Mercurial > hg > CbC > CbC_gcc
annotate gcc/tree-ssa-loop-prefetch.c @ 67:f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
author | nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp> |
---|---|
date | Tue, 22 Mar 2011 17:18:12 +0900 |
parents | b7f97abdc517 |
children | 04ced10e8804 |
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" | |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
28 #include "tree-pretty-print.h" |
0 | 29 #include "tree-flow.h" |
30 #include "tree-dump.h" | |
31 #include "timevar.h" | |
32 #include "cfgloop.h" | |
33 #include "tree-pass.h" | |
34 #include "insn-config.h" | |
35 #include "recog.h" | |
36 #include "hashtab.h" | |
37 #include "tree-chrec.h" | |
38 #include "tree-scalar-evolution.h" | |
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
39 #include "diagnostic-core.h" |
0 | 40 #include "params.h" |
41 #include "langhooks.h" | |
42 #include "tree-inline.h" | |
43 #include "tree-data-ref.h" | |
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
44 |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
45 |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
46 /* FIXME: Needed for optabs, but this should all be moved to a TBD interface |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
47 between the GIMPLE and RTL worlds. */ |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
48 #include "expr.h" |
0 | 49 #include "optabs.h" |
50 | |
51 /* This pass inserts prefetch instructions to optimize cache usage during | |
52 accesses to arrays in loops. It processes loops sequentially and: | |
53 | |
54 1) Gathers all memory references in the single loop. | |
55 2) For each of the references it decides when it is profitable to prefetch | |
56 it. To do it, we evaluate the reuse among the accesses, and determines | |
57 two values: PREFETCH_BEFORE (meaning that it only makes sense to do | |
58 prefetching in the first PREFETCH_BEFORE iterations of the loop) and | |
59 PREFETCH_MOD (meaning that it only makes sense to prefetch in the | |
60 iterations of the loop that are zero modulo PREFETCH_MOD). For example | |
61 (assuming cache line size is 64 bytes, char has size 1 byte and there | |
62 is no hardware sequential prefetch): | |
63 | |
64 char *a; | |
65 for (i = 0; i < max; i++) | |
66 { | |
67 a[255] = ...; (0) | |
68 a[i] = ...; (1) | |
69 a[i + 64] = ...; (2) | |
70 a[16*i] = ...; (3) | |
71 a[187*i] = ...; (4) | |
72 a[187*i + 50] = ...; (5) | |
73 } | |
74 | |
75 (0) obviously has PREFETCH_BEFORE 1 | |
76 (1) has PREFETCH_BEFORE 64, since (2) accesses the same memory | |
77 location 64 iterations before it, and PREFETCH_MOD 64 (since | |
78 it hits the same cache line otherwise). | |
79 (2) has PREFETCH_MOD 64 | |
80 (3) has PREFETCH_MOD 4 | |
81 (4) has PREFETCH_MOD 1. We do not set PREFETCH_BEFORE here, since | |
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
82 the cache line accessed by (5) is the same with probability only |
0 | 83 7/32. |
84 (5) has PREFETCH_MOD 1 as well. | |
85 | |
86 Additionally, we use data dependence analysis to determine for each | |
87 reference the distance till the first reuse; this information is used | |
88 to determine the temporality of the issued prefetch instruction. | |
89 | |
90 3) We determine how much ahead we need to prefetch. The number of | |
91 iterations needed is time to fetch / time spent in one iteration of | |
92 the loop. The problem is that we do not know either of these values, | |
93 so we just make a heuristic guess based on a magic (possibly) | |
94 target-specific constant and size of the loop. | |
95 | |
96 4) Determine which of the references we prefetch. We take into account | |
97 that there is a maximum number of simultaneous prefetches (provided | |
98 by machine description). We prefetch as many prefetches as possible | |
99 while still within this bound (starting with those with lowest | |
100 prefetch_mod, since they are responsible for most of the cache | |
101 misses). | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
102 |
0 | 103 5) We unroll and peel loops so that we are able to satisfy PREFETCH_MOD |
104 and PREFETCH_BEFORE requirements (within some bounds), and to avoid | |
105 prefetching nonaccessed memory. | |
106 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
|
107 |
0 | 108 6) We actually emit the prefetch instructions. ??? Perhaps emit the |
109 prefetch instructions with guards in cases where 5) was not sufficient | |
110 to satisfy the constraints? | |
111 | |
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
112 A cost model is implemented to determine whether or not prefetching is |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
113 profitable for a given loop. The cost model has three heuristics: |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
114 |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
115 1. Function trip_count_to_ahead_ratio_too_small_p implements a |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
116 heuristic that determines whether or not the loop has too few |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
117 iterations (compared to ahead). Prefetching is not likely to be |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
118 beneficial if the trip count to ahead ratio is below a certain |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
119 minimum. |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
120 |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
121 2. Function mem_ref_count_reasonable_p implements a heuristic that |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
122 determines whether the given loop has enough CPU ops that can be |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
123 overlapped with cache missing memory ops. If not, the loop |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
124 won't benefit from prefetching. In the implementation, |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
125 prefetching is not considered beneficial if the ratio between |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
126 the instruction count and the mem ref count is below a certain |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
127 minimum. |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
128 |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
129 3. Function insn_to_prefetch_ratio_too_small_p implements a |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
130 heuristic that disables prefetching in a loop if the prefetching |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
131 cost is above a certain limit. The relative prefetching cost is |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
132 estimated by taking the ratio between the prefetch count and the |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
133 total intruction count (this models the I-cache cost). |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
134 |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
135 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
|
136 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
|
137 added later. |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
138 |
0 | 139 Some other TODO: |
140 -- write and use more general reuse analysis (that could be also used | |
141 in other cache aimed loop optimizations) | |
142 -- make it behave sanely together with the prefetches given by user | |
143 (now we just ignore them; at the very least we should avoid | |
144 optimizing loops in that user put his own prefetches) | |
145 -- we assume cache line size alignment of arrays; this could be | |
146 improved. */ | |
147 | |
148 /* Magic constants follow. These should be replaced by machine specific | |
149 numbers. */ | |
150 | |
151 /* True if write can be prefetched by a read prefetch. */ | |
152 | |
153 #ifndef WRITE_CAN_USE_READ_PREFETCH | |
154 #define WRITE_CAN_USE_READ_PREFETCH 1 | |
155 #endif | |
156 | |
157 /* True if read can be prefetched by a write prefetch. */ | |
158 | |
159 #ifndef READ_CAN_USE_WRITE_PREFETCH | |
160 #define READ_CAN_USE_WRITE_PREFETCH 0 | |
161 #endif | |
162 | |
163 /* The size of the block loaded by a single prefetch. Usually, this is | |
164 the same as cache line size (at the moment, we only consider one level | |
165 of cache hierarchy). */ | |
166 | |
167 #ifndef PREFETCH_BLOCK | |
168 #define PREFETCH_BLOCK L1_CACHE_LINE_SIZE | |
169 #endif | |
170 | |
171 /* Do we have a forward hardware sequential prefetching? */ | |
172 | |
173 #ifndef HAVE_FORWARD_PREFETCH | |
174 #define HAVE_FORWARD_PREFETCH 0 | |
175 #endif | |
176 | |
177 /* Do we have a backward hardware sequential prefetching? */ | |
178 | |
179 #ifndef HAVE_BACKWARD_PREFETCH | |
180 #define HAVE_BACKWARD_PREFETCH 0 | |
181 #endif | |
182 | |
183 /* In some cases we are only able to determine that there is a certain | |
184 probability that the two accesses hit the same cache line. In this | |
185 case, we issue the prefetches for both of them if this probability | |
186 is less then (1000 - ACCEPTABLE_MISS_RATE) per thousand. */ | |
187 | |
188 #ifndef ACCEPTABLE_MISS_RATE | |
189 #define ACCEPTABLE_MISS_RATE 50 | |
190 #endif | |
191 | |
192 #ifndef HAVE_prefetch | |
193 #define HAVE_prefetch 0 | |
194 #endif | |
195 | |
196 #define L1_CACHE_SIZE_BYTES ((unsigned) (L1_CACHE_SIZE * 1024)) | |
197 #define L2_CACHE_SIZE_BYTES ((unsigned) (L2_CACHE_SIZE * 1024)) | |
198 | |
199 /* We consider a memory access nontemporal if it is not reused sooner than | |
200 after L2_CACHE_SIZE_BYTES of memory are accessed. However, we ignore | |
201 accesses closer than L1_CACHE_SIZE_BYTES / NONTEMPORAL_FRACTION, | |
202 so that we use nontemporal prefetches e.g. if single memory location | |
203 is accessed several times in a single iteration of the loop. */ | |
204 #define NONTEMPORAL_FRACTION 16 | |
205 | |
206 /* In case we have to emit a memory fence instruction after the loop that | |
207 uses nontemporal stores, this defines the builtin to use. */ | |
208 | |
209 #ifndef FENCE_FOLLOWING_MOVNT | |
210 #define FENCE_FOLLOWING_MOVNT NULL_TREE | |
211 #endif | |
212 | |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
213 /* 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
|
214 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
|
215 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
|
216 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
|
217 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
|
218 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
|
219 prefetching. */ |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
220 |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
221 #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
|
222 #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
|
223 #endif |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
224 |
0 | 225 /* The group of references between that reuse may occur. */ |
226 | |
227 struct mem_ref_group | |
228 { | |
229 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
|
230 tree step; /* Step of the reference. */ |
0 | 231 struct mem_ref *refs; /* References in the group. */ |
232 struct mem_ref_group *next; /* Next group of references. */ | |
233 }; | |
234 | |
235 /* Assigned to PREFETCH_BEFORE when all iterations are to be prefetched. */ | |
236 | |
237 #define PREFETCH_ALL (~(unsigned HOST_WIDE_INT) 0) | |
238 | |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
239 /* 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
|
240 than what is required by the prefetch. This is to avoid redundant |
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
241 prefetches. For example, when prefetch_mod is 16 and unroll_factor is |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
242 2, prefetching requires unrolling the loop 16 times, but |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
243 the loop is actually unrolled twice. In this case (ratio = 8), |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
244 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
|
245 |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
246 #ifndef PREFETCH_MOD_TO_UNROLL_FACTOR_RATIO |
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
247 #define PREFETCH_MOD_TO_UNROLL_FACTOR_RATIO 4 |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
248 #endif |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
249 |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
250 /* Some of the prefetch computations have quadratic complexity. We want to |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
251 avoid huge compile times and, therefore, want to limit the amount of |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
252 memory references per loop where we consider prefetching. */ |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
253 |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
254 #ifndef PREFETCH_MAX_MEM_REFS_PER_LOOP |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
255 #define PREFETCH_MAX_MEM_REFS_PER_LOOP 200 |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
256 #endif |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
257 |
0 | 258 /* The memory reference. */ |
259 | |
260 struct mem_ref | |
261 { | |
262 gimple stmt; /* Statement in that the reference appears. */ | |
263 tree mem; /* The reference. */ | |
264 HOST_WIDE_INT delta; /* Constant offset of the reference. */ | |
265 struct mem_ref_group *group; /* The group of references it belongs to. */ | |
266 unsigned HOST_WIDE_INT prefetch_mod; | |
267 /* Prefetch only each PREFETCH_MOD-th | |
268 iteration. */ | |
269 unsigned HOST_WIDE_INT prefetch_before; | |
270 /* Prefetch only first PREFETCH_BEFORE | |
271 iterations. */ | |
272 unsigned reuse_distance; /* The amount of data accessed before the first | |
273 reuse of this value. */ | |
274 struct mem_ref *next; /* The next reference in the group. */ | |
275 unsigned write_p : 1; /* Is it a write? */ | |
276 unsigned independent_p : 1; /* True if the reference is independent on | |
277 all other references inside the loop. */ | |
278 unsigned issue_prefetch_p : 1; /* Should we really issue the prefetch? */ | |
279 unsigned storent_p : 1; /* True if we changed the store to a | |
280 nontemporal one. */ | |
281 }; | |
282 | |
283 /* Dumps information about reference REF to FILE. */ | |
284 | |
285 static void | |
286 dump_mem_ref (FILE *file, struct mem_ref *ref) | |
287 { | |
288 fprintf (file, "Reference %p:\n", (void *) ref); | |
289 | |
290 fprintf (file, " group %p (base ", (void *) ref->group); | |
291 print_generic_expr (file, ref->group->base, TDF_SLIM); | |
292 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
|
293 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
|
294 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
|
295 else |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
296 print_generic_expr (file, ref->group->step, TDF_TREE); |
0 | 297 fprintf (file, ")\n"); |
298 | |
299 fprintf (file, " delta "); | |
300 fprintf (file, HOST_WIDE_INT_PRINT_DEC, ref->delta); | |
301 fprintf (file, "\n"); | |
302 | |
303 fprintf (file, " %s\n", ref->write_p ? "write" : "read"); | |
304 | |
305 fprintf (file, "\n"); | |
306 } | |
307 | |
308 /* Finds a group with BASE and STEP in GROUPS, or creates one if it does not | |
309 exist. */ | |
310 | |
311 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
|
312 find_or_create_group (struct mem_ref_group **groups, tree base, tree step) |
0 | 313 { |
314 struct mem_ref_group *group; | |
315 | |
316 for (; *groups; groups = &(*groups)->next) | |
317 { | |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
318 if (operand_equal_p ((*groups)->step, step, 0) |
0 | 319 && operand_equal_p ((*groups)->base, base, 0)) |
320 return *groups; | |
321 | |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
322 /* 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
|
323 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
|
324 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
|
325 && int_cst_value ((*groups)->step) < int_cst_value (step)) |
0 | 326 break; |
327 } | |
328 | |
329 group = XNEW (struct mem_ref_group); | |
330 group->base = base; | |
331 group->step = step; | |
332 group->refs = NULL; | |
333 group->next = *groups; | |
334 *groups = group; | |
335 | |
336 return group; | |
337 } | |
338 | |
339 /* Records a memory reference MEM in GROUP with offset DELTA and write status | |
340 WRITE_P. The reference occurs in statement STMT. */ | |
341 | |
342 static void | |
343 record_ref (struct mem_ref_group *group, gimple stmt, tree mem, | |
344 HOST_WIDE_INT delta, bool write_p) | |
345 { | |
346 struct mem_ref **aref; | |
347 | |
348 /* Do not record the same address twice. */ | |
349 for (aref = &group->refs; *aref; aref = &(*aref)->next) | |
350 { | |
351 /* It does not have to be possible for write reference to reuse the read | |
352 prefetch, or vice versa. */ | |
353 if (!WRITE_CAN_USE_READ_PREFETCH | |
354 && write_p | |
355 && !(*aref)->write_p) | |
356 continue; | |
357 if (!READ_CAN_USE_WRITE_PREFETCH | |
358 && !write_p | |
359 && (*aref)->write_p) | |
360 continue; | |
361 | |
362 if ((*aref)->delta == delta) | |
363 return; | |
364 } | |
365 | |
366 (*aref) = XNEW (struct mem_ref); | |
367 (*aref)->stmt = stmt; | |
368 (*aref)->mem = mem; | |
369 (*aref)->delta = delta; | |
370 (*aref)->write_p = write_p; | |
371 (*aref)->prefetch_before = PREFETCH_ALL; | |
372 (*aref)->prefetch_mod = 1; | |
373 (*aref)->reuse_distance = 0; | |
374 (*aref)->issue_prefetch_p = false; | |
375 (*aref)->group = group; | |
376 (*aref)->next = NULL; | |
377 (*aref)->independent_p = false; | |
378 (*aref)->storent_p = false; | |
379 | |
380 if (dump_file && (dump_flags & TDF_DETAILS)) | |
381 dump_mem_ref (dump_file, *aref); | |
382 } | |
383 | |
384 /* Release memory references in GROUPS. */ | |
385 | |
386 static void | |
387 release_mem_refs (struct mem_ref_group *groups) | |
388 { | |
389 struct mem_ref_group *next_g; | |
390 struct mem_ref *ref, *next_r; | |
391 | |
392 for (; groups; groups = next_g) | |
393 { | |
394 next_g = groups->next; | |
395 for (ref = groups->refs; ref; ref = next_r) | |
396 { | |
397 next_r = ref->next; | |
398 free (ref); | |
399 } | |
400 free (groups); | |
401 } | |
402 } | |
403 | |
404 /* A structure used to pass arguments to idx_analyze_ref. */ | |
405 | |
406 struct ar_data | |
407 { | |
408 struct loop *loop; /* Loop of the reference. */ | |
409 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
|
410 tree *step; /* Step of the memory reference. */ |
0 | 411 HOST_WIDE_INT *delta; /* Offset of the memory reference. */ |
412 }; | |
413 | |
414 /* Analyzes a single INDEX of a memory reference to obtain information | |
415 described at analyze_ref. Callback for for_each_index. */ | |
416 | |
417 static bool | |
418 idx_analyze_ref (tree base, tree *index, void *data) | |
419 { | |
420 struct ar_data *ar_data = (struct ar_data *) data; | |
421 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
|
422 HOST_WIDE_INT idelta = 0, imult = 1; |
0 | 423 affine_iv iv; |
424 | |
425 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
|
426 *index, &iv, true)) |
0 | 427 return false; |
428 ibase = iv.base; | |
429 step = iv.step; | |
430 | |
431 if (TREE_CODE (ibase) == POINTER_PLUS_EXPR | |
432 && cst_and_fits_in_hwi (TREE_OPERAND (ibase, 1))) | |
433 { | |
434 idelta = int_cst_value (TREE_OPERAND (ibase, 1)); | |
435 ibase = TREE_OPERAND (ibase, 0); | |
436 } | |
437 if (cst_and_fits_in_hwi (ibase)) | |
438 { | |
439 idelta += int_cst_value (ibase); | |
440 ibase = build_int_cst (TREE_TYPE (ibase), 0); | |
441 } | |
442 | |
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
443 if (TREE_CODE (base) == ARRAY_REF) |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
444 { |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
445 stepsize = array_ref_element_size (base); |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
446 if (!cst_and_fits_in_hwi (stepsize)) |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
447 return false; |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
448 imult = int_cst_value (stepsize); |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
449 step = fold_build2 (MULT_EXPR, sizetype, |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
450 fold_convert (sizetype, step), |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
451 fold_convert (sizetype, stepsize)); |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
452 idelta *= imult; |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
453 } |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
454 |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
455 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
|
456 *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
|
457 else |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
458 *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
|
459 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
|
460 fold_convert (sizetype, step)); |
0 | 461 *ar_data->delta += idelta; |
462 *index = ibase; | |
463 | |
464 return true; | |
465 } | |
466 | |
467 /* Tries to express REF_P in shape &BASE + STEP * iter + DELTA, where DELTA and | |
468 STEP are integer constants and iter is number of iterations of LOOP. The | |
469 reference occurs in statement STMT. Strips nonaddressable component | |
470 references from REF_P. */ | |
471 | |
472 static bool | |
473 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
|
474 tree *step, HOST_WIDE_INT *delta, |
0 | 475 gimple stmt) |
476 { | |
477 struct ar_data ar_data; | |
478 tree off; | |
479 HOST_WIDE_INT bit_offset; | |
480 tree ref = *ref_p; | |
481 | |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
482 *step = NULL_TREE; |
0 | 483 *delta = 0; |
484 | |
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
485 /* First strip off the component references. Ignore bitfields. |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
486 Also strip off the real and imagine parts of a complex, so that |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
487 they can have the same base. */ |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
488 if (TREE_CODE (ref) == REALPART_EXPR |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
489 || TREE_CODE (ref) == IMAGPART_EXPR |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
490 || (TREE_CODE (ref) == COMPONENT_REF |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
491 && DECL_NONADDRESSABLE_P (TREE_OPERAND (ref, 1)))) |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
492 { |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
493 if (TREE_CODE (ref) == IMAGPART_EXPR) |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
494 *delta += int_size_in_bytes (TREE_TYPE (ref)); |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
495 ref = TREE_OPERAND (ref, 0); |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
496 } |
0 | 497 |
498 *ref_p = ref; | |
499 | |
500 for (; TREE_CODE (ref) == COMPONENT_REF; ref = TREE_OPERAND (ref, 0)) | |
501 { | |
502 off = DECL_FIELD_BIT_OFFSET (TREE_OPERAND (ref, 1)); | |
503 bit_offset = TREE_INT_CST_LOW (off); | |
504 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
|
505 |
0 | 506 *delta += bit_offset / BITS_PER_UNIT; |
507 } | |
508 | |
509 *base = unshare_expr (ref); | |
510 ar_data.loop = loop; | |
511 ar_data.stmt = stmt; | |
512 ar_data.step = step; | |
513 ar_data.delta = delta; | |
514 return for_each_index (base, idx_analyze_ref, &ar_data); | |
515 } | |
516 | |
517 /* Record a memory reference REF to the list REFS. The reference occurs in | |
518 LOOP in statement STMT and it is write if WRITE_P. Returns true if the | |
519 reference was recorded, false otherwise. */ | |
520 | |
521 static bool | |
522 gather_memory_references_ref (struct loop *loop, struct mem_ref_group **refs, | |
523 tree ref, bool write_p, gimple stmt) | |
524 { | |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
525 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
|
526 HOST_WIDE_INT delta; |
0 | 527 struct mem_ref_group *agrp; |
528 | |
529 if (get_base_address (ref) == NULL) | |
530 return false; | |
531 | |
532 if (!analyze_ref (loop, &ref, &base, &step, &delta, stmt)) | |
533 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
|
534 /* 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
|
535 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
|
536 return false; |
0 | 537 |
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
538 /* Stop if the address of BASE could not be taken. */ |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
539 if (may_be_nonaddressable_p (base)) |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
540 return false; |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
541 |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
542 /* Limit non-constant step prefetching only to the innermost loops. */ |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
543 if (!cst_and_fits_in_hwi (step) && loop->inner != NULL) |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
544 return false; |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
545 |
0 | 546 /* Now we know that REF = &BASE + STEP * iter + DELTA, where DELTA and STEP |
547 are integer constants. */ | |
548 agrp = find_or_create_group (refs, base, step); | |
549 record_ref (agrp, stmt, ref, delta, write_p); | |
550 | |
551 return true; | |
552 } | |
553 | |
554 /* Record the suitable memory references in LOOP. NO_OTHER_REFS is set to | |
555 true if there are no other memory references inside the loop. */ | |
556 | |
557 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
|
558 gather_memory_references (struct loop *loop, bool *no_other_refs, unsigned *ref_count) |
0 | 559 { |
560 basic_block *body = get_loop_body_in_dom_order (loop); | |
561 basic_block bb; | |
562 unsigned i; | |
563 gimple_stmt_iterator bsi; | |
564 gimple stmt; | |
565 tree lhs, rhs; | |
566 struct mem_ref_group *refs = NULL; | |
567 | |
568 *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
|
569 *ref_count = 0; |
0 | 570 |
571 /* Scan the loop body in order, so that the former references precede the | |
572 later ones. */ | |
573 for (i = 0; i < loop->num_nodes; i++) | |
574 { | |
575 bb = body[i]; | |
576 if (bb->loop_father != loop) | |
577 continue; | |
578 | |
579 for (bsi = gsi_start_bb (bb); !gsi_end_p (bsi); gsi_next (&bsi)) | |
580 { | |
581 stmt = gsi_stmt (bsi); | |
582 | |
583 if (gimple_code (stmt) != GIMPLE_ASSIGN) | |
584 { | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
585 if (gimple_vuse (stmt) |
0 | 586 || (is_gimple_call (stmt) |
587 && !(gimple_call_flags (stmt) & ECF_CONST))) | |
588 *no_other_refs = false; | |
589 continue; | |
590 } | |
591 | |
592 lhs = gimple_assign_lhs (stmt); | |
593 rhs = gimple_assign_rhs1 (stmt); | |
594 | |
595 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
|
596 { |
0 | 597 *no_other_refs &= gather_memory_references_ref (loop, &refs, |
598 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
|
599 *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
|
600 } |
0 | 601 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
|
602 { |
0 | 603 *no_other_refs &= gather_memory_references_ref (loop, &refs, |
604 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
|
605 *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
|
606 } |
0 | 607 } |
608 } | |
609 free (body); | |
610 | |
611 return refs; | |
612 } | |
613 | |
614 /* Prune the prefetch candidate REF using the self-reuse. */ | |
615 | |
616 static void | |
617 prune_ref_by_self_reuse (struct mem_ref *ref) | |
618 { | |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
619 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
|
620 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
|
621 |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
622 /* 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
|
623 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
|
624 return; |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
625 |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
626 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
|
627 |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
628 backward = step < 0; |
0 | 629 |
630 if (step == 0) | |
631 { | |
632 /* Prefetch references to invariant address just once. */ | |
633 ref->prefetch_before = 1; | |
634 return; | |
635 } | |
636 | |
637 if (backward) | |
638 step = -step; | |
639 | |
640 if (step > PREFETCH_BLOCK) | |
641 return; | |
642 | |
643 if ((backward && HAVE_BACKWARD_PREFETCH) | |
644 || (!backward && HAVE_FORWARD_PREFETCH)) | |
645 { | |
646 ref->prefetch_before = 1; | |
647 return; | |
648 } | |
649 | |
650 ref->prefetch_mod = PREFETCH_BLOCK / step; | |
651 } | |
652 | |
653 /* Divides X by BY, rounding down. */ | |
654 | |
655 static HOST_WIDE_INT | |
656 ddown (HOST_WIDE_INT x, unsigned HOST_WIDE_INT by) | |
657 { | |
658 gcc_assert (by > 0); | |
659 | |
660 if (x >= 0) | |
661 return x / by; | |
662 else | |
663 return (x + by - 1) / by; | |
664 } | |
665 | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
666 /* 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
|
667 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
|
668 difference DELTA, compute the probability that they will fall |
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
669 in different cache lines. Return true if the computed miss rate |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
670 is not greater than the ACCEPTABLE_MISS_RATE. DISTINCT_ITERS is the |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
671 number of distinct iterations after which the pattern repeats itself. |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
672 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
|
673 |
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
674 static bool |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
675 is_miss_rate_acceptable (unsigned HOST_WIDE_INT cache_line_size, |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
676 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
|
677 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
|
678 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
|
679 { |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
680 unsigned align, iter; |
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
681 int total_positions, miss_positions, max_allowed_miss_positions; |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
682 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
|
683 |
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
684 /* It always misses if delta is greater than or equal to the cache |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
685 line size. */ |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
686 if (delta >= (HOST_WIDE_INT) cache_line_size) |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
687 return false; |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
688 |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
689 miss_positions = 0; |
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
690 total_positions = (cache_line_size / align_unit) * distinct_iters; |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
691 max_allowed_miss_positions = (ACCEPTABLE_MISS_RATE * total_positions) / 1000; |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
692 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
693 /* 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
|
694 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
|
695 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
|
696 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
697 /* 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
|
698 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
|
699 { |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
700 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
|
701 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
|
702 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
|
703 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
|
704 if (cache_line1 != cache_line2) |
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
705 { |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
706 miss_positions += 1; |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
707 if (miss_positions > max_allowed_miss_positions) |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
708 return false; |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
709 } |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
710 } |
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
711 return true; |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
712 } |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
713 |
0 | 714 /* Prune the prefetch candidate REF using the reuse with BY. |
715 If BY_IS_BEFORE is true, BY is before REF in the loop. */ | |
716 | |
717 static void | |
718 prune_ref_by_group_reuse (struct mem_ref *ref, struct mem_ref *by, | |
719 bool by_is_before) | |
720 { | |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
721 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
|
722 bool backward; |
0 | 723 HOST_WIDE_INT delta_r = ref->delta, delta_b = by->delta; |
724 HOST_WIDE_INT delta = delta_b - delta_r; | |
725 HOST_WIDE_INT hit_from; | |
726 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
|
727 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
|
728 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
|
729 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
|
730 int align_unit; |
0 | 731 |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
732 /* 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
|
733 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
|
734 return; |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
735 } |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
736 |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
737 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
|
738 |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
739 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
|
740 |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
741 |
0 | 742 if (delta == 0) |
743 { | |
744 /* If the references has the same address, only prefetch the | |
745 former. */ | |
746 if (by_is_before) | |
747 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
|
748 |
0 | 749 return; |
750 } | |
751 | |
752 if (!step) | |
753 { | |
754 /* If the reference addresses are invariant and fall into the | |
755 same cache line, prefetch just the first one. */ | |
756 if (!by_is_before) | |
757 return; | |
758 | |
759 if (ddown (ref->delta, PREFETCH_BLOCK) | |
760 != ddown (by->delta, PREFETCH_BLOCK)) | |
761 return; | |
762 | |
763 ref->prefetch_before = 0; | |
764 return; | |
765 } | |
766 | |
767 /* Only prune the reference that is behind in the array. */ | |
768 if (backward) | |
769 { | |
770 if (delta > 0) | |
771 return; | |
772 | |
773 /* Transform the data so that we may assume that the accesses | |
774 are forward. */ | |
775 delta = - delta; | |
776 step = -step; | |
777 delta_r = PREFETCH_BLOCK - 1 - delta_r; | |
778 delta_b = PREFETCH_BLOCK - 1 - delta_b; | |
779 } | |
780 else | |
781 { | |
782 if (delta < 0) | |
783 return; | |
784 } | |
785 | |
786 /* Check whether the two references are likely to hit the same cache | |
787 line, and how distant the iterations in that it occurs are from | |
788 each other. */ | |
789 | |
790 if (step <= PREFETCH_BLOCK) | |
791 { | |
792 /* The accesses are sure to meet. Let us check when. */ | |
793 hit_from = ddown (delta_b, PREFETCH_BLOCK) * PREFETCH_BLOCK; | |
794 prefetch_before = (hit_from - delta_r + step - 1) / step; | |
795 | |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
796 /* 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
|
797 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
|
798 prefetch_before = PREFETCH_ALL; |
0 | 799 if (prefetch_before < ref->prefetch_before) |
800 ref->prefetch_before = prefetch_before; | |
801 | |
802 return; | |
803 } | |
804 | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
805 /* 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
|
806 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
|
807 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
|
808 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
|
809 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
|
810 that PREFETCH_BLOCK is a power of two. */ |
0 | 811 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
|
812 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
|
813 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
|
814 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
|
815 && reduced_prefetch_block > 1) |
0 | 816 { |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
817 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
|
818 reduced_prefetch_block >>= 1; |
0 | 819 } |
820 | |
821 prefetch_before = delta / step; | |
822 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
|
823 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
|
824 align_unit = TYPE_ALIGN (ref_type) / 8; |
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
825 if (is_miss_rate_acceptable (prefetch_block, step, delta, |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
826 reduced_prefetch_block, align_unit)) |
0 | 827 { |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
828 /* 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
|
829 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
|
830 prefetch_before = PREFETCH_ALL; |
0 | 831 if (prefetch_before < ref->prefetch_before) |
832 ref->prefetch_before = prefetch_before; | |
833 | |
834 return; | |
835 } | |
836 | |
837 /* Try also the following iteration. */ | |
838 prefetch_before++; | |
839 delta = step - delta; | |
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
840 if (is_miss_rate_acceptable (prefetch_block, step, delta, |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
841 reduced_prefetch_block, align_unit)) |
0 | 842 { |
843 if (prefetch_before < ref->prefetch_before) | |
844 ref->prefetch_before = prefetch_before; | |
845 | |
846 return; | |
847 } | |
848 | |
849 /* The ref probably does not reuse by. */ | |
850 return; | |
851 } | |
852 | |
853 /* Prune the prefetch candidate REF using the reuses with other references | |
854 in REFS. */ | |
855 | |
856 static void | |
857 prune_ref_by_reuse (struct mem_ref *ref, struct mem_ref *refs) | |
858 { | |
859 struct mem_ref *prune_by; | |
860 bool before = true; | |
861 | |
862 prune_ref_by_self_reuse (ref); | |
863 | |
864 for (prune_by = refs; prune_by; prune_by = prune_by->next) | |
865 { | |
866 if (prune_by == ref) | |
867 { | |
868 before = false; | |
869 continue; | |
870 } | |
871 | |
872 if (!WRITE_CAN_USE_READ_PREFETCH | |
873 && ref->write_p | |
874 && !prune_by->write_p) | |
875 continue; | |
876 if (!READ_CAN_USE_WRITE_PREFETCH | |
877 && !ref->write_p | |
878 && prune_by->write_p) | |
879 continue; | |
880 | |
881 prune_ref_by_group_reuse (ref, prune_by, before); | |
882 } | |
883 } | |
884 | |
885 /* Prune the prefetch candidates in GROUP using the reuse analysis. */ | |
886 | |
887 static void | |
888 prune_group_by_reuse (struct mem_ref_group *group) | |
889 { | |
890 struct mem_ref *ref_pruned; | |
891 | |
892 for (ref_pruned = group->refs; ref_pruned; ref_pruned = ref_pruned->next) | |
893 { | |
894 prune_ref_by_reuse (ref_pruned, group->refs); | |
895 | |
896 if (dump_file && (dump_flags & TDF_DETAILS)) | |
897 { | |
898 fprintf (dump_file, "Reference %p:", (void *) ref_pruned); | |
899 | |
900 if (ref_pruned->prefetch_before == PREFETCH_ALL | |
901 && ref_pruned->prefetch_mod == 1) | |
902 fprintf (dump_file, " no restrictions"); | |
903 else if (ref_pruned->prefetch_before == 0) | |
904 fprintf (dump_file, " do not prefetch"); | |
905 else if (ref_pruned->prefetch_before <= ref_pruned->prefetch_mod) | |
906 fprintf (dump_file, " prefetch once"); | |
907 else | |
908 { | |
909 if (ref_pruned->prefetch_before != PREFETCH_ALL) | |
910 { | |
911 fprintf (dump_file, " prefetch before "); | |
912 fprintf (dump_file, HOST_WIDE_INT_PRINT_DEC, | |
913 ref_pruned->prefetch_before); | |
914 } | |
915 if (ref_pruned->prefetch_mod != 1) | |
916 { | |
917 fprintf (dump_file, " prefetch mod "); | |
918 fprintf (dump_file, HOST_WIDE_INT_PRINT_DEC, | |
919 ref_pruned->prefetch_mod); | |
920 } | |
921 } | |
922 fprintf (dump_file, "\n"); | |
923 } | |
924 } | |
925 } | |
926 | |
927 /* Prune the list of prefetch candidates GROUPS using the reuse analysis. */ | |
928 | |
929 static void | |
930 prune_by_reuse (struct mem_ref_group *groups) | |
931 { | |
932 for (; groups; groups = groups->next) | |
933 prune_group_by_reuse (groups); | |
934 } | |
935 | |
936 /* Returns true if we should issue prefetch for REF. */ | |
937 | |
938 static bool | |
939 should_issue_prefetch_p (struct mem_ref *ref) | |
940 { | |
941 /* For now do not issue prefetches for only first few of the | |
942 iterations. */ | |
943 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
|
944 { |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
945 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
|
946 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
|
947 (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
|
948 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
|
949 } |
0 | 950 |
951 /* Do not prefetch nontemporal stores. */ | |
952 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
|
953 { |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
954 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
|
955 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
|
956 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
|
957 } |
0 | 958 |
959 return true; | |
960 } | |
961 | |
962 /* Decide which of the prefetch candidates in GROUPS to prefetch. | |
963 AHEAD is the number of iterations to prefetch ahead (which corresponds | |
964 to the number of simultaneous instances of one prefetch running at a | |
965 time). UNROLL_FACTOR is the factor by that the loop is going to be | |
966 unrolled. Returns true if there is anything to prefetch. */ | |
967 | |
968 static bool | |
969 schedule_prefetches (struct mem_ref_group *groups, unsigned unroll_factor, | |
970 unsigned ahead) | |
971 { | |
972 unsigned remaining_prefetch_slots, n_prefetches, prefetch_slots; | |
973 unsigned slots_per_prefetch; | |
974 struct mem_ref *ref; | |
975 bool any = false; | |
976 | |
977 /* At most SIMULTANEOUS_PREFETCHES should be running at the same time. */ | |
978 remaining_prefetch_slots = SIMULTANEOUS_PREFETCHES; | |
979 | |
980 /* The prefetch will run for AHEAD iterations of the original loop, i.e., | |
981 AHEAD / UNROLL_FACTOR iterations of the unrolled loop. In each iteration, | |
982 it will need a prefetch slot. */ | |
983 slots_per_prefetch = (ahead + unroll_factor / 2) / unroll_factor; | |
984 if (dump_file && (dump_flags & TDF_DETAILS)) | |
985 fprintf (dump_file, "Each prefetch instruction takes %u prefetch slots.\n", | |
986 slots_per_prefetch); | |
987 | |
988 /* For now we just take memory references one by one and issue | |
989 prefetches for as many as possible. The groups are sorted | |
990 starting with the largest step, since the references with | |
991 large step are more likely to cause many cache misses. */ | |
992 | |
993 for (; groups; groups = groups->next) | |
994 for (ref = groups->refs; ref; ref = ref->next) | |
995 { | |
996 if (!should_issue_prefetch_p (ref)) | |
997 continue; | |
998 | |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
999 /* 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
|
1000 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
|
1001 prefetches. */ |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1002 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
|
1003 continue; |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1004 |
0 | 1005 /* If we need to prefetch the reference each PREFETCH_MOD iterations, |
1006 and we unroll the loop UNROLL_FACTOR times, we need to insert | |
1007 ceil (UNROLL_FACTOR / PREFETCH_MOD) instructions in each | |
1008 iteration. */ | |
1009 n_prefetches = ((unroll_factor + ref->prefetch_mod - 1) | |
1010 / ref->prefetch_mod); | |
1011 prefetch_slots = n_prefetches * slots_per_prefetch; | |
1012 | |
1013 /* If more than half of the prefetches would be lost anyway, do not | |
1014 issue the prefetch. */ | |
1015 if (2 * remaining_prefetch_slots < prefetch_slots) | |
1016 continue; | |
1017 | |
1018 ref->issue_prefetch_p = true; | |
1019 | |
1020 if (remaining_prefetch_slots <= prefetch_slots) | |
1021 return true; | |
1022 remaining_prefetch_slots -= prefetch_slots; | |
1023 any = true; | |
1024 } | |
1025 | |
1026 return any; | |
1027 } | |
1028 | |
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1029 /* Return TRUE if no prefetch is going to be generated in the given |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1030 GROUPS. */ |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1031 |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1032 static bool |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1033 nothing_to_prefetch_p (struct mem_ref_group *groups) |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1034 { |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1035 struct mem_ref *ref; |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1036 |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1037 for (; groups; groups = groups->next) |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1038 for (ref = groups->refs; ref; ref = ref->next) |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1039 if (should_issue_prefetch_p (ref)) |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1040 return false; |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1041 |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1042 return true; |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1043 } |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1044 |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1045 /* Estimate the number of prefetches in the given GROUPS. |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1046 UNROLL_FACTOR is the factor by which LOOP was unrolled. */ |
0 | 1047 |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1048 static int |
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1049 estimate_prefetch_count (struct mem_ref_group *groups, unsigned unroll_factor) |
0 | 1050 { |
1051 struct mem_ref *ref; | |
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1052 unsigned n_prefetches; |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1053 int prefetch_count = 0; |
0 | 1054 |
1055 for (; groups; groups = groups->next) | |
1056 for (ref = groups->refs; ref; ref = ref->next) | |
1057 if (should_issue_prefetch_p (ref)) | |
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1058 { |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1059 n_prefetches = ((unroll_factor + ref->prefetch_mod - 1) |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1060 / ref->prefetch_mod); |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1061 prefetch_count += n_prefetches; |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1062 } |
0 | 1063 |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1064 return prefetch_count; |
0 | 1065 } |
1066 | |
1067 /* Issue prefetches for the reference REF into loop as decided before. | |
1068 HEAD is the number of iterations to prefetch ahead. UNROLL_FACTOR | |
1069 is the factor by which LOOP was unrolled. */ | |
1070 | |
1071 static void | |
1072 issue_prefetch_ref (struct mem_ref *ref, unsigned unroll_factor, unsigned ahead) | |
1073 { | |
1074 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
|
1075 tree addr, addr_base, write_p, local, forward; |
0 | 1076 gimple prefetch; |
1077 gimple_stmt_iterator bsi; | |
1078 unsigned n_prefetches, ap; | |
1079 bool nontemporal = ref->reuse_distance >= L2_CACHE_SIZE_BYTES; | |
1080 | |
1081 if (dump_file && (dump_flags & TDF_DETAILS)) | |
1082 fprintf (dump_file, "Issued%s prefetch for %p.\n", | |
1083 nontemporal ? " nontemporal" : "", | |
1084 (void *) ref); | |
1085 | |
1086 bsi = gsi_for_stmt (ref->stmt); | |
1087 | |
1088 n_prefetches = ((unroll_factor + ref->prefetch_mod - 1) | |
1089 / ref->prefetch_mod); | |
1090 addr_base = build_fold_addr_expr_with_type (ref->mem, ptr_type_node); | |
1091 addr_base = force_gimple_operand_gsi (&bsi, unshare_expr (addr_base), | |
1092 true, NULL, true, GSI_SAME_STMT); | |
1093 write_p = ref->write_p ? integer_one_node : integer_zero_node; | |
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1094 local = nontemporal ? integer_zero_node : integer_three_node; |
0 | 1095 |
1096 for (ap = 0; ap < n_prefetches; ap++) | |
1097 { | |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1098 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
|
1099 { |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1100 /* 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
|
1101 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
|
1102 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
|
1103 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
|
1104 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
|
1105 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
|
1106 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
|
1107 } |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1108 else |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1109 { |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1110 /* 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
|
1111 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
|
1112 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
|
1113 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
|
1114 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
|
1115 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
|
1116 forward); |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1117 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
|
1118 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
|
1119 } |
0 | 1120 /* Create the prefetch instruction. */ |
1121 prefetch = gimple_build_call (built_in_decls[BUILT_IN_PREFETCH], | |
1122 3, addr, write_p, local); | |
1123 gsi_insert_before (&bsi, prefetch, GSI_SAME_STMT); | |
1124 } | |
1125 } | |
1126 | |
1127 /* Issue prefetches for the references in GROUPS into loop as decided before. | |
1128 HEAD is the number of iterations to prefetch ahead. UNROLL_FACTOR is the | |
1129 factor by that LOOP was unrolled. */ | |
1130 | |
1131 static void | |
1132 issue_prefetches (struct mem_ref_group *groups, | |
1133 unsigned unroll_factor, unsigned ahead) | |
1134 { | |
1135 struct mem_ref *ref; | |
1136 | |
1137 for (; groups; groups = groups->next) | |
1138 for (ref = groups->refs; ref; ref = ref->next) | |
1139 if (ref->issue_prefetch_p) | |
1140 issue_prefetch_ref (ref, unroll_factor, ahead); | |
1141 } | |
1142 | |
1143 /* Returns true if REF is a memory write for that a nontemporal store insn | |
1144 can be used. */ | |
1145 | |
1146 static bool | |
1147 nontemporal_store_p (struct mem_ref *ref) | |
1148 { | |
1149 enum machine_mode mode; | |
1150 enum insn_code code; | |
1151 | |
1152 /* REF must be a write that is not reused. We require it to be independent | |
1153 on all other memory references in the loop, as the nontemporal stores may | |
1154 be reordered with respect to other memory references. */ | |
1155 if (!ref->write_p | |
1156 || !ref->independent_p | |
1157 || ref->reuse_distance < L2_CACHE_SIZE_BYTES) | |
1158 return false; | |
1159 | |
1160 /* Check that we have the storent instruction for the mode. */ | |
1161 mode = TYPE_MODE (TREE_TYPE (ref->mem)); | |
1162 if (mode == BLKmode) | |
1163 return false; | |
1164 | |
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1165 code = optab_handler (storent_optab, mode); |
0 | 1166 return code != CODE_FOR_nothing; |
1167 } | |
1168 | |
1169 /* If REF is a nontemporal store, we mark the corresponding modify statement | |
1170 and return true. Otherwise, we return false. */ | |
1171 | |
1172 static bool | |
1173 mark_nontemporal_store (struct mem_ref *ref) | |
1174 { | |
1175 if (!nontemporal_store_p (ref)) | |
1176 return false; | |
1177 | |
1178 if (dump_file && (dump_flags & TDF_DETAILS)) | |
1179 fprintf (dump_file, "Marked reference %p as a nontemporal store.\n", | |
1180 (void *) ref); | |
1181 | |
1182 gimple_assign_set_nontemporal_move (ref->stmt, true); | |
1183 ref->storent_p = true; | |
1184 | |
1185 return true; | |
1186 } | |
1187 | |
1188 /* Issue a memory fence instruction after LOOP. */ | |
1189 | |
1190 static void | |
1191 emit_mfence_after_loop (struct loop *loop) | |
1192 { | |
1193 VEC (edge, heap) *exits = get_loop_exit_edges (loop); | |
1194 edge exit; | |
1195 gimple call; | |
1196 gimple_stmt_iterator bsi; | |
1197 unsigned i; | |
1198 | |
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1199 FOR_EACH_VEC_ELT (edge, exits, i, exit) |
0 | 1200 { |
1201 call = gimple_build_call (FENCE_FOLLOWING_MOVNT, 0); | |
1202 | |
1203 if (!single_pred_p (exit->dest) | |
1204 /* If possible, we prefer not to insert the fence on other paths | |
1205 in cfg. */ | |
1206 && !(exit->flags & EDGE_ABNORMAL)) | |
1207 split_loop_exit_edge (exit); | |
1208 bsi = gsi_after_labels (exit->dest); | |
1209 | |
1210 gsi_insert_before (&bsi, call, GSI_NEW_STMT); | |
1211 mark_virtual_ops_for_renaming (call); | |
1212 } | |
1213 | |
1214 VEC_free (edge, heap, exits); | |
1215 update_ssa (TODO_update_ssa_only_virtuals); | |
1216 } | |
1217 | |
1218 /* Returns true if we can use storent in loop, false otherwise. */ | |
1219 | |
1220 static bool | |
1221 may_use_storent_in_loop_p (struct loop *loop) | |
1222 { | |
1223 bool ret = true; | |
1224 | |
1225 if (loop->inner != NULL) | |
1226 return false; | |
1227 | |
1228 /* If we must issue a mfence insn after using storent, check that there | |
1229 is a suitable place for it at each of the loop exits. */ | |
1230 if (FENCE_FOLLOWING_MOVNT != NULL_TREE) | |
1231 { | |
1232 VEC (edge, heap) *exits = get_loop_exit_edges (loop); | |
1233 unsigned i; | |
1234 edge exit; | |
1235 | |
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1236 FOR_EACH_VEC_ELT (edge, exits, i, exit) |
0 | 1237 if ((exit->flags & EDGE_ABNORMAL) |
1238 && exit->dest == EXIT_BLOCK_PTR) | |
1239 ret = false; | |
1240 | |
1241 VEC_free (edge, heap, exits); | |
1242 } | |
1243 | |
1244 return ret; | |
1245 } | |
1246 | |
1247 /* Marks nontemporal stores in LOOP. GROUPS contains the description of memory | |
1248 references in the loop. */ | |
1249 | |
1250 static void | |
1251 mark_nontemporal_stores (struct loop *loop, struct mem_ref_group *groups) | |
1252 { | |
1253 struct mem_ref *ref; | |
1254 bool any = false; | |
1255 | |
1256 if (!may_use_storent_in_loop_p (loop)) | |
1257 return; | |
1258 | |
1259 for (; groups; groups = groups->next) | |
1260 for (ref = groups->refs; ref; ref = ref->next) | |
1261 any |= mark_nontemporal_store (ref); | |
1262 | |
1263 if (any && FENCE_FOLLOWING_MOVNT != NULL_TREE) | |
1264 emit_mfence_after_loop (loop); | |
1265 } | |
1266 | |
1267 /* Determines whether we can profitably unroll LOOP FACTOR times, and if | |
1268 this is the case, fill in DESC by the description of number of | |
1269 iterations. */ | |
1270 | |
1271 static bool | |
1272 should_unroll_loop_p (struct loop *loop, struct tree_niter_desc *desc, | |
1273 unsigned factor) | |
1274 { | |
1275 if (!can_unroll_loop_p (loop, factor, desc)) | |
1276 return false; | |
1277 | |
1278 /* We only consider loops without control flow for unrolling. This is not | |
1279 a hard restriction -- tree_unroll_loop works with arbitrary loops | |
1280 as well; but the unrolling/prefetching is usually more profitable for | |
1281 loops consisting of a single basic block, and we want to limit the | |
1282 code growth. */ | |
1283 if (loop->num_nodes > 2) | |
1284 return false; | |
1285 | |
1286 return true; | |
1287 } | |
1288 | |
1289 /* Determine the coefficient by that unroll LOOP, from the information | |
1290 contained in the list of memory references REFS. Description of | |
1291 umber of iterations of LOOP is stored to DESC. NINSNS is the number of | |
1292 insns of the LOOP. EST_NITER is the estimated number of iterations of | |
1293 the loop, or -1 if no estimate is available. */ | |
1294 | |
1295 static unsigned | |
1296 determine_unroll_factor (struct loop *loop, struct mem_ref_group *refs, | |
1297 unsigned ninsns, struct tree_niter_desc *desc, | |
1298 HOST_WIDE_INT est_niter) | |
1299 { | |
1300 unsigned upper_bound; | |
1301 unsigned nfactor, factor, mod_constraint; | |
1302 struct mem_ref_group *agp; | |
1303 struct mem_ref *ref; | |
1304 | |
1305 /* First check whether the loop is not too large to unroll. We ignore | |
1306 PARAM_MAX_UNROLL_TIMES, because for small loops, it prevented us | |
1307 from unrolling them enough to make exactly one cache line covered by each | |
1308 iteration. Also, the goal of PARAM_MAX_UNROLL_TIMES is to prevent | |
1309 us from unrolling the loops too many times in cases where we only expect | |
1310 gains from better scheduling and decreasing loop overhead, which is not | |
1311 the case here. */ | |
1312 upper_bound = PARAM_VALUE (PARAM_MAX_UNROLLED_INSNS) / ninsns; | |
1313 | |
1314 /* If we unrolled the loop more times than it iterates, the unrolled version | |
1315 of the loop would be never entered. */ | |
1316 if (est_niter >= 0 && est_niter < (HOST_WIDE_INT) upper_bound) | |
1317 upper_bound = est_niter; | |
1318 | |
1319 if (upper_bound <= 1) | |
1320 return 1; | |
1321 | |
1322 /* Choose the factor so that we may prefetch each cache just once, | |
1323 but bound the unrolling by UPPER_BOUND. */ | |
1324 factor = 1; | |
1325 for (agp = refs; agp; agp = agp->next) | |
1326 for (ref = agp->refs; ref; ref = ref->next) | |
1327 if (should_issue_prefetch_p (ref)) | |
1328 { | |
1329 mod_constraint = ref->prefetch_mod; | |
1330 nfactor = least_common_multiple (mod_constraint, factor); | |
1331 if (nfactor <= upper_bound) | |
1332 factor = nfactor; | |
1333 } | |
1334 | |
1335 if (!should_unroll_loop_p (loop, desc, factor)) | |
1336 return 1; | |
1337 | |
1338 return factor; | |
1339 } | |
1340 | |
1341 /* Returns the total volume of the memory references REFS, taking into account | |
1342 reuses in the innermost loop and cache line size. TODO -- we should also | |
1343 take into account reuses across the iterations of the loops in the loop | |
1344 nest. */ | |
1345 | |
1346 static unsigned | |
1347 volume_of_references (struct mem_ref_group *refs) | |
1348 { | |
1349 unsigned volume = 0; | |
1350 struct mem_ref_group *gr; | |
1351 struct mem_ref *ref; | |
1352 | |
1353 for (gr = refs; gr; gr = gr->next) | |
1354 for (ref = gr->refs; ref; ref = ref->next) | |
1355 { | |
1356 /* Almost always reuses another value? */ | |
1357 if (ref->prefetch_before != PREFETCH_ALL) | |
1358 continue; | |
1359 | |
1360 /* If several iterations access the same cache line, use the size of | |
1361 the line divided by this number. Otherwise, a cache line is | |
1362 accessed in each iteration. TODO -- in the latter case, we should | |
1363 take the size of the reference into account, rounding it up on cache | |
1364 line size multiple. */ | |
1365 volume += L1_CACHE_LINE_SIZE / ref->prefetch_mod; | |
1366 } | |
1367 return volume; | |
1368 } | |
1369 | |
1370 /* Returns the volume of memory references accessed across VEC iterations of | |
1371 loops, whose sizes are described in the LOOP_SIZES array. N is the number | |
1372 of the loops in the nest (length of VEC and LOOP_SIZES vectors). */ | |
1373 | |
1374 static unsigned | |
1375 volume_of_dist_vector (lambda_vector vec, unsigned *loop_sizes, unsigned n) | |
1376 { | |
1377 unsigned i; | |
1378 | |
1379 for (i = 0; i < n; i++) | |
1380 if (vec[i] != 0) | |
1381 break; | |
1382 | |
1383 if (i == n) | |
1384 return 0; | |
1385 | |
1386 gcc_assert (vec[i] > 0); | |
1387 | |
1388 /* We ignore the parts of the distance vector in subloops, since usually | |
1389 the numbers of iterations are much smaller. */ | |
1390 return loop_sizes[i] * vec[i]; | |
1391 } | |
1392 | |
1393 /* Add the steps of ACCESS_FN multiplied by STRIDE to the array STRIDE | |
1394 at the position corresponding to the loop of the step. N is the depth | |
1395 of the considered loop nest, and, LOOP is its innermost loop. */ | |
1396 | |
1397 static void | |
1398 add_subscript_strides (tree access_fn, unsigned stride, | |
1399 HOST_WIDE_INT *strides, unsigned n, struct loop *loop) | |
1400 { | |
1401 struct loop *aloop; | |
1402 tree step; | |
1403 HOST_WIDE_INT astep; | |
1404 unsigned min_depth = loop_depth (loop) - n; | |
1405 | |
1406 while (TREE_CODE (access_fn) == POLYNOMIAL_CHREC) | |
1407 { | |
1408 aloop = get_chrec_loop (access_fn); | |
1409 step = CHREC_RIGHT (access_fn); | |
1410 access_fn = CHREC_LEFT (access_fn); | |
1411 | |
1412 if ((unsigned) loop_depth (aloop) <= min_depth) | |
1413 continue; | |
1414 | |
1415 if (host_integerp (step, 0)) | |
1416 astep = tree_low_cst (step, 0); | |
1417 else | |
1418 astep = L1_CACHE_LINE_SIZE; | |
1419 | |
1420 strides[n - 1 - loop_depth (loop) + loop_depth (aloop)] += astep * stride; | |
1421 | |
1422 } | |
1423 } | |
1424 | |
1425 /* Returns the volume of memory references accessed between two consecutive | |
1426 self-reuses of the reference DR. We consider the subscripts of DR in N | |
1427 loops, and LOOP_SIZES contains the volumes of accesses in each of the | |
1428 loops. LOOP is the innermost loop of the current loop nest. */ | |
1429 | |
1430 static unsigned | |
1431 self_reuse_distance (data_reference_p dr, unsigned *loop_sizes, unsigned n, | |
1432 struct loop *loop) | |
1433 { | |
1434 tree stride, access_fn; | |
1435 HOST_WIDE_INT *strides, astride; | |
1436 VEC (tree, heap) *access_fns; | |
1437 tree ref = DR_REF (dr); | |
1438 unsigned i, ret = ~0u; | |
1439 | |
1440 /* In the following example: | |
1441 | |
1442 for (i = 0; i < N; i++) | |
1443 for (j = 0; j < N; j++) | |
1444 use (a[j][i]); | |
1445 the same cache line is accessed each N steps (except if the change from | |
1446 i to i + 1 crosses the boundary of the cache line). Thus, for self-reuse, | |
1447 we cannot rely purely on the results of the data dependence analysis. | |
1448 | |
1449 Instead, we compute the stride of the reference in each loop, and consider | |
1450 the innermost loop in that the stride is less than cache size. */ | |
1451 | |
1452 strides = XCNEWVEC (HOST_WIDE_INT, n); | |
1453 access_fns = DR_ACCESS_FNS (dr); | |
1454 | |
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1455 FOR_EACH_VEC_ELT (tree, access_fns, i, access_fn) |
0 | 1456 { |
1457 /* Keep track of the reference corresponding to the subscript, so that we | |
1458 know its stride. */ | |
1459 while (handled_component_p (ref) && TREE_CODE (ref) != ARRAY_REF) | |
1460 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
|
1461 |
0 | 1462 if (TREE_CODE (ref) == ARRAY_REF) |
1463 { | |
1464 stride = TYPE_SIZE_UNIT (TREE_TYPE (ref)); | |
1465 if (host_integerp (stride, 1)) | |
1466 astride = tree_low_cst (stride, 1); | |
1467 else | |
1468 astride = L1_CACHE_LINE_SIZE; | |
1469 | |
1470 ref = TREE_OPERAND (ref, 0); | |
1471 } | |
1472 else | |
1473 astride = 1; | |
1474 | |
1475 add_subscript_strides (access_fn, astride, strides, n, loop); | |
1476 } | |
1477 | |
1478 for (i = n; i-- > 0; ) | |
1479 { | |
1480 unsigned HOST_WIDE_INT s; | |
1481 | |
1482 s = strides[i] < 0 ? -strides[i] : strides[i]; | |
1483 | |
1484 if (s < (unsigned) L1_CACHE_LINE_SIZE | |
1485 && (loop_sizes[i] | |
1486 > (unsigned) (L1_CACHE_SIZE_BYTES / NONTEMPORAL_FRACTION))) | |
1487 { | |
1488 ret = loop_sizes[i]; | |
1489 break; | |
1490 } | |
1491 } | |
1492 | |
1493 free (strides); | |
1494 return ret; | |
1495 } | |
1496 | |
1497 /* Determines the distance till the first reuse of each reference in REFS | |
1498 in the loop nest of LOOP. NO_OTHER_REFS is true if there are no other | |
1499 memory references in the loop. */ | |
1500 | |
1501 static void | |
1502 determine_loop_nest_reuse (struct loop *loop, struct mem_ref_group *refs, | |
1503 bool no_other_refs) | |
1504 { | |
1505 struct loop *nest, *aloop; | |
1506 VEC (data_reference_p, heap) *datarefs = NULL; | |
1507 VEC (ddr_p, heap) *dependences = NULL; | |
1508 struct mem_ref_group *gr; | |
1509 struct mem_ref *ref, *refb; | |
1510 VEC (loop_p, heap) *vloops = NULL; | |
1511 unsigned *loop_data_size; | |
1512 unsigned i, j, n; | |
1513 unsigned volume, dist, adist; | |
1514 HOST_WIDE_INT vol; | |
1515 data_reference_p dr; | |
1516 ddr_p dep; | |
1517 | |
1518 if (loop->inner) | |
1519 return; | |
1520 | |
1521 /* Find the outermost loop of the loop nest of loop (we require that | |
1522 there are no sibling loops inside the nest). */ | |
1523 nest = loop; | |
1524 while (1) | |
1525 { | |
1526 aloop = loop_outer (nest); | |
1527 | |
1528 if (aloop == current_loops->tree_root | |
1529 || aloop->inner->next) | |
1530 break; | |
1531 | |
1532 nest = aloop; | |
1533 } | |
1534 | |
1535 /* For each loop, determine the amount of data accessed in each iteration. | |
1536 We use this to estimate whether the reference is evicted from the | |
1537 cache before its reuse. */ | |
1538 find_loop_nest (nest, &vloops); | |
1539 n = VEC_length (loop_p, vloops); | |
1540 loop_data_size = XNEWVEC (unsigned, n); | |
1541 volume = volume_of_references (refs); | |
1542 i = n; | |
1543 while (i-- != 0) | |
1544 { | |
1545 loop_data_size[i] = volume; | |
1546 /* Bound the volume by the L2 cache size, since above this bound, | |
1547 all dependence distances are equivalent. */ | |
1548 if (volume > L2_CACHE_SIZE_BYTES) | |
1549 continue; | |
1550 | |
1551 aloop = VEC_index (loop_p, vloops, i); | |
1552 vol = estimated_loop_iterations_int (aloop, false); | |
1553 if (vol < 0) | |
1554 vol = expected_loop_iterations (aloop); | |
1555 volume *= vol; | |
1556 } | |
1557 | |
1558 /* Prepare the references in the form suitable for data dependence | |
1559 analysis. We ignore unanalyzable data references (the results | |
1560 are used just as a heuristics to estimate temporality of the | |
1561 references, hence we do not need to worry about correctness). */ | |
1562 for (gr = refs; gr; gr = gr->next) | |
1563 for (ref = gr->refs; ref; ref = ref->next) | |
1564 { | |
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1565 dr = create_data_ref (nest, loop_containing_stmt (ref->stmt), |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1566 ref->mem, ref->stmt, !ref->write_p); |
0 | 1567 |
1568 if (dr) | |
1569 { | |
1570 ref->reuse_distance = volume; | |
1571 dr->aux = ref; | |
1572 VEC_safe_push (data_reference_p, heap, datarefs, dr); | |
1573 } | |
1574 else | |
1575 no_other_refs = false; | |
1576 } | |
1577 | |
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1578 FOR_EACH_VEC_ELT (data_reference_p, datarefs, i, dr) |
0 | 1579 { |
1580 dist = self_reuse_distance (dr, loop_data_size, n, loop); | |
1581 ref = (struct mem_ref *) dr->aux; | |
1582 if (ref->reuse_distance > dist) | |
1583 ref->reuse_distance = dist; | |
1584 | |
1585 if (no_other_refs) | |
1586 ref->independent_p = true; | |
1587 } | |
1588 | |
1589 compute_all_dependences (datarefs, &dependences, vloops, true); | |
1590 | |
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1591 FOR_EACH_VEC_ELT (ddr_p, dependences, i, dep) |
0 | 1592 { |
1593 if (DDR_ARE_DEPENDENT (dep) == chrec_known) | |
1594 continue; | |
1595 | |
1596 ref = (struct mem_ref *) DDR_A (dep)->aux; | |
1597 refb = (struct mem_ref *) DDR_B (dep)->aux; | |
1598 | |
1599 if (DDR_ARE_DEPENDENT (dep) == chrec_dont_know | |
1600 || DDR_NUM_DIST_VECTS (dep) == 0) | |
1601 { | |
1602 /* If the dependence cannot be analyzed, assume that there might be | |
1603 a reuse. */ | |
1604 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
|
1605 |
0 | 1606 ref->independent_p = false; |
1607 refb->independent_p = false; | |
1608 } | |
1609 else | |
1610 { | |
1611 /* The distance vectors are normalized to be always lexicographically | |
1612 positive, hence we cannot tell just from them whether DDR_A comes | |
1613 before DDR_B or vice versa. However, it is not important, | |
1614 anyway -- if DDR_A is close to DDR_B, then it is either reused in | |
1615 DDR_B (and it is not nontemporal), or it reuses the value of DDR_B | |
1616 in cache (and marking it as nontemporal would not affect | |
1617 anything). */ | |
1618 | |
1619 dist = volume; | |
1620 for (j = 0; j < DDR_NUM_DIST_VECTS (dep); j++) | |
1621 { | |
1622 adist = volume_of_dist_vector (DDR_DIST_VECT (dep, j), | |
1623 loop_data_size, n); | |
1624 | |
1625 /* If this is a dependence in the innermost loop (i.e., the | |
1626 distances in all superloops are zero) and it is not | |
1627 the trivial self-dependence with distance zero, record that | |
1628 the references are not completely independent. */ | |
1629 if (lambda_vector_zerop (DDR_DIST_VECT (dep, j), n - 1) | |
1630 && (ref != refb | |
1631 || DDR_DIST_VECT (dep, j)[n-1] != 0)) | |
1632 { | |
1633 ref->independent_p = false; | |
1634 refb->independent_p = false; | |
1635 } | |
1636 | |
1637 /* Ignore accesses closer than | |
1638 L1_CACHE_SIZE_BYTES / NONTEMPORAL_FRACTION, | |
1639 so that we use nontemporal prefetches e.g. if single memory | |
1640 location is accessed several times in a single iteration of | |
1641 the loop. */ | |
1642 if (adist < L1_CACHE_SIZE_BYTES / NONTEMPORAL_FRACTION) | |
1643 continue; | |
1644 | |
1645 if (adist < dist) | |
1646 dist = adist; | |
1647 } | |
1648 } | |
1649 | |
1650 if (ref->reuse_distance > dist) | |
1651 ref->reuse_distance = dist; | |
1652 if (refb->reuse_distance > dist) | |
1653 refb->reuse_distance = dist; | |
1654 } | |
1655 | |
1656 free_dependence_relations (dependences); | |
1657 free_data_refs (datarefs); | |
1658 free (loop_data_size); | |
1659 | |
1660 if (dump_file && (dump_flags & TDF_DETAILS)) | |
1661 { | |
1662 fprintf (dump_file, "Reuse distances:\n"); | |
1663 for (gr = refs; gr; gr = gr->next) | |
1664 for (ref = gr->refs; ref; ref = ref->next) | |
1665 fprintf (dump_file, " ref %p distance %u\n", | |
1666 (void *) ref, ref->reuse_distance); | |
1667 } | |
1668 } | |
1669 | |
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1670 /* Determine whether or not the trip count to ahead ratio is too small based |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1671 on prefitablility consideration. |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1672 AHEAD: the iteration ahead distance, |
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1673 EST_NITER: the estimated trip count. */ |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1674 |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1675 static bool |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1676 trip_count_to_ahead_ratio_too_small_p (unsigned ahead, HOST_WIDE_INT est_niter) |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1677 { |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1678 /* Assume trip count to ahead ratio is big enough if the trip count could not |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1679 be estimated at compile time. */ |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1680 if (est_niter < 0) |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1681 return false; |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1682 |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1683 if (est_niter < (HOST_WIDE_INT) (TRIP_COUNT_TO_AHEAD_RATIO * ahead)) |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1684 { |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1685 if (dump_file && (dump_flags & TDF_DETAILS)) |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1686 fprintf (dump_file, |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1687 "Not prefetching -- loop estimated to roll only %d times\n", |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1688 (int) est_niter); |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1689 return true; |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1690 } |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1691 |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1692 return false; |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1693 } |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1694 |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1695 /* Determine whether or not the number of memory references in the loop is |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1696 reasonable based on the profitablity and compilation time considerations. |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1697 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
|
1698 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
|
1699 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1700 static bool |
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1701 mem_ref_count_reasonable_p (unsigned ninsns, unsigned mem_ref_count) |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1702 { |
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1703 int insn_to_mem_ratio; |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1704 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1705 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
|
1706 return false; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1707 |
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1708 /* Miss rate computation (is_miss_rate_acceptable) and dependence analysis |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1709 (compute_all_dependences) have high costs based on quadratic complexity. |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1710 To avoid huge compilation time, we give up prefetching if mem_ref_count |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1711 is too large. */ |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1712 if (mem_ref_count > PREFETCH_MAX_MEM_REFS_PER_LOOP) |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1713 return false; |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1714 |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1715 /* 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
|
1716 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
|
1717 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
|
1718 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
|
1719 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
|
1720 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
|
1721 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
|
1722 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
|
1723 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
|
1724 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
|
1725 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1726 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
|
1727 { |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1728 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
|
1729 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
|
1730 "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
|
1731 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
|
1732 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
|
1733 } |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1734 |
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1735 return true; |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1736 } |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1737 |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1738 /* Determine whether or not the instruction to prefetch ratio in the loop is |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1739 too small based on the profitablity consideration. |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1740 NINSNS: estimated number of instructions in the loop, |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1741 PREFETCH_COUNT: an estimate of the number of prefetches, |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1742 UNROLL_FACTOR: the factor to unroll the loop if prefetching. */ |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1743 |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1744 static bool |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1745 insn_to_prefetch_ratio_too_small_p (unsigned ninsns, unsigned prefetch_count, |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1746 unsigned unroll_factor) |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1747 { |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1748 int insn_to_prefetch_ratio; |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1749 |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1750 /* 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
|
1751 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
|
1752 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
|
1753 (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
|
1754 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
|
1755 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
|
1756 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
|
1757 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
|
1758 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
|
1759 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
|
1760 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
|
1761 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
|
1762 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
|
1763 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
|
1764 { |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1765 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
|
1766 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
|
1767 "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
|
1768 insn_to_prefetch_ratio); |
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1769 return true; |
63
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1770 } |
b7f97abdc517
update gcc from gcc-4.5.0 to gcc-4.6
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
55
diff
changeset
|
1771 |
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1772 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
|
1773 } |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1774 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1775 |
0 | 1776 /* Issue prefetch instructions for array references in LOOP. Returns |
1777 true if the LOOP was unrolled. */ | |
1778 | |
1779 static bool | |
1780 loop_prefetch_arrays (struct loop *loop) | |
1781 { | |
1782 struct mem_ref_group *refs; | |
1783 unsigned ahead, ninsns, time, unroll_factor; | |
1784 HOST_WIDE_INT est_niter; | |
1785 struct tree_niter_desc desc; | |
1786 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
|
1787 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
|
1788 unsigned mem_ref_count; |
0 | 1789 |
1790 if (optimize_loop_nest_for_size_p (loop)) | |
1791 { | |
1792 if (dump_file && (dump_flags & TDF_DETAILS)) | |
1793 fprintf (dump_file, " ignored (cold area)\n"); | |
1794 return false; | |
1795 } | |
1796 | |
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1797 /* FIXME: the time should be weighted by the probabilities of the blocks in |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1798 the loop body. */ |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1799 time = tree_num_loop_insns (loop, &eni_time_weights); |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1800 if (time == 0) |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1801 return false; |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1802 |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1803 ahead = (PREFETCH_LATENCY + time - 1) / time; |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1804 est_niter = estimated_loop_iterations_int (loop, false); |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1805 |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1806 /* Prefetching is not likely to be profitable if the trip count to ahead |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1807 ratio is too small. */ |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1808 if (trip_count_to_ahead_ratio_too_small_p (ahead, est_niter)) |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1809 return false; |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1810 |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1811 ninsns = tree_num_loop_insns (loop, &eni_size_weights); |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1812 |
0 | 1813 /* 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
|
1814 refs = gather_memory_references (loop, &no_other_refs, &mem_ref_count); |
0 | 1815 |
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1816 /* Give up prefetching if the number of memory references in the |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1817 loop is not reasonable based on profitablity and compilation time |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1818 considerations. */ |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1819 if (!mem_ref_count_reasonable_p (ninsns, mem_ref_count)) |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1820 goto fail; |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1821 |
0 | 1822 /* Step 2: estimate the reuse effects. */ |
1823 prune_by_reuse (refs); | |
1824 | |
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1825 if (nothing_to_prefetch_p (refs)) |
0 | 1826 goto fail; |
1827 | |
1828 determine_loop_nest_reuse (loop, refs, no_other_refs); | |
1829 | |
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1830 /* Step 3: determine unroll factor. */ |
0 | 1831 unroll_factor = determine_unroll_factor (loop, refs, ninsns, &desc, |
1832 est_niter); | |
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1833 |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1834 /* Estimate prefetch count for the unrolled loop. */ |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1835 prefetch_count = estimate_prefetch_count (refs, unroll_factor); |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1836 if (prefetch_count == 0) |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1837 goto fail; |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1838 |
0 | 1839 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
|
1840 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
|
1841 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
|
1842 "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
|
1843 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
|
1844 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
|
1845 |
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1846 /* Prefetching is not likely to be profitable if the instruction to prefetch |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1847 ratio is too small. */ |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1848 if (insn_to_prefetch_ratio_too_small_p (ninsns, prefetch_count, |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1849 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
|
1850 goto fail; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1851 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1852 mark_nontemporal_stores (loop, refs); |
0 | 1853 |
1854 /* Step 4: what to prefetch? */ | |
1855 if (!schedule_prefetches (refs, unroll_factor, ahead)) | |
1856 goto fail; | |
1857 | |
1858 /* Step 5: unroll the loop. TODO -- peeling of first and last few | |
1859 iterations so that we do not issue superfluous prefetches. */ | |
1860 if (unroll_factor != 1) | |
1861 { | |
1862 tree_unroll_loop (loop, unroll_factor, | |
1863 single_dom_exit (loop), &desc); | |
1864 unrolled = true; | |
1865 } | |
1866 | |
1867 /* Step 6: issue the prefetches. */ | |
1868 issue_prefetches (refs, unroll_factor, ahead); | |
1869 | |
1870 fail: | |
1871 release_mem_refs (refs); | |
1872 return unrolled; | |
1873 } | |
1874 | |
1875 /* Issue prefetch instructions for array references in loops. */ | |
1876 | |
1877 unsigned int | |
1878 tree_ssa_prefetch_arrays (void) | |
1879 { | |
1880 loop_iterator li; | |
1881 struct loop *loop; | |
1882 bool unrolled = false; | |
1883 int todo_flags = 0; | |
1884 | |
1885 if (!HAVE_prefetch | |
1886 /* It is possible to ask compiler for say -mtune=i486 -march=pentium4. | |
1887 -mtune=i486 causes us having PREFETCH_BLOCK 0, since this is part | |
1888 of processor costs and i486 does not have prefetch, but | |
1889 -march=pentium4 causes HAVE_prefetch to be true. Ugh. */ | |
1890 || PREFETCH_BLOCK == 0) | |
1891 return 0; | |
1892 | |
1893 if (dump_file && (dump_flags & TDF_DETAILS)) | |
1894 { | |
1895 fprintf (dump_file, "Prefetching parameters:\n"); | |
1896 fprintf (dump_file, " simultaneous prefetches: %d\n", | |
1897 SIMULTANEOUS_PREFETCHES); | |
1898 fprintf (dump_file, " prefetch latency: %d\n", PREFETCH_LATENCY); | |
1899 fprintf (dump_file, " prefetch block size: %d\n", PREFETCH_BLOCK); | |
1900 fprintf (dump_file, " L1 cache size: %d lines, %d kB\n", | |
1901 L1_CACHE_SIZE_BYTES / L1_CACHE_LINE_SIZE, L1_CACHE_SIZE); | |
1902 fprintf (dump_file, " L1 cache line size: %d\n", L1_CACHE_LINE_SIZE); | |
1903 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
|
1904 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
|
1905 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
|
1906 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
|
1907 PREFETCH_MIN_INSN_TO_MEM_RATIO); |
0 | 1908 fprintf (dump_file, "\n"); |
1909 } | |
1910 | |
1911 initialize_original_copy_tables (); | |
1912 | |
1913 if (!built_in_decls[BUILT_IN_PREFETCH]) | |
1914 { | |
67
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1915 tree type = build_function_type_list (void_type_node, |
f6334be47118
update gcc from gcc-4.6-20100522 to gcc-4.6-20110318
nobuyasu <dimolto@cr.ie.u-ryukyu.ac.jp>
parents:
63
diff
changeset
|
1916 const_ptr_type_node, NULL_TREE); |
0 | 1917 tree decl = add_builtin_function ("__builtin_prefetch", type, |
1918 BUILT_IN_PREFETCH, BUILT_IN_NORMAL, | |
1919 NULL, NULL_TREE); | |
1920 DECL_IS_NOVOPS (decl) = true; | |
1921 built_in_decls[BUILT_IN_PREFETCH] = decl; | |
1922 } | |
1923 | |
1924 /* We assume that size of cache line is a power of two, so verify this | |
1925 here. */ | |
1926 gcc_assert ((PREFETCH_BLOCK & (PREFETCH_BLOCK - 1)) == 0); | |
1927 | |
1928 FOR_EACH_LOOP (li, loop, LI_FROM_INNERMOST) | |
1929 { | |
1930 if (dump_file && (dump_flags & TDF_DETAILS)) | |
1931 fprintf (dump_file, "Processing loop %d:\n", loop->num); | |
1932 | |
1933 unrolled |= loop_prefetch_arrays (loop); | |
1934 | |
1935 if (dump_file && (dump_flags & TDF_DETAILS)) | |
1936 fprintf (dump_file, "\n\n"); | |
1937 } | |
1938 | |
1939 if (unrolled) | |
1940 { | |
1941 scev_reset (); | |
1942 todo_flags |= TODO_cleanup_cfg; | |
1943 } | |
1944 | |
1945 free_original_copy_tables (); | |
1946 return todo_flags; | |
1947 } |