2 Copyright (C) 2006, 2007, 2008, 2009, 2010
3 Free Software Foundation, Inc.
4 Contributed by Georges-Andre Silber <Georges-Andre.Silber@ensmp.fr>
5 and Sebastian Pop <sebastian.pop@amd.com>.
7 This file is part of GCC.
9 GCC is free software; you can redistribute it and/or modify it
10 under the terms of the GNU General Public License as published by the
11 Free Software Foundation; either version 3, or (at your option) any
14 GCC is distributed in the hope that it will be useful, but WITHOUT
15 ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
16 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
19 You should have received a copy of the GNU General Public License
20 along with GCC; see the file COPYING3. If not see
21 <http://www.gnu.org/licenses/>. */
23 /* This pass performs loop distribution: for example, the loop
40 This pass uses an RDG, Reduced Dependence Graph built on top of the
41 data dependence relations. The RDG is then topologically sorted to
42 obtain a map of information producers/consumers based on which it
43 generates the new loops. */
47 #include "coretypes.h"
50 #include "basic-block.h"
51 #include "tree-flow.h"
52 #include "tree-dump.h"
55 #include "tree-chrec.h"
56 #include "tree-data-ref.h"
57 #include "tree-scalar-evolution.h"
58 #include "tree-pass.h"
60 #include "langhooks.h"
61 #include "tree-vectorizer.h"
63 /* If bit I is not set, it means that this node represents an
64 operation that has already been performed, and that should not be
65 performed again. This is the subgraph of remaining important
66 computations that is passed to the DFS algorithm for avoiding to
67 include several times the same stores in different loops. */
68 static bitmap remaining_stmts;
70 /* A node of the RDG is marked in this bitmap when it has as a
71 predecessor a node that writes to memory. */
72 static bitmap upstream_mem_writes;
74 /* Update the PHI nodes of NEW_LOOP. NEW_LOOP is a duplicate of
78 update_phis_for_loop_copy (struct loop *orig_loop, struct loop *new_loop)
81 gimple_stmt_iterator si_new, si_orig;
82 edge orig_loop_latch = loop_latch_edge (orig_loop);
83 edge orig_entry_e = loop_preheader_edge (orig_loop);
84 edge new_loop_entry_e = loop_preheader_edge (new_loop);
86 /* Scan the phis in the headers of the old and new loops
87 (they are organized in exactly the same order). */
88 for (si_new = gsi_start_phis (new_loop->header),
89 si_orig = gsi_start_phis (orig_loop->header);
90 !gsi_end_p (si_new) && !gsi_end_p (si_orig);
91 gsi_next (&si_new), gsi_next (&si_orig))
94 source_location locus;
95 gimple phi_new = gsi_stmt (si_new);
96 gimple phi_orig = gsi_stmt (si_orig);
98 /* Add the first phi argument for the phi in NEW_LOOP (the one
99 associated with the entry of NEW_LOOP) */
100 def = PHI_ARG_DEF_FROM_EDGE (phi_orig, orig_entry_e);
101 locus = gimple_phi_arg_location_from_edge (phi_orig, orig_entry_e);
102 add_phi_arg (phi_new, def, new_loop_entry_e, locus);
104 /* Add the second phi argument for the phi in NEW_LOOP (the one
105 associated with the latch of NEW_LOOP) */
106 def = PHI_ARG_DEF_FROM_EDGE (phi_orig, orig_loop_latch);
107 locus = gimple_phi_arg_location_from_edge (phi_orig, orig_loop_latch);
109 if (TREE_CODE (def) == SSA_NAME)
111 new_ssa_name = get_current_def (def);
114 /* This only happens if there are no definitions inside the
115 loop. Use the the invariant in the new loop as is. */
119 /* Could be an integer. */
122 add_phi_arg (phi_new, new_ssa_name, loop_latch_edge (new_loop), locus);
126 /* Return a copy of LOOP placed before LOOP. */
129 copy_loop_before (struct loop *loop)
132 edge preheader = loop_preheader_edge (loop);
134 if (!single_exit (loop))
137 initialize_original_copy_tables ();
138 res = slpeel_tree_duplicate_loop_to_edge_cfg (loop, preheader);
139 free_original_copy_tables ();
144 update_phis_for_loop_copy (loop, res);
145 rename_variables_in_loop (res);
150 /* Creates an empty basic block after LOOP. */
153 create_bb_after_loop (struct loop *loop)
155 edge exit = single_exit (loop);
163 /* Generate code for PARTITION from the code in LOOP. The loop is
164 copied when COPY_P is true. All the statements not flagged in the
165 PARTITION bitmap are removed from the loop or from its copy. The
166 statements are indexed in sequence inside a basic block, and the
167 basic blocks of a loop are taken in dom order. Returns true when
168 the code gen succeeded. */
171 generate_loops_for_partition (struct loop *loop, bitmap partition, bool copy_p)
174 gimple_stmt_iterator bsi;
179 loop = copy_loop_before (loop);
180 create_preheader (loop, CP_SIMPLE_PREHEADERS);
181 create_bb_after_loop (loop);
187 /* Remove stmts not in the PARTITION bitmap. The order in which we
188 visit the phi nodes and the statements is exactly as in
190 bbs = get_loop_body_in_dom_order (loop);
192 for (x = 0, i = 0; i < loop->num_nodes; i++)
194 basic_block bb = bbs[i];
196 for (bsi = gsi_start_phis (bb); !gsi_end_p (bsi);)
197 if (!bitmap_bit_p (partition, x++))
199 gimple phi = gsi_stmt (bsi);
200 if (!is_gimple_reg (gimple_phi_result (phi)))
201 mark_virtual_phi_result_for_renaming (phi);
202 remove_phi_node (&bsi, true);
207 for (bsi = gsi_start_bb (bb); !gsi_end_p (bsi);)
209 gimple stmt = gsi_stmt (bsi);
210 if (gimple_code (gsi_stmt (bsi)) != GIMPLE_LABEL
211 && !bitmap_bit_p (partition, x++))
213 unlink_stmt_vdef (stmt);
214 gsi_remove (&bsi, true);
226 /* Build the size argument for a memset call. */
229 build_size_arg_loc (location_t loc, tree nb_iter, tree op,
230 gimple_seq *stmt_list)
233 tree x = size_binop_loc (loc, MULT_EXPR,
234 fold_convert_loc (loc, sizetype, nb_iter),
235 TYPE_SIZE_UNIT (TREE_TYPE (op)));
236 x = force_gimple_operand (x, &stmts, true, NULL);
237 gimple_seq_add_seq (stmt_list, stmts);
242 /* Generate a call to memset. Return true when the operation succeeded. */
245 generate_memset_zero (gimple stmt, tree op0, tree nb_iter,
246 gimple_stmt_iterator bsi)
248 tree addr_base, nb_bytes;
250 gimple_seq stmt_list = NULL, stmts;
253 struct data_reference *dr = XCNEW (struct data_reference);
254 location_t loc = gimple_location (stmt);
258 if (!dr_analyze_innermost (dr))
261 if (!stride_of_unit_type_p (DR_STEP (dr), TREE_TYPE (op0)))
264 nb_bytes = build_size_arg_loc (loc, nb_iter, op0, &stmt_list);
265 addr_base = size_binop_loc (loc, PLUS_EXPR, DR_OFFSET (dr), DR_INIT (dr));
266 addr_base = fold_convert_loc (loc, sizetype, addr_base);
268 /* Test for a negative stride, iterating over every element. */
269 if (integer_zerop (size_binop (PLUS_EXPR,
270 TYPE_SIZE_UNIT (TREE_TYPE (op0)),
271 fold_convert (sizetype, DR_STEP (dr)))))
273 addr_base = size_binop_loc (loc, MINUS_EXPR, addr_base,
274 fold_convert_loc (loc, sizetype, nb_bytes));
275 addr_base = size_binop_loc (loc, PLUS_EXPR, addr_base,
276 TYPE_SIZE_UNIT (TREE_TYPE (op0)));
279 addr_base = fold_build2_loc (loc, POINTER_PLUS_EXPR,
280 TREE_TYPE (DR_BASE_ADDRESS (dr)),
281 DR_BASE_ADDRESS (dr), addr_base);
282 mem = force_gimple_operand (addr_base, &stmts, true, NULL);
283 gimple_seq_add_seq (&stmt_list, stmts);
285 fn = build_fold_addr_expr (implicit_built_in_decls [BUILT_IN_MEMSET]);
286 fn_call = gimple_build_call (fn, 3, mem, integer_zero_node, nb_bytes);
287 gimple_seq_add_stmt (&stmt_list, fn_call);
288 gsi_insert_seq_after (&bsi, stmt_list, GSI_CONTINUE_LINKING);
291 if (dump_file && (dump_flags & TDF_DETAILS))
292 fprintf (dump_file, "generated memset zero\n");
299 /* Tries to generate a builtin function for the instructions of LOOP
300 pointed to by the bits set in PARTITION. Returns true when the
301 operation succeeded. */
304 generate_builtin (struct loop *loop, bitmap partition, bool copy_p)
311 gimple_stmt_iterator bsi;
312 tree nb_iter = number_of_exit_cond_executions (loop);
314 if (!nb_iter || nb_iter == chrec_dont_know)
317 bbs = get_loop_body_in_dom_order (loop);
319 for (i = 0; i < loop->num_nodes; i++)
321 basic_block bb = bbs[i];
323 for (bsi = gsi_start_phis (bb); !gsi_end_p (bsi); gsi_next (&bsi))
326 for (bsi = gsi_start_bb (bb); !gsi_end_p (bsi); gsi_next (&bsi))
328 gimple stmt = gsi_stmt (bsi);
330 if (bitmap_bit_p (partition, x++)
331 && is_gimple_assign (stmt)
332 && !is_gimple_reg (gimple_assign_lhs (stmt)))
334 /* Don't generate the builtins when there are more than
340 if (bb == loop->latch)
341 nb_iter = number_of_latch_executions (loop);
349 op0 = gimple_assign_lhs (write);
350 op1 = gimple_assign_rhs1 (write);
352 if (!(TREE_CODE (op0) == ARRAY_REF
353 || TREE_CODE (op0) == MEM_REF))
356 /* The new statements will be placed before LOOP. */
357 bsi = gsi_last_bb (loop_preheader_edge (loop)->src);
359 if (gimple_assign_rhs_code (write) == INTEGER_CST
360 && (integer_zerop (op1) || real_zerop (op1)))
361 res = generate_memset_zero (write, op0, nb_iter, bsi);
363 /* If this is the last partition for which we generate code, we have
364 to destroy the loop. */
367 unsigned nbbs = loop->num_nodes;
368 edge exit = single_exit (loop);
369 basic_block src = loop_preheader_edge (loop)->src, dest = exit->dest;
370 redirect_edge_pred (exit, src);
371 exit->flags &= ~(EDGE_TRUE_VALUE|EDGE_FALSE_VALUE);
372 exit->flags |= EDGE_FALLTHRU;
373 cancel_loop_tree (loop);
374 rescan_loop_exit (exit, false, true);
376 for (i = 0; i < nbbs; i++)
377 delete_basic_block (bbs[i]);
379 set_immediate_dominator (CDI_DOMINATORS, dest,
380 recompute_dominator (CDI_DOMINATORS, dest));
388 /* Generates code for PARTITION. For simple loops, this function can
389 generate a built-in. */
392 generate_code_for_partition (struct loop *loop, bitmap partition, bool copy_p)
394 if (generate_builtin (loop, partition, copy_p))
397 return generate_loops_for_partition (loop, partition, copy_p);
401 /* Returns true if the node V of RDG cannot be recomputed. */
404 rdg_cannot_recompute_vertex_p (struct graph *rdg, int v)
406 if (RDG_MEM_WRITE_STMT (rdg, v))
412 /* Returns true when the vertex V has already been generated in the
413 current partition (V is in PROCESSED), or when V belongs to another
414 partition and cannot be recomputed (V is not in REMAINING_STMTS). */
417 already_processed_vertex_p (bitmap processed, int v)
419 return (bitmap_bit_p (processed, v)
420 || !bitmap_bit_p (remaining_stmts, v));
423 /* Returns NULL when there is no anti-dependence among the successors
424 of vertex V, otherwise returns the edge with the anti-dep. */
426 static struct graph_edge *
427 has_anti_dependence (struct vertex *v)
429 struct graph_edge *e;
432 for (e = v->succ; e; e = e->succ_next)
433 if (RDGE_TYPE (e) == anti_dd)
439 /* Returns true when V has an anti-dependence edge among its successors. */
442 predecessor_has_mem_write (struct graph *rdg, struct vertex *v)
444 struct graph_edge *e;
447 for (e = v->pred; e; e = e->pred_next)
448 if (bitmap_bit_p (upstream_mem_writes, e->src)
449 /* Don't consider flow channels: a write to memory followed
450 by a read from memory. These channels allow the split of
451 the RDG in different partitions. */
452 && !RDG_MEM_WRITE_STMT (rdg, e->src))
458 /* Initializes the upstream_mem_writes bitmap following the
459 information from RDG. */
462 mark_nodes_having_upstream_mem_writes (struct graph *rdg)
465 bitmap seen = BITMAP_ALLOC (NULL);
467 for (v = rdg->n_vertices - 1; v >= 0; v--)
468 if (!bitmap_bit_p (seen, v))
471 VEC (int, heap) *nodes = VEC_alloc (int, heap, 3);
473 graphds_dfs (rdg, &v, 1, &nodes, false, NULL);
475 FOR_EACH_VEC_ELT (int, nodes, i, x)
477 if (!bitmap_set_bit (seen, x))
480 if (RDG_MEM_WRITE_STMT (rdg, x)
481 || predecessor_has_mem_write (rdg, &(rdg->vertices[x]))
482 /* In anti dependences the read should occur before
483 the write, this is why both the read and the write
484 should be placed in the same partition. */
485 || has_anti_dependence (&(rdg->vertices[x])))
487 bitmap_set_bit (upstream_mem_writes, x);
491 VEC_free (int, heap, nodes);
495 /* Returns true when vertex u has a memory write node as a predecessor
499 has_upstream_mem_writes (int u)
501 return bitmap_bit_p (upstream_mem_writes, u);
504 static void rdg_flag_vertex_and_dependent (struct graph *, int, bitmap, bitmap,
507 /* Flag all the uses of U. */
510 rdg_flag_all_uses (struct graph *rdg, int u, bitmap partition, bitmap loops,
511 bitmap processed, bool *part_has_writes)
513 struct graph_edge *e;
515 for (e = rdg->vertices[u].succ; e; e = e->succ_next)
516 if (!bitmap_bit_p (processed, e->dest))
518 rdg_flag_vertex_and_dependent (rdg, e->dest, partition, loops,
519 processed, part_has_writes);
520 rdg_flag_all_uses (rdg, e->dest, partition, loops, processed,
525 /* Flag the uses of U stopping following the information from
526 upstream_mem_writes. */
529 rdg_flag_uses (struct graph *rdg, int u, bitmap partition, bitmap loops,
530 bitmap processed, bool *part_has_writes)
533 struct vertex *x = &(rdg->vertices[u]);
534 gimple stmt = RDGV_STMT (x);
535 struct graph_edge *anti_dep = has_anti_dependence (x);
537 /* Keep in the same partition the destination of an antidependence,
538 because this is a store to the exact same location. Putting this
539 in another partition is bad for cache locality. */
542 int v = anti_dep->dest;
544 if (!already_processed_vertex_p (processed, v))
545 rdg_flag_vertex_and_dependent (rdg, v, partition, loops,
546 processed, part_has_writes);
549 if (gimple_code (stmt) != GIMPLE_PHI)
551 if ((use_p = gimple_vuse_op (stmt)) != NULL_USE_OPERAND_P)
553 tree use = USE_FROM_PTR (use_p);
555 if (TREE_CODE (use) == SSA_NAME)
557 gimple def_stmt = SSA_NAME_DEF_STMT (use);
558 int v = rdg_vertex_for_stmt (rdg, def_stmt);
561 && !already_processed_vertex_p (processed, v))
562 rdg_flag_vertex_and_dependent (rdg, v, partition, loops,
563 processed, part_has_writes);
568 if (is_gimple_assign (stmt) && has_upstream_mem_writes (u))
570 tree op0 = gimple_assign_lhs (stmt);
572 /* Scalar channels don't have enough space for transmitting data
573 between tasks, unless we add more storage by privatizing. */
574 if (is_gimple_reg (op0))
577 imm_use_iterator iter;
579 FOR_EACH_IMM_USE_FAST (use_p, iter, op0)
581 int v = rdg_vertex_for_stmt (rdg, USE_STMT (use_p));
583 if (!already_processed_vertex_p (processed, v))
584 rdg_flag_vertex_and_dependent (rdg, v, partition, loops,
585 processed, part_has_writes);
591 /* Flag V from RDG as part of PARTITION, and also flag its loop number
595 rdg_flag_vertex (struct graph *rdg, int v, bitmap partition, bitmap loops,
596 bool *part_has_writes)
600 if (!bitmap_set_bit (partition, v))
603 loop = loop_containing_stmt (RDG_STMT (rdg, v));
604 bitmap_set_bit (loops, loop->num);
606 if (rdg_cannot_recompute_vertex_p (rdg, v))
608 *part_has_writes = true;
609 bitmap_clear_bit (remaining_stmts, v);
613 /* Flag in the bitmap PARTITION the vertex V and all its predecessors.
614 Also flag their loop number in LOOPS. */
617 rdg_flag_vertex_and_dependent (struct graph *rdg, int v, bitmap partition,
618 bitmap loops, bitmap processed,
619 bool *part_has_writes)
622 VEC (int, heap) *nodes = VEC_alloc (int, heap, 3);
625 bitmap_set_bit (processed, v);
626 rdg_flag_uses (rdg, v, partition, loops, processed, part_has_writes);
627 graphds_dfs (rdg, &v, 1, &nodes, false, remaining_stmts);
628 rdg_flag_vertex (rdg, v, partition, loops, part_has_writes);
630 FOR_EACH_VEC_ELT (int, nodes, i, x)
631 if (!already_processed_vertex_p (processed, x))
632 rdg_flag_vertex_and_dependent (rdg, x, partition, loops, processed,
635 VEC_free (int, heap, nodes);
638 /* Initialize CONDS with all the condition statements from the basic
642 collect_condition_stmts (struct loop *loop, VEC (gimple, heap) **conds)
646 VEC (edge, heap) *exits = get_loop_exit_edges (loop);
648 FOR_EACH_VEC_ELT (edge, exits, i, e)
650 gimple cond = last_stmt (e->src);
653 VEC_safe_push (gimple, heap, *conds, cond);
656 VEC_free (edge, heap, exits);
659 /* Add to PARTITION all the exit condition statements for LOOPS
660 together with all their dependent statements determined from
664 rdg_flag_loop_exits (struct graph *rdg, bitmap loops, bitmap partition,
665 bitmap processed, bool *part_has_writes)
669 VEC (gimple, heap) *conds = VEC_alloc (gimple, heap, 3);
671 EXECUTE_IF_SET_IN_BITMAP (loops, 0, i, bi)
672 collect_condition_stmts (get_loop (i), &conds);
674 while (!VEC_empty (gimple, conds))
676 gimple cond = VEC_pop (gimple, conds);
677 int v = rdg_vertex_for_stmt (rdg, cond);
678 bitmap new_loops = BITMAP_ALLOC (NULL);
680 if (!already_processed_vertex_p (processed, v))
681 rdg_flag_vertex_and_dependent (rdg, v, partition, new_loops, processed,
684 EXECUTE_IF_SET_IN_BITMAP (new_loops, 0, i, bi)
685 if (bitmap_set_bit (loops, i))
686 collect_condition_stmts (get_loop (i), &conds);
688 BITMAP_FREE (new_loops);
692 /* Flag all the nodes of RDG containing memory accesses that could
693 potentially belong to arrays already accessed in the current
697 rdg_flag_similar_memory_accesses (struct graph *rdg, bitmap partition,
698 bitmap loops, bitmap processed,
699 VEC (int, heap) **other_stores)
705 struct graph_edge *e;
707 EXECUTE_IF_SET_IN_BITMAP (partition, 0, i, ii)
708 if (RDG_MEM_WRITE_STMT (rdg, i)
709 || RDG_MEM_READS_STMT (rdg, i))
711 for (j = 0; j < rdg->n_vertices; j++)
712 if (!bitmap_bit_p (processed, j)
713 && (RDG_MEM_WRITE_STMT (rdg, j)
714 || RDG_MEM_READS_STMT (rdg, j))
715 && rdg_has_similar_memory_accesses (rdg, i, j))
717 /* Flag first the node J itself, and all the nodes that
718 are needed to compute J. */
719 rdg_flag_vertex_and_dependent (rdg, j, partition, loops,
722 /* When J is a read, we want to coalesce in the same
723 PARTITION all the nodes that are using J: this is
724 needed for better cache locality. */
725 rdg_flag_all_uses (rdg, j, partition, loops, processed, &foo);
727 /* Remove from OTHER_STORES the vertex that we flagged. */
728 if (RDG_MEM_WRITE_STMT (rdg, j))
729 FOR_EACH_VEC_ELT (int, *other_stores, k, kk)
732 VEC_unordered_remove (int, *other_stores, k);
737 /* If the node I has two uses, then keep these together in the
739 for (n = 0, e = rdg->vertices[i].succ; e; e = e->succ_next, n++);
742 rdg_flag_all_uses (rdg, i, partition, loops, processed, &foo);
746 /* Returns a bitmap in which all the statements needed for computing
747 the strongly connected component C of the RDG are flagged, also
748 including the loop exit conditions. */
751 build_rdg_partition_for_component (struct graph *rdg, rdgc c,
752 bool *part_has_writes,
753 VEC (int, heap) **other_stores)
756 bitmap partition = BITMAP_ALLOC (NULL);
757 bitmap loops = BITMAP_ALLOC (NULL);
758 bitmap processed = BITMAP_ALLOC (NULL);
760 FOR_EACH_VEC_ELT (int, c->vertices, i, v)
761 if (!already_processed_vertex_p (processed, v))
762 rdg_flag_vertex_and_dependent (rdg, v, partition, loops, processed,
765 /* Also iterate on the array of stores not in the starting vertices,
766 and determine those vertices that have some memory affinity with
767 the current nodes in the component: these are stores to the same
768 arrays, i.e. we're taking care of cache locality. */
769 if (!flag_tree_loop_distribute_patterns)
770 rdg_flag_similar_memory_accesses (rdg, partition, loops, processed,
773 rdg_flag_loop_exits (rdg, loops, partition, processed, part_has_writes);
775 BITMAP_FREE (processed);
780 /* Free memory for COMPONENTS. */
783 free_rdg_components (VEC (rdgc, heap) *components)
788 FOR_EACH_VEC_ELT (rdgc, components, i, x)
790 VEC_free (int, heap, x->vertices);
795 /* Build the COMPONENTS vector with the strongly connected components
796 of RDG in which the STARTING_VERTICES occur. */
799 rdg_build_components (struct graph *rdg, VEC (int, heap) *starting_vertices,
800 VEC (rdgc, heap) **components)
803 bitmap saved_components = BITMAP_ALLOC (NULL);
804 int n_components = graphds_scc (rdg, NULL);
805 VEC (int, heap) **all_components = XNEWVEC (VEC (int, heap) *, n_components);
807 for (i = 0; i < n_components; i++)
808 all_components[i] = VEC_alloc (int, heap, 3);
810 for (i = 0; i < rdg->n_vertices; i++)
811 VEC_safe_push (int, heap, all_components[rdg->vertices[i].component], i);
813 FOR_EACH_VEC_ELT (int, starting_vertices, i, v)
815 int c = rdg->vertices[v].component;
817 if (bitmap_set_bit (saved_components, c))
819 rdgc x = XCNEW (struct rdg_component);
821 x->vertices = all_components[c];
823 VEC_safe_push (rdgc, heap, *components, x);
827 for (i = 0; i < n_components; i++)
828 if (!bitmap_bit_p (saved_components, i))
829 VEC_free (int, heap, all_components[i]);
831 free (all_components);
832 BITMAP_FREE (saved_components);
835 /* Aggregate several components into a useful partition that is
836 registered in the PARTITIONS vector. Partitions will be
837 distributed in different loops. */
840 rdg_build_partitions (struct graph *rdg, VEC (rdgc, heap) *components,
841 VEC (int, heap) **other_stores,
842 VEC (bitmap, heap) **partitions, bitmap processed)
846 bitmap partition = BITMAP_ALLOC (NULL);
848 FOR_EACH_VEC_ELT (rdgc, components, i, x)
851 bool part_has_writes = false;
852 int v = VEC_index (int, x->vertices, 0);
854 if (bitmap_bit_p (processed, v))
857 np = build_rdg_partition_for_component (rdg, x, &part_has_writes,
859 bitmap_ior_into (partition, np);
860 bitmap_ior_into (processed, np);
865 if (dump_file && (dump_flags & TDF_DETAILS))
867 fprintf (dump_file, "ldist useful partition:\n");
868 dump_bitmap (dump_file, partition);
871 VEC_safe_push (bitmap, heap, *partitions, partition);
872 partition = BITMAP_ALLOC (NULL);
876 /* Add the nodes from the RDG that were not marked as processed, and
877 that are used outside the current loop. These are scalar
878 computations that are not yet part of previous partitions. */
879 for (i = 0; i < rdg->n_vertices; i++)
880 if (!bitmap_bit_p (processed, i)
881 && rdg_defs_used_in_other_loops_p (rdg, i))
882 VEC_safe_push (int, heap, *other_stores, i);
884 /* If there are still statements left in the OTHER_STORES array,
885 create other components and partitions with these stores and
886 their dependences. */
887 if (VEC_length (int, *other_stores) > 0)
889 VEC (rdgc, heap) *comps = VEC_alloc (rdgc, heap, 3);
890 VEC (int, heap) *foo = VEC_alloc (int, heap, 3);
892 rdg_build_components (rdg, *other_stores, &comps);
893 rdg_build_partitions (rdg, comps, &foo, partitions, processed);
895 VEC_free (int, heap, foo);
896 free_rdg_components (comps);
899 /* If there is something left in the last partition, save it. */
900 if (bitmap_count_bits (partition) > 0)
901 VEC_safe_push (bitmap, heap, *partitions, partition);
903 BITMAP_FREE (partition);
906 /* Dump to FILE the PARTITIONS. */
909 dump_rdg_partitions (FILE *file, VEC (bitmap, heap) *partitions)
914 FOR_EACH_VEC_ELT (bitmap, partitions, i, partition)
915 debug_bitmap_file (file, partition);
918 /* Debug PARTITIONS. */
919 extern void debug_rdg_partitions (VEC (bitmap, heap) *);
922 debug_rdg_partitions (VEC (bitmap, heap) *partitions)
924 dump_rdg_partitions (stderr, partitions);
927 /* Returns the number of read and write operations in the RDG. */
930 number_of_rw_in_rdg (struct graph *rdg)
934 for (i = 0; i < rdg->n_vertices; i++)
936 if (RDG_MEM_WRITE_STMT (rdg, i))
939 if (RDG_MEM_READS_STMT (rdg, i))
946 /* Returns the number of read and write operations in a PARTITION of
950 number_of_rw_in_partition (struct graph *rdg, bitmap partition)
956 EXECUTE_IF_SET_IN_BITMAP (partition, 0, i, ii)
958 if (RDG_MEM_WRITE_STMT (rdg, i))
961 if (RDG_MEM_READS_STMT (rdg, i))
968 /* Returns true when one of the PARTITIONS contains all the read or
969 write operations of RDG. */
972 partition_contains_all_rw (struct graph *rdg, VEC (bitmap, heap) *partitions)
976 int nrw = number_of_rw_in_rdg (rdg);
978 FOR_EACH_VEC_ELT (bitmap, partitions, i, partition)
979 if (nrw == number_of_rw_in_partition (rdg, partition))
985 /* Generate code from STARTING_VERTICES in RDG. Returns the number of
986 distributed loops. */
989 ldist_gen (struct loop *loop, struct graph *rdg,
990 VEC (int, heap) *starting_vertices)
993 VEC (rdgc, heap) *components = VEC_alloc (rdgc, heap, 3);
994 VEC (bitmap, heap) *partitions = VEC_alloc (bitmap, heap, 3);
995 VEC (int, heap) *other_stores = VEC_alloc (int, heap, 3);
996 bitmap partition, processed = BITMAP_ALLOC (NULL);
998 remaining_stmts = BITMAP_ALLOC (NULL);
999 upstream_mem_writes = BITMAP_ALLOC (NULL);
1001 for (i = 0; i < rdg->n_vertices; i++)
1003 bitmap_set_bit (remaining_stmts, i);
1005 /* Save in OTHER_STORES all the memory writes that are not in
1006 STARTING_VERTICES. */
1007 if (RDG_MEM_WRITE_STMT (rdg, i))
1013 FOR_EACH_VEC_ELT (int, starting_vertices, j, v)
1021 VEC_safe_push (int, heap, other_stores, i);
1025 mark_nodes_having_upstream_mem_writes (rdg);
1026 rdg_build_components (rdg, starting_vertices, &components);
1027 rdg_build_partitions (rdg, components, &other_stores, &partitions,
1029 BITMAP_FREE (processed);
1030 nbp = VEC_length (bitmap, partitions);
1033 || partition_contains_all_rw (rdg, partitions))
1036 if (dump_file && (dump_flags & TDF_DETAILS))
1037 dump_rdg_partitions (dump_file, partitions);
1039 FOR_EACH_VEC_ELT (bitmap, partitions, i, partition)
1040 if (!generate_code_for_partition (loop, partition, i < nbp - 1))
1043 rewrite_into_loop_closed_ssa (NULL, TODO_update_ssa);
1044 update_ssa (TODO_update_ssa_only_virtuals | TODO_update_ssa);
1048 BITMAP_FREE (remaining_stmts);
1049 BITMAP_FREE (upstream_mem_writes);
1051 FOR_EACH_VEC_ELT (bitmap, partitions, i, partition)
1052 BITMAP_FREE (partition);
1054 VEC_free (int, heap, other_stores);
1055 VEC_free (bitmap, heap, partitions);
1056 free_rdg_components (components);
1060 /* Distributes the code from LOOP in such a way that producer
1061 statements are placed before consumer statements. When STMTS is
1062 NULL, performs the maximal distribution, if STMTS is not NULL,
1063 tries to separate only these statements from the LOOP's body.
1064 Returns the number of distributed loops. */
1067 distribute_loop (struct loop *loop, VEC (gimple, heap) *stmts)
1073 VEC (int, heap) *vertices;
1075 if (loop->num_nodes > 2)
1077 if (dump_file && (dump_flags & TDF_DETAILS))
1079 "FIXME: Loop %d not distributed: it has more than two basic blocks.\n",
1085 rdg = build_rdg (loop);
1089 if (dump_file && (dump_flags & TDF_DETAILS))
1091 "FIXME: Loop %d not distributed: failed to build the RDG.\n",
1097 vertices = VEC_alloc (int, heap, 3);
1099 if (dump_file && (dump_flags & TDF_DETAILS))
1100 dump_rdg (dump_file, rdg);
1102 FOR_EACH_VEC_ELT (gimple, stmts, i, s)
1104 int v = rdg_vertex_for_stmt (rdg, s);
1108 VEC_safe_push (int, heap, vertices, v);
1110 if (dump_file && (dump_flags & TDF_DETAILS))
1112 "ldist asked to generate code for vertex %d\n", v);
1116 res = ldist_gen (loop, rdg, vertices);
1117 VEC_free (int, heap, vertices);
1123 /* Distribute all loops in the current function. */
1126 tree_loop_distribution (void)
1130 int nb_generated_loops = 0;
1132 FOR_EACH_LOOP (li, loop, 0)
1134 VEC (gimple, heap) *work_list = NULL;
1136 /* If the loop doesn't have a single exit we will fail anyway,
1137 so do that early. */
1138 if (!single_exit (loop))
1141 /* If both flag_tree_loop_distribute_patterns and
1142 flag_tree_loop_distribution are set, then only
1143 distribute_patterns is executed. */
1144 if (flag_tree_loop_distribute_patterns)
1146 /* With the following working list, we're asking
1147 distribute_loop to separate from the rest of the loop the
1148 stores of the form "A[i] = 0". */
1149 stores_zero_from_loop (loop, &work_list);
1151 /* Do nothing if there are no patterns to be distributed. */
1152 if (VEC_length (gimple, work_list) > 0)
1153 nb_generated_loops = distribute_loop (loop, work_list);
1155 else if (flag_tree_loop_distribution)
1157 /* With the following working list, we're asking
1158 distribute_loop to separate the stores of the loop: when
1159 dependences allow, it will end on having one store per
1161 stores_from_loop (loop, &work_list);
1163 /* A simple heuristic for cache locality is to not split
1164 stores to the same array. Without this call, an unrolled
1165 loop would be split into as many loops as unroll factor,
1166 each loop storing in the same array. */
1167 remove_similar_memory_refs (&work_list);
1169 nb_generated_loops = distribute_loop (loop, work_list);
1172 if (dump_file && (dump_flags & TDF_DETAILS))
1174 if (nb_generated_loops > 1)
1175 fprintf (dump_file, "Loop %d distributed: split to %d loops.\n",
1176 loop->num, nb_generated_loops);
1178 fprintf (dump_file, "Loop %d is the same.\n", loop->num);
1181 verify_loop_structure ();
1183 VEC_free (gimple, heap, work_list);
1190 gate_tree_loop_distribution (void)
1192 return flag_tree_loop_distribution
1193 || flag_tree_loop_distribute_patterns;
1196 struct gimple_opt_pass pass_loop_distribution =
1201 gate_tree_loop_distribution, /* gate */
1202 tree_loop_distribution, /* execute */
1205 0, /* static_pass_number */
1206 TV_TREE_LOOP_DISTRIBUTION, /* tv_id */
1207 PROP_cfg | PROP_ssa, /* properties_required */
1208 0, /* properties_provided */
1209 0, /* properties_destroyed */
1210 0, /* todo_flags_start */
1211 TODO_dump_func /* todo_flags_finish */