X-Git-Url: http://git.sourceforge.jp/view?p=pf3gnuchains%2Fgcc-fork.git;a=blobdiff_plain;f=gcc%2Ftree-ssa-loop-prefetch.c;h=264d97bc1b57e248a0f6a4849380b1e275520762;hp=f34214e42760e9924e906a91718dd6226d311c4e;hb=fa9f12bbe3098b9eb04e35ef6e346e0b85e5366e;hpb=1da72d7c2589433cf1a183fd4e336197f41f5450 diff --git a/gcc/tree-ssa-loop-prefetch.c b/gcc/tree-ssa-loop-prefetch.c index f34214e4276..264d97bc1b5 100644 --- a/gcc/tree-ssa-loop-prefetch.c +++ b/gcc/tree-ssa-loop-prefetch.c @@ -1,50 +1,53 @@ /* Array prefetching. - Copyright (C) 2005 Free Software Foundation, Inc. - + Copyright (C) 2005, 2007, 2008, 2009, 2010, 2011 + Free Software Foundation, Inc. + This file is part of GCC. - + 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 -Free Software Foundation; either version 2, or (at your option) any +Free Software Foundation; either version 3, or (at your option) any later version. - + GCC is distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for more details. - + You should have received a copy of the GNU General Public License -along with GCC; see the file COPYING. If not, write to the Free -Software Foundation, 59 Temple Place - Suite 330, Boston, MA -02111-1307, USA. */ +along with GCC; see the file COPYING3. If not see +. */ #include "config.h" #include "system.h" #include "coretypes.h" #include "tm.h" #include "tree.h" -#include "rtl.h" #include "tm_p.h" -#include "hard-reg-set.h" #include "basic-block.h" #include "output.h" -#include "diagnostic.h" +#include "tree-pretty-print.h" #include "tree-flow.h" #include "tree-dump.h" #include "timevar.h" #include "cfgloop.h" -#include "varray.h" -#include "expr.h" #include "tree-pass.h" -#include "ggc.h" #include "insn-config.h" #include "recog.h" #include "hashtab.h" #include "tree-chrec.h" #include "tree-scalar-evolution.h" -#include "toplev.h" +#include "diagnostic-core.h" #include "params.h" #include "langhooks.h" +#include "tree-inline.h" +#include "tree-data-ref.h" + + +/* FIXME: Needed for optabs, but this should all be moved to a TBD interface + between the GIMPLE and RTL worlds. */ +#include "expr.h" +#include "optabs.h" /* This pass inserts prefetch instructions to optimize cache usage during accesses to arrays in loops. It processes loops sequentially and: @@ -77,10 +80,14 @@ Software Foundation, 59 Temple Place - Suite 330, Boston, MA (2) has PREFETCH_MOD 64 (3) has PREFETCH_MOD 4 (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 (5) is the same with probability only 7/32. (5) has PREFETCH_MOD 1 as well. + Additionally, we use data dependence analysis to determine for each + reference the distance till the first reuse; this information is used + to determine the temporality of the issued prefetch instruction. + 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 the loop. The problem is that we do not know either of these values, @@ -93,16 +100,43 @@ Software Foundation, 59 Temple Place - Suite 330, Boston, MA while still within this bound (starting with those with lowest prefetch_mod, since they are responsible for most of the cache misses). - + 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 prefetching nonaccessed memory. TODO -- actually implement peeling. - + 6) We actually emit the prefetch instructions. ??? Perhaps emit the prefetch instructions with guards in cases where 5) was not sufficient to satisfy the constraints? + A cost model is implemented to determine whether or not prefetching is + profitable for a given loop. The cost model has three heuristics: + + 1. Function trip_count_to_ahead_ratio_too_small_p implements a + heuristic that determines whether or not the loop has too few + iterations (compared to ahead). Prefetching is not likely to be + beneficial if the trip count to ahead ratio is below a certain + minimum. + + 2. Function mem_ref_count_reasonable_p implements a heuristic that + determines whether the given loop has enough CPU ops that can be + overlapped with cache missing memory ops. If not, the loop + won't benefit from prefetching. In the implementation, + prefetching is not considered beneficial if the ratio between + the instruction count and the mem ref count is below a certain + minimum. + + 3. Function insn_to_prefetch_ratio_too_small_p implements a + heuristic that disables prefetching in a loop if the prefetching + cost is above a certain limit. The relative prefetching cost is + estimated by taking the ratio between the prefetch count and the + total intruction count (this models the I-cache cost). + + The limits used in these heuristics are defined as parameters with + reasonable default values. Machine-specific default values will be + added later. + Some other TODO: -- write and use more general reuse analysis (that could be also used in other cache aimed loop optimizations) @@ -115,19 +149,6 @@ Software Foundation, 59 Temple Place - Suite 330, Boston, MA /* Magic constants follow. These should be replaced by machine specific numbers. */ -/* A number that should roughly correspond to the number of instructions - executed before the prefetch is completed. */ - -#ifndef PREFETCH_LATENCY -#define PREFETCH_LATENCY 200 -#endif - -/* Number of prefetches that can run at the same time. */ - -#ifndef SIMULTANEOUS_PREFETCHES -#define SIMULTANEOUS_PREFETCHES 3 -#endif - /* True if write can be prefetched by a read prefetch. */ #ifndef WRITE_CAN_USE_READ_PREFETCH @@ -140,10 +161,12 @@ Software Foundation, 59 Temple Place - Suite 330, Boston, MA #define READ_CAN_USE_WRITE_PREFETCH 0 #endif -/* Cache line size. Assumed to be a power of two. */ +/* 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 + of cache hierarchy). */ #ifndef PREFETCH_BLOCK -#define PREFETCH_BLOCK 32 +#define PREFETCH_BLOCK L1_CACHE_LINE_SIZE #endif /* Do we have a forward hardware sequential prefetching? */ @@ -161,7 +184,7 @@ Software Foundation, 59 Temple Place - Suite 330, Boston, MA /* 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 case, we issue the prefetches for both of them if this probability - is less then (1000 - ACCEPTABLE_MISS_RATE) promile. */ + is less then (1000 - ACCEPTABLE_MISS_RATE) per thousand. */ #ifndef ACCEPTABLE_MISS_RATE #define ACCEPTABLE_MISS_RATE 50 @@ -171,12 +194,41 @@ Software Foundation, 59 Temple Place - Suite 330, Boston, MA #define HAVE_prefetch 0 #endif +#define L1_CACHE_SIZE_BYTES ((unsigned) (L1_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 + after L2_CACHE_SIZE_BYTES of memory are accessed. However, we ignore + accesses closer than L1_CACHE_SIZE_BYTES / NONTEMPORAL_FRACTION, + so that we use nontemporal prefetches e.g. if single memory location + is accessed several times in a single iteration of the loop. */ +#define NONTEMPORAL_FRACTION 16 + +/* In case we have to emit a memory fence instruction after the loop that + uses nontemporal stores, this defines the builtin to use. */ + +#ifndef FENCE_FOLLOWING_MOVNT +#define FENCE_FOLLOWING_MOVNT NULL_TREE +#endif + +/* It is not profitable to prefetch when the trip count is not at + least TRIP_COUNT_TO_AHEAD_RATIO times the prefetch ahead distance. + For example, in a loop with a prefetch ahead distance of 10, + supposing that TRIP_COUNT_TO_AHEAD_RATIO is equal to 4, it is + profitable to prefetch when the trip count is greater or equal to + 40. In that case, 30 out of the 40 iterations will benefit from + prefetching. */ + +#ifndef TRIP_COUNT_TO_AHEAD_RATIO +#define TRIP_COUNT_TO_AHEAD_RATIO 4 +#endif + /* The group of references between that reuse may occur. */ struct mem_ref_group { tree base; /* Base of the reference. */ - HOST_WIDE_INT step; /* Step of the reference. */ + tree step; /* Step of the reference. */ struct mem_ref *refs; /* References in the group. */ struct mem_ref_group *next; /* Next group of references. */ }; @@ -185,14 +237,32 @@ struct mem_ref_group #define PREFETCH_ALL (~(unsigned HOST_WIDE_INT) 0) +/* Do not generate a prefetch if the unroll factor is significantly less + than what is required by the prefetch. This is to avoid redundant + prefetches. For example, when prefetch_mod is 16 and unroll_factor is + 2, prefetching requires unrolling the loop 16 times, but + the loop is actually unrolled twice. In this case (ratio = 8), + prefetching is not likely to be beneficial. */ + +#ifndef PREFETCH_MOD_TO_UNROLL_FACTOR_RATIO +#define PREFETCH_MOD_TO_UNROLL_FACTOR_RATIO 4 +#endif + +/* Some of the prefetch computations have quadratic complexity. We want to + avoid huge compile times and, therefore, want to limit the amount of + memory references per loop where we consider prefetching. */ + +#ifndef PREFETCH_MAX_MEM_REFS_PER_LOOP +#define PREFETCH_MAX_MEM_REFS_PER_LOOP 200 +#endif + /* The memory reference. */ struct mem_ref { - tree stmt; /* Statement in that the reference appears. */ + gimple stmt; /* Statement in that the reference appears. */ tree mem; /* The reference. */ HOST_WIDE_INT delta; /* Constant offset of the reference. */ - bool write_p; /* Is it a write? */ struct mem_ref_group *group; /* The group of references it belongs to. */ unsigned HOST_WIDE_INT prefetch_mod; /* Prefetch only each PREFETCH_MOD-th @@ -200,8 +270,15 @@ struct mem_ref unsigned HOST_WIDE_INT prefetch_before; /* Prefetch only first PREFETCH_BEFORE iterations. */ - bool issue_prefetch_p; /* Should we really issue the prefetch? */ + unsigned reuse_distance; /* The amount of data accessed before the first + reuse of this value. */ struct mem_ref *next; /* The next reference in the group. */ + unsigned write_p : 1; /* Is it a write? */ + unsigned independent_p : 1; /* True if the reference is independent on + all other references inside the loop. */ + unsigned issue_prefetch_p : 1; /* Should we really issue the prefetch? */ + unsigned storent_p : 1; /* True if we changed the store to a + nontemporal one. */ }; /* Dumps information about reference REF to FILE. */ @@ -214,7 +291,10 @@ dump_mem_ref (FILE *file, struct mem_ref *ref) fprintf (file, " group %p (base ", (void *) ref->group); print_generic_expr (file, ref->group->base, TDF_SLIM); fprintf (file, ", step "); - fprintf (file, HOST_WIDE_INT_PRINT_DEC, ref->group->step); + if (cst_and_fits_in_hwi (ref->group->step)) + fprintf (file, HOST_WIDE_INT_PRINT_DEC, int_cst_value (ref->group->step)); + else + print_generic_expr (file, ref->group->step, TDF_TREE); fprintf (file, ")\n"); fprintf (file, " delta "); @@ -230,23 +310,24 @@ dump_mem_ref (FILE *file, struct mem_ref *ref) exist. */ static struct mem_ref_group * -find_or_create_group (struct mem_ref_group **groups, tree base, - HOST_WIDE_INT step) +find_or_create_group (struct mem_ref_group **groups, tree base, tree step) { struct mem_ref_group *group; for (; *groups; groups = &(*groups)->next) { - if ((*groups)->step == step + if (operand_equal_p ((*groups)->step, step, 0) && operand_equal_p ((*groups)->base, base, 0)) return *groups; - /* Keep the list of groups sorted by decreasing step. */ - if ((*groups)->step < step) + /* If step is an integer constant, keep the list of groups sorted + by decreasing step. */ + if (cst_and_fits_in_hwi ((*groups)->step) && cst_and_fits_in_hwi (step) + && int_cst_value ((*groups)->step) < int_cst_value (step)) break; } - group = xcalloc (1, sizeof (struct mem_ref_group)); + group = XNEW (struct mem_ref_group); group->base = base; group->step = step; group->refs = NULL; @@ -260,7 +341,7 @@ find_or_create_group (struct mem_ref_group **groups, tree base, WRITE_P. The reference occurs in statement STMT. */ static void -record_ref (struct mem_ref_group *group, tree stmt, tree mem, +record_ref (struct mem_ref_group *group, gimple stmt, tree mem, HOST_WIDE_INT delta, bool write_p) { struct mem_ref **aref; @@ -283,16 +364,19 @@ record_ref (struct mem_ref_group *group, tree stmt, tree mem, return; } - (*aref) = xcalloc (1, sizeof (struct mem_ref)); + (*aref) = XNEW (struct mem_ref); (*aref)->stmt = stmt; (*aref)->mem = mem; (*aref)->delta = delta; (*aref)->write_p = write_p; (*aref)->prefetch_before = PREFETCH_ALL; (*aref)->prefetch_mod = 1; + (*aref)->reuse_distance = 0; (*aref)->issue_prefetch_p = false; (*aref)->group = group; (*aref)->next = NULL; + (*aref)->independent_p = false; + (*aref)->storent_p = false; if (dump_file && (dump_flags & TDF_DETAILS)) dump_mem_ref (dump_file, *aref); @@ -323,8 +407,8 @@ release_mem_refs (struct mem_ref_group *groups) struct ar_data { struct loop *loop; /* Loop of the reference. */ - tree stmt; /* Statement of the reference. */ - HOST_WIDE_INT *step; /* Step of the memory reference. */ + gimple stmt; /* Statement of the reference. */ + tree *step; /* Step of the memory reference. */ HOST_WIDE_INT *delta; /* Offset of the memory reference. */ }; @@ -334,30 +418,18 @@ struct ar_data static bool idx_analyze_ref (tree base, tree *index, void *data) { - struct ar_data *ar_data = data; + struct ar_data *ar_data = (struct ar_data *) data; tree ibase, step, stepsize; - HOST_WIDE_INT istep, idelta = 0, imult = 1; + HOST_WIDE_INT idelta = 0, imult = 1; affine_iv iv; - if (TREE_CODE (base) == MISALIGNED_INDIRECT_REF - || TREE_CODE (base) == ALIGN_INDIRECT_REF) - return false; - - if (!simple_iv (ar_data->loop, ar_data->stmt, *index, &iv, false)) + if (!simple_iv (ar_data->loop, loop_containing_stmt (ar_data->stmt), + *index, &iv, true)) return false; ibase = iv.base; step = iv.step; - if (zero_p (step)) - istep = 0; - else - { - if (!cst_and_fits_in_hwi (step)) - return false; - istep = int_cst_value (step); - } - - if (TREE_CODE (ibase) == PLUS_EXPR + if (TREE_CODE (ibase) == POINTER_PLUS_EXPR && cst_and_fits_in_hwi (TREE_OPERAND (ibase, 1))) { idelta = int_cst_value (TREE_OPERAND (ibase, 1)); @@ -375,12 +447,18 @@ idx_analyze_ref (tree base, tree *index, void *data) if (!cst_and_fits_in_hwi (stepsize)) return false; imult = int_cst_value (stepsize); - - istep *= imult; + step = fold_build2 (MULT_EXPR, sizetype, + fold_convert (sizetype, step), + fold_convert (sizetype, stepsize)); idelta *= imult; } - *ar_data->step += istep; + if (*ar_data->step == NULL_TREE) + *ar_data->step = step; + else + *ar_data->step = fold_build2 (PLUS_EXPR, sizetype, + fold_convert (sizetype, *ar_data->step), + fold_convert (sizetype, step)); *ar_data->delta += idelta; *index = ibase; @@ -394,21 +472,29 @@ idx_analyze_ref (tree base, tree *index, void *data) static bool analyze_ref (struct loop *loop, tree *ref_p, tree *base, - HOST_WIDE_INT *step, HOST_WIDE_INT *delta, - tree stmt) + tree *step, HOST_WIDE_INT *delta, + gimple stmt) { struct ar_data ar_data; tree off; HOST_WIDE_INT bit_offset; tree ref = *ref_p; - *step = 0; + *step = NULL_TREE; *delta = 0; - /* First strip off the component references. Ignore bitfields. */ - if (TREE_CODE (ref) == COMPONENT_REF - && DECL_NONADDRESSABLE_P (TREE_OPERAND (ref, 1))) - ref = TREE_OPERAND (ref, 0); + /* First strip off the component references. Ignore bitfields. + Also strip off the real and imagine parts of a complex, so that + they can have the same base. */ + if (TREE_CODE (ref) == REALPART_EXPR + || TREE_CODE (ref) == IMAGPART_EXPR + || (TREE_CODE (ref) == COMPONENT_REF + && DECL_NONADDRESSABLE_P (TREE_OPERAND (ref, 1)))) + { + if (TREE_CODE (ref) == IMAGPART_EXPR) + *delta += int_size_in_bytes (TREE_TYPE (ref)); + ref = TREE_OPERAND (ref, 0); + } *ref_p = ref; @@ -417,7 +503,7 @@ analyze_ref (struct loop *loop, tree *ref_p, tree *base, off = DECL_FIELD_BIT_OFFSET (TREE_OPERAND (ref, 1)); bit_offset = TREE_INT_CST_LOW (off); gcc_assert (bit_offset % BITS_PER_UNIT == 0); - + *delta += bit_offset / BITS_PER_UNIT; } @@ -430,37 +516,59 @@ analyze_ref (struct loop *loop, tree *ref_p, tree *base, } /* Record a memory reference REF to the list REFS. The reference occurs in - LOOP in statement STMT and it is write if WRITE_P. */ + LOOP in statement STMT and it is write if WRITE_P. Returns true if the + reference was recorded, false otherwise. */ -static void +static bool gather_memory_references_ref (struct loop *loop, struct mem_ref_group **refs, - tree ref, bool write_p, tree stmt) + tree ref, bool write_p, gimple stmt) { - tree base; - HOST_WIDE_INT step, delta; + tree base, step; + HOST_WIDE_INT delta; struct mem_ref_group *agrp; + if (get_base_address (ref) == NULL) + return false; + if (!analyze_ref (loop, &ref, &base, &step, &delta, stmt)) - return; + return false; + /* If analyze_ref fails the default is a NULL_TREE. We can stop here. */ + if (step == NULL_TREE) + return false; + + /* Stop if the address of BASE could not be taken. */ + if (may_be_nonaddressable_p (base)) + return false; + + /* Limit non-constant step prefetching only to the innermost loops. */ + if (!cst_and_fits_in_hwi (step) && loop->inner != NULL) + return false; /* Now we know that REF = &BASE + STEP * iter + DELTA, where DELTA and STEP are integer constants. */ agrp = find_or_create_group (refs, base, step); record_ref (agrp, stmt, ref, delta, write_p); + + return true; } -/* Record the suitable memory references in LOOP. */ +/* Record the suitable memory references in LOOP. NO_OTHER_REFS is set to + true if there are no other memory references inside the loop. */ static struct mem_ref_group * -gather_memory_references (struct loop *loop) +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 bb; unsigned i; - block_stmt_iterator bsi; - tree stmt, lhs, rhs; + gimple_stmt_iterator bsi; + gimple stmt; + tree lhs, rhs; struct mem_ref_group *refs = NULL; + *no_other_refs = true; + *ref_count = 0; + /* Scan the loop body in order, so that the former references precede the later ones. */ for (i = 0; i < loop->num_nodes; i++) @@ -469,19 +577,34 @@ gather_memory_references (struct loop *loop) if (bb->loop_father != loop) continue; - for (bsi = bsi_start (bb); !bsi_end_p (bsi); bsi_next (&bsi)) + for (bsi = gsi_start_bb (bb); !gsi_end_p (bsi); gsi_next (&bsi)) { - stmt = bsi_stmt (bsi); - if (TREE_CODE (stmt) != MODIFY_EXPR) - continue; + stmt = gsi_stmt (bsi); + + if (gimple_code (stmt) != GIMPLE_ASSIGN) + { + if (gimple_vuse (stmt) + || (is_gimple_call (stmt) + && !(gimple_call_flags (stmt) & ECF_CONST))) + *no_other_refs = false; + continue; + } - lhs = TREE_OPERAND (stmt, 0); - rhs = TREE_OPERAND (stmt, 1); + lhs = gimple_assign_lhs (stmt); + rhs = gimple_assign_rhs1 (stmt); if (REFERENCE_CLASS_P (rhs)) - gather_memory_references_ref (loop, &refs, rhs, false, stmt); + { + *no_other_refs &= gather_memory_references_ref (loop, &refs, + rhs, false, stmt); + *ref_count += 1; + } if (REFERENCE_CLASS_P (lhs)) - gather_memory_references_ref (loop, &refs, lhs, true, stmt); + { + *no_other_refs &= gather_memory_references_ref (loop, &refs, + lhs, true, stmt); + *ref_count += 1; + } } } free (body); @@ -494,8 +617,16 @@ gather_memory_references (struct loop *loop) static void prune_ref_by_self_reuse (struct mem_ref *ref) { - HOST_WIDE_INT step = ref->group->step; - bool backward = step < 0; + HOST_WIDE_INT step; + bool backward; + + /* If the step size is non constant, we cannot calculate prefetch_mod. */ + if (!cst_and_fits_in_hwi (ref->group->step)) + return; + + step = int_cst_value (ref->group->step); + + backward = step < 0; if (step == 0) { @@ -533,6 +664,54 @@ ddown (HOST_WIDE_INT x, unsigned HOST_WIDE_INT by) return (x + by - 1) / by; } +/* Given a CACHE_LINE_SIZE and two inductive memory references + with a common STEP greater than CACHE_LINE_SIZE and an address + difference DELTA, compute the probability that they will fall + in different cache lines. Return true if the computed miss rate + is not greater than the ACCEPTABLE_MISS_RATE. DISTINCT_ITERS is the + number of distinct iterations after which the pattern repeats itself. + ALIGN_UNIT is the unit of alignment in bytes. */ + +static bool +is_miss_rate_acceptable (unsigned HOST_WIDE_INT cache_line_size, + HOST_WIDE_INT step, HOST_WIDE_INT delta, + unsigned HOST_WIDE_INT distinct_iters, + int align_unit) +{ + unsigned align, iter; + int total_positions, miss_positions, max_allowed_miss_positions; + int address1, address2, cache_line1, cache_line2; + + /* It always misses if delta is greater than or equal to the cache + line size. */ + if (delta >= (HOST_WIDE_INT) cache_line_size) + return false; + + miss_positions = 0; + total_positions = (cache_line_size / align_unit) * distinct_iters; + max_allowed_miss_positions = (ACCEPTABLE_MISS_RATE * total_positions) / 1000; + + /* Iterate through all possible alignments of the first + memory reference within its cache line. */ + for (align = 0; align < cache_line_size; align += align_unit) + + /* Iterate through all distinct iterations. */ + for (iter = 0; iter < distinct_iters; iter++) + { + address1 = align + step * iter; + address2 = address1 + delta; + cache_line1 = address1 / cache_line_size; + cache_line2 = address2 / cache_line_size; + if (cache_line1 != cache_line2) + { + miss_positions += 1; + if (miss_positions > max_allowed_miss_positions) + return false; + } + } + return true; +} + /* Prune the prefetch candidate REF using the reuse with BY. If BY_IS_BEFORE is true, BY is before REF in the loop. */ @@ -540,12 +719,26 @@ static void prune_ref_by_group_reuse (struct mem_ref *ref, struct mem_ref *by, bool by_is_before) { - HOST_WIDE_INT step = ref->group->step; - bool backward = step < 0; + HOST_WIDE_INT step; + bool backward; HOST_WIDE_INT delta_r = ref->delta, delta_b = by->delta; HOST_WIDE_INT delta = delta_b - delta_r; HOST_WIDE_INT hit_from; unsigned HOST_WIDE_INT prefetch_before, prefetch_block; + HOST_WIDE_INT reduced_step; + unsigned HOST_WIDE_INT reduced_prefetch_block; + tree ref_type; + int align_unit; + + /* If the step is non constant we cannot calculate prefetch_before. */ + if (!cst_and_fits_in_hwi (ref->group->step)) { + return; + } + + step = int_cst_value (ref->group->step); + + backward = step < 0; + if (delta == 0) { @@ -553,7 +746,7 @@ prune_ref_by_group_reuse (struct mem_ref *ref, struct mem_ref *by, former. */ if (by_is_before) ref->prefetch_before = 0; - + return; } @@ -601,32 +794,41 @@ prune_ref_by_group_reuse (struct mem_ref *ref, struct mem_ref *by, hit_from = ddown (delta_b, PREFETCH_BLOCK) * PREFETCH_BLOCK; prefetch_before = (hit_from - delta_r + step - 1) / step; + /* Do not reduce prefetch_before if we meet beyond cache size. */ + if (prefetch_before > absu_hwi (L2_CACHE_SIZE_BYTES / step)) + prefetch_before = PREFETCH_ALL; if (prefetch_before < ref->prefetch_before) ref->prefetch_before = prefetch_before; return; } - /* A more complicated case. First let us ensure that size of cache line - and step are coprime (here we assume that PREFETCH_BLOCK is a power - of two. */ + /* A more complicated case with step > prefetch_block. First reduce + the ratio between the step and the cache line size to its simplest + terms. The resulting denominator will then represent the number of + distinct iterations after which each address will go back to its + initial location within the cache line. This computation assumes + that PREFETCH_BLOCK is a power of two. */ prefetch_block = PREFETCH_BLOCK; - while ((step & 1) == 0 - && prefetch_block > 1) + reduced_prefetch_block = prefetch_block; + reduced_step = step; + while ((reduced_step & 1) == 0 + && reduced_prefetch_block > 1) { - step >>= 1; - prefetch_block >>= 1; - delta >>= 1; + reduced_step >>= 1; + reduced_prefetch_block >>= 1; } - /* Now step > prefetch_block, and step and prefetch_block are coprime. - Determine the probability that the accesses hit the same cache line. */ - prefetch_before = delta / step; delta %= step; - if ((unsigned HOST_WIDE_INT) delta - <= (prefetch_block * ACCEPTABLE_MISS_RATE / 1000)) + ref_type = TREE_TYPE (ref->mem); + align_unit = TYPE_ALIGN (ref_type) / 8; + if (is_miss_rate_acceptable (prefetch_block, step, delta, + reduced_prefetch_block, align_unit)) { + /* Do not reduce prefetch_before if we meet beyond cache size. */ + if (prefetch_before > L2_CACHE_SIZE_BYTES / PREFETCH_BLOCK) + prefetch_before = PREFETCH_ALL; if (prefetch_before < ref->prefetch_before) ref->prefetch_before = prefetch_before; @@ -636,8 +838,8 @@ prune_ref_by_group_reuse (struct mem_ref *ref, struct mem_ref *by, /* Try also the following iteration. */ prefetch_before++; delta = step - delta; - if ((unsigned HOST_WIDE_INT) delta - <= (prefetch_block * ACCEPTABLE_MISS_RATE / 1000)) + if (is_miss_rate_acceptable (prefetch_block, step, delta, + reduced_prefetch_block, align_unit)) { if (prefetch_before < ref->prefetch_before) ref->prefetch_before = prefetch_before; @@ -740,7 +942,20 @@ should_issue_prefetch_p (struct mem_ref *ref) /* For now do not issue prefetches for only first few of the iterations. */ if (ref->prefetch_before != PREFETCH_ALL) - return false; + { + if (dump_file && (dump_flags & TDF_DETAILS)) + fprintf (dump_file, "Ignoring %p due to prefetch_before\n", + (void *) ref); + return false; + } + + /* Do not prefetch nontemporal stores. */ + if (ref->storent_p) + { + if (dump_file && (dump_flags & TDF_DETAILS)) + fprintf (dump_file, "Ignoring nontemporal store %p\n", (void *) ref); + return false; + } return true; } @@ -755,19 +970,21 @@ static bool schedule_prefetches (struct mem_ref_group *groups, unsigned unroll_factor, unsigned ahead) { - unsigned max_prefetches, n_prefetches; + unsigned remaining_prefetch_slots, n_prefetches, prefetch_slots; + unsigned slots_per_prefetch; struct mem_ref *ref; bool any = false; - max_prefetches = (SIMULTANEOUS_PREFETCHES * unroll_factor) / ahead; - if (max_prefetches > (unsigned) SIMULTANEOUS_PREFETCHES) - max_prefetches = SIMULTANEOUS_PREFETCHES; + /* At most SIMULTANEOUS_PREFETCHES should be running at the same time. */ + remaining_prefetch_slots = SIMULTANEOUS_PREFETCHES; + /* The prefetch will run for AHEAD iterations of the original loop, i.e., + AHEAD / UNROLL_FACTOR iterations of the unrolled loop. In each iteration, + it will need a prefetch slot. */ + slots_per_prefetch = (ahead + unroll_factor / 2) / unroll_factor; if (dump_file && (dump_flags & TDF_DETAILS)) - fprintf (dump_file, "Max prefetches to issue: %d.\n", max_prefetches); - - if (!max_prefetches) - return false; + fprintf (dump_file, "Each prefetch instruction takes %u prefetch slots.\n", + slots_per_prefetch); /* For now we just take memory references one by one and issue prefetches for as many as possible. The groups are sorted @@ -780,36 +997,72 @@ schedule_prefetches (struct mem_ref_group *groups, unsigned unroll_factor, if (!should_issue_prefetch_p (ref)) continue; - ref->issue_prefetch_p = true; + /* The loop is far from being sufficiently unrolled for this + prefetch. Do not generate prefetch to avoid many redudant + prefetches. */ + if (ref->prefetch_mod / unroll_factor > PREFETCH_MOD_TO_UNROLL_FACTOR_RATIO) + continue; - /* If prefetch_mod is less then unroll_factor, we need to insert - several prefetches for the reference. */ + /* If we need to prefetch the reference each PREFETCH_MOD iterations, + and we unroll the loop UNROLL_FACTOR times, we need to insert + ceil (UNROLL_FACTOR / PREFETCH_MOD) instructions in each + iteration. */ n_prefetches = ((unroll_factor + ref->prefetch_mod - 1) / ref->prefetch_mod); - if (max_prefetches <= n_prefetches) - return true; + prefetch_slots = n_prefetches * slots_per_prefetch; - max_prefetches -= n_prefetches; + /* If more than half of the prefetches would be lost anyway, do not + issue the prefetch. */ + if (2 * remaining_prefetch_slots < prefetch_slots) + continue; + + ref->issue_prefetch_p = true; + + if (remaining_prefetch_slots <= prefetch_slots) + return true; + remaining_prefetch_slots -= prefetch_slots; any = true; } return any; } -/* Determine whether there is any reference suitable for prefetching - in GROUPS. */ +/* Return TRUE if no prefetch is going to be generated in the given + GROUPS. */ static bool -anything_to_prefetch_p (struct mem_ref_group *groups) +nothing_to_prefetch_p (struct mem_ref_group *groups) { struct mem_ref *ref; for (; groups; groups = groups->next) for (ref = groups->refs; ref; ref = ref->next) if (should_issue_prefetch_p (ref)) - return true; + return false; - return false; + return true; +} + +/* Estimate the number of prefetches in the given GROUPS. + UNROLL_FACTOR is the factor by which LOOP was unrolled. */ + +static int +estimate_prefetch_count (struct mem_ref_group *groups, unsigned unroll_factor) +{ + struct mem_ref *ref; + unsigned n_prefetches; + int prefetch_count = 0; + + for (; groups; groups = groups->next) + for (ref = groups->refs; ref; ref = ref->next) + if (should_issue_prefetch_p (ref)) + { + n_prefetches = ((unroll_factor + ref->prefetch_mod - 1) + / ref->prefetch_mod); + prefetch_count += n_prefetches; + } + + return prefetch_count; } /* Issue prefetches for the reference REF into loop as decided before. @@ -820,36 +1073,53 @@ static void issue_prefetch_ref (struct mem_ref *ref, unsigned unroll_factor, unsigned ahead) { HOST_WIDE_INT delta; - tree addr, addr_base, prefetch, params, write_p; - block_stmt_iterator bsi; + tree addr, addr_base, write_p, local, forward; + gimple prefetch; + gimple_stmt_iterator bsi; unsigned n_prefetches, ap; + bool nontemporal = ref->reuse_distance >= L2_CACHE_SIZE_BYTES; if (dump_file && (dump_flags & TDF_DETAILS)) - fprintf (dump_file, "Issued prefetch for %p.\n", (void *) ref); + fprintf (dump_file, "Issued%s prefetch for %p.\n", + nontemporal ? " nontemporal" : "", + (void *) ref); - bsi = bsi_for_stmt (ref->stmt); + bsi = gsi_for_stmt (ref->stmt); n_prefetches = ((unroll_factor + ref->prefetch_mod - 1) / ref->prefetch_mod); addr_base = build_fold_addr_expr_with_type (ref->mem, ptr_type_node); - addr_base = force_gimple_operand_bsi (&bsi, unshare_expr (addr_base), true, NULL); + addr_base = force_gimple_operand_gsi (&bsi, unshare_expr (addr_base), + true, NULL, true, GSI_SAME_STMT); + write_p = ref->write_p ? integer_one_node : integer_zero_node; + local = nontemporal ? integer_zero_node : integer_three_node; for (ap = 0; ap < n_prefetches; ap++) { - /* Determine the address to prefetch. */ - delta = (ahead + ap * ref->prefetch_mod) * ref->group->step; - addr = fold_build2 (PLUS_EXPR, ptr_type_node, - addr_base, build_int_cst (ptr_type_node, delta)); - addr = force_gimple_operand_bsi (&bsi, unshare_expr (addr), true, NULL); - + if (cst_and_fits_in_hwi (ref->group->step)) + { + /* Determine the address to prefetch. */ + delta = (ahead + ap * ref->prefetch_mod) * + int_cst_value (ref->group->step); + addr = fold_build_pointer_plus_hwi (addr_base, delta); + addr = force_gimple_operand_gsi (&bsi, unshare_expr (addr), true, NULL, + true, GSI_SAME_STMT); + } + else + { + /* The step size is non-constant but loop-invariant. We use the + heuristic to simply prefetch ahead iterations ahead. */ + forward = fold_build2 (MULT_EXPR, sizetype, + fold_convert (sizetype, ref->group->step), + fold_convert (sizetype, size_int (ahead))); + addr = fold_build_pointer_plus (addr_base, forward); + addr = force_gimple_operand_gsi (&bsi, unshare_expr (addr), true, + NULL, true, GSI_SAME_STMT); + } /* Create the prefetch instruction. */ - write_p = ref->write_p ? integer_one_node : integer_zero_node; - params = tree_cons (NULL_TREE, addr, - tree_cons (NULL_TREE, write_p, NULL_TREE)); - - prefetch = build_function_call_expr (built_in_decls[BUILT_IN_PREFETCH], - params); - bsi_insert_before (&bsi, prefetch, BSI_SAME_STMT); + prefetch = gimple_build_call (builtin_decl_explicit (BUILT_IN_PREFETCH), + 3, addr, write_p, local); + gsi_insert_before (&bsi, prefetch, GSI_SAME_STMT); } } @@ -869,6 +1139,130 @@ issue_prefetches (struct mem_ref_group *groups, issue_prefetch_ref (ref, unroll_factor, ahead); } +/* Returns true if REF is a memory write for that a nontemporal store insn + can be used. */ + +static bool +nontemporal_store_p (struct mem_ref *ref) +{ + enum machine_mode mode; + enum insn_code code; + + /* 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 + be reordered with respect to other memory references. */ + if (!ref->write_p + || !ref->independent_p + || ref->reuse_distance < L2_CACHE_SIZE_BYTES) + return false; + + /* Check that we have the storent instruction for the mode. */ + mode = TYPE_MODE (TREE_TYPE (ref->mem)); + if (mode == BLKmode) + return false; + + code = optab_handler (storent_optab, mode); + return code != CODE_FOR_nothing; +} + +/* If REF is a nontemporal store, we mark the corresponding modify statement + and return true. Otherwise, we return false. */ + +static bool +mark_nontemporal_store (struct mem_ref *ref) +{ + if (!nontemporal_store_p (ref)) + return false; + + if (dump_file && (dump_flags & TDF_DETAILS)) + fprintf (dump_file, "Marked reference %p as a nontemporal store.\n", + (void *) ref); + + gimple_assign_set_nontemporal_move (ref->stmt, true); + ref->storent_p = true; + + return true; +} + +/* Issue a memory fence instruction after LOOP. */ + +static void +emit_mfence_after_loop (struct loop *loop) +{ + VEC (edge, heap) *exits = get_loop_exit_edges (loop); + edge exit; + gimple call; + gimple_stmt_iterator bsi; + unsigned i; + + FOR_EACH_VEC_ELT (edge, exits, i, exit) + { + call = gimple_build_call (FENCE_FOLLOWING_MOVNT, 0); + + if (!single_pred_p (exit->dest) + /* If possible, we prefer not to insert the fence on other paths + in cfg. */ + && !(exit->flags & EDGE_ABNORMAL)) + split_loop_exit_edge (exit); + bsi = gsi_after_labels (exit->dest); + + gsi_insert_before (&bsi, call, GSI_NEW_STMT); + mark_virtual_ops_for_renaming (call); + } + + VEC_free (edge, heap, exits); + update_ssa (TODO_update_ssa_only_virtuals); +} + +/* Returns true if we can use storent in loop, false otherwise. */ + +static bool +may_use_storent_in_loop_p (struct loop *loop) +{ + bool ret = true; + + if (loop->inner != NULL) + return false; + + /* 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. */ + if (FENCE_FOLLOWING_MOVNT != NULL_TREE) + { + VEC (edge, heap) *exits = get_loop_exit_edges (loop); + unsigned i; + edge exit; + + FOR_EACH_VEC_ELT (edge, exits, i, exit) + if ((exit->flags & EDGE_ABNORMAL) + && exit->dest == EXIT_BLOCK_PTR) + ret = false; + + VEC_free (edge, heap, exits); + } + + return ret; +} + +/* Marks nontemporal stores in LOOP. GROUPS contains the description of memory + references in the loop. */ + +static void +mark_nontemporal_stores (struct loop *loop, struct mem_ref_group *groups) +{ + struct mem_ref *ref; + bool any = false; + + if (!may_use_storent_in_loop_p (loop)) + return; + + for (; groups; groups = groups->next) + for (ref = groups->refs; ref; ref = ref->next) + any |= mark_nontemporal_store (ref); + + if (any && FENCE_FOLLOWING_MOVNT != NULL_TREE) + emit_mfence_after_loop (loop); +} + /* Determines whether we can profitably unroll LOOP FACTOR times, and if this is the case, fill in DESC by the description of number of iterations. */ @@ -893,54 +1287,49 @@ should_unroll_loop_p (struct loop *loop, struct tree_niter_desc *desc, /* Determine the coefficient by that unroll LOOP, from the information contained in the list of memory references REFS. Description of - umber of iterations of LOOP is stored to DESC. AHEAD is the number - of iterations ahead that we need to prefetch. NINSNS is number of - insns of the LOOP. */ + 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 + the loop, or -1 if no estimate is available. */ static unsigned determine_unroll_factor (struct loop *loop, struct mem_ref_group *refs, - unsigned ahead, unsigned ninsns, - struct tree_niter_desc *desc) + unsigned ninsns, struct tree_niter_desc *desc, + HOST_WIDE_INT est_niter) { - unsigned upper_bound, size_factor, constraint_factor; - unsigned factor, max_mod_constraint, ahead_factor; + unsigned upper_bound; + unsigned nfactor, factor, mod_constraint; struct mem_ref_group *agp; struct mem_ref *ref; - upper_bound = PARAM_VALUE (PARAM_MAX_UNROLL_TIMES); - - /* First check whether the loop is not too large to unroll. */ - size_factor = PARAM_VALUE (PARAM_MAX_UNROLLED_INSNS) / ninsns; - if (size_factor <= 1) + /* First check whether the loop is not too large to unroll. We ignore + PARAM_MAX_UNROLL_TIMES, because for small loops, it prevented us + 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 + 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 + the case here. */ + upper_bound = PARAM_VALUE (PARAM_MAX_UNROLLED_INSNS) / ninsns; + + /* If we unrolled the loop more times than it iterates, the unrolled version + of the loop would be never entered. */ + if (est_niter >= 0 && est_niter < (HOST_WIDE_INT) upper_bound) + upper_bound = est_niter; + + if (upper_bound <= 1) return 1; - if (size_factor < upper_bound) - upper_bound = size_factor; - - max_mod_constraint = 1; + /* Choose the factor so that we may prefetch each cache just once, + but bound the unrolling by UPPER_BOUND. */ + factor = 1; for (agp = refs; agp; agp = agp->next) for (ref = agp->refs; ref; ref = ref->next) - if (should_issue_prefetch_p (ref) - && ref->prefetch_mod > max_mod_constraint) - max_mod_constraint = ref->prefetch_mod; - - /* Set constraint_factor as large as needed to be able to satisfy the - largest modulo constraint. */ - constraint_factor = max_mod_constraint; - - /* If ahead is too large in comparison with the number of available - prefetches, unroll the loop as much as needed to be able to prefetch - at least partially some of the references in the loop. */ - ahead_factor = ((ahead + SIMULTANEOUS_PREFETCHES - 1) - / SIMULTANEOUS_PREFETCHES); - - /* Unroll as much as useful, but bound the code size growth. */ - if (constraint_factor < ahead_factor) - factor = ahead_factor; - else - factor = constraint_factor; - if (factor > upper_bound) - factor = upper_bound; + if (should_issue_prefetch_p (ref)) + { + mod_constraint = ref->prefetch_mod; + nfactor = least_common_multiple (mod_constraint, factor); + if (nfactor <= upper_bound) + factor = nfactor; + } if (!should_unroll_loop_p (loop, desc, factor)) return 1; @@ -948,45 +1337,519 @@ determine_unroll_factor (struct loop *loop, struct mem_ref_group *refs, return factor; } +/* 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 + take into account reuses across the iterations of the loops in the loop + nest. */ + +static unsigned +volume_of_references (struct mem_ref_group *refs) +{ + unsigned volume = 0; + struct mem_ref_group *gr; + struct mem_ref *ref; + + for (gr = refs; gr; gr = gr->next) + for (ref = gr->refs; ref; ref = ref->next) + { + /* Almost always reuses another value? */ + if (ref->prefetch_before != PREFETCH_ALL) + continue; + + /* If several iterations access the same cache line, use the size of + the line divided by this number. Otherwise, a cache line is + accessed in each iteration. TODO -- in the latter case, we should + take the size of the reference into account, rounding it up on cache + line size multiple. */ + volume += L1_CACHE_LINE_SIZE / ref->prefetch_mod; + } + return volume; +} + +/* 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 + of the loops in the nest (length of VEC and LOOP_SIZES vectors). */ + +static unsigned +volume_of_dist_vector (lambda_vector vec, unsigned *loop_sizes, unsigned n) +{ + unsigned i; + + for (i = 0; i < n; i++) + if (vec[i] != 0) + break; + + if (i == n) + return 0; + + gcc_assert (vec[i] > 0); + + /* We ignore the parts of the distance vector in subloops, since usually + the numbers of iterations are much smaller. */ + return loop_sizes[i] * vec[i]; +} + +/* 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 + of the considered loop nest, and, LOOP is its innermost loop. */ + +static void +add_subscript_strides (tree access_fn, unsigned stride, + HOST_WIDE_INT *strides, unsigned n, struct loop *loop) +{ + struct loop *aloop; + tree step; + HOST_WIDE_INT astep; + unsigned min_depth = loop_depth (loop) - n; + + while (TREE_CODE (access_fn) == POLYNOMIAL_CHREC) + { + aloop = get_chrec_loop (access_fn); + step = CHREC_RIGHT (access_fn); + access_fn = CHREC_LEFT (access_fn); + + if ((unsigned) loop_depth (aloop) <= min_depth) + continue; + + if (host_integerp (step, 0)) + astep = tree_low_cst (step, 0); + else + astep = L1_CACHE_LINE_SIZE; + + strides[n - 1 - loop_depth (loop) + loop_depth (aloop)] += astep * stride; + + } +} + +/* Returns the volume of memory references accessed between two consecutive + 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. LOOP is the innermost loop of the current loop nest. */ + +static unsigned +self_reuse_distance (data_reference_p dr, unsigned *loop_sizes, unsigned n, + struct loop *loop) +{ + tree stride, access_fn; + HOST_WIDE_INT *strides, astride; + VEC (tree, heap) *access_fns; + tree ref = DR_REF (dr); + unsigned i, ret = ~0u; + + /* In the following example: + + for (i = 0; i < N; i++) + for (j = 0; j < N; j++) + use (a[j][i]); + 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, + 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 + the innermost loop in that the stride is less than cache size. */ + + strides = XCNEWVEC (HOST_WIDE_INT, n); + access_fns = DR_ACCESS_FNS (dr); + + FOR_EACH_VEC_ELT (tree, access_fns, i, access_fn) + { + /* Keep track of the reference corresponding to the subscript, so that we + know its stride. */ + while (handled_component_p (ref) && TREE_CODE (ref) != ARRAY_REF) + ref = TREE_OPERAND (ref, 0); + + if (TREE_CODE (ref) == ARRAY_REF) + { + stride = TYPE_SIZE_UNIT (TREE_TYPE (ref)); + if (host_integerp (stride, 1)) + astride = tree_low_cst (stride, 1); + else + astride = L1_CACHE_LINE_SIZE; + + ref = TREE_OPERAND (ref, 0); + } + else + astride = 1; + + add_subscript_strides (access_fn, astride, strides, n, loop); + } + + for (i = n; i-- > 0; ) + { + unsigned HOST_WIDE_INT s; + + s = strides[i] < 0 ? -strides[i] : strides[i]; + + if (s < (unsigned) L1_CACHE_LINE_SIZE + && (loop_sizes[i] + > (unsigned) (L1_CACHE_SIZE_BYTES / NONTEMPORAL_FRACTION))) + { + ret = loop_sizes[i]; + break; + } + } + + free (strides); + return ret; +} + +/* 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 + memory references in the loop. */ + +static void +determine_loop_nest_reuse (struct loop *loop, struct mem_ref_group *refs, + bool no_other_refs) +{ + struct loop *nest, *aloop; + VEC (data_reference_p, heap) *datarefs = NULL; + VEC (ddr_p, heap) *dependences = NULL; + struct mem_ref_group *gr; + struct mem_ref *ref, *refb; + VEC (loop_p, heap) *vloops = NULL; + unsigned *loop_data_size; + unsigned i, j, n; + unsigned volume, dist, adist; + HOST_WIDE_INT vol; + data_reference_p dr; + ddr_p dep; + + if (loop->inner) + return; + + /* Find the outermost loop of the loop nest of loop (we require that + there are no sibling loops inside the nest). */ + nest = loop; + while (1) + { + aloop = loop_outer (nest); + + if (aloop == current_loops->tree_root + || aloop->inner->next) + break; + + nest = aloop; + } + + /* For each loop, determine the amount of data accessed in each iteration. + We use this to estimate whether the reference is evicted from the + cache before its reuse. */ + find_loop_nest (nest, &vloops); + n = VEC_length (loop_p, vloops); + loop_data_size = XNEWVEC (unsigned, n); + volume = volume_of_references (refs); + i = n; + while (i-- != 0) + { + loop_data_size[i] = volume; + /* Bound the volume by the L2 cache size, since above this bound, + all dependence distances are equivalent. */ + if (volume > L2_CACHE_SIZE_BYTES) + continue; + + aloop = VEC_index (loop_p, vloops, i); + vol = max_stmt_executions_int (aloop, false); + if (vol < 0) + vol = expected_loop_iterations (aloop); + volume *= vol; + } + + /* Prepare the references in the form suitable for data dependence + analysis. We ignore unanalyzable data references (the results + are used just as a heuristics to estimate temporality of the + references, hence we do not need to worry about correctness). */ + for (gr = refs; gr; gr = gr->next) + for (ref = gr->refs; ref; ref = ref->next) + { + dr = create_data_ref (nest, loop_containing_stmt (ref->stmt), + ref->mem, ref->stmt, !ref->write_p); + + if (dr) + { + ref->reuse_distance = volume; + dr->aux = ref; + VEC_safe_push (data_reference_p, heap, datarefs, dr); + } + else + no_other_refs = false; + } + + FOR_EACH_VEC_ELT (data_reference_p, datarefs, i, dr) + { + dist = self_reuse_distance (dr, loop_data_size, n, loop); + ref = (struct mem_ref *) dr->aux; + if (ref->reuse_distance > dist) + ref->reuse_distance = dist; + + if (no_other_refs) + ref->independent_p = true; + } + + compute_all_dependences (datarefs, &dependences, vloops, true); + + FOR_EACH_VEC_ELT (ddr_p, dependences, i, dep) + { + if (DDR_ARE_DEPENDENT (dep) == chrec_known) + continue; + + ref = (struct mem_ref *) DDR_A (dep)->aux; + refb = (struct mem_ref *) DDR_B (dep)->aux; + + if (DDR_ARE_DEPENDENT (dep) == chrec_dont_know + || DDR_NUM_DIST_VECTS (dep) == 0) + { + /* If the dependence cannot be analyzed, assume that there might be + a reuse. */ + dist = 0; + + ref->independent_p = false; + refb->independent_p = false; + } + else + { + /* The distance vectors are normalized to be always lexicographically + positive, hence we cannot tell just from them whether DDR_A comes + 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 + 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 + anything). */ + + dist = volume; + for (j = 0; j < DDR_NUM_DIST_VECTS (dep); j++) + { + adist = volume_of_dist_vector (DDR_DIST_VECT (dep, j), + loop_data_size, n); + + /* If this is a dependence in the innermost loop (i.e., the + distances in all superloops are zero) and it is not + the trivial self-dependence with distance zero, record that + the references are not completely independent. */ + if (lambda_vector_zerop (DDR_DIST_VECT (dep, j), n - 1) + && (ref != refb + || DDR_DIST_VECT (dep, j)[n-1] != 0)) + { + ref->independent_p = false; + refb->independent_p = false; + } + + /* Ignore accesses closer than + L1_CACHE_SIZE_BYTES / NONTEMPORAL_FRACTION, + so that we use nontemporal prefetches e.g. if single memory + location is accessed several times in a single iteration of + the loop. */ + if (adist < L1_CACHE_SIZE_BYTES / NONTEMPORAL_FRACTION) + continue; + + if (adist < dist) + dist = adist; + } + } + + if (ref->reuse_distance > dist) + ref->reuse_distance = dist; + if (refb->reuse_distance > dist) + refb->reuse_distance = dist; + } + + free_dependence_relations (dependences); + free_data_refs (datarefs); + free (loop_data_size); + + if (dump_file && (dump_flags & TDF_DETAILS)) + { + fprintf (dump_file, "Reuse distances:\n"); + for (gr = refs; gr; gr = gr->next) + for (ref = gr->refs; ref; ref = ref->next) + fprintf (dump_file, " ref %p distance %u\n", + (void *) ref, ref->reuse_distance); + } +} + +/* Determine whether or not the trip count to ahead ratio is too small based + on prefitablility consideration. + AHEAD: the iteration ahead distance, + EST_NITER: the estimated trip count. */ + +static bool +trip_count_to_ahead_ratio_too_small_p (unsigned ahead, HOST_WIDE_INT est_niter) +{ + /* Assume trip count to ahead ratio is big enough if the trip count could not + be estimated at compile time. */ + if (est_niter < 0) + return false; + + if (est_niter < (HOST_WIDE_INT) (TRIP_COUNT_TO_AHEAD_RATIO * ahead)) + { + if (dump_file && (dump_flags & TDF_DETAILS)) + fprintf (dump_file, + "Not prefetching -- loop estimated to roll only %d times\n", + (int) est_niter); + return true; + } + + return false; +} + +/* Determine whether or not the number of memory references in the loop is + reasonable based on the profitablity and compilation time considerations. + NINSNS: estimated number of instructions in the loop, + MEM_REF_COUNT: total number of memory references in the loop. */ + +static bool +mem_ref_count_reasonable_p (unsigned ninsns, unsigned mem_ref_count) +{ + int insn_to_mem_ratio; + + if (mem_ref_count == 0) + return false; + + /* Miss rate computation (is_miss_rate_acceptable) and dependence analysis + (compute_all_dependences) have high costs based on quadratic complexity. + To avoid huge compilation time, we give up prefetching if mem_ref_count + is too large. */ + if (mem_ref_count > PREFETCH_MAX_MEM_REFS_PER_LOOP) + return false; + + /* Prefetching improves performance by overlapping cache missing + memory accesses with CPU operations. If the loop does not have + enough CPU operations to overlap with memory operations, prefetching + won't give a significant benefit. One approximate way of checking + this is to require the ratio of instructions to memory references to + be above a certain limit. This approximation works well in practice. + TODO: Implement a more precise computation by estimating the time + for each CPU or memory op in the loop. Time estimates for memory ops + should account for cache misses. */ + insn_to_mem_ratio = ninsns / mem_ref_count; + + if (insn_to_mem_ratio < PREFETCH_MIN_INSN_TO_MEM_RATIO) + { + if (dump_file && (dump_flags & TDF_DETAILS)) + fprintf (dump_file, + "Not prefetching -- instruction to memory reference ratio (%d) too small\n", + insn_to_mem_ratio); + return false; + } + + return true; +} + +/* Determine whether or not the instruction to prefetch ratio in the loop is + too small based on the profitablity consideration. + NINSNS: estimated number of instructions in the loop, + PREFETCH_COUNT: an estimate of the number of prefetches, + UNROLL_FACTOR: the factor to unroll the loop if prefetching. */ + +static bool +insn_to_prefetch_ratio_too_small_p (unsigned ninsns, unsigned prefetch_count, + unsigned unroll_factor) +{ + int insn_to_prefetch_ratio; + + /* Prefetching most likely causes performance degradation when the instruction + to prefetch ratio is too small. Too many prefetch instructions in a loop + may reduce the I-cache performance. + (unroll_factor * ninsns) is used to estimate the number of instructions in + the unrolled loop. This implementation is a bit simplistic -- the number + of issued prefetch instructions is also affected by unrolling. So, + prefetch_mod and the unroll factor should be taken into account when + determining prefetch_count. Also, the number of insns of the unrolled + loop will usually be significantly smaller than the number of insns of the + original loop * unroll_factor (at least the induction variable increases + and the exit branches will get eliminated), so it might be better to use + tree_estimate_loop_size + estimated_unrolled_size. */ + insn_to_prefetch_ratio = (unroll_factor * ninsns) / prefetch_count; + if (insn_to_prefetch_ratio < MIN_INSN_TO_PREFETCH_RATIO) + { + if (dump_file && (dump_flags & TDF_DETAILS)) + fprintf (dump_file, + "Not prefetching -- instruction to prefetch ratio (%d) too small\n", + insn_to_prefetch_ratio); + return true; + } + + return false; +} + + /* Issue prefetch instructions for array references in LOOP. Returns - true if the LOOP was unrolled. LOOPS is the array containing all - loops. */ + true if the LOOP was unrolled. */ static bool -loop_prefetch_arrays (struct loops *loops, struct loop *loop) +loop_prefetch_arrays (struct loop *loop) { struct mem_ref_group *refs; - unsigned ahead, ninsns, unroll_factor; + unsigned ahead, ninsns, time, unroll_factor; + HOST_WIDE_INT est_niter; struct tree_niter_desc desc; - bool unrolled = false; + bool unrolled = false, no_other_refs; + unsigned prefetch_count; + unsigned mem_ref_count; + + if (optimize_loop_nest_for_size_p (loop)) + { + if (dump_file && (dump_flags & TDF_DETAILS)) + fprintf (dump_file, " ignored (cold area)\n"); + return false; + } + + /* FIXME: the time should be weighted by the probabilities of the blocks in + the loop body. */ + time = tree_num_loop_insns (loop, &eni_time_weights); + if (time == 0) + return false; + + ahead = (PREFETCH_LATENCY + time - 1) / time; + est_niter = max_stmt_executions_int (loop, false); + + /* Prefetching is not likely to be profitable if the trip count to ahead + ratio is too small. */ + if (trip_count_to_ahead_ratio_too_small_p (ahead, est_niter)) + return false; + + ninsns = tree_num_loop_insns (loop, &eni_size_weights); /* Step 1: gather the memory references. */ - refs = gather_memory_references (loop); + refs = gather_memory_references (loop, &no_other_refs, &mem_ref_count); + + /* Give up prefetching if the number of memory references in the + loop is not reasonable based on profitablity and compilation time + considerations. */ + if (!mem_ref_count_reasonable_p (ninsns, mem_ref_count)) + goto fail; /* Step 2: estimate the reuse effects. */ prune_by_reuse (refs); - if (!anything_to_prefetch_p (refs)) + if (nothing_to_prefetch_p (refs)) goto fail; - /* Step 3: determine the ahead and unroll factor. */ + determine_loop_nest_reuse (loop, refs, no_other_refs); - /* FIXME: We should use not size of the loop, but the average number of - instructions executed per iteration of the loop. */ - ninsns = tree_num_loop_insns (loop); - ahead = (PREFETCH_LATENCY + ninsns - 1) / ninsns; - unroll_factor = determine_unroll_factor (loop, refs, ahead, ninsns, - &desc); - if (dump_file && (dump_flags & TDF_DETAILS)) - fprintf (dump_file, "Ahead %d, unroll factor %d\n", ahead, unroll_factor); + /* Step 3: determine unroll factor. */ + unroll_factor = determine_unroll_factor (loop, refs, ninsns, &desc, + est_niter); - /* If the loop rolls less than the required unroll factor, prefetching - is useless. */ - if (unroll_factor > 1 - && cst_and_fits_in_hwi (desc.niter) - && (unsigned HOST_WIDE_INT) int_cst_value (desc.niter) < unroll_factor) + /* Estimate prefetch count for the unrolled loop. */ + prefetch_count = estimate_prefetch_count (refs, unroll_factor); + if (prefetch_count == 0) goto fail; + if (dump_file && (dump_flags & TDF_DETAILS)) + fprintf (dump_file, "Ahead %d, unroll factor %d, trip count " + HOST_WIDE_INT_PRINT_DEC "\n" + "insn count %d, mem ref count %d, prefetch count %d\n", + ahead, unroll_factor, est_niter, + ninsns, mem_ref_count, prefetch_count); + + /* Prefetching is not likely to be profitable if the instruction to prefetch + ratio is too small. */ + if (insn_to_prefetch_ratio_too_small_p (ninsns, prefetch_count, + unroll_factor)) + goto fail; + + mark_nontemporal_stores (loop, refs); + /* Step 4: what to prefetch? */ if (!schedule_prefetches (refs, unroll_factor, ahead)) goto fail; @@ -995,7 +1858,7 @@ loop_prefetch_arrays (struct loops *loops, struct loop *loop) iterations so that we do not issue superfluous prefetches. */ if (unroll_factor != 1) { - tree_unroll_loop (loops, loop, unroll_factor, + tree_unroll_loop (loop, unroll_factor, single_dom_exit (loop), &desc); unrolled = true; } @@ -1008,12 +1871,12 @@ fail: return unrolled; } -/* Issue prefetch instructions for array references in LOOPS. */ +/* Issue prefetch instructions for array references in loops. */ unsigned int -tree_ssa_prefetch_arrays (struct loops *loops) +tree_ssa_prefetch_arrays (void) { - unsigned i; + loop_iterator li; struct loop *loop; bool unrolled = false; int todo_flags = 0; @@ -1026,35 +1889,47 @@ tree_ssa_prefetch_arrays (struct loops *loops) || PREFETCH_BLOCK == 0) return 0; + if (dump_file && (dump_flags & TDF_DETAILS)) + { + fprintf (dump_file, "Prefetching parameters:\n"); + fprintf (dump_file, " simultaneous prefetches: %d\n", + SIMULTANEOUS_PREFETCHES); + fprintf (dump_file, " prefetch latency: %d\n", PREFETCH_LATENCY); + fprintf (dump_file, " prefetch block size: %d\n", PREFETCH_BLOCK); + fprintf (dump_file, " L1 cache size: %d lines, %d kB\n", + 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, " L2 cache size: %d kB\n", L2_CACHE_SIZE); + fprintf (dump_file, " min insn-to-prefetch ratio: %d \n", + MIN_INSN_TO_PREFETCH_RATIO); + fprintf (dump_file, " min insn-to-mem ratio: %d \n", + PREFETCH_MIN_INSN_TO_MEM_RATIO); + fprintf (dump_file, "\n"); + } + initialize_original_copy_tables (); - if (!built_in_decls[BUILT_IN_PREFETCH]) + if (!builtin_decl_explicit_p (BUILT_IN_PREFETCH)) { - tree type = build_function_type (void_type_node, - tree_cons (NULL_TREE, - const_ptr_type_node, - NULL_TREE)); + tree type = build_function_type_list (void_type_node, + const_ptr_type_node, NULL_TREE); tree decl = add_builtin_function ("__builtin_prefetch", type, BUILT_IN_PREFETCH, BUILT_IN_NORMAL, NULL, NULL_TREE); DECL_IS_NOVOPS (decl) = true; - built_in_decls[BUILT_IN_PREFETCH] = decl; + set_builtin_decl (BUILT_IN_PREFETCH, decl, false); } /* We assume that size of cache line is a power of two, so verify this here. */ gcc_assert ((PREFETCH_BLOCK & (PREFETCH_BLOCK - 1)) == 0); - for (i = loops->num - 1; i > 0; i--) + FOR_EACH_LOOP (li, loop, LI_FROM_INNERMOST) { - loop = loops->parray[i]; - if (!loop) - continue; - if (dump_file && (dump_flags & TDF_DETAILS)) fprintf (dump_file, "Processing loop %d:\n", loop->num); - unrolled |= loop_prefetch_arrays (loops, loop); + unrolled |= loop_prefetch_arrays (loop); if (dump_file && (dump_flags & TDF_DETAILS)) fprintf (dump_file, "\n\n");