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)
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/>. */
29 #include "coretypes.h"
37 #include "fold-const.h"
38 #include "gimple-iterator.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"
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"
60 set_dump_file (FILE *f
)
66 friend debug_printer
&
67 operator<< (debug_printer
&output
, int i
)
69 fprintf (output
.dump_file
, "%d", i
);
72 friend debug_printer
&
73 operator<< (debug_printer
&output
, const char *s
)
75 fprintf (output
.dump_file
, "%s", s
);
80 #define DEBUG_PRINT(args) do \
82 if (dump_file && (dump_flags & TDF_DETAILS)) { args; } \
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.
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. */
96 dot_all_sese (FILE *file
, vec
<sese_l
>& scops
)
98 /* Disable debugging while printing graph. */
99 dump_flags_t tmp_dump_flags
= dump_flags
;
100 dump_flags
= TDF_NONE
;
102 fprintf (file
, "digraph all {\n");
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\" ",
114 fprintf (file
, "CELLSPACING=\"0\">\n");
116 /* Select color for SCoP. */
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
))
183 fprintf (file
, " <TR><TD WIDTH=\"50\" BGCOLOR=\"%s\">",
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
);
196 fprintf (file
, " %d ", bb
->index
);
198 fprintf (file
, "{lp_%d}", bb
->loop_father
->num
);
203 fprintf (file
, "</TD></TR>\n");
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
)
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. */
234 dot_sese (sese_l
& scop
)
240 scops
.safe_push (scop
);
242 dot_all_sese (stderr
, scops
);
252 dot_all_sese (stderr
, scops
);
256 /* Return true if BB is empty, contains only DEBUG_INSNs. */
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
)
270 /* Returns true when P1 and P2 are close phis with the same
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
288 remove_duplicate_close_phi (gphi
*phi
, gphi_iterator
*gsi
)
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. */
319 make_close_phi_nodes_unique (basic_block bb
)
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. */
333 while (!gsi_end_p (gsi
))
334 if (same_close_phi_node (phi
, gsi
.phi ()))
335 remove_duplicate_close_phi (phi
, &gsi
);
341 /* Return true when NAME is defined in LOOP. */
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. */
353 canonicalize_loop_closed_ssa (loop_p loop
)
355 edge e
= single_exit (loop
);
358 if (!e
|| e
->flags
& EDGE_ABNORMAL
)
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
);
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 ();
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
);
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
))
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),
401 use_p
= gimple_phi_arg_imm_use_ptr (phi
, i
);
402 replace_exp (use_p
, res
);
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
430 - the basic block containing the close phi nodes does not contain
433 - there exist only one phi node per definition in the loop.
437 canonicalize_loop_closed_ssa_form (void)
439 checking_verify_loop_closed_ssa (true);
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. */
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
);
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
)
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. */
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
;
486 if (loop_depth (pred
->loop_father
) > loop_depth (bb
->loop_father
))
489 stmt
= last_stmt (pred
);
491 if (stmt
&& gimple_code (stmt
) == GIMPLE_COND
)
492 return as_a
<gcond
*> (stmt
);
501 /* Build the maximal scop containing LOOPs and add it to SCOPS. */
506 scop_detection () : scops (vNULL
) {}
513 /* A marker for invalid sese_l. */
514 static sese_l invalid_sese
;
516 /* Return the SCOPS in this SCOP_DETECTION. */
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
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
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
,
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
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
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
);
667 sese_l
scop_detection::invalid_sese (NULL
, NULL
);
669 /* Return an sese_l around the LOOP. */
672 scop_detection::get_sese (loop_p loop
)
677 if (!loops_state_satisfies_p (LOOPS_HAVE_PREHEADERS
))
679 edge scop_end
= single_exit (loop
);
682 edge scop_begin
= loop_preheader_edge (loop
);
683 sese_l
s (scop_begin
, scop_end
);
687 /* Return the closest dominator with a single entry edge. */
690 scop_detection::get_nearest_dom_with_single_entry (basic_block dom
)
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
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
))
709 if (l
!= l2
&& l
== l1
710 && dominated_by_p (CDI_DOMINATORS
, e1
->src
, e2
->src
))
714 while (dom
->preds
->length () != 1)
716 if (dom
->preds
->length () < 1)
718 dom
= get_immediate_dominator (CDI_DOMINATORS
, dom
);
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. */
729 scop_detection::get_nearest_pdom_with_single_exit (basic_block pdom
)
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
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
))
749 if (l
!= l2
&& l
== l1
750 && dominated_by_p (CDI_POST_DOMINATORS
, e1
->dest
, e2
->dest
))
754 while (pdom
->succs
->length () != 1)
756 if (pdom
->succs
->length () < 1)
758 pdom
= get_immediate_dominator (CDI_POST_DOMINATORS
, pdom
);
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. */
770 scop_detection::merge_sese (sese_l first
, sese_l second
) const
772 /* In the trivial case first/second may be NULL. */
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
))
797 basic_block pdom
= nearest_common_dominator (CDI_POST_DOMINATORS
,
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
))
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
))
820 /* For now we just bail out when there is a loop exit in the region
821 that is not also the exit of the region. We could enlarge the
822 region to cover the loop that region exits to. See PR79977. */
823 if (loop_outer (entry
->src
->loop_father
))
825 vec
<edge
> exits
= get_loop_exit_edges (entry
->src
->loop_father
);
826 for (unsigned i
= 0; i
< exits
.length (); ++i
)
829 && bb_in_region (exits
[i
]->src
, entry
->dest
, exit
->src
))
831 DEBUG_PRINT (dp
<< "[scop-detection-fail] cannot merge seses.\n");
839 /* For now we just want to bail out when exit does not post-dominate entry.
840 TODO: We might just add a basic_block at the exit to make exit
841 post-dominate entry (the entire region). */
842 if (!dominated_by_p (CDI_POST_DOMINATORS
, get_entry_bb (combined
),
843 get_exit_bb (combined
))
844 || !dominated_by_p (CDI_DOMINATORS
, get_exit_bb (combined
),
845 get_entry_bb (combined
)))
847 DEBUG_PRINT (dp
<< "[scop-detection-fail] cannot merge seses.\n");
851 /* FIXME: We should remove this piece of code once
852 canonicalize_loop_closed_ssa has been removed, because that function
853 adds a BB with single exit. */
854 if (!trivially_empty_bb_p (get_exit_bb (combined
)))
856 /* Find the first empty succ (with single exit) of combined.exit. */
857 basic_block imm_succ
= combined
.exit
->dest
;
858 if (single_succ_p (imm_succ
)
859 && single_pred_p (imm_succ
)
860 && trivially_empty_bb_p (imm_succ
))
861 combined
.exit
= single_succ_edge (imm_succ
);
864 DEBUG_PRINT (dp
<< "[scop-detection-fail] Discarding SCoP because "
865 << "no single exit (empty succ) for sese exit";
866 print_sese (dump_file
, combined
));
871 /* Analyze all the BBs in new sese. */
872 if (harmful_loop_in_region (combined
))
875 DEBUG_PRINT (dp
<< "[merged-sese] s1: "; print_sese (dump_file
, combined
));
880 /* Build scop outer->inner if possible. */
883 scop_detection::build_scop_depth (sese_l s
, loop_p loop
)
888 DEBUG_PRINT (dp
<< "[Depth loop_" << loop
->num
<< "]\n");
889 s
= build_scop_depth (s
, loop
->inner
);
891 sese_l s2
= merge_sese (s
, get_sese (loop
));
894 /* s might be a valid scop, so return it and start analyzing from the
896 build_scop_depth (invalid_sese
, loop
->next
);
900 if (!loop_is_valid_in_scop (loop
, s2
))
901 return build_scop_depth (invalid_sese
, loop
->next
);
903 return build_scop_breadth (s2
, loop
);
906 /* If loop and loop->next are valid scops, try to merge them. */
909 scop_detection::build_scop_breadth (sese_l s1
, loop_p loop
)
913 DEBUG_PRINT (dp
<< "[Breadth loop_" << loop
->num
<< "]\n");
917 sese_l s2
= build_scop_depth (invalid_sese
, l
->next
);
925 sese_l combined
= merge_sese (s1
, s2
);
927 /* Combining adjacent loops may add unrelated loops into the
928 region so we have to check all sub-loops of the outer loop
929 that are in the combined region. */
931 for (l
= loop_outer (loop
)->inner
; l
; l
= l
->next
)
932 if (bb_in_sese_p (l
->header
, combined
)
933 && ! loop_is_valid_in_scop (l
, combined
))
935 combined
= invalid_sese
;
949 /* Returns true when Graphite can represent LOOP in SCOP.
950 FIXME: For the moment, graphite cannot be used on loops that iterate using
951 induction variables that wrap. */
954 scop_detection::can_represent_loop_1 (loop_p loop
, sese_l scop
)
957 struct tree_niter_desc niter_desc
;
959 return single_exit (loop
)
960 && !(loop_preheader_edge (loop
)->flags
& EDGE_IRREDUCIBLE_LOOP
)
961 && number_of_iterations_exit (loop
, single_exit (loop
), &niter_desc
, false)
962 && niter_desc
.control
.no_overflow
963 && (niter
= number_of_latch_executions (loop
))
964 && !chrec_contains_undetermined (niter
)
965 && !chrec_contains_undetermined (scalar_evolution_in_region (scop
,
967 && graphite_can_represent_expr (scop
, loop
, niter
);
970 /* Return true when all the loops within LOOP can be represented by
974 scop_detection::can_represent_loop (loop_p loop
, sese_l scop
)
976 if (!can_represent_loop_1 (loop
, scop
))
978 if (loop
->inner
&& !can_represent_loop (loop
->inner
, scop
))
980 if (loop
->next
&& !can_represent_loop (loop
->next
, scop
))
986 /* Return true when LOOP is a valid scop, that is a Static Control Part, a
987 region of code that can be represented in the polyhedral model. SCOP
988 defines the region we analyse. */
991 scop_detection::loop_is_valid_in_scop (loop_p loop
, sese_l scop
) const
996 if (!optimize_loop_nest_for_speed_p (loop
))
998 DEBUG_PRINT (dp
<< "[scop-detection-fail] loop_"
999 << loop
->num
<< " is not on a hot path.\n");
1003 if (!can_represent_loop (loop
, scop
))
1005 DEBUG_PRINT (dp
<< "[scop-detection-fail] cannot represent loop_"
1006 << loop
->num
<< "\n");
1010 if (loop_body_is_valid_scop (loop
, scop
))
1012 DEBUG_PRINT (dp
<< "[valid-scop] loop_" << loop
->num
1013 << " is a valid scop.\n");
1019 /* Return true when BEGIN is the preheader edge of a loop with a single exit
1023 scop_detection::region_has_one_loop (sese_l s
)
1025 edge begin
= s
.entry
;
1027 /* Check for a single perfectly nested loop. */
1028 if (begin
->dest
->loop_father
->inner
)
1031 /* Otherwise, check whether we have adjacent loops. */
1032 return begin
->dest
->loop_father
== end
->src
->loop_father
;
1035 /* Add to SCOPS a scop starting at SCOP_BEGIN and ending at SCOP_END. */
1038 scop_detection::add_scop (sese_l s
)
1042 /* Do not add scops with only one loop. */
1043 if (region_has_one_loop (s
))
1045 DEBUG_PRINT (dp
<< "[scop-detection-fail] Discarding one loop SCoP: ";
1046 print_sese (dump_file
, s
));
1050 if (get_exit_bb (s
) == EXIT_BLOCK_PTR_FOR_FN (cfun
))
1052 DEBUG_PRINT (dp
<< "[scop-detection-fail] "
1053 << "Discarding SCoP exiting to return: ";
1054 print_sese (dump_file
, s
));
1058 /* Remove all the scops which are subsumed by s. */
1059 remove_subscops (s
);
1061 /* Remove intersecting scops. FIXME: It will be a good idea to keep
1062 the non-intersecting part of the scop already in the list. */
1063 remove_intersecting_scops (s
);
1065 scops
.safe_push (s
);
1066 DEBUG_PRINT (dp
<< "[scop-detection] Adding SCoP: "; print_sese (dump_file
, s
));
1069 /* Return true when a statement in SCOP cannot be represented by Graphite.
1070 The assumptions are that L1 dominates L2, and SCOP->entry dominates L1.
1071 Limit the number of bbs between adjacent loops to
1072 PARAM_SCOP_MAX_NUM_BBS_BETWEEN_LOOPS. */
1075 scop_detection::harmful_loop_in_region (sese_l scop
) const
1077 basic_block exit_bb
= get_exit_bb (scop
);
1078 basic_block entry_bb
= get_entry_bb (scop
);
1080 DEBUG_PRINT (dp
<< "[checking-harmful-bbs] ";
1081 print_sese (dump_file
, scop
));
1082 gcc_assert (dominated_by_p (CDI_DOMINATORS
, exit_bb
, entry_bb
));
1084 auto_vec
<basic_block
> worklist
;
1087 worklist
.safe_push (entry_bb
);
1088 while (! worklist
.is_empty ())
1090 basic_block bb
= worklist
.pop ();
1091 DEBUG_PRINT (dp
<< "Visiting bb_" << bb
->index
<< "\n");
1093 /* The basic block should not be part of an irreducible loop. */
1094 if (bb
->flags
& BB_IRREDUCIBLE_LOOP
)
1097 /* Check for unstructured control flow: CFG not generated by structured
1099 if (bb
->succs
->length () > 1)
1103 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
1104 if (!dominated_by_p (CDI_POST_DOMINATORS
, bb
, e
->dest
)
1105 && !dominated_by_p (CDI_DOMINATORS
, e
->dest
, bb
))
1109 /* Collect all loops in the current region. */
1110 loop_p loop
= bb
->loop_father
;
1111 if (loop_in_sese_p (loop
, scop
))
1112 bitmap_set_bit (loops
, loop
->num
);
1115 /* We only check for harmful statements in basic blocks not part of
1116 any loop fully contained in the scop: other bbs are checked below
1117 in loop_is_valid_in_scop. */
1118 if (harmful_stmt_in_bb (scop
, bb
))
1123 for (basic_block dom
= first_dom_son (CDI_DOMINATORS
, bb
);
1125 dom
= next_dom_son (CDI_DOMINATORS
, dom
))
1126 worklist
.safe_push (dom
);
1129 /* Go through all loops and check that they are still valid in the combined
1133 EXECUTE_IF_SET_IN_BITMAP (loops
, 0, j
, bi
)
1135 loop_p loop
= (*current_loops
->larray
)[j
];
1136 gcc_assert (loop
->num
== (int) j
);
1138 if (!loop_is_valid_in_scop (loop
, scop
))
1145 /* Returns true if S1 subsumes/surrounds S2. */
1147 scop_detection::subsumes (sese_l s1
, sese_l s2
)
1149 if (dominated_by_p (CDI_DOMINATORS
, get_entry_bb (s2
),
1151 && dominated_by_p (CDI_POST_DOMINATORS
, s2
.exit
->dest
,
1157 /* Remove a SCoP which is subsumed by S1. */
1159 scop_detection::remove_subscops (sese_l s1
)
1163 FOR_EACH_VEC_ELT_REVERSE (scops
, j
, s2
)
1165 if (subsumes (s1
, *s2
))
1167 DEBUG_PRINT (dp
<< "Removing sub-SCoP";
1168 print_sese (dump_file
, *s2
));
1169 scops
.unordered_remove (j
);
1174 /* Returns true if S1 intersects with S2. Since we already know that S1 does
1175 not subsume S2 or vice-versa, we only check for entry bbs. */
1178 scop_detection::intersects (sese_l s1
, sese_l s2
)
1180 if (dominated_by_p (CDI_DOMINATORS
, get_entry_bb (s2
),
1182 && !dominated_by_p (CDI_DOMINATORS
, get_entry_bb (s2
),
1185 if ((s1
.exit
== s2
.entry
) || (s2
.exit
== s1
.entry
))
1191 /* Remove one of the scops when it intersects with any other. */
1194 scop_detection::remove_intersecting_scops (sese_l s1
)
1198 FOR_EACH_VEC_ELT_REVERSE (scops
, j
, s2
)
1200 if (intersects (s1
, *s2
))
1202 DEBUG_PRINT (dp
<< "Removing intersecting SCoP";
1203 print_sese (dump_file
, *s2
);
1204 dp
<< "Intersects with:";
1205 print_sese (dump_file
, s1
));
1206 scops
.unordered_remove (j
);
1211 /* Something like "n * m" is not allowed. */
1214 scop_detection::graphite_can_represent_init (tree e
)
1216 switch (TREE_CODE (e
))
1218 case POLYNOMIAL_CHREC
:
1219 return graphite_can_represent_init (CHREC_LEFT (e
))
1220 && graphite_can_represent_init (CHREC_RIGHT (e
));
1223 if (chrec_contains_symbols (TREE_OPERAND (e
, 0)))
1224 return graphite_can_represent_init (TREE_OPERAND (e
, 0))
1225 && tree_fits_shwi_p (TREE_OPERAND (e
, 1));
1227 return graphite_can_represent_init (TREE_OPERAND (e
, 1))
1228 && tree_fits_shwi_p (TREE_OPERAND (e
, 0));
1231 case POINTER_PLUS_EXPR
:
1233 return graphite_can_represent_init (TREE_OPERAND (e
, 0))
1234 && graphite_can_represent_init (TREE_OPERAND (e
, 1));
1239 case NON_LVALUE_EXPR
:
1240 return graphite_can_represent_init (TREE_OPERAND (e
, 0));
1249 /* Return true when SCEV can be represented in the polyhedral model.
1251 An expression can be represented, if it can be expressed as an
1252 affine expression. For loops (i, j) and parameters (m, n) all
1253 affine expressions are of the form:
1255 x1 * i + x2 * j + x3 * m + x4 * n + x5 * 1 where x1..x5 element of Z
1257 1 i + 20 j + (-2) m + 25
1259 Something like "i * n" or "n * m" is not allowed. */
1262 scop_detection::graphite_can_represent_scev (tree scev
)
1264 if (chrec_contains_undetermined (scev
))
1267 /* We disable the handling of pointer types, because it’s currently not
1268 supported by Graphite with the isl AST generator. SSA_NAME nodes are
1269 the only nodes, which are disabled in case they are pointers to object
1270 types, but this can be changed. */
1272 if (POINTER_TYPE_P (TREE_TYPE (scev
)) && TREE_CODE (scev
) == SSA_NAME
)
1275 switch (TREE_CODE (scev
))
1280 case NON_LVALUE_EXPR
:
1281 return graphite_can_represent_scev (TREE_OPERAND (scev
, 0));
1284 case POINTER_PLUS_EXPR
:
1286 return graphite_can_represent_scev (TREE_OPERAND (scev
, 0))
1287 && graphite_can_represent_scev (TREE_OPERAND (scev
, 1));
1290 return !CONVERT_EXPR_CODE_P (TREE_CODE (TREE_OPERAND (scev
, 0)))
1291 && !CONVERT_EXPR_CODE_P (TREE_CODE (TREE_OPERAND (scev
, 1)))
1292 && !(chrec_contains_symbols (TREE_OPERAND (scev
, 0))
1293 && chrec_contains_symbols (TREE_OPERAND (scev
, 1)))
1294 && graphite_can_represent_init (scev
)
1295 && graphite_can_represent_scev (TREE_OPERAND (scev
, 0))
1296 && graphite_can_represent_scev (TREE_OPERAND (scev
, 1));
1298 case POLYNOMIAL_CHREC
:
1299 /* Check for constant strides. With a non constant stride of
1300 'n' we would have a value of 'iv * n'. Also check that the
1301 initial value can represented: for example 'n * m' cannot be
1303 if (!evolution_function_right_is_integer_cst (scev
)
1304 || !graphite_can_represent_init (scev
))
1306 return graphite_can_represent_scev (CHREC_LEFT (scev
));
1312 /* Only affine functions can be represented. */
1313 if (tree_contains_chrecs (scev
, NULL
) || !scev_is_linear_expression (scev
))
1319 /* Return true when EXPR can be represented in the polyhedral model.
1321 This means an expression can be represented, if it is linear with respect to
1322 the loops and the strides are non parametric. LOOP is the place where the
1323 expr will be evaluated. SCOP defines the region we analyse. */
1326 scop_detection::graphite_can_represent_expr (sese_l scop
, loop_p loop
,
1329 tree scev
= scalar_evolution_in_region (scop
, loop
, expr
);
1330 return graphite_can_represent_scev (scev
);
1333 /* Return true if the data references of STMT can be represented by Graphite.
1334 We try to analyze the data references in a loop contained in the SCOP. */
1337 scop_detection::stmt_has_simple_data_refs_p (sese_l scop
, gimple
*stmt
)
1339 loop_p nest
= outermost_loop_in_sese (scop
, gimple_bb (stmt
));
1340 loop_p loop
= loop_containing_stmt (stmt
);
1341 vec
<data_reference_p
> drs
= vNULL
;
1343 graphite_find_data_references_in_stmt (nest
, loop
, stmt
, &drs
);
1346 data_reference_p dr
;
1347 FOR_EACH_VEC_ELT (drs
, j
, dr
)
1349 int nb_subscripts
= DR_NUM_DIMENSIONS (dr
);
1351 if (nb_subscripts
< 1)
1353 free_data_refs (drs
);
1357 tree ref
= DR_REF (dr
);
1359 for (int i
= nb_subscripts
- 1; i
>= 0; i
--)
1361 if (!graphite_can_represent_scev (DR_ACCESS_FN (dr
, i
))
1362 || (TREE_CODE (ref
) != ARRAY_REF
&& TREE_CODE (ref
) != MEM_REF
1363 && TREE_CODE (ref
) != COMPONENT_REF
))
1365 free_data_refs (drs
);
1369 ref
= TREE_OPERAND (ref
, 0);
1373 free_data_refs (drs
);
1377 /* GIMPLE_ASM and GIMPLE_CALL may embed arbitrary side effects.
1378 Calls have side-effects, except those to const or pure
1382 stmt_has_side_effects (gimple
*stmt
)
1384 if (gimple_has_volatile_ops (stmt
)
1385 || (gimple_code (stmt
) == GIMPLE_CALL
1386 && !(gimple_call_flags (stmt
) & (ECF_CONST
| ECF_PURE
)))
1387 || (gimple_code (stmt
) == GIMPLE_ASM
))
1389 DEBUG_PRINT (dp
<< "[scop-detection-fail] "
1390 << "Statement has side-effects:\n";
1391 print_gimple_stmt (dump_file
, stmt
, 0, TDF_VOPS
| TDF_MEMSYMS
));
1397 /* Returns true if STMT can be represented in polyhedral model. LABEL,
1398 simple COND stmts, pure calls, and assignments can be repesented. */
1401 scop_detection::graphite_can_represent_stmt (sese_l scop
, gimple
*stmt
,
1404 loop_p loop
= bb
->loop_father
;
1405 switch (gimple_code (stmt
))
1412 /* We can handle all binary comparisons. Inequalities are
1413 also supported as they can be represented with union of
1415 enum tree_code code
= gimple_cond_code (stmt
);
1416 if (!(code
== LT_EXPR
1421 || code
== NE_EXPR
))
1423 DEBUG_PRINT (dp
<< "[scop-detection-fail] "
1424 << "Graphite cannot handle cond stmt:\n";
1425 print_gimple_stmt (dump_file
, stmt
, 0,
1426 TDF_VOPS
| TDF_MEMSYMS
));
1430 for (unsigned i
= 0; i
< 2; ++i
)
1432 tree op
= gimple_op (stmt
, i
);
1433 if (!graphite_can_represent_expr (scop
, loop
, op
)
1434 /* We can only constrain on integer type. */
1435 || (TREE_CODE (TREE_TYPE (op
)) != INTEGER_TYPE
))
1437 DEBUG_PRINT (dp
<< "[scop-detection-fail] "
1438 << "Graphite cannot represent stmt:\n";
1439 print_gimple_stmt (dump_file
, stmt
, 0,
1440 TDF_VOPS
| TDF_MEMSYMS
));
1453 /* These nodes cut a new scope. */
1455 dp
<< "[scop-detection-fail] "
1456 << "Gimple stmt not handled in Graphite:\n";
1457 print_gimple_stmt (dump_file
, stmt
, 0, TDF_VOPS
| TDF_MEMSYMS
));
1462 /* Return true only when STMT is simple enough for being handled by Graphite.
1463 This depends on SCOP, as the parameters are initialized relatively to
1464 this basic block, the linear functions are initialized based on the outermost
1465 loop containing STMT inside the SCOP. BB is the place where we try to
1466 evaluate the STMT. */
1469 scop_detection::stmt_simple_for_scop_p (sese_l scop
, gimple
*stmt
,
1470 basic_block bb
) const
1474 if (is_gimple_debug (stmt
))
1477 if (stmt_has_side_effects (stmt
))
1480 if (!stmt_has_simple_data_refs_p (scop
, stmt
))
1482 DEBUG_PRINT (dp
<< "[scop-detection-fail] "
1483 << "Graphite cannot handle data-refs in stmt:\n";
1484 print_gimple_stmt (dump_file
, stmt
, 0, TDF_VOPS
|TDF_MEMSYMS
););
1488 return graphite_can_represent_stmt (scop
, stmt
, bb
);
1491 /* Return true when BB contains a harmful operation for a scop: that
1492 can be a function call with side effects, the induction variables
1493 are not linear with respect to SCOP, etc. The current open
1494 scop should end before this statement. */
1497 scop_detection::harmful_stmt_in_bb (sese_l scop
, basic_block bb
) const
1499 gimple_stmt_iterator gsi
;
1501 for (gsi
= gsi_start_bb (bb
); !gsi_end_p (gsi
); gsi_next (&gsi
))
1502 if (!stmt_simple_for_scop_p (scop
, gsi_stmt (gsi
), bb
))
1508 /* Return true when the body of LOOP has statements that can be represented as a
1512 scop_detection::loop_body_is_valid_scop (loop_p loop
, sese_l scop
) const
1514 if (!loop_ivs_can_be_represented (loop
))
1516 DEBUG_PRINT (dp
<< "[scop-detection-fail] loop_" << loop
->num
1517 << "IV cannot be represented.\n");
1521 if (!loop_nest_has_data_refs (loop
))
1523 DEBUG_PRINT (dp
<< "[scop-detection-fail] loop_" << loop
->num
1524 << "does not have any data reference.\n");
1528 basic_block
*bbs
= get_loop_body (loop
);
1529 for (unsigned i
= 0; i
< loop
->num_nodes
; i
++)
1531 basic_block bb
= bbs
[i
];
1533 if (harmful_stmt_in_bb (scop
, bb
))
1546 if (!loop_body_is_valid_scop (loop
, scop
))
1555 /* Returns the number of pbbs that are in loops contained in SCOP. */
1558 scop_detection::nb_pbbs_in_loops (scop_p scop
)
1564 FOR_EACH_VEC_ELT (scop
->pbbs
, i
, pbb
)
1565 if (loop_in_sese_p (gbb_loop (PBB_BLACK_BOX (pbb
)), scop
->scop_info
->region
))
1571 /* When parameter NAME is in REGION, returns its index in SESE_PARAMS.
1572 Otherwise returns -1. */
1575 parameter_index_in_region_1 (tree name
, sese_info_p region
)
1580 gcc_assert (TREE_CODE (name
) == SSA_NAME
);
1582 FOR_EACH_VEC_ELT (region
->params
, i
, p
)
1589 /* When the parameter NAME is in REGION, returns its index in
1590 SESE_PARAMS. Otherwise this function inserts NAME in SESE_PARAMS
1591 and returns the index of NAME. */
1594 parameter_index_in_region (tree name
, sese_info_p region
)
1598 gcc_assert (TREE_CODE (name
) == SSA_NAME
);
1600 /* Cannot constrain on anything else than INTEGER_TYPE parameters. */
1601 if (TREE_CODE (TREE_TYPE (name
)) != INTEGER_TYPE
)
1604 if (!invariant_in_sese_p_rec (name
, region
->region
, NULL
))
1607 i
= parameter_index_in_region_1 (name
, region
);
1611 i
= region
->params
.length ();
1612 region
->params
.safe_push (name
);
1616 /* In the context of sese S, scan the expression E and translate it to
1617 a linear expression C. When parsing a symbolic multiplication, K
1618 represents the constant multiplier of an expression containing
1622 scan_tree_for_params (sese_info_p s
, tree e
)
1624 if (e
== chrec_dont_know
)
1627 switch (TREE_CODE (e
))
1629 case POLYNOMIAL_CHREC
:
1630 scan_tree_for_params (s
, CHREC_LEFT (e
));
1634 if (chrec_contains_symbols (TREE_OPERAND (e
, 0)))
1635 scan_tree_for_params (s
, TREE_OPERAND (e
, 0));
1637 scan_tree_for_params (s
, TREE_OPERAND (e
, 1));
1641 case POINTER_PLUS_EXPR
:
1643 scan_tree_for_params (s
, TREE_OPERAND (e
, 0));
1644 scan_tree_for_params (s
, TREE_OPERAND (e
, 1));
1650 case NON_LVALUE_EXPR
:
1651 scan_tree_for_params (s
, TREE_OPERAND (e
, 0));
1655 parameter_index_in_region (e
, s
);
1671 /* Find parameters with respect to REGION in BB. We are looking in memory
1672 access functions, conditions and loop bounds. */
1675 find_params_in_bb (sese_info_p region
, gimple_poly_bb_p gbb
)
1677 /* Find parameters in the access functions of data references. */
1679 data_reference_p dr
;
1680 FOR_EACH_VEC_ELT (GBB_DATA_REFS (gbb
), i
, dr
)
1681 for (unsigned j
= 0; j
< DR_NUM_DIMENSIONS (dr
); j
++)
1682 scan_tree_for_params (region
, DR_ACCESS_FN (dr
, j
));
1684 /* Find parameters in conditional statements. */
1686 loop_p loop
= GBB_BB (gbb
)->loop_father
;
1687 FOR_EACH_VEC_ELT (GBB_CONDITIONS (gbb
), i
, stmt
)
1689 tree lhs
= scalar_evolution_in_region (region
->region
, loop
,
1690 gimple_cond_lhs (stmt
));
1691 tree rhs
= scalar_evolution_in_region (region
->region
, loop
,
1692 gimple_cond_rhs (stmt
));
1694 scan_tree_for_params (region
, lhs
);
1695 scan_tree_for_params (region
, rhs
);
1699 /* Record the parameters used in the SCOP. A variable is a parameter
1700 in a scop if it does not vary during the execution of that scop. */
1703 find_scop_parameters (scop_p scop
)
1706 sese_info_p region
= scop
->scop_info
;
1709 /* Find the parameters used in the loop bounds. */
1710 FOR_EACH_VEC_ELT (region
->loop_nest
, i
, loop
)
1712 tree nb_iters
= number_of_latch_executions (loop
);
1714 if (!chrec_contains_symbols (nb_iters
))
1717 nb_iters
= scalar_evolution_in_region (region
->region
, loop
, nb_iters
);
1718 scan_tree_for_params (region
, nb_iters
);
1721 /* Find the parameters used in data accesses. */
1723 FOR_EACH_VEC_ELT (scop
->pbbs
, i
, pbb
)
1724 find_params_in_bb (region
, PBB_BLACK_BOX (pbb
));
1726 int nbp
= sese_nb_params (region
);
1727 scop_set_nb_params (scop
, nbp
);
1730 /* Record DEF if it is used in other bbs different than DEF_BB in the SCOP. */
1733 build_cross_bb_scalars_def (scop_p scop
, tree def
, basic_block def_bb
,
1736 if (!def
|| !is_gimple_reg (def
))
1739 /* Do not gather scalar variables that can be analyzed by SCEV as they can be
1740 generated out of the induction variables. */
1741 if (scev_analyzable_p (def
, scop
->scop_info
->region
))
1745 imm_use_iterator imm_iter
;
1746 FOR_EACH_IMM_USE_STMT (use_stmt
, imm_iter
, def
)
1747 if (def_bb
!= gimple_bb (use_stmt
) && !is_gimple_debug (use_stmt
))
1749 writes
->safe_push (def
);
1750 DEBUG_PRINT (dp
<< "Adding scalar write: ";
1751 print_generic_expr (dump_file
, def
);
1752 dp
<< "\nFrom stmt: ";
1753 print_gimple_stmt (dump_file
,
1754 SSA_NAME_DEF_STMT (def
), 0));
1755 /* This is required by the FOR_EACH_IMM_USE_STMT when we want to break
1756 before all the uses have been visited. */
1757 BREAK_FROM_IMM_USE_STMT (imm_iter
);
1761 /* Record DEF if it is used in other bbs different than DEF_BB in the SCOP. */
1764 build_cross_bb_scalars_use (scop_p scop
, tree use
, gimple
*use_stmt
,
1765 vec
<scalar_use
> *reads
)
1768 if (!is_gimple_reg (use
))
1771 /* Do not gather scalar variables that can be analyzed by SCEV as they can be
1772 generated out of the induction variables. */
1773 if (scev_analyzable_p (use
, scop
->scop_info
->region
))
1776 gimple
*def_stmt
= SSA_NAME_DEF_STMT (use
);
1777 if (gimple_bb (def_stmt
) != gimple_bb (use_stmt
))
1779 DEBUG_PRINT (dp
<< "Adding scalar read: ";
1780 print_generic_expr (dump_file
, use
);
1781 dp
<< "\nFrom stmt: ";
1782 print_gimple_stmt (dump_file
, use_stmt
, 0));
1783 reads
->safe_push (std::make_pair (use_stmt
, use
));
1787 /* Record all scalar variables that are defined and used in different BBs of the
1791 graphite_find_cross_bb_scalar_vars (scop_p scop
, gimple
*stmt
,
1792 vec
<scalar_use
> *reads
, vec
<tree
> *writes
)
1796 if (gimple_code (stmt
) == GIMPLE_ASSIGN
)
1797 def
= gimple_assign_lhs (stmt
);
1798 else if (gimple_code (stmt
) == GIMPLE_CALL
)
1799 def
= gimple_call_lhs (stmt
);
1800 else if (gimple_code (stmt
) == GIMPLE_PHI
)
1801 def
= gimple_phi_result (stmt
);
1806 build_cross_bb_scalars_def (scop
, def
, gimple_bb (stmt
), writes
);
1809 use_operand_p use_p
;
1810 FOR_EACH_PHI_OR_STMT_USE (use_p
, stmt
, iter
, SSA_OP_USE
)
1812 tree use
= USE_FROM_PTR (use_p
);
1813 build_cross_bb_scalars_use (scop
, use
, stmt
, reads
);
1817 /* Generates a polyhedral black box only if the bb contains interesting
1820 static gimple_poly_bb_p
1821 try_generate_gimple_bb (scop_p scop
, basic_block bb
)
1823 vec
<data_reference_p
> drs
= vNULL
;
1824 vec
<tree
> writes
= vNULL
;
1825 vec
<scalar_use
> reads
= vNULL
;
1827 sese_l region
= scop
->scop_info
->region
;
1828 loop_p nest
= outermost_loop_in_sese (region
, bb
);
1830 loop_p loop
= bb
->loop_father
;
1831 if (!loop_in_sese_p (loop
, region
))
1834 for (gimple_stmt_iterator gsi
= gsi_start_bb (bb
); !gsi_end_p (gsi
);
1837 gimple
*stmt
= gsi_stmt (gsi
);
1838 if (is_gimple_debug (stmt
))
1841 graphite_find_data_references_in_stmt (nest
, loop
, stmt
, &drs
);
1842 graphite_find_cross_bb_scalar_vars (scop
, stmt
, &reads
, &writes
);
1845 for (gphi_iterator psi
= gsi_start_phis (bb
); !gsi_end_p (psi
);
1847 if (!virtual_operand_p (gimple_phi_result (psi
.phi ())))
1848 graphite_find_cross_bb_scalar_vars (scop
, psi
.phi (), &reads
, &writes
);
1850 if (drs
.is_empty () && writes
.is_empty () && reads
.is_empty ())
1853 return new_gimple_poly_bb (bb
, drs
, reads
, writes
);
1856 /* Compute alias-sets for all data references in DRS. */
1859 build_alias_set (scop_p scop
)
1861 int num_vertices
= scop
->drs
.length ();
1862 struct graph
*g
= new_graph (num_vertices
);
1867 FOR_EACH_VEC_ELT (scop
->drs
, i
, dr1
)
1868 for (j
= i
+1; scop
->drs
.iterate (j
, &dr2
); j
++)
1869 if (dr_may_alias_p (dr1
->dr
, dr2
->dr
, true))
1875 all_vertices
= XNEWVEC (int, num_vertices
);
1876 for (i
= 0; i
< num_vertices
; i
++)
1877 all_vertices
[i
] = i
;
1879 graphds_dfs (g
, all_vertices
, num_vertices
, NULL
, true, NULL
);
1880 free (all_vertices
);
1882 for (i
= 0; i
< g
->n_vertices
; i
++)
1883 scop
->drs
[i
].alias_set
= g
->vertices
[i
].component
+ 1;
1888 /* Gather BBs and conditions for a SCOP. */
1889 class gather_bbs
: public dom_walker
1892 gather_bbs (cdi_direction
, scop_p
);
1894 virtual edge
before_dom_children (basic_block
);
1895 virtual void after_dom_children (basic_block
);
1898 auto_vec
<gimple
*, 3> conditions
, cases
;
1902 gather_bbs::gather_bbs (cdi_direction direction
, scop_p scop
)
1903 : dom_walker (direction
), scop (scop
)
1907 /* Record in execution order the loops fully contained in the region. */
1910 record_loop_in_sese (basic_block bb
, sese_info_p region
)
1912 loop_p father
= bb
->loop_father
;
1913 if (loop_in_sese_p (father
, region
->region
))
1918 FOR_EACH_VEC_ELT (region
->loop_nest
, j
, loop0
)
1919 if (father
== loop0
)
1925 region
->loop_nest
.safe_push (father
);
1929 /* Call-back for dom_walk executed before visiting the dominated
1933 gather_bbs::before_dom_children (basic_block bb
)
1935 sese_info_p region
= scop
->scop_info
;
1936 if (!bb_in_sese_p (bb
, region
->region
))
1939 record_loop_in_sese (bb
, region
);
1941 gcond
*stmt
= single_pred_cond_non_loop_exit (bb
);
1945 edge e
= single_pred_edge (bb
);
1947 conditions
.safe_push (stmt
);
1949 if (e
->flags
& EDGE_TRUE_VALUE
)
1950 cases
.safe_push (stmt
);
1952 cases
.safe_push (NULL
);
1955 scop
->scop_info
->bbs
.safe_push (bb
);
1957 gimple_poly_bb_p gbb
= try_generate_gimple_bb (scop
, bb
);
1961 GBB_CONDITIONS (gbb
) = conditions
.copy ();
1962 GBB_CONDITION_CASES (gbb
) = cases
.copy ();
1964 poly_bb_p pbb
= new_poly_bb (scop
, gbb
);
1965 scop
->pbbs
.safe_push (pbb
);
1968 data_reference_p dr
;
1969 FOR_EACH_VEC_ELT (gbb
->data_refs
, i
, dr
)
1971 DEBUG_PRINT (dp
<< "Adding memory ";
1976 print_generic_expr (dump_file
, dr
->ref
);
1977 dp
<< "\nFrom stmt: ";
1978 print_gimple_stmt (dump_file
, dr
->stmt
, 0));
1980 scop
->drs
.safe_push (dr_info (dr
, pbb
));
1986 /* Call-back for dom_walk executed after visiting the dominated
1990 gather_bbs::after_dom_children (basic_block bb
)
1992 if (!bb_in_sese_p (bb
, scop
->scop_info
->region
))
1995 if (single_pred_cond_non_loop_exit (bb
))
2003 /* Compute sth like an execution order, dominator order with first executing
2004 edges that stay inside the current loop, delaying processing exit edges. */
2006 static vec
<unsigned> order
;
2009 get_order (scop_p scop
, basic_block bb
, vec
<unsigned> *order
, unsigned *dfs_num
)
2011 if (! bb_in_sese_p (bb
, scop
->scop_info
->region
))
2014 (*order
)[bb
->index
] = (*dfs_num
)++;
2015 for (basic_block son
= first_dom_son (CDI_DOMINATORS
, bb
);
2017 son
= next_dom_son (CDI_DOMINATORS
, son
))
2018 if (flow_bb_inside_loop_p (bb
->loop_father
, son
))
2019 get_order (scop
, son
, order
, dfs_num
);
2020 for (basic_block son
= first_dom_son (CDI_DOMINATORS
, bb
);
2022 son
= next_dom_son (CDI_DOMINATORS
, son
))
2023 if (! flow_bb_inside_loop_p (bb
->loop_father
, son
))
2024 get_order (scop
, son
, order
, dfs_num
);
2027 /* Helper for qsort, sorting after order above. */
2030 cmp_pbbs (const void *pa
, const void *pb
)
2032 poly_bb_p bb1
= *((const poly_bb_p
*)pa
);
2033 poly_bb_p bb2
= *((const poly_bb_p
*)pb
);
2034 if (order
[bb1
->black_box
->bb
->index
] < order
[bb2
->black_box
->bb
->index
])
2036 else if (order
[bb1
->black_box
->bb
->index
] > order
[bb2
->black_box
->bb
->index
])
2042 /* Find Static Control Parts (SCoP) in the current function and pushes
2046 build_scops (vec
<scop_p
> *scops
)
2049 dp
.set_dump_file (dump_file
);
2051 canonicalize_loop_closed_ssa_form ();
2054 sb
.build_scop_depth (scop_detection::invalid_sese
, current_loops
->tree_root
);
2056 /* Now create scops from the lightweight SESEs. */
2057 vec
<sese_l
> scops_l
= sb
.get_scops ();
2060 FOR_EACH_VEC_ELT (scops_l
, i
, s
)
2062 scop_p scop
= new_scop (s
->entry
, s
->exit
);
2064 /* Record all basic blocks and their conditions in REGION. */
2065 gather_bbs (CDI_DOMINATORS
, scop
).walk (s
->entry
->dest
);
2067 /* domwalk does not fulfil our code-generations constraints on the
2068 order of pbb which is to produce sth like execution order, delaying
2069 exection of loop exit edges. So compute such order and sort after
2071 order
.create (last_basic_block_for_fn (cfun
));
2072 order
.quick_grow (last_basic_block_for_fn (cfun
));
2073 unsigned dfs_num
= 0;
2074 get_order (scop
, s
->entry
->dest
, &order
, &dfs_num
);
2075 scop
->pbbs
.qsort (cmp_pbbs
);
2078 build_alias_set (scop
);
2080 /* Do not optimize a scop containing only PBBs that do not belong
2082 if (sb
.nb_pbbs_in_loops (scop
) == 0)
2084 DEBUG_PRINT (dp
<< "[scop-detection-fail] no data references.\n");
2089 unsigned max_arrays
= PARAM_VALUE (PARAM_GRAPHITE_MAX_ARRAYS_PER_SCOP
);
2090 if (scop
->drs
.length () >= max_arrays
)
2092 DEBUG_PRINT (dp
<< "[scop-detection-fail] too many data references: "
2093 << scop
->drs
.length ()
2094 << " is larger than --param graphite-max-arrays-per-scop="
2095 << max_arrays
<< ".\n");
2100 find_scop_parameters (scop
);
2101 graphite_dim_t max_dim
= PARAM_VALUE (PARAM_GRAPHITE_MAX_NB_SCOP_PARAMS
);
2103 if (scop_nb_params (scop
) > max_dim
)
2105 DEBUG_PRINT (dp
<< "[scop-detection-fail] too many parameters: "
2106 << scop_nb_params (scop
)
2107 << " larger than --param graphite-max-nb-scop-params="
2108 << max_dim
<< ".\n");
2113 scops
->safe_push (scop
);
2116 DEBUG_PRINT (dp
<< "number of SCoPs: " << (scops
? scops
->length () : 0););
2119 #endif /* HAVE_isl */