X-Git-Url: http://git.sourceforge.jp/view?a=blobdiff_plain;f=gcc%2Ftree-ssa-loop-prefetch.c;h=633dd337c74993683090ae2ba1869d6a0b2534b9;hb=6eb06263e2becc999cfc2fd06c29f1de14dcd3fc;hp=64f45a856ed97f74093d26ab0a4510451e3df195;hpb=f0d6e81c06c3a390e450aad66ac3083a175621f3;p=pf3gnuchains%2Fgcc-fork.git diff --git a/gcc/tree-ssa-loop-prefetch.c b/gcc/tree-ssa-loop-prefetch.c index 64f45a856ed..633dd337c74 100644 --- a/gcc/tree-ssa-loop-prefetch.c +++ b/gcc/tree-ssa-loop-prefetch.c @@ -1,42 +1,38 @@ /* Array prefetching. - Copyright (C) 2005 Free Software Foundation, Inc. - + Copyright (C) 2005, 2007, 2008, 2009, 2010 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" @@ -47,6 +43,7 @@ Software Foundation, 59 Temple Place - Suite 330, Boston, MA #include "langhooks.h" #include "tree-inline.h" #include "tree-data-ref.h" +#include "optabs.h" /* This pass inserts prefetch instructions to optimize cache usage during accesses to arrays in loops. It processes loops sequentially and: @@ -99,16 +96,33 @@ 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? + The function is_loop_prefetching_profitable() implements a cost model + to determine if prefetching is profitable for a given loop. The cost + model has two heuristcs: + 1. 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. This is implemented + by requirung the ratio between the instruction count and the mem ref + count to be above a certain minimum. + 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 + 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) @@ -156,7 +170,7 @@ 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 @@ -166,9 +180,8 @@ Software Foundation, 59 Temple Place - Suite 330, Boston, MA #define HAVE_prefetch 0 #endif -#define L1_CACHE_SIZE_BYTES ((unsigned) (L1_CACHE_SIZE * L1_CACHE_LINE_SIZE)) -/* TODO: Add parameter to specify L2 cache size. */ -#define L2_CACHE_SIZE_BYTES (8 * L1_CACHE_SIZE_BYTES) +#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 @@ -177,12 +190,31 @@ Software Foundation, 59 Temple Place - Suite 330, Boston, MA 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. */ }; @@ -191,14 +223,24 @@ 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, if prefetch_mod is 16 and unroll_factor is + 1, this means prefetching requires unrolling the loop 16 times, but + the loop is not going to be unrolled. In this case (ratio = 16), + prefetching is not likely to be beneficial. */ + +#ifndef PREFETCH_MOD_TO_UNROLL_FACTOR_RATIO +#define PREFETCH_MOD_TO_UNROLL_FACTOR_RATIO 8 +#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 @@ -208,8 +250,13 @@ struct mem_ref iterations. */ unsigned reuse_distance; /* The amount of data accessed before the first reuse of this value. */ - bool issue_prefetch_p; /* Should we really issue the prefetch? */ 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. */ @@ -222,7 +269,10 @@ 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 "); @@ -238,19 +288,20 @@ 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; } @@ -268,7 +319,7 @@ 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; @@ -302,6 +353,8 @@ record_ref (struct mem_ref_group *group, tree stmt, tree mem, (*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); @@ -332,8 +385,8 @@ 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. */ }; @@ -345,23 +398,20 @@ idx_analyze_ref (tree base, tree *index, void *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 (!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)); @@ -373,6 +423,12 @@ idx_analyze_ref (tree base, tree *index, void *data) ibase = build_int_cst (TREE_TYPE (ibase), 0); } + 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)); if (TREE_CODE (base) == ARRAY_REF) { stepsize = array_ref_element_size (base); @@ -380,11 +436,12 @@ idx_analyze_ref (tree base, tree *index, void *data) return false; imult = int_cst_value (stepsize); - istep *= imult; + *ar_data->step = fold_build2 (MULT_EXPR, sizetype, + fold_convert (sizetype, *ar_data->step), + fold_convert (sizetype, step)); idelta *= imult; } - *ar_data->step += istep; *ar_data->delta += idelta; *index = ibase; @@ -398,15 +455,15 @@ 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. */ @@ -421,7 +478,7 @@ 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; } @@ -434,37 +491,51 @@ 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; /* 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++) @@ -473,19 +544,34 @@ 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) != GIMPLE_MODIFY_STMT) - 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 = GIMPLE_STMT_OPERAND (stmt, 0); - rhs = GIMPLE_STMT_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); @@ -498,8 +584,16 @@ 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) { @@ -537,6 +631,45 @@ 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. DISTINCT_ITERS is the number of + distinct iterations after which the pattern repeats itself. + ALIGN_UNIT is the unit of alignment in bytes. */ + +static int +compute_miss_rate (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, miss_rate; + int address1, address2, cache_line1, cache_line2; + + total_positions = 0; + miss_positions = 0; + + /* 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; + total_positions += 1; + if (cache_line1 != cache_line2) + miss_positions += 1; + } + miss_rate = 1000 * miss_positions / total_positions; + return miss_rate; +} + /* Prune the prefetch candidate REF using the reuse with BY. If BY_IS_BEFORE is true, BY is before REF in the loop. */ @@ -544,12 +677,27 @@ 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; + int miss_rate; + 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) { @@ -557,7 +705,7 @@ prune_ref_by_group_reuse (struct mem_ref *ref, struct mem_ref *by, former. */ if (by_is_before) ref->prefetch_before = 0; - + return; } @@ -605,32 +753,42 @@ 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 > (unsigned) abs (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; + miss_rate = compute_miss_rate(prefetch_block, step, delta, + reduced_prefetch_block, align_unit); + if (miss_rate <= ACCEPTABLE_MISS_RATE) { + /* 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; @@ -640,8 +798,9 @@ 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)) + miss_rate = compute_miss_rate(prefetch_block, step, delta, + reduced_prefetch_block, align_unit); + if (miss_rate <= ACCEPTABLE_MISS_RATE) { if (prefetch_before < ref->prefetch_before) ref->prefetch_before = prefetch_before; @@ -744,7 +903,20 @@ 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; } @@ -786,6 +958,12 @@ schedule_prefetches (struct mem_ref_group *groups, unsigned unroll_factor, if (!should_issue_prefetch_p (ref)) continue; + /* 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 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 @@ -810,20 +988,20 @@ schedule_prefetches (struct mem_ref_group *groups, unsigned unroll_factor, return any; } -/* Determine whether there is any reference suitable for prefetching - in GROUPS. */ +/* Estimate the number of prefetches in the given GROUPS. */ -static bool -anything_to_prefetch_p (struct mem_ref_group *groups) +static int +estimate_prefetch_count (struct mem_ref_group *groups) { struct mem_ref *ref; + int prefetch_count = 0; for (; groups; groups = groups->next) for (ref = groups->refs; ref; ref = ref->next) if (should_issue_prefetch_p (ref)) - return true; + prefetch_count++; - return false; + return prefetch_count; } /* Issue prefetches for the reference REF into loop as decided before. @@ -834,8 +1012,9 @@ static void issue_prefetch_ref (struct mem_ref *ref, unsigned unroll_factor, unsigned ahead) { HOST_WIDE_INT delta; - tree addr, addr_base, prefetch, write_p, local; - 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; @@ -844,27 +1023,44 @@ issue_prefetch_ref (struct mem_ref *ref, unsigned unroll_factor, unsigned ahead) 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 = build_int_cst (integer_type_node, nontemporal ? 0 : 3); 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_build2 (POINTER_PLUS_EXPR, ptr_type_node, + addr_base, size_int (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_build2 (POINTER_PLUS_EXPR, ptr_type_node, addr_base, + forward); + addr = force_gimple_operand_gsi (&bsi, unshare_expr (addr), true, + NULL, true, GSI_SAME_STMT); + } /* Create the prefetch instruction. */ - prefetch = build_call_expr (built_in_decls[BUILT_IN_PREFETCH], - 3, addr, write_p, local); - bsi_insert_before (&bsi, prefetch, BSI_SAME_STMT); + prefetch = gimple_build_call (built_in_decls[BUILT_IN_PREFETCH], + 3, addr, write_p, local); + gsi_insert_before (&bsi, prefetch, GSI_SAME_STMT); } } @@ -884,6 +1080,130 @@ 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)->insn_code; + 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 (i = 0; VEC_iterate (edge, exits, i, exit); i++) + { + 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 (i = 0; VEC_iterate (edge, exits, i, exit); i++) + 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. */ @@ -1078,7 +1398,7 @@ self_reuse_distance (data_reference_p dr, unsigned *loop_sizes, unsigned n, 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)); @@ -1115,16 +1435,18 @@ self_reuse_distance (data_reference_p dr, unsigned *loop_sizes, unsigned n, } /* Determines the distance till the first reuse of each reference in REFS - in the loop nest of LOOP. */ + 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) +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; + struct mem_ref *ref, *refb; VEC (loop_p, heap) *vloops = NULL; unsigned *loop_data_size; unsigned i, j, n; @@ -1174,7 +1496,7 @@ determine_loop_nest_reuse (struct loop *loop, struct mem_ref_group *refs) } /* Prepare the references in the form suitable for data dependence - analysis. We ignore unanalysable data references (the results + 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) @@ -1188,14 +1510,19 @@ determine_loop_nest_reuse (struct loop *loop, struct mem_ref_group *refs) dr->aux = ref; VEC_safe_push (data_reference_p, heap, datarefs, dr); } + else + no_other_refs = false; } for (i = 0; VEC_iterate (data_reference_p, datarefs, i, dr); i++) { dist = self_reuse_distance (dr, loop_data_size, n, loop); - ref = dr->aux; + 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); @@ -1205,16 +1532,22 @@ determine_loop_nest_reuse (struct loop *loop, struct mem_ref_group *refs) 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 analysed, assume that there might be + /* 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 normalised 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 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 @@ -1228,6 +1561,18 @@ determine_loop_nest_reuse (struct loop *loop, struct mem_ref_group *refs) 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 @@ -1241,12 +1586,10 @@ determine_loop_nest_reuse (struct loop *loop, struct mem_ref_group *refs) } } - ref = DDR_A (dep)->aux; - if (ref->reuse_distance > dist) - ref->reuse_distance = dist; - ref = DDR_B (dep)->aux; if (ref->reuse_distance > dist) ref->reuse_distance = dist; + if (refb->reuse_distance > dist) + refb->reuse_distance = dist; } free_dependence_relations (dependences); @@ -1263,6 +1606,83 @@ determine_loop_nest_reuse (struct loop *loop, struct mem_ref_group *refs) } } +/* Do a cost-benefit analysis to determine if prefetching is profitable + for the current loop given the following parameters: + AHEAD: the iteration ahead distance, + EST_NITER: the estimated trip count, + NINSNS: estimated number of instructions in the loop, + PREFETCH_COUNT: an estimate of the number of prefetches + MEM_REF_COUNT: total number of memory references in the loop. */ + +static bool +is_loop_prefetching_profitable (unsigned ahead, HOST_WIDE_INT est_niter, + unsigned ninsns, unsigned prefetch_count, + unsigned mem_ref_count, unsigned unroll_factor) +{ + int insn_to_mem_ratio, insn_to_prefetch_ratio; + + if (mem_ref_count == 0) + 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; + } + + /* 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 false; + } + + /* Could not do further estimation if the trip count is unknown. Just assume + prefetching is profitable. Too aggressive??? */ + if (est_niter < 0) + return true; + + 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 false; + } + return true; +} + + /* Issue prefetch instructions for array references in LOOP. Returns true if the LOOP was unrolled. */ @@ -1273,9 +1693,11 @@ loop_prefetch_arrays (struct loop *loop) 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 (!maybe_hot_bb_p (loop->header)) + if (optimize_loop_nest_for_size_p (loop)) { if (dump_file && (dump_flags & TDF_DETAILS)) fprintf (dump_file, " ignored (cold area)\n"); @@ -1283,15 +1705,16 @@ loop_prefetch_arrays (struct loop *loop) } /* Step 1: gather the memory references. */ - refs = gather_memory_references (loop); + refs = gather_memory_references (loop, &no_other_refs, &mem_ref_count); /* Step 2: estimate the reuse effects. */ prune_by_reuse (refs); - if (!anything_to_prefetch_p (refs)) + prefetch_count = estimate_prefetch_count (refs); + if (prefetch_count == 0) goto fail; - determine_loop_nest_reuse (loop, refs); + determine_loop_nest_reuse (loop, refs, no_other_refs); /* Step 3: determine the ahead and unroll factor. */ @@ -1301,23 +1724,21 @@ loop_prefetch_arrays (struct loop *loop) ahead = (PREFETCH_LATENCY + time - 1) / time; est_niter = estimated_loop_iterations_int (loop, false); - /* The prefetches will run for AHEAD iterations of the original loop. Unless - the loop rolls at least AHEAD times, prefetching the references does not - make sense. */ - if (est_niter >= 0 && est_niter <= (HOST_WIDE_INT) ahead) - { - if (dump_file && (dump_flags & TDF_DETAILS)) - fprintf (dump_file, - "Not prefetching -- loop estimated to roll only %d times\n", - (int) est_niter); - goto fail; - } - ninsns = tree_num_loop_insns (loop, &eni_size_weights); unroll_factor = determine_unroll_factor (loop, refs, ninsns, &desc, est_niter); if (dump_file && (dump_flags & TDF_DETAILS)) - fprintf (dump_file, "Ahead %d, unroll factor %d\n", ahead, unroll_factor); + 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); + + if (!is_loop_prefetching_profitable (ahead, est_niter, ninsns, prefetch_count, + mem_ref_count, unroll_factor)) + goto fail; + + mark_nontemporal_stores (loop, refs); /* Step 4: what to prefetch? */ if (!schedule_prefetches (refs, unroll_factor, ahead)) @@ -1365,10 +1786,14 @@ tree_ssa_prefetch_arrays (void) 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 bytes\n", - L1_CACHE_SIZE, L1_CACHE_SIZE_BYTES); + 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 bytes\n", L2_CACHE_SIZE_BYTES); + 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"); }