Mercurial > hg > CbC > CbC_gcc
annotate gcc/tree-ssa-loop-prefetch.c @ 63:b7f97abdc517 gcc-4.6-20100522
update gcc from gcc-4.5.0 to gcc-4.6
author | ryoma <e075725@ie.u-ryukyu.ac.jp> |
---|---|
date | Mon, 24 May 2010 12:47:05 +0900 |
parents | 77e2b8dfacca |
children | f6334be47118 |
rev | line source |
---|---|
0 | 1 /* Array prefetching. |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
2 Copyright (C) 2005, 2007, 2008, 2009, 2010 Free Software Foundation, Inc. |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
3 |
0 | 4 This file is part of GCC. |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
5 |
0 | 6 GCC is free software; you can redistribute it and/or modify it |
7 under the terms of the GNU General Public License as published by the | |
8 Free Software Foundation; either version 3, or (at your option) any | |
9 later version. | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
10 |
0 | 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. | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
15 |
0 | 16 You should have received a copy of the GNU General Public License |
17 along with GCC; see the file COPYING3. If not see | |
18 <http://www.gnu.org/licenses/>. */ | |
19 | |
20 #include "config.h" | |
21 #include "system.h" | |
22 #include "coretypes.h" | |
23 #include "tm.h" | |
24 #include "tree.h" | |
25 #include "tm_p.h" | |
26 #include "basic-block.h" | |
27 #include "output.h" | |
28 #include "diagnostic.h" | |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
29 #include "tree-pretty-print.h" |
0 | 30 #include "tree-flow.h" |
31 #include "tree-dump.h" | |
32 #include "timevar.h" | |
33 #include "cfgloop.h" | |
34 #include "expr.h" | |
35 #include "tree-pass.h" | |
36 #include "insn-config.h" | |
37 #include "recog.h" | |
38 #include "hashtab.h" | |
39 #include "tree-chrec.h" | |
40 #include "tree-scalar-evolution.h" | |
41 #include "toplev.h" | |
42 #include "params.h" | |
43 #include "langhooks.h" | |
44 #include "tree-inline.h" | |
45 #include "tree-data-ref.h" | |
46 #include "optabs.h" | |
47 | |
48 /* This pass inserts prefetch instructions to optimize cache usage during | |
49 accesses to arrays in loops. It processes loops sequentially and: | |
50 | |
51 1) Gathers all memory references in the single loop. | |
52 2) For each of the references it decides when it is profitable to prefetch | |
53 it. To do it, we evaluate the reuse among the accesses, and determines | |
54 two values: PREFETCH_BEFORE (meaning that it only makes sense to do | |
55 prefetching in the first PREFETCH_BEFORE iterations of the loop) and | |
56 PREFETCH_MOD (meaning that it only makes sense to prefetch in the | |
57 iterations of the loop that are zero modulo PREFETCH_MOD). For example | |
58 (assuming cache line size is 64 bytes, char has size 1 byte and there | |
59 is no hardware sequential prefetch): | |
60 | |
61 char *a; | |
62 for (i = 0; i < max; i++) | |
63 { | |
64 a[255] = ...; (0) | |
65 a[i] = ...; (1) | |
66 a[i + 64] = ...; (2) | |
67 a[16*i] = ...; (3) | |
68 a[187*i] = ...; (4) | |
69 a[187*i + 50] = ...; (5) | |
70 } | |
71 | |
72 (0) obviously has PREFETCH_BEFORE 1 | |
73 (1) has PREFETCH_BEFORE 64, since (2) accesses the same memory | |
74 location 64 iterations before it, and PREFETCH_MOD 64 (since | |
75 it hits the same cache line otherwise). | |
76 (2) has PREFETCH_MOD 64 | |
77 (3) has PREFETCH_MOD 4 | |
78 (4) has PREFETCH_MOD 1. We do not set PREFETCH_BEFORE here, since | |
79 the cache line accessed by (4) is the same with probability only | |
80 7/32. | |
81 (5) has PREFETCH_MOD 1 as well. | |
82 | |
83 Additionally, we use data dependence analysis to determine for each | |
84 reference the distance till the first reuse; this information is used | |
85 to determine the temporality of the issued prefetch instruction. | |
86 | |
87 3) We determine how much ahead we need to prefetch. The number of | |
88 iterations needed is time to fetch / time spent in one iteration of | |
89 the loop. The problem is that we do not know either of these values, | |
90 so we just make a heuristic guess based on a magic (possibly) | |
91 target-specific constant and size of the loop. | |
92 | |
93 4) Determine which of the references we prefetch. We take into account | |
94 that there is a maximum number of simultaneous prefetches (provided | |
95 by machine description). We prefetch as many prefetches as possible | |
96 while still within this bound (starting with those with lowest | |
97 prefetch_mod, since they are responsible for most of the cache | |
98 misses). | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
99 |
0 | 100 5) We unroll and peel loops so that we are able to satisfy PREFETCH_MOD |
101 and PREFETCH_BEFORE requirements (within some bounds), and to avoid | |
102 prefetching nonaccessed memory. | |
103 TODO -- actually implement peeling. | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
104 |
0 | 105 6) We actually emit the prefetch instructions. ??? Perhaps emit the |
106 prefetch instructions with guards in cases where 5) was not sufficient | |
107 to satisfy the constraints? | |
108 | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
109 The function is_loop_prefetching_profitable() implements a cost model |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
110 to determine if prefetching is profitable for a given loop. The cost |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
111 model has two heuristcs: |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
112 1. A heuristic that determines whether the given loop has enough CPU |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
113 ops that can be overlapped with cache missing memory ops. |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
114 If not, the loop won't benefit from prefetching. This is implemented |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
115 by requirung the ratio between the instruction count and the mem ref |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
116 count to be above a certain minimum. |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
117 2. A heuristic that disables prefetching in a loop with an unknown trip |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
118 count if the prefetching cost is above a certain limit. The relative |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
119 prefetching cost is estimated by taking the ratio between the |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
120 prefetch count and the total intruction count (this models the I-cache |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
121 cost). |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
122 The limits used in these heuristics are defined as parameters with |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
123 reasonable default values. Machine-specific default values will be |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
124 added later. |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
125 |
0 | 126 Some other TODO: |
127 -- write and use more general reuse analysis (that could be also used | |
128 in other cache aimed loop optimizations) | |
129 -- make it behave sanely together with the prefetches given by user | |
130 (now we just ignore them; at the very least we should avoid | |
131 optimizing loops in that user put his own prefetches) | |
132 -- we assume cache line size alignment of arrays; this could be | |
133 improved. */ | |
134 | |
135 /* Magic constants follow. These should be replaced by machine specific | |
136 numbers. */ | |
137 | |
138 /* True if write can be prefetched by a read prefetch. */ | |
139 | |
140 #ifndef WRITE_CAN_USE_READ_PREFETCH | |
141 #define WRITE_CAN_USE_READ_PREFETCH 1 | |
142 #endif | |
143 | |
144 /* True if read can be prefetched by a write prefetch. */ | |
145 | |
146 #ifndef READ_CAN_USE_WRITE_PREFETCH | |
147 #define READ_CAN_USE_WRITE_PREFETCH 0 | |
148 #endif | |
149 | |
150 /* The size of the block loaded by a single prefetch. Usually, this is | |
151 the same as cache line size (at the moment, we only consider one level | |
152 of cache hierarchy). */ | |
153 | |
154 #ifndef PREFETCH_BLOCK | |
155 #define PREFETCH_BLOCK L1_CACHE_LINE_SIZE | |
156 #endif | |
157 | |
158 /* Do we have a forward hardware sequential prefetching? */ | |
159 | |
160 #ifndef HAVE_FORWARD_PREFETCH | |
161 #define HAVE_FORWARD_PREFETCH 0 | |
162 #endif | |
163 | |
164 /* Do we have a backward hardware sequential prefetching? */ | |
165 | |
166 #ifndef HAVE_BACKWARD_PREFETCH | |
167 #define HAVE_BACKWARD_PREFETCH 0 | |
168 #endif | |
169 | |
170 /* In some cases we are only able to determine that there is a certain | |
171 probability that the two accesses hit the same cache line. In this | |
172 case, we issue the prefetches for both of them if this probability | |
173 is less then (1000 - ACCEPTABLE_MISS_RATE) per thousand. */ | |
174 | |
175 #ifndef ACCEPTABLE_MISS_RATE | |
176 #define ACCEPTABLE_MISS_RATE 50 | |
177 #endif | |
178 | |
179 #ifndef HAVE_prefetch | |
180 #define HAVE_prefetch 0 | |
181 #endif | |
182 | |
183 #define L1_CACHE_SIZE_BYTES ((unsigned) (L1_CACHE_SIZE * 1024)) | |
184 #define L2_CACHE_SIZE_BYTES ((unsigned) (L2_CACHE_SIZE * 1024)) | |
185 | |
186 /* We consider a memory access nontemporal if it is not reused sooner than | |
187 after L2_CACHE_SIZE_BYTES of memory are accessed. However, we ignore | |
188 accesses closer than L1_CACHE_SIZE_BYTES / NONTEMPORAL_FRACTION, | |
189 so that we use nontemporal prefetches e.g. if single memory location | |
190 is accessed several times in a single iteration of the loop. */ | |
191 #define NONTEMPORAL_FRACTION 16 | |
192 | |
193 /* In case we have to emit a memory fence instruction after the loop that | |
194 uses nontemporal stores, this defines the builtin to use. */ | |
195 | |
196 #ifndef FENCE_FOLLOWING_MOVNT | |
197 #define FENCE_FOLLOWING_MOVNT NULL_TREE | |
198 #endif | |
199 | |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
200 /* It is not profitable to prefetch when the trip count is not at |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
201 least TRIP_COUNT_TO_AHEAD_RATIO times the prefetch ahead distance. |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
202 For example, in a loop with a prefetch ahead distance of 10, |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
203 supposing that TRIP_COUNT_TO_AHEAD_RATIO is equal to 4, it is |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
204 profitable to prefetch when the trip count is greater or equal to |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
205 40. In that case, 30 out of the 40 iterations will benefit from |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
206 prefetching. */ |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
207 |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
208 #ifndef TRIP_COUNT_TO_AHEAD_RATIO |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
209 #define TRIP_COUNT_TO_AHEAD_RATIO 4 |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
210 #endif |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
211 |
0 | 212 /* The group of references between that reuse may occur. */ |
213 | |
214 struct mem_ref_group | |
215 { | |
216 tree base; /* Base of the reference. */ | |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
217 tree step; /* Step of the reference. */ |
0 | 218 struct mem_ref *refs; /* References in the group. */ |
219 struct mem_ref_group *next; /* Next group of references. */ | |
220 }; | |
221 | |
222 /* Assigned to PREFETCH_BEFORE when all iterations are to be prefetched. */ | |
223 | |
224 #define PREFETCH_ALL (~(unsigned HOST_WIDE_INT) 0) | |
225 | |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
226 /* Do not generate a prefetch if the unroll factor is significantly less |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
227 than what is required by the prefetch. This is to avoid redundant |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
228 prefetches. For example, if prefetch_mod is 16 and unroll_factor is |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
229 1, this means prefetching requires unrolling the loop 16 times, but |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
230 the loop is not going to be unrolled. In this case (ratio = 16), |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
231 prefetching is not likely to be beneficial. */ |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
232 |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
233 #ifndef PREFETCH_MOD_TO_UNROLL_FACTOR_RATIO |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
234 #define PREFETCH_MOD_TO_UNROLL_FACTOR_RATIO 8 |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
235 #endif |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
236 |
0 | 237 /* The memory reference. */ |
238 | |
239 struct mem_ref | |
240 { | |
241 gimple stmt; /* Statement in that the reference appears. */ | |
242 tree mem; /* The reference. */ | |
243 HOST_WIDE_INT delta; /* Constant offset of the reference. */ | |
244 struct mem_ref_group *group; /* The group of references it belongs to. */ | |
245 unsigned HOST_WIDE_INT prefetch_mod; | |
246 /* Prefetch only each PREFETCH_MOD-th | |
247 iteration. */ | |
248 unsigned HOST_WIDE_INT prefetch_before; | |
249 /* Prefetch only first PREFETCH_BEFORE | |
250 iterations. */ | |
251 unsigned reuse_distance; /* The amount of data accessed before the first | |
252 reuse of this value. */ | |
253 struct mem_ref *next; /* The next reference in the group. */ | |
254 unsigned write_p : 1; /* Is it a write? */ | |
255 unsigned independent_p : 1; /* True if the reference is independent on | |
256 all other references inside the loop. */ | |
257 unsigned issue_prefetch_p : 1; /* Should we really issue the prefetch? */ | |
258 unsigned storent_p : 1; /* True if we changed the store to a | |
259 nontemporal one. */ | |
260 }; | |
261 | |
262 /* Dumps information about reference REF to FILE. */ | |
263 | |
264 static void | |
265 dump_mem_ref (FILE *file, struct mem_ref *ref) | |
266 { | |
267 fprintf (file, "Reference %p:\n", (void *) ref); | |
268 | |
269 fprintf (file, " group %p (base ", (void *) ref->group); | |
270 print_generic_expr (file, ref->group->base, TDF_SLIM); | |
271 fprintf (file, ", step "); | |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
272 if (cst_and_fits_in_hwi (ref->group->step)) |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
273 fprintf (file, HOST_WIDE_INT_PRINT_DEC, int_cst_value (ref->group->step)); |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
274 else |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
275 print_generic_expr (file, ref->group->step, TDF_TREE); |
0 | 276 fprintf (file, ")\n"); |
277 | |
278 fprintf (file, " delta "); | |
279 fprintf (file, HOST_WIDE_INT_PRINT_DEC, ref->delta); | |
280 fprintf (file, "\n"); | |
281 | |
282 fprintf (file, " %s\n", ref->write_p ? "write" : "read"); | |
283 | |
284 fprintf (file, "\n"); | |
285 } | |
286 | |
287 /* Finds a group with BASE and STEP in GROUPS, or creates one if it does not | |
288 exist. */ | |
289 | |
290 static struct mem_ref_group * | |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
291 find_or_create_group (struct mem_ref_group **groups, tree base, tree step) |
0 | 292 { |
293 struct mem_ref_group *group; | |
294 | |
295 for (; *groups; groups = &(*groups)->next) | |
296 { | |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
297 if (operand_equal_p ((*groups)->step, step, 0) |
0 | 298 && operand_equal_p ((*groups)->base, base, 0)) |
299 return *groups; | |
300 | |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
301 /* If step is an integer constant, keep the list of groups sorted |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
302 by decreasing step. */ |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
303 if (cst_and_fits_in_hwi ((*groups)->step) && cst_and_fits_in_hwi (step) |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
304 && int_cst_value ((*groups)->step) < int_cst_value (step)) |
0 | 305 break; |
306 } | |
307 | |
308 group = XNEW (struct mem_ref_group); | |
309 group->base = base; | |
310 group->step = step; | |
311 group->refs = NULL; | |
312 group->next = *groups; | |
313 *groups = group; | |
314 | |
315 return group; | |
316 } | |
317 | |
318 /* Records a memory reference MEM in GROUP with offset DELTA and write status | |
319 WRITE_P. The reference occurs in statement STMT. */ | |
320 | |
321 static void | |
322 record_ref (struct mem_ref_group *group, gimple stmt, tree mem, | |
323 HOST_WIDE_INT delta, bool write_p) | |
324 { | |
325 struct mem_ref **aref; | |
326 | |
327 /* Do not record the same address twice. */ | |
328 for (aref = &group->refs; *aref; aref = &(*aref)->next) | |
329 { | |
330 /* It does not have to be possible for write reference to reuse the read | |
331 prefetch, or vice versa. */ | |
332 if (!WRITE_CAN_USE_READ_PREFETCH | |
333 && write_p | |
334 && !(*aref)->write_p) | |
335 continue; | |
336 if (!READ_CAN_USE_WRITE_PREFETCH | |
337 && !write_p | |
338 && (*aref)->write_p) | |
339 continue; | |
340 | |
341 if ((*aref)->delta == delta) | |
342 return; | |
343 } | |
344 | |
345 (*aref) = XNEW (struct mem_ref); | |
346 (*aref)->stmt = stmt; | |
347 (*aref)->mem = mem; | |
348 (*aref)->delta = delta; | |
349 (*aref)->write_p = write_p; | |
350 (*aref)->prefetch_before = PREFETCH_ALL; | |
351 (*aref)->prefetch_mod = 1; | |
352 (*aref)->reuse_distance = 0; | |
353 (*aref)->issue_prefetch_p = false; | |
354 (*aref)->group = group; | |
355 (*aref)->next = NULL; | |
356 (*aref)->independent_p = false; | |
357 (*aref)->storent_p = false; | |
358 | |
359 if (dump_file && (dump_flags & TDF_DETAILS)) | |
360 dump_mem_ref (dump_file, *aref); | |
361 } | |
362 | |
363 /* Release memory references in GROUPS. */ | |
364 | |
365 static void | |
366 release_mem_refs (struct mem_ref_group *groups) | |
367 { | |
368 struct mem_ref_group *next_g; | |
369 struct mem_ref *ref, *next_r; | |
370 | |
371 for (; groups; groups = next_g) | |
372 { | |
373 next_g = groups->next; | |
374 for (ref = groups->refs; ref; ref = next_r) | |
375 { | |
376 next_r = ref->next; | |
377 free (ref); | |
378 } | |
379 free (groups); | |
380 } | |
381 } | |
382 | |
383 /* A structure used to pass arguments to idx_analyze_ref. */ | |
384 | |
385 struct ar_data | |
386 { | |
387 struct loop *loop; /* Loop of the reference. */ | |
388 gimple stmt; /* Statement of the reference. */ | |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
389 tree *step; /* Step of the memory reference. */ |
0 | 390 HOST_WIDE_INT *delta; /* Offset of the memory reference. */ |
391 }; | |
392 | |
393 /* Analyzes a single INDEX of a memory reference to obtain information | |
394 described at analyze_ref. Callback for for_each_index. */ | |
395 | |
396 static bool | |
397 idx_analyze_ref (tree base, tree *index, void *data) | |
398 { | |
399 struct ar_data *ar_data = (struct ar_data *) data; | |
400 tree ibase, step, stepsize; | |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
401 HOST_WIDE_INT idelta = 0, imult = 1; |
0 | 402 affine_iv iv; |
403 | |
404 if (TREE_CODE (base) == MISALIGNED_INDIRECT_REF | |
405 || TREE_CODE (base) == ALIGN_INDIRECT_REF) | |
406 return false; | |
407 | |
408 if (!simple_iv (ar_data->loop, loop_containing_stmt (ar_data->stmt), | |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
409 *index, &iv, true)) |
0 | 410 return false; |
411 ibase = iv.base; | |
412 step = iv.step; | |
413 | |
414 if (TREE_CODE (ibase) == POINTER_PLUS_EXPR | |
415 && cst_and_fits_in_hwi (TREE_OPERAND (ibase, 1))) | |
416 { | |
417 idelta = int_cst_value (TREE_OPERAND (ibase, 1)); | |
418 ibase = TREE_OPERAND (ibase, 0); | |
419 } | |
420 if (cst_and_fits_in_hwi (ibase)) | |
421 { | |
422 idelta += int_cst_value (ibase); | |
423 ibase = build_int_cst (TREE_TYPE (ibase), 0); | |
424 } | |
425 | |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
426 if (*ar_data->step == NULL_TREE) |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
427 *ar_data->step = step; |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
428 else |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
429 *ar_data->step = fold_build2 (PLUS_EXPR, sizetype, |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
430 fold_convert (sizetype, *ar_data->step), |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
431 fold_convert (sizetype, step)); |
0 | 432 if (TREE_CODE (base) == ARRAY_REF) |
433 { | |
434 stepsize = array_ref_element_size (base); | |
435 if (!cst_and_fits_in_hwi (stepsize)) | |
436 return false; | |
437 imult = int_cst_value (stepsize); | |
438 | |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
439 *ar_data->step = fold_build2 (MULT_EXPR, sizetype, |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
440 fold_convert (sizetype, *ar_data->step), |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
441 fold_convert (sizetype, step)); |
0 | 442 idelta *= imult; |
443 } | |
444 | |
445 *ar_data->delta += idelta; | |
446 *index = ibase; | |
447 | |
448 return true; | |
449 } | |
450 | |
451 /* Tries to express REF_P in shape &BASE + STEP * iter + DELTA, where DELTA and | |
452 STEP are integer constants and iter is number of iterations of LOOP. The | |
453 reference occurs in statement STMT. Strips nonaddressable component | |
454 references from REF_P. */ | |
455 | |
456 static bool | |
457 analyze_ref (struct loop *loop, tree *ref_p, tree *base, | |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
458 tree *step, HOST_WIDE_INT *delta, |
0 | 459 gimple stmt) |
460 { | |
461 struct ar_data ar_data; | |
462 tree off; | |
463 HOST_WIDE_INT bit_offset; | |
464 tree ref = *ref_p; | |
465 | |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
466 *step = NULL_TREE; |
0 | 467 *delta = 0; |
468 | |
469 /* First strip off the component references. Ignore bitfields. */ | |
470 if (TREE_CODE (ref) == COMPONENT_REF | |
471 && DECL_NONADDRESSABLE_P (TREE_OPERAND (ref, 1))) | |
472 ref = TREE_OPERAND (ref, 0); | |
473 | |
474 *ref_p = ref; | |
475 | |
476 for (; TREE_CODE (ref) == COMPONENT_REF; ref = TREE_OPERAND (ref, 0)) | |
477 { | |
478 off = DECL_FIELD_BIT_OFFSET (TREE_OPERAND (ref, 1)); | |
479 bit_offset = TREE_INT_CST_LOW (off); | |
480 gcc_assert (bit_offset % BITS_PER_UNIT == 0); | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
481 |
0 | 482 *delta += bit_offset / BITS_PER_UNIT; |
483 } | |
484 | |
485 *base = unshare_expr (ref); | |
486 ar_data.loop = loop; | |
487 ar_data.stmt = stmt; | |
488 ar_data.step = step; | |
489 ar_data.delta = delta; | |
490 return for_each_index (base, idx_analyze_ref, &ar_data); | |
491 } | |
492 | |
493 /* Record a memory reference REF to the list REFS. The reference occurs in | |
494 LOOP in statement STMT and it is write if WRITE_P. Returns true if the | |
495 reference was recorded, false otherwise. */ | |
496 | |
497 static bool | |
498 gather_memory_references_ref (struct loop *loop, struct mem_ref_group **refs, | |
499 tree ref, bool write_p, gimple stmt) | |
500 { | |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
501 tree base, step; |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
502 HOST_WIDE_INT delta; |
0 | 503 struct mem_ref_group *agrp; |
504 | |
505 if (get_base_address (ref) == NULL) | |
506 return false; | |
507 | |
508 if (!analyze_ref (loop, &ref, &base, &step, &delta, stmt)) | |
509 return false; | |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
510 /* If analyze_ref fails the default is a NULL_TREE. We can stop here. */ |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
511 if (step == NULL_TREE) |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
512 return false; |
0 | 513 |
514 /* Now we know that REF = &BASE + STEP * iter + DELTA, where DELTA and STEP | |
515 are integer constants. */ | |
516 agrp = find_or_create_group (refs, base, step); | |
517 record_ref (agrp, stmt, ref, delta, write_p); | |
518 | |
519 return true; | |
520 } | |
521 | |
522 /* Record the suitable memory references in LOOP. NO_OTHER_REFS is set to | |
523 true if there are no other memory references inside the loop. */ | |
524 | |
525 static struct mem_ref_group * | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
526 gather_memory_references (struct loop *loop, bool *no_other_refs, unsigned *ref_count) |
0 | 527 { |
528 basic_block *body = get_loop_body_in_dom_order (loop); | |
529 basic_block bb; | |
530 unsigned i; | |
531 gimple_stmt_iterator bsi; | |
532 gimple stmt; | |
533 tree lhs, rhs; | |
534 struct mem_ref_group *refs = NULL; | |
535 | |
536 *no_other_refs = true; | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
537 *ref_count = 0; |
0 | 538 |
539 /* Scan the loop body in order, so that the former references precede the | |
540 later ones. */ | |
541 for (i = 0; i < loop->num_nodes; i++) | |
542 { | |
543 bb = body[i]; | |
544 if (bb->loop_father != loop) | |
545 continue; | |
546 | |
547 for (bsi = gsi_start_bb (bb); !gsi_end_p (bsi); gsi_next (&bsi)) | |
548 { | |
549 stmt = gsi_stmt (bsi); | |
550 | |
551 if (gimple_code (stmt) != GIMPLE_ASSIGN) | |
552 { | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
553 if (gimple_vuse (stmt) |
0 | 554 || (is_gimple_call (stmt) |
555 && !(gimple_call_flags (stmt) & ECF_CONST))) | |
556 *no_other_refs = false; | |
557 continue; | |
558 } | |
559 | |
560 lhs = gimple_assign_lhs (stmt); | |
561 rhs = gimple_assign_rhs1 (stmt); | |
562 | |
563 if (REFERENCE_CLASS_P (rhs)) | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
564 { |
0 | 565 *no_other_refs &= gather_memory_references_ref (loop, &refs, |
566 rhs, false, stmt); | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
567 *ref_count += 1; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
568 } |
0 | 569 if (REFERENCE_CLASS_P (lhs)) |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
570 { |
0 | 571 *no_other_refs &= gather_memory_references_ref (loop, &refs, |
572 lhs, true, stmt); | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
573 *ref_count += 1; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
574 } |
0 | 575 } |
576 } | |
577 free (body); | |
578 | |
579 return refs; | |
580 } | |
581 | |
582 /* Prune the prefetch candidate REF using the self-reuse. */ | |
583 | |
584 static void | |
585 prune_ref_by_self_reuse (struct mem_ref *ref) | |
586 { | |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
587 HOST_WIDE_INT step; |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
588 bool backward; |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
589 |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
590 /* If the step size is non constant, we cannot calculate prefetch_mod. */ |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
591 if (!cst_and_fits_in_hwi (ref->group->step)) |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
592 return; |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
593 |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
594 step = int_cst_value (ref->group->step); |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
595 |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
596 backward = step < 0; |
0 | 597 |
598 if (step == 0) | |
599 { | |
600 /* Prefetch references to invariant address just once. */ | |
601 ref->prefetch_before = 1; | |
602 return; | |
603 } | |
604 | |
605 if (backward) | |
606 step = -step; | |
607 | |
608 if (step > PREFETCH_BLOCK) | |
609 return; | |
610 | |
611 if ((backward && HAVE_BACKWARD_PREFETCH) | |
612 || (!backward && HAVE_FORWARD_PREFETCH)) | |
613 { | |
614 ref->prefetch_before = 1; | |
615 return; | |
616 } | |
617 | |
618 ref->prefetch_mod = PREFETCH_BLOCK / step; | |
619 } | |
620 | |
621 /* Divides X by BY, rounding down. */ | |
622 | |
623 static HOST_WIDE_INT | |
624 ddown (HOST_WIDE_INT x, unsigned HOST_WIDE_INT by) | |
625 { | |
626 gcc_assert (by > 0); | |
627 | |
628 if (x >= 0) | |
629 return x / by; | |
630 else | |
631 return (x + by - 1) / by; | |
632 } | |
633 | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
634 /* Given a CACHE_LINE_SIZE and two inductive memory references |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
635 with a common STEP greater than CACHE_LINE_SIZE and an address |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
636 difference DELTA, compute the probability that they will fall |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
637 in different cache lines. DISTINCT_ITERS is the number of |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
638 distinct iterations after which the pattern repeats itself. |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
639 ALIGN_UNIT is the unit of alignment in bytes. */ |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
640 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
641 static int |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
642 compute_miss_rate (unsigned HOST_WIDE_INT cache_line_size, |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
643 HOST_WIDE_INT step, HOST_WIDE_INT delta, |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
644 unsigned HOST_WIDE_INT distinct_iters, |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
645 int align_unit) |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
646 { |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
647 unsigned align, iter; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
648 int total_positions, miss_positions, miss_rate; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
649 int address1, address2, cache_line1, cache_line2; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
650 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
651 total_positions = 0; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
652 miss_positions = 0; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
653 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
654 /* Iterate through all possible alignments of the first |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
655 memory reference within its cache line. */ |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
656 for (align = 0; align < cache_line_size; align += align_unit) |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
657 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
658 /* Iterate through all distinct iterations. */ |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
659 for (iter = 0; iter < distinct_iters; iter++) |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
660 { |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
661 address1 = align + step * iter; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
662 address2 = address1 + delta; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
663 cache_line1 = address1 / cache_line_size; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
664 cache_line2 = address2 / cache_line_size; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
665 total_positions += 1; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
666 if (cache_line1 != cache_line2) |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
667 miss_positions += 1; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
668 } |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
669 miss_rate = 1000 * miss_positions / total_positions; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
670 return miss_rate; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
671 } |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
672 |
0 | 673 /* Prune the prefetch candidate REF using the reuse with BY. |
674 If BY_IS_BEFORE is true, BY is before REF in the loop. */ | |
675 | |
676 static void | |
677 prune_ref_by_group_reuse (struct mem_ref *ref, struct mem_ref *by, | |
678 bool by_is_before) | |
679 { | |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
680 HOST_WIDE_INT step; |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
681 bool backward; |
0 | 682 HOST_WIDE_INT delta_r = ref->delta, delta_b = by->delta; |
683 HOST_WIDE_INT delta = delta_b - delta_r; | |
684 HOST_WIDE_INT hit_from; | |
685 unsigned HOST_WIDE_INT prefetch_before, prefetch_block; | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
686 int miss_rate; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
687 HOST_WIDE_INT reduced_step; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
688 unsigned HOST_WIDE_INT reduced_prefetch_block; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
689 tree ref_type; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
690 int align_unit; |
0 | 691 |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
692 /* If the step is non constant we cannot calculate prefetch_before. */ |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
693 if (!cst_and_fits_in_hwi (ref->group->step)) { |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
694 return; |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
695 } |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
696 |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
697 step = int_cst_value (ref->group->step); |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
698 |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
699 backward = step < 0; |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
700 |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
701 |
0 | 702 if (delta == 0) |
703 { | |
704 /* If the references has the same address, only prefetch the | |
705 former. */ | |
706 if (by_is_before) | |
707 ref->prefetch_before = 0; | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
708 |
0 | 709 return; |
710 } | |
711 | |
712 if (!step) | |
713 { | |
714 /* If the reference addresses are invariant and fall into the | |
715 same cache line, prefetch just the first one. */ | |
716 if (!by_is_before) | |
717 return; | |
718 | |
719 if (ddown (ref->delta, PREFETCH_BLOCK) | |
720 != ddown (by->delta, PREFETCH_BLOCK)) | |
721 return; | |
722 | |
723 ref->prefetch_before = 0; | |
724 return; | |
725 } | |
726 | |
727 /* Only prune the reference that is behind in the array. */ | |
728 if (backward) | |
729 { | |
730 if (delta > 0) | |
731 return; | |
732 | |
733 /* Transform the data so that we may assume that the accesses | |
734 are forward. */ | |
735 delta = - delta; | |
736 step = -step; | |
737 delta_r = PREFETCH_BLOCK - 1 - delta_r; | |
738 delta_b = PREFETCH_BLOCK - 1 - delta_b; | |
739 } | |
740 else | |
741 { | |
742 if (delta < 0) | |
743 return; | |
744 } | |
745 | |
746 /* Check whether the two references are likely to hit the same cache | |
747 line, and how distant the iterations in that it occurs are from | |
748 each other. */ | |
749 | |
750 if (step <= PREFETCH_BLOCK) | |
751 { | |
752 /* The accesses are sure to meet. Let us check when. */ | |
753 hit_from = ddown (delta_b, PREFETCH_BLOCK) * PREFETCH_BLOCK; | |
754 prefetch_before = (hit_from - delta_r + step - 1) / step; | |
755 | |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
756 /* Do not reduce prefetch_before if we meet beyond cache size. */ |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
757 if (prefetch_before > (unsigned) abs (L2_CACHE_SIZE_BYTES / step)) |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
758 prefetch_before = PREFETCH_ALL; |
0 | 759 if (prefetch_before < ref->prefetch_before) |
760 ref->prefetch_before = prefetch_before; | |
761 | |
762 return; | |
763 } | |
764 | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
765 /* A more complicated case with step > prefetch_block. First reduce |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
766 the ratio between the step and the cache line size to its simplest |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
767 terms. The resulting denominator will then represent the number of |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
768 distinct iterations after which each address will go back to its |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
769 initial location within the cache line. This computation assumes |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
770 that PREFETCH_BLOCK is a power of two. */ |
0 | 771 prefetch_block = PREFETCH_BLOCK; |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
772 reduced_prefetch_block = prefetch_block; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
773 reduced_step = step; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
774 while ((reduced_step & 1) == 0 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
775 && reduced_prefetch_block > 1) |
0 | 776 { |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
777 reduced_step >>= 1; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
778 reduced_prefetch_block >>= 1; |
0 | 779 } |
780 | |
781 prefetch_before = delta / step; | |
782 delta %= step; | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
783 ref_type = TREE_TYPE (ref->mem); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
784 align_unit = TYPE_ALIGN (ref_type) / 8; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
785 miss_rate = compute_miss_rate(prefetch_block, step, delta, |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
786 reduced_prefetch_block, align_unit); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
787 if (miss_rate <= ACCEPTABLE_MISS_RATE) |
0 | 788 { |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
789 /* Do not reduce prefetch_before if we meet beyond cache size. */ |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
790 if (prefetch_before > L2_CACHE_SIZE_BYTES / PREFETCH_BLOCK) |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
791 prefetch_before = PREFETCH_ALL; |
0 | 792 if (prefetch_before < ref->prefetch_before) |
793 ref->prefetch_before = prefetch_before; | |
794 | |
795 return; | |
796 } | |
797 | |
798 /* Try also the following iteration. */ | |
799 prefetch_before++; | |
800 delta = step - delta; | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
801 miss_rate = compute_miss_rate(prefetch_block, step, delta, |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
802 reduced_prefetch_block, align_unit); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
803 if (miss_rate <= ACCEPTABLE_MISS_RATE) |
0 | 804 { |
805 if (prefetch_before < ref->prefetch_before) | |
806 ref->prefetch_before = prefetch_before; | |
807 | |
808 return; | |
809 } | |
810 | |
811 /* The ref probably does not reuse by. */ | |
812 return; | |
813 } | |
814 | |
815 /* Prune the prefetch candidate REF using the reuses with other references | |
816 in REFS. */ | |
817 | |
818 static void | |
819 prune_ref_by_reuse (struct mem_ref *ref, struct mem_ref *refs) | |
820 { | |
821 struct mem_ref *prune_by; | |
822 bool before = true; | |
823 | |
824 prune_ref_by_self_reuse (ref); | |
825 | |
826 for (prune_by = refs; prune_by; prune_by = prune_by->next) | |
827 { | |
828 if (prune_by == ref) | |
829 { | |
830 before = false; | |
831 continue; | |
832 } | |
833 | |
834 if (!WRITE_CAN_USE_READ_PREFETCH | |
835 && ref->write_p | |
836 && !prune_by->write_p) | |
837 continue; | |
838 if (!READ_CAN_USE_WRITE_PREFETCH | |
839 && !ref->write_p | |
840 && prune_by->write_p) | |
841 continue; | |
842 | |
843 prune_ref_by_group_reuse (ref, prune_by, before); | |
844 } | |
845 } | |
846 | |
847 /* Prune the prefetch candidates in GROUP using the reuse analysis. */ | |
848 | |
849 static void | |
850 prune_group_by_reuse (struct mem_ref_group *group) | |
851 { | |
852 struct mem_ref *ref_pruned; | |
853 | |
854 for (ref_pruned = group->refs; ref_pruned; ref_pruned = ref_pruned->next) | |
855 { | |
856 prune_ref_by_reuse (ref_pruned, group->refs); | |
857 | |
858 if (dump_file && (dump_flags & TDF_DETAILS)) | |
859 { | |
860 fprintf (dump_file, "Reference %p:", (void *) ref_pruned); | |
861 | |
862 if (ref_pruned->prefetch_before == PREFETCH_ALL | |
863 && ref_pruned->prefetch_mod == 1) | |
864 fprintf (dump_file, " no restrictions"); | |
865 else if (ref_pruned->prefetch_before == 0) | |
866 fprintf (dump_file, " do not prefetch"); | |
867 else if (ref_pruned->prefetch_before <= ref_pruned->prefetch_mod) | |
868 fprintf (dump_file, " prefetch once"); | |
869 else | |
870 { | |
871 if (ref_pruned->prefetch_before != PREFETCH_ALL) | |
872 { | |
873 fprintf (dump_file, " prefetch before "); | |
874 fprintf (dump_file, HOST_WIDE_INT_PRINT_DEC, | |
875 ref_pruned->prefetch_before); | |
876 } | |
877 if (ref_pruned->prefetch_mod != 1) | |
878 { | |
879 fprintf (dump_file, " prefetch mod "); | |
880 fprintf (dump_file, HOST_WIDE_INT_PRINT_DEC, | |
881 ref_pruned->prefetch_mod); | |
882 } | |
883 } | |
884 fprintf (dump_file, "\n"); | |
885 } | |
886 } | |
887 } | |
888 | |
889 /* Prune the list of prefetch candidates GROUPS using the reuse analysis. */ | |
890 | |
891 static void | |
892 prune_by_reuse (struct mem_ref_group *groups) | |
893 { | |
894 for (; groups; groups = groups->next) | |
895 prune_group_by_reuse (groups); | |
896 } | |
897 | |
898 /* Returns true if we should issue prefetch for REF. */ | |
899 | |
900 static bool | |
901 should_issue_prefetch_p (struct mem_ref *ref) | |
902 { | |
903 /* For now do not issue prefetches for only first few of the | |
904 iterations. */ | |
905 if (ref->prefetch_before != PREFETCH_ALL) | |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
906 { |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
907 if (dump_file && (dump_flags & TDF_DETAILS)) |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
908 fprintf (dump_file, "Ignoring %p due to prefetch_before\n", |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
909 (void *) ref); |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
910 return false; |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
911 } |
0 | 912 |
913 /* Do not prefetch nontemporal stores. */ | |
914 if (ref->storent_p) | |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
915 { |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
916 if (dump_file && (dump_flags & TDF_DETAILS)) |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
917 fprintf (dump_file, "Ignoring nontemporal store %p\n", (void *) ref); |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
918 return false; |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
919 } |
0 | 920 |
921 return true; | |
922 } | |
923 | |
924 /* Decide which of the prefetch candidates in GROUPS to prefetch. | |
925 AHEAD is the number of iterations to prefetch ahead (which corresponds | |
926 to the number of simultaneous instances of one prefetch running at a | |
927 time). UNROLL_FACTOR is the factor by that the loop is going to be | |
928 unrolled. Returns true if there is anything to prefetch. */ | |
929 | |
930 static bool | |
931 schedule_prefetches (struct mem_ref_group *groups, unsigned unroll_factor, | |
932 unsigned ahead) | |
933 { | |
934 unsigned remaining_prefetch_slots, n_prefetches, prefetch_slots; | |
935 unsigned slots_per_prefetch; | |
936 struct mem_ref *ref; | |
937 bool any = false; | |
938 | |
939 /* At most SIMULTANEOUS_PREFETCHES should be running at the same time. */ | |
940 remaining_prefetch_slots = SIMULTANEOUS_PREFETCHES; | |
941 | |
942 /* The prefetch will run for AHEAD iterations of the original loop, i.e., | |
943 AHEAD / UNROLL_FACTOR iterations of the unrolled loop. In each iteration, | |
944 it will need a prefetch slot. */ | |
945 slots_per_prefetch = (ahead + unroll_factor / 2) / unroll_factor; | |
946 if (dump_file && (dump_flags & TDF_DETAILS)) | |
947 fprintf (dump_file, "Each prefetch instruction takes %u prefetch slots.\n", | |
948 slots_per_prefetch); | |
949 | |
950 /* For now we just take memory references one by one and issue | |
951 prefetches for as many as possible. The groups are sorted | |
952 starting with the largest step, since the references with | |
953 large step are more likely to cause many cache misses. */ | |
954 | |
955 for (; groups; groups = groups->next) | |
956 for (ref = groups->refs; ref; ref = ref->next) | |
957 { | |
958 if (!should_issue_prefetch_p (ref)) | |
959 continue; | |
960 | |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
961 /* The loop is far from being sufficiently unrolled for this |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
962 prefetch. Do not generate prefetch to avoid many redudant |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
963 prefetches. */ |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
964 if (ref->prefetch_mod / unroll_factor > PREFETCH_MOD_TO_UNROLL_FACTOR_RATIO) |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
965 continue; |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
966 |
0 | 967 /* If we need to prefetch the reference each PREFETCH_MOD iterations, |
968 and we unroll the loop UNROLL_FACTOR times, we need to insert | |
969 ceil (UNROLL_FACTOR / PREFETCH_MOD) instructions in each | |
970 iteration. */ | |
971 n_prefetches = ((unroll_factor + ref->prefetch_mod - 1) | |
972 / ref->prefetch_mod); | |
973 prefetch_slots = n_prefetches * slots_per_prefetch; | |
974 | |
975 /* If more than half of the prefetches would be lost anyway, do not | |
976 issue the prefetch. */ | |
977 if (2 * remaining_prefetch_slots < prefetch_slots) | |
978 continue; | |
979 | |
980 ref->issue_prefetch_p = true; | |
981 | |
982 if (remaining_prefetch_slots <= prefetch_slots) | |
983 return true; | |
984 remaining_prefetch_slots -= prefetch_slots; | |
985 any = true; | |
986 } | |
987 | |
988 return any; | |
989 } | |
990 | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
991 /* Estimate the number of prefetches in the given GROUPS. */ |
0 | 992 |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
993 static int |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
994 estimate_prefetch_count (struct mem_ref_group *groups) |
0 | 995 { |
996 struct mem_ref *ref; | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
997 int prefetch_count = 0; |
0 | 998 |
999 for (; groups; groups = groups->next) | |
1000 for (ref = groups->refs; ref; ref = ref->next) | |
1001 if (should_issue_prefetch_p (ref)) | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1002 prefetch_count++; |
0 | 1003 |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1004 return prefetch_count; |
0 | 1005 } |
1006 | |
1007 /* Issue prefetches for the reference REF into loop as decided before. | |
1008 HEAD is the number of iterations to prefetch ahead. UNROLL_FACTOR | |
1009 is the factor by which LOOP was unrolled. */ | |
1010 | |
1011 static void | |
1012 issue_prefetch_ref (struct mem_ref *ref, unsigned unroll_factor, unsigned ahead) | |
1013 { | |
1014 HOST_WIDE_INT delta; | |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1015 tree addr, addr_base, write_p, local, forward; |
0 | 1016 gimple prefetch; |
1017 gimple_stmt_iterator bsi; | |
1018 unsigned n_prefetches, ap; | |
1019 bool nontemporal = ref->reuse_distance >= L2_CACHE_SIZE_BYTES; | |
1020 | |
1021 if (dump_file && (dump_flags & TDF_DETAILS)) | |
1022 fprintf (dump_file, "Issued%s prefetch for %p.\n", | |
1023 nontemporal ? " nontemporal" : "", | |
1024 (void *) ref); | |
1025 | |
1026 bsi = gsi_for_stmt (ref->stmt); | |
1027 | |
1028 n_prefetches = ((unroll_factor + ref->prefetch_mod - 1) | |
1029 / ref->prefetch_mod); | |
1030 addr_base = build_fold_addr_expr_with_type (ref->mem, ptr_type_node); | |
1031 addr_base = force_gimple_operand_gsi (&bsi, unshare_expr (addr_base), | |
1032 true, NULL, true, GSI_SAME_STMT); | |
1033 write_p = ref->write_p ? integer_one_node : integer_zero_node; | |
1034 local = build_int_cst (integer_type_node, nontemporal ? 0 : 3); | |
1035 | |
1036 for (ap = 0; ap < n_prefetches; ap++) | |
1037 { | |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1038 if (cst_and_fits_in_hwi (ref->group->step)) |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1039 { |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1040 /* Determine the address to prefetch. */ |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1041 delta = (ahead + ap * ref->prefetch_mod) * |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1042 int_cst_value (ref->group->step); |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1043 addr = fold_build2 (POINTER_PLUS_EXPR, ptr_type_node, |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1044 addr_base, size_int (delta)); |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1045 addr = force_gimple_operand_gsi (&bsi, unshare_expr (addr), true, NULL, |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1046 true, GSI_SAME_STMT); |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1047 } |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1048 else |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1049 { |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1050 /* The step size is non-constant but loop-invariant. We use the |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1051 heuristic to simply prefetch ahead iterations ahead. */ |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1052 forward = fold_build2 (MULT_EXPR, sizetype, |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1053 fold_convert (sizetype, ref->group->step), |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1054 fold_convert (sizetype, size_int (ahead))); |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1055 addr = fold_build2 (POINTER_PLUS_EXPR, ptr_type_node, addr_base, |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1056 forward); |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1057 addr = force_gimple_operand_gsi (&bsi, unshare_expr (addr), true, |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1058 NULL, true, GSI_SAME_STMT); |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1059 } |
0 | 1060 /* Create the prefetch instruction. */ |
1061 prefetch = gimple_build_call (built_in_decls[BUILT_IN_PREFETCH], | |
1062 3, addr, write_p, local); | |
1063 gsi_insert_before (&bsi, prefetch, GSI_SAME_STMT); | |
1064 } | |
1065 } | |
1066 | |
1067 /* Issue prefetches for the references in GROUPS into loop as decided before. | |
1068 HEAD is the number of iterations to prefetch ahead. UNROLL_FACTOR is the | |
1069 factor by that LOOP was unrolled. */ | |
1070 | |
1071 static void | |
1072 issue_prefetches (struct mem_ref_group *groups, | |
1073 unsigned unroll_factor, unsigned ahead) | |
1074 { | |
1075 struct mem_ref *ref; | |
1076 | |
1077 for (; groups; groups = groups->next) | |
1078 for (ref = groups->refs; ref; ref = ref->next) | |
1079 if (ref->issue_prefetch_p) | |
1080 issue_prefetch_ref (ref, unroll_factor, ahead); | |
1081 } | |
1082 | |
1083 /* Returns true if REF is a memory write for that a nontemporal store insn | |
1084 can be used. */ | |
1085 | |
1086 static bool | |
1087 nontemporal_store_p (struct mem_ref *ref) | |
1088 { | |
1089 enum machine_mode mode; | |
1090 enum insn_code code; | |
1091 | |
1092 /* REF must be a write that is not reused. We require it to be independent | |
1093 on all other memory references in the loop, as the nontemporal stores may | |
1094 be reordered with respect to other memory references. */ | |
1095 if (!ref->write_p | |
1096 || !ref->independent_p | |
1097 || ref->reuse_distance < L2_CACHE_SIZE_BYTES) | |
1098 return false; | |
1099 | |
1100 /* Check that we have the storent instruction for the mode. */ | |
1101 mode = TYPE_MODE (TREE_TYPE (ref->mem)); | |
1102 if (mode == BLKmode) | |
1103 return false; | |
1104 | |
1105 code = optab_handler (storent_optab, mode)->insn_code; | |
1106 return code != CODE_FOR_nothing; | |
1107 } | |
1108 | |
1109 /* If REF is a nontemporal store, we mark the corresponding modify statement | |
1110 and return true. Otherwise, we return false. */ | |
1111 | |
1112 static bool | |
1113 mark_nontemporal_store (struct mem_ref *ref) | |
1114 { | |
1115 if (!nontemporal_store_p (ref)) | |
1116 return false; | |
1117 | |
1118 if (dump_file && (dump_flags & TDF_DETAILS)) | |
1119 fprintf (dump_file, "Marked reference %p as a nontemporal store.\n", | |
1120 (void *) ref); | |
1121 | |
1122 gimple_assign_set_nontemporal_move (ref->stmt, true); | |
1123 ref->storent_p = true; | |
1124 | |
1125 return true; | |
1126 } | |
1127 | |
1128 /* Issue a memory fence instruction after LOOP. */ | |
1129 | |
1130 static void | |
1131 emit_mfence_after_loop (struct loop *loop) | |
1132 { | |
1133 VEC (edge, heap) *exits = get_loop_exit_edges (loop); | |
1134 edge exit; | |
1135 gimple call; | |
1136 gimple_stmt_iterator bsi; | |
1137 unsigned i; | |
1138 | |
1139 for (i = 0; VEC_iterate (edge, exits, i, exit); i++) | |
1140 { | |
1141 call = gimple_build_call (FENCE_FOLLOWING_MOVNT, 0); | |
1142 | |
1143 if (!single_pred_p (exit->dest) | |
1144 /* If possible, we prefer not to insert the fence on other paths | |
1145 in cfg. */ | |
1146 && !(exit->flags & EDGE_ABNORMAL)) | |
1147 split_loop_exit_edge (exit); | |
1148 bsi = gsi_after_labels (exit->dest); | |
1149 | |
1150 gsi_insert_before (&bsi, call, GSI_NEW_STMT); | |
1151 mark_virtual_ops_for_renaming (call); | |
1152 } | |
1153 | |
1154 VEC_free (edge, heap, exits); | |
1155 update_ssa (TODO_update_ssa_only_virtuals); | |
1156 } | |
1157 | |
1158 /* Returns true if we can use storent in loop, false otherwise. */ | |
1159 | |
1160 static bool | |
1161 may_use_storent_in_loop_p (struct loop *loop) | |
1162 { | |
1163 bool ret = true; | |
1164 | |
1165 if (loop->inner != NULL) | |
1166 return false; | |
1167 | |
1168 /* If we must issue a mfence insn after using storent, check that there | |
1169 is a suitable place for it at each of the loop exits. */ | |
1170 if (FENCE_FOLLOWING_MOVNT != NULL_TREE) | |
1171 { | |
1172 VEC (edge, heap) *exits = get_loop_exit_edges (loop); | |
1173 unsigned i; | |
1174 edge exit; | |
1175 | |
1176 for (i = 0; VEC_iterate (edge, exits, i, exit); i++) | |
1177 if ((exit->flags & EDGE_ABNORMAL) | |
1178 && exit->dest == EXIT_BLOCK_PTR) | |
1179 ret = false; | |
1180 | |
1181 VEC_free (edge, heap, exits); | |
1182 } | |
1183 | |
1184 return ret; | |
1185 } | |
1186 | |
1187 /* Marks nontemporal stores in LOOP. GROUPS contains the description of memory | |
1188 references in the loop. */ | |
1189 | |
1190 static void | |
1191 mark_nontemporal_stores (struct loop *loop, struct mem_ref_group *groups) | |
1192 { | |
1193 struct mem_ref *ref; | |
1194 bool any = false; | |
1195 | |
1196 if (!may_use_storent_in_loop_p (loop)) | |
1197 return; | |
1198 | |
1199 for (; groups; groups = groups->next) | |
1200 for (ref = groups->refs; ref; ref = ref->next) | |
1201 any |= mark_nontemporal_store (ref); | |
1202 | |
1203 if (any && FENCE_FOLLOWING_MOVNT != NULL_TREE) | |
1204 emit_mfence_after_loop (loop); | |
1205 } | |
1206 | |
1207 /* Determines whether we can profitably unroll LOOP FACTOR times, and if | |
1208 this is the case, fill in DESC by the description of number of | |
1209 iterations. */ | |
1210 | |
1211 static bool | |
1212 should_unroll_loop_p (struct loop *loop, struct tree_niter_desc *desc, | |
1213 unsigned factor) | |
1214 { | |
1215 if (!can_unroll_loop_p (loop, factor, desc)) | |
1216 return false; | |
1217 | |
1218 /* We only consider loops without control flow for unrolling. This is not | |
1219 a hard restriction -- tree_unroll_loop works with arbitrary loops | |
1220 as well; but the unrolling/prefetching is usually more profitable for | |
1221 loops consisting of a single basic block, and we want to limit the | |
1222 code growth. */ | |
1223 if (loop->num_nodes > 2) | |
1224 return false; | |
1225 | |
1226 return true; | |
1227 } | |
1228 | |
1229 /* Determine the coefficient by that unroll LOOP, from the information | |
1230 contained in the list of memory references REFS. Description of | |
1231 umber of iterations of LOOP is stored to DESC. NINSNS is the number of | |
1232 insns of the LOOP. EST_NITER is the estimated number of iterations of | |
1233 the loop, or -1 if no estimate is available. */ | |
1234 | |
1235 static unsigned | |
1236 determine_unroll_factor (struct loop *loop, struct mem_ref_group *refs, | |
1237 unsigned ninsns, struct tree_niter_desc *desc, | |
1238 HOST_WIDE_INT est_niter) | |
1239 { | |
1240 unsigned upper_bound; | |
1241 unsigned nfactor, factor, mod_constraint; | |
1242 struct mem_ref_group *agp; | |
1243 struct mem_ref *ref; | |
1244 | |
1245 /* First check whether the loop is not too large to unroll. We ignore | |
1246 PARAM_MAX_UNROLL_TIMES, because for small loops, it prevented us | |
1247 from unrolling them enough to make exactly one cache line covered by each | |
1248 iteration. Also, the goal of PARAM_MAX_UNROLL_TIMES is to prevent | |
1249 us from unrolling the loops too many times in cases where we only expect | |
1250 gains from better scheduling and decreasing loop overhead, which is not | |
1251 the case here. */ | |
1252 upper_bound = PARAM_VALUE (PARAM_MAX_UNROLLED_INSNS) / ninsns; | |
1253 | |
1254 /* If we unrolled the loop more times than it iterates, the unrolled version | |
1255 of the loop would be never entered. */ | |
1256 if (est_niter >= 0 && est_niter < (HOST_WIDE_INT) upper_bound) | |
1257 upper_bound = est_niter; | |
1258 | |
1259 if (upper_bound <= 1) | |
1260 return 1; | |
1261 | |
1262 /* Choose the factor so that we may prefetch each cache just once, | |
1263 but bound the unrolling by UPPER_BOUND. */ | |
1264 factor = 1; | |
1265 for (agp = refs; agp; agp = agp->next) | |
1266 for (ref = agp->refs; ref; ref = ref->next) | |
1267 if (should_issue_prefetch_p (ref)) | |
1268 { | |
1269 mod_constraint = ref->prefetch_mod; | |
1270 nfactor = least_common_multiple (mod_constraint, factor); | |
1271 if (nfactor <= upper_bound) | |
1272 factor = nfactor; | |
1273 } | |
1274 | |
1275 if (!should_unroll_loop_p (loop, desc, factor)) | |
1276 return 1; | |
1277 | |
1278 return factor; | |
1279 } | |
1280 | |
1281 /* Returns the total volume of the memory references REFS, taking into account | |
1282 reuses in the innermost loop and cache line size. TODO -- we should also | |
1283 take into account reuses across the iterations of the loops in the loop | |
1284 nest. */ | |
1285 | |
1286 static unsigned | |
1287 volume_of_references (struct mem_ref_group *refs) | |
1288 { | |
1289 unsigned volume = 0; | |
1290 struct mem_ref_group *gr; | |
1291 struct mem_ref *ref; | |
1292 | |
1293 for (gr = refs; gr; gr = gr->next) | |
1294 for (ref = gr->refs; ref; ref = ref->next) | |
1295 { | |
1296 /* Almost always reuses another value? */ | |
1297 if (ref->prefetch_before != PREFETCH_ALL) | |
1298 continue; | |
1299 | |
1300 /* If several iterations access the same cache line, use the size of | |
1301 the line divided by this number. Otherwise, a cache line is | |
1302 accessed in each iteration. TODO -- in the latter case, we should | |
1303 take the size of the reference into account, rounding it up on cache | |
1304 line size multiple. */ | |
1305 volume += L1_CACHE_LINE_SIZE / ref->prefetch_mod; | |
1306 } | |
1307 return volume; | |
1308 } | |
1309 | |
1310 /* Returns the volume of memory references accessed across VEC iterations of | |
1311 loops, whose sizes are described in the LOOP_SIZES array. N is the number | |
1312 of the loops in the nest (length of VEC and LOOP_SIZES vectors). */ | |
1313 | |
1314 static unsigned | |
1315 volume_of_dist_vector (lambda_vector vec, unsigned *loop_sizes, unsigned n) | |
1316 { | |
1317 unsigned i; | |
1318 | |
1319 for (i = 0; i < n; i++) | |
1320 if (vec[i] != 0) | |
1321 break; | |
1322 | |
1323 if (i == n) | |
1324 return 0; | |
1325 | |
1326 gcc_assert (vec[i] > 0); | |
1327 | |
1328 /* We ignore the parts of the distance vector in subloops, since usually | |
1329 the numbers of iterations are much smaller. */ | |
1330 return loop_sizes[i] * vec[i]; | |
1331 } | |
1332 | |
1333 /* Add the steps of ACCESS_FN multiplied by STRIDE to the array STRIDE | |
1334 at the position corresponding to the loop of the step. N is the depth | |
1335 of the considered loop nest, and, LOOP is its innermost loop. */ | |
1336 | |
1337 static void | |
1338 add_subscript_strides (tree access_fn, unsigned stride, | |
1339 HOST_WIDE_INT *strides, unsigned n, struct loop *loop) | |
1340 { | |
1341 struct loop *aloop; | |
1342 tree step; | |
1343 HOST_WIDE_INT astep; | |
1344 unsigned min_depth = loop_depth (loop) - n; | |
1345 | |
1346 while (TREE_CODE (access_fn) == POLYNOMIAL_CHREC) | |
1347 { | |
1348 aloop = get_chrec_loop (access_fn); | |
1349 step = CHREC_RIGHT (access_fn); | |
1350 access_fn = CHREC_LEFT (access_fn); | |
1351 | |
1352 if ((unsigned) loop_depth (aloop) <= min_depth) | |
1353 continue; | |
1354 | |
1355 if (host_integerp (step, 0)) | |
1356 astep = tree_low_cst (step, 0); | |
1357 else | |
1358 astep = L1_CACHE_LINE_SIZE; | |
1359 | |
1360 strides[n - 1 - loop_depth (loop) + loop_depth (aloop)] += astep * stride; | |
1361 | |
1362 } | |
1363 } | |
1364 | |
1365 /* Returns the volume of memory references accessed between two consecutive | |
1366 self-reuses of the reference DR. We consider the subscripts of DR in N | |
1367 loops, and LOOP_SIZES contains the volumes of accesses in each of the | |
1368 loops. LOOP is the innermost loop of the current loop nest. */ | |
1369 | |
1370 static unsigned | |
1371 self_reuse_distance (data_reference_p dr, unsigned *loop_sizes, unsigned n, | |
1372 struct loop *loop) | |
1373 { | |
1374 tree stride, access_fn; | |
1375 HOST_WIDE_INT *strides, astride; | |
1376 VEC (tree, heap) *access_fns; | |
1377 tree ref = DR_REF (dr); | |
1378 unsigned i, ret = ~0u; | |
1379 | |
1380 /* In the following example: | |
1381 | |
1382 for (i = 0; i < N; i++) | |
1383 for (j = 0; j < N; j++) | |
1384 use (a[j][i]); | |
1385 the same cache line is accessed each N steps (except if the change from | |
1386 i to i + 1 crosses the boundary of the cache line). Thus, for self-reuse, | |
1387 we cannot rely purely on the results of the data dependence analysis. | |
1388 | |
1389 Instead, we compute the stride of the reference in each loop, and consider | |
1390 the innermost loop in that the stride is less than cache size. */ | |
1391 | |
1392 strides = XCNEWVEC (HOST_WIDE_INT, n); | |
1393 access_fns = DR_ACCESS_FNS (dr); | |
1394 | |
1395 for (i = 0; VEC_iterate (tree, access_fns, i, access_fn); i++) | |
1396 { | |
1397 /* Keep track of the reference corresponding to the subscript, so that we | |
1398 know its stride. */ | |
1399 while (handled_component_p (ref) && TREE_CODE (ref) != ARRAY_REF) | |
1400 ref = TREE_OPERAND (ref, 0); | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1401 |
0 | 1402 if (TREE_CODE (ref) == ARRAY_REF) |
1403 { | |
1404 stride = TYPE_SIZE_UNIT (TREE_TYPE (ref)); | |
1405 if (host_integerp (stride, 1)) | |
1406 astride = tree_low_cst (stride, 1); | |
1407 else | |
1408 astride = L1_CACHE_LINE_SIZE; | |
1409 | |
1410 ref = TREE_OPERAND (ref, 0); | |
1411 } | |
1412 else | |
1413 astride = 1; | |
1414 | |
1415 add_subscript_strides (access_fn, astride, strides, n, loop); | |
1416 } | |
1417 | |
1418 for (i = n; i-- > 0; ) | |
1419 { | |
1420 unsigned HOST_WIDE_INT s; | |
1421 | |
1422 s = strides[i] < 0 ? -strides[i] : strides[i]; | |
1423 | |
1424 if (s < (unsigned) L1_CACHE_LINE_SIZE | |
1425 && (loop_sizes[i] | |
1426 > (unsigned) (L1_CACHE_SIZE_BYTES / NONTEMPORAL_FRACTION))) | |
1427 { | |
1428 ret = loop_sizes[i]; | |
1429 break; | |
1430 } | |
1431 } | |
1432 | |
1433 free (strides); | |
1434 return ret; | |
1435 } | |
1436 | |
1437 /* Determines the distance till the first reuse of each reference in REFS | |
1438 in the loop nest of LOOP. NO_OTHER_REFS is true if there are no other | |
1439 memory references in the loop. */ | |
1440 | |
1441 static void | |
1442 determine_loop_nest_reuse (struct loop *loop, struct mem_ref_group *refs, | |
1443 bool no_other_refs) | |
1444 { | |
1445 struct loop *nest, *aloop; | |
1446 VEC (data_reference_p, heap) *datarefs = NULL; | |
1447 VEC (ddr_p, heap) *dependences = NULL; | |
1448 struct mem_ref_group *gr; | |
1449 struct mem_ref *ref, *refb; | |
1450 VEC (loop_p, heap) *vloops = NULL; | |
1451 unsigned *loop_data_size; | |
1452 unsigned i, j, n; | |
1453 unsigned volume, dist, adist; | |
1454 HOST_WIDE_INT vol; | |
1455 data_reference_p dr; | |
1456 ddr_p dep; | |
1457 | |
1458 if (loop->inner) | |
1459 return; | |
1460 | |
1461 /* Find the outermost loop of the loop nest of loop (we require that | |
1462 there are no sibling loops inside the nest). */ | |
1463 nest = loop; | |
1464 while (1) | |
1465 { | |
1466 aloop = loop_outer (nest); | |
1467 | |
1468 if (aloop == current_loops->tree_root | |
1469 || aloop->inner->next) | |
1470 break; | |
1471 | |
1472 nest = aloop; | |
1473 } | |
1474 | |
1475 /* For each loop, determine the amount of data accessed in each iteration. | |
1476 We use this to estimate whether the reference is evicted from the | |
1477 cache before its reuse. */ | |
1478 find_loop_nest (nest, &vloops); | |
1479 n = VEC_length (loop_p, vloops); | |
1480 loop_data_size = XNEWVEC (unsigned, n); | |
1481 volume = volume_of_references (refs); | |
1482 i = n; | |
1483 while (i-- != 0) | |
1484 { | |
1485 loop_data_size[i] = volume; | |
1486 /* Bound the volume by the L2 cache size, since above this bound, | |
1487 all dependence distances are equivalent. */ | |
1488 if (volume > L2_CACHE_SIZE_BYTES) | |
1489 continue; | |
1490 | |
1491 aloop = VEC_index (loop_p, vloops, i); | |
1492 vol = estimated_loop_iterations_int (aloop, false); | |
1493 if (vol < 0) | |
1494 vol = expected_loop_iterations (aloop); | |
1495 volume *= vol; | |
1496 } | |
1497 | |
1498 /* Prepare the references in the form suitable for data dependence | |
1499 analysis. We ignore unanalyzable data references (the results | |
1500 are used just as a heuristics to estimate temporality of the | |
1501 references, hence we do not need to worry about correctness). */ | |
1502 for (gr = refs; gr; gr = gr->next) | |
1503 for (ref = gr->refs; ref; ref = ref->next) | |
1504 { | |
1505 dr = create_data_ref (nest, ref->mem, ref->stmt, !ref->write_p); | |
1506 | |
1507 if (dr) | |
1508 { | |
1509 ref->reuse_distance = volume; | |
1510 dr->aux = ref; | |
1511 VEC_safe_push (data_reference_p, heap, datarefs, dr); | |
1512 } | |
1513 else | |
1514 no_other_refs = false; | |
1515 } | |
1516 | |
1517 for (i = 0; VEC_iterate (data_reference_p, datarefs, i, dr); i++) | |
1518 { | |
1519 dist = self_reuse_distance (dr, loop_data_size, n, loop); | |
1520 ref = (struct mem_ref *) dr->aux; | |
1521 if (ref->reuse_distance > dist) | |
1522 ref->reuse_distance = dist; | |
1523 | |
1524 if (no_other_refs) | |
1525 ref->independent_p = true; | |
1526 } | |
1527 | |
1528 compute_all_dependences (datarefs, &dependences, vloops, true); | |
1529 | |
1530 for (i = 0; VEC_iterate (ddr_p, dependences, i, dep); i++) | |
1531 { | |
1532 if (DDR_ARE_DEPENDENT (dep) == chrec_known) | |
1533 continue; | |
1534 | |
1535 ref = (struct mem_ref *) DDR_A (dep)->aux; | |
1536 refb = (struct mem_ref *) DDR_B (dep)->aux; | |
1537 | |
1538 if (DDR_ARE_DEPENDENT (dep) == chrec_dont_know | |
1539 || DDR_NUM_DIST_VECTS (dep) == 0) | |
1540 { | |
1541 /* If the dependence cannot be analyzed, assume that there might be | |
1542 a reuse. */ | |
1543 dist = 0; | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1544 |
0 | 1545 ref->independent_p = false; |
1546 refb->independent_p = false; | |
1547 } | |
1548 else | |
1549 { | |
1550 /* The distance vectors are normalized to be always lexicographically | |
1551 positive, hence we cannot tell just from them whether DDR_A comes | |
1552 before DDR_B or vice versa. However, it is not important, | |
1553 anyway -- if DDR_A is close to DDR_B, then it is either reused in | |
1554 DDR_B (and it is not nontemporal), or it reuses the value of DDR_B | |
1555 in cache (and marking it as nontemporal would not affect | |
1556 anything). */ | |
1557 | |
1558 dist = volume; | |
1559 for (j = 0; j < DDR_NUM_DIST_VECTS (dep); j++) | |
1560 { | |
1561 adist = volume_of_dist_vector (DDR_DIST_VECT (dep, j), | |
1562 loop_data_size, n); | |
1563 | |
1564 /* If this is a dependence in the innermost loop (i.e., the | |
1565 distances in all superloops are zero) and it is not | |
1566 the trivial self-dependence with distance zero, record that | |
1567 the references are not completely independent. */ | |
1568 if (lambda_vector_zerop (DDR_DIST_VECT (dep, j), n - 1) | |
1569 && (ref != refb | |
1570 || DDR_DIST_VECT (dep, j)[n-1] != 0)) | |
1571 { | |
1572 ref->independent_p = false; | |
1573 refb->independent_p = false; | |
1574 } | |
1575 | |
1576 /* Ignore accesses closer than | |
1577 L1_CACHE_SIZE_BYTES / NONTEMPORAL_FRACTION, | |
1578 so that we use nontemporal prefetches e.g. if single memory | |
1579 location is accessed several times in a single iteration of | |
1580 the loop. */ | |
1581 if (adist < L1_CACHE_SIZE_BYTES / NONTEMPORAL_FRACTION) | |
1582 continue; | |
1583 | |
1584 if (adist < dist) | |
1585 dist = adist; | |
1586 } | |
1587 } | |
1588 | |
1589 if (ref->reuse_distance > dist) | |
1590 ref->reuse_distance = dist; | |
1591 if (refb->reuse_distance > dist) | |
1592 refb->reuse_distance = dist; | |
1593 } | |
1594 | |
1595 free_dependence_relations (dependences); | |
1596 free_data_refs (datarefs); | |
1597 free (loop_data_size); | |
1598 | |
1599 if (dump_file && (dump_flags & TDF_DETAILS)) | |
1600 { | |
1601 fprintf (dump_file, "Reuse distances:\n"); | |
1602 for (gr = refs; gr; gr = gr->next) | |
1603 for (ref = gr->refs; ref; ref = ref->next) | |
1604 fprintf (dump_file, " ref %p distance %u\n", | |
1605 (void *) ref, ref->reuse_distance); | |
1606 } | |
1607 } | |
1608 | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1609 /* Do a cost-benefit analysis to determine if prefetching is profitable |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1610 for the current loop given the following parameters: |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1611 AHEAD: the iteration ahead distance, |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1612 EST_NITER: the estimated trip count, |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1613 NINSNS: estimated number of instructions in the loop, |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1614 PREFETCH_COUNT: an estimate of the number of prefetches |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1615 MEM_REF_COUNT: total number of memory references in the loop. */ |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1616 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1617 static bool |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1618 is_loop_prefetching_profitable (unsigned ahead, HOST_WIDE_INT est_niter, |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1619 unsigned ninsns, unsigned prefetch_count, |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1620 unsigned mem_ref_count, unsigned unroll_factor) |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1621 { |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1622 int insn_to_mem_ratio, insn_to_prefetch_ratio; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1623 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1624 if (mem_ref_count == 0) |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1625 return false; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1626 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1627 /* Prefetching improves performance by overlapping cache missing |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1628 memory accesses with CPU operations. If the loop does not have |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1629 enough CPU operations to overlap with memory operations, prefetching |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1630 won't give a significant benefit. One approximate way of checking |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1631 this is to require the ratio of instructions to memory references to |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1632 be above a certain limit. This approximation works well in practice. |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1633 TODO: Implement a more precise computation by estimating the time |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1634 for each CPU or memory op in the loop. Time estimates for memory ops |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1635 should account for cache misses. */ |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1636 insn_to_mem_ratio = ninsns / mem_ref_count; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1637 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1638 if (insn_to_mem_ratio < PREFETCH_MIN_INSN_TO_MEM_RATIO) |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1639 { |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1640 if (dump_file && (dump_flags & TDF_DETAILS)) |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1641 fprintf (dump_file, |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1642 "Not prefetching -- instruction to memory reference ratio (%d) too small\n", |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1643 insn_to_mem_ratio); |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1644 return false; |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1645 } |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1646 |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1647 /* Prefetching most likely causes performance degradation when the instruction |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1648 to prefetch ratio is too small. Too many prefetch instructions in a loop |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1649 may reduce the I-cache performance. |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1650 (unroll_factor * ninsns) is used to estimate the number of instructions in |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1651 the unrolled loop. This implementation is a bit simplistic -- the number |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1652 of issued prefetch instructions is also affected by unrolling. So, |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1653 prefetch_mod and the unroll factor should be taken into account when |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1654 determining prefetch_count. Also, the number of insns of the unrolled |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1655 loop will usually be significantly smaller than the number of insns of the |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1656 original loop * unroll_factor (at least the induction variable increases |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1657 and the exit branches will get eliminated), so it might be better to use |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1658 tree_estimate_loop_size + estimated_unrolled_size. */ |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1659 insn_to_prefetch_ratio = (unroll_factor * ninsns) / prefetch_count; |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1660 if (insn_to_prefetch_ratio < MIN_INSN_TO_PREFETCH_RATIO) |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1661 { |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1662 if (dump_file && (dump_flags & TDF_DETAILS)) |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1663 fprintf (dump_file, |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1664 "Not prefetching -- instruction to prefetch ratio (%d) too small\n", |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1665 insn_to_prefetch_ratio); |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1666 return false; |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1667 } |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1668 |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1669 /* Could not do further estimation if the trip count is unknown. Just assume |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1670 prefetching is profitable. Too aggressive??? */ |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1671 if (est_niter < 0) |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1672 return true; |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1673 |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1674 if (est_niter < (HOST_WIDE_INT) (TRIP_COUNT_TO_AHEAD_RATIO * ahead)) |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1675 { |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1676 if (dump_file && (dump_flags & TDF_DETAILS)) |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1677 fprintf (dump_file, |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1678 "Not prefetching -- loop estimated to roll only %d times\n", |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1679 (int) est_niter); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1680 return false; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1681 } |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1682 return true; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1683 } |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1684 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1685 |
0 | 1686 /* Issue prefetch instructions for array references in LOOP. Returns |
1687 true if the LOOP was unrolled. */ | |
1688 | |
1689 static bool | |
1690 loop_prefetch_arrays (struct loop *loop) | |
1691 { | |
1692 struct mem_ref_group *refs; | |
1693 unsigned ahead, ninsns, time, unroll_factor; | |
1694 HOST_WIDE_INT est_niter; | |
1695 struct tree_niter_desc desc; | |
1696 bool unrolled = false, no_other_refs; | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1697 unsigned prefetch_count; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1698 unsigned mem_ref_count; |
0 | 1699 |
1700 if (optimize_loop_nest_for_size_p (loop)) | |
1701 { | |
1702 if (dump_file && (dump_flags & TDF_DETAILS)) | |
1703 fprintf (dump_file, " ignored (cold area)\n"); | |
1704 return false; | |
1705 } | |
1706 | |
1707 /* Step 1: gather the memory references. */ | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1708 refs = gather_memory_references (loop, &no_other_refs, &mem_ref_count); |
0 | 1709 |
1710 /* Step 2: estimate the reuse effects. */ | |
1711 prune_by_reuse (refs); | |
1712 | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1713 prefetch_count = estimate_prefetch_count (refs); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1714 if (prefetch_count == 0) |
0 | 1715 goto fail; |
1716 | |
1717 determine_loop_nest_reuse (loop, refs, no_other_refs); | |
1718 | |
1719 /* Step 3: determine the ahead and unroll factor. */ | |
1720 | |
1721 /* FIXME: the time should be weighted by the probabilities of the blocks in | |
1722 the loop body. */ | |
1723 time = tree_num_loop_insns (loop, &eni_time_weights); | |
1724 ahead = (PREFETCH_LATENCY + time - 1) / time; | |
1725 est_niter = estimated_loop_iterations_int (loop, false); | |
1726 | |
1727 ninsns = tree_num_loop_insns (loop, &eni_size_weights); | |
1728 unroll_factor = determine_unroll_factor (loop, refs, ninsns, &desc, | |
1729 est_niter); | |
1730 if (dump_file && (dump_flags & TDF_DETAILS)) | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1731 fprintf (dump_file, "Ahead %d, unroll factor %d, trip count " |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1732 HOST_WIDE_INT_PRINT_DEC "\n" |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1733 "insn count %d, mem ref count %d, prefetch count %d\n", |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1734 ahead, unroll_factor, est_niter, |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1735 ninsns, mem_ref_count, prefetch_count); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1736 |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1737 if (!is_loop_prefetching_profitable (ahead, est_niter, ninsns, prefetch_count, |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1738 mem_ref_count, unroll_factor)) |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1739 goto fail; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1740 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1741 mark_nontemporal_stores (loop, refs); |
0 | 1742 |
1743 /* Step 4: what to prefetch? */ | |
1744 if (!schedule_prefetches (refs, unroll_factor, ahead)) | |
1745 goto fail; | |
1746 | |
1747 /* Step 5: unroll the loop. TODO -- peeling of first and last few | |
1748 iterations so that we do not issue superfluous prefetches. */ | |
1749 if (unroll_factor != 1) | |
1750 { | |
1751 tree_unroll_loop (loop, unroll_factor, | |
1752 single_dom_exit (loop), &desc); | |
1753 unrolled = true; | |
1754 } | |
1755 | |
1756 /* Step 6: issue the prefetches. */ | |
1757 issue_prefetches (refs, unroll_factor, ahead); | |
1758 | |
1759 fail: | |
1760 release_mem_refs (refs); | |
1761 return unrolled; | |
1762 } | |
1763 | |
1764 /* Issue prefetch instructions for array references in loops. */ | |
1765 | |
1766 unsigned int | |
1767 tree_ssa_prefetch_arrays (void) | |
1768 { | |
1769 loop_iterator li; | |
1770 struct loop *loop; | |
1771 bool unrolled = false; | |
1772 int todo_flags = 0; | |
1773 | |
1774 if (!HAVE_prefetch | |
1775 /* It is possible to ask compiler for say -mtune=i486 -march=pentium4. | |
1776 -mtune=i486 causes us having PREFETCH_BLOCK 0, since this is part | |
1777 of processor costs and i486 does not have prefetch, but | |
1778 -march=pentium4 causes HAVE_prefetch to be true. Ugh. */ | |
1779 || PREFETCH_BLOCK == 0) | |
1780 return 0; | |
1781 | |
1782 if (dump_file && (dump_flags & TDF_DETAILS)) | |
1783 { | |
1784 fprintf (dump_file, "Prefetching parameters:\n"); | |
1785 fprintf (dump_file, " simultaneous prefetches: %d\n", | |
1786 SIMULTANEOUS_PREFETCHES); | |
1787 fprintf (dump_file, " prefetch latency: %d\n", PREFETCH_LATENCY); | |
1788 fprintf (dump_file, " prefetch block size: %d\n", PREFETCH_BLOCK); | |
1789 fprintf (dump_file, " L1 cache size: %d lines, %d kB\n", | |
1790 L1_CACHE_SIZE_BYTES / L1_CACHE_LINE_SIZE, L1_CACHE_SIZE); | |
1791 fprintf (dump_file, " L1 cache line size: %d\n", L1_CACHE_LINE_SIZE); | |
1792 fprintf (dump_file, " L2 cache size: %d kB\n", L2_CACHE_SIZE); | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1793 fprintf (dump_file, " min insn-to-prefetch ratio: %d \n", |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1794 MIN_INSN_TO_PREFETCH_RATIO); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1795 fprintf (dump_file, " min insn-to-mem ratio: %d \n", |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1796 PREFETCH_MIN_INSN_TO_MEM_RATIO); |
0 | 1797 fprintf (dump_file, "\n"); |
1798 } | |
1799 | |
1800 initialize_original_copy_tables (); | |
1801 | |
1802 if (!built_in_decls[BUILT_IN_PREFETCH]) | |
1803 { | |
1804 tree type = build_function_type (void_type_node, | |
1805 tree_cons (NULL_TREE, | |
1806 const_ptr_type_node, | |
1807 NULL_TREE)); | |
1808 tree decl = add_builtin_function ("__builtin_prefetch", type, | |
1809 BUILT_IN_PREFETCH, BUILT_IN_NORMAL, | |
1810 NULL, NULL_TREE); | |
1811 DECL_IS_NOVOPS (decl) = true; | |
1812 built_in_decls[BUILT_IN_PREFETCH] = decl; | |
1813 } | |
1814 | |
1815 /* We assume that size of cache line is a power of two, so verify this | |
1816 here. */ | |
1817 gcc_assert ((PREFETCH_BLOCK & (PREFETCH_BLOCK - 1)) == 0); | |
1818 | |
1819 FOR_EACH_LOOP (li, loop, LI_FROM_INNERMOST) | |
1820 { | |
1821 if (dump_file && (dump_flags & TDF_DETAILS)) | |
1822 fprintf (dump_file, "Processing loop %d:\n", loop->num); | |
1823 | |
1824 unrolled |= loop_prefetch_arrays (loop); | |
1825 | |
1826 if (dump_file && (dump_flags & TDF_DETAILS)) | |
1827 fprintf (dump_file, "\n\n"); | |
1828 } | |
1829 | |
1830 if (unrolled) | |
1831 { | |
1832 scev_reset (); | |
1833 todo_flags |= TODO_cleanup_cfg; | |
1834 } | |
1835 | |
1836 free_original_copy_tables (); | |
1837 return todo_flags; | |
1838 } |