/* Array prefetching.
|
/* Array prefetching.
|
Copyright (C) 2005, 2007, 2008 Free Software Foundation, Inc.
|
Copyright (C) 2005, 2007, 2008 Free Software Foundation, Inc.
|
|
|
This file is part of GCC.
|
This file is part of GCC.
|
|
|
GCC is free software; you can redistribute it and/or modify it
|
GCC is free software; you can redistribute it and/or modify it
|
under the terms of the GNU General Public License as published by the
|
under the terms of the GNU General Public License as published by the
|
Free Software Foundation; either version 3, or (at your option) any
|
Free Software Foundation; either version 3, or (at your option) any
|
later version.
|
later version.
|
|
|
GCC is distributed in the hope that it will be useful, but WITHOUT
|
GCC is distributed in the hope that it will be useful, but WITHOUT
|
ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
|
ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
|
FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
|
FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
|
for more details.
|
for more details.
|
|
|
You should have received a copy of the GNU General Public License
|
You should have received a copy of the GNU General Public License
|
along with GCC; see the file COPYING3. If not see
|
along with GCC; see the file COPYING3. If not see
|
<http://www.gnu.org/licenses/>. */
|
<http://www.gnu.org/licenses/>. */
|
|
|
#include "config.h"
|
#include "config.h"
|
#include "system.h"
|
#include "system.h"
|
#include "coretypes.h"
|
#include "coretypes.h"
|
#include "tm.h"
|
#include "tm.h"
|
#include "tree.h"
|
#include "tree.h"
|
#include "rtl.h"
|
#include "rtl.h"
|
#include "tm_p.h"
|
#include "tm_p.h"
|
#include "hard-reg-set.h"
|
#include "hard-reg-set.h"
|
#include "basic-block.h"
|
#include "basic-block.h"
|
#include "output.h"
|
#include "output.h"
|
#include "diagnostic.h"
|
#include "diagnostic.h"
|
#include "tree-flow.h"
|
#include "tree-flow.h"
|
#include "tree-dump.h"
|
#include "tree-dump.h"
|
#include "timevar.h"
|
#include "timevar.h"
|
#include "cfgloop.h"
|
#include "cfgloop.h"
|
#include "varray.h"
|
#include "varray.h"
|
#include "expr.h"
|
#include "expr.h"
|
#include "tree-pass.h"
|
#include "tree-pass.h"
|
#include "ggc.h"
|
#include "ggc.h"
|
#include "insn-config.h"
|
#include "insn-config.h"
|
#include "recog.h"
|
#include "recog.h"
|
#include "hashtab.h"
|
#include "hashtab.h"
|
#include "tree-chrec.h"
|
#include "tree-chrec.h"
|
#include "tree-scalar-evolution.h"
|
#include "tree-scalar-evolution.h"
|
#include "toplev.h"
|
#include "toplev.h"
|
#include "params.h"
|
#include "params.h"
|
#include "langhooks.h"
|
#include "langhooks.h"
|
#include "tree-inline.h"
|
#include "tree-inline.h"
|
#include "tree-data-ref.h"
|
#include "tree-data-ref.h"
|
#include "optabs.h"
|
#include "optabs.h"
|
|
|
/* This pass inserts prefetch instructions to optimize cache usage during
|
/* This pass inserts prefetch instructions to optimize cache usage during
|
accesses to arrays in loops. It processes loops sequentially and:
|
accesses to arrays in loops. It processes loops sequentially and:
|
|
|
1) Gathers all memory references in the single loop.
|
1) Gathers all memory references in the single loop.
|
2) For each of the references it decides when it is profitable to prefetch
|
2) For each of the references it decides when it is profitable to prefetch
|
it. To do it, we evaluate the reuse among the accesses, and determines
|
it. To do it, we evaluate the reuse among the accesses, and determines
|
two values: PREFETCH_BEFORE (meaning that it only makes sense to do
|
two values: PREFETCH_BEFORE (meaning that it only makes sense to do
|
prefetching in the first PREFETCH_BEFORE iterations of the loop) and
|
prefetching in the first PREFETCH_BEFORE iterations of the loop) and
|
PREFETCH_MOD (meaning that it only makes sense to prefetch in the
|
PREFETCH_MOD (meaning that it only makes sense to prefetch in the
|
iterations of the loop that are zero modulo PREFETCH_MOD). For example
|
iterations of the loop that are zero modulo PREFETCH_MOD). For example
|
(assuming cache line size is 64 bytes, char has size 1 byte and there
|
(assuming cache line size is 64 bytes, char has size 1 byte and there
|
is no hardware sequential prefetch):
|
is no hardware sequential prefetch):
|
|
|
char *a;
|
char *a;
|
for (i = 0; i < max; i++)
|
for (i = 0; i < max; i++)
|
{
|
{
|
a[255] = ...; (0)
|
a[255] = ...; (0)
|
a[i] = ...; (1)
|
a[i] = ...; (1)
|
a[i + 64] = ...; (2)
|
a[i + 64] = ...; (2)
|
a[16*i] = ...; (3)
|
a[16*i] = ...; (3)
|
a[187*i] = ...; (4)
|
a[187*i] = ...; (4)
|
a[187*i + 50] = ...; (5)
|
a[187*i + 50] = ...; (5)
|
}
|
}
|
|
|
(0) obviously has PREFETCH_BEFORE 1
|
(0) obviously has PREFETCH_BEFORE 1
|
(1) has PREFETCH_BEFORE 64, since (2) accesses the same memory
|
(1) has PREFETCH_BEFORE 64, since (2) accesses the same memory
|
location 64 iterations before it, and PREFETCH_MOD 64 (since
|
location 64 iterations before it, and PREFETCH_MOD 64 (since
|
it hits the same cache line otherwise).
|
it hits the same cache line otherwise).
|
(2) has PREFETCH_MOD 64
|
(2) has PREFETCH_MOD 64
|
(3) has PREFETCH_MOD 4
|
(3) has PREFETCH_MOD 4
|
(4) has PREFETCH_MOD 1. We do not set PREFETCH_BEFORE here, since
|
(4) has PREFETCH_MOD 1. We do not set PREFETCH_BEFORE here, since
|
the cache line accessed by (4) is the same with probability only
|
the cache line accessed by (4) is the same with probability only
|
7/32.
|
7/32.
|
(5) has PREFETCH_MOD 1 as well.
|
(5) has PREFETCH_MOD 1 as well.
|
|
|
Additionally, we use data dependence analysis to determine for each
|
Additionally, we use data dependence analysis to determine for each
|
reference the distance till the first reuse; this information is used
|
reference the distance till the first reuse; this information is used
|
to determine the temporality of the issued prefetch instruction.
|
to determine the temporality of the issued prefetch instruction.
|
|
|
3) We determine how much ahead we need to prefetch. The number of
|
3) We determine how much ahead we need to prefetch. The number of
|
iterations needed is time to fetch / time spent in one iteration of
|
iterations needed is time to fetch / time spent in one iteration of
|
the loop. The problem is that we do not know either of these values,
|
the loop. The problem is that we do not know either of these values,
|
so we just make a heuristic guess based on a magic (possibly)
|
so we just make a heuristic guess based on a magic (possibly)
|
target-specific constant and size of the loop.
|
target-specific constant and size of the loop.
|
|
|
4) Determine which of the references we prefetch. We take into account
|
4) Determine which of the references we prefetch. We take into account
|
that there is a maximum number of simultaneous prefetches (provided
|
that there is a maximum number of simultaneous prefetches (provided
|
by machine description). We prefetch as many prefetches as possible
|
by machine description). We prefetch as many prefetches as possible
|
while still within this bound (starting with those with lowest
|
while still within this bound (starting with those with lowest
|
prefetch_mod, since they are responsible for most of the cache
|
prefetch_mod, since they are responsible for most of the cache
|
misses).
|
misses).
|
|
|
5) We unroll and peel loops so that we are able to satisfy PREFETCH_MOD
|
5) We unroll and peel loops so that we are able to satisfy PREFETCH_MOD
|
and PREFETCH_BEFORE requirements (within some bounds), and to avoid
|
and PREFETCH_BEFORE requirements (within some bounds), and to avoid
|
prefetching nonaccessed memory.
|
prefetching nonaccessed memory.
|
TODO -- actually implement peeling.
|
TODO -- actually implement peeling.
|
|
|
6) We actually emit the prefetch instructions. ??? Perhaps emit the
|
6) We actually emit the prefetch instructions. ??? Perhaps emit the
|
prefetch instructions with guards in cases where 5) was not sufficient
|
prefetch instructions with guards in cases where 5) was not sufficient
|
to satisfy the constraints?
|
to satisfy the constraints?
|
|
|
The function is_loop_prefetching_profitable() implements a cost model
|
The function is_loop_prefetching_profitable() implements a cost model
|
to determine if prefetching is profitable for a given loop. The cost
|
to determine if prefetching is profitable for a given loop. The cost
|
model has two heuristcs:
|
model has two heuristcs:
|
1. A heuristic that determines whether the given loop has enough CPU
|
1. A heuristic that determines whether the given loop has enough CPU
|
ops that can be overlapped with cache missing memory ops.
|
ops that can be overlapped with cache missing memory ops.
|
If not, the loop won't benefit from prefetching. This is implemented
|
If not, the loop won't benefit from prefetching. This is implemented
|
by requirung the ratio between the instruction count and the mem ref
|
by requirung the ratio between the instruction count and the mem ref
|
count to be above a certain minimum.
|
count to be above a certain minimum.
|
2. A heuristic that disables prefetching in a loop with an unknown trip
|
2. A heuristic that disables prefetching in a loop with an unknown trip
|
count if the prefetching cost is above a certain limit. The relative
|
count if the prefetching cost is above a certain limit. The relative
|
prefetching cost is estimated by taking the ratio between the
|
prefetching cost is estimated by taking the ratio between the
|
prefetch count and the total intruction count (this models the I-cache
|
prefetch count and the total intruction count (this models the I-cache
|
cost).
|
cost).
|
The limits used in these heuristics are defined as parameters with
|
The limits used in these heuristics are defined as parameters with
|
reasonable default values. Machine-specific default values will be
|
reasonable default values. Machine-specific default values will be
|
added later.
|
added later.
|
|
|
Some other TODO:
|
Some other TODO:
|
-- write and use more general reuse analysis (that could be also used
|
-- write and use more general reuse analysis (that could be also used
|
in other cache aimed loop optimizations)
|
in other cache aimed loop optimizations)
|
-- make it behave sanely together with the prefetches given by user
|
-- make it behave sanely together with the prefetches given by user
|
(now we just ignore them; at the very least we should avoid
|
(now we just ignore them; at the very least we should avoid
|
optimizing loops in that user put his own prefetches)
|
optimizing loops in that user put his own prefetches)
|
-- we assume cache line size alignment of arrays; this could be
|
-- we assume cache line size alignment of arrays; this could be
|
improved. */
|
improved. */
|
|
|
/* Magic constants follow. These should be replaced by machine specific
|
/* Magic constants follow. These should be replaced by machine specific
|
numbers. */
|
numbers. */
|
|
|
/* True if write can be prefetched by a read prefetch. */
|
/* True if write can be prefetched by a read prefetch. */
|
|
|
#ifndef WRITE_CAN_USE_READ_PREFETCH
|
#ifndef WRITE_CAN_USE_READ_PREFETCH
|
#define WRITE_CAN_USE_READ_PREFETCH 1
|
#define WRITE_CAN_USE_READ_PREFETCH 1
|
#endif
|
#endif
|
|
|
/* True if read can be prefetched by a write prefetch. */
|
/* True if read can be prefetched by a write prefetch. */
|
|
|
#ifndef READ_CAN_USE_WRITE_PREFETCH
|
#ifndef READ_CAN_USE_WRITE_PREFETCH
|
#define READ_CAN_USE_WRITE_PREFETCH 0
|
#define READ_CAN_USE_WRITE_PREFETCH 0
|
#endif
|
#endif
|
|
|
/* The size of the block loaded by a single prefetch. Usually, this is
|
/* The size of the block loaded by a single prefetch. Usually, this is
|
the same as cache line size (at the moment, we only consider one level
|
the same as cache line size (at the moment, we only consider one level
|
of cache hierarchy). */
|
of cache hierarchy). */
|
|
|
#ifndef PREFETCH_BLOCK
|
#ifndef PREFETCH_BLOCK
|
#define PREFETCH_BLOCK L1_CACHE_LINE_SIZE
|
#define PREFETCH_BLOCK L1_CACHE_LINE_SIZE
|
#endif
|
#endif
|
|
|
/* Do we have a forward hardware sequential prefetching? */
|
/* Do we have a forward hardware sequential prefetching? */
|
|
|
#ifndef HAVE_FORWARD_PREFETCH
|
#ifndef HAVE_FORWARD_PREFETCH
|
#define HAVE_FORWARD_PREFETCH 0
|
#define HAVE_FORWARD_PREFETCH 0
|
#endif
|
#endif
|
|
|
/* Do we have a backward hardware sequential prefetching? */
|
/* Do we have a backward hardware sequential prefetching? */
|
|
|
#ifndef HAVE_BACKWARD_PREFETCH
|
#ifndef HAVE_BACKWARD_PREFETCH
|
#define HAVE_BACKWARD_PREFETCH 0
|
#define HAVE_BACKWARD_PREFETCH 0
|
#endif
|
#endif
|
|
|
/* In some cases we are only able to determine that there is a certain
|
/* In some cases we are only able to determine that there is a certain
|
probability that the two accesses hit the same cache line. In this
|
probability that the two accesses hit the same cache line. In this
|
case, we issue the prefetches for both of them if this probability
|
case, we issue the prefetches for both of them if this probability
|
is less then (1000 - ACCEPTABLE_MISS_RATE) per thousand. */
|
is less then (1000 - ACCEPTABLE_MISS_RATE) per thousand. */
|
|
|
#ifndef ACCEPTABLE_MISS_RATE
|
#ifndef ACCEPTABLE_MISS_RATE
|
#define ACCEPTABLE_MISS_RATE 50
|
#define ACCEPTABLE_MISS_RATE 50
|
#endif
|
#endif
|
|
|
#ifndef HAVE_prefetch
|
#ifndef HAVE_prefetch
|
#define HAVE_prefetch 0
|
#define HAVE_prefetch 0
|
#endif
|
#endif
|
|
|
#define L1_CACHE_SIZE_BYTES ((unsigned) (L1_CACHE_SIZE * 1024))
|
#define L1_CACHE_SIZE_BYTES ((unsigned) (L1_CACHE_SIZE * 1024))
|
#define L2_CACHE_SIZE_BYTES ((unsigned) (L2_CACHE_SIZE * 1024))
|
#define L2_CACHE_SIZE_BYTES ((unsigned) (L2_CACHE_SIZE * 1024))
|
|
|
/* We consider a memory access nontemporal if it is not reused sooner than
|
/* We consider a memory access nontemporal if it is not reused sooner than
|
after L2_CACHE_SIZE_BYTES of memory are accessed. However, we ignore
|
after L2_CACHE_SIZE_BYTES of memory are accessed. However, we ignore
|
accesses closer than L1_CACHE_SIZE_BYTES / NONTEMPORAL_FRACTION,
|
accesses closer than L1_CACHE_SIZE_BYTES / NONTEMPORAL_FRACTION,
|
so that we use nontemporal prefetches e.g. if single memory location
|
so that we use nontemporal prefetches e.g. if single memory location
|
is accessed several times in a single iteration of the loop. */
|
is accessed several times in a single iteration of the loop. */
|
#define NONTEMPORAL_FRACTION 16
|
#define NONTEMPORAL_FRACTION 16
|
|
|
/* In case we have to emit a memory fence instruction after the loop that
|
/* In case we have to emit a memory fence instruction after the loop that
|
uses nontemporal stores, this defines the builtin to use. */
|
uses nontemporal stores, this defines the builtin to use. */
|
|
|
#ifndef FENCE_FOLLOWING_MOVNT
|
#ifndef FENCE_FOLLOWING_MOVNT
|
#define FENCE_FOLLOWING_MOVNT NULL_TREE
|
#define FENCE_FOLLOWING_MOVNT NULL_TREE
|
#endif
|
#endif
|
|
|
/* The group of references between that reuse may occur. */
|
/* The group of references between that reuse may occur. */
|
|
|
struct mem_ref_group
|
struct mem_ref_group
|
{
|
{
|
tree base; /* Base of the reference. */
|
tree base; /* Base of the reference. */
|
HOST_WIDE_INT step; /* Step of the reference. */
|
HOST_WIDE_INT step; /* Step of the reference. */
|
struct mem_ref *refs; /* References in the group. */
|
struct mem_ref *refs; /* References in the group. */
|
struct mem_ref_group *next; /* Next group of references. */
|
struct mem_ref_group *next; /* Next group of references. */
|
};
|
};
|
|
|
/* Assigned to PREFETCH_BEFORE when all iterations are to be prefetched. */
|
/* Assigned to PREFETCH_BEFORE when all iterations are to be prefetched. */
|
|
|
#define PREFETCH_ALL (~(unsigned HOST_WIDE_INT) 0)
|
#define PREFETCH_ALL (~(unsigned HOST_WIDE_INT) 0)
|
|
|
/* The memory reference. */
|
/* The memory reference. */
|
|
|
struct mem_ref
|
struct mem_ref
|
{
|
{
|
gimple stmt; /* Statement in that the reference appears. */
|
gimple stmt; /* Statement in that the reference appears. */
|
tree mem; /* The reference. */
|
tree mem; /* The reference. */
|
HOST_WIDE_INT delta; /* Constant offset of the reference. */
|
HOST_WIDE_INT delta; /* Constant offset of the reference. */
|
struct mem_ref_group *group; /* The group of references it belongs to. */
|
struct mem_ref_group *group; /* The group of references it belongs to. */
|
unsigned HOST_WIDE_INT prefetch_mod;
|
unsigned HOST_WIDE_INT prefetch_mod;
|
/* Prefetch only each PREFETCH_MOD-th
|
/* Prefetch only each PREFETCH_MOD-th
|
iteration. */
|
iteration. */
|
unsigned HOST_WIDE_INT prefetch_before;
|
unsigned HOST_WIDE_INT prefetch_before;
|
/* Prefetch only first PREFETCH_BEFORE
|
/* Prefetch only first PREFETCH_BEFORE
|
iterations. */
|
iterations. */
|
unsigned reuse_distance; /* The amount of data accessed before the first
|
unsigned reuse_distance; /* The amount of data accessed before the first
|
reuse of this value. */
|
reuse of this value. */
|
struct mem_ref *next; /* The next reference in the group. */
|
struct mem_ref *next; /* The next reference in the group. */
|
unsigned write_p : 1; /* Is it a write? */
|
unsigned write_p : 1; /* Is it a write? */
|
unsigned independent_p : 1; /* True if the reference is independent on
|
unsigned independent_p : 1; /* True if the reference is independent on
|
all other references inside the loop. */
|
all other references inside the loop. */
|
unsigned issue_prefetch_p : 1; /* Should we really issue the prefetch? */
|
unsigned issue_prefetch_p : 1; /* Should we really issue the prefetch? */
|
unsigned storent_p : 1; /* True if we changed the store to a
|
unsigned storent_p : 1; /* True if we changed the store to a
|
nontemporal one. */
|
nontemporal one. */
|
};
|
};
|
|
|
/* Dumps information about reference REF to FILE. */
|
/* Dumps information about reference REF to FILE. */
|
|
|
static void
|
static void
|
dump_mem_ref (FILE *file, struct mem_ref *ref)
|
dump_mem_ref (FILE *file, struct mem_ref *ref)
|
{
|
{
|
fprintf (file, "Reference %p:\n", (void *) ref);
|
fprintf (file, "Reference %p:\n", (void *) ref);
|
|
|
fprintf (file, " group %p (base ", (void *) ref->group);
|
fprintf (file, " group %p (base ", (void *) ref->group);
|
print_generic_expr (file, ref->group->base, TDF_SLIM);
|
print_generic_expr (file, ref->group->base, TDF_SLIM);
|
fprintf (file, ", step ");
|
fprintf (file, ", step ");
|
fprintf (file, HOST_WIDE_INT_PRINT_DEC, ref->group->step);
|
fprintf (file, HOST_WIDE_INT_PRINT_DEC, ref->group->step);
|
fprintf (file, ")\n");
|
fprintf (file, ")\n");
|
|
|
fprintf (file, " delta ");
|
fprintf (file, " delta ");
|
fprintf (file, HOST_WIDE_INT_PRINT_DEC, ref->delta);
|
fprintf (file, HOST_WIDE_INT_PRINT_DEC, ref->delta);
|
fprintf (file, "\n");
|
fprintf (file, "\n");
|
|
|
fprintf (file, " %s\n", ref->write_p ? "write" : "read");
|
fprintf (file, " %s\n", ref->write_p ? "write" : "read");
|
|
|
fprintf (file, "\n");
|
fprintf (file, "\n");
|
}
|
}
|
|
|
/* Finds a group with BASE and STEP in GROUPS, or creates one if it does not
|
/* Finds a group with BASE and STEP in GROUPS, or creates one if it does not
|
exist. */
|
exist. */
|
|
|
static struct mem_ref_group *
|
static struct mem_ref_group *
|
find_or_create_group (struct mem_ref_group **groups, tree base,
|
find_or_create_group (struct mem_ref_group **groups, tree base,
|
HOST_WIDE_INT step)
|
HOST_WIDE_INT step)
|
{
|
{
|
struct mem_ref_group *group;
|
struct mem_ref_group *group;
|
|
|
for (; *groups; groups = &(*groups)->next)
|
for (; *groups; groups = &(*groups)->next)
|
{
|
{
|
if ((*groups)->step == step
|
if ((*groups)->step == step
|
&& operand_equal_p ((*groups)->base, base, 0))
|
&& operand_equal_p ((*groups)->base, base, 0))
|
return *groups;
|
return *groups;
|
|
|
/* Keep the list of groups sorted by decreasing step. */
|
/* Keep the list of groups sorted by decreasing step. */
|
if ((*groups)->step < step)
|
if ((*groups)->step < step)
|
break;
|
break;
|
}
|
}
|
|
|
group = XNEW (struct mem_ref_group);
|
group = XNEW (struct mem_ref_group);
|
group->base = base;
|
group->base = base;
|
group->step = step;
|
group->step = step;
|
group->refs = NULL;
|
group->refs = NULL;
|
group->next = *groups;
|
group->next = *groups;
|
*groups = group;
|
*groups = group;
|
|
|
return group;
|
return group;
|
}
|
}
|
|
|
/* Records a memory reference MEM in GROUP with offset DELTA and write status
|
/* Records a memory reference MEM in GROUP with offset DELTA and write status
|
WRITE_P. The reference occurs in statement STMT. */
|
WRITE_P. The reference occurs in statement STMT. */
|
|
|
static void
|
static void
|
record_ref (struct mem_ref_group *group, gimple stmt, tree mem,
|
record_ref (struct mem_ref_group *group, gimple stmt, tree mem,
|
HOST_WIDE_INT delta, bool write_p)
|
HOST_WIDE_INT delta, bool write_p)
|
{
|
{
|
struct mem_ref **aref;
|
struct mem_ref **aref;
|
|
|
/* Do not record the same address twice. */
|
/* Do not record the same address twice. */
|
for (aref = &group->refs; *aref; aref = &(*aref)->next)
|
for (aref = &group->refs; *aref; aref = &(*aref)->next)
|
{
|
{
|
/* It does not have to be possible for write reference to reuse the read
|
/* It does not have to be possible for write reference to reuse the read
|
prefetch, or vice versa. */
|
prefetch, or vice versa. */
|
if (!WRITE_CAN_USE_READ_PREFETCH
|
if (!WRITE_CAN_USE_READ_PREFETCH
|
&& write_p
|
&& write_p
|
&& !(*aref)->write_p)
|
&& !(*aref)->write_p)
|
continue;
|
continue;
|
if (!READ_CAN_USE_WRITE_PREFETCH
|
if (!READ_CAN_USE_WRITE_PREFETCH
|
&& !write_p
|
&& !write_p
|
&& (*aref)->write_p)
|
&& (*aref)->write_p)
|
continue;
|
continue;
|
|
|
if ((*aref)->delta == delta)
|
if ((*aref)->delta == delta)
|
return;
|
return;
|
}
|
}
|
|
|
(*aref) = XNEW (struct mem_ref);
|
(*aref) = XNEW (struct mem_ref);
|
(*aref)->stmt = stmt;
|
(*aref)->stmt = stmt;
|
(*aref)->mem = mem;
|
(*aref)->mem = mem;
|
(*aref)->delta = delta;
|
(*aref)->delta = delta;
|
(*aref)->write_p = write_p;
|
(*aref)->write_p = write_p;
|
(*aref)->prefetch_before = PREFETCH_ALL;
|
(*aref)->prefetch_before = PREFETCH_ALL;
|
(*aref)->prefetch_mod = 1;
|
(*aref)->prefetch_mod = 1;
|
(*aref)->reuse_distance = 0;
|
(*aref)->reuse_distance = 0;
|
(*aref)->issue_prefetch_p = false;
|
(*aref)->issue_prefetch_p = false;
|
(*aref)->group = group;
|
(*aref)->group = group;
|
(*aref)->next = NULL;
|
(*aref)->next = NULL;
|
(*aref)->independent_p = false;
|
(*aref)->independent_p = false;
|
(*aref)->storent_p = false;
|
(*aref)->storent_p = false;
|
|
|
if (dump_file && (dump_flags & TDF_DETAILS))
|
if (dump_file && (dump_flags & TDF_DETAILS))
|
dump_mem_ref (dump_file, *aref);
|
dump_mem_ref (dump_file, *aref);
|
}
|
}
|
|
|
/* Release memory references in GROUPS. */
|
/* Release memory references in GROUPS. */
|
|
|
static void
|
static void
|
release_mem_refs (struct mem_ref_group *groups)
|
release_mem_refs (struct mem_ref_group *groups)
|
{
|
{
|
struct mem_ref_group *next_g;
|
struct mem_ref_group *next_g;
|
struct mem_ref *ref, *next_r;
|
struct mem_ref *ref, *next_r;
|
|
|
for (; groups; groups = next_g)
|
for (; groups; groups = next_g)
|
{
|
{
|
next_g = groups->next;
|
next_g = groups->next;
|
for (ref = groups->refs; ref; ref = next_r)
|
for (ref = groups->refs; ref; ref = next_r)
|
{
|
{
|
next_r = ref->next;
|
next_r = ref->next;
|
free (ref);
|
free (ref);
|
}
|
}
|
free (groups);
|
free (groups);
|
}
|
}
|
}
|
}
|
|
|
/* A structure used to pass arguments to idx_analyze_ref. */
|
/* A structure used to pass arguments to idx_analyze_ref. */
|
|
|
struct ar_data
|
struct ar_data
|
{
|
{
|
struct loop *loop; /* Loop of the reference. */
|
struct loop *loop; /* Loop of the reference. */
|
gimple stmt; /* Statement of the reference. */
|
gimple stmt; /* Statement of the reference. */
|
HOST_WIDE_INT *step; /* Step of the memory reference. */
|
HOST_WIDE_INT *step; /* Step of the memory reference. */
|
HOST_WIDE_INT *delta; /* Offset of the memory reference. */
|
HOST_WIDE_INT *delta; /* Offset of the memory reference. */
|
};
|
};
|
|
|
/* Analyzes a single INDEX of a memory reference to obtain information
|
/* Analyzes a single INDEX of a memory reference to obtain information
|
described at analyze_ref. Callback for for_each_index. */
|
described at analyze_ref. Callback for for_each_index. */
|
|
|
static bool
|
static bool
|
idx_analyze_ref (tree base, tree *index, void *data)
|
idx_analyze_ref (tree base, tree *index, void *data)
|
{
|
{
|
struct ar_data *ar_data = (struct ar_data *) data;
|
struct ar_data *ar_data = (struct ar_data *) data;
|
tree ibase, step, stepsize;
|
tree ibase, step, stepsize;
|
HOST_WIDE_INT istep, idelta = 0, imult = 1;
|
HOST_WIDE_INT istep, idelta = 0, imult = 1;
|
affine_iv iv;
|
affine_iv iv;
|
|
|
if (TREE_CODE (base) == MISALIGNED_INDIRECT_REF
|
if (TREE_CODE (base) == MISALIGNED_INDIRECT_REF
|
|| TREE_CODE (base) == ALIGN_INDIRECT_REF)
|
|| TREE_CODE (base) == ALIGN_INDIRECT_REF)
|
return false;
|
return false;
|
|
|
if (!simple_iv (ar_data->loop, loop_containing_stmt (ar_data->stmt),
|
if (!simple_iv (ar_data->loop, loop_containing_stmt (ar_data->stmt),
|
*index, &iv, false))
|
*index, &iv, false))
|
return false;
|
return false;
|
ibase = iv.base;
|
ibase = iv.base;
|
step = iv.step;
|
step = iv.step;
|
|
|
if (!cst_and_fits_in_hwi (step))
|
if (!cst_and_fits_in_hwi (step))
|
return false;
|
return false;
|
istep = int_cst_value (step);
|
istep = int_cst_value (step);
|
|
|
if (TREE_CODE (ibase) == POINTER_PLUS_EXPR
|
if (TREE_CODE (ibase) == POINTER_PLUS_EXPR
|
&& cst_and_fits_in_hwi (TREE_OPERAND (ibase, 1)))
|
&& cst_and_fits_in_hwi (TREE_OPERAND (ibase, 1)))
|
{
|
{
|
idelta = int_cst_value (TREE_OPERAND (ibase, 1));
|
idelta = int_cst_value (TREE_OPERAND (ibase, 1));
|
ibase = TREE_OPERAND (ibase, 0);
|
ibase = TREE_OPERAND (ibase, 0);
|
}
|
}
|
if (cst_and_fits_in_hwi (ibase))
|
if (cst_and_fits_in_hwi (ibase))
|
{
|
{
|
idelta += int_cst_value (ibase);
|
idelta += int_cst_value (ibase);
|
ibase = build_int_cst (TREE_TYPE (ibase), 0);
|
ibase = build_int_cst (TREE_TYPE (ibase), 0);
|
}
|
}
|
|
|
if (TREE_CODE (base) == ARRAY_REF)
|
if (TREE_CODE (base) == ARRAY_REF)
|
{
|
{
|
stepsize = array_ref_element_size (base);
|
stepsize = array_ref_element_size (base);
|
if (!cst_and_fits_in_hwi (stepsize))
|
if (!cst_and_fits_in_hwi (stepsize))
|
return false;
|
return false;
|
imult = int_cst_value (stepsize);
|
imult = int_cst_value (stepsize);
|
|
|
istep *= imult;
|
istep *= imult;
|
idelta *= imult;
|
idelta *= imult;
|
}
|
}
|
|
|
*ar_data->step += istep;
|
*ar_data->step += istep;
|
*ar_data->delta += idelta;
|
*ar_data->delta += idelta;
|
*index = ibase;
|
*index = ibase;
|
|
|
return true;
|
return true;
|
}
|
}
|
|
|
/* Tries to express REF_P in shape &BASE + STEP * iter + DELTA, where DELTA and
|
/* Tries to express REF_P in shape &BASE + STEP * iter + DELTA, where DELTA and
|
STEP are integer constants and iter is number of iterations of LOOP. The
|
STEP are integer constants and iter is number of iterations of LOOP. The
|
reference occurs in statement STMT. Strips nonaddressable component
|
reference occurs in statement STMT. Strips nonaddressable component
|
references from REF_P. */
|
references from REF_P. */
|
|
|
static bool
|
static bool
|
analyze_ref (struct loop *loop, tree *ref_p, tree *base,
|
analyze_ref (struct loop *loop, tree *ref_p, tree *base,
|
HOST_WIDE_INT *step, HOST_WIDE_INT *delta,
|
HOST_WIDE_INT *step, HOST_WIDE_INT *delta,
|
gimple stmt)
|
gimple stmt)
|
{
|
{
|
struct ar_data ar_data;
|
struct ar_data ar_data;
|
tree off;
|
tree off;
|
HOST_WIDE_INT bit_offset;
|
HOST_WIDE_INT bit_offset;
|
tree ref = *ref_p;
|
tree ref = *ref_p;
|
|
|
*step = 0;
|
*step = 0;
|
*delta = 0;
|
*delta = 0;
|
|
|
/* First strip off the component references. Ignore bitfields. */
|
/* First strip off the component references. Ignore bitfields. */
|
if (TREE_CODE (ref) == COMPONENT_REF
|
if (TREE_CODE (ref) == COMPONENT_REF
|
&& DECL_NONADDRESSABLE_P (TREE_OPERAND (ref, 1)))
|
&& DECL_NONADDRESSABLE_P (TREE_OPERAND (ref, 1)))
|
ref = TREE_OPERAND (ref, 0);
|
ref = TREE_OPERAND (ref, 0);
|
|
|
*ref_p = ref;
|
*ref_p = ref;
|
|
|
for (; TREE_CODE (ref) == COMPONENT_REF; ref = TREE_OPERAND (ref, 0))
|
for (; TREE_CODE (ref) == COMPONENT_REF; ref = TREE_OPERAND (ref, 0))
|
{
|
{
|
off = DECL_FIELD_BIT_OFFSET (TREE_OPERAND (ref, 1));
|
off = DECL_FIELD_BIT_OFFSET (TREE_OPERAND (ref, 1));
|
bit_offset = TREE_INT_CST_LOW (off);
|
bit_offset = TREE_INT_CST_LOW (off);
|
gcc_assert (bit_offset % BITS_PER_UNIT == 0);
|
gcc_assert (bit_offset % BITS_PER_UNIT == 0);
|
|
|
*delta += bit_offset / BITS_PER_UNIT;
|
*delta += bit_offset / BITS_PER_UNIT;
|
}
|
}
|
|
|
*base = unshare_expr (ref);
|
*base = unshare_expr (ref);
|
ar_data.loop = loop;
|
ar_data.loop = loop;
|
ar_data.stmt = stmt;
|
ar_data.stmt = stmt;
|
ar_data.step = step;
|
ar_data.step = step;
|
ar_data.delta = delta;
|
ar_data.delta = delta;
|
return for_each_index (base, idx_analyze_ref, &ar_data);
|
return for_each_index (base, idx_analyze_ref, &ar_data);
|
}
|
}
|
|
|
/* Record a memory reference REF to the list REFS. The reference occurs in
|
/* Record a memory reference REF to the list REFS. The reference occurs in
|
LOOP in statement STMT and it is write if WRITE_P. Returns true if the
|
LOOP in statement STMT and it is write if WRITE_P. Returns true if the
|
reference was recorded, false otherwise. */
|
reference was recorded, false otherwise. */
|
|
|
static bool
|
static bool
|
gather_memory_references_ref (struct loop *loop, struct mem_ref_group **refs,
|
gather_memory_references_ref (struct loop *loop, struct mem_ref_group **refs,
|
tree ref, bool write_p, gimple stmt)
|
tree ref, bool write_p, gimple stmt)
|
{
|
{
|
tree base;
|
tree base;
|
HOST_WIDE_INT step, delta;
|
HOST_WIDE_INT step, delta;
|
struct mem_ref_group *agrp;
|
struct mem_ref_group *agrp;
|
|
|
if (get_base_address (ref) == NULL)
|
if (get_base_address (ref) == NULL)
|
return false;
|
return false;
|
|
|
if (!analyze_ref (loop, &ref, &base, &step, &delta, stmt))
|
if (!analyze_ref (loop, &ref, &base, &step, &delta, stmt))
|
return false;
|
return false;
|
|
|
/* Now we know that REF = &BASE + STEP * iter + DELTA, where DELTA and STEP
|
/* Now we know that REF = &BASE + STEP * iter + DELTA, where DELTA and STEP
|
are integer constants. */
|
are integer constants. */
|
agrp = find_or_create_group (refs, base, step);
|
agrp = find_or_create_group (refs, base, step);
|
record_ref (agrp, stmt, ref, delta, write_p);
|
record_ref (agrp, stmt, ref, delta, write_p);
|
|
|
return true;
|
return true;
|
}
|
}
|
|
|
/* Record the suitable memory references in LOOP. NO_OTHER_REFS is set to
|
/* Record the suitable memory references in LOOP. NO_OTHER_REFS is set to
|
true if there are no other memory references inside the loop. */
|
true if there are no other memory references inside the loop. */
|
|
|
static struct mem_ref_group *
|
static struct mem_ref_group *
|
gather_memory_references (struct loop *loop, bool *no_other_refs, unsigned *ref_count)
|
gather_memory_references (struct loop *loop, bool *no_other_refs, unsigned *ref_count)
|
{
|
{
|
basic_block *body = get_loop_body_in_dom_order (loop);
|
basic_block *body = get_loop_body_in_dom_order (loop);
|
basic_block bb;
|
basic_block bb;
|
unsigned i;
|
unsigned i;
|
gimple_stmt_iterator bsi;
|
gimple_stmt_iterator bsi;
|
gimple stmt;
|
gimple stmt;
|
tree lhs, rhs;
|
tree lhs, rhs;
|
struct mem_ref_group *refs = NULL;
|
struct mem_ref_group *refs = NULL;
|
|
|
*no_other_refs = true;
|
*no_other_refs = true;
|
*ref_count = 0;
|
*ref_count = 0;
|
|
|
/* Scan the loop body in order, so that the former references precede the
|
/* Scan the loop body in order, so that the former references precede the
|
later ones. */
|
later ones. */
|
for (i = 0; i < loop->num_nodes; i++)
|
for (i = 0; i < loop->num_nodes; i++)
|
{
|
{
|
bb = body[i];
|
bb = body[i];
|
if (bb->loop_father != loop)
|
if (bb->loop_father != loop)
|
continue;
|
continue;
|
|
|
for (bsi = gsi_start_bb (bb); !gsi_end_p (bsi); gsi_next (&bsi))
|
for (bsi = gsi_start_bb (bb); !gsi_end_p (bsi); gsi_next (&bsi))
|
{
|
{
|
stmt = gsi_stmt (bsi);
|
stmt = gsi_stmt (bsi);
|
|
|
if (gimple_code (stmt) != GIMPLE_ASSIGN)
|
if (gimple_code (stmt) != GIMPLE_ASSIGN)
|
{
|
{
|
if (gimple_vuse (stmt)
|
if (gimple_vuse (stmt)
|
|| (is_gimple_call (stmt)
|
|| (is_gimple_call (stmt)
|
&& !(gimple_call_flags (stmt) & ECF_CONST)))
|
&& !(gimple_call_flags (stmt) & ECF_CONST)))
|
*no_other_refs = false;
|
*no_other_refs = false;
|
continue;
|
continue;
|
}
|
}
|
|
|
lhs = gimple_assign_lhs (stmt);
|
lhs = gimple_assign_lhs (stmt);
|
rhs = gimple_assign_rhs1 (stmt);
|
rhs = gimple_assign_rhs1 (stmt);
|
|
|
if (REFERENCE_CLASS_P (rhs))
|
if (REFERENCE_CLASS_P (rhs))
|
{
|
{
|
*no_other_refs &= gather_memory_references_ref (loop, &refs,
|
*no_other_refs &= gather_memory_references_ref (loop, &refs,
|
rhs, false, stmt);
|
rhs, false, stmt);
|
*ref_count += 1;
|
*ref_count += 1;
|
}
|
}
|
if (REFERENCE_CLASS_P (lhs))
|
if (REFERENCE_CLASS_P (lhs))
|
{
|
{
|
*no_other_refs &= gather_memory_references_ref (loop, &refs,
|
*no_other_refs &= gather_memory_references_ref (loop, &refs,
|
lhs, true, stmt);
|
lhs, true, stmt);
|
*ref_count += 1;
|
*ref_count += 1;
|
}
|
}
|
}
|
}
|
}
|
}
|
free (body);
|
free (body);
|
|
|
return refs;
|
return refs;
|
}
|
}
|
|
|
/* Prune the prefetch candidate REF using the self-reuse. */
|
/* Prune the prefetch candidate REF using the self-reuse. */
|
|
|
static void
|
static void
|
prune_ref_by_self_reuse (struct mem_ref *ref)
|
prune_ref_by_self_reuse (struct mem_ref *ref)
|
{
|
{
|
HOST_WIDE_INT step = ref->group->step;
|
HOST_WIDE_INT step = ref->group->step;
|
bool backward = step < 0;
|
bool backward = step < 0;
|
|
|
if (step == 0)
|
if (step == 0)
|
{
|
{
|
/* Prefetch references to invariant address just once. */
|
/* Prefetch references to invariant address just once. */
|
ref->prefetch_before = 1;
|
ref->prefetch_before = 1;
|
return;
|
return;
|
}
|
}
|
|
|
if (backward)
|
if (backward)
|
step = -step;
|
step = -step;
|
|
|
if (step > PREFETCH_BLOCK)
|
if (step > PREFETCH_BLOCK)
|
return;
|
return;
|
|
|
if ((backward && HAVE_BACKWARD_PREFETCH)
|
if ((backward && HAVE_BACKWARD_PREFETCH)
|
|| (!backward && HAVE_FORWARD_PREFETCH))
|
|| (!backward && HAVE_FORWARD_PREFETCH))
|
{
|
{
|
ref->prefetch_before = 1;
|
ref->prefetch_before = 1;
|
return;
|
return;
|
}
|
}
|
|
|
ref->prefetch_mod = PREFETCH_BLOCK / step;
|
ref->prefetch_mod = PREFETCH_BLOCK / step;
|
}
|
}
|
|
|
/* Divides X by BY, rounding down. */
|
/* Divides X by BY, rounding down. */
|
|
|
static HOST_WIDE_INT
|
static HOST_WIDE_INT
|
ddown (HOST_WIDE_INT x, unsigned HOST_WIDE_INT by)
|
ddown (HOST_WIDE_INT x, unsigned HOST_WIDE_INT by)
|
{
|
{
|
gcc_assert (by > 0);
|
gcc_assert (by > 0);
|
|
|
if (x >= 0)
|
if (x >= 0)
|
return x / by;
|
return x / by;
|
else
|
else
|
return (x + by - 1) / by;
|
return (x + by - 1) / by;
|
}
|
}
|
|
|
/* Given a CACHE_LINE_SIZE and two inductive memory references
|
/* Given a CACHE_LINE_SIZE and two inductive memory references
|
with a common STEP greater than CACHE_LINE_SIZE and an address
|
with a common STEP greater than CACHE_LINE_SIZE and an address
|
difference DELTA, compute the probability that they will fall
|
difference DELTA, compute the probability that they will fall
|
in different cache lines. DISTINCT_ITERS is the number of
|
in different cache lines. DISTINCT_ITERS is the number of
|
distinct iterations after which the pattern repeats itself.
|
distinct iterations after which the pattern repeats itself.
|
ALIGN_UNIT is the unit of alignment in bytes. */
|
ALIGN_UNIT is the unit of alignment in bytes. */
|
|
|
static int
|
static int
|
compute_miss_rate (unsigned HOST_WIDE_INT cache_line_size,
|
compute_miss_rate (unsigned HOST_WIDE_INT cache_line_size,
|
HOST_WIDE_INT step, HOST_WIDE_INT delta,
|
HOST_WIDE_INT step, HOST_WIDE_INT delta,
|
unsigned HOST_WIDE_INT distinct_iters,
|
unsigned HOST_WIDE_INT distinct_iters,
|
int align_unit)
|
int align_unit)
|
{
|
{
|
unsigned align, iter;
|
unsigned align, iter;
|
int total_positions, miss_positions, miss_rate;
|
int total_positions, miss_positions, miss_rate;
|
int address1, address2, cache_line1, cache_line2;
|
int address1, address2, cache_line1, cache_line2;
|
|
|
total_positions = 0;
|
total_positions = 0;
|
miss_positions = 0;
|
miss_positions = 0;
|
|
|
/* Iterate through all possible alignments of the first
|
/* Iterate through all possible alignments of the first
|
memory reference within its cache line. */
|
memory reference within its cache line. */
|
for (align = 0; align < cache_line_size; align += align_unit)
|
for (align = 0; align < cache_line_size; align += align_unit)
|
|
|
/* Iterate through all distinct iterations. */
|
/* Iterate through all distinct iterations. */
|
for (iter = 0; iter < distinct_iters; iter++)
|
for (iter = 0; iter < distinct_iters; iter++)
|
{
|
{
|
address1 = align + step * iter;
|
address1 = align + step * iter;
|
address2 = address1 + delta;
|
address2 = address1 + delta;
|
cache_line1 = address1 / cache_line_size;
|
cache_line1 = address1 / cache_line_size;
|
cache_line2 = address2 / cache_line_size;
|
cache_line2 = address2 / cache_line_size;
|
total_positions += 1;
|
total_positions += 1;
|
if (cache_line1 != cache_line2)
|
if (cache_line1 != cache_line2)
|
miss_positions += 1;
|
miss_positions += 1;
|
}
|
}
|
miss_rate = 1000 * miss_positions / total_positions;
|
miss_rate = 1000 * miss_positions / total_positions;
|
return miss_rate;
|
return miss_rate;
|
}
|
}
|
|
|
/* Prune the prefetch candidate REF using the reuse with BY.
|
/* Prune the prefetch candidate REF using the reuse with BY.
|
If BY_IS_BEFORE is true, BY is before REF in the loop. */
|
If BY_IS_BEFORE is true, BY is before REF in the loop. */
|
|
|
static void
|
static void
|
prune_ref_by_group_reuse (struct mem_ref *ref, struct mem_ref *by,
|
prune_ref_by_group_reuse (struct mem_ref *ref, struct mem_ref *by,
|
bool by_is_before)
|
bool by_is_before)
|
{
|
{
|
HOST_WIDE_INT step = ref->group->step;
|
HOST_WIDE_INT step = ref->group->step;
|
bool backward = step < 0;
|
bool backward = step < 0;
|
HOST_WIDE_INT delta_r = ref->delta, delta_b = by->delta;
|
HOST_WIDE_INT delta_r = ref->delta, delta_b = by->delta;
|
HOST_WIDE_INT delta = delta_b - delta_r;
|
HOST_WIDE_INT delta = delta_b - delta_r;
|
HOST_WIDE_INT hit_from;
|
HOST_WIDE_INT hit_from;
|
unsigned HOST_WIDE_INT prefetch_before, prefetch_block;
|
unsigned HOST_WIDE_INT prefetch_before, prefetch_block;
|
int miss_rate;
|
int miss_rate;
|
HOST_WIDE_INT reduced_step;
|
HOST_WIDE_INT reduced_step;
|
unsigned HOST_WIDE_INT reduced_prefetch_block;
|
unsigned HOST_WIDE_INT reduced_prefetch_block;
|
tree ref_type;
|
tree ref_type;
|
int align_unit;
|
int align_unit;
|
|
|
if (delta == 0)
|
if (delta == 0)
|
{
|
{
|
/* If the references has the same address, only prefetch the
|
/* If the references has the same address, only prefetch the
|
former. */
|
former. */
|
if (by_is_before)
|
if (by_is_before)
|
ref->prefetch_before = 0;
|
ref->prefetch_before = 0;
|
|
|
return;
|
return;
|
}
|
}
|
|
|
if (!step)
|
if (!step)
|
{
|
{
|
/* If the reference addresses are invariant and fall into the
|
/* If the reference addresses are invariant and fall into the
|
same cache line, prefetch just the first one. */
|
same cache line, prefetch just the first one. */
|
if (!by_is_before)
|
if (!by_is_before)
|
return;
|
return;
|
|
|
if (ddown (ref->delta, PREFETCH_BLOCK)
|
if (ddown (ref->delta, PREFETCH_BLOCK)
|
!= ddown (by->delta, PREFETCH_BLOCK))
|
!= ddown (by->delta, PREFETCH_BLOCK))
|
return;
|
return;
|
|
|
ref->prefetch_before = 0;
|
ref->prefetch_before = 0;
|
return;
|
return;
|
}
|
}
|
|
|
/* Only prune the reference that is behind in the array. */
|
/* Only prune the reference that is behind in the array. */
|
if (backward)
|
if (backward)
|
{
|
{
|
if (delta > 0)
|
if (delta > 0)
|
return;
|
return;
|
|
|
/* Transform the data so that we may assume that the accesses
|
/* Transform the data so that we may assume that the accesses
|
are forward. */
|
are forward. */
|
delta = - delta;
|
delta = - delta;
|
step = -step;
|
step = -step;
|
delta_r = PREFETCH_BLOCK - 1 - delta_r;
|
delta_r = PREFETCH_BLOCK - 1 - delta_r;
|
delta_b = PREFETCH_BLOCK - 1 - delta_b;
|
delta_b = PREFETCH_BLOCK - 1 - delta_b;
|
}
|
}
|
else
|
else
|
{
|
{
|
if (delta < 0)
|
if (delta < 0)
|
return;
|
return;
|
}
|
}
|
|
|
/* Check whether the two references are likely to hit the same cache
|
/* Check whether the two references are likely to hit the same cache
|
line, and how distant the iterations in that it occurs are from
|
line, and how distant the iterations in that it occurs are from
|
each other. */
|
each other. */
|
|
|
if (step <= PREFETCH_BLOCK)
|
if (step <= PREFETCH_BLOCK)
|
{
|
{
|
/* The accesses are sure to meet. Let us check when. */
|
/* The accesses are sure to meet. Let us check when. */
|
hit_from = ddown (delta_b, PREFETCH_BLOCK) * PREFETCH_BLOCK;
|
hit_from = ddown (delta_b, PREFETCH_BLOCK) * PREFETCH_BLOCK;
|
prefetch_before = (hit_from - delta_r + step - 1) / step;
|
prefetch_before = (hit_from - delta_r + step - 1) / step;
|
|
|
if (prefetch_before < ref->prefetch_before)
|
if (prefetch_before < ref->prefetch_before)
|
ref->prefetch_before = prefetch_before;
|
ref->prefetch_before = prefetch_before;
|
|
|
return;
|
return;
|
}
|
}
|
|
|
/* A more complicated case with step > prefetch_block. First reduce
|
/* A more complicated case with step > prefetch_block. First reduce
|
the ratio between the step and the cache line size to its simplest
|
the ratio between the step and the cache line size to its simplest
|
terms. The resulting denominator will then represent the number of
|
terms. The resulting denominator will then represent the number of
|
distinct iterations after which each address will go back to its
|
distinct iterations after which each address will go back to its
|
initial location within the cache line. This computation assumes
|
initial location within the cache line. This computation assumes
|
that PREFETCH_BLOCK is a power of two. */
|
that PREFETCH_BLOCK is a power of two. */
|
prefetch_block = PREFETCH_BLOCK;
|
prefetch_block = PREFETCH_BLOCK;
|
reduced_prefetch_block = prefetch_block;
|
reduced_prefetch_block = prefetch_block;
|
reduced_step = step;
|
reduced_step = step;
|
while ((reduced_step & 1) == 0
|
while ((reduced_step & 1) == 0
|
&& reduced_prefetch_block > 1)
|
&& reduced_prefetch_block > 1)
|
{
|
{
|
reduced_step >>= 1;
|
reduced_step >>= 1;
|
reduced_prefetch_block >>= 1;
|
reduced_prefetch_block >>= 1;
|
}
|
}
|
|
|
prefetch_before = delta / step;
|
prefetch_before = delta / step;
|
delta %= step;
|
delta %= step;
|
ref_type = TREE_TYPE (ref->mem);
|
ref_type = TREE_TYPE (ref->mem);
|
align_unit = TYPE_ALIGN (ref_type) / 8;
|
align_unit = TYPE_ALIGN (ref_type) / 8;
|
miss_rate = compute_miss_rate(prefetch_block, step, delta,
|
miss_rate = compute_miss_rate(prefetch_block, step, delta,
|
reduced_prefetch_block, align_unit);
|
reduced_prefetch_block, align_unit);
|
if (miss_rate <= ACCEPTABLE_MISS_RATE)
|
if (miss_rate <= ACCEPTABLE_MISS_RATE)
|
{
|
{
|
if (prefetch_before < ref->prefetch_before)
|
if (prefetch_before < ref->prefetch_before)
|
ref->prefetch_before = prefetch_before;
|
ref->prefetch_before = prefetch_before;
|
|
|
return;
|
return;
|
}
|
}
|
|
|
/* Try also the following iteration. */
|
/* Try also the following iteration. */
|
prefetch_before++;
|
prefetch_before++;
|
delta = step - delta;
|
delta = step - delta;
|
miss_rate = compute_miss_rate(prefetch_block, step, delta,
|
miss_rate = compute_miss_rate(prefetch_block, step, delta,
|
reduced_prefetch_block, align_unit);
|
reduced_prefetch_block, align_unit);
|
if (miss_rate <= ACCEPTABLE_MISS_RATE)
|
if (miss_rate <= ACCEPTABLE_MISS_RATE)
|
{
|
{
|
if (prefetch_before < ref->prefetch_before)
|
if (prefetch_before < ref->prefetch_before)
|
ref->prefetch_before = prefetch_before;
|
ref->prefetch_before = prefetch_before;
|
|
|
return;
|
return;
|
}
|
}
|
|
|
/* The ref probably does not reuse by. */
|
/* The ref probably does not reuse by. */
|
return;
|
return;
|
}
|
}
|
|
|
/* Prune the prefetch candidate REF using the reuses with other references
|
/* Prune the prefetch candidate REF using the reuses with other references
|
in REFS. */
|
in REFS. */
|
|
|
static void
|
static void
|
prune_ref_by_reuse (struct mem_ref *ref, struct mem_ref *refs)
|
prune_ref_by_reuse (struct mem_ref *ref, struct mem_ref *refs)
|
{
|
{
|
struct mem_ref *prune_by;
|
struct mem_ref *prune_by;
|
bool before = true;
|
bool before = true;
|
|
|
prune_ref_by_self_reuse (ref);
|
prune_ref_by_self_reuse (ref);
|
|
|
for (prune_by = refs; prune_by; prune_by = prune_by->next)
|
for (prune_by = refs; prune_by; prune_by = prune_by->next)
|
{
|
{
|
if (prune_by == ref)
|
if (prune_by == ref)
|
{
|
{
|
before = false;
|
before = false;
|
continue;
|
continue;
|
}
|
}
|
|
|
if (!WRITE_CAN_USE_READ_PREFETCH
|
if (!WRITE_CAN_USE_READ_PREFETCH
|
&& ref->write_p
|
&& ref->write_p
|
&& !prune_by->write_p)
|
&& !prune_by->write_p)
|
continue;
|
continue;
|
if (!READ_CAN_USE_WRITE_PREFETCH
|
if (!READ_CAN_USE_WRITE_PREFETCH
|
&& !ref->write_p
|
&& !ref->write_p
|
&& prune_by->write_p)
|
&& prune_by->write_p)
|
continue;
|
continue;
|
|
|
prune_ref_by_group_reuse (ref, prune_by, before);
|
prune_ref_by_group_reuse (ref, prune_by, before);
|
}
|
}
|
}
|
}
|
|
|
/* Prune the prefetch candidates in GROUP using the reuse analysis. */
|
/* Prune the prefetch candidates in GROUP using the reuse analysis. */
|
|
|
static void
|
static void
|
prune_group_by_reuse (struct mem_ref_group *group)
|
prune_group_by_reuse (struct mem_ref_group *group)
|
{
|
{
|
struct mem_ref *ref_pruned;
|
struct mem_ref *ref_pruned;
|
|
|
for (ref_pruned = group->refs; ref_pruned; ref_pruned = ref_pruned->next)
|
for (ref_pruned = group->refs; ref_pruned; ref_pruned = ref_pruned->next)
|
{
|
{
|
prune_ref_by_reuse (ref_pruned, group->refs);
|
prune_ref_by_reuse (ref_pruned, group->refs);
|
|
|
if (dump_file && (dump_flags & TDF_DETAILS))
|
if (dump_file && (dump_flags & TDF_DETAILS))
|
{
|
{
|
fprintf (dump_file, "Reference %p:", (void *) ref_pruned);
|
fprintf (dump_file, "Reference %p:", (void *) ref_pruned);
|
|
|
if (ref_pruned->prefetch_before == PREFETCH_ALL
|
if (ref_pruned->prefetch_before == PREFETCH_ALL
|
&& ref_pruned->prefetch_mod == 1)
|
&& ref_pruned->prefetch_mod == 1)
|
fprintf (dump_file, " no restrictions");
|
fprintf (dump_file, " no restrictions");
|
else if (ref_pruned->prefetch_before == 0)
|
else if (ref_pruned->prefetch_before == 0)
|
fprintf (dump_file, " do not prefetch");
|
fprintf (dump_file, " do not prefetch");
|
else if (ref_pruned->prefetch_before <= ref_pruned->prefetch_mod)
|
else if (ref_pruned->prefetch_before <= ref_pruned->prefetch_mod)
|
fprintf (dump_file, " prefetch once");
|
fprintf (dump_file, " prefetch once");
|
else
|
else
|
{
|
{
|
if (ref_pruned->prefetch_before != PREFETCH_ALL)
|
if (ref_pruned->prefetch_before != PREFETCH_ALL)
|
{
|
{
|
fprintf (dump_file, " prefetch before ");
|
fprintf (dump_file, " prefetch before ");
|
fprintf (dump_file, HOST_WIDE_INT_PRINT_DEC,
|
fprintf (dump_file, HOST_WIDE_INT_PRINT_DEC,
|
ref_pruned->prefetch_before);
|
ref_pruned->prefetch_before);
|
}
|
}
|
if (ref_pruned->prefetch_mod != 1)
|
if (ref_pruned->prefetch_mod != 1)
|
{
|
{
|
fprintf (dump_file, " prefetch mod ");
|
fprintf (dump_file, " prefetch mod ");
|
fprintf (dump_file, HOST_WIDE_INT_PRINT_DEC,
|
fprintf (dump_file, HOST_WIDE_INT_PRINT_DEC,
|
ref_pruned->prefetch_mod);
|
ref_pruned->prefetch_mod);
|
}
|
}
|
}
|
}
|
fprintf (dump_file, "\n");
|
fprintf (dump_file, "\n");
|
}
|
}
|
}
|
}
|
}
|
}
|
|
|
/* Prune the list of prefetch candidates GROUPS using the reuse analysis. */
|
/* Prune the list of prefetch candidates GROUPS using the reuse analysis. */
|
|
|
static void
|
static void
|
prune_by_reuse (struct mem_ref_group *groups)
|
prune_by_reuse (struct mem_ref_group *groups)
|
{
|
{
|
for (; groups; groups = groups->next)
|
for (; groups; groups = groups->next)
|
prune_group_by_reuse (groups);
|
prune_group_by_reuse (groups);
|
}
|
}
|
|
|
/* Returns true if we should issue prefetch for REF. */
|
/* Returns true if we should issue prefetch for REF. */
|
|
|
static bool
|
static bool
|
should_issue_prefetch_p (struct mem_ref *ref)
|
should_issue_prefetch_p (struct mem_ref *ref)
|
{
|
{
|
/* For now do not issue prefetches for only first few of the
|
/* For now do not issue prefetches for only first few of the
|
iterations. */
|
iterations. */
|
if (ref->prefetch_before != PREFETCH_ALL)
|
if (ref->prefetch_before != PREFETCH_ALL)
|
return false;
|
return false;
|
|
|
/* Do not prefetch nontemporal stores. */
|
/* Do not prefetch nontemporal stores. */
|
if (ref->storent_p)
|
if (ref->storent_p)
|
return false;
|
return false;
|
|
|
return true;
|
return true;
|
}
|
}
|
|
|
/* Decide which of the prefetch candidates in GROUPS to prefetch.
|
/* Decide which of the prefetch candidates in GROUPS to prefetch.
|
AHEAD is the number of iterations to prefetch ahead (which corresponds
|
AHEAD is the number of iterations to prefetch ahead (which corresponds
|
to the number of simultaneous instances of one prefetch running at a
|
to the number of simultaneous instances of one prefetch running at a
|
time). UNROLL_FACTOR is the factor by that the loop is going to be
|
time). UNROLL_FACTOR is the factor by that the loop is going to be
|
unrolled. Returns true if there is anything to prefetch. */
|
unrolled. Returns true if there is anything to prefetch. */
|
|
|
static bool
|
static bool
|
schedule_prefetches (struct mem_ref_group *groups, unsigned unroll_factor,
|
schedule_prefetches (struct mem_ref_group *groups, unsigned unroll_factor,
|
unsigned ahead)
|
unsigned ahead)
|
{
|
{
|
unsigned remaining_prefetch_slots, n_prefetches, prefetch_slots;
|
unsigned remaining_prefetch_slots, n_prefetches, prefetch_slots;
|
unsigned slots_per_prefetch;
|
unsigned slots_per_prefetch;
|
struct mem_ref *ref;
|
struct mem_ref *ref;
|
bool any = false;
|
bool any = false;
|
|
|
/* At most SIMULTANEOUS_PREFETCHES should be running at the same time. */
|
/* At most SIMULTANEOUS_PREFETCHES should be running at the same time. */
|
remaining_prefetch_slots = SIMULTANEOUS_PREFETCHES;
|
remaining_prefetch_slots = SIMULTANEOUS_PREFETCHES;
|
|
|
/* The prefetch will run for AHEAD iterations of the original loop, i.e.,
|
/* The prefetch will run for AHEAD iterations of the original loop, i.e.,
|
AHEAD / UNROLL_FACTOR iterations of the unrolled loop. In each iteration,
|
AHEAD / UNROLL_FACTOR iterations of the unrolled loop. In each iteration,
|
it will need a prefetch slot. */
|
it will need a prefetch slot. */
|
slots_per_prefetch = (ahead + unroll_factor / 2) / unroll_factor;
|
slots_per_prefetch = (ahead + unroll_factor / 2) / unroll_factor;
|
if (dump_file && (dump_flags & TDF_DETAILS))
|
if (dump_file && (dump_flags & TDF_DETAILS))
|
fprintf (dump_file, "Each prefetch instruction takes %u prefetch slots.\n",
|
fprintf (dump_file, "Each prefetch instruction takes %u prefetch slots.\n",
|
slots_per_prefetch);
|
slots_per_prefetch);
|
|
|
/* For now we just take memory references one by one and issue
|
/* For now we just take memory references one by one and issue
|
prefetches for as many as possible. The groups are sorted
|
prefetches for as many as possible. The groups are sorted
|
starting with the largest step, since the references with
|
starting with the largest step, since the references with
|
large step are more likely to cause many cache misses. */
|
large step are more likely to cause many cache misses. */
|
|
|
for (; groups; groups = groups->next)
|
for (; groups; groups = groups->next)
|
for (ref = groups->refs; ref; ref = ref->next)
|
for (ref = groups->refs; ref; ref = ref->next)
|
{
|
{
|
if (!should_issue_prefetch_p (ref))
|
if (!should_issue_prefetch_p (ref))
|
continue;
|
continue;
|
|
|
/* If we need to prefetch the reference each PREFETCH_MOD iterations,
|
/* If we need to prefetch the reference each PREFETCH_MOD iterations,
|
and we unroll the loop UNROLL_FACTOR times, we need to insert
|
and we unroll the loop UNROLL_FACTOR times, we need to insert
|
ceil (UNROLL_FACTOR / PREFETCH_MOD) instructions in each
|
ceil (UNROLL_FACTOR / PREFETCH_MOD) instructions in each
|
iteration. */
|
iteration. */
|
n_prefetches = ((unroll_factor + ref->prefetch_mod - 1)
|
n_prefetches = ((unroll_factor + ref->prefetch_mod - 1)
|
/ ref->prefetch_mod);
|
/ ref->prefetch_mod);
|
prefetch_slots = n_prefetches * slots_per_prefetch;
|
prefetch_slots = n_prefetches * slots_per_prefetch;
|
|
|
/* If more than half of the prefetches would be lost anyway, do not
|
/* If more than half of the prefetches would be lost anyway, do not
|
issue the prefetch. */
|
issue the prefetch. */
|
if (2 * remaining_prefetch_slots < prefetch_slots)
|
if (2 * remaining_prefetch_slots < prefetch_slots)
|
continue;
|
continue;
|
|
|
ref->issue_prefetch_p = true;
|
ref->issue_prefetch_p = true;
|
|
|
if (remaining_prefetch_slots <= prefetch_slots)
|
if (remaining_prefetch_slots <= prefetch_slots)
|
return true;
|
return true;
|
remaining_prefetch_slots -= prefetch_slots;
|
remaining_prefetch_slots -= prefetch_slots;
|
any = true;
|
any = true;
|
}
|
}
|
|
|
return any;
|
return any;
|
}
|
}
|
|
|
/* Estimate the number of prefetches in the given GROUPS. */
|
/* Estimate the number of prefetches in the given GROUPS. */
|
|
|
static int
|
static int
|
estimate_prefetch_count (struct mem_ref_group *groups)
|
estimate_prefetch_count (struct mem_ref_group *groups)
|
{
|
{
|
struct mem_ref *ref;
|
struct mem_ref *ref;
|
int prefetch_count = 0;
|
int prefetch_count = 0;
|
|
|
for (; groups; groups = groups->next)
|
for (; groups; groups = groups->next)
|
for (ref = groups->refs; ref; ref = ref->next)
|
for (ref = groups->refs; ref; ref = ref->next)
|
if (should_issue_prefetch_p (ref))
|
if (should_issue_prefetch_p (ref))
|
prefetch_count++;
|
prefetch_count++;
|
|
|
return prefetch_count;
|
return prefetch_count;
|
}
|
}
|
|
|
/* Issue prefetches for the reference REF into loop as decided before.
|
/* Issue prefetches for the reference REF into loop as decided before.
|
HEAD is the number of iterations to prefetch ahead. UNROLL_FACTOR
|
HEAD is the number of iterations to prefetch ahead. UNROLL_FACTOR
|
is the factor by which LOOP was unrolled. */
|
is the factor by which LOOP was unrolled. */
|
|
|
static void
|
static void
|
issue_prefetch_ref (struct mem_ref *ref, unsigned unroll_factor, unsigned ahead)
|
issue_prefetch_ref (struct mem_ref *ref, unsigned unroll_factor, unsigned ahead)
|
{
|
{
|
HOST_WIDE_INT delta;
|
HOST_WIDE_INT delta;
|
tree addr, addr_base, write_p, local;
|
tree addr, addr_base, write_p, local;
|
gimple prefetch;
|
gimple prefetch;
|
gimple_stmt_iterator bsi;
|
gimple_stmt_iterator bsi;
|
unsigned n_prefetches, ap;
|
unsigned n_prefetches, ap;
|
bool nontemporal = ref->reuse_distance >= L2_CACHE_SIZE_BYTES;
|
bool nontemporal = ref->reuse_distance >= L2_CACHE_SIZE_BYTES;
|
|
|
if (dump_file && (dump_flags & TDF_DETAILS))
|
if (dump_file && (dump_flags & TDF_DETAILS))
|
fprintf (dump_file, "Issued%s prefetch for %p.\n",
|
fprintf (dump_file, "Issued%s prefetch for %p.\n",
|
nontemporal ? " nontemporal" : "",
|
nontemporal ? " nontemporal" : "",
|
(void *) ref);
|
(void *) ref);
|
|
|
bsi = gsi_for_stmt (ref->stmt);
|
bsi = gsi_for_stmt (ref->stmt);
|
|
|
n_prefetches = ((unroll_factor + ref->prefetch_mod - 1)
|
n_prefetches = ((unroll_factor + ref->prefetch_mod - 1)
|
/ ref->prefetch_mod);
|
/ ref->prefetch_mod);
|
addr_base = build_fold_addr_expr_with_type (ref->mem, ptr_type_node);
|
addr_base = build_fold_addr_expr_with_type (ref->mem, ptr_type_node);
|
addr_base = force_gimple_operand_gsi (&bsi, unshare_expr (addr_base),
|
addr_base = force_gimple_operand_gsi (&bsi, unshare_expr (addr_base),
|
true, NULL, true, GSI_SAME_STMT);
|
true, NULL, true, GSI_SAME_STMT);
|
write_p = ref->write_p ? integer_one_node : integer_zero_node;
|
write_p = ref->write_p ? integer_one_node : integer_zero_node;
|
local = build_int_cst (integer_type_node, nontemporal ? 0 : 3);
|
local = build_int_cst (integer_type_node, nontemporal ? 0 : 3);
|
|
|
for (ap = 0; ap < n_prefetches; ap++)
|
for (ap = 0; ap < n_prefetches; ap++)
|
{
|
{
|
/* Determine the address to prefetch. */
|
/* Determine the address to prefetch. */
|
delta = (ahead + ap * ref->prefetch_mod) * ref->group->step;
|
delta = (ahead + ap * ref->prefetch_mod) * ref->group->step;
|
addr = fold_build2 (POINTER_PLUS_EXPR, ptr_type_node,
|
addr = fold_build2 (POINTER_PLUS_EXPR, ptr_type_node,
|
addr_base, size_int (delta));
|
addr_base, size_int (delta));
|
addr = force_gimple_operand_gsi (&bsi, unshare_expr (addr), true, NULL,
|
addr = force_gimple_operand_gsi (&bsi, unshare_expr (addr), true, NULL,
|
true, GSI_SAME_STMT);
|
true, GSI_SAME_STMT);
|
|
|
/* Create the prefetch instruction. */
|
/* Create the prefetch instruction. */
|
prefetch = gimple_build_call (built_in_decls[BUILT_IN_PREFETCH],
|
prefetch = gimple_build_call (built_in_decls[BUILT_IN_PREFETCH],
|
3, addr, write_p, local);
|
3, addr, write_p, local);
|
gsi_insert_before (&bsi, prefetch, GSI_SAME_STMT);
|
gsi_insert_before (&bsi, prefetch, GSI_SAME_STMT);
|
}
|
}
|
}
|
}
|
|
|
/* Issue prefetches for the references in GROUPS into loop as decided before.
|
/* Issue prefetches for the references in GROUPS into loop as decided before.
|
HEAD is the number of iterations to prefetch ahead. UNROLL_FACTOR is the
|
HEAD is the number of iterations to prefetch ahead. UNROLL_FACTOR is the
|
factor by that LOOP was unrolled. */
|
factor by that LOOP was unrolled. */
|
|
|
static void
|
static void
|
issue_prefetches (struct mem_ref_group *groups,
|
issue_prefetches (struct mem_ref_group *groups,
|
unsigned unroll_factor, unsigned ahead)
|
unsigned unroll_factor, unsigned ahead)
|
{
|
{
|
struct mem_ref *ref;
|
struct mem_ref *ref;
|
|
|
for (; groups; groups = groups->next)
|
for (; groups; groups = groups->next)
|
for (ref = groups->refs; ref; ref = ref->next)
|
for (ref = groups->refs; ref; ref = ref->next)
|
if (ref->issue_prefetch_p)
|
if (ref->issue_prefetch_p)
|
issue_prefetch_ref (ref, unroll_factor, ahead);
|
issue_prefetch_ref (ref, unroll_factor, ahead);
|
}
|
}
|
|
|
/* Returns true if REF is a memory write for that a nontemporal store insn
|
/* Returns true if REF is a memory write for that a nontemporal store insn
|
can be used. */
|
can be used. */
|
|
|
static bool
|
static bool
|
nontemporal_store_p (struct mem_ref *ref)
|
nontemporal_store_p (struct mem_ref *ref)
|
{
|
{
|
enum machine_mode mode;
|
enum machine_mode mode;
|
enum insn_code code;
|
enum insn_code code;
|
|
|
/* REF must be a write that is not reused. We require it to be independent
|
/* REF must be a write that is not reused. We require it to be independent
|
on all other memory references in the loop, as the nontemporal stores may
|
on all other memory references in the loop, as the nontemporal stores may
|
be reordered with respect to other memory references. */
|
be reordered with respect to other memory references. */
|
if (!ref->write_p
|
if (!ref->write_p
|
|| !ref->independent_p
|
|| !ref->independent_p
|
|| ref->reuse_distance < L2_CACHE_SIZE_BYTES)
|
|| ref->reuse_distance < L2_CACHE_SIZE_BYTES)
|
return false;
|
return false;
|
|
|
/* Check that we have the storent instruction for the mode. */
|
/* Check that we have the storent instruction for the mode. */
|
mode = TYPE_MODE (TREE_TYPE (ref->mem));
|
mode = TYPE_MODE (TREE_TYPE (ref->mem));
|
if (mode == BLKmode)
|
if (mode == BLKmode)
|
return false;
|
return false;
|
|
|
code = optab_handler (storent_optab, mode)->insn_code;
|
code = optab_handler (storent_optab, mode)->insn_code;
|
return code != CODE_FOR_nothing;
|
return code != CODE_FOR_nothing;
|
}
|
}
|
|
|
/* If REF is a nontemporal store, we mark the corresponding modify statement
|
/* If REF is a nontemporal store, we mark the corresponding modify statement
|
and return true. Otherwise, we return false. */
|
and return true. Otherwise, we return false. */
|
|
|
static bool
|
static bool
|
mark_nontemporal_store (struct mem_ref *ref)
|
mark_nontemporal_store (struct mem_ref *ref)
|
{
|
{
|
if (!nontemporal_store_p (ref))
|
if (!nontemporal_store_p (ref))
|
return false;
|
return false;
|
|
|
if (dump_file && (dump_flags & TDF_DETAILS))
|
if (dump_file && (dump_flags & TDF_DETAILS))
|
fprintf (dump_file, "Marked reference %p as a nontemporal store.\n",
|
fprintf (dump_file, "Marked reference %p as a nontemporal store.\n",
|
(void *) ref);
|
(void *) ref);
|
|
|
gimple_assign_set_nontemporal_move (ref->stmt, true);
|
gimple_assign_set_nontemporal_move (ref->stmt, true);
|
ref->storent_p = true;
|
ref->storent_p = true;
|
|
|
return true;
|
return true;
|
}
|
}
|
|
|
/* Issue a memory fence instruction after LOOP. */
|
/* Issue a memory fence instruction after LOOP. */
|
|
|
static void
|
static void
|
emit_mfence_after_loop (struct loop *loop)
|
emit_mfence_after_loop (struct loop *loop)
|
{
|
{
|
VEC (edge, heap) *exits = get_loop_exit_edges (loop);
|
VEC (edge, heap) *exits = get_loop_exit_edges (loop);
|
edge exit;
|
edge exit;
|
gimple call;
|
gimple call;
|
gimple_stmt_iterator bsi;
|
gimple_stmt_iterator bsi;
|
unsigned i;
|
unsigned i;
|
|
|
for (i = 0; VEC_iterate (edge, exits, i, exit); i++)
|
for (i = 0; VEC_iterate (edge, exits, i, exit); i++)
|
{
|
{
|
call = gimple_build_call (FENCE_FOLLOWING_MOVNT, 0);
|
call = gimple_build_call (FENCE_FOLLOWING_MOVNT, 0);
|
|
|
if (!single_pred_p (exit->dest)
|
if (!single_pred_p (exit->dest)
|
/* If possible, we prefer not to insert the fence on other paths
|
/* If possible, we prefer not to insert the fence on other paths
|
in cfg. */
|
in cfg. */
|
&& !(exit->flags & EDGE_ABNORMAL))
|
&& !(exit->flags & EDGE_ABNORMAL))
|
split_loop_exit_edge (exit);
|
split_loop_exit_edge (exit);
|
bsi = gsi_after_labels (exit->dest);
|
bsi = gsi_after_labels (exit->dest);
|
|
|
gsi_insert_before (&bsi, call, GSI_NEW_STMT);
|
gsi_insert_before (&bsi, call, GSI_NEW_STMT);
|
mark_virtual_ops_for_renaming (call);
|
mark_virtual_ops_for_renaming (call);
|
}
|
}
|
|
|
VEC_free (edge, heap, exits);
|
VEC_free (edge, heap, exits);
|
update_ssa (TODO_update_ssa_only_virtuals);
|
update_ssa (TODO_update_ssa_only_virtuals);
|
}
|
}
|
|
|
/* Returns true if we can use storent in loop, false otherwise. */
|
/* Returns true if we can use storent in loop, false otherwise. */
|
|
|
static bool
|
static bool
|
may_use_storent_in_loop_p (struct loop *loop)
|
may_use_storent_in_loop_p (struct loop *loop)
|
{
|
{
|
bool ret = true;
|
bool ret = true;
|
|
|
if (loop->inner != NULL)
|
if (loop->inner != NULL)
|
return false;
|
return false;
|
|
|
/* If we must issue a mfence insn after using storent, check that there
|
/* If we must issue a mfence insn after using storent, check that there
|
is a suitable place for it at each of the loop exits. */
|
is a suitable place for it at each of the loop exits. */
|
if (FENCE_FOLLOWING_MOVNT != NULL_TREE)
|
if (FENCE_FOLLOWING_MOVNT != NULL_TREE)
|
{
|
{
|
VEC (edge, heap) *exits = get_loop_exit_edges (loop);
|
VEC (edge, heap) *exits = get_loop_exit_edges (loop);
|
unsigned i;
|
unsigned i;
|
edge exit;
|
edge exit;
|
|
|
for (i = 0; VEC_iterate (edge, exits, i, exit); i++)
|
for (i = 0; VEC_iterate (edge, exits, i, exit); i++)
|
if ((exit->flags & EDGE_ABNORMAL)
|
if ((exit->flags & EDGE_ABNORMAL)
|
&& exit->dest == EXIT_BLOCK_PTR)
|
&& exit->dest == EXIT_BLOCK_PTR)
|
ret = false;
|
ret = false;
|
|
|
VEC_free (edge, heap, exits);
|
VEC_free (edge, heap, exits);
|
}
|
}
|
|
|
return ret;
|
return ret;
|
}
|
}
|
|
|
/* Marks nontemporal stores in LOOP. GROUPS contains the description of memory
|
/* Marks nontemporal stores in LOOP. GROUPS contains the description of memory
|
references in the loop. */
|
references in the loop. */
|
|
|
static void
|
static void
|
mark_nontemporal_stores (struct loop *loop, struct mem_ref_group *groups)
|
mark_nontemporal_stores (struct loop *loop, struct mem_ref_group *groups)
|
{
|
{
|
struct mem_ref *ref;
|
struct mem_ref *ref;
|
bool any = false;
|
bool any = false;
|
|
|
if (!may_use_storent_in_loop_p (loop))
|
if (!may_use_storent_in_loop_p (loop))
|
return;
|
return;
|
|
|
for (; groups; groups = groups->next)
|
for (; groups; groups = groups->next)
|
for (ref = groups->refs; ref; ref = ref->next)
|
for (ref = groups->refs; ref; ref = ref->next)
|
any |= mark_nontemporal_store (ref);
|
any |= mark_nontemporal_store (ref);
|
|
|
if (any && FENCE_FOLLOWING_MOVNT != NULL_TREE)
|
if (any && FENCE_FOLLOWING_MOVNT != NULL_TREE)
|
emit_mfence_after_loop (loop);
|
emit_mfence_after_loop (loop);
|
}
|
}
|
|
|
/* Determines whether we can profitably unroll LOOP FACTOR times, and if
|
/* Determines whether we can profitably unroll LOOP FACTOR times, and if
|
this is the case, fill in DESC by the description of number of
|
this is the case, fill in DESC by the description of number of
|
iterations. */
|
iterations. */
|
|
|
static bool
|
static bool
|
should_unroll_loop_p (struct loop *loop, struct tree_niter_desc *desc,
|
should_unroll_loop_p (struct loop *loop, struct tree_niter_desc *desc,
|
unsigned factor)
|
unsigned factor)
|
{
|
{
|
if (!can_unroll_loop_p (loop, factor, desc))
|
if (!can_unroll_loop_p (loop, factor, desc))
|
return false;
|
return false;
|
|
|
/* We only consider loops without control flow for unrolling. This is not
|
/* We only consider loops without control flow for unrolling. This is not
|
a hard restriction -- tree_unroll_loop works with arbitrary loops
|
a hard restriction -- tree_unroll_loop works with arbitrary loops
|
as well; but the unrolling/prefetching is usually more profitable for
|
as well; but the unrolling/prefetching is usually more profitable for
|
loops consisting of a single basic block, and we want to limit the
|
loops consisting of a single basic block, and we want to limit the
|
code growth. */
|
code growth. */
|
if (loop->num_nodes > 2)
|
if (loop->num_nodes > 2)
|
return false;
|
return false;
|
|
|
return true;
|
return true;
|
}
|
}
|
|
|
/* Determine the coefficient by that unroll LOOP, from the information
|
/* Determine the coefficient by that unroll LOOP, from the information
|
contained in the list of memory references REFS. Description of
|
contained in the list of memory references REFS. Description of
|
umber of iterations of LOOP is stored to DESC. NINSNS is the number of
|
umber of iterations of LOOP is stored to DESC. NINSNS is the number of
|
insns of the LOOP. EST_NITER is the estimated number of iterations of
|
insns of the LOOP. EST_NITER is the estimated number of iterations of
|
the loop, or -1 if no estimate is available. */
|
the loop, or -1 if no estimate is available. */
|
|
|
static unsigned
|
static unsigned
|
determine_unroll_factor (struct loop *loop, struct mem_ref_group *refs,
|
determine_unroll_factor (struct loop *loop, struct mem_ref_group *refs,
|
unsigned ninsns, struct tree_niter_desc *desc,
|
unsigned ninsns, struct tree_niter_desc *desc,
|
HOST_WIDE_INT est_niter)
|
HOST_WIDE_INT est_niter)
|
{
|
{
|
unsigned upper_bound;
|
unsigned upper_bound;
|
unsigned nfactor, factor, mod_constraint;
|
unsigned nfactor, factor, mod_constraint;
|
struct mem_ref_group *agp;
|
struct mem_ref_group *agp;
|
struct mem_ref *ref;
|
struct mem_ref *ref;
|
|
|
/* First check whether the loop is not too large to unroll. We ignore
|
/* First check whether the loop is not too large to unroll. We ignore
|
PARAM_MAX_UNROLL_TIMES, because for small loops, it prevented us
|
PARAM_MAX_UNROLL_TIMES, because for small loops, it prevented us
|
from unrolling them enough to make exactly one cache line covered by each
|
from unrolling them enough to make exactly one cache line covered by each
|
iteration. Also, the goal of PARAM_MAX_UNROLL_TIMES is to prevent
|
iteration. Also, the goal of PARAM_MAX_UNROLL_TIMES is to prevent
|
us from unrolling the loops too many times in cases where we only expect
|
us from unrolling the loops too many times in cases where we only expect
|
gains from better scheduling and decreasing loop overhead, which is not
|
gains from better scheduling and decreasing loop overhead, which is not
|
the case here. */
|
the case here. */
|
upper_bound = PARAM_VALUE (PARAM_MAX_UNROLLED_INSNS) / ninsns;
|
upper_bound = PARAM_VALUE (PARAM_MAX_UNROLLED_INSNS) / ninsns;
|
|
|
/* If we unrolled the loop more times than it iterates, the unrolled version
|
/* If we unrolled the loop more times than it iterates, the unrolled version
|
of the loop would be never entered. */
|
of the loop would be never entered. */
|
if (est_niter >= 0 && est_niter < (HOST_WIDE_INT) upper_bound)
|
if (est_niter >= 0 && est_niter < (HOST_WIDE_INT) upper_bound)
|
upper_bound = est_niter;
|
upper_bound = est_niter;
|
|
|
if (upper_bound <= 1)
|
if (upper_bound <= 1)
|
return 1;
|
return 1;
|
|
|
/* Choose the factor so that we may prefetch each cache just once,
|
/* Choose the factor so that we may prefetch each cache just once,
|
but bound the unrolling by UPPER_BOUND. */
|
but bound the unrolling by UPPER_BOUND. */
|
factor = 1;
|
factor = 1;
|
for (agp = refs; agp; agp = agp->next)
|
for (agp = refs; agp; agp = agp->next)
|
for (ref = agp->refs; ref; ref = ref->next)
|
for (ref = agp->refs; ref; ref = ref->next)
|
if (should_issue_prefetch_p (ref))
|
if (should_issue_prefetch_p (ref))
|
{
|
{
|
mod_constraint = ref->prefetch_mod;
|
mod_constraint = ref->prefetch_mod;
|
nfactor = least_common_multiple (mod_constraint, factor);
|
nfactor = least_common_multiple (mod_constraint, factor);
|
if (nfactor <= upper_bound)
|
if (nfactor <= upper_bound)
|
factor = nfactor;
|
factor = nfactor;
|
}
|
}
|
|
|
if (!should_unroll_loop_p (loop, desc, factor))
|
if (!should_unroll_loop_p (loop, desc, factor))
|
return 1;
|
return 1;
|
|
|
return factor;
|
return factor;
|
}
|
}
|
|
|
/* Returns the total volume of the memory references REFS, taking into account
|
/* Returns the total volume of the memory references REFS, taking into account
|
reuses in the innermost loop and cache line size. TODO -- we should also
|
reuses in the innermost loop and cache line size. TODO -- we should also
|
take into account reuses across the iterations of the loops in the loop
|
take into account reuses across the iterations of the loops in the loop
|
nest. */
|
nest. */
|
|
|
static unsigned
|
static unsigned
|
volume_of_references (struct mem_ref_group *refs)
|
volume_of_references (struct mem_ref_group *refs)
|
{
|
{
|
unsigned volume = 0;
|
unsigned volume = 0;
|
struct mem_ref_group *gr;
|
struct mem_ref_group *gr;
|
struct mem_ref *ref;
|
struct mem_ref *ref;
|
|
|
for (gr = refs; gr; gr = gr->next)
|
for (gr = refs; gr; gr = gr->next)
|
for (ref = gr->refs; ref; ref = ref->next)
|
for (ref = gr->refs; ref; ref = ref->next)
|
{
|
{
|
/* Almost always reuses another value? */
|
/* Almost always reuses another value? */
|
if (ref->prefetch_before != PREFETCH_ALL)
|
if (ref->prefetch_before != PREFETCH_ALL)
|
continue;
|
continue;
|
|
|
/* If several iterations access the same cache line, use the size of
|
/* If several iterations access the same cache line, use the size of
|
the line divided by this number. Otherwise, a cache line is
|
the line divided by this number. Otherwise, a cache line is
|
accessed in each iteration. TODO -- in the latter case, we should
|
accessed in each iteration. TODO -- in the latter case, we should
|
take the size of the reference into account, rounding it up on cache
|
take the size of the reference into account, rounding it up on cache
|
line size multiple. */
|
line size multiple. */
|
volume += L1_CACHE_LINE_SIZE / ref->prefetch_mod;
|
volume += L1_CACHE_LINE_SIZE / ref->prefetch_mod;
|
}
|
}
|
return volume;
|
return volume;
|
}
|
}
|
|
|
/* Returns the volume of memory references accessed across VEC iterations of
|
/* Returns the volume of memory references accessed across VEC iterations of
|
loops, whose sizes are described in the LOOP_SIZES array. N is the number
|
loops, whose sizes are described in the LOOP_SIZES array. N is the number
|
of the loops in the nest (length of VEC and LOOP_SIZES vectors). */
|
of the loops in the nest (length of VEC and LOOP_SIZES vectors). */
|
|
|
static unsigned
|
static unsigned
|
volume_of_dist_vector (lambda_vector vec, unsigned *loop_sizes, unsigned n)
|
volume_of_dist_vector (lambda_vector vec, unsigned *loop_sizes, unsigned n)
|
{
|
{
|
unsigned i;
|
unsigned i;
|
|
|
for (i = 0; i < n; i++)
|
for (i = 0; i < n; i++)
|
if (vec[i] != 0)
|
if (vec[i] != 0)
|
break;
|
break;
|
|
|
if (i == n)
|
if (i == n)
|
return 0;
|
return 0;
|
|
|
gcc_assert (vec[i] > 0);
|
gcc_assert (vec[i] > 0);
|
|
|
/* We ignore the parts of the distance vector in subloops, since usually
|
/* We ignore the parts of the distance vector in subloops, since usually
|
the numbers of iterations are much smaller. */
|
the numbers of iterations are much smaller. */
|
return loop_sizes[i] * vec[i];
|
return loop_sizes[i] * vec[i];
|
}
|
}
|
|
|
/* Add the steps of ACCESS_FN multiplied by STRIDE to the array STRIDE
|
/* Add the steps of ACCESS_FN multiplied by STRIDE to the array STRIDE
|
at the position corresponding to the loop of the step. N is the depth
|
at the position corresponding to the loop of the step. N is the depth
|
of the considered loop nest, and, LOOP is its innermost loop. */
|
of the considered loop nest, and, LOOP is its innermost loop. */
|
|
|
static void
|
static void
|
add_subscript_strides (tree access_fn, unsigned stride,
|
add_subscript_strides (tree access_fn, unsigned stride,
|
HOST_WIDE_INT *strides, unsigned n, struct loop *loop)
|
HOST_WIDE_INT *strides, unsigned n, struct loop *loop)
|
{
|
{
|
struct loop *aloop;
|
struct loop *aloop;
|
tree step;
|
tree step;
|
HOST_WIDE_INT astep;
|
HOST_WIDE_INT astep;
|
unsigned min_depth = loop_depth (loop) - n;
|
unsigned min_depth = loop_depth (loop) - n;
|
|
|
while (TREE_CODE (access_fn) == POLYNOMIAL_CHREC)
|
while (TREE_CODE (access_fn) == POLYNOMIAL_CHREC)
|
{
|
{
|
aloop = get_chrec_loop (access_fn);
|
aloop = get_chrec_loop (access_fn);
|
step = CHREC_RIGHT (access_fn);
|
step = CHREC_RIGHT (access_fn);
|
access_fn = CHREC_LEFT (access_fn);
|
access_fn = CHREC_LEFT (access_fn);
|
|
|
if ((unsigned) loop_depth (aloop) <= min_depth)
|
if ((unsigned) loop_depth (aloop) <= min_depth)
|
continue;
|
continue;
|
|
|
if (host_integerp (step, 0))
|
if (host_integerp (step, 0))
|
astep = tree_low_cst (step, 0);
|
astep = tree_low_cst (step, 0);
|
else
|
else
|
astep = L1_CACHE_LINE_SIZE;
|
astep = L1_CACHE_LINE_SIZE;
|
|
|
strides[n - 1 - loop_depth (loop) + loop_depth (aloop)] += astep * stride;
|
strides[n - 1 - loop_depth (loop) + loop_depth (aloop)] += astep * stride;
|
|
|
}
|
}
|
}
|
}
|
|
|
/* Returns the volume of memory references accessed between two consecutive
|
/* Returns the volume of memory references accessed between two consecutive
|
self-reuses of the reference DR. We consider the subscripts of DR in N
|
self-reuses of the reference DR. We consider the subscripts of DR in N
|
loops, and LOOP_SIZES contains the volumes of accesses in each of the
|
loops, and LOOP_SIZES contains the volumes of accesses in each of the
|
loops. LOOP is the innermost loop of the current loop nest. */
|
loops. LOOP is the innermost loop of the current loop nest. */
|
|
|
static unsigned
|
static unsigned
|
self_reuse_distance (data_reference_p dr, unsigned *loop_sizes, unsigned n,
|
self_reuse_distance (data_reference_p dr, unsigned *loop_sizes, unsigned n,
|
struct loop *loop)
|
struct loop *loop)
|
{
|
{
|
tree stride, access_fn;
|
tree stride, access_fn;
|
HOST_WIDE_INT *strides, astride;
|
HOST_WIDE_INT *strides, astride;
|
VEC (tree, heap) *access_fns;
|
VEC (tree, heap) *access_fns;
|
tree ref = DR_REF (dr);
|
tree ref = DR_REF (dr);
|
unsigned i, ret = ~0u;
|
unsigned i, ret = ~0u;
|
|
|
/* In the following example:
|
/* In the following example:
|
|
|
for (i = 0; i < N; i++)
|
for (i = 0; i < N; i++)
|
for (j = 0; j < N; j++)
|
for (j = 0; j < N; j++)
|
use (a[j][i]);
|
use (a[j][i]);
|
the same cache line is accessed each N steps (except if the change from
|
the same cache line is accessed each N steps (except if the change from
|
i to i + 1 crosses the boundary of the cache line). Thus, for self-reuse,
|
i to i + 1 crosses the boundary of the cache line). Thus, for self-reuse,
|
we cannot rely purely on the results of the data dependence analysis.
|
we cannot rely purely on the results of the data dependence analysis.
|
|
|
Instead, we compute the stride of the reference in each loop, and consider
|
Instead, we compute the stride of the reference in each loop, and consider
|
the innermost loop in that the stride is less than cache size. */
|
the innermost loop in that the stride is less than cache size. */
|
|
|
strides = XCNEWVEC (HOST_WIDE_INT, n);
|
strides = XCNEWVEC (HOST_WIDE_INT, n);
|
access_fns = DR_ACCESS_FNS (dr);
|
access_fns = DR_ACCESS_FNS (dr);
|
|
|
for (i = 0; VEC_iterate (tree, access_fns, i, access_fn); i++)
|
for (i = 0; VEC_iterate (tree, access_fns, i, access_fn); i++)
|
{
|
{
|
/* Keep track of the reference corresponding to the subscript, so that we
|
/* Keep track of the reference corresponding to the subscript, so that we
|
know its stride. */
|
know its stride. */
|
while (handled_component_p (ref) && TREE_CODE (ref) != ARRAY_REF)
|
while (handled_component_p (ref) && TREE_CODE (ref) != ARRAY_REF)
|
ref = TREE_OPERAND (ref, 0);
|
ref = TREE_OPERAND (ref, 0);
|
|
|
if (TREE_CODE (ref) == ARRAY_REF)
|
if (TREE_CODE (ref) == ARRAY_REF)
|
{
|
{
|
stride = TYPE_SIZE_UNIT (TREE_TYPE (ref));
|
stride = TYPE_SIZE_UNIT (TREE_TYPE (ref));
|
if (host_integerp (stride, 1))
|
if (host_integerp (stride, 1))
|
astride = tree_low_cst (stride, 1);
|
astride = tree_low_cst (stride, 1);
|
else
|
else
|
astride = L1_CACHE_LINE_SIZE;
|
astride = L1_CACHE_LINE_SIZE;
|
|
|
ref = TREE_OPERAND (ref, 0);
|
ref = TREE_OPERAND (ref, 0);
|
}
|
}
|
else
|
else
|
astride = 1;
|
astride = 1;
|
|
|
add_subscript_strides (access_fn, astride, strides, n, loop);
|
add_subscript_strides (access_fn, astride, strides, n, loop);
|
}
|
}
|
|
|
for (i = n; i-- > 0; )
|
for (i = n; i-- > 0; )
|
{
|
{
|
unsigned HOST_WIDE_INT s;
|
unsigned HOST_WIDE_INT s;
|
|
|
s = strides[i] < 0 ? -strides[i] : strides[i];
|
s = strides[i] < 0 ? -strides[i] : strides[i];
|
|
|
if (s < (unsigned) L1_CACHE_LINE_SIZE
|
if (s < (unsigned) L1_CACHE_LINE_SIZE
|
&& (loop_sizes[i]
|
&& (loop_sizes[i]
|
> (unsigned) (L1_CACHE_SIZE_BYTES / NONTEMPORAL_FRACTION)))
|
> (unsigned) (L1_CACHE_SIZE_BYTES / NONTEMPORAL_FRACTION)))
|
{
|
{
|
ret = loop_sizes[i];
|
ret = loop_sizes[i];
|
break;
|
break;
|
}
|
}
|
}
|
}
|
|
|
free (strides);
|
free (strides);
|
return ret;
|
return ret;
|
}
|
}
|
|
|
/* Determines the distance till the first reuse of each reference in REFS
|
/* Determines the distance till the first reuse of each reference in REFS
|
in the loop nest of LOOP. NO_OTHER_REFS is true if there are no other
|
in the loop nest of LOOP. NO_OTHER_REFS is true if there are no other
|
memory references in the loop. */
|
memory references in the loop. */
|
|
|
static void
|
static void
|
determine_loop_nest_reuse (struct loop *loop, struct mem_ref_group *refs,
|
determine_loop_nest_reuse (struct loop *loop, struct mem_ref_group *refs,
|
bool no_other_refs)
|
bool no_other_refs)
|
{
|
{
|
struct loop *nest, *aloop;
|
struct loop *nest, *aloop;
|
VEC (data_reference_p, heap) *datarefs = NULL;
|
VEC (data_reference_p, heap) *datarefs = NULL;
|
VEC (ddr_p, heap) *dependences = NULL;
|
VEC (ddr_p, heap) *dependences = NULL;
|
struct mem_ref_group *gr;
|
struct mem_ref_group *gr;
|
struct mem_ref *ref, *refb;
|
struct mem_ref *ref, *refb;
|
VEC (loop_p, heap) *vloops = NULL;
|
VEC (loop_p, heap) *vloops = NULL;
|
unsigned *loop_data_size;
|
unsigned *loop_data_size;
|
unsigned i, j, n;
|
unsigned i, j, n;
|
unsigned volume, dist, adist;
|
unsigned volume, dist, adist;
|
HOST_WIDE_INT vol;
|
HOST_WIDE_INT vol;
|
data_reference_p dr;
|
data_reference_p dr;
|
ddr_p dep;
|
ddr_p dep;
|
|
|
if (loop->inner)
|
if (loop->inner)
|
return;
|
return;
|
|
|
/* Find the outermost loop of the loop nest of loop (we require that
|
/* Find the outermost loop of the loop nest of loop (we require that
|
there are no sibling loops inside the nest). */
|
there are no sibling loops inside the nest). */
|
nest = loop;
|
nest = loop;
|
while (1)
|
while (1)
|
{
|
{
|
aloop = loop_outer (nest);
|
aloop = loop_outer (nest);
|
|
|
if (aloop == current_loops->tree_root
|
if (aloop == current_loops->tree_root
|
|| aloop->inner->next)
|
|| aloop->inner->next)
|
break;
|
break;
|
|
|
nest = aloop;
|
nest = aloop;
|
}
|
}
|
|
|
/* For each loop, determine the amount of data accessed in each iteration.
|
/* For each loop, determine the amount of data accessed in each iteration.
|
We use this to estimate whether the reference is evicted from the
|
We use this to estimate whether the reference is evicted from the
|
cache before its reuse. */
|
cache before its reuse. */
|
find_loop_nest (nest, &vloops);
|
find_loop_nest (nest, &vloops);
|
n = VEC_length (loop_p, vloops);
|
n = VEC_length (loop_p, vloops);
|
loop_data_size = XNEWVEC (unsigned, n);
|
loop_data_size = XNEWVEC (unsigned, n);
|
volume = volume_of_references (refs);
|
volume = volume_of_references (refs);
|
i = n;
|
i = n;
|
while (i-- != 0)
|
while (i-- != 0)
|
{
|
{
|
loop_data_size[i] = volume;
|
loop_data_size[i] = volume;
|
/* Bound the volume by the L2 cache size, since above this bound,
|
/* Bound the volume by the L2 cache size, since above this bound,
|
all dependence distances are equivalent. */
|
all dependence distances are equivalent. */
|
if (volume > L2_CACHE_SIZE_BYTES)
|
if (volume > L2_CACHE_SIZE_BYTES)
|
continue;
|
continue;
|
|
|
aloop = VEC_index (loop_p, vloops, i);
|
aloop = VEC_index (loop_p, vloops, i);
|
vol = estimated_loop_iterations_int (aloop, false);
|
vol = estimated_loop_iterations_int (aloop, false);
|
if (vol < 0)
|
if (vol < 0)
|
vol = expected_loop_iterations (aloop);
|
vol = expected_loop_iterations (aloop);
|
volume *= vol;
|
volume *= vol;
|
}
|
}
|
|
|
/* Prepare the references in the form suitable for data dependence
|
/* Prepare the references in the form suitable for data dependence
|
analysis. We ignore unanalyzable data references (the results
|
analysis. We ignore unanalyzable data references (the results
|
are used just as a heuristics to estimate temporality of the
|
are used just as a heuristics to estimate temporality of the
|
references, hence we do not need to worry about correctness). */
|
references, hence we do not need to worry about correctness). */
|
for (gr = refs; gr; gr = gr->next)
|
for (gr = refs; gr; gr = gr->next)
|
for (ref = gr->refs; ref; ref = ref->next)
|
for (ref = gr->refs; ref; ref = ref->next)
|
{
|
{
|
dr = create_data_ref (nest, ref->mem, ref->stmt, !ref->write_p);
|
dr = create_data_ref (nest, ref->mem, ref->stmt, !ref->write_p);
|
|
|
if (dr)
|
if (dr)
|
{
|
{
|
ref->reuse_distance = volume;
|
ref->reuse_distance = volume;
|
dr->aux = ref;
|
dr->aux = ref;
|
VEC_safe_push (data_reference_p, heap, datarefs, dr);
|
VEC_safe_push (data_reference_p, heap, datarefs, dr);
|
}
|
}
|
else
|
else
|
no_other_refs = false;
|
no_other_refs = false;
|
}
|
}
|
|
|
for (i = 0; VEC_iterate (data_reference_p, datarefs, i, dr); i++)
|
for (i = 0; VEC_iterate (data_reference_p, datarefs, i, dr); i++)
|
{
|
{
|
dist = self_reuse_distance (dr, loop_data_size, n, loop);
|
dist = self_reuse_distance (dr, loop_data_size, n, loop);
|
ref = (struct mem_ref *) dr->aux;
|
ref = (struct mem_ref *) dr->aux;
|
if (ref->reuse_distance > dist)
|
if (ref->reuse_distance > dist)
|
ref->reuse_distance = dist;
|
ref->reuse_distance = dist;
|
|
|
if (no_other_refs)
|
if (no_other_refs)
|
ref->independent_p = true;
|
ref->independent_p = true;
|
}
|
}
|
|
|
compute_all_dependences (datarefs, &dependences, vloops, true);
|
compute_all_dependences (datarefs, &dependences, vloops, true);
|
|
|
for (i = 0; VEC_iterate (ddr_p, dependences, i, dep); i++)
|
for (i = 0; VEC_iterate (ddr_p, dependences, i, dep); i++)
|
{
|
{
|
if (DDR_ARE_DEPENDENT (dep) == chrec_known)
|
if (DDR_ARE_DEPENDENT (dep) == chrec_known)
|
continue;
|
continue;
|
|
|
ref = (struct mem_ref *) DDR_A (dep)->aux;
|
ref = (struct mem_ref *) DDR_A (dep)->aux;
|
refb = (struct mem_ref *) DDR_B (dep)->aux;
|
refb = (struct mem_ref *) DDR_B (dep)->aux;
|
|
|
if (DDR_ARE_DEPENDENT (dep) == chrec_dont_know
|
if (DDR_ARE_DEPENDENT (dep) == chrec_dont_know
|
|| DDR_NUM_DIST_VECTS (dep) == 0)
|
|| DDR_NUM_DIST_VECTS (dep) == 0)
|
{
|
{
|
/* If the dependence cannot be analyzed, assume that there might be
|
/* If the dependence cannot be analyzed, assume that there might be
|
a reuse. */
|
a reuse. */
|
dist = 0;
|
dist = 0;
|
|
|
ref->independent_p = false;
|
ref->independent_p = false;
|
refb->independent_p = false;
|
refb->independent_p = false;
|
}
|
}
|
else
|
else
|
{
|
{
|
/* The distance vectors are normalized to be always lexicographically
|
/* The distance vectors are normalized to be always lexicographically
|
positive, hence we cannot tell just from them whether DDR_A comes
|
positive, hence we cannot tell just from them whether DDR_A comes
|
before DDR_B or vice versa. However, it is not important,
|
before DDR_B or vice versa. However, it is not important,
|
anyway -- if DDR_A is close to DDR_B, then it is either reused in
|
anyway -- if DDR_A is close to DDR_B, then it is either reused in
|
DDR_B (and it is not nontemporal), or it reuses the value of DDR_B
|
DDR_B (and it is not nontemporal), or it reuses the value of DDR_B
|
in cache (and marking it as nontemporal would not affect
|
in cache (and marking it as nontemporal would not affect
|
anything). */
|
anything). */
|
|
|
dist = volume;
|
dist = volume;
|
for (j = 0; j < DDR_NUM_DIST_VECTS (dep); j++)
|
for (j = 0; j < DDR_NUM_DIST_VECTS (dep); j++)
|
{
|
{
|
adist = volume_of_dist_vector (DDR_DIST_VECT (dep, j),
|
adist = volume_of_dist_vector (DDR_DIST_VECT (dep, j),
|
loop_data_size, n);
|
loop_data_size, n);
|
|
|
/* If this is a dependence in the innermost loop (i.e., the
|
/* If this is a dependence in the innermost loop (i.e., the
|
distances in all superloops are zero) and it is not
|
distances in all superloops are zero) and it is not
|
the trivial self-dependence with distance zero, record that
|
the trivial self-dependence with distance zero, record that
|
the references are not completely independent. */
|
the references are not completely independent. */
|
if (lambda_vector_zerop (DDR_DIST_VECT (dep, j), n - 1)
|
if (lambda_vector_zerop (DDR_DIST_VECT (dep, j), n - 1)
|
&& (ref != refb
|
&& (ref != refb
|
|| DDR_DIST_VECT (dep, j)[n-1] != 0))
|
|| DDR_DIST_VECT (dep, j)[n-1] != 0))
|
{
|
{
|
ref->independent_p = false;
|
ref->independent_p = false;
|
refb->independent_p = false;
|
refb->independent_p = false;
|
}
|
}
|
|
|
/* Ignore accesses closer than
|
/* Ignore accesses closer than
|
L1_CACHE_SIZE_BYTES / NONTEMPORAL_FRACTION,
|
L1_CACHE_SIZE_BYTES / NONTEMPORAL_FRACTION,
|
so that we use nontemporal prefetches e.g. if single memory
|
so that we use nontemporal prefetches e.g. if single memory
|
location is accessed several times in a single iteration of
|
location is accessed several times in a single iteration of
|
the loop. */
|
the loop. */
|
if (adist < L1_CACHE_SIZE_BYTES / NONTEMPORAL_FRACTION)
|
if (adist < L1_CACHE_SIZE_BYTES / NONTEMPORAL_FRACTION)
|
continue;
|
continue;
|
|
|
if (adist < dist)
|
if (adist < dist)
|
dist = adist;
|
dist = adist;
|
}
|
}
|
}
|
}
|
|
|
if (ref->reuse_distance > dist)
|
if (ref->reuse_distance > dist)
|
ref->reuse_distance = dist;
|
ref->reuse_distance = dist;
|
if (refb->reuse_distance > dist)
|
if (refb->reuse_distance > dist)
|
refb->reuse_distance = dist;
|
refb->reuse_distance = dist;
|
}
|
}
|
|
|
free_dependence_relations (dependences);
|
free_dependence_relations (dependences);
|
free_data_refs (datarefs);
|
free_data_refs (datarefs);
|
free (loop_data_size);
|
free (loop_data_size);
|
|
|
if (dump_file && (dump_flags & TDF_DETAILS))
|
if (dump_file && (dump_flags & TDF_DETAILS))
|
{
|
{
|
fprintf (dump_file, "Reuse distances:\n");
|
fprintf (dump_file, "Reuse distances:\n");
|
for (gr = refs; gr; gr = gr->next)
|
for (gr = refs; gr; gr = gr->next)
|
for (ref = gr->refs; ref; ref = ref->next)
|
for (ref = gr->refs; ref; ref = ref->next)
|
fprintf (dump_file, " ref %p distance %u\n",
|
fprintf (dump_file, " ref %p distance %u\n",
|
(void *) ref, ref->reuse_distance);
|
(void *) ref, ref->reuse_distance);
|
}
|
}
|
}
|
}
|
|
|
/* Do a cost-benefit analysis to determine if prefetching is profitable
|
/* Do a cost-benefit analysis to determine if prefetching is profitable
|
for the current loop given the following parameters:
|
for the current loop given the following parameters:
|
AHEAD: the iteration ahead distance,
|
AHEAD: the iteration ahead distance,
|
EST_NITER: the estimated trip count,
|
EST_NITER: the estimated trip count,
|
NINSNS: estimated number of instructions in the loop,
|
NINSNS: estimated number of instructions in the loop,
|
PREFETCH_COUNT: an estimate of the number of prefetches
|
PREFETCH_COUNT: an estimate of the number of prefetches
|
MEM_REF_COUNT: total number of memory references in the loop. */
|
MEM_REF_COUNT: total number of memory references in the loop. */
|
|
|
static bool
|
static bool
|
is_loop_prefetching_profitable (unsigned ahead, HOST_WIDE_INT est_niter,
|
is_loop_prefetching_profitable (unsigned ahead, HOST_WIDE_INT est_niter,
|
unsigned ninsns, unsigned prefetch_count,
|
unsigned ninsns, unsigned prefetch_count,
|
unsigned mem_ref_count)
|
unsigned mem_ref_count)
|
{
|
{
|
int insn_to_mem_ratio, insn_to_prefetch_ratio;
|
int insn_to_mem_ratio, insn_to_prefetch_ratio;
|
|
|
if (mem_ref_count == 0)
|
if (mem_ref_count == 0)
|
return false;
|
return false;
|
|
|
/* Prefetching improves performance by overlapping cache missing
|
/* Prefetching improves performance by overlapping cache missing
|
memory accesses with CPU operations. If the loop does not have
|
memory accesses with CPU operations. If the loop does not have
|
enough CPU operations to overlap with memory operations, prefetching
|
enough CPU operations to overlap with memory operations, prefetching
|
won't give a significant benefit. One approximate way of checking
|
won't give a significant benefit. One approximate way of checking
|
this is to require the ratio of instructions to memory references to
|
this is to require the ratio of instructions to memory references to
|
be above a certain limit. This approximation works well in practice.
|
be above a certain limit. This approximation works well in practice.
|
TODO: Implement a more precise computation by estimating the time
|
TODO: Implement a more precise computation by estimating the time
|
for each CPU or memory op in the loop. Time estimates for memory ops
|
for each CPU or memory op in the loop. Time estimates for memory ops
|
should account for cache misses. */
|
should account for cache misses. */
|
insn_to_mem_ratio = ninsns / mem_ref_count;
|
insn_to_mem_ratio = ninsns / mem_ref_count;
|
|
|
if (insn_to_mem_ratio < PREFETCH_MIN_INSN_TO_MEM_RATIO)
|
if (insn_to_mem_ratio < PREFETCH_MIN_INSN_TO_MEM_RATIO)
|
return false;
|
return false;
|
|
|
/* Profitability of prefetching is highly dependent on the trip count.
|
/* Profitability of prefetching is highly dependent on the trip count.
|
For a given AHEAD distance, the first AHEAD iterations do not benefit
|
For a given AHEAD distance, the first AHEAD iterations do not benefit
|
from prefetching, and the last AHEAD iterations execute useless
|
from prefetching, and the last AHEAD iterations execute useless
|
prefetches. So, if the trip count is not large enough relative to AHEAD,
|
prefetches. So, if the trip count is not large enough relative to AHEAD,
|
prefetching may cause serious performance degradation. To avoid this
|
prefetching may cause serious performance degradation. To avoid this
|
problem when the trip count is not known at compile time, we
|
problem when the trip count is not known at compile time, we
|
conservatively skip loops with high prefetching costs. For now, only
|
conservatively skip loops with high prefetching costs. For now, only
|
the I-cache cost is considered. The relative I-cache cost is estimated
|
the I-cache cost is considered. The relative I-cache cost is estimated
|
by taking the ratio between the number of prefetches and the total
|
by taking the ratio between the number of prefetches and the total
|
number of instructions. Since we are using integer arithmetic, we
|
number of instructions. Since we are using integer arithmetic, we
|
compute the reciprocal of this ratio.
|
compute the reciprocal of this ratio.
|
TODO: Account for loop unrolling, which may reduce the costs of
|
TODO: Account for loop unrolling, which may reduce the costs of
|
shorter stride prefetches. Note that not accounting for loop
|
shorter stride prefetches. Note that not accounting for loop
|
unrolling over-estimates the cost and hence gives more conservative
|
unrolling over-estimates the cost and hence gives more conservative
|
results. */
|
results. */
|
if (est_niter < 0)
|
if (est_niter < 0)
|
{
|
{
|
insn_to_prefetch_ratio = ninsns / prefetch_count;
|
insn_to_prefetch_ratio = ninsns / prefetch_count;
|
return insn_to_prefetch_ratio >= MIN_INSN_TO_PREFETCH_RATIO;
|
return insn_to_prefetch_ratio >= MIN_INSN_TO_PREFETCH_RATIO;
|
}
|
}
|
|
|
if (est_niter <= (HOST_WIDE_INT) ahead)
|
if (est_niter <= (HOST_WIDE_INT) ahead)
|
{
|
{
|
if (dump_file && (dump_flags & TDF_DETAILS))
|
if (dump_file && (dump_flags & TDF_DETAILS))
|
fprintf (dump_file,
|
fprintf (dump_file,
|
"Not prefetching -- loop estimated to roll only %d times\n",
|
"Not prefetching -- loop estimated to roll only %d times\n",
|
(int) est_niter);
|
(int) est_niter);
|
return false;
|
return false;
|
}
|
}
|
return true;
|
return true;
|
}
|
}
|
|
|
|
|
/* Issue prefetch instructions for array references in LOOP. Returns
|
/* Issue prefetch instructions for array references in LOOP. Returns
|
true if the LOOP was unrolled. */
|
true if the LOOP was unrolled. */
|
|
|
static bool
|
static bool
|
loop_prefetch_arrays (struct loop *loop)
|
loop_prefetch_arrays (struct loop *loop)
|
{
|
{
|
struct mem_ref_group *refs;
|
struct mem_ref_group *refs;
|
unsigned ahead, ninsns, time, unroll_factor;
|
unsigned ahead, ninsns, time, unroll_factor;
|
HOST_WIDE_INT est_niter;
|
HOST_WIDE_INT est_niter;
|
struct tree_niter_desc desc;
|
struct tree_niter_desc desc;
|
bool unrolled = false, no_other_refs;
|
bool unrolled = false, no_other_refs;
|
unsigned prefetch_count;
|
unsigned prefetch_count;
|
unsigned mem_ref_count;
|
unsigned mem_ref_count;
|
|
|
if (optimize_loop_nest_for_size_p (loop))
|
if (optimize_loop_nest_for_size_p (loop))
|
{
|
{
|
if (dump_file && (dump_flags & TDF_DETAILS))
|
if (dump_file && (dump_flags & TDF_DETAILS))
|
fprintf (dump_file, " ignored (cold area)\n");
|
fprintf (dump_file, " ignored (cold area)\n");
|
return false;
|
return false;
|
}
|
}
|
|
|
/* Step 1: gather the memory references. */
|
/* Step 1: gather the memory references. */
|
refs = gather_memory_references (loop, &no_other_refs, &mem_ref_count);
|
refs = gather_memory_references (loop, &no_other_refs, &mem_ref_count);
|
|
|
/* Step 2: estimate the reuse effects. */
|
/* Step 2: estimate the reuse effects. */
|
prune_by_reuse (refs);
|
prune_by_reuse (refs);
|
|
|
prefetch_count = estimate_prefetch_count (refs);
|
prefetch_count = estimate_prefetch_count (refs);
|
if (prefetch_count == 0)
|
if (prefetch_count == 0)
|
goto fail;
|
goto fail;
|
|
|
determine_loop_nest_reuse (loop, refs, no_other_refs);
|
determine_loop_nest_reuse (loop, refs, no_other_refs);
|
|
|
/* Step 3: determine the ahead and unroll factor. */
|
/* Step 3: determine the ahead and unroll factor. */
|
|
|
/* FIXME: the time should be weighted by the probabilities of the blocks in
|
/* FIXME: the time should be weighted by the probabilities of the blocks in
|
the loop body. */
|
the loop body. */
|
time = tree_num_loop_insns (loop, &eni_time_weights);
|
time = tree_num_loop_insns (loop, &eni_time_weights);
|
ahead = (PREFETCH_LATENCY + time - 1) / time;
|
ahead = (PREFETCH_LATENCY + time - 1) / time;
|
est_niter = estimated_loop_iterations_int (loop, false);
|
est_niter = estimated_loop_iterations_int (loop, false);
|
|
|
ninsns = tree_num_loop_insns (loop, &eni_size_weights);
|
ninsns = tree_num_loop_insns (loop, &eni_size_weights);
|
unroll_factor = determine_unroll_factor (loop, refs, ninsns, &desc,
|
unroll_factor = determine_unroll_factor (loop, refs, ninsns, &desc,
|
est_niter);
|
est_niter);
|
if (dump_file && (dump_flags & TDF_DETAILS))
|
if (dump_file && (dump_flags & TDF_DETAILS))
|
fprintf (dump_file, "Ahead %d, unroll factor %d, trip count "
|
fprintf (dump_file, "Ahead %d, unroll factor %d, trip count "
|
HOST_WIDE_INT_PRINT_DEC "\n"
|
HOST_WIDE_INT_PRINT_DEC "\n"
|
"insn count %d, mem ref count %d, prefetch count %d\n",
|
"insn count %d, mem ref count %d, prefetch count %d\n",
|
ahead, unroll_factor, est_niter,
|
ahead, unroll_factor, est_niter,
|
ninsns, mem_ref_count, prefetch_count);
|
ninsns, mem_ref_count, prefetch_count);
|
|
|
if (!is_loop_prefetching_profitable (ahead, est_niter, ninsns,
|
if (!is_loop_prefetching_profitable (ahead, est_niter, ninsns,
|
prefetch_count, mem_ref_count))
|
prefetch_count, mem_ref_count))
|
goto fail;
|
goto fail;
|
|
|
mark_nontemporal_stores (loop, refs);
|
mark_nontemporal_stores (loop, refs);
|
|
|
/* Step 4: what to prefetch? */
|
/* Step 4: what to prefetch? */
|
if (!schedule_prefetches (refs, unroll_factor, ahead))
|
if (!schedule_prefetches (refs, unroll_factor, ahead))
|
goto fail;
|
goto fail;
|
|
|
/* Step 5: unroll the loop. TODO -- peeling of first and last few
|
/* Step 5: unroll the loop. TODO -- peeling of first and last few
|
iterations so that we do not issue superfluous prefetches. */
|
iterations so that we do not issue superfluous prefetches. */
|
if (unroll_factor != 1)
|
if (unroll_factor != 1)
|
{
|
{
|
tree_unroll_loop (loop, unroll_factor,
|
tree_unroll_loop (loop, unroll_factor,
|
single_dom_exit (loop), &desc);
|
single_dom_exit (loop), &desc);
|
unrolled = true;
|
unrolled = true;
|
}
|
}
|
|
|
/* Step 6: issue the prefetches. */
|
/* Step 6: issue the prefetches. */
|
issue_prefetches (refs, unroll_factor, ahead);
|
issue_prefetches (refs, unroll_factor, ahead);
|
|
|
fail:
|
fail:
|
release_mem_refs (refs);
|
release_mem_refs (refs);
|
return unrolled;
|
return unrolled;
|
}
|
}
|
|
|
/* Issue prefetch instructions for array references in loops. */
|
/* Issue prefetch instructions for array references in loops. */
|
|
|
unsigned int
|
unsigned int
|
tree_ssa_prefetch_arrays (void)
|
tree_ssa_prefetch_arrays (void)
|
{
|
{
|
loop_iterator li;
|
loop_iterator li;
|
struct loop *loop;
|
struct loop *loop;
|
bool unrolled = false;
|
bool unrolled = false;
|
int todo_flags = 0;
|
int todo_flags = 0;
|
|
|
if (!HAVE_prefetch
|
if (!HAVE_prefetch
|
/* It is possible to ask compiler for say -mtune=i486 -march=pentium4.
|
/* It is possible to ask compiler for say -mtune=i486 -march=pentium4.
|
-mtune=i486 causes us having PREFETCH_BLOCK 0, since this is part
|
-mtune=i486 causes us having PREFETCH_BLOCK 0, since this is part
|
of processor costs and i486 does not have prefetch, but
|
of processor costs and i486 does not have prefetch, but
|
-march=pentium4 causes HAVE_prefetch to be true. Ugh. */
|
-march=pentium4 causes HAVE_prefetch to be true. Ugh. */
|
|| PREFETCH_BLOCK == 0)
|
|| PREFETCH_BLOCK == 0)
|
return 0;
|
return 0;
|
|
|
if (dump_file && (dump_flags & TDF_DETAILS))
|
if (dump_file && (dump_flags & TDF_DETAILS))
|
{
|
{
|
fprintf (dump_file, "Prefetching parameters:\n");
|
fprintf (dump_file, "Prefetching parameters:\n");
|
fprintf (dump_file, " simultaneous prefetches: %d\n",
|
fprintf (dump_file, " simultaneous prefetches: %d\n",
|
SIMULTANEOUS_PREFETCHES);
|
SIMULTANEOUS_PREFETCHES);
|
fprintf (dump_file, " prefetch latency: %d\n", PREFETCH_LATENCY);
|
fprintf (dump_file, " prefetch latency: %d\n", PREFETCH_LATENCY);
|
fprintf (dump_file, " prefetch block size: %d\n", PREFETCH_BLOCK);
|
fprintf (dump_file, " prefetch block size: %d\n", PREFETCH_BLOCK);
|
fprintf (dump_file, " L1 cache size: %d lines, %d kB\n",
|
fprintf (dump_file, " L1 cache size: %d lines, %d kB\n",
|
L1_CACHE_SIZE_BYTES / L1_CACHE_LINE_SIZE, L1_CACHE_SIZE);
|
L1_CACHE_SIZE_BYTES / L1_CACHE_LINE_SIZE, L1_CACHE_SIZE);
|
fprintf (dump_file, " L1 cache line size: %d\n", L1_CACHE_LINE_SIZE);
|
fprintf (dump_file, " L1 cache line size: %d\n", L1_CACHE_LINE_SIZE);
|
fprintf (dump_file, " L2 cache size: %d kB\n", L2_CACHE_SIZE);
|
fprintf (dump_file, " L2 cache size: %d kB\n", L2_CACHE_SIZE);
|
fprintf (dump_file, " min insn-to-prefetch ratio: %d \n",
|
fprintf (dump_file, " min insn-to-prefetch ratio: %d \n",
|
MIN_INSN_TO_PREFETCH_RATIO);
|
MIN_INSN_TO_PREFETCH_RATIO);
|
fprintf (dump_file, " min insn-to-mem ratio: %d \n",
|
fprintf (dump_file, " min insn-to-mem ratio: %d \n",
|
PREFETCH_MIN_INSN_TO_MEM_RATIO);
|
PREFETCH_MIN_INSN_TO_MEM_RATIO);
|
fprintf (dump_file, "\n");
|
fprintf (dump_file, "\n");
|
}
|
}
|
|
|
initialize_original_copy_tables ();
|
initialize_original_copy_tables ();
|
|
|
if (!built_in_decls[BUILT_IN_PREFETCH])
|
if (!built_in_decls[BUILT_IN_PREFETCH])
|
{
|
{
|
tree type = build_function_type (void_type_node,
|
tree type = build_function_type (void_type_node,
|
tree_cons (NULL_TREE,
|
tree_cons (NULL_TREE,
|
const_ptr_type_node,
|
const_ptr_type_node,
|
NULL_TREE));
|
NULL_TREE));
|
tree decl = add_builtin_function ("__builtin_prefetch", type,
|
tree decl = add_builtin_function ("__builtin_prefetch", type,
|
BUILT_IN_PREFETCH, BUILT_IN_NORMAL,
|
BUILT_IN_PREFETCH, BUILT_IN_NORMAL,
|
NULL, NULL_TREE);
|
NULL, NULL_TREE);
|
DECL_IS_NOVOPS (decl) = true;
|
DECL_IS_NOVOPS (decl) = true;
|
built_in_decls[BUILT_IN_PREFETCH] = decl;
|
built_in_decls[BUILT_IN_PREFETCH] = decl;
|
}
|
}
|
|
|
/* We assume that size of cache line is a power of two, so verify this
|
/* We assume that size of cache line is a power of two, so verify this
|
here. */
|
here. */
|
gcc_assert ((PREFETCH_BLOCK & (PREFETCH_BLOCK - 1)) == 0);
|
gcc_assert ((PREFETCH_BLOCK & (PREFETCH_BLOCK - 1)) == 0);
|
|
|
FOR_EACH_LOOP (li, loop, LI_FROM_INNERMOST)
|
FOR_EACH_LOOP (li, loop, LI_FROM_INNERMOST)
|
{
|
{
|
if (dump_file && (dump_flags & TDF_DETAILS))
|
if (dump_file && (dump_flags & TDF_DETAILS))
|
fprintf (dump_file, "Processing loop %d:\n", loop->num);
|
fprintf (dump_file, "Processing loop %d:\n", loop->num);
|
|
|
unrolled |= loop_prefetch_arrays (loop);
|
unrolled |= loop_prefetch_arrays (loop);
|
|
|
if (dump_file && (dump_flags & TDF_DETAILS))
|
if (dump_file && (dump_flags & TDF_DETAILS))
|
fprintf (dump_file, "\n\n");
|
fprintf (dump_file, "\n\n");
|
}
|
}
|
|
|
if (unrolled)
|
if (unrolled)
|
{
|
{
|
scev_reset ();
|
scev_reset ();
|
todo_flags |= TODO_cleanup_cfg;
|
todo_flags |= TODO_cleanup_cfg;
|
}
|
}
|
|
|
free_original_copy_tables ();
|
free_original_copy_tables ();
|
return todo_flags;
|
return todo_flags;
|
}
|
}
|
|
|
