2015-06-24 François Dumont <fdumont@gcc.gnu.org>
[official-gcc.git] / gcc / graphite-sese-to-poly.c
blob271c49925c0be17aaf5221bee0458813b7cf1cc3
1 /* Conversion of SESE regions to Polyhedra.
2 Copyright (C) 2009-2015 Free Software Foundation, Inc.
3 Contributed by Sebastian Pop <sebastian.pop@amd.com>.
5 This file is part of GCC.
7 GCC is free software; you can redistribute it and/or modify
8 it under the terms of the GNU General Public License as published by
9 the Free Software Foundation; either version 3, or (at your option)
10 any later version.
12 GCC is distributed in the hope that it will be useful,
13 but WITHOUT ANY WARRANTY; without even the implied warranty of
14 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
15 GNU General Public License for more details.
17 You should have received a copy of the GNU General Public License
18 along with GCC; see the file COPYING3. If not see
19 <http://www.gnu.org/licenses/>. */
21 #include "config.h"
23 #ifdef HAVE_isl
24 #include <isl/set.h>
25 #include <isl/map.h>
26 #include <isl/union_map.h>
27 #include <isl/constraint.h>
28 #include <isl/aff.h>
29 #include <isl/val.h>
31 /* Since ISL-0.13, the extern is in val_gmp.h. */
32 #if !defined(HAVE_ISL_SCHED_CONSTRAINTS_COMPUTE_SCHEDULE) && defined(__cplusplus)
33 extern "C" {
34 #endif
35 #include <isl/val_gmp.h>
36 #if !defined(HAVE_ISL_SCHED_CONSTRAINTS_COMPUTE_SCHEDULE) && defined(__cplusplus)
38 #endif
39 #endif
41 #include "system.h"
42 #include "coretypes.h"
43 #include "alias.h"
44 #include "symtab.h"
45 #include "options.h"
46 #include "tree.h"
47 #include "fold-const.h"
48 #include "predict.h"
49 #include "tm.h"
50 #include "hard-reg-set.h"
51 #include "function.h"
52 #include "dominance.h"
53 #include "cfg.h"
54 #include "basic-block.h"
55 #include "tree-ssa-alias.h"
56 #include "internal-fn.h"
57 #include "gimple-expr.h"
58 #include "gimple.h"
59 #include "gimple-iterator.h"
60 #include "gimplify.h"
61 #include "gimplify-me.h"
62 #include "gimple-ssa.h"
63 #include "tree-cfg.h"
64 #include "tree-phinodes.h"
65 #include "ssa-iterators.h"
66 #include "stringpool.h"
67 #include "tree-ssanames.h"
68 #include "tree-ssa-loop-manip.h"
69 #include "tree-ssa-loop-niter.h"
70 #include "tree-ssa-loop.h"
71 #include "tree-into-ssa.h"
72 #include "tree-pass.h"
73 #include "cfgloop.h"
74 #include "tree-chrec.h"
75 #include "tree-data-ref.h"
76 #include "tree-scalar-evolution.h"
77 #include "domwalk.h"
78 #include "sese.h"
79 #include "tree-ssa-propagate.h"
81 #ifdef HAVE_isl
82 #include "rtl.h"
83 #include "flags.h"
84 #include "insn-config.h"
85 #include "expmed.h"
86 #include "dojump.h"
87 #include "explow.h"
88 #include "calls.h"
89 #include "emit-rtl.h"
90 #include "varasm.h"
91 #include "stmt.h"
92 #include "expr.h"
93 #include "graphite-poly.h"
94 #include "graphite-sese-to-poly.h"
97 /* Assigns to RES the value of the INTEGER_CST T. */
99 static inline void
100 tree_int_to_gmp (tree t, mpz_t res)
102 wi::to_mpz (t, res, TYPE_SIGN (TREE_TYPE (t)));
105 /* Returns the index of the PHI argument defined in the outermost
106 loop. */
108 static size_t
109 phi_arg_in_outermost_loop (gphi *phi)
111 loop_p loop = gimple_bb (phi)->loop_father;
112 size_t i, res = 0;
114 for (i = 0; i < gimple_phi_num_args (phi); i++)
115 if (!flow_bb_inside_loop_p (loop, gimple_phi_arg_edge (phi, i)->src))
117 loop = gimple_phi_arg_edge (phi, i)->src->loop_father;
118 res = i;
121 return res;
124 /* Removes a simple copy phi node "RES = phi (INIT, RES)" at position
125 PSI by inserting on the loop ENTRY edge assignment "RES = INIT". */
127 static void
128 remove_simple_copy_phi (gphi_iterator *psi)
130 gphi *phi = psi->phi ();
131 tree res = gimple_phi_result (phi);
132 size_t entry = phi_arg_in_outermost_loop (phi);
133 tree init = gimple_phi_arg_def (phi, entry);
134 gassign *stmt = gimple_build_assign (res, init);
135 edge e = gimple_phi_arg_edge (phi, entry);
137 remove_phi_node (psi, false);
138 gsi_insert_on_edge_immediate (e, stmt);
141 /* Removes an invariant phi node at position PSI by inserting on the
142 loop ENTRY edge the assignment RES = INIT. */
144 static void
145 remove_invariant_phi (sese region, gphi_iterator *psi)
147 gphi *phi = psi->phi ();
148 loop_p loop = loop_containing_stmt (phi);
149 tree res = gimple_phi_result (phi);
150 tree scev = scalar_evolution_in_region (region, loop, res);
151 size_t entry = phi_arg_in_outermost_loop (phi);
152 edge e = gimple_phi_arg_edge (phi, entry);
153 tree var;
154 gassign *stmt;
155 gimple_seq stmts = NULL;
157 if (tree_contains_chrecs (scev, NULL))
158 scev = gimple_phi_arg_def (phi, entry);
160 var = force_gimple_operand (scev, &stmts, true, NULL_TREE);
161 stmt = gimple_build_assign (res, var);
162 remove_phi_node (psi, false);
164 gimple_seq_add_stmt (&stmts, stmt);
165 gsi_insert_seq_on_edge (e, stmts);
166 gsi_commit_edge_inserts ();
167 SSA_NAME_DEF_STMT (res) = stmt;
170 /* Returns true when the phi node at PSI is of the form "a = phi (a, x)". */
172 static inline bool
173 simple_copy_phi_p (gphi *phi)
175 tree res;
177 if (gimple_phi_num_args (phi) != 2)
178 return false;
180 res = gimple_phi_result (phi);
181 return (res == gimple_phi_arg_def (phi, 0)
182 || res == gimple_phi_arg_def (phi, 1));
185 /* Returns true when the phi node at position PSI is a reduction phi
186 node in REGION. Otherwise moves the pointer PSI to the next phi to
187 be considered. */
189 static bool
190 reduction_phi_p (sese region, gphi_iterator *psi)
192 loop_p loop;
193 gphi *phi = psi->phi ();
194 tree res = gimple_phi_result (phi);
196 loop = loop_containing_stmt (phi);
198 if (simple_copy_phi_p (phi))
200 /* PRE introduces phi nodes like these, for an example,
201 see id-5.f in the fortran graphite testsuite:
203 # prephitmp.85_265 = PHI <prephitmp.85_258(33), prephitmp.85_265(18)>
205 remove_simple_copy_phi (psi);
206 return false;
209 if (scev_analyzable_p (res, region))
211 tree scev = scalar_evolution_in_region (region, loop, res);
213 if (evolution_function_is_invariant_p (scev, loop->num))
214 remove_invariant_phi (region, psi);
215 else
216 gsi_next (psi);
218 return false;
221 /* All the other cases are considered reductions. */
222 return true;
225 /* Store the GRAPHITE representation of BB. */
227 static gimple_bb_p
228 new_gimple_bb (basic_block bb, vec<data_reference_p> drs)
230 struct gimple_bb *gbb;
232 gbb = XNEW (struct gimple_bb);
233 bb->aux = gbb;
234 GBB_BB (gbb) = bb;
235 GBB_DATA_REFS (gbb) = drs;
236 GBB_CONDITIONS (gbb).create (0);
237 GBB_CONDITION_CASES (gbb).create (0);
239 return gbb;
242 static void
243 free_data_refs_aux (vec<data_reference_p> datarefs)
245 unsigned int i;
246 struct data_reference *dr;
248 FOR_EACH_VEC_ELT (datarefs, i, dr)
249 if (dr->aux)
251 base_alias_pair *bap = (base_alias_pair *)(dr->aux);
253 free (bap->alias_set);
255 free (bap);
256 dr->aux = NULL;
259 /* Frees GBB. */
261 static void
262 free_gimple_bb (struct gimple_bb *gbb)
264 free_data_refs_aux (GBB_DATA_REFS (gbb));
265 free_data_refs (GBB_DATA_REFS (gbb));
267 GBB_CONDITIONS (gbb).release ();
268 GBB_CONDITION_CASES (gbb).release ();
269 GBB_BB (gbb)->aux = 0;
270 XDELETE (gbb);
273 /* Deletes all gimple bbs in SCOP. */
275 static void
276 remove_gbbs_in_scop (scop_p scop)
278 int i;
279 poly_bb_p pbb;
281 FOR_EACH_VEC_ELT (SCOP_BBS (scop), i, pbb)
282 free_gimple_bb (PBB_BLACK_BOX (pbb));
285 /* Deletes all scops in SCOPS. */
287 void
288 free_scops (vec<scop_p> scops)
290 int i;
291 scop_p scop;
293 FOR_EACH_VEC_ELT (scops, i, scop)
295 remove_gbbs_in_scop (scop);
296 free_sese (SCOP_REGION (scop));
297 free_scop (scop);
300 scops.release ();
303 /* Same as outermost_loop_in_sese, returns the outermost loop
304 containing BB in REGION, but makes sure that the returned loop
305 belongs to the REGION, and so this returns the first loop in the
306 REGION when the loop containing BB does not belong to REGION. */
308 static loop_p
309 outermost_loop_in_sese_1 (sese region, basic_block bb)
311 loop_p nest = outermost_loop_in_sese (region, bb);
313 if (loop_in_sese_p (nest, region))
314 return nest;
316 /* When the basic block BB does not belong to a loop in the region,
317 return the first loop in the region. */
318 nest = nest->inner;
319 while (nest)
320 if (loop_in_sese_p (nest, region))
321 break;
322 else
323 nest = nest->next;
325 gcc_assert (nest);
326 return nest;
329 /* Generates a polyhedral black box only if the bb contains interesting
330 information. */
332 static gimple_bb_p
333 try_generate_gimple_bb (scop_p scop, basic_block bb)
335 vec<data_reference_p> drs;
336 drs.create (5);
337 sese region = SCOP_REGION (scop);
338 loop_p nest = outermost_loop_in_sese_1 (region, bb);
339 gimple_stmt_iterator gsi;
341 for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi))
343 gimple stmt = gsi_stmt (gsi);
344 loop_p loop;
346 if (is_gimple_debug (stmt))
347 continue;
349 loop = loop_containing_stmt (stmt);
350 if (!loop_in_sese_p (loop, region))
351 loop = nest;
353 graphite_find_data_references_in_stmt (nest, loop, stmt, &drs);
356 return new_gimple_bb (bb, drs);
359 /* Returns true if all predecessors of BB, that are not dominated by BB, are
360 marked in MAP. The predecessors dominated by BB are loop latches and will
361 be handled after BB. */
363 static bool
364 all_non_dominated_preds_marked_p (basic_block bb, sbitmap map)
366 edge e;
367 edge_iterator ei;
369 FOR_EACH_EDGE (e, ei, bb->preds)
370 if (!bitmap_bit_p (map, e->src->index)
371 && !dominated_by_p (CDI_DOMINATORS, e->src, bb))
372 return false;
374 return true;
377 /* Compare the depth of two basic_block's P1 and P2. */
379 static int
380 compare_bb_depths (const void *p1, const void *p2)
382 const_basic_block const bb1 = *(const_basic_block const*)p1;
383 const_basic_block const bb2 = *(const_basic_block const*)p2;
384 int d1 = loop_depth (bb1->loop_father);
385 int d2 = loop_depth (bb2->loop_father);
387 if (d1 < d2)
388 return 1;
390 if (d1 > d2)
391 return -1;
393 return 0;
396 /* Sort the basic blocks from DOM such that the first are the ones at
397 a deepest loop level. */
399 static void
400 graphite_sort_dominated_info (vec<basic_block> dom)
402 dom.qsort (compare_bb_depths);
405 /* Recursive helper function for build_scops_bbs. */
407 static void
408 build_scop_bbs_1 (scop_p scop, sbitmap visited, basic_block bb)
410 sese region = SCOP_REGION (scop);
411 vec<basic_block> dom;
412 poly_bb_p pbb;
414 if (bitmap_bit_p (visited, bb->index)
415 || !bb_in_sese_p (bb, region))
416 return;
418 pbb = new_poly_bb (scop, try_generate_gimple_bb (scop, bb));
419 SCOP_BBS (scop).safe_push (pbb);
420 bitmap_set_bit (visited, bb->index);
422 dom = get_dominated_by (CDI_DOMINATORS, bb);
424 if (!dom.exists ())
425 return;
427 graphite_sort_dominated_info (dom);
429 while (!dom.is_empty ())
431 int i;
432 basic_block dom_bb;
434 FOR_EACH_VEC_ELT (dom, i, dom_bb)
435 if (all_non_dominated_preds_marked_p (dom_bb, visited))
437 build_scop_bbs_1 (scop, visited, dom_bb);
438 dom.unordered_remove (i);
439 break;
443 dom.release ();
446 /* Gather the basic blocks belonging to the SCOP. */
448 static void
449 build_scop_bbs (scop_p scop)
451 sbitmap visited = sbitmap_alloc (last_basic_block_for_fn (cfun));
452 sese region = SCOP_REGION (scop);
454 bitmap_clear (visited);
455 build_scop_bbs_1 (scop, visited, SESE_ENTRY_BB (region));
456 sbitmap_free (visited);
459 /* Return an ISL identifier for the polyhedral basic block PBB. */
461 static isl_id *
462 isl_id_for_pbb (scop_p s, poly_bb_p pbb)
464 char name[50];
465 snprintf (name, sizeof (name), "S_%d", pbb_index (pbb));
466 return isl_id_alloc (s->ctx, name, pbb);
469 /* Converts the STATIC_SCHEDULE of PBB into a scattering polyhedron.
470 We generate SCATTERING_DIMENSIONS scattering dimensions.
472 CLooG 0.15.0 and previous versions require, that all
473 scattering functions of one CloogProgram have the same number of
474 scattering dimensions, therefore we allow to specify it. This
475 should be removed in future versions of CLooG.
477 The scattering polyhedron consists of these dimensions: scattering,
478 loop_iterators, parameters.
480 Example:
482 | scattering_dimensions = 5
483 | used_scattering_dimensions = 3
484 | nb_iterators = 1
485 | scop_nb_params = 2
487 | Schedule:
489 | 4 5
491 | Scattering polyhedron:
493 | scattering: {s1, s2, s3, s4, s5}
494 | loop_iterators: {i}
495 | parameters: {p1, p2}
497 | s1 s2 s3 s4 s5 i p1 p2 1
498 | 1 0 0 0 0 0 0 0 -4 = 0
499 | 0 1 0 0 0 -1 0 0 0 = 0
500 | 0 0 1 0 0 0 0 0 -5 = 0 */
502 static void
503 build_pbb_scattering_polyhedrons (isl_aff *static_sched,
504 poly_bb_p pbb, int scattering_dimensions)
506 int i;
507 int nb_iterators = pbb_dim_iter_domain (pbb);
508 int used_scattering_dimensions = nb_iterators * 2 + 1;
509 isl_val *val;
510 isl_space *dc, *dm;
512 gcc_assert (scattering_dimensions >= used_scattering_dimensions);
514 dc = isl_set_get_space (pbb->domain);
515 dm = isl_space_add_dims (isl_space_from_domain (dc),
516 isl_dim_out, scattering_dimensions);
517 pbb->schedule = isl_map_universe (dm);
519 for (i = 0; i < scattering_dimensions; i++)
521 /* Textual order inside this loop. */
522 if ((i % 2) == 0)
524 isl_constraint *c = isl_equality_alloc
525 (isl_local_space_from_space (isl_map_get_space (pbb->schedule)));
527 val = isl_aff_get_coefficient_val (static_sched, isl_dim_in, i / 2);
529 val = isl_val_neg (val);
530 c = isl_constraint_set_constant_val (c, val);
531 c = isl_constraint_set_coefficient_si (c, isl_dim_out, i, 1);
532 pbb->schedule = isl_map_add_constraint (pbb->schedule, c);
535 /* Iterations of this loop. */
536 else /* if ((i % 2) == 1) */
538 int loop = (i - 1) / 2;
539 pbb->schedule = isl_map_equate (pbb->schedule, isl_dim_in, loop,
540 isl_dim_out, i);
544 pbb->transformed = isl_map_copy (pbb->schedule);
547 /* Build for BB the static schedule.
549 The static schedule is a Dewey numbering of the abstract syntax
550 tree: http://en.wikipedia.org/wiki/Dewey_Decimal_Classification
552 The following example informally defines the static schedule:
555 for (i: ...)
557 for (j: ...)
563 for (k: ...)
571 Static schedules for A to F:
573 DEPTH
574 0 1 2
576 B 1 0 0
577 C 1 0 1
578 D 1 1 0
579 E 1 1 1
583 static void
584 build_scop_scattering (scop_p scop)
586 int i;
587 poly_bb_p pbb;
588 gimple_bb_p previous_gbb = NULL;
589 isl_space *dc = isl_set_get_space (scop->context);
590 isl_aff *static_sched;
592 dc = isl_space_add_dims (dc, isl_dim_set, number_of_loops (cfun));
593 static_sched = isl_aff_zero_on_domain (isl_local_space_from_space (dc));
595 /* We have to start schedules at 0 on the first component and
596 because we cannot compare_prefix_loops against a previous loop,
597 prefix will be equal to zero, and that index will be
598 incremented before copying. */
599 static_sched = isl_aff_add_coefficient_si (static_sched, isl_dim_in, 0, -1);
601 FOR_EACH_VEC_ELT (SCOP_BBS (scop), i, pbb)
603 gimple_bb_p gbb = PBB_BLACK_BOX (pbb);
604 int prefix;
605 int nb_scat_dims = pbb_dim_iter_domain (pbb) * 2 + 1;
607 if (previous_gbb)
608 prefix = nb_common_loops (SCOP_REGION (scop), previous_gbb, gbb);
609 else
610 prefix = 0;
612 previous_gbb = gbb;
614 static_sched = isl_aff_add_coefficient_si (static_sched, isl_dim_in,
615 prefix, 1);
616 build_pbb_scattering_polyhedrons (static_sched, pbb, nb_scat_dims);
619 isl_aff_free (static_sched);
622 static isl_pw_aff *extract_affine (scop_p, tree, __isl_take isl_space *space);
624 /* Extract an affine expression from the chain of recurrence E. */
626 static isl_pw_aff *
627 extract_affine_chrec (scop_p s, tree e, __isl_take isl_space *space)
629 isl_pw_aff *lhs = extract_affine (s, CHREC_LEFT (e), isl_space_copy (space));
630 isl_pw_aff *rhs = extract_affine (s, CHREC_RIGHT (e), isl_space_copy (space));
631 isl_local_space *ls = isl_local_space_from_space (space);
632 unsigned pos = sese_loop_depth ((sese) s->region, get_chrec_loop (e)) - 1;
633 isl_aff *loop = isl_aff_set_coefficient_si
634 (isl_aff_zero_on_domain (ls), isl_dim_in, pos, 1);
635 isl_pw_aff *l = isl_pw_aff_from_aff (loop);
637 /* Before multiplying, make sure that the result is affine. */
638 gcc_assert (isl_pw_aff_is_cst (rhs)
639 || isl_pw_aff_is_cst (l));
641 return isl_pw_aff_add (lhs, isl_pw_aff_mul (rhs, l));
644 /* Extract an affine expression from the mult_expr E. */
646 static isl_pw_aff *
647 extract_affine_mul (scop_p s, tree e, __isl_take isl_space *space)
649 isl_pw_aff *lhs = extract_affine (s, TREE_OPERAND (e, 0),
650 isl_space_copy (space));
651 isl_pw_aff *rhs = extract_affine (s, TREE_OPERAND (e, 1), space);
653 if (!isl_pw_aff_is_cst (lhs)
654 && !isl_pw_aff_is_cst (rhs))
656 isl_pw_aff_free (lhs);
657 isl_pw_aff_free (rhs);
658 return NULL;
661 return isl_pw_aff_mul (lhs, rhs);
664 /* Return an ISL identifier from the name of the ssa_name E. */
666 static isl_id *
667 isl_id_for_ssa_name (scop_p s, tree e)
669 const char *name = get_name (e);
670 isl_id *id;
672 if (name)
673 id = isl_id_alloc (s->ctx, name, e);
674 else
676 char name1[50];
677 snprintf (name1, sizeof (name1), "P_%d", SSA_NAME_VERSION (e));
678 id = isl_id_alloc (s->ctx, name1, e);
681 return id;
684 /* Return an ISL identifier for the data reference DR. */
686 static isl_id *
687 isl_id_for_dr (scop_p s, data_reference_p dr ATTRIBUTE_UNUSED)
689 /* Data references all get the same isl_id. They need to be comparable
690 and are distinguished through the first dimension, which contains the
691 alias set number. */
692 return isl_id_alloc (s->ctx, "", 0);
695 /* Extract an affine expression from the ssa_name E. */
697 static isl_pw_aff *
698 extract_affine_name (scop_p s, tree e, __isl_take isl_space *space)
700 isl_aff *aff;
701 isl_set *dom;
702 isl_id *id;
703 int dimension;
705 id = isl_id_for_ssa_name (s, e);
706 dimension = isl_space_find_dim_by_id (space, isl_dim_param, id);
707 isl_id_free (id);
708 dom = isl_set_universe (isl_space_copy (space));
709 aff = isl_aff_zero_on_domain (isl_local_space_from_space (space));
710 aff = isl_aff_add_coefficient_si (aff, isl_dim_param, dimension, 1);
711 return isl_pw_aff_alloc (dom, aff);
714 /* Extract an affine expression from the gmp constant G. */
716 static isl_pw_aff *
717 extract_affine_gmp (mpz_t g, __isl_take isl_space *space)
719 isl_local_space *ls = isl_local_space_from_space (isl_space_copy (space));
720 isl_aff *aff = isl_aff_zero_on_domain (ls);
721 isl_set *dom = isl_set_universe (space);
722 isl_val *v;
723 isl_ctx *ct;
725 ct = isl_aff_get_ctx (aff);
726 v = isl_val_int_from_gmp (ct, g);
727 aff = isl_aff_add_constant_val (aff, v);
729 return isl_pw_aff_alloc (dom, aff);
732 /* Extract an affine expression from the integer_cst E. */
734 static isl_pw_aff *
735 extract_affine_int (tree e, __isl_take isl_space *space)
737 isl_pw_aff *res;
738 mpz_t g;
740 mpz_init (g);
741 tree_int_to_gmp (e, g);
742 res = extract_affine_gmp (g, space);
743 mpz_clear (g);
745 return res;
748 /* Compute pwaff mod 2^width. */
750 extern isl_ctx *the_isl_ctx;
752 static isl_pw_aff *
753 wrap (isl_pw_aff *pwaff, unsigned width)
755 isl_val *mod;
757 mod = isl_val_int_from_ui(the_isl_ctx, width);
758 mod = isl_val_2exp (mod);
759 pwaff = isl_pw_aff_mod_val (pwaff, mod);
761 return pwaff;
764 /* When parameter NAME is in REGION, returns its index in SESE_PARAMS.
765 Otherwise returns -1. */
767 static inline int
768 parameter_index_in_region_1 (tree name, sese region)
770 int i;
771 tree p;
773 gcc_assert (TREE_CODE (name) == SSA_NAME);
775 FOR_EACH_VEC_ELT (SESE_PARAMS (region), i, p)
776 if (p == name)
777 return i;
779 return -1;
782 /* When the parameter NAME is in REGION, returns its index in
783 SESE_PARAMS. Otherwise this function inserts NAME in SESE_PARAMS
784 and returns the index of NAME. */
786 static int
787 parameter_index_in_region (tree name, sese region)
789 int i;
791 gcc_assert (TREE_CODE (name) == SSA_NAME);
793 i = parameter_index_in_region_1 (name, region);
794 if (i != -1)
795 return i;
797 gcc_assert (SESE_ADD_PARAMS (region));
799 i = SESE_PARAMS (region).length ();
800 SESE_PARAMS (region).safe_push (name);
801 return i;
804 /* Extract an affine expression from the tree E in the scop S. */
806 static isl_pw_aff *
807 extract_affine (scop_p s, tree e, __isl_take isl_space *space)
809 isl_pw_aff *lhs, *rhs, *res;
810 tree type;
812 if (e == chrec_dont_know) {
813 isl_space_free (space);
814 return NULL;
817 switch (TREE_CODE (e))
819 case POLYNOMIAL_CHREC:
820 res = extract_affine_chrec (s, e, space);
821 break;
823 case MULT_EXPR:
824 res = extract_affine_mul (s, e, space);
825 break;
827 case PLUS_EXPR:
828 case POINTER_PLUS_EXPR:
829 lhs = extract_affine (s, TREE_OPERAND (e, 0), isl_space_copy (space));
830 rhs = extract_affine (s, TREE_OPERAND (e, 1), space);
831 res = isl_pw_aff_add (lhs, rhs);
832 break;
834 case MINUS_EXPR:
835 lhs = extract_affine (s, TREE_OPERAND (e, 0), isl_space_copy (space));
836 rhs = extract_affine (s, TREE_OPERAND (e, 1), space);
837 res = isl_pw_aff_sub (lhs, rhs);
838 break;
840 case NEGATE_EXPR:
841 case BIT_NOT_EXPR:
842 lhs = extract_affine (s, TREE_OPERAND (e, 0), isl_space_copy (space));
843 rhs = extract_affine (s, integer_minus_one_node, space);
844 res = isl_pw_aff_mul (lhs, rhs);
845 break;
847 case SSA_NAME:
848 gcc_assert (-1 != parameter_index_in_region_1 (e, SCOP_REGION (s)));
849 res = extract_affine_name (s, e, space);
850 break;
852 case INTEGER_CST:
853 res = extract_affine_int (e, space);
854 /* No need to wrap a single integer. */
855 return res;
857 CASE_CONVERT:
858 case NON_LVALUE_EXPR:
859 res = extract_affine (s, TREE_OPERAND (e, 0), space);
860 break;
862 default:
863 gcc_unreachable ();
864 break;
867 type = TREE_TYPE (e);
868 if (TYPE_UNSIGNED (type))
869 res = wrap (res, TYPE_PRECISION (type));
871 return res;
874 /* In the context of sese S, scan the expression E and translate it to
875 a linear expression C. When parsing a symbolic multiplication, K
876 represents the constant multiplier of an expression containing
877 parameters. */
879 static void
880 scan_tree_for_params (sese s, tree e)
882 if (e == chrec_dont_know)
883 return;
885 switch (TREE_CODE (e))
887 case POLYNOMIAL_CHREC:
888 scan_tree_for_params (s, CHREC_LEFT (e));
889 break;
891 case MULT_EXPR:
892 if (chrec_contains_symbols (TREE_OPERAND (e, 0)))
893 scan_tree_for_params (s, TREE_OPERAND (e, 0));
894 else
895 scan_tree_for_params (s, TREE_OPERAND (e, 1));
896 break;
898 case PLUS_EXPR:
899 case POINTER_PLUS_EXPR:
900 case MINUS_EXPR:
901 scan_tree_for_params (s, TREE_OPERAND (e, 0));
902 scan_tree_for_params (s, TREE_OPERAND (e, 1));
903 break;
905 case NEGATE_EXPR:
906 case BIT_NOT_EXPR:
907 CASE_CONVERT:
908 case NON_LVALUE_EXPR:
909 scan_tree_for_params (s, TREE_OPERAND (e, 0));
910 break;
912 case SSA_NAME:
913 parameter_index_in_region (e, s);
914 break;
916 case INTEGER_CST:
917 case ADDR_EXPR:
918 break;
920 default:
921 gcc_unreachable ();
922 break;
926 /* Find parameters with respect to REGION in BB. We are looking in memory
927 access functions, conditions and loop bounds. */
929 static void
930 find_params_in_bb (sese region, gimple_bb_p gbb)
932 int i;
933 unsigned j;
934 data_reference_p dr;
935 gimple stmt;
936 loop_p loop = GBB_BB (gbb)->loop_father;
938 /* Find parameters in the access functions of data references. */
939 FOR_EACH_VEC_ELT (GBB_DATA_REFS (gbb), i, dr)
940 for (j = 0; j < DR_NUM_DIMENSIONS (dr); j++)
941 scan_tree_for_params (region, DR_ACCESS_FN (dr, j));
943 /* Find parameters in conditional statements. */
944 FOR_EACH_VEC_ELT (GBB_CONDITIONS (gbb), i, stmt)
946 tree lhs = scalar_evolution_in_region (region, loop,
947 gimple_cond_lhs (stmt));
948 tree rhs = scalar_evolution_in_region (region, loop,
949 gimple_cond_rhs (stmt));
951 scan_tree_for_params (region, lhs);
952 scan_tree_for_params (region, rhs);
956 /* Record the parameters used in the SCOP. A variable is a parameter
957 in a scop if it does not vary during the execution of that scop. */
959 static void
960 find_scop_parameters (scop_p scop)
962 poly_bb_p pbb;
963 unsigned i;
964 sese region = SCOP_REGION (scop);
965 struct loop *loop;
966 int nbp;
968 /* Find the parameters used in the loop bounds. */
969 FOR_EACH_VEC_ELT (SESE_LOOP_NEST (region), i, loop)
971 tree nb_iters = number_of_latch_executions (loop);
973 if (!chrec_contains_symbols (nb_iters))
974 continue;
976 nb_iters = scalar_evolution_in_region (region, loop, nb_iters);
977 scan_tree_for_params (region, nb_iters);
980 /* Find the parameters used in data accesses. */
981 FOR_EACH_VEC_ELT (SCOP_BBS (scop), i, pbb)
982 find_params_in_bb (region, PBB_BLACK_BOX (pbb));
984 nbp = sese_nb_params (region);
985 scop_set_nb_params (scop, nbp);
986 SESE_ADD_PARAMS (region) = false;
989 tree e;
990 isl_space *space = isl_space_set_alloc (scop->ctx, nbp, 0);
992 FOR_EACH_VEC_ELT (SESE_PARAMS (region), i, e)
993 space = isl_space_set_dim_id (space, isl_dim_param, i,
994 isl_id_for_ssa_name (scop, e));
996 scop->context = isl_set_universe (space);
1000 /* Builds the constraint polyhedra for LOOP in SCOP. OUTER_PH gives
1001 the constraints for the surrounding loops. */
1003 static void
1004 build_loop_iteration_domains (scop_p scop, struct loop *loop,
1005 int nb,
1006 isl_set *outer, isl_set **doms)
1008 tree nb_iters = number_of_latch_executions (loop);
1009 sese region = SCOP_REGION (scop);
1011 isl_set *inner = isl_set_copy (outer);
1012 isl_space *space;
1013 isl_constraint *c;
1014 int pos = isl_set_dim (outer, isl_dim_set);
1015 isl_val *v;
1016 mpz_t g;
1018 mpz_init (g);
1020 inner = isl_set_add_dims (inner, isl_dim_set, 1);
1021 space = isl_set_get_space (inner);
1023 /* 0 <= loop_i */
1024 c = isl_inequality_alloc
1025 (isl_local_space_from_space (isl_space_copy (space)));
1026 c = isl_constraint_set_coefficient_si (c, isl_dim_set, pos, 1);
1027 inner = isl_set_add_constraint (inner, c);
1029 /* loop_i <= cst_nb_iters */
1030 if (TREE_CODE (nb_iters) == INTEGER_CST)
1032 c = isl_inequality_alloc
1033 (isl_local_space_from_space (isl_space_copy (space)));
1034 c = isl_constraint_set_coefficient_si (c, isl_dim_set, pos, -1);
1035 tree_int_to_gmp (nb_iters, g);
1036 v = isl_val_int_from_gmp (the_isl_ctx, g);
1037 c = isl_constraint_set_constant_val (c, v);
1038 inner = isl_set_add_constraint (inner, c);
1041 /* loop_i <= expr_nb_iters */
1042 else if (!chrec_contains_undetermined (nb_iters))
1044 widest_int nit;
1045 isl_pw_aff *aff;
1046 isl_set *valid;
1047 isl_local_space *ls;
1048 isl_aff *al;
1049 isl_set *le;
1051 nb_iters = scalar_evolution_in_region (region, loop, nb_iters);
1053 aff = extract_affine (scop, nb_iters, isl_set_get_space (inner));
1054 valid = isl_pw_aff_nonneg_set (isl_pw_aff_copy (aff));
1055 valid = isl_set_project_out (valid, isl_dim_set, 0,
1056 isl_set_dim (valid, isl_dim_set));
1057 scop->context = isl_set_intersect (scop->context, valid);
1059 ls = isl_local_space_from_space (isl_space_copy (space));
1060 al = isl_aff_set_coefficient_si (isl_aff_zero_on_domain (ls),
1061 isl_dim_in, pos, 1);
1062 le = isl_pw_aff_le_set (isl_pw_aff_from_aff (al),
1063 isl_pw_aff_copy (aff));
1064 inner = isl_set_intersect (inner, le);
1066 if (max_stmt_executions (loop, &nit))
1068 /* Insert in the context the constraints from the
1069 estimation of the number of iterations NIT and the
1070 symbolic number of iterations (involving parameter
1071 names) NB_ITERS. First, build the affine expression
1072 "NIT - NB_ITERS" and then say that it is positive,
1073 i.e., NIT approximates NB_ITERS: "NIT >= NB_ITERS". */
1074 isl_pw_aff *approx;
1075 mpz_t g;
1076 isl_set *x;
1077 isl_constraint *c;
1079 mpz_init (g);
1080 wi::to_mpz (nit, g, SIGNED);
1081 mpz_sub_ui (g, g, 1);
1082 approx = extract_affine_gmp (g, isl_set_get_space (inner));
1083 x = isl_pw_aff_ge_set (approx, aff);
1084 x = isl_set_project_out (x, isl_dim_set, 0,
1085 isl_set_dim (x, isl_dim_set));
1086 scop->context = isl_set_intersect (scop->context, x);
1088 c = isl_inequality_alloc
1089 (isl_local_space_from_space (isl_space_copy (space)));
1090 c = isl_constraint_set_coefficient_si (c, isl_dim_set, pos, -1);
1091 v = isl_val_int_from_gmp (the_isl_ctx, g);
1092 mpz_clear (g);
1093 c = isl_constraint_set_constant_val (c, v);
1094 inner = isl_set_add_constraint (inner, c);
1096 else
1097 isl_pw_aff_free (aff);
1099 else
1100 gcc_unreachable ();
1102 if (loop->inner && loop_in_sese_p (loop->inner, region))
1103 build_loop_iteration_domains (scop, loop->inner, nb + 1,
1104 isl_set_copy (inner), doms);
1106 if (nb != 0
1107 && loop->next
1108 && loop_in_sese_p (loop->next, region))
1109 build_loop_iteration_domains (scop, loop->next, nb,
1110 isl_set_copy (outer), doms);
1112 doms[loop->num] = inner;
1114 isl_set_free (outer);
1115 isl_space_free (space);
1116 mpz_clear (g);
1119 /* Returns a linear expression for tree T evaluated in PBB. */
1121 static isl_pw_aff *
1122 create_pw_aff_from_tree (poly_bb_p pbb, tree t)
1124 scop_p scop = PBB_SCOP (pbb);
1126 t = scalar_evolution_in_region (SCOP_REGION (scop), pbb_loop (pbb), t);
1127 gcc_assert (!automatically_generated_chrec_p (t));
1129 return extract_affine (scop, t, isl_set_get_space (pbb->domain));
1132 /* Add conditional statement STMT to pbb. CODE is used as the comparison
1133 operator. This allows us to invert the condition or to handle
1134 inequalities. */
1136 static void
1137 add_condition_to_pbb (poly_bb_p pbb, gcond *stmt, enum tree_code code)
1139 isl_pw_aff *lhs = create_pw_aff_from_tree (pbb, gimple_cond_lhs (stmt));
1140 isl_pw_aff *rhs = create_pw_aff_from_tree (pbb, gimple_cond_rhs (stmt));
1141 isl_set *cond;
1143 switch (code)
1145 case LT_EXPR:
1146 cond = isl_pw_aff_lt_set (lhs, rhs);
1147 break;
1149 case GT_EXPR:
1150 cond = isl_pw_aff_gt_set (lhs, rhs);
1151 break;
1153 case LE_EXPR:
1154 cond = isl_pw_aff_le_set (lhs, rhs);
1155 break;
1157 case GE_EXPR:
1158 cond = isl_pw_aff_ge_set (lhs, rhs);
1159 break;
1161 case EQ_EXPR:
1162 cond = isl_pw_aff_eq_set (lhs, rhs);
1163 break;
1165 case NE_EXPR:
1166 cond = isl_pw_aff_ne_set (lhs, rhs);
1167 break;
1169 default:
1170 isl_pw_aff_free (lhs);
1171 isl_pw_aff_free (rhs);
1172 return;
1175 cond = isl_set_coalesce (cond);
1176 cond = isl_set_set_tuple_id (cond, isl_set_get_tuple_id (pbb->domain));
1177 pbb->domain = isl_set_intersect (pbb->domain, cond);
1180 /* Add conditions to the domain of PBB. */
1182 static void
1183 add_conditions_to_domain (poly_bb_p pbb)
1185 unsigned int i;
1186 gimple stmt;
1187 gimple_bb_p gbb = PBB_BLACK_BOX (pbb);
1189 if (GBB_CONDITIONS (gbb).is_empty ())
1190 return;
1192 FOR_EACH_VEC_ELT (GBB_CONDITIONS (gbb), i, stmt)
1193 switch (gimple_code (stmt))
1195 case GIMPLE_COND:
1197 gcond *cond_stmt = as_a <gcond *> (stmt);
1198 enum tree_code code = gimple_cond_code (cond_stmt);
1200 /* The conditions for ELSE-branches are inverted. */
1201 if (!GBB_CONDITION_CASES (gbb)[i])
1202 code = invert_tree_comparison (code, false);
1204 add_condition_to_pbb (pbb, cond_stmt, code);
1205 break;
1208 case GIMPLE_SWITCH:
1209 /* Switch statements are not supported right now - fall through. */
1211 default:
1212 gcc_unreachable ();
1213 break;
1217 /* Traverses all the GBBs of the SCOP and add their constraints to the
1218 iteration domains. */
1220 static void
1221 add_conditions_to_constraints (scop_p scop)
1223 int i;
1224 poly_bb_p pbb;
1226 FOR_EACH_VEC_ELT (SCOP_BBS (scop), i, pbb)
1227 add_conditions_to_domain (pbb);
1230 /* Returns a COND_EXPR statement when BB has a single predecessor, the
1231 edge between BB and its predecessor is not a loop exit edge, and
1232 the last statement of the single predecessor is a COND_EXPR. */
1234 static gcond *
1235 single_pred_cond_non_loop_exit (basic_block bb)
1237 if (single_pred_p (bb))
1239 edge e = single_pred_edge (bb);
1240 basic_block pred = e->src;
1241 gimple stmt;
1243 if (loop_depth (pred->loop_father) > loop_depth (bb->loop_father))
1244 return NULL;
1246 stmt = last_stmt (pred);
1248 if (stmt && gimple_code (stmt) == GIMPLE_COND)
1249 return as_a <gcond *> (stmt);
1252 return NULL;
1255 class sese_dom_walker : public dom_walker
1257 public:
1258 sese_dom_walker (cdi_direction, sese);
1260 virtual void before_dom_children (basic_block);
1261 virtual void after_dom_children (basic_block);
1263 private:
1264 auto_vec<gimple, 3> m_conditions, m_cases;
1265 sese m_region;
1268 sese_dom_walker::sese_dom_walker (cdi_direction direction, sese region)
1269 : dom_walker (direction), m_region (region)
1273 /* Call-back for dom_walk executed before visiting the dominated
1274 blocks. */
1276 void
1277 sese_dom_walker::before_dom_children (basic_block bb)
1279 gimple_bb_p gbb;
1280 gcond *stmt;
1282 if (!bb_in_sese_p (bb, m_region))
1283 return;
1285 stmt = single_pred_cond_non_loop_exit (bb);
1287 if (stmt)
1289 edge e = single_pred_edge (bb);
1291 m_conditions.safe_push (stmt);
1293 if (e->flags & EDGE_TRUE_VALUE)
1294 m_cases.safe_push (stmt);
1295 else
1296 m_cases.safe_push (NULL);
1299 gbb = gbb_from_bb (bb);
1301 if (gbb)
1303 GBB_CONDITIONS (gbb) = m_conditions.copy ();
1304 GBB_CONDITION_CASES (gbb) = m_cases.copy ();
1308 /* Call-back for dom_walk executed after visiting the dominated
1309 blocks. */
1311 void
1312 sese_dom_walker::after_dom_children (basic_block bb)
1314 if (!bb_in_sese_p (bb, m_region))
1315 return;
1317 if (single_pred_cond_non_loop_exit (bb))
1319 m_conditions.pop ();
1320 m_cases.pop ();
1324 /* Add constraints on the possible values of parameter P from the type
1325 of P. */
1327 static void
1328 add_param_constraints (scop_p scop, graphite_dim_t p)
1330 tree parameter = SESE_PARAMS (SCOP_REGION (scop))[p];
1331 tree type = TREE_TYPE (parameter);
1332 tree lb = NULL_TREE;
1333 tree ub = NULL_TREE;
1335 if (POINTER_TYPE_P (type) || !TYPE_MIN_VALUE (type))
1336 lb = lower_bound_in_type (type, type);
1337 else
1338 lb = TYPE_MIN_VALUE (type);
1340 if (POINTER_TYPE_P (type) || !TYPE_MAX_VALUE (type))
1341 ub = upper_bound_in_type (type, type);
1342 else
1343 ub = TYPE_MAX_VALUE (type);
1345 if (lb)
1347 isl_space *space = isl_set_get_space (scop->context);
1348 isl_constraint *c;
1349 mpz_t g;
1350 isl_val *v;
1352 c = isl_inequality_alloc (isl_local_space_from_space (space));
1353 mpz_init (g);
1354 tree_int_to_gmp (lb, g);
1355 v = isl_val_int_from_gmp (the_isl_ctx, g);
1356 v = isl_val_neg (v);
1357 mpz_clear (g);
1358 c = isl_constraint_set_constant_val (c, v);
1359 c = isl_constraint_set_coefficient_si (c, isl_dim_param, p, 1);
1361 scop->context = isl_set_add_constraint (scop->context, c);
1364 if (ub)
1366 isl_space *space = isl_set_get_space (scop->context);
1367 isl_constraint *c;
1368 mpz_t g;
1369 isl_val *v;
1371 c = isl_inequality_alloc (isl_local_space_from_space (space));
1373 mpz_init (g);
1374 tree_int_to_gmp (ub, g);
1375 v = isl_val_int_from_gmp (the_isl_ctx, g);
1376 mpz_clear (g);
1377 c = isl_constraint_set_constant_val (c, v);
1378 c = isl_constraint_set_coefficient_si (c, isl_dim_param, p, -1);
1380 scop->context = isl_set_add_constraint (scop->context, c);
1384 /* Build the context of the SCOP. The context usually contains extra
1385 constraints that are added to the iteration domains that constrain
1386 some parameters. */
1388 static void
1389 build_scop_context (scop_p scop)
1391 graphite_dim_t p, n = scop_nb_params (scop);
1393 for (p = 0; p < n; p++)
1394 add_param_constraints (scop, p);
1397 /* Build the iteration domains: the loops belonging to the current
1398 SCOP, and that vary for the execution of the current basic block.
1399 Returns false if there is no loop in SCOP. */
1401 static void
1402 build_scop_iteration_domain (scop_p scop)
1404 struct loop *loop;
1405 sese region = SCOP_REGION (scop);
1406 int i;
1407 poly_bb_p pbb;
1408 int nb_loops = number_of_loops (cfun);
1409 isl_set **doms = XCNEWVEC (isl_set *, nb_loops);
1411 FOR_EACH_VEC_ELT (SESE_LOOP_NEST (region), i, loop)
1412 if (!loop_in_sese_p (loop_outer (loop), region))
1413 build_loop_iteration_domains (scop, loop, 0,
1414 isl_set_copy (scop->context), doms);
1416 FOR_EACH_VEC_ELT (SCOP_BBS (scop), i, pbb)
1418 loop = pbb_loop (pbb);
1420 if (doms[loop->num])
1421 pbb->domain = isl_set_copy (doms[loop->num]);
1422 else
1423 pbb->domain = isl_set_copy (scop->context);
1425 pbb->domain = isl_set_set_tuple_id (pbb->domain,
1426 isl_id_for_pbb (scop, pbb));
1429 for (i = 0; i < nb_loops; i++)
1430 if (doms[i])
1431 isl_set_free (doms[i]);
1433 free (doms);
1436 /* Add a constrain to the ACCESSES polyhedron for the alias set of
1437 data reference DR. ACCESSP_NB_DIMS is the dimension of the
1438 ACCESSES polyhedron, DOM_NB_DIMS is the dimension of the iteration
1439 domain. */
1441 static isl_map *
1442 pdr_add_alias_set (isl_map *acc, data_reference_p dr)
1444 isl_constraint *c;
1445 int alias_set_num = 0;
1446 base_alias_pair *bap = (base_alias_pair *)(dr->aux);
1448 if (bap && bap->alias_set)
1449 alias_set_num = *(bap->alias_set);
1451 c = isl_equality_alloc
1452 (isl_local_space_from_space (isl_map_get_space (acc)));
1453 c = isl_constraint_set_constant_si (c, -alias_set_num);
1454 c = isl_constraint_set_coefficient_si (c, isl_dim_out, 0, 1);
1456 return isl_map_add_constraint (acc, c);
1459 /* Assign the affine expression INDEX to the output dimension POS of
1460 MAP and return the result. */
1462 static isl_map *
1463 set_index (isl_map *map, int pos, isl_pw_aff *index)
1465 isl_map *index_map;
1466 int len = isl_map_dim (map, isl_dim_out);
1467 isl_id *id;
1469 index_map = isl_map_from_pw_aff (index);
1470 index_map = isl_map_insert_dims (index_map, isl_dim_out, 0, pos);
1471 index_map = isl_map_add_dims (index_map, isl_dim_out, len - pos - 1);
1473 id = isl_map_get_tuple_id (map, isl_dim_out);
1474 index_map = isl_map_set_tuple_id (index_map, isl_dim_out, id);
1475 id = isl_map_get_tuple_id (map, isl_dim_in);
1476 index_map = isl_map_set_tuple_id (index_map, isl_dim_in, id);
1478 return isl_map_intersect (map, index_map);
1481 /* Add to ACCESSES polyhedron equalities defining the access functions
1482 to the memory. ACCESSP_NB_DIMS is the dimension of the ACCESSES
1483 polyhedron, DOM_NB_DIMS is the dimension of the iteration domain.
1484 PBB is the poly_bb_p that contains the data reference DR. */
1486 static isl_map *
1487 pdr_add_memory_accesses (isl_map *acc, data_reference_p dr, poly_bb_p pbb)
1489 int i, nb_subscripts = DR_NUM_DIMENSIONS (dr);
1490 scop_p scop = PBB_SCOP (pbb);
1492 for (i = 0; i < nb_subscripts; i++)
1494 isl_pw_aff *aff;
1495 tree afn = DR_ACCESS_FN (dr, nb_subscripts - 1 - i);
1497 aff = extract_affine (scop, afn,
1498 isl_space_domain (isl_map_get_space (acc)));
1499 acc = set_index (acc, i + 1, aff);
1502 return acc;
1505 /* Add constrains representing the size of the accessed data to the
1506 ACCESSES polyhedron. ACCESSP_NB_DIMS is the dimension of the
1507 ACCESSES polyhedron, DOM_NB_DIMS is the dimension of the iteration
1508 domain. */
1510 static isl_set *
1511 pdr_add_data_dimensions (isl_set *extent, scop_p scop, data_reference_p dr)
1513 tree ref = DR_REF (dr);
1514 int i, nb_subscripts = DR_NUM_DIMENSIONS (dr);
1516 for (i = nb_subscripts - 1; i >= 0; i--, ref = TREE_OPERAND (ref, 0))
1518 tree low, high;
1520 if (TREE_CODE (ref) != ARRAY_REF)
1521 break;
1523 low = array_ref_low_bound (ref);
1524 high = array_ref_up_bound (ref);
1526 /* XXX The PPL code dealt separately with
1527 subscript - low >= 0 and high - subscript >= 0 in case one of
1528 the two bounds isn't known. Do the same here? */
1530 if (tree_fits_shwi_p (low)
1531 && high
1532 && tree_fits_shwi_p (high)
1533 /* 1-element arrays at end of structures may extend over
1534 their declared size. */
1535 && !(array_at_struct_end_p (ref)
1536 && operand_equal_p (low, high, 0)))
1538 isl_id *id;
1539 isl_aff *aff;
1540 isl_set *univ, *lbs, *ubs;
1541 isl_pw_aff *index;
1542 isl_space *space;
1543 isl_set *valid;
1544 isl_pw_aff *lb = extract_affine_int (low, isl_set_get_space (extent));
1545 isl_pw_aff *ub = extract_affine_int (high, isl_set_get_space (extent));
1547 /* high >= 0 */
1548 valid = isl_pw_aff_nonneg_set (isl_pw_aff_copy (ub));
1549 valid = isl_set_project_out (valid, isl_dim_set, 0,
1550 isl_set_dim (valid, isl_dim_set));
1551 scop->context = isl_set_intersect (scop->context, valid);
1553 space = isl_set_get_space (extent);
1554 aff = isl_aff_zero_on_domain (isl_local_space_from_space (space));
1555 aff = isl_aff_add_coefficient_si (aff, isl_dim_in, i + 1, 1);
1556 univ = isl_set_universe (isl_space_domain (isl_aff_get_space (aff)));
1557 index = isl_pw_aff_alloc (univ, aff);
1559 id = isl_set_get_tuple_id (extent);
1560 lb = isl_pw_aff_set_tuple_id (lb, isl_dim_in, isl_id_copy (id));
1561 ub = isl_pw_aff_set_tuple_id (ub, isl_dim_in, id);
1563 /* low <= sub_i <= high */
1564 lbs = isl_pw_aff_ge_set (isl_pw_aff_copy (index), lb);
1565 ubs = isl_pw_aff_le_set (index, ub);
1566 extent = isl_set_intersect (extent, lbs);
1567 extent = isl_set_intersect (extent, ubs);
1571 return extent;
1574 /* Build data accesses for DR in PBB. */
1576 static void
1577 build_poly_dr (data_reference_p dr, poly_bb_p pbb)
1579 int dr_base_object_set;
1580 isl_map *acc;
1581 isl_set *extent;
1582 scop_p scop = PBB_SCOP (pbb);
1585 isl_space *dc = isl_set_get_space (pbb->domain);
1586 int nb_out = 1 + DR_NUM_DIMENSIONS (dr);
1587 isl_space *space = isl_space_add_dims (isl_space_from_domain (dc),
1588 isl_dim_out, nb_out);
1590 acc = isl_map_universe (space);
1591 acc = isl_map_set_tuple_id (acc, isl_dim_out, isl_id_for_dr (scop, dr));
1594 acc = pdr_add_alias_set (acc, dr);
1595 acc = pdr_add_memory_accesses (acc, dr, pbb);
1598 isl_id *id = isl_id_for_dr (scop, dr);
1599 int nb = 1 + DR_NUM_DIMENSIONS (dr);
1600 isl_space *space = isl_space_set_alloc (scop->ctx, 0, nb);
1601 int alias_set_num = 0;
1602 base_alias_pair *bap = (base_alias_pair *)(dr->aux);
1604 if (bap && bap->alias_set)
1605 alias_set_num = *(bap->alias_set);
1607 space = isl_space_set_tuple_id (space, isl_dim_set, id);
1608 extent = isl_set_nat_universe (space);
1609 extent = isl_set_fix_si (extent, isl_dim_set, 0, alias_set_num);
1610 extent = pdr_add_data_dimensions (extent, scop, dr);
1613 gcc_assert (dr->aux);
1614 dr_base_object_set = ((base_alias_pair *)(dr->aux))->base_obj_set;
1616 new_poly_dr (pbb, dr_base_object_set,
1617 DR_IS_READ (dr) ? PDR_READ : PDR_WRITE,
1618 dr, DR_NUM_DIMENSIONS (dr), acc, extent);
1621 /* Write to FILE the alias graph of data references in DIMACS format. */
1623 static inline bool
1624 write_alias_graph_to_ascii_dimacs (FILE *file, char *comment,
1625 vec<data_reference_p> drs)
1627 int num_vertex = drs.length ();
1628 int edge_num = 0;
1629 data_reference_p dr1, dr2;
1630 int i, j;
1632 if (num_vertex == 0)
1633 return true;
1635 FOR_EACH_VEC_ELT (drs, i, dr1)
1636 for (j = i + 1; drs.iterate (j, &dr2); j++)
1637 if (dr_may_alias_p (dr1, dr2, true))
1638 edge_num++;
1640 fprintf (file, "$\n");
1642 if (comment)
1643 fprintf (file, "c %s\n", comment);
1645 fprintf (file, "p edge %d %d\n", num_vertex, edge_num);
1647 FOR_EACH_VEC_ELT (drs, i, dr1)
1648 for (j = i + 1; drs.iterate (j, &dr2); j++)
1649 if (dr_may_alias_p (dr1, dr2, true))
1650 fprintf (file, "e %d %d\n", i + 1, j + 1);
1652 return true;
1655 /* Write to FILE the alias graph of data references in DOT format. */
1657 static inline bool
1658 write_alias_graph_to_ascii_dot (FILE *file, char *comment,
1659 vec<data_reference_p> drs)
1661 int num_vertex = drs.length ();
1662 data_reference_p dr1, dr2;
1663 int i, j;
1665 if (num_vertex == 0)
1666 return true;
1668 fprintf (file, "$\n");
1670 if (comment)
1671 fprintf (file, "c %s\n", comment);
1673 /* First print all the vertices. */
1674 FOR_EACH_VEC_ELT (drs, i, dr1)
1675 fprintf (file, "n%d;\n", i);
1677 FOR_EACH_VEC_ELT (drs, i, dr1)
1678 for (j = i + 1; drs.iterate (j, &dr2); j++)
1679 if (dr_may_alias_p (dr1, dr2, true))
1680 fprintf (file, "n%d n%d\n", i, j);
1682 return true;
1685 /* Write to FILE the alias graph of data references in ECC format. */
1687 static inline bool
1688 write_alias_graph_to_ascii_ecc (FILE *file, char *comment,
1689 vec<data_reference_p> drs)
1691 int num_vertex = drs.length ();
1692 data_reference_p dr1, dr2;
1693 int i, j;
1695 if (num_vertex == 0)
1696 return true;
1698 fprintf (file, "$\n");
1700 if (comment)
1701 fprintf (file, "c %s\n", comment);
1703 FOR_EACH_VEC_ELT (drs, i, dr1)
1704 for (j = i + 1; drs.iterate (j, &dr2); j++)
1705 if (dr_may_alias_p (dr1, dr2, true))
1706 fprintf (file, "%d %d\n", i, j);
1708 return true;
1711 /* Check if DR1 and DR2 are in the same object set. */
1713 static bool
1714 dr_same_base_object_p (const struct data_reference *dr1,
1715 const struct data_reference *dr2)
1717 return operand_equal_p (DR_BASE_OBJECT (dr1), DR_BASE_OBJECT (dr2), 0);
1720 /* Uses DFS component number as representative of alias-sets. Also tests for
1721 optimality by verifying if every connected component is a clique. Returns
1722 true (1) if the above test is true, and false (0) otherwise. */
1724 static int
1725 build_alias_set_optimal_p (vec<data_reference_p> drs)
1727 int num_vertices = drs.length ();
1728 struct graph *g = new_graph (num_vertices);
1729 data_reference_p dr1, dr2;
1730 int i, j;
1731 int num_connected_components;
1732 int v_indx1, v_indx2, num_vertices_in_component;
1733 int *all_vertices;
1734 int *vertices;
1735 struct graph_edge *e;
1736 int this_component_is_clique;
1737 int all_components_are_cliques = 1;
1739 FOR_EACH_VEC_ELT (drs, i, dr1)
1740 for (j = i+1; drs.iterate (j, &dr2); j++)
1741 if (dr_may_alias_p (dr1, dr2, true))
1743 add_edge (g, i, j);
1744 add_edge (g, j, i);
1747 all_vertices = XNEWVEC (int, num_vertices);
1748 vertices = XNEWVEC (int, num_vertices);
1749 for (i = 0; i < num_vertices; i++)
1750 all_vertices[i] = i;
1752 num_connected_components = graphds_dfs (g, all_vertices, num_vertices,
1753 NULL, true, NULL);
1754 for (i = 0; i < g->n_vertices; i++)
1756 data_reference_p dr = drs[i];
1757 base_alias_pair *bap;
1759 gcc_assert (dr->aux);
1760 bap = (base_alias_pair *)(dr->aux);
1762 bap->alias_set = XNEW (int);
1763 *(bap->alias_set) = g->vertices[i].component + 1;
1766 /* Verify if the DFS numbering results in optimal solution. */
1767 for (i = 0; i < num_connected_components; i++)
1769 num_vertices_in_component = 0;
1770 /* Get all vertices whose DFS component number is the same as i. */
1771 for (j = 0; j < num_vertices; j++)
1772 if (g->vertices[j].component == i)
1773 vertices[num_vertices_in_component++] = j;
1775 /* Now test if the vertices in 'vertices' form a clique, by testing
1776 for edges among each pair. */
1777 this_component_is_clique = 1;
1778 for (v_indx1 = 0; v_indx1 < num_vertices_in_component; v_indx1++)
1780 for (v_indx2 = v_indx1+1; v_indx2 < num_vertices_in_component; v_indx2++)
1782 /* Check if the two vertices are connected by iterating
1783 through all the edges which have one of these are source. */
1784 e = g->vertices[vertices[v_indx2]].pred;
1785 while (e)
1787 if (e->src == vertices[v_indx1])
1788 break;
1789 e = e->pred_next;
1791 if (!e)
1793 this_component_is_clique = 0;
1794 break;
1797 if (!this_component_is_clique)
1798 all_components_are_cliques = 0;
1802 free (all_vertices);
1803 free (vertices);
1804 free_graph (g);
1805 return all_components_are_cliques;
1808 /* Group each data reference in DRS with its base object set num. */
1810 static void
1811 build_base_obj_set_for_drs (vec<data_reference_p> drs)
1813 int num_vertex = drs.length ();
1814 struct graph *g = new_graph (num_vertex);
1815 data_reference_p dr1, dr2;
1816 int i, j;
1817 int *queue;
1819 FOR_EACH_VEC_ELT (drs, i, dr1)
1820 for (j = i + 1; drs.iterate (j, &dr2); j++)
1821 if (dr_same_base_object_p (dr1, dr2))
1823 add_edge (g, i, j);
1824 add_edge (g, j, i);
1827 queue = XNEWVEC (int, num_vertex);
1828 for (i = 0; i < num_vertex; i++)
1829 queue[i] = i;
1831 graphds_dfs (g, queue, num_vertex, NULL, true, NULL);
1833 for (i = 0; i < g->n_vertices; i++)
1835 data_reference_p dr = drs[i];
1836 base_alias_pair *bap;
1838 gcc_assert (dr->aux);
1839 bap = (base_alias_pair *)(dr->aux);
1841 bap->base_obj_set = g->vertices[i].component + 1;
1844 free (queue);
1845 free_graph (g);
1848 /* Build the data references for PBB. */
1850 static void
1851 build_pbb_drs (poly_bb_p pbb)
1853 int j;
1854 data_reference_p dr;
1855 vec<data_reference_p> gbb_drs = GBB_DATA_REFS (PBB_BLACK_BOX (pbb));
1857 FOR_EACH_VEC_ELT (gbb_drs, j, dr)
1858 build_poly_dr (dr, pbb);
1861 /* Dump to file the alias graphs for the data references in DRS. */
1863 static void
1864 dump_alias_graphs (vec<data_reference_p> drs)
1866 char comment[100];
1867 FILE *file_dimacs, *file_ecc, *file_dot;
1869 file_dimacs = fopen ("/tmp/dr_alias_graph_dimacs", "ab");
1870 if (file_dimacs)
1872 snprintf (comment, sizeof (comment), "%s %s", main_input_filename,
1873 current_function_name ());
1874 write_alias_graph_to_ascii_dimacs (file_dimacs, comment, drs);
1875 fclose (file_dimacs);
1878 file_ecc = fopen ("/tmp/dr_alias_graph_ecc", "ab");
1879 if (file_ecc)
1881 snprintf (comment, sizeof (comment), "%s %s", main_input_filename,
1882 current_function_name ());
1883 write_alias_graph_to_ascii_ecc (file_ecc, comment, drs);
1884 fclose (file_ecc);
1887 file_dot = fopen ("/tmp/dr_alias_graph_dot", "ab");
1888 if (file_dot)
1890 snprintf (comment, sizeof (comment), "%s %s", main_input_filename,
1891 current_function_name ());
1892 write_alias_graph_to_ascii_dot (file_dot, comment, drs);
1893 fclose (file_dot);
1897 /* Build data references in SCOP. */
1899 static void
1900 build_scop_drs (scop_p scop)
1902 int i, j;
1903 poly_bb_p pbb;
1904 data_reference_p dr;
1905 auto_vec<data_reference_p, 3> drs;
1907 /* Remove all the PBBs that do not have data references: these basic
1908 blocks are not handled in the polyhedral representation. */
1909 for (i = 0; SCOP_BBS (scop).iterate (i, &pbb); i++)
1910 if (GBB_DATA_REFS (PBB_BLACK_BOX (pbb)).is_empty ())
1912 free_gimple_bb (PBB_BLACK_BOX (pbb));
1913 free_poly_bb (pbb);
1914 SCOP_BBS (scop).ordered_remove (i);
1915 i--;
1918 FOR_EACH_VEC_ELT (SCOP_BBS (scop), i, pbb)
1919 for (j = 0; GBB_DATA_REFS (PBB_BLACK_BOX (pbb)).iterate (j, &dr); j++)
1920 drs.safe_push (dr);
1922 FOR_EACH_VEC_ELT (drs, i, dr)
1923 dr->aux = XNEW (base_alias_pair);
1925 if (!build_alias_set_optimal_p (drs))
1927 /* TODO: Add support when building alias set is not optimal. */
1931 build_base_obj_set_for_drs (drs);
1933 /* When debugging, enable the following code. This cannot be used
1934 in production compilers. */
1935 if (0)
1936 dump_alias_graphs (drs);
1938 drs.release ();
1940 FOR_EACH_VEC_ELT (SCOP_BBS (scop), i, pbb)
1941 build_pbb_drs (pbb);
1944 /* Return a gsi at the position of the phi node STMT. */
1946 static gphi_iterator
1947 gsi_for_phi_node (gphi *stmt)
1949 gphi_iterator psi;
1950 basic_block bb = gimple_bb (stmt);
1952 for (psi = gsi_start_phis (bb); !gsi_end_p (psi); gsi_next (&psi))
1953 if (stmt == psi.phi ())
1954 return psi;
1956 gcc_unreachable ();
1957 return psi;
1960 /* Analyze all the data references of STMTS and add them to the
1961 GBB_DATA_REFS vector of BB. */
1963 static void
1964 analyze_drs_in_stmts (scop_p scop, basic_block bb, vec<gimple> stmts)
1966 loop_p nest;
1967 gimple_bb_p gbb;
1968 gimple stmt;
1969 int i;
1970 sese region = SCOP_REGION (scop);
1972 if (!bb_in_sese_p (bb, region))
1973 return;
1975 nest = outermost_loop_in_sese_1 (region, bb);
1976 gbb = gbb_from_bb (bb);
1978 FOR_EACH_VEC_ELT (stmts, i, stmt)
1980 loop_p loop;
1982 if (is_gimple_debug (stmt))
1983 continue;
1985 loop = loop_containing_stmt (stmt);
1986 if (!loop_in_sese_p (loop, region))
1987 loop = nest;
1989 graphite_find_data_references_in_stmt (nest, loop, stmt,
1990 &GBB_DATA_REFS (gbb));
1994 /* Insert STMT at the end of the STMTS sequence and then insert the
1995 statements from STMTS at INSERT_GSI and call analyze_drs_in_stmts
1996 on STMTS. */
1998 static void
1999 insert_stmts (scop_p scop, gimple stmt, gimple_seq stmts,
2000 gimple_stmt_iterator insert_gsi)
2002 gimple_stmt_iterator gsi;
2003 auto_vec<gimple, 3> x;
2005 gimple_seq_add_stmt (&stmts, stmt);
2006 for (gsi = gsi_start (stmts); !gsi_end_p (gsi); gsi_next (&gsi))
2007 x.safe_push (gsi_stmt (gsi));
2009 gsi_insert_seq_before (&insert_gsi, stmts, GSI_SAME_STMT);
2010 analyze_drs_in_stmts (scop, gsi_bb (insert_gsi), x);
2013 /* Insert the assignment "RES := EXPR" just after AFTER_STMT. */
2015 static void
2016 insert_out_of_ssa_copy (scop_p scop, tree res, tree expr, gimple after_stmt)
2018 gimple_seq stmts;
2019 gimple_stmt_iterator gsi;
2020 tree var = force_gimple_operand (expr, &stmts, true, NULL_TREE);
2021 gassign *stmt = gimple_build_assign (unshare_expr (res), var);
2022 auto_vec<gimple, 3> x;
2024 gimple_seq_add_stmt (&stmts, stmt);
2025 for (gsi = gsi_start (stmts); !gsi_end_p (gsi); gsi_next (&gsi))
2026 x.safe_push (gsi_stmt (gsi));
2028 if (gimple_code (after_stmt) == GIMPLE_PHI)
2030 gsi = gsi_after_labels (gimple_bb (after_stmt));
2031 gsi_insert_seq_before (&gsi, stmts, GSI_NEW_STMT);
2033 else
2035 gsi = gsi_for_stmt (after_stmt);
2036 gsi_insert_seq_after (&gsi, stmts, GSI_NEW_STMT);
2039 analyze_drs_in_stmts (scop, gimple_bb (after_stmt), x);
2042 /* Creates a poly_bb_p for basic_block BB from the existing PBB. */
2044 static void
2045 new_pbb_from_pbb (scop_p scop, poly_bb_p pbb, basic_block bb)
2047 vec<data_reference_p> drs;
2048 drs.create (3);
2049 gimple_bb_p gbb = PBB_BLACK_BOX (pbb);
2050 gimple_bb_p gbb1 = new_gimple_bb (bb, drs);
2051 poly_bb_p pbb1 = new_poly_bb (scop, gbb1);
2052 int index, n = SCOP_BBS (scop).length ();
2054 /* The INDEX of PBB in SCOP_BBS. */
2055 for (index = 0; index < n; index++)
2056 if (SCOP_BBS (scop)[index] == pbb)
2057 break;
2059 pbb1->domain = isl_set_copy (pbb->domain);
2060 pbb1->domain = isl_set_set_tuple_id (pbb1->domain,
2061 isl_id_for_pbb (scop, pbb1));
2063 GBB_PBB (gbb1) = pbb1;
2064 GBB_CONDITIONS (gbb1) = GBB_CONDITIONS (gbb).copy ();
2065 GBB_CONDITION_CASES (gbb1) = GBB_CONDITION_CASES (gbb).copy ();
2066 SCOP_BBS (scop).safe_insert (index + 1, pbb1);
2069 /* Insert on edge E the assignment "RES := EXPR". */
2071 static void
2072 insert_out_of_ssa_copy_on_edge (scop_p scop, edge e, tree res, tree expr)
2074 gimple_stmt_iterator gsi;
2075 gimple_seq stmts = NULL;
2076 tree var = force_gimple_operand (expr, &stmts, true, NULL_TREE);
2077 gimple stmt = gimple_build_assign (unshare_expr (res), var);
2078 basic_block bb;
2079 auto_vec<gimple, 3> x;
2081 gimple_seq_add_stmt (&stmts, stmt);
2082 for (gsi = gsi_start (stmts); !gsi_end_p (gsi); gsi_next (&gsi))
2083 x.safe_push (gsi_stmt (gsi));
2085 gsi_insert_seq_on_edge (e, stmts);
2086 gsi_commit_edge_inserts ();
2087 bb = gimple_bb (stmt);
2089 if (!bb_in_sese_p (bb, SCOP_REGION (scop)))
2090 return;
2092 if (!gbb_from_bb (bb))
2093 new_pbb_from_pbb (scop, pbb_from_bb (e->src), bb);
2095 analyze_drs_in_stmts (scop, bb, x);
2098 /* Creates a zero dimension array of the same type as VAR. */
2100 static tree
2101 create_zero_dim_array (tree var, const char *base_name)
2103 tree index_type = build_index_type (integer_zero_node);
2104 tree elt_type = TREE_TYPE (var);
2105 tree array_type = build_array_type (elt_type, index_type);
2106 tree base = create_tmp_var (array_type, base_name);
2108 return build4 (ARRAY_REF, elt_type, base, integer_zero_node, NULL_TREE,
2109 NULL_TREE);
2112 /* Returns true when PHI is a loop close phi node. */
2114 static bool
2115 scalar_close_phi_node_p (gimple phi)
2117 if (gimple_code (phi) != GIMPLE_PHI
2118 || virtual_operand_p (gimple_phi_result (phi)))
2119 return false;
2121 /* Note that loop close phi nodes should have a single argument
2122 because we translated the representation into a canonical form
2123 before Graphite: see canonicalize_loop_closed_ssa_form. */
2124 return (gimple_phi_num_args (phi) == 1);
2127 /* For a definition DEF in REGION, propagates the expression EXPR in
2128 all the uses of DEF outside REGION. */
2130 static void
2131 propagate_expr_outside_region (tree def, tree expr, sese region)
2133 imm_use_iterator imm_iter;
2134 gimple use_stmt;
2135 gimple_seq stmts;
2136 bool replaced_once = false;
2138 gcc_assert (TREE_CODE (def) == SSA_NAME);
2140 expr = force_gimple_operand (unshare_expr (expr), &stmts, true,
2141 NULL_TREE);
2143 FOR_EACH_IMM_USE_STMT (use_stmt, imm_iter, def)
2144 if (!is_gimple_debug (use_stmt)
2145 && !bb_in_sese_p (gimple_bb (use_stmt), region))
2147 ssa_op_iter iter;
2148 use_operand_p use_p;
2150 FOR_EACH_PHI_OR_STMT_USE (use_p, use_stmt, iter, SSA_OP_ALL_USES)
2151 if (operand_equal_p (def, USE_FROM_PTR (use_p), 0)
2152 && (replaced_once = true))
2153 replace_exp (use_p, expr);
2155 update_stmt (use_stmt);
2158 if (replaced_once)
2160 gsi_insert_seq_on_edge (SESE_ENTRY (region), stmts);
2161 gsi_commit_edge_inserts ();
2165 /* Rewrite out of SSA the reduction phi node at PSI by creating a zero
2166 dimension array for it. */
2168 static void
2169 rewrite_close_phi_out_of_ssa (scop_p scop, gimple_stmt_iterator *psi)
2171 sese region = SCOP_REGION (scop);
2172 gimple phi = gsi_stmt (*psi);
2173 tree res = gimple_phi_result (phi);
2174 basic_block bb = gimple_bb (phi);
2175 gimple_stmt_iterator gsi = gsi_after_labels (bb);
2176 tree arg = gimple_phi_arg_def (phi, 0);
2177 gimple stmt;
2179 /* Note that loop close phi nodes should have a single argument
2180 because we translated the representation into a canonical form
2181 before Graphite: see canonicalize_loop_closed_ssa_form. */
2182 gcc_assert (gimple_phi_num_args (phi) == 1);
2184 /* The phi node can be a non close phi node, when its argument is
2185 invariant, or a default definition. */
2186 if (is_gimple_min_invariant (arg)
2187 || SSA_NAME_IS_DEFAULT_DEF (arg))
2189 propagate_expr_outside_region (res, arg, region);
2190 gsi_next (psi);
2191 return;
2194 else if (gimple_bb (SSA_NAME_DEF_STMT (arg))->loop_father == bb->loop_father)
2196 propagate_expr_outside_region (res, arg, region);
2197 stmt = gimple_build_assign (res, arg);
2198 remove_phi_node (psi, false);
2199 gsi_insert_before (&gsi, stmt, GSI_NEW_STMT);
2200 return;
2203 /* If res is scev analyzable and is not a scalar value, it is safe
2204 to ignore the close phi node: it will be code generated in the
2205 out of Graphite pass. */
2206 else if (scev_analyzable_p (res, region))
2208 loop_p loop = loop_containing_stmt (SSA_NAME_DEF_STMT (res));
2209 tree scev;
2211 if (!loop_in_sese_p (loop, region))
2213 loop = loop_containing_stmt (SSA_NAME_DEF_STMT (arg));
2214 scev = scalar_evolution_in_region (region, loop, arg);
2215 scev = compute_overall_effect_of_inner_loop (loop, scev);
2217 else
2218 scev = scalar_evolution_in_region (region, loop, res);
2220 if (tree_does_not_contain_chrecs (scev))
2221 propagate_expr_outside_region (res, scev, region);
2223 gsi_next (psi);
2224 return;
2226 else
2228 tree zero_dim_array = create_zero_dim_array (res, "Close_Phi");
2230 stmt = gimple_build_assign (res, unshare_expr (zero_dim_array));
2232 if (TREE_CODE (arg) == SSA_NAME)
2233 insert_out_of_ssa_copy (scop, zero_dim_array, arg,
2234 SSA_NAME_DEF_STMT (arg));
2235 else
2236 insert_out_of_ssa_copy_on_edge (scop, single_pred_edge (bb),
2237 zero_dim_array, arg);
2240 remove_phi_node (psi, false);
2241 SSA_NAME_DEF_STMT (res) = stmt;
2243 insert_stmts (scop, stmt, NULL, gsi_after_labels (bb));
2246 /* Rewrite out of SSA the reduction phi node at PSI by creating a zero
2247 dimension array for it. */
2249 static void
2250 rewrite_phi_out_of_ssa (scop_p scop, gphi_iterator *psi)
2252 size_t i;
2253 gphi *phi = psi->phi ();
2254 basic_block bb = gimple_bb (phi);
2255 tree res = gimple_phi_result (phi);
2256 tree zero_dim_array = create_zero_dim_array (res, "phi_out_of_ssa");
2257 gimple stmt;
2259 for (i = 0; i < gimple_phi_num_args (phi); i++)
2261 tree arg = gimple_phi_arg_def (phi, i);
2262 edge e = gimple_phi_arg_edge (phi, i);
2264 /* Avoid the insertion of code in the loop latch to please the
2265 pattern matching of the vectorizer. */
2266 if (TREE_CODE (arg) == SSA_NAME
2267 && !SSA_NAME_IS_DEFAULT_DEF (arg)
2268 && e->src == bb->loop_father->latch)
2269 insert_out_of_ssa_copy (scop, zero_dim_array, arg,
2270 SSA_NAME_DEF_STMT (arg));
2271 else
2272 insert_out_of_ssa_copy_on_edge (scop, e, zero_dim_array, arg);
2275 stmt = gimple_build_assign (res, unshare_expr (zero_dim_array));
2276 remove_phi_node (psi, false);
2277 insert_stmts (scop, stmt, NULL, gsi_after_labels (bb));
2280 /* Rewrite the degenerate phi node at position PSI from the degenerate
2281 form "x = phi (y, y, ..., y)" to "x = y". */
2283 static void
2284 rewrite_degenerate_phi (gphi_iterator *psi)
2286 tree rhs;
2287 gimple stmt;
2288 gimple_stmt_iterator gsi;
2289 gphi *phi = psi->phi ();
2290 tree res = gimple_phi_result (phi);
2291 basic_block bb;
2293 bb = gimple_bb (phi);
2294 rhs = degenerate_phi_result (phi);
2295 gcc_assert (rhs);
2297 stmt = gimple_build_assign (res, rhs);
2298 remove_phi_node (psi, false);
2300 gsi = gsi_after_labels (bb);
2301 gsi_insert_before (&gsi, stmt, GSI_NEW_STMT);
2304 /* Rewrite out of SSA all the reduction phi nodes of SCOP. */
2306 static void
2307 rewrite_reductions_out_of_ssa (scop_p scop)
2309 basic_block bb;
2310 gphi_iterator psi;
2311 sese region = SCOP_REGION (scop);
2313 FOR_EACH_BB_FN (bb, cfun)
2314 if (bb_in_sese_p (bb, region))
2315 for (psi = gsi_start_phis (bb); !gsi_end_p (psi);)
2317 gphi *phi = psi.phi ();
2319 if (virtual_operand_p (gimple_phi_result (phi)))
2321 gsi_next (&psi);
2322 continue;
2325 if (gimple_phi_num_args (phi) > 1
2326 && degenerate_phi_result (phi))
2327 rewrite_degenerate_phi (&psi);
2329 else if (scalar_close_phi_node_p (phi))
2330 rewrite_close_phi_out_of_ssa (scop, &psi);
2332 else if (reduction_phi_p (region, &psi))
2333 rewrite_phi_out_of_ssa (scop, &psi);
2336 update_ssa (TODO_update_ssa);
2337 #ifdef ENABLE_CHECKING
2338 verify_loop_closed_ssa (true);
2339 #endif
2342 /* Rewrite the scalar dependence of DEF used in USE_STMT with a memory
2343 read from ZERO_DIM_ARRAY. */
2345 static void
2346 rewrite_cross_bb_scalar_dependence (scop_p scop, tree zero_dim_array,
2347 tree def, gimple use_stmt)
2349 gimple name_stmt;
2350 tree name;
2351 ssa_op_iter iter;
2352 use_operand_p use_p;
2354 gcc_assert (gimple_code (use_stmt) != GIMPLE_PHI);
2356 name = copy_ssa_name (def);
2357 name_stmt = gimple_build_assign (name, zero_dim_array);
2359 gimple_assign_set_lhs (name_stmt, name);
2360 insert_stmts (scop, name_stmt, NULL, gsi_for_stmt (use_stmt));
2362 FOR_EACH_SSA_USE_OPERAND (use_p, use_stmt, iter, SSA_OP_ALL_USES)
2363 if (operand_equal_p (def, USE_FROM_PTR (use_p), 0))
2364 replace_exp (use_p, name);
2366 update_stmt (use_stmt);
2369 /* For every definition DEF in the SCOP that is used outside the scop,
2370 insert a closing-scop definition in the basic block just after this
2371 SCOP. */
2373 static void
2374 handle_scalar_deps_crossing_scop_limits (scop_p scop, tree def, gimple stmt)
2376 tree var = create_tmp_reg (TREE_TYPE (def));
2377 tree new_name = make_ssa_name (var, stmt);
2378 bool needs_copy = false;
2379 use_operand_p use_p;
2380 imm_use_iterator imm_iter;
2381 gimple use_stmt;
2382 sese region = SCOP_REGION (scop);
2384 FOR_EACH_IMM_USE_STMT (use_stmt, imm_iter, def)
2386 if (!bb_in_sese_p (gimple_bb (use_stmt), region))
2388 FOR_EACH_IMM_USE_ON_STMT (use_p, imm_iter)
2390 SET_USE (use_p, new_name);
2392 update_stmt (use_stmt);
2393 needs_copy = true;
2397 /* Insert in the empty BB just after the scop a use of DEF such
2398 that the rewrite of cross_bb_scalar_dependences won't insert
2399 arrays everywhere else. */
2400 if (needs_copy)
2402 gimple assign = gimple_build_assign (new_name, def);
2403 gimple_stmt_iterator psi = gsi_after_labels (SESE_EXIT (region)->dest);
2405 update_stmt (assign);
2406 gsi_insert_before (&psi, assign, GSI_SAME_STMT);
2410 /* Rewrite the scalar dependences crossing the boundary of the BB
2411 containing STMT with an array. Return true when something has been
2412 changed. */
2414 static bool
2415 rewrite_cross_bb_scalar_deps (scop_p scop, gimple_stmt_iterator *gsi)
2417 sese region = SCOP_REGION (scop);
2418 gimple stmt = gsi_stmt (*gsi);
2419 imm_use_iterator imm_iter;
2420 tree def;
2421 basic_block def_bb;
2422 tree zero_dim_array = NULL_TREE;
2423 gimple use_stmt;
2424 bool res = false;
2426 switch (gimple_code (stmt))
2428 case GIMPLE_ASSIGN:
2429 def = gimple_assign_lhs (stmt);
2430 break;
2432 case GIMPLE_CALL:
2433 def = gimple_call_lhs (stmt);
2434 break;
2436 default:
2437 return false;
2440 if (!def
2441 || !is_gimple_reg (def))
2442 return false;
2444 if (scev_analyzable_p (def, region))
2446 loop_p loop = loop_containing_stmt (SSA_NAME_DEF_STMT (def));
2447 tree scev = scalar_evolution_in_region (region, loop, def);
2449 if (tree_contains_chrecs (scev, NULL))
2450 return false;
2452 propagate_expr_outside_region (def, scev, region);
2453 return true;
2456 def_bb = gimple_bb (stmt);
2458 handle_scalar_deps_crossing_scop_limits (scop, def, stmt);
2460 FOR_EACH_IMM_USE_STMT (use_stmt, imm_iter, def)
2461 if (gimple_code (use_stmt) == GIMPLE_PHI
2462 && (res = true))
2464 gphi_iterator psi = gsi_start_phis (gimple_bb (use_stmt));
2466 if (scalar_close_phi_node_p (gsi_stmt (psi)))
2467 rewrite_close_phi_out_of_ssa (scop, &psi);
2468 else
2469 rewrite_phi_out_of_ssa (scop, &psi);
2472 FOR_EACH_IMM_USE_STMT (use_stmt, imm_iter, def)
2473 if (gimple_code (use_stmt) != GIMPLE_PHI
2474 && def_bb != gimple_bb (use_stmt)
2475 && !is_gimple_debug (use_stmt)
2476 && (res = true))
2478 if (!zero_dim_array)
2480 zero_dim_array = create_zero_dim_array
2481 (def, "Cross_BB_scalar_dependence");
2482 insert_out_of_ssa_copy (scop, zero_dim_array, def,
2483 SSA_NAME_DEF_STMT (def));
2484 gsi_next (gsi);
2487 rewrite_cross_bb_scalar_dependence (scop, unshare_expr (zero_dim_array),
2488 def, use_stmt);
2491 return res;
2494 /* Rewrite out of SSA all the reduction phi nodes of SCOP. */
2496 static void
2497 rewrite_cross_bb_scalar_deps_out_of_ssa (scop_p scop)
2499 basic_block bb;
2500 gimple_stmt_iterator psi;
2501 sese region = SCOP_REGION (scop);
2502 bool changed = false;
2504 /* Create an extra empty BB after the scop. */
2505 split_edge (SESE_EXIT (region));
2507 FOR_EACH_BB_FN (bb, cfun)
2508 if (bb_in_sese_p (bb, region))
2509 for (psi = gsi_start_bb (bb); !gsi_end_p (psi); gsi_next (&psi))
2510 changed |= rewrite_cross_bb_scalar_deps (scop, &psi);
2512 if (changed)
2514 scev_reset_htab ();
2515 update_ssa (TODO_update_ssa);
2516 #ifdef ENABLE_CHECKING
2517 verify_loop_closed_ssa (true);
2518 #endif
2522 /* Returns the number of pbbs that are in loops contained in SCOP. */
2524 static int
2525 nb_pbbs_in_loops (scop_p scop)
2527 int i;
2528 poly_bb_p pbb;
2529 int res = 0;
2531 FOR_EACH_VEC_ELT (SCOP_BBS (scop), i, pbb)
2532 if (loop_in_sese_p (gbb_loop (PBB_BLACK_BOX (pbb)), SCOP_REGION (scop)))
2533 res++;
2535 return res;
2538 /* Return the number of data references in BB that write in
2539 memory. */
2541 static int
2542 nb_data_writes_in_bb (basic_block bb)
2544 int res = 0;
2545 gimple_stmt_iterator gsi;
2547 for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi))
2548 if (gimple_vdef (gsi_stmt (gsi)))
2549 res++;
2551 return res;
2554 /* Splits at STMT the basic block BB represented as PBB in the
2555 polyhedral form. */
2557 static edge
2558 split_pbb (scop_p scop, poly_bb_p pbb, basic_block bb, gimple stmt)
2560 edge e1 = split_block (bb, stmt);
2561 new_pbb_from_pbb (scop, pbb, e1->dest);
2562 return e1;
2565 /* Splits STMT out of its current BB. This is done for reduction
2566 statements for which we want to ignore data dependences. */
2568 static basic_block
2569 split_reduction_stmt (scop_p scop, gimple stmt)
2571 basic_block bb = gimple_bb (stmt);
2572 poly_bb_p pbb = pbb_from_bb (bb);
2573 gimple_bb_p gbb = gbb_from_bb (bb);
2574 edge e1;
2575 int i;
2576 data_reference_p dr;
2578 /* Do not split basic blocks with no writes to memory: the reduction
2579 will be the only write to memory. */
2580 if (nb_data_writes_in_bb (bb) == 0
2581 /* Or if we have already marked BB as a reduction. */
2582 || PBB_IS_REDUCTION (pbb_from_bb (bb)))
2583 return bb;
2585 e1 = split_pbb (scop, pbb, bb, stmt);
2587 /* Split once more only when the reduction stmt is not the only one
2588 left in the original BB. */
2589 if (!gsi_one_before_end_p (gsi_start_nondebug_bb (bb)))
2591 gimple_stmt_iterator gsi = gsi_last_bb (bb);
2592 gsi_prev (&gsi);
2593 e1 = split_pbb (scop, pbb, bb, gsi_stmt (gsi));
2596 /* A part of the data references will end in a different basic block
2597 after the split: move the DRs from the original GBB to the newly
2598 created GBB1. */
2599 FOR_EACH_VEC_ELT (GBB_DATA_REFS (gbb), i, dr)
2601 basic_block bb1 = gimple_bb (DR_STMT (dr));
2603 if (bb1 != bb)
2605 gimple_bb_p gbb1 = gbb_from_bb (bb1);
2606 GBB_DATA_REFS (gbb1).safe_push (dr);
2607 GBB_DATA_REFS (gbb).ordered_remove (i);
2608 i--;
2612 return e1->dest;
2615 /* Return true when stmt is a reduction operation. */
2617 static inline bool
2618 is_reduction_operation_p (gimple stmt)
2620 enum tree_code code;
2622 gcc_assert (is_gimple_assign (stmt));
2623 code = gimple_assign_rhs_code (stmt);
2625 return flag_associative_math
2626 && commutative_tree_code (code)
2627 && associative_tree_code (code);
2630 /* Returns true when PHI contains an argument ARG. */
2632 static bool
2633 phi_contains_arg (gphi *phi, tree arg)
2635 size_t i;
2637 for (i = 0; i < gimple_phi_num_args (phi); i++)
2638 if (operand_equal_p (arg, gimple_phi_arg_def (phi, i), 0))
2639 return true;
2641 return false;
2644 /* Return a loop phi node that corresponds to a reduction containing LHS. */
2646 static gphi *
2647 follow_ssa_with_commutative_ops (tree arg, tree lhs)
2649 gimple stmt;
2651 if (TREE_CODE (arg) != SSA_NAME)
2652 return NULL;
2654 stmt = SSA_NAME_DEF_STMT (arg);
2656 if (gimple_code (stmt) == GIMPLE_NOP
2657 || gimple_code (stmt) == GIMPLE_CALL)
2658 return NULL;
2660 if (gphi *phi = dyn_cast <gphi *> (stmt))
2662 if (phi_contains_arg (phi, lhs))
2663 return phi;
2664 return NULL;
2667 if (!is_gimple_assign (stmt))
2668 return NULL;
2670 if (gimple_num_ops (stmt) == 2)
2671 return follow_ssa_with_commutative_ops (gimple_assign_rhs1 (stmt), lhs);
2673 if (is_reduction_operation_p (stmt))
2675 gphi *res
2676 = follow_ssa_with_commutative_ops (gimple_assign_rhs1 (stmt), lhs);
2678 return res ? res :
2679 follow_ssa_with_commutative_ops (gimple_assign_rhs2 (stmt), lhs);
2682 return NULL;
2685 /* Detect commutative and associative scalar reductions starting at
2686 the STMT. Return the phi node of the reduction cycle, or NULL. */
2688 static gphi *
2689 detect_commutative_reduction_arg (tree lhs, gimple stmt, tree arg,
2690 vec<gimple> *in,
2691 vec<gimple> *out)
2693 gphi *phi = follow_ssa_with_commutative_ops (arg, lhs);
2695 if (!phi)
2696 return NULL;
2698 in->safe_push (stmt);
2699 out->safe_push (stmt);
2700 return phi;
2703 /* Detect commutative and associative scalar reductions starting at
2704 STMT. Return the phi node of the reduction cycle, or NULL. */
2706 static gphi *
2707 detect_commutative_reduction_assign (gimple stmt, vec<gimple> *in,
2708 vec<gimple> *out)
2710 tree lhs = gimple_assign_lhs (stmt);
2712 if (gimple_num_ops (stmt) == 2)
2713 return detect_commutative_reduction_arg (lhs, stmt,
2714 gimple_assign_rhs1 (stmt),
2715 in, out);
2717 if (is_reduction_operation_p (stmt))
2719 gphi *res = detect_commutative_reduction_arg (lhs, stmt,
2720 gimple_assign_rhs1 (stmt),
2721 in, out);
2722 return res ? res
2723 : detect_commutative_reduction_arg (lhs, stmt,
2724 gimple_assign_rhs2 (stmt),
2725 in, out);
2728 return NULL;
2731 /* Return a loop phi node that corresponds to a reduction containing LHS. */
2733 static gphi *
2734 follow_inital_value_to_phi (tree arg, tree lhs)
2736 gimple stmt;
2738 if (!arg || TREE_CODE (arg) != SSA_NAME)
2739 return NULL;
2741 stmt = SSA_NAME_DEF_STMT (arg);
2743 if (gphi *phi = dyn_cast <gphi *> (stmt))
2744 if (phi_contains_arg (phi, lhs))
2745 return phi;
2747 return NULL;
2751 /* Return the argument of the loop PHI that is the initial value coming
2752 from outside the loop. */
2754 static edge
2755 edge_initial_value_for_loop_phi (gphi *phi)
2757 size_t i;
2759 for (i = 0; i < gimple_phi_num_args (phi); i++)
2761 edge e = gimple_phi_arg_edge (phi, i);
2763 if (loop_depth (e->src->loop_father)
2764 < loop_depth (e->dest->loop_father))
2765 return e;
2768 return NULL;
2771 /* Return the argument of the loop PHI that is the initial value coming
2772 from outside the loop. */
2774 static tree
2775 initial_value_for_loop_phi (gphi *phi)
2777 size_t i;
2779 for (i = 0; i < gimple_phi_num_args (phi); i++)
2781 edge e = gimple_phi_arg_edge (phi, i);
2783 if (loop_depth (e->src->loop_father)
2784 < loop_depth (e->dest->loop_father))
2785 return gimple_phi_arg_def (phi, i);
2788 return NULL_TREE;
2791 /* Returns true when DEF is used outside the reduction cycle of
2792 LOOP_PHI. */
2794 static bool
2795 used_outside_reduction (tree def, gimple loop_phi)
2797 use_operand_p use_p;
2798 imm_use_iterator imm_iter;
2799 loop_p loop = loop_containing_stmt (loop_phi);
2801 /* In LOOP, DEF should be used only in LOOP_PHI. */
2802 FOR_EACH_IMM_USE_FAST (use_p, imm_iter, def)
2804 gimple stmt = USE_STMT (use_p);
2806 if (stmt != loop_phi
2807 && !is_gimple_debug (stmt)
2808 && flow_bb_inside_loop_p (loop, gimple_bb (stmt)))
2809 return true;
2812 return false;
2815 /* Detect commutative and associative scalar reductions belonging to
2816 the SCOP starting at the loop closed phi node STMT. Return the phi
2817 node of the reduction cycle, or NULL. */
2819 static gphi *
2820 detect_commutative_reduction (scop_p scop, gimple stmt, vec<gimple> *in,
2821 vec<gimple> *out)
2823 if (scalar_close_phi_node_p (stmt))
2825 gimple def;
2826 gphi *loop_phi, *phi, *close_phi = as_a <gphi *> (stmt);
2827 tree init, lhs, arg = gimple_phi_arg_def (close_phi, 0);
2829 if (TREE_CODE (arg) != SSA_NAME)
2830 return NULL;
2832 /* Note that loop close phi nodes should have a single argument
2833 because we translated the representation into a canonical form
2834 before Graphite: see canonicalize_loop_closed_ssa_form. */
2835 gcc_assert (gimple_phi_num_args (close_phi) == 1);
2837 def = SSA_NAME_DEF_STMT (arg);
2838 if (!stmt_in_sese_p (def, SCOP_REGION (scop))
2839 || !(loop_phi = detect_commutative_reduction (scop, def, in, out)))
2840 return NULL;
2842 lhs = gimple_phi_result (close_phi);
2843 init = initial_value_for_loop_phi (loop_phi);
2844 phi = follow_inital_value_to_phi (init, lhs);
2846 if (phi && (used_outside_reduction (lhs, phi)
2847 || !has_single_use (gimple_phi_result (phi))))
2848 return NULL;
2850 in->safe_push (loop_phi);
2851 out->safe_push (close_phi);
2852 return phi;
2855 if (gimple_code (stmt) == GIMPLE_ASSIGN)
2856 return detect_commutative_reduction_assign (stmt, in, out);
2858 return NULL;
2861 /* Translate the scalar reduction statement STMT to an array RED
2862 knowing that its recursive phi node is LOOP_PHI. */
2864 static void
2865 translate_scalar_reduction_to_array_for_stmt (scop_p scop, tree red,
2866 gimple stmt, gphi *loop_phi)
2868 tree res = gimple_phi_result (loop_phi);
2869 gassign *assign = gimple_build_assign (res, unshare_expr (red));
2870 gimple_stmt_iterator gsi;
2872 insert_stmts (scop, assign, NULL, gsi_after_labels (gimple_bb (loop_phi)));
2874 assign = gimple_build_assign (unshare_expr (red), gimple_assign_lhs (stmt));
2875 gsi = gsi_for_stmt (stmt);
2876 gsi_next (&gsi);
2877 insert_stmts (scop, assign, NULL, gsi);
2880 /* Removes the PHI node and resets all the debug stmts that are using
2881 the PHI_RESULT. */
2883 static void
2884 remove_phi (gphi *phi)
2886 imm_use_iterator imm_iter;
2887 tree def;
2888 use_operand_p use_p;
2889 gimple_stmt_iterator gsi;
2890 auto_vec<gimple, 3> update;
2891 unsigned int i;
2892 gimple stmt;
2894 def = PHI_RESULT (phi);
2895 FOR_EACH_IMM_USE_FAST (use_p, imm_iter, def)
2897 stmt = USE_STMT (use_p);
2899 if (is_gimple_debug (stmt))
2901 gimple_debug_bind_reset_value (stmt);
2902 update.safe_push (stmt);
2906 FOR_EACH_VEC_ELT (update, i, stmt)
2907 update_stmt (stmt);
2909 gsi = gsi_for_phi_node (phi);
2910 remove_phi_node (&gsi, false);
2913 /* Helper function for for_each_index. For each INDEX of the data
2914 reference REF, returns true when its indices are valid in the loop
2915 nest LOOP passed in as DATA. */
2917 static bool
2918 dr_indices_valid_in_loop (tree ref ATTRIBUTE_UNUSED, tree *index, void *data)
2920 loop_p loop;
2921 basic_block header, def_bb;
2922 gimple stmt;
2924 if (TREE_CODE (*index) != SSA_NAME)
2925 return true;
2927 loop = *((loop_p *) data);
2928 header = loop->header;
2929 stmt = SSA_NAME_DEF_STMT (*index);
2931 if (!stmt)
2932 return true;
2934 def_bb = gimple_bb (stmt);
2936 if (!def_bb)
2937 return true;
2939 return dominated_by_p (CDI_DOMINATORS, header, def_bb);
2942 /* When the result of a CLOSE_PHI is written to a memory location,
2943 return a pointer to that memory reference, otherwise return
2944 NULL_TREE. */
2946 static tree
2947 close_phi_written_to_memory (gphi *close_phi)
2949 imm_use_iterator imm_iter;
2950 use_operand_p use_p;
2951 gimple stmt;
2952 tree res, def = gimple_phi_result (close_phi);
2954 FOR_EACH_IMM_USE_FAST (use_p, imm_iter, def)
2955 if ((stmt = USE_STMT (use_p))
2956 && gimple_code (stmt) == GIMPLE_ASSIGN
2957 && (res = gimple_assign_lhs (stmt)))
2959 switch (TREE_CODE (res))
2961 case VAR_DECL:
2962 case PARM_DECL:
2963 case RESULT_DECL:
2964 return res;
2966 case ARRAY_REF:
2967 case MEM_REF:
2969 tree arg = gimple_phi_arg_def (close_phi, 0);
2970 loop_p nest = loop_containing_stmt (SSA_NAME_DEF_STMT (arg));
2972 /* FIXME: this restriction is for id-{24,25}.f and
2973 could be handled by duplicating the computation of
2974 array indices before the loop of the close_phi. */
2975 if (for_each_index (&res, dr_indices_valid_in_loop, &nest))
2976 return res;
2978 /* Fallthru. */
2980 default:
2981 continue;
2984 return NULL_TREE;
2987 /* Rewrite out of SSA the reduction described by the loop phi nodes
2988 IN, and the close phi nodes OUT. IN and OUT are structured by loop
2989 levels like this:
2991 IN: stmt, loop_n, ..., loop_0
2992 OUT: stmt, close_n, ..., close_0
2994 the first element is the reduction statement, and the next elements
2995 are the loop and close phi nodes of each of the outer loops. */
2997 static void
2998 translate_scalar_reduction_to_array (scop_p scop,
2999 vec<gimple> in,
3000 vec<gimple> out)
3002 gimple loop_stmt;
3003 unsigned int i = out.length () - 1;
3004 tree red = close_phi_written_to_memory (as_a <gphi *> (out[i]));
3006 FOR_EACH_VEC_ELT (in, i, loop_stmt)
3008 gimple close_stmt = out[i];
3010 if (i == 0)
3012 basic_block bb = split_reduction_stmt (scop, loop_stmt);
3013 poly_bb_p pbb = pbb_from_bb (bb);
3014 PBB_IS_REDUCTION (pbb) = true;
3015 gcc_assert (close_stmt == loop_stmt);
3017 if (!red)
3018 red = create_zero_dim_array
3019 (gimple_assign_lhs (loop_stmt), "Commutative_Associative_Reduction");
3021 translate_scalar_reduction_to_array_for_stmt (scop, red, loop_stmt,
3022 as_a <gphi *> (in[1]));
3023 continue;
3026 gphi *loop_phi = as_a <gphi *> (loop_stmt);
3027 gphi *close_phi = as_a <gphi *> (close_stmt);
3029 if (i == in.length () - 1)
3031 insert_out_of_ssa_copy (scop, gimple_phi_result (close_phi),
3032 unshare_expr (red), close_phi);
3033 insert_out_of_ssa_copy_on_edge
3034 (scop, edge_initial_value_for_loop_phi (loop_phi),
3035 unshare_expr (red), initial_value_for_loop_phi (loop_phi));
3038 remove_phi (loop_phi);
3039 remove_phi (close_phi);
3043 /* Rewrites out of SSA a commutative reduction at CLOSE_PHI. Returns
3044 true when something has been changed. */
3046 static bool
3047 rewrite_commutative_reductions_out_of_ssa_close_phi (scop_p scop,
3048 gphi *close_phi)
3050 bool res;
3051 auto_vec<gimple, 10> in;
3052 auto_vec<gimple, 10> out;
3054 detect_commutative_reduction (scop, close_phi, &in, &out);
3055 res = in.length () > 1;
3056 if (res)
3057 translate_scalar_reduction_to_array (scop, in, out);
3059 return res;
3062 /* Rewrites all the commutative reductions from LOOP out of SSA.
3063 Returns true when something has been changed. */
3065 static bool
3066 rewrite_commutative_reductions_out_of_ssa_loop (scop_p scop,
3067 loop_p loop)
3069 gphi_iterator gsi;
3070 edge exit = single_exit (loop);
3071 tree res;
3072 bool changed = false;
3074 if (!exit)
3075 return false;
3077 for (gsi = gsi_start_phis (exit->dest); !gsi_end_p (gsi); gsi_next (&gsi))
3078 if ((res = gimple_phi_result (gsi.phi ()))
3079 && !virtual_operand_p (res)
3080 && !scev_analyzable_p (res, SCOP_REGION (scop)))
3081 changed |= rewrite_commutative_reductions_out_of_ssa_close_phi
3082 (scop, gsi.phi ());
3084 return changed;
3087 /* Rewrites all the commutative reductions from SCOP out of SSA. */
3089 static void
3090 rewrite_commutative_reductions_out_of_ssa (scop_p scop)
3092 loop_p loop;
3093 bool changed = false;
3094 sese region = SCOP_REGION (scop);
3096 FOR_EACH_LOOP (loop, 0)
3097 if (loop_in_sese_p (loop, region))
3098 changed |= rewrite_commutative_reductions_out_of_ssa_loop (scop, loop);
3100 if (changed)
3102 scev_reset_htab ();
3103 gsi_commit_edge_inserts ();
3104 update_ssa (TODO_update_ssa);
3105 #ifdef ENABLE_CHECKING
3106 verify_loop_closed_ssa (true);
3107 #endif
3111 /* Can all ivs be represented by a signed integer?
3112 As CLooG might generate negative values in its expressions, signed loop ivs
3113 are required in the backend. */
3115 static bool
3116 scop_ivs_can_be_represented (scop_p scop)
3118 loop_p loop;
3119 gphi_iterator psi;
3120 bool result = true;
3122 FOR_EACH_LOOP (loop, 0)
3124 if (!loop_in_sese_p (loop, SCOP_REGION (scop)))
3125 continue;
3127 for (psi = gsi_start_phis (loop->header);
3128 !gsi_end_p (psi); gsi_next (&psi))
3130 gphi *phi = psi.phi ();
3131 tree res = PHI_RESULT (phi);
3132 tree type = TREE_TYPE (res);
3134 if (TYPE_UNSIGNED (type)
3135 && TYPE_PRECISION (type) >= TYPE_PRECISION (long_long_integer_type_node))
3137 result = false;
3138 break;
3141 if (!result)
3142 break;
3145 return result;
3148 /* Builds the polyhedral representation for a SESE region. */
3150 void
3151 build_poly_scop (scop_p scop)
3153 sese region = SCOP_REGION (scop);
3154 graphite_dim_t max_dim;
3156 build_scop_bbs (scop);
3158 /* FIXME: This restriction is needed to avoid a problem in CLooG.
3159 Once CLooG is fixed, remove this guard. Anyways, it makes no
3160 sense to optimize a scop containing only PBBs that do not belong
3161 to any loops. */
3162 if (nb_pbbs_in_loops (scop) == 0)
3163 return;
3165 if (!scop_ivs_can_be_represented (scop))
3166 return;
3168 if (flag_associative_math)
3169 rewrite_commutative_reductions_out_of_ssa (scop);
3171 build_sese_loop_nests (region);
3172 /* Record all conditions in REGION. */
3173 sese_dom_walker (CDI_DOMINATORS, region).walk (cfun->cfg->x_entry_block_ptr);
3174 find_scop_parameters (scop);
3176 max_dim = PARAM_VALUE (PARAM_GRAPHITE_MAX_NB_SCOP_PARAMS);
3177 if (scop_nb_params (scop) > max_dim)
3178 return;
3180 build_scop_iteration_domain (scop);
3181 build_scop_context (scop);
3182 add_conditions_to_constraints (scop);
3184 /* Rewrite out of SSA only after having translated the
3185 representation to the polyhedral representation to avoid scev
3186 analysis failures. That means that these functions will insert
3187 new data references that they create in the right place. */
3188 rewrite_reductions_out_of_ssa (scop);
3189 rewrite_cross_bb_scalar_deps_out_of_ssa (scop);
3191 build_scop_drs (scop);
3192 scop_to_lst (scop);
3193 build_scop_scattering (scop);
3195 /* This SCoP has been translated to the polyhedral
3196 representation. */
3197 POLY_SCOP_P (scop) = true;
3199 #endif