2010-12-20 Tobias Burnus <burnus@net-b.de>
[official-gcc.git] / gcc / tree-data-ref.c
bloba6d073190009f74d30bc0d7e98e70b2dbc366114
1 /* Data references and dependences detectors.
2 Copyright (C) 2003, 2004, 2005, 2006, 2007, 2008, 2009, 2010
3 Free Software Foundation, Inc.
4 Contributed by Sebastian Pop <pop@cri.ensmp.fr>
6 This file is part of GCC.
8 GCC is free software; you can redistribute it and/or modify it under
9 the terms of the GNU General Public License as published by the Free
10 Software Foundation; either version 3, or (at your option) any later
11 version.
13 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
14 WARRANTY; without even the implied warranty of MERCHANTABILITY or
15 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
16 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 /* This pass walks a given loop structure searching for array
23 references. The information about the array accesses is recorded
24 in DATA_REFERENCE structures.
26 The basic test for determining the dependences is:
27 given two access functions chrec1 and chrec2 to a same array, and
28 x and y two vectors from the iteration domain, the same element of
29 the array is accessed twice at iterations x and y if and only if:
30 | chrec1 (x) == chrec2 (y).
32 The goals of this analysis are:
34 - to determine the independence: the relation between two
35 independent accesses is qualified with the chrec_known (this
36 information allows a loop parallelization),
38 - when two data references access the same data, to qualify the
39 dependence relation with classic dependence representations:
41 - distance vectors
42 - direction vectors
43 - loop carried level dependence
44 - polyhedron dependence
45 or with the chains of recurrences based representation,
47 - to define a knowledge base for storing the data dependence
48 information,
50 - to define an interface to access this data.
53 Definitions:
55 - subscript: given two array accesses a subscript is the tuple
56 composed of the access functions for a given dimension. Example:
57 Given A[f1][f2][f3] and B[g1][g2][g3], there are three subscripts:
58 (f1, g1), (f2, g2), (f3, g3).
60 - Diophantine equation: an equation whose coefficients and
61 solutions are integer constants, for example the equation
62 | 3*x + 2*y = 1
63 has an integer solution x = 1 and y = -1.
65 References:
67 - "Advanced Compilation for High Performance Computing" by Randy
68 Allen and Ken Kennedy.
69 http://citeseer.ist.psu.edu/goff91practical.html
71 - "Loop Transformations for Restructuring Compilers - The Foundations"
72 by Utpal Banerjee.
77 #include "config.h"
78 #include "system.h"
79 #include "coretypes.h"
80 #include "tm.h"
81 #include "ggc.h"
82 #include "flags.h"
83 #include "tree.h"
84 #include "basic-block.h"
85 #include "tree-pretty-print.h"
86 #include "gimple-pretty-print.h"
87 #include "tree-flow.h"
88 #include "tree-dump.h"
89 #include "timevar.h"
90 #include "cfgloop.h"
91 #include "tree-data-ref.h"
92 #include "tree-scalar-evolution.h"
93 #include "tree-pass.h"
94 #include "langhooks.h"
96 static struct datadep_stats
98 int num_dependence_tests;
99 int num_dependence_dependent;
100 int num_dependence_independent;
101 int num_dependence_undetermined;
103 int num_subscript_tests;
104 int num_subscript_undetermined;
105 int num_same_subscript_function;
107 int num_ziv;
108 int num_ziv_independent;
109 int num_ziv_dependent;
110 int num_ziv_unimplemented;
112 int num_siv;
113 int num_siv_independent;
114 int num_siv_dependent;
115 int num_siv_unimplemented;
117 int num_miv;
118 int num_miv_independent;
119 int num_miv_dependent;
120 int num_miv_unimplemented;
121 } dependence_stats;
123 static bool subscript_dependence_tester_1 (struct data_dependence_relation *,
124 struct data_reference *,
125 struct data_reference *,
126 struct loop *);
127 /* Returns true iff A divides B. */
129 static inline bool
130 tree_fold_divides_p (const_tree a, const_tree b)
132 gcc_assert (TREE_CODE (a) == INTEGER_CST);
133 gcc_assert (TREE_CODE (b) == INTEGER_CST);
134 return integer_zerop (int_const_binop (TRUNC_MOD_EXPR, b, a, 0));
137 /* Returns true iff A divides B. */
139 static inline bool
140 int_divides_p (int a, int b)
142 return ((b % a) == 0);
147 /* Dump into FILE all the data references from DATAREFS. */
149 void
150 dump_data_references (FILE *file, VEC (data_reference_p, heap) *datarefs)
152 unsigned int i;
153 struct data_reference *dr;
155 FOR_EACH_VEC_ELT (data_reference_p, datarefs, i, dr)
156 dump_data_reference (file, dr);
159 /* Dump into STDERR all the data references from DATAREFS. */
161 DEBUG_FUNCTION void
162 debug_data_references (VEC (data_reference_p, heap) *datarefs)
164 dump_data_references (stderr, datarefs);
167 /* Dump to STDERR all the dependence relations from DDRS. */
169 DEBUG_FUNCTION void
170 debug_data_dependence_relations (VEC (ddr_p, heap) *ddrs)
172 dump_data_dependence_relations (stderr, ddrs);
175 /* Dump into FILE all the dependence relations from DDRS. */
177 void
178 dump_data_dependence_relations (FILE *file,
179 VEC (ddr_p, heap) *ddrs)
181 unsigned int i;
182 struct data_dependence_relation *ddr;
184 FOR_EACH_VEC_ELT (ddr_p, ddrs, i, ddr)
185 dump_data_dependence_relation (file, ddr);
188 /* Print to STDERR the data_reference DR. */
190 DEBUG_FUNCTION void
191 debug_data_reference (struct data_reference *dr)
193 dump_data_reference (stderr, dr);
196 /* Dump function for a DATA_REFERENCE structure. */
198 void
199 dump_data_reference (FILE *outf,
200 struct data_reference *dr)
202 unsigned int i;
204 fprintf (outf, "#(Data Ref: \n# stmt: ");
205 print_gimple_stmt (outf, DR_STMT (dr), 0, 0);
206 fprintf (outf, "# ref: ");
207 print_generic_stmt (outf, DR_REF (dr), 0);
208 fprintf (outf, "# base_object: ");
209 print_generic_stmt (outf, DR_BASE_OBJECT (dr), 0);
211 for (i = 0; i < DR_NUM_DIMENSIONS (dr); i++)
213 fprintf (outf, "# Access function %d: ", i);
214 print_generic_stmt (outf, DR_ACCESS_FN (dr, i), 0);
216 fprintf (outf, "#)\n");
219 /* Dumps the affine function described by FN to the file OUTF. */
221 static void
222 dump_affine_function (FILE *outf, affine_fn fn)
224 unsigned i;
225 tree coef;
227 print_generic_expr (outf, VEC_index (tree, fn, 0), TDF_SLIM);
228 for (i = 1; VEC_iterate (tree, fn, i, coef); i++)
230 fprintf (outf, " + ");
231 print_generic_expr (outf, coef, TDF_SLIM);
232 fprintf (outf, " * x_%u", i);
236 /* Dumps the conflict function CF to the file OUTF. */
238 static void
239 dump_conflict_function (FILE *outf, conflict_function *cf)
241 unsigned i;
243 if (cf->n == NO_DEPENDENCE)
244 fprintf (outf, "no dependence\n");
245 else if (cf->n == NOT_KNOWN)
246 fprintf (outf, "not known\n");
247 else
249 for (i = 0; i < cf->n; i++)
251 fprintf (outf, "[");
252 dump_affine_function (outf, cf->fns[i]);
253 fprintf (outf, "]\n");
258 /* Dump function for a SUBSCRIPT structure. */
260 void
261 dump_subscript (FILE *outf, struct subscript *subscript)
263 conflict_function *cf = SUB_CONFLICTS_IN_A (subscript);
265 fprintf (outf, "\n (subscript \n");
266 fprintf (outf, " iterations_that_access_an_element_twice_in_A: ");
267 dump_conflict_function (outf, cf);
268 if (CF_NONTRIVIAL_P (cf))
270 tree last_iteration = SUB_LAST_CONFLICT (subscript);
271 fprintf (outf, " last_conflict: ");
272 print_generic_stmt (outf, last_iteration, 0);
275 cf = SUB_CONFLICTS_IN_B (subscript);
276 fprintf (outf, " iterations_that_access_an_element_twice_in_B: ");
277 dump_conflict_function (outf, cf);
278 if (CF_NONTRIVIAL_P (cf))
280 tree last_iteration = SUB_LAST_CONFLICT (subscript);
281 fprintf (outf, " last_conflict: ");
282 print_generic_stmt (outf, last_iteration, 0);
285 fprintf (outf, " (Subscript distance: ");
286 print_generic_stmt (outf, SUB_DISTANCE (subscript), 0);
287 fprintf (outf, " )\n");
288 fprintf (outf, " )\n");
291 /* Print the classic direction vector DIRV to OUTF. */
293 void
294 print_direction_vector (FILE *outf,
295 lambda_vector dirv,
296 int length)
298 int eq;
300 for (eq = 0; eq < length; eq++)
302 enum data_dependence_direction dir = ((enum data_dependence_direction)
303 dirv[eq]);
305 switch (dir)
307 case dir_positive:
308 fprintf (outf, " +");
309 break;
310 case dir_negative:
311 fprintf (outf, " -");
312 break;
313 case dir_equal:
314 fprintf (outf, " =");
315 break;
316 case dir_positive_or_equal:
317 fprintf (outf, " +=");
318 break;
319 case dir_positive_or_negative:
320 fprintf (outf, " +-");
321 break;
322 case dir_negative_or_equal:
323 fprintf (outf, " -=");
324 break;
325 case dir_star:
326 fprintf (outf, " *");
327 break;
328 default:
329 fprintf (outf, "indep");
330 break;
333 fprintf (outf, "\n");
336 /* Print a vector of direction vectors. */
338 void
339 print_dir_vectors (FILE *outf, VEC (lambda_vector, heap) *dir_vects,
340 int length)
342 unsigned j;
343 lambda_vector v;
345 FOR_EACH_VEC_ELT (lambda_vector, dir_vects, j, v)
346 print_direction_vector (outf, v, length);
349 /* Print a vector of distance vectors. */
351 void
352 print_dist_vectors (FILE *outf, VEC (lambda_vector, heap) *dist_vects,
353 int length)
355 unsigned j;
356 lambda_vector v;
358 FOR_EACH_VEC_ELT (lambda_vector, dist_vects, j, v)
359 print_lambda_vector (outf, v, length);
362 /* Debug version. */
364 DEBUG_FUNCTION void
365 debug_data_dependence_relation (struct data_dependence_relation *ddr)
367 dump_data_dependence_relation (stderr, ddr);
370 /* Dump function for a DATA_DEPENDENCE_RELATION structure. */
372 void
373 dump_data_dependence_relation (FILE *outf,
374 struct data_dependence_relation *ddr)
376 struct data_reference *dra, *drb;
378 fprintf (outf, "(Data Dep: \n");
380 if (!ddr || DDR_ARE_DEPENDENT (ddr) == chrec_dont_know)
382 if (ddr)
384 dra = DDR_A (ddr);
385 drb = DDR_B (ddr);
386 if (dra)
387 dump_data_reference (outf, dra);
388 else
389 fprintf (outf, " (nil)\n");
390 if (drb)
391 dump_data_reference (outf, drb);
392 else
393 fprintf (outf, " (nil)\n");
395 fprintf (outf, " (don't know)\n)\n");
396 return;
399 dra = DDR_A (ddr);
400 drb = DDR_B (ddr);
401 dump_data_reference (outf, dra);
402 dump_data_reference (outf, drb);
404 if (DDR_ARE_DEPENDENT (ddr) == chrec_known)
405 fprintf (outf, " (no dependence)\n");
407 else if (DDR_ARE_DEPENDENT (ddr) == NULL_TREE)
409 unsigned int i;
410 struct loop *loopi;
412 for (i = 0; i < DDR_NUM_SUBSCRIPTS (ddr); i++)
414 fprintf (outf, " access_fn_A: ");
415 print_generic_stmt (outf, DR_ACCESS_FN (dra, i), 0);
416 fprintf (outf, " access_fn_B: ");
417 print_generic_stmt (outf, DR_ACCESS_FN (drb, i), 0);
418 dump_subscript (outf, DDR_SUBSCRIPT (ddr, i));
421 fprintf (outf, " inner loop index: %d\n", DDR_INNER_LOOP (ddr));
422 fprintf (outf, " loop nest: (");
423 FOR_EACH_VEC_ELT (loop_p, DDR_LOOP_NEST (ddr), i, loopi)
424 fprintf (outf, "%d ", loopi->num);
425 fprintf (outf, ")\n");
427 for (i = 0; i < DDR_NUM_DIST_VECTS (ddr); i++)
429 fprintf (outf, " distance_vector: ");
430 print_lambda_vector (outf, DDR_DIST_VECT (ddr, i),
431 DDR_NB_LOOPS (ddr));
434 for (i = 0; i < DDR_NUM_DIR_VECTS (ddr); i++)
436 fprintf (outf, " direction_vector: ");
437 print_direction_vector (outf, DDR_DIR_VECT (ddr, i),
438 DDR_NB_LOOPS (ddr));
442 fprintf (outf, ")\n");
445 /* Dump function for a DATA_DEPENDENCE_DIRECTION structure. */
447 void
448 dump_data_dependence_direction (FILE *file,
449 enum data_dependence_direction dir)
451 switch (dir)
453 case dir_positive:
454 fprintf (file, "+");
455 break;
457 case dir_negative:
458 fprintf (file, "-");
459 break;
461 case dir_equal:
462 fprintf (file, "=");
463 break;
465 case dir_positive_or_negative:
466 fprintf (file, "+-");
467 break;
469 case dir_positive_or_equal:
470 fprintf (file, "+=");
471 break;
473 case dir_negative_or_equal:
474 fprintf (file, "-=");
475 break;
477 case dir_star:
478 fprintf (file, "*");
479 break;
481 default:
482 break;
486 /* Dumps the distance and direction vectors in FILE. DDRS contains
487 the dependence relations, and VECT_SIZE is the size of the
488 dependence vectors, or in other words the number of loops in the
489 considered nest. */
491 void
492 dump_dist_dir_vectors (FILE *file, VEC (ddr_p, heap) *ddrs)
494 unsigned int i, j;
495 struct data_dependence_relation *ddr;
496 lambda_vector v;
498 FOR_EACH_VEC_ELT (ddr_p, ddrs, i, ddr)
499 if (DDR_ARE_DEPENDENT (ddr) == NULL_TREE && DDR_AFFINE_P (ddr))
501 FOR_EACH_VEC_ELT (lambda_vector, DDR_DIST_VECTS (ddr), j, v)
503 fprintf (file, "DISTANCE_V (");
504 print_lambda_vector (file, v, DDR_NB_LOOPS (ddr));
505 fprintf (file, ")\n");
508 FOR_EACH_VEC_ELT (lambda_vector, DDR_DIR_VECTS (ddr), j, v)
510 fprintf (file, "DIRECTION_V (");
511 print_direction_vector (file, v, DDR_NB_LOOPS (ddr));
512 fprintf (file, ")\n");
516 fprintf (file, "\n\n");
519 /* Dumps the data dependence relations DDRS in FILE. */
521 void
522 dump_ddrs (FILE *file, VEC (ddr_p, heap) *ddrs)
524 unsigned int i;
525 struct data_dependence_relation *ddr;
527 FOR_EACH_VEC_ELT (ddr_p, ddrs, i, ddr)
528 dump_data_dependence_relation (file, ddr);
530 fprintf (file, "\n\n");
533 /* Helper function for split_constant_offset. Expresses OP0 CODE OP1
534 (the type of the result is TYPE) as VAR + OFF, where OFF is a nonzero
535 constant of type ssizetype, and returns true. If we cannot do this
536 with OFF nonzero, OFF and VAR are set to NULL_TREE instead and false
537 is returned. */
539 static bool
540 split_constant_offset_1 (tree type, tree op0, enum tree_code code, tree op1,
541 tree *var, tree *off)
543 tree var0, var1;
544 tree off0, off1;
545 enum tree_code ocode = code;
547 *var = NULL_TREE;
548 *off = NULL_TREE;
550 switch (code)
552 case INTEGER_CST:
553 *var = build_int_cst (type, 0);
554 *off = fold_convert (ssizetype, op0);
555 return true;
557 case POINTER_PLUS_EXPR:
558 ocode = PLUS_EXPR;
559 /* FALLTHROUGH */
560 case PLUS_EXPR:
561 case MINUS_EXPR:
562 split_constant_offset (op0, &var0, &off0);
563 split_constant_offset (op1, &var1, &off1);
564 *var = fold_build2 (code, type, var0, var1);
565 *off = size_binop (ocode, off0, off1);
566 return true;
568 case MULT_EXPR:
569 if (TREE_CODE (op1) != INTEGER_CST)
570 return false;
572 split_constant_offset (op0, &var0, &off0);
573 *var = fold_build2 (MULT_EXPR, type, var0, op1);
574 *off = size_binop (MULT_EXPR, off0, fold_convert (ssizetype, op1));
575 return true;
577 case ADDR_EXPR:
579 tree base, poffset;
580 HOST_WIDE_INT pbitsize, pbitpos;
581 enum machine_mode pmode;
582 int punsignedp, pvolatilep;
584 op0 = TREE_OPERAND (op0, 0);
585 if (!handled_component_p (op0))
586 return false;
588 base = get_inner_reference (op0, &pbitsize, &pbitpos, &poffset,
589 &pmode, &punsignedp, &pvolatilep, false);
591 if (pbitpos % BITS_PER_UNIT != 0)
592 return false;
593 base = build_fold_addr_expr (base);
594 off0 = ssize_int (pbitpos / BITS_PER_UNIT);
596 if (poffset)
598 split_constant_offset (poffset, &poffset, &off1);
599 off0 = size_binop (PLUS_EXPR, off0, off1);
600 if (POINTER_TYPE_P (TREE_TYPE (base)))
601 base = fold_build2 (POINTER_PLUS_EXPR, TREE_TYPE (base),
602 base, fold_convert (sizetype, poffset));
603 else
604 base = fold_build2 (PLUS_EXPR, TREE_TYPE (base), base,
605 fold_convert (TREE_TYPE (base), poffset));
608 var0 = fold_convert (type, base);
610 /* If variable length types are involved, punt, otherwise casts
611 might be converted into ARRAY_REFs in gimplify_conversion.
612 To compute that ARRAY_REF's element size TYPE_SIZE_UNIT, which
613 possibly no longer appears in current GIMPLE, might resurface.
614 This perhaps could run
615 if (CONVERT_EXPR_P (var0))
617 gimplify_conversion (&var0);
618 // Attempt to fill in any within var0 found ARRAY_REF's
619 // element size from corresponding op embedded ARRAY_REF,
620 // if unsuccessful, just punt.
621 } */
622 while (POINTER_TYPE_P (type))
623 type = TREE_TYPE (type);
624 if (int_size_in_bytes (type) < 0)
625 return false;
627 *var = var0;
628 *off = off0;
629 return true;
632 case SSA_NAME:
634 gimple def_stmt = SSA_NAME_DEF_STMT (op0);
635 enum tree_code subcode;
637 if (gimple_code (def_stmt) != GIMPLE_ASSIGN)
638 return false;
640 var0 = gimple_assign_rhs1 (def_stmt);
641 subcode = gimple_assign_rhs_code (def_stmt);
642 var1 = gimple_assign_rhs2 (def_stmt);
644 return split_constant_offset_1 (type, var0, subcode, var1, var, off);
646 CASE_CONVERT:
648 /* We must not introduce undefined overflow, and we must not change the value.
649 Hence we're okay if the inner type doesn't overflow to start with
650 (pointer or signed), the outer type also is an integer or pointer
651 and the outer precision is at least as large as the inner. */
652 tree itype = TREE_TYPE (op0);
653 if ((POINTER_TYPE_P (itype)
654 || (INTEGRAL_TYPE_P (itype) && TYPE_OVERFLOW_UNDEFINED (itype)))
655 && TYPE_PRECISION (type) >= TYPE_PRECISION (itype)
656 && (POINTER_TYPE_P (type) || INTEGRAL_TYPE_P (type)))
658 split_constant_offset (op0, &var0, off);
659 *var = fold_convert (type, var0);
660 return true;
662 return false;
665 default:
666 return false;
670 /* Expresses EXP as VAR + OFF, where off is a constant. The type of OFF
671 will be ssizetype. */
673 void
674 split_constant_offset (tree exp, tree *var, tree *off)
676 tree type = TREE_TYPE (exp), otype, op0, op1, e, o;
677 enum tree_code code;
679 *var = exp;
680 *off = ssize_int (0);
681 STRIP_NOPS (exp);
683 if (automatically_generated_chrec_p (exp))
684 return;
686 otype = TREE_TYPE (exp);
687 code = TREE_CODE (exp);
688 extract_ops_from_tree (exp, &code, &op0, &op1);
689 if (split_constant_offset_1 (otype, op0, code, op1, &e, &o))
691 *var = fold_convert (type, e);
692 *off = o;
696 /* Returns the address ADDR of an object in a canonical shape (without nop
697 casts, and with type of pointer to the object). */
699 static tree
700 canonicalize_base_object_address (tree addr)
702 tree orig = addr;
704 STRIP_NOPS (addr);
706 /* The base address may be obtained by casting from integer, in that case
707 keep the cast. */
708 if (!POINTER_TYPE_P (TREE_TYPE (addr)))
709 return orig;
711 if (TREE_CODE (addr) != ADDR_EXPR)
712 return addr;
714 return build_fold_addr_expr (TREE_OPERAND (addr, 0));
717 /* Analyzes the behavior of the memory reference DR in the innermost loop or
718 basic block that contains it. Returns true if analysis succeed or false
719 otherwise. */
721 bool
722 dr_analyze_innermost (struct data_reference *dr)
724 gimple stmt = DR_STMT (dr);
725 struct loop *loop = loop_containing_stmt (stmt);
726 tree ref = DR_REF (dr);
727 HOST_WIDE_INT pbitsize, pbitpos;
728 tree base, poffset;
729 enum machine_mode pmode;
730 int punsignedp, pvolatilep;
731 affine_iv base_iv, offset_iv;
732 tree init, dinit, step;
733 bool in_loop = (loop && loop->num);
735 if (dump_file && (dump_flags & TDF_DETAILS))
736 fprintf (dump_file, "analyze_innermost: ");
738 base = get_inner_reference (ref, &pbitsize, &pbitpos, &poffset,
739 &pmode, &punsignedp, &pvolatilep, false);
740 gcc_assert (base != NULL_TREE);
742 if (pbitpos % BITS_PER_UNIT != 0)
744 if (dump_file && (dump_flags & TDF_DETAILS))
745 fprintf (dump_file, "failed: bit offset alignment.\n");
746 return false;
749 if (TREE_CODE (base) == MEM_REF)
751 if (!integer_zerop (TREE_OPERAND (base, 1)))
753 if (!poffset)
755 double_int moff = mem_ref_offset (base);
756 poffset = double_int_to_tree (sizetype, moff);
758 else
759 poffset = size_binop (PLUS_EXPR, poffset, TREE_OPERAND (base, 1));
761 base = TREE_OPERAND (base, 0);
763 else
764 base = build_fold_addr_expr (base);
765 if (in_loop)
767 if (!simple_iv (loop, loop_containing_stmt (stmt), base, &base_iv,
768 false))
770 if (dump_file && (dump_flags & TDF_DETAILS))
771 fprintf (dump_file, "failed: evolution of base is not affine.\n");
772 return false;
775 else
777 base_iv.base = base;
778 base_iv.step = ssize_int (0);
779 base_iv.no_overflow = true;
782 if (!poffset)
784 offset_iv.base = ssize_int (0);
785 offset_iv.step = ssize_int (0);
787 else
789 if (!in_loop)
791 offset_iv.base = poffset;
792 offset_iv.step = ssize_int (0);
794 else if (!simple_iv (loop, loop_containing_stmt (stmt),
795 poffset, &offset_iv, false))
797 if (dump_file && (dump_flags & TDF_DETAILS))
798 fprintf (dump_file, "failed: evolution of offset is not"
799 " affine.\n");
800 return false;
804 init = ssize_int (pbitpos / BITS_PER_UNIT);
805 split_constant_offset (base_iv.base, &base_iv.base, &dinit);
806 init = size_binop (PLUS_EXPR, init, dinit);
807 split_constant_offset (offset_iv.base, &offset_iv.base, &dinit);
808 init = size_binop (PLUS_EXPR, init, dinit);
810 step = size_binop (PLUS_EXPR,
811 fold_convert (ssizetype, base_iv.step),
812 fold_convert (ssizetype, offset_iv.step));
814 DR_BASE_ADDRESS (dr) = canonicalize_base_object_address (base_iv.base);
816 DR_OFFSET (dr) = fold_convert (ssizetype, offset_iv.base);
817 DR_INIT (dr) = init;
818 DR_STEP (dr) = step;
820 DR_ALIGNED_TO (dr) = size_int (highest_pow2_factor (offset_iv.base));
822 if (dump_file && (dump_flags & TDF_DETAILS))
823 fprintf (dump_file, "success.\n");
825 return true;
828 /* Determines the base object and the list of indices of memory reference
829 DR, analyzed in loop nest NEST. */
831 static void
832 dr_analyze_indices (struct data_reference *dr, struct loop *nest)
834 gimple stmt = DR_STMT (dr);
835 struct loop *loop = loop_containing_stmt (stmt);
836 VEC (tree, heap) *access_fns = NULL;
837 tree ref = unshare_expr (DR_REF (dr)), aref = ref, op;
838 tree base, off, access_fn = NULL_TREE;
839 basic_block before_loop = NULL;
841 if (nest)
842 before_loop = block_before_loop (nest);
844 while (handled_component_p (aref))
846 if (TREE_CODE (aref) == ARRAY_REF)
848 op = TREE_OPERAND (aref, 1);
849 if (nest)
851 access_fn = analyze_scalar_evolution (loop, op);
852 access_fn = instantiate_scev (before_loop, loop, access_fn);
853 VEC_safe_push (tree, heap, access_fns, access_fn);
856 TREE_OPERAND (aref, 1) = build_int_cst (TREE_TYPE (op), 0);
859 aref = TREE_OPERAND (aref, 0);
862 if (nest
863 && (INDIRECT_REF_P (aref)
864 || TREE_CODE (aref) == MEM_REF))
866 op = TREE_OPERAND (aref, 0);
867 access_fn = analyze_scalar_evolution (loop, op);
868 access_fn = instantiate_scev (before_loop, loop, access_fn);
869 base = initial_condition (access_fn);
870 split_constant_offset (base, &base, &off);
871 if (TREE_CODE (aref) == MEM_REF)
872 off = size_binop (PLUS_EXPR, off,
873 fold_convert (ssizetype, TREE_OPERAND (aref, 1)));
874 access_fn = chrec_replace_initial_condition (access_fn,
875 fold_convert (TREE_TYPE (base), off));
877 TREE_OPERAND (aref, 0) = base;
878 VEC_safe_push (tree, heap, access_fns, access_fn);
881 if (TREE_CODE (aref) == MEM_REF)
882 TREE_OPERAND (aref, 1)
883 = build_int_cst (TREE_TYPE (TREE_OPERAND (aref, 1)), 0);
885 if (TREE_CODE (ref) == MEM_REF
886 && TREE_CODE (TREE_OPERAND (ref, 0)) == ADDR_EXPR
887 && integer_zerop (TREE_OPERAND (ref, 1)))
888 ref = TREE_OPERAND (TREE_OPERAND (ref, 0), 0);
890 /* For canonicalization purposes we'd like to strip all outermost
891 zero-offset component-refs.
892 ??? For now simply handle zero-index array-refs. */
893 while (TREE_CODE (ref) == ARRAY_REF
894 && integer_zerop (TREE_OPERAND (ref, 1)))
895 ref = TREE_OPERAND (ref, 0);
897 DR_BASE_OBJECT (dr) = ref;
898 DR_ACCESS_FNS (dr) = access_fns;
901 /* Extracts the alias analysis information from the memory reference DR. */
903 static void
904 dr_analyze_alias (struct data_reference *dr)
906 tree ref = DR_REF (dr);
907 tree base = get_base_address (ref), addr;
909 if (INDIRECT_REF_P (base)
910 || TREE_CODE (base) == MEM_REF)
912 addr = TREE_OPERAND (base, 0);
913 if (TREE_CODE (addr) == SSA_NAME)
914 DR_PTR_INFO (dr) = SSA_NAME_PTR_INFO (addr);
918 /* Returns true if the address of DR is invariant. */
920 static bool
921 dr_address_invariant_p (struct data_reference *dr)
923 unsigned i;
924 tree idx;
926 FOR_EACH_VEC_ELT (tree, DR_ACCESS_FNS (dr), i, idx)
927 if (tree_contains_chrecs (idx, NULL))
928 return false;
930 return true;
933 /* Frees data reference DR. */
935 void
936 free_data_ref (data_reference_p dr)
938 VEC_free (tree, heap, DR_ACCESS_FNS (dr));
939 free (dr);
942 /* Analyzes memory reference MEMREF accessed in STMT. The reference
943 is read if IS_READ is true, write otherwise. Returns the
944 data_reference description of MEMREF. NEST is the outermost loop of the
945 loop nest in that the reference should be analyzed. */
947 struct data_reference *
948 create_data_ref (struct loop *nest, tree memref, gimple stmt, bool is_read)
950 struct data_reference *dr;
952 if (dump_file && (dump_flags & TDF_DETAILS))
954 fprintf (dump_file, "Creating dr for ");
955 print_generic_expr (dump_file, memref, TDF_SLIM);
956 fprintf (dump_file, "\n");
959 dr = XCNEW (struct data_reference);
960 DR_STMT (dr) = stmt;
961 DR_REF (dr) = memref;
962 DR_IS_READ (dr) = is_read;
964 dr_analyze_innermost (dr);
965 dr_analyze_indices (dr, nest);
966 dr_analyze_alias (dr);
968 if (dump_file && (dump_flags & TDF_DETAILS))
970 fprintf (dump_file, "\tbase_address: ");
971 print_generic_expr (dump_file, DR_BASE_ADDRESS (dr), TDF_SLIM);
972 fprintf (dump_file, "\n\toffset from base address: ");
973 print_generic_expr (dump_file, DR_OFFSET (dr), TDF_SLIM);
974 fprintf (dump_file, "\n\tconstant offset from base address: ");
975 print_generic_expr (dump_file, DR_INIT (dr), TDF_SLIM);
976 fprintf (dump_file, "\n\tstep: ");
977 print_generic_expr (dump_file, DR_STEP (dr), TDF_SLIM);
978 fprintf (dump_file, "\n\taligned to: ");
979 print_generic_expr (dump_file, DR_ALIGNED_TO (dr), TDF_SLIM);
980 fprintf (dump_file, "\n\tbase_object: ");
981 print_generic_expr (dump_file, DR_BASE_OBJECT (dr), TDF_SLIM);
982 fprintf (dump_file, "\n");
985 return dr;
988 /* Returns true if FNA == FNB. */
990 static bool
991 affine_function_equal_p (affine_fn fna, affine_fn fnb)
993 unsigned i, n = VEC_length (tree, fna);
995 if (n != VEC_length (tree, fnb))
996 return false;
998 for (i = 0; i < n; i++)
999 if (!operand_equal_p (VEC_index (tree, fna, i),
1000 VEC_index (tree, fnb, i), 0))
1001 return false;
1003 return true;
1006 /* If all the functions in CF are the same, returns one of them,
1007 otherwise returns NULL. */
1009 static affine_fn
1010 common_affine_function (conflict_function *cf)
1012 unsigned i;
1013 affine_fn comm;
1015 if (!CF_NONTRIVIAL_P (cf))
1016 return NULL;
1018 comm = cf->fns[0];
1020 for (i = 1; i < cf->n; i++)
1021 if (!affine_function_equal_p (comm, cf->fns[i]))
1022 return NULL;
1024 return comm;
1027 /* Returns the base of the affine function FN. */
1029 static tree
1030 affine_function_base (affine_fn fn)
1032 return VEC_index (tree, fn, 0);
1035 /* Returns true if FN is a constant. */
1037 static bool
1038 affine_function_constant_p (affine_fn fn)
1040 unsigned i;
1041 tree coef;
1043 for (i = 1; VEC_iterate (tree, fn, i, coef); i++)
1044 if (!integer_zerop (coef))
1045 return false;
1047 return true;
1050 /* Returns true if FN is the zero constant function. */
1052 static bool
1053 affine_function_zero_p (affine_fn fn)
1055 return (integer_zerop (affine_function_base (fn))
1056 && affine_function_constant_p (fn));
1059 /* Returns a signed integer type with the largest precision from TA
1060 and TB. */
1062 static tree
1063 signed_type_for_types (tree ta, tree tb)
1065 if (TYPE_PRECISION (ta) > TYPE_PRECISION (tb))
1066 return signed_type_for (ta);
1067 else
1068 return signed_type_for (tb);
1071 /* Applies operation OP on affine functions FNA and FNB, and returns the
1072 result. */
1074 static affine_fn
1075 affine_fn_op (enum tree_code op, affine_fn fna, affine_fn fnb)
1077 unsigned i, n, m;
1078 affine_fn ret;
1079 tree coef;
1081 if (VEC_length (tree, fnb) > VEC_length (tree, fna))
1083 n = VEC_length (tree, fna);
1084 m = VEC_length (tree, fnb);
1086 else
1088 n = VEC_length (tree, fnb);
1089 m = VEC_length (tree, fna);
1092 ret = VEC_alloc (tree, heap, m);
1093 for (i = 0; i < n; i++)
1095 tree type = signed_type_for_types (TREE_TYPE (VEC_index (tree, fna, i)),
1096 TREE_TYPE (VEC_index (tree, fnb, i)));
1098 VEC_quick_push (tree, ret,
1099 fold_build2 (op, type,
1100 VEC_index (tree, fna, i),
1101 VEC_index (tree, fnb, i)));
1104 for (; VEC_iterate (tree, fna, i, coef); i++)
1105 VEC_quick_push (tree, ret,
1106 fold_build2 (op, signed_type_for (TREE_TYPE (coef)),
1107 coef, integer_zero_node));
1108 for (; VEC_iterate (tree, fnb, i, coef); i++)
1109 VEC_quick_push (tree, ret,
1110 fold_build2 (op, signed_type_for (TREE_TYPE (coef)),
1111 integer_zero_node, coef));
1113 return ret;
1116 /* Returns the sum of affine functions FNA and FNB. */
1118 static affine_fn
1119 affine_fn_plus (affine_fn fna, affine_fn fnb)
1121 return affine_fn_op (PLUS_EXPR, fna, fnb);
1124 /* Returns the difference of affine functions FNA and FNB. */
1126 static affine_fn
1127 affine_fn_minus (affine_fn fna, affine_fn fnb)
1129 return affine_fn_op (MINUS_EXPR, fna, fnb);
1132 /* Frees affine function FN. */
1134 static void
1135 affine_fn_free (affine_fn fn)
1137 VEC_free (tree, heap, fn);
1140 /* Determine for each subscript in the data dependence relation DDR
1141 the distance. */
1143 static void
1144 compute_subscript_distance (struct data_dependence_relation *ddr)
1146 conflict_function *cf_a, *cf_b;
1147 affine_fn fn_a, fn_b, diff;
1149 if (DDR_ARE_DEPENDENT (ddr) == NULL_TREE)
1151 unsigned int i;
1153 for (i = 0; i < DDR_NUM_SUBSCRIPTS (ddr); i++)
1155 struct subscript *subscript;
1157 subscript = DDR_SUBSCRIPT (ddr, i);
1158 cf_a = SUB_CONFLICTS_IN_A (subscript);
1159 cf_b = SUB_CONFLICTS_IN_B (subscript);
1161 fn_a = common_affine_function (cf_a);
1162 fn_b = common_affine_function (cf_b);
1163 if (!fn_a || !fn_b)
1165 SUB_DISTANCE (subscript) = chrec_dont_know;
1166 return;
1168 diff = affine_fn_minus (fn_a, fn_b);
1170 if (affine_function_constant_p (diff))
1171 SUB_DISTANCE (subscript) = affine_function_base (diff);
1172 else
1173 SUB_DISTANCE (subscript) = chrec_dont_know;
1175 affine_fn_free (diff);
1180 /* Returns the conflict function for "unknown". */
1182 static conflict_function *
1183 conflict_fn_not_known (void)
1185 conflict_function *fn = XCNEW (conflict_function);
1186 fn->n = NOT_KNOWN;
1188 return fn;
1191 /* Returns the conflict function for "independent". */
1193 static conflict_function *
1194 conflict_fn_no_dependence (void)
1196 conflict_function *fn = XCNEW (conflict_function);
1197 fn->n = NO_DEPENDENCE;
1199 return fn;
1202 /* Returns true if the address of OBJ is invariant in LOOP. */
1204 static bool
1205 object_address_invariant_in_loop_p (const struct loop *loop, const_tree obj)
1207 while (handled_component_p (obj))
1209 if (TREE_CODE (obj) == ARRAY_REF)
1211 /* Index of the ARRAY_REF was zeroed in analyze_indices, thus we only
1212 need to check the stride and the lower bound of the reference. */
1213 if (chrec_contains_symbols_defined_in_loop (TREE_OPERAND (obj, 2),
1214 loop->num)
1215 || chrec_contains_symbols_defined_in_loop (TREE_OPERAND (obj, 3),
1216 loop->num))
1217 return false;
1219 else if (TREE_CODE (obj) == COMPONENT_REF)
1221 if (chrec_contains_symbols_defined_in_loop (TREE_OPERAND (obj, 2),
1222 loop->num))
1223 return false;
1225 obj = TREE_OPERAND (obj, 0);
1228 if (!INDIRECT_REF_P (obj)
1229 && TREE_CODE (obj) != MEM_REF)
1230 return true;
1232 return !chrec_contains_symbols_defined_in_loop (TREE_OPERAND (obj, 0),
1233 loop->num);
1236 /* Returns false if we can prove that data references A and B do not alias,
1237 true otherwise. */
1239 bool
1240 dr_may_alias_p (const struct data_reference *a, const struct data_reference *b)
1242 tree addr_a = DR_BASE_OBJECT (a);
1243 tree addr_b = DR_BASE_OBJECT (b);
1245 if (DR_IS_WRITE (a) && DR_IS_WRITE (b))
1246 return refs_output_dependent_p (addr_a, addr_b);
1247 else if (DR_IS_READ (a) && DR_IS_WRITE (b))
1248 return refs_anti_dependent_p (addr_a, addr_b);
1249 return refs_may_alias_p (addr_a, addr_b);
1252 static void compute_self_dependence (struct data_dependence_relation *);
1254 /* Initialize a data dependence relation between data accesses A and
1255 B. NB_LOOPS is the number of loops surrounding the references: the
1256 size of the classic distance/direction vectors. */
1258 static struct data_dependence_relation *
1259 initialize_data_dependence_relation (struct data_reference *a,
1260 struct data_reference *b,
1261 VEC (loop_p, heap) *loop_nest)
1263 struct data_dependence_relation *res;
1264 unsigned int i;
1266 res = XNEW (struct data_dependence_relation);
1267 DDR_A (res) = a;
1268 DDR_B (res) = b;
1269 DDR_LOOP_NEST (res) = NULL;
1270 DDR_REVERSED_P (res) = false;
1271 DDR_SUBSCRIPTS (res) = NULL;
1272 DDR_DIR_VECTS (res) = NULL;
1273 DDR_DIST_VECTS (res) = NULL;
1275 if (a == NULL || b == NULL)
1277 DDR_ARE_DEPENDENT (res) = chrec_dont_know;
1278 return res;
1281 /* If the data references do not alias, then they are independent. */
1282 if (!dr_may_alias_p (a, b))
1284 DDR_ARE_DEPENDENT (res) = chrec_known;
1285 return res;
1288 /* When the references are exactly the same, don't spend time doing
1289 the data dependence tests, just initialize the ddr and return. */
1290 if (operand_equal_p (DR_REF (a), DR_REF (b), 0))
1292 DDR_AFFINE_P (res) = true;
1293 DDR_ARE_DEPENDENT (res) = NULL_TREE;
1294 DDR_SUBSCRIPTS (res) = VEC_alloc (subscript_p, heap, DR_NUM_DIMENSIONS (a));
1295 DDR_LOOP_NEST (res) = loop_nest;
1296 DDR_INNER_LOOP (res) = 0;
1297 DDR_SELF_REFERENCE (res) = true;
1298 compute_self_dependence (res);
1299 return res;
1302 /* If the references do not access the same object, we do not know
1303 whether they alias or not. */
1304 if (!operand_equal_p (DR_BASE_OBJECT (a), DR_BASE_OBJECT (b), 0))
1306 DDR_ARE_DEPENDENT (res) = chrec_dont_know;
1307 return res;
1310 /* If the base of the object is not invariant in the loop nest, we cannot
1311 analyze it. TODO -- in fact, it would suffice to record that there may
1312 be arbitrary dependences in the loops where the base object varies. */
1313 if (loop_nest
1314 && !object_address_invariant_in_loop_p (VEC_index (loop_p, loop_nest, 0),
1315 DR_BASE_OBJECT (a)))
1317 DDR_ARE_DEPENDENT (res) = chrec_dont_know;
1318 return res;
1321 /* If the number of dimensions of the access to not agree we can have
1322 a pointer access to a component of the array element type and an
1323 array access while the base-objects are still the same. Punt. */
1324 if (DR_NUM_DIMENSIONS (a) != DR_NUM_DIMENSIONS (b))
1326 DDR_ARE_DEPENDENT (res) = chrec_dont_know;
1327 return res;
1330 DDR_AFFINE_P (res) = true;
1331 DDR_ARE_DEPENDENT (res) = NULL_TREE;
1332 DDR_SUBSCRIPTS (res) = VEC_alloc (subscript_p, heap, DR_NUM_DIMENSIONS (a));
1333 DDR_LOOP_NEST (res) = loop_nest;
1334 DDR_INNER_LOOP (res) = 0;
1335 DDR_SELF_REFERENCE (res) = false;
1337 for (i = 0; i < DR_NUM_DIMENSIONS (a); i++)
1339 struct subscript *subscript;
1341 subscript = XNEW (struct subscript);
1342 SUB_CONFLICTS_IN_A (subscript) = conflict_fn_not_known ();
1343 SUB_CONFLICTS_IN_B (subscript) = conflict_fn_not_known ();
1344 SUB_LAST_CONFLICT (subscript) = chrec_dont_know;
1345 SUB_DISTANCE (subscript) = chrec_dont_know;
1346 VEC_safe_push (subscript_p, heap, DDR_SUBSCRIPTS (res), subscript);
1349 return res;
1352 /* Frees memory used by the conflict function F. */
1354 static void
1355 free_conflict_function (conflict_function *f)
1357 unsigned i;
1359 if (CF_NONTRIVIAL_P (f))
1361 for (i = 0; i < f->n; i++)
1362 affine_fn_free (f->fns[i]);
1364 free (f);
1367 /* Frees memory used by SUBSCRIPTS. */
1369 static void
1370 free_subscripts (VEC (subscript_p, heap) *subscripts)
1372 unsigned i;
1373 subscript_p s;
1375 FOR_EACH_VEC_ELT (subscript_p, subscripts, i, s)
1377 free_conflict_function (s->conflicting_iterations_in_a);
1378 free_conflict_function (s->conflicting_iterations_in_b);
1379 free (s);
1381 VEC_free (subscript_p, heap, subscripts);
1384 /* Set DDR_ARE_DEPENDENT to CHREC and finalize the subscript overlap
1385 description. */
1387 static inline void
1388 finalize_ddr_dependent (struct data_dependence_relation *ddr,
1389 tree chrec)
1391 if (dump_file && (dump_flags & TDF_DETAILS))
1393 fprintf (dump_file, "(dependence classified: ");
1394 print_generic_expr (dump_file, chrec, 0);
1395 fprintf (dump_file, ")\n");
1398 DDR_ARE_DEPENDENT (ddr) = chrec;
1399 free_subscripts (DDR_SUBSCRIPTS (ddr));
1400 DDR_SUBSCRIPTS (ddr) = NULL;
1403 /* The dependence relation DDR cannot be represented by a distance
1404 vector. */
1406 static inline void
1407 non_affine_dependence_relation (struct data_dependence_relation *ddr)
1409 if (dump_file && (dump_flags & TDF_DETAILS))
1410 fprintf (dump_file, "(Dependence relation cannot be represented by distance vector.) \n");
1412 DDR_AFFINE_P (ddr) = false;
1417 /* This section contains the classic Banerjee tests. */
1419 /* Returns true iff CHREC_A and CHREC_B are not dependent on any index
1420 variables, i.e., if the ZIV (Zero Index Variable) test is true. */
1422 static inline bool
1423 ziv_subscript_p (const_tree chrec_a, const_tree chrec_b)
1425 return (evolution_function_is_constant_p (chrec_a)
1426 && evolution_function_is_constant_p (chrec_b));
1429 /* Returns true iff CHREC_A and CHREC_B are dependent on an index
1430 variable, i.e., if the SIV (Single Index Variable) test is true. */
1432 static bool
1433 siv_subscript_p (const_tree chrec_a, const_tree chrec_b)
1435 if ((evolution_function_is_constant_p (chrec_a)
1436 && evolution_function_is_univariate_p (chrec_b))
1437 || (evolution_function_is_constant_p (chrec_b)
1438 && evolution_function_is_univariate_p (chrec_a)))
1439 return true;
1441 if (evolution_function_is_univariate_p (chrec_a)
1442 && evolution_function_is_univariate_p (chrec_b))
1444 switch (TREE_CODE (chrec_a))
1446 case POLYNOMIAL_CHREC:
1447 switch (TREE_CODE (chrec_b))
1449 case POLYNOMIAL_CHREC:
1450 if (CHREC_VARIABLE (chrec_a) != CHREC_VARIABLE (chrec_b))
1451 return false;
1453 default:
1454 return true;
1457 default:
1458 return true;
1462 return false;
1465 /* Creates a conflict function with N dimensions. The affine functions
1466 in each dimension follow. */
1468 static conflict_function *
1469 conflict_fn (unsigned n, ...)
1471 unsigned i;
1472 conflict_function *ret = XCNEW (conflict_function);
1473 va_list ap;
1475 gcc_assert (0 < n && n <= MAX_DIM);
1476 va_start(ap, n);
1478 ret->n = n;
1479 for (i = 0; i < n; i++)
1480 ret->fns[i] = va_arg (ap, affine_fn);
1481 va_end(ap);
1483 return ret;
1486 /* Returns constant affine function with value CST. */
1488 static affine_fn
1489 affine_fn_cst (tree cst)
1491 affine_fn fn = VEC_alloc (tree, heap, 1);
1492 VEC_quick_push (tree, fn, cst);
1493 return fn;
1496 /* Returns affine function with single variable, CST + COEF * x_DIM. */
1498 static affine_fn
1499 affine_fn_univar (tree cst, unsigned dim, tree coef)
1501 affine_fn fn = VEC_alloc (tree, heap, dim + 1);
1502 unsigned i;
1504 gcc_assert (dim > 0);
1505 VEC_quick_push (tree, fn, cst);
1506 for (i = 1; i < dim; i++)
1507 VEC_quick_push (tree, fn, integer_zero_node);
1508 VEC_quick_push (tree, fn, coef);
1509 return fn;
1512 /* Analyze a ZIV (Zero Index Variable) subscript. *OVERLAPS_A and
1513 *OVERLAPS_B are initialized to the functions that describe the
1514 relation between the elements accessed twice by CHREC_A and
1515 CHREC_B. For k >= 0, the following property is verified:
1517 CHREC_A (*OVERLAPS_A (k)) = CHREC_B (*OVERLAPS_B (k)). */
1519 static void
1520 analyze_ziv_subscript (tree chrec_a,
1521 tree chrec_b,
1522 conflict_function **overlaps_a,
1523 conflict_function **overlaps_b,
1524 tree *last_conflicts)
1526 tree type, difference;
1527 dependence_stats.num_ziv++;
1529 if (dump_file && (dump_flags & TDF_DETAILS))
1530 fprintf (dump_file, "(analyze_ziv_subscript \n");
1532 type = signed_type_for_types (TREE_TYPE (chrec_a), TREE_TYPE (chrec_b));
1533 chrec_a = chrec_convert (type, chrec_a, NULL);
1534 chrec_b = chrec_convert (type, chrec_b, NULL);
1535 difference = chrec_fold_minus (type, chrec_a, chrec_b);
1537 switch (TREE_CODE (difference))
1539 case INTEGER_CST:
1540 if (integer_zerop (difference))
1542 /* The difference is equal to zero: the accessed index
1543 overlaps for each iteration in the loop. */
1544 *overlaps_a = conflict_fn (1, affine_fn_cst (integer_zero_node));
1545 *overlaps_b = conflict_fn (1, affine_fn_cst (integer_zero_node));
1546 *last_conflicts = chrec_dont_know;
1547 dependence_stats.num_ziv_dependent++;
1549 else
1551 /* The accesses do not overlap. */
1552 *overlaps_a = conflict_fn_no_dependence ();
1553 *overlaps_b = conflict_fn_no_dependence ();
1554 *last_conflicts = integer_zero_node;
1555 dependence_stats.num_ziv_independent++;
1557 break;
1559 default:
1560 /* We're not sure whether the indexes overlap. For the moment,
1561 conservatively answer "don't know". */
1562 if (dump_file && (dump_flags & TDF_DETAILS))
1563 fprintf (dump_file, "ziv test failed: difference is non-integer.\n");
1565 *overlaps_a = conflict_fn_not_known ();
1566 *overlaps_b = conflict_fn_not_known ();
1567 *last_conflicts = chrec_dont_know;
1568 dependence_stats.num_ziv_unimplemented++;
1569 break;
1572 if (dump_file && (dump_flags & TDF_DETAILS))
1573 fprintf (dump_file, ")\n");
1576 /* Sets NIT to the estimated number of executions of the statements in
1577 LOOP. If CONSERVATIVE is true, we must be sure that NIT is at least as
1578 large as the number of iterations. If we have no reliable estimate,
1579 the function returns false, otherwise returns true. */
1581 bool
1582 estimated_loop_iterations (struct loop *loop, bool conservative,
1583 double_int *nit)
1585 estimate_numbers_of_iterations_loop (loop, true);
1586 if (conservative)
1588 if (!loop->any_upper_bound)
1589 return false;
1591 *nit = loop->nb_iterations_upper_bound;
1593 else
1595 if (!loop->any_estimate)
1596 return false;
1598 *nit = loop->nb_iterations_estimate;
1601 return true;
1604 /* Similar to estimated_loop_iterations, but returns the estimate only
1605 if it fits to HOST_WIDE_INT. If this is not the case, or the estimate
1606 on the number of iterations of LOOP could not be derived, returns -1. */
1608 HOST_WIDE_INT
1609 estimated_loop_iterations_int (struct loop *loop, bool conservative)
1611 double_int nit;
1612 HOST_WIDE_INT hwi_nit;
1614 if (!estimated_loop_iterations (loop, conservative, &nit))
1615 return -1;
1617 if (!double_int_fits_in_shwi_p (nit))
1618 return -1;
1619 hwi_nit = double_int_to_shwi (nit);
1621 return hwi_nit < 0 ? -1 : hwi_nit;
1624 /* Similar to estimated_loop_iterations, but returns the estimate as a tree,
1625 and only if it fits to the int type. If this is not the case, or the
1626 estimate on the number of iterations of LOOP could not be derived, returns
1627 chrec_dont_know. */
1629 static tree
1630 estimated_loop_iterations_tree (struct loop *loop, bool conservative)
1632 double_int nit;
1633 tree type;
1635 if (!estimated_loop_iterations (loop, conservative, &nit))
1636 return chrec_dont_know;
1638 type = lang_hooks.types.type_for_size (INT_TYPE_SIZE, true);
1639 if (!double_int_fits_to_tree_p (type, nit))
1640 return chrec_dont_know;
1642 return double_int_to_tree (type, nit);
1645 /* Analyze a SIV (Single Index Variable) subscript where CHREC_A is a
1646 constant, and CHREC_B is an affine function. *OVERLAPS_A and
1647 *OVERLAPS_B are initialized to the functions that describe the
1648 relation between the elements accessed twice by CHREC_A and
1649 CHREC_B. For k >= 0, the following property is verified:
1651 CHREC_A (*OVERLAPS_A (k)) = CHREC_B (*OVERLAPS_B (k)). */
1653 static void
1654 analyze_siv_subscript_cst_affine (tree chrec_a,
1655 tree chrec_b,
1656 conflict_function **overlaps_a,
1657 conflict_function **overlaps_b,
1658 tree *last_conflicts)
1660 bool value0, value1, value2;
1661 tree type, difference, tmp;
1663 type = signed_type_for_types (TREE_TYPE (chrec_a), TREE_TYPE (chrec_b));
1664 chrec_a = chrec_convert (type, chrec_a, NULL);
1665 chrec_b = chrec_convert (type, chrec_b, NULL);
1666 difference = chrec_fold_minus (type, initial_condition (chrec_b), chrec_a);
1668 if (!chrec_is_positive (initial_condition (difference), &value0))
1670 if (dump_file && (dump_flags & TDF_DETAILS))
1671 fprintf (dump_file, "siv test failed: chrec is not positive.\n");
1673 dependence_stats.num_siv_unimplemented++;
1674 *overlaps_a = conflict_fn_not_known ();
1675 *overlaps_b = conflict_fn_not_known ();
1676 *last_conflicts = chrec_dont_know;
1677 return;
1679 else
1681 if (value0 == false)
1683 if (!chrec_is_positive (CHREC_RIGHT (chrec_b), &value1))
1685 if (dump_file && (dump_flags & TDF_DETAILS))
1686 fprintf (dump_file, "siv test failed: chrec not positive.\n");
1688 *overlaps_a = conflict_fn_not_known ();
1689 *overlaps_b = conflict_fn_not_known ();
1690 *last_conflicts = chrec_dont_know;
1691 dependence_stats.num_siv_unimplemented++;
1692 return;
1694 else
1696 if (value1 == true)
1698 /* Example:
1699 chrec_a = 12
1700 chrec_b = {10, +, 1}
1703 if (tree_fold_divides_p (CHREC_RIGHT (chrec_b), difference))
1705 HOST_WIDE_INT numiter;
1706 struct loop *loop = get_chrec_loop (chrec_b);
1708 *overlaps_a = conflict_fn (1, affine_fn_cst (integer_zero_node));
1709 tmp = fold_build2 (EXACT_DIV_EXPR, type,
1710 fold_build1 (ABS_EXPR, type, difference),
1711 CHREC_RIGHT (chrec_b));
1712 *overlaps_b = conflict_fn (1, affine_fn_cst (tmp));
1713 *last_conflicts = integer_one_node;
1716 /* Perform weak-zero siv test to see if overlap is
1717 outside the loop bounds. */
1718 numiter = estimated_loop_iterations_int (loop, false);
1720 if (numiter >= 0
1721 && compare_tree_int (tmp, numiter) > 0)
1723 free_conflict_function (*overlaps_a);
1724 free_conflict_function (*overlaps_b);
1725 *overlaps_a = conflict_fn_no_dependence ();
1726 *overlaps_b = conflict_fn_no_dependence ();
1727 *last_conflicts = integer_zero_node;
1728 dependence_stats.num_siv_independent++;
1729 return;
1731 dependence_stats.num_siv_dependent++;
1732 return;
1735 /* When the step does not divide the difference, there are
1736 no overlaps. */
1737 else
1739 *overlaps_a = conflict_fn_no_dependence ();
1740 *overlaps_b = conflict_fn_no_dependence ();
1741 *last_conflicts = integer_zero_node;
1742 dependence_stats.num_siv_independent++;
1743 return;
1747 else
1749 /* Example:
1750 chrec_a = 12
1751 chrec_b = {10, +, -1}
1753 In this case, chrec_a will not overlap with chrec_b. */
1754 *overlaps_a = conflict_fn_no_dependence ();
1755 *overlaps_b = conflict_fn_no_dependence ();
1756 *last_conflicts = integer_zero_node;
1757 dependence_stats.num_siv_independent++;
1758 return;
1762 else
1764 if (!chrec_is_positive (CHREC_RIGHT (chrec_b), &value2))
1766 if (dump_file && (dump_flags & TDF_DETAILS))
1767 fprintf (dump_file, "siv test failed: chrec not positive.\n");
1769 *overlaps_a = conflict_fn_not_known ();
1770 *overlaps_b = conflict_fn_not_known ();
1771 *last_conflicts = chrec_dont_know;
1772 dependence_stats.num_siv_unimplemented++;
1773 return;
1775 else
1777 if (value2 == false)
1779 /* Example:
1780 chrec_a = 3
1781 chrec_b = {10, +, -1}
1783 if (tree_fold_divides_p (CHREC_RIGHT (chrec_b), difference))
1785 HOST_WIDE_INT numiter;
1786 struct loop *loop = get_chrec_loop (chrec_b);
1788 *overlaps_a = conflict_fn (1, affine_fn_cst (integer_zero_node));
1789 tmp = fold_build2 (EXACT_DIV_EXPR, type, difference,
1790 CHREC_RIGHT (chrec_b));
1791 *overlaps_b = conflict_fn (1, affine_fn_cst (tmp));
1792 *last_conflicts = integer_one_node;
1794 /* Perform weak-zero siv test to see if overlap is
1795 outside the loop bounds. */
1796 numiter = estimated_loop_iterations_int (loop, false);
1798 if (numiter >= 0
1799 && compare_tree_int (tmp, numiter) > 0)
1801 free_conflict_function (*overlaps_a);
1802 free_conflict_function (*overlaps_b);
1803 *overlaps_a = conflict_fn_no_dependence ();
1804 *overlaps_b = conflict_fn_no_dependence ();
1805 *last_conflicts = integer_zero_node;
1806 dependence_stats.num_siv_independent++;
1807 return;
1809 dependence_stats.num_siv_dependent++;
1810 return;
1813 /* When the step does not divide the difference, there
1814 are no overlaps. */
1815 else
1817 *overlaps_a = conflict_fn_no_dependence ();
1818 *overlaps_b = conflict_fn_no_dependence ();
1819 *last_conflicts = integer_zero_node;
1820 dependence_stats.num_siv_independent++;
1821 return;
1824 else
1826 /* Example:
1827 chrec_a = 3
1828 chrec_b = {4, +, 1}
1830 In this case, chrec_a will not overlap with chrec_b. */
1831 *overlaps_a = conflict_fn_no_dependence ();
1832 *overlaps_b = conflict_fn_no_dependence ();
1833 *last_conflicts = integer_zero_node;
1834 dependence_stats.num_siv_independent++;
1835 return;
1842 /* Helper recursive function for initializing the matrix A. Returns
1843 the initial value of CHREC. */
1845 static tree
1846 initialize_matrix_A (lambda_matrix A, tree chrec, unsigned index, int mult)
1848 gcc_assert (chrec);
1850 switch (TREE_CODE (chrec))
1852 case POLYNOMIAL_CHREC:
1853 gcc_assert (TREE_CODE (CHREC_RIGHT (chrec)) == INTEGER_CST);
1855 A[index][0] = mult * int_cst_value (CHREC_RIGHT (chrec));
1856 return initialize_matrix_A (A, CHREC_LEFT (chrec), index + 1, mult);
1858 case PLUS_EXPR:
1859 case MULT_EXPR:
1860 case MINUS_EXPR:
1862 tree op0 = initialize_matrix_A (A, TREE_OPERAND (chrec, 0), index, mult);
1863 tree op1 = initialize_matrix_A (A, TREE_OPERAND (chrec, 1), index, mult);
1865 return chrec_fold_op (TREE_CODE (chrec), chrec_type (chrec), op0, op1);
1868 case NOP_EXPR:
1870 tree op = initialize_matrix_A (A, TREE_OPERAND (chrec, 0), index, mult);
1871 return chrec_convert (chrec_type (chrec), op, NULL);
1874 case BIT_NOT_EXPR:
1876 /* Handle ~X as -1 - X. */
1877 tree op = initialize_matrix_A (A, TREE_OPERAND (chrec, 0), index, mult);
1878 return chrec_fold_op (MINUS_EXPR, chrec_type (chrec),
1879 build_int_cst (TREE_TYPE (chrec), -1), op);
1882 case INTEGER_CST:
1883 return chrec;
1885 default:
1886 gcc_unreachable ();
1887 return NULL_TREE;
1891 #define FLOOR_DIV(x,y) ((x) / (y))
1893 /* Solves the special case of the Diophantine equation:
1894 | {0, +, STEP_A}_x (OVERLAPS_A) = {0, +, STEP_B}_y (OVERLAPS_B)
1896 Computes the descriptions OVERLAPS_A and OVERLAPS_B. NITER is the
1897 number of iterations that loops X and Y run. The overlaps will be
1898 constructed as evolutions in dimension DIM. */
1900 static void
1901 compute_overlap_steps_for_affine_univar (int niter, int step_a, int step_b,
1902 affine_fn *overlaps_a,
1903 affine_fn *overlaps_b,
1904 tree *last_conflicts, int dim)
1906 if (((step_a > 0 && step_b > 0)
1907 || (step_a < 0 && step_b < 0)))
1909 int step_overlaps_a, step_overlaps_b;
1910 int gcd_steps_a_b, last_conflict, tau2;
1912 gcd_steps_a_b = gcd (step_a, step_b);
1913 step_overlaps_a = step_b / gcd_steps_a_b;
1914 step_overlaps_b = step_a / gcd_steps_a_b;
1916 if (niter > 0)
1918 tau2 = FLOOR_DIV (niter, step_overlaps_a);
1919 tau2 = MIN (tau2, FLOOR_DIV (niter, step_overlaps_b));
1920 last_conflict = tau2;
1921 *last_conflicts = build_int_cst (NULL_TREE, last_conflict);
1923 else
1924 *last_conflicts = chrec_dont_know;
1926 *overlaps_a = affine_fn_univar (integer_zero_node, dim,
1927 build_int_cst (NULL_TREE,
1928 step_overlaps_a));
1929 *overlaps_b = affine_fn_univar (integer_zero_node, dim,
1930 build_int_cst (NULL_TREE,
1931 step_overlaps_b));
1934 else
1936 *overlaps_a = affine_fn_cst (integer_zero_node);
1937 *overlaps_b = affine_fn_cst (integer_zero_node);
1938 *last_conflicts = integer_zero_node;
1942 /* Solves the special case of a Diophantine equation where CHREC_A is
1943 an affine bivariate function, and CHREC_B is an affine univariate
1944 function. For example,
1946 | {{0, +, 1}_x, +, 1335}_y = {0, +, 1336}_z
1948 has the following overlapping functions:
1950 | x (t, u, v) = {{0, +, 1336}_t, +, 1}_v
1951 | y (t, u, v) = {{0, +, 1336}_u, +, 1}_v
1952 | z (t, u, v) = {{{0, +, 1}_t, +, 1335}_u, +, 1}_v
1954 FORNOW: This is a specialized implementation for a case occurring in
1955 a common benchmark. Implement the general algorithm. */
1957 static void
1958 compute_overlap_steps_for_affine_1_2 (tree chrec_a, tree chrec_b,
1959 conflict_function **overlaps_a,
1960 conflict_function **overlaps_b,
1961 tree *last_conflicts)
1963 bool xz_p, yz_p, xyz_p;
1964 int step_x, step_y, step_z;
1965 HOST_WIDE_INT niter_x, niter_y, niter_z, niter;
1966 affine_fn overlaps_a_xz, overlaps_b_xz;
1967 affine_fn overlaps_a_yz, overlaps_b_yz;
1968 affine_fn overlaps_a_xyz, overlaps_b_xyz;
1969 affine_fn ova1, ova2, ovb;
1970 tree last_conflicts_xz, last_conflicts_yz, last_conflicts_xyz;
1972 step_x = int_cst_value (CHREC_RIGHT (CHREC_LEFT (chrec_a)));
1973 step_y = int_cst_value (CHREC_RIGHT (chrec_a));
1974 step_z = int_cst_value (CHREC_RIGHT (chrec_b));
1976 niter_x =
1977 estimated_loop_iterations_int (get_chrec_loop (CHREC_LEFT (chrec_a)),
1978 false);
1979 niter_y = estimated_loop_iterations_int (get_chrec_loop (chrec_a), false);
1980 niter_z = estimated_loop_iterations_int (get_chrec_loop (chrec_b), false);
1982 if (niter_x < 0 || niter_y < 0 || niter_z < 0)
1984 if (dump_file && (dump_flags & TDF_DETAILS))
1985 fprintf (dump_file, "overlap steps test failed: no iteration counts.\n");
1987 *overlaps_a = conflict_fn_not_known ();
1988 *overlaps_b = conflict_fn_not_known ();
1989 *last_conflicts = chrec_dont_know;
1990 return;
1993 niter = MIN (niter_x, niter_z);
1994 compute_overlap_steps_for_affine_univar (niter, step_x, step_z,
1995 &overlaps_a_xz,
1996 &overlaps_b_xz,
1997 &last_conflicts_xz, 1);
1998 niter = MIN (niter_y, niter_z);
1999 compute_overlap_steps_for_affine_univar (niter, step_y, step_z,
2000 &overlaps_a_yz,
2001 &overlaps_b_yz,
2002 &last_conflicts_yz, 2);
2003 niter = MIN (niter_x, niter_z);
2004 niter = MIN (niter_y, niter);
2005 compute_overlap_steps_for_affine_univar (niter, step_x + step_y, step_z,
2006 &overlaps_a_xyz,
2007 &overlaps_b_xyz,
2008 &last_conflicts_xyz, 3);
2010 xz_p = !integer_zerop (last_conflicts_xz);
2011 yz_p = !integer_zerop (last_conflicts_yz);
2012 xyz_p = !integer_zerop (last_conflicts_xyz);
2014 if (xz_p || yz_p || xyz_p)
2016 ova1 = affine_fn_cst (integer_zero_node);
2017 ova2 = affine_fn_cst (integer_zero_node);
2018 ovb = affine_fn_cst (integer_zero_node);
2019 if (xz_p)
2021 affine_fn t0 = ova1;
2022 affine_fn t2 = ovb;
2024 ova1 = affine_fn_plus (ova1, overlaps_a_xz);
2025 ovb = affine_fn_plus (ovb, overlaps_b_xz);
2026 affine_fn_free (t0);
2027 affine_fn_free (t2);
2028 *last_conflicts = last_conflicts_xz;
2030 if (yz_p)
2032 affine_fn t0 = ova2;
2033 affine_fn t2 = ovb;
2035 ova2 = affine_fn_plus (ova2, overlaps_a_yz);
2036 ovb = affine_fn_plus (ovb, overlaps_b_yz);
2037 affine_fn_free (t0);
2038 affine_fn_free (t2);
2039 *last_conflicts = last_conflicts_yz;
2041 if (xyz_p)
2043 affine_fn t0 = ova1;
2044 affine_fn t2 = ova2;
2045 affine_fn t4 = ovb;
2047 ova1 = affine_fn_plus (ova1, overlaps_a_xyz);
2048 ova2 = affine_fn_plus (ova2, overlaps_a_xyz);
2049 ovb = affine_fn_plus (ovb, overlaps_b_xyz);
2050 affine_fn_free (t0);
2051 affine_fn_free (t2);
2052 affine_fn_free (t4);
2053 *last_conflicts = last_conflicts_xyz;
2055 *overlaps_a = conflict_fn (2, ova1, ova2);
2056 *overlaps_b = conflict_fn (1, ovb);
2058 else
2060 *overlaps_a = conflict_fn (1, affine_fn_cst (integer_zero_node));
2061 *overlaps_b = conflict_fn (1, affine_fn_cst (integer_zero_node));
2062 *last_conflicts = integer_zero_node;
2065 affine_fn_free (overlaps_a_xz);
2066 affine_fn_free (overlaps_b_xz);
2067 affine_fn_free (overlaps_a_yz);
2068 affine_fn_free (overlaps_b_yz);
2069 affine_fn_free (overlaps_a_xyz);
2070 affine_fn_free (overlaps_b_xyz);
2073 /* Determines the overlapping elements due to accesses CHREC_A and
2074 CHREC_B, that are affine functions. This function cannot handle
2075 symbolic evolution functions, ie. when initial conditions are
2076 parameters, because it uses lambda matrices of integers. */
2078 static void
2079 analyze_subscript_affine_affine (tree chrec_a,
2080 tree chrec_b,
2081 conflict_function **overlaps_a,
2082 conflict_function **overlaps_b,
2083 tree *last_conflicts)
2085 unsigned nb_vars_a, nb_vars_b, dim;
2086 HOST_WIDE_INT init_a, init_b, gamma, gcd_alpha_beta;
2087 lambda_matrix A, U, S;
2088 struct obstack scratch_obstack;
2090 if (eq_evolutions_p (chrec_a, chrec_b))
2092 /* The accessed index overlaps for each iteration in the
2093 loop. */
2094 *overlaps_a = conflict_fn (1, affine_fn_cst (integer_zero_node));
2095 *overlaps_b = conflict_fn (1, affine_fn_cst (integer_zero_node));
2096 *last_conflicts = chrec_dont_know;
2097 return;
2099 if (dump_file && (dump_flags & TDF_DETAILS))
2100 fprintf (dump_file, "(analyze_subscript_affine_affine \n");
2102 /* For determining the initial intersection, we have to solve a
2103 Diophantine equation. This is the most time consuming part.
2105 For answering to the question: "Is there a dependence?" we have
2106 to prove that there exists a solution to the Diophantine
2107 equation, and that the solution is in the iteration domain,
2108 i.e. the solution is positive or zero, and that the solution
2109 happens before the upper bound loop.nb_iterations. Otherwise
2110 there is no dependence. This function outputs a description of
2111 the iterations that hold the intersections. */
2113 nb_vars_a = nb_vars_in_chrec (chrec_a);
2114 nb_vars_b = nb_vars_in_chrec (chrec_b);
2116 gcc_obstack_init (&scratch_obstack);
2118 dim = nb_vars_a + nb_vars_b;
2119 U = lambda_matrix_new (dim, dim, &scratch_obstack);
2120 A = lambda_matrix_new (dim, 1, &scratch_obstack);
2121 S = lambda_matrix_new (dim, 1, &scratch_obstack);
2123 init_a = int_cst_value (initialize_matrix_A (A, chrec_a, 0, 1));
2124 init_b = int_cst_value (initialize_matrix_A (A, chrec_b, nb_vars_a, -1));
2125 gamma = init_b - init_a;
2127 /* Don't do all the hard work of solving the Diophantine equation
2128 when we already know the solution: for example,
2129 | {3, +, 1}_1
2130 | {3, +, 4}_2
2131 | gamma = 3 - 3 = 0.
2132 Then the first overlap occurs during the first iterations:
2133 | {3, +, 1}_1 ({0, +, 4}_x) = {3, +, 4}_2 ({0, +, 1}_x)
2135 if (gamma == 0)
2137 if (nb_vars_a == 1 && nb_vars_b == 1)
2139 HOST_WIDE_INT step_a, step_b;
2140 HOST_WIDE_INT niter, niter_a, niter_b;
2141 affine_fn ova, ovb;
2143 niter_a = estimated_loop_iterations_int (get_chrec_loop (chrec_a),
2144 false);
2145 niter_b = estimated_loop_iterations_int (get_chrec_loop (chrec_b),
2146 false);
2147 niter = MIN (niter_a, niter_b);
2148 step_a = int_cst_value (CHREC_RIGHT (chrec_a));
2149 step_b = int_cst_value (CHREC_RIGHT (chrec_b));
2151 compute_overlap_steps_for_affine_univar (niter, step_a, step_b,
2152 &ova, &ovb,
2153 last_conflicts, 1);
2154 *overlaps_a = conflict_fn (1, ova);
2155 *overlaps_b = conflict_fn (1, ovb);
2158 else if (nb_vars_a == 2 && nb_vars_b == 1)
2159 compute_overlap_steps_for_affine_1_2
2160 (chrec_a, chrec_b, overlaps_a, overlaps_b, last_conflicts);
2162 else if (nb_vars_a == 1 && nb_vars_b == 2)
2163 compute_overlap_steps_for_affine_1_2
2164 (chrec_b, chrec_a, overlaps_b, overlaps_a, last_conflicts);
2166 else
2168 if (dump_file && (dump_flags & TDF_DETAILS))
2169 fprintf (dump_file, "affine-affine test failed: too many variables.\n");
2170 *overlaps_a = conflict_fn_not_known ();
2171 *overlaps_b = conflict_fn_not_known ();
2172 *last_conflicts = chrec_dont_know;
2174 goto end_analyze_subs_aa;
2177 /* U.A = S */
2178 lambda_matrix_right_hermite (A, dim, 1, S, U);
2180 if (S[0][0] < 0)
2182 S[0][0] *= -1;
2183 lambda_matrix_row_negate (U, dim, 0);
2185 gcd_alpha_beta = S[0][0];
2187 /* Something went wrong: for example in {1, +, 0}_5 vs. {0, +, 0}_5,
2188 but that is a quite strange case. Instead of ICEing, answer
2189 don't know. */
2190 if (gcd_alpha_beta == 0)
2192 *overlaps_a = conflict_fn_not_known ();
2193 *overlaps_b = conflict_fn_not_known ();
2194 *last_conflicts = chrec_dont_know;
2195 goto end_analyze_subs_aa;
2198 /* The classic "gcd-test". */
2199 if (!int_divides_p (gcd_alpha_beta, gamma))
2201 /* The "gcd-test" has determined that there is no integer
2202 solution, i.e. there is no dependence. */
2203 *overlaps_a = conflict_fn_no_dependence ();
2204 *overlaps_b = conflict_fn_no_dependence ();
2205 *last_conflicts = integer_zero_node;
2208 /* Both access functions are univariate. This includes SIV and MIV cases. */
2209 else if (nb_vars_a == 1 && nb_vars_b == 1)
2211 /* Both functions should have the same evolution sign. */
2212 if (((A[0][0] > 0 && -A[1][0] > 0)
2213 || (A[0][0] < 0 && -A[1][0] < 0)))
2215 /* The solutions are given by:
2217 | [GAMMA/GCD_ALPHA_BETA t].[u11 u12] = [x0]
2218 | [u21 u22] [y0]
2220 For a given integer t. Using the following variables,
2222 | i0 = u11 * gamma / gcd_alpha_beta
2223 | j0 = u12 * gamma / gcd_alpha_beta
2224 | i1 = u21
2225 | j1 = u22
2227 the solutions are:
2229 | x0 = i0 + i1 * t,
2230 | y0 = j0 + j1 * t. */
2231 HOST_WIDE_INT i0, j0, i1, j1;
2233 i0 = U[0][0] * gamma / gcd_alpha_beta;
2234 j0 = U[0][1] * gamma / gcd_alpha_beta;
2235 i1 = U[1][0];
2236 j1 = U[1][1];
2238 if ((i1 == 0 && i0 < 0)
2239 || (j1 == 0 && j0 < 0))
2241 /* There is no solution.
2242 FIXME: The case "i0 > nb_iterations, j0 > nb_iterations"
2243 falls in here, but for the moment we don't look at the
2244 upper bound of the iteration domain. */
2245 *overlaps_a = conflict_fn_no_dependence ();
2246 *overlaps_b = conflict_fn_no_dependence ();
2247 *last_conflicts = integer_zero_node;
2248 goto end_analyze_subs_aa;
2251 if (i1 > 0 && j1 > 0)
2253 HOST_WIDE_INT niter_a = estimated_loop_iterations_int
2254 (get_chrec_loop (chrec_a), false);
2255 HOST_WIDE_INT niter_b = estimated_loop_iterations_int
2256 (get_chrec_loop (chrec_b), false);
2257 HOST_WIDE_INT niter = MIN (niter_a, niter_b);
2259 /* (X0, Y0) is a solution of the Diophantine equation:
2260 "chrec_a (X0) = chrec_b (Y0)". */
2261 HOST_WIDE_INT tau1 = MAX (CEIL (-i0, i1),
2262 CEIL (-j0, j1));
2263 HOST_WIDE_INT x0 = i1 * tau1 + i0;
2264 HOST_WIDE_INT y0 = j1 * tau1 + j0;
2266 /* (X1, Y1) is the smallest positive solution of the eq
2267 "chrec_a (X1) = chrec_b (Y1)", i.e. this is where the
2268 first conflict occurs. */
2269 HOST_WIDE_INT min_multiple = MIN (x0 / i1, y0 / j1);
2270 HOST_WIDE_INT x1 = x0 - i1 * min_multiple;
2271 HOST_WIDE_INT y1 = y0 - j1 * min_multiple;
2273 if (niter > 0)
2275 HOST_WIDE_INT tau2 = MIN (FLOOR_DIV (niter - i0, i1),
2276 FLOOR_DIV (niter - j0, j1));
2277 HOST_WIDE_INT last_conflict = tau2 - (x1 - i0)/i1;
2279 /* If the overlap occurs outside of the bounds of the
2280 loop, there is no dependence. */
2281 if (x1 >= niter || y1 >= niter)
2283 *overlaps_a = conflict_fn_no_dependence ();
2284 *overlaps_b = conflict_fn_no_dependence ();
2285 *last_conflicts = integer_zero_node;
2286 goto end_analyze_subs_aa;
2288 else
2289 *last_conflicts = build_int_cst (NULL_TREE, last_conflict);
2291 else
2292 *last_conflicts = chrec_dont_know;
2294 *overlaps_a
2295 = conflict_fn (1,
2296 affine_fn_univar (build_int_cst (NULL_TREE, x1),
2298 build_int_cst (NULL_TREE, i1)));
2299 *overlaps_b
2300 = conflict_fn (1,
2301 affine_fn_univar (build_int_cst (NULL_TREE, y1),
2303 build_int_cst (NULL_TREE, j1)));
2305 else
2307 /* FIXME: For the moment, the upper bound of the
2308 iteration domain for i and j is not checked. */
2309 if (dump_file && (dump_flags & TDF_DETAILS))
2310 fprintf (dump_file, "affine-affine test failed: unimplemented.\n");
2311 *overlaps_a = conflict_fn_not_known ();
2312 *overlaps_b = conflict_fn_not_known ();
2313 *last_conflicts = chrec_dont_know;
2316 else
2318 if (dump_file && (dump_flags & TDF_DETAILS))
2319 fprintf (dump_file, "affine-affine test failed: unimplemented.\n");
2320 *overlaps_a = conflict_fn_not_known ();
2321 *overlaps_b = conflict_fn_not_known ();
2322 *last_conflicts = chrec_dont_know;
2325 else
2327 if (dump_file && (dump_flags & TDF_DETAILS))
2328 fprintf (dump_file, "affine-affine test failed: unimplemented.\n");
2329 *overlaps_a = conflict_fn_not_known ();
2330 *overlaps_b = conflict_fn_not_known ();
2331 *last_conflicts = chrec_dont_know;
2334 end_analyze_subs_aa:
2335 obstack_free (&scratch_obstack, NULL);
2336 if (dump_file && (dump_flags & TDF_DETAILS))
2338 fprintf (dump_file, " (overlaps_a = ");
2339 dump_conflict_function (dump_file, *overlaps_a);
2340 fprintf (dump_file, ")\n (overlaps_b = ");
2341 dump_conflict_function (dump_file, *overlaps_b);
2342 fprintf (dump_file, ")\n");
2343 fprintf (dump_file, ")\n");
2347 /* Returns true when analyze_subscript_affine_affine can be used for
2348 determining the dependence relation between chrec_a and chrec_b,
2349 that contain symbols. This function modifies chrec_a and chrec_b
2350 such that the analysis result is the same, and such that they don't
2351 contain symbols, and then can safely be passed to the analyzer.
2353 Example: The analysis of the following tuples of evolutions produce
2354 the same results: {x+1, +, 1}_1 vs. {x+3, +, 1}_1, and {-2, +, 1}_1
2355 vs. {0, +, 1}_1
2357 {x+1, +, 1}_1 ({2, +, 1}_1) = {x+3, +, 1}_1 ({0, +, 1}_1)
2358 {-2, +, 1}_1 ({2, +, 1}_1) = {0, +, 1}_1 ({0, +, 1}_1)
2361 static bool
2362 can_use_analyze_subscript_affine_affine (tree *chrec_a, tree *chrec_b)
2364 tree diff, type, left_a, left_b, right_b;
2366 if (chrec_contains_symbols (CHREC_RIGHT (*chrec_a))
2367 || chrec_contains_symbols (CHREC_RIGHT (*chrec_b)))
2368 /* FIXME: For the moment not handled. Might be refined later. */
2369 return false;
2371 type = chrec_type (*chrec_a);
2372 left_a = CHREC_LEFT (*chrec_a);
2373 left_b = chrec_convert (type, CHREC_LEFT (*chrec_b), NULL);
2374 diff = chrec_fold_minus (type, left_a, left_b);
2376 if (!evolution_function_is_constant_p (diff))
2377 return false;
2379 if (dump_file && (dump_flags & TDF_DETAILS))
2380 fprintf (dump_file, "can_use_subscript_aff_aff_for_symbolic \n");
2382 *chrec_a = build_polynomial_chrec (CHREC_VARIABLE (*chrec_a),
2383 diff, CHREC_RIGHT (*chrec_a));
2384 right_b = chrec_convert (type, CHREC_RIGHT (*chrec_b), NULL);
2385 *chrec_b = build_polynomial_chrec (CHREC_VARIABLE (*chrec_b),
2386 build_int_cst (type, 0),
2387 right_b);
2388 return true;
2391 /* Analyze a SIV (Single Index Variable) subscript. *OVERLAPS_A and
2392 *OVERLAPS_B are initialized to the functions that describe the
2393 relation between the elements accessed twice by CHREC_A and
2394 CHREC_B. For k >= 0, the following property is verified:
2396 CHREC_A (*OVERLAPS_A (k)) = CHREC_B (*OVERLAPS_B (k)). */
2398 static void
2399 analyze_siv_subscript (tree chrec_a,
2400 tree chrec_b,
2401 conflict_function **overlaps_a,
2402 conflict_function **overlaps_b,
2403 tree *last_conflicts,
2404 int loop_nest_num)
2406 dependence_stats.num_siv++;
2408 if (dump_file && (dump_flags & TDF_DETAILS))
2409 fprintf (dump_file, "(analyze_siv_subscript \n");
2411 if (evolution_function_is_constant_p (chrec_a)
2412 && evolution_function_is_affine_in_loop (chrec_b, loop_nest_num))
2413 analyze_siv_subscript_cst_affine (chrec_a, chrec_b,
2414 overlaps_a, overlaps_b, last_conflicts);
2416 else if (evolution_function_is_affine_in_loop (chrec_a, loop_nest_num)
2417 && evolution_function_is_constant_p (chrec_b))
2418 analyze_siv_subscript_cst_affine (chrec_b, chrec_a,
2419 overlaps_b, overlaps_a, last_conflicts);
2421 else if (evolution_function_is_affine_in_loop (chrec_a, loop_nest_num)
2422 && evolution_function_is_affine_in_loop (chrec_b, loop_nest_num))
2424 if (!chrec_contains_symbols (chrec_a)
2425 && !chrec_contains_symbols (chrec_b))
2427 analyze_subscript_affine_affine (chrec_a, chrec_b,
2428 overlaps_a, overlaps_b,
2429 last_conflicts);
2431 if (CF_NOT_KNOWN_P (*overlaps_a)
2432 || CF_NOT_KNOWN_P (*overlaps_b))
2433 dependence_stats.num_siv_unimplemented++;
2434 else if (CF_NO_DEPENDENCE_P (*overlaps_a)
2435 || CF_NO_DEPENDENCE_P (*overlaps_b))
2436 dependence_stats.num_siv_independent++;
2437 else
2438 dependence_stats.num_siv_dependent++;
2440 else if (can_use_analyze_subscript_affine_affine (&chrec_a,
2441 &chrec_b))
2443 analyze_subscript_affine_affine (chrec_a, chrec_b,
2444 overlaps_a, overlaps_b,
2445 last_conflicts);
2447 if (CF_NOT_KNOWN_P (*overlaps_a)
2448 || CF_NOT_KNOWN_P (*overlaps_b))
2449 dependence_stats.num_siv_unimplemented++;
2450 else if (CF_NO_DEPENDENCE_P (*overlaps_a)
2451 || CF_NO_DEPENDENCE_P (*overlaps_b))
2452 dependence_stats.num_siv_independent++;
2453 else
2454 dependence_stats.num_siv_dependent++;
2456 else
2457 goto siv_subscript_dontknow;
2460 else
2462 siv_subscript_dontknow:;
2463 if (dump_file && (dump_flags & TDF_DETAILS))
2464 fprintf (dump_file, "siv test failed: unimplemented.\n");
2465 *overlaps_a = conflict_fn_not_known ();
2466 *overlaps_b = conflict_fn_not_known ();
2467 *last_conflicts = chrec_dont_know;
2468 dependence_stats.num_siv_unimplemented++;
2471 if (dump_file && (dump_flags & TDF_DETAILS))
2472 fprintf (dump_file, ")\n");
2475 /* Returns false if we can prove that the greatest common divisor of the steps
2476 of CHREC does not divide CST, false otherwise. */
2478 static bool
2479 gcd_of_steps_may_divide_p (const_tree chrec, const_tree cst)
2481 HOST_WIDE_INT cd = 0, val;
2482 tree step;
2484 if (!host_integerp (cst, 0))
2485 return true;
2486 val = tree_low_cst (cst, 0);
2488 while (TREE_CODE (chrec) == POLYNOMIAL_CHREC)
2490 step = CHREC_RIGHT (chrec);
2491 if (!host_integerp (step, 0))
2492 return true;
2493 cd = gcd (cd, tree_low_cst (step, 0));
2494 chrec = CHREC_LEFT (chrec);
2497 return val % cd == 0;
2500 /* Analyze a MIV (Multiple Index Variable) subscript with respect to
2501 LOOP_NEST. *OVERLAPS_A and *OVERLAPS_B are initialized to the
2502 functions that describe the relation between the elements accessed
2503 twice by CHREC_A and CHREC_B. For k >= 0, the following property
2504 is verified:
2506 CHREC_A (*OVERLAPS_A (k)) = CHREC_B (*OVERLAPS_B (k)). */
2508 static void
2509 analyze_miv_subscript (tree chrec_a,
2510 tree chrec_b,
2511 conflict_function **overlaps_a,
2512 conflict_function **overlaps_b,
2513 tree *last_conflicts,
2514 struct loop *loop_nest)
2516 /* FIXME: This is a MIV subscript, not yet handled.
2517 Example: (A[{1, +, 1}_1] vs. A[{1, +, 1}_2]) that comes from
2518 (A[i] vs. A[j]).
2520 In the SIV test we had to solve a Diophantine equation with two
2521 variables. In the MIV case we have to solve a Diophantine
2522 equation with 2*n variables (if the subscript uses n IVs).
2524 tree type, difference;
2526 dependence_stats.num_miv++;
2527 if (dump_file && (dump_flags & TDF_DETAILS))
2528 fprintf (dump_file, "(analyze_miv_subscript \n");
2530 type = signed_type_for_types (TREE_TYPE (chrec_a), TREE_TYPE (chrec_b));
2531 chrec_a = chrec_convert (type, chrec_a, NULL);
2532 chrec_b = chrec_convert (type, chrec_b, NULL);
2533 difference = chrec_fold_minus (type, chrec_a, chrec_b);
2535 if (eq_evolutions_p (chrec_a, chrec_b))
2537 /* Access functions are the same: all the elements are accessed
2538 in the same order. */
2539 *overlaps_a = conflict_fn (1, affine_fn_cst (integer_zero_node));
2540 *overlaps_b = conflict_fn (1, affine_fn_cst (integer_zero_node));
2541 *last_conflicts = estimated_loop_iterations_tree
2542 (get_chrec_loop (chrec_a), true);
2543 dependence_stats.num_miv_dependent++;
2546 else if (evolution_function_is_constant_p (difference)
2547 /* For the moment, the following is verified:
2548 evolution_function_is_affine_multivariate_p (chrec_a,
2549 loop_nest->num) */
2550 && !gcd_of_steps_may_divide_p (chrec_a, difference))
2552 /* testsuite/.../ssa-chrec-33.c
2553 {{21, +, 2}_1, +, -2}_2 vs. {{20, +, 2}_1, +, -2}_2
2555 The difference is 1, and all the evolution steps are multiples
2556 of 2, consequently there are no overlapping elements. */
2557 *overlaps_a = conflict_fn_no_dependence ();
2558 *overlaps_b = conflict_fn_no_dependence ();
2559 *last_conflicts = integer_zero_node;
2560 dependence_stats.num_miv_independent++;
2563 else if (evolution_function_is_affine_multivariate_p (chrec_a, loop_nest->num)
2564 && !chrec_contains_symbols (chrec_a)
2565 && evolution_function_is_affine_multivariate_p (chrec_b, loop_nest->num)
2566 && !chrec_contains_symbols (chrec_b))
2568 /* testsuite/.../ssa-chrec-35.c
2569 {0, +, 1}_2 vs. {0, +, 1}_3
2570 the overlapping elements are respectively located at iterations:
2571 {0, +, 1}_x and {0, +, 1}_x,
2572 in other words, we have the equality:
2573 {0, +, 1}_2 ({0, +, 1}_x) = {0, +, 1}_3 ({0, +, 1}_x)
2575 Other examples:
2576 {{0, +, 1}_1, +, 2}_2 ({0, +, 1}_x, {0, +, 1}_y) =
2577 {0, +, 1}_1 ({{0, +, 1}_x, +, 2}_y)
2579 {{0, +, 2}_1, +, 3}_2 ({0, +, 1}_y, {0, +, 1}_x) =
2580 {{0, +, 3}_1, +, 2}_2 ({0, +, 1}_x, {0, +, 1}_y)
2582 analyze_subscript_affine_affine (chrec_a, chrec_b,
2583 overlaps_a, overlaps_b, last_conflicts);
2585 if (CF_NOT_KNOWN_P (*overlaps_a)
2586 || CF_NOT_KNOWN_P (*overlaps_b))
2587 dependence_stats.num_miv_unimplemented++;
2588 else if (CF_NO_DEPENDENCE_P (*overlaps_a)
2589 || CF_NO_DEPENDENCE_P (*overlaps_b))
2590 dependence_stats.num_miv_independent++;
2591 else
2592 dependence_stats.num_miv_dependent++;
2595 else
2597 /* When the analysis is too difficult, answer "don't know". */
2598 if (dump_file && (dump_flags & TDF_DETAILS))
2599 fprintf (dump_file, "analyze_miv_subscript test failed: unimplemented.\n");
2601 *overlaps_a = conflict_fn_not_known ();
2602 *overlaps_b = conflict_fn_not_known ();
2603 *last_conflicts = chrec_dont_know;
2604 dependence_stats.num_miv_unimplemented++;
2607 if (dump_file && (dump_flags & TDF_DETAILS))
2608 fprintf (dump_file, ")\n");
2611 /* Determines the iterations for which CHREC_A is equal to CHREC_B in
2612 with respect to LOOP_NEST. OVERLAP_ITERATIONS_A and
2613 OVERLAP_ITERATIONS_B are initialized with two functions that
2614 describe the iterations that contain conflicting elements.
2616 Remark: For an integer k >= 0, the following equality is true:
2618 CHREC_A (OVERLAP_ITERATIONS_A (k)) == CHREC_B (OVERLAP_ITERATIONS_B (k)).
2621 static void
2622 analyze_overlapping_iterations (tree chrec_a,
2623 tree chrec_b,
2624 conflict_function **overlap_iterations_a,
2625 conflict_function **overlap_iterations_b,
2626 tree *last_conflicts, struct loop *loop_nest)
2628 unsigned int lnn = loop_nest->num;
2630 dependence_stats.num_subscript_tests++;
2632 if (dump_file && (dump_flags & TDF_DETAILS))
2634 fprintf (dump_file, "(analyze_overlapping_iterations \n");
2635 fprintf (dump_file, " (chrec_a = ");
2636 print_generic_expr (dump_file, chrec_a, 0);
2637 fprintf (dump_file, ")\n (chrec_b = ");
2638 print_generic_expr (dump_file, chrec_b, 0);
2639 fprintf (dump_file, ")\n");
2642 if (chrec_a == NULL_TREE
2643 || chrec_b == NULL_TREE
2644 || chrec_contains_undetermined (chrec_a)
2645 || chrec_contains_undetermined (chrec_b))
2647 dependence_stats.num_subscript_undetermined++;
2649 *overlap_iterations_a = conflict_fn_not_known ();
2650 *overlap_iterations_b = conflict_fn_not_known ();
2653 /* If they are the same chrec, and are affine, they overlap
2654 on every iteration. */
2655 else if (eq_evolutions_p (chrec_a, chrec_b)
2656 && (evolution_function_is_affine_multivariate_p (chrec_a, lnn)
2657 || operand_equal_p (chrec_a, chrec_b, 0)))
2659 dependence_stats.num_same_subscript_function++;
2660 *overlap_iterations_a = conflict_fn (1, affine_fn_cst (integer_zero_node));
2661 *overlap_iterations_b = conflict_fn (1, affine_fn_cst (integer_zero_node));
2662 *last_conflicts = chrec_dont_know;
2665 /* If they aren't the same, and aren't affine, we can't do anything
2666 yet. */
2667 else if ((chrec_contains_symbols (chrec_a)
2668 || chrec_contains_symbols (chrec_b))
2669 && (!evolution_function_is_affine_multivariate_p (chrec_a, lnn)
2670 || !evolution_function_is_affine_multivariate_p (chrec_b, lnn)))
2672 dependence_stats.num_subscript_undetermined++;
2673 *overlap_iterations_a = conflict_fn_not_known ();
2674 *overlap_iterations_b = conflict_fn_not_known ();
2677 else if (ziv_subscript_p (chrec_a, chrec_b))
2678 analyze_ziv_subscript (chrec_a, chrec_b,
2679 overlap_iterations_a, overlap_iterations_b,
2680 last_conflicts);
2682 else if (siv_subscript_p (chrec_a, chrec_b))
2683 analyze_siv_subscript (chrec_a, chrec_b,
2684 overlap_iterations_a, overlap_iterations_b,
2685 last_conflicts, lnn);
2687 else
2688 analyze_miv_subscript (chrec_a, chrec_b,
2689 overlap_iterations_a, overlap_iterations_b,
2690 last_conflicts, loop_nest);
2692 if (dump_file && (dump_flags & TDF_DETAILS))
2694 fprintf (dump_file, " (overlap_iterations_a = ");
2695 dump_conflict_function (dump_file, *overlap_iterations_a);
2696 fprintf (dump_file, ")\n (overlap_iterations_b = ");
2697 dump_conflict_function (dump_file, *overlap_iterations_b);
2698 fprintf (dump_file, ")\n");
2699 fprintf (dump_file, ")\n");
2703 /* Helper function for uniquely inserting distance vectors. */
2705 static void
2706 save_dist_v (struct data_dependence_relation *ddr, lambda_vector dist_v)
2708 unsigned i;
2709 lambda_vector v;
2711 FOR_EACH_VEC_ELT (lambda_vector, DDR_DIST_VECTS (ddr), i, v)
2712 if (lambda_vector_equal (v, dist_v, DDR_NB_LOOPS (ddr)))
2713 return;
2715 VEC_safe_push (lambda_vector, heap, DDR_DIST_VECTS (ddr), dist_v);
2718 /* Helper function for uniquely inserting direction vectors. */
2720 static void
2721 save_dir_v (struct data_dependence_relation *ddr, lambda_vector dir_v)
2723 unsigned i;
2724 lambda_vector v;
2726 FOR_EACH_VEC_ELT (lambda_vector, DDR_DIR_VECTS (ddr), i, v)
2727 if (lambda_vector_equal (v, dir_v, DDR_NB_LOOPS (ddr)))
2728 return;
2730 VEC_safe_push (lambda_vector, heap, DDR_DIR_VECTS (ddr), dir_v);
2733 /* Add a distance of 1 on all the loops outer than INDEX. If we
2734 haven't yet determined a distance for this outer loop, push a new
2735 distance vector composed of the previous distance, and a distance
2736 of 1 for this outer loop. Example:
2738 | loop_1
2739 | loop_2
2740 | A[10]
2741 | endloop_2
2742 | endloop_1
2744 Saved vectors are of the form (dist_in_1, dist_in_2). First, we
2745 save (0, 1), then we have to save (1, 0). */
2747 static void
2748 add_outer_distances (struct data_dependence_relation *ddr,
2749 lambda_vector dist_v, int index)
2751 /* For each outer loop where init_v is not set, the accesses are
2752 in dependence of distance 1 in the loop. */
2753 while (--index >= 0)
2755 lambda_vector save_v = lambda_vector_new (DDR_NB_LOOPS (ddr));
2756 lambda_vector_copy (dist_v, save_v, DDR_NB_LOOPS (ddr));
2757 save_v[index] = 1;
2758 save_dist_v (ddr, save_v);
2762 /* Return false when fail to represent the data dependence as a
2763 distance vector. INIT_B is set to true when a component has been
2764 added to the distance vector DIST_V. INDEX_CARRY is then set to
2765 the index in DIST_V that carries the dependence. */
2767 static bool
2768 build_classic_dist_vector_1 (struct data_dependence_relation *ddr,
2769 struct data_reference *ddr_a,
2770 struct data_reference *ddr_b,
2771 lambda_vector dist_v, bool *init_b,
2772 int *index_carry)
2774 unsigned i;
2775 lambda_vector init_v = lambda_vector_new (DDR_NB_LOOPS (ddr));
2777 for (i = 0; i < DDR_NUM_SUBSCRIPTS (ddr); i++)
2779 tree access_fn_a, access_fn_b;
2780 struct subscript *subscript = DDR_SUBSCRIPT (ddr, i);
2782 if (chrec_contains_undetermined (SUB_DISTANCE (subscript)))
2784 non_affine_dependence_relation (ddr);
2785 return false;
2788 access_fn_a = DR_ACCESS_FN (ddr_a, i);
2789 access_fn_b = DR_ACCESS_FN (ddr_b, i);
2791 if (TREE_CODE (access_fn_a) == POLYNOMIAL_CHREC
2792 && TREE_CODE (access_fn_b) == POLYNOMIAL_CHREC)
2794 int dist, index;
2795 int index_a = index_in_loop_nest (CHREC_VARIABLE (access_fn_a),
2796 DDR_LOOP_NEST (ddr));
2797 int index_b = index_in_loop_nest (CHREC_VARIABLE (access_fn_b),
2798 DDR_LOOP_NEST (ddr));
2800 /* The dependence is carried by the outermost loop. Example:
2801 | loop_1
2802 | A[{4, +, 1}_1]
2803 | loop_2
2804 | A[{5, +, 1}_2]
2805 | endloop_2
2806 | endloop_1
2807 In this case, the dependence is carried by loop_1. */
2808 index = index_a < index_b ? index_a : index_b;
2809 *index_carry = MIN (index, *index_carry);
2811 if (chrec_contains_undetermined (SUB_DISTANCE (subscript)))
2813 non_affine_dependence_relation (ddr);
2814 return false;
2817 dist = int_cst_value (SUB_DISTANCE (subscript));
2819 /* This is the subscript coupling test. If we have already
2820 recorded a distance for this loop (a distance coming from
2821 another subscript), it should be the same. For example,
2822 in the following code, there is no dependence:
2824 | loop i = 0, N, 1
2825 | T[i+1][i] = ...
2826 | ... = T[i][i]
2827 | endloop
2829 if (init_v[index] != 0 && dist_v[index] != dist)
2831 finalize_ddr_dependent (ddr, chrec_known);
2832 return false;
2835 dist_v[index] = dist;
2836 init_v[index] = 1;
2837 *init_b = true;
2839 else if (!operand_equal_p (access_fn_a, access_fn_b, 0))
2841 /* This can be for example an affine vs. constant dependence
2842 (T[i] vs. T[3]) that is not an affine dependence and is
2843 not representable as a distance vector. */
2844 non_affine_dependence_relation (ddr);
2845 return false;
2849 return true;
2852 /* Return true when the DDR contains only constant access functions. */
2854 static bool
2855 constant_access_functions (const struct data_dependence_relation *ddr)
2857 unsigned i;
2859 for (i = 0; i < DDR_NUM_SUBSCRIPTS (ddr); i++)
2860 if (!evolution_function_is_constant_p (DR_ACCESS_FN (DDR_A (ddr), i))
2861 || !evolution_function_is_constant_p (DR_ACCESS_FN (DDR_B (ddr), i)))
2862 return false;
2864 return true;
2867 /* Helper function for the case where DDR_A and DDR_B are the same
2868 multivariate access function with a constant step. For an example
2869 see pr34635-1.c. */
2871 static void
2872 add_multivariate_self_dist (struct data_dependence_relation *ddr, tree c_2)
2874 int x_1, x_2;
2875 tree c_1 = CHREC_LEFT (c_2);
2876 tree c_0 = CHREC_LEFT (c_1);
2877 lambda_vector dist_v;
2878 int v1, v2, cd;
2880 /* Polynomials with more than 2 variables are not handled yet. When
2881 the evolution steps are parameters, it is not possible to
2882 represent the dependence using classical distance vectors. */
2883 if (TREE_CODE (c_0) != INTEGER_CST
2884 || TREE_CODE (CHREC_RIGHT (c_1)) != INTEGER_CST
2885 || TREE_CODE (CHREC_RIGHT (c_2)) != INTEGER_CST)
2887 DDR_AFFINE_P (ddr) = false;
2888 return;
2891 x_2 = index_in_loop_nest (CHREC_VARIABLE (c_2), DDR_LOOP_NEST (ddr));
2892 x_1 = index_in_loop_nest (CHREC_VARIABLE (c_1), DDR_LOOP_NEST (ddr));
2894 /* For "{{0, +, 2}_1, +, 3}_2" the distance vector is (3, -2). */
2895 dist_v = lambda_vector_new (DDR_NB_LOOPS (ddr));
2896 v1 = int_cst_value (CHREC_RIGHT (c_1));
2897 v2 = int_cst_value (CHREC_RIGHT (c_2));
2898 cd = gcd (v1, v2);
2899 v1 /= cd;
2900 v2 /= cd;
2902 if (v2 < 0)
2904 v2 = -v2;
2905 v1 = -v1;
2908 dist_v[x_1] = v2;
2909 dist_v[x_2] = -v1;
2910 save_dist_v (ddr, dist_v);
2912 add_outer_distances (ddr, dist_v, x_1);
2915 /* Helper function for the case where DDR_A and DDR_B are the same
2916 access functions. */
2918 static void
2919 add_other_self_distances (struct data_dependence_relation *ddr)
2921 lambda_vector dist_v;
2922 unsigned i;
2923 int index_carry = DDR_NB_LOOPS (ddr);
2925 for (i = 0; i < DDR_NUM_SUBSCRIPTS (ddr); i++)
2927 tree access_fun = DR_ACCESS_FN (DDR_A (ddr), i);
2929 if (TREE_CODE (access_fun) == POLYNOMIAL_CHREC)
2931 if (!evolution_function_is_univariate_p (access_fun))
2933 if (DDR_NUM_SUBSCRIPTS (ddr) != 1)
2935 DDR_ARE_DEPENDENT (ddr) = chrec_dont_know;
2936 return;
2939 access_fun = DR_ACCESS_FN (DDR_A (ddr), 0);
2941 if (TREE_CODE (CHREC_LEFT (access_fun)) == POLYNOMIAL_CHREC)
2942 add_multivariate_self_dist (ddr, access_fun);
2943 else
2944 /* The evolution step is not constant: it varies in
2945 the outer loop, so this cannot be represented by a
2946 distance vector. For example in pr34635.c the
2947 evolution is {0, +, {0, +, 4}_1}_2. */
2948 DDR_AFFINE_P (ddr) = false;
2950 return;
2953 index_carry = MIN (index_carry,
2954 index_in_loop_nest (CHREC_VARIABLE (access_fun),
2955 DDR_LOOP_NEST (ddr)));
2959 dist_v = lambda_vector_new (DDR_NB_LOOPS (ddr));
2960 add_outer_distances (ddr, dist_v, index_carry);
2963 static void
2964 insert_innermost_unit_dist_vector (struct data_dependence_relation *ddr)
2966 lambda_vector dist_v = lambda_vector_new (DDR_NB_LOOPS (ddr));
2968 dist_v[DDR_INNER_LOOP (ddr)] = 1;
2969 save_dist_v (ddr, dist_v);
2972 /* Adds a unit distance vector to DDR when there is a 0 overlap. This
2973 is the case for example when access functions are the same and
2974 equal to a constant, as in:
2976 | loop_1
2977 | A[3] = ...
2978 | ... = A[3]
2979 | endloop_1
2981 in which case the distance vectors are (0) and (1). */
2983 static void
2984 add_distance_for_zero_overlaps (struct data_dependence_relation *ddr)
2986 unsigned i, j;
2988 for (i = 0; i < DDR_NUM_SUBSCRIPTS (ddr); i++)
2990 subscript_p sub = DDR_SUBSCRIPT (ddr, i);
2991 conflict_function *ca = SUB_CONFLICTS_IN_A (sub);
2992 conflict_function *cb = SUB_CONFLICTS_IN_B (sub);
2994 for (j = 0; j < ca->n; j++)
2995 if (affine_function_zero_p (ca->fns[j]))
2997 insert_innermost_unit_dist_vector (ddr);
2998 return;
3001 for (j = 0; j < cb->n; j++)
3002 if (affine_function_zero_p (cb->fns[j]))
3004 insert_innermost_unit_dist_vector (ddr);
3005 return;
3010 /* Compute the classic per loop distance vector. DDR is the data
3011 dependence relation to build a vector from. Return false when fail
3012 to represent the data dependence as a distance vector. */
3014 static bool
3015 build_classic_dist_vector (struct data_dependence_relation *ddr,
3016 struct loop *loop_nest)
3018 bool init_b = false;
3019 int index_carry = DDR_NB_LOOPS (ddr);
3020 lambda_vector dist_v;
3022 if (DDR_ARE_DEPENDENT (ddr) != NULL_TREE)
3023 return false;
3025 if (same_access_functions (ddr))
3027 /* Save the 0 vector. */
3028 dist_v = lambda_vector_new (DDR_NB_LOOPS (ddr));
3029 save_dist_v (ddr, dist_v);
3031 if (constant_access_functions (ddr))
3032 add_distance_for_zero_overlaps (ddr);
3034 if (DDR_NB_LOOPS (ddr) > 1)
3035 add_other_self_distances (ddr);
3037 return true;
3040 dist_v = lambda_vector_new (DDR_NB_LOOPS (ddr));
3041 if (!build_classic_dist_vector_1 (ddr, DDR_A (ddr), DDR_B (ddr),
3042 dist_v, &init_b, &index_carry))
3043 return false;
3045 /* Save the distance vector if we initialized one. */
3046 if (init_b)
3048 /* Verify a basic constraint: classic distance vectors should
3049 always be lexicographically positive.
3051 Data references are collected in the order of execution of
3052 the program, thus for the following loop
3054 | for (i = 1; i < 100; i++)
3055 | for (j = 1; j < 100; j++)
3057 | t = T[j+1][i-1]; // A
3058 | T[j][i] = t + 2; // B
3061 references are collected following the direction of the wind:
3062 A then B. The data dependence tests are performed also
3063 following this order, such that we're looking at the distance
3064 separating the elements accessed by A from the elements later
3065 accessed by B. But in this example, the distance returned by
3066 test_dep (A, B) is lexicographically negative (-1, 1), that
3067 means that the access A occurs later than B with respect to
3068 the outer loop, ie. we're actually looking upwind. In this
3069 case we solve test_dep (B, A) looking downwind to the
3070 lexicographically positive solution, that returns the
3071 distance vector (1, -1). */
3072 if (!lambda_vector_lexico_pos (dist_v, DDR_NB_LOOPS (ddr)))
3074 lambda_vector save_v = lambda_vector_new (DDR_NB_LOOPS (ddr));
3075 if (!subscript_dependence_tester_1 (ddr, DDR_B (ddr), DDR_A (ddr),
3076 loop_nest))
3077 return false;
3078 compute_subscript_distance (ddr);
3079 if (!build_classic_dist_vector_1 (ddr, DDR_B (ddr), DDR_A (ddr),
3080 save_v, &init_b, &index_carry))
3081 return false;
3082 save_dist_v (ddr, save_v);
3083 DDR_REVERSED_P (ddr) = true;
3085 /* In this case there is a dependence forward for all the
3086 outer loops:
3088 | for (k = 1; k < 100; k++)
3089 | for (i = 1; i < 100; i++)
3090 | for (j = 1; j < 100; j++)
3092 | t = T[j+1][i-1]; // A
3093 | T[j][i] = t + 2; // B
3096 the vectors are:
3097 (0, 1, -1)
3098 (1, 1, -1)
3099 (1, -1, 1)
3101 if (DDR_NB_LOOPS (ddr) > 1)
3103 add_outer_distances (ddr, save_v, index_carry);
3104 add_outer_distances (ddr, dist_v, index_carry);
3107 else
3109 lambda_vector save_v = lambda_vector_new (DDR_NB_LOOPS (ddr));
3110 lambda_vector_copy (dist_v, save_v, DDR_NB_LOOPS (ddr));
3112 if (DDR_NB_LOOPS (ddr) > 1)
3114 lambda_vector opposite_v = lambda_vector_new (DDR_NB_LOOPS (ddr));
3116 if (!subscript_dependence_tester_1 (ddr, DDR_B (ddr),
3117 DDR_A (ddr), loop_nest))
3118 return false;
3119 compute_subscript_distance (ddr);
3120 if (!build_classic_dist_vector_1 (ddr, DDR_B (ddr), DDR_A (ddr),
3121 opposite_v, &init_b,
3122 &index_carry))
3123 return false;
3125 save_dist_v (ddr, save_v);
3126 add_outer_distances (ddr, dist_v, index_carry);
3127 add_outer_distances (ddr, opposite_v, index_carry);
3129 else
3130 save_dist_v (ddr, save_v);
3133 else
3135 /* There is a distance of 1 on all the outer loops: Example:
3136 there is a dependence of distance 1 on loop_1 for the array A.
3138 | loop_1
3139 | A[5] = ...
3140 | endloop
3142 add_outer_distances (ddr, dist_v,
3143 lambda_vector_first_nz (dist_v,
3144 DDR_NB_LOOPS (ddr), 0));
3147 if (dump_file && (dump_flags & TDF_DETAILS))
3149 unsigned i;
3151 fprintf (dump_file, "(build_classic_dist_vector\n");
3152 for (i = 0; i < DDR_NUM_DIST_VECTS (ddr); i++)
3154 fprintf (dump_file, " dist_vector = (");
3155 print_lambda_vector (dump_file, DDR_DIST_VECT (ddr, i),
3156 DDR_NB_LOOPS (ddr));
3157 fprintf (dump_file, " )\n");
3159 fprintf (dump_file, ")\n");
3162 return true;
3165 /* Return the direction for a given distance.
3166 FIXME: Computing dir this way is suboptimal, since dir can catch
3167 cases that dist is unable to represent. */
3169 static inline enum data_dependence_direction
3170 dir_from_dist (int dist)
3172 if (dist > 0)
3173 return dir_positive;
3174 else if (dist < 0)
3175 return dir_negative;
3176 else
3177 return dir_equal;
3180 /* Compute the classic per loop direction vector. DDR is the data
3181 dependence relation to build a vector from. */
3183 static void
3184 build_classic_dir_vector (struct data_dependence_relation *ddr)
3186 unsigned i, j;
3187 lambda_vector dist_v;
3189 FOR_EACH_VEC_ELT (lambda_vector, DDR_DIST_VECTS (ddr), i, dist_v)
3191 lambda_vector dir_v = lambda_vector_new (DDR_NB_LOOPS (ddr));
3193 for (j = 0; j < DDR_NB_LOOPS (ddr); j++)
3194 dir_v[j] = dir_from_dist (dist_v[j]);
3196 save_dir_v (ddr, dir_v);
3200 /* Helper function. Returns true when there is a dependence between
3201 data references DRA and DRB. */
3203 static bool
3204 subscript_dependence_tester_1 (struct data_dependence_relation *ddr,
3205 struct data_reference *dra,
3206 struct data_reference *drb,
3207 struct loop *loop_nest)
3209 unsigned int i;
3210 tree last_conflicts;
3211 struct subscript *subscript;
3213 for (i = 0; VEC_iterate (subscript_p, DDR_SUBSCRIPTS (ddr), i, subscript);
3214 i++)
3216 conflict_function *overlaps_a, *overlaps_b;
3218 analyze_overlapping_iterations (DR_ACCESS_FN (dra, i),
3219 DR_ACCESS_FN (drb, i),
3220 &overlaps_a, &overlaps_b,
3221 &last_conflicts, loop_nest);
3223 if (CF_NOT_KNOWN_P (overlaps_a)
3224 || CF_NOT_KNOWN_P (overlaps_b))
3226 finalize_ddr_dependent (ddr, chrec_dont_know);
3227 dependence_stats.num_dependence_undetermined++;
3228 free_conflict_function (overlaps_a);
3229 free_conflict_function (overlaps_b);
3230 return false;
3233 else if (CF_NO_DEPENDENCE_P (overlaps_a)
3234 || CF_NO_DEPENDENCE_P (overlaps_b))
3236 finalize_ddr_dependent (ddr, chrec_known);
3237 dependence_stats.num_dependence_independent++;
3238 free_conflict_function (overlaps_a);
3239 free_conflict_function (overlaps_b);
3240 return false;
3243 else
3245 if (SUB_CONFLICTS_IN_A (subscript))
3246 free_conflict_function (SUB_CONFLICTS_IN_A (subscript));
3247 if (SUB_CONFLICTS_IN_B (subscript))
3248 free_conflict_function (SUB_CONFLICTS_IN_B (subscript));
3250 SUB_CONFLICTS_IN_A (subscript) = overlaps_a;
3251 SUB_CONFLICTS_IN_B (subscript) = overlaps_b;
3252 SUB_LAST_CONFLICT (subscript) = last_conflicts;
3256 return true;
3259 /* Computes the conflicting iterations in LOOP_NEST, and initialize DDR. */
3261 static void
3262 subscript_dependence_tester (struct data_dependence_relation *ddr,
3263 struct loop *loop_nest)
3266 if (dump_file && (dump_flags & TDF_DETAILS))
3267 fprintf (dump_file, "(subscript_dependence_tester \n");
3269 if (subscript_dependence_tester_1 (ddr, DDR_A (ddr), DDR_B (ddr), loop_nest))
3270 dependence_stats.num_dependence_dependent++;
3272 compute_subscript_distance (ddr);
3273 if (build_classic_dist_vector (ddr, loop_nest))
3274 build_classic_dir_vector (ddr);
3276 if (dump_file && (dump_flags & TDF_DETAILS))
3277 fprintf (dump_file, ")\n");
3280 /* Returns true when all the access functions of A are affine or
3281 constant with respect to LOOP_NEST. */
3283 static bool
3284 access_functions_are_affine_or_constant_p (const struct data_reference *a,
3285 const struct loop *loop_nest)
3287 unsigned int i;
3288 VEC(tree,heap) *fns = DR_ACCESS_FNS (a);
3289 tree t;
3291 FOR_EACH_VEC_ELT (tree, fns, i, t)
3292 if (!evolution_function_is_invariant_p (t, loop_nest->num)
3293 && !evolution_function_is_affine_multivariate_p (t, loop_nest->num))
3294 return false;
3296 return true;
3299 /* Initializes an equation for an OMEGA problem using the information
3300 contained in the ACCESS_FUN. Returns true when the operation
3301 succeeded.
3303 PB is the omega constraint system.
3304 EQ is the number of the equation to be initialized.
3305 OFFSET is used for shifting the variables names in the constraints:
3306 a constrain is composed of 2 * the number of variables surrounding
3307 dependence accesses. OFFSET is set either to 0 for the first n variables,
3308 then it is set to n.
3309 ACCESS_FUN is expected to be an affine chrec. */
3311 static bool
3312 init_omega_eq_with_af (omega_pb pb, unsigned eq,
3313 unsigned int offset, tree access_fun,
3314 struct data_dependence_relation *ddr)
3316 switch (TREE_CODE (access_fun))
3318 case POLYNOMIAL_CHREC:
3320 tree left = CHREC_LEFT (access_fun);
3321 tree right = CHREC_RIGHT (access_fun);
3322 int var = CHREC_VARIABLE (access_fun);
3323 unsigned var_idx;
3325 if (TREE_CODE (right) != INTEGER_CST)
3326 return false;
3328 var_idx = index_in_loop_nest (var, DDR_LOOP_NEST (ddr));
3329 pb->eqs[eq].coef[offset + var_idx + 1] = int_cst_value (right);
3331 /* Compute the innermost loop index. */
3332 DDR_INNER_LOOP (ddr) = MAX (DDR_INNER_LOOP (ddr), var_idx);
3334 if (offset == 0)
3335 pb->eqs[eq].coef[var_idx + DDR_NB_LOOPS (ddr) + 1]
3336 += int_cst_value (right);
3338 switch (TREE_CODE (left))
3340 case POLYNOMIAL_CHREC:
3341 return init_omega_eq_with_af (pb, eq, offset, left, ddr);
3343 case INTEGER_CST:
3344 pb->eqs[eq].coef[0] += int_cst_value (left);
3345 return true;
3347 default:
3348 return false;
3352 case INTEGER_CST:
3353 pb->eqs[eq].coef[0] += int_cst_value (access_fun);
3354 return true;
3356 default:
3357 return false;
3361 /* As explained in the comments preceding init_omega_for_ddr, we have
3362 to set up a system for each loop level, setting outer loops
3363 variation to zero, and current loop variation to positive or zero.
3364 Save each lexico positive distance vector. */
3366 static void
3367 omega_extract_distance_vectors (omega_pb pb,
3368 struct data_dependence_relation *ddr)
3370 int eq, geq;
3371 unsigned i, j;
3372 struct loop *loopi, *loopj;
3373 enum omega_result res;
3375 /* Set a new problem for each loop in the nest. The basis is the
3376 problem that we have initialized until now. On top of this we
3377 add new constraints. */
3378 for (i = 0; i <= DDR_INNER_LOOP (ddr)
3379 && VEC_iterate (loop_p, DDR_LOOP_NEST (ddr), i, loopi); i++)
3381 int dist = 0;
3382 omega_pb copy = omega_alloc_problem (2 * DDR_NB_LOOPS (ddr),
3383 DDR_NB_LOOPS (ddr));
3385 omega_copy_problem (copy, pb);
3387 /* For all the outer loops "loop_j", add "dj = 0". */
3388 for (j = 0;
3389 j < i && VEC_iterate (loop_p, DDR_LOOP_NEST (ddr), j, loopj); j++)
3391 eq = omega_add_zero_eq (copy, omega_black);
3392 copy->eqs[eq].coef[j + 1] = 1;
3395 /* For "loop_i", add "0 <= di". */
3396 geq = omega_add_zero_geq (copy, omega_black);
3397 copy->geqs[geq].coef[i + 1] = 1;
3399 /* Reduce the constraint system, and test that the current
3400 problem is feasible. */
3401 res = omega_simplify_problem (copy);
3402 if (res == omega_false
3403 || res == omega_unknown
3404 || copy->num_geqs > (int) DDR_NB_LOOPS (ddr))
3405 goto next_problem;
3407 for (eq = 0; eq < copy->num_subs; eq++)
3408 if (copy->subs[eq].key == (int) i + 1)
3410 dist = copy->subs[eq].coef[0];
3411 goto found_dist;
3414 if (dist == 0)
3416 /* Reinitialize problem... */
3417 omega_copy_problem (copy, pb);
3418 for (j = 0;
3419 j < i && VEC_iterate (loop_p, DDR_LOOP_NEST (ddr), j, loopj); j++)
3421 eq = omega_add_zero_eq (copy, omega_black);
3422 copy->eqs[eq].coef[j + 1] = 1;
3425 /* ..., but this time "di = 1". */
3426 eq = omega_add_zero_eq (copy, omega_black);
3427 copy->eqs[eq].coef[i + 1] = 1;
3428 copy->eqs[eq].coef[0] = -1;
3430 res = omega_simplify_problem (copy);
3431 if (res == omega_false
3432 || res == omega_unknown
3433 || copy->num_geqs > (int) DDR_NB_LOOPS (ddr))
3434 goto next_problem;
3436 for (eq = 0; eq < copy->num_subs; eq++)
3437 if (copy->subs[eq].key == (int) i + 1)
3439 dist = copy->subs[eq].coef[0];
3440 goto found_dist;
3444 found_dist:;
3445 /* Save the lexicographically positive distance vector. */
3446 if (dist >= 0)
3448 lambda_vector dist_v = lambda_vector_new (DDR_NB_LOOPS (ddr));
3449 lambda_vector dir_v = lambda_vector_new (DDR_NB_LOOPS (ddr));
3451 dist_v[i] = dist;
3453 for (eq = 0; eq < copy->num_subs; eq++)
3454 if (copy->subs[eq].key > 0)
3456 dist = copy->subs[eq].coef[0];
3457 dist_v[copy->subs[eq].key - 1] = dist;
3460 for (j = 0; j < DDR_NB_LOOPS (ddr); j++)
3461 dir_v[j] = dir_from_dist (dist_v[j]);
3463 save_dist_v (ddr, dist_v);
3464 save_dir_v (ddr, dir_v);
3467 next_problem:;
3468 omega_free_problem (copy);
3472 /* This is called for each subscript of a tuple of data references:
3473 insert an equality for representing the conflicts. */
3475 static bool
3476 omega_setup_subscript (tree access_fun_a, tree access_fun_b,
3477 struct data_dependence_relation *ddr,
3478 omega_pb pb, bool *maybe_dependent)
3480 int eq;
3481 tree type = signed_type_for_types (TREE_TYPE (access_fun_a),
3482 TREE_TYPE (access_fun_b));
3483 tree fun_a = chrec_convert (type, access_fun_a, NULL);
3484 tree fun_b = chrec_convert (type, access_fun_b, NULL);
3485 tree difference = chrec_fold_minus (type, fun_a, fun_b);
3486 tree minus_one;
3488 /* When the fun_a - fun_b is not constant, the dependence is not
3489 captured by the classic distance vector representation. */
3490 if (TREE_CODE (difference) != INTEGER_CST)
3491 return false;
3493 /* ZIV test. */
3494 if (ziv_subscript_p (fun_a, fun_b) && !integer_zerop (difference))
3496 /* There is no dependence. */
3497 *maybe_dependent = false;
3498 return true;
3501 minus_one = build_int_cst (type, -1);
3502 fun_b = chrec_fold_multiply (type, fun_b, minus_one);
3504 eq = omega_add_zero_eq (pb, omega_black);
3505 if (!init_omega_eq_with_af (pb, eq, DDR_NB_LOOPS (ddr), fun_a, ddr)
3506 || !init_omega_eq_with_af (pb, eq, 0, fun_b, ddr))
3507 /* There is probably a dependence, but the system of
3508 constraints cannot be built: answer "don't know". */
3509 return false;
3511 /* GCD test. */
3512 if (DDR_NB_LOOPS (ddr) != 0 && pb->eqs[eq].coef[0]
3513 && !int_divides_p (lambda_vector_gcd
3514 ((lambda_vector) &(pb->eqs[eq].coef[1]),
3515 2 * DDR_NB_LOOPS (ddr)),
3516 pb->eqs[eq].coef[0]))
3518 /* There is no dependence. */
3519 *maybe_dependent = false;
3520 return true;
3523 return true;
3526 /* Helper function, same as init_omega_for_ddr but specialized for
3527 data references A and B. */
3529 static bool
3530 init_omega_for_ddr_1 (struct data_reference *dra, struct data_reference *drb,
3531 struct data_dependence_relation *ddr,
3532 omega_pb pb, bool *maybe_dependent)
3534 unsigned i;
3535 int ineq;
3536 struct loop *loopi;
3537 unsigned nb_loops = DDR_NB_LOOPS (ddr);
3539 /* Insert an equality per subscript. */
3540 for (i = 0; i < DDR_NUM_SUBSCRIPTS (ddr); i++)
3542 if (!omega_setup_subscript (DR_ACCESS_FN (dra, i), DR_ACCESS_FN (drb, i),
3543 ddr, pb, maybe_dependent))
3544 return false;
3545 else if (*maybe_dependent == false)
3547 /* There is no dependence. */
3548 DDR_ARE_DEPENDENT (ddr) = chrec_known;
3549 return true;
3553 /* Insert inequalities: constraints corresponding to the iteration
3554 domain, i.e. the loops surrounding the references "loop_x" and
3555 the distance variables "dx". The layout of the OMEGA
3556 representation is as follows:
3557 - coef[0] is the constant
3558 - coef[1..nb_loops] are the protected variables that will not be
3559 removed by the solver: the "dx"
3560 - coef[nb_loops + 1, 2*nb_loops] are the loop variables: "loop_x".
3562 for (i = 0; i <= DDR_INNER_LOOP (ddr)
3563 && VEC_iterate (loop_p, DDR_LOOP_NEST (ddr), i, loopi); i++)
3565 HOST_WIDE_INT nbi = estimated_loop_iterations_int (loopi, false);
3567 /* 0 <= loop_x */
3568 ineq = omega_add_zero_geq (pb, omega_black);
3569 pb->geqs[ineq].coef[i + nb_loops + 1] = 1;
3571 /* 0 <= loop_x + dx */
3572 ineq = omega_add_zero_geq (pb, omega_black);
3573 pb->geqs[ineq].coef[i + nb_loops + 1] = 1;
3574 pb->geqs[ineq].coef[i + 1] = 1;
3576 if (nbi != -1)
3578 /* loop_x <= nb_iters */
3579 ineq = omega_add_zero_geq (pb, omega_black);
3580 pb->geqs[ineq].coef[i + nb_loops + 1] = -1;
3581 pb->geqs[ineq].coef[0] = nbi;
3583 /* loop_x + dx <= nb_iters */
3584 ineq = omega_add_zero_geq (pb, omega_black);
3585 pb->geqs[ineq].coef[i + nb_loops + 1] = -1;
3586 pb->geqs[ineq].coef[i + 1] = -1;
3587 pb->geqs[ineq].coef[0] = nbi;
3589 /* A step "dx" bigger than nb_iters is not feasible, so
3590 add "0 <= nb_iters + dx", */
3591 ineq = omega_add_zero_geq (pb, omega_black);
3592 pb->geqs[ineq].coef[i + 1] = 1;
3593 pb->geqs[ineq].coef[0] = nbi;
3594 /* and "dx <= nb_iters". */
3595 ineq = omega_add_zero_geq (pb, omega_black);
3596 pb->geqs[ineq].coef[i + 1] = -1;
3597 pb->geqs[ineq].coef[0] = nbi;
3601 omega_extract_distance_vectors (pb, ddr);
3603 return true;
3606 /* Sets up the Omega dependence problem for the data dependence
3607 relation DDR. Returns false when the constraint system cannot be
3608 built, ie. when the test answers "don't know". Returns true
3609 otherwise, and when independence has been proved (using one of the
3610 trivial dependence test), set MAYBE_DEPENDENT to false, otherwise
3611 set MAYBE_DEPENDENT to true.
3613 Example: for setting up the dependence system corresponding to the
3614 conflicting accesses
3616 | loop_i
3617 | loop_j
3618 | A[i, i+1] = ...
3619 | ... A[2*j, 2*(i + j)]
3620 | endloop_j
3621 | endloop_i
3623 the following constraints come from the iteration domain:
3625 0 <= i <= Ni
3626 0 <= i + di <= Ni
3627 0 <= j <= Nj
3628 0 <= j + dj <= Nj
3630 where di, dj are the distance variables. The constraints
3631 representing the conflicting elements are:
3633 i = 2 * (j + dj)
3634 i + 1 = 2 * (i + di + j + dj)
3636 For asking that the resulting distance vector (di, dj) be
3637 lexicographically positive, we insert the constraint "di >= 0". If
3638 "di = 0" in the solution, we fix that component to zero, and we
3639 look at the inner loops: we set a new problem where all the outer
3640 loop distances are zero, and fix this inner component to be
3641 positive. When one of the components is positive, we save that
3642 distance, and set a new problem where the distance on this loop is
3643 zero, searching for other distances in the inner loops. Here is
3644 the classic example that illustrates that we have to set for each
3645 inner loop a new problem:
3647 | loop_1
3648 | loop_2
3649 | A[10]
3650 | endloop_2
3651 | endloop_1
3653 we have to save two distances (1, 0) and (0, 1).
3655 Given two array references, refA and refB, we have to set the
3656 dependence problem twice, refA vs. refB and refB vs. refA, and we
3657 cannot do a single test, as refB might occur before refA in the
3658 inner loops, and the contrary when considering outer loops: ex.
3660 | loop_0
3661 | loop_1
3662 | loop_2
3663 | T[{1,+,1}_2][{1,+,1}_1] // refA
3664 | T[{2,+,1}_2][{0,+,1}_1] // refB
3665 | endloop_2
3666 | endloop_1
3667 | endloop_0
3669 refB touches the elements in T before refA, and thus for the same
3670 loop_0 refB precedes refA: ie. the distance vector (0, 1, -1)
3671 but for successive loop_0 iterations, we have (1, -1, 1)
3673 The Omega solver expects the distance variables ("di" in the
3674 previous example) to come first in the constraint system (as
3675 variables to be protected, or "safe" variables), the constraint
3676 system is built using the following layout:
3678 "cst | distance vars | index vars".
3681 static bool
3682 init_omega_for_ddr (struct data_dependence_relation *ddr,
3683 bool *maybe_dependent)
3685 omega_pb pb;
3686 bool res = false;
3688 *maybe_dependent = true;
3690 if (same_access_functions (ddr))
3692 unsigned j;
3693 lambda_vector dir_v;
3695 /* Save the 0 vector. */
3696 save_dist_v (ddr, lambda_vector_new (DDR_NB_LOOPS (ddr)));
3697 dir_v = lambda_vector_new (DDR_NB_LOOPS (ddr));
3698 for (j = 0; j < DDR_NB_LOOPS (ddr); j++)
3699 dir_v[j] = dir_equal;
3700 save_dir_v (ddr, dir_v);
3702 /* Save the dependences carried by outer loops. */
3703 pb = omega_alloc_problem (2 * DDR_NB_LOOPS (ddr), DDR_NB_LOOPS (ddr));
3704 res = init_omega_for_ddr_1 (DDR_A (ddr), DDR_B (ddr), ddr, pb,
3705 maybe_dependent);
3706 omega_free_problem (pb);
3707 return res;
3710 /* Omega expects the protected variables (those that have to be kept
3711 after elimination) to appear first in the constraint system.
3712 These variables are the distance variables. In the following
3713 initialization we declare NB_LOOPS safe variables, and the total
3714 number of variables for the constraint system is 2*NB_LOOPS. */
3715 pb = omega_alloc_problem (2 * DDR_NB_LOOPS (ddr), DDR_NB_LOOPS (ddr));
3716 res = init_omega_for_ddr_1 (DDR_A (ddr), DDR_B (ddr), ddr, pb,
3717 maybe_dependent);
3718 omega_free_problem (pb);
3720 /* Stop computation if not decidable, or no dependence. */
3721 if (res == false || *maybe_dependent == false)
3722 return res;
3724 pb = omega_alloc_problem (2 * DDR_NB_LOOPS (ddr), DDR_NB_LOOPS (ddr));
3725 res = init_omega_for_ddr_1 (DDR_B (ddr), DDR_A (ddr), ddr, pb,
3726 maybe_dependent);
3727 omega_free_problem (pb);
3729 return res;
3732 /* Return true when DDR contains the same information as that stored
3733 in DIR_VECTS and in DIST_VECTS, return false otherwise. */
3735 static bool
3736 ddr_consistent_p (FILE *file,
3737 struct data_dependence_relation *ddr,
3738 VEC (lambda_vector, heap) *dist_vects,
3739 VEC (lambda_vector, heap) *dir_vects)
3741 unsigned int i, j;
3743 /* If dump_file is set, output there. */
3744 if (dump_file && (dump_flags & TDF_DETAILS))
3745 file = dump_file;
3747 if (VEC_length (lambda_vector, dist_vects) != DDR_NUM_DIST_VECTS (ddr))
3749 lambda_vector b_dist_v;
3750 fprintf (file, "\n(Number of distance vectors differ: Banerjee has %d, Omega has %d.\n",
3751 VEC_length (lambda_vector, dist_vects),
3752 DDR_NUM_DIST_VECTS (ddr));
3754 fprintf (file, "Banerjee dist vectors:\n");
3755 FOR_EACH_VEC_ELT (lambda_vector, dist_vects, i, b_dist_v)
3756 print_lambda_vector (file, b_dist_v, DDR_NB_LOOPS (ddr));
3758 fprintf (file, "Omega dist vectors:\n");
3759 for (i = 0; i < DDR_NUM_DIST_VECTS (ddr); i++)
3760 print_lambda_vector (file, DDR_DIST_VECT (ddr, i), DDR_NB_LOOPS (ddr));
3762 fprintf (file, "data dependence relation:\n");
3763 dump_data_dependence_relation (file, ddr);
3765 fprintf (file, ")\n");
3766 return false;
3769 if (VEC_length (lambda_vector, dir_vects) != DDR_NUM_DIR_VECTS (ddr))
3771 fprintf (file, "\n(Number of direction vectors differ: Banerjee has %d, Omega has %d.)\n",
3772 VEC_length (lambda_vector, dir_vects),
3773 DDR_NUM_DIR_VECTS (ddr));
3774 return false;
3777 for (i = 0; i < DDR_NUM_DIST_VECTS (ddr); i++)
3779 lambda_vector a_dist_v;
3780 lambda_vector b_dist_v = DDR_DIST_VECT (ddr, i);
3782 /* Distance vectors are not ordered in the same way in the DDR
3783 and in the DIST_VECTS: search for a matching vector. */
3784 FOR_EACH_VEC_ELT (lambda_vector, dist_vects, j, a_dist_v)
3785 if (lambda_vector_equal (a_dist_v, b_dist_v, DDR_NB_LOOPS (ddr)))
3786 break;
3788 if (j == VEC_length (lambda_vector, dist_vects))
3790 fprintf (file, "\n(Dist vectors from the first dependence analyzer:\n");
3791 print_dist_vectors (file, dist_vects, DDR_NB_LOOPS (ddr));
3792 fprintf (file, "not found in Omega dist vectors:\n");
3793 print_dist_vectors (file, DDR_DIST_VECTS (ddr), DDR_NB_LOOPS (ddr));
3794 fprintf (file, "data dependence relation:\n");
3795 dump_data_dependence_relation (file, ddr);
3796 fprintf (file, ")\n");
3800 for (i = 0; i < DDR_NUM_DIR_VECTS (ddr); i++)
3802 lambda_vector a_dir_v;
3803 lambda_vector b_dir_v = DDR_DIR_VECT (ddr, i);
3805 /* Direction vectors are not ordered in the same way in the DDR
3806 and in the DIR_VECTS: search for a matching vector. */
3807 FOR_EACH_VEC_ELT (lambda_vector, dir_vects, j, a_dir_v)
3808 if (lambda_vector_equal (a_dir_v, b_dir_v, DDR_NB_LOOPS (ddr)))
3809 break;
3811 if (j == VEC_length (lambda_vector, dist_vects))
3813 fprintf (file, "\n(Dir vectors from the first dependence analyzer:\n");
3814 print_dir_vectors (file, dir_vects, DDR_NB_LOOPS (ddr));
3815 fprintf (file, "not found in Omega dir vectors:\n");
3816 print_dir_vectors (file, DDR_DIR_VECTS (ddr), DDR_NB_LOOPS (ddr));
3817 fprintf (file, "data dependence relation:\n");
3818 dump_data_dependence_relation (file, ddr);
3819 fprintf (file, ")\n");
3823 return true;
3826 /* This computes the affine dependence relation between A and B with
3827 respect to LOOP_NEST. CHREC_KNOWN is used for representing the
3828 independence between two accesses, while CHREC_DONT_KNOW is used
3829 for representing the unknown relation.
3831 Note that it is possible to stop the computation of the dependence
3832 relation the first time we detect a CHREC_KNOWN element for a given
3833 subscript. */
3835 static void
3836 compute_affine_dependence (struct data_dependence_relation *ddr,
3837 struct loop *loop_nest)
3839 struct data_reference *dra = DDR_A (ddr);
3840 struct data_reference *drb = DDR_B (ddr);
3842 if (dump_file && (dump_flags & TDF_DETAILS))
3844 fprintf (dump_file, "(compute_affine_dependence\n");
3845 fprintf (dump_file, " (stmt_a = \n");
3846 print_gimple_stmt (dump_file, DR_STMT (dra), 0, 0);
3847 fprintf (dump_file, ")\n (stmt_b = \n");
3848 print_gimple_stmt (dump_file, DR_STMT (drb), 0, 0);
3849 fprintf (dump_file, ")\n");
3852 /* Analyze only when the dependence relation is not yet known. */
3853 if (DDR_ARE_DEPENDENT (ddr) == NULL_TREE
3854 && !DDR_SELF_REFERENCE (ddr))
3856 dependence_stats.num_dependence_tests++;
3858 if (access_functions_are_affine_or_constant_p (dra, loop_nest)
3859 && access_functions_are_affine_or_constant_p (drb, loop_nest))
3861 if (flag_check_data_deps)
3863 /* Compute the dependences using the first algorithm. */
3864 subscript_dependence_tester (ddr, loop_nest);
3866 if (dump_file && (dump_flags & TDF_DETAILS))
3868 fprintf (dump_file, "\n\nBanerjee Analyzer\n");
3869 dump_data_dependence_relation (dump_file, ddr);
3872 if (DDR_ARE_DEPENDENT (ddr) == NULL_TREE)
3874 bool maybe_dependent;
3875 VEC (lambda_vector, heap) *dir_vects, *dist_vects;
3877 /* Save the result of the first DD analyzer. */
3878 dist_vects = DDR_DIST_VECTS (ddr);
3879 dir_vects = DDR_DIR_VECTS (ddr);
3881 /* Reset the information. */
3882 DDR_DIST_VECTS (ddr) = NULL;
3883 DDR_DIR_VECTS (ddr) = NULL;
3885 /* Compute the same information using Omega. */
3886 if (!init_omega_for_ddr (ddr, &maybe_dependent))
3887 goto csys_dont_know;
3889 if (dump_file && (dump_flags & TDF_DETAILS))
3891 fprintf (dump_file, "Omega Analyzer\n");
3892 dump_data_dependence_relation (dump_file, ddr);
3895 /* Check that we get the same information. */
3896 if (maybe_dependent)
3897 gcc_assert (ddr_consistent_p (stderr, ddr, dist_vects,
3898 dir_vects));
3901 else
3902 subscript_dependence_tester (ddr, loop_nest);
3905 /* As a last case, if the dependence cannot be determined, or if
3906 the dependence is considered too difficult to determine, answer
3907 "don't know". */
3908 else
3910 csys_dont_know:;
3911 dependence_stats.num_dependence_undetermined++;
3913 if (dump_file && (dump_flags & TDF_DETAILS))
3915 fprintf (dump_file, "Data ref a:\n");
3916 dump_data_reference (dump_file, dra);
3917 fprintf (dump_file, "Data ref b:\n");
3918 dump_data_reference (dump_file, drb);
3919 fprintf (dump_file, "affine dependence test not usable: access function not affine or constant.\n");
3921 finalize_ddr_dependent (ddr, chrec_dont_know);
3925 if (dump_file && (dump_flags & TDF_DETAILS))
3926 fprintf (dump_file, ")\n");
3929 /* This computes the dependence relation for the same data
3930 reference into DDR. */
3932 static void
3933 compute_self_dependence (struct data_dependence_relation *ddr)
3935 unsigned int i;
3936 struct subscript *subscript;
3938 if (DDR_ARE_DEPENDENT (ddr) != NULL_TREE)
3939 return;
3941 for (i = 0; VEC_iterate (subscript_p, DDR_SUBSCRIPTS (ddr), i, subscript);
3942 i++)
3944 if (SUB_CONFLICTS_IN_A (subscript))
3945 free_conflict_function (SUB_CONFLICTS_IN_A (subscript));
3946 if (SUB_CONFLICTS_IN_B (subscript))
3947 free_conflict_function (SUB_CONFLICTS_IN_B (subscript));
3949 /* The accessed index overlaps for each iteration. */
3950 SUB_CONFLICTS_IN_A (subscript)
3951 = conflict_fn (1, affine_fn_cst (integer_zero_node));
3952 SUB_CONFLICTS_IN_B (subscript)
3953 = conflict_fn (1, affine_fn_cst (integer_zero_node));
3954 SUB_LAST_CONFLICT (subscript) = chrec_dont_know;
3957 /* The distance vector is the zero vector. */
3958 save_dist_v (ddr, lambda_vector_new (DDR_NB_LOOPS (ddr)));
3959 save_dir_v (ddr, lambda_vector_new (DDR_NB_LOOPS (ddr)));
3962 /* Compute in DEPENDENCE_RELATIONS the data dependence graph for all
3963 the data references in DATAREFS, in the LOOP_NEST. When
3964 COMPUTE_SELF_AND_RR is FALSE, don't compute read-read and self
3965 relations. */
3967 void
3968 compute_all_dependences (VEC (data_reference_p, heap) *datarefs,
3969 VEC (ddr_p, heap) **dependence_relations,
3970 VEC (loop_p, heap) *loop_nest,
3971 bool compute_self_and_rr)
3973 struct data_dependence_relation *ddr;
3974 struct data_reference *a, *b;
3975 unsigned int i, j;
3977 FOR_EACH_VEC_ELT (data_reference_p, datarefs, i, a)
3978 for (j = i + 1; VEC_iterate (data_reference_p, datarefs, j, b); j++)
3979 if (DR_IS_WRITE (a) || DR_IS_WRITE (b) || compute_self_and_rr)
3981 ddr = initialize_data_dependence_relation (a, b, loop_nest);
3982 VEC_safe_push (ddr_p, heap, *dependence_relations, ddr);
3983 if (loop_nest)
3984 compute_affine_dependence (ddr, VEC_index (loop_p, loop_nest, 0));
3987 if (compute_self_and_rr)
3988 FOR_EACH_VEC_ELT (data_reference_p, datarefs, i, a)
3990 ddr = initialize_data_dependence_relation (a, a, loop_nest);
3991 VEC_safe_push (ddr_p, heap, *dependence_relations, ddr);
3992 compute_self_dependence (ddr);
3996 /* Stores the locations of memory references in STMT to REFERENCES. Returns
3997 true if STMT clobbers memory, false otherwise. */
3999 bool
4000 get_references_in_stmt (gimple stmt, VEC (data_ref_loc, heap) **references)
4002 bool clobbers_memory = false;
4003 data_ref_loc *ref;
4004 tree *op0, *op1;
4005 enum gimple_code stmt_code = gimple_code (stmt);
4007 *references = NULL;
4009 /* ASM_EXPR and CALL_EXPR may embed arbitrary side effects.
4010 Calls have side-effects, except those to const or pure
4011 functions. */
4012 if ((stmt_code == GIMPLE_CALL
4013 && !(gimple_call_flags (stmt) & (ECF_CONST | ECF_PURE)))
4014 || (stmt_code == GIMPLE_ASM
4015 && gimple_asm_volatile_p (stmt)))
4016 clobbers_memory = true;
4018 if (!gimple_vuse (stmt))
4019 return clobbers_memory;
4021 if (stmt_code == GIMPLE_ASSIGN)
4023 tree base;
4024 op0 = gimple_assign_lhs_ptr (stmt);
4025 op1 = gimple_assign_rhs1_ptr (stmt);
4027 if (DECL_P (*op1)
4028 || (REFERENCE_CLASS_P (*op1)
4029 && (base = get_base_address (*op1))
4030 && TREE_CODE (base) != SSA_NAME))
4032 ref = VEC_safe_push (data_ref_loc, heap, *references, NULL);
4033 ref->pos = op1;
4034 ref->is_read = true;
4037 if (DECL_P (*op0)
4038 || (REFERENCE_CLASS_P (*op0) && get_base_address (*op0)))
4040 ref = VEC_safe_push (data_ref_loc, heap, *references, NULL);
4041 ref->pos = op0;
4042 ref->is_read = false;
4045 else if (stmt_code == GIMPLE_CALL)
4047 unsigned i, n = gimple_call_num_args (stmt);
4049 for (i = 0; i < n; i++)
4051 op0 = gimple_call_arg_ptr (stmt, i);
4053 if (DECL_P (*op0)
4054 || (REFERENCE_CLASS_P (*op0) && get_base_address (*op0)))
4056 ref = VEC_safe_push (data_ref_loc, heap, *references, NULL);
4057 ref->pos = op0;
4058 ref->is_read = true;
4063 return clobbers_memory;
4066 /* Stores the data references in STMT to DATAREFS. If there is an unanalyzable
4067 reference, returns false, otherwise returns true. NEST is the outermost
4068 loop of the loop nest in which the references should be analyzed. */
4070 bool
4071 find_data_references_in_stmt (struct loop *nest, gimple stmt,
4072 VEC (data_reference_p, heap) **datarefs)
4074 unsigned i;
4075 VEC (data_ref_loc, heap) *references;
4076 data_ref_loc *ref;
4077 bool ret = true;
4078 data_reference_p dr;
4080 if (get_references_in_stmt (stmt, &references))
4082 VEC_free (data_ref_loc, heap, references);
4083 return false;
4086 FOR_EACH_VEC_ELT (data_ref_loc, references, i, ref)
4088 dr = create_data_ref (nest, *ref->pos, stmt, ref->is_read);
4089 gcc_assert (dr != NULL);
4091 /* FIXME -- data dependence analysis does not work correctly for objects
4092 with invariant addresses in loop nests. Let us fail here until the
4093 problem is fixed. */
4094 if (dr_address_invariant_p (dr) && nest)
4096 free_data_ref (dr);
4097 if (dump_file && (dump_flags & TDF_DETAILS))
4098 fprintf (dump_file, "\tFAILED as dr address is invariant\n");
4099 ret = false;
4100 break;
4103 VEC_safe_push (data_reference_p, heap, *datarefs, dr);
4105 VEC_free (data_ref_loc, heap, references);
4106 return ret;
4109 /* Stores the data references in STMT to DATAREFS. If there is an unanalyzable
4110 reference, returns false, otherwise returns true. NEST is the outermost
4111 loop of the loop nest in which the references should be analyzed. */
4113 bool
4114 graphite_find_data_references_in_stmt (struct loop *nest, gimple stmt,
4115 VEC (data_reference_p, heap) **datarefs)
4117 unsigned i;
4118 VEC (data_ref_loc, heap) *references;
4119 data_ref_loc *ref;
4120 bool ret = true;
4121 data_reference_p dr;
4123 if (get_references_in_stmt (stmt, &references))
4125 VEC_free (data_ref_loc, heap, references);
4126 return false;
4129 FOR_EACH_VEC_ELT (data_ref_loc, references, i, ref)
4131 dr = create_data_ref (nest, *ref->pos, stmt, ref->is_read);
4132 gcc_assert (dr != NULL);
4133 VEC_safe_push (data_reference_p, heap, *datarefs, dr);
4136 VEC_free (data_ref_loc, heap, references);
4137 return ret;
4140 /* Search the data references in LOOP, and record the information into
4141 DATAREFS. Returns chrec_dont_know when failing to analyze a
4142 difficult case, returns NULL_TREE otherwise. */
4144 static tree
4145 find_data_references_in_bb (struct loop *loop, basic_block bb,
4146 VEC (data_reference_p, heap) **datarefs)
4148 gimple_stmt_iterator bsi;
4150 for (bsi = gsi_start_bb (bb); !gsi_end_p (bsi); gsi_next (&bsi))
4152 gimple stmt = gsi_stmt (bsi);
4154 if (!find_data_references_in_stmt (loop, stmt, datarefs))
4156 struct data_reference *res;
4157 res = XCNEW (struct data_reference);
4158 VEC_safe_push (data_reference_p, heap, *datarefs, res);
4160 return chrec_dont_know;
4164 return NULL_TREE;
4167 /* Search the data references in LOOP, and record the information into
4168 DATAREFS. Returns chrec_dont_know when failing to analyze a
4169 difficult case, returns NULL_TREE otherwise.
4171 TODO: This function should be made smarter so that it can handle address
4172 arithmetic as if they were array accesses, etc. */
4174 tree
4175 find_data_references_in_loop (struct loop *loop,
4176 VEC (data_reference_p, heap) **datarefs)
4178 basic_block bb, *bbs;
4179 unsigned int i;
4181 bbs = get_loop_body_in_dom_order (loop);
4183 for (i = 0; i < loop->num_nodes; i++)
4185 bb = bbs[i];
4187 if (find_data_references_in_bb (loop, bb, datarefs) == chrec_dont_know)
4189 free (bbs);
4190 return chrec_dont_know;
4193 free (bbs);
4195 return NULL_TREE;
4198 /* Recursive helper function. */
4200 static bool
4201 find_loop_nest_1 (struct loop *loop, VEC (loop_p, heap) **loop_nest)
4203 /* Inner loops of the nest should not contain siblings. Example:
4204 when there are two consecutive loops,
4206 | loop_0
4207 | loop_1
4208 | A[{0, +, 1}_1]
4209 | endloop_1
4210 | loop_2
4211 | A[{0, +, 1}_2]
4212 | endloop_2
4213 | endloop_0
4215 the dependence relation cannot be captured by the distance
4216 abstraction. */
4217 if (loop->next)
4218 return false;
4220 VEC_safe_push (loop_p, heap, *loop_nest, loop);
4221 if (loop->inner)
4222 return find_loop_nest_1 (loop->inner, loop_nest);
4223 return true;
4226 /* Return false when the LOOP is not well nested. Otherwise return
4227 true and insert in LOOP_NEST the loops of the nest. LOOP_NEST will
4228 contain the loops from the outermost to the innermost, as they will
4229 appear in the classic distance vector. */
4231 bool
4232 find_loop_nest (struct loop *loop, VEC (loop_p, heap) **loop_nest)
4234 VEC_safe_push (loop_p, heap, *loop_nest, loop);
4235 if (loop->inner)
4236 return find_loop_nest_1 (loop->inner, loop_nest);
4237 return true;
4240 /* Returns true when the data dependences have been computed, false otherwise.
4241 Given a loop nest LOOP, the following vectors are returned:
4242 DATAREFS is initialized to all the array elements contained in this loop,
4243 DEPENDENCE_RELATIONS contains the relations between the data references.
4244 Compute read-read and self relations if
4245 COMPUTE_SELF_AND_READ_READ_DEPENDENCES is TRUE. */
4247 bool
4248 compute_data_dependences_for_loop (struct loop *loop,
4249 bool compute_self_and_read_read_dependences,
4250 VEC (data_reference_p, heap) **datarefs,
4251 VEC (ddr_p, heap) **dependence_relations)
4253 bool res = true;
4254 VEC (loop_p, heap) *vloops = VEC_alloc (loop_p, heap, 3);
4256 memset (&dependence_stats, 0, sizeof (dependence_stats));
4258 /* If the loop nest is not well formed, or one of the data references
4259 is not computable, give up without spending time to compute other
4260 dependences. */
4261 if (!loop
4262 || !find_loop_nest (loop, &vloops)
4263 || find_data_references_in_loop (loop, datarefs) == chrec_dont_know)
4265 struct data_dependence_relation *ddr;
4267 /* Insert a single relation into dependence_relations:
4268 chrec_dont_know. */
4269 ddr = initialize_data_dependence_relation (NULL, NULL, vloops);
4270 VEC_safe_push (ddr_p, heap, *dependence_relations, ddr);
4271 res = false;
4273 else
4274 compute_all_dependences (*datarefs, dependence_relations, vloops,
4275 compute_self_and_read_read_dependences);
4277 if (dump_file && (dump_flags & TDF_STATS))
4279 fprintf (dump_file, "Dependence tester statistics:\n");
4281 fprintf (dump_file, "Number of dependence tests: %d\n",
4282 dependence_stats.num_dependence_tests);
4283 fprintf (dump_file, "Number of dependence tests classified dependent: %d\n",
4284 dependence_stats.num_dependence_dependent);
4285 fprintf (dump_file, "Number of dependence tests classified independent: %d\n",
4286 dependence_stats.num_dependence_independent);
4287 fprintf (dump_file, "Number of undetermined dependence tests: %d\n",
4288 dependence_stats.num_dependence_undetermined);
4290 fprintf (dump_file, "Number of subscript tests: %d\n",
4291 dependence_stats.num_subscript_tests);
4292 fprintf (dump_file, "Number of undetermined subscript tests: %d\n",
4293 dependence_stats.num_subscript_undetermined);
4294 fprintf (dump_file, "Number of same subscript function: %d\n",
4295 dependence_stats.num_same_subscript_function);
4297 fprintf (dump_file, "Number of ziv tests: %d\n",
4298 dependence_stats.num_ziv);
4299 fprintf (dump_file, "Number of ziv tests returning dependent: %d\n",
4300 dependence_stats.num_ziv_dependent);
4301 fprintf (dump_file, "Number of ziv tests returning independent: %d\n",
4302 dependence_stats.num_ziv_independent);
4303 fprintf (dump_file, "Number of ziv tests unimplemented: %d\n",
4304 dependence_stats.num_ziv_unimplemented);
4306 fprintf (dump_file, "Number of siv tests: %d\n",
4307 dependence_stats.num_siv);
4308 fprintf (dump_file, "Number of siv tests returning dependent: %d\n",
4309 dependence_stats.num_siv_dependent);
4310 fprintf (dump_file, "Number of siv tests returning independent: %d\n",
4311 dependence_stats.num_siv_independent);
4312 fprintf (dump_file, "Number of siv tests unimplemented: %d\n",
4313 dependence_stats.num_siv_unimplemented);
4315 fprintf (dump_file, "Number of miv tests: %d\n",
4316 dependence_stats.num_miv);
4317 fprintf (dump_file, "Number of miv tests returning dependent: %d\n",
4318 dependence_stats.num_miv_dependent);
4319 fprintf (dump_file, "Number of miv tests returning independent: %d\n",
4320 dependence_stats.num_miv_independent);
4321 fprintf (dump_file, "Number of miv tests unimplemented: %d\n",
4322 dependence_stats.num_miv_unimplemented);
4325 return res;
4328 /* Returns true when the data dependences for the basic block BB have been
4329 computed, false otherwise.
4330 DATAREFS is initialized to all the array elements contained in this basic
4331 block, DEPENDENCE_RELATIONS contains the relations between the data
4332 references. Compute read-read and self relations if
4333 COMPUTE_SELF_AND_READ_READ_DEPENDENCES is TRUE. */
4334 bool
4335 compute_data_dependences_for_bb (basic_block bb,
4336 bool compute_self_and_read_read_dependences,
4337 VEC (data_reference_p, heap) **datarefs,
4338 VEC (ddr_p, heap) **dependence_relations)
4340 if (find_data_references_in_bb (NULL, bb, datarefs) == chrec_dont_know)
4341 return false;
4343 compute_all_dependences (*datarefs, dependence_relations, NULL,
4344 compute_self_and_read_read_dependences);
4345 return true;
4348 /* Entry point (for testing only). Analyze all the data references
4349 and the dependence relations in LOOP.
4351 The data references are computed first.
4353 A relation on these nodes is represented by a complete graph. Some
4354 of the relations could be of no interest, thus the relations can be
4355 computed on demand.
4357 In the following function we compute all the relations. This is
4358 just a first implementation that is here for:
4359 - for showing how to ask for the dependence relations,
4360 - for the debugging the whole dependence graph,
4361 - for the dejagnu testcases and maintenance.
4363 It is possible to ask only for a part of the graph, avoiding to
4364 compute the whole dependence graph. The computed dependences are
4365 stored in a knowledge base (KB) such that later queries don't
4366 recompute the same information. The implementation of this KB is
4367 transparent to the optimizer, and thus the KB can be changed with a
4368 more efficient implementation, or the KB could be disabled. */
4369 static void
4370 analyze_all_data_dependences (struct loop *loop)
4372 unsigned int i;
4373 int nb_data_refs = 10;
4374 VEC (data_reference_p, heap) *datarefs =
4375 VEC_alloc (data_reference_p, heap, nb_data_refs);
4376 VEC (ddr_p, heap) *dependence_relations =
4377 VEC_alloc (ddr_p, heap, nb_data_refs * nb_data_refs);
4379 /* Compute DDs on the whole function. */
4380 compute_data_dependences_for_loop (loop, false, &datarefs,
4381 &dependence_relations);
4383 if (dump_file)
4385 dump_data_dependence_relations (dump_file, dependence_relations);
4386 fprintf (dump_file, "\n\n");
4388 if (dump_flags & TDF_DETAILS)
4389 dump_dist_dir_vectors (dump_file, dependence_relations);
4391 if (dump_flags & TDF_STATS)
4393 unsigned nb_top_relations = 0;
4394 unsigned nb_bot_relations = 0;
4395 unsigned nb_chrec_relations = 0;
4396 struct data_dependence_relation *ddr;
4398 FOR_EACH_VEC_ELT (ddr_p, dependence_relations, i, ddr)
4400 if (chrec_contains_undetermined (DDR_ARE_DEPENDENT (ddr)))
4401 nb_top_relations++;
4403 else if (DDR_ARE_DEPENDENT (ddr) == chrec_known)
4404 nb_bot_relations++;
4406 else
4407 nb_chrec_relations++;
4410 gather_stats_on_scev_database ();
4414 free_dependence_relations (dependence_relations);
4415 free_data_refs (datarefs);
4418 /* Computes all the data dependences and check that the results of
4419 several analyzers are the same. */
4421 void
4422 tree_check_data_deps (void)
4424 loop_iterator li;
4425 struct loop *loop_nest;
4427 FOR_EACH_LOOP (li, loop_nest, 0)
4428 analyze_all_data_dependences (loop_nest);
4431 /* Free the memory used by a data dependence relation DDR. */
4433 void
4434 free_dependence_relation (struct data_dependence_relation *ddr)
4436 if (ddr == NULL)
4437 return;
4439 if (DDR_SUBSCRIPTS (ddr))
4440 free_subscripts (DDR_SUBSCRIPTS (ddr));
4441 if (DDR_DIST_VECTS (ddr))
4442 VEC_free (lambda_vector, heap, DDR_DIST_VECTS (ddr));
4443 if (DDR_DIR_VECTS (ddr))
4444 VEC_free (lambda_vector, heap, DDR_DIR_VECTS (ddr));
4446 free (ddr);
4449 /* Free the memory used by the data dependence relations from
4450 DEPENDENCE_RELATIONS. */
4452 void
4453 free_dependence_relations (VEC (ddr_p, heap) *dependence_relations)
4455 unsigned int i;
4456 struct data_dependence_relation *ddr;
4457 VEC (loop_p, heap) *loop_nest = NULL;
4459 FOR_EACH_VEC_ELT (ddr_p, dependence_relations, i, ddr)
4461 if (ddr == NULL)
4462 continue;
4463 if (loop_nest == NULL)
4464 loop_nest = DDR_LOOP_NEST (ddr);
4465 else
4466 gcc_assert (DDR_LOOP_NEST (ddr) == NULL
4467 || DDR_LOOP_NEST (ddr) == loop_nest);
4468 free_dependence_relation (ddr);
4471 if (loop_nest)
4472 VEC_free (loop_p, heap, loop_nest);
4473 VEC_free (ddr_p, heap, dependence_relations);
4476 /* Free the memory used by the data references from DATAREFS. */
4478 void
4479 free_data_refs (VEC (data_reference_p, heap) *datarefs)
4481 unsigned int i;
4482 struct data_reference *dr;
4484 FOR_EACH_VEC_ELT (data_reference_p, datarefs, i, dr)
4485 free_data_ref (dr);
4486 VEC_free (data_reference_p, heap, datarefs);
4491 /* Dump vertex I in RDG to FILE. */
4493 void
4494 dump_rdg_vertex (FILE *file, struct graph *rdg, int i)
4496 struct vertex *v = &(rdg->vertices[i]);
4497 struct graph_edge *e;
4499 fprintf (file, "(vertex %d: (%s%s) (in:", i,
4500 RDG_MEM_WRITE_STMT (rdg, i) ? "w" : "",
4501 RDG_MEM_READS_STMT (rdg, i) ? "r" : "");
4503 if (v->pred)
4504 for (e = v->pred; e; e = e->pred_next)
4505 fprintf (file, " %d", e->src);
4507 fprintf (file, ") (out:");
4509 if (v->succ)
4510 for (e = v->succ; e; e = e->succ_next)
4511 fprintf (file, " %d", e->dest);
4513 fprintf (file, ")\n");
4514 print_gimple_stmt (file, RDGV_STMT (v), 0, TDF_VOPS|TDF_MEMSYMS);
4515 fprintf (file, ")\n");
4518 /* Call dump_rdg_vertex on stderr. */
4520 DEBUG_FUNCTION void
4521 debug_rdg_vertex (struct graph *rdg, int i)
4523 dump_rdg_vertex (stderr, rdg, i);
4526 /* Dump component C of RDG to FILE. If DUMPED is non-null, set the
4527 dumped vertices to that bitmap. */
4529 void dump_rdg_component (FILE *file, struct graph *rdg, int c, bitmap dumped)
4531 int i;
4533 fprintf (file, "(%d\n", c);
4535 for (i = 0; i < rdg->n_vertices; i++)
4536 if (rdg->vertices[i].component == c)
4538 if (dumped)
4539 bitmap_set_bit (dumped, i);
4541 dump_rdg_vertex (file, rdg, i);
4544 fprintf (file, ")\n");
4547 /* Call dump_rdg_vertex on stderr. */
4549 DEBUG_FUNCTION void
4550 debug_rdg_component (struct graph *rdg, int c)
4552 dump_rdg_component (stderr, rdg, c, NULL);
4555 /* Dump the reduced dependence graph RDG to FILE. */
4557 void
4558 dump_rdg (FILE *file, struct graph *rdg)
4560 int i;
4561 bitmap dumped = BITMAP_ALLOC (NULL);
4563 fprintf (file, "(rdg\n");
4565 for (i = 0; i < rdg->n_vertices; i++)
4566 if (!bitmap_bit_p (dumped, i))
4567 dump_rdg_component (file, rdg, rdg->vertices[i].component, dumped);
4569 fprintf (file, ")\n");
4570 BITMAP_FREE (dumped);
4573 /* Call dump_rdg on stderr. */
4575 DEBUG_FUNCTION void
4576 debug_rdg (struct graph *rdg)
4578 dump_rdg (stderr, rdg);
4581 static void
4582 dot_rdg_1 (FILE *file, struct graph *rdg)
4584 int i;
4586 fprintf (file, "digraph RDG {\n");
4588 for (i = 0; i < rdg->n_vertices; i++)
4590 struct vertex *v = &(rdg->vertices[i]);
4591 struct graph_edge *e;
4593 /* Highlight reads from memory. */
4594 if (RDG_MEM_READS_STMT (rdg, i))
4595 fprintf (file, "%d [style=filled, fillcolor=green]\n", i);
4597 /* Highlight stores to memory. */
4598 if (RDG_MEM_WRITE_STMT (rdg, i))
4599 fprintf (file, "%d [style=filled, fillcolor=red]\n", i);
4601 if (v->succ)
4602 for (e = v->succ; e; e = e->succ_next)
4603 switch (RDGE_TYPE (e))
4605 case input_dd:
4606 fprintf (file, "%d -> %d [label=input] \n", i, e->dest);
4607 break;
4609 case output_dd:
4610 fprintf (file, "%d -> %d [label=output] \n", i, e->dest);
4611 break;
4613 case flow_dd:
4614 /* These are the most common dependences: don't print these. */
4615 fprintf (file, "%d -> %d \n", i, e->dest);
4616 break;
4618 case anti_dd:
4619 fprintf (file, "%d -> %d [label=anti] \n", i, e->dest);
4620 break;
4622 default:
4623 gcc_unreachable ();
4627 fprintf (file, "}\n\n");
4630 /* Display the Reduced Dependence Graph using dotty. */
4631 extern void dot_rdg (struct graph *);
4633 DEBUG_FUNCTION void
4634 dot_rdg (struct graph *rdg)
4636 /* When debugging, enable the following code. This cannot be used
4637 in production compilers because it calls "system". */
4638 #if 0
4639 FILE *file = fopen ("/tmp/rdg.dot", "w");
4640 gcc_assert (file != NULL);
4642 dot_rdg_1 (file, rdg);
4643 fclose (file);
4645 system ("dotty /tmp/rdg.dot &");
4646 #else
4647 dot_rdg_1 (stderr, rdg);
4648 #endif
4651 /* This structure is used for recording the mapping statement index in
4652 the RDG. */
4654 struct GTY(()) rdg_vertex_info
4656 gimple stmt;
4657 int index;
4660 /* Returns the index of STMT in RDG. */
4663 rdg_vertex_for_stmt (struct graph *rdg, gimple stmt)
4665 struct rdg_vertex_info rvi, *slot;
4667 rvi.stmt = stmt;
4668 slot = (struct rdg_vertex_info *) htab_find (rdg->indices, &rvi);
4670 if (!slot)
4671 return -1;
4673 return slot->index;
4676 /* Creates an edge in RDG for each distance vector from DDR. The
4677 order that we keep track of in the RDG is the order in which
4678 statements have to be executed. */
4680 static void
4681 create_rdg_edge_for_ddr (struct graph *rdg, ddr_p ddr)
4683 struct graph_edge *e;
4684 int va, vb;
4685 data_reference_p dra = DDR_A (ddr);
4686 data_reference_p drb = DDR_B (ddr);
4687 unsigned level = ddr_dependence_level (ddr);
4689 /* For non scalar dependences, when the dependence is REVERSED,
4690 statement B has to be executed before statement A. */
4691 if (level > 0
4692 && !DDR_REVERSED_P (ddr))
4694 data_reference_p tmp = dra;
4695 dra = drb;
4696 drb = tmp;
4699 va = rdg_vertex_for_stmt (rdg, DR_STMT (dra));
4700 vb = rdg_vertex_for_stmt (rdg, DR_STMT (drb));
4702 if (va < 0 || vb < 0)
4703 return;
4705 e = add_edge (rdg, va, vb);
4706 e->data = XNEW (struct rdg_edge);
4708 RDGE_LEVEL (e) = level;
4709 RDGE_RELATION (e) = ddr;
4711 /* Determines the type of the data dependence. */
4712 if (DR_IS_READ (dra) && DR_IS_READ (drb))
4713 RDGE_TYPE (e) = input_dd;
4714 else if (DR_IS_WRITE (dra) && DR_IS_WRITE (drb))
4715 RDGE_TYPE (e) = output_dd;
4716 else if (DR_IS_WRITE (dra) && DR_IS_READ (drb))
4717 RDGE_TYPE (e) = flow_dd;
4718 else if (DR_IS_READ (dra) && DR_IS_WRITE (drb))
4719 RDGE_TYPE (e) = anti_dd;
4722 /* Creates dependence edges in RDG for all the uses of DEF. IDEF is
4723 the index of DEF in RDG. */
4725 static void
4726 create_rdg_edges_for_scalar (struct graph *rdg, tree def, int idef)
4728 use_operand_p imm_use_p;
4729 imm_use_iterator iterator;
4731 FOR_EACH_IMM_USE_FAST (imm_use_p, iterator, def)
4733 struct graph_edge *e;
4734 int use = rdg_vertex_for_stmt (rdg, USE_STMT (imm_use_p));
4736 if (use < 0)
4737 continue;
4739 e = add_edge (rdg, idef, use);
4740 e->data = XNEW (struct rdg_edge);
4741 RDGE_TYPE (e) = flow_dd;
4742 RDGE_RELATION (e) = NULL;
4746 /* Creates the edges of the reduced dependence graph RDG. */
4748 static void
4749 create_rdg_edges (struct graph *rdg, VEC (ddr_p, heap) *ddrs)
4751 int i;
4752 struct data_dependence_relation *ddr;
4753 def_operand_p def_p;
4754 ssa_op_iter iter;
4756 FOR_EACH_VEC_ELT (ddr_p, ddrs, i, ddr)
4757 if (DDR_ARE_DEPENDENT (ddr) == NULL_TREE)
4758 create_rdg_edge_for_ddr (rdg, ddr);
4760 for (i = 0; i < rdg->n_vertices; i++)
4761 FOR_EACH_PHI_OR_STMT_DEF (def_p, RDG_STMT (rdg, i),
4762 iter, SSA_OP_DEF)
4763 create_rdg_edges_for_scalar (rdg, DEF_FROM_PTR (def_p), i);
4766 /* Build the vertices of the reduced dependence graph RDG. */
4768 void
4769 create_rdg_vertices (struct graph *rdg, VEC (gimple, heap) *stmts)
4771 int i, j;
4772 gimple stmt;
4774 FOR_EACH_VEC_ELT (gimple, stmts, i, stmt)
4776 VEC (data_ref_loc, heap) *references;
4777 data_ref_loc *ref;
4778 struct vertex *v = &(rdg->vertices[i]);
4779 struct rdg_vertex_info *rvi = XNEW (struct rdg_vertex_info);
4780 struct rdg_vertex_info **slot;
4782 rvi->stmt = stmt;
4783 rvi->index = i;
4784 slot = (struct rdg_vertex_info **) htab_find_slot (rdg->indices, rvi, INSERT);
4786 if (!*slot)
4787 *slot = rvi;
4788 else
4789 free (rvi);
4791 v->data = XNEW (struct rdg_vertex);
4792 RDG_STMT (rdg, i) = stmt;
4794 RDG_MEM_WRITE_STMT (rdg, i) = false;
4795 RDG_MEM_READS_STMT (rdg, i) = false;
4796 if (gimple_code (stmt) == GIMPLE_PHI)
4797 continue;
4799 get_references_in_stmt (stmt, &references);
4800 FOR_EACH_VEC_ELT (data_ref_loc, references, j, ref)
4801 if (!ref->is_read)
4802 RDG_MEM_WRITE_STMT (rdg, i) = true;
4803 else
4804 RDG_MEM_READS_STMT (rdg, i) = true;
4806 VEC_free (data_ref_loc, heap, references);
4810 /* Initialize STMTS with all the statements of LOOP. When
4811 INCLUDE_PHIS is true, include also the PHI nodes. The order in
4812 which we discover statements is important as
4813 generate_loops_for_partition is using the same traversal for
4814 identifying statements. */
4816 static void
4817 stmts_from_loop (struct loop *loop, VEC (gimple, heap) **stmts)
4819 unsigned int i;
4820 basic_block *bbs = get_loop_body_in_dom_order (loop);
4822 for (i = 0; i < loop->num_nodes; i++)
4824 basic_block bb = bbs[i];
4825 gimple_stmt_iterator bsi;
4826 gimple stmt;
4828 for (bsi = gsi_start_phis (bb); !gsi_end_p (bsi); gsi_next (&bsi))
4829 VEC_safe_push (gimple, heap, *stmts, gsi_stmt (bsi));
4831 for (bsi = gsi_start_bb (bb); !gsi_end_p (bsi); gsi_next (&bsi))
4833 stmt = gsi_stmt (bsi);
4834 if (gimple_code (stmt) != GIMPLE_LABEL)
4835 VEC_safe_push (gimple, heap, *stmts, stmt);
4839 free (bbs);
4842 /* Returns true when all the dependences are computable. */
4844 static bool
4845 known_dependences_p (VEC (ddr_p, heap) *dependence_relations)
4847 ddr_p ddr;
4848 unsigned int i;
4850 FOR_EACH_VEC_ELT (ddr_p, dependence_relations, i, ddr)
4851 if (DDR_ARE_DEPENDENT (ddr) == chrec_dont_know)
4852 return false;
4854 return true;
4857 /* Computes a hash function for element ELT. */
4859 static hashval_t
4860 hash_stmt_vertex_info (const void *elt)
4862 const struct rdg_vertex_info *const rvi =
4863 (const struct rdg_vertex_info *) elt;
4864 gimple stmt = rvi->stmt;
4866 return htab_hash_pointer (stmt);
4869 /* Compares database elements E1 and E2. */
4871 static int
4872 eq_stmt_vertex_info (const void *e1, const void *e2)
4874 const struct rdg_vertex_info *elt1 = (const struct rdg_vertex_info *) e1;
4875 const struct rdg_vertex_info *elt2 = (const struct rdg_vertex_info *) e2;
4877 return elt1->stmt == elt2->stmt;
4880 /* Free the element E. */
4882 static void
4883 hash_stmt_vertex_del (void *e)
4885 free (e);
4888 /* Build the Reduced Dependence Graph (RDG) with one vertex per
4889 statement of the loop nest, and one edge per data dependence or
4890 scalar dependence. */
4892 struct graph *
4893 build_empty_rdg (int n_stmts)
4895 int nb_data_refs = 10;
4896 struct graph *rdg = new_graph (n_stmts);
4898 rdg->indices = htab_create (nb_data_refs, hash_stmt_vertex_info,
4899 eq_stmt_vertex_info, hash_stmt_vertex_del);
4900 return rdg;
4903 /* Build the Reduced Dependence Graph (RDG) with one vertex per
4904 statement of the loop nest, and one edge per data dependence or
4905 scalar dependence. */
4907 struct graph *
4908 build_rdg (struct loop *loop)
4910 int nb_data_refs = 10;
4911 struct graph *rdg = NULL;
4912 VEC (ddr_p, heap) *dependence_relations;
4913 VEC (data_reference_p, heap) *datarefs;
4914 VEC (gimple, heap) *stmts = VEC_alloc (gimple, heap, nb_data_refs);
4916 dependence_relations = VEC_alloc (ddr_p, heap, nb_data_refs * nb_data_refs) ;
4917 datarefs = VEC_alloc (data_reference_p, heap, nb_data_refs);
4918 compute_data_dependences_for_loop (loop,
4919 false,
4920 &datarefs,
4921 &dependence_relations);
4923 if (!known_dependences_p (dependence_relations))
4925 free_dependence_relations (dependence_relations);
4926 free_data_refs (datarefs);
4927 VEC_free (gimple, heap, stmts);
4929 return rdg;
4932 stmts_from_loop (loop, &stmts);
4933 rdg = build_empty_rdg (VEC_length (gimple, stmts));
4935 rdg->indices = htab_create (nb_data_refs, hash_stmt_vertex_info,
4936 eq_stmt_vertex_info, hash_stmt_vertex_del);
4937 create_rdg_vertices (rdg, stmts);
4938 create_rdg_edges (rdg, dependence_relations);
4940 VEC_free (gimple, heap, stmts);
4941 return rdg;
4944 /* Free the reduced dependence graph RDG. */
4946 void
4947 free_rdg (struct graph *rdg)
4949 int i;
4951 for (i = 0; i < rdg->n_vertices; i++)
4952 free (rdg->vertices[i].data);
4954 htab_delete (rdg->indices);
4955 free_graph (rdg);
4958 /* Initialize STMTS with all the statements of LOOP that contain a
4959 store to memory. */
4961 void
4962 stores_from_loop (struct loop *loop, VEC (gimple, heap) **stmts)
4964 unsigned int i;
4965 basic_block *bbs = get_loop_body_in_dom_order (loop);
4967 for (i = 0; i < loop->num_nodes; i++)
4969 basic_block bb = bbs[i];
4970 gimple_stmt_iterator bsi;
4972 for (bsi = gsi_start_bb (bb); !gsi_end_p (bsi); gsi_next (&bsi))
4973 if (gimple_vdef (gsi_stmt (bsi)))
4974 VEC_safe_push (gimple, heap, *stmts, gsi_stmt (bsi));
4977 free (bbs);
4980 /* Returns true when the statement at STMT is of the form "A[i] = 0"
4981 that contains a data reference on its LHS with a stride of the same
4982 size as its unit type. */
4984 bool
4985 stmt_with_adjacent_zero_store_dr_p (gimple stmt)
4987 tree op0, op1;
4988 bool res;
4989 struct data_reference *dr;
4991 if (!stmt
4992 || !gimple_vdef (stmt)
4993 || !is_gimple_assign (stmt)
4994 || !gimple_assign_single_p (stmt)
4995 || !(op1 = gimple_assign_rhs1 (stmt))
4996 || !(integer_zerop (op1) || real_zerop (op1)))
4997 return false;
4999 dr = XCNEW (struct data_reference);
5000 op0 = gimple_assign_lhs (stmt);
5002 DR_STMT (dr) = stmt;
5003 DR_REF (dr) = op0;
5005 res = dr_analyze_innermost (dr)
5006 && stride_of_unit_type_p (DR_STEP (dr), TREE_TYPE (op0));
5008 free_data_ref (dr);
5009 return res;
5012 /* Initialize STMTS with all the statements of LOOP that contain a
5013 store to memory of the form "A[i] = 0". */
5015 void
5016 stores_zero_from_loop (struct loop *loop, VEC (gimple, heap) **stmts)
5018 unsigned int i;
5019 basic_block bb;
5020 gimple_stmt_iterator si;
5021 gimple stmt;
5022 basic_block *bbs = get_loop_body_in_dom_order (loop);
5024 for (i = 0; i < loop->num_nodes; i++)
5025 for (bb = bbs[i], si = gsi_start_bb (bb); !gsi_end_p (si); gsi_next (&si))
5026 if ((stmt = gsi_stmt (si))
5027 && stmt_with_adjacent_zero_store_dr_p (stmt))
5028 VEC_safe_push (gimple, heap, *stmts, gsi_stmt (si));
5030 free (bbs);
5033 /* For a data reference REF, return the declaration of its base
5034 address or NULL_TREE if the base is not determined. */
5036 static inline tree
5037 ref_base_address (gimple stmt, data_ref_loc *ref)
5039 tree base = NULL_TREE;
5040 tree base_address;
5041 struct data_reference *dr = XCNEW (struct data_reference);
5043 DR_STMT (dr) = stmt;
5044 DR_REF (dr) = *ref->pos;
5045 dr_analyze_innermost (dr);
5046 base_address = DR_BASE_ADDRESS (dr);
5048 if (!base_address)
5049 goto end;
5051 switch (TREE_CODE (base_address))
5053 case ADDR_EXPR:
5054 base = TREE_OPERAND (base_address, 0);
5055 break;
5057 default:
5058 base = base_address;
5059 break;
5062 end:
5063 free_data_ref (dr);
5064 return base;
5067 /* Determines whether the statement from vertex V of the RDG has a
5068 definition used outside the loop that contains this statement. */
5070 bool
5071 rdg_defs_used_in_other_loops_p (struct graph *rdg, int v)
5073 gimple stmt = RDG_STMT (rdg, v);
5074 struct loop *loop = loop_containing_stmt (stmt);
5075 use_operand_p imm_use_p;
5076 imm_use_iterator iterator;
5077 ssa_op_iter it;
5078 def_operand_p def_p;
5080 if (!loop)
5081 return true;
5083 FOR_EACH_PHI_OR_STMT_DEF (def_p, stmt, it, SSA_OP_DEF)
5085 FOR_EACH_IMM_USE_FAST (imm_use_p, iterator, DEF_FROM_PTR (def_p))
5087 if (loop_containing_stmt (USE_STMT (imm_use_p)) != loop)
5088 return true;
5092 return false;
5095 /* Determines whether statements S1 and S2 access to similar memory
5096 locations. Two memory accesses are considered similar when they
5097 have the same base address declaration, i.e. when their
5098 ref_base_address is the same. */
5100 bool
5101 have_similar_memory_accesses (gimple s1, gimple s2)
5103 bool res = false;
5104 unsigned i, j;
5105 VEC (data_ref_loc, heap) *refs1, *refs2;
5106 data_ref_loc *ref1, *ref2;
5108 get_references_in_stmt (s1, &refs1);
5109 get_references_in_stmt (s2, &refs2);
5111 FOR_EACH_VEC_ELT (data_ref_loc, refs1, i, ref1)
5113 tree base1 = ref_base_address (s1, ref1);
5115 if (base1)
5116 FOR_EACH_VEC_ELT (data_ref_loc, refs2, j, ref2)
5117 if (base1 == ref_base_address (s2, ref2))
5119 res = true;
5120 goto end;
5124 end:
5125 VEC_free (data_ref_loc, heap, refs1);
5126 VEC_free (data_ref_loc, heap, refs2);
5127 return res;
5130 /* Helper function for the hashtab. */
5132 static int
5133 have_similar_memory_accesses_1 (const void *s1, const void *s2)
5135 return have_similar_memory_accesses (CONST_CAST_GIMPLE ((const_gimple) s1),
5136 CONST_CAST_GIMPLE ((const_gimple) s2));
5139 /* Helper function for the hashtab. */
5141 static hashval_t
5142 ref_base_address_1 (const void *s)
5144 gimple stmt = CONST_CAST_GIMPLE ((const_gimple) s);
5145 unsigned i;
5146 VEC (data_ref_loc, heap) *refs;
5147 data_ref_loc *ref;
5148 hashval_t res = 0;
5150 get_references_in_stmt (stmt, &refs);
5152 FOR_EACH_VEC_ELT (data_ref_loc, refs, i, ref)
5153 if (!ref->is_read)
5155 res = htab_hash_pointer (ref_base_address (stmt, ref));
5156 break;
5159 VEC_free (data_ref_loc, heap, refs);
5160 return res;
5163 /* Try to remove duplicated write data references from STMTS. */
5165 void
5166 remove_similar_memory_refs (VEC (gimple, heap) **stmts)
5168 unsigned i;
5169 gimple stmt;
5170 htab_t seen = htab_create (VEC_length (gimple, *stmts), ref_base_address_1,
5171 have_similar_memory_accesses_1, NULL);
5173 for (i = 0; VEC_iterate (gimple, *stmts, i, stmt); )
5175 void **slot;
5177 slot = htab_find_slot (seen, stmt, INSERT);
5179 if (*slot)
5180 VEC_ordered_remove (gimple, *stmts, i);
5181 else
5183 *slot = (void *) stmt;
5184 i++;
5188 htab_delete (seen);
5191 /* Returns the index of PARAMETER in the parameters vector of the
5192 ACCESS_MATRIX. If PARAMETER does not exist return -1. */
5195 access_matrix_get_index_for_parameter (tree parameter,
5196 struct access_matrix *access_matrix)
5198 int i;
5199 VEC (tree,heap) *lambda_parameters = AM_PARAMETERS (access_matrix);
5200 tree lambda_parameter;
5202 FOR_EACH_VEC_ELT (tree, lambda_parameters, i, lambda_parameter)
5203 if (lambda_parameter == parameter)
5204 return i + AM_NB_INDUCTION_VARS (access_matrix);
5206 return -1;