Introduce gimple_omp_continue
[official-gcc.git] / gcc / graphite-sese-to-poly.c
blobc3adfe13e9e1af6ce51943930462f478ed9b1c5a
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 (gimple 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 (gimple_stmt_iterator *psi)
118 gimple phi = gsi_stmt (*psi);
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 gimple 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, gimple_stmt_iterator *psi)
135 gimple phi = gsi_stmt (*psi);
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 gimple 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 (gimple 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, gimple_stmt_iterator *psi)
180 loop_p loop;
181 gimple phi = gsi_stmt (*psi);
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, gimple 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 enum tree_code code = gimple_cond_code (stmt);
1187 /* The conditions for ELSE-branches are inverted. */
1188 if (!GBB_CONDITION_CASES (gbb)[i])
1189 code = invert_tree_comparison (code, false);
1191 add_condition_to_pbb (pbb, stmt, code);
1192 break;
1195 case GIMPLE_SWITCH:
1196 /* Switch statements are not supported right now - fall through. */
1198 default:
1199 gcc_unreachable ();
1200 break;
1204 /* Traverses all the GBBs of the SCOP and add their constraints to the
1205 iteration domains. */
1207 static void
1208 add_conditions_to_constraints (scop_p scop)
1210 int i;
1211 poly_bb_p pbb;
1213 FOR_EACH_VEC_ELT (SCOP_BBS (scop), i, pbb)
1214 add_conditions_to_domain (pbb);
1217 /* Returns a COND_EXPR statement when BB has a single predecessor, the
1218 edge between BB and its predecessor is not a loop exit edge, and
1219 the last statement of the single predecessor is a COND_EXPR. */
1221 static gimple
1222 single_pred_cond_non_loop_exit (basic_block bb)
1224 if (single_pred_p (bb))
1226 edge e = single_pred_edge (bb);
1227 basic_block pred = e->src;
1228 gimple stmt;
1230 if (loop_depth (pred->loop_father) > loop_depth (bb->loop_father))
1231 return NULL;
1233 stmt = last_stmt (pred);
1235 if (stmt && gimple_code (stmt) == GIMPLE_COND)
1236 return stmt;
1239 return NULL;
1242 class sese_dom_walker : public dom_walker
1244 public:
1245 sese_dom_walker (cdi_direction, sese);
1247 virtual void before_dom_children (basic_block);
1248 virtual void after_dom_children (basic_block);
1250 private:
1251 auto_vec<gimple, 3> m_conditions, m_cases;
1252 sese m_region;
1255 sese_dom_walker::sese_dom_walker (cdi_direction direction, sese region)
1256 : dom_walker (direction), m_region (region)
1260 /* Call-back for dom_walk executed before visiting the dominated
1261 blocks. */
1263 void
1264 sese_dom_walker::before_dom_children (basic_block bb)
1266 gimple_bb_p gbb;
1267 gimple stmt;
1269 if (!bb_in_sese_p (bb, m_region))
1270 return;
1272 stmt = single_pred_cond_non_loop_exit (bb);
1274 if (stmt)
1276 edge e = single_pred_edge (bb);
1278 m_conditions.safe_push (stmt);
1280 if (e->flags & EDGE_TRUE_VALUE)
1281 m_cases.safe_push (stmt);
1282 else
1283 m_cases.safe_push (NULL);
1286 gbb = gbb_from_bb (bb);
1288 if (gbb)
1290 GBB_CONDITIONS (gbb) = m_conditions.copy ();
1291 GBB_CONDITION_CASES (gbb) = m_cases.copy ();
1295 /* Call-back for dom_walk executed after visiting the dominated
1296 blocks. */
1298 void
1299 sese_dom_walker::after_dom_children (basic_block bb)
1301 if (!bb_in_sese_p (bb, m_region))
1302 return;
1304 if (single_pred_cond_non_loop_exit (bb))
1306 m_conditions.pop ();
1307 m_cases.pop ();
1311 /* Add constraints on the possible values of parameter P from the type
1312 of P. */
1314 static void
1315 add_param_constraints (scop_p scop, graphite_dim_t p)
1317 tree parameter = SESE_PARAMS (SCOP_REGION (scop))[p];
1318 tree type = TREE_TYPE (parameter);
1319 tree lb = NULL_TREE;
1320 tree ub = NULL_TREE;
1322 if (POINTER_TYPE_P (type) || !TYPE_MIN_VALUE (type))
1323 lb = lower_bound_in_type (type, type);
1324 else
1325 lb = TYPE_MIN_VALUE (type);
1327 if (POINTER_TYPE_P (type) || !TYPE_MAX_VALUE (type))
1328 ub = upper_bound_in_type (type, type);
1329 else
1330 ub = TYPE_MAX_VALUE (type);
1332 if (lb)
1334 isl_space *space = isl_set_get_space (scop->context);
1335 isl_constraint *c;
1336 mpz_t g;
1337 isl_val *v;
1339 c = isl_inequality_alloc (isl_local_space_from_space (space));
1340 mpz_init (g);
1341 tree_int_to_gmp (lb, g);
1342 v = isl_val_int_from_gmp (the_isl_ctx, g);
1343 v = isl_val_neg (v);
1344 mpz_clear (g);
1345 c = isl_constraint_set_constant_val (c, v);
1346 c = isl_constraint_set_coefficient_si (c, isl_dim_param, p, 1);
1348 scop->context = isl_set_add_constraint (scop->context, c);
1351 if (ub)
1353 isl_space *space = isl_set_get_space (scop->context);
1354 isl_constraint *c;
1355 mpz_t g;
1356 isl_val *v;
1358 c = isl_inequality_alloc (isl_local_space_from_space (space));
1360 mpz_init (g);
1361 tree_int_to_gmp (ub, g);
1362 v = isl_val_int_from_gmp (the_isl_ctx, g);
1363 mpz_clear (g);
1364 c = isl_constraint_set_constant_val (c, v);
1365 c = isl_constraint_set_coefficient_si (c, isl_dim_param, p, -1);
1367 scop->context = isl_set_add_constraint (scop->context, c);
1371 /* Build the context of the SCOP. The context usually contains extra
1372 constraints that are added to the iteration domains that constrain
1373 some parameters. */
1375 static void
1376 build_scop_context (scop_p scop)
1378 graphite_dim_t p, n = scop_nb_params (scop);
1380 for (p = 0; p < n; p++)
1381 add_param_constraints (scop, p);
1384 /* Build the iteration domains: the loops belonging to the current
1385 SCOP, and that vary for the execution of the current basic block.
1386 Returns false if there is no loop in SCOP. */
1388 static void
1389 build_scop_iteration_domain (scop_p scop)
1391 struct loop *loop;
1392 sese region = SCOP_REGION (scop);
1393 int i;
1394 poly_bb_p pbb;
1395 int nb_loops = number_of_loops (cfun);
1396 isl_set **doms = XCNEWVEC (isl_set *, nb_loops);
1398 FOR_EACH_VEC_ELT (SESE_LOOP_NEST (region), i, loop)
1399 if (!loop_in_sese_p (loop_outer (loop), region))
1400 build_loop_iteration_domains (scop, loop, 0,
1401 isl_set_copy (scop->context), doms);
1403 FOR_EACH_VEC_ELT (SCOP_BBS (scop), i, pbb)
1405 loop = pbb_loop (pbb);
1407 if (doms[loop->num])
1408 pbb->domain = isl_set_copy (doms[loop->num]);
1409 else
1410 pbb->domain = isl_set_copy (scop->context);
1412 pbb->domain = isl_set_set_tuple_id (pbb->domain,
1413 isl_id_for_pbb (scop, pbb));
1416 for (i = 0; i < nb_loops; i++)
1417 if (doms[i])
1418 isl_set_free (doms[i]);
1420 free (doms);
1423 /* Add a constrain to the ACCESSES polyhedron for the alias set of
1424 data reference DR. ACCESSP_NB_DIMS is the dimension of the
1425 ACCESSES polyhedron, DOM_NB_DIMS is the dimension of the iteration
1426 domain. */
1428 static isl_map *
1429 pdr_add_alias_set (isl_map *acc, data_reference_p dr)
1431 isl_constraint *c;
1432 int alias_set_num = 0;
1433 base_alias_pair *bap = (base_alias_pair *)(dr->aux);
1435 if (bap && bap->alias_set)
1436 alias_set_num = *(bap->alias_set);
1438 c = isl_equality_alloc
1439 (isl_local_space_from_space (isl_map_get_space (acc)));
1440 c = isl_constraint_set_constant_si (c, -alias_set_num);
1441 c = isl_constraint_set_coefficient_si (c, isl_dim_out, 0, 1);
1443 return isl_map_add_constraint (acc, c);
1446 /* Assign the affine expression INDEX to the output dimension POS of
1447 MAP and return the result. */
1449 static isl_map *
1450 set_index (isl_map *map, int pos, isl_pw_aff *index)
1452 isl_map *index_map;
1453 int len = isl_map_dim (map, isl_dim_out);
1454 isl_id *id;
1456 index_map = isl_map_from_pw_aff (index);
1457 index_map = isl_map_insert_dims (index_map, isl_dim_out, 0, pos);
1458 index_map = isl_map_add_dims (index_map, isl_dim_out, len - pos - 1);
1460 id = isl_map_get_tuple_id (map, isl_dim_out);
1461 index_map = isl_map_set_tuple_id (index_map, isl_dim_out, id);
1462 id = isl_map_get_tuple_id (map, isl_dim_in);
1463 index_map = isl_map_set_tuple_id (index_map, isl_dim_in, id);
1465 return isl_map_intersect (map, index_map);
1468 /* Add to ACCESSES polyhedron equalities defining the access functions
1469 to the memory. ACCESSP_NB_DIMS is the dimension of the ACCESSES
1470 polyhedron, DOM_NB_DIMS is the dimension of the iteration domain.
1471 PBB is the poly_bb_p that contains the data reference DR. */
1473 static isl_map *
1474 pdr_add_memory_accesses (isl_map *acc, data_reference_p dr, poly_bb_p pbb)
1476 int i, nb_subscripts = DR_NUM_DIMENSIONS (dr);
1477 scop_p scop = PBB_SCOP (pbb);
1479 for (i = 0; i < nb_subscripts; i++)
1481 isl_pw_aff *aff;
1482 tree afn = DR_ACCESS_FN (dr, nb_subscripts - 1 - i);
1484 aff = extract_affine (scop, afn,
1485 isl_space_domain (isl_map_get_space (acc)));
1486 acc = set_index (acc, i + 1, aff);
1489 return acc;
1492 /* Add constrains representing the size of the accessed data to the
1493 ACCESSES polyhedron. ACCESSP_NB_DIMS is the dimension of the
1494 ACCESSES polyhedron, DOM_NB_DIMS is the dimension of the iteration
1495 domain. */
1497 static isl_set *
1498 pdr_add_data_dimensions (isl_set *extent, scop_p scop, data_reference_p dr)
1500 tree ref = DR_REF (dr);
1501 int i, nb_subscripts = DR_NUM_DIMENSIONS (dr);
1503 for (i = nb_subscripts - 1; i >= 0; i--, ref = TREE_OPERAND (ref, 0))
1505 tree low, high;
1507 if (TREE_CODE (ref) != ARRAY_REF)
1508 break;
1510 low = array_ref_low_bound (ref);
1511 high = array_ref_up_bound (ref);
1513 /* XXX The PPL code dealt separately with
1514 subscript - low >= 0 and high - subscript >= 0 in case one of
1515 the two bounds isn't known. Do the same here? */
1517 if (tree_fits_shwi_p (low)
1518 && high
1519 && tree_fits_shwi_p (high)
1520 /* 1-element arrays at end of structures may extend over
1521 their declared size. */
1522 && !(array_at_struct_end_p (ref)
1523 && operand_equal_p (low, high, 0)))
1525 isl_id *id;
1526 isl_aff *aff;
1527 isl_set *univ, *lbs, *ubs;
1528 isl_pw_aff *index;
1529 isl_space *space;
1530 isl_set *valid;
1531 isl_pw_aff *lb = extract_affine_int (low, isl_set_get_space (extent));
1532 isl_pw_aff *ub = extract_affine_int (high, isl_set_get_space (extent));
1534 /* high >= 0 */
1535 valid = isl_pw_aff_nonneg_set (isl_pw_aff_copy (ub));
1536 valid = isl_set_project_out (valid, isl_dim_set, 0,
1537 isl_set_dim (valid, isl_dim_set));
1538 scop->context = isl_set_intersect (scop->context, valid);
1540 space = isl_set_get_space (extent);
1541 aff = isl_aff_zero_on_domain (isl_local_space_from_space (space));
1542 aff = isl_aff_add_coefficient_si (aff, isl_dim_in, i + 1, 1);
1543 univ = isl_set_universe (isl_space_domain (isl_aff_get_space (aff)));
1544 index = isl_pw_aff_alloc (univ, aff);
1546 id = isl_set_get_tuple_id (extent);
1547 lb = isl_pw_aff_set_tuple_id (lb, isl_dim_in, isl_id_copy (id));
1548 ub = isl_pw_aff_set_tuple_id (ub, isl_dim_in, id);
1550 /* low <= sub_i <= high */
1551 lbs = isl_pw_aff_ge_set (isl_pw_aff_copy (index), lb);
1552 ubs = isl_pw_aff_le_set (index, ub);
1553 extent = isl_set_intersect (extent, lbs);
1554 extent = isl_set_intersect (extent, ubs);
1558 return extent;
1561 /* Build data accesses for DR in PBB. */
1563 static void
1564 build_poly_dr (data_reference_p dr, poly_bb_p pbb)
1566 int dr_base_object_set;
1567 isl_map *acc;
1568 isl_set *extent;
1569 scop_p scop = PBB_SCOP (pbb);
1572 isl_space *dc = isl_set_get_space (pbb->domain);
1573 int nb_out = 1 + DR_NUM_DIMENSIONS (dr);
1574 isl_space *space = isl_space_add_dims (isl_space_from_domain (dc),
1575 isl_dim_out, nb_out);
1577 acc = isl_map_universe (space);
1578 acc = isl_map_set_tuple_id (acc, isl_dim_out, isl_id_for_dr (scop, dr));
1581 acc = pdr_add_alias_set (acc, dr);
1582 acc = pdr_add_memory_accesses (acc, dr, pbb);
1585 isl_id *id = isl_id_for_dr (scop, dr);
1586 int nb = 1 + DR_NUM_DIMENSIONS (dr);
1587 isl_space *space = isl_space_set_alloc (scop->ctx, 0, nb);
1588 int alias_set_num = 0;
1589 base_alias_pair *bap = (base_alias_pair *)(dr->aux);
1591 if (bap && bap->alias_set)
1592 alias_set_num = *(bap->alias_set);
1594 space = isl_space_set_tuple_id (space, isl_dim_set, id);
1595 extent = isl_set_nat_universe (space);
1596 extent = isl_set_fix_si (extent, isl_dim_set, 0, alias_set_num);
1597 extent = pdr_add_data_dimensions (extent, scop, dr);
1600 gcc_assert (dr->aux);
1601 dr_base_object_set = ((base_alias_pair *)(dr->aux))->base_obj_set;
1603 new_poly_dr (pbb, dr_base_object_set,
1604 DR_IS_READ (dr) ? PDR_READ : PDR_WRITE,
1605 dr, DR_NUM_DIMENSIONS (dr), acc, extent);
1608 /* Write to FILE the alias graph of data references in DIMACS format. */
1610 static inline bool
1611 write_alias_graph_to_ascii_dimacs (FILE *file, char *comment,
1612 vec<data_reference_p> drs)
1614 int num_vertex = drs.length ();
1615 int edge_num = 0;
1616 data_reference_p dr1, dr2;
1617 int i, j;
1619 if (num_vertex == 0)
1620 return true;
1622 FOR_EACH_VEC_ELT (drs, i, dr1)
1623 for (j = i + 1; drs.iterate (j, &dr2); j++)
1624 if (dr_may_alias_p (dr1, dr2, true))
1625 edge_num++;
1627 fprintf (file, "$\n");
1629 if (comment)
1630 fprintf (file, "c %s\n", comment);
1632 fprintf (file, "p edge %d %d\n", num_vertex, edge_num);
1634 FOR_EACH_VEC_ELT (drs, i, dr1)
1635 for (j = i + 1; drs.iterate (j, &dr2); j++)
1636 if (dr_may_alias_p (dr1, dr2, true))
1637 fprintf (file, "e %d %d\n", i + 1, j + 1);
1639 return true;
1642 /* Write to FILE the alias graph of data references in DOT format. */
1644 static inline bool
1645 write_alias_graph_to_ascii_dot (FILE *file, char *comment,
1646 vec<data_reference_p> drs)
1648 int num_vertex = drs.length ();
1649 data_reference_p dr1, dr2;
1650 int i, j;
1652 if (num_vertex == 0)
1653 return true;
1655 fprintf (file, "$\n");
1657 if (comment)
1658 fprintf (file, "c %s\n", comment);
1660 /* First print all the vertices. */
1661 FOR_EACH_VEC_ELT (drs, i, dr1)
1662 fprintf (file, "n%d;\n", i);
1664 FOR_EACH_VEC_ELT (drs, i, dr1)
1665 for (j = i + 1; drs.iterate (j, &dr2); j++)
1666 if (dr_may_alias_p (dr1, dr2, true))
1667 fprintf (file, "n%d n%d\n", i, j);
1669 return true;
1672 /* Write to FILE the alias graph of data references in ECC format. */
1674 static inline bool
1675 write_alias_graph_to_ascii_ecc (FILE *file, char *comment,
1676 vec<data_reference_p> drs)
1678 int num_vertex = drs.length ();
1679 data_reference_p dr1, dr2;
1680 int i, j;
1682 if (num_vertex == 0)
1683 return true;
1685 fprintf (file, "$\n");
1687 if (comment)
1688 fprintf (file, "c %s\n", comment);
1690 FOR_EACH_VEC_ELT (drs, i, dr1)
1691 for (j = i + 1; drs.iterate (j, &dr2); j++)
1692 if (dr_may_alias_p (dr1, dr2, true))
1693 fprintf (file, "%d %d\n", i, j);
1695 return true;
1698 /* Check if DR1 and DR2 are in the same object set. */
1700 static bool
1701 dr_same_base_object_p (const struct data_reference *dr1,
1702 const struct data_reference *dr2)
1704 return operand_equal_p (DR_BASE_OBJECT (dr1), DR_BASE_OBJECT (dr2), 0);
1707 /* Uses DFS component number as representative of alias-sets. Also tests for
1708 optimality by verifying if every connected component is a clique. Returns
1709 true (1) if the above test is true, and false (0) otherwise. */
1711 static int
1712 build_alias_set_optimal_p (vec<data_reference_p> drs)
1714 int num_vertices = drs.length ();
1715 struct graph *g = new_graph (num_vertices);
1716 data_reference_p dr1, dr2;
1717 int i, j;
1718 int num_connected_components;
1719 int v_indx1, v_indx2, num_vertices_in_component;
1720 int *all_vertices;
1721 int *vertices;
1722 struct graph_edge *e;
1723 int this_component_is_clique;
1724 int all_components_are_cliques = 1;
1726 FOR_EACH_VEC_ELT (drs, i, dr1)
1727 for (j = i+1; drs.iterate (j, &dr2); j++)
1728 if (dr_may_alias_p (dr1, dr2, true))
1730 add_edge (g, i, j);
1731 add_edge (g, j, i);
1734 all_vertices = XNEWVEC (int, num_vertices);
1735 vertices = XNEWVEC (int, num_vertices);
1736 for (i = 0; i < num_vertices; i++)
1737 all_vertices[i] = i;
1739 num_connected_components = graphds_dfs (g, all_vertices, num_vertices,
1740 NULL, true, NULL);
1741 for (i = 0; i < g->n_vertices; i++)
1743 data_reference_p dr = drs[i];
1744 base_alias_pair *bap;
1746 gcc_assert (dr->aux);
1747 bap = (base_alias_pair *)(dr->aux);
1749 bap->alias_set = XNEW (int);
1750 *(bap->alias_set) = g->vertices[i].component + 1;
1753 /* Verify if the DFS numbering results in optimal solution. */
1754 for (i = 0; i < num_connected_components; i++)
1756 num_vertices_in_component = 0;
1757 /* Get all vertices whose DFS component number is the same as i. */
1758 for (j = 0; j < num_vertices; j++)
1759 if (g->vertices[j].component == i)
1760 vertices[num_vertices_in_component++] = j;
1762 /* Now test if the vertices in 'vertices' form a clique, by testing
1763 for edges among each pair. */
1764 this_component_is_clique = 1;
1765 for (v_indx1 = 0; v_indx1 < num_vertices_in_component; v_indx1++)
1767 for (v_indx2 = v_indx1+1; v_indx2 < num_vertices_in_component; v_indx2++)
1769 /* Check if the two vertices are connected by iterating
1770 through all the edges which have one of these are source. */
1771 e = g->vertices[vertices[v_indx2]].pred;
1772 while (e)
1774 if (e->src == vertices[v_indx1])
1775 break;
1776 e = e->pred_next;
1778 if (!e)
1780 this_component_is_clique = 0;
1781 break;
1784 if (!this_component_is_clique)
1785 all_components_are_cliques = 0;
1789 free (all_vertices);
1790 free (vertices);
1791 free_graph (g);
1792 return all_components_are_cliques;
1795 /* Group each data reference in DRS with its base object set num. */
1797 static void
1798 build_base_obj_set_for_drs (vec<data_reference_p> drs)
1800 int num_vertex = drs.length ();
1801 struct graph *g = new_graph (num_vertex);
1802 data_reference_p dr1, dr2;
1803 int i, j;
1804 int *queue;
1806 FOR_EACH_VEC_ELT (drs, i, dr1)
1807 for (j = i + 1; drs.iterate (j, &dr2); j++)
1808 if (dr_same_base_object_p (dr1, dr2))
1810 add_edge (g, i, j);
1811 add_edge (g, j, i);
1814 queue = XNEWVEC (int, num_vertex);
1815 for (i = 0; i < num_vertex; i++)
1816 queue[i] = i;
1818 graphds_dfs (g, queue, num_vertex, NULL, true, NULL);
1820 for (i = 0; i < g->n_vertices; i++)
1822 data_reference_p dr = drs[i];
1823 base_alias_pair *bap;
1825 gcc_assert (dr->aux);
1826 bap = (base_alias_pair *)(dr->aux);
1828 bap->base_obj_set = g->vertices[i].component + 1;
1831 free (queue);
1832 free_graph (g);
1835 /* Build the data references for PBB. */
1837 static void
1838 build_pbb_drs (poly_bb_p pbb)
1840 int j;
1841 data_reference_p dr;
1842 vec<data_reference_p> gbb_drs = GBB_DATA_REFS (PBB_BLACK_BOX (pbb));
1844 FOR_EACH_VEC_ELT (gbb_drs, j, dr)
1845 build_poly_dr (dr, pbb);
1848 /* Dump to file the alias graphs for the data references in DRS. */
1850 static void
1851 dump_alias_graphs (vec<data_reference_p> drs)
1853 char comment[100];
1854 FILE *file_dimacs, *file_ecc, *file_dot;
1856 file_dimacs = fopen ("/tmp/dr_alias_graph_dimacs", "ab");
1857 if (file_dimacs)
1859 snprintf (comment, sizeof (comment), "%s %s", main_input_filename,
1860 current_function_name ());
1861 write_alias_graph_to_ascii_dimacs (file_dimacs, comment, drs);
1862 fclose (file_dimacs);
1865 file_ecc = fopen ("/tmp/dr_alias_graph_ecc", "ab");
1866 if (file_ecc)
1868 snprintf (comment, sizeof (comment), "%s %s", main_input_filename,
1869 current_function_name ());
1870 write_alias_graph_to_ascii_ecc (file_ecc, comment, drs);
1871 fclose (file_ecc);
1874 file_dot = fopen ("/tmp/dr_alias_graph_dot", "ab");
1875 if (file_dot)
1877 snprintf (comment, sizeof (comment), "%s %s", main_input_filename,
1878 current_function_name ());
1879 write_alias_graph_to_ascii_dot (file_dot, comment, drs);
1880 fclose (file_dot);
1884 /* Build data references in SCOP. */
1886 static void
1887 build_scop_drs (scop_p scop)
1889 int i, j;
1890 poly_bb_p pbb;
1891 data_reference_p dr;
1892 auto_vec<data_reference_p, 3> drs;
1894 /* Remove all the PBBs that do not have data references: these basic
1895 blocks are not handled in the polyhedral representation. */
1896 for (i = 0; SCOP_BBS (scop).iterate (i, &pbb); i++)
1897 if (GBB_DATA_REFS (PBB_BLACK_BOX (pbb)).is_empty ())
1899 free_gimple_bb (PBB_BLACK_BOX (pbb));
1900 free_poly_bb (pbb);
1901 SCOP_BBS (scop).ordered_remove (i);
1902 i--;
1905 FOR_EACH_VEC_ELT (SCOP_BBS (scop), i, pbb)
1906 for (j = 0; GBB_DATA_REFS (PBB_BLACK_BOX (pbb)).iterate (j, &dr); j++)
1907 drs.safe_push (dr);
1909 FOR_EACH_VEC_ELT (drs, i, dr)
1910 dr->aux = XNEW (base_alias_pair);
1912 if (!build_alias_set_optimal_p (drs))
1914 /* TODO: Add support when building alias set is not optimal. */
1918 build_base_obj_set_for_drs (drs);
1920 /* When debugging, enable the following code. This cannot be used
1921 in production compilers. */
1922 if (0)
1923 dump_alias_graphs (drs);
1925 drs.release ();
1927 FOR_EACH_VEC_ELT (SCOP_BBS (scop), i, pbb)
1928 build_pbb_drs (pbb);
1931 /* Return a gsi at the position of the phi node STMT. */
1933 static gimple_stmt_iterator
1934 gsi_for_phi_node (gimple stmt)
1936 gimple_stmt_iterator psi;
1937 basic_block bb = gimple_bb (stmt);
1939 for (psi = gsi_start_phis (bb); !gsi_end_p (psi); gsi_next (&psi))
1940 if (stmt == gsi_stmt (psi))
1941 return psi;
1943 gcc_unreachable ();
1944 return psi;
1947 /* Analyze all the data references of STMTS and add them to the
1948 GBB_DATA_REFS vector of BB. */
1950 static void
1951 analyze_drs_in_stmts (scop_p scop, basic_block bb, vec<gimple> stmts)
1953 loop_p nest;
1954 gimple_bb_p gbb;
1955 gimple stmt;
1956 int i;
1957 sese region = SCOP_REGION (scop);
1959 if (!bb_in_sese_p (bb, region))
1960 return;
1962 nest = outermost_loop_in_sese_1 (region, bb);
1963 gbb = gbb_from_bb (bb);
1965 FOR_EACH_VEC_ELT (stmts, i, stmt)
1967 loop_p loop;
1969 if (is_gimple_debug (stmt))
1970 continue;
1972 loop = loop_containing_stmt (stmt);
1973 if (!loop_in_sese_p (loop, region))
1974 loop = nest;
1976 graphite_find_data_references_in_stmt (nest, loop, stmt,
1977 &GBB_DATA_REFS (gbb));
1981 /* Insert STMT at the end of the STMTS sequence and then insert the
1982 statements from STMTS at INSERT_GSI and call analyze_drs_in_stmts
1983 on STMTS. */
1985 static void
1986 insert_stmts (scop_p scop, gimple stmt, gimple_seq stmts,
1987 gimple_stmt_iterator insert_gsi)
1989 gimple_stmt_iterator gsi;
1990 auto_vec<gimple, 3> x;
1992 gimple_seq_add_stmt (&stmts, stmt);
1993 for (gsi = gsi_start (stmts); !gsi_end_p (gsi); gsi_next (&gsi))
1994 x.safe_push (gsi_stmt (gsi));
1996 gsi_insert_seq_before (&insert_gsi, stmts, GSI_SAME_STMT);
1997 analyze_drs_in_stmts (scop, gsi_bb (insert_gsi), x);
2000 /* Insert the assignment "RES := EXPR" just after AFTER_STMT. */
2002 static void
2003 insert_out_of_ssa_copy (scop_p scop, tree res, tree expr, gimple after_stmt)
2005 gimple_seq stmts;
2006 gimple_stmt_iterator gsi;
2007 tree var = force_gimple_operand (expr, &stmts, true, NULL_TREE);
2008 gimple stmt = gimple_build_assign (unshare_expr (res), var);
2009 auto_vec<gimple, 3> x;
2011 gimple_seq_add_stmt (&stmts, stmt);
2012 for (gsi = gsi_start (stmts); !gsi_end_p (gsi); gsi_next (&gsi))
2013 x.safe_push (gsi_stmt (gsi));
2015 if (gimple_code (after_stmt) == GIMPLE_PHI)
2017 gsi = gsi_after_labels (gimple_bb (after_stmt));
2018 gsi_insert_seq_before (&gsi, stmts, GSI_NEW_STMT);
2020 else
2022 gsi = gsi_for_stmt (after_stmt);
2023 gsi_insert_seq_after (&gsi, stmts, GSI_NEW_STMT);
2026 analyze_drs_in_stmts (scop, gimple_bb (after_stmt), x);
2029 /* Creates a poly_bb_p for basic_block BB from the existing PBB. */
2031 static void
2032 new_pbb_from_pbb (scop_p scop, poly_bb_p pbb, basic_block bb)
2034 vec<data_reference_p> drs;
2035 drs.create (3);
2036 gimple_bb_p gbb = PBB_BLACK_BOX (pbb);
2037 gimple_bb_p gbb1 = new_gimple_bb (bb, drs);
2038 poly_bb_p pbb1 = new_poly_bb (scop, gbb1);
2039 int index, n = SCOP_BBS (scop).length ();
2041 /* The INDEX of PBB in SCOP_BBS. */
2042 for (index = 0; index < n; index++)
2043 if (SCOP_BBS (scop)[index] == pbb)
2044 break;
2046 pbb1->domain = isl_set_copy (pbb->domain);
2047 pbb1->domain = isl_set_set_tuple_id (pbb1->domain,
2048 isl_id_for_pbb (scop, pbb1));
2050 GBB_PBB (gbb1) = pbb1;
2051 GBB_CONDITIONS (gbb1) = GBB_CONDITIONS (gbb).copy ();
2052 GBB_CONDITION_CASES (gbb1) = GBB_CONDITION_CASES (gbb).copy ();
2053 SCOP_BBS (scop).safe_insert (index + 1, pbb1);
2056 /* Insert on edge E the assignment "RES := EXPR". */
2058 static void
2059 insert_out_of_ssa_copy_on_edge (scop_p scop, edge e, tree res, tree expr)
2061 gimple_stmt_iterator gsi;
2062 gimple_seq stmts = NULL;
2063 tree var = force_gimple_operand (expr, &stmts, true, NULL_TREE);
2064 gimple stmt = gimple_build_assign (unshare_expr (res), var);
2065 basic_block bb;
2066 auto_vec<gimple, 3> x;
2068 gimple_seq_add_stmt (&stmts, stmt);
2069 for (gsi = gsi_start (stmts); !gsi_end_p (gsi); gsi_next (&gsi))
2070 x.safe_push (gsi_stmt (gsi));
2072 gsi_insert_seq_on_edge (e, stmts);
2073 gsi_commit_edge_inserts ();
2074 bb = gimple_bb (stmt);
2076 if (!bb_in_sese_p (bb, SCOP_REGION (scop)))
2077 return;
2079 if (!gbb_from_bb (bb))
2080 new_pbb_from_pbb (scop, pbb_from_bb (e->src), bb);
2082 analyze_drs_in_stmts (scop, bb, x);
2085 /* Creates a zero dimension array of the same type as VAR. */
2087 static tree
2088 create_zero_dim_array (tree var, const char *base_name)
2090 tree index_type = build_index_type (integer_zero_node);
2091 tree elt_type = TREE_TYPE (var);
2092 tree array_type = build_array_type (elt_type, index_type);
2093 tree base = create_tmp_var (array_type, base_name);
2095 return build4 (ARRAY_REF, elt_type, base, integer_zero_node, NULL_TREE,
2096 NULL_TREE);
2099 /* Returns true when PHI is a loop close phi node. */
2101 static bool
2102 scalar_close_phi_node_p (gimple phi)
2104 if (gimple_code (phi) != GIMPLE_PHI
2105 || virtual_operand_p (gimple_phi_result (phi)))
2106 return false;
2108 /* Note that loop close phi nodes should have a single argument
2109 because we translated the representation into a canonical form
2110 before Graphite: see canonicalize_loop_closed_ssa_form. */
2111 return (gimple_phi_num_args (phi) == 1);
2114 /* For a definition DEF in REGION, propagates the expression EXPR in
2115 all the uses of DEF outside REGION. */
2117 static void
2118 propagate_expr_outside_region (tree def, tree expr, sese region)
2120 imm_use_iterator imm_iter;
2121 gimple use_stmt;
2122 gimple_seq stmts;
2123 bool replaced_once = false;
2125 gcc_assert (TREE_CODE (def) == SSA_NAME);
2127 expr = force_gimple_operand (unshare_expr (expr), &stmts, true,
2128 NULL_TREE);
2130 FOR_EACH_IMM_USE_STMT (use_stmt, imm_iter, def)
2131 if (!is_gimple_debug (use_stmt)
2132 && !bb_in_sese_p (gimple_bb (use_stmt), region))
2134 ssa_op_iter iter;
2135 use_operand_p use_p;
2137 FOR_EACH_PHI_OR_STMT_USE (use_p, use_stmt, iter, SSA_OP_ALL_USES)
2138 if (operand_equal_p (def, USE_FROM_PTR (use_p), 0)
2139 && (replaced_once = true))
2140 replace_exp (use_p, expr);
2142 update_stmt (use_stmt);
2145 if (replaced_once)
2147 gsi_insert_seq_on_edge (SESE_ENTRY (region), stmts);
2148 gsi_commit_edge_inserts ();
2152 /* Rewrite out of SSA the reduction phi node at PSI by creating a zero
2153 dimension array for it. */
2155 static void
2156 rewrite_close_phi_out_of_ssa (scop_p scop, gimple_stmt_iterator *psi)
2158 sese region = SCOP_REGION (scop);
2159 gimple phi = gsi_stmt (*psi);
2160 tree res = gimple_phi_result (phi);
2161 basic_block bb = gimple_bb (phi);
2162 gimple_stmt_iterator gsi = gsi_after_labels (bb);
2163 tree arg = gimple_phi_arg_def (phi, 0);
2164 gimple stmt;
2166 /* Note that loop close phi nodes should have a single argument
2167 because we translated the representation into a canonical form
2168 before Graphite: see canonicalize_loop_closed_ssa_form. */
2169 gcc_assert (gimple_phi_num_args (phi) == 1);
2171 /* The phi node can be a non close phi node, when its argument is
2172 invariant, or a default definition. */
2173 if (is_gimple_min_invariant (arg)
2174 || SSA_NAME_IS_DEFAULT_DEF (arg))
2176 propagate_expr_outside_region (res, arg, region);
2177 gsi_next (psi);
2178 return;
2181 else if (gimple_bb (SSA_NAME_DEF_STMT (arg))->loop_father == bb->loop_father)
2183 propagate_expr_outside_region (res, arg, region);
2184 stmt = gimple_build_assign (res, arg);
2185 remove_phi_node (psi, false);
2186 gsi_insert_before (&gsi, stmt, GSI_NEW_STMT);
2187 return;
2190 /* If res is scev analyzable and is not a scalar value, it is safe
2191 to ignore the close phi node: it will be code generated in the
2192 out of Graphite pass. */
2193 else if (scev_analyzable_p (res, region))
2195 loop_p loop = loop_containing_stmt (SSA_NAME_DEF_STMT (res));
2196 tree scev;
2198 if (!loop_in_sese_p (loop, region))
2200 loop = loop_containing_stmt (SSA_NAME_DEF_STMT (arg));
2201 scev = scalar_evolution_in_region (region, loop, arg);
2202 scev = compute_overall_effect_of_inner_loop (loop, scev);
2204 else
2205 scev = scalar_evolution_in_region (region, loop, res);
2207 if (tree_does_not_contain_chrecs (scev))
2208 propagate_expr_outside_region (res, scev, region);
2210 gsi_next (psi);
2211 return;
2213 else
2215 tree zero_dim_array = create_zero_dim_array (res, "Close_Phi");
2217 stmt = gimple_build_assign (res, unshare_expr (zero_dim_array));
2219 if (TREE_CODE (arg) == SSA_NAME)
2220 insert_out_of_ssa_copy (scop, zero_dim_array, arg,
2221 SSA_NAME_DEF_STMT (arg));
2222 else
2223 insert_out_of_ssa_copy_on_edge (scop, single_pred_edge (bb),
2224 zero_dim_array, arg);
2227 remove_phi_node (psi, false);
2228 SSA_NAME_DEF_STMT (res) = stmt;
2230 insert_stmts (scop, stmt, NULL, gsi_after_labels (bb));
2233 /* Rewrite out of SSA the reduction phi node at PSI by creating a zero
2234 dimension array for it. */
2236 static void
2237 rewrite_phi_out_of_ssa (scop_p scop, gimple_stmt_iterator *psi)
2239 size_t i;
2240 gimple phi = gsi_stmt (*psi);
2241 basic_block bb = gimple_bb (phi);
2242 tree res = gimple_phi_result (phi);
2243 tree zero_dim_array = create_zero_dim_array (res, "phi_out_of_ssa");
2244 gimple stmt;
2246 for (i = 0; i < gimple_phi_num_args (phi); i++)
2248 tree arg = gimple_phi_arg_def (phi, i);
2249 edge e = gimple_phi_arg_edge (phi, i);
2251 /* Avoid the insertion of code in the loop latch to please the
2252 pattern matching of the vectorizer. */
2253 if (TREE_CODE (arg) == SSA_NAME
2254 && !SSA_NAME_IS_DEFAULT_DEF (arg)
2255 && e->src == bb->loop_father->latch)
2256 insert_out_of_ssa_copy (scop, zero_dim_array, arg,
2257 SSA_NAME_DEF_STMT (arg));
2258 else
2259 insert_out_of_ssa_copy_on_edge (scop, e, zero_dim_array, arg);
2262 stmt = gimple_build_assign (res, unshare_expr (zero_dim_array));
2263 remove_phi_node (psi, false);
2264 insert_stmts (scop, stmt, NULL, gsi_after_labels (bb));
2267 /* Rewrite the degenerate phi node at position PSI from the degenerate
2268 form "x = phi (y, y, ..., y)" to "x = y". */
2270 static void
2271 rewrite_degenerate_phi (gimple_stmt_iterator *psi)
2273 tree rhs;
2274 gimple stmt;
2275 gimple_stmt_iterator gsi;
2276 gimple phi = gsi_stmt (*psi);
2277 tree res = gimple_phi_result (phi);
2278 basic_block bb;
2280 bb = gimple_bb (phi);
2281 rhs = degenerate_phi_result (phi);
2282 gcc_assert (rhs);
2284 stmt = gimple_build_assign (res, rhs);
2285 remove_phi_node (psi, false);
2287 gsi = gsi_after_labels (bb);
2288 gsi_insert_before (&gsi, stmt, GSI_NEW_STMT);
2291 /* Rewrite out of SSA all the reduction phi nodes of SCOP. */
2293 static void
2294 rewrite_reductions_out_of_ssa (scop_p scop)
2296 basic_block bb;
2297 gimple_stmt_iterator psi;
2298 sese region = SCOP_REGION (scop);
2300 FOR_EACH_BB_FN (bb, cfun)
2301 if (bb_in_sese_p (bb, region))
2302 for (psi = gsi_start_phis (bb); !gsi_end_p (psi);)
2304 gimple phi = gsi_stmt (psi);
2306 if (virtual_operand_p (gimple_phi_result (phi)))
2308 gsi_next (&psi);
2309 continue;
2312 if (gimple_phi_num_args (phi) > 1
2313 && degenerate_phi_result (phi))
2314 rewrite_degenerate_phi (&psi);
2316 else if (scalar_close_phi_node_p (phi))
2317 rewrite_close_phi_out_of_ssa (scop, &psi);
2319 else if (reduction_phi_p (region, &psi))
2320 rewrite_phi_out_of_ssa (scop, &psi);
2323 update_ssa (TODO_update_ssa);
2324 #ifdef ENABLE_CHECKING
2325 verify_loop_closed_ssa (true);
2326 #endif
2329 /* Rewrite the scalar dependence of DEF used in USE_STMT with a memory
2330 read from ZERO_DIM_ARRAY. */
2332 static void
2333 rewrite_cross_bb_scalar_dependence (scop_p scop, tree zero_dim_array,
2334 tree def, gimple use_stmt)
2336 gimple name_stmt;
2337 tree name;
2338 ssa_op_iter iter;
2339 use_operand_p use_p;
2341 gcc_assert (gimple_code (use_stmt) != GIMPLE_PHI);
2343 name = copy_ssa_name (def, NULL);
2344 name_stmt = gimple_build_assign (name, zero_dim_array);
2346 gimple_assign_set_lhs (name_stmt, name);
2347 insert_stmts (scop, name_stmt, NULL, gsi_for_stmt (use_stmt));
2349 FOR_EACH_SSA_USE_OPERAND (use_p, use_stmt, iter, SSA_OP_ALL_USES)
2350 if (operand_equal_p (def, USE_FROM_PTR (use_p), 0))
2351 replace_exp (use_p, name);
2353 update_stmt (use_stmt);
2356 /* For every definition DEF in the SCOP that is used outside the scop,
2357 insert a closing-scop definition in the basic block just after this
2358 SCOP. */
2360 static void
2361 handle_scalar_deps_crossing_scop_limits (scop_p scop, tree def, gimple stmt)
2363 tree var = create_tmp_reg (TREE_TYPE (def), NULL);
2364 tree new_name = make_ssa_name (var, stmt);
2365 bool needs_copy = false;
2366 use_operand_p use_p;
2367 imm_use_iterator imm_iter;
2368 gimple use_stmt;
2369 sese region = SCOP_REGION (scop);
2371 FOR_EACH_IMM_USE_STMT (use_stmt, imm_iter, def)
2373 if (!bb_in_sese_p (gimple_bb (use_stmt), region))
2375 FOR_EACH_IMM_USE_ON_STMT (use_p, imm_iter)
2377 SET_USE (use_p, new_name);
2379 update_stmt (use_stmt);
2380 needs_copy = true;
2384 /* Insert in the empty BB just after the scop a use of DEF such
2385 that the rewrite of cross_bb_scalar_dependences won't insert
2386 arrays everywhere else. */
2387 if (needs_copy)
2389 gimple assign = gimple_build_assign (new_name, def);
2390 gimple_stmt_iterator psi = gsi_after_labels (SESE_EXIT (region)->dest);
2392 update_stmt (assign);
2393 gsi_insert_before (&psi, assign, GSI_SAME_STMT);
2397 /* Rewrite the scalar dependences crossing the boundary of the BB
2398 containing STMT with an array. Return true when something has been
2399 changed. */
2401 static bool
2402 rewrite_cross_bb_scalar_deps (scop_p scop, gimple_stmt_iterator *gsi)
2404 sese region = SCOP_REGION (scop);
2405 gimple stmt = gsi_stmt (*gsi);
2406 imm_use_iterator imm_iter;
2407 tree def;
2408 basic_block def_bb;
2409 tree zero_dim_array = NULL_TREE;
2410 gimple use_stmt;
2411 bool res = false;
2413 switch (gimple_code (stmt))
2415 case GIMPLE_ASSIGN:
2416 def = gimple_assign_lhs (stmt);
2417 break;
2419 case GIMPLE_CALL:
2420 def = gimple_call_lhs (stmt);
2421 break;
2423 default:
2424 return false;
2427 if (!def
2428 || !is_gimple_reg (def))
2429 return false;
2431 if (scev_analyzable_p (def, region))
2433 loop_p loop = loop_containing_stmt (SSA_NAME_DEF_STMT (def));
2434 tree scev = scalar_evolution_in_region (region, loop, def);
2436 if (tree_contains_chrecs (scev, NULL))
2437 return false;
2439 propagate_expr_outside_region (def, scev, region);
2440 return true;
2443 def_bb = gimple_bb (stmt);
2445 handle_scalar_deps_crossing_scop_limits (scop, def, stmt);
2447 FOR_EACH_IMM_USE_STMT (use_stmt, imm_iter, def)
2448 if (gimple_code (use_stmt) == GIMPLE_PHI
2449 && (res = true))
2451 gimple_stmt_iterator psi = gsi_for_stmt (use_stmt);
2453 if (scalar_close_phi_node_p (gsi_stmt (psi)))
2454 rewrite_close_phi_out_of_ssa (scop, &psi);
2455 else
2456 rewrite_phi_out_of_ssa (scop, &psi);
2459 FOR_EACH_IMM_USE_STMT (use_stmt, imm_iter, def)
2460 if (gimple_code (use_stmt) != GIMPLE_PHI
2461 && def_bb != gimple_bb (use_stmt)
2462 && !is_gimple_debug (use_stmt)
2463 && (res = true))
2465 if (!zero_dim_array)
2467 zero_dim_array = create_zero_dim_array
2468 (def, "Cross_BB_scalar_dependence");
2469 insert_out_of_ssa_copy (scop, zero_dim_array, def,
2470 SSA_NAME_DEF_STMT (def));
2471 gsi_next (gsi);
2474 rewrite_cross_bb_scalar_dependence (scop, unshare_expr (zero_dim_array),
2475 def, use_stmt);
2478 return res;
2481 /* Rewrite out of SSA all the reduction phi nodes of SCOP. */
2483 static void
2484 rewrite_cross_bb_scalar_deps_out_of_ssa (scop_p scop)
2486 basic_block bb;
2487 gimple_stmt_iterator psi;
2488 sese region = SCOP_REGION (scop);
2489 bool changed = false;
2491 /* Create an extra empty BB after the scop. */
2492 split_edge (SESE_EXIT (region));
2494 FOR_EACH_BB_FN (bb, cfun)
2495 if (bb_in_sese_p (bb, region))
2496 for (psi = gsi_start_bb (bb); !gsi_end_p (psi); gsi_next (&psi))
2497 changed |= rewrite_cross_bb_scalar_deps (scop, &psi);
2499 if (changed)
2501 scev_reset_htab ();
2502 update_ssa (TODO_update_ssa);
2503 #ifdef ENABLE_CHECKING
2504 verify_loop_closed_ssa (true);
2505 #endif
2509 /* Returns the number of pbbs that are in loops contained in SCOP. */
2511 static int
2512 nb_pbbs_in_loops (scop_p scop)
2514 int i;
2515 poly_bb_p pbb;
2516 int res = 0;
2518 FOR_EACH_VEC_ELT (SCOP_BBS (scop), i, pbb)
2519 if (loop_in_sese_p (gbb_loop (PBB_BLACK_BOX (pbb)), SCOP_REGION (scop)))
2520 res++;
2522 return res;
2525 /* Return the number of data references in BB that write in
2526 memory. */
2528 static int
2529 nb_data_writes_in_bb (basic_block bb)
2531 int res = 0;
2532 gimple_stmt_iterator gsi;
2534 for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi))
2535 if (gimple_vdef (gsi_stmt (gsi)))
2536 res++;
2538 return res;
2541 /* Splits at STMT the basic block BB represented as PBB in the
2542 polyhedral form. */
2544 static edge
2545 split_pbb (scop_p scop, poly_bb_p pbb, basic_block bb, gimple stmt)
2547 edge e1 = split_block (bb, stmt);
2548 new_pbb_from_pbb (scop, pbb, e1->dest);
2549 return e1;
2552 /* Splits STMT out of its current BB. This is done for reduction
2553 statements for which we want to ignore data dependences. */
2555 static basic_block
2556 split_reduction_stmt (scop_p scop, gimple stmt)
2558 basic_block bb = gimple_bb (stmt);
2559 poly_bb_p pbb = pbb_from_bb (bb);
2560 gimple_bb_p gbb = gbb_from_bb (bb);
2561 edge e1;
2562 int i;
2563 data_reference_p dr;
2565 /* Do not split basic blocks with no writes to memory: the reduction
2566 will be the only write to memory. */
2567 if (nb_data_writes_in_bb (bb) == 0
2568 /* Or if we have already marked BB as a reduction. */
2569 || PBB_IS_REDUCTION (pbb_from_bb (bb)))
2570 return bb;
2572 e1 = split_pbb (scop, pbb, bb, stmt);
2574 /* Split once more only when the reduction stmt is not the only one
2575 left in the original BB. */
2576 if (!gsi_one_before_end_p (gsi_start_nondebug_bb (bb)))
2578 gimple_stmt_iterator gsi = gsi_last_bb (bb);
2579 gsi_prev (&gsi);
2580 e1 = split_pbb (scop, pbb, bb, gsi_stmt (gsi));
2583 /* A part of the data references will end in a different basic block
2584 after the split: move the DRs from the original GBB to the newly
2585 created GBB1. */
2586 FOR_EACH_VEC_ELT (GBB_DATA_REFS (gbb), i, dr)
2588 basic_block bb1 = gimple_bb (DR_STMT (dr));
2590 if (bb1 != bb)
2592 gimple_bb_p gbb1 = gbb_from_bb (bb1);
2593 GBB_DATA_REFS (gbb1).safe_push (dr);
2594 GBB_DATA_REFS (gbb).ordered_remove (i);
2595 i--;
2599 return e1->dest;
2602 /* Return true when stmt is a reduction operation. */
2604 static inline bool
2605 is_reduction_operation_p (gimple stmt)
2607 enum tree_code code;
2609 gcc_assert (is_gimple_assign (stmt));
2610 code = gimple_assign_rhs_code (stmt);
2612 return flag_associative_math
2613 && commutative_tree_code (code)
2614 && associative_tree_code (code);
2617 /* Returns true when PHI contains an argument ARG. */
2619 static bool
2620 phi_contains_arg (gimple phi, tree arg)
2622 size_t i;
2624 for (i = 0; i < gimple_phi_num_args (phi); i++)
2625 if (operand_equal_p (arg, gimple_phi_arg_def (phi, i), 0))
2626 return true;
2628 return false;
2631 /* Return a loop phi node that corresponds to a reduction containing LHS. */
2633 static gimple
2634 follow_ssa_with_commutative_ops (tree arg, tree lhs)
2636 gimple stmt;
2638 if (TREE_CODE (arg) != SSA_NAME)
2639 return NULL;
2641 stmt = SSA_NAME_DEF_STMT (arg);
2643 if (gimple_code (stmt) == GIMPLE_NOP
2644 || gimple_code (stmt) == GIMPLE_CALL)
2645 return NULL;
2647 if (gimple_code (stmt) == GIMPLE_PHI)
2649 if (phi_contains_arg (stmt, lhs))
2650 return stmt;
2651 return NULL;
2654 if (!is_gimple_assign (stmt))
2655 return NULL;
2657 if (gimple_num_ops (stmt) == 2)
2658 return follow_ssa_with_commutative_ops (gimple_assign_rhs1 (stmt), lhs);
2660 if (is_reduction_operation_p (stmt))
2662 gimple res = follow_ssa_with_commutative_ops (gimple_assign_rhs1 (stmt), lhs);
2664 return res ? res :
2665 follow_ssa_with_commutative_ops (gimple_assign_rhs2 (stmt), lhs);
2668 return NULL;
2671 /* Detect commutative and associative scalar reductions starting at
2672 the STMT. Return the phi node of the reduction cycle, or NULL. */
2674 static gimple
2675 detect_commutative_reduction_arg (tree lhs, gimple stmt, tree arg,
2676 vec<gimple> *in,
2677 vec<gimple> *out)
2679 gimple phi = follow_ssa_with_commutative_ops (arg, lhs);
2681 if (!phi)
2682 return NULL;
2684 in->safe_push (stmt);
2685 out->safe_push (stmt);
2686 return phi;
2689 /* Detect commutative and associative scalar reductions starting at
2690 STMT. Return the phi node of the reduction cycle, or NULL. */
2692 static gimple
2693 detect_commutative_reduction_assign (gimple stmt, vec<gimple> *in,
2694 vec<gimple> *out)
2696 tree lhs = gimple_assign_lhs (stmt);
2698 if (gimple_num_ops (stmt) == 2)
2699 return detect_commutative_reduction_arg (lhs, stmt,
2700 gimple_assign_rhs1 (stmt),
2701 in, out);
2703 if (is_reduction_operation_p (stmt))
2705 gimple res = detect_commutative_reduction_arg (lhs, stmt,
2706 gimple_assign_rhs1 (stmt),
2707 in, out);
2708 return res ? res
2709 : detect_commutative_reduction_arg (lhs, stmt,
2710 gimple_assign_rhs2 (stmt),
2711 in, out);
2714 return NULL;
2717 /* Return a loop phi node that corresponds to a reduction containing LHS. */
2719 static gimple
2720 follow_inital_value_to_phi (tree arg, tree lhs)
2722 gimple stmt;
2724 if (!arg || TREE_CODE (arg) != SSA_NAME)
2725 return NULL;
2727 stmt = SSA_NAME_DEF_STMT (arg);
2729 if (gimple_code (stmt) == GIMPLE_PHI
2730 && phi_contains_arg (stmt, lhs))
2731 return stmt;
2733 return NULL;
2737 /* Return the argument of the loop PHI that is the initial value coming
2738 from outside the loop. */
2740 static edge
2741 edge_initial_value_for_loop_phi (gimple phi)
2743 size_t i;
2745 for (i = 0; i < gimple_phi_num_args (phi); i++)
2747 edge e = gimple_phi_arg_edge (phi, i);
2749 if (loop_depth (e->src->loop_father)
2750 < loop_depth (e->dest->loop_father))
2751 return e;
2754 return NULL;
2757 /* Return the argument of the loop PHI that is the initial value coming
2758 from outside the loop. */
2760 static tree
2761 initial_value_for_loop_phi (gimple phi)
2763 size_t i;
2765 for (i = 0; i < gimple_phi_num_args (phi); i++)
2767 edge e = gimple_phi_arg_edge (phi, i);
2769 if (loop_depth (e->src->loop_father)
2770 < loop_depth (e->dest->loop_father))
2771 return gimple_phi_arg_def (phi, i);
2774 return NULL_TREE;
2777 /* Returns true when DEF is used outside the reduction cycle of
2778 LOOP_PHI. */
2780 static bool
2781 used_outside_reduction (tree def, gimple loop_phi)
2783 use_operand_p use_p;
2784 imm_use_iterator imm_iter;
2785 loop_p loop = loop_containing_stmt (loop_phi);
2787 /* In LOOP, DEF should be used only in LOOP_PHI. */
2788 FOR_EACH_IMM_USE_FAST (use_p, imm_iter, def)
2790 gimple stmt = USE_STMT (use_p);
2792 if (stmt != loop_phi
2793 && !is_gimple_debug (stmt)
2794 && flow_bb_inside_loop_p (loop, gimple_bb (stmt)))
2795 return true;
2798 return false;
2801 /* Detect commutative and associative scalar reductions belonging to
2802 the SCOP starting at the loop closed phi node STMT. Return the phi
2803 node of the reduction cycle, or NULL. */
2805 static gimple
2806 detect_commutative_reduction (scop_p scop, gimple stmt, vec<gimple> *in,
2807 vec<gimple> *out)
2809 if (scalar_close_phi_node_p (stmt))
2811 gimple def, loop_phi, phi, close_phi = stmt;
2812 tree init, lhs, arg = gimple_phi_arg_def (close_phi, 0);
2814 if (TREE_CODE (arg) != SSA_NAME)
2815 return NULL;
2817 /* Note that loop close phi nodes should have a single argument
2818 because we translated the representation into a canonical form
2819 before Graphite: see canonicalize_loop_closed_ssa_form. */
2820 gcc_assert (gimple_phi_num_args (close_phi) == 1);
2822 def = SSA_NAME_DEF_STMT (arg);
2823 if (!stmt_in_sese_p (def, SCOP_REGION (scop))
2824 || !(loop_phi = detect_commutative_reduction (scop, def, in, out)))
2825 return NULL;
2827 lhs = gimple_phi_result (close_phi);
2828 init = initial_value_for_loop_phi (loop_phi);
2829 phi = follow_inital_value_to_phi (init, lhs);
2831 if (phi && (used_outside_reduction (lhs, phi)
2832 || !has_single_use (gimple_phi_result (phi))))
2833 return NULL;
2835 in->safe_push (loop_phi);
2836 out->safe_push (close_phi);
2837 return phi;
2840 if (gimple_code (stmt) == GIMPLE_ASSIGN)
2841 return detect_commutative_reduction_assign (stmt, in, out);
2843 return NULL;
2846 /* Translate the scalar reduction statement STMT to an array RED
2847 knowing that its recursive phi node is LOOP_PHI. */
2849 static void
2850 translate_scalar_reduction_to_array_for_stmt (scop_p scop, tree red,
2851 gimple stmt, gimple loop_phi)
2853 tree res = gimple_phi_result (loop_phi);
2854 gimple assign = gimple_build_assign (res, unshare_expr (red));
2855 gimple_stmt_iterator gsi;
2857 insert_stmts (scop, assign, NULL, gsi_after_labels (gimple_bb (loop_phi)));
2859 assign = gimple_build_assign (unshare_expr (red), gimple_assign_lhs (stmt));
2860 gsi = gsi_for_stmt (stmt);
2861 gsi_next (&gsi);
2862 insert_stmts (scop, assign, NULL, gsi);
2865 /* Removes the PHI node and resets all the debug stmts that are using
2866 the PHI_RESULT. */
2868 static void
2869 remove_phi (gimple phi)
2871 imm_use_iterator imm_iter;
2872 tree def;
2873 use_operand_p use_p;
2874 gimple_stmt_iterator gsi;
2875 auto_vec<gimple, 3> update;
2876 unsigned int i;
2877 gimple stmt;
2879 def = PHI_RESULT (phi);
2880 FOR_EACH_IMM_USE_FAST (use_p, imm_iter, def)
2882 stmt = USE_STMT (use_p);
2884 if (is_gimple_debug (stmt))
2886 gimple_debug_bind_reset_value (stmt);
2887 update.safe_push (stmt);
2891 FOR_EACH_VEC_ELT (update, i, stmt)
2892 update_stmt (stmt);
2894 gsi = gsi_for_phi_node (phi);
2895 remove_phi_node (&gsi, false);
2898 /* Helper function for for_each_index. For each INDEX of the data
2899 reference REF, returns true when its indices are valid in the loop
2900 nest LOOP passed in as DATA. */
2902 static bool
2903 dr_indices_valid_in_loop (tree ref ATTRIBUTE_UNUSED, tree *index, void *data)
2905 loop_p loop;
2906 basic_block header, def_bb;
2907 gimple stmt;
2909 if (TREE_CODE (*index) != SSA_NAME)
2910 return true;
2912 loop = *((loop_p *) data);
2913 header = loop->header;
2914 stmt = SSA_NAME_DEF_STMT (*index);
2916 if (!stmt)
2917 return true;
2919 def_bb = gimple_bb (stmt);
2921 if (!def_bb)
2922 return true;
2924 return dominated_by_p (CDI_DOMINATORS, header, def_bb);
2927 /* When the result of a CLOSE_PHI is written to a memory location,
2928 return a pointer to that memory reference, otherwise return
2929 NULL_TREE. */
2931 static tree
2932 close_phi_written_to_memory (gimple close_phi)
2934 imm_use_iterator imm_iter;
2935 use_operand_p use_p;
2936 gimple stmt;
2937 tree res, def = gimple_phi_result (close_phi);
2939 FOR_EACH_IMM_USE_FAST (use_p, imm_iter, def)
2940 if ((stmt = USE_STMT (use_p))
2941 && gimple_code (stmt) == GIMPLE_ASSIGN
2942 && (res = gimple_assign_lhs (stmt)))
2944 switch (TREE_CODE (res))
2946 case VAR_DECL:
2947 case PARM_DECL:
2948 case RESULT_DECL:
2949 return res;
2951 case ARRAY_REF:
2952 case MEM_REF:
2954 tree arg = gimple_phi_arg_def (close_phi, 0);
2955 loop_p nest = loop_containing_stmt (SSA_NAME_DEF_STMT (arg));
2957 /* FIXME: this restriction is for id-{24,25}.f and
2958 could be handled by duplicating the computation of
2959 array indices before the loop of the close_phi. */
2960 if (for_each_index (&res, dr_indices_valid_in_loop, &nest))
2961 return res;
2963 /* Fallthru. */
2965 default:
2966 continue;
2969 return NULL_TREE;
2972 /* Rewrite out of SSA the reduction described by the loop phi nodes
2973 IN, and the close phi nodes OUT. IN and OUT are structured by loop
2974 levels like this:
2976 IN: stmt, loop_n, ..., loop_0
2977 OUT: stmt, close_n, ..., close_0
2979 the first element is the reduction statement, and the next elements
2980 are the loop and close phi nodes of each of the outer loops. */
2982 static void
2983 translate_scalar_reduction_to_array (scop_p scop,
2984 vec<gimple> in,
2985 vec<gimple> out)
2987 gimple loop_phi;
2988 unsigned int i = out.length () - 1;
2989 tree red = close_phi_written_to_memory (out[i]);
2991 FOR_EACH_VEC_ELT (in, i, loop_phi)
2993 gimple close_phi = out[i];
2995 if (i == 0)
2997 gimple stmt = loop_phi;
2998 basic_block bb = split_reduction_stmt (scop, stmt);
2999 poly_bb_p pbb = pbb_from_bb (bb);
3000 PBB_IS_REDUCTION (pbb) = true;
3001 gcc_assert (close_phi == loop_phi);
3003 if (!red)
3004 red = create_zero_dim_array
3005 (gimple_assign_lhs (stmt), "Commutative_Associative_Reduction");
3007 translate_scalar_reduction_to_array_for_stmt (scop, red, stmt, in[1]);
3008 continue;
3011 if (i == in.length () - 1)
3013 insert_out_of_ssa_copy (scop, gimple_phi_result (close_phi),
3014 unshare_expr (red), close_phi);
3015 insert_out_of_ssa_copy_on_edge
3016 (scop, edge_initial_value_for_loop_phi (loop_phi),
3017 unshare_expr (red), initial_value_for_loop_phi (loop_phi));
3020 remove_phi (loop_phi);
3021 remove_phi (close_phi);
3025 /* Rewrites out of SSA a commutative reduction at CLOSE_PHI. Returns
3026 true when something has been changed. */
3028 static bool
3029 rewrite_commutative_reductions_out_of_ssa_close_phi (scop_p scop,
3030 gimple close_phi)
3032 bool res;
3033 auto_vec<gimple, 10> in;
3034 auto_vec<gimple, 10> out;
3036 detect_commutative_reduction (scop, close_phi, &in, &out);
3037 res = in.length () > 1;
3038 if (res)
3039 translate_scalar_reduction_to_array (scop, in, out);
3041 return res;
3044 /* Rewrites all the commutative reductions from LOOP out of SSA.
3045 Returns true when something has been changed. */
3047 static bool
3048 rewrite_commutative_reductions_out_of_ssa_loop (scop_p scop,
3049 loop_p loop)
3051 gimple_stmt_iterator gsi;
3052 edge exit = single_exit (loop);
3053 tree res;
3054 bool changed = false;
3056 if (!exit)
3057 return false;
3059 for (gsi = gsi_start_phis (exit->dest); !gsi_end_p (gsi); gsi_next (&gsi))
3060 if ((res = gimple_phi_result (gsi_stmt (gsi)))
3061 && !virtual_operand_p (res)
3062 && !scev_analyzable_p (res, SCOP_REGION (scop)))
3063 changed |= rewrite_commutative_reductions_out_of_ssa_close_phi
3064 (scop, gsi_stmt (gsi));
3066 return changed;
3069 /* Rewrites all the commutative reductions from SCOP out of SSA. */
3071 static void
3072 rewrite_commutative_reductions_out_of_ssa (scop_p scop)
3074 loop_p loop;
3075 bool changed = false;
3076 sese region = SCOP_REGION (scop);
3078 FOR_EACH_LOOP (loop, 0)
3079 if (loop_in_sese_p (loop, region))
3080 changed |= rewrite_commutative_reductions_out_of_ssa_loop (scop, loop);
3082 if (changed)
3084 scev_reset_htab ();
3085 gsi_commit_edge_inserts ();
3086 update_ssa (TODO_update_ssa);
3087 #ifdef ENABLE_CHECKING
3088 verify_loop_closed_ssa (true);
3089 #endif
3093 /* Can all ivs be represented by a signed integer?
3094 As CLooG might generate negative values in its expressions, signed loop ivs
3095 are required in the backend. */
3097 static bool
3098 scop_ivs_can_be_represented (scop_p scop)
3100 loop_p loop;
3101 gimple_stmt_iterator psi;
3102 bool result = true;
3104 FOR_EACH_LOOP (loop, 0)
3106 if (!loop_in_sese_p (loop, SCOP_REGION (scop)))
3107 continue;
3109 for (psi = gsi_start_phis (loop->header);
3110 !gsi_end_p (psi); gsi_next (&psi))
3112 gimple phi = gsi_stmt (psi);
3113 tree res = PHI_RESULT (phi);
3114 tree type = TREE_TYPE (res);
3116 if (TYPE_UNSIGNED (type)
3117 && TYPE_PRECISION (type) >= TYPE_PRECISION (long_long_integer_type_node))
3119 result = false;
3120 break;
3123 if (!result)
3124 break;
3127 return result;
3130 /* Builds the polyhedral representation for a SESE region. */
3132 void
3133 build_poly_scop (scop_p scop)
3135 sese region = SCOP_REGION (scop);
3136 graphite_dim_t max_dim;
3138 build_scop_bbs (scop);
3140 /* FIXME: This restriction is needed to avoid a problem in CLooG.
3141 Once CLooG is fixed, remove this guard. Anyways, it makes no
3142 sense to optimize a scop containing only PBBs that do not belong
3143 to any loops. */
3144 if (nb_pbbs_in_loops (scop) == 0)
3145 return;
3147 if (!scop_ivs_can_be_represented (scop))
3148 return;
3150 if (flag_associative_math)
3151 rewrite_commutative_reductions_out_of_ssa (scop);
3153 build_sese_loop_nests (region);
3154 /* Record all conditions in REGION. */
3155 sese_dom_walker (CDI_DOMINATORS, region).walk (cfun->cfg->x_entry_block_ptr);
3156 find_scop_parameters (scop);
3158 max_dim = PARAM_VALUE (PARAM_GRAPHITE_MAX_NB_SCOP_PARAMS);
3159 if (scop_nb_params (scop) > max_dim)
3160 return;
3162 build_scop_iteration_domain (scop);
3163 build_scop_context (scop);
3164 add_conditions_to_constraints (scop);
3166 /* Rewrite out of SSA only after having translated the
3167 representation to the polyhedral representation to avoid scev
3168 analysis failures. That means that these functions will insert
3169 new data references that they create in the right place. */
3170 rewrite_reductions_out_of_ssa (scop);
3171 rewrite_cross_bb_scalar_deps_out_of_ssa (scop);
3173 build_scop_drs (scop);
3174 scop_to_lst (scop);
3175 build_scop_scattering (scop);
3177 /* This SCoP has been translated to the polyhedral
3178 representation. */
3179 POLY_SCOP_P (scop) = true;
3181 #endif