[ARM] Fix assembly comment syntax in -mprint-tune-info
[official-gcc.git] / gcc / graphite-scop-detection.c
blobc372141addf5968902a5297a6acd42c3c678ac92
1 /* Detection of Static Control Parts (SCoP) for Graphite.
2 Copyright (C) 2009-2017 Free Software Foundation, Inc.
3 Contributed by Sebastian Pop <sebastian.pop@amd.com> and
4 Tobias Grosser <grosser@fim.uni-passau.de>.
6 This file is part of GCC.
8 GCC is free software; you can redistribute it and/or modify
9 it under the terms of the GNU General Public License as published by
10 the Free Software Foundation; either version 3, or (at your option)
11 any later version.
13 GCC is distributed in the hope that it will be useful,
14 but WITHOUT ANY WARRANTY; without even the implied warranty of
15 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
16 GNU General Public License for more details.
18 You should have received a copy of the GNU General Public License
19 along with GCC; see the file COPYING3. If not see
20 <http://www.gnu.org/licenses/>. */
22 #define USES_ISL
24 #include "config.h"
26 #ifdef HAVE_isl
28 #include "system.h"
29 #include "coretypes.h"
30 #include "backend.h"
31 #include "cfghooks.h"
32 #include "domwalk.h"
33 #include "params.h"
34 #include "tree.h"
35 #include "gimple.h"
36 #include "ssa.h"
37 #include "fold-const.h"
38 #include "gimple-iterator.h"
39 #include "tree-cfg.h"
40 #include "tree-ssa-loop-manip.h"
41 #include "tree-ssa-loop-niter.h"
42 #include "tree-ssa-loop.h"
43 #include "tree-into-ssa.h"
44 #include "tree-ssa.h"
45 #include "cfgloop.h"
46 #include "tree-data-ref.h"
47 #include "tree-scalar-evolution.h"
48 #include "tree-pass.h"
49 #include "tree-ssa-propagate.h"
50 #include "gimple-pretty-print.h"
51 #include "graphite.h"
53 class debug_printer
55 private:
56 FILE *dump_file;
58 public:
59 void
60 set_dump_file (FILE *f)
62 gcc_assert (f);
63 dump_file = f;
66 friend debug_printer &
67 operator<< (debug_printer &output, int i)
69 fprintf (output.dump_file, "%d", i);
70 return output;
72 friend debug_printer &
73 operator<< (debug_printer &output, const char *s)
75 fprintf (output.dump_file, "%s", s);
76 return output;
78 } dp;
80 #define DEBUG_PRINT(args) do \
81 { \
82 if (dump_file && (dump_flags & TDF_DETAILS)) { args; } \
83 } while (0);
85 /* Pretty print to FILE all the SCoPs in DOT format and mark them with
86 different colors. If there are not enough colors, paint the
87 remaining SCoPs in gray.
89 Special nodes:
90 - "*" after the node number denotes the entry of a SCoP,
91 - "#" after the node number denotes the exit of a SCoP,
92 - "()" around the node number denotes the entry or the
93 exit nodes of the SCOP. These are not part of SCoP. */
95 DEBUG_FUNCTION void
96 dot_all_sese (FILE *file, vec<sese_l>& scops)
98 /* Disable debugging while printing graph. */
99 int tmp_dump_flags = dump_flags;
100 dump_flags = 0;
102 fprintf (file, "digraph all {\n");
104 basic_block bb;
105 FOR_ALL_BB_FN (bb, cfun)
107 int part_of_scop = false;
109 /* Use HTML for every bb label. So we are able to print bbs
110 which are part of two different SCoPs, with two different
111 background colors. */
112 fprintf (file, "%d [label=<\n <TABLE BORDER=\"0\" CELLBORDER=\"1\" ",
113 bb->index);
114 fprintf (file, "CELLSPACING=\"0\">\n");
116 /* Select color for SCoP. */
117 sese_l *region;
118 int i;
119 FOR_EACH_VEC_ELT (scops, i, region)
121 bool sese_in_region = bb_in_sese_p (bb, *region);
122 if (sese_in_region || (region->exit->dest == bb)
123 || (region->entry->dest == bb))
125 const char *color;
126 switch (i % 17)
128 case 0: /* red */
129 color = "#e41a1c";
130 break;
131 case 1: /* blue */
132 color = "#377eb8";
133 break;
134 case 2: /* green */
135 color = "#4daf4a";
136 break;
137 case 3: /* purple */
138 color = "#984ea3";
139 break;
140 case 4: /* orange */
141 color = "#ff7f00";
142 break;
143 case 5: /* yellow */
144 color = "#ffff33";
145 break;
146 case 6: /* brown */
147 color = "#a65628";
148 break;
149 case 7: /* rose */
150 color = "#f781bf";
151 break;
152 case 8:
153 color = "#8dd3c7";
154 break;
155 case 9:
156 color = "#ffffb3";
157 break;
158 case 10:
159 color = "#bebada";
160 break;
161 case 11:
162 color = "#fb8072";
163 break;
164 case 12:
165 color = "#80b1d3";
166 break;
167 case 13:
168 color = "#fdb462";
169 break;
170 case 14:
171 color = "#b3de69";
172 break;
173 case 15:
174 color = "#fccde5";
175 break;
176 case 16:
177 color = "#bc80bd";
178 break;
179 default: /* gray */
180 color = "#999999";
183 fprintf (file, " <TR><TD WIDTH=\"50\" BGCOLOR=\"%s\">",
184 color);
186 if (!sese_in_region)
187 fprintf (file, " (");
189 if (bb == region->entry->dest && bb == region->exit->dest)
190 fprintf (file, " %d*# ", bb->index);
191 else if (bb == region->entry->dest)
192 fprintf (file, " %d* ", bb->index);
193 else if (bb == region->exit->dest)
194 fprintf (file, " %d# ", bb->index);
195 else
196 fprintf (file, " %d ", bb->index);
198 fprintf (file, "{lp_%d}", bb->loop_father->num);
200 if (!sese_in_region)
201 fprintf (file, ")");
203 fprintf (file, "</TD></TR>\n");
204 part_of_scop = true;
208 if (!part_of_scop)
210 fprintf (file, " <TR><TD WIDTH=\"50\" BGCOLOR=\"#ffffff\">");
211 fprintf (file, " %d {lp_%d} </TD></TR>\n", bb->index,
212 bb->loop_father->num);
214 fprintf (file, " </TABLE>>, shape=box, style=\"setlinewidth(0)\"]\n");
217 FOR_ALL_BB_FN (bb, cfun)
219 edge e;
220 edge_iterator ei;
221 FOR_EACH_EDGE (e, ei, bb->succs)
222 fprintf (file, "%d -> %d;\n", bb->index, e->dest->index);
225 fputs ("}\n\n", file);
227 /* Enable debugging again. */
228 dump_flags = tmp_dump_flags;
231 /* Display SCoP on stderr. */
233 DEBUG_FUNCTION void
234 dot_sese (sese_l& scop)
236 vec<sese_l> scops;
237 scops.create (1);
239 if (scop)
240 scops.safe_push (scop);
242 dot_all_sese (stderr, scops);
244 scops.release ();
247 DEBUG_FUNCTION void
248 dot_cfg ()
250 vec<sese_l> scops;
251 scops.create (1);
252 dot_all_sese (stderr, scops);
253 scops.release ();
256 /* Return true if BB is empty, contains only DEBUG_INSNs. */
258 static bool
259 trivially_empty_bb_p (basic_block bb)
261 gimple_stmt_iterator gsi;
263 for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi))
264 if (gimple_code (gsi_stmt (gsi)) != GIMPLE_DEBUG)
265 return false;
267 return true;
270 /* Returns true when P1 and P2 are close phis with the same
271 argument. */
273 static inline bool
274 same_close_phi_node (gphi *p1, gphi *p2)
276 return (types_compatible_p (TREE_TYPE (gimple_phi_result (p1)),
277 TREE_TYPE (gimple_phi_result (p2)))
278 && operand_equal_p (gimple_phi_arg_def (p1, 0),
279 gimple_phi_arg_def (p2, 0), 0));
282 static void make_close_phi_nodes_unique (basic_block bb);
284 /* Remove the close phi node at GSI and replace its rhs with the rhs
285 of PHI. */
287 static void
288 remove_duplicate_close_phi (gphi *phi, gphi_iterator *gsi)
290 gimple *use_stmt;
291 use_operand_p use_p;
292 imm_use_iterator imm_iter;
293 tree res = gimple_phi_result (phi);
294 tree def = gimple_phi_result (gsi->phi ());
296 gcc_assert (same_close_phi_node (phi, gsi->phi ()));
298 FOR_EACH_IMM_USE_STMT (use_stmt, imm_iter, def)
300 FOR_EACH_IMM_USE_ON_STMT (use_p, imm_iter)
301 SET_USE (use_p, res);
303 update_stmt (use_stmt);
305 /* It is possible that we just created a duplicate close-phi
306 for an already-processed containing loop. Check for this
307 case and clean it up. */
308 if (gimple_code (use_stmt) == GIMPLE_PHI
309 && gimple_phi_num_args (use_stmt) == 1)
310 make_close_phi_nodes_unique (gimple_bb (use_stmt));
313 remove_phi_node (gsi, true);
316 /* Removes all the close phi duplicates from BB. */
318 static void
319 make_close_phi_nodes_unique (basic_block bb)
321 gphi_iterator psi;
323 for (psi = gsi_start_phis (bb); !gsi_end_p (psi); gsi_next (&psi))
325 gphi_iterator gsi = psi;
326 gphi *phi = psi.phi ();
328 /* At this point, PHI should be a close phi in normal form. */
329 gcc_assert (gimple_phi_num_args (phi) == 1);
331 /* Iterate over the next phis and remove duplicates. */
332 gsi_next (&gsi);
333 while (!gsi_end_p (gsi))
334 if (same_close_phi_node (phi, gsi.phi ()))
335 remove_duplicate_close_phi (phi, &gsi);
336 else
337 gsi_next (&gsi);
341 /* Return true when NAME is defined in LOOP. */
343 static bool
344 defined_in_loop_p (tree name, loop_p loop)
346 gcc_assert (TREE_CODE (name) == SSA_NAME);
347 return loop == loop_containing_stmt (SSA_NAME_DEF_STMT (name));
350 /* Transforms LOOP to the canonical loop closed SSA form. */
352 static void
353 canonicalize_loop_closed_ssa (loop_p loop)
355 edge e = single_exit (loop);
356 basic_block bb;
358 if (!e || e->flags & EDGE_ABNORMAL)
359 return;
361 bb = e->dest;
363 if (single_pred_p (bb))
365 e = split_block_after_labels (bb);
366 DEBUG_PRINT (dp << "Splitting bb_" << bb->index << ".\n");
367 make_close_phi_nodes_unique (e->src);
369 else
371 gphi_iterator psi;
372 basic_block close = split_edge (e);
374 e = single_succ_edge (close);
375 DEBUG_PRINT (dp << "Splitting edge (" << e->src->index << ","
376 << e->dest->index << ")\n");
378 for (psi = gsi_start_phis (bb); !gsi_end_p (psi); gsi_next (&psi))
380 gphi *phi = psi.phi ();
381 unsigned i;
383 for (i = 0; i < gimple_phi_num_args (phi); i++)
384 if (gimple_phi_arg_edge (phi, i) == e)
386 tree res, arg = gimple_phi_arg_def (phi, i);
387 use_operand_p use_p;
388 gphi *close_phi;
390 /* Only add close phi nodes for SSA_NAMEs defined in LOOP. */
391 if (TREE_CODE (arg) != SSA_NAME
392 || !defined_in_loop_p (arg, loop))
393 continue;
395 close_phi = create_phi_node (NULL_TREE, close);
396 res = create_new_def_for (arg, close_phi,
397 gimple_phi_result_ptr (close_phi));
398 add_phi_arg (close_phi, arg,
399 gimple_phi_arg_edge (close_phi, 0),
400 UNKNOWN_LOCATION);
401 use_p = gimple_phi_arg_imm_use_ptr (phi, i);
402 replace_exp (use_p, res);
403 update_stmt (phi);
407 make_close_phi_nodes_unique (close);
410 /* The code above does not properly handle changes in the post dominance
411 information (yet). */
412 recompute_all_dominators ();
415 /* Converts the current loop closed SSA form to a canonical form
416 expected by the Graphite code generation.
418 The loop closed SSA form has the following invariant: a variable
419 defined in a loop that is used outside the loop appears only in the
420 phi nodes in the destination of the loop exit. These phi nodes are
421 called close phi nodes.
423 The canonical loop closed SSA form contains the extra invariants:
425 - when the loop contains only one exit, the close phi nodes contain
426 only one argument. That implies that the basic block that contains
427 the close phi nodes has only one predecessor, that is a basic block
428 in the loop.
430 - the basic block containing the close phi nodes does not contain
431 other statements.
433 - there exist only one phi node per definition in the loop.
436 static void
437 canonicalize_loop_closed_ssa_form (void)
439 checking_verify_loop_closed_ssa (true);
441 loop_p loop;
442 FOR_EACH_LOOP (loop, 0)
443 canonicalize_loop_closed_ssa (loop);
445 rewrite_into_loop_closed_ssa (NULL, TODO_update_ssa);
446 update_ssa (TODO_update_ssa);
448 checking_verify_loop_closed_ssa (true);
451 /* Can all ivs be represented by a signed integer?
452 As isl might generate negative values in its expressions, signed loop ivs
453 are required in the backend. */
455 static bool
456 loop_ivs_can_be_represented (loop_p loop)
458 unsigned type_long_long = TYPE_PRECISION (long_long_integer_type_node);
459 for (gphi_iterator psi = gsi_start_phis (loop->header); !gsi_end_p (psi);
460 gsi_next (&psi))
462 gphi *phi = psi.phi ();
463 tree res = PHI_RESULT (phi);
464 tree type = TREE_TYPE (res);
466 if (TYPE_UNSIGNED (type) && TYPE_PRECISION (type) >= type_long_long)
467 return false;
470 return true;
473 /* Returns a COND_EXPR statement when BB has a single predecessor, the
474 edge between BB and its predecessor is not a loop exit edge, and
475 the last statement of the single predecessor is a COND_EXPR. */
477 static gcond *
478 single_pred_cond_non_loop_exit (basic_block bb)
480 if (single_pred_p (bb))
482 edge e = single_pred_edge (bb);
483 basic_block pred = e->src;
484 gimple *stmt;
486 if (loop_depth (pred->loop_father) > loop_depth (bb->loop_father))
487 return NULL;
489 stmt = last_stmt (pred);
491 if (stmt && gimple_code (stmt) == GIMPLE_COND)
492 return as_a<gcond *> (stmt);
495 return NULL;
498 namespace
501 /* Build the maximal scop containing LOOPs and add it to SCOPS. */
503 class scop_detection
505 public:
506 scop_detection () : scops (vNULL) {}
508 ~scop_detection ()
510 scops.release ();
513 /* A marker for invalid sese_l. */
514 static sese_l invalid_sese;
516 /* Return the SCOPS in this SCOP_DETECTION. */
518 vec<sese_l>
519 get_scops ()
521 return scops;
524 /* Return an sese_l around the LOOP. */
526 sese_l get_sese (loop_p loop);
528 /* Return the closest dominator with a single entry edge. In case of a
529 back-loop the back-edge is not counted. */
531 static edge get_nearest_dom_with_single_entry (basic_block dom);
533 /* Return the closest post-dominator with a single exit edge. In case of a
534 back-loop the back-edge is not counted. */
536 static edge get_nearest_pdom_with_single_exit (basic_block dom);
538 /* Merge scops at same loop depth and returns the new sese.
539 Returns a new SESE when merge was successful, INVALID_SESE otherwise. */
541 sese_l merge_sese (sese_l first, sese_l second) const;
543 /* Build scop outer->inner if possible. */
545 sese_l build_scop_depth (sese_l s, loop_p loop);
547 /* If loop and loop->next are valid scops, try to merge them. */
549 sese_l build_scop_breadth (sese_l s1, loop_p loop);
551 /* Return true when LOOP is a valid scop, that is a Static Control Part, a
552 region of code that can be represented in the polyhedral model. SCOP
553 defines the region we analyse. */
555 bool loop_is_valid_in_scop (loop_p loop, sese_l scop) const;
557 /* Return true when BEGIN is the preheader edge of a loop with a single exit
558 END. */
560 static bool region_has_one_loop (sese_l s);
562 /* Add to SCOPS a scop starting at SCOP_BEGIN and ending at SCOP_END. */
564 void add_scop (sese_l s);
566 /* Returns true if S1 subsumes/surrounds S2. */
567 static bool subsumes (sese_l s1, sese_l s2);
569 /* Remove a SCoP which is subsumed by S1. */
570 void remove_subscops (sese_l s1);
572 /* Returns true if S1 intersects with S2. Since we already know that S1 does
573 not subsume S2 or vice-versa, we only check for entry bbs. */
575 static bool intersects (sese_l s1, sese_l s2);
577 /* Remove one of the scops when it intersects with any other. */
579 void remove_intersecting_scops (sese_l s1);
581 /* Return true when the body of LOOP has statements that can be represented
582 as a valid scop. */
584 bool loop_body_is_valid_scop (loop_p loop, sese_l scop) const;
586 /* Return true when BB contains a harmful operation for a scop: that
587 can be a function call with side effects, the induction variables
588 are not linear with respect to SCOP, etc. The current open
589 scop should end before this statement. */
591 bool harmful_stmt_in_bb (sese_l scop, basic_block bb) const;
593 /* Return true when a statement in SCOP cannot be represented by Graphite.
594 The assumptions are that L1 dominates L2, and SCOP->entry dominates L1.
595 Limit the number of bbs between adjacent loops to
596 PARAM_SCOP_MAX_NUM_BBS_BETWEEN_LOOPS. */
598 bool harmful_loop_in_region (sese_l scop) const;
600 /* Return true only when STMT is simple enough for being handled by Graphite.
601 This depends on SCOP, as the parameters are initialized relatively to
602 this basic block, the linear functions are initialized based on the
603 outermost loop containing STMT inside the SCOP. BB is the place where we
604 try to evaluate the STMT. */
606 bool stmt_simple_for_scop_p (sese_l scop, gimple *stmt,
607 basic_block bb) const;
609 /* Something like "n * m" is not allowed. */
611 static bool graphite_can_represent_init (tree e);
613 /* Return true when SCEV can be represented in the polyhedral model.
615 An expression can be represented, if it can be expressed as an
616 affine expression. For loops (i, j) and parameters (m, n) all
617 affine expressions are of the form:
619 x1 * i + x2 * j + x3 * m + x4 * n + x5 * 1 where x1..x5 element of Z
621 1 i + 20 j + (-2) m + 25
623 Something like "i * n" or "n * m" is not allowed. */
625 static bool graphite_can_represent_scev (tree scev);
627 /* Return true when EXPR can be represented in the polyhedral model.
629 This means an expression can be represented, if it is linear with respect
630 to the loops and the strides are non parametric. LOOP is the place where
631 the expr will be evaluated. SCOP defines the region we analyse. */
633 static bool graphite_can_represent_expr (sese_l scop, loop_p loop,
634 tree expr);
636 /* Return true if the data references of STMT can be represented by Graphite.
637 We try to analyze the data references in a loop contained in the SCOP. */
639 static bool stmt_has_simple_data_refs_p (sese_l scop, gimple *stmt);
641 /* Remove the close phi node at GSI and replace its rhs with the rhs
642 of PHI. */
644 static void remove_duplicate_close_phi (gphi *phi, gphi_iterator *gsi);
646 /* Returns true when Graphite can represent LOOP in SCOP.
647 FIXME: For the moment, graphite cannot be used on loops that iterate using
648 induction variables that wrap. */
650 static bool can_represent_loop_1 (loop_p loop, sese_l scop);
652 /* Return true when all the loops within LOOP can be represented by
653 Graphite. */
655 static bool can_represent_loop (loop_p loop, sese_l scop);
657 /* Returns the number of pbbs that are in loops contained in SCOP. */
659 static int nb_pbbs_in_loops (scop_p scop);
661 static bool graphite_can_represent_stmt (sese_l, gimple *, basic_block);
663 private:
664 vec<sese_l> scops;
667 sese_l scop_detection::invalid_sese (NULL, NULL);
669 /* Return an sese_l around the LOOP. */
671 sese_l
672 scop_detection::get_sese (loop_p loop)
674 if (!loop)
675 return invalid_sese;
677 if (!loops_state_satisfies_p (LOOPS_HAVE_PREHEADERS))
678 return invalid_sese;
679 edge scop_end = single_exit (loop);
680 if (!scop_end)
681 return invalid_sese;
682 edge scop_begin = loop_preheader_edge (loop);
683 sese_l s (scop_begin, scop_end);
684 return s;
687 /* Return the closest dominator with a single entry edge. */
689 edge
690 scop_detection::get_nearest_dom_with_single_entry (basic_block dom)
692 if (!dom->preds)
693 return NULL;
695 /* If any of the dominators has two predecessors but one of them is a back
696 edge, then that basic block also qualifies as a dominator with single
697 entry. */
698 if (dom->preds->length () == 2)
700 /* If e1->src dominates e2->src then e1->src will also dominate dom. */
701 edge e1 = (*dom->preds)[0];
702 edge e2 = (*dom->preds)[1];
703 loop_p l = dom->loop_father;
704 loop_p l1 = e1->src->loop_father;
705 loop_p l2 = e2->src->loop_father;
706 if (l != l1 && l == l2
707 && dominated_by_p (CDI_DOMINATORS, e2->src, e1->src))
708 return e1;
709 if (l != l2 && l == l1
710 && dominated_by_p (CDI_DOMINATORS, e1->src, e2->src))
711 return e2;
714 while (dom->preds->length () != 1)
716 if (dom->preds->length () < 1)
717 return NULL;
718 dom = get_immediate_dominator (CDI_DOMINATORS, dom);
719 if (!dom->preds)
720 return NULL;
722 return (*dom->preds)[0];
725 /* Return the closest post-dominator with a single exit edge. In case of a
726 back-loop the back-edge is not counted. */
728 edge
729 scop_detection::get_nearest_pdom_with_single_exit (basic_block pdom)
731 if (!pdom->succs)
732 return NULL;
734 /* If any of the post-dominators has two successors but one of them is a back
735 edge, then that basic block also qualifies as a post-dominator with single
736 exit. */
737 if (pdom->succs->length () == 2)
739 /* If e1->dest post-dominates e2->dest then e1->dest will also
740 post-dominate pdom. */
741 edge e1 = (*pdom->succs)[0];
742 edge e2 = (*pdom->succs)[1];
743 loop_p l = pdom->loop_father;
744 loop_p l1 = e1->dest->loop_father;
745 loop_p l2 = e2->dest->loop_father;
746 if (l != l1 && l == l2
747 && dominated_by_p (CDI_POST_DOMINATORS, e2->dest, e1->dest))
748 return e1;
749 if (l != l2 && l == l1
750 && dominated_by_p (CDI_POST_DOMINATORS, e1->dest, e2->dest))
751 return e2;
754 while (pdom->succs->length () != 1)
756 if (pdom->succs->length () < 1)
757 return NULL;
758 pdom = get_immediate_dominator (CDI_POST_DOMINATORS, pdom);
759 if (!pdom->succs)
760 return NULL;
763 return (*pdom->succs)[0];
766 /* Merge scops at same loop depth and returns the new sese.
767 Returns a new SESE when merge was successful, INVALID_SESE otherwise. */
769 sese_l
770 scop_detection::merge_sese (sese_l first, sese_l second) const
772 /* In the trivial case first/second may be NULL. */
773 if (!first)
774 return second;
775 if (!second)
776 return first;
778 DEBUG_PRINT (dp << "[scop-detection] try merging sese s1: ";
779 print_sese (dump_file, first);
780 dp << "[scop-detection] try merging sese s2: ";
781 print_sese (dump_file, second));
783 /* Assumption: Both the sese's should be at the same loop depth or one scop
784 should subsume the other like in case of nested loops. */
786 /* Find the common dominators for entry,
787 and common post-dominators for the exit. */
788 basic_block dom = nearest_common_dominator (CDI_DOMINATORS,
789 get_entry_bb (first),
790 get_entry_bb (second));
792 edge entry = get_nearest_dom_with_single_entry (dom);
794 if (!entry || (entry->flags & EDGE_IRREDUCIBLE_LOOP))
795 return invalid_sese;
797 basic_block pdom = nearest_common_dominator (CDI_POST_DOMINATORS,
798 get_exit_bb (first),
799 get_exit_bb (second));
800 pdom = nearest_common_dominator (CDI_POST_DOMINATORS, dom, pdom);
802 edge exit = get_nearest_pdom_with_single_exit (pdom);
804 if (!exit || (exit->flags & EDGE_IRREDUCIBLE_LOOP))
805 return invalid_sese;
807 sese_l combined (entry, exit);
809 DEBUG_PRINT (dp << "[scop-detection] checking combined sese: ";
810 print_sese (dump_file, combined));
812 /* FIXME: We could iterate to find the dom which dominates pdom, and pdom
813 which post-dominates dom, until it stabilizes. Also, ENTRY->SRC and
814 EXIT->DEST should be in the same loop nest. */
815 if (!dominated_by_p (CDI_DOMINATORS, pdom, dom)
816 || loop_depth (entry->src->loop_father)
817 != loop_depth (exit->dest->loop_father))
818 return invalid_sese;
820 /* For now we just want to bail out when exit does not post-dominate entry.
821 TODO: We might just add a basic_block at the exit to make exit
822 post-dominate entry (the entire region). */
823 if (!dominated_by_p (CDI_POST_DOMINATORS, get_entry_bb (combined),
824 get_exit_bb (combined))
825 || !dominated_by_p (CDI_DOMINATORS, get_exit_bb (combined),
826 get_entry_bb (combined)))
828 DEBUG_PRINT (dp << "[scop-detection-fail] cannot merge seses.\n");
829 return invalid_sese;
832 /* FIXME: We should remove this piece of code once
833 canonicalize_loop_closed_ssa has been removed, because that function
834 adds a BB with single exit. */
835 if (!trivially_empty_bb_p (get_exit_bb (combined)))
837 /* Find the first empty succ (with single exit) of combined.exit. */
838 basic_block imm_succ = combined.exit->dest;
839 if (single_succ_p (imm_succ)
840 && single_pred_p (imm_succ)
841 && trivially_empty_bb_p (imm_succ))
842 combined.exit = single_succ_edge (imm_succ);
843 else
845 DEBUG_PRINT (dp << "[scop-detection-fail] Discarding SCoP because "
846 << "no single exit (empty succ) for sese exit";
847 print_sese (dump_file, combined));
848 return invalid_sese;
852 /* Analyze all the BBs in new sese. */
853 if (harmful_loop_in_region (combined))
854 return invalid_sese;
856 DEBUG_PRINT (dp << "[merged-sese] s1: "; print_sese (dump_file, combined));
858 return combined;
861 /* Build scop outer->inner if possible. */
863 sese_l
864 scop_detection::build_scop_depth (sese_l s, loop_p loop)
866 if (!loop)
867 return s;
869 DEBUG_PRINT (dp << "[Depth loop_" << loop->num << "]\n");
870 s = build_scop_depth (s, loop->inner);
872 sese_l s2 = merge_sese (s, get_sese (loop));
873 if (!s2)
875 /* s might be a valid scop, so return it and start analyzing from the
876 adjacent loop. */
877 build_scop_depth (invalid_sese, loop->next);
878 return s;
881 if (!loop_is_valid_in_scop (loop, s2))
882 return build_scop_depth (invalid_sese, loop->next);
884 return build_scop_breadth (s2, loop);
887 /* If loop and loop->next are valid scops, try to merge them. */
889 sese_l
890 scop_detection::build_scop_breadth (sese_l s1, loop_p loop)
892 if (!loop)
893 return s1;
894 DEBUG_PRINT (dp << "[Breadth loop_" << loop->num << "]\n");
895 gcc_assert (s1);
897 loop_p l = loop;
898 sese_l s2 = build_scop_depth (invalid_sese, l->next);
899 if (!s2)
901 if (s1)
902 add_scop (s1);
903 return s1;
906 sese_l combined = merge_sese (s1, s2);
908 /* Combining adjacent loops may add unrelated loops into the
909 region so we have to check all sub-loops of the outer loop
910 that are in the combined region. */
911 if (combined)
912 for (l = loop_outer (loop)->inner; l; l = l->next)
913 if (bb_in_sese_p (l->header, combined)
914 && ! loop_is_valid_in_scop (l, combined))
916 combined = invalid_sese;
917 break;
920 if (combined)
921 s1 = combined;
922 else
923 add_scop (s2);
925 if (s1)
926 add_scop (s1);
927 return s1;
930 /* Returns true when Graphite can represent LOOP in SCOP.
931 FIXME: For the moment, graphite cannot be used on loops that iterate using
932 induction variables that wrap. */
934 bool
935 scop_detection::can_represent_loop_1 (loop_p loop, sese_l scop)
937 tree niter;
938 struct tree_niter_desc niter_desc;
940 return single_exit (loop)
941 && !(loop_preheader_edge (loop)->flags & EDGE_IRREDUCIBLE_LOOP)
942 && number_of_iterations_exit (loop, single_exit (loop), &niter_desc, false)
943 && niter_desc.control.no_overflow
944 && (niter = number_of_latch_executions (loop))
945 && !chrec_contains_undetermined (niter)
946 && !chrec_contains_undetermined (scalar_evolution_in_region (scop,
947 loop, niter))
948 && graphite_can_represent_expr (scop, loop, niter);
951 /* Return true when all the loops within LOOP can be represented by
952 Graphite. */
954 bool
955 scop_detection::can_represent_loop (loop_p loop, sese_l scop)
957 if (!can_represent_loop_1 (loop, scop))
958 return false;
959 if (loop->inner && !can_represent_loop (loop->inner, scop))
960 return false;
961 if (loop->next && !can_represent_loop (loop->next, scop))
962 return false;
964 return true;
967 /* Return true when LOOP is a valid scop, that is a Static Control Part, a
968 region of code that can be represented in the polyhedral model. SCOP
969 defines the region we analyse. */
971 bool
972 scop_detection::loop_is_valid_in_scop (loop_p loop, sese_l scop) const
974 if (!scop)
975 return false;
977 if (!optimize_loop_nest_for_speed_p (loop))
979 DEBUG_PRINT (dp << "[scop-detection-fail] loop_"
980 << loop->num << " is not on a hot path.\n");
981 return false;
984 if (!can_represent_loop (loop, scop))
986 DEBUG_PRINT (dp << "[scop-detection-fail] cannot represent loop_"
987 << loop->num << "\n");
988 return false;
991 if (loop_body_is_valid_scop (loop, scop))
993 DEBUG_PRINT (dp << "[valid-scop] loop_" << loop->num
994 << " is a valid scop.\n");
995 return true;
997 return false;
1000 /* Return true when BEGIN is the preheader edge of a loop with a single exit
1001 END. */
1003 bool
1004 scop_detection::region_has_one_loop (sese_l s)
1006 edge begin = s.entry;
1007 edge end = s.exit;
1008 /* Check for a single perfectly nested loop. */
1009 if (begin->dest->loop_father->inner)
1010 return false;
1012 /* Otherwise, check whether we have adjacent loops. */
1013 return begin->dest->loop_father == end->src->loop_father;
1016 /* Add to SCOPS a scop starting at SCOP_BEGIN and ending at SCOP_END. */
1018 void
1019 scop_detection::add_scop (sese_l s)
1021 gcc_assert (s);
1023 /* Do not add scops with only one loop. */
1024 if (region_has_one_loop (s))
1026 DEBUG_PRINT (dp << "[scop-detection-fail] Discarding one loop SCoP: ";
1027 print_sese (dump_file, s));
1028 return;
1031 if (get_exit_bb (s) == EXIT_BLOCK_PTR_FOR_FN (cfun))
1033 DEBUG_PRINT (dp << "[scop-detection-fail] "
1034 << "Discarding SCoP exiting to return: ";
1035 print_sese (dump_file, s));
1036 return;
1039 /* Remove all the scops which are subsumed by s. */
1040 remove_subscops (s);
1042 /* Remove intersecting scops. FIXME: It will be a good idea to keep
1043 the non-intersecting part of the scop already in the list. */
1044 remove_intersecting_scops (s);
1046 scops.safe_push (s);
1047 DEBUG_PRINT (dp << "[scop-detection] Adding SCoP: "; print_sese (dump_file, s));
1050 /* Return true when a statement in SCOP cannot be represented by Graphite.
1051 The assumptions are that L1 dominates L2, and SCOP->entry dominates L1.
1052 Limit the number of bbs between adjacent loops to
1053 PARAM_SCOP_MAX_NUM_BBS_BETWEEN_LOOPS. */
1055 bool
1056 scop_detection::harmful_loop_in_region (sese_l scop) const
1058 basic_block exit_bb = get_exit_bb (scop);
1059 basic_block entry_bb = get_entry_bb (scop);
1061 DEBUG_PRINT (dp << "[checking-harmful-bbs] ";
1062 print_sese (dump_file, scop));
1063 gcc_assert (dominated_by_p (CDI_DOMINATORS, exit_bb, entry_bb));
1065 auto_vec<basic_block> worklist;
1066 auto_bitmap loops;
1068 worklist.safe_push (entry_bb);
1069 while (! worklist.is_empty ())
1071 basic_block bb = worklist.pop ();
1072 DEBUG_PRINT (dp << "Visiting bb_" << bb->index << "\n");
1074 /* The basic block should not be part of an irreducible loop. */
1075 if (bb->flags & BB_IRREDUCIBLE_LOOP)
1076 return true;
1078 /* Check for unstructured control flow: CFG not generated by structured
1079 if-then-else. */
1080 if (bb->succs->length () > 1)
1082 edge e;
1083 edge_iterator ei;
1084 FOR_EACH_EDGE (e, ei, bb->succs)
1085 if (!dominated_by_p (CDI_POST_DOMINATORS, bb, e->dest)
1086 && !dominated_by_p (CDI_DOMINATORS, e->dest, bb))
1087 return true;
1090 /* Collect all loops in the current region. */
1091 loop_p loop = bb->loop_father;
1092 if (loop_in_sese_p (loop, scop))
1093 bitmap_set_bit (loops, loop->num);
1094 else
1096 /* We only check for harmful statements in basic blocks not part of
1097 any loop fully contained in the scop: other bbs are checked below
1098 in loop_is_valid_in_scop. */
1099 if (harmful_stmt_in_bb (scop, bb))
1100 return true;
1103 if (bb != exit_bb)
1104 for (basic_block dom = first_dom_son (CDI_DOMINATORS, bb);
1105 dom;
1106 dom = next_dom_son (CDI_DOMINATORS, dom))
1107 worklist.safe_push (dom);
1110 /* Go through all loops and check that they are still valid in the combined
1111 scop. */
1112 unsigned j;
1113 bitmap_iterator bi;
1114 EXECUTE_IF_SET_IN_BITMAP (loops, 0, j, bi)
1116 loop_p loop = (*current_loops->larray)[j];
1117 gcc_assert (loop->num == (int) j);
1119 if (!loop_is_valid_in_scop (loop, scop))
1120 return true;
1123 return false;
1126 /* Returns true if S1 subsumes/surrounds S2. */
1127 bool
1128 scop_detection::subsumes (sese_l s1, sese_l s2)
1130 if (dominated_by_p (CDI_DOMINATORS, get_entry_bb (s2),
1131 get_entry_bb (s1))
1132 && dominated_by_p (CDI_POST_DOMINATORS, s2.exit->dest,
1133 s1.exit->dest))
1134 return true;
1135 return false;
1138 /* Remove a SCoP which is subsumed by S1. */
1139 void
1140 scop_detection::remove_subscops (sese_l s1)
1142 int j;
1143 sese_l *s2;
1144 FOR_EACH_VEC_ELT_REVERSE (scops, j, s2)
1146 if (subsumes (s1, *s2))
1148 DEBUG_PRINT (dp << "Removing sub-SCoP";
1149 print_sese (dump_file, *s2));
1150 scops.unordered_remove (j);
1155 /* Returns true if S1 intersects with S2. Since we already know that S1 does
1156 not subsume S2 or vice-versa, we only check for entry bbs. */
1158 bool
1159 scop_detection::intersects (sese_l s1, sese_l s2)
1161 if (dominated_by_p (CDI_DOMINATORS, get_entry_bb (s2),
1162 get_entry_bb (s1))
1163 && !dominated_by_p (CDI_DOMINATORS, get_entry_bb (s2),
1164 get_exit_bb (s1)))
1165 return true;
1166 if ((s1.exit == s2.entry) || (s2.exit == s1.entry))
1167 return true;
1169 return false;
1172 /* Remove one of the scops when it intersects with any other. */
1174 void
1175 scop_detection::remove_intersecting_scops (sese_l s1)
1177 int j;
1178 sese_l *s2;
1179 FOR_EACH_VEC_ELT_REVERSE (scops, j, s2)
1181 if (intersects (s1, *s2))
1183 DEBUG_PRINT (dp << "Removing intersecting SCoP";
1184 print_sese (dump_file, *s2);
1185 dp << "Intersects with:";
1186 print_sese (dump_file, s1));
1187 scops.unordered_remove (j);
1192 /* Something like "n * m" is not allowed. */
1194 bool
1195 scop_detection::graphite_can_represent_init (tree e)
1197 switch (TREE_CODE (e))
1199 case POLYNOMIAL_CHREC:
1200 return graphite_can_represent_init (CHREC_LEFT (e))
1201 && graphite_can_represent_init (CHREC_RIGHT (e));
1203 case MULT_EXPR:
1204 if (chrec_contains_symbols (TREE_OPERAND (e, 0)))
1205 return graphite_can_represent_init (TREE_OPERAND (e, 0))
1206 && tree_fits_shwi_p (TREE_OPERAND (e, 1));
1207 else
1208 return graphite_can_represent_init (TREE_OPERAND (e, 1))
1209 && tree_fits_shwi_p (TREE_OPERAND (e, 0));
1211 case PLUS_EXPR:
1212 case POINTER_PLUS_EXPR:
1213 case MINUS_EXPR:
1214 return graphite_can_represent_init (TREE_OPERAND (e, 0))
1215 && graphite_can_represent_init (TREE_OPERAND (e, 1));
1217 case NEGATE_EXPR:
1218 case BIT_NOT_EXPR:
1219 CASE_CONVERT:
1220 case NON_LVALUE_EXPR:
1221 return graphite_can_represent_init (TREE_OPERAND (e, 0));
1223 default:
1224 break;
1227 return true;
1230 /* Return true when SCEV can be represented in the polyhedral model.
1232 An expression can be represented, if it can be expressed as an
1233 affine expression. For loops (i, j) and parameters (m, n) all
1234 affine expressions are of the form:
1236 x1 * i + x2 * j + x3 * m + x4 * n + x5 * 1 where x1..x5 element of Z
1238 1 i + 20 j + (-2) m + 25
1240 Something like "i * n" or "n * m" is not allowed. */
1242 bool
1243 scop_detection::graphite_can_represent_scev (tree scev)
1245 if (chrec_contains_undetermined (scev))
1246 return false;
1248 /* We disable the handling of pointer types, because it’s currently not
1249 supported by Graphite with the isl AST generator. SSA_NAME nodes are
1250 the only nodes, which are disabled in case they are pointers to object
1251 types, but this can be changed. */
1253 if (POINTER_TYPE_P (TREE_TYPE (scev)) && TREE_CODE (scev) == SSA_NAME)
1254 return false;
1256 switch (TREE_CODE (scev))
1258 case NEGATE_EXPR:
1259 case BIT_NOT_EXPR:
1260 CASE_CONVERT:
1261 case NON_LVALUE_EXPR:
1262 return graphite_can_represent_scev (TREE_OPERAND (scev, 0));
1264 case PLUS_EXPR:
1265 case POINTER_PLUS_EXPR:
1266 case MINUS_EXPR:
1267 return graphite_can_represent_scev (TREE_OPERAND (scev, 0))
1268 && graphite_can_represent_scev (TREE_OPERAND (scev, 1));
1270 case MULT_EXPR:
1271 return !CONVERT_EXPR_CODE_P (TREE_CODE (TREE_OPERAND (scev, 0)))
1272 && !CONVERT_EXPR_CODE_P (TREE_CODE (TREE_OPERAND (scev, 1)))
1273 && !(chrec_contains_symbols (TREE_OPERAND (scev, 0))
1274 && chrec_contains_symbols (TREE_OPERAND (scev, 1)))
1275 && graphite_can_represent_init (scev)
1276 && graphite_can_represent_scev (TREE_OPERAND (scev, 0))
1277 && graphite_can_represent_scev (TREE_OPERAND (scev, 1));
1279 case POLYNOMIAL_CHREC:
1280 /* Check for constant strides. With a non constant stride of
1281 'n' we would have a value of 'iv * n'. Also check that the
1282 initial value can represented: for example 'n * m' cannot be
1283 represented. */
1284 if (!evolution_function_right_is_integer_cst (scev)
1285 || !graphite_can_represent_init (scev))
1286 return false;
1287 return graphite_can_represent_scev (CHREC_LEFT (scev));
1289 default:
1290 break;
1293 /* Only affine functions can be represented. */
1294 if (tree_contains_chrecs (scev, NULL) || !scev_is_linear_expression (scev))
1295 return false;
1297 return true;
1300 /* Return true when EXPR can be represented in the polyhedral model.
1302 This means an expression can be represented, if it is linear with respect to
1303 the loops and the strides are non parametric. LOOP is the place where the
1304 expr will be evaluated. SCOP defines the region we analyse. */
1306 bool
1307 scop_detection::graphite_can_represent_expr (sese_l scop, loop_p loop,
1308 tree expr)
1310 tree scev = scalar_evolution_in_region (scop, loop, expr);
1311 return graphite_can_represent_scev (scev);
1314 /* Return true if the data references of STMT can be represented by Graphite.
1315 We try to analyze the data references in a loop contained in the SCOP. */
1317 bool
1318 scop_detection::stmt_has_simple_data_refs_p (sese_l scop, gimple *stmt)
1320 loop_p nest = outermost_loop_in_sese (scop, gimple_bb (stmt));
1321 loop_p loop = loop_containing_stmt (stmt);
1322 vec<data_reference_p> drs = vNULL;
1324 graphite_find_data_references_in_stmt (nest, loop, stmt, &drs);
1326 int j;
1327 data_reference_p dr;
1328 FOR_EACH_VEC_ELT (drs, j, dr)
1330 int nb_subscripts = DR_NUM_DIMENSIONS (dr);
1332 if (nb_subscripts < 1)
1334 free_data_refs (drs);
1335 return false;
1338 tree ref = DR_REF (dr);
1340 for (int i = nb_subscripts - 1; i >= 0; i--)
1342 if (!graphite_can_represent_scev (DR_ACCESS_FN (dr, i))
1343 || (TREE_CODE (ref) != ARRAY_REF && TREE_CODE (ref) != MEM_REF
1344 && TREE_CODE (ref) != COMPONENT_REF))
1346 free_data_refs (drs);
1347 return false;
1350 ref = TREE_OPERAND (ref, 0);
1354 free_data_refs (drs);
1355 return true;
1358 /* GIMPLE_ASM and GIMPLE_CALL may embed arbitrary side effects.
1359 Calls have side-effects, except those to const or pure
1360 functions. */
1362 static bool
1363 stmt_has_side_effects (gimple *stmt)
1365 if (gimple_has_volatile_ops (stmt)
1366 || (gimple_code (stmt) == GIMPLE_CALL
1367 && !(gimple_call_flags (stmt) & (ECF_CONST | ECF_PURE)))
1368 || (gimple_code (stmt) == GIMPLE_ASM))
1370 DEBUG_PRINT (dp << "[scop-detection-fail] "
1371 << "Statement has side-effects:\n";
1372 print_gimple_stmt (dump_file, stmt, 0, TDF_VOPS | TDF_MEMSYMS));
1373 return true;
1375 return false;
1378 /* Returns true if STMT can be represented in polyhedral model. LABEL,
1379 simple COND stmts, pure calls, and assignments can be repesented. */
1381 bool
1382 scop_detection::graphite_can_represent_stmt (sese_l scop, gimple *stmt,
1383 basic_block bb)
1385 loop_p loop = bb->loop_father;
1386 switch (gimple_code (stmt))
1388 case GIMPLE_LABEL:
1389 return true;
1391 case GIMPLE_COND:
1393 /* We can handle all binary comparisons. Inequalities are
1394 also supported as they can be represented with union of
1395 polyhedra. */
1396 enum tree_code code = gimple_cond_code (stmt);
1397 if (!(code == LT_EXPR
1398 || code == GT_EXPR
1399 || code == LE_EXPR
1400 || code == GE_EXPR
1401 || code == EQ_EXPR
1402 || code == NE_EXPR))
1404 DEBUG_PRINT (dp << "[scop-detection-fail] "
1405 << "Graphite cannot handle cond stmt:\n";
1406 print_gimple_stmt (dump_file, stmt, 0,
1407 TDF_VOPS | TDF_MEMSYMS));
1408 return false;
1411 for (unsigned i = 0; i < 2; ++i)
1413 tree op = gimple_op (stmt, i);
1414 if (!graphite_can_represent_expr (scop, loop, op)
1415 /* We can only constrain on integer type. */
1416 || (TREE_CODE (TREE_TYPE (op)) != INTEGER_TYPE))
1418 DEBUG_PRINT (dp << "[scop-detection-fail] "
1419 << "Graphite cannot represent stmt:\n";
1420 print_gimple_stmt (dump_file, stmt, 0,
1421 TDF_VOPS | TDF_MEMSYMS));
1422 return false;
1426 return true;
1429 case GIMPLE_ASSIGN:
1430 case GIMPLE_CALL:
1431 return true;
1433 default:
1434 /* These nodes cut a new scope. */
1435 DEBUG_PRINT (
1436 dp << "[scop-detection-fail] "
1437 << "Gimple stmt not handled in Graphite:\n";
1438 print_gimple_stmt (dump_file, stmt, 0, TDF_VOPS | TDF_MEMSYMS));
1439 return false;
1443 /* Return true only when STMT is simple enough for being handled by Graphite.
1444 This depends on SCOP, as the parameters are initialized relatively to
1445 this basic block, the linear functions are initialized based on the outermost
1446 loop containing STMT inside the SCOP. BB is the place where we try to
1447 evaluate the STMT. */
1449 bool
1450 scop_detection::stmt_simple_for_scop_p (sese_l scop, gimple *stmt,
1451 basic_block bb) const
1453 gcc_assert (scop);
1455 if (is_gimple_debug (stmt))
1456 return true;
1458 if (stmt_has_side_effects (stmt))
1459 return false;
1461 if (!stmt_has_simple_data_refs_p (scop, stmt))
1463 DEBUG_PRINT (dp << "[scop-detection-fail] "
1464 << "Graphite cannot handle data-refs in stmt:\n";
1465 print_gimple_stmt (dump_file, stmt, 0, TDF_VOPS|TDF_MEMSYMS););
1466 return false;
1469 return graphite_can_represent_stmt (scop, stmt, bb);
1472 /* Return true when BB contains a harmful operation for a scop: that
1473 can be a function call with side effects, the induction variables
1474 are not linear with respect to SCOP, etc. The current open
1475 scop should end before this statement. */
1477 bool
1478 scop_detection::harmful_stmt_in_bb (sese_l scop, basic_block bb) const
1480 gimple_stmt_iterator gsi;
1482 for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi))
1483 if (!stmt_simple_for_scop_p (scop, gsi_stmt (gsi), bb))
1484 return true;
1486 return false;
1489 /* Return true when the body of LOOP has statements that can be represented as a
1490 valid scop. */
1492 bool
1493 scop_detection::loop_body_is_valid_scop (loop_p loop, sese_l scop) const
1495 if (!loop_ivs_can_be_represented (loop))
1497 DEBUG_PRINT (dp << "[scop-detection-fail] loop_" << loop->num
1498 << "IV cannot be represented.\n");
1499 return false;
1502 if (!loop_nest_has_data_refs (loop))
1504 DEBUG_PRINT (dp << "[scop-detection-fail] loop_" << loop->num
1505 << "does not have any data reference.\n");
1506 return false;
1509 basic_block *bbs = get_loop_body (loop);
1510 for (unsigned i = 0; i < loop->num_nodes; i++)
1512 basic_block bb = bbs[i];
1514 if (harmful_stmt_in_bb (scop, bb))
1516 free (bbs);
1517 return false;
1520 free (bbs);
1522 if (loop->inner)
1524 loop = loop->inner;
1525 while (loop)
1527 if (!loop_body_is_valid_scop (loop, scop))
1528 return false;
1529 loop = loop->next;
1533 return true;
1536 /* Returns the number of pbbs that are in loops contained in SCOP. */
1539 scop_detection::nb_pbbs_in_loops (scop_p scop)
1541 int i;
1542 poly_bb_p pbb;
1543 int res = 0;
1545 FOR_EACH_VEC_ELT (scop->pbbs, i, pbb)
1546 if (loop_in_sese_p (gbb_loop (PBB_BLACK_BOX (pbb)), scop->scop_info->region))
1547 res++;
1549 return res;
1552 /* When parameter NAME is in REGION, returns its index in SESE_PARAMS.
1553 Otherwise returns -1. */
1555 static inline int
1556 parameter_index_in_region_1 (tree name, sese_info_p region)
1558 int i;
1559 tree p;
1561 gcc_assert (TREE_CODE (name) == SSA_NAME);
1563 FOR_EACH_VEC_ELT (region->params, i, p)
1564 if (p == name)
1565 return i;
1567 return -1;
1570 /* When the parameter NAME is in REGION, returns its index in
1571 SESE_PARAMS. Otherwise this function inserts NAME in SESE_PARAMS
1572 and returns the index of NAME. */
1574 static int
1575 parameter_index_in_region (tree name, sese_info_p region)
1577 int i;
1579 gcc_assert (TREE_CODE (name) == SSA_NAME);
1581 /* Cannot constrain on anything else than INTEGER_TYPE parameters. */
1582 if (TREE_CODE (TREE_TYPE (name)) != INTEGER_TYPE)
1583 return -1;
1585 if (!invariant_in_sese_p_rec (name, region->region, NULL))
1586 return -1;
1588 i = parameter_index_in_region_1 (name, region);
1589 if (i != -1)
1590 return i;
1592 i = region->params.length ();
1593 region->params.safe_push (name);
1594 return i;
1597 /* In the context of sese S, scan the expression E and translate it to
1598 a linear expression C. When parsing a symbolic multiplication, K
1599 represents the constant multiplier of an expression containing
1600 parameters. */
1602 static void
1603 scan_tree_for_params (sese_info_p s, tree e)
1605 if (e == chrec_dont_know)
1606 return;
1608 switch (TREE_CODE (e))
1610 case POLYNOMIAL_CHREC:
1611 scan_tree_for_params (s, CHREC_LEFT (e));
1612 break;
1614 case MULT_EXPR:
1615 if (chrec_contains_symbols (TREE_OPERAND (e, 0)))
1616 scan_tree_for_params (s, TREE_OPERAND (e, 0));
1617 else
1618 scan_tree_for_params (s, TREE_OPERAND (e, 1));
1619 break;
1621 case PLUS_EXPR:
1622 case POINTER_PLUS_EXPR:
1623 case MINUS_EXPR:
1624 scan_tree_for_params (s, TREE_OPERAND (e, 0));
1625 scan_tree_for_params (s, TREE_OPERAND (e, 1));
1626 break;
1628 case NEGATE_EXPR:
1629 case BIT_NOT_EXPR:
1630 CASE_CONVERT:
1631 case NON_LVALUE_EXPR:
1632 scan_tree_for_params (s, TREE_OPERAND (e, 0));
1633 break;
1635 case SSA_NAME:
1636 parameter_index_in_region (e, s);
1637 break;
1639 case INTEGER_CST:
1640 case ADDR_EXPR:
1641 case REAL_CST:
1642 case COMPLEX_CST:
1643 case VECTOR_CST:
1644 break;
1646 default:
1647 gcc_unreachable ();
1648 break;
1652 /* Find parameters with respect to REGION in BB. We are looking in memory
1653 access functions, conditions and loop bounds. */
1655 static void
1656 find_params_in_bb (sese_info_p region, gimple_poly_bb_p gbb)
1658 /* Find parameters in the access functions of data references. */
1659 int i;
1660 data_reference_p dr;
1661 FOR_EACH_VEC_ELT (GBB_DATA_REFS (gbb), i, dr)
1662 for (unsigned j = 0; j < DR_NUM_DIMENSIONS (dr); j++)
1663 scan_tree_for_params (region, DR_ACCESS_FN (dr, j));
1665 /* Find parameters in conditional statements. */
1666 gimple *stmt;
1667 loop_p loop = GBB_BB (gbb)->loop_father;
1668 FOR_EACH_VEC_ELT (GBB_CONDITIONS (gbb), i, stmt)
1670 tree lhs = scalar_evolution_in_region (region->region, loop,
1671 gimple_cond_lhs (stmt));
1672 tree rhs = scalar_evolution_in_region (region->region, loop,
1673 gimple_cond_rhs (stmt));
1675 scan_tree_for_params (region, lhs);
1676 scan_tree_for_params (region, rhs);
1680 /* Record the parameters used in the SCOP. A variable is a parameter
1681 in a scop if it does not vary during the execution of that scop. */
1683 static void
1684 find_scop_parameters (scop_p scop)
1686 unsigned i;
1687 sese_info_p region = scop->scop_info;
1688 struct loop *loop;
1690 /* Find the parameters used in the loop bounds. */
1691 FOR_EACH_VEC_ELT (region->loop_nest, i, loop)
1693 tree nb_iters = number_of_latch_executions (loop);
1695 if (!chrec_contains_symbols (nb_iters))
1696 continue;
1698 nb_iters = scalar_evolution_in_region (region->region, loop, nb_iters);
1699 scan_tree_for_params (region, nb_iters);
1702 /* Find the parameters used in data accesses. */
1703 poly_bb_p pbb;
1704 FOR_EACH_VEC_ELT (scop->pbbs, i, pbb)
1705 find_params_in_bb (region, PBB_BLACK_BOX (pbb));
1707 int nbp = sese_nb_params (region);
1708 scop_set_nb_params (scop, nbp);
1711 /* Record DEF if it is used in other bbs different than DEF_BB in the SCOP. */
1713 static void
1714 build_cross_bb_scalars_def (scop_p scop, tree def, basic_block def_bb,
1715 vec<tree> *writes)
1717 if (!def || !is_gimple_reg (def))
1718 return;
1720 /* Do not gather scalar variables that can be analyzed by SCEV as they can be
1721 generated out of the induction variables. */
1722 if (scev_analyzable_p (def, scop->scop_info->region))
1723 return;
1725 gimple *use_stmt;
1726 imm_use_iterator imm_iter;
1727 FOR_EACH_IMM_USE_STMT (use_stmt, imm_iter, def)
1728 if (def_bb != gimple_bb (use_stmt) && !is_gimple_debug (use_stmt))
1730 writes->safe_push (def);
1731 DEBUG_PRINT (dp << "Adding scalar write: ";
1732 print_generic_expr (dump_file, def, 0);
1733 dp << "\nFrom stmt: ";
1734 print_gimple_stmt (dump_file,
1735 SSA_NAME_DEF_STMT (def), 0, 0));
1736 /* This is required by the FOR_EACH_IMM_USE_STMT when we want to break
1737 before all the uses have been visited. */
1738 BREAK_FROM_IMM_USE_STMT (imm_iter);
1742 /* Record DEF if it is used in other bbs different than DEF_BB in the SCOP. */
1744 static void
1745 build_cross_bb_scalars_use (scop_p scop, tree use, gimple *use_stmt,
1746 vec<scalar_use> *reads)
1748 gcc_assert (use);
1749 if (!is_gimple_reg (use))
1750 return;
1752 /* Do not gather scalar variables that can be analyzed by SCEV as they can be
1753 generated out of the induction variables. */
1754 if (scev_analyzable_p (use, scop->scop_info->region))
1755 return;
1757 gimple *def_stmt = SSA_NAME_DEF_STMT (use);
1758 if (gimple_bb (def_stmt) != gimple_bb (use_stmt))
1760 DEBUG_PRINT (dp << "Adding scalar read: ";
1761 print_generic_expr (dump_file, use, 0);
1762 dp << "\nFrom stmt: ";
1763 print_gimple_stmt (dump_file, use_stmt, 0, 0));
1764 reads->safe_push (std::make_pair (use_stmt, use));
1768 /* Record all scalar variables that are defined and used in different BBs of the
1769 SCOP. */
1771 static void
1772 graphite_find_cross_bb_scalar_vars (scop_p scop, gimple *stmt,
1773 vec<scalar_use> *reads, vec<tree> *writes)
1775 tree def;
1777 if (gimple_code (stmt) == GIMPLE_ASSIGN)
1778 def = gimple_assign_lhs (stmt);
1779 else if (gimple_code (stmt) == GIMPLE_CALL)
1780 def = gimple_call_lhs (stmt);
1781 else if (gimple_code (stmt) == GIMPLE_PHI)
1782 def = gimple_phi_result (stmt);
1783 else
1784 return;
1787 build_cross_bb_scalars_def (scop, def, gimple_bb (stmt), writes);
1789 ssa_op_iter iter;
1790 use_operand_p use_p;
1791 FOR_EACH_PHI_OR_STMT_USE (use_p, stmt, iter, SSA_OP_USE)
1793 tree use = USE_FROM_PTR (use_p);
1794 build_cross_bb_scalars_use (scop, use, stmt, reads);
1798 /* Generates a polyhedral black box only if the bb contains interesting
1799 information. */
1801 static gimple_poly_bb_p
1802 try_generate_gimple_bb (scop_p scop, basic_block bb)
1804 vec<data_reference_p> drs = vNULL;
1805 vec<tree> writes = vNULL;
1806 vec<scalar_use> reads = vNULL;
1808 sese_l region = scop->scop_info->region;
1809 loop_p nest = outermost_loop_in_sese (region, bb);
1811 loop_p loop = bb->loop_father;
1812 if (!loop_in_sese_p (loop, region))
1813 loop = nest;
1815 for (gimple_stmt_iterator gsi = gsi_start_bb (bb); !gsi_end_p (gsi);
1816 gsi_next (&gsi))
1818 gimple *stmt = gsi_stmt (gsi);
1819 if (is_gimple_debug (stmt))
1820 continue;
1822 graphite_find_data_references_in_stmt (nest, loop, stmt, &drs);
1823 graphite_find_cross_bb_scalar_vars (scop, stmt, &reads, &writes);
1826 for (gphi_iterator psi = gsi_start_phis (bb); !gsi_end_p (psi);
1827 gsi_next (&psi))
1828 if (!virtual_operand_p (gimple_phi_result (psi.phi ())))
1829 graphite_find_cross_bb_scalar_vars (scop, psi.phi (), &reads, &writes);
1831 if (drs.is_empty () && writes.is_empty () && reads.is_empty ())
1832 return NULL;
1834 return new_gimple_poly_bb (bb, drs, reads, writes);
1837 /* Compute alias-sets for all data references in DRS. */
1839 static void
1840 build_alias_set (scop_p scop)
1842 int num_vertices = scop->drs.length ();
1843 struct graph *g = new_graph (num_vertices);
1844 dr_info *dr1, *dr2;
1845 int i, j;
1846 int *all_vertices;
1848 FOR_EACH_VEC_ELT (scop->drs, i, dr1)
1849 for (j = i+1; scop->drs.iterate (j, &dr2); j++)
1850 if (dr_may_alias_p (dr1->dr, dr2->dr, true))
1852 add_edge (g, i, j);
1853 add_edge (g, j, i);
1856 all_vertices = XNEWVEC (int, num_vertices);
1857 for (i = 0; i < num_vertices; i++)
1858 all_vertices[i] = i;
1860 graphds_dfs (g, all_vertices, num_vertices, NULL, true, NULL);
1861 free (all_vertices);
1863 for (i = 0; i < g->n_vertices; i++)
1864 scop->drs[i].alias_set = g->vertices[i].component + 1;
1866 free_graph (g);
1869 /* Gather BBs and conditions for a SCOP. */
1870 class gather_bbs : public dom_walker
1872 public:
1873 gather_bbs (cdi_direction, scop_p);
1875 virtual edge before_dom_children (basic_block);
1876 virtual void after_dom_children (basic_block);
1878 private:
1879 auto_vec<gimple *, 3> conditions, cases;
1880 scop_p scop;
1883 gather_bbs::gather_bbs (cdi_direction direction, scop_p scop)
1884 : dom_walker (direction), scop (scop)
1888 /* Record in execution order the loops fully contained in the region. */
1890 static void
1891 record_loop_in_sese (basic_block bb, sese_info_p region)
1893 loop_p father = bb->loop_father;
1894 if (loop_in_sese_p (father, region->region))
1896 bool found = false;
1897 loop_p loop0;
1898 int j;
1899 FOR_EACH_VEC_ELT (region->loop_nest, j, loop0)
1900 if (father == loop0)
1902 found = true;
1903 break;
1905 if (!found)
1906 region->loop_nest.safe_push (father);
1910 /* Call-back for dom_walk executed before visiting the dominated
1911 blocks. */
1913 edge
1914 gather_bbs::before_dom_children (basic_block bb)
1916 sese_info_p region = scop->scop_info;
1917 if (!bb_in_sese_p (bb, region->region))
1918 return NULL;
1920 record_loop_in_sese (bb, region);
1922 gcond *stmt = single_pred_cond_non_loop_exit (bb);
1924 if (stmt)
1926 edge e = single_pred_edge (bb);
1928 conditions.safe_push (stmt);
1930 if (e->flags & EDGE_TRUE_VALUE)
1931 cases.safe_push (stmt);
1932 else
1933 cases.safe_push (NULL);
1936 scop->scop_info->bbs.safe_push (bb);
1938 gimple_poly_bb_p gbb = try_generate_gimple_bb (scop, bb);
1939 if (!gbb)
1940 return NULL;
1942 GBB_CONDITIONS (gbb) = conditions.copy ();
1943 GBB_CONDITION_CASES (gbb) = cases.copy ();
1945 poly_bb_p pbb = new_poly_bb (scop, gbb);
1946 scop->pbbs.safe_push (pbb);
1948 int i;
1949 data_reference_p dr;
1950 FOR_EACH_VEC_ELT (gbb->data_refs, i, dr)
1952 DEBUG_PRINT (dp << "Adding memory ";
1953 if (dr->is_read)
1954 dp << "read: ";
1955 else
1956 dp << "write: ";
1957 print_generic_expr (dump_file, dr->ref, 0);
1958 dp << "\nFrom stmt: ";
1959 print_gimple_stmt (dump_file, dr->stmt, 0, 0));
1961 scop->drs.safe_push (dr_info (dr, pbb));
1964 return NULL;
1967 /* Call-back for dom_walk executed after visiting the dominated
1968 blocks. */
1970 void
1971 gather_bbs::after_dom_children (basic_block bb)
1973 if (!bb_in_sese_p (bb, scop->scop_info->region))
1974 return;
1976 if (single_pred_cond_non_loop_exit (bb))
1978 conditions.pop ();
1979 cases.pop ();
1983 /* Find Static Control Parts (SCoP) in the current function and pushes
1984 them to SCOPS. */
1986 void
1987 build_scops (vec<scop_p> *scops)
1989 if (dump_file)
1990 dp.set_dump_file (dump_file);
1992 canonicalize_loop_closed_ssa_form ();
1994 scop_detection sb;
1995 sb.build_scop_depth (scop_detection::invalid_sese, current_loops->tree_root);
1997 /* Now create scops from the lightweight SESEs. */
1998 vec<sese_l> scops_l = sb.get_scops ();
1999 int i;
2000 sese_l *s;
2001 FOR_EACH_VEC_ELT (scops_l, i, s)
2003 scop_p scop = new_scop (s->entry, s->exit);
2005 /* Record all basic blocks and their conditions in REGION. */
2006 gather_bbs (CDI_DOMINATORS, scop).walk (cfun->cfg->x_entry_block_ptr);
2008 build_alias_set (scop);
2010 /* Do not optimize a scop containing only PBBs that do not belong
2011 to any loops. */
2012 if (sb.nb_pbbs_in_loops (scop) == 0)
2014 DEBUG_PRINT (dp << "[scop-detection-fail] no data references.\n");
2015 free_scop (scop);
2016 continue;
2019 unsigned max_arrays = PARAM_VALUE (PARAM_GRAPHITE_MAX_ARRAYS_PER_SCOP);
2020 if (scop->drs.length () >= max_arrays)
2022 DEBUG_PRINT (dp << "[scop-detection-fail] too many data references: "
2023 << scop->drs.length ()
2024 << " is larger than --param graphite-max-arrays-per-scop="
2025 << max_arrays << ".\n");
2026 free_scop (scop);
2027 continue;
2030 find_scop_parameters (scop);
2031 graphite_dim_t max_dim = PARAM_VALUE (PARAM_GRAPHITE_MAX_NB_SCOP_PARAMS);
2033 if (scop_nb_params (scop) > max_dim)
2035 DEBUG_PRINT (dp << "[scop-detection-fail] too many parameters: "
2036 << scop_nb_params (scop)
2037 << " larger than --param graphite-max-nb-scop-params="
2038 << max_dim << ".\n");
2039 free_scop (scop);
2040 continue;
2043 scops->safe_push (scop);
2046 DEBUG_PRINT (dp << "number of SCoPs: " << (scops ? scops->length () : 0););
2049 #endif /* HAVE_isl */