* include/bits/basic_string.h (getline): Qualify call to prevent ADL
[official-gcc.git] / gcc / graphite-sese-to-poly.c
blobc36420a2b78f5dddddcc2bc72efa498f66144a8a
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_cloog
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 #include <cloog/cloog.h>
42 #include <cloog/cloog.h>
43 #include <cloog/isl/domain.h>
44 #endif
46 #include "system.h"
47 #include "coretypes.h"
48 #include "tree.h"
49 #include "basic-block.h"
50 #include "tree-ssa-alias.h"
51 #include "internal-fn.h"
52 #include "gimple-expr.h"
53 #include "is-a.h"
54 #include "gimple.h"
55 #include "gimple-iterator.h"
56 #include "gimplify.h"
57 #include "gimplify-me.h"
58 #include "gimple-ssa.h"
59 #include "tree-cfg.h"
60 #include "tree-phinodes.h"
61 #include "ssa-iterators.h"
62 #include "stringpool.h"
63 #include "tree-ssanames.h"
64 #include "tree-ssa-loop-manip.h"
65 #include "tree-ssa-loop-niter.h"
66 #include "tree-ssa-loop.h"
67 #include "tree-into-ssa.h"
68 #include "tree-pass.h"
69 #include "cfgloop.h"
70 #include "tree-chrec.h"
71 #include "tree-data-ref.h"
72 #include "tree-scalar-evolution.h"
73 #include "domwalk.h"
74 #include "sese.h"
75 #include "tree-ssa-propagate.h"
77 #ifdef HAVE_cloog
78 #include "expr.h"
79 #include "graphite-poly.h"
80 #include "graphite-sese-to-poly.h"
83 /* Assigns to RES the value of the INTEGER_CST T. */
85 static inline void
86 tree_int_to_gmp (tree t, mpz_t res)
88 wi::to_mpz (t, res, TYPE_SIGN (TREE_TYPE (t)));
91 /* Returns the index of the PHI argument defined in the outermost
92 loop. */
94 static size_t
95 phi_arg_in_outermost_loop (gimple phi)
97 loop_p loop = gimple_bb (phi)->loop_father;
98 size_t i, res = 0;
100 for (i = 0; i < gimple_phi_num_args (phi); i++)
101 if (!flow_bb_inside_loop_p (loop, gimple_phi_arg_edge (phi, i)->src))
103 loop = gimple_phi_arg_edge (phi, i)->src->loop_father;
104 res = i;
107 return res;
110 /* Removes a simple copy phi node "RES = phi (INIT, RES)" at position
111 PSI by inserting on the loop ENTRY edge assignment "RES = INIT". */
113 static void
114 remove_simple_copy_phi (gimple_stmt_iterator *psi)
116 gimple phi = gsi_stmt (*psi);
117 tree res = gimple_phi_result (phi);
118 size_t entry = phi_arg_in_outermost_loop (phi);
119 tree init = gimple_phi_arg_def (phi, entry);
120 gimple stmt = gimple_build_assign (res, init);
121 edge e = gimple_phi_arg_edge (phi, entry);
123 remove_phi_node (psi, false);
124 gsi_insert_on_edge_immediate (e, stmt);
127 /* Removes an invariant phi node at position PSI by inserting on the
128 loop ENTRY edge the assignment RES = INIT. */
130 static void
131 remove_invariant_phi (sese region, gimple_stmt_iterator *psi)
133 gimple phi = gsi_stmt (*psi);
134 loop_p loop = loop_containing_stmt (phi);
135 tree res = gimple_phi_result (phi);
136 tree scev = scalar_evolution_in_region (region, loop, res);
137 size_t entry = phi_arg_in_outermost_loop (phi);
138 edge e = gimple_phi_arg_edge (phi, entry);
139 tree var;
140 gimple stmt;
141 gimple_seq stmts = NULL;
143 if (tree_contains_chrecs (scev, NULL))
144 scev = gimple_phi_arg_def (phi, entry);
146 var = force_gimple_operand (scev, &stmts, true, NULL_TREE);
147 stmt = gimple_build_assign (res, var);
148 remove_phi_node (psi, false);
150 gimple_seq_add_stmt (&stmts, stmt);
151 gsi_insert_seq_on_edge (e, stmts);
152 gsi_commit_edge_inserts ();
153 SSA_NAME_DEF_STMT (res) = stmt;
156 /* Returns true when the phi node at PSI is of the form "a = phi (a, x)". */
158 static inline bool
159 simple_copy_phi_p (gimple phi)
161 tree res;
163 if (gimple_phi_num_args (phi) != 2)
164 return false;
166 res = gimple_phi_result (phi);
167 return (res == gimple_phi_arg_def (phi, 0)
168 || res == gimple_phi_arg_def (phi, 1));
171 /* Returns true when the phi node at position PSI is a reduction phi
172 node in REGION. Otherwise moves the pointer PSI to the next phi to
173 be considered. */
175 static bool
176 reduction_phi_p (sese region, gimple_stmt_iterator *psi)
178 loop_p loop;
179 gimple phi = gsi_stmt (*psi);
180 tree res = gimple_phi_result (phi);
182 loop = loop_containing_stmt (phi);
184 if (simple_copy_phi_p (phi))
186 /* PRE introduces phi nodes like these, for an example,
187 see id-5.f in the fortran graphite testsuite:
189 # prephitmp.85_265 = PHI <prephitmp.85_258(33), prephitmp.85_265(18)>
191 remove_simple_copy_phi (psi);
192 return false;
195 if (scev_analyzable_p (res, region))
197 tree scev = scalar_evolution_in_region (region, loop, res);
199 if (evolution_function_is_invariant_p (scev, loop->num))
200 remove_invariant_phi (region, psi);
201 else
202 gsi_next (psi);
204 return false;
207 /* All the other cases are considered reductions. */
208 return true;
211 /* Store the GRAPHITE representation of BB. */
213 static gimple_bb_p
214 new_gimple_bb (basic_block bb, vec<data_reference_p> drs)
216 struct gimple_bb *gbb;
218 gbb = XNEW (struct gimple_bb);
219 bb->aux = gbb;
220 GBB_BB (gbb) = bb;
221 GBB_DATA_REFS (gbb) = drs;
222 GBB_CONDITIONS (gbb).create (0);
223 GBB_CONDITION_CASES (gbb).create (0);
225 return gbb;
228 static void
229 free_data_refs_aux (vec<data_reference_p> datarefs)
231 unsigned int i;
232 struct data_reference *dr;
234 FOR_EACH_VEC_ELT (datarefs, i, dr)
235 if (dr->aux)
237 base_alias_pair *bap = (base_alias_pair *)(dr->aux);
239 free (bap->alias_set);
241 free (bap);
242 dr->aux = NULL;
245 /* Frees GBB. */
247 static void
248 free_gimple_bb (struct gimple_bb *gbb)
250 free_data_refs_aux (GBB_DATA_REFS (gbb));
251 free_data_refs (GBB_DATA_REFS (gbb));
253 GBB_CONDITIONS (gbb).release ();
254 GBB_CONDITION_CASES (gbb).release ();
255 GBB_BB (gbb)->aux = 0;
256 XDELETE (gbb);
259 /* Deletes all gimple bbs in SCOP. */
261 static void
262 remove_gbbs_in_scop (scop_p scop)
264 int i;
265 poly_bb_p pbb;
267 FOR_EACH_VEC_ELT (SCOP_BBS (scop), i, pbb)
268 free_gimple_bb (PBB_BLACK_BOX (pbb));
271 /* Deletes all scops in SCOPS. */
273 void
274 free_scops (vec<scop_p> scops)
276 int i;
277 scop_p scop;
279 FOR_EACH_VEC_ELT (scops, i, scop)
281 remove_gbbs_in_scop (scop);
282 free_sese (SCOP_REGION (scop));
283 free_scop (scop);
286 scops.release ();
289 /* Same as outermost_loop_in_sese, returns the outermost loop
290 containing BB in REGION, but makes sure that the returned loop
291 belongs to the REGION, and so this returns the first loop in the
292 REGION when the loop containing BB does not belong to REGION. */
294 static loop_p
295 outermost_loop_in_sese_1 (sese region, basic_block bb)
297 loop_p nest = outermost_loop_in_sese (region, bb);
299 if (loop_in_sese_p (nest, region))
300 return nest;
302 /* When the basic block BB does not belong to a loop in the region,
303 return the first loop in the region. */
304 nest = nest->inner;
305 while (nest)
306 if (loop_in_sese_p (nest, region))
307 break;
308 else
309 nest = nest->next;
311 gcc_assert (nest);
312 return nest;
315 /* Generates a polyhedral black box only if the bb contains interesting
316 information. */
318 static gimple_bb_p
319 try_generate_gimple_bb (scop_p scop, basic_block bb)
321 vec<data_reference_p> drs;
322 drs.create (5);
323 sese region = SCOP_REGION (scop);
324 loop_p nest = outermost_loop_in_sese_1 (region, bb);
325 gimple_stmt_iterator gsi;
327 for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi))
329 gimple stmt = gsi_stmt (gsi);
330 loop_p loop;
332 if (is_gimple_debug (stmt))
333 continue;
335 loop = loop_containing_stmt (stmt);
336 if (!loop_in_sese_p (loop, region))
337 loop = nest;
339 graphite_find_data_references_in_stmt (nest, loop, stmt, &drs);
342 return new_gimple_bb (bb, drs);
345 /* Returns true if all predecessors of BB, that are not dominated by BB, are
346 marked in MAP. The predecessors dominated by BB are loop latches and will
347 be handled after BB. */
349 static bool
350 all_non_dominated_preds_marked_p (basic_block bb, sbitmap map)
352 edge e;
353 edge_iterator ei;
355 FOR_EACH_EDGE (e, ei, bb->preds)
356 if (!bitmap_bit_p (map, e->src->index)
357 && !dominated_by_p (CDI_DOMINATORS, e->src, bb))
358 return false;
360 return true;
363 /* Compare the depth of two basic_block's P1 and P2. */
365 static int
366 compare_bb_depths (const void *p1, const void *p2)
368 const_basic_block const bb1 = *(const_basic_block const*)p1;
369 const_basic_block const bb2 = *(const_basic_block const*)p2;
370 int d1 = loop_depth (bb1->loop_father);
371 int d2 = loop_depth (bb2->loop_father);
373 if (d1 < d2)
374 return 1;
376 if (d1 > d2)
377 return -1;
379 return 0;
382 /* Sort the basic blocks from DOM such that the first are the ones at
383 a deepest loop level. */
385 static void
386 graphite_sort_dominated_info (vec<basic_block> dom)
388 dom.qsort (compare_bb_depths);
391 /* Recursive helper function for build_scops_bbs. */
393 static void
394 build_scop_bbs_1 (scop_p scop, sbitmap visited, basic_block bb)
396 sese region = SCOP_REGION (scop);
397 vec<basic_block> dom;
398 poly_bb_p pbb;
400 if (bitmap_bit_p (visited, bb->index)
401 || !bb_in_sese_p (bb, region))
402 return;
404 pbb = new_poly_bb (scop, try_generate_gimple_bb (scop, bb));
405 SCOP_BBS (scop).safe_push (pbb);
406 bitmap_set_bit (visited, bb->index);
408 dom = get_dominated_by (CDI_DOMINATORS, bb);
410 if (!dom.exists ())
411 return;
413 graphite_sort_dominated_info (dom);
415 while (!dom.is_empty ())
417 int i;
418 basic_block dom_bb;
420 FOR_EACH_VEC_ELT (dom, i, dom_bb)
421 if (all_non_dominated_preds_marked_p (dom_bb, visited))
423 build_scop_bbs_1 (scop, visited, dom_bb);
424 dom.unordered_remove (i);
425 break;
429 dom.release ();
432 /* Gather the basic blocks belonging to the SCOP. */
434 static void
435 build_scop_bbs (scop_p scop)
437 sbitmap visited = sbitmap_alloc (last_basic_block_for_fn (cfun));
438 sese region = SCOP_REGION (scop);
440 bitmap_clear (visited);
441 build_scop_bbs_1 (scop, visited, SESE_ENTRY_BB (region));
442 sbitmap_free (visited);
445 /* Return an ISL identifier for the polyhedral basic block PBB. */
447 static isl_id *
448 isl_id_for_pbb (scop_p s, poly_bb_p pbb)
450 char name[50];
451 snprintf (name, sizeof (name), "S_%d", pbb_index (pbb));
452 return isl_id_alloc (s->ctx, name, pbb);
455 /* Converts the STATIC_SCHEDULE of PBB into a scattering polyhedron.
456 We generate SCATTERING_DIMENSIONS scattering dimensions.
458 CLooG 0.15.0 and previous versions require, that all
459 scattering functions of one CloogProgram have the same number of
460 scattering dimensions, therefore we allow to specify it. This
461 should be removed in future versions of CLooG.
463 The scattering polyhedron consists of these dimensions: scattering,
464 loop_iterators, parameters.
466 Example:
468 | scattering_dimensions = 5
469 | used_scattering_dimensions = 3
470 | nb_iterators = 1
471 | scop_nb_params = 2
473 | Schedule:
475 | 4 5
477 | Scattering polyhedron:
479 | scattering: {s1, s2, s3, s4, s5}
480 | loop_iterators: {i}
481 | parameters: {p1, p2}
483 | s1 s2 s3 s4 s5 i p1 p2 1
484 | 1 0 0 0 0 0 0 0 -4 = 0
485 | 0 1 0 0 0 -1 0 0 0 = 0
486 | 0 0 1 0 0 0 0 0 -5 = 0 */
488 static void
489 build_pbb_scattering_polyhedrons (isl_aff *static_sched,
490 poly_bb_p pbb, int scattering_dimensions)
492 int i;
493 int nb_iterators = pbb_dim_iter_domain (pbb);
494 int used_scattering_dimensions = nb_iterators * 2 + 1;
495 isl_val *val;
496 isl_space *dc, *dm;
498 gcc_assert (scattering_dimensions >= used_scattering_dimensions);
500 dc = isl_set_get_space (pbb->domain);
501 dm = isl_space_add_dims (isl_space_from_domain (dc),
502 isl_dim_out, scattering_dimensions);
503 pbb->schedule = isl_map_universe (dm);
505 for (i = 0; i < scattering_dimensions; i++)
507 /* Textual order inside this loop. */
508 if ((i % 2) == 0)
510 isl_constraint *c = isl_equality_alloc
511 (isl_local_space_from_space (isl_map_get_space (pbb->schedule)));
513 val = isl_aff_get_coefficient_val (static_sched, isl_dim_in, i / 2);
515 val = isl_val_neg (val);
516 c = isl_constraint_set_constant_val (c, val);
517 c = isl_constraint_set_coefficient_si (c, isl_dim_out, i, 1);
518 pbb->schedule = isl_map_add_constraint (pbb->schedule, c);
521 /* Iterations of this loop. */
522 else /* if ((i % 2) == 1) */
524 int loop = (i - 1) / 2;
525 pbb->schedule = isl_map_equate (pbb->schedule, isl_dim_in, loop,
526 isl_dim_out, i);
530 pbb->transformed = isl_map_copy (pbb->schedule);
533 /* Build for BB the static schedule.
535 The static schedule is a Dewey numbering of the abstract syntax
536 tree: http://en.wikipedia.org/wiki/Dewey_Decimal_Classification
538 The following example informally defines the static schedule:
541 for (i: ...)
543 for (j: ...)
549 for (k: ...)
557 Static schedules for A to F:
559 DEPTH
560 0 1 2
562 B 1 0 0
563 C 1 0 1
564 D 1 1 0
565 E 1 1 1
569 static void
570 build_scop_scattering (scop_p scop)
572 int i;
573 poly_bb_p pbb;
574 gimple_bb_p previous_gbb = NULL;
575 isl_space *dc = isl_set_get_space (scop->context);
576 isl_aff *static_sched;
578 dc = isl_space_add_dims (dc, isl_dim_set, number_of_loops (cfun));
579 static_sched = isl_aff_zero_on_domain (isl_local_space_from_space (dc));
581 /* We have to start schedules at 0 on the first component and
582 because we cannot compare_prefix_loops against a previous loop,
583 prefix will be equal to zero, and that index will be
584 incremented before copying. */
585 static_sched = isl_aff_add_coefficient_si (static_sched, isl_dim_in, 0, -1);
587 FOR_EACH_VEC_ELT (SCOP_BBS (scop), i, pbb)
589 gimple_bb_p gbb = PBB_BLACK_BOX (pbb);
590 int prefix;
591 int nb_scat_dims = pbb_dim_iter_domain (pbb) * 2 + 1;
593 if (previous_gbb)
594 prefix = nb_common_loops (SCOP_REGION (scop), previous_gbb, gbb);
595 else
596 prefix = 0;
598 previous_gbb = gbb;
600 static_sched = isl_aff_add_coefficient_si (static_sched, isl_dim_in,
601 prefix, 1);
602 build_pbb_scattering_polyhedrons (static_sched, pbb, nb_scat_dims);
605 isl_aff_free (static_sched);
608 static isl_pw_aff *extract_affine (scop_p, tree, __isl_take isl_space *space);
610 /* Extract an affine expression from the chain of recurrence E. */
612 static isl_pw_aff *
613 extract_affine_chrec (scop_p s, tree e, __isl_take isl_space *space)
615 isl_pw_aff *lhs = extract_affine (s, CHREC_LEFT (e), isl_space_copy (space));
616 isl_pw_aff *rhs = extract_affine (s, CHREC_RIGHT (e), isl_space_copy (space));
617 isl_local_space *ls = isl_local_space_from_space (space);
618 unsigned pos = sese_loop_depth ((sese) s->region, get_chrec_loop (e)) - 1;
619 isl_aff *loop = isl_aff_set_coefficient_si
620 (isl_aff_zero_on_domain (ls), isl_dim_in, pos, 1);
621 isl_pw_aff *l = isl_pw_aff_from_aff (loop);
623 /* Before multiplying, make sure that the result is affine. */
624 gcc_assert (isl_pw_aff_is_cst (rhs)
625 || isl_pw_aff_is_cst (l));
627 return isl_pw_aff_add (lhs, isl_pw_aff_mul (rhs, l));
630 /* Extract an affine expression from the mult_expr E. */
632 static isl_pw_aff *
633 extract_affine_mul (scop_p s, tree e, __isl_take isl_space *space)
635 isl_pw_aff *lhs = extract_affine (s, TREE_OPERAND (e, 0),
636 isl_space_copy (space));
637 isl_pw_aff *rhs = extract_affine (s, TREE_OPERAND (e, 1), space);
639 if (!isl_pw_aff_is_cst (lhs)
640 && !isl_pw_aff_is_cst (rhs))
642 isl_pw_aff_free (lhs);
643 isl_pw_aff_free (rhs);
644 return NULL;
647 return isl_pw_aff_mul (lhs, rhs);
650 /* Return an ISL identifier from the name of the ssa_name E. */
652 static isl_id *
653 isl_id_for_ssa_name (scop_p s, tree e)
655 const char *name = get_name (e);
656 isl_id *id;
658 if (name)
659 id = isl_id_alloc (s->ctx, name, e);
660 else
662 char name1[50];
663 snprintf (name1, sizeof (name1), "P_%d", SSA_NAME_VERSION (e));
664 id = isl_id_alloc (s->ctx, name1, e);
667 return id;
670 /* Return an ISL identifier for the data reference DR. */
672 static isl_id *
673 isl_id_for_dr (scop_p s, data_reference_p dr ATTRIBUTE_UNUSED)
675 /* Data references all get the same isl_id. They need to be comparable
676 and are distinguished through the first dimension, which contains the
677 alias set number. */
678 return isl_id_alloc (s->ctx, "", 0);
681 /* Extract an affine expression from the ssa_name E. */
683 static isl_pw_aff *
684 extract_affine_name (scop_p s, tree e, __isl_take isl_space *space)
686 isl_aff *aff;
687 isl_set *dom;
688 isl_id *id;
689 int dimension;
691 id = isl_id_for_ssa_name (s, e);
692 dimension = isl_space_find_dim_by_id (space, isl_dim_param, id);
693 isl_id_free (id);
694 dom = isl_set_universe (isl_space_copy (space));
695 aff = isl_aff_zero_on_domain (isl_local_space_from_space (space));
696 aff = isl_aff_add_coefficient_si (aff, isl_dim_param, dimension, 1);
697 return isl_pw_aff_alloc (dom, aff);
700 /* Extract an affine expression from the gmp constant G. */
702 static isl_pw_aff *
703 extract_affine_gmp (mpz_t g, __isl_take isl_space *space)
705 isl_local_space *ls = isl_local_space_from_space (isl_space_copy (space));
706 isl_aff *aff = isl_aff_zero_on_domain (ls);
707 isl_set *dom = isl_set_universe (space);
708 isl_val *v;
709 isl_ctx *ct;
711 ct = isl_aff_get_ctx (aff);
712 v = isl_val_int_from_gmp (ct, g);
713 aff = isl_aff_add_constant_val (aff, v);
715 return isl_pw_aff_alloc (dom, aff);
718 /* Extract an affine expression from the integer_cst E. */
720 static isl_pw_aff *
721 extract_affine_int (tree e, __isl_take isl_space *space)
723 isl_pw_aff *res;
724 mpz_t g;
726 mpz_init (g);
727 tree_int_to_gmp (e, g);
728 res = extract_affine_gmp (g, space);
729 mpz_clear (g);
731 return res;
734 /* Compute pwaff mod 2^width. */
736 extern isl_ctx *the_isl_ctx;
738 static isl_pw_aff *
739 wrap (isl_pw_aff *pwaff, unsigned width)
741 isl_val *mod;
743 mod = isl_val_int_from_ui(the_isl_ctx, width);
744 mod = isl_val_2exp (mod);
745 pwaff = isl_pw_aff_mod_val (pwaff, mod);
747 return pwaff;
750 /* When parameter NAME is in REGION, returns its index in SESE_PARAMS.
751 Otherwise returns -1. */
753 static inline int
754 parameter_index_in_region_1 (tree name, sese region)
756 int i;
757 tree p;
759 gcc_assert (TREE_CODE (name) == SSA_NAME);
761 FOR_EACH_VEC_ELT (SESE_PARAMS (region), i, p)
762 if (p == name)
763 return i;
765 return -1;
768 /* When the parameter NAME is in REGION, returns its index in
769 SESE_PARAMS. Otherwise this function inserts NAME in SESE_PARAMS
770 and returns the index of NAME. */
772 static int
773 parameter_index_in_region (tree name, sese region)
775 int i;
777 gcc_assert (TREE_CODE (name) == SSA_NAME);
779 i = parameter_index_in_region_1 (name, region);
780 if (i != -1)
781 return i;
783 gcc_assert (SESE_ADD_PARAMS (region));
785 i = SESE_PARAMS (region).length ();
786 SESE_PARAMS (region).safe_push (name);
787 return i;
790 /* Extract an affine expression from the tree E in the scop S. */
792 static isl_pw_aff *
793 extract_affine (scop_p s, tree e, __isl_take isl_space *space)
795 isl_pw_aff *lhs, *rhs, *res;
796 tree type;
798 if (e == chrec_dont_know) {
799 isl_space_free (space);
800 return NULL;
803 switch (TREE_CODE (e))
805 case POLYNOMIAL_CHREC:
806 res = extract_affine_chrec (s, e, space);
807 break;
809 case MULT_EXPR:
810 res = extract_affine_mul (s, e, space);
811 break;
813 case PLUS_EXPR:
814 case POINTER_PLUS_EXPR:
815 lhs = extract_affine (s, TREE_OPERAND (e, 0), isl_space_copy (space));
816 rhs = extract_affine (s, TREE_OPERAND (e, 1), space);
817 res = isl_pw_aff_add (lhs, rhs);
818 break;
820 case MINUS_EXPR:
821 lhs = extract_affine (s, TREE_OPERAND (e, 0), isl_space_copy (space));
822 rhs = extract_affine (s, TREE_OPERAND (e, 1), space);
823 res = isl_pw_aff_sub (lhs, rhs);
824 break;
826 case NEGATE_EXPR:
827 case BIT_NOT_EXPR:
828 lhs = extract_affine (s, TREE_OPERAND (e, 0), isl_space_copy (space));
829 rhs = extract_affine (s, integer_minus_one_node, space);
830 res = isl_pw_aff_mul (lhs, rhs);
831 break;
833 case SSA_NAME:
834 gcc_assert (-1 != parameter_index_in_region_1 (e, SCOP_REGION (s)));
835 res = extract_affine_name (s, e, space);
836 break;
838 case INTEGER_CST:
839 res = extract_affine_int (e, space);
840 /* No need to wrap a single integer. */
841 return res;
843 CASE_CONVERT:
844 case NON_LVALUE_EXPR:
845 res = extract_affine (s, TREE_OPERAND (e, 0), space);
846 break;
848 default:
849 gcc_unreachable ();
850 break;
853 type = TREE_TYPE (e);
854 if (TYPE_UNSIGNED (type))
855 res = wrap (res, TYPE_PRECISION (type));
857 return res;
860 /* In the context of sese S, scan the expression E and translate it to
861 a linear expression C. When parsing a symbolic multiplication, K
862 represents the constant multiplier of an expression containing
863 parameters. */
865 static void
866 scan_tree_for_params (sese s, tree e)
868 if (e == chrec_dont_know)
869 return;
871 switch (TREE_CODE (e))
873 case POLYNOMIAL_CHREC:
874 scan_tree_for_params (s, CHREC_LEFT (e));
875 break;
877 case MULT_EXPR:
878 if (chrec_contains_symbols (TREE_OPERAND (e, 0)))
879 scan_tree_for_params (s, TREE_OPERAND (e, 0));
880 else
881 scan_tree_for_params (s, TREE_OPERAND (e, 1));
882 break;
884 case PLUS_EXPR:
885 case POINTER_PLUS_EXPR:
886 case MINUS_EXPR:
887 scan_tree_for_params (s, TREE_OPERAND (e, 0));
888 scan_tree_for_params (s, TREE_OPERAND (e, 1));
889 break;
891 case NEGATE_EXPR:
892 case BIT_NOT_EXPR:
893 CASE_CONVERT:
894 case NON_LVALUE_EXPR:
895 scan_tree_for_params (s, TREE_OPERAND (e, 0));
896 break;
898 case SSA_NAME:
899 parameter_index_in_region (e, s);
900 break;
902 case INTEGER_CST:
903 case ADDR_EXPR:
904 break;
906 default:
907 gcc_unreachable ();
908 break;
912 /* Find parameters with respect to REGION in BB. We are looking in memory
913 access functions, conditions and loop bounds. */
915 static void
916 find_params_in_bb (sese region, gimple_bb_p gbb)
918 int i;
919 unsigned j;
920 data_reference_p dr;
921 gimple stmt;
922 loop_p loop = GBB_BB (gbb)->loop_father;
924 /* Find parameters in the access functions of data references. */
925 FOR_EACH_VEC_ELT (GBB_DATA_REFS (gbb), i, dr)
926 for (j = 0; j < DR_NUM_DIMENSIONS (dr); j++)
927 scan_tree_for_params (region, DR_ACCESS_FN (dr, j));
929 /* Find parameters in conditional statements. */
930 FOR_EACH_VEC_ELT (GBB_CONDITIONS (gbb), i, stmt)
932 tree lhs = scalar_evolution_in_region (region, loop,
933 gimple_cond_lhs (stmt));
934 tree rhs = scalar_evolution_in_region (region, loop,
935 gimple_cond_rhs (stmt));
937 scan_tree_for_params (region, lhs);
938 scan_tree_for_params (region, rhs);
942 /* Record the parameters used in the SCOP. A variable is a parameter
943 in a scop if it does not vary during the execution of that scop. */
945 static void
946 find_scop_parameters (scop_p scop)
948 poly_bb_p pbb;
949 unsigned i;
950 sese region = SCOP_REGION (scop);
951 struct loop *loop;
952 int nbp;
954 /* Find the parameters used in the loop bounds. */
955 FOR_EACH_VEC_ELT (SESE_LOOP_NEST (region), i, loop)
957 tree nb_iters = number_of_latch_executions (loop);
959 if (!chrec_contains_symbols (nb_iters))
960 continue;
962 nb_iters = scalar_evolution_in_region (region, loop, nb_iters);
963 scan_tree_for_params (region, nb_iters);
966 /* Find the parameters used in data accesses. */
967 FOR_EACH_VEC_ELT (SCOP_BBS (scop), i, pbb)
968 find_params_in_bb (region, PBB_BLACK_BOX (pbb));
970 nbp = sese_nb_params (region);
971 scop_set_nb_params (scop, nbp);
972 SESE_ADD_PARAMS (region) = false;
975 tree e;
976 isl_space *space = isl_space_set_alloc (scop->ctx, nbp, 0);
978 FOR_EACH_VEC_ELT (SESE_PARAMS (region), i, e)
979 space = isl_space_set_dim_id (space, isl_dim_param, i,
980 isl_id_for_ssa_name (scop, e));
982 scop->context = isl_set_universe (space);
986 /* Builds the constraint polyhedra for LOOP in SCOP. OUTER_PH gives
987 the constraints for the surrounding loops. */
989 static void
990 build_loop_iteration_domains (scop_p scop, struct loop *loop,
991 int nb,
992 isl_set *outer, isl_set **doms)
994 tree nb_iters = number_of_latch_executions (loop);
995 sese region = SCOP_REGION (scop);
997 isl_set *inner = isl_set_copy (outer);
998 isl_space *space;
999 isl_constraint *c;
1000 int pos = isl_set_dim (outer, isl_dim_set);
1001 isl_val *v;
1002 mpz_t g;
1004 mpz_init (g);
1006 inner = isl_set_add_dims (inner, isl_dim_set, 1);
1007 space = isl_set_get_space (inner);
1009 /* 0 <= loop_i */
1010 c = isl_inequality_alloc
1011 (isl_local_space_from_space (isl_space_copy (space)));
1012 c = isl_constraint_set_coefficient_si (c, isl_dim_set, pos, 1);
1013 inner = isl_set_add_constraint (inner, c);
1015 /* loop_i <= cst_nb_iters */
1016 if (TREE_CODE (nb_iters) == INTEGER_CST)
1018 c = isl_inequality_alloc
1019 (isl_local_space_from_space (isl_space_copy (space)));
1020 c = isl_constraint_set_coefficient_si (c, isl_dim_set, pos, -1);
1021 tree_int_to_gmp (nb_iters, g);
1022 v = isl_val_int_from_gmp (the_isl_ctx, g);
1023 c = isl_constraint_set_constant_val (c, v);
1024 inner = isl_set_add_constraint (inner, c);
1027 /* loop_i <= expr_nb_iters */
1028 else if (!chrec_contains_undetermined (nb_iters))
1030 widest_int nit;
1031 isl_pw_aff *aff;
1032 isl_set *valid;
1033 isl_local_space *ls;
1034 isl_aff *al;
1035 isl_set *le;
1037 nb_iters = scalar_evolution_in_region (region, loop, nb_iters);
1039 aff = extract_affine (scop, nb_iters, isl_set_get_space (inner));
1040 valid = isl_pw_aff_nonneg_set (isl_pw_aff_copy (aff));
1041 valid = isl_set_project_out (valid, isl_dim_set, 0,
1042 isl_set_dim (valid, isl_dim_set));
1043 scop->context = isl_set_intersect (scop->context, valid);
1045 ls = isl_local_space_from_space (isl_space_copy (space));
1046 al = isl_aff_set_coefficient_si (isl_aff_zero_on_domain (ls),
1047 isl_dim_in, pos, 1);
1048 le = isl_pw_aff_le_set (isl_pw_aff_from_aff (al),
1049 isl_pw_aff_copy (aff));
1050 inner = isl_set_intersect (inner, le);
1052 if (max_stmt_executions (loop, &nit))
1054 /* Insert in the context the constraints from the
1055 estimation of the number of iterations NIT and the
1056 symbolic number of iterations (involving parameter
1057 names) NB_ITERS. First, build the affine expression
1058 "NIT - NB_ITERS" and then say that it is positive,
1059 i.e., NIT approximates NB_ITERS: "NIT >= NB_ITERS". */
1060 isl_pw_aff *approx;
1061 mpz_t g;
1062 isl_set *x;
1063 isl_constraint *c;
1065 mpz_init (g);
1066 wi::to_mpz (nit, g, SIGNED);
1067 mpz_sub_ui (g, g, 1);
1068 approx = extract_affine_gmp (g, isl_set_get_space (inner));
1069 x = isl_pw_aff_ge_set (approx, aff);
1070 x = isl_set_project_out (x, isl_dim_set, 0,
1071 isl_set_dim (x, isl_dim_set));
1072 scop->context = isl_set_intersect (scop->context, x);
1074 c = isl_inequality_alloc
1075 (isl_local_space_from_space (isl_space_copy (space)));
1076 c = isl_constraint_set_coefficient_si (c, isl_dim_set, pos, -1);
1077 v = isl_val_int_from_gmp (the_isl_ctx, g);
1078 mpz_clear (g);
1079 c = isl_constraint_set_constant_val (c, v);
1080 inner = isl_set_add_constraint (inner, c);
1082 else
1083 isl_pw_aff_free (aff);
1085 else
1086 gcc_unreachable ();
1088 if (loop->inner && loop_in_sese_p (loop->inner, region))
1089 build_loop_iteration_domains (scop, loop->inner, nb + 1,
1090 isl_set_copy (inner), doms);
1092 if (nb != 0
1093 && loop->next
1094 && loop_in_sese_p (loop->next, region))
1095 build_loop_iteration_domains (scop, loop->next, nb,
1096 isl_set_copy (outer), doms);
1098 doms[loop->num] = inner;
1100 isl_set_free (outer);
1101 isl_space_free (space);
1102 mpz_clear (g);
1105 /* Returns a linear expression for tree T evaluated in PBB. */
1107 static isl_pw_aff *
1108 create_pw_aff_from_tree (poly_bb_p pbb, tree t)
1110 scop_p scop = PBB_SCOP (pbb);
1112 t = scalar_evolution_in_region (SCOP_REGION (scop), pbb_loop (pbb), t);
1113 gcc_assert (!automatically_generated_chrec_p (t));
1115 return extract_affine (scop, t, isl_set_get_space (pbb->domain));
1118 /* Add conditional statement STMT to pbb. CODE is used as the comparison
1119 operator. This allows us to invert the condition or to handle
1120 inequalities. */
1122 static void
1123 add_condition_to_pbb (poly_bb_p pbb, gimple stmt, enum tree_code code)
1125 isl_pw_aff *lhs = create_pw_aff_from_tree (pbb, gimple_cond_lhs (stmt));
1126 isl_pw_aff *rhs = create_pw_aff_from_tree (pbb, gimple_cond_rhs (stmt));
1127 isl_set *cond;
1129 switch (code)
1131 case LT_EXPR:
1132 cond = isl_pw_aff_lt_set (lhs, rhs);
1133 break;
1135 case GT_EXPR:
1136 cond = isl_pw_aff_gt_set (lhs, rhs);
1137 break;
1139 case LE_EXPR:
1140 cond = isl_pw_aff_le_set (lhs, rhs);
1141 break;
1143 case GE_EXPR:
1144 cond = isl_pw_aff_ge_set (lhs, rhs);
1145 break;
1147 case EQ_EXPR:
1148 cond = isl_pw_aff_eq_set (lhs, rhs);
1149 break;
1151 case NE_EXPR:
1152 cond = isl_pw_aff_ne_set (lhs, rhs);
1153 break;
1155 default:
1156 isl_pw_aff_free (lhs);
1157 isl_pw_aff_free (rhs);
1158 return;
1161 cond = isl_set_coalesce (cond);
1162 cond = isl_set_set_tuple_id (cond, isl_set_get_tuple_id (pbb->domain));
1163 pbb->domain = isl_set_intersect (pbb->domain, cond);
1166 /* Add conditions to the domain of PBB. */
1168 static void
1169 add_conditions_to_domain (poly_bb_p pbb)
1171 unsigned int i;
1172 gimple stmt;
1173 gimple_bb_p gbb = PBB_BLACK_BOX (pbb);
1175 if (GBB_CONDITIONS (gbb).is_empty ())
1176 return;
1178 FOR_EACH_VEC_ELT (GBB_CONDITIONS (gbb), i, stmt)
1179 switch (gimple_code (stmt))
1181 case GIMPLE_COND:
1183 enum tree_code code = gimple_cond_code (stmt);
1185 /* The conditions for ELSE-branches are inverted. */
1186 if (!GBB_CONDITION_CASES (gbb)[i])
1187 code = invert_tree_comparison (code, false);
1189 add_condition_to_pbb (pbb, stmt, code);
1190 break;
1193 case GIMPLE_SWITCH:
1194 /* Switch statements are not supported right now - fall through. */
1196 default:
1197 gcc_unreachable ();
1198 break;
1202 /* Traverses all the GBBs of the SCOP and add their constraints to the
1203 iteration domains. */
1205 static void
1206 add_conditions_to_constraints (scop_p scop)
1208 int i;
1209 poly_bb_p pbb;
1211 FOR_EACH_VEC_ELT (SCOP_BBS (scop), i, pbb)
1212 add_conditions_to_domain (pbb);
1215 /* Returns a COND_EXPR statement when BB has a single predecessor, the
1216 edge between BB and its predecessor is not a loop exit edge, and
1217 the last statement of the single predecessor is a COND_EXPR. */
1219 static gimple
1220 single_pred_cond_non_loop_exit (basic_block bb)
1222 if (single_pred_p (bb))
1224 edge e = single_pred_edge (bb);
1225 basic_block pred = e->src;
1226 gimple stmt;
1228 if (loop_depth (pred->loop_father) > loop_depth (bb->loop_father))
1229 return NULL;
1231 stmt = last_stmt (pred);
1233 if (stmt && gimple_code (stmt) == GIMPLE_COND)
1234 return stmt;
1237 return NULL;
1240 class sese_dom_walker : public dom_walker
1242 public:
1243 sese_dom_walker (cdi_direction, sese);
1245 virtual void before_dom_children (basic_block);
1246 virtual void after_dom_children (basic_block);
1248 private:
1249 auto_vec<gimple, 3> m_conditions, m_cases;
1250 sese m_region;
1253 sese_dom_walker::sese_dom_walker (cdi_direction direction, sese region)
1254 : dom_walker (direction), m_region (region)
1258 /* Call-back for dom_walk executed before visiting the dominated
1259 blocks. */
1261 void
1262 sese_dom_walker::before_dom_children (basic_block bb)
1264 gimple_bb_p gbb;
1265 gimple stmt;
1267 if (!bb_in_sese_p (bb, m_region))
1268 return;
1270 stmt = single_pred_cond_non_loop_exit (bb);
1272 if (stmt)
1274 edge e = single_pred_edge (bb);
1276 m_conditions.safe_push (stmt);
1278 if (e->flags & EDGE_TRUE_VALUE)
1279 m_cases.safe_push (stmt);
1280 else
1281 m_cases.safe_push (NULL);
1284 gbb = gbb_from_bb (bb);
1286 if (gbb)
1288 GBB_CONDITIONS (gbb) = m_conditions.copy ();
1289 GBB_CONDITION_CASES (gbb) = m_cases.copy ();
1293 /* Call-back for dom_walk executed after visiting the dominated
1294 blocks. */
1296 void
1297 sese_dom_walker::after_dom_children (basic_block bb)
1299 if (!bb_in_sese_p (bb, m_region))
1300 return;
1302 if (single_pred_cond_non_loop_exit (bb))
1304 m_conditions.pop ();
1305 m_cases.pop ();
1309 /* Add constraints on the possible values of parameter P from the type
1310 of P. */
1312 static void
1313 add_param_constraints (scop_p scop, graphite_dim_t p)
1315 tree parameter = SESE_PARAMS (SCOP_REGION (scop))[p];
1316 tree type = TREE_TYPE (parameter);
1317 tree lb = NULL_TREE;
1318 tree ub = NULL_TREE;
1320 if (POINTER_TYPE_P (type) || !TYPE_MIN_VALUE (type))
1321 lb = lower_bound_in_type (type, type);
1322 else
1323 lb = TYPE_MIN_VALUE (type);
1325 if (POINTER_TYPE_P (type) || !TYPE_MAX_VALUE (type))
1326 ub = upper_bound_in_type (type, type);
1327 else
1328 ub = TYPE_MAX_VALUE (type);
1330 if (lb)
1332 isl_space *space = isl_set_get_space (scop->context);
1333 isl_constraint *c;
1334 mpz_t g;
1335 isl_val *v;
1337 c = isl_inequality_alloc (isl_local_space_from_space (space));
1338 mpz_init (g);
1339 tree_int_to_gmp (lb, g);
1340 v = isl_val_int_from_gmp (the_isl_ctx, g);
1341 v = isl_val_neg (v);
1342 mpz_clear (g);
1343 c = isl_constraint_set_constant_val (c, v);
1344 c = isl_constraint_set_coefficient_si (c, isl_dim_param, p, 1);
1346 scop->context = isl_set_add_constraint (scop->context, c);
1349 if (ub)
1351 isl_space *space = isl_set_get_space (scop->context);
1352 isl_constraint *c;
1353 mpz_t g;
1354 isl_val *v;
1356 c = isl_inequality_alloc (isl_local_space_from_space (space));
1358 mpz_init (g);
1359 tree_int_to_gmp (ub, g);
1360 v = isl_val_int_from_gmp (the_isl_ctx, g);
1361 mpz_clear (g);
1362 c = isl_constraint_set_constant_val (c, v);
1363 c = isl_constraint_set_coefficient_si (c, isl_dim_param, p, -1);
1365 scop->context = isl_set_add_constraint (scop->context, c);
1369 /* Build the context of the SCOP. The context usually contains extra
1370 constraints that are added to the iteration domains that constrain
1371 some parameters. */
1373 static void
1374 build_scop_context (scop_p scop)
1376 graphite_dim_t p, n = scop_nb_params (scop);
1378 for (p = 0; p < n; p++)
1379 add_param_constraints (scop, p);
1382 /* Build the iteration domains: the loops belonging to the current
1383 SCOP, and that vary for the execution of the current basic block.
1384 Returns false if there is no loop in SCOP. */
1386 static void
1387 build_scop_iteration_domain (scop_p scop)
1389 struct loop *loop;
1390 sese region = SCOP_REGION (scop);
1391 int i;
1392 poly_bb_p pbb;
1393 int nb_loops = number_of_loops (cfun);
1394 isl_set **doms = XCNEWVEC (isl_set *, nb_loops);
1396 FOR_EACH_VEC_ELT (SESE_LOOP_NEST (region), i, loop)
1397 if (!loop_in_sese_p (loop_outer (loop), region))
1398 build_loop_iteration_domains (scop, loop, 0,
1399 isl_set_copy (scop->context), doms);
1401 FOR_EACH_VEC_ELT (SCOP_BBS (scop), i, pbb)
1403 loop = pbb_loop (pbb);
1405 if (doms[loop->num])
1406 pbb->domain = isl_set_copy (doms[loop->num]);
1407 else
1408 pbb->domain = isl_set_copy (scop->context);
1410 pbb->domain = isl_set_set_tuple_id (pbb->domain,
1411 isl_id_for_pbb (scop, pbb));
1414 for (i = 0; i < nb_loops; i++)
1415 if (doms[i])
1416 isl_set_free (doms[i]);
1418 free (doms);
1421 /* Add a constrain to the ACCESSES polyhedron for the alias set of
1422 data reference DR. ACCESSP_NB_DIMS is the dimension of the
1423 ACCESSES polyhedron, DOM_NB_DIMS is the dimension of the iteration
1424 domain. */
1426 static isl_map *
1427 pdr_add_alias_set (isl_map *acc, data_reference_p dr)
1429 isl_constraint *c;
1430 int alias_set_num = 0;
1431 base_alias_pair *bap = (base_alias_pair *)(dr->aux);
1433 if (bap && bap->alias_set)
1434 alias_set_num = *(bap->alias_set);
1436 c = isl_equality_alloc
1437 (isl_local_space_from_space (isl_map_get_space (acc)));
1438 c = isl_constraint_set_constant_si (c, -alias_set_num);
1439 c = isl_constraint_set_coefficient_si (c, isl_dim_out, 0, 1);
1441 return isl_map_add_constraint (acc, c);
1444 /* Assign the affine expression INDEX to the output dimension POS of
1445 MAP and return the result. */
1447 static isl_map *
1448 set_index (isl_map *map, int pos, isl_pw_aff *index)
1450 isl_map *index_map;
1451 int len = isl_map_dim (map, isl_dim_out);
1452 isl_id *id;
1454 index_map = isl_map_from_pw_aff (index);
1455 index_map = isl_map_insert_dims (index_map, isl_dim_out, 0, pos);
1456 index_map = isl_map_add_dims (index_map, isl_dim_out, len - pos - 1);
1458 id = isl_map_get_tuple_id (map, isl_dim_out);
1459 index_map = isl_map_set_tuple_id (index_map, isl_dim_out, id);
1460 id = isl_map_get_tuple_id (map, isl_dim_in);
1461 index_map = isl_map_set_tuple_id (index_map, isl_dim_in, id);
1463 return isl_map_intersect (map, index_map);
1466 /* Add to ACCESSES polyhedron equalities defining the access functions
1467 to the memory. ACCESSP_NB_DIMS is the dimension of the ACCESSES
1468 polyhedron, DOM_NB_DIMS is the dimension of the iteration domain.
1469 PBB is the poly_bb_p that contains the data reference DR. */
1471 static isl_map *
1472 pdr_add_memory_accesses (isl_map *acc, data_reference_p dr, poly_bb_p pbb)
1474 int i, nb_subscripts = DR_NUM_DIMENSIONS (dr);
1475 scop_p scop = PBB_SCOP (pbb);
1477 for (i = 0; i < nb_subscripts; i++)
1479 isl_pw_aff *aff;
1480 tree afn = DR_ACCESS_FN (dr, nb_subscripts - 1 - i);
1482 aff = extract_affine (scop, afn,
1483 isl_space_domain (isl_map_get_space (acc)));
1484 acc = set_index (acc, i + 1, aff);
1487 return acc;
1490 /* Add constrains representing the size of the accessed data to the
1491 ACCESSES polyhedron. ACCESSP_NB_DIMS is the dimension of the
1492 ACCESSES polyhedron, DOM_NB_DIMS is the dimension of the iteration
1493 domain. */
1495 static isl_set *
1496 pdr_add_data_dimensions (isl_set *extent, scop_p scop, data_reference_p dr)
1498 tree ref = DR_REF (dr);
1499 int i, nb_subscripts = DR_NUM_DIMENSIONS (dr);
1501 for (i = nb_subscripts - 1; i >= 0; i--, ref = TREE_OPERAND (ref, 0))
1503 tree low, high;
1505 if (TREE_CODE (ref) != ARRAY_REF)
1506 break;
1508 low = array_ref_low_bound (ref);
1509 high = array_ref_up_bound (ref);
1511 /* XXX The PPL code dealt separately with
1512 subscript - low >= 0 and high - subscript >= 0 in case one of
1513 the two bounds isn't known. Do the same here? */
1515 if (tree_fits_shwi_p (low)
1516 && high
1517 && tree_fits_shwi_p (high)
1518 /* 1-element arrays at end of structures may extend over
1519 their declared size. */
1520 && !(array_at_struct_end_p (ref)
1521 && operand_equal_p (low, high, 0)))
1523 isl_id *id;
1524 isl_aff *aff;
1525 isl_set *univ, *lbs, *ubs;
1526 isl_pw_aff *index;
1527 isl_space *space;
1528 isl_set *valid;
1529 isl_pw_aff *lb = extract_affine_int (low, isl_set_get_space (extent));
1530 isl_pw_aff *ub = extract_affine_int (high, isl_set_get_space (extent));
1532 /* high >= 0 */
1533 valid = isl_pw_aff_nonneg_set (isl_pw_aff_copy (ub));
1534 valid = isl_set_project_out (valid, isl_dim_set, 0,
1535 isl_set_dim (valid, isl_dim_set));
1536 scop->context = isl_set_intersect (scop->context, valid);
1538 space = isl_set_get_space (extent);
1539 aff = isl_aff_zero_on_domain (isl_local_space_from_space (space));
1540 aff = isl_aff_add_coefficient_si (aff, isl_dim_in, i + 1, 1);
1541 univ = isl_set_universe (isl_space_domain (isl_aff_get_space (aff)));
1542 index = isl_pw_aff_alloc (univ, aff);
1544 id = isl_set_get_tuple_id (extent);
1545 lb = isl_pw_aff_set_tuple_id (lb, isl_dim_in, isl_id_copy (id));
1546 ub = isl_pw_aff_set_tuple_id (ub, isl_dim_in, id);
1548 /* low <= sub_i <= high */
1549 lbs = isl_pw_aff_ge_set (isl_pw_aff_copy (index), lb);
1550 ubs = isl_pw_aff_le_set (index, ub);
1551 extent = isl_set_intersect (extent, lbs);
1552 extent = isl_set_intersect (extent, ubs);
1556 return extent;
1559 /* Build data accesses for DR in PBB. */
1561 static void
1562 build_poly_dr (data_reference_p dr, poly_bb_p pbb)
1564 int dr_base_object_set;
1565 isl_map *acc;
1566 isl_set *extent;
1567 scop_p scop = PBB_SCOP (pbb);
1570 isl_space *dc = isl_set_get_space (pbb->domain);
1571 int nb_out = 1 + DR_NUM_DIMENSIONS (dr);
1572 isl_space *space = isl_space_add_dims (isl_space_from_domain (dc),
1573 isl_dim_out, nb_out);
1575 acc = isl_map_universe (space);
1576 acc = isl_map_set_tuple_id (acc, isl_dim_out, isl_id_for_dr (scop, dr));
1579 acc = pdr_add_alias_set (acc, dr);
1580 acc = pdr_add_memory_accesses (acc, dr, pbb);
1583 isl_id *id = isl_id_for_dr (scop, dr);
1584 int nb = 1 + DR_NUM_DIMENSIONS (dr);
1585 isl_space *space = isl_space_set_alloc (scop->ctx, 0, nb);
1586 int alias_set_num = 0;
1587 base_alias_pair *bap = (base_alias_pair *)(dr->aux);
1589 if (bap && bap->alias_set)
1590 alias_set_num = *(bap->alias_set);
1592 space = isl_space_set_tuple_id (space, isl_dim_set, id);
1593 extent = isl_set_nat_universe (space);
1594 extent = isl_set_fix_si (extent, isl_dim_set, 0, alias_set_num);
1595 extent = pdr_add_data_dimensions (extent, scop, dr);
1598 gcc_assert (dr->aux);
1599 dr_base_object_set = ((base_alias_pair *)(dr->aux))->base_obj_set;
1601 new_poly_dr (pbb, dr_base_object_set,
1602 DR_IS_READ (dr) ? PDR_READ : PDR_WRITE,
1603 dr, DR_NUM_DIMENSIONS (dr), acc, extent);
1606 /* Write to FILE the alias graph of data references in DIMACS format. */
1608 static inline bool
1609 write_alias_graph_to_ascii_dimacs (FILE *file, char *comment,
1610 vec<data_reference_p> drs)
1612 int num_vertex = drs.length ();
1613 int edge_num = 0;
1614 data_reference_p dr1, dr2;
1615 int i, j;
1617 if (num_vertex == 0)
1618 return true;
1620 FOR_EACH_VEC_ELT (drs, i, dr1)
1621 for (j = i + 1; drs.iterate (j, &dr2); j++)
1622 if (dr_may_alias_p (dr1, dr2, true))
1623 edge_num++;
1625 fprintf (file, "$\n");
1627 if (comment)
1628 fprintf (file, "c %s\n", comment);
1630 fprintf (file, "p edge %d %d\n", num_vertex, edge_num);
1632 FOR_EACH_VEC_ELT (drs, i, dr1)
1633 for (j = i + 1; drs.iterate (j, &dr2); j++)
1634 if (dr_may_alias_p (dr1, dr2, true))
1635 fprintf (file, "e %d %d\n", i + 1, j + 1);
1637 return true;
1640 /* Write to FILE the alias graph of data references in DOT format. */
1642 static inline bool
1643 write_alias_graph_to_ascii_dot (FILE *file, char *comment,
1644 vec<data_reference_p> drs)
1646 int num_vertex = drs.length ();
1647 data_reference_p dr1, dr2;
1648 int i, j;
1650 if (num_vertex == 0)
1651 return true;
1653 fprintf (file, "$\n");
1655 if (comment)
1656 fprintf (file, "c %s\n", comment);
1658 /* First print all the vertices. */
1659 FOR_EACH_VEC_ELT (drs, i, dr1)
1660 fprintf (file, "n%d;\n", i);
1662 FOR_EACH_VEC_ELT (drs, i, dr1)
1663 for (j = i + 1; drs.iterate (j, &dr2); j++)
1664 if (dr_may_alias_p (dr1, dr2, true))
1665 fprintf (file, "n%d n%d\n", i, j);
1667 return true;
1670 /* Write to FILE the alias graph of data references in ECC format. */
1672 static inline bool
1673 write_alias_graph_to_ascii_ecc (FILE *file, char *comment,
1674 vec<data_reference_p> drs)
1676 int num_vertex = drs.length ();
1677 data_reference_p dr1, dr2;
1678 int i, j;
1680 if (num_vertex == 0)
1681 return true;
1683 fprintf (file, "$\n");
1685 if (comment)
1686 fprintf (file, "c %s\n", comment);
1688 FOR_EACH_VEC_ELT (drs, i, dr1)
1689 for (j = i + 1; drs.iterate (j, &dr2); j++)
1690 if (dr_may_alias_p (dr1, dr2, true))
1691 fprintf (file, "%d %d\n", i, j);
1693 return true;
1696 /* Check if DR1 and DR2 are in the same object set. */
1698 static bool
1699 dr_same_base_object_p (const struct data_reference *dr1,
1700 const struct data_reference *dr2)
1702 return operand_equal_p (DR_BASE_OBJECT (dr1), DR_BASE_OBJECT (dr2), 0);
1705 /* Uses DFS component number as representative of alias-sets. Also tests for
1706 optimality by verifying if every connected component is a clique. Returns
1707 true (1) if the above test is true, and false (0) otherwise. */
1709 static int
1710 build_alias_set_optimal_p (vec<data_reference_p> drs)
1712 int num_vertices = drs.length ();
1713 struct graph *g = new_graph (num_vertices);
1714 data_reference_p dr1, dr2;
1715 int i, j;
1716 int num_connected_components;
1717 int v_indx1, v_indx2, num_vertices_in_component;
1718 int *all_vertices;
1719 int *vertices;
1720 struct graph_edge *e;
1721 int this_component_is_clique;
1722 int all_components_are_cliques = 1;
1724 FOR_EACH_VEC_ELT (drs, i, dr1)
1725 for (j = i+1; drs.iterate (j, &dr2); j++)
1726 if (dr_may_alias_p (dr1, dr2, true))
1728 add_edge (g, i, j);
1729 add_edge (g, j, i);
1732 all_vertices = XNEWVEC (int, num_vertices);
1733 vertices = XNEWVEC (int, num_vertices);
1734 for (i = 0; i < num_vertices; i++)
1735 all_vertices[i] = i;
1737 num_connected_components = graphds_dfs (g, all_vertices, num_vertices,
1738 NULL, true, NULL);
1739 for (i = 0; i < g->n_vertices; i++)
1741 data_reference_p dr = drs[i];
1742 base_alias_pair *bap;
1744 gcc_assert (dr->aux);
1745 bap = (base_alias_pair *)(dr->aux);
1747 bap->alias_set = XNEW (int);
1748 *(bap->alias_set) = g->vertices[i].component + 1;
1751 /* Verify if the DFS numbering results in optimal solution. */
1752 for (i = 0; i < num_connected_components; i++)
1754 num_vertices_in_component = 0;
1755 /* Get all vertices whose DFS component number is the same as i. */
1756 for (j = 0; j < num_vertices; j++)
1757 if (g->vertices[j].component == i)
1758 vertices[num_vertices_in_component++] = j;
1760 /* Now test if the vertices in 'vertices' form a clique, by testing
1761 for edges among each pair. */
1762 this_component_is_clique = 1;
1763 for (v_indx1 = 0; v_indx1 < num_vertices_in_component; v_indx1++)
1765 for (v_indx2 = v_indx1+1; v_indx2 < num_vertices_in_component; v_indx2++)
1767 /* Check if the two vertices are connected by iterating
1768 through all the edges which have one of these are source. */
1769 e = g->vertices[vertices[v_indx2]].pred;
1770 while (e)
1772 if (e->src == vertices[v_indx1])
1773 break;
1774 e = e->pred_next;
1776 if (!e)
1778 this_component_is_clique = 0;
1779 break;
1782 if (!this_component_is_clique)
1783 all_components_are_cliques = 0;
1787 free (all_vertices);
1788 free (vertices);
1789 free_graph (g);
1790 return all_components_are_cliques;
1793 /* Group each data reference in DRS with its base object set num. */
1795 static void
1796 build_base_obj_set_for_drs (vec<data_reference_p> drs)
1798 int num_vertex = drs.length ();
1799 struct graph *g = new_graph (num_vertex);
1800 data_reference_p dr1, dr2;
1801 int i, j;
1802 int *queue;
1804 FOR_EACH_VEC_ELT (drs, i, dr1)
1805 for (j = i + 1; drs.iterate (j, &dr2); j++)
1806 if (dr_same_base_object_p (dr1, dr2))
1808 add_edge (g, i, j);
1809 add_edge (g, j, i);
1812 queue = XNEWVEC (int, num_vertex);
1813 for (i = 0; i < num_vertex; i++)
1814 queue[i] = i;
1816 graphds_dfs (g, queue, num_vertex, NULL, true, NULL);
1818 for (i = 0; i < g->n_vertices; i++)
1820 data_reference_p dr = drs[i];
1821 base_alias_pair *bap;
1823 gcc_assert (dr->aux);
1824 bap = (base_alias_pair *)(dr->aux);
1826 bap->base_obj_set = g->vertices[i].component + 1;
1829 free (queue);
1830 free_graph (g);
1833 /* Build the data references for PBB. */
1835 static void
1836 build_pbb_drs (poly_bb_p pbb)
1838 int j;
1839 data_reference_p dr;
1840 vec<data_reference_p> gbb_drs = GBB_DATA_REFS (PBB_BLACK_BOX (pbb));
1842 FOR_EACH_VEC_ELT (gbb_drs, j, dr)
1843 build_poly_dr (dr, pbb);
1846 /* Dump to file the alias graphs for the data references in DRS. */
1848 static void
1849 dump_alias_graphs (vec<data_reference_p> drs)
1851 char comment[100];
1852 FILE *file_dimacs, *file_ecc, *file_dot;
1854 file_dimacs = fopen ("/tmp/dr_alias_graph_dimacs", "ab");
1855 if (file_dimacs)
1857 snprintf (comment, sizeof (comment), "%s %s", main_input_filename,
1858 current_function_name ());
1859 write_alias_graph_to_ascii_dimacs (file_dimacs, comment, drs);
1860 fclose (file_dimacs);
1863 file_ecc = fopen ("/tmp/dr_alias_graph_ecc", "ab");
1864 if (file_ecc)
1866 snprintf (comment, sizeof (comment), "%s %s", main_input_filename,
1867 current_function_name ());
1868 write_alias_graph_to_ascii_ecc (file_ecc, comment, drs);
1869 fclose (file_ecc);
1872 file_dot = fopen ("/tmp/dr_alias_graph_dot", "ab");
1873 if (file_dot)
1875 snprintf (comment, sizeof (comment), "%s %s", main_input_filename,
1876 current_function_name ());
1877 write_alias_graph_to_ascii_dot (file_dot, comment, drs);
1878 fclose (file_dot);
1882 /* Build data references in SCOP. */
1884 static void
1885 build_scop_drs (scop_p scop)
1887 int i, j;
1888 poly_bb_p pbb;
1889 data_reference_p dr;
1890 auto_vec<data_reference_p, 3> drs;
1892 /* Remove all the PBBs that do not have data references: these basic
1893 blocks are not handled in the polyhedral representation. */
1894 for (i = 0; SCOP_BBS (scop).iterate (i, &pbb); i++)
1895 if (GBB_DATA_REFS (PBB_BLACK_BOX (pbb)).is_empty ())
1897 free_gimple_bb (PBB_BLACK_BOX (pbb));
1898 free_poly_bb (pbb);
1899 SCOP_BBS (scop).ordered_remove (i);
1900 i--;
1903 FOR_EACH_VEC_ELT (SCOP_BBS (scop), i, pbb)
1904 for (j = 0; GBB_DATA_REFS (PBB_BLACK_BOX (pbb)).iterate (j, &dr); j++)
1905 drs.safe_push (dr);
1907 FOR_EACH_VEC_ELT (drs, i, dr)
1908 dr->aux = XNEW (base_alias_pair);
1910 if (!build_alias_set_optimal_p (drs))
1912 /* TODO: Add support when building alias set is not optimal. */
1916 build_base_obj_set_for_drs (drs);
1918 /* When debugging, enable the following code. This cannot be used
1919 in production compilers. */
1920 if (0)
1921 dump_alias_graphs (drs);
1923 drs.release ();
1925 FOR_EACH_VEC_ELT (SCOP_BBS (scop), i, pbb)
1926 build_pbb_drs (pbb);
1929 /* Return a gsi at the position of the phi node STMT. */
1931 static gimple_stmt_iterator
1932 gsi_for_phi_node (gimple stmt)
1934 gimple_stmt_iterator psi;
1935 basic_block bb = gimple_bb (stmt);
1937 for (psi = gsi_start_phis (bb); !gsi_end_p (psi); gsi_next (&psi))
1938 if (stmt == gsi_stmt (psi))
1939 return psi;
1941 gcc_unreachable ();
1942 return psi;
1945 /* Analyze all the data references of STMTS and add them to the
1946 GBB_DATA_REFS vector of BB. */
1948 static void
1949 analyze_drs_in_stmts (scop_p scop, basic_block bb, vec<gimple> stmts)
1951 loop_p nest;
1952 gimple_bb_p gbb;
1953 gimple stmt;
1954 int i;
1955 sese region = SCOP_REGION (scop);
1957 if (!bb_in_sese_p (bb, region))
1958 return;
1960 nest = outermost_loop_in_sese_1 (region, bb);
1961 gbb = gbb_from_bb (bb);
1963 FOR_EACH_VEC_ELT (stmts, i, stmt)
1965 loop_p loop;
1967 if (is_gimple_debug (stmt))
1968 continue;
1970 loop = loop_containing_stmt (stmt);
1971 if (!loop_in_sese_p (loop, region))
1972 loop = nest;
1974 graphite_find_data_references_in_stmt (nest, loop, stmt,
1975 &GBB_DATA_REFS (gbb));
1979 /* Insert STMT at the end of the STMTS sequence and then insert the
1980 statements from STMTS at INSERT_GSI and call analyze_drs_in_stmts
1981 on STMTS. */
1983 static void
1984 insert_stmts (scop_p scop, gimple stmt, gimple_seq stmts,
1985 gimple_stmt_iterator insert_gsi)
1987 gimple_stmt_iterator gsi;
1988 auto_vec<gimple, 3> x;
1990 gimple_seq_add_stmt (&stmts, stmt);
1991 for (gsi = gsi_start (stmts); !gsi_end_p (gsi); gsi_next (&gsi))
1992 x.safe_push (gsi_stmt (gsi));
1994 gsi_insert_seq_before (&insert_gsi, stmts, GSI_SAME_STMT);
1995 analyze_drs_in_stmts (scop, gsi_bb (insert_gsi), x);
1998 /* Insert the assignment "RES := EXPR" just after AFTER_STMT. */
2000 static void
2001 insert_out_of_ssa_copy (scop_p scop, tree res, tree expr, gimple after_stmt)
2003 gimple_seq stmts;
2004 gimple_stmt_iterator gsi;
2005 tree var = force_gimple_operand (expr, &stmts, true, NULL_TREE);
2006 gimple stmt = gimple_build_assign (unshare_expr (res), var);
2007 auto_vec<gimple, 3> x;
2009 gimple_seq_add_stmt (&stmts, stmt);
2010 for (gsi = gsi_start (stmts); !gsi_end_p (gsi); gsi_next (&gsi))
2011 x.safe_push (gsi_stmt (gsi));
2013 if (gimple_code (after_stmt) == GIMPLE_PHI)
2015 gsi = gsi_after_labels (gimple_bb (after_stmt));
2016 gsi_insert_seq_before (&gsi, stmts, GSI_NEW_STMT);
2018 else
2020 gsi = gsi_for_stmt (after_stmt);
2021 gsi_insert_seq_after (&gsi, stmts, GSI_NEW_STMT);
2024 analyze_drs_in_stmts (scop, gimple_bb (after_stmt), x);
2027 /* Creates a poly_bb_p for basic_block BB from the existing PBB. */
2029 static void
2030 new_pbb_from_pbb (scop_p scop, poly_bb_p pbb, basic_block bb)
2032 vec<data_reference_p> drs;
2033 drs.create (3);
2034 gimple_bb_p gbb = PBB_BLACK_BOX (pbb);
2035 gimple_bb_p gbb1 = new_gimple_bb (bb, drs);
2036 poly_bb_p pbb1 = new_poly_bb (scop, gbb1);
2037 int index, n = SCOP_BBS (scop).length ();
2039 /* The INDEX of PBB in SCOP_BBS. */
2040 for (index = 0; index < n; index++)
2041 if (SCOP_BBS (scop)[index] == pbb)
2042 break;
2044 pbb1->domain = isl_set_copy (pbb->domain);
2045 pbb1->domain = isl_set_set_tuple_id (pbb1->domain,
2046 isl_id_for_pbb (scop, pbb1));
2048 GBB_PBB (gbb1) = pbb1;
2049 GBB_CONDITIONS (gbb1) = GBB_CONDITIONS (gbb).copy ();
2050 GBB_CONDITION_CASES (gbb1) = GBB_CONDITION_CASES (gbb).copy ();
2051 SCOP_BBS (scop).safe_insert (index + 1, pbb1);
2054 /* Insert on edge E the assignment "RES := EXPR". */
2056 static void
2057 insert_out_of_ssa_copy_on_edge (scop_p scop, edge e, tree res, tree expr)
2059 gimple_stmt_iterator gsi;
2060 gimple_seq stmts = NULL;
2061 tree var = force_gimple_operand (expr, &stmts, true, NULL_TREE);
2062 gimple stmt = gimple_build_assign (unshare_expr (res), var);
2063 basic_block bb;
2064 auto_vec<gimple, 3> x;
2066 gimple_seq_add_stmt (&stmts, stmt);
2067 for (gsi = gsi_start (stmts); !gsi_end_p (gsi); gsi_next (&gsi))
2068 x.safe_push (gsi_stmt (gsi));
2070 gsi_insert_seq_on_edge (e, stmts);
2071 gsi_commit_edge_inserts ();
2072 bb = gimple_bb (stmt);
2074 if (!bb_in_sese_p (bb, SCOP_REGION (scop)))
2075 return;
2077 if (!gbb_from_bb (bb))
2078 new_pbb_from_pbb (scop, pbb_from_bb (e->src), bb);
2080 analyze_drs_in_stmts (scop, bb, x);
2083 /* Creates a zero dimension array of the same type as VAR. */
2085 static tree
2086 create_zero_dim_array (tree var, const char *base_name)
2088 tree index_type = build_index_type (integer_zero_node);
2089 tree elt_type = TREE_TYPE (var);
2090 tree array_type = build_array_type (elt_type, index_type);
2091 tree base = create_tmp_var (array_type, base_name);
2093 return build4 (ARRAY_REF, elt_type, base, integer_zero_node, NULL_TREE,
2094 NULL_TREE);
2097 /* Returns true when PHI is a loop close phi node. */
2099 static bool
2100 scalar_close_phi_node_p (gimple phi)
2102 if (gimple_code (phi) != GIMPLE_PHI
2103 || virtual_operand_p (gimple_phi_result (phi)))
2104 return false;
2106 /* Note that loop close phi nodes should have a single argument
2107 because we translated the representation into a canonical form
2108 before Graphite: see canonicalize_loop_closed_ssa_form. */
2109 return (gimple_phi_num_args (phi) == 1);
2112 /* For a definition DEF in REGION, propagates the expression EXPR in
2113 all the uses of DEF outside REGION. */
2115 static void
2116 propagate_expr_outside_region (tree def, tree expr, sese region)
2118 imm_use_iterator imm_iter;
2119 gimple use_stmt;
2120 gimple_seq stmts;
2121 bool replaced_once = false;
2123 gcc_assert (TREE_CODE (def) == SSA_NAME);
2125 expr = force_gimple_operand (unshare_expr (expr), &stmts, true,
2126 NULL_TREE);
2128 FOR_EACH_IMM_USE_STMT (use_stmt, imm_iter, def)
2129 if (!is_gimple_debug (use_stmt)
2130 && !bb_in_sese_p (gimple_bb (use_stmt), region))
2132 ssa_op_iter iter;
2133 use_operand_p use_p;
2135 FOR_EACH_PHI_OR_STMT_USE (use_p, use_stmt, iter, SSA_OP_ALL_USES)
2136 if (operand_equal_p (def, USE_FROM_PTR (use_p), 0)
2137 && (replaced_once = true))
2138 replace_exp (use_p, expr);
2140 update_stmt (use_stmt);
2143 if (replaced_once)
2145 gsi_insert_seq_on_edge (SESE_ENTRY (region), stmts);
2146 gsi_commit_edge_inserts ();
2150 /* Rewrite out of SSA the reduction phi node at PSI by creating a zero
2151 dimension array for it. */
2153 static void
2154 rewrite_close_phi_out_of_ssa (scop_p scop, gimple_stmt_iterator *psi)
2156 sese region = SCOP_REGION (scop);
2157 gimple phi = gsi_stmt (*psi);
2158 tree res = gimple_phi_result (phi);
2159 basic_block bb = gimple_bb (phi);
2160 gimple_stmt_iterator gsi = gsi_after_labels (bb);
2161 tree arg = gimple_phi_arg_def (phi, 0);
2162 gimple stmt;
2164 /* Note that loop close phi nodes should have a single argument
2165 because we translated the representation into a canonical form
2166 before Graphite: see canonicalize_loop_closed_ssa_form. */
2167 gcc_assert (gimple_phi_num_args (phi) == 1);
2169 /* The phi node can be a non close phi node, when its argument is
2170 invariant, or a default definition. */
2171 if (is_gimple_min_invariant (arg)
2172 || SSA_NAME_IS_DEFAULT_DEF (arg))
2174 propagate_expr_outside_region (res, arg, region);
2175 gsi_next (psi);
2176 return;
2179 else if (gimple_bb (SSA_NAME_DEF_STMT (arg))->loop_father == bb->loop_father)
2181 propagate_expr_outside_region (res, arg, region);
2182 stmt = gimple_build_assign (res, arg);
2183 remove_phi_node (psi, false);
2184 gsi_insert_before (&gsi, stmt, GSI_NEW_STMT);
2185 return;
2188 /* If res is scev analyzable and is not a scalar value, it is safe
2189 to ignore the close phi node: it will be code generated in the
2190 out of Graphite pass. */
2191 else if (scev_analyzable_p (res, region))
2193 loop_p loop = loop_containing_stmt (SSA_NAME_DEF_STMT (res));
2194 tree scev;
2196 if (!loop_in_sese_p (loop, region))
2198 loop = loop_containing_stmt (SSA_NAME_DEF_STMT (arg));
2199 scev = scalar_evolution_in_region (region, loop, arg);
2200 scev = compute_overall_effect_of_inner_loop (loop, scev);
2202 else
2203 scev = scalar_evolution_in_region (region, loop, res);
2205 if (tree_does_not_contain_chrecs (scev))
2206 propagate_expr_outside_region (res, scev, region);
2208 gsi_next (psi);
2209 return;
2211 else
2213 tree zero_dim_array = create_zero_dim_array (res, "Close_Phi");
2215 stmt = gimple_build_assign (res, unshare_expr (zero_dim_array));
2217 if (TREE_CODE (arg) == SSA_NAME)
2218 insert_out_of_ssa_copy (scop, zero_dim_array, arg,
2219 SSA_NAME_DEF_STMT (arg));
2220 else
2221 insert_out_of_ssa_copy_on_edge (scop, single_pred_edge (bb),
2222 zero_dim_array, arg);
2225 remove_phi_node (psi, false);
2226 SSA_NAME_DEF_STMT (res) = stmt;
2228 insert_stmts (scop, stmt, NULL, gsi_after_labels (bb));
2231 /* Rewrite out of SSA the reduction phi node at PSI by creating a zero
2232 dimension array for it. */
2234 static void
2235 rewrite_phi_out_of_ssa (scop_p scop, gimple_stmt_iterator *psi)
2237 size_t i;
2238 gimple phi = gsi_stmt (*psi);
2239 basic_block bb = gimple_bb (phi);
2240 tree res = gimple_phi_result (phi);
2241 tree zero_dim_array = create_zero_dim_array (res, "phi_out_of_ssa");
2242 gimple stmt;
2244 for (i = 0; i < gimple_phi_num_args (phi); i++)
2246 tree arg = gimple_phi_arg_def (phi, i);
2247 edge e = gimple_phi_arg_edge (phi, i);
2249 /* Avoid the insertion of code in the loop latch to please the
2250 pattern matching of the vectorizer. */
2251 if (TREE_CODE (arg) == SSA_NAME
2252 && !SSA_NAME_IS_DEFAULT_DEF (arg)
2253 && e->src == bb->loop_father->latch)
2254 insert_out_of_ssa_copy (scop, zero_dim_array, arg,
2255 SSA_NAME_DEF_STMT (arg));
2256 else
2257 insert_out_of_ssa_copy_on_edge (scop, e, zero_dim_array, arg);
2260 stmt = gimple_build_assign (res, unshare_expr (zero_dim_array));
2261 remove_phi_node (psi, false);
2262 insert_stmts (scop, stmt, NULL, gsi_after_labels (bb));
2265 /* Rewrite the degenerate phi node at position PSI from the degenerate
2266 form "x = phi (y, y, ..., y)" to "x = y". */
2268 static void
2269 rewrite_degenerate_phi (gimple_stmt_iterator *psi)
2271 tree rhs;
2272 gimple stmt;
2273 gimple_stmt_iterator gsi;
2274 gimple phi = gsi_stmt (*psi);
2275 tree res = gimple_phi_result (phi);
2276 basic_block bb;
2278 bb = gimple_bb (phi);
2279 rhs = degenerate_phi_result (phi);
2280 gcc_assert (rhs);
2282 stmt = gimple_build_assign (res, rhs);
2283 remove_phi_node (psi, false);
2285 gsi = gsi_after_labels (bb);
2286 gsi_insert_before (&gsi, stmt, GSI_NEW_STMT);
2289 /* Rewrite out of SSA all the reduction phi nodes of SCOP. */
2291 static void
2292 rewrite_reductions_out_of_ssa (scop_p scop)
2294 basic_block bb;
2295 gimple_stmt_iterator psi;
2296 sese region = SCOP_REGION (scop);
2298 FOR_EACH_BB_FN (bb, cfun)
2299 if (bb_in_sese_p (bb, region))
2300 for (psi = gsi_start_phis (bb); !gsi_end_p (psi);)
2302 gimple phi = gsi_stmt (psi);
2304 if (virtual_operand_p (gimple_phi_result (phi)))
2306 gsi_next (&psi);
2307 continue;
2310 if (gimple_phi_num_args (phi) > 1
2311 && degenerate_phi_result (phi))
2312 rewrite_degenerate_phi (&psi);
2314 else if (scalar_close_phi_node_p (phi))
2315 rewrite_close_phi_out_of_ssa (scop, &psi);
2317 else if (reduction_phi_p (region, &psi))
2318 rewrite_phi_out_of_ssa (scop, &psi);
2321 update_ssa (TODO_update_ssa);
2322 #ifdef ENABLE_CHECKING
2323 verify_loop_closed_ssa (true);
2324 #endif
2327 /* Rewrite the scalar dependence of DEF used in USE_STMT with a memory
2328 read from ZERO_DIM_ARRAY. */
2330 static void
2331 rewrite_cross_bb_scalar_dependence (scop_p scop, tree zero_dim_array,
2332 tree def, gimple use_stmt)
2334 gimple name_stmt;
2335 tree name;
2336 ssa_op_iter iter;
2337 use_operand_p use_p;
2339 gcc_assert (gimple_code (use_stmt) != GIMPLE_PHI);
2341 name = copy_ssa_name (def, NULL);
2342 name_stmt = gimple_build_assign (name, zero_dim_array);
2344 gimple_assign_set_lhs (name_stmt, name);
2345 insert_stmts (scop, name_stmt, NULL, gsi_for_stmt (use_stmt));
2347 FOR_EACH_SSA_USE_OPERAND (use_p, use_stmt, iter, SSA_OP_ALL_USES)
2348 if (operand_equal_p (def, USE_FROM_PTR (use_p), 0))
2349 replace_exp (use_p, name);
2351 update_stmt (use_stmt);
2354 /* For every definition DEF in the SCOP that is used outside the scop,
2355 insert a closing-scop definition in the basic block just after this
2356 SCOP. */
2358 static void
2359 handle_scalar_deps_crossing_scop_limits (scop_p scop, tree def, gimple stmt)
2361 tree var = create_tmp_reg (TREE_TYPE (def), NULL);
2362 tree new_name = make_ssa_name (var, stmt);
2363 bool needs_copy = false;
2364 use_operand_p use_p;
2365 imm_use_iterator imm_iter;
2366 gimple use_stmt;
2367 sese region = SCOP_REGION (scop);
2369 FOR_EACH_IMM_USE_STMT (use_stmt, imm_iter, def)
2371 if (!bb_in_sese_p (gimple_bb (use_stmt), region))
2373 FOR_EACH_IMM_USE_ON_STMT (use_p, imm_iter)
2375 SET_USE (use_p, new_name);
2377 update_stmt (use_stmt);
2378 needs_copy = true;
2382 /* Insert in the empty BB just after the scop a use of DEF such
2383 that the rewrite of cross_bb_scalar_dependences won't insert
2384 arrays everywhere else. */
2385 if (needs_copy)
2387 gimple assign = gimple_build_assign (new_name, def);
2388 gimple_stmt_iterator psi = gsi_after_labels (SESE_EXIT (region)->dest);
2390 update_stmt (assign);
2391 gsi_insert_before (&psi, assign, GSI_SAME_STMT);
2395 /* Rewrite the scalar dependences crossing the boundary of the BB
2396 containing STMT with an array. Return true when something has been
2397 changed. */
2399 static bool
2400 rewrite_cross_bb_scalar_deps (scop_p scop, gimple_stmt_iterator *gsi)
2402 sese region = SCOP_REGION (scop);
2403 gimple stmt = gsi_stmt (*gsi);
2404 imm_use_iterator imm_iter;
2405 tree def;
2406 basic_block def_bb;
2407 tree zero_dim_array = NULL_TREE;
2408 gimple use_stmt;
2409 bool res = false;
2411 switch (gimple_code (stmt))
2413 case GIMPLE_ASSIGN:
2414 def = gimple_assign_lhs (stmt);
2415 break;
2417 case GIMPLE_CALL:
2418 def = gimple_call_lhs (stmt);
2419 break;
2421 default:
2422 return false;
2425 if (!def
2426 || !is_gimple_reg (def))
2427 return false;
2429 if (scev_analyzable_p (def, region))
2431 loop_p loop = loop_containing_stmt (SSA_NAME_DEF_STMT (def));
2432 tree scev = scalar_evolution_in_region (region, loop, def);
2434 if (tree_contains_chrecs (scev, NULL))
2435 return false;
2437 propagate_expr_outside_region (def, scev, region);
2438 return true;
2441 def_bb = gimple_bb (stmt);
2443 handle_scalar_deps_crossing_scop_limits (scop, def, stmt);
2445 FOR_EACH_IMM_USE_STMT (use_stmt, imm_iter, def)
2446 if (gimple_code (use_stmt) == GIMPLE_PHI
2447 && (res = true))
2449 gimple_stmt_iterator psi = gsi_for_stmt (use_stmt);
2451 if (scalar_close_phi_node_p (gsi_stmt (psi)))
2452 rewrite_close_phi_out_of_ssa (scop, &psi);
2453 else
2454 rewrite_phi_out_of_ssa (scop, &psi);
2457 FOR_EACH_IMM_USE_STMT (use_stmt, imm_iter, def)
2458 if (gimple_code (use_stmt) != GIMPLE_PHI
2459 && def_bb != gimple_bb (use_stmt)
2460 && !is_gimple_debug (use_stmt)
2461 && (res = true))
2463 if (!zero_dim_array)
2465 zero_dim_array = create_zero_dim_array
2466 (def, "Cross_BB_scalar_dependence");
2467 insert_out_of_ssa_copy (scop, zero_dim_array, def,
2468 SSA_NAME_DEF_STMT (def));
2469 gsi_next (gsi);
2472 rewrite_cross_bb_scalar_dependence (scop, unshare_expr (zero_dim_array),
2473 def, use_stmt);
2476 return res;
2479 /* Rewrite out of SSA all the reduction phi nodes of SCOP. */
2481 static void
2482 rewrite_cross_bb_scalar_deps_out_of_ssa (scop_p scop)
2484 basic_block bb;
2485 gimple_stmt_iterator psi;
2486 sese region = SCOP_REGION (scop);
2487 bool changed = false;
2489 /* Create an extra empty BB after the scop. */
2490 split_edge (SESE_EXIT (region));
2492 FOR_EACH_BB_FN (bb, cfun)
2493 if (bb_in_sese_p (bb, region))
2494 for (psi = gsi_start_bb (bb); !gsi_end_p (psi); gsi_next (&psi))
2495 changed |= rewrite_cross_bb_scalar_deps (scop, &psi);
2497 if (changed)
2499 scev_reset_htab ();
2500 update_ssa (TODO_update_ssa);
2501 #ifdef ENABLE_CHECKING
2502 verify_loop_closed_ssa (true);
2503 #endif
2507 /* Returns the number of pbbs that are in loops contained in SCOP. */
2509 static int
2510 nb_pbbs_in_loops (scop_p scop)
2512 int i;
2513 poly_bb_p pbb;
2514 int res = 0;
2516 FOR_EACH_VEC_ELT (SCOP_BBS (scop), i, pbb)
2517 if (loop_in_sese_p (gbb_loop (PBB_BLACK_BOX (pbb)), SCOP_REGION (scop)))
2518 res++;
2520 return res;
2523 /* Return the number of data references in BB that write in
2524 memory. */
2526 static int
2527 nb_data_writes_in_bb (basic_block bb)
2529 int res = 0;
2530 gimple_stmt_iterator gsi;
2532 for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi))
2533 if (gimple_vdef (gsi_stmt (gsi)))
2534 res++;
2536 return res;
2539 /* Splits at STMT the basic block BB represented as PBB in the
2540 polyhedral form. */
2542 static edge
2543 split_pbb (scop_p scop, poly_bb_p pbb, basic_block bb, gimple stmt)
2545 edge e1 = split_block (bb, stmt);
2546 new_pbb_from_pbb (scop, pbb, e1->dest);
2547 return e1;
2550 /* Splits STMT out of its current BB. This is done for reduction
2551 statements for which we want to ignore data dependences. */
2553 static basic_block
2554 split_reduction_stmt (scop_p scop, gimple stmt)
2556 basic_block bb = gimple_bb (stmt);
2557 poly_bb_p pbb = pbb_from_bb (bb);
2558 gimple_bb_p gbb = gbb_from_bb (bb);
2559 edge e1;
2560 int i;
2561 data_reference_p dr;
2563 /* Do not split basic blocks with no writes to memory: the reduction
2564 will be the only write to memory. */
2565 if (nb_data_writes_in_bb (bb) == 0
2566 /* Or if we have already marked BB as a reduction. */
2567 || PBB_IS_REDUCTION (pbb_from_bb (bb)))
2568 return bb;
2570 e1 = split_pbb (scop, pbb, bb, stmt);
2572 /* Split once more only when the reduction stmt is not the only one
2573 left in the original BB. */
2574 if (!gsi_one_before_end_p (gsi_start_nondebug_bb (bb)))
2576 gimple_stmt_iterator gsi = gsi_last_bb (bb);
2577 gsi_prev (&gsi);
2578 e1 = split_pbb (scop, pbb, bb, gsi_stmt (gsi));
2581 /* A part of the data references will end in a different basic block
2582 after the split: move the DRs from the original GBB to the newly
2583 created GBB1. */
2584 FOR_EACH_VEC_ELT (GBB_DATA_REFS (gbb), i, dr)
2586 basic_block bb1 = gimple_bb (DR_STMT (dr));
2588 if (bb1 != bb)
2590 gimple_bb_p gbb1 = gbb_from_bb (bb1);
2591 GBB_DATA_REFS (gbb1).safe_push (dr);
2592 GBB_DATA_REFS (gbb).ordered_remove (i);
2593 i--;
2597 return e1->dest;
2600 /* Return true when stmt is a reduction operation. */
2602 static inline bool
2603 is_reduction_operation_p (gimple stmt)
2605 enum tree_code code;
2607 gcc_assert (is_gimple_assign (stmt));
2608 code = gimple_assign_rhs_code (stmt);
2610 return flag_associative_math
2611 && commutative_tree_code (code)
2612 && associative_tree_code (code);
2615 /* Returns true when PHI contains an argument ARG. */
2617 static bool
2618 phi_contains_arg (gimple phi, tree arg)
2620 size_t i;
2622 for (i = 0; i < gimple_phi_num_args (phi); i++)
2623 if (operand_equal_p (arg, gimple_phi_arg_def (phi, i), 0))
2624 return true;
2626 return false;
2629 /* Return a loop phi node that corresponds to a reduction containing LHS. */
2631 static gimple
2632 follow_ssa_with_commutative_ops (tree arg, tree lhs)
2634 gimple stmt;
2636 if (TREE_CODE (arg) != SSA_NAME)
2637 return NULL;
2639 stmt = SSA_NAME_DEF_STMT (arg);
2641 if (gimple_code (stmt) == GIMPLE_NOP
2642 || gimple_code (stmt) == GIMPLE_CALL)
2643 return NULL;
2645 if (gimple_code (stmt) == GIMPLE_PHI)
2647 if (phi_contains_arg (stmt, lhs))
2648 return stmt;
2649 return NULL;
2652 if (!is_gimple_assign (stmt))
2653 return NULL;
2655 if (gimple_num_ops (stmt) == 2)
2656 return follow_ssa_with_commutative_ops (gimple_assign_rhs1 (stmt), lhs);
2658 if (is_reduction_operation_p (stmt))
2660 gimple res = follow_ssa_with_commutative_ops (gimple_assign_rhs1 (stmt), lhs);
2662 return res ? res :
2663 follow_ssa_with_commutative_ops (gimple_assign_rhs2 (stmt), lhs);
2666 return NULL;
2669 /* Detect commutative and associative scalar reductions starting at
2670 the STMT. Return the phi node of the reduction cycle, or NULL. */
2672 static gimple
2673 detect_commutative_reduction_arg (tree lhs, gimple stmt, tree arg,
2674 vec<gimple> *in,
2675 vec<gimple> *out)
2677 gimple phi = follow_ssa_with_commutative_ops (arg, lhs);
2679 if (!phi)
2680 return NULL;
2682 in->safe_push (stmt);
2683 out->safe_push (stmt);
2684 return phi;
2687 /* Detect commutative and associative scalar reductions starting at
2688 STMT. Return the phi node of the reduction cycle, or NULL. */
2690 static gimple
2691 detect_commutative_reduction_assign (gimple stmt, vec<gimple> *in,
2692 vec<gimple> *out)
2694 tree lhs = gimple_assign_lhs (stmt);
2696 if (gimple_num_ops (stmt) == 2)
2697 return detect_commutative_reduction_arg (lhs, stmt,
2698 gimple_assign_rhs1 (stmt),
2699 in, out);
2701 if (is_reduction_operation_p (stmt))
2703 gimple res = detect_commutative_reduction_arg (lhs, stmt,
2704 gimple_assign_rhs1 (stmt),
2705 in, out);
2706 return res ? res
2707 : detect_commutative_reduction_arg (lhs, stmt,
2708 gimple_assign_rhs2 (stmt),
2709 in, out);
2712 return NULL;
2715 /* Return a loop phi node that corresponds to a reduction containing LHS. */
2717 static gimple
2718 follow_inital_value_to_phi (tree arg, tree lhs)
2720 gimple stmt;
2722 if (!arg || TREE_CODE (arg) != SSA_NAME)
2723 return NULL;
2725 stmt = SSA_NAME_DEF_STMT (arg);
2727 if (gimple_code (stmt) == GIMPLE_PHI
2728 && phi_contains_arg (stmt, lhs))
2729 return stmt;
2731 return NULL;
2735 /* Return the argument of the loop PHI that is the initial value coming
2736 from outside the loop. */
2738 static edge
2739 edge_initial_value_for_loop_phi (gimple phi)
2741 size_t i;
2743 for (i = 0; i < gimple_phi_num_args (phi); i++)
2745 edge e = gimple_phi_arg_edge (phi, i);
2747 if (loop_depth (e->src->loop_father)
2748 < loop_depth (e->dest->loop_father))
2749 return e;
2752 return NULL;
2755 /* Return the argument of the loop PHI that is the initial value coming
2756 from outside the loop. */
2758 static tree
2759 initial_value_for_loop_phi (gimple phi)
2761 size_t i;
2763 for (i = 0; i < gimple_phi_num_args (phi); i++)
2765 edge e = gimple_phi_arg_edge (phi, i);
2767 if (loop_depth (e->src->loop_father)
2768 < loop_depth (e->dest->loop_father))
2769 return gimple_phi_arg_def (phi, i);
2772 return NULL_TREE;
2775 /* Returns true when DEF is used outside the reduction cycle of
2776 LOOP_PHI. */
2778 static bool
2779 used_outside_reduction (tree def, gimple loop_phi)
2781 use_operand_p use_p;
2782 imm_use_iterator imm_iter;
2783 loop_p loop = loop_containing_stmt (loop_phi);
2785 /* In LOOP, DEF should be used only in LOOP_PHI. */
2786 FOR_EACH_IMM_USE_FAST (use_p, imm_iter, def)
2788 gimple stmt = USE_STMT (use_p);
2790 if (stmt != loop_phi
2791 && !is_gimple_debug (stmt)
2792 && flow_bb_inside_loop_p (loop, gimple_bb (stmt)))
2793 return true;
2796 return false;
2799 /* Detect commutative and associative scalar reductions belonging to
2800 the SCOP starting at the loop closed phi node STMT. Return the phi
2801 node of the reduction cycle, or NULL. */
2803 static gimple
2804 detect_commutative_reduction (scop_p scop, gimple stmt, vec<gimple> *in,
2805 vec<gimple> *out)
2807 if (scalar_close_phi_node_p (stmt))
2809 gimple def, loop_phi, phi, close_phi = stmt;
2810 tree init, lhs, arg = gimple_phi_arg_def (close_phi, 0);
2812 if (TREE_CODE (arg) != SSA_NAME)
2813 return NULL;
2815 /* Note that loop close phi nodes should have a single argument
2816 because we translated the representation into a canonical form
2817 before Graphite: see canonicalize_loop_closed_ssa_form. */
2818 gcc_assert (gimple_phi_num_args (close_phi) == 1);
2820 def = SSA_NAME_DEF_STMT (arg);
2821 if (!stmt_in_sese_p (def, SCOP_REGION (scop))
2822 || !(loop_phi = detect_commutative_reduction (scop, def, in, out)))
2823 return NULL;
2825 lhs = gimple_phi_result (close_phi);
2826 init = initial_value_for_loop_phi (loop_phi);
2827 phi = follow_inital_value_to_phi (init, lhs);
2829 if (phi && (used_outside_reduction (lhs, phi)
2830 || !has_single_use (gimple_phi_result (phi))))
2831 return NULL;
2833 in->safe_push (loop_phi);
2834 out->safe_push (close_phi);
2835 return phi;
2838 if (gimple_code (stmt) == GIMPLE_ASSIGN)
2839 return detect_commutative_reduction_assign (stmt, in, out);
2841 return NULL;
2844 /* Translate the scalar reduction statement STMT to an array RED
2845 knowing that its recursive phi node is LOOP_PHI. */
2847 static void
2848 translate_scalar_reduction_to_array_for_stmt (scop_p scop, tree red,
2849 gimple stmt, gimple loop_phi)
2851 tree res = gimple_phi_result (loop_phi);
2852 gimple assign = gimple_build_assign (res, unshare_expr (red));
2853 gimple_stmt_iterator gsi;
2855 insert_stmts (scop, assign, NULL, gsi_after_labels (gimple_bb (loop_phi)));
2857 assign = gimple_build_assign (unshare_expr (red), gimple_assign_lhs (stmt));
2858 gsi = gsi_for_stmt (stmt);
2859 gsi_next (&gsi);
2860 insert_stmts (scop, assign, NULL, gsi);
2863 /* Removes the PHI node and resets all the debug stmts that are using
2864 the PHI_RESULT. */
2866 static void
2867 remove_phi (gimple phi)
2869 imm_use_iterator imm_iter;
2870 tree def;
2871 use_operand_p use_p;
2872 gimple_stmt_iterator gsi;
2873 auto_vec<gimple, 3> update;
2874 unsigned int i;
2875 gimple stmt;
2877 def = PHI_RESULT (phi);
2878 FOR_EACH_IMM_USE_FAST (use_p, imm_iter, def)
2880 stmt = USE_STMT (use_p);
2882 if (is_gimple_debug (stmt))
2884 gimple_debug_bind_reset_value (stmt);
2885 update.safe_push (stmt);
2889 FOR_EACH_VEC_ELT (update, i, stmt)
2890 update_stmt (stmt);
2892 gsi = gsi_for_phi_node (phi);
2893 remove_phi_node (&gsi, false);
2896 /* Helper function for for_each_index. For each INDEX of the data
2897 reference REF, returns true when its indices are valid in the loop
2898 nest LOOP passed in as DATA. */
2900 static bool
2901 dr_indices_valid_in_loop (tree ref ATTRIBUTE_UNUSED, tree *index, void *data)
2903 loop_p loop;
2904 basic_block header, def_bb;
2905 gimple stmt;
2907 if (TREE_CODE (*index) != SSA_NAME)
2908 return true;
2910 loop = *((loop_p *) data);
2911 header = loop->header;
2912 stmt = SSA_NAME_DEF_STMT (*index);
2914 if (!stmt)
2915 return true;
2917 def_bb = gimple_bb (stmt);
2919 if (!def_bb)
2920 return true;
2922 return dominated_by_p (CDI_DOMINATORS, header, def_bb);
2925 /* When the result of a CLOSE_PHI is written to a memory location,
2926 return a pointer to that memory reference, otherwise return
2927 NULL_TREE. */
2929 static tree
2930 close_phi_written_to_memory (gimple close_phi)
2932 imm_use_iterator imm_iter;
2933 use_operand_p use_p;
2934 gimple stmt;
2935 tree res, def = gimple_phi_result (close_phi);
2937 FOR_EACH_IMM_USE_FAST (use_p, imm_iter, def)
2938 if ((stmt = USE_STMT (use_p))
2939 && gimple_code (stmt) == GIMPLE_ASSIGN
2940 && (res = gimple_assign_lhs (stmt)))
2942 switch (TREE_CODE (res))
2944 case VAR_DECL:
2945 case PARM_DECL:
2946 case RESULT_DECL:
2947 return res;
2949 case ARRAY_REF:
2950 case MEM_REF:
2952 tree arg = gimple_phi_arg_def (close_phi, 0);
2953 loop_p nest = loop_containing_stmt (SSA_NAME_DEF_STMT (arg));
2955 /* FIXME: this restriction is for id-{24,25}.f and
2956 could be handled by duplicating the computation of
2957 array indices before the loop of the close_phi. */
2958 if (for_each_index (&res, dr_indices_valid_in_loop, &nest))
2959 return res;
2961 /* Fallthru. */
2963 default:
2964 continue;
2967 return NULL_TREE;
2970 /* Rewrite out of SSA the reduction described by the loop phi nodes
2971 IN, and the close phi nodes OUT. IN and OUT are structured by loop
2972 levels like this:
2974 IN: stmt, loop_n, ..., loop_0
2975 OUT: stmt, close_n, ..., close_0
2977 the first element is the reduction statement, and the next elements
2978 are the loop and close phi nodes of each of the outer loops. */
2980 static void
2981 translate_scalar_reduction_to_array (scop_p scop,
2982 vec<gimple> in,
2983 vec<gimple> out)
2985 gimple loop_phi;
2986 unsigned int i = out.length () - 1;
2987 tree red = close_phi_written_to_memory (out[i]);
2989 FOR_EACH_VEC_ELT (in, i, loop_phi)
2991 gimple close_phi = out[i];
2993 if (i == 0)
2995 gimple stmt = loop_phi;
2996 basic_block bb = split_reduction_stmt (scop, stmt);
2997 poly_bb_p pbb = pbb_from_bb (bb);
2998 PBB_IS_REDUCTION (pbb) = true;
2999 gcc_assert (close_phi == loop_phi);
3001 if (!red)
3002 red = create_zero_dim_array
3003 (gimple_assign_lhs (stmt), "Commutative_Associative_Reduction");
3005 translate_scalar_reduction_to_array_for_stmt (scop, red, stmt, in[1]);
3006 continue;
3009 if (i == in.length () - 1)
3011 insert_out_of_ssa_copy (scop, gimple_phi_result (close_phi),
3012 unshare_expr (red), close_phi);
3013 insert_out_of_ssa_copy_on_edge
3014 (scop, edge_initial_value_for_loop_phi (loop_phi),
3015 unshare_expr (red), initial_value_for_loop_phi (loop_phi));
3018 remove_phi (loop_phi);
3019 remove_phi (close_phi);
3023 /* Rewrites out of SSA a commutative reduction at CLOSE_PHI. Returns
3024 true when something has been changed. */
3026 static bool
3027 rewrite_commutative_reductions_out_of_ssa_close_phi (scop_p scop,
3028 gimple close_phi)
3030 bool res;
3031 auto_vec<gimple, 10> in;
3032 auto_vec<gimple, 10> out;
3034 detect_commutative_reduction (scop, close_phi, &in, &out);
3035 res = in.length () > 1;
3036 if (res)
3037 translate_scalar_reduction_to_array (scop, in, out);
3039 return res;
3042 /* Rewrites all the commutative reductions from LOOP out of SSA.
3043 Returns true when something has been changed. */
3045 static bool
3046 rewrite_commutative_reductions_out_of_ssa_loop (scop_p scop,
3047 loop_p loop)
3049 gimple_stmt_iterator gsi;
3050 edge exit = single_exit (loop);
3051 tree res;
3052 bool changed = false;
3054 if (!exit)
3055 return false;
3057 for (gsi = gsi_start_phis (exit->dest); !gsi_end_p (gsi); gsi_next (&gsi))
3058 if ((res = gimple_phi_result (gsi_stmt (gsi)))
3059 && !virtual_operand_p (res)
3060 && !scev_analyzable_p (res, SCOP_REGION (scop)))
3061 changed |= rewrite_commutative_reductions_out_of_ssa_close_phi
3062 (scop, gsi_stmt (gsi));
3064 return changed;
3067 /* Rewrites all the commutative reductions from SCOP out of SSA. */
3069 static void
3070 rewrite_commutative_reductions_out_of_ssa (scop_p scop)
3072 loop_p loop;
3073 bool changed = false;
3074 sese region = SCOP_REGION (scop);
3076 FOR_EACH_LOOP (loop, 0)
3077 if (loop_in_sese_p (loop, region))
3078 changed |= rewrite_commutative_reductions_out_of_ssa_loop (scop, loop);
3080 if (changed)
3082 scev_reset_htab ();
3083 gsi_commit_edge_inserts ();
3084 update_ssa (TODO_update_ssa);
3085 #ifdef ENABLE_CHECKING
3086 verify_loop_closed_ssa (true);
3087 #endif
3091 /* Can all ivs be represented by a signed integer?
3092 As CLooG might generate negative values in its expressions, signed loop ivs
3093 are required in the backend. */
3095 static bool
3096 scop_ivs_can_be_represented (scop_p scop)
3098 loop_p loop;
3099 gimple_stmt_iterator psi;
3100 bool result = true;
3102 FOR_EACH_LOOP (loop, 0)
3104 if (!loop_in_sese_p (loop, SCOP_REGION (scop)))
3105 continue;
3107 for (psi = gsi_start_phis (loop->header);
3108 !gsi_end_p (psi); gsi_next (&psi))
3110 gimple phi = gsi_stmt (psi);
3111 tree res = PHI_RESULT (phi);
3112 tree type = TREE_TYPE (res);
3114 if (TYPE_UNSIGNED (type)
3115 && TYPE_PRECISION (type) >= TYPE_PRECISION (long_long_integer_type_node))
3117 result = false;
3118 break;
3121 if (!result)
3122 break;
3125 return result;
3128 /* Builds the polyhedral representation for a SESE region. */
3130 void
3131 build_poly_scop (scop_p scop)
3133 sese region = SCOP_REGION (scop);
3134 graphite_dim_t max_dim;
3136 build_scop_bbs (scop);
3138 /* FIXME: This restriction is needed to avoid a problem in CLooG.
3139 Once CLooG is fixed, remove this guard. Anyways, it makes no
3140 sense to optimize a scop containing only PBBs that do not belong
3141 to any loops. */
3142 if (nb_pbbs_in_loops (scop) == 0)
3143 return;
3145 if (!scop_ivs_can_be_represented (scop))
3146 return;
3148 if (flag_associative_math)
3149 rewrite_commutative_reductions_out_of_ssa (scop);
3151 build_sese_loop_nests (region);
3152 /* Record all conditions in REGION. */
3153 sese_dom_walker (CDI_DOMINATORS, region).walk (cfun->cfg->x_entry_block_ptr);
3154 find_scop_parameters (scop);
3156 max_dim = PARAM_VALUE (PARAM_GRAPHITE_MAX_NB_SCOP_PARAMS);
3157 if (scop_nb_params (scop) > max_dim)
3158 return;
3160 build_scop_iteration_domain (scop);
3161 build_scop_context (scop);
3162 add_conditions_to_constraints (scop);
3164 /* Rewrite out of SSA only after having translated the
3165 representation to the polyhedral representation to avoid scev
3166 analysis failures. That means that these functions will insert
3167 new data references that they create in the right place. */
3168 rewrite_reductions_out_of_ssa (scop);
3169 rewrite_cross_bb_scalar_deps_out_of_ssa (scop);
3171 build_scop_drs (scop);
3172 scop_to_lst (scop);
3173 build_scop_scattering (scop);
3175 /* This SCoP has been translated to the polyhedral
3176 representation. */
3177 POLY_SCOP_P (scop) = true;
3179 #endif