Mercurial > hg > CbC > CbC_gcc
annotate gcc/tree-loop-distribution.c @ 127:4c56639505ff
fix function.c and add CbC-example Makefile
author | mir3636 |
---|---|
date | Wed, 11 Apr 2018 18:46:58 +0900 |
parents | 04ced10e8804 |
children | 84e7813d76e9 |
rev | line source |
---|---|
0 | 1 /* Loop distribution. |
111 | 2 Copyright (C) 2006-2017 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" | |
111 | 93 #define INCLUDE_ALGORITHM /* stable_sort */ |
0 | 94 #include "system.h" |
95 #include "coretypes.h" | |
111 | 96 #include "backend.h" |
97 #include "tree.h" | |
98 #include "gimple.h" | |
99 #include "cfghooks.h" | |
0 | 100 #include "tree-pass.h" |
111 | 101 #include "ssa.h" |
102 #include "gimple-pretty-print.h" | |
103 #include "fold-const.h" | |
104 #include "cfganal.h" | |
105 #include "gimple-iterator.h" | |
106 #include "gimplify-me.h" | |
107 #include "stor-layout.h" | |
108 #include "tree-cfg.h" | |
109 #include "tree-ssa-loop-manip.h" | |
110 #include "tree-ssa-loop-ivopts.h" | |
111 #include "tree-ssa-loop.h" | |
112 #include "tree-into-ssa.h" | |
113 #include "tree-ssa.h" | |
114 #include "cfgloop.h" | |
115 #include "tree-scalar-evolution.h" | |
116 #include "params.h" | |
117 #include "tree-vectorizer.h" | |
118 | |
119 | |
120 #define MAX_DATAREFS_NUM \ | |
121 ((unsigned) PARAM_VALUE (PARAM_LOOP_MAX_DATAREFS_FOR_DATADEPS)) | |
122 | |
123 /* Threshold controlling number of distributed partitions. Given it may | |
124 be unnecessary if a memory stream cost model is invented in the future, | |
125 we define it as a temporary macro, rather than a parameter. */ | |
126 #define NUM_PARTITION_THRESHOLD (4) | |
127 | |
128 /* Hashtable helpers. */ | |
129 | |
130 struct ddr_hasher : nofree_ptr_hash <struct data_dependence_relation> | |
131 { | |
132 static inline hashval_t hash (const data_dependence_relation *); | |
133 static inline bool equal (const data_dependence_relation *, | |
134 const data_dependence_relation *); | |
135 }; | |
136 | |
137 /* Hash function for data dependence. */ | |
138 | |
139 inline hashval_t | |
140 ddr_hasher::hash (const data_dependence_relation *ddr) | |
141 { | |
142 inchash::hash h; | |
143 h.add_ptr (DDR_A (ddr)); | |
144 h.add_ptr (DDR_B (ddr)); | |
145 return h.end (); | |
146 } | |
147 | |
148 /* Hash table equality function for data dependence. */ | |
149 | |
150 inline bool | |
151 ddr_hasher::equal (const data_dependence_relation *ddr1, | |
152 const data_dependence_relation *ddr2) | |
153 { | |
154 return (DDR_A (ddr1) == DDR_A (ddr2) && DDR_B (ddr1) == DDR_B (ddr2)); | |
155 } | |
156 | |
157 /* The loop (nest) to be distributed. */ | |
158 static vec<loop_p> loop_nest; | |
159 | |
160 /* Vector of data references in the loop to be distributed. */ | |
161 static vec<data_reference_p> datarefs_vec; | |
162 | |
163 /* Store index of data reference in aux field. */ | |
164 #define DR_INDEX(dr) ((uintptr_t) (dr)->aux) | |
165 | |
166 /* Hash table for data dependence relation in the loop to be distributed. */ | |
167 static hash_table<ddr_hasher> *ddrs_table; | |
168 | |
169 /* A Reduced Dependence Graph (RDG) vertex representing a statement. */ | |
170 struct rdg_vertex | |
171 { | |
172 /* The statement represented by this vertex. */ | |
173 gimple *stmt; | |
174 | |
175 /* Vector of data-references in this statement. */ | |
176 vec<data_reference_p> datarefs; | |
177 | |
178 /* True when the statement contains a write to memory. */ | |
179 bool has_mem_write; | |
180 | |
181 /* True when the statement contains a read from memory. */ | |
182 bool has_mem_reads; | |
183 }; | |
184 | |
185 #define RDGV_STMT(V) ((struct rdg_vertex *) ((V)->data))->stmt | |
186 #define RDGV_DATAREFS(V) ((struct rdg_vertex *) ((V)->data))->datarefs | |
187 #define RDGV_HAS_MEM_WRITE(V) ((struct rdg_vertex *) ((V)->data))->has_mem_write | |
188 #define RDGV_HAS_MEM_READS(V) ((struct rdg_vertex *) ((V)->data))->has_mem_reads | |
189 #define RDG_STMT(RDG, I) RDGV_STMT (&(RDG->vertices[I])) | |
190 #define RDG_DATAREFS(RDG, I) RDGV_DATAREFS (&(RDG->vertices[I])) | |
191 #define RDG_MEM_WRITE_STMT(RDG, I) RDGV_HAS_MEM_WRITE (&(RDG->vertices[I])) | |
192 #define RDG_MEM_READS_STMT(RDG, I) RDGV_HAS_MEM_READS (&(RDG->vertices[I])) | |
193 | |
194 /* Data dependence type. */ | |
195 | |
196 enum rdg_dep_type | |
197 { | |
198 /* Read After Write (RAW). */ | |
199 flow_dd = 'f', | |
200 | |
201 /* Control dependence (execute conditional on). */ | |
202 control_dd = 'c' | |
203 }; | |
204 | |
205 /* Dependence information attached to an edge of the RDG. */ | |
206 | |
207 struct rdg_edge | |
208 { | |
209 /* Type of the dependence. */ | |
210 enum rdg_dep_type type; | |
211 }; | |
212 | |
213 #define RDGE_TYPE(E) ((struct rdg_edge *) ((E)->data))->type | |
214 | |
215 /* Dump vertex I in RDG to FILE. */ | |
216 | |
217 static void | |
218 dump_rdg_vertex (FILE *file, struct graph *rdg, int i) | |
219 { | |
220 struct vertex *v = &(rdg->vertices[i]); | |
221 struct graph_edge *e; | |
222 | |
223 fprintf (file, "(vertex %d: (%s%s) (in:", i, | |
224 RDG_MEM_WRITE_STMT (rdg, i) ? "w" : "", | |
225 RDG_MEM_READS_STMT (rdg, i) ? "r" : ""); | |
226 | |
227 if (v->pred) | |
228 for (e = v->pred; e; e = e->pred_next) | |
229 fprintf (file, " %d", e->src); | |
230 | |
231 fprintf (file, ") (out:"); | |
232 | |
233 if (v->succ) | |
234 for (e = v->succ; e; e = e->succ_next) | |
235 fprintf (file, " %d", e->dest); | |
236 | |
237 fprintf (file, ")\n"); | |
238 print_gimple_stmt (file, RDGV_STMT (v), 0, TDF_VOPS|TDF_MEMSYMS); | |
239 fprintf (file, ")\n"); | |
240 } | |
241 | |
242 /* Call dump_rdg_vertex on stderr. */ | |
243 | |
244 DEBUG_FUNCTION void | |
245 debug_rdg_vertex (struct graph *rdg, int i) | |
246 { | |
247 dump_rdg_vertex (stderr, rdg, i); | |
248 } | |
249 | |
250 /* Dump the reduced dependence graph RDG to FILE. */ | |
251 | |
252 static void | |
253 dump_rdg (FILE *file, struct graph *rdg) | |
254 { | |
255 fprintf (file, "(rdg\n"); | |
256 for (int i = 0; i < rdg->n_vertices; i++) | |
257 dump_rdg_vertex (file, rdg, i); | |
258 fprintf (file, ")\n"); | |
259 } | |
260 | |
261 /* Call dump_rdg on stderr. */ | |
262 | |
263 DEBUG_FUNCTION void | |
264 debug_rdg (struct graph *rdg) | |
265 { | |
266 dump_rdg (stderr, rdg); | |
267 } | |
268 | |
269 static void | |
270 dot_rdg_1 (FILE *file, struct graph *rdg) | |
271 { | |
272 int i; | |
273 pretty_printer buffer; | |
274 pp_needs_newline (&buffer) = false; | |
275 buffer.buffer->stream = file; | |
276 | |
277 fprintf (file, "digraph RDG {\n"); | |
278 | |
279 for (i = 0; i < rdg->n_vertices; i++) | |
280 { | |
281 struct vertex *v = &(rdg->vertices[i]); | |
282 struct graph_edge *e; | |
283 | |
284 fprintf (file, "%d [label=\"[%d] ", i, i); | |
285 pp_gimple_stmt_1 (&buffer, RDGV_STMT (v), 0, TDF_SLIM); | |
286 pp_flush (&buffer); | |
287 fprintf (file, "\"]\n"); | |
288 | |
289 /* Highlight reads from memory. */ | |
290 if (RDG_MEM_READS_STMT (rdg, i)) | |
291 fprintf (file, "%d [style=filled, fillcolor=green]\n", i); | |
292 | |
293 /* Highlight stores to memory. */ | |
294 if (RDG_MEM_WRITE_STMT (rdg, i)) | |
295 fprintf (file, "%d [style=filled, fillcolor=red]\n", i); | |
296 | |
297 if (v->succ) | |
298 for (e = v->succ; e; e = e->succ_next) | |
299 switch (RDGE_TYPE (e)) | |
300 { | |
301 case flow_dd: | |
302 /* These are the most common dependences: don't print these. */ | |
303 fprintf (file, "%d -> %d \n", i, e->dest); | |
304 break; | |
305 | |
306 case control_dd: | |
307 fprintf (file, "%d -> %d [label=control] \n", i, e->dest); | |
308 break; | |
309 | |
310 default: | |
311 gcc_unreachable (); | |
312 } | |
313 } | |
314 | |
315 fprintf (file, "}\n\n"); | |
316 } | |
317 | |
318 /* Display the Reduced Dependence Graph using dotty. */ | |
319 | |
320 DEBUG_FUNCTION void | |
321 dot_rdg (struct graph *rdg) | |
322 { | |
323 /* When debugging, you may want to enable the following code. */ | |
324 #ifdef HAVE_POPEN | |
325 FILE *file = popen ("dot -Tx11", "w"); | |
326 if (!file) | |
327 return; | |
328 dot_rdg_1 (file, rdg); | |
329 fflush (file); | |
330 close (fileno (file)); | |
331 pclose (file); | |
332 #else | |
333 dot_rdg_1 (stderr, rdg); | |
334 #endif | |
335 } | |
336 | |
337 /* Returns the index of STMT in RDG. */ | |
338 | |
339 static int | |
340 rdg_vertex_for_stmt (struct graph *rdg ATTRIBUTE_UNUSED, gimple *stmt) | |
341 { | |
342 int index = gimple_uid (stmt); | |
343 gcc_checking_assert (index == -1 || RDG_STMT (rdg, index) == stmt); | |
344 return index; | |
345 } | |
346 | |
347 /* Creates dependence edges in RDG for all the uses of DEF. IDEF is | |
348 the index of DEF in RDG. */ | |
349 | |
350 static void | |
351 create_rdg_edges_for_scalar (struct graph *rdg, tree def, int idef) | |
352 { | |
353 use_operand_p imm_use_p; | |
354 imm_use_iterator iterator; | |
355 | |
356 FOR_EACH_IMM_USE_FAST (imm_use_p, iterator, def) | |
357 { | |
358 struct graph_edge *e; | |
359 int use = rdg_vertex_for_stmt (rdg, USE_STMT (imm_use_p)); | |
360 | |
361 if (use < 0) | |
362 continue; | |
363 | |
364 e = add_edge (rdg, idef, use); | |
365 e->data = XNEW (struct rdg_edge); | |
366 RDGE_TYPE (e) = flow_dd; | |
367 } | |
368 } | |
369 | |
370 /* Creates an edge for the control dependences of BB to the vertex V. */ | |
371 | |
372 static void | |
373 create_edge_for_control_dependence (struct graph *rdg, basic_block bb, | |
374 int v, control_dependences *cd) | |
375 { | |
376 bitmap_iterator bi; | |
377 unsigned edge_n; | |
378 EXECUTE_IF_SET_IN_BITMAP (cd->get_edges_dependent_on (bb->index), | |
379 0, edge_n, bi) | |
380 { | |
381 basic_block cond_bb = cd->get_edge_src (edge_n); | |
382 gimple *stmt = last_stmt (cond_bb); | |
383 if (stmt && is_ctrl_stmt (stmt)) | |
384 { | |
385 struct graph_edge *e; | |
386 int c = rdg_vertex_for_stmt (rdg, stmt); | |
387 if (c < 0) | |
388 continue; | |
389 | |
390 e = add_edge (rdg, c, v); | |
391 e->data = XNEW (struct rdg_edge); | |
392 RDGE_TYPE (e) = control_dd; | |
393 } | |
394 } | |
395 } | |
396 | |
397 /* Creates the edges of the reduced dependence graph RDG. */ | |
0 | 398 |
399 static void | |
111 | 400 create_rdg_flow_edges (struct graph *rdg) |
0 | 401 { |
111 | 402 int i; |
403 def_operand_p def_p; | |
404 ssa_op_iter iter; | |
405 | |
406 for (i = 0; i < rdg->n_vertices; i++) | |
407 FOR_EACH_PHI_OR_STMT_DEF (def_p, RDG_STMT (rdg, i), | |
408 iter, SSA_OP_DEF) | |
409 create_rdg_edges_for_scalar (rdg, DEF_FROM_PTR (def_p), i); | |
410 } | |
411 | |
412 /* Creates the edges of the reduced dependence graph RDG. */ | |
413 | |
414 static void | |
415 create_rdg_cd_edges (struct graph *rdg, control_dependences *cd, loop_p loop) | |
416 { | |
417 int i; | |
418 | |
419 for (i = 0; i < rdg->n_vertices; i++) | |
0 | 420 { |
111 | 421 gimple *stmt = RDG_STMT (rdg, i); |
422 if (gimple_code (stmt) == GIMPLE_PHI) | |
0 | 423 { |
111 | 424 edge_iterator ei; |
425 edge e; | |
426 FOR_EACH_EDGE (e, ei, gimple_bb (stmt)->preds) | |
427 if (flow_bb_inside_loop_p (loop, e->src)) | |
428 create_edge_for_control_dependence (rdg, e->src, i, cd); | |
0 | 429 } |
430 else | |
111 | 431 create_edge_for_control_dependence (rdg, gimple_bb (stmt), i, cd); |
432 } | |
433 } | |
434 | |
435 /* Build the vertices of the reduced dependence graph RDG. Return false | |
436 if that failed. */ | |
437 | |
438 static bool | |
439 create_rdg_vertices (struct graph *rdg, vec<gimple *> stmts, loop_p loop) | |
440 { | |
441 int i; | |
442 gimple *stmt; | |
443 | |
444 FOR_EACH_VEC_ELT (stmts, i, stmt) | |
445 { | |
446 struct vertex *v = &(rdg->vertices[i]); | |
447 | |
448 /* Record statement to vertex mapping. */ | |
449 gimple_set_uid (stmt, i); | |
450 | |
451 v->data = XNEW (struct rdg_vertex); | |
452 RDGV_STMT (v) = stmt; | |
453 RDGV_DATAREFS (v).create (0); | |
454 RDGV_HAS_MEM_WRITE (v) = false; | |
455 RDGV_HAS_MEM_READS (v) = false; | |
456 if (gimple_code (stmt) == GIMPLE_PHI) | |
457 continue; | |
458 | |
459 unsigned drp = datarefs_vec.length (); | |
460 if (!find_data_references_in_stmt (loop, stmt, &datarefs_vec)) | |
461 return false; | |
462 for (unsigned j = drp; j < datarefs_vec.length (); ++j) | |
463 { | |
464 data_reference_p dr = datarefs_vec[j]; | |
465 if (DR_IS_READ (dr)) | |
466 RDGV_HAS_MEM_READS (v) = true; | |
467 else | |
468 RDGV_HAS_MEM_WRITE (v) = true; | |
469 RDGV_DATAREFS (v).safe_push (dr); | |
470 } | |
471 } | |
472 return true; | |
473 } | |
474 | |
475 /* Array mapping basic block's index to its topological order. */ | |
476 static int *bb_top_order_index; | |
477 /* And size of the array. */ | |
478 static int bb_top_order_index_size; | |
479 | |
480 /* If X has a smaller topological sort number than Y, returns -1; | |
481 if greater, returns 1. */ | |
482 | |
483 static int | |
484 bb_top_order_cmp (const void *x, const void *y) | |
485 { | |
486 basic_block bb1 = *(const basic_block *) x; | |
487 basic_block bb2 = *(const basic_block *) y; | |
488 | |
489 gcc_assert (bb1->index < bb_top_order_index_size | |
490 && bb2->index < bb_top_order_index_size); | |
491 gcc_assert (bb1 == bb2 | |
492 || bb_top_order_index[bb1->index] | |
493 != bb_top_order_index[bb2->index]); | |
494 | |
495 return (bb_top_order_index[bb1->index] - bb_top_order_index[bb2->index]); | |
496 } | |
497 | |
498 /* Initialize STMTS with all the statements of LOOP. We use topological | |
499 order to discover all statements. The order is important because | |
500 generate_loops_for_partition is using the same traversal for identifying | |
501 statements in loop copies. */ | |
502 | |
503 static void | |
504 stmts_from_loop (struct loop *loop, vec<gimple *> *stmts) | |
505 { | |
506 unsigned int i; | |
507 basic_block *bbs = get_loop_body_in_custom_order (loop, bb_top_order_cmp); | |
508 | |
509 for (i = 0; i < loop->num_nodes; i++) | |
510 { | |
511 basic_block bb = bbs[i]; | |
512 | |
513 for (gphi_iterator bsi = gsi_start_phis (bb); !gsi_end_p (bsi); | |
514 gsi_next (&bsi)) | |
515 if (!virtual_operand_p (gimple_phi_result (bsi.phi ()))) | |
516 stmts->safe_push (bsi.phi ()); | |
517 | |
518 for (gimple_stmt_iterator bsi = gsi_start_bb (bb); !gsi_end_p (bsi); | |
519 gsi_next (&bsi)) | |
520 { | |
521 gimple *stmt = gsi_stmt (bsi); | |
522 if (gimple_code (stmt) != GIMPLE_LABEL && !is_gimple_debug (stmt)) | |
523 stmts->safe_push (stmt); | |
524 } | |
525 } | |
526 | |
527 free (bbs); | |
528 } | |
529 | |
530 /* Free the reduced dependence graph RDG. */ | |
531 | |
532 static void | |
533 free_rdg (struct graph *rdg) | |
534 { | |
535 int i; | |
536 | |
537 for (i = 0; i < rdg->n_vertices; i++) | |
538 { | |
539 struct vertex *v = &(rdg->vertices[i]); | |
540 struct graph_edge *e; | |
541 | |
542 for (e = v->succ; e; e = e->succ_next) | |
543 free (e->data); | |
544 | |
545 if (v->data) | |
546 { | |
547 gimple_set_uid (RDGV_STMT (v), -1); | |
548 (RDGV_DATAREFS (v)).release (); | |
549 free (v->data); | |
550 } | |
551 } | |
552 | |
553 free_graph (rdg); | |
554 } | |
555 | |
556 /* Build the Reduced Dependence Graph (RDG) with one vertex per statement of | |
557 LOOP, and one edge per flow dependence or control dependence from control | |
558 dependence CD. During visiting each statement, data references are also | |
559 collected and recorded in global data DATAREFS_VEC. */ | |
560 | |
561 static struct graph * | |
562 build_rdg (struct loop *loop, control_dependences *cd) | |
563 { | |
564 struct graph *rdg; | |
565 | |
566 /* Create the RDG vertices from the stmts of the loop nest. */ | |
567 auto_vec<gimple *, 10> stmts; | |
568 stmts_from_loop (loop, &stmts); | |
569 rdg = new_graph (stmts.length ()); | |
570 if (!create_rdg_vertices (rdg, stmts, loop)) | |
571 { | |
572 free_rdg (rdg); | |
573 return NULL; | |
0 | 574 } |
111 | 575 stmts.release (); |
576 | |
577 create_rdg_flow_edges (rdg); | |
578 if (cd) | |
579 create_rdg_cd_edges (rdg, cd, loop); | |
580 | |
581 return rdg; | |
582 } | |
583 | |
584 | |
585 /* Kind of distributed loop. */ | |
586 enum partition_kind { | |
587 PKIND_NORMAL, PKIND_MEMSET, PKIND_MEMCPY, PKIND_MEMMOVE | |
588 }; | |
589 | |
590 /* Type of distributed loop. */ | |
591 enum partition_type { | |
592 /* The distributed loop can be executed parallelly. */ | |
593 PTYPE_PARALLEL = 0, | |
594 /* The distributed loop has to be executed sequentially. */ | |
595 PTYPE_SEQUENTIAL | |
596 }; | |
597 | |
598 /* Builtin info for loop distribution. */ | |
599 struct builtin_info | |
600 { | |
601 /* data-references a kind != PKIND_NORMAL partition is about. */ | |
602 data_reference_p dst_dr; | |
603 data_reference_p src_dr; | |
604 /* Base address and size of memory objects operated by the builtin. Note | |
605 both dest and source memory objects must have the same size. */ | |
606 tree dst_base; | |
607 tree src_base; | |
608 tree size; | |
609 /* Base and offset part of dst_base after stripping constant offset. This | |
610 is only used in memset builtin distribution for now. */ | |
611 tree dst_base_base; | |
612 unsigned HOST_WIDE_INT dst_base_offset; | |
613 }; | |
614 | |
615 /* Partition for loop distribution. */ | |
616 struct partition | |
617 { | |
618 /* Statements of the partition. */ | |
619 bitmap stmts; | |
620 /* True if the partition defines variable which is used outside of loop. */ | |
621 bool reduction_p; | |
622 enum partition_kind kind; | |
623 enum partition_type type; | |
624 /* Data references in the partition. */ | |
625 bitmap datarefs; | |
626 /* Information of builtin parition. */ | |
627 struct builtin_info *builtin; | |
628 }; | |
629 | |
630 | |
631 /* Allocate and initialize a partition from BITMAP. */ | |
632 | |
633 static partition * | |
634 partition_alloc (void) | |
635 { | |
636 partition *partition = XCNEW (struct partition); | |
637 partition->stmts = BITMAP_ALLOC (NULL); | |
638 partition->reduction_p = false; | |
639 partition->kind = PKIND_NORMAL; | |
640 partition->datarefs = BITMAP_ALLOC (NULL); | |
641 return partition; | |
642 } | |
643 | |
644 /* Free PARTITION. */ | |
645 | |
646 static void | |
647 partition_free (partition *partition) | |
648 { | |
649 BITMAP_FREE (partition->stmts); | |
650 BITMAP_FREE (partition->datarefs); | |
651 if (partition->builtin) | |
652 free (partition->builtin); | |
653 | |
654 free (partition); | |
655 } | |
656 | |
657 /* Returns true if the partition can be generated as a builtin. */ | |
658 | |
659 static bool | |
660 partition_builtin_p (partition *partition) | |
661 { | |
662 return partition->kind != PKIND_NORMAL; | |
663 } | |
664 | |
665 /* Returns true if the partition contains a reduction. */ | |
666 | |
667 static bool | |
668 partition_reduction_p (partition *partition) | |
669 { | |
670 return partition->reduction_p; | |
671 } | |
672 | |
673 /* Partitions are fused because of different reasons. */ | |
674 enum fuse_type | |
675 { | |
676 FUSE_NON_BUILTIN = 0, | |
677 FUSE_REDUCTION = 1, | |
678 FUSE_SHARE_REF = 2, | |
679 FUSE_SAME_SCC = 3, | |
680 FUSE_FINALIZE = 4 | |
681 }; | |
682 | |
683 /* Description on different fusing reason. */ | |
684 static const char *fuse_message[] = { | |
685 "they are non-builtins", | |
686 "they have reductions", | |
687 "they have shared memory refs", | |
688 "they are in the same dependence scc", | |
689 "there is no point to distribute loop"}; | |
690 | |
691 static void | |
692 update_type_for_merge (struct graph *, partition *, partition *); | |
693 | |
694 /* Merge PARTITION into the partition DEST. RDG is the reduced dependence | |
695 graph and we update type for result partition if it is non-NULL. */ | |
696 | |
697 static void | |
698 partition_merge_into (struct graph *rdg, partition *dest, | |
699 partition *partition, enum fuse_type ft) | |
700 { | |
701 if (dump_file && (dump_flags & TDF_DETAILS)) | |
702 { | |
703 fprintf (dump_file, "Fuse partitions because %s:\n", fuse_message[ft]); | |
704 fprintf (dump_file, " Part 1: "); | |
705 dump_bitmap (dump_file, dest->stmts); | |
706 fprintf (dump_file, " Part 2: "); | |
707 dump_bitmap (dump_file, partition->stmts); | |
708 } | |
709 | |
710 dest->kind = PKIND_NORMAL; | |
711 if (dest->type == PTYPE_PARALLEL) | |
712 dest->type = partition->type; | |
713 | |
714 bitmap_ior_into (dest->stmts, partition->stmts); | |
715 if (partition_reduction_p (partition)) | |
716 dest->reduction_p = true; | |
717 | |
718 /* Further check if any data dependence prevents us from executing the | |
719 new partition parallelly. */ | |
720 if (dest->type == PTYPE_PARALLEL && rdg != NULL) | |
721 update_type_for_merge (rdg, dest, partition); | |
722 | |
723 bitmap_ior_into (dest->datarefs, partition->datarefs); | |
724 } | |
725 | |
726 | |
727 /* Returns true when DEF is an SSA_NAME defined in LOOP and used after | |
728 the LOOP. */ | |
729 | |
730 static bool | |
731 ssa_name_has_uses_outside_loop_p (tree def, loop_p loop) | |
732 { | |
733 imm_use_iterator imm_iter; | |
734 use_operand_p use_p; | |
735 | |
736 FOR_EACH_IMM_USE_FAST (use_p, imm_iter, def) | |
737 { | |
738 if (is_gimple_debug (USE_STMT (use_p))) | |
739 continue; | |
740 | |
741 basic_block use_bb = gimple_bb (USE_STMT (use_p)); | |
742 if (!flow_bb_inside_loop_p (loop, use_bb)) | |
743 return true; | |
744 } | |
745 | |
746 return false; | |
747 } | |
748 | |
749 /* Returns true when STMT defines a scalar variable used after the | |
750 loop LOOP. */ | |
751 | |
752 static bool | |
753 stmt_has_scalar_dependences_outside_loop (loop_p loop, gimple *stmt) | |
754 { | |
755 def_operand_p def_p; | |
756 ssa_op_iter op_iter; | |
757 | |
758 if (gimple_code (stmt) == GIMPLE_PHI) | |
759 return ssa_name_has_uses_outside_loop_p (gimple_phi_result (stmt), loop); | |
760 | |
761 FOR_EACH_SSA_DEF_OPERAND (def_p, stmt, op_iter, SSA_OP_DEF) | |
762 if (ssa_name_has_uses_outside_loop_p (DEF_FROM_PTR (def_p), loop)) | |
763 return true; | |
764 | |
765 return false; | |
0 | 766 } |
767 | |
768 /* Return a copy of LOOP placed before LOOP. */ | |
769 | |
770 static struct loop * | |
771 copy_loop_before (struct loop *loop) | |
772 { | |
773 struct loop *res; | |
774 edge preheader = loop_preheader_edge (loop); | |
775 | |
776 initialize_original_copy_tables (); | |
111 | 777 res = slpeel_tree_duplicate_loop_to_edge_cfg (loop, NULL, preheader); |
778 gcc_assert (res != NULL); | |
0 | 779 free_original_copy_tables (); |
111 | 780 delete_update_ssa (); |
0 | 781 |
782 return res; | |
783 } | |
784 | |
785 /* Creates an empty basic block after LOOP. */ | |
786 | |
787 static void | |
788 create_bb_after_loop (struct loop *loop) | |
789 { | |
790 edge exit = single_exit (loop); | |
791 | |
792 if (!exit) | |
793 return; | |
794 | |
795 split_edge (exit); | |
796 } | |
797 | |
798 /* Generate code for PARTITION from the code in LOOP. The loop is | |
799 copied when COPY_P is true. All the statements not flagged in the | |
800 PARTITION bitmap are removed from the loop or from its copy. The | |
801 statements are indexed in sequence inside a basic block, and the | |
111 | 802 basic blocks of a loop are taken in dom order. */ |
803 | |
804 static void | |
805 generate_loops_for_partition (struct loop *loop, partition *partition, | |
806 bool copy_p) | |
0 | 807 { |
111 | 808 unsigned i; |
0 | 809 basic_block *bbs; |
810 | |
811 if (copy_p) | |
812 { | |
111 | 813 int orig_loop_num = loop->orig_loop_num; |
0 | 814 loop = copy_loop_before (loop); |
111 | 815 gcc_assert (loop != NULL); |
816 loop->orig_loop_num = orig_loop_num; | |
0 | 817 create_preheader (loop, CP_SIMPLE_PREHEADERS); |
818 create_bb_after_loop (loop); | |
819 } | |
111 | 820 else |
0 | 821 { |
111 | 822 /* Origin number is set to the new versioned loop's num. */ |
823 gcc_assert (loop->orig_loop_num != loop->num); | |
0 | 824 } |
825 | |
111 | 826 /* Remove stmts not in the PARTITION bitmap. */ |
0 | 827 bbs = get_loop_body_in_dom_order (loop); |
828 | |
111 | 829 if (MAY_HAVE_DEBUG_STMTS) |
830 for (i = 0; i < loop->num_nodes; i++) | |
831 { | |
832 basic_block bb = bbs[i]; | |
833 | |
834 for (gphi_iterator bsi = gsi_start_phis (bb); !gsi_end_p (bsi); | |
835 gsi_next (&bsi)) | |
836 { | |
837 gphi *phi = bsi.phi (); | |
838 if (!virtual_operand_p (gimple_phi_result (phi)) | |
839 && !bitmap_bit_p (partition->stmts, gimple_uid (phi))) | |
840 reset_debug_uses (phi); | |
841 } | |
842 | |
843 for (gimple_stmt_iterator bsi = gsi_start_bb (bb); !gsi_end_p (bsi); gsi_next (&bsi)) | |
844 { | |
845 gimple *stmt = gsi_stmt (bsi); | |
846 if (gimple_code (stmt) != GIMPLE_LABEL | |
847 && !is_gimple_debug (stmt) | |
848 && !bitmap_bit_p (partition->stmts, gimple_uid (stmt))) | |
849 reset_debug_uses (stmt); | |
850 } | |
851 } | |
852 | |
0 | 853 for (i = 0; i < loop->num_nodes; i++) |
854 { | |
855 basic_block bb = bbs[i]; | |
111 | 856 edge inner_exit = NULL; |
857 | |
858 if (loop != bb->loop_father) | |
859 inner_exit = single_exit (bb->loop_father); | |
860 | |
861 for (gphi_iterator bsi = gsi_start_phis (bb); !gsi_end_p (bsi);) | |
862 { | |
863 gphi *phi = bsi.phi (); | |
864 if (!virtual_operand_p (gimple_phi_result (phi)) | |
865 && !bitmap_bit_p (partition->stmts, gimple_uid (phi))) | |
866 remove_phi_node (&bsi, true); | |
867 else | |
868 gsi_next (&bsi); | |
869 } | |
870 | |
871 for (gimple_stmt_iterator bsi = gsi_start_bb (bb); !gsi_end_p (bsi);) | |
0 | 872 { |
111 | 873 gimple *stmt = gsi_stmt (bsi); |
874 if (gimple_code (stmt) != GIMPLE_LABEL | |
875 && !is_gimple_debug (stmt) | |
876 && !bitmap_bit_p (partition->stmts, gimple_uid (stmt))) | |
0 | 877 { |
111 | 878 /* In distribution of loop nest, if bb is inner loop's exit_bb, |
879 we choose its exit edge/path in order to avoid generating | |
880 infinite loop. For all other cases, we choose an arbitrary | |
881 path through the empty CFG part that this unnecessary | |
882 control stmt controls. */ | |
883 if (gcond *cond_stmt = dyn_cast <gcond *> (stmt)) | |
884 { | |
885 if (inner_exit && inner_exit->flags & EDGE_TRUE_VALUE) | |
886 gimple_cond_make_true (cond_stmt); | |
887 else | |
888 gimple_cond_make_false (cond_stmt); | |
889 update_stmt (stmt); | |
890 } | |
891 else if (gimple_code (stmt) == GIMPLE_SWITCH) | |
892 { | |
893 gswitch *switch_stmt = as_a <gswitch *> (stmt); | |
894 gimple_switch_set_index | |
895 (switch_stmt, CASE_LOW (gimple_switch_label (switch_stmt, 1))); | |
896 update_stmt (stmt); | |
897 } | |
898 else | |
899 { | |
900 unlink_stmt_vdef (stmt); | |
901 gsi_remove (&bsi, true); | |
902 release_defs (stmt); | |
903 continue; | |
904 } | |
0 | 905 } |
111 | 906 gsi_next (&bsi); |
0 | 907 } |
908 } | |
909 | |
111 | 910 free (bbs); |
911 } | |
912 | |
913 /* If VAL memory representation contains the same value in all bytes, | |
914 return that value, otherwise return -1. | |
915 E.g. for 0x24242424 return 0x24, for IEEE double | |
916 747708026454360457216.0 return 0x44, etc. */ | |
917 | |
918 static int | |
919 const_with_all_bytes_same (tree val) | |
920 { | |
921 unsigned char buf[64]; | |
922 int i, len; | |
923 | |
924 if (integer_zerop (val) | |
925 || (TREE_CODE (val) == CONSTRUCTOR | |
926 && !TREE_CLOBBER_P (val) | |
927 && CONSTRUCTOR_NELTS (val) == 0)) | |
928 return 0; | |
929 | |
930 if (real_zerop (val)) | |
931 { | |
932 /* Only return 0 for +0.0, not for -0.0, which doesn't have | |
933 an all bytes same memory representation. Don't transform | |
934 -0.0 stores into +0.0 even for !HONOR_SIGNED_ZEROS. */ | |
935 switch (TREE_CODE (val)) | |
936 { | |
937 case REAL_CST: | |
938 if (!real_isneg (TREE_REAL_CST_PTR (val))) | |
939 return 0; | |
940 break; | |
941 case COMPLEX_CST: | |
942 if (!const_with_all_bytes_same (TREE_REALPART (val)) | |
943 && !const_with_all_bytes_same (TREE_IMAGPART (val))) | |
944 return 0; | |
945 break; | |
946 case VECTOR_CST: | |
947 unsigned int j; | |
948 for (j = 0; j < VECTOR_CST_NELTS (val); ++j) | |
949 if (const_with_all_bytes_same (VECTOR_CST_ELT (val, j))) | |
950 break; | |
951 if (j == VECTOR_CST_NELTS (val)) | |
952 return 0; | |
953 break; | |
954 default: | |
955 break; | |
956 } | |
957 } | |
958 | |
959 if (CHAR_BIT != 8 || BITS_PER_UNIT != 8) | |
960 return -1; | |
961 | |
962 len = native_encode_expr (val, buf, sizeof (buf)); | |
963 if (len == 0) | |
964 return -1; | |
965 for (i = 1; i < len; i++) | |
966 if (buf[i] != buf[0]) | |
967 return -1; | |
968 return buf[0]; | |
969 } | |
970 | |
971 /* Generate a call to memset for PARTITION in LOOP. */ | |
972 | |
973 static void | |
974 generate_memset_builtin (struct loop *loop, partition *partition) | |
975 { | |
976 gimple_stmt_iterator gsi; | |
977 tree mem, fn, nb_bytes; | |
978 tree val; | |
979 struct builtin_info *builtin = partition->builtin; | |
980 gimple *fn_call; | |
0 | 981 |
982 /* The new statements will be placed before LOOP. */ | |
111 | 983 gsi = gsi_last_bb (loop_preheader_edge (loop)->src); |
984 | |
985 nb_bytes = builtin->size; | |
986 nb_bytes = force_gimple_operand_gsi (&gsi, nb_bytes, true, NULL_TREE, | |
987 false, GSI_CONTINUE_LINKING); | |
988 mem = builtin->dst_base; | |
989 mem = force_gimple_operand_gsi (&gsi, mem, true, NULL_TREE, | |
990 false, GSI_CONTINUE_LINKING); | |
991 | |
992 /* This exactly matches the pattern recognition in classify_partition. */ | |
993 val = gimple_assign_rhs1 (DR_STMT (builtin->dst_dr)); | |
994 /* Handle constants like 0x15151515 and similarly | |
995 floating point constants etc. where all bytes are the same. */ | |
996 int bytev = const_with_all_bytes_same (val); | |
997 if (bytev != -1) | |
998 val = build_int_cst (integer_type_node, bytev); | |
999 else if (TREE_CODE (val) == INTEGER_CST) | |
1000 val = fold_convert (integer_type_node, val); | |
1001 else if (!useless_type_conversion_p (integer_type_node, TREE_TYPE (val))) | |
1002 { | |
1003 tree tem = make_ssa_name (integer_type_node); | |
1004 gimple *cstmt = gimple_build_assign (tem, NOP_EXPR, val); | |
1005 gsi_insert_after (&gsi, cstmt, GSI_CONTINUE_LINKING); | |
1006 val = tem; | |
1007 } | |
1008 | |
1009 fn = build_fold_addr_expr (builtin_decl_implicit (BUILT_IN_MEMSET)); | |
1010 fn_call = gimple_build_call (fn, 3, mem, val, nb_bytes); | |
1011 gsi_insert_after (&gsi, fn_call, GSI_CONTINUE_LINKING); | |
1012 | |
1013 if (dump_file && (dump_flags & TDF_DETAILS)) | |
0 | 1014 { |
111 | 1015 fprintf (dump_file, "generated memset"); |
1016 if (bytev == 0) | |
1017 fprintf (dump_file, " zero\n"); | |
1018 else | |
1019 fprintf (dump_file, "\n"); | |
1020 } | |
1021 } | |
1022 | |
1023 /* Generate a call to memcpy for PARTITION in LOOP. */ | |
1024 | |
1025 static void | |
1026 generate_memcpy_builtin (struct loop *loop, partition *partition) | |
1027 { | |
1028 gimple_stmt_iterator gsi; | |
1029 gimple *fn_call; | |
1030 tree dest, src, fn, nb_bytes; | |
1031 enum built_in_function kind; | |
1032 struct builtin_info *builtin = partition->builtin; | |
1033 | |
1034 /* The new statements will be placed before LOOP. */ | |
1035 gsi = gsi_last_bb (loop_preheader_edge (loop)->src); | |
1036 | |
1037 nb_bytes = builtin->size; | |
1038 nb_bytes = force_gimple_operand_gsi (&gsi, nb_bytes, true, NULL_TREE, | |
1039 false, GSI_CONTINUE_LINKING); | |
1040 dest = builtin->dst_base; | |
1041 src = builtin->src_base; | |
1042 if (partition->kind == PKIND_MEMCPY | |
1043 || ! ptr_derefs_may_alias_p (dest, src)) | |
1044 kind = BUILT_IN_MEMCPY; | |
1045 else | |
1046 kind = BUILT_IN_MEMMOVE; | |
1047 | |
1048 dest = force_gimple_operand_gsi (&gsi, dest, true, NULL_TREE, | |
1049 false, GSI_CONTINUE_LINKING); | |
1050 src = force_gimple_operand_gsi (&gsi, src, true, NULL_TREE, | |
1051 false, GSI_CONTINUE_LINKING); | |
1052 fn = build_fold_addr_expr (builtin_decl_implicit (kind)); | |
1053 fn_call = gimple_build_call (fn, 3, dest, src, nb_bytes); | |
1054 gsi_insert_after (&gsi, fn_call, GSI_CONTINUE_LINKING); | |
1055 | |
1056 if (dump_file && (dump_flags & TDF_DETAILS)) | |
1057 { | |
1058 if (kind == BUILT_IN_MEMCPY) | |
1059 fprintf (dump_file, "generated memcpy\n"); | |
1060 else | |
1061 fprintf (dump_file, "generated memmove\n"); | |
0 | 1062 } |
1063 } | |
1064 | |
111 | 1065 /* Remove and destroy the loop LOOP. */ |
1066 | |
1067 static void | |
1068 destroy_loop (struct loop *loop) | |
0 | 1069 { |
111 | 1070 unsigned nbbs = loop->num_nodes; |
1071 edge exit = single_exit (loop); | |
1072 basic_block src = loop_preheader_edge (loop)->src, dest = exit->dest; | |
1073 basic_block *bbs; | |
1074 unsigned i; | |
1075 | |
1076 bbs = get_loop_body_in_dom_order (loop); | |
1077 | |
1078 redirect_edge_pred (exit, src); | |
1079 exit->flags &= ~(EDGE_TRUE_VALUE|EDGE_FALSE_VALUE); | |
1080 exit->flags |= EDGE_FALLTHRU; | |
1081 cancel_loop_tree (loop); | |
1082 rescan_loop_exit (exit, false, true); | |
1083 | |
1084 i = nbbs; | |
1085 do | |
1086 { | |
1087 /* We have made sure to not leave any dangling uses of SSA | |
1088 names defined in the loop. With the exception of virtuals. | |
1089 Make sure we replace all uses of virtual defs that will remain | |
1090 outside of the loop with the bare symbol as delete_basic_block | |
1091 will release them. */ | |
1092 --i; | |
1093 for (gphi_iterator gsi = gsi_start_phis (bbs[i]); !gsi_end_p (gsi); | |
1094 gsi_next (&gsi)) | |
1095 { | |
1096 gphi *phi = gsi.phi (); | |
1097 if (virtual_operand_p (gimple_phi_result (phi))) | |
1098 mark_virtual_phi_result_for_renaming (phi); | |
1099 } | |
1100 for (gimple_stmt_iterator gsi = gsi_start_bb (bbs[i]); !gsi_end_p (gsi); | |
1101 gsi_next (&gsi)) | |
1102 { | |
1103 gimple *stmt = gsi_stmt (gsi); | |
1104 tree vdef = gimple_vdef (stmt); | |
1105 if (vdef && TREE_CODE (vdef) == SSA_NAME) | |
1106 mark_virtual_operand_for_renaming (vdef); | |
1107 } | |
1108 delete_basic_block (bbs[i]); | |
1109 } | |
1110 while (i != 0); | |
1111 | |
1112 free (bbs); | |
1113 | |
1114 set_immediate_dominator (CDI_DOMINATORS, dest, | |
1115 recompute_dominator (CDI_DOMINATORS, dest)); | |
1116 } | |
1117 | |
1118 /* Generates code for PARTITION. Return whether LOOP needs to be destroyed. */ | |
1119 | |
1120 static bool | |
1121 generate_code_for_partition (struct loop *loop, | |
1122 partition *partition, bool copy_p) | |
1123 { | |
1124 switch (partition->kind) | |
1125 { | |
1126 case PKIND_NORMAL: | |
1127 /* Reductions all have to be in the last partition. */ | |
1128 gcc_assert (!partition_reduction_p (partition) | |
1129 || !copy_p); | |
1130 generate_loops_for_partition (loop, partition, copy_p); | |
1131 return false; | |
1132 | |
1133 case PKIND_MEMSET: | |
1134 generate_memset_builtin (loop, partition); | |
1135 break; | |
1136 | |
1137 case PKIND_MEMCPY: | |
1138 case PKIND_MEMMOVE: | |
1139 generate_memcpy_builtin (loop, partition); | |
1140 break; | |
1141 | |
1142 default: | |
1143 gcc_unreachable (); | |
1144 } | |
1145 | |
1146 /* Common tail for partitions we turn into a call. If this was the last | |
1147 partition for which we generate code, we have to destroy the loop. */ | |
1148 if (!copy_p) | |
0 | 1149 return true; |
1150 return false; | |
1151 } | |
1152 | |
111 | 1153 /* Return data dependence relation for data references A and B. The two |
1154 data references must be in lexicographic order wrto reduced dependence | |
1155 graph RDG. We firstly try to find ddr from global ddr hash table. If | |
1156 it doesn't exist, compute the ddr and cache it. */ | |
1157 | |
1158 static data_dependence_relation * | |
1159 get_data_dependence (struct graph *rdg, data_reference_p a, data_reference_p b) | |
0 | 1160 { |
111 | 1161 struct data_dependence_relation ent, **slot; |
1162 struct data_dependence_relation *ddr; | |
1163 | |
1164 gcc_assert (DR_IS_WRITE (a) || DR_IS_WRITE (b)); | |
1165 gcc_assert (rdg_vertex_for_stmt (rdg, DR_STMT (a)) | |
1166 <= rdg_vertex_for_stmt (rdg, DR_STMT (b))); | |
1167 ent.a = a; | |
1168 ent.b = b; | |
1169 slot = ddrs_table->find_slot (&ent, INSERT); | |
1170 if (*slot == NULL) | |
1171 { | |
1172 ddr = initialize_data_dependence_relation (a, b, loop_nest); | |
1173 compute_affine_dependence (ddr, loop_nest[0]); | |
1174 *slot = ddr; | |
1175 } | |
1176 | |
1177 return *slot; | |
0 | 1178 } |
1179 | |
111 | 1180 /* In reduced dependence graph RDG for loop distribution, return true if |
1181 dependence between references DR1 and DR2 leads to a dependence cycle | |
1182 and such dependence cycle can't be resolved by runtime alias check. */ | |
0 | 1183 |
1184 static bool | |
111 | 1185 data_dep_in_cycle_p (struct graph *rdg, |
1186 data_reference_p dr1, data_reference_p dr2) | |
0 | 1187 { |
111 | 1188 struct data_dependence_relation *ddr; |
1189 | |
1190 /* Re-shuffle data-refs to be in topological order. */ | |
1191 if (rdg_vertex_for_stmt (rdg, DR_STMT (dr1)) | |
1192 > rdg_vertex_for_stmt (rdg, DR_STMT (dr2))) | |
1193 std::swap (dr1, dr2); | |
1194 | |
1195 ddr = get_data_dependence (rdg, dr1, dr2); | |
1196 | |
1197 /* In case of no data dependence. */ | |
1198 if (DDR_ARE_DEPENDENT (ddr) == chrec_known) | |
1199 return false; | |
1200 /* For unknown data dependence or known data dependence which can't be | |
1201 expressed in classic distance vector, we check if it can be resolved | |
1202 by runtime alias check. If yes, we still consider data dependence | |
1203 as won't introduce data dependence cycle. */ | |
1204 else if (DDR_ARE_DEPENDENT (ddr) == chrec_dont_know | |
1205 || DDR_NUM_DIST_VECTS (ddr) == 0) | |
1206 return !runtime_alias_check_p (ddr, NULL, true); | |
1207 else if (DDR_NUM_DIST_VECTS (ddr) > 1) | |
1208 return true; | |
1209 else if (DDR_REVERSED_P (ddr) | |
1210 || lambda_vector_zerop (DDR_DIST_VECT (ddr, 0), 1)) | |
1211 return false; | |
1212 | |
1213 return true; | |
0 | 1214 } |
1215 | |
111 | 1216 /* Given reduced dependence graph RDG, PARTITION1 and PARTITION2, update |
1217 PARTITION1's type after merging PARTITION2 into PARTITION1. */ | |
0 | 1218 |
1219 static void | |
111 | 1220 update_type_for_merge (struct graph *rdg, |
1221 partition *partition1, partition *partition2) | |
0 | 1222 { |
111 | 1223 unsigned i, j; |
1224 bitmap_iterator bi, bj; | |
1225 data_reference_p dr1, dr2; | |
1226 | |
1227 EXECUTE_IF_SET_IN_BITMAP (partition1->datarefs, 0, i, bi) | |
0 | 1228 { |
111 | 1229 unsigned start = (partition1 == partition2) ? i + 1 : 0; |
1230 | |
1231 dr1 = datarefs_vec[i]; | |
1232 EXECUTE_IF_SET_IN_BITMAP (partition2->datarefs, start, j, bj) | |
0 | 1233 { |
111 | 1234 dr2 = datarefs_vec[j]; |
1235 if (DR_IS_READ (dr1) && DR_IS_READ (dr2)) | |
1236 continue; | |
1237 | |
1238 /* Partition can only be executed sequentially if there is any | |
1239 data dependence cycle. */ | |
1240 if (data_dep_in_cycle_p (rdg, dr1, dr2)) | |
0 | 1241 { |
111 | 1242 partition1->type = PTYPE_SEQUENTIAL; |
1243 return; | |
0 | 1244 } |
1245 } | |
1246 } | |
1247 } | |
1248 | |
111 | 1249 /* Returns a partition with all the statements needed for computing |
1250 the vertex V of the RDG, also including the loop exit conditions. */ | |
1251 | |
1252 static partition * | |
1253 build_rdg_partition_for_vertex (struct graph *rdg, int v) | |
0 | 1254 { |
111 | 1255 partition *partition = partition_alloc (); |
1256 auto_vec<int, 3> nodes; | |
1257 unsigned i, j; | |
0 | 1258 int x; |
111 | 1259 data_reference_p dr; |
1260 | |
1261 graphds_dfs (rdg, &v, 1, &nodes, false, NULL); | |
1262 | |
1263 FOR_EACH_VEC_ELT (nodes, i, x) | |
1264 { | |
1265 bitmap_set_bit (partition->stmts, x); | |
1266 | |
1267 for (j = 0; RDG_DATAREFS (rdg, x).iterate (j, &dr); ++j) | |
1268 { | |
1269 unsigned idx = (unsigned) DR_INDEX (dr); | |
1270 gcc_assert (idx < datarefs_vec.length ()); | |
1271 | |
1272 /* Partition can only be executed sequentially if there is any | |
1273 unknown data reference. */ | |
1274 if (!DR_BASE_ADDRESS (dr) || !DR_OFFSET (dr) | |
1275 || !DR_INIT (dr) || !DR_STEP (dr)) | |
1276 partition->type = PTYPE_SEQUENTIAL; | |
1277 | |
1278 bitmap_set_bit (partition->datarefs, idx); | |
1279 } | |
1280 } | |
1281 | |
1282 if (partition->type == PTYPE_SEQUENTIAL) | |
1283 return partition; | |
1284 | |
1285 /* Further check if any data dependence prevents us from executing the | |
1286 partition parallelly. */ | |
1287 update_type_for_merge (rdg, partition, partition); | |
1288 | |
1289 return partition; | |
0 | 1290 } |
1291 | |
111 | 1292 /* Given PARTITION of LOOP and RDG, record single load/store data references |
1293 for builtin partition in SRC_DR/DST_DR, return false if there is no such | |
1294 data references. */ | |
1295 | |
1296 static bool | |
1297 find_single_drs (struct loop *loop, struct graph *rdg, partition *partition, | |
1298 data_reference_p *dst_dr, data_reference_p *src_dr) | |
0 | 1299 { |
1300 unsigned i; | |
111 | 1301 data_reference_p single_ld = NULL, single_st = NULL; |
0 | 1302 bitmap_iterator bi; |
111 | 1303 |
1304 EXECUTE_IF_SET_IN_BITMAP (partition->stmts, 0, i, bi) | |
0 | 1305 { |
111 | 1306 gimple *stmt = RDG_STMT (rdg, i); |
1307 data_reference_p dr; | |
1308 | |
1309 if (gimple_code (stmt) == GIMPLE_PHI) | |
1310 continue; | |
1311 | |
1312 /* Any scalar stmts are ok. */ | |
1313 if (!gimple_vuse (stmt)) | |
1314 continue; | |
1315 | |
1316 /* Otherwise just regular loads/stores. */ | |
1317 if (!gimple_assign_single_p (stmt)) | |
1318 return false; | |
1319 | |
1320 /* But exactly one store and/or load. */ | |
1321 for (unsigned j = 0; RDG_DATAREFS (rdg, i).iterate (j, &dr); ++j) | |
1322 { | |
1323 tree type = TREE_TYPE (DR_REF (dr)); | |
1324 | |
1325 /* The memset, memcpy and memmove library calls are only | |
1326 able to deal with generic address space. */ | |
1327 if (!ADDR_SPACE_GENERIC_P (TYPE_ADDR_SPACE (type))) | |
1328 return false; | |
1329 | |
1330 if (DR_IS_READ (dr)) | |
1331 { | |
1332 if (single_ld != NULL) | |
1333 return false; | |
1334 single_ld = dr; | |
1335 } | |
1336 else | |
1337 { | |
1338 if (single_st != NULL) | |
1339 return false; | |
1340 single_st = dr; | |
1341 } | |
1342 } | |
0 | 1343 } |
1344 | |
111 | 1345 if (!single_st) |
1346 return false; | |
1347 | |
1348 /* Bail out if this is a bitfield memory reference. */ | |
1349 if (TREE_CODE (DR_REF (single_st)) == COMPONENT_REF | |
1350 && DECL_BIT_FIELD (TREE_OPERAND (DR_REF (single_st), 1))) | |
1351 return false; | |
1352 | |
1353 /* Data reference must be executed exactly once per iteration of each | |
1354 loop in the loop nest. We only need to check dominance information | |
1355 against the outermost one in a perfect loop nest because a bb can't | |
1356 dominate outermost loop's latch without dominating inner loop's. */ | |
1357 basic_block bb_st = gimple_bb (DR_STMT (single_st)); | |
1358 if (!dominated_by_p (CDI_DOMINATORS, loop->latch, bb_st)) | |
1359 return false; | |
1360 | |
1361 if (single_ld) | |
1362 { | |
1363 gimple *store = DR_STMT (single_st), *load = DR_STMT (single_ld); | |
1364 /* Direct aggregate copy or via an SSA name temporary. */ | |
1365 if (load != store | |
1366 && gimple_assign_lhs (load) != gimple_assign_rhs1 (store)) | |
1367 return false; | |
1368 | |
1369 /* Bail out if this is a bitfield memory reference. */ | |
1370 if (TREE_CODE (DR_REF (single_ld)) == COMPONENT_REF | |
1371 && DECL_BIT_FIELD (TREE_OPERAND (DR_REF (single_ld), 1))) | |
1372 return false; | |
1373 | |
1374 /* Load and store must be in the same loop nest. */ | |
1375 basic_block bb_ld = gimple_bb (DR_STMT (single_ld)); | |
1376 if (bb_st->loop_father != bb_ld->loop_father) | |
1377 return false; | |
1378 | |
1379 /* Data reference must be executed exactly once per iteration. | |
1380 Same as single_st, we only need to check against the outermost | |
1381 loop. */ | |
1382 if (!dominated_by_p (CDI_DOMINATORS, loop->latch, bb_ld)) | |
1383 return false; | |
1384 | |
1385 edge e = single_exit (bb_st->loop_father); | |
1386 bool dom_ld = dominated_by_p (CDI_DOMINATORS, e->src, bb_ld); | |
1387 bool dom_st = dominated_by_p (CDI_DOMINATORS, e->src, bb_st); | |
1388 if (dom_ld != dom_st) | |
1389 return false; | |
1390 } | |
1391 | |
1392 *src_dr = single_ld; | |
1393 *dst_dr = single_st; | |
1394 return true; | |
0 | 1395 } |
1396 | |
111 | 1397 /* Given data reference DR in LOOP_NEST, this function checks the enclosing |
1398 loops from inner to outer to see if loop's step equals to access size at | |
1399 each level of loop. Return true if yes; record access base and size in | |
1400 BASE and SIZE; save loop's step at each level of loop in STEPS if it is | |
1401 not null. For example: | |
1402 | |
1403 int arr[100][100][100]; | |
1404 for (i = 0; i < 100; i++) ;steps[2] = 40000 | |
1405 for (j = 100; j > 0; j--) ;steps[1] = -400 | |
1406 for (k = 0; k < 100; k++) ;steps[0] = 4 | |
1407 arr[i][j - 1][k] = 0; ;base = &arr, size = 4000000. */ | |
1408 | |
1409 static bool | |
1410 compute_access_range (loop_p loop_nest, data_reference_p dr, tree *base, | |
1411 tree *size, vec<tree> *steps = NULL) | |
0 | 1412 { |
111 | 1413 location_t loc = gimple_location (DR_STMT (dr)); |
1414 basic_block bb = gimple_bb (DR_STMT (dr)); | |
1415 struct loop *loop = bb->loop_father; | |
1416 tree ref = DR_REF (dr); | |
1417 tree access_base = build_fold_addr_expr (ref); | |
1418 tree access_size = TYPE_SIZE_UNIT (TREE_TYPE (ref)); | |
1419 | |
1420 do { | |
1421 tree scev_fn = analyze_scalar_evolution (loop, access_base); | |
1422 if (TREE_CODE (scev_fn) != POLYNOMIAL_CHREC) | |
1423 return false; | |
1424 | |
1425 access_base = CHREC_LEFT (scev_fn); | |
1426 if (tree_contains_chrecs (access_base, NULL)) | |
1427 return false; | |
1428 | |
1429 tree scev_step = CHREC_RIGHT (scev_fn); | |
1430 /* Only support constant steps. */ | |
1431 if (TREE_CODE (scev_step) != INTEGER_CST) | |
1432 return false; | |
1433 | |
1434 enum ev_direction access_dir = scev_direction (scev_fn); | |
1435 if (access_dir == EV_DIR_UNKNOWN) | |
1436 return false; | |
1437 | |
1438 if (steps != NULL) | |
1439 steps->safe_push (scev_step); | |
1440 | |
1441 scev_step = fold_convert_loc (loc, sizetype, scev_step); | |
1442 /* Compute absolute value of scev step. */ | |
1443 if (access_dir == EV_DIR_DECREASES) | |
1444 scev_step = fold_build1_loc (loc, NEGATE_EXPR, sizetype, scev_step); | |
1445 | |
1446 /* At each level of loop, scev step must equal to access size. In other | |
1447 words, DR must access consecutive memory between loop iterations. */ | |
1448 if (!operand_equal_p (scev_step, access_size, 0)) | |
1449 return false; | |
1450 | |
1451 /* Compute DR's execution times in loop. */ | |
1452 tree niters = number_of_latch_executions (loop); | |
1453 niters = fold_convert_loc (loc, sizetype, niters); | |
1454 if (dominated_by_p (CDI_DOMINATORS, single_exit (loop)->src, bb)) | |
1455 niters = size_binop_loc (loc, PLUS_EXPR, niters, size_one_node); | |
1456 | |
1457 /* Compute DR's overall access size in loop. */ | |
1458 access_size = fold_build2_loc (loc, MULT_EXPR, sizetype, | |
1459 niters, scev_step); | |
1460 /* Adjust base address in case of negative step. */ | |
1461 if (access_dir == EV_DIR_DECREASES) | |
1462 { | |
1463 tree adj = fold_build2_loc (loc, MINUS_EXPR, sizetype, | |
1464 scev_step, access_size); | |
1465 access_base = fold_build_pointer_plus_loc (loc, access_base, adj); | |
1466 } | |
1467 } while (loop != loop_nest && (loop = loop_outer (loop)) != NULL); | |
1468 | |
1469 *base = access_base; | |
1470 *size = access_size; | |
1471 return true; | |
0 | 1472 } |
1473 | |
111 | 1474 /* Allocate and return builtin struct. Record information like DST_DR, |
1475 SRC_DR, DST_BASE, SRC_BASE and SIZE in the allocated struct. */ | |
1476 | |
1477 static struct builtin_info * | |
1478 alloc_builtin (data_reference_p dst_dr, data_reference_p src_dr, | |
1479 tree dst_base, tree src_base, tree size) | |
1480 { | |
1481 struct builtin_info *builtin = XNEW (struct builtin_info); | |
1482 builtin->dst_dr = dst_dr; | |
1483 builtin->src_dr = src_dr; | |
1484 builtin->dst_base = dst_base; | |
1485 builtin->src_base = src_base; | |
1486 builtin->size = size; | |
1487 return builtin; | |
1488 } | |
1489 | |
1490 /* Given data reference DR in loop nest LOOP, classify if it forms builtin | |
1491 memset call. */ | |
0 | 1492 |
1493 static void | |
111 | 1494 classify_builtin_st (loop_p loop, partition *partition, data_reference_p dr) |
0 | 1495 { |
111 | 1496 gimple *stmt = DR_STMT (dr); |
1497 tree base, size, rhs = gimple_assign_rhs1 (stmt); | |
1498 | |
1499 if (const_with_all_bytes_same (rhs) == -1 | |
1500 && (!INTEGRAL_TYPE_P (TREE_TYPE (rhs)) | |
1501 || (TYPE_MODE (TREE_TYPE (rhs)) | |
1502 != TYPE_MODE (unsigned_char_type_node)))) | |
1503 return; | |
1504 | |
1505 if (TREE_CODE (rhs) == SSA_NAME | |
1506 && !SSA_NAME_IS_DEFAULT_DEF (rhs) | |
1507 && flow_bb_inside_loop_p (loop, gimple_bb (SSA_NAME_DEF_STMT (rhs)))) | |
1508 return; | |
1509 | |
1510 if (!compute_access_range (loop, dr, &base, &size)) | |
1511 return; | |
1512 | |
1513 struct builtin_info *builtin; | |
1514 builtin = alloc_builtin (dr, NULL, base, NULL_TREE, size); | |
1515 builtin->dst_base_base = strip_offset (builtin->dst_base, | |
1516 &builtin->dst_base_offset); | |
1517 partition->builtin = builtin; | |
1518 partition->kind = PKIND_MEMSET; | |
0 | 1519 } |
1520 | |
111 | 1521 /* Given data references DST_DR and SRC_DR in loop nest LOOP and RDG, classify |
1522 if it forms builtin memcpy or memmove call. */ | |
0 | 1523 |
1524 static void | |
111 | 1525 classify_builtin_ldst (loop_p loop, struct graph *rdg, partition *partition, |
1526 data_reference_p dst_dr, data_reference_p src_dr) | |
0 | 1527 { |
111 | 1528 tree base, size, src_base, src_size; |
1529 auto_vec<tree> dst_steps, src_steps; | |
1530 | |
1531 /* Compute access range of both load and store. They much have the same | |
1532 access size. */ | |
1533 if (!compute_access_range (loop, dst_dr, &base, &size, &dst_steps) | |
1534 || !compute_access_range (loop, src_dr, &src_base, &src_size, &src_steps) | |
1535 || !operand_equal_p (size, src_size, 0)) | |
1536 return; | |
1537 | |
1538 /* Load and store in loop nest must access memory in the same way, i.e, | |
1539 their must have the same steps in each loop of the nest. */ | |
1540 if (dst_steps.length () != src_steps.length ()) | |
1541 return; | |
1542 for (unsigned i = 0; i < dst_steps.length (); ++i) | |
1543 if (!operand_equal_p (dst_steps[i], src_steps[i], 0)) | |
1544 return; | |
1545 | |
1546 /* Now check that if there is a dependence. */ | |
1547 ddr_p ddr = get_data_dependence (rdg, src_dr, dst_dr); | |
1548 | |
1549 /* Classify as memcpy if no dependence between load and store. */ | |
1550 if (DDR_ARE_DEPENDENT (ddr) == chrec_known) | |
1551 { | |
1552 partition->builtin = alloc_builtin (dst_dr, src_dr, base, src_base, size); | |
1553 partition->kind = PKIND_MEMCPY; | |
1554 return; | |
1555 } | |
1556 | |
1557 /* Can't do memmove in case of unknown dependence or dependence without | |
1558 classical distance vector. */ | |
1559 if (DDR_ARE_DEPENDENT (ddr) == chrec_dont_know | |
1560 || DDR_NUM_DIST_VECTS (ddr) == 0) | |
1561 return; | |
1562 | |
1563 unsigned i; | |
1564 lambda_vector dist_v; | |
1565 int num_lev = (DDR_LOOP_NEST (ddr)).length (); | |
1566 FOR_EACH_VEC_ELT (DDR_DIST_VECTS (ddr), i, dist_v) | |
0 | 1567 { |
111 | 1568 unsigned dep_lev = dependence_level (dist_v, num_lev); |
1569 /* Can't do memmove if load depends on store. */ | |
1570 if (dep_lev > 0 && dist_v[dep_lev - 1] > 0 && !DDR_REVERSED_P (ddr)) | |
1571 return; | |
1572 } | |
1573 | |
1574 partition->builtin = alloc_builtin (dst_dr, src_dr, base, src_base, size); | |
1575 partition->kind = PKIND_MEMMOVE; | |
1576 return; | |
1577 } | |
1578 | |
1579 /* Classifies the builtin kind we can generate for PARTITION of RDG and LOOP. | |
1580 For the moment we detect memset, memcpy and memmove patterns. Bitmap | |
1581 STMT_IN_ALL_PARTITIONS contains statements belonging to all partitions. */ | |
1582 | |
1583 static void | |
1584 classify_partition (loop_p loop, struct graph *rdg, partition *partition, | |
1585 bitmap stmt_in_all_partitions) | |
1586 { | |
1587 bitmap_iterator bi; | |
1588 unsigned i; | |
1589 data_reference_p single_ld = NULL, single_st = NULL; | |
1590 bool volatiles_p = false, has_reduction = false; | |
1591 | |
1592 EXECUTE_IF_SET_IN_BITMAP (partition->stmts, 0, i, bi) | |
1593 { | |
1594 gimple *stmt = RDG_STMT (rdg, i); | |
1595 | |
1596 if (gimple_has_volatile_ops (stmt)) | |
1597 volatiles_p = true; | |
1598 | |
1599 /* If the stmt is not included by all partitions and there is uses | |
1600 outside of the loop, then mark the partition as reduction. */ | |
1601 if (stmt_has_scalar_dependences_outside_loop (loop, stmt)) | |
0 | 1602 { |
111 | 1603 /* Due to limitation in the transform phase we have to fuse all |
1604 reduction partitions. As a result, this could cancel valid | |
1605 loop distribution especially for loop that induction variable | |
1606 is used outside of loop. To workaround this issue, we skip | |
1607 marking partition as reudction if the reduction stmt belongs | |
1608 to all partitions. In such case, reduction will be computed | |
1609 correctly no matter how partitions are fused/distributed. */ | |
1610 if (!bitmap_bit_p (stmt_in_all_partitions, i)) | |
1611 { | |
1612 partition->reduction_p = true; | |
1613 return; | |
1614 } | |
1615 has_reduction = true; | |
0 | 1616 } |
1617 } | |
1618 | |
111 | 1619 /* Perform general partition disqualification for builtins. */ |
1620 if (volatiles_p | |
1621 /* Simple workaround to prevent classifying the partition as builtin | |
1622 if it contains any use outside of loop. */ | |
1623 || has_reduction | |
1624 || !flag_tree_loop_distribute_patterns) | |
1625 return; | |
1626 | |
1627 /* Find single load/store data references for builtin partition. */ | |
1628 if (!find_single_drs (loop, rdg, partition, &single_st, &single_ld)) | |
1629 return; | |
1630 | |
1631 /* Classify the builtin kind. */ | |
1632 if (single_ld == NULL) | |
1633 classify_builtin_st (loop, partition, single_st); | |
1634 else | |
1635 classify_builtin_ldst (loop, rdg, partition, single_st, single_ld); | |
0 | 1636 } |
1637 | |
111 | 1638 /* Returns true when PARTITION1 and PARTITION2 access the same memory |
1639 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
|
1640 |
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
|
1641 static bool |
111 | 1642 share_memory_accesses (struct graph *rdg, |
1643 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
|
1644 { |
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
|
1645 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
|
1646 bitmap_iterator bi, bj; |
111 | 1647 data_reference_p dr1, dr2; |
1648 | |
1649 /* First check whether in the intersection of the two partitions are | |
1650 any loads or stores. Common loads are the situation that happens | |
1651 most often. */ | |
1652 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
|
1653 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
|
1654 || RDG_MEM_READS_STMT (rdg, i)) |
111 | 1655 return true; |
1656 | |
1657 /* Then check whether the two partitions access the same memory object. */ | |
1658 EXECUTE_IF_SET_IN_BITMAP (partition1->datarefs, 0, i, bi) | |
1659 { | |
1660 dr1 = datarefs_vec[i]; | |
1661 | |
1662 if (!DR_BASE_ADDRESS (dr1) | |
1663 || !DR_OFFSET (dr1) || !DR_INIT (dr1) || !DR_STEP (dr1)) | |
1664 continue; | |
1665 | |
1666 EXECUTE_IF_SET_IN_BITMAP (partition2->datarefs, 0, j, bj) | |
1667 { | |
1668 dr2 = datarefs_vec[j]; | |
1669 | |
1670 if (!DR_BASE_ADDRESS (dr2) | |
1671 || !DR_OFFSET (dr2) || !DR_INIT (dr2) || !DR_STEP (dr2)) | |
1672 continue; | |
1673 | |
1674 if (operand_equal_p (DR_BASE_ADDRESS (dr1), DR_BASE_ADDRESS (dr2), 0) | |
1675 && operand_equal_p (DR_OFFSET (dr1), DR_OFFSET (dr2), 0) | |
1676 && operand_equal_p (DR_INIT (dr1), DR_INIT (dr2), 0) | |
1677 && 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
|
1678 return true; |
111 | 1679 } |
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
|
1680 } |
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
|
1681 |
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
|
1682 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
|
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 |
111 | 1685 /* For each seed statement in STARTING_STMTS, this function builds |
1686 partition for it by adding depended statements according to RDG. | |
1687 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
|
1688 |
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 static void |
111 | 1690 rdg_build_partitions (struct graph *rdg, |
1691 vec<gimple *> starting_stmts, | |
1692 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
|
1693 { |
111 | 1694 auto_bitmap processed; |
0 | 1695 int i; |
111 | 1696 gimple *stmt; |
1697 | |
1698 FOR_EACH_VEC_ELT (starting_stmts, i, stmt) | |
0 | 1699 { |
111 | 1700 int v = rdg_vertex_for_stmt (rdg, stmt); |
1701 | |
1702 if (dump_file && (dump_flags & TDF_DETAILS)) | |
1703 fprintf (dump_file, | |
1704 "ldist asked to generate code for vertex %d\n", v); | |
1705 | |
1706 /* If the vertex is already contained in another partition so | |
1707 is the partition rooted at it. */ | |
0 | 1708 if (bitmap_bit_p (processed, v)) |
1709 continue; | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1710 |
111 | 1711 partition *partition = build_rdg_partition_for_vertex (rdg, v); |
1712 bitmap_ior_into (processed, partition->stmts); | |
1713 | |
1714 if (dump_file && (dump_flags & TDF_DETAILS)) | |
0 | 1715 { |
111 | 1716 fprintf (dump_file, "ldist creates useful %s partition:\n", |
1717 partition->type == PTYPE_PARALLEL ? "parallel" : "sequent"); | |
1718 bitmap_print (dump_file, partition->stmts, " ", "\n"); | |
0 | 1719 } |
111 | 1720 |
1721 partitions->safe_push (partition); | |
0 | 1722 } |
1723 | |
111 | 1724 /* All vertices should have been assigned to at least one partition now, |
1725 other than vertices belonging to dead code. */ | |
0 | 1726 } |
1727 | |
1728 /* Dump to FILE the PARTITIONS. */ | |
1729 | |
1730 static void | |
111 | 1731 dump_rdg_partitions (FILE *file, vec<partition *> partitions) |
0 | 1732 { |
1733 int i; | |
111 | 1734 partition *partition; |
1735 | |
1736 FOR_EACH_VEC_ELT (partitions, i, partition) | |
1737 debug_bitmap_file (file, partition->stmts); | |
0 | 1738 } |
1739 | |
1740 /* Debug PARTITIONS. */ | |
111 | 1741 extern void debug_rdg_partitions (vec<partition *> ); |
0 | 1742 |
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
|
1743 DEBUG_FUNCTION void |
111 | 1744 debug_rdg_partitions (vec<partition *> partitions) |
0 | 1745 { |
1746 dump_rdg_partitions (stderr, partitions); | |
1747 } | |
1748 | |
1749 /* Returns the number of read and write operations in the RDG. */ | |
1750 | |
1751 static int | |
1752 number_of_rw_in_rdg (struct graph *rdg) | |
1753 { | |
1754 int i, res = 0; | |
1755 | |
1756 for (i = 0; i < rdg->n_vertices; i++) | |
1757 { | |
1758 if (RDG_MEM_WRITE_STMT (rdg, i)) | |
1759 ++res; | |
1760 | |
1761 if (RDG_MEM_READS_STMT (rdg, i)) | |
1762 ++res; | |
1763 } | |
1764 | |
1765 return res; | |
1766 } | |
1767 | |
1768 /* Returns the number of read and write operations in a PARTITION of | |
1769 the RDG. */ | |
1770 | |
1771 static int | |
111 | 1772 number_of_rw_in_partition (struct graph *rdg, partition *partition) |
0 | 1773 { |
1774 int res = 0; | |
1775 unsigned i; | |
1776 bitmap_iterator ii; | |
1777 | |
111 | 1778 EXECUTE_IF_SET_IN_BITMAP (partition->stmts, 0, i, ii) |
0 | 1779 { |
1780 if (RDG_MEM_WRITE_STMT (rdg, i)) | |
1781 ++res; | |
1782 | |
1783 if (RDG_MEM_READS_STMT (rdg, i)) | |
1784 ++res; | |
1785 } | |
1786 | |
1787 return res; | |
1788 } | |
1789 | |
1790 /* Returns true when one of the PARTITIONS contains all the read or | |
1791 write operations of RDG. */ | |
1792 | |
1793 static bool | |
111 | 1794 partition_contains_all_rw (struct graph *rdg, |
1795 vec<partition *> partitions) | |
0 | 1796 { |
1797 int i; | |
111 | 1798 partition *partition; |
0 | 1799 int nrw = number_of_rw_in_rdg (rdg); |
1800 | |
111 | 1801 FOR_EACH_VEC_ELT (partitions, i, partition) |
0 | 1802 if (nrw == number_of_rw_in_partition (rdg, partition)) |
1803 return true; | |
1804 | |
1805 return false; | |
1806 } | |
1807 | |
111 | 1808 /* Compute partition dependence created by the data references in DRS1 |
1809 and DRS2, modify and return DIR according to that. IF ALIAS_DDR is | |
1810 not NULL, we record dependence introduced by possible alias between | |
1811 two data references in ALIAS_DDRS; otherwise, we simply ignore such | |
1812 dependence as if it doesn't exist at all. */ | |
1813 | |
1814 static int | |
1815 pg_add_dependence_edges (struct graph *rdg, int dir, | |
1816 bitmap drs1, bitmap drs2, vec<ddr_p> *alias_ddrs) | |
1817 { | |
1818 unsigned i, j; | |
1819 bitmap_iterator bi, bj; | |
1820 data_reference_p dr1, dr2, saved_dr1; | |
1821 | |
1822 /* dependence direction - 0 is no dependence, -1 is back, | |
1823 1 is forth, 2 is both (we can stop then, merging will occur). */ | |
1824 EXECUTE_IF_SET_IN_BITMAP (drs1, 0, i, bi) | |
1825 { | |
1826 dr1 = datarefs_vec[i]; | |
1827 | |
1828 EXECUTE_IF_SET_IN_BITMAP (drs2, 0, j, bj) | |
1829 { | |
1830 int res, this_dir = 1; | |
1831 ddr_p ddr; | |
1832 | |
1833 dr2 = datarefs_vec[j]; | |
1834 | |
1835 /* Skip all <read, read> data dependence. */ | |
1836 if (DR_IS_READ (dr1) && DR_IS_READ (dr2)) | |
1837 continue; | |
1838 | |
1839 saved_dr1 = dr1; | |
1840 /* Re-shuffle data-refs to be in topological order. */ | |
1841 if (rdg_vertex_for_stmt (rdg, DR_STMT (dr1)) | |
1842 > rdg_vertex_for_stmt (rdg, DR_STMT (dr2))) | |
1843 { | |
1844 std::swap (dr1, dr2); | |
1845 this_dir = -this_dir; | |
1846 } | |
1847 ddr = get_data_dependence (rdg, dr1, dr2); | |
1848 if (DDR_ARE_DEPENDENT (ddr) == chrec_dont_know) | |
1849 { | |
1850 this_dir = 0; | |
1851 res = data_ref_compare_tree (DR_BASE_ADDRESS (dr1), | |
1852 DR_BASE_ADDRESS (dr2)); | |
1853 /* Be conservative. If data references are not well analyzed, | |
1854 or the two data references have the same base address and | |
1855 offset, add dependence and consider it alias to each other. | |
1856 In other words, the dependence can not be resolved by | |
1857 runtime alias check. */ | |
1858 if (!DR_BASE_ADDRESS (dr1) || !DR_BASE_ADDRESS (dr2) | |
1859 || !DR_OFFSET (dr1) || !DR_OFFSET (dr2) | |
1860 || !DR_INIT (dr1) || !DR_INIT (dr2) | |
1861 || !DR_STEP (dr1) || !tree_fits_uhwi_p (DR_STEP (dr1)) | |
1862 || !DR_STEP (dr2) || !tree_fits_uhwi_p (DR_STEP (dr2)) | |
1863 || res == 0) | |
1864 this_dir = 2; | |
1865 /* Data dependence could be resolved by runtime alias check, | |
1866 record it in ALIAS_DDRS. */ | |
1867 else if (alias_ddrs != NULL) | |
1868 alias_ddrs->safe_push (ddr); | |
1869 /* Or simply ignore it. */ | |
1870 } | |
1871 else if (DDR_ARE_DEPENDENT (ddr) == NULL_TREE) | |
1872 { | |
1873 if (DDR_REVERSED_P (ddr)) | |
1874 this_dir = -this_dir; | |
1875 | |
1876 /* Known dependences can still be unordered througout the | |
1877 iteration space, see gcc.dg/tree-ssa/ldist-16.c. */ | |
1878 if (DDR_NUM_DIST_VECTS (ddr) != 1) | |
1879 this_dir = 2; | |
1880 /* If the overlap is exact preserve stmt order. */ | |
1881 else if (lambda_vector_zerop (DDR_DIST_VECT (ddr, 0), 1)) | |
1882 ; | |
1883 /* Else as the distance vector is lexicographic positive swap | |
1884 the dependence direction. */ | |
1885 else | |
1886 this_dir = -this_dir; | |
1887 } | |
1888 else | |
1889 this_dir = 0; | |
1890 if (this_dir == 2) | |
1891 return 2; | |
1892 else if (dir == 0) | |
1893 dir = this_dir; | |
1894 else if (this_dir != 0 && dir != this_dir) | |
1895 return 2; | |
1896 /* Shuffle "back" dr1. */ | |
1897 dr1 = saved_dr1; | |
1898 } | |
1899 } | |
1900 return dir; | |
1901 } | |
1902 | |
1903 /* Compare postorder number of the partition graph vertices V1 and V2. */ | |
0 | 1904 |
1905 static int | |
111 | 1906 pgcmp (const void *v1_, const void *v2_) |
1907 { | |
1908 const vertex *v1 = (const vertex *)v1_; | |
1909 const vertex *v2 = (const vertex *)v2_; | |
1910 return v2->post - v1->post; | |
1911 } | |
1912 | |
1913 /* Data attached to vertices of partition dependence graph. */ | |
1914 struct pg_vdata | |
1915 { | |
1916 /* ID of the corresponding partition. */ | |
1917 int id; | |
1918 /* The partition. */ | |
1919 struct partition *partition; | |
1920 }; | |
1921 | |
1922 /* Data attached to edges of partition dependence graph. */ | |
1923 struct pg_edata | |
1924 { | |
1925 /* If the dependence edge can be resolved by runtime alias check, | |
1926 this vector contains data dependence relations for runtime alias | |
1927 check. On the other hand, if the dependence edge is introduced | |
1928 because of compilation time known data dependence, this vector | |
1929 contains nothing. */ | |
1930 vec<ddr_p> alias_ddrs; | |
1931 }; | |
1932 | |
1933 /* Callback data for traversing edges in graph. */ | |
1934 struct pg_edge_callback_data | |
0 | 1935 { |
111 | 1936 /* Bitmap contains strong connected components should be merged. */ |
1937 bitmap sccs_to_merge; | |
1938 /* Array constains component information for all vertices. */ | |
1939 int *vertices_component; | |
1940 /* Vector to record all data dependence relations which are needed | |
1941 to break strong connected components by runtime alias checks. */ | |
1942 vec<ddr_p> *alias_ddrs; | |
1943 }; | |
1944 | |
1945 /* Initialize vertice's data for partition dependence graph PG with | |
1946 PARTITIONS. */ | |
1947 | |
1948 static void | |
1949 init_partition_graph_vertices (struct graph *pg, | |
1950 vec<struct partition *> *partitions) | |
1951 { | |
1952 int i; | |
1953 partition *partition; | |
1954 struct pg_vdata *data; | |
1955 | |
1956 for (i = 0; partitions->iterate (i, &partition); ++i) | |
1957 { | |
1958 data = new pg_vdata; | |
1959 pg->vertices[i].data = data; | |
1960 data->id = i; | |
1961 data->partition = partition; | |
1962 } | |
1963 } | |
1964 | |
1965 /* Add edge <I, J> to partition dependence graph PG. Attach vector of data | |
1966 dependence relations to the EDGE if DDRS isn't NULL. */ | |
1967 | |
1968 static void | |
1969 add_partition_graph_edge (struct graph *pg, int i, int j, vec<ddr_p> *ddrs) | |
1970 { | |
1971 struct graph_edge *e = add_edge (pg, i, j); | |
1972 | |
1973 /* If the edge is attached with data dependence relations, it means this | |
1974 dependence edge can be resolved by runtime alias checks. */ | |
1975 if (ddrs != NULL) | |
0 | 1976 { |
111 | 1977 struct pg_edata *data = new pg_edata; |
1978 | |
1979 gcc_assert (ddrs->length () > 0); | |
1980 e->data = data; | |
1981 data->alias_ddrs = vNULL; | |
1982 data->alias_ddrs.safe_splice (*ddrs); | |
1983 } | |
1984 } | |
1985 | |
1986 /* Callback function for graph travesal algorithm. It returns true | |
1987 if edge E should skipped when traversing the graph. */ | |
1988 | |
1989 static bool | |
1990 pg_skip_alias_edge (struct graph_edge *e) | |
1991 { | |
1992 struct pg_edata *data = (struct pg_edata *)e->data; | |
1993 return (data != NULL && data->alias_ddrs.length () > 0); | |
1994 } | |
1995 | |
1996 /* Callback function freeing data attached to edge E of graph. */ | |
1997 | |
1998 static void | |
1999 free_partition_graph_edata_cb (struct graph *, struct graph_edge *e, void *) | |
2000 { | |
2001 if (e->data != NULL) | |
2002 { | |
2003 struct pg_edata *data = (struct pg_edata *)e->data; | |
2004 data->alias_ddrs.release (); | |
2005 delete data; | |
2006 } | |
2007 } | |
2008 | |
2009 /* Free data attached to vertice of partition dependence graph PG. */ | |
2010 | |
2011 static void | |
2012 free_partition_graph_vdata (struct graph *pg) | |
2013 { | |
2014 int i; | |
2015 struct pg_vdata *data; | |
2016 | |
2017 for (i = 0; i < pg->n_vertices; ++i) | |
2018 { | |
2019 data = (struct pg_vdata *)pg->vertices[i].data; | |
2020 delete data; | |
2021 } | |
2022 } | |
2023 | |
2024 /* Build and return partition dependence graph for PARTITIONS. RDG is | |
2025 reduced dependence graph for the loop to be distributed. If IGNORE_ALIAS_P | |
2026 is true, data dependence caused by possible alias between references | |
2027 is ignored, as if it doesn't exist at all; otherwise all depdendences | |
2028 are considered. */ | |
2029 | |
2030 static struct graph * | |
2031 build_partition_graph (struct graph *rdg, | |
2032 vec<struct partition *> *partitions, | |
2033 bool ignore_alias_p) | |
2034 { | |
2035 int i, j; | |
2036 struct partition *partition1, *partition2; | |
2037 graph *pg = new_graph (partitions->length ()); | |
2038 auto_vec<ddr_p> alias_ddrs, *alias_ddrs_p; | |
2039 | |
2040 alias_ddrs_p = ignore_alias_p ? NULL : &alias_ddrs; | |
2041 | |
2042 init_partition_graph_vertices (pg, partitions); | |
2043 | |
2044 for (i = 0; partitions->iterate (i, &partition1); ++i) | |
2045 { | |
2046 for (j = i + 1; partitions->iterate (j, &partition2); ++j) | |
0 | 2047 { |
111 | 2048 /* dependence direction - 0 is no dependence, -1 is back, |
2049 1 is forth, 2 is both (we can stop then, merging will occur). */ | |
2050 int dir = 0; | |
2051 | |
2052 /* If the first partition has reduction, add back edge; if the | |
2053 second partition has reduction, add forth edge. This makes | |
2054 sure that reduction partition will be sorted as the last one. */ | |
2055 if (partition_reduction_p (partition1)) | |
2056 dir = -1; | |
2057 else if (partition_reduction_p (partition2)) | |
2058 dir = 1; | |
2059 | |
2060 /* Cleanup the temporary vector. */ | |
2061 alias_ddrs.truncate (0); | |
2062 | |
2063 dir = pg_add_dependence_edges (rdg, dir, partition1->datarefs, | |
2064 partition2->datarefs, alias_ddrs_p); | |
2065 | |
2066 /* Add edge to partition graph if there exists dependence. There | |
2067 are two types of edges. One type edge is caused by compilation | |
2068 time known dependence, this type can not be resolved by runtime | |
2069 alias check. The other type can be resolved by runtime alias | |
2070 check. */ | |
2071 if (dir == 1 || dir == 2 | |
2072 || alias_ddrs.length () > 0) | |
2073 { | |
2074 /* Attach data dependence relations to edge that can be resolved | |
2075 by runtime alias check. */ | |
2076 bool alias_edge_p = (dir != 1 && dir != 2); | |
2077 add_partition_graph_edge (pg, i, j, | |
2078 (alias_edge_p) ? &alias_ddrs : NULL); | |
2079 } | |
2080 if (dir == -1 || dir == 2 | |
2081 || alias_ddrs.length () > 0) | |
2082 { | |
2083 /* Attach data dependence relations to edge that can be resolved | |
2084 by runtime alias check. */ | |
2085 bool alias_edge_p = (dir != -1 && dir != 2); | |
2086 add_partition_graph_edge (pg, j, i, | |
2087 (alias_edge_p) ? &alias_ddrs : NULL); | |
2088 } | |
0 | 2089 } |
2090 } | |
111 | 2091 return pg; |
2092 } | |
2093 | |
2094 /* Sort partitions in PG in descending post order and store them in | |
2095 PARTITIONS. */ | |
2096 | |
2097 static void | |
2098 sort_partitions_by_post_order (struct graph *pg, | |
2099 vec<struct partition *> *partitions) | |
2100 { | |
2101 int i; | |
2102 struct pg_vdata *data; | |
2103 | |
2104 /* Now order the remaining nodes in descending postorder. */ | |
2105 qsort (pg->vertices, pg->n_vertices, sizeof (vertex), pgcmp); | |
2106 partitions->truncate (0); | |
2107 for (i = 0; i < pg->n_vertices; ++i) | |
2108 { | |
2109 data = (struct pg_vdata *)pg->vertices[i].data; | |
2110 if (data->partition) | |
2111 partitions->safe_push (data->partition); | |
2112 } | |
2113 } | |
2114 | |
2115 /* Given reduced dependence graph RDG merge strong connected components | |
2116 of PARTITIONS. If IGNORE_ALIAS_P is true, data dependence caused by | |
2117 possible alias between references is ignored, as if it doesn't exist | |
2118 at all; otherwise all depdendences are considered. */ | |
2119 | |
2120 static void | |
2121 merge_dep_scc_partitions (struct graph *rdg, | |
2122 vec<struct partition *> *partitions, | |
2123 bool ignore_alias_p) | |
2124 { | |
2125 struct partition *partition1, *partition2; | |
2126 struct pg_vdata *data; | |
2127 graph *pg = build_partition_graph (rdg, partitions, ignore_alias_p); | |
2128 int i, j, num_sccs = graphds_scc (pg, NULL); | |
2129 | |
2130 /* Strong connected compoenent means dependence cycle, we cannot distribute | |
2131 them. So fuse them together. */ | |
2132 if ((unsigned) num_sccs < partitions->length ()) | |
2133 { | |
2134 for (i = 0; i < num_sccs; ++i) | |
2135 { | |
2136 for (j = 0; partitions->iterate (j, &partition1); ++j) | |
2137 if (pg->vertices[j].component == i) | |
2138 break; | |
2139 for (j = j + 1; partitions->iterate (j, &partition2); ++j) | |
2140 if (pg->vertices[j].component == i) | |
2141 { | |
2142 partition_merge_into (NULL, partition1, | |
2143 partition2, FUSE_SAME_SCC); | |
2144 partition1->type = PTYPE_SEQUENTIAL; | |
2145 (*partitions)[j] = NULL; | |
2146 partition_free (partition2); | |
2147 data = (struct pg_vdata *)pg->vertices[j].data; | |
2148 data->partition = NULL; | |
2149 } | |
2150 } | |
2151 } | |
2152 | |
2153 sort_partitions_by_post_order (pg, partitions); | |
2154 gcc_assert (partitions->length () == (unsigned)num_sccs); | |
2155 free_partition_graph_vdata (pg); | |
2156 free_graph (pg); | |
2157 } | |
2158 | |
2159 /* Callback function for traversing edge E in graph G. DATA is private | |
2160 callback data. */ | |
2161 | |
2162 static void | |
2163 pg_collect_alias_ddrs (struct graph *g, struct graph_edge *e, void *data) | |
2164 { | |
2165 int i, j, component; | |
2166 struct pg_edge_callback_data *cbdata; | |
2167 struct pg_edata *edata = (struct pg_edata *) e->data; | |
2168 | |
2169 /* If the edge doesn't have attached data dependence, it represents | |
2170 compilation time known dependences. This type dependence cannot | |
2171 be resolved by runtime alias check. */ | |
2172 if (edata == NULL || edata->alias_ddrs.length () == 0) | |
2173 return; | |
2174 | |
2175 cbdata = (struct pg_edge_callback_data *) data; | |
2176 i = e->src; | |
2177 j = e->dest; | |
2178 component = cbdata->vertices_component[i]; | |
2179 /* Vertices are topologically sorted according to compilation time | |
2180 known dependences, so we can break strong connected components | |
2181 by removing edges of the opposite direction, i.e, edges pointing | |
2182 from vertice with smaller post number to vertice with bigger post | |
2183 number. */ | |
2184 if (g->vertices[i].post < g->vertices[j].post | |
2185 /* We only need to remove edges connecting vertices in the same | |
2186 strong connected component to break it. */ | |
2187 && component == cbdata->vertices_component[j] | |
2188 /* Check if we want to break the strong connected component or not. */ | |
2189 && !bitmap_bit_p (cbdata->sccs_to_merge, component)) | |
2190 cbdata->alias_ddrs->safe_splice (edata->alias_ddrs); | |
2191 } | |
2192 | |
2193 /* This is the main function breaking strong conected components in | |
2194 PARTITIONS giving reduced depdendence graph RDG. Store data dependence | |
2195 relations for runtime alias check in ALIAS_DDRS. */ | |
2196 | |
2197 static void | |
2198 break_alias_scc_partitions (struct graph *rdg, | |
2199 vec<struct partition *> *partitions, | |
2200 vec<ddr_p> *alias_ddrs) | |
2201 { | |
2202 int i, j, k, num_sccs, num_sccs_no_alias; | |
2203 /* Build partition dependence graph. */ | |
2204 graph *pg = build_partition_graph (rdg, partitions, false); | |
2205 | |
2206 alias_ddrs->truncate (0); | |
2207 /* Find strong connected components in the graph, with all dependence edges | |
2208 considered. */ | |
2209 num_sccs = graphds_scc (pg, NULL); | |
2210 /* All SCCs now can be broken by runtime alias checks because SCCs caused by | |
2211 compilation time known dependences are merged before this function. */ | |
2212 if ((unsigned) num_sccs < partitions->length ()) | |
2213 { | |
2214 struct pg_edge_callback_data cbdata; | |
2215 auto_bitmap sccs_to_merge; | |
2216 auto_vec<enum partition_type> scc_types; | |
2217 struct partition *partition, *first; | |
2218 | |
2219 /* If all partitions in a SCC have the same type, we can simply merge the | |
2220 SCC. This loop finds out such SCCS and record them in bitmap. */ | |
2221 bitmap_set_range (sccs_to_merge, 0, (unsigned) num_sccs); | |
2222 for (i = 0; i < num_sccs; ++i) | |
2223 { | |
2224 for (j = 0; partitions->iterate (j, &first); ++j) | |
2225 if (pg->vertices[j].component == i) | |
2226 break; | |
2227 for (++j; partitions->iterate (j, &partition); ++j) | |
2228 { | |
2229 if (pg->vertices[j].component != i) | |
2230 continue; | |
2231 | |
2232 /* Note we Merge partitions of parallel type on purpose, though | |
2233 the result partition is sequential. The reason is vectorizer | |
2234 can do more accurate runtime alias check in this case. Also | |
2235 it results in more conservative distribution. */ | |
2236 if (first->type != partition->type) | |
2237 { | |
2238 bitmap_clear_bit (sccs_to_merge, i); | |
2239 break; | |
2240 } | |
2241 } | |
2242 } | |
2243 | |
2244 /* Initialize callback data for traversing. */ | |
2245 cbdata.sccs_to_merge = sccs_to_merge; | |
2246 cbdata.alias_ddrs = alias_ddrs; | |
2247 cbdata.vertices_component = XNEWVEC (int, pg->n_vertices); | |
2248 /* Record the component information which will be corrupted by next | |
2249 graph scc finding call. */ | |
2250 for (i = 0; i < pg->n_vertices; ++i) | |
2251 cbdata.vertices_component[i] = pg->vertices[i].component; | |
2252 | |
2253 /* Collect data dependences for runtime alias checks to break SCCs. */ | |
2254 if (bitmap_count_bits (sccs_to_merge) != (unsigned) num_sccs) | |
2255 { | |
2256 /* Run SCC finding algorithm again, with alias dependence edges | |
2257 skipped. This is to topologically sort partitions according to | |
2258 compilation time known dependence. Note the topological order | |
2259 is stored in the form of pg's post order number. */ | |
2260 num_sccs_no_alias = graphds_scc (pg, NULL, pg_skip_alias_edge); | |
2261 gcc_assert (partitions->length () == (unsigned) num_sccs_no_alias); | |
2262 /* With topological order, we can construct two subgraphs L and R. | |
2263 L contains edge <x, y> where x < y in terms of post order, while | |
2264 R contains edge <x, y> where x > y. Edges for compilation time | |
2265 known dependence all fall in R, so we break SCCs by removing all | |
2266 (alias) edges of in subgraph L. */ | |
2267 for_each_edge (pg, pg_collect_alias_ddrs, &cbdata); | |
2268 } | |
2269 | |
2270 /* For SCC that doesn't need to be broken, merge it. */ | |
2271 for (i = 0; i < num_sccs; ++i) | |
2272 { | |
2273 if (!bitmap_bit_p (sccs_to_merge, i)) | |
2274 continue; | |
2275 | |
2276 for (j = 0; partitions->iterate (j, &first); ++j) | |
2277 if (cbdata.vertices_component[j] == i) | |
2278 break; | |
2279 for (k = j + 1; partitions->iterate (k, &partition); ++k) | |
2280 { | |
2281 struct pg_vdata *data; | |
2282 | |
2283 if (cbdata.vertices_component[k] != i) | |
2284 continue; | |
2285 | |
2286 /* Update postorder number so that merged reduction partition is | |
2287 sorted after other partitions. */ | |
2288 if (!partition_reduction_p (first) | |
2289 && partition_reduction_p (partition)) | |
2290 { | |
2291 gcc_assert (pg->vertices[k].post < pg->vertices[j].post); | |
2292 pg->vertices[j].post = pg->vertices[k].post; | |
2293 } | |
2294 partition_merge_into (NULL, first, partition, FUSE_SAME_SCC); | |
2295 (*partitions)[k] = NULL; | |
2296 partition_free (partition); | |
2297 data = (struct pg_vdata *)pg->vertices[k].data; | |
2298 gcc_assert (data->id == k); | |
2299 data->partition = NULL; | |
2300 /* The result partition of merged SCC must be sequential. */ | |
2301 first->type = PTYPE_SEQUENTIAL; | |
2302 } | |
2303 } | |
2304 } | |
2305 | |
2306 sort_partitions_by_post_order (pg, partitions); | |
2307 free_partition_graph_vdata (pg); | |
2308 for_each_edge (pg, free_partition_graph_edata_cb, NULL); | |
2309 free_graph (pg); | |
0 | 2310 |
2311 if (dump_file && (dump_flags & TDF_DETAILS)) | |
111 | 2312 { |
2313 fprintf (dump_file, "Possible alias data dependence to break:\n"); | |
2314 dump_data_dependence_relations (dump_file, *alias_ddrs); | |
2315 } | |
2316 } | |
2317 | |
2318 /* Compute and return an expression whose value is the segment length which | |
2319 will be accessed by DR in NITERS iterations. */ | |
2320 | |
2321 static tree | |
2322 data_ref_segment_size (struct data_reference *dr, tree niters) | |
2323 { | |
2324 tree segment_length; | |
2325 | |
2326 if (integer_zerop (DR_STEP (dr))) | |
2327 segment_length = TYPE_SIZE_UNIT (TREE_TYPE (DR_REF (dr))); | |
2328 else | |
2329 segment_length = size_binop (MULT_EXPR, | |
2330 fold_convert (sizetype, DR_STEP (dr)), | |
2331 fold_convert (sizetype, niters)); | |
2332 | |
2333 return segment_length; | |
2334 } | |
2335 | |
2336 /* Return true if LOOP's latch is dominated by statement for data reference | |
2337 DR. */ | |
2338 | |
2339 static inline bool | |
2340 latch_dominated_by_data_ref (struct loop *loop, data_reference *dr) | |
2341 { | |
2342 return dominated_by_p (CDI_DOMINATORS, single_exit (loop)->src, | |
2343 gimple_bb (DR_STMT (dr))); | |
2344 } | |
2345 | |
2346 /* Compute alias check pairs and store them in COMP_ALIAS_PAIRS for LOOP's | |
2347 data dependence relations ALIAS_DDRS. */ | |
2348 | |
2349 static void | |
2350 compute_alias_check_pairs (struct loop *loop, vec<ddr_p> *alias_ddrs, | |
2351 vec<dr_with_seg_len_pair_t> *comp_alias_pairs) | |
2352 { | |
2353 unsigned int i; | |
2354 unsigned HOST_WIDE_INT factor = 1; | |
2355 tree niters_plus_one, niters = number_of_latch_executions (loop); | |
2356 | |
2357 gcc_assert (niters != NULL_TREE && niters != chrec_dont_know); | |
2358 niters = fold_convert (sizetype, niters); | |
2359 niters_plus_one = size_binop (PLUS_EXPR, niters, size_one_node); | |
2360 | |
2361 if (dump_file && (dump_flags & TDF_DETAILS)) | |
2362 fprintf (dump_file, "Creating alias check pairs:\n"); | |
2363 | |
2364 /* Iterate all data dependence relations and compute alias check pairs. */ | |
2365 for (i = 0; i < alias_ddrs->length (); i++) | |
2366 { | |
2367 ddr_p ddr = (*alias_ddrs)[i]; | |
2368 struct data_reference *dr_a = DDR_A (ddr); | |
2369 struct data_reference *dr_b = DDR_B (ddr); | |
2370 tree seg_length_a, seg_length_b; | |
2371 int comp_res = data_ref_compare_tree (DR_BASE_ADDRESS (dr_a), | |
2372 DR_BASE_ADDRESS (dr_b)); | |
2373 | |
2374 if (comp_res == 0) | |
2375 comp_res = data_ref_compare_tree (DR_OFFSET (dr_a), DR_OFFSET (dr_b)); | |
2376 gcc_assert (comp_res != 0); | |
2377 | |
2378 if (latch_dominated_by_data_ref (loop, dr_a)) | |
2379 seg_length_a = data_ref_segment_size (dr_a, niters_plus_one); | |
2380 else | |
2381 seg_length_a = data_ref_segment_size (dr_a, niters); | |
2382 | |
2383 if (latch_dominated_by_data_ref (loop, dr_b)) | |
2384 seg_length_b = data_ref_segment_size (dr_b, niters_plus_one); | |
2385 else | |
2386 seg_length_b = data_ref_segment_size (dr_b, niters); | |
2387 | |
2388 dr_with_seg_len_pair_t dr_with_seg_len_pair | |
2389 (dr_with_seg_len (dr_a, seg_length_a), | |
2390 dr_with_seg_len (dr_b, seg_length_b)); | |
2391 | |
2392 /* Canonicalize pairs by sorting the two DR members. */ | |
2393 if (comp_res > 0) | |
2394 std::swap (dr_with_seg_len_pair.first, dr_with_seg_len_pair.second); | |
2395 | |
2396 comp_alias_pairs->safe_push (dr_with_seg_len_pair); | |
2397 } | |
2398 | |
2399 if (tree_fits_uhwi_p (niters)) | |
2400 factor = tree_to_uhwi (niters); | |
2401 | |
2402 /* Prune alias check pairs. */ | |
2403 prune_runtime_alias_test_list (comp_alias_pairs, factor); | |
2404 if (dump_file && (dump_flags & TDF_DETAILS)) | |
2405 fprintf (dump_file, | |
2406 "Improved number of alias checks from %d to %d\n", | |
2407 alias_ddrs->length (), comp_alias_pairs->length ()); | |
2408 } | |
2409 | |
2410 /* Given data dependence relations in ALIAS_DDRS, generate runtime alias | |
2411 checks and version LOOP under condition of these runtime alias checks. */ | |
2412 | |
2413 static void | |
2414 version_loop_by_alias_check (struct loop *loop, vec<ddr_p> *alias_ddrs) | |
2415 { | |
2416 profile_probability prob; | |
2417 basic_block cond_bb; | |
2418 struct loop *nloop; | |
2419 tree lhs, arg0, cond_expr = NULL_TREE; | |
2420 gimple_seq cond_stmts = NULL; | |
2421 gimple *call_stmt = NULL; | |
2422 auto_vec<dr_with_seg_len_pair_t> comp_alias_pairs; | |
2423 | |
2424 /* Generate code for runtime alias checks if necessary. */ | |
2425 gcc_assert (alias_ddrs->length () > 0); | |
2426 | |
2427 if (dump_file && (dump_flags & TDF_DETAILS)) | |
2428 fprintf (dump_file, | |
2429 "Version loop <%d> with runtime alias check\n", loop->num); | |
2430 | |
2431 compute_alias_check_pairs (loop, alias_ddrs, &comp_alias_pairs); | |
2432 create_runtime_alias_checks (loop, &comp_alias_pairs, &cond_expr); | |
2433 cond_expr = force_gimple_operand_1 (cond_expr, &cond_stmts, | |
2434 is_gimple_val, NULL_TREE); | |
2435 | |
2436 /* Depend on vectorizer to fold IFN_LOOP_DIST_ALIAS. */ | |
2437 if (flag_tree_loop_vectorize) | |
2438 { | |
2439 /* Generate internal function call for loop distribution alias check. */ | |
2440 call_stmt = gimple_build_call_internal (IFN_LOOP_DIST_ALIAS, | |
2441 2, NULL_TREE, cond_expr); | |
2442 lhs = make_ssa_name (boolean_type_node); | |
2443 gimple_call_set_lhs (call_stmt, lhs); | |
2444 } | |
2445 else | |
2446 lhs = cond_expr; | |
2447 | |
2448 prob = profile_probability::guessed_always ().apply_scale (9, 10); | |
2449 initialize_original_copy_tables (); | |
2450 nloop = loop_version (loop, lhs, &cond_bb, prob, prob.invert (), | |
2451 prob, prob.invert (), true); | |
2452 free_original_copy_tables (); | |
2453 /* Record the original loop number in newly generated loops. In case of | |
2454 distribution, the original loop will be distributed and the new loop | |
2455 is kept. */ | |
2456 loop->orig_loop_num = nloop->num; | |
2457 nloop->orig_loop_num = nloop->num; | |
2458 nloop->dont_vectorize = true; | |
2459 nloop->force_vectorize = false; | |
2460 | |
2461 if (call_stmt) | |
2462 { | |
2463 /* Record new loop's num in IFN_LOOP_DIST_ALIAS because the original | |
2464 loop could be destroyed. */ | |
2465 arg0 = build_int_cst (integer_type_node, loop->orig_loop_num); | |
2466 gimple_call_set_arg (call_stmt, 0, arg0); | |
2467 gimple_seq_add_stmt_without_update (&cond_stmts, call_stmt); | |
2468 } | |
2469 | |
2470 if (cond_stmts) | |
2471 { | |
2472 gimple_stmt_iterator cond_gsi = gsi_last_bb (cond_bb); | |
2473 gsi_insert_seq_before (&cond_gsi, cond_stmts, GSI_SAME_STMT); | |
2474 } | |
2475 update_ssa (TODO_update_ssa); | |
0 | 2476 } |
2477 | |
111 | 2478 /* Return true if loop versioning is needed to distrubute PARTITIONS. |
2479 ALIAS_DDRS are data dependence relations for runtime alias check. */ | |
2480 | |
2481 static inline bool | |
2482 version_for_distribution_p (vec<struct partition *> *partitions, | |
2483 vec<ddr_p> *alias_ddrs) | |
2484 { | |
2485 /* No need to version loop if we have only one partition. */ | |
2486 if (partitions->length () == 1) | |
2487 return false; | |
2488 | |
2489 /* Need to version loop if runtime alias check is necessary. */ | |
2490 return (alias_ddrs->length () > 0); | |
2491 } | |
2492 | |
2493 /* Compare base offset of builtin mem* partitions P1 and P2. */ | |
2494 | |
2495 static bool | |
2496 offset_cmp (struct partition *p1, struct partition *p2) | |
2497 { | |
2498 gcc_assert (p1 != NULL && p1->builtin != NULL); | |
2499 gcc_assert (p2 != NULL && p2->builtin != NULL); | |
2500 return p1->builtin->dst_base_offset < p2->builtin->dst_base_offset; | |
2501 } | |
2502 | |
2503 /* Fuse adjacent memset builtin PARTITIONS if possible. This is a special | |
2504 case optimization transforming below code: | |
2505 | |
2506 __builtin_memset (&obj, 0, 100); | |
2507 _1 = &obj + 100; | |
2508 __builtin_memset (_1, 0, 200); | |
2509 _2 = &obj + 300; | |
2510 __builtin_memset (_2, 0, 100); | |
2511 | |
2512 into: | |
2513 | |
2514 __builtin_memset (&obj, 0, 400); | |
2515 | |
2516 Note we don't have dependence information between different partitions | |
2517 at this point, as a result, we can't handle nonadjacent memset builtin | |
2518 partitions since dependence might be broken. */ | |
2519 | |
2520 static void | |
2521 fuse_memset_builtins (vec<struct partition *> *partitions) | |
2522 { | |
2523 unsigned i, j; | |
2524 struct partition *part1, *part2; | |
2525 | |
2526 for (i = 0; partitions->iterate (i, &part1);) | |
2527 { | |
2528 if (part1->kind != PKIND_MEMSET) | |
2529 { | |
2530 i++; | |
2531 continue; | |
2532 } | |
2533 | |
2534 /* Find sub-array of memset builtins of the same base. Index range | |
2535 of the sub-array is [i, j) with "j > i". */ | |
2536 for (j = i + 1; partitions->iterate (j, &part2); ++j) | |
2537 { | |
2538 if (part2->kind != PKIND_MEMSET | |
2539 || !operand_equal_p (part1->builtin->dst_base_base, | |
2540 part2->builtin->dst_base_base, 0)) | |
2541 break; | |
2542 } | |
2543 | |
2544 /* Stable sort is required in order to avoid breaking dependence. */ | |
2545 std::stable_sort (&(*partitions)[i], | |
2546 &(*partitions)[i] + j - i, offset_cmp); | |
2547 /* Continue with next partition. */ | |
2548 i = j; | |
2549 } | |
2550 | |
2551 /* Merge all consecutive memset builtin partitions. */ | |
2552 for (i = 0; i < partitions->length () - 1;) | |
2553 { | |
2554 part1 = (*partitions)[i]; | |
2555 if (part1->kind != PKIND_MEMSET) | |
2556 { | |
2557 i++; | |
2558 continue; | |
2559 } | |
2560 | |
2561 part2 = (*partitions)[i + 1]; | |
2562 /* Only merge memset partitions of the same base and with constant | |
2563 access sizes. */ | |
2564 if (part2->kind != PKIND_MEMSET | |
2565 || TREE_CODE (part1->builtin->size) != INTEGER_CST | |
2566 || TREE_CODE (part2->builtin->size) != INTEGER_CST | |
2567 || !operand_equal_p (part1->builtin->dst_base_base, | |
2568 part2->builtin->dst_base_base, 0)) | |
2569 { | |
2570 i++; | |
2571 continue; | |
2572 } | |
2573 tree rhs1 = gimple_assign_rhs1 (DR_STMT (part1->builtin->dst_dr)); | |
2574 tree rhs2 = gimple_assign_rhs1 (DR_STMT (part2->builtin->dst_dr)); | |
2575 int bytev1 = const_with_all_bytes_same (rhs1); | |
2576 int bytev2 = const_with_all_bytes_same (rhs2); | |
2577 /* Only merge memset partitions of the same value. */ | |
2578 if (bytev1 != bytev2 || bytev1 == -1) | |
2579 { | |
2580 i++; | |
2581 continue; | |
2582 } | |
2583 wide_int end1 = wi::add (part1->builtin->dst_base_offset, | |
2584 wi::to_wide (part1->builtin->size)); | |
2585 /* Only merge adjacent memset partitions. */ | |
2586 if (wi::ne_p (end1, part2->builtin->dst_base_offset)) | |
2587 { | |
2588 i++; | |
2589 continue; | |
2590 } | |
2591 /* Merge partitions[i] and partitions[i+1]. */ | |
2592 part1->builtin->size = fold_build2 (PLUS_EXPR, sizetype, | |
2593 part1->builtin->size, | |
2594 part2->builtin->size); | |
2595 partition_free (part2); | |
2596 partitions->ordered_remove (i + 1); | |
2597 } | |
2598 } | |
2599 | |
2600 /* Fuse PARTITIONS of LOOP if necessary before finalizing distribution. | |
2601 ALIAS_DDRS contains ddrs which need runtime alias check. */ | |
2602 | |
2603 static void | |
2604 finalize_partitions (struct loop *loop, vec<struct partition *> *partitions, | |
2605 vec<ddr_p> *alias_ddrs) | |
2606 { | |
2607 unsigned i; | |
2608 struct partition *partition, *a; | |
2609 | |
2610 if (partitions->length () == 1 | |
2611 || alias_ddrs->length () > 0) | |
2612 return; | |
2613 | |
2614 unsigned num_builtin = 0, num_normal = 0; | |
2615 bool same_type_p = true; | |
2616 enum partition_type type = ((*partitions)[0])->type; | |
2617 for (i = 0; partitions->iterate (i, &partition); ++i) | |
2618 { | |
2619 same_type_p &= (type == partition->type); | |
2620 if (partition->kind != PKIND_NORMAL) | |
2621 num_builtin++; | |
2622 else | |
2623 num_normal++; | |
2624 } | |
2625 | |
2626 /* Don't distribute current loop into too many loops given we don't have | |
2627 memory stream cost model. Be even more conservative in case of loop | |
2628 nest distribution. */ | |
2629 if ((same_type_p && num_builtin == 0) | |
2630 || (loop->inner != NULL | |
2631 && i >= NUM_PARTITION_THRESHOLD && num_normal > 1) | |
2632 || (loop->inner == NULL | |
2633 && i >= NUM_PARTITION_THRESHOLD && num_normal > num_builtin)) | |
2634 { | |
2635 a = (*partitions)[0]; | |
2636 for (i = 1; partitions->iterate (i, &partition); ++i) | |
2637 { | |
2638 partition_merge_into (NULL, a, partition, FUSE_FINALIZE); | |
2639 partition_free (partition); | |
2640 } | |
2641 partitions->truncate (1); | |
2642 } | |
2643 | |
2644 /* Fuse memset builtins if possible. */ | |
2645 if (partitions->length () > 1) | |
2646 fuse_memset_builtins (partitions); | |
2647 } | |
2648 | |
2649 /* Distributes the code from LOOP in such a way that producer statements | |
2650 are placed before consumer statements. Tries to separate only the | |
2651 statements from STMTS into separate loops. Returns the number of | |
2652 distributed loops. Set NB_CALLS to number of generated builtin calls. | |
2653 Set *DESTROY_P to whether LOOP needs to be destroyed. */ | |
0 | 2654 |
2655 static int | |
111 | 2656 distribute_loop (struct loop *loop, vec<gimple *> stmts, |
2657 control_dependences *cd, int *nb_calls, bool *destroy_p) | |
0 | 2658 { |
111 | 2659 ddrs_table = new hash_table<ddr_hasher> (389); |
0 | 2660 struct graph *rdg; |
111 | 2661 partition *partition; |
2662 bool any_builtin; | |
2663 int i, nbp; | |
2664 | |
2665 *destroy_p = false; | |
2666 *nb_calls = 0; | |
2667 loop_nest.create (0); | |
2668 if (!find_loop_nest (loop, &loop_nest)) | |
0 | 2669 { |
111 | 2670 loop_nest.release (); |
2671 delete ddrs_table; | |
2672 return 0; | |
0 | 2673 } |
2674 | |
111 | 2675 datarefs_vec.create (20); |
2676 rdg = build_rdg (loop, cd); | |
0 | 2677 if (!rdg) |
2678 { | |
2679 if (dump_file && (dump_flags & TDF_DETAILS)) | |
2680 fprintf (dump_file, | |
111 | 2681 "Loop %d not distributed: failed to build the RDG.\n", |
0 | 2682 loop->num); |
2683 | |
111 | 2684 loop_nest.release (); |
2685 free_data_refs (datarefs_vec); | |
2686 delete ddrs_table; | |
2687 return 0; | |
0 | 2688 } |
2689 | |
111 | 2690 if (datarefs_vec.length () > MAX_DATAREFS_NUM) |
2691 { | |
2692 if (dump_file && (dump_flags & TDF_DETAILS)) | |
2693 fprintf (dump_file, | |
2694 "Loop %d not distributed: too many memory references.\n", | |
2695 loop->num); | |
2696 | |
2697 free_rdg (rdg); | |
2698 loop_nest.release (); | |
2699 free_data_refs (datarefs_vec); | |
2700 delete ddrs_table; | |
2701 return 0; | |
2702 } | |
2703 | |
2704 data_reference_p dref; | |
2705 for (i = 0; datarefs_vec.iterate (i, &dref); ++i) | |
2706 dref->aux = (void *) (uintptr_t) i; | |
0 | 2707 |
2708 if (dump_file && (dump_flags & TDF_DETAILS)) | |
2709 dump_rdg (dump_file, rdg); | |
2710 | |
111 | 2711 auto_vec<struct partition *, 3> partitions; |
2712 rdg_build_partitions (rdg, stmts, &partitions); | |
2713 | |
2714 auto_vec<ddr_p> alias_ddrs; | |
2715 | |
2716 auto_bitmap stmt_in_all_partitions; | |
2717 bitmap_copy (stmt_in_all_partitions, partitions[0]->stmts); | |
2718 for (i = 1; partitions.iterate (i, &partition); ++i) | |
2719 bitmap_and_into (stmt_in_all_partitions, partitions[i]->stmts); | |
2720 | |
2721 any_builtin = false; | |
2722 FOR_EACH_VEC_ELT (partitions, i, partition) | |
2723 { | |
2724 classify_partition (loop, rdg, partition, stmt_in_all_partitions); | |
2725 any_builtin |= partition_builtin_p (partition); | |
2726 } | |
2727 | |
2728 /* If we are only distributing patterns but did not detect any, | |
2729 simply bail out. */ | |
2730 if (!flag_tree_loop_distribution | |
2731 && !any_builtin) | |
2732 { | |
2733 nbp = 0; | |
2734 goto ldist_done; | |
2735 } | |
2736 | |
2737 /* If we are only distributing patterns fuse all partitions that | |
2738 were not classified as builtins. This also avoids chopping | |
2739 a loop into pieces, separated by builtin calls. That is, we | |
2740 only want no or a single loop body remaining. */ | |
2741 struct partition *into; | |
2742 if (!flag_tree_loop_distribution) | |
0 | 2743 { |
111 | 2744 for (i = 0; partitions.iterate (i, &into); ++i) |
2745 if (!partition_builtin_p (into)) | |
2746 break; | |
2747 for (++i; partitions.iterate (i, &partition); ++i) | |
2748 if (!partition_builtin_p (partition)) | |
2749 { | |
2750 partition_merge_into (NULL, into, partition, FUSE_NON_BUILTIN); | |
2751 partitions.unordered_remove (i); | |
2752 partition_free (partition); | |
2753 i--; | |
2754 } | |
2755 } | |
2756 | |
2757 /* Due to limitations in the transform phase we have to fuse all | |
2758 reduction partitions into the last partition so the existing | |
2759 loop will contain all loop-closed PHI nodes. */ | |
2760 for (i = 0; partitions.iterate (i, &into); ++i) | |
2761 if (partition_reduction_p (into)) | |
2762 break; | |
2763 for (i = i + 1; partitions.iterate (i, &partition); ++i) | |
2764 if (partition_reduction_p (partition)) | |
2765 { | |
2766 partition_merge_into (rdg, into, partition, FUSE_REDUCTION); | |
2767 partitions.unordered_remove (i); | |
2768 partition_free (partition); | |
2769 i--; | |
2770 } | |
2771 | |
2772 /* Apply our simple cost model - fuse partitions with similar | |
2773 memory accesses. */ | |
2774 for (i = 0; partitions.iterate (i, &into); ++i) | |
2775 { | |
2776 bool changed = false; | |
2777 if (partition_builtin_p (into)) | |
2778 continue; | |
2779 for (int j = i + 1; | |
2780 partitions.iterate (j, &partition); ++j) | |
0 | 2781 { |
111 | 2782 if (share_memory_accesses (rdg, into, partition)) |
2783 { | |
2784 partition_merge_into (rdg, into, partition, FUSE_SHARE_REF); | |
2785 partitions.unordered_remove (j); | |
2786 partition_free (partition); | |
2787 j--; | |
2788 changed = true; | |
2789 } | |
2790 } | |
2791 /* If we fused 0 1 2 in step 1 to 0,2 1 as 0 and 2 have similar | |
2792 accesses when 1 and 2 have similar accesses but not 0 and 1 | |
2793 then in the next iteration we will fail to consider merging | |
2794 1 into 0,2. So try again if we did any merging into 0. */ | |
2795 if (changed) | |
2796 i--; | |
2797 } | |
2798 | |
2799 /* Build the partition dependency graph and fuse partitions in strong | |
2800 connected component. */ | |
2801 if (partitions.length () > 1) | |
2802 { | |
2803 /* Don't support loop nest distribution under runtime alias check | |
2804 since it's not likely to enable many vectorization opportunities. */ | |
2805 if (loop->inner) | |
2806 merge_dep_scc_partitions (rdg, &partitions, false); | |
2807 else | |
2808 { | |
2809 merge_dep_scc_partitions (rdg, &partitions, true); | |
2810 if (partitions.length () > 1) | |
2811 break_alias_scc_partitions (rdg, &partitions, &alias_ddrs); | |
0 | 2812 } |
2813 } | |
2814 | |
111 | 2815 finalize_partitions (loop, &partitions, &alias_ddrs); |
2816 | |
2817 nbp = partitions.length (); | |
2818 if (nbp == 0 | |
2819 || (nbp == 1 && !partition_builtin_p (partitions[0])) | |
2820 || (nbp > 1 && partition_contains_all_rw (rdg, partitions))) | |
2821 { | |
2822 nbp = 0; | |
2823 goto ldist_done; | |
2824 } | |
2825 | |
2826 if (version_for_distribution_p (&partitions, &alias_ddrs)) | |
2827 version_loop_by_alias_check (loop, &alias_ddrs); | |
2828 | |
2829 if (dump_file && (dump_flags & TDF_DETAILS)) | |
2830 { | |
2831 fprintf (dump_file, | |
2832 "distribute loop <%d> into partitions:\n", loop->num); | |
2833 dump_rdg_partitions (dump_file, partitions); | |
2834 } | |
2835 | |
2836 FOR_EACH_VEC_ELT (partitions, i, partition) | |
2837 { | |
2838 if (partition_builtin_p (partition)) | |
2839 (*nb_calls)++; | |
2840 *destroy_p |= generate_code_for_partition (loop, partition, i < nbp - 1); | |
2841 } | |
2842 | |
2843 ldist_done: | |
2844 loop_nest.release (); | |
2845 free_data_refs (datarefs_vec); | |
2846 for (hash_table<ddr_hasher>::iterator iter = ddrs_table->begin (); | |
2847 iter != ddrs_table->end (); ++iter) | |
2848 { | |
2849 free_dependence_relation (*iter); | |
2850 *iter = NULL; | |
2851 } | |
2852 delete ddrs_table; | |
2853 | |
2854 FOR_EACH_VEC_ELT (partitions, i, partition) | |
2855 partition_free (partition); | |
2856 | |
0 | 2857 free_rdg (rdg); |
111 | 2858 return nbp - *nb_calls; |
0 | 2859 } |
2860 | |
2861 /* Distribute all loops in the current function. */ | |
2862 | |
111 | 2863 namespace { |
2864 | |
2865 const pass_data pass_data_loop_distribution = | |
2866 { | |
2867 GIMPLE_PASS, /* type */ | |
2868 "ldist", /* name */ | |
2869 OPTGROUP_LOOP, /* optinfo_flags */ | |
2870 TV_TREE_LOOP_DISTRIBUTION, /* tv_id */ | |
2871 ( PROP_cfg | PROP_ssa ), /* properties_required */ | |
2872 0, /* properties_provided */ | |
2873 0, /* properties_destroyed */ | |
2874 0, /* todo_flags_start */ | |
2875 0, /* todo_flags_finish */ | |
2876 }; | |
2877 | |
2878 class pass_loop_distribution : public gimple_opt_pass | |
2879 { | |
2880 public: | |
2881 pass_loop_distribution (gcc::context *ctxt) | |
2882 : gimple_opt_pass (pass_data_loop_distribution, ctxt) | |
2883 {} | |
2884 | |
2885 /* opt_pass methods: */ | |
2886 virtual bool gate (function *) | |
2887 { | |
2888 return flag_tree_loop_distribution | |
2889 || flag_tree_loop_distribute_patterns; | |
2890 } | |
2891 | |
2892 virtual unsigned int execute (function *); | |
2893 | |
2894 }; // class pass_loop_distribution | |
2895 | |
2896 | |
2897 /* Given LOOP, this function records seed statements for distribution in | |
2898 WORK_LIST. Return false if there is nothing for distribution. */ | |
2899 | |
2900 static bool | |
2901 find_seed_stmts_for_distribution (struct loop *loop, vec<gimple *> *work_list) | |
2902 { | |
2903 basic_block *bbs = get_loop_body_in_dom_order (loop); | |
2904 | |
2905 /* Initialize the worklist with stmts we seed the partitions with. */ | |
2906 for (unsigned i = 0; i < loop->num_nodes; ++i) | |
2907 { | |
2908 for (gphi_iterator gsi = gsi_start_phis (bbs[i]); | |
2909 !gsi_end_p (gsi); gsi_next (&gsi)) | |
2910 { | |
2911 gphi *phi = gsi.phi (); | |
2912 if (virtual_operand_p (gimple_phi_result (phi))) | |
2913 continue; | |
2914 /* Distribute stmts which have defs that are used outside of | |
2915 the loop. */ | |
2916 if (!stmt_has_scalar_dependences_outside_loop (loop, phi)) | |
2917 continue; | |
2918 work_list->safe_push (phi); | |
2919 } | |
2920 for (gimple_stmt_iterator gsi = gsi_start_bb (bbs[i]); | |
2921 !gsi_end_p (gsi); gsi_next (&gsi)) | |
2922 { | |
2923 gimple *stmt = gsi_stmt (gsi); | |
2924 | |
2925 /* If there is a stmt with side-effects bail out - we | |
2926 cannot and should not distribute this loop. */ | |
2927 if (gimple_has_side_effects (stmt)) | |
2928 { | |
2929 free (bbs); | |
2930 return false; | |
2931 } | |
2932 | |
2933 /* Distribute stmts which have defs that are used outside of | |
2934 the loop. */ | |
2935 if (stmt_has_scalar_dependences_outside_loop (loop, stmt)) | |
2936 ; | |
2937 /* Otherwise only distribute stores for now. */ | |
2938 else if (!gimple_vdef (stmt)) | |
2939 continue; | |
2940 | |
2941 work_list->safe_push (stmt); | |
2942 } | |
2943 } | |
2944 free (bbs); | |
2945 return work_list->length () > 0; | |
2946 } | |
2947 | |
2948 /* Given innermost LOOP, return the outermost enclosing loop that forms a | |
2949 perfect loop nest. */ | |
2950 | |
2951 static struct loop * | |
2952 prepare_perfect_loop_nest (struct loop *loop) | |
2953 { | |
2954 struct loop *outer = loop_outer (loop); | |
2955 tree niters = number_of_latch_executions (loop); | |
2956 | |
2957 /* TODO: We only support the innermost 2-level loop nest distribution | |
2958 because of compilation time issue for now. This should be relaxed | |
2959 in the future. */ | |
2960 while (loop->inner == NULL | |
2961 && loop_outer (outer) | |
2962 && outer->inner == loop && loop->next == NULL | |
2963 && single_exit (outer) | |
2964 && optimize_loop_for_speed_p (outer) | |
2965 && !chrec_contains_symbols_defined_in_loop (niters, outer->num) | |
2966 && (niters = number_of_latch_executions (outer)) != NULL_TREE | |
2967 && niters != chrec_dont_know) | |
2968 { | |
2969 loop = outer; | |
2970 outer = loop_outer (loop); | |
2971 } | |
2972 | |
2973 return loop; | |
2974 } | |
2975 | |
2976 unsigned int | |
2977 pass_loop_distribution::execute (function *fun) | |
0 | 2978 { |
2979 struct loop *loop; | |
111 | 2980 bool changed = false; |
2981 basic_block bb; | |
2982 control_dependences *cd = NULL; | |
2983 auto_vec<loop_p> loops_to_be_destroyed; | |
2984 | |
2985 if (number_of_loops (fun) <= 1) | |
2986 return 0; | |
2987 | |
2988 /* Compute topological order for basic blocks. Topological order is | |
2989 needed because data dependence is computed for data references in | |
2990 lexicographical order. */ | |
2991 if (bb_top_order_index == NULL) | |
0 | 2992 { |
111 | 2993 int rpo_num; |
2994 int *rpo = XNEWVEC (int, last_basic_block_for_fn (cfun)); | |
2995 | |
2996 bb_top_order_index = XNEWVEC (int, last_basic_block_for_fn (cfun)); | |
2997 bb_top_order_index_size = last_basic_block_for_fn (cfun); | |
2998 rpo_num = pre_and_rev_post_order_compute_fn (cfun, NULL, rpo, true); | |
2999 for (int i = 0; i < rpo_num; i++) | |
3000 bb_top_order_index[rpo[i]] = i; | |
3001 | |
3002 free (rpo); | |
3003 } | |
3004 | |
3005 FOR_ALL_BB_FN (bb, fun) | |
3006 { | |
3007 gimple_stmt_iterator gsi; | |
3008 for (gsi = gsi_start_phis (bb); !gsi_end_p (gsi); gsi_next (&gsi)) | |
3009 gimple_set_uid (gsi_stmt (gsi), -1); | |
3010 for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi)) | |
3011 gimple_set_uid (gsi_stmt (gsi), -1); | |
3012 } | |
3013 | |
3014 /* We can at the moment only distribute non-nested loops, thus restrict | |
3015 walking to innermost loops. */ | |
3016 FOR_EACH_LOOP (loop, LI_ONLY_INNERMOST) | |
3017 { | |
3018 /* Don't distribute multiple exit edges loop, or cold loop. */ | |
3019 if (!single_exit (loop) | |
3020 || !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
|
3021 continue; |
0 | 3022 |
111 | 3023 /* Don't distribute loop if niters is unknown. */ |
3024 tree niters = number_of_latch_executions (loop); | |
3025 if (niters == NULL_TREE || niters == chrec_dont_know) | |
3026 continue; | |
3027 | |
3028 /* Get the perfect loop nest for distribution. */ | |
3029 loop = prepare_perfect_loop_nest (loop); | |
3030 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
|
3031 { |
111 | 3032 auto_vec<gimple *> work_list; |
3033 if (!find_seed_stmts_for_distribution (loop, &work_list)) | |
3034 break; | |
3035 | |
3036 const char *str = loop->inner ? " nest" : ""; | |
3037 location_t loc = find_loop_location (loop); | |
3038 if (!cd) | |
3039 { | |
3040 calculate_dominance_info (CDI_DOMINATORS); | |
3041 calculate_dominance_info (CDI_POST_DOMINATORS); | |
3042 cd = new control_dependences (); | |
3043 free_dominance_info (CDI_POST_DOMINATORS); | |
3044 } | |
3045 | |
3046 bool destroy_p; | |
3047 int nb_generated_loops, nb_generated_calls; | |
3048 nb_generated_loops = distribute_loop (loop, work_list, cd, | |
3049 &nb_generated_calls, | |
3050 &destroy_p); | |
3051 if (destroy_p) | |
3052 loops_to_be_destroyed.safe_push (loop); | |
3053 | |
3054 if (nb_generated_loops + nb_generated_calls > 0) | |
3055 { | |
3056 changed = true; | |
3057 dump_printf_loc (MSG_OPTIMIZED_LOCATIONS, | |
3058 loc, "Loop%s %d distributed: split to %d loops " | |
3059 "and %d library calls.\n", str, loop->num, | |
3060 nb_generated_loops, nb_generated_calls); | |
3061 | |
3062 break; | |
3063 } | |
3064 | |
3065 if (dump_file && (dump_flags & TDF_DETAILS)) | |
3066 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
|
3067 } |
111 | 3068 } |
3069 | |
3070 if (cd) | |
3071 delete cd; | |
3072 | |
3073 if (bb_top_order_index != NULL) | |
3074 { | |
3075 free (bb_top_order_index); | |
3076 bb_top_order_index = NULL; | |
3077 bb_top_order_index_size = 0; | |
0 | 3078 } |
3079 | |
111 | 3080 if (changed) |
3081 { | |
3082 /* Destroy loop bodies that could not be reused. Do this late as we | |
3083 otherwise can end up refering to stale data in control dependences. */ | |
3084 unsigned i; | |
3085 FOR_EACH_VEC_ELT (loops_to_be_destroyed, i, loop) | |
3086 destroy_loop (loop); | |
3087 | |
3088 /* Cached scalar evolutions now may refer to wrong or non-existing | |
3089 loops. */ | |
3090 scev_reset_htab (); | |
3091 mark_virtual_operands_for_renaming (fun); | |
3092 rewrite_into_loop_closed_ssa (NULL, TODO_update_ssa); | |
3093 } | |
3094 | |
3095 checking_verify_loop_structure (); | |
3096 | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
3097 return 0; |
0 | 3098 } |
3099 | |
111 | 3100 } // anon namespace |
3101 | |
3102 gimple_opt_pass * | |
3103 make_pass_loop_distribution (gcc::context *ctxt) | |
0 | 3104 { |
111 | 3105 return new pass_loop_distribution (ctxt); |
3106 } |