2 Copyright (C) 2006, 2007, 2008, 2009 Free Software Foundation, Inc.
3 Contributed by Georges-Andre Silber <Georges-Andre.Silber@ensmp.fr>
4 and Sebastian Pop <sebastian.pop@amd.com>.
6 This file is part of GCC.
8 GCC is free software; you can redistribute it and/or modify it
9 under the terms of the GNU General Public License as published by the
10 Free Software Foundation; either version 3, or (at your option) any
13 GCC is distributed in the hope that it will be useful, but WITHOUT
14 ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
15 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
18 You should have received a copy of the GNU General Public License
19 along with GCC; see the file COPYING3. If not see
20 <http://www.gnu.org/licenses/>. */
22 /* This pass performs loop distribution: for example, the loop
39 This pass uses an RDG, Reduced Dependence Graph built on top of the
40 data dependence relations. The RDG is then topologically sorted to
41 obtain a map of information producers/consumers based on which it
42 generates the new loops. */
46 #include "coretypes.h"
53 #include "basic-block.h"
54 #include "diagnostic.h"
55 #include "tree-flow.h"
56 #include "tree-dump.h"
61 #include "tree-chrec.h"
62 #include "tree-data-ref.h"
63 #include "tree-scalar-evolution.h"
64 #include "tree-pass.h"
66 #include "langhooks.h"
67 #include "tree-vectorizer.h"
69 /* If bit I is not set, it means that this node represents an
70 operation that has already been performed, and that should not be
71 performed again. This is the subgraph of remaining important
72 computations that is passed to the DFS algorithm for avoiding to
73 include several times the same stores in different loops. */
74 static bitmap remaining_stmts;
76 /* A node of the RDG is marked in this bitmap when it has as a
77 predecessor a node that writes to memory. */
78 static bitmap upstream_mem_writes;
80 /* Update the PHI nodes of NEW_LOOP. NEW_LOOP is a duplicate of
84 update_phis_for_loop_copy (struct loop *orig_loop, struct loop *new_loop)
87 gimple_stmt_iterator si_new, si_orig;
88 edge orig_loop_latch = loop_latch_edge (orig_loop);
89 edge orig_entry_e = loop_preheader_edge (orig_loop);
90 edge new_loop_entry_e = loop_preheader_edge (new_loop);
92 /* Scan the phis in the headers of the old and new loops
93 (they are organized in exactly the same order). */
94 for (si_new = gsi_start_phis (new_loop->header),
95 si_orig = gsi_start_phis (orig_loop->header);
96 !gsi_end_p (si_new) && !gsi_end_p (si_orig);
97 gsi_next (&si_new), gsi_next (&si_orig))
100 gimple phi_new = gsi_stmt (si_new);
101 gimple phi_orig = gsi_stmt (si_orig);
103 /* Add the first phi argument for the phi in NEW_LOOP (the one
104 associated with the entry of NEW_LOOP) */
105 def = PHI_ARG_DEF_FROM_EDGE (phi_orig, orig_entry_e);
106 add_phi_arg (phi_new, def, new_loop_entry_e);
108 /* Add the second phi argument for the phi in NEW_LOOP (the one
109 associated with the latch of NEW_LOOP) */
110 def = PHI_ARG_DEF_FROM_EDGE (phi_orig, orig_loop_latch);
112 if (TREE_CODE (def) == SSA_NAME)
114 new_ssa_name = get_current_def (def);
117 /* This only happens if there are no definitions inside the
118 loop. Use the phi_result in this case. */
119 new_ssa_name = PHI_RESULT (phi_new);
122 /* Could be an integer. */
125 add_phi_arg (phi_new, new_ssa_name, loop_latch_edge (new_loop));
129 /* Return a copy of LOOP placed before LOOP. */
132 copy_loop_before (struct loop *loop)
135 edge preheader = loop_preheader_edge (loop);
137 if (!single_exit (loop))
140 initialize_original_copy_tables ();
141 res = slpeel_tree_duplicate_loop_to_edge_cfg (loop, preheader);
142 free_original_copy_tables ();
147 update_phis_for_loop_copy (loop, res);
148 rename_variables_in_loop (res);
153 /* Creates an empty basic block after LOOP. */
156 create_bb_after_loop (struct loop *loop)
158 edge exit = single_exit (loop);
166 /* Generate code for PARTITION from the code in LOOP. The loop is
167 copied when COPY_P is true. All the statements not flagged in the
168 PARTITION bitmap are removed from the loop or from its copy. The
169 statements are indexed in sequence inside a basic block, and the
170 basic blocks of a loop are taken in dom order. Returns true when
171 the code gen succeeded. */
174 generate_loops_for_partition (struct loop *loop, bitmap partition, bool copy_p)
177 gimple_stmt_iterator bsi;
182 loop = copy_loop_before (loop);
183 create_preheader (loop, CP_SIMPLE_PREHEADERS);
184 create_bb_after_loop (loop);
190 /* Remove stmts not in the PARTITION bitmap. The order in which we
191 visit the phi nodes and the statements is exactly as in
193 bbs = get_loop_body_in_dom_order (loop);
195 for (x = 0, i = 0; i < loop->num_nodes; i++)
197 basic_block bb = bbs[i];
199 for (bsi = gsi_start_phis (bb); !gsi_end_p (bsi);)
200 if (!bitmap_bit_p (partition, x++))
201 remove_phi_node (&bsi, true);
205 for (bsi = gsi_start_bb (bb); !gsi_end_p (bsi);)
206 if (gimple_code (gsi_stmt (bsi)) != GIMPLE_LABEL
207 && !bitmap_bit_p (partition, x++))
208 gsi_remove (&bsi, false);
212 mark_virtual_ops_in_bb (bb);
219 /* Build the size argument for a memset call. */
222 build_size_arg_loc (location_t loc, tree nb_iter, tree op, gimple_seq* stmt_list)
225 gimple_seq stmts = NULL;
227 nb_bytes = fold_build2_loc (loc, MULT_EXPR, size_type_node,
228 fold_convert_loc (loc, size_type_node, nb_iter),
229 fold_convert_loc (loc, size_type_node,
230 TYPE_SIZE_UNIT (TREE_TYPE (op))));
231 nb_bytes = force_gimple_operand (nb_bytes, &stmts, true, NULL);
232 gimple_seq_add_seq (stmt_list, stmts);
237 /* Generate a call to memset. Return true when the operation succeeded. */
240 generate_memset_zero (gimple stmt, tree op0, tree nb_iter,
241 gimple_stmt_iterator bsi)
244 tree nb_bytes = NULL;
246 gimple_seq stmts = NULL, stmt_list = NULL;
248 tree mem, fndecl, fntype, fn;
249 gimple_stmt_iterator i;
250 struct data_reference *dr = XCNEW (struct data_reference);
251 location_t loc = gimple_location (stmt);
255 if (!dr_analyze_innermost (dr))
258 /* Test for a positive stride, iterating over every element. */
259 if (integer_zerop (fold_build2_loc (loc,
260 MINUS_EXPR, integer_type_node, DR_STEP (dr),
261 TYPE_SIZE_UNIT (TREE_TYPE (op0)))))
263 tree offset = fold_convert_loc (loc, sizetype,
264 size_binop_loc (loc, PLUS_EXPR,
267 addr_base = fold_build2_loc (loc, POINTER_PLUS_EXPR,
268 TREE_TYPE (DR_BASE_ADDRESS (dr)),
269 DR_BASE_ADDRESS (dr), offset);
272 /* Test for a negative stride, iterating over every element. */
273 else if (integer_zerop (fold_build2_loc (loc, PLUS_EXPR, integer_type_node,
274 TYPE_SIZE_UNIT (TREE_TYPE (op0)),
277 nb_bytes = build_size_arg_loc (loc, nb_iter, op0, &stmt_list);
278 addr_base = size_binop_loc (loc, PLUS_EXPR, DR_OFFSET (dr), DR_INIT (dr));
279 addr_base = fold_build2_loc (loc, MINUS_EXPR, sizetype, addr_base,
280 fold_convert_loc (loc, sizetype, nb_bytes));
281 addr_base = force_gimple_operand (addr_base, &stmts, true, NULL);
282 gimple_seq_add_seq (&stmt_list, stmts);
284 addr_base = fold_build2_loc (loc, POINTER_PLUS_EXPR,
285 TREE_TYPE (DR_BASE_ADDRESS (dr)),
286 DR_BASE_ADDRESS (dr), addr_base);
291 mem = force_gimple_operand (addr_base, &stmts, true, NULL);
292 gimple_seq_add_seq (&stmt_list, stmts);
294 fndecl = implicit_built_in_decls [BUILT_IN_MEMSET];
295 fntype = TREE_TYPE (fndecl);
296 fn = build1 (ADDR_EXPR, build_pointer_type (fntype), fndecl);
299 nb_bytes = build_size_arg_loc (loc, nb_iter, op0, &stmt_list);
300 fn_call = gimple_build_call (fn, 3, mem, integer_zero_node, nb_bytes);
301 gimple_seq_add_stmt (&stmt_list, fn_call);
303 for (i = gsi_start (stmt_list); !gsi_end_p (i); gsi_next (&i))
305 gimple s = gsi_stmt (i);
306 update_stmt_if_modified (s);
309 gsi_insert_seq_after (&bsi, stmt_list, GSI_CONTINUE_LINKING);
312 if (dump_file && (dump_flags & TDF_DETAILS))
313 fprintf (dump_file, "generated memset zero\n");
320 /* Propagate phis in BB b to their uses and remove them. */
323 prop_phis (basic_block b)
325 gimple_stmt_iterator psi;
326 gimple_seq phis = phi_nodes (b);
328 for (psi = gsi_start (phis); !gsi_end_p (psi); )
330 gimple phi = gsi_stmt (psi);
331 tree def = gimple_phi_result (phi), use = gimple_phi_arg_def (phi, 0);
333 gcc_assert (gimple_phi_num_args (phi) == 1);
335 if (!is_gimple_reg (def))
337 imm_use_iterator iter;
341 FOR_EACH_IMM_USE_STMT (stmt, iter, def)
342 FOR_EACH_IMM_USE_ON_STMT (use_p, iter)
343 SET_USE (use_p, use);
346 replace_uses_by (def, use);
348 remove_phi_node (&psi, true);
352 /* Tries to generate a builtin function for the instructions of LOOP
353 pointed to by the bits set in PARTITION. Returns true when the
354 operation succeeded. */
357 generate_builtin (struct loop *loop, bitmap partition, bool copy_p)
364 gimple_stmt_iterator bsi;
365 tree nb_iter = number_of_exit_cond_executions (loop);
367 if (!nb_iter || nb_iter == chrec_dont_know)
370 bbs = get_loop_body_in_dom_order (loop);
372 for (i = 0; i < loop->num_nodes; i++)
374 basic_block bb = bbs[i];
376 for (bsi = gsi_start_phis (bb); !gsi_end_p (bsi); gsi_next (&bsi))
379 for (bsi = gsi_start_bb (bb); !gsi_end_p (bsi); gsi_next (&bsi))
381 gimple stmt = gsi_stmt (bsi);
383 if (bitmap_bit_p (partition, x++)
384 && is_gimple_assign (stmt)
385 && !is_gimple_reg (gimple_assign_lhs (stmt)))
387 /* Don't generate the builtins when there are more than
400 op0 = gimple_assign_lhs (write);
401 op1 = gimple_assign_rhs1 (write);
403 if (!(TREE_CODE (op0) == ARRAY_REF
404 || TREE_CODE (op0) == INDIRECT_REF))
407 /* The new statements will be placed before LOOP. */
408 bsi = gsi_last_bb (loop_preheader_edge (loop)->src);
410 if (gimple_assign_rhs_code (write) == INTEGER_CST
411 && (integer_zerop (op1) || real_zerop (op1)))
412 res = generate_memset_zero (write, op0, nb_iter, bsi);
414 /* If this is the last partition for which we generate code, we have
415 to destroy the loop. */
418 unsigned nbbs = loop->num_nodes;
419 basic_block src = loop_preheader_edge (loop)->src;
420 basic_block dest = single_exit (loop)->dest;
422 make_edge (src, dest, EDGE_FALLTHRU);
423 cancel_loop_tree (loop);
425 for (i = 0; i < nbbs; i++)
426 delete_basic_block (bbs[i]);
428 set_immediate_dominator (CDI_DOMINATORS, dest,
429 recompute_dominator (CDI_DOMINATORS, dest));
437 /* Generates code for PARTITION. For simple loops, this function can
438 generate a built-in. */
441 generate_code_for_partition (struct loop *loop, bitmap partition, bool copy_p)
443 if (generate_builtin (loop, partition, copy_p))
446 return generate_loops_for_partition (loop, partition, copy_p);
450 /* Returns true if the node V of RDG cannot be recomputed. */
453 rdg_cannot_recompute_vertex_p (struct graph *rdg, int v)
455 if (RDG_MEM_WRITE_STMT (rdg, v))
461 /* Returns true when the vertex V has already been generated in the
462 current partition (V is in PROCESSED), or when V belongs to another
463 partition and cannot be recomputed (V is not in REMAINING_STMTS). */
466 already_processed_vertex_p (bitmap processed, int v)
468 return (bitmap_bit_p (processed, v)
469 || !bitmap_bit_p (remaining_stmts, v));
472 /* Returns NULL when there is no anti-dependence among the successors
473 of vertex V, otherwise returns the edge with the anti-dep. */
475 static struct graph_edge *
476 has_anti_dependence (struct vertex *v)
478 struct graph_edge *e;
481 for (e = v->succ; e; e = e->succ_next)
482 if (RDGE_TYPE (e) == anti_dd)
488 /* Returns true when V has an anti-dependence edge among its successors. */
491 predecessor_has_mem_write (struct graph *rdg, struct vertex *v)
493 struct graph_edge *e;
496 for (e = v->pred; e; e = e->pred_next)
497 if (bitmap_bit_p (upstream_mem_writes, e->src)
498 /* Don't consider flow channels: a write to memory followed
499 by a read from memory. These channels allow the split of
500 the RDG in different partitions. */
501 && !RDG_MEM_WRITE_STMT (rdg, e->src))
507 /* Initializes the upstream_mem_writes bitmap following the
508 information from RDG. */
511 mark_nodes_having_upstream_mem_writes (struct graph *rdg)
514 bitmap seen = BITMAP_ALLOC (NULL);
516 for (v = rdg->n_vertices - 1; v >= 0; v--)
517 if (!bitmap_bit_p (seen, v))
520 VEC (int, heap) *nodes = VEC_alloc (int, heap, 3);
521 bool has_upstream_mem_write_p = false;
523 graphds_dfs (rdg, &v, 1, &nodes, false, NULL);
525 for (i = 0; VEC_iterate (int, nodes, i, x); i++)
527 if (bitmap_bit_p (seen, x))
530 bitmap_set_bit (seen, x);
532 if (RDG_MEM_WRITE_STMT (rdg, x)
533 || predecessor_has_mem_write (rdg, &(rdg->vertices[x]))
534 /* In anti dependences the read should occur before
535 the write, this is why both the read and the write
536 should be placed in the same partition. */
537 || has_anti_dependence (&(rdg->vertices[x])))
539 has_upstream_mem_write_p = true;
540 bitmap_set_bit (upstream_mem_writes, x);
544 VEC_free (int, heap, nodes);
548 /* Returns true when vertex u has a memory write node as a predecessor
552 has_upstream_mem_writes (int u)
554 return bitmap_bit_p (upstream_mem_writes, u);
557 static void rdg_flag_vertex_and_dependent (struct graph *, int, bitmap, bitmap,
560 /* Flag all the uses of U. */
563 rdg_flag_all_uses (struct graph *rdg, int u, bitmap partition, bitmap loops,
564 bitmap processed, bool *part_has_writes)
566 struct graph_edge *e;
568 for (e = rdg->vertices[u].succ; e; e = e->succ_next)
569 if (!bitmap_bit_p (processed, e->dest))
571 rdg_flag_vertex_and_dependent (rdg, e->dest, partition, loops,
572 processed, part_has_writes);
573 rdg_flag_all_uses (rdg, e->dest, partition, loops, processed,
578 /* Flag the uses of U stopping following the information from
579 upstream_mem_writes. */
582 rdg_flag_uses (struct graph *rdg, int u, bitmap partition, bitmap loops,
583 bitmap processed, bool *part_has_writes)
586 struct vertex *x = &(rdg->vertices[u]);
587 gimple stmt = RDGV_STMT (x);
588 struct graph_edge *anti_dep = has_anti_dependence (x);
590 /* Keep in the same partition the destination of an antidependence,
591 because this is a store to the exact same location. Putting this
592 in another partition is bad for cache locality. */
595 int v = anti_dep->dest;
597 if (!already_processed_vertex_p (processed, v))
598 rdg_flag_vertex_and_dependent (rdg, v, partition, loops,
599 processed, part_has_writes);
602 if (gimple_code (stmt) != GIMPLE_PHI)
604 if ((use_p = gimple_vuse_op (stmt)) != NULL_USE_OPERAND_P)
606 tree use = USE_FROM_PTR (use_p);
608 if (TREE_CODE (use) == SSA_NAME)
610 gimple def_stmt = SSA_NAME_DEF_STMT (use);
611 int v = rdg_vertex_for_stmt (rdg, def_stmt);
614 && !already_processed_vertex_p (processed, v))
615 rdg_flag_vertex_and_dependent (rdg, v, partition, loops,
616 processed, part_has_writes);
621 if (is_gimple_assign (stmt) && has_upstream_mem_writes (u))
623 tree op0 = gimple_assign_lhs (stmt);
625 /* Scalar channels don't have enough space for transmitting data
626 between tasks, unless we add more storage by privatizing. */
627 if (is_gimple_reg (op0))
630 imm_use_iterator iter;
632 FOR_EACH_IMM_USE_FAST (use_p, iter, op0)
634 int v = rdg_vertex_for_stmt (rdg, USE_STMT (use_p));
636 if (!already_processed_vertex_p (processed, v))
637 rdg_flag_vertex_and_dependent (rdg, v, partition, loops,
638 processed, part_has_writes);
644 /* Flag V from RDG as part of PARTITION, and also flag its loop number
648 rdg_flag_vertex (struct graph *rdg, int v, bitmap partition, bitmap loops,
649 bool *part_has_writes)
653 if (bitmap_bit_p (partition, v))
656 loop = loop_containing_stmt (RDG_STMT (rdg, v));
657 bitmap_set_bit (loops, loop->num);
658 bitmap_set_bit (partition, v);
660 if (rdg_cannot_recompute_vertex_p (rdg, v))
662 *part_has_writes = true;
663 bitmap_clear_bit (remaining_stmts, v);
667 /* Flag in the bitmap PARTITION the vertex V and all its predecessors.
668 Also flag their loop number in LOOPS. */
671 rdg_flag_vertex_and_dependent (struct graph *rdg, int v, bitmap partition,
672 bitmap loops, bitmap processed,
673 bool *part_has_writes)
676 VEC (int, heap) *nodes = VEC_alloc (int, heap, 3);
679 bitmap_set_bit (processed, v);
680 rdg_flag_uses (rdg, v, partition, loops, processed, part_has_writes);
681 graphds_dfs (rdg, &v, 1, &nodes, false, remaining_stmts);
682 rdg_flag_vertex (rdg, v, partition, loops, part_has_writes);
684 for (i = 0; VEC_iterate (int, nodes, i, x); i++)
685 if (!already_processed_vertex_p (processed, x))
686 rdg_flag_vertex_and_dependent (rdg, x, partition, loops, processed,
689 VEC_free (int, heap, nodes);
692 /* Initialize CONDS with all the condition statements from the basic
696 collect_condition_stmts (struct loop *loop, VEC (gimple, heap) **conds)
700 VEC (edge, heap) *exits = get_loop_exit_edges (loop);
702 for (i = 0; VEC_iterate (edge, exits, i, e); i++)
704 gimple cond = last_stmt (e->src);
707 VEC_safe_push (gimple, heap, *conds, cond);
710 VEC_free (edge, heap, exits);
713 /* Add to PARTITION all the exit condition statements for LOOPS
714 together with all their dependent statements determined from
718 rdg_flag_loop_exits (struct graph *rdg, bitmap loops, bitmap partition,
719 bitmap processed, bool *part_has_writes)
723 VEC (gimple, heap) *conds = VEC_alloc (gimple, heap, 3);
725 EXECUTE_IF_SET_IN_BITMAP (loops, 0, i, bi)
726 collect_condition_stmts (get_loop (i), &conds);
728 while (!VEC_empty (gimple, conds))
730 gimple cond = VEC_pop (gimple, conds);
731 int v = rdg_vertex_for_stmt (rdg, cond);
732 bitmap new_loops = BITMAP_ALLOC (NULL);
734 if (!already_processed_vertex_p (processed, v))
735 rdg_flag_vertex_and_dependent (rdg, v, partition, new_loops, processed,
738 EXECUTE_IF_SET_IN_BITMAP (new_loops, 0, i, bi)
739 if (!bitmap_bit_p (loops, i))
741 bitmap_set_bit (loops, i);
742 collect_condition_stmts (get_loop (i), &conds);
745 BITMAP_FREE (new_loops);
749 /* Flag all the nodes of RDG containing memory accesses that could
750 potentially belong to arrays already accessed in the current
754 rdg_flag_similar_memory_accesses (struct graph *rdg, bitmap partition,
755 bitmap loops, bitmap processed,
756 VEC (int, heap) **other_stores)
762 struct graph_edge *e;
764 EXECUTE_IF_SET_IN_BITMAP (partition, 0, i, ii)
765 if (RDG_MEM_WRITE_STMT (rdg, i)
766 || RDG_MEM_READS_STMT (rdg, i))
768 for (j = 0; j < rdg->n_vertices; j++)
769 if (!bitmap_bit_p (processed, j)
770 && (RDG_MEM_WRITE_STMT (rdg, j)
771 || RDG_MEM_READS_STMT (rdg, j))
772 && rdg_has_similar_memory_accesses (rdg, i, j))
774 /* Flag first the node J itself, and all the nodes that
775 are needed to compute J. */
776 rdg_flag_vertex_and_dependent (rdg, j, partition, loops,
779 /* When J is a read, we want to coalesce in the same
780 PARTITION all the nodes that are using J: this is
781 needed for better cache locality. */
782 rdg_flag_all_uses (rdg, j, partition, loops, processed, &foo);
784 /* Remove from OTHER_STORES the vertex that we flagged. */
785 if (RDG_MEM_WRITE_STMT (rdg, j))
786 for (k = 0; VEC_iterate (int, *other_stores, k, kk); k++)
789 VEC_unordered_remove (int, *other_stores, k);
794 /* If the node I has two uses, then keep these together in the
796 for (n = 0, e = rdg->vertices[i].succ; e; e = e->succ_next, n++);
799 rdg_flag_all_uses (rdg, i, partition, loops, processed, &foo);
803 /* Returns a bitmap in which all the statements needed for computing
804 the strongly connected component C of the RDG are flagged, also
805 including the loop exit conditions. */
808 build_rdg_partition_for_component (struct graph *rdg, rdgc c,
809 bool *part_has_writes,
810 VEC (int, heap) **other_stores)
813 bitmap partition = BITMAP_ALLOC (NULL);
814 bitmap loops = BITMAP_ALLOC (NULL);
815 bitmap processed = BITMAP_ALLOC (NULL);
817 for (i = 0; VEC_iterate (int, c->vertices, i, v); i++)
818 if (!already_processed_vertex_p (processed, v))
819 rdg_flag_vertex_and_dependent (rdg, v, partition, loops, processed,
822 /* Also iterate on the array of stores not in the starting vertices,
823 and determine those vertices that have some memory affinity with
824 the current nodes in the component: these are stores to the same
825 arrays, i.e. we're taking care of cache locality. */
826 rdg_flag_similar_memory_accesses (rdg, partition, loops, processed,
829 rdg_flag_loop_exits (rdg, loops, partition, processed, part_has_writes);
831 BITMAP_FREE (processed);
836 /* Free memory for COMPONENTS. */
839 free_rdg_components (VEC (rdgc, heap) *components)
844 for (i = 0; VEC_iterate (rdgc, components, i, x); i++)
846 VEC_free (int, heap, x->vertices);
851 /* Build the COMPONENTS vector with the strongly connected components
852 of RDG in which the STARTING_VERTICES occur. */
855 rdg_build_components (struct graph *rdg, VEC (int, heap) *starting_vertices,
856 VEC (rdgc, heap) **components)
859 bitmap saved_components = BITMAP_ALLOC (NULL);
860 int n_components = graphds_scc (rdg, NULL);
861 VEC (int, heap) **all_components = XNEWVEC (VEC (int, heap) *, n_components);
863 for (i = 0; i < n_components; i++)
864 all_components[i] = VEC_alloc (int, heap, 3);
866 for (i = 0; i < rdg->n_vertices; i++)
867 VEC_safe_push (int, heap, all_components[rdg->vertices[i].component], i);
869 for (i = 0; VEC_iterate (int, starting_vertices, i, v); i++)
871 int c = rdg->vertices[v].component;
873 if (!bitmap_bit_p (saved_components, c))
875 rdgc x = XCNEW (struct rdg_component);
877 x->vertices = all_components[c];
879 VEC_safe_push (rdgc, heap, *components, x);
880 bitmap_set_bit (saved_components, c);
884 for (i = 0; i < n_components; i++)
885 if (!bitmap_bit_p (saved_components, i))
886 VEC_free (int, heap, all_components[i]);
888 free (all_components);
889 BITMAP_FREE (saved_components);
892 /* Aggregate several components into a useful partition that is
893 registered in the PARTITIONS vector. Partitions will be
894 distributed in different loops. */
897 rdg_build_partitions (struct graph *rdg, VEC (rdgc, heap) *components,
898 VEC (int, heap) **other_stores,
899 VEC (bitmap, heap) **partitions, bitmap processed)
903 bitmap partition = BITMAP_ALLOC (NULL);
905 for (i = 0; VEC_iterate (rdgc, components, i, x); i++)
908 bool part_has_writes = false;
909 int v = VEC_index (int, x->vertices, 0);
911 if (bitmap_bit_p (processed, v))
914 np = build_rdg_partition_for_component (rdg, x, &part_has_writes,
916 bitmap_ior_into (partition, np);
917 bitmap_ior_into (processed, np);
922 if (dump_file && (dump_flags & TDF_DETAILS))
924 fprintf (dump_file, "ldist useful partition:\n");
925 dump_bitmap (dump_file, partition);
928 VEC_safe_push (bitmap, heap, *partitions, partition);
929 partition = BITMAP_ALLOC (NULL);
933 /* Add the nodes from the RDG that were not marked as processed, and
934 that are used outside the current loop. These are scalar
935 computations that are not yet part of previous partitions. */
936 for (i = 0; i < rdg->n_vertices; i++)
937 if (!bitmap_bit_p (processed, i)
938 && rdg_defs_used_in_other_loops_p (rdg, i))
939 VEC_safe_push (int, heap, *other_stores, i);
941 /* If there are still statements left in the OTHER_STORES array,
942 create other components and partitions with these stores and
943 their dependences. */
944 if (VEC_length (int, *other_stores) > 0)
946 VEC (rdgc, heap) *comps = VEC_alloc (rdgc, heap, 3);
947 VEC (int, heap) *foo = VEC_alloc (int, heap, 3);
949 rdg_build_components (rdg, *other_stores, &comps);
950 rdg_build_partitions (rdg, comps, &foo, partitions, processed);
952 VEC_free (int, heap, foo);
953 free_rdg_components (comps);
956 /* If there is something left in the last partition, save it. */
957 if (bitmap_count_bits (partition) > 0)
958 VEC_safe_push (bitmap, heap, *partitions, partition);
960 BITMAP_FREE (partition);
963 /* Dump to FILE the PARTITIONS. */
966 dump_rdg_partitions (FILE *file, VEC (bitmap, heap) *partitions)
971 for (i = 0; VEC_iterate (bitmap, partitions, i, partition); i++)
972 debug_bitmap_file (file, partition);
975 /* Debug PARTITIONS. */
976 extern void debug_rdg_partitions (VEC (bitmap, heap) *);
979 debug_rdg_partitions (VEC (bitmap, heap) *partitions)
981 dump_rdg_partitions (stderr, partitions);
984 /* Returns the number of read and write operations in the RDG. */
987 number_of_rw_in_rdg (struct graph *rdg)
991 for (i = 0; i < rdg->n_vertices; i++)
993 if (RDG_MEM_WRITE_STMT (rdg, i))
996 if (RDG_MEM_READS_STMT (rdg, i))
1003 /* Returns the number of read and write operations in a PARTITION of
1007 number_of_rw_in_partition (struct graph *rdg, bitmap partition)
1013 EXECUTE_IF_SET_IN_BITMAP (partition, 0, i, ii)
1015 if (RDG_MEM_WRITE_STMT (rdg, i))
1018 if (RDG_MEM_READS_STMT (rdg, i))
1025 /* Returns true when one of the PARTITIONS contains all the read or
1026 write operations of RDG. */
1029 partition_contains_all_rw (struct graph *rdg, VEC (bitmap, heap) *partitions)
1033 int nrw = number_of_rw_in_rdg (rdg);
1035 for (i = 0; VEC_iterate (bitmap, partitions, i, partition); i++)
1036 if (nrw == number_of_rw_in_partition (rdg, partition))
1042 /* Generate code from STARTING_VERTICES in RDG. Returns the number of
1043 distributed loops. */
1046 ldist_gen (struct loop *loop, struct graph *rdg,
1047 VEC (int, heap) *starting_vertices)
1050 VEC (rdgc, heap) *components = VEC_alloc (rdgc, heap, 3);
1051 VEC (bitmap, heap) *partitions = VEC_alloc (bitmap, heap, 3);
1052 VEC (int, heap) *other_stores = VEC_alloc (int, heap, 3);
1053 bitmap partition, processed = BITMAP_ALLOC (NULL);
1055 remaining_stmts = BITMAP_ALLOC (NULL);
1056 upstream_mem_writes = BITMAP_ALLOC (NULL);
1058 for (i = 0; i < rdg->n_vertices; i++)
1060 bitmap_set_bit (remaining_stmts, i);
1062 /* Save in OTHER_STORES all the memory writes that are not in
1063 STARTING_VERTICES. */
1064 if (RDG_MEM_WRITE_STMT (rdg, i))
1070 for (j = 0; VEC_iterate (int, starting_vertices, j, v); j++)
1078 VEC_safe_push (int, heap, other_stores, i);
1082 mark_nodes_having_upstream_mem_writes (rdg);
1083 rdg_build_components (rdg, starting_vertices, &components);
1084 rdg_build_partitions (rdg, components, &other_stores, &partitions,
1086 BITMAP_FREE (processed);
1087 nbp = VEC_length (bitmap, partitions);
1090 || partition_contains_all_rw (rdg, partitions))
1093 if (dump_file && (dump_flags & TDF_DETAILS))
1094 dump_rdg_partitions (dump_file, partitions);
1096 for (i = 0; VEC_iterate (bitmap, partitions, i, partition); i++)
1097 if (!generate_code_for_partition (loop, partition, i < nbp - 1))
1100 rewrite_into_loop_closed_ssa (NULL, TODO_update_ssa);
1101 update_ssa (TODO_update_ssa_only_virtuals | TODO_update_ssa);
1105 BITMAP_FREE (remaining_stmts);
1106 BITMAP_FREE (upstream_mem_writes);
1108 for (i = 0; VEC_iterate (bitmap, partitions, i, partition); i++)
1109 BITMAP_FREE (partition);
1111 VEC_free (int, heap, other_stores);
1112 VEC_free (bitmap, heap, partitions);
1113 free_rdg_components (components);
1117 /* Distributes the code from LOOP in such a way that producer
1118 statements are placed before consumer statements. When STMTS is
1119 NULL, performs the maximal distribution, if STMTS is not NULL,
1120 tries to separate only these statements from the LOOP's body.
1121 Returns the number of distributed loops. */
1124 distribute_loop (struct loop *loop, VEC (gimple, heap) *stmts)
1130 VEC (int, heap) *vertices;
1132 if (loop->num_nodes > 2)
1134 if (dump_file && (dump_flags & TDF_DETAILS))
1136 "FIXME: Loop %d not distributed: it has more than two basic blocks.\n",
1142 rdg = build_rdg (loop);
1146 if (dump_file && (dump_flags & TDF_DETAILS))
1148 "FIXME: Loop %d not distributed: failed to build the RDG.\n",
1154 vertices = VEC_alloc (int, heap, 3);
1156 if (dump_file && (dump_flags & TDF_DETAILS))
1157 dump_rdg (dump_file, rdg);
1159 for (i = 0; VEC_iterate (gimple, stmts, i, s); i++)
1161 int v = rdg_vertex_for_stmt (rdg, s);
1165 VEC_safe_push (int, heap, vertices, v);
1167 if (dump_file && (dump_flags & TDF_DETAILS))
1169 "ldist asked to generate code for vertex %d\n", v);
1173 res = ldist_gen (loop, rdg, vertices);
1174 VEC_free (int, heap, vertices);
1180 /* Distribute all loops in the current function. */
1183 tree_loop_distribution (void)
1187 int nb_generated_loops = 0;
1189 FOR_EACH_LOOP (li, loop, 0)
1191 VEC (gimple, heap) *work_list = VEC_alloc (gimple, heap, 3);
1193 /* With the following working list, we're asking distribute_loop
1194 to separate the stores of the loop: when dependences allow,
1195 it will end on having one store per loop. */
1196 stores_from_loop (loop, &work_list);
1198 /* A simple heuristic for cache locality is to not split stores
1199 to the same array. Without this call, an unrolled loop would
1200 be split into as many loops as unroll factor, each loop
1201 storing in the same array. */
1202 remove_similar_memory_refs (&work_list);
1204 nb_generated_loops = distribute_loop (loop, work_list);
1206 if (dump_file && (dump_flags & TDF_DETAILS))
1208 if (nb_generated_loops > 1)
1209 fprintf (dump_file, "Loop %d distributed: split to %d loops.\n",
1210 loop->num, nb_generated_loops);
1212 fprintf (dump_file, "Loop %d is the same.\n", loop->num);
1215 verify_loop_structure ();
1217 VEC_free (gimple, heap, work_list);
1224 gate_tree_loop_distribution (void)
1226 return flag_tree_loop_distribution != 0;
1229 struct gimple_opt_pass pass_loop_distribution =
1234 gate_tree_loop_distribution, /* gate */
1235 tree_loop_distribution, /* execute */
1238 0, /* static_pass_number */
1239 TV_TREE_LOOP_DISTRIBUTION, /* tv_id */
1240 PROP_cfg | PROP_ssa, /* properties_required */
1241 0, /* properties_provided */
1242 0, /* properties_destroyed */
1243 0, /* todo_flags_start */
1244 TODO_dump_func | TODO_verify_loops /* todo_flags_finish */