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"
41 #include "langhooks.h"
42 #include "tree-inline.h"
43 #include "tree-data-ref.h"
46 /* FIXME: Needed for optabs, but this should all be moved to a TBD interface
47 between the GIMPLE and RTL worlds. */
51 /* This pass inserts prefetch instructions to optimize cache usage during
52 accesses to arrays in loops. It processes loops sequentially and:
54 1) Gathers all memory references in the single loop.
55 2) For each of the references it decides when it is profitable to prefetch
56 it. To do it, we evaluate the reuse among the accesses, and determines
57 two values: PREFETCH_BEFORE (meaning that it only makes sense to do
58 prefetching in the first PREFETCH_BEFORE iterations of the loop) and
59 PREFETCH_MOD (meaning that it only makes sense to prefetch in the
60 iterations of the loop that are zero modulo PREFETCH_MOD). For example
61 (assuming cache line size is 64 bytes, char has size 1 byte and there
62 is no hardware sequential prefetch):
65 for (i = 0; i < max; i++)
72 a[187*i + 50] = ...; (5)
75 (0) obviously has PREFETCH_BEFORE 1
76 (1) has PREFETCH_BEFORE 64, since (2) accesses the same memory
77 location 64 iterations before it, and PREFETCH_MOD 64 (since
78 it hits the same cache line otherwise).
79 (2) has PREFETCH_MOD 64
80 (3) has PREFETCH_MOD 4
81 (4) has PREFETCH_MOD 1. We do not set PREFETCH_BEFORE here, since
82 the cache line accessed by (4) is the same with probability only
84 (5) has PREFETCH_MOD 1 as well.
86 Additionally, we use data dependence analysis to determine for each
87 reference the distance till the first reuse; this information is used
88 to determine the temporality of the issued prefetch instruction.
90 3) We determine how much ahead we need to prefetch. The number of
91 iterations needed is time to fetch / time spent in one iteration of
92 the loop. The problem is that we do not know either of these values,
93 so we just make a heuristic guess based on a magic (possibly)
94 target-specific constant and size of the loop.
96 4) Determine which of the references we prefetch. We take into account
97 that there is a maximum number of simultaneous prefetches (provided
98 by machine description). We prefetch as many prefetches as possible
99 while still within this bound (starting with those with lowest
100 prefetch_mod, since they are responsible for most of the cache
103 5) We unroll and peel loops so that we are able to satisfy PREFETCH_MOD
104 and PREFETCH_BEFORE requirements (within some bounds), and to avoid
105 prefetching nonaccessed memory.
106 TODO -- actually implement peeling.
108 6) We actually emit the prefetch instructions. ??? Perhaps emit the
109 prefetch instructions with guards in cases where 5) was not sufficient
110 to satisfy the constraints?
112 The function is_loop_prefetching_profitable() implements a cost model
113 to determine if prefetching is profitable for a given loop. The cost
114 model has two heuristcs:
115 1. A heuristic that determines whether the given loop has enough CPU
116 ops that can be overlapped with cache missing memory ops.
117 If not, the loop won't benefit from prefetching. This is implemented
118 by requirung the ratio between the instruction count and the mem ref
119 count to be above a certain minimum.
120 2. A heuristic that disables prefetching in a loop with an unknown trip
121 count if the prefetching cost is above a certain limit. The relative
122 prefetching cost is estimated by taking the ratio between the
123 prefetch count and the total intruction count (this models the I-cache
125 The limits used in these heuristics are defined as parameters with
126 reasonable default values. Machine-specific default values will be
130 -- write and use more general reuse analysis (that could be also used
131 in other cache aimed loop optimizations)
132 -- make it behave sanely together with the prefetches given by user
133 (now we just ignore them; at the very least we should avoid
134 optimizing loops in that user put his own prefetches)
135 -- we assume cache line size alignment of arrays; this could be
138 /* Magic constants follow. These should be replaced by machine specific
141 /* True if write can be prefetched by a read prefetch. */
143 #ifndef WRITE_CAN_USE_READ_PREFETCH
144 #define WRITE_CAN_USE_READ_PREFETCH 1
147 /* True if read can be prefetched by a write prefetch. */
149 #ifndef READ_CAN_USE_WRITE_PREFETCH
150 #define READ_CAN_USE_WRITE_PREFETCH 0
153 /* The size of the block loaded by a single prefetch. Usually, this is
154 the same as cache line size (at the moment, we only consider one level
155 of cache hierarchy). */
157 #ifndef PREFETCH_BLOCK
158 #define PREFETCH_BLOCK L1_CACHE_LINE_SIZE
161 /* Do we have a forward hardware sequential prefetching? */
163 #ifndef HAVE_FORWARD_PREFETCH
164 #define HAVE_FORWARD_PREFETCH 0
167 /* Do we have a backward hardware sequential prefetching? */
169 #ifndef HAVE_BACKWARD_PREFETCH
170 #define HAVE_BACKWARD_PREFETCH 0
173 /* In some cases we are only able to determine that there is a certain
174 probability that the two accesses hit the same cache line. In this
175 case, we issue the prefetches for both of them if this probability
176 is less then (1000 - ACCEPTABLE_MISS_RATE) per thousand. */
178 #ifndef ACCEPTABLE_MISS_RATE
179 #define ACCEPTABLE_MISS_RATE 50
182 #ifndef HAVE_prefetch
183 #define HAVE_prefetch 0
186 #define L1_CACHE_SIZE_BYTES ((unsigned) (L1_CACHE_SIZE * 1024))
187 #define L2_CACHE_SIZE_BYTES ((unsigned) (L2_CACHE_SIZE * 1024))
189 /* We consider a memory access nontemporal if it is not reused sooner than
190 after L2_CACHE_SIZE_BYTES of memory are accessed. However, we ignore
191 accesses closer than L1_CACHE_SIZE_BYTES / NONTEMPORAL_FRACTION,
192 so that we use nontemporal prefetches e.g. if single memory location
193 is accessed several times in a single iteration of the loop. */
194 #define NONTEMPORAL_FRACTION 16
196 /* In case we have to emit a memory fence instruction after the loop that
197 uses nontemporal stores, this defines the builtin to use. */
199 #ifndef FENCE_FOLLOWING_MOVNT
200 #define FENCE_FOLLOWING_MOVNT NULL_TREE
203 /* It is not profitable to prefetch when the trip count is not at
204 least TRIP_COUNT_TO_AHEAD_RATIO times the prefetch ahead distance.
205 For example, in a loop with a prefetch ahead distance of 10,
206 supposing that TRIP_COUNT_TO_AHEAD_RATIO is equal to 4, it is
207 profitable to prefetch when the trip count is greater or equal to
208 40. In that case, 30 out of the 40 iterations will benefit from
211 #ifndef TRIP_COUNT_TO_AHEAD_RATIO
212 #define TRIP_COUNT_TO_AHEAD_RATIO 4
215 /* The group of references between that reuse may occur. */
219 tree base; /* Base of the reference. */
220 tree step; /* Step of the reference. */
221 struct mem_ref *refs; /* References in the group. */
222 struct mem_ref_group *next; /* Next group of references. */
225 /* Assigned to PREFETCH_BEFORE when all iterations are to be prefetched. */
227 #define PREFETCH_ALL (~(unsigned HOST_WIDE_INT) 0)
229 /* Do not generate a prefetch if the unroll factor is significantly less
230 than what is required by the prefetch. This is to avoid redundant
231 prefetches. For example, if prefetch_mod is 16 and unroll_factor is
232 1, this means prefetching requires unrolling the loop 16 times, but
233 the loop is not going to be unrolled. In this case (ratio = 16),
234 prefetching is not likely to be beneficial. */
236 #ifndef PREFETCH_MOD_TO_UNROLL_FACTOR_RATIO
237 #define PREFETCH_MOD_TO_UNROLL_FACTOR_RATIO 8
240 /* The memory reference. */
244 gimple stmt; /* Statement in that the reference appears. */
245 tree mem; /* The reference. */
246 HOST_WIDE_INT delta; /* Constant offset of the reference. */
247 struct mem_ref_group *group; /* The group of references it belongs to. */
248 unsigned HOST_WIDE_INT prefetch_mod;
249 /* Prefetch only each PREFETCH_MOD-th
251 unsigned HOST_WIDE_INT prefetch_before;
252 /* Prefetch only first PREFETCH_BEFORE
254 unsigned reuse_distance; /* The amount of data accessed before the first
255 reuse of this value. */
256 struct mem_ref *next; /* The next reference in the group. */
257 unsigned write_p : 1; /* Is it a write? */
258 unsigned independent_p : 1; /* True if the reference is independent on
259 all other references inside the loop. */
260 unsigned issue_prefetch_p : 1; /* Should we really issue the prefetch? */
261 unsigned storent_p : 1; /* True if we changed the store to a
265 /* Dumps information about reference REF to FILE. */
268 dump_mem_ref (FILE *file, struct mem_ref *ref)
270 fprintf (file, "Reference %p:\n", (void *) ref);
272 fprintf (file, " group %p (base ", (void *) ref->group);
273 print_generic_expr (file, ref->group->base, TDF_SLIM);
274 fprintf (file, ", step ");
275 if (cst_and_fits_in_hwi (ref->group->step))
276 fprintf (file, HOST_WIDE_INT_PRINT_DEC, int_cst_value (ref->group->step));
278 print_generic_expr (file, ref->group->step, TDF_TREE);
279 fprintf (file, ")\n");
281 fprintf (file, " delta ");
282 fprintf (file, HOST_WIDE_INT_PRINT_DEC, ref->delta);
283 fprintf (file, "\n");
285 fprintf (file, " %s\n", ref->write_p ? "write" : "read");
287 fprintf (file, "\n");
290 /* Finds a group with BASE and STEP in GROUPS, or creates one if it does not
293 static struct mem_ref_group *
294 find_or_create_group (struct mem_ref_group **groups, tree base, tree step)
296 struct mem_ref_group *group;
298 for (; *groups; groups = &(*groups)->next)
300 if (operand_equal_p ((*groups)->step, step, 0)
301 && operand_equal_p ((*groups)->base, base, 0))
304 /* If step is an integer constant, keep the list of groups sorted
305 by decreasing step. */
306 if (cst_and_fits_in_hwi ((*groups)->step) && cst_and_fits_in_hwi (step)
307 && int_cst_value ((*groups)->step) < int_cst_value (step))
311 group = XNEW (struct mem_ref_group);
315 group->next = *groups;
321 /* Records a memory reference MEM in GROUP with offset DELTA and write status
322 WRITE_P. The reference occurs in statement STMT. */
325 record_ref (struct mem_ref_group *group, gimple stmt, tree mem,
326 HOST_WIDE_INT delta, bool write_p)
328 struct mem_ref **aref;
330 /* Do not record the same address twice. */
331 for (aref = &group->refs; *aref; aref = &(*aref)->next)
333 /* It does not have to be possible for write reference to reuse the read
334 prefetch, or vice versa. */
335 if (!WRITE_CAN_USE_READ_PREFETCH
337 && !(*aref)->write_p)
339 if (!READ_CAN_USE_WRITE_PREFETCH
344 if ((*aref)->delta == delta)
348 (*aref) = XNEW (struct mem_ref);
349 (*aref)->stmt = stmt;
351 (*aref)->delta = delta;
352 (*aref)->write_p = write_p;
353 (*aref)->prefetch_before = PREFETCH_ALL;
354 (*aref)->prefetch_mod = 1;
355 (*aref)->reuse_distance = 0;
356 (*aref)->issue_prefetch_p = false;
357 (*aref)->group = group;
358 (*aref)->next = NULL;
359 (*aref)->independent_p = false;
360 (*aref)->storent_p = false;
362 if (dump_file && (dump_flags & TDF_DETAILS))
363 dump_mem_ref (dump_file, *aref);
366 /* Release memory references in GROUPS. */
369 release_mem_refs (struct mem_ref_group *groups)
371 struct mem_ref_group *next_g;
372 struct mem_ref *ref, *next_r;
374 for (; groups; groups = next_g)
376 next_g = groups->next;
377 for (ref = groups->refs; ref; ref = next_r)
386 /* A structure used to pass arguments to idx_analyze_ref. */
390 struct loop *loop; /* Loop of the reference. */
391 gimple stmt; /* Statement of the reference. */
392 tree *step; /* Step of the memory reference. */
393 HOST_WIDE_INT *delta; /* Offset of the memory reference. */
396 /* Analyzes a single INDEX of a memory reference to obtain information
397 described at analyze_ref. Callback for for_each_index. */
400 idx_analyze_ref (tree base, tree *index, void *data)
402 struct ar_data *ar_data = (struct ar_data *) data;
403 tree ibase, step, stepsize;
404 HOST_WIDE_INT idelta = 0, imult = 1;
407 if (TREE_CODE (base) == MISALIGNED_INDIRECT_REF
408 || TREE_CODE (base) == ALIGN_INDIRECT_REF)
411 if (!simple_iv (ar_data->loop, loop_containing_stmt (ar_data->stmt),
417 if (TREE_CODE (ibase) == POINTER_PLUS_EXPR
418 && cst_and_fits_in_hwi (TREE_OPERAND (ibase, 1)))
420 idelta = int_cst_value (TREE_OPERAND (ibase, 1));
421 ibase = TREE_OPERAND (ibase, 0);
423 if (cst_and_fits_in_hwi (ibase))
425 idelta += int_cst_value (ibase);
426 ibase = build_int_cst (TREE_TYPE (ibase), 0);
429 if (TREE_CODE (base) == ARRAY_REF)
431 stepsize = array_ref_element_size (base);
432 if (!cst_and_fits_in_hwi (stepsize))
434 imult = int_cst_value (stepsize);
435 step = fold_build2 (MULT_EXPR, sizetype,
436 fold_convert (sizetype, step),
437 fold_convert (sizetype, stepsize));
441 if (*ar_data->step == NULL_TREE)
442 *ar_data->step = step;
444 *ar_data->step = fold_build2 (PLUS_EXPR, sizetype,
445 fold_convert (sizetype, *ar_data->step),
446 fold_convert (sizetype, step));
447 *ar_data->delta += idelta;
453 /* Tries to express REF_P in shape &BASE + STEP * iter + DELTA, where DELTA and
454 STEP are integer constants and iter is number of iterations of LOOP. The
455 reference occurs in statement STMT. Strips nonaddressable component
456 references from REF_P. */
459 analyze_ref (struct loop *loop, tree *ref_p, tree *base,
460 tree *step, HOST_WIDE_INT *delta,
463 struct ar_data ar_data;
465 HOST_WIDE_INT bit_offset;
471 /* First strip off the component references. Ignore bitfields. */
472 if (TREE_CODE (ref) == COMPONENT_REF
473 && DECL_NONADDRESSABLE_P (TREE_OPERAND (ref, 1)))
474 ref = TREE_OPERAND (ref, 0);
478 for (; TREE_CODE (ref) == COMPONENT_REF; ref = TREE_OPERAND (ref, 0))
480 off = DECL_FIELD_BIT_OFFSET (TREE_OPERAND (ref, 1));
481 bit_offset = TREE_INT_CST_LOW (off);
482 gcc_assert (bit_offset % BITS_PER_UNIT == 0);
484 *delta += bit_offset / BITS_PER_UNIT;
487 *base = unshare_expr (ref);
491 ar_data.delta = delta;
492 return for_each_index (base, idx_analyze_ref, &ar_data);
495 /* Record a memory reference REF to the list REFS. The reference occurs in
496 LOOP in statement STMT and it is write if WRITE_P. Returns true if the
497 reference was recorded, false otherwise. */
500 gather_memory_references_ref (struct loop *loop, struct mem_ref_group **refs,
501 tree ref, bool write_p, gimple stmt)
505 struct mem_ref_group *agrp;
507 if (get_base_address (ref) == NULL)
510 if (!analyze_ref (loop, &ref, &base, &step, &delta, stmt))
512 /* If analyze_ref fails the default is a NULL_TREE. We can stop here. */
513 if (step == NULL_TREE)
516 /* Now we know that REF = &BASE + STEP * iter + DELTA, where DELTA and STEP
517 are integer constants. */
518 agrp = find_or_create_group (refs, base, step);
519 record_ref (agrp, stmt, ref, delta, write_p);
524 /* Record the suitable memory references in LOOP. NO_OTHER_REFS is set to
525 true if there are no other memory references inside the loop. */
527 static struct mem_ref_group *
528 gather_memory_references (struct loop *loop, bool *no_other_refs, unsigned *ref_count)
530 basic_block *body = get_loop_body_in_dom_order (loop);
533 gimple_stmt_iterator bsi;
536 struct mem_ref_group *refs = NULL;
538 *no_other_refs = true;
541 /* Scan the loop body in order, so that the former references precede the
543 for (i = 0; i < loop->num_nodes; i++)
546 if (bb->loop_father != loop)
549 for (bsi = gsi_start_bb (bb); !gsi_end_p (bsi); gsi_next (&bsi))
551 stmt = gsi_stmt (bsi);
553 if (gimple_code (stmt) != GIMPLE_ASSIGN)
555 if (gimple_vuse (stmt)
556 || (is_gimple_call (stmt)
557 && !(gimple_call_flags (stmt) & ECF_CONST)))
558 *no_other_refs = false;
562 lhs = gimple_assign_lhs (stmt);
563 rhs = gimple_assign_rhs1 (stmt);
565 if (REFERENCE_CLASS_P (rhs))
567 *no_other_refs &= gather_memory_references_ref (loop, &refs,
571 if (REFERENCE_CLASS_P (lhs))
573 *no_other_refs &= gather_memory_references_ref (loop, &refs,
584 /* Prune the prefetch candidate REF using the self-reuse. */
587 prune_ref_by_self_reuse (struct mem_ref *ref)
592 /* If the step size is non constant, we cannot calculate prefetch_mod. */
593 if (!cst_and_fits_in_hwi (ref->group->step))
596 step = int_cst_value (ref->group->step);
602 /* Prefetch references to invariant address just once. */
603 ref->prefetch_before = 1;
610 if (step > PREFETCH_BLOCK)
613 if ((backward && HAVE_BACKWARD_PREFETCH)
614 || (!backward && HAVE_FORWARD_PREFETCH))
616 ref->prefetch_before = 1;
620 ref->prefetch_mod = PREFETCH_BLOCK / step;
623 /* Divides X by BY, rounding down. */
626 ddown (HOST_WIDE_INT x, unsigned HOST_WIDE_INT by)
633 return (x + by - 1) / by;
636 /* Given a CACHE_LINE_SIZE and two inductive memory references
637 with a common STEP greater than CACHE_LINE_SIZE and an address
638 difference DELTA, compute the probability that they will fall
639 in different cache lines. DISTINCT_ITERS is the number of
640 distinct iterations after which the pattern repeats itself.
641 ALIGN_UNIT is the unit of alignment in bytes. */
644 compute_miss_rate (unsigned HOST_WIDE_INT cache_line_size,
645 HOST_WIDE_INT step, HOST_WIDE_INT delta,
646 unsigned HOST_WIDE_INT distinct_iters,
649 unsigned align, iter;
650 int total_positions, miss_positions, miss_rate;
651 int address1, address2, cache_line1, cache_line2;
656 /* Iterate through all possible alignments of the first
657 memory reference within its cache line. */
658 for (align = 0; align < cache_line_size; align += align_unit)
660 /* Iterate through all distinct iterations. */
661 for (iter = 0; iter < distinct_iters; iter++)
663 address1 = align + step * iter;
664 address2 = address1 + delta;
665 cache_line1 = address1 / cache_line_size;
666 cache_line2 = address2 / cache_line_size;
667 total_positions += 1;
668 if (cache_line1 != cache_line2)
671 miss_rate = 1000 * miss_positions / total_positions;
675 /* Prune the prefetch candidate REF using the reuse with BY.
676 If BY_IS_BEFORE is true, BY is before REF in the loop. */
679 prune_ref_by_group_reuse (struct mem_ref *ref, struct mem_ref *by,
684 HOST_WIDE_INT delta_r = ref->delta, delta_b = by->delta;
685 HOST_WIDE_INT delta = delta_b - delta_r;
686 HOST_WIDE_INT hit_from;
687 unsigned HOST_WIDE_INT prefetch_before, prefetch_block;
689 HOST_WIDE_INT reduced_step;
690 unsigned HOST_WIDE_INT reduced_prefetch_block;
694 /* If the step is non constant we cannot calculate prefetch_before. */
695 if (!cst_and_fits_in_hwi (ref->group->step)) {
699 step = int_cst_value (ref->group->step);
706 /* If the references has the same address, only prefetch the
709 ref->prefetch_before = 0;
716 /* If the reference addresses are invariant and fall into the
717 same cache line, prefetch just the first one. */
721 if (ddown (ref->delta, PREFETCH_BLOCK)
722 != ddown (by->delta, PREFETCH_BLOCK))
725 ref->prefetch_before = 0;
729 /* Only prune the reference that is behind in the array. */
735 /* Transform the data so that we may assume that the accesses
739 delta_r = PREFETCH_BLOCK - 1 - delta_r;
740 delta_b = PREFETCH_BLOCK - 1 - delta_b;
748 /* Check whether the two references are likely to hit the same cache
749 line, and how distant the iterations in that it occurs are from
752 if (step <= PREFETCH_BLOCK)
754 /* The accesses are sure to meet. Let us check when. */
755 hit_from = ddown (delta_b, PREFETCH_BLOCK) * PREFETCH_BLOCK;
756 prefetch_before = (hit_from - delta_r + step - 1) / step;
758 /* Do not reduce prefetch_before if we meet beyond cache size. */
759 if (prefetch_before > (unsigned) abs (L2_CACHE_SIZE_BYTES / step))
760 prefetch_before = PREFETCH_ALL;
761 if (prefetch_before < ref->prefetch_before)
762 ref->prefetch_before = prefetch_before;
767 /* A more complicated case with step > prefetch_block. First reduce
768 the ratio between the step and the cache line size to its simplest
769 terms. The resulting denominator will then represent the number of
770 distinct iterations after which each address will go back to its
771 initial location within the cache line. This computation assumes
772 that PREFETCH_BLOCK is a power of two. */
773 prefetch_block = PREFETCH_BLOCK;
774 reduced_prefetch_block = prefetch_block;
776 while ((reduced_step & 1) == 0
777 && reduced_prefetch_block > 1)
780 reduced_prefetch_block >>= 1;
783 prefetch_before = delta / step;
785 ref_type = TREE_TYPE (ref->mem);
786 align_unit = TYPE_ALIGN (ref_type) / 8;
787 miss_rate = compute_miss_rate(prefetch_block, step, delta,
788 reduced_prefetch_block, align_unit);
789 if (miss_rate <= ACCEPTABLE_MISS_RATE)
791 /* Do not reduce prefetch_before if we meet beyond cache size. */
792 if (prefetch_before > L2_CACHE_SIZE_BYTES / PREFETCH_BLOCK)
793 prefetch_before = PREFETCH_ALL;
794 if (prefetch_before < ref->prefetch_before)
795 ref->prefetch_before = prefetch_before;
800 /* Try also the following iteration. */
802 delta = step - delta;
803 miss_rate = compute_miss_rate(prefetch_block, step, delta,
804 reduced_prefetch_block, align_unit);
805 if (miss_rate <= ACCEPTABLE_MISS_RATE)
807 if (prefetch_before < ref->prefetch_before)
808 ref->prefetch_before = prefetch_before;
813 /* The ref probably does not reuse by. */
817 /* Prune the prefetch candidate REF using the reuses with other references
821 prune_ref_by_reuse (struct mem_ref *ref, struct mem_ref *refs)
823 struct mem_ref *prune_by;
826 prune_ref_by_self_reuse (ref);
828 for (prune_by = refs; prune_by; prune_by = prune_by->next)
836 if (!WRITE_CAN_USE_READ_PREFETCH
838 && !prune_by->write_p)
840 if (!READ_CAN_USE_WRITE_PREFETCH
842 && prune_by->write_p)
845 prune_ref_by_group_reuse (ref, prune_by, before);
849 /* Prune the prefetch candidates in GROUP using the reuse analysis. */
852 prune_group_by_reuse (struct mem_ref_group *group)
854 struct mem_ref *ref_pruned;
856 for (ref_pruned = group->refs; ref_pruned; ref_pruned = ref_pruned->next)
858 prune_ref_by_reuse (ref_pruned, group->refs);
860 if (dump_file && (dump_flags & TDF_DETAILS))
862 fprintf (dump_file, "Reference %p:", (void *) ref_pruned);
864 if (ref_pruned->prefetch_before == PREFETCH_ALL
865 && ref_pruned->prefetch_mod == 1)
866 fprintf (dump_file, " no restrictions");
867 else if (ref_pruned->prefetch_before == 0)
868 fprintf (dump_file, " do not prefetch");
869 else if (ref_pruned->prefetch_before <= ref_pruned->prefetch_mod)
870 fprintf (dump_file, " prefetch once");
873 if (ref_pruned->prefetch_before != PREFETCH_ALL)
875 fprintf (dump_file, " prefetch before ");
876 fprintf (dump_file, HOST_WIDE_INT_PRINT_DEC,
877 ref_pruned->prefetch_before);
879 if (ref_pruned->prefetch_mod != 1)
881 fprintf (dump_file, " prefetch mod ");
882 fprintf (dump_file, HOST_WIDE_INT_PRINT_DEC,
883 ref_pruned->prefetch_mod);
886 fprintf (dump_file, "\n");
891 /* Prune the list of prefetch candidates GROUPS using the reuse analysis. */
894 prune_by_reuse (struct mem_ref_group *groups)
896 for (; groups; groups = groups->next)
897 prune_group_by_reuse (groups);
900 /* Returns true if we should issue prefetch for REF. */
903 should_issue_prefetch_p (struct mem_ref *ref)
905 /* For now do not issue prefetches for only first few of the
907 if (ref->prefetch_before != PREFETCH_ALL)
909 if (dump_file && (dump_flags & TDF_DETAILS))
910 fprintf (dump_file, "Ignoring %p due to prefetch_before\n",
915 /* Do not prefetch nontemporal stores. */
918 if (dump_file && (dump_flags & TDF_DETAILS))
919 fprintf (dump_file, "Ignoring nontemporal store %p\n", (void *) ref);
926 /* Decide which of the prefetch candidates in GROUPS to prefetch.
927 AHEAD is the number of iterations to prefetch ahead (which corresponds
928 to the number of simultaneous instances of one prefetch running at a
929 time). UNROLL_FACTOR is the factor by that the loop is going to be
930 unrolled. Returns true if there is anything to prefetch. */
933 schedule_prefetches (struct mem_ref_group *groups, unsigned unroll_factor,
936 unsigned remaining_prefetch_slots, n_prefetches, prefetch_slots;
937 unsigned slots_per_prefetch;
941 /* At most SIMULTANEOUS_PREFETCHES should be running at the same time. */
942 remaining_prefetch_slots = SIMULTANEOUS_PREFETCHES;
944 /* The prefetch will run for AHEAD iterations of the original loop, i.e.,
945 AHEAD / UNROLL_FACTOR iterations of the unrolled loop. In each iteration,
946 it will need a prefetch slot. */
947 slots_per_prefetch = (ahead + unroll_factor / 2) / unroll_factor;
948 if (dump_file && (dump_flags & TDF_DETAILS))
949 fprintf (dump_file, "Each prefetch instruction takes %u prefetch slots.\n",
952 /* For now we just take memory references one by one and issue
953 prefetches for as many as possible. The groups are sorted
954 starting with the largest step, since the references with
955 large step are more likely to cause many cache misses. */
957 for (; groups; groups = groups->next)
958 for (ref = groups->refs; ref; ref = ref->next)
960 if (!should_issue_prefetch_p (ref))
963 /* The loop is far from being sufficiently unrolled for this
964 prefetch. Do not generate prefetch to avoid many redudant
966 if (ref->prefetch_mod / unroll_factor > PREFETCH_MOD_TO_UNROLL_FACTOR_RATIO)
969 /* If we need to prefetch the reference each PREFETCH_MOD iterations,
970 and we unroll the loop UNROLL_FACTOR times, we need to insert
971 ceil (UNROLL_FACTOR / PREFETCH_MOD) instructions in each
973 n_prefetches = ((unroll_factor + ref->prefetch_mod - 1)
974 / ref->prefetch_mod);
975 prefetch_slots = n_prefetches * slots_per_prefetch;
977 /* If more than half of the prefetches would be lost anyway, do not
978 issue the prefetch. */
979 if (2 * remaining_prefetch_slots < prefetch_slots)
982 ref->issue_prefetch_p = true;
984 if (remaining_prefetch_slots <= prefetch_slots)
986 remaining_prefetch_slots -= prefetch_slots;
993 /* Estimate the number of prefetches in the given GROUPS. */
996 estimate_prefetch_count (struct mem_ref_group *groups)
999 int prefetch_count = 0;
1001 for (; groups; groups = groups->next)
1002 for (ref = groups->refs; ref; ref = ref->next)
1003 if (should_issue_prefetch_p (ref))
1006 return prefetch_count;
1009 /* Issue prefetches for the reference REF into loop as decided before.
1010 HEAD is the number of iterations to prefetch ahead. UNROLL_FACTOR
1011 is the factor by which LOOP was unrolled. */
1014 issue_prefetch_ref (struct mem_ref *ref, unsigned unroll_factor, unsigned ahead)
1016 HOST_WIDE_INT delta;
1017 tree addr, addr_base, write_p, local, forward;
1019 gimple_stmt_iterator bsi;
1020 unsigned n_prefetches, ap;
1021 bool nontemporal = ref->reuse_distance >= L2_CACHE_SIZE_BYTES;
1023 if (dump_file && (dump_flags & TDF_DETAILS))
1024 fprintf (dump_file, "Issued%s prefetch for %p.\n",
1025 nontemporal ? " nontemporal" : "",
1028 bsi = gsi_for_stmt (ref->stmt);
1030 n_prefetches = ((unroll_factor + ref->prefetch_mod - 1)
1031 / ref->prefetch_mod);
1032 addr_base = build_fold_addr_expr_with_type (ref->mem, ptr_type_node);
1033 addr_base = force_gimple_operand_gsi (&bsi, unshare_expr (addr_base),
1034 true, NULL, true, GSI_SAME_STMT);
1035 write_p = ref->write_p ? integer_one_node : integer_zero_node;
1036 local = build_int_cst (integer_type_node, nontemporal ? 0 : 3);
1038 for (ap = 0; ap < n_prefetches; ap++)
1040 if (cst_and_fits_in_hwi (ref->group->step))
1042 /* Determine the address to prefetch. */
1043 delta = (ahead + ap * ref->prefetch_mod) *
1044 int_cst_value (ref->group->step);
1045 addr = fold_build2 (POINTER_PLUS_EXPR, ptr_type_node,
1046 addr_base, size_int (delta));
1047 addr = force_gimple_operand_gsi (&bsi, unshare_expr (addr), true, NULL,
1048 true, GSI_SAME_STMT);
1052 /* The step size is non-constant but loop-invariant. We use the
1053 heuristic to simply prefetch ahead iterations ahead. */
1054 forward = fold_build2 (MULT_EXPR, sizetype,
1055 fold_convert (sizetype, ref->group->step),
1056 fold_convert (sizetype, size_int (ahead)));
1057 addr = fold_build2 (POINTER_PLUS_EXPR, ptr_type_node, addr_base,
1059 addr = force_gimple_operand_gsi (&bsi, unshare_expr (addr), true,
1060 NULL, true, GSI_SAME_STMT);
1062 /* Create the prefetch instruction. */
1063 prefetch = gimple_build_call (built_in_decls[BUILT_IN_PREFETCH],
1064 3, addr, write_p, local);
1065 gsi_insert_before (&bsi, prefetch, GSI_SAME_STMT);
1069 /* Issue prefetches for the references in GROUPS into loop as decided before.
1070 HEAD is the number of iterations to prefetch ahead. UNROLL_FACTOR is the
1071 factor by that LOOP was unrolled. */
1074 issue_prefetches (struct mem_ref_group *groups,
1075 unsigned unroll_factor, unsigned ahead)
1077 struct mem_ref *ref;
1079 for (; groups; groups = groups->next)
1080 for (ref = groups->refs; ref; ref = ref->next)
1081 if (ref->issue_prefetch_p)
1082 issue_prefetch_ref (ref, unroll_factor, ahead);
1085 /* Returns true if REF is a memory write for that a nontemporal store insn
1089 nontemporal_store_p (struct mem_ref *ref)
1091 enum machine_mode mode;
1092 enum insn_code code;
1094 /* REF must be a write that is not reused. We require it to be independent
1095 on all other memory references in the loop, as the nontemporal stores may
1096 be reordered with respect to other memory references. */
1098 || !ref->independent_p
1099 || ref->reuse_distance < L2_CACHE_SIZE_BYTES)
1102 /* Check that we have the storent instruction for the mode. */
1103 mode = TYPE_MODE (TREE_TYPE (ref->mem));
1104 if (mode == BLKmode)
1107 code = optab_handler (storent_optab, mode)->insn_code;
1108 return code != CODE_FOR_nothing;
1111 /* If REF is a nontemporal store, we mark the corresponding modify statement
1112 and return true. Otherwise, we return false. */
1115 mark_nontemporal_store (struct mem_ref *ref)
1117 if (!nontemporal_store_p (ref))
1120 if (dump_file && (dump_flags & TDF_DETAILS))
1121 fprintf (dump_file, "Marked reference %p as a nontemporal store.\n",
1124 gimple_assign_set_nontemporal_move (ref->stmt, true);
1125 ref->storent_p = true;
1130 /* Issue a memory fence instruction after LOOP. */
1133 emit_mfence_after_loop (struct loop *loop)
1135 VEC (edge, heap) *exits = get_loop_exit_edges (loop);
1138 gimple_stmt_iterator bsi;
1141 for (i = 0; VEC_iterate (edge, exits, i, exit); i++)
1143 call = gimple_build_call (FENCE_FOLLOWING_MOVNT, 0);
1145 if (!single_pred_p (exit->dest)
1146 /* If possible, we prefer not to insert the fence on other paths
1148 && !(exit->flags & EDGE_ABNORMAL))
1149 split_loop_exit_edge (exit);
1150 bsi = gsi_after_labels (exit->dest);
1152 gsi_insert_before (&bsi, call, GSI_NEW_STMT);
1153 mark_virtual_ops_for_renaming (call);
1156 VEC_free (edge, heap, exits);
1157 update_ssa (TODO_update_ssa_only_virtuals);
1160 /* Returns true if we can use storent in loop, false otherwise. */
1163 may_use_storent_in_loop_p (struct loop *loop)
1167 if (loop->inner != NULL)
1170 /* If we must issue a mfence insn after using storent, check that there
1171 is a suitable place for it at each of the loop exits. */
1172 if (FENCE_FOLLOWING_MOVNT != NULL_TREE)
1174 VEC (edge, heap) *exits = get_loop_exit_edges (loop);
1178 for (i = 0; VEC_iterate (edge, exits, i, exit); i++)
1179 if ((exit->flags & EDGE_ABNORMAL)
1180 && exit->dest == EXIT_BLOCK_PTR)
1183 VEC_free (edge, heap, exits);
1189 /* Marks nontemporal stores in LOOP. GROUPS contains the description of memory
1190 references in the loop. */
1193 mark_nontemporal_stores (struct loop *loop, struct mem_ref_group *groups)
1195 struct mem_ref *ref;
1198 if (!may_use_storent_in_loop_p (loop))
1201 for (; groups; groups = groups->next)
1202 for (ref = groups->refs; ref; ref = ref->next)
1203 any |= mark_nontemporal_store (ref);
1205 if (any && FENCE_FOLLOWING_MOVNT != NULL_TREE)
1206 emit_mfence_after_loop (loop);
1209 /* Determines whether we can profitably unroll LOOP FACTOR times, and if
1210 this is the case, fill in DESC by the description of number of
1214 should_unroll_loop_p (struct loop *loop, struct tree_niter_desc *desc,
1217 if (!can_unroll_loop_p (loop, factor, desc))
1220 /* We only consider loops without control flow for unrolling. This is not
1221 a hard restriction -- tree_unroll_loop works with arbitrary loops
1222 as well; but the unrolling/prefetching is usually more profitable for
1223 loops consisting of a single basic block, and we want to limit the
1225 if (loop->num_nodes > 2)
1231 /* Determine the coefficient by that unroll LOOP, from the information
1232 contained in the list of memory references REFS. Description of
1233 umber of iterations of LOOP is stored to DESC. NINSNS is the number of
1234 insns of the LOOP. EST_NITER is the estimated number of iterations of
1235 the loop, or -1 if no estimate is available. */
1238 determine_unroll_factor (struct loop *loop, struct mem_ref_group *refs,
1239 unsigned ninsns, struct tree_niter_desc *desc,
1240 HOST_WIDE_INT est_niter)
1242 unsigned upper_bound;
1243 unsigned nfactor, factor, mod_constraint;
1244 struct mem_ref_group *agp;
1245 struct mem_ref *ref;
1247 /* First check whether the loop is not too large to unroll. We ignore
1248 PARAM_MAX_UNROLL_TIMES, because for small loops, it prevented us
1249 from unrolling them enough to make exactly one cache line covered by each
1250 iteration. Also, the goal of PARAM_MAX_UNROLL_TIMES is to prevent
1251 us from unrolling the loops too many times in cases where we only expect
1252 gains from better scheduling and decreasing loop overhead, which is not
1254 upper_bound = PARAM_VALUE (PARAM_MAX_UNROLLED_INSNS) / ninsns;
1256 /* If we unrolled the loop more times than it iterates, the unrolled version
1257 of the loop would be never entered. */
1258 if (est_niter >= 0 && est_niter < (HOST_WIDE_INT) upper_bound)
1259 upper_bound = est_niter;
1261 if (upper_bound <= 1)
1264 /* Choose the factor so that we may prefetch each cache just once,
1265 but bound the unrolling by UPPER_BOUND. */
1267 for (agp = refs; agp; agp = agp->next)
1268 for (ref = agp->refs; ref; ref = ref->next)
1269 if (should_issue_prefetch_p (ref))
1271 mod_constraint = ref->prefetch_mod;
1272 nfactor = least_common_multiple (mod_constraint, factor);
1273 if (nfactor <= upper_bound)
1277 if (!should_unroll_loop_p (loop, desc, factor))
1283 /* Returns the total volume of the memory references REFS, taking into account
1284 reuses in the innermost loop and cache line size. TODO -- we should also
1285 take into account reuses across the iterations of the loops in the loop
1289 volume_of_references (struct mem_ref_group *refs)
1291 unsigned volume = 0;
1292 struct mem_ref_group *gr;
1293 struct mem_ref *ref;
1295 for (gr = refs; gr; gr = gr->next)
1296 for (ref = gr->refs; ref; ref = ref->next)
1298 /* Almost always reuses another value? */
1299 if (ref->prefetch_before != PREFETCH_ALL)
1302 /* If several iterations access the same cache line, use the size of
1303 the line divided by this number. Otherwise, a cache line is
1304 accessed in each iteration. TODO -- in the latter case, we should
1305 take the size of the reference into account, rounding it up on cache
1306 line size multiple. */
1307 volume += L1_CACHE_LINE_SIZE / ref->prefetch_mod;
1312 /* Returns the volume of memory references accessed across VEC iterations of
1313 loops, whose sizes are described in the LOOP_SIZES array. N is the number
1314 of the loops in the nest (length of VEC and LOOP_SIZES vectors). */
1317 volume_of_dist_vector (lambda_vector vec, unsigned *loop_sizes, unsigned n)
1321 for (i = 0; i < n; i++)
1328 gcc_assert (vec[i] > 0);
1330 /* We ignore the parts of the distance vector in subloops, since usually
1331 the numbers of iterations are much smaller. */
1332 return loop_sizes[i] * vec[i];
1335 /* Add the steps of ACCESS_FN multiplied by STRIDE to the array STRIDE
1336 at the position corresponding to the loop of the step. N is the depth
1337 of the considered loop nest, and, LOOP is its innermost loop. */
1340 add_subscript_strides (tree access_fn, unsigned stride,
1341 HOST_WIDE_INT *strides, unsigned n, struct loop *loop)
1345 HOST_WIDE_INT astep;
1346 unsigned min_depth = loop_depth (loop) - n;
1348 while (TREE_CODE (access_fn) == POLYNOMIAL_CHREC)
1350 aloop = get_chrec_loop (access_fn);
1351 step = CHREC_RIGHT (access_fn);
1352 access_fn = CHREC_LEFT (access_fn);
1354 if ((unsigned) loop_depth (aloop) <= min_depth)
1357 if (host_integerp (step, 0))
1358 astep = tree_low_cst (step, 0);
1360 astep = L1_CACHE_LINE_SIZE;
1362 strides[n - 1 - loop_depth (loop) + loop_depth (aloop)] += astep * stride;
1367 /* Returns the volume of memory references accessed between two consecutive
1368 self-reuses of the reference DR. We consider the subscripts of DR in N
1369 loops, and LOOP_SIZES contains the volumes of accesses in each of the
1370 loops. LOOP is the innermost loop of the current loop nest. */
1373 self_reuse_distance (data_reference_p dr, unsigned *loop_sizes, unsigned n,
1376 tree stride, access_fn;
1377 HOST_WIDE_INT *strides, astride;
1378 VEC (tree, heap) *access_fns;
1379 tree ref = DR_REF (dr);
1380 unsigned i, ret = ~0u;
1382 /* In the following example:
1384 for (i = 0; i < N; i++)
1385 for (j = 0; j < N; j++)
1387 the same cache line is accessed each N steps (except if the change from
1388 i to i + 1 crosses the boundary of the cache line). Thus, for self-reuse,
1389 we cannot rely purely on the results of the data dependence analysis.
1391 Instead, we compute the stride of the reference in each loop, and consider
1392 the innermost loop in that the stride is less than cache size. */
1394 strides = XCNEWVEC (HOST_WIDE_INT, n);
1395 access_fns = DR_ACCESS_FNS (dr);
1397 for (i = 0; VEC_iterate (tree, access_fns, i, access_fn); i++)
1399 /* Keep track of the reference corresponding to the subscript, so that we
1401 while (handled_component_p (ref) && TREE_CODE (ref) != ARRAY_REF)
1402 ref = TREE_OPERAND (ref, 0);
1404 if (TREE_CODE (ref) == ARRAY_REF)
1406 stride = TYPE_SIZE_UNIT (TREE_TYPE (ref));
1407 if (host_integerp (stride, 1))
1408 astride = tree_low_cst (stride, 1);
1410 astride = L1_CACHE_LINE_SIZE;
1412 ref = TREE_OPERAND (ref, 0);
1417 add_subscript_strides (access_fn, astride, strides, n, loop);
1420 for (i = n; i-- > 0; )
1422 unsigned HOST_WIDE_INT s;
1424 s = strides[i] < 0 ? -strides[i] : strides[i];
1426 if (s < (unsigned) L1_CACHE_LINE_SIZE
1428 > (unsigned) (L1_CACHE_SIZE_BYTES / NONTEMPORAL_FRACTION)))
1430 ret = loop_sizes[i];
1439 /* Determines the distance till the first reuse of each reference in REFS
1440 in the loop nest of LOOP. NO_OTHER_REFS is true if there are no other
1441 memory references in the loop. */
1444 determine_loop_nest_reuse (struct loop *loop, struct mem_ref_group *refs,
1447 struct loop *nest, *aloop;
1448 VEC (data_reference_p, heap) *datarefs = NULL;
1449 VEC (ddr_p, heap) *dependences = NULL;
1450 struct mem_ref_group *gr;
1451 struct mem_ref *ref, *refb;
1452 VEC (loop_p, heap) *vloops = NULL;
1453 unsigned *loop_data_size;
1455 unsigned volume, dist, adist;
1457 data_reference_p dr;
1463 /* Find the outermost loop of the loop nest of loop (we require that
1464 there are no sibling loops inside the nest). */
1468 aloop = loop_outer (nest);
1470 if (aloop == current_loops->tree_root
1471 || aloop->inner->next)
1477 /* For each loop, determine the amount of data accessed in each iteration.
1478 We use this to estimate whether the reference is evicted from the
1479 cache before its reuse. */
1480 find_loop_nest (nest, &vloops);
1481 n = VEC_length (loop_p, vloops);
1482 loop_data_size = XNEWVEC (unsigned, n);
1483 volume = volume_of_references (refs);
1487 loop_data_size[i] = volume;
1488 /* Bound the volume by the L2 cache size, since above this bound,
1489 all dependence distances are equivalent. */
1490 if (volume > L2_CACHE_SIZE_BYTES)
1493 aloop = VEC_index (loop_p, vloops, i);
1494 vol = estimated_loop_iterations_int (aloop, false);
1496 vol = expected_loop_iterations (aloop);
1500 /* Prepare the references in the form suitable for data dependence
1501 analysis. We ignore unanalyzable data references (the results
1502 are used just as a heuristics to estimate temporality of the
1503 references, hence we do not need to worry about correctness). */
1504 for (gr = refs; gr; gr = gr->next)
1505 for (ref = gr->refs; ref; ref = ref->next)
1507 dr = create_data_ref (nest, ref->mem, ref->stmt, !ref->write_p);
1511 ref->reuse_distance = volume;
1513 VEC_safe_push (data_reference_p, heap, datarefs, dr);
1516 no_other_refs = false;
1519 for (i = 0; VEC_iterate (data_reference_p, datarefs, i, dr); i++)
1521 dist = self_reuse_distance (dr, loop_data_size, n, loop);
1522 ref = (struct mem_ref *) dr->aux;
1523 if (ref->reuse_distance > dist)
1524 ref->reuse_distance = dist;
1527 ref->independent_p = true;
1530 compute_all_dependences (datarefs, &dependences, vloops, true);
1532 for (i = 0; VEC_iterate (ddr_p, dependences, i, dep); i++)
1534 if (DDR_ARE_DEPENDENT (dep) == chrec_known)
1537 ref = (struct mem_ref *) DDR_A (dep)->aux;
1538 refb = (struct mem_ref *) DDR_B (dep)->aux;
1540 if (DDR_ARE_DEPENDENT (dep) == chrec_dont_know
1541 || DDR_NUM_DIST_VECTS (dep) == 0)
1543 /* If the dependence cannot be analyzed, assume that there might be
1547 ref->independent_p = false;
1548 refb->independent_p = false;
1552 /* The distance vectors are normalized to be always lexicographically
1553 positive, hence we cannot tell just from them whether DDR_A comes
1554 before DDR_B or vice versa. However, it is not important,
1555 anyway -- if DDR_A is close to DDR_B, then it is either reused in
1556 DDR_B (and it is not nontemporal), or it reuses the value of DDR_B
1557 in cache (and marking it as nontemporal would not affect
1561 for (j = 0; j < DDR_NUM_DIST_VECTS (dep); j++)
1563 adist = volume_of_dist_vector (DDR_DIST_VECT (dep, j),
1566 /* If this is a dependence in the innermost loop (i.e., the
1567 distances in all superloops are zero) and it is not
1568 the trivial self-dependence with distance zero, record that
1569 the references are not completely independent. */
1570 if (lambda_vector_zerop (DDR_DIST_VECT (dep, j), n - 1)
1572 || DDR_DIST_VECT (dep, j)[n-1] != 0))
1574 ref->independent_p = false;
1575 refb->independent_p = false;
1578 /* Ignore accesses closer than
1579 L1_CACHE_SIZE_BYTES / NONTEMPORAL_FRACTION,
1580 so that we use nontemporal prefetches e.g. if single memory
1581 location is accessed several times in a single iteration of
1583 if (adist < L1_CACHE_SIZE_BYTES / NONTEMPORAL_FRACTION)
1591 if (ref->reuse_distance > dist)
1592 ref->reuse_distance = dist;
1593 if (refb->reuse_distance > dist)
1594 refb->reuse_distance = dist;
1597 free_dependence_relations (dependences);
1598 free_data_refs (datarefs);
1599 free (loop_data_size);
1601 if (dump_file && (dump_flags & TDF_DETAILS))
1603 fprintf (dump_file, "Reuse distances:\n");
1604 for (gr = refs; gr; gr = gr->next)
1605 for (ref = gr->refs; ref; ref = ref->next)
1606 fprintf (dump_file, " ref %p distance %u\n",
1607 (void *) ref, ref->reuse_distance);
1611 /* Do a cost-benefit analysis to determine if prefetching is profitable
1612 for the current loop given the following parameters:
1613 AHEAD: the iteration ahead distance,
1614 EST_NITER: the estimated trip count,
1615 NINSNS: estimated number of instructions in the loop,
1616 PREFETCH_COUNT: an estimate of the number of prefetches
1617 MEM_REF_COUNT: total number of memory references in the loop. */
1620 is_loop_prefetching_profitable (unsigned ahead, HOST_WIDE_INT est_niter,
1621 unsigned ninsns, unsigned prefetch_count,
1622 unsigned mem_ref_count, unsigned unroll_factor)
1624 int insn_to_mem_ratio, insn_to_prefetch_ratio;
1626 if (mem_ref_count == 0)
1629 /* Prefetching improves performance by overlapping cache missing
1630 memory accesses with CPU operations. If the loop does not have
1631 enough CPU operations to overlap with memory operations, prefetching
1632 won't give a significant benefit. One approximate way of checking
1633 this is to require the ratio of instructions to memory references to
1634 be above a certain limit. This approximation works well in practice.
1635 TODO: Implement a more precise computation by estimating the time
1636 for each CPU or memory op in the loop. Time estimates for memory ops
1637 should account for cache misses. */
1638 insn_to_mem_ratio = ninsns / mem_ref_count;
1640 if (insn_to_mem_ratio < PREFETCH_MIN_INSN_TO_MEM_RATIO)
1642 if (dump_file && (dump_flags & TDF_DETAILS))
1644 "Not prefetching -- instruction to memory reference ratio (%d) too small\n",
1649 /* Prefetching most likely causes performance degradation when the instruction
1650 to prefetch ratio is too small. Too many prefetch instructions in a loop
1651 may reduce the I-cache performance.
1652 (unroll_factor * ninsns) is used to estimate the number of instructions in
1653 the unrolled loop. This implementation is a bit simplistic -- the number
1654 of issued prefetch instructions is also affected by unrolling. So,
1655 prefetch_mod and the unroll factor should be taken into account when
1656 determining prefetch_count. Also, the number of insns of the unrolled
1657 loop will usually be significantly smaller than the number of insns of the
1658 original loop * unroll_factor (at least the induction variable increases
1659 and the exit branches will get eliminated), so it might be better to use
1660 tree_estimate_loop_size + estimated_unrolled_size. */
1661 insn_to_prefetch_ratio = (unroll_factor * ninsns) / prefetch_count;
1662 if (insn_to_prefetch_ratio < MIN_INSN_TO_PREFETCH_RATIO)
1664 if (dump_file && (dump_flags & TDF_DETAILS))
1666 "Not prefetching -- instruction to prefetch ratio (%d) too small\n",
1667 insn_to_prefetch_ratio);
1671 /* Could not do further estimation if the trip count is unknown. Just assume
1672 prefetching is profitable. Too aggressive??? */
1676 if (est_niter < (HOST_WIDE_INT) (TRIP_COUNT_TO_AHEAD_RATIO * ahead))
1678 if (dump_file && (dump_flags & TDF_DETAILS))
1680 "Not prefetching -- loop estimated to roll only %d times\n",
1688 /* Issue prefetch instructions for array references in LOOP. Returns
1689 true if the LOOP was unrolled. */
1692 loop_prefetch_arrays (struct loop *loop)
1694 struct mem_ref_group *refs;
1695 unsigned ahead, ninsns, time, unroll_factor;
1696 HOST_WIDE_INT est_niter;
1697 struct tree_niter_desc desc;
1698 bool unrolled = false, no_other_refs;
1699 unsigned prefetch_count;
1700 unsigned mem_ref_count;
1702 if (optimize_loop_nest_for_size_p (loop))
1704 if (dump_file && (dump_flags & TDF_DETAILS))
1705 fprintf (dump_file, " ignored (cold area)\n");
1709 /* Step 1: gather the memory references. */
1710 refs = gather_memory_references (loop, &no_other_refs, &mem_ref_count);
1712 /* Step 2: estimate the reuse effects. */
1713 prune_by_reuse (refs);
1715 prefetch_count = estimate_prefetch_count (refs);
1716 if (prefetch_count == 0)
1719 determine_loop_nest_reuse (loop, refs, no_other_refs);
1721 /* Step 3: determine the ahead and unroll factor. */
1723 /* FIXME: the time should be weighted by the probabilities of the blocks in
1725 time = tree_num_loop_insns (loop, &eni_time_weights);
1726 ahead = (PREFETCH_LATENCY + time - 1) / time;
1727 est_niter = estimated_loop_iterations_int (loop, false);
1729 ninsns = tree_num_loop_insns (loop, &eni_size_weights);
1730 unroll_factor = determine_unroll_factor (loop, refs, ninsns, &desc,
1732 if (dump_file && (dump_flags & TDF_DETAILS))
1733 fprintf (dump_file, "Ahead %d, unroll factor %d, trip count "
1734 HOST_WIDE_INT_PRINT_DEC "\n"
1735 "insn count %d, mem ref count %d, prefetch count %d\n",
1736 ahead, unroll_factor, est_niter,
1737 ninsns, mem_ref_count, prefetch_count);
1739 if (!is_loop_prefetching_profitable (ahead, est_niter, ninsns, prefetch_count,
1740 mem_ref_count, unroll_factor))
1743 mark_nontemporal_stores (loop, refs);
1745 /* Step 4: what to prefetch? */
1746 if (!schedule_prefetches (refs, unroll_factor, ahead))
1749 /* Step 5: unroll the loop. TODO -- peeling of first and last few
1750 iterations so that we do not issue superfluous prefetches. */
1751 if (unroll_factor != 1)
1753 tree_unroll_loop (loop, unroll_factor,
1754 single_dom_exit (loop), &desc);
1758 /* Step 6: issue the prefetches. */
1759 issue_prefetches (refs, unroll_factor, ahead);
1762 release_mem_refs (refs);
1766 /* Issue prefetch instructions for array references in loops. */
1769 tree_ssa_prefetch_arrays (void)
1773 bool unrolled = false;
1777 /* It is possible to ask compiler for say -mtune=i486 -march=pentium4.
1778 -mtune=i486 causes us having PREFETCH_BLOCK 0, since this is part
1779 of processor costs and i486 does not have prefetch, but
1780 -march=pentium4 causes HAVE_prefetch to be true. Ugh. */
1781 || PREFETCH_BLOCK == 0)
1784 if (dump_file && (dump_flags & TDF_DETAILS))
1786 fprintf (dump_file, "Prefetching parameters:\n");
1787 fprintf (dump_file, " simultaneous prefetches: %d\n",
1788 SIMULTANEOUS_PREFETCHES);
1789 fprintf (dump_file, " prefetch latency: %d\n", PREFETCH_LATENCY);
1790 fprintf (dump_file, " prefetch block size: %d\n", PREFETCH_BLOCK);
1791 fprintf (dump_file, " L1 cache size: %d lines, %d kB\n",
1792 L1_CACHE_SIZE_BYTES / L1_CACHE_LINE_SIZE, L1_CACHE_SIZE);
1793 fprintf (dump_file, " L1 cache line size: %d\n", L1_CACHE_LINE_SIZE);
1794 fprintf (dump_file, " L2 cache size: %d kB\n", L2_CACHE_SIZE);
1795 fprintf (dump_file, " min insn-to-prefetch ratio: %d \n",
1796 MIN_INSN_TO_PREFETCH_RATIO);
1797 fprintf (dump_file, " min insn-to-mem ratio: %d \n",
1798 PREFETCH_MIN_INSN_TO_MEM_RATIO);
1799 fprintf (dump_file, "\n");
1802 initialize_original_copy_tables ();
1804 if (!built_in_decls[BUILT_IN_PREFETCH])
1806 tree type = build_function_type (void_type_node,
1807 tree_cons (NULL_TREE,
1808 const_ptr_type_node,
1810 tree decl = add_builtin_function ("__builtin_prefetch", type,
1811 BUILT_IN_PREFETCH, BUILT_IN_NORMAL,
1813 DECL_IS_NOVOPS (decl) = true;
1814 built_in_decls[BUILT_IN_PREFETCH] = decl;
1817 /* We assume that size of cache line is a power of two, so verify this
1819 gcc_assert ((PREFETCH_BLOCK & (PREFETCH_BLOCK - 1)) == 0);
1821 FOR_EACH_LOOP (li, loop, LI_FROM_INNERMOST)
1823 if (dump_file && (dump_flags & TDF_DETAILS))
1824 fprintf (dump_file, "Processing loop %d:\n", loop->num);
1826 unrolled |= loop_prefetch_arrays (loop);
1828 if (dump_file && (dump_flags & TDF_DETAILS))
1829 fprintf (dump_file, "\n\n");
1835 todo_flags |= TODO_cleanup_cfg;
1838 free_original_copy_tables ();