]>
Commit | Line | Data |
---|---|---|
dea61d92 | 1 | /* Loop distribution. |
cbe34bb5 | 2 | Copyright (C) 2006-2017 Free Software Foundation, Inc. |
dea61d92 SP |
3 | Contributed by Georges-Andre Silber <Georges-Andre.Silber@ensmp.fr> |
4 | and Sebastian Pop <sebastian.pop@amd.com>. | |
5 | ||
6 | This file is part of GCC. | |
b8698a0f | 7 | |
dea61d92 SP |
8 | GCC is free software; you can redistribute it and/or modify it |
9 | under the terms of the GNU General Public License as published by the | |
10 | Free Software Foundation; either version 3, or (at your option) any | |
11 | later version. | |
b8698a0f | 12 | |
dea61d92 SP |
13 | GCC is distributed in the hope that it will be useful, but WITHOUT |
14 | ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or | |
15 | FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License | |
16 | for more details. | |
b8698a0f | 17 | |
dea61d92 SP |
18 | You should have received a copy of the GNU General Public License |
19 | along with GCC; see the file COPYING3. If not see | |
20 | <http://www.gnu.org/licenses/>. */ | |
21 | ||
22 | /* This pass performs loop distribution: for example, the loop | |
23 | ||
24 | |DO I = 2, N | |
25 | | A(I) = B(I) + C | |
26 | | D(I) = A(I-1)*E | |
27 | |ENDDO | |
28 | ||
b8698a0f | 29 | is transformed to |
dea61d92 SP |
30 | |
31 | |DOALL I = 2, N | |
32 | | A(I) = B(I) + C | |
33 | |ENDDO | |
34 | | | |
35 | |DOALL I = 2, N | |
36 | | D(I) = A(I-1)*E | |
37 | |ENDDO | |
38 | ||
a8745cc2 BC |
39 | Loop distribution is the dual of loop fusion. It separates statements |
40 | of a loop (or loop nest) into multiple loops (or loop nests) with the | |
41 | same loop header. The major goal is to separate statements which may | |
42 | be vectorized from those that can't. This pass implements distribution | |
43 | in the following steps: | |
44 | ||
45 | 1) Seed partitions with specific type statements. For now we support | |
46 | two types seed statements: statement defining variable used outside | |
47 | of loop; statement storing to memory. | |
48 | 2) Build reduced dependence graph (RDG) for loop to be distributed. | |
49 | The vertices (RDG:V) model all statements in the loop and the edges | |
50 | (RDG:E) model flow and control dependencies between statements. | |
51 | 3) Apart from RDG, compute data dependencies between memory references. | |
52 | 4) Starting from seed statement, build up partition by adding depended | |
53 | statements according to RDG's dependence information. Partition is | |
54 | classified as parallel type if it can be executed paralleled; or as | |
55 | sequential type if it can't. Parallel type partition is further | |
56 | classified as different builtin kinds if it can be implemented as | |
57 | builtin function calls. | |
58 | 5) Build partition dependence graph (PG) based on data dependencies. | |
59 | The vertices (PG:V) model all partitions and the edges (PG:E) model | |
60 | all data dependencies between every partitions pair. In general, | |
61 | data dependence is either compilation time known or unknown. In C | |
62 | family languages, there exists quite amount compilation time unknown | |
63 | dependencies because of possible alias relation of data references. | |
64 | We categorize PG's edge to two types: "true" edge that represents | |
65 | compilation time known data dependencies; "alias" edge for all other | |
66 | data dependencies. | |
67 | 6) Traverse subgraph of PG as if all "alias" edges don't exist. Merge | |
68 | partitions in each strong connected component (SCC) correspondingly. | |
69 | Build new PG for merged partitions. | |
70 | 7) Traverse PG again and this time with both "true" and "alias" edges | |
71 | included. We try to break SCCs by removing some edges. Because | |
72 | SCCs by "true" edges are all fused in step 6), we can break SCCs | |
73 | by removing some "alias" edges. It's NP-hard to choose optimal | |
74 | edge set, fortunately simple approximation is good enough for us | |
75 | given the small problem scale. | |
76 | 8) Collect all data dependencies of the removed "alias" edges. Create | |
77 | runtime alias checks for collected data dependencies. | |
78 | 9) Version loop under the condition of runtime alias checks. Given | |
79 | loop distribution generally introduces additional overhead, it is | |
80 | only useful if vectorization is achieved in distributed loop. We | |
81 | version loop with internal function call IFN_LOOP_DIST_ALIAS. If | |
82 | no distributed loop can be vectorized, we simply remove distributed | |
83 | loops and recover to the original one. | |
84 | ||
85 | TODO: | |
86 | 1) We only distribute innermost loops now. This pass should handle loop | |
87 | 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. */ | |
dea61d92 SP |
91 | |
92 | #include "config.h" | |
93 | #include "system.h" | |
94 | #include "coretypes.h" | |
c7131fb2 | 95 | #include "backend.h" |
40e23961 | 96 | #include "tree.h" |
c7131fb2 | 97 | #include "gimple.h" |
957060b5 AM |
98 | #include "cfghooks.h" |
99 | #include "tree-pass.h" | |
c7131fb2 | 100 | #include "ssa.h" |
957060b5 | 101 | #include "gimple-pretty-print.h" |
c7131fb2 | 102 | #include "fold-const.h" |
60393bbc | 103 | #include "cfganal.h" |
5be5c238 | 104 | #include "gimple-iterator.h" |
18f429e2 | 105 | #include "gimplify-me.h" |
d8a2d370 | 106 | #include "stor-layout.h" |
442b4905 | 107 | #include "tree-cfg.h" |
e28030cf | 108 | #include "tree-ssa-loop-manip.h" |
442b4905 AM |
109 | #include "tree-ssa-loop.h" |
110 | #include "tree-into-ssa.h" | |
7a300452 | 111 | #include "tree-ssa.h" |
dea61d92 | 112 | #include "cfgloop.h" |
dea61d92 | 113 | #include "tree-scalar-evolution.h" |
9fafb14a | 114 | #include "params.h" |
826a536d | 115 | #include "tree-vectorizer.h" |
80ab0b19 RB |
116 | |
117 | ||
9fafb14a BC |
118 | #define MAX_DATAREFS_NUM \ |
119 | ((unsigned) PARAM_VALUE (PARAM_LOOP_MAX_DATAREFS_FOR_DATADEPS)) | |
120 | ||
17c5cbdf BC |
121 | /* Hashtable helpers. */ |
122 | ||
123 | struct ddr_hasher : nofree_ptr_hash <struct data_dependence_relation> | |
124 | { | |
125 | static inline hashval_t hash (const data_dependence_relation *); | |
126 | static inline bool equal (const data_dependence_relation *, | |
127 | const data_dependence_relation *); | |
128 | }; | |
129 | ||
130 | /* Hash function for data dependence. */ | |
131 | ||
132 | inline hashval_t | |
133 | ddr_hasher::hash (const data_dependence_relation *ddr) | |
134 | { | |
135 | inchash::hash h; | |
136 | h.add_ptr (DDR_A (ddr)); | |
137 | h.add_ptr (DDR_B (ddr)); | |
138 | return h.end (); | |
139 | } | |
140 | ||
141 | /* Hash table equality function for data dependence. */ | |
142 | ||
143 | inline bool | |
144 | ddr_hasher::equal (const data_dependence_relation *ddr1, | |
145 | const data_dependence_relation *ddr2) | |
146 | { | |
147 | return (DDR_A (ddr1) == DDR_A (ddr2) && DDR_B (ddr1) == DDR_B (ddr2)); | |
148 | } | |
149 | ||
4084ea5f BC |
150 | /* The loop (nest) to be distributed. */ |
151 | static vec<loop_p> loop_nest; | |
152 | ||
9fafb14a BC |
153 | /* Vector of data references in the loop to be distributed. */ |
154 | static vec<data_reference_p> datarefs_vec; | |
155 | ||
156 | /* Store index of data reference in aux field. */ | |
157 | #define DR_INDEX(dr) ((uintptr_t) (dr)->aux) | |
158 | ||
17c5cbdf BC |
159 | /* Hash table for data dependence relation in the loop to be distributed. */ |
160 | static hash_table<ddr_hasher> ddrs_table (389); | |
161 | ||
162 | ||
80ab0b19 | 163 | /* A Reduced Dependence Graph (RDG) vertex representing a statement. */ |
526ceb68 | 164 | struct rdg_vertex |
80ab0b19 RB |
165 | { |
166 | /* The statement represented by this vertex. */ | |
355fe088 | 167 | gimple *stmt; |
80ab0b19 RB |
168 | |
169 | /* Vector of data-references in this statement. */ | |
170 | vec<data_reference_p> datarefs; | |
171 | ||
172 | /* True when the statement contains a write to memory. */ | |
173 | bool has_mem_write; | |
174 | ||
175 | /* True when the statement contains a read from memory. */ | |
176 | bool has_mem_reads; | |
526ceb68 | 177 | }; |
80ab0b19 RB |
178 | |
179 | #define RDGV_STMT(V) ((struct rdg_vertex *) ((V)->data))->stmt | |
180 | #define RDGV_DATAREFS(V) ((struct rdg_vertex *) ((V)->data))->datarefs | |
181 | #define RDGV_HAS_MEM_WRITE(V) ((struct rdg_vertex *) ((V)->data))->has_mem_write | |
182 | #define RDGV_HAS_MEM_READS(V) ((struct rdg_vertex *) ((V)->data))->has_mem_reads | |
183 | #define RDG_STMT(RDG, I) RDGV_STMT (&(RDG->vertices[I])) | |
184 | #define RDG_DATAREFS(RDG, I) RDGV_DATAREFS (&(RDG->vertices[I])) | |
185 | #define RDG_MEM_WRITE_STMT(RDG, I) RDGV_HAS_MEM_WRITE (&(RDG->vertices[I])) | |
186 | #define RDG_MEM_READS_STMT(RDG, I) RDGV_HAS_MEM_READS (&(RDG->vertices[I])) | |
187 | ||
188 | /* Data dependence type. */ | |
189 | ||
190 | enum rdg_dep_type | |
191 | { | |
192 | /* Read After Write (RAW). */ | |
193 | flow_dd = 'f', | |
194 | ||
36875e8f RB |
195 | /* Control dependence (execute conditional on). */ |
196 | control_dd = 'c' | |
80ab0b19 RB |
197 | }; |
198 | ||
199 | /* Dependence information attached to an edge of the RDG. */ | |
200 | ||
526ceb68 | 201 | struct rdg_edge |
80ab0b19 RB |
202 | { |
203 | /* Type of the dependence. */ | |
204 | enum rdg_dep_type type; | |
526ceb68 | 205 | }; |
80ab0b19 RB |
206 | |
207 | #define RDGE_TYPE(E) ((struct rdg_edge *) ((E)->data))->type | |
80ab0b19 | 208 | |
80ab0b19 RB |
209 | /* Dump vertex I in RDG to FILE. */ |
210 | ||
211 | static void | |
212 | dump_rdg_vertex (FILE *file, struct graph *rdg, int i) | |
213 | { | |
214 | struct vertex *v = &(rdg->vertices[i]); | |
215 | struct graph_edge *e; | |
216 | ||
217 | fprintf (file, "(vertex %d: (%s%s) (in:", i, | |
218 | RDG_MEM_WRITE_STMT (rdg, i) ? "w" : "", | |
219 | RDG_MEM_READS_STMT (rdg, i) ? "r" : ""); | |
220 | ||
221 | if (v->pred) | |
222 | for (e = v->pred; e; e = e->pred_next) | |
223 | fprintf (file, " %d", e->src); | |
224 | ||
225 | fprintf (file, ") (out:"); | |
226 | ||
227 | if (v->succ) | |
228 | for (e = v->succ; e; e = e->succ_next) | |
229 | fprintf (file, " %d", e->dest); | |
230 | ||
231 | fprintf (file, ")\n"); | |
232 | print_gimple_stmt (file, RDGV_STMT (v), 0, TDF_VOPS|TDF_MEMSYMS); | |
233 | fprintf (file, ")\n"); | |
234 | } | |
235 | ||
236 | /* Call dump_rdg_vertex on stderr. */ | |
237 | ||
238 | DEBUG_FUNCTION void | |
239 | debug_rdg_vertex (struct graph *rdg, int i) | |
240 | { | |
241 | dump_rdg_vertex (stderr, rdg, i); | |
242 | } | |
243 | ||
80ab0b19 RB |
244 | /* Dump the reduced dependence graph RDG to FILE. */ |
245 | ||
246 | static void | |
247 | dump_rdg (FILE *file, struct graph *rdg) | |
248 | { | |
80ab0b19 | 249 | fprintf (file, "(rdg\n"); |
2fd5894f RB |
250 | for (int i = 0; i < rdg->n_vertices; i++) |
251 | dump_rdg_vertex (file, rdg, i); | |
80ab0b19 | 252 | fprintf (file, ")\n"); |
80ab0b19 RB |
253 | } |
254 | ||
255 | /* Call dump_rdg on stderr. */ | |
256 | ||
257 | DEBUG_FUNCTION void | |
258 | debug_rdg (struct graph *rdg) | |
259 | { | |
260 | dump_rdg (stderr, rdg); | |
261 | } | |
262 | ||
263 | static void | |
264 | dot_rdg_1 (FILE *file, struct graph *rdg) | |
265 | { | |
266 | int i; | |
174ec470 RB |
267 | pretty_printer buffer; |
268 | pp_needs_newline (&buffer) = false; | |
269 | buffer.buffer->stream = file; | |
80ab0b19 RB |
270 | |
271 | fprintf (file, "digraph RDG {\n"); | |
272 | ||
273 | for (i = 0; i < rdg->n_vertices; i++) | |
274 | { | |
275 | struct vertex *v = &(rdg->vertices[i]); | |
276 | struct graph_edge *e; | |
277 | ||
174ec470 RB |
278 | fprintf (file, "%d [label=\"[%d] ", i, i); |
279 | pp_gimple_stmt_1 (&buffer, RDGV_STMT (v), 0, TDF_SLIM); | |
280 | pp_flush (&buffer); | |
281 | fprintf (file, "\"]\n"); | |
282 | ||
80ab0b19 RB |
283 | /* Highlight reads from memory. */ |
284 | if (RDG_MEM_READS_STMT (rdg, i)) | |
285 | fprintf (file, "%d [style=filled, fillcolor=green]\n", i); | |
286 | ||
287 | /* Highlight stores to memory. */ | |
288 | if (RDG_MEM_WRITE_STMT (rdg, i)) | |
289 | fprintf (file, "%d [style=filled, fillcolor=red]\n", i); | |
290 | ||
291 | if (v->succ) | |
292 | for (e = v->succ; e; e = e->succ_next) | |
293 | switch (RDGE_TYPE (e)) | |
294 | { | |
80ab0b19 RB |
295 | case flow_dd: |
296 | /* These are the most common dependences: don't print these. */ | |
297 | fprintf (file, "%d -> %d \n", i, e->dest); | |
298 | break; | |
299 | ||
36875e8f RB |
300 | case control_dd: |
301 | fprintf (file, "%d -> %d [label=control] \n", i, e->dest); | |
302 | break; | |
303 | ||
80ab0b19 RB |
304 | default: |
305 | gcc_unreachable (); | |
306 | } | |
307 | } | |
308 | ||
309 | fprintf (file, "}\n\n"); | |
310 | } | |
311 | ||
312 | /* Display the Reduced Dependence Graph using dotty. */ | |
313 | ||
314 | DEBUG_FUNCTION void | |
315 | dot_rdg (struct graph *rdg) | |
316 | { | |
174ec470 | 317 | /* When debugging, you may want to enable the following code. */ |
b6d94045 | 318 | #ifdef HAVE_POPEN |
c3284718 | 319 | FILE *file = popen ("dot -Tx11", "w"); |
174ec470 RB |
320 | if (!file) |
321 | return; | |
80ab0b19 | 322 | dot_rdg_1 (file, rdg); |
174ec470 RB |
323 | fflush (file); |
324 | close (fileno (file)); | |
325 | pclose (file); | |
80ab0b19 RB |
326 | #else |
327 | dot_rdg_1 (stderr, rdg); | |
328 | #endif | |
329 | } | |
330 | ||
331 | /* Returns the index of STMT in RDG. */ | |
332 | ||
333 | static int | |
355fe088 | 334 | rdg_vertex_for_stmt (struct graph *rdg ATTRIBUTE_UNUSED, gimple *stmt) |
80ab0b19 RB |
335 | { |
336 | int index = gimple_uid (stmt); | |
337 | gcc_checking_assert (index == -1 || RDG_STMT (rdg, index) == stmt); | |
338 | return index; | |
339 | } | |
340 | ||
80ab0b19 RB |
341 | /* Creates dependence edges in RDG for all the uses of DEF. IDEF is |
342 | the index of DEF in RDG. */ | |
343 | ||
344 | static void | |
345 | create_rdg_edges_for_scalar (struct graph *rdg, tree def, int idef) | |
346 | { | |
347 | use_operand_p imm_use_p; | |
348 | imm_use_iterator iterator; | |
349 | ||
350 | FOR_EACH_IMM_USE_FAST (imm_use_p, iterator, def) | |
351 | { | |
352 | struct graph_edge *e; | |
353 | int use = rdg_vertex_for_stmt (rdg, USE_STMT (imm_use_p)); | |
354 | ||
355 | if (use < 0) | |
356 | continue; | |
357 | ||
358 | e = add_edge (rdg, idef, use); | |
359 | e->data = XNEW (struct rdg_edge); | |
360 | RDGE_TYPE (e) = flow_dd; | |
80ab0b19 RB |
361 | } |
362 | } | |
363 | ||
36875e8f RB |
364 | /* Creates an edge for the control dependences of BB to the vertex V. */ |
365 | ||
366 | static void | |
367 | create_edge_for_control_dependence (struct graph *rdg, basic_block bb, | |
368 | int v, control_dependences *cd) | |
369 | { | |
370 | bitmap_iterator bi; | |
371 | unsigned edge_n; | |
372 | EXECUTE_IF_SET_IN_BITMAP (cd->get_edges_dependent_on (bb->index), | |
373 | 0, edge_n, bi) | |
374 | { | |
30fd2977 | 375 | basic_block cond_bb = cd->get_edge_src (edge_n); |
355fe088 | 376 | gimple *stmt = last_stmt (cond_bb); |
36875e8f RB |
377 | if (stmt && is_ctrl_stmt (stmt)) |
378 | { | |
379 | struct graph_edge *e; | |
380 | int c = rdg_vertex_for_stmt (rdg, stmt); | |
381 | if (c < 0) | |
382 | continue; | |
383 | ||
384 | e = add_edge (rdg, c, v); | |
385 | e->data = XNEW (struct rdg_edge); | |
386 | RDGE_TYPE (e) = control_dd; | |
36875e8f RB |
387 | } |
388 | } | |
389 | } | |
390 | ||
80ab0b19 RB |
391 | /* Creates the edges of the reduced dependence graph RDG. */ |
392 | ||
393 | static void | |
447f3223 | 394 | create_rdg_flow_edges (struct graph *rdg) |
80ab0b19 RB |
395 | { |
396 | int i; | |
80ab0b19 RB |
397 | def_operand_p def_p; |
398 | ssa_op_iter iter; | |
399 | ||
80ab0b19 RB |
400 | for (i = 0; i < rdg->n_vertices; i++) |
401 | FOR_EACH_PHI_OR_STMT_DEF (def_p, RDG_STMT (rdg, i), | |
402 | iter, SSA_OP_DEF) | |
403 | create_rdg_edges_for_scalar (rdg, DEF_FROM_PTR (def_p), i); | |
447f3223 | 404 | } |
36875e8f | 405 | |
447f3223 RB |
406 | /* Creates the edges of the reduced dependence graph RDG. */ |
407 | ||
408 | static void | |
b71b7a8e | 409 | create_rdg_cd_edges (struct graph *rdg, control_dependences *cd, loop_p loop) |
447f3223 RB |
410 | { |
411 | int i; | |
412 | ||
413 | for (i = 0; i < rdg->n_vertices; i++) | |
414 | { | |
355fe088 | 415 | gimple *stmt = RDG_STMT (rdg, i); |
447f3223 RB |
416 | if (gimple_code (stmt) == GIMPLE_PHI) |
417 | { | |
418 | edge_iterator ei; | |
419 | edge e; | |
420 | FOR_EACH_EDGE (e, ei, gimple_bb (stmt)->preds) | |
b71b7a8e | 421 | if (flow_bb_inside_loop_p (loop, e->src)) |
36875e8f | 422 | create_edge_for_control_dependence (rdg, e->src, i, cd); |
447f3223 RB |
423 | } |
424 | else | |
425 | create_edge_for_control_dependence (rdg, gimple_bb (stmt), i, cd); | |
426 | } | |
80ab0b19 RB |
427 | } |
428 | ||
429 | /* Build the vertices of the reduced dependence graph RDG. Return false | |
430 | if that failed. */ | |
431 | ||
432 | static bool | |
9fafb14a | 433 | create_rdg_vertices (struct graph *rdg, vec<gimple *> stmts, loop_p loop) |
80ab0b19 RB |
434 | { |
435 | int i; | |
355fe088 | 436 | gimple *stmt; |
80ab0b19 RB |
437 | |
438 | FOR_EACH_VEC_ELT (stmts, i, stmt) | |
439 | { | |
440 | struct vertex *v = &(rdg->vertices[i]); | |
441 | ||
442 | /* Record statement to vertex mapping. */ | |
443 | gimple_set_uid (stmt, i); | |
444 | ||
445 | v->data = XNEW (struct rdg_vertex); | |
446 | RDGV_STMT (v) = stmt; | |
447 | RDGV_DATAREFS (v).create (0); | |
448 | RDGV_HAS_MEM_WRITE (v) = false; | |
449 | RDGV_HAS_MEM_READS (v) = false; | |
450 | if (gimple_code (stmt) == GIMPLE_PHI) | |
451 | continue; | |
452 | ||
9fafb14a BC |
453 | unsigned drp = datarefs_vec.length (); |
454 | if (!find_data_references_in_stmt (loop, stmt, &datarefs_vec)) | |
80ab0b19 | 455 | return false; |
9fafb14a | 456 | for (unsigned j = drp; j < datarefs_vec.length (); ++j) |
80ab0b19 | 457 | { |
9fafb14a | 458 | data_reference_p dr = datarefs_vec[j]; |
80ab0b19 RB |
459 | if (DR_IS_READ (dr)) |
460 | RDGV_HAS_MEM_READS (v) = true; | |
461 | else | |
462 | RDGV_HAS_MEM_WRITE (v) = true; | |
463 | RDGV_DATAREFS (v).safe_push (dr); | |
464 | } | |
465 | } | |
466 | return true; | |
467 | } | |
468 | ||
3be57c56 BC |
469 | /* Array mapping basic block's index to its topological order. */ |
470 | static int *bb_top_order_index; | |
471 | /* And size of the array. */ | |
472 | static int bb_top_order_index_size; | |
473 | ||
474 | /* If X has a smaller topological sort number than Y, returns -1; | |
475 | if greater, returns 1. */ | |
476 | ||
477 | static int | |
478 | bb_top_order_cmp (const void *x, const void *y) | |
479 | { | |
480 | basic_block bb1 = *(const basic_block *) x; | |
481 | basic_block bb2 = *(const basic_block *) y; | |
482 | ||
483 | gcc_assert (bb1->index < bb_top_order_index_size | |
484 | && bb2->index < bb_top_order_index_size); | |
485 | gcc_assert (bb1 == bb2 | |
486 | || bb_top_order_index[bb1->index] | |
487 | != bb_top_order_index[bb2->index]); | |
488 | ||
489 | return (bb_top_order_index[bb1->index] - bb_top_order_index[bb2->index]); | |
490 | } | |
491 | ||
492 | /* Initialize STMTS with all the statements of LOOP. We use topological | |
493 | order to discover all statements. The order is important because | |
494 | generate_loops_for_partition is using the same traversal for identifying | |
495 | statements in loop copies. */ | |
80ab0b19 RB |
496 | |
497 | static void | |
355fe088 | 498 | stmts_from_loop (struct loop *loop, vec<gimple *> *stmts) |
80ab0b19 RB |
499 | { |
500 | unsigned int i; | |
3be57c56 | 501 | basic_block *bbs = get_loop_body_in_custom_order (loop, bb_top_order_cmp); |
80ab0b19 RB |
502 | |
503 | for (i = 0; i < loop->num_nodes; i++) | |
504 | { | |
505 | basic_block bb = bbs[i]; | |
80ab0b19 | 506 | |
538dd0b7 DM |
507 | for (gphi_iterator bsi = gsi_start_phis (bb); !gsi_end_p (bsi); |
508 | gsi_next (&bsi)) | |
509 | if (!virtual_operand_p (gimple_phi_result (bsi.phi ()))) | |
510 | stmts->safe_push (bsi.phi ()); | |
80ab0b19 | 511 | |
538dd0b7 DM |
512 | for (gimple_stmt_iterator bsi = gsi_start_bb (bb); !gsi_end_p (bsi); |
513 | gsi_next (&bsi)) | |
80ab0b19 | 514 | { |
355fe088 | 515 | gimple *stmt = gsi_stmt (bsi); |
80ab0b19 RB |
516 | if (gimple_code (stmt) != GIMPLE_LABEL && !is_gimple_debug (stmt)) |
517 | stmts->safe_push (stmt); | |
518 | } | |
519 | } | |
520 | ||
521 | free (bbs); | |
522 | } | |
523 | ||
80ab0b19 RB |
524 | /* Free the reduced dependence graph RDG. */ |
525 | ||
526 | static void | |
527 | free_rdg (struct graph *rdg) | |
528 | { | |
529 | int i; | |
530 | ||
531 | for (i = 0; i < rdg->n_vertices; i++) | |
532 | { | |
533 | struct vertex *v = &(rdg->vertices[i]); | |
534 | struct graph_edge *e; | |
535 | ||
536 | for (e = v->succ; e; e = e->succ_next) | |
447f3223 | 537 | free (e->data); |
80ab0b19 RB |
538 | |
539 | if (v->data) | |
540 | { | |
541 | gimple_set_uid (RDGV_STMT (v), -1); | |
9fafb14a | 542 | (RDGV_DATAREFS (v)).release (); |
80ab0b19 RB |
543 | free (v->data); |
544 | } | |
545 | } | |
546 | ||
547 | free_graph (rdg); | |
548 | } | |
549 | ||
4084ea5f BC |
550 | /* Build the Reduced Dependence Graph (RDG) with one vertex per statement of |
551 | LOOP, and one edge per flow dependence or control dependence from control | |
9fafb14a BC |
552 | dependence CD. During visiting each statement, data references are also |
553 | collected and recorded in global data DATAREFS_VEC. */ | |
80ab0b19 RB |
554 | |
555 | static struct graph * | |
4084ea5f | 556 | build_rdg (struct loop *loop, control_dependences *cd) |
80ab0b19 RB |
557 | { |
558 | struct graph *rdg; | |
80ab0b19 | 559 | |
97463b2b | 560 | /* Create the RDG vertices from the stmts of the loop nest. */ |
355fe088 | 561 | auto_vec<gimple *, 10> stmts; |
4084ea5f | 562 | stmts_from_loop (loop, &stmts); |
24f161fd | 563 | rdg = new_graph (stmts.length ()); |
9fafb14a | 564 | if (!create_rdg_vertices (rdg, stmts, loop)) |
80ab0b19 | 565 | { |
80ab0b19 RB |
566 | free_rdg (rdg); |
567 | return NULL; | |
568 | } | |
569 | stmts.release (); | |
97463b2b | 570 | |
447f3223 RB |
571 | create_rdg_flow_edges (rdg); |
572 | if (cd) | |
4084ea5f | 573 | create_rdg_cd_edges (rdg, cd, loop); |
447f3223 | 574 | |
80ab0b19 RB |
575 | return rdg; |
576 | } | |
577 | ||
80ab0b19 | 578 | |
f1eb4621 | 579 | /* Kind of distributed loop. */ |
b9fc0497 | 580 | enum partition_kind { |
510d73a0 | 581 | PKIND_NORMAL, PKIND_MEMSET, PKIND_MEMCPY, PKIND_MEMMOVE |
b9fc0497 | 582 | }; |
30d55936 | 583 | |
f1eb4621 BC |
584 | /* Type of distributed loop. */ |
585 | enum partition_type { | |
586 | /* The distributed loop can be executed parallelly. */ | |
587 | PTYPE_PARALLEL = 0, | |
588 | /* The distributed loop has to be executed sequentially. */ | |
589 | PTYPE_SEQUENTIAL | |
590 | }; | |
591 | ||
a7a44c07 | 592 | /* Partition for loop distribution. */ |
526ceb68 | 593 | struct partition |
c61f8985 | 594 | { |
a7a44c07 | 595 | /* Statements of the partition. */ |
c61f8985 | 596 | bitmap stmts; |
a7a44c07 | 597 | /* Loops of the partition. */ |
80ab0b19 | 598 | bitmap loops; |
a7a44c07 | 599 | /* True if the partition defines variable which is used outside of loop. */ |
826a536d | 600 | bool reduction_p; |
a7a44c07 BC |
601 | /* For builtin partition, true if it executes one iteration more than |
602 | number of loop (latch) iterations. */ | |
34e82342 | 603 | bool plus_one; |
30d55936 | 604 | enum partition_kind kind; |
f1eb4621 | 605 | enum partition_type type; |
d0582dc1 RG |
606 | /* data-references a kind != PKIND_NORMAL partition is about. */ |
607 | data_reference_p main_dr; | |
608 | data_reference_p secondary_dr; | |
a7a44c07 | 609 | /* Number of loop (latch) iterations. */ |
818625cf | 610 | tree niter; |
a7a44c07 BC |
611 | /* Data references in the partition. */ |
612 | bitmap datarefs; | |
526ceb68 | 613 | }; |
c61f8985 | 614 | |
c61f8985 RG |
615 | |
616 | /* Allocate and initialize a partition from BITMAP. */ | |
617 | ||
526ceb68 | 618 | static partition * |
a7a44c07 | 619 | partition_alloc (void) |
c61f8985 | 620 | { |
526ceb68 | 621 | partition *partition = XCNEW (struct partition); |
a7a44c07 BC |
622 | partition->stmts = BITMAP_ALLOC (NULL); |
623 | partition->loops = BITMAP_ALLOC (NULL); | |
826a536d | 624 | partition->reduction_p = false; |
30d55936 | 625 | partition->kind = PKIND_NORMAL; |
a7a44c07 | 626 | partition->datarefs = BITMAP_ALLOC (NULL); |
c61f8985 RG |
627 | return partition; |
628 | } | |
629 | ||
630 | /* Free PARTITION. */ | |
631 | ||
632 | static void | |
526ceb68 | 633 | partition_free (partition *partition) |
c61f8985 RG |
634 | { |
635 | BITMAP_FREE (partition->stmts); | |
80ab0b19 | 636 | BITMAP_FREE (partition->loops); |
a7a44c07 | 637 | BITMAP_FREE (partition->datarefs); |
c61f8985 RG |
638 | free (partition); |
639 | } | |
640 | ||
30d55936 RG |
641 | /* Returns true if the partition can be generated as a builtin. */ |
642 | ||
643 | static bool | |
526ceb68 | 644 | partition_builtin_p (partition *partition) |
30d55936 | 645 | { |
826a536d | 646 | return partition->kind != PKIND_NORMAL; |
30d55936 | 647 | } |
c61f8985 | 648 | |
826a536d | 649 | /* Returns true if the partition contains a reduction. */ |
7ad672e4 RG |
650 | |
651 | static bool | |
526ceb68 | 652 | partition_reduction_p (partition *partition) |
7ad672e4 | 653 | { |
826a536d | 654 | return partition->reduction_p; |
7ad672e4 RG |
655 | } |
656 | ||
821dbeef BC |
657 | /* Partitions are fused because of different reasons. */ |
658 | enum fuse_type | |
659 | { | |
660 | FUSE_NON_BUILTIN = 0, | |
661 | FUSE_REDUCTION = 1, | |
662 | FUSE_SHARE_REF = 2, | |
663 | FUSE_SAME_SCC = 3, | |
664 | FUSE_FINALIZE = 4 | |
665 | }; | |
666 | ||
667 | /* Description on different fusing reason. */ | |
668 | static const char *fuse_message[] = { | |
669 | "they are non-builtins", | |
670 | "they have reductions", | |
671 | "they have shared memory refs", | |
672 | "they are in the same dependence scc", | |
673 | "there is no point to distribute loop"}; | |
674 | ||
826a536d | 675 | static void |
f1eb4621 | 676 | update_type_for_merge (struct graph *, partition *, partition *); |
821dbeef | 677 | |
f1eb4621 BC |
678 | /* Merge PARTITION into the partition DEST. RDG is the reduced dependence |
679 | graph and we update type for result partition if it is non-NULL. */ | |
a7a44c07 | 680 | |
f1eb4621 BC |
681 | static void |
682 | partition_merge_into (struct graph *rdg, partition *dest, | |
683 | partition *partition, enum fuse_type ft) | |
684 | { | |
821dbeef BC |
685 | if (dump_file && (dump_flags & TDF_DETAILS)) |
686 | { | |
687 | fprintf (dump_file, "Fuse partitions because %s:\n", fuse_message[ft]); | |
688 | fprintf (dump_file, " Part 1: "); | |
689 | dump_bitmap (dump_file, dest->stmts); | |
690 | fprintf (dump_file, " Part 2: "); | |
691 | dump_bitmap (dump_file, partition->stmts); | |
692 | } | |
f1eb4621 BC |
693 | |
694 | dest->kind = PKIND_NORMAL; | |
695 | if (dest->type == PTYPE_PARALLEL) | |
696 | dest->type = partition->type; | |
697 | ||
698 | bitmap_ior_into (dest->stmts, partition->stmts); | |
699 | if (partition_reduction_p (partition)) | |
700 | dest->reduction_p = true; | |
701 | ||
702 | /* Further check if any data dependence prevents us from executing the | |
703 | new partition parallelly. */ | |
704 | if (dest->type == PTYPE_PARALLEL && rdg != NULL) | |
705 | update_type_for_merge (rdg, dest, partition); | |
706 | ||
707 | bitmap_ior_into (dest->datarefs, partition->datarefs); | |
826a536d RB |
708 | } |
709 | ||
710 | ||
c07a8cb3 RG |
711 | /* Returns true when DEF is an SSA_NAME defined in LOOP and used after |
712 | the LOOP. */ | |
713 | ||
714 | static bool | |
715 | ssa_name_has_uses_outside_loop_p (tree def, loop_p loop) | |
716 | { | |
717 | imm_use_iterator imm_iter; | |
718 | use_operand_p use_p; | |
719 | ||
720 | FOR_EACH_IMM_USE_FAST (use_p, imm_iter, def) | |
e665269a | 721 | { |
355fe088 | 722 | gimple *use_stmt = USE_STMT (use_p); |
e665269a RG |
723 | if (!is_gimple_debug (use_stmt) |
724 | && loop != loop_containing_stmt (use_stmt)) | |
725 | return true; | |
726 | } | |
c07a8cb3 RG |
727 | |
728 | return false; | |
729 | } | |
730 | ||
731 | /* Returns true when STMT defines a scalar variable used after the | |
88af7c1a | 732 | loop LOOP. */ |
c07a8cb3 RG |
733 | |
734 | static bool | |
355fe088 | 735 | stmt_has_scalar_dependences_outside_loop (loop_p loop, gimple *stmt) |
c07a8cb3 | 736 | { |
88af7c1a RG |
737 | def_operand_p def_p; |
738 | ssa_op_iter op_iter; | |
c07a8cb3 | 739 | |
9ca86fc3 RG |
740 | if (gimple_code (stmt) == GIMPLE_PHI) |
741 | return ssa_name_has_uses_outside_loop_p (gimple_phi_result (stmt), loop); | |
742 | ||
88af7c1a RG |
743 | FOR_EACH_SSA_DEF_OPERAND (def_p, stmt, op_iter, SSA_OP_DEF) |
744 | if (ssa_name_has_uses_outside_loop_p (DEF_FROM_PTR (def_p), loop)) | |
745 | return true; | |
c07a8cb3 | 746 | |
88af7c1a | 747 | return false; |
c07a8cb3 RG |
748 | } |
749 | ||
dea61d92 SP |
750 | /* Return a copy of LOOP placed before LOOP. */ |
751 | ||
752 | static struct loop * | |
753 | copy_loop_before (struct loop *loop) | |
754 | { | |
755 | struct loop *res; | |
756 | edge preheader = loop_preheader_edge (loop); | |
757 | ||
dea61d92 | 758 | initialize_original_copy_tables (); |
5ce9450f | 759 | res = slpeel_tree_duplicate_loop_to_edge_cfg (loop, NULL, preheader); |
30d55936 | 760 | gcc_assert (res != NULL); |
dea61d92 | 761 | free_original_copy_tables (); |
2cfc56b9 | 762 | delete_update_ssa (); |
dea61d92 SP |
763 | |
764 | return res; | |
765 | } | |
766 | ||
767 | /* Creates an empty basic block after LOOP. */ | |
768 | ||
769 | static void | |
770 | create_bb_after_loop (struct loop *loop) | |
771 | { | |
772 | edge exit = single_exit (loop); | |
773 | ||
774 | if (!exit) | |
775 | return; | |
776 | ||
777 | split_edge (exit); | |
778 | } | |
779 | ||
780 | /* Generate code for PARTITION from the code in LOOP. The loop is | |
781 | copied when COPY_P is true. All the statements not flagged in the | |
782 | PARTITION bitmap are removed from the loop or from its copy. The | |
783 | statements are indexed in sequence inside a basic block, and the | |
30d55936 | 784 | basic blocks of a loop are taken in dom order. */ |
dea61d92 | 785 | |
30d55936 | 786 | static void |
526ceb68 | 787 | generate_loops_for_partition (struct loop *loop, partition *partition, |
c61f8985 | 788 | bool copy_p) |
dea61d92 | 789 | { |
2fd5894f | 790 | unsigned i; |
dea61d92 SP |
791 | basic_block *bbs; |
792 | ||
793 | if (copy_p) | |
794 | { | |
a8745cc2 | 795 | int orig_loop_num = loop->orig_loop_num; |
dea61d92 | 796 | loop = copy_loop_before (loop); |
30d55936 | 797 | gcc_assert (loop != NULL); |
a8745cc2 | 798 | loop->orig_loop_num = orig_loop_num; |
dea61d92 SP |
799 | create_preheader (loop, CP_SIMPLE_PREHEADERS); |
800 | create_bb_after_loop (loop); | |
801 | } | |
a8745cc2 BC |
802 | else |
803 | { | |
804 | /* Origin number is set to the new versioned loop's num. */ | |
805 | gcc_assert (loop->orig_loop_num != loop->num); | |
806 | } | |
dea61d92 | 807 | |
2fd5894f | 808 | /* Remove stmts not in the PARTITION bitmap. */ |
dea61d92 SP |
809 | bbs = get_loop_body_in_dom_order (loop); |
810 | ||
b03c3082 | 811 | if (MAY_HAVE_DEBUG_STMTS) |
2fd5894f | 812 | for (i = 0; i < loop->num_nodes; i++) |
b03c3082 JJ |
813 | { |
814 | basic_block bb = bbs[i]; | |
815 | ||
538dd0b7 DM |
816 | for (gphi_iterator bsi = gsi_start_phis (bb); !gsi_end_p (bsi); |
817 | gsi_next (&bsi)) | |
2fd5894f | 818 | { |
538dd0b7 | 819 | gphi *phi = bsi.phi (); |
2fd5894f RB |
820 | if (!virtual_operand_p (gimple_phi_result (phi)) |
821 | && !bitmap_bit_p (partition->stmts, gimple_uid (phi))) | |
822 | reset_debug_uses (phi); | |
823 | } | |
b03c3082 | 824 | |
538dd0b7 | 825 | for (gimple_stmt_iterator bsi = gsi_start_bb (bb); !gsi_end_p (bsi); gsi_next (&bsi)) |
b03c3082 | 826 | { |
355fe088 | 827 | gimple *stmt = gsi_stmt (bsi); |
b03c3082 JJ |
828 | if (gimple_code (stmt) != GIMPLE_LABEL |
829 | && !is_gimple_debug (stmt) | |
2fd5894f | 830 | && !bitmap_bit_p (partition->stmts, gimple_uid (stmt))) |
b03c3082 JJ |
831 | reset_debug_uses (stmt); |
832 | } | |
833 | } | |
834 | ||
2fd5894f | 835 | for (i = 0; i < loop->num_nodes; i++) |
dea61d92 SP |
836 | { |
837 | basic_block bb = bbs[i]; | |
dea61d92 | 838 | |
538dd0b7 | 839 | for (gphi_iterator bsi = gsi_start_phis (bb); !gsi_end_p (bsi);) |
2fd5894f | 840 | { |
538dd0b7 | 841 | gphi *phi = bsi.phi (); |
2fd5894f RB |
842 | if (!virtual_operand_p (gimple_phi_result (phi)) |
843 | && !bitmap_bit_p (partition->stmts, gimple_uid (phi))) | |
2706a615 | 844 | remove_phi_node (&bsi, true); |
2fd5894f RB |
845 | else |
846 | gsi_next (&bsi); | |
847 | } | |
dea61d92 | 848 | |
538dd0b7 | 849 | for (gimple_stmt_iterator bsi = gsi_start_bb (bb); !gsi_end_p (bsi);) |
2706a615 | 850 | { |
355fe088 | 851 | gimple *stmt = gsi_stmt (bsi); |
b03c3082 JJ |
852 | if (gimple_code (stmt) != GIMPLE_LABEL |
853 | && !is_gimple_debug (stmt) | |
2fd5894f | 854 | && !bitmap_bit_p (partition->stmts, gimple_uid (stmt))) |
2706a615 | 855 | { |
36875e8f RB |
856 | /* Choose an arbitrary path through the empty CFG part |
857 | that this unnecessary control stmt controls. */ | |
538dd0b7 | 858 | if (gcond *cond_stmt = dyn_cast <gcond *> (stmt)) |
36875e8f | 859 | { |
538dd0b7 | 860 | gimple_cond_make_false (cond_stmt); |
36875e8f RB |
861 | update_stmt (stmt); |
862 | } | |
863 | else if (gimple_code (stmt) == GIMPLE_SWITCH) | |
864 | { | |
538dd0b7 | 865 | gswitch *switch_stmt = as_a <gswitch *> (stmt); |
36875e8f | 866 | gimple_switch_set_index |
538dd0b7 | 867 | (switch_stmt, CASE_LOW (gimple_switch_label (switch_stmt, 1))); |
36875e8f RB |
868 | update_stmt (stmt); |
869 | } | |
870 | else | |
871 | { | |
872 | unlink_stmt_vdef (stmt); | |
873 | gsi_remove (&bsi, true); | |
874 | release_defs (stmt); | |
875 | continue; | |
876 | } | |
2706a615 | 877 | } |
36875e8f | 878 | gsi_next (&bsi); |
2706a615 | 879 | } |
dea61d92 SP |
880 | } |
881 | ||
882 | free (bbs); | |
dea61d92 SP |
883 | } |
884 | ||
d0582dc1 | 885 | /* Build the size argument for a memory operation call. */ |
3661e899 | 886 | |
d0582dc1 | 887 | static tree |
d995e887 RB |
888 | build_size_arg_loc (location_t loc, data_reference_p dr, tree nb_iter, |
889 | bool plus_one) | |
3661e899 | 890 | { |
d995e887 RB |
891 | tree size = fold_convert_loc (loc, sizetype, nb_iter); |
892 | if (plus_one) | |
893 | size = size_binop (PLUS_EXPR, size, size_one_node); | |
894 | size = fold_build2_loc (loc, MULT_EXPR, sizetype, size, | |
d0582dc1 | 895 | TYPE_SIZE_UNIT (TREE_TYPE (DR_REF (dr)))); |
d995e887 RB |
896 | size = fold_convert_loc (loc, size_type_node, size); |
897 | return size; | |
d0582dc1 RG |
898 | } |
899 | ||
900 | /* Build an address argument for a memory operation call. */ | |
901 | ||
902 | static tree | |
903 | build_addr_arg_loc (location_t loc, data_reference_p dr, tree nb_bytes) | |
904 | { | |
905 | tree addr_base; | |
906 | ||
907 | addr_base = size_binop_loc (loc, PLUS_EXPR, DR_OFFSET (dr), DR_INIT (dr)); | |
908 | addr_base = fold_convert_loc (loc, sizetype, addr_base); | |
909 | ||
910 | /* Test for a negative stride, iterating over every element. */ | |
911 | if (tree_int_cst_sgn (DR_STEP (dr)) == -1) | |
912 | { | |
913 | addr_base = size_binop_loc (loc, MINUS_EXPR, addr_base, | |
914 | fold_convert_loc (loc, sizetype, nb_bytes)); | |
915 | addr_base = size_binop_loc (loc, PLUS_EXPR, addr_base, | |
916 | TYPE_SIZE_UNIT (TREE_TYPE (DR_REF (dr)))); | |
917 | } | |
918 | ||
919 | return fold_build_pointer_plus_loc (loc, DR_BASE_ADDRESS (dr), addr_base); | |
3661e899 TB |
920 | } |
921 | ||
401f3a81 JJ |
922 | /* If VAL memory representation contains the same value in all bytes, |
923 | return that value, otherwise return -1. | |
924 | E.g. for 0x24242424 return 0x24, for IEEE double | |
925 | 747708026454360457216.0 return 0x44, etc. */ | |
926 | ||
927 | static int | |
928 | const_with_all_bytes_same (tree val) | |
929 | { | |
930 | unsigned char buf[64]; | |
931 | int i, len; | |
932 | ||
933 | if (integer_zerop (val) | |
401f3a81 JJ |
934 | || (TREE_CODE (val) == CONSTRUCTOR |
935 | && !TREE_CLOBBER_P (val) | |
936 | && CONSTRUCTOR_NELTS (val) == 0)) | |
937 | return 0; | |
938 | ||
9e207d6f JJ |
939 | if (real_zerop (val)) |
940 | { | |
941 | /* Only return 0 for +0.0, not for -0.0, which doesn't have | |
942 | an all bytes same memory representation. Don't transform | |
943 | -0.0 stores into +0.0 even for !HONOR_SIGNED_ZEROS. */ | |
944 | switch (TREE_CODE (val)) | |
945 | { | |
946 | case REAL_CST: | |
947 | if (!real_isneg (TREE_REAL_CST_PTR (val))) | |
948 | return 0; | |
949 | break; | |
950 | case COMPLEX_CST: | |
951 | if (!const_with_all_bytes_same (TREE_REALPART (val)) | |
952 | && !const_with_all_bytes_same (TREE_IMAGPART (val))) | |
953 | return 0; | |
954 | break; | |
955 | case VECTOR_CST: | |
956 | unsigned int j; | |
957 | for (j = 0; j < VECTOR_CST_NELTS (val); ++j) | |
66088065 | 958 | if (const_with_all_bytes_same (VECTOR_CST_ELT (val, j))) |
9e207d6f JJ |
959 | break; |
960 | if (j == VECTOR_CST_NELTS (val)) | |
961 | return 0; | |
962 | break; | |
963 | default: | |
964 | break; | |
965 | } | |
966 | } | |
967 | ||
401f3a81 JJ |
968 | if (CHAR_BIT != 8 || BITS_PER_UNIT != 8) |
969 | return -1; | |
970 | ||
971 | len = native_encode_expr (val, buf, sizeof (buf)); | |
972 | if (len == 0) | |
973 | return -1; | |
974 | for (i = 1; i < len; i++) | |
975 | if (buf[i] != buf[0]) | |
976 | return -1; | |
977 | return buf[0]; | |
978 | } | |
979 | ||
30d55936 | 980 | /* Generate a call to memset for PARTITION in LOOP. */ |
dea61d92 | 981 | |
cfee318d | 982 | static void |
526ceb68 | 983 | generate_memset_builtin (struct loop *loop, partition *partition) |
dea61d92 | 984 | { |
30d55936 | 985 | gimple_stmt_iterator gsi; |
355fe088 | 986 | gimple *stmt, *fn_call; |
818625cf | 987 | tree mem, fn, nb_bytes; |
30d55936 | 988 | location_t loc; |
b6dd5261 | 989 | tree val; |
30d55936 | 990 | |
d0582dc1 | 991 | stmt = DR_STMT (partition->main_dr); |
30d55936 | 992 | loc = gimple_location (stmt); |
30d55936 RG |
993 | |
994 | /* The new statements will be placed before LOOP. */ | |
995 | gsi = gsi_last_bb (loop_preheader_edge (loop)->src); | |
dea61d92 | 996 | |
d995e887 RB |
997 | nb_bytes = build_size_arg_loc (loc, partition->main_dr, partition->niter, |
998 | partition->plus_one); | |
d0582dc1 RG |
999 | nb_bytes = force_gimple_operand_gsi (&gsi, nb_bytes, true, NULL_TREE, |
1000 | false, GSI_CONTINUE_LINKING); | |
1001 | mem = build_addr_arg_loc (loc, partition->main_dr, nb_bytes); | |
1002 | mem = force_gimple_operand_gsi (&gsi, mem, true, NULL_TREE, | |
1003 | false, GSI_CONTINUE_LINKING); | |
dea61d92 | 1004 | |
b6dd5261 RG |
1005 | /* This exactly matches the pattern recognition in classify_partition. */ |
1006 | val = gimple_assign_rhs1 (stmt); | |
401f3a81 JJ |
1007 | /* Handle constants like 0x15151515 and similarly |
1008 | floating point constants etc. where all bytes are the same. */ | |
1009 | int bytev = const_with_all_bytes_same (val); | |
1010 | if (bytev != -1) | |
1011 | val = build_int_cst (integer_type_node, bytev); | |
1012 | else if (TREE_CODE (val) == INTEGER_CST) | |
1013 | val = fold_convert (integer_type_node, val); | |
1014 | else if (!useless_type_conversion_p (integer_type_node, TREE_TYPE (val))) | |
b6dd5261 | 1015 | { |
b731b390 | 1016 | tree tem = make_ssa_name (integer_type_node); |
355fe088 | 1017 | gimple *cstmt = gimple_build_assign (tem, NOP_EXPR, val); |
401f3a81 JJ |
1018 | gsi_insert_after (&gsi, cstmt, GSI_CONTINUE_LINKING); |
1019 | val = tem; | |
b6dd5261 RG |
1020 | } |
1021 | ||
e79983f4 | 1022 | fn = build_fold_addr_expr (builtin_decl_implicit (BUILT_IN_MEMSET)); |
b6dd5261 | 1023 | fn_call = gimple_build_call (fn, 3, mem, val, nb_bytes); |
d0582dc1 | 1024 | gsi_insert_after (&gsi, fn_call, GSI_CONTINUE_LINKING); |
dea61d92 SP |
1025 | |
1026 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
b6dd5261 RG |
1027 | { |
1028 | fprintf (dump_file, "generated memset"); | |
401f3a81 | 1029 | if (bytev == 0) |
b6dd5261 | 1030 | fprintf (dump_file, " zero\n"); |
b6dd5261 RG |
1031 | else |
1032 | fprintf (dump_file, "\n"); | |
1033 | } | |
dea61d92 SP |
1034 | } |
1035 | ||
d0582dc1 RG |
1036 | /* Generate a call to memcpy for PARTITION in LOOP. */ |
1037 | ||
1038 | static void | |
526ceb68 | 1039 | generate_memcpy_builtin (struct loop *loop, partition *partition) |
d0582dc1 RG |
1040 | { |
1041 | gimple_stmt_iterator gsi; | |
355fe088 | 1042 | gimple *stmt, *fn_call; |
818625cf | 1043 | tree dest, src, fn, nb_bytes; |
d0582dc1 RG |
1044 | location_t loc; |
1045 | enum built_in_function kind; | |
1046 | ||
1047 | stmt = DR_STMT (partition->main_dr); | |
1048 | loc = gimple_location (stmt); | |
d0582dc1 RG |
1049 | |
1050 | /* The new statements will be placed before LOOP. */ | |
1051 | gsi = gsi_last_bb (loop_preheader_edge (loop)->src); | |
1052 | ||
d995e887 RB |
1053 | nb_bytes = build_size_arg_loc (loc, partition->main_dr, partition->niter, |
1054 | partition->plus_one); | |
d0582dc1 RG |
1055 | nb_bytes = force_gimple_operand_gsi (&gsi, nb_bytes, true, NULL_TREE, |
1056 | false, GSI_CONTINUE_LINKING); | |
1057 | dest = build_addr_arg_loc (loc, partition->main_dr, nb_bytes); | |
1058 | src = build_addr_arg_loc (loc, partition->secondary_dr, nb_bytes); | |
510d73a0 RB |
1059 | if (partition->kind == PKIND_MEMCPY |
1060 | || ! ptr_derefs_may_alias_p (dest, src)) | |
d0582dc1 | 1061 | kind = BUILT_IN_MEMCPY; |
510d73a0 RB |
1062 | else |
1063 | kind = BUILT_IN_MEMMOVE; | |
d0582dc1 RG |
1064 | |
1065 | dest = force_gimple_operand_gsi (&gsi, dest, true, NULL_TREE, | |
1066 | false, GSI_CONTINUE_LINKING); | |
1067 | src = force_gimple_operand_gsi (&gsi, src, true, NULL_TREE, | |
1068 | false, GSI_CONTINUE_LINKING); | |
1069 | fn = build_fold_addr_expr (builtin_decl_implicit (kind)); | |
1070 | fn_call = gimple_build_call (fn, 3, dest, src, nb_bytes); | |
1071 | gsi_insert_after (&gsi, fn_call, GSI_CONTINUE_LINKING); | |
1072 | ||
1073 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
1074 | { | |
1075 | if (kind == BUILT_IN_MEMCPY) | |
1076 | fprintf (dump_file, "generated memcpy\n"); | |
1077 | else | |
1078 | fprintf (dump_file, "generated memmove\n"); | |
1079 | } | |
1080 | } | |
1081 | ||
30d55936 | 1082 | /* Remove and destroy the loop LOOP. */ |
dea61d92 | 1083 | |
30d55936 RG |
1084 | static void |
1085 | destroy_loop (struct loop *loop) | |
dea61d92 | 1086 | { |
30d55936 RG |
1087 | unsigned nbbs = loop->num_nodes; |
1088 | edge exit = single_exit (loop); | |
1089 | basic_block src = loop_preheader_edge (loop)->src, dest = exit->dest; | |
dea61d92 | 1090 | basic_block *bbs; |
30d55936 | 1091 | unsigned i; |
dea61d92 SP |
1092 | |
1093 | bbs = get_loop_body_in_dom_order (loop); | |
1094 | ||
30d55936 RG |
1095 | redirect_edge_pred (exit, src); |
1096 | exit->flags &= ~(EDGE_TRUE_VALUE|EDGE_FALSE_VALUE); | |
1097 | exit->flags |= EDGE_FALLTHRU; | |
1098 | cancel_loop_tree (loop); | |
1099 | rescan_loop_exit (exit, false, true); | |
dea61d92 | 1100 | |
b9aba0a0 RB |
1101 | i = nbbs; |
1102 | do | |
c014f6f5 RG |
1103 | { |
1104 | /* We have made sure to not leave any dangling uses of SSA | |
1105 | names defined in the loop. With the exception of virtuals. | |
1106 | Make sure we replace all uses of virtual defs that will remain | |
1107 | outside of the loop with the bare symbol as delete_basic_block | |
1108 | will release them. */ | |
b9aba0a0 | 1109 | --i; |
538dd0b7 DM |
1110 | for (gphi_iterator gsi = gsi_start_phis (bbs[i]); !gsi_end_p (gsi); |
1111 | gsi_next (&gsi)) | |
c014f6f5 | 1112 | { |
538dd0b7 | 1113 | gphi *phi = gsi.phi (); |
ea057359 | 1114 | if (virtual_operand_p (gimple_phi_result (phi))) |
c014f6f5 RG |
1115 | mark_virtual_phi_result_for_renaming (phi); |
1116 | } | |
538dd0b7 DM |
1117 | for (gimple_stmt_iterator gsi = gsi_start_bb (bbs[i]); !gsi_end_p (gsi); |
1118 | gsi_next (&gsi)) | |
c014f6f5 | 1119 | { |
355fe088 | 1120 | gimple *stmt = gsi_stmt (gsi); |
c014f6f5 RG |
1121 | tree vdef = gimple_vdef (stmt); |
1122 | if (vdef && TREE_CODE (vdef) == SSA_NAME) | |
1123 | mark_virtual_operand_for_renaming (vdef); | |
1124 | } | |
1125 | delete_basic_block (bbs[i]); | |
1126 | } | |
b9aba0a0 RB |
1127 | while (i != 0); |
1128 | ||
dea61d92 | 1129 | free (bbs); |
30d55936 RG |
1130 | |
1131 | set_immediate_dominator (CDI_DOMINATORS, dest, | |
1132 | recompute_dominator (CDI_DOMINATORS, dest)); | |
dea61d92 SP |
1133 | } |
1134 | ||
b71b7a8e | 1135 | /* Generates code for PARTITION. Return whether LOOP needs to be destroyed. */ |
dea61d92 | 1136 | |
b71b7a8e | 1137 | static bool |
d0582dc1 | 1138 | generate_code_for_partition (struct loop *loop, |
526ceb68 | 1139 | partition *partition, bool copy_p) |
dea61d92 | 1140 | { |
30d55936 RG |
1141 | switch (partition->kind) |
1142 | { | |
826a536d RB |
1143 | case PKIND_NORMAL: |
1144 | /* Reductions all have to be in the last partition. */ | |
1145 | gcc_assert (!partition_reduction_p (partition) | |
1146 | || !copy_p); | |
1147 | generate_loops_for_partition (loop, partition, copy_p); | |
b71b7a8e | 1148 | return false; |
826a536d | 1149 | |
30d55936 | 1150 | case PKIND_MEMSET: |
d0582dc1 | 1151 | generate_memset_builtin (loop, partition); |
d0582dc1 RG |
1152 | break; |
1153 | ||
1154 | case PKIND_MEMCPY: | |
510d73a0 | 1155 | case PKIND_MEMMOVE: |
d0582dc1 | 1156 | generate_memcpy_builtin (loop, partition); |
30d55936 RG |
1157 | break; |
1158 | ||
1159 | default: | |
1160 | gcc_unreachable (); | |
1161 | } | |
dea61d92 | 1162 | |
826a536d RB |
1163 | /* Common tail for partitions we turn into a call. If this was the last |
1164 | partition for which we generate code, we have to destroy the loop. */ | |
1165 | if (!copy_p) | |
b71b7a8e RB |
1166 | return true; |
1167 | return false; | |
dea61d92 SP |
1168 | } |
1169 | ||
17c5cbdf BC |
1170 | /* Return data dependence relation for data references A and B. The two |
1171 | data references must be in lexicographic order wrto reduced dependence | |
1172 | graph RDG. We firstly try to find ddr from global ddr hash table. If | |
1173 | it doesn't exist, compute the ddr and cache it. */ | |
1174 | ||
1175 | static data_dependence_relation * | |
1176 | get_data_dependence (struct graph *rdg, data_reference_p a, data_reference_p b) | |
1177 | { | |
1178 | struct data_dependence_relation ent, **slot; | |
1179 | struct data_dependence_relation *ddr; | |
1180 | ||
1181 | gcc_assert (DR_IS_WRITE (a) || DR_IS_WRITE (b)); | |
1182 | gcc_assert (rdg_vertex_for_stmt (rdg, DR_STMT (a)) | |
1183 | <= rdg_vertex_for_stmt (rdg, DR_STMT (b))); | |
1184 | ent.a = a; | |
1185 | ent.b = b; | |
1186 | slot = ddrs_table.find_slot (&ent, INSERT); | |
1187 | if (*slot == NULL) | |
1188 | { | |
1189 | ddr = initialize_data_dependence_relation (a, b, loop_nest); | |
1190 | compute_affine_dependence (ddr, loop_nest[0]); | |
1191 | *slot = ddr; | |
1192 | } | |
1193 | ||
1194 | return *slot; | |
1195 | } | |
dea61d92 | 1196 | |
f1eb4621 BC |
1197 | /* In reduced dependence graph RDG for loop distribution, return true if |
1198 | dependence between references DR1 and DR2 leads to a dependence cycle | |
1199 | and such dependence cycle can't be resolved by runtime alias check. */ | |
1200 | ||
1201 | static bool | |
1202 | data_dep_in_cycle_p (struct graph *rdg, | |
1203 | data_reference_p dr1, data_reference_p dr2) | |
1204 | { | |
1205 | struct data_dependence_relation *ddr; | |
1206 | ||
1207 | /* Re-shuffle data-refs to be in topological order. */ | |
1208 | if (rdg_vertex_for_stmt (rdg, DR_STMT (dr1)) | |
1209 | > rdg_vertex_for_stmt (rdg, DR_STMT (dr2))) | |
1210 | std::swap (dr1, dr2); | |
1211 | ||
1212 | ddr = get_data_dependence (rdg, dr1, dr2); | |
1213 | ||
1214 | /* In case of no data dependence. */ | |
1215 | if (DDR_ARE_DEPENDENT (ddr) == chrec_known) | |
1216 | return false; | |
1217 | /* For unknown data dependence or known data dependence which can't be | |
1218 | expressed in classic distance vector, we check if it can be resolved | |
1219 | by runtime alias check. If yes, we still consider data dependence | |
1220 | as won't introduce data dependence cycle. */ | |
1221 | else if (DDR_ARE_DEPENDENT (ddr) == chrec_dont_know | |
1222 | || DDR_NUM_DIST_VECTS (ddr) == 0) | |
1223 | return !runtime_alias_check_p (ddr, NULL, true); | |
1224 | else if (DDR_NUM_DIST_VECTS (ddr) > 1) | |
1225 | return true; | |
1226 | else if (DDR_REVERSED_P (ddr) | |
1227 | || lambda_vector_zerop (DDR_DIST_VECT (ddr, 0), 1)) | |
1228 | return false; | |
1229 | ||
1230 | return true; | |
1231 | } | |
1232 | ||
1233 | /* Given reduced dependence graph RDG, PARTITION1 and PARTITION2, update | |
1234 | PARTITION1's type after merging PARTITION2 into PARTITION1. */ | |
1235 | ||
1236 | static void | |
1237 | update_type_for_merge (struct graph *rdg, | |
1238 | partition *partition1, partition *partition2) | |
1239 | { | |
1240 | unsigned i, j; | |
1241 | bitmap_iterator bi, bj; | |
1242 | data_reference_p dr1, dr2; | |
1243 | ||
1244 | EXECUTE_IF_SET_IN_BITMAP (partition1->datarefs, 0, i, bi) | |
1245 | { | |
1246 | unsigned start = (partition1 == partition2) ? i + 1 : 0; | |
1247 | ||
1248 | dr1 = datarefs_vec[i]; | |
1249 | EXECUTE_IF_SET_IN_BITMAP (partition2->datarefs, start, j, bj) | |
1250 | { | |
1251 | dr2 = datarefs_vec[j]; | |
1252 | if (DR_IS_READ (dr1) && DR_IS_READ (dr2)) | |
1253 | continue; | |
1254 | ||
1255 | /* Partition can only be executed sequentially if there is any | |
1256 | data dependence cycle. */ | |
1257 | if (data_dep_in_cycle_p (rdg, dr1, dr2)) | |
1258 | { | |
1259 | partition1->type = PTYPE_SEQUENTIAL; | |
1260 | return; | |
1261 | } | |
1262 | } | |
1263 | } | |
1264 | } | |
1265 | ||
24f161fd RB |
1266 | /* Returns a partition with all the statements needed for computing |
1267 | the vertex V of the RDG, also including the loop exit conditions. */ | |
dea61d92 | 1268 | |
526ceb68 | 1269 | static partition * |
24f161fd | 1270 | build_rdg_partition_for_vertex (struct graph *rdg, int v) |
dea61d92 | 1271 | { |
a7a44c07 | 1272 | partition *partition = partition_alloc (); |
00f96dc9 | 1273 | auto_vec<int, 3> nodes; |
a7a44c07 | 1274 | unsigned i, j; |
dea61d92 | 1275 | int x; |
a7a44c07 | 1276 | data_reference_p dr; |
dea61d92 | 1277 | |
174ec470 | 1278 | graphds_dfs (rdg, &v, 1, &nodes, false, NULL); |
dea61d92 | 1279 | |
9771b263 | 1280 | FOR_EACH_VEC_ELT (nodes, i, x) |
24f161fd RB |
1281 | { |
1282 | bitmap_set_bit (partition->stmts, x); | |
1283 | bitmap_set_bit (partition->loops, | |
1284 | loop_containing_stmt (RDG_STMT (rdg, x))->num); | |
a7a44c07 BC |
1285 | |
1286 | for (j = 0; RDG_DATAREFS (rdg, x).iterate (j, &dr); ++j) | |
1287 | { | |
1288 | unsigned idx = (unsigned) DR_INDEX (dr); | |
1289 | gcc_assert (idx < datarefs_vec.length ()); | |
1290 | ||
f1eb4621 BC |
1291 | /* Partition can only be executed sequentially if there is any |
1292 | unknown data reference. */ | |
1293 | if (!DR_BASE_ADDRESS (dr) || !DR_OFFSET (dr) | |
1294 | || !DR_INIT (dr) || !DR_STEP (dr)) | |
1295 | partition->type = PTYPE_SEQUENTIAL; | |
1296 | ||
a7a44c07 BC |
1297 | bitmap_set_bit (partition->datarefs, idx); |
1298 | } | |
24f161fd | 1299 | } |
dea61d92 | 1300 | |
f1eb4621 BC |
1301 | if (partition->type == PTYPE_SEQUENTIAL) |
1302 | return partition; | |
1303 | ||
1304 | /* Further check if any data dependence prevents us from executing the | |
1305 | partition parallelly. */ | |
1306 | update_type_for_merge (rdg, partition, partition); | |
1307 | ||
dea61d92 SP |
1308 | return partition; |
1309 | } | |
1310 | ||
30d55936 | 1311 | /* Classifies the builtin kind we can generate for PARTITION of RDG and LOOP. |
4a52eb19 BC |
1312 | For the moment we detect memset, memcpy and memmove patterns. Bitmap |
1313 | STMT_IN_ALL_PARTITIONS contains statements belonging to all partitions. */ | |
cfee318d | 1314 | |
30d55936 | 1315 | static void |
4a52eb19 BC |
1316 | classify_partition (loop_p loop, struct graph *rdg, partition *partition, |
1317 | bitmap stmt_in_all_partitions) | |
cfee318d | 1318 | { |
cfee318d | 1319 | bitmap_iterator bi; |
30d55936 RG |
1320 | unsigned i; |
1321 | tree nb_iter; | |
d0582dc1 | 1322 | data_reference_p single_load, single_store; |
4a52eb19 | 1323 | bool volatiles_p = false, plus_one = false, has_reduction = false; |
30d55936 RG |
1324 | |
1325 | partition->kind = PKIND_NORMAL; | |
d0582dc1 RG |
1326 | partition->main_dr = NULL; |
1327 | partition->secondary_dr = NULL; | |
818625cf | 1328 | partition->niter = NULL_TREE; |
d995e887 | 1329 | partition->plus_one = false; |
30d55936 | 1330 | |
c61f8985 | 1331 | EXECUTE_IF_SET_IN_BITMAP (partition->stmts, 0, i, bi) |
30d55936 | 1332 | { |
355fe088 | 1333 | gimple *stmt = RDG_STMT (rdg, i); |
30d55936 RG |
1334 | |
1335 | if (gimple_has_volatile_ops (stmt)) | |
b9fc0497 | 1336 | volatiles_p = true; |
cfee318d | 1337 | |
4a52eb19 BC |
1338 | /* If the stmt is not included by all partitions and there is uses |
1339 | outside of the loop, then mark the partition as reduction. */ | |
9ca86fc3 | 1340 | if (stmt_has_scalar_dependences_outside_loop (loop, stmt)) |
30d55936 | 1341 | { |
4a52eb19 BC |
1342 | /* Due to limitation in the transform phase we have to fuse all |
1343 | reduction partitions. As a result, this could cancel valid | |
1344 | loop distribution especially for loop that induction variable | |
1345 | is used outside of loop. To workaround this issue, we skip | |
1346 | marking partition as reudction if the reduction stmt belongs | |
1347 | to all partitions. In such case, reduction will be computed | |
1348 | correctly no matter how partitions are fused/distributed. */ | |
1349 | if (!bitmap_bit_p (stmt_in_all_partitions, i)) | |
1350 | { | |
1351 | partition->reduction_p = true; | |
1352 | return; | |
1353 | } | |
1354 | has_reduction = true; | |
30d55936 RG |
1355 | } |
1356 | } | |
1357 | ||
b9fc0497 RB |
1358 | /* Perform general partition disqualification for builtins. */ |
1359 | if (volatiles_p | |
4a52eb19 BC |
1360 | /* Simple workaround to prevent classifying the partition as builtin |
1361 | if it contains any use outside of loop. */ | |
1362 | || has_reduction | |
b9fc0497 RB |
1363 | || !flag_tree_loop_distribute_patterns) |
1364 | return; | |
1365 | ||
d0582dc1 RG |
1366 | /* Detect memset and memcpy. */ |
1367 | single_load = NULL; | |
1368 | single_store = NULL; | |
30d55936 RG |
1369 | EXECUTE_IF_SET_IN_BITMAP (partition->stmts, 0, i, bi) |
1370 | { | |
355fe088 | 1371 | gimple *stmt = RDG_STMT (rdg, i); |
d0582dc1 RG |
1372 | data_reference_p dr; |
1373 | unsigned j; | |
30d55936 RG |
1374 | |
1375 | if (gimple_code (stmt) == GIMPLE_PHI) | |
1376 | continue; | |
1377 | ||
1378 | /* Any scalar stmts are ok. */ | |
1379 | if (!gimple_vuse (stmt)) | |
1380 | continue; | |
1381 | ||
d0582dc1 RG |
1382 | /* Otherwise just regular loads/stores. */ |
1383 | if (!gimple_assign_single_p (stmt)) | |
1384 | return; | |
1385 | ||
1386 | /* But exactly one store and/or load. */ | |
9771b263 | 1387 | for (j = 0; RDG_DATAREFS (rdg, i).iterate (j, &dr); ++j) |
30d55936 | 1388 | { |
d002d099 JJ |
1389 | tree type = TREE_TYPE (DR_REF (dr)); |
1390 | ||
1391 | /* The memset, memcpy and memmove library calls are only | |
1392 | able to deal with generic address space. */ | |
1393 | if (!ADDR_SPACE_GENERIC_P (TYPE_ADDR_SPACE (type))) | |
1394 | return; | |
1395 | ||
d0582dc1 RG |
1396 | if (DR_IS_READ (dr)) |
1397 | { | |
1398 | if (single_load != NULL) | |
1399 | return; | |
1400 | single_load = dr; | |
1401 | } | |
1402 | else | |
1403 | { | |
1404 | if (single_store != NULL) | |
1405 | return; | |
1406 | single_store = dr; | |
1407 | } | |
30d55936 | 1408 | } |
30d55936 RG |
1409 | } |
1410 | ||
818625cf RB |
1411 | if (!single_store) |
1412 | return; | |
1413 | ||
d995e887 | 1414 | nb_iter = number_of_latch_executions (loop); |
6ff37519 | 1415 | gcc_assert (nb_iter && nb_iter != chrec_dont_know); |
d995e887 RB |
1416 | if (dominated_by_p (CDI_DOMINATORS, single_exit (loop)->src, |
1417 | gimple_bb (DR_STMT (single_store)))) | |
1418 | plus_one = true; | |
818625cf | 1419 | |
d0582dc1 RG |
1420 | if (single_store && !single_load) |
1421 | { | |
355fe088 | 1422 | gimple *stmt = DR_STMT (single_store); |
d0582dc1 | 1423 | tree rhs = gimple_assign_rhs1 (stmt); |
401f3a81 JJ |
1424 | if (const_with_all_bytes_same (rhs) == -1 |
1425 | && (!INTEGRAL_TYPE_P (TREE_TYPE (rhs)) | |
1426 | || (TYPE_MODE (TREE_TYPE (rhs)) | |
1427 | != TYPE_MODE (unsigned_char_type_node)))) | |
d0582dc1 RG |
1428 | return; |
1429 | if (TREE_CODE (rhs) == SSA_NAME | |
1430 | && !SSA_NAME_IS_DEFAULT_DEF (rhs) | |
1431 | && flow_bb_inside_loop_p (loop, gimple_bb (SSA_NAME_DEF_STMT (rhs)))) | |
1432 | return; | |
ca406576 RB |
1433 | if (!adjacent_dr_p (single_store) |
1434 | || !dominated_by_p (CDI_DOMINATORS, | |
1435 | loop->latch, gimple_bb (stmt))) | |
d0582dc1 RG |
1436 | return; |
1437 | partition->kind = PKIND_MEMSET; | |
1438 | partition->main_dr = single_store; | |
818625cf | 1439 | partition->niter = nb_iter; |
d995e887 | 1440 | partition->plus_one = plus_one; |
d0582dc1 RG |
1441 | } |
1442 | else if (single_store && single_load) | |
1443 | { | |
355fe088 TS |
1444 | gimple *store = DR_STMT (single_store); |
1445 | gimple *load = DR_STMT (single_load); | |
d0582dc1 RG |
1446 | /* Direct aggregate copy or via an SSA name temporary. */ |
1447 | if (load != store | |
1448 | && gimple_assign_lhs (load) != gimple_assign_rhs1 (store)) | |
1449 | return; | |
1450 | if (!adjacent_dr_p (single_store) | |
1451 | || !adjacent_dr_p (single_load) | |
1452 | || !operand_equal_p (DR_STEP (single_store), | |
ca406576 RB |
1453 | DR_STEP (single_load), 0) |
1454 | || !dominated_by_p (CDI_DOMINATORS, | |
1455 | loop->latch, gimple_bb (store))) | |
d0582dc1 | 1456 | return; |
f20132e7 RG |
1457 | /* Now check that if there is a dependence this dependence is |
1458 | of a suitable form for memmove. */ | |
17c5cbdf | 1459 | ddr_p ddr = get_data_dependence (rdg, single_load, single_store); |
f20132e7 | 1460 | if (DDR_ARE_DEPENDENT (ddr) == chrec_dont_know) |
17c5cbdf BC |
1461 | return; |
1462 | ||
f20132e7 RG |
1463 | if (DDR_ARE_DEPENDENT (ddr) != chrec_known) |
1464 | { | |
1465 | if (DDR_NUM_DIST_VECTS (ddr) == 0) | |
17c5cbdf BC |
1466 | return; |
1467 | ||
f20132e7 | 1468 | lambda_vector dist_v; |
9771b263 | 1469 | FOR_EACH_VEC_ELT (DDR_DIST_VECTS (ddr), i, dist_v) |
f20132e7 RG |
1470 | { |
1471 | int dist = dist_v[index_in_loop_nest (loop->num, | |
1472 | DDR_LOOP_NEST (ddr))]; | |
1473 | if (dist > 0 && !DDR_REVERSED_P (ddr)) | |
17c5cbdf | 1474 | return; |
f20132e7 | 1475 | } |
510d73a0 | 1476 | partition->kind = PKIND_MEMMOVE; |
f20132e7 | 1477 | } |
510d73a0 RB |
1478 | else |
1479 | partition->kind = PKIND_MEMCPY; | |
d0582dc1 RG |
1480 | partition->main_dr = single_store; |
1481 | partition->secondary_dr = single_load; | |
818625cf | 1482 | partition->niter = nb_iter; |
d995e887 | 1483 | partition->plus_one = plus_one; |
d0582dc1 | 1484 | } |
cfee318d SP |
1485 | } |
1486 | ||
95f7d11b BC |
1487 | /* Returns true when PARTITION1 and PARTITION2 access the same memory |
1488 | object in RDG. */ | |
cfee318d SP |
1489 | |
1490 | static bool | |
95f7d11b BC |
1491 | share_memory_accesses (struct graph *rdg, |
1492 | partition *partition1, partition *partition2) | |
cfee318d | 1493 | { |
95f7d11b | 1494 | unsigned i, j; |
cfee318d | 1495 | bitmap_iterator bi, bj; |
95f7d11b | 1496 | data_reference_p dr1, dr2; |
1fa0c180 RG |
1497 | |
1498 | /* First check whether in the intersection of the two partitions are | |
1499 | any loads or stores. Common loads are the situation that happens | |
1500 | most often. */ | |
1501 | EXECUTE_IF_AND_IN_BITMAP (partition1->stmts, partition2->stmts, 0, i, bi) | |
1502 | if (RDG_MEM_WRITE_STMT (rdg, i) | |
1503 | || RDG_MEM_READS_STMT (rdg, i)) | |
1504 | return true; | |
cfee318d | 1505 | |
95f7d11b BC |
1506 | /* Then check whether the two partitions access the same memory object. */ |
1507 | EXECUTE_IF_SET_IN_BITMAP (partition1->datarefs, 0, i, bi) | |
1508 | { | |
1509 | dr1 = datarefs_vec[i]; | |
1510 | ||
1511 | if (!DR_BASE_ADDRESS (dr1) | |
1512 | || !DR_OFFSET (dr1) || !DR_INIT (dr1) || !DR_STEP (dr1)) | |
1513 | continue; | |
1514 | ||
1515 | EXECUTE_IF_SET_IN_BITMAP (partition2->datarefs, 0, j, bj) | |
1516 | { | |
1517 | dr2 = datarefs_vec[j]; | |
1518 | ||
1519 | if (!DR_BASE_ADDRESS (dr2) | |
1520 | || !DR_OFFSET (dr2) || !DR_INIT (dr2) || !DR_STEP (dr2)) | |
1521 | continue; | |
1522 | ||
1523 | if (operand_equal_p (DR_BASE_ADDRESS (dr1), DR_BASE_ADDRESS (dr2), 0) | |
1524 | && operand_equal_p (DR_OFFSET (dr1), DR_OFFSET (dr2), 0) | |
1525 | && operand_equal_p (DR_INIT (dr1), DR_INIT (dr2), 0) | |
1526 | && operand_equal_p (DR_STEP (dr1), DR_STEP (dr2), 0)) | |
1527 | return true; | |
1528 | } | |
1529 | } | |
cfee318d SP |
1530 | |
1531 | return false; | |
1532 | } | |
1533 | ||
4a52eb19 BC |
1534 | /* For each seed statement in STARTING_STMTS, this function builds |
1535 | partition for it by adding depended statements according to RDG. | |
1536 | All partitions are recorded in PARTITIONS. */ | |
dea61d92 SP |
1537 | |
1538 | static void | |
83a95546 | 1539 | rdg_build_partitions (struct graph *rdg, |
355fe088 | 1540 | vec<gimple *> starting_stmts, |
526ceb68 | 1541 | vec<partition *> *partitions) |
dea61d92 | 1542 | { |
0e3de1d4 | 1543 | auto_bitmap processed; |
2fd5894f | 1544 | int i; |
355fe088 | 1545 | gimple *stmt; |
dea61d92 | 1546 | |
2fd5894f | 1547 | FOR_EACH_VEC_ELT (starting_stmts, i, stmt) |
dea61d92 | 1548 | { |
2fd5894f RB |
1549 | int v = rdg_vertex_for_stmt (rdg, stmt); |
1550 | ||
1551 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
1552 | fprintf (dump_file, | |
1553 | "ldist asked to generate code for vertex %d\n", v); | |
b8698a0f | 1554 | |
24f161fd RB |
1555 | /* If the vertex is already contained in another partition so |
1556 | is the partition rooted at it. */ | |
dea61d92 SP |
1557 | if (bitmap_bit_p (processed, v)) |
1558 | continue; | |
b8698a0f | 1559 | |
526ceb68 | 1560 | partition *partition = build_rdg_partition_for_vertex (rdg, v); |
24f161fd | 1561 | bitmap_ior_into (processed, partition->stmts); |
dea61d92 | 1562 | |
826a536d | 1563 | if (dump_file && (dump_flags & TDF_DETAILS)) |
dea61d92 | 1564 | { |
f1eb4621 BC |
1565 | fprintf (dump_file, "ldist creates useful %s partition:\n", |
1566 | partition->type == PTYPE_PARALLEL ? "parallel" : "sequent"); | |
1567 | bitmap_print (dump_file, partition->stmts, " ", "\n"); | |
dea61d92 | 1568 | } |
826a536d RB |
1569 | |
1570 | partitions->safe_push (partition); | |
dea61d92 SP |
1571 | } |
1572 | ||
83a95546 RB |
1573 | /* All vertices should have been assigned to at least one partition now, |
1574 | other than vertices belonging to dead code. */ | |
dea61d92 SP |
1575 | } |
1576 | ||
1577 | /* Dump to FILE the PARTITIONS. */ | |
1578 | ||
1579 | static void | |
526ceb68 | 1580 | dump_rdg_partitions (FILE *file, vec<partition *> partitions) |
dea61d92 SP |
1581 | { |
1582 | int i; | |
526ceb68 | 1583 | partition *partition; |
dea61d92 | 1584 | |
9771b263 | 1585 | FOR_EACH_VEC_ELT (partitions, i, partition) |
c61f8985 | 1586 | debug_bitmap_file (file, partition->stmts); |
dea61d92 SP |
1587 | } |
1588 | ||
1589 | /* Debug PARTITIONS. */ | |
526ceb68 | 1590 | extern void debug_rdg_partitions (vec<partition *> ); |
dea61d92 | 1591 | |
24e47c76 | 1592 | DEBUG_FUNCTION void |
526ceb68 | 1593 | debug_rdg_partitions (vec<partition *> partitions) |
dea61d92 SP |
1594 | { |
1595 | dump_rdg_partitions (stderr, partitions); | |
1596 | } | |
1597 | ||
2b8aee8e SP |
1598 | /* Returns the number of read and write operations in the RDG. */ |
1599 | ||
1600 | static int | |
1601 | number_of_rw_in_rdg (struct graph *rdg) | |
1602 | { | |
1603 | int i, res = 0; | |
1604 | ||
1605 | for (i = 0; i < rdg->n_vertices; i++) | |
1606 | { | |
1607 | if (RDG_MEM_WRITE_STMT (rdg, i)) | |
1608 | ++res; | |
1609 | ||
1610 | if (RDG_MEM_READS_STMT (rdg, i)) | |
1611 | ++res; | |
1612 | } | |
1613 | ||
1614 | return res; | |
1615 | } | |
1616 | ||
1617 | /* Returns the number of read and write operations in a PARTITION of | |
1618 | the RDG. */ | |
1619 | ||
1620 | static int | |
526ceb68 | 1621 | number_of_rw_in_partition (struct graph *rdg, partition *partition) |
2b8aee8e SP |
1622 | { |
1623 | int res = 0; | |
1624 | unsigned i; | |
1625 | bitmap_iterator ii; | |
1626 | ||
c61f8985 | 1627 | EXECUTE_IF_SET_IN_BITMAP (partition->stmts, 0, i, ii) |
2b8aee8e SP |
1628 | { |
1629 | if (RDG_MEM_WRITE_STMT (rdg, i)) | |
1630 | ++res; | |
1631 | ||
1632 | if (RDG_MEM_READS_STMT (rdg, i)) | |
1633 | ++res; | |
1634 | } | |
1635 | ||
1636 | return res; | |
1637 | } | |
1638 | ||
1639 | /* Returns true when one of the PARTITIONS contains all the read or | |
1640 | write operations of RDG. */ | |
1641 | ||
1642 | static bool | |
9771b263 | 1643 | partition_contains_all_rw (struct graph *rdg, |
526ceb68 | 1644 | vec<partition *> partitions) |
2b8aee8e SP |
1645 | { |
1646 | int i; | |
526ceb68 | 1647 | partition *partition; |
2b8aee8e SP |
1648 | int nrw = number_of_rw_in_rdg (rdg); |
1649 | ||
9771b263 | 1650 | FOR_EACH_VEC_ELT (partitions, i, partition) |
2b8aee8e SP |
1651 | if (nrw == number_of_rw_in_partition (rdg, partition)) |
1652 | return true; | |
1653 | ||
1654 | return false; | |
1655 | } | |
1656 | ||
447f3223 | 1657 | /* Compute partition dependence created by the data references in DRS1 |
a8745cc2 BC |
1658 | and DRS2, modify and return DIR according to that. IF ALIAS_DDR is |
1659 | not NULL, we record dependence introduced by possible alias between | |
1660 | two data references in ALIAS_DDRS; otherwise, we simply ignore such | |
1661 | dependence as if it doesn't exist at all. */ | |
447f3223 RB |
1662 | |
1663 | static int | |
4084ea5f | 1664 | pg_add_dependence_edges (struct graph *rdg, int dir, |
a8745cc2 | 1665 | bitmap drs1, bitmap drs2, vec<ddr_p> *alias_ddrs) |
447f3223 | 1666 | { |
a7a44c07 BC |
1667 | unsigned i, j; |
1668 | bitmap_iterator bi, bj; | |
1669 | data_reference_p dr1, dr2, saved_dr1; | |
447f3223 RB |
1670 | |
1671 | /* dependence direction - 0 is no dependence, -1 is back, | |
1672 | 1 is forth, 2 is both (we can stop then, merging will occur). */ | |
a7a44c07 BC |
1673 | EXECUTE_IF_SET_IN_BITMAP (drs1, 0, i, bi) |
1674 | { | |
1675 | dr1 = datarefs_vec[i]; | |
1676 | ||
1677 | EXECUTE_IF_SET_IN_BITMAP (drs2, 0, j, bj) | |
1678 | { | |
a8745cc2 BC |
1679 | int res, this_dir = 1; |
1680 | ddr_p ddr; | |
1681 | ||
a7a44c07 BC |
1682 | dr2 = datarefs_vec[j]; |
1683 | ||
1684 | /* Skip all <read, read> data dependence. */ | |
1685 | if (DR_IS_READ (dr1) && DR_IS_READ (dr2)) | |
1686 | continue; | |
1687 | ||
1688 | saved_dr1 = dr1; | |
a8745cc2 | 1689 | /* Re-shuffle data-refs to be in topological order. */ |
a7a44c07 BC |
1690 | if (rdg_vertex_for_stmt (rdg, DR_STMT (dr1)) |
1691 | > rdg_vertex_for_stmt (rdg, DR_STMT (dr2))) | |
1692 | { | |
1693 | std::swap (dr1, dr2); | |
1694 | this_dir = -this_dir; | |
1695 | } | |
17c5cbdf | 1696 | ddr = get_data_dependence (rdg, dr1, dr2); |
a7a44c07 | 1697 | if (DDR_ARE_DEPENDENT (ddr) == chrec_dont_know) |
a8745cc2 BC |
1698 | { |
1699 | this_dir = 0; | |
1700 | res = data_ref_compare_tree (DR_BASE_ADDRESS (dr1), | |
1701 | DR_BASE_ADDRESS (dr2)); | |
1702 | /* Be conservative. If data references are not well analyzed, | |
1703 | or the two data references have the same base address and | |
1704 | offset, add dependence and consider it alias to each other. | |
1705 | In other words, the dependence can not be resolved by | |
1706 | runtime alias check. */ | |
1707 | if (!DR_BASE_ADDRESS (dr1) || !DR_BASE_ADDRESS (dr2) | |
1708 | || !DR_OFFSET (dr1) || !DR_OFFSET (dr2) | |
1709 | || !DR_INIT (dr1) || !DR_INIT (dr2) | |
1710 | || !DR_STEP (dr1) || !tree_fits_uhwi_p (DR_STEP (dr1)) | |
1711 | || !DR_STEP (dr2) || !tree_fits_uhwi_p (DR_STEP (dr2)) | |
1712 | || res == 0) | |
1713 | this_dir = 2; | |
1714 | /* Data dependence could be resolved by runtime alias check, | |
1715 | record it in ALIAS_DDRS. */ | |
1716 | else if (alias_ddrs != NULL) | |
1717 | alias_ddrs->safe_push (ddr); | |
1718 | /* Or simply ignore it. */ | |
1719 | } | |
a7a44c07 BC |
1720 | else if (DDR_ARE_DEPENDENT (ddr) == NULL_TREE) |
1721 | { | |
1722 | if (DDR_REVERSED_P (ddr)) | |
a8745cc2 BC |
1723 | this_dir = -this_dir; |
1724 | ||
a7a44c07 BC |
1725 | /* Known dependences can still be unordered througout the |
1726 | iteration space, see gcc.dg/tree-ssa/ldist-16.c. */ | |
1727 | if (DDR_NUM_DIST_VECTS (ddr) != 1) | |
1728 | this_dir = 2; | |
1729 | /* If the overlap is exact preserve stmt order. */ | |
1730 | else if (lambda_vector_zerop (DDR_DIST_VECT (ddr, 0), 1)) | |
1731 | ; | |
a8745cc2 BC |
1732 | /* Else as the distance vector is lexicographic positive swap |
1733 | the dependence direction. */ | |
a7a44c07 | 1734 | else |
a8745cc2 | 1735 | this_dir = -this_dir; |
a7a44c07 BC |
1736 | } |
1737 | else | |
1738 | this_dir = 0; | |
a7a44c07 BC |
1739 | if (this_dir == 2) |
1740 | return 2; | |
1741 | else if (dir == 0) | |
1742 | dir = this_dir; | |
1743 | else if (this_dir != 0 && dir != this_dir) | |
1744 | return 2; | |
1745 | /* Shuffle "back" dr1. */ | |
1746 | dr1 = saved_dr1; | |
1747 | } | |
1748 | } | |
447f3223 RB |
1749 | return dir; |
1750 | } | |
1751 | ||
1752 | /* Compare postorder number of the partition graph vertices V1 and V2. */ | |
1753 | ||
1754 | static int | |
1755 | pgcmp (const void *v1_, const void *v2_) | |
1756 | { | |
1757 | const vertex *v1 = (const vertex *)v1_; | |
1758 | const vertex *v2 = (const vertex *)v2_; | |
1759 | return v2->post - v1->post; | |
1760 | } | |
2fd5894f | 1761 | |
a8745cc2 BC |
1762 | /* Data attached to vertices of partition dependence graph. */ |
1763 | struct pg_vdata | |
1764 | { | |
1765 | /* ID of the corresponding partition. */ | |
1766 | int id; | |
1767 | /* The partition. */ | |
1768 | struct partition *partition; | |
1769 | }; | |
1770 | ||
1771 | /* Data attached to edges of partition dependence graph. */ | |
1772 | struct pg_edata | |
1773 | { | |
1774 | /* If the dependence edge can be resolved by runtime alias check, | |
1775 | this vector contains data dependence relations for runtime alias | |
1776 | check. On the other hand, if the dependence edge is introduced | |
1777 | because of compilation time known data dependence, this vector | |
1778 | contains nothing. */ | |
1779 | vec<ddr_p> alias_ddrs; | |
1780 | }; | |
1781 | ||
1782 | /* Callback data for traversing edges in graph. */ | |
1783 | struct pg_edge_callback_data | |
1784 | { | |
1785 | /* Bitmap contains strong connected components should be merged. */ | |
1786 | bitmap sccs_to_merge; | |
1787 | /* Array constains component information for all vertices. */ | |
1788 | int *vertices_component; | |
1789 | /* Vector to record all data dependence relations which are needed | |
1790 | to break strong connected components by runtime alias checks. */ | |
1791 | vec<ddr_p> *alias_ddrs; | |
1792 | }; | |
1793 | ||
1794 | /* Initialize vertice's data for partition dependence graph PG with | |
1795 | PARTITIONS. */ | |
1796 | ||
1797 | static void | |
1798 | init_partition_graph_vertices (struct graph *pg, | |
1799 | vec<struct partition *> *partitions) | |
1800 | { | |
1801 | int i; | |
1802 | partition *partition; | |
1803 | struct pg_vdata *data; | |
1804 | ||
1805 | for (i = 0; partitions->iterate (i, &partition); ++i) | |
1806 | { | |
1807 | data = new pg_vdata; | |
1808 | pg->vertices[i].data = data; | |
1809 | data->id = i; | |
1810 | data->partition = partition; | |
1811 | } | |
1812 | } | |
1813 | ||
1814 | /* Add edge <I, J> to partition dependence graph PG. Attach vector of data | |
1815 | dependence relations to the EDGE if DDRS isn't NULL. */ | |
1816 | ||
1817 | static void | |
1818 | add_partition_graph_edge (struct graph *pg, int i, int j, vec<ddr_p> *ddrs) | |
1819 | { | |
1820 | struct graph_edge *e = add_edge (pg, i, j); | |
1821 | ||
1822 | /* If the edge is attached with data dependence relations, it means this | |
1823 | dependence edge can be resolved by runtime alias checks. */ | |
1824 | if (ddrs != NULL) | |
1825 | { | |
1826 | struct pg_edata *data = new pg_edata; | |
1827 | ||
1828 | gcc_assert (ddrs->length () > 0); | |
1829 | e->data = data; | |
1830 | data->alias_ddrs = vNULL; | |
1831 | data->alias_ddrs.safe_splice (*ddrs); | |
1832 | } | |
1833 | } | |
1834 | ||
1835 | /* Callback function for graph travesal algorithm. It returns true | |
1836 | if edge E should skipped when traversing the graph. */ | |
1837 | ||
1838 | static bool | |
1839 | pg_skip_alias_edge (struct graph_edge *e) | |
1840 | { | |
1841 | struct pg_edata *data = (struct pg_edata *)e->data; | |
1842 | return (data != NULL && data->alias_ddrs.length () > 0); | |
1843 | } | |
1844 | ||
1845 | /* Callback function freeing data attached to edge E of graph. */ | |
1846 | ||
1847 | static void | |
1848 | free_partition_graph_edata_cb (struct graph *, struct graph_edge *e, void *) | |
1849 | { | |
1850 | if (e->data != NULL) | |
1851 | { | |
1852 | struct pg_edata *data = (struct pg_edata *)e->data; | |
1853 | data->alias_ddrs.release (); | |
1854 | delete data; | |
1855 | } | |
1856 | } | |
1857 | ||
1858 | /* Free data attached to vertice of partition dependence graph PG. */ | |
1859 | ||
1860 | static void | |
1861 | free_partition_graph_vdata (struct graph *pg) | |
1862 | { | |
1863 | int i; | |
1864 | struct pg_vdata *data; | |
1865 | ||
1866 | for (i = 0; i < pg->n_vertices; ++i) | |
1867 | { | |
1868 | data = (struct pg_vdata *)pg->vertices[i].data; | |
1869 | delete data; | |
1870 | } | |
1871 | } | |
1872 | ||
1873 | /* Build and return partition dependence graph for PARTITIONS. RDG is | |
1874 | reduced dependence graph for the loop to be distributed. If IGNORE_ALIAS_P | |
1875 | is true, data dependence caused by possible alias between references | |
1876 | is ignored, as if it doesn't exist at all; otherwise all depdendences | |
1877 | are considered. */ | |
1878 | ||
1879 | static struct graph * | |
1880 | build_partition_graph (struct graph *rdg, | |
1881 | vec<struct partition *> *partitions, | |
1882 | bool ignore_alias_p) | |
1883 | { | |
1884 | int i, j; | |
1885 | struct partition *partition1, *partition2; | |
1886 | graph *pg = new_graph (partitions->length ()); | |
1887 | auto_vec<ddr_p> alias_ddrs, *alias_ddrs_p; | |
1888 | ||
1889 | alias_ddrs_p = ignore_alias_p ? NULL : &alias_ddrs; | |
1890 | ||
1891 | init_partition_graph_vertices (pg, partitions); | |
1892 | ||
1893 | for (i = 0; partitions->iterate (i, &partition1); ++i) | |
1894 | { | |
1895 | for (j = i + 1; partitions->iterate (j, &partition2); ++j) | |
1896 | { | |
1897 | /* dependence direction - 0 is no dependence, -1 is back, | |
1898 | 1 is forth, 2 is both (we can stop then, merging will occur). */ | |
1899 | int dir = 0; | |
1900 | ||
1901 | /* If the first partition has reduction, add back edge; if the | |
1902 | second partition has reduction, add forth edge. This makes | |
1903 | sure that reduction partition will be sorted as the last one. */ | |
1904 | if (partition_reduction_p (partition1)) | |
1905 | dir = -1; | |
1906 | else if (partition_reduction_p (partition2)) | |
1907 | dir = 1; | |
1908 | ||
1909 | /* Cleanup the temporary vector. */ | |
1910 | alias_ddrs.truncate (0); | |
1911 | ||
1912 | dir = pg_add_dependence_edges (rdg, dir, partition1->datarefs, | |
1913 | partition2->datarefs, alias_ddrs_p); | |
1914 | ||
1915 | /* Add edge to partition graph if there exists dependence. There | |
1916 | are two types of edges. One type edge is caused by compilation | |
1917 | time known dependence, this type can not be resolved by runtime | |
1918 | alias check. The other type can be resolved by runtime alias | |
1919 | check. */ | |
1920 | if (dir == 1 || dir == 2 | |
1921 | || alias_ddrs.length () > 0) | |
1922 | { | |
1923 | /* Attach data dependence relations to edge that can be resolved | |
1924 | by runtime alias check. */ | |
1925 | bool alias_edge_p = (dir != 1 && dir != 2); | |
1926 | add_partition_graph_edge (pg, i, j, | |
1927 | (alias_edge_p) ? &alias_ddrs : NULL); | |
1928 | } | |
1929 | if (dir == -1 || dir == 2 | |
1930 | || alias_ddrs.length () > 0) | |
1931 | { | |
1932 | /* Attach data dependence relations to edge that can be resolved | |
1933 | by runtime alias check. */ | |
1934 | bool alias_edge_p = (dir != -1 && dir != 2); | |
1935 | add_partition_graph_edge (pg, j, i, | |
1936 | (alias_edge_p) ? &alias_ddrs : NULL); | |
1937 | } | |
1938 | } | |
1939 | } | |
1940 | return pg; | |
1941 | } | |
1942 | ||
1943 | /* Sort partitions in PG by post order and store them in PARTITIONS. */ | |
1944 | ||
1945 | static void | |
1946 | sort_partitions_by_post_order (struct graph *pg, | |
1947 | vec<struct partition *> *partitions) | |
1948 | { | |
1949 | int i; | |
1950 | struct pg_vdata *data; | |
1951 | ||
1952 | /* Now order the remaining nodes in postorder. */ | |
1953 | qsort (pg->vertices, pg->n_vertices, sizeof (vertex), pgcmp); | |
1954 | partitions->truncate (0); | |
1955 | for (i = 0; i < pg->n_vertices; ++i) | |
1956 | { | |
1957 | data = (struct pg_vdata *)pg->vertices[i].data; | |
1958 | if (data->partition) | |
1959 | partitions->safe_push (data->partition); | |
1960 | } | |
1961 | } | |
1962 | ||
1963 | /* Given reduced dependence graph RDG merge strong connected components | |
6ff37519 BC |
1964 | of PARTITIONS. In this function, data dependence caused by possible |
1965 | alias between references is ignored, as if it doesn't exist at all. */ | |
a8745cc2 BC |
1966 | |
1967 | static void | |
1968 | merge_dep_scc_partitions (struct graph *rdg, | |
6ff37519 | 1969 | vec<struct partition *> *partitions) |
a8745cc2 BC |
1970 | { |
1971 | struct partition *partition1, *partition2; | |
1972 | struct pg_vdata *data; | |
6ff37519 | 1973 | graph *pg = build_partition_graph (rdg, partitions, true); |
a8745cc2 BC |
1974 | int i, j, num_sccs = graphds_scc (pg, NULL); |
1975 | ||
1976 | /* Strong connected compoenent means dependence cycle, we cannot distribute | |
1977 | them. So fuse them together. */ | |
1978 | if ((unsigned) num_sccs < partitions->length ()) | |
1979 | { | |
1980 | for (i = 0; i < num_sccs; ++i) | |
1981 | { | |
1982 | for (j = 0; partitions->iterate (j, &partition1); ++j) | |
1983 | if (pg->vertices[j].component == i) | |
1984 | break; | |
1985 | for (j = j + 1; partitions->iterate (j, &partition2); ++j) | |
1986 | if (pg->vertices[j].component == i) | |
1987 | { | |
1988 | partition_merge_into (NULL, partition1, | |
1989 | partition2, FUSE_SAME_SCC); | |
1990 | partition1->type = PTYPE_SEQUENTIAL; | |
1991 | (*partitions)[j] = NULL; | |
1992 | partition_free (partition2); | |
1993 | data = (struct pg_vdata *)pg->vertices[j].data; | |
1994 | data->partition = NULL; | |
1995 | } | |
1996 | } | |
a8745cc2 | 1997 | } |
aa1528b5 BC |
1998 | |
1999 | sort_partitions_by_post_order (pg, partitions); | |
a8745cc2 BC |
2000 | gcc_assert (partitions->length () == (unsigned)num_sccs); |
2001 | free_partition_graph_vdata (pg); | |
2002 | free_graph (pg); | |
2003 | } | |
2004 | ||
2005 | /* Callback function for traversing edge E in graph G. DATA is private | |
2006 | callback data. */ | |
2007 | ||
2008 | static void | |
2009 | pg_collect_alias_ddrs (struct graph *g, struct graph_edge *e, void *data) | |
2010 | { | |
2011 | int i, j, component; | |
2012 | struct pg_edge_callback_data *cbdata; | |
2013 | struct pg_edata *edata = (struct pg_edata *) e->data; | |
2014 | ||
2015 | /* If the edge doesn't have attached data dependence, it represents | |
2016 | compilation time known dependences. This type dependence cannot | |
2017 | be resolved by runtime alias check. */ | |
2018 | if (edata == NULL || edata->alias_ddrs.length () == 0) | |
2019 | return; | |
2020 | ||
2021 | cbdata = (struct pg_edge_callback_data *) data; | |
2022 | i = e->src; | |
2023 | j = e->dest; | |
2024 | component = cbdata->vertices_component[i]; | |
2025 | /* Vertices are topologically sorted according to compilation time | |
2026 | known dependences, so we can break strong connected components | |
2027 | by removing edges of the opposite direction, i.e, edges pointing | |
2028 | from vertice with smaller post number to vertice with bigger post | |
2029 | number. */ | |
2030 | if (g->vertices[i].post < g->vertices[j].post | |
2031 | /* We only need to remove edges connecting vertices in the same | |
2032 | strong connected component to break it. */ | |
2033 | && component == cbdata->vertices_component[j] | |
2034 | /* Check if we want to break the strong connected component or not. */ | |
2035 | && !bitmap_bit_p (cbdata->sccs_to_merge, component)) | |
2036 | cbdata->alias_ddrs->safe_splice (edata->alias_ddrs); | |
2037 | } | |
2038 | ||
2039 | /* This is the main function breaking strong conected components in | |
2040 | PARTITIONS giving reduced depdendence graph RDG. Store data dependence | |
2041 | relations for runtime alias check in ALIAS_DDRS. */ | |
2042 | ||
2043 | static void | |
2044 | break_alias_scc_partitions (struct graph *rdg, | |
2045 | vec<struct partition *> *partitions, | |
2046 | vec<ddr_p> *alias_ddrs) | |
2047 | { | |
2048 | int i, j, num_sccs, num_sccs_no_alias; | |
2049 | /* Build partition dependence graph. */ | |
2050 | graph *pg = build_partition_graph (rdg, partitions, false); | |
2051 | ||
2052 | alias_ddrs->truncate (0); | |
2053 | /* Find strong connected components in the graph, with all dependence edges | |
2054 | considered. */ | |
2055 | num_sccs = graphds_scc (pg, NULL); | |
2056 | /* All SCCs now can be broken by runtime alias checks because SCCs caused by | |
2057 | compilation time known dependences are merged before this function. */ | |
2058 | if ((unsigned) num_sccs < partitions->length ()) | |
2059 | { | |
2060 | struct pg_edge_callback_data cbdata; | |
2061 | auto_bitmap sccs_to_merge; | |
2062 | auto_vec<enum partition_type> scc_types; | |
2063 | struct partition *partition, *first; | |
2064 | ||
2065 | /* If all paritions in a SCC has the same type, we can simply merge the | |
2066 | SCC. This loop finds out such SCCS and record them in bitmap. */ | |
2067 | bitmap_set_range (sccs_to_merge, 0, (unsigned) num_sccs); | |
2068 | for (i = 0; i < num_sccs; ++i) | |
2069 | { | |
2070 | for (j = 0; partitions->iterate (j, &first); ++j) | |
2071 | if (pg->vertices[j].component == i) | |
2072 | break; | |
2073 | for (++j; partitions->iterate (j, &partition); ++j) | |
2074 | { | |
2075 | if (pg->vertices[j].component != i) | |
2076 | continue; | |
2077 | ||
2078 | if (first->type != partition->type) | |
2079 | { | |
2080 | bitmap_clear_bit (sccs_to_merge, i); | |
2081 | break; | |
2082 | } | |
2083 | } | |
2084 | } | |
2085 | ||
2086 | /* Initialize callback data for traversing. */ | |
2087 | cbdata.sccs_to_merge = sccs_to_merge; | |
2088 | cbdata.alias_ddrs = alias_ddrs; | |
2089 | cbdata.vertices_component = XNEWVEC (int, pg->n_vertices); | |
2090 | /* Record the component information which will be corrupted by next | |
2091 | graph scc finding call. */ | |
2092 | for (i = 0; i < pg->n_vertices; ++i) | |
2093 | cbdata.vertices_component[i] = pg->vertices[i].component; | |
2094 | ||
2095 | /* Collect data dependences for runtime alias checks to break SCCs. */ | |
2096 | if (bitmap_count_bits (sccs_to_merge) != (unsigned) num_sccs) | |
2097 | { | |
2098 | /* Run SCC finding algorithm again, with alias dependence edges | |
2099 | skipped. This is to topologically sort paritions according to | |
2100 | compilation time known dependence. Note the topological order | |
2101 | is stored in the form of pg's post order number. */ | |
2102 | num_sccs_no_alias = graphds_scc (pg, NULL, pg_skip_alias_edge); | |
2103 | gcc_assert (partitions->length () == (unsigned) num_sccs_no_alias); | |
2104 | /* With topological order, we can construct two subgraphs L and R. | |
2105 | L contains edge <x, y> where x < y in terms of post order, while | |
2106 | R contains edge <x, y> where x > y. Edges for compilation time | |
2107 | known dependence all fall in R, so we break SCCs by removing all | |
2108 | (alias) edges of in subgraph L. */ | |
2109 | for_each_edge (pg, pg_collect_alias_ddrs, &cbdata); | |
2110 | } | |
2111 | ||
2112 | /* For SCC that doesn't need to be broken, merge it. */ | |
2113 | for (i = 0; i < num_sccs; ++i) | |
2114 | { | |
2115 | if (!bitmap_bit_p (sccs_to_merge, i)) | |
2116 | continue; | |
2117 | ||
2118 | for (j = 0; partitions->iterate (j, &first); ++j) | |
2119 | if (cbdata.vertices_component[j] == i) | |
2120 | break; | |
2121 | for (++j; partitions->iterate (j, &partition); ++j) | |
2122 | { | |
2123 | struct pg_vdata *data; | |
2124 | ||
2125 | if (cbdata.vertices_component[j] != i) | |
2126 | continue; | |
2127 | ||
2128 | partition_merge_into (NULL, first, partition, FUSE_SAME_SCC); | |
2129 | (*partitions)[j] = NULL; | |
2130 | partition_free (partition); | |
2131 | data = (struct pg_vdata *)pg->vertices[j].data; | |
2132 | gcc_assert (data->id == j); | |
2133 | data->partition = NULL; | |
2134 | } | |
2135 | } | |
2136 | } | |
2137 | ||
2138 | sort_partitions_by_post_order (pg, partitions); | |
2139 | free_partition_graph_vdata (pg); | |
2140 | for_each_edge (pg, free_partition_graph_edata_cb, NULL); | |
2141 | free_graph (pg); | |
2142 | ||
2143 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
2144 | { | |
2145 | fprintf (dump_file, "Possible alias data dependence to break:\n"); | |
2146 | dump_data_dependence_relations (dump_file, *alias_ddrs); | |
2147 | } | |
2148 | } | |
2149 | ||
2150 | /* Compute and return an expression whose value is the segment length which | |
2151 | will be accessed by DR in NITERS iterations. */ | |
2152 | ||
2153 | static tree | |
2154 | data_ref_segment_size (struct data_reference *dr, tree niters) | |
2155 | { | |
2156 | tree segment_length; | |
2157 | ||
2158 | if (integer_zerop (DR_STEP (dr))) | |
2159 | segment_length = TYPE_SIZE_UNIT (TREE_TYPE (DR_REF (dr))); | |
2160 | else | |
2161 | segment_length = size_binop (MULT_EXPR, | |
2162 | fold_convert (sizetype, DR_STEP (dr)), | |
2163 | fold_convert (sizetype, niters)); | |
2164 | ||
2165 | return segment_length; | |
2166 | } | |
2167 | ||
2168 | /* Return true if LOOP's latch is dominated by statement for data reference | |
2169 | DR. */ | |
2170 | ||
2171 | static inline bool | |
2172 | latch_dominated_by_data_ref (struct loop *loop, data_reference *dr) | |
2173 | { | |
2174 | return dominated_by_p (CDI_DOMINATORS, single_exit (loop)->src, | |
2175 | gimple_bb (DR_STMT (dr))); | |
2176 | } | |
2177 | ||
2178 | /* Compute alias check pairs and store them in COMP_ALIAS_PAIRS for LOOP's | |
2179 | data dependence relations ALIAS_DDRS. */ | |
2180 | ||
2181 | static void | |
2182 | compute_alias_check_pairs (struct loop *loop, vec<ddr_p> *alias_ddrs, | |
2183 | vec<dr_with_seg_len_pair_t> *comp_alias_pairs) | |
2184 | { | |
2185 | unsigned int i; | |
2186 | unsigned HOST_WIDE_INT factor = 1; | |
2187 | tree niters_plus_one, niters = number_of_latch_executions (loop); | |
2188 | ||
2189 | gcc_assert (niters != NULL_TREE && niters != chrec_dont_know); | |
2190 | niters = fold_convert (sizetype, niters); | |
2191 | niters_plus_one = size_binop (PLUS_EXPR, niters, size_one_node); | |
2192 | ||
2193 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
2194 | fprintf (dump_file, "Creating alias check pairs:\n"); | |
2195 | ||
2196 | /* Iterate all data dependence relations and compute alias check pairs. */ | |
2197 | for (i = 0; i < alias_ddrs->length (); i++) | |
2198 | { | |
2199 | ddr_p ddr = (*alias_ddrs)[i]; | |
2200 | struct data_reference *dr_a = DDR_A (ddr); | |
2201 | struct data_reference *dr_b = DDR_B (ddr); | |
2202 | tree seg_length_a, seg_length_b; | |
2203 | int comp_res = data_ref_compare_tree (DR_BASE_ADDRESS (dr_a), | |
2204 | DR_BASE_ADDRESS (dr_b)); | |
2205 | ||
2206 | if (comp_res == 0) | |
2207 | comp_res = data_ref_compare_tree (DR_OFFSET (dr_a), DR_OFFSET (dr_b)); | |
2208 | gcc_assert (comp_res != 0); | |
2209 | ||
2210 | if (latch_dominated_by_data_ref (loop, dr_a)) | |
2211 | seg_length_a = data_ref_segment_size (dr_a, niters_plus_one); | |
2212 | else | |
2213 | seg_length_a = data_ref_segment_size (dr_a, niters); | |
2214 | ||
2215 | if (latch_dominated_by_data_ref (loop, dr_b)) | |
2216 | seg_length_b = data_ref_segment_size (dr_b, niters_plus_one); | |
2217 | else | |
2218 | seg_length_b = data_ref_segment_size (dr_b, niters); | |
2219 | ||
2220 | dr_with_seg_len_pair_t dr_with_seg_len_pair | |
2221 | (dr_with_seg_len (dr_a, seg_length_a), | |
2222 | dr_with_seg_len (dr_b, seg_length_b)); | |
2223 | ||
2224 | /* Canonicalize pairs by sorting the two DR members. */ | |
2225 | if (comp_res > 0) | |
2226 | std::swap (dr_with_seg_len_pair.first, dr_with_seg_len_pair.second); | |
2227 | ||
2228 | comp_alias_pairs->safe_push (dr_with_seg_len_pair); | |
2229 | } | |
2230 | ||
2231 | if (tree_fits_uhwi_p (niters)) | |
2232 | factor = tree_to_uhwi (niters); | |
2233 | ||
2234 | /* Prune alias check pairs. */ | |
2235 | prune_runtime_alias_test_list (comp_alias_pairs, factor); | |
2236 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
2237 | fprintf (dump_file, | |
2238 | "Improved number of alias checks from %d to %d\n", | |
2239 | alias_ddrs->length (), comp_alias_pairs->length ()); | |
2240 | } | |
2241 | ||
2242 | /* Given data dependence relations in ALIAS_DDRS, generate runtime alias | |
2243 | checks and version LOOP under condition of these runtime alias checks. */ | |
2244 | ||
2245 | static void | |
2246 | version_loop_by_alias_check (struct loop *loop, vec<ddr_p> *alias_ddrs) | |
2247 | { | |
2248 | profile_probability prob; | |
2249 | basic_block cond_bb; | |
2250 | struct loop *nloop; | |
2251 | tree lhs, arg0, cond_expr = NULL_TREE; | |
2252 | gimple_seq cond_stmts = NULL; | |
2253 | gimple *call_stmt = NULL; | |
2254 | auto_vec<dr_with_seg_len_pair_t> comp_alias_pairs; | |
2255 | ||
2256 | /* Generate code for runtime alias checks if necessary. */ | |
2257 | gcc_assert (alias_ddrs->length () > 0); | |
2258 | ||
2259 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
2260 | fprintf (dump_file, | |
2261 | "Version loop <%d> with runtime alias check\n", loop->num); | |
2262 | ||
2263 | compute_alias_check_pairs (loop, alias_ddrs, &comp_alias_pairs); | |
2264 | create_runtime_alias_checks (loop, &comp_alias_pairs, &cond_expr); | |
2265 | cond_expr = force_gimple_operand_1 (cond_expr, &cond_stmts, | |
8d2d0de9 | 2266 | is_gimple_val, NULL_TREE); |
a8745cc2 BC |
2267 | |
2268 | /* Depend on vectorizer to fold IFN_LOOP_DIST_ALIAS. */ | |
2269 | if (flag_tree_loop_vectorize) | |
2270 | { | |
2271 | /* Generate internal function call for loop distribution alias check. */ | |
2272 | call_stmt = gimple_build_call_internal (IFN_LOOP_DIST_ALIAS, | |
2273 | 2, NULL_TREE, cond_expr); | |
2274 | lhs = make_ssa_name (boolean_type_node); | |
2275 | gimple_call_set_lhs (call_stmt, lhs); | |
2276 | } | |
2277 | else | |
2278 | lhs = cond_expr; | |
2279 | ||
2280 | prob = profile_probability::guessed_always ().apply_scale (9, 10); | |
2281 | initialize_original_copy_tables (); | |
2282 | nloop = loop_version (loop, lhs, &cond_bb, prob, prob.invert (), | |
2283 | prob, prob.invert (), true); | |
2284 | free_original_copy_tables (); | |
2285 | /* Record the original loop number in newly generated loops. In case of | |
2286 | distribution, the original loop will be distributed and the new loop | |
2287 | is kept. */ | |
2288 | loop->orig_loop_num = nloop->num; | |
2289 | nloop->orig_loop_num = nloop->num; | |
2290 | nloop->dont_vectorize = true; | |
2291 | nloop->force_vectorize = false; | |
2292 | ||
2293 | if (call_stmt) | |
2294 | { | |
2295 | /* Record new loop's num in IFN_LOOP_DIST_ALIAS because the original | |
2296 | loop could be destroyed. */ | |
2297 | arg0 = build_int_cst (integer_type_node, loop->orig_loop_num); | |
2298 | gimple_call_set_arg (call_stmt, 0, arg0); | |
2299 | gimple_seq_add_stmt_without_update (&cond_stmts, call_stmt); | |
2300 | } | |
2301 | ||
2302 | if (cond_stmts) | |
2303 | { | |
2304 | gimple_stmt_iterator cond_gsi = gsi_last_bb (cond_bb); | |
2305 | gsi_insert_seq_before (&cond_gsi, cond_stmts, GSI_SAME_STMT); | |
2306 | } | |
2307 | update_ssa (TODO_update_ssa); | |
2308 | } | |
2309 | ||
2310 | /* Return true if loop versioning is needed to distrubute PARTITIONS. | |
2311 | ALIAS_DDRS are data dependence relations for runtime alias check. */ | |
2312 | ||
2313 | static inline bool | |
2314 | version_for_distribution_p (vec<struct partition *> *partitions, | |
2315 | vec<ddr_p> *alias_ddrs) | |
2316 | { | |
2317 | /* No need to version loop if we have only one partition. */ | |
2318 | if (partitions->length () == 1) | |
2319 | return false; | |
2320 | ||
2321 | /* Need to version loop if runtime alias check is necessary. */ | |
2322 | return (alias_ddrs->length () > 0); | |
2323 | } | |
2324 | ||
2325 | /* Fuse all partitions if necessary before finalizing distribution. */ | |
2326 | ||
2327 | static void | |
2328 | finalize_partitions (vec<struct partition *> *partitions, | |
2329 | vec<ddr_p> *alias_ddrs) | |
2330 | { | |
2331 | unsigned i; | |
2332 | struct partition *a, *partition; | |
2333 | ||
2334 | if (partitions->length () == 1 | |
2335 | || alias_ddrs->length () > 0) | |
2336 | return; | |
2337 | ||
2338 | a = (*partitions)[0]; | |
2339 | if (a->kind != PKIND_NORMAL) | |
2340 | return; | |
2341 | ||
2342 | for (i = 1; partitions->iterate (i, &partition); ++i) | |
2343 | { | |
2344 | /* Don't fuse if partition has different type or it is a builtin. */ | |
2345 | if (partition->type != a->type | |
2346 | || partition->kind != PKIND_NORMAL) | |
2347 | return; | |
2348 | } | |
2349 | ||
2350 | /* Fuse all partitions. */ | |
2351 | for (i = 1; partitions->iterate (i, &partition); ++i) | |
2352 | { | |
2353 | partition_merge_into (NULL, a, partition, FUSE_FINALIZE); | |
2354 | partition_free (partition); | |
2355 | } | |
2356 | partitions->truncate (1); | |
2357 | } | |
2358 | ||
2359 | /* Distributes the code from LOOP in such a way that producer statements | |
2360 | are placed before consumer statements. Tries to separate only the | |
2361 | statements from STMTS into separate loops. Returns the number of | |
2362 | distributed loops. Set NB_CALLS to number of generated builtin calls. | |
2363 | Set *DESTROY_P to whether LOOP needs to be destroyed. */ | |
dea61d92 SP |
2364 | |
2365 | static int | |
355fe088 | 2366 | distribute_loop (struct loop *loop, vec<gimple *> stmts, |
b71b7a8e | 2367 | control_dependences *cd, int *nb_calls, bool *destroy_p) |
dea61d92 | 2368 | { |
2fd5894f | 2369 | struct graph *rdg; |
526ceb68 | 2370 | partition *partition; |
be6b029b | 2371 | bool any_builtin; |
2fd5894f | 2372 | int i, nbp; |
dea61d92 | 2373 | |
c9326aef | 2374 | *destroy_p = false; |
826a536d | 2375 | *nb_calls = 0; |
4084ea5f | 2376 | loop_nest.create (0); |
2fd5894f | 2377 | if (!find_loop_nest (loop, &loop_nest)) |
4084ea5f BC |
2378 | { |
2379 | loop_nest.release (); | |
2380 | return 0; | |
2381 | } | |
2fd5894f | 2382 | |
9fafb14a | 2383 | datarefs_vec.create (20); |
4084ea5f | 2384 | rdg = build_rdg (loop, cd); |
2fd5894f RB |
2385 | if (!rdg) |
2386 | { | |
2387 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
2388 | fprintf (dump_file, | |
2389 | "Loop %d not distributed: failed to build the RDG.\n", | |
2390 | loop->num); | |
2391 | ||
4084ea5f | 2392 | loop_nest.release (); |
9fafb14a BC |
2393 | free_data_refs (datarefs_vec); |
2394 | return 0; | |
2395 | } | |
2396 | ||
2397 | if (datarefs_vec.length () > MAX_DATAREFS_NUM) | |
2398 | { | |
2399 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
2400 | fprintf (dump_file, | |
2401 | "Loop %d not distributed: too many memory references.\n", | |
2402 | loop->num); | |
2403 | ||
2404 | free_rdg (rdg); | |
2405 | loop_nest.release (); | |
2406 | free_data_refs (datarefs_vec); | |
2fd5894f RB |
2407 | return 0; |
2408 | } | |
2409 | ||
9fafb14a BC |
2410 | data_reference_p dref; |
2411 | for (i = 0; datarefs_vec.iterate (i, &dref); ++i) | |
2412 | dref->aux = (void *) (uintptr_t) i; | |
2413 | ||
2fd5894f RB |
2414 | if (dump_file && (dump_flags & TDF_DETAILS)) |
2415 | dump_rdg (dump_file, rdg); | |
2416 | ||
526ceb68 | 2417 | auto_vec<struct partition *, 3> partitions; |
2fd5894f | 2418 | rdg_build_partitions (rdg, stmts, &partitions); |
dea61d92 | 2419 | |
a8745cc2 BC |
2420 | auto_vec<ddr_p> alias_ddrs; |
2421 | ||
4a52eb19 BC |
2422 | auto_bitmap stmt_in_all_partitions; |
2423 | bitmap_copy (stmt_in_all_partitions, partitions[0]->stmts); | |
2424 | for (i = 1; partitions.iterate (i, &partition); ++i) | |
2425 | bitmap_and_into (stmt_in_all_partitions, partitions[i]->stmts); | |
2426 | ||
be6b029b | 2427 | any_builtin = false; |
9771b263 | 2428 | FOR_EACH_VEC_ELT (partitions, i, partition) |
be6b029b | 2429 | { |
4a52eb19 | 2430 | classify_partition (loop, rdg, partition, stmt_in_all_partitions); |
be6b029b RG |
2431 | any_builtin |= partition_builtin_p (partition); |
2432 | } | |
30d55936 | 2433 | |
447f3223 RB |
2434 | /* If we are only distributing patterns but did not detect any, |
2435 | simply bail out. */ | |
9fed7f3a RB |
2436 | if (!flag_tree_loop_distribution |
2437 | && !any_builtin) | |
2438 | { | |
2439 | nbp = 0; | |
2440 | goto ldist_done; | |
2441 | } | |
2442 | ||
447f3223 RB |
2443 | /* If we are only distributing patterns fuse all partitions that |
2444 | were not classified as builtins. This also avoids chopping | |
2445 | a loop into pieces, separated by builtin calls. That is, we | |
2446 | only want no or a single loop body remaining. */ | |
526ceb68 | 2447 | struct partition *into; |
447f3223 RB |
2448 | if (!flag_tree_loop_distribution) |
2449 | { | |
2450 | for (i = 0; partitions.iterate (i, &into); ++i) | |
2451 | if (!partition_builtin_p (into)) | |
2452 | break; | |
2453 | for (++i; partitions.iterate (i, &partition); ++i) | |
2454 | if (!partition_builtin_p (partition)) | |
2455 | { | |
f1eb4621 | 2456 | partition_merge_into (NULL, into, partition, FUSE_NON_BUILTIN); |
447f3223 RB |
2457 | partitions.unordered_remove (i); |
2458 | partition_free (partition); | |
2459 | i--; | |
2460 | } | |
2461 | } | |
2462 | ||
2463 | /* Due to limitations in the transform phase we have to fuse all | |
2464 | reduction partitions into the last partition so the existing | |
2465 | loop will contain all loop-closed PHI nodes. */ | |
2466 | for (i = 0; partitions.iterate (i, &into); ++i) | |
2467 | if (partition_reduction_p (into)) | |
2468 | break; | |
2469 | for (i = i + 1; partitions.iterate (i, &partition); ++i) | |
2470 | if (partition_reduction_p (partition)) | |
2471 | { | |
f1eb4621 | 2472 | partition_merge_into (rdg, into, partition, FUSE_REDUCTION); |
447f3223 RB |
2473 | partitions.unordered_remove (i); |
2474 | partition_free (partition); | |
2475 | i--; | |
2476 | } | |
2477 | ||
9fed7f3a RB |
2478 | /* Apply our simple cost model - fuse partitions with similar |
2479 | memory accesses. */ | |
9fed7f3a RB |
2480 | for (i = 0; partitions.iterate (i, &into); ++i) |
2481 | { | |
16eba420 | 2482 | bool changed = false; |
9fed7f3a RB |
2483 | if (partition_builtin_p (into)) |
2484 | continue; | |
2485 | for (int j = i + 1; | |
2486 | partitions.iterate (j, &partition); ++j) | |
2487 | { | |
95f7d11b | 2488 | if (share_memory_accesses (rdg, into, partition)) |
9fed7f3a | 2489 | { |
f1eb4621 | 2490 | partition_merge_into (rdg, into, partition, FUSE_SHARE_REF); |
447f3223 | 2491 | partitions.unordered_remove (j); |
9fed7f3a RB |
2492 | partition_free (partition); |
2493 | j--; | |
16eba420 | 2494 | changed = true; |
9fed7f3a RB |
2495 | } |
2496 | } | |
16eba420 RB |
2497 | /* If we fused 0 1 2 in step 1 to 0,2 1 as 0 and 2 have similar |
2498 | accesses when 1 and 2 have similar accesses but not 0 and 1 | |
2499 | then in the next iteration we will fail to consider merging | |
2500 | 1 into 0,2. So try again if we did any merging into 0. */ | |
2501 | if (changed) | |
2502 | i--; | |
9fed7f3a RB |
2503 | } |
2504 | ||
447f3223 RB |
2505 | /* Build the partition dependency graph. */ |
2506 | if (partitions.length () > 1) | |
c014f6f5 | 2507 | { |
6ff37519 | 2508 | merge_dep_scc_partitions (rdg, &partitions); |
a8745cc2 | 2509 | alias_ddrs.truncate (0); |
6ff37519 | 2510 | if (partitions.length () > 1) |
a8745cc2 | 2511 | break_alias_scc_partitions (rdg, &partitions, &alias_ddrs); |
b9fc0497 RB |
2512 | } |
2513 | ||
a8745cc2 BC |
2514 | finalize_partitions (&partitions, &alias_ddrs); |
2515 | ||
9771b263 | 2516 | nbp = partitions.length (); |
a4293fa6 | 2517 | if (nbp == 0 |
9771b263 DN |
2518 | || (nbp == 1 && !partition_builtin_p (partitions[0])) |
2519 | || (nbp > 1 && partition_contains_all_rw (rdg, partitions))) | |
c014f6f5 RG |
2520 | { |
2521 | nbp = 0; | |
2522 | goto ldist_done; | |
2523 | } | |
dea61d92 | 2524 | |
a8745cc2 BC |
2525 | if (version_for_distribution_p (&partitions, &alias_ddrs)) |
2526 | version_loop_by_alias_check (loop, &alias_ddrs); | |
2527 | ||
dea61d92 | 2528 | if (dump_file && (dump_flags & TDF_DETAILS)) |
a8745cc2 BC |
2529 | { |
2530 | fprintf (dump_file, | |
2531 | "distribute loop <%d> into partitions:\n", loop->num); | |
2532 | dump_rdg_partitions (dump_file, partitions); | |
2533 | } | |
dea61d92 | 2534 | |
9771b263 | 2535 | FOR_EACH_VEC_ELT (partitions, i, partition) |
826a536d RB |
2536 | { |
2537 | if (partition_builtin_p (partition)) | |
2538 | (*nb_calls)++; | |
b71b7a8e | 2539 | *destroy_p |= generate_code_for_partition (loop, partition, i < nbp - 1); |
826a536d | 2540 | } |
dea61d92 | 2541 | |
dea61d92 | 2542 | ldist_done: |
4084ea5f | 2543 | loop_nest.release (); |
9fafb14a | 2544 | free_data_refs (datarefs_vec); |
17c5cbdf BC |
2545 | for (hash_table<ddr_hasher>::iterator iter = ddrs_table.begin (); |
2546 | iter != ddrs_table.end (); ++iter) | |
2547 | { | |
2548 | free_dependence_relation (*iter); | |
2549 | *iter = NULL; | |
2550 | } | |
2551 | ddrs_table.empty (); | |
dea61d92 | 2552 | |
9771b263 | 2553 | FOR_EACH_VEC_ELT (partitions, i, partition) |
c61f8985 | 2554 | partition_free (partition); |
dea61d92 | 2555 | |
dea61d92 | 2556 | free_rdg (rdg); |
826a536d | 2557 | return nbp - *nb_calls; |
dea61d92 SP |
2558 | } |
2559 | ||
2560 | /* Distribute all loops in the current function. */ | |
2561 | ||
be55bfe6 TS |
2562 | namespace { |
2563 | ||
2564 | const pass_data pass_data_loop_distribution = | |
2565 | { | |
2566 | GIMPLE_PASS, /* type */ | |
2567 | "ldist", /* name */ | |
2568 | OPTGROUP_LOOP, /* optinfo_flags */ | |
be55bfe6 TS |
2569 | TV_TREE_LOOP_DISTRIBUTION, /* tv_id */ |
2570 | ( PROP_cfg | PROP_ssa ), /* properties_required */ | |
2571 | 0, /* properties_provided */ | |
2572 | 0, /* properties_destroyed */ | |
2573 | 0, /* todo_flags_start */ | |
3bea341f | 2574 | 0, /* todo_flags_finish */ |
be55bfe6 TS |
2575 | }; |
2576 | ||
2577 | class pass_loop_distribution : public gimple_opt_pass | |
2578 | { | |
2579 | public: | |
2580 | pass_loop_distribution (gcc::context *ctxt) | |
2581 | : gimple_opt_pass (pass_data_loop_distribution, ctxt) | |
2582 | {} | |
2583 | ||
2584 | /* opt_pass methods: */ | |
2585 | virtual bool gate (function *) | |
2586 | { | |
2587 | return flag_tree_loop_distribution | |
2588 | || flag_tree_loop_distribute_patterns; | |
2589 | } | |
2590 | ||
2591 | virtual unsigned int execute (function *); | |
2592 | ||
2593 | }; // class pass_loop_distribution | |
2594 | ||
2595 | unsigned int | |
2596 | pass_loop_distribution::execute (function *fun) | |
dea61d92 SP |
2597 | { |
2598 | struct loop *loop; | |
c014f6f5 | 2599 | bool changed = false; |
1fa0c180 | 2600 | basic_block bb; |
36875e8f | 2601 | control_dependences *cd = NULL; |
b71b7a8e | 2602 | auto_vec<loop_p> loops_to_be_destroyed; |
1fa0c180 | 2603 | |
773d9217 BC |
2604 | if (number_of_loops (fun) <= 1) |
2605 | return 0; | |
2606 | ||
3be57c56 BC |
2607 | /* Compute topological order for basic blocks. Topological order is |
2608 | needed because data dependence is computed for data references in | |
2609 | lexicographical order. */ | |
2610 | if (bb_top_order_index == NULL) | |
2611 | { | |
3fb82452 | 2612 | int rpo_num; |
3be57c56 BC |
2613 | int *rpo = XNEWVEC (int, last_basic_block_for_fn (cfun)); |
2614 | ||
2615 | bb_top_order_index = XNEWVEC (int, last_basic_block_for_fn (cfun)); | |
3fb82452 BC |
2616 | bb_top_order_index_size = last_basic_block_for_fn (cfun); |
2617 | rpo_num = pre_and_rev_post_order_compute_fn (cfun, NULL, rpo, true); | |
2618 | for (int i = 0; i < rpo_num; i++) | |
3be57c56 BC |
2619 | bb_top_order_index[rpo[i]] = i; |
2620 | ||
2621 | free (rpo); | |
2622 | } | |
2623 | ||
be55bfe6 | 2624 | FOR_ALL_BB_FN (bb, fun) |
1fa0c180 RG |
2625 | { |
2626 | gimple_stmt_iterator gsi; | |
2627 | for (gsi = gsi_start_phis (bb); !gsi_end_p (gsi); gsi_next (&gsi)) | |
2628 | gimple_set_uid (gsi_stmt (gsi), -1); | |
2629 | for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi)) | |
2630 | gimple_set_uid (gsi_stmt (gsi), -1); | |
2631 | } | |
dea61d92 | 2632 | |
c014f6f5 RG |
2633 | /* We can at the moment only distribute non-nested loops, thus restrict |
2634 | walking to innermost loops. */ | |
f0bd40b1 | 2635 | FOR_EACH_LOOP (loop, LI_ONLY_INNERMOST) |
dea61d92 | 2636 | { |
355fe088 | 2637 | auto_vec<gimple *> work_list; |
be6b029b | 2638 | basic_block *bbs; |
0e20c89f | 2639 | int num = loop->num; |
be6b029b | 2640 | unsigned int i; |
a3357f7d RG |
2641 | |
2642 | /* If the loop doesn't have a single exit we will fail anyway, | |
2643 | so do that early. */ | |
2644 | if (!single_exit (loop)) | |
2645 | continue; | |
dea61d92 | 2646 | |
f56f2d33 JH |
2647 | /* Only optimize hot loops. */ |
2648 | if (!optimize_loop_for_speed_p (loop)) | |
2649 | continue; | |
2650 | ||
6ff37519 BC |
2651 | /* Don't distribute loop if niters is unknown. */ |
2652 | tree niters = number_of_latch_executions (loop); | |
2653 | if (niters == NULL_TREE || niters == chrec_dont_know) | |
2654 | continue; | |
2655 | ||
be6b029b RG |
2656 | /* Initialize the worklist with stmts we seed the partitions with. */ |
2657 | bbs = get_loop_body_in_dom_order (loop); | |
2658 | for (i = 0; i < loop->num_nodes; ++i) | |
2659 | { | |
538dd0b7 DM |
2660 | for (gphi_iterator gsi = gsi_start_phis (bbs[i]); |
2661 | !gsi_end_p (gsi); | |
2662 | gsi_next (&gsi)) | |
deb6c11a | 2663 | { |
538dd0b7 | 2664 | gphi *phi = gsi.phi (); |
deb6c11a RB |
2665 | if (virtual_operand_p (gimple_phi_result (phi))) |
2666 | continue; | |
2667 | /* Distribute stmts which have defs that are used outside of | |
be55bfe6 | 2668 | the loop. */ |
deb6c11a RB |
2669 | if (!stmt_has_scalar_dependences_outside_loop (loop, phi)) |
2670 | continue; | |
2671 | work_list.safe_push (phi); | |
2672 | } | |
538dd0b7 DM |
2673 | for (gimple_stmt_iterator gsi = gsi_start_bb (bbs[i]); |
2674 | !gsi_end_p (gsi); | |
2675 | gsi_next (&gsi)) | |
be6b029b | 2676 | { |
355fe088 | 2677 | gimple *stmt = gsi_stmt (gsi); |
83a95546 RB |
2678 | |
2679 | /* If there is a stmt with side-effects bail out - we | |
be55bfe6 | 2680 | cannot and should not distribute this loop. */ |
83a95546 RB |
2681 | if (gimple_has_side_effects (stmt)) |
2682 | { | |
2683 | work_list.truncate (0); | |
2684 | goto out; | |
2685 | } | |
2686 | ||
b9fc0497 | 2687 | /* Distribute stmts which have defs that are used outside of |
be55bfe6 | 2688 | the loop. */ |
b9fc0497 RB |
2689 | if (stmt_has_scalar_dependences_outside_loop (loop, stmt)) |
2690 | ; | |
2691 | /* Otherwise only distribute stores for now. */ | |
e179190c | 2692 | else if (!gimple_vdef (stmt)) |
be6b029b RG |
2693 | continue; |
2694 | ||
9771b263 | 2695 | work_list.safe_push (stmt); |
be6b029b RG |
2696 | } |
2697 | } | |
83a95546 | 2698 | out: |
be6b029b | 2699 | free (bbs); |
c014f6f5 | 2700 | |
826a536d RB |
2701 | int nb_generated_loops = 0; |
2702 | int nb_generated_calls = 0; | |
2703 | location_t loc = find_loop_location (loop); | |
9771b263 | 2704 | if (work_list.length () > 0) |
36875e8f RB |
2705 | { |
2706 | if (!cd) | |
2707 | { | |
ca406576 | 2708 | calculate_dominance_info (CDI_DOMINATORS); |
36875e8f | 2709 | calculate_dominance_info (CDI_POST_DOMINATORS); |
30fd2977 | 2710 | cd = new control_dependences (); |
36875e8f RB |
2711 | free_dominance_info (CDI_POST_DOMINATORS); |
2712 | } | |
b71b7a8e | 2713 | bool destroy_p; |
826a536d | 2714 | nb_generated_loops = distribute_loop (loop, work_list, cd, |
b71b7a8e RB |
2715 | &nb_generated_calls, |
2716 | &destroy_p); | |
2717 | if (destroy_p) | |
2718 | loops_to_be_destroyed.safe_push (loop); | |
36875e8f | 2719 | } |
c014f6f5 | 2720 | |
826a536d | 2721 | if (nb_generated_loops + nb_generated_calls > 0) |
dea61d92 | 2722 | { |
826a536d RB |
2723 | changed = true; |
2724 | dump_printf_loc (MSG_OPTIMIZED_LOCATIONS, | |
2725 | loc, "Loop %d distributed: split to %d loops " | |
2726 | "and %d library calls.\n", | |
2727 | num, nb_generated_loops, nb_generated_calls); | |
dea61d92 | 2728 | } |
826a536d RB |
2729 | else if (dump_file && (dump_flags & TDF_DETAILS)) |
2730 | fprintf (dump_file, "Loop %d is the same.\n", num); | |
dea61d92 SP |
2731 | } |
2732 | ||
36875e8f RB |
2733 | if (cd) |
2734 | delete cd; | |
2735 | ||
3be57c56 BC |
2736 | if (bb_top_order_index != NULL) |
2737 | { | |
2738 | free (bb_top_order_index); | |
2739 | bb_top_order_index = NULL; | |
2740 | bb_top_order_index_size = 0; | |
2741 | } | |
2742 | ||
c014f6f5 RG |
2743 | if (changed) |
2744 | { | |
30fd2977 RB |
2745 | /* Destroy loop bodies that could not be reused. Do this late as we |
2746 | otherwise can end up refering to stale data in control dependences. */ | |
2747 | unsigned i; | |
2748 | FOR_EACH_VEC_ELT (loops_to_be_destroyed, i, loop) | |
3be57c56 | 2749 | destroy_loop (loop); |
30fd2977 | 2750 | |
d0ed943c RB |
2751 | /* Cached scalar evolutions now may refer to wrong or non-existing |
2752 | loops. */ | |
2753 | scev_reset_htab (); | |
be55bfe6 | 2754 | mark_virtual_operands_for_renaming (fun); |
c014f6f5 RG |
2755 | rewrite_into_loop_closed_ssa (NULL, TODO_update_ssa); |
2756 | } | |
2757 | ||
b2b29377 | 2758 | checking_verify_loop_structure (); |
c014f6f5 | 2759 | |
5006671f | 2760 | return 0; |
dea61d92 SP |
2761 | } |
2762 | ||
27a4cd48 DM |
2763 | } // anon namespace |
2764 | ||
2765 | gimple_opt_pass * | |
2766 | make_pass_loop_distribution (gcc::context *ctxt) | |
2767 | { | |
2768 | return new pass_loop_distribution (ctxt); | |
2769 | } |