1 /* Detection of Static Control Parts (SCoP) for Graphite.
2 Copyright (C) 2009-2018 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"
61 set_dump_file (FILE *f
)
67 friend debug_printer
&
68 operator<< (debug_printer
&output
, int i
)
70 fprintf (output
.dump_file
, "%d", i
);
73 friend debug_printer
&
74 operator<< (debug_printer
&output
, const char *s
)
76 fprintf (output
.dump_file
, "%s", s
);
81 #define DEBUG_PRINT(args) do \
83 if (dump_file && (dump_flags & TDF_DETAILS)) { args; } \
86 /* Pretty print to FILE all the SCoPs in DOT format and mark them with
87 different colors. If there are not enough colors, paint the
88 remaining SCoPs in gray.
91 - "*" after the node number denotes the entry of a SCoP,
92 - "#" after the node number denotes the exit of a SCoP,
93 - "()" around the node number denotes the entry or the
94 exit nodes of the SCOP. These are not part of SCoP. */
97 dot_all_sese (FILE *file
, vec
<sese_l
>& scops
)
99 /* Disable debugging while printing graph. */
100 dump_flags_t tmp_dump_flags
= dump_flags
;
101 dump_flags
= TDF_NONE
;
103 fprintf (file
, "digraph all {\n");
106 FOR_ALL_BB_FN (bb
, cfun
)
108 int part_of_scop
= false;
110 /* Use HTML for every bb label. So we are able to print bbs
111 which are part of two different SCoPs, with two different
112 background colors. */
113 fprintf (file
, "%d [label=<\n <TABLE BORDER=\"0\" CELLBORDER=\"1\" ",
115 fprintf (file
, "CELLSPACING=\"0\">\n");
117 /* Select color for SCoP. */
120 FOR_EACH_VEC_ELT (scops
, i
, region
)
122 bool sese_in_region
= bb_in_sese_p (bb
, *region
);
123 if (sese_in_region
|| (region
->exit
->dest
== bb
)
124 || (region
->entry
->dest
== bb
))
184 fprintf (file
, " <TR><TD WIDTH=\"50\" BGCOLOR=\"%s\">",
188 fprintf (file
, " (");
190 if (bb
== region
->entry
->dest
&& bb
== region
->exit
->dest
)
191 fprintf (file
, " %d*# ", bb
->index
);
192 else if (bb
== region
->entry
->dest
)
193 fprintf (file
, " %d* ", bb
->index
);
194 else if (bb
== region
->exit
->dest
)
195 fprintf (file
, " %d# ", bb
->index
);
197 fprintf (file
, " %d ", bb
->index
);
199 fprintf (file
, "{lp_%d}", bb
->loop_father
->num
);
204 fprintf (file
, "</TD></TR>\n");
211 fprintf (file
, " <TR><TD WIDTH=\"50\" BGCOLOR=\"#ffffff\">");
212 fprintf (file
, " %d {lp_%d} </TD></TR>\n", bb
->index
,
213 bb
->loop_father
->num
);
215 fprintf (file
, " </TABLE>>, shape=box, style=\"setlinewidth(0)\"]\n");
218 FOR_ALL_BB_FN (bb
, cfun
)
222 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
223 fprintf (file
, "%d -> %d;\n", bb
->index
, e
->dest
->index
);
226 fputs ("}\n\n", file
);
228 /* Enable debugging again. */
229 dump_flags
= tmp_dump_flags
;
232 /* Display SCoP on stderr. */
235 dot_sese (sese_l
& scop
)
241 scops
.safe_push (scop
);
243 dot_all_sese (stderr
, scops
);
253 dot_all_sese (stderr
, scops
);
257 /* Returns a COND_EXPR statement when BB has a single predecessor, the
258 edge between BB and its predecessor is not a loop exit edge, and
259 the last statement of the single predecessor is a COND_EXPR. */
262 single_pred_cond_non_loop_exit (basic_block bb
)
264 if (single_pred_p (bb
))
266 edge e
= single_pred_edge (bb
);
267 basic_block pred
= e
->src
;
270 if (loop_depth (pred
->loop_father
) > loop_depth (bb
->loop_father
))
273 stmt
= last_stmt (pred
);
275 if (stmt
&& gimple_code (stmt
) == GIMPLE_COND
)
276 return as_a
<gcond
*> (stmt
);
285 /* Build the maximal scop containing LOOPs and add it to SCOPS. */
290 scop_detection () : scops (vNULL
) {}
297 /* A marker for invalid sese_l. */
298 static sese_l invalid_sese
;
300 /* Return the SCOPS in this SCOP_DETECTION. */
308 /* Return an sese_l around the LOOP. */
310 sese_l
get_sese (loop_p loop
);
312 /* Merge scops at same loop depth and returns the new sese.
313 Returns a new SESE when merge was successful, INVALID_SESE otherwise. */
315 sese_l
merge_sese (sese_l first
, sese_l second
) const;
317 /* Build scop outer->inner if possible. */
319 void build_scop_depth (loop_p loop
);
321 /* Return true when BEGIN is the preheader edge of a loop with a single exit
324 static bool region_has_one_loop (sese_l s
);
326 /* Add to SCOPS a scop starting at SCOP_BEGIN and ending at SCOP_END. */
328 void add_scop (sese_l s
);
330 /* Returns true if S1 subsumes/surrounds S2. */
331 static bool subsumes (sese_l s1
, sese_l s2
);
333 /* Remove a SCoP which is subsumed by S1. */
334 void remove_subscops (sese_l s1
);
336 /* Returns true if S1 intersects with S2. Since we already know that S1 does
337 not subsume S2 or vice-versa, we only check for entry bbs. */
339 static bool intersects (sese_l s1
, sese_l s2
);
341 /* Remove one of the scops when it intersects with any other. */
343 void remove_intersecting_scops (sese_l s1
);
345 /* Return true when a statement in SCOP cannot be represented by Graphite. */
347 bool harmful_loop_in_region (sese_l scop
) const;
349 /* Return true only when STMT is simple enough for being handled by Graphite.
350 This depends on SCOP, as the parameters are initialized relatively to
351 this basic block, the linear functions are initialized based on the
352 outermost loop containing STMT inside the SCOP. BB is the place where we
353 try to evaluate the STMT. */
355 bool stmt_simple_for_scop_p (sese_l scop
, gimple
*stmt
,
356 basic_block bb
) const;
358 /* Something like "n * m" is not allowed. */
360 static bool graphite_can_represent_init (tree e
);
362 /* Return true when SCEV can be represented in the polyhedral model.
364 An expression can be represented, if it can be expressed as an
365 affine expression. For loops (i, j) and parameters (m, n) all
366 affine expressions are of the form:
368 x1 * i + x2 * j + x3 * m + x4 * n + x5 * 1 where x1..x5 element of Z
370 1 i + 20 j + (-2) m + 25
372 Something like "i * n" or "n * m" is not allowed. */
374 static bool graphite_can_represent_scev (sese_l scop
, tree scev
);
376 /* Return true when EXPR can be represented in the polyhedral model.
378 This means an expression can be represented, if it is linear with respect
379 to the loops and the strides are non parametric. LOOP is the place where
380 the expr will be evaluated. SCOP defines the region we analyse. */
382 static bool graphite_can_represent_expr (sese_l scop
, loop_p loop
,
385 /* Return true if the data references of STMT can be represented by Graphite.
386 We try to analyze the data references in a loop contained in the SCOP. */
388 static bool stmt_has_simple_data_refs_p (sese_l scop
, gimple
*stmt
);
390 /* Remove the close phi node at GSI and replace its rhs with the rhs
393 static void remove_duplicate_close_phi (gphi
*phi
, gphi_iterator
*gsi
);
395 /* Returns true when Graphite can represent LOOP in SCOP.
396 FIXME: For the moment, graphite cannot be used on loops that iterate using
397 induction variables that wrap. */
399 static bool can_represent_loop (loop_p loop
, sese_l scop
);
401 /* Returns the number of pbbs that are in loops contained in SCOP. */
403 static int nb_pbbs_in_loops (scop_p scop
);
409 sese_l
scop_detection::invalid_sese (NULL
, NULL
);
411 /* Return an sese_l around the LOOP. */
414 scop_detection::get_sese (loop_p loop
)
419 edge scop_begin
= loop_preheader_edge (loop
);
420 edge scop_end
= single_exit (loop
);
421 if (!scop_end
|| (scop_end
->flags
& (EDGE_COMPLEX
|EDGE_FAKE
)))
424 return sese_l (scop_begin
, scop_end
);
427 /* Merge scops at same loop depth and returns the new sese.
428 Returns a new SESE when merge was successful, INVALID_SESE otherwise. */
431 scop_detection::merge_sese (sese_l first
, sese_l second
) const
433 /* In the trivial case first/second may be NULL. */
439 DEBUG_PRINT (dp
<< "[scop-detection] try merging sese s1: ";
440 print_sese (dump_file
, first
);
441 dp
<< "[scop-detection] try merging sese s2: ";
442 print_sese (dump_file
, second
));
444 auto_bitmap worklist
, in_sese_region
;
445 bitmap_set_bit (worklist
, get_entry_bb (first
)->index
);
446 bitmap_set_bit (worklist
, get_exit_bb (first
)->index
);
447 bitmap_set_bit (worklist
, get_entry_bb (second
)->index
);
448 bitmap_set_bit (worklist
, get_exit_bb (second
)->index
);
449 edge entry
= NULL
, exit
= NULL
;
451 /* We can optimize the case of adding a loop entry dest or exit
452 src to the worklist (for single-exit loops) by skipping
453 directly to the exit dest / entry src. in_sese_region
454 doesn't have to cover all blocks in the region but merely
455 its border it acts more like a visited bitmap. */
458 int index
= bitmap_first_set_bit (worklist
);
459 bitmap_clear_bit (worklist
, index
);
460 basic_block bb
= BASIC_BLOCK_FOR_FN (cfun
, index
);
464 /* With fake exit edges we can end up with no possible exit. */
465 if (index
== EXIT_BLOCK
)
467 DEBUG_PRINT (dp
<< "[scop-detection-fail] cannot merge seses.\n");
471 bitmap_set_bit (in_sese_region
, bb
->index
);
473 basic_block dom
= get_immediate_dominator (CDI_DOMINATORS
, bb
);
474 FOR_EACH_EDGE (e
, ei
, bb
->preds
)
477 || dominated_by_p (CDI_DOMINATORS
, entry
->dest
, bb
)))
480 && ! bitmap_bit_p (in_sese_region
, entry
->src
->index
))
481 bitmap_set_bit (worklist
, entry
->src
->index
);
484 else if (! bitmap_bit_p (in_sese_region
, e
->src
->index
))
485 bitmap_set_bit (worklist
, e
->src
->index
);
487 basic_block pdom
= get_immediate_dominator (CDI_POST_DOMINATORS
, bb
);
488 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
491 || dominated_by_p (CDI_POST_DOMINATORS
, exit
->src
, bb
)))
494 && ! bitmap_bit_p (in_sese_region
, exit
->dest
->index
))
495 bitmap_set_bit (worklist
, exit
->dest
->index
);
498 else if (! bitmap_bit_p (in_sese_region
, e
->dest
->index
))
499 bitmap_set_bit (worklist
, e
->dest
->index
);
501 while (! bitmap_empty_p (worklist
));
503 sese_l
combined (entry
, exit
);
505 DEBUG_PRINT (dp
<< "[merged-sese] s1: "; print_sese (dump_file
, combined
));
510 /* Build scop outer->inner if possible. */
513 scop_detection::build_scop_depth (loop_p loop
)
515 sese_l s
= invalid_sese
;
519 sese_l next
= get_sese (loop
);
521 || harmful_loop_in_region (next
))
525 build_scop_depth (loop
);
532 sese_l combined
= merge_sese (s
, next
);
534 || harmful_loop_in_region (combined
))
548 /* Returns true when Graphite can represent LOOP in SCOP.
549 FIXME: For the moment, graphite cannot be used on loops that iterate using
550 induction variables that wrap. */
553 scop_detection::can_represent_loop (loop_p loop
, sese_l scop
)
556 struct tree_niter_desc niter_desc
;
558 return single_exit (loop
)
559 && !(loop_preheader_edge (loop
)->flags
& EDGE_IRREDUCIBLE_LOOP
)
560 && number_of_iterations_exit (loop
, single_exit (loop
), &niter_desc
, false)
561 && niter_desc
.control
.no_overflow
562 && (niter
= number_of_latch_executions (loop
))
563 && !chrec_contains_undetermined (niter
)
564 && !chrec_contains_undetermined (scalar_evolution_in_region (scop
,
566 && graphite_can_represent_expr (scop
, loop
, niter
);
569 /* Return true when BEGIN is the preheader edge of a loop with a single exit
573 scop_detection::region_has_one_loop (sese_l s
)
575 edge begin
= s
.entry
;
577 /* Check for a single perfectly nested loop. */
578 if (begin
->dest
->loop_father
->inner
)
581 /* Otherwise, check whether we have adjacent loops. */
582 return (single_pred_p (end
->src
)
583 && begin
->dest
->loop_father
== single_pred (end
->src
)->loop_father
);
586 /* Add to SCOPS a scop starting at SCOP_BEGIN and ending at SCOP_END. */
589 scop_detection::add_scop (sese_l s
)
593 /* If the exit edge is fake discard the SCoP for now as we're removing the
594 fake edges again after analysis. */
595 if (s
.exit
->flags
& EDGE_FAKE
)
597 DEBUG_PRINT (dp
<< "[scop-detection-fail] Discarding infinite loop SCoP: ";
598 print_sese (dump_file
, s
));
602 /* Include the BB with the loop-closed SSA PHI nodes, we need this
603 block in the region for code-generating out-of-SSA copies.
604 canonicalize_loop_closed_ssa makes sure that is in proper shape. */
605 if (s
.exit
->dest
!= EXIT_BLOCK_PTR_FOR_FN (cfun
)
606 && loop_exit_edge_p (s
.exit
->src
->loop_father
, s
.exit
))
608 gcc_assert (single_pred_p (s
.exit
->dest
)
609 && single_succ_p (s
.exit
->dest
)
610 && sese_trivially_empty_bb_p (s
.exit
->dest
));
611 s
.exit
= single_succ_edge (s
.exit
->dest
);
614 /* Do not add scops with only one loop. */
615 if (region_has_one_loop (s
))
617 DEBUG_PRINT (dp
<< "[scop-detection-fail] Discarding one loop SCoP: ";
618 print_sese (dump_file
, s
));
622 if (get_exit_bb (s
) == EXIT_BLOCK_PTR_FOR_FN (cfun
))
624 DEBUG_PRINT (dp
<< "[scop-detection-fail] "
625 << "Discarding SCoP exiting to return: ";
626 print_sese (dump_file
, s
));
630 /* Remove all the scops which are subsumed by s. */
633 /* Remove intersecting scops. FIXME: It will be a good idea to keep
634 the non-intersecting part of the scop already in the list. */
635 remove_intersecting_scops (s
);
638 DEBUG_PRINT (dp
<< "[scop-detection] Adding SCoP: "; print_sese (dump_file
, s
));
641 /* Return true when a statement in SCOP cannot be represented by Graphite. */
644 scop_detection::harmful_loop_in_region (sese_l scop
) const
646 basic_block exit_bb
= get_exit_bb (scop
);
647 basic_block entry_bb
= get_entry_bb (scop
);
649 DEBUG_PRINT (dp
<< "[checking-harmful-bbs] ";
650 print_sese (dump_file
, scop
));
651 gcc_assert (dominated_by_p (CDI_DOMINATORS
, exit_bb
, entry_bb
));
653 auto_vec
<basic_block
> worklist
;
656 worklist
.safe_push (entry_bb
);
657 while (! worklist
.is_empty ())
659 basic_block bb
= worklist
.pop ();
660 DEBUG_PRINT (dp
<< "Visiting bb_" << bb
->index
<< "\n");
662 /* The basic block should not be part of an irreducible loop. */
663 if (bb
->flags
& BB_IRREDUCIBLE_LOOP
)
666 /* Check for unstructured control flow: CFG not generated by structured
668 if (bb
->succs
->length () > 1)
672 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
673 if (!dominated_by_p (CDI_POST_DOMINATORS
, bb
, e
->dest
)
674 && !dominated_by_p (CDI_DOMINATORS
, e
->dest
, bb
))
678 /* Collect all loops in the current region. */
679 loop_p loop
= bb
->loop_father
;
680 if (loop_in_sese_p (loop
, scop
))
681 bitmap_set_bit (loops
, loop
->num
);
683 /* Check for harmful statements in basic blocks part of the region. */
684 for (gimple_stmt_iterator gsi
= gsi_start_bb (bb
);
685 !gsi_end_p (gsi
); gsi_next (&gsi
))
686 if (!stmt_simple_for_scop_p (scop
, gsi_stmt (gsi
), bb
))
689 for (basic_block dom
= first_dom_son (CDI_DOMINATORS
, bb
);
691 dom
= next_dom_son (CDI_DOMINATORS
, dom
))
692 if (dom
!= scop
.exit
->dest
)
693 worklist
.safe_push (dom
);
696 /* Go through all loops and check that they are still valid in the combined
700 EXECUTE_IF_SET_IN_BITMAP (loops
, 0, j
, bi
)
702 loop_p loop
= (*current_loops
->larray
)[j
];
703 gcc_assert (loop
->num
== (int) j
);
705 /* Check if the loop nests are to be optimized for speed. */
707 && ! optimize_loop_for_speed_p (loop
))
709 DEBUG_PRINT (dp
<< "[scop-detection-fail] loop_"
710 << loop
->num
<< " is not on a hot path.\n");
714 if (! can_represent_loop (loop
, scop
))
716 DEBUG_PRINT (dp
<< "[scop-detection-fail] cannot represent loop_"
717 << loop
->num
<< "\n");
721 /* Check if all loop nests have at least one data reference.
722 ??? This check is expensive and loops premature at this point.
723 If important to retain we can pre-compute this for all innermost
724 loops and reject those when we build a SESE region for a loop
725 during SESE discovery. */
727 && ! loop_nest_has_data_refs (loop
))
729 DEBUG_PRINT (dp
<< "[scop-detection-fail] loop_" << loop
->num
730 << "does not have any data reference.\n");
738 /* Returns true if S1 subsumes/surrounds S2. */
740 scop_detection::subsumes (sese_l s1
, sese_l s2
)
742 if (dominated_by_p (CDI_DOMINATORS
, get_entry_bb (s2
),
744 && dominated_by_p (CDI_POST_DOMINATORS
, s2
.exit
->dest
,
750 /* Remove a SCoP which is subsumed by S1. */
752 scop_detection::remove_subscops (sese_l s1
)
756 FOR_EACH_VEC_ELT_REVERSE (scops
, j
, s2
)
758 if (subsumes (s1
, *s2
))
760 DEBUG_PRINT (dp
<< "Removing sub-SCoP";
761 print_sese (dump_file
, *s2
));
762 scops
.unordered_remove (j
);
767 /* Returns true if S1 intersects with S2. Since we already know that S1 does
768 not subsume S2 or vice-versa, we only check for entry bbs. */
771 scop_detection::intersects (sese_l s1
, sese_l s2
)
773 if (dominated_by_p (CDI_DOMINATORS
, get_entry_bb (s2
),
775 && !dominated_by_p (CDI_DOMINATORS
, get_entry_bb (s2
),
778 if ((s1
.exit
== s2
.entry
) || (s2
.exit
== s1
.entry
))
784 /* Remove one of the scops when it intersects with any other. */
787 scop_detection::remove_intersecting_scops (sese_l s1
)
791 FOR_EACH_VEC_ELT_REVERSE (scops
, j
, s2
)
793 if (intersects (s1
, *s2
))
795 DEBUG_PRINT (dp
<< "Removing intersecting SCoP";
796 print_sese (dump_file
, *s2
);
797 dp
<< "Intersects with:";
798 print_sese (dump_file
, s1
));
799 scops
.unordered_remove (j
);
804 /* Something like "n * m" is not allowed. */
807 scop_detection::graphite_can_represent_init (tree e
)
809 switch (TREE_CODE (e
))
811 case POLYNOMIAL_CHREC
:
812 return graphite_can_represent_init (CHREC_LEFT (e
))
813 && graphite_can_represent_init (CHREC_RIGHT (e
));
816 if (chrec_contains_symbols (TREE_OPERAND (e
, 0)))
817 return graphite_can_represent_init (TREE_OPERAND (e
, 0))
818 && tree_fits_shwi_p (TREE_OPERAND (e
, 1));
820 return graphite_can_represent_init (TREE_OPERAND (e
, 1))
821 && tree_fits_shwi_p (TREE_OPERAND (e
, 0));
824 case POINTER_PLUS_EXPR
:
826 return graphite_can_represent_init (TREE_OPERAND (e
, 0))
827 && graphite_can_represent_init (TREE_OPERAND (e
, 1));
832 case NON_LVALUE_EXPR
:
833 return graphite_can_represent_init (TREE_OPERAND (e
, 0));
842 /* Return true when SCEV can be represented in the polyhedral model.
844 An expression can be represented, if it can be expressed as an
845 affine expression. For loops (i, j) and parameters (m, n) all
846 affine expressions are of the form:
848 x1 * i + x2 * j + x3 * m + x4 * n + x5 * 1 where x1..x5 element of Z
850 1 i + 20 j + (-2) m + 25
852 Something like "i * n" or "n * m" is not allowed. */
855 scop_detection::graphite_can_represent_scev (sese_l scop
, tree scev
)
857 if (chrec_contains_undetermined (scev
))
860 switch (TREE_CODE (scev
))
865 case NON_LVALUE_EXPR
:
866 return graphite_can_represent_scev (scop
, TREE_OPERAND (scev
, 0));
869 case POINTER_PLUS_EXPR
:
871 return graphite_can_represent_scev (scop
, TREE_OPERAND (scev
, 0))
872 && graphite_can_represent_scev (scop
, TREE_OPERAND (scev
, 1));
875 return !CONVERT_EXPR_CODE_P (TREE_CODE (TREE_OPERAND (scev
, 0)))
876 && !CONVERT_EXPR_CODE_P (TREE_CODE (TREE_OPERAND (scev
, 1)))
877 && !(chrec_contains_symbols (TREE_OPERAND (scev
, 0))
878 && chrec_contains_symbols (TREE_OPERAND (scev
, 1)))
879 && graphite_can_represent_init (scev
)
880 && graphite_can_represent_scev (scop
, TREE_OPERAND (scev
, 0))
881 && graphite_can_represent_scev (scop
, TREE_OPERAND (scev
, 1));
883 case POLYNOMIAL_CHREC
:
884 /* Check for constant strides. With a non constant stride of
885 'n' we would have a value of 'iv * n'. Also check that the
886 initial value can represented: for example 'n * m' cannot be
888 gcc_assert (loop_in_sese_p (get_loop (cfun
,
889 CHREC_VARIABLE (scev
)), scop
));
890 if (!evolution_function_right_is_integer_cst (scev
)
891 || !graphite_can_represent_init (scev
))
893 return graphite_can_represent_scev (scop
, CHREC_LEFT (scev
));
899 /* Only affine functions can be represented. */
900 if (tree_contains_chrecs (scev
, NULL
) || !scev_is_linear_expression (scev
))
906 /* Return true when EXPR can be represented in the polyhedral model.
908 This means an expression can be represented, if it is linear with respect to
909 the loops and the strides are non parametric. LOOP is the place where the
910 expr will be evaluated. SCOP defines the region we analyse. */
913 scop_detection::graphite_can_represent_expr (sese_l scop
, loop_p loop
,
916 tree scev
= scalar_evolution_in_region (scop
, loop
, expr
);
917 return graphite_can_represent_scev (scop
, scev
);
920 /* Return true if the data references of STMT can be represented by Graphite.
921 We try to analyze the data references in a loop contained in the SCOP. */
924 scop_detection::stmt_has_simple_data_refs_p (sese_l scop
, gimple
*stmt
)
926 edge nest
= scop
.entry
;
927 loop_p loop
= loop_containing_stmt (stmt
);
928 if (!loop_in_sese_p (loop
, scop
))
931 auto_vec
<data_reference_p
> drs
;
932 if (! graphite_find_data_references_in_stmt (nest
, loop
, stmt
, &drs
))
937 FOR_EACH_VEC_ELT (drs
, j
, dr
)
939 for (unsigned i
= 0; i
< DR_NUM_DIMENSIONS (dr
); ++i
)
940 if (! graphite_can_represent_scev (scop
, DR_ACCESS_FN (dr
, i
)))
947 /* GIMPLE_ASM and GIMPLE_CALL may embed arbitrary side effects.
948 Calls have side-effects, except those to const or pure
952 stmt_has_side_effects (gimple
*stmt
)
954 if (gimple_has_volatile_ops (stmt
)
955 || (gimple_code (stmt
) == GIMPLE_CALL
956 && !(gimple_call_flags (stmt
) & (ECF_CONST
| ECF_PURE
)))
957 || (gimple_code (stmt
) == GIMPLE_ASM
))
959 DEBUG_PRINT (dp
<< "[scop-detection-fail] "
960 << "Statement has side-effects:\n";
961 print_gimple_stmt (dump_file
, stmt
, 0, TDF_VOPS
| TDF_MEMSYMS
));
967 /* Return true only when STMT is simple enough for being handled by Graphite.
968 This depends on SCOP, as the parameters are initialized relatively to
969 this basic block, the linear functions are initialized based on the outermost
970 loop containing STMT inside the SCOP. BB is the place where we try to
971 evaluate the STMT. */
974 scop_detection::stmt_simple_for_scop_p (sese_l scop
, gimple
*stmt
,
975 basic_block bb
) const
979 if (is_gimple_debug (stmt
))
982 if (stmt_has_side_effects (stmt
))
985 if (!stmt_has_simple_data_refs_p (scop
, stmt
))
987 DEBUG_PRINT (dp
<< "[scop-detection-fail] "
988 << "Graphite cannot handle data-refs in stmt:\n";
989 print_gimple_stmt (dump_file
, stmt
, 0, TDF_VOPS
|TDF_MEMSYMS
););
993 switch (gimple_code (stmt
))
1000 /* We can handle all binary comparisons. Inequalities are
1001 also supported as they can be represented with union of
1003 enum tree_code code
= gimple_cond_code (stmt
);
1004 if (!(code
== LT_EXPR
1009 || code
== NE_EXPR
))
1011 DEBUG_PRINT (dp
<< "[scop-detection-fail] "
1012 << "Graphite cannot handle cond stmt:\n";
1013 print_gimple_stmt (dump_file
, stmt
, 0,
1014 TDF_VOPS
| TDF_MEMSYMS
));
1018 loop_p loop
= bb
->loop_father
;
1019 for (unsigned i
= 0; i
< 2; ++i
)
1021 tree op
= gimple_op (stmt
, i
);
1022 if (!graphite_can_represent_expr (scop
, loop
, op
)
1023 /* We can only constrain on integer type. */
1024 || ! INTEGRAL_TYPE_P (TREE_TYPE (op
)))
1026 DEBUG_PRINT (dp
<< "[scop-detection-fail] "
1027 << "Graphite cannot represent stmt:\n";
1028 print_gimple_stmt (dump_file
, stmt
, 0,
1029 TDF_VOPS
| TDF_MEMSYMS
));
1040 tree op
, lhs
= gimple_get_lhs (stmt
);
1042 /* If we are not going to instantiate the stmt do not require
1043 its operands to be instantiatable at this point. */
1045 && TREE_CODE (lhs
) == SSA_NAME
1046 && scev_analyzable_p (lhs
, scop
))
1048 /* Verify that if we can analyze operands at their def site we
1049 also can represent them when analyzed at their uses. */
1050 FOR_EACH_SSA_TREE_OPERAND (op
, stmt
, i
, SSA_OP_USE
)
1051 if (scev_analyzable_p (op
, scop
)
1052 && chrec_contains_undetermined
1053 (scalar_evolution_in_region (scop
, bb
->loop_father
, op
)))
1055 DEBUG_PRINT (dp
<< "[scop-detection-fail] "
1056 << "Graphite cannot code-gen stmt:\n";
1057 print_gimple_stmt (dump_file
, stmt
, 0,
1058 TDF_VOPS
| TDF_MEMSYMS
));
1065 /* These nodes cut a new scope. */
1067 dp
<< "[scop-detection-fail] "
1068 << "Gimple stmt not handled in Graphite:\n";
1069 print_gimple_stmt (dump_file
, stmt
, 0, TDF_VOPS
| TDF_MEMSYMS
));
1074 /* Returns the number of pbbs that are in loops contained in SCOP. */
1077 scop_detection::nb_pbbs_in_loops (scop_p scop
)
1083 FOR_EACH_VEC_ELT (scop
->pbbs
, i
, pbb
)
1084 if (loop_in_sese_p (gbb_loop (PBB_BLACK_BOX (pbb
)), scop
->scop_info
->region
))
1090 /* Assigns the parameter NAME an index in REGION. */
1093 assign_parameter_index_in_region (tree name
, sese_info_p region
)
1095 gcc_assert (TREE_CODE (name
) == SSA_NAME
1096 && INTEGRAL_TYPE_P (TREE_TYPE (name
))
1097 && ! defined_in_sese_p (name
, region
->region
));
1101 FOR_EACH_VEC_ELT (region
->params
, i
, p
)
1105 i
= region
->params
.length ();
1106 region
->params
.safe_push (name
);
1109 /* In the context of sese S, scan the expression E and translate it to
1110 a linear expression C. When parsing a symbolic multiplication, K
1111 represents the constant multiplier of an expression containing
1115 scan_tree_for_params (sese_info_p s
, tree e
)
1117 if (e
== chrec_dont_know
)
1120 switch (TREE_CODE (e
))
1122 case POLYNOMIAL_CHREC
:
1123 scan_tree_for_params (s
, CHREC_LEFT (e
));
1127 if (chrec_contains_symbols (TREE_OPERAND (e
, 0)))
1128 scan_tree_for_params (s
, TREE_OPERAND (e
, 0));
1130 scan_tree_for_params (s
, TREE_OPERAND (e
, 1));
1134 case POINTER_PLUS_EXPR
:
1136 scan_tree_for_params (s
, TREE_OPERAND (e
, 0));
1137 scan_tree_for_params (s
, TREE_OPERAND (e
, 1));
1143 case NON_LVALUE_EXPR
:
1144 scan_tree_for_params (s
, TREE_OPERAND (e
, 0));
1148 assign_parameter_index_in_region (e
, s
);
1164 /* Find parameters with respect to REGION in BB. We are looking in memory
1165 access functions, conditions and loop bounds. */
1168 find_params_in_bb (sese_info_p region
, gimple_poly_bb_p gbb
)
1170 /* Find parameters in the access functions of data references. */
1172 data_reference_p dr
;
1173 FOR_EACH_VEC_ELT (GBB_DATA_REFS (gbb
), i
, dr
)
1174 for (unsigned j
= 0; j
< DR_NUM_DIMENSIONS (dr
); j
++)
1175 scan_tree_for_params (region
, DR_ACCESS_FN (dr
, j
));
1177 /* Find parameters in conditional statements. */
1179 FOR_EACH_VEC_ELT (GBB_CONDITIONS (gbb
), i
, stmt
)
1181 loop_p loop
= gimple_bb (stmt
)->loop_father
;
1182 tree lhs
= scalar_evolution_in_region (region
->region
, loop
,
1183 gimple_cond_lhs (stmt
));
1184 tree rhs
= scalar_evolution_in_region (region
->region
, loop
,
1185 gimple_cond_rhs (stmt
));
1186 gcc_assert (!chrec_contains_undetermined (lhs
)
1187 && !chrec_contains_undetermined (rhs
));
1189 scan_tree_for_params (region
, lhs
);
1190 scan_tree_for_params (region
, rhs
);
1194 /* Record the parameters used in the SCOP BBs. A variable is a parameter
1195 in a scop if it does not vary during the execution of that scop. */
1198 find_scop_parameters (scop_p scop
)
1201 sese_info_p region
= scop
->scop_info
;
1203 /* Parameters used in loop bounds are processed during gather_bbs. */
1205 /* Find the parameters used in data accesses. */
1207 FOR_EACH_VEC_ELT (scop
->pbbs
, i
, pbb
)
1208 find_params_in_bb (region
, PBB_BLACK_BOX (pbb
));
1210 int nbp
= sese_nb_params (region
);
1211 scop_set_nb_params (scop
, nbp
);
1215 add_write (vec
<tree
> *writes
, tree def
)
1217 writes
->safe_push (def
);
1218 DEBUG_PRINT (dp
<< "Adding scalar write: ";
1219 print_generic_expr (dump_file
, def
);
1220 dp
<< "\nFrom stmt: ";
1221 print_gimple_stmt (dump_file
,
1222 SSA_NAME_DEF_STMT (def
), 0));
1226 add_read (vec
<scalar_use
> *reads
, tree use
, gimple
*use_stmt
)
1228 DEBUG_PRINT (dp
<< "Adding scalar read: ";
1229 print_generic_expr (dump_file
, use
);
1230 dp
<< "\nFrom stmt: ";
1231 print_gimple_stmt (dump_file
, use_stmt
, 0));
1232 reads
->safe_push (std::make_pair (use_stmt
, use
));
1236 /* Record DEF if it is used in other bbs different than DEF_BB in the SCOP. */
1239 build_cross_bb_scalars_def (scop_p scop
, tree def
, basic_block def_bb
,
1242 if (!is_gimple_reg (def
))
1245 bool scev_analyzable
= scev_analyzable_p (def
, scop
->scop_info
->region
);
1248 imm_use_iterator imm_iter
;
1249 FOR_EACH_IMM_USE_STMT (use_stmt
, imm_iter
, def
)
1250 /* Do not gather scalar variables that can be analyzed by SCEV as they can
1251 be generated out of the induction variables. */
1252 if ((! scev_analyzable
1253 /* But gather SESE liveouts as we otherwise fail to rewrite their
1255 || ! bb_in_sese_p (gimple_bb (use_stmt
), scop
->scop_info
->region
))
1256 && (def_bb
!= gimple_bb (use_stmt
) && !is_gimple_debug (use_stmt
)))
1258 add_write (writes
, def
);
1259 /* This is required by the FOR_EACH_IMM_USE_STMT when we want to break
1260 before all the uses have been visited. */
1261 BREAK_FROM_IMM_USE_STMT (imm_iter
);
1265 /* Record USE if it is defined in other bbs different than USE_STMT
1269 build_cross_bb_scalars_use (scop_p scop
, tree use
, gimple
*use_stmt
,
1270 vec
<scalar_use
> *reads
)
1272 if (!is_gimple_reg (use
))
1275 /* Do not gather scalar variables that can be analyzed by SCEV as they can be
1276 generated out of the induction variables. */
1277 if (scev_analyzable_p (use
, scop
->scop_info
->region
))
1280 gimple
*def_stmt
= SSA_NAME_DEF_STMT (use
);
1281 if (gimple_bb (def_stmt
) != gimple_bb (use_stmt
))
1282 add_read (reads
, use
, use_stmt
);
1285 /* Generates a polyhedral black box only if the bb contains interesting
1288 static gimple_poly_bb_p
1289 try_generate_gimple_bb (scop_p scop
, basic_block bb
)
1291 vec
<data_reference_p
> drs
= vNULL
;
1292 vec
<tree
> writes
= vNULL
;
1293 vec
<scalar_use
> reads
= vNULL
;
1295 sese_l region
= scop
->scop_info
->region
;
1296 edge nest
= region
.entry
;
1297 loop_p loop
= bb
->loop_father
;
1298 if (!loop_in_sese_p (loop
, region
))
1301 for (gimple_stmt_iterator gsi
= gsi_start_bb (bb
); !gsi_end_p (gsi
);
1304 gimple
*stmt
= gsi_stmt (gsi
);
1305 if (is_gimple_debug (stmt
))
1308 graphite_find_data_references_in_stmt (nest
, loop
, stmt
, &drs
);
1310 tree def
= gimple_get_lhs (stmt
);
1312 build_cross_bb_scalars_def (scop
, def
, gimple_bb (stmt
), &writes
);
1316 FOR_EACH_SSA_TREE_OPERAND (use
, stmt
, iter
, SSA_OP_USE
)
1317 build_cross_bb_scalars_use (scop
, use
, stmt
, &reads
);
1320 /* Handle defs and uses in PHIs. Those need special treatment given
1321 that we have to present ISL with sth that looks like we've rewritten
1322 the IL out-of-SSA. */
1323 for (gphi_iterator psi
= gsi_start_phis (bb
); !gsi_end_p (psi
);
1326 gphi
*phi
= psi
.phi ();
1327 tree res
= gimple_phi_result (phi
);
1328 if (virtual_operand_p (res
)
1329 || scev_analyzable_p (res
, scop
->scop_info
->region
))
1331 /* To simulate out-of-SSA the block containing the PHI node has
1332 reads of the PHI destination. And to preserve SSA dependences
1333 we also write to it (the out-of-SSA decl and the SSA result
1334 are coalesced for dependence purposes which is good enough). */
1335 add_read (&reads
, res
, phi
);
1336 add_write (&writes
, res
);
1338 basic_block bb_for_succs
= bb
;
1339 if (bb_for_succs
== bb_for_succs
->loop_father
->latch
1340 && bb_in_sese_p (bb_for_succs
, scop
->scop_info
->region
)
1341 && sese_trivially_empty_bb_p (bb_for_succs
))
1342 bb_for_succs
= NULL
;
1343 while (bb_for_succs
)
1345 basic_block latch
= NULL
;
1348 FOR_EACH_EDGE (e
, ei
, bb_for_succs
->succs
)
1350 for (gphi_iterator psi
= gsi_start_phis (e
->dest
); !gsi_end_p (psi
);
1353 gphi
*phi
= psi
.phi ();
1354 tree res
= gimple_phi_result (phi
);
1355 if (virtual_operand_p (res
))
1357 /* To simulate out-of-SSA the predecessor of edges into PHI nodes
1358 has a copy from the PHI argument to the PHI destination. */
1359 if (! scev_analyzable_p (res
, scop
->scop_info
->region
))
1360 add_write (&writes
, res
);
1361 tree use
= PHI_ARG_DEF_FROM_EDGE (phi
, e
);
1362 if (TREE_CODE (use
) == SSA_NAME
1363 && ! SSA_NAME_IS_DEFAULT_DEF (use
)
1364 && gimple_bb (SSA_NAME_DEF_STMT (use
)) != bb_for_succs
1365 && ! scev_analyzable_p (use
, scop
->scop_info
->region
))
1366 add_read (&reads
, use
, phi
);
1368 if (e
->dest
== bb_for_succs
->loop_father
->latch
1369 && bb_in_sese_p (e
->dest
, scop
->scop_info
->region
)
1370 && sese_trivially_empty_bb_p (e
->dest
))
1373 /* Handle empty latch block PHIs here, otherwise we confuse ISL
1374 with extra conditional code where it then peels off the last
1375 iteration just because of that. It would be simplest if we
1376 just didn't force simple latches (thus remove the forwarder). */
1377 bb_for_succs
= latch
;
1380 /* For the region exit block add reads for all live-out vars. */
1381 if (bb
== scop
->scop_info
->region
.exit
->src
)
1383 sese_build_liveouts (scop
->scop_info
);
1386 EXECUTE_IF_SET_IN_BITMAP (scop
->scop_info
->liveout
, 0, i
, bi
)
1388 tree use
= ssa_name (i
);
1389 add_read (&reads
, use
, NULL
);
1393 if (drs
.is_empty () && writes
.is_empty () && reads
.is_empty ())
1396 return new_gimple_poly_bb (bb
, drs
, reads
, writes
);
1399 /* Compute alias-sets for all data references in DRS. */
1402 build_alias_set (scop_p scop
)
1404 int num_vertices
= scop
->drs
.length ();
1405 struct graph
*g
= new_graph (num_vertices
);
1410 FOR_EACH_VEC_ELT (scop
->drs
, i
, dr1
)
1411 for (j
= i
+1; scop
->drs
.iterate (j
, &dr2
); j
++)
1412 if (dr_may_alias_p (dr1
->dr
, dr2
->dr
, true))
1414 /* Dependences in the same alias set need to be handled
1415 by just looking at DR_ACCESS_FNs. */
1416 if (DR_NUM_DIMENSIONS (dr1
->dr
) == 0
1417 || DR_NUM_DIMENSIONS (dr1
->dr
) != DR_NUM_DIMENSIONS (dr2
->dr
)
1418 || ! operand_equal_p (DR_BASE_OBJECT (dr1
->dr
),
1419 DR_BASE_OBJECT (dr2
->dr
),
1421 || ! types_compatible_p (TREE_TYPE (DR_BASE_OBJECT (dr1
->dr
)),
1422 TREE_TYPE (DR_BASE_OBJECT (dr2
->dr
))))
1431 all_vertices
= XNEWVEC (int, num_vertices
);
1432 for (i
= 0; i
< num_vertices
; i
++)
1433 all_vertices
[i
] = i
;
1436 = graphds_dfs (g
, all_vertices
, num_vertices
, NULL
, true, NULL
) + 1;
1437 free (all_vertices
);
1439 for (i
= 0; i
< g
->n_vertices
; i
++)
1440 scop
->drs
[i
].alias_set
= g
->vertices
[i
].component
+ 1;
1446 /* Gather BBs and conditions for a SCOP. */
1447 class gather_bbs
: public dom_walker
1450 gather_bbs (cdi_direction
, scop_p
, int *);
1452 virtual edge
before_dom_children (basic_block
);
1453 virtual void after_dom_children (basic_block
);
1456 auto_vec
<gimple
*, 3> conditions
, cases
;
1460 gather_bbs::gather_bbs (cdi_direction direction
, scop_p scop
, int *bb_to_rpo
)
1461 : dom_walker (direction
, ALL_BLOCKS
, bb_to_rpo
), scop (scop
)
1465 /* Call-back for dom_walk executed before visiting the dominated
1469 gather_bbs::before_dom_children (basic_block bb
)
1471 sese_info_p region
= scop
->scop_info
;
1472 if (!bb_in_sese_p (bb
, region
->region
))
1473 return dom_walker::STOP
;
1475 /* For loops fully contained in the region record parameters in the
1477 loop_p loop
= bb
->loop_father
;
1478 if (loop
->header
== bb
1479 && loop_in_sese_p (loop
, region
->region
))
1481 tree nb_iters
= number_of_latch_executions (loop
);
1482 if (chrec_contains_symbols (nb_iters
))
1484 nb_iters
= scalar_evolution_in_region (region
->region
,
1486 scan_tree_for_params (region
, nb_iters
);
1490 if (gcond
*stmt
= single_pred_cond_non_loop_exit (bb
))
1492 edge e
= single_pred_edge (bb
);
1493 /* Make sure the condition is in the region and thus was verified
1495 if (e
!= region
->region
.entry
)
1497 conditions
.safe_push (stmt
);
1498 if (e
->flags
& EDGE_TRUE_VALUE
)
1499 cases
.safe_push (stmt
);
1501 cases
.safe_push (NULL
);
1505 scop
->scop_info
->bbs
.safe_push (bb
);
1507 gimple_poly_bb_p gbb
= try_generate_gimple_bb (scop
, bb
);
1511 GBB_CONDITIONS (gbb
) = conditions
.copy ();
1512 GBB_CONDITION_CASES (gbb
) = cases
.copy ();
1514 poly_bb_p pbb
= new_poly_bb (scop
, gbb
);
1515 scop
->pbbs
.safe_push (pbb
);
1518 data_reference_p dr
;
1519 FOR_EACH_VEC_ELT (gbb
->data_refs
, i
, dr
)
1521 DEBUG_PRINT (dp
<< "Adding memory ";
1526 print_generic_expr (dump_file
, dr
->ref
);
1527 dp
<< "\nFrom stmt: ";
1528 print_gimple_stmt (dump_file
, dr
->stmt
, 0));
1530 scop
->drs
.safe_push (dr_info (dr
, pbb
));
1536 /* Call-back for dom_walk executed after visiting the dominated
1540 gather_bbs::after_dom_children (basic_block bb
)
1542 if (!bb_in_sese_p (bb
, scop
->scop_info
->region
))
1545 if (single_pred_cond_non_loop_exit (bb
))
1547 edge e
= single_pred_edge (bb
);
1548 if (e
!= scop
->scop_info
->region
.entry
)
1557 /* Compute sth like an execution order, dominator order with first executing
1558 edges that stay inside the current loop, delaying processing exit edges. */
1560 static int *bb_to_rpo
;
1562 /* Helper for qsort, sorting after order above. */
1565 cmp_pbbs (const void *pa
, const void *pb
)
1567 poly_bb_p bb1
= *((const poly_bb_p
*)pa
);
1568 poly_bb_p bb2
= *((const poly_bb_p
*)pb
);
1569 if (bb_to_rpo
[bb1
->black_box
->bb
->index
]
1570 < bb_to_rpo
[bb2
->black_box
->bb
->index
])
1572 else if (bb_to_rpo
[bb1
->black_box
->bb
->index
]
1573 > bb_to_rpo
[bb2
->black_box
->bb
->index
])
1579 /* Find Static Control Parts (SCoP) in the current function and pushes
1583 build_scops (vec
<scop_p
> *scops
)
1586 dp
.set_dump_file (dump_file
);
1589 sb
.build_scop_depth (current_loops
->tree_root
);
1591 /* Now create scops from the lightweight SESEs. */
1592 vec
<sese_l
> scops_l
= sb
.get_scops ();
1594 /* Domwalk needs a bb to RPO mapping. Compute it once here. */
1595 int *postorder
= XNEWVEC (int, n_basic_blocks_for_fn (cfun
));
1596 int postorder_num
= pre_and_rev_post_order_compute (NULL
, postorder
, true);
1597 bb_to_rpo
= XNEWVEC (int, last_basic_block_for_fn (cfun
));
1598 for (int i
= 0; i
< postorder_num
; ++i
)
1599 bb_to_rpo
[postorder
[i
]] = i
;
1604 FOR_EACH_VEC_ELT (scops_l
, i
, s
)
1606 scop_p scop
= new_scop (s
->entry
, s
->exit
);
1608 /* Record all basic blocks and their conditions in REGION. */
1609 gather_bbs (CDI_DOMINATORS
, scop
, bb_to_rpo
).walk (s
->entry
->dest
);
1611 /* Sort pbbs after execution order for initial schedule generation. */
1612 scop
->pbbs
.qsort (cmp_pbbs
);
1614 if (! build_alias_set (scop
))
1616 DEBUG_PRINT (dp
<< "[scop-detection-fail] cannot handle dependences\n");
1621 /* Do not optimize a scop containing only PBBs that do not belong
1623 if (sb
.nb_pbbs_in_loops (scop
) == 0)
1625 DEBUG_PRINT (dp
<< "[scop-detection-fail] no data references.\n");
1630 unsigned max_arrays
= PARAM_VALUE (PARAM_GRAPHITE_MAX_ARRAYS_PER_SCOP
);
1632 && scop
->drs
.length () >= max_arrays
)
1634 DEBUG_PRINT (dp
<< "[scop-detection-fail] too many data references: "
1635 << scop
->drs
.length ()
1636 << " is larger than --param graphite-max-arrays-per-scop="
1637 << max_arrays
<< ".\n");
1642 find_scop_parameters (scop
);
1643 graphite_dim_t max_dim
= PARAM_VALUE (PARAM_GRAPHITE_MAX_NB_SCOP_PARAMS
);
1645 && scop_nb_params (scop
) > max_dim
)
1647 DEBUG_PRINT (dp
<< "[scop-detection-fail] too many parameters: "
1648 << scop_nb_params (scop
)
1649 << " larger than --param graphite-max-nb-scop-params="
1650 << max_dim
<< ".\n");
1655 scops
->safe_push (scop
);
1660 DEBUG_PRINT (dp
<< "number of SCoPs: " << (scops
? scops
->length () : 0););
1663 #endif /* HAVE_isl */