libgomp: Use pthread mutexes in the nvptx plugin.
[official-gcc.git] / gcc / graphite-sese-to-poly.c
blob5bb521ff1032c982741f825d0ce8d3efa5f279ac
1 /* Conversion of SESE regions to Polyhedra.
2 Copyright (C) 2009-2015 Free Software Foundation, Inc.
3 Contributed by Sebastian Pop <sebastian.pop@amd.com>.
5 This file is part of GCC.
7 GCC is free software; you can redistribute it and/or modify
8 it under the terms of the GNU General Public License as published by
9 the Free Software Foundation; either version 3, or (at your option)
10 any later version.
12 GCC is distributed in the hope that it will be useful,
13 but WITHOUT ANY WARRANTY; without even the implied warranty of
14 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
15 GNU General Public License for more details.
17 You should have received a copy of the GNU General Public License
18 along with GCC; see the file COPYING3. If not see
19 <http://www.gnu.org/licenses/>. */
21 #include "config.h"
23 #ifdef HAVE_isl
24 #include <isl/set.h>
25 #include <isl/map.h>
26 #include <isl/union_map.h>
27 #include <isl/constraint.h>
28 #include <isl/aff.h>
29 #include <isl/val.h>
31 /* Since ISL-0.13, the extern is in val_gmp.h. */
32 #if !defined(HAVE_ISL_SCHED_CONSTRAINTS_COMPUTE_SCHEDULE) && defined(__cplusplus)
33 extern "C" {
34 #endif
35 #include <isl/val_gmp.h>
36 #if !defined(HAVE_ISL_SCHED_CONSTRAINTS_COMPUTE_SCHEDULE) && defined(__cplusplus)
38 #endif
39 #endif
41 #include "system.h"
42 #include "coretypes.h"
43 #include "hash-set.h"
44 #include "machmode.h"
45 #include "vec.h"
46 #include "double-int.h"
47 #include "input.h"
48 #include "alias.h"
49 #include "symtab.h"
50 #include "options.h"
51 #include "wide-int.h"
52 #include "inchash.h"
53 #include "tree.h"
54 #include "fold-const.h"
55 #include "predict.h"
56 #include "tm.h"
57 #include "hard-reg-set.h"
58 #include "input.h"
59 #include "function.h"
60 #include "dominance.h"
61 #include "cfg.h"
62 #include "basic-block.h"
63 #include "tree-ssa-alias.h"
64 #include "internal-fn.h"
65 #include "gimple-expr.h"
66 #include "is-a.h"
67 #include "gimple.h"
68 #include "gimple-iterator.h"
69 #include "gimplify.h"
70 #include "gimplify-me.h"
71 #include "gimple-ssa.h"
72 #include "tree-cfg.h"
73 #include "tree-phinodes.h"
74 #include "ssa-iterators.h"
75 #include "stringpool.h"
76 #include "tree-ssanames.h"
77 #include "tree-ssa-loop-manip.h"
78 #include "tree-ssa-loop-niter.h"
79 #include "tree-ssa-loop.h"
80 #include "tree-into-ssa.h"
81 #include "tree-pass.h"
82 #include "cfgloop.h"
83 #include "tree-chrec.h"
84 #include "tree-data-ref.h"
85 #include "tree-scalar-evolution.h"
86 #include "domwalk.h"
87 #include "sese.h"
88 #include "tree-ssa-propagate.h"
90 #ifdef HAVE_isl
91 #include "expr.h"
92 #include "graphite-poly.h"
93 #include "graphite-sese-to-poly.h"
96 /* Assigns to RES the value of the INTEGER_CST T. */
98 static inline void
99 tree_int_to_gmp (tree t, mpz_t res)
101 wi::to_mpz (t, res, TYPE_SIGN (TREE_TYPE (t)));
104 /* Returns the index of the PHI argument defined in the outermost
105 loop. */
107 static size_t
108 phi_arg_in_outermost_loop (gphi *phi)
110 loop_p loop = gimple_bb (phi)->loop_father;
111 size_t i, res = 0;
113 for (i = 0; i < gimple_phi_num_args (phi); i++)
114 if (!flow_bb_inside_loop_p (loop, gimple_phi_arg_edge (phi, i)->src))
116 loop = gimple_phi_arg_edge (phi, i)->src->loop_father;
117 res = i;
120 return res;
123 /* Removes a simple copy phi node "RES = phi (INIT, RES)" at position
124 PSI by inserting on the loop ENTRY edge assignment "RES = INIT". */
126 static void
127 remove_simple_copy_phi (gphi_iterator *psi)
129 gphi *phi = psi->phi ();
130 tree res = gimple_phi_result (phi);
131 size_t entry = phi_arg_in_outermost_loop (phi);
132 tree init = gimple_phi_arg_def (phi, entry);
133 gassign *stmt = gimple_build_assign (res, init);
134 edge e = gimple_phi_arg_edge (phi, entry);
136 remove_phi_node (psi, false);
137 gsi_insert_on_edge_immediate (e, stmt);
140 /* Removes an invariant phi node at position PSI by inserting on the
141 loop ENTRY edge the assignment RES = INIT. */
143 static void
144 remove_invariant_phi (sese region, gphi_iterator *psi)
146 gphi *phi = psi->phi ();
147 loop_p loop = loop_containing_stmt (phi);
148 tree res = gimple_phi_result (phi);
149 tree scev = scalar_evolution_in_region (region, loop, res);
150 size_t entry = phi_arg_in_outermost_loop (phi);
151 edge e = gimple_phi_arg_edge (phi, entry);
152 tree var;
153 gassign *stmt;
154 gimple_seq stmts = NULL;
156 if (tree_contains_chrecs (scev, NULL))
157 scev = gimple_phi_arg_def (phi, entry);
159 var = force_gimple_operand (scev, &stmts, true, NULL_TREE);
160 stmt = gimple_build_assign (res, var);
161 remove_phi_node (psi, false);
163 gimple_seq_add_stmt (&stmts, stmt);
164 gsi_insert_seq_on_edge (e, stmts);
165 gsi_commit_edge_inserts ();
166 SSA_NAME_DEF_STMT (res) = stmt;
169 /* Returns true when the phi node at PSI is of the form "a = phi (a, x)". */
171 static inline bool
172 simple_copy_phi_p (gphi *phi)
174 tree res;
176 if (gimple_phi_num_args (phi) != 2)
177 return false;
179 res = gimple_phi_result (phi);
180 return (res == gimple_phi_arg_def (phi, 0)
181 || res == gimple_phi_arg_def (phi, 1));
184 /* Returns true when the phi node at position PSI is a reduction phi
185 node in REGION. Otherwise moves the pointer PSI to the next phi to
186 be considered. */
188 static bool
189 reduction_phi_p (sese region, gphi_iterator *psi)
191 loop_p loop;
192 gphi *phi = psi->phi ();
193 tree res = gimple_phi_result (phi);
195 loop = loop_containing_stmt (phi);
197 if (simple_copy_phi_p (phi))
199 /* PRE introduces phi nodes like these, for an example,
200 see id-5.f in the fortran graphite testsuite:
202 # prephitmp.85_265 = PHI <prephitmp.85_258(33), prephitmp.85_265(18)>
204 remove_simple_copy_phi (psi);
205 return false;
208 if (scev_analyzable_p (res, region))
210 tree scev = scalar_evolution_in_region (region, loop, res);
212 if (evolution_function_is_invariant_p (scev, loop->num))
213 remove_invariant_phi (region, psi);
214 else
215 gsi_next (psi);
217 return false;
220 /* All the other cases are considered reductions. */
221 return true;
224 /* Store the GRAPHITE representation of BB. */
226 static gimple_bb_p
227 new_gimple_bb (basic_block bb, vec<data_reference_p> drs)
229 struct gimple_bb *gbb;
231 gbb = XNEW (struct gimple_bb);
232 bb->aux = gbb;
233 GBB_BB (gbb) = bb;
234 GBB_DATA_REFS (gbb) = drs;
235 GBB_CONDITIONS (gbb).create (0);
236 GBB_CONDITION_CASES (gbb).create (0);
238 return gbb;
241 static void
242 free_data_refs_aux (vec<data_reference_p> datarefs)
244 unsigned int i;
245 struct data_reference *dr;
247 FOR_EACH_VEC_ELT (datarefs, i, dr)
248 if (dr->aux)
250 base_alias_pair *bap = (base_alias_pair *)(dr->aux);
252 free (bap->alias_set);
254 free (bap);
255 dr->aux = NULL;
258 /* Frees GBB. */
260 static void
261 free_gimple_bb (struct gimple_bb *gbb)
263 free_data_refs_aux (GBB_DATA_REFS (gbb));
264 free_data_refs (GBB_DATA_REFS (gbb));
266 GBB_CONDITIONS (gbb).release ();
267 GBB_CONDITION_CASES (gbb).release ();
268 GBB_BB (gbb)->aux = 0;
269 XDELETE (gbb);
272 /* Deletes all gimple bbs in SCOP. */
274 static void
275 remove_gbbs_in_scop (scop_p scop)
277 int i;
278 poly_bb_p pbb;
280 FOR_EACH_VEC_ELT (SCOP_BBS (scop), i, pbb)
281 free_gimple_bb (PBB_BLACK_BOX (pbb));
284 /* Deletes all scops in SCOPS. */
286 void
287 free_scops (vec<scop_p> scops)
289 int i;
290 scop_p scop;
292 FOR_EACH_VEC_ELT (scops, i, scop)
294 remove_gbbs_in_scop (scop);
295 free_sese (SCOP_REGION (scop));
296 free_scop (scop);
299 scops.release ();
302 /* Same as outermost_loop_in_sese, returns the outermost loop
303 containing BB in REGION, but makes sure that the returned loop
304 belongs to the REGION, and so this returns the first loop in the
305 REGION when the loop containing BB does not belong to REGION. */
307 static loop_p
308 outermost_loop_in_sese_1 (sese region, basic_block bb)
310 loop_p nest = outermost_loop_in_sese (region, bb);
312 if (loop_in_sese_p (nest, region))
313 return nest;
315 /* When the basic block BB does not belong to a loop in the region,
316 return the first loop in the region. */
317 nest = nest->inner;
318 while (nest)
319 if (loop_in_sese_p (nest, region))
320 break;
321 else
322 nest = nest->next;
324 gcc_assert (nest);
325 return nest;
328 /* Generates a polyhedral black box only if the bb contains interesting
329 information. */
331 static gimple_bb_p
332 try_generate_gimple_bb (scop_p scop, basic_block bb)
334 vec<data_reference_p> drs;
335 drs.create (5);
336 sese region = SCOP_REGION (scop);
337 loop_p nest = outermost_loop_in_sese_1 (region, bb);
338 gimple_stmt_iterator gsi;
340 for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi))
342 gimple stmt = gsi_stmt (gsi);
343 loop_p loop;
345 if (is_gimple_debug (stmt))
346 continue;
348 loop = loop_containing_stmt (stmt);
349 if (!loop_in_sese_p (loop, region))
350 loop = nest;
352 graphite_find_data_references_in_stmt (nest, loop, stmt, &drs);
355 return new_gimple_bb (bb, drs);
358 /* Returns true if all predecessors of BB, that are not dominated by BB, are
359 marked in MAP. The predecessors dominated by BB are loop latches and will
360 be handled after BB. */
362 static bool
363 all_non_dominated_preds_marked_p (basic_block bb, sbitmap map)
365 edge e;
366 edge_iterator ei;
368 FOR_EACH_EDGE (e, ei, bb->preds)
369 if (!bitmap_bit_p (map, e->src->index)
370 && !dominated_by_p (CDI_DOMINATORS, e->src, bb))
371 return false;
373 return true;
376 /* Compare the depth of two basic_block's P1 and P2. */
378 static int
379 compare_bb_depths (const void *p1, const void *p2)
381 const_basic_block const bb1 = *(const_basic_block const*)p1;
382 const_basic_block const bb2 = *(const_basic_block const*)p2;
383 int d1 = loop_depth (bb1->loop_father);
384 int d2 = loop_depth (bb2->loop_father);
386 if (d1 < d2)
387 return 1;
389 if (d1 > d2)
390 return -1;
392 return 0;
395 /* Sort the basic blocks from DOM such that the first are the ones at
396 a deepest loop level. */
398 static void
399 graphite_sort_dominated_info (vec<basic_block> dom)
401 dom.qsort (compare_bb_depths);
404 /* Recursive helper function for build_scops_bbs. */
406 static void
407 build_scop_bbs_1 (scop_p scop, sbitmap visited, basic_block bb)
409 sese region = SCOP_REGION (scop);
410 vec<basic_block> dom;
411 poly_bb_p pbb;
413 if (bitmap_bit_p (visited, bb->index)
414 || !bb_in_sese_p (bb, region))
415 return;
417 pbb = new_poly_bb (scop, try_generate_gimple_bb (scop, bb));
418 SCOP_BBS (scop).safe_push (pbb);
419 bitmap_set_bit (visited, bb->index);
421 dom = get_dominated_by (CDI_DOMINATORS, bb);
423 if (!dom.exists ())
424 return;
426 graphite_sort_dominated_info (dom);
428 while (!dom.is_empty ())
430 int i;
431 basic_block dom_bb;
433 FOR_EACH_VEC_ELT (dom, i, dom_bb)
434 if (all_non_dominated_preds_marked_p (dom_bb, visited))
436 build_scop_bbs_1 (scop, visited, dom_bb);
437 dom.unordered_remove (i);
438 break;
442 dom.release ();
445 /* Gather the basic blocks belonging to the SCOP. */
447 static void
448 build_scop_bbs (scop_p scop)
450 sbitmap visited = sbitmap_alloc (last_basic_block_for_fn (cfun));
451 sese region = SCOP_REGION (scop);
453 bitmap_clear (visited);
454 build_scop_bbs_1 (scop, visited, SESE_ENTRY_BB (region));
455 sbitmap_free (visited);
458 /* Return an ISL identifier for the polyhedral basic block PBB. */
460 static isl_id *
461 isl_id_for_pbb (scop_p s, poly_bb_p pbb)
463 char name[50];
464 snprintf (name, sizeof (name), "S_%d", pbb_index (pbb));
465 return isl_id_alloc (s->ctx, name, pbb);
468 /* Converts the STATIC_SCHEDULE of PBB into a scattering polyhedron.
469 We generate SCATTERING_DIMENSIONS scattering dimensions.
471 CLooG 0.15.0 and previous versions require, that all
472 scattering functions of one CloogProgram have the same number of
473 scattering dimensions, therefore we allow to specify it. This
474 should be removed in future versions of CLooG.
476 The scattering polyhedron consists of these dimensions: scattering,
477 loop_iterators, parameters.
479 Example:
481 | scattering_dimensions = 5
482 | used_scattering_dimensions = 3
483 | nb_iterators = 1
484 | scop_nb_params = 2
486 | Schedule:
488 | 4 5
490 | Scattering polyhedron:
492 | scattering: {s1, s2, s3, s4, s5}
493 | loop_iterators: {i}
494 | parameters: {p1, p2}
496 | s1 s2 s3 s4 s5 i p1 p2 1
497 | 1 0 0 0 0 0 0 0 -4 = 0
498 | 0 1 0 0 0 -1 0 0 0 = 0
499 | 0 0 1 0 0 0 0 0 -5 = 0 */
501 static void
502 build_pbb_scattering_polyhedrons (isl_aff *static_sched,
503 poly_bb_p pbb, int scattering_dimensions)
505 int i;
506 int nb_iterators = pbb_dim_iter_domain (pbb);
507 int used_scattering_dimensions = nb_iterators * 2 + 1;
508 isl_val *val;
509 isl_space *dc, *dm;
511 gcc_assert (scattering_dimensions >= used_scattering_dimensions);
513 dc = isl_set_get_space (pbb->domain);
514 dm = isl_space_add_dims (isl_space_from_domain (dc),
515 isl_dim_out, scattering_dimensions);
516 pbb->schedule = isl_map_universe (dm);
518 for (i = 0; i < scattering_dimensions; i++)
520 /* Textual order inside this loop. */
521 if ((i % 2) == 0)
523 isl_constraint *c = isl_equality_alloc
524 (isl_local_space_from_space (isl_map_get_space (pbb->schedule)));
526 val = isl_aff_get_coefficient_val (static_sched, isl_dim_in, i / 2);
528 val = isl_val_neg (val);
529 c = isl_constraint_set_constant_val (c, val);
530 c = isl_constraint_set_coefficient_si (c, isl_dim_out, i, 1);
531 pbb->schedule = isl_map_add_constraint (pbb->schedule, c);
534 /* Iterations of this loop. */
535 else /* if ((i % 2) == 1) */
537 int loop = (i - 1) / 2;
538 pbb->schedule = isl_map_equate (pbb->schedule, isl_dim_in, loop,
539 isl_dim_out, i);
543 pbb->transformed = isl_map_copy (pbb->schedule);
546 /* Build for BB the static schedule.
548 The static schedule is a Dewey numbering of the abstract syntax
549 tree: http://en.wikipedia.org/wiki/Dewey_Decimal_Classification
551 The following example informally defines the static schedule:
554 for (i: ...)
556 for (j: ...)
562 for (k: ...)
570 Static schedules for A to F:
572 DEPTH
573 0 1 2
575 B 1 0 0
576 C 1 0 1
577 D 1 1 0
578 E 1 1 1
582 static void
583 build_scop_scattering (scop_p scop)
585 int i;
586 poly_bb_p pbb;
587 gimple_bb_p previous_gbb = NULL;
588 isl_space *dc = isl_set_get_space (scop->context);
589 isl_aff *static_sched;
591 dc = isl_space_add_dims (dc, isl_dim_set, number_of_loops (cfun));
592 static_sched = isl_aff_zero_on_domain (isl_local_space_from_space (dc));
594 /* We have to start schedules at 0 on the first component and
595 because we cannot compare_prefix_loops against a previous loop,
596 prefix will be equal to zero, and that index will be
597 incremented before copying. */
598 static_sched = isl_aff_add_coefficient_si (static_sched, isl_dim_in, 0, -1);
600 FOR_EACH_VEC_ELT (SCOP_BBS (scop), i, pbb)
602 gimple_bb_p gbb = PBB_BLACK_BOX (pbb);
603 int prefix;
604 int nb_scat_dims = pbb_dim_iter_domain (pbb) * 2 + 1;
606 if (previous_gbb)
607 prefix = nb_common_loops (SCOP_REGION (scop), previous_gbb, gbb);
608 else
609 prefix = 0;
611 previous_gbb = gbb;
613 static_sched = isl_aff_add_coefficient_si (static_sched, isl_dim_in,
614 prefix, 1);
615 build_pbb_scattering_polyhedrons (static_sched, pbb, nb_scat_dims);
618 isl_aff_free (static_sched);
621 static isl_pw_aff *extract_affine (scop_p, tree, __isl_take isl_space *space);
623 /* Extract an affine expression from the chain of recurrence E. */
625 static isl_pw_aff *
626 extract_affine_chrec (scop_p s, tree e, __isl_take isl_space *space)
628 isl_pw_aff *lhs = extract_affine (s, CHREC_LEFT (e), isl_space_copy (space));
629 isl_pw_aff *rhs = extract_affine (s, CHREC_RIGHT (e), isl_space_copy (space));
630 isl_local_space *ls = isl_local_space_from_space (space);
631 unsigned pos = sese_loop_depth ((sese) s->region, get_chrec_loop (e)) - 1;
632 isl_aff *loop = isl_aff_set_coefficient_si
633 (isl_aff_zero_on_domain (ls), isl_dim_in, pos, 1);
634 isl_pw_aff *l = isl_pw_aff_from_aff (loop);
636 /* Before multiplying, make sure that the result is affine. */
637 gcc_assert (isl_pw_aff_is_cst (rhs)
638 || isl_pw_aff_is_cst (l));
640 return isl_pw_aff_add (lhs, isl_pw_aff_mul (rhs, l));
643 /* Extract an affine expression from the mult_expr E. */
645 static isl_pw_aff *
646 extract_affine_mul (scop_p s, tree e, __isl_take isl_space *space)
648 isl_pw_aff *lhs = extract_affine (s, TREE_OPERAND (e, 0),
649 isl_space_copy (space));
650 isl_pw_aff *rhs = extract_affine (s, TREE_OPERAND (e, 1), space);
652 if (!isl_pw_aff_is_cst (lhs)
653 && !isl_pw_aff_is_cst (rhs))
655 isl_pw_aff_free (lhs);
656 isl_pw_aff_free (rhs);
657 return NULL;
660 return isl_pw_aff_mul (lhs, rhs);
663 /* Return an ISL identifier from the name of the ssa_name E. */
665 static isl_id *
666 isl_id_for_ssa_name (scop_p s, tree e)
668 const char *name = get_name (e);
669 isl_id *id;
671 if (name)
672 id = isl_id_alloc (s->ctx, name, e);
673 else
675 char name1[50];
676 snprintf (name1, sizeof (name1), "P_%d", SSA_NAME_VERSION (e));
677 id = isl_id_alloc (s->ctx, name1, e);
680 return id;
683 /* Return an ISL identifier for the data reference DR. */
685 static isl_id *
686 isl_id_for_dr (scop_p s, data_reference_p dr ATTRIBUTE_UNUSED)
688 /* Data references all get the same isl_id. They need to be comparable
689 and are distinguished through the first dimension, which contains the
690 alias set number. */
691 return isl_id_alloc (s->ctx, "", 0);
694 /* Extract an affine expression from the ssa_name E. */
696 static isl_pw_aff *
697 extract_affine_name (scop_p s, tree e, __isl_take isl_space *space)
699 isl_aff *aff;
700 isl_set *dom;
701 isl_id *id;
702 int dimension;
704 id = isl_id_for_ssa_name (s, e);
705 dimension = isl_space_find_dim_by_id (space, isl_dim_param, id);
706 isl_id_free (id);
707 dom = isl_set_universe (isl_space_copy (space));
708 aff = isl_aff_zero_on_domain (isl_local_space_from_space (space));
709 aff = isl_aff_add_coefficient_si (aff, isl_dim_param, dimension, 1);
710 return isl_pw_aff_alloc (dom, aff);
713 /* Extract an affine expression from the gmp constant G. */
715 static isl_pw_aff *
716 extract_affine_gmp (mpz_t g, __isl_take isl_space *space)
718 isl_local_space *ls = isl_local_space_from_space (isl_space_copy (space));
719 isl_aff *aff = isl_aff_zero_on_domain (ls);
720 isl_set *dom = isl_set_universe (space);
721 isl_val *v;
722 isl_ctx *ct;
724 ct = isl_aff_get_ctx (aff);
725 v = isl_val_int_from_gmp (ct, g);
726 aff = isl_aff_add_constant_val (aff, v);
728 return isl_pw_aff_alloc (dom, aff);
731 /* Extract an affine expression from the integer_cst E. */
733 static isl_pw_aff *
734 extract_affine_int (tree e, __isl_take isl_space *space)
736 isl_pw_aff *res;
737 mpz_t g;
739 mpz_init (g);
740 tree_int_to_gmp (e, g);
741 res = extract_affine_gmp (g, space);
742 mpz_clear (g);
744 return res;
747 /* Compute pwaff mod 2^width. */
749 extern isl_ctx *the_isl_ctx;
751 static isl_pw_aff *
752 wrap (isl_pw_aff *pwaff, unsigned width)
754 isl_val *mod;
756 mod = isl_val_int_from_ui(the_isl_ctx, width);
757 mod = isl_val_2exp (mod);
758 pwaff = isl_pw_aff_mod_val (pwaff, mod);
760 return pwaff;
763 /* When parameter NAME is in REGION, returns its index in SESE_PARAMS.
764 Otherwise returns -1. */
766 static inline int
767 parameter_index_in_region_1 (tree name, sese region)
769 int i;
770 tree p;
772 gcc_assert (TREE_CODE (name) == SSA_NAME);
774 FOR_EACH_VEC_ELT (SESE_PARAMS (region), i, p)
775 if (p == name)
776 return i;
778 return -1;
781 /* When the parameter NAME is in REGION, returns its index in
782 SESE_PARAMS. Otherwise this function inserts NAME in SESE_PARAMS
783 and returns the index of NAME. */
785 static int
786 parameter_index_in_region (tree name, sese region)
788 int i;
790 gcc_assert (TREE_CODE (name) == SSA_NAME);
792 i = parameter_index_in_region_1 (name, region);
793 if (i != -1)
794 return i;
796 gcc_assert (SESE_ADD_PARAMS (region));
798 i = SESE_PARAMS (region).length ();
799 SESE_PARAMS (region).safe_push (name);
800 return i;
803 /* Extract an affine expression from the tree E in the scop S. */
805 static isl_pw_aff *
806 extract_affine (scop_p s, tree e, __isl_take isl_space *space)
808 isl_pw_aff *lhs, *rhs, *res;
809 tree type;
811 if (e == chrec_dont_know) {
812 isl_space_free (space);
813 return NULL;
816 switch (TREE_CODE (e))
818 case POLYNOMIAL_CHREC:
819 res = extract_affine_chrec (s, e, space);
820 break;
822 case MULT_EXPR:
823 res = extract_affine_mul (s, e, space);
824 break;
826 case PLUS_EXPR:
827 case POINTER_PLUS_EXPR:
828 lhs = extract_affine (s, TREE_OPERAND (e, 0), isl_space_copy (space));
829 rhs = extract_affine (s, TREE_OPERAND (e, 1), space);
830 res = isl_pw_aff_add (lhs, rhs);
831 break;
833 case MINUS_EXPR:
834 lhs = extract_affine (s, TREE_OPERAND (e, 0), isl_space_copy (space));
835 rhs = extract_affine (s, TREE_OPERAND (e, 1), space);
836 res = isl_pw_aff_sub (lhs, rhs);
837 break;
839 case NEGATE_EXPR:
840 case BIT_NOT_EXPR:
841 lhs = extract_affine (s, TREE_OPERAND (e, 0), isl_space_copy (space));
842 rhs = extract_affine (s, integer_minus_one_node, space);
843 res = isl_pw_aff_mul (lhs, rhs);
844 break;
846 case SSA_NAME:
847 gcc_assert (-1 != parameter_index_in_region_1 (e, SCOP_REGION (s)));
848 res = extract_affine_name (s, e, space);
849 break;
851 case INTEGER_CST:
852 res = extract_affine_int (e, space);
853 /* No need to wrap a single integer. */
854 return res;
856 CASE_CONVERT:
857 case NON_LVALUE_EXPR:
858 res = extract_affine (s, TREE_OPERAND (e, 0), space);
859 break;
861 default:
862 gcc_unreachable ();
863 break;
866 type = TREE_TYPE (e);
867 if (TYPE_UNSIGNED (type))
868 res = wrap (res, TYPE_PRECISION (type));
870 return res;
873 /* In the context of sese S, scan the expression E and translate it to
874 a linear expression C. When parsing a symbolic multiplication, K
875 represents the constant multiplier of an expression containing
876 parameters. */
878 static void
879 scan_tree_for_params (sese s, tree e)
881 if (e == chrec_dont_know)
882 return;
884 switch (TREE_CODE (e))
886 case POLYNOMIAL_CHREC:
887 scan_tree_for_params (s, CHREC_LEFT (e));
888 break;
890 case MULT_EXPR:
891 if (chrec_contains_symbols (TREE_OPERAND (e, 0)))
892 scan_tree_for_params (s, TREE_OPERAND (e, 0));
893 else
894 scan_tree_for_params (s, TREE_OPERAND (e, 1));
895 break;
897 case PLUS_EXPR:
898 case POINTER_PLUS_EXPR:
899 case MINUS_EXPR:
900 scan_tree_for_params (s, TREE_OPERAND (e, 0));
901 scan_tree_for_params (s, TREE_OPERAND (e, 1));
902 break;
904 case NEGATE_EXPR:
905 case BIT_NOT_EXPR:
906 CASE_CONVERT:
907 case NON_LVALUE_EXPR:
908 scan_tree_for_params (s, TREE_OPERAND (e, 0));
909 break;
911 case SSA_NAME:
912 parameter_index_in_region (e, s);
913 break;
915 case INTEGER_CST:
916 case ADDR_EXPR:
917 break;
919 default:
920 gcc_unreachable ();
921 break;
925 /* Find parameters with respect to REGION in BB. We are looking in memory
926 access functions, conditions and loop bounds. */
928 static void
929 find_params_in_bb (sese region, gimple_bb_p gbb)
931 int i;
932 unsigned j;
933 data_reference_p dr;
934 gimple stmt;
935 loop_p loop = GBB_BB (gbb)->loop_father;
937 /* Find parameters in the access functions of data references. */
938 FOR_EACH_VEC_ELT (GBB_DATA_REFS (gbb), i, dr)
939 for (j = 0; j < DR_NUM_DIMENSIONS (dr); j++)
940 scan_tree_for_params (region, DR_ACCESS_FN (dr, j));
942 /* Find parameters in conditional statements. */
943 FOR_EACH_VEC_ELT (GBB_CONDITIONS (gbb), i, stmt)
945 tree lhs = scalar_evolution_in_region (region, loop,
946 gimple_cond_lhs (stmt));
947 tree rhs = scalar_evolution_in_region (region, loop,
948 gimple_cond_rhs (stmt));
950 scan_tree_for_params (region, lhs);
951 scan_tree_for_params (region, rhs);
955 /* Record the parameters used in the SCOP. A variable is a parameter
956 in a scop if it does not vary during the execution of that scop. */
958 static void
959 find_scop_parameters (scop_p scop)
961 poly_bb_p pbb;
962 unsigned i;
963 sese region = SCOP_REGION (scop);
964 struct loop *loop;
965 int nbp;
967 /* Find the parameters used in the loop bounds. */
968 FOR_EACH_VEC_ELT (SESE_LOOP_NEST (region), i, loop)
970 tree nb_iters = number_of_latch_executions (loop);
972 if (!chrec_contains_symbols (nb_iters))
973 continue;
975 nb_iters = scalar_evolution_in_region (region, loop, nb_iters);
976 scan_tree_for_params (region, nb_iters);
979 /* Find the parameters used in data accesses. */
980 FOR_EACH_VEC_ELT (SCOP_BBS (scop), i, pbb)
981 find_params_in_bb (region, PBB_BLACK_BOX (pbb));
983 nbp = sese_nb_params (region);
984 scop_set_nb_params (scop, nbp);
985 SESE_ADD_PARAMS (region) = false;
988 tree e;
989 isl_space *space = isl_space_set_alloc (scop->ctx, nbp, 0);
991 FOR_EACH_VEC_ELT (SESE_PARAMS (region), i, e)
992 space = isl_space_set_dim_id (space, isl_dim_param, i,
993 isl_id_for_ssa_name (scop, e));
995 scop->context = isl_set_universe (space);
999 /* Builds the constraint polyhedra for LOOP in SCOP. OUTER_PH gives
1000 the constraints for the surrounding loops. */
1002 static void
1003 build_loop_iteration_domains (scop_p scop, struct loop *loop,
1004 int nb,
1005 isl_set *outer, isl_set **doms)
1007 tree nb_iters = number_of_latch_executions (loop);
1008 sese region = SCOP_REGION (scop);
1010 isl_set *inner = isl_set_copy (outer);
1011 isl_space *space;
1012 isl_constraint *c;
1013 int pos = isl_set_dim (outer, isl_dim_set);
1014 isl_val *v;
1015 mpz_t g;
1017 mpz_init (g);
1019 inner = isl_set_add_dims (inner, isl_dim_set, 1);
1020 space = isl_set_get_space (inner);
1022 /* 0 <= loop_i */
1023 c = isl_inequality_alloc
1024 (isl_local_space_from_space (isl_space_copy (space)));
1025 c = isl_constraint_set_coefficient_si (c, isl_dim_set, pos, 1);
1026 inner = isl_set_add_constraint (inner, c);
1028 /* loop_i <= cst_nb_iters */
1029 if (TREE_CODE (nb_iters) == INTEGER_CST)
1031 c = isl_inequality_alloc
1032 (isl_local_space_from_space (isl_space_copy (space)));
1033 c = isl_constraint_set_coefficient_si (c, isl_dim_set, pos, -1);
1034 tree_int_to_gmp (nb_iters, g);
1035 v = isl_val_int_from_gmp (the_isl_ctx, g);
1036 c = isl_constraint_set_constant_val (c, v);
1037 inner = isl_set_add_constraint (inner, c);
1040 /* loop_i <= expr_nb_iters */
1041 else if (!chrec_contains_undetermined (nb_iters))
1043 widest_int nit;
1044 isl_pw_aff *aff;
1045 isl_set *valid;
1046 isl_local_space *ls;
1047 isl_aff *al;
1048 isl_set *le;
1050 nb_iters = scalar_evolution_in_region (region, loop, nb_iters);
1052 aff = extract_affine (scop, nb_iters, isl_set_get_space (inner));
1053 valid = isl_pw_aff_nonneg_set (isl_pw_aff_copy (aff));
1054 valid = isl_set_project_out (valid, isl_dim_set, 0,
1055 isl_set_dim (valid, isl_dim_set));
1056 scop->context = isl_set_intersect (scop->context, valid);
1058 ls = isl_local_space_from_space (isl_space_copy (space));
1059 al = isl_aff_set_coefficient_si (isl_aff_zero_on_domain (ls),
1060 isl_dim_in, pos, 1);
1061 le = isl_pw_aff_le_set (isl_pw_aff_from_aff (al),
1062 isl_pw_aff_copy (aff));
1063 inner = isl_set_intersect (inner, le);
1065 if (max_stmt_executions (loop, &nit))
1067 /* Insert in the context the constraints from the
1068 estimation of the number of iterations NIT and the
1069 symbolic number of iterations (involving parameter
1070 names) NB_ITERS. First, build the affine expression
1071 "NIT - NB_ITERS" and then say that it is positive,
1072 i.e., NIT approximates NB_ITERS: "NIT >= NB_ITERS". */
1073 isl_pw_aff *approx;
1074 mpz_t g;
1075 isl_set *x;
1076 isl_constraint *c;
1078 mpz_init (g);
1079 wi::to_mpz (nit, g, SIGNED);
1080 mpz_sub_ui (g, g, 1);
1081 approx = extract_affine_gmp (g, isl_set_get_space (inner));
1082 x = isl_pw_aff_ge_set (approx, aff);
1083 x = isl_set_project_out (x, isl_dim_set, 0,
1084 isl_set_dim (x, isl_dim_set));
1085 scop->context = isl_set_intersect (scop->context, x);
1087 c = isl_inequality_alloc
1088 (isl_local_space_from_space (isl_space_copy (space)));
1089 c = isl_constraint_set_coefficient_si (c, isl_dim_set, pos, -1);
1090 v = isl_val_int_from_gmp (the_isl_ctx, g);
1091 mpz_clear (g);
1092 c = isl_constraint_set_constant_val (c, v);
1093 inner = isl_set_add_constraint (inner, c);
1095 else
1096 isl_pw_aff_free (aff);
1098 else
1099 gcc_unreachable ();
1101 if (loop->inner && loop_in_sese_p (loop->inner, region))
1102 build_loop_iteration_domains (scop, loop->inner, nb + 1,
1103 isl_set_copy (inner), doms);
1105 if (nb != 0
1106 && loop->next
1107 && loop_in_sese_p (loop->next, region))
1108 build_loop_iteration_domains (scop, loop->next, nb,
1109 isl_set_copy (outer), doms);
1111 doms[loop->num] = inner;
1113 isl_set_free (outer);
1114 isl_space_free (space);
1115 mpz_clear (g);
1118 /* Returns a linear expression for tree T evaluated in PBB. */
1120 static isl_pw_aff *
1121 create_pw_aff_from_tree (poly_bb_p pbb, tree t)
1123 scop_p scop = PBB_SCOP (pbb);
1125 t = scalar_evolution_in_region (SCOP_REGION (scop), pbb_loop (pbb), t);
1126 gcc_assert (!automatically_generated_chrec_p (t));
1128 return extract_affine (scop, t, isl_set_get_space (pbb->domain));
1131 /* Add conditional statement STMT to pbb. CODE is used as the comparison
1132 operator. This allows us to invert the condition or to handle
1133 inequalities. */
1135 static void
1136 add_condition_to_pbb (poly_bb_p pbb, gcond *stmt, enum tree_code code)
1138 isl_pw_aff *lhs = create_pw_aff_from_tree (pbb, gimple_cond_lhs (stmt));
1139 isl_pw_aff *rhs = create_pw_aff_from_tree (pbb, gimple_cond_rhs (stmt));
1140 isl_set *cond;
1142 switch (code)
1144 case LT_EXPR:
1145 cond = isl_pw_aff_lt_set (lhs, rhs);
1146 break;
1148 case GT_EXPR:
1149 cond = isl_pw_aff_gt_set (lhs, rhs);
1150 break;
1152 case LE_EXPR:
1153 cond = isl_pw_aff_le_set (lhs, rhs);
1154 break;
1156 case GE_EXPR:
1157 cond = isl_pw_aff_ge_set (lhs, rhs);
1158 break;
1160 case EQ_EXPR:
1161 cond = isl_pw_aff_eq_set (lhs, rhs);
1162 break;
1164 case NE_EXPR:
1165 cond = isl_pw_aff_ne_set (lhs, rhs);
1166 break;
1168 default:
1169 isl_pw_aff_free (lhs);
1170 isl_pw_aff_free (rhs);
1171 return;
1174 cond = isl_set_coalesce (cond);
1175 cond = isl_set_set_tuple_id (cond, isl_set_get_tuple_id (pbb->domain));
1176 pbb->domain = isl_set_intersect (pbb->domain, cond);
1179 /* Add conditions to the domain of PBB. */
1181 static void
1182 add_conditions_to_domain (poly_bb_p pbb)
1184 unsigned int i;
1185 gimple stmt;
1186 gimple_bb_p gbb = PBB_BLACK_BOX (pbb);
1188 if (GBB_CONDITIONS (gbb).is_empty ())
1189 return;
1191 FOR_EACH_VEC_ELT (GBB_CONDITIONS (gbb), i, stmt)
1192 switch (gimple_code (stmt))
1194 case GIMPLE_COND:
1196 gcond *cond_stmt = as_a <gcond *> (stmt);
1197 enum tree_code code = gimple_cond_code (cond_stmt);
1199 /* The conditions for ELSE-branches are inverted. */
1200 if (!GBB_CONDITION_CASES (gbb)[i])
1201 code = invert_tree_comparison (code, false);
1203 add_condition_to_pbb (pbb, cond_stmt, code);
1204 break;
1207 case GIMPLE_SWITCH:
1208 /* Switch statements are not supported right now - fall through. */
1210 default:
1211 gcc_unreachable ();
1212 break;
1216 /* Traverses all the GBBs of the SCOP and add their constraints to the
1217 iteration domains. */
1219 static void
1220 add_conditions_to_constraints (scop_p scop)
1222 int i;
1223 poly_bb_p pbb;
1225 FOR_EACH_VEC_ELT (SCOP_BBS (scop), i, pbb)
1226 add_conditions_to_domain (pbb);
1229 /* Returns a COND_EXPR statement when BB has a single predecessor, the
1230 edge between BB and its predecessor is not a loop exit edge, and
1231 the last statement of the single predecessor is a COND_EXPR. */
1233 static gcond *
1234 single_pred_cond_non_loop_exit (basic_block bb)
1236 if (single_pred_p (bb))
1238 edge e = single_pred_edge (bb);
1239 basic_block pred = e->src;
1240 gimple stmt;
1242 if (loop_depth (pred->loop_father) > loop_depth (bb->loop_father))
1243 return NULL;
1245 stmt = last_stmt (pred);
1247 if (stmt && gimple_code (stmt) == GIMPLE_COND)
1248 return as_a <gcond *> (stmt);
1251 return NULL;
1254 class sese_dom_walker : public dom_walker
1256 public:
1257 sese_dom_walker (cdi_direction, sese);
1259 virtual void before_dom_children (basic_block);
1260 virtual void after_dom_children (basic_block);
1262 private:
1263 auto_vec<gimple, 3> m_conditions, m_cases;
1264 sese m_region;
1267 sese_dom_walker::sese_dom_walker (cdi_direction direction, sese region)
1268 : dom_walker (direction), m_region (region)
1272 /* Call-back for dom_walk executed before visiting the dominated
1273 blocks. */
1275 void
1276 sese_dom_walker::before_dom_children (basic_block bb)
1278 gimple_bb_p gbb;
1279 gcond *stmt;
1281 if (!bb_in_sese_p (bb, m_region))
1282 return;
1284 stmt = single_pred_cond_non_loop_exit (bb);
1286 if (stmt)
1288 edge e = single_pred_edge (bb);
1290 m_conditions.safe_push (stmt);
1292 if (e->flags & EDGE_TRUE_VALUE)
1293 m_cases.safe_push (stmt);
1294 else
1295 m_cases.safe_push (NULL);
1298 gbb = gbb_from_bb (bb);
1300 if (gbb)
1302 GBB_CONDITIONS (gbb) = m_conditions.copy ();
1303 GBB_CONDITION_CASES (gbb) = m_cases.copy ();
1307 /* Call-back for dom_walk executed after visiting the dominated
1308 blocks. */
1310 void
1311 sese_dom_walker::after_dom_children (basic_block bb)
1313 if (!bb_in_sese_p (bb, m_region))
1314 return;
1316 if (single_pred_cond_non_loop_exit (bb))
1318 m_conditions.pop ();
1319 m_cases.pop ();
1323 /* Add constraints on the possible values of parameter P from the type
1324 of P. */
1326 static void
1327 add_param_constraints (scop_p scop, graphite_dim_t p)
1329 tree parameter = SESE_PARAMS (SCOP_REGION (scop))[p];
1330 tree type = TREE_TYPE (parameter);
1331 tree lb = NULL_TREE;
1332 tree ub = NULL_TREE;
1334 if (POINTER_TYPE_P (type) || !TYPE_MIN_VALUE (type))
1335 lb = lower_bound_in_type (type, type);
1336 else
1337 lb = TYPE_MIN_VALUE (type);
1339 if (POINTER_TYPE_P (type) || !TYPE_MAX_VALUE (type))
1340 ub = upper_bound_in_type (type, type);
1341 else
1342 ub = TYPE_MAX_VALUE (type);
1344 if (lb)
1346 isl_space *space = isl_set_get_space (scop->context);
1347 isl_constraint *c;
1348 mpz_t g;
1349 isl_val *v;
1351 c = isl_inequality_alloc (isl_local_space_from_space (space));
1352 mpz_init (g);
1353 tree_int_to_gmp (lb, g);
1354 v = isl_val_int_from_gmp (the_isl_ctx, g);
1355 v = isl_val_neg (v);
1356 mpz_clear (g);
1357 c = isl_constraint_set_constant_val (c, v);
1358 c = isl_constraint_set_coefficient_si (c, isl_dim_param, p, 1);
1360 scop->context = isl_set_add_constraint (scop->context, c);
1363 if (ub)
1365 isl_space *space = isl_set_get_space (scop->context);
1366 isl_constraint *c;
1367 mpz_t g;
1368 isl_val *v;
1370 c = isl_inequality_alloc (isl_local_space_from_space (space));
1372 mpz_init (g);
1373 tree_int_to_gmp (ub, g);
1374 v = isl_val_int_from_gmp (the_isl_ctx, g);
1375 mpz_clear (g);
1376 c = isl_constraint_set_constant_val (c, v);
1377 c = isl_constraint_set_coefficient_si (c, isl_dim_param, p, -1);
1379 scop->context = isl_set_add_constraint (scop->context, c);
1383 /* Build the context of the SCOP. The context usually contains extra
1384 constraints that are added to the iteration domains that constrain
1385 some parameters. */
1387 static void
1388 build_scop_context (scop_p scop)
1390 graphite_dim_t p, n = scop_nb_params (scop);
1392 for (p = 0; p < n; p++)
1393 add_param_constraints (scop, p);
1396 /* Build the iteration domains: the loops belonging to the current
1397 SCOP, and that vary for the execution of the current basic block.
1398 Returns false if there is no loop in SCOP. */
1400 static void
1401 build_scop_iteration_domain (scop_p scop)
1403 struct loop *loop;
1404 sese region = SCOP_REGION (scop);
1405 int i;
1406 poly_bb_p pbb;
1407 int nb_loops = number_of_loops (cfun);
1408 isl_set **doms = XCNEWVEC (isl_set *, nb_loops);
1410 FOR_EACH_VEC_ELT (SESE_LOOP_NEST (region), i, loop)
1411 if (!loop_in_sese_p (loop_outer (loop), region))
1412 build_loop_iteration_domains (scop, loop, 0,
1413 isl_set_copy (scop->context), doms);
1415 FOR_EACH_VEC_ELT (SCOP_BBS (scop), i, pbb)
1417 loop = pbb_loop (pbb);
1419 if (doms[loop->num])
1420 pbb->domain = isl_set_copy (doms[loop->num]);
1421 else
1422 pbb->domain = isl_set_copy (scop->context);
1424 pbb->domain = isl_set_set_tuple_id (pbb->domain,
1425 isl_id_for_pbb (scop, pbb));
1428 for (i = 0; i < nb_loops; i++)
1429 if (doms[i])
1430 isl_set_free (doms[i]);
1432 free (doms);
1435 /* Add a constrain to the ACCESSES polyhedron for the alias set of
1436 data reference DR. ACCESSP_NB_DIMS is the dimension of the
1437 ACCESSES polyhedron, DOM_NB_DIMS is the dimension of the iteration
1438 domain. */
1440 static isl_map *
1441 pdr_add_alias_set (isl_map *acc, data_reference_p dr)
1443 isl_constraint *c;
1444 int alias_set_num = 0;
1445 base_alias_pair *bap = (base_alias_pair *)(dr->aux);
1447 if (bap && bap->alias_set)
1448 alias_set_num = *(bap->alias_set);
1450 c = isl_equality_alloc
1451 (isl_local_space_from_space (isl_map_get_space (acc)));
1452 c = isl_constraint_set_constant_si (c, -alias_set_num);
1453 c = isl_constraint_set_coefficient_si (c, isl_dim_out, 0, 1);
1455 return isl_map_add_constraint (acc, c);
1458 /* Assign the affine expression INDEX to the output dimension POS of
1459 MAP and return the result. */
1461 static isl_map *
1462 set_index (isl_map *map, int pos, isl_pw_aff *index)
1464 isl_map *index_map;
1465 int len = isl_map_dim (map, isl_dim_out);
1466 isl_id *id;
1468 index_map = isl_map_from_pw_aff (index);
1469 index_map = isl_map_insert_dims (index_map, isl_dim_out, 0, pos);
1470 index_map = isl_map_add_dims (index_map, isl_dim_out, len - pos - 1);
1472 id = isl_map_get_tuple_id (map, isl_dim_out);
1473 index_map = isl_map_set_tuple_id (index_map, isl_dim_out, id);
1474 id = isl_map_get_tuple_id (map, isl_dim_in);
1475 index_map = isl_map_set_tuple_id (index_map, isl_dim_in, id);
1477 return isl_map_intersect (map, index_map);
1480 /* Add to ACCESSES polyhedron equalities defining the access functions
1481 to the memory. ACCESSP_NB_DIMS is the dimension of the ACCESSES
1482 polyhedron, DOM_NB_DIMS is the dimension of the iteration domain.
1483 PBB is the poly_bb_p that contains the data reference DR. */
1485 static isl_map *
1486 pdr_add_memory_accesses (isl_map *acc, data_reference_p dr, poly_bb_p pbb)
1488 int i, nb_subscripts = DR_NUM_DIMENSIONS (dr);
1489 scop_p scop = PBB_SCOP (pbb);
1491 for (i = 0; i < nb_subscripts; i++)
1493 isl_pw_aff *aff;
1494 tree afn = DR_ACCESS_FN (dr, nb_subscripts - 1 - i);
1496 aff = extract_affine (scop, afn,
1497 isl_space_domain (isl_map_get_space (acc)));
1498 acc = set_index (acc, i + 1, aff);
1501 return acc;
1504 /* Add constrains representing the size of the accessed data to the
1505 ACCESSES polyhedron. ACCESSP_NB_DIMS is the dimension of the
1506 ACCESSES polyhedron, DOM_NB_DIMS is the dimension of the iteration
1507 domain. */
1509 static isl_set *
1510 pdr_add_data_dimensions (isl_set *extent, scop_p scop, data_reference_p dr)
1512 tree ref = DR_REF (dr);
1513 int i, nb_subscripts = DR_NUM_DIMENSIONS (dr);
1515 for (i = nb_subscripts - 1; i >= 0; i--, ref = TREE_OPERAND (ref, 0))
1517 tree low, high;
1519 if (TREE_CODE (ref) != ARRAY_REF)
1520 break;
1522 low = array_ref_low_bound (ref);
1523 high = array_ref_up_bound (ref);
1525 /* XXX The PPL code dealt separately with
1526 subscript - low >= 0 and high - subscript >= 0 in case one of
1527 the two bounds isn't known. Do the same here? */
1529 if (tree_fits_shwi_p (low)
1530 && high
1531 && tree_fits_shwi_p (high)
1532 /* 1-element arrays at end of structures may extend over
1533 their declared size. */
1534 && !(array_at_struct_end_p (ref)
1535 && operand_equal_p (low, high, 0)))
1537 isl_id *id;
1538 isl_aff *aff;
1539 isl_set *univ, *lbs, *ubs;
1540 isl_pw_aff *index;
1541 isl_space *space;
1542 isl_set *valid;
1543 isl_pw_aff *lb = extract_affine_int (low, isl_set_get_space (extent));
1544 isl_pw_aff *ub = extract_affine_int (high, isl_set_get_space (extent));
1546 /* high >= 0 */
1547 valid = isl_pw_aff_nonneg_set (isl_pw_aff_copy (ub));
1548 valid = isl_set_project_out (valid, isl_dim_set, 0,
1549 isl_set_dim (valid, isl_dim_set));
1550 scop->context = isl_set_intersect (scop->context, valid);
1552 space = isl_set_get_space (extent);
1553 aff = isl_aff_zero_on_domain (isl_local_space_from_space (space));
1554 aff = isl_aff_add_coefficient_si (aff, isl_dim_in, i + 1, 1);
1555 univ = isl_set_universe (isl_space_domain (isl_aff_get_space (aff)));
1556 index = isl_pw_aff_alloc (univ, aff);
1558 id = isl_set_get_tuple_id (extent);
1559 lb = isl_pw_aff_set_tuple_id (lb, isl_dim_in, isl_id_copy (id));
1560 ub = isl_pw_aff_set_tuple_id (ub, isl_dim_in, id);
1562 /* low <= sub_i <= high */
1563 lbs = isl_pw_aff_ge_set (isl_pw_aff_copy (index), lb);
1564 ubs = isl_pw_aff_le_set (index, ub);
1565 extent = isl_set_intersect (extent, lbs);
1566 extent = isl_set_intersect (extent, ubs);
1570 return extent;
1573 /* Build data accesses for DR in PBB. */
1575 static void
1576 build_poly_dr (data_reference_p dr, poly_bb_p pbb)
1578 int dr_base_object_set;
1579 isl_map *acc;
1580 isl_set *extent;
1581 scop_p scop = PBB_SCOP (pbb);
1584 isl_space *dc = isl_set_get_space (pbb->domain);
1585 int nb_out = 1 + DR_NUM_DIMENSIONS (dr);
1586 isl_space *space = isl_space_add_dims (isl_space_from_domain (dc),
1587 isl_dim_out, nb_out);
1589 acc = isl_map_universe (space);
1590 acc = isl_map_set_tuple_id (acc, isl_dim_out, isl_id_for_dr (scop, dr));
1593 acc = pdr_add_alias_set (acc, dr);
1594 acc = pdr_add_memory_accesses (acc, dr, pbb);
1597 isl_id *id = isl_id_for_dr (scop, dr);
1598 int nb = 1 + DR_NUM_DIMENSIONS (dr);
1599 isl_space *space = isl_space_set_alloc (scop->ctx, 0, nb);
1600 int alias_set_num = 0;
1601 base_alias_pair *bap = (base_alias_pair *)(dr->aux);
1603 if (bap && bap->alias_set)
1604 alias_set_num = *(bap->alias_set);
1606 space = isl_space_set_tuple_id (space, isl_dim_set, id);
1607 extent = isl_set_nat_universe (space);
1608 extent = isl_set_fix_si (extent, isl_dim_set, 0, alias_set_num);
1609 extent = pdr_add_data_dimensions (extent, scop, dr);
1612 gcc_assert (dr->aux);
1613 dr_base_object_set = ((base_alias_pair *)(dr->aux))->base_obj_set;
1615 new_poly_dr (pbb, dr_base_object_set,
1616 DR_IS_READ (dr) ? PDR_READ : PDR_WRITE,
1617 dr, DR_NUM_DIMENSIONS (dr), acc, extent);
1620 /* Write to FILE the alias graph of data references in DIMACS format. */
1622 static inline bool
1623 write_alias_graph_to_ascii_dimacs (FILE *file, char *comment,
1624 vec<data_reference_p> drs)
1626 int num_vertex = drs.length ();
1627 int edge_num = 0;
1628 data_reference_p dr1, dr2;
1629 int i, j;
1631 if (num_vertex == 0)
1632 return true;
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 edge_num++;
1639 fprintf (file, "$\n");
1641 if (comment)
1642 fprintf (file, "c %s\n", comment);
1644 fprintf (file, "p edge %d %d\n", num_vertex, edge_num);
1646 FOR_EACH_VEC_ELT (drs, i, dr1)
1647 for (j = i + 1; drs.iterate (j, &dr2); j++)
1648 if (dr_may_alias_p (dr1, dr2, true))
1649 fprintf (file, "e %d %d\n", i + 1, j + 1);
1651 return true;
1654 /* Write to FILE the alias graph of data references in DOT format. */
1656 static inline bool
1657 write_alias_graph_to_ascii_dot (FILE *file, char *comment,
1658 vec<data_reference_p> drs)
1660 int num_vertex = drs.length ();
1661 data_reference_p dr1, dr2;
1662 int i, j;
1664 if (num_vertex == 0)
1665 return true;
1667 fprintf (file, "$\n");
1669 if (comment)
1670 fprintf (file, "c %s\n", comment);
1672 /* First print all the vertices. */
1673 FOR_EACH_VEC_ELT (drs, i, dr1)
1674 fprintf (file, "n%d;\n", i);
1676 FOR_EACH_VEC_ELT (drs, i, dr1)
1677 for (j = i + 1; drs.iterate (j, &dr2); j++)
1678 if (dr_may_alias_p (dr1, dr2, true))
1679 fprintf (file, "n%d n%d\n", i, j);
1681 return true;
1684 /* Write to FILE the alias graph of data references in ECC format. */
1686 static inline bool
1687 write_alias_graph_to_ascii_ecc (FILE *file, char *comment,
1688 vec<data_reference_p> drs)
1690 int num_vertex = drs.length ();
1691 data_reference_p dr1, dr2;
1692 int i, j;
1694 if (num_vertex == 0)
1695 return true;
1697 fprintf (file, "$\n");
1699 if (comment)
1700 fprintf (file, "c %s\n", comment);
1702 FOR_EACH_VEC_ELT (drs, i, dr1)
1703 for (j = i + 1; drs.iterate (j, &dr2); j++)
1704 if (dr_may_alias_p (dr1, dr2, true))
1705 fprintf (file, "%d %d\n", i, j);
1707 return true;
1710 /* Check if DR1 and DR2 are in the same object set. */
1712 static bool
1713 dr_same_base_object_p (const struct data_reference *dr1,
1714 const struct data_reference *dr2)
1716 return operand_equal_p (DR_BASE_OBJECT (dr1), DR_BASE_OBJECT (dr2), 0);
1719 /* Uses DFS component number as representative of alias-sets. Also tests for
1720 optimality by verifying if every connected component is a clique. Returns
1721 true (1) if the above test is true, and false (0) otherwise. */
1723 static int
1724 build_alias_set_optimal_p (vec<data_reference_p> drs)
1726 int num_vertices = drs.length ();
1727 struct graph *g = new_graph (num_vertices);
1728 data_reference_p dr1, dr2;
1729 int i, j;
1730 int num_connected_components;
1731 int v_indx1, v_indx2, num_vertices_in_component;
1732 int *all_vertices;
1733 int *vertices;
1734 struct graph_edge *e;
1735 int this_component_is_clique;
1736 int all_components_are_cliques = 1;
1738 FOR_EACH_VEC_ELT (drs, i, dr1)
1739 for (j = i+1; drs.iterate (j, &dr2); j++)
1740 if (dr_may_alias_p (dr1, dr2, true))
1742 add_edge (g, i, j);
1743 add_edge (g, j, i);
1746 all_vertices = XNEWVEC (int, num_vertices);
1747 vertices = XNEWVEC (int, num_vertices);
1748 for (i = 0; i < num_vertices; i++)
1749 all_vertices[i] = i;
1751 num_connected_components = graphds_dfs (g, all_vertices, num_vertices,
1752 NULL, true, NULL);
1753 for (i = 0; i < g->n_vertices; i++)
1755 data_reference_p dr = drs[i];
1756 base_alias_pair *bap;
1758 gcc_assert (dr->aux);
1759 bap = (base_alias_pair *)(dr->aux);
1761 bap->alias_set = XNEW (int);
1762 *(bap->alias_set) = g->vertices[i].component + 1;
1765 /* Verify if the DFS numbering results in optimal solution. */
1766 for (i = 0; i < num_connected_components; i++)
1768 num_vertices_in_component = 0;
1769 /* Get all vertices whose DFS component number is the same as i. */
1770 for (j = 0; j < num_vertices; j++)
1771 if (g->vertices[j].component == i)
1772 vertices[num_vertices_in_component++] = j;
1774 /* Now test if the vertices in 'vertices' form a clique, by testing
1775 for edges among each pair. */
1776 this_component_is_clique = 1;
1777 for (v_indx1 = 0; v_indx1 < num_vertices_in_component; v_indx1++)
1779 for (v_indx2 = v_indx1+1; v_indx2 < num_vertices_in_component; v_indx2++)
1781 /* Check if the two vertices are connected by iterating
1782 through all the edges which have one of these are source. */
1783 e = g->vertices[vertices[v_indx2]].pred;
1784 while (e)
1786 if (e->src == vertices[v_indx1])
1787 break;
1788 e = e->pred_next;
1790 if (!e)
1792 this_component_is_clique = 0;
1793 break;
1796 if (!this_component_is_clique)
1797 all_components_are_cliques = 0;
1801 free (all_vertices);
1802 free (vertices);
1803 free_graph (g);
1804 return all_components_are_cliques;
1807 /* Group each data reference in DRS with its base object set num. */
1809 static void
1810 build_base_obj_set_for_drs (vec<data_reference_p> drs)
1812 int num_vertex = drs.length ();
1813 struct graph *g = new_graph (num_vertex);
1814 data_reference_p dr1, dr2;
1815 int i, j;
1816 int *queue;
1818 FOR_EACH_VEC_ELT (drs, i, dr1)
1819 for (j = i + 1; drs.iterate (j, &dr2); j++)
1820 if (dr_same_base_object_p (dr1, dr2))
1822 add_edge (g, i, j);
1823 add_edge (g, j, i);
1826 queue = XNEWVEC (int, num_vertex);
1827 for (i = 0; i < num_vertex; i++)
1828 queue[i] = i;
1830 graphds_dfs (g, queue, num_vertex, NULL, true, NULL);
1832 for (i = 0; i < g->n_vertices; i++)
1834 data_reference_p dr = drs[i];
1835 base_alias_pair *bap;
1837 gcc_assert (dr->aux);
1838 bap = (base_alias_pair *)(dr->aux);
1840 bap->base_obj_set = g->vertices[i].component + 1;
1843 free (queue);
1844 free_graph (g);
1847 /* Build the data references for PBB. */
1849 static void
1850 build_pbb_drs (poly_bb_p pbb)
1852 int j;
1853 data_reference_p dr;
1854 vec<data_reference_p> gbb_drs = GBB_DATA_REFS (PBB_BLACK_BOX (pbb));
1856 FOR_EACH_VEC_ELT (gbb_drs, j, dr)
1857 build_poly_dr (dr, pbb);
1860 /* Dump to file the alias graphs for the data references in DRS. */
1862 static void
1863 dump_alias_graphs (vec<data_reference_p> drs)
1865 char comment[100];
1866 FILE *file_dimacs, *file_ecc, *file_dot;
1868 file_dimacs = fopen ("/tmp/dr_alias_graph_dimacs", "ab");
1869 if (file_dimacs)
1871 snprintf (comment, sizeof (comment), "%s %s", main_input_filename,
1872 current_function_name ());
1873 write_alias_graph_to_ascii_dimacs (file_dimacs, comment, drs);
1874 fclose (file_dimacs);
1877 file_ecc = fopen ("/tmp/dr_alias_graph_ecc", "ab");
1878 if (file_ecc)
1880 snprintf (comment, sizeof (comment), "%s %s", main_input_filename,
1881 current_function_name ());
1882 write_alias_graph_to_ascii_ecc (file_ecc, comment, drs);
1883 fclose (file_ecc);
1886 file_dot = fopen ("/tmp/dr_alias_graph_dot", "ab");
1887 if (file_dot)
1889 snprintf (comment, sizeof (comment), "%s %s", main_input_filename,
1890 current_function_name ());
1891 write_alias_graph_to_ascii_dot (file_dot, comment, drs);
1892 fclose (file_dot);
1896 /* Build data references in SCOP. */
1898 static void
1899 build_scop_drs (scop_p scop)
1901 int i, j;
1902 poly_bb_p pbb;
1903 data_reference_p dr;
1904 auto_vec<data_reference_p, 3> drs;
1906 /* Remove all the PBBs that do not have data references: these basic
1907 blocks are not handled in the polyhedral representation. */
1908 for (i = 0; SCOP_BBS (scop).iterate (i, &pbb); i++)
1909 if (GBB_DATA_REFS (PBB_BLACK_BOX (pbb)).is_empty ())
1911 free_gimple_bb (PBB_BLACK_BOX (pbb));
1912 free_poly_bb (pbb);
1913 SCOP_BBS (scop).ordered_remove (i);
1914 i--;
1917 FOR_EACH_VEC_ELT (SCOP_BBS (scop), i, pbb)
1918 for (j = 0; GBB_DATA_REFS (PBB_BLACK_BOX (pbb)).iterate (j, &dr); j++)
1919 drs.safe_push (dr);
1921 FOR_EACH_VEC_ELT (drs, i, dr)
1922 dr->aux = XNEW (base_alias_pair);
1924 if (!build_alias_set_optimal_p (drs))
1926 /* TODO: Add support when building alias set is not optimal. */
1930 build_base_obj_set_for_drs (drs);
1932 /* When debugging, enable the following code. This cannot be used
1933 in production compilers. */
1934 if (0)
1935 dump_alias_graphs (drs);
1937 drs.release ();
1939 FOR_EACH_VEC_ELT (SCOP_BBS (scop), i, pbb)
1940 build_pbb_drs (pbb);
1943 /* Return a gsi at the position of the phi node STMT. */
1945 static gphi_iterator
1946 gsi_for_phi_node (gphi *stmt)
1948 gphi_iterator psi;
1949 basic_block bb = gimple_bb (stmt);
1951 for (psi = gsi_start_phis (bb); !gsi_end_p (psi); gsi_next (&psi))
1952 if (stmt == psi.phi ())
1953 return psi;
1955 gcc_unreachable ();
1956 return psi;
1959 /* Analyze all the data references of STMTS and add them to the
1960 GBB_DATA_REFS vector of BB. */
1962 static void
1963 analyze_drs_in_stmts (scop_p scop, basic_block bb, vec<gimple> stmts)
1965 loop_p nest;
1966 gimple_bb_p gbb;
1967 gimple stmt;
1968 int i;
1969 sese region = SCOP_REGION (scop);
1971 if (!bb_in_sese_p (bb, region))
1972 return;
1974 nest = outermost_loop_in_sese_1 (region, bb);
1975 gbb = gbb_from_bb (bb);
1977 FOR_EACH_VEC_ELT (stmts, i, stmt)
1979 loop_p loop;
1981 if (is_gimple_debug (stmt))
1982 continue;
1984 loop = loop_containing_stmt (stmt);
1985 if (!loop_in_sese_p (loop, region))
1986 loop = nest;
1988 graphite_find_data_references_in_stmt (nest, loop, stmt,
1989 &GBB_DATA_REFS (gbb));
1993 /* Insert STMT at the end of the STMTS sequence and then insert the
1994 statements from STMTS at INSERT_GSI and call analyze_drs_in_stmts
1995 on STMTS. */
1997 static void
1998 insert_stmts (scop_p scop, gimple stmt, gimple_seq stmts,
1999 gimple_stmt_iterator insert_gsi)
2001 gimple_stmt_iterator gsi;
2002 auto_vec<gimple, 3> x;
2004 gimple_seq_add_stmt (&stmts, stmt);
2005 for (gsi = gsi_start (stmts); !gsi_end_p (gsi); gsi_next (&gsi))
2006 x.safe_push (gsi_stmt (gsi));
2008 gsi_insert_seq_before (&insert_gsi, stmts, GSI_SAME_STMT);
2009 analyze_drs_in_stmts (scop, gsi_bb (insert_gsi), x);
2012 /* Insert the assignment "RES := EXPR" just after AFTER_STMT. */
2014 static void
2015 insert_out_of_ssa_copy (scop_p scop, tree res, tree expr, gimple after_stmt)
2017 gimple_seq stmts;
2018 gimple_stmt_iterator gsi;
2019 tree var = force_gimple_operand (expr, &stmts, true, NULL_TREE);
2020 gassign *stmt = gimple_build_assign (unshare_expr (res), var);
2021 auto_vec<gimple, 3> x;
2023 gimple_seq_add_stmt (&stmts, stmt);
2024 for (gsi = gsi_start (stmts); !gsi_end_p (gsi); gsi_next (&gsi))
2025 x.safe_push (gsi_stmt (gsi));
2027 if (gimple_code (after_stmt) == GIMPLE_PHI)
2029 gsi = gsi_after_labels (gimple_bb (after_stmt));
2030 gsi_insert_seq_before (&gsi, stmts, GSI_NEW_STMT);
2032 else
2034 gsi = gsi_for_stmt (after_stmt);
2035 gsi_insert_seq_after (&gsi, stmts, GSI_NEW_STMT);
2038 analyze_drs_in_stmts (scop, gimple_bb (after_stmt), x);
2041 /* Creates a poly_bb_p for basic_block BB from the existing PBB. */
2043 static void
2044 new_pbb_from_pbb (scop_p scop, poly_bb_p pbb, basic_block bb)
2046 vec<data_reference_p> drs;
2047 drs.create (3);
2048 gimple_bb_p gbb = PBB_BLACK_BOX (pbb);
2049 gimple_bb_p gbb1 = new_gimple_bb (bb, drs);
2050 poly_bb_p pbb1 = new_poly_bb (scop, gbb1);
2051 int index, n = SCOP_BBS (scop).length ();
2053 /* The INDEX of PBB in SCOP_BBS. */
2054 for (index = 0; index < n; index++)
2055 if (SCOP_BBS (scop)[index] == pbb)
2056 break;
2058 pbb1->domain = isl_set_copy (pbb->domain);
2059 pbb1->domain = isl_set_set_tuple_id (pbb1->domain,
2060 isl_id_for_pbb (scop, pbb1));
2062 GBB_PBB (gbb1) = pbb1;
2063 GBB_CONDITIONS (gbb1) = GBB_CONDITIONS (gbb).copy ();
2064 GBB_CONDITION_CASES (gbb1) = GBB_CONDITION_CASES (gbb).copy ();
2065 SCOP_BBS (scop).safe_insert (index + 1, pbb1);
2068 /* Insert on edge E the assignment "RES := EXPR". */
2070 static void
2071 insert_out_of_ssa_copy_on_edge (scop_p scop, edge e, tree res, tree expr)
2073 gimple_stmt_iterator gsi;
2074 gimple_seq stmts = NULL;
2075 tree var = force_gimple_operand (expr, &stmts, true, NULL_TREE);
2076 gimple stmt = gimple_build_assign (unshare_expr (res), var);
2077 basic_block bb;
2078 auto_vec<gimple, 3> x;
2080 gimple_seq_add_stmt (&stmts, stmt);
2081 for (gsi = gsi_start (stmts); !gsi_end_p (gsi); gsi_next (&gsi))
2082 x.safe_push (gsi_stmt (gsi));
2084 gsi_insert_seq_on_edge (e, stmts);
2085 gsi_commit_edge_inserts ();
2086 bb = gimple_bb (stmt);
2088 if (!bb_in_sese_p (bb, SCOP_REGION (scop)))
2089 return;
2091 if (!gbb_from_bb (bb))
2092 new_pbb_from_pbb (scop, pbb_from_bb (e->src), bb);
2094 analyze_drs_in_stmts (scop, bb, x);
2097 /* Creates a zero dimension array of the same type as VAR. */
2099 static tree
2100 create_zero_dim_array (tree var, const char *base_name)
2102 tree index_type = build_index_type (integer_zero_node);
2103 tree elt_type = TREE_TYPE (var);
2104 tree array_type = build_array_type (elt_type, index_type);
2105 tree base = create_tmp_var (array_type, base_name);
2107 return build4 (ARRAY_REF, elt_type, base, integer_zero_node, NULL_TREE,
2108 NULL_TREE);
2111 /* Returns true when PHI is a loop close phi node. */
2113 static bool
2114 scalar_close_phi_node_p (gimple phi)
2116 if (gimple_code (phi) != GIMPLE_PHI
2117 || virtual_operand_p (gimple_phi_result (phi)))
2118 return false;
2120 /* Note that loop close phi nodes should have a single argument
2121 because we translated the representation into a canonical form
2122 before Graphite: see canonicalize_loop_closed_ssa_form. */
2123 return (gimple_phi_num_args (phi) == 1);
2126 /* For a definition DEF in REGION, propagates the expression EXPR in
2127 all the uses of DEF outside REGION. */
2129 static void
2130 propagate_expr_outside_region (tree def, tree expr, sese region)
2132 imm_use_iterator imm_iter;
2133 gimple use_stmt;
2134 gimple_seq stmts;
2135 bool replaced_once = false;
2137 gcc_assert (TREE_CODE (def) == SSA_NAME);
2139 expr = force_gimple_operand (unshare_expr (expr), &stmts, true,
2140 NULL_TREE);
2142 FOR_EACH_IMM_USE_STMT (use_stmt, imm_iter, def)
2143 if (!is_gimple_debug (use_stmt)
2144 && !bb_in_sese_p (gimple_bb (use_stmt), region))
2146 ssa_op_iter iter;
2147 use_operand_p use_p;
2149 FOR_EACH_PHI_OR_STMT_USE (use_p, use_stmt, iter, SSA_OP_ALL_USES)
2150 if (operand_equal_p (def, USE_FROM_PTR (use_p), 0)
2151 && (replaced_once = true))
2152 replace_exp (use_p, expr);
2154 update_stmt (use_stmt);
2157 if (replaced_once)
2159 gsi_insert_seq_on_edge (SESE_ENTRY (region), stmts);
2160 gsi_commit_edge_inserts ();
2164 /* Rewrite out of SSA the reduction phi node at PSI by creating a zero
2165 dimension array for it. */
2167 static void
2168 rewrite_close_phi_out_of_ssa (scop_p scop, gimple_stmt_iterator *psi)
2170 sese region = SCOP_REGION (scop);
2171 gimple phi = gsi_stmt (*psi);
2172 tree res = gimple_phi_result (phi);
2173 basic_block bb = gimple_bb (phi);
2174 gimple_stmt_iterator gsi = gsi_after_labels (bb);
2175 tree arg = gimple_phi_arg_def (phi, 0);
2176 gimple stmt;
2178 /* Note that loop close phi nodes should have a single argument
2179 because we translated the representation into a canonical form
2180 before Graphite: see canonicalize_loop_closed_ssa_form. */
2181 gcc_assert (gimple_phi_num_args (phi) == 1);
2183 /* The phi node can be a non close phi node, when its argument is
2184 invariant, or a default definition. */
2185 if (is_gimple_min_invariant (arg)
2186 || SSA_NAME_IS_DEFAULT_DEF (arg))
2188 propagate_expr_outside_region (res, arg, region);
2189 gsi_next (psi);
2190 return;
2193 else if (gimple_bb (SSA_NAME_DEF_STMT (arg))->loop_father == bb->loop_father)
2195 propagate_expr_outside_region (res, arg, region);
2196 stmt = gimple_build_assign (res, arg);
2197 remove_phi_node (psi, false);
2198 gsi_insert_before (&gsi, stmt, GSI_NEW_STMT);
2199 return;
2202 /* If res is scev analyzable and is not a scalar value, it is safe
2203 to ignore the close phi node: it will be code generated in the
2204 out of Graphite pass. */
2205 else if (scev_analyzable_p (res, region))
2207 loop_p loop = loop_containing_stmt (SSA_NAME_DEF_STMT (res));
2208 tree scev;
2210 if (!loop_in_sese_p (loop, region))
2212 loop = loop_containing_stmt (SSA_NAME_DEF_STMT (arg));
2213 scev = scalar_evolution_in_region (region, loop, arg);
2214 scev = compute_overall_effect_of_inner_loop (loop, scev);
2216 else
2217 scev = scalar_evolution_in_region (region, loop, res);
2219 if (tree_does_not_contain_chrecs (scev))
2220 propagate_expr_outside_region (res, scev, region);
2222 gsi_next (psi);
2223 return;
2225 else
2227 tree zero_dim_array = create_zero_dim_array (res, "Close_Phi");
2229 stmt = gimple_build_assign (res, unshare_expr (zero_dim_array));
2231 if (TREE_CODE (arg) == SSA_NAME)
2232 insert_out_of_ssa_copy (scop, zero_dim_array, arg,
2233 SSA_NAME_DEF_STMT (arg));
2234 else
2235 insert_out_of_ssa_copy_on_edge (scop, single_pred_edge (bb),
2236 zero_dim_array, arg);
2239 remove_phi_node (psi, false);
2240 SSA_NAME_DEF_STMT (res) = stmt;
2242 insert_stmts (scop, stmt, NULL, gsi_after_labels (bb));
2245 /* Rewrite out of SSA the reduction phi node at PSI by creating a zero
2246 dimension array for it. */
2248 static void
2249 rewrite_phi_out_of_ssa (scop_p scop, gphi_iterator *psi)
2251 size_t i;
2252 gphi *phi = psi->phi ();
2253 basic_block bb = gimple_bb (phi);
2254 tree res = gimple_phi_result (phi);
2255 tree zero_dim_array = create_zero_dim_array (res, "phi_out_of_ssa");
2256 gimple stmt;
2258 for (i = 0; i < gimple_phi_num_args (phi); i++)
2260 tree arg = gimple_phi_arg_def (phi, i);
2261 edge e = gimple_phi_arg_edge (phi, i);
2263 /* Avoid the insertion of code in the loop latch to please the
2264 pattern matching of the vectorizer. */
2265 if (TREE_CODE (arg) == SSA_NAME
2266 && !SSA_NAME_IS_DEFAULT_DEF (arg)
2267 && e->src == bb->loop_father->latch)
2268 insert_out_of_ssa_copy (scop, zero_dim_array, arg,
2269 SSA_NAME_DEF_STMT (arg));
2270 else
2271 insert_out_of_ssa_copy_on_edge (scop, e, zero_dim_array, arg);
2274 stmt = gimple_build_assign (res, unshare_expr (zero_dim_array));
2275 remove_phi_node (psi, false);
2276 insert_stmts (scop, stmt, NULL, gsi_after_labels (bb));
2279 /* Rewrite the degenerate phi node at position PSI from the degenerate
2280 form "x = phi (y, y, ..., y)" to "x = y". */
2282 static void
2283 rewrite_degenerate_phi (gphi_iterator *psi)
2285 tree rhs;
2286 gimple stmt;
2287 gimple_stmt_iterator gsi;
2288 gphi *phi = psi->phi ();
2289 tree res = gimple_phi_result (phi);
2290 basic_block bb;
2292 bb = gimple_bb (phi);
2293 rhs = degenerate_phi_result (phi);
2294 gcc_assert (rhs);
2296 stmt = gimple_build_assign (res, rhs);
2297 remove_phi_node (psi, false);
2299 gsi = gsi_after_labels (bb);
2300 gsi_insert_before (&gsi, stmt, GSI_NEW_STMT);
2303 /* Rewrite out of SSA all the reduction phi nodes of SCOP. */
2305 static void
2306 rewrite_reductions_out_of_ssa (scop_p scop)
2308 basic_block bb;
2309 gphi_iterator psi;
2310 sese region = SCOP_REGION (scop);
2312 FOR_EACH_BB_FN (bb, cfun)
2313 if (bb_in_sese_p (bb, region))
2314 for (psi = gsi_start_phis (bb); !gsi_end_p (psi);)
2316 gphi *phi = psi.phi ();
2318 if (virtual_operand_p (gimple_phi_result (phi)))
2320 gsi_next (&psi);
2321 continue;
2324 if (gimple_phi_num_args (phi) > 1
2325 && degenerate_phi_result (phi))
2326 rewrite_degenerate_phi (&psi);
2328 else if (scalar_close_phi_node_p (phi))
2329 rewrite_close_phi_out_of_ssa (scop, &psi);
2331 else if (reduction_phi_p (region, &psi))
2332 rewrite_phi_out_of_ssa (scop, &psi);
2335 update_ssa (TODO_update_ssa);
2336 #ifdef ENABLE_CHECKING
2337 verify_loop_closed_ssa (true);
2338 #endif
2341 /* Rewrite the scalar dependence of DEF used in USE_STMT with a memory
2342 read from ZERO_DIM_ARRAY. */
2344 static void
2345 rewrite_cross_bb_scalar_dependence (scop_p scop, tree zero_dim_array,
2346 tree def, gimple use_stmt)
2348 gimple name_stmt;
2349 tree name;
2350 ssa_op_iter iter;
2351 use_operand_p use_p;
2353 gcc_assert (gimple_code (use_stmt) != GIMPLE_PHI);
2355 name = copy_ssa_name (def);
2356 name_stmt = gimple_build_assign (name, zero_dim_array);
2358 gimple_assign_set_lhs (name_stmt, name);
2359 insert_stmts (scop, name_stmt, NULL, gsi_for_stmt (use_stmt));
2361 FOR_EACH_SSA_USE_OPERAND (use_p, use_stmt, iter, SSA_OP_ALL_USES)
2362 if (operand_equal_p (def, USE_FROM_PTR (use_p), 0))
2363 replace_exp (use_p, name);
2365 update_stmt (use_stmt);
2368 /* For every definition DEF in the SCOP that is used outside the scop,
2369 insert a closing-scop definition in the basic block just after this
2370 SCOP. */
2372 static void
2373 handle_scalar_deps_crossing_scop_limits (scop_p scop, tree def, gimple stmt)
2375 tree var = create_tmp_reg (TREE_TYPE (def));
2376 tree new_name = make_ssa_name (var, stmt);
2377 bool needs_copy = false;
2378 use_operand_p use_p;
2379 imm_use_iterator imm_iter;
2380 gimple use_stmt;
2381 sese region = SCOP_REGION (scop);
2383 FOR_EACH_IMM_USE_STMT (use_stmt, imm_iter, def)
2385 if (!bb_in_sese_p (gimple_bb (use_stmt), region))
2387 FOR_EACH_IMM_USE_ON_STMT (use_p, imm_iter)
2389 SET_USE (use_p, new_name);
2391 update_stmt (use_stmt);
2392 needs_copy = true;
2396 /* Insert in the empty BB just after the scop a use of DEF such
2397 that the rewrite of cross_bb_scalar_dependences won't insert
2398 arrays everywhere else. */
2399 if (needs_copy)
2401 gimple assign = gimple_build_assign (new_name, def);
2402 gimple_stmt_iterator psi = gsi_after_labels (SESE_EXIT (region)->dest);
2404 update_stmt (assign);
2405 gsi_insert_before (&psi, assign, GSI_SAME_STMT);
2409 /* Rewrite the scalar dependences crossing the boundary of the BB
2410 containing STMT with an array. Return true when something has been
2411 changed. */
2413 static bool
2414 rewrite_cross_bb_scalar_deps (scop_p scop, gimple_stmt_iterator *gsi)
2416 sese region = SCOP_REGION (scop);
2417 gimple stmt = gsi_stmt (*gsi);
2418 imm_use_iterator imm_iter;
2419 tree def;
2420 basic_block def_bb;
2421 tree zero_dim_array = NULL_TREE;
2422 gimple use_stmt;
2423 bool res = false;
2425 switch (gimple_code (stmt))
2427 case GIMPLE_ASSIGN:
2428 def = gimple_assign_lhs (stmt);
2429 break;
2431 case GIMPLE_CALL:
2432 def = gimple_call_lhs (stmt);
2433 break;
2435 default:
2436 return false;
2439 if (!def
2440 || !is_gimple_reg (def))
2441 return false;
2443 if (scev_analyzable_p (def, region))
2445 loop_p loop = loop_containing_stmt (SSA_NAME_DEF_STMT (def));
2446 tree scev = scalar_evolution_in_region (region, loop, def);
2448 if (tree_contains_chrecs (scev, NULL))
2449 return false;
2451 propagate_expr_outside_region (def, scev, region);
2452 return true;
2455 def_bb = gimple_bb (stmt);
2457 handle_scalar_deps_crossing_scop_limits (scop, def, stmt);
2459 FOR_EACH_IMM_USE_STMT (use_stmt, imm_iter, def)
2460 if (gimple_code (use_stmt) == GIMPLE_PHI
2461 && (res = true))
2463 gphi_iterator psi = gsi_start_phis (gimple_bb (use_stmt));
2465 if (scalar_close_phi_node_p (gsi_stmt (psi)))
2466 rewrite_close_phi_out_of_ssa (scop, &psi);
2467 else
2468 rewrite_phi_out_of_ssa (scop, &psi);
2471 FOR_EACH_IMM_USE_STMT (use_stmt, imm_iter, def)
2472 if (gimple_code (use_stmt) != GIMPLE_PHI
2473 && def_bb != gimple_bb (use_stmt)
2474 && !is_gimple_debug (use_stmt)
2475 && (res = true))
2477 if (!zero_dim_array)
2479 zero_dim_array = create_zero_dim_array
2480 (def, "Cross_BB_scalar_dependence");
2481 insert_out_of_ssa_copy (scop, zero_dim_array, def,
2482 SSA_NAME_DEF_STMT (def));
2483 gsi_next (gsi);
2486 rewrite_cross_bb_scalar_dependence (scop, unshare_expr (zero_dim_array),
2487 def, use_stmt);
2490 return res;
2493 /* Rewrite out of SSA all the reduction phi nodes of SCOP. */
2495 static void
2496 rewrite_cross_bb_scalar_deps_out_of_ssa (scop_p scop)
2498 basic_block bb;
2499 gimple_stmt_iterator psi;
2500 sese region = SCOP_REGION (scop);
2501 bool changed = false;
2503 /* Create an extra empty BB after the scop. */
2504 split_edge (SESE_EXIT (region));
2506 FOR_EACH_BB_FN (bb, cfun)
2507 if (bb_in_sese_p (bb, region))
2508 for (psi = gsi_start_bb (bb); !gsi_end_p (psi); gsi_next (&psi))
2509 changed |= rewrite_cross_bb_scalar_deps (scop, &psi);
2511 if (changed)
2513 scev_reset_htab ();
2514 update_ssa (TODO_update_ssa);
2515 #ifdef ENABLE_CHECKING
2516 verify_loop_closed_ssa (true);
2517 #endif
2521 /* Returns the number of pbbs that are in loops contained in SCOP. */
2523 static int
2524 nb_pbbs_in_loops (scop_p scop)
2526 int i;
2527 poly_bb_p pbb;
2528 int res = 0;
2530 FOR_EACH_VEC_ELT (SCOP_BBS (scop), i, pbb)
2531 if (loop_in_sese_p (gbb_loop (PBB_BLACK_BOX (pbb)), SCOP_REGION (scop)))
2532 res++;
2534 return res;
2537 /* Return the number of data references in BB that write in
2538 memory. */
2540 static int
2541 nb_data_writes_in_bb (basic_block bb)
2543 int res = 0;
2544 gimple_stmt_iterator gsi;
2546 for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi))
2547 if (gimple_vdef (gsi_stmt (gsi)))
2548 res++;
2550 return res;
2553 /* Splits at STMT the basic block BB represented as PBB in the
2554 polyhedral form. */
2556 static edge
2557 split_pbb (scop_p scop, poly_bb_p pbb, basic_block bb, gimple stmt)
2559 edge e1 = split_block (bb, stmt);
2560 new_pbb_from_pbb (scop, pbb, e1->dest);
2561 return e1;
2564 /* Splits STMT out of its current BB. This is done for reduction
2565 statements for which we want to ignore data dependences. */
2567 static basic_block
2568 split_reduction_stmt (scop_p scop, gimple stmt)
2570 basic_block bb = gimple_bb (stmt);
2571 poly_bb_p pbb = pbb_from_bb (bb);
2572 gimple_bb_p gbb = gbb_from_bb (bb);
2573 edge e1;
2574 int i;
2575 data_reference_p dr;
2577 /* Do not split basic blocks with no writes to memory: the reduction
2578 will be the only write to memory. */
2579 if (nb_data_writes_in_bb (bb) == 0
2580 /* Or if we have already marked BB as a reduction. */
2581 || PBB_IS_REDUCTION (pbb_from_bb (bb)))
2582 return bb;
2584 e1 = split_pbb (scop, pbb, bb, stmt);
2586 /* Split once more only when the reduction stmt is not the only one
2587 left in the original BB. */
2588 if (!gsi_one_before_end_p (gsi_start_nondebug_bb (bb)))
2590 gimple_stmt_iterator gsi = gsi_last_bb (bb);
2591 gsi_prev (&gsi);
2592 e1 = split_pbb (scop, pbb, bb, gsi_stmt (gsi));
2595 /* A part of the data references will end in a different basic block
2596 after the split: move the DRs from the original GBB to the newly
2597 created GBB1. */
2598 FOR_EACH_VEC_ELT (GBB_DATA_REFS (gbb), i, dr)
2600 basic_block bb1 = gimple_bb (DR_STMT (dr));
2602 if (bb1 != bb)
2604 gimple_bb_p gbb1 = gbb_from_bb (bb1);
2605 GBB_DATA_REFS (gbb1).safe_push (dr);
2606 GBB_DATA_REFS (gbb).ordered_remove (i);
2607 i--;
2611 return e1->dest;
2614 /* Return true when stmt is a reduction operation. */
2616 static inline bool
2617 is_reduction_operation_p (gimple stmt)
2619 enum tree_code code;
2621 gcc_assert (is_gimple_assign (stmt));
2622 code = gimple_assign_rhs_code (stmt);
2624 return flag_associative_math
2625 && commutative_tree_code (code)
2626 && associative_tree_code (code);
2629 /* Returns true when PHI contains an argument ARG. */
2631 static bool
2632 phi_contains_arg (gphi *phi, tree arg)
2634 size_t i;
2636 for (i = 0; i < gimple_phi_num_args (phi); i++)
2637 if (operand_equal_p (arg, gimple_phi_arg_def (phi, i), 0))
2638 return true;
2640 return false;
2643 /* Return a loop phi node that corresponds to a reduction containing LHS. */
2645 static gphi *
2646 follow_ssa_with_commutative_ops (tree arg, tree lhs)
2648 gimple stmt;
2650 if (TREE_CODE (arg) != SSA_NAME)
2651 return NULL;
2653 stmt = SSA_NAME_DEF_STMT (arg);
2655 if (gimple_code (stmt) == GIMPLE_NOP
2656 || gimple_code (stmt) == GIMPLE_CALL)
2657 return NULL;
2659 if (gphi *phi = dyn_cast <gphi *> (stmt))
2661 if (phi_contains_arg (phi, lhs))
2662 return phi;
2663 return NULL;
2666 if (!is_gimple_assign (stmt))
2667 return NULL;
2669 if (gimple_num_ops (stmt) == 2)
2670 return follow_ssa_with_commutative_ops (gimple_assign_rhs1 (stmt), lhs);
2672 if (is_reduction_operation_p (stmt))
2674 gphi *res
2675 = follow_ssa_with_commutative_ops (gimple_assign_rhs1 (stmt), lhs);
2677 return res ? res :
2678 follow_ssa_with_commutative_ops (gimple_assign_rhs2 (stmt), lhs);
2681 return NULL;
2684 /* Detect commutative and associative scalar reductions starting at
2685 the STMT. Return the phi node of the reduction cycle, or NULL. */
2687 static gphi *
2688 detect_commutative_reduction_arg (tree lhs, gimple stmt, tree arg,
2689 vec<gimple> *in,
2690 vec<gimple> *out)
2692 gphi *phi = follow_ssa_with_commutative_ops (arg, lhs);
2694 if (!phi)
2695 return NULL;
2697 in->safe_push (stmt);
2698 out->safe_push (stmt);
2699 return phi;
2702 /* Detect commutative and associative scalar reductions starting at
2703 STMT. Return the phi node of the reduction cycle, or NULL. */
2705 static gphi *
2706 detect_commutative_reduction_assign (gimple stmt, vec<gimple> *in,
2707 vec<gimple> *out)
2709 tree lhs = gimple_assign_lhs (stmt);
2711 if (gimple_num_ops (stmt) == 2)
2712 return detect_commutative_reduction_arg (lhs, stmt,
2713 gimple_assign_rhs1 (stmt),
2714 in, out);
2716 if (is_reduction_operation_p (stmt))
2718 gphi *res = detect_commutative_reduction_arg (lhs, stmt,
2719 gimple_assign_rhs1 (stmt),
2720 in, out);
2721 return res ? res
2722 : detect_commutative_reduction_arg (lhs, stmt,
2723 gimple_assign_rhs2 (stmt),
2724 in, out);
2727 return NULL;
2730 /* Return a loop phi node that corresponds to a reduction containing LHS. */
2732 static gphi *
2733 follow_inital_value_to_phi (tree arg, tree lhs)
2735 gimple stmt;
2737 if (!arg || TREE_CODE (arg) != SSA_NAME)
2738 return NULL;
2740 stmt = SSA_NAME_DEF_STMT (arg);
2742 if (gphi *phi = dyn_cast <gphi *> (stmt))
2743 if (phi_contains_arg (phi, lhs))
2744 return phi;
2746 return NULL;
2750 /* Return the argument of the loop PHI that is the initial value coming
2751 from outside the loop. */
2753 static edge
2754 edge_initial_value_for_loop_phi (gphi *phi)
2756 size_t i;
2758 for (i = 0; i < gimple_phi_num_args (phi); i++)
2760 edge e = gimple_phi_arg_edge (phi, i);
2762 if (loop_depth (e->src->loop_father)
2763 < loop_depth (e->dest->loop_father))
2764 return e;
2767 return NULL;
2770 /* Return the argument of the loop PHI that is the initial value coming
2771 from outside the loop. */
2773 static tree
2774 initial_value_for_loop_phi (gphi *phi)
2776 size_t i;
2778 for (i = 0; i < gimple_phi_num_args (phi); i++)
2780 edge e = gimple_phi_arg_edge (phi, i);
2782 if (loop_depth (e->src->loop_father)
2783 < loop_depth (e->dest->loop_father))
2784 return gimple_phi_arg_def (phi, i);
2787 return NULL_TREE;
2790 /* Returns true when DEF is used outside the reduction cycle of
2791 LOOP_PHI. */
2793 static bool
2794 used_outside_reduction (tree def, gimple loop_phi)
2796 use_operand_p use_p;
2797 imm_use_iterator imm_iter;
2798 loop_p loop = loop_containing_stmt (loop_phi);
2800 /* In LOOP, DEF should be used only in LOOP_PHI. */
2801 FOR_EACH_IMM_USE_FAST (use_p, imm_iter, def)
2803 gimple stmt = USE_STMT (use_p);
2805 if (stmt != loop_phi
2806 && !is_gimple_debug (stmt)
2807 && flow_bb_inside_loop_p (loop, gimple_bb (stmt)))
2808 return true;
2811 return false;
2814 /* Detect commutative and associative scalar reductions belonging to
2815 the SCOP starting at the loop closed phi node STMT. Return the phi
2816 node of the reduction cycle, or NULL. */
2818 static gphi *
2819 detect_commutative_reduction (scop_p scop, gimple stmt, vec<gimple> *in,
2820 vec<gimple> *out)
2822 if (scalar_close_phi_node_p (stmt))
2824 gimple def;
2825 gphi *loop_phi, *phi, *close_phi = as_a <gphi *> (stmt);
2826 tree init, lhs, arg = gimple_phi_arg_def (close_phi, 0);
2828 if (TREE_CODE (arg) != SSA_NAME)
2829 return NULL;
2831 /* Note that loop close phi nodes should have a single argument
2832 because we translated the representation into a canonical form
2833 before Graphite: see canonicalize_loop_closed_ssa_form. */
2834 gcc_assert (gimple_phi_num_args (close_phi) == 1);
2836 def = SSA_NAME_DEF_STMT (arg);
2837 if (!stmt_in_sese_p (def, SCOP_REGION (scop))
2838 || !(loop_phi = detect_commutative_reduction (scop, def, in, out)))
2839 return NULL;
2841 lhs = gimple_phi_result (close_phi);
2842 init = initial_value_for_loop_phi (loop_phi);
2843 phi = follow_inital_value_to_phi (init, lhs);
2845 if (phi && (used_outside_reduction (lhs, phi)
2846 || !has_single_use (gimple_phi_result (phi))))
2847 return NULL;
2849 in->safe_push (loop_phi);
2850 out->safe_push (close_phi);
2851 return phi;
2854 if (gimple_code (stmt) == GIMPLE_ASSIGN)
2855 return detect_commutative_reduction_assign (stmt, in, out);
2857 return NULL;
2860 /* Translate the scalar reduction statement STMT to an array RED
2861 knowing that its recursive phi node is LOOP_PHI. */
2863 static void
2864 translate_scalar_reduction_to_array_for_stmt (scop_p scop, tree red,
2865 gimple stmt, gphi *loop_phi)
2867 tree res = gimple_phi_result (loop_phi);
2868 gassign *assign = gimple_build_assign (res, unshare_expr (red));
2869 gimple_stmt_iterator gsi;
2871 insert_stmts (scop, assign, NULL, gsi_after_labels (gimple_bb (loop_phi)));
2873 assign = gimple_build_assign (unshare_expr (red), gimple_assign_lhs (stmt));
2874 gsi = gsi_for_stmt (stmt);
2875 gsi_next (&gsi);
2876 insert_stmts (scop, assign, NULL, gsi);
2879 /* Removes the PHI node and resets all the debug stmts that are using
2880 the PHI_RESULT. */
2882 static void
2883 remove_phi (gphi *phi)
2885 imm_use_iterator imm_iter;
2886 tree def;
2887 use_operand_p use_p;
2888 gimple_stmt_iterator gsi;
2889 auto_vec<gimple, 3> update;
2890 unsigned int i;
2891 gimple stmt;
2893 def = PHI_RESULT (phi);
2894 FOR_EACH_IMM_USE_FAST (use_p, imm_iter, def)
2896 stmt = USE_STMT (use_p);
2898 if (is_gimple_debug (stmt))
2900 gimple_debug_bind_reset_value (stmt);
2901 update.safe_push (stmt);
2905 FOR_EACH_VEC_ELT (update, i, stmt)
2906 update_stmt (stmt);
2908 gsi = gsi_for_phi_node (phi);
2909 remove_phi_node (&gsi, false);
2912 /* Helper function for for_each_index. For each INDEX of the data
2913 reference REF, returns true when its indices are valid in the loop
2914 nest LOOP passed in as DATA. */
2916 static bool
2917 dr_indices_valid_in_loop (tree ref ATTRIBUTE_UNUSED, tree *index, void *data)
2919 loop_p loop;
2920 basic_block header, def_bb;
2921 gimple stmt;
2923 if (TREE_CODE (*index) != SSA_NAME)
2924 return true;
2926 loop = *((loop_p *) data);
2927 header = loop->header;
2928 stmt = SSA_NAME_DEF_STMT (*index);
2930 if (!stmt)
2931 return true;
2933 def_bb = gimple_bb (stmt);
2935 if (!def_bb)
2936 return true;
2938 return dominated_by_p (CDI_DOMINATORS, header, def_bb);
2941 /* When the result of a CLOSE_PHI is written to a memory location,
2942 return a pointer to that memory reference, otherwise return
2943 NULL_TREE. */
2945 static tree
2946 close_phi_written_to_memory (gphi *close_phi)
2948 imm_use_iterator imm_iter;
2949 use_operand_p use_p;
2950 gimple stmt;
2951 tree res, def = gimple_phi_result (close_phi);
2953 FOR_EACH_IMM_USE_FAST (use_p, imm_iter, def)
2954 if ((stmt = USE_STMT (use_p))
2955 && gimple_code (stmt) == GIMPLE_ASSIGN
2956 && (res = gimple_assign_lhs (stmt)))
2958 switch (TREE_CODE (res))
2960 case VAR_DECL:
2961 case PARM_DECL:
2962 case RESULT_DECL:
2963 return res;
2965 case ARRAY_REF:
2966 case MEM_REF:
2968 tree arg = gimple_phi_arg_def (close_phi, 0);
2969 loop_p nest = loop_containing_stmt (SSA_NAME_DEF_STMT (arg));
2971 /* FIXME: this restriction is for id-{24,25}.f and
2972 could be handled by duplicating the computation of
2973 array indices before the loop of the close_phi. */
2974 if (for_each_index (&res, dr_indices_valid_in_loop, &nest))
2975 return res;
2977 /* Fallthru. */
2979 default:
2980 continue;
2983 return NULL_TREE;
2986 /* Rewrite out of SSA the reduction described by the loop phi nodes
2987 IN, and the close phi nodes OUT. IN and OUT are structured by loop
2988 levels like this:
2990 IN: stmt, loop_n, ..., loop_0
2991 OUT: stmt, close_n, ..., close_0
2993 the first element is the reduction statement, and the next elements
2994 are the loop and close phi nodes of each of the outer loops. */
2996 static void
2997 translate_scalar_reduction_to_array (scop_p scop,
2998 vec<gimple> in,
2999 vec<gimple> out)
3001 gimple loop_stmt;
3002 unsigned int i = out.length () - 1;
3003 tree red = close_phi_written_to_memory (as_a <gphi *> (out[i]));
3005 FOR_EACH_VEC_ELT (in, i, loop_stmt)
3007 gimple close_stmt = out[i];
3009 if (i == 0)
3011 basic_block bb = split_reduction_stmt (scop, loop_stmt);
3012 poly_bb_p pbb = pbb_from_bb (bb);
3013 PBB_IS_REDUCTION (pbb) = true;
3014 gcc_assert (close_stmt == loop_stmt);
3016 if (!red)
3017 red = create_zero_dim_array
3018 (gimple_assign_lhs (loop_stmt), "Commutative_Associative_Reduction");
3020 translate_scalar_reduction_to_array_for_stmt (scop, red, loop_stmt,
3021 as_a <gphi *> (in[1]));
3022 continue;
3025 gphi *loop_phi = as_a <gphi *> (loop_stmt);
3026 gphi *close_phi = as_a <gphi *> (close_stmt);
3028 if (i == in.length () - 1)
3030 insert_out_of_ssa_copy (scop, gimple_phi_result (close_phi),
3031 unshare_expr (red), close_phi);
3032 insert_out_of_ssa_copy_on_edge
3033 (scop, edge_initial_value_for_loop_phi (loop_phi),
3034 unshare_expr (red), initial_value_for_loop_phi (loop_phi));
3037 remove_phi (loop_phi);
3038 remove_phi (close_phi);
3042 /* Rewrites out of SSA a commutative reduction at CLOSE_PHI. Returns
3043 true when something has been changed. */
3045 static bool
3046 rewrite_commutative_reductions_out_of_ssa_close_phi (scop_p scop,
3047 gphi *close_phi)
3049 bool res;
3050 auto_vec<gimple, 10> in;
3051 auto_vec<gimple, 10> out;
3053 detect_commutative_reduction (scop, close_phi, &in, &out);
3054 res = in.length () > 1;
3055 if (res)
3056 translate_scalar_reduction_to_array (scop, in, out);
3058 return res;
3061 /* Rewrites all the commutative reductions from LOOP out of SSA.
3062 Returns true when something has been changed. */
3064 static bool
3065 rewrite_commutative_reductions_out_of_ssa_loop (scop_p scop,
3066 loop_p loop)
3068 gphi_iterator gsi;
3069 edge exit = single_exit (loop);
3070 tree res;
3071 bool changed = false;
3073 if (!exit)
3074 return false;
3076 for (gsi = gsi_start_phis (exit->dest); !gsi_end_p (gsi); gsi_next (&gsi))
3077 if ((res = gimple_phi_result (gsi.phi ()))
3078 && !virtual_operand_p (res)
3079 && !scev_analyzable_p (res, SCOP_REGION (scop)))
3080 changed |= rewrite_commutative_reductions_out_of_ssa_close_phi
3081 (scop, gsi.phi ());
3083 return changed;
3086 /* Rewrites all the commutative reductions from SCOP out of SSA. */
3088 static void
3089 rewrite_commutative_reductions_out_of_ssa (scop_p scop)
3091 loop_p loop;
3092 bool changed = false;
3093 sese region = SCOP_REGION (scop);
3095 FOR_EACH_LOOP (loop, 0)
3096 if (loop_in_sese_p (loop, region))
3097 changed |= rewrite_commutative_reductions_out_of_ssa_loop (scop, loop);
3099 if (changed)
3101 scev_reset_htab ();
3102 gsi_commit_edge_inserts ();
3103 update_ssa (TODO_update_ssa);
3104 #ifdef ENABLE_CHECKING
3105 verify_loop_closed_ssa (true);
3106 #endif
3110 /* Can all ivs be represented by a signed integer?
3111 As CLooG might generate negative values in its expressions, signed loop ivs
3112 are required in the backend. */
3114 static bool
3115 scop_ivs_can_be_represented (scop_p scop)
3117 loop_p loop;
3118 gphi_iterator psi;
3119 bool result = true;
3121 FOR_EACH_LOOP (loop, 0)
3123 if (!loop_in_sese_p (loop, SCOP_REGION (scop)))
3124 continue;
3126 for (psi = gsi_start_phis (loop->header);
3127 !gsi_end_p (psi); gsi_next (&psi))
3129 gphi *phi = psi.phi ();
3130 tree res = PHI_RESULT (phi);
3131 tree type = TREE_TYPE (res);
3133 if (TYPE_UNSIGNED (type)
3134 && TYPE_PRECISION (type) >= TYPE_PRECISION (long_long_integer_type_node))
3136 result = false;
3137 break;
3140 if (!result)
3141 break;
3144 return result;
3147 /* Builds the polyhedral representation for a SESE region. */
3149 void
3150 build_poly_scop (scop_p scop)
3152 sese region = SCOP_REGION (scop);
3153 graphite_dim_t max_dim;
3155 build_scop_bbs (scop);
3157 /* FIXME: This restriction is needed to avoid a problem in CLooG.
3158 Once CLooG is fixed, remove this guard. Anyways, it makes no
3159 sense to optimize a scop containing only PBBs that do not belong
3160 to any loops. */
3161 if (nb_pbbs_in_loops (scop) == 0)
3162 return;
3164 if (!scop_ivs_can_be_represented (scop))
3165 return;
3167 if (flag_associative_math)
3168 rewrite_commutative_reductions_out_of_ssa (scop);
3170 build_sese_loop_nests (region);
3171 /* Record all conditions in REGION. */
3172 sese_dom_walker (CDI_DOMINATORS, region).walk (cfun->cfg->x_entry_block_ptr);
3173 find_scop_parameters (scop);
3175 max_dim = PARAM_VALUE (PARAM_GRAPHITE_MAX_NB_SCOP_PARAMS);
3176 if (scop_nb_params (scop) > max_dim)
3177 return;
3179 build_scop_iteration_domain (scop);
3180 build_scop_context (scop);
3181 add_conditions_to_constraints (scop);
3183 /* Rewrite out of SSA only after having translated the
3184 representation to the polyhedral representation to avoid scev
3185 analysis failures. That means that these functions will insert
3186 new data references that they create in the right place. */
3187 rewrite_reductions_out_of_ssa (scop);
3188 rewrite_cross_bb_scalar_deps_out_of_ssa (scop);
3190 build_scop_drs (scop);
3191 scop_to_lst (scop);
3192 build_scop_scattering (scop);
3194 /* This SCoP has been translated to the polyhedral
3195 representation. */
3196 POLY_SCOP_P (scop) = true;
3198 #endif