]>
Commit | Line | Data |
---|---|---|
801c5610 | 1 | /* Loop distribution. |
fbd26352 | 2 | Copyright (C) 2006-2019 Free Software Foundation, Inc. |
801c5610 | 3 | Contributed by Georges-Andre Silber <Georges-Andre.Silber@ensmp.fr> |
4 | and Sebastian Pop <sebastian.pop@amd.com>. | |
5 | ||
6 | This file is part of GCC. | |
48e1416a | 7 | |
801c5610 | 8 | GCC is free software; you can redistribute it and/or modify it |
9 | under the terms of the GNU General Public License as published by the | |
10 | Free Software Foundation; either version 3, or (at your option) any | |
11 | later version. | |
48e1416a | 12 | |
801c5610 | 13 | GCC is distributed in the hope that it will be useful, but WITHOUT |
14 | ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or | |
15 | FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License | |
16 | for more details. | |
48e1416a | 17 | |
801c5610 | 18 | You should have received a copy of the GNU General Public License |
19 | along with GCC; see the file COPYING3. If not see | |
20 | <http://www.gnu.org/licenses/>. */ | |
21 | ||
22 | /* This pass performs loop distribution: for example, the loop | |
23 | ||
24 | |DO I = 2, N | |
25 | | A(I) = B(I) + C | |
26 | | D(I) = A(I-1)*E | |
27 | |ENDDO | |
28 | ||
48e1416a | 29 | is transformed to |
801c5610 | 30 | |
31 | |DOALL I = 2, N | |
32 | | A(I) = B(I) + C | |
33 | |ENDDO | |
34 | | | |
35 | |DOALL I = 2, N | |
36 | | D(I) = A(I-1)*E | |
37 | |ENDDO | |
38 | ||
f562e2ea | 39 | Loop distribution is the dual of loop fusion. It separates statements |
40 | of a loop (or loop nest) into multiple loops (or loop nests) with the | |
41 | same loop header. The major goal is to separate statements which may | |
42 | be vectorized from those that can't. This pass implements distribution | |
43 | in the following steps: | |
44 | ||
45 | 1) Seed partitions with specific type statements. For now we support | |
46 | two types seed statements: statement defining variable used outside | |
47 | of loop; statement storing to memory. | |
48 | 2) Build reduced dependence graph (RDG) for loop to be distributed. | |
49 | The vertices (RDG:V) model all statements in the loop and the edges | |
50 | (RDG:E) model flow and control dependencies between statements. | |
51 | 3) Apart from RDG, compute data dependencies between memory references. | |
52 | 4) Starting from seed statement, build up partition by adding depended | |
53 | statements according to RDG's dependence information. Partition is | |
54 | classified as parallel type if it can be executed paralleled; or as | |
55 | sequential type if it can't. Parallel type partition is further | |
56 | classified as different builtin kinds if it can be implemented as | |
57 | builtin function calls. | |
58 | 5) Build partition dependence graph (PG) based on data dependencies. | |
59 | The vertices (PG:V) model all partitions and the edges (PG:E) model | |
60 | all data dependencies between every partitions pair. In general, | |
61 | data dependence is either compilation time known or unknown. In C | |
62 | family languages, there exists quite amount compilation time unknown | |
63 | dependencies because of possible alias relation of data references. | |
64 | We categorize PG's edge to two types: "true" edge that represents | |
65 | compilation time known data dependencies; "alias" edge for all other | |
66 | data dependencies. | |
67 | 6) Traverse subgraph of PG as if all "alias" edges don't exist. Merge | |
68 | partitions in each strong connected component (SCC) correspondingly. | |
69 | Build new PG for merged partitions. | |
70 | 7) Traverse PG again and this time with both "true" and "alias" edges | |
71 | included. We try to break SCCs by removing some edges. Because | |
72 | SCCs by "true" edges are all fused in step 6), we can break SCCs | |
73 | by removing some "alias" edges. It's NP-hard to choose optimal | |
74 | edge set, fortunately simple approximation is good enough for us | |
75 | given the small problem scale. | |
76 | 8) Collect all data dependencies of the removed "alias" edges. Create | |
77 | runtime alias checks for collected data dependencies. | |
78 | 9) Version loop under the condition of runtime alias checks. Given | |
79 | loop distribution generally introduces additional overhead, it is | |
80 | only useful if vectorization is achieved in distributed loop. We | |
81 | version loop with internal function call IFN_LOOP_DIST_ALIAS. If | |
82 | no distributed loop can be vectorized, we simply remove distributed | |
83 | loops and recover to the original one. | |
84 | ||
85 | TODO: | |
883b4905 | 86 | 1) We only distribute innermost two-level loop nest now. We should |
87 | extend it for arbitrary loop nests in the future. | |
f562e2ea | 88 | 2) We only fuse partitions in SCC now. A better fusion algorithm is |
89 | desired to minimize loop overhead, maximize parallelism and maximize | |
90 | data reuse. */ | |
801c5610 | 91 | |
92 | #include "config.h" | |
93 | #include "system.h" | |
94 | #include "coretypes.h" | |
9ef16211 | 95 | #include "backend.h" |
b20a8bb4 | 96 | #include "tree.h" |
9ef16211 | 97 | #include "gimple.h" |
7c29e30e | 98 | #include "cfghooks.h" |
99 | #include "tree-pass.h" | |
9ef16211 | 100 | #include "ssa.h" |
7c29e30e | 101 | #include "gimple-pretty-print.h" |
9ef16211 | 102 | #include "fold-const.h" |
94ea8568 | 103 | #include "cfganal.h" |
dcf1a1ec | 104 | #include "gimple-iterator.h" |
e795d6e1 | 105 | #include "gimplify-me.h" |
9ed99284 | 106 | #include "stor-layout.h" |
073c1fd5 | 107 | #include "tree-cfg.h" |
05d9c18a | 108 | #include "tree-ssa-loop-manip.h" |
05ebeee6 | 109 | #include "tree-ssa-loop-ivopts.h" |
073c1fd5 | 110 | #include "tree-ssa-loop.h" |
111 | #include "tree-into-ssa.h" | |
69ee5dbb | 112 | #include "tree-ssa.h" |
801c5610 | 113 | #include "cfgloop.h" |
801c5610 | 114 | #include "tree-scalar-evolution.h" |
f3754041 | 115 | #include "params.h" |
df9892ff | 116 | #include "tree-vectorizer.h" |
bc6e884f | 117 | #include "tree-eh.h" |
a2740310 | 118 | |
119 | ||
f3754041 | 120 | #define MAX_DATAREFS_NUM \ |
121 | ((unsigned) PARAM_VALUE (PARAM_LOOP_MAX_DATAREFS_FOR_DATADEPS)) | |
122 | ||
883b4905 | 123 | /* Threshold controlling number of distributed partitions. Given it may |
124 | be unnecessary if a memory stream cost model is invented in the future, | |
125 | we define it as a temporary macro, rather than a parameter. */ | |
126 | #define NUM_PARTITION_THRESHOLD (4) | |
127 | ||
50f5937e | 128 | /* Hashtable helpers. */ |
129 | ||
130 | struct ddr_hasher : nofree_ptr_hash <struct data_dependence_relation> | |
131 | { | |
132 | static inline hashval_t hash (const data_dependence_relation *); | |
133 | static inline bool equal (const data_dependence_relation *, | |
134 | const data_dependence_relation *); | |
135 | }; | |
136 | ||
137 | /* Hash function for data dependence. */ | |
138 | ||
139 | inline hashval_t | |
140 | ddr_hasher::hash (const data_dependence_relation *ddr) | |
141 | { | |
142 | inchash::hash h; | |
143 | h.add_ptr (DDR_A (ddr)); | |
144 | h.add_ptr (DDR_B (ddr)); | |
145 | return h.end (); | |
146 | } | |
147 | ||
148 | /* Hash table equality function for data dependence. */ | |
149 | ||
150 | inline bool | |
151 | ddr_hasher::equal (const data_dependence_relation *ddr1, | |
152 | const data_dependence_relation *ddr2) | |
153 | { | |
154 | return (DDR_A (ddr1) == DDR_A (ddr2) && DDR_B (ddr1) == DDR_B (ddr2)); | |
155 | } | |
156 | ||
209a62a6 | 157 | /* The loop (nest) to be distributed. */ |
158 | static vec<loop_p> loop_nest; | |
159 | ||
f3754041 | 160 | /* Vector of data references in the loop to be distributed. */ |
161 | static vec<data_reference_p> datarefs_vec; | |
162 | ||
6079e9be | 163 | /* If there is nonaddressable data reference in above vector. */ |
164 | static bool has_nonaddressable_dataref_p; | |
165 | ||
f3754041 | 166 | /* Store index of data reference in aux field. */ |
167 | #define DR_INDEX(dr) ((uintptr_t) (dr)->aux) | |
168 | ||
50f5937e | 169 | /* Hash table for data dependence relation in the loop to be distributed. */ |
bbb229ef | 170 | static hash_table<ddr_hasher> *ddrs_table; |
50f5937e | 171 | |
a2740310 | 172 | /* A Reduced Dependence Graph (RDG) vertex representing a statement. */ |
04009ada | 173 | struct rdg_vertex |
a2740310 | 174 | { |
175 | /* The statement represented by this vertex. */ | |
42acab1c | 176 | gimple *stmt; |
a2740310 | 177 | |
178 | /* Vector of data-references in this statement. */ | |
179 | vec<data_reference_p> datarefs; | |
180 | ||
181 | /* True when the statement contains a write to memory. */ | |
182 | bool has_mem_write; | |
183 | ||
184 | /* True when the statement contains a read from memory. */ | |
185 | bool has_mem_reads; | |
04009ada | 186 | }; |
a2740310 | 187 | |
188 | #define RDGV_STMT(V) ((struct rdg_vertex *) ((V)->data))->stmt | |
189 | #define RDGV_DATAREFS(V) ((struct rdg_vertex *) ((V)->data))->datarefs | |
190 | #define RDGV_HAS_MEM_WRITE(V) ((struct rdg_vertex *) ((V)->data))->has_mem_write | |
191 | #define RDGV_HAS_MEM_READS(V) ((struct rdg_vertex *) ((V)->data))->has_mem_reads | |
192 | #define RDG_STMT(RDG, I) RDGV_STMT (&(RDG->vertices[I])) | |
193 | #define RDG_DATAREFS(RDG, I) RDGV_DATAREFS (&(RDG->vertices[I])) | |
194 | #define RDG_MEM_WRITE_STMT(RDG, I) RDGV_HAS_MEM_WRITE (&(RDG->vertices[I])) | |
195 | #define RDG_MEM_READS_STMT(RDG, I) RDGV_HAS_MEM_READS (&(RDG->vertices[I])) | |
196 | ||
197 | /* Data dependence type. */ | |
198 | ||
199 | enum rdg_dep_type | |
200 | { | |
201 | /* Read After Write (RAW). */ | |
202 | flow_dd = 'f', | |
203 | ||
f1ce84d9 | 204 | /* Control dependence (execute conditional on). */ |
205 | control_dd = 'c' | |
a2740310 | 206 | }; |
207 | ||
208 | /* Dependence information attached to an edge of the RDG. */ | |
209 | ||
04009ada | 210 | struct rdg_edge |
a2740310 | 211 | { |
212 | /* Type of the dependence. */ | |
213 | enum rdg_dep_type type; | |
04009ada | 214 | }; |
a2740310 | 215 | |
216 | #define RDGE_TYPE(E) ((struct rdg_edge *) ((E)->data))->type | |
a2740310 | 217 | |
a2740310 | 218 | /* Dump vertex I in RDG to FILE. */ |
219 | ||
220 | static void | |
221 | dump_rdg_vertex (FILE *file, struct graph *rdg, int i) | |
222 | { | |
223 | struct vertex *v = &(rdg->vertices[i]); | |
224 | struct graph_edge *e; | |
225 | ||
226 | fprintf (file, "(vertex %d: (%s%s) (in:", i, | |
227 | RDG_MEM_WRITE_STMT (rdg, i) ? "w" : "", | |
228 | RDG_MEM_READS_STMT (rdg, i) ? "r" : ""); | |
229 | ||
230 | if (v->pred) | |
231 | for (e = v->pred; e; e = e->pred_next) | |
232 | fprintf (file, " %d", e->src); | |
233 | ||
234 | fprintf (file, ") (out:"); | |
235 | ||
236 | if (v->succ) | |
237 | for (e = v->succ; e; e = e->succ_next) | |
238 | fprintf (file, " %d", e->dest); | |
239 | ||
240 | fprintf (file, ")\n"); | |
241 | print_gimple_stmt (file, RDGV_STMT (v), 0, TDF_VOPS|TDF_MEMSYMS); | |
242 | fprintf (file, ")\n"); | |
243 | } | |
244 | ||
245 | /* Call dump_rdg_vertex on stderr. */ | |
246 | ||
247 | DEBUG_FUNCTION void | |
248 | debug_rdg_vertex (struct graph *rdg, int i) | |
249 | { | |
250 | dump_rdg_vertex (stderr, rdg, i); | |
251 | } | |
252 | ||
a2740310 | 253 | /* Dump the reduced dependence graph RDG to FILE. */ |
254 | ||
255 | static void | |
256 | dump_rdg (FILE *file, struct graph *rdg) | |
257 | { | |
a2740310 | 258 | fprintf (file, "(rdg\n"); |
15c8650d | 259 | for (int i = 0; i < rdg->n_vertices; i++) |
260 | dump_rdg_vertex (file, rdg, i); | |
a2740310 | 261 | fprintf (file, ")\n"); |
a2740310 | 262 | } |
263 | ||
264 | /* Call dump_rdg on stderr. */ | |
265 | ||
266 | DEBUG_FUNCTION void | |
267 | debug_rdg (struct graph *rdg) | |
268 | { | |
269 | dump_rdg (stderr, rdg); | |
270 | } | |
271 | ||
272 | static void | |
273 | dot_rdg_1 (FILE *file, struct graph *rdg) | |
274 | { | |
275 | int i; | |
5529df60 | 276 | pretty_printer buffer; |
277 | pp_needs_newline (&buffer) = false; | |
278 | buffer.buffer->stream = file; | |
a2740310 | 279 | |
280 | fprintf (file, "digraph RDG {\n"); | |
281 | ||
282 | for (i = 0; i < rdg->n_vertices; i++) | |
283 | { | |
284 | struct vertex *v = &(rdg->vertices[i]); | |
285 | struct graph_edge *e; | |
286 | ||
5529df60 | 287 | fprintf (file, "%d [label=\"[%d] ", i, i); |
288 | pp_gimple_stmt_1 (&buffer, RDGV_STMT (v), 0, TDF_SLIM); | |
289 | pp_flush (&buffer); | |
290 | fprintf (file, "\"]\n"); | |
291 | ||
a2740310 | 292 | /* Highlight reads from memory. */ |
293 | if (RDG_MEM_READS_STMT (rdg, i)) | |
294 | fprintf (file, "%d [style=filled, fillcolor=green]\n", i); | |
295 | ||
296 | /* Highlight stores to memory. */ | |
297 | if (RDG_MEM_WRITE_STMT (rdg, i)) | |
298 | fprintf (file, "%d [style=filled, fillcolor=red]\n", i); | |
299 | ||
300 | if (v->succ) | |
301 | for (e = v->succ; e; e = e->succ_next) | |
302 | switch (RDGE_TYPE (e)) | |
303 | { | |
a2740310 | 304 | case flow_dd: |
305 | /* These are the most common dependences: don't print these. */ | |
306 | fprintf (file, "%d -> %d \n", i, e->dest); | |
307 | break; | |
308 | ||
f1ce84d9 | 309 | case control_dd: |
310 | fprintf (file, "%d -> %d [label=control] \n", i, e->dest); | |
311 | break; | |
312 | ||
a2740310 | 313 | default: |
314 | gcc_unreachable (); | |
315 | } | |
316 | } | |
317 | ||
318 | fprintf (file, "}\n\n"); | |
319 | } | |
320 | ||
321 | /* Display the Reduced Dependence Graph using dotty. */ | |
322 | ||
323 | DEBUG_FUNCTION void | |
324 | dot_rdg (struct graph *rdg) | |
325 | { | |
5529df60 | 326 | /* When debugging, you may want to enable the following code. */ |
6cc8a944 | 327 | #ifdef HAVE_POPEN |
9af5ce0c | 328 | FILE *file = popen ("dot -Tx11", "w"); |
5529df60 | 329 | if (!file) |
330 | return; | |
a2740310 | 331 | dot_rdg_1 (file, rdg); |
5529df60 | 332 | fflush (file); |
333 | close (fileno (file)); | |
334 | pclose (file); | |
a2740310 | 335 | #else |
336 | dot_rdg_1 (stderr, rdg); | |
337 | #endif | |
338 | } | |
339 | ||
340 | /* Returns the index of STMT in RDG. */ | |
341 | ||
342 | static int | |
42acab1c | 343 | rdg_vertex_for_stmt (struct graph *rdg ATTRIBUTE_UNUSED, gimple *stmt) |
a2740310 | 344 | { |
345 | int index = gimple_uid (stmt); | |
346 | gcc_checking_assert (index == -1 || RDG_STMT (rdg, index) == stmt); | |
347 | return index; | |
348 | } | |
349 | ||
a2740310 | 350 | /* Creates dependence edges in RDG for all the uses of DEF. IDEF is |
351 | the index of DEF in RDG. */ | |
352 | ||
353 | static void | |
354 | create_rdg_edges_for_scalar (struct graph *rdg, tree def, int idef) | |
355 | { | |
356 | use_operand_p imm_use_p; | |
357 | imm_use_iterator iterator; | |
358 | ||
359 | FOR_EACH_IMM_USE_FAST (imm_use_p, iterator, def) | |
360 | { | |
361 | struct graph_edge *e; | |
362 | int use = rdg_vertex_for_stmt (rdg, USE_STMT (imm_use_p)); | |
363 | ||
364 | if (use < 0) | |
365 | continue; | |
366 | ||
367 | e = add_edge (rdg, idef, use); | |
368 | e->data = XNEW (struct rdg_edge); | |
369 | RDGE_TYPE (e) = flow_dd; | |
a2740310 | 370 | } |
371 | } | |
372 | ||
f1ce84d9 | 373 | /* Creates an edge for the control dependences of BB to the vertex V. */ |
374 | ||
375 | static void | |
376 | create_edge_for_control_dependence (struct graph *rdg, basic_block bb, | |
377 | int v, control_dependences *cd) | |
378 | { | |
379 | bitmap_iterator bi; | |
380 | unsigned edge_n; | |
381 | EXECUTE_IF_SET_IN_BITMAP (cd->get_edges_dependent_on (bb->index), | |
382 | 0, edge_n, bi) | |
383 | { | |
ce143ff0 | 384 | basic_block cond_bb = cd->get_edge_src (edge_n); |
42acab1c | 385 | gimple *stmt = last_stmt (cond_bb); |
f1ce84d9 | 386 | if (stmt && is_ctrl_stmt (stmt)) |
387 | { | |
388 | struct graph_edge *e; | |
389 | int c = rdg_vertex_for_stmt (rdg, stmt); | |
390 | if (c < 0) | |
391 | continue; | |
392 | ||
393 | e = add_edge (rdg, c, v); | |
394 | e->data = XNEW (struct rdg_edge); | |
395 | RDGE_TYPE (e) = control_dd; | |
f1ce84d9 | 396 | } |
397 | } | |
398 | } | |
399 | ||
a2740310 | 400 | /* Creates the edges of the reduced dependence graph RDG. */ |
401 | ||
402 | static void | |
7103facc | 403 | create_rdg_flow_edges (struct graph *rdg) |
a2740310 | 404 | { |
405 | int i; | |
a2740310 | 406 | def_operand_p def_p; |
407 | ssa_op_iter iter; | |
408 | ||
a2740310 | 409 | for (i = 0; i < rdg->n_vertices; i++) |
410 | FOR_EACH_PHI_OR_STMT_DEF (def_p, RDG_STMT (rdg, i), | |
411 | iter, SSA_OP_DEF) | |
412 | create_rdg_edges_for_scalar (rdg, DEF_FROM_PTR (def_p), i); | |
7103facc | 413 | } |
f1ce84d9 | 414 | |
7103facc | 415 | /* Creates the edges of the reduced dependence graph RDG. */ |
416 | ||
417 | static void | |
77d095c0 | 418 | create_rdg_cd_edges (struct graph *rdg, control_dependences *cd, loop_p loop) |
7103facc | 419 | { |
420 | int i; | |
421 | ||
422 | for (i = 0; i < rdg->n_vertices; i++) | |
423 | { | |
42acab1c | 424 | gimple *stmt = RDG_STMT (rdg, i); |
7103facc | 425 | if (gimple_code (stmt) == GIMPLE_PHI) |
426 | { | |
427 | edge_iterator ei; | |
428 | edge e; | |
429 | FOR_EACH_EDGE (e, ei, gimple_bb (stmt)->preds) | |
77d095c0 | 430 | if (flow_bb_inside_loop_p (loop, e->src)) |
f1ce84d9 | 431 | create_edge_for_control_dependence (rdg, e->src, i, cd); |
7103facc | 432 | } |
433 | else | |
434 | create_edge_for_control_dependence (rdg, gimple_bb (stmt), i, cd); | |
435 | } | |
a2740310 | 436 | } |
437 | ||
438 | /* Build the vertices of the reduced dependence graph RDG. Return false | |
439 | if that failed. */ | |
440 | ||
441 | static bool | |
f3754041 | 442 | create_rdg_vertices (struct graph *rdg, vec<gimple *> stmts, loop_p loop) |
a2740310 | 443 | { |
444 | int i; | |
42acab1c | 445 | gimple *stmt; |
a2740310 | 446 | |
447 | FOR_EACH_VEC_ELT (stmts, i, stmt) | |
448 | { | |
449 | struct vertex *v = &(rdg->vertices[i]); | |
450 | ||
451 | /* Record statement to vertex mapping. */ | |
452 | gimple_set_uid (stmt, i); | |
453 | ||
454 | v->data = XNEW (struct rdg_vertex); | |
455 | RDGV_STMT (v) = stmt; | |
456 | RDGV_DATAREFS (v).create (0); | |
457 | RDGV_HAS_MEM_WRITE (v) = false; | |
458 | RDGV_HAS_MEM_READS (v) = false; | |
459 | if (gimple_code (stmt) == GIMPLE_PHI) | |
460 | continue; | |
461 | ||
f3754041 | 462 | unsigned drp = datarefs_vec.length (); |
463 | if (!find_data_references_in_stmt (loop, stmt, &datarefs_vec)) | |
a2740310 | 464 | return false; |
f3754041 | 465 | for (unsigned j = drp; j < datarefs_vec.length (); ++j) |
a2740310 | 466 | { |
f3754041 | 467 | data_reference_p dr = datarefs_vec[j]; |
a2740310 | 468 | if (DR_IS_READ (dr)) |
469 | RDGV_HAS_MEM_READS (v) = true; | |
470 | else | |
471 | RDGV_HAS_MEM_WRITE (v) = true; | |
472 | RDGV_DATAREFS (v).safe_push (dr); | |
6079e9be | 473 | has_nonaddressable_dataref_p |= may_be_nonaddressable_p (dr->ref); |
a2740310 | 474 | } |
475 | } | |
476 | return true; | |
477 | } | |
478 | ||
50eda3a8 | 479 | /* Array mapping basic block's index to its topological order. */ |
480 | static int *bb_top_order_index; | |
481 | /* And size of the array. */ | |
482 | static int bb_top_order_index_size; | |
483 | ||
484 | /* If X has a smaller topological sort number than Y, returns -1; | |
485 | if greater, returns 1. */ | |
486 | ||
487 | static int | |
488 | bb_top_order_cmp (const void *x, const void *y) | |
489 | { | |
490 | basic_block bb1 = *(const basic_block *) x; | |
491 | basic_block bb2 = *(const basic_block *) y; | |
492 | ||
493 | gcc_assert (bb1->index < bb_top_order_index_size | |
494 | && bb2->index < bb_top_order_index_size); | |
495 | gcc_assert (bb1 == bb2 | |
496 | || bb_top_order_index[bb1->index] | |
497 | != bb_top_order_index[bb2->index]); | |
498 | ||
499 | return (bb_top_order_index[bb1->index] - bb_top_order_index[bb2->index]); | |
500 | } | |
501 | ||
502 | /* Initialize STMTS with all the statements of LOOP. We use topological | |
503 | order to discover all statements. The order is important because | |
504 | generate_loops_for_partition is using the same traversal for identifying | |
505 | statements in loop copies. */ | |
a2740310 | 506 | |
507 | static void | |
42acab1c | 508 | stmts_from_loop (struct loop *loop, vec<gimple *> *stmts) |
a2740310 | 509 | { |
510 | unsigned int i; | |
50eda3a8 | 511 | basic_block *bbs = get_loop_body_in_custom_order (loop, bb_top_order_cmp); |
a2740310 | 512 | |
513 | for (i = 0; i < loop->num_nodes; i++) | |
514 | { | |
515 | basic_block bb = bbs[i]; | |
a2740310 | 516 | |
1a91d914 | 517 | for (gphi_iterator bsi = gsi_start_phis (bb); !gsi_end_p (bsi); |
518 | gsi_next (&bsi)) | |
519 | if (!virtual_operand_p (gimple_phi_result (bsi.phi ()))) | |
520 | stmts->safe_push (bsi.phi ()); | |
a2740310 | 521 | |
1a91d914 | 522 | for (gimple_stmt_iterator bsi = gsi_start_bb (bb); !gsi_end_p (bsi); |
523 | gsi_next (&bsi)) | |
a2740310 | 524 | { |
42acab1c | 525 | gimple *stmt = gsi_stmt (bsi); |
a2740310 | 526 | if (gimple_code (stmt) != GIMPLE_LABEL && !is_gimple_debug (stmt)) |
527 | stmts->safe_push (stmt); | |
528 | } | |
529 | } | |
530 | ||
531 | free (bbs); | |
532 | } | |
533 | ||
a2740310 | 534 | /* Free the reduced dependence graph RDG. */ |
535 | ||
536 | static void | |
537 | free_rdg (struct graph *rdg) | |
538 | { | |
539 | int i; | |
540 | ||
541 | for (i = 0; i < rdg->n_vertices; i++) | |
542 | { | |
543 | struct vertex *v = &(rdg->vertices[i]); | |
544 | struct graph_edge *e; | |
545 | ||
546 | for (e = v->succ; e; e = e->succ_next) | |
7103facc | 547 | free (e->data); |
a2740310 | 548 | |
549 | if (v->data) | |
550 | { | |
551 | gimple_set_uid (RDGV_STMT (v), -1); | |
f3754041 | 552 | (RDGV_DATAREFS (v)).release (); |
a2740310 | 553 | free (v->data); |
554 | } | |
555 | } | |
556 | ||
557 | free_graph (rdg); | |
558 | } | |
559 | ||
209a62a6 | 560 | /* Build the Reduced Dependence Graph (RDG) with one vertex per statement of |
561 | LOOP, and one edge per flow dependence or control dependence from control | |
f3754041 | 562 | dependence CD. During visiting each statement, data references are also |
563 | collected and recorded in global data DATAREFS_VEC. */ | |
a2740310 | 564 | |
565 | static struct graph * | |
209a62a6 | 566 | build_rdg (struct loop *loop, control_dependences *cd) |
a2740310 | 567 | { |
568 | struct graph *rdg; | |
a2740310 | 569 | |
eeaec025 | 570 | /* Create the RDG vertices from the stmts of the loop nest. */ |
42acab1c | 571 | auto_vec<gimple *, 10> stmts; |
209a62a6 | 572 | stmts_from_loop (loop, &stmts); |
b223e75a | 573 | rdg = new_graph (stmts.length ()); |
f3754041 | 574 | if (!create_rdg_vertices (rdg, stmts, loop)) |
a2740310 | 575 | { |
a2740310 | 576 | free_rdg (rdg); |
577 | return NULL; | |
578 | } | |
579 | stmts.release (); | |
eeaec025 | 580 | |
7103facc | 581 | create_rdg_flow_edges (rdg); |
582 | if (cd) | |
209a62a6 | 583 | create_rdg_cd_edges (rdg, cd, loop); |
7103facc | 584 | |
a2740310 | 585 | return rdg; |
586 | } | |
587 | ||
a2740310 | 588 | |
f024aa04 | 589 | /* Kind of distributed loop. */ |
ac7a1007 | 590 | enum partition_kind { |
b1c062d8 | 591 | PKIND_NORMAL, |
592 | /* Partial memset stands for a paritition can be distributed into a loop | |
593 | of memset calls, rather than a single memset call. It's handled just | |
594 | like a normal parition, i.e, distributed as separate loop, no memset | |
595 | call is generated. | |
596 | ||
597 | Note: This is a hacking fix trying to distribute ZERO-ing stmt in a | |
598 | loop nest as deep as possible. As a result, parloop achieves better | |
599 | parallelization by parallelizing deeper loop nest. This hack should | |
600 | be unnecessary and removed once distributed memset can be understood | |
601 | and analyzed in data reference analysis. See PR82604 for more. */ | |
602 | PKIND_PARTIAL_MEMSET, | |
603 | PKIND_MEMSET, PKIND_MEMCPY, PKIND_MEMMOVE | |
ac7a1007 | 604 | }; |
d32bc1d7 | 605 | |
f024aa04 | 606 | /* Type of distributed loop. */ |
607 | enum partition_type { | |
608 | /* The distributed loop can be executed parallelly. */ | |
609 | PTYPE_PARALLEL = 0, | |
610 | /* The distributed loop has to be executed sequentially. */ | |
611 | PTYPE_SEQUENTIAL | |
612 | }; | |
613 | ||
506fcb4f | 614 | /* Builtin info for loop distribution. */ |
615 | struct builtin_info | |
616 | { | |
617 | /* data-references a kind != PKIND_NORMAL partition is about. */ | |
618 | data_reference_p dst_dr; | |
619 | data_reference_p src_dr; | |
620 | /* Base address and size of memory objects operated by the builtin. Note | |
621 | both dest and source memory objects must have the same size. */ | |
622 | tree dst_base; | |
623 | tree src_base; | |
624 | tree size; | |
05ebeee6 | 625 | /* Base and offset part of dst_base after stripping constant offset. This |
626 | is only used in memset builtin distribution for now. */ | |
627 | tree dst_base_base; | |
628 | unsigned HOST_WIDE_INT dst_base_offset; | |
506fcb4f | 629 | }; |
630 | ||
889a3926 | 631 | /* Partition for loop distribution. */ |
04009ada | 632 | struct partition |
543506e0 | 633 | { |
889a3926 | 634 | /* Statements of the partition. */ |
543506e0 | 635 | bitmap stmts; |
889a3926 | 636 | /* True if the partition defines variable which is used outside of loop. */ |
df9892ff | 637 | bool reduction_p; |
d32bc1d7 | 638 | enum partition_kind kind; |
f024aa04 | 639 | enum partition_type type; |
889a3926 | 640 | /* Data references in the partition. */ |
641 | bitmap datarefs; | |
506fcb4f | 642 | /* Information of builtin parition. */ |
643 | struct builtin_info *builtin; | |
04009ada | 644 | }; |
543506e0 | 645 | |
543506e0 | 646 | |
647 | /* Allocate and initialize a partition from BITMAP. */ | |
648 | ||
04009ada | 649 | static partition * |
889a3926 | 650 | partition_alloc (void) |
543506e0 | 651 | { |
04009ada | 652 | partition *partition = XCNEW (struct partition); |
889a3926 | 653 | partition->stmts = BITMAP_ALLOC (NULL); |
df9892ff | 654 | partition->reduction_p = false; |
d32bc1d7 | 655 | partition->kind = PKIND_NORMAL; |
889a3926 | 656 | partition->datarefs = BITMAP_ALLOC (NULL); |
543506e0 | 657 | return partition; |
658 | } | |
659 | ||
660 | /* Free PARTITION. */ | |
661 | ||
662 | static void | |
04009ada | 663 | partition_free (partition *partition) |
543506e0 | 664 | { |
665 | BITMAP_FREE (partition->stmts); | |
889a3926 | 666 | BITMAP_FREE (partition->datarefs); |
506fcb4f | 667 | if (partition->builtin) |
668 | free (partition->builtin); | |
669 | ||
543506e0 | 670 | free (partition); |
671 | } | |
672 | ||
d32bc1d7 | 673 | /* Returns true if the partition can be generated as a builtin. */ |
674 | ||
675 | static bool | |
04009ada | 676 | partition_builtin_p (partition *partition) |
d32bc1d7 | 677 | { |
b1c062d8 | 678 | return partition->kind > PKIND_PARTIAL_MEMSET; |
d32bc1d7 | 679 | } |
543506e0 | 680 | |
df9892ff | 681 | /* Returns true if the partition contains a reduction. */ |
e5edce84 | 682 | |
683 | static bool | |
04009ada | 684 | partition_reduction_p (partition *partition) |
e5edce84 | 685 | { |
df9892ff | 686 | return partition->reduction_p; |
e5edce84 | 687 | } |
688 | ||
f1edc00d | 689 | /* Partitions are fused because of different reasons. */ |
690 | enum fuse_type | |
691 | { | |
692 | FUSE_NON_BUILTIN = 0, | |
693 | FUSE_REDUCTION = 1, | |
694 | FUSE_SHARE_REF = 2, | |
695 | FUSE_SAME_SCC = 3, | |
696 | FUSE_FINALIZE = 4 | |
697 | }; | |
698 | ||
699 | /* Description on different fusing reason. */ | |
700 | static const char *fuse_message[] = { | |
701 | "they are non-builtins", | |
702 | "they have reductions", | |
703 | "they have shared memory refs", | |
704 | "they are in the same dependence scc", | |
705 | "there is no point to distribute loop"}; | |
706 | ||
df9892ff | 707 | static void |
f024aa04 | 708 | update_type_for_merge (struct graph *, partition *, partition *); |
f1edc00d | 709 | |
f024aa04 | 710 | /* Merge PARTITION into the partition DEST. RDG is the reduced dependence |
711 | graph and we update type for result partition if it is non-NULL. */ | |
889a3926 | 712 | |
f024aa04 | 713 | static void |
714 | partition_merge_into (struct graph *rdg, partition *dest, | |
715 | partition *partition, enum fuse_type ft) | |
716 | { | |
f1edc00d | 717 | if (dump_file && (dump_flags & TDF_DETAILS)) |
718 | { | |
719 | fprintf (dump_file, "Fuse partitions because %s:\n", fuse_message[ft]); | |
720 | fprintf (dump_file, " Part 1: "); | |
721 | dump_bitmap (dump_file, dest->stmts); | |
722 | fprintf (dump_file, " Part 2: "); | |
723 | dump_bitmap (dump_file, partition->stmts); | |
724 | } | |
f024aa04 | 725 | |
726 | dest->kind = PKIND_NORMAL; | |
727 | if (dest->type == PTYPE_PARALLEL) | |
728 | dest->type = partition->type; | |
729 | ||
730 | bitmap_ior_into (dest->stmts, partition->stmts); | |
731 | if (partition_reduction_p (partition)) | |
732 | dest->reduction_p = true; | |
733 | ||
734 | /* Further check if any data dependence prevents us from executing the | |
735 | new partition parallelly. */ | |
736 | if (dest->type == PTYPE_PARALLEL && rdg != NULL) | |
737 | update_type_for_merge (rdg, dest, partition); | |
738 | ||
739 | bitmap_ior_into (dest->datarefs, partition->datarefs); | |
df9892ff | 740 | } |
741 | ||
742 | ||
b5698e04 | 743 | /* Returns true when DEF is an SSA_NAME defined in LOOP and used after |
744 | the LOOP. */ | |
745 | ||
746 | static bool | |
747 | ssa_name_has_uses_outside_loop_p (tree def, loop_p loop) | |
748 | { | |
749 | imm_use_iterator imm_iter; | |
750 | use_operand_p use_p; | |
751 | ||
752 | FOR_EACH_IMM_USE_FAST (use_p, imm_iter, def) | |
64a31469 | 753 | { |
883b4905 | 754 | if (is_gimple_debug (USE_STMT (use_p))) |
755 | continue; | |
756 | ||
757 | basic_block use_bb = gimple_bb (USE_STMT (use_p)); | |
758 | if (!flow_bb_inside_loop_p (loop, use_bb)) | |
64a31469 | 759 | return true; |
760 | } | |
b5698e04 | 761 | |
762 | return false; | |
763 | } | |
764 | ||
765 | /* Returns true when STMT defines a scalar variable used after the | |
bfcf35ff | 766 | loop LOOP. */ |
b5698e04 | 767 | |
768 | static bool | |
42acab1c | 769 | stmt_has_scalar_dependences_outside_loop (loop_p loop, gimple *stmt) |
b5698e04 | 770 | { |
bfcf35ff | 771 | def_operand_p def_p; |
772 | ssa_op_iter op_iter; | |
b5698e04 | 773 | |
35ec0372 | 774 | if (gimple_code (stmt) == GIMPLE_PHI) |
775 | return ssa_name_has_uses_outside_loop_p (gimple_phi_result (stmt), loop); | |
776 | ||
bfcf35ff | 777 | FOR_EACH_SSA_DEF_OPERAND (def_p, stmt, op_iter, SSA_OP_DEF) |
778 | if (ssa_name_has_uses_outside_loop_p (DEF_FROM_PTR (def_p), loop)) | |
779 | return true; | |
b5698e04 | 780 | |
bfcf35ff | 781 | return false; |
b5698e04 | 782 | } |
783 | ||
801c5610 | 784 | /* Return a copy of LOOP placed before LOOP. */ |
785 | ||
786 | static struct loop * | |
787 | copy_loop_before (struct loop *loop) | |
788 | { | |
789 | struct loop *res; | |
790 | edge preheader = loop_preheader_edge (loop); | |
791 | ||
801c5610 | 792 | initialize_original_copy_tables (); |
c71d3c24 | 793 | res = slpeel_tree_duplicate_loop_to_edge_cfg (loop, NULL, preheader); |
d32bc1d7 | 794 | gcc_assert (res != NULL); |
801c5610 | 795 | free_original_copy_tables (); |
c9b2c569 | 796 | delete_update_ssa (); |
801c5610 | 797 | |
798 | return res; | |
799 | } | |
800 | ||
801 | /* Creates an empty basic block after LOOP. */ | |
802 | ||
803 | static void | |
804 | create_bb_after_loop (struct loop *loop) | |
805 | { | |
806 | edge exit = single_exit (loop); | |
807 | ||
808 | if (!exit) | |
809 | return; | |
810 | ||
811 | split_edge (exit); | |
812 | } | |
813 | ||
814 | /* Generate code for PARTITION from the code in LOOP. The loop is | |
815 | copied when COPY_P is true. All the statements not flagged in the | |
816 | PARTITION bitmap are removed from the loop or from its copy. The | |
817 | statements are indexed in sequence inside a basic block, and the | |
d32bc1d7 | 818 | basic blocks of a loop are taken in dom order. */ |
801c5610 | 819 | |
d32bc1d7 | 820 | static void |
04009ada | 821 | generate_loops_for_partition (struct loop *loop, partition *partition, |
543506e0 | 822 | bool copy_p) |
801c5610 | 823 | { |
15c8650d | 824 | unsigned i; |
801c5610 | 825 | basic_block *bbs; |
826 | ||
827 | if (copy_p) | |
828 | { | |
f562e2ea | 829 | int orig_loop_num = loop->orig_loop_num; |
801c5610 | 830 | loop = copy_loop_before (loop); |
d32bc1d7 | 831 | gcc_assert (loop != NULL); |
f562e2ea | 832 | loop->orig_loop_num = orig_loop_num; |
801c5610 | 833 | create_preheader (loop, CP_SIMPLE_PREHEADERS); |
834 | create_bb_after_loop (loop); | |
835 | } | |
f562e2ea | 836 | else |
837 | { | |
838 | /* Origin number is set to the new versioned loop's num. */ | |
839 | gcc_assert (loop->orig_loop_num != loop->num); | |
840 | } | |
801c5610 | 841 | |
15c8650d | 842 | /* Remove stmts not in the PARTITION bitmap. */ |
801c5610 | 843 | bbs = get_loop_body_in_dom_order (loop); |
844 | ||
c64f38bf | 845 | if (MAY_HAVE_DEBUG_BIND_STMTS) |
15c8650d | 846 | for (i = 0; i < loop->num_nodes; i++) |
8ebd8f29 | 847 | { |
848 | basic_block bb = bbs[i]; | |
849 | ||
1a91d914 | 850 | for (gphi_iterator bsi = gsi_start_phis (bb); !gsi_end_p (bsi); |
851 | gsi_next (&bsi)) | |
15c8650d | 852 | { |
1a91d914 | 853 | gphi *phi = bsi.phi (); |
15c8650d | 854 | if (!virtual_operand_p (gimple_phi_result (phi)) |
855 | && !bitmap_bit_p (partition->stmts, gimple_uid (phi))) | |
856 | reset_debug_uses (phi); | |
857 | } | |
8ebd8f29 | 858 | |
1a91d914 | 859 | for (gimple_stmt_iterator bsi = gsi_start_bb (bb); !gsi_end_p (bsi); gsi_next (&bsi)) |
8ebd8f29 | 860 | { |
42acab1c | 861 | gimple *stmt = gsi_stmt (bsi); |
8ebd8f29 | 862 | if (gimple_code (stmt) != GIMPLE_LABEL |
863 | && !is_gimple_debug (stmt) | |
15c8650d | 864 | && !bitmap_bit_p (partition->stmts, gimple_uid (stmt))) |
8ebd8f29 | 865 | reset_debug_uses (stmt); |
866 | } | |
867 | } | |
868 | ||
15c8650d | 869 | for (i = 0; i < loop->num_nodes; i++) |
801c5610 | 870 | { |
871 | basic_block bb = bbs[i]; | |
142b4ffb | 872 | edge inner_exit = NULL; |
873 | ||
874 | if (loop != bb->loop_father) | |
875 | inner_exit = single_exit (bb->loop_father); | |
801c5610 | 876 | |
1a91d914 | 877 | for (gphi_iterator bsi = gsi_start_phis (bb); !gsi_end_p (bsi);) |
15c8650d | 878 | { |
1a91d914 | 879 | gphi *phi = bsi.phi (); |
15c8650d | 880 | if (!virtual_operand_p (gimple_phi_result (phi)) |
881 | && !bitmap_bit_p (partition->stmts, gimple_uid (phi))) | |
ff24e1e2 | 882 | remove_phi_node (&bsi, true); |
15c8650d | 883 | else |
884 | gsi_next (&bsi); | |
885 | } | |
801c5610 | 886 | |
1a91d914 | 887 | for (gimple_stmt_iterator bsi = gsi_start_bb (bb); !gsi_end_p (bsi);) |
ff24e1e2 | 888 | { |
42acab1c | 889 | gimple *stmt = gsi_stmt (bsi); |
8ebd8f29 | 890 | if (gimple_code (stmt) != GIMPLE_LABEL |
891 | && !is_gimple_debug (stmt) | |
15c8650d | 892 | && !bitmap_bit_p (partition->stmts, gimple_uid (stmt))) |
ff24e1e2 | 893 | { |
142b4ffb | 894 | /* In distribution of loop nest, if bb is inner loop's exit_bb, |
895 | we choose its exit edge/path in order to avoid generating | |
896 | infinite loop. For all other cases, we choose an arbitrary | |
897 | path through the empty CFG part that this unnecessary | |
898 | control stmt controls. */ | |
1a91d914 | 899 | if (gcond *cond_stmt = dyn_cast <gcond *> (stmt)) |
f1ce84d9 | 900 | { |
142b4ffb | 901 | if (inner_exit && inner_exit->flags & EDGE_TRUE_VALUE) |
902 | gimple_cond_make_true (cond_stmt); | |
903 | else | |
904 | gimple_cond_make_false (cond_stmt); | |
f1ce84d9 | 905 | update_stmt (stmt); |
906 | } | |
907 | else if (gimple_code (stmt) == GIMPLE_SWITCH) | |
908 | { | |
1a91d914 | 909 | gswitch *switch_stmt = as_a <gswitch *> (stmt); |
f1ce84d9 | 910 | gimple_switch_set_index |
1a91d914 | 911 | (switch_stmt, CASE_LOW (gimple_switch_label (switch_stmt, 1))); |
f1ce84d9 | 912 | update_stmt (stmt); |
913 | } | |
914 | else | |
915 | { | |
916 | unlink_stmt_vdef (stmt); | |
917 | gsi_remove (&bsi, true); | |
918 | release_defs (stmt); | |
919 | continue; | |
920 | } | |
ff24e1e2 | 921 | } |
f1ce84d9 | 922 | gsi_next (&bsi); |
ff24e1e2 | 923 | } |
801c5610 | 924 | } |
925 | ||
926 | free (bbs); | |
801c5610 | 927 | } |
928 | ||
0dc30b0c | 929 | /* If VAL memory representation contains the same value in all bytes, |
930 | return that value, otherwise return -1. | |
931 | E.g. for 0x24242424 return 0x24, for IEEE double | |
932 | 747708026454360457216.0 return 0x44, etc. */ | |
933 | ||
934 | static int | |
935 | const_with_all_bytes_same (tree val) | |
936 | { | |
937 | unsigned char buf[64]; | |
938 | int i, len; | |
939 | ||
940 | if (integer_zerop (val) | |
0dc30b0c | 941 | || (TREE_CODE (val) == CONSTRUCTOR |
942 | && !TREE_CLOBBER_P (val) | |
943 | && CONSTRUCTOR_NELTS (val) == 0)) | |
944 | return 0; | |
945 | ||
53e7aca1 | 946 | if (real_zerop (val)) |
947 | { | |
948 | /* Only return 0 for +0.0, not for -0.0, which doesn't have | |
949 | an all bytes same memory representation. Don't transform | |
950 | -0.0 stores into +0.0 even for !HONOR_SIGNED_ZEROS. */ | |
951 | switch (TREE_CODE (val)) | |
952 | { | |
953 | case REAL_CST: | |
954 | if (!real_isneg (TREE_REAL_CST_PTR (val))) | |
955 | return 0; | |
956 | break; | |
957 | case COMPLEX_CST: | |
958 | if (!const_with_all_bytes_same (TREE_REALPART (val)) | |
959 | && !const_with_all_bytes_same (TREE_IMAGPART (val))) | |
960 | return 0; | |
961 | break; | |
962 | case VECTOR_CST: | |
0a2b1323 | 963 | { |
964 | unsigned int count = vector_cst_encoded_nelts (val); | |
965 | unsigned int j; | |
966 | for (j = 0; j < count; ++j) | |
967 | if (const_with_all_bytes_same (VECTOR_CST_ENCODED_ELT (val, j))) | |
968 | break; | |
969 | if (j == count) | |
970 | return 0; | |
971 | break; | |
972 | } | |
53e7aca1 | 973 | default: |
974 | break; | |
975 | } | |
976 | } | |
977 | ||
0dc30b0c | 978 | if (CHAR_BIT != 8 || BITS_PER_UNIT != 8) |
979 | return -1; | |
980 | ||
981 | len = native_encode_expr (val, buf, sizeof (buf)); | |
982 | if (len == 0) | |
983 | return -1; | |
984 | for (i = 1; i < len; i++) | |
985 | if (buf[i] != buf[0]) | |
986 | return -1; | |
987 | return buf[0]; | |
988 | } | |
989 | ||
d32bc1d7 | 990 | /* Generate a call to memset for PARTITION in LOOP. */ |
801c5610 | 991 | |
a136cad6 | 992 | static void |
04009ada | 993 | generate_memset_builtin (struct loop *loop, partition *partition) |
801c5610 | 994 | { |
d32bc1d7 | 995 | gimple_stmt_iterator gsi; |
f59ae127 | 996 | tree mem, fn, nb_bytes; |
0644fcba | 997 | tree val; |
506fcb4f | 998 | struct builtin_info *builtin = partition->builtin; |
999 | gimple *fn_call; | |
d32bc1d7 | 1000 | |
1001 | /* The new statements will be placed before LOOP. */ | |
1002 | gsi = gsi_last_bb (loop_preheader_edge (loop)->src); | |
801c5610 | 1003 | |
bc6e884f | 1004 | nb_bytes = rewrite_to_non_trapping_overflow (builtin->size); |
f689d33d | 1005 | nb_bytes = force_gimple_operand_gsi (&gsi, nb_bytes, true, NULL_TREE, |
1006 | false, GSI_CONTINUE_LINKING); | |
506fcb4f | 1007 | mem = builtin->dst_base; |
f689d33d | 1008 | mem = force_gimple_operand_gsi (&gsi, mem, true, NULL_TREE, |
1009 | false, GSI_CONTINUE_LINKING); | |
801c5610 | 1010 | |
0644fcba | 1011 | /* This exactly matches the pattern recognition in classify_partition. */ |
506fcb4f | 1012 | val = gimple_assign_rhs1 (DR_STMT (builtin->dst_dr)); |
0dc30b0c | 1013 | /* Handle constants like 0x15151515 and similarly |
1014 | floating point constants etc. where all bytes are the same. */ | |
1015 | int bytev = const_with_all_bytes_same (val); | |
1016 | if (bytev != -1) | |
1017 | val = build_int_cst (integer_type_node, bytev); | |
1018 | else if (TREE_CODE (val) == INTEGER_CST) | |
1019 | val = fold_convert (integer_type_node, val); | |
1020 | else if (!useless_type_conversion_p (integer_type_node, TREE_TYPE (val))) | |
0644fcba | 1021 | { |
f9e245b2 | 1022 | tree tem = make_ssa_name (integer_type_node); |
42acab1c | 1023 | gimple *cstmt = gimple_build_assign (tem, NOP_EXPR, val); |
0dc30b0c | 1024 | gsi_insert_after (&gsi, cstmt, GSI_CONTINUE_LINKING); |
1025 | val = tem; | |
0644fcba | 1026 | } |
1027 | ||
b9a16870 | 1028 | fn = build_fold_addr_expr (builtin_decl_implicit (BUILT_IN_MEMSET)); |
0644fcba | 1029 | fn_call = gimple_build_call (fn, 3, mem, val, nb_bytes); |
f689d33d | 1030 | gsi_insert_after (&gsi, fn_call, GSI_CONTINUE_LINKING); |
801c5610 | 1031 | |
1032 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
0644fcba | 1033 | { |
1034 | fprintf (dump_file, "generated memset"); | |
0dc30b0c | 1035 | if (bytev == 0) |
0644fcba | 1036 | fprintf (dump_file, " zero\n"); |
0644fcba | 1037 | else |
1038 | fprintf (dump_file, "\n"); | |
1039 | } | |
801c5610 | 1040 | } |
1041 | ||
f689d33d | 1042 | /* Generate a call to memcpy for PARTITION in LOOP. */ |
1043 | ||
1044 | static void | |
04009ada | 1045 | generate_memcpy_builtin (struct loop *loop, partition *partition) |
f689d33d | 1046 | { |
1047 | gimple_stmt_iterator gsi; | |
506fcb4f | 1048 | gimple *fn_call; |
f59ae127 | 1049 | tree dest, src, fn, nb_bytes; |
f689d33d | 1050 | enum built_in_function kind; |
506fcb4f | 1051 | struct builtin_info *builtin = partition->builtin; |
f689d33d | 1052 | |
1053 | /* The new statements will be placed before LOOP. */ | |
1054 | gsi = gsi_last_bb (loop_preheader_edge (loop)->src); | |
1055 | ||
bc6e884f | 1056 | nb_bytes = rewrite_to_non_trapping_overflow (builtin->size); |
f689d33d | 1057 | nb_bytes = force_gimple_operand_gsi (&gsi, nb_bytes, true, NULL_TREE, |
1058 | false, GSI_CONTINUE_LINKING); | |
506fcb4f | 1059 | dest = builtin->dst_base; |
1060 | src = builtin->src_base; | |
5d4e2409 | 1061 | if (partition->kind == PKIND_MEMCPY |
1062 | || ! ptr_derefs_may_alias_p (dest, src)) | |
f689d33d | 1063 | kind = BUILT_IN_MEMCPY; |
5d4e2409 | 1064 | else |
1065 | kind = BUILT_IN_MEMMOVE; | |
f689d33d | 1066 | |
1067 | dest = force_gimple_operand_gsi (&gsi, dest, true, NULL_TREE, | |
1068 | false, GSI_CONTINUE_LINKING); | |
1069 | src = force_gimple_operand_gsi (&gsi, src, true, NULL_TREE, | |
1070 | false, GSI_CONTINUE_LINKING); | |
1071 | fn = build_fold_addr_expr (builtin_decl_implicit (kind)); | |
1072 | fn_call = gimple_build_call (fn, 3, dest, src, nb_bytes); | |
1073 | gsi_insert_after (&gsi, fn_call, GSI_CONTINUE_LINKING); | |
1074 | ||
1075 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
1076 | { | |
1077 | if (kind == BUILT_IN_MEMCPY) | |
1078 | fprintf (dump_file, "generated memcpy\n"); | |
1079 | else | |
1080 | fprintf (dump_file, "generated memmove\n"); | |
1081 | } | |
1082 | } | |
1083 | ||
d32bc1d7 | 1084 | /* Remove and destroy the loop LOOP. */ |
801c5610 | 1085 | |
d32bc1d7 | 1086 | static void |
1087 | destroy_loop (struct loop *loop) | |
801c5610 | 1088 | { |
d32bc1d7 | 1089 | unsigned nbbs = loop->num_nodes; |
1090 | edge exit = single_exit (loop); | |
1091 | basic_block src = loop_preheader_edge (loop)->src, dest = exit->dest; | |
801c5610 | 1092 | basic_block *bbs; |
d32bc1d7 | 1093 | unsigned i; |
801c5610 | 1094 | |
1095 | bbs = get_loop_body_in_dom_order (loop); | |
1096 | ||
d32bc1d7 | 1097 | redirect_edge_pred (exit, src); |
1098 | exit->flags &= ~(EDGE_TRUE_VALUE|EDGE_FALSE_VALUE); | |
1099 | exit->flags |= EDGE_FALLTHRU; | |
1100 | cancel_loop_tree (loop); | |
1101 | rescan_loop_exit (exit, false, true); | |
801c5610 | 1102 | |
2084e279 | 1103 | i = nbbs; |
1104 | do | |
54459dd6 | 1105 | { |
1106 | /* We have made sure to not leave any dangling uses of SSA | |
1107 | names defined in the loop. With the exception of virtuals. | |
1108 | Make sure we replace all uses of virtual defs that will remain | |
1109 | outside of the loop with the bare symbol as delete_basic_block | |
1110 | will release them. */ | |
2084e279 | 1111 | --i; |
1a91d914 | 1112 | for (gphi_iterator gsi = gsi_start_phis (bbs[i]); !gsi_end_p (gsi); |
1113 | gsi_next (&gsi)) | |
54459dd6 | 1114 | { |
1a91d914 | 1115 | gphi *phi = gsi.phi (); |
7c782c9b | 1116 | if (virtual_operand_p (gimple_phi_result (phi))) |
54459dd6 | 1117 | mark_virtual_phi_result_for_renaming (phi); |
1118 | } | |
1a91d914 | 1119 | for (gimple_stmt_iterator gsi = gsi_start_bb (bbs[i]); !gsi_end_p (gsi); |
1120 | gsi_next (&gsi)) | |
54459dd6 | 1121 | { |
42acab1c | 1122 | gimple *stmt = gsi_stmt (gsi); |
54459dd6 | 1123 | tree vdef = gimple_vdef (stmt); |
1124 | if (vdef && TREE_CODE (vdef) == SSA_NAME) | |
1125 | mark_virtual_operand_for_renaming (vdef); | |
1126 | } | |
1127 | delete_basic_block (bbs[i]); | |
1128 | } | |
2084e279 | 1129 | while (i != 0); |
1130 | ||
801c5610 | 1131 | free (bbs); |
d32bc1d7 | 1132 | |
1133 | set_immediate_dominator (CDI_DOMINATORS, dest, | |
1134 | recompute_dominator (CDI_DOMINATORS, dest)); | |
801c5610 | 1135 | } |
1136 | ||
77d095c0 | 1137 | /* Generates code for PARTITION. Return whether LOOP needs to be destroyed. */ |
801c5610 | 1138 | |
77d095c0 | 1139 | static bool |
f689d33d | 1140 | generate_code_for_partition (struct loop *loop, |
04009ada | 1141 | partition *partition, bool copy_p) |
801c5610 | 1142 | { |
d32bc1d7 | 1143 | switch (partition->kind) |
1144 | { | |
df9892ff | 1145 | case PKIND_NORMAL: |
b1c062d8 | 1146 | case PKIND_PARTIAL_MEMSET: |
df9892ff | 1147 | /* Reductions all have to be in the last partition. */ |
1148 | gcc_assert (!partition_reduction_p (partition) | |
1149 | || !copy_p); | |
1150 | generate_loops_for_partition (loop, partition, copy_p); | |
77d095c0 | 1151 | return false; |
df9892ff | 1152 | |
d32bc1d7 | 1153 | case PKIND_MEMSET: |
f689d33d | 1154 | generate_memset_builtin (loop, partition); |
f689d33d | 1155 | break; |
1156 | ||
1157 | case PKIND_MEMCPY: | |
5d4e2409 | 1158 | case PKIND_MEMMOVE: |
f689d33d | 1159 | generate_memcpy_builtin (loop, partition); |
d32bc1d7 | 1160 | break; |
1161 | ||
1162 | default: | |
1163 | gcc_unreachable (); | |
1164 | } | |
801c5610 | 1165 | |
df9892ff | 1166 | /* Common tail for partitions we turn into a call. If this was the last |
1167 | partition for which we generate code, we have to destroy the loop. */ | |
1168 | if (!copy_p) | |
77d095c0 | 1169 | return true; |
1170 | return false; | |
801c5610 | 1171 | } |
1172 | ||
50f5937e | 1173 | /* Return data dependence relation for data references A and B. The two |
1174 | data references must be in lexicographic order wrto reduced dependence | |
1175 | graph RDG. We firstly try to find ddr from global ddr hash table. If | |
1176 | it doesn't exist, compute the ddr and cache it. */ | |
1177 | ||
1178 | static data_dependence_relation * | |
1179 | get_data_dependence (struct graph *rdg, data_reference_p a, data_reference_p b) | |
1180 | { | |
1181 | struct data_dependence_relation ent, **slot; | |
1182 | struct data_dependence_relation *ddr; | |
1183 | ||
1184 | gcc_assert (DR_IS_WRITE (a) || DR_IS_WRITE (b)); | |
1185 | gcc_assert (rdg_vertex_for_stmt (rdg, DR_STMT (a)) | |
1186 | <= rdg_vertex_for_stmt (rdg, DR_STMT (b))); | |
1187 | ent.a = a; | |
1188 | ent.b = b; | |
bbb229ef | 1189 | slot = ddrs_table->find_slot (&ent, INSERT); |
50f5937e | 1190 | if (*slot == NULL) |
1191 | { | |
1192 | ddr = initialize_data_dependence_relation (a, b, loop_nest); | |
1193 | compute_affine_dependence (ddr, loop_nest[0]); | |
1194 | *slot = ddr; | |
1195 | } | |
1196 | ||
1197 | return *slot; | |
1198 | } | |
801c5610 | 1199 | |
f024aa04 | 1200 | /* In reduced dependence graph RDG for loop distribution, return true if |
1201 | dependence between references DR1 and DR2 leads to a dependence cycle | |
1202 | and such dependence cycle can't be resolved by runtime alias check. */ | |
1203 | ||
1204 | static bool | |
1205 | data_dep_in_cycle_p (struct graph *rdg, | |
1206 | data_reference_p dr1, data_reference_p dr2) | |
1207 | { | |
1208 | struct data_dependence_relation *ddr; | |
1209 | ||
1210 | /* Re-shuffle data-refs to be in topological order. */ | |
1211 | if (rdg_vertex_for_stmt (rdg, DR_STMT (dr1)) | |
1212 | > rdg_vertex_for_stmt (rdg, DR_STMT (dr2))) | |
1213 | std::swap (dr1, dr2); | |
1214 | ||
1215 | ddr = get_data_dependence (rdg, dr1, dr2); | |
1216 | ||
1217 | /* In case of no data dependence. */ | |
1218 | if (DDR_ARE_DEPENDENT (ddr) == chrec_known) | |
1219 | return false; | |
1220 | /* For unknown data dependence or known data dependence which can't be | |
1221 | expressed in classic distance vector, we check if it can be resolved | |
1222 | by runtime alias check. If yes, we still consider data dependence | |
1223 | as won't introduce data dependence cycle. */ | |
1224 | else if (DDR_ARE_DEPENDENT (ddr) == chrec_dont_know | |
1225 | || DDR_NUM_DIST_VECTS (ddr) == 0) | |
1226 | return !runtime_alias_check_p (ddr, NULL, true); | |
1227 | else if (DDR_NUM_DIST_VECTS (ddr) > 1) | |
1228 | return true; | |
1229 | else if (DDR_REVERSED_P (ddr) | |
1230 | || lambda_vector_zerop (DDR_DIST_VECT (ddr, 0), 1)) | |
1231 | return false; | |
1232 | ||
1233 | return true; | |
1234 | } | |
1235 | ||
1236 | /* Given reduced dependence graph RDG, PARTITION1 and PARTITION2, update | |
1237 | PARTITION1's type after merging PARTITION2 into PARTITION1. */ | |
1238 | ||
1239 | static void | |
1240 | update_type_for_merge (struct graph *rdg, | |
1241 | partition *partition1, partition *partition2) | |
1242 | { | |
1243 | unsigned i, j; | |
1244 | bitmap_iterator bi, bj; | |
1245 | data_reference_p dr1, dr2; | |
1246 | ||
1247 | EXECUTE_IF_SET_IN_BITMAP (partition1->datarefs, 0, i, bi) | |
1248 | { | |
1249 | unsigned start = (partition1 == partition2) ? i + 1 : 0; | |
1250 | ||
1251 | dr1 = datarefs_vec[i]; | |
1252 | EXECUTE_IF_SET_IN_BITMAP (partition2->datarefs, start, j, bj) | |
1253 | { | |
1254 | dr2 = datarefs_vec[j]; | |
1255 | if (DR_IS_READ (dr1) && DR_IS_READ (dr2)) | |
1256 | continue; | |
1257 | ||
1258 | /* Partition can only be executed sequentially if there is any | |
1259 | data dependence cycle. */ | |
1260 | if (data_dep_in_cycle_p (rdg, dr1, dr2)) | |
1261 | { | |
1262 | partition1->type = PTYPE_SEQUENTIAL; | |
1263 | return; | |
1264 | } | |
1265 | } | |
1266 | } | |
1267 | } | |
1268 | ||
b223e75a | 1269 | /* Returns a partition with all the statements needed for computing |
1270 | the vertex V of the RDG, also including the loop exit conditions. */ | |
801c5610 | 1271 | |
04009ada | 1272 | static partition * |
b223e75a | 1273 | build_rdg_partition_for_vertex (struct graph *rdg, int v) |
801c5610 | 1274 | { |
889a3926 | 1275 | partition *partition = partition_alloc (); |
4997014d | 1276 | auto_vec<int, 3> nodes; |
889a3926 | 1277 | unsigned i, j; |
801c5610 | 1278 | int x; |
889a3926 | 1279 | data_reference_p dr; |
801c5610 | 1280 | |
5529df60 | 1281 | graphds_dfs (rdg, &v, 1, &nodes, false, NULL); |
801c5610 | 1282 | |
f1f41a6c | 1283 | FOR_EACH_VEC_ELT (nodes, i, x) |
b223e75a | 1284 | { |
1285 | bitmap_set_bit (partition->stmts, x); | |
889a3926 | 1286 | |
1287 | for (j = 0; RDG_DATAREFS (rdg, x).iterate (j, &dr); ++j) | |
1288 | { | |
1289 | unsigned idx = (unsigned) DR_INDEX (dr); | |
1290 | gcc_assert (idx < datarefs_vec.length ()); | |
1291 | ||
f024aa04 | 1292 | /* Partition can only be executed sequentially if there is any |
1293 | unknown data reference. */ | |
1294 | if (!DR_BASE_ADDRESS (dr) || !DR_OFFSET (dr) | |
1295 | || !DR_INIT (dr) || !DR_STEP (dr)) | |
1296 | partition->type = PTYPE_SEQUENTIAL; | |
1297 | ||
889a3926 | 1298 | bitmap_set_bit (partition->datarefs, idx); |
1299 | } | |
b223e75a | 1300 | } |
801c5610 | 1301 | |
f024aa04 | 1302 | if (partition->type == PTYPE_SEQUENTIAL) |
1303 | return partition; | |
1304 | ||
1305 | /* Further check if any data dependence prevents us from executing the | |
1306 | partition parallelly. */ | |
1307 | update_type_for_merge (rdg, partition, partition); | |
1308 | ||
801c5610 | 1309 | return partition; |
1310 | } | |
1311 | ||
b363c31b | 1312 | /* Given PARTITION of LOOP and RDG, record single load/store data references |
1313 | for builtin partition in SRC_DR/DST_DR, return false if there is no such | |
506fcb4f | 1314 | data references. */ |
a136cad6 | 1315 | |
506fcb4f | 1316 | static bool |
b363c31b | 1317 | find_single_drs (struct loop *loop, struct graph *rdg, partition *partition, |
506fcb4f | 1318 | data_reference_p *dst_dr, data_reference_p *src_dr) |
a136cad6 | 1319 | { |
d32bc1d7 | 1320 | unsigned i; |
506fcb4f | 1321 | data_reference_p single_ld = NULL, single_st = NULL; |
1322 | bitmap_iterator bi; | |
ac7a1007 | 1323 | |
d32bc1d7 | 1324 | EXECUTE_IF_SET_IN_BITMAP (partition->stmts, 0, i, bi) |
1325 | { | |
42acab1c | 1326 | gimple *stmt = RDG_STMT (rdg, i); |
f689d33d | 1327 | data_reference_p dr; |
d32bc1d7 | 1328 | |
1329 | if (gimple_code (stmt) == GIMPLE_PHI) | |
1330 | continue; | |
1331 | ||
1332 | /* Any scalar stmts are ok. */ | |
1333 | if (!gimple_vuse (stmt)) | |
1334 | continue; | |
1335 | ||
f689d33d | 1336 | /* Otherwise just regular loads/stores. */ |
1337 | if (!gimple_assign_single_p (stmt)) | |
506fcb4f | 1338 | return false; |
f689d33d | 1339 | |
1340 | /* But exactly one store and/or load. */ | |
506fcb4f | 1341 | for (unsigned j = 0; RDG_DATAREFS (rdg, i).iterate (j, &dr); ++j) |
d32bc1d7 | 1342 | { |
c8a8b2cf | 1343 | tree type = TREE_TYPE (DR_REF (dr)); |
1344 | ||
1345 | /* The memset, memcpy and memmove library calls are only | |
1346 | able to deal with generic address space. */ | |
1347 | if (!ADDR_SPACE_GENERIC_P (TYPE_ADDR_SPACE (type))) | |
506fcb4f | 1348 | return false; |
c8a8b2cf | 1349 | |
f689d33d | 1350 | if (DR_IS_READ (dr)) |
1351 | { | |
506fcb4f | 1352 | if (single_ld != NULL) |
1353 | return false; | |
1354 | single_ld = dr; | |
f689d33d | 1355 | } |
1356 | else | |
1357 | { | |
506fcb4f | 1358 | if (single_st != NULL) |
1359 | return false; | |
1360 | single_st = dr; | |
f689d33d | 1361 | } |
d32bc1d7 | 1362 | } |
d32bc1d7 | 1363 | } |
1364 | ||
506fcb4f | 1365 | if (!single_st) |
1366 | return false; | |
1367 | ||
1368 | /* Bail out if this is a bitfield memory reference. */ | |
1369 | if (TREE_CODE (DR_REF (single_st)) == COMPONENT_REF | |
1370 | && DECL_BIT_FIELD (TREE_OPERAND (DR_REF (single_st), 1))) | |
1371 | return false; | |
f59ae127 | 1372 | |
b363c31b | 1373 | /* Data reference must be executed exactly once per iteration of each |
1374 | loop in the loop nest. We only need to check dominance information | |
1375 | against the outermost one in a perfect loop nest because a bb can't | |
1376 | dominate outermost loop's latch without dominating inner loop's. */ | |
506fcb4f | 1377 | basic_block bb_st = gimple_bb (DR_STMT (single_st)); |
b363c31b | 1378 | if (!dominated_by_p (CDI_DOMINATORS, loop->latch, bb_st)) |
506fcb4f | 1379 | return false; |
1380 | ||
1381 | if (single_ld) | |
883b4905 | 1382 | { |
506fcb4f | 1383 | gimple *store = DR_STMT (single_st), *load = DR_STMT (single_ld); |
1384 | /* Direct aggregate copy or via an SSA name temporary. */ | |
1385 | if (load != store | |
1386 | && gimple_assign_lhs (load) != gimple_assign_rhs1 (store)) | |
1387 | return false; | |
883b4905 | 1388 | |
506fcb4f | 1389 | /* Bail out if this is a bitfield memory reference. */ |
1390 | if (TREE_CODE (DR_REF (single_ld)) == COMPONENT_REF | |
1391 | && DECL_BIT_FIELD (TREE_OPERAND (DR_REF (single_ld), 1))) | |
1392 | return false; | |
1393 | ||
1394 | /* Load and store must be in the same loop nest. */ | |
1395 | basic_block bb_ld = gimple_bb (DR_STMT (single_ld)); | |
b363c31b | 1396 | if (bb_st->loop_father != bb_ld->loop_father) |
506fcb4f | 1397 | return false; |
1398 | ||
b363c31b | 1399 | /* Data reference must be executed exactly once per iteration. |
1400 | Same as single_st, we only need to check against the outermost | |
1401 | loop. */ | |
1402 | if (!dominated_by_p (CDI_DOMINATORS, loop->latch, bb_ld)) | |
506fcb4f | 1403 | return false; |
1404 | ||
b363c31b | 1405 | edge e = single_exit (bb_st->loop_father); |
506fcb4f | 1406 | bool dom_ld = dominated_by_p (CDI_DOMINATORS, e->src, bb_ld); |
1407 | bool dom_st = dominated_by_p (CDI_DOMINATORS, e->src, bb_st); | |
1408 | if (dom_ld != dom_st) | |
1409 | return false; | |
1410 | } | |
1411 | ||
1412 | *src_dr = single_ld; | |
1413 | *dst_dr = single_st; | |
1414 | return true; | |
1415 | } | |
1416 | ||
1417 | /* Given data reference DR in LOOP_NEST, this function checks the enclosing | |
1418 | loops from inner to outer to see if loop's step equals to access size at | |
b1c062d8 | 1419 | each level of loop. Return 2 if we can prove this at all level loops; |
1420 | record access base and size in BASE and SIZE; save loop's step at each | |
1421 | level of loop in STEPS if it is not null. For example: | |
506fcb4f | 1422 | |
1423 | int arr[100][100][100]; | |
1424 | for (i = 0; i < 100; i++) ;steps[2] = 40000 | |
1425 | for (j = 100; j > 0; j--) ;steps[1] = -400 | |
1426 | for (k = 0; k < 100; k++) ;steps[0] = 4 | |
b1c062d8 | 1427 | arr[i][j - 1][k] = 0; ;base = &arr, size = 4000000 |
506fcb4f | 1428 | |
b1c062d8 | 1429 | Return 1 if we can prove the equality at the innermost loop, but not all |
1430 | level loops. In this case, no information is recorded. | |
1431 | ||
1432 | Return 0 if no equality can be proven at any level loops. */ | |
1433 | ||
1434 | static int | |
506fcb4f | 1435 | compute_access_range (loop_p loop_nest, data_reference_p dr, tree *base, |
1436 | tree *size, vec<tree> *steps = NULL) | |
1437 | { | |
1438 | location_t loc = gimple_location (DR_STMT (dr)); | |
1439 | basic_block bb = gimple_bb (DR_STMT (dr)); | |
1440 | struct loop *loop = bb->loop_father; | |
1441 | tree ref = DR_REF (dr); | |
1442 | tree access_base = build_fold_addr_expr (ref); | |
1443 | tree access_size = TYPE_SIZE_UNIT (TREE_TYPE (ref)); | |
b1c062d8 | 1444 | int res = 0; |
506fcb4f | 1445 | |
1446 | do { | |
1447 | tree scev_fn = analyze_scalar_evolution (loop, access_base); | |
1448 | if (TREE_CODE (scev_fn) != POLYNOMIAL_CHREC) | |
b1c062d8 | 1449 | return res; |
883b4905 | 1450 | |
506fcb4f | 1451 | access_base = CHREC_LEFT (scev_fn); |
1452 | if (tree_contains_chrecs (access_base, NULL)) | |
b1c062d8 | 1453 | return res; |
506fcb4f | 1454 | |
1455 | tree scev_step = CHREC_RIGHT (scev_fn); | |
1456 | /* Only support constant steps. */ | |
1457 | if (TREE_CODE (scev_step) != INTEGER_CST) | |
b1c062d8 | 1458 | return res; |
506fcb4f | 1459 | |
1460 | enum ev_direction access_dir = scev_direction (scev_fn); | |
1461 | if (access_dir == EV_DIR_UNKNOWN) | |
b1c062d8 | 1462 | return res; |
506fcb4f | 1463 | |
1464 | if (steps != NULL) | |
1465 | steps->safe_push (scev_step); | |
1466 | ||
1467 | scev_step = fold_convert_loc (loc, sizetype, scev_step); | |
1468 | /* Compute absolute value of scev step. */ | |
1469 | if (access_dir == EV_DIR_DECREASES) | |
1470 | scev_step = fold_build1_loc (loc, NEGATE_EXPR, sizetype, scev_step); | |
1471 | ||
1472 | /* At each level of loop, scev step must equal to access size. In other | |
1473 | words, DR must access consecutive memory between loop iterations. */ | |
1474 | if (!operand_equal_p (scev_step, access_size, 0)) | |
b1c062d8 | 1475 | return res; |
1476 | ||
1477 | /* Access stride can be computed for data reference at least for the | |
1478 | innermost loop. */ | |
1479 | res = 1; | |
506fcb4f | 1480 | |
1481 | /* Compute DR's execution times in loop. */ | |
1482 | tree niters = number_of_latch_executions (loop); | |
1483 | niters = fold_convert_loc (loc, sizetype, niters); | |
1484 | if (dominated_by_p (CDI_DOMINATORS, single_exit (loop)->src, bb)) | |
1485 | niters = size_binop_loc (loc, PLUS_EXPR, niters, size_one_node); | |
1486 | ||
1487 | /* Compute DR's overall access size in loop. */ | |
1488 | access_size = fold_build2_loc (loc, MULT_EXPR, sizetype, | |
1489 | niters, scev_step); | |
1490 | /* Adjust base address in case of negative step. */ | |
1491 | if (access_dir == EV_DIR_DECREASES) | |
883b4905 | 1492 | { |
506fcb4f | 1493 | tree adj = fold_build2_loc (loc, MINUS_EXPR, sizetype, |
1494 | scev_step, access_size); | |
1495 | access_base = fold_build_pointer_plus_loc (loc, access_base, adj); | |
883b4905 | 1496 | } |
506fcb4f | 1497 | } while (loop != loop_nest && (loop = loop_outer (loop)) != NULL); |
1498 | ||
1499 | *base = access_base; | |
1500 | *size = access_size; | |
b1c062d8 | 1501 | /* Access stride can be computed for data reference at each level loop. */ |
1502 | return 2; | |
506fcb4f | 1503 | } |
1504 | ||
1505 | /* Allocate and return builtin struct. Record information like DST_DR, | |
1506 | SRC_DR, DST_BASE, SRC_BASE and SIZE in the allocated struct. */ | |
1507 | ||
1508 | static struct builtin_info * | |
1509 | alloc_builtin (data_reference_p dst_dr, data_reference_p src_dr, | |
1510 | tree dst_base, tree src_base, tree size) | |
1511 | { | |
1512 | struct builtin_info *builtin = XNEW (struct builtin_info); | |
1513 | builtin->dst_dr = dst_dr; | |
1514 | builtin->src_dr = src_dr; | |
1515 | builtin->dst_base = dst_base; | |
1516 | builtin->src_base = src_base; | |
1517 | builtin->size = size; | |
1518 | return builtin; | |
1519 | } | |
1520 | ||
1521 | /* Given data reference DR in loop nest LOOP, classify if it forms builtin | |
1522 | memset call. */ | |
1523 | ||
1524 | static void | |
1525 | classify_builtin_st (loop_p loop, partition *partition, data_reference_p dr) | |
1526 | { | |
1527 | gimple *stmt = DR_STMT (dr); | |
1528 | tree base, size, rhs = gimple_assign_rhs1 (stmt); | |
1529 | ||
1530 | if (const_with_all_bytes_same (rhs) == -1 | |
1531 | && (!INTEGRAL_TYPE_P (TREE_TYPE (rhs)) | |
1532 | || (TYPE_MODE (TREE_TYPE (rhs)) | |
1533 | != TYPE_MODE (unsigned_char_type_node)))) | |
1534 | return; | |
1535 | ||
1536 | if (TREE_CODE (rhs) == SSA_NAME | |
1537 | && !SSA_NAME_IS_DEFAULT_DEF (rhs) | |
1538 | && flow_bb_inside_loop_p (loop, gimple_bb (SSA_NAME_DEF_STMT (rhs)))) | |
1539 | return; | |
1540 | ||
b1c062d8 | 1541 | int res = compute_access_range (loop, dr, &base, &size); |
1542 | if (res == 0) | |
506fcb4f | 1543 | return; |
b1c062d8 | 1544 | if (res == 1) |
1545 | { | |
1546 | partition->kind = PKIND_PARTIAL_MEMSET; | |
1547 | return; | |
1548 | } | |
506fcb4f | 1549 | |
caf8edd6 | 1550 | poly_uint64 base_offset; |
1551 | unsigned HOST_WIDE_INT const_base_offset; | |
1552 | tree base_base = strip_offset (base, &base_offset); | |
1553 | if (!base_offset.is_constant (&const_base_offset)) | |
1554 | return; | |
1555 | ||
05ebeee6 | 1556 | struct builtin_info *builtin; |
1557 | builtin = alloc_builtin (dr, NULL, base, NULL_TREE, size); | |
caf8edd6 | 1558 | builtin->dst_base_base = base_base; |
1559 | builtin->dst_base_offset = const_base_offset; | |
05ebeee6 | 1560 | partition->builtin = builtin; |
506fcb4f | 1561 | partition->kind = PKIND_MEMSET; |
1562 | } | |
1563 | ||
1564 | /* Given data references DST_DR and SRC_DR in loop nest LOOP and RDG, classify | |
1565 | if it forms builtin memcpy or memmove call. */ | |
1566 | ||
1567 | static void | |
1568 | classify_builtin_ldst (loop_p loop, struct graph *rdg, partition *partition, | |
1569 | data_reference_p dst_dr, data_reference_p src_dr) | |
1570 | { | |
1571 | tree base, size, src_base, src_size; | |
1572 | auto_vec<tree> dst_steps, src_steps; | |
1573 | ||
b1c062d8 | 1574 | /* Compute access range of both load and store. */ |
1575 | int res = compute_access_range (loop, dst_dr, &base, &size, &dst_steps); | |
1576 | if (res != 2) | |
1577 | return; | |
1578 | res = compute_access_range (loop, src_dr, &src_base, &src_size, &src_steps); | |
1579 | if (res != 2) | |
1580 | return; | |
1581 | ||
1582 | /* They much have the same access size. */ | |
1583 | if (!operand_equal_p (size, src_size, 0)) | |
506fcb4f | 1584 | return; |
1585 | ||
1586 | /* Load and store in loop nest must access memory in the same way, i.e, | |
1587 | their must have the same steps in each loop of the nest. */ | |
1588 | if (dst_steps.length () != src_steps.length ()) | |
1589 | return; | |
1590 | for (unsigned i = 0; i < dst_steps.length (); ++i) | |
1591 | if (!operand_equal_p (dst_steps[i], src_steps[i], 0)) | |
1592 | return; | |
1593 | ||
1594 | /* Now check that if there is a dependence. */ | |
1595 | ddr_p ddr = get_data_dependence (rdg, src_dr, dst_dr); | |
1596 | ||
1597 | /* Classify as memcpy if no dependence between load and store. */ | |
1598 | if (DDR_ARE_DEPENDENT (ddr) == chrec_known) | |
1599 | { | |
1600 | partition->builtin = alloc_builtin (dst_dr, src_dr, base, src_base, size); | |
1601 | partition->kind = PKIND_MEMCPY; | |
1602 | return; | |
883b4905 | 1603 | } |
1604 | ||
506fcb4f | 1605 | /* Can't do memmove in case of unknown dependence or dependence without |
1606 | classical distance vector. */ | |
1607 | if (DDR_ARE_DEPENDENT (ddr) == chrec_dont_know | |
1608 | || DDR_NUM_DIST_VECTS (ddr) == 0) | |
1609 | return; | |
f59ae127 | 1610 | |
506fcb4f | 1611 | unsigned i; |
1612 | lambda_vector dist_v; | |
1613 | int num_lev = (DDR_LOOP_NEST (ddr)).length (); | |
1614 | FOR_EACH_VEC_ELT (DDR_DIST_VECTS (ddr), i, dist_v) | |
f689d33d | 1615 | { |
506fcb4f | 1616 | unsigned dep_lev = dependence_level (dist_v, num_lev); |
1617 | /* Can't do memmove if load depends on store. */ | |
1618 | if (dep_lev > 0 && dist_v[dep_lev - 1] > 0 && !DDR_REVERSED_P (ddr)) | |
f689d33d | 1619 | return; |
f689d33d | 1620 | } |
506fcb4f | 1621 | |
1622 | partition->builtin = alloc_builtin (dst_dr, src_dr, base, src_base, size); | |
1623 | partition->kind = PKIND_MEMMOVE; | |
1624 | return; | |
1625 | } | |
1626 | ||
1627 | /* Classifies the builtin kind we can generate for PARTITION of RDG and LOOP. | |
1628 | For the moment we detect memset, memcpy and memmove patterns. Bitmap | |
1629 | STMT_IN_ALL_PARTITIONS contains statements belonging to all partitions. */ | |
1630 | ||
1631 | static void | |
1632 | classify_partition (loop_p loop, struct graph *rdg, partition *partition, | |
1633 | bitmap stmt_in_all_partitions) | |
1634 | { | |
1635 | bitmap_iterator bi; | |
1636 | unsigned i; | |
1637 | data_reference_p single_ld = NULL, single_st = NULL; | |
1638 | bool volatiles_p = false, has_reduction = false; | |
1639 | ||
1640 | EXECUTE_IF_SET_IN_BITMAP (partition->stmts, 0, i, bi) | |
f689d33d | 1641 | { |
506fcb4f | 1642 | gimple *stmt = RDG_STMT (rdg, i); |
50f5937e | 1643 | |
506fcb4f | 1644 | if (gimple_has_volatile_ops (stmt)) |
1645 | volatiles_p = true; | |
50f5937e | 1646 | |
506fcb4f | 1647 | /* If the stmt is not included by all partitions and there is uses |
1648 | outside of the loop, then mark the partition as reduction. */ | |
1649 | if (stmt_has_scalar_dependences_outside_loop (loop, stmt)) | |
1650 | { | |
1651 | /* Due to limitation in the transform phase we have to fuse all | |
1652 | reduction partitions. As a result, this could cancel valid | |
1653 | loop distribution especially for loop that induction variable | |
1654 | is used outside of loop. To workaround this issue, we skip | |
1655 | marking partition as reudction if the reduction stmt belongs | |
1656 | to all partitions. In such case, reduction will be computed | |
1657 | correctly no matter how partitions are fused/distributed. */ | |
1658 | if (!bitmap_bit_p (stmt_in_all_partitions, i)) | |
7b6f8db4 | 1659 | { |
506fcb4f | 1660 | partition->reduction_p = true; |
1661 | return; | |
7b6f8db4 | 1662 | } |
506fcb4f | 1663 | has_reduction = true; |
7b6f8db4 | 1664 | } |
f689d33d | 1665 | } |
506fcb4f | 1666 | |
1667 | /* Perform general partition disqualification for builtins. */ | |
1668 | if (volatiles_p | |
1669 | /* Simple workaround to prevent classifying the partition as builtin | |
1670 | if it contains any use outside of loop. */ | |
1671 | || has_reduction | |
1672 | || !flag_tree_loop_distribute_patterns) | |
1673 | return; | |
1674 | ||
1675 | /* Find single load/store data references for builtin partition. */ | |
b363c31b | 1676 | if (!find_single_drs (loop, rdg, partition, &single_st, &single_ld)) |
506fcb4f | 1677 | return; |
1678 | ||
1679 | /* Classify the builtin kind. */ | |
1680 | if (single_ld == NULL) | |
1681 | classify_builtin_st (loop, partition, single_st); | |
1682 | else | |
1683 | classify_builtin_ldst (loop, rdg, partition, single_st, single_ld); | |
a136cad6 | 1684 | } |
1685 | ||
fd34627b | 1686 | /* Returns true when PARTITION1 and PARTITION2 access the same memory |
1687 | object in RDG. */ | |
a136cad6 | 1688 | |
1689 | static bool | |
fd34627b | 1690 | share_memory_accesses (struct graph *rdg, |
1691 | partition *partition1, partition *partition2) | |
a136cad6 | 1692 | { |
fd34627b | 1693 | unsigned i, j; |
a136cad6 | 1694 | bitmap_iterator bi, bj; |
fd34627b | 1695 | data_reference_p dr1, dr2; |
f83623cc | 1696 | |
1697 | /* First check whether in the intersection of the two partitions are | |
1698 | any loads or stores. Common loads are the situation that happens | |
1699 | most often. */ | |
1700 | EXECUTE_IF_AND_IN_BITMAP (partition1->stmts, partition2->stmts, 0, i, bi) | |
1701 | if (RDG_MEM_WRITE_STMT (rdg, i) | |
1702 | || RDG_MEM_READS_STMT (rdg, i)) | |
1703 | return true; | |
a136cad6 | 1704 | |
fd34627b | 1705 | /* Then check whether the two partitions access the same memory object. */ |
1706 | EXECUTE_IF_SET_IN_BITMAP (partition1->datarefs, 0, i, bi) | |
1707 | { | |
1708 | dr1 = datarefs_vec[i]; | |
1709 | ||
1710 | if (!DR_BASE_ADDRESS (dr1) | |
1711 | || !DR_OFFSET (dr1) || !DR_INIT (dr1) || !DR_STEP (dr1)) | |
1712 | continue; | |
1713 | ||
1714 | EXECUTE_IF_SET_IN_BITMAP (partition2->datarefs, 0, j, bj) | |
1715 | { | |
1716 | dr2 = datarefs_vec[j]; | |
1717 | ||
1718 | if (!DR_BASE_ADDRESS (dr2) | |
1719 | || !DR_OFFSET (dr2) || !DR_INIT (dr2) || !DR_STEP (dr2)) | |
1720 | continue; | |
1721 | ||
1722 | if (operand_equal_p (DR_BASE_ADDRESS (dr1), DR_BASE_ADDRESS (dr2), 0) | |
1723 | && operand_equal_p (DR_OFFSET (dr1), DR_OFFSET (dr2), 0) | |
1724 | && operand_equal_p (DR_INIT (dr1), DR_INIT (dr2), 0) | |
1725 | && operand_equal_p (DR_STEP (dr1), DR_STEP (dr2), 0)) | |
1726 | return true; | |
1727 | } | |
1728 | } | |
a136cad6 | 1729 | |
1730 | return false; | |
1731 | } | |
1732 | ||
8d95fe31 | 1733 | /* For each seed statement in STARTING_STMTS, this function builds |
1734 | partition for it by adding depended statements according to RDG. | |
1735 | All partitions are recorded in PARTITIONS. */ | |
801c5610 | 1736 | |
1737 | static void | |
88e0cdd9 | 1738 | rdg_build_partitions (struct graph *rdg, |
42acab1c | 1739 | vec<gimple *> starting_stmts, |
04009ada | 1740 | vec<partition *> *partitions) |
801c5610 | 1741 | { |
035def86 | 1742 | auto_bitmap processed; |
15c8650d | 1743 | int i; |
42acab1c | 1744 | gimple *stmt; |
801c5610 | 1745 | |
15c8650d | 1746 | FOR_EACH_VEC_ELT (starting_stmts, i, stmt) |
801c5610 | 1747 | { |
15c8650d | 1748 | int v = rdg_vertex_for_stmt (rdg, stmt); |
1749 | ||
1750 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
1751 | fprintf (dump_file, | |
1752 | "ldist asked to generate code for vertex %d\n", v); | |
48e1416a | 1753 | |
b223e75a | 1754 | /* If the vertex is already contained in another partition so |
1755 | is the partition rooted at it. */ | |
801c5610 | 1756 | if (bitmap_bit_p (processed, v)) |
1757 | continue; | |
48e1416a | 1758 | |
04009ada | 1759 | partition *partition = build_rdg_partition_for_vertex (rdg, v); |
b223e75a | 1760 | bitmap_ior_into (processed, partition->stmts); |
801c5610 | 1761 | |
df9892ff | 1762 | if (dump_file && (dump_flags & TDF_DETAILS)) |
801c5610 | 1763 | { |
f024aa04 | 1764 | fprintf (dump_file, "ldist creates useful %s partition:\n", |
1765 | partition->type == PTYPE_PARALLEL ? "parallel" : "sequent"); | |
1766 | bitmap_print (dump_file, partition->stmts, " ", "\n"); | |
801c5610 | 1767 | } |
df9892ff | 1768 | |
1769 | partitions->safe_push (partition); | |
801c5610 | 1770 | } |
1771 | ||
88e0cdd9 | 1772 | /* All vertices should have been assigned to at least one partition now, |
1773 | other than vertices belonging to dead code. */ | |
801c5610 | 1774 | } |
1775 | ||
1776 | /* Dump to FILE the PARTITIONS. */ | |
1777 | ||
1778 | static void | |
04009ada | 1779 | dump_rdg_partitions (FILE *file, vec<partition *> partitions) |
801c5610 | 1780 | { |
1781 | int i; | |
04009ada | 1782 | partition *partition; |
801c5610 | 1783 | |
f1f41a6c | 1784 | FOR_EACH_VEC_ELT (partitions, i, partition) |
543506e0 | 1785 | debug_bitmap_file (file, partition->stmts); |
801c5610 | 1786 | } |
1787 | ||
1788 | /* Debug PARTITIONS. */ | |
04009ada | 1789 | extern void debug_rdg_partitions (vec<partition *> ); |
801c5610 | 1790 | |
4b987fac | 1791 | DEBUG_FUNCTION void |
04009ada | 1792 | debug_rdg_partitions (vec<partition *> partitions) |
801c5610 | 1793 | { |
1794 | dump_rdg_partitions (stderr, partitions); | |
1795 | } | |
1796 | ||
577982d8 | 1797 | /* Returns the number of read and write operations in the RDG. */ |
1798 | ||
1799 | static int | |
1800 | number_of_rw_in_rdg (struct graph *rdg) | |
1801 | { | |
1802 | int i, res = 0; | |
1803 | ||
1804 | for (i = 0; i < rdg->n_vertices; i++) | |
1805 | { | |
1806 | if (RDG_MEM_WRITE_STMT (rdg, i)) | |
1807 | ++res; | |
1808 | ||
1809 | if (RDG_MEM_READS_STMT (rdg, i)) | |
1810 | ++res; | |
1811 | } | |
1812 | ||
1813 | return res; | |
1814 | } | |
1815 | ||
1816 | /* Returns the number of read and write operations in a PARTITION of | |
1817 | the RDG. */ | |
1818 | ||
1819 | static int | |
04009ada | 1820 | number_of_rw_in_partition (struct graph *rdg, partition *partition) |
577982d8 | 1821 | { |
1822 | int res = 0; | |
1823 | unsigned i; | |
1824 | bitmap_iterator ii; | |
1825 | ||
543506e0 | 1826 | EXECUTE_IF_SET_IN_BITMAP (partition->stmts, 0, i, ii) |
577982d8 | 1827 | { |
1828 | if (RDG_MEM_WRITE_STMT (rdg, i)) | |
1829 | ++res; | |
1830 | ||
1831 | if (RDG_MEM_READS_STMT (rdg, i)) | |
1832 | ++res; | |
1833 | } | |
1834 | ||
1835 | return res; | |
1836 | } | |
1837 | ||
1838 | /* Returns true when one of the PARTITIONS contains all the read or | |
1839 | write operations of RDG. */ | |
1840 | ||
1841 | static bool | |
f1f41a6c | 1842 | partition_contains_all_rw (struct graph *rdg, |
04009ada | 1843 | vec<partition *> partitions) |
577982d8 | 1844 | { |
1845 | int i; | |
04009ada | 1846 | partition *partition; |
577982d8 | 1847 | int nrw = number_of_rw_in_rdg (rdg); |
1848 | ||
f1f41a6c | 1849 | FOR_EACH_VEC_ELT (partitions, i, partition) |
577982d8 | 1850 | if (nrw == number_of_rw_in_partition (rdg, partition)) |
1851 | return true; | |
1852 | ||
1853 | return false; | |
1854 | } | |
1855 | ||
7103facc | 1856 | /* Compute partition dependence created by the data references in DRS1 |
f562e2ea | 1857 | and DRS2, modify and return DIR according to that. IF ALIAS_DDR is |
1858 | not NULL, we record dependence introduced by possible alias between | |
1859 | two data references in ALIAS_DDRS; otherwise, we simply ignore such | |
1860 | dependence as if it doesn't exist at all. */ | |
7103facc | 1861 | |
1862 | static int | |
209a62a6 | 1863 | pg_add_dependence_edges (struct graph *rdg, int dir, |
f562e2ea | 1864 | bitmap drs1, bitmap drs2, vec<ddr_p> *alias_ddrs) |
7103facc | 1865 | { |
889a3926 | 1866 | unsigned i, j; |
1867 | bitmap_iterator bi, bj; | |
1868 | data_reference_p dr1, dr2, saved_dr1; | |
7103facc | 1869 | |
1870 | /* dependence direction - 0 is no dependence, -1 is back, | |
1871 | 1 is forth, 2 is both (we can stop then, merging will occur). */ | |
889a3926 | 1872 | EXECUTE_IF_SET_IN_BITMAP (drs1, 0, i, bi) |
1873 | { | |
1874 | dr1 = datarefs_vec[i]; | |
1875 | ||
1876 | EXECUTE_IF_SET_IN_BITMAP (drs2, 0, j, bj) | |
1877 | { | |
f562e2ea | 1878 | int res, this_dir = 1; |
1879 | ddr_p ddr; | |
1880 | ||
889a3926 | 1881 | dr2 = datarefs_vec[j]; |
1882 | ||
1883 | /* Skip all <read, read> data dependence. */ | |
1884 | if (DR_IS_READ (dr1) && DR_IS_READ (dr2)) | |
1885 | continue; | |
1886 | ||
1887 | saved_dr1 = dr1; | |
f562e2ea | 1888 | /* Re-shuffle data-refs to be in topological order. */ |
889a3926 | 1889 | if (rdg_vertex_for_stmt (rdg, DR_STMT (dr1)) |
1890 | > rdg_vertex_for_stmt (rdg, DR_STMT (dr2))) | |
1891 | { | |
1892 | std::swap (dr1, dr2); | |
1893 | this_dir = -this_dir; | |
1894 | } | |
50f5937e | 1895 | ddr = get_data_dependence (rdg, dr1, dr2); |
889a3926 | 1896 | if (DDR_ARE_DEPENDENT (ddr) == chrec_dont_know) |
f562e2ea | 1897 | { |
1898 | this_dir = 0; | |
1899 | res = data_ref_compare_tree (DR_BASE_ADDRESS (dr1), | |
1900 | DR_BASE_ADDRESS (dr2)); | |
1901 | /* Be conservative. If data references are not well analyzed, | |
1902 | or the two data references have the same base address and | |
1903 | offset, add dependence and consider it alias to each other. | |
f4d3c071 | 1904 | In other words, the dependence cannot be resolved by |
f562e2ea | 1905 | runtime alias check. */ |
1906 | if (!DR_BASE_ADDRESS (dr1) || !DR_BASE_ADDRESS (dr2) | |
1907 | || !DR_OFFSET (dr1) || !DR_OFFSET (dr2) | |
1908 | || !DR_INIT (dr1) || !DR_INIT (dr2) | |
1909 | || !DR_STEP (dr1) || !tree_fits_uhwi_p (DR_STEP (dr1)) | |
1910 | || !DR_STEP (dr2) || !tree_fits_uhwi_p (DR_STEP (dr2)) | |
1911 | || res == 0) | |
1912 | this_dir = 2; | |
1913 | /* Data dependence could be resolved by runtime alias check, | |
1914 | record it in ALIAS_DDRS. */ | |
1915 | else if (alias_ddrs != NULL) | |
1916 | alias_ddrs->safe_push (ddr); | |
1917 | /* Or simply ignore it. */ | |
1918 | } | |
889a3926 | 1919 | else if (DDR_ARE_DEPENDENT (ddr) == NULL_TREE) |
1920 | { | |
1921 | if (DDR_REVERSED_P (ddr)) | |
f562e2ea | 1922 | this_dir = -this_dir; |
1923 | ||
889a3926 | 1924 | /* Known dependences can still be unordered througout the |
1925 | iteration space, see gcc.dg/tree-ssa/ldist-16.c. */ | |
1926 | if (DDR_NUM_DIST_VECTS (ddr) != 1) | |
1927 | this_dir = 2; | |
1928 | /* If the overlap is exact preserve stmt order. */ | |
21e5a1b6 | 1929 | else if (lambda_vector_zerop (DDR_DIST_VECT (ddr, 0), |
1930 | DDR_NB_LOOPS (ddr))) | |
889a3926 | 1931 | ; |
f562e2ea | 1932 | /* Else as the distance vector is lexicographic positive swap |
1933 | the dependence direction. */ | |
889a3926 | 1934 | else |
f562e2ea | 1935 | this_dir = -this_dir; |
889a3926 | 1936 | } |
1937 | else | |
1938 | this_dir = 0; | |
889a3926 | 1939 | if (this_dir == 2) |
1940 | return 2; | |
1941 | else if (dir == 0) | |
1942 | dir = this_dir; | |
1943 | else if (this_dir != 0 && dir != this_dir) | |
1944 | return 2; | |
1945 | /* Shuffle "back" dr1. */ | |
1946 | dr1 = saved_dr1; | |
1947 | } | |
1948 | } | |
7103facc | 1949 | return dir; |
1950 | } | |
1951 | ||
1952 | /* Compare postorder number of the partition graph vertices V1 and V2. */ | |
1953 | ||
1954 | static int | |
1955 | pgcmp (const void *v1_, const void *v2_) | |
1956 | { | |
1957 | const vertex *v1 = (const vertex *)v1_; | |
1958 | const vertex *v2 = (const vertex *)v2_; | |
1959 | return v2->post - v1->post; | |
1960 | } | |
15c8650d | 1961 | |
f562e2ea | 1962 | /* Data attached to vertices of partition dependence graph. */ |
1963 | struct pg_vdata | |
1964 | { | |
1965 | /* ID of the corresponding partition. */ | |
1966 | int id; | |
1967 | /* The partition. */ | |
1968 | struct partition *partition; | |
1969 | }; | |
1970 | ||
1971 | /* Data attached to edges of partition dependence graph. */ | |
1972 | struct pg_edata | |
1973 | { | |
1974 | /* If the dependence edge can be resolved by runtime alias check, | |
1975 | this vector contains data dependence relations for runtime alias | |
1976 | check. On the other hand, if the dependence edge is introduced | |
1977 | because of compilation time known data dependence, this vector | |
1978 | contains nothing. */ | |
1979 | vec<ddr_p> alias_ddrs; | |
1980 | }; | |
1981 | ||
1982 | /* Callback data for traversing edges in graph. */ | |
1983 | struct pg_edge_callback_data | |
1984 | { | |
1985 | /* Bitmap contains strong connected components should be merged. */ | |
1986 | bitmap sccs_to_merge; | |
1987 | /* Array constains component information for all vertices. */ | |
1988 | int *vertices_component; | |
1989 | /* Vector to record all data dependence relations which are needed | |
1990 | to break strong connected components by runtime alias checks. */ | |
1991 | vec<ddr_p> *alias_ddrs; | |
1992 | }; | |
1993 | ||
1994 | /* Initialize vertice's data for partition dependence graph PG with | |
1995 | PARTITIONS. */ | |
1996 | ||
1997 | static void | |
1998 | init_partition_graph_vertices (struct graph *pg, | |
1999 | vec<struct partition *> *partitions) | |
2000 | { | |
2001 | int i; | |
2002 | partition *partition; | |
2003 | struct pg_vdata *data; | |
2004 | ||
2005 | for (i = 0; partitions->iterate (i, &partition); ++i) | |
2006 | { | |
2007 | data = new pg_vdata; | |
2008 | pg->vertices[i].data = data; | |
2009 | data->id = i; | |
2010 | data->partition = partition; | |
2011 | } | |
2012 | } | |
2013 | ||
2014 | /* Add edge <I, J> to partition dependence graph PG. Attach vector of data | |
2015 | dependence relations to the EDGE if DDRS isn't NULL. */ | |
2016 | ||
2017 | static void | |
2018 | add_partition_graph_edge (struct graph *pg, int i, int j, vec<ddr_p> *ddrs) | |
2019 | { | |
2020 | struct graph_edge *e = add_edge (pg, i, j); | |
2021 | ||
2022 | /* If the edge is attached with data dependence relations, it means this | |
2023 | dependence edge can be resolved by runtime alias checks. */ | |
2024 | if (ddrs != NULL) | |
2025 | { | |
2026 | struct pg_edata *data = new pg_edata; | |
2027 | ||
2028 | gcc_assert (ddrs->length () > 0); | |
2029 | e->data = data; | |
2030 | data->alias_ddrs = vNULL; | |
2031 | data->alias_ddrs.safe_splice (*ddrs); | |
2032 | } | |
2033 | } | |
2034 | ||
2035 | /* Callback function for graph travesal algorithm. It returns true | |
2036 | if edge E should skipped when traversing the graph. */ | |
2037 | ||
2038 | static bool | |
2039 | pg_skip_alias_edge (struct graph_edge *e) | |
2040 | { | |
2041 | struct pg_edata *data = (struct pg_edata *)e->data; | |
2042 | return (data != NULL && data->alias_ddrs.length () > 0); | |
2043 | } | |
2044 | ||
2045 | /* Callback function freeing data attached to edge E of graph. */ | |
2046 | ||
2047 | static void | |
2048 | free_partition_graph_edata_cb (struct graph *, struct graph_edge *e, void *) | |
2049 | { | |
2050 | if (e->data != NULL) | |
2051 | { | |
2052 | struct pg_edata *data = (struct pg_edata *)e->data; | |
2053 | data->alias_ddrs.release (); | |
2054 | delete data; | |
2055 | } | |
2056 | } | |
2057 | ||
2058 | /* Free data attached to vertice of partition dependence graph PG. */ | |
2059 | ||
2060 | static void | |
2061 | free_partition_graph_vdata (struct graph *pg) | |
2062 | { | |
2063 | int i; | |
2064 | struct pg_vdata *data; | |
2065 | ||
2066 | for (i = 0; i < pg->n_vertices; ++i) | |
2067 | { | |
2068 | data = (struct pg_vdata *)pg->vertices[i].data; | |
2069 | delete data; | |
2070 | } | |
2071 | } | |
2072 | ||
2073 | /* Build and return partition dependence graph for PARTITIONS. RDG is | |
2074 | reduced dependence graph for the loop to be distributed. If IGNORE_ALIAS_P | |
2075 | is true, data dependence caused by possible alias between references | |
2076 | is ignored, as if it doesn't exist at all; otherwise all depdendences | |
2077 | are considered. */ | |
2078 | ||
2079 | static struct graph * | |
2080 | build_partition_graph (struct graph *rdg, | |
2081 | vec<struct partition *> *partitions, | |
2082 | bool ignore_alias_p) | |
2083 | { | |
2084 | int i, j; | |
2085 | struct partition *partition1, *partition2; | |
2086 | graph *pg = new_graph (partitions->length ()); | |
2087 | auto_vec<ddr_p> alias_ddrs, *alias_ddrs_p; | |
2088 | ||
2089 | alias_ddrs_p = ignore_alias_p ? NULL : &alias_ddrs; | |
2090 | ||
2091 | init_partition_graph_vertices (pg, partitions); | |
2092 | ||
2093 | for (i = 0; partitions->iterate (i, &partition1); ++i) | |
2094 | { | |
2095 | for (j = i + 1; partitions->iterate (j, &partition2); ++j) | |
2096 | { | |
2097 | /* dependence direction - 0 is no dependence, -1 is back, | |
2098 | 1 is forth, 2 is both (we can stop then, merging will occur). */ | |
2099 | int dir = 0; | |
2100 | ||
2101 | /* If the first partition has reduction, add back edge; if the | |
2102 | second partition has reduction, add forth edge. This makes | |
2103 | sure that reduction partition will be sorted as the last one. */ | |
2104 | if (partition_reduction_p (partition1)) | |
2105 | dir = -1; | |
2106 | else if (partition_reduction_p (partition2)) | |
2107 | dir = 1; | |
2108 | ||
2109 | /* Cleanup the temporary vector. */ | |
2110 | alias_ddrs.truncate (0); | |
2111 | ||
2112 | dir = pg_add_dependence_edges (rdg, dir, partition1->datarefs, | |
2113 | partition2->datarefs, alias_ddrs_p); | |
2114 | ||
2115 | /* Add edge to partition graph if there exists dependence. There | |
2116 | are two types of edges. One type edge is caused by compilation | |
f4d3c071 | 2117 | time known dependence, this type cannot be resolved by runtime |
f562e2ea | 2118 | alias check. The other type can be resolved by runtime alias |
2119 | check. */ | |
2120 | if (dir == 1 || dir == 2 | |
2121 | || alias_ddrs.length () > 0) | |
2122 | { | |
2123 | /* Attach data dependence relations to edge that can be resolved | |
2124 | by runtime alias check. */ | |
2125 | bool alias_edge_p = (dir != 1 && dir != 2); | |
2126 | add_partition_graph_edge (pg, i, j, | |
2127 | (alias_edge_p) ? &alias_ddrs : NULL); | |
2128 | } | |
2129 | if (dir == -1 || dir == 2 | |
2130 | || alias_ddrs.length () > 0) | |
2131 | { | |
2132 | /* Attach data dependence relations to edge that can be resolved | |
2133 | by runtime alias check. */ | |
2134 | bool alias_edge_p = (dir != -1 && dir != 2); | |
2135 | add_partition_graph_edge (pg, j, i, | |
2136 | (alias_edge_p) ? &alias_ddrs : NULL); | |
2137 | } | |
2138 | } | |
2139 | } | |
2140 | return pg; | |
2141 | } | |
2142 | ||
cb072485 | 2143 | /* Sort partitions in PG in descending post order and store them in |
2144 | PARTITIONS. */ | |
f562e2ea | 2145 | |
2146 | static void | |
2147 | sort_partitions_by_post_order (struct graph *pg, | |
2148 | vec<struct partition *> *partitions) | |
2149 | { | |
2150 | int i; | |
2151 | struct pg_vdata *data; | |
2152 | ||
cb072485 | 2153 | /* Now order the remaining nodes in descending postorder. */ |
f562e2ea | 2154 | qsort (pg->vertices, pg->n_vertices, sizeof (vertex), pgcmp); |
2155 | partitions->truncate (0); | |
2156 | for (i = 0; i < pg->n_vertices; ++i) | |
2157 | { | |
2158 | data = (struct pg_vdata *)pg->vertices[i].data; | |
2159 | if (data->partition) | |
2160 | partitions->safe_push (data->partition); | |
2161 | } | |
2162 | } | |
2163 | ||
2164 | /* Given reduced dependence graph RDG merge strong connected components | |
883b4905 | 2165 | of PARTITIONS. If IGNORE_ALIAS_P is true, data dependence caused by |
2166 | possible alias between references is ignored, as if it doesn't exist | |
2167 | at all; otherwise all depdendences are considered. */ | |
f562e2ea | 2168 | |
2169 | static void | |
2170 | merge_dep_scc_partitions (struct graph *rdg, | |
883b4905 | 2171 | vec<struct partition *> *partitions, |
2172 | bool ignore_alias_p) | |
f562e2ea | 2173 | { |
2174 | struct partition *partition1, *partition2; | |
2175 | struct pg_vdata *data; | |
883b4905 | 2176 | graph *pg = build_partition_graph (rdg, partitions, ignore_alias_p); |
f562e2ea | 2177 | int i, j, num_sccs = graphds_scc (pg, NULL); |
2178 | ||
2179 | /* Strong connected compoenent means dependence cycle, we cannot distribute | |
2180 | them. So fuse them together. */ | |
2181 | if ((unsigned) num_sccs < partitions->length ()) | |
2182 | { | |
2183 | for (i = 0; i < num_sccs; ++i) | |
2184 | { | |
2185 | for (j = 0; partitions->iterate (j, &partition1); ++j) | |
2186 | if (pg->vertices[j].component == i) | |
2187 | break; | |
2188 | for (j = j + 1; partitions->iterate (j, &partition2); ++j) | |
2189 | if (pg->vertices[j].component == i) | |
2190 | { | |
2191 | partition_merge_into (NULL, partition1, | |
2192 | partition2, FUSE_SAME_SCC); | |
2193 | partition1->type = PTYPE_SEQUENTIAL; | |
2194 | (*partitions)[j] = NULL; | |
2195 | partition_free (partition2); | |
2196 | data = (struct pg_vdata *)pg->vertices[j].data; | |
2197 | data->partition = NULL; | |
2198 | } | |
2199 | } | |
f562e2ea | 2200 | } |
1df7b42b | 2201 | |
2202 | sort_partitions_by_post_order (pg, partitions); | |
f562e2ea | 2203 | gcc_assert (partitions->length () == (unsigned)num_sccs); |
2204 | free_partition_graph_vdata (pg); | |
2205 | free_graph (pg); | |
2206 | } | |
2207 | ||
2208 | /* Callback function for traversing edge E in graph G. DATA is private | |
2209 | callback data. */ | |
2210 | ||
2211 | static void | |
2212 | pg_collect_alias_ddrs (struct graph *g, struct graph_edge *e, void *data) | |
2213 | { | |
2214 | int i, j, component; | |
2215 | struct pg_edge_callback_data *cbdata; | |
2216 | struct pg_edata *edata = (struct pg_edata *) e->data; | |
2217 | ||
2218 | /* If the edge doesn't have attached data dependence, it represents | |
2219 | compilation time known dependences. This type dependence cannot | |
2220 | be resolved by runtime alias check. */ | |
2221 | if (edata == NULL || edata->alias_ddrs.length () == 0) | |
2222 | return; | |
2223 | ||
2224 | cbdata = (struct pg_edge_callback_data *) data; | |
2225 | i = e->src; | |
2226 | j = e->dest; | |
2227 | component = cbdata->vertices_component[i]; | |
2228 | /* Vertices are topologically sorted according to compilation time | |
2229 | known dependences, so we can break strong connected components | |
2230 | by removing edges of the opposite direction, i.e, edges pointing | |
2231 | from vertice with smaller post number to vertice with bigger post | |
2232 | number. */ | |
2233 | if (g->vertices[i].post < g->vertices[j].post | |
2234 | /* We only need to remove edges connecting vertices in the same | |
2235 | strong connected component to break it. */ | |
2236 | && component == cbdata->vertices_component[j] | |
2237 | /* Check if we want to break the strong connected component or not. */ | |
2238 | && !bitmap_bit_p (cbdata->sccs_to_merge, component)) | |
2239 | cbdata->alias_ddrs->safe_splice (edata->alias_ddrs); | |
2240 | } | |
2241 | ||
2242 | /* This is the main function breaking strong conected components in | |
2243 | PARTITIONS giving reduced depdendence graph RDG. Store data dependence | |
2244 | relations for runtime alias check in ALIAS_DDRS. */ | |
2245 | ||
2246 | static void | |
2247 | break_alias_scc_partitions (struct graph *rdg, | |
2248 | vec<struct partition *> *partitions, | |
2249 | vec<ddr_p> *alias_ddrs) | |
2250 | { | |
cb072485 | 2251 | int i, j, k, num_sccs, num_sccs_no_alias; |
f562e2ea | 2252 | /* Build partition dependence graph. */ |
2253 | graph *pg = build_partition_graph (rdg, partitions, false); | |
2254 | ||
2255 | alias_ddrs->truncate (0); | |
2256 | /* Find strong connected components in the graph, with all dependence edges | |
2257 | considered. */ | |
2258 | num_sccs = graphds_scc (pg, NULL); | |
2259 | /* All SCCs now can be broken by runtime alias checks because SCCs caused by | |
2260 | compilation time known dependences are merged before this function. */ | |
2261 | if ((unsigned) num_sccs < partitions->length ()) | |
2262 | { | |
2263 | struct pg_edge_callback_data cbdata; | |
2264 | auto_bitmap sccs_to_merge; | |
2265 | auto_vec<enum partition_type> scc_types; | |
2266 | struct partition *partition, *first; | |
2267 | ||
85676b62 | 2268 | /* If all partitions in a SCC have the same type, we can simply merge the |
f562e2ea | 2269 | SCC. This loop finds out such SCCS and record them in bitmap. */ |
2270 | bitmap_set_range (sccs_to_merge, 0, (unsigned) num_sccs); | |
2271 | for (i = 0; i < num_sccs; ++i) | |
2272 | { | |
2273 | for (j = 0; partitions->iterate (j, &first); ++j) | |
2274 | if (pg->vertices[j].component == i) | |
2275 | break; | |
a1e9c80f | 2276 | |
2277 | bool same_type = true, all_builtins = partition_builtin_p (first); | |
f562e2ea | 2278 | for (++j; partitions->iterate (j, &partition); ++j) |
2279 | { | |
2280 | if (pg->vertices[j].component != i) | |
2281 | continue; | |
2282 | ||
2283 | if (first->type != partition->type) | |
2284 | { | |
a1e9c80f | 2285 | same_type = false; |
f562e2ea | 2286 | break; |
2287 | } | |
a1e9c80f | 2288 | all_builtins &= partition_builtin_p (partition); |
f562e2ea | 2289 | } |
a1e9c80f | 2290 | /* Merge SCC if all partitions in SCC have the same type, though the |
2291 | result partition is sequential, because vectorizer can do better | |
2292 | runtime alias check. One expecption is all partitions in SCC are | |
2293 | builtins. */ | |
2294 | if (!same_type || all_builtins) | |
2295 | bitmap_clear_bit (sccs_to_merge, i); | |
f562e2ea | 2296 | } |
2297 | ||
2298 | /* Initialize callback data for traversing. */ | |
2299 | cbdata.sccs_to_merge = sccs_to_merge; | |
2300 | cbdata.alias_ddrs = alias_ddrs; | |
2301 | cbdata.vertices_component = XNEWVEC (int, pg->n_vertices); | |
2302 | /* Record the component information which will be corrupted by next | |
2303 | graph scc finding call. */ | |
2304 | for (i = 0; i < pg->n_vertices; ++i) | |
2305 | cbdata.vertices_component[i] = pg->vertices[i].component; | |
2306 | ||
2307 | /* Collect data dependences for runtime alias checks to break SCCs. */ | |
2308 | if (bitmap_count_bits (sccs_to_merge) != (unsigned) num_sccs) | |
2309 | { | |
2310 | /* Run SCC finding algorithm again, with alias dependence edges | |
85676b62 | 2311 | skipped. This is to topologically sort partitions according to |
f562e2ea | 2312 | compilation time known dependence. Note the topological order |
2313 | is stored in the form of pg's post order number. */ | |
2314 | num_sccs_no_alias = graphds_scc (pg, NULL, pg_skip_alias_edge); | |
2315 | gcc_assert (partitions->length () == (unsigned) num_sccs_no_alias); | |
2316 | /* With topological order, we can construct two subgraphs L and R. | |
2317 | L contains edge <x, y> where x < y in terms of post order, while | |
2318 | R contains edge <x, y> where x > y. Edges for compilation time | |
2319 | known dependence all fall in R, so we break SCCs by removing all | |
2320 | (alias) edges of in subgraph L. */ | |
2321 | for_each_edge (pg, pg_collect_alias_ddrs, &cbdata); | |
2322 | } | |
2323 | ||
2324 | /* For SCC that doesn't need to be broken, merge it. */ | |
2325 | for (i = 0; i < num_sccs; ++i) | |
2326 | { | |
2327 | if (!bitmap_bit_p (sccs_to_merge, i)) | |
2328 | continue; | |
2329 | ||
2330 | for (j = 0; partitions->iterate (j, &first); ++j) | |
2331 | if (cbdata.vertices_component[j] == i) | |
2332 | break; | |
cb072485 | 2333 | for (k = j + 1; partitions->iterate (k, &partition); ++k) |
f562e2ea | 2334 | { |
2335 | struct pg_vdata *data; | |
2336 | ||
cb072485 | 2337 | if (cbdata.vertices_component[k] != i) |
f562e2ea | 2338 | continue; |
2339 | ||
cb072485 | 2340 | /* Update postorder number so that merged reduction partition is |
2341 | sorted after other partitions. */ | |
2342 | if (!partition_reduction_p (first) | |
2343 | && partition_reduction_p (partition)) | |
2344 | { | |
2345 | gcc_assert (pg->vertices[k].post < pg->vertices[j].post); | |
2346 | pg->vertices[j].post = pg->vertices[k].post; | |
2347 | } | |
f562e2ea | 2348 | partition_merge_into (NULL, first, partition, FUSE_SAME_SCC); |
cb072485 | 2349 | (*partitions)[k] = NULL; |
f562e2ea | 2350 | partition_free (partition); |
cb072485 | 2351 | data = (struct pg_vdata *)pg->vertices[k].data; |
2352 | gcc_assert (data->id == k); | |
f562e2ea | 2353 | data->partition = NULL; |
85676b62 | 2354 | /* The result partition of merged SCC must be sequential. */ |
2355 | first->type = PTYPE_SEQUENTIAL; | |
f562e2ea | 2356 | } |
2357 | } | |
2358 | } | |
2359 | ||
2360 | sort_partitions_by_post_order (pg, partitions); | |
2361 | free_partition_graph_vdata (pg); | |
2362 | for_each_edge (pg, free_partition_graph_edata_cb, NULL); | |
2363 | free_graph (pg); | |
2364 | ||
2365 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
2366 | { | |
2367 | fprintf (dump_file, "Possible alias data dependence to break:\n"); | |
2368 | dump_data_dependence_relations (dump_file, *alias_ddrs); | |
2369 | } | |
2370 | } | |
2371 | ||
2372 | /* Compute and return an expression whose value is the segment length which | |
2373 | will be accessed by DR in NITERS iterations. */ | |
2374 | ||
2375 | static tree | |
2376 | data_ref_segment_size (struct data_reference *dr, tree niters) | |
2377 | { | |
e85b4a5e | 2378 | niters = size_binop (MINUS_EXPR, |
2379 | fold_convert (sizetype, niters), | |
2380 | size_one_node); | |
2381 | return size_binop (MULT_EXPR, | |
2382 | fold_convert (sizetype, DR_STEP (dr)), | |
2383 | fold_convert (sizetype, niters)); | |
f562e2ea | 2384 | } |
2385 | ||
2386 | /* Return true if LOOP's latch is dominated by statement for data reference | |
2387 | DR. */ | |
2388 | ||
2389 | static inline bool | |
2390 | latch_dominated_by_data_ref (struct loop *loop, data_reference *dr) | |
2391 | { | |
2392 | return dominated_by_p (CDI_DOMINATORS, single_exit (loop)->src, | |
2393 | gimple_bb (DR_STMT (dr))); | |
2394 | } | |
2395 | ||
2396 | /* Compute alias check pairs and store them in COMP_ALIAS_PAIRS for LOOP's | |
2397 | data dependence relations ALIAS_DDRS. */ | |
2398 | ||
2399 | static void | |
2400 | compute_alias_check_pairs (struct loop *loop, vec<ddr_p> *alias_ddrs, | |
2401 | vec<dr_with_seg_len_pair_t> *comp_alias_pairs) | |
2402 | { | |
2403 | unsigned int i; | |
2404 | unsigned HOST_WIDE_INT factor = 1; | |
2405 | tree niters_plus_one, niters = number_of_latch_executions (loop); | |
2406 | ||
2407 | gcc_assert (niters != NULL_TREE && niters != chrec_dont_know); | |
2408 | niters = fold_convert (sizetype, niters); | |
2409 | niters_plus_one = size_binop (PLUS_EXPR, niters, size_one_node); | |
2410 | ||
2411 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
2412 | fprintf (dump_file, "Creating alias check pairs:\n"); | |
2413 | ||
2414 | /* Iterate all data dependence relations and compute alias check pairs. */ | |
2415 | for (i = 0; i < alias_ddrs->length (); i++) | |
2416 | { | |
2417 | ddr_p ddr = (*alias_ddrs)[i]; | |
2418 | struct data_reference *dr_a = DDR_A (ddr); | |
2419 | struct data_reference *dr_b = DDR_B (ddr); | |
2420 | tree seg_length_a, seg_length_b; | |
2421 | int comp_res = data_ref_compare_tree (DR_BASE_ADDRESS (dr_a), | |
2422 | DR_BASE_ADDRESS (dr_b)); | |
2423 | ||
2424 | if (comp_res == 0) | |
2425 | comp_res = data_ref_compare_tree (DR_OFFSET (dr_a), DR_OFFSET (dr_b)); | |
2426 | gcc_assert (comp_res != 0); | |
2427 | ||
2428 | if (latch_dominated_by_data_ref (loop, dr_a)) | |
2429 | seg_length_a = data_ref_segment_size (dr_a, niters_plus_one); | |
2430 | else | |
2431 | seg_length_a = data_ref_segment_size (dr_a, niters); | |
2432 | ||
2433 | if (latch_dominated_by_data_ref (loop, dr_b)) | |
2434 | seg_length_b = data_ref_segment_size (dr_b, niters_plus_one); | |
2435 | else | |
2436 | seg_length_b = data_ref_segment_size (dr_b, niters); | |
2437 | ||
e85b4a5e | 2438 | unsigned HOST_WIDE_INT access_size_a |
2439 | = tree_to_uhwi (TYPE_SIZE_UNIT (TREE_TYPE (DR_REF (dr_a)))); | |
2440 | unsigned HOST_WIDE_INT access_size_b | |
2441 | = tree_to_uhwi (TYPE_SIZE_UNIT (TREE_TYPE (DR_REF (dr_b)))); | |
2442 | unsigned int align_a = TYPE_ALIGN_UNIT (TREE_TYPE (DR_REF (dr_a))); | |
2443 | unsigned int align_b = TYPE_ALIGN_UNIT (TREE_TYPE (DR_REF (dr_b))); | |
2444 | ||
f562e2ea | 2445 | dr_with_seg_len_pair_t dr_with_seg_len_pair |
e85b4a5e | 2446 | (dr_with_seg_len (dr_a, seg_length_a, access_size_a, align_a), |
2447 | dr_with_seg_len (dr_b, seg_length_b, access_size_b, align_b)); | |
f562e2ea | 2448 | |
2449 | /* Canonicalize pairs by sorting the two DR members. */ | |
2450 | if (comp_res > 0) | |
2451 | std::swap (dr_with_seg_len_pair.first, dr_with_seg_len_pair.second); | |
2452 | ||
2453 | comp_alias_pairs->safe_push (dr_with_seg_len_pair); | |
2454 | } | |
2455 | ||
2456 | if (tree_fits_uhwi_p (niters)) | |
2457 | factor = tree_to_uhwi (niters); | |
2458 | ||
2459 | /* Prune alias check pairs. */ | |
2460 | prune_runtime_alias_test_list (comp_alias_pairs, factor); | |
2461 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
2462 | fprintf (dump_file, | |
2463 | "Improved number of alias checks from %d to %d\n", | |
2464 | alias_ddrs->length (), comp_alias_pairs->length ()); | |
2465 | } | |
2466 | ||
2467 | /* Given data dependence relations in ALIAS_DDRS, generate runtime alias | |
2468 | checks and version LOOP under condition of these runtime alias checks. */ | |
2469 | ||
2470 | static void | |
a1e9c80f | 2471 | version_loop_by_alias_check (vec<struct partition *> *partitions, |
2472 | struct loop *loop, vec<ddr_p> *alias_ddrs) | |
f562e2ea | 2473 | { |
2474 | profile_probability prob; | |
2475 | basic_block cond_bb; | |
2476 | struct loop *nloop; | |
2477 | tree lhs, arg0, cond_expr = NULL_TREE; | |
2478 | gimple_seq cond_stmts = NULL; | |
2479 | gimple *call_stmt = NULL; | |
2480 | auto_vec<dr_with_seg_len_pair_t> comp_alias_pairs; | |
2481 | ||
2482 | /* Generate code for runtime alias checks if necessary. */ | |
2483 | gcc_assert (alias_ddrs->length () > 0); | |
2484 | ||
2485 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
2486 | fprintf (dump_file, | |
2487 | "Version loop <%d> with runtime alias check\n", loop->num); | |
2488 | ||
2489 | compute_alias_check_pairs (loop, alias_ddrs, &comp_alias_pairs); | |
2490 | create_runtime_alias_checks (loop, &comp_alias_pairs, &cond_expr); | |
2491 | cond_expr = force_gimple_operand_1 (cond_expr, &cond_stmts, | |
b48c230a | 2492 | is_gimple_val, NULL_TREE); |
f562e2ea | 2493 | |
2494 | /* Depend on vectorizer to fold IFN_LOOP_DIST_ALIAS. */ | |
a1e9c80f | 2495 | bool cancelable_p = flag_tree_loop_vectorize; |
2496 | if (cancelable_p) | |
2497 | { | |
2498 | unsigned i = 0; | |
2499 | struct partition *partition; | |
2500 | for (; partitions->iterate (i, &partition); ++i) | |
2501 | if (!partition_builtin_p (partition)) | |
2502 | break; | |
2503 | ||
2504 | /* If all partitions are builtins, distributing it would be profitable and | |
2505 | we don't want to cancel the runtime alias checks. */ | |
2506 | if (i == partitions->length ()) | |
2507 | cancelable_p = false; | |
2508 | } | |
2509 | ||
2510 | /* Generate internal function call for loop distribution alias check if the | |
2511 | runtime alias check should be cancelable. */ | |
2512 | if (cancelable_p) | |
f562e2ea | 2513 | { |
f562e2ea | 2514 | call_stmt = gimple_build_call_internal (IFN_LOOP_DIST_ALIAS, |
2515 | 2, NULL_TREE, cond_expr); | |
2516 | lhs = make_ssa_name (boolean_type_node); | |
2517 | gimple_call_set_lhs (call_stmt, lhs); | |
2518 | } | |
2519 | else | |
2520 | lhs = cond_expr; | |
2521 | ||
2522 | prob = profile_probability::guessed_always ().apply_scale (9, 10); | |
2523 | initialize_original_copy_tables (); | |
2524 | nloop = loop_version (loop, lhs, &cond_bb, prob, prob.invert (), | |
2525 | prob, prob.invert (), true); | |
2526 | free_original_copy_tables (); | |
2527 | /* Record the original loop number in newly generated loops. In case of | |
2528 | distribution, the original loop will be distributed and the new loop | |
2529 | is kept. */ | |
2530 | loop->orig_loop_num = nloop->num; | |
2531 | nloop->orig_loop_num = nloop->num; | |
2532 | nloop->dont_vectorize = true; | |
2533 | nloop->force_vectorize = false; | |
2534 | ||
2535 | if (call_stmt) | |
2536 | { | |
2537 | /* Record new loop's num in IFN_LOOP_DIST_ALIAS because the original | |
2538 | loop could be destroyed. */ | |
2539 | arg0 = build_int_cst (integer_type_node, loop->orig_loop_num); | |
2540 | gimple_call_set_arg (call_stmt, 0, arg0); | |
2541 | gimple_seq_add_stmt_without_update (&cond_stmts, call_stmt); | |
2542 | } | |
2543 | ||
2544 | if (cond_stmts) | |
2545 | { | |
2546 | gimple_stmt_iterator cond_gsi = gsi_last_bb (cond_bb); | |
2547 | gsi_insert_seq_before (&cond_gsi, cond_stmts, GSI_SAME_STMT); | |
2548 | } | |
2549 | update_ssa (TODO_update_ssa); | |
2550 | } | |
2551 | ||
2552 | /* Return true if loop versioning is needed to distrubute PARTITIONS. | |
2553 | ALIAS_DDRS are data dependence relations for runtime alias check. */ | |
2554 | ||
2555 | static inline bool | |
2556 | version_for_distribution_p (vec<struct partition *> *partitions, | |
2557 | vec<ddr_p> *alias_ddrs) | |
2558 | { | |
2559 | /* No need to version loop if we have only one partition. */ | |
2560 | if (partitions->length () == 1) | |
2561 | return false; | |
2562 | ||
2563 | /* Need to version loop if runtime alias check is necessary. */ | |
2564 | return (alias_ddrs->length () > 0); | |
2565 | } | |
2566 | ||
05ebeee6 | 2567 | /* Compare base offset of builtin mem* partitions P1 and P2. */ |
2568 | ||
207c43a2 | 2569 | static int |
2570 | offset_cmp (const void *vp1, const void *vp2) | |
05ebeee6 | 2571 | { |
207c43a2 | 2572 | struct partition *p1 = *(struct partition *const *) vp1; |
2573 | struct partition *p2 = *(struct partition *const *) vp2; | |
2574 | unsigned HOST_WIDE_INT o1 = p1->builtin->dst_base_offset; | |
2575 | unsigned HOST_WIDE_INT o2 = p2->builtin->dst_base_offset; | |
2576 | return (o2 < o1) - (o1 < o2); | |
05ebeee6 | 2577 | } |
2578 | ||
2579 | /* Fuse adjacent memset builtin PARTITIONS if possible. This is a special | |
2580 | case optimization transforming below code: | |
2581 | ||
2582 | __builtin_memset (&obj, 0, 100); | |
2583 | _1 = &obj + 100; | |
2584 | __builtin_memset (_1, 0, 200); | |
2585 | _2 = &obj + 300; | |
2586 | __builtin_memset (_2, 0, 100); | |
2587 | ||
2588 | into: | |
2589 | ||
2590 | __builtin_memset (&obj, 0, 400); | |
2591 | ||
2592 | Note we don't have dependence information between different partitions | |
2593 | at this point, as a result, we can't handle nonadjacent memset builtin | |
2594 | partitions since dependence might be broken. */ | |
2595 | ||
2596 | static void | |
2597 | fuse_memset_builtins (vec<struct partition *> *partitions) | |
2598 | { | |
2599 | unsigned i, j; | |
2600 | struct partition *part1, *part2; | |
79f9738a | 2601 | tree rhs1, rhs2; |
05ebeee6 | 2602 | |
2603 | for (i = 0; partitions->iterate (i, &part1);) | |
2604 | { | |
2605 | if (part1->kind != PKIND_MEMSET) | |
2606 | { | |
2607 | i++; | |
2608 | continue; | |
2609 | } | |
2610 | ||
2611 | /* Find sub-array of memset builtins of the same base. Index range | |
2612 | of the sub-array is [i, j) with "j > i". */ | |
2613 | for (j = i + 1; partitions->iterate (j, &part2); ++j) | |
2614 | { | |
2615 | if (part2->kind != PKIND_MEMSET | |
2616 | || !operand_equal_p (part1->builtin->dst_base_base, | |
2617 | part2->builtin->dst_base_base, 0)) | |
2618 | break; | |
79f9738a | 2619 | |
2620 | /* Memset calls setting different values can't be merged. */ | |
2621 | rhs1 = gimple_assign_rhs1 (DR_STMT (part1->builtin->dst_dr)); | |
2622 | rhs2 = gimple_assign_rhs1 (DR_STMT (part2->builtin->dst_dr)); | |
2623 | if (!operand_equal_p (rhs1, rhs2, 0)) | |
2624 | break; | |
05ebeee6 | 2625 | } |
2626 | ||
2627 | /* Stable sort is required in order to avoid breaking dependence. */ | |
207c43a2 | 2628 | gcc_stablesort (&(*partitions)[i], j - i, sizeof (*partitions)[i], |
2629 | offset_cmp); | |
05ebeee6 | 2630 | /* Continue with next partition. */ |
2631 | i = j; | |
2632 | } | |
2633 | ||
2634 | /* Merge all consecutive memset builtin partitions. */ | |
2635 | for (i = 0; i < partitions->length () - 1;) | |
2636 | { | |
2637 | part1 = (*partitions)[i]; | |
2638 | if (part1->kind != PKIND_MEMSET) | |
2639 | { | |
2640 | i++; | |
2641 | continue; | |
2642 | } | |
2643 | ||
2644 | part2 = (*partitions)[i + 1]; | |
2645 | /* Only merge memset partitions of the same base and with constant | |
2646 | access sizes. */ | |
2647 | if (part2->kind != PKIND_MEMSET | |
2648 | || TREE_CODE (part1->builtin->size) != INTEGER_CST | |
2649 | || TREE_CODE (part2->builtin->size) != INTEGER_CST | |
2650 | || !operand_equal_p (part1->builtin->dst_base_base, | |
2651 | part2->builtin->dst_base_base, 0)) | |
2652 | { | |
2653 | i++; | |
2654 | continue; | |
2655 | } | |
79f9738a | 2656 | rhs1 = gimple_assign_rhs1 (DR_STMT (part1->builtin->dst_dr)); |
2657 | rhs2 = gimple_assign_rhs1 (DR_STMT (part2->builtin->dst_dr)); | |
05ebeee6 | 2658 | int bytev1 = const_with_all_bytes_same (rhs1); |
2659 | int bytev2 = const_with_all_bytes_same (rhs2); | |
2660 | /* Only merge memset partitions of the same value. */ | |
2661 | if (bytev1 != bytev2 || bytev1 == -1) | |
2662 | { | |
2663 | i++; | |
2664 | continue; | |
2665 | } | |
2666 | wide_int end1 = wi::add (part1->builtin->dst_base_offset, | |
2667 | wi::to_wide (part1->builtin->size)); | |
2668 | /* Only merge adjacent memset partitions. */ | |
2669 | if (wi::ne_p (end1, part2->builtin->dst_base_offset)) | |
2670 | { | |
2671 | i++; | |
2672 | continue; | |
2673 | } | |
2674 | /* Merge partitions[i] and partitions[i+1]. */ | |
2675 | part1->builtin->size = fold_build2 (PLUS_EXPR, sizetype, | |
2676 | part1->builtin->size, | |
2677 | part2->builtin->size); | |
2678 | partition_free (part2); | |
2679 | partitions->ordered_remove (i + 1); | |
2680 | } | |
2681 | } | |
2682 | ||
883b4905 | 2683 | /* Fuse PARTITIONS of LOOP if necessary before finalizing distribution. |
2684 | ALIAS_DDRS contains ddrs which need runtime alias check. */ | |
f562e2ea | 2685 | |
2686 | static void | |
883b4905 | 2687 | finalize_partitions (struct loop *loop, vec<struct partition *> *partitions, |
f562e2ea | 2688 | vec<ddr_p> *alias_ddrs) |
2689 | { | |
2690 | unsigned i; | |
883b4905 | 2691 | struct partition *partition, *a; |
f562e2ea | 2692 | |
2693 | if (partitions->length () == 1 | |
2694 | || alias_ddrs->length () > 0) | |
2695 | return; | |
2696 | ||
b1c062d8 | 2697 | unsigned num_builtin = 0, num_normal = 0, num_partial_memset = 0; |
883b4905 | 2698 | bool same_type_p = true; |
2699 | enum partition_type type = ((*partitions)[0])->type; | |
2700 | for (i = 0; partitions->iterate (i, &partition); ++i) | |
f562e2ea | 2701 | { |
883b4905 | 2702 | same_type_p &= (type == partition->type); |
b1c062d8 | 2703 | if (partition_builtin_p (partition)) |
2704 | { | |
2705 | num_builtin++; | |
2706 | continue; | |
2707 | } | |
2708 | num_normal++; | |
2709 | if (partition->kind == PKIND_PARTIAL_MEMSET) | |
2710 | num_partial_memset++; | |
f562e2ea | 2711 | } |
2712 | ||
883b4905 | 2713 | /* Don't distribute current loop into too many loops given we don't have |
2714 | memory stream cost model. Be even more conservative in case of loop | |
2715 | nest distribution. */ | |
b1c062d8 | 2716 | if ((same_type_p && num_builtin == 0 |
2717 | && (loop->inner == NULL || num_normal != 2 || num_partial_memset != 1)) | |
883b4905 | 2718 | || (loop->inner != NULL |
2719 | && i >= NUM_PARTITION_THRESHOLD && num_normal > 1) | |
2720 | || (loop->inner == NULL | |
2721 | && i >= NUM_PARTITION_THRESHOLD && num_normal > num_builtin)) | |
f562e2ea | 2722 | { |
883b4905 | 2723 | a = (*partitions)[0]; |
2724 | for (i = 1; partitions->iterate (i, &partition); ++i) | |
2725 | { | |
2726 | partition_merge_into (NULL, a, partition, FUSE_FINALIZE); | |
2727 | partition_free (partition); | |
2728 | } | |
2729 | partitions->truncate (1); | |
f562e2ea | 2730 | } |
05ebeee6 | 2731 | |
2732 | /* Fuse memset builtins if possible. */ | |
2733 | if (partitions->length () > 1) | |
2734 | fuse_memset_builtins (partitions); | |
f562e2ea | 2735 | } |
2736 | ||
2737 | /* Distributes the code from LOOP in such a way that producer statements | |
2738 | are placed before consumer statements. Tries to separate only the | |
2739 | statements from STMTS into separate loops. Returns the number of | |
2740 | distributed loops. Set NB_CALLS to number of generated builtin calls. | |
2741 | Set *DESTROY_P to whether LOOP needs to be destroyed. */ | |
801c5610 | 2742 | |
2743 | static int | |
42acab1c | 2744 | distribute_loop (struct loop *loop, vec<gimple *> stmts, |
77d095c0 | 2745 | control_dependences *cd, int *nb_calls, bool *destroy_p) |
801c5610 | 2746 | { |
bbb229ef | 2747 | ddrs_table = new hash_table<ddr_hasher> (389); |
15c8650d | 2748 | struct graph *rdg; |
04009ada | 2749 | partition *partition; |
6198d968 | 2750 | bool any_builtin; |
15c8650d | 2751 | int i, nbp; |
801c5610 | 2752 | |
303e9267 | 2753 | *destroy_p = false; |
df9892ff | 2754 | *nb_calls = 0; |
209a62a6 | 2755 | loop_nest.create (0); |
15c8650d | 2756 | if (!find_loop_nest (loop, &loop_nest)) |
209a62a6 | 2757 | { |
2758 | loop_nest.release (); | |
bbb229ef | 2759 | delete ddrs_table; |
209a62a6 | 2760 | return 0; |
2761 | } | |
15c8650d | 2762 | |
f3754041 | 2763 | datarefs_vec.create (20); |
6079e9be | 2764 | has_nonaddressable_dataref_p = false; |
209a62a6 | 2765 | rdg = build_rdg (loop, cd); |
15c8650d | 2766 | if (!rdg) |
2767 | { | |
2768 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
2769 | fprintf (dump_file, | |
2770 | "Loop %d not distributed: failed to build the RDG.\n", | |
2771 | loop->num); | |
2772 | ||
209a62a6 | 2773 | loop_nest.release (); |
f3754041 | 2774 | free_data_refs (datarefs_vec); |
bbb229ef | 2775 | delete ddrs_table; |
f3754041 | 2776 | return 0; |
2777 | } | |
2778 | ||
2779 | if (datarefs_vec.length () > MAX_DATAREFS_NUM) | |
2780 | { | |
2781 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
2782 | fprintf (dump_file, | |
2783 | "Loop %d not distributed: too many memory references.\n", | |
2784 | loop->num); | |
2785 | ||
2786 | free_rdg (rdg); | |
2787 | loop_nest.release (); | |
2788 | free_data_refs (datarefs_vec); | |
bbb229ef | 2789 | delete ddrs_table; |
15c8650d | 2790 | return 0; |
2791 | } | |
2792 | ||
f3754041 | 2793 | data_reference_p dref; |
2794 | for (i = 0; datarefs_vec.iterate (i, &dref); ++i) | |
2795 | dref->aux = (void *) (uintptr_t) i; | |
2796 | ||
15c8650d | 2797 | if (dump_file && (dump_flags & TDF_DETAILS)) |
2798 | dump_rdg (dump_file, rdg); | |
2799 | ||
04009ada | 2800 | auto_vec<struct partition *, 3> partitions; |
15c8650d | 2801 | rdg_build_partitions (rdg, stmts, &partitions); |
801c5610 | 2802 | |
f562e2ea | 2803 | auto_vec<ddr_p> alias_ddrs; |
2804 | ||
8d95fe31 | 2805 | auto_bitmap stmt_in_all_partitions; |
2806 | bitmap_copy (stmt_in_all_partitions, partitions[0]->stmts); | |
2807 | for (i = 1; partitions.iterate (i, &partition); ++i) | |
2808 | bitmap_and_into (stmt_in_all_partitions, partitions[i]->stmts); | |
2809 | ||
6198d968 | 2810 | any_builtin = false; |
f1f41a6c | 2811 | FOR_EACH_VEC_ELT (partitions, i, partition) |
6198d968 | 2812 | { |
8d95fe31 | 2813 | classify_partition (loop, rdg, partition, stmt_in_all_partitions); |
6198d968 | 2814 | any_builtin |= partition_builtin_p (partition); |
2815 | } | |
d32bc1d7 | 2816 | |
7103facc | 2817 | /* If we are only distributing patterns but did not detect any, |
2818 | simply bail out. */ | |
0c58611e | 2819 | if (!flag_tree_loop_distribution |
2820 | && !any_builtin) | |
2821 | { | |
2822 | nbp = 0; | |
2823 | goto ldist_done; | |
2824 | } | |
2825 | ||
7103facc | 2826 | /* If we are only distributing patterns fuse all partitions that |
2827 | were not classified as builtins. This also avoids chopping | |
2828 | a loop into pieces, separated by builtin calls. That is, we | |
2829 | only want no or a single loop body remaining. */ | |
04009ada | 2830 | struct partition *into; |
7103facc | 2831 | if (!flag_tree_loop_distribution) |
2832 | { | |
2833 | for (i = 0; partitions.iterate (i, &into); ++i) | |
2834 | if (!partition_builtin_p (into)) | |
2835 | break; | |
2836 | for (++i; partitions.iterate (i, &partition); ++i) | |
2837 | if (!partition_builtin_p (partition)) | |
2838 | { | |
f024aa04 | 2839 | partition_merge_into (NULL, into, partition, FUSE_NON_BUILTIN); |
7103facc | 2840 | partitions.unordered_remove (i); |
2841 | partition_free (partition); | |
2842 | i--; | |
2843 | } | |
2844 | } | |
2845 | ||
2846 | /* Due to limitations in the transform phase we have to fuse all | |
2847 | reduction partitions into the last partition so the existing | |
2848 | loop will contain all loop-closed PHI nodes. */ | |
2849 | for (i = 0; partitions.iterate (i, &into); ++i) | |
2850 | if (partition_reduction_p (into)) | |
2851 | break; | |
2852 | for (i = i + 1; partitions.iterate (i, &partition); ++i) | |
2853 | if (partition_reduction_p (partition)) | |
2854 | { | |
f024aa04 | 2855 | partition_merge_into (rdg, into, partition, FUSE_REDUCTION); |
7103facc | 2856 | partitions.unordered_remove (i); |
2857 | partition_free (partition); | |
2858 | i--; | |
2859 | } | |
2860 | ||
0c58611e | 2861 | /* Apply our simple cost model - fuse partitions with similar |
2862 | memory accesses. */ | |
0c58611e | 2863 | for (i = 0; partitions.iterate (i, &into); ++i) |
2864 | { | |
facf7bb7 | 2865 | bool changed = false; |
b1c062d8 | 2866 | if (partition_builtin_p (into) || into->kind == PKIND_PARTIAL_MEMSET) |
0c58611e | 2867 | continue; |
2868 | for (int j = i + 1; | |
2869 | partitions.iterate (j, &partition); ++j) | |
2870 | { | |
fd34627b | 2871 | if (share_memory_accesses (rdg, into, partition)) |
0c58611e | 2872 | { |
f024aa04 | 2873 | partition_merge_into (rdg, into, partition, FUSE_SHARE_REF); |
7103facc | 2874 | partitions.unordered_remove (j); |
0c58611e | 2875 | partition_free (partition); |
2876 | j--; | |
facf7bb7 | 2877 | changed = true; |
0c58611e | 2878 | } |
2879 | } | |
facf7bb7 | 2880 | /* If we fused 0 1 2 in step 1 to 0,2 1 as 0 and 2 have similar |
2881 | accesses when 1 and 2 have similar accesses but not 0 and 1 | |
2882 | then in the next iteration we will fail to consider merging | |
2883 | 1 into 0,2. So try again if we did any merging into 0. */ | |
2884 | if (changed) | |
2885 | i--; | |
0c58611e | 2886 | } |
2887 | ||
883b4905 | 2888 | /* Build the partition dependency graph and fuse partitions in strong |
2889 | connected component. */ | |
7103facc | 2890 | if (partitions.length () > 1) |
54459dd6 | 2891 | { |
883b4905 | 2892 | /* Don't support loop nest distribution under runtime alias check |
6079e9be | 2893 | since it's not likely to enable many vectorization opportunities. |
2894 | Also if loop has any data reference which may be not addressable | |
2895 | since alias check needs to take, compare address of the object. */ | |
2896 | if (loop->inner || has_nonaddressable_dataref_p) | |
883b4905 | 2897 | merge_dep_scc_partitions (rdg, &partitions, false); |
2898 | else | |
2899 | { | |
2900 | merge_dep_scc_partitions (rdg, &partitions, true); | |
2901 | if (partitions.length () > 1) | |
2902 | break_alias_scc_partitions (rdg, &partitions, &alias_ddrs); | |
2903 | } | |
ac7a1007 | 2904 | } |
2905 | ||
883b4905 | 2906 | finalize_partitions (loop, &partitions, &alias_ddrs); |
f562e2ea | 2907 | |
f1f41a6c | 2908 | nbp = partitions.length (); |
58ccfbea | 2909 | if (nbp == 0 |
f1f41a6c | 2910 | || (nbp == 1 && !partition_builtin_p (partitions[0])) |
2911 | || (nbp > 1 && partition_contains_all_rw (rdg, partitions))) | |
54459dd6 | 2912 | { |
2913 | nbp = 0; | |
2914 | goto ldist_done; | |
2915 | } | |
801c5610 | 2916 | |
f562e2ea | 2917 | if (version_for_distribution_p (&partitions, &alias_ddrs)) |
a1e9c80f | 2918 | version_loop_by_alias_check (&partitions, loop, &alias_ddrs); |
f562e2ea | 2919 | |
801c5610 | 2920 | if (dump_file && (dump_flags & TDF_DETAILS)) |
f562e2ea | 2921 | { |
2922 | fprintf (dump_file, | |
2923 | "distribute loop <%d> into partitions:\n", loop->num); | |
2924 | dump_rdg_partitions (dump_file, partitions); | |
2925 | } | |
801c5610 | 2926 | |
f1f41a6c | 2927 | FOR_EACH_VEC_ELT (partitions, i, partition) |
df9892ff | 2928 | { |
2929 | if (partition_builtin_p (partition)) | |
2930 | (*nb_calls)++; | |
77d095c0 | 2931 | *destroy_p |= generate_code_for_partition (loop, partition, i < nbp - 1); |
df9892ff | 2932 | } |
801c5610 | 2933 | |
801c5610 | 2934 | ldist_done: |
209a62a6 | 2935 | loop_nest.release (); |
f3754041 | 2936 | free_data_refs (datarefs_vec); |
bbb229ef | 2937 | for (hash_table<ddr_hasher>::iterator iter = ddrs_table->begin (); |
2938 | iter != ddrs_table->end (); ++iter) | |
50f5937e | 2939 | { |
2940 | free_dependence_relation (*iter); | |
2941 | *iter = NULL; | |
2942 | } | |
bbb229ef | 2943 | delete ddrs_table; |
801c5610 | 2944 | |
f1f41a6c | 2945 | FOR_EACH_VEC_ELT (partitions, i, partition) |
543506e0 | 2946 | partition_free (partition); |
801c5610 | 2947 | |
801c5610 | 2948 | free_rdg (rdg); |
df9892ff | 2949 | return nbp - *nb_calls; |
801c5610 | 2950 | } |
2951 | ||
2952 | /* Distribute all loops in the current function. */ | |
2953 | ||
65b0537f | 2954 | namespace { |
2955 | ||
2956 | const pass_data pass_data_loop_distribution = | |
2957 | { | |
2958 | GIMPLE_PASS, /* type */ | |
2959 | "ldist", /* name */ | |
2960 | OPTGROUP_LOOP, /* optinfo_flags */ | |
65b0537f | 2961 | TV_TREE_LOOP_DISTRIBUTION, /* tv_id */ |
2962 | ( PROP_cfg | PROP_ssa ), /* properties_required */ | |
2963 | 0, /* properties_provided */ | |
2964 | 0, /* properties_destroyed */ | |
2965 | 0, /* todo_flags_start */ | |
8b88439e | 2966 | 0, /* todo_flags_finish */ |
65b0537f | 2967 | }; |
2968 | ||
2969 | class pass_loop_distribution : public gimple_opt_pass | |
2970 | { | |
2971 | public: | |
2972 | pass_loop_distribution (gcc::context *ctxt) | |
2973 | : gimple_opt_pass (pass_data_loop_distribution, ctxt) | |
2974 | {} | |
2975 | ||
2976 | /* opt_pass methods: */ | |
2977 | virtual bool gate (function *) | |
2978 | { | |
2979 | return flag_tree_loop_distribution | |
2980 | || flag_tree_loop_distribute_patterns; | |
2981 | } | |
2982 | ||
2983 | virtual unsigned int execute (function *); | |
2984 | ||
2985 | }; // class pass_loop_distribution | |
2986 | ||
883b4905 | 2987 | |
2988 | /* Given LOOP, this function records seed statements for distribution in | |
2989 | WORK_LIST. Return false if there is nothing for distribution. */ | |
2990 | ||
2991 | static bool | |
2992 | find_seed_stmts_for_distribution (struct loop *loop, vec<gimple *> *work_list) | |
2993 | { | |
2994 | basic_block *bbs = get_loop_body_in_dom_order (loop); | |
2995 | ||
2996 | /* Initialize the worklist with stmts we seed the partitions with. */ | |
2997 | for (unsigned i = 0; i < loop->num_nodes; ++i) | |
2998 | { | |
2999 | for (gphi_iterator gsi = gsi_start_phis (bbs[i]); | |
3000 | !gsi_end_p (gsi); gsi_next (&gsi)) | |
3001 | { | |
3002 | gphi *phi = gsi.phi (); | |
3003 | if (virtual_operand_p (gimple_phi_result (phi))) | |
3004 | continue; | |
3005 | /* Distribute stmts which have defs that are used outside of | |
3006 | the loop. */ | |
3007 | if (!stmt_has_scalar_dependences_outside_loop (loop, phi)) | |
3008 | continue; | |
3009 | work_list->safe_push (phi); | |
3010 | } | |
3011 | for (gimple_stmt_iterator gsi = gsi_start_bb (bbs[i]); | |
3012 | !gsi_end_p (gsi); gsi_next (&gsi)) | |
3013 | { | |
3014 | gimple *stmt = gsi_stmt (gsi); | |
3015 | ||
3016 | /* If there is a stmt with side-effects bail out - we | |
3017 | cannot and should not distribute this loop. */ | |
3018 | if (gimple_has_side_effects (stmt)) | |
3019 | { | |
3020 | free (bbs); | |
3021 | return false; | |
3022 | } | |
3023 | ||
3024 | /* Distribute stmts which have defs that are used outside of | |
3025 | the loop. */ | |
3026 | if (stmt_has_scalar_dependences_outside_loop (loop, stmt)) | |
3027 | ; | |
3028 | /* Otherwise only distribute stores for now. */ | |
3029 | else if (!gimple_vdef (stmt)) | |
3030 | continue; | |
3031 | ||
3032 | work_list->safe_push (stmt); | |
3033 | } | |
3034 | } | |
3035 | free (bbs); | |
3036 | return work_list->length () > 0; | |
3037 | } | |
3038 | ||
3039 | /* Given innermost LOOP, return the outermost enclosing loop that forms a | |
3040 | perfect loop nest. */ | |
3041 | ||
3042 | static struct loop * | |
3043 | prepare_perfect_loop_nest (struct loop *loop) | |
3044 | { | |
3045 | struct loop *outer = loop_outer (loop); | |
3046 | tree niters = number_of_latch_executions (loop); | |
3047 | ||
b1c062d8 | 3048 | /* TODO: We only support the innermost 3-level loop nest distribution |
883b4905 | 3049 | because of compilation time issue for now. This should be relaxed |
b1c062d8 | 3050 | in the future. Note we only allow 3-level loop nest distribution |
3051 | when parallelizing loops. */ | |
3052 | while ((loop->inner == NULL | |
3053 | || (loop->inner->inner == NULL && flag_tree_parallelize_loops > 1)) | |
883b4905 | 3054 | && loop_outer (outer) |
3055 | && outer->inner == loop && loop->next == NULL | |
3056 | && single_exit (outer) | |
3057 | && optimize_loop_for_speed_p (outer) | |
3058 | && !chrec_contains_symbols_defined_in_loop (niters, outer->num) | |
3059 | && (niters = number_of_latch_executions (outer)) != NULL_TREE | |
3060 | && niters != chrec_dont_know) | |
3061 | { | |
3062 | loop = outer; | |
3063 | outer = loop_outer (loop); | |
3064 | } | |
3065 | ||
3066 | return loop; | |
3067 | } | |
3068 | ||
65b0537f | 3069 | unsigned int |
3070 | pass_loop_distribution::execute (function *fun) | |
801c5610 | 3071 | { |
3072 | struct loop *loop; | |
54459dd6 | 3073 | bool changed = false; |
f83623cc | 3074 | basic_block bb; |
f1ce84d9 | 3075 | control_dependences *cd = NULL; |
77d095c0 | 3076 | auto_vec<loop_p> loops_to_be_destroyed; |
f83623cc | 3077 | |
c4b1b865 | 3078 | if (number_of_loops (fun) <= 1) |
3079 | return 0; | |
3080 | ||
50eda3a8 | 3081 | /* Compute topological order for basic blocks. Topological order is |
3082 | needed because data dependence is computed for data references in | |
3083 | lexicographical order. */ | |
3084 | if (bb_top_order_index == NULL) | |
3085 | { | |
5c49e6ea | 3086 | int rpo_num; |
50eda3a8 | 3087 | int *rpo = XNEWVEC (int, last_basic_block_for_fn (cfun)); |
3088 | ||
3089 | bb_top_order_index = XNEWVEC (int, last_basic_block_for_fn (cfun)); | |
5c49e6ea | 3090 | bb_top_order_index_size = last_basic_block_for_fn (cfun); |
3091 | rpo_num = pre_and_rev_post_order_compute_fn (cfun, NULL, rpo, true); | |
3092 | for (int i = 0; i < rpo_num; i++) | |
50eda3a8 | 3093 | bb_top_order_index[rpo[i]] = i; |
3094 | ||
3095 | free (rpo); | |
3096 | } | |
3097 | ||
65b0537f | 3098 | FOR_ALL_BB_FN (bb, fun) |
f83623cc | 3099 | { |
3100 | gimple_stmt_iterator gsi; | |
3101 | for (gsi = gsi_start_phis (bb); !gsi_end_p (gsi); gsi_next (&gsi)) | |
3102 | gimple_set_uid (gsi_stmt (gsi), -1); | |
3103 | for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi)) | |
3104 | gimple_set_uid (gsi_stmt (gsi), -1); | |
3105 | } | |
801c5610 | 3106 | |
54459dd6 | 3107 | /* We can at the moment only distribute non-nested loops, thus restrict |
3108 | walking to innermost loops. */ | |
f21d4d00 | 3109 | FOR_EACH_LOOP (loop, LI_ONLY_INNERMOST) |
801c5610 | 3110 | { |
883b4905 | 3111 | /* Don't distribute multiple exit edges loop, or cold loop. */ |
3112 | if (!single_exit (loop) | |
3113 | || !optimize_loop_for_speed_p (loop)) | |
5f38d9ef | 3114 | continue; |
3115 | ||
1c4ee769 | 3116 | /* Don't distribute loop if niters is unknown. */ |
3117 | tree niters = number_of_latch_executions (loop); | |
3118 | if (niters == NULL_TREE || niters == chrec_dont_know) | |
3119 | continue; | |
3120 | ||
883b4905 | 3121 | /* Get the perfect loop nest for distribution. */ |
3122 | loop = prepare_perfect_loop_nest (loop); | |
3123 | for (; loop; loop = loop->inner) | |
6198d968 | 3124 | { |
883b4905 | 3125 | auto_vec<gimple *> work_list; |
3126 | if (!find_seed_stmts_for_distribution (loop, &work_list)) | |
3127 | break; | |
54459dd6 | 3128 | |
883b4905 | 3129 | const char *str = loop->inner ? " nest" : ""; |
c309657f | 3130 | dump_user_location_t loc = find_loop_location (loop); |
f1ce84d9 | 3131 | if (!cd) |
3132 | { | |
73423315 | 3133 | calculate_dominance_info (CDI_DOMINATORS); |
f1ce84d9 | 3134 | calculate_dominance_info (CDI_POST_DOMINATORS); |
ce143ff0 | 3135 | cd = new control_dependences (); |
f1ce84d9 | 3136 | free_dominance_info (CDI_POST_DOMINATORS); |
3137 | } | |
883b4905 | 3138 | |
77d095c0 | 3139 | bool destroy_p; |
883b4905 | 3140 | int nb_generated_loops, nb_generated_calls; |
df9892ff | 3141 | nb_generated_loops = distribute_loop (loop, work_list, cd, |
77d095c0 | 3142 | &nb_generated_calls, |
3143 | &destroy_p); | |
3144 | if (destroy_p) | |
3145 | loops_to_be_destroyed.safe_push (loop); | |
54459dd6 | 3146 | |
883b4905 | 3147 | if (nb_generated_loops + nb_generated_calls > 0) |
3148 | { | |
3149 | changed = true; | |
91f42adc | 3150 | if (dump_enabled_p ()) |
3151 | dump_printf_loc (MSG_OPTIMIZED_LOCATIONS, | |
3152 | loc, "Loop%s %d distributed: split to %d loops " | |
3153 | "and %d library calls.\n", str, loop->num, | |
3154 | nb_generated_loops, nb_generated_calls); | |
883b4905 | 3155 | |
3156 | break; | |
3157 | } | |
3158 | ||
3159 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
3160 | fprintf (dump_file, "Loop%s %d not distributed.\n", str, loop->num); | |
801c5610 | 3161 | } |
801c5610 | 3162 | } |
3163 | ||
f1ce84d9 | 3164 | if (cd) |
3165 | delete cd; | |
3166 | ||
50eda3a8 | 3167 | if (bb_top_order_index != NULL) |
3168 | { | |
3169 | free (bb_top_order_index); | |
3170 | bb_top_order_index = NULL; | |
3171 | bb_top_order_index_size = 0; | |
3172 | } | |
3173 | ||
54459dd6 | 3174 | if (changed) |
3175 | { | |
ce143ff0 | 3176 | /* Destroy loop bodies that could not be reused. Do this late as we |
3177 | otherwise can end up refering to stale data in control dependences. */ | |
3178 | unsigned i; | |
3179 | FOR_EACH_VEC_ELT (loops_to_be_destroyed, i, loop) | |
50eda3a8 | 3180 | destroy_loop (loop); |
ce143ff0 | 3181 | |
505d3633 | 3182 | /* Cached scalar evolutions now may refer to wrong or non-existing |
3183 | loops. */ | |
3184 | scev_reset_htab (); | |
65b0537f | 3185 | mark_virtual_operands_for_renaming (fun); |
54459dd6 | 3186 | rewrite_into_loop_closed_ssa (NULL, TODO_update_ssa); |
3187 | } | |
3188 | ||
382ecba7 | 3189 | checking_verify_loop_structure (); |
54459dd6 | 3190 | |
ade75d0f | 3191 | return changed ? TODO_cleanup_cfg : 0; |
801c5610 | 3192 | } |
3193 | ||
cbe8bda8 | 3194 | } // anon namespace |
3195 | ||
3196 | gimple_opt_pass * | |
3197 | make_pass_loop_distribution (gcc::context *ctxt) | |
3198 | { | |
3199 | return new pass_loop_distribution (ctxt); | |
3200 | } |