]>
Commit | Line | Data |
---|---|---|
bbc8a8dc | 1 | /* Predictive commoning. |
e08120b1 | 2 | Copyright (C) 2005, 2007, 2008, 2009, 2010, 2011, 2012 |
c75c517d | 3 | Free Software Foundation, Inc. |
b8698a0f | 4 | |
bbc8a8dc | 5 | This file is part of GCC. |
b8698a0f | 6 | |
bbc8a8dc ZD |
7 | GCC is free software; you can redistribute it and/or modify it |
8 | under the terms of the GNU General Public License as published by the | |
9dcd6f09 | 9 | Free Software Foundation; either version 3, or (at your option) any |
bbc8a8dc | 10 | later version. |
b8698a0f | 11 | |
bbc8a8dc ZD |
12 | GCC is distributed in the hope that it will be useful, but WITHOUT |
13 | ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or | |
14 | FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License | |
15 | for more details. | |
b8698a0f | 16 | |
bbc8a8dc | 17 | You should have received a copy of the GNU General Public License |
9dcd6f09 NC |
18 | along with GCC; see the file COPYING3. If not see |
19 | <http://www.gnu.org/licenses/>. */ | |
bbc8a8dc ZD |
20 | |
21 | /* This file implements the predictive commoning optimization. Predictive | |
22 | commoning can be viewed as CSE around a loop, and with some improvements, | |
23 | as generalized strength reduction-- i.e., reusing values computed in | |
24 | earlier iterations of a loop in the later ones. So far, the pass only | |
25 | handles the most useful case, that is, reusing values of memory references. | |
26 | If you think this is all just a special case of PRE, you are sort of right; | |
27 | however, concentrating on loops is simpler, and makes it possible to | |
28 | incorporate data dependence analysis to detect the opportunities, perform | |
29 | loop unrolling to avoid copies together with renaming immediately, | |
30 | and if needed, we could also take register pressure into account. | |
31 | ||
32 | Let us demonstrate what is done on an example: | |
b8698a0f | 33 | |
bbc8a8dc ZD |
34 | for (i = 0; i < 100; i++) |
35 | { | |
36 | a[i+2] = a[i] + a[i+1]; | |
37 | b[10] = b[10] + i; | |
38 | c[i] = c[99 - i]; | |
39 | d[i] = d[i + 1]; | |
40 | } | |
41 | ||
42 | 1) We find data references in the loop, and split them to mutually | |
43 | independent groups (i.e., we find components of a data dependence | |
44 | graph). We ignore read-read dependences whose distance is not constant. | |
45 | (TODO -- we could also ignore antidependences). In this example, we | |
46 | find the following groups: | |
47 | ||
48 | a[i]{read}, a[i+1]{read}, a[i+2]{write} | |
49 | b[10]{read}, b[10]{write} | |
50 | c[99 - i]{read}, c[i]{write} | |
51 | d[i + 1]{read}, d[i]{write} | |
52 | ||
53 | 2) Inside each of the group, we verify several conditions: | |
54 | a) all the references must differ in indices only, and the indices | |
55 | must all have the same step | |
56 | b) the references must dominate loop latch (and thus, they must be | |
57 | ordered by dominance relation). | |
58 | c) the distance of the indices must be a small multiple of the step | |
59 | We are then able to compute the difference of the references (# of | |
60 | iterations before they point to the same place as the first of them). | |
61 | Also, in case there are writes in the loop, we split the groups into | |
62 | chains whose head is the write whose values are used by the reads in | |
63 | the same chain. The chains are then processed independently, | |
64 | making the further transformations simpler. Also, the shorter chains | |
65 | need the same number of registers, but may require lower unrolling | |
66 | factor in order to get rid of the copies on the loop latch. | |
b8698a0f | 67 | |
bbc8a8dc ZD |
68 | In our example, we get the following chains (the chain for c is invalid). |
69 | ||
70 | a[i]{read,+0}, a[i+1]{read,-1}, a[i+2]{write,-2} | |
71 | b[10]{read,+0}, b[10]{write,+0} | |
72 | d[i + 1]{read,+0}, d[i]{write,+1} | |
73 | ||
74 | 3) For each read, we determine the read or write whose value it reuses, | |
75 | together with the distance of this reuse. I.e. we take the last | |
76 | reference before it with distance 0, or the last of the references | |
77 | with the smallest positive distance to the read. Then, we remove | |
78 | the references that are not used in any of these chains, discard the | |
79 | empty groups, and propagate all the links so that they point to the | |
b8698a0f | 80 | single root reference of the chain (adjusting their distance |
bbc8a8dc ZD |
81 | appropriately). Some extra care needs to be taken for references with |
82 | step 0. In our example (the numbers indicate the distance of the | |
83 | reuse), | |
84 | ||
85 | a[i] --> (*) 2, a[i+1] --> (*) 1, a[i+2] (*) | |
86 | b[10] --> (*) 1, b[10] (*) | |
87 | ||
88 | 4) The chains are combined together if possible. If the corresponding | |
89 | elements of two chains are always combined together with the same | |
90 | operator, we remember just the result of this combination, instead | |
91 | of remembering the values separately. We may need to perform | |
92 | reassociation to enable combining, for example | |
93 | ||
94 | e[i] + f[i+1] + e[i+1] + f[i] | |
95 | ||
96 | can be reassociated as | |
97 | ||
98 | (e[i] + f[i]) + (e[i+1] + f[i+1]) | |
99 | ||
100 | and we can combine the chains for e and f into one chain. | |
101 | ||
102 | 5) For each root reference (end of the chain) R, let N be maximum distance | |
073a8998 | 103 | of a reference reusing its value. Variables R0 up to RN are created, |
bbc8a8dc ZD |
104 | together with phi nodes that transfer values from R1 .. RN to |
105 | R0 .. R(N-1). | |
106 | Initial values are loaded to R0..R(N-1) (in case not all references | |
107 | must necessarily be accessed and they may trap, we may fail here; | |
108 | TODO sometimes, the loads could be guarded by a check for the number | |
109 | of iterations). Values loaded/stored in roots are also copied to | |
110 | RN. Other reads are replaced with the appropriate variable Ri. | |
111 | Everything is put to SSA form. | |
112 | ||
113 | As a small improvement, if R0 is dead after the root (i.e., all uses of | |
114 | the value with the maximum distance dominate the root), we can avoid | |
115 | creating RN and use R0 instead of it. | |
116 | ||
117 | In our example, we get (only the parts concerning a and b are shown): | |
118 | for (i = 0; i < 100; i++) | |
119 | { | |
120 | f = phi (a[0], s); | |
121 | s = phi (a[1], f); | |
122 | x = phi (b[10], x); | |
123 | ||
124 | f = f + s; | |
125 | a[i+2] = f; | |
126 | x = x + i; | |
127 | b[10] = x; | |
128 | } | |
129 | ||
130 | 6) Factor F for unrolling is determined as the smallest common multiple of | |
131 | (N + 1) for each root reference (N for references for that we avoided | |
132 | creating RN). If F and the loop is small enough, loop is unrolled F | |
133 | times. The stores to RN (R0) in the copies of the loop body are | |
134 | periodically replaced with R0, R1, ... (R1, R2, ...), so that they can | |
135 | be coalesced and the copies can be eliminated. | |
b8698a0f | 136 | |
bbc8a8dc ZD |
137 | TODO -- copy propagation and other optimizations may change the live |
138 | ranges of the temporary registers and prevent them from being coalesced; | |
139 | this may increase the register pressure. | |
140 | ||
141 | In our case, F = 2 and the (main loop of the) result is | |
142 | ||
143 | for (i = 0; i < ...; i += 2) | |
144 | { | |
145 | f = phi (a[0], f); | |
146 | s = phi (a[1], s); | |
147 | x = phi (b[10], x); | |
148 | ||
149 | f = f + s; | |
150 | a[i+2] = f; | |
151 | x = x + i; | |
152 | b[10] = x; | |
153 | ||
154 | s = s + f; | |
155 | a[i+3] = s; | |
156 | x = x + i; | |
157 | b[10] = x; | |
158 | } | |
159 | ||
160 | TODO -- stores killing other stores can be taken into account, e.g., | |
161 | for (i = 0; i < n; i++) | |
162 | { | |
163 | a[i] = 1; | |
164 | a[i+2] = 2; | |
165 | } | |
166 | ||
167 | can be replaced with | |
168 | ||
169 | t0 = a[0]; | |
170 | t1 = a[1]; | |
171 | for (i = 0; i < n; i++) | |
172 | { | |
173 | a[i] = 1; | |
174 | t2 = 2; | |
175 | t0 = t1; | |
176 | t1 = t2; | |
177 | } | |
178 | a[n] = t0; | |
179 | a[n+1] = t1; | |
180 | ||
181 | The interesting part is that this would generalize store motion; still, since | |
182 | sm is performed elsewhere, it does not seem that important. | |
183 | ||
184 | Predictive commoning can be generalized for arbitrary computations (not | |
185 | just memory loads), and also nontrivial transfer functions (e.g., replacing | |
186 | i * i with ii_last + 2 * i + 1), to generalize strength reduction. */ | |
187 | ||
188 | #include "config.h" | |
189 | #include "system.h" | |
190 | #include "coretypes.h" | |
191 | #include "tm.h" | |
192 | #include "tree.h" | |
193 | #include "tm_p.h" | |
194 | #include "cfgloop.h" | |
195 | #include "tree-flow.h" | |
196 | #include "ggc.h" | |
197 | #include "tree-data-ref.h" | |
198 | #include "tree-scalar-evolution.h" | |
199 | #include "tree-chrec.h" | |
200 | #include "params.h" | |
cf835838 | 201 | #include "gimple-pretty-print.h" |
bbc8a8dc ZD |
202 | #include "tree-pass.h" |
203 | #include "tree-affine.h" | |
204 | #include "tree-inline.h" | |
205 | ||
206 | /* The maximum number of iterations between the considered memory | |
207 | references. */ | |
208 | ||
209 | #define MAX_DISTANCE (target_avail_regs < 16 ? 4 : 8) | |
b8698a0f | 210 | |
726a989a RB |
211 | /* Data references (or phi nodes that carry data reference values across |
212 | loop iterations). */ | |
bbc8a8dc | 213 | |
7e5487a2 | 214 | typedef struct dref_d |
bbc8a8dc ZD |
215 | { |
216 | /* The reference itself. */ | |
217 | struct data_reference *ref; | |
218 | ||
219 | /* The statement in that the reference appears. */ | |
726a989a RB |
220 | gimple stmt; |
221 | ||
222 | /* In case that STMT is a phi node, this field is set to the SSA name | |
223 | defined by it in replace_phis_by_defined_names (in order to avoid | |
224 | pointing to phi node that got reallocated in the meantime). */ | |
225 | tree name_defined_by_phi; | |
bbc8a8dc ZD |
226 | |
227 | /* Distance of the reference from the root of the chain (in number of | |
228 | iterations of the loop). */ | |
229 | unsigned distance; | |
230 | ||
231 | /* Number of iterations offset from the first reference in the component. */ | |
232 | double_int offset; | |
233 | ||
234 | /* Number of the reference in a component, in dominance ordering. */ | |
235 | unsigned pos; | |
236 | ||
237 | /* True if the memory reference is always accessed when the loop is | |
238 | entered. */ | |
239 | unsigned always_accessed : 1; | |
240 | } *dref; | |
241 | ||
242 | DEF_VEC_P (dref); | |
243 | DEF_VEC_ALLOC_P (dref, heap); | |
244 | ||
245 | /* Type of the chain of the references. */ | |
246 | ||
247 | enum chain_type | |
248 | { | |
249 | /* The addresses of the references in the chain are constant. */ | |
250 | CT_INVARIANT, | |
251 | ||
252 | /* There are only loads in the chain. */ | |
253 | CT_LOAD, | |
254 | ||
255 | /* Root of the chain is store, the rest are loads. */ | |
256 | CT_STORE_LOAD, | |
257 | ||
258 | /* A combination of two chains. */ | |
259 | CT_COMBINATION | |
260 | }; | |
261 | ||
262 | /* Chains of data references. */ | |
263 | ||
264 | typedef struct chain | |
265 | { | |
266 | /* Type of the chain. */ | |
267 | enum chain_type type; | |
268 | ||
269 | /* For combination chains, the operator and the two chains that are | |
270 | combined, and the type of the result. */ | |
82d6e6fc | 271 | enum tree_code op; |
bbc8a8dc ZD |
272 | tree rslt_type; |
273 | struct chain *ch1, *ch2; | |
274 | ||
275 | /* The references in the chain. */ | |
276 | VEC(dref,heap) *refs; | |
277 | ||
278 | /* The maximum distance of the reference in the chain from the root. */ | |
279 | unsigned length; | |
280 | ||
281 | /* The variables used to copy the value throughout iterations. */ | |
282 | VEC(tree,heap) *vars; | |
283 | ||
284 | /* Initializers for the variables. */ | |
285 | VEC(tree,heap) *inits; | |
286 | ||
287 | /* True if there is a use of a variable with the maximal distance | |
288 | that comes after the root in the loop. */ | |
289 | unsigned has_max_use_after : 1; | |
290 | ||
291 | /* True if all the memory references in the chain are always accessed. */ | |
292 | unsigned all_always_accessed : 1; | |
293 | ||
294 | /* True if this chain was combined together with some other chain. */ | |
295 | unsigned combined : 1; | |
296 | } *chain_p; | |
297 | ||
298 | DEF_VEC_P (chain_p); | |
299 | DEF_VEC_ALLOC_P (chain_p, heap); | |
300 | ||
301 | /* Describes the knowledge about the step of the memory references in | |
302 | the component. */ | |
303 | ||
304 | enum ref_step_type | |
305 | { | |
306 | /* The step is zero. */ | |
307 | RS_INVARIANT, | |
308 | ||
309 | /* The step is nonzero. */ | |
310 | RS_NONZERO, | |
311 | ||
312 | /* The step may or may not be nonzero. */ | |
313 | RS_ANY | |
314 | }; | |
315 | ||
316 | /* Components of the data dependence graph. */ | |
317 | ||
318 | struct component | |
319 | { | |
320 | /* The references in the component. */ | |
321 | VEC(dref,heap) *refs; | |
322 | ||
323 | /* What we know about the step of the references in the component. */ | |
324 | enum ref_step_type comp_step; | |
325 | ||
326 | /* Next component in the list. */ | |
327 | struct component *next; | |
328 | }; | |
329 | ||
330 | /* Bitmap of ssa names defined by looparound phi nodes covered by chains. */ | |
331 | ||
332 | static bitmap looparound_phis; | |
333 | ||
334 | /* Cache used by tree_to_aff_combination_expand. */ | |
335 | ||
336 | static struct pointer_map_t *name_expansions; | |
337 | ||
338 | /* Dumps data reference REF to FILE. */ | |
339 | ||
340 | extern void dump_dref (FILE *, dref); | |
341 | void | |
342 | dump_dref (FILE *file, dref ref) | |
343 | { | |
344 | if (ref->ref) | |
345 | { | |
346 | fprintf (file, " "); | |
347 | print_generic_expr (file, DR_REF (ref->ref), TDF_SLIM); | |
348 | fprintf (file, " (id %u%s)\n", ref->pos, | |
349 | DR_IS_READ (ref->ref) ? "" : ", write"); | |
350 | ||
351 | fprintf (file, " offset "); | |
352 | dump_double_int (file, ref->offset, false); | |
353 | fprintf (file, "\n"); | |
354 | ||
355 | fprintf (file, " distance %u\n", ref->distance); | |
356 | } | |
357 | else | |
358 | { | |
726a989a | 359 | if (gimple_code (ref->stmt) == GIMPLE_PHI) |
bbc8a8dc ZD |
360 | fprintf (file, " looparound ref\n"); |
361 | else | |
362 | fprintf (file, " combination ref\n"); | |
363 | fprintf (file, " in statement "); | |
726a989a | 364 | print_gimple_stmt (file, ref->stmt, 0, TDF_SLIM); |
bbc8a8dc ZD |
365 | fprintf (file, "\n"); |
366 | fprintf (file, " distance %u\n", ref->distance); | |
367 | } | |
368 | ||
369 | } | |
370 | ||
371 | /* Dumps CHAIN to FILE. */ | |
372 | ||
373 | extern void dump_chain (FILE *, chain_p); | |
374 | void | |
375 | dump_chain (FILE *file, chain_p chain) | |
376 | { | |
377 | dref a; | |
378 | const char *chain_type; | |
379 | unsigned i; | |
380 | tree var; | |
381 | ||
382 | switch (chain->type) | |
383 | { | |
384 | case CT_INVARIANT: | |
385 | chain_type = "Load motion"; | |
386 | break; | |
387 | ||
388 | case CT_LOAD: | |
389 | chain_type = "Loads-only"; | |
390 | break; | |
391 | ||
392 | case CT_STORE_LOAD: | |
393 | chain_type = "Store-loads"; | |
394 | break; | |
395 | ||
396 | case CT_COMBINATION: | |
397 | chain_type = "Combination"; | |
398 | break; | |
399 | ||
400 | default: | |
401 | gcc_unreachable (); | |
402 | } | |
403 | ||
404 | fprintf (file, "%s chain %p%s\n", chain_type, (void *) chain, | |
405 | chain->combined ? " (combined)" : ""); | |
406 | if (chain->type != CT_INVARIANT) | |
407 | fprintf (file, " max distance %u%s\n", chain->length, | |
408 | chain->has_max_use_after ? "" : ", may reuse first"); | |
409 | ||
410 | if (chain->type == CT_COMBINATION) | |
411 | { | |
412 | fprintf (file, " equal to %p %s %p in type ", | |
82d6e6fc | 413 | (void *) chain->ch1, op_symbol_code (chain->op), |
bbc8a8dc ZD |
414 | (void *) chain->ch2); |
415 | print_generic_expr (file, chain->rslt_type, TDF_SLIM); | |
416 | fprintf (file, "\n"); | |
417 | } | |
418 | ||
419 | if (chain->vars) | |
420 | { | |
421 | fprintf (file, " vars"); | |
ac47786e | 422 | FOR_EACH_VEC_ELT (tree, chain->vars, i, var) |
bbc8a8dc ZD |
423 | { |
424 | fprintf (file, " "); | |
425 | print_generic_expr (file, var, TDF_SLIM); | |
426 | } | |
427 | fprintf (file, "\n"); | |
428 | } | |
429 | ||
430 | if (chain->inits) | |
431 | { | |
432 | fprintf (file, " inits"); | |
ac47786e | 433 | FOR_EACH_VEC_ELT (tree, chain->inits, i, var) |
bbc8a8dc ZD |
434 | { |
435 | fprintf (file, " "); | |
436 | print_generic_expr (file, var, TDF_SLIM); | |
437 | } | |
438 | fprintf (file, "\n"); | |
439 | } | |
440 | ||
441 | fprintf (file, " references:\n"); | |
ac47786e | 442 | FOR_EACH_VEC_ELT (dref, chain->refs, i, a) |
bbc8a8dc ZD |
443 | dump_dref (file, a); |
444 | ||
445 | fprintf (file, "\n"); | |
446 | } | |
447 | ||
448 | /* Dumps CHAINS to FILE. */ | |
449 | ||
450 | extern void dump_chains (FILE *, VEC (chain_p, heap) *); | |
451 | void | |
452 | dump_chains (FILE *file, VEC (chain_p, heap) *chains) | |
453 | { | |
454 | chain_p chain; | |
455 | unsigned i; | |
456 | ||
ac47786e | 457 | FOR_EACH_VEC_ELT (chain_p, chains, i, chain) |
bbc8a8dc ZD |
458 | dump_chain (file, chain); |
459 | } | |
460 | ||
461 | /* Dumps COMP to FILE. */ | |
462 | ||
463 | extern void dump_component (FILE *, struct component *); | |
464 | void | |
465 | dump_component (FILE *file, struct component *comp) | |
466 | { | |
467 | dref a; | |
468 | unsigned i; | |
469 | ||
470 | fprintf (file, "Component%s:\n", | |
471 | comp->comp_step == RS_INVARIANT ? " (invariant)" : ""); | |
ac47786e | 472 | FOR_EACH_VEC_ELT (dref, comp->refs, i, a) |
bbc8a8dc ZD |
473 | dump_dref (file, a); |
474 | fprintf (file, "\n"); | |
475 | } | |
476 | ||
477 | /* Dumps COMPS to FILE. */ | |
478 | ||
479 | extern void dump_components (FILE *, struct component *); | |
480 | void | |
481 | dump_components (FILE *file, struct component *comps) | |
482 | { | |
483 | struct component *comp; | |
484 | ||
485 | for (comp = comps; comp; comp = comp->next) | |
486 | dump_component (file, comp); | |
487 | } | |
488 | ||
489 | /* Frees a chain CHAIN. */ | |
490 | ||
491 | static void | |
492 | release_chain (chain_p chain) | |
493 | { | |
494 | dref ref; | |
495 | unsigned i; | |
496 | ||
497 | if (chain == NULL) | |
498 | return; | |
499 | ||
ac47786e | 500 | FOR_EACH_VEC_ELT (dref, chain->refs, i, ref) |
bbc8a8dc ZD |
501 | free (ref); |
502 | ||
503 | VEC_free (dref, heap, chain->refs); | |
504 | VEC_free (tree, heap, chain->vars); | |
505 | VEC_free (tree, heap, chain->inits); | |
506 | ||
507 | free (chain); | |
508 | } | |
509 | ||
510 | /* Frees CHAINS. */ | |
511 | ||
512 | static void | |
513 | release_chains (VEC (chain_p, heap) *chains) | |
514 | { | |
515 | unsigned i; | |
516 | chain_p chain; | |
517 | ||
ac47786e | 518 | FOR_EACH_VEC_ELT (chain_p, chains, i, chain) |
bbc8a8dc ZD |
519 | release_chain (chain); |
520 | VEC_free (chain_p, heap, chains); | |
521 | } | |
522 | ||
523 | /* Frees a component COMP. */ | |
524 | ||
525 | static void | |
526 | release_component (struct component *comp) | |
527 | { | |
528 | VEC_free (dref, heap, comp->refs); | |
529 | free (comp); | |
530 | } | |
531 | ||
532 | /* Frees list of components COMPS. */ | |
533 | ||
534 | static void | |
535 | release_components (struct component *comps) | |
536 | { | |
537 | struct component *act, *next; | |
538 | ||
539 | for (act = comps; act; act = next) | |
540 | { | |
541 | next = act->next; | |
542 | release_component (act); | |
543 | } | |
544 | } | |
545 | ||
546 | /* Finds a root of tree given by FATHERS containing A, and performs path | |
547 | shortening. */ | |
548 | ||
549 | static unsigned | |
550 | component_of (unsigned fathers[], unsigned a) | |
551 | { | |
552 | unsigned root, n; | |
553 | ||
554 | for (root = a; root != fathers[root]; root = fathers[root]) | |
555 | continue; | |
556 | ||
557 | for (; a != root; a = n) | |
558 | { | |
559 | n = fathers[a]; | |
560 | fathers[a] = root; | |
561 | } | |
562 | ||
563 | return root; | |
564 | } | |
565 | ||
566 | /* Join operation for DFU. FATHERS gives the tree, SIZES are sizes of the | |
567 | components, A and B are components to merge. */ | |
568 | ||
569 | static void | |
570 | merge_comps (unsigned fathers[], unsigned sizes[], unsigned a, unsigned b) | |
571 | { | |
572 | unsigned ca = component_of (fathers, a); | |
573 | unsigned cb = component_of (fathers, b); | |
574 | ||
575 | if (ca == cb) | |
576 | return; | |
577 | ||
578 | if (sizes[ca] < sizes[cb]) | |
579 | { | |
580 | sizes[cb] += sizes[ca]; | |
581 | fathers[ca] = cb; | |
582 | } | |
583 | else | |
584 | { | |
585 | sizes[ca] += sizes[cb]; | |
586 | fathers[cb] = ca; | |
587 | } | |
588 | } | |
589 | ||
590 | /* Returns true if A is a reference that is suitable for predictive commoning | |
591 | in the innermost loop that contains it. REF_STEP is set according to the | |
592 | step of the reference A. */ | |
593 | ||
594 | static bool | |
595 | suitable_reference_p (struct data_reference *a, enum ref_step_type *ref_step) | |
596 | { | |
597 | tree ref = DR_REF (a), step = DR_STEP (a); | |
598 | ||
599 | if (!step | |
64fb0d3a | 600 | || TREE_THIS_VOLATILE (ref) |
7e80c6bf EB |
601 | || !is_gimple_reg_type (TREE_TYPE (ref)) |
602 | || tree_could_throw_p (ref)) | |
bbc8a8dc ZD |
603 | return false; |
604 | ||
605 | if (integer_zerop (step)) | |
606 | *ref_step = RS_INVARIANT; | |
607 | else if (integer_nonzerop (step)) | |
608 | *ref_step = RS_NONZERO; | |
609 | else | |
610 | *ref_step = RS_ANY; | |
611 | ||
612 | return true; | |
613 | } | |
614 | ||
615 | /* Stores DR_OFFSET (DR) + DR_INIT (DR) to OFFSET. */ | |
616 | ||
617 | static void | |
618 | aff_combination_dr_offset (struct data_reference *dr, aff_tree *offset) | |
619 | { | |
0d82a1c8 | 620 | tree type = TREE_TYPE (DR_OFFSET (dr)); |
bbc8a8dc ZD |
621 | aff_tree delta; |
622 | ||
0d82a1c8 | 623 | tree_to_aff_combination_expand (DR_OFFSET (dr), type, offset, |
bbc8a8dc | 624 | &name_expansions); |
0d82a1c8 | 625 | aff_combination_const (&delta, type, tree_to_double_int (DR_INIT (dr))); |
bbc8a8dc ZD |
626 | aff_combination_add (offset, &delta); |
627 | } | |
628 | ||
629 | /* Determines number of iterations of the innermost enclosing loop before B | |
630 | refers to exactly the same location as A and stores it to OFF. If A and | |
631 | B do not have the same step, they never meet, or anything else fails, | |
632 | returns false, otherwise returns true. Both A and B are assumed to | |
633 | satisfy suitable_reference_p. */ | |
634 | ||
635 | static bool | |
636 | determine_offset (struct data_reference *a, struct data_reference *b, | |
637 | double_int *off) | |
638 | { | |
639 | aff_tree diff, baseb, step; | |
49379cb1 ZD |
640 | tree typea, typeb; |
641 | ||
642 | /* Check that both the references access the location in the same type. */ | |
643 | typea = TREE_TYPE (DR_REF (a)); | |
644 | typeb = TREE_TYPE (DR_REF (b)); | |
36618b93 | 645 | if (!useless_type_conversion_p (typeb, typea)) |
49379cb1 | 646 | return false; |
bbc8a8dc ZD |
647 | |
648 | /* Check whether the base address and the step of both references is the | |
649 | same. */ | |
650 | if (!operand_equal_p (DR_STEP (a), DR_STEP (b), 0) | |
651 | || !operand_equal_p (DR_BASE_ADDRESS (a), DR_BASE_ADDRESS (b), 0)) | |
652 | return false; | |
653 | ||
654 | if (integer_zerop (DR_STEP (a))) | |
655 | { | |
656 | /* If the references have loop invariant address, check that they access | |
657 | exactly the same location. */ | |
658 | *off = double_int_zero; | |
659 | return (operand_equal_p (DR_OFFSET (a), DR_OFFSET (b), 0) | |
660 | && operand_equal_p (DR_INIT (a), DR_INIT (b), 0)); | |
661 | } | |
662 | ||
663 | /* Compare the offsets of the addresses, and check whether the difference | |
664 | is a multiple of step. */ | |
665 | aff_combination_dr_offset (a, &diff); | |
666 | aff_combination_dr_offset (b, &baseb); | |
667 | aff_combination_scale (&baseb, double_int_minus_one); | |
668 | aff_combination_add (&diff, &baseb); | |
669 | ||
0d82a1c8 | 670 | tree_to_aff_combination_expand (DR_STEP (a), TREE_TYPE (DR_STEP (a)), |
bbc8a8dc ZD |
671 | &step, &name_expansions); |
672 | return aff_combination_constant_multiple_p (&diff, &step, off); | |
673 | } | |
674 | ||
675 | /* Returns the last basic block in LOOP for that we are sure that | |
676 | it is executed whenever the loop is entered. */ | |
677 | ||
678 | static basic_block | |
679 | last_always_executed_block (struct loop *loop) | |
680 | { | |
681 | unsigned i; | |
682 | VEC (edge, heap) *exits = get_loop_exit_edges (loop); | |
683 | edge ex; | |
684 | basic_block last = loop->latch; | |
685 | ||
ac47786e | 686 | FOR_EACH_VEC_ELT (edge, exits, i, ex) |
bbc8a8dc ZD |
687 | last = nearest_common_dominator (CDI_DOMINATORS, last, ex->src); |
688 | VEC_free (edge, heap, exits); | |
689 | ||
690 | return last; | |
691 | } | |
692 | ||
693 | /* Splits dependence graph on DATAREFS described by DEPENDS to components. */ | |
694 | ||
695 | static struct component * | |
696 | split_data_refs_to_components (struct loop *loop, | |
697 | VEC (data_reference_p, heap) *datarefs, | |
698 | VEC (ddr_p, heap) *depends) | |
699 | { | |
700 | unsigned i, n = VEC_length (data_reference_p, datarefs); | |
701 | unsigned ca, ia, ib, bad; | |
702 | unsigned *comp_father = XNEWVEC (unsigned, n + 1); | |
703 | unsigned *comp_size = XNEWVEC (unsigned, n + 1); | |
704 | struct component **comps; | |
705 | struct data_reference *dr, *dra, *drb; | |
706 | struct data_dependence_relation *ddr; | |
707 | struct component *comp_list = NULL, *comp; | |
708 | dref dataref; | |
709 | basic_block last_always_executed = last_always_executed_block (loop); | |
b8698a0f | 710 | |
ac47786e | 711 | FOR_EACH_VEC_ELT (data_reference_p, datarefs, i, dr) |
bbc8a8dc ZD |
712 | { |
713 | if (!DR_REF (dr)) | |
714 | { | |
715 | /* A fake reference for call or asm_expr that may clobber memory; | |
716 | just fail. */ | |
717 | goto end; | |
718 | } | |
5417e022 | 719 | dr->aux = (void *) (size_t) i; |
bbc8a8dc ZD |
720 | comp_father[i] = i; |
721 | comp_size[i] = 1; | |
722 | } | |
723 | ||
724 | /* A component reserved for the "bad" data references. */ | |
725 | comp_father[n] = n; | |
726 | comp_size[n] = 1; | |
727 | ||
ac47786e | 728 | FOR_EACH_VEC_ELT (data_reference_p, datarefs, i, dr) |
bbc8a8dc ZD |
729 | { |
730 | enum ref_step_type dummy; | |
731 | ||
732 | if (!suitable_reference_p (dr, &dummy)) | |
733 | { | |
5417e022 | 734 | ia = (unsigned) (size_t) dr->aux; |
bbc8a8dc ZD |
735 | merge_comps (comp_father, comp_size, n, ia); |
736 | } | |
737 | } | |
738 | ||
ac47786e | 739 | FOR_EACH_VEC_ELT (ddr_p, depends, i, ddr) |
bbc8a8dc ZD |
740 | { |
741 | double_int dummy_off; | |
742 | ||
743 | if (DDR_ARE_DEPENDENT (ddr) == chrec_known) | |
744 | continue; | |
745 | ||
746 | dra = DDR_A (ddr); | |
747 | drb = DDR_B (ddr); | |
5417e022 ZD |
748 | ia = component_of (comp_father, (unsigned) (size_t) dra->aux); |
749 | ib = component_of (comp_father, (unsigned) (size_t) drb->aux); | |
bbc8a8dc ZD |
750 | if (ia == ib) |
751 | continue; | |
752 | ||
753 | bad = component_of (comp_father, n); | |
754 | ||
755 | /* If both A and B are reads, we may ignore unsuitable dependences. */ | |
756 | if (DR_IS_READ (dra) && DR_IS_READ (drb) | |
757 | && (ia == bad || ib == bad | |
758 | || !determine_offset (dra, drb, &dummy_off))) | |
759 | continue; | |
b8698a0f | 760 | |
bbc8a8dc ZD |
761 | merge_comps (comp_father, comp_size, ia, ib); |
762 | } | |
763 | ||
764 | comps = XCNEWVEC (struct component *, n); | |
765 | bad = component_of (comp_father, n); | |
ac47786e | 766 | FOR_EACH_VEC_ELT (data_reference_p, datarefs, i, dr) |
bbc8a8dc | 767 | { |
5417e022 | 768 | ia = (unsigned) (size_t) dr->aux; |
bbc8a8dc ZD |
769 | ca = component_of (comp_father, ia); |
770 | if (ca == bad) | |
771 | continue; | |
772 | ||
773 | comp = comps[ca]; | |
774 | if (!comp) | |
775 | { | |
776 | comp = XCNEW (struct component); | |
777 | comp->refs = VEC_alloc (dref, heap, comp_size[ca]); | |
778 | comps[ca] = comp; | |
779 | } | |
780 | ||
7e5487a2 | 781 | dataref = XCNEW (struct dref_d); |
bbc8a8dc ZD |
782 | dataref->ref = dr; |
783 | dataref->stmt = DR_STMT (dr); | |
784 | dataref->offset = double_int_zero; | |
785 | dataref->distance = 0; | |
786 | ||
787 | dataref->always_accessed | |
788 | = dominated_by_p (CDI_DOMINATORS, last_always_executed, | |
726a989a | 789 | gimple_bb (dataref->stmt)); |
bbc8a8dc ZD |
790 | dataref->pos = VEC_length (dref, comp->refs); |
791 | VEC_quick_push (dref, comp->refs, dataref); | |
792 | } | |
793 | ||
794 | for (i = 0; i < n; i++) | |
795 | { | |
796 | comp = comps[i]; | |
797 | if (comp) | |
798 | { | |
799 | comp->next = comp_list; | |
800 | comp_list = comp; | |
801 | } | |
802 | } | |
803 | free (comps); | |
804 | ||
805 | end: | |
806 | free (comp_father); | |
807 | free (comp_size); | |
808 | return comp_list; | |
809 | } | |
810 | ||
811 | /* Returns true if the component COMP satisfies the conditions | |
c80b4100 | 812 | described in 2) at the beginning of this file. LOOP is the current |
bbc8a8dc | 813 | loop. */ |
b8698a0f | 814 | |
bbc8a8dc ZD |
815 | static bool |
816 | suitable_component_p (struct loop *loop, struct component *comp) | |
817 | { | |
818 | unsigned i; | |
819 | dref a, first; | |
820 | basic_block ba, bp = loop->header; | |
821 | bool ok, has_write = false; | |
822 | ||
ac47786e | 823 | FOR_EACH_VEC_ELT (dref, comp->refs, i, a) |
bbc8a8dc | 824 | { |
726a989a | 825 | ba = gimple_bb (a->stmt); |
bbc8a8dc ZD |
826 | |
827 | if (!just_once_each_iteration_p (loop, ba)) | |
828 | return false; | |
829 | ||
830 | gcc_assert (dominated_by_p (CDI_DOMINATORS, ba, bp)); | |
831 | bp = ba; | |
832 | ||
b0af49c4 | 833 | if (DR_IS_WRITE (a->ref)) |
bbc8a8dc ZD |
834 | has_write = true; |
835 | } | |
836 | ||
837 | first = VEC_index (dref, comp->refs, 0); | |
838 | ok = suitable_reference_p (first->ref, &comp->comp_step); | |
839 | gcc_assert (ok); | |
840 | first->offset = double_int_zero; | |
841 | ||
842 | for (i = 1; VEC_iterate (dref, comp->refs, i, a); i++) | |
843 | { | |
844 | if (!determine_offset (first->ref, a->ref, &a->offset)) | |
845 | return false; | |
846 | ||
847 | #ifdef ENABLE_CHECKING | |
848 | { | |
849 | enum ref_step_type a_step; | |
850 | ok = suitable_reference_p (a->ref, &a_step); | |
851 | gcc_assert (ok && a_step == comp->comp_step); | |
852 | } | |
853 | #endif | |
854 | } | |
855 | ||
856 | /* If there is a write inside the component, we must know whether the | |
857 | step is nonzero or not -- we would not otherwise be able to recognize | |
858 | whether the value accessed by reads comes from the OFFSET-th iteration | |
859 | or the previous one. */ | |
860 | if (has_write && comp->comp_step == RS_ANY) | |
861 | return false; | |
862 | ||
863 | return true; | |
864 | } | |
b8698a0f | 865 | |
bbc8a8dc ZD |
866 | /* Check the conditions on references inside each of components COMPS, |
867 | and remove the unsuitable components from the list. The new list | |
868 | of components is returned. The conditions are described in 2) at | |
c80b4100 | 869 | the beginning of this file. LOOP is the current loop. */ |
bbc8a8dc ZD |
870 | |
871 | static struct component * | |
872 | filter_suitable_components (struct loop *loop, struct component *comps) | |
873 | { | |
874 | struct component **comp, *act; | |
875 | ||
876 | for (comp = &comps; *comp; ) | |
877 | { | |
878 | act = *comp; | |
879 | if (suitable_component_p (loop, act)) | |
880 | comp = &act->next; | |
881 | else | |
882 | { | |
a0044be5 JJ |
883 | dref ref; |
884 | unsigned i; | |
885 | ||
bbc8a8dc | 886 | *comp = act->next; |
ac47786e | 887 | FOR_EACH_VEC_ELT (dref, act->refs, i, ref) |
a0044be5 | 888 | free (ref); |
bbc8a8dc ZD |
889 | release_component (act); |
890 | } | |
891 | } | |
892 | ||
893 | return comps; | |
894 | } | |
895 | ||
896 | /* Compares two drefs A and B by their offset and position. Callback for | |
897 | qsort. */ | |
898 | ||
899 | static int | |
900 | order_drefs (const void *a, const void *b) | |
901 | { | |
3d9a9f94 KG |
902 | const dref *const da = (const dref *) a; |
903 | const dref *const db = (const dref *) b; | |
bbc8a8dc ZD |
904 | int offcmp = double_int_scmp ((*da)->offset, (*db)->offset); |
905 | ||
906 | if (offcmp != 0) | |
907 | return offcmp; | |
908 | ||
909 | return (*da)->pos - (*db)->pos; | |
910 | } | |
911 | ||
912 | /* Returns root of the CHAIN. */ | |
913 | ||
914 | static inline dref | |
915 | get_chain_root (chain_p chain) | |
916 | { | |
917 | return VEC_index (dref, chain->refs, 0); | |
918 | } | |
919 | ||
920 | /* Adds REF to the chain CHAIN. */ | |
921 | ||
922 | static void | |
923 | add_ref_to_chain (chain_p chain, dref ref) | |
924 | { | |
925 | dref root = get_chain_root (chain); | |
926 | double_int dist; | |
927 | ||
928 | gcc_assert (double_int_scmp (root->offset, ref->offset) <= 0); | |
bdc45386 | 929 | dist = double_int_sub (ref->offset, root->offset); |
bbc8a8dc | 930 | if (double_int_ucmp (uhwi_to_double_int (MAX_DISTANCE), dist) <= 0) |
a0044be5 JJ |
931 | { |
932 | free (ref); | |
933 | return; | |
934 | } | |
bbc8a8dc ZD |
935 | gcc_assert (double_int_fits_in_uhwi_p (dist)); |
936 | ||
937 | VEC_safe_push (dref, heap, chain->refs, ref); | |
938 | ||
939 | ref->distance = double_int_to_uhwi (dist); | |
940 | ||
941 | if (ref->distance >= chain->length) | |
942 | { | |
943 | chain->length = ref->distance; | |
944 | chain->has_max_use_after = false; | |
945 | } | |
946 | ||
947 | if (ref->distance == chain->length | |
948 | && ref->pos > root->pos) | |
949 | chain->has_max_use_after = true; | |
950 | ||
951 | chain->all_always_accessed &= ref->always_accessed; | |
952 | } | |
953 | ||
954 | /* Returns the chain for invariant component COMP. */ | |
955 | ||
956 | static chain_p | |
957 | make_invariant_chain (struct component *comp) | |
958 | { | |
959 | chain_p chain = XCNEW (struct chain); | |
960 | unsigned i; | |
961 | dref ref; | |
962 | ||
963 | chain->type = CT_INVARIANT; | |
964 | ||
965 | chain->all_always_accessed = true; | |
966 | ||
ac47786e | 967 | FOR_EACH_VEC_ELT (dref, comp->refs, i, ref) |
bbc8a8dc ZD |
968 | { |
969 | VEC_safe_push (dref, heap, chain->refs, ref); | |
970 | chain->all_always_accessed &= ref->always_accessed; | |
971 | } | |
972 | ||
973 | return chain; | |
974 | } | |
975 | ||
976 | /* Make a new chain rooted at REF. */ | |
977 | ||
978 | static chain_p | |
979 | make_rooted_chain (dref ref) | |
980 | { | |
981 | chain_p chain = XCNEW (struct chain); | |
982 | ||
983 | chain->type = DR_IS_READ (ref->ref) ? CT_LOAD : CT_STORE_LOAD; | |
984 | ||
985 | VEC_safe_push (dref, heap, chain->refs, ref); | |
986 | chain->all_always_accessed = ref->always_accessed; | |
987 | ||
988 | ref->distance = 0; | |
989 | ||
990 | return chain; | |
991 | } | |
992 | ||
993 | /* Returns true if CHAIN is not trivial. */ | |
994 | ||
995 | static bool | |
996 | nontrivial_chain_p (chain_p chain) | |
997 | { | |
998 | return chain != NULL && VEC_length (dref, chain->refs) > 1; | |
999 | } | |
1000 | ||
1001 | /* Returns the ssa name that contains the value of REF, or NULL_TREE if there | |
1002 | is no such name. */ | |
1003 | ||
1004 | static tree | |
1005 | name_for_ref (dref ref) | |
1006 | { | |
1007 | tree name; | |
1008 | ||
726a989a | 1009 | if (is_gimple_assign (ref->stmt)) |
bbc8a8dc ZD |
1010 | { |
1011 | if (!ref->ref || DR_IS_READ (ref->ref)) | |
726a989a | 1012 | name = gimple_assign_lhs (ref->stmt); |
bbc8a8dc | 1013 | else |
726a989a | 1014 | name = gimple_assign_rhs1 (ref->stmt); |
bbc8a8dc ZD |
1015 | } |
1016 | else | |
1017 | name = PHI_RESULT (ref->stmt); | |
1018 | ||
1019 | return (TREE_CODE (name) == SSA_NAME ? name : NULL_TREE); | |
1020 | } | |
1021 | ||
1022 | /* Returns true if REF is a valid initializer for ROOT with given DISTANCE (in | |
1023 | iterations of the innermost enclosing loop). */ | |
1024 | ||
1025 | static bool | |
1026 | valid_initializer_p (struct data_reference *ref, | |
1027 | unsigned distance, struct data_reference *root) | |
1028 | { | |
1029 | aff_tree diff, base, step; | |
1030 | double_int off; | |
1031 | ||
bbc8a8dc ZD |
1032 | /* Both REF and ROOT must be accessing the same object. */ |
1033 | if (!operand_equal_p (DR_BASE_ADDRESS (ref), DR_BASE_ADDRESS (root), 0)) | |
1034 | return false; | |
1035 | ||
1036 | /* The initializer is defined outside of loop, hence its address must be | |
1037 | invariant inside the loop. */ | |
1038 | gcc_assert (integer_zerop (DR_STEP (ref))); | |
1039 | ||
1040 | /* If the address of the reference is invariant, initializer must access | |
1041 | exactly the same location. */ | |
1042 | if (integer_zerop (DR_STEP (root))) | |
1043 | return (operand_equal_p (DR_OFFSET (ref), DR_OFFSET (root), 0) | |
1044 | && operand_equal_p (DR_INIT (ref), DR_INIT (root), 0)); | |
1045 | ||
1046 | /* Verify that this index of REF is equal to the root's index at | |
1047 | -DISTANCE-th iteration. */ | |
1048 | aff_combination_dr_offset (root, &diff); | |
1049 | aff_combination_dr_offset (ref, &base); | |
1050 | aff_combination_scale (&base, double_int_minus_one); | |
1051 | aff_combination_add (&diff, &base); | |
1052 | ||
0d82a1c8 RG |
1053 | tree_to_aff_combination_expand (DR_STEP (root), TREE_TYPE (DR_STEP (root)), |
1054 | &step, &name_expansions); | |
bbc8a8dc ZD |
1055 | if (!aff_combination_constant_multiple_p (&diff, &step, &off)) |
1056 | return false; | |
1057 | ||
1058 | if (!double_int_equal_p (off, uhwi_to_double_int (distance))) | |
1059 | return false; | |
1060 | ||
1061 | return true; | |
1062 | } | |
1063 | ||
1064 | /* Finds looparound phi node of LOOP that copies the value of REF, and if its | |
1065 | initial value is correct (equal to initial value of REF shifted by one | |
1066 | iteration), returns the phi node. Otherwise, NULL_TREE is returned. ROOT | |
1067 | is the root of the current chain. */ | |
1068 | ||
726a989a | 1069 | static gimple |
bbc8a8dc ZD |
1070 | find_looparound_phi (struct loop *loop, dref ref, dref root) |
1071 | { | |
726a989a RB |
1072 | tree name, init, init_ref; |
1073 | gimple phi = NULL, init_stmt; | |
bbc8a8dc ZD |
1074 | edge latch = loop_latch_edge (loop); |
1075 | struct data_reference init_dr; | |
726a989a | 1076 | gimple_stmt_iterator psi; |
bbc8a8dc | 1077 | |
726a989a | 1078 | if (is_gimple_assign (ref->stmt)) |
bbc8a8dc ZD |
1079 | { |
1080 | if (DR_IS_READ (ref->ref)) | |
726a989a | 1081 | name = gimple_assign_lhs (ref->stmt); |
bbc8a8dc | 1082 | else |
726a989a | 1083 | name = gimple_assign_rhs1 (ref->stmt); |
bbc8a8dc ZD |
1084 | } |
1085 | else | |
1086 | name = PHI_RESULT (ref->stmt); | |
1087 | if (!name) | |
726a989a | 1088 | return NULL; |
bbc8a8dc | 1089 | |
726a989a RB |
1090 | for (psi = gsi_start_phis (loop->header); !gsi_end_p (psi); gsi_next (&psi)) |
1091 | { | |
1092 | phi = gsi_stmt (psi); | |
1093 | if (PHI_ARG_DEF_FROM_EDGE (phi, latch) == name) | |
1094 | break; | |
1095 | } | |
bbc8a8dc | 1096 | |
726a989a RB |
1097 | if (gsi_end_p (psi)) |
1098 | return NULL; | |
bbc8a8dc ZD |
1099 | |
1100 | init = PHI_ARG_DEF_FROM_EDGE (phi, loop_preheader_edge (loop)); | |
1101 | if (TREE_CODE (init) != SSA_NAME) | |
726a989a | 1102 | return NULL; |
bbc8a8dc | 1103 | init_stmt = SSA_NAME_DEF_STMT (init); |
726a989a RB |
1104 | if (gimple_code (init_stmt) != GIMPLE_ASSIGN) |
1105 | return NULL; | |
1106 | gcc_assert (gimple_assign_lhs (init_stmt) == init); | |
bbc8a8dc | 1107 | |
726a989a | 1108 | init_ref = gimple_assign_rhs1 (init_stmt); |
bbc8a8dc ZD |
1109 | if (!REFERENCE_CLASS_P (init_ref) |
1110 | && !DECL_P (init_ref)) | |
726a989a | 1111 | return NULL; |
bbc8a8dc ZD |
1112 | |
1113 | /* Analyze the behavior of INIT_REF with respect to LOOP (innermost | |
1114 | loop enclosing PHI). */ | |
1115 | memset (&init_dr, 0, sizeof (struct data_reference)); | |
1116 | DR_REF (&init_dr) = init_ref; | |
1117 | DR_STMT (&init_dr) = phi; | |
4e4452b6 | 1118 | if (!dr_analyze_innermost (&init_dr, loop)) |
3661e899 | 1119 | return NULL; |
bbc8a8dc ZD |
1120 | |
1121 | if (!valid_initializer_p (&init_dr, ref->distance + 1, root->ref)) | |
726a989a | 1122 | return NULL; |
bbc8a8dc ZD |
1123 | |
1124 | return phi; | |
1125 | } | |
1126 | ||
1127 | /* Adds a reference for the looparound copy of REF in PHI to CHAIN. */ | |
1128 | ||
1129 | static void | |
726a989a | 1130 | insert_looparound_copy (chain_p chain, dref ref, gimple phi) |
bbc8a8dc | 1131 | { |
7e5487a2 | 1132 | dref nw = XCNEW (struct dref_d), aref; |
bbc8a8dc ZD |
1133 | unsigned i; |
1134 | ||
1135 | nw->stmt = phi; | |
1136 | nw->distance = ref->distance + 1; | |
1137 | nw->always_accessed = 1; | |
1138 | ||
ac47786e | 1139 | FOR_EACH_VEC_ELT (dref, chain->refs, i, aref) |
bbc8a8dc ZD |
1140 | if (aref->distance >= nw->distance) |
1141 | break; | |
1142 | VEC_safe_insert (dref, heap, chain->refs, i, nw); | |
1143 | ||
1144 | if (nw->distance > chain->length) | |
1145 | { | |
1146 | chain->length = nw->distance; | |
1147 | chain->has_max_use_after = false; | |
1148 | } | |
1149 | } | |
1150 | ||
1151 | /* For references in CHAIN that are copied around the LOOP (created previously | |
1152 | by PRE, or by user), add the results of such copies to the chain. This | |
1153 | enables us to remove the copies by unrolling, and may need less registers | |
1154 | (also, it may allow us to combine chains together). */ | |
1155 | ||
1156 | static void | |
1157 | add_looparound_copies (struct loop *loop, chain_p chain) | |
1158 | { | |
1159 | unsigned i; | |
1160 | dref ref, root = get_chain_root (chain); | |
726a989a | 1161 | gimple phi; |
bbc8a8dc | 1162 | |
ac47786e | 1163 | FOR_EACH_VEC_ELT (dref, chain->refs, i, ref) |
bbc8a8dc ZD |
1164 | { |
1165 | phi = find_looparound_phi (loop, ref, root); | |
1166 | if (!phi) | |
1167 | continue; | |
1168 | ||
1169 | bitmap_set_bit (looparound_phis, SSA_NAME_VERSION (PHI_RESULT (phi))); | |
1170 | insert_looparound_copy (chain, ref, phi); | |
1171 | } | |
1172 | } | |
1173 | ||
1174 | /* Find roots of the values and determine distances in the component COMP. | |
1175 | The references are redistributed into CHAINS. LOOP is the current | |
1176 | loop. */ | |
1177 | ||
1178 | static void | |
1179 | determine_roots_comp (struct loop *loop, | |
1180 | struct component *comp, | |
1181 | VEC (chain_p, heap) **chains) | |
1182 | { | |
1183 | unsigned i; | |
1184 | dref a; | |
1185 | chain_p chain = NULL; | |
b61b1f17 | 1186 | double_int last_ofs = double_int_zero; |
bbc8a8dc ZD |
1187 | |
1188 | /* Invariants are handled specially. */ | |
1189 | if (comp->comp_step == RS_INVARIANT) | |
1190 | { | |
1191 | chain = make_invariant_chain (comp); | |
1192 | VEC_safe_push (chain_p, heap, *chains, chain); | |
1193 | return; | |
1194 | } | |
1195 | ||
5095da95 | 1196 | VEC_qsort (dref, comp->refs, order_drefs); |
bbc8a8dc | 1197 | |
ac47786e | 1198 | FOR_EACH_VEC_ELT (dref, comp->refs, i, a) |
bbc8a8dc | 1199 | { |
b0af49c4 | 1200 | if (!chain || DR_IS_WRITE (a->ref) |
b61b1f17 | 1201 | || double_int_ucmp (uhwi_to_double_int (MAX_DISTANCE), |
bdc45386 | 1202 | double_int_sub (a->offset, last_ofs)) <= 0) |
bbc8a8dc ZD |
1203 | { |
1204 | if (nontrivial_chain_p (chain)) | |
b61b1f17 MM |
1205 | { |
1206 | add_looparound_copies (loop, chain); | |
1207 | VEC_safe_push (chain_p, heap, *chains, chain); | |
1208 | } | |
bbc8a8dc ZD |
1209 | else |
1210 | release_chain (chain); | |
1211 | chain = make_rooted_chain (a); | |
b61b1f17 | 1212 | last_ofs = a->offset; |
bbc8a8dc ZD |
1213 | continue; |
1214 | } | |
1215 | ||
1216 | add_ref_to_chain (chain, a); | |
1217 | } | |
1218 | ||
1219 | if (nontrivial_chain_p (chain)) | |
1220 | { | |
1221 | add_looparound_copies (loop, chain); | |
1222 | VEC_safe_push (chain_p, heap, *chains, chain); | |
1223 | } | |
1224 | else | |
1225 | release_chain (chain); | |
1226 | } | |
1227 | ||
1228 | /* Find roots of the values and determine distances in components COMPS, and | |
1229 | separates the references to CHAINS. LOOP is the current loop. */ | |
1230 | ||
1231 | static void | |
1232 | determine_roots (struct loop *loop, | |
1233 | struct component *comps, VEC (chain_p, heap) **chains) | |
1234 | { | |
1235 | struct component *comp; | |
1236 | ||
1237 | for (comp = comps; comp; comp = comp->next) | |
1238 | determine_roots_comp (loop, comp, chains); | |
1239 | } | |
1240 | ||
1241 | /* Replace the reference in statement STMT with temporary variable | |
82d6e6fc | 1242 | NEW_TREE. If SET is true, NEW_TREE is instead initialized to the value of |
bbc8a8dc ZD |
1243 | the reference in the statement. IN_LHS is true if the reference |
1244 | is in the lhs of STMT, false if it is in rhs. */ | |
1245 | ||
1246 | static void | |
82d6e6fc | 1247 | replace_ref_with (gimple stmt, tree new_tree, bool set, bool in_lhs) |
bbc8a8dc | 1248 | { |
726a989a RB |
1249 | tree val; |
1250 | gimple new_stmt; | |
1251 | gimple_stmt_iterator bsi, psi; | |
bbc8a8dc | 1252 | |
726a989a | 1253 | if (gimple_code (stmt) == GIMPLE_PHI) |
bbc8a8dc ZD |
1254 | { |
1255 | gcc_assert (!in_lhs && !set); | |
1256 | ||
1257 | val = PHI_RESULT (stmt); | |
726a989a RB |
1258 | bsi = gsi_after_labels (gimple_bb (stmt)); |
1259 | psi = gsi_for_stmt (stmt); | |
1260 | remove_phi_node (&psi, false); | |
bbc8a8dc | 1261 | |
726a989a | 1262 | /* Turn the phi node into GIMPLE_ASSIGN. */ |
82d6e6fc | 1263 | new_stmt = gimple_build_assign (val, new_tree); |
726a989a | 1264 | gsi_insert_before (&bsi, new_stmt, GSI_NEW_STMT); |
bbc8a8dc ZD |
1265 | return; |
1266 | } | |
b8698a0f | 1267 | |
bbc8a8dc ZD |
1268 | /* Since the reference is of gimple_reg type, it should only |
1269 | appear as lhs or rhs of modify statement. */ | |
726a989a RB |
1270 | gcc_assert (is_gimple_assign (stmt)); |
1271 | ||
1272 | bsi = gsi_for_stmt (stmt); | |
bbc8a8dc | 1273 | |
82d6e6fc | 1274 | /* If we do not need to initialize NEW_TREE, just replace the use of OLD. */ |
bbc8a8dc ZD |
1275 | if (!set) |
1276 | { | |
1277 | gcc_assert (!in_lhs); | |
82d6e6fc | 1278 | gimple_assign_set_rhs_from_tree (&bsi, new_tree); |
726a989a | 1279 | stmt = gsi_stmt (bsi); |
bbc8a8dc ZD |
1280 | update_stmt (stmt); |
1281 | return; | |
1282 | } | |
1283 | ||
bbc8a8dc ZD |
1284 | if (in_lhs) |
1285 | { | |
726a989a | 1286 | /* We have statement |
b8698a0f | 1287 | |
726a989a | 1288 | OLD = VAL |
bbc8a8dc | 1289 | |
726a989a RB |
1290 | If OLD is a memory reference, then VAL is gimple_val, and we transform |
1291 | this to | |
bbc8a8dc ZD |
1292 | |
1293 | OLD = VAL | |
1294 | NEW = VAL | |
1295 | ||
b8698a0f | 1296 | Otherwise, we are replacing a combination chain, |
726a989a RB |
1297 | VAL is the expression that performs the combination, and OLD is an |
1298 | SSA name. In this case, we transform the assignment to | |
1299 | ||
1300 | OLD = VAL | |
1301 | NEW = OLD | |
1302 | ||
1303 | */ | |
1304 | ||
1305 | val = gimple_assign_lhs (stmt); | |
1306 | if (TREE_CODE (val) != SSA_NAME) | |
1307 | { | |
726a989a | 1308 | val = gimple_assign_rhs1 (stmt); |
7e8b1c4b JJ |
1309 | gcc_assert (gimple_assign_single_p (stmt)); |
1310 | if (TREE_CLOBBER_P (val)) | |
32244553 | 1311 | val = get_or_create_ssa_default_def (cfun, SSA_NAME_VAR (new_tree)); |
7e8b1c4b JJ |
1312 | else |
1313 | gcc_assert (gimple_assign_copy_p (stmt)); | |
726a989a | 1314 | } |
bbc8a8dc ZD |
1315 | } |
1316 | else | |
1317 | { | |
bbc8a8dc ZD |
1318 | /* VAL = OLD |
1319 | ||
1320 | is transformed to | |
1321 | ||
1322 | VAL = OLD | |
1323 | NEW = VAL */ | |
726a989a RB |
1324 | |
1325 | val = gimple_assign_lhs (stmt); | |
bbc8a8dc ZD |
1326 | } |
1327 | ||
82d6e6fc | 1328 | new_stmt = gimple_build_assign (new_tree, unshare_expr (val)); |
726a989a | 1329 | gsi_insert_after (&bsi, new_stmt, GSI_NEW_STMT); |
bbc8a8dc ZD |
1330 | } |
1331 | ||
1332 | /* Returns the reference to the address of REF in the ITER-th iteration of | |
1333 | LOOP, or NULL if we fail to determine it (ITER may be negative). We | |
1334 | try to preserve the original shape of the reference (not rewrite it | |
1335 | as an indirect ref to the address), to make tree_could_trap_p in | |
1336 | prepare_initializers_chain return false more often. */ | |
1337 | ||
1338 | static tree | |
1339 | ref_at_iteration (struct loop *loop, tree ref, int iter) | |
1340 | { | |
1341 | tree idx, *idx_p, type, val, op0 = NULL_TREE, ret; | |
1342 | affine_iv iv; | |
1343 | bool ok; | |
1344 | ||
1345 | if (handled_component_p (ref)) | |
1346 | { | |
1347 | op0 = ref_at_iteration (loop, TREE_OPERAND (ref, 0), iter); | |
1348 | if (!op0) | |
1349 | return NULL_TREE; | |
1350 | } | |
70f34814 RG |
1351 | else if (!INDIRECT_REF_P (ref) |
1352 | && TREE_CODE (ref) != MEM_REF) | |
bbc8a8dc ZD |
1353 | return unshare_expr (ref); |
1354 | ||
be1ac4ec | 1355 | if (TREE_CODE (ref) == MEM_REF) |
bbc8a8dc | 1356 | { |
70f34814 | 1357 | ret = unshare_expr (ref); |
bbc8a8dc ZD |
1358 | idx = TREE_OPERAND (ref, 0); |
1359 | idx_p = &TREE_OPERAND (ret, 0); | |
1360 | } | |
1361 | else if (TREE_CODE (ref) == COMPONENT_REF) | |
1362 | { | |
1363 | /* Check that the offset is loop invariant. */ | |
1364 | if (TREE_OPERAND (ref, 2) | |
1365 | && !expr_invariant_in_loop_p (loop, TREE_OPERAND (ref, 2))) | |
1366 | return NULL_TREE; | |
1367 | ||
1368 | return build3 (COMPONENT_REF, TREE_TYPE (ref), op0, | |
1369 | unshare_expr (TREE_OPERAND (ref, 1)), | |
1370 | unshare_expr (TREE_OPERAND (ref, 2))); | |
1371 | } | |
1372 | else if (TREE_CODE (ref) == ARRAY_REF) | |
1373 | { | |
1374 | /* Check that the lower bound and the step are loop invariant. */ | |
1375 | if (TREE_OPERAND (ref, 2) | |
1376 | && !expr_invariant_in_loop_p (loop, TREE_OPERAND (ref, 2))) | |
1377 | return NULL_TREE; | |
1378 | if (TREE_OPERAND (ref, 3) | |
1379 | && !expr_invariant_in_loop_p (loop, TREE_OPERAND (ref, 3))) | |
1380 | return NULL_TREE; | |
1381 | ||
1382 | ret = build4 (ARRAY_REF, TREE_TYPE (ref), op0, NULL_TREE, | |
1383 | unshare_expr (TREE_OPERAND (ref, 2)), | |
1384 | unshare_expr (TREE_OPERAND (ref, 3))); | |
1385 | idx = TREE_OPERAND (ref, 1); | |
1386 | idx_p = &TREE_OPERAND (ret, 1); | |
1387 | } | |
1388 | else | |
1389 | return NULL_TREE; | |
1390 | ||
f017bf5e | 1391 | ok = simple_iv (loop, loop, idx, &iv, true); |
bbc8a8dc ZD |
1392 | if (!ok) |
1393 | return NULL_TREE; | |
1394 | iv.base = expand_simple_operations (iv.base); | |
1395 | if (integer_zerop (iv.step)) | |
1396 | *idx_p = unshare_expr (iv.base); | |
1397 | else | |
1398 | { | |
1399 | type = TREE_TYPE (iv.base); | |
5be014d5 AP |
1400 | if (POINTER_TYPE_P (type)) |
1401 | { | |
1402 | val = fold_build2 (MULT_EXPR, sizetype, iv.step, | |
1403 | size_int (iter)); | |
5d49b6a7 | 1404 | val = fold_build_pointer_plus (iv.base, val); |
5be014d5 AP |
1405 | } |
1406 | else | |
1407 | { | |
1408 | val = fold_build2 (MULT_EXPR, type, iv.step, | |
1409 | build_int_cst_type (type, iter)); | |
1410 | val = fold_build2 (PLUS_EXPR, type, iv.base, val); | |
1411 | } | |
bbc8a8dc ZD |
1412 | *idx_p = unshare_expr (val); |
1413 | } | |
1414 | ||
1415 | return ret; | |
1416 | } | |
1417 | ||
1418 | /* Get the initialization expression for the INDEX-th temporary variable | |
1419 | of CHAIN. */ | |
1420 | ||
1421 | static tree | |
1422 | get_init_expr (chain_p chain, unsigned index) | |
1423 | { | |
1424 | if (chain->type == CT_COMBINATION) | |
1425 | { | |
1426 | tree e1 = get_init_expr (chain->ch1, index); | |
1427 | tree e2 = get_init_expr (chain->ch2, index); | |
1428 | ||
82d6e6fc | 1429 | return fold_build2 (chain->op, chain->rslt_type, e1, e2); |
bbc8a8dc ZD |
1430 | } |
1431 | else | |
1432 | return VEC_index (tree, chain->inits, index); | |
1433 | } | |
1434 | ||
2664efb6 ZD |
1435 | /* Returns a new temporary variable used for the I-th variable carrying |
1436 | value of REF. The variable's uid is marked in TMP_VARS. */ | |
1437 | ||
1438 | static tree | |
1439 | predcom_tmp_var (tree ref, unsigned i, bitmap tmp_vars) | |
1440 | { | |
1441 | tree type = TREE_TYPE (ref); | |
2664efb6 ZD |
1442 | /* We never access the components of the temporary variable in predictive |
1443 | commoning. */ | |
acd63801 | 1444 | tree var = create_tmp_reg (type, get_lsm_tmp_name (ref, i)); |
2664efb6 ZD |
1445 | bitmap_set_bit (tmp_vars, DECL_UID (var)); |
1446 | return var; | |
1447 | } | |
1448 | ||
bbc8a8dc ZD |
1449 | /* Creates the variables for CHAIN, as well as phi nodes for them and |
1450 | initialization on entry to LOOP. Uids of the newly created | |
1451 | temporary variables are marked in TMP_VARS. */ | |
1452 | ||
1453 | static void | |
1454 | initialize_root_vars (struct loop *loop, chain_p chain, bitmap tmp_vars) | |
1455 | { | |
1456 | unsigned i; | |
1457 | unsigned n = chain->length; | |
1458 | dref root = get_chain_root (chain); | |
1459 | bool reuse_first = !chain->has_max_use_after; | |
726a989a RB |
1460 | tree ref, init, var, next; |
1461 | gimple phi; | |
1462 | gimple_seq stmts; | |
bbc8a8dc ZD |
1463 | edge entry = loop_preheader_edge (loop), latch = loop_latch_edge (loop); |
1464 | ||
1465 | /* If N == 0, then all the references are within the single iteration. And | |
1466 | since this is an nonempty chain, reuse_first cannot be true. */ | |
1467 | gcc_assert (n > 0 || !reuse_first); | |
1468 | ||
1469 | chain->vars = VEC_alloc (tree, heap, n + 1); | |
1470 | ||
1471 | if (chain->type == CT_COMBINATION) | |
726a989a | 1472 | ref = gimple_assign_lhs (root->stmt); |
bbc8a8dc ZD |
1473 | else |
1474 | ref = DR_REF (root->ref); | |
1475 | ||
1476 | for (i = 0; i < n + (reuse_first ? 0 : 1); i++) | |
1477 | { | |
2664efb6 | 1478 | var = predcom_tmp_var (ref, i, tmp_vars); |
bbc8a8dc ZD |
1479 | VEC_quick_push (tree, chain->vars, var); |
1480 | } | |
1481 | if (reuse_first) | |
1482 | VEC_quick_push (tree, chain->vars, VEC_index (tree, chain->vars, 0)); | |
b8698a0f | 1483 | |
ac47786e | 1484 | FOR_EACH_VEC_ELT (tree, chain->vars, i, var) |
726a989a | 1485 | VEC_replace (tree, chain->vars, i, make_ssa_name (var, NULL)); |
bbc8a8dc ZD |
1486 | |
1487 | for (i = 0; i < n; i++) | |
1488 | { | |
1489 | var = VEC_index (tree, chain->vars, i); | |
1490 | next = VEC_index (tree, chain->vars, i + 1); | |
1491 | init = get_init_expr (chain, i); | |
1492 | ||
1493 | init = force_gimple_operand (init, &stmts, true, NULL_TREE); | |
1494 | if (stmts) | |
5006671f | 1495 | gsi_insert_seq_on_edge_immediate (entry, stmts); |
bbc8a8dc ZD |
1496 | |
1497 | phi = create_phi_node (var, loop->header); | |
9e227d60 DC |
1498 | add_phi_arg (phi, init, entry, UNKNOWN_LOCATION); |
1499 | add_phi_arg (phi, next, latch, UNKNOWN_LOCATION); | |
bbc8a8dc ZD |
1500 | } |
1501 | } | |
1502 | ||
1503 | /* Create the variables and initialization statement for root of chain | |
1504 | CHAIN. Uids of the newly created temporary variables are marked | |
1505 | in TMP_VARS. */ | |
1506 | ||
1507 | static void | |
1508 | initialize_root (struct loop *loop, chain_p chain, bitmap tmp_vars) | |
1509 | { | |
1510 | dref root = get_chain_root (chain); | |
1511 | bool in_lhs = (chain->type == CT_STORE_LOAD | |
1512 | || chain->type == CT_COMBINATION); | |
1513 | ||
1514 | initialize_root_vars (loop, chain, tmp_vars); | |
1515 | replace_ref_with (root->stmt, | |
1516 | VEC_index (tree, chain->vars, chain->length), | |
1517 | true, in_lhs); | |
1518 | } | |
1519 | ||
1520 | /* Initializes a variable for load motion for ROOT and prepares phi nodes and | |
1521 | initialization on entry to LOOP if necessary. The ssa name for the variable | |
1522 | is stored in VARS. If WRITTEN is true, also a phi node to copy its value | |
1523 | around the loop is created. Uid of the newly created temporary variable | |
1524 | is marked in TMP_VARS. INITS is the list containing the (single) | |
1525 | initializer. */ | |
1526 | ||
1527 | static void | |
1528 | initialize_root_vars_lm (struct loop *loop, dref root, bool written, | |
1529 | VEC(tree, heap) **vars, VEC(tree, heap) *inits, | |
1530 | bitmap tmp_vars) | |
1531 | { | |
1532 | unsigned i; | |
726a989a RB |
1533 | tree ref = DR_REF (root->ref), init, var, next; |
1534 | gimple_seq stmts; | |
1535 | gimple phi; | |
bbc8a8dc ZD |
1536 | edge entry = loop_preheader_edge (loop), latch = loop_latch_edge (loop); |
1537 | ||
1538 | /* Find the initializer for the variable, and check that it cannot | |
1539 | trap. */ | |
1540 | init = VEC_index (tree, inits, 0); | |
1541 | ||
1542 | *vars = VEC_alloc (tree, heap, written ? 2 : 1); | |
2664efb6 | 1543 | var = predcom_tmp_var (ref, 0, tmp_vars); |
bbc8a8dc ZD |
1544 | VEC_quick_push (tree, *vars, var); |
1545 | if (written) | |
1546 | VEC_quick_push (tree, *vars, VEC_index (tree, *vars, 0)); | |
b8698a0f | 1547 | |
ac47786e | 1548 | FOR_EACH_VEC_ELT (tree, *vars, i, var) |
726a989a | 1549 | VEC_replace (tree, *vars, i, make_ssa_name (var, NULL)); |
bbc8a8dc ZD |
1550 | |
1551 | var = VEC_index (tree, *vars, 0); | |
b8698a0f | 1552 | |
bbc8a8dc ZD |
1553 | init = force_gimple_operand (init, &stmts, written, NULL_TREE); |
1554 | if (stmts) | |
5006671f | 1555 | gsi_insert_seq_on_edge_immediate (entry, stmts); |
bbc8a8dc ZD |
1556 | |
1557 | if (written) | |
1558 | { | |
1559 | next = VEC_index (tree, *vars, 1); | |
1560 | phi = create_phi_node (var, loop->header); | |
9e227d60 DC |
1561 | add_phi_arg (phi, init, entry, UNKNOWN_LOCATION); |
1562 | add_phi_arg (phi, next, latch, UNKNOWN_LOCATION); | |
bbc8a8dc ZD |
1563 | } |
1564 | else | |
1565 | { | |
726a989a | 1566 | gimple init_stmt = gimple_build_assign (var, init); |
726a989a | 1567 | gsi_insert_on_edge_immediate (entry, init_stmt); |
bbc8a8dc ZD |
1568 | } |
1569 | } | |
1570 | ||
1571 | ||
1572 | /* Execute load motion for references in chain CHAIN. Uids of the newly | |
1573 | created temporary variables are marked in TMP_VARS. */ | |
1574 | ||
1575 | static void | |
1576 | execute_load_motion (struct loop *loop, chain_p chain, bitmap tmp_vars) | |
1577 | { | |
1578 | VEC (tree, heap) *vars; | |
1579 | dref a; | |
1580 | unsigned n_writes = 0, ridx, i; | |
1581 | tree var; | |
1582 | ||
1583 | gcc_assert (chain->type == CT_INVARIANT); | |
1584 | gcc_assert (!chain->combined); | |
ac47786e | 1585 | FOR_EACH_VEC_ELT (dref, chain->refs, i, a) |
b0af49c4 | 1586 | if (DR_IS_WRITE (a->ref)) |
bbc8a8dc | 1587 | n_writes++; |
b8698a0f | 1588 | |
bbc8a8dc ZD |
1589 | /* If there are no reads in the loop, there is nothing to do. */ |
1590 | if (n_writes == VEC_length (dref, chain->refs)) | |
1591 | return; | |
1592 | ||
1593 | initialize_root_vars_lm (loop, get_chain_root (chain), n_writes > 0, | |
1594 | &vars, chain->inits, tmp_vars); | |
1595 | ||
1596 | ridx = 0; | |
ac47786e | 1597 | FOR_EACH_VEC_ELT (dref, chain->refs, i, a) |
bbc8a8dc ZD |
1598 | { |
1599 | bool is_read = DR_IS_READ (a->ref); | |
bbc8a8dc | 1600 | |
b0af49c4 | 1601 | if (DR_IS_WRITE (a->ref)) |
bbc8a8dc ZD |
1602 | { |
1603 | n_writes--; | |
1604 | if (n_writes) | |
1605 | { | |
1606 | var = VEC_index (tree, vars, 0); | |
726a989a | 1607 | var = make_ssa_name (SSA_NAME_VAR (var), NULL); |
bbc8a8dc ZD |
1608 | VEC_replace (tree, vars, 0, var); |
1609 | } | |
1610 | else | |
1611 | ridx = 1; | |
1612 | } | |
b8698a0f | 1613 | |
bbc8a8dc ZD |
1614 | replace_ref_with (a->stmt, VEC_index (tree, vars, ridx), |
1615 | !is_read, !is_read); | |
1616 | } | |
1617 | ||
1618 | VEC_free (tree, heap, vars); | |
1619 | } | |
1620 | ||
1621 | /* Returns the single statement in that NAME is used, excepting | |
1622 | the looparound phi nodes contained in one of the chains. If there is no | |
726a989a | 1623 | such statement, or more statements, NULL is returned. */ |
bbc8a8dc | 1624 | |
726a989a | 1625 | static gimple |
bbc8a8dc ZD |
1626 | single_nonlooparound_use (tree name) |
1627 | { | |
1628 | use_operand_p use; | |
1629 | imm_use_iterator it; | |
726a989a | 1630 | gimple stmt, ret = NULL; |
bbc8a8dc ZD |
1631 | |
1632 | FOR_EACH_IMM_USE_FAST (use, it, name) | |
1633 | { | |
1634 | stmt = USE_STMT (use); | |
1635 | ||
726a989a | 1636 | if (gimple_code (stmt) == GIMPLE_PHI) |
bbc8a8dc ZD |
1637 | { |
1638 | /* Ignore uses in looparound phi nodes. Uses in other phi nodes | |
1639 | could not be processed anyway, so just fail for them. */ | |
1640 | if (bitmap_bit_p (looparound_phis, | |
1641 | SSA_NAME_VERSION (PHI_RESULT (stmt)))) | |
1642 | continue; | |
1643 | ||
726a989a | 1644 | return NULL; |
bbc8a8dc | 1645 | } |
b63f974e JJ |
1646 | else if (is_gimple_debug (stmt)) |
1647 | continue; | |
726a989a RB |
1648 | else if (ret != NULL) |
1649 | return NULL; | |
bbc8a8dc ZD |
1650 | else |
1651 | ret = stmt; | |
1652 | } | |
1653 | ||
1654 | return ret; | |
1655 | } | |
1656 | ||
1657 | /* Remove statement STMT, as well as the chain of assignments in that it is | |
1658 | used. */ | |
1659 | ||
1660 | static void | |
726a989a | 1661 | remove_stmt (gimple stmt) |
bbc8a8dc | 1662 | { |
726a989a RB |
1663 | tree name; |
1664 | gimple next; | |
1665 | gimple_stmt_iterator psi; | |
bbc8a8dc | 1666 | |
726a989a | 1667 | if (gimple_code (stmt) == GIMPLE_PHI) |
bbc8a8dc ZD |
1668 | { |
1669 | name = PHI_RESULT (stmt); | |
1670 | next = single_nonlooparound_use (name); | |
273ccb6d | 1671 | reset_debug_uses (stmt); |
726a989a RB |
1672 | psi = gsi_for_stmt (stmt); |
1673 | remove_phi_node (&psi, true); | |
bbc8a8dc ZD |
1674 | |
1675 | if (!next | |
5f8ecf45 | 1676 | || !gimple_assign_ssa_name_copy_p (next) |
726a989a | 1677 | || gimple_assign_rhs1 (next) != name) |
bbc8a8dc ZD |
1678 | return; |
1679 | ||
1680 | stmt = next; | |
1681 | } | |
1682 | ||
1683 | while (1) | |
1684 | { | |
726a989a | 1685 | gimple_stmt_iterator bsi; |
b8698a0f | 1686 | |
726a989a | 1687 | bsi = gsi_for_stmt (stmt); |
bbc8a8dc | 1688 | |
726a989a | 1689 | name = gimple_assign_lhs (stmt); |
bbc8a8dc ZD |
1690 | gcc_assert (TREE_CODE (name) == SSA_NAME); |
1691 | ||
1692 | next = single_nonlooparound_use (name); | |
273ccb6d | 1693 | reset_debug_uses (stmt); |
bbc8a8dc | 1694 | |
13714310 | 1695 | unlink_stmt_vdef (stmt); |
726a989a | 1696 | gsi_remove (&bsi, true); |
5f8ecf45 | 1697 | release_defs (stmt); |
bbc8a8dc ZD |
1698 | |
1699 | if (!next | |
5f8ecf45 | 1700 | || !gimple_assign_ssa_name_copy_p (next) |
726a989a | 1701 | || gimple_assign_rhs1 (next) != name) |
bbc8a8dc ZD |
1702 | return; |
1703 | ||
1704 | stmt = next; | |
1705 | } | |
1706 | } | |
1707 | ||
1708 | /* Perform the predictive commoning optimization for a chain CHAIN. | |
1709 | Uids of the newly created temporary variables are marked in TMP_VARS.*/ | |
1710 | ||
1711 | static void | |
1712 | execute_pred_commoning_chain (struct loop *loop, chain_p chain, | |
1713 | bitmap tmp_vars) | |
1714 | { | |
1715 | unsigned i; | |
13714310 | 1716 | dref a; |
bbc8a8dc ZD |
1717 | tree var; |
1718 | ||
1719 | if (chain->combined) | |
1720 | { | |
1721 | /* For combined chains, just remove the statements that are used to | |
1722 | compute the values of the expression (except for the root one). */ | |
1723 | for (i = 1; VEC_iterate (dref, chain->refs, i, a); i++) | |
1724 | remove_stmt (a->stmt); | |
1725 | } | |
1726 | else | |
1727 | { | |
1728 | /* For non-combined chains, set up the variables that hold its value, | |
1729 | and replace the uses of the original references by these | |
1730 | variables. */ | |
bbc8a8dc ZD |
1731 | initialize_root (loop, chain, tmp_vars); |
1732 | for (i = 1; VEC_iterate (dref, chain->refs, i, a); i++) | |
1733 | { | |
bbc8a8dc ZD |
1734 | var = VEC_index (tree, chain->vars, chain->length - a->distance); |
1735 | replace_ref_with (a->stmt, var, false, false); | |
1736 | } | |
1737 | } | |
1738 | } | |
1739 | ||
1740 | /* Determines the unroll factor necessary to remove as many temporary variable | |
1741 | copies as possible. CHAINS is the list of chains that will be | |
1742 | optimized. */ | |
1743 | ||
1744 | static unsigned | |
1745 | determine_unroll_factor (VEC (chain_p, heap) *chains) | |
1746 | { | |
1747 | chain_p chain; | |
1748 | unsigned factor = 1, af, nfactor, i; | |
1749 | unsigned max = PARAM_VALUE (PARAM_MAX_UNROLL_TIMES); | |
1750 | ||
ac47786e | 1751 | FOR_EACH_VEC_ELT (chain_p, chains, i, chain) |
bbc8a8dc ZD |
1752 | { |
1753 | if (chain->type == CT_INVARIANT || chain->combined) | |
1754 | continue; | |
1755 | ||
1756 | /* The best unroll factor for this chain is equal to the number of | |
1757 | temporary variables that we create for it. */ | |
1758 | af = chain->length; | |
1759 | if (chain->has_max_use_after) | |
1760 | af++; | |
1761 | ||
1762 | nfactor = factor * af / gcd (factor, af); | |
1763 | if (nfactor <= max) | |
1764 | factor = nfactor; | |
1765 | } | |
1766 | ||
1767 | return factor; | |
1768 | } | |
1769 | ||
1770 | /* Perform the predictive commoning optimization for CHAINS. | |
1771 | Uids of the newly created temporary variables are marked in TMP_VARS. */ | |
1772 | ||
1773 | static void | |
1774 | execute_pred_commoning (struct loop *loop, VEC (chain_p, heap) *chains, | |
1775 | bitmap tmp_vars) | |
1776 | { | |
1777 | chain_p chain; | |
1778 | unsigned i; | |
1779 | ||
ac47786e | 1780 | FOR_EACH_VEC_ELT (chain_p, chains, i, chain) |
bbc8a8dc ZD |
1781 | { |
1782 | if (chain->type == CT_INVARIANT) | |
1783 | execute_load_motion (loop, chain, tmp_vars); | |
1784 | else | |
1785 | execute_pred_commoning_chain (loop, chain, tmp_vars); | |
1786 | } | |
b8698a0f | 1787 | |
bbc8a8dc ZD |
1788 | update_ssa (TODO_update_ssa_only_virtuals); |
1789 | } | |
1790 | ||
c80b4100 | 1791 | /* For each reference in CHAINS, if its defining statement is |
726a989a | 1792 | phi node, record the ssa name that is defined by it. */ |
bbc8a8dc ZD |
1793 | |
1794 | static void | |
1795 | replace_phis_by_defined_names (VEC (chain_p, heap) *chains) | |
1796 | { | |
1797 | chain_p chain; | |
1798 | dref a; | |
1799 | unsigned i, j; | |
1800 | ||
ac47786e NF |
1801 | FOR_EACH_VEC_ELT (chain_p, chains, i, chain) |
1802 | FOR_EACH_VEC_ELT (dref, chain->refs, j, a) | |
bbc8a8dc | 1803 | { |
726a989a RB |
1804 | if (gimple_code (a->stmt) == GIMPLE_PHI) |
1805 | { | |
1806 | a->name_defined_by_phi = PHI_RESULT (a->stmt); | |
1807 | a->stmt = NULL; | |
1808 | } | |
bbc8a8dc ZD |
1809 | } |
1810 | } | |
1811 | ||
726a989a RB |
1812 | /* For each reference in CHAINS, if name_defined_by_phi is not |
1813 | NULL, use it to set the stmt field. */ | |
bbc8a8dc ZD |
1814 | |
1815 | static void | |
1816 | replace_names_by_phis (VEC (chain_p, heap) *chains) | |
1817 | { | |
1818 | chain_p chain; | |
1819 | dref a; | |
1820 | unsigned i, j; | |
1821 | ||
ac47786e NF |
1822 | FOR_EACH_VEC_ELT (chain_p, chains, i, chain) |
1823 | FOR_EACH_VEC_ELT (dref, chain->refs, j, a) | |
726a989a | 1824 | if (a->stmt == NULL) |
bbc8a8dc | 1825 | { |
726a989a RB |
1826 | a->stmt = SSA_NAME_DEF_STMT (a->name_defined_by_phi); |
1827 | gcc_assert (gimple_code (a->stmt) == GIMPLE_PHI); | |
1828 | a->name_defined_by_phi = NULL_TREE; | |
bbc8a8dc ZD |
1829 | } |
1830 | } | |
1831 | ||
1832 | /* Wrapper over execute_pred_commoning, to pass it as a callback | |
1833 | to tree_transform_and_unroll_loop. */ | |
1834 | ||
1835 | struct epcc_data | |
1836 | { | |
1837 | VEC (chain_p, heap) *chains; | |
1838 | bitmap tmp_vars; | |
1839 | }; | |
1840 | ||
1841 | static void | |
1842 | execute_pred_commoning_cbck (struct loop *loop, void *data) | |
1843 | { | |
3d9a9f94 | 1844 | struct epcc_data *const dta = (struct epcc_data *) data; |
bbc8a8dc ZD |
1845 | |
1846 | /* Restore phi nodes that were replaced by ssa names before | |
1847 | tree_transform_and_unroll_loop (see detailed description in | |
1848 | tree_predictive_commoning_loop). */ | |
1849 | replace_names_by_phis (dta->chains); | |
1850 | execute_pred_commoning (loop, dta->chains, dta->tmp_vars); | |
1851 | } | |
1852 | ||
bbc8a8dc ZD |
1853 | /* Base NAME and all the names in the chain of phi nodes that use it |
1854 | on variable VAR. The phi nodes are recognized by being in the copies of | |
1855 | the header of the LOOP. */ | |
1856 | ||
1857 | static void | |
1858 | base_names_in_chain_on (struct loop *loop, tree name, tree var) | |
1859 | { | |
726a989a | 1860 | gimple stmt, phi; |
bbc8a8dc | 1861 | imm_use_iterator iter; |
bbc8a8dc | 1862 | |
b2ec94d4 | 1863 | replace_ssa_name_symbol (name, var); |
bbc8a8dc ZD |
1864 | |
1865 | while (1) | |
1866 | { | |
1867 | phi = NULL; | |
1868 | FOR_EACH_IMM_USE_STMT (stmt, iter, name) | |
1869 | { | |
726a989a RB |
1870 | if (gimple_code (stmt) == GIMPLE_PHI |
1871 | && flow_bb_inside_loop_p (loop, gimple_bb (stmt))) | |
bbc8a8dc ZD |
1872 | { |
1873 | phi = stmt; | |
1874 | BREAK_FROM_IMM_USE_STMT (iter); | |
1875 | } | |
1876 | } | |
1877 | if (!phi) | |
1878 | return; | |
1879 | ||
bbc8a8dc | 1880 | name = PHI_RESULT (phi); |
b2ec94d4 | 1881 | replace_ssa_name_symbol (name, var); |
bbc8a8dc ZD |
1882 | } |
1883 | } | |
1884 | ||
1885 | /* Given an unrolled LOOP after predictive commoning, remove the | |
1886 | register copies arising from phi nodes by changing the base | |
1887 | variables of SSA names. TMP_VARS is the set of the temporary variables | |
1888 | for those we want to perform this. */ | |
1889 | ||
1890 | static void | |
1891 | eliminate_temp_copies (struct loop *loop, bitmap tmp_vars) | |
1892 | { | |
1893 | edge e; | |
726a989a RB |
1894 | gimple phi, stmt; |
1895 | tree name, use, var; | |
1896 | gimple_stmt_iterator psi; | |
bbc8a8dc ZD |
1897 | |
1898 | e = loop_latch_edge (loop); | |
726a989a | 1899 | for (psi = gsi_start_phis (loop->header); !gsi_end_p (psi); gsi_next (&psi)) |
bbc8a8dc | 1900 | { |
726a989a | 1901 | phi = gsi_stmt (psi); |
bbc8a8dc ZD |
1902 | name = PHI_RESULT (phi); |
1903 | var = SSA_NAME_VAR (name); | |
70b5e7dc | 1904 | if (!var || !bitmap_bit_p (tmp_vars, DECL_UID (var))) |
bbc8a8dc ZD |
1905 | continue; |
1906 | use = PHI_ARG_DEF_FROM_EDGE (phi, e); | |
1907 | gcc_assert (TREE_CODE (use) == SSA_NAME); | |
1908 | ||
1909 | /* Base all the ssa names in the ud and du chain of NAME on VAR. */ | |
1910 | stmt = SSA_NAME_DEF_STMT (use); | |
726a989a | 1911 | while (gimple_code (stmt) == GIMPLE_PHI |
1b0cfaa6 ZD |
1912 | /* In case we could not unroll the loop enough to eliminate |
1913 | all copies, we may reach the loop header before the defining | |
1914 | statement (in that case, some register copies will be present | |
1915 | in loop latch in the final code, corresponding to the newly | |
1916 | created looparound phi nodes). */ | |
726a989a | 1917 | && gimple_bb (stmt) != loop->header) |
bbc8a8dc | 1918 | { |
726a989a | 1919 | gcc_assert (single_pred_p (gimple_bb (stmt))); |
bbc8a8dc ZD |
1920 | use = PHI_ARG_DEF (stmt, 0); |
1921 | stmt = SSA_NAME_DEF_STMT (use); | |
1922 | } | |
1923 | ||
1924 | base_names_in_chain_on (loop, use, var); | |
1925 | } | |
1926 | } | |
1927 | ||
1928 | /* Returns true if CHAIN is suitable to be combined. */ | |
1929 | ||
1930 | static bool | |
1931 | chain_can_be_combined_p (chain_p chain) | |
1932 | { | |
1933 | return (!chain->combined | |
1934 | && (chain->type == CT_LOAD || chain->type == CT_COMBINATION)); | |
1935 | } | |
1936 | ||
1937 | /* Returns the modify statement that uses NAME. Skips over assignment | |
1938 | statements, NAME is replaced with the actual name used in the returned | |
1939 | statement. */ | |
1940 | ||
726a989a | 1941 | static gimple |
bbc8a8dc ZD |
1942 | find_use_stmt (tree *name) |
1943 | { | |
726a989a RB |
1944 | gimple stmt; |
1945 | tree rhs, lhs; | |
bbc8a8dc ZD |
1946 | |
1947 | /* Skip over assignments. */ | |
1948 | while (1) | |
1949 | { | |
1950 | stmt = single_nonlooparound_use (*name); | |
1951 | if (!stmt) | |
726a989a | 1952 | return NULL; |
bbc8a8dc | 1953 | |
726a989a RB |
1954 | if (gimple_code (stmt) != GIMPLE_ASSIGN) |
1955 | return NULL; | |
bbc8a8dc | 1956 | |
726a989a | 1957 | lhs = gimple_assign_lhs (stmt); |
bbc8a8dc | 1958 | if (TREE_CODE (lhs) != SSA_NAME) |
726a989a | 1959 | return NULL; |
bbc8a8dc | 1960 | |
726a989a RB |
1961 | if (gimple_assign_copy_p (stmt)) |
1962 | { | |
1963 | rhs = gimple_assign_rhs1 (stmt); | |
1964 | if (rhs != *name) | |
1965 | return NULL; | |
bbc8a8dc | 1966 | |
726a989a RB |
1967 | *name = lhs; |
1968 | } | |
1969 | else if (get_gimple_rhs_class (gimple_assign_rhs_code (stmt)) | |
1970 | == GIMPLE_BINARY_RHS) | |
1971 | return stmt; | |
1972 | else | |
1973 | return NULL; | |
bbc8a8dc | 1974 | } |
bbc8a8dc ZD |
1975 | } |
1976 | ||
1977 | /* Returns true if we may perform reassociation for operation CODE in TYPE. */ | |
1978 | ||
1979 | static bool | |
1980 | may_reassociate_p (tree type, enum tree_code code) | |
1981 | { | |
1982 | if (FLOAT_TYPE_P (type) | |
1983 | && !flag_unsafe_math_optimizations) | |
1984 | return false; | |
1985 | ||
1986 | return (commutative_tree_code (code) | |
1987 | && associative_tree_code (code)); | |
1988 | } | |
1989 | ||
1990 | /* If the operation used in STMT is associative and commutative, go through the | |
1991 | tree of the same operations and returns its root. Distance to the root | |
1992 | is stored in DISTANCE. */ | |
1993 | ||
726a989a RB |
1994 | static gimple |
1995 | find_associative_operation_root (gimple stmt, unsigned *distance) | |
bbc8a8dc | 1996 | { |
726a989a RB |
1997 | tree lhs; |
1998 | gimple next; | |
1999 | enum tree_code code = gimple_assign_rhs_code (stmt); | |
2000 | tree type = TREE_TYPE (gimple_assign_lhs (stmt)); | |
bbc8a8dc ZD |
2001 | unsigned dist = 0; |
2002 | ||
726a989a RB |
2003 | if (!may_reassociate_p (type, code)) |
2004 | return NULL; | |
bbc8a8dc ZD |
2005 | |
2006 | while (1) | |
2007 | { | |
726a989a | 2008 | lhs = gimple_assign_lhs (stmt); |
bbc8a8dc ZD |
2009 | gcc_assert (TREE_CODE (lhs) == SSA_NAME); |
2010 | ||
2011 | next = find_use_stmt (&lhs); | |
726a989a RB |
2012 | if (!next |
2013 | || gimple_assign_rhs_code (next) != code) | |
bbc8a8dc ZD |
2014 | break; |
2015 | ||
2016 | stmt = next; | |
2017 | dist++; | |
2018 | } | |
2019 | ||
2020 | if (distance) | |
2021 | *distance = dist; | |
2022 | return stmt; | |
2023 | } | |
2024 | ||
2025 | /* Returns the common statement in that NAME1 and NAME2 have a use. If there | |
2026 | is no such statement, returns NULL_TREE. In case the operation used on | |
c80b4100 | 2027 | NAME1 and NAME2 is associative and commutative, returns the root of the |
bbc8a8dc ZD |
2028 | tree formed by this operation instead of the statement that uses NAME1 or |
2029 | NAME2. */ | |
2030 | ||
726a989a | 2031 | static gimple |
bbc8a8dc ZD |
2032 | find_common_use_stmt (tree *name1, tree *name2) |
2033 | { | |
726a989a | 2034 | gimple stmt1, stmt2; |
bbc8a8dc ZD |
2035 | |
2036 | stmt1 = find_use_stmt (name1); | |
2037 | if (!stmt1) | |
726a989a | 2038 | return NULL; |
bbc8a8dc ZD |
2039 | |
2040 | stmt2 = find_use_stmt (name2); | |
2041 | if (!stmt2) | |
726a989a | 2042 | return NULL; |
bbc8a8dc ZD |
2043 | |
2044 | if (stmt1 == stmt2) | |
2045 | return stmt1; | |
2046 | ||
2047 | stmt1 = find_associative_operation_root (stmt1, NULL); | |
2048 | if (!stmt1) | |
726a989a | 2049 | return NULL; |
bbc8a8dc ZD |
2050 | stmt2 = find_associative_operation_root (stmt2, NULL); |
2051 | if (!stmt2) | |
726a989a | 2052 | return NULL; |
bbc8a8dc | 2053 | |
726a989a | 2054 | return (stmt1 == stmt2 ? stmt1 : NULL); |
bbc8a8dc ZD |
2055 | } |
2056 | ||
2057 | /* Checks whether R1 and R2 are combined together using CODE, with the result | |
2058 | in RSLT_TYPE, in order R1 CODE R2 if SWAP is false and in order R2 CODE R1 | |
2059 | if it is true. If CODE is ERROR_MARK, set these values instead. */ | |
2060 | ||
2061 | static bool | |
2062 | combinable_refs_p (dref r1, dref r2, | |
2063 | enum tree_code *code, bool *swap, tree *rslt_type) | |
2064 | { | |
2065 | enum tree_code acode; | |
2066 | bool aswap; | |
2067 | tree atype; | |
726a989a RB |
2068 | tree name1, name2; |
2069 | gimple stmt; | |
bbc8a8dc ZD |
2070 | |
2071 | name1 = name_for_ref (r1); | |
2072 | name2 = name_for_ref (r2); | |
2073 | gcc_assert (name1 != NULL_TREE && name2 != NULL_TREE); | |
2074 | ||
2075 | stmt = find_common_use_stmt (&name1, &name2); | |
2076 | ||
2077 | if (!stmt) | |
2078 | return false; | |
2079 | ||
726a989a | 2080 | acode = gimple_assign_rhs_code (stmt); |
bbc8a8dc | 2081 | aswap = (!commutative_tree_code (acode) |
726a989a RB |
2082 | && gimple_assign_rhs1 (stmt) != name1); |
2083 | atype = TREE_TYPE (gimple_assign_lhs (stmt)); | |
bbc8a8dc ZD |
2084 | |
2085 | if (*code == ERROR_MARK) | |
2086 | { | |
2087 | *code = acode; | |
2088 | *swap = aswap; | |
2089 | *rslt_type = atype; | |
2090 | return true; | |
2091 | } | |
2092 | ||
2093 | return (*code == acode | |
2094 | && *swap == aswap | |
2095 | && *rslt_type == atype); | |
2096 | } | |
2097 | ||
2098 | /* Remove OP from the operation on rhs of STMT, and replace STMT with | |
2099 | an assignment of the remaining operand. */ | |
2100 | ||
2101 | static void | |
726a989a | 2102 | remove_name_from_operation (gimple stmt, tree op) |
bbc8a8dc | 2103 | { |
726a989a RB |
2104 | tree other_op; |
2105 | gimple_stmt_iterator si; | |
bbc8a8dc | 2106 | |
726a989a | 2107 | gcc_assert (is_gimple_assign (stmt)); |
bbc8a8dc | 2108 | |
726a989a RB |
2109 | if (gimple_assign_rhs1 (stmt) == op) |
2110 | other_op = gimple_assign_rhs2 (stmt); | |
bbc8a8dc | 2111 | else |
726a989a RB |
2112 | other_op = gimple_assign_rhs1 (stmt); |
2113 | ||
2114 | si = gsi_for_stmt (stmt); | |
2115 | gimple_assign_set_rhs_from_tree (&si, other_op); | |
2116 | ||
2117 | /* We should not have reallocated STMT. */ | |
2118 | gcc_assert (gsi_stmt (si) == stmt); | |
2119 | ||
bbc8a8dc ZD |
2120 | update_stmt (stmt); |
2121 | } | |
2122 | ||
2123 | /* Reassociates the expression in that NAME1 and NAME2 are used so that they | |
2124 | are combined in a single statement, and returns this statement. */ | |
2125 | ||
726a989a | 2126 | static gimple |
bbc8a8dc ZD |
2127 | reassociate_to_the_same_stmt (tree name1, tree name2) |
2128 | { | |
726a989a RB |
2129 | gimple stmt1, stmt2, root1, root2, s1, s2; |
2130 | gimple new_stmt, tmp_stmt; | |
2131 | tree new_name, tmp_name, var, r1, r2; | |
bbc8a8dc ZD |
2132 | unsigned dist1, dist2; |
2133 | enum tree_code code; | |
2134 | tree type = TREE_TYPE (name1); | |
726a989a | 2135 | gimple_stmt_iterator bsi; |
bbc8a8dc ZD |
2136 | |
2137 | stmt1 = find_use_stmt (&name1); | |
2138 | stmt2 = find_use_stmt (&name2); | |
2139 | root1 = find_associative_operation_root (stmt1, &dist1); | |
2140 | root2 = find_associative_operation_root (stmt2, &dist2); | |
726a989a | 2141 | code = gimple_assign_rhs_code (stmt1); |
bbc8a8dc ZD |
2142 | |
2143 | gcc_assert (root1 && root2 && root1 == root2 | |
726a989a | 2144 | && code == gimple_assign_rhs_code (stmt2)); |
bbc8a8dc ZD |
2145 | |
2146 | /* Find the root of the nearest expression in that both NAME1 and NAME2 | |
2147 | are used. */ | |
2148 | r1 = name1; | |
2149 | s1 = stmt1; | |
2150 | r2 = name2; | |
2151 | s2 = stmt2; | |
2152 | ||
2153 | while (dist1 > dist2) | |
2154 | { | |
2155 | s1 = find_use_stmt (&r1); | |
726a989a | 2156 | r1 = gimple_assign_lhs (s1); |
bbc8a8dc ZD |
2157 | dist1--; |
2158 | } | |
2159 | while (dist2 > dist1) | |
2160 | { | |
2161 | s2 = find_use_stmt (&r2); | |
726a989a | 2162 | r2 = gimple_assign_lhs (s2); |
bbc8a8dc ZD |
2163 | dist2--; |
2164 | } | |
2165 | ||
2166 | while (s1 != s2) | |
2167 | { | |
2168 | s1 = find_use_stmt (&r1); | |
726a989a | 2169 | r1 = gimple_assign_lhs (s1); |
bbc8a8dc | 2170 | s2 = find_use_stmt (&r2); |
726a989a | 2171 | r2 = gimple_assign_lhs (s2); |
bbc8a8dc ZD |
2172 | } |
2173 | ||
2174 | /* Remove NAME1 and NAME2 from the statements in that they are used | |
2175 | currently. */ | |
2176 | remove_name_from_operation (stmt1, name1); | |
2177 | remove_name_from_operation (stmt2, name2); | |
2178 | ||
2179 | /* Insert the new statement combining NAME1 and NAME2 before S1, and | |
2180 | combine it with the rhs of S1. */ | |
acd63801 | 2181 | var = create_tmp_reg (type, "predreastmp"); |
726a989a RB |
2182 | new_name = make_ssa_name (var, NULL); |
2183 | new_stmt = gimple_build_assign_with_ops (code, new_name, name1, name2); | |
bbc8a8dc | 2184 | |
acd63801 | 2185 | var = create_tmp_reg (type, "predreastmp"); |
726a989a RB |
2186 | tmp_name = make_ssa_name (var, NULL); |
2187 | ||
2188 | /* Rhs of S1 may now be either a binary expression with operation | |
2189 | CODE, or gimple_val (in case that stmt1 == s1 or stmt2 == s1, | |
2190 | so that name1 or name2 was removed from it). */ | |
2191 | tmp_stmt = gimple_build_assign_with_ops (gimple_assign_rhs_code (s1), | |
2192 | tmp_name, | |
2193 | gimple_assign_rhs1 (s1), | |
2194 | gimple_assign_rhs2 (s1)); | |
2195 | ||
2196 | bsi = gsi_for_stmt (s1); | |
2197 | gimple_assign_set_rhs_with_ops (&bsi, code, new_name, tmp_name); | |
2198 | s1 = gsi_stmt (bsi); | |
bbc8a8dc ZD |
2199 | update_stmt (s1); |
2200 | ||
726a989a RB |
2201 | gsi_insert_before (&bsi, new_stmt, GSI_SAME_STMT); |
2202 | gsi_insert_before (&bsi, tmp_stmt, GSI_SAME_STMT); | |
bbc8a8dc ZD |
2203 | |
2204 | return new_stmt; | |
2205 | } | |
2206 | ||
2207 | /* Returns the statement that combines references R1 and R2. In case R1 | |
2208 | and R2 are not used in the same statement, but they are used with an | |
2209 | associative and commutative operation in the same expression, reassociate | |
2210 | the expression so that they are used in the same statement. */ | |
2211 | ||
726a989a | 2212 | static gimple |
bbc8a8dc ZD |
2213 | stmt_combining_refs (dref r1, dref r2) |
2214 | { | |
726a989a | 2215 | gimple stmt1, stmt2; |
bbc8a8dc ZD |
2216 | tree name1 = name_for_ref (r1); |
2217 | tree name2 = name_for_ref (r2); | |
2218 | ||
2219 | stmt1 = find_use_stmt (&name1); | |
2220 | stmt2 = find_use_stmt (&name2); | |
2221 | if (stmt1 == stmt2) | |
2222 | return stmt1; | |
2223 | ||
2224 | return reassociate_to_the_same_stmt (name1, name2); | |
2225 | } | |
2226 | ||
2227 | /* Tries to combine chains CH1 and CH2 together. If this succeeds, the | |
2228 | description of the new chain is returned, otherwise we return NULL. */ | |
2229 | ||
2230 | static chain_p | |
2231 | combine_chains (chain_p ch1, chain_p ch2) | |
2232 | { | |
2233 | dref r1, r2, nw; | |
2234 | enum tree_code op = ERROR_MARK; | |
2235 | bool swap = false; | |
2236 | chain_p new_chain; | |
2237 | unsigned i; | |
726a989a | 2238 | gimple root_stmt; |
bbc8a8dc ZD |
2239 | tree rslt_type = NULL_TREE; |
2240 | ||
2241 | if (ch1 == ch2) | |
a90352a0 | 2242 | return NULL; |
bbc8a8dc ZD |
2243 | if (ch1->length != ch2->length) |
2244 | return NULL; | |
2245 | ||
2246 | if (VEC_length (dref, ch1->refs) != VEC_length (dref, ch2->refs)) | |
2247 | return NULL; | |
2248 | ||
2249 | for (i = 0; (VEC_iterate (dref, ch1->refs, i, r1) | |
2250 | && VEC_iterate (dref, ch2->refs, i, r2)); i++) | |
2251 | { | |
2252 | if (r1->distance != r2->distance) | |
2253 | return NULL; | |
2254 | ||
2255 | if (!combinable_refs_p (r1, r2, &op, &swap, &rslt_type)) | |
2256 | return NULL; | |
2257 | } | |
2258 | ||
2259 | if (swap) | |
2260 | { | |
2261 | chain_p tmp = ch1; | |
2262 | ch1 = ch2; | |
2263 | ch2 = tmp; | |
2264 | } | |
2265 | ||
2266 | new_chain = XCNEW (struct chain); | |
2267 | new_chain->type = CT_COMBINATION; | |
82d6e6fc | 2268 | new_chain->op = op; |
bbc8a8dc ZD |
2269 | new_chain->ch1 = ch1; |
2270 | new_chain->ch2 = ch2; | |
2271 | new_chain->rslt_type = rslt_type; | |
2272 | new_chain->length = ch1->length; | |
2273 | ||
2274 | for (i = 0; (VEC_iterate (dref, ch1->refs, i, r1) | |
2275 | && VEC_iterate (dref, ch2->refs, i, r2)); i++) | |
2276 | { | |
7e5487a2 | 2277 | nw = XCNEW (struct dref_d); |
bbc8a8dc ZD |
2278 | nw->stmt = stmt_combining_refs (r1, r2); |
2279 | nw->distance = r1->distance; | |
2280 | ||
2281 | VEC_safe_push (dref, heap, new_chain->refs, nw); | |
2282 | } | |
2283 | ||
2284 | new_chain->has_max_use_after = false; | |
2285 | root_stmt = get_chain_root (new_chain)->stmt; | |
2286 | for (i = 1; VEC_iterate (dref, new_chain->refs, i, nw); i++) | |
2287 | { | |
2288 | if (nw->distance == new_chain->length | |
2289 | && !stmt_dominates_stmt_p (nw->stmt, root_stmt)) | |
2290 | { | |
2291 | new_chain->has_max_use_after = true; | |
2292 | break; | |
2293 | } | |
2294 | } | |
2295 | ||
2296 | ch1->combined = true; | |
2297 | ch2->combined = true; | |
2298 | return new_chain; | |
2299 | } | |
2300 | ||
2301 | /* Try to combine the CHAINS. */ | |
2302 | ||
2303 | static void | |
2304 | try_combine_chains (VEC (chain_p, heap) **chains) | |
2305 | { | |
2306 | unsigned i, j; | |
2307 | chain_p ch1, ch2, cch; | |
2308 | VEC (chain_p, heap) *worklist = NULL; | |
2309 | ||
ac47786e | 2310 | FOR_EACH_VEC_ELT (chain_p, *chains, i, ch1) |
bbc8a8dc ZD |
2311 | if (chain_can_be_combined_p (ch1)) |
2312 | VEC_safe_push (chain_p, heap, worklist, ch1); | |
2313 | ||
2314 | while (!VEC_empty (chain_p, worklist)) | |
2315 | { | |
2316 | ch1 = VEC_pop (chain_p, worklist); | |
2317 | if (!chain_can_be_combined_p (ch1)) | |
2318 | continue; | |
2319 | ||
ac47786e | 2320 | FOR_EACH_VEC_ELT (chain_p, *chains, j, ch2) |
bbc8a8dc ZD |
2321 | { |
2322 | if (!chain_can_be_combined_p (ch2)) | |
2323 | continue; | |
2324 | ||
2325 | cch = combine_chains (ch1, ch2); | |
2326 | if (cch) | |
2327 | { | |
2328 | VEC_safe_push (chain_p, heap, worklist, cch); | |
2329 | VEC_safe_push (chain_p, heap, *chains, cch); | |
2330 | break; | |
2331 | } | |
2332 | } | |
2333 | } | |
2334 | } | |
2335 | ||
bbc8a8dc ZD |
2336 | /* Prepare initializers for CHAIN in LOOP. Returns false if this is |
2337 | impossible because one of these initializers may trap, true otherwise. */ | |
2338 | ||
2339 | static bool | |
2340 | prepare_initializers_chain (struct loop *loop, chain_p chain) | |
2341 | { | |
2342 | unsigned i, n = (chain->type == CT_INVARIANT) ? 1 : chain->length; | |
2343 | struct data_reference *dr = get_chain_root (chain)->ref; | |
726a989a RB |
2344 | tree init; |
2345 | gimple_seq stmts; | |
bbc8a8dc ZD |
2346 | dref laref; |
2347 | edge entry = loop_preheader_edge (loop); | |
2348 | ||
2349 | /* Find the initializers for the variables, and check that they cannot | |
2350 | trap. */ | |
2351 | chain->inits = VEC_alloc (tree, heap, n); | |
2352 | for (i = 0; i < n; i++) | |
2353 | VEC_quick_push (tree, chain->inits, NULL_TREE); | |
2354 | ||
2355 | /* If we have replaced some looparound phi nodes, use their initializers | |
2356 | instead of creating our own. */ | |
ac47786e | 2357 | FOR_EACH_VEC_ELT (dref, chain->refs, i, laref) |
bbc8a8dc | 2358 | { |
726a989a | 2359 | if (gimple_code (laref->stmt) != GIMPLE_PHI) |
bbc8a8dc ZD |
2360 | continue; |
2361 | ||
2362 | gcc_assert (laref->distance > 0); | |
2363 | VEC_replace (tree, chain->inits, n - laref->distance, | |
2364 | PHI_ARG_DEF_FROM_EDGE (laref->stmt, entry)); | |
2365 | } | |
2366 | ||
2367 | for (i = 0; i < n; i++) | |
2368 | { | |
2369 | if (VEC_index (tree, chain->inits, i) != NULL_TREE) | |
2370 | continue; | |
2371 | ||
2372 | init = ref_at_iteration (loop, DR_REF (dr), (int) i - n); | |
2373 | if (!init) | |
2374 | return false; | |
b8698a0f | 2375 | |
bbc8a8dc ZD |
2376 | if (!chain->all_always_accessed && tree_could_trap_p (init)) |
2377 | return false; | |
2378 | ||
2379 | init = force_gimple_operand (init, &stmts, false, NULL_TREE); | |
2380 | if (stmts) | |
5006671f | 2381 | gsi_insert_seq_on_edge_immediate (entry, stmts); |
bbc8a8dc ZD |
2382 | |
2383 | VEC_replace (tree, chain->inits, i, init); | |
2384 | } | |
2385 | ||
2386 | return true; | |
2387 | } | |
2388 | ||
2389 | /* Prepare initializers for CHAINS in LOOP, and free chains that cannot | |
2390 | be used because the initializers might trap. */ | |
2391 | ||
2392 | static void | |
2393 | prepare_initializers (struct loop *loop, VEC (chain_p, heap) *chains) | |
2394 | { | |
2395 | chain_p chain; | |
2396 | unsigned i; | |
2397 | ||
2398 | for (i = 0; i < VEC_length (chain_p, chains); ) | |
2399 | { | |
2400 | chain = VEC_index (chain_p, chains, i); | |
2401 | if (prepare_initializers_chain (loop, chain)) | |
2402 | i++; | |
2403 | else | |
2404 | { | |
2405 | release_chain (chain); | |
2406 | VEC_unordered_remove (chain_p, chains, i); | |
2407 | } | |
2408 | } | |
2409 | } | |
2410 | ||
2411 | /* Performs predictive commoning for LOOP. Returns true if LOOP was | |
2412 | unrolled. */ | |
2413 | ||
2414 | static bool | |
2415 | tree_predictive_commoning_loop (struct loop *loop) | |
2416 | { | |
01be8516 | 2417 | VEC (loop_p, heap) *loop_nest; |
bbc8a8dc ZD |
2418 | VEC (data_reference_p, heap) *datarefs; |
2419 | VEC (ddr_p, heap) *dependences; | |
2420 | struct component *components; | |
2421 | VEC (chain_p, heap) *chains = NULL; | |
2422 | unsigned unroll_factor; | |
2423 | struct tree_niter_desc desc; | |
2424 | bool unroll = false; | |
2425 | edge exit; | |
2426 | bitmap tmp_vars; | |
2427 | ||
2428 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
2429 | fprintf (dump_file, "Processing loop %d\n", loop->num); | |
2430 | ||
2431 | /* Find the data references and split them into components according to their | |
2432 | dependence relations. */ | |
2433 | datarefs = VEC_alloc (data_reference_p, heap, 10); | |
2434 | dependences = VEC_alloc (ddr_p, heap, 10); | |
01be8516 | 2435 | loop_nest = VEC_alloc (loop_p, heap, 3); |
9ca3d00e AB |
2436 | if (! compute_data_dependences_for_loop (loop, true, &loop_nest, &datarefs, |
2437 | &dependences)) | |
2438 | { | |
2439 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
2440 | fprintf (dump_file, "Cannot analyze data dependencies\n"); | |
2441 | VEC_free (loop_p, heap, loop_nest); | |
2442 | free_data_refs (datarefs); | |
2443 | free_dependence_relations (dependences); | |
2444 | return false; | |
2445 | } | |
2446 | ||
bbc8a8dc ZD |
2447 | if (dump_file && (dump_flags & TDF_DETAILS)) |
2448 | dump_data_dependence_relations (dump_file, dependences); | |
2449 | ||
2450 | components = split_data_refs_to_components (loop, datarefs, dependences); | |
01be8516 | 2451 | VEC_free (loop_p, heap, loop_nest); |
bbc8a8dc ZD |
2452 | free_dependence_relations (dependences); |
2453 | if (!components) | |
2454 | { | |
2455 | free_data_refs (datarefs); | |
2456 | return false; | |
2457 | } | |
2458 | ||
2459 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
2460 | { | |
2461 | fprintf (dump_file, "Initial state:\n\n"); | |
2462 | dump_components (dump_file, components); | |
2463 | } | |
2464 | ||
2465 | /* Find the suitable components and split them into chains. */ | |
2466 | components = filter_suitable_components (loop, components); | |
2467 | ||
2468 | tmp_vars = BITMAP_ALLOC (NULL); | |
2469 | looparound_phis = BITMAP_ALLOC (NULL); | |
2470 | determine_roots (loop, components, &chains); | |
2471 | release_components (components); | |
2472 | ||
2473 | if (!chains) | |
2474 | { | |
2475 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
2476 | fprintf (dump_file, | |
2477 | "Predictive commoning failed: no suitable chains\n"); | |
2478 | goto end; | |
2479 | } | |
2480 | prepare_initializers (loop, chains); | |
2481 | ||
2482 | /* Try to combine the chains that are always worked with together. */ | |
2483 | try_combine_chains (&chains); | |
2484 | ||
2485 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
2486 | { | |
2487 | fprintf (dump_file, "Before commoning:\n\n"); | |
2488 | dump_chains (dump_file, chains); | |
2489 | } | |
2490 | ||
2491 | /* Determine the unroll factor, and if the loop should be unrolled, ensure | |
2492 | that its number of iterations is divisible by the factor. */ | |
2493 | unroll_factor = determine_unroll_factor (chains); | |
2494 | scev_reset (); | |
e1e6dc73 RG |
2495 | unroll = (unroll_factor > 1 |
2496 | && can_unroll_loop_p (loop, unroll_factor, &desc)); | |
bbc8a8dc ZD |
2497 | exit = single_dom_exit (loop); |
2498 | ||
2499 | /* Execute the predictive commoning transformations, and possibly unroll the | |
2500 | loop. */ | |
2501 | if (unroll) | |
2502 | { | |
2503 | struct epcc_data dta; | |
2504 | ||
2505 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
2506 | fprintf (dump_file, "Unrolling %u times.\n", unroll_factor); | |
2507 | ||
2508 | dta.chains = chains; | |
2509 | dta.tmp_vars = tmp_vars; | |
b8698a0f | 2510 | |
bbc8a8dc ZD |
2511 | update_ssa (TODO_update_ssa_only_virtuals); |
2512 | ||
2513 | /* Cfg manipulations performed in tree_transform_and_unroll_loop before | |
2514 | execute_pred_commoning_cbck is called may cause phi nodes to be | |
2515 | reallocated, which is a problem since CHAINS may point to these | |
2516 | statements. To fix this, we store the ssa names defined by the | |
2517 | phi nodes here instead of the phi nodes themselves, and restore | |
2518 | the phi nodes in execute_pred_commoning_cbck. A bit hacky. */ | |
2519 | replace_phis_by_defined_names (chains); | |
2520 | ||
2521 | tree_transform_and_unroll_loop (loop, unroll_factor, exit, &desc, | |
2522 | execute_pred_commoning_cbck, &dta); | |
2523 | eliminate_temp_copies (loop, tmp_vars); | |
2524 | } | |
2525 | else | |
2526 | { | |
2527 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
2528 | fprintf (dump_file, | |
2529 | "Executing predictive commoning without unrolling.\n"); | |
2530 | execute_pred_commoning (loop, chains, tmp_vars); | |
2531 | } | |
2532 | ||
2533 | end: ; | |
2534 | release_chains (chains); | |
2535 | free_data_refs (datarefs); | |
2536 | BITMAP_FREE (tmp_vars); | |
2537 | BITMAP_FREE (looparound_phis); | |
2538 | ||
2539 | free_affine_expand_cache (&name_expansions); | |
2540 | ||
2541 | return unroll; | |
2542 | } | |
2543 | ||
2544 | /* Runs predictive commoning. */ | |
2545 | ||
592c303d | 2546 | unsigned |
bbc8a8dc ZD |
2547 | tree_predictive_commoning (void) |
2548 | { | |
2549 | bool unrolled = false; | |
2550 | struct loop *loop; | |
2551 | loop_iterator li; | |
592c303d | 2552 | unsigned ret = 0; |
bbc8a8dc ZD |
2553 | |
2554 | initialize_original_copy_tables (); | |
2555 | FOR_EACH_LOOP (li, loop, LI_ONLY_INNERMOST) | |
8bcf15f6 JH |
2556 | if (optimize_loop_for_speed_p (loop)) |
2557 | { | |
2558 | unrolled |= tree_predictive_commoning_loop (loop); | |
2559 | } | |
bbc8a8dc ZD |
2560 | |
2561 | if (unrolled) | |
2562 | { | |
2563 | scev_reset (); | |
592c303d | 2564 | ret = TODO_cleanup_cfg; |
bbc8a8dc ZD |
2565 | } |
2566 | free_original_copy_tables (); | |
592c303d ZD |
2567 | |
2568 | return ret; | |
bbc8a8dc | 2569 | } |