2 Copyright (C) 2006, 2007, 2008 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 size argument. */
222 build_size_arg (tree nb_iter, tree op, gimple_seq* stmt_list)
225 gimple_seq stmts = NULL;
227 nb_bytes = fold_build2 (MULT_EXPR, TREE_TYPE (nb_iter),
228 nb_iter, TYPE_SIZE_UNIT (TREE_TYPE (op)));
229 nb_bytes = force_gimple_operand (nb_bytes, &stmts, true, NULL);
230 gimple_seq_add_seq (stmt_list, stmts);
235 /* Generate a call to memset. Return true when the operation succeeded. */
238 generate_memset_zero (gimple stmt, tree op0, tree nb_iter,
239 gimple_stmt_iterator bsi)
242 tree nb_bytes = NULL;
244 gimple_seq stmts = NULL, stmt_list = NULL;
246 tree mem, fndecl, fntype, fn;
247 gimple_stmt_iterator i;
249 struct data_reference *dr = XCNEW (struct data_reference);
253 if (!dr_analyze_innermost (dr))
256 /* Test for a positive stride, iterating over every element. */
257 if (integer_zerop (fold_build2 (MINUS_EXPR, integer_type_node, DR_STEP (dr),
258 TYPE_SIZE_UNIT (TREE_TYPE (op0)))))
260 tree offset = fold_convert (sizetype,
261 size_binop (PLUS_EXPR,
264 addr_base = fold_build2 (POINTER_PLUS_EXPR,
265 TREE_TYPE (DR_BASE_ADDRESS (dr)),
266 DR_BASE_ADDRESS (dr), offset);
269 /* Test for a negative stride, iterating over every element. */
270 else if (integer_zerop (fold_build2 (PLUS_EXPR, integer_type_node,
271 TYPE_SIZE_UNIT (TREE_TYPE (op0)),
274 nb_bytes = build_size_arg (nb_iter, op0, &stmt_list);
275 addr_base = size_binop (PLUS_EXPR, DR_OFFSET (dr), DR_INIT (dr));
276 addr_base = fold_build2 (MINUS_EXPR, sizetype, addr_base, nb_bytes);
277 addr_base = force_gimple_operand (addr_base, &stmts, true, NULL);
278 gimple_seq_add_seq (&stmt_list, stmts);
280 addr_base = fold_build2 (POINTER_PLUS_EXPR,
281 TREE_TYPE (DR_BASE_ADDRESS (dr)),
282 DR_BASE_ADDRESS (dr), addr_base);
287 mem = force_gimple_operand (addr_base, &stmts, true, NULL);
288 gimple_seq_add_seq (&stmt_list, stmts);
290 fndecl = implicit_built_in_decls [BUILT_IN_MEMSET];
291 fntype = TREE_TYPE (fndecl);
292 fn = build1 (ADDR_EXPR, build_pointer_type (fntype), fndecl);
295 nb_bytes = build_size_arg (nb_iter, op0, &stmt_list);
296 fn_call = gimple_build_call (fn, 3, mem, integer_zero_node, nb_bytes);
297 gimple_seq_add_stmt (&stmt_list, fn_call);
299 for (i = gsi_start (stmt_list); !gsi_end_p (i); gsi_next (&i))
301 gimple s = gsi_stmt (i);
302 update_stmt_if_modified (s);
304 FOR_EACH_SSA_TREE_OPERAND (t, s, iter, SSA_OP_VIRTUAL_DEFS)
306 if (TREE_CODE (t) == SSA_NAME)
307 t = SSA_NAME_VAR (t);
308 mark_sym_for_renaming (t);
312 /* Mark also the uses of the VDEFS of STMT to be renamed. */
313 FOR_EACH_SSA_TREE_OPERAND (t, stmt, iter, SSA_OP_VIRTUAL_DEFS)
315 if (TREE_CODE (t) == SSA_NAME)
318 imm_use_iterator imm_iter;
320 FOR_EACH_IMM_USE_STMT (s, imm_iter, t)
323 t = SSA_NAME_VAR (t);
325 mark_sym_for_renaming (t);
328 gsi_insert_seq_after (&bsi, stmt_list, GSI_CONTINUE_LINKING);
331 if (dump_file && (dump_flags & TDF_DETAILS))
332 fprintf (dump_file, "generated memset zero\n");
339 /* Propagate phis in BB b to their uses and remove them. */
342 prop_phis (basic_block b)
344 gimple_stmt_iterator psi;
345 gimple_seq phis = phi_nodes (b);
347 for (psi = gsi_start (phis); !gsi_end_p (psi); )
349 gimple phi = gsi_stmt (psi);
350 tree def = gimple_phi_result (phi), use = gimple_phi_arg_def (phi, 0);
352 gcc_assert (gimple_phi_num_args (phi) == 1);
354 if (!is_gimple_reg (def))
356 imm_use_iterator iter;
360 FOR_EACH_IMM_USE_STMT (stmt, iter, def)
361 FOR_EACH_IMM_USE_ON_STMT (use_p, iter)
362 SET_USE (use_p, use);
365 replace_uses_by (def, use);
367 remove_phi_node (&psi, true);
371 /* Tries to generate a builtin function for the instructions of LOOP
372 pointed to by the bits set in PARTITION. Returns true when the
373 operation succeeded. */
376 generate_builtin (struct loop *loop, bitmap partition, bool copy_p)
383 gimple_stmt_iterator bsi;
384 tree nb_iter = number_of_exit_cond_executions (loop);
386 if (!nb_iter || nb_iter == chrec_dont_know)
389 bbs = get_loop_body_in_dom_order (loop);
391 for (i = 0; i < loop->num_nodes; i++)
393 basic_block bb = bbs[i];
395 for (bsi = gsi_start_phis (bb); !gsi_end_p (bsi); gsi_next (&bsi))
398 for (bsi = gsi_start_bb (bb); !gsi_end_p (bsi); gsi_next (&bsi))
400 gimple stmt = gsi_stmt (bsi);
402 if (bitmap_bit_p (partition, x++)
403 && is_gimple_assign (stmt)
404 && !is_gimple_reg (gimple_assign_lhs (stmt)))
406 /* Don't generate the builtins when there are more than
419 op0 = gimple_assign_lhs (write);
420 op1 = gimple_assign_rhs1 (write);
422 if (!(TREE_CODE (op0) == ARRAY_REF
423 || TREE_CODE (op0) == INDIRECT_REF))
426 /* The new statements will be placed before LOOP. */
427 bsi = gsi_last_bb (loop_preheader_edge (loop)->src);
429 if (gimple_assign_rhs_code (write) == INTEGER_CST
430 && (integer_zerop (op1) || real_zerop (op1)))
431 res = generate_memset_zero (write, op0, nb_iter, bsi);
433 /* If this is the last partition for which we generate code, we have
434 to destroy the loop. */
437 unsigned nbbs = loop->num_nodes;
438 basic_block src = loop_preheader_edge (loop)->src;
439 basic_block dest = single_exit (loop)->dest;
441 make_edge (src, dest, EDGE_FALLTHRU);
442 cancel_loop_tree (loop);
444 for (i = 0; i < nbbs; i++)
445 delete_basic_block (bbs[i]);
447 set_immediate_dominator (CDI_DOMINATORS, dest,
448 recompute_dominator (CDI_DOMINATORS, dest));
456 /* Generates code for PARTITION. For simple loops, this function can
457 generate a built-in. */
460 generate_code_for_partition (struct loop *loop, bitmap partition, bool copy_p)
462 if (generate_builtin (loop, partition, copy_p))
465 return generate_loops_for_partition (loop, partition, copy_p);
469 /* Returns true if the node V of RDG cannot be recomputed. */
472 rdg_cannot_recompute_vertex_p (struct graph *rdg, int v)
474 if (RDG_MEM_WRITE_STMT (rdg, v))
480 /* Returns true when the vertex V has already been generated in the
481 current partition (V is in PROCESSED), or when V belongs to another
482 partition and cannot be recomputed (V is not in REMAINING_STMTS). */
485 already_processed_vertex_p (bitmap processed, int v)
487 return (bitmap_bit_p (processed, v)
488 || !bitmap_bit_p (remaining_stmts, v));
491 /* Returns NULL when there is no anti-dependence among the successors
492 of vertex V, otherwise returns the edge with the anti-dep. */
494 static struct graph_edge *
495 has_anti_dependence (struct vertex *v)
497 struct graph_edge *e;
500 for (e = v->succ; e; e = e->succ_next)
501 if (RDGE_TYPE (e) == anti_dd)
507 /* Returns true when V has an anti-dependence edge among its successors. */
510 predecessor_has_mem_write (struct graph *rdg, struct vertex *v)
512 struct graph_edge *e;
515 for (e = v->pred; e; e = e->pred_next)
516 if (bitmap_bit_p (upstream_mem_writes, e->src)
517 /* Don't consider flow channels: a write to memory followed
518 by a read from memory. These channels allow the split of
519 the RDG in different partitions. */
520 && !RDG_MEM_WRITE_STMT (rdg, e->src))
526 /* Initializes the upstream_mem_writes bitmap following the
527 information from RDG. */
530 mark_nodes_having_upstream_mem_writes (struct graph *rdg)
533 bitmap seen = BITMAP_ALLOC (NULL);
535 for (v = rdg->n_vertices - 1; v >= 0; v--)
536 if (!bitmap_bit_p (seen, v))
539 VEC (int, heap) *nodes = VEC_alloc (int, heap, 3);
540 bool has_upstream_mem_write_p = false;
542 graphds_dfs (rdg, &v, 1, &nodes, false, NULL);
544 for (i = 0; VEC_iterate (int, nodes, i, x); i++)
546 if (bitmap_bit_p (seen, x))
549 bitmap_set_bit (seen, x);
551 if (RDG_MEM_WRITE_STMT (rdg, x)
552 || predecessor_has_mem_write (rdg, &(rdg->vertices[x]))
553 /* In anti dependences the read should occur before
554 the write, this is why both the read and the write
555 should be placed in the same partition. */
556 || has_anti_dependence (&(rdg->vertices[x])))
558 has_upstream_mem_write_p = true;
559 bitmap_set_bit (upstream_mem_writes, x);
563 VEC_free (int, heap, nodes);
567 /* Returns true when vertex u has a memory write node as a predecessor
571 has_upstream_mem_writes (int u)
573 return bitmap_bit_p (upstream_mem_writes, u);
576 static void rdg_flag_vertex_and_dependent (struct graph *, int, bitmap, bitmap,
579 /* Flag all the uses of U. */
582 rdg_flag_all_uses (struct graph *rdg, int u, bitmap partition, bitmap loops,
583 bitmap processed, bool *part_has_writes)
585 struct graph_edge *e;
587 for (e = rdg->vertices[u].succ; e; e = e->succ_next)
588 if (!bitmap_bit_p (processed, e->dest))
590 rdg_flag_vertex_and_dependent (rdg, e->dest, partition, loops,
591 processed, part_has_writes);
592 rdg_flag_all_uses (rdg, e->dest, partition, loops, processed,
597 /* Flag the uses of U stopping following the information from
598 upstream_mem_writes. */
601 rdg_flag_uses (struct graph *rdg, int u, bitmap partition, bitmap loops,
602 bitmap processed, bool *part_has_writes)
606 struct vertex *x = &(rdg->vertices[u]);
607 gimple stmt = RDGV_STMT (x);
608 struct graph_edge *anti_dep = has_anti_dependence (x);
610 /* Keep in the same partition the destination of an antidependence,
611 because this is a store to the exact same location. Putting this
612 in another partition is bad for cache locality. */
615 int v = anti_dep->dest;
617 if (!already_processed_vertex_p (processed, v))
618 rdg_flag_vertex_and_dependent (rdg, v, partition, loops,
619 processed, part_has_writes);
622 if (gimple_code (stmt) != GIMPLE_PHI)
624 FOR_EACH_SSA_USE_OPERAND (use_p, stmt, iter, SSA_OP_VIRTUAL_USES)
626 tree use = USE_FROM_PTR (use_p);
628 if (TREE_CODE (use) == SSA_NAME)
630 gimple def_stmt = SSA_NAME_DEF_STMT (use);
631 int v = rdg_vertex_for_stmt (rdg, def_stmt);
634 && !already_processed_vertex_p (processed, v))
635 rdg_flag_vertex_and_dependent (rdg, v, partition, loops,
636 processed, part_has_writes);
641 if (is_gimple_assign (stmt) && has_upstream_mem_writes (u))
643 tree op0 = gimple_assign_lhs (stmt);
645 /* Scalar channels don't have enough space for transmitting data
646 between tasks, unless we add more storage by privatizing. */
647 if (is_gimple_reg (op0))
650 imm_use_iterator iter;
652 FOR_EACH_IMM_USE_FAST (use_p, iter, op0)
654 int v = rdg_vertex_for_stmt (rdg, USE_STMT (use_p));
656 if (!already_processed_vertex_p (processed, v))
657 rdg_flag_vertex_and_dependent (rdg, v, partition, loops,
658 processed, part_has_writes);
664 /* Flag V from RDG as part of PARTITION, and also flag its loop number
668 rdg_flag_vertex (struct graph *rdg, int v, bitmap partition, bitmap loops,
669 bool *part_has_writes)
673 if (bitmap_bit_p (partition, v))
676 loop = loop_containing_stmt (RDG_STMT (rdg, v));
677 bitmap_set_bit (loops, loop->num);
678 bitmap_set_bit (partition, v);
680 if (rdg_cannot_recompute_vertex_p (rdg, v))
682 *part_has_writes = true;
683 bitmap_clear_bit (remaining_stmts, v);
687 /* Flag in the bitmap PARTITION the vertex V and all its predecessors.
688 Also flag their loop number in LOOPS. */
691 rdg_flag_vertex_and_dependent (struct graph *rdg, int v, bitmap partition,
692 bitmap loops, bitmap processed,
693 bool *part_has_writes)
696 VEC (int, heap) *nodes = VEC_alloc (int, heap, 3);
699 bitmap_set_bit (processed, v);
700 rdg_flag_uses (rdg, v, partition, loops, processed, part_has_writes);
701 graphds_dfs (rdg, &v, 1, &nodes, false, remaining_stmts);
702 rdg_flag_vertex (rdg, v, partition, loops, part_has_writes);
704 for (i = 0; VEC_iterate (int, nodes, i, x); i++)
705 if (!already_processed_vertex_p (processed, x))
706 rdg_flag_vertex_and_dependent (rdg, x, partition, loops, processed,
709 VEC_free (int, heap, nodes);
712 /* Initialize CONDS with all the condition statements from the basic
716 collect_condition_stmts (struct loop *loop, VEC (gimple, heap) **conds)
720 VEC (edge, heap) *exits = get_loop_exit_edges (loop);
722 for (i = 0; VEC_iterate (edge, exits, i, e); i++)
724 gimple cond = last_stmt (e->src);
727 VEC_safe_push (gimple, heap, *conds, cond);
730 VEC_free (edge, heap, exits);
733 /* Add to PARTITION all the exit condition statements for LOOPS
734 together with all their dependent statements determined from
738 rdg_flag_loop_exits (struct graph *rdg, bitmap loops, bitmap partition,
739 bitmap processed, bool *part_has_writes)
743 VEC (gimple, heap) *conds = VEC_alloc (gimple, heap, 3);
745 EXECUTE_IF_SET_IN_BITMAP (loops, 0, i, bi)
746 collect_condition_stmts (get_loop (i), &conds);
748 while (!VEC_empty (gimple, conds))
750 gimple cond = VEC_pop (gimple, conds);
751 int v = rdg_vertex_for_stmt (rdg, cond);
752 bitmap new_loops = BITMAP_ALLOC (NULL);
754 if (!already_processed_vertex_p (processed, v))
755 rdg_flag_vertex_and_dependent (rdg, v, partition, new_loops, processed,
758 EXECUTE_IF_SET_IN_BITMAP (new_loops, 0, i, bi)
759 if (!bitmap_bit_p (loops, i))
761 bitmap_set_bit (loops, i);
762 collect_condition_stmts (get_loop (i), &conds);
765 BITMAP_FREE (new_loops);
769 /* Flag all the nodes of RDG containing memory accesses that could
770 potentially belong to arrays already accessed in the current
774 rdg_flag_similar_memory_accesses (struct graph *rdg, bitmap partition,
775 bitmap loops, bitmap processed,
776 VEC (int, heap) **other_stores)
782 struct graph_edge *e;
784 EXECUTE_IF_SET_IN_BITMAP (partition, 0, i, ii)
785 if (RDG_MEM_WRITE_STMT (rdg, i)
786 || RDG_MEM_READS_STMT (rdg, i))
788 for (j = 0; j < rdg->n_vertices; j++)
789 if (!bitmap_bit_p (processed, j)
790 && (RDG_MEM_WRITE_STMT (rdg, j)
791 || RDG_MEM_READS_STMT (rdg, j))
792 && rdg_has_similar_memory_accesses (rdg, i, j))
794 /* Flag first the node J itself, and all the nodes that
795 are needed to compute J. */
796 rdg_flag_vertex_and_dependent (rdg, j, partition, loops,
799 /* When J is a read, we want to coalesce in the same
800 PARTITION all the nodes that are using J: this is
801 needed for better cache locality. */
802 rdg_flag_all_uses (rdg, j, partition, loops, processed, &foo);
804 /* Remove from OTHER_STORES the vertex that we flagged. */
805 if (RDG_MEM_WRITE_STMT (rdg, j))
806 for (k = 0; VEC_iterate (int, *other_stores, k, kk); k++)
809 VEC_unordered_remove (int, *other_stores, k);
814 /* If the node I has two uses, then keep these together in the
816 for (n = 0, e = rdg->vertices[i].succ; e; e = e->succ_next, n++);
819 rdg_flag_all_uses (rdg, i, partition, loops, processed, &foo);
823 /* Returns a bitmap in which all the statements needed for computing
824 the strongly connected component C of the RDG are flagged, also
825 including the loop exit conditions. */
828 build_rdg_partition_for_component (struct graph *rdg, rdgc c,
829 bool *part_has_writes,
830 VEC (int, heap) **other_stores)
833 bitmap partition = BITMAP_ALLOC (NULL);
834 bitmap loops = BITMAP_ALLOC (NULL);
835 bitmap processed = BITMAP_ALLOC (NULL);
837 for (i = 0; VEC_iterate (int, c->vertices, i, v); i++)
838 if (!already_processed_vertex_p (processed, v))
839 rdg_flag_vertex_and_dependent (rdg, v, partition, loops, processed,
842 /* Also iterate on the array of stores not in the starting vertices,
843 and determine those vertices that have some memory affinity with
844 the current nodes in the component: these are stores to the same
845 arrays, i.e. we're taking care of cache locality. */
846 rdg_flag_similar_memory_accesses (rdg, partition, loops, processed,
849 rdg_flag_loop_exits (rdg, loops, partition, processed, part_has_writes);
851 BITMAP_FREE (processed);
856 /* Free memory for COMPONENTS. */
859 free_rdg_components (VEC (rdgc, heap) *components)
864 for (i = 0; VEC_iterate (rdgc, components, i, x); i++)
866 VEC_free (int, heap, x->vertices);
871 /* Build the COMPONENTS vector with the strongly connected components
872 of RDG in which the STARTING_VERTICES occur. */
875 rdg_build_components (struct graph *rdg, VEC (int, heap) *starting_vertices,
876 VEC (rdgc, heap) **components)
879 bitmap saved_components = BITMAP_ALLOC (NULL);
880 int n_components = graphds_scc (rdg, NULL);
881 VEC (int, heap) **all_components = XNEWVEC (VEC (int, heap) *, n_components);
883 for (i = 0; i < n_components; i++)
884 all_components[i] = VEC_alloc (int, heap, 3);
886 for (i = 0; i < rdg->n_vertices; i++)
887 VEC_safe_push (int, heap, all_components[rdg->vertices[i].component], i);
889 for (i = 0; VEC_iterate (int, starting_vertices, i, v); i++)
891 int c = rdg->vertices[v].component;
893 if (!bitmap_bit_p (saved_components, c))
895 rdgc x = XCNEW (struct rdg_component);
897 x->vertices = all_components[c];
899 VEC_safe_push (rdgc, heap, *components, x);
900 bitmap_set_bit (saved_components, c);
904 for (i = 0; i < n_components; i++)
905 if (!bitmap_bit_p (saved_components, i))
906 VEC_free (int, heap, all_components[i]);
908 free (all_components);
909 BITMAP_FREE (saved_components);
912 /* Aggregate several components into a useful partition that is
913 registered in the PARTITIONS vector. Partitions will be
914 distributed in different loops. */
917 rdg_build_partitions (struct graph *rdg, VEC (rdgc, heap) *components,
918 VEC (int, heap) **other_stores,
919 VEC (bitmap, heap) **partitions, bitmap processed)
923 bitmap partition = BITMAP_ALLOC (NULL);
925 for (i = 0; VEC_iterate (rdgc, components, i, x); i++)
928 bool part_has_writes = false;
929 int v = VEC_index (int, x->vertices, 0);
931 if (bitmap_bit_p (processed, v))
934 np = build_rdg_partition_for_component (rdg, x, &part_has_writes,
936 bitmap_ior_into (partition, np);
937 bitmap_ior_into (processed, np);
942 if (dump_file && (dump_flags & TDF_DETAILS))
944 fprintf (dump_file, "ldist useful partition:\n");
945 dump_bitmap (dump_file, partition);
948 VEC_safe_push (bitmap, heap, *partitions, partition);
949 partition = BITMAP_ALLOC (NULL);
953 /* Add the nodes from the RDG that were not marked as processed, and
954 that are used outside the current loop. These are scalar
955 computations that are not yet part of previous partitions. */
956 for (i = 0; i < rdg->n_vertices; i++)
957 if (!bitmap_bit_p (processed, i)
958 && rdg_defs_used_in_other_loops_p (rdg, i))
959 VEC_safe_push (int, heap, *other_stores, i);
961 /* If there are still statements left in the OTHER_STORES array,
962 create other components and partitions with these stores and
963 their dependences. */
964 if (VEC_length (int, *other_stores) > 0)
966 VEC (rdgc, heap) *comps = VEC_alloc (rdgc, heap, 3);
967 VEC (int, heap) *foo = VEC_alloc (int, heap, 3);
969 rdg_build_components (rdg, *other_stores, &comps);
970 rdg_build_partitions (rdg, comps, &foo, partitions, processed);
972 VEC_free (int, heap, foo);
973 free_rdg_components (comps);
976 /* If there is something left in the last partition, save it. */
977 if (bitmap_count_bits (partition) > 0)
978 VEC_safe_push (bitmap, heap, *partitions, partition);
980 BITMAP_FREE (partition);
983 /* Dump to FILE the PARTITIONS. */
986 dump_rdg_partitions (FILE *file, VEC (bitmap, heap) *partitions)
991 for (i = 0; VEC_iterate (bitmap, partitions, i, partition); i++)
992 debug_bitmap_file (file, partition);
995 /* Debug PARTITIONS. */
996 extern void debug_rdg_partitions (VEC (bitmap, heap) *);
999 debug_rdg_partitions (VEC (bitmap, heap) *partitions)
1001 dump_rdg_partitions (stderr, partitions);
1004 /* Returns the number of read and write operations in the RDG. */
1007 number_of_rw_in_rdg (struct graph *rdg)
1011 for (i = 0; i < rdg->n_vertices; i++)
1013 if (RDG_MEM_WRITE_STMT (rdg, i))
1016 if (RDG_MEM_READS_STMT (rdg, i))
1023 /* Returns the number of read and write operations in a PARTITION of
1027 number_of_rw_in_partition (struct graph *rdg, bitmap partition)
1033 EXECUTE_IF_SET_IN_BITMAP (partition, 0, i, ii)
1035 if (RDG_MEM_WRITE_STMT (rdg, i))
1038 if (RDG_MEM_READS_STMT (rdg, i))
1045 /* Returns true when one of the PARTITIONS contains all the read or
1046 write operations of RDG. */
1049 partition_contains_all_rw (struct graph *rdg, VEC (bitmap, heap) *partitions)
1053 int nrw = number_of_rw_in_rdg (rdg);
1055 for (i = 0; VEC_iterate (bitmap, partitions, i, partition); i++)
1056 if (nrw == number_of_rw_in_partition (rdg, partition))
1062 /* Generate code from STARTING_VERTICES in RDG. Returns the number of
1063 distributed loops. */
1066 ldist_gen (struct loop *loop, struct graph *rdg,
1067 VEC (int, heap) *starting_vertices)
1070 VEC (rdgc, heap) *components = VEC_alloc (rdgc, heap, 3);
1071 VEC (bitmap, heap) *partitions = VEC_alloc (bitmap, heap, 3);
1072 VEC (int, heap) *other_stores = VEC_alloc (int, heap, 3);
1073 bitmap partition, processed = BITMAP_ALLOC (NULL);
1075 remaining_stmts = BITMAP_ALLOC (NULL);
1076 upstream_mem_writes = BITMAP_ALLOC (NULL);
1078 for (i = 0; i < rdg->n_vertices; i++)
1080 bitmap_set_bit (remaining_stmts, i);
1082 /* Save in OTHER_STORES all the memory writes that are not in
1083 STARTING_VERTICES. */
1084 if (RDG_MEM_WRITE_STMT (rdg, i))
1090 for (j = 0; VEC_iterate (int, starting_vertices, j, v); j++)
1098 VEC_safe_push (int, heap, other_stores, i);
1102 mark_nodes_having_upstream_mem_writes (rdg);
1103 rdg_build_components (rdg, starting_vertices, &components);
1104 rdg_build_partitions (rdg, components, &other_stores, &partitions,
1106 BITMAP_FREE (processed);
1107 nbp = VEC_length (bitmap, partitions);
1110 || partition_contains_all_rw (rdg, partitions))
1113 if (dump_file && (dump_flags & TDF_DETAILS))
1114 dump_rdg_partitions (dump_file, partitions);
1116 for (i = 0; VEC_iterate (bitmap, partitions, i, partition); i++)
1117 if (!generate_code_for_partition (loop, partition, i < nbp - 1))
1120 rewrite_into_loop_closed_ssa (NULL, TODO_update_ssa);
1121 update_ssa (TODO_update_ssa_only_virtuals | TODO_update_ssa);
1125 BITMAP_FREE (remaining_stmts);
1126 BITMAP_FREE (upstream_mem_writes);
1128 for (i = 0; VEC_iterate (bitmap, partitions, i, partition); i++)
1129 BITMAP_FREE (partition);
1131 VEC_free (int, heap, other_stores);
1132 VEC_free (bitmap, heap, partitions);
1133 free_rdg_components (components);
1137 /* Distributes the code from LOOP in such a way that producer
1138 statements are placed before consumer statements. When STMTS is
1139 NULL, performs the maximal distribution, if STMTS is not NULL,
1140 tries to separate only these statements from the LOOP's body.
1141 Returns the number of distributed loops. */
1144 distribute_loop (struct loop *loop, VEC (gimple, heap) *stmts)
1150 VEC (int, heap) *vertices;
1152 if (loop->num_nodes > 2)
1154 if (dump_file && (dump_flags & TDF_DETAILS))
1156 "FIXME: Loop %d not distributed: it has more than two basic blocks.\n",
1162 rdg = build_rdg (loop);
1166 if (dump_file && (dump_flags & TDF_DETAILS))
1168 "FIXME: Loop %d not distributed: failed to build the RDG.\n",
1174 vertices = VEC_alloc (int, heap, 3);
1176 if (dump_file && (dump_flags & TDF_DETAILS))
1177 dump_rdg (dump_file, rdg);
1179 for (i = 0; VEC_iterate (gimple, stmts, i, s); i++)
1181 int v = rdg_vertex_for_stmt (rdg, s);
1185 VEC_safe_push (int, heap, vertices, v);
1187 if (dump_file && (dump_flags & TDF_DETAILS))
1189 "ldist asked to generate code for vertex %d\n", v);
1193 res = ldist_gen (loop, rdg, vertices);
1194 VEC_free (int, heap, vertices);
1200 /* Distribute all loops in the current function. */
1203 tree_loop_distribution (void)
1207 int nb_generated_loops = 0;
1209 FOR_EACH_LOOP (li, loop, 0)
1211 VEC (gimple, heap) *work_list = VEC_alloc (gimple, heap, 3);
1213 /* With the following working list, we're asking distribute_loop
1214 to separate the stores of the loop: when dependences allow,
1215 it will end on having one store per loop. */
1216 stores_from_loop (loop, &work_list);
1218 /* A simple heuristic for cache locality is to not split stores
1219 to the same array. Without this call, an unrolled loop would
1220 be split into as many loops as unroll factor, each loop
1221 storing in the same array. */
1222 remove_similar_memory_refs (&work_list);
1224 nb_generated_loops = distribute_loop (loop, work_list);
1226 if (dump_file && (dump_flags & TDF_DETAILS))
1228 if (nb_generated_loops > 1)
1229 fprintf (dump_file, "Loop %d distributed: split to %d loops.\n",
1230 loop->num, nb_generated_loops);
1232 fprintf (dump_file, "Loop %d is the same.\n", loop->num);
1235 verify_loop_structure ();
1237 VEC_free (gimple, heap, work_list);
1244 gate_tree_loop_distribution (void)
1246 return flag_tree_loop_distribution != 0;
1249 struct gimple_opt_pass pass_loop_distribution =
1254 gate_tree_loop_distribution, /* gate */
1255 tree_loop_distribution, /* execute */
1258 0, /* static_pass_number */
1259 TV_TREE_LOOP_DISTRIBUTION, /* tv_id */
1260 PROP_cfg | PROP_ssa, /* properties_required */
1261 0, /* properties_provided */
1262 0, /* properties_destroyed */
1263 0, /* todo_flags_start */
1264 TODO_dump_func | TODO_verify_loops /* todo_flags_finish */