]>
Commit | Line | Data |
---|---|---|
b076a3fd ZD |
1 | /* Array prefetching. |
2 | Copyright (C) 2005 Free Software Foundation, Inc. | |
3 | ||
4 | This file is part of GCC. | |
5 | ||
6 | GCC is free software; you can redistribute it and/or modify it | |
7 | under the terms of the GNU General Public License as published by the | |
8 | Free Software Foundation; either version 2, or (at your option) any | |
9 | later version. | |
10 | ||
11 | GCC is distributed in the hope that it will be useful, but WITHOUT | |
12 | ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or | |
13 | FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License | |
14 | for more details. | |
15 | ||
16 | You should have received a copy of the GNU General Public License | |
17 | along with GCC; see the file COPYING. If not, write to the Free | |
18 | Software Foundation, 59 Temple Place - Suite 330, Boston, MA | |
19 | 02111-1307, USA. */ | |
20 | ||
21 | #include "config.h" | |
22 | #include "system.h" | |
23 | #include "coretypes.h" | |
24 | #include "tm.h" | |
25 | #include "tree.h" | |
26 | #include "rtl.h" | |
27 | #include "tm_p.h" | |
28 | #include "hard-reg-set.h" | |
29 | #include "basic-block.h" | |
30 | #include "output.h" | |
31 | #include "diagnostic.h" | |
32 | #include "tree-flow.h" | |
33 | #include "tree-dump.h" | |
34 | #include "timevar.h" | |
35 | #include "cfgloop.h" | |
36 | #include "varray.h" | |
37 | #include "expr.h" | |
38 | #include "tree-pass.h" | |
39 | #include "ggc.h" | |
40 | #include "insn-config.h" | |
41 | #include "recog.h" | |
42 | #include "hashtab.h" | |
43 | #include "tree-chrec.h" | |
44 | #include "tree-scalar-evolution.h" | |
45 | #include "toplev.h" | |
46 | #include "params.h" | |
47 | #include "langhooks.h" | |
48 | ||
49 | /* This pass inserts prefetch instructions to optimize cache usage during | |
50 | accesses to arrays in loops. It processes loops sequentially and: | |
51 | ||
52 | 1) Gathers all memory references in the single loop. | |
53 | 2) For each of the references it decides when it is profitable to prefetch | |
54 | it. To do it, we evaluate the reuse among the accesses, and determines | |
55 | two values: PREFETCH_BEFORE (meaning that it only makes sense to do | |
56 | prefetching in the first PREFETCH_BEFORE iterations of the loop) and | |
57 | PREFETCH_MOD (meaning that it only makes sense to prefetch in the | |
58 | iterations of the loop that are zero modulo PREFETCH_MOD). For example | |
59 | (assuming cache line size is 64 bytes, char has size 1 byte and there | |
60 | is no hardware sequential prefetch): | |
61 | ||
62 | char *a; | |
63 | for (i = 0; i < max; i++) | |
64 | { | |
65 | a[255] = ...; (0) | |
66 | a[i] = ...; (1) | |
67 | a[i + 64] = ...; (2) | |
68 | a[16*i] = ...; (3) | |
69 | a[187*i] = ...; (4) | |
70 | a[187*i + 50] = ...; (5) | |
71 | } | |
72 | ||
73 | (0) obviously has PREFETCH_BEFORE 1 | |
74 | (1) has PREFETCH_BEFORE 64, since (2) accesses the same memory | |
75 | location 64 iterations before it, and PREFETCH_MOD 64 (since | |
76 | it hits the same cache line otherwise). | |
77 | (2) has PREFETCH_MOD 64 | |
78 | (3) has PREFETCH_MOD 4 | |
79 | (4) has PREFETCH_MOD 1. We do not set PREFETCH_BEFORE here, since | |
80 | the cache line accessed by (4) is the same with probability only | |
81 | 7/32. | |
82 | (5) has PREFETCH_MOD 1 as well. | |
83 | ||
84 | 3) We determine how much ahead we need to prefetch. The number of | |
85 | iterations needed is time to fetch / time spent in one iteration of | |
86 | the loop. The problem is that we do not know either of these values, | |
87 | so we just make a heuristic guess based on a magic (possibly) | |
88 | target-specific constant and size of the loop. | |
89 | ||
90 | 4) Determine which of the references we prefetch. We take into account | |
91 | that there is a maximum number of simultaneous prefetches (provided | |
92 | by machine description). We prefetch as many prefetches as possible | |
93 | while still within this bound (starting with those with lowest | |
94 | prefetch_mod, since they are responsible for most of the cache | |
95 | misses). | |
96 | ||
97 | 5) We unroll and peel loops so that we are able to satisfy PREFETCH_MOD | |
98 | and PREFETCH_BEFORE requirements (within some bounds), and to avoid | |
99 | prefetching nonaccessed memory. | |
100 | TODO -- actually implement peeling. | |
101 | ||
102 | 6) We actually emit the prefetch instructions. ??? Perhaps emit the | |
103 | prefetch instructions with guards in cases where 5) was not sufficient | |
104 | to satisfy the constraints? | |
105 | ||
106 | Some other TODO: | |
107 | -- write and use more general reuse analysis (that could be also used | |
108 | in other cache aimed loop optimizations) | |
109 | -- make it behave sanely together with the prefetches given by user | |
110 | (now we just ignore them; at the very least we should avoid | |
111 | optimizing loops in that user put his own prefetches) | |
112 | -- we assume cache line size alignment of arrays; this could be | |
113 | improved. */ | |
114 | ||
115 | /* Magic constants follow. These should be replaced by machine specific | |
116 | numbers. */ | |
117 | ||
b076a3fd ZD |
118 | /* True if write can be prefetched by a read prefetch. */ |
119 | ||
120 | #ifndef WRITE_CAN_USE_READ_PREFETCH | |
121 | #define WRITE_CAN_USE_READ_PREFETCH 1 | |
122 | #endif | |
123 | ||
124 | /* True if read can be prefetched by a write prefetch. */ | |
125 | ||
126 | #ifndef READ_CAN_USE_WRITE_PREFETCH | |
127 | #define READ_CAN_USE_WRITE_PREFETCH 0 | |
128 | #endif | |
129 | ||
47eb5b32 ZD |
130 | /* The size of the block loaded by a single prefetch. Usually, this is |
131 | the same as cache line size (at the moment, we only consider one level | |
132 | of cache hierarchy). */ | |
b076a3fd ZD |
133 | |
134 | #ifndef PREFETCH_BLOCK | |
47eb5b32 | 135 | #define PREFETCH_BLOCK L1_CACHE_LINE_SIZE |
b076a3fd ZD |
136 | #endif |
137 | ||
138 | /* Do we have a forward hardware sequential prefetching? */ | |
139 | ||
140 | #ifndef HAVE_FORWARD_PREFETCH | |
141 | #define HAVE_FORWARD_PREFETCH 0 | |
142 | #endif | |
143 | ||
144 | /* Do we have a backward hardware sequential prefetching? */ | |
145 | ||
146 | #ifndef HAVE_BACKWARD_PREFETCH | |
147 | #define HAVE_BACKWARD_PREFETCH 0 | |
148 | #endif | |
149 | ||
150 | /* In some cases we are only able to determine that there is a certain | |
151 | probability that the two accesses hit the same cache line. In this | |
152 | case, we issue the prefetches for both of them if this probability | |
153 | is less then (1000 - ACCEPTABLE_MISS_RATE) promile. */ | |
154 | ||
155 | #ifndef ACCEPTABLE_MISS_RATE | |
156 | #define ACCEPTABLE_MISS_RATE 50 | |
157 | #endif | |
158 | ||
159 | #ifndef HAVE_prefetch | |
160 | #define HAVE_prefetch 0 | |
161 | #endif | |
162 | ||
163 | /* The group of references between that reuse may occur. */ | |
164 | ||
165 | struct mem_ref_group | |
166 | { | |
167 | tree base; /* Base of the reference. */ | |
168 | HOST_WIDE_INT step; /* Step of the reference. */ | |
169 | struct mem_ref *refs; /* References in the group. */ | |
170 | struct mem_ref_group *next; /* Next group of references. */ | |
171 | }; | |
172 | ||
173 | /* Assigned to PREFETCH_BEFORE when all iterations are to be prefetched. */ | |
174 | ||
175 | #define PREFETCH_ALL (~(unsigned HOST_WIDE_INT) 0) | |
176 | ||
177 | /* The memory reference. */ | |
178 | ||
179 | struct mem_ref | |
180 | { | |
181 | tree stmt; /* Statement in that the reference appears. */ | |
182 | tree mem; /* The reference. */ | |
183 | HOST_WIDE_INT delta; /* Constant offset of the reference. */ | |
184 | bool write_p; /* Is it a write? */ | |
185 | struct mem_ref_group *group; /* The group of references it belongs to. */ | |
186 | unsigned HOST_WIDE_INT prefetch_mod; | |
187 | /* Prefetch only each PREFETCH_MOD-th | |
188 | iteration. */ | |
189 | unsigned HOST_WIDE_INT prefetch_before; | |
190 | /* Prefetch only first PREFETCH_BEFORE | |
191 | iterations. */ | |
192 | bool issue_prefetch_p; /* Should we really issue the prefetch? */ | |
193 | struct mem_ref *next; /* The next reference in the group. */ | |
194 | }; | |
195 | ||
75c40d56 | 196 | /* Dumps information about reference REF to FILE. */ |
b076a3fd ZD |
197 | |
198 | static void | |
199 | dump_mem_ref (FILE *file, struct mem_ref *ref) | |
200 | { | |
201 | fprintf (file, "Reference %p:\n", (void *) ref); | |
202 | ||
203 | fprintf (file, " group %p (base ", (void *) ref->group); | |
204 | print_generic_expr (file, ref->group->base, TDF_SLIM); | |
205 | fprintf (file, ", step "); | |
206 | fprintf (file, HOST_WIDE_INT_PRINT_DEC, ref->group->step); | |
207 | fprintf (file, ")\n"); | |
208 | ||
e324a72f | 209 | fprintf (file, " delta "); |
b076a3fd ZD |
210 | fprintf (file, HOST_WIDE_INT_PRINT_DEC, ref->delta); |
211 | fprintf (file, "\n"); | |
212 | ||
213 | fprintf (file, " %s\n", ref->write_p ? "write" : "read"); | |
214 | ||
215 | fprintf (file, "\n"); | |
216 | } | |
217 | ||
218 | /* Finds a group with BASE and STEP in GROUPS, or creates one if it does not | |
219 | exist. */ | |
220 | ||
221 | static struct mem_ref_group * | |
222 | find_or_create_group (struct mem_ref_group **groups, tree base, | |
223 | HOST_WIDE_INT step) | |
224 | { | |
225 | struct mem_ref_group *group; | |
226 | ||
227 | for (; *groups; groups = &(*groups)->next) | |
228 | { | |
229 | if ((*groups)->step == step | |
230 | && operand_equal_p ((*groups)->base, base, 0)) | |
231 | return *groups; | |
232 | ||
233 | /* Keep the list of groups sorted by decreasing step. */ | |
234 | if ((*groups)->step < step) | |
235 | break; | |
236 | } | |
237 | ||
238 | group = xcalloc (1, sizeof (struct mem_ref_group)); | |
239 | group->base = base; | |
240 | group->step = step; | |
241 | group->refs = NULL; | |
242 | group->next = *groups; | |
243 | *groups = group; | |
244 | ||
245 | return group; | |
246 | } | |
247 | ||
248 | /* Records a memory reference MEM in GROUP with offset DELTA and write status | |
249 | WRITE_P. The reference occurs in statement STMT. */ | |
250 | ||
251 | static void | |
252 | record_ref (struct mem_ref_group *group, tree stmt, tree mem, | |
253 | HOST_WIDE_INT delta, bool write_p) | |
254 | { | |
255 | struct mem_ref **aref; | |
256 | ||
257 | /* Do not record the same address twice. */ | |
258 | for (aref = &group->refs; *aref; aref = &(*aref)->next) | |
259 | { | |
260 | /* It does not have to be possible for write reference to reuse the read | |
261 | prefetch, or vice versa. */ | |
262 | if (!WRITE_CAN_USE_READ_PREFETCH | |
263 | && write_p | |
264 | && !(*aref)->write_p) | |
265 | continue; | |
266 | if (!READ_CAN_USE_WRITE_PREFETCH | |
267 | && !write_p | |
268 | && (*aref)->write_p) | |
269 | continue; | |
270 | ||
271 | if ((*aref)->delta == delta) | |
272 | return; | |
273 | } | |
274 | ||
275 | (*aref) = xcalloc (1, sizeof (struct mem_ref)); | |
276 | (*aref)->stmt = stmt; | |
277 | (*aref)->mem = mem; | |
278 | (*aref)->delta = delta; | |
279 | (*aref)->write_p = write_p; | |
280 | (*aref)->prefetch_before = PREFETCH_ALL; | |
281 | (*aref)->prefetch_mod = 1; | |
282 | (*aref)->issue_prefetch_p = false; | |
283 | (*aref)->group = group; | |
284 | (*aref)->next = NULL; | |
285 | ||
286 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
287 | dump_mem_ref (dump_file, *aref); | |
288 | } | |
289 | ||
290 | /* Release memory references in GROUPS. */ | |
291 | ||
292 | static void | |
293 | release_mem_refs (struct mem_ref_group *groups) | |
294 | { | |
295 | struct mem_ref_group *next_g; | |
296 | struct mem_ref *ref, *next_r; | |
297 | ||
298 | for (; groups; groups = next_g) | |
299 | { | |
300 | next_g = groups->next; | |
301 | for (ref = groups->refs; ref; ref = next_r) | |
302 | { | |
303 | next_r = ref->next; | |
304 | free (ref); | |
305 | } | |
306 | free (groups); | |
307 | } | |
308 | } | |
309 | ||
310 | /* A structure used to pass arguments to idx_analyze_ref. */ | |
311 | ||
312 | struct ar_data | |
313 | { | |
314 | struct loop *loop; /* Loop of the reference. */ | |
315 | tree stmt; /* Statement of the reference. */ | |
316 | HOST_WIDE_INT *step; /* Step of the memory reference. */ | |
317 | HOST_WIDE_INT *delta; /* Offset of the memory reference. */ | |
318 | }; | |
319 | ||
320 | /* Analyzes a single INDEX of a memory reference to obtain information | |
321 | described at analyze_ref. Callback for for_each_index. */ | |
322 | ||
323 | static bool | |
324 | idx_analyze_ref (tree base, tree *index, void *data) | |
325 | { | |
326 | struct ar_data *ar_data = data; | |
327 | tree ibase, step, stepsize; | |
328 | HOST_WIDE_INT istep, idelta = 0, imult = 1; | |
329 | affine_iv iv; | |
330 | ||
331 | if (TREE_CODE (base) == MISALIGNED_INDIRECT_REF | |
332 | || TREE_CODE (base) == ALIGN_INDIRECT_REF) | |
333 | return false; | |
334 | ||
335 | if (!simple_iv (ar_data->loop, ar_data->stmt, *index, &iv, false)) | |
336 | return false; | |
337 | ibase = iv.base; | |
338 | step = iv.step; | |
339 | ||
6e42ce54 ZD |
340 | if (!cst_and_fits_in_hwi (step)) |
341 | return false; | |
342 | istep = int_cst_value (step); | |
b076a3fd ZD |
343 | |
344 | if (TREE_CODE (ibase) == PLUS_EXPR | |
345 | && cst_and_fits_in_hwi (TREE_OPERAND (ibase, 1))) | |
346 | { | |
347 | idelta = int_cst_value (TREE_OPERAND (ibase, 1)); | |
348 | ibase = TREE_OPERAND (ibase, 0); | |
349 | } | |
350 | if (cst_and_fits_in_hwi (ibase)) | |
351 | { | |
352 | idelta += int_cst_value (ibase); | |
ff5e9a94 | 353 | ibase = build_int_cst (TREE_TYPE (ibase), 0); |
b076a3fd ZD |
354 | } |
355 | ||
356 | if (TREE_CODE (base) == ARRAY_REF) | |
357 | { | |
358 | stepsize = array_ref_element_size (base); | |
359 | if (!cst_and_fits_in_hwi (stepsize)) | |
360 | return false; | |
361 | imult = int_cst_value (stepsize); | |
362 | ||
363 | istep *= imult; | |
364 | idelta *= imult; | |
365 | } | |
366 | ||
367 | *ar_data->step += istep; | |
368 | *ar_data->delta += idelta; | |
369 | *index = ibase; | |
370 | ||
371 | return true; | |
372 | } | |
373 | ||
aac8b8ed | 374 | /* Tries to express REF_P in shape &BASE + STEP * iter + DELTA, where DELTA and |
b076a3fd | 375 | STEP are integer constants and iter is number of iterations of LOOP. The |
aac8b8ed RS |
376 | reference occurs in statement STMT. Strips nonaddressable component |
377 | references from REF_P. */ | |
b076a3fd ZD |
378 | |
379 | static bool | |
aac8b8ed | 380 | analyze_ref (struct loop *loop, tree *ref_p, tree *base, |
b076a3fd ZD |
381 | HOST_WIDE_INT *step, HOST_WIDE_INT *delta, |
382 | tree stmt) | |
383 | { | |
384 | struct ar_data ar_data; | |
385 | tree off; | |
386 | HOST_WIDE_INT bit_offset; | |
aac8b8ed | 387 | tree ref = *ref_p; |
b076a3fd ZD |
388 | |
389 | *step = 0; | |
390 | *delta = 0; | |
391 | ||
392 | /* First strip off the component references. Ignore bitfields. */ | |
393 | if (TREE_CODE (ref) == COMPONENT_REF | |
394 | && DECL_NONADDRESSABLE_P (TREE_OPERAND (ref, 1))) | |
395 | ref = TREE_OPERAND (ref, 0); | |
396 | ||
aac8b8ed RS |
397 | *ref_p = ref; |
398 | ||
b076a3fd ZD |
399 | for (; TREE_CODE (ref) == COMPONENT_REF; ref = TREE_OPERAND (ref, 0)) |
400 | { | |
401 | off = DECL_FIELD_BIT_OFFSET (TREE_OPERAND (ref, 1)); | |
402 | bit_offset = TREE_INT_CST_LOW (off); | |
403 | gcc_assert (bit_offset % BITS_PER_UNIT == 0); | |
404 | ||
405 | *delta += bit_offset / BITS_PER_UNIT; | |
406 | } | |
407 | ||
408 | *base = unshare_expr (ref); | |
409 | ar_data.loop = loop; | |
410 | ar_data.stmt = stmt; | |
411 | ar_data.step = step; | |
412 | ar_data.delta = delta; | |
413 | return for_each_index (base, idx_analyze_ref, &ar_data); | |
414 | } | |
415 | ||
416 | /* Record a memory reference REF to the list REFS. The reference occurs in | |
417 | LOOP in statement STMT and it is write if WRITE_P. */ | |
418 | ||
419 | static void | |
420 | gather_memory_references_ref (struct loop *loop, struct mem_ref_group **refs, | |
421 | tree ref, bool write_p, tree stmt) | |
422 | { | |
423 | tree base; | |
424 | HOST_WIDE_INT step, delta; | |
425 | struct mem_ref_group *agrp; | |
426 | ||
aac8b8ed | 427 | if (!analyze_ref (loop, &ref, &base, &step, &delta, stmt)) |
b076a3fd ZD |
428 | return; |
429 | ||
430 | /* Now we know that REF = &BASE + STEP * iter + DELTA, where DELTA and STEP | |
431 | are integer constants. */ | |
432 | agrp = find_or_create_group (refs, base, step); | |
433 | record_ref (agrp, stmt, ref, delta, write_p); | |
434 | } | |
435 | ||
436 | /* Record the suitable memory references in LOOP. */ | |
437 | ||
438 | static struct mem_ref_group * | |
439 | gather_memory_references (struct loop *loop) | |
440 | { | |
441 | basic_block *body = get_loop_body_in_dom_order (loop); | |
442 | basic_block bb; | |
443 | unsigned i; | |
444 | block_stmt_iterator bsi; | |
445 | tree stmt, lhs, rhs; | |
446 | struct mem_ref_group *refs = NULL; | |
447 | ||
448 | /* Scan the loop body in order, so that the former references precede the | |
449 | later ones. */ | |
450 | for (i = 0; i < loop->num_nodes; i++) | |
451 | { | |
452 | bb = body[i]; | |
453 | if (bb->loop_father != loop) | |
454 | continue; | |
455 | ||
456 | for (bsi = bsi_start (bb); !bsi_end_p (bsi); bsi_next (&bsi)) | |
457 | { | |
458 | stmt = bsi_stmt (bsi); | |
07beea0d | 459 | if (TREE_CODE (stmt) != GIMPLE_MODIFY_STMT) |
b076a3fd ZD |
460 | continue; |
461 | ||
07beea0d AH |
462 | lhs = GIMPLE_STMT_OPERAND (stmt, 0); |
463 | rhs = GIMPLE_STMT_OPERAND (stmt, 1); | |
b076a3fd ZD |
464 | |
465 | if (REFERENCE_CLASS_P (rhs)) | |
466 | gather_memory_references_ref (loop, &refs, rhs, false, stmt); | |
467 | if (REFERENCE_CLASS_P (lhs)) | |
468 | gather_memory_references_ref (loop, &refs, lhs, true, stmt); | |
469 | } | |
470 | } | |
471 | free (body); | |
472 | ||
473 | return refs; | |
474 | } | |
475 | ||
476 | /* Prune the prefetch candidate REF using the self-reuse. */ | |
477 | ||
478 | static void | |
479 | prune_ref_by_self_reuse (struct mem_ref *ref) | |
480 | { | |
481 | HOST_WIDE_INT step = ref->group->step; | |
482 | bool backward = step < 0; | |
483 | ||
484 | if (step == 0) | |
485 | { | |
486 | /* Prefetch references to invariant address just once. */ | |
487 | ref->prefetch_before = 1; | |
488 | return; | |
489 | } | |
490 | ||
491 | if (backward) | |
492 | step = -step; | |
493 | ||
494 | if (step > PREFETCH_BLOCK) | |
495 | return; | |
496 | ||
497 | if ((backward && HAVE_BACKWARD_PREFETCH) | |
498 | || (!backward && HAVE_FORWARD_PREFETCH)) | |
499 | { | |
500 | ref->prefetch_before = 1; | |
501 | return; | |
502 | } | |
503 | ||
504 | ref->prefetch_mod = PREFETCH_BLOCK / step; | |
505 | } | |
506 | ||
507 | /* Divides X by BY, rounding down. */ | |
508 | ||
509 | static HOST_WIDE_INT | |
510 | ddown (HOST_WIDE_INT x, unsigned HOST_WIDE_INT by) | |
511 | { | |
512 | gcc_assert (by > 0); | |
513 | ||
514 | if (x >= 0) | |
515 | return x / by; | |
516 | else | |
517 | return (x + by - 1) / by; | |
518 | } | |
519 | ||
520 | /* Prune the prefetch candidate REF using the reuse with BY. | |
521 | If BY_IS_BEFORE is true, BY is before REF in the loop. */ | |
522 | ||
523 | static void | |
524 | prune_ref_by_group_reuse (struct mem_ref *ref, struct mem_ref *by, | |
525 | bool by_is_before) | |
526 | { | |
527 | HOST_WIDE_INT step = ref->group->step; | |
528 | bool backward = step < 0; | |
529 | HOST_WIDE_INT delta_r = ref->delta, delta_b = by->delta; | |
530 | HOST_WIDE_INT delta = delta_b - delta_r; | |
531 | HOST_WIDE_INT hit_from; | |
532 | unsigned HOST_WIDE_INT prefetch_before, prefetch_block; | |
533 | ||
534 | if (delta == 0) | |
535 | { | |
536 | /* If the references has the same address, only prefetch the | |
537 | former. */ | |
538 | if (by_is_before) | |
539 | ref->prefetch_before = 0; | |
540 | ||
541 | return; | |
542 | } | |
543 | ||
544 | if (!step) | |
545 | { | |
546 | /* If the reference addresses are invariant and fall into the | |
547 | same cache line, prefetch just the first one. */ | |
548 | if (!by_is_before) | |
549 | return; | |
550 | ||
551 | if (ddown (ref->delta, PREFETCH_BLOCK) | |
552 | != ddown (by->delta, PREFETCH_BLOCK)) | |
553 | return; | |
554 | ||
555 | ref->prefetch_before = 0; | |
556 | return; | |
557 | } | |
558 | ||
559 | /* Only prune the reference that is behind in the array. */ | |
560 | if (backward) | |
561 | { | |
562 | if (delta > 0) | |
563 | return; | |
564 | ||
565 | /* Transform the data so that we may assume that the accesses | |
566 | are forward. */ | |
567 | delta = - delta; | |
568 | step = -step; | |
569 | delta_r = PREFETCH_BLOCK - 1 - delta_r; | |
570 | delta_b = PREFETCH_BLOCK - 1 - delta_b; | |
571 | } | |
572 | else | |
573 | { | |
574 | if (delta < 0) | |
575 | return; | |
576 | } | |
577 | ||
578 | /* Check whether the two references are likely to hit the same cache | |
579 | line, and how distant the iterations in that it occurs are from | |
580 | each other. */ | |
581 | ||
582 | if (step <= PREFETCH_BLOCK) | |
583 | { | |
584 | /* The accesses are sure to meet. Let us check when. */ | |
585 | hit_from = ddown (delta_b, PREFETCH_BLOCK) * PREFETCH_BLOCK; | |
586 | prefetch_before = (hit_from - delta_r + step - 1) / step; | |
587 | ||
588 | if (prefetch_before < ref->prefetch_before) | |
589 | ref->prefetch_before = prefetch_before; | |
590 | ||
591 | return; | |
592 | } | |
593 | ||
594 | /* A more complicated case. First let us ensure that size of cache line | |
595 | and step are coprime (here we assume that PREFETCH_BLOCK is a power | |
596 | of two. */ | |
597 | prefetch_block = PREFETCH_BLOCK; | |
598 | while ((step & 1) == 0 | |
599 | && prefetch_block > 1) | |
600 | { | |
601 | step >>= 1; | |
602 | prefetch_block >>= 1; | |
603 | delta >>= 1; | |
604 | } | |
605 | ||
606 | /* Now step > prefetch_block, and step and prefetch_block are coprime. | |
607 | Determine the probability that the accesses hit the same cache line. */ | |
608 | ||
609 | prefetch_before = delta / step; | |
610 | delta %= step; | |
611 | if ((unsigned HOST_WIDE_INT) delta | |
612 | <= (prefetch_block * ACCEPTABLE_MISS_RATE / 1000)) | |
613 | { | |
614 | if (prefetch_before < ref->prefetch_before) | |
615 | ref->prefetch_before = prefetch_before; | |
616 | ||
617 | return; | |
618 | } | |
619 | ||
620 | /* Try also the following iteration. */ | |
621 | prefetch_before++; | |
622 | delta = step - delta; | |
623 | if ((unsigned HOST_WIDE_INT) delta | |
624 | <= (prefetch_block * ACCEPTABLE_MISS_RATE / 1000)) | |
625 | { | |
626 | if (prefetch_before < ref->prefetch_before) | |
627 | ref->prefetch_before = prefetch_before; | |
628 | ||
629 | return; | |
630 | } | |
631 | ||
632 | /* The ref probably does not reuse by. */ | |
633 | return; | |
634 | } | |
635 | ||
636 | /* Prune the prefetch candidate REF using the reuses with other references | |
637 | in REFS. */ | |
638 | ||
639 | static void | |
640 | prune_ref_by_reuse (struct mem_ref *ref, struct mem_ref *refs) | |
641 | { | |
642 | struct mem_ref *prune_by; | |
643 | bool before = true; | |
644 | ||
645 | prune_ref_by_self_reuse (ref); | |
646 | ||
647 | for (prune_by = refs; prune_by; prune_by = prune_by->next) | |
648 | { | |
649 | if (prune_by == ref) | |
650 | { | |
651 | before = false; | |
652 | continue; | |
653 | } | |
654 | ||
655 | if (!WRITE_CAN_USE_READ_PREFETCH | |
656 | && ref->write_p | |
657 | && !prune_by->write_p) | |
658 | continue; | |
659 | if (!READ_CAN_USE_WRITE_PREFETCH | |
660 | && !ref->write_p | |
661 | && prune_by->write_p) | |
662 | continue; | |
663 | ||
664 | prune_ref_by_group_reuse (ref, prune_by, before); | |
665 | } | |
666 | } | |
667 | ||
668 | /* Prune the prefetch candidates in GROUP using the reuse analysis. */ | |
669 | ||
670 | static void | |
671 | prune_group_by_reuse (struct mem_ref_group *group) | |
672 | { | |
673 | struct mem_ref *ref_pruned; | |
674 | ||
675 | for (ref_pruned = group->refs; ref_pruned; ref_pruned = ref_pruned->next) | |
676 | { | |
677 | prune_ref_by_reuse (ref_pruned, group->refs); | |
678 | ||
679 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
680 | { | |
681 | fprintf (dump_file, "Reference %p:", (void *) ref_pruned); | |
682 | ||
683 | if (ref_pruned->prefetch_before == PREFETCH_ALL | |
684 | && ref_pruned->prefetch_mod == 1) | |
685 | fprintf (dump_file, " no restrictions"); | |
686 | else if (ref_pruned->prefetch_before == 0) | |
687 | fprintf (dump_file, " do not prefetch"); | |
688 | else if (ref_pruned->prefetch_before <= ref_pruned->prefetch_mod) | |
689 | fprintf (dump_file, " prefetch once"); | |
690 | else | |
691 | { | |
692 | if (ref_pruned->prefetch_before != PREFETCH_ALL) | |
693 | { | |
694 | fprintf (dump_file, " prefetch before "); | |
695 | fprintf (dump_file, HOST_WIDE_INT_PRINT_DEC, | |
696 | ref_pruned->prefetch_before); | |
697 | } | |
698 | if (ref_pruned->prefetch_mod != 1) | |
699 | { | |
700 | fprintf (dump_file, " prefetch mod "); | |
701 | fprintf (dump_file, HOST_WIDE_INT_PRINT_DEC, | |
702 | ref_pruned->prefetch_mod); | |
703 | } | |
704 | } | |
705 | fprintf (dump_file, "\n"); | |
706 | } | |
707 | } | |
708 | } | |
709 | ||
710 | /* Prune the list of prefetch candidates GROUPS using the reuse analysis. */ | |
711 | ||
712 | static void | |
713 | prune_by_reuse (struct mem_ref_group *groups) | |
714 | { | |
715 | for (; groups; groups = groups->next) | |
716 | prune_group_by_reuse (groups); | |
717 | } | |
718 | ||
719 | /* Returns true if we should issue prefetch for REF. */ | |
720 | ||
721 | static bool | |
722 | should_issue_prefetch_p (struct mem_ref *ref) | |
723 | { | |
724 | /* For now do not issue prefetches for only first few of the | |
725 | iterations. */ | |
726 | if (ref->prefetch_before != PREFETCH_ALL) | |
727 | return false; | |
728 | ||
729 | return true; | |
730 | } | |
731 | ||
732 | /* Decide which of the prefetch candidates in GROUPS to prefetch. | |
733 | AHEAD is the number of iterations to prefetch ahead (which corresponds | |
734 | to the number of simultaneous instances of one prefetch running at a | |
735 | time). UNROLL_FACTOR is the factor by that the loop is going to be | |
736 | unrolled. Returns true if there is anything to prefetch. */ | |
737 | ||
738 | static bool | |
739 | schedule_prefetches (struct mem_ref_group *groups, unsigned unroll_factor, | |
740 | unsigned ahead) | |
741 | { | |
911b3fdb ZD |
742 | unsigned remaining_prefetch_slots, n_prefetches, prefetch_slots; |
743 | unsigned slots_per_prefetch; | |
b076a3fd ZD |
744 | struct mem_ref *ref; |
745 | bool any = false; | |
746 | ||
911b3fdb ZD |
747 | /* At most SIMULTANEOUS_PREFETCHES should be running at the same time. */ |
748 | remaining_prefetch_slots = SIMULTANEOUS_PREFETCHES; | |
b076a3fd | 749 | |
911b3fdb ZD |
750 | /* The prefetch will run for AHEAD iterations of the original loop, i.e., |
751 | AHEAD / UNROLL_FACTOR iterations of the unrolled loop. In each iteration, | |
752 | it will need a prefetch slot. */ | |
753 | slots_per_prefetch = (ahead + unroll_factor / 2) / unroll_factor; | |
b076a3fd | 754 | if (dump_file && (dump_flags & TDF_DETAILS)) |
911b3fdb ZD |
755 | fprintf (dump_file, "Each prefetch instruction takes %u prefetch slots.\n", |
756 | slots_per_prefetch); | |
b076a3fd ZD |
757 | |
758 | /* For now we just take memory references one by one and issue | |
759 | prefetches for as many as possible. The groups are sorted | |
760 | starting with the largest step, since the references with | |
c0220ea4 | 761 | large step are more likely to cause many cache misses. */ |
b076a3fd ZD |
762 | |
763 | for (; groups; groups = groups->next) | |
764 | for (ref = groups->refs; ref; ref = ref->next) | |
765 | { | |
766 | if (!should_issue_prefetch_p (ref)) | |
767 | continue; | |
768 | ||
911b3fdb ZD |
769 | /* If we need to prefetch the reference each PREFETCH_MOD iterations, |
770 | and we unroll the loop UNROLL_FACTOR times, we need to insert | |
771 | ceil (UNROLL_FACTOR / PREFETCH_MOD) instructions in each | |
772 | iteration. */ | |
b076a3fd ZD |
773 | n_prefetches = ((unroll_factor + ref->prefetch_mod - 1) |
774 | / ref->prefetch_mod); | |
911b3fdb ZD |
775 | prefetch_slots = n_prefetches * slots_per_prefetch; |
776 | ||
777 | /* If more than half of the prefetches would be lost anyway, do not | |
778 | issue the prefetch. */ | |
779 | if (2 * remaining_prefetch_slots < prefetch_slots) | |
780 | continue; | |
781 | ||
782 | ref->issue_prefetch_p = true; | |
b076a3fd | 783 | |
911b3fdb ZD |
784 | if (remaining_prefetch_slots <= prefetch_slots) |
785 | return true; | |
786 | remaining_prefetch_slots -= prefetch_slots; | |
b076a3fd ZD |
787 | any = true; |
788 | } | |
789 | ||
790 | return any; | |
791 | } | |
792 | ||
793 | /* Determine whether there is any reference suitable for prefetching | |
794 | in GROUPS. */ | |
795 | ||
796 | static bool | |
797 | anything_to_prefetch_p (struct mem_ref_group *groups) | |
798 | { | |
799 | struct mem_ref *ref; | |
800 | ||
801 | for (; groups; groups = groups->next) | |
802 | for (ref = groups->refs; ref; ref = ref->next) | |
803 | if (should_issue_prefetch_p (ref)) | |
804 | return true; | |
805 | ||
806 | return false; | |
807 | } | |
808 | ||
809 | /* Issue prefetches for the reference REF into loop as decided before. | |
810 | HEAD is the number of iterations to prefetch ahead. UNROLL_FACTOR | |
917f1b7e | 811 | is the factor by which LOOP was unrolled. */ |
b076a3fd ZD |
812 | |
813 | static void | |
814 | issue_prefetch_ref (struct mem_ref *ref, unsigned unroll_factor, unsigned ahead) | |
815 | { | |
816 | HOST_WIDE_INT delta; | |
817 | tree addr, addr_base, prefetch, params, write_p; | |
818 | block_stmt_iterator bsi; | |
819 | unsigned n_prefetches, ap; | |
820 | ||
821 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
822 | fprintf (dump_file, "Issued prefetch for %p.\n", (void *) ref); | |
823 | ||
824 | bsi = bsi_for_stmt (ref->stmt); | |
825 | ||
826 | n_prefetches = ((unroll_factor + ref->prefetch_mod - 1) | |
827 | / ref->prefetch_mod); | |
828 | addr_base = build_fold_addr_expr_with_type (ref->mem, ptr_type_node); | |
829 | addr_base = force_gimple_operand_bsi (&bsi, unshare_expr (addr_base), true, NULL); | |
911b3fdb | 830 | write_p = ref->write_p ? integer_one_node : integer_zero_node; |
b076a3fd ZD |
831 | |
832 | for (ap = 0; ap < n_prefetches; ap++) | |
833 | { | |
834 | /* Determine the address to prefetch. */ | |
835 | delta = (ahead + ap * ref->prefetch_mod) * ref->group->step; | |
836 | addr = fold_build2 (PLUS_EXPR, ptr_type_node, | |
837 | addr_base, build_int_cst (ptr_type_node, delta)); | |
838 | addr = force_gimple_operand_bsi (&bsi, unshare_expr (addr), true, NULL); | |
839 | ||
840 | /* Create the prefetch instruction. */ | |
b076a3fd ZD |
841 | params = tree_cons (NULL_TREE, addr, |
842 | tree_cons (NULL_TREE, write_p, NULL_TREE)); | |
911b3fdb | 843 | |
b076a3fd ZD |
844 | prefetch = build_function_call_expr (built_in_decls[BUILT_IN_PREFETCH], |
845 | params); | |
846 | bsi_insert_before (&bsi, prefetch, BSI_SAME_STMT); | |
847 | } | |
848 | } | |
849 | ||
850 | /* Issue prefetches for the references in GROUPS into loop as decided before. | |
851 | HEAD is the number of iterations to prefetch ahead. UNROLL_FACTOR is the | |
852 | factor by that LOOP was unrolled. */ | |
853 | ||
854 | static void | |
855 | issue_prefetches (struct mem_ref_group *groups, | |
856 | unsigned unroll_factor, unsigned ahead) | |
857 | { | |
858 | struct mem_ref *ref; | |
859 | ||
860 | for (; groups; groups = groups->next) | |
861 | for (ref = groups->refs; ref; ref = ref->next) | |
862 | if (ref->issue_prefetch_p) | |
863 | issue_prefetch_ref (ref, unroll_factor, ahead); | |
864 | } | |
865 | ||
866 | /* Determines whether we can profitably unroll LOOP FACTOR times, and if | |
867 | this is the case, fill in DESC by the description of number of | |
868 | iterations. */ | |
869 | ||
870 | static bool | |
871 | should_unroll_loop_p (struct loop *loop, struct tree_niter_desc *desc, | |
872 | unsigned factor) | |
873 | { | |
874 | if (!can_unroll_loop_p (loop, factor, desc)) | |
875 | return false; | |
876 | ||
877 | /* We only consider loops without control flow for unrolling. This is not | |
878 | a hard restriction -- tree_unroll_loop works with arbitrary loops | |
879 | as well; but the unrolling/prefetching is usually more profitable for | |
880 | loops consisting of a single basic block, and we want to limit the | |
881 | code growth. */ | |
882 | if (loop->num_nodes > 2) | |
883 | return false; | |
884 | ||
885 | return true; | |
886 | } | |
887 | ||
888 | /* Determine the coefficient by that unroll LOOP, from the information | |
889 | contained in the list of memory references REFS. Description of | |
890 | umber of iterations of LOOP is stored to DESC. AHEAD is the number | |
891 | of iterations ahead that we need to prefetch. NINSNS is number of | |
892 | insns of the LOOP. */ | |
893 | ||
894 | static unsigned | |
895 | determine_unroll_factor (struct loop *loop, struct mem_ref_group *refs, | |
911b3fdb | 896 | unsigned ninsns, struct tree_niter_desc *desc) |
b076a3fd | 897 | { |
911b3fdb ZD |
898 | unsigned upper_bound; |
899 | unsigned nfactor, factor, mod_constraint; | |
b076a3fd ZD |
900 | struct mem_ref_group *agp; |
901 | struct mem_ref *ref; | |
902 | ||
911b3fdb ZD |
903 | /* First check whether the loop is not too large to unroll. We ignore |
904 | PARAM_MAX_UNROLL_TIMES, because for small loops, it prevented us | |
905 | from unrolling them enough to make exactly one cache line covered by each | |
906 | iteration. Also, the goal of PARAM_MAX_UNROLL_TIMES is to prevent | |
907 | us from unrolling the loops too many times in cases where we only expect | |
908 | gains from better scheduling and decreasing loop overhead, which is not | |
909 | the case here. */ | |
910 | upper_bound = PARAM_VALUE (PARAM_MAX_UNROLLED_INSNS) / ninsns; | |
911 | if (upper_bound <= 1) | |
b076a3fd ZD |
912 | return 1; |
913 | ||
911b3fdb ZD |
914 | /* Choose the factor so that we may prefetch each cache just once, |
915 | but bound the unrolling by UPPER_BOUND. */ | |
916 | factor = 1; | |
b076a3fd ZD |
917 | for (agp = refs; agp; agp = agp->next) |
918 | for (ref = agp->refs; ref; ref = ref->next) | |
911b3fdb ZD |
919 | if (should_issue_prefetch_p (ref)) |
920 | { | |
921 | mod_constraint = ref->prefetch_mod; | |
922 | nfactor = least_common_multiple (mod_constraint, factor); | |
923 | if (nfactor <= upper_bound) | |
924 | factor = nfactor; | |
925 | } | |
b076a3fd ZD |
926 | |
927 | if (!should_unroll_loop_p (loop, desc, factor)) | |
928 | return 1; | |
929 | ||
930 | return factor; | |
931 | } | |
932 | ||
933 | /* Issue prefetch instructions for array references in LOOP. Returns | |
d73be268 | 934 | true if the LOOP was unrolled. */ |
b076a3fd ZD |
935 | |
936 | static bool | |
d73be268 | 937 | loop_prefetch_arrays (struct loop *loop) |
b076a3fd ZD |
938 | { |
939 | struct mem_ref_group *refs; | |
940 | unsigned ahead, ninsns, unroll_factor; | |
941 | struct tree_niter_desc desc; | |
942 | bool unrolled = false; | |
943 | ||
944 | /* Step 1: gather the memory references. */ | |
945 | refs = gather_memory_references (loop); | |
946 | ||
947 | /* Step 2: estimate the reuse effects. */ | |
948 | prune_by_reuse (refs); | |
949 | ||
950 | if (!anything_to_prefetch_p (refs)) | |
951 | goto fail; | |
952 | ||
953 | /* Step 3: determine the ahead and unroll factor. */ | |
954 | ||
955 | /* FIXME: We should use not size of the loop, but the average number of | |
956 | instructions executed per iteration of the loop. */ | |
957 | ninsns = tree_num_loop_insns (loop); | |
958 | ahead = (PREFETCH_LATENCY + ninsns - 1) / ninsns; | |
911b3fdb | 959 | unroll_factor = determine_unroll_factor (loop, refs, ninsns, &desc); |
b076a3fd ZD |
960 | if (dump_file && (dump_flags & TDF_DETAILS)) |
961 | fprintf (dump_file, "Ahead %d, unroll factor %d\n", ahead, unroll_factor); | |
962 | ||
963 | /* If the loop rolls less than the required unroll factor, prefetching | |
964 | is useless. */ | |
965 | if (unroll_factor > 1 | |
966 | && cst_and_fits_in_hwi (desc.niter) | |
967 | && (unsigned HOST_WIDE_INT) int_cst_value (desc.niter) < unroll_factor) | |
968 | goto fail; | |
969 | ||
970 | /* Step 4: what to prefetch? */ | |
971 | if (!schedule_prefetches (refs, unroll_factor, ahead)) | |
972 | goto fail; | |
973 | ||
974 | /* Step 5: unroll the loop. TODO -- peeling of first and last few | |
975 | iterations so that we do not issue superfluous prefetches. */ | |
976 | if (unroll_factor != 1) | |
977 | { | |
d73be268 | 978 | tree_unroll_loop (loop, unroll_factor, |
b076a3fd ZD |
979 | single_dom_exit (loop), &desc); |
980 | unrolled = true; | |
981 | } | |
982 | ||
983 | /* Step 6: issue the prefetches. */ | |
984 | issue_prefetches (refs, unroll_factor, ahead); | |
985 | ||
986 | fail: | |
987 | release_mem_refs (refs); | |
988 | return unrolled; | |
989 | } | |
990 | ||
d73be268 | 991 | /* Issue prefetch instructions for array references in loops. */ |
b076a3fd | 992 | |
c7f965b6 | 993 | unsigned int |
d73be268 | 994 | tree_ssa_prefetch_arrays (void) |
b076a3fd | 995 | { |
42fd6772 | 996 | loop_iterator li; |
b076a3fd ZD |
997 | struct loop *loop; |
998 | bool unrolled = false; | |
c7f965b6 | 999 | int todo_flags = 0; |
b076a3fd ZD |
1000 | |
1001 | if (!HAVE_prefetch | |
1002 | /* It is possible to ask compiler for say -mtune=i486 -march=pentium4. | |
1003 | -mtune=i486 causes us having PREFETCH_BLOCK 0, since this is part | |
1004 | of processor costs and i486 does not have prefetch, but | |
1005 | -march=pentium4 causes HAVE_prefetch to be true. Ugh. */ | |
1006 | || PREFETCH_BLOCK == 0) | |
c7f965b6 | 1007 | return 0; |
b076a3fd | 1008 | |
47eb5b32 ZD |
1009 | if (dump_file && (dump_flags & TDF_DETAILS)) |
1010 | { | |
1011 | fprintf (dump_file, "Prefetching parameters:\n"); | |
1012 | fprintf (dump_file, " simultaneous prefetches: %d\n", | |
1013 | SIMULTANEOUS_PREFETCHES); | |
1014 | fprintf (dump_file, " prefetch latency: %d\n", PREFETCH_LATENCY); | |
1015 | fprintf (dump_file, " L1 cache size: %d (%d bytes)\n", | |
1016 | L1_CACHE_SIZE, L1_CACHE_SIZE * L1_CACHE_LINE_SIZE); | |
1017 | fprintf (dump_file, " L1 cache line size: %d\n", L1_CACHE_LINE_SIZE); | |
1018 | fprintf (dump_file, " prefetch block size: %d\n", PREFETCH_BLOCK); | |
1019 | fprintf (dump_file, "\n"); | |
1020 | } | |
1021 | ||
b076a3fd ZD |
1022 | initialize_original_copy_tables (); |
1023 | ||
1024 | if (!built_in_decls[BUILT_IN_PREFETCH]) | |
1025 | { | |
1026 | tree type = build_function_type (void_type_node, | |
1027 | tree_cons (NULL_TREE, | |
1028 | const_ptr_type_node, | |
1029 | NULL_TREE)); | |
c79efc4d RÁE |
1030 | tree decl = add_builtin_function ("__builtin_prefetch", type, |
1031 | BUILT_IN_PREFETCH, BUILT_IN_NORMAL, | |
1032 | NULL, NULL_TREE); | |
b076a3fd ZD |
1033 | DECL_IS_NOVOPS (decl) = true; |
1034 | built_in_decls[BUILT_IN_PREFETCH] = decl; | |
1035 | } | |
1036 | ||
1037 | /* We assume that size of cache line is a power of two, so verify this | |
1038 | here. */ | |
1039 | gcc_assert ((PREFETCH_BLOCK & (PREFETCH_BLOCK - 1)) == 0); | |
1040 | ||
42fd6772 | 1041 | FOR_EACH_LOOP (li, loop, LI_FROM_INNERMOST) |
b076a3fd | 1042 | { |
b076a3fd ZD |
1043 | if (dump_file && (dump_flags & TDF_DETAILS)) |
1044 | fprintf (dump_file, "Processing loop %d:\n", loop->num); | |
1045 | ||
d73be268 | 1046 | unrolled |= loop_prefetch_arrays (loop); |
b076a3fd ZD |
1047 | |
1048 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
1049 | fprintf (dump_file, "\n\n"); | |
1050 | } | |
1051 | ||
1052 | if (unrolled) | |
1053 | { | |
1054 | scev_reset (); | |
c7f965b6 | 1055 | todo_flags |= TODO_cleanup_cfg; |
b076a3fd ZD |
1056 | } |
1057 | ||
1058 | free_original_copy_tables (); | |
c7f965b6 | 1059 | return todo_flags; |
b076a3fd | 1060 | } |