2 Copyright (C) 2006, 2007, 2008, 2009, 2010, 2011
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"
48 #include "tree-flow.h"
50 #include "tree-chrec.h"
51 #include "tree-data-ref.h"
52 #include "tree-scalar-evolution.h"
53 #include "tree-pass.h"
55 /* If bit I is not set, it means that this node represents an
56 operation that has already been performed, and that should not be
57 performed again. This is the subgraph of remaining important
58 computations that is passed to the DFS algorithm for avoiding to
59 include several times the same stores in different loops. */
60 static bitmap remaining_stmts;
62 /* A node of the RDG is marked in this bitmap when it has as a
63 predecessor a node that writes to memory. */
64 static bitmap upstream_mem_writes;
66 /* Update the PHI nodes of NEW_LOOP. NEW_LOOP is a duplicate of
70 update_phis_for_loop_copy (struct loop *orig_loop, struct loop *new_loop)
73 gimple_stmt_iterator si_new, si_orig;
74 edge orig_loop_latch = loop_latch_edge (orig_loop);
75 edge orig_entry_e = loop_preheader_edge (orig_loop);
76 edge new_loop_entry_e = loop_preheader_edge (new_loop);
78 /* Scan the phis in the headers of the old and new loops
79 (they are organized in exactly the same order). */
80 for (si_new = gsi_start_phis (new_loop->header),
81 si_orig = gsi_start_phis (orig_loop->header);
82 !gsi_end_p (si_new) && !gsi_end_p (si_orig);
83 gsi_next (&si_new), gsi_next (&si_orig))
86 source_location locus;
87 gimple phi_new = gsi_stmt (si_new);
88 gimple phi_orig = gsi_stmt (si_orig);
90 /* Add the first phi argument for the phi in NEW_LOOP (the one
91 associated with the entry of NEW_LOOP) */
92 def = PHI_ARG_DEF_FROM_EDGE (phi_orig, orig_entry_e);
93 locus = gimple_phi_arg_location_from_edge (phi_orig, orig_entry_e);
94 add_phi_arg (phi_new, def, new_loop_entry_e, locus);
96 /* Add the second phi argument for the phi in NEW_LOOP (the one
97 associated with the latch of NEW_LOOP) */
98 def = PHI_ARG_DEF_FROM_EDGE (phi_orig, orig_loop_latch);
99 locus = gimple_phi_arg_location_from_edge (phi_orig, orig_loop_latch);
101 if (TREE_CODE (def) == SSA_NAME)
103 new_ssa_name = get_current_def (def);
106 /* This only happens if there are no definitions inside the
107 loop. Use the the invariant in the new loop as is. */
111 /* Could be an integer. */
114 add_phi_arg (phi_new, new_ssa_name, loop_latch_edge (new_loop), locus);
118 /* Return a copy of LOOP placed before LOOP. */
121 copy_loop_before (struct loop *loop)
124 edge preheader = loop_preheader_edge (loop);
126 if (!single_exit (loop))
129 initialize_original_copy_tables ();
130 res = slpeel_tree_duplicate_loop_to_edge_cfg (loop, preheader);
131 free_original_copy_tables ();
136 update_phis_for_loop_copy (loop, res);
137 rename_variables_in_loop (res);
142 /* Creates an empty basic block after LOOP. */
145 create_bb_after_loop (struct loop *loop)
147 edge exit = single_exit (loop);
155 /* Generate code for PARTITION from the code in LOOP. The loop is
156 copied when COPY_P is true. All the statements not flagged in the
157 PARTITION bitmap are removed from the loop or from its copy. The
158 statements are indexed in sequence inside a basic block, and the
159 basic blocks of a loop are taken in dom order. Returns true when
160 the code gen succeeded. */
163 generate_loops_for_partition (struct loop *loop, bitmap partition, bool copy_p)
166 gimple_stmt_iterator bsi;
171 loop = copy_loop_before (loop);
172 create_preheader (loop, CP_SIMPLE_PREHEADERS);
173 create_bb_after_loop (loop);
179 /* Remove stmts not in the PARTITION bitmap. The order in which we
180 visit the phi nodes and the statements is exactly as in
182 bbs = get_loop_body_in_dom_order (loop);
184 if (MAY_HAVE_DEBUG_STMTS)
185 for (x = 0, i = 0; i < loop->num_nodes; i++)
187 basic_block bb = bbs[i];
189 for (bsi = gsi_start_phis (bb); !gsi_end_p (bsi); gsi_next (&bsi))
190 if (!bitmap_bit_p (partition, x++))
191 reset_debug_uses (gsi_stmt (bsi));
193 for (bsi = gsi_start_bb (bb); !gsi_end_p (bsi); gsi_next (&bsi))
195 gimple stmt = gsi_stmt (bsi);
196 if (gimple_code (stmt) != GIMPLE_LABEL
197 && !is_gimple_debug (stmt)
198 && !bitmap_bit_p (partition, x++))
199 reset_debug_uses (stmt);
203 for (x = 0, i = 0; i < loop->num_nodes; i++)
205 basic_block bb = bbs[i];
207 for (bsi = gsi_start_phis (bb); !gsi_end_p (bsi);)
208 if (!bitmap_bit_p (partition, x++))
210 gimple phi = gsi_stmt (bsi);
211 if (!is_gimple_reg (gimple_phi_result (phi)))
212 mark_virtual_phi_result_for_renaming (phi);
213 remove_phi_node (&bsi, true);
218 for (bsi = gsi_start_bb (bb); !gsi_end_p (bsi);)
220 gimple stmt = gsi_stmt (bsi);
221 if (gimple_code (stmt) != GIMPLE_LABEL
222 && !is_gimple_debug (stmt)
223 && !bitmap_bit_p (partition, x++))
225 unlink_stmt_vdef (stmt);
226 gsi_remove (&bsi, true);
238 /* Build the size argument for a memset call. */
241 build_size_arg_loc (location_t loc, tree nb_iter, tree op,
242 gimple_seq *stmt_list)
245 tree x = size_binop_loc (loc, MULT_EXPR,
246 fold_convert_loc (loc, sizetype, nb_iter),
247 TYPE_SIZE_UNIT (TREE_TYPE (op)));
248 x = force_gimple_operand (x, &stmts, true, NULL);
249 gimple_seq_add_seq (stmt_list, stmts);
254 /* Generate a call to memset. Return true when the operation succeeded. */
257 generate_memset_zero (gimple stmt, tree op0, tree nb_iter,
258 gimple_stmt_iterator bsi)
260 tree addr_base, nb_bytes;
262 gimple_seq stmt_list = NULL, stmts;
265 struct data_reference *dr = XCNEW (struct data_reference);
266 location_t loc = gimple_location (stmt);
270 res = dr_analyze_innermost (dr);
271 gcc_assert (res && stride_of_unit_type_p (DR_STEP (dr), TREE_TYPE (op0)));
273 nb_bytes = build_size_arg_loc (loc, nb_iter, op0, &stmt_list);
274 addr_base = size_binop_loc (loc, PLUS_EXPR, DR_OFFSET (dr), DR_INIT (dr));
275 addr_base = fold_convert_loc (loc, sizetype, addr_base);
277 /* Test for a negative stride, iterating over every element. */
278 if (integer_zerop (size_binop (PLUS_EXPR,
279 TYPE_SIZE_UNIT (TREE_TYPE (op0)),
280 fold_convert (sizetype, DR_STEP (dr)))))
282 addr_base = size_binop_loc (loc, MINUS_EXPR, addr_base,
283 fold_convert_loc (loc, sizetype, nb_bytes));
284 addr_base = size_binop_loc (loc, PLUS_EXPR, addr_base,
285 TYPE_SIZE_UNIT (TREE_TYPE (op0)));
288 addr_base = fold_build_pointer_plus_loc (loc,
289 DR_BASE_ADDRESS (dr), addr_base);
290 mem = force_gimple_operand (addr_base, &stmts, true, NULL);
291 gimple_seq_add_seq (&stmt_list, stmts);
293 fn = build_fold_addr_expr (implicit_built_in_decls [BUILT_IN_MEMSET]);
294 fn_call = gimple_build_call (fn, 3, mem, integer_zero_node, nb_bytes);
295 gimple_seq_add_stmt (&stmt_list, fn_call);
296 gsi_insert_seq_after (&bsi, stmt_list, GSI_CONTINUE_LINKING);
298 if (dump_file && (dump_flags & TDF_DETAILS))
299 fprintf (dump_file, "generated memset zero\n");
304 /* Tries to generate a builtin function for the instructions of LOOP
305 pointed to by the bits set in PARTITION. Returns true when the
306 operation succeeded. */
309 generate_builtin (struct loop *loop, bitmap partition, bool copy_p)
315 gimple_stmt_iterator bsi;
316 tree nb_iter = number_of_exit_cond_executions (loop);
318 if (!nb_iter || nb_iter == chrec_dont_know)
321 bbs = get_loop_body_in_dom_order (loop);
323 for (i = 0; i < loop->num_nodes; i++)
325 basic_block bb = bbs[i];
327 for (bsi = gsi_start_phis (bb); !gsi_end_p (bsi); gsi_next (&bsi))
330 for (bsi = gsi_start_bb (bb); !gsi_end_p (bsi); gsi_next (&bsi))
332 gimple stmt = gsi_stmt (bsi);
334 if (gimple_code (stmt) != GIMPLE_LABEL
335 && !is_gimple_debug (stmt)
336 && bitmap_bit_p (partition, x++)
337 && is_gimple_assign (stmt)
338 && !is_gimple_reg (gimple_assign_lhs (stmt)))
340 /* Don't generate the builtins when there are more than
346 if (bb == loop->latch)
347 nb_iter = number_of_latch_executions (loop);
352 if (!stmt_with_adjacent_zero_store_dr_p (write))
355 /* The new statements will be placed before LOOP. */
356 bsi = gsi_last_bb (loop_preheader_edge (loop)->src);
357 generate_memset_zero (write, gimple_assign_lhs (write), nb_iter, bsi);
360 /* If this is the last partition for which we generate code, we have
361 to destroy the loop. */
364 unsigned nbbs = loop->num_nodes;
365 edge exit = single_exit (loop);
366 basic_block src = loop_preheader_edge (loop)->src, dest = exit->dest;
367 redirect_edge_pred (exit, src);
368 exit->flags &= ~(EDGE_TRUE_VALUE|EDGE_FALSE_VALUE);
369 exit->flags |= EDGE_FALLTHRU;
370 cancel_loop_tree (loop);
371 rescan_loop_exit (exit, false, true);
373 for (i = 0; i < nbbs; i++)
374 delete_basic_block (bbs[i]);
376 set_immediate_dominator (CDI_DOMINATORS, dest,
377 recompute_dominator (CDI_DOMINATORS, dest));
385 /* Generates code for PARTITION. For simple loops, this function can
386 generate a built-in. */
389 generate_code_for_partition (struct loop *loop, bitmap partition, bool copy_p)
391 if (generate_builtin (loop, partition, copy_p))
394 return generate_loops_for_partition (loop, partition, copy_p);
398 /* Returns true if the node V of RDG cannot be recomputed. */
401 rdg_cannot_recompute_vertex_p (struct graph *rdg, int v)
403 if (RDG_MEM_WRITE_STMT (rdg, v))
409 /* Returns true when the vertex V has already been generated in the
410 current partition (V is in PROCESSED), or when V belongs to another
411 partition and cannot be recomputed (V is not in REMAINING_STMTS). */
414 already_processed_vertex_p (bitmap processed, int v)
416 return (bitmap_bit_p (processed, v)
417 || !bitmap_bit_p (remaining_stmts, v));
420 /* Returns NULL when there is no anti-dependence among the successors
421 of vertex V, otherwise returns the edge with the anti-dep. */
423 static struct graph_edge *
424 has_anti_dependence (struct vertex *v)
426 struct graph_edge *e;
429 for (e = v->succ; e; e = e->succ_next)
430 if (RDGE_TYPE (e) == anti_dd)
436 /* Returns true when V has an anti-dependence edge among its successors. */
439 predecessor_has_mem_write (struct graph *rdg, struct vertex *v)
441 struct graph_edge *e;
444 for (e = v->pred; e; e = e->pred_next)
445 if (bitmap_bit_p (upstream_mem_writes, e->src)
446 /* Don't consider flow channels: a write to memory followed
447 by a read from memory. These channels allow the split of
448 the RDG in different partitions. */
449 && !RDG_MEM_WRITE_STMT (rdg, e->src))
455 /* Initializes the upstream_mem_writes bitmap following the
456 information from RDG. */
459 mark_nodes_having_upstream_mem_writes (struct graph *rdg)
462 bitmap seen = BITMAP_ALLOC (NULL);
464 for (v = rdg->n_vertices - 1; v >= 0; v--)
465 if (!bitmap_bit_p (seen, v))
468 VEC (int, heap) *nodes = VEC_alloc (int, heap, 3);
470 graphds_dfs (rdg, &v, 1, &nodes, false, NULL);
472 FOR_EACH_VEC_ELT (int, nodes, i, x)
474 if (!bitmap_set_bit (seen, x))
477 if (RDG_MEM_WRITE_STMT (rdg, x)
478 || predecessor_has_mem_write (rdg, &(rdg->vertices[x]))
479 /* In anti dependences the read should occur before
480 the write, this is why both the read and the write
481 should be placed in the same partition. */
482 || has_anti_dependence (&(rdg->vertices[x])))
484 bitmap_set_bit (upstream_mem_writes, x);
488 VEC_free (int, heap, nodes);
492 /* Returns true when vertex u has a memory write node as a predecessor
496 has_upstream_mem_writes (int u)
498 return bitmap_bit_p (upstream_mem_writes, u);
501 static void rdg_flag_vertex_and_dependent (struct graph *, int, bitmap, bitmap,
504 /* Flag the uses of U stopping following the information from
505 upstream_mem_writes. */
508 rdg_flag_uses (struct graph *rdg, int u, bitmap partition, bitmap loops,
509 bitmap processed, bool *part_has_writes)
512 struct vertex *x = &(rdg->vertices[u]);
513 gimple stmt = RDGV_STMT (x);
514 struct graph_edge *anti_dep = has_anti_dependence (x);
516 /* Keep in the same partition the destination of an antidependence,
517 because this is a store to the exact same location. Putting this
518 in another partition is bad for cache locality. */
521 int v = anti_dep->dest;
523 if (!already_processed_vertex_p (processed, v))
524 rdg_flag_vertex_and_dependent (rdg, v, partition, loops,
525 processed, part_has_writes);
528 if (gimple_code (stmt) != GIMPLE_PHI)
530 if ((use_p = gimple_vuse_op (stmt)) != NULL_USE_OPERAND_P)
532 tree use = USE_FROM_PTR (use_p);
534 if (TREE_CODE (use) == SSA_NAME)
536 gimple def_stmt = SSA_NAME_DEF_STMT (use);
537 int v = rdg_vertex_for_stmt (rdg, def_stmt);
540 && !already_processed_vertex_p (processed, v))
541 rdg_flag_vertex_and_dependent (rdg, v, partition, loops,
542 processed, part_has_writes);
547 if (is_gimple_assign (stmt) && has_upstream_mem_writes (u))
549 tree op0 = gimple_assign_lhs (stmt);
551 /* Scalar channels don't have enough space for transmitting data
552 between tasks, unless we add more storage by privatizing. */
553 if (is_gimple_reg (op0))
556 imm_use_iterator iter;
558 FOR_EACH_IMM_USE_FAST (use_p, iter, op0)
560 int v = rdg_vertex_for_stmt (rdg, USE_STMT (use_p));
562 if (!already_processed_vertex_p (processed, v))
563 rdg_flag_vertex_and_dependent (rdg, v, partition, loops,
564 processed, part_has_writes);
570 /* Flag V from RDG as part of PARTITION, and also flag its loop number
574 rdg_flag_vertex (struct graph *rdg, int v, bitmap partition, bitmap loops,
575 bool *part_has_writes)
579 if (!bitmap_set_bit (partition, v))
582 loop = loop_containing_stmt (RDG_STMT (rdg, v));
583 bitmap_set_bit (loops, loop->num);
585 if (rdg_cannot_recompute_vertex_p (rdg, v))
587 *part_has_writes = true;
588 bitmap_clear_bit (remaining_stmts, v);
592 /* Flag in the bitmap PARTITION the vertex V and all its predecessors.
593 Also flag their loop number in LOOPS. */
596 rdg_flag_vertex_and_dependent (struct graph *rdg, int v, bitmap partition,
597 bitmap loops, bitmap processed,
598 bool *part_has_writes)
601 VEC (int, heap) *nodes = VEC_alloc (int, heap, 3);
604 bitmap_set_bit (processed, v);
605 rdg_flag_uses (rdg, v, partition, loops, processed, part_has_writes);
606 graphds_dfs (rdg, &v, 1, &nodes, false, remaining_stmts);
607 rdg_flag_vertex (rdg, v, partition, loops, part_has_writes);
609 FOR_EACH_VEC_ELT (int, nodes, i, x)
610 if (!already_processed_vertex_p (processed, x))
611 rdg_flag_vertex_and_dependent (rdg, x, partition, loops, processed,
614 VEC_free (int, heap, nodes);
617 /* Initialize CONDS with all the condition statements from the basic
621 collect_condition_stmts (struct loop *loop, VEC (gimple, heap) **conds)
625 VEC (edge, heap) *exits = get_loop_exit_edges (loop);
627 FOR_EACH_VEC_ELT (edge, exits, i, e)
629 gimple cond = last_stmt (e->src);
632 VEC_safe_push (gimple, heap, *conds, cond);
635 VEC_free (edge, heap, exits);
638 /* Add to PARTITION all the exit condition statements for LOOPS
639 together with all their dependent statements determined from
643 rdg_flag_loop_exits (struct graph *rdg, bitmap loops, bitmap partition,
644 bitmap processed, bool *part_has_writes)
648 VEC (gimple, heap) *conds = VEC_alloc (gimple, heap, 3);
650 EXECUTE_IF_SET_IN_BITMAP (loops, 0, i, bi)
651 collect_condition_stmts (get_loop (i), &conds);
653 while (!VEC_empty (gimple, conds))
655 gimple cond = VEC_pop (gimple, conds);
656 int v = rdg_vertex_for_stmt (rdg, cond);
657 bitmap new_loops = BITMAP_ALLOC (NULL);
659 if (!already_processed_vertex_p (processed, v))
660 rdg_flag_vertex_and_dependent (rdg, v, partition, new_loops, processed,
663 EXECUTE_IF_SET_IN_BITMAP (new_loops, 0, i, bi)
664 if (bitmap_set_bit (loops, i))
665 collect_condition_stmts (get_loop (i), &conds);
667 BITMAP_FREE (new_loops);
670 VEC_free (gimple, heap, conds);
673 /* Returns a bitmap in which all the statements needed for computing
674 the strongly connected component C of the RDG are flagged, also
675 including the loop exit conditions. */
678 build_rdg_partition_for_component (struct graph *rdg, rdgc c,
679 bool *part_has_writes)
682 bitmap partition = BITMAP_ALLOC (NULL);
683 bitmap loops = BITMAP_ALLOC (NULL);
684 bitmap processed = BITMAP_ALLOC (NULL);
686 FOR_EACH_VEC_ELT (int, c->vertices, i, v)
687 if (!already_processed_vertex_p (processed, v))
688 rdg_flag_vertex_and_dependent (rdg, v, partition, loops, processed,
691 rdg_flag_loop_exits (rdg, loops, partition, processed, part_has_writes);
693 BITMAP_FREE (processed);
698 /* Free memory for COMPONENTS. */
701 free_rdg_components (VEC (rdgc, heap) *components)
706 FOR_EACH_VEC_ELT (rdgc, components, i, x)
708 VEC_free (int, heap, x->vertices);
712 VEC_free (rdgc, heap, components);
715 /* Build the COMPONENTS vector with the strongly connected components
716 of RDG in which the STARTING_VERTICES occur. */
719 rdg_build_components (struct graph *rdg, VEC (int, heap) *starting_vertices,
720 VEC (rdgc, heap) **components)
723 bitmap saved_components = BITMAP_ALLOC (NULL);
724 int n_components = graphds_scc (rdg, NULL);
725 VEC (int, heap) **all_components = XNEWVEC (VEC (int, heap) *, n_components);
727 for (i = 0; i < n_components; i++)
728 all_components[i] = VEC_alloc (int, heap, 3);
730 for (i = 0; i < rdg->n_vertices; i++)
731 VEC_safe_push (int, heap, all_components[rdg->vertices[i].component], i);
733 FOR_EACH_VEC_ELT (int, starting_vertices, i, v)
735 int c = rdg->vertices[v].component;
737 if (bitmap_set_bit (saved_components, c))
739 rdgc x = XCNEW (struct rdg_component);
741 x->vertices = all_components[c];
743 VEC_safe_push (rdgc, heap, *components, x);
747 for (i = 0; i < n_components; i++)
748 if (!bitmap_bit_p (saved_components, i))
749 VEC_free (int, heap, all_components[i]);
751 free (all_components);
752 BITMAP_FREE (saved_components);
755 /* Returns true when it is possible to generate a builtin pattern for
756 the PARTITION of RDG. For the moment we detect only the memset
760 can_generate_builtin (struct graph *rdg, bitmap partition)
768 EXECUTE_IF_SET_IN_BITMAP (partition, 0, i, bi)
769 if (RDG_MEM_READS_STMT (rdg, i))
771 else if (RDG_MEM_WRITE_STMT (rdg, i))
774 if (stmt_with_adjacent_zero_store_dr_p (RDG_STMT (rdg, i)))
778 return stores_zero == 1 && nb_writes == 1 && nb_reads == 0;
781 /* Returns true when PARTITION1 and PARTITION2 have similar memory
785 similar_memory_accesses (struct graph *rdg, bitmap partition1,
789 bitmap_iterator bi, bj;
791 EXECUTE_IF_SET_IN_BITMAP (partition1, 0, i, bi)
792 if (RDG_MEM_WRITE_STMT (rdg, i)
793 || RDG_MEM_READS_STMT (rdg, i))
794 EXECUTE_IF_SET_IN_BITMAP (partition2, 0, j, bj)
795 if (RDG_MEM_WRITE_STMT (rdg, j)
796 || RDG_MEM_READS_STMT (rdg, j))
797 if (rdg_has_similar_memory_accesses (rdg, i, j))
803 /* Fuse all the partitions from PARTITIONS that contain similar memory
804 references, i.e., we're taking care of cache locality. This
805 function does not fuse those partitions that contain patterns that
806 can be code generated with builtins. */
809 fuse_partitions_with_similar_memory_accesses (struct graph *rdg,
810 VEC (bitmap, heap) **partitions)
813 bitmap partition1, partition2;
815 FOR_EACH_VEC_ELT (bitmap, *partitions, p1, partition1)
816 if (!can_generate_builtin (rdg, partition1))
817 FOR_EACH_VEC_ELT (bitmap, *partitions, p2, partition2)
819 && !can_generate_builtin (rdg, partition2)
820 && similar_memory_accesses (rdg, partition1, partition2))
822 bitmap_ior_into (partition1, partition2);
823 VEC_ordered_remove (bitmap, *partitions, p2);
828 /* Returns true when DEF is an SSA_NAME defined in LOOP and used after
832 ssa_name_has_uses_outside_loop_p (tree def, loop_p loop)
834 imm_use_iterator imm_iter;
837 FOR_EACH_IMM_USE_FAST (use_p, imm_iter, def)
838 if (loop != loop_containing_stmt (USE_STMT (use_p)))
844 /* Returns true when STMT defines a scalar variable used after the
848 stmt_has_scalar_dependences_outside_loop (gimple stmt)
852 switch (gimple_code (stmt))
855 name = gimple_assign_lhs (stmt);
859 name = gimple_phi_result (stmt);
866 return TREE_CODE (name) == SSA_NAME
867 && ssa_name_has_uses_outside_loop_p (name, loop_containing_stmt (stmt));
870 /* Returns true when STMT will be code generated in a partition of RDG
871 different than PART and that will not be code generated as a
875 stmt_generated_in_another_partition (struct graph *rdg, gimple stmt, int part,
876 VEC (bitmap, heap) *partitions)
883 FOR_EACH_VEC_ELT (bitmap, partitions, p, pp)
885 && !can_generate_builtin (rdg, pp))
886 EXECUTE_IF_SET_IN_BITMAP (pp, 0, i, bi)
887 if (stmt == RDG_STMT (rdg, i))
893 /* For each partition in PARTITIONS that will be code generated using
894 a builtin, add its scalar computations used after the loop to
898 add_scalar_computations_to_partition (struct graph *rdg,
899 VEC (bitmap, heap) *partitions,
906 bitmap l = BITMAP_ALLOC (NULL);
907 bitmap pr = BITMAP_ALLOC (NULL);
910 FOR_EACH_VEC_ELT (bitmap, partitions, p, pp)
911 if (can_generate_builtin (rdg, pp))
912 EXECUTE_IF_SET_IN_BITMAP (pp, 0, i, bi)
913 if (stmt_has_scalar_dependences_outside_loop (RDG_STMT (rdg, i))
914 && !stmt_generated_in_another_partition (rdg, RDG_STMT (rdg, i), p,
916 rdg_flag_vertex_and_dependent (rdg, i, partition, l, pr, &f);
918 rdg_flag_loop_exits (rdg, l, partition, pr, &f);
924 /* Aggregate several components into a useful partition that is
925 registered in the PARTITIONS vector. Partitions will be
926 distributed in different loops. */
929 rdg_build_partitions (struct graph *rdg, VEC (rdgc, heap) *components,
930 VEC (int, heap) **other_stores,
931 VEC (bitmap, heap) **partitions, bitmap processed)
935 bitmap partition = BITMAP_ALLOC (NULL);
937 FOR_EACH_VEC_ELT (rdgc, components, i, x)
940 bool part_has_writes = false;
941 int v = VEC_index (int, x->vertices, 0);
943 if (bitmap_bit_p (processed, v))
946 np = build_rdg_partition_for_component (rdg, x, &part_has_writes);
947 bitmap_ior_into (partition, np);
948 bitmap_ior_into (processed, np);
953 if (dump_file && (dump_flags & TDF_DETAILS))
955 fprintf (dump_file, "ldist useful partition:\n");
956 dump_bitmap (dump_file, partition);
959 VEC_safe_push (bitmap, heap, *partitions, partition);
960 partition = BITMAP_ALLOC (NULL);
964 /* Add the nodes from the RDG that were not marked as processed, and
965 that are used outside the current loop. These are scalar
966 computations that are not yet part of previous partitions. */
967 for (i = 0; i < rdg->n_vertices; i++)
968 if (!bitmap_bit_p (processed, i)
969 && rdg_defs_used_in_other_loops_p (rdg, i))
970 VEC_safe_push (int, heap, *other_stores, i);
972 /* If there are still statements left in the OTHER_STORES array,
973 create other components and partitions with these stores and
974 their dependences. */
975 if (VEC_length (int, *other_stores) > 0)
977 VEC (rdgc, heap) *comps = VEC_alloc (rdgc, heap, 3);
978 VEC (int, heap) *foo = VEC_alloc (int, heap, 3);
980 rdg_build_components (rdg, *other_stores, &comps);
981 rdg_build_partitions (rdg, comps, &foo, partitions, processed);
983 VEC_free (int, heap, foo);
984 free_rdg_components (comps);
987 add_scalar_computations_to_partition (rdg, *partitions, partition);
989 /* If there is something left in the last partition, save it. */
990 if (bitmap_count_bits (partition) > 0)
991 VEC_safe_push (bitmap, heap, *partitions, partition);
993 BITMAP_FREE (partition);
995 fuse_partitions_with_similar_memory_accesses (rdg, partitions);
998 /* Dump to FILE the PARTITIONS. */
1001 dump_rdg_partitions (FILE *file, VEC (bitmap, heap) *partitions)
1006 FOR_EACH_VEC_ELT (bitmap, partitions, i, partition)
1007 debug_bitmap_file (file, partition);
1010 /* Debug PARTITIONS. */
1011 extern void debug_rdg_partitions (VEC (bitmap, heap) *);
1014 debug_rdg_partitions (VEC (bitmap, heap) *partitions)
1016 dump_rdg_partitions (stderr, partitions);
1019 /* Returns the number of read and write operations in the RDG. */
1022 number_of_rw_in_rdg (struct graph *rdg)
1026 for (i = 0; i < rdg->n_vertices; i++)
1028 if (RDG_MEM_WRITE_STMT (rdg, i))
1031 if (RDG_MEM_READS_STMT (rdg, i))
1038 /* Returns the number of read and write operations in a PARTITION of
1042 number_of_rw_in_partition (struct graph *rdg, bitmap partition)
1048 EXECUTE_IF_SET_IN_BITMAP (partition, 0, i, ii)
1050 if (RDG_MEM_WRITE_STMT (rdg, i))
1053 if (RDG_MEM_READS_STMT (rdg, i))
1060 /* Returns true when one of the PARTITIONS contains all the read or
1061 write operations of RDG. */
1064 partition_contains_all_rw (struct graph *rdg, VEC (bitmap, heap) *partitions)
1068 int nrw = number_of_rw_in_rdg (rdg);
1070 FOR_EACH_VEC_ELT (bitmap, partitions, i, partition)
1071 if (nrw == number_of_rw_in_partition (rdg, partition))
1077 /* Generate code from STARTING_VERTICES in RDG. Returns the number of
1078 distributed loops. */
1081 ldist_gen (struct loop *loop, struct graph *rdg,
1082 VEC (int, heap) *starting_vertices)
1085 VEC (rdgc, heap) *components = VEC_alloc (rdgc, heap, 3);
1086 VEC (bitmap, heap) *partitions = VEC_alloc (bitmap, heap, 3);
1087 VEC (int, heap) *other_stores = VEC_alloc (int, heap, 3);
1088 bitmap partition, processed = BITMAP_ALLOC (NULL);
1090 remaining_stmts = BITMAP_ALLOC (NULL);
1091 upstream_mem_writes = BITMAP_ALLOC (NULL);
1093 for (i = 0; i < rdg->n_vertices; i++)
1095 bitmap_set_bit (remaining_stmts, i);
1097 /* Save in OTHER_STORES all the memory writes that are not in
1098 STARTING_VERTICES. */
1099 if (RDG_MEM_WRITE_STMT (rdg, i))
1105 FOR_EACH_VEC_ELT (int, starting_vertices, j, v)
1113 VEC_safe_push (int, heap, other_stores, i);
1117 mark_nodes_having_upstream_mem_writes (rdg);
1118 rdg_build_components (rdg, starting_vertices, &components);
1119 rdg_build_partitions (rdg, components, &other_stores, &partitions,
1121 BITMAP_FREE (processed);
1122 nbp = VEC_length (bitmap, partitions);
1125 || partition_contains_all_rw (rdg, partitions))
1128 if (dump_file && (dump_flags & TDF_DETAILS))
1129 dump_rdg_partitions (dump_file, partitions);
1131 FOR_EACH_VEC_ELT (bitmap, partitions, i, partition)
1132 if (!generate_code_for_partition (loop, partition, i < nbp - 1))
1135 rewrite_into_loop_closed_ssa (NULL, TODO_update_ssa);
1136 update_ssa (TODO_update_ssa_only_virtuals | TODO_update_ssa);
1140 BITMAP_FREE (remaining_stmts);
1141 BITMAP_FREE (upstream_mem_writes);
1143 FOR_EACH_VEC_ELT (bitmap, partitions, i, partition)
1144 BITMAP_FREE (partition);
1146 VEC_free (int, heap, other_stores);
1147 VEC_free (bitmap, heap, partitions);
1148 free_rdg_components (components);
1152 /* Distributes the code from LOOP in such a way that producer
1153 statements are placed before consumer statements. When STMTS is
1154 NULL, performs the maximal distribution, if STMTS is not NULL,
1155 tries to separate only these statements from the LOOP's body.
1156 Returns the number of distributed loops. */
1159 distribute_loop (struct loop *loop, VEC (gimple, heap) *stmts)
1165 VEC (int, heap) *vertices;
1166 VEC (ddr_p, heap) *dependence_relations;
1167 VEC (data_reference_p, heap) *datarefs;
1168 VEC (loop_p, heap) *loop_nest;
1170 if (loop->num_nodes > 2)
1172 if (dump_file && (dump_flags & TDF_DETAILS))
1174 "FIXME: Loop %d not distributed: it has more than two basic blocks.\n",
1180 datarefs = VEC_alloc (data_reference_p, heap, 10);
1181 dependence_relations = VEC_alloc (ddr_p, heap, 100);
1182 loop_nest = VEC_alloc (loop_p, heap, 3);
1183 rdg = build_rdg (loop, &loop_nest, &dependence_relations, &datarefs);
1187 if (dump_file && (dump_flags & TDF_DETAILS))
1189 "FIXME: Loop %d not distributed: failed to build the RDG.\n",
1192 free_dependence_relations (dependence_relations);
1193 free_data_refs (datarefs);
1194 VEC_free (loop_p, heap, loop_nest);
1198 vertices = VEC_alloc (int, heap, 3);
1200 if (dump_file && (dump_flags & TDF_DETAILS))
1201 dump_rdg (dump_file, rdg);
1203 FOR_EACH_VEC_ELT (gimple, stmts, i, s)
1205 int v = rdg_vertex_for_stmt (rdg, s);
1209 VEC_safe_push (int, heap, vertices, v);
1211 if (dump_file && (dump_flags & TDF_DETAILS))
1213 "ldist asked to generate code for vertex %d\n", v);
1217 res = ldist_gen (loop, rdg, vertices);
1218 VEC_free (int, heap, vertices);
1220 free_dependence_relations (dependence_relations);
1221 free_data_refs (datarefs);
1222 VEC_free (loop_p, heap, loop_nest);
1226 /* Distribute all loops in the current function. */
1229 tree_loop_distribution (void)
1233 int nb_generated_loops = 0;
1235 FOR_EACH_LOOP (li, loop, 0)
1237 VEC (gimple, heap) *work_list = NULL;
1238 int num = loop->num;
1240 /* If the loop doesn't have a single exit we will fail anyway,
1241 so do that early. */
1242 if (!single_exit (loop))
1245 /* If both flag_tree_loop_distribute_patterns and
1246 flag_tree_loop_distribution are set, then only
1247 distribute_patterns is executed. */
1248 if (flag_tree_loop_distribute_patterns)
1250 /* With the following working list, we're asking
1251 distribute_loop to separate from the rest of the loop the
1252 stores of the form "A[i] = 0". */
1253 stores_zero_from_loop (loop, &work_list);
1255 /* Do nothing if there are no patterns to be distributed. */
1256 if (VEC_length (gimple, work_list) > 0)
1257 nb_generated_loops = distribute_loop (loop, work_list);
1259 else if (flag_tree_loop_distribution)
1261 /* With the following working list, we're asking
1262 distribute_loop to separate the stores of the loop: when
1263 dependences allow, it will end on having one store per
1265 stores_from_loop (loop, &work_list);
1267 /* A simple heuristic for cache locality is to not split
1268 stores to the same array. Without this call, an unrolled
1269 loop would be split into as many loops as unroll factor,
1270 each loop storing in the same array. */
1271 remove_similar_memory_refs (&work_list);
1273 nb_generated_loops = distribute_loop (loop, work_list);
1276 if (dump_file && (dump_flags & TDF_DETAILS))
1278 if (nb_generated_loops > 1)
1279 fprintf (dump_file, "Loop %d distributed: split to %d loops.\n",
1280 num, nb_generated_loops);
1282 fprintf (dump_file, "Loop %d is the same.\n", num);
1285 verify_loop_structure ();
1287 VEC_free (gimple, heap, work_list);
1294 gate_tree_loop_distribution (void)
1296 return flag_tree_loop_distribution
1297 || flag_tree_loop_distribute_patterns;
1300 struct gimple_opt_pass pass_loop_distribution =
1305 gate_tree_loop_distribution, /* gate */
1306 tree_loop_distribution, /* execute */
1309 0, /* static_pass_number */
1310 TV_TREE_LOOP_DISTRIBUTION, /* tv_id */
1311 PROP_cfg | PROP_ssa, /* properties_required */
1312 0, /* properties_provided */
1313 0, /* properties_destroyed */
1314 0, /* todo_flags_start */
1315 0 /* todo_flags_finish */