1 /* Predictive commoning.
2 Copyright (C) 2005, 2007, 2008, 2009, 2010, 2011
3 Free Software Foundation, Inc.
5 This file is part of GCC.
7 GCC is free software; you can redistribute it and/or modify it
8 under the terms of the GNU General Public License as published by the
9 Free Software Foundation; either version 3, or (at your option) any
12 GCC is distributed in the hope that it will be useful, but WITHOUT
13 ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
14 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
17 You should have received a copy of the GNU General Public License
18 along with GCC; see the file COPYING3. If not see
19 <http://www.gnu.org/licenses/>. */
21 /* This file implements the predictive commoning optimization. Predictive
22 commoning can be viewed as CSE around a loop, and with some improvements,
23 as generalized strength reduction-- i.e., reusing values computed in
24 earlier iterations of a loop in the later ones. So far, the pass only
25 handles the most useful case, that is, reusing values of memory references.
26 If you think this is all just a special case of PRE, you are sort of right;
27 however, concentrating on loops is simpler, and makes it possible to
28 incorporate data dependence analysis to detect the opportunities, perform
29 loop unrolling to avoid copies together with renaming immediately,
30 and if needed, we could also take register pressure into account.
32 Let us demonstrate what is done on an example:
34 for (i = 0; i < 100; i++)
36 a[i+2] = a[i] + a[i+1];
42 1) We find data references in the loop, and split them to mutually
43 independent groups (i.e., we find components of a data dependence
44 graph). We ignore read-read dependences whose distance is not constant.
45 (TODO -- we could also ignore antidependences). In this example, we
46 find the following groups:
48 a[i]{read}, a[i+1]{read}, a[i+2]{write}
49 b[10]{read}, b[10]{write}
50 c[99 - i]{read}, c[i]{write}
51 d[i + 1]{read}, d[i]{write}
53 2) Inside each of the group, we verify several conditions:
54 a) all the references must differ in indices only, and the indices
55 must all have the same step
56 b) the references must dominate loop latch (and thus, they must be
57 ordered by dominance relation).
58 c) the distance of the indices must be a small multiple of the step
59 We are then able to compute the difference of the references (# of
60 iterations before they point to the same place as the first of them).
61 Also, in case there are writes in the loop, we split the groups into
62 chains whose head is the write whose values are used by the reads in
63 the same chain. The chains are then processed independently,
64 making the further transformations simpler. Also, the shorter chains
65 need the same number of registers, but may require lower unrolling
66 factor in order to get rid of the copies on the loop latch.
68 In our example, we get the following chains (the chain for c is invalid).
70 a[i]{read,+0}, a[i+1]{read,-1}, a[i+2]{write,-2}
71 b[10]{read,+0}, b[10]{write,+0}
72 d[i + 1]{read,+0}, d[i]{write,+1}
74 3) For each read, we determine the read or write whose value it reuses,
75 together with the distance of this reuse. I.e. we take the last
76 reference before it with distance 0, or the last of the references
77 with the smallest positive distance to the read. Then, we remove
78 the references that are not used in any of these chains, discard the
79 empty groups, and propagate all the links so that they point to the
80 single root reference of the chain (adjusting their distance
81 appropriately). Some extra care needs to be taken for references with
82 step 0. In our example (the numbers indicate the distance of the
85 a[i] --> (*) 2, a[i+1] --> (*) 1, a[i+2] (*)
86 b[10] --> (*) 1, b[10] (*)
88 4) The chains are combined together if possible. If the corresponding
89 elements of two chains are always combined together with the same
90 operator, we remember just the result of this combination, instead
91 of remembering the values separately. We may need to perform
92 reassociation to enable combining, for example
94 e[i] + f[i+1] + e[i+1] + f[i]
96 can be reassociated as
98 (e[i] + f[i]) + (e[i+1] + f[i+1])
100 and we can combine the chains for e and f into one chain.
102 5) For each root reference (end of the chain) R, let N be maximum distance
103 of a reference reusing its value. Variables R0 upto RN are created,
104 together with phi nodes that transfer values from R1 .. RN to
106 Initial values are loaded to R0..R(N-1) (in case not all references
107 must necessarily be accessed and they may trap, we may fail here;
108 TODO sometimes, the loads could be guarded by a check for the number
109 of iterations). Values loaded/stored in roots are also copied to
110 RN. Other reads are replaced with the appropriate variable Ri.
111 Everything is put to SSA form.
113 As a small improvement, if R0 is dead after the root (i.e., all uses of
114 the value with the maximum distance dominate the root), we can avoid
115 creating RN and use R0 instead of it.
117 In our example, we get (only the parts concerning a and b are shown):
118 for (i = 0; i < 100; i++)
130 6) Factor F for unrolling is determined as the smallest common multiple of
131 (N + 1) for each root reference (N for references for that we avoided
132 creating RN). If F and the loop is small enough, loop is unrolled F
133 times. The stores to RN (R0) in the copies of the loop body are
134 periodically replaced with R0, R1, ... (R1, R2, ...), so that they can
135 be coalesced and the copies can be eliminated.
137 TODO -- copy propagation and other optimizations may change the live
138 ranges of the temporary registers and prevent them from being coalesced;
139 this may increase the register pressure.
141 In our case, F = 2 and the (main loop of the) result is
143 for (i = 0; i < ...; i += 2)
160 TODO -- stores killing other stores can be taken into account, e.g.,
161 for (i = 0; i < n; i++)
171 for (i = 0; i < n; i++)
181 The interesting part is that this would generalize store motion; still, since
182 sm is performed elsewhere, it does not seem that important.
184 Predictive commoning can be generalized for arbitrary computations (not
185 just memory loads), and also nontrivial transfer functions (e.g., replacing
186 i * i with ii_last + 2 * i + 1), to generalize strength reduction. */
190 #include "coretypes.h"
195 #include "tree-flow.h"
197 #include "tree-data-ref.h"
198 #include "tree-scalar-evolution.h"
199 #include "tree-chrec.h"
201 #include "tree-pretty-print.h"
202 #include "gimple-pretty-print.h"
203 #include "tree-pass.h"
204 #include "tree-affine.h"
205 #include "tree-inline.h"
207 /* The maximum number of iterations between the considered memory
210 #define MAX_DISTANCE (target_avail_regs < 16 ? 4 : 8)
212 /* Data references (or phi nodes that carry data reference values across
215 typedef struct dref_d
217 /* The reference itself. */
218 struct data_reference
*ref
;
220 /* The statement in that the reference appears. */
223 /* In case that STMT is a phi node, this field is set to the SSA name
224 defined by it in replace_phis_by_defined_names (in order to avoid
225 pointing to phi node that got reallocated in the meantime). */
226 tree name_defined_by_phi
;
228 /* Distance of the reference from the root of the chain (in number of
229 iterations of the loop). */
232 /* Number of iterations offset from the first reference in the component. */
235 /* Number of the reference in a component, in dominance ordering. */
238 /* True if the memory reference is always accessed when the loop is
240 unsigned always_accessed
: 1;
244 DEF_VEC_ALLOC_P (dref
, heap
);
246 /* Type of the chain of the references. */
250 /* The addresses of the references in the chain are constant. */
253 /* There are only loads in the chain. */
256 /* Root of the chain is store, the rest are loads. */
259 /* A combination of two chains. */
263 /* Chains of data references. */
267 /* Type of the chain. */
268 enum chain_type type
;
270 /* For combination chains, the operator and the two chains that are
271 combined, and the type of the result. */
274 struct chain
*ch1
, *ch2
;
276 /* The references in the chain. */
277 VEC(dref
,heap
) *refs
;
279 /* The maximum distance of the reference in the chain from the root. */
282 /* The variables used to copy the value throughout iterations. */
283 VEC(tree
,heap
) *vars
;
285 /* Initializers for the variables. */
286 VEC(tree
,heap
) *inits
;
288 /* True if there is a use of a variable with the maximal distance
289 that comes after the root in the loop. */
290 unsigned has_max_use_after
: 1;
292 /* True if all the memory references in the chain are always accessed. */
293 unsigned all_always_accessed
: 1;
295 /* True if this chain was combined together with some other chain. */
296 unsigned combined
: 1;
300 DEF_VEC_ALLOC_P (chain_p
, heap
);
302 /* Describes the knowledge about the step of the memory references in
307 /* The step is zero. */
310 /* The step is nonzero. */
313 /* The step may or may not be nonzero. */
317 /* Components of the data dependence graph. */
321 /* The references in the component. */
322 VEC(dref
,heap
) *refs
;
324 /* What we know about the step of the references in the component. */
325 enum ref_step_type comp_step
;
327 /* Next component in the list. */
328 struct component
*next
;
331 /* Bitmap of ssa names defined by looparound phi nodes covered by chains. */
333 static bitmap looparound_phis
;
335 /* Cache used by tree_to_aff_combination_expand. */
337 static struct pointer_map_t
*name_expansions
;
339 /* Dumps data reference REF to FILE. */
341 extern void dump_dref (FILE *, dref
);
343 dump_dref (FILE *file
, dref ref
)
348 print_generic_expr (file
, DR_REF (ref
->ref
), TDF_SLIM
);
349 fprintf (file
, " (id %u%s)\n", ref
->pos
,
350 DR_IS_READ (ref
->ref
) ? "" : ", write");
352 fprintf (file
, " offset ");
353 dump_double_int (file
, ref
->offset
, false);
354 fprintf (file
, "\n");
356 fprintf (file
, " distance %u\n", ref
->distance
);
360 if (gimple_code (ref
->stmt
) == GIMPLE_PHI
)
361 fprintf (file
, " looparound ref\n");
363 fprintf (file
, " combination ref\n");
364 fprintf (file
, " in statement ");
365 print_gimple_stmt (file
, ref
->stmt
, 0, TDF_SLIM
);
366 fprintf (file
, "\n");
367 fprintf (file
, " distance %u\n", ref
->distance
);
372 /* Dumps CHAIN to FILE. */
374 extern void dump_chain (FILE *, chain_p
);
376 dump_chain (FILE *file
, chain_p chain
)
379 const char *chain_type
;
386 chain_type
= "Load motion";
390 chain_type
= "Loads-only";
394 chain_type
= "Store-loads";
398 chain_type
= "Combination";
405 fprintf (file
, "%s chain %p%s\n", chain_type
, (void *) chain
,
406 chain
->combined
? " (combined)" : "");
407 if (chain
->type
!= CT_INVARIANT
)
408 fprintf (file
, " max distance %u%s\n", chain
->length
,
409 chain
->has_max_use_after
? "" : ", may reuse first");
411 if (chain
->type
== CT_COMBINATION
)
413 fprintf (file
, " equal to %p %s %p in type ",
414 (void *) chain
->ch1
, op_symbol_code (chain
->op
),
415 (void *) chain
->ch2
);
416 print_generic_expr (file
, chain
->rslt_type
, TDF_SLIM
);
417 fprintf (file
, "\n");
422 fprintf (file
, " vars");
423 FOR_EACH_VEC_ELT (tree
, chain
->vars
, i
, var
)
426 print_generic_expr (file
, var
, TDF_SLIM
);
428 fprintf (file
, "\n");
433 fprintf (file
, " inits");
434 FOR_EACH_VEC_ELT (tree
, chain
->inits
, i
, var
)
437 print_generic_expr (file
, var
, TDF_SLIM
);
439 fprintf (file
, "\n");
442 fprintf (file
, " references:\n");
443 FOR_EACH_VEC_ELT (dref
, chain
->refs
, i
, a
)
446 fprintf (file
, "\n");
449 /* Dumps CHAINS to FILE. */
451 extern void dump_chains (FILE *, VEC (chain_p
, heap
) *);
453 dump_chains (FILE *file
, VEC (chain_p
, heap
) *chains
)
458 FOR_EACH_VEC_ELT (chain_p
, chains
, i
, chain
)
459 dump_chain (file
, chain
);
462 /* Dumps COMP to FILE. */
464 extern void dump_component (FILE *, struct component
*);
466 dump_component (FILE *file
, struct component
*comp
)
471 fprintf (file
, "Component%s:\n",
472 comp
->comp_step
== RS_INVARIANT
? " (invariant)" : "");
473 FOR_EACH_VEC_ELT (dref
, comp
->refs
, i
, a
)
475 fprintf (file
, "\n");
478 /* Dumps COMPS to FILE. */
480 extern void dump_components (FILE *, struct component
*);
482 dump_components (FILE *file
, struct component
*comps
)
484 struct component
*comp
;
486 for (comp
= comps
; comp
; comp
= comp
->next
)
487 dump_component (file
, comp
);
490 /* Frees a chain CHAIN. */
493 release_chain (chain_p chain
)
501 FOR_EACH_VEC_ELT (dref
, chain
->refs
, i
, ref
)
504 VEC_free (dref
, heap
, chain
->refs
);
505 VEC_free (tree
, heap
, chain
->vars
);
506 VEC_free (tree
, heap
, chain
->inits
);
514 release_chains (VEC (chain_p
, heap
) *chains
)
519 FOR_EACH_VEC_ELT (chain_p
, chains
, i
, chain
)
520 release_chain (chain
);
521 VEC_free (chain_p
, heap
, chains
);
524 /* Frees a component COMP. */
527 release_component (struct component
*comp
)
529 VEC_free (dref
, heap
, comp
->refs
);
533 /* Frees list of components COMPS. */
536 release_components (struct component
*comps
)
538 struct component
*act
, *next
;
540 for (act
= comps
; act
; act
= next
)
543 release_component (act
);
547 /* Finds a root of tree given by FATHERS containing A, and performs path
551 component_of (unsigned fathers
[], unsigned a
)
555 for (root
= a
; root
!= fathers
[root
]; root
= fathers
[root
])
558 for (; a
!= root
; a
= n
)
567 /* Join operation for DFU. FATHERS gives the tree, SIZES are sizes of the
568 components, A and B are components to merge. */
571 merge_comps (unsigned fathers
[], unsigned sizes
[], unsigned a
, unsigned b
)
573 unsigned ca
= component_of (fathers
, a
);
574 unsigned cb
= component_of (fathers
, b
);
579 if (sizes
[ca
] < sizes
[cb
])
581 sizes
[cb
] += sizes
[ca
];
586 sizes
[ca
] += sizes
[cb
];
591 /* Returns true if A is a reference that is suitable for predictive commoning
592 in the innermost loop that contains it. REF_STEP is set according to the
593 step of the reference A. */
596 suitable_reference_p (struct data_reference
*a
, enum ref_step_type
*ref_step
)
598 tree ref
= DR_REF (a
), step
= DR_STEP (a
);
601 || TREE_THIS_VOLATILE (ref
)
602 || !is_gimple_reg_type (TREE_TYPE (ref
))
603 || tree_could_throw_p (ref
))
606 if (integer_zerop (step
))
607 *ref_step
= RS_INVARIANT
;
608 else if (integer_nonzerop (step
))
609 *ref_step
= RS_NONZERO
;
616 /* Stores DR_OFFSET (DR) + DR_INIT (DR) to OFFSET. */
619 aff_combination_dr_offset (struct data_reference
*dr
, aff_tree
*offset
)
623 tree_to_aff_combination_expand (DR_OFFSET (dr
), sizetype
, offset
,
625 aff_combination_const (&delta
, sizetype
, tree_to_double_int (DR_INIT (dr
)));
626 aff_combination_add (offset
, &delta
);
629 /* Determines number of iterations of the innermost enclosing loop before B
630 refers to exactly the same location as A and stores it to OFF. If A and
631 B do not have the same step, they never meet, or anything else fails,
632 returns false, otherwise returns true. Both A and B are assumed to
633 satisfy suitable_reference_p. */
636 determine_offset (struct data_reference
*a
, struct data_reference
*b
,
639 aff_tree diff
, baseb
, step
;
642 /* Check that both the references access the location in the same type. */
643 typea
= TREE_TYPE (DR_REF (a
));
644 typeb
= TREE_TYPE (DR_REF (b
));
645 if (!useless_type_conversion_p (typeb
, typea
))
648 /* Check whether the base address and the step of both references is the
650 if (!operand_equal_p (DR_STEP (a
), DR_STEP (b
), 0)
651 || !operand_equal_p (DR_BASE_ADDRESS (a
), DR_BASE_ADDRESS (b
), 0))
654 if (integer_zerop (DR_STEP (a
)))
656 /* If the references have loop invariant address, check that they access
657 exactly the same location. */
658 *off
= double_int_zero
;
659 return (operand_equal_p (DR_OFFSET (a
), DR_OFFSET (b
), 0)
660 && operand_equal_p (DR_INIT (a
), DR_INIT (b
), 0));
663 /* Compare the offsets of the addresses, and check whether the difference
664 is a multiple of step. */
665 aff_combination_dr_offset (a
, &diff
);
666 aff_combination_dr_offset (b
, &baseb
);
667 aff_combination_scale (&baseb
, double_int_minus_one
);
668 aff_combination_add (&diff
, &baseb
);
670 tree_to_aff_combination_expand (DR_STEP (a
), sizetype
,
671 &step
, &name_expansions
);
672 return aff_combination_constant_multiple_p (&diff
, &step
, off
);
675 /* Returns the last basic block in LOOP for that we are sure that
676 it is executed whenever the loop is entered. */
679 last_always_executed_block (struct loop
*loop
)
682 VEC (edge
, heap
) *exits
= get_loop_exit_edges (loop
);
684 basic_block last
= loop
->latch
;
686 FOR_EACH_VEC_ELT (edge
, exits
, i
, ex
)
687 last
= nearest_common_dominator (CDI_DOMINATORS
, last
, ex
->src
);
688 VEC_free (edge
, heap
, exits
);
693 /* Splits dependence graph on DATAREFS described by DEPENDS to components. */
695 static struct component
*
696 split_data_refs_to_components (struct loop
*loop
,
697 VEC (data_reference_p
, heap
) *datarefs
,
698 VEC (ddr_p
, heap
) *depends
)
700 unsigned i
, n
= VEC_length (data_reference_p
, datarefs
);
701 unsigned ca
, ia
, ib
, bad
;
702 unsigned *comp_father
= XNEWVEC (unsigned, n
+ 1);
703 unsigned *comp_size
= XNEWVEC (unsigned, n
+ 1);
704 struct component
**comps
;
705 struct data_reference
*dr
, *dra
, *drb
;
706 struct data_dependence_relation
*ddr
;
707 struct component
*comp_list
= NULL
, *comp
;
709 basic_block last_always_executed
= last_always_executed_block (loop
);
711 FOR_EACH_VEC_ELT (data_reference_p
, datarefs
, i
, dr
)
715 /* A fake reference for call or asm_expr that may clobber memory;
719 dr
->aux
= (void *) (size_t) i
;
724 /* A component reserved for the "bad" data references. */
728 FOR_EACH_VEC_ELT (data_reference_p
, datarefs
, i
, dr
)
730 enum ref_step_type dummy
;
732 if (!suitable_reference_p (dr
, &dummy
))
734 ia
= (unsigned) (size_t) dr
->aux
;
735 merge_comps (comp_father
, comp_size
, n
, ia
);
739 FOR_EACH_VEC_ELT (ddr_p
, depends
, i
, ddr
)
741 double_int dummy_off
;
743 if (DDR_ARE_DEPENDENT (ddr
) == chrec_known
)
748 ia
= component_of (comp_father
, (unsigned) (size_t) dra
->aux
);
749 ib
= component_of (comp_father
, (unsigned) (size_t) drb
->aux
);
753 bad
= component_of (comp_father
, n
);
755 /* If both A and B are reads, we may ignore unsuitable dependences. */
756 if (DR_IS_READ (dra
) && DR_IS_READ (drb
)
757 && (ia
== bad
|| ib
== bad
758 || !determine_offset (dra
, drb
, &dummy_off
)))
761 merge_comps (comp_father
, comp_size
, ia
, ib
);
764 comps
= XCNEWVEC (struct component
*, n
);
765 bad
= component_of (comp_father
, n
);
766 FOR_EACH_VEC_ELT (data_reference_p
, datarefs
, i
, dr
)
768 ia
= (unsigned) (size_t) dr
->aux
;
769 ca
= component_of (comp_father
, ia
);
776 comp
= XCNEW (struct component
);
777 comp
->refs
= VEC_alloc (dref
, heap
, comp_size
[ca
]);
781 dataref
= XCNEW (struct dref_d
);
783 dataref
->stmt
= DR_STMT (dr
);
784 dataref
->offset
= double_int_zero
;
785 dataref
->distance
= 0;
787 dataref
->always_accessed
788 = dominated_by_p (CDI_DOMINATORS
, last_always_executed
,
789 gimple_bb (dataref
->stmt
));
790 dataref
->pos
= VEC_length (dref
, comp
->refs
);
791 VEC_quick_push (dref
, comp
->refs
, dataref
);
794 for (i
= 0; i
< n
; i
++)
799 comp
->next
= comp_list
;
811 /* Returns true if the component COMP satisfies the conditions
812 described in 2) at the beginning of this file. LOOP is the current
816 suitable_component_p (struct loop
*loop
, struct component
*comp
)
820 basic_block ba
, bp
= loop
->header
;
821 bool ok
, has_write
= false;
823 FOR_EACH_VEC_ELT (dref
, comp
->refs
, i
, a
)
825 ba
= gimple_bb (a
->stmt
);
827 if (!just_once_each_iteration_p (loop
, ba
))
830 gcc_assert (dominated_by_p (CDI_DOMINATORS
, ba
, bp
));
833 if (DR_IS_WRITE (a
->ref
))
837 first
= VEC_index (dref
, comp
->refs
, 0);
838 ok
= suitable_reference_p (first
->ref
, &comp
->comp_step
);
840 first
->offset
= double_int_zero
;
842 for (i
= 1; VEC_iterate (dref
, comp
->refs
, i
, a
); i
++)
844 if (!determine_offset (first
->ref
, a
->ref
, &a
->offset
))
847 #ifdef ENABLE_CHECKING
849 enum ref_step_type a_step
;
850 ok
= suitable_reference_p (a
->ref
, &a_step
);
851 gcc_assert (ok
&& a_step
== comp
->comp_step
);
856 /* If there is a write inside the component, we must know whether the
857 step is nonzero or not -- we would not otherwise be able to recognize
858 whether the value accessed by reads comes from the OFFSET-th iteration
859 or the previous one. */
860 if (has_write
&& comp
->comp_step
== RS_ANY
)
866 /* Check the conditions on references inside each of components COMPS,
867 and remove the unsuitable components from the list. The new list
868 of components is returned. The conditions are described in 2) at
869 the beginning of this file. LOOP is the current loop. */
871 static struct component
*
872 filter_suitable_components (struct loop
*loop
, struct component
*comps
)
874 struct component
**comp
, *act
;
876 for (comp
= &comps
; *comp
; )
879 if (suitable_component_p (loop
, act
))
887 FOR_EACH_VEC_ELT (dref
, act
->refs
, i
, ref
)
889 release_component (act
);
896 /* Compares two drefs A and B by their offset and position. Callback for
900 order_drefs (const void *a
, const void *b
)
902 const dref
*const da
= (const dref
*) a
;
903 const dref
*const db
= (const dref
*) b
;
904 int offcmp
= double_int_scmp ((*da
)->offset
, (*db
)->offset
);
909 return (*da
)->pos
- (*db
)->pos
;
912 /* Returns root of the CHAIN. */
915 get_chain_root (chain_p chain
)
917 return VEC_index (dref
, chain
->refs
, 0);
920 /* Adds REF to the chain CHAIN. */
923 add_ref_to_chain (chain_p chain
, dref ref
)
925 dref root
= get_chain_root (chain
);
928 gcc_assert (double_int_scmp (root
->offset
, ref
->offset
) <= 0);
929 dist
= double_int_sub (ref
->offset
, root
->offset
);
930 if (double_int_ucmp (uhwi_to_double_int (MAX_DISTANCE
), dist
) <= 0)
935 gcc_assert (double_int_fits_in_uhwi_p (dist
));
937 VEC_safe_push (dref
, heap
, chain
->refs
, ref
);
939 ref
->distance
= double_int_to_uhwi (dist
);
941 if (ref
->distance
>= chain
->length
)
943 chain
->length
= ref
->distance
;
944 chain
->has_max_use_after
= false;
947 if (ref
->distance
== chain
->length
948 && ref
->pos
> root
->pos
)
949 chain
->has_max_use_after
= true;
951 chain
->all_always_accessed
&= ref
->always_accessed
;
954 /* Returns the chain for invariant component COMP. */
957 make_invariant_chain (struct component
*comp
)
959 chain_p chain
= XCNEW (struct chain
);
963 chain
->type
= CT_INVARIANT
;
965 chain
->all_always_accessed
= true;
967 FOR_EACH_VEC_ELT (dref
, comp
->refs
, i
, ref
)
969 VEC_safe_push (dref
, heap
, chain
->refs
, ref
);
970 chain
->all_always_accessed
&= ref
->always_accessed
;
976 /* Make a new chain rooted at REF. */
979 make_rooted_chain (dref ref
)
981 chain_p chain
= XCNEW (struct chain
);
983 chain
->type
= DR_IS_READ (ref
->ref
) ? CT_LOAD
: CT_STORE_LOAD
;
985 VEC_safe_push (dref
, heap
, chain
->refs
, ref
);
986 chain
->all_always_accessed
= ref
->always_accessed
;
993 /* Returns true if CHAIN is not trivial. */
996 nontrivial_chain_p (chain_p chain
)
998 return chain
!= NULL
&& VEC_length (dref
, chain
->refs
) > 1;
1001 /* Returns the ssa name that contains the value of REF, or NULL_TREE if there
1005 name_for_ref (dref ref
)
1009 if (is_gimple_assign (ref
->stmt
))
1011 if (!ref
->ref
|| DR_IS_READ (ref
->ref
))
1012 name
= gimple_assign_lhs (ref
->stmt
);
1014 name
= gimple_assign_rhs1 (ref
->stmt
);
1017 name
= PHI_RESULT (ref
->stmt
);
1019 return (TREE_CODE (name
) == SSA_NAME
? name
: NULL_TREE
);
1022 /* Returns true if REF is a valid initializer for ROOT with given DISTANCE (in
1023 iterations of the innermost enclosing loop). */
1026 valid_initializer_p (struct data_reference
*ref
,
1027 unsigned distance
, struct data_reference
*root
)
1029 aff_tree diff
, base
, step
;
1032 /* Both REF and ROOT must be accessing the same object. */
1033 if (!operand_equal_p (DR_BASE_ADDRESS (ref
), DR_BASE_ADDRESS (root
), 0))
1036 /* The initializer is defined outside of loop, hence its address must be
1037 invariant inside the loop. */
1038 gcc_assert (integer_zerop (DR_STEP (ref
)));
1040 /* If the address of the reference is invariant, initializer must access
1041 exactly the same location. */
1042 if (integer_zerop (DR_STEP (root
)))
1043 return (operand_equal_p (DR_OFFSET (ref
), DR_OFFSET (root
), 0)
1044 && operand_equal_p (DR_INIT (ref
), DR_INIT (root
), 0));
1046 /* Verify that this index of REF is equal to the root's index at
1047 -DISTANCE-th iteration. */
1048 aff_combination_dr_offset (root
, &diff
);
1049 aff_combination_dr_offset (ref
, &base
);
1050 aff_combination_scale (&base
, double_int_minus_one
);
1051 aff_combination_add (&diff
, &base
);
1053 tree_to_aff_combination_expand (DR_STEP (root
), sizetype
, &step
,
1055 if (!aff_combination_constant_multiple_p (&diff
, &step
, &off
))
1058 if (!double_int_equal_p (off
, uhwi_to_double_int (distance
)))
1064 /* Finds looparound phi node of LOOP that copies the value of REF, and if its
1065 initial value is correct (equal to initial value of REF shifted by one
1066 iteration), returns the phi node. Otherwise, NULL_TREE is returned. ROOT
1067 is the root of the current chain. */
1070 find_looparound_phi (struct loop
*loop
, dref ref
, dref root
)
1072 tree name
, init
, init_ref
;
1073 gimple phi
= NULL
, init_stmt
;
1074 edge latch
= loop_latch_edge (loop
);
1075 struct data_reference init_dr
;
1076 gimple_stmt_iterator psi
;
1078 if (is_gimple_assign (ref
->stmt
))
1080 if (DR_IS_READ (ref
->ref
))
1081 name
= gimple_assign_lhs (ref
->stmt
);
1083 name
= gimple_assign_rhs1 (ref
->stmt
);
1086 name
= PHI_RESULT (ref
->stmt
);
1090 for (psi
= gsi_start_phis (loop
->header
); !gsi_end_p (psi
); gsi_next (&psi
))
1092 phi
= gsi_stmt (psi
);
1093 if (PHI_ARG_DEF_FROM_EDGE (phi
, latch
) == name
)
1097 if (gsi_end_p (psi
))
1100 init
= PHI_ARG_DEF_FROM_EDGE (phi
, loop_preheader_edge (loop
));
1101 if (TREE_CODE (init
) != SSA_NAME
)
1103 init_stmt
= SSA_NAME_DEF_STMT (init
);
1104 if (gimple_code (init_stmt
) != GIMPLE_ASSIGN
)
1106 gcc_assert (gimple_assign_lhs (init_stmt
) == init
);
1108 init_ref
= gimple_assign_rhs1 (init_stmt
);
1109 if (!REFERENCE_CLASS_P (init_ref
)
1110 && !DECL_P (init_ref
))
1113 /* Analyze the behavior of INIT_REF with respect to LOOP (innermost
1114 loop enclosing PHI). */
1115 memset (&init_dr
, 0, sizeof (struct data_reference
));
1116 DR_REF (&init_dr
) = init_ref
;
1117 DR_STMT (&init_dr
) = phi
;
1118 if (!dr_analyze_innermost (&init_dr
))
1121 if (!valid_initializer_p (&init_dr
, ref
->distance
+ 1, root
->ref
))
1127 /* Adds a reference for the looparound copy of REF in PHI to CHAIN. */
1130 insert_looparound_copy (chain_p chain
, dref ref
, gimple phi
)
1132 dref nw
= XCNEW (struct dref_d
), aref
;
1136 nw
->distance
= ref
->distance
+ 1;
1137 nw
->always_accessed
= 1;
1139 FOR_EACH_VEC_ELT (dref
, chain
->refs
, i
, aref
)
1140 if (aref
->distance
>= nw
->distance
)
1142 VEC_safe_insert (dref
, heap
, chain
->refs
, i
, nw
);
1144 if (nw
->distance
> chain
->length
)
1146 chain
->length
= nw
->distance
;
1147 chain
->has_max_use_after
= false;
1151 /* For references in CHAIN that are copied around the LOOP (created previously
1152 by PRE, or by user), add the results of such copies to the chain. This
1153 enables us to remove the copies by unrolling, and may need less registers
1154 (also, it may allow us to combine chains together). */
1157 add_looparound_copies (struct loop
*loop
, chain_p chain
)
1160 dref ref
, root
= get_chain_root (chain
);
1163 FOR_EACH_VEC_ELT (dref
, chain
->refs
, i
, ref
)
1165 phi
= find_looparound_phi (loop
, ref
, root
);
1169 bitmap_set_bit (looparound_phis
, SSA_NAME_VERSION (PHI_RESULT (phi
)));
1170 insert_looparound_copy (chain
, ref
, phi
);
1174 /* Find roots of the values and determine distances in the component COMP.
1175 The references are redistributed into CHAINS. LOOP is the current
1179 determine_roots_comp (struct loop
*loop
,
1180 struct component
*comp
,
1181 VEC (chain_p
, heap
) **chains
)
1185 chain_p chain
= NULL
;
1186 double_int last_ofs
= double_int_zero
;
1188 /* Invariants are handled specially. */
1189 if (comp
->comp_step
== RS_INVARIANT
)
1191 chain
= make_invariant_chain (comp
);
1192 VEC_safe_push (chain_p
, heap
, *chains
, chain
);
1196 VEC_qsort (dref
, comp
->refs
, order_drefs
);
1198 FOR_EACH_VEC_ELT (dref
, comp
->refs
, i
, a
)
1200 if (!chain
|| DR_IS_WRITE (a
->ref
)
1201 || double_int_ucmp (uhwi_to_double_int (MAX_DISTANCE
),
1202 double_int_sub (a
->offset
, last_ofs
)) <= 0)
1204 if (nontrivial_chain_p (chain
))
1206 add_looparound_copies (loop
, chain
);
1207 VEC_safe_push (chain_p
, heap
, *chains
, chain
);
1210 release_chain (chain
);
1211 chain
= make_rooted_chain (a
);
1212 last_ofs
= a
->offset
;
1216 add_ref_to_chain (chain
, a
);
1219 if (nontrivial_chain_p (chain
))
1221 add_looparound_copies (loop
, chain
);
1222 VEC_safe_push (chain_p
, heap
, *chains
, chain
);
1225 release_chain (chain
);
1228 /* Find roots of the values and determine distances in components COMPS, and
1229 separates the references to CHAINS. LOOP is the current loop. */
1232 determine_roots (struct loop
*loop
,
1233 struct component
*comps
, VEC (chain_p
, heap
) **chains
)
1235 struct component
*comp
;
1237 for (comp
= comps
; comp
; comp
= comp
->next
)
1238 determine_roots_comp (loop
, comp
, chains
);
1241 /* Replace the reference in statement STMT with temporary variable
1242 NEW_TREE. If SET is true, NEW_TREE is instead initialized to the value of
1243 the reference in the statement. IN_LHS is true if the reference
1244 is in the lhs of STMT, false if it is in rhs. */
1247 replace_ref_with (gimple stmt
, tree new_tree
, bool set
, bool in_lhs
)
1251 gimple_stmt_iterator bsi
, psi
;
1253 if (gimple_code (stmt
) == GIMPLE_PHI
)
1255 gcc_assert (!in_lhs
&& !set
);
1257 val
= PHI_RESULT (stmt
);
1258 bsi
= gsi_after_labels (gimple_bb (stmt
));
1259 psi
= gsi_for_stmt (stmt
);
1260 remove_phi_node (&psi
, false);
1262 /* Turn the phi node into GIMPLE_ASSIGN. */
1263 new_stmt
= gimple_build_assign (val
, new_tree
);
1264 gsi_insert_before (&bsi
, new_stmt
, GSI_NEW_STMT
);
1268 /* Since the reference is of gimple_reg type, it should only
1269 appear as lhs or rhs of modify statement. */
1270 gcc_assert (is_gimple_assign (stmt
));
1272 bsi
= gsi_for_stmt (stmt
);
1274 /* If we do not need to initialize NEW_TREE, just replace the use of OLD. */
1277 gcc_assert (!in_lhs
);
1278 gimple_assign_set_rhs_from_tree (&bsi
, new_tree
);
1279 stmt
= gsi_stmt (bsi
);
1286 /* We have statement
1290 If OLD is a memory reference, then VAL is gimple_val, and we transform
1296 Otherwise, we are replacing a combination chain,
1297 VAL is the expression that performs the combination, and OLD is an
1298 SSA name. In this case, we transform the assignment to
1305 val
= gimple_assign_lhs (stmt
);
1306 if (TREE_CODE (val
) != SSA_NAME
)
1308 gcc_assert (gimple_assign_copy_p (stmt
));
1309 val
= gimple_assign_rhs1 (stmt
);
1321 val
= gimple_assign_lhs (stmt
);
1324 new_stmt
= gimple_build_assign (new_tree
, unshare_expr (val
));
1325 gsi_insert_after (&bsi
, new_stmt
, GSI_NEW_STMT
);
1328 /* Returns the reference to the address of REF in the ITER-th iteration of
1329 LOOP, or NULL if we fail to determine it (ITER may be negative). We
1330 try to preserve the original shape of the reference (not rewrite it
1331 as an indirect ref to the address), to make tree_could_trap_p in
1332 prepare_initializers_chain return false more often. */
1335 ref_at_iteration (struct loop
*loop
, tree ref
, int iter
)
1337 tree idx
, *idx_p
, type
, val
, op0
= NULL_TREE
, ret
;
1341 if (handled_component_p (ref
))
1343 op0
= ref_at_iteration (loop
, TREE_OPERAND (ref
, 0), iter
);
1347 else if (!INDIRECT_REF_P (ref
)
1348 && TREE_CODE (ref
) != MEM_REF
)
1349 return unshare_expr (ref
);
1351 if (TREE_CODE (ref
) == MEM_REF
)
1353 ret
= unshare_expr (ref
);
1354 idx
= TREE_OPERAND (ref
, 0);
1355 idx_p
= &TREE_OPERAND (ret
, 0);
1357 else if (TREE_CODE (ref
) == COMPONENT_REF
)
1359 /* Check that the offset is loop invariant. */
1360 if (TREE_OPERAND (ref
, 2)
1361 && !expr_invariant_in_loop_p (loop
, TREE_OPERAND (ref
, 2)))
1364 return build3 (COMPONENT_REF
, TREE_TYPE (ref
), op0
,
1365 unshare_expr (TREE_OPERAND (ref
, 1)),
1366 unshare_expr (TREE_OPERAND (ref
, 2)));
1368 else if (TREE_CODE (ref
) == ARRAY_REF
)
1370 /* Check that the lower bound and the step are loop invariant. */
1371 if (TREE_OPERAND (ref
, 2)
1372 && !expr_invariant_in_loop_p (loop
, TREE_OPERAND (ref
, 2)))
1374 if (TREE_OPERAND (ref
, 3)
1375 && !expr_invariant_in_loop_p (loop
, TREE_OPERAND (ref
, 3)))
1378 ret
= build4 (ARRAY_REF
, TREE_TYPE (ref
), op0
, NULL_TREE
,
1379 unshare_expr (TREE_OPERAND (ref
, 2)),
1380 unshare_expr (TREE_OPERAND (ref
, 3)));
1381 idx
= TREE_OPERAND (ref
, 1);
1382 idx_p
= &TREE_OPERAND (ret
, 1);
1387 ok
= simple_iv (loop
, loop
, idx
, &iv
, true);
1390 iv
.base
= expand_simple_operations (iv
.base
);
1391 if (integer_zerop (iv
.step
))
1392 *idx_p
= unshare_expr (iv
.base
);
1395 type
= TREE_TYPE (iv
.base
);
1396 if (POINTER_TYPE_P (type
))
1398 val
= fold_build2 (MULT_EXPR
, sizetype
, iv
.step
,
1400 val
= fold_build_pointer_plus (iv
.base
, val
);
1404 val
= fold_build2 (MULT_EXPR
, type
, iv
.step
,
1405 build_int_cst_type (type
, iter
));
1406 val
= fold_build2 (PLUS_EXPR
, type
, iv
.base
, val
);
1408 *idx_p
= unshare_expr (val
);
1414 /* Get the initialization expression for the INDEX-th temporary variable
1418 get_init_expr (chain_p chain
, unsigned index
)
1420 if (chain
->type
== CT_COMBINATION
)
1422 tree e1
= get_init_expr (chain
->ch1
, index
);
1423 tree e2
= get_init_expr (chain
->ch2
, index
);
1425 return fold_build2 (chain
->op
, chain
->rslt_type
, e1
, e2
);
1428 return VEC_index (tree
, chain
->inits
, index
);
1431 /* Marks all virtual operands of statement STMT for renaming. */
1434 mark_virtual_ops_for_renaming (gimple stmt
)
1438 if (gimple_code (stmt
) == GIMPLE_PHI
)
1440 var
= PHI_RESULT (stmt
);
1441 if (is_gimple_reg (var
))
1444 if (TREE_CODE (var
) == SSA_NAME
)
1445 var
= SSA_NAME_VAR (var
);
1446 mark_sym_for_renaming (var
);
1451 if (gimple_vuse (stmt
))
1452 mark_sym_for_renaming (gimple_vop (cfun
));
1455 /* Returns a new temporary variable used for the I-th variable carrying
1456 value of REF. The variable's uid is marked in TMP_VARS. */
1459 predcom_tmp_var (tree ref
, unsigned i
, bitmap tmp_vars
)
1461 tree type
= TREE_TYPE (ref
);
1462 /* We never access the components of the temporary variable in predictive
1464 tree var
= create_tmp_reg (type
, get_lsm_tmp_name (ref
, i
));
1466 add_referenced_var (var
);
1467 bitmap_set_bit (tmp_vars
, DECL_UID (var
));
1471 /* Creates the variables for CHAIN, as well as phi nodes for them and
1472 initialization on entry to LOOP. Uids of the newly created
1473 temporary variables are marked in TMP_VARS. */
1476 initialize_root_vars (struct loop
*loop
, chain_p chain
, bitmap tmp_vars
)
1479 unsigned n
= chain
->length
;
1480 dref root
= get_chain_root (chain
);
1481 bool reuse_first
= !chain
->has_max_use_after
;
1482 tree ref
, init
, var
, next
;
1485 edge entry
= loop_preheader_edge (loop
), latch
= loop_latch_edge (loop
);
1487 /* If N == 0, then all the references are within the single iteration. And
1488 since this is an nonempty chain, reuse_first cannot be true. */
1489 gcc_assert (n
> 0 || !reuse_first
);
1491 chain
->vars
= VEC_alloc (tree
, heap
, n
+ 1);
1493 if (chain
->type
== CT_COMBINATION
)
1494 ref
= gimple_assign_lhs (root
->stmt
);
1496 ref
= DR_REF (root
->ref
);
1498 for (i
= 0; i
< n
+ (reuse_first
? 0 : 1); i
++)
1500 var
= predcom_tmp_var (ref
, i
, tmp_vars
);
1501 VEC_quick_push (tree
, chain
->vars
, var
);
1504 VEC_quick_push (tree
, chain
->vars
, VEC_index (tree
, chain
->vars
, 0));
1506 FOR_EACH_VEC_ELT (tree
, chain
->vars
, i
, var
)
1507 VEC_replace (tree
, chain
->vars
, i
, make_ssa_name (var
, NULL
));
1509 for (i
= 0; i
< n
; i
++)
1511 var
= VEC_index (tree
, chain
->vars
, i
);
1512 next
= VEC_index (tree
, chain
->vars
, i
+ 1);
1513 init
= get_init_expr (chain
, i
);
1515 init
= force_gimple_operand (init
, &stmts
, true, NULL_TREE
);
1517 gsi_insert_seq_on_edge_immediate (entry
, stmts
);
1519 phi
= create_phi_node (var
, loop
->header
);
1520 SSA_NAME_DEF_STMT (var
) = phi
;
1521 add_phi_arg (phi
, init
, entry
, UNKNOWN_LOCATION
);
1522 add_phi_arg (phi
, next
, latch
, UNKNOWN_LOCATION
);
1526 /* Create the variables and initialization statement for root of chain
1527 CHAIN. Uids of the newly created temporary variables are marked
1531 initialize_root (struct loop
*loop
, chain_p chain
, bitmap tmp_vars
)
1533 dref root
= get_chain_root (chain
);
1534 bool in_lhs
= (chain
->type
== CT_STORE_LOAD
1535 || chain
->type
== CT_COMBINATION
);
1537 initialize_root_vars (loop
, chain
, tmp_vars
);
1538 replace_ref_with (root
->stmt
,
1539 VEC_index (tree
, chain
->vars
, chain
->length
),
1543 /* Initializes a variable for load motion for ROOT and prepares phi nodes and
1544 initialization on entry to LOOP if necessary. The ssa name for the variable
1545 is stored in VARS. If WRITTEN is true, also a phi node to copy its value
1546 around the loop is created. Uid of the newly created temporary variable
1547 is marked in TMP_VARS. INITS is the list containing the (single)
1551 initialize_root_vars_lm (struct loop
*loop
, dref root
, bool written
,
1552 VEC(tree
, heap
) **vars
, VEC(tree
, heap
) *inits
,
1556 tree ref
= DR_REF (root
->ref
), init
, var
, next
;
1559 edge entry
= loop_preheader_edge (loop
), latch
= loop_latch_edge (loop
);
1561 /* Find the initializer for the variable, and check that it cannot
1563 init
= VEC_index (tree
, inits
, 0);
1565 *vars
= VEC_alloc (tree
, heap
, written
? 2 : 1);
1566 var
= predcom_tmp_var (ref
, 0, tmp_vars
);
1567 VEC_quick_push (tree
, *vars
, var
);
1569 VEC_quick_push (tree
, *vars
, VEC_index (tree
, *vars
, 0));
1571 FOR_EACH_VEC_ELT (tree
, *vars
, i
, var
)
1572 VEC_replace (tree
, *vars
, i
, make_ssa_name (var
, NULL
));
1574 var
= VEC_index (tree
, *vars
, 0);
1576 init
= force_gimple_operand (init
, &stmts
, written
, NULL_TREE
);
1578 gsi_insert_seq_on_edge_immediate (entry
, stmts
);
1582 next
= VEC_index (tree
, *vars
, 1);
1583 phi
= create_phi_node (var
, loop
->header
);
1584 SSA_NAME_DEF_STMT (var
) = phi
;
1585 add_phi_arg (phi
, init
, entry
, UNKNOWN_LOCATION
);
1586 add_phi_arg (phi
, next
, latch
, UNKNOWN_LOCATION
);
1590 gimple init_stmt
= gimple_build_assign (var
, init
);
1591 mark_virtual_ops_for_renaming (init_stmt
);
1592 gsi_insert_on_edge_immediate (entry
, init_stmt
);
1597 /* Execute load motion for references in chain CHAIN. Uids of the newly
1598 created temporary variables are marked in TMP_VARS. */
1601 execute_load_motion (struct loop
*loop
, chain_p chain
, bitmap tmp_vars
)
1603 VEC (tree
, heap
) *vars
;
1605 unsigned n_writes
= 0, ridx
, i
;
1608 gcc_assert (chain
->type
== CT_INVARIANT
);
1609 gcc_assert (!chain
->combined
);
1610 FOR_EACH_VEC_ELT (dref
, chain
->refs
, i
, a
)
1611 if (DR_IS_WRITE (a
->ref
))
1614 /* If there are no reads in the loop, there is nothing to do. */
1615 if (n_writes
== VEC_length (dref
, chain
->refs
))
1618 initialize_root_vars_lm (loop
, get_chain_root (chain
), n_writes
> 0,
1619 &vars
, chain
->inits
, tmp_vars
);
1622 FOR_EACH_VEC_ELT (dref
, chain
->refs
, i
, a
)
1624 bool is_read
= DR_IS_READ (a
->ref
);
1625 mark_virtual_ops_for_renaming (a
->stmt
);
1627 if (DR_IS_WRITE (a
->ref
))
1632 var
= VEC_index (tree
, vars
, 0);
1633 var
= make_ssa_name (SSA_NAME_VAR (var
), NULL
);
1634 VEC_replace (tree
, vars
, 0, var
);
1640 replace_ref_with (a
->stmt
, VEC_index (tree
, vars
, ridx
),
1641 !is_read
, !is_read
);
1644 VEC_free (tree
, heap
, vars
);
1647 /* Returns the single statement in that NAME is used, excepting
1648 the looparound phi nodes contained in one of the chains. If there is no
1649 such statement, or more statements, NULL is returned. */
1652 single_nonlooparound_use (tree name
)
1655 imm_use_iterator it
;
1656 gimple stmt
, ret
= NULL
;
1658 FOR_EACH_IMM_USE_FAST (use
, it
, name
)
1660 stmt
= USE_STMT (use
);
1662 if (gimple_code (stmt
) == GIMPLE_PHI
)
1664 /* Ignore uses in looparound phi nodes. Uses in other phi nodes
1665 could not be processed anyway, so just fail for them. */
1666 if (bitmap_bit_p (looparound_phis
,
1667 SSA_NAME_VERSION (PHI_RESULT (stmt
))))
1672 else if (is_gimple_debug (stmt
))
1674 else if (ret
!= NULL
)
1683 /* Remove statement STMT, as well as the chain of assignments in that it is
1687 remove_stmt (gimple stmt
)
1691 gimple_stmt_iterator psi
;
1693 if (gimple_code (stmt
) == GIMPLE_PHI
)
1695 name
= PHI_RESULT (stmt
);
1696 next
= single_nonlooparound_use (name
);
1697 psi
= gsi_for_stmt (stmt
);
1698 remove_phi_node (&psi
, true);
1701 || !gimple_assign_ssa_name_copy_p (next
)
1702 || gimple_assign_rhs1 (next
) != name
)
1710 gimple_stmt_iterator bsi
;
1712 bsi
= gsi_for_stmt (stmt
);
1714 name
= gimple_assign_lhs (stmt
);
1715 gcc_assert (TREE_CODE (name
) == SSA_NAME
);
1717 next
= single_nonlooparound_use (name
);
1719 mark_virtual_ops_for_renaming (stmt
);
1720 gsi_remove (&bsi
, true);
1721 release_defs (stmt
);
1724 || !gimple_assign_ssa_name_copy_p (next
)
1725 || gimple_assign_rhs1 (next
) != name
)
1732 /* Perform the predictive commoning optimization for a chain CHAIN.
1733 Uids of the newly created temporary variables are marked in TMP_VARS.*/
1736 execute_pred_commoning_chain (struct loop
*loop
, chain_p chain
,
1743 if (chain
->combined
)
1745 /* For combined chains, just remove the statements that are used to
1746 compute the values of the expression (except for the root one). */
1747 for (i
= 1; VEC_iterate (dref
, chain
->refs
, i
, a
); i
++)
1748 remove_stmt (a
->stmt
);
1752 /* For non-combined chains, set up the variables that hold its value,
1753 and replace the uses of the original references by these
1755 root
= get_chain_root (chain
);
1756 mark_virtual_ops_for_renaming (root
->stmt
);
1758 initialize_root (loop
, chain
, tmp_vars
);
1759 for (i
= 1; VEC_iterate (dref
, chain
->refs
, i
, a
); i
++)
1761 mark_virtual_ops_for_renaming (a
->stmt
);
1762 var
= VEC_index (tree
, chain
->vars
, chain
->length
- a
->distance
);
1763 replace_ref_with (a
->stmt
, var
, false, false);
1768 /* Determines the unroll factor necessary to remove as many temporary variable
1769 copies as possible. CHAINS is the list of chains that will be
1773 determine_unroll_factor (VEC (chain_p
, heap
) *chains
)
1776 unsigned factor
= 1, af
, nfactor
, i
;
1777 unsigned max
= PARAM_VALUE (PARAM_MAX_UNROLL_TIMES
);
1779 FOR_EACH_VEC_ELT (chain_p
, chains
, i
, chain
)
1781 if (chain
->type
== CT_INVARIANT
|| chain
->combined
)
1784 /* The best unroll factor for this chain is equal to the number of
1785 temporary variables that we create for it. */
1787 if (chain
->has_max_use_after
)
1790 nfactor
= factor
* af
/ gcd (factor
, af
);
1798 /* Perform the predictive commoning optimization for CHAINS.
1799 Uids of the newly created temporary variables are marked in TMP_VARS. */
1802 execute_pred_commoning (struct loop
*loop
, VEC (chain_p
, heap
) *chains
,
1808 FOR_EACH_VEC_ELT (chain_p
, chains
, i
, chain
)
1810 if (chain
->type
== CT_INVARIANT
)
1811 execute_load_motion (loop
, chain
, tmp_vars
);
1813 execute_pred_commoning_chain (loop
, chain
, tmp_vars
);
1816 update_ssa (TODO_update_ssa_only_virtuals
);
1819 /* For each reference in CHAINS, if its defining statement is
1820 phi node, record the ssa name that is defined by it. */
1823 replace_phis_by_defined_names (VEC (chain_p
, heap
) *chains
)
1829 FOR_EACH_VEC_ELT (chain_p
, chains
, i
, chain
)
1830 FOR_EACH_VEC_ELT (dref
, chain
->refs
, j
, a
)
1832 if (gimple_code (a
->stmt
) == GIMPLE_PHI
)
1834 a
->name_defined_by_phi
= PHI_RESULT (a
->stmt
);
1840 /* For each reference in CHAINS, if name_defined_by_phi is not
1841 NULL, use it to set the stmt field. */
1844 replace_names_by_phis (VEC (chain_p
, heap
) *chains
)
1850 FOR_EACH_VEC_ELT (chain_p
, chains
, i
, chain
)
1851 FOR_EACH_VEC_ELT (dref
, chain
->refs
, j
, a
)
1852 if (a
->stmt
== NULL
)
1854 a
->stmt
= SSA_NAME_DEF_STMT (a
->name_defined_by_phi
);
1855 gcc_assert (gimple_code (a
->stmt
) == GIMPLE_PHI
);
1856 a
->name_defined_by_phi
= NULL_TREE
;
1860 /* Wrapper over execute_pred_commoning, to pass it as a callback
1861 to tree_transform_and_unroll_loop. */
1865 VEC (chain_p
, heap
) *chains
;
1870 execute_pred_commoning_cbck (struct loop
*loop
, void *data
)
1872 struct epcc_data
*const dta
= (struct epcc_data
*) data
;
1874 /* Restore phi nodes that were replaced by ssa names before
1875 tree_transform_and_unroll_loop (see detailed description in
1876 tree_predictive_commoning_loop). */
1877 replace_names_by_phis (dta
->chains
);
1878 execute_pred_commoning (loop
, dta
->chains
, dta
->tmp_vars
);
1881 /* Base NAME and all the names in the chain of phi nodes that use it
1882 on variable VAR. The phi nodes are recognized by being in the copies of
1883 the header of the LOOP. */
1886 base_names_in_chain_on (struct loop
*loop
, tree name
, tree var
)
1889 imm_use_iterator iter
;
1891 SSA_NAME_VAR (name
) = var
;
1896 FOR_EACH_IMM_USE_STMT (stmt
, iter
, name
)
1898 if (gimple_code (stmt
) == GIMPLE_PHI
1899 && flow_bb_inside_loop_p (loop
, gimple_bb (stmt
)))
1902 BREAK_FROM_IMM_USE_STMT (iter
);
1908 name
= PHI_RESULT (phi
);
1909 SSA_NAME_VAR (name
) = var
;
1913 /* Given an unrolled LOOP after predictive commoning, remove the
1914 register copies arising from phi nodes by changing the base
1915 variables of SSA names. TMP_VARS is the set of the temporary variables
1916 for those we want to perform this. */
1919 eliminate_temp_copies (struct loop
*loop
, bitmap tmp_vars
)
1923 tree name
, use
, var
;
1924 gimple_stmt_iterator psi
;
1926 e
= loop_latch_edge (loop
);
1927 for (psi
= gsi_start_phis (loop
->header
); !gsi_end_p (psi
); gsi_next (&psi
))
1929 phi
= gsi_stmt (psi
);
1930 name
= PHI_RESULT (phi
);
1931 var
= SSA_NAME_VAR (name
);
1932 if (!bitmap_bit_p (tmp_vars
, DECL_UID (var
)))
1934 use
= PHI_ARG_DEF_FROM_EDGE (phi
, e
);
1935 gcc_assert (TREE_CODE (use
) == SSA_NAME
);
1937 /* Base all the ssa names in the ud and du chain of NAME on VAR. */
1938 stmt
= SSA_NAME_DEF_STMT (use
);
1939 while (gimple_code (stmt
) == GIMPLE_PHI
1940 /* In case we could not unroll the loop enough to eliminate
1941 all copies, we may reach the loop header before the defining
1942 statement (in that case, some register copies will be present
1943 in loop latch in the final code, corresponding to the newly
1944 created looparound phi nodes). */
1945 && gimple_bb (stmt
) != loop
->header
)
1947 gcc_assert (single_pred_p (gimple_bb (stmt
)));
1948 use
= PHI_ARG_DEF (stmt
, 0);
1949 stmt
= SSA_NAME_DEF_STMT (use
);
1952 base_names_in_chain_on (loop
, use
, var
);
1956 /* Returns true if CHAIN is suitable to be combined. */
1959 chain_can_be_combined_p (chain_p chain
)
1961 return (!chain
->combined
1962 && (chain
->type
== CT_LOAD
|| chain
->type
== CT_COMBINATION
));
1965 /* Returns the modify statement that uses NAME. Skips over assignment
1966 statements, NAME is replaced with the actual name used in the returned
1970 find_use_stmt (tree
*name
)
1975 /* Skip over assignments. */
1978 stmt
= single_nonlooparound_use (*name
);
1982 if (gimple_code (stmt
) != GIMPLE_ASSIGN
)
1985 lhs
= gimple_assign_lhs (stmt
);
1986 if (TREE_CODE (lhs
) != SSA_NAME
)
1989 if (gimple_assign_copy_p (stmt
))
1991 rhs
= gimple_assign_rhs1 (stmt
);
1997 else if (get_gimple_rhs_class (gimple_assign_rhs_code (stmt
))
1998 == GIMPLE_BINARY_RHS
)
2005 /* Returns true if we may perform reassociation for operation CODE in TYPE. */
2008 may_reassociate_p (tree type
, enum tree_code code
)
2010 if (FLOAT_TYPE_P (type
)
2011 && !flag_unsafe_math_optimizations
)
2014 return (commutative_tree_code (code
)
2015 && associative_tree_code (code
));
2018 /* If the operation used in STMT is associative and commutative, go through the
2019 tree of the same operations and returns its root. Distance to the root
2020 is stored in DISTANCE. */
2023 find_associative_operation_root (gimple stmt
, unsigned *distance
)
2027 enum tree_code code
= gimple_assign_rhs_code (stmt
);
2028 tree type
= TREE_TYPE (gimple_assign_lhs (stmt
));
2031 if (!may_reassociate_p (type
, code
))
2036 lhs
= gimple_assign_lhs (stmt
);
2037 gcc_assert (TREE_CODE (lhs
) == SSA_NAME
);
2039 next
= find_use_stmt (&lhs
);
2041 || gimple_assign_rhs_code (next
) != code
)
2053 /* Returns the common statement in that NAME1 and NAME2 have a use. If there
2054 is no such statement, returns NULL_TREE. In case the operation used on
2055 NAME1 and NAME2 is associative and commutative, returns the root of the
2056 tree formed by this operation instead of the statement that uses NAME1 or
2060 find_common_use_stmt (tree
*name1
, tree
*name2
)
2062 gimple stmt1
, stmt2
;
2064 stmt1
= find_use_stmt (name1
);
2068 stmt2
= find_use_stmt (name2
);
2075 stmt1
= find_associative_operation_root (stmt1
, NULL
);
2078 stmt2
= find_associative_operation_root (stmt2
, NULL
);
2082 return (stmt1
== stmt2
? stmt1
: NULL
);
2085 /* Checks whether R1 and R2 are combined together using CODE, with the result
2086 in RSLT_TYPE, in order R1 CODE R2 if SWAP is false and in order R2 CODE R1
2087 if it is true. If CODE is ERROR_MARK, set these values instead. */
2090 combinable_refs_p (dref r1
, dref r2
,
2091 enum tree_code
*code
, bool *swap
, tree
*rslt_type
)
2093 enum tree_code acode
;
2099 name1
= name_for_ref (r1
);
2100 name2
= name_for_ref (r2
);
2101 gcc_assert (name1
!= NULL_TREE
&& name2
!= NULL_TREE
);
2103 stmt
= find_common_use_stmt (&name1
, &name2
);
2108 acode
= gimple_assign_rhs_code (stmt
);
2109 aswap
= (!commutative_tree_code (acode
)
2110 && gimple_assign_rhs1 (stmt
) != name1
);
2111 atype
= TREE_TYPE (gimple_assign_lhs (stmt
));
2113 if (*code
== ERROR_MARK
)
2121 return (*code
== acode
2123 && *rslt_type
== atype
);
2126 /* Remove OP from the operation on rhs of STMT, and replace STMT with
2127 an assignment of the remaining operand. */
2130 remove_name_from_operation (gimple stmt
, tree op
)
2133 gimple_stmt_iterator si
;
2135 gcc_assert (is_gimple_assign (stmt
));
2137 if (gimple_assign_rhs1 (stmt
) == op
)
2138 other_op
= gimple_assign_rhs2 (stmt
);
2140 other_op
= gimple_assign_rhs1 (stmt
);
2142 si
= gsi_for_stmt (stmt
);
2143 gimple_assign_set_rhs_from_tree (&si
, other_op
);
2145 /* We should not have reallocated STMT. */
2146 gcc_assert (gsi_stmt (si
) == stmt
);
2151 /* Reassociates the expression in that NAME1 and NAME2 are used so that they
2152 are combined in a single statement, and returns this statement. */
2155 reassociate_to_the_same_stmt (tree name1
, tree name2
)
2157 gimple stmt1
, stmt2
, root1
, root2
, s1
, s2
;
2158 gimple new_stmt
, tmp_stmt
;
2159 tree new_name
, tmp_name
, var
, r1
, r2
;
2160 unsigned dist1
, dist2
;
2161 enum tree_code code
;
2162 tree type
= TREE_TYPE (name1
);
2163 gimple_stmt_iterator bsi
;
2165 stmt1
= find_use_stmt (&name1
);
2166 stmt2
= find_use_stmt (&name2
);
2167 root1
= find_associative_operation_root (stmt1
, &dist1
);
2168 root2
= find_associative_operation_root (stmt2
, &dist2
);
2169 code
= gimple_assign_rhs_code (stmt1
);
2171 gcc_assert (root1
&& root2
&& root1
== root2
2172 && code
== gimple_assign_rhs_code (stmt2
));
2174 /* Find the root of the nearest expression in that both NAME1 and NAME2
2181 while (dist1
> dist2
)
2183 s1
= find_use_stmt (&r1
);
2184 r1
= gimple_assign_lhs (s1
);
2187 while (dist2
> dist1
)
2189 s2
= find_use_stmt (&r2
);
2190 r2
= gimple_assign_lhs (s2
);
2196 s1
= find_use_stmt (&r1
);
2197 r1
= gimple_assign_lhs (s1
);
2198 s2
= find_use_stmt (&r2
);
2199 r2
= gimple_assign_lhs (s2
);
2202 /* Remove NAME1 and NAME2 from the statements in that they are used
2204 remove_name_from_operation (stmt1
, name1
);
2205 remove_name_from_operation (stmt2
, name2
);
2207 /* Insert the new statement combining NAME1 and NAME2 before S1, and
2208 combine it with the rhs of S1. */
2209 var
= create_tmp_reg (type
, "predreastmp");
2210 add_referenced_var (var
);
2211 new_name
= make_ssa_name (var
, NULL
);
2212 new_stmt
= gimple_build_assign_with_ops (code
, new_name
, name1
, name2
);
2214 var
= create_tmp_reg (type
, "predreastmp");
2215 add_referenced_var (var
);
2216 tmp_name
= make_ssa_name (var
, NULL
);
2218 /* Rhs of S1 may now be either a binary expression with operation
2219 CODE, or gimple_val (in case that stmt1 == s1 or stmt2 == s1,
2220 so that name1 or name2 was removed from it). */
2221 tmp_stmt
= gimple_build_assign_with_ops (gimple_assign_rhs_code (s1
),
2223 gimple_assign_rhs1 (s1
),
2224 gimple_assign_rhs2 (s1
));
2226 bsi
= gsi_for_stmt (s1
);
2227 gimple_assign_set_rhs_with_ops (&bsi
, code
, new_name
, tmp_name
);
2228 s1
= gsi_stmt (bsi
);
2231 gsi_insert_before (&bsi
, new_stmt
, GSI_SAME_STMT
);
2232 gsi_insert_before (&bsi
, tmp_stmt
, GSI_SAME_STMT
);
2237 /* Returns the statement that combines references R1 and R2. In case R1
2238 and R2 are not used in the same statement, but they are used with an
2239 associative and commutative operation in the same expression, reassociate
2240 the expression so that they are used in the same statement. */
2243 stmt_combining_refs (dref r1
, dref r2
)
2245 gimple stmt1
, stmt2
;
2246 tree name1
= name_for_ref (r1
);
2247 tree name2
= name_for_ref (r2
);
2249 stmt1
= find_use_stmt (&name1
);
2250 stmt2
= find_use_stmt (&name2
);
2254 return reassociate_to_the_same_stmt (name1
, name2
);
2257 /* Tries to combine chains CH1 and CH2 together. If this succeeds, the
2258 description of the new chain is returned, otherwise we return NULL. */
2261 combine_chains (chain_p ch1
, chain_p ch2
)
2264 enum tree_code op
= ERROR_MARK
;
2269 tree rslt_type
= NULL_TREE
;
2273 if (ch1
->length
!= ch2
->length
)
2276 if (VEC_length (dref
, ch1
->refs
) != VEC_length (dref
, ch2
->refs
))
2279 for (i
= 0; (VEC_iterate (dref
, ch1
->refs
, i
, r1
)
2280 && VEC_iterate (dref
, ch2
->refs
, i
, r2
)); i
++)
2282 if (r1
->distance
!= r2
->distance
)
2285 if (!combinable_refs_p (r1
, r2
, &op
, &swap
, &rslt_type
))
2296 new_chain
= XCNEW (struct chain
);
2297 new_chain
->type
= CT_COMBINATION
;
2299 new_chain
->ch1
= ch1
;
2300 new_chain
->ch2
= ch2
;
2301 new_chain
->rslt_type
= rslt_type
;
2302 new_chain
->length
= ch1
->length
;
2304 for (i
= 0; (VEC_iterate (dref
, ch1
->refs
, i
, r1
)
2305 && VEC_iterate (dref
, ch2
->refs
, i
, r2
)); i
++)
2307 nw
= XCNEW (struct dref_d
);
2308 nw
->stmt
= stmt_combining_refs (r1
, r2
);
2309 nw
->distance
= r1
->distance
;
2311 VEC_safe_push (dref
, heap
, new_chain
->refs
, nw
);
2314 new_chain
->has_max_use_after
= false;
2315 root_stmt
= get_chain_root (new_chain
)->stmt
;
2316 for (i
= 1; VEC_iterate (dref
, new_chain
->refs
, i
, nw
); i
++)
2318 if (nw
->distance
== new_chain
->length
2319 && !stmt_dominates_stmt_p (nw
->stmt
, root_stmt
))
2321 new_chain
->has_max_use_after
= true;
2326 ch1
->combined
= true;
2327 ch2
->combined
= true;
2331 /* Try to combine the CHAINS. */
2334 try_combine_chains (VEC (chain_p
, heap
) **chains
)
2337 chain_p ch1
, ch2
, cch
;
2338 VEC (chain_p
, heap
) *worklist
= NULL
;
2340 FOR_EACH_VEC_ELT (chain_p
, *chains
, i
, ch1
)
2341 if (chain_can_be_combined_p (ch1
))
2342 VEC_safe_push (chain_p
, heap
, worklist
, ch1
);
2344 while (!VEC_empty (chain_p
, worklist
))
2346 ch1
= VEC_pop (chain_p
, worklist
);
2347 if (!chain_can_be_combined_p (ch1
))
2350 FOR_EACH_VEC_ELT (chain_p
, *chains
, j
, ch2
)
2352 if (!chain_can_be_combined_p (ch2
))
2355 cch
= combine_chains (ch1
, ch2
);
2358 VEC_safe_push (chain_p
, heap
, worklist
, cch
);
2359 VEC_safe_push (chain_p
, heap
, *chains
, cch
);
2366 /* Prepare initializers for CHAIN in LOOP. Returns false if this is
2367 impossible because one of these initializers may trap, true otherwise. */
2370 prepare_initializers_chain (struct loop
*loop
, chain_p chain
)
2372 unsigned i
, n
= (chain
->type
== CT_INVARIANT
) ? 1 : chain
->length
;
2373 struct data_reference
*dr
= get_chain_root (chain
)->ref
;
2377 edge entry
= loop_preheader_edge (loop
);
2379 /* Find the initializers for the variables, and check that they cannot
2381 chain
->inits
= VEC_alloc (tree
, heap
, n
);
2382 for (i
= 0; i
< n
; i
++)
2383 VEC_quick_push (tree
, chain
->inits
, NULL_TREE
);
2385 /* If we have replaced some looparound phi nodes, use their initializers
2386 instead of creating our own. */
2387 FOR_EACH_VEC_ELT (dref
, chain
->refs
, i
, laref
)
2389 if (gimple_code (laref
->stmt
) != GIMPLE_PHI
)
2392 gcc_assert (laref
->distance
> 0);
2393 VEC_replace (tree
, chain
->inits
, n
- laref
->distance
,
2394 PHI_ARG_DEF_FROM_EDGE (laref
->stmt
, entry
));
2397 for (i
= 0; i
< n
; i
++)
2399 if (VEC_index (tree
, chain
->inits
, i
) != NULL_TREE
)
2402 init
= ref_at_iteration (loop
, DR_REF (dr
), (int) i
- n
);
2406 if (!chain
->all_always_accessed
&& tree_could_trap_p (init
))
2409 init
= force_gimple_operand (init
, &stmts
, false, NULL_TREE
);
2411 gsi_insert_seq_on_edge_immediate (entry
, stmts
);
2413 VEC_replace (tree
, chain
->inits
, i
, init
);
2419 /* Prepare initializers for CHAINS in LOOP, and free chains that cannot
2420 be used because the initializers might trap. */
2423 prepare_initializers (struct loop
*loop
, VEC (chain_p
, heap
) *chains
)
2428 for (i
= 0; i
< VEC_length (chain_p
, chains
); )
2430 chain
= VEC_index (chain_p
, chains
, i
);
2431 if (prepare_initializers_chain (loop
, chain
))
2435 release_chain (chain
);
2436 VEC_unordered_remove (chain_p
, chains
, i
);
2441 /* Performs predictive commoning for LOOP. Returns true if LOOP was
2445 tree_predictive_commoning_loop (struct loop
*loop
)
2447 VEC (loop_p
, heap
) *loop_nest
;
2448 VEC (data_reference_p
, heap
) *datarefs
;
2449 VEC (ddr_p
, heap
) *dependences
;
2450 struct component
*components
;
2451 VEC (chain_p
, heap
) *chains
= NULL
;
2452 unsigned unroll_factor
;
2453 struct tree_niter_desc desc
;
2454 bool unroll
= false;
2458 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2459 fprintf (dump_file
, "Processing loop %d\n", loop
->num
);
2461 /* Find the data references and split them into components according to their
2462 dependence relations. */
2463 datarefs
= VEC_alloc (data_reference_p
, heap
, 10);
2464 dependences
= VEC_alloc (ddr_p
, heap
, 10);
2465 loop_nest
= VEC_alloc (loop_p
, heap
, 3);
2466 compute_data_dependences_for_loop (loop
, true, &loop_nest
, &datarefs
,
2468 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2469 dump_data_dependence_relations (dump_file
, dependences
);
2471 components
= split_data_refs_to_components (loop
, datarefs
, dependences
);
2472 VEC_free (loop_p
, heap
, loop_nest
);
2473 free_dependence_relations (dependences
);
2476 free_data_refs (datarefs
);
2480 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2482 fprintf (dump_file
, "Initial state:\n\n");
2483 dump_components (dump_file
, components
);
2486 /* Find the suitable components and split them into chains. */
2487 components
= filter_suitable_components (loop
, components
);
2489 tmp_vars
= BITMAP_ALLOC (NULL
);
2490 looparound_phis
= BITMAP_ALLOC (NULL
);
2491 determine_roots (loop
, components
, &chains
);
2492 release_components (components
);
2496 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2498 "Predictive commoning failed: no suitable chains\n");
2501 prepare_initializers (loop
, chains
);
2503 /* Try to combine the chains that are always worked with together. */
2504 try_combine_chains (&chains
);
2506 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2508 fprintf (dump_file
, "Before commoning:\n\n");
2509 dump_chains (dump_file
, chains
);
2512 /* Determine the unroll factor, and if the loop should be unrolled, ensure
2513 that its number of iterations is divisible by the factor. */
2514 unroll_factor
= determine_unroll_factor (chains
);
2516 unroll
= (unroll_factor
> 1
2517 && can_unroll_loop_p (loop
, unroll_factor
, &desc
));
2518 exit
= single_dom_exit (loop
);
2520 /* Execute the predictive commoning transformations, and possibly unroll the
2524 struct epcc_data dta
;
2526 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2527 fprintf (dump_file
, "Unrolling %u times.\n", unroll_factor
);
2529 dta
.chains
= chains
;
2530 dta
.tmp_vars
= tmp_vars
;
2532 update_ssa (TODO_update_ssa_only_virtuals
);
2534 /* Cfg manipulations performed in tree_transform_and_unroll_loop before
2535 execute_pred_commoning_cbck is called may cause phi nodes to be
2536 reallocated, which is a problem since CHAINS may point to these
2537 statements. To fix this, we store the ssa names defined by the
2538 phi nodes here instead of the phi nodes themselves, and restore
2539 the phi nodes in execute_pred_commoning_cbck. A bit hacky. */
2540 replace_phis_by_defined_names (chains
);
2542 tree_transform_and_unroll_loop (loop
, unroll_factor
, exit
, &desc
,
2543 execute_pred_commoning_cbck
, &dta
);
2544 eliminate_temp_copies (loop
, tmp_vars
);
2548 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2550 "Executing predictive commoning without unrolling.\n");
2551 execute_pred_commoning (loop
, chains
, tmp_vars
);
2555 release_chains (chains
);
2556 free_data_refs (datarefs
);
2557 BITMAP_FREE (tmp_vars
);
2558 BITMAP_FREE (looparound_phis
);
2560 free_affine_expand_cache (&name_expansions
);
2565 /* Runs predictive commoning. */
2568 tree_predictive_commoning (void)
2570 bool unrolled
= false;
2575 initialize_original_copy_tables ();
2576 FOR_EACH_LOOP (li
, loop
, LI_ONLY_INNERMOST
)
2577 if (optimize_loop_for_speed_p (loop
))
2579 unrolled
|= tree_predictive_commoning_loop (loop
);
2585 ret
= TODO_cleanup_cfg
;
2587 free_original_copy_tables ();