]>
Commit | Line | Data |
---|---|---|
b076a3fd | 1 | /* Array prefetching. |
d1e082c2 | 2 | Copyright (C) 2005-2013 Free Software Foundation, Inc. |
b8698a0f | 3 | |
b076a3fd | 4 | This file is part of GCC. |
b8698a0f | 5 | |
b076a3fd ZD |
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 | |
9dcd6f09 | 8 | Free Software Foundation; either version 3, or (at your option) any |
b076a3fd | 9 | later version. |
b8698a0f | 10 | |
b076a3fd ZD |
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. | |
b8698a0f | 15 | |
b076a3fd | 16 | You should have received a copy of the GNU General Public License |
9dcd6f09 NC |
17 | along with GCC; see the file COPYING3. If not see |
18 | <http://www.gnu.org/licenses/>. */ | |
b076a3fd ZD |
19 | |
20 | #include "config.h" | |
21 | #include "system.h" | |
22 | #include "coretypes.h" | |
23 | #include "tm.h" | |
24 | #include "tree.h" | |
d8a2d370 | 25 | #include "stor-layout.h" |
b076a3fd | 26 | #include "tm_p.h" |
b076a3fd | 27 | #include "basic-block.h" |
cf835838 | 28 | #include "tree-pretty-print.h" |
18f429e2 | 29 | #include "gimple.h" |
45b0be94 | 30 | #include "gimplify.h" |
5be5c238 | 31 | #include "gimple-iterator.h" |
18f429e2 | 32 | #include "gimplify-me.h" |
442b4905 | 33 | #include "gimple-ssa.h" |
e28030cf AM |
34 | #include "tree-ssa-loop-ivopts.h" |
35 | #include "tree-ssa-loop-manip.h" | |
36 | #include "tree-ssa-loop-niter.h" | |
442b4905 AM |
37 | #include "tree-ssa-loop.h" |
38 | #include "tree-into-ssa.h" | |
b076a3fd | 39 | #include "cfgloop.h" |
b076a3fd | 40 | #include "tree-pass.h" |
b076a3fd | 41 | #include "insn-config.h" |
b076a3fd ZD |
42 | #include "hashtab.h" |
43 | #include "tree-chrec.h" | |
44 | #include "tree-scalar-evolution.h" | |
718f9c0f | 45 | #include "diagnostic-core.h" |
b076a3fd ZD |
46 | #include "params.h" |
47 | #include "langhooks.h" | |
7f9bc51b | 48 | #include "tree-inline.h" |
5417e022 | 49 | #include "tree-data-ref.h" |
2eb79bbb SB |
50 | |
51 | ||
52 | /* FIXME: Needed for optabs, but this should all be moved to a TBD interface | |
53 | between the GIMPLE and RTL worlds. */ | |
54 | #include "expr.h" | |
79f5e442 | 55 | #include "optabs.h" |
1c1ad7bb | 56 | #include "recog.h" |
b076a3fd ZD |
57 | |
58 | /* This pass inserts prefetch instructions to optimize cache usage during | |
59 | accesses to arrays in loops. It processes loops sequentially and: | |
60 | ||
61 | 1) Gathers all memory references in the single loop. | |
62 | 2) For each of the references it decides when it is profitable to prefetch | |
63 | it. To do it, we evaluate the reuse among the accesses, and determines | |
64 | two values: PREFETCH_BEFORE (meaning that it only makes sense to do | |
65 | prefetching in the first PREFETCH_BEFORE iterations of the loop) and | |
66 | PREFETCH_MOD (meaning that it only makes sense to prefetch in the | |
67 | iterations of the loop that are zero modulo PREFETCH_MOD). For example | |
68 | (assuming cache line size is 64 bytes, char has size 1 byte and there | |
69 | is no hardware sequential prefetch): | |
70 | ||
71 | char *a; | |
72 | for (i = 0; i < max; i++) | |
73 | { | |
74 | a[255] = ...; (0) | |
75 | a[i] = ...; (1) | |
76 | a[i + 64] = ...; (2) | |
77 | a[16*i] = ...; (3) | |
78 | a[187*i] = ...; (4) | |
79 | a[187*i + 50] = ...; (5) | |
80 | } | |
81 | ||
82 | (0) obviously has PREFETCH_BEFORE 1 | |
83 | (1) has PREFETCH_BEFORE 64, since (2) accesses the same memory | |
84 | location 64 iterations before it, and PREFETCH_MOD 64 (since | |
85 | it hits the same cache line otherwise). | |
86 | (2) has PREFETCH_MOD 64 | |
87 | (3) has PREFETCH_MOD 4 | |
88 | (4) has PREFETCH_MOD 1. We do not set PREFETCH_BEFORE here, since | |
bae077dc | 89 | the cache line accessed by (5) is the same with probability only |
b076a3fd ZD |
90 | 7/32. |
91 | (5) has PREFETCH_MOD 1 as well. | |
92 | ||
5417e022 ZD |
93 | Additionally, we use data dependence analysis to determine for each |
94 | reference the distance till the first reuse; this information is used | |
95 | to determine the temporality of the issued prefetch instruction. | |
96 | ||
b076a3fd ZD |
97 | 3) We determine how much ahead we need to prefetch. The number of |
98 | iterations needed is time to fetch / time spent in one iteration of | |
99 | the loop. The problem is that we do not know either of these values, | |
100 | so we just make a heuristic guess based on a magic (possibly) | |
101 | target-specific constant and size of the loop. | |
102 | ||
103 | 4) Determine which of the references we prefetch. We take into account | |
104 | that there is a maximum number of simultaneous prefetches (provided | |
105 | by machine description). We prefetch as many prefetches as possible | |
106 | while still within this bound (starting with those with lowest | |
107 | prefetch_mod, since they are responsible for most of the cache | |
108 | misses). | |
b8698a0f | 109 | |
b076a3fd ZD |
110 | 5) We unroll and peel loops so that we are able to satisfy PREFETCH_MOD |
111 | and PREFETCH_BEFORE requirements (within some bounds), and to avoid | |
112 | prefetching nonaccessed memory. | |
113 | TODO -- actually implement peeling. | |
b8698a0f | 114 | |
b076a3fd ZD |
115 | 6) We actually emit the prefetch instructions. ??? Perhaps emit the |
116 | prefetch instructions with guards in cases where 5) was not sufficient | |
117 | to satisfy the constraints? | |
118 | ||
0bbe50f6 CF |
119 | A cost model is implemented to determine whether or not prefetching is |
120 | profitable for a given loop. The cost model has three heuristics: | |
121 | ||
122 | 1. Function trip_count_to_ahead_ratio_too_small_p implements a | |
123 | heuristic that determines whether or not the loop has too few | |
124 | iterations (compared to ahead). Prefetching is not likely to be | |
125 | beneficial if the trip count to ahead ratio is below a certain | |
126 | minimum. | |
127 | ||
128 | 2. Function mem_ref_count_reasonable_p implements a heuristic that | |
129 | determines whether the given loop has enough CPU ops that can be | |
130 | overlapped with cache missing memory ops. If not, the loop | |
131 | won't benefit from prefetching. In the implementation, | |
132 | prefetching is not considered beneficial if the ratio between | |
133 | the instruction count and the mem ref count is below a certain | |
134 | minimum. | |
135 | ||
136 | 3. Function insn_to_prefetch_ratio_too_small_p implements a | |
137 | heuristic that disables prefetching in a loop if the prefetching | |
138 | cost is above a certain limit. The relative prefetching cost is | |
139 | estimated by taking the ratio between the prefetch count and the | |
140 | total intruction count (this models the I-cache cost). | |
141 | ||
db34470d | 142 | The limits used in these heuristics are defined as parameters with |
b8698a0f | 143 | reasonable default values. Machine-specific default values will be |
db34470d | 144 | added later. |
b8698a0f | 145 | |
b076a3fd ZD |
146 | Some other TODO: |
147 | -- write and use more general reuse analysis (that could be also used | |
148 | in other cache aimed loop optimizations) | |
149 | -- make it behave sanely together with the prefetches given by user | |
150 | (now we just ignore them; at the very least we should avoid | |
151 | optimizing loops in that user put his own prefetches) | |
152 | -- we assume cache line size alignment of arrays; this could be | |
153 | improved. */ | |
154 | ||
155 | /* Magic constants follow. These should be replaced by machine specific | |
156 | numbers. */ | |
157 | ||
b076a3fd ZD |
158 | /* True if write can be prefetched by a read prefetch. */ |
159 | ||
160 | #ifndef WRITE_CAN_USE_READ_PREFETCH | |
161 | #define WRITE_CAN_USE_READ_PREFETCH 1 | |
162 | #endif | |
163 | ||
164 | /* True if read can be prefetched by a write prefetch. */ | |
165 | ||
166 | #ifndef READ_CAN_USE_WRITE_PREFETCH | |
167 | #define READ_CAN_USE_WRITE_PREFETCH 0 | |
168 | #endif | |
169 | ||
47eb5b32 ZD |
170 | /* The size of the block loaded by a single prefetch. Usually, this is |
171 | the same as cache line size (at the moment, we only consider one level | |
172 | of cache hierarchy). */ | |
b076a3fd ZD |
173 | |
174 | #ifndef PREFETCH_BLOCK | |
47eb5b32 | 175 | #define PREFETCH_BLOCK L1_CACHE_LINE_SIZE |
b076a3fd ZD |
176 | #endif |
177 | ||
178 | /* Do we have a forward hardware sequential prefetching? */ | |
179 | ||
180 | #ifndef HAVE_FORWARD_PREFETCH | |
181 | #define HAVE_FORWARD_PREFETCH 0 | |
182 | #endif | |
183 | ||
184 | /* Do we have a backward hardware sequential prefetching? */ | |
185 | ||
186 | #ifndef HAVE_BACKWARD_PREFETCH | |
187 | #define HAVE_BACKWARD_PREFETCH 0 | |
188 | #endif | |
189 | ||
190 | /* In some cases we are only able to determine that there is a certain | |
191 | probability that the two accesses hit the same cache line. In this | |
192 | case, we issue the prefetches for both of them if this probability | |
fa10beec | 193 | is less then (1000 - ACCEPTABLE_MISS_RATE) per thousand. */ |
b076a3fd ZD |
194 | |
195 | #ifndef ACCEPTABLE_MISS_RATE | |
196 | #define ACCEPTABLE_MISS_RATE 50 | |
197 | #endif | |
198 | ||
199 | #ifndef HAVE_prefetch | |
200 | #define HAVE_prefetch 0 | |
201 | #endif | |
202 | ||
46cb0441 ZD |
203 | #define L1_CACHE_SIZE_BYTES ((unsigned) (L1_CACHE_SIZE * 1024)) |
204 | #define L2_CACHE_SIZE_BYTES ((unsigned) (L2_CACHE_SIZE * 1024)) | |
5417e022 ZD |
205 | |
206 | /* We consider a memory access nontemporal if it is not reused sooner than | |
207 | after L2_CACHE_SIZE_BYTES of memory are accessed. However, we ignore | |
208 | accesses closer than L1_CACHE_SIZE_BYTES / NONTEMPORAL_FRACTION, | |
209 | so that we use nontemporal prefetches e.g. if single memory location | |
210 | is accessed several times in a single iteration of the loop. */ | |
211 | #define NONTEMPORAL_FRACTION 16 | |
212 | ||
79f5e442 ZD |
213 | /* In case we have to emit a memory fence instruction after the loop that |
214 | uses nontemporal stores, this defines the builtin to use. */ | |
215 | ||
216 | #ifndef FENCE_FOLLOWING_MOVNT | |
217 | #define FENCE_FOLLOWING_MOVNT NULL_TREE | |
218 | #endif | |
219 | ||
9bf4598b CF |
220 | /* It is not profitable to prefetch when the trip count is not at |
221 | least TRIP_COUNT_TO_AHEAD_RATIO times the prefetch ahead distance. | |
222 | For example, in a loop with a prefetch ahead distance of 10, | |
223 | supposing that TRIP_COUNT_TO_AHEAD_RATIO is equal to 4, it is | |
224 | profitable to prefetch when the trip count is greater or equal to | |
225 | 40. In that case, 30 out of the 40 iterations will benefit from | |
226 | prefetching. */ | |
227 | ||
228 | #ifndef TRIP_COUNT_TO_AHEAD_RATIO | |
229 | #define TRIP_COUNT_TO_AHEAD_RATIO 4 | |
230 | #endif | |
231 | ||
b076a3fd ZD |
232 | /* The group of references between that reuse may occur. */ |
233 | ||
234 | struct mem_ref_group | |
235 | { | |
236 | tree base; /* Base of the reference. */ | |
81f32326 | 237 | tree step; /* Step of the reference. */ |
b076a3fd ZD |
238 | struct mem_ref *refs; /* References in the group. */ |
239 | struct mem_ref_group *next; /* Next group of references. */ | |
240 | }; | |
241 | ||
242 | /* Assigned to PREFETCH_BEFORE when all iterations are to be prefetched. */ | |
243 | ||
244 | #define PREFETCH_ALL (~(unsigned HOST_WIDE_INT) 0) | |
245 | ||
8532678c CF |
246 | /* Do not generate a prefetch if the unroll factor is significantly less |
247 | than what is required by the prefetch. This is to avoid redundant | |
f7963a7c CF |
248 | prefetches. For example, when prefetch_mod is 16 and unroll_factor is |
249 | 2, prefetching requires unrolling the loop 16 times, but | |
250 | the loop is actually unrolled twice. In this case (ratio = 8), | |
8532678c CF |
251 | prefetching is not likely to be beneficial. */ |
252 | ||
253 | #ifndef PREFETCH_MOD_TO_UNROLL_FACTOR_RATIO | |
f7963a7c | 254 | #define PREFETCH_MOD_TO_UNROLL_FACTOR_RATIO 4 |
8532678c CF |
255 | #endif |
256 | ||
0bbe50f6 CF |
257 | /* Some of the prefetch computations have quadratic complexity. We want to |
258 | avoid huge compile times and, therefore, want to limit the amount of | |
259 | memory references per loop where we consider prefetching. */ | |
260 | ||
261 | #ifndef PREFETCH_MAX_MEM_REFS_PER_LOOP | |
262 | #define PREFETCH_MAX_MEM_REFS_PER_LOOP 200 | |
263 | #endif | |
264 | ||
b076a3fd ZD |
265 | /* The memory reference. */ |
266 | ||
267 | struct mem_ref | |
268 | { | |
726a989a | 269 | gimple stmt; /* Statement in that the reference appears. */ |
b076a3fd ZD |
270 | tree mem; /* The reference. */ |
271 | HOST_WIDE_INT delta; /* Constant offset of the reference. */ | |
b076a3fd ZD |
272 | struct mem_ref_group *group; /* The group of references it belongs to. */ |
273 | unsigned HOST_WIDE_INT prefetch_mod; | |
274 | /* Prefetch only each PREFETCH_MOD-th | |
275 | iteration. */ | |
276 | unsigned HOST_WIDE_INT prefetch_before; | |
277 | /* Prefetch only first PREFETCH_BEFORE | |
278 | iterations. */ | |
5417e022 ZD |
279 | unsigned reuse_distance; /* The amount of data accessed before the first |
280 | reuse of this value. */ | |
b076a3fd | 281 | struct mem_ref *next; /* The next reference in the group. */ |
79f5e442 ZD |
282 | unsigned write_p : 1; /* Is it a write? */ |
283 | unsigned independent_p : 1; /* True if the reference is independent on | |
284 | all other references inside the loop. */ | |
285 | unsigned issue_prefetch_p : 1; /* Should we really issue the prefetch? */ | |
286 | unsigned storent_p : 1; /* True if we changed the store to a | |
287 | nontemporal one. */ | |
b076a3fd ZD |
288 | }; |
289 | ||
a5497b12 | 290 | /* Dumps information about memory reference */ |
b076a3fd | 291 | static void |
a5497b12 VK |
292 | dump_mem_details (FILE *file, tree base, tree step, |
293 | HOST_WIDE_INT delta, bool write_p) | |
b076a3fd | 294 | { |
a5497b12 VK |
295 | fprintf (file, "(base "); |
296 | print_generic_expr (file, base, TDF_SLIM); | |
b076a3fd | 297 | fprintf (file, ", step "); |
a5497b12 VK |
298 | if (cst_and_fits_in_hwi (step)) |
299 | fprintf (file, HOST_WIDE_INT_PRINT_DEC, int_cst_value (step)); | |
81f32326 | 300 | else |
a5497b12 | 301 | print_generic_expr (file, step, TDF_TREE); |
b076a3fd | 302 | fprintf (file, ")\n"); |
e324a72f | 303 | fprintf (file, " delta "); |
a5497b12 VK |
304 | fprintf (file, HOST_WIDE_INT_PRINT_DEC, delta); |
305 | fprintf (file, "\n"); | |
306 | fprintf (file, " %s\n", write_p ? "write" : "read"); | |
b076a3fd | 307 | fprintf (file, "\n"); |
a5497b12 | 308 | } |
b076a3fd | 309 | |
a5497b12 | 310 | /* Dumps information about reference REF to FILE. */ |
b076a3fd | 311 | |
a5497b12 VK |
312 | static void |
313 | dump_mem_ref (FILE *file, struct mem_ref *ref) | |
314 | { | |
315 | fprintf (file, "Reference %p:\n", (void *) ref); | |
316 | ||
317 | fprintf (file, " group %p ", (void *) ref->group); | |
318 | ||
319 | dump_mem_details (file, ref->group->base, ref->group->step, ref->delta, | |
320 | ref->write_p); | |
b076a3fd ZD |
321 | } |
322 | ||
323 | /* Finds a group with BASE and STEP in GROUPS, or creates one if it does not | |
324 | exist. */ | |
325 | ||
326 | static struct mem_ref_group * | |
81f32326 | 327 | find_or_create_group (struct mem_ref_group **groups, tree base, tree step) |
b076a3fd ZD |
328 | { |
329 | struct mem_ref_group *group; | |
330 | ||
331 | for (; *groups; groups = &(*groups)->next) | |
332 | { | |
81f32326 | 333 | if (operand_equal_p ((*groups)->step, step, 0) |
b076a3fd ZD |
334 | && operand_equal_p ((*groups)->base, base, 0)) |
335 | return *groups; | |
336 | ||
81f32326 CB |
337 | /* If step is an integer constant, keep the list of groups sorted |
338 | by decreasing step. */ | |
339 | if (cst_and_fits_in_hwi ((*groups)->step) && cst_and_fits_in_hwi (step) | |
340 | && int_cst_value ((*groups)->step) < int_cst_value (step)) | |
b076a3fd ZD |
341 | break; |
342 | } | |
343 | ||
5417e022 | 344 | group = XNEW (struct mem_ref_group); |
b076a3fd ZD |
345 | group->base = base; |
346 | group->step = step; | |
347 | group->refs = NULL; | |
348 | group->next = *groups; | |
349 | *groups = group; | |
350 | ||
351 | return group; | |
352 | } | |
353 | ||
354 | /* Records a memory reference MEM in GROUP with offset DELTA and write status | |
355 | WRITE_P. The reference occurs in statement STMT. */ | |
356 | ||
357 | static void | |
726a989a | 358 | record_ref (struct mem_ref_group *group, gimple stmt, tree mem, |
b076a3fd ZD |
359 | HOST_WIDE_INT delta, bool write_p) |
360 | { | |
361 | struct mem_ref **aref; | |
362 | ||
363 | /* Do not record the same address twice. */ | |
364 | for (aref = &group->refs; *aref; aref = &(*aref)->next) | |
365 | { | |
366 | /* It does not have to be possible for write reference to reuse the read | |
367 | prefetch, or vice versa. */ | |
368 | if (!WRITE_CAN_USE_READ_PREFETCH | |
369 | && write_p | |
370 | && !(*aref)->write_p) | |
371 | continue; | |
372 | if (!READ_CAN_USE_WRITE_PREFETCH | |
373 | && !write_p | |
374 | && (*aref)->write_p) | |
375 | continue; | |
376 | ||
377 | if ((*aref)->delta == delta) | |
378 | return; | |
379 | } | |
380 | ||
5417e022 | 381 | (*aref) = XNEW (struct mem_ref); |
b076a3fd ZD |
382 | (*aref)->stmt = stmt; |
383 | (*aref)->mem = mem; | |
384 | (*aref)->delta = delta; | |
385 | (*aref)->write_p = write_p; | |
386 | (*aref)->prefetch_before = PREFETCH_ALL; | |
387 | (*aref)->prefetch_mod = 1; | |
5417e022 | 388 | (*aref)->reuse_distance = 0; |
b076a3fd ZD |
389 | (*aref)->issue_prefetch_p = false; |
390 | (*aref)->group = group; | |
391 | (*aref)->next = NULL; | |
79f5e442 ZD |
392 | (*aref)->independent_p = false; |
393 | (*aref)->storent_p = false; | |
b076a3fd ZD |
394 | |
395 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
396 | dump_mem_ref (dump_file, *aref); | |
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 (ref); | |
414 | } | |
415 | free (groups); | |
416 | } | |
417 | } | |
418 | ||
419 | /* A structure used to pass arguments to idx_analyze_ref. */ | |
420 | ||
421 | struct ar_data | |
422 | { | |
423 | struct loop *loop; /* Loop of the reference. */ | |
726a989a | 424 | gimple stmt; /* Statement of the reference. */ |
81f32326 | 425 | tree *step; /* Step of the memory reference. */ |
b076a3fd ZD |
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 | { | |
c22940cd | 435 | struct ar_data *ar_data = (struct ar_data *) data; |
b076a3fd | 436 | tree ibase, step, stepsize; |
81f32326 | 437 | HOST_WIDE_INT idelta = 0, imult = 1; |
b076a3fd ZD |
438 | affine_iv iv; |
439 | ||
f017bf5e | 440 | if (!simple_iv (ar_data->loop, loop_containing_stmt (ar_data->stmt), |
81f32326 | 441 | *index, &iv, true)) |
b076a3fd ZD |
442 | return false; |
443 | ibase = iv.base; | |
444 | step = iv.step; | |
445 | ||
5be014d5 | 446 | if (TREE_CODE (ibase) == POINTER_PLUS_EXPR |
b076a3fd ZD |
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); | |
ff5e9a94 | 455 | ibase = build_int_cst (TREE_TYPE (ibase), 0); |
b076a3fd ZD |
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); | |
8fde8b40 CB |
464 | step = fold_build2 (MULT_EXPR, sizetype, |
465 | fold_convert (sizetype, step), | |
466 | fold_convert (sizetype, stepsize)); | |
b076a3fd ZD |
467 | idelta *= imult; |
468 | } | |
469 | ||
8fde8b40 CB |
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)); | |
b076a3fd ZD |
476 | *ar_data->delta += idelta; |
477 | *index = ibase; | |
478 | ||
479 | return true; | |
480 | } | |
481 | ||
aac8b8ed | 482 | /* Tries to express REF_P in shape &BASE + STEP * iter + DELTA, where DELTA and |
b076a3fd | 483 | STEP are integer constants and iter is number of iterations of LOOP. The |
aac8b8ed RS |
484 | reference occurs in statement STMT. Strips nonaddressable component |
485 | references from REF_P. */ | |
b076a3fd ZD |
486 | |
487 | static bool | |
aac8b8ed | 488 | analyze_ref (struct loop *loop, tree *ref_p, tree *base, |
81f32326 | 489 | tree *step, HOST_WIDE_INT *delta, |
726a989a | 490 | gimple stmt) |
b076a3fd ZD |
491 | { |
492 | struct ar_data ar_data; | |
493 | tree off; | |
494 | HOST_WIDE_INT bit_offset; | |
aac8b8ed | 495 | tree ref = *ref_p; |
b076a3fd | 496 | |
81f32326 | 497 | *step = NULL_TREE; |
b076a3fd ZD |
498 | *delta = 0; |
499 | ||
7c6dafac CF |
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 | } | |
b076a3fd | 512 | |
aac8b8ed RS |
513 | *ref_p = ref; |
514 | ||
b076a3fd ZD |
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); | |
b8698a0f | 520 | |
b076a3fd ZD |
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 | |
79f5e442 ZD |
533 | LOOP in statement STMT and it is write if WRITE_P. Returns true if the |
534 | reference was recorded, false otherwise. */ | |
b076a3fd | 535 | |
79f5e442 | 536 | static bool |
b076a3fd | 537 | gather_memory_references_ref (struct loop *loop, struct mem_ref_group **refs, |
726a989a | 538 | tree ref, bool write_p, gimple stmt) |
b076a3fd | 539 | { |
81f32326 CB |
540 | tree base, step; |
541 | HOST_WIDE_INT delta; | |
b076a3fd ZD |
542 | struct mem_ref_group *agrp; |
543 | ||
a80a2701 JJ |
544 | if (get_base_address (ref) == NULL) |
545 | return false; | |
546 | ||
aac8b8ed | 547 | if (!analyze_ref (loop, &ref, &base, &step, &delta, stmt)) |
79f5e442 | 548 | return false; |
81f32326 CB |
549 | /* If analyze_ref fails the default is a NULL_TREE. We can stop here. */ |
550 | if (step == NULL_TREE) | |
551 | return false; | |
b076a3fd | 552 | |
756f50ce | 553 | /* Stop if the address of BASE could not be taken. */ |
bc068a23 CF |
554 | if (may_be_nonaddressable_p (base)) |
555 | return false; | |
556 | ||
a5497b12 VK |
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 (dump_file, "Memory expression %p\n",(void *) ref ); | |
566 | print_generic_expr (dump_file, ref, TDF_TREE); | |
567 | fprintf (dump_file,":"); | |
c3284718 | 568 | dump_mem_details (dump_file, base, step, delta, write_p); |
a5497b12 VK |
569 | fprintf (dump_file, |
570 | "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 (dump_file, "Memory expression %p\n",(void *) ref ); | |
583 | print_generic_expr (dump_file, ref, TDF_TREE); | |
584 | fprintf (dump_file,":"); | |
c3284718 | 585 | dump_mem_details (dump_file, base, step, delta, write_p); |
a5497b12 VK |
586 | fprintf (dump_file, |
587 | "Not prefetching, ignoring %p due to " | |
588 | "loop variant step\n", | |
589 | (void *) ref); | |
590 | } | |
591 | return false; | |
592 | } | |
593 | } | |
594 | } | |
50814135 | 595 | |
b076a3fd ZD |
596 | /* Now we know that REF = &BASE + STEP * iter + DELTA, where DELTA and STEP |
597 | are integer constants. */ | |
598 | agrp = find_or_create_group (refs, base, step); | |
599 | record_ref (agrp, stmt, ref, delta, write_p); | |
79f5e442 ZD |
600 | |
601 | return true; | |
b076a3fd ZD |
602 | } |
603 | ||
79f5e442 ZD |
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. */ | |
b076a3fd ZD |
606 | |
607 | static struct mem_ref_group * | |
db34470d | 608 | gather_memory_references (struct loop *loop, bool *no_other_refs, unsigned *ref_count) |
b076a3fd ZD |
609 | { |
610 | basic_block *body = get_loop_body_in_dom_order (loop); | |
611 | basic_block bb; | |
612 | unsigned i; | |
726a989a RB |
613 | gimple_stmt_iterator bsi; |
614 | gimple stmt; | |
615 | tree lhs, rhs; | |
b076a3fd ZD |
616 | struct mem_ref_group *refs = NULL; |
617 | ||
79f5e442 | 618 | *no_other_refs = true; |
db34470d | 619 | *ref_count = 0; |
79f5e442 | 620 | |
b076a3fd ZD |
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 | ||
726a989a | 629 | for (bsi = gsi_start_bb (bb); !gsi_end_p (bsi); gsi_next (&bsi)) |
b076a3fd | 630 | { |
726a989a | 631 | stmt = gsi_stmt (bsi); |
79f5e442 | 632 | |
726a989a | 633 | if (gimple_code (stmt) != GIMPLE_ASSIGN) |
79f5e442 | 634 | { |
5006671f | 635 | if (gimple_vuse (stmt) |
726a989a RB |
636 | || (is_gimple_call (stmt) |
637 | && !(gimple_call_flags (stmt) & ECF_CONST))) | |
79f5e442 ZD |
638 | *no_other_refs = false; |
639 | continue; | |
640 | } | |
b076a3fd | 641 | |
726a989a RB |
642 | lhs = gimple_assign_lhs (stmt); |
643 | rhs = gimple_assign_rhs1 (stmt); | |
b076a3fd ZD |
644 | |
645 | if (REFERENCE_CLASS_P (rhs)) | |
db34470d | 646 | { |
79f5e442 ZD |
647 | *no_other_refs &= gather_memory_references_ref (loop, &refs, |
648 | rhs, false, stmt); | |
db34470d GS |
649 | *ref_count += 1; |
650 | } | |
b076a3fd | 651 | if (REFERENCE_CLASS_P (lhs)) |
db34470d | 652 | { |
79f5e442 ZD |
653 | *no_other_refs &= gather_memory_references_ref (loop, &refs, |
654 | lhs, true, stmt); | |
db34470d GS |
655 | *ref_count += 1; |
656 | } | |
b076a3fd ZD |
657 | } |
658 | } | |
659 | free (body); | |
660 | ||
661 | return refs; | |
662 | } | |
663 | ||
664 | /* Prune the prefetch candidate REF using the self-reuse. */ | |
665 | ||
666 | static void | |
667 | prune_ref_by_self_reuse (struct mem_ref *ref) | |
668 | { | |
81f32326 CB |
669 | HOST_WIDE_INT step; |
670 | bool backward; | |
671 | ||
672 | /* If the step size is non constant, we cannot calculate prefetch_mod. */ | |
673 | if (!cst_and_fits_in_hwi (ref->group->step)) | |
674 | return; | |
675 | ||
676 | step = int_cst_value (ref->group->step); | |
677 | ||
678 | backward = step < 0; | |
b076a3fd ZD |
679 | |
680 | if (step == 0) | |
681 | { | |
682 | /* Prefetch references to invariant address just once. */ | |
683 | ref->prefetch_before = 1; | |
684 | return; | |
685 | } | |
686 | ||
687 | if (backward) | |
688 | step = -step; | |
689 | ||
690 | if (step > PREFETCH_BLOCK) | |
691 | return; | |
692 | ||
693 | if ((backward && HAVE_BACKWARD_PREFETCH) | |
694 | || (!backward && HAVE_FORWARD_PREFETCH)) | |
695 | { | |
696 | ref->prefetch_before = 1; | |
697 | return; | |
698 | } | |
699 | ||
700 | ref->prefetch_mod = PREFETCH_BLOCK / step; | |
701 | } | |
702 | ||
703 | /* Divides X by BY, rounding down. */ | |
704 | ||
705 | static HOST_WIDE_INT | |
706 | ddown (HOST_WIDE_INT x, unsigned HOST_WIDE_INT by) | |
707 | { | |
708 | gcc_assert (by > 0); | |
709 | ||
710 | if (x >= 0) | |
711 | return x / by; | |
712 | else | |
713 | return (x + by - 1) / by; | |
714 | } | |
715 | ||
b8698a0f L |
716 | /* Given a CACHE_LINE_SIZE and two inductive memory references |
717 | with a common STEP greater than CACHE_LINE_SIZE and an address | |
718 | difference DELTA, compute the probability that they will fall | |
14e444c3 CF |
719 | in different cache lines. Return true if the computed miss rate |
720 | is not greater than the ACCEPTABLE_MISS_RATE. DISTINCT_ITERS is the | |
721 | number of distinct iterations after which the pattern repeats itself. | |
2c6dd136 GS |
722 | ALIGN_UNIT is the unit of alignment in bytes. */ |
723 | ||
14e444c3 CF |
724 | static bool |
725 | is_miss_rate_acceptable (unsigned HOST_WIDE_INT cache_line_size, | |
2c6dd136 GS |
726 | HOST_WIDE_INT step, HOST_WIDE_INT delta, |
727 | unsigned HOST_WIDE_INT distinct_iters, | |
728 | int align_unit) | |
729 | { | |
730 | unsigned align, iter; | |
14e444c3 | 731 | int total_positions, miss_positions, max_allowed_miss_positions; |
2c6dd136 GS |
732 | int address1, address2, cache_line1, cache_line2; |
733 | ||
a245c04b CF |
734 | /* It always misses if delta is greater than or equal to the cache |
735 | line size. */ | |
14e444c3 CF |
736 | if (delta >= (HOST_WIDE_INT) cache_line_size) |
737 | return false; | |
a245c04b | 738 | |
2c6dd136 | 739 | miss_positions = 0; |
14e444c3 CF |
740 | total_positions = (cache_line_size / align_unit) * distinct_iters; |
741 | max_allowed_miss_positions = (ACCEPTABLE_MISS_RATE * total_positions) / 1000; | |
b8698a0f | 742 | |
2c6dd136 GS |
743 | /* Iterate through all possible alignments of the first |
744 | memory reference within its cache line. */ | |
745 | for (align = 0; align < cache_line_size; align += align_unit) | |
746 | ||
747 | /* Iterate through all distinct iterations. */ | |
748 | for (iter = 0; iter < distinct_iters; iter++) | |
749 | { | |
750 | address1 = align + step * iter; | |
751 | address2 = address1 + delta; | |
752 | cache_line1 = address1 / cache_line_size; | |
753 | cache_line2 = address2 / cache_line_size; | |
2c6dd136 | 754 | if (cache_line1 != cache_line2) |
14e444c3 CF |
755 | { |
756 | miss_positions += 1; | |
757 | if (miss_positions > max_allowed_miss_positions) | |
758 | return false; | |
759 | } | |
2c6dd136 | 760 | } |
14e444c3 | 761 | return true; |
2c6dd136 GS |
762 | } |
763 | ||
b076a3fd ZD |
764 | /* Prune the prefetch candidate REF using the reuse with BY. |
765 | If BY_IS_BEFORE is true, BY is before REF in the loop. */ | |
766 | ||
767 | static void | |
768 | prune_ref_by_group_reuse (struct mem_ref *ref, struct mem_ref *by, | |
769 | bool by_is_before) | |
770 | { | |
81f32326 CB |
771 | HOST_WIDE_INT step; |
772 | bool backward; | |
b076a3fd ZD |
773 | HOST_WIDE_INT delta_r = ref->delta, delta_b = by->delta; |
774 | HOST_WIDE_INT delta = delta_b - delta_r; | |
775 | HOST_WIDE_INT hit_from; | |
776 | unsigned HOST_WIDE_INT prefetch_before, prefetch_block; | |
2c6dd136 GS |
777 | HOST_WIDE_INT reduced_step; |
778 | unsigned HOST_WIDE_INT reduced_prefetch_block; | |
779 | tree ref_type; | |
780 | int align_unit; | |
b076a3fd | 781 | |
81f32326 CB |
782 | /* If the step is non constant we cannot calculate prefetch_before. */ |
783 | if (!cst_and_fits_in_hwi (ref->group->step)) { | |
784 | return; | |
785 | } | |
786 | ||
787 | step = int_cst_value (ref->group->step); | |
788 | ||
789 | backward = step < 0; | |
790 | ||
791 | ||
b076a3fd ZD |
792 | if (delta == 0) |
793 | { | |
794 | /* If the references has the same address, only prefetch the | |
795 | former. */ | |
796 | if (by_is_before) | |
797 | ref->prefetch_before = 0; | |
b8698a0f | 798 | |
b076a3fd ZD |
799 | return; |
800 | } | |
801 | ||
802 | if (!step) | |
803 | { | |
804 | /* If the reference addresses are invariant and fall into the | |
805 | same cache line, prefetch just the first one. */ | |
806 | if (!by_is_before) | |
807 | return; | |
808 | ||
809 | if (ddown (ref->delta, PREFETCH_BLOCK) | |
810 | != ddown (by->delta, PREFETCH_BLOCK)) | |
811 | return; | |
812 | ||
813 | ref->prefetch_before = 0; | |
814 | return; | |
815 | } | |
816 | ||
817 | /* Only prune the reference that is behind in the array. */ | |
818 | if (backward) | |
819 | { | |
820 | if (delta > 0) | |
821 | return; | |
822 | ||
823 | /* Transform the data so that we may assume that the accesses | |
824 | are forward. */ | |
825 | delta = - delta; | |
826 | step = -step; | |
827 | delta_r = PREFETCH_BLOCK - 1 - delta_r; | |
828 | delta_b = PREFETCH_BLOCK - 1 - delta_b; | |
829 | } | |
830 | else | |
831 | { | |
832 | if (delta < 0) | |
833 | return; | |
834 | } | |
835 | ||
836 | /* Check whether the two references are likely to hit the same cache | |
837 | line, and how distant the iterations in that it occurs are from | |
838 | each other. */ | |
839 | ||
840 | if (step <= PREFETCH_BLOCK) | |
841 | { | |
842 | /* The accesses are sure to meet. Let us check when. */ | |
843 | hit_from = ddown (delta_b, PREFETCH_BLOCK) * PREFETCH_BLOCK; | |
844 | prefetch_before = (hit_from - delta_r + step - 1) / step; | |
845 | ||
57762e97 | 846 | /* Do not reduce prefetch_before if we meet beyond cache size. */ |
4c9cf7af | 847 | if (prefetch_before > absu_hwi (L2_CACHE_SIZE_BYTES / step)) |
57762e97 | 848 | prefetch_before = PREFETCH_ALL; |
b076a3fd ZD |
849 | if (prefetch_before < ref->prefetch_before) |
850 | ref->prefetch_before = prefetch_before; | |
851 | ||
852 | return; | |
853 | } | |
854 | ||
b8698a0f | 855 | /* A more complicated case with step > prefetch_block. First reduce |
2c6dd136 | 856 | the ratio between the step and the cache line size to its simplest |
b8698a0f L |
857 | terms. The resulting denominator will then represent the number of |
858 | distinct iterations after which each address will go back to its | |
859 | initial location within the cache line. This computation assumes | |
2c6dd136 | 860 | that PREFETCH_BLOCK is a power of two. */ |
b076a3fd | 861 | prefetch_block = PREFETCH_BLOCK; |
2c6dd136 GS |
862 | reduced_prefetch_block = prefetch_block; |
863 | reduced_step = step; | |
864 | while ((reduced_step & 1) == 0 | |
865 | && reduced_prefetch_block > 1) | |
b076a3fd | 866 | { |
2c6dd136 GS |
867 | reduced_step >>= 1; |
868 | reduced_prefetch_block >>= 1; | |
b076a3fd ZD |
869 | } |
870 | ||
b076a3fd ZD |
871 | prefetch_before = delta / step; |
872 | delta %= step; | |
2c6dd136 GS |
873 | ref_type = TREE_TYPE (ref->mem); |
874 | align_unit = TYPE_ALIGN (ref_type) / 8; | |
14e444c3 CF |
875 | if (is_miss_rate_acceptable (prefetch_block, step, delta, |
876 | reduced_prefetch_block, align_unit)) | |
b076a3fd | 877 | { |
57762e97 CB |
878 | /* Do not reduce prefetch_before if we meet beyond cache size. */ |
879 | if (prefetch_before > L2_CACHE_SIZE_BYTES / PREFETCH_BLOCK) | |
880 | prefetch_before = PREFETCH_ALL; | |
b076a3fd ZD |
881 | if (prefetch_before < ref->prefetch_before) |
882 | ref->prefetch_before = prefetch_before; | |
883 | ||
884 | return; | |
885 | } | |
886 | ||
887 | /* Try also the following iteration. */ | |
888 | prefetch_before++; | |
889 | delta = step - delta; | |
14e444c3 CF |
890 | if (is_miss_rate_acceptable (prefetch_block, step, delta, |
891 | reduced_prefetch_block, align_unit)) | |
b076a3fd ZD |
892 | { |
893 | if (prefetch_before < ref->prefetch_before) | |
894 | ref->prefetch_before = prefetch_before; | |
895 | ||
896 | return; | |
897 | } | |
898 | ||
899 | /* The ref probably does not reuse by. */ | |
900 | return; | |
901 | } | |
902 | ||
903 | /* Prune the prefetch candidate REF using the reuses with other references | |
904 | in REFS. */ | |
905 | ||
906 | static void | |
907 | prune_ref_by_reuse (struct mem_ref *ref, struct mem_ref *refs) | |
908 | { | |
909 | struct mem_ref *prune_by; | |
910 | bool before = true; | |
911 | ||
912 | prune_ref_by_self_reuse (ref); | |
913 | ||
914 | for (prune_by = refs; prune_by; prune_by = prune_by->next) | |
915 | { | |
916 | if (prune_by == ref) | |
917 | { | |
918 | before = false; | |
919 | continue; | |
920 | } | |
921 | ||
922 | if (!WRITE_CAN_USE_READ_PREFETCH | |
923 | && ref->write_p | |
924 | && !prune_by->write_p) | |
925 | continue; | |
926 | if (!READ_CAN_USE_WRITE_PREFETCH | |
927 | && !ref->write_p | |
928 | && prune_by->write_p) | |
929 | continue; | |
930 | ||
931 | prune_ref_by_group_reuse (ref, prune_by, before); | |
932 | } | |
933 | } | |
934 | ||
935 | /* Prune the prefetch candidates in GROUP using the reuse analysis. */ | |
936 | ||
937 | static void | |
938 | prune_group_by_reuse (struct mem_ref_group *group) | |
939 | { | |
940 | struct mem_ref *ref_pruned; | |
941 | ||
942 | for (ref_pruned = group->refs; ref_pruned; ref_pruned = ref_pruned->next) | |
943 | { | |
944 | prune_ref_by_reuse (ref_pruned, group->refs); | |
945 | ||
946 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
947 | { | |
948 | fprintf (dump_file, "Reference %p:", (void *) ref_pruned); | |
949 | ||
950 | if (ref_pruned->prefetch_before == PREFETCH_ALL | |
951 | && ref_pruned->prefetch_mod == 1) | |
952 | fprintf (dump_file, " no restrictions"); | |
953 | else if (ref_pruned->prefetch_before == 0) | |
954 | fprintf (dump_file, " do not prefetch"); | |
955 | else if (ref_pruned->prefetch_before <= ref_pruned->prefetch_mod) | |
956 | fprintf (dump_file, " prefetch once"); | |
957 | else | |
958 | { | |
959 | if (ref_pruned->prefetch_before != PREFETCH_ALL) | |
960 | { | |
961 | fprintf (dump_file, " prefetch before "); | |
962 | fprintf (dump_file, HOST_WIDE_INT_PRINT_DEC, | |
963 | ref_pruned->prefetch_before); | |
964 | } | |
965 | if (ref_pruned->prefetch_mod != 1) | |
966 | { | |
967 | fprintf (dump_file, " prefetch mod "); | |
968 | fprintf (dump_file, HOST_WIDE_INT_PRINT_DEC, | |
969 | ref_pruned->prefetch_mod); | |
970 | } | |
971 | } | |
972 | fprintf (dump_file, "\n"); | |
973 | } | |
974 | } | |
975 | } | |
976 | ||
977 | /* Prune the list of prefetch candidates GROUPS using the reuse analysis. */ | |
978 | ||
979 | static void | |
980 | prune_by_reuse (struct mem_ref_group *groups) | |
981 | { | |
982 | for (; groups; groups = groups->next) | |
983 | prune_group_by_reuse (groups); | |
984 | } | |
985 | ||
986 | /* Returns true if we should issue prefetch for REF. */ | |
987 | ||
988 | static bool | |
989 | should_issue_prefetch_p (struct mem_ref *ref) | |
990 | { | |
991 | /* For now do not issue prefetches for only first few of the | |
992 | iterations. */ | |
993 | if (ref->prefetch_before != PREFETCH_ALL) | |
a8beb3a7 CB |
994 | { |
995 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
996 | fprintf (dump_file, "Ignoring %p due to prefetch_before\n", | |
997 | (void *) ref); | |
998 | return false; | |
999 | } | |
b076a3fd | 1000 | |
79f5e442 ZD |
1001 | /* Do not prefetch nontemporal stores. */ |
1002 | if (ref->storent_p) | |
a8beb3a7 CB |
1003 | { |
1004 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
1005 | fprintf (dump_file, "Ignoring nontemporal store %p\n", (void *) ref); | |
1006 | return false; | |
1007 | } | |
79f5e442 | 1008 | |
b076a3fd ZD |
1009 | return true; |
1010 | } | |
1011 | ||
1012 | /* Decide which of the prefetch candidates in GROUPS to prefetch. | |
1013 | AHEAD is the number of iterations to prefetch ahead (which corresponds | |
1014 | to the number of simultaneous instances of one prefetch running at a | |
1015 | time). UNROLL_FACTOR is the factor by that the loop is going to be | |
1016 | unrolled. Returns true if there is anything to prefetch. */ | |
1017 | ||
1018 | static bool | |
1019 | schedule_prefetches (struct mem_ref_group *groups, unsigned unroll_factor, | |
1020 | unsigned ahead) | |
1021 | { | |
911b3fdb ZD |
1022 | unsigned remaining_prefetch_slots, n_prefetches, prefetch_slots; |
1023 | unsigned slots_per_prefetch; | |
b076a3fd ZD |
1024 | struct mem_ref *ref; |
1025 | bool any = false; | |
1026 | ||
911b3fdb ZD |
1027 | /* At most SIMULTANEOUS_PREFETCHES should be running at the same time. */ |
1028 | remaining_prefetch_slots = SIMULTANEOUS_PREFETCHES; | |
b076a3fd | 1029 | |
911b3fdb ZD |
1030 | /* The prefetch will run for AHEAD iterations of the original loop, i.e., |
1031 | AHEAD / UNROLL_FACTOR iterations of the unrolled loop. In each iteration, | |
1032 | it will need a prefetch slot. */ | |
1033 | slots_per_prefetch = (ahead + unroll_factor / 2) / unroll_factor; | |
b076a3fd | 1034 | if (dump_file && (dump_flags & TDF_DETAILS)) |
911b3fdb ZD |
1035 | fprintf (dump_file, "Each prefetch instruction takes %u prefetch slots.\n", |
1036 | slots_per_prefetch); | |
b076a3fd ZD |
1037 | |
1038 | /* For now we just take memory references one by one and issue | |
1039 | prefetches for as many as possible. The groups are sorted | |
1040 | starting with the largest step, since the references with | |
c0220ea4 | 1041 | large step are more likely to cause many cache misses. */ |
b076a3fd ZD |
1042 | |
1043 | for (; groups; groups = groups->next) | |
1044 | for (ref = groups->refs; ref; ref = ref->next) | |
1045 | { | |
1046 | if (!should_issue_prefetch_p (ref)) | |
1047 | continue; | |
1048 | ||
8532678c CF |
1049 | /* The loop is far from being sufficiently unrolled for this |
1050 | prefetch. Do not generate prefetch to avoid many redudant | |
1051 | prefetches. */ | |
1052 | if (ref->prefetch_mod / unroll_factor > PREFETCH_MOD_TO_UNROLL_FACTOR_RATIO) | |
1053 | continue; | |
1054 | ||
911b3fdb ZD |
1055 | /* If we need to prefetch the reference each PREFETCH_MOD iterations, |
1056 | and we unroll the loop UNROLL_FACTOR times, we need to insert | |
1057 | ceil (UNROLL_FACTOR / PREFETCH_MOD) instructions in each | |
1058 | iteration. */ | |
b076a3fd ZD |
1059 | n_prefetches = ((unroll_factor + ref->prefetch_mod - 1) |
1060 | / ref->prefetch_mod); | |
911b3fdb ZD |
1061 | prefetch_slots = n_prefetches * slots_per_prefetch; |
1062 | ||
1063 | /* If more than half of the prefetches would be lost anyway, do not | |
1064 | issue the prefetch. */ | |
1065 | if (2 * remaining_prefetch_slots < prefetch_slots) | |
1066 | continue; | |
1067 | ||
1068 | ref->issue_prefetch_p = true; | |
b076a3fd | 1069 | |
911b3fdb ZD |
1070 | if (remaining_prefetch_slots <= prefetch_slots) |
1071 | return true; | |
1072 | remaining_prefetch_slots -= prefetch_slots; | |
b076a3fd ZD |
1073 | any = true; |
1074 | } | |
1075 | ||
1076 | return any; | |
1077 | } | |
1078 | ||
d5058523 CF |
1079 | /* Return TRUE if no prefetch is going to be generated in the given |
1080 | GROUPS. */ | |
1081 | ||
1082 | static bool | |
1083 | nothing_to_prefetch_p (struct mem_ref_group *groups) | |
1084 | { | |
1085 | struct mem_ref *ref; | |
1086 | ||
1087 | for (; groups; groups = groups->next) | |
1088 | for (ref = groups->refs; ref; ref = ref->next) | |
1089 | if (should_issue_prefetch_p (ref)) | |
1090 | return false; | |
1091 | ||
1092 | return true; | |
1093 | } | |
1094 | ||
1095 | /* Estimate the number of prefetches in the given GROUPS. | |
1096 | UNROLL_FACTOR is the factor by which LOOP was unrolled. */ | |
b076a3fd | 1097 | |
db34470d | 1098 | static int |
d5058523 | 1099 | estimate_prefetch_count (struct mem_ref_group *groups, unsigned unroll_factor) |
b076a3fd ZD |
1100 | { |
1101 | struct mem_ref *ref; | |
d5058523 | 1102 | unsigned n_prefetches; |
db34470d | 1103 | int prefetch_count = 0; |
b076a3fd ZD |
1104 | |
1105 | for (; groups; groups = groups->next) | |
1106 | for (ref = groups->refs; ref; ref = ref->next) | |
1107 | if (should_issue_prefetch_p (ref)) | |
d5058523 CF |
1108 | { |
1109 | n_prefetches = ((unroll_factor + ref->prefetch_mod - 1) | |
1110 | / ref->prefetch_mod); | |
1111 | prefetch_count += n_prefetches; | |
1112 | } | |
b076a3fd | 1113 | |
db34470d | 1114 | return prefetch_count; |
b076a3fd ZD |
1115 | } |
1116 | ||
1117 | /* Issue prefetches for the reference REF into loop as decided before. | |
1118 | HEAD is the number of iterations to prefetch ahead. UNROLL_FACTOR | |
917f1b7e | 1119 | is the factor by which LOOP was unrolled. */ |
b076a3fd ZD |
1120 | |
1121 | static void | |
1122 | issue_prefetch_ref (struct mem_ref *ref, unsigned unroll_factor, unsigned ahead) | |
1123 | { | |
1124 | HOST_WIDE_INT delta; | |
81f32326 | 1125 | tree addr, addr_base, write_p, local, forward; |
726a989a RB |
1126 | gimple prefetch; |
1127 | gimple_stmt_iterator bsi; | |
b076a3fd | 1128 | unsigned n_prefetches, ap; |
5417e022 | 1129 | bool nontemporal = ref->reuse_distance >= L2_CACHE_SIZE_BYTES; |
b076a3fd ZD |
1130 | |
1131 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
5417e022 ZD |
1132 | fprintf (dump_file, "Issued%s prefetch for %p.\n", |
1133 | nontemporal ? " nontemporal" : "", | |
1134 | (void *) ref); | |
b076a3fd | 1135 | |
726a989a | 1136 | bsi = gsi_for_stmt (ref->stmt); |
b076a3fd ZD |
1137 | |
1138 | n_prefetches = ((unroll_factor + ref->prefetch_mod - 1) | |
1139 | / ref->prefetch_mod); | |
1140 | addr_base = build_fold_addr_expr_with_type (ref->mem, ptr_type_node); | |
726a989a RB |
1141 | addr_base = force_gimple_operand_gsi (&bsi, unshare_expr (addr_base), |
1142 | true, NULL, true, GSI_SAME_STMT); | |
911b3fdb | 1143 | write_p = ref->write_p ? integer_one_node : integer_zero_node; |
9a9d280e | 1144 | local = nontemporal ? integer_zero_node : integer_three_node; |
b076a3fd ZD |
1145 | |
1146 | for (ap = 0; ap < n_prefetches; ap++) | |
1147 | { | |
81f32326 CB |
1148 | if (cst_and_fits_in_hwi (ref->group->step)) |
1149 | { | |
1150 | /* Determine the address to prefetch. */ | |
1151 | delta = (ahead + ap * ref->prefetch_mod) * | |
1152 | int_cst_value (ref->group->step); | |
5d49b6a7 | 1153 | addr = fold_build_pointer_plus_hwi (addr_base, delta); |
81f32326 CB |
1154 | addr = force_gimple_operand_gsi (&bsi, unshare_expr (addr), true, NULL, |
1155 | true, GSI_SAME_STMT); | |
1156 | } | |
1157 | else | |
1158 | { | |
1159 | /* The step size is non-constant but loop-invariant. We use the | |
1160 | heuristic to simply prefetch ahead iterations ahead. */ | |
1161 | forward = fold_build2 (MULT_EXPR, sizetype, | |
1162 | fold_convert (sizetype, ref->group->step), | |
1163 | fold_convert (sizetype, size_int (ahead))); | |
5d49b6a7 | 1164 | addr = fold_build_pointer_plus (addr_base, forward); |
81f32326 CB |
1165 | addr = force_gimple_operand_gsi (&bsi, unshare_expr (addr), true, |
1166 | NULL, true, GSI_SAME_STMT); | |
1167 | } | |
b076a3fd | 1168 | /* Create the prefetch instruction. */ |
e79983f4 | 1169 | prefetch = gimple_build_call (builtin_decl_explicit (BUILT_IN_PREFETCH), |
726a989a RB |
1170 | 3, addr, write_p, local); |
1171 | gsi_insert_before (&bsi, prefetch, GSI_SAME_STMT); | |
b076a3fd ZD |
1172 | } |
1173 | } | |
1174 | ||
1175 | /* Issue prefetches for the references in GROUPS into loop as decided before. | |
1176 | HEAD is the number of iterations to prefetch ahead. UNROLL_FACTOR is the | |
1177 | factor by that LOOP was unrolled. */ | |
1178 | ||
1179 | static void | |
1180 | issue_prefetches (struct mem_ref_group *groups, | |
1181 | unsigned unroll_factor, unsigned ahead) | |
1182 | { | |
1183 | struct mem_ref *ref; | |
1184 | ||
1185 | for (; groups; groups = groups->next) | |
1186 | for (ref = groups->refs; ref; ref = ref->next) | |
1187 | if (ref->issue_prefetch_p) | |
1188 | issue_prefetch_ref (ref, unroll_factor, ahead); | |
1189 | } | |
1190 | ||
79f5e442 ZD |
1191 | /* Returns true if REF is a memory write for that a nontemporal store insn |
1192 | can be used. */ | |
1193 | ||
1194 | static bool | |
1195 | nontemporal_store_p (struct mem_ref *ref) | |
1196 | { | |
1197 | enum machine_mode mode; | |
1198 | enum insn_code code; | |
1199 | ||
1200 | /* REF must be a write that is not reused. We require it to be independent | |
1201 | on all other memory references in the loop, as the nontemporal stores may | |
1202 | be reordered with respect to other memory references. */ | |
1203 | if (!ref->write_p | |
1204 | || !ref->independent_p | |
1205 | || ref->reuse_distance < L2_CACHE_SIZE_BYTES) | |
1206 | return false; | |
1207 | ||
1208 | /* Check that we have the storent instruction for the mode. */ | |
1209 | mode = TYPE_MODE (TREE_TYPE (ref->mem)); | |
1210 | if (mode == BLKmode) | |
1211 | return false; | |
1212 | ||
947131ba | 1213 | code = optab_handler (storent_optab, mode); |
79f5e442 ZD |
1214 | return code != CODE_FOR_nothing; |
1215 | } | |
1216 | ||
1217 | /* If REF is a nontemporal store, we mark the corresponding modify statement | |
1218 | and return true. Otherwise, we return false. */ | |
1219 | ||
1220 | static bool | |
1221 | mark_nontemporal_store (struct mem_ref *ref) | |
1222 | { | |
1223 | if (!nontemporal_store_p (ref)) | |
1224 | return false; | |
1225 | ||
1226 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
1227 | fprintf (dump_file, "Marked reference %p as a nontemporal store.\n", | |
1228 | (void *) ref); | |
1229 | ||
726a989a | 1230 | gimple_assign_set_nontemporal_move (ref->stmt, true); |
79f5e442 ZD |
1231 | ref->storent_p = true; |
1232 | ||
1233 | return true; | |
1234 | } | |
1235 | ||
1236 | /* Issue a memory fence instruction after LOOP. */ | |
1237 | ||
1238 | static void | |
1239 | emit_mfence_after_loop (struct loop *loop) | |
1240 | { | |
9771b263 | 1241 | vec<edge> exits = get_loop_exit_edges (loop); |
79f5e442 | 1242 | edge exit; |
726a989a RB |
1243 | gimple call; |
1244 | gimple_stmt_iterator bsi; | |
79f5e442 ZD |
1245 | unsigned i; |
1246 | ||
9771b263 | 1247 | FOR_EACH_VEC_ELT (exits, i, exit) |
79f5e442 | 1248 | { |
726a989a | 1249 | call = gimple_build_call (FENCE_FOLLOWING_MOVNT, 0); |
79f5e442 ZD |
1250 | |
1251 | if (!single_pred_p (exit->dest) | |
1252 | /* If possible, we prefer not to insert the fence on other paths | |
1253 | in cfg. */ | |
1254 | && !(exit->flags & EDGE_ABNORMAL)) | |
1255 | split_loop_exit_edge (exit); | |
726a989a | 1256 | bsi = gsi_after_labels (exit->dest); |
79f5e442 | 1257 | |
726a989a | 1258 | gsi_insert_before (&bsi, call, GSI_NEW_STMT); |
79f5e442 ZD |
1259 | } |
1260 | ||
9771b263 | 1261 | exits.release (); |
79f5e442 ZD |
1262 | update_ssa (TODO_update_ssa_only_virtuals); |
1263 | } | |
1264 | ||
1265 | /* Returns true if we can use storent in loop, false otherwise. */ | |
1266 | ||
1267 | static bool | |
1268 | may_use_storent_in_loop_p (struct loop *loop) | |
1269 | { | |
1270 | bool ret = true; | |
1271 | ||
1272 | if (loop->inner != NULL) | |
1273 | return false; | |
1274 | ||
1275 | /* If we must issue a mfence insn after using storent, check that there | |
1276 | is a suitable place for it at each of the loop exits. */ | |
1277 | if (FENCE_FOLLOWING_MOVNT != NULL_TREE) | |
1278 | { | |
9771b263 | 1279 | vec<edge> exits = get_loop_exit_edges (loop); |
79f5e442 ZD |
1280 | unsigned i; |
1281 | edge exit; | |
1282 | ||
9771b263 | 1283 | FOR_EACH_VEC_ELT (exits, i, exit) |
79f5e442 | 1284 | if ((exit->flags & EDGE_ABNORMAL) |
fefa31b5 | 1285 | && exit->dest == EXIT_BLOCK_PTR_FOR_FN (cfun)) |
79f5e442 ZD |
1286 | ret = false; |
1287 | ||
9771b263 | 1288 | exits.release (); |
79f5e442 ZD |
1289 | } |
1290 | ||
1291 | return ret; | |
1292 | } | |
1293 | ||
1294 | /* Marks nontemporal stores in LOOP. GROUPS contains the description of memory | |
1295 | references in the loop. */ | |
1296 | ||
1297 | static void | |
1298 | mark_nontemporal_stores (struct loop *loop, struct mem_ref_group *groups) | |
1299 | { | |
1300 | struct mem_ref *ref; | |
1301 | bool any = false; | |
1302 | ||
1303 | if (!may_use_storent_in_loop_p (loop)) | |
1304 | return; | |
1305 | ||
1306 | for (; groups; groups = groups->next) | |
1307 | for (ref = groups->refs; ref; ref = ref->next) | |
1308 | any |= mark_nontemporal_store (ref); | |
1309 | ||
1310 | if (any && FENCE_FOLLOWING_MOVNT != NULL_TREE) | |
1311 | emit_mfence_after_loop (loop); | |
1312 | } | |
1313 | ||
b076a3fd ZD |
1314 | /* Determines whether we can profitably unroll LOOP FACTOR times, and if |
1315 | this is the case, fill in DESC by the description of number of | |
1316 | iterations. */ | |
1317 | ||
1318 | static bool | |
1319 | should_unroll_loop_p (struct loop *loop, struct tree_niter_desc *desc, | |
1320 | unsigned factor) | |
1321 | { | |
1322 | if (!can_unroll_loop_p (loop, factor, desc)) | |
1323 | return false; | |
1324 | ||
1325 | /* We only consider loops without control flow for unrolling. This is not | |
1326 | a hard restriction -- tree_unroll_loop works with arbitrary loops | |
1327 | as well; but the unrolling/prefetching is usually more profitable for | |
1328 | loops consisting of a single basic block, and we want to limit the | |
1329 | code growth. */ | |
1330 | if (loop->num_nodes > 2) | |
1331 | return false; | |
1332 | ||
1333 | return true; | |
1334 | } | |
1335 | ||
1336 | /* Determine the coefficient by that unroll LOOP, from the information | |
1337 | contained in the list of memory references REFS. Description of | |
2711355f ZD |
1338 | umber of iterations of LOOP is stored to DESC. NINSNS is the number of |
1339 | insns of the LOOP. EST_NITER is the estimated number of iterations of | |
1340 | the loop, or -1 if no estimate is available. */ | |
b076a3fd ZD |
1341 | |
1342 | static unsigned | |
1343 | determine_unroll_factor (struct loop *loop, struct mem_ref_group *refs, | |
2711355f ZD |
1344 | unsigned ninsns, struct tree_niter_desc *desc, |
1345 | HOST_WIDE_INT est_niter) | |
b076a3fd | 1346 | { |
911b3fdb ZD |
1347 | unsigned upper_bound; |
1348 | unsigned nfactor, factor, mod_constraint; | |
b076a3fd ZD |
1349 | struct mem_ref_group *agp; |
1350 | struct mem_ref *ref; | |
1351 | ||
911b3fdb ZD |
1352 | /* First check whether the loop is not too large to unroll. We ignore |
1353 | PARAM_MAX_UNROLL_TIMES, because for small loops, it prevented us | |
1354 | from unrolling them enough to make exactly one cache line covered by each | |
1355 | iteration. Also, the goal of PARAM_MAX_UNROLL_TIMES is to prevent | |
1356 | us from unrolling the loops too many times in cases where we only expect | |
1357 | gains from better scheduling and decreasing loop overhead, which is not | |
1358 | the case here. */ | |
1359 | upper_bound = PARAM_VALUE (PARAM_MAX_UNROLLED_INSNS) / ninsns; | |
2711355f ZD |
1360 | |
1361 | /* If we unrolled the loop more times than it iterates, the unrolled version | |
1362 | of the loop would be never entered. */ | |
1363 | if (est_niter >= 0 && est_niter < (HOST_WIDE_INT) upper_bound) | |
1364 | upper_bound = est_niter; | |
1365 | ||
911b3fdb | 1366 | if (upper_bound <= 1) |
b076a3fd ZD |
1367 | return 1; |
1368 | ||
911b3fdb ZD |
1369 | /* Choose the factor so that we may prefetch each cache just once, |
1370 | but bound the unrolling by UPPER_BOUND. */ | |
1371 | factor = 1; | |
b076a3fd ZD |
1372 | for (agp = refs; agp; agp = agp->next) |
1373 | for (ref = agp->refs; ref; ref = ref->next) | |
911b3fdb ZD |
1374 | if (should_issue_prefetch_p (ref)) |
1375 | { | |
1376 | mod_constraint = ref->prefetch_mod; | |
1377 | nfactor = least_common_multiple (mod_constraint, factor); | |
1378 | if (nfactor <= upper_bound) | |
1379 | factor = nfactor; | |
1380 | } | |
b076a3fd ZD |
1381 | |
1382 | if (!should_unroll_loop_p (loop, desc, factor)) | |
1383 | return 1; | |
1384 | ||
1385 | return factor; | |
1386 | } | |
1387 | ||
5417e022 ZD |
1388 | /* Returns the total volume of the memory references REFS, taking into account |
1389 | reuses in the innermost loop and cache line size. TODO -- we should also | |
1390 | take into account reuses across the iterations of the loops in the loop | |
1391 | nest. */ | |
1392 | ||
1393 | static unsigned | |
1394 | volume_of_references (struct mem_ref_group *refs) | |
1395 | { | |
1396 | unsigned volume = 0; | |
1397 | struct mem_ref_group *gr; | |
1398 | struct mem_ref *ref; | |
1399 | ||
1400 | for (gr = refs; gr; gr = gr->next) | |
1401 | for (ref = gr->refs; ref; ref = ref->next) | |
1402 | { | |
1403 | /* Almost always reuses another value? */ | |
1404 | if (ref->prefetch_before != PREFETCH_ALL) | |
1405 | continue; | |
1406 | ||
1407 | /* If several iterations access the same cache line, use the size of | |
1408 | the line divided by this number. Otherwise, a cache line is | |
1409 | accessed in each iteration. TODO -- in the latter case, we should | |
1410 | take the size of the reference into account, rounding it up on cache | |
1411 | line size multiple. */ | |
1412 | volume += L1_CACHE_LINE_SIZE / ref->prefetch_mod; | |
1413 | } | |
1414 | return volume; | |
1415 | } | |
1416 | ||
1417 | /* Returns the volume of memory references accessed across VEC iterations of | |
1418 | loops, whose sizes are described in the LOOP_SIZES array. N is the number | |
1419 | of the loops in the nest (length of VEC and LOOP_SIZES vectors). */ | |
1420 | ||
1421 | static unsigned | |
1422 | volume_of_dist_vector (lambda_vector vec, unsigned *loop_sizes, unsigned n) | |
1423 | { | |
1424 | unsigned i; | |
1425 | ||
1426 | for (i = 0; i < n; i++) | |
1427 | if (vec[i] != 0) | |
1428 | break; | |
1429 | ||
1430 | if (i == n) | |
1431 | return 0; | |
1432 | ||
1433 | gcc_assert (vec[i] > 0); | |
1434 | ||
1435 | /* We ignore the parts of the distance vector in subloops, since usually | |
1436 | the numbers of iterations are much smaller. */ | |
1437 | return loop_sizes[i] * vec[i]; | |
1438 | } | |
1439 | ||
1440 | /* Add the steps of ACCESS_FN multiplied by STRIDE to the array STRIDE | |
1441 | at the position corresponding to the loop of the step. N is the depth | |
1442 | of the considered loop nest, and, LOOP is its innermost loop. */ | |
1443 | ||
1444 | static void | |
1445 | add_subscript_strides (tree access_fn, unsigned stride, | |
1446 | HOST_WIDE_INT *strides, unsigned n, struct loop *loop) | |
1447 | { | |
1448 | struct loop *aloop; | |
1449 | tree step; | |
1450 | HOST_WIDE_INT astep; | |
1451 | unsigned min_depth = loop_depth (loop) - n; | |
1452 | ||
1453 | while (TREE_CODE (access_fn) == POLYNOMIAL_CHREC) | |
1454 | { | |
1455 | aloop = get_chrec_loop (access_fn); | |
1456 | step = CHREC_RIGHT (access_fn); | |
1457 | access_fn = CHREC_LEFT (access_fn); | |
1458 | ||
1459 | if ((unsigned) loop_depth (aloop) <= min_depth) | |
1460 | continue; | |
1461 | ||
9541ffee | 1462 | if (tree_fits_shwi_p (step)) |
9439e9a1 | 1463 | astep = tree_to_shwi (step); |
5417e022 ZD |
1464 | else |
1465 | astep = L1_CACHE_LINE_SIZE; | |
1466 | ||
1467 | strides[n - 1 - loop_depth (loop) + loop_depth (aloop)] += astep * stride; | |
1468 | ||
1469 | } | |
1470 | } | |
1471 | ||
1472 | /* Returns the volume of memory references accessed between two consecutive | |
1473 | self-reuses of the reference DR. We consider the subscripts of DR in N | |
1474 | loops, and LOOP_SIZES contains the volumes of accesses in each of the | |
1475 | loops. LOOP is the innermost loop of the current loop nest. */ | |
1476 | ||
1477 | static unsigned | |
1478 | self_reuse_distance (data_reference_p dr, unsigned *loop_sizes, unsigned n, | |
1479 | struct loop *loop) | |
1480 | { | |
1481 | tree stride, access_fn; | |
1482 | HOST_WIDE_INT *strides, astride; | |
9771b263 | 1483 | vec<tree> access_fns; |
5417e022 ZD |
1484 | tree ref = DR_REF (dr); |
1485 | unsigned i, ret = ~0u; | |
1486 | ||
1487 | /* In the following example: | |
1488 | ||
1489 | for (i = 0; i < N; i++) | |
1490 | for (j = 0; j < N; j++) | |
1491 | use (a[j][i]); | |
1492 | the same cache line is accessed each N steps (except if the change from | |
1493 | i to i + 1 crosses the boundary of the cache line). Thus, for self-reuse, | |
1494 | we cannot rely purely on the results of the data dependence analysis. | |
1495 | ||
1496 | Instead, we compute the stride of the reference in each loop, and consider | |
1497 | the innermost loop in that the stride is less than cache size. */ | |
1498 | ||
1499 | strides = XCNEWVEC (HOST_WIDE_INT, n); | |
1500 | access_fns = DR_ACCESS_FNS (dr); | |
1501 | ||
9771b263 | 1502 | FOR_EACH_VEC_ELT (access_fns, i, access_fn) |
5417e022 ZD |
1503 | { |
1504 | /* Keep track of the reference corresponding to the subscript, so that we | |
1505 | know its stride. */ | |
1506 | while (handled_component_p (ref) && TREE_CODE (ref) != ARRAY_REF) | |
1507 | ref = TREE_OPERAND (ref, 0); | |
b8698a0f | 1508 | |
5417e022 ZD |
1509 | if (TREE_CODE (ref) == ARRAY_REF) |
1510 | { | |
1511 | stride = TYPE_SIZE_UNIT (TREE_TYPE (ref)); | |
cc269bb6 | 1512 | if (tree_fits_uhwi_p (stride)) |
ae7e9ddd | 1513 | astride = tree_to_uhwi (stride); |
5417e022 ZD |
1514 | else |
1515 | astride = L1_CACHE_LINE_SIZE; | |
1516 | ||
1517 | ref = TREE_OPERAND (ref, 0); | |
1518 | } | |
1519 | else | |
1520 | astride = 1; | |
1521 | ||
1522 | add_subscript_strides (access_fn, astride, strides, n, loop); | |
1523 | } | |
1524 | ||
1525 | for (i = n; i-- > 0; ) | |
1526 | { | |
1527 | unsigned HOST_WIDE_INT s; | |
1528 | ||
1529 | s = strides[i] < 0 ? -strides[i] : strides[i]; | |
1530 | ||
1531 | if (s < (unsigned) L1_CACHE_LINE_SIZE | |
1532 | && (loop_sizes[i] | |
1533 | > (unsigned) (L1_CACHE_SIZE_BYTES / NONTEMPORAL_FRACTION))) | |
1534 | { | |
1535 | ret = loop_sizes[i]; | |
1536 | break; | |
1537 | } | |
1538 | } | |
1539 | ||
1540 | free (strides); | |
1541 | return ret; | |
1542 | } | |
1543 | ||
1544 | /* Determines the distance till the first reuse of each reference in REFS | |
79f5e442 | 1545 | in the loop nest of LOOP. NO_OTHER_REFS is true if there are no other |
1e373390 | 1546 | memory references in the loop. Return false if the analysis fails. */ |
5417e022 | 1547 | |
1e373390 | 1548 | static bool |
79f5e442 ZD |
1549 | determine_loop_nest_reuse (struct loop *loop, struct mem_ref_group *refs, |
1550 | bool no_other_refs) | |
5417e022 ZD |
1551 | { |
1552 | struct loop *nest, *aloop; | |
6e1aa848 DN |
1553 | vec<data_reference_p> datarefs = vNULL; |
1554 | vec<ddr_p> dependences = vNULL; | |
5417e022 | 1555 | struct mem_ref_group *gr; |
79f5e442 | 1556 | struct mem_ref *ref, *refb; |
6e1aa848 | 1557 | vec<loop_p> vloops = vNULL; |
5417e022 ZD |
1558 | unsigned *loop_data_size; |
1559 | unsigned i, j, n; | |
1560 | unsigned volume, dist, adist; | |
1561 | HOST_WIDE_INT vol; | |
1562 | data_reference_p dr; | |
1563 | ddr_p dep; | |
1564 | ||
1565 | if (loop->inner) | |
1e373390 | 1566 | return true; |
5417e022 ZD |
1567 | |
1568 | /* Find the outermost loop of the loop nest of loop (we require that | |
1569 | there are no sibling loops inside the nest). */ | |
1570 | nest = loop; | |
1571 | while (1) | |
1572 | { | |
1573 | aloop = loop_outer (nest); | |
1574 | ||
1575 | if (aloop == current_loops->tree_root | |
1576 | || aloop->inner->next) | |
1577 | break; | |
1578 | ||
1579 | nest = aloop; | |
1580 | } | |
1581 | ||
1582 | /* For each loop, determine the amount of data accessed in each iteration. | |
1583 | We use this to estimate whether the reference is evicted from the | |
1584 | cache before its reuse. */ | |
1585 | find_loop_nest (nest, &vloops); | |
9771b263 | 1586 | n = vloops.length (); |
5417e022 ZD |
1587 | loop_data_size = XNEWVEC (unsigned, n); |
1588 | volume = volume_of_references (refs); | |
1589 | i = n; | |
1590 | while (i-- != 0) | |
1591 | { | |
1592 | loop_data_size[i] = volume; | |
1593 | /* Bound the volume by the L2 cache size, since above this bound, | |
1594 | all dependence distances are equivalent. */ | |
1595 | if (volume > L2_CACHE_SIZE_BYTES) | |
1596 | continue; | |
1597 | ||
9771b263 | 1598 | aloop = vloops[i]; |
652c4c71 | 1599 | vol = estimated_stmt_executions_int (aloop); |
e5b332cd | 1600 | if (vol == -1) |
5417e022 ZD |
1601 | vol = expected_loop_iterations (aloop); |
1602 | volume *= vol; | |
1603 | } | |
1604 | ||
1605 | /* Prepare the references in the form suitable for data dependence | |
0d52bcc1 | 1606 | analysis. We ignore unanalyzable data references (the results |
5417e022 ZD |
1607 | are used just as a heuristics to estimate temporality of the |
1608 | references, hence we do not need to worry about correctness). */ | |
1609 | for (gr = refs; gr; gr = gr->next) | |
1610 | for (ref = gr->refs; ref; ref = ref->next) | |
1611 | { | |
5c640e29 SP |
1612 | dr = create_data_ref (nest, loop_containing_stmt (ref->stmt), |
1613 | ref->mem, ref->stmt, !ref->write_p); | |
5417e022 ZD |
1614 | |
1615 | if (dr) | |
1616 | { | |
1617 | ref->reuse_distance = volume; | |
1618 | dr->aux = ref; | |
9771b263 | 1619 | datarefs.safe_push (dr); |
5417e022 | 1620 | } |
79f5e442 ZD |
1621 | else |
1622 | no_other_refs = false; | |
5417e022 ZD |
1623 | } |
1624 | ||
9771b263 | 1625 | FOR_EACH_VEC_ELT (datarefs, i, dr) |
5417e022 ZD |
1626 | { |
1627 | dist = self_reuse_distance (dr, loop_data_size, n, loop); | |
3d9a9f94 | 1628 | ref = (struct mem_ref *) dr->aux; |
5417e022 ZD |
1629 | if (ref->reuse_distance > dist) |
1630 | ref->reuse_distance = dist; | |
79f5e442 ZD |
1631 | |
1632 | if (no_other_refs) | |
1633 | ref->independent_p = true; | |
5417e022 ZD |
1634 | } |
1635 | ||
1e373390 RG |
1636 | if (!compute_all_dependences (datarefs, &dependences, vloops, true)) |
1637 | return false; | |
5417e022 | 1638 | |
9771b263 | 1639 | FOR_EACH_VEC_ELT (dependences, i, dep) |
5417e022 ZD |
1640 | { |
1641 | if (DDR_ARE_DEPENDENT (dep) == chrec_known) | |
1642 | continue; | |
1643 | ||
3d9a9f94 KG |
1644 | ref = (struct mem_ref *) DDR_A (dep)->aux; |
1645 | refb = (struct mem_ref *) DDR_B (dep)->aux; | |
79f5e442 | 1646 | |
5417e022 ZD |
1647 | if (DDR_ARE_DEPENDENT (dep) == chrec_dont_know |
1648 | || DDR_NUM_DIST_VECTS (dep) == 0) | |
1649 | { | |
0d52bcc1 | 1650 | /* If the dependence cannot be analyzed, assume that there might be |
5417e022 ZD |
1651 | a reuse. */ |
1652 | dist = 0; | |
b8698a0f | 1653 | |
79f5e442 ZD |
1654 | ref->independent_p = false; |
1655 | refb->independent_p = false; | |
5417e022 ZD |
1656 | } |
1657 | else | |
1658 | { | |
0d52bcc1 | 1659 | /* The distance vectors are normalized to be always lexicographically |
5417e022 ZD |
1660 | positive, hence we cannot tell just from them whether DDR_A comes |
1661 | before DDR_B or vice versa. However, it is not important, | |
1662 | anyway -- if DDR_A is close to DDR_B, then it is either reused in | |
1663 | DDR_B (and it is not nontemporal), or it reuses the value of DDR_B | |
1664 | in cache (and marking it as nontemporal would not affect | |
1665 | anything). */ | |
1666 | ||
1667 | dist = volume; | |
1668 | for (j = 0; j < DDR_NUM_DIST_VECTS (dep); j++) | |
1669 | { | |
1670 | adist = volume_of_dist_vector (DDR_DIST_VECT (dep, j), | |
1671 | loop_data_size, n); | |
1672 | ||
79f5e442 ZD |
1673 | /* If this is a dependence in the innermost loop (i.e., the |
1674 | distances in all superloops are zero) and it is not | |
1675 | the trivial self-dependence with distance zero, record that | |
1676 | the references are not completely independent. */ | |
1677 | if (lambda_vector_zerop (DDR_DIST_VECT (dep, j), n - 1) | |
1678 | && (ref != refb | |
1679 | || DDR_DIST_VECT (dep, j)[n-1] != 0)) | |
1680 | { | |
1681 | ref->independent_p = false; | |
1682 | refb->independent_p = false; | |
1683 | } | |
1684 | ||
5417e022 ZD |
1685 | /* Ignore accesses closer than |
1686 | L1_CACHE_SIZE_BYTES / NONTEMPORAL_FRACTION, | |
1687 | so that we use nontemporal prefetches e.g. if single memory | |
1688 | location is accessed several times in a single iteration of | |
1689 | the loop. */ | |
1690 | if (adist < L1_CACHE_SIZE_BYTES / NONTEMPORAL_FRACTION) | |
1691 | continue; | |
1692 | ||
1693 | if (adist < dist) | |
1694 | dist = adist; | |
1695 | } | |
1696 | } | |
1697 | ||
5417e022 ZD |
1698 | if (ref->reuse_distance > dist) |
1699 | ref->reuse_distance = dist; | |
79f5e442 ZD |
1700 | if (refb->reuse_distance > dist) |
1701 | refb->reuse_distance = dist; | |
5417e022 ZD |
1702 | } |
1703 | ||
1704 | free_dependence_relations (dependences); | |
1705 | free_data_refs (datarefs); | |
1706 | free (loop_data_size); | |
1707 | ||
1708 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
1709 | { | |
1710 | fprintf (dump_file, "Reuse distances:\n"); | |
1711 | for (gr = refs; gr; gr = gr->next) | |
1712 | for (ref = gr->refs; ref; ref = ref->next) | |
1713 | fprintf (dump_file, " ref %p distance %u\n", | |
1714 | (void *) ref, ref->reuse_distance); | |
1715 | } | |
1e373390 RG |
1716 | |
1717 | return true; | |
5417e022 ZD |
1718 | } |
1719 | ||
0bbe50f6 CF |
1720 | /* Determine whether or not the trip count to ahead ratio is too small based |
1721 | on prefitablility consideration. | |
db34470d | 1722 | AHEAD: the iteration ahead distance, |
0bbe50f6 CF |
1723 | EST_NITER: the estimated trip count. */ |
1724 | ||
1725 | static bool | |
1726 | trip_count_to_ahead_ratio_too_small_p (unsigned ahead, HOST_WIDE_INT est_niter) | |
1727 | { | |
1728 | /* Assume trip count to ahead ratio is big enough if the trip count could not | |
1729 | be estimated at compile time. */ | |
1730 | if (est_niter < 0) | |
1731 | return false; | |
1732 | ||
1733 | if (est_niter < (HOST_WIDE_INT) (TRIP_COUNT_TO_AHEAD_RATIO * ahead)) | |
1734 | { | |
1735 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
1736 | fprintf (dump_file, | |
1737 | "Not prefetching -- loop estimated to roll only %d times\n", | |
1738 | (int) est_niter); | |
1739 | return true; | |
1740 | } | |
1741 | ||
1742 | return false; | |
1743 | } | |
1744 | ||
1745 | /* Determine whether or not the number of memory references in the loop is | |
1746 | reasonable based on the profitablity and compilation time considerations. | |
db34470d | 1747 | NINSNS: estimated number of instructions in the loop, |
db34470d GS |
1748 | MEM_REF_COUNT: total number of memory references in the loop. */ |
1749 | ||
b8698a0f | 1750 | static bool |
0bbe50f6 | 1751 | mem_ref_count_reasonable_p (unsigned ninsns, unsigned mem_ref_count) |
db34470d | 1752 | { |
0bbe50f6 | 1753 | int insn_to_mem_ratio; |
db34470d GS |
1754 | |
1755 | if (mem_ref_count == 0) | |
1756 | return false; | |
1757 | ||
0bbe50f6 CF |
1758 | /* Miss rate computation (is_miss_rate_acceptable) and dependence analysis |
1759 | (compute_all_dependences) have high costs based on quadratic complexity. | |
1760 | To avoid huge compilation time, we give up prefetching if mem_ref_count | |
1761 | is too large. */ | |
1762 | if (mem_ref_count > PREFETCH_MAX_MEM_REFS_PER_LOOP) | |
1763 | return false; | |
1764 | ||
b8698a0f L |
1765 | /* Prefetching improves performance by overlapping cache missing |
1766 | memory accesses with CPU operations. If the loop does not have | |
1767 | enough CPU operations to overlap with memory operations, prefetching | |
1768 | won't give a significant benefit. One approximate way of checking | |
1769 | this is to require the ratio of instructions to memory references to | |
db34470d GS |
1770 | be above a certain limit. This approximation works well in practice. |
1771 | TODO: Implement a more precise computation by estimating the time | |
1772 | for each CPU or memory op in the loop. Time estimates for memory ops | |
1773 | should account for cache misses. */ | |
b8698a0f | 1774 | insn_to_mem_ratio = ninsns / mem_ref_count; |
db34470d GS |
1775 | |
1776 | if (insn_to_mem_ratio < PREFETCH_MIN_INSN_TO_MEM_RATIO) | |
55e5a2eb CF |
1777 | { |
1778 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
1779 | fprintf (dump_file, | |
1780 | "Not prefetching -- instruction to memory reference ratio (%d) too small\n", | |
1781 | insn_to_mem_ratio); | |
1782 | return false; | |
1783 | } | |
db34470d | 1784 | |
0bbe50f6 CF |
1785 | return true; |
1786 | } | |
1787 | ||
1788 | /* Determine whether or not the instruction to prefetch ratio in the loop is | |
1789 | too small based on the profitablity consideration. | |
1790 | NINSNS: estimated number of instructions in the loop, | |
1791 | PREFETCH_COUNT: an estimate of the number of prefetches, | |
1792 | UNROLL_FACTOR: the factor to unroll the loop if prefetching. */ | |
1793 | ||
1794 | static bool | |
1795 | insn_to_prefetch_ratio_too_small_p (unsigned ninsns, unsigned prefetch_count, | |
1796 | unsigned unroll_factor) | |
1797 | { | |
1798 | int insn_to_prefetch_ratio; | |
1799 | ||
d3a9b459 CF |
1800 | /* Prefetching most likely causes performance degradation when the instruction |
1801 | to prefetch ratio is too small. Too many prefetch instructions in a loop | |
1802 | may reduce the I-cache performance. | |
ccacf0e1 CF |
1803 | (unroll_factor * ninsns) is used to estimate the number of instructions in |
1804 | the unrolled loop. This implementation is a bit simplistic -- the number | |
1805 | of issued prefetch instructions is also affected by unrolling. So, | |
1806 | prefetch_mod and the unroll factor should be taken into account when | |
1807 | determining prefetch_count. Also, the number of insns of the unrolled | |
1808 | loop will usually be significantly smaller than the number of insns of the | |
1809 | original loop * unroll_factor (at least the induction variable increases | |
1810 | and the exit branches will get eliminated), so it might be better to use | |
1811 | tree_estimate_loop_size + estimated_unrolled_size. */ | |
d3a9b459 CF |
1812 | insn_to_prefetch_ratio = (unroll_factor * ninsns) / prefetch_count; |
1813 | if (insn_to_prefetch_ratio < MIN_INSN_TO_PREFETCH_RATIO) | |
db34470d | 1814 | { |
d3a9b459 CF |
1815 | if (dump_file && (dump_flags & TDF_DETAILS)) |
1816 | fprintf (dump_file, | |
1817 | "Not prefetching -- instruction to prefetch ratio (%d) too small\n", | |
1818 | insn_to_prefetch_ratio); | |
0bbe50f6 | 1819 | return true; |
db34470d | 1820 | } |
b8698a0f | 1821 | |
0bbe50f6 | 1822 | return false; |
db34470d GS |
1823 | } |
1824 | ||
1825 | ||
b076a3fd | 1826 | /* Issue prefetch instructions for array references in LOOP. Returns |
d73be268 | 1827 | true if the LOOP was unrolled. */ |
b076a3fd ZD |
1828 | |
1829 | static bool | |
d73be268 | 1830 | loop_prefetch_arrays (struct loop *loop) |
b076a3fd ZD |
1831 | { |
1832 | struct mem_ref_group *refs; | |
2711355f ZD |
1833 | unsigned ahead, ninsns, time, unroll_factor; |
1834 | HOST_WIDE_INT est_niter; | |
b076a3fd | 1835 | struct tree_niter_desc desc; |
79f5e442 | 1836 | bool unrolled = false, no_other_refs; |
db34470d GS |
1837 | unsigned prefetch_count; |
1838 | unsigned mem_ref_count; | |
b076a3fd | 1839 | |
efd8f750 | 1840 | if (optimize_loop_nest_for_size_p (loop)) |
2732d767 ZD |
1841 | { |
1842 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
1843 | fprintf (dump_file, " ignored (cold area)\n"); | |
1844 | return false; | |
1845 | } | |
1846 | ||
0bbe50f6 CF |
1847 | /* FIXME: the time should be weighted by the probabilities of the blocks in |
1848 | the loop body. */ | |
1849 | time = tree_num_loop_insns (loop, &eni_time_weights); | |
1850 | if (time == 0) | |
1851 | return false; | |
1852 | ||
1853 | ahead = (PREFETCH_LATENCY + time - 1) / time; | |
652c4c71 | 1854 | est_niter = estimated_stmt_executions_int (loop); |
e5b332cd RG |
1855 | if (est_niter == -1) |
1856 | est_niter = max_stmt_executions_int (loop); | |
0bbe50f6 CF |
1857 | |
1858 | /* Prefetching is not likely to be profitable if the trip count to ahead | |
1859 | ratio is too small. */ | |
1860 | if (trip_count_to_ahead_ratio_too_small_p (ahead, est_niter)) | |
1861 | return false; | |
1862 | ||
1863 | ninsns = tree_num_loop_insns (loop, &eni_size_weights); | |
1864 | ||
b076a3fd | 1865 | /* Step 1: gather the memory references. */ |
db34470d | 1866 | refs = gather_memory_references (loop, &no_other_refs, &mem_ref_count); |
b076a3fd | 1867 | |
0bbe50f6 CF |
1868 | /* Give up prefetching if the number of memory references in the |
1869 | loop is not reasonable based on profitablity and compilation time | |
1870 | considerations. */ | |
1871 | if (!mem_ref_count_reasonable_p (ninsns, mem_ref_count)) | |
1872 | goto fail; | |
1873 | ||
b076a3fd ZD |
1874 | /* Step 2: estimate the reuse effects. */ |
1875 | prune_by_reuse (refs); | |
1876 | ||
d5058523 | 1877 | if (nothing_to_prefetch_p (refs)) |
b076a3fd ZD |
1878 | goto fail; |
1879 | ||
1e373390 RG |
1880 | if (!determine_loop_nest_reuse (loop, refs, no_other_refs)) |
1881 | goto fail; | |
5417e022 | 1882 | |
0bbe50f6 | 1883 | /* Step 3: determine unroll factor. */ |
2711355f ZD |
1884 | unroll_factor = determine_unroll_factor (loop, refs, ninsns, &desc, |
1885 | est_niter); | |
d5058523 CF |
1886 | |
1887 | /* Estimate prefetch count for the unrolled loop. */ | |
1888 | prefetch_count = estimate_prefetch_count (refs, unroll_factor); | |
1889 | if (prefetch_count == 0) | |
1890 | goto fail; | |
1891 | ||
2711355f | 1892 | if (dump_file && (dump_flags & TDF_DETAILS)) |
b8698a0f | 1893 | fprintf (dump_file, "Ahead %d, unroll factor %d, trip count " |
d81f5387 | 1894 | HOST_WIDE_INT_PRINT_DEC "\n" |
b8698a0f L |
1895 | "insn count %d, mem ref count %d, prefetch count %d\n", |
1896 | ahead, unroll_factor, est_niter, | |
1897 | ninsns, mem_ref_count, prefetch_count); | |
db34470d | 1898 | |
0bbe50f6 CF |
1899 | /* Prefetching is not likely to be profitable if the instruction to prefetch |
1900 | ratio is too small. */ | |
1901 | if (insn_to_prefetch_ratio_too_small_p (ninsns, prefetch_count, | |
1902 | unroll_factor)) | |
db34470d GS |
1903 | goto fail; |
1904 | ||
1905 | mark_nontemporal_stores (loop, refs); | |
2711355f | 1906 | |
b076a3fd ZD |
1907 | /* Step 4: what to prefetch? */ |
1908 | if (!schedule_prefetches (refs, unroll_factor, ahead)) | |
1909 | goto fail; | |
1910 | ||
1911 | /* Step 5: unroll the loop. TODO -- peeling of first and last few | |
1912 | iterations so that we do not issue superfluous prefetches. */ | |
1913 | if (unroll_factor != 1) | |
1914 | { | |
d73be268 | 1915 | tree_unroll_loop (loop, unroll_factor, |
b076a3fd ZD |
1916 | single_dom_exit (loop), &desc); |
1917 | unrolled = true; | |
1918 | } | |
1919 | ||
1920 | /* Step 6: issue the prefetches. */ | |
1921 | issue_prefetches (refs, unroll_factor, ahead); | |
1922 | ||
1923 | fail: | |
1924 | release_mem_refs (refs); | |
1925 | return unrolled; | |
1926 | } | |
1927 | ||
d73be268 | 1928 | /* Issue prefetch instructions for array references in loops. */ |
b076a3fd | 1929 | |
c7f965b6 | 1930 | unsigned int |
d73be268 | 1931 | tree_ssa_prefetch_arrays (void) |
b076a3fd | 1932 | { |
b076a3fd ZD |
1933 | struct loop *loop; |
1934 | bool unrolled = false; | |
c7f965b6 | 1935 | int todo_flags = 0; |
b076a3fd ZD |
1936 | |
1937 | if (!HAVE_prefetch | |
1938 | /* It is possible to ask compiler for say -mtune=i486 -march=pentium4. | |
1939 | -mtune=i486 causes us having PREFETCH_BLOCK 0, since this is part | |
1940 | of processor costs and i486 does not have prefetch, but | |
1941 | -march=pentium4 causes HAVE_prefetch to be true. Ugh. */ | |
1942 | || PREFETCH_BLOCK == 0) | |
c7f965b6 | 1943 | return 0; |
b076a3fd | 1944 | |
47eb5b32 ZD |
1945 | if (dump_file && (dump_flags & TDF_DETAILS)) |
1946 | { | |
1947 | fprintf (dump_file, "Prefetching parameters:\n"); | |
1948 | fprintf (dump_file, " simultaneous prefetches: %d\n", | |
1949 | SIMULTANEOUS_PREFETCHES); | |
1950 | fprintf (dump_file, " prefetch latency: %d\n", PREFETCH_LATENCY); | |
47eb5b32 | 1951 | fprintf (dump_file, " prefetch block size: %d\n", PREFETCH_BLOCK); |
46cb0441 ZD |
1952 | fprintf (dump_file, " L1 cache size: %d lines, %d kB\n", |
1953 | L1_CACHE_SIZE_BYTES / L1_CACHE_LINE_SIZE, L1_CACHE_SIZE); | |
5417e022 | 1954 | fprintf (dump_file, " L1 cache line size: %d\n", L1_CACHE_LINE_SIZE); |
b8698a0f L |
1955 | fprintf (dump_file, " L2 cache size: %d kB\n", L2_CACHE_SIZE); |
1956 | fprintf (dump_file, " min insn-to-prefetch ratio: %d \n", | |
db34470d | 1957 | MIN_INSN_TO_PREFETCH_RATIO); |
b8698a0f | 1958 | fprintf (dump_file, " min insn-to-mem ratio: %d \n", |
db34470d | 1959 | PREFETCH_MIN_INSN_TO_MEM_RATIO); |
47eb5b32 ZD |
1960 | fprintf (dump_file, "\n"); |
1961 | } | |
1962 | ||
b076a3fd ZD |
1963 | initialize_original_copy_tables (); |
1964 | ||
e79983f4 | 1965 | if (!builtin_decl_explicit_p (BUILT_IN_PREFETCH)) |
b076a3fd | 1966 | { |
6a4825bd NF |
1967 | tree type = build_function_type_list (void_type_node, |
1968 | const_ptr_type_node, NULL_TREE); | |
c79efc4d RÁE |
1969 | tree decl = add_builtin_function ("__builtin_prefetch", type, |
1970 | BUILT_IN_PREFETCH, BUILT_IN_NORMAL, | |
1971 | NULL, NULL_TREE); | |
b076a3fd | 1972 | DECL_IS_NOVOPS (decl) = true; |
e79983f4 | 1973 | set_builtin_decl (BUILT_IN_PREFETCH, decl, false); |
b076a3fd ZD |
1974 | } |
1975 | ||
1976 | /* We assume that size of cache line is a power of two, so verify this | |
1977 | here. */ | |
1978 | gcc_assert ((PREFETCH_BLOCK & (PREFETCH_BLOCK - 1)) == 0); | |
1979 | ||
f0bd40b1 | 1980 | FOR_EACH_LOOP (loop, LI_FROM_INNERMOST) |
b076a3fd | 1981 | { |
b076a3fd ZD |
1982 | if (dump_file && (dump_flags & TDF_DETAILS)) |
1983 | fprintf (dump_file, "Processing loop %d:\n", loop->num); | |
1984 | ||
d73be268 | 1985 | unrolled |= loop_prefetch_arrays (loop); |
b076a3fd ZD |
1986 | |
1987 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
1988 | fprintf (dump_file, "\n\n"); | |
1989 | } | |
1990 | ||
1991 | if (unrolled) | |
1992 | { | |
1993 | scev_reset (); | |
c7f965b6 | 1994 | todo_flags |= TODO_cleanup_cfg; |
b076a3fd ZD |
1995 | } |
1996 | ||
1997 | free_original_copy_tables (); | |
c7f965b6 | 1998 | return todo_flags; |
b076a3fd | 1999 | } |
71343877 AM |
2000 | |
2001 | /* Prefetching. */ | |
2002 | ||
2003 | static unsigned int | |
2004 | tree_ssa_loop_prefetch (void) | |
2005 | { | |
2006 | if (number_of_loops (cfun) <= 1) | |
2007 | return 0; | |
2008 | ||
2009 | return tree_ssa_prefetch_arrays (); | |
2010 | } | |
2011 | ||
2012 | static bool | |
2013 | gate_tree_ssa_loop_prefetch (void) | |
2014 | { | |
2015 | return flag_prefetch_loop_arrays > 0; | |
2016 | } | |
2017 | ||
2018 | namespace { | |
2019 | ||
2020 | const pass_data pass_data_loop_prefetch = | |
2021 | { | |
2022 | GIMPLE_PASS, /* type */ | |
2023 | "aprefetch", /* name */ | |
2024 | OPTGROUP_LOOP, /* optinfo_flags */ | |
2025 | true, /* has_gate */ | |
2026 | true, /* has_execute */ | |
2027 | TV_TREE_PREFETCH, /* tv_id */ | |
2028 | ( PROP_cfg | PROP_ssa ), /* properties_required */ | |
2029 | 0, /* properties_provided */ | |
2030 | 0, /* properties_destroyed */ | |
2031 | 0, /* todo_flags_start */ | |
2032 | 0, /* todo_flags_finish */ | |
2033 | }; | |
2034 | ||
2035 | class pass_loop_prefetch : public gimple_opt_pass | |
2036 | { | |
2037 | public: | |
2038 | pass_loop_prefetch (gcc::context *ctxt) | |
2039 | : gimple_opt_pass (pass_data_loop_prefetch, ctxt) | |
2040 | {} | |
2041 | ||
2042 | /* opt_pass methods: */ | |
2043 | bool gate () { return gate_tree_ssa_loop_prefetch (); } | |
2044 | unsigned int execute () { return tree_ssa_loop_prefetch (); } | |
2045 | ||
2046 | }; // class pass_loop_prefetch | |
2047 | ||
2048 | } // anon namespace | |
2049 | ||
2050 | gimple_opt_pass * | |
2051 | make_pass_loop_prefetch (gcc::context *ctxt) | |
2052 | { | |
2053 | return new pass_loop_prefetch (ctxt); | |
2054 | } | |
2055 | ||
2056 |