1 /* Predictive commoning.
2 Copyright (C) 2005, 2007, 2008, 2009, 2010
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 "diagnostic.h"
202 #include "tree-pretty-print.h"
203 #include "gimple-pretty-print.h"
204 #include "tree-pass.h"
205 #include "tree-affine.h"
206 #include "tree-inline.h"
208 /* The maximum number of iterations between the considered memory
211 #define MAX_DISTANCE (target_avail_regs < 16 ? 4 : 8)
213 /* Data references (or phi nodes that carry data reference values across
216 typedef struct dref_d
218 /* The reference itself. */
219 struct data_reference *ref;
221 /* The statement in that the reference appears. */
224 /* In case that STMT is a phi node, this field is set to the SSA name
225 defined by it in replace_phis_by_defined_names (in order to avoid
226 pointing to phi node that got reallocated in the meantime). */
227 tree name_defined_by_phi;
229 /* Distance of the reference from the root of the chain (in number of
230 iterations of the loop). */
233 /* Number of iterations offset from the first reference in the component. */
236 /* Number of the reference in a component, in dominance ordering. */
239 /* True if the memory reference is always accessed when the loop is
241 unsigned always_accessed : 1;
245 DEF_VEC_ALLOC_P (dref, heap);
247 /* Type of the chain of the references. */
251 /* The addresses of the references in the chain are constant. */
254 /* There are only loads in the chain. */
257 /* Root of the chain is store, the rest are loads. */
260 /* A combination of two chains. */
264 /* Chains of data references. */
268 /* Type of the chain. */
269 enum chain_type type;
271 /* For combination chains, the operator and the two chains that are
272 combined, and the type of the result. */
275 struct chain *ch1, *ch2;
277 /* The references in the chain. */
278 VEC(dref,heap) *refs;
280 /* The maximum distance of the reference in the chain from the root. */
283 /* The variables used to copy the value throughout iterations. */
284 VEC(tree,heap) *vars;
286 /* Initializers for the variables. */
287 VEC(tree,heap) *inits;
289 /* True if there is a use of a variable with the maximal distance
290 that comes after the root in the loop. */
291 unsigned has_max_use_after : 1;
293 /* True if all the memory references in the chain are always accessed. */
294 unsigned all_always_accessed : 1;
296 /* True if this chain was combined together with some other chain. */
297 unsigned combined : 1;
301 DEF_VEC_ALLOC_P (chain_p, heap);
303 /* Describes the knowledge about the step of the memory references in
308 /* The step is zero. */
311 /* The step is nonzero. */
314 /* The step may or may not be nonzero. */
318 /* Components of the data dependence graph. */
322 /* The references in the component. */
323 VEC(dref,heap) *refs;
325 /* What we know about the step of the references in the component. */
326 enum ref_step_type comp_step;
328 /* Next component in the list. */
329 struct component *next;
332 /* Bitmap of ssa names defined by looparound phi nodes covered by chains. */
334 static bitmap looparound_phis;
336 /* Cache used by tree_to_aff_combination_expand. */
338 static struct pointer_map_t *name_expansions;
340 /* Dumps data reference REF to FILE. */
342 extern void dump_dref (FILE *, dref);
344 dump_dref (FILE *file, dref ref)
349 print_generic_expr (file, DR_REF (ref->ref), TDF_SLIM);
350 fprintf (file, " (id %u%s)\n", ref->pos,
351 DR_IS_READ (ref->ref) ? "" : ", write");
353 fprintf (file, " offset ");
354 dump_double_int (file, ref->offset, false);
355 fprintf (file, "\n");
357 fprintf (file, " distance %u\n", ref->distance);
361 if (gimple_code (ref->stmt) == GIMPLE_PHI)
362 fprintf (file, " looparound ref\n");
364 fprintf (file, " combination ref\n");
365 fprintf (file, " in statement ");
366 print_gimple_stmt (file, ref->stmt, 0, TDF_SLIM);
367 fprintf (file, "\n");
368 fprintf (file, " distance %u\n", ref->distance);
373 /* Dumps CHAIN to FILE. */
375 extern void dump_chain (FILE *, chain_p);
377 dump_chain (FILE *file, chain_p chain)
380 const char *chain_type;
387 chain_type = "Load motion";
391 chain_type = "Loads-only";
395 chain_type = "Store-loads";
399 chain_type = "Combination";
406 fprintf (file, "%s chain %p%s\n", chain_type, (void *) chain,
407 chain->combined ? " (combined)" : "");
408 if (chain->type != CT_INVARIANT)
409 fprintf (file, " max distance %u%s\n", chain->length,
410 chain->has_max_use_after ? "" : ", may reuse first");
412 if (chain->type == CT_COMBINATION)
414 fprintf (file, " equal to %p %s %p in type ",
415 (void *) chain->ch1, op_symbol_code (chain->op),
416 (void *) chain->ch2);
417 print_generic_expr (file, chain->rslt_type, TDF_SLIM);
418 fprintf (file, "\n");
423 fprintf (file, " vars");
424 for (i = 0; VEC_iterate (tree, chain->vars, i, var); i++)
427 print_generic_expr (file, var, TDF_SLIM);
429 fprintf (file, "\n");
434 fprintf (file, " inits");
435 for (i = 0; VEC_iterate (tree, chain->inits, i, var); i++)
438 print_generic_expr (file, var, TDF_SLIM);
440 fprintf (file, "\n");
443 fprintf (file, " references:\n");
444 for (i = 0; VEC_iterate (dref, chain->refs, i, a); i++)
447 fprintf (file, "\n");
450 /* Dumps CHAINS to FILE. */
452 extern void dump_chains (FILE *, VEC (chain_p, heap) *);
454 dump_chains (FILE *file, VEC (chain_p, heap) *chains)
459 for (i = 0; VEC_iterate (chain_p, chains, i, chain); i++)
460 dump_chain (file, chain);
463 /* Dumps COMP to FILE. */
465 extern void dump_component (FILE *, struct component *);
467 dump_component (FILE *file, struct component *comp)
472 fprintf (file, "Component%s:\n",
473 comp->comp_step == RS_INVARIANT ? " (invariant)" : "");
474 for (i = 0; VEC_iterate (dref, comp->refs, i, a); i++)
476 fprintf (file, "\n");
479 /* Dumps COMPS to FILE. */
481 extern void dump_components (FILE *, struct component *);
483 dump_components (FILE *file, struct component *comps)
485 struct component *comp;
487 for (comp = comps; comp; comp = comp->next)
488 dump_component (file, comp);
491 /* Frees a chain CHAIN. */
494 release_chain (chain_p chain)
502 for (i = 0; VEC_iterate (dref, chain->refs, i, ref); i++)
505 VEC_free (dref, heap, chain->refs);
506 VEC_free (tree, heap, chain->vars);
507 VEC_free (tree, heap, chain->inits);
515 release_chains (VEC (chain_p, heap) *chains)
520 for (i = 0; VEC_iterate (chain_p, chains, i, chain); i++)
521 release_chain (chain);
522 VEC_free (chain_p, heap, chains);
525 /* Frees a component COMP. */
528 release_component (struct component *comp)
530 VEC_free (dref, heap, comp->refs);
534 /* Frees list of components COMPS. */
537 release_components (struct component *comps)
539 struct component *act, *next;
541 for (act = comps; act; act = next)
544 release_component (act);
548 /* Finds a root of tree given by FATHERS containing A, and performs path
552 component_of (unsigned fathers[], unsigned a)
556 for (root = a; root != fathers[root]; root = fathers[root])
559 for (; a != root; a = n)
568 /* Join operation for DFU. FATHERS gives the tree, SIZES are sizes of the
569 components, A and B are components to merge. */
572 merge_comps (unsigned fathers[], unsigned sizes[], unsigned a, unsigned b)
574 unsigned ca = component_of (fathers, a);
575 unsigned cb = component_of (fathers, b);
580 if (sizes[ca] < sizes[cb])
582 sizes[cb] += sizes[ca];
587 sizes[ca] += sizes[cb];
592 /* Returns true if A is a reference that is suitable for predictive commoning
593 in the innermost loop that contains it. REF_STEP is set according to the
594 step of the reference A. */
597 suitable_reference_p (struct data_reference *a, enum ref_step_type *ref_step)
599 tree ref = DR_REF (a), step = DR_STEP (a);
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 (i = 0; VEC_iterate (edge, exits, i, ex); i++)
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 (i = 0; VEC_iterate (data_reference_p, datarefs, i, dr); i++)
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 (i = 0; VEC_iterate (data_reference_p, datarefs, i, dr); i++)
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 (i = 0; VEC_iterate (ddr_p, depends, i, ddr); i++)
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 (i = 0; VEC_iterate (data_reference_p, datarefs, i, dr); i++)
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 (i = 0; VEC_iterate (dref, comp->refs, i, a); i++)
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_READ (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 (i = 0; VEC_iterate (dref, act->refs, i, ref); i++)
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_add (ref->offset, double_int_neg (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 (i = 0; VEC_iterate (dref, comp->refs, i, ref); i++)
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 (i = 0; VEC_iterate (dref, chain->refs, i, aref); i++)
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 (i = 0; VEC_iterate (dref, chain->refs, i, ref); i++)
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 qsort (VEC_address (dref, comp->refs), VEC_length (dref, comp->refs),
1197 sizeof (dref), order_drefs);
1199 for (i = 0; VEC_iterate (dref, comp->refs, i, a); i++)
1201 if (!chain || !DR_IS_READ (a->ref)
1202 || double_int_ucmp (uhwi_to_double_int (MAX_DISTANCE),
1203 double_int_add (a->offset,
1204 double_int_neg (last_ofs))) <= 0)
1206 if (nontrivial_chain_p (chain))
1208 add_looparound_copies (loop, chain);
1209 VEC_safe_push (chain_p, heap, *chains, chain);
1212 release_chain (chain);
1213 chain = make_rooted_chain (a);
1214 last_ofs = a->offset;
1218 add_ref_to_chain (chain, a);
1221 if (nontrivial_chain_p (chain))
1223 add_looparound_copies (loop, chain);
1224 VEC_safe_push (chain_p, heap, *chains, chain);
1227 release_chain (chain);
1230 /* Find roots of the values and determine distances in components COMPS, and
1231 separates the references to CHAINS. LOOP is the current loop. */
1234 determine_roots (struct loop *loop,
1235 struct component *comps, VEC (chain_p, heap) **chains)
1237 struct component *comp;
1239 for (comp = comps; comp; comp = comp->next)
1240 determine_roots_comp (loop, comp, chains);
1243 /* Replace the reference in statement STMT with temporary variable
1244 NEW_TREE. If SET is true, NEW_TREE is instead initialized to the value of
1245 the reference in the statement. IN_LHS is true if the reference
1246 is in the lhs of STMT, false if it is in rhs. */
1249 replace_ref_with (gimple stmt, tree new_tree, bool set, bool in_lhs)
1253 gimple_stmt_iterator bsi, psi;
1255 if (gimple_code (stmt) == GIMPLE_PHI)
1257 gcc_assert (!in_lhs && !set);
1259 val = PHI_RESULT (stmt);
1260 bsi = gsi_after_labels (gimple_bb (stmt));
1261 psi = gsi_for_stmt (stmt);
1262 remove_phi_node (&psi, false);
1264 /* Turn the phi node into GIMPLE_ASSIGN. */
1265 new_stmt = gimple_build_assign (val, new_tree);
1266 gsi_insert_before (&bsi, new_stmt, GSI_NEW_STMT);
1270 /* Since the reference is of gimple_reg type, it should only
1271 appear as lhs or rhs of modify statement. */
1272 gcc_assert (is_gimple_assign (stmt));
1274 bsi = gsi_for_stmt (stmt);
1276 /* If we do not need to initialize NEW_TREE, just replace the use of OLD. */
1279 gcc_assert (!in_lhs);
1280 gimple_assign_set_rhs_from_tree (&bsi, new_tree);
1281 stmt = gsi_stmt (bsi);
1288 /* We have statement
1292 If OLD is a memory reference, then VAL is gimple_val, and we transform
1298 Otherwise, we are replacing a combination chain,
1299 VAL is the expression that performs the combination, and OLD is an
1300 SSA name. In this case, we transform the assignment to
1307 val = gimple_assign_lhs (stmt);
1308 if (TREE_CODE (val) != SSA_NAME)
1310 gcc_assert (gimple_assign_copy_p (stmt));
1311 val = gimple_assign_rhs1 (stmt);
1323 val = gimple_assign_lhs (stmt);
1326 new_stmt = gimple_build_assign (new_tree, unshare_expr (val));
1327 gsi_insert_after (&bsi, new_stmt, GSI_NEW_STMT);
1330 /* Returns the reference to the address of REF in the ITER-th iteration of
1331 LOOP, or NULL if we fail to determine it (ITER may be negative). We
1332 try to preserve the original shape of the reference (not rewrite it
1333 as an indirect ref to the address), to make tree_could_trap_p in
1334 prepare_initializers_chain return false more often. */
1337 ref_at_iteration (struct loop *loop, tree ref, int iter)
1339 tree idx, *idx_p, type, val, op0 = NULL_TREE, ret;
1343 if (handled_component_p (ref))
1345 op0 = ref_at_iteration (loop, TREE_OPERAND (ref, 0), iter);
1349 else if (!INDIRECT_REF_P (ref))
1350 return unshare_expr (ref);
1352 if (INDIRECT_REF_P (ref))
1354 /* Take care for INDIRECT_REF and MISALIGNED_INDIRECT_REF at
1356 ret = copy_node (ref);
1357 idx = TREE_OPERAND (ref, 0);
1358 idx_p = &TREE_OPERAND (ret, 0);
1360 else if (TREE_CODE (ref) == COMPONENT_REF)
1362 /* Check that the offset is loop invariant. */
1363 if (TREE_OPERAND (ref, 2)
1364 && !expr_invariant_in_loop_p (loop, TREE_OPERAND (ref, 2)))
1367 return build3 (COMPONENT_REF, TREE_TYPE (ref), op0,
1368 unshare_expr (TREE_OPERAND (ref, 1)),
1369 unshare_expr (TREE_OPERAND (ref, 2)));
1371 else if (TREE_CODE (ref) == ARRAY_REF)
1373 /* Check that the lower bound and the step are loop invariant. */
1374 if (TREE_OPERAND (ref, 2)
1375 && !expr_invariant_in_loop_p (loop, TREE_OPERAND (ref, 2)))
1377 if (TREE_OPERAND (ref, 3)
1378 && !expr_invariant_in_loop_p (loop, TREE_OPERAND (ref, 3)))
1381 ret = build4 (ARRAY_REF, TREE_TYPE (ref), op0, NULL_TREE,
1382 unshare_expr (TREE_OPERAND (ref, 2)),
1383 unshare_expr (TREE_OPERAND (ref, 3)));
1384 idx = TREE_OPERAND (ref, 1);
1385 idx_p = &TREE_OPERAND (ret, 1);
1390 ok = simple_iv (loop, loop, idx, &iv, true);
1393 iv.base = expand_simple_operations (iv.base);
1394 if (integer_zerop (iv.step))
1395 *idx_p = unshare_expr (iv.base);
1398 type = TREE_TYPE (iv.base);
1399 if (POINTER_TYPE_P (type))
1401 val = fold_build2 (MULT_EXPR, sizetype, iv.step,
1403 val = fold_build2 (POINTER_PLUS_EXPR, type, iv.base, val);
1407 val = fold_build2 (MULT_EXPR, type, iv.step,
1408 build_int_cst_type (type, iter));
1409 val = fold_build2 (PLUS_EXPR, type, iv.base, val);
1411 *idx_p = unshare_expr (val);
1417 /* Get the initialization expression for the INDEX-th temporary variable
1421 get_init_expr (chain_p chain, unsigned index)
1423 if (chain->type == CT_COMBINATION)
1425 tree e1 = get_init_expr (chain->ch1, index);
1426 tree e2 = get_init_expr (chain->ch2, index);
1428 return fold_build2 (chain->op, chain->rslt_type, e1, e2);
1431 return VEC_index (tree, chain->inits, index);
1434 /* Marks all virtual operands of statement STMT for renaming. */
1437 mark_virtual_ops_for_renaming (gimple stmt)
1441 if (gimple_code (stmt) == GIMPLE_PHI)
1443 var = PHI_RESULT (stmt);
1444 if (is_gimple_reg (var))
1447 if (TREE_CODE (var) == SSA_NAME)
1448 var = SSA_NAME_VAR (var);
1449 mark_sym_for_renaming (var);
1454 if (gimple_vuse (stmt))
1455 mark_sym_for_renaming (gimple_vop (cfun));
1458 /* Returns a new temporary variable used for the I-th variable carrying
1459 value of REF. The variable's uid is marked in TMP_VARS. */
1462 predcom_tmp_var (tree ref, unsigned i, bitmap tmp_vars)
1464 tree type = TREE_TYPE (ref);
1465 /* We never access the components of the temporary variable in predictive
1467 tree var = create_tmp_reg (type, get_lsm_tmp_name (ref, i));
1469 add_referenced_var (var);
1470 bitmap_set_bit (tmp_vars, DECL_UID (var));
1474 /* Creates the variables for CHAIN, as well as phi nodes for them and
1475 initialization on entry to LOOP. Uids of the newly created
1476 temporary variables are marked in TMP_VARS. */
1479 initialize_root_vars (struct loop *loop, chain_p chain, bitmap tmp_vars)
1482 unsigned n = chain->length;
1483 dref root = get_chain_root (chain);
1484 bool reuse_first = !chain->has_max_use_after;
1485 tree ref, init, var, next;
1488 edge entry = loop_preheader_edge (loop), latch = loop_latch_edge (loop);
1490 /* If N == 0, then all the references are within the single iteration. And
1491 since this is an nonempty chain, reuse_first cannot be true. */
1492 gcc_assert (n > 0 || !reuse_first);
1494 chain->vars = VEC_alloc (tree, heap, n + 1);
1496 if (chain->type == CT_COMBINATION)
1497 ref = gimple_assign_lhs (root->stmt);
1499 ref = DR_REF (root->ref);
1501 for (i = 0; i < n + (reuse_first ? 0 : 1); i++)
1503 var = predcom_tmp_var (ref, i, tmp_vars);
1504 VEC_quick_push (tree, chain->vars, var);
1507 VEC_quick_push (tree, chain->vars, VEC_index (tree, chain->vars, 0));
1509 for (i = 0; VEC_iterate (tree, chain->vars, i, var); i++)
1510 VEC_replace (tree, chain->vars, i, make_ssa_name (var, NULL));
1512 for (i = 0; i < n; i++)
1514 var = VEC_index (tree, chain->vars, i);
1515 next = VEC_index (tree, chain->vars, i + 1);
1516 init = get_init_expr (chain, i);
1518 init = force_gimple_operand (init, &stmts, true, NULL_TREE);
1520 gsi_insert_seq_on_edge_immediate (entry, stmts);
1522 phi = create_phi_node (var, loop->header);
1523 SSA_NAME_DEF_STMT (var) = phi;
1524 add_phi_arg (phi, init, entry, UNKNOWN_LOCATION);
1525 add_phi_arg (phi, next, latch, UNKNOWN_LOCATION);
1529 /* Create the variables and initialization statement for root of chain
1530 CHAIN. Uids of the newly created temporary variables are marked
1534 initialize_root (struct loop *loop, chain_p chain, bitmap tmp_vars)
1536 dref root = get_chain_root (chain);
1537 bool in_lhs = (chain->type == CT_STORE_LOAD
1538 || chain->type == CT_COMBINATION);
1540 initialize_root_vars (loop, chain, tmp_vars);
1541 replace_ref_with (root->stmt,
1542 VEC_index (tree, chain->vars, chain->length),
1546 /* Initializes a variable for load motion for ROOT and prepares phi nodes and
1547 initialization on entry to LOOP if necessary. The ssa name for the variable
1548 is stored in VARS. If WRITTEN is true, also a phi node to copy its value
1549 around the loop is created. Uid of the newly created temporary variable
1550 is marked in TMP_VARS. INITS is the list containing the (single)
1554 initialize_root_vars_lm (struct loop *loop, dref root, bool written,
1555 VEC(tree, heap) **vars, VEC(tree, heap) *inits,
1559 tree ref = DR_REF (root->ref), init, var, next;
1562 edge entry = loop_preheader_edge (loop), latch = loop_latch_edge (loop);
1564 /* Find the initializer for the variable, and check that it cannot
1566 init = VEC_index (tree, inits, 0);
1568 *vars = VEC_alloc (tree, heap, written ? 2 : 1);
1569 var = predcom_tmp_var (ref, 0, tmp_vars);
1570 VEC_quick_push (tree, *vars, var);
1572 VEC_quick_push (tree, *vars, VEC_index (tree, *vars, 0));
1574 for (i = 0; VEC_iterate (tree, *vars, i, var); i++)
1575 VEC_replace (tree, *vars, i, make_ssa_name (var, NULL));
1577 var = VEC_index (tree, *vars, 0);
1579 init = force_gimple_operand (init, &stmts, written, NULL_TREE);
1581 gsi_insert_seq_on_edge_immediate (entry, stmts);
1585 next = VEC_index (tree, *vars, 1);
1586 phi = create_phi_node (var, loop->header);
1587 SSA_NAME_DEF_STMT (var) = phi;
1588 add_phi_arg (phi, init, entry, UNKNOWN_LOCATION);
1589 add_phi_arg (phi, next, latch, UNKNOWN_LOCATION);
1593 gimple init_stmt = gimple_build_assign (var, init);
1594 mark_virtual_ops_for_renaming (init_stmt);
1595 gsi_insert_on_edge_immediate (entry, init_stmt);
1600 /* Execute load motion for references in chain CHAIN. Uids of the newly
1601 created temporary variables are marked in TMP_VARS. */
1604 execute_load_motion (struct loop *loop, chain_p chain, bitmap tmp_vars)
1606 VEC (tree, heap) *vars;
1608 unsigned n_writes = 0, ridx, i;
1611 gcc_assert (chain->type == CT_INVARIANT);
1612 gcc_assert (!chain->combined);
1613 for (i = 0; VEC_iterate (dref, chain->refs, i, a); i++)
1614 if (!DR_IS_READ (a->ref))
1617 /* If there are no reads in the loop, there is nothing to do. */
1618 if (n_writes == VEC_length (dref, chain->refs))
1621 initialize_root_vars_lm (loop, get_chain_root (chain), n_writes > 0,
1622 &vars, chain->inits, tmp_vars);
1625 for (i = 0; VEC_iterate (dref, chain->refs, i, a); i++)
1627 bool is_read = DR_IS_READ (a->ref);
1628 mark_virtual_ops_for_renaming (a->stmt);
1630 if (!DR_IS_READ (a->ref))
1635 var = VEC_index (tree, vars, 0);
1636 var = make_ssa_name (SSA_NAME_VAR (var), NULL);
1637 VEC_replace (tree, vars, 0, var);
1643 replace_ref_with (a->stmt, VEC_index (tree, vars, ridx),
1644 !is_read, !is_read);
1647 VEC_free (tree, heap, vars);
1650 /* Returns the single statement in that NAME is used, excepting
1651 the looparound phi nodes contained in one of the chains. If there is no
1652 such statement, or more statements, NULL is returned. */
1655 single_nonlooparound_use (tree name)
1658 imm_use_iterator it;
1659 gimple stmt, ret = NULL;
1661 FOR_EACH_IMM_USE_FAST (use, it, name)
1663 stmt = USE_STMT (use);
1665 if (gimple_code (stmt) == GIMPLE_PHI)
1667 /* Ignore uses in looparound phi nodes. Uses in other phi nodes
1668 could not be processed anyway, so just fail for them. */
1669 if (bitmap_bit_p (looparound_phis,
1670 SSA_NAME_VERSION (PHI_RESULT (stmt))))
1675 else if (ret != NULL)
1684 /* Remove statement STMT, as well as the chain of assignments in that it is
1688 remove_stmt (gimple stmt)
1692 gimple_stmt_iterator psi;
1694 if (gimple_code (stmt) == GIMPLE_PHI)
1696 name = PHI_RESULT (stmt);
1697 next = single_nonlooparound_use (name);
1698 psi = gsi_for_stmt (stmt);
1699 remove_phi_node (&psi, true);
1702 || !gimple_assign_ssa_name_copy_p (next)
1703 || gimple_assign_rhs1 (next) != name)
1711 gimple_stmt_iterator bsi;
1713 bsi = gsi_for_stmt (stmt);
1715 name = gimple_assign_lhs (stmt);
1716 gcc_assert (TREE_CODE (name) == SSA_NAME);
1718 next = single_nonlooparound_use (name);
1720 mark_virtual_ops_for_renaming (stmt);
1721 gsi_remove (&bsi, true);
1722 release_defs (stmt);
1725 || !gimple_assign_ssa_name_copy_p (next)
1726 || gimple_assign_rhs1 (next) != name)
1733 /* Perform the predictive commoning optimization for a chain CHAIN.
1734 Uids of the newly created temporary variables are marked in TMP_VARS.*/
1737 execute_pred_commoning_chain (struct loop *loop, chain_p chain,
1744 if (chain->combined)
1746 /* For combined chains, just remove the statements that are used to
1747 compute the values of the expression (except for the root one). */
1748 for (i = 1; VEC_iterate (dref, chain->refs, i, a); i++)
1749 remove_stmt (a->stmt);
1753 /* For non-combined chains, set up the variables that hold its value,
1754 and replace the uses of the original references by these
1756 root = get_chain_root (chain);
1757 mark_virtual_ops_for_renaming (root->stmt);
1759 initialize_root (loop, chain, tmp_vars);
1760 for (i = 1; VEC_iterate (dref, chain->refs, i, a); i++)
1762 mark_virtual_ops_for_renaming (a->stmt);
1763 var = VEC_index (tree, chain->vars, chain->length - a->distance);
1764 replace_ref_with (a->stmt, var, false, false);
1769 /* Determines the unroll factor necessary to remove as many temporary variable
1770 copies as possible. CHAINS is the list of chains that will be
1774 determine_unroll_factor (VEC (chain_p, heap) *chains)
1777 unsigned factor = 1, af, nfactor, i;
1778 unsigned max = PARAM_VALUE (PARAM_MAX_UNROLL_TIMES);
1780 for (i = 0; VEC_iterate (chain_p, chains, i, chain); i++)
1782 if (chain->type == CT_INVARIANT || chain->combined)
1785 /* The best unroll factor for this chain is equal to the number of
1786 temporary variables that we create for it. */
1788 if (chain->has_max_use_after)
1791 nfactor = factor * af / gcd (factor, af);
1799 /* Perform the predictive commoning optimization for CHAINS.
1800 Uids of the newly created temporary variables are marked in TMP_VARS. */
1803 execute_pred_commoning (struct loop *loop, VEC (chain_p, heap) *chains,
1809 for (i = 0; VEC_iterate (chain_p, chains, i, chain); i++)
1811 if (chain->type == CT_INVARIANT)
1812 execute_load_motion (loop, chain, tmp_vars);
1814 execute_pred_commoning_chain (loop, chain, tmp_vars);
1817 update_ssa (TODO_update_ssa_only_virtuals);
1820 /* For each reference in CHAINS, if its defining statement is
1821 phi node, record the ssa name that is defined by it. */
1824 replace_phis_by_defined_names (VEC (chain_p, heap) *chains)
1830 for (i = 0; VEC_iterate (chain_p, chains, i, chain); i++)
1831 for (j = 0; VEC_iterate (dref, chain->refs, j, a); j++)
1833 if (gimple_code (a->stmt) == GIMPLE_PHI)
1835 a->name_defined_by_phi = PHI_RESULT (a->stmt);
1841 /* For each reference in CHAINS, if name_defined_by_phi is not
1842 NULL, use it to set the stmt field. */
1845 replace_names_by_phis (VEC (chain_p, heap) *chains)
1851 for (i = 0; VEC_iterate (chain_p, chains, i, chain); i++)
1852 for (j = 0; VEC_iterate (dref, chain->refs, j, a); j++)
1853 if (a->stmt == NULL)
1855 a->stmt = SSA_NAME_DEF_STMT (a->name_defined_by_phi);
1856 gcc_assert (gimple_code (a->stmt) == GIMPLE_PHI);
1857 a->name_defined_by_phi = NULL_TREE;
1861 /* Wrapper over execute_pred_commoning, to pass it as a callback
1862 to tree_transform_and_unroll_loop. */
1866 VEC (chain_p, heap) *chains;
1871 execute_pred_commoning_cbck (struct loop *loop, void *data)
1873 struct epcc_data *const dta = (struct epcc_data *) data;
1875 /* Restore phi nodes that were replaced by ssa names before
1876 tree_transform_and_unroll_loop (see detailed description in
1877 tree_predictive_commoning_loop). */
1878 replace_names_by_phis (dta->chains);
1879 execute_pred_commoning (loop, dta->chains, dta->tmp_vars);
1882 /* Base NAME and all the names in the chain of phi nodes that use it
1883 on variable VAR. The phi nodes are recognized by being in the copies of
1884 the header of the LOOP. */
1887 base_names_in_chain_on (struct loop *loop, tree name, tree var)
1890 imm_use_iterator iter;
1892 SSA_NAME_VAR (name) = var;
1897 FOR_EACH_IMM_USE_STMT (stmt, iter, name)
1899 if (gimple_code (stmt) == GIMPLE_PHI
1900 && flow_bb_inside_loop_p (loop, gimple_bb (stmt)))
1903 BREAK_FROM_IMM_USE_STMT (iter);
1909 name = PHI_RESULT (phi);
1910 SSA_NAME_VAR (name) = var;
1914 /* Given an unrolled LOOP after predictive commoning, remove the
1915 register copies arising from phi nodes by changing the base
1916 variables of SSA names. TMP_VARS is the set of the temporary variables
1917 for those we want to perform this. */
1920 eliminate_temp_copies (struct loop *loop, bitmap tmp_vars)
1924 tree name, use, var;
1925 gimple_stmt_iterator psi;
1927 e = loop_latch_edge (loop);
1928 for (psi = gsi_start_phis (loop->header); !gsi_end_p (psi); gsi_next (&psi))
1930 phi = gsi_stmt (psi);
1931 name = PHI_RESULT (phi);
1932 var = SSA_NAME_VAR (name);
1933 if (!bitmap_bit_p (tmp_vars, DECL_UID (var)))
1935 use = PHI_ARG_DEF_FROM_EDGE (phi, e);
1936 gcc_assert (TREE_CODE (use) == SSA_NAME);
1938 /* Base all the ssa names in the ud and du chain of NAME on VAR. */
1939 stmt = SSA_NAME_DEF_STMT (use);
1940 while (gimple_code (stmt) == GIMPLE_PHI
1941 /* In case we could not unroll the loop enough to eliminate
1942 all copies, we may reach the loop header before the defining
1943 statement (in that case, some register copies will be present
1944 in loop latch in the final code, corresponding to the newly
1945 created looparound phi nodes). */
1946 && gimple_bb (stmt) != loop->header)
1948 gcc_assert (single_pred_p (gimple_bb (stmt)));
1949 use = PHI_ARG_DEF (stmt, 0);
1950 stmt = SSA_NAME_DEF_STMT (use);
1953 base_names_in_chain_on (loop, use, var);
1957 /* Returns true if CHAIN is suitable to be combined. */
1960 chain_can_be_combined_p (chain_p chain)
1962 return (!chain->combined
1963 && (chain->type == CT_LOAD || chain->type == CT_COMBINATION));
1966 /* Returns the modify statement that uses NAME. Skips over assignment
1967 statements, NAME is replaced with the actual name used in the returned
1971 find_use_stmt (tree *name)
1976 /* Skip over assignments. */
1979 stmt = single_nonlooparound_use (*name);
1983 if (gimple_code (stmt) != GIMPLE_ASSIGN)
1986 lhs = gimple_assign_lhs (stmt);
1987 if (TREE_CODE (lhs) != SSA_NAME)
1990 if (gimple_assign_copy_p (stmt))
1992 rhs = gimple_assign_rhs1 (stmt);
1998 else if (get_gimple_rhs_class (gimple_assign_rhs_code (stmt))
1999 == GIMPLE_BINARY_RHS)
2006 /* Returns true if we may perform reassociation for operation CODE in TYPE. */
2009 may_reassociate_p (tree type, enum tree_code code)
2011 if (FLOAT_TYPE_P (type)
2012 && !flag_unsafe_math_optimizations)
2015 return (commutative_tree_code (code)
2016 && associative_tree_code (code));
2019 /* If the operation used in STMT is associative and commutative, go through the
2020 tree of the same operations and returns its root. Distance to the root
2021 is stored in DISTANCE. */
2024 find_associative_operation_root (gimple stmt, unsigned *distance)
2028 enum tree_code code = gimple_assign_rhs_code (stmt);
2029 tree type = TREE_TYPE (gimple_assign_lhs (stmt));
2032 if (!may_reassociate_p (type, code))
2037 lhs = gimple_assign_lhs (stmt);
2038 gcc_assert (TREE_CODE (lhs) == SSA_NAME);
2040 next = find_use_stmt (&lhs);
2042 || gimple_assign_rhs_code (next) != code)
2054 /* Returns the common statement in that NAME1 and NAME2 have a use. If there
2055 is no such statement, returns NULL_TREE. In case the operation used on
2056 NAME1 and NAME2 is associative and commutative, returns the root of the
2057 tree formed by this operation instead of the statement that uses NAME1 or
2061 find_common_use_stmt (tree *name1, tree *name2)
2063 gimple stmt1, stmt2;
2065 stmt1 = find_use_stmt (name1);
2069 stmt2 = find_use_stmt (name2);
2076 stmt1 = find_associative_operation_root (stmt1, NULL);
2079 stmt2 = find_associative_operation_root (stmt2, NULL);
2083 return (stmt1 == stmt2 ? stmt1 : NULL);
2086 /* Checks whether R1 and R2 are combined together using CODE, with the result
2087 in RSLT_TYPE, in order R1 CODE R2 if SWAP is false and in order R2 CODE R1
2088 if it is true. If CODE is ERROR_MARK, set these values instead. */
2091 combinable_refs_p (dref r1, dref r2,
2092 enum tree_code *code, bool *swap, tree *rslt_type)
2094 enum tree_code acode;
2100 name1 = name_for_ref (r1);
2101 name2 = name_for_ref (r2);
2102 gcc_assert (name1 != NULL_TREE && name2 != NULL_TREE);
2104 stmt = find_common_use_stmt (&name1, &name2);
2109 acode = gimple_assign_rhs_code (stmt);
2110 aswap = (!commutative_tree_code (acode)
2111 && gimple_assign_rhs1 (stmt) != name1);
2112 atype = TREE_TYPE (gimple_assign_lhs (stmt));
2114 if (*code == ERROR_MARK)
2122 return (*code == acode
2124 && *rslt_type == atype);
2127 /* Remove OP from the operation on rhs of STMT, and replace STMT with
2128 an assignment of the remaining operand. */
2131 remove_name_from_operation (gimple stmt, tree op)
2134 gimple_stmt_iterator si;
2136 gcc_assert (is_gimple_assign (stmt));
2138 if (gimple_assign_rhs1 (stmt) == op)
2139 other_op = gimple_assign_rhs2 (stmt);
2141 other_op = gimple_assign_rhs1 (stmt);
2143 si = gsi_for_stmt (stmt);
2144 gimple_assign_set_rhs_from_tree (&si, other_op);
2146 /* We should not have reallocated STMT. */
2147 gcc_assert (gsi_stmt (si) == stmt);
2152 /* Reassociates the expression in that NAME1 and NAME2 are used so that they
2153 are combined in a single statement, and returns this statement. */
2156 reassociate_to_the_same_stmt (tree name1, tree name2)
2158 gimple stmt1, stmt2, root1, root2, s1, s2;
2159 gimple new_stmt, tmp_stmt;
2160 tree new_name, tmp_name, var, r1, r2;
2161 unsigned dist1, dist2;
2162 enum tree_code code;
2163 tree type = TREE_TYPE (name1);
2164 gimple_stmt_iterator bsi;
2166 stmt1 = find_use_stmt (&name1);
2167 stmt2 = find_use_stmt (&name2);
2168 root1 = find_associative_operation_root (stmt1, &dist1);
2169 root2 = find_associative_operation_root (stmt2, &dist2);
2170 code = gimple_assign_rhs_code (stmt1);
2172 gcc_assert (root1 && root2 && root1 == root2
2173 && code == gimple_assign_rhs_code (stmt2));
2175 /* Find the root of the nearest expression in that both NAME1 and NAME2
2182 while (dist1 > dist2)
2184 s1 = find_use_stmt (&r1);
2185 r1 = gimple_assign_lhs (s1);
2188 while (dist2 > dist1)
2190 s2 = find_use_stmt (&r2);
2191 r2 = gimple_assign_lhs (s2);
2197 s1 = find_use_stmt (&r1);
2198 r1 = gimple_assign_lhs (s1);
2199 s2 = find_use_stmt (&r2);
2200 r2 = gimple_assign_lhs (s2);
2203 /* Remove NAME1 and NAME2 from the statements in that they are used
2205 remove_name_from_operation (stmt1, name1);
2206 remove_name_from_operation (stmt2, name2);
2208 /* Insert the new statement combining NAME1 and NAME2 before S1, and
2209 combine it with the rhs of S1. */
2210 var = create_tmp_reg (type, "predreastmp");
2211 add_referenced_var (var);
2212 new_name = make_ssa_name (var, NULL);
2213 new_stmt = gimple_build_assign_with_ops (code, new_name, name1, name2);
2215 var = create_tmp_reg (type, "predreastmp");
2216 add_referenced_var (var);
2217 tmp_name = make_ssa_name (var, NULL);
2219 /* Rhs of S1 may now be either a binary expression with operation
2220 CODE, or gimple_val (in case that stmt1 == s1 or stmt2 == s1,
2221 so that name1 or name2 was removed from it). */
2222 tmp_stmt = gimple_build_assign_with_ops (gimple_assign_rhs_code (s1),
2224 gimple_assign_rhs1 (s1),
2225 gimple_assign_rhs2 (s1));
2227 bsi = gsi_for_stmt (s1);
2228 gimple_assign_set_rhs_with_ops (&bsi, code, new_name, tmp_name);
2229 s1 = gsi_stmt (bsi);
2232 gsi_insert_before (&bsi, new_stmt, GSI_SAME_STMT);
2233 gsi_insert_before (&bsi, tmp_stmt, GSI_SAME_STMT);
2238 /* Returns the statement that combines references R1 and R2. In case R1
2239 and R2 are not used in the same statement, but they are used with an
2240 associative and commutative operation in the same expression, reassociate
2241 the expression so that they are used in the same statement. */
2244 stmt_combining_refs (dref r1, dref r2)
2246 gimple stmt1, stmt2;
2247 tree name1 = name_for_ref (r1);
2248 tree name2 = name_for_ref (r2);
2250 stmt1 = find_use_stmt (&name1);
2251 stmt2 = find_use_stmt (&name2);
2255 return reassociate_to_the_same_stmt (name1, name2);
2258 /* Tries to combine chains CH1 and CH2 together. If this succeeds, the
2259 description of the new chain is returned, otherwise we return NULL. */
2262 combine_chains (chain_p ch1, chain_p ch2)
2265 enum tree_code op = ERROR_MARK;
2270 tree rslt_type = NULL_TREE;
2274 if (ch1->length != ch2->length)
2277 if (VEC_length (dref, ch1->refs) != VEC_length (dref, ch2->refs))
2280 for (i = 0; (VEC_iterate (dref, ch1->refs, i, r1)
2281 && VEC_iterate (dref, ch2->refs, i, r2)); i++)
2283 if (r1->distance != r2->distance)
2286 if (!combinable_refs_p (r1, r2, &op, &swap, &rslt_type))
2297 new_chain = XCNEW (struct chain);
2298 new_chain->type = CT_COMBINATION;
2300 new_chain->ch1 = ch1;
2301 new_chain->ch2 = ch2;
2302 new_chain->rslt_type = rslt_type;
2303 new_chain->length = ch1->length;
2305 for (i = 0; (VEC_iterate (dref, ch1->refs, i, r1)
2306 && VEC_iterate (dref, ch2->refs, i, r2)); i++)
2308 nw = XCNEW (struct dref_d);
2309 nw->stmt = stmt_combining_refs (r1, r2);
2310 nw->distance = r1->distance;
2312 VEC_safe_push (dref, heap, new_chain->refs, nw);
2315 new_chain->has_max_use_after = false;
2316 root_stmt = get_chain_root (new_chain)->stmt;
2317 for (i = 1; VEC_iterate (dref, new_chain->refs, i, nw); i++)
2319 if (nw->distance == new_chain->length
2320 && !stmt_dominates_stmt_p (nw->stmt, root_stmt))
2322 new_chain->has_max_use_after = true;
2327 ch1->combined = true;
2328 ch2->combined = true;
2332 /* Try to combine the CHAINS. */
2335 try_combine_chains (VEC (chain_p, heap) **chains)
2338 chain_p ch1, ch2, cch;
2339 VEC (chain_p, heap) *worklist = NULL;
2341 for (i = 0; VEC_iterate (chain_p, *chains, i, ch1); i++)
2342 if (chain_can_be_combined_p (ch1))
2343 VEC_safe_push (chain_p, heap, worklist, ch1);
2345 while (!VEC_empty (chain_p, worklist))
2347 ch1 = VEC_pop (chain_p, worklist);
2348 if (!chain_can_be_combined_p (ch1))
2351 for (j = 0; VEC_iterate (chain_p, *chains, j, ch2); j++)
2353 if (!chain_can_be_combined_p (ch2))
2356 cch = combine_chains (ch1, ch2);
2359 VEC_safe_push (chain_p, heap, worklist, cch);
2360 VEC_safe_push (chain_p, heap, *chains, cch);
2367 /* Prepare initializers for CHAIN in LOOP. Returns false if this is
2368 impossible because one of these initializers may trap, true otherwise. */
2371 prepare_initializers_chain (struct loop *loop, chain_p chain)
2373 unsigned i, n = (chain->type == CT_INVARIANT) ? 1 : chain->length;
2374 struct data_reference *dr = get_chain_root (chain)->ref;
2378 edge entry = loop_preheader_edge (loop);
2380 /* Find the initializers for the variables, and check that they cannot
2382 chain->inits = VEC_alloc (tree, heap, n);
2383 for (i = 0; i < n; i++)
2384 VEC_quick_push (tree, chain->inits, NULL_TREE);
2386 /* If we have replaced some looparound phi nodes, use their initializers
2387 instead of creating our own. */
2388 for (i = 0; VEC_iterate (dref, chain->refs, i, laref); i++)
2390 if (gimple_code (laref->stmt) != GIMPLE_PHI)
2393 gcc_assert (laref->distance > 0);
2394 VEC_replace (tree, chain->inits, n - laref->distance,
2395 PHI_ARG_DEF_FROM_EDGE (laref->stmt, entry));
2398 for (i = 0; i < n; i++)
2400 if (VEC_index (tree, chain->inits, i) != NULL_TREE)
2403 init = ref_at_iteration (loop, DR_REF (dr), (int) i - n);
2407 if (!chain->all_always_accessed && tree_could_trap_p (init))
2410 init = force_gimple_operand (init, &stmts, false, NULL_TREE);
2412 gsi_insert_seq_on_edge_immediate (entry, stmts);
2414 VEC_replace (tree, chain->inits, i, init);
2420 /* Prepare initializers for CHAINS in LOOP, and free chains that cannot
2421 be used because the initializers might trap. */
2424 prepare_initializers (struct loop *loop, VEC (chain_p, heap) *chains)
2429 for (i = 0; i < VEC_length (chain_p, chains); )
2431 chain = VEC_index (chain_p, chains, i);
2432 if (prepare_initializers_chain (loop, chain))
2436 release_chain (chain);
2437 VEC_unordered_remove (chain_p, chains, i);
2442 /* Performs predictive commoning for LOOP. Returns true if LOOP was
2446 tree_predictive_commoning_loop (struct loop *loop)
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 compute_data_dependences_for_loop (loop, true, &datarefs, &dependences);
2466 if (dump_file && (dump_flags & TDF_DETAILS))
2467 dump_data_dependence_relations (dump_file, dependences);
2469 components = split_data_refs_to_components (loop, datarefs, dependences);
2470 free_dependence_relations (dependences);
2473 free_data_refs (datarefs);
2477 if (dump_file && (dump_flags & TDF_DETAILS))
2479 fprintf (dump_file, "Initial state:\n\n");
2480 dump_components (dump_file, components);
2483 /* Find the suitable components and split them into chains. */
2484 components = filter_suitable_components (loop, components);
2486 tmp_vars = BITMAP_ALLOC (NULL);
2487 looparound_phis = BITMAP_ALLOC (NULL);
2488 determine_roots (loop, components, &chains);
2489 release_components (components);
2493 if (dump_file && (dump_flags & TDF_DETAILS))
2495 "Predictive commoning failed: no suitable chains\n");
2498 prepare_initializers (loop, chains);
2500 /* Try to combine the chains that are always worked with together. */
2501 try_combine_chains (&chains);
2503 if (dump_file && (dump_flags & TDF_DETAILS))
2505 fprintf (dump_file, "Before commoning:\n\n");
2506 dump_chains (dump_file, chains);
2509 /* Determine the unroll factor, and if the loop should be unrolled, ensure
2510 that its number of iterations is divisible by the factor. */
2511 unroll_factor = determine_unroll_factor (chains);
2513 unroll = (unroll_factor > 1
2514 && can_unroll_loop_p (loop, unroll_factor, &desc));
2515 exit = single_dom_exit (loop);
2517 /* Execute the predictive commoning transformations, and possibly unroll the
2521 struct epcc_data dta;
2523 if (dump_file && (dump_flags & TDF_DETAILS))
2524 fprintf (dump_file, "Unrolling %u times.\n", unroll_factor);
2526 dta.chains = chains;
2527 dta.tmp_vars = tmp_vars;
2529 update_ssa (TODO_update_ssa_only_virtuals);
2531 /* Cfg manipulations performed in tree_transform_and_unroll_loop before
2532 execute_pred_commoning_cbck is called may cause phi nodes to be
2533 reallocated, which is a problem since CHAINS may point to these
2534 statements. To fix this, we store the ssa names defined by the
2535 phi nodes here instead of the phi nodes themselves, and restore
2536 the phi nodes in execute_pred_commoning_cbck. A bit hacky. */
2537 replace_phis_by_defined_names (chains);
2539 tree_transform_and_unroll_loop (loop, unroll_factor, exit, &desc,
2540 execute_pred_commoning_cbck, &dta);
2541 eliminate_temp_copies (loop, tmp_vars);
2545 if (dump_file && (dump_flags & TDF_DETAILS))
2547 "Executing predictive commoning without unrolling.\n");
2548 execute_pred_commoning (loop, chains, tmp_vars);
2552 release_chains (chains);
2553 free_data_refs (datarefs);
2554 BITMAP_FREE (tmp_vars);
2555 BITMAP_FREE (looparound_phis);
2557 free_affine_expand_cache (&name_expansions);
2562 /* Runs predictive commoning. */
2565 tree_predictive_commoning (void)
2567 bool unrolled = false;
2572 initialize_original_copy_tables ();
2573 FOR_EACH_LOOP (li, loop, LI_ONLY_INNERMOST)
2574 if (optimize_loop_for_speed_p (loop))
2576 unrolled |= tree_predictive_commoning_loop (loop);
2582 ret = TODO_cleanup_cfg;
2584 free_original_copy_tables ();