]>
Commit | Line | Data |
---|---|---|
dea61d92 | 1 | /* Loop distribution. |
8d9254fc | 2 | Copyright (C) 2006-2020 Free Software Foundation, Inc. |
dea61d92 SP |
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. | |
b8698a0f | 7 | |
dea61d92 SP |
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. | |
b8698a0f | 12 | |
dea61d92 SP |
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. | |
b8698a0f | 17 | |
dea61d92 SP |
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 | ||
b8698a0f | 29 | is transformed to |
dea61d92 SP |
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 | ||
a8745cc2 BC |
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: | |
163aa51b BC |
86 | 1) We only distribute innermost two-level loop nest now. We should |
87 | extend it for arbitrary loop nests in the future. | |
a8745cc2 BC |
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. */ | |
dea61d92 SP |
91 | |
92 | #include "config.h" | |
93 | #include "system.h" | |
94 | #include "coretypes.h" | |
c7131fb2 | 95 | #include "backend.h" |
40e23961 | 96 | #include "tree.h" |
c7131fb2 | 97 | #include "gimple.h" |
957060b5 AM |
98 | #include "cfghooks.h" |
99 | #include "tree-pass.h" | |
c7131fb2 | 100 | #include "ssa.h" |
957060b5 | 101 | #include "gimple-pretty-print.h" |
c7131fb2 | 102 | #include "fold-const.h" |
60393bbc | 103 | #include "cfganal.h" |
5be5c238 | 104 | #include "gimple-iterator.h" |
18f429e2 | 105 | #include "gimplify-me.h" |
d8a2d370 | 106 | #include "stor-layout.h" |
442b4905 | 107 | #include "tree-cfg.h" |
e28030cf | 108 | #include "tree-ssa-loop-manip.h" |
957f0d8f | 109 | #include "tree-ssa-loop-ivopts.h" |
442b4905 AM |
110 | #include "tree-ssa-loop.h" |
111 | #include "tree-into-ssa.h" | |
7a300452 | 112 | #include "tree-ssa.h" |
dea61d92 | 113 | #include "cfgloop.h" |
dea61d92 | 114 | #include "tree-scalar-evolution.h" |
826a536d | 115 | #include "tree-vectorizer.h" |
4c9ed22a | 116 | #include "tree-eh.h" |
5879ab5f | 117 | #include "gimple-fold.h" |
80ab0b19 RB |
118 | |
119 | ||
9fafb14a | 120 | #define MAX_DATAREFS_NUM \ |
028d4092 | 121 | ((unsigned) param_loop_max_datarefs_for_datadeps) |
9fafb14a | 122 | |
163aa51b BC |
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 | ||
17c5cbdf BC |
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 | ||
4084ea5f | 157 | |
9fafb14a | 158 | |
9fafb14a BC |
159 | #define DR_INDEX(dr) ((uintptr_t) (dr)->aux) |
160 | ||
80ab0b19 | 161 | /* A Reduced Dependence Graph (RDG) vertex representing a statement. */ |
526ceb68 | 162 | struct rdg_vertex |
80ab0b19 RB |
163 | { |
164 | /* The statement represented by this vertex. */ | |
355fe088 | 165 | gimple *stmt; |
80ab0b19 RB |
166 | |
167 | /* Vector of data-references in this statement. */ | |
168 | vec<data_reference_p> datarefs; | |
169 | ||
170 | /* True when the statement contains a write to memory. */ | |
171 | bool has_mem_write; | |
172 | ||
173 | /* True when the statement contains a read from memory. */ | |
174 | bool has_mem_reads; | |
526ceb68 | 175 | }; |
80ab0b19 RB |
176 | |
177 | #define RDGV_STMT(V) ((struct rdg_vertex *) ((V)->data))->stmt | |
178 | #define RDGV_DATAREFS(V) ((struct rdg_vertex *) ((V)->data))->datarefs | |
179 | #define RDGV_HAS_MEM_WRITE(V) ((struct rdg_vertex *) ((V)->data))->has_mem_write | |
180 | #define RDGV_HAS_MEM_READS(V) ((struct rdg_vertex *) ((V)->data))->has_mem_reads | |
181 | #define RDG_STMT(RDG, I) RDGV_STMT (&(RDG->vertices[I])) | |
182 | #define RDG_DATAREFS(RDG, I) RDGV_DATAREFS (&(RDG->vertices[I])) | |
183 | #define RDG_MEM_WRITE_STMT(RDG, I) RDGV_HAS_MEM_WRITE (&(RDG->vertices[I])) | |
184 | #define RDG_MEM_READS_STMT(RDG, I) RDGV_HAS_MEM_READS (&(RDG->vertices[I])) | |
185 | ||
186 | /* Data dependence type. */ | |
187 | ||
188 | enum rdg_dep_type | |
189 | { | |
190 | /* Read After Write (RAW). */ | |
191 | flow_dd = 'f', | |
192 | ||
36875e8f RB |
193 | /* Control dependence (execute conditional on). */ |
194 | control_dd = 'c' | |
80ab0b19 RB |
195 | }; |
196 | ||
197 | /* Dependence information attached to an edge of the RDG. */ | |
198 | ||
526ceb68 | 199 | struct rdg_edge |
80ab0b19 RB |
200 | { |
201 | /* Type of the dependence. */ | |
202 | enum rdg_dep_type type; | |
526ceb68 | 203 | }; |
80ab0b19 RB |
204 | |
205 | #define RDGE_TYPE(E) ((struct rdg_edge *) ((E)->data))->type | |
80ab0b19 | 206 | |
eef99cd9 GB |
207 | /* Kind of distributed loop. */ |
208 | enum partition_kind { | |
209 | PKIND_NORMAL, | |
210 | /* Partial memset stands for a paritition can be distributed into a loop | |
211 | of memset calls, rather than a single memset call. It's handled just | |
212 | like a normal parition, i.e, distributed as separate loop, no memset | |
213 | call is generated. | |
214 | ||
215 | Note: This is a hacking fix trying to distribute ZERO-ing stmt in a | |
216 | loop nest as deep as possible. As a result, parloop achieves better | |
217 | parallelization by parallelizing deeper loop nest. This hack should | |
218 | be unnecessary and removed once distributed memset can be understood | |
219 | and analyzed in data reference analysis. See PR82604 for more. */ | |
220 | PKIND_PARTIAL_MEMSET, | |
221 | PKIND_MEMSET, PKIND_MEMCPY, PKIND_MEMMOVE | |
222 | }; | |
223 | ||
224 | /* Type of distributed loop. */ | |
225 | enum partition_type { | |
226 | /* The distributed loop can be executed parallelly. */ | |
227 | PTYPE_PARALLEL = 0, | |
228 | /* The distributed loop has to be executed sequentially. */ | |
229 | PTYPE_SEQUENTIAL | |
230 | }; | |
231 | ||
232 | /* Builtin info for loop distribution. */ | |
233 | struct builtin_info | |
234 | { | |
235 | /* data-references a kind != PKIND_NORMAL partition is about. */ | |
236 | data_reference_p dst_dr; | |
237 | data_reference_p src_dr; | |
238 | /* Base address and size of memory objects operated by the builtin. Note | |
239 | both dest and source memory objects must have the same size. */ | |
240 | tree dst_base; | |
241 | tree src_base; | |
242 | tree size; | |
243 | /* Base and offset part of dst_base after stripping constant offset. This | |
244 | is only used in memset builtin distribution for now. */ | |
245 | tree dst_base_base; | |
246 | unsigned HOST_WIDE_INT dst_base_offset; | |
247 | }; | |
248 | ||
249 | /* Partition for loop distribution. */ | |
250 | struct partition | |
251 | { | |
252 | /* Statements of the partition. */ | |
253 | bitmap stmts; | |
254 | /* True if the partition defines variable which is used outside of loop. */ | |
255 | bool reduction_p; | |
256 | location_t loc; | |
257 | enum partition_kind kind; | |
258 | enum partition_type type; | |
259 | /* Data references in the partition. */ | |
260 | bitmap datarefs; | |
261 | /* Information of builtin parition. */ | |
262 | struct builtin_info *builtin; | |
263 | }; | |
264 | ||
265 | /* Partitions are fused because of different reasons. */ | |
266 | enum fuse_type | |
267 | { | |
268 | FUSE_NON_BUILTIN = 0, | |
269 | FUSE_REDUCTION = 1, | |
270 | FUSE_SHARE_REF = 2, | |
271 | FUSE_SAME_SCC = 3, | |
272 | FUSE_FINALIZE = 4 | |
273 | }; | |
274 | ||
275 | /* Description on different fusing reason. */ | |
276 | static const char *fuse_message[] = { | |
277 | "they are non-builtins", | |
278 | "they have reductions", | |
279 | "they have shared memory refs", | |
280 | "they are in the same dependence scc", | |
281 | "there is no point to distribute loop"}; | |
282 | ||
283 | ||
80ab0b19 RB |
284 | /* Dump vertex I in RDG to FILE. */ |
285 | ||
286 | static void | |
287 | dump_rdg_vertex (FILE *file, struct graph *rdg, int i) | |
288 | { | |
289 | struct vertex *v = &(rdg->vertices[i]); | |
290 | struct graph_edge *e; | |
291 | ||
292 | fprintf (file, "(vertex %d: (%s%s) (in:", i, | |
293 | RDG_MEM_WRITE_STMT (rdg, i) ? "w" : "", | |
294 | RDG_MEM_READS_STMT (rdg, i) ? "r" : ""); | |
295 | ||
296 | if (v->pred) | |
297 | for (e = v->pred; e; e = e->pred_next) | |
298 | fprintf (file, " %d", e->src); | |
299 | ||
300 | fprintf (file, ") (out:"); | |
301 | ||
302 | if (v->succ) | |
303 | for (e = v->succ; e; e = e->succ_next) | |
304 | fprintf (file, " %d", e->dest); | |
305 | ||
306 | fprintf (file, ")\n"); | |
307 | print_gimple_stmt (file, RDGV_STMT (v), 0, TDF_VOPS|TDF_MEMSYMS); | |
308 | fprintf (file, ")\n"); | |
309 | } | |
310 | ||
311 | /* Call dump_rdg_vertex on stderr. */ | |
312 | ||
313 | DEBUG_FUNCTION void | |
314 | debug_rdg_vertex (struct graph *rdg, int i) | |
315 | { | |
316 | dump_rdg_vertex (stderr, rdg, i); | |
317 | } | |
318 | ||
80ab0b19 RB |
319 | /* Dump the reduced dependence graph RDG to FILE. */ |
320 | ||
321 | static void | |
322 | dump_rdg (FILE *file, struct graph *rdg) | |
323 | { | |
80ab0b19 | 324 | fprintf (file, "(rdg\n"); |
2fd5894f RB |
325 | for (int i = 0; i < rdg->n_vertices; i++) |
326 | dump_rdg_vertex (file, rdg, i); | |
80ab0b19 | 327 | fprintf (file, ")\n"); |
80ab0b19 RB |
328 | } |
329 | ||
330 | /* Call dump_rdg on stderr. */ | |
331 | ||
332 | DEBUG_FUNCTION void | |
333 | debug_rdg (struct graph *rdg) | |
334 | { | |
335 | dump_rdg (stderr, rdg); | |
336 | } | |
337 | ||
338 | static void | |
339 | dot_rdg_1 (FILE *file, struct graph *rdg) | |
340 | { | |
341 | int i; | |
174ec470 RB |
342 | pretty_printer buffer; |
343 | pp_needs_newline (&buffer) = false; | |
344 | buffer.buffer->stream = file; | |
80ab0b19 RB |
345 | |
346 | fprintf (file, "digraph RDG {\n"); | |
347 | ||
348 | for (i = 0; i < rdg->n_vertices; i++) | |
349 | { | |
350 | struct vertex *v = &(rdg->vertices[i]); | |
351 | struct graph_edge *e; | |
352 | ||
174ec470 RB |
353 | fprintf (file, "%d [label=\"[%d] ", i, i); |
354 | pp_gimple_stmt_1 (&buffer, RDGV_STMT (v), 0, TDF_SLIM); | |
355 | pp_flush (&buffer); | |
356 | fprintf (file, "\"]\n"); | |
357 | ||
80ab0b19 RB |
358 | /* Highlight reads from memory. */ |
359 | if (RDG_MEM_READS_STMT (rdg, i)) | |
360 | fprintf (file, "%d [style=filled, fillcolor=green]\n", i); | |
361 | ||
362 | /* Highlight stores to memory. */ | |
363 | if (RDG_MEM_WRITE_STMT (rdg, i)) | |
364 | fprintf (file, "%d [style=filled, fillcolor=red]\n", i); | |
365 | ||
366 | if (v->succ) | |
367 | for (e = v->succ; e; e = e->succ_next) | |
368 | switch (RDGE_TYPE (e)) | |
369 | { | |
80ab0b19 RB |
370 | case flow_dd: |
371 | /* These are the most common dependences: don't print these. */ | |
372 | fprintf (file, "%d -> %d \n", i, e->dest); | |
373 | break; | |
374 | ||
36875e8f RB |
375 | case control_dd: |
376 | fprintf (file, "%d -> %d [label=control] \n", i, e->dest); | |
377 | break; | |
378 | ||
80ab0b19 RB |
379 | default: |
380 | gcc_unreachable (); | |
381 | } | |
382 | } | |
383 | ||
384 | fprintf (file, "}\n\n"); | |
385 | } | |
386 | ||
387 | /* Display the Reduced Dependence Graph using dotty. */ | |
388 | ||
389 | DEBUG_FUNCTION void | |
390 | dot_rdg (struct graph *rdg) | |
391 | { | |
174ec470 | 392 | /* When debugging, you may want to enable the following code. */ |
b6d94045 | 393 | #ifdef HAVE_POPEN |
c3284718 | 394 | FILE *file = popen ("dot -Tx11", "w"); |
174ec470 RB |
395 | if (!file) |
396 | return; | |
80ab0b19 | 397 | dot_rdg_1 (file, rdg); |
174ec470 RB |
398 | fflush (file); |
399 | close (fileno (file)); | |
400 | pclose (file); | |
80ab0b19 RB |
401 | #else |
402 | dot_rdg_1 (stderr, rdg); | |
403 | #endif | |
404 | } | |
405 | ||
406 | /* Returns the index of STMT in RDG. */ | |
407 | ||
408 | static int | |
355fe088 | 409 | rdg_vertex_for_stmt (struct graph *rdg ATTRIBUTE_UNUSED, gimple *stmt) |
80ab0b19 RB |
410 | { |
411 | int index = gimple_uid (stmt); | |
412 | gcc_checking_assert (index == -1 || RDG_STMT (rdg, index) == stmt); | |
413 | return index; | |
414 | } | |
415 | ||
80ab0b19 RB |
416 | /* Creates dependence edges in RDG for all the uses of DEF. IDEF is |
417 | the index of DEF in RDG. */ | |
418 | ||
419 | static void | |
420 | create_rdg_edges_for_scalar (struct graph *rdg, tree def, int idef) | |
421 | { | |
422 | use_operand_p imm_use_p; | |
423 | imm_use_iterator iterator; | |
424 | ||
425 | FOR_EACH_IMM_USE_FAST (imm_use_p, iterator, def) | |
426 | { | |
427 | struct graph_edge *e; | |
428 | int use = rdg_vertex_for_stmt (rdg, USE_STMT (imm_use_p)); | |
429 | ||
430 | if (use < 0) | |
431 | continue; | |
432 | ||
433 | e = add_edge (rdg, idef, use); | |
434 | e->data = XNEW (struct rdg_edge); | |
435 | RDGE_TYPE (e) = flow_dd; | |
80ab0b19 RB |
436 | } |
437 | } | |
438 | ||
36875e8f RB |
439 | /* Creates an edge for the control dependences of BB to the vertex V. */ |
440 | ||
441 | static void | |
442 | create_edge_for_control_dependence (struct graph *rdg, basic_block bb, | |
443 | int v, control_dependences *cd) | |
444 | { | |
445 | bitmap_iterator bi; | |
446 | unsigned edge_n; | |
447 | EXECUTE_IF_SET_IN_BITMAP (cd->get_edges_dependent_on (bb->index), | |
448 | 0, edge_n, bi) | |
449 | { | |
30fd2977 | 450 | basic_block cond_bb = cd->get_edge_src (edge_n); |
355fe088 | 451 | gimple *stmt = last_stmt (cond_bb); |
36875e8f RB |
452 | if (stmt && is_ctrl_stmt (stmt)) |
453 | { | |
454 | struct graph_edge *e; | |
455 | int c = rdg_vertex_for_stmt (rdg, stmt); | |
456 | if (c < 0) | |
457 | continue; | |
458 | ||
459 | e = add_edge (rdg, c, v); | |
460 | e->data = XNEW (struct rdg_edge); | |
461 | RDGE_TYPE (e) = control_dd; | |
36875e8f RB |
462 | } |
463 | } | |
464 | } | |
465 | ||
80ab0b19 RB |
466 | /* Creates the edges of the reduced dependence graph RDG. */ |
467 | ||
468 | static void | |
447f3223 | 469 | create_rdg_flow_edges (struct graph *rdg) |
80ab0b19 RB |
470 | { |
471 | int i; | |
80ab0b19 RB |
472 | def_operand_p def_p; |
473 | ssa_op_iter iter; | |
474 | ||
80ab0b19 RB |
475 | for (i = 0; i < rdg->n_vertices; i++) |
476 | FOR_EACH_PHI_OR_STMT_DEF (def_p, RDG_STMT (rdg, i), | |
477 | iter, SSA_OP_DEF) | |
478 | create_rdg_edges_for_scalar (rdg, DEF_FROM_PTR (def_p), i); | |
447f3223 | 479 | } |
36875e8f | 480 | |
447f3223 RB |
481 | /* Creates the edges of the reduced dependence graph RDG. */ |
482 | ||
483 | static void | |
b71b7a8e | 484 | create_rdg_cd_edges (struct graph *rdg, control_dependences *cd, loop_p loop) |
447f3223 RB |
485 | { |
486 | int i; | |
487 | ||
488 | for (i = 0; i < rdg->n_vertices; i++) | |
489 | { | |
355fe088 | 490 | gimple *stmt = RDG_STMT (rdg, i); |
447f3223 RB |
491 | if (gimple_code (stmt) == GIMPLE_PHI) |
492 | { | |
493 | edge_iterator ei; | |
494 | edge e; | |
495 | FOR_EACH_EDGE (e, ei, gimple_bb (stmt)->preds) | |
b71b7a8e | 496 | if (flow_bb_inside_loop_p (loop, e->src)) |
36875e8f | 497 | create_edge_for_control_dependence (rdg, e->src, i, cd); |
447f3223 RB |
498 | } |
499 | else | |
500 | create_edge_for_control_dependence (rdg, gimple_bb (stmt), i, cd); | |
501 | } | |
80ab0b19 RB |
502 | } |
503 | ||
80ab0b19 | 504 | |
eef99cd9 GB |
505 | class loop_distribution |
506 | { | |
507 | private: | |
508 | /* The loop (nest) to be distributed. */ | |
509 | vec<loop_p> loop_nest; | |
510 | ||
511 | /* Vector of data references in the loop to be distributed. */ | |
512 | vec<data_reference_p> datarefs_vec; | |
513 | ||
514 | /* If there is nonaddressable data reference in above vector. */ | |
515 | bool has_nonaddressable_dataref_p; | |
516 | ||
517 | /* Store index of data reference in aux field. */ | |
518 | ||
519 | /* Hash table for data dependence relation in the loop to be distributed. */ | |
520 | hash_table<ddr_hasher> *ddrs_table; | |
521 | ||
522 | /* Array mapping basic block's index to its topological order. */ | |
523 | int *bb_top_order_index; | |
524 | /* And size of the array. */ | |
525 | int bb_top_order_index_size; | |
526 | ||
527 | /* Build the vertices of the reduced dependence graph RDG. Return false | |
528 | if that failed. */ | |
529 | bool create_rdg_vertices (struct graph *rdg, vec<gimple *> stmts, loop_p loop); | |
530 | ||
531 | /* Initialize STMTS with all the statements of LOOP. We use topological | |
532 | order to discover all statements. The order is important because | |
533 | generate_loops_for_partition is using the same traversal for identifying | |
534 | statements in loop copies. */ | |
535 | void stmts_from_loop (class loop *loop, vec<gimple *> *stmts); | |
536 | ||
537 | ||
538 | /* Build the Reduced Dependence Graph (RDG) with one vertex per statement of | |
539 | LOOP, and one edge per flow dependence or control dependence from control | |
540 | dependence CD. During visiting each statement, data references are also | |
541 | collected and recorded in global data DATAREFS_VEC. */ | |
542 | struct graph * build_rdg (class loop *loop, control_dependences *cd); | |
543 | ||
544 | /* Merge PARTITION into the partition DEST. RDG is the reduced dependence | |
545 | graph and we update type for result partition if it is non-NULL. */ | |
546 | void partition_merge_into (struct graph *rdg, | |
547 | partition *dest, partition *partition, | |
548 | enum fuse_type ft); | |
549 | ||
550 | ||
551 | /* Return data dependence relation for data references A and B. The two | |
552 | data references must be in lexicographic order wrto reduced dependence | |
553 | graph RDG. We firstly try to find ddr from global ddr hash table. If | |
554 | it doesn't exist, compute the ddr and cache it. */ | |
555 | data_dependence_relation * get_data_dependence (struct graph *rdg, | |
556 | data_reference_p a, | |
557 | data_reference_p b); | |
558 | ||
559 | ||
560 | /* In reduced dependence graph RDG for loop distribution, return true if | |
561 | dependence between references DR1 and DR2 leads to a dependence cycle | |
562 | and such dependence cycle can't be resolved by runtime alias check. */ | |
563 | bool data_dep_in_cycle_p (struct graph *rdg, data_reference_p dr1, | |
564 | data_reference_p dr2); | |
565 | ||
566 | ||
567 | /* Given reduced dependence graph RDG, PARTITION1 and PARTITION2, update | |
568 | PARTITION1's type after merging PARTITION2 into PARTITION1. */ | |
569 | void update_type_for_merge (struct graph *rdg, | |
570 | partition *partition1, partition *partition2); | |
571 | ||
572 | ||
573 | /* Returns a partition with all the statements needed for computing | |
574 | the vertex V of the RDG, also including the loop exit conditions. */ | |
575 | partition *build_rdg_partition_for_vertex (struct graph *rdg, int v); | |
576 | ||
577 | /* Given data references DST_DR and SRC_DR in loop nest LOOP and RDG, classify | |
578 | if it forms builtin memcpy or memmove call. */ | |
579 | void classify_builtin_ldst (loop_p loop, struct graph *rdg, partition *partition, | |
580 | data_reference_p dst_dr, data_reference_p src_dr); | |
581 | ||
582 | /* Classifies the builtin kind we can generate for PARTITION of RDG and LOOP. | |
583 | For the moment we detect memset, memcpy and memmove patterns. Bitmap | |
584 | STMT_IN_ALL_PARTITIONS contains statements belonging to all partitions. | |
585 | Returns true if there is a reduction in all partitions and we | |
586 | possibly did not mark PARTITION as having one for this reason. */ | |
587 | ||
588 | bool | |
589 | classify_partition (loop_p loop, | |
590 | struct graph *rdg, partition *partition, | |
591 | bitmap stmt_in_all_partitions); | |
592 | ||
593 | ||
594 | /* Returns true when PARTITION1 and PARTITION2 access the same memory | |
595 | object in RDG. */ | |
596 | bool share_memory_accesses (struct graph *rdg, | |
597 | partition *partition1, partition *partition2); | |
598 | ||
599 | /* For each seed statement in STARTING_STMTS, this function builds | |
600 | partition for it by adding depended statements according to RDG. | |
601 | All partitions are recorded in PARTITIONS. */ | |
602 | void rdg_build_partitions (struct graph *rdg, | |
603 | vec<gimple *> starting_stmts, | |
604 | vec<partition *> *partitions); | |
605 | ||
606 | /* Compute partition dependence created by the data references in DRS1 | |
607 | and DRS2, modify and return DIR according to that. IF ALIAS_DDR is | |
608 | not NULL, we record dependence introduced by possible alias between | |
609 | two data references in ALIAS_DDRS; otherwise, we simply ignore such | |
610 | dependence as if it doesn't exist at all. */ | |
611 | int pg_add_dependence_edges (struct graph *rdg, int dir, bitmap drs1, | |
612 | bitmap drs2, vec<ddr_p> *alias_ddrs); | |
613 | ||
614 | ||
615 | /* Build and return partition dependence graph for PARTITIONS. RDG is | |
616 | reduced dependence graph for the loop to be distributed. If IGNORE_ALIAS_P | |
617 | is true, data dependence caused by possible alias between references | |
618 | is ignored, as if it doesn't exist at all; otherwise all depdendences | |
619 | are considered. */ | |
620 | struct graph *build_partition_graph (struct graph *rdg, | |
621 | vec<struct partition *> *partitions, | |
622 | bool ignore_alias_p); | |
623 | ||
624 | /* Given reduced dependence graph RDG merge strong connected components | |
625 | of PARTITIONS. If IGNORE_ALIAS_P is true, data dependence caused by | |
626 | possible alias between references is ignored, as if it doesn't exist | |
627 | at all; otherwise all depdendences are considered. */ | |
628 | void merge_dep_scc_partitions (struct graph *rdg, vec<struct partition *> | |
629 | *partitions, bool ignore_alias_p); | |
630 | ||
631 | /* This is the main function breaking strong conected components in | |
632 | PARTITIONS giving reduced depdendence graph RDG. Store data dependence | |
633 | relations for runtime alias check in ALIAS_DDRS. */ | |
634 | void break_alias_scc_partitions (struct graph *rdg, vec<struct partition *> | |
635 | *partitions, vec<ddr_p> *alias_ddrs); | |
636 | ||
637 | ||
638 | /* Fuse PARTITIONS of LOOP if necessary before finalizing distribution. | |
639 | ALIAS_DDRS contains ddrs which need runtime alias check. */ | |
640 | void finalize_partitions (class loop *loop, vec<struct partition *> | |
641 | *partitions, vec<ddr_p> *alias_ddrs); | |
642 | ||
643 | /* Distributes the code from LOOP in such a way that producer statements | |
644 | are placed before consumer statements. Tries to separate only the | |
645 | statements from STMTS into separate loops. Returns the number of | |
646 | distributed loops. Set NB_CALLS to number of generated builtin calls. | |
647 | Set *DESTROY_P to whether LOOP needs to be destroyed. */ | |
648 | int distribute_loop (class loop *loop, vec<gimple *> stmts, | |
649 | control_dependences *cd, int *nb_calls, bool *destroy_p, | |
650 | bool only_patterns_p); | |
651 | ||
652 | /* Compute topological order for basic blocks. Topological order is | |
653 | needed because data dependence is computed for data references in | |
654 | lexicographical order. */ | |
655 | void bb_top_order_init (void); | |
656 | ||
657 | void bb_top_order_destroy (void); | |
658 | ||
659 | public: | |
660 | ||
661 | /* Getter for bb_top_order. */ | |
662 | ||
663 | inline int get_bb_top_order_index_size (void) | |
664 | { | |
665 | return bb_top_order_index_size; | |
666 | } | |
667 | ||
668 | inline int get_bb_top_order_index (int i) | |
669 | { | |
670 | return bb_top_order_index[i]; | |
671 | } | |
672 | ||
673 | unsigned int execute (function *fun); | |
674 | }; | |
675 | ||
676 | ||
677 | /* If X has a smaller topological sort number than Y, returns -1; | |
678 | if greater, returns 1. */ | |
679 | static int | |
680 | bb_top_order_cmp_r (const void *x, const void *y, void *loop) | |
681 | { | |
682 | loop_distribution *_loop = | |
683 | (loop_distribution *) loop; | |
684 | ||
685 | basic_block bb1 = *(const basic_block *) x; | |
686 | basic_block bb2 = *(const basic_block *) y; | |
687 | ||
688 | int bb_top_order_index_size = _loop->get_bb_top_order_index_size (); | |
689 | ||
690 | gcc_assert (bb1->index < bb_top_order_index_size | |
691 | && bb2->index < bb_top_order_index_size); | |
692 | gcc_assert (bb1 == bb2 | |
693 | || _loop->get_bb_top_order_index(bb1->index) | |
694 | != _loop->get_bb_top_order_index(bb2->index)); | |
695 | ||
696 | return (_loop->get_bb_top_order_index(bb1->index) - | |
697 | _loop->get_bb_top_order_index(bb2->index)); | |
698 | } | |
699 | ||
700 | bool | |
701 | loop_distribution::create_rdg_vertices (struct graph *rdg, vec<gimple *> stmts, | |
702 | loop_p loop) | |
80ab0b19 RB |
703 | { |
704 | int i; | |
355fe088 | 705 | gimple *stmt; |
80ab0b19 RB |
706 | |
707 | FOR_EACH_VEC_ELT (stmts, i, stmt) | |
708 | { | |
709 | struct vertex *v = &(rdg->vertices[i]); | |
710 | ||
711 | /* Record statement to vertex mapping. */ | |
712 | gimple_set_uid (stmt, i); | |
713 | ||
714 | v->data = XNEW (struct rdg_vertex); | |
715 | RDGV_STMT (v) = stmt; | |
716 | RDGV_DATAREFS (v).create (0); | |
717 | RDGV_HAS_MEM_WRITE (v) = false; | |
718 | RDGV_HAS_MEM_READS (v) = false; | |
719 | if (gimple_code (stmt) == GIMPLE_PHI) | |
720 | continue; | |
721 | ||
9fafb14a BC |
722 | unsigned drp = datarefs_vec.length (); |
723 | if (!find_data_references_in_stmt (loop, stmt, &datarefs_vec)) | |
80ab0b19 | 724 | return false; |
9fafb14a | 725 | for (unsigned j = drp; j < datarefs_vec.length (); ++j) |
80ab0b19 | 726 | { |
9fafb14a | 727 | data_reference_p dr = datarefs_vec[j]; |
80ab0b19 RB |
728 | if (DR_IS_READ (dr)) |
729 | RDGV_HAS_MEM_READS (v) = true; | |
730 | else | |
731 | RDGV_HAS_MEM_WRITE (v) = true; | |
732 | RDGV_DATAREFS (v).safe_push (dr); | |
c4450491 | 733 | has_nonaddressable_dataref_p |= may_be_nonaddressable_p (dr->ref); |
80ab0b19 RB |
734 | } |
735 | } | |
736 | return true; | |
737 | } | |
738 | ||
eef99cd9 GB |
739 | void |
740 | loop_distribution::stmts_from_loop (class loop *loop, vec<gimple *> *stmts) | |
80ab0b19 RB |
741 | { |
742 | unsigned int i; | |
eef99cd9 | 743 | basic_block *bbs = get_loop_body_in_custom_order (loop, this, bb_top_order_cmp_r); |
80ab0b19 RB |
744 | |
745 | for (i = 0; i < loop->num_nodes; i++) | |
746 | { | |
747 | basic_block bb = bbs[i]; | |
80ab0b19 | 748 | |
538dd0b7 DM |
749 | for (gphi_iterator bsi = gsi_start_phis (bb); !gsi_end_p (bsi); |
750 | gsi_next (&bsi)) | |
751 | if (!virtual_operand_p (gimple_phi_result (bsi.phi ()))) | |
752 | stmts->safe_push (bsi.phi ()); | |
80ab0b19 | 753 | |
538dd0b7 DM |
754 | for (gimple_stmt_iterator bsi = gsi_start_bb (bb); !gsi_end_p (bsi); |
755 | gsi_next (&bsi)) | |
80ab0b19 | 756 | { |
355fe088 | 757 | gimple *stmt = gsi_stmt (bsi); |
80ab0b19 RB |
758 | if (gimple_code (stmt) != GIMPLE_LABEL && !is_gimple_debug (stmt)) |
759 | stmts->safe_push (stmt); | |
760 | } | |
761 | } | |
762 | ||
763 | free (bbs); | |
764 | } | |
765 | ||
80ab0b19 RB |
766 | /* Free the reduced dependence graph RDG. */ |
767 | ||
768 | static void | |
769 | free_rdg (struct graph *rdg) | |
770 | { | |
771 | int i; | |
772 | ||
773 | for (i = 0; i < rdg->n_vertices; i++) | |
774 | { | |
775 | struct vertex *v = &(rdg->vertices[i]); | |
776 | struct graph_edge *e; | |
777 | ||
778 | for (e = v->succ; e; e = e->succ_next) | |
447f3223 | 779 | free (e->data); |
80ab0b19 RB |
780 | |
781 | if (v->data) | |
782 | { | |
783 | gimple_set_uid (RDGV_STMT (v), -1); | |
9fafb14a | 784 | (RDGV_DATAREFS (v)).release (); |
80ab0b19 RB |
785 | free (v->data); |
786 | } | |
787 | } | |
788 | ||
789 | free_graph (rdg); | |
790 | } | |
791 | ||
eef99cd9 GB |
792 | struct graph * |
793 | loop_distribution::build_rdg (class loop *loop, control_dependences *cd) | |
80ab0b19 RB |
794 | { |
795 | struct graph *rdg; | |
80ab0b19 | 796 | |
97463b2b | 797 | /* Create the RDG vertices from the stmts of the loop nest. */ |
355fe088 | 798 | auto_vec<gimple *, 10> stmts; |
4084ea5f | 799 | stmts_from_loop (loop, &stmts); |
24f161fd | 800 | rdg = new_graph (stmts.length ()); |
9fafb14a | 801 | if (!create_rdg_vertices (rdg, stmts, loop)) |
80ab0b19 | 802 | { |
80ab0b19 RB |
803 | free_rdg (rdg); |
804 | return NULL; | |
805 | } | |
806 | stmts.release (); | |
97463b2b | 807 | |
447f3223 RB |
808 | create_rdg_flow_edges (rdg); |
809 | if (cd) | |
4084ea5f | 810 | create_rdg_cd_edges (rdg, cd, loop); |
447f3223 | 811 | |
80ab0b19 RB |
812 | return rdg; |
813 | } | |
814 | ||
80ab0b19 | 815 | |
c61f8985 RG |
816 | /* Allocate and initialize a partition from BITMAP. */ |
817 | ||
526ceb68 | 818 | static partition * |
a7a44c07 | 819 | partition_alloc (void) |
c61f8985 | 820 | { |
526ceb68 | 821 | partition *partition = XCNEW (struct partition); |
a7a44c07 | 822 | partition->stmts = BITMAP_ALLOC (NULL); |
826a536d | 823 | partition->reduction_p = false; |
dbd59515 | 824 | partition->loc = UNKNOWN_LOCATION; |
30d55936 | 825 | partition->kind = PKIND_NORMAL; |
dbd59515 | 826 | partition->type = PTYPE_PARALLEL; |
a7a44c07 | 827 | partition->datarefs = BITMAP_ALLOC (NULL); |
c61f8985 RG |
828 | return partition; |
829 | } | |
830 | ||
831 | /* Free PARTITION. */ | |
832 | ||
833 | static void | |
526ceb68 | 834 | partition_free (partition *partition) |
c61f8985 RG |
835 | { |
836 | BITMAP_FREE (partition->stmts); | |
a7a44c07 | 837 | BITMAP_FREE (partition->datarefs); |
939cf90f BC |
838 | if (partition->builtin) |
839 | free (partition->builtin); | |
840 | ||
c61f8985 RG |
841 | free (partition); |
842 | } | |
843 | ||
30d55936 RG |
844 | /* Returns true if the partition can be generated as a builtin. */ |
845 | ||
846 | static bool | |
526ceb68 | 847 | partition_builtin_p (partition *partition) |
30d55936 | 848 | { |
5955438a | 849 | return partition->kind > PKIND_PARTIAL_MEMSET; |
30d55936 | 850 | } |
c61f8985 | 851 | |
826a536d | 852 | /* Returns true if the partition contains a reduction. */ |
7ad672e4 RG |
853 | |
854 | static bool | |
526ceb68 | 855 | partition_reduction_p (partition *partition) |
7ad672e4 | 856 | { |
826a536d | 857 | return partition->reduction_p; |
7ad672e4 RG |
858 | } |
859 | ||
eef99cd9 GB |
860 | void |
861 | loop_distribution::partition_merge_into (struct graph *rdg, | |
862 | partition *dest, partition *partition, enum fuse_type ft) | |
f1eb4621 | 863 | { |
821dbeef BC |
864 | if (dump_file && (dump_flags & TDF_DETAILS)) |
865 | { | |
866 | fprintf (dump_file, "Fuse partitions because %s:\n", fuse_message[ft]); | |
867 | fprintf (dump_file, " Part 1: "); | |
868 | dump_bitmap (dump_file, dest->stmts); | |
869 | fprintf (dump_file, " Part 2: "); | |
870 | dump_bitmap (dump_file, partition->stmts); | |
871 | } | |
f1eb4621 BC |
872 | |
873 | dest->kind = PKIND_NORMAL; | |
874 | if (dest->type == PTYPE_PARALLEL) | |
875 | dest->type = partition->type; | |
876 | ||
877 | bitmap_ior_into (dest->stmts, partition->stmts); | |
878 | if (partition_reduction_p (partition)) | |
879 | dest->reduction_p = true; | |
880 | ||
881 | /* Further check if any data dependence prevents us from executing the | |
882 | new partition parallelly. */ | |
883 | if (dest->type == PTYPE_PARALLEL && rdg != NULL) | |
884 | update_type_for_merge (rdg, dest, partition); | |
885 | ||
886 | bitmap_ior_into (dest->datarefs, partition->datarefs); | |
826a536d RB |
887 | } |
888 | ||
889 | ||
c07a8cb3 RG |
890 | /* Returns true when DEF is an SSA_NAME defined in LOOP and used after |
891 | the LOOP. */ | |
892 | ||
893 | static bool | |
894 | ssa_name_has_uses_outside_loop_p (tree def, loop_p loop) | |
895 | { | |
896 | imm_use_iterator imm_iter; | |
897 | use_operand_p use_p; | |
898 | ||
899 | FOR_EACH_IMM_USE_FAST (use_p, imm_iter, def) | |
e665269a | 900 | { |
163aa51b BC |
901 | if (is_gimple_debug (USE_STMT (use_p))) |
902 | continue; | |
903 | ||
904 | basic_block use_bb = gimple_bb (USE_STMT (use_p)); | |
905 | if (!flow_bb_inside_loop_p (loop, use_bb)) | |
e665269a RG |
906 | return true; |
907 | } | |
c07a8cb3 RG |
908 | |
909 | return false; | |
910 | } | |
911 | ||
912 | /* Returns true when STMT defines a scalar variable used after the | |
88af7c1a | 913 | loop LOOP. */ |
c07a8cb3 RG |
914 | |
915 | static bool | |
355fe088 | 916 | stmt_has_scalar_dependences_outside_loop (loop_p loop, gimple *stmt) |
c07a8cb3 | 917 | { |
88af7c1a RG |
918 | def_operand_p def_p; |
919 | ssa_op_iter op_iter; | |
c07a8cb3 | 920 | |
9ca86fc3 RG |
921 | if (gimple_code (stmt) == GIMPLE_PHI) |
922 | return ssa_name_has_uses_outside_loop_p (gimple_phi_result (stmt), loop); | |
923 | ||
88af7c1a RG |
924 | FOR_EACH_SSA_DEF_OPERAND (def_p, stmt, op_iter, SSA_OP_DEF) |
925 | if (ssa_name_has_uses_outside_loop_p (DEF_FROM_PTR (def_p), loop)) | |
926 | return true; | |
c07a8cb3 | 927 | |
88af7c1a | 928 | return false; |
c07a8cb3 RG |
929 | } |
930 | ||
dea61d92 SP |
931 | /* Return a copy of LOOP placed before LOOP. */ |
932 | ||
99b1c316 MS |
933 | static class loop * |
934 | copy_loop_before (class loop *loop) | |
dea61d92 | 935 | { |
99b1c316 | 936 | class loop *res; |
dea61d92 SP |
937 | edge preheader = loop_preheader_edge (loop); |
938 | ||
dea61d92 | 939 | initialize_original_copy_tables (); |
5ce9450f | 940 | res = slpeel_tree_duplicate_loop_to_edge_cfg (loop, NULL, preheader); |
30d55936 | 941 | gcc_assert (res != NULL); |
dea61d92 | 942 | free_original_copy_tables (); |
2cfc56b9 | 943 | delete_update_ssa (); |
dea61d92 SP |
944 | |
945 | return res; | |
946 | } | |
947 | ||
948 | /* Creates an empty basic block after LOOP. */ | |
949 | ||
950 | static void | |
99b1c316 | 951 | create_bb_after_loop (class loop *loop) |
dea61d92 SP |
952 | { |
953 | edge exit = single_exit (loop); | |
954 | ||
955 | if (!exit) | |
956 | return; | |
957 | ||
958 | split_edge (exit); | |
959 | } | |
960 | ||
961 | /* Generate code for PARTITION from the code in LOOP. The loop is | |
962 | copied when COPY_P is true. All the statements not flagged in the | |
963 | PARTITION bitmap are removed from the loop or from its copy. The | |
964 | statements are indexed in sequence inside a basic block, and the | |
30d55936 | 965 | basic blocks of a loop are taken in dom order. */ |
dea61d92 | 966 | |
30d55936 | 967 | static void |
99b1c316 | 968 | generate_loops_for_partition (class loop *loop, partition *partition, |
c61f8985 | 969 | bool copy_p) |
dea61d92 | 970 | { |
2fd5894f | 971 | unsigned i; |
dea61d92 SP |
972 | basic_block *bbs; |
973 | ||
974 | if (copy_p) | |
975 | { | |
a8745cc2 | 976 | int orig_loop_num = loop->orig_loop_num; |
dea61d92 | 977 | loop = copy_loop_before (loop); |
30d55936 | 978 | gcc_assert (loop != NULL); |
a8745cc2 | 979 | loop->orig_loop_num = orig_loop_num; |
dea61d92 SP |
980 | create_preheader (loop, CP_SIMPLE_PREHEADERS); |
981 | create_bb_after_loop (loop); | |
982 | } | |
a8745cc2 BC |
983 | else |
984 | { | |
985 | /* Origin number is set to the new versioned loop's num. */ | |
986 | gcc_assert (loop->orig_loop_num != loop->num); | |
987 | } | |
dea61d92 | 988 | |
2fd5894f | 989 | /* Remove stmts not in the PARTITION bitmap. */ |
dea61d92 SP |
990 | bbs = get_loop_body_in_dom_order (loop); |
991 | ||
36f52e8f | 992 | if (MAY_HAVE_DEBUG_BIND_STMTS) |
2fd5894f | 993 | for (i = 0; i < loop->num_nodes; i++) |
b03c3082 JJ |
994 | { |
995 | basic_block bb = bbs[i]; | |
996 | ||
538dd0b7 DM |
997 | for (gphi_iterator bsi = gsi_start_phis (bb); !gsi_end_p (bsi); |
998 | gsi_next (&bsi)) | |
2fd5894f | 999 | { |
538dd0b7 | 1000 | gphi *phi = bsi.phi (); |
2fd5894f RB |
1001 | if (!virtual_operand_p (gimple_phi_result (phi)) |
1002 | && !bitmap_bit_p (partition->stmts, gimple_uid (phi))) | |
1003 | reset_debug_uses (phi); | |
1004 | } | |
b03c3082 | 1005 | |
538dd0b7 | 1006 | for (gimple_stmt_iterator bsi = gsi_start_bb (bb); !gsi_end_p (bsi); gsi_next (&bsi)) |
b03c3082 | 1007 | { |
355fe088 | 1008 | gimple *stmt = gsi_stmt (bsi); |
b03c3082 JJ |
1009 | if (gimple_code (stmt) != GIMPLE_LABEL |
1010 | && !is_gimple_debug (stmt) | |
2fd5894f | 1011 | && !bitmap_bit_p (partition->stmts, gimple_uid (stmt))) |
b03c3082 JJ |
1012 | reset_debug_uses (stmt); |
1013 | } | |
1014 | } | |
1015 | ||
2fd5894f | 1016 | for (i = 0; i < loop->num_nodes; i++) |
dea61d92 SP |
1017 | { |
1018 | basic_block bb = bbs[i]; | |
efe040bf BC |
1019 | edge inner_exit = NULL; |
1020 | ||
1021 | if (loop != bb->loop_father) | |
1022 | inner_exit = single_exit (bb->loop_father); | |
dea61d92 | 1023 | |
538dd0b7 | 1024 | for (gphi_iterator bsi = gsi_start_phis (bb); !gsi_end_p (bsi);) |
2fd5894f | 1025 | { |
538dd0b7 | 1026 | gphi *phi = bsi.phi (); |
2fd5894f RB |
1027 | if (!virtual_operand_p (gimple_phi_result (phi)) |
1028 | && !bitmap_bit_p (partition->stmts, gimple_uid (phi))) | |
2706a615 | 1029 | remove_phi_node (&bsi, true); |
2fd5894f RB |
1030 | else |
1031 | gsi_next (&bsi); | |
1032 | } | |
dea61d92 | 1033 | |
538dd0b7 | 1034 | for (gimple_stmt_iterator bsi = gsi_start_bb (bb); !gsi_end_p (bsi);) |
2706a615 | 1035 | { |
355fe088 | 1036 | gimple *stmt = gsi_stmt (bsi); |
b03c3082 JJ |
1037 | if (gimple_code (stmt) != GIMPLE_LABEL |
1038 | && !is_gimple_debug (stmt) | |
2fd5894f | 1039 | && !bitmap_bit_p (partition->stmts, gimple_uid (stmt))) |
2706a615 | 1040 | { |
efe040bf BC |
1041 | /* In distribution of loop nest, if bb is inner loop's exit_bb, |
1042 | we choose its exit edge/path in order to avoid generating | |
1043 | infinite loop. For all other cases, we choose an arbitrary | |
1044 | path through the empty CFG part that this unnecessary | |
1045 | control stmt controls. */ | |
538dd0b7 | 1046 | if (gcond *cond_stmt = dyn_cast <gcond *> (stmt)) |
36875e8f | 1047 | { |
efe040bf BC |
1048 | if (inner_exit && inner_exit->flags & EDGE_TRUE_VALUE) |
1049 | gimple_cond_make_true (cond_stmt); | |
1050 | else | |
1051 | gimple_cond_make_false (cond_stmt); | |
36875e8f RB |
1052 | update_stmt (stmt); |
1053 | } | |
1054 | else if (gimple_code (stmt) == GIMPLE_SWITCH) | |
1055 | { | |
538dd0b7 | 1056 | gswitch *switch_stmt = as_a <gswitch *> (stmt); |
36875e8f | 1057 | gimple_switch_set_index |
538dd0b7 | 1058 | (switch_stmt, CASE_LOW (gimple_switch_label (switch_stmt, 1))); |
36875e8f RB |
1059 | update_stmt (stmt); |
1060 | } | |
1061 | else | |
1062 | { | |
1063 | unlink_stmt_vdef (stmt); | |
1064 | gsi_remove (&bsi, true); | |
1065 | release_defs (stmt); | |
1066 | continue; | |
1067 | } | |
2706a615 | 1068 | } |
36875e8f | 1069 | gsi_next (&bsi); |
2706a615 | 1070 | } |
dea61d92 SP |
1071 | } |
1072 | ||
1073 | free (bbs); | |
dea61d92 SP |
1074 | } |
1075 | ||
401f3a81 JJ |
1076 | /* If VAL memory representation contains the same value in all bytes, |
1077 | return that value, otherwise return -1. | |
1078 | E.g. for 0x24242424 return 0x24, for IEEE double | |
1079 | 747708026454360457216.0 return 0x44, etc. */ | |
1080 | ||
1081 | static int | |
1082 | const_with_all_bytes_same (tree val) | |
1083 | { | |
1084 | unsigned char buf[64]; | |
1085 | int i, len; | |
1086 | ||
1087 | if (integer_zerop (val) | |
401f3a81 JJ |
1088 | || (TREE_CODE (val) == CONSTRUCTOR |
1089 | && !TREE_CLOBBER_P (val) | |
1090 | && CONSTRUCTOR_NELTS (val) == 0)) | |
1091 | return 0; | |
1092 | ||
9e207d6f JJ |
1093 | if (real_zerop (val)) |
1094 | { | |
1095 | /* Only return 0 for +0.0, not for -0.0, which doesn't have | |
1096 | an all bytes same memory representation. Don't transform | |
1097 | -0.0 stores into +0.0 even for !HONOR_SIGNED_ZEROS. */ | |
1098 | switch (TREE_CODE (val)) | |
1099 | { | |
1100 | case REAL_CST: | |
1101 | if (!real_isneg (TREE_REAL_CST_PTR (val))) | |
1102 | return 0; | |
1103 | break; | |
1104 | case COMPLEX_CST: | |
1105 | if (!const_with_all_bytes_same (TREE_REALPART (val)) | |
1106 | && !const_with_all_bytes_same (TREE_IMAGPART (val))) | |
1107 | return 0; | |
1108 | break; | |
1109 | case VECTOR_CST: | |
63570af0 RS |
1110 | { |
1111 | unsigned int count = vector_cst_encoded_nelts (val); | |
1112 | unsigned int j; | |
1113 | for (j = 0; j < count; ++j) | |
1114 | if (const_with_all_bytes_same (VECTOR_CST_ENCODED_ELT (val, j))) | |
1115 | break; | |
1116 | if (j == count) | |
1117 | return 0; | |
1118 | break; | |
1119 | } | |
9e207d6f JJ |
1120 | default: |
1121 | break; | |
1122 | } | |
1123 | } | |
1124 | ||
401f3a81 JJ |
1125 | if (CHAR_BIT != 8 || BITS_PER_UNIT != 8) |
1126 | return -1; | |
1127 | ||
1128 | len = native_encode_expr (val, buf, sizeof (buf)); | |
1129 | if (len == 0) | |
1130 | return -1; | |
1131 | for (i = 1; i < len; i++) | |
1132 | if (buf[i] != buf[0]) | |
1133 | return -1; | |
1134 | return buf[0]; | |
1135 | } | |
1136 | ||
30d55936 | 1137 | /* Generate a call to memset for PARTITION in LOOP. */ |
dea61d92 | 1138 | |
cfee318d | 1139 | static void |
99b1c316 | 1140 | generate_memset_builtin (class loop *loop, partition *partition) |
dea61d92 | 1141 | { |
30d55936 | 1142 | gimple_stmt_iterator gsi; |
818625cf | 1143 | tree mem, fn, nb_bytes; |
b6dd5261 | 1144 | tree val; |
939cf90f BC |
1145 | struct builtin_info *builtin = partition->builtin; |
1146 | gimple *fn_call; | |
30d55936 RG |
1147 | |
1148 | /* The new statements will be placed before LOOP. */ | |
1149 | gsi = gsi_last_bb (loop_preheader_edge (loop)->src); | |
dea61d92 | 1150 | |
4c9ed22a | 1151 | nb_bytes = rewrite_to_non_trapping_overflow (builtin->size); |
d0582dc1 RG |
1152 | nb_bytes = force_gimple_operand_gsi (&gsi, nb_bytes, true, NULL_TREE, |
1153 | false, GSI_CONTINUE_LINKING); | |
05ac4d9c | 1154 | mem = rewrite_to_non_trapping_overflow (builtin->dst_base); |
d0582dc1 RG |
1155 | mem = force_gimple_operand_gsi (&gsi, mem, true, NULL_TREE, |
1156 | false, GSI_CONTINUE_LINKING); | |
dea61d92 | 1157 | |
b6dd5261 | 1158 | /* This exactly matches the pattern recognition in classify_partition. */ |
939cf90f | 1159 | val = gimple_assign_rhs1 (DR_STMT (builtin->dst_dr)); |
401f3a81 JJ |
1160 | /* Handle constants like 0x15151515 and similarly |
1161 | floating point constants etc. where all bytes are the same. */ | |
1162 | int bytev = const_with_all_bytes_same (val); | |
1163 | if (bytev != -1) | |
1164 | val = build_int_cst (integer_type_node, bytev); | |
1165 | else if (TREE_CODE (val) == INTEGER_CST) | |
1166 | val = fold_convert (integer_type_node, val); | |
1167 | else if (!useless_type_conversion_p (integer_type_node, TREE_TYPE (val))) | |
b6dd5261 | 1168 | { |
b731b390 | 1169 | tree tem = make_ssa_name (integer_type_node); |
355fe088 | 1170 | gimple *cstmt = gimple_build_assign (tem, NOP_EXPR, val); |
401f3a81 JJ |
1171 | gsi_insert_after (&gsi, cstmt, GSI_CONTINUE_LINKING); |
1172 | val = tem; | |
b6dd5261 RG |
1173 | } |
1174 | ||
e79983f4 | 1175 | fn = build_fold_addr_expr (builtin_decl_implicit (BUILT_IN_MEMSET)); |
b6dd5261 | 1176 | fn_call = gimple_build_call (fn, 3, mem, val, nb_bytes); |
dbd59515 | 1177 | gimple_set_location (fn_call, partition->loc); |
d0582dc1 | 1178 | gsi_insert_after (&gsi, fn_call, GSI_CONTINUE_LINKING); |
5879ab5f | 1179 | fold_stmt (&gsi); |
dea61d92 SP |
1180 | |
1181 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
b6dd5261 RG |
1182 | { |
1183 | fprintf (dump_file, "generated memset"); | |
401f3a81 | 1184 | if (bytev == 0) |
b6dd5261 | 1185 | fprintf (dump_file, " zero\n"); |
b6dd5261 RG |
1186 | else |
1187 | fprintf (dump_file, "\n"); | |
1188 | } | |
dea61d92 SP |
1189 | } |
1190 | ||
d0582dc1 RG |
1191 | /* Generate a call to memcpy for PARTITION in LOOP. */ |
1192 | ||
1193 | static void | |
99b1c316 | 1194 | generate_memcpy_builtin (class loop *loop, partition *partition) |
d0582dc1 RG |
1195 | { |
1196 | gimple_stmt_iterator gsi; | |
939cf90f | 1197 | gimple *fn_call; |
818625cf | 1198 | tree dest, src, fn, nb_bytes; |
d0582dc1 | 1199 | enum built_in_function kind; |
939cf90f | 1200 | struct builtin_info *builtin = partition->builtin; |
d0582dc1 RG |
1201 | |
1202 | /* The new statements will be placed before LOOP. */ | |
1203 | gsi = gsi_last_bb (loop_preheader_edge (loop)->src); | |
1204 | ||
4c9ed22a | 1205 | nb_bytes = rewrite_to_non_trapping_overflow (builtin->size); |
d0582dc1 RG |
1206 | nb_bytes = force_gimple_operand_gsi (&gsi, nb_bytes, true, NULL_TREE, |
1207 | false, GSI_CONTINUE_LINKING); | |
05ac4d9c JJ |
1208 | dest = rewrite_to_non_trapping_overflow (builtin->dst_base); |
1209 | src = rewrite_to_non_trapping_overflow (builtin->src_base); | |
510d73a0 RB |
1210 | if (partition->kind == PKIND_MEMCPY |
1211 | || ! ptr_derefs_may_alias_p (dest, src)) | |
d0582dc1 | 1212 | kind = BUILT_IN_MEMCPY; |
510d73a0 RB |
1213 | else |
1214 | kind = BUILT_IN_MEMMOVE; | |
d0582dc1 RG |
1215 | |
1216 | dest = force_gimple_operand_gsi (&gsi, dest, true, NULL_TREE, | |
1217 | false, GSI_CONTINUE_LINKING); | |
1218 | src = force_gimple_operand_gsi (&gsi, src, true, NULL_TREE, | |
1219 | false, GSI_CONTINUE_LINKING); | |
1220 | fn = build_fold_addr_expr (builtin_decl_implicit (kind)); | |
1221 | fn_call = gimple_build_call (fn, 3, dest, src, nb_bytes); | |
dbd59515 | 1222 | gimple_set_location (fn_call, partition->loc); |
d0582dc1 | 1223 | gsi_insert_after (&gsi, fn_call, GSI_CONTINUE_LINKING); |
5879ab5f | 1224 | fold_stmt (&gsi); |
d0582dc1 RG |
1225 | |
1226 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
1227 | { | |
1228 | if (kind == BUILT_IN_MEMCPY) | |
1229 | fprintf (dump_file, "generated memcpy\n"); | |
1230 | else | |
1231 | fprintf (dump_file, "generated memmove\n"); | |
1232 | } | |
1233 | } | |
1234 | ||
30d55936 | 1235 | /* Remove and destroy the loop LOOP. */ |
dea61d92 | 1236 | |
30d55936 | 1237 | static void |
99b1c316 | 1238 | destroy_loop (class loop *loop) |
dea61d92 | 1239 | { |
30d55936 RG |
1240 | unsigned nbbs = loop->num_nodes; |
1241 | edge exit = single_exit (loop); | |
1242 | basic_block src = loop_preheader_edge (loop)->src, dest = exit->dest; | |
dea61d92 | 1243 | basic_block *bbs; |
30d55936 | 1244 | unsigned i; |
dea61d92 SP |
1245 | |
1246 | bbs = get_loop_body_in_dom_order (loop); | |
1247 | ||
b7a9e9f4 RB |
1248 | gimple_stmt_iterator dst_gsi = gsi_after_labels (exit->dest); |
1249 | bool safe_p = single_pred_p (exit->dest); | |
efb34006 | 1250 | for (unsigned i = 0; i < nbbs; ++i) |
c014f6f5 RG |
1251 | { |
1252 | /* We have made sure to not leave any dangling uses of SSA | |
1253 | names defined in the loop. With the exception of virtuals. | |
1254 | Make sure we replace all uses of virtual defs that will remain | |
1255 | outside of the loop with the bare symbol as delete_basic_block | |
1256 | will release them. */ | |
538dd0b7 DM |
1257 | for (gphi_iterator gsi = gsi_start_phis (bbs[i]); !gsi_end_p (gsi); |
1258 | gsi_next (&gsi)) | |
c014f6f5 | 1259 | { |
538dd0b7 | 1260 | gphi *phi = gsi.phi (); |
ea057359 | 1261 | if (virtual_operand_p (gimple_phi_result (phi))) |
c014f6f5 RG |
1262 | mark_virtual_phi_result_for_renaming (phi); |
1263 | } | |
b7a9e9f4 | 1264 | for (gimple_stmt_iterator gsi = gsi_start_bb (bbs[i]); !gsi_end_p (gsi);) |
c014f6f5 | 1265 | { |
355fe088 | 1266 | gimple *stmt = gsi_stmt (gsi); |
c014f6f5 RG |
1267 | tree vdef = gimple_vdef (stmt); |
1268 | if (vdef && TREE_CODE (vdef) == SSA_NAME) | |
1269 | mark_virtual_operand_for_renaming (vdef); | |
b7a9e9f4 RB |
1270 | /* Also move and eventually reset debug stmts. We can leave |
1271 | constant values in place in case the stmt dominates the exit. | |
1272 | ??? Non-constant values from the last iteration can be | |
1273 | replaced with final values if we can compute them. */ | |
1274 | if (gimple_debug_bind_p (stmt)) | |
1275 | { | |
1276 | tree val = gimple_debug_bind_get_value (stmt); | |
1277 | gsi_move_before (&gsi, &dst_gsi); | |
1278 | if (val | |
1279 | && (!safe_p | |
1280 | || !is_gimple_min_invariant (val) | |
1281 | || !dominated_by_p (CDI_DOMINATORS, exit->src, bbs[i]))) | |
1282 | { | |
1283 | gimple_debug_bind_reset_value (stmt); | |
1284 | update_stmt (stmt); | |
1285 | } | |
1286 | } | |
1287 | else | |
1288 | gsi_next (&gsi); | |
c014f6f5 | 1289 | } |
b7a9e9f4 | 1290 | } |
b7a9e9f4 RB |
1291 | |
1292 | redirect_edge_pred (exit, src); | |
1293 | exit->flags &= ~(EDGE_TRUE_VALUE|EDGE_FALSE_VALUE); | |
1294 | exit->flags |= EDGE_FALLTHRU; | |
1295 | cancel_loop_tree (loop); | |
1296 | rescan_loop_exit (exit, false, true); | |
1297 | ||
1298 | i = nbbs; | |
1299 | do | |
1300 | { | |
1301 | --i; | |
c014f6f5 RG |
1302 | delete_basic_block (bbs[i]); |
1303 | } | |
b9aba0a0 RB |
1304 | while (i != 0); |
1305 | ||
dea61d92 | 1306 | free (bbs); |
30d55936 RG |
1307 | |
1308 | set_immediate_dominator (CDI_DOMINATORS, dest, | |
1309 | recompute_dominator (CDI_DOMINATORS, dest)); | |
dea61d92 SP |
1310 | } |
1311 | ||
b71b7a8e | 1312 | /* Generates code for PARTITION. Return whether LOOP needs to be destroyed. */ |
dea61d92 | 1313 | |
b71b7a8e | 1314 | static bool |
99b1c316 | 1315 | generate_code_for_partition (class loop *loop, |
526ceb68 | 1316 | partition *partition, bool copy_p) |
dea61d92 | 1317 | { |
30d55936 RG |
1318 | switch (partition->kind) |
1319 | { | |
826a536d | 1320 | case PKIND_NORMAL: |
5955438a | 1321 | case PKIND_PARTIAL_MEMSET: |
826a536d RB |
1322 | /* Reductions all have to be in the last partition. */ |
1323 | gcc_assert (!partition_reduction_p (partition) | |
1324 | || !copy_p); | |
1325 | generate_loops_for_partition (loop, partition, copy_p); | |
b71b7a8e | 1326 | return false; |
826a536d | 1327 | |
30d55936 | 1328 | case PKIND_MEMSET: |
d0582dc1 | 1329 | generate_memset_builtin (loop, partition); |
d0582dc1 RG |
1330 | break; |
1331 | ||
1332 | case PKIND_MEMCPY: | |
510d73a0 | 1333 | case PKIND_MEMMOVE: |
d0582dc1 | 1334 | generate_memcpy_builtin (loop, partition); |
30d55936 RG |
1335 | break; |
1336 | ||
1337 | default: | |
1338 | gcc_unreachable (); | |
1339 | } | |
dea61d92 | 1340 | |
826a536d RB |
1341 | /* Common tail for partitions we turn into a call. If this was the last |
1342 | partition for which we generate code, we have to destroy the loop. */ | |
1343 | if (!copy_p) | |
b71b7a8e RB |
1344 | return true; |
1345 | return false; | |
dea61d92 SP |
1346 | } |
1347 | ||
eef99cd9 GB |
1348 | data_dependence_relation * |
1349 | loop_distribution::get_data_dependence (struct graph *rdg, data_reference_p a, | |
1350 | data_reference_p b) | |
17c5cbdf BC |
1351 | { |
1352 | struct data_dependence_relation ent, **slot; | |
1353 | struct data_dependence_relation *ddr; | |
1354 | ||
1355 | gcc_assert (DR_IS_WRITE (a) || DR_IS_WRITE (b)); | |
1356 | gcc_assert (rdg_vertex_for_stmt (rdg, DR_STMT (a)) | |
1357 | <= rdg_vertex_for_stmt (rdg, DR_STMT (b))); | |
1358 | ent.a = a; | |
1359 | ent.b = b; | |
1e485f89 | 1360 | slot = ddrs_table->find_slot (&ent, INSERT); |
17c5cbdf BC |
1361 | if (*slot == NULL) |
1362 | { | |
1363 | ddr = initialize_data_dependence_relation (a, b, loop_nest); | |
1364 | compute_affine_dependence (ddr, loop_nest[0]); | |
1365 | *slot = ddr; | |
1366 | } | |
1367 | ||
1368 | return *slot; | |
1369 | } | |
dea61d92 | 1370 | |
eef99cd9 GB |
1371 | bool |
1372 | loop_distribution::data_dep_in_cycle_p (struct graph *rdg, | |
1373 | data_reference_p dr1, | |
1374 | data_reference_p dr2) | |
f1eb4621 BC |
1375 | { |
1376 | struct data_dependence_relation *ddr; | |
1377 | ||
1378 | /* Re-shuffle data-refs to be in topological order. */ | |
1379 | if (rdg_vertex_for_stmt (rdg, DR_STMT (dr1)) | |
1380 | > rdg_vertex_for_stmt (rdg, DR_STMT (dr2))) | |
1381 | std::swap (dr1, dr2); | |
1382 | ||
1383 | ddr = get_data_dependence (rdg, dr1, dr2); | |
1384 | ||
1385 | /* In case of no data dependence. */ | |
1386 | if (DDR_ARE_DEPENDENT (ddr) == chrec_known) | |
1387 | return false; | |
1388 | /* For unknown data dependence or known data dependence which can't be | |
1389 | expressed in classic distance vector, we check if it can be resolved | |
1390 | by runtime alias check. If yes, we still consider data dependence | |
1391 | as won't introduce data dependence cycle. */ | |
1392 | else if (DDR_ARE_DEPENDENT (ddr) == chrec_dont_know | |
1393 | || DDR_NUM_DIST_VECTS (ddr) == 0) | |
1394 | return !runtime_alias_check_p (ddr, NULL, true); | |
1395 | else if (DDR_NUM_DIST_VECTS (ddr) > 1) | |
1396 | return true; | |
1397 | else if (DDR_REVERSED_P (ddr) | |
1398 | || lambda_vector_zerop (DDR_DIST_VECT (ddr, 0), 1)) | |
1399 | return false; | |
1400 | ||
1401 | return true; | |
1402 | } | |
1403 | ||
eef99cd9 GB |
1404 | void |
1405 | loop_distribution::update_type_for_merge (struct graph *rdg, | |
1406 | partition *partition1, | |
1407 | partition *partition2) | |
f1eb4621 BC |
1408 | { |
1409 | unsigned i, j; | |
1410 | bitmap_iterator bi, bj; | |
1411 | data_reference_p dr1, dr2; | |
1412 | ||
1413 | EXECUTE_IF_SET_IN_BITMAP (partition1->datarefs, 0, i, bi) | |
1414 | { | |
1415 | unsigned start = (partition1 == partition2) ? i + 1 : 0; | |
1416 | ||
1417 | dr1 = datarefs_vec[i]; | |
1418 | EXECUTE_IF_SET_IN_BITMAP (partition2->datarefs, start, j, bj) | |
1419 | { | |
1420 | dr2 = datarefs_vec[j]; | |
1421 | if (DR_IS_READ (dr1) && DR_IS_READ (dr2)) | |
1422 | continue; | |
1423 | ||
1424 | /* Partition can only be executed sequentially if there is any | |
1425 | data dependence cycle. */ | |
1426 | if (data_dep_in_cycle_p (rdg, dr1, dr2)) | |
1427 | { | |
1428 | partition1->type = PTYPE_SEQUENTIAL; | |
1429 | return; | |
1430 | } | |
1431 | } | |
1432 | } | |
1433 | } | |
1434 | ||
eef99cd9 GB |
1435 | partition * |
1436 | loop_distribution::build_rdg_partition_for_vertex (struct graph *rdg, int v) | |
dea61d92 | 1437 | { |
a7a44c07 | 1438 | partition *partition = partition_alloc (); |
00f96dc9 | 1439 | auto_vec<int, 3> nodes; |
a7a44c07 | 1440 | unsigned i, j; |
dea61d92 | 1441 | int x; |
a7a44c07 | 1442 | data_reference_p dr; |
dea61d92 | 1443 | |
174ec470 | 1444 | graphds_dfs (rdg, &v, 1, &nodes, false, NULL); |
dea61d92 | 1445 | |
9771b263 | 1446 | FOR_EACH_VEC_ELT (nodes, i, x) |
24f161fd RB |
1447 | { |
1448 | bitmap_set_bit (partition->stmts, x); | |
a7a44c07 BC |
1449 | |
1450 | for (j = 0; RDG_DATAREFS (rdg, x).iterate (j, &dr); ++j) | |
1451 | { | |
1452 | unsigned idx = (unsigned) DR_INDEX (dr); | |
1453 | gcc_assert (idx < datarefs_vec.length ()); | |
1454 | ||
f1eb4621 BC |
1455 | /* Partition can only be executed sequentially if there is any |
1456 | unknown data reference. */ | |
1457 | if (!DR_BASE_ADDRESS (dr) || !DR_OFFSET (dr) | |
1458 | || !DR_INIT (dr) || !DR_STEP (dr)) | |
1459 | partition->type = PTYPE_SEQUENTIAL; | |
1460 | ||
a7a44c07 BC |
1461 | bitmap_set_bit (partition->datarefs, idx); |
1462 | } | |
24f161fd | 1463 | } |
dea61d92 | 1464 | |
f1eb4621 BC |
1465 | if (partition->type == PTYPE_SEQUENTIAL) |
1466 | return partition; | |
1467 | ||
1468 | /* Further check if any data dependence prevents us from executing the | |
1469 | partition parallelly. */ | |
1470 | update_type_for_merge (rdg, partition, partition); | |
1471 | ||
dea61d92 SP |
1472 | return partition; |
1473 | } | |
1474 | ||
85aa9ed6 BC |
1475 | /* Given PARTITION of LOOP and RDG, record single load/store data references |
1476 | for builtin partition in SRC_DR/DST_DR, return false if there is no such | |
939cf90f | 1477 | data references. */ |
cfee318d | 1478 | |
939cf90f | 1479 | static bool |
99b1c316 | 1480 | find_single_drs (class loop *loop, struct graph *rdg, partition *partition, |
939cf90f | 1481 | data_reference_p *dst_dr, data_reference_p *src_dr) |
cfee318d | 1482 | { |
30d55936 | 1483 | unsigned i; |
939cf90f BC |
1484 | data_reference_p single_ld = NULL, single_st = NULL; |
1485 | bitmap_iterator bi; | |
b9fc0497 | 1486 | |
30d55936 RG |
1487 | EXECUTE_IF_SET_IN_BITMAP (partition->stmts, 0, i, bi) |
1488 | { | |
355fe088 | 1489 | gimple *stmt = RDG_STMT (rdg, i); |
d0582dc1 | 1490 | data_reference_p dr; |
30d55936 RG |
1491 | |
1492 | if (gimple_code (stmt) == GIMPLE_PHI) | |
1493 | continue; | |
1494 | ||
1495 | /* Any scalar stmts are ok. */ | |
1496 | if (!gimple_vuse (stmt)) | |
1497 | continue; | |
1498 | ||
d0582dc1 RG |
1499 | /* Otherwise just regular loads/stores. */ |
1500 | if (!gimple_assign_single_p (stmt)) | |
939cf90f | 1501 | return false; |
d0582dc1 RG |
1502 | |
1503 | /* But exactly one store and/or load. */ | |
939cf90f | 1504 | for (unsigned j = 0; RDG_DATAREFS (rdg, i).iterate (j, &dr); ++j) |
30d55936 | 1505 | { |
d002d099 JJ |
1506 | tree type = TREE_TYPE (DR_REF (dr)); |
1507 | ||
1508 | /* The memset, memcpy and memmove library calls are only | |
1509 | able to deal with generic address space. */ | |
1510 | if (!ADDR_SPACE_GENERIC_P (TYPE_ADDR_SPACE (type))) | |
939cf90f | 1511 | return false; |
d002d099 | 1512 | |
d0582dc1 RG |
1513 | if (DR_IS_READ (dr)) |
1514 | { | |
939cf90f BC |
1515 | if (single_ld != NULL) |
1516 | return false; | |
1517 | single_ld = dr; | |
d0582dc1 RG |
1518 | } |
1519 | else | |
1520 | { | |
939cf90f BC |
1521 | if (single_st != NULL) |
1522 | return false; | |
1523 | single_st = dr; | |
d0582dc1 | 1524 | } |
30d55936 | 1525 | } |
30d55936 RG |
1526 | } |
1527 | ||
939cf90f BC |
1528 | if (!single_st) |
1529 | return false; | |
1530 | ||
1531 | /* Bail out if this is a bitfield memory reference. */ | |
1532 | if (TREE_CODE (DR_REF (single_st)) == COMPONENT_REF | |
1533 | && DECL_BIT_FIELD (TREE_OPERAND (DR_REF (single_st), 1))) | |
1534 | return false; | |
818625cf | 1535 | |
85aa9ed6 BC |
1536 | /* Data reference must be executed exactly once per iteration of each |
1537 | loop in the loop nest. We only need to check dominance information | |
1538 | against the outermost one in a perfect loop nest because a bb can't | |
1539 | dominate outermost loop's latch without dominating inner loop's. */ | |
939cf90f | 1540 | basic_block bb_st = gimple_bb (DR_STMT (single_st)); |
85aa9ed6 | 1541 | if (!dominated_by_p (CDI_DOMINATORS, loop->latch, bb_st)) |
939cf90f BC |
1542 | return false; |
1543 | ||
1544 | if (single_ld) | |
163aa51b | 1545 | { |
939cf90f BC |
1546 | gimple *store = DR_STMT (single_st), *load = DR_STMT (single_ld); |
1547 | /* Direct aggregate copy or via an SSA name temporary. */ | |
1548 | if (load != store | |
1549 | && gimple_assign_lhs (load) != gimple_assign_rhs1 (store)) | |
1550 | return false; | |
163aa51b | 1551 | |
939cf90f BC |
1552 | /* Bail out if this is a bitfield memory reference. */ |
1553 | if (TREE_CODE (DR_REF (single_ld)) == COMPONENT_REF | |
1554 | && DECL_BIT_FIELD (TREE_OPERAND (DR_REF (single_ld), 1))) | |
1555 | return false; | |
1556 | ||
1557 | /* Load and store must be in the same loop nest. */ | |
1558 | basic_block bb_ld = gimple_bb (DR_STMT (single_ld)); | |
85aa9ed6 | 1559 | if (bb_st->loop_father != bb_ld->loop_father) |
939cf90f BC |
1560 | return false; |
1561 | ||
85aa9ed6 BC |
1562 | /* Data reference must be executed exactly once per iteration. |
1563 | Same as single_st, we only need to check against the outermost | |
1564 | loop. */ | |
1565 | if (!dominated_by_p (CDI_DOMINATORS, loop->latch, bb_ld)) | |
939cf90f BC |
1566 | return false; |
1567 | ||
85aa9ed6 | 1568 | edge e = single_exit (bb_st->loop_father); |
939cf90f BC |
1569 | bool dom_ld = dominated_by_p (CDI_DOMINATORS, e->src, bb_ld); |
1570 | bool dom_st = dominated_by_p (CDI_DOMINATORS, e->src, bb_st); | |
1571 | if (dom_ld != dom_st) | |
1572 | return false; | |
1573 | } | |
1574 | ||
1575 | *src_dr = single_ld; | |
1576 | *dst_dr = single_st; | |
1577 | return true; | |
1578 | } | |
1579 | ||
1580 | /* Given data reference DR in LOOP_NEST, this function checks the enclosing | |
1581 | loops from inner to outer to see if loop's step equals to access size at | |
5955438a BC |
1582 | each level of loop. Return 2 if we can prove this at all level loops; |
1583 | record access base and size in BASE and SIZE; save loop's step at each | |
1584 | level of loop in STEPS if it is not null. For example: | |
939cf90f BC |
1585 | |
1586 | int arr[100][100][100]; | |
1587 | for (i = 0; i < 100; i++) ;steps[2] = 40000 | |
1588 | for (j = 100; j > 0; j--) ;steps[1] = -400 | |
1589 | for (k = 0; k < 100; k++) ;steps[0] = 4 | |
5955438a | 1590 | arr[i][j - 1][k] = 0; ;base = &arr, size = 4000000 |
939cf90f | 1591 | |
5955438a BC |
1592 | Return 1 if we can prove the equality at the innermost loop, but not all |
1593 | level loops. In this case, no information is recorded. | |
1594 | ||
1595 | Return 0 if no equality can be proven at any level loops. */ | |
1596 | ||
1597 | static int | |
939cf90f BC |
1598 | compute_access_range (loop_p loop_nest, data_reference_p dr, tree *base, |
1599 | tree *size, vec<tree> *steps = NULL) | |
1600 | { | |
1601 | location_t loc = gimple_location (DR_STMT (dr)); | |
1602 | basic_block bb = gimple_bb (DR_STMT (dr)); | |
99b1c316 | 1603 | class loop *loop = bb->loop_father; |
939cf90f BC |
1604 | tree ref = DR_REF (dr); |
1605 | tree access_base = build_fold_addr_expr (ref); | |
1606 | tree access_size = TYPE_SIZE_UNIT (TREE_TYPE (ref)); | |
5955438a | 1607 | int res = 0; |
939cf90f BC |
1608 | |
1609 | do { | |
1610 | tree scev_fn = analyze_scalar_evolution (loop, access_base); | |
1611 | if (TREE_CODE (scev_fn) != POLYNOMIAL_CHREC) | |
5955438a | 1612 | return res; |
163aa51b | 1613 | |
939cf90f BC |
1614 | access_base = CHREC_LEFT (scev_fn); |
1615 | if (tree_contains_chrecs (access_base, NULL)) | |
5955438a | 1616 | return res; |
939cf90f BC |
1617 | |
1618 | tree scev_step = CHREC_RIGHT (scev_fn); | |
1619 | /* Only support constant steps. */ | |
1620 | if (TREE_CODE (scev_step) != INTEGER_CST) | |
5955438a | 1621 | return res; |
939cf90f BC |
1622 | |
1623 | enum ev_direction access_dir = scev_direction (scev_fn); | |
1624 | if (access_dir == EV_DIR_UNKNOWN) | |
5955438a | 1625 | return res; |
939cf90f BC |
1626 | |
1627 | if (steps != NULL) | |
1628 | steps->safe_push (scev_step); | |
1629 | ||
1630 | scev_step = fold_convert_loc (loc, sizetype, scev_step); | |
1631 | /* Compute absolute value of scev step. */ | |
1632 | if (access_dir == EV_DIR_DECREASES) | |
1633 | scev_step = fold_build1_loc (loc, NEGATE_EXPR, sizetype, scev_step); | |
1634 | ||
1635 | /* At each level of loop, scev step must equal to access size. In other | |
1636 | words, DR must access consecutive memory between loop iterations. */ | |
1637 | if (!operand_equal_p (scev_step, access_size, 0)) | |
5955438a BC |
1638 | return res; |
1639 | ||
1640 | /* Access stride can be computed for data reference at least for the | |
1641 | innermost loop. */ | |
1642 | res = 1; | |
939cf90f BC |
1643 | |
1644 | /* Compute DR's execution times in loop. */ | |
1645 | tree niters = number_of_latch_executions (loop); | |
1646 | niters = fold_convert_loc (loc, sizetype, niters); | |
1647 | if (dominated_by_p (CDI_DOMINATORS, single_exit (loop)->src, bb)) | |
1648 | niters = size_binop_loc (loc, PLUS_EXPR, niters, size_one_node); | |
1649 | ||
1650 | /* Compute DR's overall access size in loop. */ | |
1651 | access_size = fold_build2_loc (loc, MULT_EXPR, sizetype, | |
1652 | niters, scev_step); | |
1653 | /* Adjust base address in case of negative step. */ | |
1654 | if (access_dir == EV_DIR_DECREASES) | |
163aa51b | 1655 | { |
939cf90f BC |
1656 | tree adj = fold_build2_loc (loc, MINUS_EXPR, sizetype, |
1657 | scev_step, access_size); | |
1658 | access_base = fold_build_pointer_plus_loc (loc, access_base, adj); | |
163aa51b | 1659 | } |
939cf90f BC |
1660 | } while (loop != loop_nest && (loop = loop_outer (loop)) != NULL); |
1661 | ||
1662 | *base = access_base; | |
1663 | *size = access_size; | |
5955438a BC |
1664 | /* Access stride can be computed for data reference at each level loop. */ |
1665 | return 2; | |
939cf90f BC |
1666 | } |
1667 | ||
1668 | /* Allocate and return builtin struct. Record information like DST_DR, | |
1669 | SRC_DR, DST_BASE, SRC_BASE and SIZE in the allocated struct. */ | |
1670 | ||
1671 | static struct builtin_info * | |
1672 | alloc_builtin (data_reference_p dst_dr, data_reference_p src_dr, | |
1673 | tree dst_base, tree src_base, tree size) | |
1674 | { | |
1675 | struct builtin_info *builtin = XNEW (struct builtin_info); | |
1676 | builtin->dst_dr = dst_dr; | |
1677 | builtin->src_dr = src_dr; | |
1678 | builtin->dst_base = dst_base; | |
1679 | builtin->src_base = src_base; | |
1680 | builtin->size = size; | |
1681 | return builtin; | |
1682 | } | |
1683 | ||
1684 | /* Given data reference DR in loop nest LOOP, classify if it forms builtin | |
1685 | memset call. */ | |
1686 | ||
1687 | static void | |
1688 | classify_builtin_st (loop_p loop, partition *partition, data_reference_p dr) | |
1689 | { | |
1690 | gimple *stmt = DR_STMT (dr); | |
1691 | tree base, size, rhs = gimple_assign_rhs1 (stmt); | |
1692 | ||
1693 | if (const_with_all_bytes_same (rhs) == -1 | |
1694 | && (!INTEGRAL_TYPE_P (TREE_TYPE (rhs)) | |
1695 | || (TYPE_MODE (TREE_TYPE (rhs)) | |
1696 | != TYPE_MODE (unsigned_char_type_node)))) | |
1697 | return; | |
1698 | ||
1699 | if (TREE_CODE (rhs) == SSA_NAME | |
1700 | && !SSA_NAME_IS_DEFAULT_DEF (rhs) | |
1701 | && flow_bb_inside_loop_p (loop, gimple_bb (SSA_NAME_DEF_STMT (rhs)))) | |
1702 | return; | |
1703 | ||
5955438a BC |
1704 | int res = compute_access_range (loop, dr, &base, &size); |
1705 | if (res == 0) | |
939cf90f | 1706 | return; |
5955438a BC |
1707 | if (res == 1) |
1708 | { | |
1709 | partition->kind = PKIND_PARTIAL_MEMSET; | |
1710 | return; | |
1711 | } | |
939cf90f | 1712 | |
d2fd6a04 RS |
1713 | poly_uint64 base_offset; |
1714 | unsigned HOST_WIDE_INT const_base_offset; | |
1715 | tree base_base = strip_offset (base, &base_offset); | |
1716 | if (!base_offset.is_constant (&const_base_offset)) | |
1717 | return; | |
1718 | ||
957f0d8f BC |
1719 | struct builtin_info *builtin; |
1720 | builtin = alloc_builtin (dr, NULL, base, NULL_TREE, size); | |
d2fd6a04 RS |
1721 | builtin->dst_base_base = base_base; |
1722 | builtin->dst_base_offset = const_base_offset; | |
957f0d8f | 1723 | partition->builtin = builtin; |
939cf90f BC |
1724 | partition->kind = PKIND_MEMSET; |
1725 | } | |
1726 | ||
1727 | /* Given data references DST_DR and SRC_DR in loop nest LOOP and RDG, classify | |
1728 | if it forms builtin memcpy or memmove call. */ | |
1729 | ||
eef99cd9 GB |
1730 | void |
1731 | loop_distribution::classify_builtin_ldst (loop_p loop, struct graph *rdg, | |
1732 | partition *partition, | |
1733 | data_reference_p dst_dr, | |
1734 | data_reference_p src_dr) | |
939cf90f BC |
1735 | { |
1736 | tree base, size, src_base, src_size; | |
1737 | auto_vec<tree> dst_steps, src_steps; | |
1738 | ||
5955438a BC |
1739 | /* Compute access range of both load and store. */ |
1740 | int res = compute_access_range (loop, dst_dr, &base, &size, &dst_steps); | |
1741 | if (res != 2) | |
1742 | return; | |
1743 | res = compute_access_range (loop, src_dr, &src_base, &src_size, &src_steps); | |
1744 | if (res != 2) | |
1745 | return; | |
1746 | ||
1747 | /* They much have the same access size. */ | |
1748 | if (!operand_equal_p (size, src_size, 0)) | |
939cf90f BC |
1749 | return; |
1750 | ||
1751 | /* Load and store in loop nest must access memory in the same way, i.e, | |
1752 | their must have the same steps in each loop of the nest. */ | |
1753 | if (dst_steps.length () != src_steps.length ()) | |
1754 | return; | |
1755 | for (unsigned i = 0; i < dst_steps.length (); ++i) | |
1756 | if (!operand_equal_p (dst_steps[i], src_steps[i], 0)) | |
1757 | return; | |
1758 | ||
1759 | /* Now check that if there is a dependence. */ | |
1760 | ddr_p ddr = get_data_dependence (rdg, src_dr, dst_dr); | |
1761 | ||
1762 | /* Classify as memcpy if no dependence between load and store. */ | |
1763 | if (DDR_ARE_DEPENDENT (ddr) == chrec_known) | |
1764 | { | |
1765 | partition->builtin = alloc_builtin (dst_dr, src_dr, base, src_base, size); | |
1766 | partition->kind = PKIND_MEMCPY; | |
1767 | return; | |
163aa51b BC |
1768 | } |
1769 | ||
939cf90f BC |
1770 | /* Can't do memmove in case of unknown dependence or dependence without |
1771 | classical distance vector. */ | |
1772 | if (DDR_ARE_DEPENDENT (ddr) == chrec_dont_know | |
1773 | || DDR_NUM_DIST_VECTS (ddr) == 0) | |
1774 | return; | |
818625cf | 1775 | |
939cf90f BC |
1776 | unsigned i; |
1777 | lambda_vector dist_v; | |
1778 | int num_lev = (DDR_LOOP_NEST (ddr)).length (); | |
1779 | FOR_EACH_VEC_ELT (DDR_DIST_VECTS (ddr), i, dist_v) | |
d0582dc1 | 1780 | { |
939cf90f BC |
1781 | unsigned dep_lev = dependence_level (dist_v, num_lev); |
1782 | /* Can't do memmove if load depends on store. */ | |
1783 | if (dep_lev > 0 && dist_v[dep_lev - 1] > 0 && !DDR_REVERSED_P (ddr)) | |
d0582dc1 | 1784 | return; |
d0582dc1 | 1785 | } |
939cf90f BC |
1786 | |
1787 | partition->builtin = alloc_builtin (dst_dr, src_dr, base, src_base, size); | |
1788 | partition->kind = PKIND_MEMMOVE; | |
1789 | return; | |
1790 | } | |
1791 | ||
eef99cd9 GB |
1792 | bool |
1793 | loop_distribution::classify_partition (loop_p loop, | |
1794 | struct graph *rdg, partition *partition, | |
1795 | bitmap stmt_in_all_partitions) | |
939cf90f BC |
1796 | { |
1797 | bitmap_iterator bi; | |
1798 | unsigned i; | |
1799 | data_reference_p single_ld = NULL, single_st = NULL; | |
1800 | bool volatiles_p = false, has_reduction = false; | |
1801 | ||
1802 | EXECUTE_IF_SET_IN_BITMAP (partition->stmts, 0, i, bi) | |
d0582dc1 | 1803 | { |
939cf90f | 1804 | gimple *stmt = RDG_STMT (rdg, i); |
17c5cbdf | 1805 | |
939cf90f BC |
1806 | if (gimple_has_volatile_ops (stmt)) |
1807 | volatiles_p = true; | |
17c5cbdf | 1808 | |
939cf90f BC |
1809 | /* If the stmt is not included by all partitions and there is uses |
1810 | outside of the loop, then mark the partition as reduction. */ | |
1811 | if (stmt_has_scalar_dependences_outside_loop (loop, stmt)) | |
1812 | { | |
1813 | /* Due to limitation in the transform phase we have to fuse all | |
1814 | reduction partitions. As a result, this could cancel valid | |
1815 | loop distribution especially for loop that induction variable | |
1816 | is used outside of loop. To workaround this issue, we skip | |
1817 | marking partition as reudction if the reduction stmt belongs | |
1818 | to all partitions. In such case, reduction will be computed | |
1819 | correctly no matter how partitions are fused/distributed. */ | |
1820 | if (!bitmap_bit_p (stmt_in_all_partitions, i)) | |
e7484357 RB |
1821 | partition->reduction_p = true; |
1822 | else | |
1823 | has_reduction = true; | |
f20132e7 | 1824 | } |
d0582dc1 | 1825 | } |
939cf90f | 1826 | |
e7484357 RB |
1827 | /* Simple workaround to prevent classifying the partition as builtin |
1828 | if it contains any use outside of loop. For the case where all | |
1829 | partitions have the reduction this simple workaround is delayed | |
1830 | to only affect the last partition. */ | |
1831 | if (partition->reduction_p) | |
1832 | return has_reduction; | |
1833 | ||
939cf90f BC |
1834 | /* Perform general partition disqualification for builtins. */ |
1835 | if (volatiles_p | |
939cf90f | 1836 | || !flag_tree_loop_distribute_patterns) |
e7484357 | 1837 | return has_reduction; |
939cf90f BC |
1838 | |
1839 | /* Find single load/store data references for builtin partition. */ | |
85aa9ed6 | 1840 | if (!find_single_drs (loop, rdg, partition, &single_st, &single_ld)) |
e7484357 | 1841 | return has_reduction; |
939cf90f | 1842 | |
dbd59515 RB |
1843 | partition->loc = gimple_location (DR_STMT (single_st)); |
1844 | ||
939cf90f BC |
1845 | /* Classify the builtin kind. */ |
1846 | if (single_ld == NULL) | |
1847 | classify_builtin_st (loop, partition, single_st); | |
1848 | else | |
1849 | classify_builtin_ldst (loop, rdg, partition, single_st, single_ld); | |
e7484357 | 1850 | return has_reduction; |
cfee318d SP |
1851 | } |
1852 | ||
eef99cd9 GB |
1853 | bool |
1854 | loop_distribution::share_memory_accesses (struct graph *rdg, | |
95f7d11b | 1855 | partition *partition1, partition *partition2) |
cfee318d | 1856 | { |
95f7d11b | 1857 | unsigned i, j; |
cfee318d | 1858 | bitmap_iterator bi, bj; |
95f7d11b | 1859 | data_reference_p dr1, dr2; |
1fa0c180 RG |
1860 | |
1861 | /* First check whether in the intersection of the two partitions are | |
1862 | any loads or stores. Common loads are the situation that happens | |
1863 | most often. */ | |
1864 | EXECUTE_IF_AND_IN_BITMAP (partition1->stmts, partition2->stmts, 0, i, bi) | |
1865 | if (RDG_MEM_WRITE_STMT (rdg, i) | |
1866 | || RDG_MEM_READS_STMT (rdg, i)) | |
1867 | return true; | |
cfee318d | 1868 | |
95f7d11b BC |
1869 | /* Then check whether the two partitions access the same memory object. */ |
1870 | EXECUTE_IF_SET_IN_BITMAP (partition1->datarefs, 0, i, bi) | |
1871 | { | |
1872 | dr1 = datarefs_vec[i]; | |
1873 | ||
1874 | if (!DR_BASE_ADDRESS (dr1) | |
1875 | || !DR_OFFSET (dr1) || !DR_INIT (dr1) || !DR_STEP (dr1)) | |
1876 | continue; | |
1877 | ||
1878 | EXECUTE_IF_SET_IN_BITMAP (partition2->datarefs, 0, j, bj) | |
1879 | { | |
1880 | dr2 = datarefs_vec[j]; | |
1881 | ||
1882 | if (!DR_BASE_ADDRESS (dr2) | |
1883 | || !DR_OFFSET (dr2) || !DR_INIT (dr2) || !DR_STEP (dr2)) | |
1884 | continue; | |
1885 | ||
1886 | if (operand_equal_p (DR_BASE_ADDRESS (dr1), DR_BASE_ADDRESS (dr2), 0) | |
1887 | && operand_equal_p (DR_OFFSET (dr1), DR_OFFSET (dr2), 0) | |
1888 | && operand_equal_p (DR_INIT (dr1), DR_INIT (dr2), 0) | |
1889 | && operand_equal_p (DR_STEP (dr1), DR_STEP (dr2), 0)) | |
1890 | return true; | |
1891 | } | |
1892 | } | |
cfee318d SP |
1893 | |
1894 | return false; | |
1895 | } | |
1896 | ||
4a52eb19 BC |
1897 | /* For each seed statement in STARTING_STMTS, this function builds |
1898 | partition for it by adding depended statements according to RDG. | |
1899 | All partitions are recorded in PARTITIONS. */ | |
dea61d92 | 1900 | |
eef99cd9 GB |
1901 | void |
1902 | loop_distribution::rdg_build_partitions (struct graph *rdg, | |
1903 | vec<gimple *> starting_stmts, | |
1904 | vec<partition *> *partitions) | |
dea61d92 | 1905 | { |
0e3de1d4 | 1906 | auto_bitmap processed; |
2fd5894f | 1907 | int i; |
355fe088 | 1908 | gimple *stmt; |
dea61d92 | 1909 | |
2fd5894f | 1910 | FOR_EACH_VEC_ELT (starting_stmts, i, stmt) |
dea61d92 | 1911 | { |
2fd5894f RB |
1912 | int v = rdg_vertex_for_stmt (rdg, stmt); |
1913 | ||
1914 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
1915 | fprintf (dump_file, | |
1916 | "ldist asked to generate code for vertex %d\n", v); | |
b8698a0f | 1917 | |
24f161fd RB |
1918 | /* If the vertex is already contained in another partition so |
1919 | is the partition rooted at it. */ | |
dea61d92 SP |
1920 | if (bitmap_bit_p (processed, v)) |
1921 | continue; | |
b8698a0f | 1922 | |
526ceb68 | 1923 | partition *partition = build_rdg_partition_for_vertex (rdg, v); |
24f161fd | 1924 | bitmap_ior_into (processed, partition->stmts); |
dea61d92 | 1925 | |
826a536d | 1926 | if (dump_file && (dump_flags & TDF_DETAILS)) |
dea61d92 | 1927 | { |
f1eb4621 BC |
1928 | fprintf (dump_file, "ldist creates useful %s partition:\n", |
1929 | partition->type == PTYPE_PARALLEL ? "parallel" : "sequent"); | |
1930 | bitmap_print (dump_file, partition->stmts, " ", "\n"); | |
dea61d92 | 1931 | } |
826a536d RB |
1932 | |
1933 | partitions->safe_push (partition); | |
dea61d92 SP |
1934 | } |
1935 | ||
83a95546 RB |
1936 | /* All vertices should have been assigned to at least one partition now, |
1937 | other than vertices belonging to dead code. */ | |
dea61d92 SP |
1938 | } |
1939 | ||
1940 | /* Dump to FILE the PARTITIONS. */ | |
1941 | ||
1942 | static void | |
526ceb68 | 1943 | dump_rdg_partitions (FILE *file, vec<partition *> partitions) |
dea61d92 SP |
1944 | { |
1945 | int i; | |
526ceb68 | 1946 | partition *partition; |
dea61d92 | 1947 | |
9771b263 | 1948 | FOR_EACH_VEC_ELT (partitions, i, partition) |
c61f8985 | 1949 | debug_bitmap_file (file, partition->stmts); |
dea61d92 SP |
1950 | } |
1951 | ||
1952 | /* Debug PARTITIONS. */ | |
526ceb68 | 1953 | extern void debug_rdg_partitions (vec<partition *> ); |
dea61d92 | 1954 | |
24e47c76 | 1955 | DEBUG_FUNCTION void |
526ceb68 | 1956 | debug_rdg_partitions (vec<partition *> partitions) |
dea61d92 SP |
1957 | { |
1958 | dump_rdg_partitions (stderr, partitions); | |
1959 | } | |
1960 | ||
2b8aee8e SP |
1961 | /* Returns the number of read and write operations in the RDG. */ |
1962 | ||
1963 | static int | |
1964 | number_of_rw_in_rdg (struct graph *rdg) | |
1965 | { | |
1966 | int i, res = 0; | |
1967 | ||
1968 | for (i = 0; i < rdg->n_vertices; i++) | |
1969 | { | |
1970 | if (RDG_MEM_WRITE_STMT (rdg, i)) | |
1971 | ++res; | |
1972 | ||
1973 | if (RDG_MEM_READS_STMT (rdg, i)) | |
1974 | ++res; | |
1975 | } | |
1976 | ||
1977 | return res; | |
1978 | } | |
1979 | ||
1980 | /* Returns the number of read and write operations in a PARTITION of | |
1981 | the RDG. */ | |
1982 | ||
1983 | static int | |
526ceb68 | 1984 | number_of_rw_in_partition (struct graph *rdg, partition *partition) |
2b8aee8e SP |
1985 | { |
1986 | int res = 0; | |
1987 | unsigned i; | |
1988 | bitmap_iterator ii; | |
1989 | ||
c61f8985 | 1990 | EXECUTE_IF_SET_IN_BITMAP (partition->stmts, 0, i, ii) |
2b8aee8e SP |
1991 | { |
1992 | if (RDG_MEM_WRITE_STMT (rdg, i)) | |
1993 | ++res; | |
1994 | ||
1995 | if (RDG_MEM_READS_STMT (rdg, i)) | |
1996 | ++res; | |
1997 | } | |
1998 | ||
1999 | return res; | |
2000 | } | |
2001 | ||
2002 | /* Returns true when one of the PARTITIONS contains all the read or | |
2003 | write operations of RDG. */ | |
2004 | ||
2005 | static bool | |
9771b263 | 2006 | partition_contains_all_rw (struct graph *rdg, |
526ceb68 | 2007 | vec<partition *> partitions) |
2b8aee8e SP |
2008 | { |
2009 | int i; | |
526ceb68 | 2010 | partition *partition; |
2b8aee8e SP |
2011 | int nrw = number_of_rw_in_rdg (rdg); |
2012 | ||
9771b263 | 2013 | FOR_EACH_VEC_ELT (partitions, i, partition) |
2b8aee8e SP |
2014 | if (nrw == number_of_rw_in_partition (rdg, partition)) |
2015 | return true; | |
2016 | ||
2017 | return false; | |
2018 | } | |
2019 | ||
eef99cd9 GB |
2020 | int |
2021 | loop_distribution::pg_add_dependence_edges (struct graph *rdg, int dir, | |
a8745cc2 | 2022 | bitmap drs1, bitmap drs2, vec<ddr_p> *alias_ddrs) |
447f3223 | 2023 | { |
a7a44c07 BC |
2024 | unsigned i, j; |
2025 | bitmap_iterator bi, bj; | |
2026 | data_reference_p dr1, dr2, saved_dr1; | |
447f3223 RB |
2027 | |
2028 | /* dependence direction - 0 is no dependence, -1 is back, | |
2029 | 1 is forth, 2 is both (we can stop then, merging will occur). */ | |
a7a44c07 BC |
2030 | EXECUTE_IF_SET_IN_BITMAP (drs1, 0, i, bi) |
2031 | { | |
2032 | dr1 = datarefs_vec[i]; | |
2033 | ||
2034 | EXECUTE_IF_SET_IN_BITMAP (drs2, 0, j, bj) | |
2035 | { | |
a8745cc2 BC |
2036 | int res, this_dir = 1; |
2037 | ddr_p ddr; | |
2038 | ||
a7a44c07 BC |
2039 | dr2 = datarefs_vec[j]; |
2040 | ||
2041 | /* Skip all <read, read> data dependence. */ | |
2042 | if (DR_IS_READ (dr1) && DR_IS_READ (dr2)) | |
2043 | continue; | |
2044 | ||
2045 | saved_dr1 = dr1; | |
a8745cc2 | 2046 | /* Re-shuffle data-refs to be in topological order. */ |
a7a44c07 BC |
2047 | if (rdg_vertex_for_stmt (rdg, DR_STMT (dr1)) |
2048 | > rdg_vertex_for_stmt (rdg, DR_STMT (dr2))) | |
2049 | { | |
2050 | std::swap (dr1, dr2); | |
2051 | this_dir = -this_dir; | |
2052 | } | |
17c5cbdf | 2053 | ddr = get_data_dependence (rdg, dr1, dr2); |
a7a44c07 | 2054 | if (DDR_ARE_DEPENDENT (ddr) == chrec_dont_know) |
a8745cc2 BC |
2055 | { |
2056 | this_dir = 0; | |
2057 | res = data_ref_compare_tree (DR_BASE_ADDRESS (dr1), | |
2058 | DR_BASE_ADDRESS (dr2)); | |
2059 | /* Be conservative. If data references are not well analyzed, | |
2060 | or the two data references have the same base address and | |
2061 | offset, add dependence and consider it alias to each other. | |
67914693 | 2062 | In other words, the dependence cannot be resolved by |
a8745cc2 BC |
2063 | runtime alias check. */ |
2064 | if (!DR_BASE_ADDRESS (dr1) || !DR_BASE_ADDRESS (dr2) | |
2065 | || !DR_OFFSET (dr1) || !DR_OFFSET (dr2) | |
2066 | || !DR_INIT (dr1) || !DR_INIT (dr2) | |
2067 | || !DR_STEP (dr1) || !tree_fits_uhwi_p (DR_STEP (dr1)) | |
2068 | || !DR_STEP (dr2) || !tree_fits_uhwi_p (DR_STEP (dr2)) | |
2069 | || res == 0) | |
2070 | this_dir = 2; | |
2071 | /* Data dependence could be resolved by runtime alias check, | |
2072 | record it in ALIAS_DDRS. */ | |
2073 | else if (alias_ddrs != NULL) | |
2074 | alias_ddrs->safe_push (ddr); | |
2075 | /* Or simply ignore it. */ | |
2076 | } | |
a7a44c07 BC |
2077 | else if (DDR_ARE_DEPENDENT (ddr) == NULL_TREE) |
2078 | { | |
2079 | if (DDR_REVERSED_P (ddr)) | |
a8745cc2 BC |
2080 | this_dir = -this_dir; |
2081 | ||
a7a44c07 | 2082 | /* Known dependences can still be unordered througout the |
f6e1a4cd BC |
2083 | iteration space, see gcc.dg/tree-ssa/ldist-16.c and |
2084 | gcc.dg/tree-ssa/pr94969.c. */ | |
a7a44c07 BC |
2085 | if (DDR_NUM_DIST_VECTS (ddr) != 1) |
2086 | this_dir = 2; | |
2087 | /* If the overlap is exact preserve stmt order. */ | |
0df7c778 BC |
2088 | else if (lambda_vector_zerop (DDR_DIST_VECT (ddr, 0), |
2089 | DDR_NB_LOOPS (ddr))) | |
a7a44c07 | 2090 | ; |
a8745cc2 BC |
2091 | /* Else as the distance vector is lexicographic positive swap |
2092 | the dependence direction. */ | |
a7a44c07 | 2093 | else |
a8745cc2 | 2094 | this_dir = -this_dir; |
a7a44c07 BC |
2095 | } |
2096 | else | |
2097 | this_dir = 0; | |
a7a44c07 BC |
2098 | if (this_dir == 2) |
2099 | return 2; | |
2100 | else if (dir == 0) | |
2101 | dir = this_dir; | |
2102 | else if (this_dir != 0 && dir != this_dir) | |
2103 | return 2; | |
2104 | /* Shuffle "back" dr1. */ | |
2105 | dr1 = saved_dr1; | |
2106 | } | |
2107 | } | |
447f3223 RB |
2108 | return dir; |
2109 | } | |
2110 | ||
2111 | /* Compare postorder number of the partition graph vertices V1 and V2. */ | |
2112 | ||
2113 | static int | |
2114 | pgcmp (const void *v1_, const void *v2_) | |
2115 | { | |
2116 | const vertex *v1 = (const vertex *)v1_; | |
2117 | const vertex *v2 = (const vertex *)v2_; | |
2118 | return v2->post - v1->post; | |
2119 | } | |
2fd5894f | 2120 | |
a8745cc2 BC |
2121 | /* Data attached to vertices of partition dependence graph. */ |
2122 | struct pg_vdata | |
2123 | { | |
2124 | /* ID of the corresponding partition. */ | |
2125 | int id; | |
2126 | /* The partition. */ | |
2127 | struct partition *partition; | |
2128 | }; | |
2129 | ||
2130 | /* Data attached to edges of partition dependence graph. */ | |
2131 | struct pg_edata | |
2132 | { | |
2133 | /* If the dependence edge can be resolved by runtime alias check, | |
2134 | this vector contains data dependence relations for runtime alias | |
2135 | check. On the other hand, if the dependence edge is introduced | |
2136 | because of compilation time known data dependence, this vector | |
2137 | contains nothing. */ | |
2138 | vec<ddr_p> alias_ddrs; | |
2139 | }; | |
2140 | ||
2141 | /* Callback data for traversing edges in graph. */ | |
2142 | struct pg_edge_callback_data | |
2143 | { | |
2144 | /* Bitmap contains strong connected components should be merged. */ | |
2145 | bitmap sccs_to_merge; | |
2146 | /* Array constains component information for all vertices. */ | |
2147 | int *vertices_component; | |
2148 | /* Vector to record all data dependence relations which are needed | |
2149 | to break strong connected components by runtime alias checks. */ | |
2150 | vec<ddr_p> *alias_ddrs; | |
2151 | }; | |
2152 | ||
2153 | /* Initialize vertice's data for partition dependence graph PG with | |
2154 | PARTITIONS. */ | |
2155 | ||
2156 | static void | |
2157 | init_partition_graph_vertices (struct graph *pg, | |
2158 | vec<struct partition *> *partitions) | |
2159 | { | |
2160 | int i; | |
2161 | partition *partition; | |
2162 | struct pg_vdata *data; | |
2163 | ||
2164 | for (i = 0; partitions->iterate (i, &partition); ++i) | |
2165 | { | |
2166 | data = new pg_vdata; | |
2167 | pg->vertices[i].data = data; | |
2168 | data->id = i; | |
2169 | data->partition = partition; | |
2170 | } | |
2171 | } | |
2172 | ||
2173 | /* Add edge <I, J> to partition dependence graph PG. Attach vector of data | |
2174 | dependence relations to the EDGE if DDRS isn't NULL. */ | |
2175 | ||
2176 | static void | |
2177 | add_partition_graph_edge (struct graph *pg, int i, int j, vec<ddr_p> *ddrs) | |
2178 | { | |
2179 | struct graph_edge *e = add_edge (pg, i, j); | |
2180 | ||
2181 | /* If the edge is attached with data dependence relations, it means this | |
2182 | dependence edge can be resolved by runtime alias checks. */ | |
2183 | if (ddrs != NULL) | |
2184 | { | |
2185 | struct pg_edata *data = new pg_edata; | |
2186 | ||
2187 | gcc_assert (ddrs->length () > 0); | |
2188 | e->data = data; | |
2189 | data->alias_ddrs = vNULL; | |
2190 | data->alias_ddrs.safe_splice (*ddrs); | |
2191 | } | |
2192 | } | |
2193 | ||
2194 | /* Callback function for graph travesal algorithm. It returns true | |
2195 | if edge E should skipped when traversing the graph. */ | |
2196 | ||
2197 | static bool | |
2198 | pg_skip_alias_edge (struct graph_edge *e) | |
2199 | { | |
2200 | struct pg_edata *data = (struct pg_edata *)e->data; | |
2201 | return (data != NULL && data->alias_ddrs.length () > 0); | |
2202 | } | |
2203 | ||
2204 | /* Callback function freeing data attached to edge E of graph. */ | |
2205 | ||
2206 | static void | |
2207 | free_partition_graph_edata_cb (struct graph *, struct graph_edge *e, void *) | |
2208 | { | |
2209 | if (e->data != NULL) | |
2210 | { | |
2211 | struct pg_edata *data = (struct pg_edata *)e->data; | |
2212 | data->alias_ddrs.release (); | |
2213 | delete data; | |
2214 | } | |
2215 | } | |
2216 | ||
2217 | /* Free data attached to vertice of partition dependence graph PG. */ | |
2218 | ||
2219 | static void | |
2220 | free_partition_graph_vdata (struct graph *pg) | |
2221 | { | |
2222 | int i; | |
2223 | struct pg_vdata *data; | |
2224 | ||
2225 | for (i = 0; i < pg->n_vertices; ++i) | |
2226 | { | |
2227 | data = (struct pg_vdata *)pg->vertices[i].data; | |
2228 | delete data; | |
2229 | } | |
2230 | } | |
2231 | ||
2232 | /* Build and return partition dependence graph for PARTITIONS. RDG is | |
2233 | reduced dependence graph for the loop to be distributed. If IGNORE_ALIAS_P | |
2234 | is true, data dependence caused by possible alias between references | |
2235 | is ignored, as if it doesn't exist at all; otherwise all depdendences | |
2236 | are considered. */ | |
2237 | ||
eef99cd9 GB |
2238 | struct graph * |
2239 | loop_distribution::build_partition_graph (struct graph *rdg, | |
2240 | vec<struct partition *> *partitions, | |
2241 | bool ignore_alias_p) | |
a8745cc2 BC |
2242 | { |
2243 | int i, j; | |
2244 | struct partition *partition1, *partition2; | |
2245 | graph *pg = new_graph (partitions->length ()); | |
2246 | auto_vec<ddr_p> alias_ddrs, *alias_ddrs_p; | |
2247 | ||
2248 | alias_ddrs_p = ignore_alias_p ? NULL : &alias_ddrs; | |
2249 | ||
2250 | init_partition_graph_vertices (pg, partitions); | |
2251 | ||
2252 | for (i = 0; partitions->iterate (i, &partition1); ++i) | |
2253 | { | |
2254 | for (j = i + 1; partitions->iterate (j, &partition2); ++j) | |
2255 | { | |
2256 | /* dependence direction - 0 is no dependence, -1 is back, | |
2257 | 1 is forth, 2 is both (we can stop then, merging will occur). */ | |
2258 | int dir = 0; | |
2259 | ||
2260 | /* If the first partition has reduction, add back edge; if the | |
2261 | second partition has reduction, add forth edge. This makes | |
2262 | sure that reduction partition will be sorted as the last one. */ | |
2263 | if (partition_reduction_p (partition1)) | |
2264 | dir = -1; | |
2265 | else if (partition_reduction_p (partition2)) | |
2266 | dir = 1; | |
2267 | ||
2268 | /* Cleanup the temporary vector. */ | |
2269 | alias_ddrs.truncate (0); | |
2270 | ||
2271 | dir = pg_add_dependence_edges (rdg, dir, partition1->datarefs, | |
2272 | partition2->datarefs, alias_ddrs_p); | |
2273 | ||
2274 | /* Add edge to partition graph if there exists dependence. There | |
2275 | are two types of edges. One type edge is caused by compilation | |
67914693 | 2276 | time known dependence, this type cannot be resolved by runtime |
a8745cc2 BC |
2277 | alias check. The other type can be resolved by runtime alias |
2278 | check. */ | |
2279 | if (dir == 1 || dir == 2 | |
2280 | || alias_ddrs.length () > 0) | |
2281 | { | |
2282 | /* Attach data dependence relations to edge that can be resolved | |
2283 | by runtime alias check. */ | |
2284 | bool alias_edge_p = (dir != 1 && dir != 2); | |
2285 | add_partition_graph_edge (pg, i, j, | |
2286 | (alias_edge_p) ? &alias_ddrs : NULL); | |
2287 | } | |
2288 | if (dir == -1 || dir == 2 | |
2289 | || alias_ddrs.length () > 0) | |
2290 | { | |
2291 | /* Attach data dependence relations to edge that can be resolved | |
2292 | by runtime alias check. */ | |
2293 | bool alias_edge_p = (dir != -1 && dir != 2); | |
2294 | add_partition_graph_edge (pg, j, i, | |
2295 | (alias_edge_p) ? &alias_ddrs : NULL); | |
2296 | } | |
2297 | } | |
2298 | } | |
2299 | return pg; | |
2300 | } | |
2301 | ||
b4ec1d31 BC |
2302 | /* Sort partitions in PG in descending post order and store them in |
2303 | PARTITIONS. */ | |
a8745cc2 BC |
2304 | |
2305 | static void | |
2306 | sort_partitions_by_post_order (struct graph *pg, | |
2307 | vec<struct partition *> *partitions) | |
2308 | { | |
2309 | int i; | |
2310 | struct pg_vdata *data; | |
2311 | ||
b4ec1d31 | 2312 | /* Now order the remaining nodes in descending postorder. */ |
a8745cc2 BC |
2313 | qsort (pg->vertices, pg->n_vertices, sizeof (vertex), pgcmp); |
2314 | partitions->truncate (0); | |
2315 | for (i = 0; i < pg->n_vertices; ++i) | |
2316 | { | |
2317 | data = (struct pg_vdata *)pg->vertices[i].data; | |
2318 | if (data->partition) | |
2319 | partitions->safe_push (data->partition); | |
2320 | } | |
2321 | } | |
2322 | ||
eef99cd9 GB |
2323 | void |
2324 | loop_distribution::merge_dep_scc_partitions (struct graph *rdg, | |
2325 | vec<struct partition *> *partitions, | |
2326 | bool ignore_alias_p) | |
a8745cc2 BC |
2327 | { |
2328 | struct partition *partition1, *partition2; | |
2329 | struct pg_vdata *data; | |
163aa51b | 2330 | graph *pg = build_partition_graph (rdg, partitions, ignore_alias_p); |
a8745cc2 BC |
2331 | int i, j, num_sccs = graphds_scc (pg, NULL); |
2332 | ||
2333 | /* Strong connected compoenent means dependence cycle, we cannot distribute | |
2334 | them. So fuse them together. */ | |
2335 | if ((unsigned) num_sccs < partitions->length ()) | |
2336 | { | |
2337 | for (i = 0; i < num_sccs; ++i) | |
2338 | { | |
2339 | for (j = 0; partitions->iterate (j, &partition1); ++j) | |
2340 | if (pg->vertices[j].component == i) | |
2341 | break; | |
2342 | for (j = j + 1; partitions->iterate (j, &partition2); ++j) | |
2343 | if (pg->vertices[j].component == i) | |
2344 | { | |
2345 | partition_merge_into (NULL, partition1, | |
2346 | partition2, FUSE_SAME_SCC); | |
2347 | partition1->type = PTYPE_SEQUENTIAL; | |
2348 | (*partitions)[j] = NULL; | |
2349 | partition_free (partition2); | |
2350 | data = (struct pg_vdata *)pg->vertices[j].data; | |
2351 | data->partition = NULL; | |
2352 | } | |
2353 | } | |
a8745cc2 | 2354 | } |
aa1528b5 BC |
2355 | |
2356 | sort_partitions_by_post_order (pg, partitions); | |
a8745cc2 BC |
2357 | gcc_assert (partitions->length () == (unsigned)num_sccs); |
2358 | free_partition_graph_vdata (pg); | |
2359 | free_graph (pg); | |
2360 | } | |
2361 | ||
2362 | /* Callback function for traversing edge E in graph G. DATA is private | |
2363 | callback data. */ | |
2364 | ||
2365 | static void | |
2366 | pg_collect_alias_ddrs (struct graph *g, struct graph_edge *e, void *data) | |
2367 | { | |
2368 | int i, j, component; | |
2369 | struct pg_edge_callback_data *cbdata; | |
2370 | struct pg_edata *edata = (struct pg_edata *) e->data; | |
2371 | ||
2372 | /* If the edge doesn't have attached data dependence, it represents | |
2373 | compilation time known dependences. This type dependence cannot | |
2374 | be resolved by runtime alias check. */ | |
2375 | if (edata == NULL || edata->alias_ddrs.length () == 0) | |
2376 | return; | |
2377 | ||
2378 | cbdata = (struct pg_edge_callback_data *) data; | |
2379 | i = e->src; | |
2380 | j = e->dest; | |
2381 | component = cbdata->vertices_component[i]; | |
2382 | /* Vertices are topologically sorted according to compilation time | |
2383 | known dependences, so we can break strong connected components | |
2384 | by removing edges of the opposite direction, i.e, edges pointing | |
2385 | from vertice with smaller post number to vertice with bigger post | |
2386 | number. */ | |
2387 | if (g->vertices[i].post < g->vertices[j].post | |
2388 | /* We only need to remove edges connecting vertices in the same | |
2389 | strong connected component to break it. */ | |
2390 | && component == cbdata->vertices_component[j] | |
2391 | /* Check if we want to break the strong connected component or not. */ | |
2392 | && !bitmap_bit_p (cbdata->sccs_to_merge, component)) | |
2393 | cbdata->alias_ddrs->safe_splice (edata->alias_ddrs); | |
2394 | } | |
2395 | ||
2396 | /* This is the main function breaking strong conected components in | |
2397 | PARTITIONS giving reduced depdendence graph RDG. Store data dependence | |
2398 | relations for runtime alias check in ALIAS_DDRS. */ | |
eef99cd9 GB |
2399 | void |
2400 | loop_distribution::break_alias_scc_partitions (struct graph *rdg, | |
2401 | vec<struct partition *> *partitions, | |
2402 | vec<ddr_p> *alias_ddrs) | |
a8745cc2 | 2403 | { |
b4ec1d31 | 2404 | int i, j, k, num_sccs, num_sccs_no_alias; |
a8745cc2 BC |
2405 | /* Build partition dependence graph. */ |
2406 | graph *pg = build_partition_graph (rdg, partitions, false); | |
2407 | ||
2408 | alias_ddrs->truncate (0); | |
2409 | /* Find strong connected components in the graph, with all dependence edges | |
2410 | considered. */ | |
2411 | num_sccs = graphds_scc (pg, NULL); | |
2412 | /* All SCCs now can be broken by runtime alias checks because SCCs caused by | |
2413 | compilation time known dependences are merged before this function. */ | |
2414 | if ((unsigned) num_sccs < partitions->length ()) | |
2415 | { | |
2416 | struct pg_edge_callback_data cbdata; | |
2417 | auto_bitmap sccs_to_merge; | |
2418 | auto_vec<enum partition_type> scc_types; | |
2419 | struct partition *partition, *first; | |
2420 | ||
6dc29d3a | 2421 | /* If all partitions in a SCC have the same type, we can simply merge the |
a8745cc2 BC |
2422 | SCC. This loop finds out such SCCS and record them in bitmap. */ |
2423 | bitmap_set_range (sccs_to_merge, 0, (unsigned) num_sccs); | |
2424 | for (i = 0; i < num_sccs; ++i) | |
2425 | { | |
2426 | for (j = 0; partitions->iterate (j, &first); ++j) | |
2427 | if (pg->vertices[j].component == i) | |
2428 | break; | |
1623d9f3 BC |
2429 | |
2430 | bool same_type = true, all_builtins = partition_builtin_p (first); | |
a8745cc2 BC |
2431 | for (++j; partitions->iterate (j, &partition); ++j) |
2432 | { | |
2433 | if (pg->vertices[j].component != i) | |
2434 | continue; | |
2435 | ||
2436 | if (first->type != partition->type) | |
2437 | { | |
1623d9f3 | 2438 | same_type = false; |
a8745cc2 BC |
2439 | break; |
2440 | } | |
1623d9f3 | 2441 | all_builtins &= partition_builtin_p (partition); |
a8745cc2 | 2442 | } |
1623d9f3 BC |
2443 | /* Merge SCC if all partitions in SCC have the same type, though the |
2444 | result partition is sequential, because vectorizer can do better | |
2445 | runtime alias check. One expecption is all partitions in SCC are | |
2446 | builtins. */ | |
2447 | if (!same_type || all_builtins) | |
2448 | bitmap_clear_bit (sccs_to_merge, i); | |
a8745cc2 BC |
2449 | } |
2450 | ||
2451 | /* Initialize callback data for traversing. */ | |
2452 | cbdata.sccs_to_merge = sccs_to_merge; | |
2453 | cbdata.alias_ddrs = alias_ddrs; | |
2454 | cbdata.vertices_component = XNEWVEC (int, pg->n_vertices); | |
2455 | /* Record the component information which will be corrupted by next | |
2456 | graph scc finding call. */ | |
2457 | for (i = 0; i < pg->n_vertices; ++i) | |
2458 | cbdata.vertices_component[i] = pg->vertices[i].component; | |
2459 | ||
2460 | /* Collect data dependences for runtime alias checks to break SCCs. */ | |
2461 | if (bitmap_count_bits (sccs_to_merge) != (unsigned) num_sccs) | |
2462 | { | |
2463 | /* Run SCC finding algorithm again, with alias dependence edges | |
6dc29d3a | 2464 | skipped. This is to topologically sort partitions according to |
a8745cc2 BC |
2465 | compilation time known dependence. Note the topological order |
2466 | is stored in the form of pg's post order number. */ | |
2467 | num_sccs_no_alias = graphds_scc (pg, NULL, pg_skip_alias_edge); | |
2468 | gcc_assert (partitions->length () == (unsigned) num_sccs_no_alias); | |
2469 | /* With topological order, we can construct two subgraphs L and R. | |
2470 | L contains edge <x, y> where x < y in terms of post order, while | |
2471 | R contains edge <x, y> where x > y. Edges for compilation time | |
2472 | known dependence all fall in R, so we break SCCs by removing all | |
2473 | (alias) edges of in subgraph L. */ | |
2474 | for_each_edge (pg, pg_collect_alias_ddrs, &cbdata); | |
2475 | } | |
2476 | ||
2477 | /* For SCC that doesn't need to be broken, merge it. */ | |
2478 | for (i = 0; i < num_sccs; ++i) | |
2479 | { | |
2480 | if (!bitmap_bit_p (sccs_to_merge, i)) | |
2481 | continue; | |
2482 | ||
2483 | for (j = 0; partitions->iterate (j, &first); ++j) | |
2484 | if (cbdata.vertices_component[j] == i) | |
2485 | break; | |
b4ec1d31 | 2486 | for (k = j + 1; partitions->iterate (k, &partition); ++k) |
a8745cc2 BC |
2487 | { |
2488 | struct pg_vdata *data; | |
2489 | ||
b4ec1d31 | 2490 | if (cbdata.vertices_component[k] != i) |
a8745cc2 BC |
2491 | continue; |
2492 | ||
e4e9a591 BC |
2493 | /* Update to the minimal postordeer number of vertices in scc so |
2494 | that merged partition is sorted correctly against others. */ | |
2495 | if (pg->vertices[j].post > pg->vertices[k].post) | |
2496 | pg->vertices[j].post = pg->vertices[k].post; | |
2497 | ||
a8745cc2 | 2498 | partition_merge_into (NULL, first, partition, FUSE_SAME_SCC); |
b4ec1d31 | 2499 | (*partitions)[k] = NULL; |
a8745cc2 | 2500 | partition_free (partition); |
b4ec1d31 BC |
2501 | data = (struct pg_vdata *)pg->vertices[k].data; |
2502 | gcc_assert (data->id == k); | |
a8745cc2 | 2503 | data->partition = NULL; |
6dc29d3a BC |
2504 | /* The result partition of merged SCC must be sequential. */ |
2505 | first->type = PTYPE_SEQUENTIAL; | |
a8745cc2 BC |
2506 | } |
2507 | } | |
2508 | } | |
2509 | ||
2510 | sort_partitions_by_post_order (pg, partitions); | |
2511 | free_partition_graph_vdata (pg); | |
2512 | for_each_edge (pg, free_partition_graph_edata_cb, NULL); | |
2513 | free_graph (pg); | |
2514 | ||
2515 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
2516 | { | |
2517 | fprintf (dump_file, "Possible alias data dependence to break:\n"); | |
2518 | dump_data_dependence_relations (dump_file, *alias_ddrs); | |
2519 | } | |
2520 | } | |
2521 | ||
2522 | /* Compute and return an expression whose value is the segment length which | |
2523 | will be accessed by DR in NITERS iterations. */ | |
2524 | ||
2525 | static tree | |
2526 | data_ref_segment_size (struct data_reference *dr, tree niters) | |
2527 | { | |
a57776a1 RS |
2528 | niters = size_binop (MINUS_EXPR, |
2529 | fold_convert (sizetype, niters), | |
2530 | size_one_node); | |
2531 | return size_binop (MULT_EXPR, | |
2532 | fold_convert (sizetype, DR_STEP (dr)), | |
2533 | fold_convert (sizetype, niters)); | |
a8745cc2 BC |
2534 | } |
2535 | ||
2536 | /* Return true if LOOP's latch is dominated by statement for data reference | |
2537 | DR. */ | |
2538 | ||
2539 | static inline bool | |
99b1c316 | 2540 | latch_dominated_by_data_ref (class loop *loop, data_reference *dr) |
a8745cc2 BC |
2541 | { |
2542 | return dominated_by_p (CDI_DOMINATORS, single_exit (loop)->src, | |
2543 | gimple_bb (DR_STMT (dr))); | |
2544 | } | |
2545 | ||
2546 | /* Compute alias check pairs and store them in COMP_ALIAS_PAIRS for LOOP's | |
2547 | data dependence relations ALIAS_DDRS. */ | |
2548 | ||
2549 | static void | |
99b1c316 | 2550 | compute_alias_check_pairs (class loop *loop, vec<ddr_p> *alias_ddrs, |
a8745cc2 BC |
2551 | vec<dr_with_seg_len_pair_t> *comp_alias_pairs) |
2552 | { | |
2553 | unsigned int i; | |
2554 | unsigned HOST_WIDE_INT factor = 1; | |
2555 | tree niters_plus_one, niters = number_of_latch_executions (loop); | |
2556 | ||
2557 | gcc_assert (niters != NULL_TREE && niters != chrec_dont_know); | |
2558 | niters = fold_convert (sizetype, niters); | |
2559 | niters_plus_one = size_binop (PLUS_EXPR, niters, size_one_node); | |
2560 | ||
2561 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
2562 | fprintf (dump_file, "Creating alias check pairs:\n"); | |
2563 | ||
2564 | /* Iterate all data dependence relations and compute alias check pairs. */ | |
2565 | for (i = 0; i < alias_ddrs->length (); i++) | |
2566 | { | |
2567 | ddr_p ddr = (*alias_ddrs)[i]; | |
2568 | struct data_reference *dr_a = DDR_A (ddr); | |
2569 | struct data_reference *dr_b = DDR_B (ddr); | |
2570 | tree seg_length_a, seg_length_b; | |
a8745cc2 BC |
2571 | |
2572 | if (latch_dominated_by_data_ref (loop, dr_a)) | |
2573 | seg_length_a = data_ref_segment_size (dr_a, niters_plus_one); | |
2574 | else | |
2575 | seg_length_a = data_ref_segment_size (dr_a, niters); | |
2576 | ||
2577 | if (latch_dominated_by_data_ref (loop, dr_b)) | |
2578 | seg_length_b = data_ref_segment_size (dr_b, niters_plus_one); | |
2579 | else | |
2580 | seg_length_b = data_ref_segment_size (dr_b, niters); | |
2581 | ||
a57776a1 RS |
2582 | unsigned HOST_WIDE_INT access_size_a |
2583 | = tree_to_uhwi (TYPE_SIZE_UNIT (TREE_TYPE (DR_REF (dr_a)))); | |
2584 | unsigned HOST_WIDE_INT access_size_b | |
2585 | = tree_to_uhwi (TYPE_SIZE_UNIT (TREE_TYPE (DR_REF (dr_b)))); | |
2586 | unsigned int align_a = TYPE_ALIGN_UNIT (TREE_TYPE (DR_REF (dr_a))); | |
2587 | unsigned int align_b = TYPE_ALIGN_UNIT (TREE_TYPE (DR_REF (dr_b))); | |
2588 | ||
a8745cc2 | 2589 | dr_with_seg_len_pair_t dr_with_seg_len_pair |
a57776a1 | 2590 | (dr_with_seg_len (dr_a, seg_length_a, access_size_a, align_a), |
e9acf80c RS |
2591 | dr_with_seg_len (dr_b, seg_length_b, access_size_b, align_b), |
2592 | /* ??? Would WELL_ORDERED be safe? */ | |
2593 | dr_with_seg_len_pair_t::REORDERED); | |
a8745cc2 | 2594 | |
a8745cc2 BC |
2595 | comp_alias_pairs->safe_push (dr_with_seg_len_pair); |
2596 | } | |
2597 | ||
2598 | if (tree_fits_uhwi_p (niters)) | |
2599 | factor = tree_to_uhwi (niters); | |
2600 | ||
2601 | /* Prune alias check pairs. */ | |
2602 | prune_runtime_alias_test_list (comp_alias_pairs, factor); | |
2603 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
2604 | fprintf (dump_file, | |
2605 | "Improved number of alias checks from %d to %d\n", | |
2606 | alias_ddrs->length (), comp_alias_pairs->length ()); | |
2607 | } | |
2608 | ||
2609 | /* Given data dependence relations in ALIAS_DDRS, generate runtime alias | |
2610 | checks and version LOOP under condition of these runtime alias checks. */ | |
2611 | ||
2612 | static void | |
1623d9f3 | 2613 | version_loop_by_alias_check (vec<struct partition *> *partitions, |
99b1c316 | 2614 | class loop *loop, vec<ddr_p> *alias_ddrs) |
a8745cc2 BC |
2615 | { |
2616 | profile_probability prob; | |
2617 | basic_block cond_bb; | |
99b1c316 | 2618 | class loop *nloop; |
a8745cc2 BC |
2619 | tree lhs, arg0, cond_expr = NULL_TREE; |
2620 | gimple_seq cond_stmts = NULL; | |
2621 | gimple *call_stmt = NULL; | |
2622 | auto_vec<dr_with_seg_len_pair_t> comp_alias_pairs; | |
2623 | ||
2624 | /* Generate code for runtime alias checks if necessary. */ | |
2625 | gcc_assert (alias_ddrs->length () > 0); | |
2626 | ||
2627 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
2628 | fprintf (dump_file, | |
2629 | "Version loop <%d> with runtime alias check\n", loop->num); | |
2630 | ||
2631 | compute_alias_check_pairs (loop, alias_ddrs, &comp_alias_pairs); | |
2632 | create_runtime_alias_checks (loop, &comp_alias_pairs, &cond_expr); | |
2633 | cond_expr = force_gimple_operand_1 (cond_expr, &cond_stmts, | |
8d2d0de9 | 2634 | is_gimple_val, NULL_TREE); |
a8745cc2 BC |
2635 | |
2636 | /* Depend on vectorizer to fold IFN_LOOP_DIST_ALIAS. */ | |
1623d9f3 BC |
2637 | bool cancelable_p = flag_tree_loop_vectorize; |
2638 | if (cancelable_p) | |
2639 | { | |
2640 | unsigned i = 0; | |
2641 | struct partition *partition; | |
2642 | for (; partitions->iterate (i, &partition); ++i) | |
2643 | if (!partition_builtin_p (partition)) | |
2644 | break; | |
2645 | ||
2646 | /* If all partitions are builtins, distributing it would be profitable and | |
2647 | we don't want to cancel the runtime alias checks. */ | |
2648 | if (i == partitions->length ()) | |
2649 | cancelable_p = false; | |
2650 | } | |
2651 | ||
2652 | /* Generate internal function call for loop distribution alias check if the | |
2653 | runtime alias check should be cancelable. */ | |
2654 | if (cancelable_p) | |
a8745cc2 | 2655 | { |
a8745cc2 BC |
2656 | call_stmt = gimple_build_call_internal (IFN_LOOP_DIST_ALIAS, |
2657 | 2, NULL_TREE, cond_expr); | |
2658 | lhs = make_ssa_name (boolean_type_node); | |
2659 | gimple_call_set_lhs (call_stmt, lhs); | |
2660 | } | |
2661 | else | |
2662 | lhs = cond_expr; | |
2663 | ||
2664 | prob = profile_probability::guessed_always ().apply_scale (9, 10); | |
2665 | initialize_original_copy_tables (); | |
2666 | nloop = loop_version (loop, lhs, &cond_bb, prob, prob.invert (), | |
2667 | prob, prob.invert (), true); | |
2668 | free_original_copy_tables (); | |
2669 | /* Record the original loop number in newly generated loops. In case of | |
2670 | distribution, the original loop will be distributed and the new loop | |
2671 | is kept. */ | |
2672 | loop->orig_loop_num = nloop->num; | |
2673 | nloop->orig_loop_num = nloop->num; | |
2674 | nloop->dont_vectorize = true; | |
2675 | nloop->force_vectorize = false; | |
2676 | ||
2677 | if (call_stmt) | |
2678 | { | |
2679 | /* Record new loop's num in IFN_LOOP_DIST_ALIAS because the original | |
2680 | loop could be destroyed. */ | |
2681 | arg0 = build_int_cst (integer_type_node, loop->orig_loop_num); | |
2682 | gimple_call_set_arg (call_stmt, 0, arg0); | |
2683 | gimple_seq_add_stmt_without_update (&cond_stmts, call_stmt); | |
2684 | } | |
2685 | ||
2686 | if (cond_stmts) | |
2687 | { | |
2688 | gimple_stmt_iterator cond_gsi = gsi_last_bb (cond_bb); | |
2689 | gsi_insert_seq_before (&cond_gsi, cond_stmts, GSI_SAME_STMT); | |
2690 | } | |
2691 | update_ssa (TODO_update_ssa); | |
2692 | } | |
2693 | ||
2694 | /* Return true if loop versioning is needed to distrubute PARTITIONS. | |
2695 | ALIAS_DDRS are data dependence relations for runtime alias check. */ | |
2696 | ||
2697 | static inline bool | |
2698 | version_for_distribution_p (vec<struct partition *> *partitions, | |
2699 | vec<ddr_p> *alias_ddrs) | |
2700 | { | |
2701 | /* No need to version loop if we have only one partition. */ | |
2702 | if (partitions->length () == 1) | |
2703 | return false; | |
2704 | ||
2705 | /* Need to version loop if runtime alias check is necessary. */ | |
2706 | return (alias_ddrs->length () > 0); | |
2707 | } | |
2708 | ||
957f0d8f BC |
2709 | /* Compare base offset of builtin mem* partitions P1 and P2. */ |
2710 | ||
d2391983 AM |
2711 | static int |
2712 | offset_cmp (const void *vp1, const void *vp2) | |
957f0d8f | 2713 | { |
d2391983 AM |
2714 | struct partition *p1 = *(struct partition *const *) vp1; |
2715 | struct partition *p2 = *(struct partition *const *) vp2; | |
2716 | unsigned HOST_WIDE_INT o1 = p1->builtin->dst_base_offset; | |
2717 | unsigned HOST_WIDE_INT o2 = p2->builtin->dst_base_offset; | |
2718 | return (o2 < o1) - (o1 < o2); | |
957f0d8f BC |
2719 | } |
2720 | ||
2721 | /* Fuse adjacent memset builtin PARTITIONS if possible. This is a special | |
2722 | case optimization transforming below code: | |
2723 | ||
2724 | __builtin_memset (&obj, 0, 100); | |
2725 | _1 = &obj + 100; | |
2726 | __builtin_memset (_1, 0, 200); | |
2727 | _2 = &obj + 300; | |
2728 | __builtin_memset (_2, 0, 100); | |
2729 | ||
2730 | into: | |
2731 | ||
2732 | __builtin_memset (&obj, 0, 400); | |
2733 | ||
2734 | Note we don't have dependence information between different partitions | |
2735 | at this point, as a result, we can't handle nonadjacent memset builtin | |
2736 | partitions since dependence might be broken. */ | |
2737 | ||
2738 | static void | |
2739 | fuse_memset_builtins (vec<struct partition *> *partitions) | |
2740 | { | |
2741 | unsigned i, j; | |
2742 | struct partition *part1, *part2; | |
49e4ca31 | 2743 | tree rhs1, rhs2; |
957f0d8f BC |
2744 | |
2745 | for (i = 0; partitions->iterate (i, &part1);) | |
2746 | { | |
2747 | if (part1->kind != PKIND_MEMSET) | |
2748 | { | |
2749 | i++; | |
2750 | continue; | |
2751 | } | |
2752 | ||
2753 | /* Find sub-array of memset builtins of the same base. Index range | |
2754 | of the sub-array is [i, j) with "j > i". */ | |
2755 | for (j = i + 1; partitions->iterate (j, &part2); ++j) | |
2756 | { | |
2757 | if (part2->kind != PKIND_MEMSET | |
2758 | || !operand_equal_p (part1->builtin->dst_base_base, | |
2759 | part2->builtin->dst_base_base, 0)) | |
2760 | break; | |
49e4ca31 BC |
2761 | |
2762 | /* Memset calls setting different values can't be merged. */ | |
2763 | rhs1 = gimple_assign_rhs1 (DR_STMT (part1->builtin->dst_dr)); | |
2764 | rhs2 = gimple_assign_rhs1 (DR_STMT (part2->builtin->dst_dr)); | |
2765 | if (!operand_equal_p (rhs1, rhs2, 0)) | |
2766 | break; | |
957f0d8f BC |
2767 | } |
2768 | ||
2769 | /* Stable sort is required in order to avoid breaking dependence. */ | |
d2391983 AM |
2770 | gcc_stablesort (&(*partitions)[i], j - i, sizeof (*partitions)[i], |
2771 | offset_cmp); | |
957f0d8f BC |
2772 | /* Continue with next partition. */ |
2773 | i = j; | |
2774 | } | |
2775 | ||
2776 | /* Merge all consecutive memset builtin partitions. */ | |
2777 | for (i = 0; i < partitions->length () - 1;) | |
2778 | { | |
2779 | part1 = (*partitions)[i]; | |
2780 | if (part1->kind != PKIND_MEMSET) | |
2781 | { | |
2782 | i++; | |
2783 | continue; | |
2784 | } | |
2785 | ||
2786 | part2 = (*partitions)[i + 1]; | |
2787 | /* Only merge memset partitions of the same base and with constant | |
2788 | access sizes. */ | |
2789 | if (part2->kind != PKIND_MEMSET | |
2790 | || TREE_CODE (part1->builtin->size) != INTEGER_CST | |
2791 | || TREE_CODE (part2->builtin->size) != INTEGER_CST | |
2792 | || !operand_equal_p (part1->builtin->dst_base_base, | |
2793 | part2->builtin->dst_base_base, 0)) | |
2794 | { | |
2795 | i++; | |
2796 | continue; | |
2797 | } | |
49e4ca31 BC |
2798 | rhs1 = gimple_assign_rhs1 (DR_STMT (part1->builtin->dst_dr)); |
2799 | rhs2 = gimple_assign_rhs1 (DR_STMT (part2->builtin->dst_dr)); | |
957f0d8f BC |
2800 | int bytev1 = const_with_all_bytes_same (rhs1); |
2801 | int bytev2 = const_with_all_bytes_same (rhs2); | |
2802 | /* Only merge memset partitions of the same value. */ | |
2803 | if (bytev1 != bytev2 || bytev1 == -1) | |
2804 | { | |
2805 | i++; | |
2806 | continue; | |
2807 | } | |
2808 | wide_int end1 = wi::add (part1->builtin->dst_base_offset, | |
2809 | wi::to_wide (part1->builtin->size)); | |
2810 | /* Only merge adjacent memset partitions. */ | |
2811 | if (wi::ne_p (end1, part2->builtin->dst_base_offset)) | |
2812 | { | |
2813 | i++; | |
2814 | continue; | |
2815 | } | |
2816 | /* Merge partitions[i] and partitions[i+1]. */ | |
2817 | part1->builtin->size = fold_build2 (PLUS_EXPR, sizetype, | |
2818 | part1->builtin->size, | |
2819 | part2->builtin->size); | |
2820 | partition_free (part2); | |
2821 | partitions->ordered_remove (i + 1); | |
2822 | } | |
2823 | } | |
2824 | ||
eef99cd9 GB |
2825 | void |
2826 | loop_distribution::finalize_partitions (class loop *loop, | |
2827 | vec<struct partition *> *partitions, | |
2828 | vec<ddr_p> *alias_ddrs) | |
a8745cc2 BC |
2829 | { |
2830 | unsigned i; | |
163aa51b | 2831 | struct partition *partition, *a; |
a8745cc2 BC |
2832 | |
2833 | if (partitions->length () == 1 | |
2834 | || alias_ddrs->length () > 0) | |
2835 | return; | |
2836 | ||
5955438a | 2837 | unsigned num_builtin = 0, num_normal = 0, num_partial_memset = 0; |
163aa51b BC |
2838 | bool same_type_p = true; |
2839 | enum partition_type type = ((*partitions)[0])->type; | |
2840 | for (i = 0; partitions->iterate (i, &partition); ++i) | |
a8745cc2 | 2841 | { |
163aa51b | 2842 | same_type_p &= (type == partition->type); |
5955438a BC |
2843 | if (partition_builtin_p (partition)) |
2844 | { | |
2845 | num_builtin++; | |
2846 | continue; | |
2847 | } | |
2848 | num_normal++; | |
2849 | if (partition->kind == PKIND_PARTIAL_MEMSET) | |
2850 | num_partial_memset++; | |
a8745cc2 BC |
2851 | } |
2852 | ||
163aa51b BC |
2853 | /* Don't distribute current loop into too many loops given we don't have |
2854 | memory stream cost model. Be even more conservative in case of loop | |
2855 | nest distribution. */ | |
5955438a BC |
2856 | if ((same_type_p && num_builtin == 0 |
2857 | && (loop->inner == NULL || num_normal != 2 || num_partial_memset != 1)) | |
163aa51b BC |
2858 | || (loop->inner != NULL |
2859 | && i >= NUM_PARTITION_THRESHOLD && num_normal > 1) | |
2860 | || (loop->inner == NULL | |
2861 | && i >= NUM_PARTITION_THRESHOLD && num_normal > num_builtin)) | |
a8745cc2 | 2862 | { |
163aa51b BC |
2863 | a = (*partitions)[0]; |
2864 | for (i = 1; partitions->iterate (i, &partition); ++i) | |
2865 | { | |
2866 | partition_merge_into (NULL, a, partition, FUSE_FINALIZE); | |
2867 | partition_free (partition); | |
2868 | } | |
2869 | partitions->truncate (1); | |
a8745cc2 | 2870 | } |
957f0d8f BC |
2871 | |
2872 | /* Fuse memset builtins if possible. */ | |
2873 | if (partitions->length () > 1) | |
2874 | fuse_memset_builtins (partitions); | |
a8745cc2 BC |
2875 | } |
2876 | ||
2877 | /* Distributes the code from LOOP in such a way that producer statements | |
2878 | are placed before consumer statements. Tries to separate only the | |
2879 | statements from STMTS into separate loops. Returns the number of | |
2880 | distributed loops. Set NB_CALLS to number of generated builtin calls. | |
2881 | Set *DESTROY_P to whether LOOP needs to be destroyed. */ | |
dea61d92 | 2882 | |
eef99cd9 GB |
2883 | int |
2884 | loop_distribution::distribute_loop (class loop *loop, vec<gimple *> stmts, | |
5879ab5f RB |
2885 | control_dependences *cd, int *nb_calls, bool *destroy_p, |
2886 | bool only_patterns_p) | |
dea61d92 | 2887 | { |
1e485f89 | 2888 | ddrs_table = new hash_table<ddr_hasher> (389); |
2fd5894f | 2889 | struct graph *rdg; |
526ceb68 | 2890 | partition *partition; |
2fd5894f | 2891 | int i, nbp; |
dea61d92 | 2892 | |
c9326aef | 2893 | *destroy_p = false; |
826a536d | 2894 | *nb_calls = 0; |
4084ea5f | 2895 | loop_nest.create (0); |
2fd5894f | 2896 | if (!find_loop_nest (loop, &loop_nest)) |
4084ea5f BC |
2897 | { |
2898 | loop_nest.release (); | |
1e485f89 | 2899 | delete ddrs_table; |
4084ea5f BC |
2900 | return 0; |
2901 | } | |
2fd5894f | 2902 | |
9fafb14a | 2903 | datarefs_vec.create (20); |
c4450491 | 2904 | has_nonaddressable_dataref_p = false; |
4084ea5f | 2905 | rdg = build_rdg (loop, cd); |
2fd5894f RB |
2906 | if (!rdg) |
2907 | { | |
2908 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
2909 | fprintf (dump_file, | |
2910 | "Loop %d not distributed: failed to build the RDG.\n", | |
2911 | loop->num); | |
2912 | ||
4084ea5f | 2913 | loop_nest.release (); |
9fafb14a | 2914 | free_data_refs (datarefs_vec); |
1e485f89 | 2915 | delete ddrs_table; |
9fafb14a BC |
2916 | return 0; |
2917 | } | |
2918 | ||
2919 | if (datarefs_vec.length () > MAX_DATAREFS_NUM) | |
2920 | { | |
2921 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
2922 | fprintf (dump_file, | |
2923 | "Loop %d not distributed: too many memory references.\n", | |
2924 | loop->num); | |
2925 | ||
2926 | free_rdg (rdg); | |
2927 | loop_nest.release (); | |
2928 | free_data_refs (datarefs_vec); | |
1e485f89 | 2929 | delete ddrs_table; |
2fd5894f RB |
2930 | return 0; |
2931 | } | |
2932 | ||
9fafb14a BC |
2933 | data_reference_p dref; |
2934 | for (i = 0; datarefs_vec.iterate (i, &dref); ++i) | |
2935 | dref->aux = (void *) (uintptr_t) i; | |
2936 | ||
2fd5894f RB |
2937 | if (dump_file && (dump_flags & TDF_DETAILS)) |
2938 | dump_rdg (dump_file, rdg); | |
2939 | ||
526ceb68 | 2940 | auto_vec<struct partition *, 3> partitions; |
2fd5894f | 2941 | rdg_build_partitions (rdg, stmts, &partitions); |
dea61d92 | 2942 | |
a8745cc2 BC |
2943 | auto_vec<ddr_p> alias_ddrs; |
2944 | ||
4a52eb19 BC |
2945 | auto_bitmap stmt_in_all_partitions; |
2946 | bitmap_copy (stmt_in_all_partitions, partitions[0]->stmts); | |
2947 | for (i = 1; partitions.iterate (i, &partition); ++i) | |
2948 | bitmap_and_into (stmt_in_all_partitions, partitions[i]->stmts); | |
2949 | ||
e7484357 RB |
2950 | bool any_builtin = false; |
2951 | bool reduction_in_all = false; | |
9771b263 | 2952 | FOR_EACH_VEC_ELT (partitions, i, partition) |
be6b029b | 2953 | { |
e7484357 RB |
2954 | reduction_in_all |
2955 | |= classify_partition (loop, rdg, partition, stmt_in_all_partitions); | |
be6b029b RG |
2956 | any_builtin |= partition_builtin_p (partition); |
2957 | } | |
30d55936 | 2958 | |
447f3223 RB |
2959 | /* If we are only distributing patterns but did not detect any, |
2960 | simply bail out. */ | |
5879ab5f | 2961 | if (only_patterns_p |
9fed7f3a RB |
2962 | && !any_builtin) |
2963 | { | |
2964 | nbp = 0; | |
2965 | goto ldist_done; | |
2966 | } | |
2967 | ||
447f3223 RB |
2968 | /* If we are only distributing patterns fuse all partitions that |
2969 | were not classified as builtins. This also avoids chopping | |
2970 | a loop into pieces, separated by builtin calls. That is, we | |
2971 | only want no or a single loop body remaining. */ | |
526ceb68 | 2972 | struct partition *into; |
5879ab5f | 2973 | if (only_patterns_p) |
447f3223 RB |
2974 | { |
2975 | for (i = 0; partitions.iterate (i, &into); ++i) | |
2976 | if (!partition_builtin_p (into)) | |
2977 | break; | |
2978 | for (++i; partitions.iterate (i, &partition); ++i) | |
2979 | if (!partition_builtin_p (partition)) | |
2980 | { | |
f1eb4621 | 2981 | partition_merge_into (NULL, into, partition, FUSE_NON_BUILTIN); |
447f3223 RB |
2982 | partitions.unordered_remove (i); |
2983 | partition_free (partition); | |
2984 | i--; | |
2985 | } | |
2986 | } | |
2987 | ||
2988 | /* Due to limitations in the transform phase we have to fuse all | |
2989 | reduction partitions into the last partition so the existing | |
2990 | loop will contain all loop-closed PHI nodes. */ | |
2991 | for (i = 0; partitions.iterate (i, &into); ++i) | |
2992 | if (partition_reduction_p (into)) | |
2993 | break; | |
2994 | for (i = i + 1; partitions.iterate (i, &partition); ++i) | |
2995 | if (partition_reduction_p (partition)) | |
2996 | { | |
f1eb4621 | 2997 | partition_merge_into (rdg, into, partition, FUSE_REDUCTION); |
447f3223 RB |
2998 | partitions.unordered_remove (i); |
2999 | partition_free (partition); | |
3000 | i--; | |
3001 | } | |
3002 | ||
9fed7f3a RB |
3003 | /* Apply our simple cost model - fuse partitions with similar |
3004 | memory accesses. */ | |
9fed7f3a RB |
3005 | for (i = 0; partitions.iterate (i, &into); ++i) |
3006 | { | |
16eba420 | 3007 | bool changed = false; |
5955438a | 3008 | if (partition_builtin_p (into) || into->kind == PKIND_PARTIAL_MEMSET) |
9fed7f3a RB |
3009 | continue; |
3010 | for (int j = i + 1; | |
3011 | partitions.iterate (j, &partition); ++j) | |
3012 | { | |
95f7d11b | 3013 | if (share_memory_accesses (rdg, into, partition)) |
9fed7f3a | 3014 | { |
f1eb4621 | 3015 | partition_merge_into (rdg, into, partition, FUSE_SHARE_REF); |
447f3223 | 3016 | partitions.unordered_remove (j); |
9fed7f3a RB |
3017 | partition_free (partition); |
3018 | j--; | |
16eba420 | 3019 | changed = true; |
9fed7f3a RB |
3020 | } |
3021 | } | |
16eba420 RB |
3022 | /* If we fused 0 1 2 in step 1 to 0,2 1 as 0 and 2 have similar |
3023 | accesses when 1 and 2 have similar accesses but not 0 and 1 | |
3024 | then in the next iteration we will fail to consider merging | |
3025 | 1 into 0,2. So try again if we did any merging into 0. */ | |
3026 | if (changed) | |
3027 | i--; | |
9fed7f3a RB |
3028 | } |
3029 | ||
e7484357 RB |
3030 | /* Put a non-builtin partition last if we need to preserve a reduction. |
3031 | ??? This is a workaround that makes sort_partitions_by_post_order do | |
3032 | the correct thing while in reality it should sort each component | |
3033 | separately and then put the component with a reduction or a non-builtin | |
3034 | last. */ | |
3035 | if (reduction_in_all | |
3036 | && partition_builtin_p (partitions.last())) | |
3037 | FOR_EACH_VEC_ELT (partitions, i, partition) | |
3038 | if (!partition_builtin_p (partition)) | |
3039 | { | |
3040 | partitions.unordered_remove (i); | |
3041 | partitions.quick_push (partition); | |
3042 | break; | |
3043 | } | |
3044 | ||
163aa51b BC |
3045 | /* Build the partition dependency graph and fuse partitions in strong |
3046 | connected component. */ | |
447f3223 | 3047 | if (partitions.length () > 1) |
c014f6f5 | 3048 | { |
163aa51b | 3049 | /* Don't support loop nest distribution under runtime alias check |
c4450491 BC |
3050 | since it's not likely to enable many vectorization opportunities. |
3051 | Also if loop has any data reference which may be not addressable | |
3052 | since alias check needs to take, compare address of the object. */ | |
3053 | if (loop->inner || has_nonaddressable_dataref_p) | |
163aa51b BC |
3054 | merge_dep_scc_partitions (rdg, &partitions, false); |
3055 | else | |
3056 | { | |
3057 | merge_dep_scc_partitions (rdg, &partitions, true); | |
3058 | if (partitions.length () > 1) | |
3059 | break_alias_scc_partitions (rdg, &partitions, &alias_ddrs); | |
3060 | } | |
b9fc0497 RB |
3061 | } |
3062 | ||
163aa51b | 3063 | finalize_partitions (loop, &partitions, &alias_ddrs); |
a8745cc2 | 3064 | |
e7484357 RB |
3065 | /* If there is a reduction in all partitions make sure the last one |
3066 | is not classified for builtin code generation. */ | |
3067 | if (reduction_in_all) | |
3068 | { | |
3069 | partition = partitions.last (); | |
3070 | if (only_patterns_p | |
3071 | && partition_builtin_p (partition) | |
3072 | && !partition_builtin_p (partitions[0])) | |
3073 | { | |
3074 | nbp = 0; | |
3075 | goto ldist_done; | |
3076 | } | |
3077 | partition->kind = PKIND_NORMAL; | |
3078 | } | |
3079 | ||
9771b263 | 3080 | nbp = partitions.length (); |
a4293fa6 | 3081 | if (nbp == 0 |
9771b263 DN |
3082 | || (nbp == 1 && !partition_builtin_p (partitions[0])) |
3083 | || (nbp > 1 && partition_contains_all_rw (rdg, partitions))) | |
c014f6f5 RG |
3084 | { |
3085 | nbp = 0; | |
3086 | goto ldist_done; | |
3087 | } | |
dea61d92 | 3088 | |
a8745cc2 | 3089 | if (version_for_distribution_p (&partitions, &alias_ddrs)) |
1623d9f3 | 3090 | version_loop_by_alias_check (&partitions, loop, &alias_ddrs); |
a8745cc2 | 3091 | |
dea61d92 | 3092 | if (dump_file && (dump_flags & TDF_DETAILS)) |
a8745cc2 BC |
3093 | { |
3094 | fprintf (dump_file, | |
3095 | "distribute loop <%d> into partitions:\n", loop->num); | |
3096 | dump_rdg_partitions (dump_file, partitions); | |
3097 | } | |
dea61d92 | 3098 | |
9771b263 | 3099 | FOR_EACH_VEC_ELT (partitions, i, partition) |
826a536d RB |
3100 | { |
3101 | if (partition_builtin_p (partition)) | |
3102 | (*nb_calls)++; | |
b71b7a8e | 3103 | *destroy_p |= generate_code_for_partition (loop, partition, i < nbp - 1); |
826a536d | 3104 | } |
dea61d92 | 3105 | |
dea61d92 | 3106 | ldist_done: |
4084ea5f | 3107 | loop_nest.release (); |
9fafb14a | 3108 | free_data_refs (datarefs_vec); |
1e485f89 ML |
3109 | for (hash_table<ddr_hasher>::iterator iter = ddrs_table->begin (); |
3110 | iter != ddrs_table->end (); ++iter) | |
17c5cbdf BC |
3111 | { |
3112 | free_dependence_relation (*iter); | |
3113 | *iter = NULL; | |
3114 | } | |
1e485f89 | 3115 | delete ddrs_table; |
dea61d92 | 3116 | |
9771b263 | 3117 | FOR_EACH_VEC_ELT (partitions, i, partition) |
c61f8985 | 3118 | partition_free (partition); |
dea61d92 | 3119 | |
dea61d92 | 3120 | free_rdg (rdg); |
826a536d | 3121 | return nbp - *nb_calls; |
dea61d92 SP |
3122 | } |
3123 | ||
be55bfe6 | 3124 | |
eef99cd9 | 3125 | void loop_distribution::bb_top_order_init (void) |
be55bfe6 | 3126 | { |
eef99cd9 GB |
3127 | int rpo_num; |
3128 | int *rpo = XNEWVEC (int, last_basic_block_for_fn (cfun)); | |
be55bfe6 | 3129 | |
eef99cd9 GB |
3130 | bb_top_order_index = XNEWVEC (int, last_basic_block_for_fn (cfun)); |
3131 | bb_top_order_index_size = last_basic_block_for_fn (cfun); | |
3132 | rpo_num = pre_and_rev_post_order_compute_fn (cfun, NULL, rpo, true); | |
3133 | for (int i = 0; i < rpo_num; i++) | |
3134 | bb_top_order_index[rpo[i]] = i; | |
be55bfe6 | 3135 | |
eef99cd9 GB |
3136 | free (rpo); |
3137 | } | |
be55bfe6 | 3138 | |
eef99cd9 GB |
3139 | void loop_distribution::bb_top_order_destroy () |
3140 | { | |
3141 | free (bb_top_order_index); | |
3142 | bb_top_order_index = NULL; | |
3143 | bb_top_order_index_size = 0; | |
3144 | } | |
be55bfe6 | 3145 | |
163aa51b BC |
3146 | |
3147 | /* Given LOOP, this function records seed statements for distribution in | |
3148 | WORK_LIST. Return false if there is nothing for distribution. */ | |
3149 | ||
3150 | static bool | |
99b1c316 | 3151 | find_seed_stmts_for_distribution (class loop *loop, vec<gimple *> *work_list) |
163aa51b BC |
3152 | { |
3153 | basic_block *bbs = get_loop_body_in_dom_order (loop); | |
3154 | ||
3155 | /* Initialize the worklist with stmts we seed the partitions with. */ | |
3156 | for (unsigned i = 0; i < loop->num_nodes; ++i) | |
3157 | { | |
3158 | for (gphi_iterator gsi = gsi_start_phis (bbs[i]); | |
3159 | !gsi_end_p (gsi); gsi_next (&gsi)) | |
3160 | { | |
3161 | gphi *phi = gsi.phi (); | |
3162 | if (virtual_operand_p (gimple_phi_result (phi))) | |
3163 | continue; | |
3164 | /* Distribute stmts which have defs that are used outside of | |
3165 | the loop. */ | |
3166 | if (!stmt_has_scalar_dependences_outside_loop (loop, phi)) | |
3167 | continue; | |
3168 | work_list->safe_push (phi); | |
3169 | } | |
3170 | for (gimple_stmt_iterator gsi = gsi_start_bb (bbs[i]); | |
3171 | !gsi_end_p (gsi); gsi_next (&gsi)) | |
3172 | { | |
3173 | gimple *stmt = gsi_stmt (gsi); | |
3174 | ||
051ef623 MG |
3175 | /* Ignore clobbers, they do not have true side effects. */ |
3176 | if (gimple_clobber_p (stmt)) | |
3177 | continue; | |
3178 | ||
163aa51b BC |
3179 | /* If there is a stmt with side-effects bail out - we |
3180 | cannot and should not distribute this loop. */ | |
3181 | if (gimple_has_side_effects (stmt)) | |
3182 | { | |
3183 | free (bbs); | |
3184 | return false; | |
3185 | } | |
3186 | ||
3187 | /* Distribute stmts which have defs that are used outside of | |
3188 | the loop. */ | |
3189 | if (stmt_has_scalar_dependences_outside_loop (loop, stmt)) | |
3190 | ; | |
3191 | /* Otherwise only distribute stores for now. */ | |
3192 | else if (!gimple_vdef (stmt)) | |
3193 | continue; | |
3194 | ||
3195 | work_list->safe_push (stmt); | |
3196 | } | |
3197 | } | |
3198 | free (bbs); | |
3199 | return work_list->length () > 0; | |
3200 | } | |
3201 | ||
3202 | /* Given innermost LOOP, return the outermost enclosing loop that forms a | |
3203 | perfect loop nest. */ | |
3204 | ||
99b1c316 MS |
3205 | static class loop * |
3206 | prepare_perfect_loop_nest (class loop *loop) | |
163aa51b | 3207 | { |
99b1c316 | 3208 | class loop *outer = loop_outer (loop); |
163aa51b BC |
3209 | tree niters = number_of_latch_executions (loop); |
3210 | ||
5955438a | 3211 | /* TODO: We only support the innermost 3-level loop nest distribution |
163aa51b | 3212 | because of compilation time issue for now. This should be relaxed |
5955438a BC |
3213 | in the future. Note we only allow 3-level loop nest distribution |
3214 | when parallelizing loops. */ | |
3215 | while ((loop->inner == NULL | |
3216 | || (loop->inner->inner == NULL && flag_tree_parallelize_loops > 1)) | |
163aa51b BC |
3217 | && loop_outer (outer) |
3218 | && outer->inner == loop && loop->next == NULL | |
3219 | && single_exit (outer) | |
163aa51b BC |
3220 | && !chrec_contains_symbols_defined_in_loop (niters, outer->num) |
3221 | && (niters = number_of_latch_executions (outer)) != NULL_TREE | |
3222 | && niters != chrec_dont_know) | |
3223 | { | |
3224 | loop = outer; | |
3225 | outer = loop_outer (loop); | |
3226 | } | |
3227 | ||
3228 | return loop; | |
3229 | } | |
3230 | ||
eef99cd9 | 3231 | |
be55bfe6 | 3232 | unsigned int |
eef99cd9 | 3233 | loop_distribution::execute (function *fun) |
dea61d92 | 3234 | { |
99b1c316 | 3235 | class loop *loop; |
c014f6f5 | 3236 | bool changed = false; |
1fa0c180 | 3237 | basic_block bb; |
36875e8f | 3238 | control_dependences *cd = NULL; |
b71b7a8e | 3239 | auto_vec<loop_p> loops_to_be_destroyed; |
1fa0c180 | 3240 | |
773d9217 BC |
3241 | if (number_of_loops (fun) <= 1) |
3242 | return 0; | |
3243 | ||
eef99cd9 | 3244 | bb_top_order_init (); |
3be57c56 | 3245 | |
be55bfe6 | 3246 | FOR_ALL_BB_FN (bb, fun) |
1fa0c180 RG |
3247 | { |
3248 | gimple_stmt_iterator gsi; | |
3249 | for (gsi = gsi_start_phis (bb); !gsi_end_p (gsi); gsi_next (&gsi)) | |
3250 | gimple_set_uid (gsi_stmt (gsi), -1); | |
3251 | for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi)) | |
3252 | gimple_set_uid (gsi_stmt (gsi), -1); | |
3253 | } | |
dea61d92 | 3254 | |
c014f6f5 RG |
3255 | /* We can at the moment only distribute non-nested loops, thus restrict |
3256 | walking to innermost loops. */ | |
f0bd40b1 | 3257 | FOR_EACH_LOOP (loop, LI_ONLY_INNERMOST) |
dea61d92 | 3258 | { |
5879ab5f RB |
3259 | /* Don't distribute multiple exit edges loop, or cold loop when |
3260 | not doing pattern detection. */ | |
163aa51b | 3261 | if (!single_exit (loop) |
5879ab5f RB |
3262 | || (!flag_tree_loop_distribute_patterns |
3263 | && !optimize_loop_for_speed_p (loop))) | |
f56f2d33 JH |
3264 | continue; |
3265 | ||
6ff37519 BC |
3266 | /* Don't distribute loop if niters is unknown. */ |
3267 | tree niters = number_of_latch_executions (loop); | |
3268 | if (niters == NULL_TREE || niters == chrec_dont_know) | |
3269 | continue; | |
3270 | ||
163aa51b BC |
3271 | /* Get the perfect loop nest for distribution. */ |
3272 | loop = prepare_perfect_loop_nest (loop); | |
3273 | for (; loop; loop = loop->inner) | |
be6b029b | 3274 | { |
163aa51b BC |
3275 | auto_vec<gimple *> work_list; |
3276 | if (!find_seed_stmts_for_distribution (loop, &work_list)) | |
3277 | break; | |
c014f6f5 | 3278 | |
163aa51b | 3279 | const char *str = loop->inner ? " nest" : ""; |
4f5b9c80 | 3280 | dump_user_location_t loc = find_loop_location (loop); |
36875e8f RB |
3281 | if (!cd) |
3282 | { | |
ca406576 | 3283 | calculate_dominance_info (CDI_DOMINATORS); |
36875e8f | 3284 | calculate_dominance_info (CDI_POST_DOMINATORS); |
30fd2977 | 3285 | cd = new control_dependences (); |
36875e8f RB |
3286 | free_dominance_info (CDI_POST_DOMINATORS); |
3287 | } | |
163aa51b | 3288 | |
b71b7a8e | 3289 | bool destroy_p; |
163aa51b | 3290 | int nb_generated_loops, nb_generated_calls; |
5879ab5f RB |
3291 | nb_generated_loops |
3292 | = distribute_loop (loop, work_list, cd, &nb_generated_calls, | |
3293 | &destroy_p, (!optimize_loop_for_speed_p (loop) | |
3294 | || !flag_tree_loop_distribution)); | |
b71b7a8e RB |
3295 | if (destroy_p) |
3296 | loops_to_be_destroyed.safe_push (loop); | |
c014f6f5 | 3297 | |
163aa51b BC |
3298 | if (nb_generated_loops + nb_generated_calls > 0) |
3299 | { | |
3300 | changed = true; | |
bbeeac91 DM |
3301 | if (dump_enabled_p ()) |
3302 | dump_printf_loc (MSG_OPTIMIZED_LOCATIONS, | |
3303 | loc, "Loop%s %d distributed: split to %d loops " | |
3304 | "and %d library calls.\n", str, loop->num, | |
3305 | nb_generated_loops, nb_generated_calls); | |
163aa51b BC |
3306 | |
3307 | break; | |
3308 | } | |
3309 | ||
3310 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
3311 | fprintf (dump_file, "Loop%s %d not distributed.\n", str, loop->num); | |
dea61d92 | 3312 | } |
dea61d92 SP |
3313 | } |
3314 | ||
36875e8f RB |
3315 | if (cd) |
3316 | delete cd; | |
3317 | ||
3be57c56 | 3318 | if (bb_top_order_index != NULL) |
eef99cd9 | 3319 | bb_top_order_destroy (); |
3be57c56 | 3320 | |
c014f6f5 RG |
3321 | if (changed) |
3322 | { | |
30fd2977 RB |
3323 | /* Destroy loop bodies that could not be reused. Do this late as we |
3324 | otherwise can end up refering to stale data in control dependences. */ | |
3325 | unsigned i; | |
3326 | FOR_EACH_VEC_ELT (loops_to_be_destroyed, i, loop) | |
3be57c56 | 3327 | destroy_loop (loop); |
30fd2977 | 3328 | |
d0ed943c RB |
3329 | /* Cached scalar evolutions now may refer to wrong or non-existing |
3330 | loops. */ | |
3331 | scev_reset_htab (); | |
be55bfe6 | 3332 | mark_virtual_operands_for_renaming (fun); |
c014f6f5 RG |
3333 | rewrite_into_loop_closed_ssa (NULL, TODO_update_ssa); |
3334 | } | |
3335 | ||
b2b29377 | 3336 | checking_verify_loop_structure (); |
c014f6f5 | 3337 | |
1bfb3b8b | 3338 | return changed ? TODO_cleanup_cfg : 0; |
dea61d92 SP |
3339 | } |
3340 | ||
eef99cd9 GB |
3341 | |
3342 | /* Distribute all loops in the current function. */ | |
3343 | ||
3344 | namespace { | |
3345 | ||
3346 | const pass_data pass_data_loop_distribution = | |
3347 | { | |
3348 | GIMPLE_PASS, /* type */ | |
3349 | "ldist", /* name */ | |
3350 | OPTGROUP_LOOP, /* optinfo_flags */ | |
3351 | TV_TREE_LOOP_DISTRIBUTION, /* tv_id */ | |
3352 | ( PROP_cfg | PROP_ssa ), /* properties_required */ | |
3353 | 0, /* properties_provided */ | |
3354 | 0, /* properties_destroyed */ | |
3355 | 0, /* todo_flags_start */ | |
3356 | 0, /* todo_flags_finish */ | |
3357 | }; | |
3358 | ||
3359 | class pass_loop_distribution : public gimple_opt_pass | |
3360 | { | |
3361 | public: | |
3362 | pass_loop_distribution (gcc::context *ctxt) | |
3363 | : gimple_opt_pass (pass_data_loop_distribution, ctxt) | |
3364 | {} | |
3365 | ||
3366 | /* opt_pass methods: */ | |
3367 | virtual bool gate (function *) | |
3368 | { | |
3369 | return flag_tree_loop_distribution | |
3370 | || flag_tree_loop_distribute_patterns; | |
3371 | } | |
3372 | ||
3373 | virtual unsigned int execute (function *); | |
3374 | ||
3375 | }; // class pass_loop_distribution | |
3376 | ||
3377 | unsigned int | |
3378 | pass_loop_distribution::execute (function *fun) | |
3379 | { | |
3380 | return loop_distribution ().execute (fun); | |
3381 | } | |
3382 | ||
27a4cd48 DM |
3383 | } // anon namespace |
3384 | ||
3385 | gimple_opt_pass * | |
3386 | make_pass_loop_distribution (gcc::context *ctxt) | |
3387 | { | |
3388 | return new pass_loop_distribution (ctxt); | |
3389 | } |