2 Copyright (C) 2005, 2007, 2008, 2009, 2010 Free Software Foundation, Inc.
4 This file is part of GCC.
6 GCC is free software; you can redistribute it and/or modify it
7 under the terms of the GNU General Public License as published by the
8 Free Software Foundation; either version 3, or (at your option) any
11 GCC is distributed in the hope that it will be useful, but WITHOUT
12 ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
13 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
16 You should have received a copy of the GNU General Public License
17 along with GCC; see the file COPYING3. If not see
18 <http://www.gnu.org/licenses/>. */
22 #include "coretypes.h"
26 #include "basic-block.h"
28 #include "tree-pretty-print.h"
29 #include "tree-flow.h"
30 #include "tree-dump.h"
33 #include "tree-pass.h"
34 #include "insn-config.h"
37 #include "tree-chrec.h"
38 #include "tree-scalar-evolution.h"
39 #include "diagnostic-core.h"
42 #include "langhooks.h"
43 #include "tree-inline.h"
44 #include "tree-data-ref.h"
47 /* FIXME: Needed for optabs, but this should all be moved to a TBD interface
48 between the GIMPLE and RTL worlds. */
52 /* This pass inserts prefetch instructions to optimize cache usage during
53 accesses to arrays in loops. It processes loops sequentially and:
55 1) Gathers all memory references in the single loop.
56 2) For each of the references it decides when it is profitable to prefetch
57 it. To do it, we evaluate the reuse among the accesses, and determines
58 two values: PREFETCH_BEFORE (meaning that it only makes sense to do
59 prefetching in the first PREFETCH_BEFORE iterations of the loop) and
60 PREFETCH_MOD (meaning that it only makes sense to prefetch in the
61 iterations of the loop that are zero modulo PREFETCH_MOD). For example
62 (assuming cache line size is 64 bytes, char has size 1 byte and there
63 is no hardware sequential prefetch):
66 for (i = 0; i < max; i++)
73 a[187*i + 50] = ...; (5)
76 (0) obviously has PREFETCH_BEFORE 1
77 (1) has PREFETCH_BEFORE 64, since (2) accesses the same memory
78 location 64 iterations before it, and PREFETCH_MOD 64 (since
79 it hits the same cache line otherwise).
80 (2) has PREFETCH_MOD 64
81 (3) has PREFETCH_MOD 4
82 (4) has PREFETCH_MOD 1. We do not set PREFETCH_BEFORE here, since
83 the cache line accessed by (4) is the same with probability only
85 (5) has PREFETCH_MOD 1 as well.
87 Additionally, we use data dependence analysis to determine for each
88 reference the distance till the first reuse; this information is used
89 to determine the temporality of the issued prefetch instruction.
91 3) We determine how much ahead we need to prefetch. The number of
92 iterations needed is time to fetch / time spent in one iteration of
93 the loop. The problem is that we do not know either of these values,
94 so we just make a heuristic guess based on a magic (possibly)
95 target-specific constant and size of the loop.
97 4) Determine which of the references we prefetch. We take into account
98 that there is a maximum number of simultaneous prefetches (provided
99 by machine description). We prefetch as many prefetches as possible
100 while still within this bound (starting with those with lowest
101 prefetch_mod, since they are responsible for most of the cache
104 5) We unroll and peel loops so that we are able to satisfy PREFETCH_MOD
105 and PREFETCH_BEFORE requirements (within some bounds), and to avoid
106 prefetching nonaccessed memory.
107 TODO -- actually implement peeling.
109 6) We actually emit the prefetch instructions. ??? Perhaps emit the
110 prefetch instructions with guards in cases where 5) was not sufficient
111 to satisfy the constraints?
113 A cost model is implemented to determine whether or not prefetching is
114 profitable for a given loop. The cost model has three heuristics:
116 1. Function trip_count_to_ahead_ratio_too_small_p implements a
117 heuristic that determines whether or not the loop has too few
118 iterations (compared to ahead). Prefetching is not likely to be
119 beneficial if the trip count to ahead ratio is below a certain
122 2. Function mem_ref_count_reasonable_p implements a heuristic that
123 determines whether the given loop has enough CPU ops that can be
124 overlapped with cache missing memory ops. If not, the loop
125 won't benefit from prefetching. In the implementation,
126 prefetching is not considered beneficial if the ratio between
127 the instruction count and the mem ref count is below a certain
130 3. Function insn_to_prefetch_ratio_too_small_p implements a
131 heuristic that disables prefetching in a loop if the prefetching
132 cost is above a certain limit. The relative prefetching cost is
133 estimated by taking the ratio between the prefetch count and the
134 total intruction count (this models the I-cache cost).
136 The limits used in these heuristics are defined as parameters with
137 reasonable default values. Machine-specific default values will be
141 -- write and use more general reuse analysis (that could be also used
142 in other cache aimed loop optimizations)
143 -- make it behave sanely together with the prefetches given by user
144 (now we just ignore them; at the very least we should avoid
145 optimizing loops in that user put his own prefetches)
146 -- we assume cache line size alignment of arrays; this could be
149 /* Magic constants follow. These should be replaced by machine specific
152 /* True if write can be prefetched by a read prefetch. */
154 #ifndef WRITE_CAN_USE_READ_PREFETCH
155 #define WRITE_CAN_USE_READ_PREFETCH 1
158 /* True if read can be prefetched by a write prefetch. */
160 #ifndef READ_CAN_USE_WRITE_PREFETCH
161 #define READ_CAN_USE_WRITE_PREFETCH 0
164 /* The size of the block loaded by a single prefetch. Usually, this is
165 the same as cache line size (at the moment, we only consider one level
166 of cache hierarchy). */
168 #ifndef PREFETCH_BLOCK
169 #define PREFETCH_BLOCK L1_CACHE_LINE_SIZE
172 /* Do we have a forward hardware sequential prefetching? */
174 #ifndef HAVE_FORWARD_PREFETCH
175 #define HAVE_FORWARD_PREFETCH 0
178 /* Do we have a backward hardware sequential prefetching? */
180 #ifndef HAVE_BACKWARD_PREFETCH
181 #define HAVE_BACKWARD_PREFETCH 0
184 /* In some cases we are only able to determine that there is a certain
185 probability that the two accesses hit the same cache line. In this
186 case, we issue the prefetches for both of them if this probability
187 is less then (1000 - ACCEPTABLE_MISS_RATE) per thousand. */
189 #ifndef ACCEPTABLE_MISS_RATE
190 #define ACCEPTABLE_MISS_RATE 50
193 #ifndef HAVE_prefetch
194 #define HAVE_prefetch 0
197 #define L1_CACHE_SIZE_BYTES ((unsigned) (L1_CACHE_SIZE * 1024))
198 #define L2_CACHE_SIZE_BYTES ((unsigned) (L2_CACHE_SIZE * 1024))
200 /* We consider a memory access nontemporal if it is not reused sooner than
201 after L2_CACHE_SIZE_BYTES of memory are accessed. However, we ignore
202 accesses closer than L1_CACHE_SIZE_BYTES / NONTEMPORAL_FRACTION,
203 so that we use nontemporal prefetches e.g. if single memory location
204 is accessed several times in a single iteration of the loop. */
205 #define NONTEMPORAL_FRACTION 16
207 /* In case we have to emit a memory fence instruction after the loop that
208 uses nontemporal stores, this defines the builtin to use. */
210 #ifndef FENCE_FOLLOWING_MOVNT
211 #define FENCE_FOLLOWING_MOVNT NULL_TREE
214 /* It is not profitable to prefetch when the trip count is not at
215 least TRIP_COUNT_TO_AHEAD_RATIO times the prefetch ahead distance.
216 For example, in a loop with a prefetch ahead distance of 10,
217 supposing that TRIP_COUNT_TO_AHEAD_RATIO is equal to 4, it is
218 profitable to prefetch when the trip count is greater or equal to
219 40. In that case, 30 out of the 40 iterations will benefit from
222 #ifndef TRIP_COUNT_TO_AHEAD_RATIO
223 #define TRIP_COUNT_TO_AHEAD_RATIO 4
226 /* The group of references between that reuse may occur. */
230 tree base; /* Base of the reference. */
231 tree step; /* Step of the reference. */
232 struct mem_ref *refs; /* References in the group. */
233 struct mem_ref_group *next; /* Next group of references. */
236 /* Assigned to PREFETCH_BEFORE when all iterations are to be prefetched. */
238 #define PREFETCH_ALL (~(unsigned HOST_WIDE_INT) 0)
240 /* Do not generate a prefetch if the unroll factor is significantly less
241 than what is required by the prefetch. This is to avoid redundant
242 prefetches. For example, when prefetch_mod is 16 and unroll_factor is
243 2, prefetching requires unrolling the loop 16 times, but
244 the loop is actually unrolled twice. In this case (ratio = 8),
245 prefetching is not likely to be beneficial. */
247 #ifndef PREFETCH_MOD_TO_UNROLL_FACTOR_RATIO
248 #define PREFETCH_MOD_TO_UNROLL_FACTOR_RATIO 4
251 /* Some of the prefetch computations have quadratic complexity. We want to
252 avoid huge compile times and, therefore, want to limit the amount of
253 memory references per loop where we consider prefetching. */
255 #ifndef PREFETCH_MAX_MEM_REFS_PER_LOOP
256 #define PREFETCH_MAX_MEM_REFS_PER_LOOP 200
259 /* The memory reference. */
263 gimple stmt; /* Statement in that the reference appears. */
264 tree mem; /* The reference. */
265 HOST_WIDE_INT delta; /* Constant offset of the reference. */
266 struct mem_ref_group *group; /* The group of references it belongs to. */
267 unsigned HOST_WIDE_INT prefetch_mod;
268 /* Prefetch only each PREFETCH_MOD-th
270 unsigned HOST_WIDE_INT prefetch_before;
271 /* Prefetch only first PREFETCH_BEFORE
273 unsigned reuse_distance; /* The amount of data accessed before the first
274 reuse of this value. */
275 struct mem_ref *next; /* The next reference in the group. */
276 unsigned write_p : 1; /* Is it a write? */
277 unsigned independent_p : 1; /* True if the reference is independent on
278 all other references inside the loop. */
279 unsigned issue_prefetch_p : 1; /* Should we really issue the prefetch? */
280 unsigned storent_p : 1; /* True if we changed the store to a
284 /* Dumps information about reference REF to FILE. */
287 dump_mem_ref (FILE *file, struct mem_ref *ref)
289 fprintf (file, "Reference %p:\n", (void *) ref);
291 fprintf (file, " group %p (base ", (void *) ref->group);
292 print_generic_expr (file, ref->group->base, TDF_SLIM);
293 fprintf (file, ", step ");
294 if (cst_and_fits_in_hwi (ref->group->step))
295 fprintf (file, HOST_WIDE_INT_PRINT_DEC, int_cst_value (ref->group->step));
297 print_generic_expr (file, ref->group->step, TDF_TREE);
298 fprintf (file, ")\n");
300 fprintf (file, " delta ");
301 fprintf (file, HOST_WIDE_INT_PRINT_DEC, ref->delta);
302 fprintf (file, "\n");
304 fprintf (file, " %s\n", ref->write_p ? "write" : "read");
306 fprintf (file, "\n");
309 /* Finds a group with BASE and STEP in GROUPS, or creates one if it does not
312 static struct mem_ref_group *
313 find_or_create_group (struct mem_ref_group **groups, tree base, tree step)
315 struct mem_ref_group *group;
317 for (; *groups; groups = &(*groups)->next)
319 if (operand_equal_p ((*groups)->step, step, 0)
320 && operand_equal_p ((*groups)->base, base, 0))
323 /* If step is an integer constant, keep the list of groups sorted
324 by decreasing step. */
325 if (cst_and_fits_in_hwi ((*groups)->step) && cst_and_fits_in_hwi (step)
326 && int_cst_value ((*groups)->step) < int_cst_value (step))
330 group = XNEW (struct mem_ref_group);
334 group->next = *groups;
340 /* Records a memory reference MEM in GROUP with offset DELTA and write status
341 WRITE_P. The reference occurs in statement STMT. */
344 record_ref (struct mem_ref_group *group, gimple stmt, tree mem,
345 HOST_WIDE_INT delta, bool write_p)
347 struct mem_ref **aref;
349 /* Do not record the same address twice. */
350 for (aref = &group->refs; *aref; aref = &(*aref)->next)
352 /* It does not have to be possible for write reference to reuse the read
353 prefetch, or vice versa. */
354 if (!WRITE_CAN_USE_READ_PREFETCH
356 && !(*aref)->write_p)
358 if (!READ_CAN_USE_WRITE_PREFETCH
363 if ((*aref)->delta == delta)
367 (*aref) = XNEW (struct mem_ref);
368 (*aref)->stmt = stmt;
370 (*aref)->delta = delta;
371 (*aref)->write_p = write_p;
372 (*aref)->prefetch_before = PREFETCH_ALL;
373 (*aref)->prefetch_mod = 1;
374 (*aref)->reuse_distance = 0;
375 (*aref)->issue_prefetch_p = false;
376 (*aref)->group = group;
377 (*aref)->next = NULL;
378 (*aref)->independent_p = false;
379 (*aref)->storent_p = false;
381 if (dump_file && (dump_flags & TDF_DETAILS))
382 dump_mem_ref (dump_file, *aref);
385 /* Release memory references in GROUPS. */
388 release_mem_refs (struct mem_ref_group *groups)
390 struct mem_ref_group *next_g;
391 struct mem_ref *ref, *next_r;
393 for (; groups; groups = next_g)
395 next_g = groups->next;
396 for (ref = groups->refs; ref; ref = next_r)
405 /* A structure used to pass arguments to idx_analyze_ref. */
409 struct loop *loop; /* Loop of the reference. */
410 gimple stmt; /* Statement of the reference. */
411 tree *step; /* Step of the memory reference. */
412 HOST_WIDE_INT *delta; /* Offset of the memory reference. */
415 /* Analyzes a single INDEX of a memory reference to obtain information
416 described at analyze_ref. Callback for for_each_index. */
419 idx_analyze_ref (tree base, tree *index, void *data)
421 struct ar_data *ar_data = (struct ar_data *) data;
422 tree ibase, step, stepsize;
423 HOST_WIDE_INT idelta = 0, imult = 1;
426 if (TREE_CODE (base) == MISALIGNED_INDIRECT_REF)
429 if (!simple_iv (ar_data->loop, loop_containing_stmt (ar_data->stmt),
435 if (TREE_CODE (ibase) == POINTER_PLUS_EXPR
436 && cst_and_fits_in_hwi (TREE_OPERAND (ibase, 1)))
438 idelta = int_cst_value (TREE_OPERAND (ibase, 1));
439 ibase = TREE_OPERAND (ibase, 0);
441 if (cst_and_fits_in_hwi (ibase))
443 idelta += int_cst_value (ibase);
444 ibase = build_int_cst (TREE_TYPE (ibase), 0);
447 if (TREE_CODE (base) == ARRAY_REF)
449 stepsize = array_ref_element_size (base);
450 if (!cst_and_fits_in_hwi (stepsize))
452 imult = int_cst_value (stepsize);
453 step = fold_build2 (MULT_EXPR, sizetype,
454 fold_convert (sizetype, step),
455 fold_convert (sizetype, stepsize));
459 if (*ar_data->step == NULL_TREE)
460 *ar_data->step = step;
462 *ar_data->step = fold_build2 (PLUS_EXPR, sizetype,
463 fold_convert (sizetype, *ar_data->step),
464 fold_convert (sizetype, step));
465 *ar_data->delta += idelta;
471 /* Tries to express REF_P in shape &BASE + STEP * iter + DELTA, where DELTA and
472 STEP are integer constants and iter is number of iterations of LOOP. The
473 reference occurs in statement STMT. Strips nonaddressable component
474 references from REF_P. */
477 analyze_ref (struct loop *loop, tree *ref_p, tree *base,
478 tree *step, HOST_WIDE_INT *delta,
481 struct ar_data ar_data;
483 HOST_WIDE_INT bit_offset;
489 /* First strip off the component references. Ignore bitfields. */
490 if (TREE_CODE (ref) == COMPONENT_REF
491 && DECL_NONADDRESSABLE_P (TREE_OPERAND (ref, 1)))
492 ref = TREE_OPERAND (ref, 0);
496 for (; TREE_CODE (ref) == COMPONENT_REF; ref = TREE_OPERAND (ref, 0))
498 off = DECL_FIELD_BIT_OFFSET (TREE_OPERAND (ref, 1));
499 bit_offset = TREE_INT_CST_LOW (off);
500 gcc_assert (bit_offset % BITS_PER_UNIT == 0);
502 *delta += bit_offset / BITS_PER_UNIT;
505 *base = unshare_expr (ref);
509 ar_data.delta = delta;
510 return for_each_index (base, idx_analyze_ref, &ar_data);
513 /* Record a memory reference REF to the list REFS. The reference occurs in
514 LOOP in statement STMT and it is write if WRITE_P. Returns true if the
515 reference was recorded, false otherwise. */
518 gather_memory_references_ref (struct loop *loop, struct mem_ref_group **refs,
519 tree ref, bool write_p, gimple stmt)
523 struct mem_ref_group *agrp;
525 if (get_base_address (ref) == NULL)
528 if (!analyze_ref (loop, &ref, &base, &step, &delta, stmt))
530 /* If analyze_ref fails the default is a NULL_TREE. We can stop here. */
531 if (step == NULL_TREE)
534 /* Limit non-constant step prefetching only to the innermost loops. */
535 if (!cst_and_fits_in_hwi (step) && loop->inner != NULL)
538 /* Now we know that REF = &BASE + STEP * iter + DELTA, where DELTA and STEP
539 are integer constants. */
540 agrp = find_or_create_group (refs, base, step);
541 record_ref (agrp, stmt, ref, delta, write_p);
546 /* Record the suitable memory references in LOOP. NO_OTHER_REFS is set to
547 true if there are no other memory references inside the loop. */
549 static struct mem_ref_group *
550 gather_memory_references (struct loop *loop, bool *no_other_refs, unsigned *ref_count)
552 basic_block *body = get_loop_body_in_dom_order (loop);
555 gimple_stmt_iterator bsi;
558 struct mem_ref_group *refs = NULL;
560 *no_other_refs = true;
563 /* Scan the loop body in order, so that the former references precede the
565 for (i = 0; i < loop->num_nodes; i++)
568 if (bb->loop_father != loop)
571 for (bsi = gsi_start_bb (bb); !gsi_end_p (bsi); gsi_next (&bsi))
573 stmt = gsi_stmt (bsi);
575 if (gimple_code (stmt) != GIMPLE_ASSIGN)
577 if (gimple_vuse (stmt)
578 || (is_gimple_call (stmt)
579 && !(gimple_call_flags (stmt) & ECF_CONST)))
580 *no_other_refs = false;
584 lhs = gimple_assign_lhs (stmt);
585 rhs = gimple_assign_rhs1 (stmt);
587 if (REFERENCE_CLASS_P (rhs))
589 *no_other_refs &= gather_memory_references_ref (loop, &refs,
593 if (REFERENCE_CLASS_P (lhs))
595 *no_other_refs &= gather_memory_references_ref (loop, &refs,
606 /* Prune the prefetch candidate REF using the self-reuse. */
609 prune_ref_by_self_reuse (struct mem_ref *ref)
614 /* If the step size is non constant, we cannot calculate prefetch_mod. */
615 if (!cst_and_fits_in_hwi (ref->group->step))
618 step = int_cst_value (ref->group->step);
624 /* Prefetch references to invariant address just once. */
625 ref->prefetch_before = 1;
632 if (step > PREFETCH_BLOCK)
635 if ((backward && HAVE_BACKWARD_PREFETCH)
636 || (!backward && HAVE_FORWARD_PREFETCH))
638 ref->prefetch_before = 1;
642 ref->prefetch_mod = PREFETCH_BLOCK / step;
645 /* Divides X by BY, rounding down. */
648 ddown (HOST_WIDE_INT x, unsigned HOST_WIDE_INT by)
655 return (x + by - 1) / by;
658 /* Given a CACHE_LINE_SIZE and two inductive memory references
659 with a common STEP greater than CACHE_LINE_SIZE and an address
660 difference DELTA, compute the probability that they will fall
661 in different cache lines. Return true if the computed miss rate
662 is not greater than the ACCEPTABLE_MISS_RATE. DISTINCT_ITERS is the
663 number of distinct iterations after which the pattern repeats itself.
664 ALIGN_UNIT is the unit of alignment in bytes. */
667 is_miss_rate_acceptable (unsigned HOST_WIDE_INT cache_line_size,
668 HOST_WIDE_INT step, HOST_WIDE_INT delta,
669 unsigned HOST_WIDE_INT distinct_iters,
672 unsigned align, iter;
673 int total_positions, miss_positions, max_allowed_miss_positions;
674 int address1, address2, cache_line1, cache_line2;
676 /* It always misses if delta is greater than or equal to the cache
678 if (delta >= (HOST_WIDE_INT) cache_line_size)
682 total_positions = (cache_line_size / align_unit) * distinct_iters;
683 max_allowed_miss_positions = (ACCEPTABLE_MISS_RATE * total_positions) / 1000;
685 /* Iterate through all possible alignments of the first
686 memory reference within its cache line. */
687 for (align = 0; align < cache_line_size; align += align_unit)
689 /* Iterate through all distinct iterations. */
690 for (iter = 0; iter < distinct_iters; iter++)
692 address1 = align + step * iter;
693 address2 = address1 + delta;
694 cache_line1 = address1 / cache_line_size;
695 cache_line2 = address2 / cache_line_size;
696 if (cache_line1 != cache_line2)
699 if (miss_positions > max_allowed_miss_positions)
706 /* Prune the prefetch candidate REF using the reuse with BY.
707 If BY_IS_BEFORE is true, BY is before REF in the loop. */
710 prune_ref_by_group_reuse (struct mem_ref *ref, struct mem_ref *by,
715 HOST_WIDE_INT delta_r = ref->delta, delta_b = by->delta;
716 HOST_WIDE_INT delta = delta_b - delta_r;
717 HOST_WIDE_INT hit_from;
718 unsigned HOST_WIDE_INT prefetch_before, prefetch_block;
719 HOST_WIDE_INT reduced_step;
720 unsigned HOST_WIDE_INT reduced_prefetch_block;
724 /* If the step is non constant we cannot calculate prefetch_before. */
725 if (!cst_and_fits_in_hwi (ref->group->step)) {
729 step = int_cst_value (ref->group->step);
736 /* If the references has the same address, only prefetch the
739 ref->prefetch_before = 0;
746 /* If the reference addresses are invariant and fall into the
747 same cache line, prefetch just the first one. */
751 if (ddown (ref->delta, PREFETCH_BLOCK)
752 != ddown (by->delta, PREFETCH_BLOCK))
755 ref->prefetch_before = 0;
759 /* Only prune the reference that is behind in the array. */
765 /* Transform the data so that we may assume that the accesses
769 delta_r = PREFETCH_BLOCK - 1 - delta_r;
770 delta_b = PREFETCH_BLOCK - 1 - delta_b;
778 /* Check whether the two references are likely to hit the same cache
779 line, and how distant the iterations in that it occurs are from
782 if (step <= PREFETCH_BLOCK)
784 /* The accesses are sure to meet. Let us check when. */
785 hit_from = ddown (delta_b, PREFETCH_BLOCK) * PREFETCH_BLOCK;
786 prefetch_before = (hit_from - delta_r + step - 1) / step;
788 /* Do not reduce prefetch_before if we meet beyond cache size. */
789 if (prefetch_before > (unsigned) abs (L2_CACHE_SIZE_BYTES / step))
790 prefetch_before = PREFETCH_ALL;
791 if (prefetch_before < ref->prefetch_before)
792 ref->prefetch_before = prefetch_before;
797 /* A more complicated case with step > prefetch_block. First reduce
798 the ratio between the step and the cache line size to its simplest
799 terms. The resulting denominator will then represent the number of
800 distinct iterations after which each address will go back to its
801 initial location within the cache line. This computation assumes
802 that PREFETCH_BLOCK is a power of two. */
803 prefetch_block = PREFETCH_BLOCK;
804 reduced_prefetch_block = prefetch_block;
806 while ((reduced_step & 1) == 0
807 && reduced_prefetch_block > 1)
810 reduced_prefetch_block >>= 1;
813 prefetch_before = delta / step;
815 ref_type = TREE_TYPE (ref->mem);
816 align_unit = TYPE_ALIGN (ref_type) / 8;
817 if (is_miss_rate_acceptable (prefetch_block, step, delta,
818 reduced_prefetch_block, align_unit))
820 /* Do not reduce prefetch_before if we meet beyond cache size. */
821 if (prefetch_before > L2_CACHE_SIZE_BYTES / PREFETCH_BLOCK)
822 prefetch_before = PREFETCH_ALL;
823 if (prefetch_before < ref->prefetch_before)
824 ref->prefetch_before = prefetch_before;
829 /* Try also the following iteration. */
831 delta = step - delta;
832 if (is_miss_rate_acceptable (prefetch_block, step, delta,
833 reduced_prefetch_block, align_unit))
835 if (prefetch_before < ref->prefetch_before)
836 ref->prefetch_before = prefetch_before;
841 /* The ref probably does not reuse by. */
845 /* Prune the prefetch candidate REF using the reuses with other references
849 prune_ref_by_reuse (struct mem_ref *ref, struct mem_ref *refs)
851 struct mem_ref *prune_by;
854 prune_ref_by_self_reuse (ref);
856 for (prune_by = refs; prune_by; prune_by = prune_by->next)
864 if (!WRITE_CAN_USE_READ_PREFETCH
866 && !prune_by->write_p)
868 if (!READ_CAN_USE_WRITE_PREFETCH
870 && prune_by->write_p)
873 prune_ref_by_group_reuse (ref, prune_by, before);
877 /* Prune the prefetch candidates in GROUP using the reuse analysis. */
880 prune_group_by_reuse (struct mem_ref_group *group)
882 struct mem_ref *ref_pruned;
884 for (ref_pruned = group->refs; ref_pruned; ref_pruned = ref_pruned->next)
886 prune_ref_by_reuse (ref_pruned, group->refs);
888 if (dump_file && (dump_flags & TDF_DETAILS))
890 fprintf (dump_file, "Reference %p:", (void *) ref_pruned);
892 if (ref_pruned->prefetch_before == PREFETCH_ALL
893 && ref_pruned->prefetch_mod == 1)
894 fprintf (dump_file, " no restrictions");
895 else if (ref_pruned->prefetch_before == 0)
896 fprintf (dump_file, " do not prefetch");
897 else if (ref_pruned->prefetch_before <= ref_pruned->prefetch_mod)
898 fprintf (dump_file, " prefetch once");
901 if (ref_pruned->prefetch_before != PREFETCH_ALL)
903 fprintf (dump_file, " prefetch before ");
904 fprintf (dump_file, HOST_WIDE_INT_PRINT_DEC,
905 ref_pruned->prefetch_before);
907 if (ref_pruned->prefetch_mod != 1)
909 fprintf (dump_file, " prefetch mod ");
910 fprintf (dump_file, HOST_WIDE_INT_PRINT_DEC,
911 ref_pruned->prefetch_mod);
914 fprintf (dump_file, "\n");
919 /* Prune the list of prefetch candidates GROUPS using the reuse analysis. */
922 prune_by_reuse (struct mem_ref_group *groups)
924 for (; groups; groups = groups->next)
925 prune_group_by_reuse (groups);
928 /* Returns true if we should issue prefetch for REF. */
931 should_issue_prefetch_p (struct mem_ref *ref)
933 /* For now do not issue prefetches for only first few of the
935 if (ref->prefetch_before != PREFETCH_ALL)
937 if (dump_file && (dump_flags & TDF_DETAILS))
938 fprintf (dump_file, "Ignoring %p due to prefetch_before\n",
943 /* Do not prefetch nontemporal stores. */
946 if (dump_file && (dump_flags & TDF_DETAILS))
947 fprintf (dump_file, "Ignoring nontemporal store %p\n", (void *) ref);
954 /* Decide which of the prefetch candidates in GROUPS to prefetch.
955 AHEAD is the number of iterations to prefetch ahead (which corresponds
956 to the number of simultaneous instances of one prefetch running at a
957 time). UNROLL_FACTOR is the factor by that the loop is going to be
958 unrolled. Returns true if there is anything to prefetch. */
961 schedule_prefetches (struct mem_ref_group *groups, unsigned unroll_factor,
964 unsigned remaining_prefetch_slots, n_prefetches, prefetch_slots;
965 unsigned slots_per_prefetch;
969 /* At most SIMULTANEOUS_PREFETCHES should be running at the same time. */
970 remaining_prefetch_slots = SIMULTANEOUS_PREFETCHES;
972 /* The prefetch will run for AHEAD iterations of the original loop, i.e.,
973 AHEAD / UNROLL_FACTOR iterations of the unrolled loop. In each iteration,
974 it will need a prefetch slot. */
975 slots_per_prefetch = (ahead + unroll_factor / 2) / unroll_factor;
976 if (dump_file && (dump_flags & TDF_DETAILS))
977 fprintf (dump_file, "Each prefetch instruction takes %u prefetch slots.\n",
980 /* For now we just take memory references one by one and issue
981 prefetches for as many as possible. The groups are sorted
982 starting with the largest step, since the references with
983 large step are more likely to cause many cache misses. */
985 for (; groups; groups = groups->next)
986 for (ref = groups->refs; ref; ref = ref->next)
988 if (!should_issue_prefetch_p (ref))
991 /* The loop is far from being sufficiently unrolled for this
992 prefetch. Do not generate prefetch to avoid many redudant
994 if (ref->prefetch_mod / unroll_factor > PREFETCH_MOD_TO_UNROLL_FACTOR_RATIO)
997 /* If we need to prefetch the reference each PREFETCH_MOD iterations,
998 and we unroll the loop UNROLL_FACTOR times, we need to insert
999 ceil (UNROLL_FACTOR / PREFETCH_MOD) instructions in each
1001 n_prefetches = ((unroll_factor + ref->prefetch_mod - 1)
1002 / ref->prefetch_mod);
1003 prefetch_slots = n_prefetches * slots_per_prefetch;
1005 /* If more than half of the prefetches would be lost anyway, do not
1006 issue the prefetch. */
1007 if (2 * remaining_prefetch_slots < prefetch_slots)
1010 ref->issue_prefetch_p = true;
1012 if (remaining_prefetch_slots <= prefetch_slots)
1014 remaining_prefetch_slots -= prefetch_slots;
1021 /* Return TRUE if no prefetch is going to be generated in the given
1025 nothing_to_prefetch_p (struct mem_ref_group *groups)
1027 struct mem_ref *ref;
1029 for (; groups; groups = groups->next)
1030 for (ref = groups->refs; ref; ref = ref->next)
1031 if (should_issue_prefetch_p (ref))
1037 /* Estimate the number of prefetches in the given GROUPS.
1038 UNROLL_FACTOR is the factor by which LOOP was unrolled. */
1041 estimate_prefetch_count (struct mem_ref_group *groups, unsigned unroll_factor)
1043 struct mem_ref *ref;
1044 unsigned n_prefetches;
1045 int prefetch_count = 0;
1047 for (; groups; groups = groups->next)
1048 for (ref = groups->refs; ref; ref = ref->next)
1049 if (should_issue_prefetch_p (ref))
1051 n_prefetches = ((unroll_factor + ref->prefetch_mod - 1)
1052 / ref->prefetch_mod);
1053 prefetch_count += n_prefetches;
1056 return prefetch_count;
1059 /* Issue prefetches for the reference REF into loop as decided before.
1060 HEAD is the number of iterations to prefetch ahead. UNROLL_FACTOR
1061 is the factor by which LOOP was unrolled. */
1064 issue_prefetch_ref (struct mem_ref *ref, unsigned unroll_factor, unsigned ahead)
1066 HOST_WIDE_INT delta;
1067 tree addr, addr_base, write_p, local, forward;
1069 gimple_stmt_iterator bsi;
1070 unsigned n_prefetches, ap;
1071 bool nontemporal = ref->reuse_distance >= L2_CACHE_SIZE_BYTES;
1073 if (dump_file && (dump_flags & TDF_DETAILS))
1074 fprintf (dump_file, "Issued%s prefetch for %p.\n",
1075 nontemporal ? " nontemporal" : "",
1078 bsi = gsi_for_stmt (ref->stmt);
1080 n_prefetches = ((unroll_factor + ref->prefetch_mod - 1)
1081 / ref->prefetch_mod);
1082 addr_base = build_fold_addr_expr_with_type (ref->mem, ptr_type_node);
1083 addr_base = force_gimple_operand_gsi (&bsi, unshare_expr (addr_base),
1084 true, NULL, true, GSI_SAME_STMT);
1085 write_p = ref->write_p ? integer_one_node : integer_zero_node;
1086 local = nontemporal ? integer_zero_node : integer_three_node;
1088 for (ap = 0; ap < n_prefetches; ap++)
1090 if (cst_and_fits_in_hwi (ref->group->step))
1092 /* Determine the address to prefetch. */
1093 delta = (ahead + ap * ref->prefetch_mod) *
1094 int_cst_value (ref->group->step);
1095 addr = fold_build2 (POINTER_PLUS_EXPR, ptr_type_node,
1096 addr_base, size_int (delta));
1097 addr = force_gimple_operand_gsi (&bsi, unshare_expr (addr), true, NULL,
1098 true, GSI_SAME_STMT);
1102 /* The step size is non-constant but loop-invariant. We use the
1103 heuristic to simply prefetch ahead iterations ahead. */
1104 forward = fold_build2 (MULT_EXPR, sizetype,
1105 fold_convert (sizetype, ref->group->step),
1106 fold_convert (sizetype, size_int (ahead)));
1107 addr = fold_build2 (POINTER_PLUS_EXPR, ptr_type_node, addr_base,
1109 addr = force_gimple_operand_gsi (&bsi, unshare_expr (addr), true,
1110 NULL, true, GSI_SAME_STMT);
1112 /* Create the prefetch instruction. */
1113 prefetch = gimple_build_call (built_in_decls[BUILT_IN_PREFETCH],
1114 3, addr, write_p, local);
1115 gsi_insert_before (&bsi, prefetch, GSI_SAME_STMT);
1119 /* Issue prefetches for the references in GROUPS into loop as decided before.
1120 HEAD is the number of iterations to prefetch ahead. UNROLL_FACTOR is the
1121 factor by that LOOP was unrolled. */
1124 issue_prefetches (struct mem_ref_group *groups,
1125 unsigned unroll_factor, unsigned ahead)
1127 struct mem_ref *ref;
1129 for (; groups; groups = groups->next)
1130 for (ref = groups->refs; ref; ref = ref->next)
1131 if (ref->issue_prefetch_p)
1132 issue_prefetch_ref (ref, unroll_factor, ahead);
1135 /* Returns true if REF is a memory write for that a nontemporal store insn
1139 nontemporal_store_p (struct mem_ref *ref)
1141 enum machine_mode mode;
1142 enum insn_code code;
1144 /* REF must be a write that is not reused. We require it to be independent
1145 on all other memory references in the loop, as the nontemporal stores may
1146 be reordered with respect to other memory references. */
1148 || !ref->independent_p
1149 || ref->reuse_distance < L2_CACHE_SIZE_BYTES)
1152 /* Check that we have the storent instruction for the mode. */
1153 mode = TYPE_MODE (TREE_TYPE (ref->mem));
1154 if (mode == BLKmode)
1157 code = optab_handler (storent_optab, mode);
1158 return code != CODE_FOR_nothing;
1161 /* If REF is a nontemporal store, we mark the corresponding modify statement
1162 and return true. Otherwise, we return false. */
1165 mark_nontemporal_store (struct mem_ref *ref)
1167 if (!nontemporal_store_p (ref))
1170 if (dump_file && (dump_flags & TDF_DETAILS))
1171 fprintf (dump_file, "Marked reference %p as a nontemporal store.\n",
1174 gimple_assign_set_nontemporal_move (ref->stmt, true);
1175 ref->storent_p = true;
1180 /* Issue a memory fence instruction after LOOP. */
1183 emit_mfence_after_loop (struct loop *loop)
1185 VEC (edge, heap) *exits = get_loop_exit_edges (loop);
1188 gimple_stmt_iterator bsi;
1191 for (i = 0; VEC_iterate (edge, exits, i, exit); i++)
1193 call = gimple_build_call (FENCE_FOLLOWING_MOVNT, 0);
1195 if (!single_pred_p (exit->dest)
1196 /* If possible, we prefer not to insert the fence on other paths
1198 && !(exit->flags & EDGE_ABNORMAL))
1199 split_loop_exit_edge (exit);
1200 bsi = gsi_after_labels (exit->dest);
1202 gsi_insert_before (&bsi, call, GSI_NEW_STMT);
1203 mark_virtual_ops_for_renaming (call);
1206 VEC_free (edge, heap, exits);
1207 update_ssa (TODO_update_ssa_only_virtuals);
1210 /* Returns true if we can use storent in loop, false otherwise. */
1213 may_use_storent_in_loop_p (struct loop *loop)
1217 if (loop->inner != NULL)
1220 /* If we must issue a mfence insn after using storent, check that there
1221 is a suitable place for it at each of the loop exits. */
1222 if (FENCE_FOLLOWING_MOVNT != NULL_TREE)
1224 VEC (edge, heap) *exits = get_loop_exit_edges (loop);
1228 for (i = 0; VEC_iterate (edge, exits, i, exit); i++)
1229 if ((exit->flags & EDGE_ABNORMAL)
1230 && exit->dest == EXIT_BLOCK_PTR)
1233 VEC_free (edge, heap, exits);
1239 /* Marks nontemporal stores in LOOP. GROUPS contains the description of memory
1240 references in the loop. */
1243 mark_nontemporal_stores (struct loop *loop, struct mem_ref_group *groups)
1245 struct mem_ref *ref;
1248 if (!may_use_storent_in_loop_p (loop))
1251 for (; groups; groups = groups->next)
1252 for (ref = groups->refs; ref; ref = ref->next)
1253 any |= mark_nontemporal_store (ref);
1255 if (any && FENCE_FOLLOWING_MOVNT != NULL_TREE)
1256 emit_mfence_after_loop (loop);
1259 /* Determines whether we can profitably unroll LOOP FACTOR times, and if
1260 this is the case, fill in DESC by the description of number of
1264 should_unroll_loop_p (struct loop *loop, struct tree_niter_desc *desc,
1267 if (!can_unroll_loop_p (loop, factor, desc))
1270 /* We only consider loops without control flow for unrolling. This is not
1271 a hard restriction -- tree_unroll_loop works with arbitrary loops
1272 as well; but the unrolling/prefetching is usually more profitable for
1273 loops consisting of a single basic block, and we want to limit the
1275 if (loop->num_nodes > 2)
1281 /* Determine the coefficient by that unroll LOOP, from the information
1282 contained in the list of memory references REFS. Description of
1283 umber of iterations of LOOP is stored to DESC. NINSNS is the number of
1284 insns of the LOOP. EST_NITER is the estimated number of iterations of
1285 the loop, or -1 if no estimate is available. */
1288 determine_unroll_factor (struct loop *loop, struct mem_ref_group *refs,
1289 unsigned ninsns, struct tree_niter_desc *desc,
1290 HOST_WIDE_INT est_niter)
1292 unsigned upper_bound;
1293 unsigned nfactor, factor, mod_constraint;
1294 struct mem_ref_group *agp;
1295 struct mem_ref *ref;
1297 /* First check whether the loop is not too large to unroll. We ignore
1298 PARAM_MAX_UNROLL_TIMES, because for small loops, it prevented us
1299 from unrolling them enough to make exactly one cache line covered by each
1300 iteration. Also, the goal of PARAM_MAX_UNROLL_TIMES is to prevent
1301 us from unrolling the loops too many times in cases where we only expect
1302 gains from better scheduling and decreasing loop overhead, which is not
1304 upper_bound = PARAM_VALUE (PARAM_MAX_UNROLLED_INSNS) / ninsns;
1306 /* If we unrolled the loop more times than it iterates, the unrolled version
1307 of the loop would be never entered. */
1308 if (est_niter >= 0 && est_niter < (HOST_WIDE_INT) upper_bound)
1309 upper_bound = est_niter;
1311 if (upper_bound <= 1)
1314 /* Choose the factor so that we may prefetch each cache just once,
1315 but bound the unrolling by UPPER_BOUND. */
1317 for (agp = refs; agp; agp = agp->next)
1318 for (ref = agp->refs; ref; ref = ref->next)
1319 if (should_issue_prefetch_p (ref))
1321 mod_constraint = ref->prefetch_mod;
1322 nfactor = least_common_multiple (mod_constraint, factor);
1323 if (nfactor <= upper_bound)
1327 if (!should_unroll_loop_p (loop, desc, factor))
1333 /* Returns the total volume of the memory references REFS, taking into account
1334 reuses in the innermost loop and cache line size. TODO -- we should also
1335 take into account reuses across the iterations of the loops in the loop
1339 volume_of_references (struct mem_ref_group *refs)
1341 unsigned volume = 0;
1342 struct mem_ref_group *gr;
1343 struct mem_ref *ref;
1345 for (gr = refs; gr; gr = gr->next)
1346 for (ref = gr->refs; ref; ref = ref->next)
1348 /* Almost always reuses another value? */
1349 if (ref->prefetch_before != PREFETCH_ALL)
1352 /* If several iterations access the same cache line, use the size of
1353 the line divided by this number. Otherwise, a cache line is
1354 accessed in each iteration. TODO -- in the latter case, we should
1355 take the size of the reference into account, rounding it up on cache
1356 line size multiple. */
1357 volume += L1_CACHE_LINE_SIZE / ref->prefetch_mod;
1362 /* Returns the volume of memory references accessed across VEC iterations of
1363 loops, whose sizes are described in the LOOP_SIZES array. N is the number
1364 of the loops in the nest (length of VEC and LOOP_SIZES vectors). */
1367 volume_of_dist_vector (lambda_vector vec, unsigned *loop_sizes, unsigned n)
1371 for (i = 0; i < n; i++)
1378 gcc_assert (vec[i] > 0);
1380 /* We ignore the parts of the distance vector in subloops, since usually
1381 the numbers of iterations are much smaller. */
1382 return loop_sizes[i] * vec[i];
1385 /* Add the steps of ACCESS_FN multiplied by STRIDE to the array STRIDE
1386 at the position corresponding to the loop of the step. N is the depth
1387 of the considered loop nest, and, LOOP is its innermost loop. */
1390 add_subscript_strides (tree access_fn, unsigned stride,
1391 HOST_WIDE_INT *strides, unsigned n, struct loop *loop)
1395 HOST_WIDE_INT astep;
1396 unsigned min_depth = loop_depth (loop) - n;
1398 while (TREE_CODE (access_fn) == POLYNOMIAL_CHREC)
1400 aloop = get_chrec_loop (access_fn);
1401 step = CHREC_RIGHT (access_fn);
1402 access_fn = CHREC_LEFT (access_fn);
1404 if ((unsigned) loop_depth (aloop) <= min_depth)
1407 if (host_integerp (step, 0))
1408 astep = tree_low_cst (step, 0);
1410 astep = L1_CACHE_LINE_SIZE;
1412 strides[n - 1 - loop_depth (loop) + loop_depth (aloop)] += astep * stride;
1417 /* Returns the volume of memory references accessed between two consecutive
1418 self-reuses of the reference DR. We consider the subscripts of DR in N
1419 loops, and LOOP_SIZES contains the volumes of accesses in each of the
1420 loops. LOOP is the innermost loop of the current loop nest. */
1423 self_reuse_distance (data_reference_p dr, unsigned *loop_sizes, unsigned n,
1426 tree stride, access_fn;
1427 HOST_WIDE_INT *strides, astride;
1428 VEC (tree, heap) *access_fns;
1429 tree ref = DR_REF (dr);
1430 unsigned i, ret = ~0u;
1432 /* In the following example:
1434 for (i = 0; i < N; i++)
1435 for (j = 0; j < N; j++)
1437 the same cache line is accessed each N steps (except if the change from
1438 i to i + 1 crosses the boundary of the cache line). Thus, for self-reuse,
1439 we cannot rely purely on the results of the data dependence analysis.
1441 Instead, we compute the stride of the reference in each loop, and consider
1442 the innermost loop in that the stride is less than cache size. */
1444 strides = XCNEWVEC (HOST_WIDE_INT, n);
1445 access_fns = DR_ACCESS_FNS (dr);
1447 for (i = 0; VEC_iterate (tree, access_fns, i, access_fn); i++)
1449 /* Keep track of the reference corresponding to the subscript, so that we
1451 while (handled_component_p (ref) && TREE_CODE (ref) != ARRAY_REF)
1452 ref = TREE_OPERAND (ref, 0);
1454 if (TREE_CODE (ref) == ARRAY_REF)
1456 stride = TYPE_SIZE_UNIT (TREE_TYPE (ref));
1457 if (host_integerp (stride, 1))
1458 astride = tree_low_cst (stride, 1);
1460 astride = L1_CACHE_LINE_SIZE;
1462 ref = TREE_OPERAND (ref, 0);
1467 add_subscript_strides (access_fn, astride, strides, n, loop);
1470 for (i = n; i-- > 0; )
1472 unsigned HOST_WIDE_INT s;
1474 s = strides[i] < 0 ? -strides[i] : strides[i];
1476 if (s < (unsigned) L1_CACHE_LINE_SIZE
1478 > (unsigned) (L1_CACHE_SIZE_BYTES / NONTEMPORAL_FRACTION)))
1480 ret = loop_sizes[i];
1489 /* Determines the distance till the first reuse of each reference in REFS
1490 in the loop nest of LOOP. NO_OTHER_REFS is true if there are no other
1491 memory references in the loop. */
1494 determine_loop_nest_reuse (struct loop *loop, struct mem_ref_group *refs,
1497 struct loop *nest, *aloop;
1498 VEC (data_reference_p, heap) *datarefs = NULL;
1499 VEC (ddr_p, heap) *dependences = NULL;
1500 struct mem_ref_group *gr;
1501 struct mem_ref *ref, *refb;
1502 VEC (loop_p, heap) *vloops = NULL;
1503 unsigned *loop_data_size;
1505 unsigned volume, dist, adist;
1507 data_reference_p dr;
1513 /* Find the outermost loop of the loop nest of loop (we require that
1514 there are no sibling loops inside the nest). */
1518 aloop = loop_outer (nest);
1520 if (aloop == current_loops->tree_root
1521 || aloop->inner->next)
1527 /* For each loop, determine the amount of data accessed in each iteration.
1528 We use this to estimate whether the reference is evicted from the
1529 cache before its reuse. */
1530 find_loop_nest (nest, &vloops);
1531 n = VEC_length (loop_p, vloops);
1532 loop_data_size = XNEWVEC (unsigned, n);
1533 volume = volume_of_references (refs);
1537 loop_data_size[i] = volume;
1538 /* Bound the volume by the L2 cache size, since above this bound,
1539 all dependence distances are equivalent. */
1540 if (volume > L2_CACHE_SIZE_BYTES)
1543 aloop = VEC_index (loop_p, vloops, i);
1544 vol = estimated_loop_iterations_int (aloop, false);
1546 vol = expected_loop_iterations (aloop);
1550 /* Prepare the references in the form suitable for data dependence
1551 analysis. We ignore unanalyzable data references (the results
1552 are used just as a heuristics to estimate temporality of the
1553 references, hence we do not need to worry about correctness). */
1554 for (gr = refs; gr; gr = gr->next)
1555 for (ref = gr->refs; ref; ref = ref->next)
1557 dr = create_data_ref (nest, ref->mem, ref->stmt, !ref->write_p);
1561 ref->reuse_distance = volume;
1563 VEC_safe_push (data_reference_p, heap, datarefs, dr);
1566 no_other_refs = false;
1569 for (i = 0; VEC_iterate (data_reference_p, datarefs, i, dr); i++)
1571 dist = self_reuse_distance (dr, loop_data_size, n, loop);
1572 ref = (struct mem_ref *) dr->aux;
1573 if (ref->reuse_distance > dist)
1574 ref->reuse_distance = dist;
1577 ref->independent_p = true;
1580 compute_all_dependences (datarefs, &dependences, vloops, true);
1582 for (i = 0; VEC_iterate (ddr_p, dependences, i, dep); i++)
1584 if (DDR_ARE_DEPENDENT (dep) == chrec_known)
1587 ref = (struct mem_ref *) DDR_A (dep)->aux;
1588 refb = (struct mem_ref *) DDR_B (dep)->aux;
1590 if (DDR_ARE_DEPENDENT (dep) == chrec_dont_know
1591 || DDR_NUM_DIST_VECTS (dep) == 0)
1593 /* If the dependence cannot be analyzed, assume that there might be
1597 ref->independent_p = false;
1598 refb->independent_p = false;
1602 /* The distance vectors are normalized to be always lexicographically
1603 positive, hence we cannot tell just from them whether DDR_A comes
1604 before DDR_B or vice versa. However, it is not important,
1605 anyway -- if DDR_A is close to DDR_B, then it is either reused in
1606 DDR_B (and it is not nontemporal), or it reuses the value of DDR_B
1607 in cache (and marking it as nontemporal would not affect
1611 for (j = 0; j < DDR_NUM_DIST_VECTS (dep); j++)
1613 adist = volume_of_dist_vector (DDR_DIST_VECT (dep, j),
1616 /* If this is a dependence in the innermost loop (i.e., the
1617 distances in all superloops are zero) and it is not
1618 the trivial self-dependence with distance zero, record that
1619 the references are not completely independent. */
1620 if (lambda_vector_zerop (DDR_DIST_VECT (dep, j), n - 1)
1622 || DDR_DIST_VECT (dep, j)[n-1] != 0))
1624 ref->independent_p = false;
1625 refb->independent_p = false;
1628 /* Ignore accesses closer than
1629 L1_CACHE_SIZE_BYTES / NONTEMPORAL_FRACTION,
1630 so that we use nontemporal prefetches e.g. if single memory
1631 location is accessed several times in a single iteration of
1633 if (adist < L1_CACHE_SIZE_BYTES / NONTEMPORAL_FRACTION)
1641 if (ref->reuse_distance > dist)
1642 ref->reuse_distance = dist;
1643 if (refb->reuse_distance > dist)
1644 refb->reuse_distance = dist;
1647 free_dependence_relations (dependences);
1648 free_data_refs (datarefs);
1649 free (loop_data_size);
1651 if (dump_file && (dump_flags & TDF_DETAILS))
1653 fprintf (dump_file, "Reuse distances:\n");
1654 for (gr = refs; gr; gr = gr->next)
1655 for (ref = gr->refs; ref; ref = ref->next)
1656 fprintf (dump_file, " ref %p distance %u\n",
1657 (void *) ref, ref->reuse_distance);
1661 /* Determine whether or not the trip count to ahead ratio is too small based
1662 on prefitablility consideration.
1663 AHEAD: the iteration ahead distance,
1664 EST_NITER: the estimated trip count. */
1667 trip_count_to_ahead_ratio_too_small_p (unsigned ahead, HOST_WIDE_INT est_niter)
1669 /* Assume trip count to ahead ratio is big enough if the trip count could not
1670 be estimated at compile time. */
1674 if (est_niter < (HOST_WIDE_INT) (TRIP_COUNT_TO_AHEAD_RATIO * ahead))
1676 if (dump_file && (dump_flags & TDF_DETAILS))
1678 "Not prefetching -- loop estimated to roll only %d times\n",
1686 /* Determine whether or not the number of memory references in the loop is
1687 reasonable based on the profitablity and compilation time considerations.
1688 NINSNS: estimated number of instructions in the loop,
1689 MEM_REF_COUNT: total number of memory references in the loop. */
1692 mem_ref_count_reasonable_p (unsigned ninsns, unsigned mem_ref_count)
1694 int insn_to_mem_ratio;
1696 if (mem_ref_count == 0)
1699 /* Miss rate computation (is_miss_rate_acceptable) and dependence analysis
1700 (compute_all_dependences) have high costs based on quadratic complexity.
1701 To avoid huge compilation time, we give up prefetching if mem_ref_count
1703 if (mem_ref_count > PREFETCH_MAX_MEM_REFS_PER_LOOP)
1706 /* Prefetching improves performance by overlapping cache missing
1707 memory accesses with CPU operations. If the loop does not have
1708 enough CPU operations to overlap with memory operations, prefetching
1709 won't give a significant benefit. One approximate way of checking
1710 this is to require the ratio of instructions to memory references to
1711 be above a certain limit. This approximation works well in practice.
1712 TODO: Implement a more precise computation by estimating the time
1713 for each CPU or memory op in the loop. Time estimates for memory ops
1714 should account for cache misses. */
1715 insn_to_mem_ratio = ninsns / mem_ref_count;
1717 if (insn_to_mem_ratio < PREFETCH_MIN_INSN_TO_MEM_RATIO)
1719 if (dump_file && (dump_flags & TDF_DETAILS))
1721 "Not prefetching -- instruction to memory reference ratio (%d) too small\n",
1729 /* Determine whether or not the instruction to prefetch ratio in the loop is
1730 too small based on the profitablity consideration.
1731 NINSNS: estimated number of instructions in the loop,
1732 PREFETCH_COUNT: an estimate of the number of prefetches,
1733 UNROLL_FACTOR: the factor to unroll the loop if prefetching. */
1736 insn_to_prefetch_ratio_too_small_p (unsigned ninsns, unsigned prefetch_count,
1737 unsigned unroll_factor)
1739 int insn_to_prefetch_ratio;
1741 /* Prefetching most likely causes performance degradation when the instruction
1742 to prefetch ratio is too small. Too many prefetch instructions in a loop
1743 may reduce the I-cache performance.
1744 (unroll_factor * ninsns) is used to estimate the number of instructions in
1745 the unrolled loop. This implementation is a bit simplistic -- the number
1746 of issued prefetch instructions is also affected by unrolling. So,
1747 prefetch_mod and the unroll factor should be taken into account when
1748 determining prefetch_count. Also, the number of insns of the unrolled
1749 loop will usually be significantly smaller than the number of insns of the
1750 original loop * unroll_factor (at least the induction variable increases
1751 and the exit branches will get eliminated), so it might be better to use
1752 tree_estimate_loop_size + estimated_unrolled_size. */
1753 insn_to_prefetch_ratio = (unroll_factor * ninsns) / prefetch_count;
1754 if (insn_to_prefetch_ratio < MIN_INSN_TO_PREFETCH_RATIO)
1756 if (dump_file && (dump_flags & TDF_DETAILS))
1758 "Not prefetching -- instruction to prefetch ratio (%d) too small\n",
1759 insn_to_prefetch_ratio);
1767 /* Issue prefetch instructions for array references in LOOP. Returns
1768 true if the LOOP was unrolled. */
1771 loop_prefetch_arrays (struct loop *loop)
1773 struct mem_ref_group *refs;
1774 unsigned ahead, ninsns, time, unroll_factor;
1775 HOST_WIDE_INT est_niter;
1776 struct tree_niter_desc desc;
1777 bool unrolled = false, no_other_refs;
1778 unsigned prefetch_count;
1779 unsigned mem_ref_count;
1781 if (optimize_loop_nest_for_size_p (loop))
1783 if (dump_file && (dump_flags & TDF_DETAILS))
1784 fprintf (dump_file, " ignored (cold area)\n");
1788 /* FIXME: the time should be weighted by the probabilities of the blocks in
1790 time = tree_num_loop_insns (loop, &eni_time_weights);
1794 ahead = (PREFETCH_LATENCY + time - 1) / time;
1795 est_niter = estimated_loop_iterations_int (loop, false);
1797 /* Prefetching is not likely to be profitable if the trip count to ahead
1798 ratio is too small. */
1799 if (trip_count_to_ahead_ratio_too_small_p (ahead, est_niter))
1802 ninsns = tree_num_loop_insns (loop, &eni_size_weights);
1804 /* Step 1: gather the memory references. */
1805 refs = gather_memory_references (loop, &no_other_refs, &mem_ref_count);
1807 /* Give up prefetching if the number of memory references in the
1808 loop is not reasonable based on profitablity and compilation time
1810 if (!mem_ref_count_reasonable_p (ninsns, mem_ref_count))
1813 /* Step 2: estimate the reuse effects. */
1814 prune_by_reuse (refs);
1816 if (nothing_to_prefetch_p (refs))
1819 determine_loop_nest_reuse (loop, refs, no_other_refs);
1821 /* Step 3: determine unroll factor. */
1822 unroll_factor = determine_unroll_factor (loop, refs, ninsns, &desc,
1825 /* Estimate prefetch count for the unrolled loop. */
1826 prefetch_count = estimate_prefetch_count (refs, unroll_factor);
1827 if (prefetch_count == 0)
1830 if (dump_file && (dump_flags & TDF_DETAILS))
1831 fprintf (dump_file, "Ahead %d, unroll factor %d, trip count "
1832 HOST_WIDE_INT_PRINT_DEC "\n"
1833 "insn count %d, mem ref count %d, prefetch count %d\n",
1834 ahead, unroll_factor, est_niter,
1835 ninsns, mem_ref_count, prefetch_count);
1837 /* Prefetching is not likely to be profitable if the instruction to prefetch
1838 ratio is too small. */
1839 if (insn_to_prefetch_ratio_too_small_p (ninsns, prefetch_count,
1843 mark_nontemporal_stores (loop, refs);
1845 /* Step 4: what to prefetch? */
1846 if (!schedule_prefetches (refs, unroll_factor, ahead))
1849 /* Step 5: unroll the loop. TODO -- peeling of first and last few
1850 iterations so that we do not issue superfluous prefetches. */
1851 if (unroll_factor != 1)
1853 tree_unroll_loop (loop, unroll_factor,
1854 single_dom_exit (loop), &desc);
1858 /* Step 6: issue the prefetches. */
1859 issue_prefetches (refs, unroll_factor, ahead);
1862 release_mem_refs (refs);
1866 /* Issue prefetch instructions for array references in loops. */
1869 tree_ssa_prefetch_arrays (void)
1873 bool unrolled = false;
1877 /* It is possible to ask compiler for say -mtune=i486 -march=pentium4.
1878 -mtune=i486 causes us having PREFETCH_BLOCK 0, since this is part
1879 of processor costs and i486 does not have prefetch, but
1880 -march=pentium4 causes HAVE_prefetch to be true. Ugh. */
1881 || PREFETCH_BLOCK == 0)
1884 if (dump_file && (dump_flags & TDF_DETAILS))
1886 fprintf (dump_file, "Prefetching parameters:\n");
1887 fprintf (dump_file, " simultaneous prefetches: %d\n",
1888 SIMULTANEOUS_PREFETCHES);
1889 fprintf (dump_file, " prefetch latency: %d\n", PREFETCH_LATENCY);
1890 fprintf (dump_file, " prefetch block size: %d\n", PREFETCH_BLOCK);
1891 fprintf (dump_file, " L1 cache size: %d lines, %d kB\n",
1892 L1_CACHE_SIZE_BYTES / L1_CACHE_LINE_SIZE, L1_CACHE_SIZE);
1893 fprintf (dump_file, " L1 cache line size: %d\n", L1_CACHE_LINE_SIZE);
1894 fprintf (dump_file, " L2 cache size: %d kB\n", L2_CACHE_SIZE);
1895 fprintf (dump_file, " min insn-to-prefetch ratio: %d \n",
1896 MIN_INSN_TO_PREFETCH_RATIO);
1897 fprintf (dump_file, " min insn-to-mem ratio: %d \n",
1898 PREFETCH_MIN_INSN_TO_MEM_RATIO);
1899 fprintf (dump_file, "\n");
1902 initialize_original_copy_tables ();
1904 if (!built_in_decls[BUILT_IN_PREFETCH])
1906 tree type = build_function_type_list (void_type_node,
1907 const_ptr_type_node, NULL_TREE);
1908 tree decl = add_builtin_function ("__builtin_prefetch", type,
1909 BUILT_IN_PREFETCH, BUILT_IN_NORMAL,
1911 DECL_IS_NOVOPS (decl) = true;
1912 built_in_decls[BUILT_IN_PREFETCH] = decl;
1915 /* We assume that size of cache line is a power of two, so verify this
1917 gcc_assert ((PREFETCH_BLOCK & (PREFETCH_BLOCK - 1)) == 0);
1919 FOR_EACH_LOOP (li, loop, LI_FROM_INNERMOST)
1921 if (dump_file && (dump_flags & TDF_DETAILS))
1922 fprintf (dump_file, "Processing loop %d:\n", loop->num);
1924 unrolled |= loop_prefetch_arrays (loop);
1926 if (dump_file && (dump_flags & TDF_DETAILS))
1927 fprintf (dump_file, "\n\n");
1933 todo_flags |= TODO_cleanup_cfg;
1936 free_original_copy_tables ();