PR go/67101
[official-gcc.git] / gcc / graphite-sese-to-poly.c
blobfdcc790336ca87b4fd95f049061755944a4dd6e1
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 /* Workaround for GMP 5.1.3 bug, see PR56019. */
25 #include <stddef.h>
27 #include <isl/constraint.h>
28 #include <isl/set.h>
29 #include <isl/map.h>
30 #include <isl/union_map.h>
31 #include <isl/constraint.h>
32 #include <isl/aff.h>
33 #include <isl/val.h>
35 /* Since ISL-0.13, the extern is in val_gmp.h. */
36 #if !defined(HAVE_ISL_SCHED_CONSTRAINTS_COMPUTE_SCHEDULE) && defined(__cplusplus)
37 extern "C" {
38 #endif
39 #include <isl/val_gmp.h>
40 #if !defined(HAVE_ISL_SCHED_CONSTRAINTS_COMPUTE_SCHEDULE) && defined(__cplusplus)
42 #endif
44 #include "system.h"
45 #include "coretypes.h"
46 #include "backend.h"
47 #include "cfghooks.h"
48 #include "tree.h"
49 #include "gimple.h"
50 #include "ssa.h"
51 #include "params.h"
52 #include "fold-const.h"
53 #include "gimple-iterator.h"
54 #include "gimplify.h"
55 #include "gimplify-me.h"
56 #include "tree-cfg.h"
57 #include "tree-ssa-loop-manip.h"
58 #include "tree-ssa-loop-niter.h"
59 #include "tree-ssa-loop.h"
60 #include "tree-into-ssa.h"
61 #include "tree-pass.h"
62 #include "cfgloop.h"
63 #include "tree-data-ref.h"
64 #include "tree-scalar-evolution.h"
65 #include "domwalk.h"
66 #include "graphite-poly.h"
67 #include "tree-ssa-propagate.h"
68 #include "graphite-sese-to-poly.h"
71 /* Assigns to RES the value of the INTEGER_CST T. */
73 static inline void
74 tree_int_to_gmp (tree t, mpz_t res)
76 wi::to_mpz (t, res, TYPE_SIGN (TREE_TYPE (t)));
79 /* Returns the index of the PHI argument defined in the outermost
80 loop. */
82 static size_t
83 phi_arg_in_outermost_loop (gphi *phi)
85 loop_p loop = gimple_bb (phi)->loop_father;
86 size_t i, res = 0;
88 for (i = 0; i < gimple_phi_num_args (phi); i++)
89 if (!flow_bb_inside_loop_p (loop, gimple_phi_arg_edge (phi, i)->src))
91 loop = gimple_phi_arg_edge (phi, i)->src->loop_father;
92 res = i;
95 return res;
98 /* Removes a simple copy phi node "RES = phi (INIT, RES)" at position
99 PSI by inserting on the loop ENTRY edge assignment "RES = INIT". */
101 static void
102 remove_simple_copy_phi (gphi_iterator *psi)
104 gphi *phi = psi->phi ();
105 tree res = gimple_phi_result (phi);
106 size_t entry = phi_arg_in_outermost_loop (phi);
107 tree init = gimple_phi_arg_def (phi, entry);
108 gassign *stmt = gimple_build_assign (res, init);
109 edge e = gimple_phi_arg_edge (phi, entry);
111 remove_phi_node (psi, false);
112 gsi_insert_on_edge_immediate (e, stmt);
115 /* Removes an invariant phi node at position PSI by inserting on the
116 loop ENTRY edge the assignment RES = INIT. */
118 static void
119 remove_invariant_phi (sese region, gphi_iterator *psi)
121 gphi *phi = psi->phi ();
122 loop_p loop = loop_containing_stmt (phi);
123 tree res = gimple_phi_result (phi);
124 tree scev = scalar_evolution_in_region (region, loop, res);
125 size_t entry = phi_arg_in_outermost_loop (phi);
126 edge e = gimple_phi_arg_edge (phi, entry);
127 tree var;
128 gassign *stmt;
129 gimple_seq stmts = NULL;
131 if (tree_contains_chrecs (scev, NULL))
132 scev = gimple_phi_arg_def (phi, entry);
134 var = force_gimple_operand (scev, &stmts, true, NULL_TREE);
135 stmt = gimple_build_assign (res, var);
136 remove_phi_node (psi, false);
138 gimple_seq_add_stmt (&stmts, stmt);
139 gsi_insert_seq_on_edge (e, stmts);
140 gsi_commit_edge_inserts ();
141 SSA_NAME_DEF_STMT (res) = stmt;
144 /* Returns true when the phi node at PSI is of the form "a = phi (a, x)". */
146 static inline bool
147 simple_copy_phi_p (gphi *phi)
149 tree res;
151 if (gimple_phi_num_args (phi) != 2)
152 return false;
154 res = gimple_phi_result (phi);
155 return (res == gimple_phi_arg_def (phi, 0)
156 || res == gimple_phi_arg_def (phi, 1));
159 /* Returns true when the phi node at position PSI is a reduction phi
160 node in REGION. Otherwise moves the pointer PSI to the next phi to
161 be considered. */
163 static bool
164 reduction_phi_p (sese region, gphi_iterator *psi)
166 loop_p loop;
167 gphi *phi = psi->phi ();
168 tree res = gimple_phi_result (phi);
170 loop = loop_containing_stmt (phi);
172 if (simple_copy_phi_p (phi))
174 /* PRE introduces phi nodes like these, for an example,
175 see id-5.f in the fortran graphite testsuite:
177 # prephitmp.85_265 = PHI <prephitmp.85_258(33), prephitmp.85_265(18)>
179 remove_simple_copy_phi (psi);
180 return false;
183 if (scev_analyzable_p (res, region))
185 tree scev = scalar_evolution_in_region (region, loop, res);
187 if (evolution_function_is_invariant_p (scev, loop->num))
188 remove_invariant_phi (region, psi);
189 else
190 gsi_next (psi);
192 return false;
195 /* All the other cases are considered reductions. */
196 return true;
199 /* Store the GRAPHITE representation of BB. */
201 static gimple_bb_p
202 new_gimple_bb (basic_block bb, vec<data_reference_p> drs)
204 struct gimple_bb *gbb;
206 gbb = XNEW (struct gimple_bb);
207 bb->aux = gbb;
208 GBB_BB (gbb) = bb;
209 GBB_DATA_REFS (gbb) = drs;
210 GBB_CONDITIONS (gbb).create (0);
211 GBB_CONDITION_CASES (gbb).create (0);
213 return gbb;
216 static void
217 free_data_refs_aux (vec<data_reference_p> datarefs)
219 unsigned int i;
220 struct data_reference *dr;
222 FOR_EACH_VEC_ELT (datarefs, i, dr)
223 if (dr->aux)
225 base_alias_pair *bap = (base_alias_pair *)(dr->aux);
227 free (bap->alias_set);
229 free (bap);
230 dr->aux = NULL;
233 /* Frees GBB. */
235 static void
236 free_gimple_bb (struct gimple_bb *gbb)
238 free_data_refs_aux (GBB_DATA_REFS (gbb));
239 free_data_refs (GBB_DATA_REFS (gbb));
241 GBB_CONDITIONS (gbb).release ();
242 GBB_CONDITION_CASES (gbb).release ();
243 GBB_BB (gbb)->aux = 0;
244 XDELETE (gbb);
247 /* Deletes all gimple bbs in SCOP. */
249 static void
250 remove_gbbs_in_scop (scop_p scop)
252 int i;
253 poly_bb_p pbb;
255 FOR_EACH_VEC_ELT (SCOP_BBS (scop), i, pbb)
256 free_gimple_bb (PBB_BLACK_BOX (pbb));
259 /* Deletes all scops in SCOPS. */
261 void
262 free_scops (vec<scop_p> scops)
264 int i;
265 scop_p scop;
267 FOR_EACH_VEC_ELT (scops, i, scop)
269 remove_gbbs_in_scop (scop);
270 free_sese (SCOP_REGION (scop));
271 free_scop (scop);
274 scops.release ();
277 /* Same as outermost_loop_in_sese, returns the outermost loop
278 containing BB in REGION, but makes sure that the returned loop
279 belongs to the REGION, and so this returns the first loop in the
280 REGION when the loop containing BB does not belong to REGION. */
282 static loop_p
283 outermost_loop_in_sese_1 (sese region, basic_block bb)
285 loop_p nest = outermost_loop_in_sese (region, bb);
287 if (loop_in_sese_p (nest, region))
288 return nest;
290 /* When the basic block BB does not belong to a loop in the region,
291 return the first loop in the region. */
292 nest = nest->inner;
293 while (nest)
294 if (loop_in_sese_p (nest, region))
295 break;
296 else
297 nest = nest->next;
299 gcc_assert (nest);
300 return nest;
303 /* Generates a polyhedral black box only if the bb contains interesting
304 information. */
306 static gimple_bb_p
307 try_generate_gimple_bb (scop_p scop, basic_block bb)
309 vec<data_reference_p> drs;
310 drs.create (5);
311 sese region = SCOP_REGION (scop);
312 loop_p nest = outermost_loop_in_sese_1 (region, bb);
313 gimple_stmt_iterator gsi;
315 for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi))
317 gimple stmt = gsi_stmt (gsi);
318 loop_p loop;
320 if (is_gimple_debug (stmt))
321 continue;
323 loop = loop_containing_stmt (stmt);
324 if (!loop_in_sese_p (loop, region))
325 loop = nest;
327 graphite_find_data_references_in_stmt (nest, loop, stmt, &drs);
330 return new_gimple_bb (bb, drs);
333 /* Returns true if all predecessors of BB, that are not dominated by BB, are
334 marked in MAP. The predecessors dominated by BB are loop latches and will
335 be handled after BB. */
337 static bool
338 all_non_dominated_preds_marked_p (basic_block bb, sbitmap map)
340 edge e;
341 edge_iterator ei;
343 FOR_EACH_EDGE (e, ei, bb->preds)
344 if (!bitmap_bit_p (map, e->src->index)
345 && !dominated_by_p (CDI_DOMINATORS, e->src, bb))
346 return false;
348 return true;
351 /* Compare the depth of two basic_block's P1 and P2. */
353 static int
354 compare_bb_depths (const void *p1, const void *p2)
356 const_basic_block const bb1 = *(const_basic_block const*)p1;
357 const_basic_block const bb2 = *(const_basic_block const*)p2;
358 int d1 = loop_depth (bb1->loop_father);
359 int d2 = loop_depth (bb2->loop_father);
361 if (d1 < d2)
362 return 1;
364 if (d1 > d2)
365 return -1;
367 return 0;
370 /* Sort the basic blocks from DOM such that the first are the ones at
371 a deepest loop level. */
373 static void
374 graphite_sort_dominated_info (vec<basic_block> dom)
376 dom.qsort (compare_bb_depths);
379 /* Recursive helper function for build_scops_bbs. */
381 static void
382 build_scop_bbs_1 (scop_p scop, sbitmap visited, basic_block bb)
384 sese region = SCOP_REGION (scop);
385 vec<basic_block> dom;
386 poly_bb_p pbb;
388 if (bitmap_bit_p (visited, bb->index)
389 || !bb_in_sese_p (bb, region))
390 return;
392 pbb = new_poly_bb (scop, try_generate_gimple_bb (scop, bb));
393 SCOP_BBS (scop).safe_push (pbb);
394 bitmap_set_bit (visited, bb->index);
396 dom = get_dominated_by (CDI_DOMINATORS, bb);
398 if (!dom.exists ())
399 return;
401 graphite_sort_dominated_info (dom);
403 while (!dom.is_empty ())
405 int i;
406 basic_block dom_bb;
408 FOR_EACH_VEC_ELT (dom, i, dom_bb)
409 if (all_non_dominated_preds_marked_p (dom_bb, visited))
411 build_scop_bbs_1 (scop, visited, dom_bb);
412 dom.unordered_remove (i);
413 break;
417 dom.release ();
420 /* Gather the basic blocks belonging to the SCOP. */
422 static void
423 build_scop_bbs (scop_p scop)
425 sbitmap visited = sbitmap_alloc (last_basic_block_for_fn (cfun));
426 sese region = SCOP_REGION (scop);
428 bitmap_clear (visited);
429 build_scop_bbs_1 (scop, visited, SESE_ENTRY_BB (region));
430 sbitmap_free (visited);
433 /* Return an ISL identifier for the polyhedral basic block PBB. */
435 static isl_id *
436 isl_id_for_pbb (scop_p s, poly_bb_p pbb)
438 char name[50];
439 snprintf (name, sizeof (name), "S_%d", pbb_index (pbb));
440 return isl_id_alloc (s->ctx, name, pbb);
443 /* Converts the STATIC_SCHEDULE of PBB into a scattering polyhedron.
444 We generate SCATTERING_DIMENSIONS scattering dimensions.
446 CLooG 0.15.0 and previous versions require, that all
447 scattering functions of one CloogProgram have the same number of
448 scattering dimensions, therefore we allow to specify it. This
449 should be removed in future versions of CLooG.
451 The scattering polyhedron consists of these dimensions: scattering,
452 loop_iterators, parameters.
454 Example:
456 | scattering_dimensions = 5
457 | used_scattering_dimensions = 3
458 | nb_iterators = 1
459 | scop_nb_params = 2
461 | Schedule:
463 | 4 5
465 | Scattering polyhedron:
467 | scattering: {s1, s2, s3, s4, s5}
468 | loop_iterators: {i}
469 | parameters: {p1, p2}
471 | s1 s2 s3 s4 s5 i p1 p2 1
472 | 1 0 0 0 0 0 0 0 -4 = 0
473 | 0 1 0 0 0 -1 0 0 0 = 0
474 | 0 0 1 0 0 0 0 0 -5 = 0 */
476 static void
477 build_pbb_scattering_polyhedrons (isl_aff *static_sched,
478 poly_bb_p pbb, int scattering_dimensions)
480 int i;
481 int nb_iterators = pbb_dim_iter_domain (pbb);
482 int used_scattering_dimensions = nb_iterators * 2 + 1;
483 isl_val *val;
484 isl_space *dc, *dm;
486 gcc_assert (scattering_dimensions >= used_scattering_dimensions);
488 dc = isl_set_get_space (pbb->domain);
489 dm = isl_space_add_dims (isl_space_from_domain (dc),
490 isl_dim_out, scattering_dimensions);
491 pbb->schedule = isl_map_universe (dm);
493 for (i = 0; i < scattering_dimensions; i++)
495 /* Textual order inside this loop. */
496 if ((i % 2) == 0)
498 isl_constraint *c = isl_equality_alloc
499 (isl_local_space_from_space (isl_map_get_space (pbb->schedule)));
501 val = isl_aff_get_coefficient_val (static_sched, isl_dim_in, i / 2);
502 gcc_assert (val && isl_val_is_int (val));
504 val = isl_val_neg (val);
505 c = isl_constraint_set_constant_val (c, val);
506 c = isl_constraint_set_coefficient_si (c, isl_dim_out, i, 1);
507 pbb->schedule = isl_map_add_constraint (pbb->schedule, c);
510 /* Iterations of this loop. */
511 else /* if ((i % 2) == 1) */
513 int loop = (i - 1) / 2;
514 pbb->schedule = isl_map_equate (pbb->schedule, isl_dim_in, loop,
515 isl_dim_out, i);
519 pbb->transformed = isl_map_copy (pbb->schedule);
522 /* Build for BB the static schedule.
524 The static schedule is a Dewey numbering of the abstract syntax
525 tree: http://en.wikipedia.org/wiki/Dewey_Decimal_Classification
527 The following example informally defines the static schedule:
530 for (i: ...)
532 for (j: ...)
538 for (k: ...)
546 Static schedules for A to F:
548 DEPTH
549 0 1 2
551 B 1 0 0
552 C 1 0 1
553 D 1 1 0
554 E 1 1 1
558 static void
559 build_scop_scattering (scop_p scop)
561 int i;
562 poly_bb_p pbb;
563 gimple_bb_p previous_gbb = NULL;
564 isl_space *dc = isl_set_get_space (scop->context);
565 isl_aff *static_sched;
567 dc = isl_space_add_dims (dc, isl_dim_set, number_of_loops (cfun));
568 static_sched = isl_aff_zero_on_domain (isl_local_space_from_space (dc));
570 /* We have to start schedules at 0 on the first component and
571 because we cannot compare_prefix_loops against a previous loop,
572 prefix will be equal to zero, and that index will be
573 incremented before copying. */
574 static_sched = isl_aff_add_coefficient_si (static_sched, isl_dim_in, 0, -1);
576 FOR_EACH_VEC_ELT (SCOP_BBS (scop), i, pbb)
578 gimple_bb_p gbb = PBB_BLACK_BOX (pbb);
579 int prefix;
580 int nb_scat_dims = pbb_dim_iter_domain (pbb) * 2 + 1;
582 if (previous_gbb)
583 prefix = nb_common_loops (SCOP_REGION (scop), previous_gbb, gbb);
584 else
585 prefix = 0;
587 previous_gbb = gbb;
589 static_sched = isl_aff_add_coefficient_si (static_sched, isl_dim_in,
590 prefix, 1);
591 build_pbb_scattering_polyhedrons (static_sched, pbb, nb_scat_dims);
594 isl_aff_free (static_sched);
597 static isl_pw_aff *extract_affine (scop_p, tree, __isl_take isl_space *space);
599 /* Extract an affine expression from the chain of recurrence E. */
601 static isl_pw_aff *
602 extract_affine_chrec (scop_p s, tree e, __isl_take isl_space *space)
604 isl_pw_aff *lhs = extract_affine (s, CHREC_LEFT (e), isl_space_copy (space));
605 isl_pw_aff *rhs = extract_affine (s, CHREC_RIGHT (e), isl_space_copy (space));
606 isl_local_space *ls = isl_local_space_from_space (space);
607 unsigned pos = sese_loop_depth ((sese) s->region, get_chrec_loop (e)) - 1;
608 isl_aff *loop = isl_aff_set_coefficient_si
609 (isl_aff_zero_on_domain (ls), isl_dim_in, pos, 1);
610 isl_pw_aff *l = isl_pw_aff_from_aff (loop);
612 /* Before multiplying, make sure that the result is affine. */
613 gcc_assert (isl_pw_aff_is_cst (rhs)
614 || isl_pw_aff_is_cst (l));
616 return isl_pw_aff_add (lhs, isl_pw_aff_mul (rhs, l));
619 /* Extract an affine expression from the mult_expr E. */
621 static isl_pw_aff *
622 extract_affine_mul (scop_p s, tree e, __isl_take isl_space *space)
624 isl_pw_aff *lhs = extract_affine (s, TREE_OPERAND (e, 0),
625 isl_space_copy (space));
626 isl_pw_aff *rhs = extract_affine (s, TREE_OPERAND (e, 1), space);
628 if (!isl_pw_aff_is_cst (lhs)
629 && !isl_pw_aff_is_cst (rhs))
631 isl_pw_aff_free (lhs);
632 isl_pw_aff_free (rhs);
633 return NULL;
636 return isl_pw_aff_mul (lhs, rhs);
639 /* Return an ISL identifier from the name of the ssa_name E. */
641 static isl_id *
642 isl_id_for_ssa_name (scop_p s, tree e)
644 const char *name = get_name (e);
645 isl_id *id;
647 if (name)
648 id = isl_id_alloc (s->ctx, name, e);
649 else
651 char name1[50];
652 snprintf (name1, sizeof (name1), "P_%d", SSA_NAME_VERSION (e));
653 id = isl_id_alloc (s->ctx, name1, e);
656 return id;
659 /* Return an ISL identifier for the data reference DR. */
661 static isl_id *
662 isl_id_for_dr (scop_p s, data_reference_p dr ATTRIBUTE_UNUSED)
664 /* Data references all get the same isl_id. They need to be comparable
665 and are distinguished through the first dimension, which contains the
666 alias set number. */
667 return isl_id_alloc (s->ctx, "", 0);
670 /* Extract an affine expression from the ssa_name E. */
672 static isl_pw_aff *
673 extract_affine_name (scop_p s, tree e, __isl_take isl_space *space)
675 isl_aff *aff;
676 isl_set *dom;
677 isl_id *id;
678 int dimension;
680 id = isl_id_for_ssa_name (s, e);
681 dimension = isl_space_find_dim_by_id (space, isl_dim_param, id);
682 isl_id_free (id);
683 dom = isl_set_universe (isl_space_copy (space));
684 aff = isl_aff_zero_on_domain (isl_local_space_from_space (space));
685 aff = isl_aff_add_coefficient_si (aff, isl_dim_param, dimension, 1);
686 return isl_pw_aff_alloc (dom, aff);
689 /* Extract an affine expression from the gmp constant G. */
691 static isl_pw_aff *
692 extract_affine_gmp (mpz_t g, __isl_take isl_space *space)
694 isl_local_space *ls = isl_local_space_from_space (isl_space_copy (space));
695 isl_aff *aff = isl_aff_zero_on_domain (ls);
696 isl_set *dom = isl_set_universe (space);
697 isl_val *v;
698 isl_ctx *ct;
700 ct = isl_aff_get_ctx (aff);
701 v = isl_val_int_from_gmp (ct, g);
702 aff = isl_aff_add_constant_val (aff, v);
704 return isl_pw_aff_alloc (dom, aff);
707 /* Extract an affine expression from the integer_cst E. */
709 static isl_pw_aff *
710 extract_affine_int (tree e, __isl_take isl_space *space)
712 isl_pw_aff *res;
713 mpz_t g;
715 mpz_init (g);
716 tree_int_to_gmp (e, g);
717 res = extract_affine_gmp (g, space);
718 mpz_clear (g);
720 return res;
723 /* Compute pwaff mod 2^width. */
725 static isl_pw_aff *
726 wrap (isl_pw_aff *pwaff, unsigned width)
728 isl_val *mod;
730 mod = isl_val_int_from_ui (isl_pw_aff_get_ctx (pwaff), width);
731 mod = isl_val_2exp (mod);
732 pwaff = isl_pw_aff_mod_val (pwaff, mod);
734 return pwaff;
737 /* When parameter NAME is in REGION, returns its index in SESE_PARAMS.
738 Otherwise returns -1. */
740 static inline int
741 parameter_index_in_region_1 (tree name, sese region)
743 int i;
744 tree p;
746 gcc_assert (TREE_CODE (name) == SSA_NAME);
748 FOR_EACH_VEC_ELT (SESE_PARAMS (region), i, p)
749 if (p == name)
750 return i;
752 return -1;
755 /* When the parameter NAME is in REGION, returns its index in
756 SESE_PARAMS. Otherwise this function inserts NAME in SESE_PARAMS
757 and returns the index of NAME. */
759 static int
760 parameter_index_in_region (tree name, sese region)
762 int i;
764 gcc_assert (TREE_CODE (name) == SSA_NAME);
766 /* Cannot constrain on anything else than INTEGER_TYPE parameters. */
767 if (TREE_CODE (TREE_TYPE (name)) != INTEGER_TYPE)
768 return -1;
770 i = parameter_index_in_region_1 (name, region);
771 if (i != -1)
772 return i;
774 gcc_assert (SESE_ADD_PARAMS (region));
776 i = SESE_PARAMS (region).length ();
777 SESE_PARAMS (region).safe_push (name);
778 return i;
781 /* Extract an affine expression from the tree E in the scop S. */
783 static isl_pw_aff *
784 extract_affine (scop_p s, tree e, __isl_take isl_space *space)
786 isl_pw_aff *lhs, *rhs, *res;
787 tree type;
789 if (e == chrec_dont_know) {
790 isl_space_free (space);
791 return NULL;
794 switch (TREE_CODE (e))
796 case POLYNOMIAL_CHREC:
797 res = extract_affine_chrec (s, e, space);
798 break;
800 case MULT_EXPR:
801 res = extract_affine_mul (s, e, space);
802 break;
804 case PLUS_EXPR:
805 case POINTER_PLUS_EXPR:
806 lhs = extract_affine (s, TREE_OPERAND (e, 0), isl_space_copy (space));
807 rhs = extract_affine (s, TREE_OPERAND (e, 1), space);
808 res = isl_pw_aff_add (lhs, rhs);
809 break;
811 case MINUS_EXPR:
812 lhs = extract_affine (s, TREE_OPERAND (e, 0), isl_space_copy (space));
813 rhs = extract_affine (s, TREE_OPERAND (e, 1), space);
814 res = isl_pw_aff_sub (lhs, rhs);
815 break;
817 case NEGATE_EXPR:
818 case BIT_NOT_EXPR:
819 lhs = extract_affine (s, TREE_OPERAND (e, 0), isl_space_copy (space));
820 rhs = extract_affine (s, integer_minus_one_node, space);
821 res = isl_pw_aff_mul (lhs, rhs);
822 break;
824 case SSA_NAME:
825 gcc_assert (-1 != parameter_index_in_region_1 (e, SCOP_REGION (s)));
826 res = extract_affine_name (s, e, space);
827 break;
829 case INTEGER_CST:
830 res = extract_affine_int (e, space);
831 /* No need to wrap a single integer. */
832 return res;
834 CASE_CONVERT:
835 case NON_LVALUE_EXPR:
836 res = extract_affine (s, TREE_OPERAND (e, 0), space);
837 break;
839 default:
840 gcc_unreachable ();
841 break;
844 type = TREE_TYPE (e);
845 if (TYPE_UNSIGNED (type))
846 res = wrap (res, TYPE_PRECISION (type));
848 return res;
851 /* In the context of sese S, scan the expression E and translate it to
852 a linear expression C. When parsing a symbolic multiplication, K
853 represents the constant multiplier of an expression containing
854 parameters. */
856 static void
857 scan_tree_for_params (sese s, tree e)
859 if (e == chrec_dont_know)
860 return;
862 switch (TREE_CODE (e))
864 case POLYNOMIAL_CHREC:
865 scan_tree_for_params (s, CHREC_LEFT (e));
866 break;
868 case MULT_EXPR:
869 if (chrec_contains_symbols (TREE_OPERAND (e, 0)))
870 scan_tree_for_params (s, TREE_OPERAND (e, 0));
871 else
872 scan_tree_for_params (s, TREE_OPERAND (e, 1));
873 break;
875 case PLUS_EXPR:
876 case POINTER_PLUS_EXPR:
877 case MINUS_EXPR:
878 scan_tree_for_params (s, TREE_OPERAND (e, 0));
879 scan_tree_for_params (s, TREE_OPERAND (e, 1));
880 break;
882 case NEGATE_EXPR:
883 case BIT_NOT_EXPR:
884 CASE_CONVERT:
885 case NON_LVALUE_EXPR:
886 scan_tree_for_params (s, TREE_OPERAND (e, 0));
887 break;
889 case SSA_NAME:
890 parameter_index_in_region (e, s);
891 break;
893 case INTEGER_CST:
894 case ADDR_EXPR:
895 case REAL_CST:
896 case COMPLEX_CST:
897 case VECTOR_CST:
898 break;
900 default:
901 gcc_unreachable ();
902 break;
906 /* Find parameters with respect to REGION in BB. We are looking in memory
907 access functions, conditions and loop bounds. */
909 static void
910 find_params_in_bb (sese region, gimple_bb_p gbb)
912 int i;
913 unsigned j;
914 data_reference_p dr;
915 gimple stmt;
916 loop_p loop = GBB_BB (gbb)->loop_father;
918 /* Find parameters in the access functions of data references. */
919 FOR_EACH_VEC_ELT (GBB_DATA_REFS (gbb), i, dr)
920 for (j = 0; j < DR_NUM_DIMENSIONS (dr); j++)
921 scan_tree_for_params (region, DR_ACCESS_FN (dr, j));
923 /* Find parameters in conditional statements. */
924 FOR_EACH_VEC_ELT (GBB_CONDITIONS (gbb), i, stmt)
926 tree lhs = scalar_evolution_in_region (region, loop,
927 gimple_cond_lhs (stmt));
928 tree rhs = scalar_evolution_in_region (region, loop,
929 gimple_cond_rhs (stmt));
931 scan_tree_for_params (region, lhs);
932 scan_tree_for_params (region, rhs);
936 /* Record the parameters used in the SCOP. A variable is a parameter
937 in a scop if it does not vary during the execution of that scop. */
939 static void
940 find_scop_parameters (scop_p scop)
942 poly_bb_p pbb;
943 unsigned i;
944 sese region = SCOP_REGION (scop);
945 struct loop *loop;
946 int nbp;
948 /* Find the parameters used in the loop bounds. */
949 FOR_EACH_VEC_ELT (SESE_LOOP_NEST (region), i, loop)
951 tree nb_iters = number_of_latch_executions (loop);
953 if (!chrec_contains_symbols (nb_iters))
954 continue;
956 nb_iters = scalar_evolution_in_region (region, loop, nb_iters);
957 scan_tree_for_params (region, nb_iters);
960 /* Find the parameters used in data accesses. */
961 FOR_EACH_VEC_ELT (SCOP_BBS (scop), i, pbb)
962 find_params_in_bb (region, PBB_BLACK_BOX (pbb));
964 nbp = sese_nb_params (region);
965 scop_set_nb_params (scop, nbp);
966 SESE_ADD_PARAMS (region) = false;
969 tree e;
970 isl_space *space = isl_space_set_alloc (scop->ctx, nbp, 0);
972 FOR_EACH_VEC_ELT (SESE_PARAMS (region), i, e)
973 space = isl_space_set_dim_id (space, isl_dim_param, i,
974 isl_id_for_ssa_name (scop, e));
976 scop->context = isl_set_universe (space);
980 /* Builds the constraint polyhedra for LOOP in SCOP. OUTER_PH gives
981 the constraints for the surrounding loops. */
983 static void
984 build_loop_iteration_domains (scop_p scop, struct loop *loop,
985 int nb,
986 isl_set *outer, isl_set **doms)
988 tree nb_iters = number_of_latch_executions (loop);
989 sese region = SCOP_REGION (scop);
991 isl_set *inner = isl_set_copy (outer);
992 isl_space *space;
993 isl_constraint *c;
994 int pos = isl_set_dim (outer, isl_dim_set);
995 isl_val *v;
996 mpz_t g;
998 mpz_init (g);
1000 inner = isl_set_add_dims (inner, isl_dim_set, 1);
1001 space = isl_set_get_space (inner);
1003 /* 0 <= loop_i */
1004 c = isl_inequality_alloc
1005 (isl_local_space_from_space (isl_space_copy (space)));
1006 c = isl_constraint_set_coefficient_si (c, isl_dim_set, pos, 1);
1007 inner = isl_set_add_constraint (inner, c);
1009 /* loop_i <= cst_nb_iters */
1010 if (TREE_CODE (nb_iters) == INTEGER_CST)
1012 c = isl_inequality_alloc
1013 (isl_local_space_from_space (isl_space_copy (space)));
1014 c = isl_constraint_set_coefficient_si (c, isl_dim_set, pos, -1);
1015 tree_int_to_gmp (nb_iters, g);
1016 v = isl_val_int_from_gmp (scop->ctx, g);
1017 c = isl_constraint_set_constant_val (c, v);
1018 inner = isl_set_add_constraint (inner, c);
1021 /* loop_i <= expr_nb_iters */
1022 else if (!chrec_contains_undetermined (nb_iters))
1024 widest_int nit;
1025 isl_pw_aff *aff;
1026 isl_set *valid;
1027 isl_local_space *ls;
1028 isl_aff *al;
1029 isl_set *le;
1031 nb_iters = scalar_evolution_in_region (region, loop, nb_iters);
1033 aff = extract_affine (scop, nb_iters, isl_set_get_space (inner));
1034 valid = isl_pw_aff_nonneg_set (isl_pw_aff_copy (aff));
1035 valid = isl_set_project_out (valid, isl_dim_set, 0,
1036 isl_set_dim (valid, isl_dim_set));
1037 scop->context = isl_set_intersect (scop->context, valid);
1039 ls = isl_local_space_from_space (isl_space_copy (space));
1040 al = isl_aff_set_coefficient_si (isl_aff_zero_on_domain (ls),
1041 isl_dim_in, pos, 1);
1042 le = isl_pw_aff_le_set (isl_pw_aff_from_aff (al),
1043 isl_pw_aff_copy (aff));
1044 inner = isl_set_intersect (inner, le);
1046 if (max_stmt_executions (loop, &nit))
1048 /* Insert in the context the constraints from the
1049 estimation of the number of iterations NIT and the
1050 symbolic number of iterations (involving parameter
1051 names) NB_ITERS. First, build the affine expression
1052 "NIT - NB_ITERS" and then say that it is positive,
1053 i.e., NIT approximates NB_ITERS: "NIT >= NB_ITERS". */
1054 isl_pw_aff *approx;
1055 mpz_t g;
1056 isl_set *x;
1057 isl_constraint *c;
1059 mpz_init (g);
1060 wi::to_mpz (nit, g, SIGNED);
1061 mpz_sub_ui (g, g, 1);
1062 approx = extract_affine_gmp (g, isl_set_get_space (inner));
1063 x = isl_pw_aff_ge_set (approx, aff);
1064 x = isl_set_project_out (x, isl_dim_set, 0,
1065 isl_set_dim (x, isl_dim_set));
1066 scop->context = isl_set_intersect (scop->context, x);
1068 c = isl_inequality_alloc
1069 (isl_local_space_from_space (isl_space_copy (space)));
1070 c = isl_constraint_set_coefficient_si (c, isl_dim_set, pos, -1);
1071 v = isl_val_int_from_gmp (scop->ctx, g);
1072 mpz_clear (g);
1073 c = isl_constraint_set_constant_val (c, v);
1074 inner = isl_set_add_constraint (inner, c);
1076 else
1077 isl_pw_aff_free (aff);
1079 else
1080 gcc_unreachable ();
1082 if (loop->inner && loop_in_sese_p (loop->inner, region))
1083 build_loop_iteration_domains (scop, loop->inner, nb + 1,
1084 isl_set_copy (inner), doms);
1086 if (nb != 0
1087 && loop->next
1088 && loop_in_sese_p (loop->next, region))
1089 build_loop_iteration_domains (scop, loop->next, nb,
1090 isl_set_copy (outer), doms);
1092 doms[loop->num] = inner;
1094 isl_set_free (outer);
1095 isl_space_free (space);
1096 mpz_clear (g);
1099 /* Returns a linear expression for tree T evaluated in PBB. */
1101 static isl_pw_aff *
1102 create_pw_aff_from_tree (poly_bb_p pbb, tree t)
1104 scop_p scop = PBB_SCOP (pbb);
1106 t = scalar_evolution_in_region (SCOP_REGION (scop), pbb_loop (pbb), t);
1107 gcc_assert (!automatically_generated_chrec_p (t));
1109 return extract_affine (scop, t, isl_set_get_space (pbb->domain));
1112 /* Add conditional statement STMT to pbb. CODE is used as the comparison
1113 operator. This allows us to invert the condition or to handle
1114 inequalities. */
1116 static void
1117 add_condition_to_pbb (poly_bb_p pbb, gcond *stmt, enum tree_code code)
1119 isl_pw_aff *lhs = create_pw_aff_from_tree (pbb, gimple_cond_lhs (stmt));
1120 isl_pw_aff *rhs = create_pw_aff_from_tree (pbb, gimple_cond_rhs (stmt));
1121 isl_set *cond;
1123 switch (code)
1125 case LT_EXPR:
1126 cond = isl_pw_aff_lt_set (lhs, rhs);
1127 break;
1129 case GT_EXPR:
1130 cond = isl_pw_aff_gt_set (lhs, rhs);
1131 break;
1133 case LE_EXPR:
1134 cond = isl_pw_aff_le_set (lhs, rhs);
1135 break;
1137 case GE_EXPR:
1138 cond = isl_pw_aff_ge_set (lhs, rhs);
1139 break;
1141 case EQ_EXPR:
1142 cond = isl_pw_aff_eq_set (lhs, rhs);
1143 break;
1145 case NE_EXPR:
1146 cond = isl_pw_aff_ne_set (lhs, rhs);
1147 break;
1149 default:
1150 isl_pw_aff_free (lhs);
1151 isl_pw_aff_free (rhs);
1152 return;
1155 cond = isl_set_coalesce (cond);
1156 cond = isl_set_set_tuple_id (cond, isl_set_get_tuple_id (pbb->domain));
1157 pbb->domain = isl_set_intersect (pbb->domain, cond);
1160 /* Add conditions to the domain of PBB. */
1162 static void
1163 add_conditions_to_domain (poly_bb_p pbb)
1165 unsigned int i;
1166 gimple stmt;
1167 gimple_bb_p gbb = PBB_BLACK_BOX (pbb);
1169 if (GBB_CONDITIONS (gbb).is_empty ())
1170 return;
1172 FOR_EACH_VEC_ELT (GBB_CONDITIONS (gbb), i, stmt)
1173 switch (gimple_code (stmt))
1175 case GIMPLE_COND:
1177 /* Don't constrain on anything else than INTEGER_TYPE. */
1178 if (TREE_CODE (TREE_TYPE (gimple_cond_lhs (stmt))) != INTEGER_TYPE)
1179 break;
1181 gcond *cond_stmt = as_a <gcond *> (stmt);
1182 enum tree_code code = gimple_cond_code (cond_stmt);
1184 /* The conditions for ELSE-branches are inverted. */
1185 if (!GBB_CONDITION_CASES (gbb)[i])
1186 code = invert_tree_comparison (code, false);
1188 add_condition_to_pbb (pbb, cond_stmt, code);
1189 break;
1192 case GIMPLE_SWITCH:
1193 /* Switch statements are not supported right now - fall through. */
1195 default:
1196 gcc_unreachable ();
1197 break;
1201 /* Traverses all the GBBs of the SCOP and add their constraints to the
1202 iteration domains. */
1204 static void
1205 add_conditions_to_constraints (scop_p scop)
1207 int i;
1208 poly_bb_p pbb;
1210 FOR_EACH_VEC_ELT (SCOP_BBS (scop), i, pbb)
1211 add_conditions_to_domain (pbb);
1214 /* Returns a COND_EXPR statement when BB has a single predecessor, the
1215 edge between BB and its predecessor is not a loop exit edge, and
1216 the last statement of the single predecessor is a COND_EXPR. */
1218 static gcond *
1219 single_pred_cond_non_loop_exit (basic_block bb)
1221 if (single_pred_p (bb))
1223 edge e = single_pred_edge (bb);
1224 basic_block pred = e->src;
1225 gimple stmt;
1227 if (loop_depth (pred->loop_father) > loop_depth (bb->loop_father))
1228 return NULL;
1230 stmt = last_stmt (pred);
1232 if (stmt && gimple_code (stmt) == GIMPLE_COND)
1233 return as_a <gcond *> (stmt);
1236 return NULL;
1239 class sese_dom_walker : public dom_walker
1241 public:
1242 sese_dom_walker (cdi_direction, sese);
1244 virtual void before_dom_children (basic_block);
1245 virtual void after_dom_children (basic_block);
1247 private:
1248 auto_vec<gimple, 3> m_conditions, m_cases;
1249 sese m_region;
1252 sese_dom_walker::sese_dom_walker (cdi_direction direction, sese region)
1253 : dom_walker (direction), m_region (region)
1257 /* Call-back for dom_walk executed before visiting the dominated
1258 blocks. */
1260 void
1261 sese_dom_walker::before_dom_children (basic_block bb)
1263 gimple_bb_p gbb;
1264 gcond *stmt;
1266 if (!bb_in_sese_p (bb, m_region))
1267 return;
1269 stmt = single_pred_cond_non_loop_exit (bb);
1271 if (stmt)
1273 edge e = single_pred_edge (bb);
1275 m_conditions.safe_push (stmt);
1277 if (e->flags & EDGE_TRUE_VALUE)
1278 m_cases.safe_push (stmt);
1279 else
1280 m_cases.safe_push (NULL);
1283 gbb = gbb_from_bb (bb);
1285 if (gbb)
1287 GBB_CONDITIONS (gbb) = m_conditions.copy ();
1288 GBB_CONDITION_CASES (gbb) = m_cases.copy ();
1292 /* Call-back for dom_walk executed after visiting the dominated
1293 blocks. */
1295 void
1296 sese_dom_walker::after_dom_children (basic_block bb)
1298 if (!bb_in_sese_p (bb, m_region))
1299 return;
1301 if (single_pred_cond_non_loop_exit (bb))
1303 m_conditions.pop ();
1304 m_cases.pop ();
1308 /* Add constraints on the possible values of parameter P from the type
1309 of P. */
1311 static void
1312 add_param_constraints (scop_p scop, graphite_dim_t p)
1314 tree parameter = SESE_PARAMS (SCOP_REGION (scop))[p];
1315 tree type = TREE_TYPE (parameter);
1316 tree lb = NULL_TREE;
1317 tree ub = NULL_TREE;
1319 if (POINTER_TYPE_P (type) || !TYPE_MIN_VALUE (type))
1320 lb = lower_bound_in_type (type, type);
1321 else
1322 lb = TYPE_MIN_VALUE (type);
1324 if (POINTER_TYPE_P (type) || !TYPE_MAX_VALUE (type))
1325 ub = upper_bound_in_type (type, type);
1326 else
1327 ub = TYPE_MAX_VALUE (type);
1329 if (lb)
1331 isl_space *space = isl_set_get_space (scop->context);
1332 isl_constraint *c;
1333 mpz_t g;
1334 isl_val *v;
1336 c = isl_inequality_alloc (isl_local_space_from_space (space));
1337 mpz_init (g);
1338 tree_int_to_gmp (lb, g);
1339 v = isl_val_int_from_gmp (scop->ctx, g);
1340 v = isl_val_neg (v);
1341 mpz_clear (g);
1342 c = isl_constraint_set_constant_val (c, v);
1343 c = isl_constraint_set_coefficient_si (c, isl_dim_param, p, 1);
1345 scop->context = isl_set_add_constraint (scop->context, c);
1348 if (ub)
1350 isl_space *space = isl_set_get_space (scop->context);
1351 isl_constraint *c;
1352 mpz_t g;
1353 isl_val *v;
1355 c = isl_inequality_alloc (isl_local_space_from_space (space));
1357 mpz_init (g);
1358 tree_int_to_gmp (ub, g);
1359 v = isl_val_int_from_gmp (scop->ctx, g);
1360 mpz_clear (g);
1361 c = isl_constraint_set_constant_val (c, v);
1362 c = isl_constraint_set_coefficient_si (c, isl_dim_param, p, -1);
1364 scop->context = isl_set_add_constraint (scop->context, c);
1368 /* Build the context of the SCOP. The context usually contains extra
1369 constraints that are added to the iteration domains that constrain
1370 some parameters. */
1372 static void
1373 build_scop_context (scop_p scop)
1375 graphite_dim_t p, n = scop_nb_params (scop);
1377 for (p = 0; p < n; p++)
1378 add_param_constraints (scop, p);
1381 /* Build the iteration domains: the loops belonging to the current
1382 SCOP, and that vary for the execution of the current basic block.
1383 Returns false if there is no loop in SCOP. */
1385 static void
1386 build_scop_iteration_domain (scop_p scop)
1388 struct loop *loop;
1389 sese region = SCOP_REGION (scop);
1390 int i;
1391 poly_bb_p pbb;
1392 int nb_loops = number_of_loops (cfun);
1393 isl_set **doms = XCNEWVEC (isl_set *, nb_loops);
1395 FOR_EACH_VEC_ELT (SESE_LOOP_NEST (region), i, loop)
1396 if (!loop_in_sese_p (loop_outer (loop), region))
1397 build_loop_iteration_domains (scop, loop, 0,
1398 isl_set_copy (scop->context), doms);
1400 FOR_EACH_VEC_ELT (SCOP_BBS (scop), i, pbb)
1402 loop = pbb_loop (pbb);
1404 if (doms[loop->num])
1405 pbb->domain = isl_set_copy (doms[loop->num]);
1406 else
1407 pbb->domain = isl_set_copy (scop->context);
1409 pbb->domain = isl_set_set_tuple_id (pbb->domain,
1410 isl_id_for_pbb (scop, pbb));
1413 for (i = 0; i < nb_loops; i++)
1414 if (doms[i])
1415 isl_set_free (doms[i]);
1417 free (doms);
1420 /* Add a constrain to the ACCESSES polyhedron for the alias set of
1421 data reference DR. ACCESSP_NB_DIMS is the dimension of the
1422 ACCESSES polyhedron, DOM_NB_DIMS is the dimension of the iteration
1423 domain. */
1425 static isl_map *
1426 pdr_add_alias_set (isl_map *acc, data_reference_p dr)
1428 isl_constraint *c;
1429 int alias_set_num = 0;
1430 base_alias_pair *bap = (base_alias_pair *)(dr->aux);
1432 if (bap && bap->alias_set)
1433 alias_set_num = *(bap->alias_set);
1435 c = isl_equality_alloc
1436 (isl_local_space_from_space (isl_map_get_space (acc)));
1437 c = isl_constraint_set_constant_si (c, -alias_set_num);
1438 c = isl_constraint_set_coefficient_si (c, isl_dim_out, 0, 1);
1440 return isl_map_add_constraint (acc, c);
1443 /* Assign the affine expression INDEX to the output dimension POS of
1444 MAP and return the result. */
1446 static isl_map *
1447 set_index (isl_map *map, int pos, isl_pw_aff *index)
1449 isl_map *index_map;
1450 int len = isl_map_dim (map, isl_dim_out);
1451 isl_id *id;
1453 index_map = isl_map_from_pw_aff (index);
1454 index_map = isl_map_insert_dims (index_map, isl_dim_out, 0, pos);
1455 index_map = isl_map_add_dims (index_map, isl_dim_out, len - pos - 1);
1457 id = isl_map_get_tuple_id (map, isl_dim_out);
1458 index_map = isl_map_set_tuple_id (index_map, isl_dim_out, id);
1459 id = isl_map_get_tuple_id (map, isl_dim_in);
1460 index_map = isl_map_set_tuple_id (index_map, isl_dim_in, id);
1462 return isl_map_intersect (map, index_map);
1465 /* Add to ACCESSES polyhedron equalities defining the access functions
1466 to the memory. ACCESSP_NB_DIMS is the dimension of the ACCESSES
1467 polyhedron, DOM_NB_DIMS is the dimension of the iteration domain.
1468 PBB is the poly_bb_p that contains the data reference DR. */
1470 static isl_map *
1471 pdr_add_memory_accesses (isl_map *acc, data_reference_p dr, poly_bb_p pbb)
1473 int i, nb_subscripts = DR_NUM_DIMENSIONS (dr);
1474 scop_p scop = PBB_SCOP (pbb);
1476 for (i = 0; i < nb_subscripts; i++)
1478 isl_pw_aff *aff;
1479 tree afn = DR_ACCESS_FN (dr, nb_subscripts - 1 - i);
1481 aff = extract_affine (scop, afn,
1482 isl_space_domain (isl_map_get_space (acc)));
1483 acc = set_index (acc, i + 1, aff);
1486 return acc;
1489 /* Add constrains representing the size of the accessed data to the
1490 ACCESSES polyhedron. ACCESSP_NB_DIMS is the dimension of the
1491 ACCESSES polyhedron, DOM_NB_DIMS is the dimension of the iteration
1492 domain. */
1494 static isl_set *
1495 pdr_add_data_dimensions (isl_set *subscript_sizes, scop_p scop,
1496 data_reference_p dr)
1498 tree ref = DR_REF (dr);
1500 int nb_subscripts = DR_NUM_DIMENSIONS (dr);
1501 for (int i = nb_subscripts - 1; i >= 0; i--, ref = TREE_OPERAND (ref, 0))
1503 if (TREE_CODE (ref) != ARRAY_REF)
1504 return subscript_sizes;
1506 tree low = array_ref_low_bound (ref);
1507 tree high = array_ref_up_bound (ref);
1509 /* XXX The PPL code dealt separately with
1510 subscript - low >= 0 and high - subscript >= 0 in case one of
1511 the two bounds isn't known. Do the same here? */
1513 if (tree_fits_shwi_p (low)
1514 && high
1515 && tree_fits_shwi_p (high)
1516 /* 1-element arrays at end of structures may extend over
1517 their declared size. */
1518 && !(array_at_struct_end_p (ref)
1519 && operand_equal_p (low, high, 0)))
1521 isl_id *id;
1522 isl_aff *aff;
1523 isl_set *univ, *lbs, *ubs;
1524 isl_pw_aff *index;
1525 isl_set *valid;
1526 isl_space *space = isl_set_get_space (subscript_sizes);
1527 isl_pw_aff *lb = extract_affine_int (low, isl_space_copy (space));
1528 isl_pw_aff *ub = extract_affine_int (high, isl_space_copy (space));
1530 /* high >= 0 */
1531 valid = isl_pw_aff_nonneg_set (isl_pw_aff_copy (ub));
1532 valid = isl_set_project_out (valid, isl_dim_set, 0,
1533 isl_set_dim (valid, isl_dim_set));
1534 scop->context = isl_set_intersect (scop->context, valid);
1536 aff = isl_aff_zero_on_domain (isl_local_space_from_space (space));
1537 aff = isl_aff_add_coefficient_si (aff, isl_dim_in, i + 1, 1);
1538 univ = isl_set_universe (isl_space_domain (isl_aff_get_space (aff)));
1539 index = isl_pw_aff_alloc (univ, aff);
1541 id = isl_set_get_tuple_id (subscript_sizes);
1542 lb = isl_pw_aff_set_tuple_id (lb, isl_dim_in, isl_id_copy (id));
1543 ub = isl_pw_aff_set_tuple_id (ub, isl_dim_in, id);
1545 /* low <= sub_i <= high */
1546 lbs = isl_pw_aff_ge_set (isl_pw_aff_copy (index), lb);
1547 ubs = isl_pw_aff_le_set (index, ub);
1548 subscript_sizes = isl_set_intersect (subscript_sizes, lbs);
1549 subscript_sizes = isl_set_intersect (subscript_sizes, ubs);
1553 return subscript_sizes;
1556 /* Build data accesses for DR in PBB. */
1558 static void
1559 build_poly_dr (data_reference_p dr, poly_bb_p pbb)
1561 int dr_base_object_set;
1562 isl_map *acc;
1563 isl_set *subscript_sizes;
1564 scop_p scop = PBB_SCOP (pbb);
1567 isl_space *dc = isl_set_get_space (pbb->domain);
1568 int nb_out = 1 + DR_NUM_DIMENSIONS (dr);
1569 isl_space *space = isl_space_add_dims (isl_space_from_domain (dc),
1570 isl_dim_out, nb_out);
1572 acc = isl_map_universe (space);
1573 acc = isl_map_set_tuple_id (acc, isl_dim_out, isl_id_for_dr (scop, dr));
1576 acc = pdr_add_alias_set (acc, dr);
1577 acc = pdr_add_memory_accesses (acc, dr, pbb);
1580 isl_id *id = isl_id_for_dr (scop, dr);
1581 int nb = 1 + DR_NUM_DIMENSIONS (dr);
1582 isl_space *space = isl_space_set_alloc (scop->ctx, 0, nb);
1583 int alias_set_num = 0;
1584 base_alias_pair *bap = (base_alias_pair *)(dr->aux);
1586 if (bap && bap->alias_set)
1587 alias_set_num = *(bap->alias_set);
1589 space = isl_space_set_tuple_id (space, isl_dim_set, id);
1590 subscript_sizes = isl_set_nat_universe (space);
1591 subscript_sizes = isl_set_fix_si (subscript_sizes, isl_dim_set, 0,
1592 alias_set_num);
1593 subscript_sizes = pdr_add_data_dimensions (subscript_sizes, scop, dr);
1596 gcc_assert (dr->aux);
1597 dr_base_object_set = ((base_alias_pair *)(dr->aux))->base_obj_set;
1599 new_poly_dr (pbb, dr_base_object_set,
1600 DR_IS_READ (dr) ? PDR_READ : PDR_WRITE,
1601 dr, DR_NUM_DIMENSIONS (dr), acc, subscript_sizes);
1604 /* Write to FILE the alias graph of data references in DIMACS format. */
1606 static inline bool
1607 write_alias_graph_to_ascii_dimacs (FILE *file, char *comment,
1608 vec<data_reference_p> drs)
1610 int num_vertex = drs.length ();
1611 int edge_num = 0;
1612 data_reference_p dr1, dr2;
1613 int i, j;
1615 if (num_vertex == 0)
1616 return true;
1618 FOR_EACH_VEC_ELT (drs, i, dr1)
1619 for (j = i + 1; drs.iterate (j, &dr2); j++)
1620 if (dr_may_alias_p (dr1, dr2, true))
1621 edge_num++;
1623 fprintf (file, "$\n");
1625 if (comment)
1626 fprintf (file, "c %s\n", comment);
1628 fprintf (file, "p edge %d %d\n", num_vertex, edge_num);
1630 FOR_EACH_VEC_ELT (drs, i, dr1)
1631 for (j = i + 1; drs.iterate (j, &dr2); j++)
1632 if (dr_may_alias_p (dr1, dr2, true))
1633 fprintf (file, "e %d %d\n", i + 1, j + 1);
1635 return true;
1638 /* Write to FILE the alias graph of data references in DOT format. */
1640 static inline bool
1641 write_alias_graph_to_ascii_dot (FILE *file, char *comment,
1642 vec<data_reference_p> drs)
1644 int num_vertex = drs.length ();
1645 data_reference_p dr1, dr2;
1646 int i, j;
1648 if (num_vertex == 0)
1649 return true;
1651 fprintf (file, "$\n");
1653 if (comment)
1654 fprintf (file, "c %s\n", comment);
1656 /* First print all the vertices. */
1657 FOR_EACH_VEC_ELT (drs, i, dr1)
1658 fprintf (file, "n%d;\n", i);
1660 FOR_EACH_VEC_ELT (drs, i, dr1)
1661 for (j = i + 1; drs.iterate (j, &dr2); j++)
1662 if (dr_may_alias_p (dr1, dr2, true))
1663 fprintf (file, "n%d n%d\n", i, j);
1665 return true;
1668 /* Write to FILE the alias graph of data references in ECC format. */
1670 static inline bool
1671 write_alias_graph_to_ascii_ecc (FILE *file, char *comment,
1672 vec<data_reference_p> drs)
1674 int num_vertex = drs.length ();
1675 data_reference_p dr1, dr2;
1676 int i, j;
1678 if (num_vertex == 0)
1679 return true;
1681 fprintf (file, "$\n");
1683 if (comment)
1684 fprintf (file, "c %s\n", comment);
1686 FOR_EACH_VEC_ELT (drs, i, dr1)
1687 for (j = i + 1; drs.iterate (j, &dr2); j++)
1688 if (dr_may_alias_p (dr1, dr2, true))
1689 fprintf (file, "%d %d\n", i, j);
1691 return true;
1694 /* Check if DR1 and DR2 are in the same object set. */
1696 static bool
1697 dr_same_base_object_p (const struct data_reference *dr1,
1698 const struct data_reference *dr2)
1700 return operand_equal_p (DR_BASE_OBJECT (dr1), DR_BASE_OBJECT (dr2), 0);
1703 /* Uses DFS component number as representative of alias-sets. Also tests for
1704 optimality by verifying if every connected component is a clique. Returns
1705 true (1) if the above test is true, and false (0) otherwise. */
1707 static int
1708 build_alias_set_optimal_p (vec<data_reference_p> drs)
1710 int num_vertices = drs.length ();
1711 struct graph *g = new_graph (num_vertices);
1712 data_reference_p dr1, dr2;
1713 int i, j;
1714 int num_connected_components;
1715 int v_indx1, v_indx2, num_vertices_in_component;
1716 int *all_vertices;
1717 int *vertices;
1718 struct graph_edge *e;
1719 int this_component_is_clique;
1720 int all_components_are_cliques = 1;
1722 FOR_EACH_VEC_ELT (drs, i, dr1)
1723 for (j = i+1; drs.iterate (j, &dr2); j++)
1724 if (dr_may_alias_p (dr1, dr2, true))
1726 add_edge (g, i, j);
1727 add_edge (g, j, i);
1730 all_vertices = XNEWVEC (int, num_vertices);
1731 vertices = XNEWVEC (int, num_vertices);
1732 for (i = 0; i < num_vertices; i++)
1733 all_vertices[i] = i;
1735 num_connected_components = graphds_dfs (g, all_vertices, num_vertices,
1736 NULL, true, NULL);
1737 for (i = 0; i < g->n_vertices; i++)
1739 data_reference_p dr = drs[i];
1740 base_alias_pair *bap;
1742 gcc_assert (dr->aux);
1743 bap = (base_alias_pair *)(dr->aux);
1745 bap->alias_set = XNEW (int);
1746 *(bap->alias_set) = g->vertices[i].component + 1;
1749 /* Verify if the DFS numbering results in optimal solution. */
1750 for (i = 0; i < num_connected_components; i++)
1752 num_vertices_in_component = 0;
1753 /* Get all vertices whose DFS component number is the same as i. */
1754 for (j = 0; j < num_vertices; j++)
1755 if (g->vertices[j].component == i)
1756 vertices[num_vertices_in_component++] = j;
1758 /* Now test if the vertices in 'vertices' form a clique, by testing
1759 for edges among each pair. */
1760 this_component_is_clique = 1;
1761 for (v_indx1 = 0; v_indx1 < num_vertices_in_component; v_indx1++)
1763 for (v_indx2 = v_indx1+1; v_indx2 < num_vertices_in_component; v_indx2++)
1765 /* Check if the two vertices are connected by iterating
1766 through all the edges which have one of these are source. */
1767 e = g->vertices[vertices[v_indx2]].pred;
1768 while (e)
1770 if (e->src == vertices[v_indx1])
1771 break;
1772 e = e->pred_next;
1774 if (!e)
1776 this_component_is_clique = 0;
1777 break;
1780 if (!this_component_is_clique)
1781 all_components_are_cliques = 0;
1785 free (all_vertices);
1786 free (vertices);
1787 free_graph (g);
1788 return all_components_are_cliques;
1791 /* Group each data reference in DRS with its base object set num. */
1793 static void
1794 build_base_obj_set_for_drs (vec<data_reference_p> drs)
1796 int num_vertex = drs.length ();
1797 struct graph *g = new_graph (num_vertex);
1798 data_reference_p dr1, dr2;
1799 int i, j;
1800 int *queue;
1802 FOR_EACH_VEC_ELT (drs, i, dr1)
1803 for (j = i + 1; drs.iterate (j, &dr2); j++)
1804 if (dr_same_base_object_p (dr1, dr2))
1806 add_edge (g, i, j);
1807 add_edge (g, j, i);
1810 queue = XNEWVEC (int, num_vertex);
1811 for (i = 0; i < num_vertex; i++)
1812 queue[i] = i;
1814 graphds_dfs (g, queue, num_vertex, NULL, true, NULL);
1816 for (i = 0; i < g->n_vertices; i++)
1818 data_reference_p dr = drs[i];
1819 base_alias_pair *bap;
1821 gcc_assert (dr->aux);
1822 bap = (base_alias_pair *)(dr->aux);
1824 bap->base_obj_set = g->vertices[i].component + 1;
1827 free (queue);
1828 free_graph (g);
1831 /* Build the data references for PBB. */
1833 static void
1834 build_pbb_drs (poly_bb_p pbb)
1836 int j;
1837 data_reference_p dr;
1838 vec<data_reference_p> gbb_drs = GBB_DATA_REFS (PBB_BLACK_BOX (pbb));
1840 FOR_EACH_VEC_ELT (gbb_drs, j, dr)
1841 build_poly_dr (dr, pbb);
1844 /* Dump to file the alias graphs for the data references in DRS. */
1846 static void
1847 dump_alias_graphs (vec<data_reference_p> drs)
1849 char comment[100];
1850 FILE *file_dimacs, *file_ecc, *file_dot;
1852 file_dimacs = fopen ("/tmp/dr_alias_graph_dimacs", "ab");
1853 if (file_dimacs)
1855 snprintf (comment, sizeof (comment), "%s %s", main_input_filename,
1856 current_function_name ());
1857 write_alias_graph_to_ascii_dimacs (file_dimacs, comment, drs);
1858 fclose (file_dimacs);
1861 file_ecc = fopen ("/tmp/dr_alias_graph_ecc", "ab");
1862 if (file_ecc)
1864 snprintf (comment, sizeof (comment), "%s %s", main_input_filename,
1865 current_function_name ());
1866 write_alias_graph_to_ascii_ecc (file_ecc, comment, drs);
1867 fclose (file_ecc);
1870 file_dot = fopen ("/tmp/dr_alias_graph_dot", "ab");
1871 if (file_dot)
1873 snprintf (comment, sizeof (comment), "%s %s", main_input_filename,
1874 current_function_name ());
1875 write_alias_graph_to_ascii_dot (file_dot, comment, drs);
1876 fclose (file_dot);
1880 /* Build data references in SCOP. */
1882 static void
1883 build_scop_drs (scop_p scop)
1885 int i, j;
1886 poly_bb_p pbb;
1887 data_reference_p dr;
1888 auto_vec<data_reference_p, 3> drs;
1890 /* Remove all the PBBs that do not have data references: these basic
1891 blocks are not handled in the polyhedral representation. */
1892 for (i = 0; SCOP_BBS (scop).iterate (i, &pbb); i++)
1893 if (GBB_DATA_REFS (PBB_BLACK_BOX (pbb)).is_empty ())
1895 free_gimple_bb (PBB_BLACK_BOX (pbb));
1896 free_poly_bb (pbb);
1897 SCOP_BBS (scop).ordered_remove (i);
1898 i--;
1901 FOR_EACH_VEC_ELT (SCOP_BBS (scop), i, pbb)
1902 for (j = 0; GBB_DATA_REFS (PBB_BLACK_BOX (pbb)).iterate (j, &dr); j++)
1903 drs.safe_push (dr);
1905 FOR_EACH_VEC_ELT (drs, i, dr)
1906 dr->aux = XNEW (base_alias_pair);
1908 if (!build_alias_set_optimal_p (drs))
1910 /* TODO: Add support when building alias set is not optimal. */
1914 build_base_obj_set_for_drs (drs);
1916 /* When debugging, enable the following code. This cannot be used
1917 in production compilers. */
1918 if (0)
1919 dump_alias_graphs (drs);
1921 drs.release ();
1923 FOR_EACH_VEC_ELT (SCOP_BBS (scop), i, pbb)
1924 build_pbb_drs (pbb);
1927 /* Return a gsi at the position of the phi node STMT. */
1929 static gphi_iterator
1930 gsi_for_phi_node (gphi *stmt)
1932 gphi_iterator psi;
1933 basic_block bb = gimple_bb (stmt);
1935 for (psi = gsi_start_phis (bb); !gsi_end_p (psi); gsi_next (&psi))
1936 if (stmt == psi.phi ())
1937 return psi;
1939 gcc_unreachable ();
1940 return psi;
1943 /* Analyze all the data references of STMTS and add them to the
1944 GBB_DATA_REFS vector of BB. */
1946 static void
1947 analyze_drs_in_stmts (scop_p scop, basic_block bb, vec<gimple> stmts)
1949 loop_p nest;
1950 gimple_bb_p gbb;
1951 gimple stmt;
1952 int i;
1953 sese region = SCOP_REGION (scop);
1955 if (!bb_in_sese_p (bb, region))
1956 return;
1958 nest = outermost_loop_in_sese_1 (region, bb);
1959 gbb = gbb_from_bb (bb);
1961 FOR_EACH_VEC_ELT (stmts, i, stmt)
1963 loop_p loop;
1965 if (is_gimple_debug (stmt))
1966 continue;
1968 loop = loop_containing_stmt (stmt);
1969 if (!loop_in_sese_p (loop, region))
1970 loop = nest;
1972 graphite_find_data_references_in_stmt (nest, loop, stmt,
1973 &GBB_DATA_REFS (gbb));
1977 /* Insert STMT at the end of the STMTS sequence and then insert the
1978 statements from STMTS at INSERT_GSI and call analyze_drs_in_stmts
1979 on STMTS. */
1981 static void
1982 insert_stmts (scop_p scop, gimple stmt, gimple_seq stmts,
1983 gimple_stmt_iterator insert_gsi)
1985 gimple_stmt_iterator gsi;
1986 auto_vec<gimple, 3> x;
1988 gimple_seq_add_stmt (&stmts, stmt);
1989 for (gsi = gsi_start (stmts); !gsi_end_p (gsi); gsi_next (&gsi))
1990 x.safe_push (gsi_stmt (gsi));
1992 gsi_insert_seq_before (&insert_gsi, stmts, GSI_SAME_STMT);
1993 analyze_drs_in_stmts (scop, gsi_bb (insert_gsi), x);
1996 /* Insert the assignment "RES := EXPR" just after AFTER_STMT. */
1998 static void
1999 insert_out_of_ssa_copy (scop_p scop, tree res, tree expr, gimple after_stmt)
2001 gimple_seq stmts;
2002 gimple_stmt_iterator gsi;
2003 tree var = force_gimple_operand (expr, &stmts, true, NULL_TREE);
2004 gassign *stmt = gimple_build_assign (unshare_expr (res), var);
2005 auto_vec<gimple, 3> x;
2007 gimple_seq_add_stmt (&stmts, stmt);
2008 for (gsi = gsi_start (stmts); !gsi_end_p (gsi); gsi_next (&gsi))
2009 x.safe_push (gsi_stmt (gsi));
2011 if (gimple_code (after_stmt) == GIMPLE_PHI)
2013 gsi = gsi_after_labels (gimple_bb (after_stmt));
2014 gsi_insert_seq_before (&gsi, stmts, GSI_NEW_STMT);
2016 else
2018 gsi = gsi_for_stmt (after_stmt);
2019 gsi_insert_seq_after (&gsi, stmts, GSI_NEW_STMT);
2022 analyze_drs_in_stmts (scop, gimple_bb (after_stmt), x);
2025 /* Creates a poly_bb_p for basic_block BB from the existing PBB. */
2027 static void
2028 new_pbb_from_pbb (scop_p scop, poly_bb_p pbb, basic_block bb)
2030 vec<data_reference_p> drs;
2031 drs.create (3);
2032 gimple_bb_p gbb = PBB_BLACK_BOX (pbb);
2033 gimple_bb_p gbb1 = new_gimple_bb (bb, drs);
2034 poly_bb_p pbb1 = new_poly_bb (scop, gbb1);
2035 int index, n = SCOP_BBS (scop).length ();
2037 /* The INDEX of PBB in SCOP_BBS. */
2038 for (index = 0; index < n; index++)
2039 if (SCOP_BBS (scop)[index] == pbb)
2040 break;
2042 pbb1->domain = isl_set_copy (pbb->domain);
2043 pbb1->domain = isl_set_set_tuple_id (pbb1->domain,
2044 isl_id_for_pbb (scop, pbb1));
2046 GBB_PBB (gbb1) = pbb1;
2047 GBB_CONDITIONS (gbb1) = GBB_CONDITIONS (gbb).copy ();
2048 GBB_CONDITION_CASES (gbb1) = GBB_CONDITION_CASES (gbb).copy ();
2049 SCOP_BBS (scop).safe_insert (index + 1, pbb1);
2052 /* Insert on edge E the assignment "RES := EXPR". */
2054 static void
2055 insert_out_of_ssa_copy_on_edge (scop_p scop, edge e, tree res, tree expr)
2057 gimple_stmt_iterator gsi;
2058 gimple_seq stmts = NULL;
2059 tree var = force_gimple_operand (expr, &stmts, true, NULL_TREE);
2060 gimple stmt = gimple_build_assign (unshare_expr (res), var);
2061 basic_block bb;
2062 auto_vec<gimple, 3> x;
2064 gimple_seq_add_stmt (&stmts, stmt);
2065 for (gsi = gsi_start (stmts); !gsi_end_p (gsi); gsi_next (&gsi))
2066 x.safe_push (gsi_stmt (gsi));
2068 gsi_insert_seq_on_edge (e, stmts);
2069 gsi_commit_edge_inserts ();
2070 bb = gimple_bb (stmt);
2072 if (!bb_in_sese_p (bb, SCOP_REGION (scop)))
2073 return;
2075 if (!gbb_from_bb (bb))
2076 new_pbb_from_pbb (scop, pbb_from_bb (e->src), bb);
2078 analyze_drs_in_stmts (scop, bb, x);
2081 /* Creates a zero dimension array of the same type as VAR. */
2083 static tree
2084 create_zero_dim_array (tree var, const char *base_name)
2086 tree index_type = build_index_type (integer_zero_node);
2087 tree elt_type = TREE_TYPE (var);
2088 tree array_type = build_array_type (elt_type, index_type);
2089 tree base = create_tmp_var (array_type, base_name);
2091 return build4 (ARRAY_REF, elt_type, base, integer_zero_node, NULL_TREE,
2092 NULL_TREE);
2095 /* Returns true when PHI is a loop close phi node. */
2097 static bool
2098 scalar_close_phi_node_p (gimple phi)
2100 if (gimple_code (phi) != GIMPLE_PHI
2101 || virtual_operand_p (gimple_phi_result (phi)))
2102 return false;
2104 /* Note that loop close phi nodes should have a single argument
2105 because we translated the representation into a canonical form
2106 before Graphite: see canonicalize_loop_closed_ssa_form. */
2107 return (gimple_phi_num_args (phi) == 1);
2110 /* For a definition DEF in REGION, propagates the expression EXPR in
2111 all the uses of DEF outside REGION. */
2113 static void
2114 propagate_expr_outside_region (tree def, tree expr, sese region)
2116 imm_use_iterator imm_iter;
2117 gimple use_stmt;
2118 gimple_seq stmts;
2119 bool replaced_once = false;
2121 gcc_assert (TREE_CODE (def) == SSA_NAME);
2123 expr = force_gimple_operand (unshare_expr (expr), &stmts, true,
2124 NULL_TREE);
2126 FOR_EACH_IMM_USE_STMT (use_stmt, imm_iter, def)
2127 if (!is_gimple_debug (use_stmt)
2128 && !bb_in_sese_p (gimple_bb (use_stmt), region))
2130 ssa_op_iter iter;
2131 use_operand_p use_p;
2133 FOR_EACH_PHI_OR_STMT_USE (use_p, use_stmt, iter, SSA_OP_ALL_USES)
2134 if (operand_equal_p (def, USE_FROM_PTR (use_p), 0)
2135 && (replaced_once = true))
2136 replace_exp (use_p, expr);
2138 update_stmt (use_stmt);
2141 if (replaced_once)
2143 gsi_insert_seq_on_edge (SESE_ENTRY (region), stmts);
2144 gsi_commit_edge_inserts ();
2148 /* Rewrite out of SSA the reduction phi node at PSI by creating a zero
2149 dimension array for it. */
2151 static void
2152 rewrite_close_phi_out_of_ssa (scop_p scop, gimple_stmt_iterator *psi)
2154 sese region = SCOP_REGION (scop);
2155 gimple phi = gsi_stmt (*psi);
2156 tree res = gimple_phi_result (phi);
2157 basic_block bb = gimple_bb (phi);
2158 gimple_stmt_iterator gsi = gsi_after_labels (bb);
2159 tree arg = gimple_phi_arg_def (phi, 0);
2160 gimple stmt;
2162 /* Note that loop close phi nodes should have a single argument
2163 because we translated the representation into a canonical form
2164 before Graphite: see canonicalize_loop_closed_ssa_form. */
2165 gcc_assert (gimple_phi_num_args (phi) == 1);
2167 /* The phi node can be a non close phi node, when its argument is
2168 invariant, or a default definition. */
2169 if (is_gimple_min_invariant (arg)
2170 || SSA_NAME_IS_DEFAULT_DEF (arg))
2172 propagate_expr_outside_region (res, arg, region);
2173 gsi_next (psi);
2174 return;
2177 else if (gimple_bb (SSA_NAME_DEF_STMT (arg))->loop_father == bb->loop_father)
2179 propagate_expr_outside_region (res, arg, region);
2180 stmt = gimple_build_assign (res, arg);
2181 remove_phi_node (psi, false);
2182 gsi_insert_before (&gsi, stmt, GSI_NEW_STMT);
2183 return;
2186 /* If res is scev analyzable and is not a scalar value, it is safe
2187 to ignore the close phi node: it will be code generated in the
2188 out of Graphite pass. */
2189 else if (scev_analyzable_p (res, region))
2191 loop_p loop = loop_containing_stmt (SSA_NAME_DEF_STMT (res));
2192 tree scev;
2194 if (!loop_in_sese_p (loop, region))
2196 loop = loop_containing_stmt (SSA_NAME_DEF_STMT (arg));
2197 scev = scalar_evolution_in_region (region, loop, arg);
2198 scev = compute_overall_effect_of_inner_loop (loop, scev);
2200 else
2201 scev = scalar_evolution_in_region (region, loop, res);
2203 if (tree_does_not_contain_chrecs (scev))
2204 propagate_expr_outside_region (res, scev, region);
2206 gsi_next (psi);
2207 return;
2209 else
2211 tree zero_dim_array = create_zero_dim_array (res, "Close_Phi");
2213 stmt = gimple_build_assign (res, unshare_expr (zero_dim_array));
2215 if (TREE_CODE (arg) == SSA_NAME)
2216 insert_out_of_ssa_copy (scop, zero_dim_array, arg,
2217 SSA_NAME_DEF_STMT (arg));
2218 else
2219 insert_out_of_ssa_copy_on_edge (scop, single_pred_edge (bb),
2220 zero_dim_array, arg);
2223 remove_phi_node (psi, false);
2224 SSA_NAME_DEF_STMT (res) = stmt;
2226 insert_stmts (scop, stmt, NULL, gsi_after_labels (bb));
2229 /* Rewrite out of SSA the reduction phi node at PSI by creating a zero
2230 dimension array for it. */
2232 static void
2233 rewrite_phi_out_of_ssa (scop_p scop, gphi_iterator *psi)
2235 size_t i;
2236 gphi *phi = psi->phi ();
2237 basic_block bb = gimple_bb (phi);
2238 tree res = gimple_phi_result (phi);
2239 tree zero_dim_array = create_zero_dim_array (res, "phi_out_of_ssa");
2240 gimple stmt;
2242 for (i = 0; i < gimple_phi_num_args (phi); i++)
2244 tree arg = gimple_phi_arg_def (phi, i);
2245 edge e = gimple_phi_arg_edge (phi, i);
2247 /* Avoid the insertion of code in the loop latch to please the
2248 pattern matching of the vectorizer. */
2249 if (TREE_CODE (arg) == SSA_NAME
2250 && !SSA_NAME_IS_DEFAULT_DEF (arg)
2251 && e->src == bb->loop_father->latch)
2252 insert_out_of_ssa_copy (scop, zero_dim_array, arg,
2253 SSA_NAME_DEF_STMT (arg));
2254 else
2255 insert_out_of_ssa_copy_on_edge (scop, e, zero_dim_array, arg);
2258 stmt = gimple_build_assign (res, unshare_expr (zero_dim_array));
2259 remove_phi_node (psi, false);
2260 insert_stmts (scop, stmt, NULL, gsi_after_labels (bb));
2263 /* Rewrite the degenerate phi node at position PSI from the degenerate
2264 form "x = phi (y, y, ..., y)" to "x = y". */
2266 static void
2267 rewrite_degenerate_phi (gphi_iterator *psi)
2269 tree rhs;
2270 gimple stmt;
2271 gimple_stmt_iterator gsi;
2272 gphi *phi = psi->phi ();
2273 tree res = gimple_phi_result (phi);
2274 basic_block bb;
2276 bb = gimple_bb (phi);
2277 rhs = degenerate_phi_result (phi);
2278 gcc_assert (rhs);
2280 stmt = gimple_build_assign (res, rhs);
2281 remove_phi_node (psi, false);
2283 gsi = gsi_after_labels (bb);
2284 gsi_insert_before (&gsi, stmt, GSI_NEW_STMT);
2287 /* Rewrite out of SSA all the reduction phi nodes of SCOP. */
2289 static void
2290 rewrite_reductions_out_of_ssa (scop_p scop)
2292 basic_block bb;
2293 gphi_iterator psi;
2294 sese region = SCOP_REGION (scop);
2296 FOR_EACH_BB_FN (bb, cfun)
2297 if (bb_in_sese_p (bb, region))
2298 for (psi = gsi_start_phis (bb); !gsi_end_p (psi);)
2300 gphi *phi = psi.phi ();
2302 if (virtual_operand_p (gimple_phi_result (phi)))
2304 gsi_next (&psi);
2305 continue;
2308 if (gimple_phi_num_args (phi) > 1
2309 && degenerate_phi_result (phi))
2310 rewrite_degenerate_phi (&psi);
2312 else if (scalar_close_phi_node_p (phi))
2313 rewrite_close_phi_out_of_ssa (scop, &psi);
2315 else if (reduction_phi_p (region, &psi))
2316 rewrite_phi_out_of_ssa (scop, &psi);
2319 update_ssa (TODO_update_ssa);
2320 #ifdef ENABLE_CHECKING
2321 verify_loop_closed_ssa (true);
2322 #endif
2325 /* Rewrite the scalar dependence of DEF used in USE_STMT with a memory
2326 read from ZERO_DIM_ARRAY. */
2328 static void
2329 rewrite_cross_bb_scalar_dependence (scop_p scop, tree zero_dim_array,
2330 tree def, gimple use_stmt)
2332 gimple name_stmt;
2333 tree name;
2334 ssa_op_iter iter;
2335 use_operand_p use_p;
2337 gcc_assert (gimple_code (use_stmt) != GIMPLE_PHI);
2339 name = copy_ssa_name (def);
2340 name_stmt = gimple_build_assign (name, zero_dim_array);
2342 gimple_assign_set_lhs (name_stmt, name);
2343 insert_stmts (scop, name_stmt, NULL, gsi_for_stmt (use_stmt));
2345 FOR_EACH_SSA_USE_OPERAND (use_p, use_stmt, iter, SSA_OP_ALL_USES)
2346 if (operand_equal_p (def, USE_FROM_PTR (use_p), 0))
2347 replace_exp (use_p, name);
2349 update_stmt (use_stmt);
2352 /* For every definition DEF in the SCOP that is used outside the scop,
2353 insert a closing-scop definition in the basic block just after this
2354 SCOP. */
2356 static void
2357 handle_scalar_deps_crossing_scop_limits (scop_p scop, tree def, gimple stmt)
2359 tree var = create_tmp_reg (TREE_TYPE (def));
2360 tree new_name = make_ssa_name (var, stmt);
2361 bool needs_copy = false;
2362 use_operand_p use_p;
2363 imm_use_iterator imm_iter;
2364 gimple use_stmt;
2365 sese region = SCOP_REGION (scop);
2367 FOR_EACH_IMM_USE_STMT (use_stmt, imm_iter, def)
2369 if (!bb_in_sese_p (gimple_bb (use_stmt), region))
2371 FOR_EACH_IMM_USE_ON_STMT (use_p, imm_iter)
2373 SET_USE (use_p, new_name);
2375 update_stmt (use_stmt);
2376 needs_copy = true;
2380 /* Insert in the empty BB just after the scop a use of DEF such
2381 that the rewrite of cross_bb_scalar_dependences won't insert
2382 arrays everywhere else. */
2383 if (needs_copy)
2385 gimple assign = gimple_build_assign (new_name, def);
2386 gimple_stmt_iterator psi = gsi_after_labels (SESE_EXIT (region)->dest);
2388 update_stmt (assign);
2389 gsi_insert_before (&psi, assign, GSI_SAME_STMT);
2393 /* Rewrite the scalar dependences crossing the boundary of the BB
2394 containing STMT with an array. Return true when something has been
2395 changed. */
2397 static bool
2398 rewrite_cross_bb_scalar_deps (scop_p scop, gimple_stmt_iterator *gsi)
2400 sese region = SCOP_REGION (scop);
2401 gimple stmt = gsi_stmt (*gsi);
2402 imm_use_iterator imm_iter;
2403 tree def;
2404 basic_block def_bb;
2405 tree zero_dim_array = NULL_TREE;
2406 gimple use_stmt;
2407 bool res = false;
2409 switch (gimple_code (stmt))
2411 case GIMPLE_ASSIGN:
2412 def = gimple_assign_lhs (stmt);
2413 break;
2415 case GIMPLE_CALL:
2416 def = gimple_call_lhs (stmt);
2417 break;
2419 default:
2420 return false;
2423 if (!def
2424 || !is_gimple_reg (def))
2425 return false;
2427 if (scev_analyzable_p (def, region))
2429 loop_p loop = loop_containing_stmt (SSA_NAME_DEF_STMT (def));
2430 tree scev = scalar_evolution_in_region (region, loop, def);
2432 if (tree_contains_chrecs (scev, NULL))
2433 return false;
2435 propagate_expr_outside_region (def, scev, region);
2436 return true;
2439 def_bb = gimple_bb (stmt);
2441 handle_scalar_deps_crossing_scop_limits (scop, def, stmt);
2443 FOR_EACH_IMM_USE_STMT (use_stmt, imm_iter, def)
2444 if (gphi *phi = dyn_cast <gphi *> (use_stmt))
2446 res = true;
2447 gphi_iterator psi = gsi_for_phi (phi);
2449 if (scalar_close_phi_node_p (gsi_stmt (psi)))
2450 rewrite_close_phi_out_of_ssa (scop, &psi);
2451 else
2452 rewrite_phi_out_of_ssa (scop, &psi);
2455 FOR_EACH_IMM_USE_STMT (use_stmt, imm_iter, def)
2456 if (gimple_code (use_stmt) != GIMPLE_PHI
2457 && def_bb != gimple_bb (use_stmt)
2458 && !is_gimple_debug (use_stmt)
2459 && (res = true))
2461 if (!zero_dim_array)
2463 zero_dim_array = create_zero_dim_array
2464 (def, "Cross_BB_scalar_dependence");
2465 insert_out_of_ssa_copy (scop, zero_dim_array, def,
2466 SSA_NAME_DEF_STMT (def));
2467 gsi_next (gsi);
2470 rewrite_cross_bb_scalar_dependence (scop, unshare_expr (zero_dim_array),
2471 def, use_stmt);
2474 return res;
2477 /* Rewrite out of SSA all the reduction phi nodes of SCOP. */
2479 static void
2480 rewrite_cross_bb_scalar_deps_out_of_ssa (scop_p scop)
2482 basic_block bb;
2483 gimple_stmt_iterator psi;
2484 sese region = SCOP_REGION (scop);
2485 bool changed = false;
2487 /* Create an extra empty BB after the scop. */
2488 split_edge (SESE_EXIT (region));
2490 FOR_EACH_BB_FN (bb, cfun)
2491 if (bb_in_sese_p (bb, region))
2492 for (psi = gsi_start_bb (bb); !gsi_end_p (psi); gsi_next (&psi))
2493 changed |= rewrite_cross_bb_scalar_deps (scop, &psi);
2495 if (changed)
2497 scev_reset_htab ();
2498 update_ssa (TODO_update_ssa);
2499 #ifdef ENABLE_CHECKING
2500 verify_loop_closed_ssa (true);
2501 #endif
2505 /* Returns the number of pbbs that are in loops contained in SCOP. */
2507 static int
2508 nb_pbbs_in_loops (scop_p scop)
2510 int i;
2511 poly_bb_p pbb;
2512 int res = 0;
2514 FOR_EACH_VEC_ELT (SCOP_BBS (scop), i, pbb)
2515 if (loop_in_sese_p (gbb_loop (PBB_BLACK_BOX (pbb)), SCOP_REGION (scop)))
2516 res++;
2518 return res;
2521 /* Return the number of data references in BB that write in
2522 memory. */
2524 static int
2525 nb_data_writes_in_bb (basic_block bb)
2527 int res = 0;
2528 gimple_stmt_iterator gsi;
2530 for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi))
2531 if (gimple_vdef (gsi_stmt (gsi)))
2532 res++;
2534 return res;
2537 /* Splits at STMT the basic block BB represented as PBB in the
2538 polyhedral form. */
2540 static edge
2541 split_pbb (scop_p scop, poly_bb_p pbb, basic_block bb, gimple stmt)
2543 edge e1 = split_block (bb, stmt);
2544 new_pbb_from_pbb (scop, pbb, e1->dest);
2545 return e1;
2548 /* Splits STMT out of its current BB. This is done for reduction
2549 statements for which we want to ignore data dependences. */
2551 static basic_block
2552 split_reduction_stmt (scop_p scop, gimple stmt)
2554 basic_block bb = gimple_bb (stmt);
2555 poly_bb_p pbb = pbb_from_bb (bb);
2556 gimple_bb_p gbb = gbb_from_bb (bb);
2557 edge e1;
2558 int i;
2559 data_reference_p dr;
2561 /* Do not split basic blocks with no writes to memory: the reduction
2562 will be the only write to memory. */
2563 if (nb_data_writes_in_bb (bb) == 0
2564 /* Or if we have already marked BB as a reduction. */
2565 || PBB_IS_REDUCTION (pbb_from_bb (bb)))
2566 return bb;
2568 e1 = split_pbb (scop, pbb, bb, stmt);
2570 /* Split once more only when the reduction stmt is not the only one
2571 left in the original BB. */
2572 if (!gsi_one_before_end_p (gsi_start_nondebug_bb (bb)))
2574 gimple_stmt_iterator gsi = gsi_last_bb (bb);
2575 gsi_prev (&gsi);
2576 e1 = split_pbb (scop, pbb, bb, gsi_stmt (gsi));
2579 /* A part of the data references will end in a different basic block
2580 after the split: move the DRs from the original GBB to the newly
2581 created GBB1. */
2582 FOR_EACH_VEC_ELT (GBB_DATA_REFS (gbb), i, dr)
2584 basic_block bb1 = gimple_bb (DR_STMT (dr));
2586 if (bb1 != bb)
2588 gimple_bb_p gbb1 = gbb_from_bb (bb1);
2589 GBB_DATA_REFS (gbb1).safe_push (dr);
2590 GBB_DATA_REFS (gbb).ordered_remove (i);
2591 i--;
2595 return e1->dest;
2598 /* Return true when stmt is a reduction operation. */
2600 static inline bool
2601 is_reduction_operation_p (gimple stmt)
2603 enum tree_code code;
2605 gcc_assert (is_gimple_assign (stmt));
2606 code = gimple_assign_rhs_code (stmt);
2608 if (!commutative_tree_code (code)
2609 || !associative_tree_code (code))
2610 return false;
2612 tree type = TREE_TYPE (gimple_assign_lhs (stmt));
2614 if (FLOAT_TYPE_P (type))
2615 return flag_associative_math;
2617 if (ANY_INTEGRAL_TYPE_P (type))
2618 return (TYPE_OVERFLOW_WRAPS (type)
2619 || !operation_can_overflow (code));
2621 return false;
2624 /* Returns true when PHI contains an argument ARG. */
2626 static bool
2627 phi_contains_arg (gphi *phi, tree arg)
2629 size_t i;
2631 for (i = 0; i < gimple_phi_num_args (phi); i++)
2632 if (operand_equal_p (arg, gimple_phi_arg_def (phi, i), 0))
2633 return true;
2635 return false;
2638 /* Return a loop phi node that corresponds to a reduction containing LHS. */
2640 static gphi *
2641 follow_ssa_with_commutative_ops (tree arg, tree lhs)
2643 gimple stmt;
2645 if (TREE_CODE (arg) != SSA_NAME)
2646 return NULL;
2648 stmt = SSA_NAME_DEF_STMT (arg);
2650 if (gimple_code (stmt) == GIMPLE_NOP
2651 || gimple_code (stmt) == GIMPLE_CALL)
2652 return NULL;
2654 if (gphi *phi = dyn_cast <gphi *> (stmt))
2656 if (phi_contains_arg (phi, lhs))
2657 return phi;
2658 return NULL;
2661 if (!is_gimple_assign (stmt))
2662 return NULL;
2664 if (gimple_num_ops (stmt) == 2)
2665 return follow_ssa_with_commutative_ops (gimple_assign_rhs1 (stmt), lhs);
2667 if (is_reduction_operation_p (stmt))
2669 gphi *res
2670 = follow_ssa_with_commutative_ops (gimple_assign_rhs1 (stmt), lhs);
2672 return res ? res :
2673 follow_ssa_with_commutative_ops (gimple_assign_rhs2 (stmt), lhs);
2676 return NULL;
2679 /* Detect commutative and associative scalar reductions starting at
2680 the STMT. Return the phi node of the reduction cycle, or NULL. */
2682 static gphi *
2683 detect_commutative_reduction_arg (tree lhs, gimple stmt, tree arg,
2684 vec<gimple> *in,
2685 vec<gimple> *out)
2687 gphi *phi = follow_ssa_with_commutative_ops (arg, lhs);
2689 if (!phi)
2690 return NULL;
2692 in->safe_push (stmt);
2693 out->safe_push (stmt);
2694 return phi;
2697 /* Detect commutative and associative scalar reductions starting at
2698 STMT. Return the phi node of the reduction cycle, or NULL. */
2700 static gphi *
2701 detect_commutative_reduction_assign (gimple stmt, vec<gimple> *in,
2702 vec<gimple> *out)
2704 tree lhs = gimple_assign_lhs (stmt);
2706 if (gimple_num_ops (stmt) == 2)
2707 return detect_commutative_reduction_arg (lhs, stmt,
2708 gimple_assign_rhs1 (stmt),
2709 in, out);
2711 if (is_reduction_operation_p (stmt))
2713 gphi *res = detect_commutative_reduction_arg (lhs, stmt,
2714 gimple_assign_rhs1 (stmt),
2715 in, out);
2716 return res ? res
2717 : detect_commutative_reduction_arg (lhs, stmt,
2718 gimple_assign_rhs2 (stmt),
2719 in, out);
2722 return NULL;
2725 /* Return a loop phi node that corresponds to a reduction containing LHS. */
2727 static gphi *
2728 follow_inital_value_to_phi (tree arg, tree lhs)
2730 gimple stmt;
2732 if (!arg || TREE_CODE (arg) != SSA_NAME)
2733 return NULL;
2735 stmt = SSA_NAME_DEF_STMT (arg);
2737 if (gphi *phi = dyn_cast <gphi *> (stmt))
2738 if (phi_contains_arg (phi, lhs))
2739 return phi;
2741 return NULL;
2745 /* Return the argument of the loop PHI that is the initial value coming
2746 from outside the loop. */
2748 static edge
2749 edge_initial_value_for_loop_phi (gphi *phi)
2751 size_t i;
2753 for (i = 0; i < gimple_phi_num_args (phi); i++)
2755 edge e = gimple_phi_arg_edge (phi, i);
2757 if (loop_depth (e->src->loop_father)
2758 < loop_depth (e->dest->loop_father))
2759 return e;
2762 return NULL;
2765 /* Return the argument of the loop PHI that is the initial value coming
2766 from outside the loop. */
2768 static tree
2769 initial_value_for_loop_phi (gphi *phi)
2771 size_t i;
2773 for (i = 0; i < gimple_phi_num_args (phi); i++)
2775 edge e = gimple_phi_arg_edge (phi, i);
2777 if (loop_depth (e->src->loop_father)
2778 < loop_depth (e->dest->loop_father))
2779 return gimple_phi_arg_def (phi, i);
2782 return NULL_TREE;
2785 /* Returns true when DEF is used outside the reduction cycle of
2786 LOOP_PHI. */
2788 static bool
2789 used_outside_reduction (tree def, gimple loop_phi)
2791 use_operand_p use_p;
2792 imm_use_iterator imm_iter;
2793 loop_p loop = loop_containing_stmt (loop_phi);
2795 /* In LOOP, DEF should be used only in LOOP_PHI. */
2796 FOR_EACH_IMM_USE_FAST (use_p, imm_iter, def)
2798 gimple stmt = USE_STMT (use_p);
2800 if (stmt != loop_phi
2801 && !is_gimple_debug (stmt)
2802 && flow_bb_inside_loop_p (loop, gimple_bb (stmt)))
2803 return true;
2806 return false;
2809 /* Detect commutative and associative scalar reductions belonging to
2810 the SCOP starting at the loop closed phi node STMT. Return the phi
2811 node of the reduction cycle, or NULL. */
2813 static gphi *
2814 detect_commutative_reduction (scop_p scop, gimple stmt, vec<gimple> *in,
2815 vec<gimple> *out)
2817 if (scalar_close_phi_node_p (stmt))
2819 gimple def;
2820 gphi *loop_phi, *phi, *close_phi = as_a <gphi *> (stmt);
2821 tree init, lhs, arg = gimple_phi_arg_def (close_phi, 0);
2823 if (TREE_CODE (arg) != SSA_NAME)
2824 return NULL;
2826 /* Note that loop close phi nodes should have a single argument
2827 because we translated the representation into a canonical form
2828 before Graphite: see canonicalize_loop_closed_ssa_form. */
2829 gcc_assert (gimple_phi_num_args (close_phi) == 1);
2831 def = SSA_NAME_DEF_STMT (arg);
2832 if (!stmt_in_sese_p (def, SCOP_REGION (scop))
2833 || !(loop_phi = detect_commutative_reduction (scop, def, in, out)))
2834 return NULL;
2836 lhs = gimple_phi_result (close_phi);
2837 init = initial_value_for_loop_phi (loop_phi);
2838 phi = follow_inital_value_to_phi (init, lhs);
2840 if (phi && (used_outside_reduction (lhs, phi)
2841 || !has_single_use (gimple_phi_result (phi))))
2842 return NULL;
2844 in->safe_push (loop_phi);
2845 out->safe_push (close_phi);
2846 return phi;
2849 if (gimple_code (stmt) == GIMPLE_ASSIGN)
2850 return detect_commutative_reduction_assign (stmt, in, out);
2852 return NULL;
2855 /* Translate the scalar reduction statement STMT to an array RED
2856 knowing that its recursive phi node is LOOP_PHI. */
2858 static void
2859 translate_scalar_reduction_to_array_for_stmt (scop_p scop, tree red,
2860 gimple stmt, gphi *loop_phi)
2862 tree res = gimple_phi_result (loop_phi);
2863 gassign *assign = gimple_build_assign (res, unshare_expr (red));
2864 gimple_stmt_iterator gsi;
2866 insert_stmts (scop, assign, NULL, gsi_after_labels (gimple_bb (loop_phi)));
2868 assign = gimple_build_assign (unshare_expr (red), gimple_assign_lhs (stmt));
2869 gsi = gsi_for_stmt (stmt);
2870 gsi_next (&gsi);
2871 insert_stmts (scop, assign, NULL, gsi);
2874 /* Removes the PHI node and resets all the debug stmts that are using
2875 the PHI_RESULT. */
2877 static void
2878 remove_phi (gphi *phi)
2880 imm_use_iterator imm_iter;
2881 tree def;
2882 use_operand_p use_p;
2883 gimple_stmt_iterator gsi;
2884 auto_vec<gimple, 3> update;
2885 unsigned int i;
2886 gimple stmt;
2888 def = PHI_RESULT (phi);
2889 FOR_EACH_IMM_USE_FAST (use_p, imm_iter, def)
2891 stmt = USE_STMT (use_p);
2893 if (is_gimple_debug (stmt))
2895 gimple_debug_bind_reset_value (stmt);
2896 update.safe_push (stmt);
2900 FOR_EACH_VEC_ELT (update, i, stmt)
2901 update_stmt (stmt);
2903 gsi = gsi_for_phi_node (phi);
2904 remove_phi_node (&gsi, false);
2907 /* Helper function for for_each_index. For each INDEX of the data
2908 reference REF, returns true when its indices are valid in the loop
2909 nest LOOP passed in as DATA. */
2911 static bool
2912 dr_indices_valid_in_loop (tree ref ATTRIBUTE_UNUSED, tree *index, void *data)
2914 loop_p loop;
2915 basic_block header, def_bb;
2916 gimple stmt;
2918 if (TREE_CODE (*index) != SSA_NAME)
2919 return true;
2921 loop = *((loop_p *) data);
2922 header = loop->header;
2923 stmt = SSA_NAME_DEF_STMT (*index);
2925 if (!stmt)
2926 return true;
2928 def_bb = gimple_bb (stmt);
2930 if (!def_bb)
2931 return true;
2933 return dominated_by_p (CDI_DOMINATORS, header, def_bb);
2936 /* When the result of a CLOSE_PHI is written to a memory location,
2937 return a pointer to that memory reference, otherwise return
2938 NULL_TREE. */
2940 static tree
2941 close_phi_written_to_memory (gphi *close_phi)
2943 imm_use_iterator imm_iter;
2944 use_operand_p use_p;
2945 gimple stmt;
2946 tree res, def = gimple_phi_result (close_phi);
2948 FOR_EACH_IMM_USE_FAST (use_p, imm_iter, def)
2949 if ((stmt = USE_STMT (use_p))
2950 && gimple_code (stmt) == GIMPLE_ASSIGN
2951 && (res = gimple_assign_lhs (stmt)))
2953 switch (TREE_CODE (res))
2955 case VAR_DECL:
2956 case PARM_DECL:
2957 case RESULT_DECL:
2958 return res;
2960 case ARRAY_REF:
2961 case MEM_REF:
2963 tree arg = gimple_phi_arg_def (close_phi, 0);
2964 loop_p nest = loop_containing_stmt (SSA_NAME_DEF_STMT (arg));
2966 /* FIXME: this restriction is for id-{24,25}.f and
2967 could be handled by duplicating the computation of
2968 array indices before the loop of the close_phi. */
2969 if (for_each_index (&res, dr_indices_valid_in_loop, &nest))
2970 return res;
2972 /* Fallthru. */
2974 default:
2975 continue;
2978 return NULL_TREE;
2981 /* Rewrite out of SSA the reduction described by the loop phi nodes
2982 IN, and the close phi nodes OUT. IN and OUT are structured by loop
2983 levels like this:
2985 IN: stmt, loop_n, ..., loop_0
2986 OUT: stmt, close_n, ..., close_0
2988 the first element is the reduction statement, and the next elements
2989 are the loop and close phi nodes of each of the outer loops. */
2991 static void
2992 translate_scalar_reduction_to_array (scop_p scop,
2993 vec<gimple> in,
2994 vec<gimple> out)
2996 gimple loop_stmt;
2997 unsigned int i = out.length () - 1;
2998 tree red = close_phi_written_to_memory (as_a <gphi *> (out[i]));
3000 FOR_EACH_VEC_ELT (in, i, loop_stmt)
3002 gimple close_stmt = out[i];
3004 if (i == 0)
3006 basic_block bb = split_reduction_stmt (scop, loop_stmt);
3007 poly_bb_p pbb = pbb_from_bb (bb);
3008 PBB_IS_REDUCTION (pbb) = true;
3009 gcc_assert (close_stmt == loop_stmt);
3011 if (!red)
3012 red = create_zero_dim_array
3013 (gimple_assign_lhs (loop_stmt), "Commutative_Associative_Reduction");
3015 translate_scalar_reduction_to_array_for_stmt (scop, red, loop_stmt,
3016 as_a <gphi *> (in[1]));
3017 continue;
3020 gphi *loop_phi = as_a <gphi *> (loop_stmt);
3021 gphi *close_phi = as_a <gphi *> (close_stmt);
3023 if (i == in.length () - 1)
3025 insert_out_of_ssa_copy (scop, gimple_phi_result (close_phi),
3026 unshare_expr (red), close_phi);
3027 insert_out_of_ssa_copy_on_edge
3028 (scop, edge_initial_value_for_loop_phi (loop_phi),
3029 unshare_expr (red), initial_value_for_loop_phi (loop_phi));
3032 remove_phi (loop_phi);
3033 remove_phi (close_phi);
3037 /* Rewrites out of SSA a commutative reduction at CLOSE_PHI. Returns
3038 true when something has been changed. */
3040 static bool
3041 rewrite_commutative_reductions_out_of_ssa_close_phi (scop_p scop,
3042 gphi *close_phi)
3044 bool res;
3045 auto_vec<gimple, 10> in;
3046 auto_vec<gimple, 10> out;
3048 detect_commutative_reduction (scop, close_phi, &in, &out);
3049 res = in.length () > 1;
3050 if (res)
3051 translate_scalar_reduction_to_array (scop, in, out);
3053 return res;
3056 /* Rewrites all the commutative reductions from LOOP out of SSA.
3057 Returns true when something has been changed. */
3059 static bool
3060 rewrite_commutative_reductions_out_of_ssa_loop (scop_p scop,
3061 loop_p loop)
3063 gphi_iterator gsi;
3064 edge exit = single_exit (loop);
3065 tree res;
3066 bool changed = false;
3068 if (!exit)
3069 return false;
3071 for (gsi = gsi_start_phis (exit->dest); !gsi_end_p (gsi); gsi_next (&gsi))
3072 if ((res = gimple_phi_result (gsi.phi ()))
3073 && !virtual_operand_p (res)
3074 && !scev_analyzable_p (res, SCOP_REGION (scop)))
3075 changed |= rewrite_commutative_reductions_out_of_ssa_close_phi
3076 (scop, gsi.phi ());
3078 return changed;
3081 /* Rewrites all the commutative reductions from SCOP out of SSA. */
3083 static void
3084 rewrite_commutative_reductions_out_of_ssa (scop_p scop)
3086 loop_p loop;
3087 bool changed = false;
3088 sese region = SCOP_REGION (scop);
3090 FOR_EACH_LOOP (loop, 0)
3091 if (loop_in_sese_p (loop, region))
3092 changed |= rewrite_commutative_reductions_out_of_ssa_loop (scop, loop);
3094 if (changed)
3096 scev_reset_htab ();
3097 gsi_commit_edge_inserts ();
3098 update_ssa (TODO_update_ssa);
3099 #ifdef ENABLE_CHECKING
3100 verify_loop_closed_ssa (true);
3101 #endif
3105 /* Can all ivs be represented by a signed integer?
3106 As CLooG might generate negative values in its expressions, signed loop ivs
3107 are required in the backend. */
3109 static bool
3110 scop_ivs_can_be_represented (scop_p scop)
3112 loop_p loop;
3113 gphi_iterator psi;
3114 bool result = true;
3116 FOR_EACH_LOOP (loop, 0)
3118 if (!loop_in_sese_p (loop, SCOP_REGION (scop)))
3119 continue;
3121 for (psi = gsi_start_phis (loop->header);
3122 !gsi_end_p (psi); gsi_next (&psi))
3124 gphi *phi = psi.phi ();
3125 tree res = PHI_RESULT (phi);
3126 tree type = TREE_TYPE (res);
3128 if (TYPE_UNSIGNED (type)
3129 && TYPE_PRECISION (type) >= TYPE_PRECISION (long_long_integer_type_node))
3131 result = false;
3132 break;
3135 if (!result)
3136 break;
3139 return result;
3142 /* Builds the polyhedral representation for a SESE region. */
3144 void
3145 build_poly_scop (scop_p scop)
3147 sese region = SCOP_REGION (scop);
3148 graphite_dim_t max_dim;
3150 build_scop_bbs (scop);
3152 /* FIXME: This restriction is needed to avoid a problem in CLooG.
3153 Once CLooG is fixed, remove this guard. Anyways, it makes no
3154 sense to optimize a scop containing only PBBs that do not belong
3155 to any loops. */
3156 if (nb_pbbs_in_loops (scop) == 0)
3157 return;
3159 if (!scop_ivs_can_be_represented (scop))
3160 return;
3162 rewrite_commutative_reductions_out_of_ssa (scop);
3164 build_sese_loop_nests (region);
3165 /* Record all conditions in REGION. */
3166 sese_dom_walker (CDI_DOMINATORS, region).walk (cfun->cfg->x_entry_block_ptr);
3167 find_scop_parameters (scop);
3169 max_dim = PARAM_VALUE (PARAM_GRAPHITE_MAX_NB_SCOP_PARAMS);
3170 if (scop_nb_params (scop) > max_dim)
3171 return;
3173 build_scop_iteration_domain (scop);
3174 build_scop_context (scop);
3175 add_conditions_to_constraints (scop);
3177 /* Rewrite out of SSA only after having translated the
3178 representation to the polyhedral representation to avoid scev
3179 analysis failures. That means that these functions will insert
3180 new data references that they create in the right place. */
3181 rewrite_reductions_out_of_ssa (scop);
3182 rewrite_cross_bb_scalar_deps_out_of_ssa (scop);
3184 build_scop_drs (scop);
3185 scop_to_lst (scop);
3186 build_scop_scattering (scop);
3188 /* This SCoP has been translated to the polyhedral
3189 representation. */
3190 POLY_SCOP_P (scop) = true;
3192 #endif /* HAVE_isl */