Update gimple.texi class hierarchy diagram
[official-gcc.git] / gcc / graphite-sese-to-poly.c
blobd9487a7566abb143a8b2c097eb8f6abc97544372
1 /* Conversion of SESE regions to Polyhedra.
2 Copyright (C) 2009-2014 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>
30 /* For C++ linkage of C functions.
31 Missing from isl/val_gmp.h in isl 0.12 versions.
32 Appearing in isl/val_gmp.h in isl 0.13.
33 To be removed when passing to isl 0.13. */
34 #if defined(__cplusplus)
35 extern "C" {
36 #endif
37 #include <isl/val_gmp.h>
38 #if defined(__cplusplus)
40 #endif
41 #ifdef HAVE_cloog
42 #include <cloog/cloog.h>
43 #include <cloog/cloog.h>
44 #include <cloog/isl/domain.h>
45 #endif
46 #endif
48 #include "system.h"
49 #include "coretypes.h"
50 #include "tree.h"
51 #include "basic-block.h"
52 #include "tree-ssa-alias.h"
53 #include "internal-fn.h"
54 #include "gimple-expr.h"
55 #include "is-a.h"
56 #include "gimple.h"
57 #include "gimple-iterator.h"
58 #include "gimplify.h"
59 #include "gimplify-me.h"
60 #include "gimple-ssa.h"
61 #include "tree-cfg.h"
62 #include "tree-phinodes.h"
63 #include "ssa-iterators.h"
64 #include "stringpool.h"
65 #include "tree-ssanames.h"
66 #include "tree-ssa-loop-manip.h"
67 #include "tree-ssa-loop-niter.h"
68 #include "tree-ssa-loop.h"
69 #include "tree-into-ssa.h"
70 #include "tree-pass.h"
71 #include "cfgloop.h"
72 #include "tree-chrec.h"
73 #include "tree-data-ref.h"
74 #include "tree-scalar-evolution.h"
75 #include "domwalk.h"
76 #include "sese.h"
77 #include "tree-ssa-propagate.h"
79 #ifdef HAVE_isl
80 #include "expr.h"
81 #include "graphite-poly.h"
82 #include "graphite-sese-to-poly.h"
85 /* Assigns to RES the value of the INTEGER_CST T. */
87 static inline void
88 tree_int_to_gmp (tree t, mpz_t res)
90 wi::to_mpz (t, res, TYPE_SIGN (TREE_TYPE (t)));
93 /* Returns the index of the PHI argument defined in the outermost
94 loop. */
96 static size_t
97 phi_arg_in_outermost_loop (gphi *phi)
99 loop_p loop = gimple_bb (phi)->loop_father;
100 size_t i, res = 0;
102 for (i = 0; i < gimple_phi_num_args (phi); i++)
103 if (!flow_bb_inside_loop_p (loop, gimple_phi_arg_edge (phi, i)->src))
105 loop = gimple_phi_arg_edge (phi, i)->src->loop_father;
106 res = i;
109 return res;
112 /* Removes a simple copy phi node "RES = phi (INIT, RES)" at position
113 PSI by inserting on the loop ENTRY edge assignment "RES = INIT". */
115 static void
116 remove_simple_copy_phi (gphi_iterator *psi)
118 gphi *phi = psi->phi ();
119 tree res = gimple_phi_result (phi);
120 size_t entry = phi_arg_in_outermost_loop (phi);
121 tree init = gimple_phi_arg_def (phi, entry);
122 gassign *stmt = gimple_build_assign (res, init);
123 edge e = gimple_phi_arg_edge (phi, entry);
125 remove_phi_node (psi, false);
126 gsi_insert_on_edge_immediate (e, stmt);
129 /* Removes an invariant phi node at position PSI by inserting on the
130 loop ENTRY edge the assignment RES = INIT. */
132 static void
133 remove_invariant_phi (sese region, gphi_iterator *psi)
135 gphi *phi = psi->phi ();
136 loop_p loop = loop_containing_stmt (phi);
137 tree res = gimple_phi_result (phi);
138 tree scev = scalar_evolution_in_region (region, loop, res);
139 size_t entry = phi_arg_in_outermost_loop (phi);
140 edge e = gimple_phi_arg_edge (phi, entry);
141 tree var;
142 gassign *stmt;
143 gimple_seq stmts = NULL;
145 if (tree_contains_chrecs (scev, NULL))
146 scev = gimple_phi_arg_def (phi, entry);
148 var = force_gimple_operand (scev, &stmts, true, NULL_TREE);
149 stmt = gimple_build_assign (res, var);
150 remove_phi_node (psi, false);
152 gimple_seq_add_stmt (&stmts, stmt);
153 gsi_insert_seq_on_edge (e, stmts);
154 gsi_commit_edge_inserts ();
155 SSA_NAME_DEF_STMT (res) = stmt;
158 /* Returns true when the phi node at PSI is of the form "a = phi (a, x)". */
160 static inline bool
161 simple_copy_phi_p (gphi *phi)
163 tree res;
165 if (gimple_phi_num_args (phi) != 2)
166 return false;
168 res = gimple_phi_result (phi);
169 return (res == gimple_phi_arg_def (phi, 0)
170 || res == gimple_phi_arg_def (phi, 1));
173 /* Returns true when the phi node at position PSI is a reduction phi
174 node in REGION. Otherwise moves the pointer PSI to the next phi to
175 be considered. */
177 static bool
178 reduction_phi_p (sese region, gphi_iterator *psi)
180 loop_p loop;
181 gphi *phi = psi->phi ();
182 tree res = gimple_phi_result (phi);
184 loop = loop_containing_stmt (phi);
186 if (simple_copy_phi_p (phi))
188 /* PRE introduces phi nodes like these, for an example,
189 see id-5.f in the fortran graphite testsuite:
191 # prephitmp.85_265 = PHI <prephitmp.85_258(33), prephitmp.85_265(18)>
193 remove_simple_copy_phi (psi);
194 return false;
197 if (scev_analyzable_p (res, region))
199 tree scev = scalar_evolution_in_region (region, loop, res);
201 if (evolution_function_is_invariant_p (scev, loop->num))
202 remove_invariant_phi (region, psi);
203 else
204 gsi_next (psi);
206 return false;
209 /* All the other cases are considered reductions. */
210 return true;
213 /* Store the GRAPHITE representation of BB. */
215 static gimple_bb_p
216 new_gimple_bb (basic_block bb, vec<data_reference_p> drs)
218 struct gimple_bb *gbb;
220 gbb = XNEW (struct gimple_bb);
221 bb->aux = gbb;
222 GBB_BB (gbb) = bb;
223 GBB_DATA_REFS (gbb) = drs;
224 GBB_CONDITIONS (gbb).create (0);
225 GBB_CONDITION_CASES (gbb).create (0);
227 return gbb;
230 static void
231 free_data_refs_aux (vec<data_reference_p> datarefs)
233 unsigned int i;
234 struct data_reference *dr;
236 FOR_EACH_VEC_ELT (datarefs, i, dr)
237 if (dr->aux)
239 base_alias_pair *bap = (base_alias_pair *)(dr->aux);
241 free (bap->alias_set);
243 free (bap);
244 dr->aux = NULL;
247 /* Frees GBB. */
249 static void
250 free_gimple_bb (struct gimple_bb *gbb)
252 free_data_refs_aux (GBB_DATA_REFS (gbb));
253 free_data_refs (GBB_DATA_REFS (gbb));
255 GBB_CONDITIONS (gbb).release ();
256 GBB_CONDITION_CASES (gbb).release ();
257 GBB_BB (gbb)->aux = 0;
258 XDELETE (gbb);
261 /* Deletes all gimple bbs in SCOP. */
263 static void
264 remove_gbbs_in_scop (scop_p scop)
266 int i;
267 poly_bb_p pbb;
269 FOR_EACH_VEC_ELT (SCOP_BBS (scop), i, pbb)
270 free_gimple_bb (PBB_BLACK_BOX (pbb));
273 /* Deletes all scops in SCOPS. */
275 void
276 free_scops (vec<scop_p> scops)
278 int i;
279 scop_p scop;
281 FOR_EACH_VEC_ELT (scops, i, scop)
283 remove_gbbs_in_scop (scop);
284 free_sese (SCOP_REGION (scop));
285 free_scop (scop);
288 scops.release ();
291 /* Same as outermost_loop_in_sese, returns the outermost loop
292 containing BB in REGION, but makes sure that the returned loop
293 belongs to the REGION, and so this returns the first loop in the
294 REGION when the loop containing BB does not belong to REGION. */
296 static loop_p
297 outermost_loop_in_sese_1 (sese region, basic_block bb)
299 loop_p nest = outermost_loop_in_sese (region, bb);
301 if (loop_in_sese_p (nest, region))
302 return nest;
304 /* When the basic block BB does not belong to a loop in the region,
305 return the first loop in the region. */
306 nest = nest->inner;
307 while (nest)
308 if (loop_in_sese_p (nest, region))
309 break;
310 else
311 nest = nest->next;
313 gcc_assert (nest);
314 return nest;
317 /* Generates a polyhedral black box only if the bb contains interesting
318 information. */
320 static gimple_bb_p
321 try_generate_gimple_bb (scop_p scop, basic_block bb)
323 vec<data_reference_p> drs;
324 drs.create (5);
325 sese region = SCOP_REGION (scop);
326 loop_p nest = outermost_loop_in_sese_1 (region, bb);
327 gimple_stmt_iterator gsi;
329 for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi))
331 gimple stmt = gsi_stmt (gsi);
332 loop_p loop;
334 if (is_gimple_debug (stmt))
335 continue;
337 loop = loop_containing_stmt (stmt);
338 if (!loop_in_sese_p (loop, region))
339 loop = nest;
341 graphite_find_data_references_in_stmt (nest, loop, stmt, &drs);
344 return new_gimple_bb (bb, drs);
347 /* Returns true if all predecessors of BB, that are not dominated by BB, are
348 marked in MAP. The predecessors dominated by BB are loop latches and will
349 be handled after BB. */
351 static bool
352 all_non_dominated_preds_marked_p (basic_block bb, sbitmap map)
354 edge e;
355 edge_iterator ei;
357 FOR_EACH_EDGE (e, ei, bb->preds)
358 if (!bitmap_bit_p (map, e->src->index)
359 && !dominated_by_p (CDI_DOMINATORS, e->src, bb))
360 return false;
362 return true;
365 /* Compare the depth of two basic_block's P1 and P2. */
367 static int
368 compare_bb_depths (const void *p1, const void *p2)
370 const_basic_block const bb1 = *(const_basic_block const*)p1;
371 const_basic_block const bb2 = *(const_basic_block const*)p2;
372 int d1 = loop_depth (bb1->loop_father);
373 int d2 = loop_depth (bb2->loop_father);
375 if (d1 < d2)
376 return 1;
378 if (d1 > d2)
379 return -1;
381 return 0;
384 /* Sort the basic blocks from DOM such that the first are the ones at
385 a deepest loop level. */
387 static void
388 graphite_sort_dominated_info (vec<basic_block> dom)
390 dom.qsort (compare_bb_depths);
393 /* Recursive helper function for build_scops_bbs. */
395 static void
396 build_scop_bbs_1 (scop_p scop, sbitmap visited, basic_block bb)
398 sese region = SCOP_REGION (scop);
399 vec<basic_block> dom;
400 poly_bb_p pbb;
402 if (bitmap_bit_p (visited, bb->index)
403 || !bb_in_sese_p (bb, region))
404 return;
406 pbb = new_poly_bb (scop, try_generate_gimple_bb (scop, bb));
407 SCOP_BBS (scop).safe_push (pbb);
408 bitmap_set_bit (visited, bb->index);
410 dom = get_dominated_by (CDI_DOMINATORS, bb);
412 if (!dom.exists ())
413 return;
415 graphite_sort_dominated_info (dom);
417 while (!dom.is_empty ())
419 int i;
420 basic_block dom_bb;
422 FOR_EACH_VEC_ELT (dom, i, dom_bb)
423 if (all_non_dominated_preds_marked_p (dom_bb, visited))
425 build_scop_bbs_1 (scop, visited, dom_bb);
426 dom.unordered_remove (i);
427 break;
431 dom.release ();
434 /* Gather the basic blocks belonging to the SCOP. */
436 static void
437 build_scop_bbs (scop_p scop)
439 sbitmap visited = sbitmap_alloc (last_basic_block_for_fn (cfun));
440 sese region = SCOP_REGION (scop);
442 bitmap_clear (visited);
443 build_scop_bbs_1 (scop, visited, SESE_ENTRY_BB (region));
444 sbitmap_free (visited);
447 /* Return an ISL identifier for the polyhedral basic block PBB. */
449 static isl_id *
450 isl_id_for_pbb (scop_p s, poly_bb_p pbb)
452 char name[50];
453 snprintf (name, sizeof (name), "S_%d", pbb_index (pbb));
454 return isl_id_alloc (s->ctx, name, pbb);
457 /* Converts the STATIC_SCHEDULE of PBB into a scattering polyhedron.
458 We generate SCATTERING_DIMENSIONS scattering dimensions.
460 CLooG 0.15.0 and previous versions require, that all
461 scattering functions of one CloogProgram have the same number of
462 scattering dimensions, therefore we allow to specify it. This
463 should be removed in future versions of CLooG.
465 The scattering polyhedron consists of these dimensions: scattering,
466 loop_iterators, parameters.
468 Example:
470 | scattering_dimensions = 5
471 | used_scattering_dimensions = 3
472 | nb_iterators = 1
473 | scop_nb_params = 2
475 | Schedule:
477 | 4 5
479 | Scattering polyhedron:
481 | scattering: {s1, s2, s3, s4, s5}
482 | loop_iterators: {i}
483 | parameters: {p1, p2}
485 | s1 s2 s3 s4 s5 i p1 p2 1
486 | 1 0 0 0 0 0 0 0 -4 = 0
487 | 0 1 0 0 0 -1 0 0 0 = 0
488 | 0 0 1 0 0 0 0 0 -5 = 0 */
490 static void
491 build_pbb_scattering_polyhedrons (isl_aff *static_sched,
492 poly_bb_p pbb, int scattering_dimensions)
494 int i;
495 int nb_iterators = pbb_dim_iter_domain (pbb);
496 int used_scattering_dimensions = nb_iterators * 2 + 1;
497 isl_val *val;
498 isl_space *dc, *dm;
500 gcc_assert (scattering_dimensions >= used_scattering_dimensions);
502 dc = isl_set_get_space (pbb->domain);
503 dm = isl_space_add_dims (isl_space_from_domain (dc),
504 isl_dim_out, scattering_dimensions);
505 pbb->schedule = isl_map_universe (dm);
507 for (i = 0; i < scattering_dimensions; i++)
509 /* Textual order inside this loop. */
510 if ((i % 2) == 0)
512 isl_constraint *c = isl_equality_alloc
513 (isl_local_space_from_space (isl_map_get_space (pbb->schedule)));
515 val = isl_aff_get_coefficient_val (static_sched, isl_dim_in, i / 2);
517 val = isl_val_neg (val);
518 c = isl_constraint_set_constant_val (c, val);
519 c = isl_constraint_set_coefficient_si (c, isl_dim_out, i, 1);
520 pbb->schedule = isl_map_add_constraint (pbb->schedule, c);
523 /* Iterations of this loop. */
524 else /* if ((i % 2) == 1) */
526 int loop = (i - 1) / 2;
527 pbb->schedule = isl_map_equate (pbb->schedule, isl_dim_in, loop,
528 isl_dim_out, i);
532 pbb->transformed = isl_map_copy (pbb->schedule);
535 /* Build for BB the static schedule.
537 The static schedule is a Dewey numbering of the abstract syntax
538 tree: http://en.wikipedia.org/wiki/Dewey_Decimal_Classification
540 The following example informally defines the static schedule:
543 for (i: ...)
545 for (j: ...)
551 for (k: ...)
559 Static schedules for A to F:
561 DEPTH
562 0 1 2
564 B 1 0 0
565 C 1 0 1
566 D 1 1 0
567 E 1 1 1
571 static void
572 build_scop_scattering (scop_p scop)
574 int i;
575 poly_bb_p pbb;
576 gimple_bb_p previous_gbb = NULL;
577 isl_space *dc = isl_set_get_space (scop->context);
578 isl_aff *static_sched;
580 dc = isl_space_add_dims (dc, isl_dim_set, number_of_loops (cfun));
581 static_sched = isl_aff_zero_on_domain (isl_local_space_from_space (dc));
583 /* We have to start schedules at 0 on the first component and
584 because we cannot compare_prefix_loops against a previous loop,
585 prefix will be equal to zero, and that index will be
586 incremented before copying. */
587 static_sched = isl_aff_add_coefficient_si (static_sched, isl_dim_in, 0, -1);
589 FOR_EACH_VEC_ELT (SCOP_BBS (scop), i, pbb)
591 gimple_bb_p gbb = PBB_BLACK_BOX (pbb);
592 int prefix;
593 int nb_scat_dims = pbb_dim_iter_domain (pbb) * 2 + 1;
595 if (previous_gbb)
596 prefix = nb_common_loops (SCOP_REGION (scop), previous_gbb, gbb);
597 else
598 prefix = 0;
600 previous_gbb = gbb;
602 static_sched = isl_aff_add_coefficient_si (static_sched, isl_dim_in,
603 prefix, 1);
604 build_pbb_scattering_polyhedrons (static_sched, pbb, nb_scat_dims);
607 isl_aff_free (static_sched);
610 static isl_pw_aff *extract_affine (scop_p, tree, __isl_take isl_space *space);
612 /* Extract an affine expression from the chain of recurrence E. */
614 static isl_pw_aff *
615 extract_affine_chrec (scop_p s, tree e, __isl_take isl_space *space)
617 isl_pw_aff *lhs = extract_affine (s, CHREC_LEFT (e), isl_space_copy (space));
618 isl_pw_aff *rhs = extract_affine (s, CHREC_RIGHT (e), isl_space_copy (space));
619 isl_local_space *ls = isl_local_space_from_space (space);
620 unsigned pos = sese_loop_depth ((sese) s->region, get_chrec_loop (e)) - 1;
621 isl_aff *loop = isl_aff_set_coefficient_si
622 (isl_aff_zero_on_domain (ls), isl_dim_in, pos, 1);
623 isl_pw_aff *l = isl_pw_aff_from_aff (loop);
625 /* Before multiplying, make sure that the result is affine. */
626 gcc_assert (isl_pw_aff_is_cst (rhs)
627 || isl_pw_aff_is_cst (l));
629 return isl_pw_aff_add (lhs, isl_pw_aff_mul (rhs, l));
632 /* Extract an affine expression from the mult_expr E. */
634 static isl_pw_aff *
635 extract_affine_mul (scop_p s, tree e, __isl_take isl_space *space)
637 isl_pw_aff *lhs = extract_affine (s, TREE_OPERAND (e, 0),
638 isl_space_copy (space));
639 isl_pw_aff *rhs = extract_affine (s, TREE_OPERAND (e, 1), space);
641 if (!isl_pw_aff_is_cst (lhs)
642 && !isl_pw_aff_is_cst (rhs))
644 isl_pw_aff_free (lhs);
645 isl_pw_aff_free (rhs);
646 return NULL;
649 return isl_pw_aff_mul (lhs, rhs);
652 /* Return an ISL identifier from the name of the ssa_name E. */
654 static isl_id *
655 isl_id_for_ssa_name (scop_p s, tree e)
657 const char *name = get_name (e);
658 isl_id *id;
660 if (name)
661 id = isl_id_alloc (s->ctx, name, e);
662 else
664 char name1[50];
665 snprintf (name1, sizeof (name1), "P_%d", SSA_NAME_VERSION (e));
666 id = isl_id_alloc (s->ctx, name1, e);
669 return id;
672 /* Return an ISL identifier for the data reference DR. */
674 static isl_id *
675 isl_id_for_dr (scop_p s, data_reference_p dr ATTRIBUTE_UNUSED)
677 /* Data references all get the same isl_id. They need to be comparable
678 and are distinguished through the first dimension, which contains the
679 alias set number. */
680 return isl_id_alloc (s->ctx, "", 0);
683 /* Extract an affine expression from the ssa_name E. */
685 static isl_pw_aff *
686 extract_affine_name (scop_p s, tree e, __isl_take isl_space *space)
688 isl_aff *aff;
689 isl_set *dom;
690 isl_id *id;
691 int dimension;
693 id = isl_id_for_ssa_name (s, e);
694 dimension = isl_space_find_dim_by_id (space, isl_dim_param, id);
695 isl_id_free (id);
696 dom = isl_set_universe (isl_space_copy (space));
697 aff = isl_aff_zero_on_domain (isl_local_space_from_space (space));
698 aff = isl_aff_add_coefficient_si (aff, isl_dim_param, dimension, 1);
699 return isl_pw_aff_alloc (dom, aff);
702 /* Extract an affine expression from the gmp constant G. */
704 static isl_pw_aff *
705 extract_affine_gmp (mpz_t g, __isl_take isl_space *space)
707 isl_local_space *ls = isl_local_space_from_space (isl_space_copy (space));
708 isl_aff *aff = isl_aff_zero_on_domain (ls);
709 isl_set *dom = isl_set_universe (space);
710 isl_val *v;
711 isl_ctx *ct;
713 ct = isl_aff_get_ctx (aff);
714 v = isl_val_int_from_gmp (ct, g);
715 aff = isl_aff_add_constant_val (aff, v);
717 return isl_pw_aff_alloc (dom, aff);
720 /* Extract an affine expression from the integer_cst E. */
722 static isl_pw_aff *
723 extract_affine_int (tree e, __isl_take isl_space *space)
725 isl_pw_aff *res;
726 mpz_t g;
728 mpz_init (g);
729 tree_int_to_gmp (e, g);
730 res = extract_affine_gmp (g, space);
731 mpz_clear (g);
733 return res;
736 /* Compute pwaff mod 2^width. */
738 extern isl_ctx *the_isl_ctx;
740 static isl_pw_aff *
741 wrap (isl_pw_aff *pwaff, unsigned width)
743 isl_val *mod;
745 mod = isl_val_int_from_ui(the_isl_ctx, width);
746 mod = isl_val_2exp (mod);
747 pwaff = isl_pw_aff_mod_val (pwaff, mod);
749 return pwaff;
752 /* When parameter NAME is in REGION, returns its index in SESE_PARAMS.
753 Otherwise returns -1. */
755 static inline int
756 parameter_index_in_region_1 (tree name, sese region)
758 int i;
759 tree p;
761 gcc_assert (TREE_CODE (name) == SSA_NAME);
763 FOR_EACH_VEC_ELT (SESE_PARAMS (region), i, p)
764 if (p == name)
765 return i;
767 return -1;
770 /* When the parameter NAME is in REGION, returns its index in
771 SESE_PARAMS. Otherwise this function inserts NAME in SESE_PARAMS
772 and returns the index of NAME. */
774 static int
775 parameter_index_in_region (tree name, sese region)
777 int i;
779 gcc_assert (TREE_CODE (name) == SSA_NAME);
781 i = parameter_index_in_region_1 (name, region);
782 if (i != -1)
783 return i;
785 gcc_assert (SESE_ADD_PARAMS (region));
787 i = SESE_PARAMS (region).length ();
788 SESE_PARAMS (region).safe_push (name);
789 return i;
792 /* Extract an affine expression from the tree E in the scop S. */
794 static isl_pw_aff *
795 extract_affine (scop_p s, tree e, __isl_take isl_space *space)
797 isl_pw_aff *lhs, *rhs, *res;
798 tree type;
800 if (e == chrec_dont_know) {
801 isl_space_free (space);
802 return NULL;
805 switch (TREE_CODE (e))
807 case POLYNOMIAL_CHREC:
808 res = extract_affine_chrec (s, e, space);
809 break;
811 case MULT_EXPR:
812 res = extract_affine_mul (s, e, space);
813 break;
815 case PLUS_EXPR:
816 case POINTER_PLUS_EXPR:
817 lhs = extract_affine (s, TREE_OPERAND (e, 0), isl_space_copy (space));
818 rhs = extract_affine (s, TREE_OPERAND (e, 1), space);
819 res = isl_pw_aff_add (lhs, rhs);
820 break;
822 case MINUS_EXPR:
823 lhs = extract_affine (s, TREE_OPERAND (e, 0), isl_space_copy (space));
824 rhs = extract_affine (s, TREE_OPERAND (e, 1), space);
825 res = isl_pw_aff_sub (lhs, rhs);
826 break;
828 case NEGATE_EXPR:
829 case BIT_NOT_EXPR:
830 lhs = extract_affine (s, TREE_OPERAND (e, 0), isl_space_copy (space));
831 rhs = extract_affine (s, integer_minus_one_node, space);
832 res = isl_pw_aff_mul (lhs, rhs);
833 break;
835 case SSA_NAME:
836 gcc_assert (-1 != parameter_index_in_region_1 (e, SCOP_REGION (s)));
837 res = extract_affine_name (s, e, space);
838 break;
840 case INTEGER_CST:
841 res = extract_affine_int (e, space);
842 /* No need to wrap a single integer. */
843 return res;
845 CASE_CONVERT:
846 case NON_LVALUE_EXPR:
847 res = extract_affine (s, TREE_OPERAND (e, 0), space);
848 break;
850 default:
851 gcc_unreachable ();
852 break;
855 type = TREE_TYPE (e);
856 if (TYPE_UNSIGNED (type))
857 res = wrap (res, TYPE_PRECISION (type));
859 return res;
862 /* In the context of sese S, scan the expression E and translate it to
863 a linear expression C. When parsing a symbolic multiplication, K
864 represents the constant multiplier of an expression containing
865 parameters. */
867 static void
868 scan_tree_for_params (sese s, tree e)
870 if (e == chrec_dont_know)
871 return;
873 switch (TREE_CODE (e))
875 case POLYNOMIAL_CHREC:
876 scan_tree_for_params (s, CHREC_LEFT (e));
877 break;
879 case MULT_EXPR:
880 if (chrec_contains_symbols (TREE_OPERAND (e, 0)))
881 scan_tree_for_params (s, TREE_OPERAND (e, 0));
882 else
883 scan_tree_for_params (s, TREE_OPERAND (e, 1));
884 break;
886 case PLUS_EXPR:
887 case POINTER_PLUS_EXPR:
888 case MINUS_EXPR:
889 scan_tree_for_params (s, TREE_OPERAND (e, 0));
890 scan_tree_for_params (s, TREE_OPERAND (e, 1));
891 break;
893 case NEGATE_EXPR:
894 case BIT_NOT_EXPR:
895 CASE_CONVERT:
896 case NON_LVALUE_EXPR:
897 scan_tree_for_params (s, TREE_OPERAND (e, 0));
898 break;
900 case SSA_NAME:
901 parameter_index_in_region (e, s);
902 break;
904 case INTEGER_CST:
905 case ADDR_EXPR:
906 break;
908 default:
909 gcc_unreachable ();
910 break;
914 /* Find parameters with respect to REGION in BB. We are looking in memory
915 access functions, conditions and loop bounds. */
917 static void
918 find_params_in_bb (sese region, gimple_bb_p gbb)
920 int i;
921 unsigned j;
922 data_reference_p dr;
923 gimple stmt;
924 loop_p loop = GBB_BB (gbb)->loop_father;
926 /* Find parameters in the access functions of data references. */
927 FOR_EACH_VEC_ELT (GBB_DATA_REFS (gbb), i, dr)
928 for (j = 0; j < DR_NUM_DIMENSIONS (dr); j++)
929 scan_tree_for_params (region, DR_ACCESS_FN (dr, j));
931 /* Find parameters in conditional statements. */
932 FOR_EACH_VEC_ELT (GBB_CONDITIONS (gbb), i, stmt)
934 tree lhs = scalar_evolution_in_region (region, loop,
935 gimple_cond_lhs (stmt));
936 tree rhs = scalar_evolution_in_region (region, loop,
937 gimple_cond_rhs (stmt));
939 scan_tree_for_params (region, lhs);
940 scan_tree_for_params (region, rhs);
944 /* Record the parameters used in the SCOP. A variable is a parameter
945 in a scop if it does not vary during the execution of that scop. */
947 static void
948 find_scop_parameters (scop_p scop)
950 poly_bb_p pbb;
951 unsigned i;
952 sese region = SCOP_REGION (scop);
953 struct loop *loop;
954 int nbp;
956 /* Find the parameters used in the loop bounds. */
957 FOR_EACH_VEC_ELT (SESE_LOOP_NEST (region), i, loop)
959 tree nb_iters = number_of_latch_executions (loop);
961 if (!chrec_contains_symbols (nb_iters))
962 continue;
964 nb_iters = scalar_evolution_in_region (region, loop, nb_iters);
965 scan_tree_for_params (region, nb_iters);
968 /* Find the parameters used in data accesses. */
969 FOR_EACH_VEC_ELT (SCOP_BBS (scop), i, pbb)
970 find_params_in_bb (region, PBB_BLACK_BOX (pbb));
972 nbp = sese_nb_params (region);
973 scop_set_nb_params (scop, nbp);
974 SESE_ADD_PARAMS (region) = false;
977 tree e;
978 isl_space *space = isl_space_set_alloc (scop->ctx, nbp, 0);
980 FOR_EACH_VEC_ELT (SESE_PARAMS (region), i, e)
981 space = isl_space_set_dim_id (space, isl_dim_param, i,
982 isl_id_for_ssa_name (scop, e));
984 scop->context = isl_set_universe (space);
988 /* Builds the constraint polyhedra for LOOP in SCOP. OUTER_PH gives
989 the constraints for the surrounding loops. */
991 static void
992 build_loop_iteration_domains (scop_p scop, struct loop *loop,
993 int nb,
994 isl_set *outer, isl_set **doms)
996 tree nb_iters = number_of_latch_executions (loop);
997 sese region = SCOP_REGION (scop);
999 isl_set *inner = isl_set_copy (outer);
1000 isl_space *space;
1001 isl_constraint *c;
1002 int pos = isl_set_dim (outer, isl_dim_set);
1003 isl_val *v;
1004 mpz_t g;
1006 mpz_init (g);
1008 inner = isl_set_add_dims (inner, isl_dim_set, 1);
1009 space = isl_set_get_space (inner);
1011 /* 0 <= loop_i */
1012 c = isl_inequality_alloc
1013 (isl_local_space_from_space (isl_space_copy (space)));
1014 c = isl_constraint_set_coefficient_si (c, isl_dim_set, pos, 1);
1015 inner = isl_set_add_constraint (inner, c);
1017 /* loop_i <= cst_nb_iters */
1018 if (TREE_CODE (nb_iters) == INTEGER_CST)
1020 c = isl_inequality_alloc
1021 (isl_local_space_from_space (isl_space_copy (space)));
1022 c = isl_constraint_set_coefficient_si (c, isl_dim_set, pos, -1);
1023 tree_int_to_gmp (nb_iters, g);
1024 v = isl_val_int_from_gmp (the_isl_ctx, g);
1025 c = isl_constraint_set_constant_val (c, v);
1026 inner = isl_set_add_constraint (inner, c);
1029 /* loop_i <= expr_nb_iters */
1030 else if (!chrec_contains_undetermined (nb_iters))
1032 widest_int nit;
1033 isl_pw_aff *aff;
1034 isl_set *valid;
1035 isl_local_space *ls;
1036 isl_aff *al;
1037 isl_set *le;
1039 nb_iters = scalar_evolution_in_region (region, loop, nb_iters);
1041 aff = extract_affine (scop, nb_iters, isl_set_get_space (inner));
1042 valid = isl_pw_aff_nonneg_set (isl_pw_aff_copy (aff));
1043 valid = isl_set_project_out (valid, isl_dim_set, 0,
1044 isl_set_dim (valid, isl_dim_set));
1045 scop->context = isl_set_intersect (scop->context, valid);
1047 ls = isl_local_space_from_space (isl_space_copy (space));
1048 al = isl_aff_set_coefficient_si (isl_aff_zero_on_domain (ls),
1049 isl_dim_in, pos, 1);
1050 le = isl_pw_aff_le_set (isl_pw_aff_from_aff (al),
1051 isl_pw_aff_copy (aff));
1052 inner = isl_set_intersect (inner, le);
1054 if (max_stmt_executions (loop, &nit))
1056 /* Insert in the context the constraints from the
1057 estimation of the number of iterations NIT and the
1058 symbolic number of iterations (involving parameter
1059 names) NB_ITERS. First, build the affine expression
1060 "NIT - NB_ITERS" and then say that it is positive,
1061 i.e., NIT approximates NB_ITERS: "NIT >= NB_ITERS". */
1062 isl_pw_aff *approx;
1063 mpz_t g;
1064 isl_set *x;
1065 isl_constraint *c;
1067 mpz_init (g);
1068 wi::to_mpz (nit, g, SIGNED);
1069 mpz_sub_ui (g, g, 1);
1070 approx = extract_affine_gmp (g, isl_set_get_space (inner));
1071 x = isl_pw_aff_ge_set (approx, aff);
1072 x = isl_set_project_out (x, isl_dim_set, 0,
1073 isl_set_dim (x, isl_dim_set));
1074 scop->context = isl_set_intersect (scop->context, x);
1076 c = isl_inequality_alloc
1077 (isl_local_space_from_space (isl_space_copy (space)));
1078 c = isl_constraint_set_coefficient_si (c, isl_dim_set, pos, -1);
1079 v = isl_val_int_from_gmp (the_isl_ctx, g);
1080 mpz_clear (g);
1081 c = isl_constraint_set_constant_val (c, v);
1082 inner = isl_set_add_constraint (inner, c);
1084 else
1085 isl_pw_aff_free (aff);
1087 else
1088 gcc_unreachable ();
1090 if (loop->inner && loop_in_sese_p (loop->inner, region))
1091 build_loop_iteration_domains (scop, loop->inner, nb + 1,
1092 isl_set_copy (inner), doms);
1094 if (nb != 0
1095 && loop->next
1096 && loop_in_sese_p (loop->next, region))
1097 build_loop_iteration_domains (scop, loop->next, nb,
1098 isl_set_copy (outer), doms);
1100 doms[loop->num] = inner;
1102 isl_set_free (outer);
1103 isl_space_free (space);
1104 mpz_clear (g);
1107 /* Returns a linear expression for tree T evaluated in PBB. */
1109 static isl_pw_aff *
1110 create_pw_aff_from_tree (poly_bb_p pbb, tree t)
1112 scop_p scop = PBB_SCOP (pbb);
1114 t = scalar_evolution_in_region (SCOP_REGION (scop), pbb_loop (pbb), t);
1115 gcc_assert (!automatically_generated_chrec_p (t));
1117 return extract_affine (scop, t, isl_set_get_space (pbb->domain));
1120 /* Add conditional statement STMT to pbb. CODE is used as the comparison
1121 operator. This allows us to invert the condition or to handle
1122 inequalities. */
1124 static void
1125 add_condition_to_pbb (poly_bb_p pbb, gcond *stmt, enum tree_code code)
1127 isl_pw_aff *lhs = create_pw_aff_from_tree (pbb, gimple_cond_lhs (stmt));
1128 isl_pw_aff *rhs = create_pw_aff_from_tree (pbb, gimple_cond_rhs (stmt));
1129 isl_set *cond;
1131 switch (code)
1133 case LT_EXPR:
1134 cond = isl_pw_aff_lt_set (lhs, rhs);
1135 break;
1137 case GT_EXPR:
1138 cond = isl_pw_aff_gt_set (lhs, rhs);
1139 break;
1141 case LE_EXPR:
1142 cond = isl_pw_aff_le_set (lhs, rhs);
1143 break;
1145 case GE_EXPR:
1146 cond = isl_pw_aff_ge_set (lhs, rhs);
1147 break;
1149 case EQ_EXPR:
1150 cond = isl_pw_aff_eq_set (lhs, rhs);
1151 break;
1153 case NE_EXPR:
1154 cond = isl_pw_aff_ne_set (lhs, rhs);
1155 break;
1157 default:
1158 isl_pw_aff_free (lhs);
1159 isl_pw_aff_free (rhs);
1160 return;
1163 cond = isl_set_coalesce (cond);
1164 cond = isl_set_set_tuple_id (cond, isl_set_get_tuple_id (pbb->domain));
1165 pbb->domain = isl_set_intersect (pbb->domain, cond);
1168 /* Add conditions to the domain of PBB. */
1170 static void
1171 add_conditions_to_domain (poly_bb_p pbb)
1173 unsigned int i;
1174 gimple stmt;
1175 gimple_bb_p gbb = PBB_BLACK_BOX (pbb);
1177 if (GBB_CONDITIONS (gbb).is_empty ())
1178 return;
1180 FOR_EACH_VEC_ELT (GBB_CONDITIONS (gbb), i, stmt)
1181 switch (gimple_code (stmt))
1183 case GIMPLE_COND:
1185 gcond *cond_stmt = as_a <gcond *> (stmt);
1186 enum tree_code code = gimple_cond_code (cond_stmt);
1188 /* The conditions for ELSE-branches are inverted. */
1189 if (!GBB_CONDITION_CASES (gbb)[i])
1190 code = invert_tree_comparison (code, false);
1192 add_condition_to_pbb (pbb, cond_stmt, code);
1193 break;
1196 case GIMPLE_SWITCH:
1197 /* Switch statements are not supported right now - fall through. */
1199 default:
1200 gcc_unreachable ();
1201 break;
1205 /* Traverses all the GBBs of the SCOP and add their constraints to the
1206 iteration domains. */
1208 static void
1209 add_conditions_to_constraints (scop_p scop)
1211 int i;
1212 poly_bb_p pbb;
1214 FOR_EACH_VEC_ELT (SCOP_BBS (scop), i, pbb)
1215 add_conditions_to_domain (pbb);
1218 /* Returns a COND_EXPR statement when BB has a single predecessor, the
1219 edge between BB and its predecessor is not a loop exit edge, and
1220 the last statement of the single predecessor is a COND_EXPR. */
1222 static gcond *
1223 single_pred_cond_non_loop_exit (basic_block bb)
1225 if (single_pred_p (bb))
1227 edge e = single_pred_edge (bb);
1228 basic_block pred = e->src;
1229 gimple stmt;
1231 if (loop_depth (pred->loop_father) > loop_depth (bb->loop_father))
1232 return NULL;
1234 stmt = last_stmt (pred);
1236 if (stmt && gimple_code (stmt) == GIMPLE_COND)
1237 return as_a <gcond *> (stmt);
1240 return NULL;
1243 class sese_dom_walker : public dom_walker
1245 public:
1246 sese_dom_walker (cdi_direction, sese);
1248 virtual void before_dom_children (basic_block);
1249 virtual void after_dom_children (basic_block);
1251 private:
1252 auto_vec<gimple, 3> m_conditions, m_cases;
1253 sese m_region;
1256 sese_dom_walker::sese_dom_walker (cdi_direction direction, sese region)
1257 : dom_walker (direction), m_region (region)
1261 /* Call-back for dom_walk executed before visiting the dominated
1262 blocks. */
1264 void
1265 sese_dom_walker::before_dom_children (basic_block bb)
1267 gimple_bb_p gbb;
1268 gcond *stmt;
1270 if (!bb_in_sese_p (bb, m_region))
1271 return;
1273 stmt = single_pred_cond_non_loop_exit (bb);
1275 if (stmt)
1277 edge e = single_pred_edge (bb);
1279 m_conditions.safe_push (stmt);
1281 if (e->flags & EDGE_TRUE_VALUE)
1282 m_cases.safe_push (stmt);
1283 else
1284 m_cases.safe_push (NULL);
1287 gbb = gbb_from_bb (bb);
1289 if (gbb)
1291 GBB_CONDITIONS (gbb) = m_conditions.copy ();
1292 GBB_CONDITION_CASES (gbb) = m_cases.copy ();
1296 /* Call-back for dom_walk executed after visiting the dominated
1297 blocks. */
1299 void
1300 sese_dom_walker::after_dom_children (basic_block bb)
1302 if (!bb_in_sese_p (bb, m_region))
1303 return;
1305 if (single_pred_cond_non_loop_exit (bb))
1307 m_conditions.pop ();
1308 m_cases.pop ();
1312 /* Add constraints on the possible values of parameter P from the type
1313 of P. */
1315 static void
1316 add_param_constraints (scop_p scop, graphite_dim_t p)
1318 tree parameter = SESE_PARAMS (SCOP_REGION (scop))[p];
1319 tree type = TREE_TYPE (parameter);
1320 tree lb = NULL_TREE;
1321 tree ub = NULL_TREE;
1323 if (POINTER_TYPE_P (type) || !TYPE_MIN_VALUE (type))
1324 lb = lower_bound_in_type (type, type);
1325 else
1326 lb = TYPE_MIN_VALUE (type);
1328 if (POINTER_TYPE_P (type) || !TYPE_MAX_VALUE (type))
1329 ub = upper_bound_in_type (type, type);
1330 else
1331 ub = TYPE_MAX_VALUE (type);
1333 if (lb)
1335 isl_space *space = isl_set_get_space (scop->context);
1336 isl_constraint *c;
1337 mpz_t g;
1338 isl_val *v;
1340 c = isl_inequality_alloc (isl_local_space_from_space (space));
1341 mpz_init (g);
1342 tree_int_to_gmp (lb, g);
1343 v = isl_val_int_from_gmp (the_isl_ctx, g);
1344 v = isl_val_neg (v);
1345 mpz_clear (g);
1346 c = isl_constraint_set_constant_val (c, v);
1347 c = isl_constraint_set_coefficient_si (c, isl_dim_param, p, 1);
1349 scop->context = isl_set_add_constraint (scop->context, c);
1352 if (ub)
1354 isl_space *space = isl_set_get_space (scop->context);
1355 isl_constraint *c;
1356 mpz_t g;
1357 isl_val *v;
1359 c = isl_inequality_alloc (isl_local_space_from_space (space));
1361 mpz_init (g);
1362 tree_int_to_gmp (ub, g);
1363 v = isl_val_int_from_gmp (the_isl_ctx, g);
1364 mpz_clear (g);
1365 c = isl_constraint_set_constant_val (c, v);
1366 c = isl_constraint_set_coefficient_si (c, isl_dim_param, p, -1);
1368 scop->context = isl_set_add_constraint (scop->context, c);
1372 /* Build the context of the SCOP. The context usually contains extra
1373 constraints that are added to the iteration domains that constrain
1374 some parameters. */
1376 static void
1377 build_scop_context (scop_p scop)
1379 graphite_dim_t p, n = scop_nb_params (scop);
1381 for (p = 0; p < n; p++)
1382 add_param_constraints (scop, p);
1385 /* Build the iteration domains: the loops belonging to the current
1386 SCOP, and that vary for the execution of the current basic block.
1387 Returns false if there is no loop in SCOP. */
1389 static void
1390 build_scop_iteration_domain (scop_p scop)
1392 struct loop *loop;
1393 sese region = SCOP_REGION (scop);
1394 int i;
1395 poly_bb_p pbb;
1396 int nb_loops = number_of_loops (cfun);
1397 isl_set **doms = XCNEWVEC (isl_set *, nb_loops);
1399 FOR_EACH_VEC_ELT (SESE_LOOP_NEST (region), i, loop)
1400 if (!loop_in_sese_p (loop_outer (loop), region))
1401 build_loop_iteration_domains (scop, loop, 0,
1402 isl_set_copy (scop->context), doms);
1404 FOR_EACH_VEC_ELT (SCOP_BBS (scop), i, pbb)
1406 loop = pbb_loop (pbb);
1408 if (doms[loop->num])
1409 pbb->domain = isl_set_copy (doms[loop->num]);
1410 else
1411 pbb->domain = isl_set_copy (scop->context);
1413 pbb->domain = isl_set_set_tuple_id (pbb->domain,
1414 isl_id_for_pbb (scop, pbb));
1417 for (i = 0; i < nb_loops; i++)
1418 if (doms[i])
1419 isl_set_free (doms[i]);
1421 free (doms);
1424 /* Add a constrain to the ACCESSES polyhedron for the alias set of
1425 data reference DR. ACCESSP_NB_DIMS is the dimension of the
1426 ACCESSES polyhedron, DOM_NB_DIMS is the dimension of the iteration
1427 domain. */
1429 static isl_map *
1430 pdr_add_alias_set (isl_map *acc, data_reference_p dr)
1432 isl_constraint *c;
1433 int alias_set_num = 0;
1434 base_alias_pair *bap = (base_alias_pair *)(dr->aux);
1436 if (bap && bap->alias_set)
1437 alias_set_num = *(bap->alias_set);
1439 c = isl_equality_alloc
1440 (isl_local_space_from_space (isl_map_get_space (acc)));
1441 c = isl_constraint_set_constant_si (c, -alias_set_num);
1442 c = isl_constraint_set_coefficient_si (c, isl_dim_out, 0, 1);
1444 return isl_map_add_constraint (acc, c);
1447 /* Assign the affine expression INDEX to the output dimension POS of
1448 MAP and return the result. */
1450 static isl_map *
1451 set_index (isl_map *map, int pos, isl_pw_aff *index)
1453 isl_map *index_map;
1454 int len = isl_map_dim (map, isl_dim_out);
1455 isl_id *id;
1457 index_map = isl_map_from_pw_aff (index);
1458 index_map = isl_map_insert_dims (index_map, isl_dim_out, 0, pos);
1459 index_map = isl_map_add_dims (index_map, isl_dim_out, len - pos - 1);
1461 id = isl_map_get_tuple_id (map, isl_dim_out);
1462 index_map = isl_map_set_tuple_id (index_map, isl_dim_out, id);
1463 id = isl_map_get_tuple_id (map, isl_dim_in);
1464 index_map = isl_map_set_tuple_id (index_map, isl_dim_in, id);
1466 return isl_map_intersect (map, index_map);
1469 /* Add to ACCESSES polyhedron equalities defining the access functions
1470 to the memory. ACCESSP_NB_DIMS is the dimension of the ACCESSES
1471 polyhedron, DOM_NB_DIMS is the dimension of the iteration domain.
1472 PBB is the poly_bb_p that contains the data reference DR. */
1474 static isl_map *
1475 pdr_add_memory_accesses (isl_map *acc, data_reference_p dr, poly_bb_p pbb)
1477 int i, nb_subscripts = DR_NUM_DIMENSIONS (dr);
1478 scop_p scop = PBB_SCOP (pbb);
1480 for (i = 0; i < nb_subscripts; i++)
1482 isl_pw_aff *aff;
1483 tree afn = DR_ACCESS_FN (dr, nb_subscripts - 1 - i);
1485 aff = extract_affine (scop, afn,
1486 isl_space_domain (isl_map_get_space (acc)));
1487 acc = set_index (acc, i + 1, aff);
1490 return acc;
1493 /* Add constrains representing the size of the accessed data to the
1494 ACCESSES polyhedron. ACCESSP_NB_DIMS is the dimension of the
1495 ACCESSES polyhedron, DOM_NB_DIMS is the dimension of the iteration
1496 domain. */
1498 static isl_set *
1499 pdr_add_data_dimensions (isl_set *extent, scop_p scop, data_reference_p dr)
1501 tree ref = DR_REF (dr);
1502 int i, nb_subscripts = DR_NUM_DIMENSIONS (dr);
1504 for (i = nb_subscripts - 1; i >= 0; i--, ref = TREE_OPERAND (ref, 0))
1506 tree low, high;
1508 if (TREE_CODE (ref) != ARRAY_REF)
1509 break;
1511 low = array_ref_low_bound (ref);
1512 high = array_ref_up_bound (ref);
1514 /* XXX The PPL code dealt separately with
1515 subscript - low >= 0 and high - subscript >= 0 in case one of
1516 the two bounds isn't known. Do the same here? */
1518 if (tree_fits_shwi_p (low)
1519 && high
1520 && tree_fits_shwi_p (high)
1521 /* 1-element arrays at end of structures may extend over
1522 their declared size. */
1523 && !(array_at_struct_end_p (ref)
1524 && operand_equal_p (low, high, 0)))
1526 isl_id *id;
1527 isl_aff *aff;
1528 isl_set *univ, *lbs, *ubs;
1529 isl_pw_aff *index;
1530 isl_space *space;
1531 isl_set *valid;
1532 isl_pw_aff *lb = extract_affine_int (low, isl_set_get_space (extent));
1533 isl_pw_aff *ub = extract_affine_int (high, isl_set_get_space (extent));
1535 /* high >= 0 */
1536 valid = isl_pw_aff_nonneg_set (isl_pw_aff_copy (ub));
1537 valid = isl_set_project_out (valid, isl_dim_set, 0,
1538 isl_set_dim (valid, isl_dim_set));
1539 scop->context = isl_set_intersect (scop->context, valid);
1541 space = isl_set_get_space (extent);
1542 aff = isl_aff_zero_on_domain (isl_local_space_from_space (space));
1543 aff = isl_aff_add_coefficient_si (aff, isl_dim_in, i + 1, 1);
1544 univ = isl_set_universe (isl_space_domain (isl_aff_get_space (aff)));
1545 index = isl_pw_aff_alloc (univ, aff);
1547 id = isl_set_get_tuple_id (extent);
1548 lb = isl_pw_aff_set_tuple_id (lb, isl_dim_in, isl_id_copy (id));
1549 ub = isl_pw_aff_set_tuple_id (ub, isl_dim_in, id);
1551 /* low <= sub_i <= high */
1552 lbs = isl_pw_aff_ge_set (isl_pw_aff_copy (index), lb);
1553 ubs = isl_pw_aff_le_set (index, ub);
1554 extent = isl_set_intersect (extent, lbs);
1555 extent = isl_set_intersect (extent, ubs);
1559 return extent;
1562 /* Build data accesses for DR in PBB. */
1564 static void
1565 build_poly_dr (data_reference_p dr, poly_bb_p pbb)
1567 int dr_base_object_set;
1568 isl_map *acc;
1569 isl_set *extent;
1570 scop_p scop = PBB_SCOP (pbb);
1573 isl_space *dc = isl_set_get_space (pbb->domain);
1574 int nb_out = 1 + DR_NUM_DIMENSIONS (dr);
1575 isl_space *space = isl_space_add_dims (isl_space_from_domain (dc),
1576 isl_dim_out, nb_out);
1578 acc = isl_map_universe (space);
1579 acc = isl_map_set_tuple_id (acc, isl_dim_out, isl_id_for_dr (scop, dr));
1582 acc = pdr_add_alias_set (acc, dr);
1583 acc = pdr_add_memory_accesses (acc, dr, pbb);
1586 isl_id *id = isl_id_for_dr (scop, dr);
1587 int nb = 1 + DR_NUM_DIMENSIONS (dr);
1588 isl_space *space = isl_space_set_alloc (scop->ctx, 0, nb);
1589 int alias_set_num = 0;
1590 base_alias_pair *bap = (base_alias_pair *)(dr->aux);
1592 if (bap && bap->alias_set)
1593 alias_set_num = *(bap->alias_set);
1595 space = isl_space_set_tuple_id (space, isl_dim_set, id);
1596 extent = isl_set_nat_universe (space);
1597 extent = isl_set_fix_si (extent, isl_dim_set, 0, alias_set_num);
1598 extent = pdr_add_data_dimensions (extent, scop, dr);
1601 gcc_assert (dr->aux);
1602 dr_base_object_set = ((base_alias_pair *)(dr->aux))->base_obj_set;
1604 new_poly_dr (pbb, dr_base_object_set,
1605 DR_IS_READ (dr) ? PDR_READ : PDR_WRITE,
1606 dr, DR_NUM_DIMENSIONS (dr), acc, extent);
1609 /* Write to FILE the alias graph of data references in DIMACS format. */
1611 static inline bool
1612 write_alias_graph_to_ascii_dimacs (FILE *file, char *comment,
1613 vec<data_reference_p> drs)
1615 int num_vertex = drs.length ();
1616 int edge_num = 0;
1617 data_reference_p dr1, dr2;
1618 int i, j;
1620 if (num_vertex == 0)
1621 return true;
1623 FOR_EACH_VEC_ELT (drs, i, dr1)
1624 for (j = i + 1; drs.iterate (j, &dr2); j++)
1625 if (dr_may_alias_p (dr1, dr2, true))
1626 edge_num++;
1628 fprintf (file, "$\n");
1630 if (comment)
1631 fprintf (file, "c %s\n", comment);
1633 fprintf (file, "p edge %d %d\n", num_vertex, edge_num);
1635 FOR_EACH_VEC_ELT (drs, i, dr1)
1636 for (j = i + 1; drs.iterate (j, &dr2); j++)
1637 if (dr_may_alias_p (dr1, dr2, true))
1638 fprintf (file, "e %d %d\n", i + 1, j + 1);
1640 return true;
1643 /* Write to FILE the alias graph of data references in DOT format. */
1645 static inline bool
1646 write_alias_graph_to_ascii_dot (FILE *file, char *comment,
1647 vec<data_reference_p> drs)
1649 int num_vertex = drs.length ();
1650 data_reference_p dr1, dr2;
1651 int i, j;
1653 if (num_vertex == 0)
1654 return true;
1656 fprintf (file, "$\n");
1658 if (comment)
1659 fprintf (file, "c %s\n", comment);
1661 /* First print all the vertices. */
1662 FOR_EACH_VEC_ELT (drs, i, dr1)
1663 fprintf (file, "n%d;\n", i);
1665 FOR_EACH_VEC_ELT (drs, i, dr1)
1666 for (j = i + 1; drs.iterate (j, &dr2); j++)
1667 if (dr_may_alias_p (dr1, dr2, true))
1668 fprintf (file, "n%d n%d\n", i, j);
1670 return true;
1673 /* Write to FILE the alias graph of data references in ECC format. */
1675 static inline bool
1676 write_alias_graph_to_ascii_ecc (FILE *file, char *comment,
1677 vec<data_reference_p> drs)
1679 int num_vertex = drs.length ();
1680 data_reference_p dr1, dr2;
1681 int i, j;
1683 if (num_vertex == 0)
1684 return true;
1686 fprintf (file, "$\n");
1688 if (comment)
1689 fprintf (file, "c %s\n", comment);
1691 FOR_EACH_VEC_ELT (drs, i, dr1)
1692 for (j = i + 1; drs.iterate (j, &dr2); j++)
1693 if (dr_may_alias_p (dr1, dr2, true))
1694 fprintf (file, "%d %d\n", i, j);
1696 return true;
1699 /* Check if DR1 and DR2 are in the same object set. */
1701 static bool
1702 dr_same_base_object_p (const struct data_reference *dr1,
1703 const struct data_reference *dr2)
1705 return operand_equal_p (DR_BASE_OBJECT (dr1), DR_BASE_OBJECT (dr2), 0);
1708 /* Uses DFS component number as representative of alias-sets. Also tests for
1709 optimality by verifying if every connected component is a clique. Returns
1710 true (1) if the above test is true, and false (0) otherwise. */
1712 static int
1713 build_alias_set_optimal_p (vec<data_reference_p> drs)
1715 int num_vertices = drs.length ();
1716 struct graph *g = new_graph (num_vertices);
1717 data_reference_p dr1, dr2;
1718 int i, j;
1719 int num_connected_components;
1720 int v_indx1, v_indx2, num_vertices_in_component;
1721 int *all_vertices;
1722 int *vertices;
1723 struct graph_edge *e;
1724 int this_component_is_clique;
1725 int all_components_are_cliques = 1;
1727 FOR_EACH_VEC_ELT (drs, i, dr1)
1728 for (j = i+1; drs.iterate (j, &dr2); j++)
1729 if (dr_may_alias_p (dr1, dr2, true))
1731 add_edge (g, i, j);
1732 add_edge (g, j, i);
1735 all_vertices = XNEWVEC (int, num_vertices);
1736 vertices = XNEWVEC (int, num_vertices);
1737 for (i = 0; i < num_vertices; i++)
1738 all_vertices[i] = i;
1740 num_connected_components = graphds_dfs (g, all_vertices, num_vertices,
1741 NULL, true, NULL);
1742 for (i = 0; i < g->n_vertices; i++)
1744 data_reference_p dr = drs[i];
1745 base_alias_pair *bap;
1747 gcc_assert (dr->aux);
1748 bap = (base_alias_pair *)(dr->aux);
1750 bap->alias_set = XNEW (int);
1751 *(bap->alias_set) = g->vertices[i].component + 1;
1754 /* Verify if the DFS numbering results in optimal solution. */
1755 for (i = 0; i < num_connected_components; i++)
1757 num_vertices_in_component = 0;
1758 /* Get all vertices whose DFS component number is the same as i. */
1759 for (j = 0; j < num_vertices; j++)
1760 if (g->vertices[j].component == i)
1761 vertices[num_vertices_in_component++] = j;
1763 /* Now test if the vertices in 'vertices' form a clique, by testing
1764 for edges among each pair. */
1765 this_component_is_clique = 1;
1766 for (v_indx1 = 0; v_indx1 < num_vertices_in_component; v_indx1++)
1768 for (v_indx2 = v_indx1+1; v_indx2 < num_vertices_in_component; v_indx2++)
1770 /* Check if the two vertices are connected by iterating
1771 through all the edges which have one of these are source. */
1772 e = g->vertices[vertices[v_indx2]].pred;
1773 while (e)
1775 if (e->src == vertices[v_indx1])
1776 break;
1777 e = e->pred_next;
1779 if (!e)
1781 this_component_is_clique = 0;
1782 break;
1785 if (!this_component_is_clique)
1786 all_components_are_cliques = 0;
1790 free (all_vertices);
1791 free (vertices);
1792 free_graph (g);
1793 return all_components_are_cliques;
1796 /* Group each data reference in DRS with its base object set num. */
1798 static void
1799 build_base_obj_set_for_drs (vec<data_reference_p> drs)
1801 int num_vertex = drs.length ();
1802 struct graph *g = new_graph (num_vertex);
1803 data_reference_p dr1, dr2;
1804 int i, j;
1805 int *queue;
1807 FOR_EACH_VEC_ELT (drs, i, dr1)
1808 for (j = i + 1; drs.iterate (j, &dr2); j++)
1809 if (dr_same_base_object_p (dr1, dr2))
1811 add_edge (g, i, j);
1812 add_edge (g, j, i);
1815 queue = XNEWVEC (int, num_vertex);
1816 for (i = 0; i < num_vertex; i++)
1817 queue[i] = i;
1819 graphds_dfs (g, queue, num_vertex, NULL, true, NULL);
1821 for (i = 0; i < g->n_vertices; i++)
1823 data_reference_p dr = drs[i];
1824 base_alias_pair *bap;
1826 gcc_assert (dr->aux);
1827 bap = (base_alias_pair *)(dr->aux);
1829 bap->base_obj_set = g->vertices[i].component + 1;
1832 free (queue);
1833 free_graph (g);
1836 /* Build the data references for PBB. */
1838 static void
1839 build_pbb_drs (poly_bb_p pbb)
1841 int j;
1842 data_reference_p dr;
1843 vec<data_reference_p> gbb_drs = GBB_DATA_REFS (PBB_BLACK_BOX (pbb));
1845 FOR_EACH_VEC_ELT (gbb_drs, j, dr)
1846 build_poly_dr (dr, pbb);
1849 /* Dump to file the alias graphs for the data references in DRS. */
1851 static void
1852 dump_alias_graphs (vec<data_reference_p> drs)
1854 char comment[100];
1855 FILE *file_dimacs, *file_ecc, *file_dot;
1857 file_dimacs = fopen ("/tmp/dr_alias_graph_dimacs", "ab");
1858 if (file_dimacs)
1860 snprintf (comment, sizeof (comment), "%s %s", main_input_filename,
1861 current_function_name ());
1862 write_alias_graph_to_ascii_dimacs (file_dimacs, comment, drs);
1863 fclose (file_dimacs);
1866 file_ecc = fopen ("/tmp/dr_alias_graph_ecc", "ab");
1867 if (file_ecc)
1869 snprintf (comment, sizeof (comment), "%s %s", main_input_filename,
1870 current_function_name ());
1871 write_alias_graph_to_ascii_ecc (file_ecc, comment, drs);
1872 fclose (file_ecc);
1875 file_dot = fopen ("/tmp/dr_alias_graph_dot", "ab");
1876 if (file_dot)
1878 snprintf (comment, sizeof (comment), "%s %s", main_input_filename,
1879 current_function_name ());
1880 write_alias_graph_to_ascii_dot (file_dot, comment, drs);
1881 fclose (file_dot);
1885 /* Build data references in SCOP. */
1887 static void
1888 build_scop_drs (scop_p scop)
1890 int i, j;
1891 poly_bb_p pbb;
1892 data_reference_p dr;
1893 auto_vec<data_reference_p, 3> drs;
1895 /* Remove all the PBBs that do not have data references: these basic
1896 blocks are not handled in the polyhedral representation. */
1897 for (i = 0; SCOP_BBS (scop).iterate (i, &pbb); i++)
1898 if (GBB_DATA_REFS (PBB_BLACK_BOX (pbb)).is_empty ())
1900 free_gimple_bb (PBB_BLACK_BOX (pbb));
1901 free_poly_bb (pbb);
1902 SCOP_BBS (scop).ordered_remove (i);
1903 i--;
1906 FOR_EACH_VEC_ELT (SCOP_BBS (scop), i, pbb)
1907 for (j = 0; GBB_DATA_REFS (PBB_BLACK_BOX (pbb)).iterate (j, &dr); j++)
1908 drs.safe_push (dr);
1910 FOR_EACH_VEC_ELT (drs, i, dr)
1911 dr->aux = XNEW (base_alias_pair);
1913 if (!build_alias_set_optimal_p (drs))
1915 /* TODO: Add support when building alias set is not optimal. */
1919 build_base_obj_set_for_drs (drs);
1921 /* When debugging, enable the following code. This cannot be used
1922 in production compilers. */
1923 if (0)
1924 dump_alias_graphs (drs);
1926 drs.release ();
1928 FOR_EACH_VEC_ELT (SCOP_BBS (scop), i, pbb)
1929 build_pbb_drs (pbb);
1932 /* Return a gsi at the position of the phi node STMT. */
1934 static gphi_iterator
1935 gsi_for_phi_node (gphi *stmt)
1937 gphi_iterator psi;
1938 basic_block bb = gimple_bb (stmt);
1940 for (psi = gsi_start_phis (bb); !gsi_end_p (psi); gsi_next (&psi))
1941 if (stmt == psi.phi ())
1942 return psi;
1944 gcc_unreachable ();
1945 return psi;
1948 /* Analyze all the data references of STMTS and add them to the
1949 GBB_DATA_REFS vector of BB. */
1951 static void
1952 analyze_drs_in_stmts (scop_p scop, basic_block bb, vec<gimple> stmts)
1954 loop_p nest;
1955 gimple_bb_p gbb;
1956 gimple stmt;
1957 int i;
1958 sese region = SCOP_REGION (scop);
1960 if (!bb_in_sese_p (bb, region))
1961 return;
1963 nest = outermost_loop_in_sese_1 (region, bb);
1964 gbb = gbb_from_bb (bb);
1966 FOR_EACH_VEC_ELT (stmts, i, stmt)
1968 loop_p loop;
1970 if (is_gimple_debug (stmt))
1971 continue;
1973 loop = loop_containing_stmt (stmt);
1974 if (!loop_in_sese_p (loop, region))
1975 loop = nest;
1977 graphite_find_data_references_in_stmt (nest, loop, stmt,
1978 &GBB_DATA_REFS (gbb));
1982 /* Insert STMT at the end of the STMTS sequence and then insert the
1983 statements from STMTS at INSERT_GSI and call analyze_drs_in_stmts
1984 on STMTS. */
1986 static void
1987 insert_stmts (scop_p scop, gimple stmt, gimple_seq stmts,
1988 gimple_stmt_iterator insert_gsi)
1990 gimple_stmt_iterator gsi;
1991 auto_vec<gimple, 3> x;
1993 gimple_seq_add_stmt (&stmts, stmt);
1994 for (gsi = gsi_start (stmts); !gsi_end_p (gsi); gsi_next (&gsi))
1995 x.safe_push (gsi_stmt (gsi));
1997 gsi_insert_seq_before (&insert_gsi, stmts, GSI_SAME_STMT);
1998 analyze_drs_in_stmts (scop, gsi_bb (insert_gsi), x);
2001 /* Insert the assignment "RES := EXPR" just after AFTER_STMT. */
2003 static void
2004 insert_out_of_ssa_copy (scop_p scop, tree res, tree expr, gimple after_stmt)
2006 gimple_seq stmts;
2007 gimple_stmt_iterator gsi;
2008 tree var = force_gimple_operand (expr, &stmts, true, NULL_TREE);
2009 gassign *stmt = gimple_build_assign (unshare_expr (res), var);
2010 auto_vec<gimple, 3> x;
2012 gimple_seq_add_stmt (&stmts, stmt);
2013 for (gsi = gsi_start (stmts); !gsi_end_p (gsi); gsi_next (&gsi))
2014 x.safe_push (gsi_stmt (gsi));
2016 if (gimple_code (after_stmt) == GIMPLE_PHI)
2018 gsi = gsi_after_labels (gimple_bb (after_stmt));
2019 gsi_insert_seq_before (&gsi, stmts, GSI_NEW_STMT);
2021 else
2023 gsi = gsi_for_stmt (after_stmt);
2024 gsi_insert_seq_after (&gsi, stmts, GSI_NEW_STMT);
2027 analyze_drs_in_stmts (scop, gimple_bb (after_stmt), x);
2030 /* Creates a poly_bb_p for basic_block BB from the existing PBB. */
2032 static void
2033 new_pbb_from_pbb (scop_p scop, poly_bb_p pbb, basic_block bb)
2035 vec<data_reference_p> drs;
2036 drs.create (3);
2037 gimple_bb_p gbb = PBB_BLACK_BOX (pbb);
2038 gimple_bb_p gbb1 = new_gimple_bb (bb, drs);
2039 poly_bb_p pbb1 = new_poly_bb (scop, gbb1);
2040 int index, n = SCOP_BBS (scop).length ();
2042 /* The INDEX of PBB in SCOP_BBS. */
2043 for (index = 0; index < n; index++)
2044 if (SCOP_BBS (scop)[index] == pbb)
2045 break;
2047 pbb1->domain = isl_set_copy (pbb->domain);
2048 pbb1->domain = isl_set_set_tuple_id (pbb1->domain,
2049 isl_id_for_pbb (scop, pbb1));
2051 GBB_PBB (gbb1) = pbb1;
2052 GBB_CONDITIONS (gbb1) = GBB_CONDITIONS (gbb).copy ();
2053 GBB_CONDITION_CASES (gbb1) = GBB_CONDITION_CASES (gbb).copy ();
2054 SCOP_BBS (scop).safe_insert (index + 1, pbb1);
2057 /* Insert on edge E the assignment "RES := EXPR". */
2059 static void
2060 insert_out_of_ssa_copy_on_edge (scop_p scop, edge e, tree res, tree expr)
2062 gimple_stmt_iterator gsi;
2063 gimple_seq stmts = NULL;
2064 tree var = force_gimple_operand (expr, &stmts, true, NULL_TREE);
2065 gimple stmt = gimple_build_assign (unshare_expr (res), var);
2066 basic_block bb;
2067 auto_vec<gimple, 3> x;
2069 gimple_seq_add_stmt (&stmts, stmt);
2070 for (gsi = gsi_start (stmts); !gsi_end_p (gsi); gsi_next (&gsi))
2071 x.safe_push (gsi_stmt (gsi));
2073 gsi_insert_seq_on_edge (e, stmts);
2074 gsi_commit_edge_inserts ();
2075 bb = gimple_bb (stmt);
2077 if (!bb_in_sese_p (bb, SCOP_REGION (scop)))
2078 return;
2080 if (!gbb_from_bb (bb))
2081 new_pbb_from_pbb (scop, pbb_from_bb (e->src), bb);
2083 analyze_drs_in_stmts (scop, bb, x);
2086 /* Creates a zero dimension array of the same type as VAR. */
2088 static tree
2089 create_zero_dim_array (tree var, const char *base_name)
2091 tree index_type = build_index_type (integer_zero_node);
2092 tree elt_type = TREE_TYPE (var);
2093 tree array_type = build_array_type (elt_type, index_type);
2094 tree base = create_tmp_var (array_type, base_name);
2096 return build4 (ARRAY_REF, elt_type, base, integer_zero_node, NULL_TREE,
2097 NULL_TREE);
2100 /* Returns true when PHI is a loop close phi node. */
2102 static bool
2103 scalar_close_phi_node_p (gimple phi)
2105 if (gimple_code (phi) != GIMPLE_PHI
2106 || virtual_operand_p (gimple_phi_result (phi)))
2107 return false;
2109 /* Note that loop close phi nodes should have a single argument
2110 because we translated the representation into a canonical form
2111 before Graphite: see canonicalize_loop_closed_ssa_form. */
2112 return (gimple_phi_num_args (phi) == 1);
2115 /* For a definition DEF in REGION, propagates the expression EXPR in
2116 all the uses of DEF outside REGION. */
2118 static void
2119 propagate_expr_outside_region (tree def, tree expr, sese region)
2121 imm_use_iterator imm_iter;
2122 gimple use_stmt;
2123 gimple_seq stmts;
2124 bool replaced_once = false;
2126 gcc_assert (TREE_CODE (def) == SSA_NAME);
2128 expr = force_gimple_operand (unshare_expr (expr), &stmts, true,
2129 NULL_TREE);
2131 FOR_EACH_IMM_USE_STMT (use_stmt, imm_iter, def)
2132 if (!is_gimple_debug (use_stmt)
2133 && !bb_in_sese_p (gimple_bb (use_stmt), region))
2135 ssa_op_iter iter;
2136 use_operand_p use_p;
2138 FOR_EACH_PHI_OR_STMT_USE (use_p, use_stmt, iter, SSA_OP_ALL_USES)
2139 if (operand_equal_p (def, USE_FROM_PTR (use_p), 0)
2140 && (replaced_once = true))
2141 replace_exp (use_p, expr);
2143 update_stmt (use_stmt);
2146 if (replaced_once)
2148 gsi_insert_seq_on_edge (SESE_ENTRY (region), stmts);
2149 gsi_commit_edge_inserts ();
2153 /* Rewrite out of SSA the reduction phi node at PSI by creating a zero
2154 dimension array for it. */
2156 static void
2157 rewrite_close_phi_out_of_ssa (scop_p scop, gimple_stmt_iterator *psi)
2159 sese region = SCOP_REGION (scop);
2160 gimple phi = gsi_stmt (*psi);
2161 tree res = gimple_phi_result (phi);
2162 basic_block bb = gimple_bb (phi);
2163 gimple_stmt_iterator gsi = gsi_after_labels (bb);
2164 tree arg = gimple_phi_arg_def (phi, 0);
2165 gimple stmt;
2167 /* Note that loop close phi nodes should have a single argument
2168 because we translated the representation into a canonical form
2169 before Graphite: see canonicalize_loop_closed_ssa_form. */
2170 gcc_assert (gimple_phi_num_args (phi) == 1);
2172 /* The phi node can be a non close phi node, when its argument is
2173 invariant, or a default definition. */
2174 if (is_gimple_min_invariant (arg)
2175 || SSA_NAME_IS_DEFAULT_DEF (arg))
2177 propagate_expr_outside_region (res, arg, region);
2178 gsi_next (psi);
2179 return;
2182 else if (gimple_bb (SSA_NAME_DEF_STMT (arg))->loop_father == bb->loop_father)
2184 propagate_expr_outside_region (res, arg, region);
2185 stmt = gimple_build_assign (res, arg);
2186 remove_phi_node (psi, false);
2187 gsi_insert_before (&gsi, stmt, GSI_NEW_STMT);
2188 return;
2191 /* If res is scev analyzable and is not a scalar value, it is safe
2192 to ignore the close phi node: it will be code generated in the
2193 out of Graphite pass. */
2194 else if (scev_analyzable_p (res, region))
2196 loop_p loop = loop_containing_stmt (SSA_NAME_DEF_STMT (res));
2197 tree scev;
2199 if (!loop_in_sese_p (loop, region))
2201 loop = loop_containing_stmt (SSA_NAME_DEF_STMT (arg));
2202 scev = scalar_evolution_in_region (region, loop, arg);
2203 scev = compute_overall_effect_of_inner_loop (loop, scev);
2205 else
2206 scev = scalar_evolution_in_region (region, loop, res);
2208 if (tree_does_not_contain_chrecs (scev))
2209 propagate_expr_outside_region (res, scev, region);
2211 gsi_next (psi);
2212 return;
2214 else
2216 tree zero_dim_array = create_zero_dim_array (res, "Close_Phi");
2218 stmt = gimple_build_assign (res, unshare_expr (zero_dim_array));
2220 if (TREE_CODE (arg) == SSA_NAME)
2221 insert_out_of_ssa_copy (scop, zero_dim_array, arg,
2222 SSA_NAME_DEF_STMT (arg));
2223 else
2224 insert_out_of_ssa_copy_on_edge (scop, single_pred_edge (bb),
2225 zero_dim_array, arg);
2228 remove_phi_node (psi, false);
2229 SSA_NAME_DEF_STMT (res) = stmt;
2231 insert_stmts (scop, stmt, NULL, gsi_after_labels (bb));
2234 /* Rewrite out of SSA the reduction phi node at PSI by creating a zero
2235 dimension array for it. */
2237 static void
2238 rewrite_phi_out_of_ssa (scop_p scop, gphi_iterator *psi)
2240 size_t i;
2241 gphi *phi = psi->phi ();
2242 basic_block bb = gimple_bb (phi);
2243 tree res = gimple_phi_result (phi);
2244 tree zero_dim_array = create_zero_dim_array (res, "phi_out_of_ssa");
2245 gimple stmt;
2247 for (i = 0; i < gimple_phi_num_args (phi); i++)
2249 tree arg = gimple_phi_arg_def (phi, i);
2250 edge e = gimple_phi_arg_edge (phi, i);
2252 /* Avoid the insertion of code in the loop latch to please the
2253 pattern matching of the vectorizer. */
2254 if (TREE_CODE (arg) == SSA_NAME
2255 && !SSA_NAME_IS_DEFAULT_DEF (arg)
2256 && e->src == bb->loop_father->latch)
2257 insert_out_of_ssa_copy (scop, zero_dim_array, arg,
2258 SSA_NAME_DEF_STMT (arg));
2259 else
2260 insert_out_of_ssa_copy_on_edge (scop, e, zero_dim_array, arg);
2263 stmt = gimple_build_assign (res, unshare_expr (zero_dim_array));
2264 remove_phi_node (psi, false);
2265 insert_stmts (scop, stmt, NULL, gsi_after_labels (bb));
2268 /* Rewrite the degenerate phi node at position PSI from the degenerate
2269 form "x = phi (y, y, ..., y)" to "x = y". */
2271 static void
2272 rewrite_degenerate_phi (gphi_iterator *psi)
2274 tree rhs;
2275 gimple stmt;
2276 gimple_stmt_iterator gsi;
2277 gphi *phi = psi->phi ();
2278 tree res = gimple_phi_result (phi);
2279 basic_block bb;
2281 bb = gimple_bb (phi);
2282 rhs = degenerate_phi_result (phi);
2283 gcc_assert (rhs);
2285 stmt = gimple_build_assign (res, rhs);
2286 remove_phi_node (psi, false);
2288 gsi = gsi_after_labels (bb);
2289 gsi_insert_before (&gsi, stmt, GSI_NEW_STMT);
2292 /* Rewrite out of SSA all the reduction phi nodes of SCOP. */
2294 static void
2295 rewrite_reductions_out_of_ssa (scop_p scop)
2297 basic_block bb;
2298 gphi_iterator psi;
2299 sese region = SCOP_REGION (scop);
2301 FOR_EACH_BB_FN (bb, cfun)
2302 if (bb_in_sese_p (bb, region))
2303 for (psi = gsi_start_phis (bb); !gsi_end_p (psi);)
2305 gphi *phi = psi.phi ();
2307 if (virtual_operand_p (gimple_phi_result (phi)))
2309 gsi_next (&psi);
2310 continue;
2313 if (gimple_phi_num_args (phi) > 1
2314 && degenerate_phi_result (phi))
2315 rewrite_degenerate_phi (&psi);
2317 else if (scalar_close_phi_node_p (phi))
2318 rewrite_close_phi_out_of_ssa (scop, &psi);
2320 else if (reduction_phi_p (region, &psi))
2321 rewrite_phi_out_of_ssa (scop, &psi);
2324 update_ssa (TODO_update_ssa);
2325 #ifdef ENABLE_CHECKING
2326 verify_loop_closed_ssa (true);
2327 #endif
2330 /* Rewrite the scalar dependence of DEF used in USE_STMT with a memory
2331 read from ZERO_DIM_ARRAY. */
2333 static void
2334 rewrite_cross_bb_scalar_dependence (scop_p scop, tree zero_dim_array,
2335 tree def, gimple use_stmt)
2337 gimple name_stmt;
2338 tree name;
2339 ssa_op_iter iter;
2340 use_operand_p use_p;
2342 gcc_assert (gimple_code (use_stmt) != GIMPLE_PHI);
2344 name = copy_ssa_name (def, NULL);
2345 name_stmt = gimple_build_assign (name, zero_dim_array);
2347 gimple_assign_set_lhs (name_stmt, name);
2348 insert_stmts (scop, name_stmt, NULL, gsi_for_stmt (use_stmt));
2350 FOR_EACH_SSA_USE_OPERAND (use_p, use_stmt, iter, SSA_OP_ALL_USES)
2351 if (operand_equal_p (def, USE_FROM_PTR (use_p), 0))
2352 replace_exp (use_p, name);
2354 update_stmt (use_stmt);
2357 /* For every definition DEF in the SCOP that is used outside the scop,
2358 insert a closing-scop definition in the basic block just after this
2359 SCOP. */
2361 static void
2362 handle_scalar_deps_crossing_scop_limits (scop_p scop, tree def, gimple stmt)
2364 tree var = create_tmp_reg (TREE_TYPE (def), NULL);
2365 tree new_name = make_ssa_name (var, stmt);
2366 bool needs_copy = false;
2367 use_operand_p use_p;
2368 imm_use_iterator imm_iter;
2369 gimple use_stmt;
2370 sese region = SCOP_REGION (scop);
2372 FOR_EACH_IMM_USE_STMT (use_stmt, imm_iter, def)
2374 if (!bb_in_sese_p (gimple_bb (use_stmt), region))
2376 FOR_EACH_IMM_USE_ON_STMT (use_p, imm_iter)
2378 SET_USE (use_p, new_name);
2380 update_stmt (use_stmt);
2381 needs_copy = true;
2385 /* Insert in the empty BB just after the scop a use of DEF such
2386 that the rewrite of cross_bb_scalar_dependences won't insert
2387 arrays everywhere else. */
2388 if (needs_copy)
2390 gimple assign = gimple_build_assign (new_name, def);
2391 gimple_stmt_iterator psi = gsi_after_labels (SESE_EXIT (region)->dest);
2393 update_stmt (assign);
2394 gsi_insert_before (&psi, assign, GSI_SAME_STMT);
2398 /* Rewrite the scalar dependences crossing the boundary of the BB
2399 containing STMT with an array. Return true when something has been
2400 changed. */
2402 static bool
2403 rewrite_cross_bb_scalar_deps (scop_p scop, gimple_stmt_iterator *gsi)
2405 sese region = SCOP_REGION (scop);
2406 gimple stmt = gsi_stmt (*gsi);
2407 imm_use_iterator imm_iter;
2408 tree def;
2409 basic_block def_bb;
2410 tree zero_dim_array = NULL_TREE;
2411 gimple use_stmt;
2412 bool res = false;
2414 switch (gimple_code (stmt))
2416 case GIMPLE_ASSIGN:
2417 def = gimple_assign_lhs (stmt);
2418 break;
2420 case GIMPLE_CALL:
2421 def = gimple_call_lhs (stmt);
2422 break;
2424 default:
2425 return false;
2428 if (!def
2429 || !is_gimple_reg (def))
2430 return false;
2432 if (scev_analyzable_p (def, region))
2434 loop_p loop = loop_containing_stmt (SSA_NAME_DEF_STMT (def));
2435 tree scev = scalar_evolution_in_region (region, loop, def);
2437 if (tree_contains_chrecs (scev, NULL))
2438 return false;
2440 propagate_expr_outside_region (def, scev, region);
2441 return true;
2444 def_bb = gimple_bb (stmt);
2446 handle_scalar_deps_crossing_scop_limits (scop, def, stmt);
2448 FOR_EACH_IMM_USE_STMT (use_stmt, imm_iter, def)
2449 if (gimple_code (use_stmt) == GIMPLE_PHI
2450 && (res = true))
2452 gphi_iterator psi = gsi_start_phis (gimple_bb (use_stmt));
2454 if (scalar_close_phi_node_p (gsi_stmt (psi)))
2455 rewrite_close_phi_out_of_ssa (scop, &psi);
2456 else
2457 rewrite_phi_out_of_ssa (scop, &psi);
2460 FOR_EACH_IMM_USE_STMT (use_stmt, imm_iter, def)
2461 if (gimple_code (use_stmt) != GIMPLE_PHI
2462 && def_bb != gimple_bb (use_stmt)
2463 && !is_gimple_debug (use_stmt)
2464 && (res = true))
2466 if (!zero_dim_array)
2468 zero_dim_array = create_zero_dim_array
2469 (def, "Cross_BB_scalar_dependence");
2470 insert_out_of_ssa_copy (scop, zero_dim_array, def,
2471 SSA_NAME_DEF_STMT (def));
2472 gsi_next (gsi);
2475 rewrite_cross_bb_scalar_dependence (scop, unshare_expr (zero_dim_array),
2476 def, use_stmt);
2479 return res;
2482 /* Rewrite out of SSA all the reduction phi nodes of SCOP. */
2484 static void
2485 rewrite_cross_bb_scalar_deps_out_of_ssa (scop_p scop)
2487 basic_block bb;
2488 gimple_stmt_iterator psi;
2489 sese region = SCOP_REGION (scop);
2490 bool changed = false;
2492 /* Create an extra empty BB after the scop. */
2493 split_edge (SESE_EXIT (region));
2495 FOR_EACH_BB_FN (bb, cfun)
2496 if (bb_in_sese_p (bb, region))
2497 for (psi = gsi_start_bb (bb); !gsi_end_p (psi); gsi_next (&psi))
2498 changed |= rewrite_cross_bb_scalar_deps (scop, &psi);
2500 if (changed)
2502 scev_reset_htab ();
2503 update_ssa (TODO_update_ssa);
2504 #ifdef ENABLE_CHECKING
2505 verify_loop_closed_ssa (true);
2506 #endif
2510 /* Returns the number of pbbs that are in loops contained in SCOP. */
2512 static int
2513 nb_pbbs_in_loops (scop_p scop)
2515 int i;
2516 poly_bb_p pbb;
2517 int res = 0;
2519 FOR_EACH_VEC_ELT (SCOP_BBS (scop), i, pbb)
2520 if (loop_in_sese_p (gbb_loop (PBB_BLACK_BOX (pbb)), SCOP_REGION (scop)))
2521 res++;
2523 return res;
2526 /* Return the number of data references in BB that write in
2527 memory. */
2529 static int
2530 nb_data_writes_in_bb (basic_block bb)
2532 int res = 0;
2533 gimple_stmt_iterator gsi;
2535 for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi))
2536 if (gimple_vdef (gsi_stmt (gsi)))
2537 res++;
2539 return res;
2542 /* Splits at STMT the basic block BB represented as PBB in the
2543 polyhedral form. */
2545 static edge
2546 split_pbb (scop_p scop, poly_bb_p pbb, basic_block bb, gimple stmt)
2548 edge e1 = split_block (bb, stmt);
2549 new_pbb_from_pbb (scop, pbb, e1->dest);
2550 return e1;
2553 /* Splits STMT out of its current BB. This is done for reduction
2554 statements for which we want to ignore data dependences. */
2556 static basic_block
2557 split_reduction_stmt (scop_p scop, gimple stmt)
2559 basic_block bb = gimple_bb (stmt);
2560 poly_bb_p pbb = pbb_from_bb (bb);
2561 gimple_bb_p gbb = gbb_from_bb (bb);
2562 edge e1;
2563 int i;
2564 data_reference_p dr;
2566 /* Do not split basic blocks with no writes to memory: the reduction
2567 will be the only write to memory. */
2568 if (nb_data_writes_in_bb (bb) == 0
2569 /* Or if we have already marked BB as a reduction. */
2570 || PBB_IS_REDUCTION (pbb_from_bb (bb)))
2571 return bb;
2573 e1 = split_pbb (scop, pbb, bb, stmt);
2575 /* Split once more only when the reduction stmt is not the only one
2576 left in the original BB. */
2577 if (!gsi_one_before_end_p (gsi_start_nondebug_bb (bb)))
2579 gimple_stmt_iterator gsi = gsi_last_bb (bb);
2580 gsi_prev (&gsi);
2581 e1 = split_pbb (scop, pbb, bb, gsi_stmt (gsi));
2584 /* A part of the data references will end in a different basic block
2585 after the split: move the DRs from the original GBB to the newly
2586 created GBB1. */
2587 FOR_EACH_VEC_ELT (GBB_DATA_REFS (gbb), i, dr)
2589 basic_block bb1 = gimple_bb (DR_STMT (dr));
2591 if (bb1 != bb)
2593 gimple_bb_p gbb1 = gbb_from_bb (bb1);
2594 GBB_DATA_REFS (gbb1).safe_push (dr);
2595 GBB_DATA_REFS (gbb).ordered_remove (i);
2596 i--;
2600 return e1->dest;
2603 /* Return true when stmt is a reduction operation. */
2605 static inline bool
2606 is_reduction_operation_p (gimple stmt)
2608 enum tree_code code;
2610 gcc_assert (is_gimple_assign (stmt));
2611 code = gimple_assign_rhs_code (stmt);
2613 return flag_associative_math
2614 && commutative_tree_code (code)
2615 && associative_tree_code (code);
2618 /* Returns true when PHI contains an argument ARG. */
2620 static bool
2621 phi_contains_arg (gphi *phi, tree arg)
2623 size_t i;
2625 for (i = 0; i < gimple_phi_num_args (phi); i++)
2626 if (operand_equal_p (arg, gimple_phi_arg_def (phi, i), 0))
2627 return true;
2629 return false;
2632 /* Return a loop phi node that corresponds to a reduction containing LHS. */
2634 static gphi *
2635 follow_ssa_with_commutative_ops (tree arg, tree lhs)
2637 gimple stmt;
2639 if (TREE_CODE (arg) != SSA_NAME)
2640 return NULL;
2642 stmt = SSA_NAME_DEF_STMT (arg);
2644 if (gimple_code (stmt) == GIMPLE_NOP
2645 || gimple_code (stmt) == GIMPLE_CALL)
2646 return NULL;
2648 if (gphi *phi = dyn_cast <gphi *> (stmt))
2650 if (phi_contains_arg (phi, lhs))
2651 return phi;
2652 return NULL;
2655 if (!is_gimple_assign (stmt))
2656 return NULL;
2658 if (gimple_num_ops (stmt) == 2)
2659 return follow_ssa_with_commutative_ops (gimple_assign_rhs1 (stmt), lhs);
2661 if (is_reduction_operation_p (stmt))
2663 gphi *res =
2664 follow_ssa_with_commutative_ops (gimple_assign_rhs1 (stmt), lhs);
2666 return res ? res :
2667 follow_ssa_with_commutative_ops (gimple_assign_rhs2 (stmt), lhs);
2670 return NULL;
2673 /* Detect commutative and associative scalar reductions starting at
2674 the STMT. Return the phi node of the reduction cycle, or NULL. */
2676 static gphi *
2677 detect_commutative_reduction_arg (tree lhs, gimple stmt, tree arg,
2678 vec<gimple> *in,
2679 vec<gimple> *out)
2681 gphi *phi = follow_ssa_with_commutative_ops (arg, lhs);
2683 if (!phi)
2684 return NULL;
2686 in->safe_push (stmt);
2687 out->safe_push (stmt);
2688 return phi;
2691 /* Detect commutative and associative scalar reductions starting at
2692 STMT. Return the phi node of the reduction cycle, or NULL. */
2694 static gphi *
2695 detect_commutative_reduction_assign (gimple stmt, vec<gimple> *in,
2696 vec<gimple> *out)
2698 tree lhs = gimple_assign_lhs (stmt);
2700 if (gimple_num_ops (stmt) == 2)
2701 return detect_commutative_reduction_arg (lhs, stmt,
2702 gimple_assign_rhs1 (stmt),
2703 in, out);
2705 if (is_reduction_operation_p (stmt))
2707 gphi *res =
2708 detect_commutative_reduction_arg (lhs, stmt,
2709 gimple_assign_rhs1 (stmt),
2710 in, out);
2711 return res ? res
2712 : detect_commutative_reduction_arg (lhs, stmt,
2713 gimple_assign_rhs2 (stmt),
2714 in, out);
2717 return NULL;
2720 /* Return a loop phi node that corresponds to a reduction containing LHS. */
2722 static gphi *
2723 follow_inital_value_to_phi (tree arg, tree lhs)
2725 gimple stmt;
2727 if (!arg || TREE_CODE (arg) != SSA_NAME)
2728 return NULL;
2730 stmt = SSA_NAME_DEF_STMT (arg);
2732 if (gphi *phi = dyn_cast <gphi *> (stmt))
2733 if (phi_contains_arg (phi, lhs))
2734 return phi;
2736 return NULL;
2740 /* Return the argument of the loop PHI that is the initial value coming
2741 from outside the loop. */
2743 static edge
2744 edge_initial_value_for_loop_phi (gphi *phi)
2746 size_t i;
2748 for (i = 0; i < gimple_phi_num_args (phi); i++)
2750 edge e = gimple_phi_arg_edge (phi, i);
2752 if (loop_depth (e->src->loop_father)
2753 < loop_depth (e->dest->loop_father))
2754 return e;
2757 return NULL;
2760 /* Return the argument of the loop PHI that is the initial value coming
2761 from outside the loop. */
2763 static tree
2764 initial_value_for_loop_phi (gphi *phi)
2766 size_t i;
2768 for (i = 0; i < gimple_phi_num_args (phi); i++)
2770 edge e = gimple_phi_arg_edge (phi, i);
2772 if (loop_depth (e->src->loop_father)
2773 < loop_depth (e->dest->loop_father))
2774 return gimple_phi_arg_def (phi, i);
2777 return NULL_TREE;
2780 /* Returns true when DEF is used outside the reduction cycle of
2781 LOOP_PHI. */
2783 static bool
2784 used_outside_reduction (tree def, gimple loop_phi)
2786 use_operand_p use_p;
2787 imm_use_iterator imm_iter;
2788 loop_p loop = loop_containing_stmt (loop_phi);
2790 /* In LOOP, DEF should be used only in LOOP_PHI. */
2791 FOR_EACH_IMM_USE_FAST (use_p, imm_iter, def)
2793 gimple stmt = USE_STMT (use_p);
2795 if (stmt != loop_phi
2796 && !is_gimple_debug (stmt)
2797 && flow_bb_inside_loop_p (loop, gimple_bb (stmt)))
2798 return true;
2801 return false;
2804 /* Detect commutative and associative scalar reductions belonging to
2805 the SCOP starting at the loop closed phi node STMT. Return the phi
2806 node of the reduction cycle, or NULL. */
2808 static gphi *
2809 detect_commutative_reduction (scop_p scop, gimple stmt, vec<gimple> *in,
2810 vec<gimple> *out)
2812 if (scalar_close_phi_node_p (stmt))
2814 gimple def;
2815 gphi *loop_phi, *phi, *close_phi = as_a <gphi *> (stmt);
2816 tree init, lhs, arg = gimple_phi_arg_def (close_phi, 0);
2818 if (TREE_CODE (arg) != SSA_NAME)
2819 return NULL;
2821 /* Note that loop close phi nodes should have a single argument
2822 because we translated the representation into a canonical form
2823 before Graphite: see canonicalize_loop_closed_ssa_form. */
2824 gcc_assert (gimple_phi_num_args (close_phi) == 1);
2826 def = SSA_NAME_DEF_STMT (arg);
2827 if (!stmt_in_sese_p (def, SCOP_REGION (scop))
2828 || !(loop_phi = detect_commutative_reduction (scop, def, in, out)))
2829 return NULL;
2831 lhs = gimple_phi_result (close_phi);
2832 init = initial_value_for_loop_phi (loop_phi);
2833 phi = follow_inital_value_to_phi (init, lhs);
2835 if (phi && (used_outside_reduction (lhs, phi)
2836 || !has_single_use (gimple_phi_result (phi))))
2837 return NULL;
2839 in->safe_push (loop_phi);
2840 out->safe_push (close_phi);
2841 return phi;
2844 if (gimple_code (stmt) == GIMPLE_ASSIGN)
2845 return detect_commutative_reduction_assign (stmt, in, out);
2847 return NULL;
2850 /* Translate the scalar reduction statement STMT to an array RED
2851 knowing that its recursive phi node is LOOP_PHI. */
2853 static void
2854 translate_scalar_reduction_to_array_for_stmt (scop_p scop, tree red,
2855 gimple stmt, gphi *loop_phi)
2857 tree res = gimple_phi_result (loop_phi);
2858 gassign *assign = gimple_build_assign (res, unshare_expr (red));
2859 gimple_stmt_iterator gsi;
2861 insert_stmts (scop, assign, NULL, gsi_after_labels (gimple_bb (loop_phi)));
2863 assign = gimple_build_assign (unshare_expr (red), gimple_assign_lhs (stmt));
2864 gsi = gsi_for_stmt (stmt);
2865 gsi_next (&gsi);
2866 insert_stmts (scop, assign, NULL, gsi);
2869 /* Removes the PHI node and resets all the debug stmts that are using
2870 the PHI_RESULT. */
2872 static void
2873 remove_phi (gphi *phi)
2875 imm_use_iterator imm_iter;
2876 tree def;
2877 use_operand_p use_p;
2878 gimple_stmt_iterator gsi;
2879 auto_vec<gimple, 3> update;
2880 unsigned int i;
2881 gimple stmt;
2883 def = PHI_RESULT (phi);
2884 FOR_EACH_IMM_USE_FAST (use_p, imm_iter, def)
2886 stmt = USE_STMT (use_p);
2888 if (is_gimple_debug (stmt))
2890 gimple_debug_bind_reset_value (stmt);
2891 update.safe_push (stmt);
2895 FOR_EACH_VEC_ELT (update, i, stmt)
2896 update_stmt (stmt);
2898 gsi = gsi_for_phi_node (phi);
2899 remove_phi_node (&gsi, false);
2902 /* Helper function for for_each_index. For each INDEX of the data
2903 reference REF, returns true when its indices are valid in the loop
2904 nest LOOP passed in as DATA. */
2906 static bool
2907 dr_indices_valid_in_loop (tree ref ATTRIBUTE_UNUSED, tree *index, void *data)
2909 loop_p loop;
2910 basic_block header, def_bb;
2911 gimple stmt;
2913 if (TREE_CODE (*index) != SSA_NAME)
2914 return true;
2916 loop = *((loop_p *) data);
2917 header = loop->header;
2918 stmt = SSA_NAME_DEF_STMT (*index);
2920 if (!stmt)
2921 return true;
2923 def_bb = gimple_bb (stmt);
2925 if (!def_bb)
2926 return true;
2928 return dominated_by_p (CDI_DOMINATORS, header, def_bb);
2931 /* When the result of a CLOSE_PHI is written to a memory location,
2932 return a pointer to that memory reference, otherwise return
2933 NULL_TREE. */
2935 static tree
2936 close_phi_written_to_memory (gphi *close_phi)
2938 imm_use_iterator imm_iter;
2939 use_operand_p use_p;
2940 gimple stmt;
2941 tree res, def = gimple_phi_result (close_phi);
2943 FOR_EACH_IMM_USE_FAST (use_p, imm_iter, def)
2944 if ((stmt = USE_STMT (use_p))
2945 && gimple_code (stmt) == GIMPLE_ASSIGN
2946 && (res = gimple_assign_lhs (stmt)))
2948 switch (TREE_CODE (res))
2950 case VAR_DECL:
2951 case PARM_DECL:
2952 case RESULT_DECL:
2953 return res;
2955 case ARRAY_REF:
2956 case MEM_REF:
2958 tree arg = gimple_phi_arg_def (close_phi, 0);
2959 loop_p nest = loop_containing_stmt (SSA_NAME_DEF_STMT (arg));
2961 /* FIXME: this restriction is for id-{24,25}.f and
2962 could be handled by duplicating the computation of
2963 array indices before the loop of the close_phi. */
2964 if (for_each_index (&res, dr_indices_valid_in_loop, &nest))
2965 return res;
2967 /* Fallthru. */
2969 default:
2970 continue;
2973 return NULL_TREE;
2976 /* Rewrite out of SSA the reduction described by the loop phi nodes
2977 IN, and the close phi nodes OUT. IN and OUT are structured by loop
2978 levels like this:
2980 IN: stmt, loop_n, ..., loop_0
2981 OUT: stmt, close_n, ..., close_0
2983 the first element is the reduction statement, and the next elements
2984 are the loop and close phi nodes of each of the outer loops. */
2986 static void
2987 translate_scalar_reduction_to_array (scop_p scop,
2988 vec<gimple> in,
2989 vec<gimple> out)
2991 gimple loop_stmt;
2992 unsigned int i = out.length () - 1;
2993 tree red = close_phi_written_to_memory (as_a <gphi *> (out[i]));
2995 FOR_EACH_VEC_ELT (in, i, loop_stmt)
2997 gimple close_stmt = out[i];
2999 if (i == 0)
3001 basic_block bb = split_reduction_stmt (scop, loop_stmt);
3002 poly_bb_p pbb = pbb_from_bb (bb);
3003 PBB_IS_REDUCTION (pbb) = true;
3004 gcc_assert (close_stmt == loop_stmt);
3006 if (!red)
3007 red = create_zero_dim_array
3008 (gimple_assign_lhs (loop_stmt), "Commutative_Associative_Reduction");
3010 translate_scalar_reduction_to_array_for_stmt (scop, red, loop_stmt,
3011 as_a <gphi *> (in[1]));
3012 continue;
3015 gphi *loop_phi = as_a <gphi *> (loop_stmt);
3016 gphi *close_phi = as_a <gphi *> (close_stmt);
3018 if (i == in.length () - 1)
3020 insert_out_of_ssa_copy (scop, gimple_phi_result (close_phi),
3021 unshare_expr (red), close_phi);
3022 insert_out_of_ssa_copy_on_edge
3023 (scop, edge_initial_value_for_loop_phi (loop_phi),
3024 unshare_expr (red), initial_value_for_loop_phi (loop_phi));
3027 remove_phi (loop_phi);
3028 remove_phi (close_phi);
3032 /* Rewrites out of SSA a commutative reduction at CLOSE_PHI. Returns
3033 true when something has been changed. */
3035 static bool
3036 rewrite_commutative_reductions_out_of_ssa_close_phi (scop_p scop,
3037 gphi *close_phi)
3039 bool res;
3040 auto_vec<gimple, 10> in;
3041 auto_vec<gimple, 10> out;
3043 detect_commutative_reduction (scop, close_phi, &in, &out);
3044 res = in.length () > 1;
3045 if (res)
3046 translate_scalar_reduction_to_array (scop, in, out);
3048 return res;
3051 /* Rewrites all the commutative reductions from LOOP out of SSA.
3052 Returns true when something has been changed. */
3054 static bool
3055 rewrite_commutative_reductions_out_of_ssa_loop (scop_p scop,
3056 loop_p loop)
3058 gphi_iterator gsi;
3059 edge exit = single_exit (loop);
3060 tree res;
3061 bool changed = false;
3063 if (!exit)
3064 return false;
3066 for (gsi = gsi_start_phis (exit->dest); !gsi_end_p (gsi); gsi_next (&gsi))
3067 if ((res = gimple_phi_result (gsi.phi ()))
3068 && !virtual_operand_p (res)
3069 && !scev_analyzable_p (res, SCOP_REGION (scop)))
3070 changed |= rewrite_commutative_reductions_out_of_ssa_close_phi
3071 (scop, gsi.phi ());
3073 return changed;
3076 /* Rewrites all the commutative reductions from SCOP out of SSA. */
3078 static void
3079 rewrite_commutative_reductions_out_of_ssa (scop_p scop)
3081 loop_p loop;
3082 bool changed = false;
3083 sese region = SCOP_REGION (scop);
3085 FOR_EACH_LOOP (loop, 0)
3086 if (loop_in_sese_p (loop, region))
3087 changed |= rewrite_commutative_reductions_out_of_ssa_loop (scop, loop);
3089 if (changed)
3091 scev_reset_htab ();
3092 gsi_commit_edge_inserts ();
3093 update_ssa (TODO_update_ssa);
3094 #ifdef ENABLE_CHECKING
3095 verify_loop_closed_ssa (true);
3096 #endif
3100 /* Can all ivs be represented by a signed integer?
3101 As CLooG might generate negative values in its expressions, signed loop ivs
3102 are required in the backend. */
3104 static bool
3105 scop_ivs_can_be_represented (scop_p scop)
3107 loop_p loop;
3108 gphi_iterator psi;
3109 bool result = true;
3111 FOR_EACH_LOOP (loop, 0)
3113 if (!loop_in_sese_p (loop, SCOP_REGION (scop)))
3114 continue;
3116 for (psi = gsi_start_phis (loop->header);
3117 !gsi_end_p (psi); gsi_next (&psi))
3119 gphi *phi = psi.phi ();
3120 tree res = PHI_RESULT (phi);
3121 tree type = TREE_TYPE (res);
3123 if (TYPE_UNSIGNED (type)
3124 && TYPE_PRECISION (type) >= TYPE_PRECISION (long_long_integer_type_node))
3126 result = false;
3127 break;
3130 if (!result)
3131 break;
3134 return result;
3137 /* Builds the polyhedral representation for a SESE region. */
3139 void
3140 build_poly_scop (scop_p scop)
3142 sese region = SCOP_REGION (scop);
3143 graphite_dim_t max_dim;
3145 build_scop_bbs (scop);
3147 /* FIXME: This restriction is needed to avoid a problem in CLooG.
3148 Once CLooG is fixed, remove this guard. Anyways, it makes no
3149 sense to optimize a scop containing only PBBs that do not belong
3150 to any loops. */
3151 if (nb_pbbs_in_loops (scop) == 0)
3152 return;
3154 if (!scop_ivs_can_be_represented (scop))
3155 return;
3157 if (flag_associative_math)
3158 rewrite_commutative_reductions_out_of_ssa (scop);
3160 build_sese_loop_nests (region);
3161 /* Record all conditions in REGION. */
3162 sese_dom_walker (CDI_DOMINATORS, region).walk (cfun->cfg->x_entry_block_ptr);
3163 find_scop_parameters (scop);
3165 max_dim = PARAM_VALUE (PARAM_GRAPHITE_MAX_NB_SCOP_PARAMS);
3166 if (scop_nb_params (scop) > max_dim)
3167 return;
3169 build_scop_iteration_domain (scop);
3170 build_scop_context (scop);
3171 add_conditions_to_constraints (scop);
3173 /* Rewrite out of SSA only after having translated the
3174 representation to the polyhedral representation to avoid scev
3175 analysis failures. That means that these functions will insert
3176 new data references that they create in the right place. */
3177 rewrite_reductions_out_of_ssa (scop);
3178 rewrite_cross_bb_scalar_deps_out_of_ssa (scop);
3180 build_scop_drs (scop);
3181 scop_to_lst (scop);
3182 build_scop_scattering (scop);
3184 /* This SCoP has been translated to the polyhedral
3185 representation. */
3186 POLY_SCOP_P (scop) = true;
3188 #endif