]> git.ipfire.org Git - thirdparty/gcc.git/blame - gcc/graphite-scop-detection.c
[Ada] Fix documentation for GNAT.Command_Line.Exit_From_Command_Line
[thirdparty/gcc.git] / gcc / graphite-scop-detection.c
CommitLineData
c6bb733d 1/* Detection of Static Control Parts (SCoP) for Graphite.
fbd26352 2 Copyright (C) 2009-2019 Free Software Foundation, Inc.
c6bb733d 3 Contributed by Sebastian Pop <sebastian.pop@amd.com> and
4 Tobias Grosser <grosser@fim.uni-passau.de>.
5
6This file is part of GCC.
7
8GCC is free software; you can redistribute it and/or modify
9it under the terms of the GNU General Public License as published by
10the Free Software Foundation; either version 3, or (at your option)
11any later version.
12
13GCC is distributed in the hope that it will be useful,
14but WITHOUT ANY WARRANTY; without even the implied warranty of
15MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
16GNU General Public License for more details.
17
18You should have received a copy of the GNU General Public License
19along with GCC; see the file COPYING3. If not see
20<http://www.gnu.org/licenses/>. */
21
d8a2c81e 22#define USES_ISL
23
c6bb733d 24#include "config.h"
87e20041 25
429cca51 26#ifdef HAVE_isl
87e20041 27
c6bb733d 28#include "system.h"
29#include "coretypes.h"
9ef16211 30#include "backend.h"
d040a5b0 31#include "cfghooks.h"
edbec012 32#include "domwalk.h"
7eb20e71 33#include "params.h"
b20a8bb4 34#include "tree.h"
9ef16211 35#include "gimple.h"
9ef16211 36#include "ssa.h"
9ef16211 37#include "fold-const.h"
dcf1a1ec 38#include "gimple-iterator.h"
edbec012 39#include "tree-cfg.h"
05d9c18a 40#include "tree-ssa-loop-manip.h"
41#include "tree-ssa-loop-niter.h"
073c1fd5 42#include "tree-ssa-loop.h"
43#include "tree-into-ssa.h"
69ee5dbb 44#include "tree-ssa.h"
c6bb733d 45#include "cfgloop.h"
c6bb733d 46#include "tree-data-ref.h"
47#include "tree-scalar-evolution.h"
48#include "tree-pass.h"
26bd1f19 49#include "tree-ssa-propagate.h"
4773ab25 50#include "gimple-pretty-print.h"
0fcd2c46 51#include "cfganal.h"
39b8c5f3 52#include "graphite.h"
d8a2c81e 53
edbec012 54class debug_printer
55{
56private:
57 FILE *dump_file;
58
59public:
60 void
61 set_dump_file (FILE *f)
62 {
63 gcc_assert (f);
64 dump_file = f;
65 }
66
67 friend debug_printer &
68 operator<< (debug_printer &output, int i)
69 {
70 fprintf (output.dump_file, "%d", i);
71 return output;
72 }
73 friend debug_printer &
74 operator<< (debug_printer &output, const char *s)
75 {
76 fprintf (output.dump_file, "%s", s);
77 return output;
78 }
79} dp;
80
81#define DEBUG_PRINT(args) do \
82 { \
83 if (dump_file && (dump_flags & TDF_DETAILS)) { args; } \
fa57650a 84 } while (0)
edbec012 85
e057353f 86/* Pretty print to FILE all the SCoPs in DOT format and mark them with
87 different colors. If there are not enough colors, paint the
88 remaining SCoPs in gray.
89
90 Special nodes:
91 - "*" after the node number denotes the entry of a SCoP,
92 - "#" after the node number denotes the exit of a SCoP,
93 - "()" around the node number denotes the entry or the
94 exit nodes of the SCOP. These are not part of SCoP. */
95
6cfed2d0 96DEBUG_FUNCTION void
97dot_all_sese (FILE *file, vec<sese_l>& scops)
e057353f 98{
e057353f 99 /* Disable debugging while printing graph. */
3f6e5ced 100 dump_flags_t tmp_dump_flags = dump_flags;
101 dump_flags = TDF_NONE;
e057353f 102
103 fprintf (file, "digraph all {\n");
104
6cfed2d0 105 basic_block bb;
e057353f 106 FOR_ALL_BB_FN (bb, cfun)
107 {
108 int part_of_scop = false;
109
110 /* Use HTML for every bb label. So we are able to print bbs
111 which are part of two different SCoPs, with two different
112 background colors. */
113 fprintf (file, "%d [label=<\n <TABLE BORDER=\"0\" CELLBORDER=\"1\" ",
114 bb->index);
115 fprintf (file, "CELLSPACING=\"0\">\n");
116
117 /* Select color for SCoP. */
6cfed2d0 118 sese_l *region;
119 int i;
120 FOR_EACH_VEC_ELT (scops, i, region)
e057353f 121 {
6cfed2d0 122 bool sese_in_region = bb_in_sese_p (bb, *region);
123 if (sese_in_region || (region->exit->dest == bb)
124 || (region->entry->dest == bb))
e057353f 125 {
6cfed2d0 126 const char *color;
e057353f 127 switch (i % 17)
128 {
129 case 0: /* red */
130 color = "#e41a1c";
131 break;
132 case 1: /* blue */
133 color = "#377eb8";
134 break;
135 case 2: /* green */
136 color = "#4daf4a";
137 break;
138 case 3: /* purple */
139 color = "#984ea3";
140 break;
141 case 4: /* orange */
142 color = "#ff7f00";
143 break;
144 case 5: /* yellow */
145 color = "#ffff33";
146 break;
147 case 6: /* brown */
148 color = "#a65628";
149 break;
150 case 7: /* rose */
151 color = "#f781bf";
152 break;
153 case 8:
154 color = "#8dd3c7";
155 break;
156 case 9:
157 color = "#ffffb3";
158 break;
159 case 10:
160 color = "#bebada";
161 break;
162 case 11:
163 color = "#fb8072";
164 break;
165 case 12:
166 color = "#80b1d3";
167 break;
168 case 13:
169 color = "#fdb462";
170 break;
171 case 14:
172 color = "#b3de69";
173 break;
174 case 15:
175 color = "#fccde5";
176 break;
177 case 16:
178 color = "#bc80bd";
179 break;
180 default: /* gray */
181 color = "#999999";
182 }
183
184 fprintf (file, " <TR><TD WIDTH=\"50\" BGCOLOR=\"%s\">",
185 color);
186
6cfed2d0 187 if (!sese_in_region)
e057353f 188 fprintf (file, " (");
189
6cfed2d0 190 if (bb == region->entry->dest && bb == region->exit->dest)
e057353f 191 fprintf (file, " %d*# ", bb->index);
6cfed2d0 192 else if (bb == region->entry->dest)
e057353f 193 fprintf (file, " %d* ", bb->index);
6cfed2d0 194 else if (bb == region->exit->dest)
e057353f 195 fprintf (file, " %d# ", bb->index);
196 else
197 fprintf (file, " %d ", bb->index);
198
199 fprintf (file, "{lp_%d}", bb->loop_father->num);
200
6cfed2d0 201 if (!sese_in_region)
e057353f 202 fprintf (file, ")");
203
204 fprintf (file, "</TD></TR>\n");
205 part_of_scop = true;
206 }
207 }
208
209 if (!part_of_scop)
210 {
211 fprintf (file, " <TR><TD WIDTH=\"50\" BGCOLOR=\"#ffffff\">");
212 fprintf (file, " %d {lp_%d} </TD></TR>\n", bb->index,
213 bb->loop_father->num);
214 }
215 fprintf (file, " </TABLE>>, shape=box, style=\"setlinewidth(0)\"]\n");
216 }
217
218 FOR_ALL_BB_FN (bb, cfun)
219 {
6cfed2d0 220 edge e;
221 edge_iterator ei;
e057353f 222 FOR_EACH_EDGE (e, ei, bb->succs)
223 fprintf (file, "%d -> %d;\n", bb->index, e->dest->index);
224 }
225
226 fputs ("}\n\n", file);
227
228 /* Enable debugging again. */
229 dump_flags = tmp_dump_flags;
230}
231
6cfed2d0 232/* Display SCoP on stderr. */
e057353f 233
234DEBUG_FUNCTION void
6cfed2d0 235dot_sese (sese_l& scop)
e057353f 236{
6cfed2d0 237 vec<sese_l> scops;
238 scops.create (1);
e057353f 239
240 if (scop)
241 scops.safe_push (scop);
242
6cfed2d0 243 dot_all_sese (stderr, scops);
e057353f 244
6cfed2d0 245 scops.release ();
246}
247
248DEBUG_FUNCTION void
249dot_cfg ()
250{
251 vec<sese_l> scops;
252 scops.create (1);
253 dot_all_sese (stderr, scops);
254 scops.release ();
e057353f 255}
edbec012 256
edbec012 257/* Returns a COND_EXPR statement when BB has a single predecessor, the
258 edge between BB and its predecessor is not a loop exit edge, and
259 the last statement of the single predecessor is a COND_EXPR. */
260
261static gcond *
262single_pred_cond_non_loop_exit (basic_block bb)
263{
264 if (single_pred_p (bb))
265 {
266 edge e = single_pred_edge (bb);
267 basic_block pred = e->src;
268 gimple *stmt;
269
270 if (loop_depth (pred->loop_father) > loop_depth (bb->loop_father))
271 return NULL;
272
273 stmt = last_stmt (pred);
274
275 if (stmt && gimple_code (stmt) == GIMPLE_COND)
276 return as_a<gcond *> (stmt);
277 }
278
279 return NULL;
280}
281
282namespace
283{
284
285/* Build the maximal scop containing LOOPs and add it to SCOPS. */
286
287class scop_detection
288{
289public:
290 scop_detection () : scops (vNULL) {}
291
5e6359b9 292 ~scop_detection ()
293 {
294 scops.release ();
295 }
296
edbec012 297 /* A marker for invalid sese_l. */
298 static sese_l invalid_sese;
299
300 /* Return the SCOPS in this SCOP_DETECTION. */
301
302 vec<sese_l>
303 get_scops ()
304 {
305 return scops;
306 }
307
308 /* Return an sese_l around the LOOP. */
309
310 sese_l get_sese (loop_p loop);
311
edbec012 312 /* Merge scops at same loop depth and returns the new sese.
313 Returns a new SESE when merge was successful, INVALID_SESE otherwise. */
314
315 sese_l merge_sese (sese_l first, sese_l second) const;
316
317 /* Build scop outer->inner if possible. */
318
cb44892e 319 void build_scop_depth (loop_p loop);
edbec012 320
321 /* Return true when BEGIN is the preheader edge of a loop with a single exit
322 END. */
323
324 static bool region_has_one_loop (sese_l s);
325
326 /* Add to SCOPS a scop starting at SCOP_BEGIN and ending at SCOP_END. */
327
328 void add_scop (sese_l s);
329
330 /* Returns true if S1 subsumes/surrounds S2. */
331 static bool subsumes (sese_l s1, sese_l s2);
332
333 /* Remove a SCoP which is subsumed by S1. */
334 void remove_subscops (sese_l s1);
335
336 /* Returns true if S1 intersects with S2. Since we already know that S1 does
337 not subsume S2 or vice-versa, we only check for entry bbs. */
338
339 static bool intersects (sese_l s1, sese_l s2);
340
341 /* Remove one of the scops when it intersects with any other. */
342
343 void remove_intersecting_scops (sese_l s1);
344
8affe2f6 345 /* Return true when a statement in SCOP cannot be represented by Graphite. */
edbec012 346
b8830caa 347 bool harmful_loop_in_region (sese_l scop) const;
edbec012 348
349 /* Return true only when STMT is simple enough for being handled by Graphite.
350 This depends on SCOP, as the parameters are initialized relatively to
351 this basic block, the linear functions are initialized based on the
352 outermost loop containing STMT inside the SCOP. BB is the place where we
353 try to evaluate the STMT. */
354
355 bool stmt_simple_for_scop_p (sese_l scop, gimple *stmt,
356 basic_block bb) const;
357
358 /* Something like "n * m" is not allowed. */
359
360 static bool graphite_can_represent_init (tree e);
361
362 /* Return true when SCEV can be represented in the polyhedral model.
363
364 An expression can be represented, if it can be expressed as an
365 affine expression. For loops (i, j) and parameters (m, n) all
366 affine expressions are of the form:
367
368 x1 * i + x2 * j + x3 * m + x4 * n + x5 * 1 where x1..x5 element of Z
369
370 1 i + 20 j + (-2) m + 25
371
372 Something like "i * n" or "n * m" is not allowed. */
373
453841f9 374 static bool graphite_can_represent_scev (sese_l scop, tree scev);
edbec012 375
376 /* Return true when EXPR can be represented in the polyhedral model.
377
378 This means an expression can be represented, if it is linear with respect
379 to the loops and the strides are non parametric. LOOP is the place where
380 the expr will be evaluated. SCOP defines the region we analyse. */
381
382 static bool graphite_can_represent_expr (sese_l scop, loop_p loop,
383 tree expr);
384
385 /* Return true if the data references of STMT can be represented by Graphite.
386 We try to analyze the data references in a loop contained in the SCOP. */
387
388 static bool stmt_has_simple_data_refs_p (sese_l scop, gimple *stmt);
389
390 /* Remove the close phi node at GSI and replace its rhs with the rhs
391 of PHI. */
392
393 static void remove_duplicate_close_phi (gphi *phi, gphi_iterator *gsi);
394
395 /* Returns true when Graphite can represent LOOP in SCOP.
396 FIXME: For the moment, graphite cannot be used on loops that iterate using
397 induction variables that wrap. */
398
edbec012 399 static bool can_represent_loop (loop_p loop, sese_l scop);
400
edbec012 401 /* Returns the number of pbbs that are in loops contained in SCOP. */
402
403 static int nb_pbbs_in_loops (scop_p scop);
404
edbec012 405private:
406 vec<sese_l> scops;
407};
408
5828c94d 409sese_l scop_detection::invalid_sese (NULL, NULL);
edbec012 410
411/* Return an sese_l around the LOOP. */
412
413sese_l
414scop_detection::get_sese (loop_p loop)
415{
416 if (!loop)
417 return invalid_sese;
418
c8459b28 419 edge scop_begin = loop_preheader_edge (loop);
edbec012 420 edge scop_end = single_exit (loop);
b42450b3 421 if (!scop_end || (scop_end->flags & (EDGE_COMPLEX|EDGE_FAKE)))
edbec012 422 return invalid_sese;
c8459b28 423
424 return sese_l (scop_begin, scop_end);
edbec012 425}
426
edbec012 427/* Merge scops at same loop depth and returns the new sese.
428 Returns a new SESE when merge was successful, INVALID_SESE otherwise. */
429
430sese_l
431scop_detection::merge_sese (sese_l first, sese_l second) const
432{
433 /* In the trivial case first/second may be NULL. */
434 if (!first)
435 return second;
436 if (!second)
437 return first;
438
b8830caa 439 DEBUG_PRINT (dp << "[scop-detection] try merging sese s1: ";
440 print_sese (dump_file, first);
441 dp << "[scop-detection] try merging sese s2: ";
edbec012 442 print_sese (dump_file, second));
443
1290f093 444 auto_bitmap worklist, in_sese_region;
445 bitmap_set_bit (worklist, get_entry_bb (first)->index);
446 bitmap_set_bit (worklist, get_exit_bb (first)->index);
447 bitmap_set_bit (worklist, get_entry_bb (second)->index);
448 bitmap_set_bit (worklist, get_exit_bb (second)->index);
449 edge entry = NULL, exit = NULL;
450
451 /* We can optimize the case of adding a loop entry dest or exit
452 src to the worklist (for single-exit loops) by skipping
453 directly to the exit dest / entry src. in_sese_region
454 doesn't have to cover all blocks in the region but merely
455 its border it acts more like a visited bitmap. */
456 do
55fec184 457 {
1290f093 458 int index = bitmap_first_set_bit (worklist);
459 bitmap_clear_bit (worklist, index);
460 basic_block bb = BASIC_BLOCK_FOR_FN (cfun, index);
461 edge_iterator ei;
462 edge e;
463
464 /* With fake exit edges we can end up with no possible exit. */
465 if (index == EXIT_BLOCK)
55fec184 466 {
1290f093 467 DEBUG_PRINT (dp << "[scop-detection-fail] cannot merge seses.\n");
468 return invalid_sese;
55fec184 469 }
1290f093 470
471 bitmap_set_bit (in_sese_region, bb->index);
472
473 basic_block dom = get_immediate_dominator (CDI_DOMINATORS, bb);
474 FOR_EACH_EDGE (e, ei, bb->preds)
475 if (e->src == dom
476 && (! entry
477 || dominated_by_p (CDI_DOMINATORS, entry->dest, bb)))
478 {
479 if (entry
480 && ! bitmap_bit_p (in_sese_region, entry->src->index))
481 bitmap_set_bit (worklist, entry->src->index);
482 entry = e;
483 }
484 else if (! bitmap_bit_p (in_sese_region, e->src->index))
485 bitmap_set_bit (worklist, e->src->index);
486
487 basic_block pdom = get_immediate_dominator (CDI_POST_DOMINATORS, bb);
488 FOR_EACH_EDGE (e, ei, bb->succs)
489 if (e->dest == pdom
490 && (! exit
491 || dominated_by_p (CDI_POST_DOMINATORS, exit->src, bb)))
492 {
493 if (exit
494 && ! bitmap_bit_p (in_sese_region, exit->dest->index))
495 bitmap_set_bit (worklist, exit->dest->index);
496 exit = e;
497 }
498 else if (! bitmap_bit_p (in_sese_region, e->dest->index))
499 bitmap_set_bit (worklist, e->dest->index);
55fec184 500 }
1290f093 501 while (! bitmap_empty_p (worklist));
55fec184 502
1290f093 503 sese_l combined (entry, exit);
504
edbec012 505 DEBUG_PRINT (dp << "[merged-sese] s1: "; print_sese (dump_file, combined));
506
507 return combined;
508}
509
510/* Build scop outer->inner if possible. */
511
cb44892e 512void
513scop_detection::build_scop_depth (loop_p loop)
edbec012 514{
cb44892e 515 sese_l s = invalid_sese;
516 loop = loop->inner;
517 while (loop)
edbec012 518 {
cb44892e 519 sese_l next = get_sese (loop);
520 if (! next
521 || harmful_loop_in_region (next))
d5697ffe 522 {
cb44892e 523 if (s)
524 add_scop (s);
525 build_scop_depth (loop);
526 s = invalid_sese;
d5697ffe 527 }
cb44892e 528 else if (! s)
529 s = next;
530 else
531 {
532 sese_l combined = merge_sese (s, next);
533 if (! combined
534 || harmful_loop_in_region (combined))
535 {
536 add_scop (s);
537 s = next;
538 }
539 else
540 s = combined;
541 }
542 loop = loop->next;
543 }
544 if (s)
545 add_scop (s);
edbec012 546}
547
548/* Returns true when Graphite can represent LOOP in SCOP.
549 FIXME: For the moment, graphite cannot be used on loops that iterate using
550 induction variables that wrap. */
551
552bool
cb44892e 553scop_detection::can_represent_loop (loop_p loop, sese_l scop)
edbec012 554{
555 tree niter;
556 struct tree_niter_desc niter_desc;
557
794fefda 558 /* We can only handle do {} while () style loops correctly. */
559 edge exit = single_exit (loop);
560 if (!exit
561 || !single_pred_p (loop->latch)
562 || exit->src != single_pred (loop->latch)
563 || !empty_block_p (loop->latch))
564 return false;
565
566 return !(loop_preheader_edge (loop)->flags & EDGE_IRREDUCIBLE_LOOP)
edbec012 567 && number_of_iterations_exit (loop, single_exit (loop), &niter_desc, false)
568 && niter_desc.control.no_overflow
569 && (niter = number_of_latch_executions (loop))
570 && !chrec_contains_undetermined (niter)
571 && graphite_can_represent_expr (scop, loop, niter);
572}
573
edbec012 574/* Return true when BEGIN is the preheader edge of a loop with a single exit
575 END. */
576
577bool
578scop_detection::region_has_one_loop (sese_l s)
579{
580 edge begin = s.entry;
581 edge end = s.exit;
582 /* Check for a single perfectly nested loop. */
583 if (begin->dest->loop_father->inner)
584 return false;
585
586 /* Otherwise, check whether we have adjacent loops. */
c8459b28 587 return (single_pred_p (end->src)
588 && begin->dest->loop_father == single_pred (end->src)->loop_father);
edbec012 589}
590
591/* Add to SCOPS a scop starting at SCOP_BEGIN and ending at SCOP_END. */
592
593void
594scop_detection::add_scop (sese_l s)
595{
596 gcc_assert (s);
597
ca74fb2b 598 /* If the exit edge is fake discard the SCoP for now as we're removing the
599 fake edges again after analysis. */
600 if (s.exit->flags & EDGE_FAKE)
601 {
602 DEBUG_PRINT (dp << "[scop-detection-fail] Discarding infinite loop SCoP: ";
603 print_sese (dump_file, s));
604 return;
605 }
606
c56f248d 607 /* Include the BB with the loop-closed SSA PHI nodes, we need this
608 block in the region for code-generating out-of-SSA copies.
609 canonicalize_loop_closed_ssa makes sure that is in proper shape. */
610 if (s.exit->dest != EXIT_BLOCK_PTR_FOR_FN (cfun)
611 && loop_exit_edge_p (s.exit->src->loop_father, s.exit))
612 {
613 gcc_assert (single_pred_p (s.exit->dest)
614 && single_succ_p (s.exit->dest)
615 && sese_trivially_empty_bb_p (s.exit->dest));
616 s.exit = single_succ_edge (s.exit->dest);
617 }
618
edbec012 619 /* Do not add scops with only one loop. */
620 if (region_has_one_loop (s))
621 {
b8830caa 622 DEBUG_PRINT (dp << "[scop-detection-fail] Discarding one loop SCoP: ";
edbec012 623 print_sese (dump_file, s));
624 return;
625 }
626
f032380c 627 if (get_exit_bb (s) == EXIT_BLOCK_PTR_FOR_FN (cfun))
edbec012 628 {
958d01f3 629 DEBUG_PRINT (dp << "[scop-detection-fail] "
b8830caa 630 << "Discarding SCoP exiting to return: ";
edbec012 631 print_sese (dump_file, s));
632 return;
633 }
634
635 /* Remove all the scops which are subsumed by s. */
636 remove_subscops (s);
637
4caef4a5 638 /* Remove intersecting scops. FIXME: It will be a good idea to keep
639 the non-intersecting part of the scop already in the list. */
edbec012 640 remove_intersecting_scops (s);
c6bb733d 641
edbec012 642 scops.safe_push (s);
b8830caa 643 DEBUG_PRINT (dp << "[scop-detection] Adding SCoP: "; print_sese (dump_file, s));
7eb20e71 644}
c6bb733d 645
8affe2f6 646/* Return true when a statement in SCOP cannot be represented by Graphite. */
edbec012 647
648bool
b8830caa 649scop_detection::harmful_loop_in_region (sese_l scop) const
c6bb733d 650{
f032380c 651 basic_block exit_bb = get_exit_bb (scop);
652 basic_block entry_bb = get_entry_bb (scop);
c6bb733d 653
958d01f3 654 DEBUG_PRINT (dp << "[checking-harmful-bbs] ";
edbec012 655 print_sese (dump_file, scop));
656 gcc_assert (dominated_by_p (CDI_DOMINATORS, exit_bb, entry_bb));
c6bb733d 657
67477b79 658 auto_vec<basic_block> worklist;
659 auto_bitmap loops;
c6bb733d 660
67477b79 661 worklist.safe_push (entry_bb);
662 while (! worklist.is_empty ())
edbec012 663 {
67477b79 664 basic_block bb = worklist.pop ();
ac0b9d89 665 DEBUG_PRINT (dp << "Visiting bb_" << bb->index << "\n");
c6bb733d 666
9e3b8c11 667 /* The basic block should not be part of an irreducible loop. */
668 if (bb->flags & BB_IRREDUCIBLE_LOOP)
67477b79 669 return true;
9e3b8c11 670
dc06f294 671 /* Check for unstructured control flow: CFG not generated by structured
672 if-then-else. */
673 if (bb->succs->length () > 1)
674 {
675 edge e;
676 edge_iterator ei;
677 FOR_EACH_EDGE (e, ei, bb->succs)
678 if (!dominated_by_p (CDI_POST_DOMINATORS, bb, e->dest)
679 && !dominated_by_p (CDI_DOMINATORS, e->dest, bb))
680 return true;
681 }
682
b8830caa 683 /* Collect all loops in the current region. */
684 loop_p loop = bb->loop_father;
685 if (loop_in_sese_p (loop, scop))
686 bitmap_set_bit (loops, loop->num);
cb44892e 687
688 /* Check for harmful statements in basic blocks part of the region. */
689 for (gimple_stmt_iterator gsi = gsi_start_bb (bb);
690 !gsi_end_p (gsi); gsi_next (&gsi))
691 if (!stmt_simple_for_scop_p (scop, gsi_stmt (gsi), bb))
692 return true;
b8830caa 693
e7e5bab6 694 for (basic_block dom = first_dom_son (CDI_DOMINATORS, bb);
695 dom;
696 dom = next_dom_son (CDI_DOMINATORS, dom))
697 if (dom != scop.exit->dest)
67477b79 698 worklist.safe_push (dom);
b8830caa 699 }
700
701 /* Go through all loops and check that they are still valid in the combined
702 scop. */
703 unsigned j;
704 bitmap_iterator bi;
705 EXECUTE_IF_SET_IN_BITMAP (loops, 0, j, bi)
706 {
707 loop_p loop = (*current_loops->larray)[j];
708 gcc_assert (loop->num == (int) j);
709
cb44892e 710 /* Check if the loop nests are to be optimized for speed. */
711 if (! loop->inner
712 && ! optimize_loop_for_speed_p (loop))
713 {
714 DEBUG_PRINT (dp << "[scop-detection-fail] loop_"
715 << loop->num << " is not on a hot path.\n");
716 return true;
717 }
718
719 if (! can_represent_loop (loop, scop))
720 {
721 DEBUG_PRINT (dp << "[scop-detection-fail] cannot represent loop_"
722 << loop->num << "\n");
723 return true;
724 }
725
cb44892e 726 /* Check if all loop nests have at least one data reference.
727 ??? This check is expensive and loops premature at this point.
728 If important to retain we can pre-compute this for all innermost
729 loops and reject those when we build a SESE region for a loop
730 during SESE discovery. */
731 if (! loop->inner
732 && ! loop_nest_has_data_refs (loop))
733 {
734 DEBUG_PRINT (dp << "[scop-detection-fail] loop_" << loop->num
735 << "does not have any data reference.\n");
736 return true;
737 }
edbec012 738 }
739
5e6359b9 740 return false;
edbec012 741}
742
743/* Returns true if S1 subsumes/surrounds S2. */
744bool
745scop_detection::subsumes (sese_l s1, sese_l s2)
c6bb733d 746{
f032380c 747 if (dominated_by_p (CDI_DOMINATORS, get_entry_bb (s2),
748 get_entry_bb (s1))
749 && dominated_by_p (CDI_POST_DOMINATORS, s2.exit->dest,
750 s1.exit->dest))
edbec012 751 return true;
752 return false;
753}
c6bb733d 754
edbec012 755/* Remove a SCoP which is subsumed by S1. */
756void
757scop_detection::remove_subscops (sese_l s1)
758{
759 int j;
5828c94d 760 sese_l *s2;
edbec012 761 FOR_EACH_VEC_ELT_REVERSE (scops, j, s2)
762 {
5828c94d 763 if (subsumes (s1, *s2))
edbec012 764 {
958d01f3 765 DEBUG_PRINT (dp << "Removing sub-SCoP";
5828c94d 766 print_sese (dump_file, *s2));
edbec012 767 scops.unordered_remove (j);
768 }
769 }
770}
c6bb733d 771
edbec012 772/* Returns true if S1 intersects with S2. Since we already know that S1 does
773 not subsume S2 or vice-versa, we only check for entry bbs. */
774
775bool
776scop_detection::intersects (sese_l s1, sese_l s2)
777{
f032380c 778 if (dominated_by_p (CDI_DOMINATORS, get_entry_bb (s2),
779 get_entry_bb (s1))
780 && !dominated_by_p (CDI_DOMINATORS, get_entry_bb (s2),
781 get_exit_bb (s1)))
edbec012 782 return true;
783 if ((s1.exit == s2.entry) || (s2.exit == s1.entry))
784 return true;
785
786 return false;
c6bb733d 787}
788
edbec012 789/* Remove one of the scops when it intersects with any other. */
7eb20e71 790
edbec012 791void
792scop_detection::remove_intersecting_scops (sese_l s1)
793{
794 int j;
5828c94d 795 sese_l *s2;
edbec012 796 FOR_EACH_VEC_ELT_REVERSE (scops, j, s2)
797 {
5828c94d 798 if (intersects (s1, *s2))
edbec012 799 {
958d01f3 800 DEBUG_PRINT (dp << "Removing intersecting SCoP";
801 print_sese (dump_file, *s2);
802 dp << "Intersects with:";
edbec012 803 print_sese (dump_file, s1));
804 scops.unordered_remove (j);
805 }
806 }
807}
7eb20e71 808
c6bb733d 809/* Something like "n * m" is not allowed. */
810
edbec012 811bool
812scop_detection::graphite_can_represent_init (tree e)
c6bb733d 813{
814 switch (TREE_CODE (e))
815 {
816 case POLYNOMIAL_CHREC:
817 return graphite_can_represent_init (CHREC_LEFT (e))
818 && graphite_can_represent_init (CHREC_RIGHT (e));
819
820 case MULT_EXPR:
821 if (chrec_contains_symbols (TREE_OPERAND (e, 0)))
7464e753 822 return graphite_can_represent_init (TREE_OPERAND (e, 0))
35ec552a 823 && tree_fits_shwi_p (TREE_OPERAND (e, 1));
c6bb733d 824 else
7464e753 825 return graphite_can_represent_init (TREE_OPERAND (e, 1))
35ec552a 826 && tree_fits_shwi_p (TREE_OPERAND (e, 0));
c6bb733d 827
828 case PLUS_EXPR:
829 case POINTER_PLUS_EXPR:
830 case MINUS_EXPR:
831 return graphite_can_represent_init (TREE_OPERAND (e, 0))
832 && graphite_can_represent_init (TREE_OPERAND (e, 1));
833
834 case NEGATE_EXPR:
835 case BIT_NOT_EXPR:
836 CASE_CONVERT:
837 case NON_LVALUE_EXPR:
838 return graphite_can_represent_init (TREE_OPERAND (e, 0));
839
edbec012 840 default:
841 break;
c6bb733d 842 }
843
844 return true;
845}
846
847/* Return true when SCEV can be represented in the polyhedral model.
848
849 An expression can be represented, if it can be expressed as an
850 affine expression. For loops (i, j) and parameters (m, n) all
851 affine expressions are of the form:
852
853 x1 * i + x2 * j + x3 * m + x4 * n + x5 * 1 where x1..x5 element of Z
854
855 1 i + 20 j + (-2) m + 25
856
e3135850 857 Something like "i * n" or "n * m" is not allowed. */
c6bb733d 858
edbec012 859bool
453841f9 860scop_detection::graphite_can_represent_scev (sese_l scop, tree scev)
c6bb733d 861{
862 if (chrec_contains_undetermined (scev))
863 return false;
864
99c136a5 865 switch (TREE_CODE (scev))
866 {
98acb419 867 case NEGATE_EXPR:
868 case BIT_NOT_EXPR:
869 CASE_CONVERT:
870 case NON_LVALUE_EXPR:
453841f9 871 return graphite_can_represent_scev (scop, TREE_OPERAND (scev, 0));
98acb419 872
99c136a5 873 case PLUS_EXPR:
98acb419 874 case POINTER_PLUS_EXPR:
99c136a5 875 case MINUS_EXPR:
453841f9 876 return graphite_can_represent_scev (scop, TREE_OPERAND (scev, 0))
877 && graphite_can_represent_scev (scop, TREE_OPERAND (scev, 1));
c6bb733d 878
99c136a5 879 case MULT_EXPR:
880 return !CONVERT_EXPR_CODE_P (TREE_CODE (TREE_OPERAND (scev, 0)))
881 && !CONVERT_EXPR_CODE_P (TREE_CODE (TREE_OPERAND (scev, 1)))
882 && !(chrec_contains_symbols (TREE_OPERAND (scev, 0))
883 && chrec_contains_symbols (TREE_OPERAND (scev, 1)))
ce0ae3b6 884 && graphite_can_represent_init (scev)
453841f9 885 && graphite_can_represent_scev (scop, TREE_OPERAND (scev, 0))
886 && graphite_can_represent_scev (scop, TREE_OPERAND (scev, 1));
c6bb733d 887
99c136a5 888 case POLYNOMIAL_CHREC:
889 /* Check for constant strides. With a non constant stride of
890 'n' we would have a value of 'iv * n'. Also check that the
891 initial value can represented: for example 'n * m' cannot be
892 represented. */
453841f9 893 gcc_assert (loop_in_sese_p (get_loop (cfun,
894 CHREC_VARIABLE (scev)), scop));
99c136a5 895 if (!evolution_function_right_is_integer_cst (scev)
896 || !graphite_can_represent_init (scev))
897 return false;
453841f9 898 return graphite_can_represent_scev (scop, CHREC_LEFT (scev));
99c136a5 899
b9ee1c20 900 case ADDR_EXPR:
901 /* We cannot encode addresses for ISL. */
902 return false;
903
99c136a5 904 default:
905 break;
906 }
c6bb733d 907
908 /* Only affine functions can be represented. */
edbec012 909 if (tree_contains_chrecs (scev, NULL) || !scev_is_linear_expression (scev))
c6bb733d 910 return false;
911
629787af 912 return true;
c6bb733d 913}
914
c6bb733d 915/* Return true when EXPR can be represented in the polyhedral model.
916
7eb20e71 917 This means an expression can be represented, if it is linear with respect to
918 the loops and the strides are non parametric. LOOP is the place where the
919 expr will be evaluated. SCOP defines the region we analyse. */
c6bb733d 920
edbec012 921bool
922scop_detection::graphite_can_represent_expr (sese_l scop, loop_p loop,
923 tree expr)
c6bb733d 924{
d75855b4 925 tree scev = cached_scalar_evolution_in_region (scop, loop, expr);
453841f9 926 return graphite_can_represent_scev (scop, scev);
c6bb733d 927}
928
7eb20e71 929/* Return true if the data references of STMT can be represented by Graphite.
930 We try to analyze the data references in a loop contained in the SCOP. */
c6bb733d 931
edbec012 932bool
933scop_detection::stmt_has_simple_data_refs_p (sese_l scop, gimple *stmt)
c6bb733d 934{
7f38a6aa 935 edge nest = scop.entry;
443b5bd0 936 loop_p loop = loop_containing_stmt (stmt);
28e7ffc9 937 if (!loop_in_sese_p (loop, scop))
44e2f332 938 loop = NULL;
e97c4b0d 939
28e7ffc9 940 auto_vec<data_reference_p> drs;
941 if (! graphite_find_data_references_in_stmt (nest, loop, stmt, &drs))
942 return false;
443b5bd0 943
944 int j;
945 data_reference_p dr;
946 FOR_EACH_VEC_ELT (drs, j, dr)
e97c4b0d 947 {
28e7ffc9 948 for (unsigned i = 0; i < DR_NUM_DIMENSIONS (dr); ++i)
453841f9 949 if (! graphite_can_represent_scev (scop, DR_ACCESS_FN (dr, i)))
b98a7d5b 950 return false;
e97c4b0d 951 }
c6bb733d 952
28e7ffc9 953 return true;
c6bb733d 954}
955
29663955 956/* GIMPLE_ASM and GIMPLE_CALL may embed arbitrary side effects.
957 Calls have side-effects, except those to const or pure
958 functions. */
c6bb733d 959
960static bool
29663955 961stmt_has_side_effects (gimple *stmt)
c6bb733d 962{
c6bb733d 963 if (gimple_has_volatile_ops (stmt)
964 || (gimple_code (stmt) == GIMPLE_CALL
965 && !(gimple_call_flags (stmt) & (ECF_CONST | ECF_PURE)))
966 || (gimple_code (stmt) == GIMPLE_ASM))
4773ab25 967 {
7eb20e71 968 DEBUG_PRINT (dp << "[scop-detection-fail] "
29663955 969 << "Statement has side-effects:\n";
edbec012 970 print_gimple_stmt (dump_file, stmt, 0, TDF_VOPS | TDF_MEMSYMS));
29663955 971 return true;
4773ab25 972 }
29663955 973 return false;
974}
c6bb733d 975
cb44892e 976/* Return true only when STMT is simple enough for being handled by Graphite.
977 This depends on SCOP, as the parameters are initialized relatively to
978 this basic block, the linear functions are initialized based on the outermost
979 loop containing STMT inside the SCOP. BB is the place where we try to
980 evaluate the STMT. */
c6bb733d 981
edbec012 982bool
cb44892e 983scop_detection::stmt_simple_for_scop_p (sese_l scop, gimple *stmt,
984 basic_block bb) const
29663955 985{
cb44892e 986 gcc_assert (scop);
987
988 if (is_gimple_debug (stmt))
989 return true;
990
991 if (stmt_has_side_effects (stmt))
992 return false;
993
994 if (!stmt_has_simple_data_refs_p (scop, stmt))
995 {
996 DEBUG_PRINT (dp << "[scop-detection-fail] "
997 << "Graphite cannot handle data-refs in stmt:\n";
998 print_gimple_stmt (dump_file, stmt, 0, TDF_VOPS|TDF_MEMSYMS););
999 return false;
1000 }
1001
c6bb733d 1002 switch (gimple_code (stmt))
1003 {
c6bb733d 1004 case GIMPLE_LABEL:
1005 return true;
1006
1007 case GIMPLE_COND:
1008 {
c6bb733d 1009 /* We can handle all binary comparisons. Inequalities are
1010 also supported as they can be represented with union of
1011 polyhedra. */
29663955 1012 enum tree_code code = gimple_cond_code (stmt);
1013 if (!(code == LT_EXPR
c6bb733d 1014 || code == GT_EXPR
1015 || code == LE_EXPR
1016 || code == GE_EXPR
1017 || code == EQ_EXPR
1018 || code == NE_EXPR))
29663955 1019 {
1020 DEBUG_PRINT (dp << "[scop-detection-fail] "
7eb20e71 1021 << "Graphite cannot handle cond stmt:\n";
edbec012 1022 print_gimple_stmt (dump_file, stmt, 0,
1023 TDF_VOPS | TDF_MEMSYMS));
4773ab25 1024 return false;
1025 }
c6bb733d 1026
cb44892e 1027 loop_p loop = bb->loop_father;
5da4c394 1028 for (unsigned i = 0; i < 2; ++i)
1029 {
1030 tree op = gimple_op (stmt, i);
7eb20e71 1031 if (!graphite_can_represent_expr (scop, loop, op)
9852e66f 1032 /* We can only constrain on integer type. */
6e2a6380 1033 || ! INTEGRAL_TYPE_P (TREE_TYPE (op)))
4773ab25 1034 {
edbec012 1035 DEBUG_PRINT (dp << "[scop-detection-fail] "
1036 << "Graphite cannot represent stmt:\n";
1037 print_gimple_stmt (dump_file, stmt, 0,
1038 TDF_VOPS | TDF_MEMSYMS));
4773ab25 1039 return false;
1040 }
5da4c394 1041 }
c6bb733d 1042
1043 return true;
1044 }
1045
1046 case GIMPLE_ASSIGN:
c6bb733d 1047 case GIMPLE_CALL:
97744494 1048 {
583ab4de 1049 tree op, lhs = gimple_get_lhs (stmt);
97744494 1050 ssa_op_iter i;
583ab4de 1051 /* If we are not going to instantiate the stmt do not require
1052 its operands to be instantiatable at this point. */
1053 if (lhs
1054 && TREE_CODE (lhs) == SSA_NAME
1055 && scev_analyzable_p (lhs, scop))
1056 return true;
97744494 1057 /* Verify that if we can analyze operands at their def site we
1058 also can represent them when analyzed at their uses. */
1059 FOR_EACH_SSA_TREE_OPERAND (op, stmt, i, SSA_OP_USE)
1060 if (scev_analyzable_p (op, scop)
583ab4de 1061 && chrec_contains_undetermined
d75855b4 1062 (cached_scalar_evolution_in_region (scop,
1063 bb->loop_father, op)))
97744494 1064 {
1065 DEBUG_PRINT (dp << "[scop-detection-fail] "
583ab4de 1066 << "Graphite cannot code-gen stmt:\n";
97744494 1067 print_gimple_stmt (dump_file, stmt, 0,
1068 TDF_VOPS | TDF_MEMSYMS));
1069 return false;
1070 }
1071 return true;
1072 }
c6bb733d 1073
1074 default:
1075 /* These nodes cut a new scope. */
edbec012 1076 DEBUG_PRINT (
1077 dp << "[scop-detection-fail] "
1078 << "Gimple stmt not handled in Graphite:\n";
1079 print_gimple_stmt (dump_file, stmt, 0, TDF_VOPS | TDF_MEMSYMS));
c6bb733d 1080 return false;
1081 }
29663955 1082}
c6bb733d 1083
edbec012 1084/* Returns the number of pbbs that are in loops contained in SCOP. */
7eb20e71 1085
edbec012 1086int
1087scop_detection::nb_pbbs_in_loops (scop_p scop)
1088{
1089 int i;
1090 poly_bb_p pbb;
1091 int res = 0;
7eb20e71 1092
0e526381 1093 FOR_EACH_VEC_ELT (scop->pbbs, i, pbb)
5828c94d 1094 if (loop_in_sese_p (gbb_loop (PBB_BLACK_BOX (pbb)), scop->scop_info->region))
edbec012 1095 res++;
7eb20e71 1096
edbec012 1097 return res;
1098}
7eb20e71 1099
6e2a6380 1100/* Assigns the parameter NAME an index in REGION. */
118a202b 1101
6e2a6380 1102static void
1103assign_parameter_index_in_region (tree name, sese_info_p region)
118a202b 1104{
6e2a6380 1105 gcc_assert (TREE_CODE (name) == SSA_NAME
1106 && INTEGRAL_TYPE_P (TREE_TYPE (name))
1107 && ! defined_in_sese_p (name, region->region));
118a202b 1108 int i;
1109 tree p;
30162daa 1110 FOR_EACH_VEC_ELT (region->params, i, p)
118a202b 1111 if (p == name)
6e2a6380 1112 return;
118a202b 1113
30162daa 1114 region->params.safe_push (name);
118a202b 1115}
1116
1117/* In the context of sese S, scan the expression E and translate it to
1118 a linear expression C. When parsing a symbolic multiplication, K
1119 represents the constant multiplier of an expression containing
1120 parameters. */
1121
1122static void
f032380c 1123scan_tree_for_params (sese_info_p s, tree e)
118a202b 1124{
1125 if (e == chrec_dont_know)
1126 return;
1127
1128 switch (TREE_CODE (e))
1129 {
1130 case POLYNOMIAL_CHREC:
1131 scan_tree_for_params (s, CHREC_LEFT (e));
1132 break;
1133
1134 case MULT_EXPR:
1135 if (chrec_contains_symbols (TREE_OPERAND (e, 0)))
1136 scan_tree_for_params (s, TREE_OPERAND (e, 0));
1137 else
1138 scan_tree_for_params (s, TREE_OPERAND (e, 1));
1139 break;
1140
1141 case PLUS_EXPR:
1142 case POINTER_PLUS_EXPR:
1143 case MINUS_EXPR:
1144 scan_tree_for_params (s, TREE_OPERAND (e, 0));
1145 scan_tree_for_params (s, TREE_OPERAND (e, 1));
1146 break;
1147
1148 case NEGATE_EXPR:
1149 case BIT_NOT_EXPR:
1150 CASE_CONVERT:
1151 case NON_LVALUE_EXPR:
1152 scan_tree_for_params (s, TREE_OPERAND (e, 0));
1153 break;
1154
1155 case SSA_NAME:
6e2a6380 1156 assign_parameter_index_in_region (e, s);
118a202b 1157 break;
1158
1159 case INTEGER_CST:
1160 case ADDR_EXPR:
1161 case REAL_CST:
1162 case COMPLEX_CST:
1163 case VECTOR_CST:
1164 break;
1165
1166 default:
1167 gcc_unreachable ();
1168 break;
1169 }
1170}
1171
1172/* Find parameters with respect to REGION in BB. We are looking in memory
1173 access functions, conditions and loop bounds. */
1174
1175static void
f032380c 1176find_params_in_bb (sese_info_p region, gimple_poly_bb_p gbb)
118a202b 1177{
3b9ce1aa 1178 /* Find parameters in the access functions of data references. */
118a202b 1179 int i;
118a202b 1180 data_reference_p dr;
118a202b 1181 FOR_EACH_VEC_ELT (GBB_DATA_REFS (gbb), i, dr)
3b9ce1aa 1182 for (unsigned j = 0; j < DR_NUM_DIMENSIONS (dr); j++)
118a202b 1183 scan_tree_for_params (region, DR_ACCESS_FN (dr, j));
1184
1185 /* Find parameters in conditional statements. */
3b9ce1aa 1186 gimple *stmt;
118a202b 1187 FOR_EACH_VEC_ELT (GBB_CONDITIONS (gbb), i, stmt)
1188 {
78395978 1189 loop_p loop = gimple_bb (stmt)->loop_father;
d75855b4 1190 tree lhs = cached_scalar_evolution_in_region (region->region, loop,
1191 gimple_cond_lhs (stmt));
1192 tree rhs = cached_scalar_evolution_in_region (region->region, loop,
1193 gimple_cond_rhs (stmt));
78395978 1194 gcc_assert (!chrec_contains_undetermined (lhs)
1195 && !chrec_contains_undetermined (rhs));
118a202b 1196
1197 scan_tree_for_params (region, lhs);
1198 scan_tree_for_params (region, rhs);
1199 }
1200}
1201
f47117d1 1202/* Record the parameters used in the SCOP BBs. A variable is a parameter
118a202b 1203 in a scop if it does not vary during the execution of that scop. */
1204
1205static void
1206find_scop_parameters (scop_p scop)
1207{
118a202b 1208 unsigned i;
5828c94d 1209 sese_info_p region = scop->scop_info;
118a202b 1210
f47117d1 1211 /* Parameters used in loop bounds are processed during gather_bbs. */
118a202b 1212
1213 /* Find the parameters used in data accesses. */
3b9ce1aa 1214 poly_bb_p pbb;
0e526381 1215 FOR_EACH_VEC_ELT (scop->pbbs, i, pbb)
118a202b 1216 find_params_in_bb (region, PBB_BLACK_BOX (pbb));
1217
3b9ce1aa 1218 int nbp = sese_nb_params (region);
118a202b 1219 scop_set_nb_params (scop, nbp);
118a202b 1220}
1221
74936b22 1222static void
1223add_write (vec<tree> *writes, tree def)
1224{
1225 writes->safe_push (def);
1226 DEBUG_PRINT (dp << "Adding scalar write: ";
1227 print_generic_expr (dump_file, def);
1228 dp << "\nFrom stmt: ";
1229 print_gimple_stmt (dump_file,
1230 SSA_NAME_DEF_STMT (def), 0));
1231}
1232
1233static void
1234add_read (vec<scalar_use> *reads, tree use, gimple *use_stmt)
1235{
1236 DEBUG_PRINT (dp << "Adding scalar read: ";
1237 print_generic_expr (dump_file, use);
1238 dp << "\nFrom stmt: ";
1239 print_gimple_stmt (dump_file, use_stmt, 0));
1240 reads->safe_push (std::make_pair (use_stmt, use));
1241}
1242
1243
30162daa 1244/* Record DEF if it is used in other bbs different than DEF_BB in the SCOP. */
1245
1246static void
1247build_cross_bb_scalars_def (scop_p scop, tree def, basic_block def_bb,
1248 vec<tree> *writes)
1249{
74936b22 1250 if (!is_gimple_reg (def))
30162daa 1251 return;
1252
38725f99 1253 bool scev_analyzable = scev_analyzable_p (def, scop->scop_info->region);
30162daa 1254
1255 gimple *use_stmt;
1256 imm_use_iterator imm_iter;
1257 FOR_EACH_IMM_USE_STMT (use_stmt, imm_iter, def)
38725f99 1258 /* Do not gather scalar variables that can be analyzed by SCEV as they can
1259 be generated out of the induction variables. */
1260 if ((! scev_analyzable
1261 /* But gather SESE liveouts as we otherwise fail to rewrite their
1262 exit PHIs. */
1263 || ! bb_in_sese_p (gimple_bb (use_stmt), scop->scop_info->region))
74936b22 1264 && (def_bb != gimple_bb (use_stmt) && !is_gimple_debug (use_stmt)))
30162daa 1265 {
74936b22 1266 add_write (writes, def);
30162daa 1267 /* This is required by the FOR_EACH_IMM_USE_STMT when we want to break
1268 before all the uses have been visited. */
1269 BREAK_FROM_IMM_USE_STMT (imm_iter);
1270 }
1271}
1272
ba372f2c 1273/* Record USE if it is defined in other bbs different than USE_STMT
1274 in the SCOP. */
30162daa 1275
1276static void
1277build_cross_bb_scalars_use (scop_p scop, tree use, gimple *use_stmt,
1278 vec<scalar_use> *reads)
1279{
30162daa 1280 if (!is_gimple_reg (use))
1281 return;
1282
1283 /* Do not gather scalar variables that can be analyzed by SCEV as they can be
1284 generated out of the induction variables. */
1285 if (scev_analyzable_p (use, scop->scop_info->region))
1286 return;
1287
1288 gimple *def_stmt = SSA_NAME_DEF_STMT (use);
74936b22 1289 if (gimple_bb (def_stmt) != gimple_bb (use_stmt))
1290 add_read (reads, use, use_stmt);
30162daa 1291}
1292
0e526381 1293/* Generates a polyhedral black box only if the bb contains interesting
1294 information. */
1295
1296static gimple_poly_bb_p
1297try_generate_gimple_bb (scop_p scop, basic_block bb)
1298{
5e6359b9 1299 vec<data_reference_p> drs = vNULL;
1300 vec<tree> writes = vNULL;
1301 vec<scalar_use> reads = vNULL;
30162daa 1302
5828c94d 1303 sese_l region = scop->scop_info->region;
44e2f332 1304 edge nest = region.entry;
0e526381 1305 loop_p loop = bb->loop_father;
1306 if (!loop_in_sese_p (loop, region))
44e2f332 1307 loop = NULL;
0e526381 1308
30162daa 1309 for (gimple_stmt_iterator gsi = gsi_start_bb (bb); !gsi_end_p (gsi);
1310 gsi_next (&gsi))
0e526381 1311 {
1312 gimple *stmt = gsi_stmt (gsi);
1313 if (is_gimple_debug (stmt))
1314 continue;
1315
1316 graphite_find_data_references_in_stmt (nest, loop, stmt, &drs);
74936b22 1317
1318 tree def = gimple_get_lhs (stmt);
1319 if (def)
1320 build_cross_bb_scalars_def (scop, def, gimple_bb (stmt), &writes);
1321
1322 ssa_op_iter iter;
1323 tree use;
1324 FOR_EACH_SSA_TREE_OPERAND (use, stmt, iter, SSA_OP_USE)
1325 build_cross_bb_scalars_use (scop, use, stmt, &reads);
0e526381 1326 }
1327
74936b22 1328 /* Handle defs and uses in PHIs. Those need special treatment given
1329 that we have to present ISL with sth that looks like we've rewritten
1330 the IL out-of-SSA. */
30162daa 1331 for (gphi_iterator psi = gsi_start_phis (bb); !gsi_end_p (psi);
1332 gsi_next (&psi))
74936b22 1333 {
1334 gphi *phi = psi.phi ();
1335 tree res = gimple_phi_result (phi);
1336 if (virtual_operand_p (res)
1337 || scev_analyzable_p (res, scop->scop_info->region))
1338 continue;
1339 /* To simulate out-of-SSA the block containing the PHI node has
1340 reads of the PHI destination. And to preserve SSA dependences
1341 we also write to it (the out-of-SSA decl and the SSA result
1342 are coalesced for dependence purposes which is good enough). */
1343 add_read (&reads, res, phi);
1344 add_write (&writes, res);
1345 }
1346 basic_block bb_for_succs = bb;
1347 if (bb_for_succs == bb_for_succs->loop_father->latch
1348 && bb_in_sese_p (bb_for_succs, scop->scop_info->region)
1349 && sese_trivially_empty_bb_p (bb_for_succs))
1350 bb_for_succs = NULL;
1351 while (bb_for_succs)
1352 {
1353 basic_block latch = NULL;
1354 edge_iterator ei;
1355 edge e;
1356 FOR_EACH_EDGE (e, ei, bb_for_succs->succs)
1357 {
1358 for (gphi_iterator psi = gsi_start_phis (e->dest); !gsi_end_p (psi);
1359 gsi_next (&psi))
1360 {
1361 gphi *phi = psi.phi ();
1362 tree res = gimple_phi_result (phi);
1363 if (virtual_operand_p (res))
1364 continue;
1365 /* To simulate out-of-SSA the predecessor of edges into PHI nodes
1366 has a copy from the PHI argument to the PHI destination. */
1367 if (! scev_analyzable_p (res, scop->scop_info->region))
1368 add_write (&writes, res);
1369 tree use = PHI_ARG_DEF_FROM_EDGE (phi, e);
1370 if (TREE_CODE (use) == SSA_NAME
1371 && ! SSA_NAME_IS_DEFAULT_DEF (use)
1372 && gimple_bb (SSA_NAME_DEF_STMT (use)) != bb_for_succs
1373 && ! scev_analyzable_p (use, scop->scop_info->region))
1374 add_read (&reads, use, phi);
1375 }
1376 if (e->dest == bb_for_succs->loop_father->latch
1377 && bb_in_sese_p (e->dest, scop->scop_info->region)
1378 && sese_trivially_empty_bb_p (e->dest))
1379 latch = e->dest;
1380 }
1381 /* Handle empty latch block PHIs here, otherwise we confuse ISL
1382 with extra conditional code where it then peels off the last
1383 iteration just because of that. It would be simplest if we
1384 just didn't force simple latches (thus remove the forwarder). */
1385 bb_for_succs = latch;
1386 }
1387
1388 /* For the region exit block add reads for all live-out vars. */
1389 if (bb == scop->scop_info->region.exit->src)
1390 {
1391 sese_build_liveouts (scop->scop_info);
1392 unsigned i;
1393 bitmap_iterator bi;
1394 EXECUTE_IF_SET_IN_BITMAP (scop->scop_info->liveout, 0, i, bi)
1395 {
1396 tree use = ssa_name (i);
1397 add_read (&reads, use, NULL);
1398 }
1399 }
30162daa 1400
1401 if (drs.is_empty () && writes.is_empty () && reads.is_empty ())
1402 return NULL;
1403
1404 return new_gimple_poly_bb (bb, drs, reads, writes);
1405}
1406
1407/* Compute alias-sets for all data references in DRS. */
1408
b0fae515 1409static bool
30162daa 1410build_alias_set (scop_p scop)
1411{
1412 int num_vertices = scop->drs.length ();
1413 struct graph *g = new_graph (num_vertices);
1414 dr_info *dr1, *dr2;
1415 int i, j;
1416 int *all_vertices;
1417
1b4e6584 1418 struct loop *nest
1419 = find_common_loop (scop->scop_info->region.entry->dest->loop_father,
1420 scop->scop_info->region.exit->src->loop_father);
1421
30162daa 1422 FOR_EACH_VEC_ELT (scop->drs, i, dr1)
1423 for (j = i+1; scop->drs.iterate (j, &dr2); j++)
1b4e6584 1424 if (dr_may_alias_p (dr1->dr, dr2->dr, nest))
30162daa 1425 {
b0fae515 1426 /* Dependences in the same alias set need to be handled
1427 by just looking at DR_ACCESS_FNs. */
28e7ffc9 1428 if (DR_NUM_DIMENSIONS (dr1->dr) == 0
1429 || DR_NUM_DIMENSIONS (dr1->dr) != DR_NUM_DIMENSIONS (dr2->dr)
b0fae515 1430 || ! operand_equal_p (DR_BASE_OBJECT (dr1->dr),
1431 DR_BASE_OBJECT (dr2->dr),
1432 OEP_ADDRESS_OF)
1433 || ! types_compatible_p (TREE_TYPE (DR_BASE_OBJECT (dr1->dr)),
1434 TREE_TYPE (DR_BASE_OBJECT (dr2->dr))))
1435 {
1436 free_graph (g);
1437 return false;
1438 }
30162daa 1439 add_edge (g, i, j);
1440 add_edge (g, j, i);
1441 }
1442
1443 all_vertices = XNEWVEC (int, num_vertices);
1444 for (i = 0; i < num_vertices; i++)
1445 all_vertices[i] = i;
1446
8affe2f6 1447 scop->max_alias_set
1448 = graphds_dfs (g, all_vertices, num_vertices, NULL, true, NULL) + 1;
30162daa 1449 free (all_vertices);
1450
1451 for (i = 0; i < g->n_vertices; i++)
1452 scop->drs[i].alias_set = g->vertices[i].component + 1;
1453
1454 free_graph (g);
b0fae515 1455 return true;
0e526381 1456}
1457
1458/* Gather BBs and conditions for a SCOP. */
1459class gather_bbs : public dom_walker
edbec012 1460{
1461public:
0fcd2c46 1462 gather_bbs (cdi_direction, scop_p, int *);
7eb20e71 1463
7b1a8d9e 1464 virtual edge before_dom_children (basic_block);
edbec012 1465 virtual void after_dom_children (basic_block);
7eb20e71 1466
edbec012 1467private:
0e526381 1468 auto_vec<gimple *, 3> conditions, cases;
1469 scop_p scop;
edbec012 1470};
1471}
0fcd2c46 1472gather_bbs::gather_bbs (cdi_direction direction, scop_p scop, int *bb_to_rpo)
c9d75619 1473 : dom_walker (direction, ALL_BLOCKS, bb_to_rpo), scop (scop)
edbec012 1474{
1475}
7eb20e71 1476
edbec012 1477/* Call-back for dom_walk executed before visiting the dominated
1478 blocks. */
7eb20e71 1479
7b1a8d9e 1480edge
0e526381 1481gather_bbs::before_dom_children (basic_block bb)
edbec012 1482{
86ee769e 1483 sese_info_p region = scop->scop_info;
1484 if (!bb_in_sese_p (bb, region->region))
0fcd2c46 1485 return dom_walker::STOP;
7eb20e71 1486
f47117d1 1487 /* For loops fully contained in the region record parameters in the
1488 loop bounds. */
1489 loop_p loop = bb->loop_father;
1490 if (loop->header == bb
1491 && loop_in_sese_p (loop, region->region))
1492 {
1493 tree nb_iters = number_of_latch_executions (loop);
1494 if (chrec_contains_symbols (nb_iters))
1495 {
d75855b4 1496 nb_iters = cached_scalar_evolution_in_region (region->region,
1497 loop, nb_iters);
f47117d1 1498 scan_tree_for_params (region, nb_iters);
1499 }
1500 }
86ee769e 1501
45105e0e 1502 if (gcond *stmt = single_pred_cond_non_loop_exit (bb))
edbec012 1503 {
1504 edge e = single_pred_edge (bb);
45105e0e 1505 /* Make sure the condition is in the region and thus was verified
1506 to be handled. */
1507 if (e != region->region.entry)
1508 {
1509 conditions.safe_push (stmt);
1510 if (e->flags & EDGE_TRUE_VALUE)
1511 cases.safe_push (stmt);
1512 else
1513 cases.safe_push (NULL);
1514 }
edbec012 1515 }
7eb20e71 1516
5828c94d 1517 scop->scop_info->bbs.safe_push (bb);
7eb20e71 1518
0e526381 1519 gimple_poly_bb_p gbb = try_generate_gimple_bb (scop, bb);
30162daa 1520 if (!gbb)
7b1a8d9e 1521 return NULL;
30162daa 1522
0e526381 1523 GBB_CONDITIONS (gbb) = conditions.copy ();
1524 GBB_CONDITION_CASES (gbb) = cases.copy ();
1525
1526 poly_bb_p pbb = new_poly_bb (scop, gbb);
1527 scop->pbbs.safe_push (pbb);
30162daa 1528
1529 int i;
1530 data_reference_p dr;
1531 FOR_EACH_VEC_ELT (gbb->data_refs, i, dr)
6fb10557 1532 {
1533 DEBUG_PRINT (dp << "Adding memory ";
1534 if (dr->is_read)
1535 dp << "read: ";
1536 else
1537 dp << "write: ";
1ffa4346 1538 print_generic_expr (dump_file, dr->ref);
6fb10557 1539 dp << "\nFrom stmt: ";
1ffa4346 1540 print_gimple_stmt (dump_file, dr->stmt, 0));
6fb10557 1541
1542 scop->drs.safe_push (dr_info (dr, pbb));
1543 }
7b1a8d9e 1544
1545 return NULL;
edbec012 1546}
7eb20e71 1547
edbec012 1548/* Call-back for dom_walk executed after visiting the dominated
1549 blocks. */
7eb20e71 1550
edbec012 1551void
0e526381 1552gather_bbs::after_dom_children (basic_block bb)
edbec012 1553{
5828c94d 1554 if (!bb_in_sese_p (bb, scop->scop_info->region))
edbec012 1555 return;
7eb20e71 1556
edbec012 1557 if (single_pred_cond_non_loop_exit (bb))
1558 {
45105e0e 1559 edge e = single_pred_edge (bb);
1560 if (e != scop->scop_info->region.entry)
1561 {
1562 conditions.pop ();
1563 cases.pop ();
1564 }
edbec012 1565 }
1566}
7eb20e71 1567
f857d1b7 1568
1569/* Compute sth like an execution order, dominator order with first executing
1570 edges that stay inside the current loop, delaying processing exit edges. */
1571
56adbb23 1572static int *bb_to_rpo;
f857d1b7 1573
1574/* Helper for qsort, sorting after order above. */
1575
1576static int
1577cmp_pbbs (const void *pa, const void *pb)
1578{
1579 poly_bb_p bb1 = *((const poly_bb_p *)pa);
1580 poly_bb_p bb2 = *((const poly_bb_p *)pb);
56adbb23 1581 if (bb_to_rpo[bb1->black_box->bb->index]
1582 < bb_to_rpo[bb2->black_box->bb->index])
f857d1b7 1583 return -1;
56adbb23 1584 else if (bb_to_rpo[bb1->black_box->bb->index]
1585 > bb_to_rpo[bb2->black_box->bb->index])
f857d1b7 1586 return 1;
1587 else
1588 return 0;
1589}
1590
7eb20e71 1591/* Find Static Control Parts (SCoP) in the current function and pushes
1592 them to SCOPS. */
1593
1594void
1595build_scops (vec<scop_p> *scops)
1596{
1597 if (dump_file)
1598 dp.set_dump_file (dump_file);
1599
edbec012 1600 scop_detection sb;
cb44892e 1601 sb.build_scop_depth (current_loops->tree_root);
16cbd7c3 1602
1603 /* Now create scops from the lightweight SESEs. */
1604 vec<sese_l> scops_l = sb.get_scops ();
0fcd2c46 1605
1606 /* Domwalk needs a bb to RPO mapping. Compute it once here. */
1607 int *postorder = XNEWVEC (int, n_basic_blocks_for_fn (cfun));
1608 int postorder_num = pre_and_rev_post_order_compute (NULL, postorder, true);
56adbb23 1609 bb_to_rpo = XNEWVEC (int, last_basic_block_for_fn (cfun));
0fcd2c46 1610 for (int i = 0; i < postorder_num; ++i)
1611 bb_to_rpo[postorder[i]] = i;
1612 free (postorder);
1613
16cbd7c3 1614 int i;
5828c94d 1615 sese_l *s;
16cbd7c3 1616 FOR_EACH_VEC_ELT (scops_l, i, s)
edbec012 1617 {
5828c94d 1618 scop_p scop = new_scop (s->entry, s->exit);
edbec012 1619
0e526381 1620 /* Record all basic blocks and their conditions in REGION. */
0fcd2c46 1621 gather_bbs (CDI_DOMINATORS, scop, bb_to_rpo).walk (s->entry->dest);
f857d1b7 1622
56adbb23 1623 /* Sort pbbs after execution order for initial schedule generation. */
f857d1b7 1624 scop->pbbs.qsort (cmp_pbbs);
0e526381 1625
b0fae515 1626 if (! build_alias_set (scop))
1627 {
1628 DEBUG_PRINT (dp << "[scop-detection-fail] cannot handle dependences\n");
1629 free_scop (scop);
1630 continue;
1631 }
30162daa 1632
edbec012 1633 /* Do not optimize a scop containing only PBBs that do not belong
1634 to any loops. */
1635 if (sb.nb_pbbs_in_loops (scop) == 0)
1636 {
118a202b 1637 DEBUG_PRINT (dp << "[scop-detection-fail] no data references.\n");
1638 free_scop (scop);
1639 continue;
1640 }
1641
84e96705 1642 unsigned max_arrays = PARAM_VALUE (PARAM_GRAPHITE_MAX_ARRAYS_PER_SCOP);
8affe2f6 1643 if (max_arrays > 0
1644 && scop->drs.length () >= max_arrays)
84e96705 1645 {
1646 DEBUG_PRINT (dp << "[scop-detection-fail] too many data references: "
1647 << scop->drs.length ()
1648 << " is larger than --param graphite-max-arrays-per-scop="
1649 << max_arrays << ".\n");
1650 free_scop (scop);
1651 continue;
1652 }
1653
118a202b 1654 find_scop_parameters (scop);
1655 graphite_dim_t max_dim = PARAM_VALUE (PARAM_GRAPHITE_MAX_NB_SCOP_PARAMS);
0fcd2c46 1656 if (max_dim > 0
1657 && scop_nb_params (scop) > max_dim)
118a202b 1658 {
1659 DEBUG_PRINT (dp << "[scop-detection-fail] too many parameters: "
4caef4a5 1660 << scop_nb_params (scop)
1661 << " larger than --param graphite-max-nb-scop-params="
1662 << max_dim << ".\n");
edbec012 1663 free_scop (scop);
1664 continue;
1665 }
1666
1667 scops->safe_push (scop);
1668 }
1669
0fcd2c46 1670 free (bb_to_rpo);
56adbb23 1671 bb_to_rpo = NULL;
7eb20e71 1672 DEBUG_PRINT (dp << "number of SCoPs: " << (scops ? scops->length () : 0););
1673}
1674
edbec012 1675#endif /* HAVE_isl */