1 | /* Array prefetching. |
2 | Copyright (C) 2005-2023 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 3, 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 COPYING3. If not see |
18 | <http://www.gnu.org/licenses/>. */ |
19 | |
20 | #include "config.h" |
21 | #include "system.h" |
22 | #include "coretypes.h" |
23 | #include "backend.h" |
24 | #include "target.h" |
25 | #include "rtl.h" |
26 | #include "tree.h" |
27 | #include "gimple.h" |
28 | #include "predict.h" |
29 | #include "tree-pass.h" |
30 | #include "gimple-ssa.h" |
31 | #include "optabs-query.h" |
32 | #include "tree-pretty-print.h" |
33 | #include "fold-const.h" |
34 | #include "stor-layout.h" |
35 | #include "gimplify.h" |
36 | #include "gimple-iterator.h" |
37 | #include "gimplify-me.h" |
38 | #include "tree-ssa-loop-ivopts.h" |
39 | #include "tree-ssa-loop-manip.h" |
40 | #include "tree-ssa-loop-niter.h" |
41 | #include "tree-ssa-loop.h" |
42 | #include "ssa.h" |
43 | #include "tree-into-ssa.h" |
44 | #include "cfgloop.h" |
45 | #include "tree-scalar-evolution.h" |
46 | #include "langhooks.h" |
47 | #include "tree-inline.h" |
48 | #include "tree-data-ref.h" |
49 | #include "diagnostic-core.h" |
50 | #include "dbgcnt.h" |
51 | |
52 | /* This pass inserts prefetch instructions to optimize cache usage during |
53 | accesses to arrays in loops. It processes loops sequentially and: |
54 | |
55 | 1) Gathers all memory references in the single loop. |
56 | 2) For each of the references it decides when it is profitable to prefetch |
57 | it. To do it, we evaluate the reuse among the accesses, and determines |
58 | two values: PREFETCH_BEFORE (meaning that it only makes sense to do |
59 | prefetching in the first PREFETCH_BEFORE iterations of the loop) and |
60 | PREFETCH_MOD (meaning that it only makes sense to prefetch in the |
61 | iterations of the loop that are zero modulo PREFETCH_MOD). For example |
62 | (assuming cache line size is 64 bytes, char has size 1 byte and there |
63 | is no hardware sequential prefetch): |
64 | |
65 | char *a; |
66 | for (i = 0; i < max; i++) |
67 | { |
68 | a[255] = ...; (0) |
69 | a[i] = ...; (1) |
70 | a[i + 64] = ...; (2) |
71 | a[16*i] = ...; (3) |
72 | a[187*i] = ...; (4) |
73 | a[187*i + 50] = ...; (5) |
74 | } |
75 | |
76 | (0) obviously has PREFETCH_BEFORE 1 |
77 | (1) has PREFETCH_BEFORE 64, since (2) accesses the same memory |
78 | location 64 iterations before it, and PREFETCH_MOD 64 (since |
79 | it hits the same cache line otherwise). |
80 | (2) has PREFETCH_MOD 64 |
81 | (3) has PREFETCH_MOD 4 |
82 | (4) has PREFETCH_MOD 1. We do not set PREFETCH_BEFORE here, since |
83 | the cache line accessed by (5) is the same with probability only |
84 | 7/32. |
85 | (5) has PREFETCH_MOD 1 as well. |
86 | |
87 | Additionally, we use data dependence analysis to determine for each |
88 | reference the distance till the first reuse; this information is used |
89 | to determine the temporality of the issued prefetch instruction. |
90 | |
91 | 3) We determine how much ahead we need to prefetch. The number of |
92 | iterations needed is time to fetch / time spent in one iteration of |
93 | the loop. The problem is that we do not know either of these values, |
94 | so we just make a heuristic guess based on a magic (possibly) |
95 | target-specific constant and size of the loop. |
96 | |
97 | 4) Determine which of the references we prefetch. We take into account |
98 | that there is a maximum number of simultaneous prefetches (provided |
99 | by machine description). We prefetch as many prefetches as possible |
100 | while still within this bound (starting with those with lowest |
101 | prefetch_mod, since they are responsible for most of the cache |
102 | misses). |
103 | |
104 | 5) We unroll and peel loops so that we are able to satisfy PREFETCH_MOD |
105 | and PREFETCH_BEFORE requirements (within some bounds), and to avoid |
106 | prefetching nonaccessed memory. |
107 | TODO -- actually implement peeling. |
108 | |
109 | 6) We actually emit the prefetch instructions. ??? Perhaps emit the |
110 | prefetch instructions with guards in cases where 5) was not sufficient |
111 | to satisfy the constraints? |
112 | |
113 | A cost model is implemented to determine whether or not prefetching is |
114 | profitable for a given loop. The cost model has three heuristics: |
115 | |
116 | 1. Function trip_count_to_ahead_ratio_too_small_p implements a |
117 | heuristic that determines whether or not the loop has too few |
118 | iterations (compared to ahead). Prefetching is not likely to be |
119 | beneficial if the trip count to ahead ratio is below a certain |
120 | minimum. |
121 | |
122 | 2. Function mem_ref_count_reasonable_p implements a heuristic that |
123 | determines whether the given loop has enough CPU ops that can be |
124 | overlapped with cache missing memory ops. If not, the loop |
125 | won't benefit from prefetching. In the implementation, |
126 | prefetching is not considered beneficial if the ratio between |
127 | the instruction count and the mem ref count is below a certain |
128 | minimum. |
129 | |
130 | 3. Function insn_to_prefetch_ratio_too_small_p implements a |
131 | heuristic that disables prefetching in a loop if the prefetching |
132 | cost is above a certain limit. The relative prefetching cost is |
133 | estimated by taking the ratio between the prefetch count and the |
134 | total intruction count (this models the I-cache cost). |
135 | |
136 | The limits used in these heuristics are defined as parameters with |
137 | reasonable default values. Machine-specific default values will be |
138 | added later. |
139 | |
140 | Some other TODO: |
141 | -- write and use more general reuse analysis (that could be also used |
142 | in other cache aimed loop optimizations) |
143 | -- make it behave sanely together with the prefetches given by user |
144 | (now we just ignore them; at the very least we should avoid |
145 | optimizing loops in that user put his own prefetches) |
146 | -- we assume cache line size alignment of arrays; this could be |
147 | improved. */ |
148 | |
149 | /* Magic constants follow. These should be replaced by machine specific |
150 | numbers. */ |
151 | |
152 | /* True if write can be prefetched by a read prefetch. */ |
153 | |
154 | #ifndef WRITE_CAN_USE_READ_PREFETCH |
155 | #define WRITE_CAN_USE_READ_PREFETCH 1 |
156 | #endif |
157 | |
158 | /* True if read can be prefetched by a write prefetch. */ |
159 | |
160 | #ifndef READ_CAN_USE_WRITE_PREFETCH |
161 | #define READ_CAN_USE_WRITE_PREFETCH 0 |
162 | #endif |
163 | |
164 | /* The size of the block loaded by a single prefetch. Usually, this is |
165 | the same as cache line size (at the moment, we only consider one level |
166 | of cache hierarchy). */ |
167 | |
168 | #ifndef PREFETCH_BLOCK |
169 | #define PREFETCH_BLOCK param_l1_cache_line_size |
170 | #endif |
171 | |
172 | /* Do we have a forward hardware sequential prefetching? */ |
173 | |
174 | #ifndef HAVE_FORWARD_PREFETCH |
175 | #define HAVE_FORWARD_PREFETCH 0 |
176 | #endif |
177 | |
178 | /* Do we have a backward hardware sequential prefetching? */ |
179 | |
180 | #ifndef HAVE_BACKWARD_PREFETCH |
181 | #define HAVE_BACKWARD_PREFETCH 0 |
182 | #endif |
183 | |
184 | /* In some cases we are only able to determine that there is a certain |
185 | probability that the two accesses hit the same cache line. In this |
186 | case, we issue the prefetches for both of them if this probability |
187 | is less then (1000 - ACCEPTABLE_MISS_RATE) per thousand. */ |
188 | |
189 | #ifndef ACCEPTABLE_MISS_RATE |
190 | #define ACCEPTABLE_MISS_RATE 50 |
191 | #endif |
192 | |
193 | #define L1_CACHE_SIZE_BYTES ((unsigned) (param_l1_cache_size * 1024)) |
194 | #define L2_CACHE_SIZE_BYTES ((unsigned) (param_l2_cache_size * 1024)) |
195 | |
196 | /* We consider a memory access nontemporal if it is not reused sooner than |
197 | after L2_CACHE_SIZE_BYTES of memory are accessed. However, we ignore |
198 | accesses closer than L1_CACHE_SIZE_BYTES / NONTEMPORAL_FRACTION, |
199 | so that we use nontemporal prefetches e.g. if single memory location |
200 | is accessed several times in a single iteration of the loop. */ |
201 | #define NONTEMPORAL_FRACTION 16 |
202 | |
203 | /* In case we have to emit a memory fence instruction after the loop that |
204 | uses nontemporal stores, this defines the builtin to use. */ |
205 | |
206 | #ifndef FENCE_FOLLOWING_MOVNT |
207 | #define FENCE_FOLLOWING_MOVNT NULL_TREE |
208 | #endif |
209 | |
210 | /* It is not profitable to prefetch when the trip count is not at |
211 | least TRIP_COUNT_TO_AHEAD_RATIO times the prefetch ahead distance. |
212 | For example, in a loop with a prefetch ahead distance of 10, |
213 | supposing that TRIP_COUNT_TO_AHEAD_RATIO is equal to 4, it is |
214 | profitable to prefetch when the trip count is greater or equal to |
215 | 40. In that case, 30 out of the 40 iterations will benefit from |
216 | prefetching. */ |
217 | |
218 | #ifndef TRIP_COUNT_TO_AHEAD_RATIO |
219 | #define TRIP_COUNT_TO_AHEAD_RATIO 4 |
220 | #endif |
221 | |
222 | /* The group of references between that reuse may occur. */ |
223 | |
224 | struct mem_ref_group |
225 | { |
226 | tree base; /* Base of the reference. */ |
227 | tree step; /* Step of the reference. */ |
228 | struct mem_ref *refs; /* References in the group. */ |
229 | struct mem_ref_group *next; /* Next group of references. */ |
230 | unsigned int uid; /* Group UID, used only for debugging. */ |
231 | }; |
232 | |
233 | /* Assigned to PREFETCH_BEFORE when all iterations are to be prefetched. */ |
234 | |
235 | #define PREFETCH_ALL HOST_WIDE_INT_M1U |
236 | |
237 | /* Do not generate a prefetch if the unroll factor is significantly less |
238 | than what is required by the prefetch. This is to avoid redundant |
239 | prefetches. For example, when prefetch_mod is 16 and unroll_factor is |
240 | 2, prefetching requires unrolling the loop 16 times, but |
241 | the loop is actually unrolled twice. In this case (ratio = 8), |
242 | prefetching is not likely to be beneficial. */ |
243 | |
244 | #ifndef PREFETCH_MOD_TO_UNROLL_FACTOR_RATIO |
245 | #define PREFETCH_MOD_TO_UNROLL_FACTOR_RATIO 4 |
246 | #endif |
247 | |
248 | /* Some of the prefetch computations have quadratic complexity. We want to |
249 | avoid huge compile times and, therefore, want to limit the amount of |
250 | memory references per loop where we consider prefetching. */ |
251 | |
252 | #ifndef PREFETCH_MAX_MEM_REFS_PER_LOOP |
253 | #define PREFETCH_MAX_MEM_REFS_PER_LOOP 200 |
254 | #endif |
255 | |
256 | /* The memory reference. */ |
257 | |
258 | struct mem_ref |
259 | { |
260 | gimple *stmt; /* Statement in that the reference appears. */ |
261 | tree mem; /* The reference. */ |
262 | HOST_WIDE_INT delta; /* Constant offset of the reference. */ |
263 | struct mem_ref_group *group; /* The group of references it belongs to. */ |
264 | unsigned HOST_WIDE_INT prefetch_mod; |
265 | /* Prefetch only each PREFETCH_MOD-th |
266 | iteration. */ |
267 | unsigned HOST_WIDE_INT prefetch_before; |
268 | /* Prefetch only first PREFETCH_BEFORE |
269 | iterations. */ |
270 | unsigned reuse_distance; /* The amount of data accessed before the first |
271 | reuse of this value. */ |
272 | struct mem_ref *next; /* The next reference in the group. */ |
273 | unsigned int uid; /* Ref UID, used only for debugging. */ |
274 | unsigned write_p : 1; /* Is it a write? */ |
275 | unsigned independent_p : 1; /* True if the reference is independent on |
276 | all other references inside the loop. */ |
277 | unsigned issue_prefetch_p : 1; /* Should we really issue the prefetch? */ |
278 | unsigned storent_p : 1; /* True if we changed the store to a |
279 | nontemporal one. */ |
280 | }; |
281 | |
282 | /* Dumps information about memory reference */ |
283 | static void |
284 | dump_mem_details (FILE *file, tree base, tree step, |
285 | HOST_WIDE_INT delta, bool write_p) |
286 | { |
287 | fprintf (stream: file, format: "(base " ); |
288 | print_generic_expr (file, base, TDF_SLIM); |
289 | fprintf (stream: file, format: ", step " ); |
290 | if (cst_and_fits_in_hwi (step)) |
291 | fprintf (stream: file, HOST_WIDE_INT_PRINT_DEC, int_cst_value (step)); |
292 | else |
293 | print_generic_expr (file, step, TDF_SLIM); |
294 | fprintf (stream: file, format: ")\n" ); |
295 | fprintf (stream: file, format: " delta " HOST_WIDE_INT_PRINT_DEC "\n" , delta); |
296 | fprintf (stream: file, format: " %s\n\n" , write_p ? "write" : "read" ); |
297 | } |
298 | |
299 | /* Dumps information about reference REF to FILE. */ |
300 | |
301 | static void |
302 | dump_mem_ref (FILE *file, struct mem_ref *ref) |
303 | { |
304 | fprintf (stream: file, format: "reference %u:%u (" , ref->group->uid, ref->uid); |
305 | print_generic_expr (file, ref->mem, TDF_SLIM); |
306 | fprintf (stream: file, format: ")\n" ); |
307 | } |
308 | |
309 | /* Finds a group with BASE and STEP in GROUPS, or creates one if it does not |
310 | exist. */ |
311 | |
312 | static struct mem_ref_group * |
313 | find_or_create_group (struct mem_ref_group **groups, tree base, tree step) |
314 | { |
315 | /* Global count for setting struct mem_ref_group->uid. */ |
316 | static unsigned int last_mem_ref_group_uid = 0; |
317 | |
318 | struct mem_ref_group *group; |
319 | |
320 | for (; *groups; groups = &(*groups)->next) |
321 | { |
322 | if (operand_equal_p ((*groups)->step, step, flags: 0) |
323 | && operand_equal_p ((*groups)->base, base, flags: 0)) |
324 | return *groups; |
325 | |
326 | /* If step is an integer constant, keep the list of groups sorted |
327 | by decreasing step. */ |
328 | if (cst_and_fits_in_hwi ((*groups)->step) && cst_and_fits_in_hwi (step) |
329 | && int_cst_value ((*groups)->step) < int_cst_value (step)) |
330 | break; |
331 | } |
332 | |
333 | group = XNEW (struct mem_ref_group); |
334 | group->base = base; |
335 | group->step = step; |
336 | group->refs = NULL; |
337 | group->uid = ++last_mem_ref_group_uid; |
338 | group->next = *groups; |
339 | *groups = group; |
340 | |
341 | return group; |
342 | } |
343 | |
344 | /* Records a memory reference MEM in GROUP with offset DELTA and write status |
345 | WRITE_P. The reference occurs in statement STMT. */ |
346 | |
347 | static void |
348 | record_ref (struct mem_ref_group *group, gimple *stmt, tree mem, |
349 | HOST_WIDE_INT delta, bool write_p) |
350 | { |
351 | unsigned int last_mem_ref_uid = 0; |
352 | struct mem_ref **aref; |
353 | |
354 | /* Do not record the same address twice. */ |
355 | for (aref = &group->refs; *aref; aref = &(*aref)->next) |
356 | { |
357 | last_mem_ref_uid = (*aref)->uid; |
358 | |
359 | /* It does not have to be possible for write reference to reuse the read |
360 | prefetch, or vice versa. */ |
361 | if (!WRITE_CAN_USE_READ_PREFETCH |
362 | && write_p |
363 | && !(*aref)->write_p) |
364 | continue; |
365 | if (!READ_CAN_USE_WRITE_PREFETCH |
366 | && !write_p |
367 | && (*aref)->write_p) |
368 | continue; |
369 | |
370 | if ((*aref)->delta == delta) |
371 | return; |
372 | } |
373 | |
374 | (*aref) = XNEW (struct mem_ref); |
375 | (*aref)->stmt = stmt; |
376 | (*aref)->mem = mem; |
377 | (*aref)->delta = delta; |
378 | (*aref)->write_p = write_p; |
379 | (*aref)->prefetch_before = PREFETCH_ALL; |
380 | (*aref)->prefetch_mod = 1; |
381 | (*aref)->reuse_distance = 0; |
382 | (*aref)->issue_prefetch_p = false; |
383 | (*aref)->group = group; |
384 | (*aref)->next = NULL; |
385 | (*aref)->independent_p = false; |
386 | (*aref)->storent_p = false; |
387 | (*aref)->uid = last_mem_ref_uid + 1; |
388 | |
389 | if (dump_file && (dump_flags & TDF_DETAILS)) |
390 | { |
391 | dump_mem_ref (file: dump_file, ref: *aref); |
392 | |
393 | fprintf (stream: dump_file, format: " group %u " , group->uid); |
394 | dump_mem_details (file: dump_file, base: group->base, step: group->step, delta, |
395 | write_p); |
396 | } |
397 | } |
398 | |
399 | /* Release memory references in GROUPS. */ |
400 | |
401 | static void |
402 | release_mem_refs (struct mem_ref_group *groups) |
403 | { |
404 | struct mem_ref_group *next_g; |
405 | struct mem_ref *ref, *next_r; |
406 | |
407 | for (; groups; groups = next_g) |
408 | { |
409 | next_g = groups->next; |
410 | for (ref = groups->refs; ref; ref = next_r) |
411 | { |
412 | next_r = ref->next; |
413 | free (ptr: ref); |
414 | } |
415 | free (ptr: groups); |
416 | } |
417 | } |
418 | |
419 | /* A structure used to pass arguments to idx_analyze_ref. */ |
420 | |
421 | struct ar_data |
422 | { |
423 | class loop *loop; /* Loop of the reference. */ |
424 | gimple *stmt; /* Statement of the reference. */ |
425 | tree *step; /* Step of the memory reference. */ |
426 | HOST_WIDE_INT *delta; /* Offset of the memory reference. */ |
427 | }; |
428 | |
429 | /* Analyzes a single INDEX of a memory reference to obtain information |
430 | described at analyze_ref. Callback for for_each_index. */ |
431 | |
432 | static bool |
433 | idx_analyze_ref (tree base, tree *index, void *data) |
434 | { |
435 | struct ar_data *ar_data = (struct ar_data *) data; |
436 | tree ibase, step, stepsize; |
437 | HOST_WIDE_INT idelta = 0, imult = 1; |
438 | affine_iv iv; |
439 | |
440 | if (!simple_iv (ar_data->loop, loop_containing_stmt (stmt: ar_data->stmt), |
441 | *index, &iv, true)) |
442 | return false; |
443 | ibase = iv.base; |
444 | step = iv.step; |
445 | |
446 | if (TREE_CODE (ibase) == POINTER_PLUS_EXPR |
447 | && cst_and_fits_in_hwi (TREE_OPERAND (ibase, 1))) |
448 | { |
449 | idelta = int_cst_value (TREE_OPERAND (ibase, 1)); |
450 | ibase = TREE_OPERAND (ibase, 0); |
451 | } |
452 | if (cst_and_fits_in_hwi (ibase)) |
453 | { |
454 | idelta += int_cst_value (ibase); |
455 | ibase = build_int_cst (TREE_TYPE (ibase), 0); |
456 | } |
457 | |
458 | if (TREE_CODE (base) == ARRAY_REF) |
459 | { |
460 | stepsize = array_ref_element_size (base); |
461 | if (!cst_and_fits_in_hwi (stepsize)) |
462 | return false; |
463 | imult = int_cst_value (stepsize); |
464 | step = fold_build2 (MULT_EXPR, sizetype, |
465 | fold_convert (sizetype, step), |
466 | fold_convert (sizetype, stepsize)); |
467 | idelta *= imult; |
468 | } |
469 | |
470 | if (*ar_data->step == NULL_TREE) |
471 | *ar_data->step = step; |
472 | else |
473 | *ar_data->step = fold_build2 (PLUS_EXPR, sizetype, |
474 | fold_convert (sizetype, *ar_data->step), |
475 | fold_convert (sizetype, step)); |
476 | *ar_data->delta += idelta; |
477 | *index = ibase; |
478 | |
479 | return true; |
480 | } |
481 | |
482 | /* Tries to express REF_P in shape &BASE + STEP * iter + DELTA, where DELTA and |
483 | STEP are integer constants and iter is number of iterations of LOOP. The |
484 | reference occurs in statement STMT. Strips nonaddressable component |
485 | references from REF_P. */ |
486 | |
487 | static bool |
488 | analyze_ref (class loop *loop, tree *ref_p, tree *base, |
489 | tree *step, HOST_WIDE_INT *delta, |
490 | gimple *stmt) |
491 | { |
492 | struct ar_data ar_data; |
493 | tree off; |
494 | HOST_WIDE_INT bit_offset; |
495 | tree ref = *ref_p; |
496 | |
497 | *step = NULL_TREE; |
498 | *delta = 0; |
499 | |
500 | /* First strip off the component references. Ignore bitfields. |
501 | Also strip off the real and imagine parts of a complex, so that |
502 | they can have the same base. */ |
503 | if (TREE_CODE (ref) == REALPART_EXPR |
504 | || TREE_CODE (ref) == IMAGPART_EXPR |
505 | || (TREE_CODE (ref) == COMPONENT_REF |
506 | && DECL_NONADDRESSABLE_P (TREE_OPERAND (ref, 1)))) |
507 | { |
508 | if (TREE_CODE (ref) == IMAGPART_EXPR) |
509 | *delta += int_size_in_bytes (TREE_TYPE (ref)); |
510 | ref = TREE_OPERAND (ref, 0); |
511 | } |
512 | |
513 | *ref_p = ref; |
514 | |
515 | for (; TREE_CODE (ref) == COMPONENT_REF; ref = TREE_OPERAND (ref, 0)) |
516 | { |
517 | off = DECL_FIELD_BIT_OFFSET (TREE_OPERAND (ref, 1)); |
518 | bit_offset = TREE_INT_CST_LOW (off); |
519 | gcc_assert (bit_offset % BITS_PER_UNIT == 0); |
520 | |
521 | *delta += bit_offset / BITS_PER_UNIT; |
522 | } |
523 | |
524 | *base = unshare_expr (ref); |
525 | ar_data.loop = loop; |
526 | ar_data.stmt = stmt; |
527 | ar_data.step = step; |
528 | ar_data.delta = delta; |
529 | return for_each_index (base, idx_analyze_ref, &ar_data); |
530 | } |
531 | |
532 | /* Record a memory reference REF to the list REFS. The reference occurs in |
533 | LOOP in statement STMT and it is write if WRITE_P. Returns true if the |
534 | reference was recorded, false otherwise. */ |
535 | |
536 | static bool |
537 | gather_memory_references_ref (class loop *loop, struct mem_ref_group **refs, |
538 | tree ref, bool write_p, gimple *stmt) |
539 | { |
540 | tree base, step; |
541 | HOST_WIDE_INT delta; |
542 | struct mem_ref_group *agrp; |
543 | |
544 | if (get_base_address (t: ref) == NULL) |
545 | return false; |
546 | |
547 | if (!analyze_ref (loop, ref_p: &ref, base: &base, step: &step, delta: &delta, stmt)) |
548 | return false; |
549 | /* If analyze_ref fails the default is a NULL_TREE. We can stop here. */ |
550 | if (step == NULL_TREE) |
551 | return false; |
552 | |
553 | /* Stop if the address of BASE could not be taken. */ |
554 | if (may_be_nonaddressable_p (expr: base)) |
555 | return false; |
556 | |
557 | /* Limit non-constant step prefetching only to the innermost loops and |
558 | only when the step is loop invariant in the entire loop nest. */ |
559 | if (!cst_and_fits_in_hwi (step)) |
560 | { |
561 | if (loop->inner != NULL) |
562 | { |
563 | if (dump_file && (dump_flags & TDF_DETAILS)) |
564 | { |
565 | fprintf (stream: dump_file, format: "Memory expression %p\n" ,(void *) ref ); |
566 | print_generic_expr (dump_file, ref, TDF_SLIM); |
567 | fprintf (stream: dump_file,format: ":" ); |
568 | dump_mem_details (file: dump_file, base, step, delta, write_p); |
569 | fprintf (stream: dump_file, |
570 | format: "Ignoring %p, non-constant step prefetching is " |
571 | "limited to inner most loops \n" , |
572 | (void *) ref); |
573 | } |
574 | return false; |
575 | } |
576 | else |
577 | { |
578 | if (!expr_invariant_in_loop_p (loop_outermost (loop), step)) |
579 | { |
580 | if (dump_file && (dump_flags & TDF_DETAILS)) |
581 | { |
582 | fprintf (stream: dump_file, format: "Memory expression %p\n" ,(void *) ref ); |
583 | print_generic_expr (dump_file, ref, TDF_SLIM); |
584 | fprintf (stream: dump_file,format: ":" ); |
585 | dump_mem_details (file: dump_file, base, step, delta, write_p); |
586 | fprintf (stream: dump_file, |
587 | format: "Not prefetching, ignoring %p due to " |
588 | "loop variant step\n" , |
589 | (void *) ref); |
590 | } |
591 | return false; |
592 | } |
593 | } |
594 | } |
595 | |
596 | /* Now we know that REF = &BASE + STEP * iter + DELTA, where DELTA and STEP |
597 | are integer constants. */ |
598 | agrp = find_or_create_group (groups: refs, base, step); |
599 | record_ref (group: agrp, stmt, mem: ref, delta, write_p); |
600 | |
601 | return true; |
602 | } |
603 | |
604 | /* Record the suitable memory references in LOOP. NO_OTHER_REFS is set to |
605 | true if there are no other memory references inside the loop. */ |
606 | |
607 | static struct mem_ref_group * |
608 | gather_memory_references (class loop *loop, bool *no_other_refs, unsigned *ref_count) |
609 | { |
610 | basic_block *body = get_loop_body_in_dom_order (loop); |
611 | basic_block bb; |
612 | unsigned i; |
613 | gimple_stmt_iterator bsi; |
614 | gimple *stmt; |
615 | tree lhs, rhs; |
616 | struct mem_ref_group *refs = NULL; |
617 | |
618 | *no_other_refs = true; |
619 | *ref_count = 0; |
620 | |
621 | /* Scan the loop body in order, so that the former references precede the |
622 | later ones. */ |
623 | for (i = 0; i < loop->num_nodes; i++) |
624 | { |
625 | bb = body[i]; |
626 | if (bb->loop_father != loop) |
627 | continue; |
628 | |
629 | for (bsi = gsi_start_bb (bb); !gsi_end_p (i: bsi); gsi_next (i: &bsi)) |
630 | { |
631 | stmt = gsi_stmt (i: bsi); |
632 | |
633 | if (gimple_code (g: stmt) != GIMPLE_ASSIGN) |
634 | { |
635 | if (gimple_vuse (g: stmt) |
636 | || (is_gimple_call (gs: stmt) |
637 | && !(gimple_call_flags (stmt) & ECF_CONST))) |
638 | *no_other_refs = false; |
639 | continue; |
640 | } |
641 | |
642 | if (! gimple_vuse (g: stmt)) |
643 | continue; |
644 | |
645 | lhs = gimple_assign_lhs (gs: stmt); |
646 | rhs = gimple_assign_rhs1 (gs: stmt); |
647 | |
648 | if (REFERENCE_CLASS_P (rhs)) |
649 | { |
650 | *no_other_refs &= gather_memory_references_ref (loop, refs: &refs, |
651 | ref: rhs, write_p: false, stmt); |
652 | *ref_count += 1; |
653 | } |
654 | if (REFERENCE_CLASS_P (lhs)) |
655 | { |
656 | *no_other_refs &= gather_memory_references_ref (loop, refs: &refs, |
657 | ref: lhs, write_p: true, stmt); |
658 | *ref_count += 1; |
659 | } |
660 | } |
661 | } |
662 | free (ptr: body); |
663 | |
664 | return refs; |
665 | } |
666 | |
667 | /* Prune the prefetch candidate REF using the self-reuse. */ |
668 | |
669 | static void |
670 | prune_ref_by_self_reuse (struct mem_ref *ref) |
671 | { |
672 | HOST_WIDE_INT step; |
673 | bool backward; |
674 | |
675 | /* If the step size is non constant, we cannot calculate prefetch_mod. */ |
676 | if (!cst_and_fits_in_hwi (ref->group->step)) |
677 | return; |
678 | |
679 | step = int_cst_value (ref->group->step); |
680 | |
681 | backward = step < 0; |
682 | |
683 | if (step == 0) |
684 | { |
685 | /* Prefetch references to invariant address just once. */ |
686 | ref->prefetch_before = 1; |
687 | return; |
688 | } |
689 | |
690 | if (backward) |
691 | step = -step; |
692 | |
693 | if (step > PREFETCH_BLOCK) |
694 | return; |
695 | |
696 | if ((backward && HAVE_BACKWARD_PREFETCH) |
697 | || (!backward && HAVE_FORWARD_PREFETCH)) |
698 | { |
699 | ref->prefetch_before = 1; |
700 | return; |
701 | } |
702 | |
703 | ref->prefetch_mod = PREFETCH_BLOCK / step; |
704 | } |
705 | |
706 | /* Divides X by BY, rounding down. */ |
707 | |
708 | static HOST_WIDE_INT |
709 | ddown (HOST_WIDE_INT x, unsigned HOST_WIDE_INT by) |
710 | { |
711 | gcc_assert (by > 0); |
712 | |
713 | if (x >= 0) |
714 | return x / (HOST_WIDE_INT) by; |
715 | else |
716 | return (x + (HOST_WIDE_INT) by - 1) / (HOST_WIDE_INT) by; |
717 | } |
718 | |
719 | /* Given a CACHE_LINE_SIZE and two inductive memory references |
720 | with a common STEP greater than CACHE_LINE_SIZE and an address |
721 | difference DELTA, compute the probability that they will fall |
722 | in different cache lines. Return true if the computed miss rate |
723 | is not greater than the ACCEPTABLE_MISS_RATE. DISTINCT_ITERS is the |
724 | number of distinct iterations after which the pattern repeats itself. |
725 | ALIGN_UNIT is the unit of alignment in bytes. */ |
726 | |
727 | static bool |
728 | is_miss_rate_acceptable (unsigned HOST_WIDE_INT cache_line_size, |
729 | HOST_WIDE_INT step, HOST_WIDE_INT delta, |
730 | unsigned HOST_WIDE_INT distinct_iters, |
731 | int align_unit) |
732 | { |
733 | unsigned align, iter; |
734 | int total_positions, miss_positions, max_allowed_miss_positions; |
735 | int address1, address2, cache_line1, cache_line2; |
736 | |
737 | /* It always misses if delta is greater than or equal to the cache |
738 | line size. */ |
739 | if (delta >= (HOST_WIDE_INT) cache_line_size) |
740 | return false; |
741 | |
742 | miss_positions = 0; |
743 | total_positions = (cache_line_size / align_unit) * distinct_iters; |
744 | max_allowed_miss_positions = (ACCEPTABLE_MISS_RATE * total_positions) / 1000; |
745 | |
746 | /* Iterate through all possible alignments of the first |
747 | memory reference within its cache line. */ |
748 | for (align = 0; align < cache_line_size; align += align_unit) |
749 | |
750 | /* Iterate through all distinct iterations. */ |
751 | for (iter = 0; iter < distinct_iters; iter++) |
752 | { |
753 | address1 = align + step * iter; |
754 | address2 = address1 + delta; |
755 | cache_line1 = address1 / cache_line_size; |
756 | cache_line2 = address2 / cache_line_size; |
757 | if (cache_line1 != cache_line2) |
758 | { |
759 | miss_positions += 1; |
760 | if (miss_positions > max_allowed_miss_positions) |
761 | return false; |
762 | } |
763 | } |
764 | return true; |
765 | } |
766 | |
767 | /* Prune the prefetch candidate REF using the reuse with BY. |
768 | If BY_IS_BEFORE is true, BY is before REF in the loop. */ |
769 | |
770 | static void |
771 | prune_ref_by_group_reuse (struct mem_ref *ref, struct mem_ref *by, |
772 | bool by_is_before) |
773 | { |
774 | HOST_WIDE_INT step; |
775 | bool backward; |
776 | HOST_WIDE_INT delta_r = ref->delta, delta_b = by->delta; |
777 | HOST_WIDE_INT delta = delta_b - delta_r; |
778 | HOST_WIDE_INT hit_from; |
779 | unsigned HOST_WIDE_INT prefetch_before, prefetch_block; |
780 | HOST_WIDE_INT reduced_step; |
781 | unsigned HOST_WIDE_INT reduced_prefetch_block; |
782 | tree ref_type; |
783 | int align_unit; |
784 | |
785 | /* If the step is non constant we cannot calculate prefetch_before. */ |
786 | if (!cst_and_fits_in_hwi (ref->group->step)) { |
787 | return; |
788 | } |
789 | |
790 | step = int_cst_value (ref->group->step); |
791 | |
792 | backward = step < 0; |
793 | |
794 | |
795 | if (delta == 0) |
796 | { |
797 | /* If the references has the same address, only prefetch the |
798 | former. */ |
799 | if (by_is_before) |
800 | ref->prefetch_before = 0; |
801 | |
802 | return; |
803 | } |
804 | |
805 | if (!step) |
806 | { |
807 | /* If the reference addresses are invariant and fall into the |
808 | same cache line, prefetch just the first one. */ |
809 | if (!by_is_before) |
810 | return; |
811 | |
812 | if (ddown (x: ref->delta, PREFETCH_BLOCK) |
813 | != ddown (x: by->delta, PREFETCH_BLOCK)) |
814 | return; |
815 | |
816 | ref->prefetch_before = 0; |
817 | return; |
818 | } |
819 | |
820 | /* Only prune the reference that is behind in the array. */ |
821 | if (backward) |
822 | { |
823 | if (delta > 0) |
824 | return; |
825 | |
826 | /* Transform the data so that we may assume that the accesses |
827 | are forward. */ |
828 | delta = - delta; |
829 | step = -step; |
830 | delta_r = PREFETCH_BLOCK - 1 - delta_r; |
831 | delta_b = PREFETCH_BLOCK - 1 - delta_b; |
832 | } |
833 | else |
834 | { |
835 | if (delta < 0) |
836 | return; |
837 | } |
838 | |
839 | /* Check whether the two references are likely to hit the same cache |
840 | line, and how distant the iterations in that it occurs are from |
841 | each other. */ |
842 | |
843 | if (step <= PREFETCH_BLOCK) |
844 | { |
845 | /* The accesses are sure to meet. Let us check when. */ |
846 | hit_from = ddown (x: delta_b, PREFETCH_BLOCK) * PREFETCH_BLOCK; |
847 | prefetch_before = (hit_from - delta_r + step - 1) / step; |
848 | |
849 | /* Do not reduce prefetch_before if we meet beyond cache size. */ |
850 | if (prefetch_before > absu_hwi (L2_CACHE_SIZE_BYTES / step)) |
851 | prefetch_before = PREFETCH_ALL; |
852 | if (prefetch_before < ref->prefetch_before) |
853 | ref->prefetch_before = prefetch_before; |
854 | |
855 | return; |
856 | } |
857 | |
858 | /* A more complicated case with step > prefetch_block. First reduce |
859 | the ratio between the step and the cache line size to its simplest |
860 | terms. The resulting denominator will then represent the number of |
861 | distinct iterations after which each address will go back to its |
862 | initial location within the cache line. This computation assumes |
863 | that PREFETCH_BLOCK is a power of two. */ |
864 | prefetch_block = PREFETCH_BLOCK; |
865 | reduced_prefetch_block = prefetch_block; |
866 | reduced_step = step; |
867 | while ((reduced_step & 1) == 0 |
868 | && reduced_prefetch_block > 1) |
869 | { |
870 | reduced_step >>= 1; |
871 | reduced_prefetch_block >>= 1; |
872 | } |
873 | |
874 | prefetch_before = delta / step; |
875 | delta %= step; |
876 | ref_type = TREE_TYPE (ref->mem); |
877 | align_unit = TYPE_ALIGN (ref_type) / 8; |
878 | if (is_miss_rate_acceptable (cache_line_size: prefetch_block, step, delta, |
879 | distinct_iters: reduced_prefetch_block, align_unit)) |
880 | { |
881 | /* Do not reduce prefetch_before if we meet beyond cache size. */ |
882 | if (prefetch_before > L2_CACHE_SIZE_BYTES / PREFETCH_BLOCK) |
883 | prefetch_before = PREFETCH_ALL; |
884 | if (prefetch_before < ref->prefetch_before) |
885 | ref->prefetch_before = prefetch_before; |
886 | |
887 | return; |
888 | } |
889 | |
890 | /* Try also the following iteration. */ |
891 | prefetch_before++; |
892 | delta = step - delta; |
893 | if (is_miss_rate_acceptable (cache_line_size: prefetch_block, step, delta, |
894 | distinct_iters: reduced_prefetch_block, align_unit)) |
895 | { |
896 | if (prefetch_before < ref->prefetch_before) |
897 | ref->prefetch_before = prefetch_before; |
898 | |
899 | return; |
900 | } |
901 | |
902 | /* The ref probably does not reuse by. */ |
903 | return; |
904 | } |
905 | |
906 | /* Prune the prefetch candidate REF using the reuses with other references |
907 | in REFS. */ |
908 | |
909 | static void |
910 | prune_ref_by_reuse (struct mem_ref *ref, struct mem_ref *refs) |
911 | { |
912 | struct mem_ref *prune_by; |
913 | bool before = true; |
914 | |
915 | prune_ref_by_self_reuse (ref); |
916 | |
917 | for (prune_by = refs; prune_by; prune_by = prune_by->next) |
918 | { |
919 | if (prune_by == ref) |
920 | { |
921 | before = false; |
922 | continue; |
923 | } |
924 | |
925 | if (!WRITE_CAN_USE_READ_PREFETCH |
926 | && ref->write_p |
927 | && !prune_by->write_p) |
928 | continue; |
929 | if (!READ_CAN_USE_WRITE_PREFETCH |
930 | && !ref->write_p |
931 | && prune_by->write_p) |
932 | continue; |
933 | |
934 | prune_ref_by_group_reuse (ref, by: prune_by, by_is_before: before); |
935 | } |
936 | } |
937 | |
938 | /* Prune the prefetch candidates in GROUP using the reuse analysis. */ |
939 | |
940 | static void |
941 | prune_group_by_reuse (struct mem_ref_group *group) |
942 | { |
943 | struct mem_ref *ref_pruned; |
944 | |
945 | for (ref_pruned = group->refs; ref_pruned; ref_pruned = ref_pruned->next) |
946 | { |
947 | prune_ref_by_reuse (ref: ref_pruned, refs: group->refs); |
948 | |
949 | if (dump_file && (dump_flags & TDF_DETAILS)) |
950 | { |
951 | dump_mem_ref (file: dump_file, ref: ref_pruned); |
952 | |
953 | if (ref_pruned->prefetch_before == PREFETCH_ALL |
954 | && ref_pruned->prefetch_mod == 1) |
955 | fprintf (stream: dump_file, format: " no restrictions" ); |
956 | else if (ref_pruned->prefetch_before == 0) |
957 | fprintf (stream: dump_file, format: " do not prefetch" ); |
958 | else if (ref_pruned->prefetch_before <= ref_pruned->prefetch_mod) |
959 | fprintf (stream: dump_file, format: " prefetch once" ); |
960 | else |
961 | { |
962 | if (ref_pruned->prefetch_before != PREFETCH_ALL) |
963 | { |
964 | fprintf (stream: dump_file, format: " prefetch before " ); |
965 | fprintf (stream: dump_file, HOST_WIDE_INT_PRINT_DEC, |
966 | ref_pruned->prefetch_before); |
967 | } |
968 | if (ref_pruned->prefetch_mod != 1) |
969 | { |
970 | fprintf (stream: dump_file, format: " prefetch mod " ); |
971 | fprintf (stream: dump_file, HOST_WIDE_INT_PRINT_DEC, |
972 | ref_pruned->prefetch_mod); |
973 | } |
974 | } |
975 | fprintf (stream: dump_file, format: "\n" ); |
976 | } |
977 | } |
978 | } |
979 | |
980 | /* Prune the list of prefetch candidates GROUPS using the reuse analysis. */ |
981 | |
982 | static void |
983 | prune_by_reuse (struct mem_ref_group *groups) |
984 | { |
985 | for (; groups; groups = groups->next) |
986 | prune_group_by_reuse (group: groups); |
987 | } |
988 | |
989 | /* Returns true if we should issue prefetch for REF. */ |
990 | |
991 | static bool |
992 | should_issue_prefetch_p (struct mem_ref *ref) |
993 | { |
994 | /* Do we want to issue prefetches for non-constant strides? */ |
995 | if (!cst_and_fits_in_hwi (ref->group->step) |
996 | && param_prefetch_dynamic_strides == 0) |
997 | { |
998 | if (dump_file && (dump_flags & TDF_DETAILS)) |
999 | fprintf (stream: dump_file, |
1000 | format: "Skipping non-constant step for reference %u:%u\n" , |
1001 | ref->group->uid, ref->uid); |
1002 | return false; |
1003 | } |
1004 | |
1005 | /* Some processors may have a hardware prefetcher that may conflict with |
1006 | prefetch hints for a range of strides. Make sure we don't issue |
1007 | prefetches for such cases if the stride is within this particular |
1008 | range. */ |
1009 | if (cst_and_fits_in_hwi (ref->group->step) |
1010 | && abs_hwi (x: int_cst_value (ref->group->step)) |
1011 | < (HOST_WIDE_INT) param_prefetch_minimum_stride) |
1012 | { |
1013 | if (dump_file && (dump_flags & TDF_DETAILS)) |
1014 | fprintf (stream: dump_file, |
1015 | format: "Step for reference %u:%u (" HOST_WIDE_INT_PRINT_DEC |
1016 | ") is less than the mininum required stride of %d\n" , |
1017 | ref->group->uid, ref->uid, int_cst_value (ref->group->step), |
1018 | param_prefetch_minimum_stride); |
1019 | return false; |
1020 | } |
1021 | |
1022 | /* For now do not issue prefetches for only first few of the |
1023 | iterations. */ |
1024 | if (ref->prefetch_before != PREFETCH_ALL) |
1025 | { |
1026 | if (dump_file && (dump_flags & TDF_DETAILS)) |
1027 | fprintf (stream: dump_file, format: "Ignoring reference %u:%u due to prefetch_before\n" , |
1028 | ref->group->uid, ref->uid); |
1029 | return false; |
1030 | } |
1031 | |
1032 | /* Do not prefetch nontemporal stores. */ |
1033 | if (ref->storent_p) |
1034 | { |
1035 | if (dump_file && (dump_flags & TDF_DETAILS)) |
1036 | fprintf (stream: dump_file, format: "Ignoring nontemporal store reference %u:%u\n" , ref->group->uid, ref->uid); |
1037 | return false; |
1038 | } |
1039 | |
1040 | return true; |
1041 | } |
1042 | |
1043 | /* Decide which of the prefetch candidates in GROUPS to prefetch. |
1044 | AHEAD is the number of iterations to prefetch ahead (which corresponds |
1045 | to the number of simultaneous instances of one prefetch running at a |
1046 | time). UNROLL_FACTOR is the factor by that the loop is going to be |
1047 | unrolled. Returns true if there is anything to prefetch. */ |
1048 | |
1049 | static bool |
1050 | schedule_prefetches (struct mem_ref_group *groups, unsigned unroll_factor, |
1051 | unsigned ahead) |
1052 | { |
1053 | unsigned remaining_prefetch_slots, n_prefetches, prefetch_slots; |
1054 | unsigned slots_per_prefetch; |
1055 | struct mem_ref *ref; |
1056 | bool any = false; |
1057 | |
1058 | /* At most param_simultaneous_prefetches should be running |
1059 | at the same time. */ |
1060 | remaining_prefetch_slots = param_simultaneous_prefetches; |
1061 | |
1062 | /* The prefetch will run for AHEAD iterations of the original loop, i.e., |
1063 | AHEAD / UNROLL_FACTOR iterations of the unrolled loop. In each iteration, |
1064 | it will need a prefetch slot. */ |
1065 | slots_per_prefetch = (ahead + unroll_factor / 2) / unroll_factor; |
1066 | if (dump_file && (dump_flags & TDF_DETAILS)) |
1067 | fprintf (stream: dump_file, format: "Each prefetch instruction takes %u prefetch slots.\n" , |
1068 | slots_per_prefetch); |
1069 | |
1070 | /* For now we just take memory references one by one and issue |
1071 | prefetches for as many as possible. The groups are sorted |
1072 | starting with the largest step, since the references with |
1073 | large step are more likely to cause many cache misses. */ |
1074 | |
1075 | for (; groups; groups = groups->next) |
1076 | for (ref = groups->refs; ref; ref = ref->next) |
1077 | { |
1078 | if (!should_issue_prefetch_p (ref)) |
1079 | continue; |
1080 | |
1081 | /* The loop is far from being sufficiently unrolled for this |
1082 | prefetch. Do not generate prefetch to avoid many redudant |
1083 | prefetches. */ |
1084 | if (ref->prefetch_mod / unroll_factor > PREFETCH_MOD_TO_UNROLL_FACTOR_RATIO) |
1085 | continue; |
1086 | |
1087 | /* If we need to prefetch the reference each PREFETCH_MOD iterations, |
1088 | and we unroll the loop UNROLL_FACTOR times, we need to insert |
1089 | ceil (UNROLL_FACTOR / PREFETCH_MOD) instructions in each |
1090 | iteration. */ |
1091 | n_prefetches = ((unroll_factor + ref->prefetch_mod - 1) |
1092 | / ref->prefetch_mod); |
1093 | prefetch_slots = n_prefetches * slots_per_prefetch; |
1094 | |
1095 | /* If more than half of the prefetches would be lost anyway, do not |
1096 | issue the prefetch. */ |
1097 | if (2 * remaining_prefetch_slots < prefetch_slots) |
1098 | continue; |
1099 | |
1100 | /* Stop prefetching if debug counter is activated. */ |
1101 | if (!dbg_cnt (index: prefetch)) |
1102 | continue; |
1103 | |
1104 | ref->issue_prefetch_p = true; |
1105 | if (dump_file && (dump_flags & TDF_DETAILS)) |
1106 | fprintf (stream: dump_file, format: "Decided to issue prefetch for reference %u:%u\n" , |
1107 | ref->group->uid, ref->uid); |
1108 | |
1109 | if (remaining_prefetch_slots <= prefetch_slots) |
1110 | return true; |
1111 | remaining_prefetch_slots -= prefetch_slots; |
1112 | any = true; |
1113 | } |
1114 | |
1115 | return any; |
1116 | } |
1117 | |
1118 | /* Return TRUE if no prefetch is going to be generated in the given |
1119 | GROUPS. */ |
1120 | |
1121 | static bool |
1122 | nothing_to_prefetch_p (struct mem_ref_group *groups) |
1123 | { |
1124 | struct mem_ref *ref; |
1125 | |
1126 | for (; groups; groups = groups->next) |
1127 | for (ref = groups->refs; ref; ref = ref->next) |
1128 | if (should_issue_prefetch_p (ref)) |
1129 | return false; |
1130 | |
1131 | return true; |
1132 | } |
1133 | |
1134 | /* Estimate the number of prefetches in the given GROUPS. |
1135 | UNROLL_FACTOR is the factor by which LOOP was unrolled. */ |
1136 | |
1137 | static int |
1138 | estimate_prefetch_count (struct mem_ref_group *groups, unsigned unroll_factor) |
1139 | { |
1140 | struct mem_ref *ref; |
1141 | unsigned n_prefetches; |
1142 | int prefetch_count = 0; |
1143 | |
1144 | for (; groups; groups = groups->next) |
1145 | for (ref = groups->refs; ref; ref = ref->next) |
1146 | if (should_issue_prefetch_p (ref)) |
1147 | { |
1148 | n_prefetches = ((unroll_factor + ref->prefetch_mod - 1) |
1149 | / ref->prefetch_mod); |
1150 | prefetch_count += n_prefetches; |
1151 | } |
1152 | |
1153 | return prefetch_count; |
1154 | } |
1155 | |
1156 | /* Issue prefetches for the reference REF into loop as decided before. |
1157 | HEAD is the number of iterations to prefetch ahead. UNROLL_FACTOR |
1158 | is the factor by which LOOP was unrolled. */ |
1159 | |
1160 | static void |
1161 | issue_prefetch_ref (struct mem_ref *ref, unsigned unroll_factor, unsigned ahead) |
1162 | { |
1163 | HOST_WIDE_INT delta; |
1164 | tree addr, addr_base, write_p, local, forward; |
1165 | gcall *prefetch; |
1166 | gimple_stmt_iterator bsi; |
1167 | unsigned n_prefetches, ap; |
1168 | bool nontemporal = ref->reuse_distance >= L2_CACHE_SIZE_BYTES; |
1169 | |
1170 | if (dump_file && (dump_flags & TDF_DETAILS)) |
1171 | fprintf (stream: dump_file, format: "Issued%s prefetch for reference %u:%u.\n" , |
1172 | nontemporal ? " nontemporal" : "" , |
1173 | ref->group->uid, ref->uid); |
1174 | |
1175 | bsi = gsi_for_stmt (ref->stmt); |
1176 | |
1177 | n_prefetches = ((unroll_factor + ref->prefetch_mod - 1) |
1178 | / ref->prefetch_mod); |
1179 | addr_base = build_fold_addr_expr_with_type (ref->mem, ptr_type_node); |
1180 | addr_base = force_gimple_operand_gsi (&bsi, unshare_expr (addr_base), |
1181 | true, NULL, true, GSI_SAME_STMT); |
1182 | write_p = ref->write_p ? integer_one_node : integer_zero_node; |
1183 | local = nontemporal ? integer_zero_node : integer_three_node; |
1184 | |
1185 | for (ap = 0; ap < n_prefetches; ap++) |
1186 | { |
1187 | if (cst_and_fits_in_hwi (ref->group->step)) |
1188 | { |
1189 | /* Determine the address to prefetch. */ |
1190 | delta = (ahead + ap * ref->prefetch_mod) * |
1191 | int_cst_value (ref->group->step); |
1192 | addr = fold_build_pointer_plus_hwi (addr_base, delta); |
1193 | addr = force_gimple_operand_gsi (&bsi, unshare_expr (addr), true, |
1194 | NULL, true, GSI_SAME_STMT); |
1195 | } |
1196 | else |
1197 | { |
1198 | /* The step size is non-constant but loop-invariant. We use the |
1199 | heuristic to simply prefetch ahead iterations ahead. */ |
1200 | forward = fold_build2 (MULT_EXPR, sizetype, |
1201 | fold_convert (sizetype, ref->group->step), |
1202 | fold_convert (sizetype, size_int (ahead))); |
1203 | addr = fold_build_pointer_plus (addr_base, forward); |
1204 | addr = force_gimple_operand_gsi (&bsi, unshare_expr (addr), true, |
1205 | NULL, true, GSI_SAME_STMT); |
1206 | } |
1207 | |
1208 | if (addr_base != addr |
1209 | && TREE_CODE (addr_base) == SSA_NAME |
1210 | && TREE_CODE (addr) == SSA_NAME) |
1211 | { |
1212 | duplicate_ssa_name_ptr_info (addr, SSA_NAME_PTR_INFO (addr_base)); |
1213 | /* As this isn't a plain copy we have to reset alignment |
1214 | information. */ |
1215 | if (SSA_NAME_PTR_INFO (addr)) |
1216 | mark_ptr_info_alignment_unknown (SSA_NAME_PTR_INFO (addr)); |
1217 | } |
1218 | |
1219 | /* Create the prefetch instruction. */ |
1220 | prefetch = gimple_build_call (builtin_decl_explicit (fncode: BUILT_IN_PREFETCH), |
1221 | 3, addr, write_p, local); |
1222 | gsi_insert_before (&bsi, prefetch, GSI_SAME_STMT); |
1223 | } |
1224 | } |
1225 | |
1226 | /* Issue prefetches for the references in GROUPS into loop as decided before. |
1227 | HEAD is the number of iterations to prefetch ahead. UNROLL_FACTOR is the |
1228 | factor by that LOOP was unrolled. */ |
1229 | |
1230 | static void |
1231 | issue_prefetches (struct mem_ref_group *groups, |
1232 | unsigned unroll_factor, unsigned ahead) |
1233 | { |
1234 | struct mem_ref *ref; |
1235 | |
1236 | for (; groups; groups = groups->next) |
1237 | for (ref = groups->refs; ref; ref = ref->next) |
1238 | if (ref->issue_prefetch_p) |
1239 | issue_prefetch_ref (ref, unroll_factor, ahead); |
1240 | } |
1241 | |
1242 | /* Returns true if REF is a memory write for that a nontemporal store insn |
1243 | can be used. */ |
1244 | |
1245 | static bool |
1246 | nontemporal_store_p (struct mem_ref *ref) |
1247 | { |
1248 | machine_mode mode; |
1249 | enum insn_code code; |
1250 | |
1251 | /* REF must be a write that is not reused. We require it to be independent |
1252 | on all other memory references in the loop, as the nontemporal stores may |
1253 | be reordered with respect to other memory references. */ |
1254 | if (!ref->write_p |
1255 | || !ref->independent_p |
1256 | || ref->reuse_distance < L2_CACHE_SIZE_BYTES) |
1257 | return false; |
1258 | |
1259 | /* Check that we have the storent instruction for the mode. */ |
1260 | mode = TYPE_MODE (TREE_TYPE (ref->mem)); |
1261 | if (mode == BLKmode) |
1262 | return false; |
1263 | |
1264 | code = optab_handler (op: storent_optab, mode); |
1265 | return code != CODE_FOR_nothing; |
1266 | } |
1267 | |
1268 | /* If REF is a nontemporal store, we mark the corresponding modify statement |
1269 | and return true. Otherwise, we return false. */ |
1270 | |
1271 | static bool |
1272 | mark_nontemporal_store (struct mem_ref *ref) |
1273 | { |
1274 | if (!nontemporal_store_p (ref)) |
1275 | return false; |
1276 | |
1277 | if (dump_file && (dump_flags & TDF_DETAILS)) |
1278 | fprintf (stream: dump_file, format: "Marked reference %u:%u as a nontemporal store.\n" , |
1279 | ref->group->uid, ref->uid); |
1280 | |
1281 | gimple_assign_set_nontemporal_move (gs: ref->stmt, nontemporal: true); |
1282 | ref->storent_p = true; |
1283 | |
1284 | return true; |
1285 | } |
1286 | |
1287 | /* Issue a memory fence instruction after LOOP. */ |
1288 | |
1289 | static void |
1290 | emit_mfence_after_loop (class loop *loop) |
1291 | { |
1292 | auto_vec<edge> exits = get_loop_exit_edges (loop); |
1293 | edge exit; |
1294 | gcall *call; |
1295 | gimple_stmt_iterator bsi; |
1296 | unsigned i; |
1297 | |
1298 | FOR_EACH_VEC_ELT (exits, i, exit) |
1299 | { |
1300 | call = gimple_build_call (FENCE_FOLLOWING_MOVNT, 0); |
1301 | |
1302 | if (!single_pred_p (bb: exit->dest) |
1303 | /* If possible, we prefer not to insert the fence on other paths |
1304 | in cfg. */ |
1305 | && !(exit->flags & EDGE_ABNORMAL)) |
1306 | split_loop_exit_edge (exit); |
1307 | bsi = gsi_after_labels (bb: exit->dest); |
1308 | |
1309 | gsi_insert_before (&bsi, call, GSI_NEW_STMT); |
1310 | } |
1311 | } |
1312 | |
1313 | /* Returns true if we can use storent in loop, false otherwise. */ |
1314 | |
1315 | static bool |
1316 | may_use_storent_in_loop_p (class loop *loop) |
1317 | { |
1318 | bool ret = true; |
1319 | |
1320 | if (loop->inner != NULL) |
1321 | return false; |
1322 | |
1323 | /* If we must issue a mfence insn after using storent, check that there |
1324 | is a suitable place for it at each of the loop exits. */ |
1325 | if (FENCE_FOLLOWING_MOVNT != NULL_TREE) |
1326 | { |
1327 | auto_vec<edge> exits = get_loop_exit_edges (loop); |
1328 | unsigned i; |
1329 | edge exit; |
1330 | |
1331 | FOR_EACH_VEC_ELT (exits, i, exit) |
1332 | if ((exit->flags & EDGE_ABNORMAL) |
1333 | && exit->dest == EXIT_BLOCK_PTR_FOR_FN (cfun)) |
1334 | ret = false; |
1335 | } |
1336 | |
1337 | return ret; |
1338 | } |
1339 | |
1340 | /* Marks nontemporal stores in LOOP. GROUPS contains the description of memory |
1341 | references in the loop. Returns whether we inserted any mfence call. */ |
1342 | |
1343 | static bool |
1344 | mark_nontemporal_stores (class loop *loop, struct mem_ref_group *groups) |
1345 | { |
1346 | struct mem_ref *ref; |
1347 | bool any = false; |
1348 | |
1349 | if (!may_use_storent_in_loop_p (loop)) |
1350 | return false; |
1351 | |
1352 | for (; groups; groups = groups->next) |
1353 | for (ref = groups->refs; ref; ref = ref->next) |
1354 | any |= mark_nontemporal_store (ref); |
1355 | |
1356 | if (any && FENCE_FOLLOWING_MOVNT != NULL_TREE) |
1357 | { |
1358 | emit_mfence_after_loop (loop); |
1359 | return true; |
1360 | } |
1361 | return false; |
1362 | } |
1363 | |
1364 | /* Determines whether we can profitably unroll LOOP FACTOR times, and if |
1365 | this is the case, fill in DESC by the description of number of |
1366 | iterations. */ |
1367 | |
1368 | static bool |
1369 | should_unroll_loop_p (class loop *loop, class tree_niter_desc *desc, |
1370 | unsigned factor) |
1371 | { |
1372 | if (!can_unroll_loop_p (loop, factor, niter: desc)) |
1373 | return false; |
1374 | |
1375 | /* We only consider loops without control flow for unrolling. This is not |
1376 | a hard restriction -- tree_unroll_loop works with arbitrary loops |
1377 | as well; but the unrolling/prefetching is usually more profitable for |
1378 | loops consisting of a single basic block, and we want to limit the |
1379 | code growth. */ |
1380 | if (loop->num_nodes > 2) |
1381 | return false; |
1382 | |
1383 | return true; |
1384 | } |
1385 | |
1386 | /* Determine the coefficient by that unroll LOOP, from the information |
1387 | contained in the list of memory references REFS. Description of |
1388 | number of iterations of LOOP is stored to DESC. NINSNS is the number of |
1389 | insns of the LOOP. EST_NITER is the estimated number of iterations of |
1390 | the loop, or -1 if no estimate is available. */ |
1391 | |
1392 | static unsigned |
1393 | determine_unroll_factor (class loop *loop, struct mem_ref_group *refs, |
1394 | unsigned ninsns, class tree_niter_desc *desc, |
1395 | HOST_WIDE_INT est_niter) |
1396 | { |
1397 | unsigned upper_bound; |
1398 | unsigned nfactor, factor, mod_constraint; |
1399 | struct mem_ref_group *agp; |
1400 | struct mem_ref *ref; |
1401 | |
1402 | /* First check whether the loop is not too large to unroll. We ignore |
1403 | PARAM_MAX_UNROLL_TIMES, because for small loops, it prevented us |
1404 | from unrolling them enough to make exactly one cache line covered by each |
1405 | iteration. Also, the goal of PARAM_MAX_UNROLL_TIMES is to prevent |
1406 | us from unrolling the loops too many times in cases where we only expect |
1407 | gains from better scheduling and decreasing loop overhead, which is not |
1408 | the case here. */ |
1409 | upper_bound = param_max_unrolled_insns / ninsns; |
1410 | |
1411 | /* If we unrolled the loop more times than it iterates, the unrolled version |
1412 | of the loop would be never entered. */ |
1413 | if (est_niter >= 0 && est_niter < (HOST_WIDE_INT) upper_bound) |
1414 | upper_bound = est_niter; |
1415 | |
1416 | if (upper_bound <= 1) |
1417 | return 1; |
1418 | |
1419 | /* Choose the factor so that we may prefetch each cache just once, |
1420 | but bound the unrolling by UPPER_BOUND. */ |
1421 | factor = 1; |
1422 | for (agp = refs; agp; agp = agp->next) |
1423 | for (ref = agp->refs; ref; ref = ref->next) |
1424 | if (should_issue_prefetch_p (ref)) |
1425 | { |
1426 | mod_constraint = ref->prefetch_mod; |
1427 | nfactor = least_common_multiple (mod_constraint, factor); |
1428 | if (nfactor <= upper_bound) |
1429 | factor = nfactor; |
1430 | } |
1431 | |
1432 | if (!should_unroll_loop_p (loop, desc, factor)) |
1433 | return 1; |
1434 | |
1435 | return factor; |
1436 | } |
1437 | |
1438 | /* Returns the total volume of the memory references REFS, taking into account |
1439 | reuses in the innermost loop and cache line size. TODO -- we should also |
1440 | take into account reuses across the iterations of the loops in the loop |
1441 | nest. */ |
1442 | |
1443 | static unsigned |
1444 | volume_of_references (struct mem_ref_group *refs) |
1445 | { |
1446 | unsigned volume = 0; |
1447 | struct mem_ref_group *gr; |
1448 | struct mem_ref *ref; |
1449 | |
1450 | for (gr = refs; gr; gr = gr->next) |
1451 | for (ref = gr->refs; ref; ref = ref->next) |
1452 | { |
1453 | /* Almost always reuses another value? */ |
1454 | if (ref->prefetch_before != PREFETCH_ALL) |
1455 | continue; |
1456 | |
1457 | /* If several iterations access the same cache line, use the size of |
1458 | the line divided by this number. Otherwise, a cache line is |
1459 | accessed in each iteration. TODO -- in the latter case, we should |
1460 | take the size of the reference into account, rounding it up on cache |
1461 | line size multiple. */ |
1462 | volume += param_l1_cache_line_size / ref->prefetch_mod; |
1463 | } |
1464 | return volume; |
1465 | } |
1466 | |
1467 | /* Returns the volume of memory references accessed across VEC iterations of |
1468 | loops, whose sizes are described in the LOOP_SIZES array. N is the number |
1469 | of the loops in the nest (length of VEC and LOOP_SIZES vectors). */ |
1470 | |
1471 | static unsigned |
1472 | volume_of_dist_vector (lambda_vector vec, unsigned *loop_sizes, unsigned n) |
1473 | { |
1474 | unsigned i; |
1475 | |
1476 | for (i = 0; i < n; i++) |
1477 | if (vec[i] != 0) |
1478 | break; |
1479 | |
1480 | if (i == n) |
1481 | return 0; |
1482 | |
1483 | gcc_assert (vec[i] > 0); |
1484 | |
1485 | /* We ignore the parts of the distance vector in subloops, since usually |
1486 | the numbers of iterations are much smaller. */ |
1487 | return loop_sizes[i] * vec[i]; |
1488 | } |
1489 | |
1490 | /* Add the steps of ACCESS_FN multiplied by STRIDE to the array STRIDE |
1491 | at the position corresponding to the loop of the step. N is the depth |
1492 | of the considered loop nest, and, LOOP is its innermost loop. */ |
1493 | |
1494 | static void |
1495 | add_subscript_strides (tree access_fn, unsigned stride, |
1496 | HOST_WIDE_INT *strides, unsigned n, class loop *loop) |
1497 | { |
1498 | class loop *aloop; |
1499 | tree step; |
1500 | HOST_WIDE_INT astep; |
1501 | unsigned min_depth = loop_depth (loop) - n; |
1502 | |
1503 | while (TREE_CODE (access_fn) == POLYNOMIAL_CHREC) |
1504 | { |
1505 | aloop = get_chrec_loop (chrec: access_fn); |
1506 | step = CHREC_RIGHT (access_fn); |
1507 | access_fn = CHREC_LEFT (access_fn); |
1508 | |
1509 | if ((unsigned) loop_depth (loop: aloop) <= min_depth) |
1510 | continue; |
1511 | |
1512 | if (tree_fits_shwi_p (step)) |
1513 | astep = tree_to_shwi (step); |
1514 | else |
1515 | astep = param_l1_cache_line_size; |
1516 | |
1517 | strides[n - 1 - loop_depth (loop) + loop_depth (loop: aloop)] += astep * stride; |
1518 | |
1519 | } |
1520 | } |
1521 | |
1522 | /* Returns the volume of memory references accessed between two consecutive |
1523 | self-reuses of the reference DR. We consider the subscripts of DR in N |
1524 | loops, and LOOP_SIZES contains the volumes of accesses in each of the |
1525 | loops. LOOP is the innermost loop of the current loop nest. */ |
1526 | |
1527 | static unsigned |
1528 | self_reuse_distance (data_reference_p dr, unsigned *loop_sizes, unsigned n, |
1529 | class loop *loop) |
1530 | { |
1531 | tree stride, access_fn; |
1532 | HOST_WIDE_INT *strides, astride; |
1533 | vec<tree> access_fns; |
1534 | tree ref = DR_REF (dr); |
1535 | unsigned i, ret = ~0u; |
1536 | |
1537 | /* In the following example: |
1538 | |
1539 | for (i = 0; i < N; i++) |
1540 | for (j = 0; j < N; j++) |
1541 | use (a[j][i]); |
1542 | the same cache line is accessed each N steps (except if the change from |
1543 | i to i + 1 crosses the boundary of the cache line). Thus, for self-reuse, |
1544 | we cannot rely purely on the results of the data dependence analysis. |
1545 | |
1546 | Instead, we compute the stride of the reference in each loop, and consider |
1547 | the innermost loop in that the stride is less than cache size. */ |
1548 | |
1549 | strides = XCNEWVEC (HOST_WIDE_INT, n); |
1550 | access_fns = DR_ACCESS_FNS (dr); |
1551 | |
1552 | FOR_EACH_VEC_ELT (access_fns, i, access_fn) |
1553 | { |
1554 | /* Keep track of the reference corresponding to the subscript, so that we |
1555 | know its stride. */ |
1556 | while (handled_component_p (t: ref) && TREE_CODE (ref) != ARRAY_REF) |
1557 | ref = TREE_OPERAND (ref, 0); |
1558 | |
1559 | if (TREE_CODE (ref) == ARRAY_REF) |
1560 | { |
1561 | stride = TYPE_SIZE_UNIT (TREE_TYPE (ref)); |
1562 | if (tree_fits_uhwi_p (stride)) |
1563 | astride = tree_to_uhwi (stride); |
1564 | else |
1565 | astride = param_l1_cache_line_size; |
1566 | |
1567 | ref = TREE_OPERAND (ref, 0); |
1568 | } |
1569 | else |
1570 | astride = 1; |
1571 | |
1572 | add_subscript_strides (access_fn, stride: astride, strides, n, loop); |
1573 | } |
1574 | |
1575 | for (i = n; i-- > 0; ) |
1576 | { |
1577 | unsigned HOST_WIDE_INT s; |
1578 | |
1579 | s = strides[i] < 0 ? -strides[i] : strides[i]; |
1580 | |
1581 | if (s < (unsigned) param_l1_cache_line_size |
1582 | && (loop_sizes[i] |
1583 | > (unsigned) (L1_CACHE_SIZE_BYTES / NONTEMPORAL_FRACTION))) |
1584 | { |
1585 | ret = loop_sizes[i]; |
1586 | break; |
1587 | } |
1588 | } |
1589 | |
1590 | free (ptr: strides); |
1591 | return ret; |
1592 | } |
1593 | |
1594 | /* Determines the distance till the first reuse of each reference in REFS |
1595 | in the loop nest of LOOP. NO_OTHER_REFS is true if there are no other |
1596 | memory references in the loop. Return false if the analysis fails. */ |
1597 | |
1598 | static bool |
1599 | determine_loop_nest_reuse (class loop *loop, struct mem_ref_group *refs, |
1600 | bool no_other_refs) |
1601 | { |
1602 | class loop *nest, *aloop; |
1603 | vec<data_reference_p> datarefs = vNULL; |
1604 | vec<ddr_p> dependences = vNULL; |
1605 | struct mem_ref_group *gr; |
1606 | struct mem_ref *ref, *refb; |
1607 | auto_vec<loop_p> vloops; |
1608 | unsigned *loop_data_size; |
1609 | unsigned i, j, n; |
1610 | unsigned volume, dist, adist; |
1611 | HOST_WIDE_INT vol; |
1612 | data_reference_p dr; |
1613 | ddr_p dep; |
1614 | |
1615 | if (loop->inner) |
1616 | return true; |
1617 | |
1618 | /* Find the outermost loop of the loop nest of loop (we require that |
1619 | there are no sibling loops inside the nest). */ |
1620 | nest = loop; |
1621 | while (1) |
1622 | { |
1623 | aloop = loop_outer (loop: nest); |
1624 | |
1625 | if (aloop == current_loops->tree_root |
1626 | || aloop->inner->next) |
1627 | break; |
1628 | |
1629 | nest = aloop; |
1630 | } |
1631 | |
1632 | /* For each loop, determine the amount of data accessed in each iteration. |
1633 | We use this to estimate whether the reference is evicted from the |
1634 | cache before its reuse. */ |
1635 | find_loop_nest (nest, &vloops); |
1636 | n = vloops.length (); |
1637 | loop_data_size = XNEWVEC (unsigned, n); |
1638 | volume = volume_of_references (refs); |
1639 | i = n; |
1640 | while (i-- != 0) |
1641 | { |
1642 | loop_data_size[i] = volume; |
1643 | /* Bound the volume by the L2 cache size, since above this bound, |
1644 | all dependence distances are equivalent. */ |
1645 | if (volume > L2_CACHE_SIZE_BYTES) |
1646 | continue; |
1647 | |
1648 | aloop = vloops[i]; |
1649 | vol = estimated_stmt_executions_int (aloop); |
1650 | if (vol == -1) |
1651 | vol = expected_loop_iterations (aloop); |
1652 | volume *= vol; |
1653 | } |
1654 | |
1655 | /* Prepare the references in the form suitable for data dependence |
1656 | analysis. We ignore unanalyzable data references (the results |
1657 | are used just as a heuristics to estimate temporality of the |
1658 | references, hence we do not need to worry about correctness). */ |
1659 | for (gr = refs; gr; gr = gr->next) |
1660 | for (ref = gr->refs; ref; ref = ref->next) |
1661 | { |
1662 | dr = create_data_ref (loop_preheader_edge (nest), |
1663 | loop_containing_stmt (stmt: ref->stmt), |
1664 | ref->mem, ref->stmt, !ref->write_p, false); |
1665 | |
1666 | if (dr) |
1667 | { |
1668 | ref->reuse_distance = volume; |
1669 | dr->aux = ref; |
1670 | datarefs.safe_push (obj: dr); |
1671 | } |
1672 | else |
1673 | no_other_refs = false; |
1674 | } |
1675 | |
1676 | FOR_EACH_VEC_ELT (datarefs, i, dr) |
1677 | { |
1678 | dist = self_reuse_distance (dr, loop_sizes: loop_data_size, n, loop); |
1679 | ref = (struct mem_ref *) dr->aux; |
1680 | if (ref->reuse_distance > dist) |
1681 | ref->reuse_distance = dist; |
1682 | |
1683 | if (no_other_refs) |
1684 | ref->independent_p = true; |
1685 | } |
1686 | |
1687 | if (!compute_all_dependences (datarefs, &dependences, vloops, true)) |
1688 | return false; |
1689 | |
1690 | FOR_EACH_VEC_ELT (dependences, i, dep) |
1691 | { |
1692 | if (DDR_ARE_DEPENDENT (dep) == chrec_known) |
1693 | continue; |
1694 | |
1695 | ref = (struct mem_ref *) DDR_A (dep)->aux; |
1696 | refb = (struct mem_ref *) DDR_B (dep)->aux; |
1697 | |
1698 | if (DDR_ARE_DEPENDENT (dep) == chrec_dont_know |
1699 | || DDR_COULD_BE_INDEPENDENT_P (dep) |
1700 | || DDR_NUM_DIST_VECTS (dep) == 0) |
1701 | { |
1702 | /* If the dependence cannot be analyzed, assume that there might be |
1703 | a reuse. */ |
1704 | dist = 0; |
1705 | |
1706 | ref->independent_p = false; |
1707 | refb->independent_p = false; |
1708 | } |
1709 | else |
1710 | { |
1711 | /* The distance vectors are normalized to be always lexicographically |
1712 | positive, hence we cannot tell just from them whether DDR_A comes |
1713 | before DDR_B or vice versa. However, it is not important, |
1714 | anyway -- if DDR_A is close to DDR_B, then it is either reused in |
1715 | DDR_B (and it is not nontemporal), or it reuses the value of DDR_B |
1716 | in cache (and marking it as nontemporal would not affect |
1717 | anything). */ |
1718 | |
1719 | dist = volume; |
1720 | for (j = 0; j < DDR_NUM_DIST_VECTS (dep); j++) |
1721 | { |
1722 | adist = volume_of_dist_vector (DDR_DIST_VECT (dep, j), |
1723 | loop_sizes: loop_data_size, n); |
1724 | |
1725 | /* If this is a dependence in the innermost loop (i.e., the |
1726 | distances in all superloops are zero) and it is not |
1727 | the trivial self-dependence with distance zero, record that |
1728 | the references are not completely independent. */ |
1729 | if (lambda_vector_zerop (DDR_DIST_VECT (dep, j), size: n - 1) |
1730 | && (ref != refb |
1731 | || DDR_DIST_VECT (dep, j)[n-1] != 0)) |
1732 | { |
1733 | ref->independent_p = false; |
1734 | refb->independent_p = false; |
1735 | } |
1736 | |
1737 | /* Ignore accesses closer than |
1738 | L1_CACHE_SIZE_BYTES / NONTEMPORAL_FRACTION, |
1739 | so that we use nontemporal prefetches e.g. if single memory |
1740 | location is accessed several times in a single iteration of |
1741 | the loop. */ |
1742 | if (adist < L1_CACHE_SIZE_BYTES / NONTEMPORAL_FRACTION) |
1743 | continue; |
1744 | |
1745 | if (adist < dist) |
1746 | dist = adist; |
1747 | } |
1748 | } |
1749 | |
1750 | if (ref->reuse_distance > dist) |
1751 | ref->reuse_distance = dist; |
1752 | if (refb->reuse_distance > dist) |
1753 | refb->reuse_distance = dist; |
1754 | } |
1755 | |
1756 | free_dependence_relations (dependences); |
1757 | free_data_refs (datarefs); |
1758 | free (ptr: loop_data_size); |
1759 | |
1760 | if (dump_file && (dump_flags & TDF_DETAILS)) |
1761 | { |
1762 | fprintf (stream: dump_file, format: "Reuse distances:\n" ); |
1763 | for (gr = refs; gr; gr = gr->next) |
1764 | for (ref = gr->refs; ref; ref = ref->next) |
1765 | fprintf (stream: dump_file, format: " reference %u:%u distance %u\n" , |
1766 | ref->group->uid, ref->uid, ref->reuse_distance); |
1767 | } |
1768 | |
1769 | return true; |
1770 | } |
1771 | |
1772 | /* Determine whether or not the trip count to ahead ratio is too small based |
1773 | on prefitablility consideration. |
1774 | AHEAD: the iteration ahead distance, |
1775 | EST_NITER: the estimated trip count. */ |
1776 | |
1777 | static bool |
1778 | trip_count_to_ahead_ratio_too_small_p (unsigned ahead, HOST_WIDE_INT est_niter) |
1779 | { |
1780 | /* Assume trip count to ahead ratio is big enough if the trip count could not |
1781 | be estimated at compile time. */ |
1782 | if (est_niter < 0) |
1783 | return false; |
1784 | |
1785 | if (est_niter < (HOST_WIDE_INT) (TRIP_COUNT_TO_AHEAD_RATIO * ahead)) |
1786 | { |
1787 | if (dump_file && (dump_flags & TDF_DETAILS)) |
1788 | fprintf (stream: dump_file, |
1789 | format: "Not prefetching -- loop estimated to roll only %d times\n" , |
1790 | (int) est_niter); |
1791 | return true; |
1792 | } |
1793 | |
1794 | return false; |
1795 | } |
1796 | |
1797 | /* Determine whether or not the number of memory references in the loop is |
1798 | reasonable based on the profitablity and compilation time considerations. |
1799 | NINSNS: estimated number of instructions in the loop, |
1800 | MEM_REF_COUNT: total number of memory references in the loop. */ |
1801 | |
1802 | static bool |
1803 | mem_ref_count_reasonable_p (unsigned ninsns, unsigned mem_ref_count) |
1804 | { |
1805 | int insn_to_mem_ratio; |
1806 | |
1807 | if (mem_ref_count == 0) |
1808 | return false; |
1809 | |
1810 | /* Miss rate computation (is_miss_rate_acceptable) and dependence analysis |
1811 | (compute_all_dependences) have high costs based on quadratic complexity. |
1812 | To avoid huge compilation time, we give up prefetching if mem_ref_count |
1813 | is too large. */ |
1814 | if (mem_ref_count > PREFETCH_MAX_MEM_REFS_PER_LOOP) |
1815 | return false; |
1816 | |
1817 | /* Prefetching improves performance by overlapping cache missing |
1818 | memory accesses with CPU operations. If the loop does not have |
1819 | enough CPU operations to overlap with memory operations, prefetching |
1820 | won't give a significant benefit. One approximate way of checking |
1821 | this is to require the ratio of instructions to memory references to |
1822 | be above a certain limit. This approximation works well in practice. |
1823 | TODO: Implement a more precise computation by estimating the time |
1824 | for each CPU or memory op in the loop. Time estimates for memory ops |
1825 | should account for cache misses. */ |
1826 | insn_to_mem_ratio = ninsns / mem_ref_count; |
1827 | |
1828 | if (insn_to_mem_ratio < param_prefetch_min_insn_to_mem_ratio) |
1829 | { |
1830 | if (dump_file && (dump_flags & TDF_DETAILS)) |
1831 | fprintf (stream: dump_file, |
1832 | format: "Not prefetching -- instruction to memory reference ratio (%d) too small\n" , |
1833 | insn_to_mem_ratio); |
1834 | return false; |
1835 | } |
1836 | |
1837 | return true; |
1838 | } |
1839 | |
1840 | /* Determine whether or not the instruction to prefetch ratio in the loop is |
1841 | too small based on the profitablity consideration. |
1842 | NINSNS: estimated number of instructions in the loop, |
1843 | PREFETCH_COUNT: an estimate of the number of prefetches, |
1844 | UNROLL_FACTOR: the factor to unroll the loop if prefetching. */ |
1845 | |
1846 | static bool |
1847 | insn_to_prefetch_ratio_too_small_p (unsigned ninsns, unsigned prefetch_count, |
1848 | unsigned unroll_factor) |
1849 | { |
1850 | int insn_to_prefetch_ratio; |
1851 | |
1852 | /* Prefetching most likely causes performance degradation when the instruction |
1853 | to prefetch ratio is too small. Too many prefetch instructions in a loop |
1854 | may reduce the I-cache performance. |
1855 | (unroll_factor * ninsns) is used to estimate the number of instructions in |
1856 | the unrolled loop. This implementation is a bit simplistic -- the number |
1857 | of issued prefetch instructions is also affected by unrolling. So, |
1858 | prefetch_mod and the unroll factor should be taken into account when |
1859 | determining prefetch_count. Also, the number of insns of the unrolled |
1860 | loop will usually be significantly smaller than the number of insns of the |
1861 | original loop * unroll_factor (at least the induction variable increases |
1862 | and the exit branches will get eliminated), so it might be better to use |
1863 | tree_estimate_loop_size + estimated_unrolled_size. */ |
1864 | insn_to_prefetch_ratio = (unroll_factor * ninsns) / prefetch_count; |
1865 | if (insn_to_prefetch_ratio < param_min_insn_to_prefetch_ratio) |
1866 | { |
1867 | if (dump_file && (dump_flags & TDF_DETAILS)) |
1868 | fprintf (stream: dump_file, |
1869 | format: "Not prefetching -- instruction to prefetch ratio (%d) too small\n" , |
1870 | insn_to_prefetch_ratio); |
1871 | return true; |
1872 | } |
1873 | |
1874 | return false; |
1875 | } |
1876 | |
1877 | |
1878 | /* Issue prefetch instructions for array references in LOOP. Returns |
1879 | true if the LOOP was unrolled and updates NEED_LC_SSA_UPDATE if we need |
1880 | to update SSA for virtual operands and LC SSA for a split edge. */ |
1881 | |
1882 | static bool |
1883 | loop_prefetch_arrays (class loop *loop, bool &need_lc_ssa_update) |
1884 | { |
1885 | struct mem_ref_group *refs; |
1886 | unsigned ahead, ninsns, time, unroll_factor; |
1887 | HOST_WIDE_INT est_niter; |
1888 | class tree_niter_desc desc; |
1889 | bool unrolled = false, no_other_refs; |
1890 | unsigned prefetch_count; |
1891 | unsigned mem_ref_count; |
1892 | |
1893 | if (optimize_loop_nest_for_size_p (loop)) |
1894 | { |
1895 | if (dump_file && (dump_flags & TDF_DETAILS)) |
1896 | fprintf (stream: dump_file, format: " ignored (cold area)\n" ); |
1897 | return false; |
1898 | } |
1899 | |
1900 | /* FIXME: the time should be weighted by the probabilities of the blocks in |
1901 | the loop body. */ |
1902 | time = tree_num_loop_insns (loop, &eni_time_weights); |
1903 | if (time == 0) |
1904 | return false; |
1905 | |
1906 | ahead = (param_prefetch_latency + time - 1) / time; |
1907 | est_niter = estimated_stmt_executions_int (loop); |
1908 | if (est_niter == -1) |
1909 | est_niter = likely_max_stmt_executions_int (loop); |
1910 | |
1911 | /* Prefetching is not likely to be profitable if the trip count to ahead |
1912 | ratio is too small. */ |
1913 | if (trip_count_to_ahead_ratio_too_small_p (ahead, est_niter)) |
1914 | return false; |
1915 | |
1916 | ninsns = tree_num_loop_insns (loop, &eni_size_weights); |
1917 | |
1918 | /* Step 1: gather the memory references. */ |
1919 | refs = gather_memory_references (loop, no_other_refs: &no_other_refs, ref_count: &mem_ref_count); |
1920 | |
1921 | /* Give up prefetching if the number of memory references in the |
1922 | loop is not reasonable based on profitablity and compilation time |
1923 | considerations. */ |
1924 | if (!mem_ref_count_reasonable_p (ninsns, mem_ref_count)) |
1925 | goto fail; |
1926 | |
1927 | /* Step 2: estimate the reuse effects. */ |
1928 | prune_by_reuse (groups: refs); |
1929 | |
1930 | if (nothing_to_prefetch_p (groups: refs)) |
1931 | goto fail; |
1932 | |
1933 | if (!determine_loop_nest_reuse (loop, refs, no_other_refs)) |
1934 | goto fail; |
1935 | |
1936 | /* Step 3: determine unroll factor. */ |
1937 | unroll_factor = determine_unroll_factor (loop, refs, ninsns, desc: &desc, |
1938 | est_niter); |
1939 | |
1940 | /* Estimate prefetch count for the unrolled loop. */ |
1941 | prefetch_count = estimate_prefetch_count (groups: refs, unroll_factor); |
1942 | if (prefetch_count == 0) |
1943 | goto fail; |
1944 | |
1945 | if (dump_file && (dump_flags & TDF_DETAILS)) |
1946 | fprintf (stream: dump_file, format: "Ahead %d, unroll factor %d, trip count " |
1947 | HOST_WIDE_INT_PRINT_DEC "\n" |
1948 | "insn count %d, mem ref count %d, prefetch count %d\n" , |
1949 | ahead, unroll_factor, est_niter, |
1950 | ninsns, mem_ref_count, prefetch_count); |
1951 | |
1952 | /* Prefetching is not likely to be profitable if the instruction to prefetch |
1953 | ratio is too small. */ |
1954 | if (insn_to_prefetch_ratio_too_small_p (ninsns, prefetch_count, |
1955 | unroll_factor)) |
1956 | goto fail; |
1957 | |
1958 | need_lc_ssa_update |= mark_nontemporal_stores (loop, groups: refs); |
1959 | |
1960 | /* Step 4: what to prefetch? */ |
1961 | if (!schedule_prefetches (groups: refs, unroll_factor, ahead)) |
1962 | goto fail; |
1963 | |
1964 | /* Step 5: unroll the loop. TODO -- peeling of first and last few |
1965 | iterations so that we do not issue superfluous prefetches. */ |
1966 | if (unroll_factor != 1) |
1967 | { |
1968 | tree_unroll_loop (loop, unroll_factor, &desc); |
1969 | unrolled = true; |
1970 | } |
1971 | |
1972 | /* Step 6: issue the prefetches. */ |
1973 | issue_prefetches (groups: refs, unroll_factor, ahead); |
1974 | |
1975 | fail: |
1976 | release_mem_refs (groups: refs); |
1977 | return unrolled; |
1978 | } |
1979 | |
1980 | /* Issue prefetch instructions for array references in loops. */ |
1981 | |
1982 | unsigned int |
1983 | tree_ssa_prefetch_arrays (void) |
1984 | { |
1985 | bool unrolled = false; |
1986 | bool need_lc_ssa_update = false; |
1987 | int todo_flags = 0; |
1988 | |
1989 | if (!targetm.have_prefetch () |
1990 | /* It is possible to ask compiler for say -mtune=i486 -march=pentium4. |
1991 | -mtune=i486 causes us having PREFETCH_BLOCK 0, since this is part |
1992 | of processor costs and i486 does not have prefetch, but |
1993 | -march=pentium4 causes targetm.have_prefetch to be true. Ugh. */ |
1994 | || PREFETCH_BLOCK == 0) |
1995 | return 0; |
1996 | |
1997 | if (dump_file && (dump_flags & TDF_DETAILS)) |
1998 | { |
1999 | fprintf (stream: dump_file, format: "Prefetching parameters:\n" ); |
2000 | fprintf (stream: dump_file, format: " simultaneous prefetches: %d\n" , |
2001 | param_simultaneous_prefetches); |
2002 | fprintf (stream: dump_file, format: " prefetch latency: %d\n" , param_prefetch_latency); |
2003 | fprintf (stream: dump_file, format: " prefetch block size: %d\n" , PREFETCH_BLOCK); |
2004 | fprintf (stream: dump_file, format: " L1 cache size: %d lines, %d kB\n" , |
2005 | L1_CACHE_SIZE_BYTES / param_l1_cache_line_size, |
2006 | param_l1_cache_size); |
2007 | fprintf (stream: dump_file, format: " L1 cache line size: %d\n" , |
2008 | param_l1_cache_line_size); |
2009 | fprintf (stream: dump_file, format: " L2 cache size: %d kB\n" , param_l2_cache_size); |
2010 | fprintf (stream: dump_file, format: " min insn-to-prefetch ratio: %d \n" , |
2011 | param_min_insn_to_prefetch_ratio); |
2012 | fprintf (stream: dump_file, format: " min insn-to-mem ratio: %d \n" , |
2013 | param_prefetch_min_insn_to_mem_ratio); |
2014 | fprintf (stream: dump_file, format: "\n" ); |
2015 | } |
2016 | |
2017 | initialize_original_copy_tables (); |
2018 | |
2019 | if (!builtin_decl_explicit_p (fncode: BUILT_IN_PREFETCH)) |
2020 | { |
2021 | tree type = build_function_type_list (void_type_node, |
2022 | const_ptr_type_node, NULL_TREE); |
2023 | tree decl = add_builtin_function (name: "__builtin_prefetch" , type, |
2024 | function_code: BUILT_IN_PREFETCH, cl: BUILT_IN_NORMAL, |
2025 | NULL, NULL_TREE); |
2026 | DECL_IS_NOVOPS (decl) = true; |
2027 | set_builtin_decl (fncode: BUILT_IN_PREFETCH, decl, implicit_p: false); |
2028 | } |
2029 | |
2030 | for (auto loop : loops_list (cfun, LI_FROM_INNERMOST)) |
2031 | { |
2032 | if (dump_file && (dump_flags & TDF_DETAILS)) |
2033 | fprintf (stream: dump_file, format: "Processing loop %d:\n" , loop->num); |
2034 | |
2035 | unrolled |= loop_prefetch_arrays (loop, need_lc_ssa_update); |
2036 | |
2037 | if (dump_file && (dump_flags & TDF_DETAILS)) |
2038 | fprintf (stream: dump_file, format: "\n\n" ); |
2039 | } |
2040 | |
2041 | if (need_lc_ssa_update) |
2042 | rewrite_into_loop_closed_ssa (NULL, TODO_update_ssa_only_virtuals); |
2043 | |
2044 | if (unrolled) |
2045 | { |
2046 | scev_reset (); |
2047 | todo_flags |= TODO_cleanup_cfg; |
2048 | } |
2049 | |
2050 | free_original_copy_tables (); |
2051 | return todo_flags; |
2052 | } |
2053 | |
2054 | /* Prefetching. */ |
2055 | |
2056 | namespace { |
2057 | |
2058 | const pass_data pass_data_loop_prefetch = |
2059 | { |
2060 | .type: GIMPLE_PASS, /* type */ |
2061 | .name: "aprefetch" , /* name */ |
2062 | .optinfo_flags: OPTGROUP_LOOP, /* optinfo_flags */ |
2063 | .tv_id: TV_TREE_PREFETCH, /* tv_id */ |
2064 | .properties_required: ( PROP_cfg | PROP_ssa ), /* properties_required */ |
2065 | .properties_provided: 0, /* properties_provided */ |
2066 | .properties_destroyed: 0, /* properties_destroyed */ |
2067 | .todo_flags_start: 0, /* todo_flags_start */ |
2068 | .todo_flags_finish: 0, /* todo_flags_finish */ |
2069 | }; |
2070 | |
2071 | class pass_loop_prefetch : public gimple_opt_pass |
2072 | { |
2073 | public: |
2074 | pass_loop_prefetch (gcc::context *ctxt) |
2075 | : gimple_opt_pass (pass_data_loop_prefetch, ctxt) |
2076 | {} |
2077 | |
2078 | /* opt_pass methods: */ |
2079 | bool gate (function *) final override |
2080 | { |
2081 | return flag_prefetch_loop_arrays > 0; |
2082 | } |
2083 | unsigned int execute (function *) final override; |
2084 | |
2085 | }; // class pass_loop_prefetch |
2086 | |
2087 | unsigned int |
2088 | pass_loop_prefetch::execute (function *fun) |
2089 | { |
2090 | if (number_of_loops (fn: fun) <= 1) |
2091 | return 0; |
2092 | |
2093 | if ((PREFETCH_BLOCK & (PREFETCH_BLOCK - 1)) != 0) |
2094 | { |
2095 | static bool warned = false; |
2096 | |
2097 | if (!warned) |
2098 | { |
2099 | warning (OPT_Wdisabled_optimization, |
2100 | "%<l1-cache-size%> parameter is not a power of two %d" , |
2101 | PREFETCH_BLOCK); |
2102 | warned = true; |
2103 | } |
2104 | return 0; |
2105 | } |
2106 | |
2107 | return tree_ssa_prefetch_arrays (); |
2108 | } |
2109 | |
2110 | } // anon namespace |
2111 | |
2112 | gimple_opt_pass * |
2113 | make_pass_loop_prefetch (gcc::context *ctxt) |
2114 | { |
2115 | return new pass_loop_prefetch (ctxt); |
2116 | } |
2117 | |
2118 | |
2119 | |