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 res = dr_analyze_innermost (dr);
259 gcc_assert (res && stride_of_unit_type_p (DR_STEP (dr), TREE_TYPE (op0)));
261 nb_bytes = build_size_arg_loc (loc, nb_iter, op0, &stmt_list);
262 addr_base = size_binop_loc (loc, PLUS_EXPR, DR_OFFSET (dr), DR_INIT (dr));
263 addr_base = fold_convert_loc (loc, sizetype, addr_base);
265 /* Test for a negative stride, iterating over every element. */
266 if (integer_zerop (size_binop (PLUS_EXPR,
267 TYPE_SIZE_UNIT (TREE_TYPE (op0)),
268 fold_convert (sizetype, DR_STEP (dr)))))
270 addr_base = size_binop_loc (loc, MINUS_EXPR, addr_base,
271 fold_convert_loc (loc, sizetype, nb_bytes));
272 addr_base = size_binop_loc (loc, PLUS_EXPR, addr_base,
273 TYPE_SIZE_UNIT (TREE_TYPE (op0)));
276 addr_base = fold_build2_loc (loc, POINTER_PLUS_EXPR,
277 TREE_TYPE (DR_BASE_ADDRESS (dr)),
278 DR_BASE_ADDRESS (dr), addr_base);
279 mem = force_gimple_operand (addr_base, &stmts, true, NULL);
280 gimple_seq_add_seq (&stmt_list, stmts);
282 fn = build_fold_addr_expr (implicit_built_in_decls [BUILT_IN_MEMSET]);
283 fn_call = gimple_build_call (fn, 3, mem, integer_zero_node, nb_bytes);
284 gimple_seq_add_stmt (&stmt_list, fn_call);
285 gsi_insert_seq_after (&bsi, stmt_list, GSI_CONTINUE_LINKING);
287 if (dump_file && (dump_flags & TDF_DETAILS))
288 fprintf (dump_file, "generated memset zero\n");
293 /* Tries to generate a builtin function for the instructions of LOOP
294 pointed to by the bits set in PARTITION. Returns true when the
295 operation succeeded. */
298 generate_builtin (struct loop *loop, bitmap partition, bool copy_p)
304 gimple_stmt_iterator bsi;
305 tree nb_iter = number_of_exit_cond_executions (loop);
307 if (!nb_iter || nb_iter == chrec_dont_know)
310 bbs = get_loop_body_in_dom_order (loop);
312 for (i = 0; i < loop->num_nodes; i++)
314 basic_block bb = bbs[i];
316 for (bsi = gsi_start_phis (bb); !gsi_end_p (bsi); gsi_next (&bsi))
319 for (bsi = gsi_start_bb (bb); !gsi_end_p (bsi); gsi_next (&bsi))
321 gimple stmt = gsi_stmt (bsi);
323 if (bitmap_bit_p (partition, x++)
324 && is_gimple_assign (stmt)
325 && !is_gimple_reg (gimple_assign_lhs (stmt)))
327 /* Don't generate the builtins when there are more than
333 if (bb == loop->latch)
334 nb_iter = number_of_latch_executions (loop);
339 if (!stmt_with_adjacent_zero_store_dr_p (write))
342 /* The new statements will be placed before LOOP. */
343 bsi = gsi_last_bb (loop_preheader_edge (loop)->src);
344 generate_memset_zero (write, gimple_assign_lhs (write), nb_iter, bsi);
347 /* If this is the last partition for which we generate code, we have
348 to destroy the loop. */
351 unsigned nbbs = loop->num_nodes;
352 edge exit = single_exit (loop);
353 basic_block src = loop_preheader_edge (loop)->src, dest = exit->dest;
354 redirect_edge_pred (exit, src);
355 exit->flags &= ~(EDGE_TRUE_VALUE|EDGE_FALSE_VALUE);
356 exit->flags |= EDGE_FALLTHRU;
357 cancel_loop_tree (loop);
358 rescan_loop_exit (exit, false, true);
360 for (i = 0; i < nbbs; i++)
361 delete_basic_block (bbs[i]);
363 set_immediate_dominator (CDI_DOMINATORS, dest,
364 recompute_dominator (CDI_DOMINATORS, dest));
372 /* Generates code for PARTITION. For simple loops, this function can
373 generate a built-in. */
376 generate_code_for_partition (struct loop *loop, bitmap partition, bool copy_p)
378 if (generate_builtin (loop, partition, copy_p))
381 return generate_loops_for_partition (loop, partition, copy_p);
385 /* Returns true if the node V of RDG cannot be recomputed. */
388 rdg_cannot_recompute_vertex_p (struct graph *rdg, int v)
390 if (RDG_MEM_WRITE_STMT (rdg, v))
396 /* Returns true when the vertex V has already been generated in the
397 current partition (V is in PROCESSED), or when V belongs to another
398 partition and cannot be recomputed (V is not in REMAINING_STMTS). */
401 already_processed_vertex_p (bitmap processed, int v)
403 return (bitmap_bit_p (processed, v)
404 || !bitmap_bit_p (remaining_stmts, v));
407 /* Returns NULL when there is no anti-dependence among the successors
408 of vertex V, otherwise returns the edge with the anti-dep. */
410 static struct graph_edge *
411 has_anti_dependence (struct vertex *v)
413 struct graph_edge *e;
416 for (e = v->succ; e; e = e->succ_next)
417 if (RDGE_TYPE (e) == anti_dd)
423 /* Returns true when V has an anti-dependence edge among its successors. */
426 predecessor_has_mem_write (struct graph *rdg, struct vertex *v)
428 struct graph_edge *e;
431 for (e = v->pred; e; e = e->pred_next)
432 if (bitmap_bit_p (upstream_mem_writes, e->src)
433 /* Don't consider flow channels: a write to memory followed
434 by a read from memory. These channels allow the split of
435 the RDG in different partitions. */
436 && !RDG_MEM_WRITE_STMT (rdg, e->src))
442 /* Initializes the upstream_mem_writes bitmap following the
443 information from RDG. */
446 mark_nodes_having_upstream_mem_writes (struct graph *rdg)
449 bitmap seen = BITMAP_ALLOC (NULL);
451 for (v = rdg->n_vertices - 1; v >= 0; v--)
452 if (!bitmap_bit_p (seen, v))
455 VEC (int, heap) *nodes = VEC_alloc (int, heap, 3);
457 graphds_dfs (rdg, &v, 1, &nodes, false, NULL);
459 FOR_EACH_VEC_ELT (int, nodes, i, x)
461 if (!bitmap_set_bit (seen, x))
464 if (RDG_MEM_WRITE_STMT (rdg, x)
465 || predecessor_has_mem_write (rdg, &(rdg->vertices[x]))
466 /* In anti dependences the read should occur before
467 the write, this is why both the read and the write
468 should be placed in the same partition. */
469 || has_anti_dependence (&(rdg->vertices[x])))
471 bitmap_set_bit (upstream_mem_writes, x);
475 VEC_free (int, heap, nodes);
479 /* Returns true when vertex u has a memory write node as a predecessor
483 has_upstream_mem_writes (int u)
485 return bitmap_bit_p (upstream_mem_writes, u);
488 static void rdg_flag_vertex_and_dependent (struct graph *, int, bitmap, bitmap,
491 /* Flag the uses of U stopping following the information from
492 upstream_mem_writes. */
495 rdg_flag_uses (struct graph *rdg, int u, bitmap partition, bitmap loops,
496 bitmap processed, bool *part_has_writes)
499 struct vertex *x = &(rdg->vertices[u]);
500 gimple stmt = RDGV_STMT (x);
501 struct graph_edge *anti_dep = has_anti_dependence (x);
503 /* Keep in the same partition the destination of an antidependence,
504 because this is a store to the exact same location. Putting this
505 in another partition is bad for cache locality. */
508 int v = anti_dep->dest;
510 if (!already_processed_vertex_p (processed, v))
511 rdg_flag_vertex_and_dependent (rdg, v, partition, loops,
512 processed, part_has_writes);
515 if (gimple_code (stmt) != GIMPLE_PHI)
517 if ((use_p = gimple_vuse_op (stmt)) != NULL_USE_OPERAND_P)
519 tree use = USE_FROM_PTR (use_p);
521 if (TREE_CODE (use) == SSA_NAME)
523 gimple def_stmt = SSA_NAME_DEF_STMT (use);
524 int v = rdg_vertex_for_stmt (rdg, def_stmt);
527 && !already_processed_vertex_p (processed, v))
528 rdg_flag_vertex_and_dependent (rdg, v, partition, loops,
529 processed, part_has_writes);
534 if (is_gimple_assign (stmt) && has_upstream_mem_writes (u))
536 tree op0 = gimple_assign_lhs (stmt);
538 /* Scalar channels don't have enough space for transmitting data
539 between tasks, unless we add more storage by privatizing. */
540 if (is_gimple_reg (op0))
543 imm_use_iterator iter;
545 FOR_EACH_IMM_USE_FAST (use_p, iter, op0)
547 int v = rdg_vertex_for_stmt (rdg, USE_STMT (use_p));
549 if (!already_processed_vertex_p (processed, v))
550 rdg_flag_vertex_and_dependent (rdg, v, partition, loops,
551 processed, part_has_writes);
557 /* Flag V from RDG as part of PARTITION, and also flag its loop number
561 rdg_flag_vertex (struct graph *rdg, int v, bitmap partition, bitmap loops,
562 bool *part_has_writes)
566 if (!bitmap_set_bit (partition, v))
569 loop = loop_containing_stmt (RDG_STMT (rdg, v));
570 bitmap_set_bit (loops, loop->num);
572 if (rdg_cannot_recompute_vertex_p (rdg, v))
574 *part_has_writes = true;
575 bitmap_clear_bit (remaining_stmts, v);
579 /* Flag in the bitmap PARTITION the vertex V and all its predecessors.
580 Also flag their loop number in LOOPS. */
583 rdg_flag_vertex_and_dependent (struct graph *rdg, int v, bitmap partition,
584 bitmap loops, bitmap processed,
585 bool *part_has_writes)
588 VEC (int, heap) *nodes = VEC_alloc (int, heap, 3);
591 bitmap_set_bit (processed, v);
592 rdg_flag_uses (rdg, v, partition, loops, processed, part_has_writes);
593 graphds_dfs (rdg, &v, 1, &nodes, false, remaining_stmts);
594 rdg_flag_vertex (rdg, v, partition, loops, part_has_writes);
596 FOR_EACH_VEC_ELT (int, nodes, i, x)
597 if (!already_processed_vertex_p (processed, x))
598 rdg_flag_vertex_and_dependent (rdg, x, partition, loops, processed,
601 VEC_free (int, heap, nodes);
604 /* Initialize CONDS with all the condition statements from the basic
608 collect_condition_stmts (struct loop *loop, VEC (gimple, heap) **conds)
612 VEC (edge, heap) *exits = get_loop_exit_edges (loop);
614 FOR_EACH_VEC_ELT (edge, exits, i, e)
616 gimple cond = last_stmt (e->src);
619 VEC_safe_push (gimple, heap, *conds, cond);
622 VEC_free (edge, heap, exits);
625 /* Add to PARTITION all the exit condition statements for LOOPS
626 together with all their dependent statements determined from
630 rdg_flag_loop_exits (struct graph *rdg, bitmap loops, bitmap partition,
631 bitmap processed, bool *part_has_writes)
635 VEC (gimple, heap) *conds = VEC_alloc (gimple, heap, 3);
637 EXECUTE_IF_SET_IN_BITMAP (loops, 0, i, bi)
638 collect_condition_stmts (get_loop (i), &conds);
640 while (!VEC_empty (gimple, conds))
642 gimple cond = VEC_pop (gimple, conds);
643 int v = rdg_vertex_for_stmt (rdg, cond);
644 bitmap new_loops = BITMAP_ALLOC (NULL);
646 if (!already_processed_vertex_p (processed, v))
647 rdg_flag_vertex_and_dependent (rdg, v, partition, new_loops, processed,
650 EXECUTE_IF_SET_IN_BITMAP (new_loops, 0, i, bi)
651 if (bitmap_set_bit (loops, i))
652 collect_condition_stmts (get_loop (i), &conds);
654 BITMAP_FREE (new_loops);
658 /* Returns a bitmap in which all the statements needed for computing
659 the strongly connected component C of the RDG are flagged, also
660 including the loop exit conditions. */
663 build_rdg_partition_for_component (struct graph *rdg, rdgc c,
664 bool *part_has_writes)
667 bitmap partition = BITMAP_ALLOC (NULL);
668 bitmap loops = BITMAP_ALLOC (NULL);
669 bitmap processed = BITMAP_ALLOC (NULL);
671 FOR_EACH_VEC_ELT (int, c->vertices, i, v)
672 if (!already_processed_vertex_p (processed, v))
673 rdg_flag_vertex_and_dependent (rdg, v, partition, loops, processed,
676 rdg_flag_loop_exits (rdg, loops, partition, processed, part_has_writes);
678 BITMAP_FREE (processed);
683 /* Free memory for COMPONENTS. */
686 free_rdg_components (VEC (rdgc, heap) *components)
691 FOR_EACH_VEC_ELT (rdgc, components, i, x)
693 VEC_free (int, heap, x->vertices);
698 /* Build the COMPONENTS vector with the strongly connected components
699 of RDG in which the STARTING_VERTICES occur. */
702 rdg_build_components (struct graph *rdg, VEC (int, heap) *starting_vertices,
703 VEC (rdgc, heap) **components)
706 bitmap saved_components = BITMAP_ALLOC (NULL);
707 int n_components = graphds_scc (rdg, NULL);
708 VEC (int, heap) **all_components = XNEWVEC (VEC (int, heap) *, n_components);
710 for (i = 0; i < n_components; i++)
711 all_components[i] = VEC_alloc (int, heap, 3);
713 for (i = 0; i < rdg->n_vertices; i++)
714 VEC_safe_push (int, heap, all_components[rdg->vertices[i].component], i);
716 FOR_EACH_VEC_ELT (int, starting_vertices, i, v)
718 int c = rdg->vertices[v].component;
720 if (bitmap_set_bit (saved_components, c))
722 rdgc x = XCNEW (struct rdg_component);
724 x->vertices = all_components[c];
726 VEC_safe_push (rdgc, heap, *components, x);
730 for (i = 0; i < n_components; i++)
731 if (!bitmap_bit_p (saved_components, i))
732 VEC_free (int, heap, all_components[i]);
734 free (all_components);
735 BITMAP_FREE (saved_components);
738 /* Returns true when it is possible to generate a builtin pattern for
739 the PARTITION of RDG. For the moment we detect only the memset
743 can_generate_builtin (struct graph *rdg, bitmap partition)
751 EXECUTE_IF_SET_IN_BITMAP (partition, 0, i, bi)
752 if (RDG_MEM_READS_STMT (rdg, i))
754 else if (RDG_MEM_WRITE_STMT (rdg, i))
757 if (stmt_with_adjacent_zero_store_dr_p (RDG_STMT (rdg, i)))
761 return stores_zero == 1 && nb_writes == 1 && nb_reads == 0;
764 /* Returns true when PARTITION1 and PARTITION2 have similar memory
768 similar_memory_accesses (struct graph *rdg, bitmap partition1,
772 bitmap_iterator bi, bj;
774 EXECUTE_IF_SET_IN_BITMAP (partition1, 0, i, bi)
775 if (RDG_MEM_WRITE_STMT (rdg, i)
776 || RDG_MEM_READS_STMT (rdg, i))
777 EXECUTE_IF_SET_IN_BITMAP (partition2, 0, j, bj)
778 if (RDG_MEM_WRITE_STMT (rdg, j)
779 || RDG_MEM_READS_STMT (rdg, j))
780 if (rdg_has_similar_memory_accesses (rdg, i, j))
786 /* Fuse all the partitions from PARTITIONS that contain similar memory
787 references, i.e., we're taking care of cache locality. This
788 function does not fuse those partitions that contain patterns that
789 can be code generated with builtins. */
792 fuse_partitions_with_similar_memory_accesses (struct graph *rdg,
793 VEC (bitmap, heap) **partitions)
796 bitmap partition1, partition2;
798 FOR_EACH_VEC_ELT (bitmap, *partitions, p1, partition1)
799 if (!can_generate_builtin (rdg, partition1))
800 FOR_EACH_VEC_ELT (bitmap, *partitions, p2, partition2)
802 && !can_generate_builtin (rdg, partition2)
803 && similar_memory_accesses (rdg, partition1, partition2))
805 bitmap_ior_into (partition1, partition2);
806 VEC_ordered_remove (bitmap, *partitions, p2);
811 /* Returns true when DEF is an SSA_NAME defined in LOOP and used after
815 ssa_name_has_uses_outside_loop_p (tree def, loop_p loop)
817 imm_use_iterator imm_iter;
820 FOR_EACH_IMM_USE_FAST (use_p, imm_iter, def)
821 if (loop != loop_containing_stmt (USE_STMT (use_p)))
827 /* Returns true when STMT defines a scalar variable used after the
831 stmt_has_scalar_dependences_outside_loop (gimple stmt)
835 switch (gimple_code (stmt))
838 name = gimple_assign_lhs (stmt);
842 name = gimple_phi_result (stmt);
849 return TREE_CODE (name) == SSA_NAME
850 && ssa_name_has_uses_outside_loop_p (name, loop_containing_stmt (stmt));
853 /* Returns true when STMT will be code generated in a partition of RDG
854 different than PART and that will not be code generated as a
858 stmt_generated_in_another_partition (struct graph *rdg, gimple stmt, int part,
859 VEC (bitmap, heap) *partitions)
866 FOR_EACH_VEC_ELT (bitmap, partitions, p, pp)
868 && !can_generate_builtin (rdg, pp))
869 EXECUTE_IF_SET_IN_BITMAP (pp, 0, i, bi)
870 if (stmt == RDG_STMT (rdg, i))
876 /* For each partition in PARTITIONS that will be code generated using
877 a builtin, add its scalar computations used after the loop to
881 add_scalar_computations_to_partition (struct graph *rdg,
882 VEC (bitmap, heap) *partitions,
889 bitmap l = BITMAP_ALLOC (NULL);
890 bitmap pr = BITMAP_ALLOC (NULL);
893 FOR_EACH_VEC_ELT (bitmap, partitions, p, pp)
894 if (can_generate_builtin (rdg, pp))
895 EXECUTE_IF_SET_IN_BITMAP (pp, 0, i, bi)
896 if (stmt_has_scalar_dependences_outside_loop (RDG_STMT (rdg, i))
897 && !stmt_generated_in_another_partition (rdg, RDG_STMT (rdg, i), p,
899 rdg_flag_vertex_and_dependent (rdg, i, partition, l, pr, &f);
901 rdg_flag_loop_exits (rdg, l, partition, pr, &f);
907 /* Aggregate several components into a useful partition that is
908 registered in the PARTITIONS vector. Partitions will be
909 distributed in different loops. */
912 rdg_build_partitions (struct graph *rdg, VEC (rdgc, heap) *components,
913 VEC (int, heap) **other_stores,
914 VEC (bitmap, heap) **partitions, bitmap processed)
918 bitmap partition = BITMAP_ALLOC (NULL);
920 FOR_EACH_VEC_ELT (rdgc, components, i, x)
923 bool part_has_writes = false;
924 int v = VEC_index (int, x->vertices, 0);
926 if (bitmap_bit_p (processed, v))
929 np = build_rdg_partition_for_component (rdg, x, &part_has_writes);
930 bitmap_ior_into (partition, np);
931 bitmap_ior_into (processed, np);
936 if (dump_file && (dump_flags & TDF_DETAILS))
938 fprintf (dump_file, "ldist useful partition:\n");
939 dump_bitmap (dump_file, partition);
942 VEC_safe_push (bitmap, heap, *partitions, partition);
943 partition = BITMAP_ALLOC (NULL);
947 /* Add the nodes from the RDG that were not marked as processed, and
948 that are used outside the current loop. These are scalar
949 computations that are not yet part of previous partitions. */
950 for (i = 0; i < rdg->n_vertices; i++)
951 if (!bitmap_bit_p (processed, i)
952 && rdg_defs_used_in_other_loops_p (rdg, i))
953 VEC_safe_push (int, heap, *other_stores, i);
955 /* If there are still statements left in the OTHER_STORES array,
956 create other components and partitions with these stores and
957 their dependences. */
958 if (VEC_length (int, *other_stores) > 0)
960 VEC (rdgc, heap) *comps = VEC_alloc (rdgc, heap, 3);
961 VEC (int, heap) *foo = VEC_alloc (int, heap, 3);
963 rdg_build_components (rdg, *other_stores, &comps);
964 rdg_build_partitions (rdg, comps, &foo, partitions, processed);
966 VEC_free (int, heap, foo);
967 free_rdg_components (comps);
970 add_scalar_computations_to_partition (rdg, *partitions, partition);
972 /* If there is something left in the last partition, save it. */
973 if (bitmap_count_bits (partition) > 0)
974 VEC_safe_push (bitmap, heap, *partitions, partition);
976 BITMAP_FREE (partition);
978 fuse_partitions_with_similar_memory_accesses (rdg, partitions);
981 /* Dump to FILE the PARTITIONS. */
984 dump_rdg_partitions (FILE *file, VEC (bitmap, heap) *partitions)
989 FOR_EACH_VEC_ELT (bitmap, partitions, i, partition)
990 debug_bitmap_file (file, partition);
993 /* Debug PARTITIONS. */
994 extern void debug_rdg_partitions (VEC (bitmap, heap) *);
997 debug_rdg_partitions (VEC (bitmap, heap) *partitions)
999 dump_rdg_partitions (stderr, partitions);
1002 /* Returns the number of read and write operations in the RDG. */
1005 number_of_rw_in_rdg (struct graph *rdg)
1009 for (i = 0; i < rdg->n_vertices; i++)
1011 if (RDG_MEM_WRITE_STMT (rdg, i))
1014 if (RDG_MEM_READS_STMT (rdg, i))
1021 /* Returns the number of read and write operations in a PARTITION of
1025 number_of_rw_in_partition (struct graph *rdg, bitmap partition)
1031 EXECUTE_IF_SET_IN_BITMAP (partition, 0, i, ii)
1033 if (RDG_MEM_WRITE_STMT (rdg, i))
1036 if (RDG_MEM_READS_STMT (rdg, i))
1043 /* Returns true when one of the PARTITIONS contains all the read or
1044 write operations of RDG. */
1047 partition_contains_all_rw (struct graph *rdg, VEC (bitmap, heap) *partitions)
1051 int nrw = number_of_rw_in_rdg (rdg);
1053 FOR_EACH_VEC_ELT (bitmap, partitions, i, partition)
1054 if (nrw == number_of_rw_in_partition (rdg, partition))
1060 /* Generate code from STARTING_VERTICES in RDG. Returns the number of
1061 distributed loops. */
1064 ldist_gen (struct loop *loop, struct graph *rdg,
1065 VEC (int, heap) *starting_vertices)
1068 VEC (rdgc, heap) *components = VEC_alloc (rdgc, heap, 3);
1069 VEC (bitmap, heap) *partitions = VEC_alloc (bitmap, heap, 3);
1070 VEC (int, heap) *other_stores = VEC_alloc (int, heap, 3);
1071 bitmap partition, processed = BITMAP_ALLOC (NULL);
1073 remaining_stmts = BITMAP_ALLOC (NULL);
1074 upstream_mem_writes = BITMAP_ALLOC (NULL);
1076 for (i = 0; i < rdg->n_vertices; i++)
1078 bitmap_set_bit (remaining_stmts, i);
1080 /* Save in OTHER_STORES all the memory writes that are not in
1081 STARTING_VERTICES. */
1082 if (RDG_MEM_WRITE_STMT (rdg, i))
1088 FOR_EACH_VEC_ELT (int, starting_vertices, j, v)
1096 VEC_safe_push (int, heap, other_stores, i);
1100 mark_nodes_having_upstream_mem_writes (rdg);
1101 rdg_build_components (rdg, starting_vertices, &components);
1102 rdg_build_partitions (rdg, components, &other_stores, &partitions,
1104 BITMAP_FREE (processed);
1105 nbp = VEC_length (bitmap, partitions);
1108 || partition_contains_all_rw (rdg, partitions))
1111 if (dump_file && (dump_flags & TDF_DETAILS))
1112 dump_rdg_partitions (dump_file, partitions);
1114 FOR_EACH_VEC_ELT (bitmap, partitions, i, partition)
1115 if (!generate_code_for_partition (loop, partition, i < nbp - 1))
1118 rewrite_into_loop_closed_ssa (NULL, TODO_update_ssa);
1119 update_ssa (TODO_update_ssa_only_virtuals | TODO_update_ssa);
1123 BITMAP_FREE (remaining_stmts);
1124 BITMAP_FREE (upstream_mem_writes);
1126 FOR_EACH_VEC_ELT (bitmap, partitions, i, partition)
1127 BITMAP_FREE (partition);
1129 VEC_free (int, heap, other_stores);
1130 VEC_free (bitmap, heap, partitions);
1131 free_rdg_components (components);
1135 /* Distributes the code from LOOP in such a way that producer
1136 statements are placed before consumer statements. When STMTS is
1137 NULL, performs the maximal distribution, if STMTS is not NULL,
1138 tries to separate only these statements from the LOOP's body.
1139 Returns the number of distributed loops. */
1142 distribute_loop (struct loop *loop, VEC (gimple, heap) *stmts)
1148 VEC (int, heap) *vertices;
1150 if (loop->num_nodes > 2)
1152 if (dump_file && (dump_flags & TDF_DETAILS))
1154 "FIXME: Loop %d not distributed: it has more than two basic blocks.\n",
1160 rdg = build_rdg (loop);
1164 if (dump_file && (dump_flags & TDF_DETAILS))
1166 "FIXME: Loop %d not distributed: failed to build the RDG.\n",
1172 vertices = VEC_alloc (int, heap, 3);
1174 if (dump_file && (dump_flags & TDF_DETAILS))
1175 dump_rdg (dump_file, rdg);
1177 FOR_EACH_VEC_ELT (gimple, stmts, i, s)
1179 int v = rdg_vertex_for_stmt (rdg, s);
1183 VEC_safe_push (int, heap, vertices, v);
1185 if (dump_file && (dump_flags & TDF_DETAILS))
1187 "ldist asked to generate code for vertex %d\n", v);
1191 res = ldist_gen (loop, rdg, vertices);
1192 VEC_free (int, heap, vertices);
1198 /* Distribute all loops in the current function. */
1201 tree_loop_distribution (void)
1205 int nb_generated_loops = 0;
1207 FOR_EACH_LOOP (li, loop, 0)
1209 VEC (gimple, heap) *work_list = NULL;
1211 /* If the loop doesn't have a single exit we will fail anyway,
1212 so do that early. */
1213 if (!single_exit (loop))
1216 /* If both flag_tree_loop_distribute_patterns and
1217 flag_tree_loop_distribution are set, then only
1218 distribute_patterns is executed. */
1219 if (flag_tree_loop_distribute_patterns)
1221 /* With the following working list, we're asking
1222 distribute_loop to separate from the rest of the loop the
1223 stores of the form "A[i] = 0". */
1224 stores_zero_from_loop (loop, &work_list);
1226 /* Do nothing if there are no patterns to be distributed. */
1227 if (VEC_length (gimple, work_list) > 0)
1228 nb_generated_loops = distribute_loop (loop, work_list);
1230 else if (flag_tree_loop_distribution)
1232 /* With the following working list, we're asking
1233 distribute_loop to separate the stores of the loop: when
1234 dependences allow, it will end on having one store per
1236 stores_from_loop (loop, &work_list);
1238 /* A simple heuristic for cache locality is to not split
1239 stores to the same array. Without this call, an unrolled
1240 loop would be split into as many loops as unroll factor,
1241 each loop storing in the same array. */
1242 remove_similar_memory_refs (&work_list);
1244 nb_generated_loops = distribute_loop (loop, work_list);
1247 if (dump_file && (dump_flags & TDF_DETAILS))
1249 if (nb_generated_loops > 1)
1250 fprintf (dump_file, "Loop %d distributed: split to %d loops.\n",
1251 loop->num, nb_generated_loops);
1253 fprintf (dump_file, "Loop %d is the same.\n", loop->num);
1256 verify_loop_structure ();
1258 VEC_free (gimple, heap, work_list);
1265 gate_tree_loop_distribution (void)
1267 return flag_tree_loop_distribution
1268 || flag_tree_loop_distribute_patterns;
1271 struct gimple_opt_pass pass_loop_distribution =
1276 gate_tree_loop_distribution, /* gate */
1277 tree_loop_distribution, /* execute */
1280 0, /* static_pass_number */
1281 TV_TREE_LOOP_DISTRIBUTION, /* tv_id */
1282 PROP_cfg | PROP_ssa, /* properties_required */
1283 0, /* properties_provided */
1284 0, /* properties_destroyed */
1285 0, /* todo_flags_start */
1286 TODO_dump_func /* todo_flags_finish */