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, when prefetch_mod is 16 and unroll_factor is
232 2, prefetching requires unrolling the loop 16 times, but
233 the loop is actually unrolled twice. In this case (ratio = 8),
234 prefetching is not likely to be beneficial. */
236 #ifndef PREFETCH_MOD_TO_UNROLL_FACTOR_RATIO
237 #define PREFETCH_MOD_TO_UNROLL_FACTOR_RATIO 4
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)
410 if (!simple_iv (ar_data->loop, loop_containing_stmt (ar_data->stmt),
416 if (TREE_CODE (ibase) == POINTER_PLUS_EXPR
417 && cst_and_fits_in_hwi (TREE_OPERAND (ibase, 1)))
419 idelta = int_cst_value (TREE_OPERAND (ibase, 1));
420 ibase = TREE_OPERAND (ibase, 0);
422 if (cst_and_fits_in_hwi (ibase))
424 idelta += int_cst_value (ibase);
425 ibase = build_int_cst (TREE_TYPE (ibase), 0);
428 if (TREE_CODE (base) == ARRAY_REF)
430 stepsize = array_ref_element_size (base);
431 if (!cst_and_fits_in_hwi (stepsize))
433 imult = int_cst_value (stepsize);
434 step = fold_build2 (MULT_EXPR, sizetype,
435 fold_convert (sizetype, step),
436 fold_convert (sizetype, stepsize));
440 if (*ar_data->step == NULL_TREE)
441 *ar_data->step = step;
443 *ar_data->step = fold_build2 (PLUS_EXPR, sizetype,
444 fold_convert (sizetype, *ar_data->step),
445 fold_convert (sizetype, step));
446 *ar_data->delta += idelta;
452 /* Tries to express REF_P in shape &BASE + STEP * iter + DELTA, where DELTA and
453 STEP are integer constants and iter is number of iterations of LOOP. The
454 reference occurs in statement STMT. Strips nonaddressable component
455 references from REF_P. */
458 analyze_ref (struct loop *loop, tree *ref_p, tree *base,
459 tree *step, HOST_WIDE_INT *delta,
462 struct ar_data ar_data;
464 HOST_WIDE_INT bit_offset;
470 /* First strip off the component references. Ignore bitfields. */
471 if (TREE_CODE (ref) == COMPONENT_REF
472 && DECL_NONADDRESSABLE_P (TREE_OPERAND (ref, 1)))
473 ref = TREE_OPERAND (ref, 0);
477 for (; TREE_CODE (ref) == COMPONENT_REF; ref = TREE_OPERAND (ref, 0))
479 off = DECL_FIELD_BIT_OFFSET (TREE_OPERAND (ref, 1));
480 bit_offset = TREE_INT_CST_LOW (off);
481 gcc_assert (bit_offset % BITS_PER_UNIT == 0);
483 *delta += bit_offset / BITS_PER_UNIT;
486 *base = unshare_expr (ref);
490 ar_data.delta = delta;
491 return for_each_index (base, idx_analyze_ref, &ar_data);
494 /* Record a memory reference REF to the list REFS. The reference occurs in
495 LOOP in statement STMT and it is write if WRITE_P. Returns true if the
496 reference was recorded, false otherwise. */
499 gather_memory_references_ref (struct loop *loop, struct mem_ref_group **refs,
500 tree ref, bool write_p, gimple stmt)
504 struct mem_ref_group *agrp;
506 if (get_base_address (ref) == NULL)
509 if (!analyze_ref (loop, &ref, &base, &step, &delta, stmt))
511 /* If analyze_ref fails the default is a NULL_TREE. We can stop here. */
512 if (step == NULL_TREE)
515 /* Limit non-constant step prefetching only to the innermost loops. */
516 if (!cst_and_fits_in_hwi (step) && loop->inner != NULL)
519 /* Now we know that REF = &BASE + STEP * iter + DELTA, where DELTA and STEP
520 are integer constants. */
521 agrp = find_or_create_group (refs, base, step);
522 record_ref (agrp, stmt, ref, delta, write_p);
527 /* Record the suitable memory references in LOOP. NO_OTHER_REFS is set to
528 true if there are no other memory references inside the loop. */
530 static struct mem_ref_group *
531 gather_memory_references (struct loop *loop, bool *no_other_refs, unsigned *ref_count)
533 basic_block *body = get_loop_body_in_dom_order (loop);
536 gimple_stmt_iterator bsi;
539 struct mem_ref_group *refs = NULL;
541 *no_other_refs = true;
544 /* Scan the loop body in order, so that the former references precede the
546 for (i = 0; i < loop->num_nodes; i++)
549 if (bb->loop_father != loop)
552 for (bsi = gsi_start_bb (bb); !gsi_end_p (bsi); gsi_next (&bsi))
554 stmt = gsi_stmt (bsi);
556 if (gimple_code (stmt) != GIMPLE_ASSIGN)
558 if (gimple_vuse (stmt)
559 || (is_gimple_call (stmt)
560 && !(gimple_call_flags (stmt) & ECF_CONST)))
561 *no_other_refs = false;
565 lhs = gimple_assign_lhs (stmt);
566 rhs = gimple_assign_rhs1 (stmt);
568 if (REFERENCE_CLASS_P (rhs))
570 *no_other_refs &= gather_memory_references_ref (loop, &refs,
574 if (REFERENCE_CLASS_P (lhs))
576 *no_other_refs &= gather_memory_references_ref (loop, &refs,
587 /* Prune the prefetch candidate REF using the self-reuse. */
590 prune_ref_by_self_reuse (struct mem_ref *ref)
595 /* If the step size is non constant, we cannot calculate prefetch_mod. */
596 if (!cst_and_fits_in_hwi (ref->group->step))
599 step = int_cst_value (ref->group->step);
605 /* Prefetch references to invariant address just once. */
606 ref->prefetch_before = 1;
613 if (step > PREFETCH_BLOCK)
616 if ((backward && HAVE_BACKWARD_PREFETCH)
617 || (!backward && HAVE_FORWARD_PREFETCH))
619 ref->prefetch_before = 1;
623 ref->prefetch_mod = PREFETCH_BLOCK / step;
626 /* Divides X by BY, rounding down. */
629 ddown (HOST_WIDE_INT x, unsigned HOST_WIDE_INT by)
636 return (x + by - 1) / by;
639 /* Given a CACHE_LINE_SIZE and two inductive memory references
640 with a common STEP greater than CACHE_LINE_SIZE and an address
641 difference DELTA, compute the probability that they will fall
642 in different cache lines. Return true if the computed miss rate
643 is not greater than the ACCEPTABLE_MISS_RATE. DISTINCT_ITERS is the
644 number of distinct iterations after which the pattern repeats itself.
645 ALIGN_UNIT is the unit of alignment in bytes. */
648 is_miss_rate_acceptable (unsigned HOST_WIDE_INT cache_line_size,
649 HOST_WIDE_INT step, HOST_WIDE_INT delta,
650 unsigned HOST_WIDE_INT distinct_iters,
653 unsigned align, iter;
654 int total_positions, miss_positions, max_allowed_miss_positions;
655 int address1, address2, cache_line1, cache_line2;
657 /* It always misses if delta is greater than or equal to the cache
659 if (delta >= (HOST_WIDE_INT) cache_line_size)
663 total_positions = (cache_line_size / align_unit) * distinct_iters;
664 max_allowed_miss_positions = (ACCEPTABLE_MISS_RATE * total_positions) / 1000;
666 /* Iterate through all possible alignments of the first
667 memory reference within its cache line. */
668 for (align = 0; align < cache_line_size; align += align_unit)
670 /* Iterate through all distinct iterations. */
671 for (iter = 0; iter < distinct_iters; iter++)
673 address1 = align + step * iter;
674 address2 = address1 + delta;
675 cache_line1 = address1 / cache_line_size;
676 cache_line2 = address2 / cache_line_size;
677 if (cache_line1 != cache_line2)
680 if (miss_positions > max_allowed_miss_positions)
687 /* Prune the prefetch candidate REF using the reuse with BY.
688 If BY_IS_BEFORE is true, BY is before REF in the loop. */
691 prune_ref_by_group_reuse (struct mem_ref *ref, struct mem_ref *by,
696 HOST_WIDE_INT delta_r = ref->delta, delta_b = by->delta;
697 HOST_WIDE_INT delta = delta_b - delta_r;
698 HOST_WIDE_INT hit_from;
699 unsigned HOST_WIDE_INT prefetch_before, prefetch_block;
700 HOST_WIDE_INT reduced_step;
701 unsigned HOST_WIDE_INT reduced_prefetch_block;
705 /* If the step is non constant we cannot calculate prefetch_before. */
706 if (!cst_and_fits_in_hwi (ref->group->step)) {
710 step = int_cst_value (ref->group->step);
717 /* If the references has the same address, only prefetch the
720 ref->prefetch_before = 0;
727 /* If the reference addresses are invariant and fall into the
728 same cache line, prefetch just the first one. */
732 if (ddown (ref->delta, PREFETCH_BLOCK)
733 != ddown (by->delta, PREFETCH_BLOCK))
736 ref->prefetch_before = 0;
740 /* Only prune the reference that is behind in the array. */
746 /* Transform the data so that we may assume that the accesses
750 delta_r = PREFETCH_BLOCK - 1 - delta_r;
751 delta_b = PREFETCH_BLOCK - 1 - delta_b;
759 /* Check whether the two references are likely to hit the same cache
760 line, and how distant the iterations in that it occurs are from
763 if (step <= PREFETCH_BLOCK)
765 /* The accesses are sure to meet. Let us check when. */
766 hit_from = ddown (delta_b, PREFETCH_BLOCK) * PREFETCH_BLOCK;
767 prefetch_before = (hit_from - delta_r + step - 1) / step;
769 /* Do not reduce prefetch_before if we meet beyond cache size. */
770 if (prefetch_before > (unsigned) abs (L2_CACHE_SIZE_BYTES / step))
771 prefetch_before = PREFETCH_ALL;
772 if (prefetch_before < ref->prefetch_before)
773 ref->prefetch_before = prefetch_before;
778 /* A more complicated case with step > prefetch_block. First reduce
779 the ratio between the step and the cache line size to its simplest
780 terms. The resulting denominator will then represent the number of
781 distinct iterations after which each address will go back to its
782 initial location within the cache line. This computation assumes
783 that PREFETCH_BLOCK is a power of two. */
784 prefetch_block = PREFETCH_BLOCK;
785 reduced_prefetch_block = prefetch_block;
787 while ((reduced_step & 1) == 0
788 && reduced_prefetch_block > 1)
791 reduced_prefetch_block >>= 1;
794 prefetch_before = delta / step;
796 ref_type = TREE_TYPE (ref->mem);
797 align_unit = TYPE_ALIGN (ref_type) / 8;
798 if (is_miss_rate_acceptable (prefetch_block, step, delta,
799 reduced_prefetch_block, align_unit))
801 /* Do not reduce prefetch_before if we meet beyond cache size. */
802 if (prefetch_before > L2_CACHE_SIZE_BYTES / PREFETCH_BLOCK)
803 prefetch_before = PREFETCH_ALL;
804 if (prefetch_before < ref->prefetch_before)
805 ref->prefetch_before = prefetch_before;
810 /* Try also the following iteration. */
812 delta = step - delta;
813 if (is_miss_rate_acceptable (prefetch_block, step, delta,
814 reduced_prefetch_block, align_unit))
816 if (prefetch_before < ref->prefetch_before)
817 ref->prefetch_before = prefetch_before;
822 /* The ref probably does not reuse by. */
826 /* Prune the prefetch candidate REF using the reuses with other references
830 prune_ref_by_reuse (struct mem_ref *ref, struct mem_ref *refs)
832 struct mem_ref *prune_by;
835 prune_ref_by_self_reuse (ref);
837 for (prune_by = refs; prune_by; prune_by = prune_by->next)
845 if (!WRITE_CAN_USE_READ_PREFETCH
847 && !prune_by->write_p)
849 if (!READ_CAN_USE_WRITE_PREFETCH
851 && prune_by->write_p)
854 prune_ref_by_group_reuse (ref, prune_by, before);
858 /* Prune the prefetch candidates in GROUP using the reuse analysis. */
861 prune_group_by_reuse (struct mem_ref_group *group)
863 struct mem_ref *ref_pruned;
865 for (ref_pruned = group->refs; ref_pruned; ref_pruned = ref_pruned->next)
867 prune_ref_by_reuse (ref_pruned, group->refs);
869 if (dump_file && (dump_flags & TDF_DETAILS))
871 fprintf (dump_file, "Reference %p:", (void *) ref_pruned);
873 if (ref_pruned->prefetch_before == PREFETCH_ALL
874 && ref_pruned->prefetch_mod == 1)
875 fprintf (dump_file, " no restrictions");
876 else if (ref_pruned->prefetch_before == 0)
877 fprintf (dump_file, " do not prefetch");
878 else if (ref_pruned->prefetch_before <= ref_pruned->prefetch_mod)
879 fprintf (dump_file, " prefetch once");
882 if (ref_pruned->prefetch_before != PREFETCH_ALL)
884 fprintf (dump_file, " prefetch before ");
885 fprintf (dump_file, HOST_WIDE_INT_PRINT_DEC,
886 ref_pruned->prefetch_before);
888 if (ref_pruned->prefetch_mod != 1)
890 fprintf (dump_file, " prefetch mod ");
891 fprintf (dump_file, HOST_WIDE_INT_PRINT_DEC,
892 ref_pruned->prefetch_mod);
895 fprintf (dump_file, "\n");
900 /* Prune the list of prefetch candidates GROUPS using the reuse analysis. */
903 prune_by_reuse (struct mem_ref_group *groups)
905 for (; groups; groups = groups->next)
906 prune_group_by_reuse (groups);
909 /* Returns true if we should issue prefetch for REF. */
912 should_issue_prefetch_p (struct mem_ref *ref)
914 /* For now do not issue prefetches for only first few of the
916 if (ref->prefetch_before != PREFETCH_ALL)
918 if (dump_file && (dump_flags & TDF_DETAILS))
919 fprintf (dump_file, "Ignoring %p due to prefetch_before\n",
924 /* Do not prefetch nontemporal stores. */
927 if (dump_file && (dump_flags & TDF_DETAILS))
928 fprintf (dump_file, "Ignoring nontemporal store %p\n", (void *) ref);
935 /* Decide which of the prefetch candidates in GROUPS to prefetch.
936 AHEAD is the number of iterations to prefetch ahead (which corresponds
937 to the number of simultaneous instances of one prefetch running at a
938 time). UNROLL_FACTOR is the factor by that the loop is going to be
939 unrolled. Returns true if there is anything to prefetch. */
942 schedule_prefetches (struct mem_ref_group *groups, unsigned unroll_factor,
945 unsigned remaining_prefetch_slots, n_prefetches, prefetch_slots;
946 unsigned slots_per_prefetch;
950 /* At most SIMULTANEOUS_PREFETCHES should be running at the same time. */
951 remaining_prefetch_slots = SIMULTANEOUS_PREFETCHES;
953 /* The prefetch will run for AHEAD iterations of the original loop, i.e.,
954 AHEAD / UNROLL_FACTOR iterations of the unrolled loop. In each iteration,
955 it will need a prefetch slot. */
956 slots_per_prefetch = (ahead + unroll_factor / 2) / unroll_factor;
957 if (dump_file && (dump_flags & TDF_DETAILS))
958 fprintf (dump_file, "Each prefetch instruction takes %u prefetch slots.\n",
961 /* For now we just take memory references one by one and issue
962 prefetches for as many as possible. The groups are sorted
963 starting with the largest step, since the references with
964 large step are more likely to cause many cache misses. */
966 for (; groups; groups = groups->next)
967 for (ref = groups->refs; ref; ref = ref->next)
969 if (!should_issue_prefetch_p (ref))
972 /* The loop is far from being sufficiently unrolled for this
973 prefetch. Do not generate prefetch to avoid many redudant
975 if (ref->prefetch_mod / unroll_factor > PREFETCH_MOD_TO_UNROLL_FACTOR_RATIO)
978 /* If we need to prefetch the reference each PREFETCH_MOD iterations,
979 and we unroll the loop UNROLL_FACTOR times, we need to insert
980 ceil (UNROLL_FACTOR / PREFETCH_MOD) instructions in each
982 n_prefetches = ((unroll_factor + ref->prefetch_mod - 1)
983 / ref->prefetch_mod);
984 prefetch_slots = n_prefetches * slots_per_prefetch;
986 /* If more than half of the prefetches would be lost anyway, do not
987 issue the prefetch. */
988 if (2 * remaining_prefetch_slots < prefetch_slots)
991 ref->issue_prefetch_p = true;
993 if (remaining_prefetch_slots <= prefetch_slots)
995 remaining_prefetch_slots -= prefetch_slots;
1002 /* Return TRUE if no prefetch is going to be generated in the given
1006 nothing_to_prefetch_p (struct mem_ref_group *groups)
1008 struct mem_ref *ref;
1010 for (; groups; groups = groups->next)
1011 for (ref = groups->refs; ref; ref = ref->next)
1012 if (should_issue_prefetch_p (ref))
1018 /* Estimate the number of prefetches in the given GROUPS.
1019 UNROLL_FACTOR is the factor by which LOOP was unrolled. */
1022 estimate_prefetch_count (struct mem_ref_group *groups, unsigned unroll_factor)
1024 struct mem_ref *ref;
1025 unsigned n_prefetches;
1026 int prefetch_count = 0;
1028 for (; groups; groups = groups->next)
1029 for (ref = groups->refs; ref; ref = ref->next)
1030 if (should_issue_prefetch_p (ref))
1032 n_prefetches = ((unroll_factor + ref->prefetch_mod - 1)
1033 / ref->prefetch_mod);
1034 prefetch_count += n_prefetches;
1037 return prefetch_count;
1040 /* Issue prefetches for the reference REF into loop as decided before.
1041 HEAD is the number of iterations to prefetch ahead. UNROLL_FACTOR
1042 is the factor by which LOOP was unrolled. */
1045 issue_prefetch_ref (struct mem_ref *ref, unsigned unroll_factor, unsigned ahead)
1047 HOST_WIDE_INT delta;
1048 tree addr, addr_base, write_p, local, forward;
1050 gimple_stmt_iterator bsi;
1051 unsigned n_prefetches, ap;
1052 bool nontemporal = ref->reuse_distance >= L2_CACHE_SIZE_BYTES;
1054 if (dump_file && (dump_flags & TDF_DETAILS))
1055 fprintf (dump_file, "Issued%s prefetch for %p.\n",
1056 nontemporal ? " nontemporal" : "",
1059 bsi = gsi_for_stmt (ref->stmt);
1061 n_prefetches = ((unroll_factor + ref->prefetch_mod - 1)
1062 / ref->prefetch_mod);
1063 addr_base = build_fold_addr_expr_with_type (ref->mem, ptr_type_node);
1064 addr_base = force_gimple_operand_gsi (&bsi, unshare_expr (addr_base),
1065 true, NULL, true, GSI_SAME_STMT);
1066 write_p = ref->write_p ? integer_one_node : integer_zero_node;
1067 local = build_int_cst (integer_type_node, nontemporal ? 0 : 3);
1069 for (ap = 0; ap < n_prefetches; ap++)
1071 if (cst_and_fits_in_hwi (ref->group->step))
1073 /* Determine the address to prefetch. */
1074 delta = (ahead + ap * ref->prefetch_mod) *
1075 int_cst_value (ref->group->step);
1076 addr = fold_build2 (POINTER_PLUS_EXPR, ptr_type_node,
1077 addr_base, size_int (delta));
1078 addr = force_gimple_operand_gsi (&bsi, unshare_expr (addr), true, NULL,
1079 true, GSI_SAME_STMT);
1083 /* The step size is non-constant but loop-invariant. We use the
1084 heuristic to simply prefetch ahead iterations ahead. */
1085 forward = fold_build2 (MULT_EXPR, sizetype,
1086 fold_convert (sizetype, ref->group->step),
1087 fold_convert (sizetype, size_int (ahead)));
1088 addr = fold_build2 (POINTER_PLUS_EXPR, ptr_type_node, addr_base,
1090 addr = force_gimple_operand_gsi (&bsi, unshare_expr (addr), true,
1091 NULL, true, GSI_SAME_STMT);
1093 /* Create the prefetch instruction. */
1094 prefetch = gimple_build_call (built_in_decls[BUILT_IN_PREFETCH],
1095 3, addr, write_p, local);
1096 gsi_insert_before (&bsi, prefetch, GSI_SAME_STMT);
1100 /* Issue prefetches for the references in GROUPS into loop as decided before.
1101 HEAD is the number of iterations to prefetch ahead. UNROLL_FACTOR is the
1102 factor by that LOOP was unrolled. */
1105 issue_prefetches (struct mem_ref_group *groups,
1106 unsigned unroll_factor, unsigned ahead)
1108 struct mem_ref *ref;
1110 for (; groups; groups = groups->next)
1111 for (ref = groups->refs; ref; ref = ref->next)
1112 if (ref->issue_prefetch_p)
1113 issue_prefetch_ref (ref, unroll_factor, ahead);
1116 /* Returns true if REF is a memory write for that a nontemporal store insn
1120 nontemporal_store_p (struct mem_ref *ref)
1122 enum machine_mode mode;
1123 enum insn_code code;
1125 /* REF must be a write that is not reused. We require it to be independent
1126 on all other memory references in the loop, as the nontemporal stores may
1127 be reordered with respect to other memory references. */
1129 || !ref->independent_p
1130 || ref->reuse_distance < L2_CACHE_SIZE_BYTES)
1133 /* Check that we have the storent instruction for the mode. */
1134 mode = TYPE_MODE (TREE_TYPE (ref->mem));
1135 if (mode == BLKmode)
1138 code = optab_handler (storent_optab, mode);
1139 return code != CODE_FOR_nothing;
1142 /* If REF is a nontemporal store, we mark the corresponding modify statement
1143 and return true. Otherwise, we return false. */
1146 mark_nontemporal_store (struct mem_ref *ref)
1148 if (!nontemporal_store_p (ref))
1151 if (dump_file && (dump_flags & TDF_DETAILS))
1152 fprintf (dump_file, "Marked reference %p as a nontemporal store.\n",
1155 gimple_assign_set_nontemporal_move (ref->stmt, true);
1156 ref->storent_p = true;
1161 /* Issue a memory fence instruction after LOOP. */
1164 emit_mfence_after_loop (struct loop *loop)
1166 VEC (edge, heap) *exits = get_loop_exit_edges (loop);
1169 gimple_stmt_iterator bsi;
1172 for (i = 0; VEC_iterate (edge, exits, i, exit); i++)
1174 call = gimple_build_call (FENCE_FOLLOWING_MOVNT, 0);
1176 if (!single_pred_p (exit->dest)
1177 /* If possible, we prefer not to insert the fence on other paths
1179 && !(exit->flags & EDGE_ABNORMAL))
1180 split_loop_exit_edge (exit);
1181 bsi = gsi_after_labels (exit->dest);
1183 gsi_insert_before (&bsi, call, GSI_NEW_STMT);
1184 mark_virtual_ops_for_renaming (call);
1187 VEC_free (edge, heap, exits);
1188 update_ssa (TODO_update_ssa_only_virtuals);
1191 /* Returns true if we can use storent in loop, false otherwise. */
1194 may_use_storent_in_loop_p (struct loop *loop)
1198 if (loop->inner != NULL)
1201 /* If we must issue a mfence insn after using storent, check that there
1202 is a suitable place for it at each of the loop exits. */
1203 if (FENCE_FOLLOWING_MOVNT != NULL_TREE)
1205 VEC (edge, heap) *exits = get_loop_exit_edges (loop);
1209 for (i = 0; VEC_iterate (edge, exits, i, exit); i++)
1210 if ((exit->flags & EDGE_ABNORMAL)
1211 && exit->dest == EXIT_BLOCK_PTR)
1214 VEC_free (edge, heap, exits);
1220 /* Marks nontemporal stores in LOOP. GROUPS contains the description of memory
1221 references in the loop. */
1224 mark_nontemporal_stores (struct loop *loop, struct mem_ref_group *groups)
1226 struct mem_ref *ref;
1229 if (!may_use_storent_in_loop_p (loop))
1232 for (; groups; groups = groups->next)
1233 for (ref = groups->refs; ref; ref = ref->next)
1234 any |= mark_nontemporal_store (ref);
1236 if (any && FENCE_FOLLOWING_MOVNT != NULL_TREE)
1237 emit_mfence_after_loop (loop);
1240 /* Determines whether we can profitably unroll LOOP FACTOR times, and if
1241 this is the case, fill in DESC by the description of number of
1245 should_unroll_loop_p (struct loop *loop, struct tree_niter_desc *desc,
1248 if (!can_unroll_loop_p (loop, factor, desc))
1251 /* We only consider loops without control flow for unrolling. This is not
1252 a hard restriction -- tree_unroll_loop works with arbitrary loops
1253 as well; but the unrolling/prefetching is usually more profitable for
1254 loops consisting of a single basic block, and we want to limit the
1256 if (loop->num_nodes > 2)
1262 /* Determine the coefficient by that unroll LOOP, from the information
1263 contained in the list of memory references REFS. Description of
1264 umber of iterations of LOOP is stored to DESC. NINSNS is the number of
1265 insns of the LOOP. EST_NITER is the estimated number of iterations of
1266 the loop, or -1 if no estimate is available. */
1269 determine_unroll_factor (struct loop *loop, struct mem_ref_group *refs,
1270 unsigned ninsns, struct tree_niter_desc *desc,
1271 HOST_WIDE_INT est_niter)
1273 unsigned upper_bound;
1274 unsigned nfactor, factor, mod_constraint;
1275 struct mem_ref_group *agp;
1276 struct mem_ref *ref;
1278 /* First check whether the loop is not too large to unroll. We ignore
1279 PARAM_MAX_UNROLL_TIMES, because for small loops, it prevented us
1280 from unrolling them enough to make exactly one cache line covered by each
1281 iteration. Also, the goal of PARAM_MAX_UNROLL_TIMES is to prevent
1282 us from unrolling the loops too many times in cases where we only expect
1283 gains from better scheduling and decreasing loop overhead, which is not
1285 upper_bound = PARAM_VALUE (PARAM_MAX_UNROLLED_INSNS) / ninsns;
1287 /* If we unrolled the loop more times than it iterates, the unrolled version
1288 of the loop would be never entered. */
1289 if (est_niter >= 0 && est_niter < (HOST_WIDE_INT) upper_bound)
1290 upper_bound = est_niter;
1292 if (upper_bound <= 1)
1295 /* Choose the factor so that we may prefetch each cache just once,
1296 but bound the unrolling by UPPER_BOUND. */
1298 for (agp = refs; agp; agp = agp->next)
1299 for (ref = agp->refs; ref; ref = ref->next)
1300 if (should_issue_prefetch_p (ref))
1302 mod_constraint = ref->prefetch_mod;
1303 nfactor = least_common_multiple (mod_constraint, factor);
1304 if (nfactor <= upper_bound)
1308 if (!should_unroll_loop_p (loop, desc, factor))
1314 /* Returns the total volume of the memory references REFS, taking into account
1315 reuses in the innermost loop and cache line size. TODO -- we should also
1316 take into account reuses across the iterations of the loops in the loop
1320 volume_of_references (struct mem_ref_group *refs)
1322 unsigned volume = 0;
1323 struct mem_ref_group *gr;
1324 struct mem_ref *ref;
1326 for (gr = refs; gr; gr = gr->next)
1327 for (ref = gr->refs; ref; ref = ref->next)
1329 /* Almost always reuses another value? */
1330 if (ref->prefetch_before != PREFETCH_ALL)
1333 /* If several iterations access the same cache line, use the size of
1334 the line divided by this number. Otherwise, a cache line is
1335 accessed in each iteration. TODO -- in the latter case, we should
1336 take the size of the reference into account, rounding it up on cache
1337 line size multiple. */
1338 volume += L1_CACHE_LINE_SIZE / ref->prefetch_mod;
1343 /* Returns the volume of memory references accessed across VEC iterations of
1344 loops, whose sizes are described in the LOOP_SIZES array. N is the number
1345 of the loops in the nest (length of VEC and LOOP_SIZES vectors). */
1348 volume_of_dist_vector (lambda_vector vec, unsigned *loop_sizes, unsigned n)
1352 for (i = 0; i < n; i++)
1359 gcc_assert (vec[i] > 0);
1361 /* We ignore the parts of the distance vector in subloops, since usually
1362 the numbers of iterations are much smaller. */
1363 return loop_sizes[i] * vec[i];
1366 /* Add the steps of ACCESS_FN multiplied by STRIDE to the array STRIDE
1367 at the position corresponding to the loop of the step. N is the depth
1368 of the considered loop nest, and, LOOP is its innermost loop. */
1371 add_subscript_strides (tree access_fn, unsigned stride,
1372 HOST_WIDE_INT *strides, unsigned n, struct loop *loop)
1376 HOST_WIDE_INT astep;
1377 unsigned min_depth = loop_depth (loop) - n;
1379 while (TREE_CODE (access_fn) == POLYNOMIAL_CHREC)
1381 aloop = get_chrec_loop (access_fn);
1382 step = CHREC_RIGHT (access_fn);
1383 access_fn = CHREC_LEFT (access_fn);
1385 if ((unsigned) loop_depth (aloop) <= min_depth)
1388 if (host_integerp (step, 0))
1389 astep = tree_low_cst (step, 0);
1391 astep = L1_CACHE_LINE_SIZE;
1393 strides[n - 1 - loop_depth (loop) + loop_depth (aloop)] += astep * stride;
1398 /* Returns the volume of memory references accessed between two consecutive
1399 self-reuses of the reference DR. We consider the subscripts of DR in N
1400 loops, and LOOP_SIZES contains the volumes of accesses in each of the
1401 loops. LOOP is the innermost loop of the current loop nest. */
1404 self_reuse_distance (data_reference_p dr, unsigned *loop_sizes, unsigned n,
1407 tree stride, access_fn;
1408 HOST_WIDE_INT *strides, astride;
1409 VEC (tree, heap) *access_fns;
1410 tree ref = DR_REF (dr);
1411 unsigned i, ret = ~0u;
1413 /* In the following example:
1415 for (i = 0; i < N; i++)
1416 for (j = 0; j < N; j++)
1418 the same cache line is accessed each N steps (except if the change from
1419 i to i + 1 crosses the boundary of the cache line). Thus, for self-reuse,
1420 we cannot rely purely on the results of the data dependence analysis.
1422 Instead, we compute the stride of the reference in each loop, and consider
1423 the innermost loop in that the stride is less than cache size. */
1425 strides = XCNEWVEC (HOST_WIDE_INT, n);
1426 access_fns = DR_ACCESS_FNS (dr);
1428 for (i = 0; VEC_iterate (tree, access_fns, i, access_fn); i++)
1430 /* Keep track of the reference corresponding to the subscript, so that we
1432 while (handled_component_p (ref) && TREE_CODE (ref) != ARRAY_REF)
1433 ref = TREE_OPERAND (ref, 0);
1435 if (TREE_CODE (ref) == ARRAY_REF)
1437 stride = TYPE_SIZE_UNIT (TREE_TYPE (ref));
1438 if (host_integerp (stride, 1))
1439 astride = tree_low_cst (stride, 1);
1441 astride = L1_CACHE_LINE_SIZE;
1443 ref = TREE_OPERAND (ref, 0);
1448 add_subscript_strides (access_fn, astride, strides, n, loop);
1451 for (i = n; i-- > 0; )
1453 unsigned HOST_WIDE_INT s;
1455 s = strides[i] < 0 ? -strides[i] : strides[i];
1457 if (s < (unsigned) L1_CACHE_LINE_SIZE
1459 > (unsigned) (L1_CACHE_SIZE_BYTES / NONTEMPORAL_FRACTION)))
1461 ret = loop_sizes[i];
1470 /* Determines the distance till the first reuse of each reference in REFS
1471 in the loop nest of LOOP. NO_OTHER_REFS is true if there are no other
1472 memory references in the loop. */
1475 determine_loop_nest_reuse (struct loop *loop, struct mem_ref_group *refs,
1478 struct loop *nest, *aloop;
1479 VEC (data_reference_p, heap) *datarefs = NULL;
1480 VEC (ddr_p, heap) *dependences = NULL;
1481 struct mem_ref_group *gr;
1482 struct mem_ref *ref, *refb;
1483 VEC (loop_p, heap) *vloops = NULL;
1484 unsigned *loop_data_size;
1486 unsigned volume, dist, adist;
1488 data_reference_p dr;
1494 /* Find the outermost loop of the loop nest of loop (we require that
1495 there are no sibling loops inside the nest). */
1499 aloop = loop_outer (nest);
1501 if (aloop == current_loops->tree_root
1502 || aloop->inner->next)
1508 /* For each loop, determine the amount of data accessed in each iteration.
1509 We use this to estimate whether the reference is evicted from the
1510 cache before its reuse. */
1511 find_loop_nest (nest, &vloops);
1512 n = VEC_length (loop_p, vloops);
1513 loop_data_size = XNEWVEC (unsigned, n);
1514 volume = volume_of_references (refs);
1518 loop_data_size[i] = volume;
1519 /* Bound the volume by the L2 cache size, since above this bound,
1520 all dependence distances are equivalent. */
1521 if (volume > L2_CACHE_SIZE_BYTES)
1524 aloop = VEC_index (loop_p, vloops, i);
1525 vol = estimated_loop_iterations_int (aloop, false);
1527 vol = expected_loop_iterations (aloop);
1531 /* Prepare the references in the form suitable for data dependence
1532 analysis. We ignore unanalyzable data references (the results
1533 are used just as a heuristics to estimate temporality of the
1534 references, hence we do not need to worry about correctness). */
1535 for (gr = refs; gr; gr = gr->next)
1536 for (ref = gr->refs; ref; ref = ref->next)
1538 dr = create_data_ref (nest, ref->mem, ref->stmt, !ref->write_p);
1542 ref->reuse_distance = volume;
1544 VEC_safe_push (data_reference_p, heap, datarefs, dr);
1547 no_other_refs = false;
1550 for (i = 0; VEC_iterate (data_reference_p, datarefs, i, dr); i++)
1552 dist = self_reuse_distance (dr, loop_data_size, n, loop);
1553 ref = (struct mem_ref *) dr->aux;
1554 if (ref->reuse_distance > dist)
1555 ref->reuse_distance = dist;
1558 ref->independent_p = true;
1561 compute_all_dependences (datarefs, &dependences, vloops, true);
1563 for (i = 0; VEC_iterate (ddr_p, dependences, i, dep); i++)
1565 if (DDR_ARE_DEPENDENT (dep) == chrec_known)
1568 ref = (struct mem_ref *) DDR_A (dep)->aux;
1569 refb = (struct mem_ref *) DDR_B (dep)->aux;
1571 if (DDR_ARE_DEPENDENT (dep) == chrec_dont_know
1572 || DDR_NUM_DIST_VECTS (dep) == 0)
1574 /* If the dependence cannot be analyzed, assume that there might be
1578 ref->independent_p = false;
1579 refb->independent_p = false;
1583 /* The distance vectors are normalized to be always lexicographically
1584 positive, hence we cannot tell just from them whether DDR_A comes
1585 before DDR_B or vice versa. However, it is not important,
1586 anyway -- if DDR_A is close to DDR_B, then it is either reused in
1587 DDR_B (and it is not nontemporal), or it reuses the value of DDR_B
1588 in cache (and marking it as nontemporal would not affect
1592 for (j = 0; j < DDR_NUM_DIST_VECTS (dep); j++)
1594 adist = volume_of_dist_vector (DDR_DIST_VECT (dep, j),
1597 /* If this is a dependence in the innermost loop (i.e., the
1598 distances in all superloops are zero) and it is not
1599 the trivial self-dependence with distance zero, record that
1600 the references are not completely independent. */
1601 if (lambda_vector_zerop (DDR_DIST_VECT (dep, j), n - 1)
1603 || DDR_DIST_VECT (dep, j)[n-1] != 0))
1605 ref->independent_p = false;
1606 refb->independent_p = false;
1609 /* Ignore accesses closer than
1610 L1_CACHE_SIZE_BYTES / NONTEMPORAL_FRACTION,
1611 so that we use nontemporal prefetches e.g. if single memory
1612 location is accessed several times in a single iteration of
1614 if (adist < L1_CACHE_SIZE_BYTES / NONTEMPORAL_FRACTION)
1622 if (ref->reuse_distance > dist)
1623 ref->reuse_distance = dist;
1624 if (refb->reuse_distance > dist)
1625 refb->reuse_distance = dist;
1628 free_dependence_relations (dependences);
1629 free_data_refs (datarefs);
1630 free (loop_data_size);
1632 if (dump_file && (dump_flags & TDF_DETAILS))
1634 fprintf (dump_file, "Reuse distances:\n");
1635 for (gr = refs; gr; gr = gr->next)
1636 for (ref = gr->refs; ref; ref = ref->next)
1637 fprintf (dump_file, " ref %p distance %u\n",
1638 (void *) ref, ref->reuse_distance);
1642 /* Do a cost-benefit analysis to determine if prefetching is profitable
1643 for the current loop given the following parameters:
1644 AHEAD: the iteration ahead distance,
1645 EST_NITER: the estimated trip count,
1646 NINSNS: estimated number of instructions in the loop,
1647 PREFETCH_COUNT: an estimate of the number of prefetches
1648 MEM_REF_COUNT: total number of memory references in the loop. */
1651 is_loop_prefetching_profitable (unsigned ahead, HOST_WIDE_INT est_niter,
1652 unsigned ninsns, unsigned prefetch_count,
1653 unsigned mem_ref_count, unsigned unroll_factor)
1655 int insn_to_mem_ratio, insn_to_prefetch_ratio;
1657 if (mem_ref_count == 0)
1660 /* Prefetching improves performance by overlapping cache missing
1661 memory accesses with CPU operations. If the loop does not have
1662 enough CPU operations to overlap with memory operations, prefetching
1663 won't give a significant benefit. One approximate way of checking
1664 this is to require the ratio of instructions to memory references to
1665 be above a certain limit. This approximation works well in practice.
1666 TODO: Implement a more precise computation by estimating the time
1667 for each CPU or memory op in the loop. Time estimates for memory ops
1668 should account for cache misses. */
1669 insn_to_mem_ratio = ninsns / mem_ref_count;
1671 if (insn_to_mem_ratio < PREFETCH_MIN_INSN_TO_MEM_RATIO)
1673 if (dump_file && (dump_flags & TDF_DETAILS))
1675 "Not prefetching -- instruction to memory reference ratio (%d) too small\n",
1680 /* Prefetching most likely causes performance degradation when the instruction
1681 to prefetch ratio is too small. Too many prefetch instructions in a loop
1682 may reduce the I-cache performance.
1683 (unroll_factor * ninsns) is used to estimate the number of instructions in
1684 the unrolled loop. This implementation is a bit simplistic -- the number
1685 of issued prefetch instructions is also affected by unrolling. So,
1686 prefetch_mod and the unroll factor should be taken into account when
1687 determining prefetch_count. Also, the number of insns of the unrolled
1688 loop will usually be significantly smaller than the number of insns of the
1689 original loop * unroll_factor (at least the induction variable increases
1690 and the exit branches will get eliminated), so it might be better to use
1691 tree_estimate_loop_size + estimated_unrolled_size. */
1692 insn_to_prefetch_ratio = (unroll_factor * ninsns) / prefetch_count;
1693 if (insn_to_prefetch_ratio < MIN_INSN_TO_PREFETCH_RATIO)
1695 if (dump_file && (dump_flags & TDF_DETAILS))
1697 "Not prefetching -- instruction to prefetch ratio (%d) too small\n",
1698 insn_to_prefetch_ratio);
1702 /* Could not do further estimation if the trip count is unknown. Just assume
1703 prefetching is profitable. Too aggressive??? */
1707 if (est_niter < (HOST_WIDE_INT) (TRIP_COUNT_TO_AHEAD_RATIO * ahead))
1709 if (dump_file && (dump_flags & TDF_DETAILS))
1711 "Not prefetching -- loop estimated to roll only %d times\n",
1719 /* Issue prefetch instructions for array references in LOOP. Returns
1720 true if the LOOP was unrolled. */
1723 loop_prefetch_arrays (struct loop *loop)
1725 struct mem_ref_group *refs;
1726 unsigned ahead, ninsns, time, unroll_factor;
1727 HOST_WIDE_INT est_niter;
1728 struct tree_niter_desc desc;
1729 bool unrolled = false, no_other_refs;
1730 unsigned prefetch_count;
1731 unsigned mem_ref_count;
1733 if (optimize_loop_nest_for_size_p (loop))
1735 if (dump_file && (dump_flags & TDF_DETAILS))
1736 fprintf (dump_file, " ignored (cold area)\n");
1740 /* Step 1: gather the memory references. */
1741 refs = gather_memory_references (loop, &no_other_refs, &mem_ref_count);
1743 /* Step 2: estimate the reuse effects. */
1744 prune_by_reuse (refs);
1746 if (nothing_to_prefetch_p (refs))
1749 determine_loop_nest_reuse (loop, refs, no_other_refs);
1751 /* Step 3: determine the ahead and unroll factor. */
1753 /* FIXME: the time should be weighted by the probabilities of the blocks in
1755 time = tree_num_loop_insns (loop, &eni_time_weights);
1756 ahead = (PREFETCH_LATENCY + time - 1) / time;
1757 est_niter = estimated_loop_iterations_int (loop, false);
1759 ninsns = tree_num_loop_insns (loop, &eni_size_weights);
1760 unroll_factor = determine_unroll_factor (loop, refs, ninsns, &desc,
1763 /* Estimate prefetch count for the unrolled loop. */
1764 prefetch_count = estimate_prefetch_count (refs, unroll_factor);
1765 if (prefetch_count == 0)
1768 if (dump_file && (dump_flags & TDF_DETAILS))
1769 fprintf (dump_file, "Ahead %d, unroll factor %d, trip count "
1770 HOST_WIDE_INT_PRINT_DEC "\n"
1771 "insn count %d, mem ref count %d, prefetch count %d\n",
1772 ahead, unroll_factor, est_niter,
1773 ninsns, mem_ref_count, prefetch_count);
1775 if (!is_loop_prefetching_profitable (ahead, est_niter, ninsns, prefetch_count,
1776 mem_ref_count, unroll_factor))
1779 mark_nontemporal_stores (loop, refs);
1781 /* Step 4: what to prefetch? */
1782 if (!schedule_prefetches (refs, unroll_factor, ahead))
1785 /* Step 5: unroll the loop. TODO -- peeling of first and last few
1786 iterations so that we do not issue superfluous prefetches. */
1787 if (unroll_factor != 1)
1789 tree_unroll_loop (loop, unroll_factor,
1790 single_dom_exit (loop), &desc);
1794 /* Step 6: issue the prefetches. */
1795 issue_prefetches (refs, unroll_factor, ahead);
1798 release_mem_refs (refs);
1802 /* Issue prefetch instructions for array references in loops. */
1805 tree_ssa_prefetch_arrays (void)
1809 bool unrolled = false;
1813 /* It is possible to ask compiler for say -mtune=i486 -march=pentium4.
1814 -mtune=i486 causes us having PREFETCH_BLOCK 0, since this is part
1815 of processor costs and i486 does not have prefetch, but
1816 -march=pentium4 causes HAVE_prefetch to be true. Ugh. */
1817 || PREFETCH_BLOCK == 0)
1820 if (dump_file && (dump_flags & TDF_DETAILS))
1822 fprintf (dump_file, "Prefetching parameters:\n");
1823 fprintf (dump_file, " simultaneous prefetches: %d\n",
1824 SIMULTANEOUS_PREFETCHES);
1825 fprintf (dump_file, " prefetch latency: %d\n", PREFETCH_LATENCY);
1826 fprintf (dump_file, " prefetch block size: %d\n", PREFETCH_BLOCK);
1827 fprintf (dump_file, " L1 cache size: %d lines, %d kB\n",
1828 L1_CACHE_SIZE_BYTES / L1_CACHE_LINE_SIZE, L1_CACHE_SIZE);
1829 fprintf (dump_file, " L1 cache line size: %d\n", L1_CACHE_LINE_SIZE);
1830 fprintf (dump_file, " L2 cache size: %d kB\n", L2_CACHE_SIZE);
1831 fprintf (dump_file, " min insn-to-prefetch ratio: %d \n",
1832 MIN_INSN_TO_PREFETCH_RATIO);
1833 fprintf (dump_file, " min insn-to-mem ratio: %d \n",
1834 PREFETCH_MIN_INSN_TO_MEM_RATIO);
1835 fprintf (dump_file, "\n");
1838 initialize_original_copy_tables ();
1840 if (!built_in_decls[BUILT_IN_PREFETCH])
1842 tree type = build_function_type (void_type_node,
1843 tree_cons (NULL_TREE,
1844 const_ptr_type_node,
1846 tree decl = add_builtin_function ("__builtin_prefetch", type,
1847 BUILT_IN_PREFETCH, BUILT_IN_NORMAL,
1849 DECL_IS_NOVOPS (decl) = true;
1850 built_in_decls[BUILT_IN_PREFETCH] = decl;
1853 /* We assume that size of cache line is a power of two, so verify this
1855 gcc_assert ((PREFETCH_BLOCK & (PREFETCH_BLOCK - 1)) == 0);
1857 FOR_EACH_LOOP (li, loop, LI_FROM_INNERMOST)
1859 if (dump_file && (dump_flags & TDF_DETAILS))
1860 fprintf (dump_file, "Processing loop %d:\n", loop->num);
1862 unrolled |= loop_prefetch_arrays (loop);
1864 if (dump_file && (dump_flags & TDF_DETAILS))
1865 fprintf (dump_file, "\n\n");
1871 todo_flags |= TODO_cleanup_cfg;
1874 free_original_copy_tables ();