]>
Commit | Line | Data |
---|---|---|
b076a3fd | 1 | /* Array prefetching. |
5624e564 | 2 | Copyright (C) 2005-2015 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" |
b076a3fd | 24 | #include "tree.h" |
c7131fb2 AM |
25 | #include "gimple.h" |
26 | #include "rtl.h" | |
27 | #include "alias.h" | |
40e23961 | 28 | #include "fold-const.h" |
d8a2d370 | 29 | #include "stor-layout.h" |
b076a3fd | 30 | #include "tm_p.h" |
cf835838 | 31 | #include "tree-pretty-print.h" |
2fb9a547 | 32 | #include "internal-fn.h" |
45b0be94 | 33 | #include "gimplify.h" |
5be5c238 | 34 | #include "gimple-iterator.h" |
18f429e2 | 35 | #include "gimplify-me.h" |
442b4905 | 36 | #include "gimple-ssa.h" |
e28030cf AM |
37 | #include "tree-ssa-loop-ivopts.h" |
38 | #include "tree-ssa-loop-manip.h" | |
39 | #include "tree-ssa-loop-niter.h" | |
442b4905 AM |
40 | #include "tree-ssa-loop.h" |
41 | #include "tree-into-ssa.h" | |
b076a3fd | 42 | #include "cfgloop.h" |
b076a3fd | 43 | #include "tree-pass.h" |
b076a3fd | 44 | #include "insn-config.h" |
b076a3fd ZD |
45 | #include "tree-chrec.h" |
46 | #include "tree-scalar-evolution.h" | |
718f9c0f | 47 | #include "diagnostic-core.h" |
b076a3fd ZD |
48 | #include "params.h" |
49 | #include "langhooks.h" | |
7f9bc51b | 50 | #include "tree-inline.h" |
5417e022 | 51 | #include "tree-data-ref.h" |
134b044d | 52 | #include "target.h" |
2eb79bbb SB |
53 | |
54 | ||
55 | /* FIXME: Needed for optabs, but this should all be moved to a TBD interface | |
56 | between the GIMPLE and RTL worlds. */ | |
36566b39 | 57 | #include "flags.h" |
36566b39 PK |
58 | #include "expmed.h" |
59 | #include "dojump.h" | |
60 | #include "explow.h" | |
61 | #include "calls.h" | |
62 | #include "emit-rtl.h" | |
63 | #include "varasm.h" | |
64 | #include "stmt.h" | |
2eb79bbb | 65 | #include "expr.h" |
b0710fe1 | 66 | #include "insn-codes.h" |
79f5e442 | 67 | #include "optabs.h" |
1c1ad7bb | 68 | #include "recog.h" |
b076a3fd ZD |
69 | |
70 | /* This pass inserts prefetch instructions to optimize cache usage during | |
71 | accesses to arrays in loops. It processes loops sequentially and: | |
72 | ||
73 | 1) Gathers all memory references in the single loop. | |
74 | 2) For each of the references it decides when it is profitable to prefetch | |
75 | it. To do it, we evaluate the reuse among the accesses, and determines | |
76 | two values: PREFETCH_BEFORE (meaning that it only makes sense to do | |
77 | prefetching in the first PREFETCH_BEFORE iterations of the loop) and | |
78 | PREFETCH_MOD (meaning that it only makes sense to prefetch in the | |
79 | iterations of the loop that are zero modulo PREFETCH_MOD). For example | |
80 | (assuming cache line size is 64 bytes, char has size 1 byte and there | |
81 | is no hardware sequential prefetch): | |
82 | ||
83 | char *a; | |
84 | for (i = 0; i < max; i++) | |
85 | { | |
86 | a[255] = ...; (0) | |
87 | a[i] = ...; (1) | |
88 | a[i + 64] = ...; (2) | |
89 | a[16*i] = ...; (3) | |
90 | a[187*i] = ...; (4) | |
91 | a[187*i + 50] = ...; (5) | |
92 | } | |
93 | ||
94 | (0) obviously has PREFETCH_BEFORE 1 | |
95 | (1) has PREFETCH_BEFORE 64, since (2) accesses the same memory | |
96 | location 64 iterations before it, and PREFETCH_MOD 64 (since | |
97 | it hits the same cache line otherwise). | |
98 | (2) has PREFETCH_MOD 64 | |
99 | (3) has PREFETCH_MOD 4 | |
100 | (4) has PREFETCH_MOD 1. We do not set PREFETCH_BEFORE here, since | |
bae077dc | 101 | the cache line accessed by (5) is the same with probability only |
b076a3fd ZD |
102 | 7/32. |
103 | (5) has PREFETCH_MOD 1 as well. | |
104 | ||
5417e022 ZD |
105 | Additionally, we use data dependence analysis to determine for each |
106 | reference the distance till the first reuse; this information is used | |
107 | to determine the temporality of the issued prefetch instruction. | |
108 | ||
b076a3fd ZD |
109 | 3) We determine how much ahead we need to prefetch. The number of |
110 | iterations needed is time to fetch / time spent in one iteration of | |
111 | the loop. The problem is that we do not know either of these values, | |
112 | so we just make a heuristic guess based on a magic (possibly) | |
113 | target-specific constant and size of the loop. | |
114 | ||
115 | 4) Determine which of the references we prefetch. We take into account | |
116 | that there is a maximum number of simultaneous prefetches (provided | |
117 | by machine description). We prefetch as many prefetches as possible | |
118 | while still within this bound (starting with those with lowest | |
119 | prefetch_mod, since they are responsible for most of the cache | |
120 | misses). | |
b8698a0f | 121 | |
b076a3fd ZD |
122 | 5) We unroll and peel loops so that we are able to satisfy PREFETCH_MOD |
123 | and PREFETCH_BEFORE requirements (within some bounds), and to avoid | |
124 | prefetching nonaccessed memory. | |
125 | TODO -- actually implement peeling. | |
b8698a0f | 126 | |
b076a3fd ZD |
127 | 6) We actually emit the prefetch instructions. ??? Perhaps emit the |
128 | prefetch instructions with guards in cases where 5) was not sufficient | |
129 | to satisfy the constraints? | |
130 | ||
0bbe50f6 CF |
131 | A cost model is implemented to determine whether or not prefetching is |
132 | profitable for a given loop. The cost model has three heuristics: | |
133 | ||
134 | 1. Function trip_count_to_ahead_ratio_too_small_p implements a | |
135 | heuristic that determines whether or not the loop has too few | |
136 | iterations (compared to ahead). Prefetching is not likely to be | |
137 | beneficial if the trip count to ahead ratio is below a certain | |
138 | minimum. | |
139 | ||
140 | 2. Function mem_ref_count_reasonable_p implements a heuristic that | |
141 | determines whether the given loop has enough CPU ops that can be | |
142 | overlapped with cache missing memory ops. If not, the loop | |
143 | won't benefit from prefetching. In the implementation, | |
144 | prefetching is not considered beneficial if the ratio between | |
145 | the instruction count and the mem ref count is below a certain | |
146 | minimum. | |
147 | ||
148 | 3. Function insn_to_prefetch_ratio_too_small_p implements a | |
149 | heuristic that disables prefetching in a loop if the prefetching | |
150 | cost is above a certain limit. The relative prefetching cost is | |
151 | estimated by taking the ratio between the prefetch count and the | |
152 | total intruction count (this models the I-cache cost). | |
153 | ||
db34470d | 154 | The limits used in these heuristics are defined as parameters with |
b8698a0f | 155 | reasonable default values. Machine-specific default values will be |
db34470d | 156 | added later. |
b8698a0f | 157 | |
b076a3fd ZD |
158 | Some other TODO: |
159 | -- write and use more general reuse analysis (that could be also used | |
160 | in other cache aimed loop optimizations) | |
161 | -- make it behave sanely together with the prefetches given by user | |
162 | (now we just ignore them; at the very least we should avoid | |
163 | optimizing loops in that user put his own prefetches) | |
164 | -- we assume cache line size alignment of arrays; this could be | |
165 | improved. */ | |
166 | ||
167 | /* Magic constants follow. These should be replaced by machine specific | |
168 | numbers. */ | |
169 | ||
b076a3fd ZD |
170 | /* True if write can be prefetched by a read prefetch. */ |
171 | ||
172 | #ifndef WRITE_CAN_USE_READ_PREFETCH | |
173 | #define WRITE_CAN_USE_READ_PREFETCH 1 | |
174 | #endif | |
175 | ||
176 | /* True if read can be prefetched by a write prefetch. */ | |
177 | ||
178 | #ifndef READ_CAN_USE_WRITE_PREFETCH | |
179 | #define READ_CAN_USE_WRITE_PREFETCH 0 | |
180 | #endif | |
181 | ||
47eb5b32 ZD |
182 | /* The size of the block loaded by a single prefetch. Usually, this is |
183 | the same as cache line size (at the moment, we only consider one level | |
184 | of cache hierarchy). */ | |
b076a3fd ZD |
185 | |
186 | #ifndef PREFETCH_BLOCK | |
47eb5b32 | 187 | #define PREFETCH_BLOCK L1_CACHE_LINE_SIZE |
b076a3fd ZD |
188 | #endif |
189 | ||
190 | /* Do we have a forward hardware sequential prefetching? */ | |
191 | ||
192 | #ifndef HAVE_FORWARD_PREFETCH | |
193 | #define HAVE_FORWARD_PREFETCH 0 | |
194 | #endif | |
195 | ||
196 | /* Do we have a backward hardware sequential prefetching? */ | |
197 | ||
198 | #ifndef HAVE_BACKWARD_PREFETCH | |
199 | #define HAVE_BACKWARD_PREFETCH 0 | |
200 | #endif | |
201 | ||
202 | /* In some cases we are only able to determine that there is a certain | |
203 | probability that the two accesses hit the same cache line. In this | |
204 | case, we issue the prefetches for both of them if this probability | |
fa10beec | 205 | is less then (1000 - ACCEPTABLE_MISS_RATE) per thousand. */ |
b076a3fd ZD |
206 | |
207 | #ifndef ACCEPTABLE_MISS_RATE | |
208 | #define ACCEPTABLE_MISS_RATE 50 | |
209 | #endif | |
210 | ||
46cb0441 ZD |
211 | #define L1_CACHE_SIZE_BYTES ((unsigned) (L1_CACHE_SIZE * 1024)) |
212 | #define L2_CACHE_SIZE_BYTES ((unsigned) (L2_CACHE_SIZE * 1024)) | |
5417e022 ZD |
213 | |
214 | /* We consider a memory access nontemporal if it is not reused sooner than | |
215 | after L2_CACHE_SIZE_BYTES of memory are accessed. However, we ignore | |
216 | accesses closer than L1_CACHE_SIZE_BYTES / NONTEMPORAL_FRACTION, | |
217 | so that we use nontemporal prefetches e.g. if single memory location | |
218 | is accessed several times in a single iteration of the loop. */ | |
219 | #define NONTEMPORAL_FRACTION 16 | |
220 | ||
79f5e442 ZD |
221 | /* In case we have to emit a memory fence instruction after the loop that |
222 | uses nontemporal stores, this defines the builtin to use. */ | |
223 | ||
224 | #ifndef FENCE_FOLLOWING_MOVNT | |
225 | #define FENCE_FOLLOWING_MOVNT NULL_TREE | |
226 | #endif | |
227 | ||
9bf4598b CF |
228 | /* It is not profitable to prefetch when the trip count is not at |
229 | least TRIP_COUNT_TO_AHEAD_RATIO times the prefetch ahead distance. | |
230 | For example, in a loop with a prefetch ahead distance of 10, | |
231 | supposing that TRIP_COUNT_TO_AHEAD_RATIO is equal to 4, it is | |
232 | profitable to prefetch when the trip count is greater or equal to | |
233 | 40. In that case, 30 out of the 40 iterations will benefit from | |
234 | prefetching. */ | |
235 | ||
236 | #ifndef TRIP_COUNT_TO_AHEAD_RATIO | |
237 | #define TRIP_COUNT_TO_AHEAD_RATIO 4 | |
238 | #endif | |
239 | ||
b076a3fd ZD |
240 | /* The group of references between that reuse may occur. */ |
241 | ||
242 | struct mem_ref_group | |
243 | { | |
244 | tree base; /* Base of the reference. */ | |
81f32326 | 245 | tree step; /* Step of the reference. */ |
b076a3fd ZD |
246 | struct mem_ref *refs; /* References in the group. */ |
247 | struct mem_ref_group *next; /* Next group of references. */ | |
248 | }; | |
249 | ||
250 | /* Assigned to PREFETCH_BEFORE when all iterations are to be prefetched. */ | |
251 | ||
252 | #define PREFETCH_ALL (~(unsigned HOST_WIDE_INT) 0) | |
253 | ||
8532678c CF |
254 | /* Do not generate a prefetch if the unroll factor is significantly less |
255 | than what is required by the prefetch. This is to avoid redundant | |
f7963a7c CF |
256 | prefetches. For example, when prefetch_mod is 16 and unroll_factor is |
257 | 2, prefetching requires unrolling the loop 16 times, but | |
258 | the loop is actually unrolled twice. In this case (ratio = 8), | |
8532678c CF |
259 | prefetching is not likely to be beneficial. */ |
260 | ||
261 | #ifndef PREFETCH_MOD_TO_UNROLL_FACTOR_RATIO | |
f7963a7c | 262 | #define PREFETCH_MOD_TO_UNROLL_FACTOR_RATIO 4 |
8532678c CF |
263 | #endif |
264 | ||
0bbe50f6 CF |
265 | /* Some of the prefetch computations have quadratic complexity. We want to |
266 | avoid huge compile times and, therefore, want to limit the amount of | |
267 | memory references per loop where we consider prefetching. */ | |
268 | ||
269 | #ifndef PREFETCH_MAX_MEM_REFS_PER_LOOP | |
270 | #define PREFETCH_MAX_MEM_REFS_PER_LOOP 200 | |
271 | #endif | |
272 | ||
b076a3fd ZD |
273 | /* The memory reference. */ |
274 | ||
275 | struct mem_ref | |
276 | { | |
726a989a | 277 | gimple stmt; /* Statement in that the reference appears. */ |
b076a3fd ZD |
278 | tree mem; /* The reference. */ |
279 | HOST_WIDE_INT delta; /* Constant offset of the reference. */ | |
b076a3fd ZD |
280 | struct mem_ref_group *group; /* The group of references it belongs to. */ |
281 | unsigned HOST_WIDE_INT prefetch_mod; | |
282 | /* Prefetch only each PREFETCH_MOD-th | |
283 | iteration. */ | |
284 | unsigned HOST_WIDE_INT prefetch_before; | |
285 | /* Prefetch only first PREFETCH_BEFORE | |
286 | iterations. */ | |
5417e022 ZD |
287 | unsigned reuse_distance; /* The amount of data accessed before the first |
288 | reuse of this value. */ | |
b076a3fd | 289 | struct mem_ref *next; /* The next reference in the group. */ |
79f5e442 ZD |
290 | unsigned write_p : 1; /* Is it a write? */ |
291 | unsigned independent_p : 1; /* True if the reference is independent on | |
292 | all other references inside the loop. */ | |
293 | unsigned issue_prefetch_p : 1; /* Should we really issue the prefetch? */ | |
294 | unsigned storent_p : 1; /* True if we changed the store to a | |
295 | nontemporal one. */ | |
b076a3fd ZD |
296 | }; |
297 | ||
a5497b12 | 298 | /* Dumps information about memory reference */ |
b076a3fd | 299 | static void |
a5497b12 VK |
300 | dump_mem_details (FILE *file, tree base, tree step, |
301 | HOST_WIDE_INT delta, bool write_p) | |
b076a3fd | 302 | { |
a5497b12 VK |
303 | fprintf (file, "(base "); |
304 | print_generic_expr (file, base, TDF_SLIM); | |
b076a3fd | 305 | fprintf (file, ", step "); |
a5497b12 VK |
306 | if (cst_and_fits_in_hwi (step)) |
307 | fprintf (file, HOST_WIDE_INT_PRINT_DEC, int_cst_value (step)); | |
81f32326 | 308 | else |
a5497b12 | 309 | print_generic_expr (file, step, TDF_TREE); |
b076a3fd | 310 | fprintf (file, ")\n"); |
e324a72f | 311 | fprintf (file, " delta "); |
a5497b12 VK |
312 | fprintf (file, HOST_WIDE_INT_PRINT_DEC, delta); |
313 | fprintf (file, "\n"); | |
314 | fprintf (file, " %s\n", write_p ? "write" : "read"); | |
b076a3fd | 315 | fprintf (file, "\n"); |
a5497b12 | 316 | } |
b076a3fd | 317 | |
a5497b12 | 318 | /* Dumps information about reference REF to FILE. */ |
b076a3fd | 319 | |
a5497b12 VK |
320 | static void |
321 | dump_mem_ref (FILE *file, struct mem_ref *ref) | |
322 | { | |
323 | fprintf (file, "Reference %p:\n", (void *) ref); | |
324 | ||
325 | fprintf (file, " group %p ", (void *) ref->group); | |
326 | ||
327 | dump_mem_details (file, ref->group->base, ref->group->step, ref->delta, | |
328 | ref->write_p); | |
b076a3fd ZD |
329 | } |
330 | ||
331 | /* Finds a group with BASE and STEP in GROUPS, or creates one if it does not | |
332 | exist. */ | |
333 | ||
334 | static struct mem_ref_group * | |
81f32326 | 335 | find_or_create_group (struct mem_ref_group **groups, tree base, tree step) |
b076a3fd ZD |
336 | { |
337 | struct mem_ref_group *group; | |
338 | ||
339 | for (; *groups; groups = &(*groups)->next) | |
340 | { | |
81f32326 | 341 | if (operand_equal_p ((*groups)->step, step, 0) |
b076a3fd ZD |
342 | && operand_equal_p ((*groups)->base, base, 0)) |
343 | return *groups; | |
344 | ||
81f32326 CB |
345 | /* If step is an integer constant, keep the list of groups sorted |
346 | by decreasing step. */ | |
21c0a521 DM |
347 | if (cst_and_fits_in_hwi ((*groups)->step) && cst_and_fits_in_hwi (step) |
348 | && int_cst_value ((*groups)->step) < int_cst_value (step)) | |
b076a3fd ZD |
349 | break; |
350 | } | |
351 | ||
5417e022 | 352 | group = XNEW (struct mem_ref_group); |
b076a3fd ZD |
353 | group->base = base; |
354 | group->step = step; | |
355 | group->refs = NULL; | |
356 | group->next = *groups; | |
357 | *groups = group; | |
358 | ||
359 | return group; | |
360 | } | |
361 | ||
362 | /* Records a memory reference MEM in GROUP with offset DELTA and write status | |
363 | WRITE_P. The reference occurs in statement STMT. */ | |
364 | ||
365 | static void | |
726a989a | 366 | record_ref (struct mem_ref_group *group, gimple stmt, tree mem, |
b076a3fd ZD |
367 | HOST_WIDE_INT delta, bool write_p) |
368 | { | |
369 | struct mem_ref **aref; | |
370 | ||
371 | /* Do not record the same address twice. */ | |
372 | for (aref = &group->refs; *aref; aref = &(*aref)->next) | |
373 | { | |
374 | /* It does not have to be possible for write reference to reuse the read | |
375 | prefetch, or vice versa. */ | |
376 | if (!WRITE_CAN_USE_READ_PREFETCH | |
377 | && write_p | |
378 | && !(*aref)->write_p) | |
379 | continue; | |
380 | if (!READ_CAN_USE_WRITE_PREFETCH | |
381 | && !write_p | |
382 | && (*aref)->write_p) | |
383 | continue; | |
384 | ||
385 | if ((*aref)->delta == delta) | |
386 | return; | |
387 | } | |
388 | ||
5417e022 | 389 | (*aref) = XNEW (struct mem_ref); |
b076a3fd ZD |
390 | (*aref)->stmt = stmt; |
391 | (*aref)->mem = mem; | |
392 | (*aref)->delta = delta; | |
393 | (*aref)->write_p = write_p; | |
394 | (*aref)->prefetch_before = PREFETCH_ALL; | |
395 | (*aref)->prefetch_mod = 1; | |
5417e022 | 396 | (*aref)->reuse_distance = 0; |
b076a3fd ZD |
397 | (*aref)->issue_prefetch_p = false; |
398 | (*aref)->group = group; | |
399 | (*aref)->next = NULL; | |
79f5e442 ZD |
400 | (*aref)->independent_p = false; |
401 | (*aref)->storent_p = false; | |
b076a3fd ZD |
402 | |
403 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
404 | dump_mem_ref (dump_file, *aref); | |
405 | } | |
406 | ||
407 | /* Release memory references in GROUPS. */ | |
408 | ||
409 | static void | |
410 | release_mem_refs (struct mem_ref_group *groups) | |
411 | { | |
412 | struct mem_ref_group *next_g; | |
413 | struct mem_ref *ref, *next_r; | |
414 | ||
415 | for (; groups; groups = next_g) | |
416 | { | |
417 | next_g = groups->next; | |
418 | for (ref = groups->refs; ref; ref = next_r) | |
419 | { | |
420 | next_r = ref->next; | |
421 | free (ref); | |
422 | } | |
423 | free (groups); | |
424 | } | |
425 | } | |
426 | ||
427 | /* A structure used to pass arguments to idx_analyze_ref. */ | |
428 | ||
429 | struct ar_data | |
430 | { | |
431 | struct loop *loop; /* Loop of the reference. */ | |
726a989a | 432 | gimple stmt; /* Statement of the reference. */ |
81f32326 | 433 | tree *step; /* Step of the memory reference. */ |
b076a3fd ZD |
434 | HOST_WIDE_INT *delta; /* Offset of the memory reference. */ |
435 | }; | |
436 | ||
437 | /* Analyzes a single INDEX of a memory reference to obtain information | |
438 | described at analyze_ref. Callback for for_each_index. */ | |
439 | ||
440 | static bool | |
441 | idx_analyze_ref (tree base, tree *index, void *data) | |
442 | { | |
c22940cd | 443 | struct ar_data *ar_data = (struct ar_data *) data; |
b076a3fd | 444 | tree ibase, step, stepsize; |
81f32326 | 445 | HOST_WIDE_INT idelta = 0, imult = 1; |
b076a3fd ZD |
446 | affine_iv iv; |
447 | ||
f017bf5e | 448 | if (!simple_iv (ar_data->loop, loop_containing_stmt (ar_data->stmt), |
81f32326 | 449 | *index, &iv, true)) |
b076a3fd ZD |
450 | return false; |
451 | ibase = iv.base; | |
452 | step = iv.step; | |
453 | ||
5be014d5 | 454 | if (TREE_CODE (ibase) == POINTER_PLUS_EXPR |
b076a3fd ZD |
455 | && cst_and_fits_in_hwi (TREE_OPERAND (ibase, 1))) |
456 | { | |
457 | idelta = int_cst_value (TREE_OPERAND (ibase, 1)); | |
458 | ibase = TREE_OPERAND (ibase, 0); | |
459 | } | |
460 | if (cst_and_fits_in_hwi (ibase)) | |
461 | { | |
462 | idelta += int_cst_value (ibase); | |
ff5e9a94 | 463 | ibase = build_int_cst (TREE_TYPE (ibase), 0); |
b076a3fd ZD |
464 | } |
465 | ||
466 | if (TREE_CODE (base) == ARRAY_REF) | |
467 | { | |
468 | stepsize = array_ref_element_size (base); | |
469 | if (!cst_and_fits_in_hwi (stepsize)) | |
470 | return false; | |
471 | imult = int_cst_value (stepsize); | |
8fde8b40 CB |
472 | step = fold_build2 (MULT_EXPR, sizetype, |
473 | fold_convert (sizetype, step), | |
474 | fold_convert (sizetype, stepsize)); | |
b076a3fd ZD |
475 | idelta *= imult; |
476 | } | |
477 | ||
8fde8b40 CB |
478 | if (*ar_data->step == NULL_TREE) |
479 | *ar_data->step = step; | |
480 | else | |
481 | *ar_data->step = fold_build2 (PLUS_EXPR, sizetype, | |
482 | fold_convert (sizetype, *ar_data->step), | |
483 | fold_convert (sizetype, step)); | |
b076a3fd ZD |
484 | *ar_data->delta += idelta; |
485 | *index = ibase; | |
486 | ||
487 | return true; | |
488 | } | |
489 | ||
aac8b8ed | 490 | /* Tries to express REF_P in shape &BASE + STEP * iter + DELTA, where DELTA and |
b076a3fd | 491 | STEP are integer constants and iter is number of iterations of LOOP. The |
aac8b8ed RS |
492 | reference occurs in statement STMT. Strips nonaddressable component |
493 | references from REF_P. */ | |
b076a3fd ZD |
494 | |
495 | static bool | |
aac8b8ed | 496 | analyze_ref (struct loop *loop, tree *ref_p, tree *base, |
81f32326 | 497 | tree *step, HOST_WIDE_INT *delta, |
726a989a | 498 | gimple stmt) |
b076a3fd ZD |
499 | { |
500 | struct ar_data ar_data; | |
501 | tree off; | |
502 | HOST_WIDE_INT bit_offset; | |
aac8b8ed | 503 | tree ref = *ref_p; |
b076a3fd | 504 | |
81f32326 | 505 | *step = NULL_TREE; |
b076a3fd ZD |
506 | *delta = 0; |
507 | ||
7c6dafac CF |
508 | /* First strip off the component references. Ignore bitfields. |
509 | Also strip off the real and imagine parts of a complex, so that | |
510 | they can have the same base. */ | |
511 | if (TREE_CODE (ref) == REALPART_EXPR | |
512 | || TREE_CODE (ref) == IMAGPART_EXPR | |
513 | || (TREE_CODE (ref) == COMPONENT_REF | |
514 | && DECL_NONADDRESSABLE_P (TREE_OPERAND (ref, 1)))) | |
515 | { | |
516 | if (TREE_CODE (ref) == IMAGPART_EXPR) | |
517 | *delta += int_size_in_bytes (TREE_TYPE (ref)); | |
518 | ref = TREE_OPERAND (ref, 0); | |
519 | } | |
b076a3fd | 520 | |
aac8b8ed RS |
521 | *ref_p = ref; |
522 | ||
b076a3fd ZD |
523 | for (; TREE_CODE (ref) == COMPONENT_REF; ref = TREE_OPERAND (ref, 0)) |
524 | { | |
525 | off = DECL_FIELD_BIT_OFFSET (TREE_OPERAND (ref, 1)); | |
526 | bit_offset = TREE_INT_CST_LOW (off); | |
527 | gcc_assert (bit_offset % BITS_PER_UNIT == 0); | |
b8698a0f | 528 | |
b076a3fd ZD |
529 | *delta += bit_offset / BITS_PER_UNIT; |
530 | } | |
531 | ||
532 | *base = unshare_expr (ref); | |
533 | ar_data.loop = loop; | |
534 | ar_data.stmt = stmt; | |
535 | ar_data.step = step; | |
536 | ar_data.delta = delta; | |
537 | return for_each_index (base, idx_analyze_ref, &ar_data); | |
538 | } | |
539 | ||
540 | /* Record a memory reference REF to the list REFS. The reference occurs in | |
79f5e442 ZD |
541 | LOOP in statement STMT and it is write if WRITE_P. Returns true if the |
542 | reference was recorded, false otherwise. */ | |
b076a3fd | 543 | |
79f5e442 | 544 | static bool |
b076a3fd | 545 | gather_memory_references_ref (struct loop *loop, struct mem_ref_group **refs, |
726a989a | 546 | tree ref, bool write_p, gimple stmt) |
b076a3fd | 547 | { |
81f32326 CB |
548 | tree base, step; |
549 | HOST_WIDE_INT delta; | |
b076a3fd ZD |
550 | struct mem_ref_group *agrp; |
551 | ||
a80a2701 JJ |
552 | if (get_base_address (ref) == NULL) |
553 | return false; | |
554 | ||
aac8b8ed | 555 | if (!analyze_ref (loop, &ref, &base, &step, &delta, stmt)) |
79f5e442 | 556 | return false; |
81f32326 CB |
557 | /* If analyze_ref fails the default is a NULL_TREE. We can stop here. */ |
558 | if (step == NULL_TREE) | |
559 | return false; | |
b076a3fd | 560 | |
756f50ce | 561 | /* Stop if the address of BASE could not be taken. */ |
bc068a23 CF |
562 | if (may_be_nonaddressable_p (base)) |
563 | return false; | |
564 | ||
a5497b12 VK |
565 | /* Limit non-constant step prefetching only to the innermost loops and |
566 | only when the step is loop invariant in the entire loop nest. */ | |
567 | if (!cst_and_fits_in_hwi (step)) | |
568 | { | |
569 | if (loop->inner != NULL) | |
570 | { | |
571 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
572 | { | |
573 | fprintf (dump_file, "Memory expression %p\n",(void *) ref ); | |
574 | print_generic_expr (dump_file, ref, TDF_TREE); | |
575 | fprintf (dump_file,":"); | |
c3284718 | 576 | dump_mem_details (dump_file, base, step, delta, write_p); |
a5497b12 VK |
577 | fprintf (dump_file, |
578 | "Ignoring %p, non-constant step prefetching is " | |
579 | "limited to inner most loops \n", | |
580 | (void *) ref); | |
581 | } | |
582 | return false; | |
583 | } | |
584 | else | |
585 | { | |
586 | if (!expr_invariant_in_loop_p (loop_outermost (loop), step)) | |
587 | { | |
588 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
589 | { | |
590 | fprintf (dump_file, "Memory expression %p\n",(void *) ref ); | |
591 | print_generic_expr (dump_file, ref, TDF_TREE); | |
592 | fprintf (dump_file,":"); | |
c3284718 | 593 | dump_mem_details (dump_file, base, step, delta, write_p); |
a5497b12 VK |
594 | fprintf (dump_file, |
595 | "Not prefetching, ignoring %p due to " | |
596 | "loop variant step\n", | |
597 | (void *) ref); | |
598 | } | |
599 | return false; | |
600 | } | |
601 | } | |
602 | } | |
50814135 | 603 | |
b076a3fd ZD |
604 | /* Now we know that REF = &BASE + STEP * iter + DELTA, where DELTA and STEP |
605 | are integer constants. */ | |
606 | agrp = find_or_create_group (refs, base, step); | |
607 | record_ref (agrp, stmt, ref, delta, write_p); | |
79f5e442 ZD |
608 | |
609 | return true; | |
b076a3fd ZD |
610 | } |
611 | ||
79f5e442 ZD |
612 | /* Record the suitable memory references in LOOP. NO_OTHER_REFS is set to |
613 | true if there are no other memory references inside the loop. */ | |
b076a3fd ZD |
614 | |
615 | static struct mem_ref_group * | |
db34470d | 616 | gather_memory_references (struct loop *loop, bool *no_other_refs, unsigned *ref_count) |
b076a3fd ZD |
617 | { |
618 | basic_block *body = get_loop_body_in_dom_order (loop); | |
619 | basic_block bb; | |
620 | unsigned i; | |
726a989a RB |
621 | gimple_stmt_iterator bsi; |
622 | gimple stmt; | |
623 | tree lhs, rhs; | |
b076a3fd ZD |
624 | struct mem_ref_group *refs = NULL; |
625 | ||
79f5e442 | 626 | *no_other_refs = true; |
db34470d | 627 | *ref_count = 0; |
79f5e442 | 628 | |
b076a3fd ZD |
629 | /* Scan the loop body in order, so that the former references precede the |
630 | later ones. */ | |
631 | for (i = 0; i < loop->num_nodes; i++) | |
632 | { | |
633 | bb = body[i]; | |
634 | if (bb->loop_father != loop) | |
635 | continue; | |
636 | ||
726a989a | 637 | for (bsi = gsi_start_bb (bb); !gsi_end_p (bsi); gsi_next (&bsi)) |
b076a3fd | 638 | { |
726a989a | 639 | stmt = gsi_stmt (bsi); |
79f5e442 | 640 | |
726a989a | 641 | if (gimple_code (stmt) != GIMPLE_ASSIGN) |
79f5e442 | 642 | { |
5006671f | 643 | if (gimple_vuse (stmt) |
726a989a RB |
644 | || (is_gimple_call (stmt) |
645 | && !(gimple_call_flags (stmt) & ECF_CONST))) | |
79f5e442 ZD |
646 | *no_other_refs = false; |
647 | continue; | |
648 | } | |
b076a3fd | 649 | |
726a989a RB |
650 | lhs = gimple_assign_lhs (stmt); |
651 | rhs = gimple_assign_rhs1 (stmt); | |
b076a3fd ZD |
652 | |
653 | if (REFERENCE_CLASS_P (rhs)) | |
db34470d | 654 | { |
79f5e442 ZD |
655 | *no_other_refs &= gather_memory_references_ref (loop, &refs, |
656 | rhs, false, stmt); | |
db34470d GS |
657 | *ref_count += 1; |
658 | } | |
b076a3fd | 659 | if (REFERENCE_CLASS_P (lhs)) |
db34470d | 660 | { |
79f5e442 ZD |
661 | *no_other_refs &= gather_memory_references_ref (loop, &refs, |
662 | lhs, true, stmt); | |
db34470d GS |
663 | *ref_count += 1; |
664 | } | |
b076a3fd ZD |
665 | } |
666 | } | |
667 | free (body); | |
668 | ||
669 | return refs; | |
670 | } | |
671 | ||
672 | /* Prune the prefetch candidate REF using the self-reuse. */ | |
673 | ||
674 | static void | |
675 | prune_ref_by_self_reuse (struct mem_ref *ref) | |
676 | { | |
81f32326 CB |
677 | HOST_WIDE_INT step; |
678 | bool backward; | |
679 | ||
680 | /* If the step size is non constant, we cannot calculate prefetch_mod. */ | |
681 | if (!cst_and_fits_in_hwi (ref->group->step)) | |
682 | return; | |
683 | ||
684 | step = int_cst_value (ref->group->step); | |
685 | ||
686 | backward = step < 0; | |
b076a3fd ZD |
687 | |
688 | if (step == 0) | |
689 | { | |
690 | /* Prefetch references to invariant address just once. */ | |
691 | ref->prefetch_before = 1; | |
692 | return; | |
693 | } | |
694 | ||
695 | if (backward) | |
696 | step = -step; | |
697 | ||
698 | if (step > PREFETCH_BLOCK) | |
699 | return; | |
700 | ||
701 | if ((backward && HAVE_BACKWARD_PREFETCH) | |
702 | || (!backward && HAVE_FORWARD_PREFETCH)) | |
703 | { | |
704 | ref->prefetch_before = 1; | |
705 | return; | |
706 | } | |
707 | ||
708 | ref->prefetch_mod = PREFETCH_BLOCK / step; | |
709 | } | |
710 | ||
711 | /* Divides X by BY, rounding down. */ | |
712 | ||
713 | static HOST_WIDE_INT | |
714 | ddown (HOST_WIDE_INT x, unsigned HOST_WIDE_INT by) | |
715 | { | |
716 | gcc_assert (by > 0); | |
717 | ||
718 | if (x >= 0) | |
719 | return x / by; | |
720 | else | |
721 | return (x + by - 1) / by; | |
722 | } | |
723 | ||
b8698a0f L |
724 | /* Given a CACHE_LINE_SIZE and two inductive memory references |
725 | with a common STEP greater than CACHE_LINE_SIZE and an address | |
726 | difference DELTA, compute the probability that they will fall | |
14e444c3 CF |
727 | in different cache lines. Return true if the computed miss rate |
728 | is not greater than the ACCEPTABLE_MISS_RATE. DISTINCT_ITERS is the | |
729 | number of distinct iterations after which the pattern repeats itself. | |
2c6dd136 GS |
730 | ALIGN_UNIT is the unit of alignment in bytes. */ |
731 | ||
14e444c3 CF |
732 | static bool |
733 | is_miss_rate_acceptable (unsigned HOST_WIDE_INT cache_line_size, | |
2c6dd136 GS |
734 | HOST_WIDE_INT step, HOST_WIDE_INT delta, |
735 | unsigned HOST_WIDE_INT distinct_iters, | |
736 | int align_unit) | |
737 | { | |
738 | unsigned align, iter; | |
14e444c3 | 739 | int total_positions, miss_positions, max_allowed_miss_positions; |
2c6dd136 GS |
740 | int address1, address2, cache_line1, cache_line2; |
741 | ||
a245c04b CF |
742 | /* It always misses if delta is greater than or equal to the cache |
743 | line size. */ | |
14e444c3 CF |
744 | if (delta >= (HOST_WIDE_INT) cache_line_size) |
745 | return false; | |
a245c04b | 746 | |
2c6dd136 | 747 | miss_positions = 0; |
14e444c3 CF |
748 | total_positions = (cache_line_size / align_unit) * distinct_iters; |
749 | max_allowed_miss_positions = (ACCEPTABLE_MISS_RATE * total_positions) / 1000; | |
b8698a0f | 750 | |
2c6dd136 GS |
751 | /* Iterate through all possible alignments of the first |
752 | memory reference within its cache line. */ | |
753 | for (align = 0; align < cache_line_size; align += align_unit) | |
754 | ||
755 | /* Iterate through all distinct iterations. */ | |
756 | for (iter = 0; iter < distinct_iters; iter++) | |
757 | { | |
758 | address1 = align + step * iter; | |
759 | address2 = address1 + delta; | |
760 | cache_line1 = address1 / cache_line_size; | |
761 | cache_line2 = address2 / cache_line_size; | |
2c6dd136 | 762 | if (cache_line1 != cache_line2) |
14e444c3 CF |
763 | { |
764 | miss_positions += 1; | |
765 | if (miss_positions > max_allowed_miss_positions) | |
766 | return false; | |
767 | } | |
2c6dd136 | 768 | } |
14e444c3 | 769 | return true; |
2c6dd136 GS |
770 | } |
771 | ||
b076a3fd ZD |
772 | /* Prune the prefetch candidate REF using the reuse with BY. |
773 | If BY_IS_BEFORE is true, BY is before REF in the loop. */ | |
774 | ||
775 | static void | |
776 | prune_ref_by_group_reuse (struct mem_ref *ref, struct mem_ref *by, | |
777 | bool by_is_before) | |
778 | { | |
81f32326 CB |
779 | HOST_WIDE_INT step; |
780 | bool backward; | |
b076a3fd ZD |
781 | HOST_WIDE_INT delta_r = ref->delta, delta_b = by->delta; |
782 | HOST_WIDE_INT delta = delta_b - delta_r; | |
783 | HOST_WIDE_INT hit_from; | |
784 | unsigned HOST_WIDE_INT prefetch_before, prefetch_block; | |
2c6dd136 GS |
785 | HOST_WIDE_INT reduced_step; |
786 | unsigned HOST_WIDE_INT reduced_prefetch_block; | |
787 | tree ref_type; | |
788 | int align_unit; | |
b076a3fd | 789 | |
81f32326 CB |
790 | /* If the step is non constant we cannot calculate prefetch_before. */ |
791 | if (!cst_and_fits_in_hwi (ref->group->step)) { | |
792 | return; | |
793 | } | |
794 | ||
795 | step = int_cst_value (ref->group->step); | |
796 | ||
797 | backward = step < 0; | |
798 | ||
799 | ||
b076a3fd ZD |
800 | if (delta == 0) |
801 | { | |
802 | /* If the references has the same address, only prefetch the | |
803 | former. */ | |
804 | if (by_is_before) | |
805 | ref->prefetch_before = 0; | |
b8698a0f | 806 | |
b076a3fd ZD |
807 | return; |
808 | } | |
809 | ||
810 | if (!step) | |
811 | { | |
812 | /* If the reference addresses are invariant and fall into the | |
813 | same cache line, prefetch just the first one. */ | |
814 | if (!by_is_before) | |
815 | return; | |
816 | ||
817 | if (ddown (ref->delta, PREFETCH_BLOCK) | |
818 | != ddown (by->delta, PREFETCH_BLOCK)) | |
819 | return; | |
820 | ||
821 | ref->prefetch_before = 0; | |
822 | return; | |
823 | } | |
824 | ||
825 | /* Only prune the reference that is behind in the array. */ | |
826 | if (backward) | |
827 | { | |
828 | if (delta > 0) | |
829 | return; | |
830 | ||
831 | /* Transform the data so that we may assume that the accesses | |
832 | are forward. */ | |
833 | delta = - delta; | |
834 | step = -step; | |
835 | delta_r = PREFETCH_BLOCK - 1 - delta_r; | |
836 | delta_b = PREFETCH_BLOCK - 1 - delta_b; | |
837 | } | |
838 | else | |
839 | { | |
840 | if (delta < 0) | |
841 | return; | |
842 | } | |
843 | ||
844 | /* Check whether the two references are likely to hit the same cache | |
845 | line, and how distant the iterations in that it occurs are from | |
846 | each other. */ | |
847 | ||
848 | if (step <= PREFETCH_BLOCK) | |
849 | { | |
850 | /* The accesses are sure to meet. Let us check when. */ | |
851 | hit_from = ddown (delta_b, PREFETCH_BLOCK) * PREFETCH_BLOCK; | |
852 | prefetch_before = (hit_from - delta_r + step - 1) / step; | |
853 | ||
57762e97 | 854 | /* Do not reduce prefetch_before if we meet beyond cache size. */ |
4c9cf7af | 855 | if (prefetch_before > absu_hwi (L2_CACHE_SIZE_BYTES / step)) |
57762e97 | 856 | prefetch_before = PREFETCH_ALL; |
b076a3fd ZD |
857 | if (prefetch_before < ref->prefetch_before) |
858 | ref->prefetch_before = prefetch_before; | |
859 | ||
860 | return; | |
861 | } | |
862 | ||
b8698a0f | 863 | /* A more complicated case with step > prefetch_block. First reduce |
2c6dd136 | 864 | the ratio between the step and the cache line size to its simplest |
b8698a0f L |
865 | terms. The resulting denominator will then represent the number of |
866 | distinct iterations after which each address will go back to its | |
867 | initial location within the cache line. This computation assumes | |
2c6dd136 | 868 | that PREFETCH_BLOCK is a power of two. */ |
b076a3fd | 869 | prefetch_block = PREFETCH_BLOCK; |
2c6dd136 GS |
870 | reduced_prefetch_block = prefetch_block; |
871 | reduced_step = step; | |
872 | while ((reduced_step & 1) == 0 | |
873 | && reduced_prefetch_block > 1) | |
b076a3fd | 874 | { |
2c6dd136 GS |
875 | reduced_step >>= 1; |
876 | reduced_prefetch_block >>= 1; | |
b076a3fd ZD |
877 | } |
878 | ||
b076a3fd ZD |
879 | prefetch_before = delta / step; |
880 | delta %= step; | |
2c6dd136 GS |
881 | ref_type = TREE_TYPE (ref->mem); |
882 | align_unit = TYPE_ALIGN (ref_type) / 8; | |
14e444c3 CF |
883 | if (is_miss_rate_acceptable (prefetch_block, step, delta, |
884 | reduced_prefetch_block, align_unit)) | |
b076a3fd | 885 | { |
57762e97 CB |
886 | /* Do not reduce prefetch_before if we meet beyond cache size. */ |
887 | if (prefetch_before > L2_CACHE_SIZE_BYTES / PREFETCH_BLOCK) | |
888 | prefetch_before = PREFETCH_ALL; | |
b076a3fd ZD |
889 | if (prefetch_before < ref->prefetch_before) |
890 | ref->prefetch_before = prefetch_before; | |
891 | ||
892 | return; | |
893 | } | |
894 | ||
895 | /* Try also the following iteration. */ | |
896 | prefetch_before++; | |
897 | delta = step - delta; | |
14e444c3 CF |
898 | if (is_miss_rate_acceptable (prefetch_block, step, delta, |
899 | reduced_prefetch_block, align_unit)) | |
b076a3fd ZD |
900 | { |
901 | if (prefetch_before < ref->prefetch_before) | |
902 | ref->prefetch_before = prefetch_before; | |
903 | ||
904 | return; | |
905 | } | |
906 | ||
907 | /* The ref probably does not reuse by. */ | |
908 | return; | |
909 | } | |
910 | ||
911 | /* Prune the prefetch candidate REF using the reuses with other references | |
912 | in REFS. */ | |
913 | ||
914 | static void | |
915 | prune_ref_by_reuse (struct mem_ref *ref, struct mem_ref *refs) | |
916 | { | |
917 | struct mem_ref *prune_by; | |
918 | bool before = true; | |
919 | ||
920 | prune_ref_by_self_reuse (ref); | |
921 | ||
922 | for (prune_by = refs; prune_by; prune_by = prune_by->next) | |
923 | { | |
924 | if (prune_by == ref) | |
925 | { | |
926 | before = false; | |
927 | continue; | |
928 | } | |
929 | ||
930 | if (!WRITE_CAN_USE_READ_PREFETCH | |
931 | && ref->write_p | |
932 | && !prune_by->write_p) | |
933 | continue; | |
934 | if (!READ_CAN_USE_WRITE_PREFETCH | |
935 | && !ref->write_p | |
936 | && prune_by->write_p) | |
937 | continue; | |
938 | ||
939 | prune_ref_by_group_reuse (ref, prune_by, before); | |
940 | } | |
941 | } | |
942 | ||
943 | /* Prune the prefetch candidates in GROUP using the reuse analysis. */ | |
944 | ||
945 | static void | |
946 | prune_group_by_reuse (struct mem_ref_group *group) | |
947 | { | |
948 | struct mem_ref *ref_pruned; | |
949 | ||
950 | for (ref_pruned = group->refs; ref_pruned; ref_pruned = ref_pruned->next) | |
951 | { | |
952 | prune_ref_by_reuse (ref_pruned, group->refs); | |
953 | ||
954 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
955 | { | |
956 | fprintf (dump_file, "Reference %p:", (void *) ref_pruned); | |
957 | ||
958 | if (ref_pruned->prefetch_before == PREFETCH_ALL | |
959 | && ref_pruned->prefetch_mod == 1) | |
960 | fprintf (dump_file, " no restrictions"); | |
961 | else if (ref_pruned->prefetch_before == 0) | |
962 | fprintf (dump_file, " do not prefetch"); | |
963 | else if (ref_pruned->prefetch_before <= ref_pruned->prefetch_mod) | |
964 | fprintf (dump_file, " prefetch once"); | |
965 | else | |
966 | { | |
967 | if (ref_pruned->prefetch_before != PREFETCH_ALL) | |
968 | { | |
969 | fprintf (dump_file, " prefetch before "); | |
970 | fprintf (dump_file, HOST_WIDE_INT_PRINT_DEC, | |
971 | ref_pruned->prefetch_before); | |
972 | } | |
973 | if (ref_pruned->prefetch_mod != 1) | |
974 | { | |
975 | fprintf (dump_file, " prefetch mod "); | |
976 | fprintf (dump_file, HOST_WIDE_INT_PRINT_DEC, | |
977 | ref_pruned->prefetch_mod); | |
978 | } | |
979 | } | |
980 | fprintf (dump_file, "\n"); | |
981 | } | |
982 | } | |
983 | } | |
984 | ||
985 | /* Prune the list of prefetch candidates GROUPS using the reuse analysis. */ | |
986 | ||
987 | static void | |
988 | prune_by_reuse (struct mem_ref_group *groups) | |
989 | { | |
990 | for (; groups; groups = groups->next) | |
991 | prune_group_by_reuse (groups); | |
992 | } | |
993 | ||
994 | /* Returns true if we should issue prefetch for REF. */ | |
995 | ||
996 | static bool | |
997 | should_issue_prefetch_p (struct mem_ref *ref) | |
998 | { | |
999 | /* For now do not issue prefetches for only first few of the | |
1000 | iterations. */ | |
1001 | if (ref->prefetch_before != PREFETCH_ALL) | |
a8beb3a7 CB |
1002 | { |
1003 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
1004 | fprintf (dump_file, "Ignoring %p due to prefetch_before\n", | |
1005 | (void *) ref); | |
1006 | return false; | |
1007 | } | |
b076a3fd | 1008 | |
79f5e442 ZD |
1009 | /* Do not prefetch nontemporal stores. */ |
1010 | if (ref->storent_p) | |
a8beb3a7 CB |
1011 | { |
1012 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
1013 | fprintf (dump_file, "Ignoring nontemporal store %p\n", (void *) ref); | |
1014 | return false; | |
1015 | } | |
79f5e442 | 1016 | |
b076a3fd ZD |
1017 | return true; |
1018 | } | |
1019 | ||
1020 | /* Decide which of the prefetch candidates in GROUPS to prefetch. | |
1021 | AHEAD is the number of iterations to prefetch ahead (which corresponds | |
1022 | to the number of simultaneous instances of one prefetch running at a | |
1023 | time). UNROLL_FACTOR is the factor by that the loop is going to be | |
1024 | unrolled. Returns true if there is anything to prefetch. */ | |
1025 | ||
1026 | static bool | |
1027 | schedule_prefetches (struct mem_ref_group *groups, unsigned unroll_factor, | |
1028 | unsigned ahead) | |
1029 | { | |
911b3fdb ZD |
1030 | unsigned remaining_prefetch_slots, n_prefetches, prefetch_slots; |
1031 | unsigned slots_per_prefetch; | |
b076a3fd ZD |
1032 | struct mem_ref *ref; |
1033 | bool any = false; | |
1034 | ||
911b3fdb ZD |
1035 | /* At most SIMULTANEOUS_PREFETCHES should be running at the same time. */ |
1036 | remaining_prefetch_slots = SIMULTANEOUS_PREFETCHES; | |
b076a3fd | 1037 | |
911b3fdb ZD |
1038 | /* The prefetch will run for AHEAD iterations of the original loop, i.e., |
1039 | AHEAD / UNROLL_FACTOR iterations of the unrolled loop. In each iteration, | |
1040 | it will need a prefetch slot. */ | |
1041 | slots_per_prefetch = (ahead + unroll_factor / 2) / unroll_factor; | |
b076a3fd | 1042 | if (dump_file && (dump_flags & TDF_DETAILS)) |
911b3fdb ZD |
1043 | fprintf (dump_file, "Each prefetch instruction takes %u prefetch slots.\n", |
1044 | slots_per_prefetch); | |
b076a3fd ZD |
1045 | |
1046 | /* For now we just take memory references one by one and issue | |
1047 | prefetches for as many as possible. The groups are sorted | |
1048 | starting with the largest step, since the references with | |
c0220ea4 | 1049 | large step are more likely to cause many cache misses. */ |
b076a3fd ZD |
1050 | |
1051 | for (; groups; groups = groups->next) | |
1052 | for (ref = groups->refs; ref; ref = ref->next) | |
1053 | { | |
1054 | if (!should_issue_prefetch_p (ref)) | |
1055 | continue; | |
1056 | ||
8532678c CF |
1057 | /* The loop is far from being sufficiently unrolled for this |
1058 | prefetch. Do not generate prefetch to avoid many redudant | |
1059 | prefetches. */ | |
1060 | if (ref->prefetch_mod / unroll_factor > PREFETCH_MOD_TO_UNROLL_FACTOR_RATIO) | |
1061 | continue; | |
1062 | ||
911b3fdb ZD |
1063 | /* If we need to prefetch the reference each PREFETCH_MOD iterations, |
1064 | and we unroll the loop UNROLL_FACTOR times, we need to insert | |
1065 | ceil (UNROLL_FACTOR / PREFETCH_MOD) instructions in each | |
1066 | iteration. */ | |
b076a3fd ZD |
1067 | n_prefetches = ((unroll_factor + ref->prefetch_mod - 1) |
1068 | / ref->prefetch_mod); | |
911b3fdb ZD |
1069 | prefetch_slots = n_prefetches * slots_per_prefetch; |
1070 | ||
1071 | /* If more than half of the prefetches would be lost anyway, do not | |
1072 | issue the prefetch. */ | |
1073 | if (2 * remaining_prefetch_slots < prefetch_slots) | |
1074 | continue; | |
1075 | ||
1076 | ref->issue_prefetch_p = true; | |
b076a3fd | 1077 | |
911b3fdb ZD |
1078 | if (remaining_prefetch_slots <= prefetch_slots) |
1079 | return true; | |
1080 | remaining_prefetch_slots -= prefetch_slots; | |
b076a3fd ZD |
1081 | any = true; |
1082 | } | |
1083 | ||
1084 | return any; | |
1085 | } | |
1086 | ||
d5058523 CF |
1087 | /* Return TRUE if no prefetch is going to be generated in the given |
1088 | GROUPS. */ | |
1089 | ||
1090 | static bool | |
1091 | nothing_to_prefetch_p (struct mem_ref_group *groups) | |
1092 | { | |
1093 | struct mem_ref *ref; | |
1094 | ||
1095 | for (; groups; groups = groups->next) | |
1096 | for (ref = groups->refs; ref; ref = ref->next) | |
1097 | if (should_issue_prefetch_p (ref)) | |
1098 | return false; | |
1099 | ||
1100 | return true; | |
1101 | } | |
1102 | ||
1103 | /* Estimate the number of prefetches in the given GROUPS. | |
1104 | UNROLL_FACTOR is the factor by which LOOP was unrolled. */ | |
b076a3fd | 1105 | |
db34470d | 1106 | static int |
d5058523 | 1107 | estimate_prefetch_count (struct mem_ref_group *groups, unsigned unroll_factor) |
b076a3fd ZD |
1108 | { |
1109 | struct mem_ref *ref; | |
d5058523 | 1110 | unsigned n_prefetches; |
db34470d | 1111 | int prefetch_count = 0; |
b076a3fd ZD |
1112 | |
1113 | for (; groups; groups = groups->next) | |
1114 | for (ref = groups->refs; ref; ref = ref->next) | |
1115 | if (should_issue_prefetch_p (ref)) | |
d5058523 CF |
1116 | { |
1117 | n_prefetches = ((unroll_factor + ref->prefetch_mod - 1) | |
1118 | / ref->prefetch_mod); | |
1119 | prefetch_count += n_prefetches; | |
1120 | } | |
b076a3fd | 1121 | |
db34470d | 1122 | return prefetch_count; |
b076a3fd ZD |
1123 | } |
1124 | ||
1125 | /* Issue prefetches for the reference REF into loop as decided before. | |
1126 | HEAD is the number of iterations to prefetch ahead. UNROLL_FACTOR | |
917f1b7e | 1127 | is the factor by which LOOP was unrolled. */ |
b076a3fd ZD |
1128 | |
1129 | static void | |
1130 | issue_prefetch_ref (struct mem_ref *ref, unsigned unroll_factor, unsigned ahead) | |
1131 | { | |
1132 | HOST_WIDE_INT delta; | |
81f32326 | 1133 | tree addr, addr_base, write_p, local, forward; |
538dd0b7 | 1134 | gcall *prefetch; |
726a989a | 1135 | gimple_stmt_iterator bsi; |
b076a3fd | 1136 | unsigned n_prefetches, ap; |
5417e022 | 1137 | bool nontemporal = ref->reuse_distance >= L2_CACHE_SIZE_BYTES; |
b076a3fd ZD |
1138 | |
1139 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
5417e022 ZD |
1140 | fprintf (dump_file, "Issued%s prefetch for %p.\n", |
1141 | nontemporal ? " nontemporal" : "", | |
1142 | (void *) ref); | |
b076a3fd | 1143 | |
726a989a | 1144 | bsi = gsi_for_stmt (ref->stmt); |
b076a3fd ZD |
1145 | |
1146 | n_prefetches = ((unroll_factor + ref->prefetch_mod - 1) | |
1147 | / ref->prefetch_mod); | |
1148 | addr_base = build_fold_addr_expr_with_type (ref->mem, ptr_type_node); | |
726a989a RB |
1149 | addr_base = force_gimple_operand_gsi (&bsi, unshare_expr (addr_base), |
1150 | true, NULL, true, GSI_SAME_STMT); | |
911b3fdb | 1151 | write_p = ref->write_p ? integer_one_node : integer_zero_node; |
9a9d280e | 1152 | local = nontemporal ? integer_zero_node : integer_three_node; |
b076a3fd ZD |
1153 | |
1154 | for (ap = 0; ap < n_prefetches; ap++) | |
1155 | { | |
81f32326 CB |
1156 | if (cst_and_fits_in_hwi (ref->group->step)) |
1157 | { | |
1158 | /* Determine the address to prefetch. */ | |
1159 | delta = (ahead + ap * ref->prefetch_mod) * | |
1160 | int_cst_value (ref->group->step); | |
5d49b6a7 | 1161 | addr = fold_build_pointer_plus_hwi (addr_base, delta); |
81f32326 CB |
1162 | addr = force_gimple_operand_gsi (&bsi, unshare_expr (addr), true, NULL, |
1163 | true, GSI_SAME_STMT); | |
1164 | } | |
1165 | else | |
1166 | { | |
1167 | /* The step size is non-constant but loop-invariant. We use the | |
1168 | heuristic to simply prefetch ahead iterations ahead. */ | |
1169 | forward = fold_build2 (MULT_EXPR, sizetype, | |
1170 | fold_convert (sizetype, ref->group->step), | |
1171 | fold_convert (sizetype, size_int (ahead))); | |
5d49b6a7 | 1172 | addr = fold_build_pointer_plus (addr_base, forward); |
81f32326 CB |
1173 | addr = force_gimple_operand_gsi (&bsi, unshare_expr (addr), true, |
1174 | NULL, true, GSI_SAME_STMT); | |
1175 | } | |
b076a3fd | 1176 | /* Create the prefetch instruction. */ |
e79983f4 | 1177 | prefetch = gimple_build_call (builtin_decl_explicit (BUILT_IN_PREFETCH), |
726a989a RB |
1178 | 3, addr, write_p, local); |
1179 | gsi_insert_before (&bsi, prefetch, GSI_SAME_STMT); | |
b076a3fd ZD |
1180 | } |
1181 | } | |
1182 | ||
1183 | /* Issue prefetches for the references in GROUPS into loop as decided before. | |
1184 | HEAD is the number of iterations to prefetch ahead. UNROLL_FACTOR is the | |
1185 | factor by that LOOP was unrolled. */ | |
1186 | ||
1187 | static void | |
1188 | issue_prefetches (struct mem_ref_group *groups, | |
1189 | unsigned unroll_factor, unsigned ahead) | |
1190 | { | |
1191 | struct mem_ref *ref; | |
1192 | ||
1193 | for (; groups; groups = groups->next) | |
1194 | for (ref = groups->refs; ref; ref = ref->next) | |
1195 | if (ref->issue_prefetch_p) | |
1196 | issue_prefetch_ref (ref, unroll_factor, ahead); | |
1197 | } | |
1198 | ||
79f5e442 ZD |
1199 | /* Returns true if REF is a memory write for that a nontemporal store insn |
1200 | can be used. */ | |
1201 | ||
1202 | static bool | |
1203 | nontemporal_store_p (struct mem_ref *ref) | |
1204 | { | |
ef4bddc2 | 1205 | machine_mode mode; |
79f5e442 ZD |
1206 | enum insn_code code; |
1207 | ||
1208 | /* REF must be a write that is not reused. We require it to be independent | |
1209 | on all other memory references in the loop, as the nontemporal stores may | |
1210 | be reordered with respect to other memory references. */ | |
1211 | if (!ref->write_p | |
1212 | || !ref->independent_p | |
1213 | || ref->reuse_distance < L2_CACHE_SIZE_BYTES) | |
1214 | return false; | |
1215 | ||
1216 | /* Check that we have the storent instruction for the mode. */ | |
1217 | mode = TYPE_MODE (TREE_TYPE (ref->mem)); | |
1218 | if (mode == BLKmode) | |
1219 | return false; | |
1220 | ||
947131ba | 1221 | code = optab_handler (storent_optab, mode); |
79f5e442 ZD |
1222 | return code != CODE_FOR_nothing; |
1223 | } | |
1224 | ||
1225 | /* If REF is a nontemporal store, we mark the corresponding modify statement | |
1226 | and return true. Otherwise, we return false. */ | |
1227 | ||
1228 | static bool | |
1229 | mark_nontemporal_store (struct mem_ref *ref) | |
1230 | { | |
1231 | if (!nontemporal_store_p (ref)) | |
1232 | return false; | |
1233 | ||
1234 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
1235 | fprintf (dump_file, "Marked reference %p as a nontemporal store.\n", | |
1236 | (void *) ref); | |
1237 | ||
726a989a | 1238 | gimple_assign_set_nontemporal_move (ref->stmt, true); |
79f5e442 ZD |
1239 | ref->storent_p = true; |
1240 | ||
1241 | return true; | |
1242 | } | |
1243 | ||
1244 | /* Issue a memory fence instruction after LOOP. */ | |
1245 | ||
1246 | static void | |
1247 | emit_mfence_after_loop (struct loop *loop) | |
1248 | { | |
9771b263 | 1249 | vec<edge> exits = get_loop_exit_edges (loop); |
79f5e442 | 1250 | edge exit; |
538dd0b7 | 1251 | gcall *call; |
726a989a | 1252 | gimple_stmt_iterator bsi; |
79f5e442 ZD |
1253 | unsigned i; |
1254 | ||
9771b263 | 1255 | FOR_EACH_VEC_ELT (exits, i, exit) |
79f5e442 | 1256 | { |
726a989a | 1257 | call = gimple_build_call (FENCE_FOLLOWING_MOVNT, 0); |
79f5e442 ZD |
1258 | |
1259 | if (!single_pred_p (exit->dest) | |
1260 | /* If possible, we prefer not to insert the fence on other paths | |
1261 | in cfg. */ | |
1262 | && !(exit->flags & EDGE_ABNORMAL)) | |
1263 | split_loop_exit_edge (exit); | |
726a989a | 1264 | bsi = gsi_after_labels (exit->dest); |
79f5e442 | 1265 | |
726a989a | 1266 | gsi_insert_before (&bsi, call, GSI_NEW_STMT); |
79f5e442 ZD |
1267 | } |
1268 | ||
9771b263 | 1269 | exits.release (); |
79f5e442 ZD |
1270 | update_ssa (TODO_update_ssa_only_virtuals); |
1271 | } | |
1272 | ||
1273 | /* Returns true if we can use storent in loop, false otherwise. */ | |
1274 | ||
1275 | static bool | |
1276 | may_use_storent_in_loop_p (struct loop *loop) | |
1277 | { | |
1278 | bool ret = true; | |
1279 | ||
1280 | if (loop->inner != NULL) | |
1281 | return false; | |
1282 | ||
1283 | /* If we must issue a mfence insn after using storent, check that there | |
1284 | is a suitable place for it at each of the loop exits. */ | |
1285 | if (FENCE_FOLLOWING_MOVNT != NULL_TREE) | |
1286 | { | |
9771b263 | 1287 | vec<edge> exits = get_loop_exit_edges (loop); |
79f5e442 ZD |
1288 | unsigned i; |
1289 | edge exit; | |
1290 | ||
9771b263 | 1291 | FOR_EACH_VEC_ELT (exits, i, exit) |
79f5e442 | 1292 | if ((exit->flags & EDGE_ABNORMAL) |
fefa31b5 | 1293 | && exit->dest == EXIT_BLOCK_PTR_FOR_FN (cfun)) |
79f5e442 ZD |
1294 | ret = false; |
1295 | ||
9771b263 | 1296 | exits.release (); |
79f5e442 ZD |
1297 | } |
1298 | ||
1299 | return ret; | |
1300 | } | |
1301 | ||
1302 | /* Marks nontemporal stores in LOOP. GROUPS contains the description of memory | |
1303 | references in the loop. */ | |
1304 | ||
1305 | static void | |
1306 | mark_nontemporal_stores (struct loop *loop, struct mem_ref_group *groups) | |
1307 | { | |
1308 | struct mem_ref *ref; | |
1309 | bool any = false; | |
1310 | ||
1311 | if (!may_use_storent_in_loop_p (loop)) | |
1312 | return; | |
1313 | ||
1314 | for (; groups; groups = groups->next) | |
1315 | for (ref = groups->refs; ref; ref = ref->next) | |
1316 | any |= mark_nontemporal_store (ref); | |
1317 | ||
1318 | if (any && FENCE_FOLLOWING_MOVNT != NULL_TREE) | |
1319 | emit_mfence_after_loop (loop); | |
1320 | } | |
1321 | ||
b076a3fd ZD |
1322 | /* Determines whether we can profitably unroll LOOP FACTOR times, and if |
1323 | this is the case, fill in DESC by the description of number of | |
1324 | iterations. */ | |
1325 | ||
1326 | static bool | |
1327 | should_unroll_loop_p (struct loop *loop, struct tree_niter_desc *desc, | |
1328 | unsigned factor) | |
1329 | { | |
1330 | if (!can_unroll_loop_p (loop, factor, desc)) | |
1331 | return false; | |
1332 | ||
1333 | /* We only consider loops without control flow for unrolling. This is not | |
1334 | a hard restriction -- tree_unroll_loop works with arbitrary loops | |
1335 | as well; but the unrolling/prefetching is usually more profitable for | |
1336 | loops consisting of a single basic block, and we want to limit the | |
1337 | code growth. */ | |
1338 | if (loop->num_nodes > 2) | |
1339 | return false; | |
1340 | ||
1341 | return true; | |
1342 | } | |
1343 | ||
1344 | /* Determine the coefficient by that unroll LOOP, from the information | |
1345 | contained in the list of memory references REFS. Description of | |
2711355f ZD |
1346 | umber of iterations of LOOP is stored to DESC. NINSNS is the number of |
1347 | insns of the LOOP. EST_NITER is the estimated number of iterations of | |
1348 | the loop, or -1 if no estimate is available. */ | |
b076a3fd ZD |
1349 | |
1350 | static unsigned | |
1351 | determine_unroll_factor (struct loop *loop, struct mem_ref_group *refs, | |
2711355f ZD |
1352 | unsigned ninsns, struct tree_niter_desc *desc, |
1353 | HOST_WIDE_INT est_niter) | |
b076a3fd | 1354 | { |
911b3fdb ZD |
1355 | unsigned upper_bound; |
1356 | unsigned nfactor, factor, mod_constraint; | |
b076a3fd ZD |
1357 | struct mem_ref_group *agp; |
1358 | struct mem_ref *ref; | |
1359 | ||
911b3fdb ZD |
1360 | /* First check whether the loop is not too large to unroll. We ignore |
1361 | PARAM_MAX_UNROLL_TIMES, because for small loops, it prevented us | |
1362 | from unrolling them enough to make exactly one cache line covered by each | |
1363 | iteration. Also, the goal of PARAM_MAX_UNROLL_TIMES is to prevent | |
1364 | us from unrolling the loops too many times in cases where we only expect | |
1365 | gains from better scheduling and decreasing loop overhead, which is not | |
1366 | the case here. */ | |
1367 | upper_bound = PARAM_VALUE (PARAM_MAX_UNROLLED_INSNS) / ninsns; | |
2711355f ZD |
1368 | |
1369 | /* If we unrolled the loop more times than it iterates, the unrolled version | |
1370 | of the loop would be never entered. */ | |
1371 | if (est_niter >= 0 && est_niter < (HOST_WIDE_INT) upper_bound) | |
1372 | upper_bound = est_niter; | |
1373 | ||
911b3fdb | 1374 | if (upper_bound <= 1) |
b076a3fd ZD |
1375 | return 1; |
1376 | ||
911b3fdb ZD |
1377 | /* Choose the factor so that we may prefetch each cache just once, |
1378 | but bound the unrolling by UPPER_BOUND. */ | |
1379 | factor = 1; | |
b076a3fd ZD |
1380 | for (agp = refs; agp; agp = agp->next) |
1381 | for (ref = agp->refs; ref; ref = ref->next) | |
911b3fdb ZD |
1382 | if (should_issue_prefetch_p (ref)) |
1383 | { | |
1384 | mod_constraint = ref->prefetch_mod; | |
1385 | nfactor = least_common_multiple (mod_constraint, factor); | |
1386 | if (nfactor <= upper_bound) | |
1387 | factor = nfactor; | |
1388 | } | |
b076a3fd ZD |
1389 | |
1390 | if (!should_unroll_loop_p (loop, desc, factor)) | |
1391 | return 1; | |
1392 | ||
1393 | return factor; | |
1394 | } | |
1395 | ||
5417e022 ZD |
1396 | /* Returns the total volume of the memory references REFS, taking into account |
1397 | reuses in the innermost loop and cache line size. TODO -- we should also | |
1398 | take into account reuses across the iterations of the loops in the loop | |
1399 | nest. */ | |
1400 | ||
1401 | static unsigned | |
1402 | volume_of_references (struct mem_ref_group *refs) | |
1403 | { | |
1404 | unsigned volume = 0; | |
1405 | struct mem_ref_group *gr; | |
1406 | struct mem_ref *ref; | |
1407 | ||
1408 | for (gr = refs; gr; gr = gr->next) | |
1409 | for (ref = gr->refs; ref; ref = ref->next) | |
1410 | { | |
1411 | /* Almost always reuses another value? */ | |
1412 | if (ref->prefetch_before != PREFETCH_ALL) | |
1413 | continue; | |
1414 | ||
1415 | /* If several iterations access the same cache line, use the size of | |
1416 | the line divided by this number. Otherwise, a cache line is | |
1417 | accessed in each iteration. TODO -- in the latter case, we should | |
1418 | take the size of the reference into account, rounding it up on cache | |
1419 | line size multiple. */ | |
1420 | volume += L1_CACHE_LINE_SIZE / ref->prefetch_mod; | |
1421 | } | |
1422 | return volume; | |
1423 | } | |
1424 | ||
1425 | /* Returns the volume of memory references accessed across VEC iterations of | |
1426 | loops, whose sizes are described in the LOOP_SIZES array. N is the number | |
1427 | of the loops in the nest (length of VEC and LOOP_SIZES vectors). */ | |
1428 | ||
1429 | static unsigned | |
1430 | volume_of_dist_vector (lambda_vector vec, unsigned *loop_sizes, unsigned n) | |
1431 | { | |
1432 | unsigned i; | |
1433 | ||
1434 | for (i = 0; i < n; i++) | |
1435 | if (vec[i] != 0) | |
1436 | break; | |
1437 | ||
1438 | if (i == n) | |
1439 | return 0; | |
1440 | ||
1441 | gcc_assert (vec[i] > 0); | |
1442 | ||
1443 | /* We ignore the parts of the distance vector in subloops, since usually | |
1444 | the numbers of iterations are much smaller. */ | |
1445 | return loop_sizes[i] * vec[i]; | |
1446 | } | |
1447 | ||
1448 | /* Add the steps of ACCESS_FN multiplied by STRIDE to the array STRIDE | |
1449 | at the position corresponding to the loop of the step. N is the depth | |
1450 | of the considered loop nest, and, LOOP is its innermost loop. */ | |
1451 | ||
1452 | static void | |
1453 | add_subscript_strides (tree access_fn, unsigned stride, | |
1454 | HOST_WIDE_INT *strides, unsigned n, struct loop *loop) | |
1455 | { | |
1456 | struct loop *aloop; | |
1457 | tree step; | |
1458 | HOST_WIDE_INT astep; | |
1459 | unsigned min_depth = loop_depth (loop) - n; | |
1460 | ||
1461 | while (TREE_CODE (access_fn) == POLYNOMIAL_CHREC) | |
1462 | { | |
1463 | aloop = get_chrec_loop (access_fn); | |
1464 | step = CHREC_RIGHT (access_fn); | |
1465 | access_fn = CHREC_LEFT (access_fn); | |
1466 | ||
1467 | if ((unsigned) loop_depth (aloop) <= min_depth) | |
1468 | continue; | |
1469 | ||
9541ffee | 1470 | if (tree_fits_shwi_p (step)) |
9439e9a1 | 1471 | astep = tree_to_shwi (step); |
5417e022 ZD |
1472 | else |
1473 | astep = L1_CACHE_LINE_SIZE; | |
1474 | ||
1475 | strides[n - 1 - loop_depth (loop) + loop_depth (aloop)] += astep * stride; | |
1476 | ||
1477 | } | |
1478 | } | |
1479 | ||
1480 | /* Returns the volume of memory references accessed between two consecutive | |
1481 | self-reuses of the reference DR. We consider the subscripts of DR in N | |
1482 | loops, and LOOP_SIZES contains the volumes of accesses in each of the | |
1483 | loops. LOOP is the innermost loop of the current loop nest. */ | |
1484 | ||
1485 | static unsigned | |
1486 | self_reuse_distance (data_reference_p dr, unsigned *loop_sizes, unsigned n, | |
1487 | struct loop *loop) | |
1488 | { | |
1489 | tree stride, access_fn; | |
1490 | HOST_WIDE_INT *strides, astride; | |
9771b263 | 1491 | vec<tree> access_fns; |
5417e022 ZD |
1492 | tree ref = DR_REF (dr); |
1493 | unsigned i, ret = ~0u; | |
1494 | ||
1495 | /* In the following example: | |
1496 | ||
1497 | for (i = 0; i < N; i++) | |
1498 | for (j = 0; j < N; j++) | |
1499 | use (a[j][i]); | |
1500 | the same cache line is accessed each N steps (except if the change from | |
1501 | i to i + 1 crosses the boundary of the cache line). Thus, for self-reuse, | |
1502 | we cannot rely purely on the results of the data dependence analysis. | |
1503 | ||
1504 | Instead, we compute the stride of the reference in each loop, and consider | |
1505 | the innermost loop in that the stride is less than cache size. */ | |
1506 | ||
1507 | strides = XCNEWVEC (HOST_WIDE_INT, n); | |
1508 | access_fns = DR_ACCESS_FNS (dr); | |
1509 | ||
9771b263 | 1510 | FOR_EACH_VEC_ELT (access_fns, i, access_fn) |
5417e022 ZD |
1511 | { |
1512 | /* Keep track of the reference corresponding to the subscript, so that we | |
1513 | know its stride. */ | |
1514 | while (handled_component_p (ref) && TREE_CODE (ref) != ARRAY_REF) | |
1515 | ref = TREE_OPERAND (ref, 0); | |
b8698a0f | 1516 | |
5417e022 ZD |
1517 | if (TREE_CODE (ref) == ARRAY_REF) |
1518 | { | |
1519 | stride = TYPE_SIZE_UNIT (TREE_TYPE (ref)); | |
cc269bb6 | 1520 | if (tree_fits_uhwi_p (stride)) |
ae7e9ddd | 1521 | astride = tree_to_uhwi (stride); |
5417e022 ZD |
1522 | else |
1523 | astride = L1_CACHE_LINE_SIZE; | |
1524 | ||
1525 | ref = TREE_OPERAND (ref, 0); | |
1526 | } | |
1527 | else | |
1528 | astride = 1; | |
1529 | ||
1530 | add_subscript_strides (access_fn, astride, strides, n, loop); | |
1531 | } | |
1532 | ||
1533 | for (i = n; i-- > 0; ) | |
1534 | { | |
1535 | unsigned HOST_WIDE_INT s; | |
1536 | ||
1537 | s = strides[i] < 0 ? -strides[i] : strides[i]; | |
1538 | ||
1539 | if (s < (unsigned) L1_CACHE_LINE_SIZE | |
1540 | && (loop_sizes[i] | |
1541 | > (unsigned) (L1_CACHE_SIZE_BYTES / NONTEMPORAL_FRACTION))) | |
1542 | { | |
1543 | ret = loop_sizes[i]; | |
1544 | break; | |
1545 | } | |
1546 | } | |
1547 | ||
1548 | free (strides); | |
1549 | return ret; | |
1550 | } | |
1551 | ||
1552 | /* Determines the distance till the first reuse of each reference in REFS | |
79f5e442 | 1553 | in the loop nest of LOOP. NO_OTHER_REFS is true if there are no other |
1e373390 | 1554 | memory references in the loop. Return false if the analysis fails. */ |
5417e022 | 1555 | |
1e373390 | 1556 | static bool |
79f5e442 ZD |
1557 | determine_loop_nest_reuse (struct loop *loop, struct mem_ref_group *refs, |
1558 | bool no_other_refs) | |
5417e022 ZD |
1559 | { |
1560 | struct loop *nest, *aloop; | |
6e1aa848 DN |
1561 | vec<data_reference_p> datarefs = vNULL; |
1562 | vec<ddr_p> dependences = vNULL; | |
5417e022 | 1563 | struct mem_ref_group *gr; |
79f5e442 | 1564 | struct mem_ref *ref, *refb; |
6e1aa848 | 1565 | vec<loop_p> vloops = vNULL; |
5417e022 ZD |
1566 | unsigned *loop_data_size; |
1567 | unsigned i, j, n; | |
1568 | unsigned volume, dist, adist; | |
1569 | HOST_WIDE_INT vol; | |
1570 | data_reference_p dr; | |
1571 | ddr_p dep; | |
1572 | ||
1573 | if (loop->inner) | |
1e373390 | 1574 | return true; |
5417e022 ZD |
1575 | |
1576 | /* Find the outermost loop of the loop nest of loop (we require that | |
1577 | there are no sibling loops inside the nest). */ | |
1578 | nest = loop; | |
1579 | while (1) | |
1580 | { | |
1581 | aloop = loop_outer (nest); | |
1582 | ||
1583 | if (aloop == current_loops->tree_root | |
1584 | || aloop->inner->next) | |
1585 | break; | |
1586 | ||
1587 | nest = aloop; | |
1588 | } | |
1589 | ||
1590 | /* For each loop, determine the amount of data accessed in each iteration. | |
1591 | We use this to estimate whether the reference is evicted from the | |
1592 | cache before its reuse. */ | |
1593 | find_loop_nest (nest, &vloops); | |
9771b263 | 1594 | n = vloops.length (); |
5417e022 ZD |
1595 | loop_data_size = XNEWVEC (unsigned, n); |
1596 | volume = volume_of_references (refs); | |
1597 | i = n; | |
1598 | while (i-- != 0) | |
1599 | { | |
1600 | loop_data_size[i] = volume; | |
1601 | /* Bound the volume by the L2 cache size, since above this bound, | |
1602 | all dependence distances are equivalent. */ | |
1603 | if (volume > L2_CACHE_SIZE_BYTES) | |
1604 | continue; | |
1605 | ||
9771b263 | 1606 | aloop = vloops[i]; |
652c4c71 | 1607 | vol = estimated_stmt_executions_int (aloop); |
e5b332cd | 1608 | if (vol == -1) |
5417e022 ZD |
1609 | vol = expected_loop_iterations (aloop); |
1610 | volume *= vol; | |
1611 | } | |
1612 | ||
1613 | /* Prepare the references in the form suitable for data dependence | |
0d52bcc1 | 1614 | analysis. We ignore unanalyzable data references (the results |
5417e022 ZD |
1615 | are used just as a heuristics to estimate temporality of the |
1616 | references, hence we do not need to worry about correctness). */ | |
1617 | for (gr = refs; gr; gr = gr->next) | |
1618 | for (ref = gr->refs; ref; ref = ref->next) | |
1619 | { | |
5c640e29 SP |
1620 | dr = create_data_ref (nest, loop_containing_stmt (ref->stmt), |
1621 | ref->mem, ref->stmt, !ref->write_p); | |
5417e022 ZD |
1622 | |
1623 | if (dr) | |
1624 | { | |
1625 | ref->reuse_distance = volume; | |
1626 | dr->aux = ref; | |
9771b263 | 1627 | datarefs.safe_push (dr); |
5417e022 | 1628 | } |
79f5e442 ZD |
1629 | else |
1630 | no_other_refs = false; | |
5417e022 ZD |
1631 | } |
1632 | ||
9771b263 | 1633 | FOR_EACH_VEC_ELT (datarefs, i, dr) |
5417e022 ZD |
1634 | { |
1635 | dist = self_reuse_distance (dr, loop_data_size, n, loop); | |
3d9a9f94 | 1636 | ref = (struct mem_ref *) dr->aux; |
5417e022 ZD |
1637 | if (ref->reuse_distance > dist) |
1638 | ref->reuse_distance = dist; | |
79f5e442 ZD |
1639 | |
1640 | if (no_other_refs) | |
1641 | ref->independent_p = true; | |
5417e022 ZD |
1642 | } |
1643 | ||
1e373390 RG |
1644 | if (!compute_all_dependences (datarefs, &dependences, vloops, true)) |
1645 | return false; | |
5417e022 | 1646 | |
9771b263 | 1647 | FOR_EACH_VEC_ELT (dependences, i, dep) |
5417e022 ZD |
1648 | { |
1649 | if (DDR_ARE_DEPENDENT (dep) == chrec_known) | |
1650 | continue; | |
1651 | ||
3d9a9f94 KG |
1652 | ref = (struct mem_ref *) DDR_A (dep)->aux; |
1653 | refb = (struct mem_ref *) DDR_B (dep)->aux; | |
79f5e442 | 1654 | |
5417e022 ZD |
1655 | if (DDR_ARE_DEPENDENT (dep) == chrec_dont_know |
1656 | || DDR_NUM_DIST_VECTS (dep) == 0) | |
1657 | { | |
0d52bcc1 | 1658 | /* If the dependence cannot be analyzed, assume that there might be |
5417e022 ZD |
1659 | a reuse. */ |
1660 | dist = 0; | |
b8698a0f | 1661 | |
79f5e442 ZD |
1662 | ref->independent_p = false; |
1663 | refb->independent_p = false; | |
5417e022 ZD |
1664 | } |
1665 | else | |
1666 | { | |
0d52bcc1 | 1667 | /* The distance vectors are normalized to be always lexicographically |
5417e022 ZD |
1668 | positive, hence we cannot tell just from them whether DDR_A comes |
1669 | before DDR_B or vice versa. However, it is not important, | |
1670 | anyway -- if DDR_A is close to DDR_B, then it is either reused in | |
1671 | DDR_B (and it is not nontemporal), or it reuses the value of DDR_B | |
1672 | in cache (and marking it as nontemporal would not affect | |
1673 | anything). */ | |
1674 | ||
1675 | dist = volume; | |
1676 | for (j = 0; j < DDR_NUM_DIST_VECTS (dep); j++) | |
1677 | { | |
1678 | adist = volume_of_dist_vector (DDR_DIST_VECT (dep, j), | |
1679 | loop_data_size, n); | |
1680 | ||
79f5e442 ZD |
1681 | /* If this is a dependence in the innermost loop (i.e., the |
1682 | distances in all superloops are zero) and it is not | |
1683 | the trivial self-dependence with distance zero, record that | |
1684 | the references are not completely independent. */ | |
1685 | if (lambda_vector_zerop (DDR_DIST_VECT (dep, j), n - 1) | |
1686 | && (ref != refb | |
1687 | || DDR_DIST_VECT (dep, j)[n-1] != 0)) | |
1688 | { | |
1689 | ref->independent_p = false; | |
1690 | refb->independent_p = false; | |
1691 | } | |
1692 | ||
5417e022 ZD |
1693 | /* Ignore accesses closer than |
1694 | L1_CACHE_SIZE_BYTES / NONTEMPORAL_FRACTION, | |
1695 | so that we use nontemporal prefetches e.g. if single memory | |
1696 | location is accessed several times in a single iteration of | |
1697 | the loop. */ | |
1698 | if (adist < L1_CACHE_SIZE_BYTES / NONTEMPORAL_FRACTION) | |
1699 | continue; | |
1700 | ||
1701 | if (adist < dist) | |
1702 | dist = adist; | |
1703 | } | |
1704 | } | |
1705 | ||
5417e022 ZD |
1706 | if (ref->reuse_distance > dist) |
1707 | ref->reuse_distance = dist; | |
79f5e442 ZD |
1708 | if (refb->reuse_distance > dist) |
1709 | refb->reuse_distance = dist; | |
5417e022 ZD |
1710 | } |
1711 | ||
1712 | free_dependence_relations (dependences); | |
1713 | free_data_refs (datarefs); | |
1714 | free (loop_data_size); | |
1715 | ||
1716 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
1717 | { | |
1718 | fprintf (dump_file, "Reuse distances:\n"); | |
1719 | for (gr = refs; gr; gr = gr->next) | |
1720 | for (ref = gr->refs; ref; ref = ref->next) | |
1721 | fprintf (dump_file, " ref %p distance %u\n", | |
1722 | (void *) ref, ref->reuse_distance); | |
1723 | } | |
1e373390 RG |
1724 | |
1725 | return true; | |
5417e022 ZD |
1726 | } |
1727 | ||
0bbe50f6 CF |
1728 | /* Determine whether or not the trip count to ahead ratio is too small based |
1729 | on prefitablility consideration. | |
db34470d | 1730 | AHEAD: the iteration ahead distance, |
0bbe50f6 CF |
1731 | EST_NITER: the estimated trip count. */ |
1732 | ||
1733 | static bool | |
1734 | trip_count_to_ahead_ratio_too_small_p (unsigned ahead, HOST_WIDE_INT est_niter) | |
1735 | { | |
1736 | /* Assume trip count to ahead ratio is big enough if the trip count could not | |
1737 | be estimated at compile time. */ | |
1738 | if (est_niter < 0) | |
1739 | return false; | |
1740 | ||
1741 | if (est_niter < (HOST_WIDE_INT) (TRIP_COUNT_TO_AHEAD_RATIO * ahead)) | |
1742 | { | |
1743 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
1744 | fprintf (dump_file, | |
1745 | "Not prefetching -- loop estimated to roll only %d times\n", | |
1746 | (int) est_niter); | |
1747 | return true; | |
1748 | } | |
1749 | ||
1750 | return false; | |
1751 | } | |
1752 | ||
1753 | /* Determine whether or not the number of memory references in the loop is | |
1754 | reasonable based on the profitablity and compilation time considerations. | |
db34470d | 1755 | NINSNS: estimated number of instructions in the loop, |
db34470d GS |
1756 | MEM_REF_COUNT: total number of memory references in the loop. */ |
1757 | ||
b8698a0f | 1758 | static bool |
0bbe50f6 | 1759 | mem_ref_count_reasonable_p (unsigned ninsns, unsigned mem_ref_count) |
db34470d | 1760 | { |
0bbe50f6 | 1761 | int insn_to_mem_ratio; |
db34470d GS |
1762 | |
1763 | if (mem_ref_count == 0) | |
1764 | return false; | |
1765 | ||
0bbe50f6 CF |
1766 | /* Miss rate computation (is_miss_rate_acceptable) and dependence analysis |
1767 | (compute_all_dependences) have high costs based on quadratic complexity. | |
1768 | To avoid huge compilation time, we give up prefetching if mem_ref_count | |
1769 | is too large. */ | |
1770 | if (mem_ref_count > PREFETCH_MAX_MEM_REFS_PER_LOOP) | |
1771 | return false; | |
1772 | ||
b8698a0f L |
1773 | /* Prefetching improves performance by overlapping cache missing |
1774 | memory accesses with CPU operations. If the loop does not have | |
1775 | enough CPU operations to overlap with memory operations, prefetching | |
1776 | won't give a significant benefit. One approximate way of checking | |
1777 | this is to require the ratio of instructions to memory references to | |
db34470d GS |
1778 | be above a certain limit. This approximation works well in practice. |
1779 | TODO: Implement a more precise computation by estimating the time | |
1780 | for each CPU or memory op in the loop. Time estimates for memory ops | |
1781 | should account for cache misses. */ | |
b8698a0f | 1782 | insn_to_mem_ratio = ninsns / mem_ref_count; |
db34470d GS |
1783 | |
1784 | if (insn_to_mem_ratio < PREFETCH_MIN_INSN_TO_MEM_RATIO) | |
55e5a2eb CF |
1785 | { |
1786 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
1787 | fprintf (dump_file, | |
1788 | "Not prefetching -- instruction to memory reference ratio (%d) too small\n", | |
1789 | insn_to_mem_ratio); | |
1790 | return false; | |
1791 | } | |
db34470d | 1792 | |
0bbe50f6 CF |
1793 | return true; |
1794 | } | |
1795 | ||
1796 | /* Determine whether or not the instruction to prefetch ratio in the loop is | |
1797 | too small based on the profitablity consideration. | |
1798 | NINSNS: estimated number of instructions in the loop, | |
1799 | PREFETCH_COUNT: an estimate of the number of prefetches, | |
1800 | UNROLL_FACTOR: the factor to unroll the loop if prefetching. */ | |
1801 | ||
1802 | static bool | |
1803 | insn_to_prefetch_ratio_too_small_p (unsigned ninsns, unsigned prefetch_count, | |
1804 | unsigned unroll_factor) | |
1805 | { | |
1806 | int insn_to_prefetch_ratio; | |
1807 | ||
d3a9b459 CF |
1808 | /* Prefetching most likely causes performance degradation when the instruction |
1809 | to prefetch ratio is too small. Too many prefetch instructions in a loop | |
1810 | may reduce the I-cache performance. | |
ccacf0e1 CF |
1811 | (unroll_factor * ninsns) is used to estimate the number of instructions in |
1812 | the unrolled loop. This implementation is a bit simplistic -- the number | |
1813 | of issued prefetch instructions is also affected by unrolling. So, | |
1814 | prefetch_mod and the unroll factor should be taken into account when | |
1815 | determining prefetch_count. Also, the number of insns of the unrolled | |
1816 | loop will usually be significantly smaller than the number of insns of the | |
1817 | original loop * unroll_factor (at least the induction variable increases | |
1818 | and the exit branches will get eliminated), so it might be better to use | |
1819 | tree_estimate_loop_size + estimated_unrolled_size. */ | |
d3a9b459 CF |
1820 | insn_to_prefetch_ratio = (unroll_factor * ninsns) / prefetch_count; |
1821 | if (insn_to_prefetch_ratio < MIN_INSN_TO_PREFETCH_RATIO) | |
db34470d | 1822 | { |
d3a9b459 CF |
1823 | if (dump_file && (dump_flags & TDF_DETAILS)) |
1824 | fprintf (dump_file, | |
1825 | "Not prefetching -- instruction to prefetch ratio (%d) too small\n", | |
1826 | insn_to_prefetch_ratio); | |
0bbe50f6 | 1827 | return true; |
db34470d | 1828 | } |
b8698a0f | 1829 | |
0bbe50f6 | 1830 | return false; |
db34470d GS |
1831 | } |
1832 | ||
1833 | ||
b076a3fd | 1834 | /* Issue prefetch instructions for array references in LOOP. Returns |
d73be268 | 1835 | true if the LOOP was unrolled. */ |
b076a3fd ZD |
1836 | |
1837 | static bool | |
d73be268 | 1838 | loop_prefetch_arrays (struct loop *loop) |
b076a3fd ZD |
1839 | { |
1840 | struct mem_ref_group *refs; | |
2711355f ZD |
1841 | unsigned ahead, ninsns, time, unroll_factor; |
1842 | HOST_WIDE_INT est_niter; | |
b076a3fd | 1843 | struct tree_niter_desc desc; |
79f5e442 | 1844 | bool unrolled = false, no_other_refs; |
db34470d GS |
1845 | unsigned prefetch_count; |
1846 | unsigned mem_ref_count; | |
b076a3fd | 1847 | |
efd8f750 | 1848 | if (optimize_loop_nest_for_size_p (loop)) |
2732d767 ZD |
1849 | { |
1850 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
1851 | fprintf (dump_file, " ignored (cold area)\n"); | |
1852 | return false; | |
1853 | } | |
1854 | ||
0bbe50f6 CF |
1855 | /* FIXME: the time should be weighted by the probabilities of the blocks in |
1856 | the loop body. */ | |
1857 | time = tree_num_loop_insns (loop, &eni_time_weights); | |
1858 | if (time == 0) | |
1859 | return false; | |
1860 | ||
1861 | ahead = (PREFETCH_LATENCY + time - 1) / time; | |
652c4c71 | 1862 | est_niter = estimated_stmt_executions_int (loop); |
e5b332cd RG |
1863 | if (est_niter == -1) |
1864 | est_niter = max_stmt_executions_int (loop); | |
0bbe50f6 CF |
1865 | |
1866 | /* Prefetching is not likely to be profitable if the trip count to ahead | |
1867 | ratio is too small. */ | |
1868 | if (trip_count_to_ahead_ratio_too_small_p (ahead, est_niter)) | |
1869 | return false; | |
1870 | ||
1871 | ninsns = tree_num_loop_insns (loop, &eni_size_weights); | |
1872 | ||
b076a3fd | 1873 | /* Step 1: gather the memory references. */ |
db34470d | 1874 | refs = gather_memory_references (loop, &no_other_refs, &mem_ref_count); |
b076a3fd | 1875 | |
0bbe50f6 CF |
1876 | /* Give up prefetching if the number of memory references in the |
1877 | loop is not reasonable based on profitablity and compilation time | |
1878 | considerations. */ | |
1879 | if (!mem_ref_count_reasonable_p (ninsns, mem_ref_count)) | |
1880 | goto fail; | |
1881 | ||
b076a3fd ZD |
1882 | /* Step 2: estimate the reuse effects. */ |
1883 | prune_by_reuse (refs); | |
1884 | ||
d5058523 | 1885 | if (nothing_to_prefetch_p (refs)) |
b076a3fd ZD |
1886 | goto fail; |
1887 | ||
1e373390 RG |
1888 | if (!determine_loop_nest_reuse (loop, refs, no_other_refs)) |
1889 | goto fail; | |
5417e022 | 1890 | |
0bbe50f6 | 1891 | /* Step 3: determine unroll factor. */ |
2711355f ZD |
1892 | unroll_factor = determine_unroll_factor (loop, refs, ninsns, &desc, |
1893 | est_niter); | |
d5058523 CF |
1894 | |
1895 | /* Estimate prefetch count for the unrolled loop. */ | |
1896 | prefetch_count = estimate_prefetch_count (refs, unroll_factor); | |
1897 | if (prefetch_count == 0) | |
1898 | goto fail; | |
1899 | ||
2711355f | 1900 | if (dump_file && (dump_flags & TDF_DETAILS)) |
b8698a0f | 1901 | fprintf (dump_file, "Ahead %d, unroll factor %d, trip count " |
d81f5387 | 1902 | HOST_WIDE_INT_PRINT_DEC "\n" |
b8698a0f L |
1903 | "insn count %d, mem ref count %d, prefetch count %d\n", |
1904 | ahead, unroll_factor, est_niter, | |
1905 | ninsns, mem_ref_count, prefetch_count); | |
db34470d | 1906 | |
0bbe50f6 CF |
1907 | /* Prefetching is not likely to be profitable if the instruction to prefetch |
1908 | ratio is too small. */ | |
1909 | if (insn_to_prefetch_ratio_too_small_p (ninsns, prefetch_count, | |
1910 | unroll_factor)) | |
db34470d GS |
1911 | goto fail; |
1912 | ||
1913 | mark_nontemporal_stores (loop, refs); | |
2711355f | 1914 | |
b076a3fd ZD |
1915 | /* Step 4: what to prefetch? */ |
1916 | if (!schedule_prefetches (refs, unroll_factor, ahead)) | |
1917 | goto fail; | |
1918 | ||
1919 | /* Step 5: unroll the loop. TODO -- peeling of first and last few | |
1920 | iterations so that we do not issue superfluous prefetches. */ | |
1921 | if (unroll_factor != 1) | |
1922 | { | |
d73be268 | 1923 | tree_unroll_loop (loop, unroll_factor, |
b076a3fd ZD |
1924 | single_dom_exit (loop), &desc); |
1925 | unrolled = true; | |
1926 | } | |
1927 | ||
1928 | /* Step 6: issue the prefetches. */ | |
1929 | issue_prefetches (refs, unroll_factor, ahead); | |
1930 | ||
1931 | fail: | |
1932 | release_mem_refs (refs); | |
1933 | return unrolled; | |
1934 | } | |
1935 | ||
d73be268 | 1936 | /* Issue prefetch instructions for array references in loops. */ |
b076a3fd | 1937 | |
c7f965b6 | 1938 | unsigned int |
d73be268 | 1939 | tree_ssa_prefetch_arrays (void) |
b076a3fd | 1940 | { |
b076a3fd ZD |
1941 | struct loop *loop; |
1942 | bool unrolled = false; | |
c7f965b6 | 1943 | int todo_flags = 0; |
b076a3fd | 1944 | |
134b044d | 1945 | if (!targetm.have_prefetch () |
b076a3fd ZD |
1946 | /* It is possible to ask compiler for say -mtune=i486 -march=pentium4. |
1947 | -mtune=i486 causes us having PREFETCH_BLOCK 0, since this is part | |
1948 | of processor costs and i486 does not have prefetch, but | |
134b044d | 1949 | -march=pentium4 causes targetm.have_prefetch to be true. Ugh. */ |
b076a3fd | 1950 | || PREFETCH_BLOCK == 0) |
c7f965b6 | 1951 | return 0; |
b076a3fd | 1952 | |
47eb5b32 ZD |
1953 | if (dump_file && (dump_flags & TDF_DETAILS)) |
1954 | { | |
1955 | fprintf (dump_file, "Prefetching parameters:\n"); | |
1956 | fprintf (dump_file, " simultaneous prefetches: %d\n", | |
1957 | SIMULTANEOUS_PREFETCHES); | |
1958 | fprintf (dump_file, " prefetch latency: %d\n", PREFETCH_LATENCY); | |
47eb5b32 | 1959 | fprintf (dump_file, " prefetch block size: %d\n", PREFETCH_BLOCK); |
46cb0441 ZD |
1960 | fprintf (dump_file, " L1 cache size: %d lines, %d kB\n", |
1961 | L1_CACHE_SIZE_BYTES / L1_CACHE_LINE_SIZE, L1_CACHE_SIZE); | |
5417e022 | 1962 | fprintf (dump_file, " L1 cache line size: %d\n", L1_CACHE_LINE_SIZE); |
b8698a0f L |
1963 | fprintf (dump_file, " L2 cache size: %d kB\n", L2_CACHE_SIZE); |
1964 | fprintf (dump_file, " min insn-to-prefetch ratio: %d \n", | |
db34470d | 1965 | MIN_INSN_TO_PREFETCH_RATIO); |
b8698a0f | 1966 | fprintf (dump_file, " min insn-to-mem ratio: %d \n", |
db34470d | 1967 | PREFETCH_MIN_INSN_TO_MEM_RATIO); |
47eb5b32 ZD |
1968 | fprintf (dump_file, "\n"); |
1969 | } | |
1970 | ||
b076a3fd ZD |
1971 | initialize_original_copy_tables (); |
1972 | ||
e79983f4 | 1973 | if (!builtin_decl_explicit_p (BUILT_IN_PREFETCH)) |
b076a3fd | 1974 | { |
6a4825bd NF |
1975 | tree type = build_function_type_list (void_type_node, |
1976 | const_ptr_type_node, NULL_TREE); | |
c79efc4d RÁE |
1977 | tree decl = add_builtin_function ("__builtin_prefetch", type, |
1978 | BUILT_IN_PREFETCH, BUILT_IN_NORMAL, | |
1979 | NULL, NULL_TREE); | |
b076a3fd | 1980 | DECL_IS_NOVOPS (decl) = true; |
e79983f4 | 1981 | set_builtin_decl (BUILT_IN_PREFETCH, decl, false); |
b076a3fd ZD |
1982 | } |
1983 | ||
1984 | /* We assume that size of cache line is a power of two, so verify this | |
1985 | here. */ | |
1986 | gcc_assert ((PREFETCH_BLOCK & (PREFETCH_BLOCK - 1)) == 0); | |
1987 | ||
f0bd40b1 | 1988 | FOR_EACH_LOOP (loop, LI_FROM_INNERMOST) |
b076a3fd | 1989 | { |
b076a3fd ZD |
1990 | if (dump_file && (dump_flags & TDF_DETAILS)) |
1991 | fprintf (dump_file, "Processing loop %d:\n", loop->num); | |
1992 | ||
d73be268 | 1993 | unrolled |= loop_prefetch_arrays (loop); |
b076a3fd ZD |
1994 | |
1995 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
1996 | fprintf (dump_file, "\n\n"); | |
1997 | } | |
1998 | ||
1999 | if (unrolled) | |
2000 | { | |
2001 | scev_reset (); | |
c7f965b6 | 2002 | todo_flags |= TODO_cleanup_cfg; |
b076a3fd ZD |
2003 | } |
2004 | ||
2005 | free_original_copy_tables (); | |
c7f965b6 | 2006 | return todo_flags; |
b076a3fd | 2007 | } |
71343877 AM |
2008 | |
2009 | /* Prefetching. */ | |
2010 | ||
71343877 AM |
2011 | namespace { |
2012 | ||
2013 | const pass_data pass_data_loop_prefetch = | |
2014 | { | |
2015 | GIMPLE_PASS, /* type */ | |
2016 | "aprefetch", /* name */ | |
2017 | OPTGROUP_LOOP, /* optinfo_flags */ | |
71343877 AM |
2018 | TV_TREE_PREFETCH, /* tv_id */ |
2019 | ( PROP_cfg | PROP_ssa ), /* properties_required */ | |
2020 | 0, /* properties_provided */ | |
2021 | 0, /* properties_destroyed */ | |
2022 | 0, /* todo_flags_start */ | |
2023 | 0, /* todo_flags_finish */ | |
2024 | }; | |
2025 | ||
2026 | class pass_loop_prefetch : public gimple_opt_pass | |
2027 | { | |
2028 | public: | |
2029 | pass_loop_prefetch (gcc::context *ctxt) | |
2030 | : gimple_opt_pass (pass_data_loop_prefetch, ctxt) | |
2031 | {} | |
2032 | ||
2033 | /* opt_pass methods: */ | |
1a3d085c | 2034 | virtual bool gate (function *) { return flag_prefetch_loop_arrays > 0; } |
be55bfe6 | 2035 | virtual unsigned int execute (function *); |
71343877 AM |
2036 | |
2037 | }; // class pass_loop_prefetch | |
2038 | ||
be55bfe6 TS |
2039 | unsigned int |
2040 | pass_loop_prefetch::execute (function *fun) | |
2041 | { | |
2042 | if (number_of_loops (fun) <= 1) | |
2043 | return 0; | |
2044 | ||
2045 | return tree_ssa_prefetch_arrays (); | |
2046 | } | |
2047 | ||
71343877 AM |
2048 | } // anon namespace |
2049 | ||
2050 | gimple_opt_pass * | |
2051 | make_pass_loop_prefetch (gcc::context *ctxt) | |
2052 | { | |
2053 | return new pass_loop_prefetch (ctxt); | |
2054 | } | |
2055 | ||
2056 |