If we had an evolution in a pointer-based MEM_REF BASE_OBJECT we |
1428 | do not know the size of the base-object. So we cannot do any | |
1429 | offset/overlap based analysis but have to rely on points-to | |
1430 | information only. */ | |
95539e1d | 1431 | if (TREE_CODE (addr_a) == MEM_REF |
f146c442 | 1432 | && (DR_UNCONSTRAINED_BASE (a) |
1433 | || TREE_CODE (TREE_OPERAND (addr_a, 0)) == SSA_NAME)) | |
95539e1d | 1434 | { |
c8efcccb | 1435 | /* For true dependences we can apply TBAA. */ |
1436 | if (flag_strict_aliasing | |
1437 | && DR_IS_WRITE (a) && DR_IS_READ (b) | |
1438 | && !alias_sets_conflict_p (get_alias_set (DR_REF (a)), | |
1439 | get_alias_set (DR_REF (b)))) | |
1440 | return false; | |
1441 | if (TREE_CODE (addr_b) == MEM_REF) | |
95539e1d | 1442 | return ptr_derefs_may_alias_p (TREE_OPERAND (addr_a, 0), |
1443 | TREE_OPERAND (addr_b, 0)); | |
1444 | else | |
1445 | return ptr_derefs_may_alias_p (TREE_OPERAND (addr_a, 0), | |
1446 | build_fold_addr_expr (addr_b)); | |
1447 | } | |
1448 | else if (TREE_CODE (addr_b) == MEM_REF | |
f146c442 | 1449 | && (DR_UNCONSTRAINED_BASE (b) |
1450 | || TREE_CODE (TREE_OPERAND (addr_b, 0)) == SSA_NAME)) | |
c8efcccb | 1451 | { |
1452 | /* For true dependences we can apply TBAA. */ | |
1453 | if (flag_strict_aliasing | |
1454 | && DR_IS_WRITE (a) && DR_IS_READ (b) | |
1455 | && !alias_sets_conflict_p (get_alias_set (DR_REF (a)), | |
1456 | get_alias_set (DR_REF (b)))) | |
1457 | return false; | |
1458 | if (TREE_CODE (addr_a) == MEM_REF) | |
1459 | return ptr_derefs_may_alias_p (TREE_OPERAND (addr_a, 0), | |
1460 | TREE_OPERAND (addr_b, 0)); | |
1461 | else | |
1462 | return ptr_derefs_may_alias_p (build_fold_addr_expr (addr_a), | |
1463 | TREE_OPERAND (addr_b, 0)); | |
1464 | } | |
95539e1d | 1465 | |
1466 | /* Otherwise DR_BASE_OBJECT is an access that covers the whole object | |
1467 | that is being subsetted in the loop nest. */ | |
9ff25603 | 1468 | if (DR_IS_WRITE (a) && DR_IS_WRITE (b)) |
2760c0e3 | 1469 | return refs_output_dependent_p (addr_a, addr_b); |
9ff25603 | 1470 | else if (DR_IS_READ (a) && DR_IS_WRITE (b)) |
2760c0e3 | 1471 | return refs_anti_dependent_p (addr_a, addr_b); |
1472 | return refs_may_alias_p (addr_a, addr_b); | |
80bb306a | 1473 | } |
1474 | ||
6b421feb | 1475 | /* Initialize a data dependence relation between data accesses A and |
1476 | B. NB_LOOPS is the number of loops surrounding the references: the | |
1477 | size of the classic distance/direction vectors. */ | |
2146e26d | 1478 | |
16dfb112 | 1479 | struct data_dependence_relation * |
48e1416a | 1480 | initialize_data_dependence_relation (struct data_reference *a, |
6b421feb | 1481 | struct data_reference *b, |
f1f41a6c | 1482 | vec<loop_p> loop_nest) |
2146e26d | 1483 | { |
1484 | struct data_dependence_relation *res; | |
6b421feb | 1485 | unsigned int i; |
48e1416a | 1486 | |
4c36ffe6 | 1487 | res = XNEW (struct data_dependence_relation); |
2146e26d | 1488 | DDR_A (res) = a; |
1489 | DDR_B (res) = b; | |
f1f41a6c | 1490 | DDR_LOOP_NEST (res).create (0); |
0ecb94cf | 1491 | DDR_REVERSED_P (res) = false; |
f1f41a6c | 1492 | DDR_SUBSCRIPTS (res).create (0); |
1493 | DDR_DIR_VECTS (res).create (0); | |
1494 | DDR_DIST_VECTS (res).create (0); | |
2146e26d | 1495 | |
516849c7 | 1496 | if (a == NULL || b == NULL) |
1497 | { | |
48e1416a | 1498 | DDR_ARE_DEPENDENT (res) = chrec_dont_know; |
516849c7 | 1499 | return res; |
48e1416a | 1500 | } |
516849c7 | 1501 | |
80bb306a | 1502 | /* If the data references do not alias, then they are independent. */ |
f1f41a6c | 1503 | if (!dr_may_alias_p (a, b, loop_nest.exists ())) |
516849c7 | 1504 | { |
48e1416a | 1505 | DDR_ARE_DEPENDENT (res) = chrec_known; |
516849c7 | 1506 | return res; |
1507 | } | |
2146e26d | 1508 | |
cc76857b | 1509 | /* The case where the references are exactly the same. */ |
c127dd86 | 1510 | if (operand_equal_p (DR_REF (a), DR_REF (b), 0)) |
1511 | { | |
f1f41a6c | 1512 | if (loop_nest.exists () |
1513 | && !object_address_invariant_in_loop_p (loop_nest[0], | |
c76b51be | 1514 | DR_BASE_OBJECT (a))) |
1515 | { | |
1516 | DDR_ARE_DEPENDENT (res) = chrec_dont_know; | |
1517 | return res; | |
1518 | } | |
c127dd86 | 1519 | DDR_AFFINE_P (res) = true; |
1520 | DDR_ARE_DEPENDENT (res) = NULL_TREE; | |
f1f41a6c | 1521 | DDR_SUBSCRIPTS (res).create (DR_NUM_DIMENSIONS (a)); |
c127dd86 | 1522 | DDR_LOOP_NEST (res) = loop_nest; |
1523 | DDR_INNER_LOOP (res) = 0; | |
1524 | DDR_SELF_REFERENCE (res) = true; | |
c76b51be | 1525 | for (i = 0; i < DR_NUM_DIMENSIONS (a); i++) |
1526 | { | |
1527 | struct subscript *subscript; | |
1528 | ||
1529 | subscript = XNEW (struct subscript); | |
1530 | SUB_CONFLICTS_IN_A (subscript) = conflict_fn_not_known (); | |
1531 | SUB_CONFLICTS_IN_B (subscript) = conflict_fn_not_known (); | |
1532 | SUB_LAST_CONFLICT (subscript) = chrec_dont_know; | |
1533 | SUB_DISTANCE (subscript) = chrec_dont_know; | |
f1f41a6c | 1534 | DDR_SUBSCRIPTS (res).safe_push (subscript); |
c76b51be | 1535 | } |
c127dd86 | 1536 | return res; |
1537 | } | |
1538 | ||
80bb306a | 1539 | /* If the references do not access the same object, we do not know |
1540 | whether they alias or not. */ | |
1541 | if (!operand_equal_p (DR_BASE_OBJECT (a), DR_BASE_OBJECT (b), 0)) | |
2146e26d | 1542 | { |
48e1416a | 1543 | DDR_ARE_DEPENDENT (res) = chrec_dont_know; |
516849c7 | 1544 | return res; |
1545 | } | |
6b421feb | 1546 | |
80bb306a | 1547 | /* If the base of the object is not invariant in the loop nest, we cannot |
bef304b8 | 1548 | analyze it. TODO -- in fact, it would suffice to record that there may |
310d2511 | 1549 | be arbitrary dependences in the loops where the base object varies. */ |
f1f41a6c | 1550 | if (loop_nest.exists () |
1551 | && !object_address_invariant_in_loop_p (loop_nest[0], | |
37545e54 | 1552 | DR_BASE_OBJECT (a))) |
516849c7 | 1553 | { |
48e1416a | 1554 | DDR_ARE_DEPENDENT (res) = chrec_dont_know; |
516849c7 | 1555 | return res; |
1556 | } | |
80bb306a | 1557 | |
fabd4538 | 1558 | /* If the number of dimensions of the access to not agree we can have |
1559 | a pointer access to a component of the array element type and an | |
1560 | array access while the base-objects are still the same. Punt. */ | |
1561 | if (DR_NUM_DIMENSIONS (a) != DR_NUM_DIMENSIONS (b)) | |
1562 | { | |
1563 | DDR_ARE_DEPENDENT (res) = chrec_dont_know; | |
1564 | return res; | |
1565 | } | |
80bb306a | 1566 | |
516849c7 | 1567 | DDR_AFFINE_P (res) = true; |
1568 | DDR_ARE_DEPENDENT (res) = NULL_TREE; | |
f1f41a6c | 1569 | DDR_SUBSCRIPTS (res).create (DR_NUM_DIMENSIONS (a)); |
b44d1046 | 1570 | DDR_LOOP_NEST (res) = loop_nest; |
355572cc | 1571 | DDR_INNER_LOOP (res) = 0; |
c127dd86 | 1572 | DDR_SELF_REFERENCE (res) = false; |
1532ec98 | 1573 | |
516849c7 | 1574 | for (i = 0; i < DR_NUM_DIMENSIONS (a); i++) |
1575 | { | |
1576 | struct subscript *subscript; | |
48e1416a | 1577 | |
4c36ffe6 | 1578 | subscript = XNEW (struct subscript); |
87da4f2e | 1579 | SUB_CONFLICTS_IN_A (subscript) = conflict_fn_not_known (); |
1580 | SUB_CONFLICTS_IN_B (subscript) = conflict_fn_not_known (); | |
516849c7 | 1581 | SUB_LAST_CONFLICT (subscript) = chrec_dont_know; |
1582 | SUB_DISTANCE (subscript) = chrec_dont_know; | |
f1f41a6c | 1583 | DDR_SUBSCRIPTS (res).safe_push (subscript); |
2146e26d | 1584 | } |
41c7a324 | 1585 | |
2146e26d | 1586 | return res; |
1587 | } | |
1588 | ||
87da4f2e | 1589 | /* Frees memory used by the conflict function F. */ |
1590 | ||
1591 | static void | |
1592 | free_conflict_function (conflict_function *f) | |
1593 | { | |
1594 | unsigned i; | |
1595 | ||
1596 | if (CF_NONTRIVIAL_P (f)) | |
1597 | { | |
1598 | for (i = 0; i < f->n; i++) | |
1599 | affine_fn_free (f->fns[i]); | |
1600 | } | |
1601 | free (f); | |
1602 | } | |
1603 | ||
1604 | /* Frees memory used by SUBSCRIPTS. */ | |
1605 | ||
1606 | static void | |
f1f41a6c | 1607 | free_subscripts (vec<subscript_p> subscripts) |
87da4f2e | 1608 | { |
1609 | unsigned i; | |
1610 | subscript_p s; | |
1611 | ||
f1f41a6c | 1612 | FOR_EACH_VEC_ELT (subscripts, i, s) |
87da4f2e | 1613 | { |
1614 | free_conflict_function (s->conflicting_iterations_in_a); | |
1615 | free_conflict_function (s->conflicting_iterations_in_b); | |
19af51e2 | 1616 | free (s); |
87da4f2e | 1617 | } |
f1f41a6c | 1618 | subscripts.release (); |
87da4f2e | 1619 | } |
1620 | ||
2146e26d | 1621 | /* Set DDR_ARE_DEPENDENT to CHREC and finalize the subscript overlap |
1622 | description. */ | |
1623 | ||
1624 | static inline void | |
48e1416a | 1625 | finalize_ddr_dependent (struct data_dependence_relation *ddr, |
2146e26d | 1626 | tree chrec) |
1627 | { | |
48e1416a | 1628 | DDR_ARE_DEPENDENT (ddr) = chrec; |
87da4f2e | 1629 | free_subscripts (DDR_SUBSCRIPTS (ddr)); |
f1f41a6c | 1630 | DDR_SUBSCRIPTS (ddr).create (0); |
2146e26d | 1631 | } |
1632 | ||
bc3c8ad4 | 1633 | /* The dependence relation DDR cannot be represented by a distance |
1634 | vector. */ | |
1635 | ||
1636 | static inline void | |
1637 | non_affine_dependence_relation (struct data_dependence_relation *ddr) | |
1638 | { | |
1639 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
1640 | fprintf (dump_file, "(Dependence relation cannot be represented by distance vector.) \n"); | |
1641 | ||
1642 | DDR_AFFINE_P (ddr) = false; | |
1643 | } | |
1644 | ||
2146e26d | 1645 | \f |
1646 | ||
1647 | /* This section contains the classic Banerjee tests. */ | |
1648 | ||
1649 | /* Returns true iff CHREC_A and CHREC_B are not dependent on any index | |
1650 | variables, i.e., if the ZIV (Zero Index Variable) test is true. */ | |
1651 | ||
1652 | static inline bool | |
7ecb5bb2 | 1653 | ziv_subscript_p (const_tree chrec_a, const_tree chrec_b) |
2146e26d | 1654 | { |
1655 | return (evolution_function_is_constant_p (chrec_a) | |
1656 | && evolution_function_is_constant_p (chrec_b)); | |
1657 | } | |
1658 | ||
1659 | /* Returns true iff CHREC_A and CHREC_B are dependent on an index | |
1660 | variable, i.e., if the SIV (Single Index Variable) test is true. */ | |
1661 | ||
1662 | static bool | |
7ecb5bb2 | 1663 | siv_subscript_p (const_tree chrec_a, const_tree chrec_b) |
2146e26d | 1664 | { |
1665 | if ((evolution_function_is_constant_p (chrec_a) | |
1666 | && evolution_function_is_univariate_p (chrec_b)) | |
1667 | || (evolution_function_is_constant_p (chrec_b) | |
1668 | && evolution_function_is_univariate_p (chrec_a))) | |
1669 | return true; | |
48e1416a | 1670 | |
2146e26d | 1671 | if (evolution_function_is_univariate_p (chrec_a) |
1672 | && evolution_function_is_univariate_p (chrec_b)) | |
1673 | { | |
1674 | switch (TREE_CODE (chrec_a)) | |
1675 | { | |
1676 | case POLYNOMIAL_CHREC: | |
1677 | switch (TREE_CODE (chrec_b)) | |
1678 | { | |
1679 | case POLYNOMIAL_CHREC: | |
1680 | if (CHREC_VARIABLE (chrec_a) != CHREC_VARIABLE (chrec_b)) | |
1681 | return false; | |
48e1416a | 1682 | |
2146e26d | 1683 | default: |
1684 | return true; | |
1685 | } | |
48e1416a | 1686 | |
2146e26d | 1687 | default: |
1688 | return true; | |
1689 | } | |
1690 | } | |
48e1416a | 1691 | |
2146e26d | 1692 | return false; |
1693 | } | |
1694 | ||
87da4f2e | 1695 | /* Creates a conflict function with N dimensions. The affine functions |
1696 | in each dimension follow. */ | |
1697 | ||
1698 | static conflict_function * | |
1699 | conflict_fn (unsigned n, ...) | |
1700 | { | |
1701 | unsigned i; | |
1702 | conflict_function *ret = XCNEW (conflict_function); | |
1703 | va_list ap; | |
1704 | ||
a86c75a6 | 1705 | gcc_assert (0 < n && n <= MAX_DIM); |
9af5ce0c | 1706 | va_start (ap, n); |
48e1416a | 1707 | |
87da4f2e | 1708 | ret->n = n; |
1709 | for (i = 0; i < n; i++) | |
1710 | ret->fns[i] = va_arg (ap, affine_fn); | |
9af5ce0c | 1711 | va_end (ap); |
87da4f2e | 1712 | |
1713 | return ret; | |
1714 | } | |
1715 | ||
1716 | /* Returns constant affine function with value CST. */ | |
1717 | ||
1718 | static affine_fn | |
1719 | affine_fn_cst (tree cst) | |
1720 | { | |
f1f41a6c | 1721 | affine_fn fn; |
1722 | fn.create (1); | |
1723 | fn.quick_push (cst); | |
87da4f2e | 1724 | return fn; |
1725 | } | |
1726 | ||
1727 | /* Returns affine function with single variable, CST + COEF * x_DIM. */ | |
1728 | ||
1729 | static affine_fn | |
1730 | affine_fn_univar (tree cst, unsigned dim, tree coef) | |
1731 | { | |
f1f41a6c | 1732 | affine_fn fn; |
1733 | fn.create (dim + 1); | |
87da4f2e | 1734 | unsigned i; |
1735 | ||
1736 | gcc_assert (dim > 0); | |
f1f41a6c | 1737 | fn.quick_push (cst); |
87da4f2e | 1738 | for (i = 1; i < dim; i++) |
f1f41a6c | 1739 | fn.quick_push (integer_zero_node); |
1740 | fn.quick_push (coef); | |
87da4f2e | 1741 | return fn; |
1742 | } | |
1743 | ||
2146e26d | 1744 | /* Analyze a ZIV (Zero Index Variable) subscript. *OVERLAPS_A and |
1745 | *OVERLAPS_B are initialized to the functions that describe the | |
1746 | relation between the elements accessed twice by CHREC_A and | |
1747 | CHREC_B. For k >= 0, the following property is verified: | |
1748 | ||
1749 | CHREC_A (*OVERLAPS_A (k)) = CHREC_B (*OVERLAPS_B (k)). */ | |
1750 | ||
48e1416a | 1751 | static void |
1752 | analyze_ziv_subscript (tree chrec_a, | |
1753 | tree chrec_b, | |
87da4f2e | 1754 | conflict_function **overlaps_a, |
48e1416a | 1755 | conflict_function **overlaps_b, |
bc3c8ad4 | 1756 | tree *last_conflicts) |
2146e26d | 1757 | { |
9a3dad1c | 1758 | tree type, difference; |
6b421feb | 1759 | dependence_stats.num_ziv++; |
48e1416a | 1760 | |
2146e26d | 1761 | if (dump_file && (dump_flags & TDF_DETAILS)) |
1762 | fprintf (dump_file, "(analyze_ziv_subscript \n"); | |
9a3dad1c | 1763 | |
1764 | type = signed_type_for_types (TREE_TYPE (chrec_a), TREE_TYPE (chrec_b)); | |
75a70cf9 | 1765 | chrec_a = chrec_convert (type, chrec_a, NULL); |
1766 | chrec_b = chrec_convert (type, chrec_b, NULL); | |
9a3dad1c | 1767 | difference = chrec_fold_minus (type, chrec_a, chrec_b); |
48e1416a | 1768 | |
2146e26d | 1769 | switch (TREE_CODE (difference)) |
1770 | { | |
1771 | case INTEGER_CST: | |
1772 | if (integer_zerop (difference)) | |
1773 | { | |
1774 | /* The difference is equal to zero: the accessed index | |
1775 | overlaps for each iteration in the loop. */ | |
87da4f2e | 1776 | *overlaps_a = conflict_fn (1, affine_fn_cst (integer_zero_node)); |
1777 | *overlaps_b = conflict_fn (1, affine_fn_cst (integer_zero_node)); | |
bc3c8ad4 | 1778 | *last_conflicts = chrec_dont_know; |
6b421feb | 1779 | dependence_stats.num_ziv_dependent++; |
2146e26d | 1780 | } |
1781 | else | |
1782 | { | |
1783 | /* The accesses do not overlap. */ | |
87da4f2e | 1784 | *overlaps_a = conflict_fn_no_dependence (); |
1785 | *overlaps_b = conflict_fn_no_dependence (); | |
bc3c8ad4 | 1786 | *last_conflicts = integer_zero_node; |
6b421feb | 1787 | dependence_stats.num_ziv_independent++; |
2146e26d | 1788 | } |
1789 | break; | |
48e1416a | 1790 | |
2146e26d | 1791 | default: |
48e1416a | 1792 | /* We're not sure whether the indexes overlap. For the moment, |
2146e26d | 1793 | conservatively answer "don't know". */ |
6b421feb | 1794 | if (dump_file && (dump_flags & TDF_DETAILS)) |
1795 | fprintf (dump_file, "ziv test failed: difference is non-integer.\n"); | |
1796 | ||
87da4f2e | 1797 | *overlaps_a = conflict_fn_not_known (); |
1798 | *overlaps_b = conflict_fn_not_known (); | |
bc3c8ad4 | 1799 | *last_conflicts = chrec_dont_know; |
6b421feb | 1800 | dependence_stats.num_ziv_unimplemented++; |
2146e26d | 1801 | break; |
1802 | } | |
48e1416a | 1803 | |
2146e26d | 1804 | if (dump_file && (dump_flags & TDF_DETAILS)) |
1805 | fprintf (dump_file, ")\n"); | |
1806 | } | |
1807 | ||
8fe79ba5 | 1808 | /* Similar to max_stmt_executions_int, but returns the bound as a tree, |
d500fef3 | 1809 | and only if it fits to the int type. If this is not the case, or the |
8fe79ba5 | 1810 | bound on the number of iterations of LOOP could not be derived, returns |
d500fef3 | 1811 | chrec_dont_know. */ |
1812 | ||
1813 | static tree | |
8fe79ba5 | 1814 | max_stmt_executions_tree (struct loop *loop) |
d500fef3 | 1815 | { |
5de9d3ed | 1816 | widest_int nit; |
d500fef3 | 1817 | |
fee017b3 | 1818 | if (!max_stmt_executions (loop, &nit)) |
d500fef3 | 1819 | return chrec_dont_know; |
1820 | ||
796b6678 | 1821 | if (!wi::fits_to_tree_p (nit, unsigned_type_node)) |
d500fef3 | 1822 | return chrec_dont_know; |
1823 | ||
e913b5cd | 1824 | return wide_int_to_tree (unsigned_type_node, nit); |
d500fef3 | 1825 | } |
1826 | ||
06a9142d | 1827 | /* Determine whether the CHREC is always positive/negative. If the expression |
1828 | cannot be statically analyzed, return false, otherwise set the answer into | |
1829 | VALUE. */ | |
1830 | ||
1831 | static bool | |
1832 | chrec_is_positive (tree chrec, bool *value) | |
1833 | { | |
1834 | bool value0, value1, value2; | |
1835 | tree end_value, nb_iter; | |
1836 | ||
1837 | switch (TREE_CODE (chrec)) | |
1838 | { | |
1839 | case POLYNOMIAL_CHREC: | |
1840 | if (!chrec_is_positive (CHREC_LEFT (chrec), &value0) | |
1841 | || !chrec_is_positive (CHREC_RIGHT (chrec), &value1)) | |
1842 | return false; | |
1843 | ||
1844 | /* FIXME -- overflows. */ | |
1845 | if (value0 == value1) | |
1846 | { | |
1847 | *value = value0; | |
1848 | return true; | |
1849 | } | |
1850 | ||
1851 | /* Otherwise the chrec is under the form: "{-197, +, 2}_1", | |
1852 | and the proof consists in showing that the sign never | |
1853 | changes during the execution of the loop, from 0 to | |
1854 | loop->nb_iterations. */ | |
1855 | if (!evolution_function_is_affine_p (chrec)) | |
1856 | return false; | |
1857 | ||
1858 | nb_iter = number_of_latch_executions (get_chrec_loop (chrec)); | |
1859 | if (chrec_contains_undetermined (nb_iter)) | |
1860 | return false; | |
1861 | ||
1862 | #if 0 | |
1863 | /* TODO -- If the test is after the exit, we may decrease the number of | |
1864 | iterations by one. */ | |
1865 | if (after_exit) | |
1866 | nb_iter = chrec_fold_minus (type, nb_iter, build_int_cst (type, 1)); | |
1867 | #endif | |
1868 | ||
1869 | end_value = chrec_apply (CHREC_VARIABLE (chrec), chrec, nb_iter); | |
1870 | ||
1871 | if (!chrec_is_positive (end_value, &value2)) | |
1872 | return false; | |
1873 | ||
1874 | *value = value0; | |
1875 | return value0 == value1; | |
1876 | ||
1877 | case INTEGER_CST: | |
1878 | switch (tree_int_cst_sgn (chrec)) | |
1879 | { | |
1880 | case -1: | |
1881 | *value = false; | |
1882 | break; | |
1883 | case 1: | |
1884 | *value = true; | |
1885 | break; | |
1886 | default: | |
1887 | return false; | |
1888 | } | |
1889 | return true; | |
1890 | ||
1891 | default: | |
1892 | return false; | |
1893 | } | |
1894 | } | |
1895 | ||
1896 | ||
2146e26d | 1897 | /* Analyze a SIV (Single Index Variable) subscript where CHREC_A is a |
1898 | constant, and CHREC_B is an affine function. *OVERLAPS_A and | |
1899 | *OVERLAPS_B are initialized to the functions that describe the | |
1900 | relation between the elements accessed twice by CHREC_A and | |
1901 | CHREC_B. For k >= 0, the following property is verified: | |
1902 | ||
1903 | CHREC_A (*OVERLAPS_A (k)) = CHREC_B (*OVERLAPS_B (k)). */ | |
1904 | ||
1905 | static void | |
48e1416a | 1906 | analyze_siv_subscript_cst_affine (tree chrec_a, |
2146e26d | 1907 | tree chrec_b, |
48e1416a | 1908 | conflict_function **overlaps_a, |
1909 | conflict_function **overlaps_b, | |
bc3c8ad4 | 1910 | tree *last_conflicts) |
2146e26d | 1911 | { |
1912 | bool value0, value1, value2; | |
9a3dad1c | 1913 | tree type, difference, tmp; |
f84a688a | 1914 | |
9a3dad1c | 1915 | type = signed_type_for_types (TREE_TYPE (chrec_a), TREE_TYPE (chrec_b)); |
75a70cf9 | 1916 | chrec_a = chrec_convert (type, chrec_a, NULL); |
1917 | chrec_b = chrec_convert (type, chrec_b, NULL); | |
9a3dad1c | 1918 | difference = chrec_fold_minus (type, initial_condition (chrec_b), chrec_a); |
48e1416a | 1919 | |
06a9142d | 1920 | /* Special case overlap in the first iteration. */ |
1921 | if (integer_zerop (difference)) | |
1922 | { | |
1923 | *overlaps_a = conflict_fn (1, affine_fn_cst (integer_zero_node)); | |
1924 | *overlaps_b = conflict_fn (1, affine_fn_cst (integer_zero_node)); | |
1925 | *last_conflicts = integer_one_node; | |
1926 | return; | |
1927 | } | |
1928 | ||
2146e26d | 1929 | if (!chrec_is_positive (initial_condition (difference), &value0)) |
1930 | { | |
6b421feb | 1931 | if (dump_file && (dump_flags & TDF_DETAILS)) |
48e1416a | 1932 | fprintf (dump_file, "siv test failed: chrec is not positive.\n"); |
6b421feb | 1933 | |
1934 | dependence_stats.num_siv_unimplemented++; | |
87da4f2e | 1935 | *overlaps_a = conflict_fn_not_known (); |
1936 | *overlaps_b = conflict_fn_not_known (); | |
bc3c8ad4 | 1937 | *last_conflicts = chrec_dont_know; |
2146e26d | 1938 | return; |
1939 | } | |
1940 | else | |
1941 | { | |
1942 | if (value0 == false) | |
1943 | { | |
1944 | if (!chrec_is_positive (CHREC_RIGHT (chrec_b), &value1)) | |
1945 | { | |
6b421feb | 1946 | if (dump_file && (dump_flags & TDF_DETAILS)) |
1947 | fprintf (dump_file, "siv test failed: chrec not positive.\n"); | |
1948 | ||
87da4f2e | 1949 | *overlaps_a = conflict_fn_not_known (); |
48e1416a | 1950 | *overlaps_b = conflict_fn_not_known (); |
bc3c8ad4 | 1951 | *last_conflicts = chrec_dont_know; |
6b421feb | 1952 | dependence_stats.num_siv_unimplemented++; |
2146e26d | 1953 | return; |
1954 | } | |
1955 | else | |
1956 | { | |
1957 | if (value1 == true) | |
1958 | { | |
48e1416a | 1959 | /* Example: |
2146e26d | 1960 | chrec_a = 12 |
1961 | chrec_b = {10, +, 1} | |
1962 | */ | |
48e1416a | 1963 | |
e7be49a3 | 1964 | if (tree_fold_divides_p (CHREC_RIGHT (chrec_b), difference)) |
2146e26d | 1965 | { |
d500fef3 | 1966 | HOST_WIDE_INT numiter; |
1967 | struct loop *loop = get_chrec_loop (chrec_b); | |
b6d41046 | 1968 | |
87da4f2e | 1969 | *overlaps_a = conflict_fn (1, affine_fn_cst (integer_zero_node)); |
9a3dad1c | 1970 | tmp = fold_build2 (EXACT_DIV_EXPR, type, |
1971 | fold_build1 (ABS_EXPR, type, difference), | |
87da4f2e | 1972 | CHREC_RIGHT (chrec_b)); |
1973 | *overlaps_b = conflict_fn (1, affine_fn_cst (tmp)); | |
bc3c8ad4 | 1974 | *last_conflicts = integer_one_node; |
48e1416a | 1975 | |
b6d41046 | 1976 | |
1977 | /* Perform weak-zero siv test to see if overlap is | |
1978 | outside the loop bounds. */ | |
fee017b3 | 1979 | numiter = max_stmt_executions_int (loop); |
b6d41046 | 1980 | |
d500fef3 | 1981 | if (numiter >= 0 |
1982 | && compare_tree_int (tmp, numiter) > 0) | |
b6d41046 | 1983 | { |
87da4f2e | 1984 | free_conflict_function (*overlaps_a); |
1985 | free_conflict_function (*overlaps_b); | |
1986 | *overlaps_a = conflict_fn_no_dependence (); | |
1987 | *overlaps_b = conflict_fn_no_dependence (); | |
b6d41046 | 1988 | *last_conflicts = integer_zero_node; |
6b421feb | 1989 | dependence_stats.num_siv_independent++; |
b6d41046 | 1990 | return; |
48e1416a | 1991 | } |
6b421feb | 1992 | dependence_stats.num_siv_dependent++; |
2146e26d | 1993 | return; |
1994 | } | |
48e1416a | 1995 | |
e7be49a3 | 1996 | /* When the step does not divide the difference, there are |
2146e26d | 1997 | no overlaps. */ |
1998 | else | |
1999 | { | |
87da4f2e | 2000 | *overlaps_a = conflict_fn_no_dependence (); |
48e1416a | 2001 | *overlaps_b = conflict_fn_no_dependence (); |
bc3c8ad4 | 2002 | *last_conflicts = integer_zero_node; |
6b421feb | 2003 | dependence_stats.num_siv_independent++; |
2146e26d | 2004 | return; |
2005 | } | |
2006 | } | |
48e1416a | 2007 | |
2146e26d | 2008 | else |
2009 | { | |
48e1416a | 2010 | /* Example: |
2146e26d | 2011 | chrec_a = 12 |
2012 | chrec_b = {10, +, -1} | |
48e1416a | 2013 | |
2146e26d | 2014 | In this case, chrec_a will not overlap with chrec_b. */ |
87da4f2e | 2015 | *overlaps_a = conflict_fn_no_dependence (); |
2016 | *overlaps_b = conflict_fn_no_dependence (); | |
bc3c8ad4 | 2017 | *last_conflicts = integer_zero_node; |
6b421feb | 2018 | dependence_stats.num_siv_independent++; |
2146e26d | 2019 | return; |
2020 | } | |
2021 | } | |
2022 | } | |
48e1416a | 2023 | else |
2146e26d | 2024 | { |
2025 | if (!chrec_is_positive (CHREC_RIGHT (chrec_b), &value2)) | |
2026 | { | |
6b421feb | 2027 | if (dump_file && (dump_flags & TDF_DETAILS)) |
2028 | fprintf (dump_file, "siv test failed: chrec not positive.\n"); | |
2029 | ||
87da4f2e | 2030 | *overlaps_a = conflict_fn_not_known (); |
48e1416a | 2031 | *overlaps_b = conflict_fn_not_known (); |
bc3c8ad4 | 2032 | *last_conflicts = chrec_dont_know; |
6b421feb | 2033 | dependence_stats.num_siv_unimplemented++; |
2146e26d | 2034 | return; |
2035 | } | |
2036 | else | |
2037 | { | |
2038 | if (value2 == false) | |
2039 | { | |
48e1416a | 2040 | /* Example: |
2146e26d | 2041 | chrec_a = 3 |
2042 | chrec_b = {10, +, -1} | |
2043 | */ | |
e7be49a3 | 2044 | if (tree_fold_divides_p (CHREC_RIGHT (chrec_b), difference)) |
2146e26d | 2045 | { |
d500fef3 | 2046 | HOST_WIDE_INT numiter; |
2047 | struct loop *loop = get_chrec_loop (chrec_b); | |
b6d41046 | 2048 | |
87da4f2e | 2049 | *overlaps_a = conflict_fn (1, affine_fn_cst (integer_zero_node)); |
9a3dad1c | 2050 | tmp = fold_build2 (EXACT_DIV_EXPR, type, difference, |
87da4f2e | 2051 | CHREC_RIGHT (chrec_b)); |
2052 | *overlaps_b = conflict_fn (1, affine_fn_cst (tmp)); | |
bc3c8ad4 | 2053 | *last_conflicts = integer_one_node; |
b6d41046 | 2054 | |
2055 | /* Perform weak-zero siv test to see if overlap is | |
2056 | outside the loop bounds. */ | |
fee017b3 | 2057 | numiter = max_stmt_executions_int (loop); |
b6d41046 | 2058 | |
d500fef3 | 2059 | if (numiter >= 0 |
2060 | && compare_tree_int (tmp, numiter) > 0) | |
b6d41046 | 2061 | { |
87da4f2e | 2062 | free_conflict_function (*overlaps_a); |
2063 | free_conflict_function (*overlaps_b); | |
2064 | *overlaps_a = conflict_fn_no_dependence (); | |
2065 | *overlaps_b = conflict_fn_no_dependence (); | |
b6d41046 | 2066 | *last_conflicts = integer_zero_node; |
6b421feb | 2067 | dependence_stats.num_siv_independent++; |
b6d41046 | 2068 | return; |
48e1416a | 2069 | } |
6b421feb | 2070 | dependence_stats.num_siv_dependent++; |
2146e26d | 2071 | return; |
2072 | } | |
48e1416a | 2073 | |
c4243fe7 | 2074 | /* When the step does not divide the difference, there |
2146e26d | 2075 | are no overlaps. */ |
2076 | else | |
2077 | { | |
87da4f2e | 2078 | *overlaps_a = conflict_fn_no_dependence (); |
48e1416a | 2079 | *overlaps_b = conflict_fn_no_dependence (); |
bc3c8ad4 | 2080 | *last_conflicts = integer_zero_node; |
6b421feb | 2081 | dependence_stats.num_siv_independent++; |
2146e26d | 2082 | return; |
2083 | } | |
2084 | } | |
2085 | else | |
2086 | { | |
48e1416a | 2087 | /* Example: |
2088 | chrec_a = 3 | |
2146e26d | 2089 | chrec_b = {4, +, 1} |
48e1416a | 2090 | |
2146e26d | 2091 | In this case, chrec_a will not overlap with chrec_b. */ |
87da4f2e | 2092 | *overlaps_a = conflict_fn_no_dependence (); |
2093 | *overlaps_b = conflict_fn_no_dependence (); | |
bc3c8ad4 | 2094 | *last_conflicts = integer_zero_node; |
6b421feb | 2095 | dependence_stats.num_siv_independent++; |
2146e26d | 2096 | return; |
2097 | } | |
2098 | } | |
2099 | } | |
2100 | } | |
2101 | } | |
2102 | ||
fbd828d4 | 2103 | /* Helper recursive function for initializing the matrix A. Returns |
bc3c8ad4 | 2104 | the initial value of CHREC. */ |
2146e26d | 2105 | |
8ceddf57 | 2106 | static tree |
bc3c8ad4 | 2107 | initialize_matrix_A (lambda_matrix A, tree chrec, unsigned index, int mult) |
2108 | { | |
2109 | gcc_assert (chrec); | |
2110 | ||
8ceddf57 | 2111 | switch (TREE_CODE (chrec)) |
2112 | { | |
2113 | case POLYNOMIAL_CHREC: | |
2114 | gcc_assert (TREE_CODE (CHREC_RIGHT (chrec)) == INTEGER_CST); | |
2115 | ||
2116 | A[index][0] = mult * int_cst_value (CHREC_RIGHT (chrec)); | |
2117 | return initialize_matrix_A (A, CHREC_LEFT (chrec), index + 1, mult); | |
2118 | ||
2119 | case PLUS_EXPR: | |
2120 | case MULT_EXPR: | |
2121 | case MINUS_EXPR: | |
2122 | { | |
2123 | tree op0 = initialize_matrix_A (A, TREE_OPERAND (chrec, 0), index, mult); | |
2124 | tree op1 = initialize_matrix_A (A, TREE_OPERAND (chrec, 1), index, mult); | |
2125 | ||
2126 | return chrec_fold_op (TREE_CODE (chrec), chrec_type (chrec), op0, op1); | |
2127 | } | |
2128 | ||
d09ef31a | 2129 | CASE_CONVERT: |
8ceddf57 | 2130 | { |
2131 | tree op = initialize_matrix_A (A, TREE_OPERAND (chrec, 0), index, mult); | |
75a70cf9 | 2132 | return chrec_convert (chrec_type (chrec), op, NULL); |
8ceddf57 | 2133 | } |
2134 | ||
b6eab06c | 2135 | case BIT_NOT_EXPR: |
2136 | { | |
2137 | /* Handle ~X as -1 - X. */ | |
2138 | tree op = initialize_matrix_A (A, TREE_OPERAND (chrec, 0), index, mult); | |
2139 | return chrec_fold_op (MINUS_EXPR, chrec_type (chrec), | |
2140 | build_int_cst (TREE_TYPE (chrec), -1), op); | |
2141 | } | |
2142 | ||
8ceddf57 | 2143 | case INTEGER_CST: |
2144 | return chrec; | |
c782188f | 2145 | |
8ceddf57 | 2146 | default: |
2147 | gcc_unreachable (); | |
2148 | return NULL_TREE; | |
2149 | } | |
bc3c8ad4 | 2150 | } |
2151 | ||
2152 | #define FLOOR_DIV(x,y) ((x) / (y)) | |
2153 | ||
48e1416a | 2154 | /* Solves the special case of the Diophantine equation: |
bc3c8ad4 | 2155 | | {0, +, STEP_A}_x (OVERLAPS_A) = {0, +, STEP_B}_y (OVERLAPS_B) |
2156 | ||
2157 | Computes the descriptions OVERLAPS_A and OVERLAPS_B. NITER is the | |
2158 | number of iterations that loops X and Y run. The overlaps will be | |
2159 | constructed as evolutions in dimension DIM. */ | |
2146e26d | 2160 | |
2161 | static void | |
48e1416a | 2162 | compute_overlap_steps_for_affine_univar (int niter, int step_a, int step_b, |
87da4f2e | 2163 | affine_fn *overlaps_a, |
48e1416a | 2164 | affine_fn *overlaps_b, |
bc3c8ad4 | 2165 | tree *last_conflicts, int dim) |
2166 | { | |
2167 | if (((step_a > 0 && step_b > 0) | |
2168 | || (step_a < 0 && step_b < 0))) | |
2169 | { | |
2170 | int step_overlaps_a, step_overlaps_b; | |
2171 | int gcd_steps_a_b, last_conflict, tau2; | |
2172 | ||
2173 | gcd_steps_a_b = gcd (step_a, step_b); | |
2174 | step_overlaps_a = step_b / gcd_steps_a_b; | |
2175 | step_overlaps_b = step_a / gcd_steps_a_b; | |
2176 | ||
e781c550 | 2177 | if (niter > 0) |
2178 | { | |
2179 | tau2 = FLOOR_DIV (niter, step_overlaps_a); | |
2180 | tau2 = MIN (tau2, FLOOR_DIV (niter, step_overlaps_b)); | |
2181 | last_conflict = tau2; | |
2182 | *last_conflicts = build_int_cst (NULL_TREE, last_conflict); | |
2183 | } | |
2184 | else | |
2185 | *last_conflicts = chrec_dont_know; | |
bc3c8ad4 | 2186 | |
48e1416a | 2187 | *overlaps_a = affine_fn_univar (integer_zero_node, dim, |
87da4f2e | 2188 | build_int_cst (NULL_TREE, |
2189 | step_overlaps_a)); | |
48e1416a | 2190 | *overlaps_b = affine_fn_univar (integer_zero_node, dim, |
2191 | build_int_cst (NULL_TREE, | |
87da4f2e | 2192 | step_overlaps_b)); |
bc3c8ad4 | 2193 | } |
2194 | ||
2195 | else | |
2196 | { | |
87da4f2e | 2197 | *overlaps_a = affine_fn_cst (integer_zero_node); |
2198 | *overlaps_b = affine_fn_cst (integer_zero_node); | |
bc3c8ad4 | 2199 | *last_conflicts = integer_zero_node; |
2200 | } | |
2201 | } | |
2202 | ||
bc3c8ad4 | 2203 | /* Solves the special case of a Diophantine equation where CHREC_A is |
2204 | an affine bivariate function, and CHREC_B is an affine univariate | |
48e1416a | 2205 | function. For example, |
bc3c8ad4 | 2206 | |
2207 | | {{0, +, 1}_x, +, 1335}_y = {0, +, 1336}_z | |
48e1416a | 2208 | |
2209 | has the following overlapping functions: | |
bc3c8ad4 | 2210 | |
2211 | | x (t, u, v) = {{0, +, 1336}_t, +, 1}_v | |
2212 | | y (t, u, v) = {{0, +, 1336}_u, +, 1}_v | |
2213 | | z (t, u, v) = {{{0, +, 1}_t, +, 1335}_u, +, 1}_v | |
2214 | ||
0975351b | 2215 | FORNOW: This is a specialized implementation for a case occurring in |
bc3c8ad4 | 2216 | a common benchmark. Implement the general algorithm. */ |
2217 | ||
2218 | static void | |
48e1416a | 2219 | compute_overlap_steps_for_affine_1_2 (tree chrec_a, tree chrec_b, |
87da4f2e | 2220 | conflict_function **overlaps_a, |
48e1416a | 2221 | conflict_function **overlaps_b, |
bc3c8ad4 | 2222 | tree *last_conflicts) |
2146e26d | 2223 | { |
bc3c8ad4 | 2224 | bool xz_p, yz_p, xyz_p; |
2225 | int step_x, step_y, step_z; | |
d500fef3 | 2226 | HOST_WIDE_INT niter_x, niter_y, niter_z, niter; |
87da4f2e | 2227 | affine_fn overlaps_a_xz, overlaps_b_xz; |
2228 | affine_fn overlaps_a_yz, overlaps_b_yz; | |
2229 | affine_fn overlaps_a_xyz, overlaps_b_xyz; | |
2230 | affine_fn ova1, ova2, ovb; | |
2231 | tree last_conflicts_xz, last_conflicts_yz, last_conflicts_xyz; | |
bc3c8ad4 | 2232 | |
58f77a87 | 2233 | step_x = int_cst_value (CHREC_RIGHT (CHREC_LEFT (chrec_a))); |
2234 | step_y = int_cst_value (CHREC_RIGHT (chrec_a)); | |
2235 | step_z = int_cst_value (CHREC_RIGHT (chrec_b)); | |
bc3c8ad4 | 2236 | |
fee017b3 | 2237 | niter_x = max_stmt_executions_int (get_chrec_loop (CHREC_LEFT (chrec_a))); |
2238 | niter_y = max_stmt_executions_int (get_chrec_loop (chrec_a)); | |
2239 | niter_z = max_stmt_executions_int (get_chrec_loop (chrec_b)); | |
48e1416a | 2240 | |
d500fef3 | 2241 | if (niter_x < 0 || niter_y < 0 || niter_z < 0) |
bc3c8ad4 | 2242 | { |
6b421feb | 2243 | if (dump_file && (dump_flags & TDF_DETAILS)) |
2244 | fprintf (dump_file, "overlap steps test failed: no iteration counts.\n"); | |
48e1416a | 2245 | |
87da4f2e | 2246 | *overlaps_a = conflict_fn_not_known (); |
2247 | *overlaps_b = conflict_fn_not_known (); | |
bc3c8ad4 | 2248 | *last_conflicts = chrec_dont_know; |
2249 | return; | |
2250 | } | |
2251 | ||
bc3c8ad4 | 2252 | niter = MIN (niter_x, niter_z); |
2253 | compute_overlap_steps_for_affine_univar (niter, step_x, step_z, | |
2254 | &overlaps_a_xz, | |
2255 | &overlaps_b_xz, | |
2256 | &last_conflicts_xz, 1); | |
2257 | niter = MIN (niter_y, niter_z); | |
2258 | compute_overlap_steps_for_affine_univar (niter, step_y, step_z, | |
2259 | &overlaps_a_yz, | |
2260 | &overlaps_b_yz, | |
2261 | &last_conflicts_yz, 2); | |
2262 | niter = MIN (niter_x, niter_z); | |
2263 | niter = MIN (niter_y, niter); | |
2264 | compute_overlap_steps_for_affine_univar (niter, step_x + step_y, step_z, | |
2265 | &overlaps_a_xyz, | |
2266 | &overlaps_b_xyz, | |
2267 | &last_conflicts_xyz, 3); | |
2268 | ||
2269 | xz_p = !integer_zerop (last_conflicts_xz); | |
2270 | yz_p = !integer_zerop (last_conflicts_yz); | |
2271 | xyz_p = !integer_zerop (last_conflicts_xyz); | |
2272 | ||
2273 | if (xz_p || yz_p || xyz_p) | |
2274 | { | |
87da4f2e | 2275 | ova1 = affine_fn_cst (integer_zero_node); |
2276 | ova2 = affine_fn_cst (integer_zero_node); | |
2277 | ovb = affine_fn_cst (integer_zero_node); | |
bc3c8ad4 | 2278 | if (xz_p) |
2279 | { | |
87da4f2e | 2280 | affine_fn t0 = ova1; |
2281 | affine_fn t2 = ovb; | |
2282 | ||
2283 | ova1 = affine_fn_plus (ova1, overlaps_a_xz); | |
2284 | ovb = affine_fn_plus (ovb, overlaps_b_xz); | |
2285 | affine_fn_free (t0); | |
2286 | affine_fn_free (t2); | |
bc3c8ad4 | 2287 | *last_conflicts = last_conflicts_xz; |
2288 | } | |
2289 | if (yz_p) | |
2290 | { | |
87da4f2e | 2291 | affine_fn t0 = ova2; |
2292 | affine_fn t2 = ovb; | |
2293 | ||
2294 | ova2 = affine_fn_plus (ova2, overlaps_a_yz); | |
2295 | ovb = affine_fn_plus (ovb, overlaps_b_yz); | |
2296 | affine_fn_free (t0); | |
2297 | affine_fn_free (t2); | |
bc3c8ad4 | 2298 | *last_conflicts = last_conflicts_yz; |
2299 | } | |
2300 | if (xyz_p) | |
2301 | { | |
87da4f2e | 2302 | affine_fn t0 = ova1; |
2303 | affine_fn t2 = ova2; | |
2304 | affine_fn t4 = ovb; | |
2305 | ||
2306 | ova1 = affine_fn_plus (ova1, overlaps_a_xyz); | |
2307 | ova2 = affine_fn_plus (ova2, overlaps_a_xyz); | |
2308 | ovb = affine_fn_plus (ovb, overlaps_b_xyz); | |
2309 | affine_fn_free (t0); | |
2310 | affine_fn_free (t2); | |
2311 | affine_fn_free (t4); | |
bc3c8ad4 | 2312 | *last_conflicts = last_conflicts_xyz; |
2313 | } | |
87da4f2e | 2314 | *overlaps_a = conflict_fn (2, ova1, ova2); |
2315 | *overlaps_b = conflict_fn (1, ovb); | |
bc3c8ad4 | 2316 | } |
2317 | else | |
2318 | { | |
87da4f2e | 2319 | *overlaps_a = conflict_fn (1, affine_fn_cst (integer_zero_node)); |
2320 | *overlaps_b = conflict_fn (1, affine_fn_cst (integer_zero_node)); | |
bc3c8ad4 | 2321 | *last_conflicts = integer_zero_node; |
2322 | } | |
87da4f2e | 2323 | |
2324 | affine_fn_free (overlaps_a_xz); | |
2325 | affine_fn_free (overlaps_b_xz); | |
2326 | affine_fn_free (overlaps_a_yz); | |
2327 | affine_fn_free (overlaps_b_yz); | |
2328 | affine_fn_free (overlaps_a_xyz); | |
2329 | affine_fn_free (overlaps_b_xyz); | |
2146e26d | 2330 | } |
2331 | ||
e01f9f1f | 2332 | /* Copy the elements of vector VEC1 with length SIZE to VEC2. */ |
2333 | ||
2334 | static void | |
2335 | lambda_vector_copy (lambda_vector vec1, lambda_vector vec2, | |
2336 | int size) | |
2337 | { | |
2338 | memcpy (vec2, vec1, size * sizeof (*vec1)); | |
2339 | } | |
2340 | ||
2341 | /* Copy the elements of M x N matrix MAT1 to MAT2. */ | |
2342 | ||
2343 | static void | |
2344 | lambda_matrix_copy (lambda_matrix mat1, lambda_matrix mat2, | |
2345 | int m, int n) | |
2346 | { | |
2347 | int i; | |
2348 | ||
2349 | for (i = 0; i < m; i++) | |
2350 | lambda_vector_copy (mat1[i], mat2[i], n); | |
2351 | } | |
2352 | ||
2353 | /* Store the N x N identity matrix in MAT. */ | |
2354 | ||
2355 | static void | |
2356 | lambda_matrix_id (lambda_matrix mat, int size) | |
2357 | { | |
2358 | int i, j; | |
2359 | ||
2360 | for (i = 0; i < size; i++) | |
2361 | for (j = 0; j < size; j++) | |
2362 | mat[i][j] = (i == j) ? 1 : 0; | |
2363 | } | |
2364 | ||
2365 | /* Return the first nonzero element of vector VEC1 between START and N. | |
2366 | We must have START <= N. Returns N if VEC1 is the zero vector. */ | |
2367 | ||
2368 | static int | |
2369 | lambda_vector_first_nz (lambda_vector vec1, int n, int start) | |
2370 | { | |
2371 | int j = start; | |
2372 | while (j < n && vec1[j] == 0) | |
2373 | j++; | |
2374 | return j; | |
2375 | } | |
2376 | ||
2377 | /* Add a multiple of row R1 of matrix MAT with N columns to row R2: | |
2378 | R2 = R2 + CONST1 * R1. */ | |
2379 | ||
2380 | static void | |
2381 | lambda_matrix_row_add (lambda_matrix mat, int n, int r1, int r2, int const1) | |
2382 | { | |
2383 | int i; | |
2384 | ||
2385 | if (const1 == 0) | |
2386 | return; | |
2387 | ||
2388 | for (i = 0; i < n; i++) | |
2389 | mat[r2][i] += const1 * mat[r1][i]; | |
2390 | } | |
2391 | ||
e01f9f1f | 2392 | /* Multiply vector VEC1 of length SIZE by a constant CONST1, |
2393 | and store the result in VEC2. */ | |
2394 | ||
2395 | static void | |
2396 | lambda_vector_mult_const (lambda_vector vec1, lambda_vector vec2, | |
2397 | int size, int const1) | |
2398 | { | |
2399 | int i; | |
2400 | ||
2401 | if (const1 == 0) | |
2402 | lambda_vector_clear (vec2, size); | |
2403 | else | |
2404 | for (i = 0; i < size; i++) | |
2405 | vec2[i] = const1 * vec1[i]; | |
2406 | } | |
2407 | ||
2408 | /* Negate vector VEC1 with length SIZE and store it in VEC2. */ | |
2409 | ||
2410 | static void | |
2411 | lambda_vector_negate (lambda_vector vec1, lambda_vector vec2, | |
2412 | int size) | |
2413 | { | |
2414 | lambda_vector_mult_const (vec1, vec2, size, -1); | |
2415 | } | |
2416 | ||
2417 | /* Negate row R1 of matrix MAT which has N columns. */ | |
2418 | ||
2419 | static void | |
2420 | lambda_matrix_row_negate (lambda_matrix mat, int n, int r1) | |
2421 | { | |
2422 | lambda_vector_negate (mat[r1], mat[r1], n); | |
2423 | } | |
2424 | ||
2425 | /* Return true if two vectors are equal. */ | |
2426 | ||
2427 | static bool | |
2428 | lambda_vector_equal (lambda_vector vec1, lambda_vector vec2, int size) | |
2429 | { | |
2430 | int i; | |
2431 | for (i = 0; i < size; i++) | |
2432 | if (vec1[i] != vec2[i]) | |
2433 | return false; | |
2434 | return true; | |
2435 | } | |
2436 | ||
2437 | /* Given an M x N integer matrix A, this function determines an M x | |
2438 | M unimodular matrix U, and an M x N echelon matrix S such that | |
2439 | "U.A = S". This decomposition is also known as "right Hermite". | |
2440 | ||
2441 | Ref: Algorithm 2.1 page 33 in "Loop Transformations for | |
2442 | Restructuring Compilers" Utpal Banerjee. */ | |
2443 | ||
2444 | static void | |
2445 | lambda_matrix_right_hermite (lambda_matrix A, int m, int n, | |
2446 | lambda_matrix S, lambda_matrix U) | |
2447 | { | |
2448 | int i, j, i0 = 0; | |
2449 | ||
2450 | lambda_matrix_copy (A, S, m, n); | |
2451 | lambda_matrix_id (U, m); | |
2452 | ||
2453 | for (j = 0; j < n; j++) | |
2454 | { | |
2455 | if (lambda_vector_first_nz (S[j], m, i0) < m) | |
2456 | { | |
2457 | ++i0; | |
2458 | for (i = m - 1; i >= i0; i--) | |
2459 | { | |
2460 | while (S[i][j] != 0) | |
2461 | { | |
2462 | int sigma, factor, a, b; | |
2463 | ||
2464 | a = S[i-1][j]; | |
2465 | b = S[i][j]; | |
2466 | sigma = (a * b < 0) ? -1: 1; | |
2467 | a = abs (a); | |
2468 | b = abs (b); | |
2469 | factor = sigma * (a / b); | |
2470 | ||
2471 | lambda_matrix_row_add (S, n, i, i-1, -factor); | |
dfcf26a5 | 2472 | std::swap (S[i], S[i-1]); |
e01f9f1f | 2473 | |
2474 | lambda_matrix_row_add (U, m, i, i-1, -factor); | |
dfcf26a5 | 2475 | std::swap (U[i], U[i-1]); |
e01f9f1f | 2476 | } |
2477 | } | |
2478 | } | |
2479 | } | |
2480 | } | |
2481 | ||
2146e26d | 2482 | /* Determines the overlapping elements due to accesses CHREC_A and |
6b421feb | 2483 | CHREC_B, that are affine functions. This function cannot handle |
2484 | symbolic evolution functions, ie. when initial conditions are | |
2485 | parameters, because it uses lambda matrices of integers. */ | |
2146e26d | 2486 | |
2487 | static void | |
48e1416a | 2488 | analyze_subscript_affine_affine (tree chrec_a, |
2146e26d | 2489 | tree chrec_b, |
48e1416a | 2490 | conflict_function **overlaps_a, |
2491 | conflict_function **overlaps_b, | |
bc3c8ad4 | 2492 | tree *last_conflicts) |
2146e26d | 2493 | { |
bc3c8ad4 | 2494 | unsigned nb_vars_a, nb_vars_b, dim; |
4ee81b10 | 2495 | HOST_WIDE_INT init_a, init_b, gamma, gcd_alpha_beta; |
bc3c8ad4 | 2496 | lambda_matrix A, U, S; |
1e33ad50 | 2497 | struct obstack scratch_obstack; |
bc3c8ad4 | 2498 | |
f84a688a | 2499 | if (eq_evolutions_p (chrec_a, chrec_b)) |
b6d41046 | 2500 | { |
f84a688a | 2501 | /* The accessed index overlaps for each iteration in the |
2502 | loop. */ | |
87da4f2e | 2503 | *overlaps_a = conflict_fn (1, affine_fn_cst (integer_zero_node)); |
2504 | *overlaps_b = conflict_fn (1, affine_fn_cst (integer_zero_node)); | |
b6d41046 | 2505 | *last_conflicts = chrec_dont_know; |
2506 | return; | |
2507 | } | |
2146e26d | 2508 | if (dump_file && (dump_flags & TDF_DETAILS)) |
2509 | fprintf (dump_file, "(analyze_subscript_affine_affine \n"); | |
48e1416a | 2510 | |
2146e26d | 2511 | /* For determining the initial intersection, we have to solve a |
2512 | Diophantine equation. This is the most time consuming part. | |
48e1416a | 2513 | |
2146e26d | 2514 | For answering to the question: "Is there a dependence?" we have |
2515 | to prove that there exists a solution to the Diophantine | |
2516 | equation, and that the solution is in the iteration domain, | |
5c9dae64 | 2517 | i.e. the solution is positive or zero, and that the solution |
2146e26d | 2518 | happens before the upper bound loop.nb_iterations. Otherwise |
2519 | there is no dependence. This function outputs a description of | |
2520 | the iterations that hold the intersections. */ | |
2521 | ||
bc3c8ad4 | 2522 | nb_vars_a = nb_vars_in_chrec (chrec_a); |
2523 | nb_vars_b = nb_vars_in_chrec (chrec_b); | |
2524 | ||
1e33ad50 | 2525 | gcc_obstack_init (&scratch_obstack); |
2526 | ||
bc3c8ad4 | 2527 | dim = nb_vars_a + nb_vars_b; |
1e33ad50 | 2528 | U = lambda_matrix_new (dim, dim, &scratch_obstack); |
2529 | A = lambda_matrix_new (dim, 1, &scratch_obstack); | |
2530 | S = lambda_matrix_new (dim, 1, &scratch_obstack); | |
bc3c8ad4 | 2531 | |
8ceddf57 | 2532 | init_a = int_cst_value (initialize_matrix_A (A, chrec_a, 0, 1)); |
2533 | init_b = int_cst_value (initialize_matrix_A (A, chrec_b, nb_vars_a, -1)); | |
bc3c8ad4 | 2534 | gamma = init_b - init_a; |
2535 | ||
2536 | /* Don't do all the hard work of solving the Diophantine equation | |
48e1416a | 2537 | when we already know the solution: for example, |
bc3c8ad4 | 2538 | | {3, +, 1}_1 |
2539 | | {3, +, 4}_2 | |
2540 | | gamma = 3 - 3 = 0. | |
48e1416a | 2541 | Then the first overlap occurs during the first iterations: |
bc3c8ad4 | 2542 | | {3, +, 1}_1 ({0, +, 4}_x) = {3, +, 4}_2 ({0, +, 1}_x) |
2543 | */ | |
2544 | if (gamma == 0) | |
2146e26d | 2545 | { |
bc3c8ad4 | 2546 | if (nb_vars_a == 1 && nb_vars_b == 1) |
2146e26d | 2547 | { |
4ee81b10 | 2548 | HOST_WIDE_INT step_a, step_b; |
d500fef3 | 2549 | HOST_WIDE_INT niter, niter_a, niter_b; |
87da4f2e | 2550 | affine_fn ova, ovb; |
bc3c8ad4 | 2551 | |
fee017b3 | 2552 | niter_a = max_stmt_executions_int (get_chrec_loop (chrec_a)); |
2553 | niter_b = max_stmt_executions_int (get_chrec_loop (chrec_b)); | |
bc3c8ad4 | 2554 | niter = MIN (niter_a, niter_b); |
58f77a87 | 2555 | step_a = int_cst_value (CHREC_RIGHT (chrec_a)); |
2556 | step_b = int_cst_value (CHREC_RIGHT (chrec_b)); | |
bc3c8ad4 | 2557 | |
48e1416a | 2558 | compute_overlap_steps_for_affine_univar (niter, step_a, step_b, |
2559 | &ova, &ovb, | |
bc3c8ad4 | 2560 | last_conflicts, 1); |
87da4f2e | 2561 | *overlaps_a = conflict_fn (1, ova); |
2562 | *overlaps_b = conflict_fn (1, ovb); | |
2146e26d | 2563 | } |
bc3c8ad4 | 2564 | |
2565 | else if (nb_vars_a == 2 && nb_vars_b == 1) | |
2566 | compute_overlap_steps_for_affine_1_2 | |
2567 | (chrec_a, chrec_b, overlaps_a, overlaps_b, last_conflicts); | |
2568 | ||
2569 | else if (nb_vars_a == 1 && nb_vars_b == 2) | |
2570 | compute_overlap_steps_for_affine_1_2 | |
2571 | (chrec_b, chrec_a, overlaps_b, overlaps_a, last_conflicts); | |
2572 | ||
2573 | else | |
2146e26d | 2574 | { |
6b421feb | 2575 | if (dump_file && (dump_flags & TDF_DETAILS)) |
2576 | fprintf (dump_file, "affine-affine test failed: too many variables.\n"); | |
87da4f2e | 2577 | *overlaps_a = conflict_fn_not_known (); |
2578 | *overlaps_b = conflict_fn_not_known (); | |
bc3c8ad4 | 2579 | *last_conflicts = chrec_dont_know; |
2146e26d | 2580 | } |
6b421feb | 2581 | goto end_analyze_subs_aa; |
bc3c8ad4 | 2582 | } |
2583 | ||
2584 | /* U.A = S */ | |
2585 | lambda_matrix_right_hermite (A, dim, 1, S, U); | |
2586 | ||
2587 | if (S[0][0] < 0) | |
2588 | { | |
2589 | S[0][0] *= -1; | |
2590 | lambda_matrix_row_negate (U, dim, 0); | |
2591 | } | |
2592 | gcd_alpha_beta = S[0][0]; | |
2593 | ||
b44d1046 | 2594 | /* Something went wrong: for example in {1, +, 0}_5 vs. {0, +, 0}_5, |
2595 | but that is a quite strange case. Instead of ICEing, answer | |
2596 | don't know. */ | |
2597 | if (gcd_alpha_beta == 0) | |
2598 | { | |
87da4f2e | 2599 | *overlaps_a = conflict_fn_not_known (); |
2600 | *overlaps_b = conflict_fn_not_known (); | |
b44d1046 | 2601 | *last_conflicts = chrec_dont_know; |
2602 | goto end_analyze_subs_aa; | |
2603 | } | |
2604 | ||
bc3c8ad4 | 2605 | /* The classic "gcd-test". */ |
2606 | if (!int_divides_p (gcd_alpha_beta, gamma)) | |
2607 | { | |
2608 | /* The "gcd-test" has determined that there is no integer | |
2609 | solution, i.e. there is no dependence. */ | |
87da4f2e | 2610 | *overlaps_a = conflict_fn_no_dependence (); |
2611 | *overlaps_b = conflict_fn_no_dependence (); | |
bc3c8ad4 | 2612 | *last_conflicts = integer_zero_node; |
2613 | } | |
2614 | ||
2615 | /* Both access functions are univariate. This includes SIV and MIV cases. */ | |
2616 | else if (nb_vars_a == 1 && nb_vars_b == 1) | |
2617 | { | |
2618 | /* Both functions should have the same evolution sign. */ | |
2619 | if (((A[0][0] > 0 && -A[1][0] > 0) | |
2620 | || (A[0][0] < 0 && -A[1][0] < 0))) | |
2146e26d | 2621 | { |
2622 | /* The solutions are given by: | |
48e1416a | 2623 | | |
bc3c8ad4 | 2624 | | [GAMMA/GCD_ALPHA_BETA t].[u11 u12] = [x0] |
2625 | | [u21 u22] [y0] | |
48e1416a | 2626 | |
2146e26d | 2627 | For a given integer t. Using the following variables, |
48e1416a | 2628 | |
2146e26d | 2629 | | i0 = u11 * gamma / gcd_alpha_beta |
2630 | | j0 = u12 * gamma / gcd_alpha_beta | |
2631 | | i1 = u21 | |
2632 | | j1 = u22 | |
48e1416a | 2633 | |
2146e26d | 2634 | the solutions are: |
48e1416a | 2635 | |
2636 | | x0 = i0 + i1 * t, | |
bc3c8ad4 | 2637 | | y0 = j0 + j1 * t. */ |
e781c550 | 2638 | HOST_WIDE_INT i0, j0, i1, j1; |
bc3c8ad4 | 2639 | |
2640 | i0 = U[0][0] * gamma / gcd_alpha_beta; | |
2641 | j0 = U[0][1] * gamma / gcd_alpha_beta; | |
2642 | i1 = U[1][0]; | |
2643 | j1 = U[1][1]; | |
2644 | ||
2645 | if ((i1 == 0 && i0 < 0) | |
2646 | || (j1 == 0 && j0 < 0)) | |
2146e26d | 2647 | { |
48e1416a | 2648 | /* There is no solution. |
2649 | FIXME: The case "i0 > nb_iterations, j0 > nb_iterations" | |
2650 | falls in here, but for the moment we don't look at the | |
2146e26d | 2651 | upper bound of the iteration domain. */ |
87da4f2e | 2652 | *overlaps_a = conflict_fn_no_dependence (); |
2653 | *overlaps_b = conflict_fn_no_dependence (); | |
bc3c8ad4 | 2654 | *last_conflicts = integer_zero_node; |
e781c550 | 2655 | goto end_analyze_subs_aa; |
bc3c8ad4 | 2656 | } |
2657 | ||
e781c550 | 2658 | if (i1 > 0 && j1 > 0) |
2146e26d | 2659 | { |
fee017b3 | 2660 | HOST_WIDE_INT niter_a |
2661 | = max_stmt_executions_int (get_chrec_loop (chrec_a)); | |
2662 | HOST_WIDE_INT niter_b | |
2663 | = max_stmt_executions_int (get_chrec_loop (chrec_b)); | |
e781c550 | 2664 | HOST_WIDE_INT niter = MIN (niter_a, niter_b); |
2665 | ||
2666 | /* (X0, Y0) is a solution of the Diophantine equation: | |
2667 | "chrec_a (X0) = chrec_b (Y0)". */ | |
2668 | HOST_WIDE_INT tau1 = MAX (CEIL (-i0, i1), | |
2669 | CEIL (-j0, j1)); | |
2670 | HOST_WIDE_INT x0 = i1 * tau1 + i0; | |
2671 | HOST_WIDE_INT y0 = j1 * tau1 + j0; | |
2672 | ||
2673 | /* (X1, Y1) is the smallest positive solution of the eq | |
2674 | "chrec_a (X1) = chrec_b (Y1)", i.e. this is where the | |
2675 | first conflict occurs. */ | |
2676 | HOST_WIDE_INT min_multiple = MIN (x0 / i1, y0 / j1); | |
2677 | HOST_WIDE_INT x1 = x0 - i1 * min_multiple; | |
2678 | HOST_WIDE_INT y1 = y0 - j1 * min_multiple; | |
2679 | ||
2680 | if (niter > 0) | |
2146e26d | 2681 | { |
e781c550 | 2682 | HOST_WIDE_INT tau2 = MIN (FLOOR_DIV (niter - i0, i1), |
2683 | FLOOR_DIV (niter - j0, j1)); | |
2684 | HOST_WIDE_INT last_conflict = tau2 - (x1 - i0)/i1; | |
bc3c8ad4 | 2685 | |
e781c550 | 2686 | /* If the overlap occurs outside of the bounds of the |
2687 | loop, there is no dependence. */ | |
63ca8934 | 2688 | if (x1 >= niter || y1 >= niter) |
2146e26d | 2689 | { |
e781c550 | 2690 | *overlaps_a = conflict_fn_no_dependence (); |
2691 | *overlaps_b = conflict_fn_no_dependence (); | |
2692 | *last_conflicts = integer_zero_node; | |
2693 | goto end_analyze_subs_aa; | |
2146e26d | 2694 | } |
2695 | else | |
e781c550 | 2696 | *last_conflicts = build_int_cst (NULL_TREE, last_conflict); |
2146e26d | 2697 | } |
2146e26d | 2698 | else |
e781c550 | 2699 | *last_conflicts = chrec_dont_know; |
2700 | ||
2701 | *overlaps_a | |
2702 | = conflict_fn (1, | |
2703 | affine_fn_univar (build_int_cst (NULL_TREE, x1), | |
2704 | 1, | |
2705 | build_int_cst (NULL_TREE, i1))); | |
2706 | *overlaps_b | |
2707 | = conflict_fn (1, | |
2708 | affine_fn_univar (build_int_cst (NULL_TREE, y1), | |
2709 | 1, | |
2710 | build_int_cst (NULL_TREE, j1))); | |
2711 | } | |
2712 | else | |
2713 | { | |
2714 | /* FIXME: For the moment, the upper bound of the | |
2715 | iteration domain for i and j is not checked. */ | |
2716 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
2717 | fprintf (dump_file, "affine-affine test failed: unimplemented.\n"); | |
2718 | *overlaps_a = conflict_fn_not_known (); | |
2719 | *overlaps_b = conflict_fn_not_known (); | |
2720 | *last_conflicts = chrec_dont_know; | |
2146e26d | 2721 | } |
2722 | } | |
bc3c8ad4 | 2723 | else |
2724 | { | |
6b421feb | 2725 | if (dump_file && (dump_flags & TDF_DETAILS)) |
2726 | fprintf (dump_file, "affine-affine test failed: unimplemented.\n"); | |
87da4f2e | 2727 | *overlaps_a = conflict_fn_not_known (); |
2728 | *overlaps_b = conflict_fn_not_known (); | |
bc3c8ad4 | 2729 | *last_conflicts = chrec_dont_know; |
2730 | } | |
2146e26d | 2731 | } |
2146e26d | 2732 | else |
2733 | { | |
6b421feb | 2734 | if (dump_file && (dump_flags & TDF_DETAILS)) |
2735 | fprintf (dump_file, "affine-affine test failed: unimplemented.\n"); | |
87da4f2e | 2736 | *overlaps_a = conflict_fn_not_known (); |
2737 | *overlaps_b = conflict_fn_not_known (); | |
bc3c8ad4 | 2738 | *last_conflicts = chrec_dont_know; |
2146e26d | 2739 | } |
bc3c8ad4 | 2740 | |
48e1416a | 2741 | end_analyze_subs_aa: |
1e33ad50 | 2742 | obstack_free (&scratch_obstack, NULL); |
2146e26d | 2743 | if (dump_file && (dump_flags & TDF_DETAILS)) |
2744 | { | |
2745 | fprintf (dump_file, " (overlaps_a = "); | |
87da4f2e | 2746 | dump_conflict_function (dump_file, *overlaps_a); |
2146e26d | 2747 | fprintf (dump_file, ")\n (overlaps_b = "); |
87da4f2e | 2748 | dump_conflict_function (dump_file, *overlaps_b); |
984cebc1 | 2749 | fprintf (dump_file, "))\n"); |
2146e26d | 2750 | } |
6b421feb | 2751 | } |
2752 | ||
2753 | /* Returns true when analyze_subscript_affine_affine can be used for | |
2754 | determining the dependence relation between chrec_a and chrec_b, | |
2755 | that contain symbols. This function modifies chrec_a and chrec_b | |
2756 | such that the analysis result is the same, and such that they don't | |
48e1416a | 2757 | contain symbols, and then can safely be passed to the analyzer. |
6b421feb | 2758 | |
2759 | Example: The analysis of the following tuples of evolutions produce | |
2760 | the same results: {x+1, +, 1}_1 vs. {x+3, +, 1}_1, and {-2, +, 1}_1 | |
2761 | vs. {0, +, 1}_1 | |
48e1416a | 2762 | |
6b421feb | 2763 | {x+1, +, 1}_1 ({2, +, 1}_1) = {x+3, +, 1}_1 ({0, +, 1}_1) |
2764 | {-2, +, 1}_1 ({2, +, 1}_1) = {0, +, 1}_1 ({0, +, 1}_1) | |
2765 | */ | |
2766 | ||
2767 | static bool | |
2768 | can_use_analyze_subscript_affine_affine (tree *chrec_a, tree *chrec_b) | |
2769 | { | |
b92381b6 | 2770 | tree diff, type, left_a, left_b, right_b; |
6b421feb | 2771 | |
2772 | if (chrec_contains_symbols (CHREC_RIGHT (*chrec_a)) | |
2773 | || chrec_contains_symbols (CHREC_RIGHT (*chrec_b))) | |
2774 | /* FIXME: For the moment not handled. Might be refined later. */ | |
2775 | return false; | |
2776 | ||
b92381b6 | 2777 | type = chrec_type (*chrec_a); |
2778 | left_a = CHREC_LEFT (*chrec_a); | |
75a70cf9 | 2779 | left_b = chrec_convert (type, CHREC_LEFT (*chrec_b), NULL); |
b92381b6 | 2780 | diff = chrec_fold_minus (type, left_a, left_b); |
2781 | ||
6b421feb | 2782 | if (!evolution_function_is_constant_p (diff)) |
2783 | return false; | |
2784 | ||
2146e26d | 2785 | if (dump_file && (dump_flags & TDF_DETAILS)) |
6b421feb | 2786 | fprintf (dump_file, "can_use_subscript_aff_aff_for_symbolic \n"); |
2787 | ||
48e1416a | 2788 | *chrec_a = build_polynomial_chrec (CHREC_VARIABLE (*chrec_a), |
6b421feb | 2789 | diff, CHREC_RIGHT (*chrec_a)); |
75a70cf9 | 2790 | right_b = chrec_convert (type, CHREC_RIGHT (*chrec_b), NULL); |
6b421feb | 2791 | *chrec_b = build_polynomial_chrec (CHREC_VARIABLE (*chrec_b), |
7c010d6a | 2792 | build_int_cst (type, 0), |
b92381b6 | 2793 | right_b); |
6b421feb | 2794 | return true; |
2146e26d | 2795 | } |
2796 | ||
2797 | /* Analyze a SIV (Single Index Variable) subscript. *OVERLAPS_A and | |
2798 | *OVERLAPS_B are initialized to the functions that describe the | |
2799 | relation between the elements accessed twice by CHREC_A and | |
2800 | CHREC_B. For k >= 0, the following property is verified: | |
2801 | ||
2802 | CHREC_A (*OVERLAPS_A (k)) = CHREC_B (*OVERLAPS_B (k)). */ | |
2803 | ||
2804 | static void | |
48e1416a | 2805 | analyze_siv_subscript (tree chrec_a, |
2146e26d | 2806 | tree chrec_b, |
48e1416a | 2807 | conflict_function **overlaps_a, |
2808 | conflict_function **overlaps_b, | |
8ceddf57 | 2809 | tree *last_conflicts, |
2810 | int loop_nest_num) | |
2146e26d | 2811 | { |
6b421feb | 2812 | dependence_stats.num_siv++; |
48e1416a | 2813 | |
2146e26d | 2814 | if (dump_file && (dump_flags & TDF_DETAILS)) |
2815 | fprintf (dump_file, "(analyze_siv_subscript \n"); | |
48e1416a | 2816 | |
2146e26d | 2817 | if (evolution_function_is_constant_p (chrec_a) |
8ceddf57 | 2818 | && evolution_function_is_affine_in_loop (chrec_b, loop_nest_num)) |
48e1416a | 2819 | analyze_siv_subscript_cst_affine (chrec_a, chrec_b, |
bc3c8ad4 | 2820 | overlaps_a, overlaps_b, last_conflicts); |
48e1416a | 2821 | |
8ceddf57 | 2822 | else if (evolution_function_is_affine_in_loop (chrec_a, loop_nest_num) |
2146e26d | 2823 | && evolution_function_is_constant_p (chrec_b)) |
48e1416a | 2824 | analyze_siv_subscript_cst_affine (chrec_b, chrec_a, |
bc3c8ad4 | 2825 | overlaps_b, overlaps_a, last_conflicts); |
48e1416a | 2826 | |
8ceddf57 | 2827 | else if (evolution_function_is_affine_in_loop (chrec_a, loop_nest_num) |
2828 | && evolution_function_is_affine_in_loop (chrec_b, loop_nest_num)) | |
6b421feb | 2829 | { |
2830 | if (!chrec_contains_symbols (chrec_a) | |
2831 | && !chrec_contains_symbols (chrec_b)) | |
2832 | { | |
48e1416a | 2833 | analyze_subscript_affine_affine (chrec_a, chrec_b, |
2834 | overlaps_a, overlaps_b, | |
6b421feb | 2835 | last_conflicts); |
2836 | ||
87da4f2e | 2837 | if (CF_NOT_KNOWN_P (*overlaps_a) |
2838 | || CF_NOT_KNOWN_P (*overlaps_b)) | |
6b421feb | 2839 | dependence_stats.num_siv_unimplemented++; |
87da4f2e | 2840 | else if (CF_NO_DEPENDENCE_P (*overlaps_a) |
2841 | || CF_NO_DEPENDENCE_P (*overlaps_b)) | |
6b421feb | 2842 | dependence_stats.num_siv_independent++; |
2843 | else | |
2844 | dependence_stats.num_siv_dependent++; | |
2845 | } | |
48e1416a | 2846 | else if (can_use_analyze_subscript_affine_affine (&chrec_a, |
6b421feb | 2847 | &chrec_b)) |
2848 | { | |
48e1416a | 2849 | analyze_subscript_affine_affine (chrec_a, chrec_b, |
2850 | overlaps_a, overlaps_b, | |
6b421feb | 2851 | last_conflicts); |
6b421feb | 2852 | |
87da4f2e | 2853 | if (CF_NOT_KNOWN_P (*overlaps_a) |
2854 | || CF_NOT_KNOWN_P (*overlaps_b)) | |
6b421feb | 2855 | dependence_stats.num_siv_unimplemented++; |
87da4f2e | 2856 | else if (CF_NO_DEPENDENCE_P (*overlaps_a) |
2857 | || CF_NO_DEPENDENCE_P (*overlaps_b)) | |
6b421feb | 2858 | dependence_stats.num_siv_independent++; |
2859 | else | |
2860 | dependence_stats.num_siv_dependent++; | |
2861 | } | |
2862 | else | |
2863 | goto siv_subscript_dontknow; | |
2864 | } | |
2865 | ||
2146e26d | 2866 | else |
2867 | { | |
6b421feb | 2868 | siv_subscript_dontknow:; |
2869 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
984cebc1 | 2870 | fprintf (dump_file, " siv test failed: unimplemented"); |
87da4f2e | 2871 | *overlaps_a = conflict_fn_not_known (); |
2872 | *overlaps_b = conflict_fn_not_known (); | |
bc3c8ad4 | 2873 | *last_conflicts = chrec_dont_know; |
6b421feb | 2874 | dependence_stats.num_siv_unimplemented++; |
2146e26d | 2875 | } |
48e1416a | 2876 | |
2146e26d | 2877 | if (dump_file && (dump_flags & TDF_DETAILS)) |
2878 | fprintf (dump_file, ")\n"); | |
2879 | } | |
2880 | ||
cabedb72 | 2881 | /* Returns false if we can prove that the greatest common divisor of the steps |
2882 | of CHREC does not divide CST, false otherwise. */ | |
2146e26d | 2883 | |
2884 | static bool | |
7ecb5bb2 | 2885 | gcd_of_steps_may_divide_p (const_tree chrec, const_tree cst) |
2146e26d | 2886 | { |
cabedb72 | 2887 | HOST_WIDE_INT cd = 0, val; |
2888 | tree step; | |
6b421feb | 2889 | |
e913b5cd | 2890 | if (!tree_fits_shwi_p (cst)) |
cabedb72 | 2891 | return true; |
e913b5cd | 2892 | val = tree_to_shwi (cst); |
cabedb72 | 2893 | |
2894 | while (TREE_CODE (chrec) == POLYNOMIAL_CHREC) | |
2895 | { | |
2896 | step = CHREC_RIGHT (chrec); | |
e913b5cd | 2897 | if (!tree_fits_shwi_p (step)) |
cabedb72 | 2898 | return true; |
e913b5cd | 2899 | cd = gcd (cd, tree_to_shwi (step)); |
cabedb72 | 2900 | chrec = CHREC_LEFT (chrec); |
2146e26d | 2901 | } |
cabedb72 | 2902 | |
2903 | return val % cd == 0; | |
2146e26d | 2904 | } |
2905 | ||
8c7b1d8e | 2906 | /* Analyze a MIV (Multiple Index Variable) subscript with respect to |
2907 | LOOP_NEST. *OVERLAPS_A and *OVERLAPS_B are initialized to the | |
2908 | functions that describe the relation between the elements accessed | |
2909 | twice by CHREC_A and CHREC_B. For k >= 0, the following property | |
2910 | is verified: | |
2146e26d | 2911 | |
2912 | CHREC_A (*OVERLAPS_A (k)) = CHREC_B (*OVERLAPS_B (k)). */ | |
2913 | ||
2914 | static void | |
48e1416a | 2915 | analyze_miv_subscript (tree chrec_a, |
2916 | tree chrec_b, | |
2917 | conflict_function **overlaps_a, | |
2918 | conflict_function **overlaps_b, | |
8c7b1d8e | 2919 | tree *last_conflicts, |
2920 | struct loop *loop_nest) | |
2146e26d | 2921 | { |
9a3dad1c | 2922 | tree type, difference; |
2923 | ||
6b421feb | 2924 | dependence_stats.num_miv++; |
2146e26d | 2925 | if (dump_file && (dump_flags & TDF_DETAILS)) |
2926 | fprintf (dump_file, "(analyze_miv_subscript \n"); | |
f84a688a | 2927 | |
9a3dad1c | 2928 | type = signed_type_for_types (TREE_TYPE (chrec_a), TREE_TYPE (chrec_b)); |
75a70cf9 | 2929 | chrec_a = chrec_convert (type, chrec_a, NULL); |
2930 | chrec_b = chrec_convert (type, chrec_b, NULL); | |
9a3dad1c | 2931 | difference = chrec_fold_minus (type, chrec_a, chrec_b); |
48e1416a | 2932 | |
f84a688a | 2933 | if (eq_evolutions_p (chrec_a, chrec_b)) |
2146e26d | 2934 | { |
2935 | /* Access functions are the same: all the elements are accessed | |
2936 | in the same order. */ | |
87da4f2e | 2937 | *overlaps_a = conflict_fn (1, affine_fn_cst (integer_zero_node)); |
2938 | *overlaps_b = conflict_fn (1, affine_fn_cst (integer_zero_node)); | |
8fe79ba5 | 2939 | *last_conflicts = max_stmt_executions_tree (get_chrec_loop (chrec_a)); |
6b421feb | 2940 | dependence_stats.num_miv_dependent++; |
2146e26d | 2941 | } |
48e1416a | 2942 | |
2146e26d | 2943 | else if (evolution_function_is_constant_p (difference) |
2944 | /* For the moment, the following is verified: | |
8c7b1d8e | 2945 | evolution_function_is_affine_multivariate_p (chrec_a, |
2946 | loop_nest->num) */ | |
cabedb72 | 2947 | && !gcd_of_steps_may_divide_p (chrec_a, difference)) |
2146e26d | 2948 | { |
2949 | /* testsuite/.../ssa-chrec-33.c | |
48e1416a | 2950 | {{21, +, 2}_1, +, -2}_2 vs. {{20, +, 2}_1, +, -2}_2 |
2951 | ||
cabedb72 | 2952 | The difference is 1, and all the evolution steps are multiples |
2953 | of 2, consequently there are no overlapping elements. */ | |
87da4f2e | 2954 | *overlaps_a = conflict_fn_no_dependence (); |
2955 | *overlaps_b = conflict_fn_no_dependence (); | |
bc3c8ad4 | 2956 | *last_conflicts = integer_zero_node; |
6b421feb | 2957 | dependence_stats.num_miv_independent++; |
2146e26d | 2958 | } |
48e1416a | 2959 | |
8c7b1d8e | 2960 | else if (evolution_function_is_affine_multivariate_p (chrec_a, loop_nest->num) |
6b421feb | 2961 | && !chrec_contains_symbols (chrec_a) |
8c7b1d8e | 2962 | && evolution_function_is_affine_multivariate_p (chrec_b, loop_nest->num) |
6b421feb | 2963 | && !chrec_contains_symbols (chrec_b)) |
2146e26d | 2964 | { |
2965 | /* testsuite/.../ssa-chrec-35.c | |
2966 | {0, +, 1}_2 vs. {0, +, 1}_3 | |
2967 | the overlapping elements are respectively located at iterations: | |
48e1416a | 2968 | {0, +, 1}_x and {0, +, 1}_x, |
2969 | in other words, we have the equality: | |
bc3c8ad4 | 2970 | {0, +, 1}_2 ({0, +, 1}_x) = {0, +, 1}_3 ({0, +, 1}_x) |
48e1416a | 2971 | |
2972 | Other examples: | |
2973 | {{0, +, 1}_1, +, 2}_2 ({0, +, 1}_x, {0, +, 1}_y) = | |
bc3c8ad4 | 2974 | {0, +, 1}_1 ({{0, +, 1}_x, +, 2}_y) |
2975 | ||
48e1416a | 2976 | {{0, +, 2}_1, +, 3}_2 ({0, +, 1}_y, {0, +, 1}_x) = |
bc3c8ad4 | 2977 | {{0, +, 3}_1, +, 2}_2 ({0, +, 1}_x, {0, +, 1}_y) |
2146e26d | 2978 | */ |
48e1416a | 2979 | analyze_subscript_affine_affine (chrec_a, chrec_b, |
bc3c8ad4 | 2980 | overlaps_a, overlaps_b, last_conflicts); |
6b421feb | 2981 | |
87da4f2e | 2982 | if (CF_NOT_KNOWN_P (*overlaps_a) |
2983 | || CF_NOT_KNOWN_P (*overlaps_b)) | |
6b421feb | 2984 | dependence_stats.num_miv_unimplemented++; |
87da4f2e | 2985 | else if (CF_NO_DEPENDENCE_P (*overlaps_a) |
2986 | || CF_NO_DEPENDENCE_P (*overlaps_b)) | |
6b421feb | 2987 | dependence_stats.num_miv_independent++; |
2988 | else | |
2989 | dependence_stats.num_miv_dependent++; | |
2146e26d | 2990 | } |
48e1416a | 2991 | |
2146e26d | 2992 | else |
2993 | { | |
2994 | /* When the analysis is too difficult, answer "don't know". */ | |
6b421feb | 2995 | if (dump_file && (dump_flags & TDF_DETAILS)) |
2996 | fprintf (dump_file, "analyze_miv_subscript test failed: unimplemented.\n"); | |
2997 | ||
87da4f2e | 2998 | *overlaps_a = conflict_fn_not_known (); |
2999 | *overlaps_b = conflict_fn_not_known (); | |
bc3c8ad4 | 3000 | *last_conflicts = chrec_dont_know; |
6b421feb | 3001 | dependence_stats.num_miv_unimplemented++; |
2146e26d | 3002 | } |
48e1416a | 3003 | |
2146e26d | 3004 | if (dump_file && (dump_flags & TDF_DETAILS)) |
3005 | fprintf (dump_file, ")\n"); | |
3006 | } | |
3007 | ||
8c7b1d8e | 3008 | /* Determines the iterations for which CHREC_A is equal to CHREC_B in |
3009 | with respect to LOOP_NEST. OVERLAP_ITERATIONS_A and | |
3010 | OVERLAP_ITERATIONS_B are initialized with two functions that | |
3011 | describe the iterations that contain conflicting elements. | |
48e1416a | 3012 | |
2146e26d | 3013 | Remark: For an integer k >= 0, the following equality is true: |
48e1416a | 3014 | |
2146e26d | 3015 | CHREC_A (OVERLAP_ITERATIONS_A (k)) == CHREC_B (OVERLAP_ITERATIONS_B (k)). |
3016 | */ | |
3017 | ||
48e1416a | 3018 | static void |
3019 | analyze_overlapping_iterations (tree chrec_a, | |
3020 | tree chrec_b, | |
3021 | conflict_function **overlap_iterations_a, | |
3022 | conflict_function **overlap_iterations_b, | |
8c7b1d8e | 3023 | tree *last_conflicts, struct loop *loop_nest) |
2146e26d | 3024 | { |
8c7b1d8e | 3025 | unsigned int lnn = loop_nest->num; |
3026 | ||
6b421feb | 3027 | dependence_stats.num_subscript_tests++; |
48e1416a | 3028 | |
2146e26d | 3029 | if (dump_file && (dump_flags & TDF_DETAILS)) |
3030 | { | |
3031 | fprintf (dump_file, "(analyze_overlapping_iterations \n"); | |
3032 | fprintf (dump_file, " (chrec_a = "); | |
3033 | print_generic_expr (dump_file, chrec_a, 0); | |
6b421feb | 3034 | fprintf (dump_file, ")\n (chrec_b = "); |
2146e26d | 3035 | print_generic_expr (dump_file, chrec_b, 0); |
3036 | fprintf (dump_file, ")\n"); | |
3037 | } | |
6b421feb | 3038 | |
2146e26d | 3039 | if (chrec_a == NULL_TREE |
3040 | || chrec_b == NULL_TREE | |
3041 | || chrec_contains_undetermined (chrec_a) | |
6b421feb | 3042 | || chrec_contains_undetermined (chrec_b)) |
2146e26d | 3043 | { |
6b421feb | 3044 | dependence_stats.num_subscript_undetermined++; |
48e1416a | 3045 | |
87da4f2e | 3046 | *overlap_iterations_a = conflict_fn_not_known (); |
3047 | *overlap_iterations_b = conflict_fn_not_known (); | |
2146e26d | 3048 | } |
6b421feb | 3049 | |
48e1416a | 3050 | /* If they are the same chrec, and are affine, they overlap |
6b421feb | 3051 | on every iteration. */ |
3052 | else if (eq_evolutions_p (chrec_a, chrec_b) | |
328218c2 | 3053 | && (evolution_function_is_affine_multivariate_p (chrec_a, lnn) |
3054 | || operand_equal_p (chrec_a, chrec_b, 0))) | |
6b421feb | 3055 | { |
3056 | dependence_stats.num_same_subscript_function++; | |
87da4f2e | 3057 | *overlap_iterations_a = conflict_fn (1, affine_fn_cst (integer_zero_node)); |
3058 | *overlap_iterations_b = conflict_fn (1, affine_fn_cst (integer_zero_node)); | |
6b421feb | 3059 | *last_conflicts = chrec_dont_know; |
3060 | } | |
3061 | ||
3062 | /* If they aren't the same, and aren't affine, we can't do anything | |
328218c2 | 3063 | yet. */ |
48e1416a | 3064 | else if ((chrec_contains_symbols (chrec_a) |
6b421feb | 3065 | || chrec_contains_symbols (chrec_b)) |
8c7b1d8e | 3066 | && (!evolution_function_is_affine_multivariate_p (chrec_a, lnn) |
3067 | || !evolution_function_is_affine_multivariate_p (chrec_b, lnn))) | |
6b421feb | 3068 | { |
3069 | dependence_stats.num_subscript_undetermined++; | |
87da4f2e | 3070 | *overlap_iterations_a = conflict_fn_not_known (); |
3071 | *overlap_iterations_b = conflict_fn_not_known (); | |
6b421feb | 3072 | } |
3073 | ||
2146e26d | 3074 | else if (ziv_subscript_p (chrec_a, chrec_b)) |
48e1416a | 3075 | analyze_ziv_subscript (chrec_a, chrec_b, |
bc3c8ad4 | 3076 | overlap_iterations_a, overlap_iterations_b, |
3077 | last_conflicts); | |
48e1416a | 3078 | |
2146e26d | 3079 | else if (siv_subscript_p (chrec_a, chrec_b)) |
48e1416a | 3080 | analyze_siv_subscript (chrec_a, chrec_b, |
3081 | overlap_iterations_a, overlap_iterations_b, | |
8ceddf57 | 3082 | last_conflicts, lnn); |
48e1416a | 3083 | |
2146e26d | 3084 | else |
48e1416a | 3085 | analyze_miv_subscript (chrec_a, chrec_b, |
bc3c8ad4 | 3086 | overlap_iterations_a, overlap_iterations_b, |
8c7b1d8e | 3087 | last_conflicts, loop_nest); |
48e1416a | 3088 | |
2146e26d | 3089 | if (dump_file && (dump_flags & TDF_DETAILS)) |
3090 | { | |
3091 | fprintf (dump_file, " (overlap_iterations_a = "); | |
87da4f2e | 3092 | dump_conflict_function (dump_file, *overlap_iterations_a); |
2146e26d | 3093 | fprintf (dump_file, ")\n (overlap_iterations_b = "); |
87da4f2e | 3094 | dump_conflict_function (dump_file, *overlap_iterations_b); |
984cebc1 | 3095 | fprintf (dump_file, "))\n"); |
2146e26d | 3096 | } |
3097 | } | |
3098 | ||
b44d1046 | 3099 | /* Helper function for uniquely inserting distance vectors. */ |
2146e26d | 3100 | |
b44d1046 | 3101 | static void |
3102 | save_dist_v (struct data_dependence_relation *ddr, lambda_vector dist_v) | |
3103 | { | |
3104 | unsigned i; | |
3105 | lambda_vector v; | |
2146e26d | 3106 | |
f1f41a6c | 3107 | FOR_EACH_VEC_ELT (DDR_DIST_VECTS (ddr), i, v) |
b44d1046 | 3108 | if (lambda_vector_equal (v, dist_v, DDR_NB_LOOPS (ddr))) |
3109 | return; | |
2146e26d | 3110 | |
f1f41a6c | 3111 | DDR_DIST_VECTS (ddr).safe_push (dist_v); |
b44d1046 | 3112 | } |
2146e26d | 3113 | |
b44d1046 | 3114 | /* Helper function for uniquely inserting direction vectors. */ |
3115 | ||
3116 | static void | |
3117 | save_dir_v (struct data_dependence_relation *ddr, lambda_vector dir_v) | |
2146e26d | 3118 | { |
3119 | unsigned i; | |
b44d1046 | 3120 | lambda_vector v; |
6b421feb | 3121 | |
f1f41a6c | 3122 | FOR_EACH_VEC_ELT (DDR_DIR_VECTS (ddr), i, v) |
b44d1046 | 3123 | if (lambda_vector_equal (v, dir_v, DDR_NB_LOOPS (ddr))) |
3124 | return; | |
3125 | ||
f1f41a6c | 3126 | DDR_DIR_VECTS (ddr).safe_push (dir_v); |
b44d1046 | 3127 | } |
3128 | ||
3129 | /* Add a distance of 1 on all the loops outer than INDEX. If we | |
3130 | haven't yet determined a distance for this outer loop, push a new | |
3131 | distance vector composed of the previous distance, and a distance | |
3132 | of 1 for this outer loop. Example: | |
3133 | ||
3134 | | loop_1 | |
3135 | | loop_2 | |
3136 | | A[10] | |
3137 | | endloop_2 | |
3138 | | endloop_1 | |
3139 | ||
3140 | Saved vectors are of the form (dist_in_1, dist_in_2). First, we | |
3141 | save (0, 1), then we have to save (1, 0). */ | |
3142 | ||
3143 | static void | |
3144 | add_outer_distances (struct data_dependence_relation *ddr, | |
3145 | lambda_vector dist_v, int index) | |
3146 | { | |
3147 | /* For each outer loop where init_v is not set, the accesses are | |
3148 | in dependence of distance 1 in the loop. */ | |
3149 | while (--index >= 0) | |
3150 | { | |
3151 | lambda_vector save_v = lambda_vector_new (DDR_NB_LOOPS (ddr)); | |
3152 | lambda_vector_copy (dist_v, save_v, DDR_NB_LOOPS (ddr)); | |
3153 | save_v[index] = 1; | |
3154 | save_dist_v (ddr, save_v); | |
3155 | } | |
3156 | } | |
3157 | ||
3158 | /* Return false when fail to represent the data dependence as a | |
3159 | distance vector. INIT_B is set to true when a component has been | |
3160 | added to the distance vector DIST_V. INDEX_CARRY is then set to | |
3161 | the index in DIST_V that carries the dependence. */ | |
3162 | ||
3163 | static bool | |
3164 | build_classic_dist_vector_1 (struct data_dependence_relation *ddr, | |
3165 | struct data_reference *ddr_a, | |
3166 | struct data_reference *ddr_b, | |
3167 | lambda_vector dist_v, bool *init_b, | |
3168 | int *index_carry) | |
3169 | { | |
3170 | unsigned i; | |
3171 | lambda_vector init_v = lambda_vector_new (DDR_NB_LOOPS (ddr)); | |
6b421feb | 3172 | |
6b6f234c | 3173 | for (i = 0; i < DDR_NUM_SUBSCRIPTS (ddr); i++) |
2146e26d | 3174 | { |
bc3c8ad4 | 3175 | tree access_fn_a, access_fn_b; |
6b6f234c | 3176 | struct subscript *subscript = DDR_SUBSCRIPT (ddr, i); |
2146e26d | 3177 | |
3178 | if (chrec_contains_undetermined (SUB_DISTANCE (subscript))) | |
bc3c8ad4 | 3179 | { |
3180 | non_affine_dependence_relation (ddr); | |
b44d1046 | 3181 | return false; |
bc3c8ad4 | 3182 | } |
3183 | ||
b44d1046 | 3184 | access_fn_a = DR_ACCESS_FN (ddr_a, i); |
3185 | access_fn_b = DR_ACCESS_FN (ddr_b, i); | |
2146e26d | 3186 | |
48e1416a | 3187 | if (TREE_CODE (access_fn_a) == POLYNOMIAL_CHREC |
bc3c8ad4 | 3188 | && TREE_CODE (access_fn_b) == POLYNOMIAL_CHREC) |
2146e26d | 3189 | { |
b44d1046 | 3190 | int dist, index; |
ba3fc4d0 | 3191 | int var_a = CHREC_VARIABLE (access_fn_a); |
3192 | int var_b = CHREC_VARIABLE (access_fn_b); | |
b44d1046 | 3193 | |
ba3fc4d0 | 3194 | if (var_a != var_b |
3195 | || chrec_contains_undetermined (SUB_DISTANCE (subscript))) | |
bc3c8ad4 | 3196 | { |
3197 | non_affine_dependence_relation (ddr); | |
b44d1046 | 3198 | return false; |
bc3c8ad4 | 3199 | } |
48e1416a | 3200 | |
58f77a87 | 3201 | dist = int_cst_value (SUB_DISTANCE (subscript)); |
ba3fc4d0 | 3202 | index = index_in_loop_nest (var_a, DDR_LOOP_NEST (ddr)); |
3203 | *index_carry = MIN (index, *index_carry); | |
2146e26d | 3204 | |
b44d1046 | 3205 | /* This is the subscript coupling test. If we have already |
3206 | recorded a distance for this loop (a distance coming from | |
3207 | another subscript), it should be the same. For example, | |
3208 | in the following code, there is no dependence: | |
3209 | ||
2146e26d | 3210 | | loop i = 0, N, 1 |
3211 | | T[i+1][i] = ... | |
3212 | | ... = T[i][i] | |
3213 | | endloop | |
b44d1046 | 3214 | */ |
3215 | if (init_v[index] != 0 && dist_v[index] != dist) | |
2146e26d | 3216 | { |
6b6f234c | 3217 | finalize_ddr_dependent (ddr, chrec_known); |
b44d1046 | 3218 | return false; |
2146e26d | 3219 | } |
3220 | ||
b44d1046 | 3221 | dist_v[index] = dist; |
3222 | init_v[index] = 1; | |
3223 | *init_b = true; | |
3224 | } | |
9c77efff | 3225 | else if (!operand_equal_p (access_fn_a, access_fn_b, 0)) |
b44d1046 | 3226 | { |
3227 | /* This can be for example an affine vs. constant dependence | |
3228 | (T[i] vs. T[3]) that is not an affine dependence and is | |
3229 | not representable as a distance vector. */ | |
3230 | non_affine_dependence_relation (ddr); | |
3231 | return false; | |
2146e26d | 3232 | } |
3233 | } | |
1532ec98 | 3234 | |
b44d1046 | 3235 | return true; |
3236 | } | |
1532ec98 | 3237 | |
2a21a5df | 3238 | /* Return true when the DDR contains only constant access functions. */ |
3239 | ||
3240 | static bool | |
7ecb5bb2 | 3241 | constant_access_functions (const struct data_dependence_relation *ddr) |
2a21a5df | 3242 | { |
3243 | unsigned i; | |
3244 | ||
3245 | for (i = 0; i < DDR_NUM_SUBSCRIPTS (ddr); i++) | |
3246 | if (!evolution_function_is_constant_p (DR_ACCESS_FN (DDR_A (ddr), i)) | |
3247 | || !evolution_function_is_constant_p (DR_ACCESS_FN (DDR_B (ddr), i))) | |
3248 | return false; | |
3249 | ||
3250 | return true; | |
3251 | } | |
3252 | ||
b44d1046 | 3253 | /* Helper function for the case where DDR_A and DDR_B are the same |
bfe20447 | 3254 | multivariate access function with a constant step. For an example |
3255 | see pr34635-1.c. */ | |
bc3c8ad4 | 3256 | |
b44d1046 | 3257 | static void |
3258 | add_multivariate_self_dist (struct data_dependence_relation *ddr, tree c_2) | |
3259 | { | |
3260 | int x_1, x_2; | |
3261 | tree c_1 = CHREC_LEFT (c_2); | |
3262 | tree c_0 = CHREC_LEFT (c_1); | |
3263 | lambda_vector dist_v; | |
8fb896c4 | 3264 | int v1, v2, cd; |
bc3c8ad4 | 3265 | |
bb11a0e8 | 3266 | /* Polynomials with more than 2 variables are not handled yet. When |
3267 | the evolution steps are parameters, it is not possible to | |
3268 | represent the dependence using classical distance vectors. */ | |
3269 | if (TREE_CODE (c_0) != INTEGER_CST | |
3270 | || TREE_CODE (CHREC_RIGHT (c_1)) != INTEGER_CST | |
3271 | || TREE_CODE (CHREC_RIGHT (c_2)) != INTEGER_CST) | |
3272 | { | |
3273 | DDR_AFFINE_P (ddr) = false; | |
b44d1046 | 3274 | return; |
3275 | } | |
1532ec98 | 3276 | |
b44d1046 | 3277 | x_2 = index_in_loop_nest (CHREC_VARIABLE (c_2), DDR_LOOP_NEST (ddr)); |
3278 | x_1 = index_in_loop_nest (CHREC_VARIABLE (c_1), DDR_LOOP_NEST (ddr)); | |
1532ec98 | 3279 | |
b44d1046 | 3280 | /* For "{{0, +, 2}_1, +, 3}_2" the distance vector is (3, -2). */ |
3281 | dist_v = lambda_vector_new (DDR_NB_LOOPS (ddr)); | |
58f77a87 | 3282 | v1 = int_cst_value (CHREC_RIGHT (c_1)); |
3283 | v2 = int_cst_value (CHREC_RIGHT (c_2)); | |
8fb896c4 | 3284 | cd = gcd (v1, v2); |
3285 | v1 /= cd; | |
3286 | v2 /= cd; | |
3287 | ||
3288 | if (v2 < 0) | |
3289 | { | |
3290 | v2 = -v2; | |
3291 | v1 = -v1; | |
3292 | } | |
3293 | ||
3294 | dist_v[x_1] = v2; | |
3295 | dist_v[x_2] = -v1; | |
b44d1046 | 3296 | save_dist_v (ddr, dist_v); |
1532ec98 | 3297 | |
b44d1046 | 3298 | add_outer_distances (ddr, dist_v, x_1); |
3299 | } | |
1532ec98 | 3300 | |
b44d1046 | 3301 | /* Helper function for the case where DDR_A and DDR_B are the same |
3302 | access functions. */ | |
ffe944fa | 3303 | |
b44d1046 | 3304 | static void |
3305 | add_other_self_distances (struct data_dependence_relation *ddr) | |
3306 | { | |
3307 | lambda_vector dist_v; | |
3308 | unsigned i; | |
3309 | int index_carry = DDR_NB_LOOPS (ddr); | |
1532ec98 | 3310 | |
b44d1046 | 3311 | for (i = 0; i < DDR_NUM_SUBSCRIPTS (ddr); i++) |
ffe944fa | 3312 | { |
b44d1046 | 3313 | tree access_fun = DR_ACCESS_FN (DDR_A (ddr), i); |
1532ec98 | 3314 | |
b44d1046 | 3315 | if (TREE_CODE (access_fun) == POLYNOMIAL_CHREC) |
1532ec98 | 3316 | { |
b44d1046 | 3317 | if (!evolution_function_is_univariate_p (access_fun)) |
3318 | { | |
3319 | if (DDR_NUM_SUBSCRIPTS (ddr) != 1) | |
3320 | { | |
3321 | DDR_ARE_DEPENDENT (ddr) = chrec_dont_know; | |
3322 | return; | |
3323 | } | |
3324 | ||
bfe20447 | 3325 | access_fun = DR_ACCESS_FN (DDR_A (ddr), 0); |
3326 | ||
3327 | if (TREE_CODE (CHREC_LEFT (access_fun)) == POLYNOMIAL_CHREC) | |
3328 | add_multivariate_self_dist (ddr, access_fun); | |
3329 | else | |
3330 | /* The evolution step is not constant: it varies in | |
3331 | the outer loop, so this cannot be represented by a | |
3332 | distance vector. For example in pr34635.c the | |
3333 | evolution is {0, +, {0, +, 4}_1}_2. */ | |
3334 | DDR_AFFINE_P (ddr) = false; | |
3335 | ||
b44d1046 | 3336 | return; |
3337 | } | |
3338 | ||
3339 | index_carry = MIN (index_carry, | |
3340 | index_in_loop_nest (CHREC_VARIABLE (access_fun), | |
3341 | DDR_LOOP_NEST (ddr))); | |
1532ec98 | 3342 | } |
ffe944fa | 3343 | } |
3344 | ||
b44d1046 | 3345 | dist_v = lambda_vector_new (DDR_NB_LOOPS (ddr)); |
3346 | add_outer_distances (ddr, dist_v, index_carry); | |
2146e26d | 3347 | } |
3348 | ||
2a21a5df | 3349 | static void |
3350 | insert_innermost_unit_dist_vector (struct data_dependence_relation *ddr) | |
3351 | { | |
3352 | lambda_vector dist_v = lambda_vector_new (DDR_NB_LOOPS (ddr)); | |
3353 | ||
3354 | dist_v[DDR_INNER_LOOP (ddr)] = 1; | |
3355 | save_dist_v (ddr, dist_v); | |
3356 | } | |
3357 | ||
3358 | /* Adds a unit distance vector to DDR when there is a 0 overlap. This | |
3359 | is the case for example when access functions are the same and | |
3360 | equal to a constant, as in: | |
3361 | ||
3362 | | loop_1 | |
3363 | | A[3] = ... | |
3364 | | ... = A[3] | |
3365 | | endloop_1 | |
3366 | ||
3367 | in which case the distance vectors are (0) and (1). */ | |
3368 | ||
3369 | static void | |
3370 | add_distance_for_zero_overlaps (struct data_dependence_relation *ddr) | |
3371 | { | |
3372 | unsigned i, j; | |
3373 | ||
3374 | for (i = 0; i < DDR_NUM_SUBSCRIPTS (ddr); i++) | |
3375 | { | |
3376 | subscript_p sub = DDR_SUBSCRIPT (ddr, i); | |
3377 | conflict_function *ca = SUB_CONFLICTS_IN_A (sub); | |
3378 | conflict_function *cb = SUB_CONFLICTS_IN_B (sub); | |
3379 | ||
3380 | for (j = 0; j < ca->n; j++) | |
3381 | if (affine_function_zero_p (ca->fns[j])) | |
3382 | { | |
3383 | insert_innermost_unit_dist_vector (ddr); | |
3384 | return; | |
3385 | } | |
3386 | ||
3387 | for (j = 0; j < cb->n; j++) | |
3388 | if (affine_function_zero_p (cb->fns[j])) | |
3389 | { | |
3390 | insert_innermost_unit_dist_vector (ddr); | |
3391 | return; | |
3392 | } | |
3393 | } | |
3394 | } | |
3395 | ||
b44d1046 | 3396 | /* Compute the classic per loop distance vector. DDR is the data |
3397 | dependence relation to build a vector from. Return false when fail | |
3398 | to represent the data dependence as a distance vector. */ | |
2146e26d | 3399 | |
557ef5d8 | 3400 | static bool |
8c7b1d8e | 3401 | build_classic_dist_vector (struct data_dependence_relation *ddr, |
3402 | struct loop *loop_nest) | |
2146e26d | 3403 | { |
1532ec98 | 3404 | bool init_b = false; |
b44d1046 | 3405 | int index_carry = DDR_NB_LOOPS (ddr); |
3406 | lambda_vector dist_v; | |
1532ec98 | 3407 | |
6b6f234c | 3408 | if (DDR_ARE_DEPENDENT (ddr) != NULL_TREE) |
d534eb36 | 3409 | return false; |
b44d1046 | 3410 | |
3411 | if (same_access_functions (ddr)) | |
2146e26d | 3412 | { |
b44d1046 | 3413 | /* Save the 0 vector. */ |
3414 | dist_v = lambda_vector_new (DDR_NB_LOOPS (ddr)); | |
3415 | save_dist_v (ddr, dist_v); | |
2146e26d | 3416 | |
2a21a5df | 3417 | if (constant_access_functions (ddr)) |
3418 | add_distance_for_zero_overlaps (ddr); | |
3419 | ||
b44d1046 | 3420 | if (DDR_NB_LOOPS (ddr) > 1) |
3421 | add_other_self_distances (ddr); | |
bc3c8ad4 | 3422 | |
b44d1046 | 3423 | return true; |
3424 | } | |
bc3c8ad4 | 3425 | |
b44d1046 | 3426 | dist_v = lambda_vector_new (DDR_NB_LOOPS (ddr)); |
3427 | if (!build_classic_dist_vector_1 (ddr, DDR_A (ddr), DDR_B (ddr), | |
3428 | dist_v, &init_b, &index_carry)) | |
3429 | return false; | |
bc3c8ad4 | 3430 | |
b44d1046 | 3431 | /* Save the distance vector if we initialized one. */ |
3432 | if (init_b) | |
3433 | { | |
3434 | /* Verify a basic constraint: classic distance vectors should | |
3435 | always be lexicographically positive. | |
3436 | ||
3437 | Data references are collected in the order of execution of | |
3438 | the program, thus for the following loop | |
3439 | ||
3440 | | for (i = 1; i < 100; i++) | |
3441 | | for (j = 1; j < 100; j++) | |
3442 | | { | |
3443 | | t = T[j+1][i-1]; // A | |
3444 | | T[j][i] = t + 2; // B | |
3445 | | } | |
3446 | ||
3447 | references are collected following the direction of the wind: | |
3448 | A then B. The data dependence tests are performed also | |
3449 | following this order, such that we're looking at the distance | |
3450 | separating the elements accessed by A from the elements later | |
3451 | accessed by B. But in this example, the distance returned by | |
3452 | test_dep (A, B) is lexicographically negative (-1, 1), that | |
3453 | means that the access A occurs later than B with respect to | |
3454 | the outer loop, ie. we're actually looking upwind. In this | |
3455 | case we solve test_dep (B, A) looking downwind to the | |
3456 | lexicographically positive solution, that returns the | |
3457 | distance vector (1, -1). */ | |
3458 | if (!lambda_vector_lexico_pos (dist_v, DDR_NB_LOOPS (ddr))) | |
3459 | { | |
3460 | lambda_vector save_v = lambda_vector_new (DDR_NB_LOOPS (ddr)); | |
d534eb36 | 3461 | if (!subscript_dependence_tester_1 (ddr, DDR_B (ddr), DDR_A (ddr), |
3462 | loop_nest)) | |
3463 | return false; | |
b44d1046 | 3464 | compute_subscript_distance (ddr); |
d534eb36 | 3465 | if (!build_classic_dist_vector_1 (ddr, DDR_B (ddr), DDR_A (ddr), |
3466 | save_v, &init_b, &index_carry)) | |
3467 | return false; | |
b44d1046 | 3468 | save_dist_v (ddr, save_v); |
0ecb94cf | 3469 | DDR_REVERSED_P (ddr) = true; |
b44d1046 | 3470 | |
3471 | /* In this case there is a dependence forward for all the | |
3472 | outer loops: | |
3473 | ||
3474 | | for (k = 1; k < 100; k++) | |
3475 | | for (i = 1; i < 100; i++) | |
3476 | | for (j = 1; j < 100; j++) | |
3477 | | { | |
3478 | | t = T[j+1][i-1]; // A | |
3479 | | T[j][i] = t + 2; // B | |
3480 | | } | |
3481 | ||
48e1416a | 3482 | the vectors are: |
b44d1046 | 3483 | (0, 1, -1) |
3484 | (1, 1, -1) | |
3485 | (1, -1, 1) | |
3486 | */ | |
3487 | if (DDR_NB_LOOPS (ddr) > 1) | |
3488 | { | |
3489 | add_outer_distances (ddr, save_v, index_carry); | |
3490 | add_outer_distances (ddr, dist_v, index_carry); | |
bc3c8ad4 | 3491 | } |
b44d1046 | 3492 | } |
3493 | else | |
3494 | { | |
3495 | lambda_vector save_v = lambda_vector_new (DDR_NB_LOOPS (ddr)); | |
3496 | lambda_vector_copy (dist_v, save_v, DDR_NB_LOOPS (ddr)); | |
bc3c8ad4 | 3497 | |
b44d1046 | 3498 | if (DDR_NB_LOOPS (ddr) > 1) |
2146e26d | 3499 | { |
b44d1046 | 3500 | lambda_vector opposite_v = lambda_vector_new (DDR_NB_LOOPS (ddr)); |
bc3c8ad4 | 3501 | |
d534eb36 | 3502 | if (!subscript_dependence_tester_1 (ddr, DDR_B (ddr), |
3503 | DDR_A (ddr), loop_nest)) | |
3504 | return false; | |
b44d1046 | 3505 | compute_subscript_distance (ddr); |
d534eb36 | 3506 | if (!build_classic_dist_vector_1 (ddr, DDR_B (ddr), DDR_A (ddr), |
3507 | opposite_v, &init_b, | |
3508 | &index_carry)) | |
3509 | return false; | |
bc3c8ad4 | 3510 | |
d534eb36 | 3511 | save_dist_v (ddr, save_v); |
b44d1046 | 3512 | add_outer_distances (ddr, dist_v, index_carry); |
3513 | add_outer_distances (ddr, opposite_v, index_carry); | |
2146e26d | 3514 | } |
d534eb36 | 3515 | else |
3516 | save_dist_v (ddr, save_v); | |
2146e26d | 3517 | } |
3518 | } | |
b44d1046 | 3519 | else |
3520 | { | |
3521 | /* There is a distance of 1 on all the outer loops: Example: | |
3522 | there is a dependence of distance 1 on loop_1 for the array A. | |
1532ec98 | 3523 | |
b44d1046 | 3524 | | loop_1 |
3525 | | A[5] = ... | |
3526 | | endloop | |
3527 | */ | |
3528 | add_outer_distances (ddr, dist_v, | |
3529 | lambda_vector_first_nz (dist_v, | |
3530 | DDR_NB_LOOPS (ddr), 0)); | |
3531 | } | |
3532 | ||
3533 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
1532ec98 | 3534 | { |
b44d1046 | 3535 | unsigned i; |
1532ec98 | 3536 | |
b44d1046 | 3537 | fprintf (dump_file, "(build_classic_dist_vector\n"); |
3538 | for (i = 0; i < DDR_NUM_DIST_VECTS (ddr); i++) | |
3539 | { | |
3540 | fprintf (dump_file, " dist_vector = ("); | |
3541 | print_lambda_vector (dump_file, DDR_DIST_VECT (ddr, i), | |
3542 | DDR_NB_LOOPS (ddr)); | |
3543 | fprintf (dump_file, " )\n"); | |
3544 | } | |
3545 | fprintf (dump_file, ")\n"); | |
1532ec98 | 3546 | } |
3547 | ||
b44d1046 | 3548 | return true; |
3549 | } | |
2146e26d | 3550 | |
b44d1046 | 3551 | /* Return the direction for a given distance. |
3552 | FIXME: Computing dir this way is suboptimal, since dir can catch | |
3553 | cases that dist is unable to represent. */ | |
bc3c8ad4 | 3554 | |
b44d1046 | 3555 | static inline enum data_dependence_direction |
3556 | dir_from_dist (int dist) | |
3557 | { | |
3558 | if (dist > 0) | |
3559 | return dir_positive; | |
3560 | else if (dist < 0) | |
3561 | return dir_negative; | |
3562 | else | |
3563 | return dir_equal; | |
3564 | } | |
1532ec98 | 3565 | |
b44d1046 | 3566 | /* Compute the classic per loop direction vector. DDR is the data |
3567 | dependence relation to build a vector from. */ | |
1532ec98 | 3568 | |
b44d1046 | 3569 | static void |
3570 | build_classic_dir_vector (struct data_dependence_relation *ddr) | |
3571 | { | |
3572 | unsigned i, j; | |
3573 | lambda_vector dist_v; | |
bc3c8ad4 | 3574 | |
f1f41a6c | 3575 | FOR_EACH_VEC_ELT (DDR_DIST_VECTS (ddr), i, dist_v) |
b44d1046 | 3576 | { |
3577 | lambda_vector dir_v = lambda_vector_new (DDR_NB_LOOPS (ddr)); | |
bc3c8ad4 | 3578 | |
b44d1046 | 3579 | for (j = 0; j < DDR_NB_LOOPS (ddr); j++) |
3580 | dir_v[j] = dir_from_dist (dist_v[j]); | |
1532ec98 | 3581 | |
b44d1046 | 3582 | save_dir_v (ddr, dir_v); |
3583 | } | |
2146e26d | 3584 | } |
3585 | ||
b44d1046 | 3586 | /* Helper function. Returns true when there is a dependence between |
3587 | data references DRA and DRB. */ | |
6b421feb | 3588 | |
b44d1046 | 3589 | static bool |
3590 | subscript_dependence_tester_1 (struct data_dependence_relation *ddr, | |
3591 | struct data_reference *dra, | |
8c7b1d8e | 3592 | struct data_reference *drb, |
3593 | struct loop *loop_nest) | |
6b421feb | 3594 | { |
3595 | unsigned int i; | |
6b421feb | 3596 | tree last_conflicts; |
41c7a324 | 3597 | struct subscript *subscript; |
2ecb06bc | 3598 | tree res = NULL_TREE; |
b44d1046 | 3599 | |
f1f41a6c | 3600 | for (i = 0; DDR_SUBSCRIPTS (ddr).iterate (i, &subscript); i++) |
6b421feb | 3601 | { |
87da4f2e | 3602 | conflict_function *overlaps_a, *overlaps_b; |
41c7a324 | 3603 | |
48e1416a | 3604 | analyze_overlapping_iterations (DR_ACCESS_FN (dra, i), |
6b421feb | 3605 | DR_ACCESS_FN (drb, i), |
48e1416a | 3606 | &overlaps_a, &overlaps_b, |
8c7b1d8e | 3607 | &last_conflicts, loop_nest); |
41c7a324 | 3608 | |
2ecb06bc | 3609 | if (SUB_CONFLICTS_IN_A (subscript)) |
3610 | free_conflict_function (SUB_CONFLICTS_IN_A (subscript)); | |
3611 | if (SUB_CONFLICTS_IN_B (subscript)) | |
3612 | free_conflict_function (SUB_CONFLICTS_IN_B (subscript)); | |
3613 | ||
3614 | SUB_CONFLICTS_IN_A (subscript) = overlaps_a; | |
3615 | SUB_CONFLICTS_IN_B (subscript) = overlaps_b; | |
3616 | SUB_LAST_CONFLICT (subscript) = last_conflicts; | |
3617 | ||
3618 | /* If there is any undetermined conflict function we have to | |
3619 | give a conservative answer in case we cannot prove that | |
3620 | no dependence exists when analyzing another subscript. */ | |
87da4f2e | 3621 | if (CF_NOT_KNOWN_P (overlaps_a) |
3622 | || CF_NOT_KNOWN_P (overlaps_b)) | |
6b421feb | 3623 | { |
2ecb06bc | 3624 | res = chrec_dont_know; |
3625 | continue; | |
6b421feb | 3626 | } |
41c7a324 | 3627 | |
2ecb06bc | 3628 | /* When there is a subscript with no dependence we can stop. */ |
87da4f2e | 3629 | else if (CF_NO_DEPENDENCE_P (overlaps_a) |
3630 | || CF_NO_DEPENDENCE_P (overlaps_b)) | |
6b421feb | 3631 | { |
2ecb06bc | 3632 | res = chrec_known; |
3633 | break; | |
6b421feb | 3634 | } |
3635 | } | |
3636 | ||
2ecb06bc | 3637 | if (res == NULL_TREE) |
3638 | return true; | |
3639 | ||
3640 | if (res == chrec_known) | |
3641 | dependence_stats.num_dependence_independent++; | |
3642 | else | |
3643 | dependence_stats.num_dependence_undetermined++; | |
3644 | finalize_ddr_dependent (ddr, res); | |
3645 | return false; | |
b44d1046 | 3646 | } |
3647 | ||
8c7b1d8e | 3648 | /* Computes the conflicting iterations in LOOP_NEST, and initialize DDR. */ |
b44d1046 | 3649 | |
3650 | static void | |
8c7b1d8e | 3651 | subscript_dependence_tester (struct data_dependence_relation *ddr, |
3652 | struct loop *loop_nest) | |
b44d1046 | 3653 | { |
8c7b1d8e | 3654 | if (subscript_dependence_tester_1 (ddr, DDR_A (ddr), DDR_B (ddr), loop_nest)) |
b44d1046 | 3655 | dependence_stats.num_dependence_dependent++; |
6b421feb | 3656 | |
6b421feb | 3657 | compute_subscript_distance (ddr); |
8c7b1d8e | 3658 | if (build_classic_dist_vector (ddr, loop_nest)) |
b44d1046 | 3659 | build_classic_dir_vector (ddr); |
6b421feb | 3660 | } |
3661 | ||
2146e26d | 3662 | /* Returns true when all the access functions of A are affine or |
8c7b1d8e | 3663 | constant with respect to LOOP_NEST. */ |
2146e26d | 3664 | |
48e1416a | 3665 | static bool |
7ecb5bb2 | 3666 | access_functions_are_affine_or_constant_p (const struct data_reference *a, |
3667 | const struct loop *loop_nest) | |
2146e26d | 3668 | { |
3669 | unsigned int i; | |
f1f41a6c | 3670 | vec<tree> fns = DR_ACCESS_FNS (a); |
e38f1ca4 | 3671 | tree t; |
355572cc | 3672 | |
f1f41a6c | 3673 | FOR_EACH_VEC_ELT (fns, i, t) |
8c7b1d8e | 3674 | if (!evolution_function_is_invariant_p (t, loop_nest->num) |
3675 | && !evolution_function_is_affine_multivariate_p (t, loop_nest->num)) | |
2146e26d | 3676 | return false; |
48e1416a | 3677 | |
2146e26d | 3678 | return true; |
3679 | } | |
3680 | ||
8c7b1d8e | 3681 | /* This computes the affine dependence relation between A and B with |
3682 | respect to LOOP_NEST. CHREC_KNOWN is used for representing the | |
3683 | independence between two accesses, while CHREC_DONT_KNOW is used | |
3684 | for representing the unknown relation. | |
48e1416a | 3685 | |
2146e26d | 3686 | Note that it is possible to stop the computation of the dependence |
3687 | relation the first time we detect a CHREC_KNOWN element for a given | |
3688 | subscript. */ | |
3689 | ||
7b6f8db4 | 3690 | void |
8c7b1d8e | 3691 | compute_affine_dependence (struct data_dependence_relation *ddr, |
3692 | struct loop *loop_nest) | |
2146e26d | 3693 | { |
3694 | struct data_reference *dra = DDR_A (ddr); | |
3695 | struct data_reference *drb = DDR_B (ddr); | |
48e1416a | 3696 | |
2146e26d | 3697 | if (dump_file && (dump_flags & TDF_DETAILS)) |
3698 | { | |
6b6f234c | 3699 | fprintf (dump_file, "(compute_affine_dependence\n"); |
95539e1d | 3700 | fprintf (dump_file, " stmt_a: "); |
3701 | print_gimple_stmt (dump_file, DR_STMT (dra), 0, TDF_SLIM); | |
3702 | fprintf (dump_file, " stmt_b: "); | |
3703 | print_gimple_stmt (dump_file, DR_STMT (drb), 0, TDF_SLIM); | |
2146e26d | 3704 | } |
6b421feb | 3705 | |
2146e26d | 3706 | /* Analyze only when the dependence relation is not yet known. */ |
c76b51be | 3707 | if (DDR_ARE_DEPENDENT (ddr) == NULL_TREE) |
2146e26d | 3708 | { |
6b421feb | 3709 | dependence_stats.num_dependence_tests++; |
3710 | ||
8c7b1d8e | 3711 | if (access_functions_are_affine_or_constant_p (dra, loop_nest) |
3712 | && access_functions_are_affine_or_constant_p (drb, loop_nest)) | |
0d8001a7 | 3713 | subscript_dependence_tester (ddr, loop_nest); |
48e1416a | 3714 | |
2146e26d | 3715 | /* As a last case, if the dependence cannot be determined, or if |
3716 | the dependence is considered too difficult to determine, answer | |
3717 | "don't know". */ | |
3718 | else | |
6b421feb | 3719 | { |
3720 | dependence_stats.num_dependence_undetermined++; | |
3721 | ||
3722 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
3723 | { | |
3724 | fprintf (dump_file, "Data ref a:\n"); | |
3725 | dump_data_reference (dump_file, dra); | |
3726 | fprintf (dump_file, "Data ref b:\n"); | |
3727 | dump_data_reference (dump_file, drb); | |
3728 | fprintf (dump_file, "affine dependence test not usable: access function not affine or constant.\n"); | |
3729 | } | |
3730 | finalize_ddr_dependent (ddr, chrec_dont_know); | |
3731 | } | |
2146e26d | 3732 | } |
48e1416a | 3733 | |
2146e26d | 3734 | if (dump_file && (dump_flags & TDF_DETAILS)) |
95539e1d | 3735 | { |
3736 | if (DDR_ARE_DEPENDENT (ddr) == chrec_known) | |
3737 | fprintf (dump_file, ") -> no dependence\n"); | |
3738 | else if (DDR_ARE_DEPENDENT (ddr) == chrec_dont_know) | |
3739 | fprintf (dump_file, ") -> dependence analysis failed\n"); | |
3740 | else | |
3741 | fprintf (dump_file, ")\n"); | |
3742 | } | |
2146e26d | 3743 | } |
3744 | ||
b44d1046 | 3745 | /* Compute in DEPENDENCE_RELATIONS the data dependence graph for all |
3746 | the data references in DATAREFS, in the LOOP_NEST. When | |
41c7a324 | 3747 | COMPUTE_SELF_AND_RR is FALSE, don't compute read-read and self |
c7af8ae7 | 3748 | relations. Return true when successful, i.e. data references number |
3749 | is small enough to be handled. */ | |
2146e26d | 3750 | |
c7af8ae7 | 3751 | bool |
f1f41a6c | 3752 | compute_all_dependences (vec<data_reference_p> datarefs, |
3753 | vec<ddr_p> *dependence_relations, | |
3754 | vec<loop_p> loop_nest, | |
41c7a324 | 3755 | bool compute_self_and_rr) |
2146e26d | 3756 | { |
41c7a324 | 3757 | struct data_dependence_relation *ddr; |
3758 | struct data_reference *a, *b; | |
3759 | unsigned int i, j; | |
2146e26d | 3760 | |
f1f41a6c | 3761 | if ((int) datarefs.length () |
c7af8ae7 | 3762 | > PARAM_VALUE (PARAM_LOOP_MAX_DATAREFS_FOR_DATADEPS)) |
3763 | { | |
3764 | struct data_dependence_relation *ddr; | |
3765 | ||
3766 | /* Insert a single relation into dependence_relations: | |
3767 | chrec_dont_know. */ | |
3768 | ddr = initialize_data_dependence_relation (NULL, NULL, loop_nest); | |
f1f41a6c | 3769 | dependence_relations->safe_push (ddr); |
c7af8ae7 | 3770 | return false; |
3771 | } | |
3772 | ||
f1f41a6c | 3773 | FOR_EACH_VEC_ELT (datarefs, i, a) |
3774 | for (j = i + 1; datarefs.iterate (j, &b); j++) | |
9ff25603 | 3775 | if (DR_IS_WRITE (a) || DR_IS_WRITE (b) || compute_self_and_rr) |
41c7a324 | 3776 | { |
3777 | ddr = initialize_data_dependence_relation (a, b, loop_nest); | |
f1f41a6c | 3778 | dependence_relations->safe_push (ddr); |
3779 | if (loop_nest.exists ()) | |
3780 | compute_affine_dependence (ddr, loop_nest[0]); | |
41c7a324 | 3781 | } |
02d46b9c | 3782 | |
41c7a324 | 3783 | if (compute_self_and_rr) |
f1f41a6c | 3784 | FOR_EACH_VEC_ELT (datarefs, i, a) |
2146e26d | 3785 | { |
41c7a324 | 3786 | ddr = initialize_data_dependence_relation (a, a, loop_nest); |
f1f41a6c | 3787 | dependence_relations->safe_push (ddr); |
3788 | if (loop_nest.exists ()) | |
3789 | compute_affine_dependence (ddr, loop_nest[0]); | |
2146e26d | 3790 | } |
c7af8ae7 | 3791 | |
3792 | return true; | |
2146e26d | 3793 | } |
3794 | ||
d129ada0 | 3795 | /* Describes a location of a memory reference. */ |
3796 | ||
df8eb490 | 3797 | struct data_ref_loc |
d129ada0 | 3798 | { |
239dd34a | 3799 | /* The memory reference. */ |
3800 | tree ref; | |
d129ada0 | 3801 | |
07e3bcbf | 3802 | /* True if the memory reference is read. */ |
3803 | bool is_read; | |
df8eb490 | 3804 | }; |
d129ada0 | 3805 | |
d129ada0 | 3806 | |
faa56cf9 | 3807 | /* Stores the locations of memory references in STMT to REFERENCES. Returns |
3808 | true if STMT clobbers memory, false otherwise. */ | |
3809 | ||
d129ada0 | 3810 | static bool |
42acab1c | 3811 | get_references_in_stmt (gimple *stmt, vec<data_ref_loc, va_heap> *references) |
faa56cf9 | 3812 | { |
3813 | bool clobbers_memory = false; | |
e82e4eb5 | 3814 | data_ref_loc ref; |
239dd34a | 3815 | tree op0, op1; |
75a70cf9 | 3816 | enum gimple_code stmt_code = gimple_code (stmt); |
faa56cf9 | 3817 | |
faa56cf9 | 3818 | /* ASM_EXPR and CALL_EXPR may embed arbitrary side effects. |
0ef2a6d1 | 3819 | As we cannot model data-references to not spelled out |
3820 | accesses give up if they may occur. */ | |
3d483a94 | 3821 | if (stmt_code == GIMPLE_CALL |
3822 | && !(gimple_call_flags (stmt) & ECF_CONST)) | |
3823 | { | |
3824 | /* Allow IFN_GOMP_SIMD_LANE in their own loops. */ | |
c71d3c24 | 3825 | if (gimple_call_internal_p (stmt)) |
3826 | switch (gimple_call_internal_fn (stmt)) | |
3827 | { | |
3828 | case IFN_GOMP_SIMD_LANE: | |
3829 | { | |
3830 | struct loop *loop = gimple_bb (stmt)->loop_father; | |
3831 | tree uid = gimple_call_arg (stmt, 0); | |
3832 | gcc_assert (TREE_CODE (uid) == SSA_NAME); | |
3833 | if (loop == NULL | |
3834 | || loop->simduid != SSA_NAME_VAR (uid)) | |
3835 | clobbers_memory = true; | |
3836 | break; | |
3837 | } | |
3838 | case IFN_MASK_LOAD: | |
3839 | case IFN_MASK_STORE: | |
3840 | break; | |
3841 | default: | |
3d483a94 | 3842 | clobbers_memory = true; |
c71d3c24 | 3843 | break; |
3844 | } | |
3d483a94 | 3845 | else |
3846 | clobbers_memory = true; | |
3847 | } | |
3848 | else if (stmt_code == GIMPLE_ASM | |
1a91d914 | 3849 | && (gimple_asm_volatile_p (as_a <gasm *> (stmt)) |
3850 | || gimple_vuse (stmt))) | |
faa56cf9 | 3851 | clobbers_memory = true; |
3852 | ||
dd277d48 | 3853 | if (!gimple_vuse (stmt)) |
faa56cf9 | 3854 | return clobbers_memory; |
3855 | ||
75a70cf9 | 3856 | if (stmt_code == GIMPLE_ASSIGN) |
faa56cf9 | 3857 | { |
ffb04365 | 3858 | tree base; |
239dd34a | 3859 | op0 = gimple_assign_lhs (stmt); |
3860 | op1 = gimple_assign_rhs1 (stmt); | |
48e1416a | 3861 | |
239dd34a | 3862 | if (DECL_P (op1) |
3863 | || (REFERENCE_CLASS_P (op1) | |
3864 | && (base = get_base_address (op1)) | |
ffb04365 | 3865 | && TREE_CODE (base) != SSA_NAME)) |
faa56cf9 | 3866 | { |
239dd34a | 3867 | ref.ref = op1; |
e82e4eb5 | 3868 | ref.is_read = true; |
f1f41a6c | 3869 | references->safe_push (ref); |
faa56cf9 | 3870 | } |
faa56cf9 | 3871 | } |
75a70cf9 | 3872 | else if (stmt_code == GIMPLE_CALL) |
faa56cf9 | 3873 | { |
9f2e27ae | 3874 | unsigned i, n; |
d73c1238 | 3875 | tree ptr, type; |
3876 | unsigned int align; | |
d97e22fb | 3877 | |
c71d3c24 | 3878 | ref.is_read = false; |
3879 | if (gimple_call_internal_p (stmt)) | |
3880 | switch (gimple_call_internal_fn (stmt)) | |
3881 | { | |
3882 | case IFN_MASK_LOAD: | |
4e459157 | 3883 | if (gimple_call_lhs (stmt) == NULL_TREE) |
3884 | break; | |
c71d3c24 | 3885 | ref.is_read = true; |
3886 | case IFN_MASK_STORE: | |
d73c1238 | 3887 | ptr = build_int_cst (TREE_TYPE (gimple_call_arg (stmt, 1)), 0); |
3888 | align = tree_to_shwi (gimple_call_arg (stmt, 1)); | |
3889 | if (ref.is_read) | |
3890 | type = TREE_TYPE (gimple_call_lhs (stmt)); | |
3891 | else | |
3892 | type = TREE_TYPE (gimple_call_arg (stmt, 3)); | |
3893 | if (TYPE_ALIGN (type) != align) | |
3894 | type = build_aligned_type (type, align); | |
3895 | ref.ref = fold_build2 (MEM_REF, type, gimple_call_arg (stmt, 0), | |
3896 | ptr); | |
c71d3c24 | 3897 | references->safe_push (ref); |
3898 | return false; | |
3899 | default: | |
3900 | break; | |
3901 | } | |
3902 | ||
239dd34a | 3903 | op0 = gimple_call_lhs (stmt); |
9f2e27ae | 3904 | n = gimple_call_num_args (stmt); |
d97e22fb | 3905 | for (i = 0; i < n; i++) |
faa56cf9 | 3906 | { |
239dd34a | 3907 | op1 = gimple_call_arg (stmt, i); |
d97e22fb | 3908 | |
239dd34a | 3909 | if (DECL_P (op1) |
3910 | || (REFERENCE_CLASS_P (op1) && get_base_address (op1))) | |
faa56cf9 | 3911 | { |
239dd34a | 3912 | ref.ref = op1; |
e82e4eb5 | 3913 | ref.is_read = true; |
f1f41a6c | 3914 | references->safe_push (ref); |
faa56cf9 | 3915 | } |
3916 | } | |
3917 | } | |
9f2e27ae | 3918 | else |
3919 | return clobbers_memory; | |
faa56cf9 | 3920 | |
239dd34a | 3921 | if (op0 |
3922 | && (DECL_P (op0) | |
3923 | || (REFERENCE_CLASS_P (op0) && get_base_address (op0)))) | |
9f2e27ae | 3924 | { |
239dd34a | 3925 | ref.ref = op0; |
e82e4eb5 | 3926 | ref.is_read = false; |
f1f41a6c | 3927 | references->safe_push (ref); |
9f2e27ae | 3928 | } |
faa56cf9 | 3929 | return clobbers_memory; |
3930 | } | |
3931 | ||
fa4dba85 | 3932 | |
3933 | /* Returns true if the loop-nest has any data reference. */ | |
3934 | ||
3935 | bool | |
3936 | loop_nest_has_data_refs (loop_p loop) | |
3937 | { | |
3938 | basic_block *bbs = get_loop_body (loop); | |
3939 | vec<data_ref_loc> references; | |
3940 | references.create (3); | |
3941 | ||
3942 | for (unsigned i = 0; i < loop->num_nodes; i++) | |
3943 | { | |
3944 | basic_block bb = bbs[i]; | |
3945 | gimple_stmt_iterator bsi; | |
3946 | ||
3947 | for (bsi = gsi_start_bb (bb); !gsi_end_p (bsi); gsi_next (&bsi)) | |
3948 | { | |
42acab1c | 3949 | gimple *stmt = gsi_stmt (bsi); |
fa4dba85 | 3950 | get_references_in_stmt (stmt, &references); |
3951 | if (references.length ()) | |
3952 | { | |
3953 | free (bbs); | |
3954 | references.release (); | |
3955 | return true; | |
3956 | } | |
3957 | } | |
3958 | } | |
3959 | free (bbs); | |
3960 | references.release (); | |
3961 | ||
3962 | if (loop->inner) | |
3963 | { | |
3964 | loop = loop->inner; | |
3965 | while (loop) | |
3966 | { | |
3967 | if (loop_nest_has_data_refs (loop)) | |
3968 | return true; | |
3969 | loop = loop->next; | |
3970 | } | |
3971 | } | |
3972 | return false; | |
3973 | } | |
3974 | ||
faa56cf9 | 3975 | /* Stores the data references in STMT to DATAREFS. If there is an unanalyzable |
80bb306a | 3976 | reference, returns false, otherwise returns true. NEST is the outermost |
255b6be7 | 3977 | loop of the loop nest in which the references should be analyzed. */ |
faa56cf9 | 3978 | |
255b6be7 | 3979 | bool |
42acab1c | 3980 | find_data_references_in_stmt (struct loop *nest, gimple *stmt, |
f1f41a6c | 3981 | vec<data_reference_p> *datarefs) |
faa56cf9 | 3982 | { |
3983 | unsigned i; | |
4997014d | 3984 | auto_vec<data_ref_loc, 2> references; |
faa56cf9 | 3985 | data_ref_loc *ref; |
3986 | bool ret = true; | |
3987 | data_reference_p dr; | |
3988 | ||
3989 | if (get_references_in_stmt (stmt, &references)) | |
d70aebca | 3990 | return false; |
faa56cf9 | 3991 | |
f1f41a6c | 3992 | FOR_EACH_VEC_ELT (references, i, ref) |
faa56cf9 | 3993 | { |
221a697e | 3994 | dr = create_data_ref (nest, loop_containing_stmt (stmt), |
239dd34a | 3995 | ref->ref, stmt, ref->is_read); |
ad4a85ad | 3996 | gcc_assert (dr != NULL); |
f1f41a6c | 3997 | datarefs->safe_push (dr); |
faa56cf9 | 3998 | } |
f1f41a6c | 3999 | references.release (); |
faa56cf9 | 4000 | return ret; |
4001 | } | |
4002 | ||
221a697e | 4003 | /* Stores the data references in STMT to DATAREFS. If there is an |
4004 | unanalyzable reference, returns false, otherwise returns true. | |
4005 | NEST is the outermost loop of the loop nest in which the references | |
4006 | should be instantiated, LOOP is the loop in which the references | |
4007 | should be analyzed. */ | |
c794d738 | 4008 | |
4009 | bool | |
42acab1c | 4010 | graphite_find_data_references_in_stmt (loop_p nest, loop_p loop, gimple *stmt, |
f1f41a6c | 4011 | vec<data_reference_p> *datarefs) |
c794d738 | 4012 | { |
4013 | unsigned i; | |
4997014d | 4014 | auto_vec<data_ref_loc, 2> references; |
c794d738 | 4015 | data_ref_loc *ref; |
4016 | bool ret = true; | |
4017 | data_reference_p dr; | |
4018 | ||
4019 | if (get_references_in_stmt (stmt, &references)) | |
d70aebca | 4020 | return false; |
c794d738 | 4021 | |
f1f41a6c | 4022 | FOR_EACH_VEC_ELT (references, i, ref) |
c794d738 | 4023 | { |
239dd34a | 4024 | dr = create_data_ref (nest, loop, ref->ref, stmt, ref->is_read); |
c794d738 | 4025 | gcc_assert (dr != NULL); |
f1f41a6c | 4026 | datarefs->safe_push (dr); |
c794d738 | 4027 | } |
4028 | ||
f1f41a6c | 4029 | references.release (); |
c794d738 | 4030 | return ret; |
4031 | } | |
4032 | ||
37545e54 | 4033 | /* Search the data references in LOOP, and record the information into |
4034 | DATAREFS. Returns chrec_dont_know when failing to analyze a | |
4035 | difficult case, returns NULL_TREE otherwise. */ | |
4036 | ||
ec611e12 | 4037 | tree |
37545e54 | 4038 | find_data_references_in_bb (struct loop *loop, basic_block bb, |
f1f41a6c | 4039 | vec<data_reference_p> *datarefs) |
37545e54 | 4040 | { |
4041 | gimple_stmt_iterator bsi; | |
4042 | ||
4043 | for (bsi = gsi_start_bb (bb); !gsi_end_p (bsi); gsi_next (&bsi)) | |
4044 | { | |
42acab1c | 4045 | gimple *stmt = gsi_stmt (bsi); |
37545e54 | 4046 | |
4047 | if (!find_data_references_in_stmt (loop, stmt, datarefs)) | |
4048 | { | |
4049 | struct data_reference *res; | |
4050 | res = XCNEW (struct data_reference); | |
f1f41a6c | 4051 | datarefs->safe_push (res); |
37545e54 | 4052 | |
4053 | return chrec_dont_know; | |
4054 | } | |
4055 | } | |
4056 | ||
4057 | return NULL_TREE; | |
4058 | } | |
4059 | ||
2146e26d | 4060 | /* Search the data references in LOOP, and record the information into |
4061 | DATAREFS. Returns chrec_dont_know when failing to analyze a | |
4062 | difficult case, returns NULL_TREE otherwise. | |
80bb306a | 4063 | |
557ef5d8 | 4064 | TODO: This function should be made smarter so that it can handle address |
4065 | arithmetic as if they were array accesses, etc. */ | |
2146e26d | 4066 | |
07e3bcbf | 4067 | tree |
41c7a324 | 4068 | find_data_references_in_loop (struct loop *loop, |
f1f41a6c | 4069 | vec<data_reference_p> *datarefs) |
2146e26d | 4070 | { |
0332ea54 | 4071 | basic_block bb, *bbs; |
4072 | unsigned int i; | |
bc3c8ad4 | 4073 | |
ad4a85ad | 4074 | bbs = get_loop_body_in_dom_order (loop); |
0332ea54 | 4075 | |
4076 | for (i = 0; i < loop->num_nodes; i++) | |
2146e26d | 4077 | { |
0332ea54 | 4078 | bb = bbs[i]; |
4079 | ||
37545e54 | 4080 | if (find_data_references_in_bb (loop, bb, datarefs) == chrec_dont_know) |
4081 | { | |
4082 | free (bbs); | |
4083 | return chrec_dont_know; | |
4084 | } | |
2146e26d | 4085 | } |
0332ea54 | 4086 | free (bbs); |
4087 | ||
4a012ac6 | 4088 | return NULL_TREE; |
2146e26d | 4089 | } |
4090 | ||
b44d1046 | 4091 | /* Recursive helper function. */ |
4092 | ||
4093 | static bool | |
f1f41a6c | 4094 | find_loop_nest_1 (struct loop *loop, vec<loop_p> *loop_nest) |
b44d1046 | 4095 | { |
4096 | /* Inner loops of the nest should not contain siblings. Example: | |
4097 | when there are two consecutive loops, | |
4098 | ||
4099 | | loop_0 | |
4100 | | loop_1 | |
4101 | | A[{0, +, 1}_1] | |
4102 | | endloop_1 | |
4103 | | loop_2 | |
4104 | | A[{0, +, 1}_2] | |
4105 | | endloop_2 | |
4106 | | endloop_0 | |
4107 | ||
4108 | the dependence relation cannot be captured by the distance | |
4109 | abstraction. */ | |
4110 | if (loop->next) | |
4111 | return false; | |
2146e26d | 4112 | |
f1f41a6c | 4113 | loop_nest->safe_push (loop); |
b44d1046 | 4114 | if (loop->inner) |
4115 | return find_loop_nest_1 (loop->inner, loop_nest); | |
4116 | return true; | |
4117 | } | |
4118 | ||
4119 | /* Return false when the LOOP is not well nested. Otherwise return | |
4120 | true and insert in LOOP_NEST the loops of the nest. LOOP_NEST will | |
4121 | contain the loops from the outermost to the innermost, as they will | |
4122 | appear in the classic distance vector. */ | |
4123 | ||
5c205353 | 4124 | bool |
f1f41a6c | 4125 | find_loop_nest (struct loop *loop, vec<loop_p> *loop_nest) |
b44d1046 | 4126 | { |
f1f41a6c | 4127 | loop_nest->safe_push (loop); |
b44d1046 | 4128 | if (loop->inner) |
4129 | return find_loop_nest_1 (loop->inner, loop_nest); | |
4130 | return true; | |
4131 | } | |
2146e26d | 4132 | |
b79b3386 | 4133 | /* Returns true when the data dependences have been computed, false otherwise. |
4134 | Given a loop nest LOOP, the following vectors are returned: | |
48e1416a | 4135 | DATAREFS is initialized to all the array elements contained in this loop, |
4136 | DEPENDENCE_RELATIONS contains the relations between the data references. | |
4137 | Compute read-read and self relations if | |
516849c7 | 4138 | COMPUTE_SELF_AND_READ_READ_DEPENDENCES is TRUE. */ |
2146e26d | 4139 | |
b79b3386 | 4140 | bool |
48e1416a | 4141 | compute_data_dependences_for_loop (struct loop *loop, |
516849c7 | 4142 | bool compute_self_and_read_read_dependences, |
f1f41a6c | 4143 | vec<loop_p> *loop_nest, |
4144 | vec<data_reference_p> *datarefs, | |
4145 | vec<ddr_p> *dependence_relations) | |
2146e26d | 4146 | { |
b79b3386 | 4147 | bool res = true; |
516849c7 | 4148 | |
6b421feb | 4149 | memset (&dependence_stats, 0, sizeof (dependence_stats)); |
2146e26d | 4150 | |
48e1416a | 4151 | /* If the loop nest is not well formed, or one of the data references |
b44d1046 | 4152 | is not computable, give up without spending time to compute other |
4153 | dependences. */ | |
80bb306a | 4154 | if (!loop |
a8af2e86 | 4155 | || !find_loop_nest (loop, loop_nest) |
c7af8ae7 | 4156 | || find_data_references_in_loop (loop, datarefs) == chrec_dont_know |
4157 | || !compute_all_dependences (*datarefs, dependence_relations, *loop_nest, | |
4158 | compute_self_and_read_read_dependences)) | |
4159 | res = false; | |
6b421feb | 4160 | |
4161 | if (dump_file && (dump_flags & TDF_STATS)) | |
2146e26d | 4162 | { |
6b421feb | 4163 | fprintf (dump_file, "Dependence tester statistics:\n"); |
4164 | ||
48e1416a | 4165 | fprintf (dump_file, "Number of dependence tests: %d\n", |
6b421feb | 4166 | dependence_stats.num_dependence_tests); |
48e1416a | 4167 | fprintf (dump_file, "Number of dependence tests classified dependent: %d\n", |
6b421feb | 4168 | dependence_stats.num_dependence_dependent); |
48e1416a | 4169 | fprintf (dump_file, "Number of dependence tests classified independent: %d\n", |
6b421feb | 4170 | dependence_stats.num_dependence_independent); |
48e1416a | 4171 | fprintf (dump_file, "Number of undetermined dependence tests: %d\n", |
6b421feb | 4172 | dependence_stats.num_dependence_undetermined); |
4173 | ||
48e1416a | 4174 | fprintf (dump_file, "Number of subscript tests: %d\n", |
6b421feb | 4175 | dependence_stats.num_subscript_tests); |
48e1416a | 4176 | fprintf (dump_file, "Number of undetermined subscript tests: %d\n", |
6b421feb | 4177 | dependence_stats.num_subscript_undetermined); |
48e1416a | 4178 | fprintf (dump_file, "Number of same subscript function: %d\n", |
6b421feb | 4179 | dependence_stats.num_same_subscript_function); |
4180 | ||
4181 | fprintf (dump_file, "Number of ziv tests: %d\n", | |
4182 | dependence_stats.num_ziv); | |
4183 | fprintf (dump_file, "Number of ziv tests returning dependent: %d\n", | |
4184 | dependence_stats.num_ziv_dependent); | |
4185 | fprintf (dump_file, "Number of ziv tests returning independent: %d\n", | |
4186 | dependence_stats.num_ziv_independent); | |
4187 | fprintf (dump_file, "Number of ziv tests unimplemented: %d\n", | |
48e1416a | 4188 | dependence_stats.num_ziv_unimplemented); |
6b421feb | 4189 | |
48e1416a | 4190 | fprintf (dump_file, "Number of siv tests: %d\n", |
6b421feb | 4191 | dependence_stats.num_siv); |
4192 | fprintf (dump_file, "Number of siv tests returning dependent: %d\n", | |
4193 | dependence_stats.num_siv_dependent); | |
4194 | fprintf (dump_file, "Number of siv tests returning independent: %d\n", | |
4195 | dependence_stats.num_siv_independent); | |
4196 | fprintf (dump_file, "Number of siv tests unimplemented: %d\n", | |
4197 | dependence_stats.num_siv_unimplemented); | |
4198 | ||
48e1416a | 4199 | fprintf (dump_file, "Number of miv tests: %d\n", |
6b421feb | 4200 | dependence_stats.num_miv); |
4201 | fprintf (dump_file, "Number of miv tests returning dependent: %d\n", | |
4202 | dependence_stats.num_miv_dependent); | |
4203 | fprintf (dump_file, "Number of miv tests returning independent: %d\n", | |
4204 | dependence_stats.num_miv_independent); | |
4205 | fprintf (dump_file, "Number of miv tests unimplemented: %d\n", | |
4206 | dependence_stats.num_miv_unimplemented); | |
b79b3386 | 4207 | } |
4208 | ||
4209 | return res; | |
2146e26d | 4210 | } |
4211 | ||
6b6f234c | 4212 | /* Free the memory used by a data dependence relation DDR. */ |
4213 | ||
4214 | void | |
4215 | free_dependence_relation (struct data_dependence_relation *ddr) | |
4216 | { | |
4217 | if (ddr == NULL) | |
4218 | return; | |
4219 | ||
f1f41a6c | 4220 | if (DDR_SUBSCRIPTS (ddr).exists ()) |
87da4f2e | 4221 | free_subscripts (DDR_SUBSCRIPTS (ddr)); |
f1f41a6c | 4222 | DDR_DIST_VECTS (ddr).release (); |
4223 | DDR_DIR_VECTS (ddr).release (); | |
41c7a324 | 4224 | |
6b6f234c | 4225 | free (ddr); |
4226 | } | |
4227 | ||
4228 | /* Free the memory used by the data dependence relations from | |
4229 | DEPENDENCE_RELATIONS. */ | |
4230 | ||
48e1416a | 4231 | void |
f1f41a6c | 4232 | free_dependence_relations (vec<ddr_p> dependence_relations) |
6b6f234c | 4233 | { |
4234 | unsigned int i; | |
41c7a324 | 4235 | struct data_dependence_relation *ddr; |
6b6f234c | 4236 | |
f1f41a6c | 4237 | FOR_EACH_VEC_ELT (dependence_relations, i, ddr) |
a8af2e86 | 4238 | if (ddr) |
62761183 | 4239 | free_dependence_relation (ddr); |
41c7a324 | 4240 | |
f1f41a6c | 4241 | dependence_relations.release (); |
2146e26d | 4242 | } |
4243 | ||
6b6f234c | 4244 | /* Free the memory used by the data references from DATAREFS. */ |
4245 | ||
4246 | void | |
f1f41a6c | 4247 | free_data_refs (vec<data_reference_p> datarefs) |
6b6f234c | 4248 | { |
4249 | unsigned int i; | |
41c7a324 | 4250 | struct data_reference *dr; |
2146e26d | 4251 | |
f1f41a6c | 4252 | FOR_EACH_VEC_ELT (datarefs, i, dr) |
31e389a2 | 4253 | free_data_ref (dr); |
f1f41a6c | 4254 | datarefs.release (); |
6b6f234c | 4255 | } |