1 |
38 |
julius |
/* Array prefetching.
|
2 |
|
|
Copyright (C) 2005, 2007 Free Software Foundation, Inc.
|
3 |
|
|
|
4 |
|
|
This file is part of GCC.
|
5 |
|
|
|
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.
|
10 |
|
|
|
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.
|
15 |
|
|
|
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 "rtl.h"
|
26 |
|
|
#include "tm_p.h"
|
27 |
|
|
#include "hard-reg-set.h"
|
28 |
|
|
#include "basic-block.h"
|
29 |
|
|
#include "output.h"
|
30 |
|
|
#include "diagnostic.h"
|
31 |
|
|
#include "tree-flow.h"
|
32 |
|
|
#include "tree-dump.h"
|
33 |
|
|
#include "timevar.h"
|
34 |
|
|
#include "cfgloop.h"
|
35 |
|
|
#include "varray.h"
|
36 |
|
|
#include "expr.h"
|
37 |
|
|
#include "tree-pass.h"
|
38 |
|
|
#include "ggc.h"
|
39 |
|
|
#include "insn-config.h"
|
40 |
|
|
#include "recog.h"
|
41 |
|
|
#include "hashtab.h"
|
42 |
|
|
#include "tree-chrec.h"
|
43 |
|
|
#include "tree-scalar-evolution.h"
|
44 |
|
|
#include "toplev.h"
|
45 |
|
|
#include "params.h"
|
46 |
|
|
#include "langhooks.h"
|
47 |
|
|
|
48 |
|
|
/* This pass inserts prefetch instructions to optimize cache usage during
|
49 |
|
|
accesses to arrays in loops. It processes loops sequentially and:
|
50 |
|
|
|
51 |
|
|
1) Gathers all memory references in the single loop.
|
52 |
|
|
2) For each of the references it decides when it is profitable to prefetch
|
53 |
|
|
it. To do it, we evaluate the reuse among the accesses, and determines
|
54 |
|
|
two values: PREFETCH_BEFORE (meaning that it only makes sense to do
|
55 |
|
|
prefetching in the first PREFETCH_BEFORE iterations of the loop) and
|
56 |
|
|
PREFETCH_MOD (meaning that it only makes sense to prefetch in the
|
57 |
|
|
iterations of the loop that are zero modulo PREFETCH_MOD). For example
|
58 |
|
|
(assuming cache line size is 64 bytes, char has size 1 byte and there
|
59 |
|
|
is no hardware sequential prefetch):
|
60 |
|
|
|
61 |
|
|
char *a;
|
62 |
|
|
for (i = 0; i < max; i++)
|
63 |
|
|
{
|
64 |
|
|
a[255] = ...; (0)
|
65 |
|
|
a[i] = ...; (1)
|
66 |
|
|
a[i + 64] = ...; (2)
|
67 |
|
|
a[16*i] = ...; (3)
|
68 |
|
|
a[187*i] = ...; (4)
|
69 |
|
|
a[187*i + 50] = ...; (5)
|
70 |
|
|
}
|
71 |
|
|
|
72 |
|
|
(0) obviously has PREFETCH_BEFORE 1
|
73 |
|
|
(1) has PREFETCH_BEFORE 64, since (2) accesses the same memory
|
74 |
|
|
location 64 iterations before it, and PREFETCH_MOD 64 (since
|
75 |
|
|
it hits the same cache line otherwise).
|
76 |
|
|
(2) has PREFETCH_MOD 64
|
77 |
|
|
(3) has PREFETCH_MOD 4
|
78 |
|
|
(4) has PREFETCH_MOD 1. We do not set PREFETCH_BEFORE here, since
|
79 |
|
|
the cache line accessed by (4) is the same with probability only
|
80 |
|
|
7/32.
|
81 |
|
|
(5) has PREFETCH_MOD 1 as well.
|
82 |
|
|
|
83 |
|
|
3) We determine how much ahead we need to prefetch. The number of
|
84 |
|
|
iterations needed is time to fetch / time spent in one iteration of
|
85 |
|
|
the loop. The problem is that we do not know either of these values,
|
86 |
|
|
so we just make a heuristic guess based on a magic (possibly)
|
87 |
|
|
target-specific constant and size of the loop.
|
88 |
|
|
|
89 |
|
|
4) Determine which of the references we prefetch. We take into account
|
90 |
|
|
that there is a maximum number of simultaneous prefetches (provided
|
91 |
|
|
by machine description). We prefetch as many prefetches as possible
|
92 |
|
|
while still within this bound (starting with those with lowest
|
93 |
|
|
prefetch_mod, since they are responsible for most of the cache
|
94 |
|
|
misses).
|
95 |
|
|
|
96 |
|
|
5) We unroll and peel loops so that we are able to satisfy PREFETCH_MOD
|
97 |
|
|
and PREFETCH_BEFORE requirements (within some bounds), and to avoid
|
98 |
|
|
prefetching nonaccessed memory.
|
99 |
|
|
TODO -- actually implement peeling.
|
100 |
|
|
|
101 |
|
|
6) We actually emit the prefetch instructions. ??? Perhaps emit the
|
102 |
|
|
prefetch instructions with guards in cases where 5) was not sufficient
|
103 |
|
|
to satisfy the constraints?
|
104 |
|
|
|
105 |
|
|
Some other TODO:
|
106 |
|
|
-- write and use more general reuse analysis (that could be also used
|
107 |
|
|
in other cache aimed loop optimizations)
|
108 |
|
|
-- make it behave sanely together with the prefetches given by user
|
109 |
|
|
(now we just ignore them; at the very least we should avoid
|
110 |
|
|
optimizing loops in that user put his own prefetches)
|
111 |
|
|
-- we assume cache line size alignment of arrays; this could be
|
112 |
|
|
improved. */
|
113 |
|
|
|
114 |
|
|
/* Magic constants follow. These should be replaced by machine specific
|
115 |
|
|
numbers. */
|
116 |
|
|
|
117 |
|
|
/* A number that should roughly correspond to the number of instructions
|
118 |
|
|
executed before the prefetch is completed. */
|
119 |
|
|
|
120 |
|
|
#ifndef PREFETCH_LATENCY
|
121 |
|
|
#define PREFETCH_LATENCY 200
|
122 |
|
|
#endif
|
123 |
|
|
|
124 |
|
|
/* Number of prefetches that can run at the same time. */
|
125 |
|
|
|
126 |
|
|
#ifndef SIMULTANEOUS_PREFETCHES
|
127 |
|
|
#define SIMULTANEOUS_PREFETCHES 3
|
128 |
|
|
#endif
|
129 |
|
|
|
130 |
|
|
/* True if write can be prefetched by a read prefetch. */
|
131 |
|
|
|
132 |
|
|
#ifndef WRITE_CAN_USE_READ_PREFETCH
|
133 |
|
|
#define WRITE_CAN_USE_READ_PREFETCH 1
|
134 |
|
|
#endif
|
135 |
|
|
|
136 |
|
|
/* True if read can be prefetched by a write prefetch. */
|
137 |
|
|
|
138 |
|
|
#ifndef READ_CAN_USE_WRITE_PREFETCH
|
139 |
|
|
#define READ_CAN_USE_WRITE_PREFETCH 0
|
140 |
|
|
#endif
|
141 |
|
|
|
142 |
|
|
/* Cache line size. Assumed to be a power of two. */
|
143 |
|
|
|
144 |
|
|
#ifndef PREFETCH_BLOCK
|
145 |
|
|
#define PREFETCH_BLOCK 32
|
146 |
|
|
#endif
|
147 |
|
|
|
148 |
|
|
/* Do we have a forward hardware sequential prefetching? */
|
149 |
|
|
|
150 |
|
|
#ifndef HAVE_FORWARD_PREFETCH
|
151 |
|
|
#define HAVE_FORWARD_PREFETCH 0
|
152 |
|
|
#endif
|
153 |
|
|
|
154 |
|
|
/* Do we have a backward hardware sequential prefetching? */
|
155 |
|
|
|
156 |
|
|
#ifndef HAVE_BACKWARD_PREFETCH
|
157 |
|
|
#define HAVE_BACKWARD_PREFETCH 0
|
158 |
|
|
#endif
|
159 |
|
|
|
160 |
|
|
/* In some cases we are only able to determine that there is a certain
|
161 |
|
|
probability that the two accesses hit the same cache line. In this
|
162 |
|
|
case, we issue the prefetches for both of them if this probability
|
163 |
|
|
is less then (1000 - ACCEPTABLE_MISS_RATE) promile. */
|
164 |
|
|
|
165 |
|
|
#ifndef ACCEPTABLE_MISS_RATE
|
166 |
|
|
#define ACCEPTABLE_MISS_RATE 50
|
167 |
|
|
#endif
|
168 |
|
|
|
169 |
|
|
#ifndef HAVE_prefetch
|
170 |
|
|
#define HAVE_prefetch 0
|
171 |
|
|
#endif
|
172 |
|
|
|
173 |
|
|
/* The group of references between that reuse may occur. */
|
174 |
|
|
|
175 |
|
|
struct mem_ref_group
|
176 |
|
|
{
|
177 |
|
|
tree base; /* Base of the reference. */
|
178 |
|
|
HOST_WIDE_INT step; /* Step of the reference. */
|
179 |
|
|
struct mem_ref *refs; /* References in the group. */
|
180 |
|
|
struct mem_ref_group *next; /* Next group of references. */
|
181 |
|
|
};
|
182 |
|
|
|
183 |
|
|
/* Assigned to PREFETCH_BEFORE when all iterations are to be prefetched. */
|
184 |
|
|
|
185 |
|
|
#define PREFETCH_ALL (~(unsigned HOST_WIDE_INT) 0)
|
186 |
|
|
|
187 |
|
|
/* The memory reference. */
|
188 |
|
|
|
189 |
|
|
struct mem_ref
|
190 |
|
|
{
|
191 |
|
|
tree stmt; /* Statement in that the reference appears. */
|
192 |
|
|
tree mem; /* The reference. */
|
193 |
|
|
HOST_WIDE_INT delta; /* Constant offset of the reference. */
|
194 |
|
|
bool write_p; /* Is it a write? */
|
195 |
|
|
struct mem_ref_group *group; /* The group of references it belongs to. */
|
196 |
|
|
unsigned HOST_WIDE_INT prefetch_mod;
|
197 |
|
|
/* Prefetch only each PREFETCH_MOD-th
|
198 |
|
|
iteration. */
|
199 |
|
|
unsigned HOST_WIDE_INT prefetch_before;
|
200 |
|
|
/* Prefetch only first PREFETCH_BEFORE
|
201 |
|
|
iterations. */
|
202 |
|
|
bool issue_prefetch_p; /* Should we really issue the prefetch? */
|
203 |
|
|
struct mem_ref *next; /* The next reference in the group. */
|
204 |
|
|
};
|
205 |
|
|
|
206 |
|
|
/* Dumps information about reference REF to FILE. */
|
207 |
|
|
|
208 |
|
|
static void
|
209 |
|
|
dump_mem_ref (FILE *file, struct mem_ref *ref)
|
210 |
|
|
{
|
211 |
|
|
fprintf (file, "Reference %p:\n", (void *) ref);
|
212 |
|
|
|
213 |
|
|
fprintf (file, " group %p (base ", (void *) ref->group);
|
214 |
|
|
print_generic_expr (file, ref->group->base, TDF_SLIM);
|
215 |
|
|
fprintf (file, ", step ");
|
216 |
|
|
fprintf (file, HOST_WIDE_INT_PRINT_DEC, ref->group->step);
|
217 |
|
|
fprintf (file, ")\n");
|
218 |
|
|
|
219 |
|
|
fprintf (dump_file, " delta ");
|
220 |
|
|
fprintf (file, HOST_WIDE_INT_PRINT_DEC, ref->delta);
|
221 |
|
|
fprintf (file, "\n");
|
222 |
|
|
|
223 |
|
|
fprintf (file, " %s\n", ref->write_p ? "write" : "read");
|
224 |
|
|
|
225 |
|
|
fprintf (file, "\n");
|
226 |
|
|
}
|
227 |
|
|
|
228 |
|
|
/* Finds a group with BASE and STEP in GROUPS, or creates one if it does not
|
229 |
|
|
exist. */
|
230 |
|
|
|
231 |
|
|
static struct mem_ref_group *
|
232 |
|
|
find_or_create_group (struct mem_ref_group **groups, tree base,
|
233 |
|
|
HOST_WIDE_INT step)
|
234 |
|
|
{
|
235 |
|
|
struct mem_ref_group *group;
|
236 |
|
|
|
237 |
|
|
for (; *groups; groups = &(*groups)->next)
|
238 |
|
|
{
|
239 |
|
|
if ((*groups)->step == step
|
240 |
|
|
&& operand_equal_p ((*groups)->base, base, 0))
|
241 |
|
|
return *groups;
|
242 |
|
|
|
243 |
|
|
/* Keep the list of groups sorted by decreasing step. */
|
244 |
|
|
if ((*groups)->step < step)
|
245 |
|
|
break;
|
246 |
|
|
}
|
247 |
|
|
|
248 |
|
|
group = xcalloc (1, sizeof (struct mem_ref_group));
|
249 |
|
|
group->base = base;
|
250 |
|
|
group->step = step;
|
251 |
|
|
group->refs = NULL;
|
252 |
|
|
group->next = *groups;
|
253 |
|
|
*groups = group;
|
254 |
|
|
|
255 |
|
|
return group;
|
256 |
|
|
}
|
257 |
|
|
|
258 |
|
|
/* Records a memory reference MEM in GROUP with offset DELTA and write status
|
259 |
|
|
WRITE_P. The reference occurs in statement STMT. */
|
260 |
|
|
|
261 |
|
|
static void
|
262 |
|
|
record_ref (struct mem_ref_group *group, tree stmt, tree mem,
|
263 |
|
|
HOST_WIDE_INT delta, bool write_p)
|
264 |
|
|
{
|
265 |
|
|
struct mem_ref **aref;
|
266 |
|
|
|
267 |
|
|
/* Do not record the same address twice. */
|
268 |
|
|
for (aref = &group->refs; *aref; aref = &(*aref)->next)
|
269 |
|
|
{
|
270 |
|
|
/* It does not have to be possible for write reference to reuse the read
|
271 |
|
|
prefetch, or vice versa. */
|
272 |
|
|
if (!WRITE_CAN_USE_READ_PREFETCH
|
273 |
|
|
&& write_p
|
274 |
|
|
&& !(*aref)->write_p)
|
275 |
|
|
continue;
|
276 |
|
|
if (!READ_CAN_USE_WRITE_PREFETCH
|
277 |
|
|
&& !write_p
|
278 |
|
|
&& (*aref)->write_p)
|
279 |
|
|
continue;
|
280 |
|
|
|
281 |
|
|
if ((*aref)->delta == delta)
|
282 |
|
|
return;
|
283 |
|
|
}
|
284 |
|
|
|
285 |
|
|
(*aref) = xcalloc (1, sizeof (struct mem_ref));
|
286 |
|
|
(*aref)->stmt = stmt;
|
287 |
|
|
(*aref)->mem = mem;
|
288 |
|
|
(*aref)->delta = delta;
|
289 |
|
|
(*aref)->write_p = write_p;
|
290 |
|
|
(*aref)->prefetch_before = PREFETCH_ALL;
|
291 |
|
|
(*aref)->prefetch_mod = 1;
|
292 |
|
|
(*aref)->issue_prefetch_p = false;
|
293 |
|
|
(*aref)->group = group;
|
294 |
|
|
(*aref)->next = NULL;
|
295 |
|
|
|
296 |
|
|
if (dump_file && (dump_flags & TDF_DETAILS))
|
297 |
|
|
dump_mem_ref (dump_file, *aref);
|
298 |
|
|
}
|
299 |
|
|
|
300 |
|
|
/* Release memory references in GROUPS. */
|
301 |
|
|
|
302 |
|
|
static void
|
303 |
|
|
release_mem_refs (struct mem_ref_group *groups)
|
304 |
|
|
{
|
305 |
|
|
struct mem_ref_group *next_g;
|
306 |
|
|
struct mem_ref *ref, *next_r;
|
307 |
|
|
|
308 |
|
|
for (; groups; groups = next_g)
|
309 |
|
|
{
|
310 |
|
|
next_g = groups->next;
|
311 |
|
|
for (ref = groups->refs; ref; ref = next_r)
|
312 |
|
|
{
|
313 |
|
|
next_r = ref->next;
|
314 |
|
|
free (ref);
|
315 |
|
|
}
|
316 |
|
|
free (groups);
|
317 |
|
|
}
|
318 |
|
|
}
|
319 |
|
|
|
320 |
|
|
/* A structure used to pass arguments to idx_analyze_ref. */
|
321 |
|
|
|
322 |
|
|
struct ar_data
|
323 |
|
|
{
|
324 |
|
|
struct loop *loop; /* Loop of the reference. */
|
325 |
|
|
tree stmt; /* Statement of the reference. */
|
326 |
|
|
HOST_WIDE_INT *step; /* Step of the memory reference. */
|
327 |
|
|
HOST_WIDE_INT *delta; /* Offset of the memory reference. */
|
328 |
|
|
};
|
329 |
|
|
|
330 |
|
|
/* Analyzes a single INDEX of a memory reference to obtain information
|
331 |
|
|
described at analyze_ref. Callback for for_each_index. */
|
332 |
|
|
|
333 |
|
|
static bool
|
334 |
|
|
idx_analyze_ref (tree base, tree *index, void *data)
|
335 |
|
|
{
|
336 |
|
|
struct ar_data *ar_data = data;
|
337 |
|
|
tree ibase, step, stepsize;
|
338 |
|
|
HOST_WIDE_INT istep, idelta = 0, imult = 1;
|
339 |
|
|
affine_iv iv;
|
340 |
|
|
|
341 |
|
|
if (TREE_CODE (base) == MISALIGNED_INDIRECT_REF
|
342 |
|
|
|| TREE_CODE (base) == ALIGN_INDIRECT_REF)
|
343 |
|
|
return false;
|
344 |
|
|
|
345 |
|
|
if (!simple_iv (ar_data->loop, ar_data->stmt, *index, &iv, false))
|
346 |
|
|
return false;
|
347 |
|
|
ibase = iv.base;
|
348 |
|
|
step = iv.step;
|
349 |
|
|
|
350 |
|
|
if (zero_p (step))
|
351 |
|
|
istep = 0;
|
352 |
|
|
else
|
353 |
|
|
{
|
354 |
|
|
if (!cst_and_fits_in_hwi (step))
|
355 |
|
|
return false;
|
356 |
|
|
istep = int_cst_value (step);
|
357 |
|
|
}
|
358 |
|
|
|
359 |
|
|
if (TREE_CODE (ibase) == PLUS_EXPR
|
360 |
|
|
&& cst_and_fits_in_hwi (TREE_OPERAND (ibase, 1)))
|
361 |
|
|
{
|
362 |
|
|
idelta = int_cst_value (TREE_OPERAND (ibase, 1));
|
363 |
|
|
ibase = TREE_OPERAND (ibase, 0);
|
364 |
|
|
}
|
365 |
|
|
if (cst_and_fits_in_hwi (ibase))
|
366 |
|
|
{
|
367 |
|
|
idelta += int_cst_value (ibase);
|
368 |
|
|
ibase = build_int_cst (TREE_TYPE (ibase), 0);
|
369 |
|
|
}
|
370 |
|
|
|
371 |
|
|
if (TREE_CODE (base) == ARRAY_REF)
|
372 |
|
|
{
|
373 |
|
|
stepsize = array_ref_element_size (base);
|
374 |
|
|
if (!cst_and_fits_in_hwi (stepsize))
|
375 |
|
|
return false;
|
376 |
|
|
imult = int_cst_value (stepsize);
|
377 |
|
|
|
378 |
|
|
istep *= imult;
|
379 |
|
|
idelta *= imult;
|
380 |
|
|
}
|
381 |
|
|
|
382 |
|
|
*ar_data->step += istep;
|
383 |
|
|
*ar_data->delta += idelta;
|
384 |
|
|
*index = ibase;
|
385 |
|
|
|
386 |
|
|
return true;
|
387 |
|
|
}
|
388 |
|
|
|
389 |
|
|
/* Tries to express REF_P in shape &BASE + STEP * iter + DELTA, where DELTA and
|
390 |
|
|
STEP are integer constants and iter is number of iterations of LOOP. The
|
391 |
|
|
reference occurs in statement STMT. Strips nonaddressable component
|
392 |
|
|
references from REF_P. */
|
393 |
|
|
|
394 |
|
|
static bool
|
395 |
|
|
analyze_ref (struct loop *loop, tree *ref_p, tree *base,
|
396 |
|
|
HOST_WIDE_INT *step, HOST_WIDE_INT *delta,
|
397 |
|
|
tree stmt)
|
398 |
|
|
{
|
399 |
|
|
struct ar_data ar_data;
|
400 |
|
|
tree off;
|
401 |
|
|
HOST_WIDE_INT bit_offset;
|
402 |
|
|
tree ref = *ref_p;
|
403 |
|
|
|
404 |
|
|
*step = 0;
|
405 |
|
|
*delta = 0;
|
406 |
|
|
|
407 |
|
|
/* First strip off the component references. Ignore bitfields. */
|
408 |
|
|
if (TREE_CODE (ref) == COMPONENT_REF
|
409 |
|
|
&& DECL_NONADDRESSABLE_P (TREE_OPERAND (ref, 1)))
|
410 |
|
|
ref = TREE_OPERAND (ref, 0);
|
411 |
|
|
|
412 |
|
|
*ref_p = ref;
|
413 |
|
|
|
414 |
|
|
for (; TREE_CODE (ref) == COMPONENT_REF; ref = TREE_OPERAND (ref, 0))
|
415 |
|
|
{
|
416 |
|
|
off = DECL_FIELD_BIT_OFFSET (TREE_OPERAND (ref, 1));
|
417 |
|
|
bit_offset = TREE_INT_CST_LOW (off);
|
418 |
|
|
gcc_assert (bit_offset % BITS_PER_UNIT == 0);
|
419 |
|
|
|
420 |
|
|
*delta += bit_offset / BITS_PER_UNIT;
|
421 |
|
|
}
|
422 |
|
|
|
423 |
|
|
*base = unshare_expr (ref);
|
424 |
|
|
ar_data.loop = loop;
|
425 |
|
|
ar_data.stmt = stmt;
|
426 |
|
|
ar_data.step = step;
|
427 |
|
|
ar_data.delta = delta;
|
428 |
|
|
return for_each_index (base, idx_analyze_ref, &ar_data);
|
429 |
|
|
}
|
430 |
|
|
|
431 |
|
|
/* Record a memory reference REF to the list REFS. The reference occurs in
|
432 |
|
|
LOOP in statement STMT and it is write if WRITE_P. */
|
433 |
|
|
|
434 |
|
|
static void
|
435 |
|
|
gather_memory_references_ref (struct loop *loop, struct mem_ref_group **refs,
|
436 |
|
|
tree ref, bool write_p, tree stmt)
|
437 |
|
|
{
|
438 |
|
|
tree base;
|
439 |
|
|
HOST_WIDE_INT step, delta;
|
440 |
|
|
struct mem_ref_group *agrp;
|
441 |
|
|
|
442 |
|
|
if (!analyze_ref (loop, &ref, &base, &step, &delta, stmt))
|
443 |
|
|
return;
|
444 |
|
|
|
445 |
|
|
/* Now we know that REF = &BASE + STEP * iter + DELTA, where DELTA and STEP
|
446 |
|
|
are integer constants. */
|
447 |
|
|
agrp = find_or_create_group (refs, base, step);
|
448 |
|
|
record_ref (agrp, stmt, ref, delta, write_p);
|
449 |
|
|
}
|
450 |
|
|
|
451 |
|
|
/* Record the suitable memory references in LOOP. */
|
452 |
|
|
|
453 |
|
|
static struct mem_ref_group *
|
454 |
|
|
gather_memory_references (struct loop *loop)
|
455 |
|
|
{
|
456 |
|
|
basic_block *body = get_loop_body_in_dom_order (loop);
|
457 |
|
|
basic_block bb;
|
458 |
|
|
unsigned i;
|
459 |
|
|
block_stmt_iterator bsi;
|
460 |
|
|
tree stmt, lhs, rhs;
|
461 |
|
|
struct mem_ref_group *refs = NULL;
|
462 |
|
|
|
463 |
|
|
/* Scan the loop body in order, so that the former references precede the
|
464 |
|
|
later ones. */
|
465 |
|
|
for (i = 0; i < loop->num_nodes; i++)
|
466 |
|
|
{
|
467 |
|
|
bb = body[i];
|
468 |
|
|
if (bb->loop_father != loop)
|
469 |
|
|
continue;
|
470 |
|
|
|
471 |
|
|
for (bsi = bsi_start (bb); !bsi_end_p (bsi); bsi_next (&bsi))
|
472 |
|
|
{
|
473 |
|
|
stmt = bsi_stmt (bsi);
|
474 |
|
|
if (TREE_CODE (stmt) != MODIFY_EXPR)
|
475 |
|
|
continue;
|
476 |
|
|
|
477 |
|
|
lhs = TREE_OPERAND (stmt, 0);
|
478 |
|
|
rhs = TREE_OPERAND (stmt, 1);
|
479 |
|
|
|
480 |
|
|
if (REFERENCE_CLASS_P (rhs))
|
481 |
|
|
gather_memory_references_ref (loop, &refs, rhs, false, stmt);
|
482 |
|
|
if (REFERENCE_CLASS_P (lhs))
|
483 |
|
|
gather_memory_references_ref (loop, &refs, lhs, true, stmt);
|
484 |
|
|
}
|
485 |
|
|
}
|
486 |
|
|
free (body);
|
487 |
|
|
|
488 |
|
|
return refs;
|
489 |
|
|
}
|
490 |
|
|
|
491 |
|
|
/* Prune the prefetch candidate REF using the self-reuse. */
|
492 |
|
|
|
493 |
|
|
static void
|
494 |
|
|
prune_ref_by_self_reuse (struct mem_ref *ref)
|
495 |
|
|
{
|
496 |
|
|
HOST_WIDE_INT step = ref->group->step;
|
497 |
|
|
bool backward = step < 0;
|
498 |
|
|
|
499 |
|
|
if (step == 0)
|
500 |
|
|
{
|
501 |
|
|
/* Prefetch references to invariant address just once. */
|
502 |
|
|
ref->prefetch_before = 1;
|
503 |
|
|
return;
|
504 |
|
|
}
|
505 |
|
|
|
506 |
|
|
if (backward)
|
507 |
|
|
step = -step;
|
508 |
|
|
|
509 |
|
|
if (step > PREFETCH_BLOCK)
|
510 |
|
|
return;
|
511 |
|
|
|
512 |
|
|
if ((backward && HAVE_BACKWARD_PREFETCH)
|
513 |
|
|
|| (!backward && HAVE_FORWARD_PREFETCH))
|
514 |
|
|
{
|
515 |
|
|
ref->prefetch_before = 1;
|
516 |
|
|
return;
|
517 |
|
|
}
|
518 |
|
|
|
519 |
|
|
ref->prefetch_mod = PREFETCH_BLOCK / step;
|
520 |
|
|
}
|
521 |
|
|
|
522 |
|
|
/* Divides X by BY, rounding down. */
|
523 |
|
|
|
524 |
|
|
static HOST_WIDE_INT
|
525 |
|
|
ddown (HOST_WIDE_INT x, unsigned HOST_WIDE_INT by)
|
526 |
|
|
{
|
527 |
|
|
gcc_assert (by > 0);
|
528 |
|
|
|
529 |
|
|
if (x >= 0)
|
530 |
|
|
return x / by;
|
531 |
|
|
else
|
532 |
|
|
return (x + by - 1) / by;
|
533 |
|
|
}
|
534 |
|
|
|
535 |
|
|
/* Prune the prefetch candidate REF using the reuse with BY.
|
536 |
|
|
If BY_IS_BEFORE is true, BY is before REF in the loop. */
|
537 |
|
|
|
538 |
|
|
static void
|
539 |
|
|
prune_ref_by_group_reuse (struct mem_ref *ref, struct mem_ref *by,
|
540 |
|
|
bool by_is_before)
|
541 |
|
|
{
|
542 |
|
|
HOST_WIDE_INT step = ref->group->step;
|
543 |
|
|
bool backward = step < 0;
|
544 |
|
|
HOST_WIDE_INT delta_r = ref->delta, delta_b = by->delta;
|
545 |
|
|
HOST_WIDE_INT delta = delta_b - delta_r;
|
546 |
|
|
HOST_WIDE_INT hit_from;
|
547 |
|
|
unsigned HOST_WIDE_INT prefetch_before, prefetch_block;
|
548 |
|
|
|
549 |
|
|
if (delta == 0)
|
550 |
|
|
{
|
551 |
|
|
/* If the references has the same address, only prefetch the
|
552 |
|
|
former. */
|
553 |
|
|
if (by_is_before)
|
554 |
|
|
ref->prefetch_before = 0;
|
555 |
|
|
|
556 |
|
|
return;
|
557 |
|
|
}
|
558 |
|
|
|
559 |
|
|
if (!step)
|
560 |
|
|
{
|
561 |
|
|
/* If the reference addresses are invariant and fall into the
|
562 |
|
|
same cache line, prefetch just the first one. */
|
563 |
|
|
if (!by_is_before)
|
564 |
|
|
return;
|
565 |
|
|
|
566 |
|
|
if (ddown (ref->delta, PREFETCH_BLOCK)
|
567 |
|
|
!= ddown (by->delta, PREFETCH_BLOCK))
|
568 |
|
|
return;
|
569 |
|
|
|
570 |
|
|
ref->prefetch_before = 0;
|
571 |
|
|
return;
|
572 |
|
|
}
|
573 |
|
|
|
574 |
|
|
/* Only prune the reference that is behind in the array. */
|
575 |
|
|
if (backward)
|
576 |
|
|
{
|
577 |
|
|
if (delta > 0)
|
578 |
|
|
return;
|
579 |
|
|
|
580 |
|
|
/* Transform the data so that we may assume that the accesses
|
581 |
|
|
are forward. */
|
582 |
|
|
delta = - delta;
|
583 |
|
|
step = -step;
|
584 |
|
|
delta_r = PREFETCH_BLOCK - 1 - delta_r;
|
585 |
|
|
delta_b = PREFETCH_BLOCK - 1 - delta_b;
|
586 |
|
|
}
|
587 |
|
|
else
|
588 |
|
|
{
|
589 |
|
|
if (delta < 0)
|
590 |
|
|
return;
|
591 |
|
|
}
|
592 |
|
|
|
593 |
|
|
/* Check whether the two references are likely to hit the same cache
|
594 |
|
|
line, and how distant the iterations in that it occurs are from
|
595 |
|
|
each other. */
|
596 |
|
|
|
597 |
|
|
if (step <= PREFETCH_BLOCK)
|
598 |
|
|
{
|
599 |
|
|
/* The accesses are sure to meet. Let us check when. */
|
600 |
|
|
hit_from = ddown (delta_b, PREFETCH_BLOCK) * PREFETCH_BLOCK;
|
601 |
|
|
prefetch_before = (hit_from - delta_r + step - 1) / step;
|
602 |
|
|
|
603 |
|
|
if (prefetch_before < ref->prefetch_before)
|
604 |
|
|
ref->prefetch_before = prefetch_before;
|
605 |
|
|
|
606 |
|
|
return;
|
607 |
|
|
}
|
608 |
|
|
|
609 |
|
|
/* A more complicated case. First let us ensure that size of cache line
|
610 |
|
|
and step are coprime (here we assume that PREFETCH_BLOCK is a power
|
611 |
|
|
of two. */
|
612 |
|
|
prefetch_block = PREFETCH_BLOCK;
|
613 |
|
|
while ((step & 1) == 0
|
614 |
|
|
&& prefetch_block > 1)
|
615 |
|
|
{
|
616 |
|
|
step >>= 1;
|
617 |
|
|
prefetch_block >>= 1;
|
618 |
|
|
delta >>= 1;
|
619 |
|
|
}
|
620 |
|
|
|
621 |
|
|
/* Now step > prefetch_block, and step and prefetch_block are coprime.
|
622 |
|
|
Determine the probability that the accesses hit the same cache line. */
|
623 |
|
|
|
624 |
|
|
prefetch_before = delta / step;
|
625 |
|
|
delta %= step;
|
626 |
|
|
if ((unsigned HOST_WIDE_INT) delta
|
627 |
|
|
<= (prefetch_block * ACCEPTABLE_MISS_RATE / 1000))
|
628 |
|
|
{
|
629 |
|
|
if (prefetch_before < ref->prefetch_before)
|
630 |
|
|
ref->prefetch_before = prefetch_before;
|
631 |
|
|
|
632 |
|
|
return;
|
633 |
|
|
}
|
634 |
|
|
|
635 |
|
|
/* Try also the following iteration. */
|
636 |
|
|
prefetch_before++;
|
637 |
|
|
delta = step - delta;
|
638 |
|
|
if ((unsigned HOST_WIDE_INT) delta
|
639 |
|
|
<= (prefetch_block * ACCEPTABLE_MISS_RATE / 1000))
|
640 |
|
|
{
|
641 |
|
|
if (prefetch_before < ref->prefetch_before)
|
642 |
|
|
ref->prefetch_before = prefetch_before;
|
643 |
|
|
|
644 |
|
|
return;
|
645 |
|
|
}
|
646 |
|
|
|
647 |
|
|
/* The ref probably does not reuse by. */
|
648 |
|
|
return;
|
649 |
|
|
}
|
650 |
|
|
|
651 |
|
|
/* Prune the prefetch candidate REF using the reuses with other references
|
652 |
|
|
in REFS. */
|
653 |
|
|
|
654 |
|
|
static void
|
655 |
|
|
prune_ref_by_reuse (struct mem_ref *ref, struct mem_ref *refs)
|
656 |
|
|
{
|
657 |
|
|
struct mem_ref *prune_by;
|
658 |
|
|
bool before = true;
|
659 |
|
|
|
660 |
|
|
prune_ref_by_self_reuse (ref);
|
661 |
|
|
|
662 |
|
|
for (prune_by = refs; prune_by; prune_by = prune_by->next)
|
663 |
|
|
{
|
664 |
|
|
if (prune_by == ref)
|
665 |
|
|
{
|
666 |
|
|
before = false;
|
667 |
|
|
continue;
|
668 |
|
|
}
|
669 |
|
|
|
670 |
|
|
if (!WRITE_CAN_USE_READ_PREFETCH
|
671 |
|
|
&& ref->write_p
|
672 |
|
|
&& !prune_by->write_p)
|
673 |
|
|
continue;
|
674 |
|
|
if (!READ_CAN_USE_WRITE_PREFETCH
|
675 |
|
|
&& !ref->write_p
|
676 |
|
|
&& prune_by->write_p)
|
677 |
|
|
continue;
|
678 |
|
|
|
679 |
|
|
prune_ref_by_group_reuse (ref, prune_by, before);
|
680 |
|
|
}
|
681 |
|
|
}
|
682 |
|
|
|
683 |
|
|
/* Prune the prefetch candidates in GROUP using the reuse analysis. */
|
684 |
|
|
|
685 |
|
|
static void
|
686 |
|
|
prune_group_by_reuse (struct mem_ref_group *group)
|
687 |
|
|
{
|
688 |
|
|
struct mem_ref *ref_pruned;
|
689 |
|
|
|
690 |
|
|
for (ref_pruned = group->refs; ref_pruned; ref_pruned = ref_pruned->next)
|
691 |
|
|
{
|
692 |
|
|
prune_ref_by_reuse (ref_pruned, group->refs);
|
693 |
|
|
|
694 |
|
|
if (dump_file && (dump_flags & TDF_DETAILS))
|
695 |
|
|
{
|
696 |
|
|
fprintf (dump_file, "Reference %p:", (void *) ref_pruned);
|
697 |
|
|
|
698 |
|
|
if (ref_pruned->prefetch_before == PREFETCH_ALL
|
699 |
|
|
&& ref_pruned->prefetch_mod == 1)
|
700 |
|
|
fprintf (dump_file, " no restrictions");
|
701 |
|
|
else if (ref_pruned->prefetch_before == 0)
|
702 |
|
|
fprintf (dump_file, " do not prefetch");
|
703 |
|
|
else if (ref_pruned->prefetch_before <= ref_pruned->prefetch_mod)
|
704 |
|
|
fprintf (dump_file, " prefetch once");
|
705 |
|
|
else
|
706 |
|
|
{
|
707 |
|
|
if (ref_pruned->prefetch_before != PREFETCH_ALL)
|
708 |
|
|
{
|
709 |
|
|
fprintf (dump_file, " prefetch before ");
|
710 |
|
|
fprintf (dump_file, HOST_WIDE_INT_PRINT_DEC,
|
711 |
|
|
ref_pruned->prefetch_before);
|
712 |
|
|
}
|
713 |
|
|
if (ref_pruned->prefetch_mod != 1)
|
714 |
|
|
{
|
715 |
|
|
fprintf (dump_file, " prefetch mod ");
|
716 |
|
|
fprintf (dump_file, HOST_WIDE_INT_PRINT_DEC,
|
717 |
|
|
ref_pruned->prefetch_mod);
|
718 |
|
|
}
|
719 |
|
|
}
|
720 |
|
|
fprintf (dump_file, "\n");
|
721 |
|
|
}
|
722 |
|
|
}
|
723 |
|
|
}
|
724 |
|
|
|
725 |
|
|
/* Prune the list of prefetch candidates GROUPS using the reuse analysis. */
|
726 |
|
|
|
727 |
|
|
static void
|
728 |
|
|
prune_by_reuse (struct mem_ref_group *groups)
|
729 |
|
|
{
|
730 |
|
|
for (; groups; groups = groups->next)
|
731 |
|
|
prune_group_by_reuse (groups);
|
732 |
|
|
}
|
733 |
|
|
|
734 |
|
|
/* Returns true if we should issue prefetch for REF. */
|
735 |
|
|
|
736 |
|
|
static bool
|
737 |
|
|
should_issue_prefetch_p (struct mem_ref *ref)
|
738 |
|
|
{
|
739 |
|
|
/* For now do not issue prefetches for only first few of the
|
740 |
|
|
iterations. */
|
741 |
|
|
if (ref->prefetch_before != PREFETCH_ALL)
|
742 |
|
|
return false;
|
743 |
|
|
|
744 |
|
|
return true;
|
745 |
|
|
}
|
746 |
|
|
|
747 |
|
|
/* Decide which of the prefetch candidates in GROUPS to prefetch.
|
748 |
|
|
AHEAD is the number of iterations to prefetch ahead (which corresponds
|
749 |
|
|
to the number of simultaneous instances of one prefetch running at a
|
750 |
|
|
time). UNROLL_FACTOR is the factor by that the loop is going to be
|
751 |
|
|
unrolled. Returns true if there is anything to prefetch. */
|
752 |
|
|
|
753 |
|
|
static bool
|
754 |
|
|
schedule_prefetches (struct mem_ref_group *groups, unsigned unroll_factor,
|
755 |
|
|
unsigned ahead)
|
756 |
|
|
{
|
757 |
|
|
unsigned max_prefetches, n_prefetches;
|
758 |
|
|
struct mem_ref *ref;
|
759 |
|
|
bool any = false;
|
760 |
|
|
|
761 |
|
|
max_prefetches = (SIMULTANEOUS_PREFETCHES * unroll_factor) / ahead;
|
762 |
|
|
if (max_prefetches > (unsigned) SIMULTANEOUS_PREFETCHES)
|
763 |
|
|
max_prefetches = SIMULTANEOUS_PREFETCHES;
|
764 |
|
|
|
765 |
|
|
if (dump_file && (dump_flags & TDF_DETAILS))
|
766 |
|
|
fprintf (dump_file, "Max prefetches to issue: %d.\n", max_prefetches);
|
767 |
|
|
|
768 |
|
|
if (!max_prefetches)
|
769 |
|
|
return false;
|
770 |
|
|
|
771 |
|
|
/* For now we just take memory references one by one and issue
|
772 |
|
|
prefetches for as many as possible. The groups are sorted
|
773 |
|
|
starting with the largest step, since the references with
|
774 |
|
|
large step are more likely to cause many cache misses. */
|
775 |
|
|
|
776 |
|
|
for (; groups; groups = groups->next)
|
777 |
|
|
for (ref = groups->refs; ref; ref = ref->next)
|
778 |
|
|
{
|
779 |
|
|
if (!should_issue_prefetch_p (ref))
|
780 |
|
|
continue;
|
781 |
|
|
|
782 |
|
|
ref->issue_prefetch_p = true;
|
783 |
|
|
|
784 |
|
|
/* If prefetch_mod is less then unroll_factor, we need to insert
|
785 |
|
|
several prefetches for the reference. */
|
786 |
|
|
n_prefetches = ((unroll_factor + ref->prefetch_mod - 1)
|
787 |
|
|
/ ref->prefetch_mod);
|
788 |
|
|
if (max_prefetches <= n_prefetches)
|
789 |
|
|
return true;
|
790 |
|
|
|
791 |
|
|
max_prefetches -= n_prefetches;
|
792 |
|
|
any = true;
|
793 |
|
|
}
|
794 |
|
|
|
795 |
|
|
return any;
|
796 |
|
|
}
|
797 |
|
|
|
798 |
|
|
/* Determine whether there is any reference suitable for prefetching
|
799 |
|
|
in GROUPS. */
|
800 |
|
|
|
801 |
|
|
static bool
|
802 |
|
|
anything_to_prefetch_p (struct mem_ref_group *groups)
|
803 |
|
|
{
|
804 |
|
|
struct mem_ref *ref;
|
805 |
|
|
|
806 |
|
|
for (; groups; groups = groups->next)
|
807 |
|
|
for (ref = groups->refs; ref; ref = ref->next)
|
808 |
|
|
if (should_issue_prefetch_p (ref))
|
809 |
|
|
return true;
|
810 |
|
|
|
811 |
|
|
return false;
|
812 |
|
|
}
|
813 |
|
|
|
814 |
|
|
/* Issue prefetches for the reference REF into loop as decided before.
|
815 |
|
|
HEAD is the number of iterations to prefetch ahead. UNROLL_FACTOR
|
816 |
|
|
is the factor by which LOOP was unrolled. */
|
817 |
|
|
|
818 |
|
|
static void
|
819 |
|
|
issue_prefetch_ref (struct mem_ref *ref, unsigned unroll_factor, unsigned ahead)
|
820 |
|
|
{
|
821 |
|
|
HOST_WIDE_INT delta;
|
822 |
|
|
tree addr, addr_base, prefetch, params, write_p;
|
823 |
|
|
block_stmt_iterator bsi;
|
824 |
|
|
unsigned n_prefetches, ap;
|
825 |
|
|
|
826 |
|
|
if (dump_file && (dump_flags & TDF_DETAILS))
|
827 |
|
|
fprintf (dump_file, "Issued prefetch for %p.\n", (void *) ref);
|
828 |
|
|
|
829 |
|
|
bsi = bsi_for_stmt (ref->stmt);
|
830 |
|
|
|
831 |
|
|
n_prefetches = ((unroll_factor + ref->prefetch_mod - 1)
|
832 |
|
|
/ ref->prefetch_mod);
|
833 |
|
|
addr_base = build_fold_addr_expr_with_type (ref->mem, ptr_type_node);
|
834 |
|
|
addr_base = force_gimple_operand_bsi (&bsi, unshare_expr (addr_base), true, NULL);
|
835 |
|
|
|
836 |
|
|
for (ap = 0; ap < n_prefetches; ap++)
|
837 |
|
|
{
|
838 |
|
|
/* Determine the address to prefetch. */
|
839 |
|
|
delta = (ahead + ap * ref->prefetch_mod) * ref->group->step;
|
840 |
|
|
addr = fold_build2 (PLUS_EXPR, ptr_type_node,
|
841 |
|
|
addr_base, build_int_cst (ptr_type_node, delta));
|
842 |
|
|
addr = force_gimple_operand_bsi (&bsi, unshare_expr (addr), true, NULL);
|
843 |
|
|
|
844 |
|
|
/* Create the prefetch instruction. */
|
845 |
|
|
write_p = ref->write_p ? integer_one_node : integer_zero_node;
|
846 |
|
|
params = tree_cons (NULL_TREE, addr,
|
847 |
|
|
tree_cons (NULL_TREE, write_p, NULL_TREE));
|
848 |
|
|
|
849 |
|
|
prefetch = build_function_call_expr (built_in_decls[BUILT_IN_PREFETCH],
|
850 |
|
|
params);
|
851 |
|
|
bsi_insert_before (&bsi, prefetch, BSI_SAME_STMT);
|
852 |
|
|
}
|
853 |
|
|
}
|
854 |
|
|
|
855 |
|
|
/* Issue prefetches for the references in GROUPS into loop as decided before.
|
856 |
|
|
HEAD is the number of iterations to prefetch ahead. UNROLL_FACTOR is the
|
857 |
|
|
factor by that LOOP was unrolled. */
|
858 |
|
|
|
859 |
|
|
static void
|
860 |
|
|
issue_prefetches (struct mem_ref_group *groups,
|
861 |
|
|
unsigned unroll_factor, unsigned ahead)
|
862 |
|
|
{
|
863 |
|
|
struct mem_ref *ref;
|
864 |
|
|
|
865 |
|
|
for (; groups; groups = groups->next)
|
866 |
|
|
for (ref = groups->refs; ref; ref = ref->next)
|
867 |
|
|
if (ref->issue_prefetch_p)
|
868 |
|
|
issue_prefetch_ref (ref, unroll_factor, ahead);
|
869 |
|
|
}
|
870 |
|
|
|
871 |
|
|
/* Determines whether we can profitably unroll LOOP FACTOR times, and if
|
872 |
|
|
this is the case, fill in DESC by the description of number of
|
873 |
|
|
iterations. */
|
874 |
|
|
|
875 |
|
|
static bool
|
876 |
|
|
should_unroll_loop_p (struct loop *loop, struct tree_niter_desc *desc,
|
877 |
|
|
unsigned factor)
|
878 |
|
|
{
|
879 |
|
|
if (!can_unroll_loop_p (loop, factor, desc))
|
880 |
|
|
return false;
|
881 |
|
|
|
882 |
|
|
/* We only consider loops without control flow for unrolling. This is not
|
883 |
|
|
a hard restriction -- tree_unroll_loop works with arbitrary loops
|
884 |
|
|
as well; but the unrolling/prefetching is usually more profitable for
|
885 |
|
|
loops consisting of a single basic block, and we want to limit the
|
886 |
|
|
code growth. */
|
887 |
|
|
if (loop->num_nodes > 2)
|
888 |
|
|
return false;
|
889 |
|
|
|
890 |
|
|
return true;
|
891 |
|
|
}
|
892 |
|
|
|
893 |
|
|
/* Determine the coefficient by that unroll LOOP, from the information
|
894 |
|
|
contained in the list of memory references REFS. Description of
|
895 |
|
|
umber of iterations of LOOP is stored to DESC. AHEAD is the number
|
896 |
|
|
of iterations ahead that we need to prefetch. NINSNS is number of
|
897 |
|
|
insns of the LOOP. */
|
898 |
|
|
|
899 |
|
|
static unsigned
|
900 |
|
|
determine_unroll_factor (struct loop *loop, struct mem_ref_group *refs,
|
901 |
|
|
unsigned ahead, unsigned ninsns,
|
902 |
|
|
struct tree_niter_desc *desc)
|
903 |
|
|
{
|
904 |
|
|
unsigned upper_bound, size_factor, constraint_factor;
|
905 |
|
|
unsigned factor, max_mod_constraint, ahead_factor;
|
906 |
|
|
struct mem_ref_group *agp;
|
907 |
|
|
struct mem_ref *ref;
|
908 |
|
|
|
909 |
|
|
upper_bound = PARAM_VALUE (PARAM_MAX_UNROLL_TIMES);
|
910 |
|
|
|
911 |
|
|
/* First check whether the loop is not too large to unroll. */
|
912 |
|
|
size_factor = PARAM_VALUE (PARAM_MAX_UNROLLED_INSNS) / ninsns;
|
913 |
|
|
if (size_factor <= 1)
|
914 |
|
|
return 1;
|
915 |
|
|
|
916 |
|
|
if (size_factor < upper_bound)
|
917 |
|
|
upper_bound = size_factor;
|
918 |
|
|
|
919 |
|
|
max_mod_constraint = 1;
|
920 |
|
|
for (agp = refs; agp; agp = agp->next)
|
921 |
|
|
for (ref = agp->refs; ref; ref = ref->next)
|
922 |
|
|
if (should_issue_prefetch_p (ref)
|
923 |
|
|
&& ref->prefetch_mod > max_mod_constraint)
|
924 |
|
|
max_mod_constraint = ref->prefetch_mod;
|
925 |
|
|
|
926 |
|
|
/* Set constraint_factor as large as needed to be able to satisfy the
|
927 |
|
|
largest modulo constraint. */
|
928 |
|
|
constraint_factor = max_mod_constraint;
|
929 |
|
|
|
930 |
|
|
/* If ahead is too large in comparison with the number of available
|
931 |
|
|
prefetches, unroll the loop as much as needed to be able to prefetch
|
932 |
|
|
at least partially some of the references in the loop. */
|
933 |
|
|
ahead_factor = ((ahead + SIMULTANEOUS_PREFETCHES - 1)
|
934 |
|
|
/ SIMULTANEOUS_PREFETCHES);
|
935 |
|
|
|
936 |
|
|
/* Unroll as much as useful, but bound the code size growth. */
|
937 |
|
|
if (constraint_factor < ahead_factor)
|
938 |
|
|
factor = ahead_factor;
|
939 |
|
|
else
|
940 |
|
|
factor = constraint_factor;
|
941 |
|
|
if (factor > upper_bound)
|
942 |
|
|
factor = upper_bound;
|
943 |
|
|
|
944 |
|
|
if (!should_unroll_loop_p (loop, desc, factor))
|
945 |
|
|
return 1;
|
946 |
|
|
|
947 |
|
|
return factor;
|
948 |
|
|
}
|
949 |
|
|
|
950 |
|
|
/* Issue prefetch instructions for array references in LOOP. Returns
|
951 |
|
|
true if the LOOP was unrolled. LOOPS is the array containing all
|
952 |
|
|
loops. */
|
953 |
|
|
|
954 |
|
|
static bool
|
955 |
|
|
loop_prefetch_arrays (struct loops *loops, struct loop *loop)
|
956 |
|
|
{
|
957 |
|
|
struct mem_ref_group *refs;
|
958 |
|
|
unsigned ahead, ninsns, unroll_factor;
|
959 |
|
|
struct tree_niter_desc desc;
|
960 |
|
|
bool unrolled = false;
|
961 |
|
|
|
962 |
|
|
/* Step 1: gather the memory references. */
|
963 |
|
|
refs = gather_memory_references (loop);
|
964 |
|
|
|
965 |
|
|
/* Step 2: estimate the reuse effects. */
|
966 |
|
|
prune_by_reuse (refs);
|
967 |
|
|
|
968 |
|
|
if (!anything_to_prefetch_p (refs))
|
969 |
|
|
goto fail;
|
970 |
|
|
|
971 |
|
|
/* Step 3: determine the ahead and unroll factor. */
|
972 |
|
|
|
973 |
|
|
/* FIXME: We should use not size of the loop, but the average number of
|
974 |
|
|
instructions executed per iteration of the loop. */
|
975 |
|
|
ninsns = tree_num_loop_insns (loop);
|
976 |
|
|
ahead = (PREFETCH_LATENCY + ninsns - 1) / ninsns;
|
977 |
|
|
unroll_factor = determine_unroll_factor (loop, refs, ahead, ninsns,
|
978 |
|
|
&desc);
|
979 |
|
|
if (dump_file && (dump_flags & TDF_DETAILS))
|
980 |
|
|
fprintf (dump_file, "Ahead %d, unroll factor %d\n", ahead, unroll_factor);
|
981 |
|
|
|
982 |
|
|
/* If the loop rolls less than the required unroll factor, prefetching
|
983 |
|
|
is useless. */
|
984 |
|
|
if (unroll_factor > 1
|
985 |
|
|
&& cst_and_fits_in_hwi (desc.niter)
|
986 |
|
|
&& (unsigned HOST_WIDE_INT) int_cst_value (desc.niter) < unroll_factor)
|
987 |
|
|
goto fail;
|
988 |
|
|
|
989 |
|
|
/* Step 4: what to prefetch? */
|
990 |
|
|
if (!schedule_prefetches (refs, unroll_factor, ahead))
|
991 |
|
|
goto fail;
|
992 |
|
|
|
993 |
|
|
/* Step 5: unroll the loop. TODO -- peeling of first and last few
|
994 |
|
|
iterations so that we do not issue superfluous prefetches. */
|
995 |
|
|
if (unroll_factor != 1)
|
996 |
|
|
{
|
997 |
|
|
tree_unroll_loop (loops, loop, unroll_factor,
|
998 |
|
|
single_dom_exit (loop), &desc);
|
999 |
|
|
unrolled = true;
|
1000 |
|
|
}
|
1001 |
|
|
|
1002 |
|
|
/* Step 6: issue the prefetches. */
|
1003 |
|
|
issue_prefetches (refs, unroll_factor, ahead);
|
1004 |
|
|
|
1005 |
|
|
fail:
|
1006 |
|
|
release_mem_refs (refs);
|
1007 |
|
|
return unrolled;
|
1008 |
|
|
}
|
1009 |
|
|
|
1010 |
|
|
/* Issue prefetch instructions for array references in LOOPS. */
|
1011 |
|
|
|
1012 |
|
|
unsigned int
|
1013 |
|
|
tree_ssa_prefetch_arrays (struct loops *loops)
|
1014 |
|
|
{
|
1015 |
|
|
unsigned i;
|
1016 |
|
|
struct loop *loop;
|
1017 |
|
|
bool unrolled = false;
|
1018 |
|
|
int todo_flags = 0;
|
1019 |
|
|
|
1020 |
|
|
if (!HAVE_prefetch
|
1021 |
|
|
/* It is possible to ask compiler for say -mtune=i486 -march=pentium4.
|
1022 |
|
|
-mtune=i486 causes us having PREFETCH_BLOCK 0, since this is part
|
1023 |
|
|
of processor costs and i486 does not have prefetch, but
|
1024 |
|
|
-march=pentium4 causes HAVE_prefetch to be true. Ugh. */
|
1025 |
|
|
|| PREFETCH_BLOCK == 0)
|
1026 |
|
|
return 0;
|
1027 |
|
|
|
1028 |
|
|
initialize_original_copy_tables ();
|
1029 |
|
|
|
1030 |
|
|
if (!built_in_decls[BUILT_IN_PREFETCH])
|
1031 |
|
|
{
|
1032 |
|
|
tree type = build_function_type (void_type_node,
|
1033 |
|
|
tree_cons (NULL_TREE,
|
1034 |
|
|
const_ptr_type_node,
|
1035 |
|
|
NULL_TREE));
|
1036 |
|
|
tree decl = lang_hooks.builtin_function ("__builtin_prefetch", type,
|
1037 |
|
|
BUILT_IN_PREFETCH, BUILT_IN_NORMAL,
|
1038 |
|
|
NULL, NULL_TREE);
|
1039 |
|
|
DECL_IS_NOVOPS (decl) = true;
|
1040 |
|
|
built_in_decls[BUILT_IN_PREFETCH] = decl;
|
1041 |
|
|
}
|
1042 |
|
|
|
1043 |
|
|
/* We assume that size of cache line is a power of two, so verify this
|
1044 |
|
|
here. */
|
1045 |
|
|
gcc_assert ((PREFETCH_BLOCK & (PREFETCH_BLOCK - 1)) == 0);
|
1046 |
|
|
|
1047 |
|
|
for (i = loops->num - 1; i > 0; i--)
|
1048 |
|
|
{
|
1049 |
|
|
loop = loops->parray[i];
|
1050 |
|
|
|
1051 |
|
|
if (dump_file && (dump_flags & TDF_DETAILS))
|
1052 |
|
|
fprintf (dump_file, "Processing loop %d:\n", loop->num);
|
1053 |
|
|
|
1054 |
|
|
if (loop)
|
1055 |
|
|
unrolled |= loop_prefetch_arrays (loops, loop);
|
1056 |
|
|
|
1057 |
|
|
if (dump_file && (dump_flags & TDF_DETAILS))
|
1058 |
|
|
fprintf (dump_file, "\n\n");
|
1059 |
|
|
}
|
1060 |
|
|
|
1061 |
|
|
if (unrolled)
|
1062 |
|
|
{
|
1063 |
|
|
scev_reset ();
|
1064 |
|
|
todo_flags |= TODO_cleanup_cfg;
|
1065 |
|
|
}
|
1066 |
|
|
|
1067 |
|
|
free_original_copy_tables ();
|
1068 |
|
|
return todo_flags;
|
1069 |
|
|
}
|