]>
Commit | Line | Data |
---|---|---|
b076a3fd | 1 | /* Array prefetching. |
8d9254fc | 2 | Copyright (C) 2005-2020 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 | 46 | #include "langhooks.h" |
7f9bc51b | 47 | #include "tree-inline.h" |
5417e022 | 48 | #include "tree-data-ref.h" |
d78a1c01 | 49 | #include "diagnostic-core.h" |
1a70c8d5 | 50 | #include "dbgcnt.h" |
2eb79bbb | 51 | |
b076a3fd ZD |
52 | /* This pass inserts prefetch instructions to optimize cache usage during |
53 | accesses to arrays in loops. It processes loops sequentially and: | |
54 | ||
55 | 1) Gathers all memory references in the single loop. | |
56 | 2) For each of the references it decides when it is profitable to prefetch | |
57 | it. To do it, we evaluate the reuse among the accesses, and determines | |
58 | two values: PREFETCH_BEFORE (meaning that it only makes sense to do | |
59 | prefetching in the first PREFETCH_BEFORE iterations of the loop) and | |
60 | PREFETCH_MOD (meaning that it only makes sense to prefetch in the | |
61 | iterations of the loop that are zero modulo PREFETCH_MOD). For example | |
62 | (assuming cache line size is 64 bytes, char has size 1 byte and there | |
63 | is no hardware sequential prefetch): | |
64 | ||
65 | char *a; | |
66 | for (i = 0; i < max; i++) | |
67 | { | |
68 | a[255] = ...; (0) | |
69 | a[i] = ...; (1) | |
70 | a[i + 64] = ...; (2) | |
71 | a[16*i] = ...; (3) | |
72 | a[187*i] = ...; (4) | |
73 | a[187*i + 50] = ...; (5) | |
74 | } | |
75 | ||
76 | (0) obviously has PREFETCH_BEFORE 1 | |
77 | (1) has PREFETCH_BEFORE 64, since (2) accesses the same memory | |
78 | location 64 iterations before it, and PREFETCH_MOD 64 (since | |
79 | it hits the same cache line otherwise). | |
80 | (2) has PREFETCH_MOD 64 | |
81 | (3) has PREFETCH_MOD 4 | |
82 | (4) has PREFETCH_MOD 1. We do not set PREFETCH_BEFORE here, since | |
bae077dc | 83 | the cache line accessed by (5) is the same with probability only |
b076a3fd ZD |
84 | 7/32. |
85 | (5) has PREFETCH_MOD 1 as well. | |
86 | ||
5417e022 ZD |
87 | Additionally, we use data dependence analysis to determine for each |
88 | reference the distance till the first reuse; this information is used | |
89 | to determine the temporality of the issued prefetch instruction. | |
90 | ||
b076a3fd ZD |
91 | 3) We determine how much ahead we need to prefetch. The number of |
92 | iterations needed is time to fetch / time spent in one iteration of | |
93 | the loop. The problem is that we do not know either of these values, | |
94 | so we just make a heuristic guess based on a magic (possibly) | |
95 | target-specific constant and size of the loop. | |
96 | ||
97 | 4) Determine which of the references we prefetch. We take into account | |
98 | that there is a maximum number of simultaneous prefetches (provided | |
99 | by machine description). We prefetch as many prefetches as possible | |
100 | while still within this bound (starting with those with lowest | |
101 | prefetch_mod, since they are responsible for most of the cache | |
102 | misses). | |
b8698a0f | 103 | |
b076a3fd ZD |
104 | 5) We unroll and peel loops so that we are able to satisfy PREFETCH_MOD |
105 | and PREFETCH_BEFORE requirements (within some bounds), and to avoid | |
106 | prefetching nonaccessed memory. | |
107 | TODO -- actually implement peeling. | |
b8698a0f | 108 | |
b076a3fd ZD |
109 | 6) We actually emit the prefetch instructions. ??? Perhaps emit the |
110 | prefetch instructions with guards in cases where 5) was not sufficient | |
111 | to satisfy the constraints? | |
112 | ||
0bbe50f6 CF |
113 | A cost model is implemented to determine whether or not prefetching is |
114 | profitable for a given loop. The cost model has three heuristics: | |
115 | ||
116 | 1. Function trip_count_to_ahead_ratio_too_small_p implements a | |
117 | heuristic that determines whether or not the loop has too few | |
118 | iterations (compared to ahead). Prefetching is not likely to be | |
119 | beneficial if the trip count to ahead ratio is below a certain | |
120 | minimum. | |
121 | ||
122 | 2. Function mem_ref_count_reasonable_p implements a heuristic that | |
123 | determines whether the given loop has enough CPU ops that can be | |
124 | overlapped with cache missing memory ops. If not, the loop | |
125 | won't benefit from prefetching. In the implementation, | |
126 | prefetching is not considered beneficial if the ratio between | |
127 | the instruction count and the mem ref count is below a certain | |
128 | minimum. | |
129 | ||
130 | 3. Function insn_to_prefetch_ratio_too_small_p implements a | |
131 | heuristic that disables prefetching in a loop if the prefetching | |
132 | cost is above a certain limit. The relative prefetching cost is | |
133 | estimated by taking the ratio between the prefetch count and the | |
134 | total intruction count (this models the I-cache cost). | |
135 | ||
db34470d | 136 | The limits used in these heuristics are defined as parameters with |
b8698a0f | 137 | reasonable default values. Machine-specific default values will be |
db34470d | 138 | added later. |
b8698a0f | 139 | |
b076a3fd ZD |
140 | Some other TODO: |
141 | -- write and use more general reuse analysis (that could be also used | |
142 | in other cache aimed loop optimizations) | |
143 | -- make it behave sanely together with the prefetches given by user | |
144 | (now we just ignore them; at the very least we should avoid | |
145 | optimizing loops in that user put his own prefetches) | |
146 | -- we assume cache line size alignment of arrays; this could be | |
147 | improved. */ | |
148 | ||
149 | /* Magic constants follow. These should be replaced by machine specific | |
150 | numbers. */ | |
151 | ||
b076a3fd ZD |
152 | /* True if write can be prefetched by a read prefetch. */ |
153 | ||
154 | #ifndef WRITE_CAN_USE_READ_PREFETCH | |
155 | #define WRITE_CAN_USE_READ_PREFETCH 1 | |
156 | #endif | |
157 | ||
158 | /* True if read can be prefetched by a write prefetch. */ | |
159 | ||
160 | #ifndef READ_CAN_USE_WRITE_PREFETCH | |
161 | #define READ_CAN_USE_WRITE_PREFETCH 0 | |
162 | #endif | |
163 | ||
47eb5b32 ZD |
164 | /* The size of the block loaded by a single prefetch. Usually, this is |
165 | the same as cache line size (at the moment, we only consider one level | |
166 | of cache hierarchy). */ | |
b076a3fd ZD |
167 | |
168 | #ifndef PREFETCH_BLOCK | |
028d4092 | 169 | #define PREFETCH_BLOCK param_l1_cache_line_size |
b076a3fd ZD |
170 | #endif |
171 | ||
172 | /* Do we have a forward hardware sequential prefetching? */ | |
173 | ||
174 | #ifndef HAVE_FORWARD_PREFETCH | |
175 | #define HAVE_FORWARD_PREFETCH 0 | |
176 | #endif | |
177 | ||
178 | /* Do we have a backward hardware sequential prefetching? */ | |
179 | ||
180 | #ifndef HAVE_BACKWARD_PREFETCH | |
181 | #define HAVE_BACKWARD_PREFETCH 0 | |
182 | #endif | |
183 | ||
184 | /* In some cases we are only able to determine that there is a certain | |
185 | probability that the two accesses hit the same cache line. In this | |
186 | case, we issue the prefetches for both of them if this probability | |
fa10beec | 187 | is less then (1000 - ACCEPTABLE_MISS_RATE) per thousand. */ |
b076a3fd ZD |
188 | |
189 | #ifndef ACCEPTABLE_MISS_RATE | |
190 | #define ACCEPTABLE_MISS_RATE 50 | |
191 | #endif | |
192 | ||
028d4092 ML |
193 | #define L1_CACHE_SIZE_BYTES ((unsigned) (param_l1_cache_size * 1024)) |
194 | #define L2_CACHE_SIZE_BYTES ((unsigned) (param_l2_cache_size * 1024)) | |
5417e022 ZD |
195 | |
196 | /* We consider a memory access nontemporal if it is not reused sooner than | |
197 | after L2_CACHE_SIZE_BYTES of memory are accessed. However, we ignore | |
198 | accesses closer than L1_CACHE_SIZE_BYTES / NONTEMPORAL_FRACTION, | |
199 | so that we use nontemporal prefetches e.g. if single memory location | |
200 | is accessed several times in a single iteration of the loop. */ | |
201 | #define NONTEMPORAL_FRACTION 16 | |
202 | ||
79f5e442 ZD |
203 | /* In case we have to emit a memory fence instruction after the loop that |
204 | uses nontemporal stores, this defines the builtin to use. */ | |
205 | ||
206 | #ifndef FENCE_FOLLOWING_MOVNT | |
207 | #define FENCE_FOLLOWING_MOVNT NULL_TREE | |
208 | #endif | |
209 | ||
9bf4598b CF |
210 | /* It is not profitable to prefetch when the trip count is not at |
211 | least TRIP_COUNT_TO_AHEAD_RATIO times the prefetch ahead distance. | |
212 | For example, in a loop with a prefetch ahead distance of 10, | |
213 | supposing that TRIP_COUNT_TO_AHEAD_RATIO is equal to 4, it is | |
214 | profitable to prefetch when the trip count is greater or equal to | |
215 | 40. In that case, 30 out of the 40 iterations will benefit from | |
216 | prefetching. */ | |
217 | ||
218 | #ifndef TRIP_COUNT_TO_AHEAD_RATIO | |
219 | #define TRIP_COUNT_TO_AHEAD_RATIO 4 | |
220 | #endif | |
221 | ||
b076a3fd ZD |
222 | /* The group of references between that reuse may occur. */ |
223 | ||
224 | struct mem_ref_group | |
225 | { | |
226 | tree base; /* Base of the reference. */ | |
81f32326 | 227 | tree step; /* Step of the reference. */ |
b076a3fd ZD |
228 | struct mem_ref *refs; /* References in the group. */ |
229 | struct mem_ref_group *next; /* Next group of references. */ | |
23b0f9f8 | 230 | unsigned int uid; /* Group UID, used only for debugging. */ |
b076a3fd ZD |
231 | }; |
232 | ||
233 | /* Assigned to PREFETCH_BEFORE when all iterations are to be prefetched. */ | |
234 | ||
dd4786fe | 235 | #define PREFETCH_ALL HOST_WIDE_INT_M1U |
b076a3fd | 236 | |
8532678c CF |
237 | /* Do not generate a prefetch if the unroll factor is significantly less |
238 | than what is required by the prefetch. This is to avoid redundant | |
f7963a7c CF |
239 | prefetches. For example, when prefetch_mod is 16 and unroll_factor is |
240 | 2, prefetching requires unrolling the loop 16 times, but | |
241 | the loop is actually unrolled twice. In this case (ratio = 8), | |
8532678c CF |
242 | prefetching is not likely to be beneficial. */ |
243 | ||
244 | #ifndef PREFETCH_MOD_TO_UNROLL_FACTOR_RATIO | |
f7963a7c | 245 | #define PREFETCH_MOD_TO_UNROLL_FACTOR_RATIO 4 |
8532678c CF |
246 | #endif |
247 | ||
0bbe50f6 CF |
248 | /* Some of the prefetch computations have quadratic complexity. We want to |
249 | avoid huge compile times and, therefore, want to limit the amount of | |
250 | memory references per loop where we consider prefetching. */ | |
251 | ||
252 | #ifndef PREFETCH_MAX_MEM_REFS_PER_LOOP | |
253 | #define PREFETCH_MAX_MEM_REFS_PER_LOOP 200 | |
254 | #endif | |
255 | ||
b076a3fd ZD |
256 | /* The memory reference. */ |
257 | ||
258 | struct mem_ref | |
259 | { | |
355fe088 | 260 | gimple *stmt; /* Statement in that the reference appears. */ |
b076a3fd ZD |
261 | tree mem; /* The reference. */ |
262 | HOST_WIDE_INT delta; /* Constant offset of the reference. */ | |
b076a3fd ZD |
263 | struct mem_ref_group *group; /* The group of references it belongs to. */ |
264 | unsigned HOST_WIDE_INT prefetch_mod; | |
265 | /* Prefetch only each PREFETCH_MOD-th | |
266 | iteration. */ | |
267 | unsigned HOST_WIDE_INT prefetch_before; | |
268 | /* Prefetch only first PREFETCH_BEFORE | |
269 | iterations. */ | |
5417e022 ZD |
270 | unsigned reuse_distance; /* The amount of data accessed before the first |
271 | reuse of this value. */ | |
b076a3fd | 272 | struct mem_ref *next; /* The next reference in the group. */ |
23b0f9f8 | 273 | unsigned int uid; /* Ref UID, used only for debugging. */ |
79f5e442 ZD |
274 | unsigned write_p : 1; /* Is it a write? */ |
275 | unsigned independent_p : 1; /* True if the reference is independent on | |
276 | all other references inside the loop. */ | |
277 | unsigned issue_prefetch_p : 1; /* Should we really issue the prefetch? */ | |
278 | unsigned storent_p : 1; /* True if we changed the store to a | |
279 | nontemporal one. */ | |
b076a3fd ZD |
280 | }; |
281 | ||
a5497b12 | 282 | /* Dumps information about memory reference */ |
b076a3fd | 283 | static void |
a5497b12 VK |
284 | dump_mem_details (FILE *file, tree base, tree step, |
285 | HOST_WIDE_INT delta, bool write_p) | |
b076a3fd | 286 | { |
a5497b12 VK |
287 | fprintf (file, "(base "); |
288 | print_generic_expr (file, base, TDF_SLIM); | |
b076a3fd | 289 | fprintf (file, ", step "); |
a5497b12 VK |
290 | if (cst_and_fits_in_hwi (step)) |
291 | fprintf (file, HOST_WIDE_INT_PRINT_DEC, int_cst_value (step)); | |
81f32326 | 292 | else |
8264c84d | 293 | print_generic_expr (file, step, TDF_SLIM); |
b076a3fd | 294 | fprintf (file, ")\n"); |
23b0f9f8 MK |
295 | fprintf (file, " delta " HOST_WIDE_INT_PRINT_DEC "\n", delta); |
296 | fprintf (file, " %s\n\n", write_p ? "write" : "read"); | |
a5497b12 | 297 | } |
b076a3fd | 298 | |
a5497b12 | 299 | /* Dumps information about reference REF to FILE. */ |
b076a3fd | 300 | |
a5497b12 VK |
301 | static void |
302 | dump_mem_ref (FILE *file, struct mem_ref *ref) | |
303 | { | |
23b0f9f8 MK |
304 | fprintf (file, "reference %u:%u (", ref->group->uid, ref->uid); |
305 | print_generic_expr (file, ref->mem, TDF_SLIM); | |
306 | fprintf (file, ")\n"); | |
b076a3fd ZD |
307 | } |
308 | ||
309 | /* Finds a group with BASE and STEP in GROUPS, or creates one if it does not | |
310 | exist. */ | |
311 | ||
312 | static struct mem_ref_group * | |
81f32326 | 313 | find_or_create_group (struct mem_ref_group **groups, tree base, tree step) |
b076a3fd | 314 | { |
23b0f9f8 MK |
315 | /* Global count for setting struct mem_ref_group->uid. */ |
316 | static unsigned int last_mem_ref_group_uid = 0; | |
317 | ||
b076a3fd ZD |
318 | struct mem_ref_group *group; |
319 | ||
320 | for (; *groups; groups = &(*groups)->next) | |
321 | { | |
81f32326 | 322 | if (operand_equal_p ((*groups)->step, step, 0) |
b076a3fd ZD |
323 | && operand_equal_p ((*groups)->base, base, 0)) |
324 | return *groups; | |
325 | ||
81f32326 CB |
326 | /* If step is an integer constant, keep the list of groups sorted |
327 | by decreasing step. */ | |
21c0a521 DM |
328 | if (cst_and_fits_in_hwi ((*groups)->step) && cst_and_fits_in_hwi (step) |
329 | && int_cst_value ((*groups)->step) < int_cst_value (step)) | |
b076a3fd ZD |
330 | break; |
331 | } | |
332 | ||
5417e022 | 333 | group = XNEW (struct mem_ref_group); |
b076a3fd ZD |
334 | group->base = base; |
335 | group->step = step; | |
336 | group->refs = NULL; | |
23b0f9f8 | 337 | group->uid = ++last_mem_ref_group_uid; |
b076a3fd ZD |
338 | group->next = *groups; |
339 | *groups = group; | |
340 | ||
341 | return group; | |
342 | } | |
343 | ||
344 | /* Records a memory reference MEM in GROUP with offset DELTA and write status | |
345 | WRITE_P. The reference occurs in statement STMT. */ | |
346 | ||
347 | static void | |
355fe088 | 348 | record_ref (struct mem_ref_group *group, gimple *stmt, tree mem, |
b076a3fd ZD |
349 | HOST_WIDE_INT delta, bool write_p) |
350 | { | |
23b0f9f8 | 351 | unsigned int last_mem_ref_uid = 0; |
b076a3fd ZD |
352 | struct mem_ref **aref; |
353 | ||
354 | /* Do not record the same address twice. */ | |
355 | for (aref = &group->refs; *aref; aref = &(*aref)->next) | |
356 | { | |
23b0f9f8 MK |
357 | last_mem_ref_uid = (*aref)->uid; |
358 | ||
b076a3fd ZD |
359 | /* It does not have to be possible for write reference to reuse the read |
360 | prefetch, or vice versa. */ | |
361 | if (!WRITE_CAN_USE_READ_PREFETCH | |
362 | && write_p | |
363 | && !(*aref)->write_p) | |
364 | continue; | |
365 | if (!READ_CAN_USE_WRITE_PREFETCH | |
366 | && !write_p | |
367 | && (*aref)->write_p) | |
368 | continue; | |
369 | ||
370 | if ((*aref)->delta == delta) | |
371 | return; | |
372 | } | |
373 | ||
5417e022 | 374 | (*aref) = XNEW (struct mem_ref); |
b076a3fd ZD |
375 | (*aref)->stmt = stmt; |
376 | (*aref)->mem = mem; | |
377 | (*aref)->delta = delta; | |
378 | (*aref)->write_p = write_p; | |
379 | (*aref)->prefetch_before = PREFETCH_ALL; | |
380 | (*aref)->prefetch_mod = 1; | |
5417e022 | 381 | (*aref)->reuse_distance = 0; |
b076a3fd ZD |
382 | (*aref)->issue_prefetch_p = false; |
383 | (*aref)->group = group; | |
384 | (*aref)->next = NULL; | |
79f5e442 ZD |
385 | (*aref)->independent_p = false; |
386 | (*aref)->storent_p = false; | |
23b0f9f8 | 387 | (*aref)->uid = last_mem_ref_uid + 1; |
b076a3fd ZD |
388 | |
389 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
23b0f9f8 MK |
390 | { |
391 | dump_mem_ref (dump_file, *aref); | |
392 | ||
393 | fprintf (dump_file, " group %u ", group->uid); | |
394 | dump_mem_details (dump_file, group->base, group->step, delta, | |
395 | write_p); | |
396 | } | |
b076a3fd ZD |
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 | { | |
99b1c316 | 423 | class loop *loop; /* Loop of the reference. */ |
355fe088 | 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 | |
99b1c316 | 488 | analyze_ref (class loop *loop, tree *ref_p, tree *base, |
81f32326 | 489 | tree *step, HOST_WIDE_INT *delta, |
355fe088 | 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 |
99b1c316 | 537 | gather_memory_references_ref (class loop *loop, struct mem_ref_group **refs, |
355fe088 | 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 ); | |
8264c84d ML |
566 | print_generic_expr (dump_file, ref, TDF_SLIM); |
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 ); | |
8264c84d | 583 | print_generic_expr (dump_file, ref, TDF_SLIM); |
a5497b12 | 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 * | |
99b1c316 | 608 | gather_memory_references (class 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 | 613 | gimple_stmt_iterator bsi; |
355fe088 | 614 | gimple *stmt; |
726a989a | 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 | |
eb066284 RB |
642 | if (! gimple_vuse (stmt)) |
643 | continue; | |
644 | ||
726a989a RB |
645 | lhs = gimple_assign_lhs (stmt); |
646 | rhs = gimple_assign_rhs1 (stmt); | |
b076a3fd ZD |
647 | |
648 | if (REFERENCE_CLASS_P (rhs)) | |
db34470d | 649 | { |
79f5e442 ZD |
650 | *no_other_refs &= gather_memory_references_ref (loop, &refs, |
651 | rhs, false, stmt); | |
db34470d GS |
652 | *ref_count += 1; |
653 | } | |
b076a3fd | 654 | if (REFERENCE_CLASS_P (lhs)) |
db34470d | 655 | { |
79f5e442 ZD |
656 | *no_other_refs &= gather_memory_references_ref (loop, &refs, |
657 | lhs, true, stmt); | |
db34470d GS |
658 | *ref_count += 1; |
659 | } | |
b076a3fd ZD |
660 | } |
661 | } | |
662 | free (body); | |
663 | ||
664 | return refs; | |
665 | } | |
666 | ||
667 | /* Prune the prefetch candidate REF using the self-reuse. */ | |
668 | ||
669 | static void | |
670 | prune_ref_by_self_reuse (struct mem_ref *ref) | |
671 | { | |
81f32326 CB |
672 | HOST_WIDE_INT step; |
673 | bool backward; | |
674 | ||
675 | /* If the step size is non constant, we cannot calculate prefetch_mod. */ | |
676 | if (!cst_and_fits_in_hwi (ref->group->step)) | |
677 | return; | |
678 | ||
679 | step = int_cst_value (ref->group->step); | |
680 | ||
681 | backward = step < 0; | |
b076a3fd ZD |
682 | |
683 | if (step == 0) | |
684 | { | |
685 | /* Prefetch references to invariant address just once. */ | |
686 | ref->prefetch_before = 1; | |
687 | return; | |
688 | } | |
689 | ||
690 | if (backward) | |
691 | step = -step; | |
692 | ||
693 | if (step > PREFETCH_BLOCK) | |
694 | return; | |
695 | ||
696 | if ((backward && HAVE_BACKWARD_PREFETCH) | |
697 | || (!backward && HAVE_FORWARD_PREFETCH)) | |
698 | { | |
699 | ref->prefetch_before = 1; | |
700 | return; | |
701 | } | |
702 | ||
703 | ref->prefetch_mod = PREFETCH_BLOCK / step; | |
704 | } | |
705 | ||
706 | /* Divides X by BY, rounding down. */ | |
707 | ||
708 | static HOST_WIDE_INT | |
709 | ddown (HOST_WIDE_INT x, unsigned HOST_WIDE_INT by) | |
710 | { | |
711 | gcc_assert (by > 0); | |
712 | ||
713 | if (x >= 0) | |
56e1a4d7 | 714 | return x / (HOST_WIDE_INT) by; |
b076a3fd | 715 | else |
56e1a4d7 | 716 | return (x + (HOST_WIDE_INT) by - 1) / (HOST_WIDE_INT) by; |
b076a3fd ZD |
717 | } |
718 | ||
b8698a0f L |
719 | /* Given a CACHE_LINE_SIZE and two inductive memory references |
720 | with a common STEP greater than CACHE_LINE_SIZE and an address | |
721 | difference DELTA, compute the probability that they will fall | |
14e444c3 CF |
722 | in different cache lines. Return true if the computed miss rate |
723 | is not greater than the ACCEPTABLE_MISS_RATE. DISTINCT_ITERS is the | |
724 | number of distinct iterations after which the pattern repeats itself. | |
2c6dd136 GS |
725 | ALIGN_UNIT is the unit of alignment in bytes. */ |
726 | ||
14e444c3 CF |
727 | static bool |
728 | is_miss_rate_acceptable (unsigned HOST_WIDE_INT cache_line_size, | |
2c6dd136 GS |
729 | HOST_WIDE_INT step, HOST_WIDE_INT delta, |
730 | unsigned HOST_WIDE_INT distinct_iters, | |
731 | int align_unit) | |
732 | { | |
733 | unsigned align, iter; | |
14e444c3 | 734 | int total_positions, miss_positions, max_allowed_miss_positions; |
2c6dd136 GS |
735 | int address1, address2, cache_line1, cache_line2; |
736 | ||
a245c04b CF |
737 | /* It always misses if delta is greater than or equal to the cache |
738 | line size. */ | |
14e444c3 CF |
739 | if (delta >= (HOST_WIDE_INT) cache_line_size) |
740 | return false; | |
a245c04b | 741 | |
2c6dd136 | 742 | miss_positions = 0; |
14e444c3 CF |
743 | total_positions = (cache_line_size / align_unit) * distinct_iters; |
744 | max_allowed_miss_positions = (ACCEPTABLE_MISS_RATE * total_positions) / 1000; | |
b8698a0f | 745 | |
2c6dd136 GS |
746 | /* Iterate through all possible alignments of the first |
747 | memory reference within its cache line. */ | |
748 | for (align = 0; align < cache_line_size; align += align_unit) | |
749 | ||
750 | /* Iterate through all distinct iterations. */ | |
751 | for (iter = 0; iter < distinct_iters; iter++) | |
752 | { | |
753 | address1 = align + step * iter; | |
754 | address2 = address1 + delta; | |
755 | cache_line1 = address1 / cache_line_size; | |
756 | cache_line2 = address2 / cache_line_size; | |
2c6dd136 | 757 | if (cache_line1 != cache_line2) |
14e444c3 CF |
758 | { |
759 | miss_positions += 1; | |
760 | if (miss_positions > max_allowed_miss_positions) | |
761 | return false; | |
762 | } | |
2c6dd136 | 763 | } |
14e444c3 | 764 | return true; |
2c6dd136 GS |
765 | } |
766 | ||
b076a3fd ZD |
767 | /* Prune the prefetch candidate REF using the reuse with BY. |
768 | If BY_IS_BEFORE is true, BY is before REF in the loop. */ | |
769 | ||
770 | static void | |
771 | prune_ref_by_group_reuse (struct mem_ref *ref, struct mem_ref *by, | |
772 | bool by_is_before) | |
773 | { | |
81f32326 CB |
774 | HOST_WIDE_INT step; |
775 | bool backward; | |
b076a3fd ZD |
776 | HOST_WIDE_INT delta_r = ref->delta, delta_b = by->delta; |
777 | HOST_WIDE_INT delta = delta_b - delta_r; | |
778 | HOST_WIDE_INT hit_from; | |
779 | unsigned HOST_WIDE_INT prefetch_before, prefetch_block; | |
2c6dd136 GS |
780 | HOST_WIDE_INT reduced_step; |
781 | unsigned HOST_WIDE_INT reduced_prefetch_block; | |
782 | tree ref_type; | |
783 | int align_unit; | |
b076a3fd | 784 | |
81f32326 CB |
785 | /* If the step is non constant we cannot calculate prefetch_before. */ |
786 | if (!cst_and_fits_in_hwi (ref->group->step)) { | |
787 | return; | |
788 | } | |
789 | ||
790 | step = int_cst_value (ref->group->step); | |
791 | ||
792 | backward = step < 0; | |
793 | ||
794 | ||
b076a3fd ZD |
795 | if (delta == 0) |
796 | { | |
797 | /* If the references has the same address, only prefetch the | |
798 | former. */ | |
799 | if (by_is_before) | |
800 | ref->prefetch_before = 0; | |
b8698a0f | 801 | |
b076a3fd ZD |
802 | return; |
803 | } | |
804 | ||
805 | if (!step) | |
806 | { | |
807 | /* If the reference addresses are invariant and fall into the | |
808 | same cache line, prefetch just the first one. */ | |
809 | if (!by_is_before) | |
810 | return; | |
811 | ||
812 | if (ddown (ref->delta, PREFETCH_BLOCK) | |
813 | != ddown (by->delta, PREFETCH_BLOCK)) | |
814 | return; | |
815 | ||
816 | ref->prefetch_before = 0; | |
817 | return; | |
818 | } | |
819 | ||
820 | /* Only prune the reference that is behind in the array. */ | |
821 | if (backward) | |
822 | { | |
823 | if (delta > 0) | |
824 | return; | |
825 | ||
826 | /* Transform the data so that we may assume that the accesses | |
827 | are forward. */ | |
828 | delta = - delta; | |
829 | step = -step; | |
830 | delta_r = PREFETCH_BLOCK - 1 - delta_r; | |
831 | delta_b = PREFETCH_BLOCK - 1 - delta_b; | |
832 | } | |
833 | else | |
834 | { | |
835 | if (delta < 0) | |
836 | return; | |
837 | } | |
838 | ||
839 | /* Check whether the two references are likely to hit the same cache | |
840 | line, and how distant the iterations in that it occurs are from | |
841 | each other. */ | |
842 | ||
843 | if (step <= PREFETCH_BLOCK) | |
844 | { | |
845 | /* The accesses are sure to meet. Let us check when. */ | |
846 | hit_from = ddown (delta_b, PREFETCH_BLOCK) * PREFETCH_BLOCK; | |
847 | prefetch_before = (hit_from - delta_r + step - 1) / step; | |
848 | ||
57762e97 | 849 | /* Do not reduce prefetch_before if we meet beyond cache size. */ |
4c9cf7af | 850 | if (prefetch_before > absu_hwi (L2_CACHE_SIZE_BYTES / step)) |
57762e97 | 851 | prefetch_before = PREFETCH_ALL; |
b076a3fd ZD |
852 | if (prefetch_before < ref->prefetch_before) |
853 | ref->prefetch_before = prefetch_before; | |
854 | ||
855 | return; | |
856 | } | |
857 | ||
b8698a0f | 858 | /* A more complicated case with step > prefetch_block. First reduce |
2c6dd136 | 859 | the ratio between the step and the cache line size to its simplest |
b8698a0f L |
860 | terms. The resulting denominator will then represent the number of |
861 | distinct iterations after which each address will go back to its | |
862 | initial location within the cache line. This computation assumes | |
2c6dd136 | 863 | that PREFETCH_BLOCK is a power of two. */ |
b076a3fd | 864 | prefetch_block = PREFETCH_BLOCK; |
2c6dd136 GS |
865 | reduced_prefetch_block = prefetch_block; |
866 | reduced_step = step; | |
867 | while ((reduced_step & 1) == 0 | |
868 | && reduced_prefetch_block > 1) | |
b076a3fd | 869 | { |
2c6dd136 GS |
870 | reduced_step >>= 1; |
871 | reduced_prefetch_block >>= 1; | |
b076a3fd ZD |
872 | } |
873 | ||
b076a3fd ZD |
874 | prefetch_before = delta / step; |
875 | delta %= step; | |
2c6dd136 GS |
876 | ref_type = TREE_TYPE (ref->mem); |
877 | align_unit = TYPE_ALIGN (ref_type) / 8; | |
14e444c3 CF |
878 | if (is_miss_rate_acceptable (prefetch_block, step, delta, |
879 | reduced_prefetch_block, align_unit)) | |
b076a3fd | 880 | { |
57762e97 CB |
881 | /* Do not reduce prefetch_before if we meet beyond cache size. */ |
882 | if (prefetch_before > L2_CACHE_SIZE_BYTES / PREFETCH_BLOCK) | |
883 | prefetch_before = PREFETCH_ALL; | |
b076a3fd ZD |
884 | if (prefetch_before < ref->prefetch_before) |
885 | ref->prefetch_before = prefetch_before; | |
886 | ||
887 | return; | |
888 | } | |
889 | ||
890 | /* Try also the following iteration. */ | |
891 | prefetch_before++; | |
892 | delta = step - delta; | |
14e444c3 CF |
893 | if (is_miss_rate_acceptable (prefetch_block, step, delta, |
894 | reduced_prefetch_block, align_unit)) | |
b076a3fd ZD |
895 | { |
896 | if (prefetch_before < ref->prefetch_before) | |
897 | ref->prefetch_before = prefetch_before; | |
898 | ||
899 | return; | |
900 | } | |
901 | ||
902 | /* The ref probably does not reuse by. */ | |
903 | return; | |
904 | } | |
905 | ||
906 | /* Prune the prefetch candidate REF using the reuses with other references | |
907 | in REFS. */ | |
908 | ||
909 | static void | |
910 | prune_ref_by_reuse (struct mem_ref *ref, struct mem_ref *refs) | |
911 | { | |
912 | struct mem_ref *prune_by; | |
913 | bool before = true; | |
914 | ||
915 | prune_ref_by_self_reuse (ref); | |
916 | ||
917 | for (prune_by = refs; prune_by; prune_by = prune_by->next) | |
918 | { | |
919 | if (prune_by == ref) | |
920 | { | |
921 | before = false; | |
922 | continue; | |
923 | } | |
924 | ||
925 | if (!WRITE_CAN_USE_READ_PREFETCH | |
926 | && ref->write_p | |
927 | && !prune_by->write_p) | |
928 | continue; | |
929 | if (!READ_CAN_USE_WRITE_PREFETCH | |
930 | && !ref->write_p | |
931 | && prune_by->write_p) | |
932 | continue; | |
933 | ||
934 | prune_ref_by_group_reuse (ref, prune_by, before); | |
935 | } | |
936 | } | |
937 | ||
938 | /* Prune the prefetch candidates in GROUP using the reuse analysis. */ | |
939 | ||
940 | static void | |
941 | prune_group_by_reuse (struct mem_ref_group *group) | |
942 | { | |
943 | struct mem_ref *ref_pruned; | |
944 | ||
945 | for (ref_pruned = group->refs; ref_pruned; ref_pruned = ref_pruned->next) | |
946 | { | |
947 | prune_ref_by_reuse (ref_pruned, group->refs); | |
948 | ||
949 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
950 | { | |
23b0f9f8 | 951 | dump_mem_ref (dump_file, ref_pruned); |
b076a3fd ZD |
952 | |
953 | if (ref_pruned->prefetch_before == PREFETCH_ALL | |
954 | && ref_pruned->prefetch_mod == 1) | |
955 | fprintf (dump_file, " no restrictions"); | |
956 | else if (ref_pruned->prefetch_before == 0) | |
957 | fprintf (dump_file, " do not prefetch"); | |
958 | else if (ref_pruned->prefetch_before <= ref_pruned->prefetch_mod) | |
959 | fprintf (dump_file, " prefetch once"); | |
960 | else | |
961 | { | |
962 | if (ref_pruned->prefetch_before != PREFETCH_ALL) | |
963 | { | |
964 | fprintf (dump_file, " prefetch before "); | |
965 | fprintf (dump_file, HOST_WIDE_INT_PRINT_DEC, | |
966 | ref_pruned->prefetch_before); | |
967 | } | |
968 | if (ref_pruned->prefetch_mod != 1) | |
969 | { | |
970 | fprintf (dump_file, " prefetch mod "); | |
971 | fprintf (dump_file, HOST_WIDE_INT_PRINT_DEC, | |
972 | ref_pruned->prefetch_mod); | |
973 | } | |
974 | } | |
975 | fprintf (dump_file, "\n"); | |
976 | } | |
977 | } | |
978 | } | |
979 | ||
980 | /* Prune the list of prefetch candidates GROUPS using the reuse analysis. */ | |
981 | ||
982 | static void | |
983 | prune_by_reuse (struct mem_ref_group *groups) | |
984 | { | |
985 | for (; groups; groups = groups->next) | |
986 | prune_group_by_reuse (groups); | |
987 | } | |
988 | ||
989 | /* Returns true if we should issue prefetch for REF. */ | |
990 | ||
991 | static bool | |
992 | should_issue_prefetch_p (struct mem_ref *ref) | |
993 | { | |
d2ff35c0 | 994 | /* Do we want to issue prefetches for non-constant strides? */ |
028d4092 ML |
995 | if (!cst_and_fits_in_hwi (ref->group->step) |
996 | && param_prefetch_dynamic_strides == 0) | |
d2ff35c0 LM |
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)) | |
028d4092 | 1011 | < (HOST_WIDE_INT) param_prefetch_minimum_stride) |
59100dfc LM |
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 | 1017 | ref->group->uid, ref->uid, int_cst_value (ref->group->step), |
028d4092 | 1018 | param_prefetch_minimum_stride); |
59100dfc LM |
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 | ||
028d4092 ML |
1058 | /* At most param_simultaneous_prefetches should be running |
1059 | at the same time. */ | |
1060 | remaining_prefetch_slots = param_simultaneous_prefetches; | |
b076a3fd | 1061 | |
911b3fdb ZD |
1062 | /* The prefetch will run for AHEAD iterations of the original loop, i.e., |
1063 | AHEAD / UNROLL_FACTOR iterations of the unrolled loop. In each iteration, | |
1064 | it will need a prefetch slot. */ | |
1065 | slots_per_prefetch = (ahead + unroll_factor / 2) / unroll_factor; | |
b076a3fd | 1066 | if (dump_file && (dump_flags & TDF_DETAILS)) |
911b3fdb ZD |
1067 | fprintf (dump_file, "Each prefetch instruction takes %u prefetch slots.\n", |
1068 | slots_per_prefetch); | |
b076a3fd ZD |
1069 | |
1070 | /* For now we just take memory references one by one and issue | |
1071 | prefetches for as many as possible. The groups are sorted | |
1072 | starting with the largest step, since the references with | |
c0220ea4 | 1073 | large step are more likely to cause many cache misses. */ |
b076a3fd ZD |
1074 | |
1075 | for (; groups; groups = groups->next) | |
1076 | for (ref = groups->refs; ref; ref = ref->next) | |
1077 | { | |
1078 | if (!should_issue_prefetch_p (ref)) | |
1079 | continue; | |
1080 | ||
8532678c CF |
1081 | /* The loop is far from being sufficiently unrolled for this |
1082 | prefetch. Do not generate prefetch to avoid many redudant | |
1083 | prefetches. */ | |
1084 | if (ref->prefetch_mod / unroll_factor > PREFETCH_MOD_TO_UNROLL_FACTOR_RATIO) | |
1085 | continue; | |
1086 | ||
911b3fdb ZD |
1087 | /* If we need to prefetch the reference each PREFETCH_MOD iterations, |
1088 | and we unroll the loop UNROLL_FACTOR times, we need to insert | |
1089 | ceil (UNROLL_FACTOR / PREFETCH_MOD) instructions in each | |
1090 | iteration. */ | |
b076a3fd ZD |
1091 | n_prefetches = ((unroll_factor + ref->prefetch_mod - 1) |
1092 | / ref->prefetch_mod); | |
911b3fdb ZD |
1093 | prefetch_slots = n_prefetches * slots_per_prefetch; |
1094 | ||
1095 | /* If more than half of the prefetches would be lost anyway, do not | |
1096 | issue the prefetch. */ | |
1097 | if (2 * remaining_prefetch_slots < prefetch_slots) | |
1098 | continue; | |
1099 | ||
1a70c8d5 MK |
1100 | /* Stop prefetching if debug counter is activated. */ |
1101 | if (!dbg_cnt (prefetch)) | |
1102 | continue; | |
1103 | ||
911b3fdb | 1104 | ref->issue_prefetch_p = true; |
23b0f9f8 MK |
1105 | if (dump_file && (dump_flags & TDF_DETAILS)) |
1106 | fprintf (dump_file, "Decided to issue prefetch for reference %u:%u\n", | |
1107 | ref->group->uid, ref->uid); | |
b076a3fd | 1108 | |
911b3fdb ZD |
1109 | if (remaining_prefetch_slots <= prefetch_slots) |
1110 | return true; | |
1111 | remaining_prefetch_slots -= prefetch_slots; | |
b076a3fd ZD |
1112 | any = true; |
1113 | } | |
1114 | ||
1115 | return any; | |
1116 | } | |
1117 | ||
d5058523 CF |
1118 | /* Return TRUE if no prefetch is going to be generated in the given |
1119 | GROUPS. */ | |
1120 | ||
1121 | static bool | |
1122 | nothing_to_prefetch_p (struct mem_ref_group *groups) | |
1123 | { | |
1124 | struct mem_ref *ref; | |
1125 | ||
1126 | for (; groups; groups = groups->next) | |
1127 | for (ref = groups->refs; ref; ref = ref->next) | |
1128 | if (should_issue_prefetch_p (ref)) | |
1129 | return false; | |
1130 | ||
1131 | return true; | |
1132 | } | |
1133 | ||
1134 | /* Estimate the number of prefetches in the given GROUPS. | |
1135 | UNROLL_FACTOR is the factor by which LOOP was unrolled. */ | |
b076a3fd | 1136 | |
db34470d | 1137 | static int |
d5058523 | 1138 | estimate_prefetch_count (struct mem_ref_group *groups, unsigned unroll_factor) |
b076a3fd ZD |
1139 | { |
1140 | struct mem_ref *ref; | |
d5058523 | 1141 | unsigned n_prefetches; |
db34470d | 1142 | int prefetch_count = 0; |
b076a3fd ZD |
1143 | |
1144 | for (; groups; groups = groups->next) | |
1145 | for (ref = groups->refs; ref; ref = ref->next) | |
1146 | if (should_issue_prefetch_p (ref)) | |
d5058523 CF |
1147 | { |
1148 | n_prefetches = ((unroll_factor + ref->prefetch_mod - 1) | |
1149 | / ref->prefetch_mod); | |
1150 | prefetch_count += n_prefetches; | |
1151 | } | |
b076a3fd | 1152 | |
db34470d | 1153 | return prefetch_count; |
b076a3fd ZD |
1154 | } |
1155 | ||
1156 | /* Issue prefetches for the reference REF into loop as decided before. | |
1157 | HEAD is the number of iterations to prefetch ahead. UNROLL_FACTOR | |
917f1b7e | 1158 | is the factor by which LOOP was unrolled. */ |
b076a3fd ZD |
1159 | |
1160 | static void | |
1161 | issue_prefetch_ref (struct mem_ref *ref, unsigned unroll_factor, unsigned ahead) | |
1162 | { | |
1163 | HOST_WIDE_INT delta; | |
81f32326 | 1164 | tree addr, addr_base, write_p, local, forward; |
538dd0b7 | 1165 | gcall *prefetch; |
726a989a | 1166 | gimple_stmt_iterator bsi; |
b076a3fd | 1167 | unsigned n_prefetches, ap; |
5417e022 | 1168 | bool nontemporal = ref->reuse_distance >= L2_CACHE_SIZE_BYTES; |
b076a3fd ZD |
1169 | |
1170 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
23b0f9f8 | 1171 | fprintf (dump_file, "Issued%s prefetch for reference %u:%u.\n", |
5417e022 | 1172 | nontemporal ? " nontemporal" : "", |
23b0f9f8 | 1173 | ref->group->uid, ref->uid); |
b076a3fd | 1174 | |
726a989a | 1175 | bsi = gsi_for_stmt (ref->stmt); |
b076a3fd ZD |
1176 | |
1177 | n_prefetches = ((unroll_factor + ref->prefetch_mod - 1) | |
1178 | / ref->prefetch_mod); | |
1179 | addr_base = build_fold_addr_expr_with_type (ref->mem, ptr_type_node); | |
726a989a RB |
1180 | addr_base = force_gimple_operand_gsi (&bsi, unshare_expr (addr_base), |
1181 | true, NULL, true, GSI_SAME_STMT); | |
911b3fdb | 1182 | write_p = ref->write_p ? integer_one_node : integer_zero_node; |
9a9d280e | 1183 | local = nontemporal ? integer_zero_node : integer_three_node; |
b076a3fd ZD |
1184 | |
1185 | for (ap = 0; ap < n_prefetches; ap++) | |
1186 | { | |
81f32326 CB |
1187 | if (cst_and_fits_in_hwi (ref->group->step)) |
1188 | { | |
1189 | /* Determine the address to prefetch. */ | |
1190 | delta = (ahead + ap * ref->prefetch_mod) * | |
1191 | int_cst_value (ref->group->step); | |
5d49b6a7 | 1192 | addr = fold_build_pointer_plus_hwi (addr_base, delta); |
23b0f9f8 MK |
1193 | addr = force_gimple_operand_gsi (&bsi, unshare_expr (addr), true, |
1194 | NULL, true, GSI_SAME_STMT); | |
81f32326 CB |
1195 | } |
1196 | else | |
1197 | { | |
1198 | /* The step size is non-constant but loop-invariant. We use the | |
1199 | heuristic to simply prefetch ahead iterations ahead. */ | |
1200 | forward = fold_build2 (MULT_EXPR, sizetype, | |
1201 | fold_convert (sizetype, ref->group->step), | |
1202 | fold_convert (sizetype, size_int (ahead))); | |
5d49b6a7 | 1203 | addr = fold_build_pointer_plus (addr_base, forward); |
81f32326 CB |
1204 | addr = force_gimple_operand_gsi (&bsi, unshare_expr (addr), true, |
1205 | NULL, true, GSI_SAME_STMT); | |
1206 | } | |
ed481942 | 1207 | |
fe8b1e83 DG |
1208 | if (addr_base != addr |
1209 | && TREE_CODE (addr_base) == SSA_NAME | |
ed481942 DG |
1210 | && TREE_CODE (addr) == SSA_NAME) |
1211 | { | |
1212 | duplicate_ssa_name_ptr_info (addr, SSA_NAME_PTR_INFO (addr_base)); | |
1213 | /* As this isn't a plain copy we have to reset alignment | |
1214 | information. */ | |
1215 | if (SSA_NAME_PTR_INFO (addr)) | |
1216 | mark_ptr_info_alignment_unknown (SSA_NAME_PTR_INFO (addr)); | |
1217 | } | |
1218 | ||
b076a3fd | 1219 | /* Create the prefetch instruction. */ |
e79983f4 | 1220 | prefetch = gimple_build_call (builtin_decl_explicit (BUILT_IN_PREFETCH), |
726a989a RB |
1221 | 3, addr, write_p, local); |
1222 | gsi_insert_before (&bsi, prefetch, GSI_SAME_STMT); | |
b076a3fd ZD |
1223 | } |
1224 | } | |
1225 | ||
1226 | /* Issue prefetches for the references in GROUPS into loop as decided before. | |
1227 | HEAD is the number of iterations to prefetch ahead. UNROLL_FACTOR is the | |
1228 | factor by that LOOP was unrolled. */ | |
1229 | ||
1230 | static void | |
1231 | issue_prefetches (struct mem_ref_group *groups, | |
1232 | unsigned unroll_factor, unsigned ahead) | |
1233 | { | |
1234 | struct mem_ref *ref; | |
1235 | ||
1236 | for (; groups; groups = groups->next) | |
1237 | for (ref = groups->refs; ref; ref = ref->next) | |
1238 | if (ref->issue_prefetch_p) | |
1239 | issue_prefetch_ref (ref, unroll_factor, ahead); | |
1240 | } | |
1241 | ||
79f5e442 ZD |
1242 | /* Returns true if REF is a memory write for that a nontemporal store insn |
1243 | can be used. */ | |
1244 | ||
1245 | static bool | |
1246 | nontemporal_store_p (struct mem_ref *ref) | |
1247 | { | |
ef4bddc2 | 1248 | machine_mode mode; |
79f5e442 ZD |
1249 | enum insn_code code; |
1250 | ||
1251 | /* REF must be a write that is not reused. We require it to be independent | |
1252 | on all other memory references in the loop, as the nontemporal stores may | |
1253 | be reordered with respect to other memory references. */ | |
1254 | if (!ref->write_p | |
1255 | || !ref->independent_p | |
1256 | || ref->reuse_distance < L2_CACHE_SIZE_BYTES) | |
1257 | return false; | |
1258 | ||
1259 | /* Check that we have the storent instruction for the mode. */ | |
1260 | mode = TYPE_MODE (TREE_TYPE (ref->mem)); | |
1261 | if (mode == BLKmode) | |
1262 | return false; | |
1263 | ||
947131ba | 1264 | code = optab_handler (storent_optab, mode); |
79f5e442 ZD |
1265 | return code != CODE_FOR_nothing; |
1266 | } | |
1267 | ||
1268 | /* If REF is a nontemporal store, we mark the corresponding modify statement | |
1269 | and return true. Otherwise, we return false. */ | |
1270 | ||
1271 | static bool | |
1272 | mark_nontemporal_store (struct mem_ref *ref) | |
1273 | { | |
1274 | if (!nontemporal_store_p (ref)) | |
1275 | return false; | |
1276 | ||
1277 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
23b0f9f8 MK |
1278 | fprintf (dump_file, "Marked reference %u:%u as a nontemporal store.\n", |
1279 | ref->group->uid, ref->uid); | |
79f5e442 | 1280 | |
726a989a | 1281 | gimple_assign_set_nontemporal_move (ref->stmt, true); |
79f5e442 ZD |
1282 | ref->storent_p = true; |
1283 | ||
1284 | return true; | |
1285 | } | |
1286 | ||
1287 | /* Issue a memory fence instruction after LOOP. */ | |
1288 | ||
1289 | static void | |
99b1c316 | 1290 | emit_mfence_after_loop (class loop *loop) |
79f5e442 | 1291 | { |
9771b263 | 1292 | vec<edge> exits = get_loop_exit_edges (loop); |
79f5e442 | 1293 | edge exit; |
538dd0b7 | 1294 | gcall *call; |
726a989a | 1295 | gimple_stmt_iterator bsi; |
79f5e442 ZD |
1296 | unsigned i; |
1297 | ||
9771b263 | 1298 | FOR_EACH_VEC_ELT (exits, i, exit) |
79f5e442 | 1299 | { |
726a989a | 1300 | call = gimple_build_call (FENCE_FOLLOWING_MOVNT, 0); |
79f5e442 ZD |
1301 | |
1302 | if (!single_pred_p (exit->dest) | |
1303 | /* If possible, we prefer not to insert the fence on other paths | |
1304 | in cfg. */ | |
1305 | && !(exit->flags & EDGE_ABNORMAL)) | |
1306 | split_loop_exit_edge (exit); | |
726a989a | 1307 | bsi = gsi_after_labels (exit->dest); |
79f5e442 | 1308 | |
726a989a | 1309 | gsi_insert_before (&bsi, call, GSI_NEW_STMT); |
79f5e442 ZD |
1310 | } |
1311 | ||
9771b263 | 1312 | exits.release (); |
79f5e442 ZD |
1313 | update_ssa (TODO_update_ssa_only_virtuals); |
1314 | } | |
1315 | ||
1316 | /* Returns true if we can use storent in loop, false otherwise. */ | |
1317 | ||
1318 | static bool | |
99b1c316 | 1319 | may_use_storent_in_loop_p (class loop *loop) |
79f5e442 ZD |
1320 | { |
1321 | bool ret = true; | |
1322 | ||
1323 | if (loop->inner != NULL) | |
1324 | return false; | |
1325 | ||
1326 | /* If we must issue a mfence insn after using storent, check that there | |
1327 | is a suitable place for it at each of the loop exits. */ | |
1328 | if (FENCE_FOLLOWING_MOVNT != NULL_TREE) | |
1329 | { | |
9771b263 | 1330 | vec<edge> exits = get_loop_exit_edges (loop); |
79f5e442 ZD |
1331 | unsigned i; |
1332 | edge exit; | |
1333 | ||
9771b263 | 1334 | FOR_EACH_VEC_ELT (exits, i, exit) |
79f5e442 | 1335 | if ((exit->flags & EDGE_ABNORMAL) |
fefa31b5 | 1336 | && exit->dest == EXIT_BLOCK_PTR_FOR_FN (cfun)) |
79f5e442 ZD |
1337 | ret = false; |
1338 | ||
9771b263 | 1339 | exits.release (); |
79f5e442 ZD |
1340 | } |
1341 | ||
1342 | return ret; | |
1343 | } | |
1344 | ||
1345 | /* Marks nontemporal stores in LOOP. GROUPS contains the description of memory | |
1346 | references in the loop. */ | |
1347 | ||
1348 | static void | |
99b1c316 | 1349 | mark_nontemporal_stores (class loop *loop, struct mem_ref_group *groups) |
79f5e442 ZD |
1350 | { |
1351 | struct mem_ref *ref; | |
1352 | bool any = false; | |
1353 | ||
1354 | if (!may_use_storent_in_loop_p (loop)) | |
1355 | return; | |
1356 | ||
1357 | for (; groups; groups = groups->next) | |
1358 | for (ref = groups->refs; ref; ref = ref->next) | |
1359 | any |= mark_nontemporal_store (ref); | |
1360 | ||
1361 | if (any && FENCE_FOLLOWING_MOVNT != NULL_TREE) | |
1362 | emit_mfence_after_loop (loop); | |
1363 | } | |
1364 | ||
b076a3fd ZD |
1365 | /* Determines whether we can profitably unroll LOOP FACTOR times, and if |
1366 | this is the case, fill in DESC by the description of number of | |
1367 | iterations. */ | |
1368 | ||
1369 | static bool | |
99b1c316 | 1370 | should_unroll_loop_p (class loop *loop, class tree_niter_desc *desc, |
b076a3fd ZD |
1371 | unsigned factor) |
1372 | { | |
1373 | if (!can_unroll_loop_p (loop, factor, desc)) | |
1374 | return false; | |
1375 | ||
1376 | /* We only consider loops without control flow for unrolling. This is not | |
1377 | a hard restriction -- tree_unroll_loop works with arbitrary loops | |
1378 | as well; but the unrolling/prefetching is usually more profitable for | |
1379 | loops consisting of a single basic block, and we want to limit the | |
1380 | code growth. */ | |
1381 | if (loop->num_nodes > 2) | |
1382 | return false; | |
1383 | ||
1384 | return true; | |
1385 | } | |
1386 | ||
1387 | /* Determine the coefficient by that unroll LOOP, from the information | |
1388 | contained in the list of memory references REFS. Description of | |
23b0f9f8 | 1389 | number of iterations of LOOP is stored to DESC. NINSNS is the number of |
2711355f ZD |
1390 | insns of the LOOP. EST_NITER is the estimated number of iterations of |
1391 | the loop, or -1 if no estimate is available. */ | |
b076a3fd ZD |
1392 | |
1393 | static unsigned | |
99b1c316 MS |
1394 | determine_unroll_factor (class loop *loop, struct mem_ref_group *refs, |
1395 | unsigned ninsns, class tree_niter_desc *desc, | |
2711355f | 1396 | HOST_WIDE_INT est_niter) |
b076a3fd | 1397 | { |
911b3fdb ZD |
1398 | unsigned upper_bound; |
1399 | unsigned nfactor, factor, mod_constraint; | |
b076a3fd ZD |
1400 | struct mem_ref_group *agp; |
1401 | struct mem_ref *ref; | |
1402 | ||
911b3fdb ZD |
1403 | /* First check whether the loop is not too large to unroll. We ignore |
1404 | PARAM_MAX_UNROLL_TIMES, because for small loops, it prevented us | |
1405 | from unrolling them enough to make exactly one cache line covered by each | |
1406 | iteration. Also, the goal of PARAM_MAX_UNROLL_TIMES is to prevent | |
1407 | us from unrolling the loops too many times in cases where we only expect | |
1408 | gains from better scheduling and decreasing loop overhead, which is not | |
1409 | the case here. */ | |
028d4092 | 1410 | upper_bound = param_max_unrolled_insns / ninsns; |
2711355f ZD |
1411 | |
1412 | /* If we unrolled the loop more times than it iterates, the unrolled version | |
1413 | of the loop would be never entered. */ | |
1414 | if (est_niter >= 0 && est_niter < (HOST_WIDE_INT) upper_bound) | |
1415 | upper_bound = est_niter; | |
1416 | ||
911b3fdb | 1417 | if (upper_bound <= 1) |
b076a3fd ZD |
1418 | return 1; |
1419 | ||
911b3fdb ZD |
1420 | /* Choose the factor so that we may prefetch each cache just once, |
1421 | but bound the unrolling by UPPER_BOUND. */ | |
1422 | factor = 1; | |
b076a3fd ZD |
1423 | for (agp = refs; agp; agp = agp->next) |
1424 | for (ref = agp->refs; ref; ref = ref->next) | |
911b3fdb ZD |
1425 | if (should_issue_prefetch_p (ref)) |
1426 | { | |
1427 | mod_constraint = ref->prefetch_mod; | |
1428 | nfactor = least_common_multiple (mod_constraint, factor); | |
1429 | if (nfactor <= upper_bound) | |
1430 | factor = nfactor; | |
1431 | } | |
b076a3fd ZD |
1432 | |
1433 | if (!should_unroll_loop_p (loop, desc, factor)) | |
1434 | return 1; | |
1435 | ||
1436 | return factor; | |
1437 | } | |
1438 | ||
5417e022 ZD |
1439 | /* Returns the total volume of the memory references REFS, taking into account |
1440 | reuses in the innermost loop and cache line size. TODO -- we should also | |
1441 | take into account reuses across the iterations of the loops in the loop | |
1442 | nest. */ | |
1443 | ||
1444 | static unsigned | |
1445 | volume_of_references (struct mem_ref_group *refs) | |
1446 | { | |
1447 | unsigned volume = 0; | |
1448 | struct mem_ref_group *gr; | |
1449 | struct mem_ref *ref; | |
1450 | ||
1451 | for (gr = refs; gr; gr = gr->next) | |
1452 | for (ref = gr->refs; ref; ref = ref->next) | |
1453 | { | |
1454 | /* Almost always reuses another value? */ | |
1455 | if (ref->prefetch_before != PREFETCH_ALL) | |
1456 | continue; | |
1457 | ||
1458 | /* If several iterations access the same cache line, use the size of | |
1459 | the line divided by this number. Otherwise, a cache line is | |
1460 | accessed in each iteration. TODO -- in the latter case, we should | |
1461 | take the size of the reference into account, rounding it up on cache | |
1462 | line size multiple. */ | |
028d4092 | 1463 | volume += param_l1_cache_line_size / ref->prefetch_mod; |
5417e022 ZD |
1464 | } |
1465 | return volume; | |
1466 | } | |
1467 | ||
1468 | /* Returns the volume of memory references accessed across VEC iterations of | |
1469 | loops, whose sizes are described in the LOOP_SIZES array. N is the number | |
1470 | of the loops in the nest (length of VEC and LOOP_SIZES vectors). */ | |
1471 | ||
1472 | static unsigned | |
1473 | volume_of_dist_vector (lambda_vector vec, unsigned *loop_sizes, unsigned n) | |
1474 | { | |
1475 | unsigned i; | |
1476 | ||
1477 | for (i = 0; i < n; i++) | |
1478 | if (vec[i] != 0) | |
1479 | break; | |
1480 | ||
1481 | if (i == n) | |
1482 | return 0; | |
1483 | ||
1484 | gcc_assert (vec[i] > 0); | |
1485 | ||
1486 | /* We ignore the parts of the distance vector in subloops, since usually | |
1487 | the numbers of iterations are much smaller. */ | |
1488 | return loop_sizes[i] * vec[i]; | |
1489 | } | |
1490 | ||
1491 | /* Add the steps of ACCESS_FN multiplied by STRIDE to the array STRIDE | |
1492 | at the position corresponding to the loop of the step. N is the depth | |
1493 | of the considered loop nest, and, LOOP is its innermost loop. */ | |
1494 | ||
1495 | static void | |
1496 | add_subscript_strides (tree access_fn, unsigned stride, | |
99b1c316 | 1497 | HOST_WIDE_INT *strides, unsigned n, class loop *loop) |
5417e022 | 1498 | { |
99b1c316 | 1499 | class loop *aloop; |
5417e022 ZD |
1500 | tree step; |
1501 | HOST_WIDE_INT astep; | |
1502 | unsigned min_depth = loop_depth (loop) - n; | |
1503 | ||
1504 | while (TREE_CODE (access_fn) == POLYNOMIAL_CHREC) | |
1505 | { | |
1506 | aloop = get_chrec_loop (access_fn); | |
1507 | step = CHREC_RIGHT (access_fn); | |
1508 | access_fn = CHREC_LEFT (access_fn); | |
1509 | ||
1510 | if ((unsigned) loop_depth (aloop) <= min_depth) | |
1511 | continue; | |
1512 | ||
9541ffee | 1513 | if (tree_fits_shwi_p (step)) |
9439e9a1 | 1514 | astep = tree_to_shwi (step); |
5417e022 | 1515 | else |
028d4092 | 1516 | astep = param_l1_cache_line_size; |
5417e022 ZD |
1517 | |
1518 | strides[n - 1 - loop_depth (loop) + loop_depth (aloop)] += astep * stride; | |
1519 | ||
1520 | } | |
1521 | } | |
1522 | ||
1523 | /* Returns the volume of memory references accessed between two consecutive | |
1524 | self-reuses of the reference DR. We consider the subscripts of DR in N | |
1525 | loops, and LOOP_SIZES contains the volumes of accesses in each of the | |
1526 | loops. LOOP is the innermost loop of the current loop nest. */ | |
1527 | ||
1528 | static unsigned | |
1529 | self_reuse_distance (data_reference_p dr, unsigned *loop_sizes, unsigned n, | |
99b1c316 | 1530 | class loop *loop) |
5417e022 ZD |
1531 | { |
1532 | tree stride, access_fn; | |
1533 | HOST_WIDE_INT *strides, astride; | |
9771b263 | 1534 | vec<tree> access_fns; |
5417e022 ZD |
1535 | tree ref = DR_REF (dr); |
1536 | unsigned i, ret = ~0u; | |
1537 | ||
1538 | /* In the following example: | |
1539 | ||
1540 | for (i = 0; i < N; i++) | |
1541 | for (j = 0; j < N; j++) | |
1542 | use (a[j][i]); | |
1543 | the same cache line is accessed each N steps (except if the change from | |
1544 | i to i + 1 crosses the boundary of the cache line). Thus, for self-reuse, | |
1545 | we cannot rely purely on the results of the data dependence analysis. | |
1546 | ||
1547 | Instead, we compute the stride of the reference in each loop, and consider | |
1548 | the innermost loop in that the stride is less than cache size. */ | |
1549 | ||
1550 | strides = XCNEWVEC (HOST_WIDE_INT, n); | |
1551 | access_fns = DR_ACCESS_FNS (dr); | |
1552 | ||
9771b263 | 1553 | FOR_EACH_VEC_ELT (access_fns, i, access_fn) |
5417e022 ZD |
1554 | { |
1555 | /* Keep track of the reference corresponding to the subscript, so that we | |
1556 | know its stride. */ | |
1557 | while (handled_component_p (ref) && TREE_CODE (ref) != ARRAY_REF) | |
1558 | ref = TREE_OPERAND (ref, 0); | |
b8698a0f | 1559 | |
5417e022 ZD |
1560 | if (TREE_CODE (ref) == ARRAY_REF) |
1561 | { | |
1562 | stride = TYPE_SIZE_UNIT (TREE_TYPE (ref)); | |
cc269bb6 | 1563 | if (tree_fits_uhwi_p (stride)) |
ae7e9ddd | 1564 | astride = tree_to_uhwi (stride); |
5417e022 | 1565 | else |
028d4092 | 1566 | astride = param_l1_cache_line_size; |
5417e022 ZD |
1567 | |
1568 | ref = TREE_OPERAND (ref, 0); | |
1569 | } | |
1570 | else | |
1571 | astride = 1; | |
1572 | ||
1573 | add_subscript_strides (access_fn, astride, strides, n, loop); | |
1574 | } | |
1575 | ||
1576 | for (i = n; i-- > 0; ) | |
1577 | { | |
1578 | unsigned HOST_WIDE_INT s; | |
1579 | ||
1580 | s = strides[i] < 0 ? -strides[i] : strides[i]; | |
1581 | ||
028d4092 | 1582 | if (s < (unsigned) param_l1_cache_line_size |
5417e022 ZD |
1583 | && (loop_sizes[i] |
1584 | > (unsigned) (L1_CACHE_SIZE_BYTES / NONTEMPORAL_FRACTION))) | |
1585 | { | |
1586 | ret = loop_sizes[i]; | |
1587 | break; | |
1588 | } | |
1589 | } | |
1590 | ||
1591 | free (strides); | |
1592 | return ret; | |
1593 | } | |
1594 | ||
1595 | /* Determines the distance till the first reuse of each reference in REFS | |
79f5e442 | 1596 | in the loop nest of LOOP. NO_OTHER_REFS is true if there are no other |
1e373390 | 1597 | memory references in the loop. Return false if the analysis fails. */ |
5417e022 | 1598 | |
1e373390 | 1599 | static bool |
99b1c316 | 1600 | determine_loop_nest_reuse (class loop *loop, struct mem_ref_group *refs, |
79f5e442 | 1601 | bool no_other_refs) |
5417e022 | 1602 | { |
99b1c316 | 1603 | class loop *nest, *aloop; |
6e1aa848 DN |
1604 | vec<data_reference_p> datarefs = vNULL; |
1605 | vec<ddr_p> dependences = vNULL; | |
5417e022 | 1606 | struct mem_ref_group *gr; |
79f5e442 | 1607 | struct mem_ref *ref, *refb; |
843374f2 | 1608 | auto_vec<loop_p> vloops; |
5417e022 ZD |
1609 | unsigned *loop_data_size; |
1610 | unsigned i, j, n; | |
1611 | unsigned volume, dist, adist; | |
1612 | HOST_WIDE_INT vol; | |
1613 | data_reference_p dr; | |
1614 | ddr_p dep; | |
1615 | ||
1616 | if (loop->inner) | |
1e373390 | 1617 | return true; |
5417e022 ZD |
1618 | |
1619 | /* Find the outermost loop of the loop nest of loop (we require that | |
1620 | there are no sibling loops inside the nest). */ | |
1621 | nest = loop; | |
1622 | while (1) | |
1623 | { | |
1624 | aloop = loop_outer (nest); | |
1625 | ||
1626 | if (aloop == current_loops->tree_root | |
1627 | || aloop->inner->next) | |
1628 | break; | |
1629 | ||
1630 | nest = aloop; | |
1631 | } | |
1632 | ||
1633 | /* For each loop, determine the amount of data accessed in each iteration. | |
1634 | We use this to estimate whether the reference is evicted from the | |
1635 | cache before its reuse. */ | |
1636 | find_loop_nest (nest, &vloops); | |
9771b263 | 1637 | n = vloops.length (); |
5417e022 ZD |
1638 | loop_data_size = XNEWVEC (unsigned, n); |
1639 | volume = volume_of_references (refs); | |
1640 | i = n; | |
1641 | while (i-- != 0) | |
1642 | { | |
1643 | loop_data_size[i] = volume; | |
1644 | /* Bound the volume by the L2 cache size, since above this bound, | |
1645 | all dependence distances are equivalent. */ | |
1646 | if (volume > L2_CACHE_SIZE_BYTES) | |
1647 | continue; | |
1648 | ||
9771b263 | 1649 | aloop = vloops[i]; |
652c4c71 | 1650 | vol = estimated_stmt_executions_int (aloop); |
e5b332cd | 1651 | if (vol == -1) |
5417e022 ZD |
1652 | vol = expected_loop_iterations (aloop); |
1653 | volume *= vol; | |
1654 | } | |
1655 | ||
1656 | /* Prepare the references in the form suitable for data dependence | |
0d52bcc1 | 1657 | analysis. We ignore unanalyzable data references (the results |
5417e022 ZD |
1658 | are used just as a heuristics to estimate temporality of the |
1659 | references, hence we do not need to worry about correctness). */ | |
1660 | for (gr = refs; gr; gr = gr->next) | |
1661 | for (ref = gr->refs; ref; ref = ref->next) | |
1662 | { | |
a68f286c RB |
1663 | dr = create_data_ref (loop_preheader_edge (nest), |
1664 | loop_containing_stmt (ref->stmt), | |
62c8a2cf | 1665 | ref->mem, ref->stmt, !ref->write_p, false); |
5417e022 ZD |
1666 | |
1667 | if (dr) | |
1668 | { | |
1669 | ref->reuse_distance = volume; | |
1670 | dr->aux = ref; | |
9771b263 | 1671 | datarefs.safe_push (dr); |
5417e022 | 1672 | } |
79f5e442 ZD |
1673 | else |
1674 | no_other_refs = false; | |
5417e022 ZD |
1675 | } |
1676 | ||
9771b263 | 1677 | FOR_EACH_VEC_ELT (datarefs, i, dr) |
5417e022 ZD |
1678 | { |
1679 | dist = self_reuse_distance (dr, loop_data_size, n, loop); | |
3d9a9f94 | 1680 | ref = (struct mem_ref *) dr->aux; |
5417e022 ZD |
1681 | if (ref->reuse_distance > dist) |
1682 | ref->reuse_distance = dist; | |
79f5e442 ZD |
1683 | |
1684 | if (no_other_refs) | |
1685 | ref->independent_p = true; | |
5417e022 ZD |
1686 | } |
1687 | ||
1e373390 RG |
1688 | if (!compute_all_dependences (datarefs, &dependences, vloops, true)) |
1689 | return false; | |
5417e022 | 1690 | |
9771b263 | 1691 | FOR_EACH_VEC_ELT (dependences, i, dep) |
5417e022 ZD |
1692 | { |
1693 | if (DDR_ARE_DEPENDENT (dep) == chrec_known) | |
1694 | continue; | |
1695 | ||
3d9a9f94 KG |
1696 | ref = (struct mem_ref *) DDR_A (dep)->aux; |
1697 | refb = (struct mem_ref *) DDR_B (dep)->aux; | |
79f5e442 | 1698 | |
5417e022 | 1699 | if (DDR_ARE_DEPENDENT (dep) == chrec_dont_know |
dfbddbeb | 1700 | || DDR_COULD_BE_INDEPENDENT_P (dep) |
5417e022 ZD |
1701 | || DDR_NUM_DIST_VECTS (dep) == 0) |
1702 | { | |
0d52bcc1 | 1703 | /* If the dependence cannot be analyzed, assume that there might be |
5417e022 ZD |
1704 | a reuse. */ |
1705 | dist = 0; | |
b8698a0f | 1706 | |
79f5e442 ZD |
1707 | ref->independent_p = false; |
1708 | refb->independent_p = false; | |
5417e022 ZD |
1709 | } |
1710 | else | |
1711 | { | |
0d52bcc1 | 1712 | /* The distance vectors are normalized to be always lexicographically |
5417e022 ZD |
1713 | positive, hence we cannot tell just from them whether DDR_A comes |
1714 | before DDR_B or vice versa. However, it is not important, | |
1715 | anyway -- if DDR_A is close to DDR_B, then it is either reused in | |
1716 | DDR_B (and it is not nontemporal), or it reuses the value of DDR_B | |
1717 | in cache (and marking it as nontemporal would not affect | |
1718 | anything). */ | |
1719 | ||
1720 | dist = volume; | |
1721 | for (j = 0; j < DDR_NUM_DIST_VECTS (dep); j++) | |
1722 | { | |
1723 | adist = volume_of_dist_vector (DDR_DIST_VECT (dep, j), | |
1724 | loop_data_size, n); | |
1725 | ||
79f5e442 ZD |
1726 | /* If this is a dependence in the innermost loop (i.e., the |
1727 | distances in all superloops are zero) and it is not | |
1728 | the trivial self-dependence with distance zero, record that | |
1729 | the references are not completely independent. */ | |
1730 | if (lambda_vector_zerop (DDR_DIST_VECT (dep, j), n - 1) | |
1731 | && (ref != refb | |
1732 | || DDR_DIST_VECT (dep, j)[n-1] != 0)) | |
1733 | { | |
1734 | ref->independent_p = false; | |
1735 | refb->independent_p = false; | |
1736 | } | |
1737 | ||
5417e022 ZD |
1738 | /* Ignore accesses closer than |
1739 | L1_CACHE_SIZE_BYTES / NONTEMPORAL_FRACTION, | |
1740 | so that we use nontemporal prefetches e.g. if single memory | |
1741 | location is accessed several times in a single iteration of | |
1742 | the loop. */ | |
1743 | if (adist < L1_CACHE_SIZE_BYTES / NONTEMPORAL_FRACTION) | |
1744 | continue; | |
1745 | ||
1746 | if (adist < dist) | |
1747 | dist = adist; | |
1748 | } | |
1749 | } | |
1750 | ||
5417e022 ZD |
1751 | if (ref->reuse_distance > dist) |
1752 | ref->reuse_distance = dist; | |
79f5e442 ZD |
1753 | if (refb->reuse_distance > dist) |
1754 | refb->reuse_distance = dist; | |
5417e022 ZD |
1755 | } |
1756 | ||
1757 | free_dependence_relations (dependences); | |
1758 | free_data_refs (datarefs); | |
1759 | free (loop_data_size); | |
1760 | ||
1761 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
1762 | { | |
1763 | fprintf (dump_file, "Reuse distances:\n"); | |
1764 | for (gr = refs; gr; gr = gr->next) | |
1765 | for (ref = gr->refs; ref; ref = ref->next) | |
23b0f9f8 MK |
1766 | fprintf (dump_file, " reference %u:%u distance %u\n", |
1767 | ref->group->uid, ref->uid, ref->reuse_distance); | |
5417e022 | 1768 | } |
1e373390 RG |
1769 | |
1770 | return true; | |
5417e022 ZD |
1771 | } |
1772 | ||
0bbe50f6 CF |
1773 | /* Determine whether or not the trip count to ahead ratio is too small based |
1774 | on prefitablility consideration. | |
db34470d | 1775 | AHEAD: the iteration ahead distance, |
0bbe50f6 CF |
1776 | EST_NITER: the estimated trip count. */ |
1777 | ||
1778 | static bool | |
1779 | trip_count_to_ahead_ratio_too_small_p (unsigned ahead, HOST_WIDE_INT est_niter) | |
1780 | { | |
1781 | /* Assume trip count to ahead ratio is big enough if the trip count could not | |
1782 | be estimated at compile time. */ | |
1783 | if (est_niter < 0) | |
1784 | return false; | |
1785 | ||
1786 | if (est_niter < (HOST_WIDE_INT) (TRIP_COUNT_TO_AHEAD_RATIO * ahead)) | |
1787 | { | |
1788 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
1789 | fprintf (dump_file, | |
1790 | "Not prefetching -- loop estimated to roll only %d times\n", | |
1791 | (int) est_niter); | |
1792 | return true; | |
1793 | } | |
1794 | ||
1795 | return false; | |
1796 | } | |
1797 | ||
1798 | /* Determine whether or not the number of memory references in the loop is | |
1799 | reasonable based on the profitablity and compilation time considerations. | |
db34470d | 1800 | NINSNS: estimated number of instructions in the loop, |
db34470d GS |
1801 | MEM_REF_COUNT: total number of memory references in the loop. */ |
1802 | ||
b8698a0f | 1803 | static bool |
0bbe50f6 | 1804 | mem_ref_count_reasonable_p (unsigned ninsns, unsigned mem_ref_count) |
db34470d | 1805 | { |
0bbe50f6 | 1806 | int insn_to_mem_ratio; |
db34470d GS |
1807 | |
1808 | if (mem_ref_count == 0) | |
1809 | return false; | |
1810 | ||
0bbe50f6 CF |
1811 | /* Miss rate computation (is_miss_rate_acceptable) and dependence analysis |
1812 | (compute_all_dependences) have high costs based on quadratic complexity. | |
1813 | To avoid huge compilation time, we give up prefetching if mem_ref_count | |
1814 | is too large. */ | |
1815 | if (mem_ref_count > PREFETCH_MAX_MEM_REFS_PER_LOOP) | |
1816 | return false; | |
1817 | ||
b8698a0f L |
1818 | /* Prefetching improves performance by overlapping cache missing |
1819 | memory accesses with CPU operations. If the loop does not have | |
1820 | enough CPU operations to overlap with memory operations, prefetching | |
1821 | won't give a significant benefit. One approximate way of checking | |
1822 | this is to require the ratio of instructions to memory references to | |
db34470d GS |
1823 | be above a certain limit. This approximation works well in practice. |
1824 | TODO: Implement a more precise computation by estimating the time | |
1825 | for each CPU or memory op in the loop. Time estimates for memory ops | |
1826 | should account for cache misses. */ | |
b8698a0f | 1827 | insn_to_mem_ratio = ninsns / mem_ref_count; |
db34470d | 1828 | |
028d4092 | 1829 | if (insn_to_mem_ratio < param_prefetch_min_insn_to_mem_ratio) |
55e5a2eb CF |
1830 | { |
1831 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
1832 | fprintf (dump_file, | |
1833 | "Not prefetching -- instruction to memory reference ratio (%d) too small\n", | |
1834 | insn_to_mem_ratio); | |
1835 | return false; | |
1836 | } | |
db34470d | 1837 | |
0bbe50f6 CF |
1838 | return true; |
1839 | } | |
1840 | ||
1841 | /* Determine whether or not the instruction to prefetch ratio in the loop is | |
1842 | too small based on the profitablity consideration. | |
1843 | NINSNS: estimated number of instructions in the loop, | |
1844 | PREFETCH_COUNT: an estimate of the number of prefetches, | |
1845 | UNROLL_FACTOR: the factor to unroll the loop if prefetching. */ | |
1846 | ||
1847 | static bool | |
1848 | insn_to_prefetch_ratio_too_small_p (unsigned ninsns, unsigned prefetch_count, | |
1849 | unsigned unroll_factor) | |
1850 | { | |
1851 | int insn_to_prefetch_ratio; | |
1852 | ||
d3a9b459 CF |
1853 | /* Prefetching most likely causes performance degradation when the instruction |
1854 | to prefetch ratio is too small. Too many prefetch instructions in a loop | |
1855 | may reduce the I-cache performance. | |
ccacf0e1 CF |
1856 | (unroll_factor * ninsns) is used to estimate the number of instructions in |
1857 | the unrolled loop. This implementation is a bit simplistic -- the number | |
1858 | of issued prefetch instructions is also affected by unrolling. So, | |
1859 | prefetch_mod and the unroll factor should be taken into account when | |
1860 | determining prefetch_count. Also, the number of insns of the unrolled | |
1861 | loop will usually be significantly smaller than the number of insns of the | |
1862 | original loop * unroll_factor (at least the induction variable increases | |
1863 | and the exit branches will get eliminated), so it might be better to use | |
1864 | tree_estimate_loop_size + estimated_unrolled_size. */ | |
d3a9b459 | 1865 | insn_to_prefetch_ratio = (unroll_factor * ninsns) / prefetch_count; |
028d4092 | 1866 | if (insn_to_prefetch_ratio < param_min_insn_to_prefetch_ratio) |
db34470d | 1867 | { |
d3a9b459 CF |
1868 | if (dump_file && (dump_flags & TDF_DETAILS)) |
1869 | fprintf (dump_file, | |
1870 | "Not prefetching -- instruction to prefetch ratio (%d) too small\n", | |
1871 | insn_to_prefetch_ratio); | |
0bbe50f6 | 1872 | return true; |
db34470d | 1873 | } |
b8698a0f | 1874 | |
0bbe50f6 | 1875 | return false; |
db34470d GS |
1876 | } |
1877 | ||
1878 | ||
b076a3fd | 1879 | /* Issue prefetch instructions for array references in LOOP. Returns |
d73be268 | 1880 | true if the LOOP was unrolled. */ |
b076a3fd ZD |
1881 | |
1882 | static bool | |
99b1c316 | 1883 | loop_prefetch_arrays (class loop *loop) |
b076a3fd ZD |
1884 | { |
1885 | struct mem_ref_group *refs; | |
2711355f ZD |
1886 | unsigned ahead, ninsns, time, unroll_factor; |
1887 | HOST_WIDE_INT est_niter; | |
99b1c316 | 1888 | class tree_niter_desc desc; |
79f5e442 | 1889 | bool unrolled = false, no_other_refs; |
db34470d GS |
1890 | unsigned prefetch_count; |
1891 | unsigned mem_ref_count; | |
b076a3fd | 1892 | |
efd8f750 | 1893 | if (optimize_loop_nest_for_size_p (loop)) |
2732d767 ZD |
1894 | { |
1895 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
1896 | fprintf (dump_file, " ignored (cold area)\n"); | |
1897 | return false; | |
1898 | } | |
1899 | ||
0bbe50f6 CF |
1900 | /* FIXME: the time should be weighted by the probabilities of the blocks in |
1901 | the loop body. */ | |
1902 | time = tree_num_loop_insns (loop, &eni_time_weights); | |
1903 | if (time == 0) | |
1904 | return false; | |
1905 | ||
028d4092 | 1906 | ahead = (param_prefetch_latency + time - 1) / time; |
652c4c71 | 1907 | est_niter = estimated_stmt_executions_int (loop); |
e5b332cd | 1908 | if (est_niter == -1) |
00022058 | 1909 | est_niter = likely_max_stmt_executions_int (loop); |
0bbe50f6 CF |
1910 | |
1911 | /* Prefetching is not likely to be profitable if the trip count to ahead | |
1912 | ratio is too small. */ | |
1913 | if (trip_count_to_ahead_ratio_too_small_p (ahead, est_niter)) | |
1914 | return false; | |
1915 | ||
1916 | ninsns = tree_num_loop_insns (loop, &eni_size_weights); | |
1917 | ||
b076a3fd | 1918 | /* Step 1: gather the memory references. */ |
db34470d | 1919 | refs = gather_memory_references (loop, &no_other_refs, &mem_ref_count); |
b076a3fd | 1920 | |
0bbe50f6 CF |
1921 | /* Give up prefetching if the number of memory references in the |
1922 | loop is not reasonable based on profitablity and compilation time | |
1923 | considerations. */ | |
1924 | if (!mem_ref_count_reasonable_p (ninsns, mem_ref_count)) | |
1925 | goto fail; | |
1926 | ||
b076a3fd ZD |
1927 | /* Step 2: estimate the reuse effects. */ |
1928 | prune_by_reuse (refs); | |
1929 | ||
d5058523 | 1930 | if (nothing_to_prefetch_p (refs)) |
b076a3fd ZD |
1931 | goto fail; |
1932 | ||
1e373390 RG |
1933 | if (!determine_loop_nest_reuse (loop, refs, no_other_refs)) |
1934 | goto fail; | |
5417e022 | 1935 | |
0bbe50f6 | 1936 | /* Step 3: determine unroll factor. */ |
2711355f ZD |
1937 | unroll_factor = determine_unroll_factor (loop, refs, ninsns, &desc, |
1938 | est_niter); | |
d5058523 CF |
1939 | |
1940 | /* Estimate prefetch count for the unrolled loop. */ | |
1941 | prefetch_count = estimate_prefetch_count (refs, unroll_factor); | |
1942 | if (prefetch_count == 0) | |
1943 | goto fail; | |
1944 | ||
2711355f | 1945 | if (dump_file && (dump_flags & TDF_DETAILS)) |
b8698a0f | 1946 | fprintf (dump_file, "Ahead %d, unroll factor %d, trip count " |
d81f5387 | 1947 | HOST_WIDE_INT_PRINT_DEC "\n" |
b8698a0f L |
1948 | "insn count %d, mem ref count %d, prefetch count %d\n", |
1949 | ahead, unroll_factor, est_niter, | |
1950 | ninsns, mem_ref_count, prefetch_count); | |
db34470d | 1951 | |
0bbe50f6 CF |
1952 | /* Prefetching is not likely to be profitable if the instruction to prefetch |
1953 | ratio is too small. */ | |
1954 | if (insn_to_prefetch_ratio_too_small_p (ninsns, prefetch_count, | |
1955 | unroll_factor)) | |
db34470d GS |
1956 | goto fail; |
1957 | ||
1958 | mark_nontemporal_stores (loop, refs); | |
2711355f | 1959 | |
b076a3fd ZD |
1960 | /* Step 4: what to prefetch? */ |
1961 | if (!schedule_prefetches (refs, unroll_factor, ahead)) | |
1962 | goto fail; | |
1963 | ||
1964 | /* Step 5: unroll the loop. TODO -- peeling of first and last few | |
1965 | iterations so that we do not issue superfluous prefetches. */ | |
1966 | if (unroll_factor != 1) | |
1967 | { | |
d73be268 | 1968 | tree_unroll_loop (loop, unroll_factor, |
b076a3fd ZD |
1969 | single_dom_exit (loop), &desc); |
1970 | unrolled = true; | |
1971 | } | |
1972 | ||
1973 | /* Step 6: issue the prefetches. */ | |
1974 | issue_prefetches (refs, unroll_factor, ahead); | |
1975 | ||
1976 | fail: | |
1977 | release_mem_refs (refs); | |
1978 | return unrolled; | |
1979 | } | |
1980 | ||
d73be268 | 1981 | /* Issue prefetch instructions for array references in loops. */ |
b076a3fd | 1982 | |
c7f965b6 | 1983 | unsigned int |
d73be268 | 1984 | tree_ssa_prefetch_arrays (void) |
b076a3fd | 1985 | { |
99b1c316 | 1986 | class loop *loop; |
b076a3fd | 1987 | bool unrolled = false; |
c7f965b6 | 1988 | int todo_flags = 0; |
b076a3fd | 1989 | |
134b044d | 1990 | if (!targetm.have_prefetch () |
b076a3fd ZD |
1991 | /* It is possible to ask compiler for say -mtune=i486 -march=pentium4. |
1992 | -mtune=i486 causes us having PREFETCH_BLOCK 0, since this is part | |
1993 | of processor costs and i486 does not have prefetch, but | |
134b044d | 1994 | -march=pentium4 causes targetm.have_prefetch to be true. Ugh. */ |
b076a3fd | 1995 | || PREFETCH_BLOCK == 0) |
c7f965b6 | 1996 | return 0; |
b076a3fd | 1997 | |
47eb5b32 ZD |
1998 | if (dump_file && (dump_flags & TDF_DETAILS)) |
1999 | { | |
2000 | fprintf (dump_file, "Prefetching parameters:\n"); | |
2001 | fprintf (dump_file, " simultaneous prefetches: %d\n", | |
028d4092 ML |
2002 | param_simultaneous_prefetches); |
2003 | fprintf (dump_file, " prefetch latency: %d\n", param_prefetch_latency); | |
47eb5b32 | 2004 | fprintf (dump_file, " prefetch block size: %d\n", PREFETCH_BLOCK); |
46cb0441 | 2005 | fprintf (dump_file, " L1 cache size: %d lines, %d kB\n", |
028d4092 ML |
2006 | L1_CACHE_SIZE_BYTES / param_l1_cache_line_size, |
2007 | param_l1_cache_size); | |
2008 | fprintf (dump_file, " L1 cache line size: %d\n", | |
2009 | param_l1_cache_line_size); | |
2010 | fprintf (dump_file, " L2 cache size: %d kB\n", param_l2_cache_size); | |
b8698a0f | 2011 | fprintf (dump_file, " min insn-to-prefetch ratio: %d \n", |
028d4092 | 2012 | param_min_insn_to_prefetch_ratio); |
b8698a0f | 2013 | fprintf (dump_file, " min insn-to-mem ratio: %d \n", |
028d4092 | 2014 | param_prefetch_min_insn_to_mem_ratio); |
47eb5b32 ZD |
2015 | fprintf (dump_file, "\n"); |
2016 | } | |
2017 | ||
b076a3fd ZD |
2018 | initialize_original_copy_tables (); |
2019 | ||
e79983f4 | 2020 | if (!builtin_decl_explicit_p (BUILT_IN_PREFETCH)) |
b076a3fd | 2021 | { |
6a4825bd NF |
2022 | tree type = build_function_type_list (void_type_node, |
2023 | const_ptr_type_node, NULL_TREE); | |
c79efc4d RÁE |
2024 | tree decl = add_builtin_function ("__builtin_prefetch", type, |
2025 | BUILT_IN_PREFETCH, BUILT_IN_NORMAL, | |
2026 | NULL, NULL_TREE); | |
b076a3fd | 2027 | DECL_IS_NOVOPS (decl) = true; |
e79983f4 | 2028 | set_builtin_decl (BUILT_IN_PREFETCH, decl, false); |
b076a3fd ZD |
2029 | } |
2030 | ||
f0bd40b1 | 2031 | FOR_EACH_LOOP (loop, LI_FROM_INNERMOST) |
b076a3fd | 2032 | { |
b076a3fd ZD |
2033 | if (dump_file && (dump_flags & TDF_DETAILS)) |
2034 | fprintf (dump_file, "Processing loop %d:\n", loop->num); | |
2035 | ||
d73be268 | 2036 | unrolled |= loop_prefetch_arrays (loop); |
b076a3fd ZD |
2037 | |
2038 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
2039 | fprintf (dump_file, "\n\n"); | |
2040 | } | |
2041 | ||
2042 | if (unrolled) | |
2043 | { | |
2044 | scev_reset (); | |
c7f965b6 | 2045 | todo_flags |= TODO_cleanup_cfg; |
b076a3fd ZD |
2046 | } |
2047 | ||
2048 | free_original_copy_tables (); | |
c7f965b6 | 2049 | return todo_flags; |
b076a3fd | 2050 | } |
71343877 AM |
2051 | |
2052 | /* Prefetching. */ | |
2053 | ||
71343877 AM |
2054 | namespace { |
2055 | ||
2056 | const pass_data pass_data_loop_prefetch = | |
2057 | { | |
2058 | GIMPLE_PASS, /* type */ | |
2059 | "aprefetch", /* name */ | |
2060 | OPTGROUP_LOOP, /* optinfo_flags */ | |
71343877 AM |
2061 | TV_TREE_PREFETCH, /* tv_id */ |
2062 | ( PROP_cfg | PROP_ssa ), /* properties_required */ | |
2063 | 0, /* properties_provided */ | |
2064 | 0, /* properties_destroyed */ | |
2065 | 0, /* todo_flags_start */ | |
2066 | 0, /* todo_flags_finish */ | |
2067 | }; | |
2068 | ||
2069 | class pass_loop_prefetch : public gimple_opt_pass | |
2070 | { | |
2071 | public: | |
2072 | pass_loop_prefetch (gcc::context *ctxt) | |
2073 | : gimple_opt_pass (pass_data_loop_prefetch, ctxt) | |
2074 | {} | |
2075 | ||
2076 | /* opt_pass methods: */ | |
1a3d085c | 2077 | virtual bool gate (function *) { return flag_prefetch_loop_arrays > 0; } |
be55bfe6 | 2078 | virtual unsigned int execute (function *); |
71343877 AM |
2079 | |
2080 | }; // class pass_loop_prefetch | |
2081 | ||
be55bfe6 TS |
2082 | unsigned int |
2083 | pass_loop_prefetch::execute (function *fun) | |
2084 | { | |
2085 | if (number_of_loops (fun) <= 1) | |
2086 | return 0; | |
2087 | ||
d78a1c01 ML |
2088 | if ((PREFETCH_BLOCK & (PREFETCH_BLOCK - 1)) != 0) |
2089 | { | |
2090 | static bool warned = false; | |
2091 | ||
2092 | if (!warned) | |
2093 | { | |
2094 | warning (OPT_Wdisabled_optimization, | |
2095 | "%<l1-cache-size%> parameter is not a power of two %d", | |
2096 | PREFETCH_BLOCK); | |
2097 | warned = true; | |
2098 | } | |
2099 | return 0; | |
2100 | } | |
2101 | ||
be55bfe6 TS |
2102 | return tree_ssa_prefetch_arrays (); |
2103 | } | |
2104 | ||
71343877 AM |
2105 | } // anon namespace |
2106 | ||
2107 | gimple_opt_pass * | |
2108 | make_pass_loop_prefetch (gcc::context *ctxt) | |
2109 | { | |
2110 | return new pass_loop_prefetch (ctxt); | |
2111 | } | |
2112 | ||
2113 |