2 Copyright (C) 2005, 2007, 2008 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"
27 #include "hard-reg-set.h"
28 #include "basic-block.h"
30 #include "diagnostic.h"
31 #include "tree-flow.h"
32 #include "tree-dump.h"
37 #include "tree-pass.h"
39 #include "insn-config.h"
42 #include "tree-chrec.h"
43 #include "tree-scalar-evolution.h"
46 #include "langhooks.h"
47 #include "tree-inline.h"
48 #include "tree-data-ref.h"
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?
113 -- write and use more general reuse analysis (that could be also used
114 in other cache aimed loop optimizations)
115 -- make it behave sanely together with the prefetches given by user
116 (now we just ignore them; at the very least we should avoid
117 optimizing loops in that user put his own prefetches)
118 -- we assume cache line size alignment of arrays; this could be
121 /* Magic constants follow. These should be replaced by machine specific
124 /* True if write can be prefetched by a read prefetch. */
126 #ifndef WRITE_CAN_USE_READ_PREFETCH
127 #define WRITE_CAN_USE_READ_PREFETCH 1
130 /* True if read can be prefetched by a write prefetch. */
132 #ifndef READ_CAN_USE_WRITE_PREFETCH
133 #define READ_CAN_USE_WRITE_PREFETCH 0
136 /* The size of the block loaded by a single prefetch. Usually, this is
137 the same as cache line size (at the moment, we only consider one level
138 of cache hierarchy). */
140 #ifndef PREFETCH_BLOCK
141 #define PREFETCH_BLOCK L1_CACHE_LINE_SIZE
144 /* Do we have a forward hardware sequential prefetching? */
146 #ifndef HAVE_FORWARD_PREFETCH
147 #define HAVE_FORWARD_PREFETCH 0
150 /* Do we have a backward hardware sequential prefetching? */
152 #ifndef HAVE_BACKWARD_PREFETCH
153 #define HAVE_BACKWARD_PREFETCH 0
156 /* In some cases we are only able to determine that there is a certain
157 probability that the two accesses hit the same cache line. In this
158 case, we issue the prefetches for both of them if this probability
159 is less then (1000 - ACCEPTABLE_MISS_RATE) per thousand. */
161 #ifndef ACCEPTABLE_MISS_RATE
162 #define ACCEPTABLE_MISS_RATE 50
165 #ifndef HAVE_prefetch
166 #define HAVE_prefetch 0
169 #define L1_CACHE_SIZE_BYTES ((unsigned) (L1_CACHE_SIZE * 1024))
170 #define L2_CACHE_SIZE_BYTES ((unsigned) (L2_CACHE_SIZE * 1024))
172 /* We consider a memory access nontemporal if it is not reused sooner than
173 after L2_CACHE_SIZE_BYTES of memory are accessed. However, we ignore
174 accesses closer than L1_CACHE_SIZE_BYTES / NONTEMPORAL_FRACTION,
175 so that we use nontemporal prefetches e.g. if single memory location
176 is accessed several times in a single iteration of the loop. */
177 #define NONTEMPORAL_FRACTION 16
179 /* In case we have to emit a memory fence instruction after the loop that
180 uses nontemporal stores, this defines the builtin to use. */
182 #ifndef FENCE_FOLLOWING_MOVNT
183 #define FENCE_FOLLOWING_MOVNT NULL_TREE
186 /* The group of references between that reuse may occur. */
190 tree base
; /* Base of the reference. */
191 HOST_WIDE_INT step
; /* Step of the reference. */
192 struct mem_ref
*refs
; /* References in the group. */
193 struct mem_ref_group
*next
; /* Next group of references. */
196 /* Assigned to PREFETCH_BEFORE when all iterations are to be prefetched. */
198 #define PREFETCH_ALL (~(unsigned HOST_WIDE_INT) 0)
200 /* The memory reference. */
204 tree stmt
; /* Statement in that the reference appears. */
205 tree mem
; /* The reference. */
206 HOST_WIDE_INT delta
; /* Constant offset of the reference. */
207 struct mem_ref_group
*group
; /* The group of references it belongs to. */
208 unsigned HOST_WIDE_INT prefetch_mod
;
209 /* Prefetch only each PREFETCH_MOD-th
211 unsigned HOST_WIDE_INT prefetch_before
;
212 /* Prefetch only first PREFETCH_BEFORE
214 unsigned reuse_distance
; /* The amount of data accessed before the first
215 reuse of this value. */
216 struct mem_ref
*next
; /* The next reference in the group. */
217 unsigned write_p
: 1; /* Is it a write? */
218 unsigned independent_p
: 1; /* True if the reference is independent on
219 all other references inside the loop. */
220 unsigned issue_prefetch_p
: 1; /* Should we really issue the prefetch? */
221 unsigned storent_p
: 1; /* True if we changed the store to a
225 /* Dumps information about reference REF to FILE. */
228 dump_mem_ref (FILE *file
, struct mem_ref
*ref
)
230 fprintf (file
, "Reference %p:\n", (void *) ref
);
232 fprintf (file
, " group %p (base ", (void *) ref
->group
);
233 print_generic_expr (file
, ref
->group
->base
, TDF_SLIM
);
234 fprintf (file
, ", step ");
235 fprintf (file
, HOST_WIDE_INT_PRINT_DEC
, ref
->group
->step
);
236 fprintf (file
, ")\n");
238 fprintf (file
, " delta ");
239 fprintf (file
, HOST_WIDE_INT_PRINT_DEC
, ref
->delta
);
240 fprintf (file
, "\n");
242 fprintf (file
, " %s\n", ref
->write_p
? "write" : "read");
244 fprintf (file
, "\n");
247 /* Finds a group with BASE and STEP in GROUPS, or creates one if it does not
250 static struct mem_ref_group
*
251 find_or_create_group (struct mem_ref_group
**groups
, tree base
,
254 struct mem_ref_group
*group
;
256 for (; *groups
; groups
= &(*groups
)->next
)
258 if ((*groups
)->step
== step
259 && operand_equal_p ((*groups
)->base
, base
, 0))
262 /* Keep the list of groups sorted by decreasing step. */
263 if ((*groups
)->step
< step
)
267 group
= XNEW (struct mem_ref_group
);
271 group
->next
= *groups
;
277 /* Records a memory reference MEM in GROUP with offset DELTA and write status
278 WRITE_P. The reference occurs in statement STMT. */
281 record_ref (struct mem_ref_group
*group
, tree stmt
, tree mem
,
282 HOST_WIDE_INT delta
, bool write_p
)
284 struct mem_ref
**aref
;
286 /* Do not record the same address twice. */
287 for (aref
= &group
->refs
; *aref
; aref
= &(*aref
)->next
)
289 /* It does not have to be possible for write reference to reuse the read
290 prefetch, or vice versa. */
291 if (!WRITE_CAN_USE_READ_PREFETCH
293 && !(*aref
)->write_p
)
295 if (!READ_CAN_USE_WRITE_PREFETCH
300 if ((*aref
)->delta
== delta
)
304 (*aref
) = XNEW (struct mem_ref
);
305 (*aref
)->stmt
= stmt
;
307 (*aref
)->delta
= delta
;
308 (*aref
)->write_p
= write_p
;
309 (*aref
)->prefetch_before
= PREFETCH_ALL
;
310 (*aref
)->prefetch_mod
= 1;
311 (*aref
)->reuse_distance
= 0;
312 (*aref
)->issue_prefetch_p
= false;
313 (*aref
)->group
= group
;
314 (*aref
)->next
= NULL
;
315 (*aref
)->independent_p
= false;
316 (*aref
)->storent_p
= false;
318 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
319 dump_mem_ref (dump_file
, *aref
);
322 /* Release memory references in GROUPS. */
325 release_mem_refs (struct mem_ref_group
*groups
)
327 struct mem_ref_group
*next_g
;
328 struct mem_ref
*ref
, *next_r
;
330 for (; groups
; groups
= next_g
)
332 next_g
= groups
->next
;
333 for (ref
= groups
->refs
; ref
; ref
= next_r
)
342 /* A structure used to pass arguments to idx_analyze_ref. */
346 struct loop
*loop
; /* Loop of the reference. */
347 tree stmt
; /* Statement of the reference. */
348 HOST_WIDE_INT
*step
; /* Step of the memory reference. */
349 HOST_WIDE_INT
*delta
; /* Offset of the memory reference. */
352 /* Analyzes a single INDEX of a memory reference to obtain information
353 described at analyze_ref. Callback for for_each_index. */
356 idx_analyze_ref (tree base
, tree
*index
, void *data
)
358 struct ar_data
*ar_data
= (struct ar_data
*) data
;
359 tree ibase
, step
, stepsize
;
360 HOST_WIDE_INT istep
, idelta
= 0, imult
= 1;
363 if (TREE_CODE (base
) == MISALIGNED_INDIRECT_REF
364 || TREE_CODE (base
) == ALIGN_INDIRECT_REF
)
367 if (!simple_iv (ar_data
->loop
, ar_data
->stmt
, *index
, &iv
, false))
372 if (!cst_and_fits_in_hwi (step
))
374 istep
= int_cst_value (step
);
376 if (TREE_CODE (ibase
) == POINTER_PLUS_EXPR
377 && cst_and_fits_in_hwi (TREE_OPERAND (ibase
, 1)))
379 idelta
= int_cst_value (TREE_OPERAND (ibase
, 1));
380 ibase
= TREE_OPERAND (ibase
, 0);
382 if (cst_and_fits_in_hwi (ibase
))
384 idelta
+= int_cst_value (ibase
);
385 ibase
= build_int_cst (TREE_TYPE (ibase
), 0);
388 if (TREE_CODE (base
) == ARRAY_REF
)
390 stepsize
= array_ref_element_size (base
);
391 if (!cst_and_fits_in_hwi (stepsize
))
393 imult
= int_cst_value (stepsize
);
399 *ar_data
->step
+= istep
;
400 *ar_data
->delta
+= idelta
;
406 /* Tries to express REF_P in shape &BASE + STEP * iter + DELTA, where DELTA and
407 STEP are integer constants and iter is number of iterations of LOOP. The
408 reference occurs in statement STMT. Strips nonaddressable component
409 references from REF_P. */
412 analyze_ref (struct loop
*loop
, tree
*ref_p
, tree
*base
,
413 HOST_WIDE_INT
*step
, HOST_WIDE_INT
*delta
,
416 struct ar_data ar_data
;
418 HOST_WIDE_INT bit_offset
;
424 /* First strip off the component references. Ignore bitfields. */
425 if (TREE_CODE (ref
) == COMPONENT_REF
426 && DECL_NONADDRESSABLE_P (TREE_OPERAND (ref
, 1)))
427 ref
= TREE_OPERAND (ref
, 0);
431 for (; TREE_CODE (ref
) == COMPONENT_REF
; ref
= TREE_OPERAND (ref
, 0))
433 off
= DECL_FIELD_BIT_OFFSET (TREE_OPERAND (ref
, 1));
434 bit_offset
= TREE_INT_CST_LOW (off
);
435 gcc_assert (bit_offset
% BITS_PER_UNIT
== 0);
437 *delta
+= bit_offset
/ BITS_PER_UNIT
;
440 *base
= unshare_expr (ref
);
444 ar_data
.delta
= delta
;
445 return for_each_index (base
, idx_analyze_ref
, &ar_data
);
448 /* Record a memory reference REF to the list REFS. The reference occurs in
449 LOOP in statement STMT and it is write if WRITE_P. Returns true if the
450 reference was recorded, false otherwise. */
453 gather_memory_references_ref (struct loop
*loop
, struct mem_ref_group
**refs
,
454 tree ref
, bool write_p
, tree stmt
)
457 HOST_WIDE_INT step
, delta
;
458 struct mem_ref_group
*agrp
;
460 if (get_base_address (ref
) == NULL
)
463 if (!analyze_ref (loop
, &ref
, &base
, &step
, &delta
, stmt
))
466 /* Now we know that REF = &BASE + STEP * iter + DELTA, where DELTA and STEP
467 are integer constants. */
468 agrp
= find_or_create_group (refs
, base
, step
);
469 record_ref (agrp
, stmt
, ref
, delta
, write_p
);
474 /* Record the suitable memory references in LOOP. NO_OTHER_REFS is set to
475 true if there are no other memory references inside the loop. */
477 static struct mem_ref_group
*
478 gather_memory_references (struct loop
*loop
, bool *no_other_refs
)
480 basic_block
*body
= get_loop_body_in_dom_order (loop
);
483 block_stmt_iterator bsi
;
484 tree stmt
, lhs
, rhs
, call
;
485 struct mem_ref_group
*refs
= NULL
;
487 *no_other_refs
= true;
489 /* Scan the loop body in order, so that the former references precede the
491 for (i
= 0; i
< loop
->num_nodes
; i
++)
494 if (bb
->loop_father
!= loop
)
497 for (bsi
= bsi_start (bb
); !bsi_end_p (bsi
); bsi_next (&bsi
))
499 stmt
= bsi_stmt (bsi
);
500 call
= get_call_expr_in (stmt
);
501 if (call
&& !(call_expr_flags (call
) & ECF_CONST
))
502 *no_other_refs
= false;
504 if (TREE_CODE (stmt
) != GIMPLE_MODIFY_STMT
)
506 if (!ZERO_SSA_OPERANDS (stmt
, SSA_OP_ALL_VIRTUALS
))
507 *no_other_refs
= false;
511 lhs
= GIMPLE_STMT_OPERAND (stmt
, 0);
512 rhs
= GIMPLE_STMT_OPERAND (stmt
, 1);
514 if (REFERENCE_CLASS_P (rhs
))
515 *no_other_refs
&= gather_memory_references_ref (loop
, &refs
,
517 if (REFERENCE_CLASS_P (lhs
))
518 *no_other_refs
&= gather_memory_references_ref (loop
, &refs
,
527 /* Prune the prefetch candidate REF using the self-reuse. */
530 prune_ref_by_self_reuse (struct mem_ref
*ref
)
532 HOST_WIDE_INT step
= ref
->group
->step
;
533 bool backward
= step
< 0;
537 /* Prefetch references to invariant address just once. */
538 ref
->prefetch_before
= 1;
545 if (step
> PREFETCH_BLOCK
)
548 if ((backward
&& HAVE_BACKWARD_PREFETCH
)
549 || (!backward
&& HAVE_FORWARD_PREFETCH
))
551 ref
->prefetch_before
= 1;
555 ref
->prefetch_mod
= PREFETCH_BLOCK
/ step
;
558 /* Divides X by BY, rounding down. */
561 ddown (HOST_WIDE_INT x
, unsigned HOST_WIDE_INT by
)
568 return (x
+ by
- 1) / by
;
571 /* Prune the prefetch candidate REF using the reuse with BY.
572 If BY_IS_BEFORE is true, BY is before REF in the loop. */
575 prune_ref_by_group_reuse (struct mem_ref
*ref
, struct mem_ref
*by
,
578 HOST_WIDE_INT step
= ref
->group
->step
;
579 bool backward
= step
< 0;
580 HOST_WIDE_INT delta_r
= ref
->delta
, delta_b
= by
->delta
;
581 HOST_WIDE_INT delta
= delta_b
- delta_r
;
582 HOST_WIDE_INT hit_from
;
583 unsigned HOST_WIDE_INT prefetch_before
, prefetch_block
;
587 /* If the references has the same address, only prefetch the
590 ref
->prefetch_before
= 0;
597 /* If the reference addresses are invariant and fall into the
598 same cache line, prefetch just the first one. */
602 if (ddown (ref
->delta
, PREFETCH_BLOCK
)
603 != ddown (by
->delta
, PREFETCH_BLOCK
))
606 ref
->prefetch_before
= 0;
610 /* Only prune the reference that is behind in the array. */
616 /* Transform the data so that we may assume that the accesses
620 delta_r
= PREFETCH_BLOCK
- 1 - delta_r
;
621 delta_b
= PREFETCH_BLOCK
- 1 - delta_b
;
629 /* Check whether the two references are likely to hit the same cache
630 line, and how distant the iterations in that it occurs are from
633 if (step
<= PREFETCH_BLOCK
)
635 /* The accesses are sure to meet. Let us check when. */
636 hit_from
= ddown (delta_b
, PREFETCH_BLOCK
) * PREFETCH_BLOCK
;
637 prefetch_before
= (hit_from
- delta_r
+ step
- 1) / step
;
639 if (prefetch_before
< ref
->prefetch_before
)
640 ref
->prefetch_before
= prefetch_before
;
645 /* A more complicated case. First let us ensure that size of cache line
646 and step are coprime (here we assume that PREFETCH_BLOCK is a power
648 prefetch_block
= PREFETCH_BLOCK
;
649 while ((step
& 1) == 0
650 && prefetch_block
> 1)
653 prefetch_block
>>= 1;
657 /* Now step > prefetch_block, and step and prefetch_block are coprime.
658 Determine the probability that the accesses hit the same cache line. */
660 prefetch_before
= delta
/ step
;
662 if ((unsigned HOST_WIDE_INT
) delta
663 <= (prefetch_block
* ACCEPTABLE_MISS_RATE
/ 1000))
665 if (prefetch_before
< ref
->prefetch_before
)
666 ref
->prefetch_before
= prefetch_before
;
671 /* Try also the following iteration. */
673 delta
= step
- delta
;
674 if ((unsigned HOST_WIDE_INT
) delta
675 <= (prefetch_block
* ACCEPTABLE_MISS_RATE
/ 1000))
677 if (prefetch_before
< ref
->prefetch_before
)
678 ref
->prefetch_before
= prefetch_before
;
683 /* The ref probably does not reuse by. */
687 /* Prune the prefetch candidate REF using the reuses with other references
691 prune_ref_by_reuse (struct mem_ref
*ref
, struct mem_ref
*refs
)
693 struct mem_ref
*prune_by
;
696 prune_ref_by_self_reuse (ref
);
698 for (prune_by
= refs
; prune_by
; prune_by
= prune_by
->next
)
706 if (!WRITE_CAN_USE_READ_PREFETCH
708 && !prune_by
->write_p
)
710 if (!READ_CAN_USE_WRITE_PREFETCH
712 && prune_by
->write_p
)
715 prune_ref_by_group_reuse (ref
, prune_by
, before
);
719 /* Prune the prefetch candidates in GROUP using the reuse analysis. */
722 prune_group_by_reuse (struct mem_ref_group
*group
)
724 struct mem_ref
*ref_pruned
;
726 for (ref_pruned
= group
->refs
; ref_pruned
; ref_pruned
= ref_pruned
->next
)
728 prune_ref_by_reuse (ref_pruned
, group
->refs
);
730 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
732 fprintf (dump_file
, "Reference %p:", (void *) ref_pruned
);
734 if (ref_pruned
->prefetch_before
== PREFETCH_ALL
735 && ref_pruned
->prefetch_mod
== 1)
736 fprintf (dump_file
, " no restrictions");
737 else if (ref_pruned
->prefetch_before
== 0)
738 fprintf (dump_file
, " do not prefetch");
739 else if (ref_pruned
->prefetch_before
<= ref_pruned
->prefetch_mod
)
740 fprintf (dump_file
, " prefetch once");
743 if (ref_pruned
->prefetch_before
!= PREFETCH_ALL
)
745 fprintf (dump_file
, " prefetch before ");
746 fprintf (dump_file
, HOST_WIDE_INT_PRINT_DEC
,
747 ref_pruned
->prefetch_before
);
749 if (ref_pruned
->prefetch_mod
!= 1)
751 fprintf (dump_file
, " prefetch mod ");
752 fprintf (dump_file
, HOST_WIDE_INT_PRINT_DEC
,
753 ref_pruned
->prefetch_mod
);
756 fprintf (dump_file
, "\n");
761 /* Prune the list of prefetch candidates GROUPS using the reuse analysis. */
764 prune_by_reuse (struct mem_ref_group
*groups
)
766 for (; groups
; groups
= groups
->next
)
767 prune_group_by_reuse (groups
);
770 /* Returns true if we should issue prefetch for REF. */
773 should_issue_prefetch_p (struct mem_ref
*ref
)
775 /* For now do not issue prefetches for only first few of the
777 if (ref
->prefetch_before
!= PREFETCH_ALL
)
780 /* Do not prefetch nontemporal stores. */
787 /* Decide which of the prefetch candidates in GROUPS to prefetch.
788 AHEAD is the number of iterations to prefetch ahead (which corresponds
789 to the number of simultaneous instances of one prefetch running at a
790 time). UNROLL_FACTOR is the factor by that the loop is going to be
791 unrolled. Returns true if there is anything to prefetch. */
794 schedule_prefetches (struct mem_ref_group
*groups
, unsigned unroll_factor
,
797 unsigned remaining_prefetch_slots
, n_prefetches
, prefetch_slots
;
798 unsigned slots_per_prefetch
;
802 /* At most SIMULTANEOUS_PREFETCHES should be running at the same time. */
803 remaining_prefetch_slots
= SIMULTANEOUS_PREFETCHES
;
805 /* The prefetch will run for AHEAD iterations of the original loop, i.e.,
806 AHEAD / UNROLL_FACTOR iterations of the unrolled loop. In each iteration,
807 it will need a prefetch slot. */
808 slots_per_prefetch
= (ahead
+ unroll_factor
/ 2) / unroll_factor
;
809 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
810 fprintf (dump_file
, "Each prefetch instruction takes %u prefetch slots.\n",
813 /* For now we just take memory references one by one and issue
814 prefetches for as many as possible. The groups are sorted
815 starting with the largest step, since the references with
816 large step are more likely to cause many cache misses. */
818 for (; groups
; groups
= groups
->next
)
819 for (ref
= groups
->refs
; ref
; ref
= ref
->next
)
821 if (!should_issue_prefetch_p (ref
))
824 /* If we need to prefetch the reference each PREFETCH_MOD iterations,
825 and we unroll the loop UNROLL_FACTOR times, we need to insert
826 ceil (UNROLL_FACTOR / PREFETCH_MOD) instructions in each
828 n_prefetches
= ((unroll_factor
+ ref
->prefetch_mod
- 1)
829 / ref
->prefetch_mod
);
830 prefetch_slots
= n_prefetches
* slots_per_prefetch
;
832 /* If more than half of the prefetches would be lost anyway, do not
833 issue the prefetch. */
834 if (2 * remaining_prefetch_slots
< prefetch_slots
)
837 ref
->issue_prefetch_p
= true;
839 if (remaining_prefetch_slots
<= prefetch_slots
)
841 remaining_prefetch_slots
-= prefetch_slots
;
848 /* Determine whether there is any reference suitable for prefetching
852 anything_to_prefetch_p (struct mem_ref_group
*groups
)
856 for (; groups
; groups
= groups
->next
)
857 for (ref
= groups
->refs
; ref
; ref
= ref
->next
)
858 if (should_issue_prefetch_p (ref
))
864 /* Issue prefetches for the reference REF into loop as decided before.
865 HEAD is the number of iterations to prefetch ahead. UNROLL_FACTOR
866 is the factor by which LOOP was unrolled. */
869 issue_prefetch_ref (struct mem_ref
*ref
, unsigned unroll_factor
, unsigned ahead
)
872 tree addr
, addr_base
, prefetch
, write_p
, local
;
873 block_stmt_iterator bsi
;
874 unsigned n_prefetches
, ap
;
875 bool nontemporal
= ref
->reuse_distance
>= L2_CACHE_SIZE_BYTES
;
877 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
878 fprintf (dump_file
, "Issued%s prefetch for %p.\n",
879 nontemporal
? " nontemporal" : "",
882 bsi
= bsi_for_stmt (ref
->stmt
);
884 n_prefetches
= ((unroll_factor
+ ref
->prefetch_mod
- 1)
885 / ref
->prefetch_mod
);
886 addr_base
= build_fold_addr_expr_with_type (ref
->mem
, ptr_type_node
);
887 addr_base
= force_gimple_operand_bsi (&bsi
, unshare_expr (addr_base
),
888 true, NULL
, true, BSI_SAME_STMT
);
889 write_p
= ref
->write_p
? integer_one_node
: integer_zero_node
;
890 local
= build_int_cst (integer_type_node
, nontemporal
? 0 : 3);
892 for (ap
= 0; ap
< n_prefetches
; ap
++)
894 /* Determine the address to prefetch. */
895 delta
= (ahead
+ ap
* ref
->prefetch_mod
) * ref
->group
->step
;
896 addr
= fold_build2 (POINTER_PLUS_EXPR
, ptr_type_node
,
897 addr_base
, size_int (delta
));
898 addr
= force_gimple_operand_bsi (&bsi
, unshare_expr (addr
), true, NULL
,
899 true, BSI_SAME_STMT
);
901 /* Create the prefetch instruction. */
902 prefetch
= build_call_expr (built_in_decls
[BUILT_IN_PREFETCH
],
903 3, addr
, write_p
, local
);
904 bsi_insert_before (&bsi
, prefetch
, BSI_SAME_STMT
);
908 /* Issue prefetches for the references in GROUPS into loop as decided before.
909 HEAD is the number of iterations to prefetch ahead. UNROLL_FACTOR is the
910 factor by that LOOP was unrolled. */
913 issue_prefetches (struct mem_ref_group
*groups
,
914 unsigned unroll_factor
, unsigned ahead
)
918 for (; groups
; groups
= groups
->next
)
919 for (ref
= groups
->refs
; ref
; ref
= ref
->next
)
920 if (ref
->issue_prefetch_p
)
921 issue_prefetch_ref (ref
, unroll_factor
, ahead
);
924 /* Returns true if REF is a memory write for that a nontemporal store insn
928 nontemporal_store_p (struct mem_ref
*ref
)
930 enum machine_mode mode
;
933 /* REF must be a write that is not reused. We require it to be independent
934 on all other memory references in the loop, as the nontemporal stores may
935 be reordered with respect to other memory references. */
937 || !ref
->independent_p
938 || ref
->reuse_distance
< L2_CACHE_SIZE_BYTES
)
941 /* Check that we have the storent instruction for the mode. */
942 mode
= TYPE_MODE (TREE_TYPE (ref
->mem
));
946 code
= optab_handler (storent_optab
, mode
)->insn_code
;
947 return code
!= CODE_FOR_nothing
;
950 /* If REF is a nontemporal store, we mark the corresponding modify statement
951 and return true. Otherwise, we return false. */
954 mark_nontemporal_store (struct mem_ref
*ref
)
956 if (!nontemporal_store_p (ref
))
959 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
960 fprintf (dump_file
, "Marked reference %p as a nontemporal store.\n",
963 MOVE_NONTEMPORAL (ref
->stmt
) = true;
964 ref
->storent_p
= true;
969 /* Issue a memory fence instruction after LOOP. */
972 emit_mfence_after_loop (struct loop
*loop
)
974 VEC (edge
, heap
) *exits
= get_loop_exit_edges (loop
);
977 block_stmt_iterator bsi
;
980 for (i
= 0; VEC_iterate (edge
, exits
, i
, exit
); i
++)
982 call
= build_function_call_expr (FENCE_FOLLOWING_MOVNT
, NULL_TREE
);
984 if (!single_pred_p (exit
->dest
)
985 /* If possible, we prefer not to insert the fence on other paths
987 && !(exit
->flags
& EDGE_ABNORMAL
))
988 split_loop_exit_edge (exit
);
989 bsi
= bsi_after_labels (exit
->dest
);
991 bsi_insert_before (&bsi
, call
, BSI_NEW_STMT
);
992 mark_virtual_ops_for_renaming (call
);
995 VEC_free (edge
, heap
, exits
);
996 update_ssa (TODO_update_ssa_only_virtuals
);
999 /* Returns true if we can use storent in loop, false otherwise. */
1002 may_use_storent_in_loop_p (struct loop
*loop
)
1006 if (loop
->inner
!= NULL
)
1009 /* If we must issue a mfence insn after using storent, check that there
1010 is a suitable place for it at each of the loop exits. */
1011 if (FENCE_FOLLOWING_MOVNT
!= NULL_TREE
)
1013 VEC (edge
, heap
) *exits
= get_loop_exit_edges (loop
);
1017 for (i
= 0; VEC_iterate (edge
, exits
, i
, exit
); i
++)
1018 if ((exit
->flags
& EDGE_ABNORMAL
)
1019 && exit
->dest
== EXIT_BLOCK_PTR
)
1022 VEC_free (edge
, heap
, exits
);
1028 /* Marks nontemporal stores in LOOP. GROUPS contains the description of memory
1029 references in the loop. */
1032 mark_nontemporal_stores (struct loop
*loop
, struct mem_ref_group
*groups
)
1034 struct mem_ref
*ref
;
1037 if (!may_use_storent_in_loop_p (loop
))
1040 for (; groups
; groups
= groups
->next
)
1041 for (ref
= groups
->refs
; ref
; ref
= ref
->next
)
1042 any
|= mark_nontemporal_store (ref
);
1044 if (any
&& FENCE_FOLLOWING_MOVNT
!= NULL_TREE
)
1045 emit_mfence_after_loop (loop
);
1048 /* Determines whether we can profitably unroll LOOP FACTOR times, and if
1049 this is the case, fill in DESC by the description of number of
1053 should_unroll_loop_p (struct loop
*loop
, struct tree_niter_desc
*desc
,
1056 if (!can_unroll_loop_p (loop
, factor
, desc
))
1059 /* We only consider loops without control flow for unrolling. This is not
1060 a hard restriction -- tree_unroll_loop works with arbitrary loops
1061 as well; but the unrolling/prefetching is usually more profitable for
1062 loops consisting of a single basic block, and we want to limit the
1064 if (loop
->num_nodes
> 2)
1070 /* Determine the coefficient by that unroll LOOP, from the information
1071 contained in the list of memory references REFS. Description of
1072 umber of iterations of LOOP is stored to DESC. NINSNS is the number of
1073 insns of the LOOP. EST_NITER is the estimated number of iterations of
1074 the loop, or -1 if no estimate is available. */
1077 determine_unroll_factor (struct loop
*loop
, struct mem_ref_group
*refs
,
1078 unsigned ninsns
, struct tree_niter_desc
*desc
,
1079 HOST_WIDE_INT est_niter
)
1081 unsigned upper_bound
;
1082 unsigned nfactor
, factor
, mod_constraint
;
1083 struct mem_ref_group
*agp
;
1084 struct mem_ref
*ref
;
1086 /* First check whether the loop is not too large to unroll. We ignore
1087 PARAM_MAX_UNROLL_TIMES, because for small loops, it prevented us
1088 from unrolling them enough to make exactly one cache line covered by each
1089 iteration. Also, the goal of PARAM_MAX_UNROLL_TIMES is to prevent
1090 us from unrolling the loops too many times in cases where we only expect
1091 gains from better scheduling and decreasing loop overhead, which is not
1093 upper_bound
= PARAM_VALUE (PARAM_MAX_UNROLLED_INSNS
) / ninsns
;
1095 /* If we unrolled the loop more times than it iterates, the unrolled version
1096 of the loop would be never entered. */
1097 if (est_niter
>= 0 && est_niter
< (HOST_WIDE_INT
) upper_bound
)
1098 upper_bound
= est_niter
;
1100 if (upper_bound
<= 1)
1103 /* Choose the factor so that we may prefetch each cache just once,
1104 but bound the unrolling by UPPER_BOUND. */
1106 for (agp
= refs
; agp
; agp
= agp
->next
)
1107 for (ref
= agp
->refs
; ref
; ref
= ref
->next
)
1108 if (should_issue_prefetch_p (ref
))
1110 mod_constraint
= ref
->prefetch_mod
;
1111 nfactor
= least_common_multiple (mod_constraint
, factor
);
1112 if (nfactor
<= upper_bound
)
1116 if (!should_unroll_loop_p (loop
, desc
, factor
))
1122 /* Returns the total volume of the memory references REFS, taking into account
1123 reuses in the innermost loop and cache line size. TODO -- we should also
1124 take into account reuses across the iterations of the loops in the loop
1128 volume_of_references (struct mem_ref_group
*refs
)
1130 unsigned volume
= 0;
1131 struct mem_ref_group
*gr
;
1132 struct mem_ref
*ref
;
1134 for (gr
= refs
; gr
; gr
= gr
->next
)
1135 for (ref
= gr
->refs
; ref
; ref
= ref
->next
)
1137 /* Almost always reuses another value? */
1138 if (ref
->prefetch_before
!= PREFETCH_ALL
)
1141 /* If several iterations access the same cache line, use the size of
1142 the line divided by this number. Otherwise, a cache line is
1143 accessed in each iteration. TODO -- in the latter case, we should
1144 take the size of the reference into account, rounding it up on cache
1145 line size multiple. */
1146 volume
+= L1_CACHE_LINE_SIZE
/ ref
->prefetch_mod
;
1151 /* Returns the volume of memory references accessed across VEC iterations of
1152 loops, whose sizes are described in the LOOP_SIZES array. N is the number
1153 of the loops in the nest (length of VEC and LOOP_SIZES vectors). */
1156 volume_of_dist_vector (lambda_vector vec
, unsigned *loop_sizes
, unsigned n
)
1160 for (i
= 0; i
< n
; i
++)
1167 gcc_assert (vec
[i
] > 0);
1169 /* We ignore the parts of the distance vector in subloops, since usually
1170 the numbers of iterations are much smaller. */
1171 return loop_sizes
[i
] * vec
[i
];
1174 /* Add the steps of ACCESS_FN multiplied by STRIDE to the array STRIDE
1175 at the position corresponding to the loop of the step. N is the depth
1176 of the considered loop nest, and, LOOP is its innermost loop. */
1179 add_subscript_strides (tree access_fn
, unsigned stride
,
1180 HOST_WIDE_INT
*strides
, unsigned n
, struct loop
*loop
)
1184 HOST_WIDE_INT astep
;
1185 unsigned min_depth
= loop_depth (loop
) - n
;
1187 while (TREE_CODE (access_fn
) == POLYNOMIAL_CHREC
)
1189 aloop
= get_chrec_loop (access_fn
);
1190 step
= CHREC_RIGHT (access_fn
);
1191 access_fn
= CHREC_LEFT (access_fn
);
1193 if ((unsigned) loop_depth (aloop
) <= min_depth
)
1196 if (host_integerp (step
, 0))
1197 astep
= tree_low_cst (step
, 0);
1199 astep
= L1_CACHE_LINE_SIZE
;
1201 strides
[n
- 1 - loop_depth (loop
) + loop_depth (aloop
)] += astep
* stride
;
1206 /* Returns the volume of memory references accessed between two consecutive
1207 self-reuses of the reference DR. We consider the subscripts of DR in N
1208 loops, and LOOP_SIZES contains the volumes of accesses in each of the
1209 loops. LOOP is the innermost loop of the current loop nest. */
1212 self_reuse_distance (data_reference_p dr
, unsigned *loop_sizes
, unsigned n
,
1215 tree stride
, access_fn
;
1216 HOST_WIDE_INT
*strides
, astride
;
1217 VEC (tree
, heap
) *access_fns
;
1218 tree ref
= DR_REF (dr
);
1219 unsigned i
, ret
= ~0u;
1221 /* In the following example:
1223 for (i = 0; i < N; i++)
1224 for (j = 0; j < N; j++)
1226 the same cache line is accessed each N steps (except if the change from
1227 i to i + 1 crosses the boundary of the cache line). Thus, for self-reuse,
1228 we cannot rely purely on the results of the data dependence analysis.
1230 Instead, we compute the stride of the reference in each loop, and consider
1231 the innermost loop in that the stride is less than cache size. */
1233 strides
= XCNEWVEC (HOST_WIDE_INT
, n
);
1234 access_fns
= DR_ACCESS_FNS (dr
);
1236 for (i
= 0; VEC_iterate (tree
, access_fns
, i
, access_fn
); i
++)
1238 /* Keep track of the reference corresponding to the subscript, so that we
1240 while (handled_component_p (ref
) && TREE_CODE (ref
) != ARRAY_REF
)
1241 ref
= TREE_OPERAND (ref
, 0);
1243 if (TREE_CODE (ref
) == ARRAY_REF
)
1245 stride
= TYPE_SIZE_UNIT (TREE_TYPE (ref
));
1246 if (host_integerp (stride
, 1))
1247 astride
= tree_low_cst (stride
, 1);
1249 astride
= L1_CACHE_LINE_SIZE
;
1251 ref
= TREE_OPERAND (ref
, 0);
1256 add_subscript_strides (access_fn
, astride
, strides
, n
, loop
);
1259 for (i
= n
; i
-- > 0; )
1261 unsigned HOST_WIDE_INT s
;
1263 s
= strides
[i
] < 0 ? -strides
[i
] : strides
[i
];
1265 if (s
< (unsigned) L1_CACHE_LINE_SIZE
1267 > (unsigned) (L1_CACHE_SIZE_BYTES
/ NONTEMPORAL_FRACTION
)))
1269 ret
= loop_sizes
[i
];
1278 /* Determines the distance till the first reuse of each reference in REFS
1279 in the loop nest of LOOP. NO_OTHER_REFS is true if there are no other
1280 memory references in the loop. */
1283 determine_loop_nest_reuse (struct loop
*loop
, struct mem_ref_group
*refs
,
1286 struct loop
*nest
, *aloop
;
1287 VEC (data_reference_p
, heap
) *datarefs
= NULL
;
1288 VEC (ddr_p
, heap
) *dependences
= NULL
;
1289 struct mem_ref_group
*gr
;
1290 struct mem_ref
*ref
, *refb
;
1291 VEC (loop_p
, heap
) *vloops
= NULL
;
1292 unsigned *loop_data_size
;
1294 unsigned volume
, dist
, adist
;
1296 data_reference_p dr
;
1302 /* Find the outermost loop of the loop nest of loop (we require that
1303 there are no sibling loops inside the nest). */
1307 aloop
= loop_outer (nest
);
1309 if (aloop
== current_loops
->tree_root
1310 || aloop
->inner
->next
)
1316 /* For each loop, determine the amount of data accessed in each iteration.
1317 We use this to estimate whether the reference is evicted from the
1318 cache before its reuse. */
1319 find_loop_nest (nest
, &vloops
);
1320 n
= VEC_length (loop_p
, vloops
);
1321 loop_data_size
= XNEWVEC (unsigned, n
);
1322 volume
= volume_of_references (refs
);
1326 loop_data_size
[i
] = volume
;
1327 /* Bound the volume by the L2 cache size, since above this bound,
1328 all dependence distances are equivalent. */
1329 if (volume
> L2_CACHE_SIZE_BYTES
)
1332 aloop
= VEC_index (loop_p
, vloops
, i
);
1333 vol
= estimated_loop_iterations_int (aloop
, false);
1335 vol
= expected_loop_iterations (aloop
);
1339 /* Prepare the references in the form suitable for data dependence
1340 analysis. We ignore unanalyzable data references (the results
1341 are used just as a heuristics to estimate temporality of the
1342 references, hence we do not need to worry about correctness). */
1343 for (gr
= refs
; gr
; gr
= gr
->next
)
1344 for (ref
= gr
->refs
; ref
; ref
= ref
->next
)
1346 dr
= create_data_ref (nest
, ref
->mem
, ref
->stmt
, !ref
->write_p
);
1350 ref
->reuse_distance
= volume
;
1352 VEC_safe_push (data_reference_p
, heap
, datarefs
, dr
);
1355 no_other_refs
= false;
1358 for (i
= 0; VEC_iterate (data_reference_p
, datarefs
, i
, dr
); i
++)
1360 dist
= self_reuse_distance (dr
, loop_data_size
, n
, loop
);
1361 ref
= (struct mem_ref
*) dr
->aux
;
1362 if (ref
->reuse_distance
> dist
)
1363 ref
->reuse_distance
= dist
;
1366 ref
->independent_p
= true;
1369 compute_all_dependences (datarefs
, &dependences
, vloops
, true);
1371 for (i
= 0; VEC_iterate (ddr_p
, dependences
, i
, dep
); i
++)
1373 if (DDR_ARE_DEPENDENT (dep
) == chrec_known
)
1376 ref
= (struct mem_ref
*) DDR_A (dep
)->aux
;
1377 refb
= (struct mem_ref
*) DDR_B (dep
)->aux
;
1379 if (DDR_ARE_DEPENDENT (dep
) == chrec_dont_know
1380 || DDR_NUM_DIST_VECTS (dep
) == 0)
1382 /* If the dependence cannot be analyzed, assume that there might be
1386 ref
->independent_p
= false;
1387 refb
->independent_p
= false;
1391 /* The distance vectors are normalized to be always lexicographically
1392 positive, hence we cannot tell just from them whether DDR_A comes
1393 before DDR_B or vice versa. However, it is not important,
1394 anyway -- if DDR_A is close to DDR_B, then it is either reused in
1395 DDR_B (and it is not nontemporal), or it reuses the value of DDR_B
1396 in cache (and marking it as nontemporal would not affect
1400 for (j
= 0; j
< DDR_NUM_DIST_VECTS (dep
); j
++)
1402 adist
= volume_of_dist_vector (DDR_DIST_VECT (dep
, j
),
1405 /* If this is a dependence in the innermost loop (i.e., the
1406 distances in all superloops are zero) and it is not
1407 the trivial self-dependence with distance zero, record that
1408 the references are not completely independent. */
1409 if (lambda_vector_zerop (DDR_DIST_VECT (dep
, j
), n
- 1)
1411 || DDR_DIST_VECT (dep
, j
)[n
-1] != 0))
1413 ref
->independent_p
= false;
1414 refb
->independent_p
= false;
1417 /* Ignore accesses closer than
1418 L1_CACHE_SIZE_BYTES / NONTEMPORAL_FRACTION,
1419 so that we use nontemporal prefetches e.g. if single memory
1420 location is accessed several times in a single iteration of
1422 if (adist
< L1_CACHE_SIZE_BYTES
/ NONTEMPORAL_FRACTION
)
1430 if (ref
->reuse_distance
> dist
)
1431 ref
->reuse_distance
= dist
;
1432 if (refb
->reuse_distance
> dist
)
1433 refb
->reuse_distance
= dist
;
1436 free_dependence_relations (dependences
);
1437 free_data_refs (datarefs
);
1438 free (loop_data_size
);
1440 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1442 fprintf (dump_file
, "Reuse distances:\n");
1443 for (gr
= refs
; gr
; gr
= gr
->next
)
1444 for (ref
= gr
->refs
; ref
; ref
= ref
->next
)
1445 fprintf (dump_file
, " ref %p distance %u\n",
1446 (void *) ref
, ref
->reuse_distance
);
1450 /* Issue prefetch instructions for array references in LOOP. Returns
1451 true if the LOOP was unrolled. */
1454 loop_prefetch_arrays (struct loop
*loop
)
1456 struct mem_ref_group
*refs
;
1457 unsigned ahead
, ninsns
, time
, unroll_factor
;
1458 HOST_WIDE_INT est_niter
;
1459 struct tree_niter_desc desc
;
1460 bool unrolled
= false, no_other_refs
;
1462 if (!maybe_hot_bb_p (loop
->header
))
1464 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1465 fprintf (dump_file
, " ignored (cold area)\n");
1469 /* Step 1: gather the memory references. */
1470 refs
= gather_memory_references (loop
, &no_other_refs
);
1472 /* Step 2: estimate the reuse effects. */
1473 prune_by_reuse (refs
);
1475 if (!anything_to_prefetch_p (refs
))
1478 determine_loop_nest_reuse (loop
, refs
, no_other_refs
);
1480 /* Step 3: determine the ahead and unroll factor. */
1482 /* FIXME: the time should be weighted by the probabilities of the blocks in
1484 time
= tree_num_loop_insns (loop
, &eni_time_weights
);
1485 ahead
= (PREFETCH_LATENCY
+ time
- 1) / time
;
1486 est_niter
= estimated_loop_iterations_int (loop
, false);
1488 /* The prefetches will run for AHEAD iterations of the original loop. Unless
1489 the loop rolls at least AHEAD times, prefetching the references does not
1491 if (est_niter
>= 0 && est_niter
<= (HOST_WIDE_INT
) ahead
)
1493 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1495 "Not prefetching -- loop estimated to roll only %d times\n",
1500 mark_nontemporal_stores (loop
, refs
);
1502 ninsns
= tree_num_loop_insns (loop
, &eni_size_weights
);
1503 unroll_factor
= determine_unroll_factor (loop
, refs
, ninsns
, &desc
,
1505 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1506 fprintf (dump_file
, "Ahead %d, unroll factor %d\n", ahead
, unroll_factor
);
1508 /* Step 4: what to prefetch? */
1509 if (!schedule_prefetches (refs
, unroll_factor
, ahead
))
1512 /* Step 5: unroll the loop. TODO -- peeling of first and last few
1513 iterations so that we do not issue superfluous prefetches. */
1514 if (unroll_factor
!= 1)
1516 tree_unroll_loop (loop
, unroll_factor
,
1517 single_dom_exit (loop
), &desc
);
1521 /* Step 6: issue the prefetches. */
1522 issue_prefetches (refs
, unroll_factor
, ahead
);
1525 release_mem_refs (refs
);
1529 /* Issue prefetch instructions for array references in loops. */
1532 tree_ssa_prefetch_arrays (void)
1536 bool unrolled
= false;
1540 /* It is possible to ask compiler for say -mtune=i486 -march=pentium4.
1541 -mtune=i486 causes us having PREFETCH_BLOCK 0, since this is part
1542 of processor costs and i486 does not have prefetch, but
1543 -march=pentium4 causes HAVE_prefetch to be true. Ugh. */
1544 || PREFETCH_BLOCK
== 0)
1547 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1549 fprintf (dump_file
, "Prefetching parameters:\n");
1550 fprintf (dump_file
, " simultaneous prefetches: %d\n",
1551 SIMULTANEOUS_PREFETCHES
);
1552 fprintf (dump_file
, " prefetch latency: %d\n", PREFETCH_LATENCY
);
1553 fprintf (dump_file
, " prefetch block size: %d\n", PREFETCH_BLOCK
);
1554 fprintf (dump_file
, " L1 cache size: %d lines, %d kB\n",
1555 L1_CACHE_SIZE_BYTES
/ L1_CACHE_LINE_SIZE
, L1_CACHE_SIZE
);
1556 fprintf (dump_file
, " L1 cache line size: %d\n", L1_CACHE_LINE_SIZE
);
1557 fprintf (dump_file
, " L2 cache size: %d kB\n", L2_CACHE_SIZE
);
1558 fprintf (dump_file
, "\n");
1561 initialize_original_copy_tables ();
1563 if (!built_in_decls
[BUILT_IN_PREFETCH
])
1565 tree type
= build_function_type (void_type_node
,
1566 tree_cons (NULL_TREE
,
1567 const_ptr_type_node
,
1569 tree decl
= add_builtin_function ("__builtin_prefetch", type
,
1570 BUILT_IN_PREFETCH
, BUILT_IN_NORMAL
,
1572 DECL_IS_NOVOPS (decl
) = true;
1573 built_in_decls
[BUILT_IN_PREFETCH
] = decl
;
1576 /* We assume that size of cache line is a power of two, so verify this
1578 gcc_assert ((PREFETCH_BLOCK
& (PREFETCH_BLOCK
- 1)) == 0);
1580 FOR_EACH_LOOP (li
, loop
, LI_FROM_INNERMOST
)
1582 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1583 fprintf (dump_file
, "Processing loop %d:\n", loop
->num
);
1585 unrolled
|= loop_prefetch_arrays (loop
);
1587 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1588 fprintf (dump_file
, "\n\n");
1594 todo_flags
|= TODO_cleanup_cfg
;
1597 free_original_copy_tables ();