]>
Commit | Line | Data |
---|---|---|
b076a3fd | 1 | /* Array prefetching. |
a5544970 | 2 | Copyright (C) 2005-2019 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" | |
c7131fb2 | 23 | #include "backend.h" |
957060b5 AM |
24 | #include "target.h" |
25 | #include "rtl.h" | |
b076a3fd | 26 | #include "tree.h" |
c7131fb2 | 27 | #include "gimple.h" |
957060b5 AM |
28 | #include "predict.h" |
29 | #include "tree-pass.h" | |
957060b5 | 30 | #include "gimple-ssa.h" |
957060b5 | 31 | #include "optabs-query.h" |
957060b5 | 32 | #include "tree-pretty-print.h" |
40e23961 | 33 | #include "fold-const.h" |
d8a2d370 | 34 | #include "stor-layout.h" |
45b0be94 | 35 | #include "gimplify.h" |
5be5c238 | 36 | #include "gimple-iterator.h" |
18f429e2 | 37 | #include "gimplify-me.h" |
e28030cf AM |
38 | #include "tree-ssa-loop-ivopts.h" |
39 | #include "tree-ssa-loop-manip.h" | |
40 | #include "tree-ssa-loop-niter.h" | |
442b4905 | 41 | #include "tree-ssa-loop.h" |
ed481942 | 42 | #include "ssa.h" |
442b4905 | 43 | #include "tree-into-ssa.h" |
b076a3fd | 44 | #include "cfgloop.h" |
b076a3fd | 45 | #include "tree-scalar-evolution.h" |
b076a3fd ZD |
46 | #include "params.h" |
47 | #include "langhooks.h" | |
7f9bc51b | 48 | #include "tree-inline.h" |
5417e022 | 49 | #include "tree-data-ref.h" |
d78a1c01 | 50 | #include "diagnostic-core.h" |
1a70c8d5 | 51 | #include "dbgcnt.h" |
2eb79bbb | 52 | |
b076a3fd ZD |
53 | /* This pass inserts prefetch instructions to optimize cache usage during |
54 | accesses to arrays in loops. It processes loops sequentially and: | |
55 | ||
56 | 1) Gathers all memory references in the single loop. | |
57 | 2) For each of the references it decides when it is profitable to prefetch | |
58 | it. To do it, we evaluate the reuse among the accesses, and determines | |
59 | two values: PREFETCH_BEFORE (meaning that it only makes sense to do | |
60 | prefetching in the first PREFETCH_BEFORE iterations of the loop) and | |
61 | PREFETCH_MOD (meaning that it only makes sense to prefetch in the | |
62 | iterations of the loop that are zero modulo PREFETCH_MOD). For example | |
63 | (assuming cache line size is 64 bytes, char has size 1 byte and there | |
64 | is no hardware sequential prefetch): | |
65 | ||
66 | char *a; | |
67 | for (i = 0; i < max; i++) | |
68 | { | |
69 | a[255] = ...; (0) | |
70 | a[i] = ...; (1) | |
71 | a[i + 64] = ...; (2) | |
72 | a[16*i] = ...; (3) | |
73 | a[187*i] = ...; (4) | |
74 | a[187*i + 50] = ...; (5) | |
75 | } | |
76 | ||
77 | (0) obviously has PREFETCH_BEFORE 1 | |
78 | (1) has PREFETCH_BEFORE 64, since (2) accesses the same memory | |
79 | location 64 iterations before it, and PREFETCH_MOD 64 (since | |
80 | it hits the same cache line otherwise). | |
81 | (2) has PREFETCH_MOD 64 | |
82 | (3) has PREFETCH_MOD 4 | |
83 | (4) has PREFETCH_MOD 1. We do not set PREFETCH_BEFORE here, since | |
bae077dc | 84 | the cache line accessed by (5) is the same with probability only |
b076a3fd ZD |
85 | 7/32. |
86 | (5) has PREFETCH_MOD 1 as well. | |
87 | ||
5417e022 ZD |
88 | Additionally, we use data dependence analysis to determine for each |
89 | reference the distance till the first reuse; this information is used | |
90 | to determine the temporality of the issued prefetch instruction. | |
91 | ||
b076a3fd ZD |
92 | 3) We determine how much ahead we need to prefetch. The number of |
93 | iterations needed is time to fetch / time spent in one iteration of | |
94 | the loop. The problem is that we do not know either of these values, | |
95 | so we just make a heuristic guess based on a magic (possibly) | |
96 | target-specific constant and size of the loop. | |
97 | ||
98 | 4) Determine which of the references we prefetch. We take into account | |
99 | that there is a maximum number of simultaneous prefetches (provided | |
100 | by machine description). We prefetch as many prefetches as possible | |
101 | while still within this bound (starting with those with lowest | |
102 | prefetch_mod, since they are responsible for most of the cache | |
103 | misses). | |
b8698a0f | 104 | |
b076a3fd ZD |
105 | 5) We unroll and peel loops so that we are able to satisfy PREFETCH_MOD |
106 | and PREFETCH_BEFORE requirements (within some bounds), and to avoid | |
107 | prefetching nonaccessed memory. | |
108 | TODO -- actually implement peeling. | |
b8698a0f | 109 | |
b076a3fd ZD |
110 | 6) We actually emit the prefetch instructions. ??? Perhaps emit the |
111 | prefetch instructions with guards in cases where 5) was not sufficient | |
112 | to satisfy the constraints? | |
113 | ||
0bbe50f6 CF |
114 | A cost model is implemented to determine whether or not prefetching is |
115 | profitable for a given loop. The cost model has three heuristics: | |
116 | ||
117 | 1. Function trip_count_to_ahead_ratio_too_small_p implements a | |
118 | heuristic that determines whether or not the loop has too few | |
119 | iterations (compared to ahead). Prefetching is not likely to be | |
120 | beneficial if the trip count to ahead ratio is below a certain | |
121 | minimum. | |
122 | ||
123 | 2. Function mem_ref_count_reasonable_p implements a heuristic that | |
124 | determines whether the given loop has enough CPU ops that can be | |
125 | overlapped with cache missing memory ops. If not, the loop | |
126 | won't benefit from prefetching. In the implementation, | |
127 | prefetching is not considered beneficial if the ratio between | |
128 | the instruction count and the mem ref count is below a certain | |
129 | minimum. | |
130 | ||
131 | 3. Function insn_to_prefetch_ratio_too_small_p implements a | |
132 | heuristic that disables prefetching in a loop if the prefetching | |
133 | cost is above a certain limit. The relative prefetching cost is | |
134 | estimated by taking the ratio between the prefetch count and the | |
135 | total intruction count (this models the I-cache cost). | |
136 | ||
db34470d | 137 | The limits used in these heuristics are defined as parameters with |
b8698a0f | 138 | reasonable default values. Machine-specific default values will be |
db34470d | 139 | added later. |
b8698a0f | 140 | |
b076a3fd ZD |
141 | Some other TODO: |
142 | -- write and use more general reuse analysis (that could be also used | |
143 | in other cache aimed loop optimizations) | |
144 | -- make it behave sanely together with the prefetches given by user | |
145 | (now we just ignore them; at the very least we should avoid | |
146 | optimizing loops in that user put his own prefetches) | |
147 | -- we assume cache line size alignment of arrays; this could be | |
148 | improved. */ | |
149 | ||
150 | /* Magic constants follow. These should be replaced by machine specific | |
151 | numbers. */ | |
152 | ||
b076a3fd ZD |
153 | /* True if write can be prefetched by a read prefetch. */ |
154 | ||
155 | #ifndef WRITE_CAN_USE_READ_PREFETCH | |
156 | #define WRITE_CAN_USE_READ_PREFETCH 1 | |
157 | #endif | |
158 | ||
159 | /* True if read can be prefetched by a write prefetch. */ | |
160 | ||
161 | #ifndef READ_CAN_USE_WRITE_PREFETCH | |
162 | #define READ_CAN_USE_WRITE_PREFETCH 0 | |
163 | #endif | |
164 | ||
47eb5b32 ZD |
165 | /* The size of the block loaded by a single prefetch. Usually, this is |
166 | the same as cache line size (at the moment, we only consider one level | |
167 | of cache hierarchy). */ | |
b076a3fd ZD |
168 | |
169 | #ifndef PREFETCH_BLOCK | |
47eb5b32 | 170 | #define PREFETCH_BLOCK L1_CACHE_LINE_SIZE |
b076a3fd ZD |
171 | #endif |
172 | ||
173 | /* Do we have a forward hardware sequential prefetching? */ | |
174 | ||
175 | #ifndef HAVE_FORWARD_PREFETCH | |
176 | #define HAVE_FORWARD_PREFETCH 0 | |
177 | #endif | |
178 | ||
179 | /* Do we have a backward hardware sequential prefetching? */ | |
180 | ||
181 | #ifndef HAVE_BACKWARD_PREFETCH | |
182 | #define HAVE_BACKWARD_PREFETCH 0 | |
183 | #endif | |
184 | ||
185 | /* In some cases we are only able to determine that there is a certain | |
186 | probability that the two accesses hit the same cache line. In this | |
187 | case, we issue the prefetches for both of them if this probability | |
fa10beec | 188 | is less then (1000 - ACCEPTABLE_MISS_RATE) per thousand. */ |
b076a3fd ZD |
189 | |
190 | #ifndef ACCEPTABLE_MISS_RATE | |
191 | #define ACCEPTABLE_MISS_RATE 50 | |
192 | #endif | |
193 | ||
46cb0441 ZD |
194 | #define L1_CACHE_SIZE_BYTES ((unsigned) (L1_CACHE_SIZE * 1024)) |
195 | #define L2_CACHE_SIZE_BYTES ((unsigned) (L2_CACHE_SIZE * 1024)) | |
5417e022 ZD |
196 | |
197 | /* We consider a memory access nontemporal if it is not reused sooner than | |
198 | after L2_CACHE_SIZE_BYTES of memory are accessed. However, we ignore | |
199 | accesses closer than L1_CACHE_SIZE_BYTES / NONTEMPORAL_FRACTION, | |
200 | so that we use nontemporal prefetches e.g. if single memory location | |
201 | is accessed several times in a single iteration of the loop. */ | |
202 | #define NONTEMPORAL_FRACTION 16 | |
203 | ||
79f5e442 ZD |
204 | /* In case we have to emit a memory fence instruction after the loop that |
205 | uses nontemporal stores, this defines the builtin to use. */ | |
206 | ||
207 | #ifndef FENCE_FOLLOWING_MOVNT | |
208 | #define FENCE_FOLLOWING_MOVNT NULL_TREE | |
209 | #endif | |
210 | ||
9bf4598b CF |
211 | /* It is not profitable to prefetch when the trip count is not at |
212 | least TRIP_COUNT_TO_AHEAD_RATIO times the prefetch ahead distance. | |
213 | For example, in a loop with a prefetch ahead distance of 10, | |
214 | supposing that TRIP_COUNT_TO_AHEAD_RATIO is equal to 4, it is | |
215 | profitable to prefetch when the trip count is greater or equal to | |
216 | 40. In that case, 30 out of the 40 iterations will benefit from | |
217 | prefetching. */ | |
218 | ||
219 | #ifndef TRIP_COUNT_TO_AHEAD_RATIO | |
220 | #define TRIP_COUNT_TO_AHEAD_RATIO 4 | |
221 | #endif | |
222 | ||
b076a3fd ZD |
223 | /* The group of references between that reuse may occur. */ |
224 | ||
225 | struct mem_ref_group | |
226 | { | |
227 | tree base; /* Base of the reference. */ | |
81f32326 | 228 | tree step; /* Step of the reference. */ |
b076a3fd ZD |
229 | struct mem_ref *refs; /* References in the group. */ |
230 | struct mem_ref_group *next; /* Next group of references. */ | |
23b0f9f8 | 231 | unsigned int uid; /* Group UID, used only for debugging. */ |
b076a3fd ZD |
232 | }; |
233 | ||
234 | /* Assigned to PREFETCH_BEFORE when all iterations are to be prefetched. */ | |
235 | ||
dd4786fe | 236 | #define PREFETCH_ALL HOST_WIDE_INT_M1U |
b076a3fd | 237 | |
8532678c CF |
238 | /* Do not generate a prefetch if the unroll factor is significantly less |
239 | than what is required by the prefetch. This is to avoid redundant | |
f7963a7c CF |
240 | prefetches. For example, when prefetch_mod is 16 and unroll_factor is |
241 | 2, prefetching requires unrolling the loop 16 times, but | |
242 | the loop is actually unrolled twice. In this case (ratio = 8), | |
8532678c CF |
243 | prefetching is not likely to be beneficial. */ |
244 | ||
245 | #ifndef PREFETCH_MOD_TO_UNROLL_FACTOR_RATIO | |
f7963a7c | 246 | #define PREFETCH_MOD_TO_UNROLL_FACTOR_RATIO 4 |
8532678c CF |
247 | #endif |
248 | ||
0bbe50f6 CF |
249 | /* Some of the prefetch computations have quadratic complexity. We want to |
250 | avoid huge compile times and, therefore, want to limit the amount of | |
251 | memory references per loop where we consider prefetching. */ | |
252 | ||
253 | #ifndef PREFETCH_MAX_MEM_REFS_PER_LOOP | |
254 | #define PREFETCH_MAX_MEM_REFS_PER_LOOP 200 | |
255 | #endif | |
256 | ||
b076a3fd ZD |
257 | /* The memory reference. */ |
258 | ||
259 | struct mem_ref | |
260 | { | |
355fe088 | 261 | gimple *stmt; /* Statement in that the reference appears. */ |
b076a3fd ZD |
262 | tree mem; /* The reference. */ |
263 | HOST_WIDE_INT delta; /* Constant offset of the reference. */ | |
b076a3fd ZD |
264 | struct mem_ref_group *group; /* The group of references it belongs to. */ |
265 | unsigned HOST_WIDE_INT prefetch_mod; | |
266 | /* Prefetch only each PREFETCH_MOD-th | |
267 | iteration. */ | |
268 | unsigned HOST_WIDE_INT prefetch_before; | |
269 | /* Prefetch only first PREFETCH_BEFORE | |
270 | iterations. */ | |
5417e022 ZD |
271 | unsigned reuse_distance; /* The amount of data accessed before the first |
272 | reuse of this value. */ | |
b076a3fd | 273 | struct mem_ref *next; /* The next reference in the group. */ |
23b0f9f8 | 274 | unsigned int uid; /* Ref UID, used only for debugging. */ |
79f5e442 ZD |
275 | unsigned write_p : 1; /* Is it a write? */ |
276 | unsigned independent_p : 1; /* True if the reference is independent on | |
277 | all other references inside the loop. */ | |
278 | unsigned issue_prefetch_p : 1; /* Should we really issue the prefetch? */ | |
279 | unsigned storent_p : 1; /* True if we changed the store to a | |
280 | nontemporal one. */ | |
b076a3fd ZD |
281 | }; |
282 | ||
a5497b12 | 283 | /* Dumps information about memory reference */ |
b076a3fd | 284 | static void |
a5497b12 VK |
285 | dump_mem_details (FILE *file, tree base, tree step, |
286 | HOST_WIDE_INT delta, bool write_p) | |
b076a3fd | 287 | { |
a5497b12 VK |
288 | fprintf (file, "(base "); |
289 | print_generic_expr (file, base, TDF_SLIM); | |
b076a3fd | 290 | fprintf (file, ", step "); |
a5497b12 VK |
291 | if (cst_and_fits_in_hwi (step)) |
292 | fprintf (file, HOST_WIDE_INT_PRINT_DEC, int_cst_value (step)); | |
81f32326 | 293 | else |
8264c84d | 294 | print_generic_expr (file, step, TDF_SLIM); |
b076a3fd | 295 | fprintf (file, ")\n"); |
23b0f9f8 MK |
296 | fprintf (file, " delta " HOST_WIDE_INT_PRINT_DEC "\n", delta); |
297 | fprintf (file, " %s\n\n", write_p ? "write" : "read"); | |
a5497b12 | 298 | } |
b076a3fd | 299 | |
a5497b12 | 300 | /* Dumps information about reference REF to FILE. */ |
b076a3fd | 301 | |
a5497b12 VK |
302 | static void |
303 | dump_mem_ref (FILE *file, struct mem_ref *ref) | |
304 | { | |
23b0f9f8 MK |
305 | fprintf (file, "reference %u:%u (", ref->group->uid, ref->uid); |
306 | print_generic_expr (file, ref->mem, TDF_SLIM); | |
307 | fprintf (file, ")\n"); | |
b076a3fd ZD |
308 | } |
309 | ||
310 | /* Finds a group with BASE and STEP in GROUPS, or creates one if it does not | |
311 | exist. */ | |
312 | ||
313 | static struct mem_ref_group * | |
81f32326 | 314 | find_or_create_group (struct mem_ref_group **groups, tree base, tree step) |
b076a3fd | 315 | { |
23b0f9f8 MK |
316 | /* Global count for setting struct mem_ref_group->uid. */ |
317 | static unsigned int last_mem_ref_group_uid = 0; | |
318 | ||
b076a3fd ZD |
319 | struct mem_ref_group *group; |
320 | ||
321 | for (; *groups; groups = &(*groups)->next) | |
322 | { | |
81f32326 | 323 | if (operand_equal_p ((*groups)->step, step, 0) |
b076a3fd ZD |
324 | && operand_equal_p ((*groups)->base, base, 0)) |
325 | return *groups; | |
326 | ||
81f32326 CB |
327 | /* If step is an integer constant, keep the list of groups sorted |
328 | by decreasing step. */ | |
21c0a521 DM |
329 | if (cst_and_fits_in_hwi ((*groups)->step) && cst_and_fits_in_hwi (step) |
330 | && int_cst_value ((*groups)->step) < int_cst_value (step)) | |
b076a3fd ZD |
331 | break; |
332 | } | |
333 | ||
5417e022 | 334 | group = XNEW (struct mem_ref_group); |
b076a3fd ZD |
335 | group->base = base; |
336 | group->step = step; | |
337 | group->refs = NULL; | |
23b0f9f8 | 338 | group->uid = ++last_mem_ref_group_uid; |
b076a3fd ZD |
339 | group->next = *groups; |
340 | *groups = group; | |
341 | ||
342 | return group; | |
343 | } | |
344 | ||
345 | /* Records a memory reference MEM in GROUP with offset DELTA and write status | |
346 | WRITE_P. The reference occurs in statement STMT. */ | |
347 | ||
348 | static void | |
355fe088 | 349 | record_ref (struct mem_ref_group *group, gimple *stmt, tree mem, |
b076a3fd ZD |
350 | HOST_WIDE_INT delta, bool write_p) |
351 | { | |
23b0f9f8 | 352 | unsigned int last_mem_ref_uid = 0; |
b076a3fd ZD |
353 | struct mem_ref **aref; |
354 | ||
355 | /* Do not record the same address twice. */ | |
356 | for (aref = &group->refs; *aref; aref = &(*aref)->next) | |
357 | { | |
23b0f9f8 MK |
358 | last_mem_ref_uid = (*aref)->uid; |
359 | ||
b076a3fd ZD |
360 | /* It does not have to be possible for write reference to reuse the read |
361 | prefetch, or vice versa. */ | |
362 | if (!WRITE_CAN_USE_READ_PREFETCH | |
363 | && write_p | |
364 | && !(*aref)->write_p) | |
365 | continue; | |
366 | if (!READ_CAN_USE_WRITE_PREFETCH | |
367 | && !write_p | |
368 | && (*aref)->write_p) | |
369 | continue; | |
370 | ||
371 | if ((*aref)->delta == delta) | |
372 | return; | |
373 | } | |
374 | ||
5417e022 | 375 | (*aref) = XNEW (struct mem_ref); |
b076a3fd ZD |
376 | (*aref)->stmt = stmt; |
377 | (*aref)->mem = mem; | |
378 | (*aref)->delta = delta; | |
379 | (*aref)->write_p = write_p; | |
380 | (*aref)->prefetch_before = PREFETCH_ALL; | |
381 | (*aref)->prefetch_mod = 1; | |
5417e022 | 382 | (*aref)->reuse_distance = 0; |
b076a3fd ZD |
383 | (*aref)->issue_prefetch_p = false; |
384 | (*aref)->group = group; | |
385 | (*aref)->next = NULL; | |
79f5e442 ZD |
386 | (*aref)->independent_p = false; |
387 | (*aref)->storent_p = false; | |
23b0f9f8 | 388 | (*aref)->uid = last_mem_ref_uid + 1; |
b076a3fd ZD |
389 | |
390 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
23b0f9f8 MK |
391 | { |
392 | dump_mem_ref (dump_file, *aref); | |
393 | ||
394 | fprintf (dump_file, " group %u ", group->uid); | |
395 | dump_mem_details (dump_file, group->base, group->step, delta, | |
396 | write_p); | |
397 | } | |
b076a3fd ZD |
398 | } |
399 | ||
400 | /* Release memory references in GROUPS. */ | |
401 | ||
402 | static void | |
403 | release_mem_refs (struct mem_ref_group *groups) | |
404 | { | |
405 | struct mem_ref_group *next_g; | |
406 | struct mem_ref *ref, *next_r; | |
407 | ||
408 | for (; groups; groups = next_g) | |
409 | { | |
410 | next_g = groups->next; | |
411 | for (ref = groups->refs; ref; ref = next_r) | |
412 | { | |
413 | next_r = ref->next; | |
414 | free (ref); | |
415 | } | |
416 | free (groups); | |
417 | } | |
418 | } | |
419 | ||
420 | /* A structure used to pass arguments to idx_analyze_ref. */ | |
421 | ||
422 | struct ar_data | |
423 | { | |
424 | struct loop *loop; /* Loop of the reference. */ | |
355fe088 | 425 | gimple *stmt; /* Statement of the reference. */ |
81f32326 | 426 | tree *step; /* Step of the memory reference. */ |
b076a3fd ZD |
427 | HOST_WIDE_INT *delta; /* Offset of the memory reference. */ |
428 | }; | |
429 | ||
430 | /* Analyzes a single INDEX of a memory reference to obtain information | |
431 | described at analyze_ref. Callback for for_each_index. */ | |
432 | ||
433 | static bool | |
434 | idx_analyze_ref (tree base, tree *index, void *data) | |
435 | { | |
c22940cd | 436 | struct ar_data *ar_data = (struct ar_data *) data; |
b076a3fd | 437 | tree ibase, step, stepsize; |
81f32326 | 438 | HOST_WIDE_INT idelta = 0, imult = 1; |
b076a3fd ZD |
439 | affine_iv iv; |
440 | ||
f017bf5e | 441 | if (!simple_iv (ar_data->loop, loop_containing_stmt (ar_data->stmt), |
81f32326 | 442 | *index, &iv, true)) |
b076a3fd ZD |
443 | return false; |
444 | ibase = iv.base; | |
445 | step = iv.step; | |
446 | ||
5be014d5 | 447 | if (TREE_CODE (ibase) == POINTER_PLUS_EXPR |
b076a3fd ZD |
448 | && cst_and_fits_in_hwi (TREE_OPERAND (ibase, 1))) |
449 | { | |
450 | idelta = int_cst_value (TREE_OPERAND (ibase, 1)); | |
451 | ibase = TREE_OPERAND (ibase, 0); | |
452 | } | |
453 | if (cst_and_fits_in_hwi (ibase)) | |
454 | { | |
455 | idelta += int_cst_value (ibase); | |
ff5e9a94 | 456 | ibase = build_int_cst (TREE_TYPE (ibase), 0); |
b076a3fd ZD |
457 | } |
458 | ||
459 | if (TREE_CODE (base) == ARRAY_REF) | |
460 | { | |
461 | stepsize = array_ref_element_size (base); | |
462 | if (!cst_and_fits_in_hwi (stepsize)) | |
463 | return false; | |
464 | imult = int_cst_value (stepsize); | |
8fde8b40 CB |
465 | step = fold_build2 (MULT_EXPR, sizetype, |
466 | fold_convert (sizetype, step), | |
467 | fold_convert (sizetype, stepsize)); | |
b076a3fd ZD |
468 | idelta *= imult; |
469 | } | |
470 | ||
8fde8b40 CB |
471 | if (*ar_data->step == NULL_TREE) |
472 | *ar_data->step = step; | |
473 | else | |
474 | *ar_data->step = fold_build2 (PLUS_EXPR, sizetype, | |
475 | fold_convert (sizetype, *ar_data->step), | |
476 | fold_convert (sizetype, step)); | |
b076a3fd ZD |
477 | *ar_data->delta += idelta; |
478 | *index = ibase; | |
479 | ||
480 | return true; | |
481 | } | |
482 | ||
aac8b8ed | 483 | /* Tries to express REF_P in shape &BASE + STEP * iter + DELTA, where DELTA and |
b076a3fd | 484 | STEP are integer constants and iter is number of iterations of LOOP. The |
aac8b8ed RS |
485 | reference occurs in statement STMT. Strips nonaddressable component |
486 | references from REF_P. */ | |
b076a3fd ZD |
487 | |
488 | static bool | |
aac8b8ed | 489 | analyze_ref (struct loop *loop, tree *ref_p, tree *base, |
81f32326 | 490 | tree *step, HOST_WIDE_INT *delta, |
355fe088 | 491 | gimple *stmt) |
b076a3fd ZD |
492 | { |
493 | struct ar_data ar_data; | |
494 | tree off; | |
495 | HOST_WIDE_INT bit_offset; | |
aac8b8ed | 496 | tree ref = *ref_p; |
b076a3fd | 497 | |
81f32326 | 498 | *step = NULL_TREE; |
b076a3fd ZD |
499 | *delta = 0; |
500 | ||
7c6dafac CF |
501 | /* First strip off the component references. Ignore bitfields. |
502 | Also strip off the real and imagine parts of a complex, so that | |
503 | they can have the same base. */ | |
504 | if (TREE_CODE (ref) == REALPART_EXPR | |
505 | || TREE_CODE (ref) == IMAGPART_EXPR | |
506 | || (TREE_CODE (ref) == COMPONENT_REF | |
507 | && DECL_NONADDRESSABLE_P (TREE_OPERAND (ref, 1)))) | |
508 | { | |
509 | if (TREE_CODE (ref) == IMAGPART_EXPR) | |
510 | *delta += int_size_in_bytes (TREE_TYPE (ref)); | |
511 | ref = TREE_OPERAND (ref, 0); | |
512 | } | |
b076a3fd | 513 | |
aac8b8ed RS |
514 | *ref_p = ref; |
515 | ||
b076a3fd ZD |
516 | for (; TREE_CODE (ref) == COMPONENT_REF; ref = TREE_OPERAND (ref, 0)) |
517 | { | |
518 | off = DECL_FIELD_BIT_OFFSET (TREE_OPERAND (ref, 1)); | |
519 | bit_offset = TREE_INT_CST_LOW (off); | |
520 | gcc_assert (bit_offset % BITS_PER_UNIT == 0); | |
b8698a0f | 521 | |
b076a3fd ZD |
522 | *delta += bit_offset / BITS_PER_UNIT; |
523 | } | |
524 | ||
525 | *base = unshare_expr (ref); | |
526 | ar_data.loop = loop; | |
527 | ar_data.stmt = stmt; | |
528 | ar_data.step = step; | |
529 | ar_data.delta = delta; | |
530 | return for_each_index (base, idx_analyze_ref, &ar_data); | |
531 | } | |
532 | ||
533 | /* Record a memory reference REF to the list REFS. The reference occurs in | |
79f5e442 ZD |
534 | LOOP in statement STMT and it is write if WRITE_P. Returns true if the |
535 | reference was recorded, false otherwise. */ | |
b076a3fd | 536 | |
79f5e442 | 537 | static bool |
b076a3fd | 538 | gather_memory_references_ref (struct loop *loop, struct mem_ref_group **refs, |
355fe088 | 539 | tree ref, bool write_p, gimple *stmt) |
b076a3fd | 540 | { |
81f32326 CB |
541 | tree base, step; |
542 | HOST_WIDE_INT delta; | |
b076a3fd ZD |
543 | struct mem_ref_group *agrp; |
544 | ||
a80a2701 JJ |
545 | if (get_base_address (ref) == NULL) |
546 | return false; | |
547 | ||
aac8b8ed | 548 | if (!analyze_ref (loop, &ref, &base, &step, &delta, stmt)) |
79f5e442 | 549 | return false; |
81f32326 CB |
550 | /* If analyze_ref fails the default is a NULL_TREE. We can stop here. */ |
551 | if (step == NULL_TREE) | |
552 | return false; | |
b076a3fd | 553 | |
756f50ce | 554 | /* Stop if the address of BASE could not be taken. */ |
bc068a23 CF |
555 | if (may_be_nonaddressable_p (base)) |
556 | return false; | |
557 | ||
a5497b12 VK |
558 | /* Limit non-constant step prefetching only to the innermost loops and |
559 | only when the step is loop invariant in the entire loop nest. */ | |
560 | if (!cst_and_fits_in_hwi (step)) | |
561 | { | |
562 | if (loop->inner != NULL) | |
563 | { | |
564 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
565 | { | |
566 | fprintf (dump_file, "Memory expression %p\n",(void *) ref ); | |
8264c84d ML |
567 | print_generic_expr (dump_file, ref, TDF_SLIM); |
568 | fprintf (dump_file,":"); | |
c3284718 | 569 | dump_mem_details (dump_file, base, step, delta, write_p); |
a5497b12 VK |
570 | fprintf (dump_file, |
571 | "Ignoring %p, non-constant step prefetching is " | |
572 | "limited to inner most loops \n", | |
573 | (void *) ref); | |
574 | } | |
575 | return false; | |
576 | } | |
577 | else | |
578 | { | |
579 | if (!expr_invariant_in_loop_p (loop_outermost (loop), step)) | |
580 | { | |
581 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
582 | { | |
583 | fprintf (dump_file, "Memory expression %p\n",(void *) ref ); | |
8264c84d | 584 | print_generic_expr (dump_file, ref, TDF_SLIM); |
a5497b12 | 585 | fprintf (dump_file,":"); |
c3284718 | 586 | dump_mem_details (dump_file, base, step, delta, write_p); |
a5497b12 VK |
587 | fprintf (dump_file, |
588 | "Not prefetching, ignoring %p due to " | |
589 | "loop variant step\n", | |
590 | (void *) ref); | |
591 | } | |
592 | return false; | |
593 | } | |
594 | } | |
595 | } | |
50814135 | 596 | |
b076a3fd ZD |
597 | /* Now we know that REF = &BASE + STEP * iter + DELTA, where DELTA and STEP |
598 | are integer constants. */ | |
599 | agrp = find_or_create_group (refs, base, step); | |
600 | record_ref (agrp, stmt, ref, delta, write_p); | |
79f5e442 ZD |
601 | |
602 | return true; | |
b076a3fd ZD |
603 | } |
604 | ||
79f5e442 ZD |
605 | /* Record the suitable memory references in LOOP. NO_OTHER_REFS is set to |
606 | true if there are no other memory references inside the loop. */ | |
b076a3fd ZD |
607 | |
608 | static struct mem_ref_group * | |
db34470d | 609 | gather_memory_references (struct loop *loop, bool *no_other_refs, unsigned *ref_count) |
b076a3fd ZD |
610 | { |
611 | basic_block *body = get_loop_body_in_dom_order (loop); | |
612 | basic_block bb; | |
613 | unsigned i; | |
726a989a | 614 | gimple_stmt_iterator bsi; |
355fe088 | 615 | gimple *stmt; |
726a989a | 616 | tree lhs, rhs; |
b076a3fd ZD |
617 | struct mem_ref_group *refs = NULL; |
618 | ||
79f5e442 | 619 | *no_other_refs = true; |
db34470d | 620 | *ref_count = 0; |
79f5e442 | 621 | |
b076a3fd ZD |
622 | /* Scan the loop body in order, so that the former references precede the |
623 | later ones. */ | |
624 | for (i = 0; i < loop->num_nodes; i++) | |
625 | { | |
626 | bb = body[i]; | |
627 | if (bb->loop_father != loop) | |
628 | continue; | |
629 | ||
726a989a | 630 | for (bsi = gsi_start_bb (bb); !gsi_end_p (bsi); gsi_next (&bsi)) |
b076a3fd | 631 | { |
726a989a | 632 | stmt = gsi_stmt (bsi); |
79f5e442 | 633 | |
726a989a | 634 | if (gimple_code (stmt) != GIMPLE_ASSIGN) |
79f5e442 | 635 | { |
5006671f | 636 | if (gimple_vuse (stmt) |
726a989a RB |
637 | || (is_gimple_call (stmt) |
638 | && !(gimple_call_flags (stmt) & ECF_CONST))) | |
79f5e442 ZD |
639 | *no_other_refs = false; |
640 | continue; | |
641 | } | |
b076a3fd | 642 | |
eb066284 RB |
643 | if (! gimple_vuse (stmt)) |
644 | continue; | |
645 | ||
726a989a RB |
646 | lhs = gimple_assign_lhs (stmt); |
647 | rhs = gimple_assign_rhs1 (stmt); | |
b076a3fd ZD |
648 | |
649 | if (REFERENCE_CLASS_P (rhs)) | |
db34470d | 650 | { |
79f5e442 ZD |
651 | *no_other_refs &= gather_memory_references_ref (loop, &refs, |
652 | rhs, false, stmt); | |
db34470d GS |
653 | *ref_count += 1; |
654 | } | |
b076a3fd | 655 | if (REFERENCE_CLASS_P (lhs)) |
db34470d | 656 | { |
79f5e442 ZD |
657 | *no_other_refs &= gather_memory_references_ref (loop, &refs, |
658 | lhs, true, stmt); | |
db34470d GS |
659 | *ref_count += 1; |
660 | } | |
b076a3fd ZD |
661 | } |
662 | } | |
663 | free (body); | |
664 | ||
665 | return refs; | |
666 | } | |
667 | ||
668 | /* Prune the prefetch candidate REF using the self-reuse. */ | |
669 | ||
670 | static void | |
671 | prune_ref_by_self_reuse (struct mem_ref *ref) | |
672 | { | |
81f32326 CB |
673 | HOST_WIDE_INT step; |
674 | bool backward; | |
675 | ||
676 | /* If the step size is non constant, we cannot calculate prefetch_mod. */ | |
677 | if (!cst_and_fits_in_hwi (ref->group->step)) | |
678 | return; | |
679 | ||
680 | step = int_cst_value (ref->group->step); | |
681 | ||
682 | backward = step < 0; | |
b076a3fd ZD |
683 | |
684 | if (step == 0) | |
685 | { | |
686 | /* Prefetch references to invariant address just once. */ | |
687 | ref->prefetch_before = 1; | |
688 | return; | |
689 | } | |
690 | ||
691 | if (backward) | |
692 | step = -step; | |
693 | ||
694 | if (step > PREFETCH_BLOCK) | |
695 | return; | |
696 | ||
697 | if ((backward && HAVE_BACKWARD_PREFETCH) | |
698 | || (!backward && HAVE_FORWARD_PREFETCH)) | |
699 | { | |
700 | ref->prefetch_before = 1; | |
701 | return; | |
702 | } | |
703 | ||
704 | ref->prefetch_mod = PREFETCH_BLOCK / step; | |
705 | } | |
706 | ||
707 | /* Divides X by BY, rounding down. */ | |
708 | ||
709 | static HOST_WIDE_INT | |
710 | ddown (HOST_WIDE_INT x, unsigned HOST_WIDE_INT by) | |
711 | { | |
712 | gcc_assert (by > 0); | |
713 | ||
714 | if (x >= 0) | |
56e1a4d7 | 715 | return x / (HOST_WIDE_INT) by; |
b076a3fd | 716 | else |
56e1a4d7 | 717 | return (x + (HOST_WIDE_INT) by - 1) / (HOST_WIDE_INT) by; |
b076a3fd ZD |
718 | } |
719 | ||
b8698a0f L |
720 | /* Given a CACHE_LINE_SIZE and two inductive memory references |
721 | with a common STEP greater than CACHE_LINE_SIZE and an address | |
722 | difference DELTA, compute the probability that they will fall | |
14e444c3 CF |
723 | in different cache lines. Return true if the computed miss rate |
724 | is not greater than the ACCEPTABLE_MISS_RATE. DISTINCT_ITERS is the | |
725 | number of distinct iterations after which the pattern repeats itself. | |
2c6dd136 GS |
726 | ALIGN_UNIT is the unit of alignment in bytes. */ |
727 | ||
14e444c3 CF |
728 | static bool |
729 | is_miss_rate_acceptable (unsigned HOST_WIDE_INT cache_line_size, | |
2c6dd136 GS |
730 | HOST_WIDE_INT step, HOST_WIDE_INT delta, |
731 | unsigned HOST_WIDE_INT distinct_iters, | |
732 | int align_unit) | |
733 | { | |
734 | unsigned align, iter; | |
14e444c3 | 735 | int total_positions, miss_positions, max_allowed_miss_positions; |
2c6dd136 GS |
736 | int address1, address2, cache_line1, cache_line2; |
737 | ||
a245c04b CF |
738 | /* It always misses if delta is greater than or equal to the cache |
739 | line size. */ | |
14e444c3 CF |
740 | if (delta >= (HOST_WIDE_INT) cache_line_size) |
741 | return false; | |
a245c04b | 742 | |
2c6dd136 | 743 | miss_positions = 0; |
14e444c3 CF |
744 | total_positions = (cache_line_size / align_unit) * distinct_iters; |
745 | max_allowed_miss_positions = (ACCEPTABLE_MISS_RATE * total_positions) / 1000; | |
b8698a0f | 746 | |
2c6dd136 GS |
747 | /* Iterate through all possible alignments of the first |
748 | memory reference within its cache line. */ | |
749 | for (align = 0; align < cache_line_size; align += align_unit) | |
750 | ||
751 | /* Iterate through all distinct iterations. */ | |
752 | for (iter = 0; iter < distinct_iters; iter++) | |
753 | { | |
754 | address1 = align + step * iter; | |
755 | address2 = address1 + delta; | |
756 | cache_line1 = address1 / cache_line_size; | |
757 | cache_line2 = address2 / cache_line_size; | |
2c6dd136 | 758 | if (cache_line1 != cache_line2) |
14e444c3 CF |
759 | { |
760 | miss_positions += 1; | |
761 | if (miss_positions > max_allowed_miss_positions) | |
762 | return false; | |
763 | } | |
2c6dd136 | 764 | } |
14e444c3 | 765 | return true; |
2c6dd136 GS |
766 | } |
767 | ||
b076a3fd ZD |
768 | /* Prune the prefetch candidate REF using the reuse with BY. |
769 | If BY_IS_BEFORE is true, BY is before REF in the loop. */ | |
770 | ||
771 | static void | |
772 | prune_ref_by_group_reuse (struct mem_ref *ref, struct mem_ref *by, | |
773 | bool by_is_before) | |
774 | { | |
81f32326 CB |
775 | HOST_WIDE_INT step; |
776 | bool backward; | |
b076a3fd ZD |
777 | HOST_WIDE_INT delta_r = ref->delta, delta_b = by->delta; |
778 | HOST_WIDE_INT delta = delta_b - delta_r; | |
779 | HOST_WIDE_INT hit_from; | |
780 | unsigned HOST_WIDE_INT prefetch_before, prefetch_block; | |
2c6dd136 GS |
781 | HOST_WIDE_INT reduced_step; |
782 | unsigned HOST_WIDE_INT reduced_prefetch_block; | |
783 | tree ref_type; | |
784 | int align_unit; | |
b076a3fd | 785 | |
81f32326 CB |
786 | /* If the step is non constant we cannot calculate prefetch_before. */ |
787 | if (!cst_and_fits_in_hwi (ref->group->step)) { | |
788 | return; | |
789 | } | |
790 | ||
791 | step = int_cst_value (ref->group->step); | |
792 | ||
793 | backward = step < 0; | |
794 | ||
795 | ||
b076a3fd ZD |
796 | if (delta == 0) |
797 | { | |
798 | /* If the references has the same address, only prefetch the | |
799 | former. */ | |
800 | if (by_is_before) | |
801 | ref->prefetch_before = 0; | |
b8698a0f | 802 | |
b076a3fd ZD |
803 | return; |
804 | } | |
805 | ||
806 | if (!step) | |
807 | { | |
808 | /* If the reference addresses are invariant and fall into the | |
809 | same cache line, prefetch just the first one. */ | |
810 | if (!by_is_before) | |
811 | return; | |
812 | ||
813 | if (ddown (ref->delta, PREFETCH_BLOCK) | |
814 | != ddown (by->delta, PREFETCH_BLOCK)) | |
815 | return; | |
816 | ||
817 | ref->prefetch_before = 0; | |
818 | return; | |
819 | } | |
820 | ||
821 | /* Only prune the reference that is behind in the array. */ | |
822 | if (backward) | |
823 | { | |
824 | if (delta > 0) | |
825 | return; | |
826 | ||
827 | /* Transform the data so that we may assume that the accesses | |
828 | are forward. */ | |
829 | delta = - delta; | |
830 | step = -step; | |
831 | delta_r = PREFETCH_BLOCK - 1 - delta_r; | |
832 | delta_b = PREFETCH_BLOCK - 1 - delta_b; | |
833 | } | |
834 | else | |
835 | { | |
836 | if (delta < 0) | |
837 | return; | |
838 | } | |
839 | ||
840 | /* Check whether the two references are likely to hit the same cache | |
841 | line, and how distant the iterations in that it occurs are from | |
842 | each other. */ | |
843 | ||
844 | if (step <= PREFETCH_BLOCK) | |
845 | { | |
846 | /* The accesses are sure to meet. Let us check when. */ | |
847 | hit_from = ddown (delta_b, PREFETCH_BLOCK) * PREFETCH_BLOCK; | |
848 | prefetch_before = (hit_from - delta_r + step - 1) / step; | |
849 | ||
57762e97 | 850 | /* Do not reduce prefetch_before if we meet beyond cache size. */ |
4c9cf7af | 851 | if (prefetch_before > absu_hwi (L2_CACHE_SIZE_BYTES / step)) |
57762e97 | 852 | prefetch_before = PREFETCH_ALL; |
b076a3fd ZD |
853 | if (prefetch_before < ref->prefetch_before) |
854 | ref->prefetch_before = prefetch_before; | |
855 | ||
856 | return; | |
857 | } | |
858 | ||
b8698a0f | 859 | /* A more complicated case with step > prefetch_block. First reduce |
2c6dd136 | 860 | the ratio between the step and the cache line size to its simplest |
b8698a0f L |
861 | terms. The resulting denominator will then represent the number of |
862 | distinct iterations after which each address will go back to its | |
863 | initial location within the cache line. This computation assumes | |
2c6dd136 | 864 | that PREFETCH_BLOCK is a power of two. */ |
b076a3fd | 865 | prefetch_block = PREFETCH_BLOCK; |
2c6dd136 GS |
866 | reduced_prefetch_block = prefetch_block; |
867 | reduced_step = step; | |
868 | while ((reduced_step & 1) == 0 | |
869 | && reduced_prefetch_block > 1) | |
b076a3fd | 870 | { |
2c6dd136 GS |
871 | reduced_step >>= 1; |
872 | reduced_prefetch_block >>= 1; | |
b076a3fd ZD |
873 | } |
874 | ||
b076a3fd ZD |
875 | prefetch_before = delta / step; |
876 | delta %= step; | |
2c6dd136 GS |
877 | ref_type = TREE_TYPE (ref->mem); |
878 | align_unit = TYPE_ALIGN (ref_type) / 8; | |
14e444c3 CF |
879 | if (is_miss_rate_acceptable (prefetch_block, step, delta, |
880 | reduced_prefetch_block, align_unit)) | |
b076a3fd | 881 | { |
57762e97 CB |
882 | /* Do not reduce prefetch_before if we meet beyond cache size. */ |
883 | if (prefetch_before > L2_CACHE_SIZE_BYTES / PREFETCH_BLOCK) | |
884 | prefetch_before = PREFETCH_ALL; | |
b076a3fd ZD |
885 | if (prefetch_before < ref->prefetch_before) |
886 | ref->prefetch_before = prefetch_before; | |
887 | ||
888 | return; | |
889 | } | |
890 | ||
891 | /* Try also the following iteration. */ | |
892 | prefetch_before++; | |
893 | delta = step - delta; | |
14e444c3 CF |
894 | if (is_miss_rate_acceptable (prefetch_block, step, delta, |
895 | reduced_prefetch_block, align_unit)) | |
b076a3fd ZD |
896 | { |
897 | if (prefetch_before < ref->prefetch_before) | |
898 | ref->prefetch_before = prefetch_before; | |
899 | ||
900 | return; | |
901 | } | |
902 | ||
903 | /* The ref probably does not reuse by. */ | |
904 | return; | |
905 | } | |
906 | ||
907 | /* Prune the prefetch candidate REF using the reuses with other references | |
908 | in REFS. */ | |
909 | ||
910 | static void | |
911 | prune_ref_by_reuse (struct mem_ref *ref, struct mem_ref *refs) | |
912 | { | |
913 | struct mem_ref *prune_by; | |
914 | bool before = true; | |
915 | ||
916 | prune_ref_by_self_reuse (ref); | |
917 | ||
918 | for (prune_by = refs; prune_by; prune_by = prune_by->next) | |
919 | { | |
920 | if (prune_by == ref) | |
921 | { | |
922 | before = false; | |
923 | continue; | |
924 | } | |
925 | ||
926 | if (!WRITE_CAN_USE_READ_PREFETCH | |
927 | && ref->write_p | |
928 | && !prune_by->write_p) | |
929 | continue; | |
930 | if (!READ_CAN_USE_WRITE_PREFETCH | |
931 | && !ref->write_p | |
932 | && prune_by->write_p) | |
933 | continue; | |
934 | ||
935 | prune_ref_by_group_reuse (ref, prune_by, before); | |
936 | } | |
937 | } | |
938 | ||
939 | /* Prune the prefetch candidates in GROUP using the reuse analysis. */ | |
940 | ||
941 | static void | |
942 | prune_group_by_reuse (struct mem_ref_group *group) | |
943 | { | |
944 | struct mem_ref *ref_pruned; | |
945 | ||
946 | for (ref_pruned = group->refs; ref_pruned; ref_pruned = ref_pruned->next) | |
947 | { | |
948 | prune_ref_by_reuse (ref_pruned, group->refs); | |
949 | ||
950 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
951 | { | |
23b0f9f8 | 952 | dump_mem_ref (dump_file, ref_pruned); |
b076a3fd ZD |
953 | |
954 | if (ref_pruned->prefetch_before == PREFETCH_ALL | |
955 | && ref_pruned->prefetch_mod == 1) | |
956 | fprintf (dump_file, " no restrictions"); | |
957 | else if (ref_pruned->prefetch_before == 0) | |
958 | fprintf (dump_file, " do not prefetch"); | |
959 | else if (ref_pruned->prefetch_before <= ref_pruned->prefetch_mod) | |
960 | fprintf (dump_file, " prefetch once"); | |
961 | else | |
962 | { | |
963 | if (ref_pruned->prefetch_before != PREFETCH_ALL) | |
964 | { | |
965 | fprintf (dump_file, " prefetch before "); | |
966 | fprintf (dump_file, HOST_WIDE_INT_PRINT_DEC, | |
967 | ref_pruned->prefetch_before); | |
968 | } | |
969 | if (ref_pruned->prefetch_mod != 1) | |
970 | { | |
971 | fprintf (dump_file, " prefetch mod "); | |
972 | fprintf (dump_file, HOST_WIDE_INT_PRINT_DEC, | |
973 | ref_pruned->prefetch_mod); | |
974 | } | |
975 | } | |
976 | fprintf (dump_file, "\n"); | |
977 | } | |
978 | } | |
979 | } | |
980 | ||
981 | /* Prune the list of prefetch candidates GROUPS using the reuse analysis. */ | |
982 | ||
983 | static void | |
984 | prune_by_reuse (struct mem_ref_group *groups) | |
985 | { | |
986 | for (; groups; groups = groups->next) | |
987 | prune_group_by_reuse (groups); | |
988 | } | |
989 | ||
990 | /* Returns true if we should issue prefetch for REF. */ | |
991 | ||
992 | static bool | |
993 | should_issue_prefetch_p (struct mem_ref *ref) | |
994 | { | |
d2ff35c0 LM |
995 | /* Do we want to issue prefetches for non-constant strides? */ |
996 | if (!cst_and_fits_in_hwi (ref->group->step) && PREFETCH_DYNAMIC_STRIDES == 0) | |
997 | { | |
998 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
999 | fprintf (dump_file, | |
1000 | "Skipping non-constant step for reference %u:%u\n", | |
1001 | ref->group->uid, ref->uid); | |
1002 | return false; | |
1003 | } | |
1004 | ||
59100dfc LM |
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 (int_cst_value (ref->group->step)) | |
1011 | < (HOST_WIDE_INT) PREFETCH_MINIMUM_STRIDE) | |
1012 | { | |
1013 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
1014 | fprintf (dump_file, | |
79bf1170 LM |
1015 | "Step for reference %u:%u (" HOST_WIDE_INT_PRINT_DEC |
1016 | ") is less than the mininum required stride of %d\n", | |
59100dfc LM |
1017 | ref->group->uid, ref->uid, int_cst_value (ref->group->step), |
1018 | PREFETCH_MINIMUM_STRIDE); | |
1019 | return false; | |
1020 | } | |
1021 | ||
b076a3fd ZD |
1022 | /* For now do not issue prefetches for only first few of the |
1023 | iterations. */ | |
1024 | if (ref->prefetch_before != PREFETCH_ALL) | |
a8beb3a7 CB |
1025 | { |
1026 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
23b0f9f8 MK |
1027 | fprintf (dump_file, "Ignoring reference %u:%u due to prefetch_before\n", |
1028 | ref->group->uid, ref->uid); | |
a8beb3a7 CB |
1029 | return false; |
1030 | } | |
b076a3fd | 1031 | |
79f5e442 ZD |
1032 | /* Do not prefetch nontemporal stores. */ |
1033 | if (ref->storent_p) | |
a8beb3a7 CB |
1034 | { |
1035 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
23b0f9f8 | 1036 | fprintf (dump_file, "Ignoring nontemporal store reference %u:%u\n", ref->group->uid, ref->uid); |
a8beb3a7 CB |
1037 | return false; |
1038 | } | |
79f5e442 | 1039 | |
b076a3fd ZD |
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 | { | |
911b3fdb ZD |
1053 | unsigned remaining_prefetch_slots, n_prefetches, prefetch_slots; |
1054 | unsigned slots_per_prefetch; | |
b076a3fd ZD |
1055 | struct mem_ref *ref; |
1056 | bool any = false; | |
1057 | ||
911b3fdb ZD |
1058 | /* At most SIMULTANEOUS_PREFETCHES should be running at the same time. */ |
1059 | remaining_prefetch_slots = SIMULTANEOUS_PREFETCHES; | |
b076a3fd | 1060 | |
911b3fdb ZD |
1061 | /* The prefetch will run for AHEAD iterations of the original loop, i.e., |
1062 | AHEAD / UNROLL_FACTOR iterations of the unrolled loop. In each iteration, | |
1063 | it will need a prefetch slot. */ | |
1064 | slots_per_prefetch = (ahead + unroll_factor / 2) / unroll_factor; | |
b076a3fd | 1065 | if (dump_file && (dump_flags & TDF_DETAILS)) |
911b3fdb ZD |
1066 | fprintf (dump_file, "Each prefetch instruction takes %u prefetch slots.\n", |
1067 | slots_per_prefetch); | |
b076a3fd ZD |
1068 | |
1069 | /* For now we just take memory references one by one and issue | |
1070 | prefetches for as many as possible. The groups are sorted | |
1071 | starting with the largest step, since the references with | |
c0220ea4 | 1072 | large step are more likely to cause many cache misses. */ |
b076a3fd ZD |
1073 | |
1074 | for (; groups; groups = groups->next) | |
1075 | for (ref = groups->refs; ref; ref = ref->next) | |
1076 | { | |
1077 | if (!should_issue_prefetch_p (ref)) | |
1078 | continue; | |
1079 | ||
8532678c CF |
1080 | /* The loop is far from being sufficiently unrolled for this |
1081 | prefetch. Do not generate prefetch to avoid many redudant | |
1082 | prefetches. */ | |
1083 | if (ref->prefetch_mod / unroll_factor > PREFETCH_MOD_TO_UNROLL_FACTOR_RATIO) | |
1084 | continue; | |
1085 | ||
911b3fdb ZD |
1086 | /* If we need to prefetch the reference each PREFETCH_MOD iterations, |
1087 | and we unroll the loop UNROLL_FACTOR times, we need to insert | |
1088 | ceil (UNROLL_FACTOR / PREFETCH_MOD) instructions in each | |
1089 | iteration. */ | |
b076a3fd ZD |
1090 | n_prefetches = ((unroll_factor + ref->prefetch_mod - 1) |
1091 | / ref->prefetch_mod); | |
911b3fdb ZD |
1092 | prefetch_slots = n_prefetches * slots_per_prefetch; |
1093 | ||
1094 | /* If more than half of the prefetches would be lost anyway, do not | |
1095 | issue the prefetch. */ | |
1096 | if (2 * remaining_prefetch_slots < prefetch_slots) | |
1097 | continue; | |
1098 | ||
1a70c8d5 MK |
1099 | /* Stop prefetching if debug counter is activated. */ |
1100 | if (!dbg_cnt (prefetch)) | |
1101 | continue; | |
1102 | ||
911b3fdb | 1103 | ref->issue_prefetch_p = true; |
23b0f9f8 MK |
1104 | if (dump_file && (dump_flags & TDF_DETAILS)) |
1105 | fprintf (dump_file, "Decided to issue prefetch for reference %u:%u\n", | |
1106 | ref->group->uid, ref->uid); | |
b076a3fd | 1107 | |
911b3fdb ZD |
1108 | if (remaining_prefetch_slots <= prefetch_slots) |
1109 | return true; | |
1110 | remaining_prefetch_slots -= prefetch_slots; | |
b076a3fd ZD |
1111 | any = true; |
1112 | } | |
1113 | ||
1114 | return any; | |
1115 | } | |
1116 | ||
d5058523 CF |
1117 | /* Return TRUE if no prefetch is going to be generated in the given |
1118 | GROUPS. */ | |
1119 | ||
1120 | static bool | |
1121 | nothing_to_prefetch_p (struct mem_ref_group *groups) | |
1122 | { | |
1123 | struct mem_ref *ref; | |
1124 | ||
1125 | for (; groups; groups = groups->next) | |
1126 | for (ref = groups->refs; ref; ref = ref->next) | |
1127 | if (should_issue_prefetch_p (ref)) | |
1128 | return false; | |
1129 | ||
1130 | return true; | |
1131 | } | |
1132 | ||
1133 | /* Estimate the number of prefetches in the given GROUPS. | |
1134 | UNROLL_FACTOR is the factor by which LOOP was unrolled. */ | |
b076a3fd | 1135 | |
db34470d | 1136 | static int |
d5058523 | 1137 | estimate_prefetch_count (struct mem_ref_group *groups, unsigned unroll_factor) |
b076a3fd ZD |
1138 | { |
1139 | struct mem_ref *ref; | |
d5058523 | 1140 | unsigned n_prefetches; |
db34470d | 1141 | int prefetch_count = 0; |
b076a3fd ZD |
1142 | |
1143 | for (; groups; groups = groups->next) | |
1144 | for (ref = groups->refs; ref; ref = ref->next) | |
1145 | if (should_issue_prefetch_p (ref)) | |
d5058523 CF |
1146 | { |
1147 | n_prefetches = ((unroll_factor + ref->prefetch_mod - 1) | |
1148 | / ref->prefetch_mod); | |
1149 | prefetch_count += n_prefetches; | |
1150 | } | |
b076a3fd | 1151 | |
db34470d | 1152 | return prefetch_count; |
b076a3fd ZD |
1153 | } |
1154 | ||
1155 | /* Issue prefetches for the reference REF into loop as decided before. | |
1156 | HEAD is the number of iterations to prefetch ahead. UNROLL_FACTOR | |
917f1b7e | 1157 | is the factor by which LOOP was unrolled. */ |
b076a3fd ZD |
1158 | |
1159 | static void | |
1160 | issue_prefetch_ref (struct mem_ref *ref, unsigned unroll_factor, unsigned ahead) | |
1161 | { | |
1162 | HOST_WIDE_INT delta; | |
81f32326 | 1163 | tree addr, addr_base, write_p, local, forward; |
538dd0b7 | 1164 | gcall *prefetch; |
726a989a | 1165 | gimple_stmt_iterator bsi; |
b076a3fd | 1166 | unsigned n_prefetches, ap; |
5417e022 | 1167 | bool nontemporal = ref->reuse_distance >= L2_CACHE_SIZE_BYTES; |
b076a3fd ZD |
1168 | |
1169 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
23b0f9f8 | 1170 | fprintf (dump_file, "Issued%s prefetch for reference %u:%u.\n", |
5417e022 | 1171 | nontemporal ? " nontemporal" : "", |
23b0f9f8 | 1172 | ref->group->uid, ref->uid); |
b076a3fd | 1173 | |
726a989a | 1174 | bsi = gsi_for_stmt (ref->stmt); |
b076a3fd ZD |
1175 | |
1176 | n_prefetches = ((unroll_factor + ref->prefetch_mod - 1) | |
1177 | / ref->prefetch_mod); | |
1178 | addr_base = build_fold_addr_expr_with_type (ref->mem, ptr_type_node); | |
726a989a RB |
1179 | addr_base = force_gimple_operand_gsi (&bsi, unshare_expr (addr_base), |
1180 | true, NULL, true, GSI_SAME_STMT); | |
911b3fdb | 1181 | write_p = ref->write_p ? integer_one_node : integer_zero_node; |
9a9d280e | 1182 | local = nontemporal ? integer_zero_node : integer_three_node; |
b076a3fd ZD |
1183 | |
1184 | for (ap = 0; ap < n_prefetches; ap++) | |
1185 | { | |
81f32326 CB |
1186 | if (cst_and_fits_in_hwi (ref->group->step)) |
1187 | { | |
1188 | /* Determine the address to prefetch. */ | |
1189 | delta = (ahead + ap * ref->prefetch_mod) * | |
1190 | int_cst_value (ref->group->step); | |
5d49b6a7 | 1191 | addr = fold_build_pointer_plus_hwi (addr_base, delta); |
23b0f9f8 MK |
1192 | addr = force_gimple_operand_gsi (&bsi, unshare_expr (addr), true, |
1193 | NULL, true, GSI_SAME_STMT); | |
81f32326 CB |
1194 | } |
1195 | else | |
1196 | { | |
1197 | /* The step size is non-constant but loop-invariant. We use the | |
1198 | heuristic to simply prefetch ahead iterations ahead. */ | |
1199 | forward = fold_build2 (MULT_EXPR, sizetype, | |
1200 | fold_convert (sizetype, ref->group->step), | |
1201 | fold_convert (sizetype, size_int (ahead))); | |
5d49b6a7 | 1202 | addr = fold_build_pointer_plus (addr_base, forward); |
81f32326 CB |
1203 | addr = force_gimple_operand_gsi (&bsi, unshare_expr (addr), true, |
1204 | NULL, true, GSI_SAME_STMT); | |
1205 | } | |
ed481942 | 1206 | |
fe8b1e83 DG |
1207 | if (addr_base != addr |
1208 | && TREE_CODE (addr_base) == SSA_NAME | |
ed481942 DG |
1209 | && TREE_CODE (addr) == SSA_NAME) |
1210 | { | |
1211 | duplicate_ssa_name_ptr_info (addr, SSA_NAME_PTR_INFO (addr_base)); | |
1212 | /* As this isn't a plain copy we have to reset alignment | |
1213 | information. */ | |
1214 | if (SSA_NAME_PTR_INFO (addr)) | |
1215 | mark_ptr_info_alignment_unknown (SSA_NAME_PTR_INFO (addr)); | |
1216 | } | |
1217 | ||
b076a3fd | 1218 | /* Create the prefetch instruction. */ |
e79983f4 | 1219 | prefetch = gimple_build_call (builtin_decl_explicit (BUILT_IN_PREFETCH), |
726a989a RB |
1220 | 3, addr, write_p, local); |
1221 | gsi_insert_before (&bsi, prefetch, GSI_SAME_STMT); | |
b076a3fd ZD |
1222 | } |
1223 | } | |
1224 | ||
1225 | /* Issue prefetches for the references in GROUPS into loop as decided before. | |
1226 | HEAD is the number of iterations to prefetch ahead. UNROLL_FACTOR is the | |
1227 | factor by that LOOP was unrolled. */ | |
1228 | ||
1229 | static void | |
1230 | issue_prefetches (struct mem_ref_group *groups, | |
1231 | unsigned unroll_factor, unsigned ahead) | |
1232 | { | |
1233 | struct mem_ref *ref; | |
1234 | ||
1235 | for (; groups; groups = groups->next) | |
1236 | for (ref = groups->refs; ref; ref = ref->next) | |
1237 | if (ref->issue_prefetch_p) | |
1238 | issue_prefetch_ref (ref, unroll_factor, ahead); | |
1239 | } | |
1240 | ||
79f5e442 ZD |
1241 | /* Returns true if REF is a memory write for that a nontemporal store insn |
1242 | can be used. */ | |
1243 | ||
1244 | static bool | |
1245 | nontemporal_store_p (struct mem_ref *ref) | |
1246 | { | |
ef4bddc2 | 1247 | machine_mode mode; |
79f5e442 ZD |
1248 | enum insn_code code; |
1249 | ||
1250 | /* REF must be a write that is not reused. We require it to be independent | |
1251 | on all other memory references in the loop, as the nontemporal stores may | |
1252 | be reordered with respect to other memory references. */ | |
1253 | if (!ref->write_p | |
1254 | || !ref->independent_p | |
1255 | || ref->reuse_distance < L2_CACHE_SIZE_BYTES) | |
1256 | return false; | |
1257 | ||
1258 | /* Check that we have the storent instruction for the mode. */ | |
1259 | mode = TYPE_MODE (TREE_TYPE (ref->mem)); | |
1260 | if (mode == BLKmode) | |
1261 | return false; | |
1262 | ||
947131ba | 1263 | code = optab_handler (storent_optab, mode); |
79f5e442 ZD |
1264 | return code != CODE_FOR_nothing; |
1265 | } | |
1266 | ||
1267 | /* If REF is a nontemporal store, we mark the corresponding modify statement | |
1268 | and return true. Otherwise, we return false. */ | |
1269 | ||
1270 | static bool | |
1271 | mark_nontemporal_store (struct mem_ref *ref) | |
1272 | { | |
1273 | if (!nontemporal_store_p (ref)) | |
1274 | return false; | |
1275 | ||
1276 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
23b0f9f8 MK |
1277 | fprintf (dump_file, "Marked reference %u:%u as a nontemporal store.\n", |
1278 | ref->group->uid, ref->uid); | |
79f5e442 | 1279 | |
726a989a | 1280 | gimple_assign_set_nontemporal_move (ref->stmt, true); |
79f5e442 ZD |
1281 | ref->storent_p = true; |
1282 | ||
1283 | return true; | |
1284 | } | |
1285 | ||
1286 | /* Issue a memory fence instruction after LOOP. */ | |
1287 | ||
1288 | static void | |
1289 | emit_mfence_after_loop (struct loop *loop) | |
1290 | { | |
9771b263 | 1291 | vec<edge> exits = get_loop_exit_edges (loop); |
79f5e442 | 1292 | edge exit; |
538dd0b7 | 1293 | gcall *call; |
726a989a | 1294 | gimple_stmt_iterator bsi; |
79f5e442 ZD |
1295 | unsigned i; |
1296 | ||
9771b263 | 1297 | FOR_EACH_VEC_ELT (exits, i, exit) |
79f5e442 | 1298 | { |
726a989a | 1299 | call = gimple_build_call (FENCE_FOLLOWING_MOVNT, 0); |
79f5e442 ZD |
1300 | |
1301 | if (!single_pred_p (exit->dest) | |
1302 | /* If possible, we prefer not to insert the fence on other paths | |
1303 | in cfg. */ | |
1304 | && !(exit->flags & EDGE_ABNORMAL)) | |
1305 | split_loop_exit_edge (exit); | |
726a989a | 1306 | bsi = gsi_after_labels (exit->dest); |
79f5e442 | 1307 | |
726a989a | 1308 | gsi_insert_before (&bsi, call, GSI_NEW_STMT); |
79f5e442 ZD |
1309 | } |
1310 | ||
9771b263 | 1311 | exits.release (); |
79f5e442 ZD |
1312 | update_ssa (TODO_update_ssa_only_virtuals); |
1313 | } | |
1314 | ||
1315 | /* Returns true if we can use storent in loop, false otherwise. */ | |
1316 | ||
1317 | static bool | |
1318 | may_use_storent_in_loop_p (struct loop *loop) | |
1319 | { | |
1320 | bool ret = true; | |
1321 | ||
1322 | if (loop->inner != NULL) | |
1323 | return false; | |
1324 | ||
1325 | /* If we must issue a mfence insn after using storent, check that there | |
1326 | is a suitable place for it at each of the loop exits. */ | |
1327 | if (FENCE_FOLLOWING_MOVNT != NULL_TREE) | |
1328 | { | |
9771b263 | 1329 | vec<edge> exits = get_loop_exit_edges (loop); |
79f5e442 ZD |
1330 | unsigned i; |
1331 | edge exit; | |
1332 | ||
9771b263 | 1333 | FOR_EACH_VEC_ELT (exits, i, exit) |
79f5e442 | 1334 | if ((exit->flags & EDGE_ABNORMAL) |
fefa31b5 | 1335 | && exit->dest == EXIT_BLOCK_PTR_FOR_FN (cfun)) |
79f5e442 ZD |
1336 | ret = false; |
1337 | ||
9771b263 | 1338 | exits.release (); |
79f5e442 ZD |
1339 | } |
1340 | ||
1341 | return ret; | |
1342 | } | |
1343 | ||
1344 | /* Marks nontemporal stores in LOOP. GROUPS contains the description of memory | |
1345 | references in the loop. */ | |
1346 | ||
1347 | static void | |
1348 | mark_nontemporal_stores (struct loop *loop, struct mem_ref_group *groups) | |
1349 | { | |
1350 | struct mem_ref *ref; | |
1351 | bool any = false; | |
1352 | ||
1353 | if (!may_use_storent_in_loop_p (loop)) | |
1354 | return; | |
1355 | ||
1356 | for (; groups; groups = groups->next) | |
1357 | for (ref = groups->refs; ref; ref = ref->next) | |
1358 | any |= mark_nontemporal_store (ref); | |
1359 | ||
1360 | if (any && FENCE_FOLLOWING_MOVNT != NULL_TREE) | |
1361 | emit_mfence_after_loop (loop); | |
1362 | } | |
1363 | ||
b076a3fd ZD |
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 (struct loop *loop, struct tree_niter_desc *desc, | |
1370 | unsigned factor) | |
1371 | { | |
1372 | if (!can_unroll_loop_p (loop, factor, 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 | |
23b0f9f8 | 1388 | number of iterations of LOOP is stored to DESC. NINSNS is the number of |
2711355f ZD |
1389 | insns of the LOOP. EST_NITER is the estimated number of iterations of |
1390 | the loop, or -1 if no estimate is available. */ | |
b076a3fd ZD |
1391 | |
1392 | static unsigned | |
1393 | determine_unroll_factor (struct loop *loop, struct mem_ref_group *refs, | |
2711355f ZD |
1394 | unsigned ninsns, struct tree_niter_desc *desc, |
1395 | HOST_WIDE_INT est_niter) | |
b076a3fd | 1396 | { |
911b3fdb ZD |
1397 | unsigned upper_bound; |
1398 | unsigned nfactor, factor, mod_constraint; | |
b076a3fd ZD |
1399 | struct mem_ref_group *agp; |
1400 | struct mem_ref *ref; | |
1401 | ||
911b3fdb ZD |
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_VALUE (PARAM_MAX_UNROLLED_INSNS) / ninsns; | |
2711355f ZD |
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 | ||
911b3fdb | 1416 | if (upper_bound <= 1) |
b076a3fd ZD |
1417 | return 1; |
1418 | ||
911b3fdb ZD |
1419 | /* Choose the factor so that we may prefetch each cache just once, |
1420 | but bound the unrolling by UPPER_BOUND. */ | |
1421 | factor = 1; | |
b076a3fd ZD |
1422 | for (agp = refs; agp; agp = agp->next) |
1423 | for (ref = agp->refs; ref; ref = ref->next) | |
911b3fdb ZD |
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 | } | |
b076a3fd ZD |
1431 | |
1432 | if (!should_unroll_loop_p (loop, desc, factor)) | |
1433 | return 1; | |
1434 | ||
1435 | return factor; | |
1436 | } | |
1437 | ||
5417e022 ZD |
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 += 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, struct loop *loop) | |
1497 | { | |
1498 | struct 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 (access_fn); | |
1506 | step = CHREC_RIGHT (access_fn); | |
1507 | access_fn = CHREC_LEFT (access_fn); | |
1508 | ||
1509 | if ((unsigned) loop_depth (aloop) <= min_depth) | |
1510 | continue; | |
1511 | ||
9541ffee | 1512 | if (tree_fits_shwi_p (step)) |
9439e9a1 | 1513 | astep = tree_to_shwi (step); |
5417e022 ZD |
1514 | else |
1515 | astep = L1_CACHE_LINE_SIZE; | |
1516 | ||
1517 | strides[n - 1 - loop_depth (loop) + loop_depth (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 | struct loop *loop) | |
1530 | { | |
1531 | tree stride, access_fn; | |
1532 | HOST_WIDE_INT *strides, astride; | |
9771b263 | 1533 | vec<tree> access_fns; |
5417e022 ZD |
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 | ||
9771b263 | 1552 | FOR_EACH_VEC_ELT (access_fns, i, access_fn) |
5417e022 ZD |
1553 | { |
1554 | /* Keep track of the reference corresponding to the subscript, so that we | |
1555 | know its stride. */ | |
1556 | while (handled_component_p (ref) && TREE_CODE (ref) != ARRAY_REF) | |
1557 | ref = TREE_OPERAND (ref, 0); | |
b8698a0f | 1558 | |
5417e022 ZD |
1559 | if (TREE_CODE (ref) == ARRAY_REF) |
1560 | { | |
1561 | stride = TYPE_SIZE_UNIT (TREE_TYPE (ref)); | |
cc269bb6 | 1562 | if (tree_fits_uhwi_p (stride)) |
ae7e9ddd | 1563 | astride = tree_to_uhwi (stride); |
5417e022 ZD |
1564 | else |
1565 | astride = L1_CACHE_LINE_SIZE; | |
1566 | ||
1567 | ref = TREE_OPERAND (ref, 0); | |
1568 | } | |
1569 | else | |
1570 | astride = 1; | |
1571 | ||
1572 | add_subscript_strides (access_fn, 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) 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 (strides); | |
1591 | return ret; | |
1592 | } | |
1593 | ||
1594 | /* Determines the distance till the first reuse of each reference in REFS | |
79f5e442 | 1595 | in the loop nest of LOOP. NO_OTHER_REFS is true if there are no other |
1e373390 | 1596 | memory references in the loop. Return false if the analysis fails. */ |
5417e022 | 1597 | |
1e373390 | 1598 | static bool |
79f5e442 ZD |
1599 | determine_loop_nest_reuse (struct loop *loop, struct mem_ref_group *refs, |
1600 | bool no_other_refs) | |
5417e022 ZD |
1601 | { |
1602 | struct loop *nest, *aloop; | |
6e1aa848 DN |
1603 | vec<data_reference_p> datarefs = vNULL; |
1604 | vec<ddr_p> dependences = vNULL; | |
5417e022 | 1605 | struct mem_ref_group *gr; |
79f5e442 | 1606 | struct mem_ref *ref, *refb; |
843374f2 | 1607 | auto_vec<loop_p> vloops; |
5417e022 ZD |
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) | |
1e373390 | 1616 | return true; |
5417e022 ZD |
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 (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); | |
9771b263 | 1636 | n = vloops.length (); |
5417e022 ZD |
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 | ||
9771b263 | 1648 | aloop = vloops[i]; |
652c4c71 | 1649 | vol = estimated_stmt_executions_int (aloop); |
e5b332cd | 1650 | if (vol == -1) |
5417e022 ZD |
1651 | vol = expected_loop_iterations (aloop); |
1652 | volume *= vol; | |
1653 | } | |
1654 | ||
1655 | /* Prepare the references in the form suitable for data dependence | |
0d52bcc1 | 1656 | analysis. We ignore unanalyzable data references (the results |
5417e022 ZD |
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 | { | |
a68f286c RB |
1662 | dr = create_data_ref (loop_preheader_edge (nest), |
1663 | loop_containing_stmt (ref->stmt), | |
62c8a2cf | 1664 | ref->mem, ref->stmt, !ref->write_p, false); |
5417e022 ZD |
1665 | |
1666 | if (dr) | |
1667 | { | |
1668 | ref->reuse_distance = volume; | |
1669 | dr->aux = ref; | |
9771b263 | 1670 | datarefs.safe_push (dr); |
5417e022 | 1671 | } |
79f5e442 ZD |
1672 | else |
1673 | no_other_refs = false; | |
5417e022 ZD |
1674 | } |
1675 | ||
9771b263 | 1676 | FOR_EACH_VEC_ELT (datarefs, i, dr) |
5417e022 ZD |
1677 | { |
1678 | dist = self_reuse_distance (dr, loop_data_size, n, loop); | |
3d9a9f94 | 1679 | ref = (struct mem_ref *) dr->aux; |
5417e022 ZD |
1680 | if (ref->reuse_distance > dist) |
1681 | ref->reuse_distance = dist; | |
79f5e442 ZD |
1682 | |
1683 | if (no_other_refs) | |
1684 | ref->independent_p = true; | |
5417e022 ZD |
1685 | } |
1686 | ||
1e373390 RG |
1687 | if (!compute_all_dependences (datarefs, &dependences, vloops, true)) |
1688 | return false; | |
5417e022 | 1689 | |
9771b263 | 1690 | FOR_EACH_VEC_ELT (dependences, i, dep) |
5417e022 ZD |
1691 | { |
1692 | if (DDR_ARE_DEPENDENT (dep) == chrec_known) | |
1693 | continue; | |
1694 | ||
3d9a9f94 KG |
1695 | ref = (struct mem_ref *) DDR_A (dep)->aux; |
1696 | refb = (struct mem_ref *) DDR_B (dep)->aux; | |
79f5e442 | 1697 | |
5417e022 | 1698 | if (DDR_ARE_DEPENDENT (dep) == chrec_dont_know |
dfbddbeb | 1699 | || DDR_COULD_BE_INDEPENDENT_P (dep) |
5417e022 ZD |
1700 | || DDR_NUM_DIST_VECTS (dep) == 0) |
1701 | { | |
0d52bcc1 | 1702 | /* If the dependence cannot be analyzed, assume that there might be |
5417e022 ZD |
1703 | a reuse. */ |
1704 | dist = 0; | |
b8698a0f | 1705 | |
79f5e442 ZD |
1706 | ref->independent_p = false; |
1707 | refb->independent_p = false; | |
5417e022 ZD |
1708 | } |
1709 | else | |
1710 | { | |
0d52bcc1 | 1711 | /* The distance vectors are normalized to be always lexicographically |
5417e022 ZD |
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_data_size, n); | |
1724 | ||
79f5e442 ZD |
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), 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 | ||
5417e022 ZD |
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 | ||
5417e022 ZD |
1750 | if (ref->reuse_distance > dist) |
1751 | ref->reuse_distance = dist; | |
79f5e442 ZD |
1752 | if (refb->reuse_distance > dist) |
1753 | refb->reuse_distance = dist; | |
5417e022 ZD |
1754 | } |
1755 | ||
1756 | free_dependence_relations (dependences); | |
1757 | free_data_refs (datarefs); | |
1758 | free (loop_data_size); | |
1759 | ||
1760 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
1761 | { | |
1762 | fprintf (dump_file, "Reuse distances:\n"); | |
1763 | for (gr = refs; gr; gr = gr->next) | |
1764 | for (ref = gr->refs; ref; ref = ref->next) | |
23b0f9f8 MK |
1765 | fprintf (dump_file, " reference %u:%u distance %u\n", |
1766 | ref->group->uid, ref->uid, ref->reuse_distance); | |
5417e022 | 1767 | } |
1e373390 RG |
1768 | |
1769 | return true; | |
5417e022 ZD |
1770 | } |
1771 | ||
0bbe50f6 CF |
1772 | /* Determine whether or not the trip count to ahead ratio is too small based |
1773 | on prefitablility consideration. | |
db34470d | 1774 | AHEAD: the iteration ahead distance, |
0bbe50f6 CF |
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 (dump_file, | |
1789 | "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. | |
db34470d | 1799 | NINSNS: estimated number of instructions in the loop, |
db34470d GS |
1800 | MEM_REF_COUNT: total number of memory references in the loop. */ |
1801 | ||
b8698a0f | 1802 | static bool |
0bbe50f6 | 1803 | mem_ref_count_reasonable_p (unsigned ninsns, unsigned mem_ref_count) |
db34470d | 1804 | { |
0bbe50f6 | 1805 | int insn_to_mem_ratio; |
db34470d GS |
1806 | |
1807 | if (mem_ref_count == 0) | |
1808 | return false; | |
1809 | ||
0bbe50f6 CF |
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 | ||
b8698a0f L |
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 | |
db34470d GS |
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. */ | |
b8698a0f | 1826 | insn_to_mem_ratio = ninsns / mem_ref_count; |
db34470d GS |
1827 | |
1828 | if (insn_to_mem_ratio < PREFETCH_MIN_INSN_TO_MEM_RATIO) | |
55e5a2eb CF |
1829 | { |
1830 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
1831 | fprintf (dump_file, | |
1832 | "Not prefetching -- instruction to memory reference ratio (%d) too small\n", | |
1833 | insn_to_mem_ratio); | |
1834 | return false; | |
1835 | } | |
db34470d | 1836 | |
0bbe50f6 CF |
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 | ||
d3a9b459 CF |
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. | |
ccacf0e1 CF |
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. */ | |
d3a9b459 CF |
1864 | insn_to_prefetch_ratio = (unroll_factor * ninsns) / prefetch_count; |
1865 | if (insn_to_prefetch_ratio < MIN_INSN_TO_PREFETCH_RATIO) | |
db34470d | 1866 | { |
d3a9b459 CF |
1867 | if (dump_file && (dump_flags & TDF_DETAILS)) |
1868 | fprintf (dump_file, | |
1869 | "Not prefetching -- instruction to prefetch ratio (%d) too small\n", | |
1870 | insn_to_prefetch_ratio); | |
0bbe50f6 | 1871 | return true; |
db34470d | 1872 | } |
b8698a0f | 1873 | |
0bbe50f6 | 1874 | return false; |
db34470d GS |
1875 | } |
1876 | ||
1877 | ||
b076a3fd | 1878 | /* Issue prefetch instructions for array references in LOOP. Returns |
d73be268 | 1879 | true if the LOOP was unrolled. */ |
b076a3fd ZD |
1880 | |
1881 | static bool | |
d73be268 | 1882 | loop_prefetch_arrays (struct loop *loop) |
b076a3fd ZD |
1883 | { |
1884 | struct mem_ref_group *refs; | |
2711355f ZD |
1885 | unsigned ahead, ninsns, time, unroll_factor; |
1886 | HOST_WIDE_INT est_niter; | |
b076a3fd | 1887 | struct tree_niter_desc desc; |
79f5e442 | 1888 | bool unrolled = false, no_other_refs; |
db34470d GS |
1889 | unsigned prefetch_count; |
1890 | unsigned mem_ref_count; | |
b076a3fd | 1891 | |
efd8f750 | 1892 | if (optimize_loop_nest_for_size_p (loop)) |
2732d767 ZD |
1893 | { |
1894 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
1895 | fprintf (dump_file, " ignored (cold area)\n"); | |
1896 | return false; | |
1897 | } | |
1898 | ||
0bbe50f6 CF |
1899 | /* FIXME: the time should be weighted by the probabilities of the blocks in |
1900 | the loop body. */ | |
1901 | time = tree_num_loop_insns (loop, &eni_time_weights); | |
1902 | if (time == 0) | |
1903 | return false; | |
1904 | ||
1905 | ahead = (PREFETCH_LATENCY + time - 1) / time; | |
652c4c71 | 1906 | est_niter = estimated_stmt_executions_int (loop); |
e5b332cd | 1907 | if (est_niter == -1) |
00022058 | 1908 | est_niter = likely_max_stmt_executions_int (loop); |
0bbe50f6 CF |
1909 | |
1910 | /* Prefetching is not likely to be profitable if the trip count to ahead | |
1911 | ratio is too small. */ | |
1912 | if (trip_count_to_ahead_ratio_too_small_p (ahead, est_niter)) | |
1913 | return false; | |
1914 | ||
1915 | ninsns = tree_num_loop_insns (loop, &eni_size_weights); | |
1916 | ||
b076a3fd | 1917 | /* Step 1: gather the memory references. */ |
db34470d | 1918 | refs = gather_memory_references (loop, &no_other_refs, &mem_ref_count); |
b076a3fd | 1919 | |
0bbe50f6 CF |
1920 | /* Give up prefetching if the number of memory references in the |
1921 | loop is not reasonable based on profitablity and compilation time | |
1922 | considerations. */ | |
1923 | if (!mem_ref_count_reasonable_p (ninsns, mem_ref_count)) | |
1924 | goto fail; | |
1925 | ||
b076a3fd ZD |
1926 | /* Step 2: estimate the reuse effects. */ |
1927 | prune_by_reuse (refs); | |
1928 | ||
d5058523 | 1929 | if (nothing_to_prefetch_p (refs)) |
b076a3fd ZD |
1930 | goto fail; |
1931 | ||
1e373390 RG |
1932 | if (!determine_loop_nest_reuse (loop, refs, no_other_refs)) |
1933 | goto fail; | |
5417e022 | 1934 | |
0bbe50f6 | 1935 | /* Step 3: determine unroll factor. */ |
2711355f ZD |
1936 | unroll_factor = determine_unroll_factor (loop, refs, ninsns, &desc, |
1937 | est_niter); | |
d5058523 CF |
1938 | |
1939 | /* Estimate prefetch count for the unrolled loop. */ | |
1940 | prefetch_count = estimate_prefetch_count (refs, unroll_factor); | |
1941 | if (prefetch_count == 0) | |
1942 | goto fail; | |
1943 | ||
2711355f | 1944 | if (dump_file && (dump_flags & TDF_DETAILS)) |
b8698a0f | 1945 | fprintf (dump_file, "Ahead %d, unroll factor %d, trip count " |
d81f5387 | 1946 | HOST_WIDE_INT_PRINT_DEC "\n" |
b8698a0f L |
1947 | "insn count %d, mem ref count %d, prefetch count %d\n", |
1948 | ahead, unroll_factor, est_niter, | |
1949 | ninsns, mem_ref_count, prefetch_count); | |
db34470d | 1950 | |
0bbe50f6 CF |
1951 | /* Prefetching is not likely to be profitable if the instruction to prefetch |
1952 | ratio is too small. */ | |
1953 | if (insn_to_prefetch_ratio_too_small_p (ninsns, prefetch_count, | |
1954 | unroll_factor)) | |
db34470d GS |
1955 | goto fail; |
1956 | ||
1957 | mark_nontemporal_stores (loop, refs); | |
2711355f | 1958 | |
b076a3fd ZD |
1959 | /* Step 4: what to prefetch? */ |
1960 | if (!schedule_prefetches (refs, unroll_factor, ahead)) | |
1961 | goto fail; | |
1962 | ||
1963 | /* Step 5: unroll the loop. TODO -- peeling of first and last few | |
1964 | iterations so that we do not issue superfluous prefetches. */ | |
1965 | if (unroll_factor != 1) | |
1966 | { | |
d73be268 | 1967 | tree_unroll_loop (loop, unroll_factor, |
b076a3fd ZD |
1968 | single_dom_exit (loop), &desc); |
1969 | unrolled = true; | |
1970 | } | |
1971 | ||
1972 | /* Step 6: issue the prefetches. */ | |
1973 | issue_prefetches (refs, unroll_factor, ahead); | |
1974 | ||
1975 | fail: | |
1976 | release_mem_refs (refs); | |
1977 | return unrolled; | |
1978 | } | |
1979 | ||
d73be268 | 1980 | /* Issue prefetch instructions for array references in loops. */ |
b076a3fd | 1981 | |
c7f965b6 | 1982 | unsigned int |
d73be268 | 1983 | tree_ssa_prefetch_arrays (void) |
b076a3fd | 1984 | { |
b076a3fd ZD |
1985 | struct loop *loop; |
1986 | bool unrolled = false; | |
c7f965b6 | 1987 | int todo_flags = 0; |
b076a3fd | 1988 | |
134b044d | 1989 | if (!targetm.have_prefetch () |
b076a3fd ZD |
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 | |
134b044d | 1993 | -march=pentium4 causes targetm.have_prefetch to be true. Ugh. */ |
b076a3fd | 1994 | || PREFETCH_BLOCK == 0) |
c7f965b6 | 1995 | return 0; |
b076a3fd | 1996 | |
47eb5b32 ZD |
1997 | if (dump_file && (dump_flags & TDF_DETAILS)) |
1998 | { | |
1999 | fprintf (dump_file, "Prefetching parameters:\n"); | |
2000 | fprintf (dump_file, " simultaneous prefetches: %d\n", | |
2001 | SIMULTANEOUS_PREFETCHES); | |
2002 | fprintf (dump_file, " prefetch latency: %d\n", PREFETCH_LATENCY); | |
47eb5b32 | 2003 | fprintf (dump_file, " prefetch block size: %d\n", PREFETCH_BLOCK); |
46cb0441 ZD |
2004 | fprintf (dump_file, " L1 cache size: %d lines, %d kB\n", |
2005 | L1_CACHE_SIZE_BYTES / L1_CACHE_LINE_SIZE, L1_CACHE_SIZE); | |
5417e022 | 2006 | fprintf (dump_file, " L1 cache line size: %d\n", L1_CACHE_LINE_SIZE); |
b8698a0f L |
2007 | fprintf (dump_file, " L2 cache size: %d kB\n", L2_CACHE_SIZE); |
2008 | fprintf (dump_file, " min insn-to-prefetch ratio: %d \n", | |
db34470d | 2009 | MIN_INSN_TO_PREFETCH_RATIO); |
b8698a0f | 2010 | fprintf (dump_file, " min insn-to-mem ratio: %d \n", |
db34470d | 2011 | PREFETCH_MIN_INSN_TO_MEM_RATIO); |
47eb5b32 ZD |
2012 | fprintf (dump_file, "\n"); |
2013 | } | |
2014 | ||
b076a3fd ZD |
2015 | initialize_original_copy_tables (); |
2016 | ||
e79983f4 | 2017 | if (!builtin_decl_explicit_p (BUILT_IN_PREFETCH)) |
b076a3fd | 2018 | { |
6a4825bd NF |
2019 | tree type = build_function_type_list (void_type_node, |
2020 | const_ptr_type_node, NULL_TREE); | |
c79efc4d RÁE |
2021 | tree decl = add_builtin_function ("__builtin_prefetch", type, |
2022 | BUILT_IN_PREFETCH, BUILT_IN_NORMAL, | |
2023 | NULL, NULL_TREE); | |
b076a3fd | 2024 | DECL_IS_NOVOPS (decl) = true; |
e79983f4 | 2025 | set_builtin_decl (BUILT_IN_PREFETCH, decl, false); |
b076a3fd ZD |
2026 | } |
2027 | ||
f0bd40b1 | 2028 | FOR_EACH_LOOP (loop, LI_FROM_INNERMOST) |
b076a3fd | 2029 | { |
b076a3fd ZD |
2030 | if (dump_file && (dump_flags & TDF_DETAILS)) |
2031 | fprintf (dump_file, "Processing loop %d:\n", loop->num); | |
2032 | ||
d73be268 | 2033 | unrolled |= loop_prefetch_arrays (loop); |
b076a3fd ZD |
2034 | |
2035 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
2036 | fprintf (dump_file, "\n\n"); | |
2037 | } | |
2038 | ||
2039 | if (unrolled) | |
2040 | { | |
2041 | scev_reset (); | |
c7f965b6 | 2042 | todo_flags |= TODO_cleanup_cfg; |
b076a3fd ZD |
2043 | } |
2044 | ||
2045 | free_original_copy_tables (); | |
c7f965b6 | 2046 | return todo_flags; |
b076a3fd | 2047 | } |
71343877 AM |
2048 | |
2049 | /* Prefetching. */ | |
2050 | ||
71343877 AM |
2051 | namespace { |
2052 | ||
2053 | const pass_data pass_data_loop_prefetch = | |
2054 | { | |
2055 | GIMPLE_PASS, /* type */ | |
2056 | "aprefetch", /* name */ | |
2057 | OPTGROUP_LOOP, /* optinfo_flags */ | |
71343877 AM |
2058 | TV_TREE_PREFETCH, /* tv_id */ |
2059 | ( PROP_cfg | PROP_ssa ), /* properties_required */ | |
2060 | 0, /* properties_provided */ | |
2061 | 0, /* properties_destroyed */ | |
2062 | 0, /* todo_flags_start */ | |
2063 | 0, /* todo_flags_finish */ | |
2064 | }; | |
2065 | ||
2066 | class pass_loop_prefetch : public gimple_opt_pass | |
2067 | { | |
2068 | public: | |
2069 | pass_loop_prefetch (gcc::context *ctxt) | |
2070 | : gimple_opt_pass (pass_data_loop_prefetch, ctxt) | |
2071 | {} | |
2072 | ||
2073 | /* opt_pass methods: */ | |
1a3d085c | 2074 | virtual bool gate (function *) { return flag_prefetch_loop_arrays > 0; } |
be55bfe6 | 2075 | virtual unsigned int execute (function *); |
71343877 AM |
2076 | |
2077 | }; // class pass_loop_prefetch | |
2078 | ||
be55bfe6 TS |
2079 | unsigned int |
2080 | pass_loop_prefetch::execute (function *fun) | |
2081 | { | |
2082 | if (number_of_loops (fun) <= 1) | |
2083 | return 0; | |
2084 | ||
d78a1c01 ML |
2085 | if ((PREFETCH_BLOCK & (PREFETCH_BLOCK - 1)) != 0) |
2086 | { | |
2087 | static bool warned = false; | |
2088 | ||
2089 | if (!warned) | |
2090 | { | |
2091 | warning (OPT_Wdisabled_optimization, | |
2092 | "%<l1-cache-size%> parameter is not a power of two %d", | |
2093 | PREFETCH_BLOCK); | |
2094 | warned = true; | |
2095 | } | |
2096 | return 0; | |
2097 | } | |
2098 | ||
be55bfe6 TS |
2099 | return tree_ssa_prefetch_arrays (); |
2100 | } | |
2101 | ||
71343877 AM |
2102 | } // anon namespace |
2103 | ||
2104 | gimple_opt_pass * | |
2105 | make_pass_loop_prefetch (gcc::context *ctxt) | |
2106 | { | |
2107 | return new pass_loop_prefetch (ctxt); | |
2108 | } | |
2109 | ||
2110 |