Mercurial > hg > CbC > CbC_gcc
annotate gcc/tree-loop-distribution.c @ 136:4627f235cf2a
fix c-next example
author | Shinji KONO <kono@ie.u-ryukyu.ac.jp> |
---|---|
date | Thu, 08 Nov 2018 14:11:56 +0900 |
parents | 84e7813d76e9 |
children | 1830386684a0 |
rev | line source |
---|---|
0 | 1 /* Loop distribution. |
131 | 2 Copyright (C) 2006-2018 Free Software Foundation, Inc. |
0 | 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. | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
7 |
0 | 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. | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
12 |
0 | 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. | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
17 |
0 | 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 | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
29 is transformed to |
0 | 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 | |
111 | 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: | |
86 1) We only distribute innermost two-level loop nest now. We should | |
87 extend it for arbitrary loop nests in the future. | |
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. */ | |
0 | 91 |
92 #include "config.h" | |
93 #include "system.h" | |
94 #include "coretypes.h" | |
111 | 95 #include "backend.h" |
96 #include "tree.h" | |
97 #include "gimple.h" | |
98 #include "cfghooks.h" | |
0 | 99 #include "tree-pass.h" |
111 | 100 #include "ssa.h" |
101 #include "gimple-pretty-print.h" | |
102 #include "fold-const.h" | |
103 #include "cfganal.h" | |
104 #include "gimple-iterator.h" | |
105 #include "gimplify-me.h" | |
106 #include "stor-layout.h" | |
107 #include "tree-cfg.h" | |
108 #include "tree-ssa-loop-manip.h" | |
109 #include "tree-ssa-loop-ivopts.h" | |
110 #include "tree-ssa-loop.h" | |
111 #include "tree-into-ssa.h" | |
112 #include "tree-ssa.h" | |
113 #include "cfgloop.h" | |
114 #include "tree-scalar-evolution.h" | |
115 #include "params.h" | |
116 #include "tree-vectorizer.h" | |
117 | |
118 | |
119 #define MAX_DATAREFS_NUM \ | |
120 ((unsigned) PARAM_VALUE (PARAM_LOOP_MAX_DATAREFS_FOR_DATADEPS)) | |
121 | |
122 /* Threshold controlling number of distributed partitions. Given it may | |
123 be unnecessary if a memory stream cost model is invented in the future, | |
124 we define it as a temporary macro, rather than a parameter. */ | |
125 #define NUM_PARTITION_THRESHOLD (4) | |
126 | |
127 /* Hashtable helpers. */ | |
128 | |
129 struct ddr_hasher : nofree_ptr_hash <struct data_dependence_relation> | |
130 { | |
131 static inline hashval_t hash (const data_dependence_relation *); | |
132 static inline bool equal (const data_dependence_relation *, | |
133 const data_dependence_relation *); | |
134 }; | |
135 | |
136 /* Hash function for data dependence. */ | |
137 | |
138 inline hashval_t | |
139 ddr_hasher::hash (const data_dependence_relation *ddr) | |
140 { | |
141 inchash::hash h; | |
142 h.add_ptr (DDR_A (ddr)); | |
143 h.add_ptr (DDR_B (ddr)); | |
144 return h.end (); | |
145 } | |
146 | |
147 /* Hash table equality function for data dependence. */ | |
148 | |
149 inline bool | |
150 ddr_hasher::equal (const data_dependence_relation *ddr1, | |
151 const data_dependence_relation *ddr2) | |
152 { | |
153 return (DDR_A (ddr1) == DDR_A (ddr2) && DDR_B (ddr1) == DDR_B (ddr2)); | |
154 } | |
155 | |
156 /* The loop (nest) to be distributed. */ | |
157 static vec<loop_p> loop_nest; | |
158 | |
159 /* Vector of data references in the loop to be distributed. */ | |
160 static vec<data_reference_p> datarefs_vec; | |
161 | |
162 /* Store index of data reference in aux field. */ | |
163 #define DR_INDEX(dr) ((uintptr_t) (dr)->aux) | |
164 | |
165 /* Hash table for data dependence relation in the loop to be distributed. */ | |
166 static hash_table<ddr_hasher> *ddrs_table; | |
167 | |
168 /* A Reduced Dependence Graph (RDG) vertex representing a statement. */ | |
169 struct rdg_vertex | |
170 { | |
171 /* The statement represented by this vertex. */ | |
172 gimple *stmt; | |
173 | |
174 /* Vector of data-references in this statement. */ | |
175 vec<data_reference_p> datarefs; | |
176 | |
177 /* True when the statement contains a write to memory. */ | |
178 bool has_mem_write; | |
179 | |
180 /* True when the statement contains a read from memory. */ | |
181 bool has_mem_reads; | |
182 }; | |
183 | |
184 #define RDGV_STMT(V) ((struct rdg_vertex *) ((V)->data))->stmt | |
185 #define RDGV_DATAREFS(V) ((struct rdg_vertex *) ((V)->data))->datarefs | |
186 #define RDGV_HAS_MEM_WRITE(V) ((struct rdg_vertex *) ((V)->data))->has_mem_write | |
187 #define RDGV_HAS_MEM_READS(V) ((struct rdg_vertex *) ((V)->data))->has_mem_reads | |
188 #define RDG_STMT(RDG, I) RDGV_STMT (&(RDG->vertices[I])) | |
189 #define RDG_DATAREFS(RDG, I) RDGV_DATAREFS (&(RDG->vertices[I])) | |
190 #define RDG_MEM_WRITE_STMT(RDG, I) RDGV_HAS_MEM_WRITE (&(RDG->vertices[I])) | |
191 #define RDG_MEM_READS_STMT(RDG, I) RDGV_HAS_MEM_READS (&(RDG->vertices[I])) | |
192 | |
193 /* Data dependence type. */ | |
194 | |
195 enum rdg_dep_type | |
196 { | |
197 /* Read After Write (RAW). */ | |
198 flow_dd = 'f', | |
199 | |
200 /* Control dependence (execute conditional on). */ | |
201 control_dd = 'c' | |
202 }; | |
203 | |
204 /* Dependence information attached to an edge of the RDG. */ | |
205 | |
206 struct rdg_edge | |
207 { | |
208 /* Type of the dependence. */ | |
209 enum rdg_dep_type type; | |
210 }; | |
211 | |
212 #define RDGE_TYPE(E) ((struct rdg_edge *) ((E)->data))->type | |
213 | |
214 /* Dump vertex I in RDG to FILE. */ | |
215 | |
216 static void | |
217 dump_rdg_vertex (FILE *file, struct graph *rdg, int i) | |
218 { | |
219 struct vertex *v = &(rdg->vertices[i]); | |
220 struct graph_edge *e; | |
221 | |
222 fprintf (file, "(vertex %d: (%s%s) (in:", i, | |
223 RDG_MEM_WRITE_STMT (rdg, i) ? "w" : "", | |
224 RDG_MEM_READS_STMT (rdg, i) ? "r" : ""); | |
225 | |
226 if (v->pred) | |
227 for (e = v->pred; e; e = e->pred_next) | |
228 fprintf (file, " %d", e->src); | |
229 | |
230 fprintf (file, ") (out:"); | |
231 | |
232 if (v->succ) | |
233 for (e = v->succ; e; e = e->succ_next) | |
234 fprintf (file, " %d", e->dest); | |
235 | |
236 fprintf (file, ")\n"); | |
237 print_gimple_stmt (file, RDGV_STMT (v), 0, TDF_VOPS|TDF_MEMSYMS); | |
238 fprintf (file, ")\n"); | |
239 } | |
240 | |
241 /* Call dump_rdg_vertex on stderr. */ | |
242 | |
243 DEBUG_FUNCTION void | |
244 debug_rdg_vertex (struct graph *rdg, int i) | |
245 { | |
246 dump_rdg_vertex (stderr, rdg, i); | |
247 } | |
248 | |
249 /* Dump the reduced dependence graph RDG to FILE. */ | |
250 | |
251 static void | |
252 dump_rdg (FILE *file, struct graph *rdg) | |
253 { | |
254 fprintf (file, "(rdg\n"); | |
255 for (int i = 0; i < rdg->n_vertices; i++) | |
256 dump_rdg_vertex (file, rdg, i); | |
257 fprintf (file, ")\n"); | |
258 } | |
259 | |
260 /* Call dump_rdg on stderr. */ | |
261 | |
262 DEBUG_FUNCTION void | |
263 debug_rdg (struct graph *rdg) | |
264 { | |
265 dump_rdg (stderr, rdg); | |
266 } | |
267 | |
268 static void | |
269 dot_rdg_1 (FILE *file, struct graph *rdg) | |
270 { | |
271 int i; | |
272 pretty_printer buffer; | |
273 pp_needs_newline (&buffer) = false; | |
274 buffer.buffer->stream = file; | |
275 | |
276 fprintf (file, "digraph RDG {\n"); | |
277 | |
278 for (i = 0; i < rdg->n_vertices; i++) | |
279 { | |
280 struct vertex *v = &(rdg->vertices[i]); | |
281 struct graph_edge *e; | |
282 | |
283 fprintf (file, "%d [label=\"[%d] ", i, i); | |
284 pp_gimple_stmt_1 (&buffer, RDGV_STMT (v), 0, TDF_SLIM); | |
285 pp_flush (&buffer); | |
286 fprintf (file, "\"]\n"); | |
287 | |
288 /* Highlight reads from memory. */ | |
289 if (RDG_MEM_READS_STMT (rdg, i)) | |
290 fprintf (file, "%d [style=filled, fillcolor=green]\n", i); | |
291 | |
292 /* Highlight stores to memory. */ | |
293 if (RDG_MEM_WRITE_STMT (rdg, i)) | |
294 fprintf (file, "%d [style=filled, fillcolor=red]\n", i); | |
295 | |
296 if (v->succ) | |
297 for (e = v->succ; e; e = e->succ_next) | |
298 switch (RDGE_TYPE (e)) | |
299 { | |
300 case flow_dd: | |
301 /* These are the most common dependences: don't print these. */ | |
302 fprintf (file, "%d -> %d \n", i, e->dest); | |
303 break; | |
304 | |
305 case control_dd: | |
306 fprintf (file, "%d -> %d [label=control] \n", i, e->dest); | |
307 break; | |
308 | |
309 default: | |
310 gcc_unreachable (); | |
311 } | |
312 } | |
313 | |
314 fprintf (file, "}\n\n"); | |
315 } | |
316 | |
317 /* Display the Reduced Dependence Graph using dotty. */ | |
318 | |
319 DEBUG_FUNCTION void | |
320 dot_rdg (struct graph *rdg) | |
321 { | |
322 /* When debugging, you may want to enable the following code. */ | |
323 #ifdef HAVE_POPEN | |
324 FILE *file = popen ("dot -Tx11", "w"); | |
325 if (!file) | |
326 return; | |
327 dot_rdg_1 (file, rdg); | |
328 fflush (file); | |
329 close (fileno (file)); | |
330 pclose (file); | |
331 #else | |
332 dot_rdg_1 (stderr, rdg); | |
333 #endif | |
334 } | |
335 | |
336 /* Returns the index of STMT in RDG. */ | |
337 | |
338 static int | |
339 rdg_vertex_for_stmt (struct graph *rdg ATTRIBUTE_UNUSED, gimple *stmt) | |
340 { | |
341 int index = gimple_uid (stmt); | |
342 gcc_checking_assert (index == -1 || RDG_STMT (rdg, index) == stmt); | |
343 return index; | |
344 } | |
345 | |
346 /* Creates dependence edges in RDG for all the uses of DEF. IDEF is | |
347 the index of DEF in RDG. */ | |
348 | |
349 static void | |
350 create_rdg_edges_for_scalar (struct graph *rdg, tree def, int idef) | |
351 { | |
352 use_operand_p imm_use_p; | |
353 imm_use_iterator iterator; | |
354 | |
355 FOR_EACH_IMM_USE_FAST (imm_use_p, iterator, def) | |
356 { | |
357 struct graph_edge *e; | |
358 int use = rdg_vertex_for_stmt (rdg, USE_STMT (imm_use_p)); | |
359 | |
360 if (use < 0) | |
361 continue; | |
362 | |
363 e = add_edge (rdg, idef, use); | |
364 e->data = XNEW (struct rdg_edge); | |
365 RDGE_TYPE (e) = flow_dd; | |
366 } | |
367 } | |
368 | |
369 /* Creates an edge for the control dependences of BB to the vertex V. */ | |
370 | |
371 static void | |
372 create_edge_for_control_dependence (struct graph *rdg, basic_block bb, | |
373 int v, control_dependences *cd) | |
374 { | |
375 bitmap_iterator bi; | |
376 unsigned edge_n; | |
377 EXECUTE_IF_SET_IN_BITMAP (cd->get_edges_dependent_on (bb->index), | |
378 0, edge_n, bi) | |
379 { | |
380 basic_block cond_bb = cd->get_edge_src (edge_n); | |
381 gimple *stmt = last_stmt (cond_bb); | |
382 if (stmt && is_ctrl_stmt (stmt)) | |
383 { | |
384 struct graph_edge *e; | |
385 int c = rdg_vertex_for_stmt (rdg, stmt); | |
386 if (c < 0) | |
387 continue; | |
388 | |
389 e = add_edge (rdg, c, v); | |
390 e->data = XNEW (struct rdg_edge); | |
391 RDGE_TYPE (e) = control_dd; | |
392 } | |
393 } | |
394 } | |
395 | |
396 /* Creates the edges of the reduced dependence graph RDG. */ | |
0 | 397 |
398 static void | |
111 | 399 create_rdg_flow_edges (struct graph *rdg) |
0 | 400 { |
111 | 401 int i; |
402 def_operand_p def_p; | |
403 ssa_op_iter iter; | |
404 | |
405 for (i = 0; i < rdg->n_vertices; i++) | |
406 FOR_EACH_PHI_OR_STMT_DEF (def_p, RDG_STMT (rdg, i), | |
407 iter, SSA_OP_DEF) | |
408 create_rdg_edges_for_scalar (rdg, DEF_FROM_PTR (def_p), i); | |
409 } | |
410 | |
411 /* Creates the edges of the reduced dependence graph RDG. */ | |
412 | |
413 static void | |
414 create_rdg_cd_edges (struct graph *rdg, control_dependences *cd, loop_p loop) | |
415 { | |
416 int i; | |
417 | |
418 for (i = 0; i < rdg->n_vertices; i++) | |
0 | 419 { |
111 | 420 gimple *stmt = RDG_STMT (rdg, i); |
421 if (gimple_code (stmt) == GIMPLE_PHI) | |
0 | 422 { |
111 | 423 edge_iterator ei; |
424 edge e; | |
425 FOR_EACH_EDGE (e, ei, gimple_bb (stmt)->preds) | |
426 if (flow_bb_inside_loop_p (loop, e->src)) | |
427 create_edge_for_control_dependence (rdg, e->src, i, cd); | |
0 | 428 } |
429 else | |
111 | 430 create_edge_for_control_dependence (rdg, gimple_bb (stmt), i, cd); |
431 } | |
432 } | |
433 | |
434 /* Build the vertices of the reduced dependence graph RDG. Return false | |
435 if that failed. */ | |
436 | |
437 static bool | |
438 create_rdg_vertices (struct graph *rdg, vec<gimple *> stmts, loop_p loop) | |
439 { | |
440 int i; | |
441 gimple *stmt; | |
442 | |
443 FOR_EACH_VEC_ELT (stmts, i, stmt) | |
444 { | |
445 struct vertex *v = &(rdg->vertices[i]); | |
446 | |
447 /* Record statement to vertex mapping. */ | |
448 gimple_set_uid (stmt, i); | |
449 | |
450 v->data = XNEW (struct rdg_vertex); | |
451 RDGV_STMT (v) = stmt; | |
452 RDGV_DATAREFS (v).create (0); | |
453 RDGV_HAS_MEM_WRITE (v) = false; | |
454 RDGV_HAS_MEM_READS (v) = false; | |
455 if (gimple_code (stmt) == GIMPLE_PHI) | |
456 continue; | |
457 | |
458 unsigned drp = datarefs_vec.length (); | |
459 if (!find_data_references_in_stmt (loop, stmt, &datarefs_vec)) | |
460 return false; | |
461 for (unsigned j = drp; j < datarefs_vec.length (); ++j) | |
462 { | |
463 data_reference_p dr = datarefs_vec[j]; | |
464 if (DR_IS_READ (dr)) | |
465 RDGV_HAS_MEM_READS (v) = true; | |
466 else | |
467 RDGV_HAS_MEM_WRITE (v) = true; | |
468 RDGV_DATAREFS (v).safe_push (dr); | |
469 } | |
470 } | |
471 return true; | |
472 } | |
473 | |
474 /* Array mapping basic block's index to its topological order. */ | |
475 static int *bb_top_order_index; | |
476 /* And size of the array. */ | |
477 static int bb_top_order_index_size; | |
478 | |
479 /* If X has a smaller topological sort number than Y, returns -1; | |
480 if greater, returns 1. */ | |
481 | |
482 static int | |
483 bb_top_order_cmp (const void *x, const void *y) | |
484 { | |
485 basic_block bb1 = *(const basic_block *) x; | |
486 basic_block bb2 = *(const basic_block *) y; | |
487 | |
488 gcc_assert (bb1->index < bb_top_order_index_size | |
489 && bb2->index < bb_top_order_index_size); | |
490 gcc_assert (bb1 == bb2 | |
491 || bb_top_order_index[bb1->index] | |
492 != bb_top_order_index[bb2->index]); | |
493 | |
494 return (bb_top_order_index[bb1->index] - bb_top_order_index[bb2->index]); | |
495 } | |
496 | |
497 /* Initialize STMTS with all the statements of LOOP. We use topological | |
498 order to discover all statements. The order is important because | |
499 generate_loops_for_partition is using the same traversal for identifying | |
500 statements in loop copies. */ | |
501 | |
502 static void | |
503 stmts_from_loop (struct loop *loop, vec<gimple *> *stmts) | |
504 { | |
505 unsigned int i; | |
506 basic_block *bbs = get_loop_body_in_custom_order (loop, bb_top_order_cmp); | |
507 | |
508 for (i = 0; i < loop->num_nodes; i++) | |
509 { | |
510 basic_block bb = bbs[i]; | |
511 | |
512 for (gphi_iterator bsi = gsi_start_phis (bb); !gsi_end_p (bsi); | |
513 gsi_next (&bsi)) | |
514 if (!virtual_operand_p (gimple_phi_result (bsi.phi ()))) | |
515 stmts->safe_push (bsi.phi ()); | |
516 | |
517 for (gimple_stmt_iterator bsi = gsi_start_bb (bb); !gsi_end_p (bsi); | |
518 gsi_next (&bsi)) | |
519 { | |
520 gimple *stmt = gsi_stmt (bsi); | |
521 if (gimple_code (stmt) != GIMPLE_LABEL && !is_gimple_debug (stmt)) | |
522 stmts->safe_push (stmt); | |
523 } | |
524 } | |
525 | |
526 free (bbs); | |
527 } | |
528 | |
529 /* Free the reduced dependence graph RDG. */ | |
530 | |
531 static void | |
532 free_rdg (struct graph *rdg) | |
533 { | |
534 int i; | |
535 | |
536 for (i = 0; i < rdg->n_vertices; i++) | |
537 { | |
538 struct vertex *v = &(rdg->vertices[i]); | |
539 struct graph_edge *e; | |
540 | |
541 for (e = v->succ; e; e = e->succ_next) | |
542 free (e->data); | |
543 | |
544 if (v->data) | |
545 { | |
546 gimple_set_uid (RDGV_STMT (v), -1); | |
547 (RDGV_DATAREFS (v)).release (); | |
548 free (v->data); | |
549 } | |
550 } | |
551 | |
552 free_graph (rdg); | |
553 } | |
554 | |
555 /* Build the Reduced Dependence Graph (RDG) with one vertex per statement of | |
556 LOOP, and one edge per flow dependence or control dependence from control | |
557 dependence CD. During visiting each statement, data references are also | |
558 collected and recorded in global data DATAREFS_VEC. */ | |
559 | |
560 static struct graph * | |
561 build_rdg (struct loop *loop, control_dependences *cd) | |
562 { | |
563 struct graph *rdg; | |
564 | |
565 /* Create the RDG vertices from the stmts of the loop nest. */ | |
566 auto_vec<gimple *, 10> stmts; | |
567 stmts_from_loop (loop, &stmts); | |
568 rdg = new_graph (stmts.length ()); | |
569 if (!create_rdg_vertices (rdg, stmts, loop)) | |
570 { | |
571 free_rdg (rdg); | |
572 return NULL; | |
0 | 573 } |
111 | 574 stmts.release (); |
575 | |
576 create_rdg_flow_edges (rdg); | |
577 if (cd) | |
578 create_rdg_cd_edges (rdg, cd, loop); | |
579 | |
580 return rdg; | |
581 } | |
582 | |
583 | |
584 /* Kind of distributed loop. */ | |
585 enum partition_kind { | |
131 | 586 PKIND_NORMAL, |
587 /* Partial memset stands for a paritition can be distributed into a loop | |
588 of memset calls, rather than a single memset call. It's handled just | |
589 like a normal parition, i.e, distributed as separate loop, no memset | |
590 call is generated. | |
591 | |
592 Note: This is a hacking fix trying to distribute ZERO-ing stmt in a | |
593 loop nest as deep as possible. As a result, parloop achieves better | |
594 parallelization by parallelizing deeper loop nest. This hack should | |
595 be unnecessary and removed once distributed memset can be understood | |
596 and analyzed in data reference analysis. See PR82604 for more. */ | |
597 PKIND_PARTIAL_MEMSET, | |
598 PKIND_MEMSET, PKIND_MEMCPY, PKIND_MEMMOVE | |
111 | 599 }; |
600 | |
601 /* Type of distributed loop. */ | |
602 enum partition_type { | |
603 /* The distributed loop can be executed parallelly. */ | |
604 PTYPE_PARALLEL = 0, | |
605 /* The distributed loop has to be executed sequentially. */ | |
606 PTYPE_SEQUENTIAL | |
607 }; | |
608 | |
609 /* Builtin info for loop distribution. */ | |
610 struct builtin_info | |
611 { | |
612 /* data-references a kind != PKIND_NORMAL partition is about. */ | |
613 data_reference_p dst_dr; | |
614 data_reference_p src_dr; | |
615 /* Base address and size of memory objects operated by the builtin. Note | |
616 both dest and source memory objects must have the same size. */ | |
617 tree dst_base; | |
618 tree src_base; | |
619 tree size; | |
620 /* Base and offset part of dst_base after stripping constant offset. This | |
621 is only used in memset builtin distribution for now. */ | |
622 tree dst_base_base; | |
623 unsigned HOST_WIDE_INT dst_base_offset; | |
624 }; | |
625 | |
626 /* Partition for loop distribution. */ | |
627 struct partition | |
628 { | |
629 /* Statements of the partition. */ | |
630 bitmap stmts; | |
631 /* True if the partition defines variable which is used outside of loop. */ | |
632 bool reduction_p; | |
633 enum partition_kind kind; | |
634 enum partition_type type; | |
635 /* Data references in the partition. */ | |
636 bitmap datarefs; | |
637 /* Information of builtin parition. */ | |
638 struct builtin_info *builtin; | |
639 }; | |
640 | |
641 | |
642 /* Allocate and initialize a partition from BITMAP. */ | |
643 | |
644 static partition * | |
645 partition_alloc (void) | |
646 { | |
647 partition *partition = XCNEW (struct partition); | |
648 partition->stmts = BITMAP_ALLOC (NULL); | |
649 partition->reduction_p = false; | |
650 partition->kind = PKIND_NORMAL; | |
651 partition->datarefs = BITMAP_ALLOC (NULL); | |
652 return partition; | |
653 } | |
654 | |
655 /* Free PARTITION. */ | |
656 | |
657 static void | |
658 partition_free (partition *partition) | |
659 { | |
660 BITMAP_FREE (partition->stmts); | |
661 BITMAP_FREE (partition->datarefs); | |
662 if (partition->builtin) | |
663 free (partition->builtin); | |
664 | |
665 free (partition); | |
666 } | |
667 | |
668 /* Returns true if the partition can be generated as a builtin. */ | |
669 | |
670 static bool | |
671 partition_builtin_p (partition *partition) | |
672 { | |
131 | 673 return partition->kind > PKIND_PARTIAL_MEMSET; |
111 | 674 } |
675 | |
676 /* Returns true if the partition contains a reduction. */ | |
677 | |
678 static bool | |
679 partition_reduction_p (partition *partition) | |
680 { | |
681 return partition->reduction_p; | |
682 } | |
683 | |
684 /* Partitions are fused because of different reasons. */ | |
685 enum fuse_type | |
686 { | |
687 FUSE_NON_BUILTIN = 0, | |
688 FUSE_REDUCTION = 1, | |
689 FUSE_SHARE_REF = 2, | |
690 FUSE_SAME_SCC = 3, | |
691 FUSE_FINALIZE = 4 | |
692 }; | |
693 | |
694 /* Description on different fusing reason. */ | |
695 static const char *fuse_message[] = { | |
696 "they are non-builtins", | |
697 "they have reductions", | |
698 "they have shared memory refs", | |
699 "they are in the same dependence scc", | |
700 "there is no point to distribute loop"}; | |
701 | |
702 static void | |
703 update_type_for_merge (struct graph *, partition *, partition *); | |
704 | |
705 /* Merge PARTITION into the partition DEST. RDG is the reduced dependence | |
706 graph and we update type for result partition if it is non-NULL. */ | |
707 | |
708 static void | |
709 partition_merge_into (struct graph *rdg, partition *dest, | |
710 partition *partition, enum fuse_type ft) | |
711 { | |
712 if (dump_file && (dump_flags & TDF_DETAILS)) | |
713 { | |
714 fprintf (dump_file, "Fuse partitions because %s:\n", fuse_message[ft]); | |
715 fprintf (dump_file, " Part 1: "); | |
716 dump_bitmap (dump_file, dest->stmts); | |
717 fprintf (dump_file, " Part 2: "); | |
718 dump_bitmap (dump_file, partition->stmts); | |
719 } | |
720 | |
721 dest->kind = PKIND_NORMAL; | |
722 if (dest->type == PTYPE_PARALLEL) | |
723 dest->type = partition->type; | |
724 | |
725 bitmap_ior_into (dest->stmts, partition->stmts); | |
726 if (partition_reduction_p (partition)) | |
727 dest->reduction_p = true; | |
728 | |
729 /* Further check if any data dependence prevents us from executing the | |
730 new partition parallelly. */ | |
731 if (dest->type == PTYPE_PARALLEL && rdg != NULL) | |
732 update_type_for_merge (rdg, dest, partition); | |
733 | |
734 bitmap_ior_into (dest->datarefs, partition->datarefs); | |
735 } | |
736 | |
737 | |
738 /* Returns true when DEF is an SSA_NAME defined in LOOP and used after | |
739 the LOOP. */ | |
740 | |
741 static bool | |
742 ssa_name_has_uses_outside_loop_p (tree def, loop_p loop) | |
743 { | |
744 imm_use_iterator imm_iter; | |
745 use_operand_p use_p; | |
746 | |
747 FOR_EACH_IMM_USE_FAST (use_p, imm_iter, def) | |
748 { | |
749 if (is_gimple_debug (USE_STMT (use_p))) | |
750 continue; | |
751 | |
752 basic_block use_bb = gimple_bb (USE_STMT (use_p)); | |
753 if (!flow_bb_inside_loop_p (loop, use_bb)) | |
754 return true; | |
755 } | |
756 | |
757 return false; | |
758 } | |
759 | |
760 /* Returns true when STMT defines a scalar variable used after the | |
761 loop LOOP. */ | |
762 | |
763 static bool | |
764 stmt_has_scalar_dependences_outside_loop (loop_p loop, gimple *stmt) | |
765 { | |
766 def_operand_p def_p; | |
767 ssa_op_iter op_iter; | |
768 | |
769 if (gimple_code (stmt) == GIMPLE_PHI) | |
770 return ssa_name_has_uses_outside_loop_p (gimple_phi_result (stmt), loop); | |
771 | |
772 FOR_EACH_SSA_DEF_OPERAND (def_p, stmt, op_iter, SSA_OP_DEF) | |
773 if (ssa_name_has_uses_outside_loop_p (DEF_FROM_PTR (def_p), loop)) | |
774 return true; | |
775 | |
776 return false; | |
0 | 777 } |
778 | |
779 /* Return a copy of LOOP placed before LOOP. */ | |
780 | |
781 static struct loop * | |
782 copy_loop_before (struct loop *loop) | |
783 { | |
784 struct loop *res; | |
785 edge preheader = loop_preheader_edge (loop); | |
786 | |
787 initialize_original_copy_tables (); | |
111 | 788 res = slpeel_tree_duplicate_loop_to_edge_cfg (loop, NULL, preheader); |
789 gcc_assert (res != NULL); | |
0 | 790 free_original_copy_tables (); |
111 | 791 delete_update_ssa (); |
0 | 792 |
793 return res; | |
794 } | |
795 | |
796 /* Creates an empty basic block after LOOP. */ | |
797 | |
798 static void | |
799 create_bb_after_loop (struct loop *loop) | |
800 { | |
801 edge exit = single_exit (loop); | |
802 | |
803 if (!exit) | |
804 return; | |
805 | |
806 split_edge (exit); | |
807 } | |
808 | |
809 /* Generate code for PARTITION from the code in LOOP. The loop is | |
810 copied when COPY_P is true. All the statements not flagged in the | |
811 PARTITION bitmap are removed from the loop or from its copy. The | |
812 statements are indexed in sequence inside a basic block, and the | |
111 | 813 basic blocks of a loop are taken in dom order. */ |
814 | |
815 static void | |
816 generate_loops_for_partition (struct loop *loop, partition *partition, | |
817 bool copy_p) | |
0 | 818 { |
111 | 819 unsigned i; |
0 | 820 basic_block *bbs; |
821 | |
822 if (copy_p) | |
823 { | |
111 | 824 int orig_loop_num = loop->orig_loop_num; |
0 | 825 loop = copy_loop_before (loop); |
111 | 826 gcc_assert (loop != NULL); |
827 loop->orig_loop_num = orig_loop_num; | |
0 | 828 create_preheader (loop, CP_SIMPLE_PREHEADERS); |
829 create_bb_after_loop (loop); | |
830 } | |
111 | 831 else |
0 | 832 { |
111 | 833 /* Origin number is set to the new versioned loop's num. */ |
834 gcc_assert (loop->orig_loop_num != loop->num); | |
0 | 835 } |
836 | |
111 | 837 /* Remove stmts not in the PARTITION bitmap. */ |
0 | 838 bbs = get_loop_body_in_dom_order (loop); |
839 | |
131 | 840 if (MAY_HAVE_DEBUG_BIND_STMTS) |
111 | 841 for (i = 0; i < loop->num_nodes; i++) |
842 { | |
843 basic_block bb = bbs[i]; | |
844 | |
845 for (gphi_iterator bsi = gsi_start_phis (bb); !gsi_end_p (bsi); | |
846 gsi_next (&bsi)) | |
847 { | |
848 gphi *phi = bsi.phi (); | |
849 if (!virtual_operand_p (gimple_phi_result (phi)) | |
850 && !bitmap_bit_p (partition->stmts, gimple_uid (phi))) | |
851 reset_debug_uses (phi); | |
852 } | |
853 | |
854 for (gimple_stmt_iterator bsi = gsi_start_bb (bb); !gsi_end_p (bsi); gsi_next (&bsi)) | |
855 { | |
856 gimple *stmt = gsi_stmt (bsi); | |
857 if (gimple_code (stmt) != GIMPLE_LABEL | |
858 && !is_gimple_debug (stmt) | |
859 && !bitmap_bit_p (partition->stmts, gimple_uid (stmt))) | |
860 reset_debug_uses (stmt); | |
861 } | |
862 } | |
863 | |
0 | 864 for (i = 0; i < loop->num_nodes; i++) |
865 { | |
866 basic_block bb = bbs[i]; | |
111 | 867 edge inner_exit = NULL; |
868 | |
869 if (loop != bb->loop_father) | |
870 inner_exit = single_exit (bb->loop_father); | |
871 | |
872 for (gphi_iterator bsi = gsi_start_phis (bb); !gsi_end_p (bsi);) | |
873 { | |
874 gphi *phi = bsi.phi (); | |
875 if (!virtual_operand_p (gimple_phi_result (phi)) | |
876 && !bitmap_bit_p (partition->stmts, gimple_uid (phi))) | |
877 remove_phi_node (&bsi, true); | |
878 else | |
879 gsi_next (&bsi); | |
880 } | |
881 | |
882 for (gimple_stmt_iterator bsi = gsi_start_bb (bb); !gsi_end_p (bsi);) | |
0 | 883 { |
111 | 884 gimple *stmt = gsi_stmt (bsi); |
885 if (gimple_code (stmt) != GIMPLE_LABEL | |
886 && !is_gimple_debug (stmt) | |
887 && !bitmap_bit_p (partition->stmts, gimple_uid (stmt))) | |
0 | 888 { |
111 | 889 /* In distribution of loop nest, if bb is inner loop's exit_bb, |
890 we choose its exit edge/path in order to avoid generating | |
891 infinite loop. For all other cases, we choose an arbitrary | |
892 path through the empty CFG part that this unnecessary | |
893 control stmt controls. */ | |
894 if (gcond *cond_stmt = dyn_cast <gcond *> (stmt)) | |
895 { | |
896 if (inner_exit && inner_exit->flags & EDGE_TRUE_VALUE) | |
897 gimple_cond_make_true (cond_stmt); | |
898 else | |
899 gimple_cond_make_false (cond_stmt); | |
900 update_stmt (stmt); | |
901 } | |
902 else if (gimple_code (stmt) == GIMPLE_SWITCH) | |
903 { | |
904 gswitch *switch_stmt = as_a <gswitch *> (stmt); | |
905 gimple_switch_set_index | |
906 (switch_stmt, CASE_LOW (gimple_switch_label (switch_stmt, 1))); | |
907 update_stmt (stmt); | |
908 } | |
909 else | |
910 { | |
911 unlink_stmt_vdef (stmt); | |
912 gsi_remove (&bsi, true); | |
913 release_defs (stmt); | |
914 continue; | |
915 } | |
0 | 916 } |
111 | 917 gsi_next (&bsi); |
0 | 918 } |
919 } | |
920 | |
111 | 921 free (bbs); |
922 } | |
923 | |
924 /* If VAL memory representation contains the same value in all bytes, | |
925 return that value, otherwise return -1. | |
926 E.g. for 0x24242424 return 0x24, for IEEE double | |
927 747708026454360457216.0 return 0x44, etc. */ | |
928 | |
929 static int | |
930 const_with_all_bytes_same (tree val) | |
931 { | |
932 unsigned char buf[64]; | |
933 int i, len; | |
934 | |
935 if (integer_zerop (val) | |
936 || (TREE_CODE (val) == CONSTRUCTOR | |
937 && !TREE_CLOBBER_P (val) | |
938 && CONSTRUCTOR_NELTS (val) == 0)) | |
939 return 0; | |
940 | |
941 if (real_zerop (val)) | |
942 { | |
943 /* Only return 0 for +0.0, not for -0.0, which doesn't have | |
944 an all bytes same memory representation. Don't transform | |
945 -0.0 stores into +0.0 even for !HONOR_SIGNED_ZEROS. */ | |
946 switch (TREE_CODE (val)) | |
947 { | |
948 case REAL_CST: | |
949 if (!real_isneg (TREE_REAL_CST_PTR (val))) | |
950 return 0; | |
951 break; | |
952 case COMPLEX_CST: | |
953 if (!const_with_all_bytes_same (TREE_REALPART (val)) | |
954 && !const_with_all_bytes_same (TREE_IMAGPART (val))) | |
955 return 0; | |
956 break; | |
957 case VECTOR_CST: | |
131 | 958 { |
959 unsigned int count = vector_cst_encoded_nelts (val); | |
960 unsigned int j; | |
961 for (j = 0; j < count; ++j) | |
962 if (const_with_all_bytes_same (VECTOR_CST_ENCODED_ELT (val, j))) | |
963 break; | |
964 if (j == count) | |
965 return 0; | |
966 break; | |
967 } | |
111 | 968 default: |
969 break; | |
970 } | |
971 } | |
972 | |
973 if (CHAR_BIT != 8 || BITS_PER_UNIT != 8) | |
974 return -1; | |
975 | |
976 len = native_encode_expr (val, buf, sizeof (buf)); | |
977 if (len == 0) | |
978 return -1; | |
979 for (i = 1; i < len; i++) | |
980 if (buf[i] != buf[0]) | |
981 return -1; | |
982 return buf[0]; | |
983 } | |
984 | |
985 /* Generate a call to memset for PARTITION in LOOP. */ | |
986 | |
987 static void | |
988 generate_memset_builtin (struct loop *loop, partition *partition) | |
989 { | |
990 gimple_stmt_iterator gsi; | |
991 tree mem, fn, nb_bytes; | |
992 tree val; | |
993 struct builtin_info *builtin = partition->builtin; | |
994 gimple *fn_call; | |
0 | 995 |
996 /* The new statements will be placed before LOOP. */ | |
111 | 997 gsi = gsi_last_bb (loop_preheader_edge (loop)->src); |
998 | |
999 nb_bytes = builtin->size; | |
1000 nb_bytes = force_gimple_operand_gsi (&gsi, nb_bytes, true, NULL_TREE, | |
1001 false, GSI_CONTINUE_LINKING); | |
1002 mem = builtin->dst_base; | |
1003 mem = force_gimple_operand_gsi (&gsi, mem, true, NULL_TREE, | |
1004 false, GSI_CONTINUE_LINKING); | |
1005 | |
1006 /* This exactly matches the pattern recognition in classify_partition. */ | |
1007 val = gimple_assign_rhs1 (DR_STMT (builtin->dst_dr)); | |
1008 /* Handle constants like 0x15151515 and similarly | |
1009 floating point constants etc. where all bytes are the same. */ | |
1010 int bytev = const_with_all_bytes_same (val); | |
1011 if (bytev != -1) | |
1012 val = build_int_cst (integer_type_node, bytev); | |
1013 else if (TREE_CODE (val) == INTEGER_CST) | |
1014 val = fold_convert (integer_type_node, val); | |
1015 else if (!useless_type_conversion_p (integer_type_node, TREE_TYPE (val))) | |
1016 { | |
1017 tree tem = make_ssa_name (integer_type_node); | |
1018 gimple *cstmt = gimple_build_assign (tem, NOP_EXPR, val); | |
1019 gsi_insert_after (&gsi, cstmt, GSI_CONTINUE_LINKING); | |
1020 val = tem; | |
1021 } | |
1022 | |
1023 fn = build_fold_addr_expr (builtin_decl_implicit (BUILT_IN_MEMSET)); | |
1024 fn_call = gimple_build_call (fn, 3, mem, val, nb_bytes); | |
1025 gsi_insert_after (&gsi, fn_call, GSI_CONTINUE_LINKING); | |
1026 | |
1027 if (dump_file && (dump_flags & TDF_DETAILS)) | |
0 | 1028 { |
111 | 1029 fprintf (dump_file, "generated memset"); |
1030 if (bytev == 0) | |
1031 fprintf (dump_file, " zero\n"); | |
1032 else | |
1033 fprintf (dump_file, "\n"); | |
1034 } | |
1035 } | |
1036 | |
1037 /* Generate a call to memcpy for PARTITION in LOOP. */ | |
1038 | |
1039 static void | |
1040 generate_memcpy_builtin (struct loop *loop, partition *partition) | |
1041 { | |
1042 gimple_stmt_iterator gsi; | |
1043 gimple *fn_call; | |
1044 tree dest, src, fn, nb_bytes; | |
1045 enum built_in_function kind; | |
1046 struct builtin_info *builtin = partition->builtin; | |
1047 | |
1048 /* The new statements will be placed before LOOP. */ | |
1049 gsi = gsi_last_bb (loop_preheader_edge (loop)->src); | |
1050 | |
1051 nb_bytes = builtin->size; | |
1052 nb_bytes = force_gimple_operand_gsi (&gsi, nb_bytes, true, NULL_TREE, | |
1053 false, GSI_CONTINUE_LINKING); | |
1054 dest = builtin->dst_base; | |
1055 src = builtin->src_base; | |
1056 if (partition->kind == PKIND_MEMCPY | |
1057 || ! ptr_derefs_may_alias_p (dest, src)) | |
1058 kind = BUILT_IN_MEMCPY; | |
1059 else | |
1060 kind = BUILT_IN_MEMMOVE; | |
1061 | |
1062 dest = force_gimple_operand_gsi (&gsi, dest, true, NULL_TREE, | |
1063 false, GSI_CONTINUE_LINKING); | |
1064 src = force_gimple_operand_gsi (&gsi, src, true, NULL_TREE, | |
1065 false, GSI_CONTINUE_LINKING); | |
1066 fn = build_fold_addr_expr (builtin_decl_implicit (kind)); | |
1067 fn_call = gimple_build_call (fn, 3, dest, src, nb_bytes); | |
1068 gsi_insert_after (&gsi, fn_call, GSI_CONTINUE_LINKING); | |
1069 | |
1070 if (dump_file && (dump_flags & TDF_DETAILS)) | |
1071 { | |
1072 if (kind == BUILT_IN_MEMCPY) | |
1073 fprintf (dump_file, "generated memcpy\n"); | |
1074 else | |
1075 fprintf (dump_file, "generated memmove\n"); | |
0 | 1076 } |
1077 } | |
1078 | |
111 | 1079 /* Remove and destroy the loop LOOP. */ |
1080 | |
1081 static void | |
1082 destroy_loop (struct loop *loop) | |
0 | 1083 { |
111 | 1084 unsigned nbbs = loop->num_nodes; |
1085 edge exit = single_exit (loop); | |
1086 basic_block src = loop_preheader_edge (loop)->src, dest = exit->dest; | |
1087 basic_block *bbs; | |
1088 unsigned i; | |
1089 | |
1090 bbs = get_loop_body_in_dom_order (loop); | |
1091 | |
1092 redirect_edge_pred (exit, src); | |
1093 exit->flags &= ~(EDGE_TRUE_VALUE|EDGE_FALSE_VALUE); | |
1094 exit->flags |= EDGE_FALLTHRU; | |
1095 cancel_loop_tree (loop); | |
1096 rescan_loop_exit (exit, false, true); | |
1097 | |
1098 i = nbbs; | |
1099 do | |
1100 { | |
1101 /* We have made sure to not leave any dangling uses of SSA | |
1102 names defined in the loop. With the exception of virtuals. | |
1103 Make sure we replace all uses of virtual defs that will remain | |
1104 outside of the loop with the bare symbol as delete_basic_block | |
1105 will release them. */ | |
1106 --i; | |
1107 for (gphi_iterator gsi = gsi_start_phis (bbs[i]); !gsi_end_p (gsi); | |
1108 gsi_next (&gsi)) | |
1109 { | |
1110 gphi *phi = gsi.phi (); | |
1111 if (virtual_operand_p (gimple_phi_result (phi))) | |
1112 mark_virtual_phi_result_for_renaming (phi); | |
1113 } | |
1114 for (gimple_stmt_iterator gsi = gsi_start_bb (bbs[i]); !gsi_end_p (gsi); | |
1115 gsi_next (&gsi)) | |
1116 { | |
1117 gimple *stmt = gsi_stmt (gsi); | |
1118 tree vdef = gimple_vdef (stmt); | |
1119 if (vdef && TREE_CODE (vdef) == SSA_NAME) | |
1120 mark_virtual_operand_for_renaming (vdef); | |
1121 } | |
1122 delete_basic_block (bbs[i]); | |
1123 } | |
1124 while (i != 0); | |
1125 | |
1126 free (bbs); | |
1127 | |
1128 set_immediate_dominator (CDI_DOMINATORS, dest, | |
1129 recompute_dominator (CDI_DOMINATORS, dest)); | |
1130 } | |
1131 | |
1132 /* Generates code for PARTITION. Return whether LOOP needs to be destroyed. */ | |
1133 | |
1134 static bool | |
1135 generate_code_for_partition (struct loop *loop, | |
1136 partition *partition, bool copy_p) | |
1137 { | |
1138 switch (partition->kind) | |
1139 { | |
1140 case PKIND_NORMAL: | |
131 | 1141 case PKIND_PARTIAL_MEMSET: |
111 | 1142 /* Reductions all have to be in the last partition. */ |
1143 gcc_assert (!partition_reduction_p (partition) | |
1144 || !copy_p); | |
1145 generate_loops_for_partition (loop, partition, copy_p); | |
1146 return false; | |
1147 | |
1148 case PKIND_MEMSET: | |
1149 generate_memset_builtin (loop, partition); | |
1150 break; | |
1151 | |
1152 case PKIND_MEMCPY: | |
1153 case PKIND_MEMMOVE: | |
1154 generate_memcpy_builtin (loop, partition); | |
1155 break; | |
1156 | |
1157 default: | |
1158 gcc_unreachable (); | |
1159 } | |
1160 | |
1161 /* Common tail for partitions we turn into a call. If this was the last | |
1162 partition for which we generate code, we have to destroy the loop. */ | |
1163 if (!copy_p) | |
0 | 1164 return true; |
1165 return false; | |
1166 } | |
1167 | |
111 | 1168 /* Return data dependence relation for data references A and B. The two |
1169 data references must be in lexicographic order wrto reduced dependence | |
1170 graph RDG. We firstly try to find ddr from global ddr hash table. If | |
1171 it doesn't exist, compute the ddr and cache it. */ | |
1172 | |
1173 static data_dependence_relation * | |
1174 get_data_dependence (struct graph *rdg, data_reference_p a, data_reference_p b) | |
0 | 1175 { |
111 | 1176 struct data_dependence_relation ent, **slot; |
1177 struct data_dependence_relation *ddr; | |
1178 | |
1179 gcc_assert (DR_IS_WRITE (a) || DR_IS_WRITE (b)); | |
1180 gcc_assert (rdg_vertex_for_stmt (rdg, DR_STMT (a)) | |
1181 <= rdg_vertex_for_stmt (rdg, DR_STMT (b))); | |
1182 ent.a = a; | |
1183 ent.b = b; | |
1184 slot = ddrs_table->find_slot (&ent, INSERT); | |
1185 if (*slot == NULL) | |
1186 { | |
1187 ddr = initialize_data_dependence_relation (a, b, loop_nest); | |
1188 compute_affine_dependence (ddr, loop_nest[0]); | |
1189 *slot = ddr; | |
1190 } | |
1191 | |
1192 return *slot; | |
0 | 1193 } |
1194 | |
111 | 1195 /* In reduced dependence graph RDG for loop distribution, return true if |
1196 dependence between references DR1 and DR2 leads to a dependence cycle | |
1197 and such dependence cycle can't be resolved by runtime alias check. */ | |
0 | 1198 |
1199 static bool | |
111 | 1200 data_dep_in_cycle_p (struct graph *rdg, |
1201 data_reference_p dr1, data_reference_p dr2) | |
0 | 1202 { |
111 | 1203 struct data_dependence_relation *ddr; |
1204 | |
1205 /* Re-shuffle data-refs to be in topological order. */ | |
1206 if (rdg_vertex_for_stmt (rdg, DR_STMT (dr1)) | |
1207 > rdg_vertex_for_stmt (rdg, DR_STMT (dr2))) | |
1208 std::swap (dr1, dr2); | |
1209 | |
1210 ddr = get_data_dependence (rdg, dr1, dr2); | |
1211 | |
1212 /* In case of no data dependence. */ | |
1213 if (DDR_ARE_DEPENDENT (ddr) == chrec_known) | |
1214 return false; | |
1215 /* For unknown data dependence or known data dependence which can't be | |
1216 expressed in classic distance vector, we check if it can be resolved | |
1217 by runtime alias check. If yes, we still consider data dependence | |
1218 as won't introduce data dependence cycle. */ | |
1219 else if (DDR_ARE_DEPENDENT (ddr) == chrec_dont_know | |
1220 || DDR_NUM_DIST_VECTS (ddr) == 0) | |
1221 return !runtime_alias_check_p (ddr, NULL, true); | |
1222 else if (DDR_NUM_DIST_VECTS (ddr) > 1) | |
1223 return true; | |
1224 else if (DDR_REVERSED_P (ddr) | |
1225 || lambda_vector_zerop (DDR_DIST_VECT (ddr, 0), 1)) | |
1226 return false; | |
1227 | |
1228 return true; | |
0 | 1229 } |
1230 | |
111 | 1231 /* Given reduced dependence graph RDG, PARTITION1 and PARTITION2, update |
1232 PARTITION1's type after merging PARTITION2 into PARTITION1. */ | |
0 | 1233 |
1234 static void | |
111 | 1235 update_type_for_merge (struct graph *rdg, |
1236 partition *partition1, partition *partition2) | |
0 | 1237 { |
111 | 1238 unsigned i, j; |
1239 bitmap_iterator bi, bj; | |
1240 data_reference_p dr1, dr2; | |
1241 | |
1242 EXECUTE_IF_SET_IN_BITMAP (partition1->datarefs, 0, i, bi) | |
0 | 1243 { |
111 | 1244 unsigned start = (partition1 == partition2) ? i + 1 : 0; |
1245 | |
1246 dr1 = datarefs_vec[i]; | |
1247 EXECUTE_IF_SET_IN_BITMAP (partition2->datarefs, start, j, bj) | |
0 | 1248 { |
111 | 1249 dr2 = datarefs_vec[j]; |
1250 if (DR_IS_READ (dr1) && DR_IS_READ (dr2)) | |
1251 continue; | |
1252 | |
1253 /* Partition can only be executed sequentially if there is any | |
1254 data dependence cycle. */ | |
1255 if (data_dep_in_cycle_p (rdg, dr1, dr2)) | |
0 | 1256 { |
111 | 1257 partition1->type = PTYPE_SEQUENTIAL; |
1258 return; | |
0 | 1259 } |
1260 } | |
1261 } | |
1262 } | |
1263 | |
111 | 1264 /* Returns a partition with all the statements needed for computing |
1265 the vertex V of the RDG, also including the loop exit conditions. */ | |
1266 | |
1267 static partition * | |
1268 build_rdg_partition_for_vertex (struct graph *rdg, int v) | |
0 | 1269 { |
111 | 1270 partition *partition = partition_alloc (); |
1271 auto_vec<int, 3> nodes; | |
1272 unsigned i, j; | |
0 | 1273 int x; |
111 | 1274 data_reference_p dr; |
1275 | |
1276 graphds_dfs (rdg, &v, 1, &nodes, false, NULL); | |
1277 | |
1278 FOR_EACH_VEC_ELT (nodes, i, x) | |
1279 { | |
1280 bitmap_set_bit (partition->stmts, x); | |
1281 | |
1282 for (j = 0; RDG_DATAREFS (rdg, x).iterate (j, &dr); ++j) | |
1283 { | |
1284 unsigned idx = (unsigned) DR_INDEX (dr); | |
1285 gcc_assert (idx < datarefs_vec.length ()); | |
1286 | |
1287 /* Partition can only be executed sequentially if there is any | |
1288 unknown data reference. */ | |
1289 if (!DR_BASE_ADDRESS (dr) || !DR_OFFSET (dr) | |
1290 || !DR_INIT (dr) || !DR_STEP (dr)) | |
1291 partition->type = PTYPE_SEQUENTIAL; | |
1292 | |
1293 bitmap_set_bit (partition->datarefs, idx); | |
1294 } | |
1295 } | |
1296 | |
1297 if (partition->type == PTYPE_SEQUENTIAL) | |
1298 return partition; | |
1299 | |
1300 /* Further check if any data dependence prevents us from executing the | |
1301 partition parallelly. */ | |
1302 update_type_for_merge (rdg, partition, partition); | |
1303 | |
1304 return partition; | |
0 | 1305 } |
1306 | |
111 | 1307 /* Given PARTITION of LOOP and RDG, record single load/store data references |
1308 for builtin partition in SRC_DR/DST_DR, return false if there is no such | |
1309 data references. */ | |
1310 | |
1311 static bool | |
1312 find_single_drs (struct loop *loop, struct graph *rdg, partition *partition, | |
1313 data_reference_p *dst_dr, data_reference_p *src_dr) | |
0 | 1314 { |
1315 unsigned i; | |
111 | 1316 data_reference_p single_ld = NULL, single_st = NULL; |
0 | 1317 bitmap_iterator bi; |
111 | 1318 |
1319 EXECUTE_IF_SET_IN_BITMAP (partition->stmts, 0, i, bi) | |
0 | 1320 { |
111 | 1321 gimple *stmt = RDG_STMT (rdg, i); |
1322 data_reference_p dr; | |
1323 | |
1324 if (gimple_code (stmt) == GIMPLE_PHI) | |
1325 continue; | |
1326 | |
1327 /* Any scalar stmts are ok. */ | |
1328 if (!gimple_vuse (stmt)) | |
1329 continue; | |
1330 | |
1331 /* Otherwise just regular loads/stores. */ | |
1332 if (!gimple_assign_single_p (stmt)) | |
1333 return false; | |
1334 | |
1335 /* But exactly one store and/or load. */ | |
1336 for (unsigned j = 0; RDG_DATAREFS (rdg, i).iterate (j, &dr); ++j) | |
1337 { | |
1338 tree type = TREE_TYPE (DR_REF (dr)); | |
1339 | |
1340 /* The memset, memcpy and memmove library calls are only | |
1341 able to deal with generic address space. */ | |
1342 if (!ADDR_SPACE_GENERIC_P (TYPE_ADDR_SPACE (type))) | |
1343 return false; | |
1344 | |
1345 if (DR_IS_READ (dr)) | |
1346 { | |
1347 if (single_ld != NULL) | |
1348 return false; | |
1349 single_ld = dr; | |
1350 } | |
1351 else | |
1352 { | |
1353 if (single_st != NULL) | |
1354 return false; | |
1355 single_st = dr; | |
1356 } | |
1357 } | |
0 | 1358 } |
1359 | |
111 | 1360 if (!single_st) |
1361 return false; | |
1362 | |
1363 /* Bail out if this is a bitfield memory reference. */ | |
1364 if (TREE_CODE (DR_REF (single_st)) == COMPONENT_REF | |
1365 && DECL_BIT_FIELD (TREE_OPERAND (DR_REF (single_st), 1))) | |
1366 return false; | |
1367 | |
1368 /* Data reference must be executed exactly once per iteration of each | |
1369 loop in the loop nest. We only need to check dominance information | |
1370 against the outermost one in a perfect loop nest because a bb can't | |
1371 dominate outermost loop's latch without dominating inner loop's. */ | |
1372 basic_block bb_st = gimple_bb (DR_STMT (single_st)); | |
1373 if (!dominated_by_p (CDI_DOMINATORS, loop->latch, bb_st)) | |
1374 return false; | |
1375 | |
1376 if (single_ld) | |
1377 { | |
1378 gimple *store = DR_STMT (single_st), *load = DR_STMT (single_ld); | |
1379 /* Direct aggregate copy or via an SSA name temporary. */ | |
1380 if (load != store | |
1381 && gimple_assign_lhs (load) != gimple_assign_rhs1 (store)) | |
1382 return false; | |
1383 | |
1384 /* Bail out if this is a bitfield memory reference. */ | |
1385 if (TREE_CODE (DR_REF (single_ld)) == COMPONENT_REF | |
1386 && DECL_BIT_FIELD (TREE_OPERAND (DR_REF (single_ld), 1))) | |
1387 return false; | |
1388 | |
1389 /* Load and store must be in the same loop nest. */ | |
1390 basic_block bb_ld = gimple_bb (DR_STMT (single_ld)); | |
1391 if (bb_st->loop_father != bb_ld->loop_father) | |
1392 return false; | |
1393 | |
1394 /* Data reference must be executed exactly once per iteration. | |
1395 Same as single_st, we only need to check against the outermost | |
1396 loop. */ | |
1397 if (!dominated_by_p (CDI_DOMINATORS, loop->latch, bb_ld)) | |
1398 return false; | |
1399 | |
1400 edge e = single_exit (bb_st->loop_father); | |
1401 bool dom_ld = dominated_by_p (CDI_DOMINATORS, e->src, bb_ld); | |
1402 bool dom_st = dominated_by_p (CDI_DOMINATORS, e->src, bb_st); | |
1403 if (dom_ld != dom_st) | |
1404 return false; | |
1405 } | |
1406 | |
1407 *src_dr = single_ld; | |
1408 *dst_dr = single_st; | |
1409 return true; | |
0 | 1410 } |
1411 | |
111 | 1412 /* Given data reference DR in LOOP_NEST, this function checks the enclosing |
1413 loops from inner to outer to see if loop's step equals to access size at | |
131 | 1414 each level of loop. Return 2 if we can prove this at all level loops; |
1415 record access base and size in BASE and SIZE; save loop's step at each | |
1416 level of loop in STEPS if it is not null. For example: | |
111 | 1417 |
1418 int arr[100][100][100]; | |
1419 for (i = 0; i < 100; i++) ;steps[2] = 40000 | |
1420 for (j = 100; j > 0; j--) ;steps[1] = -400 | |
1421 for (k = 0; k < 100; k++) ;steps[0] = 4 | |
131 | 1422 arr[i][j - 1][k] = 0; ;base = &arr, size = 4000000 |
1423 | |
1424 Return 1 if we can prove the equality at the innermost loop, but not all | |
1425 level loops. In this case, no information is recorded. | |
1426 | |
1427 Return 0 if no equality can be proven at any level loops. */ | |
1428 | |
1429 static int | |
111 | 1430 compute_access_range (loop_p loop_nest, data_reference_p dr, tree *base, |
1431 tree *size, vec<tree> *steps = NULL) | |
0 | 1432 { |
111 | 1433 location_t loc = gimple_location (DR_STMT (dr)); |
1434 basic_block bb = gimple_bb (DR_STMT (dr)); | |
1435 struct loop *loop = bb->loop_father; | |
1436 tree ref = DR_REF (dr); | |
1437 tree access_base = build_fold_addr_expr (ref); | |
1438 tree access_size = TYPE_SIZE_UNIT (TREE_TYPE (ref)); | |
131 | 1439 int res = 0; |
111 | 1440 |
1441 do { | |
1442 tree scev_fn = analyze_scalar_evolution (loop, access_base); | |
1443 if (TREE_CODE (scev_fn) != POLYNOMIAL_CHREC) | |
131 | 1444 return res; |
111 | 1445 |
1446 access_base = CHREC_LEFT (scev_fn); | |
1447 if (tree_contains_chrecs (access_base, NULL)) | |
131 | 1448 return res; |
111 | 1449 |
1450 tree scev_step = CHREC_RIGHT (scev_fn); | |
1451 /* Only support constant steps. */ | |
1452 if (TREE_CODE (scev_step) != INTEGER_CST) | |
131 | 1453 return res; |
111 | 1454 |
1455 enum ev_direction access_dir = scev_direction (scev_fn); | |
1456 if (access_dir == EV_DIR_UNKNOWN) | |
131 | 1457 return res; |
111 | 1458 |
1459 if (steps != NULL) | |
1460 steps->safe_push (scev_step); | |
1461 | |
1462 scev_step = fold_convert_loc (loc, sizetype, scev_step); | |
1463 /* Compute absolute value of scev step. */ | |
1464 if (access_dir == EV_DIR_DECREASES) | |
1465 scev_step = fold_build1_loc (loc, NEGATE_EXPR, sizetype, scev_step); | |
1466 | |
1467 /* At each level of loop, scev step must equal to access size. In other | |
1468 words, DR must access consecutive memory between loop iterations. */ | |
1469 if (!operand_equal_p (scev_step, access_size, 0)) | |
131 | 1470 return res; |
1471 | |
1472 /* Access stride can be computed for data reference at least for the | |
1473 innermost loop. */ | |
1474 res = 1; | |
111 | 1475 |
1476 /* Compute DR's execution times in loop. */ | |
1477 tree niters = number_of_latch_executions (loop); | |
1478 niters = fold_convert_loc (loc, sizetype, niters); | |
1479 if (dominated_by_p (CDI_DOMINATORS, single_exit (loop)->src, bb)) | |
1480 niters = size_binop_loc (loc, PLUS_EXPR, niters, size_one_node); | |
1481 | |
1482 /* Compute DR's overall access size in loop. */ | |
1483 access_size = fold_build2_loc (loc, MULT_EXPR, sizetype, | |
1484 niters, scev_step); | |
1485 /* Adjust base address in case of negative step. */ | |
1486 if (access_dir == EV_DIR_DECREASES) | |
1487 { | |
1488 tree adj = fold_build2_loc (loc, MINUS_EXPR, sizetype, | |
1489 scev_step, access_size); | |
1490 access_base = fold_build_pointer_plus_loc (loc, access_base, adj); | |
1491 } | |
1492 } while (loop != loop_nest && (loop = loop_outer (loop)) != NULL); | |
1493 | |
1494 *base = access_base; | |
1495 *size = access_size; | |
131 | 1496 /* Access stride can be computed for data reference at each level loop. */ |
1497 return 2; | |
0 | 1498 } |
1499 | |
111 | 1500 /* Allocate and return builtin struct. Record information like DST_DR, |
1501 SRC_DR, DST_BASE, SRC_BASE and SIZE in the allocated struct. */ | |
1502 | |
1503 static struct builtin_info * | |
1504 alloc_builtin (data_reference_p dst_dr, data_reference_p src_dr, | |
1505 tree dst_base, tree src_base, tree size) | |
1506 { | |
1507 struct builtin_info *builtin = XNEW (struct builtin_info); | |
1508 builtin->dst_dr = dst_dr; | |
1509 builtin->src_dr = src_dr; | |
1510 builtin->dst_base = dst_base; | |
1511 builtin->src_base = src_base; | |
1512 builtin->size = size; | |
1513 return builtin; | |
1514 } | |
1515 | |
1516 /* Given data reference DR in loop nest LOOP, classify if it forms builtin | |
1517 memset call. */ | |
0 | 1518 |
1519 static void | |
111 | 1520 classify_builtin_st (loop_p loop, partition *partition, data_reference_p dr) |
0 | 1521 { |
111 | 1522 gimple *stmt = DR_STMT (dr); |
1523 tree base, size, rhs = gimple_assign_rhs1 (stmt); | |
1524 | |
1525 if (const_with_all_bytes_same (rhs) == -1 | |
1526 && (!INTEGRAL_TYPE_P (TREE_TYPE (rhs)) | |
1527 || (TYPE_MODE (TREE_TYPE (rhs)) | |
1528 != TYPE_MODE (unsigned_char_type_node)))) | |
1529 return; | |
1530 | |
1531 if (TREE_CODE (rhs) == SSA_NAME | |
1532 && !SSA_NAME_IS_DEFAULT_DEF (rhs) | |
1533 && flow_bb_inside_loop_p (loop, gimple_bb (SSA_NAME_DEF_STMT (rhs)))) | |
1534 return; | |
1535 | |
131 | 1536 int res = compute_access_range (loop, dr, &base, &size); |
1537 if (res == 0) | |
1538 return; | |
1539 if (res == 1) | |
1540 { | |
1541 partition->kind = PKIND_PARTIAL_MEMSET; | |
1542 return; | |
1543 } | |
1544 | |
1545 poly_uint64 base_offset; | |
1546 unsigned HOST_WIDE_INT const_base_offset; | |
1547 tree base_base = strip_offset (base, &base_offset); | |
1548 if (!base_offset.is_constant (&const_base_offset)) | |
111 | 1549 return; |
1550 | |
1551 struct builtin_info *builtin; | |
1552 builtin = alloc_builtin (dr, NULL, base, NULL_TREE, size); | |
131 | 1553 builtin->dst_base_base = base_base; |
1554 builtin->dst_base_offset = const_base_offset; | |
111 | 1555 partition->builtin = builtin; |
1556 partition->kind = PKIND_MEMSET; | |
0 | 1557 } |
1558 | |
111 | 1559 /* Given data references DST_DR and SRC_DR in loop nest LOOP and RDG, classify |
1560 if it forms builtin memcpy or memmove call. */ | |
0 | 1561 |
1562 static void | |
111 | 1563 classify_builtin_ldst (loop_p loop, struct graph *rdg, partition *partition, |
1564 data_reference_p dst_dr, data_reference_p src_dr) | |
0 | 1565 { |
111 | 1566 tree base, size, src_base, src_size; |
1567 auto_vec<tree> dst_steps, src_steps; | |
1568 | |
131 | 1569 /* Compute access range of both load and store. */ |
1570 int res = compute_access_range (loop, dst_dr, &base, &size, &dst_steps); | |
1571 if (res != 2) | |
1572 return; | |
1573 res = compute_access_range (loop, src_dr, &src_base, &src_size, &src_steps); | |
1574 if (res != 2) | |
1575 return; | |
1576 | |
1577 /* They much have the same access size. */ | |
1578 if (!operand_equal_p (size, src_size, 0)) | |
111 | 1579 return; |
1580 | |
1581 /* Load and store in loop nest must access memory in the same way, i.e, | |
1582 their must have the same steps in each loop of the nest. */ | |
1583 if (dst_steps.length () != src_steps.length ()) | |
1584 return; | |
1585 for (unsigned i = 0; i < dst_steps.length (); ++i) | |
1586 if (!operand_equal_p (dst_steps[i], src_steps[i], 0)) | |
1587 return; | |
1588 | |
1589 /* Now check that if there is a dependence. */ | |
1590 ddr_p ddr = get_data_dependence (rdg, src_dr, dst_dr); | |
1591 | |
1592 /* Classify as memcpy if no dependence between load and store. */ | |
1593 if (DDR_ARE_DEPENDENT (ddr) == chrec_known) | |
1594 { | |
1595 partition->builtin = alloc_builtin (dst_dr, src_dr, base, src_base, size); | |
1596 partition->kind = PKIND_MEMCPY; | |
1597 return; | |
1598 } | |
1599 | |
1600 /* Can't do memmove in case of unknown dependence or dependence without | |
1601 classical distance vector. */ | |
1602 if (DDR_ARE_DEPENDENT (ddr) == chrec_dont_know | |
1603 || DDR_NUM_DIST_VECTS (ddr) == 0) | |
1604 return; | |
1605 | |
1606 unsigned i; | |
1607 lambda_vector dist_v; | |
1608 int num_lev = (DDR_LOOP_NEST (ddr)).length (); | |
1609 FOR_EACH_VEC_ELT (DDR_DIST_VECTS (ddr), i, dist_v) | |
0 | 1610 { |
111 | 1611 unsigned dep_lev = dependence_level (dist_v, num_lev); |
1612 /* Can't do memmove if load depends on store. */ | |
1613 if (dep_lev > 0 && dist_v[dep_lev - 1] > 0 && !DDR_REVERSED_P (ddr)) | |
1614 return; | |
1615 } | |
1616 | |
1617 partition->builtin = alloc_builtin (dst_dr, src_dr, base, src_base, size); | |
1618 partition->kind = PKIND_MEMMOVE; | |
1619 return; | |
1620 } | |
1621 | |
1622 /* Classifies the builtin kind we can generate for PARTITION of RDG and LOOP. | |
1623 For the moment we detect memset, memcpy and memmove patterns. Bitmap | |
1624 STMT_IN_ALL_PARTITIONS contains statements belonging to all partitions. */ | |
1625 | |
1626 static void | |
1627 classify_partition (loop_p loop, struct graph *rdg, partition *partition, | |
1628 bitmap stmt_in_all_partitions) | |
1629 { | |
1630 bitmap_iterator bi; | |
1631 unsigned i; | |
1632 data_reference_p single_ld = NULL, single_st = NULL; | |
1633 bool volatiles_p = false, has_reduction = false; | |
1634 | |
1635 EXECUTE_IF_SET_IN_BITMAP (partition->stmts, 0, i, bi) | |
1636 { | |
1637 gimple *stmt = RDG_STMT (rdg, i); | |
1638 | |
1639 if (gimple_has_volatile_ops (stmt)) | |
1640 volatiles_p = true; | |
1641 | |
1642 /* If the stmt is not included by all partitions and there is uses | |
1643 outside of the loop, then mark the partition as reduction. */ | |
1644 if (stmt_has_scalar_dependences_outside_loop (loop, stmt)) | |
0 | 1645 { |
111 | 1646 /* Due to limitation in the transform phase we have to fuse all |
1647 reduction partitions. As a result, this could cancel valid | |
1648 loop distribution especially for loop that induction variable | |
1649 is used outside of loop. To workaround this issue, we skip | |
1650 marking partition as reudction if the reduction stmt belongs | |
1651 to all partitions. In such case, reduction will be computed | |
1652 correctly no matter how partitions are fused/distributed. */ | |
1653 if (!bitmap_bit_p (stmt_in_all_partitions, i)) | |
1654 { | |
1655 partition->reduction_p = true; | |
1656 return; | |
1657 } | |
1658 has_reduction = true; | |
0 | 1659 } |
1660 } | |
1661 | |
111 | 1662 /* Perform general partition disqualification for builtins. */ |
1663 if (volatiles_p | |
1664 /* Simple workaround to prevent classifying the partition as builtin | |
1665 if it contains any use outside of loop. */ | |
1666 || has_reduction | |
1667 || !flag_tree_loop_distribute_patterns) | |
1668 return; | |
1669 | |
1670 /* Find single load/store data references for builtin partition. */ | |
1671 if (!find_single_drs (loop, rdg, partition, &single_st, &single_ld)) | |
1672 return; | |
1673 | |
1674 /* Classify the builtin kind. */ | |
1675 if (single_ld == NULL) | |
1676 classify_builtin_st (loop, partition, single_st); | |
1677 else | |
1678 classify_builtin_ldst (loop, rdg, partition, single_st, single_ld); | |
0 | 1679 } |
1680 | |
111 | 1681 /* Returns true when PARTITION1 and PARTITION2 access the same memory |
1682 object in RDG. */ | |
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1683 |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1684 static bool |
111 | 1685 share_memory_accesses (struct graph *rdg, |
1686 partition *partition1, partition *partition2) | |
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1687 { |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1688 unsigned i, j; |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1689 bitmap_iterator bi, bj; |
111 | 1690 data_reference_p dr1, dr2; |
1691 | |
1692 /* First check whether in the intersection of the two partitions are | |
1693 any loads or stores. Common loads are the situation that happens | |
1694 most often. */ | |
1695 EXECUTE_IF_AND_IN_BITMAP (partition1->stmts, partition2->stmts, 0, i, bi) | |
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1696 if (RDG_MEM_WRITE_STMT (rdg, i) |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1697 || RDG_MEM_READS_STMT (rdg, i)) |
111 | 1698 return true; |
1699 | |
1700 /* Then check whether the two partitions access the same memory object. */ | |
1701 EXECUTE_IF_SET_IN_BITMAP (partition1->datarefs, 0, i, bi) | |
1702 { | |
1703 dr1 = datarefs_vec[i]; | |
1704 | |
1705 if (!DR_BASE_ADDRESS (dr1) | |
1706 || !DR_OFFSET (dr1) || !DR_INIT (dr1) || !DR_STEP (dr1)) | |
1707 continue; | |
1708 | |
1709 EXECUTE_IF_SET_IN_BITMAP (partition2->datarefs, 0, j, bj) | |
1710 { | |
1711 dr2 = datarefs_vec[j]; | |
1712 | |
1713 if (!DR_BASE_ADDRESS (dr2) | |
1714 || !DR_OFFSET (dr2) || !DR_INIT (dr2) || !DR_STEP (dr2)) | |
1715 continue; | |
1716 | |
1717 if (operand_equal_p (DR_BASE_ADDRESS (dr1), DR_BASE_ADDRESS (dr2), 0) | |
1718 && operand_equal_p (DR_OFFSET (dr1), DR_OFFSET (dr2), 0) | |
1719 && operand_equal_p (DR_INIT (dr1), DR_INIT (dr2), 0) | |
1720 && operand_equal_p (DR_STEP (dr1), DR_STEP (dr2), 0)) | |
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1721 return true; |
111 | 1722 } |
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1723 } |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1724 |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1725 return false; |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1726 } |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1727 |
111 | 1728 /* For each seed statement in STARTING_STMTS, this function builds |
1729 partition for it by adding depended statements according to RDG. | |
1730 All partitions are recorded in PARTITIONS. */ | |
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1731 |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1732 static void |
111 | 1733 rdg_build_partitions (struct graph *rdg, |
1734 vec<gimple *> starting_stmts, | |
1735 vec<partition *> *partitions) | |
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1736 { |
111 | 1737 auto_bitmap processed; |
0 | 1738 int i; |
111 | 1739 gimple *stmt; |
1740 | |
1741 FOR_EACH_VEC_ELT (starting_stmts, i, stmt) | |
0 | 1742 { |
111 | 1743 int v = rdg_vertex_for_stmt (rdg, stmt); |
1744 | |
1745 if (dump_file && (dump_flags & TDF_DETAILS)) | |
1746 fprintf (dump_file, | |
1747 "ldist asked to generate code for vertex %d\n", v); | |
1748 | |
1749 /* If the vertex is already contained in another partition so | |
1750 is the partition rooted at it. */ | |
0 | 1751 if (bitmap_bit_p (processed, v)) |
1752 continue; | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1753 |
111 | 1754 partition *partition = build_rdg_partition_for_vertex (rdg, v); |
1755 bitmap_ior_into (processed, partition->stmts); | |
1756 | |
1757 if (dump_file && (dump_flags & TDF_DETAILS)) | |
0 | 1758 { |
111 | 1759 fprintf (dump_file, "ldist creates useful %s partition:\n", |
1760 partition->type == PTYPE_PARALLEL ? "parallel" : "sequent"); | |
1761 bitmap_print (dump_file, partition->stmts, " ", "\n"); | |
0 | 1762 } |
111 | 1763 |
1764 partitions->safe_push (partition); | |
0 | 1765 } |
1766 | |
111 | 1767 /* All vertices should have been assigned to at least one partition now, |
1768 other than vertices belonging to dead code. */ | |
0 | 1769 } |
1770 | |
1771 /* Dump to FILE the PARTITIONS. */ | |
1772 | |
1773 static void | |
111 | 1774 dump_rdg_partitions (FILE *file, vec<partition *> partitions) |
0 | 1775 { |
1776 int i; | |
111 | 1777 partition *partition; |
1778 | |
1779 FOR_EACH_VEC_ELT (partitions, i, partition) | |
1780 debug_bitmap_file (file, partition->stmts); | |
0 | 1781 } |
1782 | |
1783 /* Debug PARTITIONS. */ | |
111 | 1784 extern void debug_rdg_partitions (vec<partition *> ); |
0 | 1785 |
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1786 DEBUG_FUNCTION void |
111 | 1787 debug_rdg_partitions (vec<partition *> partitions) |
0 | 1788 { |
1789 dump_rdg_partitions (stderr, partitions); | |
1790 } | |
1791 | |
1792 /* Returns the number of read and write operations in the RDG. */ | |
1793 | |
1794 static int | |
1795 number_of_rw_in_rdg (struct graph *rdg) | |
1796 { | |
1797 int i, res = 0; | |
1798 | |
1799 for (i = 0; i < rdg->n_vertices; i++) | |
1800 { | |
1801 if (RDG_MEM_WRITE_STMT (rdg, i)) | |
1802 ++res; | |
1803 | |
1804 if (RDG_MEM_READS_STMT (rdg, i)) | |
1805 ++res; | |
1806 } | |
1807 | |
1808 return res; | |
1809 } | |
1810 | |
1811 /* Returns the number of read and write operations in a PARTITION of | |
1812 the RDG. */ | |
1813 | |
1814 static int | |
111 | 1815 number_of_rw_in_partition (struct graph *rdg, partition *partition) |
0 | 1816 { |
1817 int res = 0; | |
1818 unsigned i; | |
1819 bitmap_iterator ii; | |
1820 | |
111 | 1821 EXECUTE_IF_SET_IN_BITMAP (partition->stmts, 0, i, ii) |
0 | 1822 { |
1823 if (RDG_MEM_WRITE_STMT (rdg, i)) | |
1824 ++res; | |
1825 | |
1826 if (RDG_MEM_READS_STMT (rdg, i)) | |
1827 ++res; | |
1828 } | |
1829 | |
1830 return res; | |
1831 } | |
1832 | |
1833 /* Returns true when one of the PARTITIONS contains all the read or | |
1834 write operations of RDG. */ | |
1835 | |
1836 static bool | |
111 | 1837 partition_contains_all_rw (struct graph *rdg, |
1838 vec<partition *> partitions) | |
0 | 1839 { |
1840 int i; | |
111 | 1841 partition *partition; |
0 | 1842 int nrw = number_of_rw_in_rdg (rdg); |
1843 | |
111 | 1844 FOR_EACH_VEC_ELT (partitions, i, partition) |
0 | 1845 if (nrw == number_of_rw_in_partition (rdg, partition)) |
1846 return true; | |
1847 | |
1848 return false; | |
1849 } | |
1850 | |
111 | 1851 /* Compute partition dependence created by the data references in DRS1 |
1852 and DRS2, modify and return DIR according to that. IF ALIAS_DDR is | |
1853 not NULL, we record dependence introduced by possible alias between | |
1854 two data references in ALIAS_DDRS; otherwise, we simply ignore such | |
1855 dependence as if it doesn't exist at all. */ | |
1856 | |
1857 static int | |
1858 pg_add_dependence_edges (struct graph *rdg, int dir, | |
1859 bitmap drs1, bitmap drs2, vec<ddr_p> *alias_ddrs) | |
1860 { | |
1861 unsigned i, j; | |
1862 bitmap_iterator bi, bj; | |
1863 data_reference_p dr1, dr2, saved_dr1; | |
1864 | |
1865 /* dependence direction - 0 is no dependence, -1 is back, | |
1866 1 is forth, 2 is both (we can stop then, merging will occur). */ | |
1867 EXECUTE_IF_SET_IN_BITMAP (drs1, 0, i, bi) | |
1868 { | |
1869 dr1 = datarefs_vec[i]; | |
1870 | |
1871 EXECUTE_IF_SET_IN_BITMAP (drs2, 0, j, bj) | |
1872 { | |
1873 int res, this_dir = 1; | |
1874 ddr_p ddr; | |
1875 | |
1876 dr2 = datarefs_vec[j]; | |
1877 | |
1878 /* Skip all <read, read> data dependence. */ | |
1879 if (DR_IS_READ (dr1) && DR_IS_READ (dr2)) | |
1880 continue; | |
1881 | |
1882 saved_dr1 = dr1; | |
1883 /* Re-shuffle data-refs to be in topological order. */ | |
1884 if (rdg_vertex_for_stmt (rdg, DR_STMT (dr1)) | |
1885 > rdg_vertex_for_stmt (rdg, DR_STMT (dr2))) | |
1886 { | |
1887 std::swap (dr1, dr2); | |
1888 this_dir = -this_dir; | |
1889 } | |
1890 ddr = get_data_dependence (rdg, dr1, dr2); | |
1891 if (DDR_ARE_DEPENDENT (ddr) == chrec_dont_know) | |
1892 { | |
1893 this_dir = 0; | |
1894 res = data_ref_compare_tree (DR_BASE_ADDRESS (dr1), | |
1895 DR_BASE_ADDRESS (dr2)); | |
1896 /* Be conservative. If data references are not well analyzed, | |
1897 or the two data references have the same base address and | |
1898 offset, add dependence and consider it alias to each other. | |
1899 In other words, the dependence can not be resolved by | |
1900 runtime alias check. */ | |
1901 if (!DR_BASE_ADDRESS (dr1) || !DR_BASE_ADDRESS (dr2) | |
1902 || !DR_OFFSET (dr1) || !DR_OFFSET (dr2) | |
1903 || !DR_INIT (dr1) || !DR_INIT (dr2) | |
1904 || !DR_STEP (dr1) || !tree_fits_uhwi_p (DR_STEP (dr1)) | |
1905 || !DR_STEP (dr2) || !tree_fits_uhwi_p (DR_STEP (dr2)) | |
1906 || res == 0) | |
1907 this_dir = 2; | |
1908 /* Data dependence could be resolved by runtime alias check, | |
1909 record it in ALIAS_DDRS. */ | |
1910 else if (alias_ddrs != NULL) | |
1911 alias_ddrs->safe_push (ddr); | |
1912 /* Or simply ignore it. */ | |
1913 } | |
1914 else if (DDR_ARE_DEPENDENT (ddr) == NULL_TREE) | |
1915 { | |
1916 if (DDR_REVERSED_P (ddr)) | |
1917 this_dir = -this_dir; | |
1918 | |
1919 /* Known dependences can still be unordered througout the | |
1920 iteration space, see gcc.dg/tree-ssa/ldist-16.c. */ | |
1921 if (DDR_NUM_DIST_VECTS (ddr) != 1) | |
1922 this_dir = 2; | |
1923 /* If the overlap is exact preserve stmt order. */ | |
131 | 1924 else if (lambda_vector_zerop (DDR_DIST_VECT (ddr, 0), |
1925 DDR_NB_LOOPS (ddr))) | |
111 | 1926 ; |
1927 /* Else as the distance vector is lexicographic positive swap | |
1928 the dependence direction. */ | |
1929 else | |
1930 this_dir = -this_dir; | |
1931 } | |
1932 else | |
1933 this_dir = 0; | |
1934 if (this_dir == 2) | |
1935 return 2; | |
1936 else if (dir == 0) | |
1937 dir = this_dir; | |
1938 else if (this_dir != 0 && dir != this_dir) | |
1939 return 2; | |
1940 /* Shuffle "back" dr1. */ | |
1941 dr1 = saved_dr1; | |
1942 } | |
1943 } | |
1944 return dir; | |
1945 } | |
1946 | |
1947 /* Compare postorder number of the partition graph vertices V1 and V2. */ | |
0 | 1948 |
1949 static int | |
111 | 1950 pgcmp (const void *v1_, const void *v2_) |
1951 { | |
1952 const vertex *v1 = (const vertex *)v1_; | |
1953 const vertex *v2 = (const vertex *)v2_; | |
1954 return v2->post - v1->post; | |
1955 } | |
1956 | |
1957 /* Data attached to vertices of partition dependence graph. */ | |
1958 struct pg_vdata | |
1959 { | |
1960 /* ID of the corresponding partition. */ | |
1961 int id; | |
1962 /* The partition. */ | |
1963 struct partition *partition; | |
1964 }; | |
1965 | |
1966 /* Data attached to edges of partition dependence graph. */ | |
1967 struct pg_edata | |
1968 { | |
1969 /* If the dependence edge can be resolved by runtime alias check, | |
1970 this vector contains data dependence relations for runtime alias | |
1971 check. On the other hand, if the dependence edge is introduced | |
1972 because of compilation time known data dependence, this vector | |
1973 contains nothing. */ | |
1974 vec<ddr_p> alias_ddrs; | |
1975 }; | |
1976 | |
1977 /* Callback data for traversing edges in graph. */ | |
1978 struct pg_edge_callback_data | |
0 | 1979 { |
111 | 1980 /* Bitmap contains strong connected components should be merged. */ |
1981 bitmap sccs_to_merge; | |
1982 /* Array constains component information for all vertices. */ | |
1983 int *vertices_component; | |
1984 /* Vector to record all data dependence relations which are needed | |
1985 to break strong connected components by runtime alias checks. */ | |
1986 vec<ddr_p> *alias_ddrs; | |
1987 }; | |
1988 | |
1989 /* Initialize vertice's data for partition dependence graph PG with | |
1990 PARTITIONS. */ | |
1991 | |
1992 static void | |
1993 init_partition_graph_vertices (struct graph *pg, | |
1994 vec<struct partition *> *partitions) | |
1995 { | |
1996 int i; | |
1997 partition *partition; | |
1998 struct pg_vdata *data; | |
1999 | |
2000 for (i = 0; partitions->iterate (i, &partition); ++i) | |
2001 { | |
2002 data = new pg_vdata; | |
2003 pg->vertices[i].data = data; | |
2004 data->id = i; | |
2005 data->partition = partition; | |
2006 } | |
2007 } | |
2008 | |
2009 /* Add edge <I, J> to partition dependence graph PG. Attach vector of data | |
2010 dependence relations to the EDGE if DDRS isn't NULL. */ | |
2011 | |
2012 static void | |
2013 add_partition_graph_edge (struct graph *pg, int i, int j, vec<ddr_p> *ddrs) | |
2014 { | |
2015 struct graph_edge *e = add_edge (pg, i, j); | |
2016 | |
2017 /* If the edge is attached with data dependence relations, it means this | |
2018 dependence edge can be resolved by runtime alias checks. */ | |
2019 if (ddrs != NULL) | |
0 | 2020 { |
111 | 2021 struct pg_edata *data = new pg_edata; |
2022 | |
2023 gcc_assert (ddrs->length () > 0); | |
2024 e->data = data; | |
2025 data->alias_ddrs = vNULL; | |
2026 data->alias_ddrs.safe_splice (*ddrs); | |
2027 } | |
2028 } | |
2029 | |
2030 /* Callback function for graph travesal algorithm. It returns true | |
2031 if edge E should skipped when traversing the graph. */ | |
2032 | |
2033 static bool | |
2034 pg_skip_alias_edge (struct graph_edge *e) | |
2035 { | |
2036 struct pg_edata *data = (struct pg_edata *)e->data; | |
2037 return (data != NULL && data->alias_ddrs.length () > 0); | |
2038 } | |
2039 | |
2040 /* Callback function freeing data attached to edge E of graph. */ | |
2041 | |
2042 static void | |
2043 free_partition_graph_edata_cb (struct graph *, struct graph_edge *e, void *) | |
2044 { | |
2045 if (e->data != NULL) | |
2046 { | |
2047 struct pg_edata *data = (struct pg_edata *)e->data; | |
2048 data->alias_ddrs.release (); | |
2049 delete data; | |
2050 } | |
2051 } | |
2052 | |
2053 /* Free data attached to vertice of partition dependence graph PG. */ | |
2054 | |
2055 static void | |
2056 free_partition_graph_vdata (struct graph *pg) | |
2057 { | |
2058 int i; | |
2059 struct pg_vdata *data; | |
2060 | |
2061 for (i = 0; i < pg->n_vertices; ++i) | |
2062 { | |
2063 data = (struct pg_vdata *)pg->vertices[i].data; | |
2064 delete data; | |
2065 } | |
2066 } | |
2067 | |
2068 /* Build and return partition dependence graph for PARTITIONS. RDG is | |
2069 reduced dependence graph for the loop to be distributed. If IGNORE_ALIAS_P | |
2070 is true, data dependence caused by possible alias between references | |
2071 is ignored, as if it doesn't exist at all; otherwise all depdendences | |
2072 are considered. */ | |
2073 | |
2074 static struct graph * | |
2075 build_partition_graph (struct graph *rdg, | |
2076 vec<struct partition *> *partitions, | |
2077 bool ignore_alias_p) | |
2078 { | |
2079 int i, j; | |
2080 struct partition *partition1, *partition2; | |
2081 graph *pg = new_graph (partitions->length ()); | |
2082 auto_vec<ddr_p> alias_ddrs, *alias_ddrs_p; | |
2083 | |
2084 alias_ddrs_p = ignore_alias_p ? NULL : &alias_ddrs; | |
2085 | |
2086 init_partition_graph_vertices (pg, partitions); | |
2087 | |
2088 for (i = 0; partitions->iterate (i, &partition1); ++i) | |
2089 { | |
2090 for (j = i + 1; partitions->iterate (j, &partition2); ++j) | |
0 | 2091 { |
111 | 2092 /* dependence direction - 0 is no dependence, -1 is back, |
2093 1 is forth, 2 is both (we can stop then, merging will occur). */ | |
2094 int dir = 0; | |
2095 | |
2096 /* If the first partition has reduction, add back edge; if the | |
2097 second partition has reduction, add forth edge. This makes | |
2098 sure that reduction partition will be sorted as the last one. */ | |
2099 if (partition_reduction_p (partition1)) | |
2100 dir = -1; | |
2101 else if (partition_reduction_p (partition2)) | |
2102 dir = 1; | |
2103 | |
2104 /* Cleanup the temporary vector. */ | |
2105 alias_ddrs.truncate (0); | |
2106 | |
2107 dir = pg_add_dependence_edges (rdg, dir, partition1->datarefs, | |
2108 partition2->datarefs, alias_ddrs_p); | |
2109 | |
2110 /* Add edge to partition graph if there exists dependence. There | |
2111 are two types of edges. One type edge is caused by compilation | |
2112 time known dependence, this type can not be resolved by runtime | |
2113 alias check. The other type can be resolved by runtime alias | |
2114 check. */ | |
2115 if (dir == 1 || dir == 2 | |
2116 || alias_ddrs.length () > 0) | |
2117 { | |
2118 /* Attach data dependence relations to edge that can be resolved | |
2119 by runtime alias check. */ | |
2120 bool alias_edge_p = (dir != 1 && dir != 2); | |
2121 add_partition_graph_edge (pg, i, j, | |
2122 (alias_edge_p) ? &alias_ddrs : NULL); | |
2123 } | |
2124 if (dir == -1 || dir == 2 | |
2125 || alias_ddrs.length () > 0) | |
2126 { | |
2127 /* Attach data dependence relations to edge that can be resolved | |
2128 by runtime alias check. */ | |
2129 bool alias_edge_p = (dir != -1 && dir != 2); | |
2130 add_partition_graph_edge (pg, j, i, | |
2131 (alias_edge_p) ? &alias_ddrs : NULL); | |
2132 } | |
0 | 2133 } |
2134 } | |
111 | 2135 return pg; |
2136 } | |
2137 | |
2138 /* Sort partitions in PG in descending post order and store them in | |
2139 PARTITIONS. */ | |
2140 | |
2141 static void | |
2142 sort_partitions_by_post_order (struct graph *pg, | |
2143 vec<struct partition *> *partitions) | |
2144 { | |
2145 int i; | |
2146 struct pg_vdata *data; | |
2147 | |
2148 /* Now order the remaining nodes in descending postorder. */ | |
2149 qsort (pg->vertices, pg->n_vertices, sizeof (vertex), pgcmp); | |
2150 partitions->truncate (0); | |
2151 for (i = 0; i < pg->n_vertices; ++i) | |
2152 { | |
2153 data = (struct pg_vdata *)pg->vertices[i].data; | |
2154 if (data->partition) | |
2155 partitions->safe_push (data->partition); | |
2156 } | |
2157 } | |
2158 | |
2159 /* Given reduced dependence graph RDG merge strong connected components | |
2160 of PARTITIONS. If IGNORE_ALIAS_P is true, data dependence caused by | |
2161 possible alias between references is ignored, as if it doesn't exist | |
2162 at all; otherwise all depdendences are considered. */ | |
2163 | |
2164 static void | |
2165 merge_dep_scc_partitions (struct graph *rdg, | |
2166 vec<struct partition *> *partitions, | |
2167 bool ignore_alias_p) | |
2168 { | |
2169 struct partition *partition1, *partition2; | |
2170 struct pg_vdata *data; | |
2171 graph *pg = build_partition_graph (rdg, partitions, ignore_alias_p); | |
2172 int i, j, num_sccs = graphds_scc (pg, NULL); | |
2173 | |
2174 /* Strong connected compoenent means dependence cycle, we cannot distribute | |
2175 them. So fuse them together. */ | |
2176 if ((unsigned) num_sccs < partitions->length ()) | |
2177 { | |
2178 for (i = 0; i < num_sccs; ++i) | |
2179 { | |
2180 for (j = 0; partitions->iterate (j, &partition1); ++j) | |
2181 if (pg->vertices[j].component == i) | |
2182 break; | |
2183 for (j = j + 1; partitions->iterate (j, &partition2); ++j) | |
2184 if (pg->vertices[j].component == i) | |
2185 { | |
2186 partition_merge_into (NULL, partition1, | |
2187 partition2, FUSE_SAME_SCC); | |
2188 partition1->type = PTYPE_SEQUENTIAL; | |
2189 (*partitions)[j] = NULL; | |
2190 partition_free (partition2); | |
2191 data = (struct pg_vdata *)pg->vertices[j].data; | |
2192 data->partition = NULL; | |
2193 } | |
2194 } | |
2195 } | |
2196 | |
2197 sort_partitions_by_post_order (pg, partitions); | |
2198 gcc_assert (partitions->length () == (unsigned)num_sccs); | |
2199 free_partition_graph_vdata (pg); | |
2200 free_graph (pg); | |
2201 } | |
2202 | |
2203 /* Callback function for traversing edge E in graph G. DATA is private | |
2204 callback data. */ | |
2205 | |
2206 static void | |
2207 pg_collect_alias_ddrs (struct graph *g, struct graph_edge *e, void *data) | |
2208 { | |
2209 int i, j, component; | |
2210 struct pg_edge_callback_data *cbdata; | |
2211 struct pg_edata *edata = (struct pg_edata *) e->data; | |
2212 | |
2213 /* If the edge doesn't have attached data dependence, it represents | |
2214 compilation time known dependences. This type dependence cannot | |
2215 be resolved by runtime alias check. */ | |
2216 if (edata == NULL || edata->alias_ddrs.length () == 0) | |
2217 return; | |
2218 | |
2219 cbdata = (struct pg_edge_callback_data *) data; | |
2220 i = e->src; | |
2221 j = e->dest; | |
2222 component = cbdata->vertices_component[i]; | |
2223 /* Vertices are topologically sorted according to compilation time | |
2224 known dependences, so we can break strong connected components | |
2225 by removing edges of the opposite direction, i.e, edges pointing | |
2226 from vertice with smaller post number to vertice with bigger post | |
2227 number. */ | |
2228 if (g->vertices[i].post < g->vertices[j].post | |
2229 /* We only need to remove edges connecting vertices in the same | |
2230 strong connected component to break it. */ | |
2231 && component == cbdata->vertices_component[j] | |
2232 /* Check if we want to break the strong connected component or not. */ | |
2233 && !bitmap_bit_p (cbdata->sccs_to_merge, component)) | |
2234 cbdata->alias_ddrs->safe_splice (edata->alias_ddrs); | |
2235 } | |
2236 | |
2237 /* This is the main function breaking strong conected components in | |
2238 PARTITIONS giving reduced depdendence graph RDG. Store data dependence | |
2239 relations for runtime alias check in ALIAS_DDRS. */ | |
2240 | |
2241 static void | |
2242 break_alias_scc_partitions (struct graph *rdg, | |
2243 vec<struct partition *> *partitions, | |
2244 vec<ddr_p> *alias_ddrs) | |
2245 { | |
2246 int i, j, k, num_sccs, num_sccs_no_alias; | |
2247 /* Build partition dependence graph. */ | |
2248 graph *pg = build_partition_graph (rdg, partitions, false); | |
2249 | |
2250 alias_ddrs->truncate (0); | |
2251 /* Find strong connected components in the graph, with all dependence edges | |
2252 considered. */ | |
2253 num_sccs = graphds_scc (pg, NULL); | |
2254 /* All SCCs now can be broken by runtime alias checks because SCCs caused by | |
2255 compilation time known dependences are merged before this function. */ | |
2256 if ((unsigned) num_sccs < partitions->length ()) | |
2257 { | |
2258 struct pg_edge_callback_data cbdata; | |
2259 auto_bitmap sccs_to_merge; | |
2260 auto_vec<enum partition_type> scc_types; | |
2261 struct partition *partition, *first; | |
2262 | |
2263 /* If all partitions in a SCC have the same type, we can simply merge the | |
2264 SCC. This loop finds out such SCCS and record them in bitmap. */ | |
2265 bitmap_set_range (sccs_to_merge, 0, (unsigned) num_sccs); | |
2266 for (i = 0; i < num_sccs; ++i) | |
2267 { | |
2268 for (j = 0; partitions->iterate (j, &first); ++j) | |
2269 if (pg->vertices[j].component == i) | |
2270 break; | |
131 | 2271 |
2272 bool same_type = true, all_builtins = partition_builtin_p (first); | |
111 | 2273 for (++j; partitions->iterate (j, &partition); ++j) |
2274 { | |
2275 if (pg->vertices[j].component != i) | |
2276 continue; | |
2277 | |
2278 if (first->type != partition->type) | |
2279 { | |
131 | 2280 same_type = false; |
111 | 2281 break; |
2282 } | |
131 | 2283 all_builtins &= partition_builtin_p (partition); |
111 | 2284 } |
131 | 2285 /* Merge SCC if all partitions in SCC have the same type, though the |
2286 result partition is sequential, because vectorizer can do better | |
2287 runtime alias check. One expecption is all partitions in SCC are | |
2288 builtins. */ | |
2289 if (!same_type || all_builtins) | |
2290 bitmap_clear_bit (sccs_to_merge, i); | |
111 | 2291 } |
2292 | |
2293 /* Initialize callback data for traversing. */ | |
2294 cbdata.sccs_to_merge = sccs_to_merge; | |
2295 cbdata.alias_ddrs = alias_ddrs; | |
2296 cbdata.vertices_component = XNEWVEC (int, pg->n_vertices); | |
2297 /* Record the component information which will be corrupted by next | |
2298 graph scc finding call. */ | |
2299 for (i = 0; i < pg->n_vertices; ++i) | |
2300 cbdata.vertices_component[i] = pg->vertices[i].component; | |
2301 | |
2302 /* Collect data dependences for runtime alias checks to break SCCs. */ | |
2303 if (bitmap_count_bits (sccs_to_merge) != (unsigned) num_sccs) | |
2304 { | |
2305 /* Run SCC finding algorithm again, with alias dependence edges | |
2306 skipped. This is to topologically sort partitions according to | |
2307 compilation time known dependence. Note the topological order | |
2308 is stored in the form of pg's post order number. */ | |
2309 num_sccs_no_alias = graphds_scc (pg, NULL, pg_skip_alias_edge); | |
2310 gcc_assert (partitions->length () == (unsigned) num_sccs_no_alias); | |
2311 /* With topological order, we can construct two subgraphs L and R. | |
2312 L contains edge <x, y> where x < y in terms of post order, while | |
2313 R contains edge <x, y> where x > y. Edges for compilation time | |
2314 known dependence all fall in R, so we break SCCs by removing all | |
2315 (alias) edges of in subgraph L. */ | |
2316 for_each_edge (pg, pg_collect_alias_ddrs, &cbdata); | |
2317 } | |
2318 | |
2319 /* For SCC that doesn't need to be broken, merge it. */ | |
2320 for (i = 0; i < num_sccs; ++i) | |
2321 { | |
2322 if (!bitmap_bit_p (sccs_to_merge, i)) | |
2323 continue; | |
2324 | |
2325 for (j = 0; partitions->iterate (j, &first); ++j) | |
2326 if (cbdata.vertices_component[j] == i) | |
2327 break; | |
2328 for (k = j + 1; partitions->iterate (k, &partition); ++k) | |
2329 { | |
2330 struct pg_vdata *data; | |
2331 | |
2332 if (cbdata.vertices_component[k] != i) | |
2333 continue; | |
2334 | |
2335 /* Update postorder number so that merged reduction partition is | |
2336 sorted after other partitions. */ | |
2337 if (!partition_reduction_p (first) | |
2338 && partition_reduction_p (partition)) | |
2339 { | |
2340 gcc_assert (pg->vertices[k].post < pg->vertices[j].post); | |
2341 pg->vertices[j].post = pg->vertices[k].post; | |
2342 } | |
2343 partition_merge_into (NULL, first, partition, FUSE_SAME_SCC); | |
2344 (*partitions)[k] = NULL; | |
2345 partition_free (partition); | |
2346 data = (struct pg_vdata *)pg->vertices[k].data; | |
2347 gcc_assert (data->id == k); | |
2348 data->partition = NULL; | |
2349 /* The result partition of merged SCC must be sequential. */ | |
2350 first->type = PTYPE_SEQUENTIAL; | |
2351 } | |
2352 } | |
2353 } | |
2354 | |
2355 sort_partitions_by_post_order (pg, partitions); | |
2356 free_partition_graph_vdata (pg); | |
2357 for_each_edge (pg, free_partition_graph_edata_cb, NULL); | |
2358 free_graph (pg); | |
0 | 2359 |
2360 if (dump_file && (dump_flags & TDF_DETAILS)) | |
111 | 2361 { |
2362 fprintf (dump_file, "Possible alias data dependence to break:\n"); | |
2363 dump_data_dependence_relations (dump_file, *alias_ddrs); | |
2364 } | |
2365 } | |
2366 | |
2367 /* Compute and return an expression whose value is the segment length which | |
2368 will be accessed by DR in NITERS iterations. */ | |
2369 | |
2370 static tree | |
2371 data_ref_segment_size (struct data_reference *dr, tree niters) | |
2372 { | |
131 | 2373 niters = size_binop (MINUS_EXPR, |
2374 fold_convert (sizetype, niters), | |
2375 size_one_node); | |
2376 return size_binop (MULT_EXPR, | |
2377 fold_convert (sizetype, DR_STEP (dr)), | |
2378 fold_convert (sizetype, niters)); | |
111 | 2379 } |
2380 | |
2381 /* Return true if LOOP's latch is dominated by statement for data reference | |
2382 DR. */ | |
2383 | |
2384 static inline bool | |
2385 latch_dominated_by_data_ref (struct loop *loop, data_reference *dr) | |
2386 { | |
2387 return dominated_by_p (CDI_DOMINATORS, single_exit (loop)->src, | |
2388 gimple_bb (DR_STMT (dr))); | |
2389 } | |
2390 | |
2391 /* Compute alias check pairs and store them in COMP_ALIAS_PAIRS for LOOP's | |
2392 data dependence relations ALIAS_DDRS. */ | |
2393 | |
2394 static void | |
2395 compute_alias_check_pairs (struct loop *loop, vec<ddr_p> *alias_ddrs, | |
2396 vec<dr_with_seg_len_pair_t> *comp_alias_pairs) | |
2397 { | |
2398 unsigned int i; | |
2399 unsigned HOST_WIDE_INT factor = 1; | |
2400 tree niters_plus_one, niters = number_of_latch_executions (loop); | |
2401 | |
2402 gcc_assert (niters != NULL_TREE && niters != chrec_dont_know); | |
2403 niters = fold_convert (sizetype, niters); | |
2404 niters_plus_one = size_binop (PLUS_EXPR, niters, size_one_node); | |
2405 | |
2406 if (dump_file && (dump_flags & TDF_DETAILS)) | |
2407 fprintf (dump_file, "Creating alias check pairs:\n"); | |
2408 | |
2409 /* Iterate all data dependence relations and compute alias check pairs. */ | |
2410 for (i = 0; i < alias_ddrs->length (); i++) | |
2411 { | |
2412 ddr_p ddr = (*alias_ddrs)[i]; | |
2413 struct data_reference *dr_a = DDR_A (ddr); | |
2414 struct data_reference *dr_b = DDR_B (ddr); | |
2415 tree seg_length_a, seg_length_b; | |
2416 int comp_res = data_ref_compare_tree (DR_BASE_ADDRESS (dr_a), | |
2417 DR_BASE_ADDRESS (dr_b)); | |
2418 | |
2419 if (comp_res == 0) | |
2420 comp_res = data_ref_compare_tree (DR_OFFSET (dr_a), DR_OFFSET (dr_b)); | |
2421 gcc_assert (comp_res != 0); | |
2422 | |
2423 if (latch_dominated_by_data_ref (loop, dr_a)) | |
2424 seg_length_a = data_ref_segment_size (dr_a, niters_plus_one); | |
2425 else | |
2426 seg_length_a = data_ref_segment_size (dr_a, niters); | |
2427 | |
2428 if (latch_dominated_by_data_ref (loop, dr_b)) | |
2429 seg_length_b = data_ref_segment_size (dr_b, niters_plus_one); | |
2430 else | |
2431 seg_length_b = data_ref_segment_size (dr_b, niters); | |
2432 | |
131 | 2433 unsigned HOST_WIDE_INT access_size_a |
2434 = tree_to_uhwi (TYPE_SIZE_UNIT (TREE_TYPE (DR_REF (dr_a)))); | |
2435 unsigned HOST_WIDE_INT access_size_b | |
2436 = tree_to_uhwi (TYPE_SIZE_UNIT (TREE_TYPE (DR_REF (dr_b)))); | |
2437 unsigned int align_a = TYPE_ALIGN_UNIT (TREE_TYPE (DR_REF (dr_a))); | |
2438 unsigned int align_b = TYPE_ALIGN_UNIT (TREE_TYPE (DR_REF (dr_b))); | |
2439 | |
111 | 2440 dr_with_seg_len_pair_t dr_with_seg_len_pair |
131 | 2441 (dr_with_seg_len (dr_a, seg_length_a, access_size_a, align_a), |
2442 dr_with_seg_len (dr_b, seg_length_b, access_size_b, align_b)); | |
111 | 2443 |
2444 /* Canonicalize pairs by sorting the two DR members. */ | |
2445 if (comp_res > 0) | |
2446 std::swap (dr_with_seg_len_pair.first, dr_with_seg_len_pair.second); | |
2447 | |
2448 comp_alias_pairs->safe_push (dr_with_seg_len_pair); | |
2449 } | |
2450 | |
2451 if (tree_fits_uhwi_p (niters)) | |
2452 factor = tree_to_uhwi (niters); | |
2453 | |
2454 /* Prune alias check pairs. */ | |
2455 prune_runtime_alias_test_list (comp_alias_pairs, factor); | |
2456 if (dump_file && (dump_flags & TDF_DETAILS)) | |
2457 fprintf (dump_file, | |
2458 "Improved number of alias checks from %d to %d\n", | |
2459 alias_ddrs->length (), comp_alias_pairs->length ()); | |
2460 } | |
2461 | |
2462 /* Given data dependence relations in ALIAS_DDRS, generate runtime alias | |
2463 checks and version LOOP under condition of these runtime alias checks. */ | |
2464 | |
2465 static void | |
131 | 2466 version_loop_by_alias_check (vec<struct partition *> *partitions, |
2467 struct loop *loop, vec<ddr_p> *alias_ddrs) | |
111 | 2468 { |
2469 profile_probability prob; | |
2470 basic_block cond_bb; | |
2471 struct loop *nloop; | |
2472 tree lhs, arg0, cond_expr = NULL_TREE; | |
2473 gimple_seq cond_stmts = NULL; | |
2474 gimple *call_stmt = NULL; | |
2475 auto_vec<dr_with_seg_len_pair_t> comp_alias_pairs; | |
2476 | |
2477 /* Generate code for runtime alias checks if necessary. */ | |
2478 gcc_assert (alias_ddrs->length () > 0); | |
2479 | |
2480 if (dump_file && (dump_flags & TDF_DETAILS)) | |
2481 fprintf (dump_file, | |
2482 "Version loop <%d> with runtime alias check\n", loop->num); | |
2483 | |
2484 compute_alias_check_pairs (loop, alias_ddrs, &comp_alias_pairs); | |
2485 create_runtime_alias_checks (loop, &comp_alias_pairs, &cond_expr); | |
2486 cond_expr = force_gimple_operand_1 (cond_expr, &cond_stmts, | |
2487 is_gimple_val, NULL_TREE); | |
2488 | |
2489 /* Depend on vectorizer to fold IFN_LOOP_DIST_ALIAS. */ | |
131 | 2490 bool cancelable_p = flag_tree_loop_vectorize; |
2491 if (cancelable_p) | |
111 | 2492 { |
131 | 2493 unsigned i = 0; |
2494 struct partition *partition; | |
2495 for (; partitions->iterate (i, &partition); ++i) | |
2496 if (!partition_builtin_p (partition)) | |
2497 break; | |
2498 | |
2499 /* If all partitions are builtins, distributing it would be profitable and | |
2500 we don't want to cancel the runtime alias checks. */ | |
2501 if (i == partitions->length ()) | |
2502 cancelable_p = false; | |
2503 } | |
2504 | |
2505 /* Generate internal function call for loop distribution alias check if the | |
2506 runtime alias check should be cancelable. */ | |
2507 if (cancelable_p) | |
2508 { | |
111 | 2509 call_stmt = gimple_build_call_internal (IFN_LOOP_DIST_ALIAS, |
2510 2, NULL_TREE, cond_expr); | |
2511 lhs = make_ssa_name (boolean_type_node); | |
2512 gimple_call_set_lhs (call_stmt, lhs); | |
2513 } | |
2514 else | |
2515 lhs = cond_expr; | |
2516 | |
2517 prob = profile_probability::guessed_always ().apply_scale (9, 10); | |
2518 initialize_original_copy_tables (); | |
2519 nloop = loop_version (loop, lhs, &cond_bb, prob, prob.invert (), | |
2520 prob, prob.invert (), true); | |
2521 free_original_copy_tables (); | |
2522 /* Record the original loop number in newly generated loops. In case of | |
2523 distribution, the original loop will be distributed and the new loop | |
2524 is kept. */ | |
2525 loop->orig_loop_num = nloop->num; | |
2526 nloop->orig_loop_num = nloop->num; | |
2527 nloop->dont_vectorize = true; | |
2528 nloop->force_vectorize = false; | |
2529 | |
2530 if (call_stmt) | |
2531 { | |
2532 /* Record new loop's num in IFN_LOOP_DIST_ALIAS because the original | |
2533 loop could be destroyed. */ | |
2534 arg0 = build_int_cst (integer_type_node, loop->orig_loop_num); | |
2535 gimple_call_set_arg (call_stmt, 0, arg0); | |
2536 gimple_seq_add_stmt_without_update (&cond_stmts, call_stmt); | |
2537 } | |
2538 | |
2539 if (cond_stmts) | |
2540 { | |
2541 gimple_stmt_iterator cond_gsi = gsi_last_bb (cond_bb); | |
2542 gsi_insert_seq_before (&cond_gsi, cond_stmts, GSI_SAME_STMT); | |
2543 } | |
2544 update_ssa (TODO_update_ssa); | |
0 | 2545 } |
2546 | |
111 | 2547 /* Return true if loop versioning is needed to distrubute PARTITIONS. |
2548 ALIAS_DDRS are data dependence relations for runtime alias check. */ | |
2549 | |
2550 static inline bool | |
2551 version_for_distribution_p (vec<struct partition *> *partitions, | |
2552 vec<ddr_p> *alias_ddrs) | |
2553 { | |
2554 /* No need to version loop if we have only one partition. */ | |
2555 if (partitions->length () == 1) | |
2556 return false; | |
2557 | |
2558 /* Need to version loop if runtime alias check is necessary. */ | |
2559 return (alias_ddrs->length () > 0); | |
2560 } | |
2561 | |
2562 /* Compare base offset of builtin mem* partitions P1 and P2. */ | |
2563 | |
131 | 2564 static int |
2565 offset_cmp (const void *vp1, const void *vp2) | |
111 | 2566 { |
131 | 2567 struct partition *p1 = *(struct partition *const *) vp1; |
2568 struct partition *p2 = *(struct partition *const *) vp2; | |
2569 unsigned HOST_WIDE_INT o1 = p1->builtin->dst_base_offset; | |
2570 unsigned HOST_WIDE_INT o2 = p2->builtin->dst_base_offset; | |
2571 return (o2 < o1) - (o1 < o2); | |
111 | 2572 } |
2573 | |
2574 /* Fuse adjacent memset builtin PARTITIONS if possible. This is a special | |
2575 case optimization transforming below code: | |
2576 | |
2577 __builtin_memset (&obj, 0, 100); | |
2578 _1 = &obj + 100; | |
2579 __builtin_memset (_1, 0, 200); | |
2580 _2 = &obj + 300; | |
2581 __builtin_memset (_2, 0, 100); | |
2582 | |
2583 into: | |
2584 | |
2585 __builtin_memset (&obj, 0, 400); | |
2586 | |
2587 Note we don't have dependence information between different partitions | |
2588 at this point, as a result, we can't handle nonadjacent memset builtin | |
2589 partitions since dependence might be broken. */ | |
2590 | |
2591 static void | |
2592 fuse_memset_builtins (vec<struct partition *> *partitions) | |
2593 { | |
2594 unsigned i, j; | |
2595 struct partition *part1, *part2; | |
131 | 2596 tree rhs1, rhs2; |
111 | 2597 |
2598 for (i = 0; partitions->iterate (i, &part1);) | |
2599 { | |
2600 if (part1->kind != PKIND_MEMSET) | |
2601 { | |
2602 i++; | |
2603 continue; | |
2604 } | |
2605 | |
2606 /* Find sub-array of memset builtins of the same base. Index range | |
2607 of the sub-array is [i, j) with "j > i". */ | |
2608 for (j = i + 1; partitions->iterate (j, &part2); ++j) | |
2609 { | |
2610 if (part2->kind != PKIND_MEMSET | |
2611 || !operand_equal_p (part1->builtin->dst_base_base, | |
2612 part2->builtin->dst_base_base, 0)) | |
2613 break; | |
131 | 2614 |
2615 /* Memset calls setting different values can't be merged. */ | |
2616 rhs1 = gimple_assign_rhs1 (DR_STMT (part1->builtin->dst_dr)); | |
2617 rhs2 = gimple_assign_rhs1 (DR_STMT (part2->builtin->dst_dr)); | |
2618 if (!operand_equal_p (rhs1, rhs2, 0)) | |
2619 break; | |
111 | 2620 } |
2621 | |
2622 /* Stable sort is required in order to avoid breaking dependence. */ | |
131 | 2623 gcc_stablesort (&(*partitions)[i], j - i, sizeof (*partitions)[i], |
2624 offset_cmp); | |
111 | 2625 /* Continue with next partition. */ |
2626 i = j; | |
2627 } | |
2628 | |
2629 /* Merge all consecutive memset builtin partitions. */ | |
2630 for (i = 0; i < partitions->length () - 1;) | |
2631 { | |
2632 part1 = (*partitions)[i]; | |
2633 if (part1->kind != PKIND_MEMSET) | |
2634 { | |
2635 i++; | |
2636 continue; | |
2637 } | |
2638 | |
2639 part2 = (*partitions)[i + 1]; | |
2640 /* Only merge memset partitions of the same base and with constant | |
2641 access sizes. */ | |
2642 if (part2->kind != PKIND_MEMSET | |
2643 || TREE_CODE (part1->builtin->size) != INTEGER_CST | |
2644 || TREE_CODE (part2->builtin->size) != INTEGER_CST | |
2645 || !operand_equal_p (part1->builtin->dst_base_base, | |
2646 part2->builtin->dst_base_base, 0)) | |
2647 { | |
2648 i++; | |
2649 continue; | |
2650 } | |
131 | 2651 rhs1 = gimple_assign_rhs1 (DR_STMT (part1->builtin->dst_dr)); |
2652 rhs2 = gimple_assign_rhs1 (DR_STMT (part2->builtin->dst_dr)); | |
111 | 2653 int bytev1 = const_with_all_bytes_same (rhs1); |
2654 int bytev2 = const_with_all_bytes_same (rhs2); | |
2655 /* Only merge memset partitions of the same value. */ | |
2656 if (bytev1 != bytev2 || bytev1 == -1) | |
2657 { | |
2658 i++; | |
2659 continue; | |
2660 } | |
2661 wide_int end1 = wi::add (part1->builtin->dst_base_offset, | |
2662 wi::to_wide (part1->builtin->size)); | |
2663 /* Only merge adjacent memset partitions. */ | |
2664 if (wi::ne_p (end1, part2->builtin->dst_base_offset)) | |
2665 { | |
2666 i++; | |
2667 continue; | |
2668 } | |
2669 /* Merge partitions[i] and partitions[i+1]. */ | |
2670 part1->builtin->size = fold_build2 (PLUS_EXPR, sizetype, | |
2671 part1->builtin->size, | |
2672 part2->builtin->size); | |
2673 partition_free (part2); | |
2674 partitions->ordered_remove (i + 1); | |
2675 } | |
2676 } | |
2677 | |
2678 /* Fuse PARTITIONS of LOOP if necessary before finalizing distribution. | |
2679 ALIAS_DDRS contains ddrs which need runtime alias check. */ | |
2680 | |
2681 static void | |
2682 finalize_partitions (struct loop *loop, vec<struct partition *> *partitions, | |
2683 vec<ddr_p> *alias_ddrs) | |
2684 { | |
2685 unsigned i; | |
2686 struct partition *partition, *a; | |
2687 | |
2688 if (partitions->length () == 1 | |
2689 || alias_ddrs->length () > 0) | |
2690 return; | |
2691 | |
131 | 2692 unsigned num_builtin = 0, num_normal = 0, num_partial_memset = 0; |
111 | 2693 bool same_type_p = true; |
2694 enum partition_type type = ((*partitions)[0])->type; | |
2695 for (i = 0; partitions->iterate (i, &partition); ++i) | |
2696 { | |
2697 same_type_p &= (type == partition->type); | |
131 | 2698 if (partition_builtin_p (partition)) |
2699 { | |
2700 num_builtin++; | |
2701 continue; | |
2702 } | |
2703 num_normal++; | |
2704 if (partition->kind == PKIND_PARTIAL_MEMSET) | |
2705 num_partial_memset++; | |
111 | 2706 } |
2707 | |
2708 /* Don't distribute current loop into too many loops given we don't have | |
2709 memory stream cost model. Be even more conservative in case of loop | |
2710 nest distribution. */ | |
131 | 2711 if ((same_type_p && num_builtin == 0 |
2712 && (loop->inner == NULL || num_normal != 2 || num_partial_memset != 1)) | |
111 | 2713 || (loop->inner != NULL |
2714 && i >= NUM_PARTITION_THRESHOLD && num_normal > 1) | |
2715 || (loop->inner == NULL | |
2716 && i >= NUM_PARTITION_THRESHOLD && num_normal > num_builtin)) | |
2717 { | |
2718 a = (*partitions)[0]; | |
2719 for (i = 1; partitions->iterate (i, &partition); ++i) | |
2720 { | |
2721 partition_merge_into (NULL, a, partition, FUSE_FINALIZE); | |
2722 partition_free (partition); | |
2723 } | |
2724 partitions->truncate (1); | |
2725 } | |
2726 | |
2727 /* Fuse memset builtins if possible. */ | |
2728 if (partitions->length () > 1) | |
2729 fuse_memset_builtins (partitions); | |
2730 } | |
2731 | |
2732 /* Distributes the code from LOOP in such a way that producer statements | |
2733 are placed before consumer statements. Tries to separate only the | |
2734 statements from STMTS into separate loops. Returns the number of | |
2735 distributed loops. Set NB_CALLS to number of generated builtin calls. | |
2736 Set *DESTROY_P to whether LOOP needs to be destroyed. */ | |
0 | 2737 |
2738 static int | |
111 | 2739 distribute_loop (struct loop *loop, vec<gimple *> stmts, |
2740 control_dependences *cd, int *nb_calls, bool *destroy_p) | |
0 | 2741 { |
111 | 2742 ddrs_table = new hash_table<ddr_hasher> (389); |
0 | 2743 struct graph *rdg; |
111 | 2744 partition *partition; |
2745 bool any_builtin; | |
2746 int i, nbp; | |
2747 | |
2748 *destroy_p = false; | |
2749 *nb_calls = 0; | |
2750 loop_nest.create (0); | |
2751 if (!find_loop_nest (loop, &loop_nest)) | |
0 | 2752 { |
111 | 2753 loop_nest.release (); |
2754 delete ddrs_table; | |
2755 return 0; | |
0 | 2756 } |
2757 | |
111 | 2758 datarefs_vec.create (20); |
2759 rdg = build_rdg (loop, cd); | |
0 | 2760 if (!rdg) |
2761 { | |
2762 if (dump_file && (dump_flags & TDF_DETAILS)) | |
2763 fprintf (dump_file, | |
111 | 2764 "Loop %d not distributed: failed to build the RDG.\n", |
0 | 2765 loop->num); |
2766 | |
111 | 2767 loop_nest.release (); |
2768 free_data_refs (datarefs_vec); | |
2769 delete ddrs_table; | |
2770 return 0; | |
0 | 2771 } |
2772 | |
111 | 2773 if (datarefs_vec.length () > MAX_DATAREFS_NUM) |
2774 { | |
2775 if (dump_file && (dump_flags & TDF_DETAILS)) | |
2776 fprintf (dump_file, | |
2777 "Loop %d not distributed: too many memory references.\n", | |
2778 loop->num); | |
2779 | |
2780 free_rdg (rdg); | |
2781 loop_nest.release (); | |
2782 free_data_refs (datarefs_vec); | |
2783 delete ddrs_table; | |
2784 return 0; | |
2785 } | |
2786 | |
2787 data_reference_p dref; | |
2788 for (i = 0; datarefs_vec.iterate (i, &dref); ++i) | |
2789 dref->aux = (void *) (uintptr_t) i; | |
0 | 2790 |
2791 if (dump_file && (dump_flags & TDF_DETAILS)) | |
2792 dump_rdg (dump_file, rdg); | |
2793 | |
111 | 2794 auto_vec<struct partition *, 3> partitions; |
2795 rdg_build_partitions (rdg, stmts, &partitions); | |
2796 | |
2797 auto_vec<ddr_p> alias_ddrs; | |
2798 | |
2799 auto_bitmap stmt_in_all_partitions; | |
2800 bitmap_copy (stmt_in_all_partitions, partitions[0]->stmts); | |
2801 for (i = 1; partitions.iterate (i, &partition); ++i) | |
2802 bitmap_and_into (stmt_in_all_partitions, partitions[i]->stmts); | |
2803 | |
2804 any_builtin = false; | |
2805 FOR_EACH_VEC_ELT (partitions, i, partition) | |
2806 { | |
2807 classify_partition (loop, rdg, partition, stmt_in_all_partitions); | |
2808 any_builtin |= partition_builtin_p (partition); | |
2809 } | |
2810 | |
2811 /* If we are only distributing patterns but did not detect any, | |
2812 simply bail out. */ | |
2813 if (!flag_tree_loop_distribution | |
2814 && !any_builtin) | |
2815 { | |
2816 nbp = 0; | |
2817 goto ldist_done; | |
2818 } | |
2819 | |
2820 /* If we are only distributing patterns fuse all partitions that | |
2821 were not classified as builtins. This also avoids chopping | |
2822 a loop into pieces, separated by builtin calls. That is, we | |
2823 only want no or a single loop body remaining. */ | |
2824 struct partition *into; | |
2825 if (!flag_tree_loop_distribution) | |
0 | 2826 { |
111 | 2827 for (i = 0; partitions.iterate (i, &into); ++i) |
2828 if (!partition_builtin_p (into)) | |
2829 break; | |
2830 for (++i; partitions.iterate (i, &partition); ++i) | |
2831 if (!partition_builtin_p (partition)) | |
2832 { | |
2833 partition_merge_into (NULL, into, partition, FUSE_NON_BUILTIN); | |
2834 partitions.unordered_remove (i); | |
2835 partition_free (partition); | |
2836 i--; | |
2837 } | |
2838 } | |
2839 | |
2840 /* Due to limitations in the transform phase we have to fuse all | |
2841 reduction partitions into the last partition so the existing | |
2842 loop will contain all loop-closed PHI nodes. */ | |
2843 for (i = 0; partitions.iterate (i, &into); ++i) | |
2844 if (partition_reduction_p (into)) | |
2845 break; | |
2846 for (i = i + 1; partitions.iterate (i, &partition); ++i) | |
2847 if (partition_reduction_p (partition)) | |
2848 { | |
2849 partition_merge_into (rdg, into, partition, FUSE_REDUCTION); | |
2850 partitions.unordered_remove (i); | |
2851 partition_free (partition); | |
2852 i--; | |
2853 } | |
2854 | |
2855 /* Apply our simple cost model - fuse partitions with similar | |
2856 memory accesses. */ | |
2857 for (i = 0; partitions.iterate (i, &into); ++i) | |
2858 { | |
2859 bool changed = false; | |
131 | 2860 if (partition_builtin_p (into) || into->kind == PKIND_PARTIAL_MEMSET) |
111 | 2861 continue; |
2862 for (int j = i + 1; | |
2863 partitions.iterate (j, &partition); ++j) | |
0 | 2864 { |
111 | 2865 if (share_memory_accesses (rdg, into, partition)) |
2866 { | |
2867 partition_merge_into (rdg, into, partition, FUSE_SHARE_REF); | |
2868 partitions.unordered_remove (j); | |
2869 partition_free (partition); | |
2870 j--; | |
2871 changed = true; | |
2872 } | |
2873 } | |
2874 /* If we fused 0 1 2 in step 1 to 0,2 1 as 0 and 2 have similar | |
2875 accesses when 1 and 2 have similar accesses but not 0 and 1 | |
2876 then in the next iteration we will fail to consider merging | |
2877 1 into 0,2. So try again if we did any merging into 0. */ | |
2878 if (changed) | |
2879 i--; | |
2880 } | |
2881 | |
2882 /* Build the partition dependency graph and fuse partitions in strong | |
2883 connected component. */ | |
2884 if (partitions.length () > 1) | |
2885 { | |
2886 /* Don't support loop nest distribution under runtime alias check | |
2887 since it's not likely to enable many vectorization opportunities. */ | |
2888 if (loop->inner) | |
2889 merge_dep_scc_partitions (rdg, &partitions, false); | |
2890 else | |
2891 { | |
2892 merge_dep_scc_partitions (rdg, &partitions, true); | |
2893 if (partitions.length () > 1) | |
2894 break_alias_scc_partitions (rdg, &partitions, &alias_ddrs); | |
0 | 2895 } |
2896 } | |
2897 | |
111 | 2898 finalize_partitions (loop, &partitions, &alias_ddrs); |
2899 | |
2900 nbp = partitions.length (); | |
2901 if (nbp == 0 | |
2902 || (nbp == 1 && !partition_builtin_p (partitions[0])) | |
2903 || (nbp > 1 && partition_contains_all_rw (rdg, partitions))) | |
2904 { | |
2905 nbp = 0; | |
2906 goto ldist_done; | |
2907 } | |
2908 | |
2909 if (version_for_distribution_p (&partitions, &alias_ddrs)) | |
131 | 2910 version_loop_by_alias_check (&partitions, loop, &alias_ddrs); |
111 | 2911 |
2912 if (dump_file && (dump_flags & TDF_DETAILS)) | |
2913 { | |
2914 fprintf (dump_file, | |
2915 "distribute loop <%d> into partitions:\n", loop->num); | |
2916 dump_rdg_partitions (dump_file, partitions); | |
2917 } | |
2918 | |
2919 FOR_EACH_VEC_ELT (partitions, i, partition) | |
2920 { | |
2921 if (partition_builtin_p (partition)) | |
2922 (*nb_calls)++; | |
2923 *destroy_p |= generate_code_for_partition (loop, partition, i < nbp - 1); | |
2924 } | |
2925 | |
2926 ldist_done: | |
2927 loop_nest.release (); | |
2928 free_data_refs (datarefs_vec); | |
2929 for (hash_table<ddr_hasher>::iterator iter = ddrs_table->begin (); | |
2930 iter != ddrs_table->end (); ++iter) | |
2931 { | |
2932 free_dependence_relation (*iter); | |
2933 *iter = NULL; | |
2934 } | |
2935 delete ddrs_table; | |
2936 | |
2937 FOR_EACH_VEC_ELT (partitions, i, partition) | |
2938 partition_free (partition); | |
2939 | |
0 | 2940 free_rdg (rdg); |
111 | 2941 return nbp - *nb_calls; |
0 | 2942 } |
2943 | |
2944 /* Distribute all loops in the current function. */ | |
2945 | |
111 | 2946 namespace { |
2947 | |
2948 const pass_data pass_data_loop_distribution = | |
2949 { | |
2950 GIMPLE_PASS, /* type */ | |
2951 "ldist", /* name */ | |
2952 OPTGROUP_LOOP, /* optinfo_flags */ | |
2953 TV_TREE_LOOP_DISTRIBUTION, /* tv_id */ | |
2954 ( PROP_cfg | PROP_ssa ), /* properties_required */ | |
2955 0, /* properties_provided */ | |
2956 0, /* properties_destroyed */ | |
2957 0, /* todo_flags_start */ | |
2958 0, /* todo_flags_finish */ | |
2959 }; | |
2960 | |
2961 class pass_loop_distribution : public gimple_opt_pass | |
2962 { | |
2963 public: | |
2964 pass_loop_distribution (gcc::context *ctxt) | |
2965 : gimple_opt_pass (pass_data_loop_distribution, ctxt) | |
2966 {} | |
2967 | |
2968 /* opt_pass methods: */ | |
2969 virtual bool gate (function *) | |
2970 { | |
2971 return flag_tree_loop_distribution | |
2972 || flag_tree_loop_distribute_patterns; | |
2973 } | |
2974 | |
2975 virtual unsigned int execute (function *); | |
2976 | |
2977 }; // class pass_loop_distribution | |
2978 | |
2979 | |
2980 /* Given LOOP, this function records seed statements for distribution in | |
2981 WORK_LIST. Return false if there is nothing for distribution. */ | |
2982 | |
2983 static bool | |
2984 find_seed_stmts_for_distribution (struct loop *loop, vec<gimple *> *work_list) | |
2985 { | |
2986 basic_block *bbs = get_loop_body_in_dom_order (loop); | |
2987 | |
2988 /* Initialize the worklist with stmts we seed the partitions with. */ | |
2989 for (unsigned i = 0; i < loop->num_nodes; ++i) | |
2990 { | |
2991 for (gphi_iterator gsi = gsi_start_phis (bbs[i]); | |
2992 !gsi_end_p (gsi); gsi_next (&gsi)) | |
2993 { | |
2994 gphi *phi = gsi.phi (); | |
2995 if (virtual_operand_p (gimple_phi_result (phi))) | |
2996 continue; | |
2997 /* Distribute stmts which have defs that are used outside of | |
2998 the loop. */ | |
2999 if (!stmt_has_scalar_dependences_outside_loop (loop, phi)) | |
3000 continue; | |
3001 work_list->safe_push (phi); | |
3002 } | |
3003 for (gimple_stmt_iterator gsi = gsi_start_bb (bbs[i]); | |
3004 !gsi_end_p (gsi); gsi_next (&gsi)) | |
3005 { | |
3006 gimple *stmt = gsi_stmt (gsi); | |
3007 | |
3008 /* If there is a stmt with side-effects bail out - we | |
3009 cannot and should not distribute this loop. */ | |
3010 if (gimple_has_side_effects (stmt)) | |
3011 { | |
3012 free (bbs); | |
3013 return false; | |
3014 } | |
3015 | |
3016 /* Distribute stmts which have defs that are used outside of | |
3017 the loop. */ | |
3018 if (stmt_has_scalar_dependences_outside_loop (loop, stmt)) | |
3019 ; | |
3020 /* Otherwise only distribute stores for now. */ | |
3021 else if (!gimple_vdef (stmt)) | |
3022 continue; | |
3023 | |
3024 work_list->safe_push (stmt); | |
3025 } | |
3026 } | |
3027 free (bbs); | |
3028 return work_list->length () > 0; | |
3029 } | |
3030 | |
3031 /* Given innermost LOOP, return the outermost enclosing loop that forms a | |
3032 perfect loop nest. */ | |
3033 | |
3034 static struct loop * | |
3035 prepare_perfect_loop_nest (struct loop *loop) | |
3036 { | |
3037 struct loop *outer = loop_outer (loop); | |
3038 tree niters = number_of_latch_executions (loop); | |
3039 | |
131 | 3040 /* TODO: We only support the innermost 3-level loop nest distribution |
111 | 3041 because of compilation time issue for now. This should be relaxed |
131 | 3042 in the future. Note we only allow 3-level loop nest distribution |
3043 when parallelizing loops. */ | |
3044 while ((loop->inner == NULL | |
3045 || (loop->inner->inner == NULL && flag_tree_parallelize_loops > 1)) | |
111 | 3046 && loop_outer (outer) |
3047 && outer->inner == loop && loop->next == NULL | |
3048 && single_exit (outer) | |
3049 && optimize_loop_for_speed_p (outer) | |
3050 && !chrec_contains_symbols_defined_in_loop (niters, outer->num) | |
3051 && (niters = number_of_latch_executions (outer)) != NULL_TREE | |
3052 && niters != chrec_dont_know) | |
3053 { | |
3054 loop = outer; | |
3055 outer = loop_outer (loop); | |
3056 } | |
3057 | |
3058 return loop; | |
3059 } | |
3060 | |
3061 unsigned int | |
3062 pass_loop_distribution::execute (function *fun) | |
0 | 3063 { |
3064 struct loop *loop; | |
111 | 3065 bool changed = false; |
3066 basic_block bb; | |
3067 control_dependences *cd = NULL; | |
3068 auto_vec<loop_p> loops_to_be_destroyed; | |
3069 | |
3070 if (number_of_loops (fun) <= 1) | |
3071 return 0; | |
3072 | |
3073 /* Compute topological order for basic blocks. Topological order is | |
3074 needed because data dependence is computed for data references in | |
3075 lexicographical order. */ | |
3076 if (bb_top_order_index == NULL) | |
0 | 3077 { |
111 | 3078 int rpo_num; |
3079 int *rpo = XNEWVEC (int, last_basic_block_for_fn (cfun)); | |
3080 | |
3081 bb_top_order_index = XNEWVEC (int, last_basic_block_for_fn (cfun)); | |
3082 bb_top_order_index_size = last_basic_block_for_fn (cfun); | |
3083 rpo_num = pre_and_rev_post_order_compute_fn (cfun, NULL, rpo, true); | |
3084 for (int i = 0; i < rpo_num; i++) | |
3085 bb_top_order_index[rpo[i]] = i; | |
3086 | |
3087 free (rpo); | |
3088 } | |
3089 | |
3090 FOR_ALL_BB_FN (bb, fun) | |
3091 { | |
3092 gimple_stmt_iterator gsi; | |
3093 for (gsi = gsi_start_phis (bb); !gsi_end_p (gsi); gsi_next (&gsi)) | |
3094 gimple_set_uid (gsi_stmt (gsi), -1); | |
3095 for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi)) | |
3096 gimple_set_uid (gsi_stmt (gsi), -1); | |
3097 } | |
3098 | |
3099 /* We can at the moment only distribute non-nested loops, thus restrict | |
3100 walking to innermost loops. */ | |
3101 FOR_EACH_LOOP (loop, LI_ONLY_INNERMOST) | |
3102 { | |
3103 /* Don't distribute multiple exit edges loop, or cold loop. */ | |
3104 if (!single_exit (loop) | |
3105 || !optimize_loop_for_speed_p (loop)) | |
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
3106 continue; |
0 | 3107 |
111 | 3108 /* Don't distribute loop if niters is unknown. */ |
3109 tree niters = number_of_latch_executions (loop); | |
3110 if (niters == NULL_TREE || niters == chrec_dont_know) | |
3111 continue; | |
3112 | |
3113 /* Get the perfect loop nest for distribution. */ | |
3114 loop = prepare_perfect_loop_nest (loop); | |
3115 for (; loop; loop = loop->inner) | |
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
3116 { |
111 | 3117 auto_vec<gimple *> work_list; |
3118 if (!find_seed_stmts_for_distribution (loop, &work_list)) | |
3119 break; | |
3120 | |
3121 const char *str = loop->inner ? " nest" : ""; | |
131 | 3122 dump_user_location_t loc = find_loop_location (loop); |
111 | 3123 if (!cd) |
3124 { | |
3125 calculate_dominance_info (CDI_DOMINATORS); | |
3126 calculate_dominance_info (CDI_POST_DOMINATORS); | |
3127 cd = new control_dependences (); | |
3128 free_dominance_info (CDI_POST_DOMINATORS); | |
3129 } | |
3130 | |
3131 bool destroy_p; | |
3132 int nb_generated_loops, nb_generated_calls; | |
3133 nb_generated_loops = distribute_loop (loop, work_list, cd, | |
3134 &nb_generated_calls, | |
3135 &destroy_p); | |
3136 if (destroy_p) | |
3137 loops_to_be_destroyed.safe_push (loop); | |
3138 | |
3139 if (nb_generated_loops + nb_generated_calls > 0) | |
3140 { | |
3141 changed = true; | |
3142 dump_printf_loc (MSG_OPTIMIZED_LOCATIONS, | |
3143 loc, "Loop%s %d distributed: split to %d loops " | |
3144 "and %d library calls.\n", str, loop->num, | |
3145 nb_generated_loops, nb_generated_calls); | |
3146 | |
3147 break; | |
3148 } | |
3149 | |
3150 if (dump_file && (dump_flags & TDF_DETAILS)) | |
3151 fprintf (dump_file, "Loop%s %d not distributed.\n", str, loop->num); | |
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
3152 } |
111 | 3153 } |
3154 | |
3155 if (cd) | |
3156 delete cd; | |
3157 | |
3158 if (bb_top_order_index != NULL) | |
3159 { | |
3160 free (bb_top_order_index); | |
3161 bb_top_order_index = NULL; | |
3162 bb_top_order_index_size = 0; | |
0 | 3163 } |
3164 | |
111 | 3165 if (changed) |
3166 { | |
3167 /* Destroy loop bodies that could not be reused. Do this late as we | |
3168 otherwise can end up refering to stale data in control dependences. */ | |
3169 unsigned i; | |
3170 FOR_EACH_VEC_ELT (loops_to_be_destroyed, i, loop) | |
3171 destroy_loop (loop); | |
3172 | |
3173 /* Cached scalar evolutions now may refer to wrong or non-existing | |
3174 loops. */ | |
3175 scev_reset_htab (); | |
3176 mark_virtual_operands_for_renaming (fun); | |
3177 rewrite_into_loop_closed_ssa (NULL, TODO_update_ssa); | |
3178 } | |
3179 | |
3180 checking_verify_loop_structure (); | |
3181 | |
131 | 3182 return changed ? TODO_cleanup_cfg : 0; |
0 | 3183 } |
3184 | |
111 | 3185 } // anon namespace |
3186 | |
3187 gimple_opt_pass * | |
3188 make_pass_loop_distribution (gcc::context *ctxt) | |
0 | 3189 { |
111 | 3190 return new pass_loop_distribution (ctxt); |
3191 } |