1 /* Data references and dependences detectors.
2 Copyright (C) 2003, 2004, 2005, 2006, 2007 Free Software Foundation, Inc.
3 Contributed by Sebastian Pop <pop@cri.ensmp.fr>
5 This file is part of GCC.
7 GCC is free software; you can redistribute it and/or modify it under
8 the terms of the GNU General Public License as published by the Free
9 Software Foundation; either version 3, or (at your option) any later
12 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
13 WARRANTY; without even the implied warranty of MERCHANTABILITY or
14 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
17 You should have received a copy of the GNU General Public License
18 along with GCC; see the file COPYING3. If not see
19 <http://www.gnu.org/licenses/>. */
21 /* This pass walks a given loop structure searching for array
22 references. The information about the array accesses is recorded
23 in DATA_REFERENCE structures.
25 The basic test for determining the dependences is:
26 given two access functions chrec1 and chrec2 to a same array, and
27 x and y two vectors from the iteration domain, the same element of
28 the array is accessed twice at iterations x and y if and only if:
29 | chrec1 (x) == chrec2 (y).
31 The goals of this analysis are:
33 - to determine the independence: the relation between two
34 independent accesses is qualified with the chrec_known (this
35 information allows a loop parallelization),
37 - when two data references access the same data, to qualify the
38 dependence relation with classic dependence representations:
42 - loop carried level dependence
43 - polyhedron dependence
44 or with the chains of recurrences based representation,
46 - to define a knowledge base for storing the data dependence
49 - to define an interface to access this data.
54 - subscript: given two array accesses a subscript is the tuple
55 composed of the access functions for a given dimension. Example:
56 Given A[f1][f2][f3] and B[g1][g2][g3], there are three subscripts:
57 (f1, g1), (f2, g2), (f3, g3).
59 - Diophantine equation: an equation whose coefficients and
60 solutions are integer constants, for example the equation
62 has an integer solution x = 1 and y = -1.
66 - "Advanced Compilation for High Performance Computing" by Randy
67 Allen and Ken Kennedy.
68 http://citeseer.ist.psu.edu/goff91practical.html
70 - "Loop Transformations for Restructuring Compilers - The Foundations"
78 #include "coretypes.h"
83 /* These RTL headers are needed for basic-block.h. */
85 #include "basic-block.h"
86 #include "diagnostic.h"
87 #include "tree-flow.h"
88 #include "tree-dump.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;
108 int num_ziv_independent;
109 int num_ziv_dependent;
110 int num_ziv_unimplemented;
113 int num_siv_independent;
114 int num_siv_dependent;
115 int num_siv_unimplemented;
118 int num_miv_independent;
119 int num_miv_dependent;
120 int num_miv_unimplemented;
123 static bool subscript_dependence_tester_1 (struct data_dependence_relation *,
124 struct data_reference *,
125 struct data_reference *,
127 /* Returns true iff A divides B. */
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. */
140 int_divides_p (int a, int b)
142 return ((b % a) == 0);
147 /* Dump into FILE all the data references from DATAREFS. */
150 dump_data_references (FILE *file, VEC (data_reference_p, heap) *datarefs)
153 struct data_reference *dr;
155 for (i = 0; VEC_iterate (data_reference_p, datarefs, i, dr); i++)
156 dump_data_reference (file, dr);
159 /* Dump to STDERR all the dependence relations from DDRS. */
162 debug_data_dependence_relations (VEC (ddr_p, heap) *ddrs)
164 dump_data_dependence_relations (stderr, ddrs);
167 /* Dump into FILE all the dependence relations from DDRS. */
170 dump_data_dependence_relations (FILE *file,
171 VEC (ddr_p, heap) *ddrs)
174 struct data_dependence_relation *ddr;
176 for (i = 0; VEC_iterate (ddr_p, ddrs, i, ddr); i++)
177 dump_data_dependence_relation (file, ddr);
180 /* Dump function for a DATA_REFERENCE structure. */
183 dump_data_reference (FILE *outf,
184 struct data_reference *dr)
188 fprintf (outf, "(Data Ref: \n stmt: ");
189 print_generic_stmt (outf, DR_STMT (dr), 0);
190 fprintf (outf, " ref: ");
191 print_generic_stmt (outf, DR_REF (dr), 0);
192 fprintf (outf, " base_object: ");
193 print_generic_stmt (outf, DR_BASE_OBJECT (dr), 0);
195 for (i = 0; i < DR_NUM_DIMENSIONS (dr); i++)
197 fprintf (outf, " Access function %d: ", i);
198 print_generic_stmt (outf, DR_ACCESS_FN (dr, i), 0);
200 fprintf (outf, ")\n");
203 /* Dumps the affine function described by FN to the file OUTF. */
206 dump_affine_function (FILE *outf, affine_fn fn)
211 print_generic_expr (outf, VEC_index (tree, fn, 0), TDF_SLIM);
212 for (i = 1; VEC_iterate (tree, fn, i, coef); i++)
214 fprintf (outf, " + ");
215 print_generic_expr (outf, coef, TDF_SLIM);
216 fprintf (outf, " * x_%u", i);
220 /* Dumps the conflict function CF to the file OUTF. */
223 dump_conflict_function (FILE *outf, conflict_function *cf)
227 if (cf->n == NO_DEPENDENCE)
228 fprintf (outf, "no dependence\n");
229 else if (cf->n == NOT_KNOWN)
230 fprintf (outf, "not known\n");
233 for (i = 0; i < cf->n; i++)
236 dump_affine_function (outf, cf->fns[i]);
237 fprintf (outf, "]\n");
242 /* Dump function for a SUBSCRIPT structure. */
245 dump_subscript (FILE *outf, struct subscript *subscript)
247 conflict_function *cf = SUB_CONFLICTS_IN_A (subscript);
249 fprintf (outf, "\n (subscript \n");
250 fprintf (outf, " iterations_that_access_an_element_twice_in_A: ");
251 dump_conflict_function (outf, cf);
252 if (CF_NONTRIVIAL_P (cf))
254 tree last_iteration = SUB_LAST_CONFLICT (subscript);
255 fprintf (outf, " last_conflict: ");
256 print_generic_stmt (outf, last_iteration, 0);
259 cf = SUB_CONFLICTS_IN_B (subscript);
260 fprintf (outf, " iterations_that_access_an_element_twice_in_B: ");
261 dump_conflict_function (outf, cf);
262 if (CF_NONTRIVIAL_P (cf))
264 tree last_iteration = SUB_LAST_CONFLICT (subscript);
265 fprintf (outf, " last_conflict: ");
266 print_generic_stmt (outf, last_iteration, 0);
269 fprintf (outf, " (Subscript distance: ");
270 print_generic_stmt (outf, SUB_DISTANCE (subscript), 0);
271 fprintf (outf, " )\n");
272 fprintf (outf, " )\n");
275 /* Print the classic direction vector DIRV to OUTF. */
278 print_direction_vector (FILE *outf,
284 for (eq = 0; eq < length; eq++)
286 enum data_dependence_direction dir = dirv[eq];
291 fprintf (outf, " +");
294 fprintf (outf, " -");
297 fprintf (outf, " =");
299 case dir_positive_or_equal:
300 fprintf (outf, " +=");
302 case dir_positive_or_negative:
303 fprintf (outf, " +-");
305 case dir_negative_or_equal:
306 fprintf (outf, " -=");
309 fprintf (outf, " *");
312 fprintf (outf, "indep");
316 fprintf (outf, "\n");
319 /* Print a vector of direction vectors. */
322 print_dir_vectors (FILE *outf, VEC (lambda_vector, heap) *dir_vects,
328 for (j = 0; VEC_iterate (lambda_vector, dir_vects, j, v); j++)
329 print_direction_vector (outf, v, length);
332 /* Print a vector of distance vectors. */
335 print_dist_vectors (FILE *outf, VEC (lambda_vector, heap) *dist_vects,
341 for (j = 0; VEC_iterate (lambda_vector, dist_vects, j, v); j++)
342 print_lambda_vector (outf, v, length);
348 debug_data_dependence_relation (struct data_dependence_relation *ddr)
350 dump_data_dependence_relation (stderr, ddr);
353 /* Dump function for a DATA_DEPENDENCE_RELATION structure. */
356 dump_data_dependence_relation (FILE *outf,
357 struct data_dependence_relation *ddr)
359 struct data_reference *dra, *drb;
361 fprintf (outf, "(Data Dep: \n");
363 if (!ddr || DDR_ARE_DEPENDENT (ddr) == chrec_dont_know)
365 fprintf (outf, " (don't know)\n)\n");
371 dump_data_reference (outf, dra);
372 dump_data_reference (outf, drb);
374 if (DDR_ARE_DEPENDENT (ddr) == chrec_known)
375 fprintf (outf, " (no dependence)\n");
377 else if (DDR_ARE_DEPENDENT (ddr) == NULL_TREE)
382 for (i = 0; i < DDR_NUM_SUBSCRIPTS (ddr); i++)
384 fprintf (outf, " access_fn_A: ");
385 print_generic_stmt (outf, DR_ACCESS_FN (dra, i), 0);
386 fprintf (outf, " access_fn_B: ");
387 print_generic_stmt (outf, DR_ACCESS_FN (drb, i), 0);
388 dump_subscript (outf, DDR_SUBSCRIPT (ddr, i));
391 fprintf (outf, " inner loop index: %d\n", DDR_INNER_LOOP (ddr));
392 fprintf (outf, " loop nest: (");
393 for (i = 0; VEC_iterate (loop_p, DDR_LOOP_NEST (ddr), i, loopi); i++)
394 fprintf (outf, "%d ", loopi->num);
395 fprintf (outf, ")\n");
397 for (i = 0; i < DDR_NUM_DIST_VECTS (ddr); i++)
399 fprintf (outf, " distance_vector: ");
400 print_lambda_vector (outf, DDR_DIST_VECT (ddr, i),
404 for (i = 0; i < DDR_NUM_DIR_VECTS (ddr); i++)
406 fprintf (outf, " direction_vector: ");
407 print_direction_vector (outf, DDR_DIR_VECT (ddr, i),
412 fprintf (outf, ")\n");
415 /* Dump function for a DATA_DEPENDENCE_DIRECTION structure. */
418 dump_data_dependence_direction (FILE *file,
419 enum data_dependence_direction dir)
435 case dir_positive_or_negative:
436 fprintf (file, "+-");
439 case dir_positive_or_equal:
440 fprintf (file, "+=");
443 case dir_negative_or_equal:
444 fprintf (file, "-=");
456 /* Dumps the distance and direction vectors in FILE. DDRS contains
457 the dependence relations, and VECT_SIZE is the size of the
458 dependence vectors, or in other words the number of loops in the
462 dump_dist_dir_vectors (FILE *file, VEC (ddr_p, heap) *ddrs)
465 struct data_dependence_relation *ddr;
468 for (i = 0; VEC_iterate (ddr_p, ddrs, i, ddr); i++)
469 if (DDR_ARE_DEPENDENT (ddr) == NULL_TREE && DDR_AFFINE_P (ddr))
471 for (j = 0; VEC_iterate (lambda_vector, DDR_DIST_VECTS (ddr), j, v); j++)
473 fprintf (file, "DISTANCE_V (");
474 print_lambda_vector (file, v, DDR_NB_LOOPS (ddr));
475 fprintf (file, ")\n");
478 for (j = 0; VEC_iterate (lambda_vector, DDR_DIR_VECTS (ddr), j, v); j++)
480 fprintf (file, "DIRECTION_V (");
481 print_direction_vector (file, v, DDR_NB_LOOPS (ddr));
482 fprintf (file, ")\n");
486 fprintf (file, "\n\n");
489 /* Dumps the data dependence relations DDRS in FILE. */
492 dump_ddrs (FILE *file, VEC (ddr_p, heap) *ddrs)
495 struct data_dependence_relation *ddr;
497 for (i = 0; VEC_iterate (ddr_p, ddrs, i, ddr); i++)
498 dump_data_dependence_relation (file, ddr);
500 fprintf (file, "\n\n");
503 /* Expresses EXP as VAR + OFF, where off is a constant. The type of OFF
504 will be ssizetype. */
507 split_constant_offset (tree exp, tree *var, tree *off)
509 tree type = TREE_TYPE (exp), otype;
516 otype = TREE_TYPE (exp);
517 code = TREE_CODE (exp);
522 *var = build_int_cst (type, 0);
523 *off = fold_convert (ssizetype, exp);
526 case POINTER_PLUS_EXPR:
531 split_constant_offset (TREE_OPERAND (exp, 0), &var0, &off0);
532 split_constant_offset (TREE_OPERAND (exp, 1), &var1, &off1);
533 *var = fold_convert (type, fold_build2 (TREE_CODE (exp), otype,
535 *off = size_binop (code, off0, off1);
539 off1 = TREE_OPERAND (exp, 1);
540 if (TREE_CODE (off1) != INTEGER_CST)
543 split_constant_offset (TREE_OPERAND (exp, 0), &var0, &off0);
544 *var = fold_convert (type, fold_build2 (MULT_EXPR, otype,
546 *off = size_binop (MULT_EXPR, off0, fold_convert (ssizetype, off1));
551 tree op, base, poffset;
552 HOST_WIDE_INT pbitsize, pbitpos;
553 enum machine_mode pmode;
554 int punsignedp, pvolatilep;
556 op = TREE_OPERAND (exp, 0);
557 if (!handled_component_p (op))
560 base = get_inner_reference (op, &pbitsize, &pbitpos, &poffset,
561 &pmode, &punsignedp, &pvolatilep, false);
563 if (pbitpos % BITS_PER_UNIT != 0)
565 base = build_fold_addr_expr (base);
566 off0 = ssize_int (pbitpos / BITS_PER_UNIT);
570 split_constant_offset (poffset, &poffset, &off1);
571 off0 = size_binop (PLUS_EXPR, off0, off1);
572 if (POINTER_TYPE_P (TREE_TYPE (base)))
573 base = fold_build2 (POINTER_PLUS_EXPR, TREE_TYPE (base),
574 base, fold_convert (sizetype, poffset));
576 base = fold_build2 (PLUS_EXPR, TREE_TYPE (base), base,
577 fold_convert (TREE_TYPE (base), poffset));
580 var0 = fold_convert (type, base);
582 /* If variable length types are involved, punt, otherwise casts
583 might be converted into ARRAY_REFs in gimplify_conversion.
584 To compute that ARRAY_REF's element size TYPE_SIZE_UNIT, which
585 possibly no longer appears in current GIMPLE, might resurface.
586 This perhaps could run
587 if (TREE_CODE (var0) == NOP_EXPR
588 || TREE_CODE (var0) == CONVERT_EXPR)
590 gimplify_conversion (&var0);
591 // Attempt to fill in any within var0 found ARRAY_REF's
592 // element size from corresponding op embedded ARRAY_REF,
593 // if unsuccessful, just punt.
595 while (POINTER_TYPE_P (type))
596 type = TREE_TYPE (type);
597 if (int_size_in_bytes (type) < 0)
607 tree def_stmt = SSA_NAME_DEF_STMT (exp);
608 if (TREE_CODE (def_stmt) == GIMPLE_MODIFY_STMT)
610 tree def_stmt_rhs = GIMPLE_STMT_OPERAND (def_stmt, 1);
612 if (!TREE_SIDE_EFFECTS (def_stmt_rhs)
613 && EXPR_P (def_stmt_rhs)
614 && !REFERENCE_CLASS_P (def_stmt_rhs)
615 && !get_call_expr_in (def_stmt_rhs))
617 split_constant_offset (def_stmt_rhs, &var0, &off0);
618 var0 = fold_convert (type, var0);
631 *off = ssize_int (0);
634 /* Returns the address ADDR of an object in a canonical shape (without nop
635 casts, and with type of pointer to the object). */
638 canonicalize_base_object_address (tree addr)
644 /* The base address may be obtained by casting from integer, in that case
646 if (!POINTER_TYPE_P (TREE_TYPE (addr)))
649 if (TREE_CODE (addr) != ADDR_EXPR)
652 return build_fold_addr_expr (TREE_OPERAND (addr, 0));
655 /* Analyzes the behavior of the memory reference DR in the innermost loop that
659 dr_analyze_innermost (struct data_reference *dr)
661 tree stmt = DR_STMT (dr);
662 struct loop *loop = loop_containing_stmt (stmt);
663 tree ref = DR_REF (dr);
664 HOST_WIDE_INT pbitsize, pbitpos;
666 enum machine_mode pmode;
667 int punsignedp, pvolatilep;
668 affine_iv base_iv, offset_iv;
669 tree init, dinit, step;
671 if (dump_file && (dump_flags & TDF_DETAILS))
672 fprintf (dump_file, "analyze_innermost: ");
674 base = get_inner_reference (ref, &pbitsize, &pbitpos, &poffset,
675 &pmode, &punsignedp, &pvolatilep, false);
676 gcc_assert (base != NULL_TREE);
678 if (pbitpos % BITS_PER_UNIT != 0)
680 if (dump_file && (dump_flags & TDF_DETAILS))
681 fprintf (dump_file, "failed: bit offset alignment.\n");
685 base = build_fold_addr_expr (base);
686 if (!simple_iv (loop, stmt, base, &base_iv, false))
688 if (dump_file && (dump_flags & TDF_DETAILS))
689 fprintf (dump_file, "failed: evolution of base is not affine.\n");
694 offset_iv.base = ssize_int (0);
695 offset_iv.step = ssize_int (0);
697 else if (!simple_iv (loop, stmt, poffset, &offset_iv, false))
699 if (dump_file && (dump_flags & TDF_DETAILS))
700 fprintf (dump_file, "failed: evolution of offset is not affine.\n");
704 init = ssize_int (pbitpos / BITS_PER_UNIT);
705 split_constant_offset (base_iv.base, &base_iv.base, &dinit);
706 init = size_binop (PLUS_EXPR, init, dinit);
707 split_constant_offset (offset_iv.base, &offset_iv.base, &dinit);
708 init = size_binop (PLUS_EXPR, init, dinit);
710 step = size_binop (PLUS_EXPR,
711 fold_convert (ssizetype, base_iv.step),
712 fold_convert (ssizetype, offset_iv.step));
714 DR_BASE_ADDRESS (dr) = canonicalize_base_object_address (base_iv.base);
716 DR_OFFSET (dr) = fold_convert (ssizetype, offset_iv.base);
720 DR_ALIGNED_TO (dr) = size_int (highest_pow2_factor (offset_iv.base));
722 if (dump_file && (dump_flags & TDF_DETAILS))
723 fprintf (dump_file, "success.\n");
726 /* Determines the base object and the list of indices of memory reference
727 DR, analyzed in loop nest NEST. */
730 dr_analyze_indices (struct data_reference *dr, struct loop *nest)
732 tree stmt = DR_STMT (dr);
733 struct loop *loop = loop_containing_stmt (stmt);
734 VEC (tree, heap) *access_fns = NULL;
735 tree ref = unshare_expr (DR_REF (dr)), aref = ref, op;
736 tree base, off, access_fn;
738 while (handled_component_p (aref))
740 if (TREE_CODE (aref) == ARRAY_REF)
742 op = TREE_OPERAND (aref, 1);
743 access_fn = analyze_scalar_evolution (loop, op);
744 access_fn = resolve_mixers (nest, access_fn);
745 VEC_safe_push (tree, heap, access_fns, access_fn);
747 TREE_OPERAND (aref, 1) = build_int_cst (TREE_TYPE (op), 0);
750 aref = TREE_OPERAND (aref, 0);
753 if (INDIRECT_REF_P (aref))
755 op = TREE_OPERAND (aref, 0);
756 access_fn = analyze_scalar_evolution (loop, op);
757 access_fn = resolve_mixers (nest, access_fn);
758 base = initial_condition (access_fn);
759 split_constant_offset (base, &base, &off);
760 access_fn = chrec_replace_initial_condition (access_fn,
761 fold_convert (TREE_TYPE (base), off));
763 TREE_OPERAND (aref, 0) = base;
764 VEC_safe_push (tree, heap, access_fns, access_fn);
767 DR_BASE_OBJECT (dr) = ref;
768 DR_ACCESS_FNS (dr) = access_fns;
771 /* Extracts the alias analysis information from the memory reference DR. */
774 dr_analyze_alias (struct data_reference *dr)
776 tree stmt = DR_STMT (dr);
777 tree ref = DR_REF (dr);
778 tree base = get_base_address (ref), addr, smt = NULL_TREE;
785 else if (INDIRECT_REF_P (base))
787 addr = TREE_OPERAND (base, 0);
788 if (TREE_CODE (addr) == SSA_NAME)
790 smt = symbol_mem_tag (SSA_NAME_VAR (addr));
791 DR_PTR_INFO (dr) = SSA_NAME_PTR_INFO (addr);
795 DR_SYMBOL_TAG (dr) = smt;
797 vops = BITMAP_ALLOC (NULL);
798 FOR_EACH_SSA_TREE_OPERAND (op, stmt, it, SSA_OP_VIRTUAL_USES)
800 bitmap_set_bit (vops, DECL_UID (SSA_NAME_VAR (op)));
806 /* Returns true if the address of DR is invariant. */
809 dr_address_invariant_p (struct data_reference *dr)
814 for (i = 0; VEC_iterate (tree, DR_ACCESS_FNS (dr), i, idx); i++)
815 if (tree_contains_chrecs (idx, NULL))
821 /* Frees data reference DR. */
824 free_data_ref (data_reference_p dr)
826 BITMAP_FREE (DR_VOPS (dr));
827 VEC_free (tree, heap, DR_ACCESS_FNS (dr));
831 /* Analyzes memory reference MEMREF accessed in STMT. The reference
832 is read if IS_READ is true, write otherwise. Returns the
833 data_reference description of MEMREF. NEST is the outermost loop of the
834 loop nest in that the reference should be analyzed. */
836 struct data_reference *
837 create_data_ref (struct loop *nest, tree memref, tree stmt, bool is_read)
839 struct data_reference *dr;
841 if (dump_file && (dump_flags & TDF_DETAILS))
843 fprintf (dump_file, "Creating dr for ");
844 print_generic_expr (dump_file, memref, TDF_SLIM);
845 fprintf (dump_file, "\n");
848 dr = XCNEW (struct data_reference);
850 DR_REF (dr) = memref;
851 DR_IS_READ (dr) = is_read;
853 dr_analyze_innermost (dr);
854 dr_analyze_indices (dr, nest);
855 dr_analyze_alias (dr);
857 if (dump_file && (dump_flags & TDF_DETAILS))
859 fprintf (dump_file, "\tbase_address: ");
860 print_generic_expr (dump_file, DR_BASE_ADDRESS (dr), TDF_SLIM);
861 fprintf (dump_file, "\n\toffset from base address: ");
862 print_generic_expr (dump_file, DR_OFFSET (dr), TDF_SLIM);
863 fprintf (dump_file, "\n\tconstant offset from base address: ");
864 print_generic_expr (dump_file, DR_INIT (dr), TDF_SLIM);
865 fprintf (dump_file, "\n\tstep: ");
866 print_generic_expr (dump_file, DR_STEP (dr), TDF_SLIM);
867 fprintf (dump_file, "\n\taligned to: ");
868 print_generic_expr (dump_file, DR_ALIGNED_TO (dr), TDF_SLIM);
869 fprintf (dump_file, "\n\tbase_object: ");
870 print_generic_expr (dump_file, DR_BASE_OBJECT (dr), TDF_SLIM);
871 fprintf (dump_file, "\n\tsymbol tag: ");
872 print_generic_expr (dump_file, DR_SYMBOL_TAG (dr), TDF_SLIM);
873 fprintf (dump_file, "\n");
879 /* Returns true if FNA == FNB. */
882 affine_function_equal_p (affine_fn fna, affine_fn fnb)
884 unsigned i, n = VEC_length (tree, fna);
886 if (n != VEC_length (tree, fnb))
889 for (i = 0; i < n; i++)
890 if (!operand_equal_p (VEC_index (tree, fna, i),
891 VEC_index (tree, fnb, i), 0))
897 /* If all the functions in CF are the same, returns one of them,
898 otherwise returns NULL. */
901 common_affine_function (conflict_function *cf)
906 if (!CF_NONTRIVIAL_P (cf))
911 for (i = 1; i < cf->n; i++)
912 if (!affine_function_equal_p (comm, cf->fns[i]))
918 /* Returns the base of the affine function FN. */
921 affine_function_base (affine_fn fn)
923 return VEC_index (tree, fn, 0);
926 /* Returns true if FN is a constant. */
929 affine_function_constant_p (affine_fn fn)
934 for (i = 1; VEC_iterate (tree, fn, i, coef); i++)
935 if (!integer_zerop (coef))
941 /* Returns true if FN is the zero constant function. */
944 affine_function_zero_p (affine_fn fn)
946 return (integer_zerop (affine_function_base (fn))
947 && affine_function_constant_p (fn));
950 /* Returns a signed integer type with the largest precision from TA
954 signed_type_for_types (tree ta, tree tb)
956 if (TYPE_PRECISION (ta) > TYPE_PRECISION (tb))
957 return signed_type_for (ta);
959 return signed_type_for (tb);
962 /* Applies operation OP on affine functions FNA and FNB, and returns the
966 affine_fn_op (enum tree_code op, affine_fn fna, affine_fn fnb)
972 if (VEC_length (tree, fnb) > VEC_length (tree, fna))
974 n = VEC_length (tree, fna);
975 m = VEC_length (tree, fnb);
979 n = VEC_length (tree, fnb);
980 m = VEC_length (tree, fna);
983 ret = VEC_alloc (tree, heap, m);
984 for (i = 0; i < n; i++)
986 tree type = signed_type_for_types (TREE_TYPE (VEC_index (tree, fna, i)),
987 TREE_TYPE (VEC_index (tree, fnb, i)));
989 VEC_quick_push (tree, ret,
990 fold_build2 (op, type,
991 VEC_index (tree, fna, i),
992 VEC_index (tree, fnb, i)));
995 for (; VEC_iterate (tree, fna, i, coef); i++)
996 VEC_quick_push (tree, ret,
997 fold_build2 (op, signed_type_for (TREE_TYPE (coef)),
998 coef, integer_zero_node));
999 for (; VEC_iterate (tree, fnb, i, coef); i++)
1000 VEC_quick_push (tree, ret,
1001 fold_build2 (op, signed_type_for (TREE_TYPE (coef)),
1002 integer_zero_node, coef));
1007 /* Returns the sum of affine functions FNA and FNB. */
1010 affine_fn_plus (affine_fn fna, affine_fn fnb)
1012 return affine_fn_op (PLUS_EXPR, fna, fnb);
1015 /* Returns the difference of affine functions FNA and FNB. */
1018 affine_fn_minus (affine_fn fna, affine_fn fnb)
1020 return affine_fn_op (MINUS_EXPR, fna, fnb);
1023 /* Frees affine function FN. */
1026 affine_fn_free (affine_fn fn)
1028 VEC_free (tree, heap, fn);
1031 /* Determine for each subscript in the data dependence relation DDR
1035 compute_subscript_distance (struct data_dependence_relation *ddr)
1037 conflict_function *cf_a, *cf_b;
1038 affine_fn fn_a, fn_b, diff;
1040 if (DDR_ARE_DEPENDENT (ddr) == NULL_TREE)
1044 for (i = 0; i < DDR_NUM_SUBSCRIPTS (ddr); i++)
1046 struct subscript *subscript;
1048 subscript = DDR_SUBSCRIPT (ddr, i);
1049 cf_a = SUB_CONFLICTS_IN_A (subscript);
1050 cf_b = SUB_CONFLICTS_IN_B (subscript);
1052 fn_a = common_affine_function (cf_a);
1053 fn_b = common_affine_function (cf_b);
1056 SUB_DISTANCE (subscript) = chrec_dont_know;
1059 diff = affine_fn_minus (fn_a, fn_b);
1061 if (affine_function_constant_p (diff))
1062 SUB_DISTANCE (subscript) = affine_function_base (diff);
1064 SUB_DISTANCE (subscript) = chrec_dont_know;
1066 affine_fn_free (diff);
1071 /* Returns the conflict function for "unknown". */
1073 static conflict_function *
1074 conflict_fn_not_known (void)
1076 conflict_function *fn = XCNEW (conflict_function);
1082 /* Returns the conflict function for "independent". */
1084 static conflict_function *
1085 conflict_fn_no_dependence (void)
1087 conflict_function *fn = XCNEW (conflict_function);
1088 fn->n = NO_DEPENDENCE;
1093 /* Returns true if the address of OBJ is invariant in LOOP. */
1096 object_address_invariant_in_loop_p (const struct loop *loop, const_tree obj)
1098 while (handled_component_p (obj))
1100 if (TREE_CODE (obj) == ARRAY_REF)
1102 /* Index of the ARRAY_REF was zeroed in analyze_indices, thus we only
1103 need to check the stride and the lower bound of the reference. */
1104 if (chrec_contains_symbols_defined_in_loop (TREE_OPERAND (obj, 2),
1106 || chrec_contains_symbols_defined_in_loop (TREE_OPERAND (obj, 3),
1110 else if (TREE_CODE (obj) == COMPONENT_REF)
1112 if (chrec_contains_symbols_defined_in_loop (TREE_OPERAND (obj, 2),
1116 obj = TREE_OPERAND (obj, 0);
1119 if (!INDIRECT_REF_P (obj))
1122 return !chrec_contains_symbols_defined_in_loop (TREE_OPERAND (obj, 0),
1126 /* Returns true if A and B are accesses to different objects, or to different
1127 fields of the same object. */
1130 disjoint_objects_p (tree a, tree b)
1132 tree base_a, base_b;
1133 VEC (tree, heap) *comp_a = NULL, *comp_b = NULL;
1136 base_a = get_base_address (a);
1137 base_b = get_base_address (b);
1141 && base_a != base_b)
1144 if (!operand_equal_p (base_a, base_b, 0))
1147 /* Compare the component references of A and B. We must start from the inner
1148 ones, so record them to the vector first. */
1149 while (handled_component_p (a))
1151 VEC_safe_push (tree, heap, comp_a, a);
1152 a = TREE_OPERAND (a, 0);
1154 while (handled_component_p (b))
1156 VEC_safe_push (tree, heap, comp_b, b);
1157 b = TREE_OPERAND (b, 0);
1163 if (VEC_length (tree, comp_a) == 0
1164 || VEC_length (tree, comp_b) == 0)
1167 a = VEC_pop (tree, comp_a);
1168 b = VEC_pop (tree, comp_b);
1170 /* Real and imaginary part of a variable do not alias. */
1171 if ((TREE_CODE (a) == REALPART_EXPR
1172 && TREE_CODE (b) == IMAGPART_EXPR)
1173 || (TREE_CODE (a) == IMAGPART_EXPR
1174 && TREE_CODE (b) == REALPART_EXPR))
1180 if (TREE_CODE (a) != TREE_CODE (b))
1183 /* Nothing to do for ARRAY_REFs, as the indices of array_refs in
1184 DR_BASE_OBJECT are always zero. */
1185 if (TREE_CODE (a) == ARRAY_REF)
1187 else if (TREE_CODE (a) == COMPONENT_REF)
1189 if (operand_equal_p (TREE_OPERAND (a, 1), TREE_OPERAND (b, 1), 0))
1192 /* Different fields of unions may overlap. */
1193 base_a = TREE_OPERAND (a, 0);
1194 if (TREE_CODE (TREE_TYPE (base_a)) == UNION_TYPE)
1197 /* Different fields of structures cannot. */
1205 VEC_free (tree, heap, comp_a);
1206 VEC_free (tree, heap, comp_b);
1211 /* Returns false if we can prove that data references A and B do not alias,
1215 dr_may_alias_p (const struct data_reference *a, const struct data_reference *b)
1217 const_tree addr_a = DR_BASE_ADDRESS (a);
1218 const_tree addr_b = DR_BASE_ADDRESS (b);
1219 const_tree type_a, type_b;
1220 const_tree decl_a = NULL_TREE, decl_b = NULL_TREE;
1222 /* If the sets of virtual operands are disjoint, the memory references do not
1224 if (!bitmap_intersect_p (DR_VOPS (a), DR_VOPS (b)))
1227 /* If the accessed objects are disjoint, the memory references do not
1229 if (disjoint_objects_p (DR_BASE_OBJECT (a), DR_BASE_OBJECT (b)))
1232 if (!addr_a || !addr_b)
1235 /* If the references are based on different static objects, they cannot alias
1236 (PTA should be able to disambiguate such accesses, but often it fails to,
1237 since currently we cannot distinguish between pointer and offset in pointer
1239 if (TREE_CODE (addr_a) == ADDR_EXPR
1240 && TREE_CODE (addr_b) == ADDR_EXPR)
1241 return TREE_OPERAND (addr_a, 0) == TREE_OPERAND (addr_b, 0);
1243 /* An instruction writing through a restricted pointer is "independent" of any
1244 instruction reading or writing through a different restricted pointer,
1245 in the same block/scope. */
1247 type_a = TREE_TYPE (addr_a);
1248 type_b = TREE_TYPE (addr_b);
1249 gcc_assert (POINTER_TYPE_P (type_a) && POINTER_TYPE_P (type_b));
1251 if (TREE_CODE (addr_a) == SSA_NAME)
1252 decl_a = SSA_NAME_VAR (addr_a);
1253 if (TREE_CODE (addr_b) == SSA_NAME)
1254 decl_b = SSA_NAME_VAR (addr_b);
1256 if (TYPE_RESTRICT (type_a) && TYPE_RESTRICT (type_b)
1257 && (!DR_IS_READ (a) || !DR_IS_READ (b))
1258 && decl_a && DECL_P (decl_a)
1259 && decl_b && DECL_P (decl_b)
1261 && TREE_CODE (DECL_CONTEXT (decl_a)) == FUNCTION_DECL
1262 && DECL_CONTEXT (decl_a) == DECL_CONTEXT (decl_b))
1268 /* Initialize a data dependence relation between data accesses A and
1269 B. NB_LOOPS is the number of loops surrounding the references: the
1270 size of the classic distance/direction vectors. */
1272 static struct data_dependence_relation *
1273 initialize_data_dependence_relation (struct data_reference *a,
1274 struct data_reference *b,
1275 VEC (loop_p, heap) *loop_nest)
1277 struct data_dependence_relation *res;
1280 res = XNEW (struct data_dependence_relation);
1283 DDR_LOOP_NEST (res) = NULL;
1284 DDR_REVERSED_P (res) = false;
1285 DDR_SUBSCRIPTS (res) = NULL;
1286 DDR_DIR_VECTS (res) = NULL;
1287 DDR_DIST_VECTS (res) = NULL;
1289 if (a == NULL || b == NULL)
1291 DDR_ARE_DEPENDENT (res) = chrec_dont_know;
1295 /* If the data references do not alias, then they are independent. */
1296 if (!dr_may_alias_p (a, b))
1298 DDR_ARE_DEPENDENT (res) = chrec_known;
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;
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 (!object_address_invariant_in_loop_p (VEC_index (loop_p, loop_nest, 0),
1314 DR_BASE_OBJECT (a)))
1316 DDR_ARE_DEPENDENT (res) = chrec_dont_know;
1320 gcc_assert (DR_NUM_DIMENSIONS (a) == DR_NUM_DIMENSIONS (b));
1322 DDR_AFFINE_P (res) = true;
1323 DDR_ARE_DEPENDENT (res) = NULL_TREE;
1324 DDR_SUBSCRIPTS (res) = VEC_alloc (subscript_p, heap, DR_NUM_DIMENSIONS (a));
1325 DDR_LOOP_NEST (res) = loop_nest;
1326 DDR_INNER_LOOP (res) = 0;
1328 for (i = 0; i < DR_NUM_DIMENSIONS (a); i++)
1330 struct subscript *subscript;
1332 subscript = XNEW (struct subscript);
1333 SUB_CONFLICTS_IN_A (subscript) = conflict_fn_not_known ();
1334 SUB_CONFLICTS_IN_B (subscript) = conflict_fn_not_known ();
1335 SUB_LAST_CONFLICT (subscript) = chrec_dont_know;
1336 SUB_DISTANCE (subscript) = chrec_dont_know;
1337 VEC_safe_push (subscript_p, heap, DDR_SUBSCRIPTS (res), subscript);
1343 /* Frees memory used by the conflict function F. */
1346 free_conflict_function (conflict_function *f)
1350 if (CF_NONTRIVIAL_P (f))
1352 for (i = 0; i < f->n; i++)
1353 affine_fn_free (f->fns[i]);
1358 /* Frees memory used by SUBSCRIPTS. */
1361 free_subscripts (VEC (subscript_p, heap) *subscripts)
1366 for (i = 0; VEC_iterate (subscript_p, subscripts, i, s); i++)
1368 free_conflict_function (s->conflicting_iterations_in_a);
1369 free_conflict_function (s->conflicting_iterations_in_b);
1371 VEC_free (subscript_p, heap, subscripts);
1374 /* Set DDR_ARE_DEPENDENT to CHREC and finalize the subscript overlap
1378 finalize_ddr_dependent (struct data_dependence_relation *ddr,
1381 if (dump_file && (dump_flags & TDF_DETAILS))
1383 fprintf (dump_file, "(dependence classified: ");
1384 print_generic_expr (dump_file, chrec, 0);
1385 fprintf (dump_file, ")\n");
1388 DDR_ARE_DEPENDENT (ddr) = chrec;
1389 free_subscripts (DDR_SUBSCRIPTS (ddr));
1390 DDR_SUBSCRIPTS (ddr) = NULL;
1393 /* The dependence relation DDR cannot be represented by a distance
1397 non_affine_dependence_relation (struct data_dependence_relation *ddr)
1399 if (dump_file && (dump_flags & TDF_DETAILS))
1400 fprintf (dump_file, "(Dependence relation cannot be represented by distance vector.) \n");
1402 DDR_AFFINE_P (ddr) = false;
1407 /* This section contains the classic Banerjee tests. */
1409 /* Returns true iff CHREC_A and CHREC_B are not dependent on any index
1410 variables, i.e., if the ZIV (Zero Index Variable) test is true. */
1413 ziv_subscript_p (const_tree chrec_a, const_tree chrec_b)
1415 return (evolution_function_is_constant_p (chrec_a)
1416 && evolution_function_is_constant_p (chrec_b));
1419 /* Returns true iff CHREC_A and CHREC_B are dependent on an index
1420 variable, i.e., if the SIV (Single Index Variable) test is true. */
1423 siv_subscript_p (const_tree chrec_a, const_tree chrec_b)
1425 if ((evolution_function_is_constant_p (chrec_a)
1426 && evolution_function_is_univariate_p (chrec_b))
1427 || (evolution_function_is_constant_p (chrec_b)
1428 && evolution_function_is_univariate_p (chrec_a)))
1431 if (evolution_function_is_univariate_p (chrec_a)
1432 && evolution_function_is_univariate_p (chrec_b))
1434 switch (TREE_CODE (chrec_a))
1436 case POLYNOMIAL_CHREC:
1437 switch (TREE_CODE (chrec_b))
1439 case POLYNOMIAL_CHREC:
1440 if (CHREC_VARIABLE (chrec_a) != CHREC_VARIABLE (chrec_b))
1455 /* Creates a conflict function with N dimensions. The affine functions
1456 in each dimension follow. */
1458 static conflict_function *
1459 conflict_fn (unsigned n, ...)
1462 conflict_function *ret = XCNEW (conflict_function);
1465 gcc_assert (0 < n && n <= MAX_DIM);
1469 for (i = 0; i < n; i++)
1470 ret->fns[i] = va_arg (ap, affine_fn);
1476 /* Returns constant affine function with value CST. */
1479 affine_fn_cst (tree cst)
1481 affine_fn fn = VEC_alloc (tree, heap, 1);
1482 VEC_quick_push (tree, fn, cst);
1486 /* Returns affine function with single variable, CST + COEF * x_DIM. */
1489 affine_fn_univar (tree cst, unsigned dim, tree coef)
1491 affine_fn fn = VEC_alloc (tree, heap, dim + 1);
1494 gcc_assert (dim > 0);
1495 VEC_quick_push (tree, fn, cst);
1496 for (i = 1; i < dim; i++)
1497 VEC_quick_push (tree, fn, integer_zero_node);
1498 VEC_quick_push (tree, fn, coef);
1502 /* Analyze a ZIV (Zero Index Variable) subscript. *OVERLAPS_A and
1503 *OVERLAPS_B are initialized to the functions that describe the
1504 relation between the elements accessed twice by CHREC_A and
1505 CHREC_B. For k >= 0, the following property is verified:
1507 CHREC_A (*OVERLAPS_A (k)) = CHREC_B (*OVERLAPS_B (k)). */
1510 analyze_ziv_subscript (tree chrec_a,
1512 conflict_function **overlaps_a,
1513 conflict_function **overlaps_b,
1514 tree *last_conflicts)
1516 tree type, difference;
1517 dependence_stats.num_ziv++;
1519 if (dump_file && (dump_flags & TDF_DETAILS))
1520 fprintf (dump_file, "(analyze_ziv_subscript \n");
1522 type = signed_type_for_types (TREE_TYPE (chrec_a), TREE_TYPE (chrec_b));
1523 chrec_a = chrec_convert (type, chrec_a, NULL_TREE);
1524 chrec_b = chrec_convert (type, chrec_b, NULL_TREE);
1525 difference = chrec_fold_minus (type, chrec_a, chrec_b);
1527 switch (TREE_CODE (difference))
1530 if (integer_zerop (difference))
1532 /* The difference is equal to zero: the accessed index
1533 overlaps for each iteration in the loop. */
1534 *overlaps_a = conflict_fn (1, affine_fn_cst (integer_zero_node));
1535 *overlaps_b = conflict_fn (1, affine_fn_cst (integer_zero_node));
1536 *last_conflicts = chrec_dont_know;
1537 dependence_stats.num_ziv_dependent++;
1541 /* The accesses do not overlap. */
1542 *overlaps_a = conflict_fn_no_dependence ();
1543 *overlaps_b = conflict_fn_no_dependence ();
1544 *last_conflicts = integer_zero_node;
1545 dependence_stats.num_ziv_independent++;
1550 /* We're not sure whether the indexes overlap. For the moment,
1551 conservatively answer "don't know". */
1552 if (dump_file && (dump_flags & TDF_DETAILS))
1553 fprintf (dump_file, "ziv test failed: difference is non-integer.\n");
1555 *overlaps_a = conflict_fn_not_known ();
1556 *overlaps_b = conflict_fn_not_known ();
1557 *last_conflicts = chrec_dont_know;
1558 dependence_stats.num_ziv_unimplemented++;
1562 if (dump_file && (dump_flags & TDF_DETAILS))
1563 fprintf (dump_file, ")\n");
1566 /* Sets NIT to the estimated number of executions of the statements in
1567 LOOP. If CONSERVATIVE is true, we must be sure that NIT is at least as
1568 large as the number of iterations. If we have no reliable estimate,
1569 the function returns false, otherwise returns true. */
1572 estimated_loop_iterations (struct loop *loop, bool conservative,
1575 estimate_numbers_of_iterations_loop (loop);
1578 if (!loop->any_upper_bound)
1581 *nit = loop->nb_iterations_upper_bound;
1585 if (!loop->any_estimate)
1588 *nit = loop->nb_iterations_estimate;
1594 /* Similar to estimated_loop_iterations, but returns the estimate only
1595 if it fits to HOST_WIDE_INT. If this is not the case, or the estimate
1596 on the number of iterations of LOOP could not be derived, returns -1. */
1599 estimated_loop_iterations_int (struct loop *loop, bool conservative)
1602 HOST_WIDE_INT hwi_nit;
1604 if (!estimated_loop_iterations (loop, conservative, &nit))
1607 if (!double_int_fits_in_shwi_p (nit))
1609 hwi_nit = double_int_to_shwi (nit);
1611 return hwi_nit < 0 ? -1 : hwi_nit;
1614 /* Similar to estimated_loop_iterations, but returns the estimate as a tree,
1615 and only if it fits to the int type. If this is not the case, or the
1616 estimate on the number of iterations of LOOP could not be derived, returns
1620 estimated_loop_iterations_tree (struct loop *loop, bool conservative)
1625 if (!estimated_loop_iterations (loop, conservative, &nit))
1626 return chrec_dont_know;
1628 type = lang_hooks.types.type_for_size (INT_TYPE_SIZE, true);
1629 if (!double_int_fits_to_tree_p (type, nit))
1630 return chrec_dont_know;
1632 return double_int_to_tree (type, nit);
1635 /* Analyze a SIV (Single Index Variable) subscript where CHREC_A is a
1636 constant, and CHREC_B is an affine function. *OVERLAPS_A and
1637 *OVERLAPS_B are initialized to the functions that describe the
1638 relation between the elements accessed twice by CHREC_A and
1639 CHREC_B. For k >= 0, the following property is verified:
1641 CHREC_A (*OVERLAPS_A (k)) = CHREC_B (*OVERLAPS_B (k)). */
1644 analyze_siv_subscript_cst_affine (tree chrec_a,
1646 conflict_function **overlaps_a,
1647 conflict_function **overlaps_b,
1648 tree *last_conflicts)
1650 bool value0, value1, value2;
1651 tree type, difference, tmp;
1653 type = signed_type_for_types (TREE_TYPE (chrec_a), TREE_TYPE (chrec_b));
1654 chrec_a = chrec_convert (type, chrec_a, NULL_TREE);
1655 chrec_b = chrec_convert (type, chrec_b, NULL_TREE);
1656 difference = chrec_fold_minus (type, initial_condition (chrec_b), chrec_a);
1658 if (!chrec_is_positive (initial_condition (difference), &value0))
1660 if (dump_file && (dump_flags & TDF_DETAILS))
1661 fprintf (dump_file, "siv test failed: chrec is not positive.\n");
1663 dependence_stats.num_siv_unimplemented++;
1664 *overlaps_a = conflict_fn_not_known ();
1665 *overlaps_b = conflict_fn_not_known ();
1666 *last_conflicts = chrec_dont_know;
1671 if (value0 == false)
1673 if (!chrec_is_positive (CHREC_RIGHT (chrec_b), &value1))
1675 if (dump_file && (dump_flags & TDF_DETAILS))
1676 fprintf (dump_file, "siv test failed: chrec not positive.\n");
1678 *overlaps_a = conflict_fn_not_known ();
1679 *overlaps_b = conflict_fn_not_known ();
1680 *last_conflicts = chrec_dont_know;
1681 dependence_stats.num_siv_unimplemented++;
1690 chrec_b = {10, +, 1}
1693 if (tree_fold_divides_p (CHREC_RIGHT (chrec_b), difference))
1695 HOST_WIDE_INT numiter;
1696 struct loop *loop = get_chrec_loop (chrec_b);
1698 *overlaps_a = conflict_fn (1, affine_fn_cst (integer_zero_node));
1699 tmp = fold_build2 (EXACT_DIV_EXPR, type,
1700 fold_build1 (ABS_EXPR, type, difference),
1701 CHREC_RIGHT (chrec_b));
1702 *overlaps_b = conflict_fn (1, affine_fn_cst (tmp));
1703 *last_conflicts = integer_one_node;
1706 /* Perform weak-zero siv test to see if overlap is
1707 outside the loop bounds. */
1708 numiter = estimated_loop_iterations_int (loop, false);
1711 && compare_tree_int (tmp, numiter) > 0)
1713 free_conflict_function (*overlaps_a);
1714 free_conflict_function (*overlaps_b);
1715 *overlaps_a = conflict_fn_no_dependence ();
1716 *overlaps_b = conflict_fn_no_dependence ();
1717 *last_conflicts = integer_zero_node;
1718 dependence_stats.num_siv_independent++;
1721 dependence_stats.num_siv_dependent++;
1725 /* When the step does not divide the difference, there are
1729 *overlaps_a = conflict_fn_no_dependence ();
1730 *overlaps_b = conflict_fn_no_dependence ();
1731 *last_conflicts = integer_zero_node;
1732 dependence_stats.num_siv_independent++;
1741 chrec_b = {10, +, -1}
1743 In this case, chrec_a will not overlap with chrec_b. */
1744 *overlaps_a = conflict_fn_no_dependence ();
1745 *overlaps_b = conflict_fn_no_dependence ();
1746 *last_conflicts = integer_zero_node;
1747 dependence_stats.num_siv_independent++;
1754 if (!chrec_is_positive (CHREC_RIGHT (chrec_b), &value2))
1756 if (dump_file && (dump_flags & TDF_DETAILS))
1757 fprintf (dump_file, "siv test failed: chrec not positive.\n");
1759 *overlaps_a = conflict_fn_not_known ();
1760 *overlaps_b = conflict_fn_not_known ();
1761 *last_conflicts = chrec_dont_know;
1762 dependence_stats.num_siv_unimplemented++;
1767 if (value2 == false)
1771 chrec_b = {10, +, -1}
1773 if (tree_fold_divides_p (CHREC_RIGHT (chrec_b), difference))
1775 HOST_WIDE_INT numiter;
1776 struct loop *loop = get_chrec_loop (chrec_b);
1778 *overlaps_a = conflict_fn (1, affine_fn_cst (integer_zero_node));
1779 tmp = fold_build2 (EXACT_DIV_EXPR, type, difference,
1780 CHREC_RIGHT (chrec_b));
1781 *overlaps_b = conflict_fn (1, affine_fn_cst (tmp));
1782 *last_conflicts = integer_one_node;
1784 /* Perform weak-zero siv test to see if overlap is
1785 outside the loop bounds. */
1786 numiter = estimated_loop_iterations_int (loop, false);
1789 && compare_tree_int (tmp, numiter) > 0)
1791 free_conflict_function (*overlaps_a);
1792 free_conflict_function (*overlaps_b);
1793 *overlaps_a = conflict_fn_no_dependence ();
1794 *overlaps_b = conflict_fn_no_dependence ();
1795 *last_conflicts = integer_zero_node;
1796 dependence_stats.num_siv_independent++;
1799 dependence_stats.num_siv_dependent++;
1803 /* When the step does not divide the difference, there
1807 *overlaps_a = conflict_fn_no_dependence ();
1808 *overlaps_b = conflict_fn_no_dependence ();
1809 *last_conflicts = integer_zero_node;
1810 dependence_stats.num_siv_independent++;
1820 In this case, chrec_a will not overlap with chrec_b. */
1821 *overlaps_a = conflict_fn_no_dependence ();
1822 *overlaps_b = conflict_fn_no_dependence ();
1823 *last_conflicts = integer_zero_node;
1824 dependence_stats.num_siv_independent++;
1832 /* Helper recursive function for initializing the matrix A. Returns
1833 the initial value of CHREC. */
1835 static HOST_WIDE_INT
1836 initialize_matrix_A (lambda_matrix A, tree chrec, unsigned index, int mult)
1840 if (TREE_CODE (chrec) != POLYNOMIAL_CHREC)
1841 return int_cst_value (chrec);
1843 A[index][0] = mult * int_cst_value (CHREC_RIGHT (chrec));
1844 return initialize_matrix_A (A, CHREC_LEFT (chrec), index + 1, mult);
1847 #define FLOOR_DIV(x,y) ((x) / (y))
1849 /* Solves the special case of the Diophantine equation:
1850 | {0, +, STEP_A}_x (OVERLAPS_A) = {0, +, STEP_B}_y (OVERLAPS_B)
1852 Computes the descriptions OVERLAPS_A and OVERLAPS_B. NITER is the
1853 number of iterations that loops X and Y run. The overlaps will be
1854 constructed as evolutions in dimension DIM. */
1857 compute_overlap_steps_for_affine_univar (int niter, int step_a, int step_b,
1858 affine_fn *overlaps_a,
1859 affine_fn *overlaps_b,
1860 tree *last_conflicts, int dim)
1862 if (((step_a > 0 && step_b > 0)
1863 || (step_a < 0 && step_b < 0)))
1865 int step_overlaps_a, step_overlaps_b;
1866 int gcd_steps_a_b, last_conflict, tau2;
1868 gcd_steps_a_b = gcd (step_a, step_b);
1869 step_overlaps_a = step_b / gcd_steps_a_b;
1870 step_overlaps_b = step_a / gcd_steps_a_b;
1874 tau2 = FLOOR_DIV (niter, step_overlaps_a);
1875 tau2 = MIN (tau2, FLOOR_DIV (niter, step_overlaps_b));
1876 last_conflict = tau2;
1877 *last_conflicts = build_int_cst (NULL_TREE, last_conflict);
1880 *last_conflicts = chrec_dont_know;
1882 *overlaps_a = affine_fn_univar (integer_zero_node, dim,
1883 build_int_cst (NULL_TREE,
1885 *overlaps_b = affine_fn_univar (integer_zero_node, dim,
1886 build_int_cst (NULL_TREE,
1892 *overlaps_a = affine_fn_cst (integer_zero_node);
1893 *overlaps_b = affine_fn_cst (integer_zero_node);
1894 *last_conflicts = integer_zero_node;
1898 /* Solves the special case of a Diophantine equation where CHREC_A is
1899 an affine bivariate function, and CHREC_B is an affine univariate
1900 function. For example,
1902 | {{0, +, 1}_x, +, 1335}_y = {0, +, 1336}_z
1904 has the following overlapping functions:
1906 | x (t, u, v) = {{0, +, 1336}_t, +, 1}_v
1907 | y (t, u, v) = {{0, +, 1336}_u, +, 1}_v
1908 | z (t, u, v) = {{{0, +, 1}_t, +, 1335}_u, +, 1}_v
1910 FORNOW: This is a specialized implementation for a case occurring in
1911 a common benchmark. Implement the general algorithm. */
1914 compute_overlap_steps_for_affine_1_2 (tree chrec_a, tree chrec_b,
1915 conflict_function **overlaps_a,
1916 conflict_function **overlaps_b,
1917 tree *last_conflicts)
1919 bool xz_p, yz_p, xyz_p;
1920 int step_x, step_y, step_z;
1921 HOST_WIDE_INT niter_x, niter_y, niter_z, niter;
1922 affine_fn overlaps_a_xz, overlaps_b_xz;
1923 affine_fn overlaps_a_yz, overlaps_b_yz;
1924 affine_fn overlaps_a_xyz, overlaps_b_xyz;
1925 affine_fn ova1, ova2, ovb;
1926 tree last_conflicts_xz, last_conflicts_yz, last_conflicts_xyz;
1928 step_x = int_cst_value (CHREC_RIGHT (CHREC_LEFT (chrec_a)));
1929 step_y = int_cst_value (CHREC_RIGHT (chrec_a));
1930 step_z = int_cst_value (CHREC_RIGHT (chrec_b));
1933 estimated_loop_iterations_int (get_chrec_loop (CHREC_LEFT (chrec_a)),
1935 niter_y = estimated_loop_iterations_int (get_chrec_loop (chrec_a), false);
1936 niter_z = estimated_loop_iterations_int (get_chrec_loop (chrec_b), false);
1938 if (niter_x < 0 || niter_y < 0 || niter_z < 0)
1940 if (dump_file && (dump_flags & TDF_DETAILS))
1941 fprintf (dump_file, "overlap steps test failed: no iteration counts.\n");
1943 *overlaps_a = conflict_fn_not_known ();
1944 *overlaps_b = conflict_fn_not_known ();
1945 *last_conflicts = chrec_dont_know;
1949 niter = MIN (niter_x, niter_z);
1950 compute_overlap_steps_for_affine_univar (niter, step_x, step_z,
1953 &last_conflicts_xz, 1);
1954 niter = MIN (niter_y, niter_z);
1955 compute_overlap_steps_for_affine_univar (niter, step_y, step_z,
1958 &last_conflicts_yz, 2);
1959 niter = MIN (niter_x, niter_z);
1960 niter = MIN (niter_y, niter);
1961 compute_overlap_steps_for_affine_univar (niter, step_x + step_y, step_z,
1964 &last_conflicts_xyz, 3);
1966 xz_p = !integer_zerop (last_conflicts_xz);
1967 yz_p = !integer_zerop (last_conflicts_yz);
1968 xyz_p = !integer_zerop (last_conflicts_xyz);
1970 if (xz_p || yz_p || xyz_p)
1972 ova1 = affine_fn_cst (integer_zero_node);
1973 ova2 = affine_fn_cst (integer_zero_node);
1974 ovb = affine_fn_cst (integer_zero_node);
1977 affine_fn t0 = ova1;
1980 ova1 = affine_fn_plus (ova1, overlaps_a_xz);
1981 ovb = affine_fn_plus (ovb, overlaps_b_xz);
1982 affine_fn_free (t0);
1983 affine_fn_free (t2);
1984 *last_conflicts = last_conflicts_xz;
1988 affine_fn t0 = ova2;
1991 ova2 = affine_fn_plus (ova2, overlaps_a_yz);
1992 ovb = affine_fn_plus (ovb, overlaps_b_yz);
1993 affine_fn_free (t0);
1994 affine_fn_free (t2);
1995 *last_conflicts = last_conflicts_yz;
1999 affine_fn t0 = ova1;
2000 affine_fn t2 = ova2;
2003 ova1 = affine_fn_plus (ova1, overlaps_a_xyz);
2004 ova2 = affine_fn_plus (ova2, overlaps_a_xyz);
2005 ovb = affine_fn_plus (ovb, overlaps_b_xyz);
2006 affine_fn_free (t0);
2007 affine_fn_free (t2);
2008 affine_fn_free (t4);
2009 *last_conflicts = last_conflicts_xyz;
2011 *overlaps_a = conflict_fn (2, ova1, ova2);
2012 *overlaps_b = conflict_fn (1, ovb);
2016 *overlaps_a = conflict_fn (1, affine_fn_cst (integer_zero_node));
2017 *overlaps_b = conflict_fn (1, affine_fn_cst (integer_zero_node));
2018 *last_conflicts = integer_zero_node;
2021 affine_fn_free (overlaps_a_xz);
2022 affine_fn_free (overlaps_b_xz);
2023 affine_fn_free (overlaps_a_yz);
2024 affine_fn_free (overlaps_b_yz);
2025 affine_fn_free (overlaps_a_xyz);
2026 affine_fn_free (overlaps_b_xyz);
2029 /* Determines the overlapping elements due to accesses CHREC_A and
2030 CHREC_B, that are affine functions. This function cannot handle
2031 symbolic evolution functions, ie. when initial conditions are
2032 parameters, because it uses lambda matrices of integers. */
2035 analyze_subscript_affine_affine (tree chrec_a,
2037 conflict_function **overlaps_a,
2038 conflict_function **overlaps_b,
2039 tree *last_conflicts)
2041 unsigned nb_vars_a, nb_vars_b, dim;
2042 HOST_WIDE_INT init_a, init_b, gamma, gcd_alpha_beta;
2043 lambda_matrix A, U, S;
2045 if (eq_evolutions_p (chrec_a, chrec_b))
2047 /* The accessed index overlaps for each iteration in the
2049 *overlaps_a = conflict_fn (1, affine_fn_cst (integer_zero_node));
2050 *overlaps_b = conflict_fn (1, affine_fn_cst (integer_zero_node));
2051 *last_conflicts = chrec_dont_know;
2054 if (dump_file && (dump_flags & TDF_DETAILS))
2055 fprintf (dump_file, "(analyze_subscript_affine_affine \n");
2057 /* For determining the initial intersection, we have to solve a
2058 Diophantine equation. This is the most time consuming part.
2060 For answering to the question: "Is there a dependence?" we have
2061 to prove that there exists a solution to the Diophantine
2062 equation, and that the solution is in the iteration domain,
2063 i.e. the solution is positive or zero, and that the solution
2064 happens before the upper bound loop.nb_iterations. Otherwise
2065 there is no dependence. This function outputs a description of
2066 the iterations that hold the intersections. */
2068 nb_vars_a = nb_vars_in_chrec (chrec_a);
2069 nb_vars_b = nb_vars_in_chrec (chrec_b);
2071 dim = nb_vars_a + nb_vars_b;
2072 U = lambda_matrix_new (dim, dim);
2073 A = lambda_matrix_new (dim, 1);
2074 S = lambda_matrix_new (dim, 1);
2076 init_a = initialize_matrix_A (A, chrec_a, 0, 1);
2077 init_b = initialize_matrix_A (A, chrec_b, nb_vars_a, -1);
2078 gamma = init_b - init_a;
2080 /* Don't do all the hard work of solving the Diophantine equation
2081 when we already know the solution: for example,
2084 | gamma = 3 - 3 = 0.
2085 Then the first overlap occurs during the first iterations:
2086 | {3, +, 1}_1 ({0, +, 4}_x) = {3, +, 4}_2 ({0, +, 1}_x)
2090 if (nb_vars_a == 1 && nb_vars_b == 1)
2092 HOST_WIDE_INT step_a, step_b;
2093 HOST_WIDE_INT niter, niter_a, niter_b;
2096 niter_a = estimated_loop_iterations_int (get_chrec_loop (chrec_a),
2098 niter_b = estimated_loop_iterations_int (get_chrec_loop (chrec_b),
2100 niter = MIN (niter_a, niter_b);
2101 step_a = int_cst_value (CHREC_RIGHT (chrec_a));
2102 step_b = int_cst_value (CHREC_RIGHT (chrec_b));
2104 compute_overlap_steps_for_affine_univar (niter, step_a, step_b,
2107 *overlaps_a = conflict_fn (1, ova);
2108 *overlaps_b = conflict_fn (1, ovb);
2111 else if (nb_vars_a == 2 && nb_vars_b == 1)
2112 compute_overlap_steps_for_affine_1_2
2113 (chrec_a, chrec_b, overlaps_a, overlaps_b, last_conflicts);
2115 else if (nb_vars_a == 1 && nb_vars_b == 2)
2116 compute_overlap_steps_for_affine_1_2
2117 (chrec_b, chrec_a, overlaps_b, overlaps_a, last_conflicts);
2121 if (dump_file && (dump_flags & TDF_DETAILS))
2122 fprintf (dump_file, "affine-affine test failed: too many variables.\n");
2123 *overlaps_a = conflict_fn_not_known ();
2124 *overlaps_b = conflict_fn_not_known ();
2125 *last_conflicts = chrec_dont_know;
2127 goto end_analyze_subs_aa;
2131 lambda_matrix_right_hermite (A, dim, 1, S, U);
2136 lambda_matrix_row_negate (U, dim, 0);
2138 gcd_alpha_beta = S[0][0];
2140 /* Something went wrong: for example in {1, +, 0}_5 vs. {0, +, 0}_5,
2141 but that is a quite strange case. Instead of ICEing, answer
2143 if (gcd_alpha_beta == 0)
2145 *overlaps_a = conflict_fn_not_known ();
2146 *overlaps_b = conflict_fn_not_known ();
2147 *last_conflicts = chrec_dont_know;
2148 goto end_analyze_subs_aa;
2151 /* The classic "gcd-test". */
2152 if (!int_divides_p (gcd_alpha_beta, gamma))
2154 /* The "gcd-test" has determined that there is no integer
2155 solution, i.e. there is no dependence. */
2156 *overlaps_a = conflict_fn_no_dependence ();
2157 *overlaps_b = conflict_fn_no_dependence ();
2158 *last_conflicts = integer_zero_node;
2161 /* Both access functions are univariate. This includes SIV and MIV cases. */
2162 else if (nb_vars_a == 1 && nb_vars_b == 1)
2164 /* Both functions should have the same evolution sign. */
2165 if (((A[0][0] > 0 && -A[1][0] > 0)
2166 || (A[0][0] < 0 && -A[1][0] < 0)))
2168 /* The solutions are given by:
2170 | [GAMMA/GCD_ALPHA_BETA t].[u11 u12] = [x0]
2173 For a given integer t. Using the following variables,
2175 | i0 = u11 * gamma / gcd_alpha_beta
2176 | j0 = u12 * gamma / gcd_alpha_beta
2183 | y0 = j0 + j1 * t. */
2184 HOST_WIDE_INT i0, j0, i1, j1;
2186 i0 = U[0][0] * gamma / gcd_alpha_beta;
2187 j0 = U[0][1] * gamma / gcd_alpha_beta;
2191 if ((i1 == 0 && i0 < 0)
2192 || (j1 == 0 && j0 < 0))
2194 /* There is no solution.
2195 FIXME: The case "i0 > nb_iterations, j0 > nb_iterations"
2196 falls in here, but for the moment we don't look at the
2197 upper bound of the iteration domain. */
2198 *overlaps_a = conflict_fn_no_dependence ();
2199 *overlaps_b = conflict_fn_no_dependence ();
2200 *last_conflicts = integer_zero_node;
2201 goto end_analyze_subs_aa;
2204 if (i1 > 0 && j1 > 0)
2206 HOST_WIDE_INT niter_a = estimated_loop_iterations_int
2207 (get_chrec_loop (chrec_a), false);
2208 HOST_WIDE_INT niter_b = estimated_loop_iterations_int
2209 (get_chrec_loop (chrec_b), false);
2210 HOST_WIDE_INT niter = MIN (niter_a, niter_b);
2212 /* (X0, Y0) is a solution of the Diophantine equation:
2213 "chrec_a (X0) = chrec_b (Y0)". */
2214 HOST_WIDE_INT tau1 = MAX (CEIL (-i0, i1),
2216 HOST_WIDE_INT x0 = i1 * tau1 + i0;
2217 HOST_WIDE_INT y0 = j1 * tau1 + j0;
2219 /* (X1, Y1) is the smallest positive solution of the eq
2220 "chrec_a (X1) = chrec_b (Y1)", i.e. this is where the
2221 first conflict occurs. */
2222 HOST_WIDE_INT min_multiple = MIN (x0 / i1, y0 / j1);
2223 HOST_WIDE_INT x1 = x0 - i1 * min_multiple;
2224 HOST_WIDE_INT y1 = y0 - j1 * min_multiple;
2228 HOST_WIDE_INT tau2 = MIN (FLOOR_DIV (niter - i0, i1),
2229 FLOOR_DIV (niter - j0, j1));
2230 HOST_WIDE_INT last_conflict = tau2 - (x1 - i0)/i1;
2232 /* If the overlap occurs outside of the bounds of the
2233 loop, there is no dependence. */
2234 if (x1 > niter || y1 > niter)
2236 *overlaps_a = conflict_fn_no_dependence ();
2237 *overlaps_b = conflict_fn_no_dependence ();
2238 *last_conflicts = integer_zero_node;
2239 goto end_analyze_subs_aa;
2242 *last_conflicts = build_int_cst (NULL_TREE, last_conflict);
2245 *last_conflicts = chrec_dont_know;
2249 affine_fn_univar (build_int_cst (NULL_TREE, x1),
2251 build_int_cst (NULL_TREE, i1)));
2254 affine_fn_univar (build_int_cst (NULL_TREE, y1),
2256 build_int_cst (NULL_TREE, j1)));
2260 /* FIXME: For the moment, the upper bound of the
2261 iteration domain for i and j is not checked. */
2262 if (dump_file && (dump_flags & TDF_DETAILS))
2263 fprintf (dump_file, "affine-affine test failed: unimplemented.\n");
2264 *overlaps_a = conflict_fn_not_known ();
2265 *overlaps_b = conflict_fn_not_known ();
2266 *last_conflicts = chrec_dont_know;
2271 if (dump_file && (dump_flags & TDF_DETAILS))
2272 fprintf (dump_file, "affine-affine test failed: unimplemented.\n");
2273 *overlaps_a = conflict_fn_not_known ();
2274 *overlaps_b = conflict_fn_not_known ();
2275 *last_conflicts = chrec_dont_know;
2280 if (dump_file && (dump_flags & TDF_DETAILS))
2281 fprintf (dump_file, "affine-affine test failed: unimplemented.\n");
2282 *overlaps_a = conflict_fn_not_known ();
2283 *overlaps_b = conflict_fn_not_known ();
2284 *last_conflicts = chrec_dont_know;
2287 end_analyze_subs_aa:
2288 if (dump_file && (dump_flags & TDF_DETAILS))
2290 fprintf (dump_file, " (overlaps_a = ");
2291 dump_conflict_function (dump_file, *overlaps_a);
2292 fprintf (dump_file, ")\n (overlaps_b = ");
2293 dump_conflict_function (dump_file, *overlaps_b);
2294 fprintf (dump_file, ")\n");
2295 fprintf (dump_file, ")\n");
2299 /* Returns true when analyze_subscript_affine_affine can be used for
2300 determining the dependence relation between chrec_a and chrec_b,
2301 that contain symbols. This function modifies chrec_a and chrec_b
2302 such that the analysis result is the same, and such that they don't
2303 contain symbols, and then can safely be passed to the analyzer.
2305 Example: The analysis of the following tuples of evolutions produce
2306 the same results: {x+1, +, 1}_1 vs. {x+3, +, 1}_1, and {-2, +, 1}_1
2309 {x+1, +, 1}_1 ({2, +, 1}_1) = {x+3, +, 1}_1 ({0, +, 1}_1)
2310 {-2, +, 1}_1 ({2, +, 1}_1) = {0, +, 1}_1 ({0, +, 1}_1)
2314 can_use_analyze_subscript_affine_affine (tree *chrec_a, tree *chrec_b)
2316 tree diff, type, left_a, left_b, right_b;
2318 if (chrec_contains_symbols (CHREC_RIGHT (*chrec_a))
2319 || chrec_contains_symbols (CHREC_RIGHT (*chrec_b)))
2320 /* FIXME: For the moment not handled. Might be refined later. */
2323 type = chrec_type (*chrec_a);
2324 left_a = CHREC_LEFT (*chrec_a);
2325 left_b = chrec_convert (type, CHREC_LEFT (*chrec_b), NULL_TREE);
2326 diff = chrec_fold_minus (type, left_a, left_b);
2328 if (!evolution_function_is_constant_p (diff))
2331 if (dump_file && (dump_flags & TDF_DETAILS))
2332 fprintf (dump_file, "can_use_subscript_aff_aff_for_symbolic \n");
2334 *chrec_a = build_polynomial_chrec (CHREC_VARIABLE (*chrec_a),
2335 diff, CHREC_RIGHT (*chrec_a));
2336 right_b = chrec_convert (type, CHREC_RIGHT (*chrec_b), NULL_TREE);
2337 *chrec_b = build_polynomial_chrec (CHREC_VARIABLE (*chrec_b),
2338 build_int_cst (type, 0),
2343 /* Analyze a SIV (Single Index Variable) subscript. *OVERLAPS_A and
2344 *OVERLAPS_B are initialized to the functions that describe the
2345 relation between the elements accessed twice by CHREC_A and
2346 CHREC_B. For k >= 0, the following property is verified:
2348 CHREC_A (*OVERLAPS_A (k)) = CHREC_B (*OVERLAPS_B (k)). */
2351 analyze_siv_subscript (tree chrec_a,
2353 conflict_function **overlaps_a,
2354 conflict_function **overlaps_b,
2355 tree *last_conflicts)
2357 dependence_stats.num_siv++;
2359 if (dump_file && (dump_flags & TDF_DETAILS))
2360 fprintf (dump_file, "(analyze_siv_subscript \n");
2362 if (evolution_function_is_constant_p (chrec_a)
2363 && evolution_function_is_affine_p (chrec_b))
2364 analyze_siv_subscript_cst_affine (chrec_a, chrec_b,
2365 overlaps_a, overlaps_b, last_conflicts);
2367 else if (evolution_function_is_affine_p (chrec_a)
2368 && evolution_function_is_constant_p (chrec_b))
2369 analyze_siv_subscript_cst_affine (chrec_b, chrec_a,
2370 overlaps_b, overlaps_a, last_conflicts);
2372 else if (evolution_function_is_affine_p (chrec_a)
2373 && evolution_function_is_affine_p (chrec_b))
2375 if (!chrec_contains_symbols (chrec_a)
2376 && !chrec_contains_symbols (chrec_b))
2378 analyze_subscript_affine_affine (chrec_a, chrec_b,
2379 overlaps_a, overlaps_b,
2382 if (CF_NOT_KNOWN_P (*overlaps_a)
2383 || CF_NOT_KNOWN_P (*overlaps_b))
2384 dependence_stats.num_siv_unimplemented++;
2385 else if (CF_NO_DEPENDENCE_P (*overlaps_a)
2386 || CF_NO_DEPENDENCE_P (*overlaps_b))
2387 dependence_stats.num_siv_independent++;
2389 dependence_stats.num_siv_dependent++;
2391 else if (can_use_analyze_subscript_affine_affine (&chrec_a,
2394 analyze_subscript_affine_affine (chrec_a, chrec_b,
2395 overlaps_a, overlaps_b,
2398 if (CF_NOT_KNOWN_P (*overlaps_a)
2399 || CF_NOT_KNOWN_P (*overlaps_b))
2400 dependence_stats.num_siv_unimplemented++;
2401 else if (CF_NO_DEPENDENCE_P (*overlaps_a)
2402 || CF_NO_DEPENDENCE_P (*overlaps_b))
2403 dependence_stats.num_siv_independent++;
2405 dependence_stats.num_siv_dependent++;
2408 goto siv_subscript_dontknow;
2413 siv_subscript_dontknow:;
2414 if (dump_file && (dump_flags & TDF_DETAILS))
2415 fprintf (dump_file, "siv test failed: unimplemented.\n");
2416 *overlaps_a = conflict_fn_not_known ();
2417 *overlaps_b = conflict_fn_not_known ();
2418 *last_conflicts = chrec_dont_know;
2419 dependence_stats.num_siv_unimplemented++;
2422 if (dump_file && (dump_flags & TDF_DETAILS))
2423 fprintf (dump_file, ")\n");
2426 /* Returns false if we can prove that the greatest common divisor of the steps
2427 of CHREC does not divide CST, false otherwise. */
2430 gcd_of_steps_may_divide_p (const_tree chrec, const_tree cst)
2432 HOST_WIDE_INT cd = 0, val;
2435 if (!host_integerp (cst, 0))
2437 val = tree_low_cst (cst, 0);
2439 while (TREE_CODE (chrec) == POLYNOMIAL_CHREC)
2441 step = CHREC_RIGHT (chrec);
2442 if (!host_integerp (step, 0))
2444 cd = gcd (cd, tree_low_cst (step, 0));
2445 chrec = CHREC_LEFT (chrec);
2448 return val % cd == 0;
2451 /* Analyze a MIV (Multiple Index Variable) subscript with respect to
2452 LOOP_NEST. *OVERLAPS_A and *OVERLAPS_B are initialized to the
2453 functions that describe the relation between the elements accessed
2454 twice by CHREC_A and CHREC_B. For k >= 0, the following property
2457 CHREC_A (*OVERLAPS_A (k)) = CHREC_B (*OVERLAPS_B (k)). */
2460 analyze_miv_subscript (tree chrec_a,
2462 conflict_function **overlaps_a,
2463 conflict_function **overlaps_b,
2464 tree *last_conflicts,
2465 struct loop *loop_nest)
2467 /* FIXME: This is a MIV subscript, not yet handled.
2468 Example: (A[{1, +, 1}_1] vs. A[{1, +, 1}_2]) that comes from
2471 In the SIV test we had to solve a Diophantine equation with two
2472 variables. In the MIV case we have to solve a Diophantine
2473 equation with 2*n variables (if the subscript uses n IVs).
2475 tree type, difference;
2477 dependence_stats.num_miv++;
2478 if (dump_file && (dump_flags & TDF_DETAILS))
2479 fprintf (dump_file, "(analyze_miv_subscript \n");
2481 type = signed_type_for_types (TREE_TYPE (chrec_a), TREE_TYPE (chrec_b));
2482 chrec_a = chrec_convert (type, chrec_a, NULL_TREE);
2483 chrec_b = chrec_convert (type, chrec_b, NULL_TREE);
2484 difference = chrec_fold_minus (type, chrec_a, chrec_b);
2486 if (eq_evolutions_p (chrec_a, chrec_b))
2488 /* Access functions are the same: all the elements are accessed
2489 in the same order. */
2490 *overlaps_a = conflict_fn (1, affine_fn_cst (integer_zero_node));
2491 *overlaps_b = conflict_fn (1, affine_fn_cst (integer_zero_node));
2492 *last_conflicts = estimated_loop_iterations_tree
2493 (get_chrec_loop (chrec_a), true);
2494 dependence_stats.num_miv_dependent++;
2497 else if (evolution_function_is_constant_p (difference)
2498 /* For the moment, the following is verified:
2499 evolution_function_is_affine_multivariate_p (chrec_a,
2501 && !gcd_of_steps_may_divide_p (chrec_a, difference))
2503 /* testsuite/.../ssa-chrec-33.c
2504 {{21, +, 2}_1, +, -2}_2 vs. {{20, +, 2}_1, +, -2}_2
2506 The difference is 1, and all the evolution steps are multiples
2507 of 2, consequently there are no overlapping elements. */
2508 *overlaps_a = conflict_fn_no_dependence ();
2509 *overlaps_b = conflict_fn_no_dependence ();
2510 *last_conflicts = integer_zero_node;
2511 dependence_stats.num_miv_independent++;
2514 else if (evolution_function_is_affine_multivariate_p (chrec_a, loop_nest->num)
2515 && !chrec_contains_symbols (chrec_a)
2516 && evolution_function_is_affine_multivariate_p (chrec_b, loop_nest->num)
2517 && !chrec_contains_symbols (chrec_b))
2519 /* testsuite/.../ssa-chrec-35.c
2520 {0, +, 1}_2 vs. {0, +, 1}_3
2521 the overlapping elements are respectively located at iterations:
2522 {0, +, 1}_x and {0, +, 1}_x,
2523 in other words, we have the equality:
2524 {0, +, 1}_2 ({0, +, 1}_x) = {0, +, 1}_3 ({0, +, 1}_x)
2527 {{0, +, 1}_1, +, 2}_2 ({0, +, 1}_x, {0, +, 1}_y) =
2528 {0, +, 1}_1 ({{0, +, 1}_x, +, 2}_y)
2530 {{0, +, 2}_1, +, 3}_2 ({0, +, 1}_y, {0, +, 1}_x) =
2531 {{0, +, 3}_1, +, 2}_2 ({0, +, 1}_x, {0, +, 1}_y)
2533 analyze_subscript_affine_affine (chrec_a, chrec_b,
2534 overlaps_a, overlaps_b, last_conflicts);
2536 if (CF_NOT_KNOWN_P (*overlaps_a)
2537 || CF_NOT_KNOWN_P (*overlaps_b))
2538 dependence_stats.num_miv_unimplemented++;
2539 else if (CF_NO_DEPENDENCE_P (*overlaps_a)
2540 || CF_NO_DEPENDENCE_P (*overlaps_b))
2541 dependence_stats.num_miv_independent++;
2543 dependence_stats.num_miv_dependent++;
2548 /* When the analysis is too difficult, answer "don't know". */
2549 if (dump_file && (dump_flags & TDF_DETAILS))
2550 fprintf (dump_file, "analyze_miv_subscript test failed: unimplemented.\n");
2552 *overlaps_a = conflict_fn_not_known ();
2553 *overlaps_b = conflict_fn_not_known ();
2554 *last_conflicts = chrec_dont_know;
2555 dependence_stats.num_miv_unimplemented++;
2558 if (dump_file && (dump_flags & TDF_DETAILS))
2559 fprintf (dump_file, ")\n");
2562 /* Determines the iterations for which CHREC_A is equal to CHREC_B in
2563 with respect to LOOP_NEST. OVERLAP_ITERATIONS_A and
2564 OVERLAP_ITERATIONS_B are initialized with two functions that
2565 describe the iterations that contain conflicting elements.
2567 Remark: For an integer k >= 0, the following equality is true:
2569 CHREC_A (OVERLAP_ITERATIONS_A (k)) == CHREC_B (OVERLAP_ITERATIONS_B (k)).
2573 analyze_overlapping_iterations (tree chrec_a,
2575 conflict_function **overlap_iterations_a,
2576 conflict_function **overlap_iterations_b,
2577 tree *last_conflicts, struct loop *loop_nest)
2579 unsigned int lnn = loop_nest->num;
2581 dependence_stats.num_subscript_tests++;
2583 if (dump_file && (dump_flags & TDF_DETAILS))
2585 fprintf (dump_file, "(analyze_overlapping_iterations \n");
2586 fprintf (dump_file, " (chrec_a = ");
2587 print_generic_expr (dump_file, chrec_a, 0);
2588 fprintf (dump_file, ")\n (chrec_b = ");
2589 print_generic_expr (dump_file, chrec_b, 0);
2590 fprintf (dump_file, ")\n");
2593 if (chrec_a == NULL_TREE
2594 || chrec_b == NULL_TREE
2595 || chrec_contains_undetermined (chrec_a)
2596 || chrec_contains_undetermined (chrec_b))
2598 dependence_stats.num_subscript_undetermined++;
2600 *overlap_iterations_a = conflict_fn_not_known ();
2601 *overlap_iterations_b = conflict_fn_not_known ();
2604 /* If they are the same chrec, and are affine, they overlap
2605 on every iteration. */
2606 else if (eq_evolutions_p (chrec_a, chrec_b)
2607 && evolution_function_is_affine_multivariate_p (chrec_a, lnn))
2609 dependence_stats.num_same_subscript_function++;
2610 *overlap_iterations_a = conflict_fn (1, affine_fn_cst (integer_zero_node));
2611 *overlap_iterations_b = conflict_fn (1, affine_fn_cst (integer_zero_node));
2612 *last_conflicts = chrec_dont_know;
2615 /* If they aren't the same, and aren't affine, we can't do anything
2617 else if ((chrec_contains_symbols (chrec_a)
2618 || chrec_contains_symbols (chrec_b))
2619 && (!evolution_function_is_affine_multivariate_p (chrec_a, lnn)
2620 || !evolution_function_is_affine_multivariate_p (chrec_b, lnn)))
2622 dependence_stats.num_subscript_undetermined++;
2623 *overlap_iterations_a = conflict_fn_not_known ();
2624 *overlap_iterations_b = conflict_fn_not_known ();
2627 else if (ziv_subscript_p (chrec_a, chrec_b))
2628 analyze_ziv_subscript (chrec_a, chrec_b,
2629 overlap_iterations_a, overlap_iterations_b,
2632 else if (siv_subscript_p (chrec_a, chrec_b))
2633 analyze_siv_subscript (chrec_a, chrec_b,
2634 overlap_iterations_a, overlap_iterations_b,
2638 analyze_miv_subscript (chrec_a, chrec_b,
2639 overlap_iterations_a, overlap_iterations_b,
2640 last_conflicts, loop_nest);
2642 if (dump_file && (dump_flags & TDF_DETAILS))
2644 fprintf (dump_file, " (overlap_iterations_a = ");
2645 dump_conflict_function (dump_file, *overlap_iterations_a);
2646 fprintf (dump_file, ")\n (overlap_iterations_b = ");
2647 dump_conflict_function (dump_file, *overlap_iterations_b);
2648 fprintf (dump_file, ")\n");
2649 fprintf (dump_file, ")\n");
2653 /* Helper function for uniquely inserting distance vectors. */
2656 save_dist_v (struct data_dependence_relation *ddr, lambda_vector dist_v)
2661 for (i = 0; VEC_iterate (lambda_vector, DDR_DIST_VECTS (ddr), i, v); i++)
2662 if (lambda_vector_equal (v, dist_v, DDR_NB_LOOPS (ddr)))
2665 VEC_safe_push (lambda_vector, heap, DDR_DIST_VECTS (ddr), dist_v);
2668 /* Helper function for uniquely inserting direction vectors. */
2671 save_dir_v (struct data_dependence_relation *ddr, lambda_vector dir_v)
2676 for (i = 0; VEC_iterate (lambda_vector, DDR_DIR_VECTS (ddr), i, v); i++)
2677 if (lambda_vector_equal (v, dir_v, DDR_NB_LOOPS (ddr)))
2680 VEC_safe_push (lambda_vector, heap, DDR_DIR_VECTS (ddr), dir_v);
2683 /* Add a distance of 1 on all the loops outer than INDEX. If we
2684 haven't yet determined a distance for this outer loop, push a new
2685 distance vector composed of the previous distance, and a distance
2686 of 1 for this outer loop. Example:
2694 Saved vectors are of the form (dist_in_1, dist_in_2). First, we
2695 save (0, 1), then we have to save (1, 0). */
2698 add_outer_distances (struct data_dependence_relation *ddr,
2699 lambda_vector dist_v, int index)
2701 /* For each outer loop where init_v is not set, the accesses are
2702 in dependence of distance 1 in the loop. */
2703 while (--index >= 0)
2705 lambda_vector save_v = lambda_vector_new (DDR_NB_LOOPS (ddr));
2706 lambda_vector_copy (dist_v, save_v, DDR_NB_LOOPS (ddr));
2708 save_dist_v (ddr, save_v);
2712 /* Return false when fail to represent the data dependence as a
2713 distance vector. INIT_B is set to true when a component has been
2714 added to the distance vector DIST_V. INDEX_CARRY is then set to
2715 the index in DIST_V that carries the dependence. */
2718 build_classic_dist_vector_1 (struct data_dependence_relation *ddr,
2719 struct data_reference *ddr_a,
2720 struct data_reference *ddr_b,
2721 lambda_vector dist_v, bool *init_b,
2725 lambda_vector init_v = lambda_vector_new (DDR_NB_LOOPS (ddr));
2727 for (i = 0; i < DDR_NUM_SUBSCRIPTS (ddr); i++)
2729 tree access_fn_a, access_fn_b;
2730 struct subscript *subscript = DDR_SUBSCRIPT (ddr, i);
2732 if (chrec_contains_undetermined (SUB_DISTANCE (subscript)))
2734 non_affine_dependence_relation (ddr);
2738 access_fn_a = DR_ACCESS_FN (ddr_a, i);
2739 access_fn_b = DR_ACCESS_FN (ddr_b, i);
2741 if (TREE_CODE (access_fn_a) == POLYNOMIAL_CHREC
2742 && TREE_CODE (access_fn_b) == POLYNOMIAL_CHREC)
2745 int index_a = index_in_loop_nest (CHREC_VARIABLE (access_fn_a),
2746 DDR_LOOP_NEST (ddr));
2747 int index_b = index_in_loop_nest (CHREC_VARIABLE (access_fn_b),
2748 DDR_LOOP_NEST (ddr));
2750 /* The dependence is carried by the outermost loop. Example:
2757 In this case, the dependence is carried by loop_1. */
2758 index = index_a < index_b ? index_a : index_b;
2759 *index_carry = MIN (index, *index_carry);
2761 if (chrec_contains_undetermined (SUB_DISTANCE (subscript)))
2763 non_affine_dependence_relation (ddr);
2767 dist = int_cst_value (SUB_DISTANCE (subscript));
2769 /* This is the subscript coupling test. If we have already
2770 recorded a distance for this loop (a distance coming from
2771 another subscript), it should be the same. For example,
2772 in the following code, there is no dependence:
2779 if (init_v[index] != 0 && dist_v[index] != dist)
2781 finalize_ddr_dependent (ddr, chrec_known);
2785 dist_v[index] = dist;
2789 else if (!operand_equal_p (access_fn_a, access_fn_b, 0))
2791 /* This can be for example an affine vs. constant dependence
2792 (T[i] vs. T[3]) that is not an affine dependence and is
2793 not representable as a distance vector. */
2794 non_affine_dependence_relation (ddr);
2802 /* Return true when the DDR contains only constant access functions. */
2805 constant_access_functions (const struct data_dependence_relation *ddr)
2809 for (i = 0; i < DDR_NUM_SUBSCRIPTS (ddr); i++)
2810 if (!evolution_function_is_constant_p (DR_ACCESS_FN (DDR_A (ddr), i))
2811 || !evolution_function_is_constant_p (DR_ACCESS_FN (DDR_B (ddr), i)))
2817 /* Helper function for the case where DDR_A and DDR_B are the same
2818 multivariate access function with a constant step. For an example
2822 add_multivariate_self_dist (struct data_dependence_relation *ddr, tree c_2)
2825 tree c_1 = CHREC_LEFT (c_2);
2826 tree c_0 = CHREC_LEFT (c_1);
2827 lambda_vector dist_v;
2830 /* Polynomials with more than 2 variables are not handled yet. When
2831 the evolution steps are parameters, it is not possible to
2832 represent the dependence using classical distance vectors. */
2833 if (TREE_CODE (c_0) != INTEGER_CST
2834 || TREE_CODE (CHREC_RIGHT (c_1)) != INTEGER_CST
2835 || TREE_CODE (CHREC_RIGHT (c_2)) != INTEGER_CST)
2837 DDR_AFFINE_P (ddr) = false;
2841 x_2 = index_in_loop_nest (CHREC_VARIABLE (c_2), DDR_LOOP_NEST (ddr));
2842 x_1 = index_in_loop_nest (CHREC_VARIABLE (c_1), DDR_LOOP_NEST (ddr));
2844 /* For "{{0, +, 2}_1, +, 3}_2" the distance vector is (3, -2). */
2845 dist_v = lambda_vector_new (DDR_NB_LOOPS (ddr));
2846 v1 = int_cst_value (CHREC_RIGHT (c_1));
2847 v2 = int_cst_value (CHREC_RIGHT (c_2));
2860 save_dist_v (ddr, dist_v);
2862 add_outer_distances (ddr, dist_v, x_1);
2865 /* Helper function for the case where DDR_A and DDR_B are the same
2866 access functions. */
2869 add_other_self_distances (struct data_dependence_relation *ddr)
2871 lambda_vector dist_v;
2873 int index_carry = DDR_NB_LOOPS (ddr);
2875 for (i = 0; i < DDR_NUM_SUBSCRIPTS (ddr); i++)
2877 tree access_fun = DR_ACCESS_FN (DDR_A (ddr), i);
2879 if (TREE_CODE (access_fun) == POLYNOMIAL_CHREC)
2881 if (!evolution_function_is_univariate_p (access_fun))
2883 if (DDR_NUM_SUBSCRIPTS (ddr) != 1)
2885 DDR_ARE_DEPENDENT (ddr) = chrec_dont_know;
2889 access_fun = DR_ACCESS_FN (DDR_A (ddr), 0);
2891 if (TREE_CODE (CHREC_LEFT (access_fun)) == POLYNOMIAL_CHREC)
2892 add_multivariate_self_dist (ddr, access_fun);
2894 /* The evolution step is not constant: it varies in
2895 the outer loop, so this cannot be represented by a
2896 distance vector. For example in pr34635.c the
2897 evolution is {0, +, {0, +, 4}_1}_2. */
2898 DDR_AFFINE_P (ddr) = false;
2903 index_carry = MIN (index_carry,
2904 index_in_loop_nest (CHREC_VARIABLE (access_fun),
2905 DDR_LOOP_NEST (ddr)));
2909 dist_v = lambda_vector_new (DDR_NB_LOOPS (ddr));
2910 add_outer_distances (ddr, dist_v, index_carry);
2914 insert_innermost_unit_dist_vector (struct data_dependence_relation *ddr)
2916 lambda_vector dist_v = lambda_vector_new (DDR_NB_LOOPS (ddr));
2918 dist_v[DDR_INNER_LOOP (ddr)] = 1;
2919 save_dist_v (ddr, dist_v);
2922 /* Adds a unit distance vector to DDR when there is a 0 overlap. This
2923 is the case for example when access functions are the same and
2924 equal to a constant, as in:
2931 in which case the distance vectors are (0) and (1). */
2934 add_distance_for_zero_overlaps (struct data_dependence_relation *ddr)
2938 for (i = 0; i < DDR_NUM_SUBSCRIPTS (ddr); i++)
2940 subscript_p sub = DDR_SUBSCRIPT (ddr, i);
2941 conflict_function *ca = SUB_CONFLICTS_IN_A (sub);
2942 conflict_function *cb = SUB_CONFLICTS_IN_B (sub);
2944 for (j = 0; j < ca->n; j++)
2945 if (affine_function_zero_p (ca->fns[j]))
2947 insert_innermost_unit_dist_vector (ddr);
2951 for (j = 0; j < cb->n; j++)
2952 if (affine_function_zero_p (cb->fns[j]))
2954 insert_innermost_unit_dist_vector (ddr);
2960 /* Compute the classic per loop distance vector. DDR is the data
2961 dependence relation to build a vector from. Return false when fail
2962 to represent the data dependence as a distance vector. */
2965 build_classic_dist_vector (struct data_dependence_relation *ddr,
2966 struct loop *loop_nest)
2968 bool init_b = false;
2969 int index_carry = DDR_NB_LOOPS (ddr);
2970 lambda_vector dist_v;
2972 if (DDR_ARE_DEPENDENT (ddr) != NULL_TREE)
2975 if (same_access_functions (ddr))
2977 /* Save the 0 vector. */
2978 dist_v = lambda_vector_new (DDR_NB_LOOPS (ddr));
2979 save_dist_v (ddr, dist_v);
2981 if (constant_access_functions (ddr))
2982 add_distance_for_zero_overlaps (ddr);
2984 if (DDR_NB_LOOPS (ddr) > 1)
2985 add_other_self_distances (ddr);
2990 dist_v = lambda_vector_new (DDR_NB_LOOPS (ddr));
2991 if (!build_classic_dist_vector_1 (ddr, DDR_A (ddr), DDR_B (ddr),
2992 dist_v, &init_b, &index_carry))
2995 /* Save the distance vector if we initialized one. */
2998 /* Verify a basic constraint: classic distance vectors should
2999 always be lexicographically positive.
3001 Data references are collected in the order of execution of
3002 the program, thus for the following loop
3004 | for (i = 1; i < 100; i++)
3005 | for (j = 1; j < 100; j++)
3007 | t = T[j+1][i-1]; // A
3008 | T[j][i] = t + 2; // B
3011 references are collected following the direction of the wind:
3012 A then B. The data dependence tests are performed also
3013 following this order, such that we're looking at the distance
3014 separating the elements accessed by A from the elements later
3015 accessed by B. But in this example, the distance returned by
3016 test_dep (A, B) is lexicographically negative (-1, 1), that
3017 means that the access A occurs later than B with respect to
3018 the outer loop, ie. we're actually looking upwind. In this
3019 case we solve test_dep (B, A) looking downwind to the
3020 lexicographically positive solution, that returns the
3021 distance vector (1, -1). */
3022 if (!lambda_vector_lexico_pos (dist_v, DDR_NB_LOOPS (ddr)))
3024 lambda_vector save_v = lambda_vector_new (DDR_NB_LOOPS (ddr));
3025 if (!subscript_dependence_tester_1 (ddr, DDR_B (ddr), DDR_A (ddr),
3028 compute_subscript_distance (ddr);
3029 if (!build_classic_dist_vector_1 (ddr, DDR_B (ddr), DDR_A (ddr),
3030 save_v, &init_b, &index_carry))
3032 save_dist_v (ddr, save_v);
3033 DDR_REVERSED_P (ddr) = true;
3035 /* In this case there is a dependence forward for all the
3038 | for (k = 1; k < 100; k++)
3039 | for (i = 1; i < 100; i++)
3040 | for (j = 1; j < 100; j++)
3042 | t = T[j+1][i-1]; // A
3043 | T[j][i] = t + 2; // B
3051 if (DDR_NB_LOOPS (ddr) > 1)
3053 add_outer_distances (ddr, save_v, index_carry);
3054 add_outer_distances (ddr, dist_v, index_carry);
3059 lambda_vector save_v = lambda_vector_new (DDR_NB_LOOPS (ddr));
3060 lambda_vector_copy (dist_v, save_v, DDR_NB_LOOPS (ddr));
3062 if (DDR_NB_LOOPS (ddr) > 1)
3064 lambda_vector opposite_v = lambda_vector_new (DDR_NB_LOOPS (ddr));
3066 if (!subscript_dependence_tester_1 (ddr, DDR_B (ddr),
3067 DDR_A (ddr), loop_nest))
3069 compute_subscript_distance (ddr);
3070 if (!build_classic_dist_vector_1 (ddr, DDR_B (ddr), DDR_A (ddr),
3071 opposite_v, &init_b,
3075 save_dist_v (ddr, save_v);
3076 add_outer_distances (ddr, dist_v, index_carry);
3077 add_outer_distances (ddr, opposite_v, index_carry);
3080 save_dist_v (ddr, save_v);
3085 /* There is a distance of 1 on all the outer loops: Example:
3086 there is a dependence of distance 1 on loop_1 for the array A.
3092 add_outer_distances (ddr, dist_v,
3093 lambda_vector_first_nz (dist_v,
3094 DDR_NB_LOOPS (ddr), 0));
3097 if (dump_file && (dump_flags & TDF_DETAILS))
3101 fprintf (dump_file, "(build_classic_dist_vector\n");
3102 for (i = 0; i < DDR_NUM_DIST_VECTS (ddr); i++)
3104 fprintf (dump_file, " dist_vector = (");
3105 print_lambda_vector (dump_file, DDR_DIST_VECT (ddr, i),
3106 DDR_NB_LOOPS (ddr));
3107 fprintf (dump_file, " )\n");
3109 fprintf (dump_file, ")\n");
3115 /* Return the direction for a given distance.
3116 FIXME: Computing dir this way is suboptimal, since dir can catch
3117 cases that dist is unable to represent. */
3119 static inline enum data_dependence_direction
3120 dir_from_dist (int dist)
3123 return dir_positive;
3125 return dir_negative;
3130 /* Compute the classic per loop direction vector. DDR is the data
3131 dependence relation to build a vector from. */
3134 build_classic_dir_vector (struct data_dependence_relation *ddr)
3137 lambda_vector dist_v;
3139 for (i = 0; VEC_iterate (lambda_vector, DDR_DIST_VECTS (ddr), i, dist_v); i++)
3141 lambda_vector dir_v = lambda_vector_new (DDR_NB_LOOPS (ddr));
3143 for (j = 0; j < DDR_NB_LOOPS (ddr); j++)
3144 dir_v[j] = dir_from_dist (dist_v[j]);
3146 save_dir_v (ddr, dir_v);
3150 /* Helper function. Returns true when there is a dependence between
3151 data references DRA and DRB. */
3154 subscript_dependence_tester_1 (struct data_dependence_relation *ddr,
3155 struct data_reference *dra,
3156 struct data_reference *drb,
3157 struct loop *loop_nest)
3160 tree last_conflicts;
3161 struct subscript *subscript;
3163 for (i = 0; VEC_iterate (subscript_p, DDR_SUBSCRIPTS (ddr), i, subscript);
3166 conflict_function *overlaps_a, *overlaps_b;
3168 analyze_overlapping_iterations (DR_ACCESS_FN (dra, i),
3169 DR_ACCESS_FN (drb, i),
3170 &overlaps_a, &overlaps_b,
3171 &last_conflicts, loop_nest);
3173 if (CF_NOT_KNOWN_P (overlaps_a)
3174 || CF_NOT_KNOWN_P (overlaps_b))
3176 finalize_ddr_dependent (ddr, chrec_dont_know);
3177 dependence_stats.num_dependence_undetermined++;
3178 free_conflict_function (overlaps_a);
3179 free_conflict_function (overlaps_b);
3183 else if (CF_NO_DEPENDENCE_P (overlaps_a)
3184 || CF_NO_DEPENDENCE_P (overlaps_b))
3186 finalize_ddr_dependent (ddr, chrec_known);
3187 dependence_stats.num_dependence_independent++;
3188 free_conflict_function (overlaps_a);
3189 free_conflict_function (overlaps_b);
3195 if (SUB_CONFLICTS_IN_A (subscript))
3196 free_conflict_function (SUB_CONFLICTS_IN_A (subscript));
3197 if (SUB_CONFLICTS_IN_B (subscript))
3198 free_conflict_function (SUB_CONFLICTS_IN_B (subscript));
3200 SUB_CONFLICTS_IN_A (subscript) = overlaps_a;
3201 SUB_CONFLICTS_IN_B (subscript) = overlaps_b;
3202 SUB_LAST_CONFLICT (subscript) = last_conflicts;
3209 /* Computes the conflicting iterations in LOOP_NEST, and initialize DDR. */
3212 subscript_dependence_tester (struct data_dependence_relation *ddr,
3213 struct loop *loop_nest)
3216 if (dump_file && (dump_flags & TDF_DETAILS))
3217 fprintf (dump_file, "(subscript_dependence_tester \n");
3219 if (subscript_dependence_tester_1 (ddr, DDR_A (ddr), DDR_B (ddr), loop_nest))
3220 dependence_stats.num_dependence_dependent++;
3222 compute_subscript_distance (ddr);
3223 if (build_classic_dist_vector (ddr, loop_nest))
3224 build_classic_dir_vector (ddr);
3226 if (dump_file && (dump_flags & TDF_DETAILS))
3227 fprintf (dump_file, ")\n");
3230 /* Returns true when all the access functions of A are affine or
3231 constant with respect to LOOP_NEST. */
3234 access_functions_are_affine_or_constant_p (const struct data_reference *a,
3235 const struct loop *loop_nest)
3238 VEC(tree,heap) *fns = DR_ACCESS_FNS (a);
3241 for (i = 0; VEC_iterate (tree, fns, i, t); i++)
3242 if (!evolution_function_is_invariant_p (t, loop_nest->num)
3243 && !evolution_function_is_affine_multivariate_p (t, loop_nest->num))
3249 /* Initializes an equation for an OMEGA problem using the information
3250 contained in the ACCESS_FUN. Returns true when the operation
3253 PB is the omega constraint system.
3254 EQ is the number of the equation to be initialized.
3255 OFFSET is used for shifting the variables names in the constraints:
3256 a constrain is composed of 2 * the number of variables surrounding
3257 dependence accesses. OFFSET is set either to 0 for the first n variables,
3258 then it is set to n.
3259 ACCESS_FUN is expected to be an affine chrec. */
3262 init_omega_eq_with_af (omega_pb pb, unsigned eq,
3263 unsigned int offset, tree access_fun,
3264 struct data_dependence_relation *ddr)
3266 switch (TREE_CODE (access_fun))
3268 case POLYNOMIAL_CHREC:
3270 tree left = CHREC_LEFT (access_fun);
3271 tree right = CHREC_RIGHT (access_fun);
3272 int var = CHREC_VARIABLE (access_fun);
3275 if (TREE_CODE (right) != INTEGER_CST)
3278 var_idx = index_in_loop_nest (var, DDR_LOOP_NEST (ddr));
3279 pb->eqs[eq].coef[offset + var_idx + 1] = int_cst_value (right);
3281 /* Compute the innermost loop index. */
3282 DDR_INNER_LOOP (ddr) = MAX (DDR_INNER_LOOP (ddr), var_idx);
3285 pb->eqs[eq].coef[var_idx + DDR_NB_LOOPS (ddr) + 1]
3286 += int_cst_value (right);
3288 switch (TREE_CODE (left))
3290 case POLYNOMIAL_CHREC:
3291 return init_omega_eq_with_af (pb, eq, offset, left, ddr);
3294 pb->eqs[eq].coef[0] += int_cst_value (left);
3303 pb->eqs[eq].coef[0] += int_cst_value (access_fun);
3311 /* As explained in the comments preceding init_omega_for_ddr, we have
3312 to set up a system for each loop level, setting outer loops
3313 variation to zero, and current loop variation to positive or zero.
3314 Save each lexico positive distance vector. */
3317 omega_extract_distance_vectors (omega_pb pb,
3318 struct data_dependence_relation *ddr)
3322 struct loop *loopi, *loopj;
3323 enum omega_result res;
3325 /* Set a new problem for each loop in the nest. The basis is the
3326 problem that we have initialized until now. On top of this we
3327 add new constraints. */
3328 for (i = 0; i <= DDR_INNER_LOOP (ddr)
3329 && VEC_iterate (loop_p, DDR_LOOP_NEST (ddr), i, loopi); i++)
3332 omega_pb copy = omega_alloc_problem (2 * DDR_NB_LOOPS (ddr),
3333 DDR_NB_LOOPS (ddr));
3335 omega_copy_problem (copy, pb);
3337 /* For all the outer loops "loop_j", add "dj = 0". */
3339 j < i && VEC_iterate (loop_p, DDR_LOOP_NEST (ddr), j, loopj); j++)
3341 eq = omega_add_zero_eq (copy, omega_black);
3342 copy->eqs[eq].coef[j + 1] = 1;
3345 /* For "loop_i", add "0 <= di". */
3346 geq = omega_add_zero_geq (copy, omega_black);
3347 copy->geqs[geq].coef[i + 1] = 1;
3349 /* Reduce the constraint system, and test that the current
3350 problem is feasible. */
3351 res = omega_simplify_problem (copy);
3352 if (res == omega_false
3353 || res == omega_unknown
3354 || copy->num_geqs > (int) DDR_NB_LOOPS (ddr))
3357 for (eq = 0; eq < copy->num_subs; eq++)
3358 if (copy->subs[eq].key == (int) i + 1)
3360 dist = copy->subs[eq].coef[0];
3366 /* Reinitialize problem... */
3367 omega_copy_problem (copy, pb);
3369 j < i && VEC_iterate (loop_p, DDR_LOOP_NEST (ddr), j, loopj); j++)
3371 eq = omega_add_zero_eq (copy, omega_black);
3372 copy->eqs[eq].coef[j + 1] = 1;
3375 /* ..., but this time "di = 1". */
3376 eq = omega_add_zero_eq (copy, omega_black);
3377 copy->eqs[eq].coef[i + 1] = 1;
3378 copy->eqs[eq].coef[0] = -1;
3380 res = omega_simplify_problem (copy);
3381 if (res == omega_false
3382 || res == omega_unknown
3383 || copy->num_geqs > (int) DDR_NB_LOOPS (ddr))
3386 for (eq = 0; eq < copy->num_subs; eq++)
3387 if (copy->subs[eq].key == (int) i + 1)
3389 dist = copy->subs[eq].coef[0];
3395 /* Save the lexicographically positive distance vector. */
3398 lambda_vector dist_v = lambda_vector_new (DDR_NB_LOOPS (ddr));
3399 lambda_vector dir_v = lambda_vector_new (DDR_NB_LOOPS (ddr));
3403 for (eq = 0; eq < copy->num_subs; eq++)
3404 if (copy->subs[eq].key > 0)
3406 dist = copy->subs[eq].coef[0];
3407 dist_v[copy->subs[eq].key - 1] = dist;
3410 for (j = 0; j < DDR_NB_LOOPS (ddr); j++)
3411 dir_v[j] = dir_from_dist (dist_v[j]);
3413 save_dist_v (ddr, dist_v);
3414 save_dir_v (ddr, dir_v);
3418 omega_free_problem (copy);
3422 /* This is called for each subscript of a tuple of data references:
3423 insert an equality for representing the conflicts. */
3426 omega_setup_subscript (tree access_fun_a, tree access_fun_b,
3427 struct data_dependence_relation *ddr,
3428 omega_pb pb, bool *maybe_dependent)
3431 tree type = signed_type_for_types (TREE_TYPE (access_fun_a),
3432 TREE_TYPE (access_fun_b));
3433 tree fun_a = chrec_convert (type, access_fun_a, NULL_TREE);
3434 tree fun_b = chrec_convert (type, access_fun_b, NULL_TREE);
3435 tree difference = chrec_fold_minus (type, fun_a, fun_b);
3437 /* When the fun_a - fun_b is not constant, the dependence is not
3438 captured by the classic distance vector representation. */
3439 if (TREE_CODE (difference) != INTEGER_CST)
3443 if (ziv_subscript_p (fun_a, fun_b) && !integer_zerop (difference))
3445 /* There is no dependence. */
3446 *maybe_dependent = false;
3450 fun_b = chrec_fold_multiply (type, fun_b, integer_minus_one_node);
3452 eq = omega_add_zero_eq (pb, omega_black);
3453 if (!init_omega_eq_with_af (pb, eq, DDR_NB_LOOPS (ddr), fun_a, ddr)
3454 || !init_omega_eq_with_af (pb, eq, 0, fun_b, ddr))
3455 /* There is probably a dependence, but the system of
3456 constraints cannot be built: answer "don't know". */
3460 if (DDR_NB_LOOPS (ddr) != 0 && pb->eqs[eq].coef[0]
3461 && !int_divides_p (lambda_vector_gcd
3462 ((lambda_vector) &(pb->eqs[eq].coef[1]),
3463 2 * DDR_NB_LOOPS (ddr)),
3464 pb->eqs[eq].coef[0]))
3466 /* There is no dependence. */
3467 *maybe_dependent = false;
3474 /* Helper function, same as init_omega_for_ddr but specialized for
3475 data references A and B. */
3478 init_omega_for_ddr_1 (struct data_reference *dra, struct data_reference *drb,
3479 struct data_dependence_relation *ddr,
3480 omega_pb pb, bool *maybe_dependent)
3485 unsigned nb_loops = DDR_NB_LOOPS (ddr);
3487 /* Insert an equality per subscript. */
3488 for (i = 0; i < DDR_NUM_SUBSCRIPTS (ddr); i++)
3490 if (!omega_setup_subscript (DR_ACCESS_FN (dra, i), DR_ACCESS_FN (drb, i),
3491 ddr, pb, maybe_dependent))
3493 else if (*maybe_dependent == false)
3495 /* There is no dependence. */
3496 DDR_ARE_DEPENDENT (ddr) = chrec_known;
3501 /* Insert inequalities: constraints corresponding to the iteration
3502 domain, i.e. the loops surrounding the references "loop_x" and
3503 the distance variables "dx". The layout of the OMEGA
3504 representation is as follows:
3505 - coef[0] is the constant
3506 - coef[1..nb_loops] are the protected variables that will not be
3507 removed by the solver: the "dx"
3508 - coef[nb_loops + 1, 2*nb_loops] are the loop variables: "loop_x".
3510 for (i = 0; i <= DDR_INNER_LOOP (ddr)
3511 && VEC_iterate (loop_p, DDR_LOOP_NEST (ddr), i, loopi); i++)
3513 HOST_WIDE_INT nbi = estimated_loop_iterations_int (loopi, false);
3516 ineq = omega_add_zero_geq (pb, omega_black);
3517 pb->geqs[ineq].coef[i + nb_loops + 1] = 1;
3519 /* 0 <= loop_x + dx */
3520 ineq = omega_add_zero_geq (pb, omega_black);
3521 pb->geqs[ineq].coef[i + nb_loops + 1] = 1;
3522 pb->geqs[ineq].coef[i + 1] = 1;
3526 /* loop_x <= nb_iters */
3527 ineq = omega_add_zero_geq (pb, omega_black);
3528 pb->geqs[ineq].coef[i + nb_loops + 1] = -1;
3529 pb->geqs[ineq].coef[0] = nbi;
3531 /* loop_x + dx <= nb_iters */
3532 ineq = omega_add_zero_geq (pb, omega_black);
3533 pb->geqs[ineq].coef[i + nb_loops + 1] = -1;
3534 pb->geqs[ineq].coef[i + 1] = -1;
3535 pb->geqs[ineq].coef[0] = nbi;
3537 /* A step "dx" bigger than nb_iters is not feasible, so
3538 add "0 <= nb_iters + dx", */
3539 ineq = omega_add_zero_geq (pb, omega_black);
3540 pb->geqs[ineq].coef[i + 1] = 1;
3541 pb->geqs[ineq].coef[0] = nbi;
3542 /* and "dx <= nb_iters". */
3543 ineq = omega_add_zero_geq (pb, omega_black);
3544 pb->geqs[ineq].coef[i + 1] = -1;
3545 pb->geqs[ineq].coef[0] = nbi;
3549 omega_extract_distance_vectors (pb, ddr);
3554 /* Sets up the Omega dependence problem for the data dependence
3555 relation DDR. Returns false when the constraint system cannot be
3556 built, ie. when the test answers "don't know". Returns true
3557 otherwise, and when independence has been proved (using one of the
3558 trivial dependence test), set MAYBE_DEPENDENT to false, otherwise
3559 set MAYBE_DEPENDENT to true.
3561 Example: for setting up the dependence system corresponding to the
3562 conflicting accesses
3567 | ... A[2*j, 2*(i + j)]
3571 the following constraints come from the iteration domain:
3578 where di, dj are the distance variables. The constraints
3579 representing the conflicting elements are:
3582 i + 1 = 2 * (i + di + j + dj)
3584 For asking that the resulting distance vector (di, dj) be
3585 lexicographically positive, we insert the constraint "di >= 0". If
3586 "di = 0" in the solution, we fix that component to zero, and we
3587 look at the inner loops: we set a new problem where all the outer
3588 loop distances are zero, and fix this inner component to be
3589 positive. When one of the components is positive, we save that
3590 distance, and set a new problem where the distance on this loop is
3591 zero, searching for other distances in the inner loops. Here is
3592 the classic example that illustrates that we have to set for each
3593 inner loop a new problem:
3601 we have to save two distances (1, 0) and (0, 1).
3603 Given two array references, refA and refB, we have to set the
3604 dependence problem twice, refA vs. refB and refB vs. refA, and we
3605 cannot do a single test, as refB might occur before refA in the
3606 inner loops, and the contrary when considering outer loops: ex.
3611 | T[{1,+,1}_2][{1,+,1}_1] // refA
3612 | T[{2,+,1}_2][{0,+,1}_1] // refB
3617 refB touches the elements in T before refA, and thus for the same
3618 loop_0 refB precedes refA: ie. the distance vector (0, 1, -1)
3619 but for successive loop_0 iterations, we have (1, -1, 1)
3621 The Omega solver expects the distance variables ("di" in the
3622 previous example) to come first in the constraint system (as
3623 variables to be protected, or "safe" variables), the constraint
3624 system is built using the following layout:
3626 "cst | distance vars | index vars".
3630 init_omega_for_ddr (struct data_dependence_relation *ddr,
3631 bool *maybe_dependent)
3636 *maybe_dependent = true;
3638 if (same_access_functions (ddr))
3641 lambda_vector dir_v;
3643 /* Save the 0 vector. */
3644 save_dist_v (ddr, lambda_vector_new (DDR_NB_LOOPS (ddr)));
3645 dir_v = lambda_vector_new (DDR_NB_LOOPS (ddr));
3646 for (j = 0; j < DDR_NB_LOOPS (ddr); j++)
3647 dir_v[j] = dir_equal;
3648 save_dir_v (ddr, dir_v);
3650 /* Save the dependences carried by outer loops. */
3651 pb = omega_alloc_problem (2 * DDR_NB_LOOPS (ddr), DDR_NB_LOOPS (ddr));
3652 res = init_omega_for_ddr_1 (DDR_A (ddr), DDR_B (ddr), ddr, pb,
3654 omega_free_problem (pb);
3658 /* Omega expects the protected variables (those that have to be kept
3659 after elimination) to appear first in the constraint system.
3660 These variables are the distance variables. In the following
3661 initialization we declare NB_LOOPS safe variables, and the total
3662 number of variables for the constraint system is 2*NB_LOOPS. */
3663 pb = omega_alloc_problem (2 * DDR_NB_LOOPS (ddr), DDR_NB_LOOPS (ddr));
3664 res = init_omega_for_ddr_1 (DDR_A (ddr), DDR_B (ddr), ddr, pb,
3666 omega_free_problem (pb);
3668 /* Stop computation if not decidable, or no dependence. */
3669 if (res == false || *maybe_dependent == false)
3672 pb = omega_alloc_problem (2 * DDR_NB_LOOPS (ddr), DDR_NB_LOOPS (ddr));
3673 res = init_omega_for_ddr_1 (DDR_B (ddr), DDR_A (ddr), ddr, pb,
3675 omega_free_problem (pb);
3680 /* Return true when DDR contains the same information as that stored
3681 in DIR_VECTS and in DIST_VECTS, return false otherwise. */
3684 ddr_consistent_p (FILE *file,
3685 struct data_dependence_relation *ddr,
3686 VEC (lambda_vector, heap) *dist_vects,
3687 VEC (lambda_vector, heap) *dir_vects)
3691 /* If dump_file is set, output there. */
3692 if (dump_file && (dump_flags & TDF_DETAILS))
3695 if (VEC_length (lambda_vector, dist_vects) != DDR_NUM_DIST_VECTS (ddr))
3697 lambda_vector b_dist_v;
3698 fprintf (file, "\n(Number of distance vectors differ: Banerjee has %d, Omega has %d.\n",
3699 VEC_length (lambda_vector, dist_vects),
3700 DDR_NUM_DIST_VECTS (ddr));
3702 fprintf (file, "Banerjee dist vectors:\n");
3703 for (i = 0; VEC_iterate (lambda_vector, dist_vects, i, b_dist_v); i++)
3704 print_lambda_vector (file, b_dist_v, DDR_NB_LOOPS (ddr));
3706 fprintf (file, "Omega dist vectors:\n");
3707 for (i = 0; i < DDR_NUM_DIST_VECTS (ddr); i++)
3708 print_lambda_vector (file, DDR_DIST_VECT (ddr, i), DDR_NB_LOOPS (ddr));
3710 fprintf (file, "data dependence relation:\n");
3711 dump_data_dependence_relation (file, ddr);
3713 fprintf (file, ")\n");
3717 if (VEC_length (lambda_vector, dir_vects) != DDR_NUM_DIR_VECTS (ddr))
3719 fprintf (file, "\n(Number of direction vectors differ: Banerjee has %d, Omega has %d.)\n",
3720 VEC_length (lambda_vector, dir_vects),
3721 DDR_NUM_DIR_VECTS (ddr));
3725 for (i = 0; i < DDR_NUM_DIST_VECTS (ddr); i++)
3727 lambda_vector a_dist_v;
3728 lambda_vector b_dist_v = DDR_DIST_VECT (ddr, i);
3730 /* Distance vectors are not ordered in the same way in the DDR
3731 and in the DIST_VECTS: search for a matching vector. */
3732 for (j = 0; VEC_iterate (lambda_vector, dist_vects, j, a_dist_v); j++)
3733 if (lambda_vector_equal (a_dist_v, b_dist_v, DDR_NB_LOOPS (ddr)))
3736 if (j == VEC_length (lambda_vector, dist_vects))
3738 fprintf (file, "\n(Dist vectors from the first dependence analyzer:\n");
3739 print_dist_vectors (file, dist_vects, DDR_NB_LOOPS (ddr));
3740 fprintf (file, "not found in Omega dist vectors:\n");
3741 print_dist_vectors (file, DDR_DIST_VECTS (ddr), DDR_NB_LOOPS (ddr));
3742 fprintf (file, "data dependence relation:\n");
3743 dump_data_dependence_relation (file, ddr);
3744 fprintf (file, ")\n");
3748 for (i = 0; i < DDR_NUM_DIR_VECTS (ddr); i++)
3750 lambda_vector a_dir_v;
3751 lambda_vector b_dir_v = DDR_DIR_VECT (ddr, i);
3753 /* Direction vectors are not ordered in the same way in the DDR
3754 and in the DIR_VECTS: search for a matching vector. */
3755 for (j = 0; VEC_iterate (lambda_vector, dir_vects, j, a_dir_v); j++)
3756 if (lambda_vector_equal (a_dir_v, b_dir_v, DDR_NB_LOOPS (ddr)))
3759 if (j == VEC_length (lambda_vector, dist_vects))
3761 fprintf (file, "\n(Dir vectors from the first dependence analyzer:\n");
3762 print_dir_vectors (file, dir_vects, DDR_NB_LOOPS (ddr));
3763 fprintf (file, "not found in Omega dir vectors:\n");
3764 print_dir_vectors (file, DDR_DIR_VECTS (ddr), DDR_NB_LOOPS (ddr));
3765 fprintf (file, "data dependence relation:\n");
3766 dump_data_dependence_relation (file, ddr);
3767 fprintf (file, ")\n");
3774 /* This computes the affine dependence relation between A and B with
3775 respect to LOOP_NEST. CHREC_KNOWN is used for representing the
3776 independence between two accesses, while CHREC_DONT_KNOW is used
3777 for representing the unknown relation.
3779 Note that it is possible to stop the computation of the dependence
3780 relation the first time we detect a CHREC_KNOWN element for a given
3784 compute_affine_dependence (struct data_dependence_relation *ddr,
3785 struct loop *loop_nest)
3787 struct data_reference *dra = DDR_A (ddr);
3788 struct data_reference *drb = DDR_B (ddr);
3790 if (dump_file && (dump_flags & TDF_DETAILS))
3792 fprintf (dump_file, "(compute_affine_dependence\n");
3793 fprintf (dump_file, " (stmt_a = \n");
3794 print_generic_expr (dump_file, DR_STMT (dra), 0);
3795 fprintf (dump_file, ")\n (stmt_b = \n");
3796 print_generic_expr (dump_file, DR_STMT (drb), 0);
3797 fprintf (dump_file, ")\n");
3800 /* Analyze only when the dependence relation is not yet known. */
3801 if (DDR_ARE_DEPENDENT (ddr) == NULL_TREE)
3803 dependence_stats.num_dependence_tests++;
3805 if (access_functions_are_affine_or_constant_p (dra, loop_nest)
3806 && access_functions_are_affine_or_constant_p (drb, loop_nest))
3808 if (flag_check_data_deps)
3810 /* Compute the dependences using the first algorithm. */
3811 subscript_dependence_tester (ddr, loop_nest);
3813 if (dump_file && (dump_flags & TDF_DETAILS))
3815 fprintf (dump_file, "\n\nBanerjee Analyzer\n");
3816 dump_data_dependence_relation (dump_file, ddr);
3819 if (DDR_ARE_DEPENDENT (ddr) == NULL_TREE)
3821 bool maybe_dependent;
3822 VEC (lambda_vector, heap) *dir_vects, *dist_vects;
3824 /* Save the result of the first DD analyzer. */
3825 dist_vects = DDR_DIST_VECTS (ddr);
3826 dir_vects = DDR_DIR_VECTS (ddr);
3828 /* Reset the information. */
3829 DDR_DIST_VECTS (ddr) = NULL;
3830 DDR_DIR_VECTS (ddr) = NULL;
3832 /* Compute the same information using Omega. */
3833 if (!init_omega_for_ddr (ddr, &maybe_dependent))
3834 goto csys_dont_know;
3836 if (dump_file && (dump_flags & TDF_DETAILS))
3838 fprintf (dump_file, "Omega Analyzer\n");
3839 dump_data_dependence_relation (dump_file, ddr);
3842 /* Check that we get the same information. */
3843 if (maybe_dependent)
3844 gcc_assert (ddr_consistent_p (stderr, ddr, dist_vects,
3849 subscript_dependence_tester (ddr, loop_nest);
3852 /* As a last case, if the dependence cannot be determined, or if
3853 the dependence is considered too difficult to determine, answer
3858 dependence_stats.num_dependence_undetermined++;
3860 if (dump_file && (dump_flags & TDF_DETAILS))
3862 fprintf (dump_file, "Data ref a:\n");
3863 dump_data_reference (dump_file, dra);
3864 fprintf (dump_file, "Data ref b:\n");
3865 dump_data_reference (dump_file, drb);
3866 fprintf (dump_file, "affine dependence test not usable: access function not affine or constant.\n");
3868 finalize_ddr_dependent (ddr, chrec_dont_know);
3872 if (dump_file && (dump_flags & TDF_DETAILS))
3873 fprintf (dump_file, ")\n");
3876 /* This computes the dependence relation for the same data
3877 reference into DDR. */
3880 compute_self_dependence (struct data_dependence_relation *ddr)
3883 struct subscript *subscript;
3885 if (DDR_ARE_DEPENDENT (ddr) != NULL_TREE)
3888 for (i = 0; VEC_iterate (subscript_p, DDR_SUBSCRIPTS (ddr), i, subscript);
3891 if (SUB_CONFLICTS_IN_A (subscript))
3892 free_conflict_function (SUB_CONFLICTS_IN_A (subscript));
3893 if (SUB_CONFLICTS_IN_B (subscript))
3894 free_conflict_function (SUB_CONFLICTS_IN_B (subscript));
3896 /* The accessed index overlaps for each iteration. */
3897 SUB_CONFLICTS_IN_A (subscript)
3898 = conflict_fn (1, affine_fn_cst (integer_zero_node));
3899 SUB_CONFLICTS_IN_B (subscript)
3900 = conflict_fn (1, affine_fn_cst (integer_zero_node));
3901 SUB_LAST_CONFLICT (subscript) = chrec_dont_know;
3904 /* The distance vector is the zero vector. */
3905 save_dist_v (ddr, lambda_vector_new (DDR_NB_LOOPS (ddr)));
3906 save_dir_v (ddr, lambda_vector_new (DDR_NB_LOOPS (ddr)));
3909 /* Compute in DEPENDENCE_RELATIONS the data dependence graph for all
3910 the data references in DATAREFS, in the LOOP_NEST. When
3911 COMPUTE_SELF_AND_RR is FALSE, don't compute read-read and self
3915 compute_all_dependences (VEC (data_reference_p, heap) *datarefs,
3916 VEC (ddr_p, heap) **dependence_relations,
3917 VEC (loop_p, heap) *loop_nest,
3918 bool compute_self_and_rr)
3920 struct data_dependence_relation *ddr;
3921 struct data_reference *a, *b;
3924 for (i = 0; VEC_iterate (data_reference_p, datarefs, i, a); i++)
3925 for (j = i + 1; VEC_iterate (data_reference_p, datarefs, j, b); j++)
3926 if (!DR_IS_READ (a) || !DR_IS_READ (b) || compute_self_and_rr)
3928 ddr = initialize_data_dependence_relation (a, b, loop_nest);
3929 VEC_safe_push (ddr_p, heap, *dependence_relations, ddr);
3930 compute_affine_dependence (ddr, VEC_index (loop_p, loop_nest, 0));
3933 if (compute_self_and_rr)
3934 for (i = 0; VEC_iterate (data_reference_p, datarefs, i, a); i++)
3936 ddr = initialize_data_dependence_relation (a, a, loop_nest);
3937 VEC_safe_push (ddr_p, heap, *dependence_relations, ddr);
3938 compute_self_dependence (ddr);
3942 /* Stores the locations of memory references in STMT to REFERENCES. Returns
3943 true if STMT clobbers memory, false otherwise. */
3946 get_references_in_stmt (tree stmt, VEC (data_ref_loc, heap) **references)
3948 bool clobbers_memory = false;
3950 tree *op0, *op1, call;
3954 /* ASM_EXPR and CALL_EXPR may embed arbitrary side effects.
3955 Calls have side-effects, except those to const or pure
3957 call = get_call_expr_in (stmt);
3959 && !(call_expr_flags (call) & (ECF_CONST | ECF_PURE)))
3960 || (TREE_CODE (stmt) == ASM_EXPR
3961 && ASM_VOLATILE_P (stmt)))
3962 clobbers_memory = true;
3964 if (ZERO_SSA_OPERANDS (stmt, SSA_OP_ALL_VIRTUALS))
3965 return clobbers_memory;
3967 if (TREE_CODE (stmt) == GIMPLE_MODIFY_STMT)
3970 op0 = &GIMPLE_STMT_OPERAND (stmt, 0);
3971 op1 = &GIMPLE_STMT_OPERAND (stmt, 1);
3974 || (REFERENCE_CLASS_P (*op1)
3975 && (base = get_base_address (*op1))
3976 && TREE_CODE (base) != SSA_NAME))
3978 ref = VEC_safe_push (data_ref_loc, heap, *references, NULL);
3980 ref->is_read = true;
3984 || (REFERENCE_CLASS_P (*op0) && get_base_address (*op0)))
3986 ref = VEC_safe_push (data_ref_loc, heap, *references, NULL);
3988 ref->is_read = false;
3994 unsigned i, n = call_expr_nargs (call);
3996 for (i = 0; i < n; i++)
3998 op0 = &CALL_EXPR_ARG (call, i);
4001 || (REFERENCE_CLASS_P (*op0) && get_base_address (*op0)))
4003 ref = VEC_safe_push (data_ref_loc, heap, *references, NULL);
4005 ref->is_read = true;
4010 return clobbers_memory;
4013 /* Stores the data references in STMT to DATAREFS. If there is an unanalyzable
4014 reference, returns false, otherwise returns true. NEST is the outermost
4015 loop of the loop nest in that the references should be analyzed. */
4018 find_data_references_in_stmt (struct loop *nest, tree stmt,
4019 VEC (data_reference_p, heap) **datarefs)
4022 VEC (data_ref_loc, heap) *references;
4025 data_reference_p dr;
4027 if (get_references_in_stmt (stmt, &references))
4029 VEC_free (data_ref_loc, heap, references);
4033 for (i = 0; VEC_iterate (data_ref_loc, references, i, ref); i++)
4035 dr = create_data_ref (nest, *ref->pos, stmt, ref->is_read);
4036 gcc_assert (dr != NULL);
4038 /* FIXME -- data dependence analysis does not work correctly for objects with
4039 invariant addresses. Let us fail here until the problem is fixed. */
4040 if (dr_address_invariant_p (dr))
4043 if (dump_file && (dump_flags & TDF_DETAILS))
4044 fprintf (dump_file, "\tFAILED as dr address is invariant\n");
4049 VEC_safe_push (data_reference_p, heap, *datarefs, dr);
4051 VEC_free (data_ref_loc, heap, references);
4055 /* Search the data references in LOOP, and record the information into
4056 DATAREFS. Returns chrec_dont_know when failing to analyze a
4057 difficult case, returns NULL_TREE otherwise.
4059 TODO: This function should be made smarter so that it can handle address
4060 arithmetic as if they were array accesses, etc. */
4063 find_data_references_in_loop (struct loop *loop,
4064 VEC (data_reference_p, heap) **datarefs)
4066 basic_block bb, *bbs;
4068 block_stmt_iterator bsi;
4070 bbs = get_loop_body_in_dom_order (loop);
4072 for (i = 0; i < loop->num_nodes; i++)
4076 for (bsi = bsi_start (bb); !bsi_end_p (bsi); bsi_next (&bsi))
4078 tree stmt = bsi_stmt (bsi);
4080 if (!find_data_references_in_stmt (loop, stmt, datarefs))
4082 struct data_reference *res;
4083 res = XCNEW (struct data_reference);
4084 VEC_safe_push (data_reference_p, heap, *datarefs, res);
4087 return chrec_dont_know;
4096 /* Recursive helper function. */
4099 find_loop_nest_1 (struct loop *loop, VEC (loop_p, heap) **loop_nest)
4101 /* Inner loops of the nest should not contain siblings. Example:
4102 when there are two consecutive loops,
4113 the dependence relation cannot be captured by the distance
4118 VEC_safe_push (loop_p, heap, *loop_nest, loop);
4120 return find_loop_nest_1 (loop->inner, loop_nest);
4124 /* Return false when the LOOP is not well nested. Otherwise return
4125 true and insert in LOOP_NEST the loops of the nest. LOOP_NEST will
4126 contain the loops from the outermost to the innermost, as they will
4127 appear in the classic distance vector. */
4130 find_loop_nest (struct loop *loop, VEC (loop_p, heap) **loop_nest)
4132 VEC_safe_push (loop_p, heap, *loop_nest, loop);
4134 return find_loop_nest_1 (loop->inner, loop_nest);
4138 /* Given a loop nest LOOP, the following vectors are returned:
4139 DATAREFS is initialized to all the array elements contained in this loop,
4140 DEPENDENCE_RELATIONS contains the relations between the data references.
4141 Compute read-read and self relations if
4142 COMPUTE_SELF_AND_READ_READ_DEPENDENCES is TRUE. */
4145 compute_data_dependences_for_loop (struct loop *loop,
4146 bool compute_self_and_read_read_dependences,
4147 VEC (data_reference_p, heap) **datarefs,
4148 VEC (ddr_p, heap) **dependence_relations)
4150 VEC (loop_p, heap) *vloops = VEC_alloc (loop_p, heap, 3);
4152 memset (&dependence_stats, 0, sizeof (dependence_stats));
4154 /* If the loop nest is not well formed, or one of the data references
4155 is not computable, give up without spending time to compute other
4158 || !find_loop_nest (loop, &vloops)
4159 || find_data_references_in_loop (loop, datarefs) == chrec_dont_know)
4161 struct data_dependence_relation *ddr;
4163 /* Insert a single relation into dependence_relations:
4165 ddr = initialize_data_dependence_relation (NULL, NULL, vloops);
4166 VEC_safe_push (ddr_p, heap, *dependence_relations, ddr);
4169 compute_all_dependences (*datarefs, dependence_relations, vloops,
4170 compute_self_and_read_read_dependences);
4172 if (dump_file && (dump_flags & TDF_STATS))
4174 fprintf (dump_file, "Dependence tester statistics:\n");
4176 fprintf (dump_file, "Number of dependence tests: %d\n",
4177 dependence_stats.num_dependence_tests);
4178 fprintf (dump_file, "Number of dependence tests classified dependent: %d\n",
4179 dependence_stats.num_dependence_dependent);
4180 fprintf (dump_file, "Number of dependence tests classified independent: %d\n",
4181 dependence_stats.num_dependence_independent);
4182 fprintf (dump_file, "Number of undetermined dependence tests: %d\n",
4183 dependence_stats.num_dependence_undetermined);
4185 fprintf (dump_file, "Number of subscript tests: %d\n",
4186 dependence_stats.num_subscript_tests);
4187 fprintf (dump_file, "Number of undetermined subscript tests: %d\n",
4188 dependence_stats.num_subscript_undetermined);
4189 fprintf (dump_file, "Number of same subscript function: %d\n",
4190 dependence_stats.num_same_subscript_function);
4192 fprintf (dump_file, "Number of ziv tests: %d\n",
4193 dependence_stats.num_ziv);
4194 fprintf (dump_file, "Number of ziv tests returning dependent: %d\n",
4195 dependence_stats.num_ziv_dependent);
4196 fprintf (dump_file, "Number of ziv tests returning independent: %d\n",
4197 dependence_stats.num_ziv_independent);
4198 fprintf (dump_file, "Number of ziv tests unimplemented: %d\n",
4199 dependence_stats.num_ziv_unimplemented);
4201 fprintf (dump_file, "Number of siv tests: %d\n",
4202 dependence_stats.num_siv);
4203 fprintf (dump_file, "Number of siv tests returning dependent: %d\n",
4204 dependence_stats.num_siv_dependent);
4205 fprintf (dump_file, "Number of siv tests returning independent: %d\n",
4206 dependence_stats.num_siv_independent);
4207 fprintf (dump_file, "Number of siv tests unimplemented: %d\n",
4208 dependence_stats.num_siv_unimplemented);
4210 fprintf (dump_file, "Number of miv tests: %d\n",
4211 dependence_stats.num_miv);
4212 fprintf (dump_file, "Number of miv tests returning dependent: %d\n",
4213 dependence_stats.num_miv_dependent);
4214 fprintf (dump_file, "Number of miv tests returning independent: %d\n",
4215 dependence_stats.num_miv_independent);
4216 fprintf (dump_file, "Number of miv tests unimplemented: %d\n",
4217 dependence_stats.num_miv_unimplemented);
4221 /* Entry point (for testing only). Analyze all the data references
4222 and the dependence relations in LOOP.
4224 The data references are computed first.
4226 A relation on these nodes is represented by a complete graph. Some
4227 of the relations could be of no interest, thus the relations can be
4230 In the following function we compute all the relations. This is
4231 just a first implementation that is here for:
4232 - for showing how to ask for the dependence relations,
4233 - for the debugging the whole dependence graph,
4234 - for the dejagnu testcases and maintenance.
4236 It is possible to ask only for a part of the graph, avoiding to
4237 compute the whole dependence graph. The computed dependences are
4238 stored in a knowledge base (KB) such that later queries don't
4239 recompute the same information. The implementation of this KB is
4240 transparent to the optimizer, and thus the KB can be changed with a
4241 more efficient implementation, or the KB could be disabled. */
4243 analyze_all_data_dependences (struct loop *loop)
4246 int nb_data_refs = 10;
4247 VEC (data_reference_p, heap) *datarefs =
4248 VEC_alloc (data_reference_p, heap, nb_data_refs);
4249 VEC (ddr_p, heap) *dependence_relations =
4250 VEC_alloc (ddr_p, heap, nb_data_refs * nb_data_refs);
4252 /* Compute DDs on the whole function. */
4253 compute_data_dependences_for_loop (loop, false, &datarefs,
4254 &dependence_relations);
4258 dump_data_dependence_relations (dump_file, dependence_relations);
4259 fprintf (dump_file, "\n\n");
4261 if (dump_flags & TDF_DETAILS)
4262 dump_dist_dir_vectors (dump_file, dependence_relations);
4264 if (dump_flags & TDF_STATS)
4266 unsigned nb_top_relations = 0;
4267 unsigned nb_bot_relations = 0;
4268 unsigned nb_basename_differ = 0;
4269 unsigned nb_chrec_relations = 0;
4270 struct data_dependence_relation *ddr;
4272 for (i = 0; VEC_iterate (ddr_p, dependence_relations, i, ddr); i++)
4274 if (chrec_contains_undetermined (DDR_ARE_DEPENDENT (ddr)))
4277 else if (DDR_ARE_DEPENDENT (ddr) == chrec_known)
4279 struct data_reference *a = DDR_A (ddr);
4280 struct data_reference *b = DDR_B (ddr);
4282 if (!bitmap_intersect_p (DR_VOPS (a), DR_VOPS (b)))
4283 nb_basename_differ++;
4289 nb_chrec_relations++;
4292 gather_stats_on_scev_database ();
4296 free_dependence_relations (dependence_relations);
4297 free_data_refs (datarefs);
4300 /* Computes all the data dependences and check that the results of
4301 several analyzers are the same. */
4304 tree_check_data_deps (void)
4307 struct loop *loop_nest;
4309 FOR_EACH_LOOP (li, loop_nest, 0)
4310 analyze_all_data_dependences (loop_nest);
4313 /* Free the memory used by a data dependence relation DDR. */
4316 free_dependence_relation (struct data_dependence_relation *ddr)
4321 if (DDR_SUBSCRIPTS (ddr))
4322 free_subscripts (DDR_SUBSCRIPTS (ddr));
4323 if (DDR_DIST_VECTS (ddr))
4324 VEC_free (lambda_vector, heap, DDR_DIST_VECTS (ddr));
4325 if (DDR_DIR_VECTS (ddr))
4326 VEC_free (lambda_vector, heap, DDR_DIR_VECTS (ddr));
4331 /* Free the memory used by the data dependence relations from
4332 DEPENDENCE_RELATIONS. */
4335 free_dependence_relations (VEC (ddr_p, heap) *dependence_relations)
4338 struct data_dependence_relation *ddr;
4339 VEC (loop_p, heap) *loop_nest = NULL;
4341 for (i = 0; VEC_iterate (ddr_p, dependence_relations, i, ddr); i++)
4345 if (loop_nest == NULL)
4346 loop_nest = DDR_LOOP_NEST (ddr);
4348 gcc_assert (DDR_LOOP_NEST (ddr) == NULL
4349 || DDR_LOOP_NEST (ddr) == loop_nest);
4350 free_dependence_relation (ddr);
4354 VEC_free (loop_p, heap, loop_nest);
4355 VEC_free (ddr_p, heap, dependence_relations);
4358 /* Free the memory used by the data references from DATAREFS. */
4361 free_data_refs (VEC (data_reference_p, heap) *datarefs)
4364 struct data_reference *dr;
4366 for (i = 0; VEC_iterate (data_reference_p, datarefs, i, dr); i++)
4368 VEC_free (data_reference_p, heap, datarefs);
4373 /* Dump vertex I in RDG to FILE. */
4376 dump_rdg_vertex (FILE *file, struct graph *rdg, int i)
4378 struct vertex *v = &(rdg->vertices[i]);
4379 struct graph_edge *e;
4381 fprintf (file, "(vertex %d: (%s%s) (in:", i,
4382 RDG_MEM_WRITE_STMT (rdg, i) ? "w" : "",
4383 RDG_MEM_READS_STMT (rdg, i) ? "r" : "");
4386 for (e = v->pred; e; e = e->pred_next)
4387 fprintf (file, " %d", e->src);
4389 fprintf (file, ") (out:");
4392 for (e = v->succ; e; e = e->succ_next)
4393 fprintf (file, " %d", e->dest);
4395 fprintf (file, ") \n");
4396 print_generic_stmt (file, RDGV_STMT (v), TDF_VOPS|TDF_MEMSYMS);
4397 fprintf (file, ")\n");
4400 /* Call dump_rdg_vertex on stderr. */
4403 debug_rdg_vertex (struct graph *rdg, int i)
4405 dump_rdg_vertex (stderr, rdg, i);
4408 /* Dump component C of RDG to FILE. If DUMPED is non-null, set the
4409 dumped vertices to that bitmap. */
4411 void dump_rdg_component (FILE *file, struct graph *rdg, int c, bitmap dumped)
4415 fprintf (file, "(%d\n", c);
4417 for (i = 0; i < rdg->n_vertices; i++)
4418 if (rdg->vertices[i].component == c)
4421 bitmap_set_bit (dumped, i);
4423 dump_rdg_vertex (file, rdg, i);
4426 fprintf (file, ")\n");
4429 /* Call dump_rdg_vertex on stderr. */
4432 debug_rdg_component (struct graph *rdg, int c)
4434 dump_rdg_component (stderr, rdg, c, NULL);
4437 /* Dump the reduced dependence graph RDG to FILE. */
4440 dump_rdg (FILE *file, struct graph *rdg)
4443 bitmap dumped = BITMAP_ALLOC (NULL);
4445 fprintf (file, "(rdg\n");
4447 for (i = 0; i < rdg->n_vertices; i++)
4448 if (!bitmap_bit_p (dumped, i))
4449 dump_rdg_component (file, rdg, rdg->vertices[i].component, dumped);
4451 fprintf (file, ")\n");
4452 BITMAP_FREE (dumped);
4455 /* Call dump_rdg on stderr. */
4458 debug_rdg (struct graph *rdg)
4460 dump_rdg (stderr, rdg);
4464 dot_rdg_1 (FILE *file, struct graph *rdg)
4468 fprintf (file, "digraph RDG {\n");
4470 for (i = 0; i < rdg->n_vertices; i++)
4472 struct vertex *v = &(rdg->vertices[i]);
4473 struct graph_edge *e;
4475 /* Highlight reads from memory. */
4476 if (RDG_MEM_READS_STMT (rdg, i))
4477 fprintf (file, "%d [style=filled, fillcolor=green]\n", i);
4479 /* Highlight stores to memory. */
4480 if (RDG_MEM_WRITE_STMT (rdg, i))
4481 fprintf (file, "%d [style=filled, fillcolor=red]\n", i);
4484 for (e = v->succ; e; e = e->succ_next)
4485 switch (RDGE_TYPE (e))
4488 fprintf (file, "%d -> %d [label=input] \n", i, e->dest);
4492 fprintf (file, "%d -> %d [label=output] \n", i, e->dest);
4496 /* These are the most common dependences: don't print these. */
4497 fprintf (file, "%d -> %d \n", i, e->dest);
4501 fprintf (file, "%d -> %d [label=anti] \n", i, e->dest);
4509 fprintf (file, "}\n\n");
4512 /* Display SCOP using dotty. */
4515 dot_rdg (struct graph *rdg)
4517 FILE *file = fopen ("/tmp/rdg.dot", "w");
4518 gcc_assert (file != NULL);
4520 dot_rdg_1 (file, rdg);
4523 system ("dotty /tmp/rdg.dot");
4527 /* This structure is used for recording the mapping statement index in
4530 struct rdg_vertex_info GTY(())
4536 /* Returns the index of STMT in RDG. */
4539 rdg_vertex_for_stmt (struct graph *rdg, tree stmt)
4541 struct rdg_vertex_info rvi, *slot;
4544 slot = (struct rdg_vertex_info *) htab_find (rdg->indices, &rvi);
4552 /* Creates an edge in RDG for each distance vector from DDR. The
4553 order that we keep track of in the RDG is the order in which
4554 statements have to be executed. */
4557 create_rdg_edge_for_ddr (struct graph *rdg, ddr_p ddr)
4559 struct graph_edge *e;
4561 data_reference_p dra = DDR_A (ddr);
4562 data_reference_p drb = DDR_B (ddr);
4563 unsigned level = ddr_dependence_level (ddr);
4565 /* For non scalar dependences, when the dependence is REVERSED,
4566 statement B has to be executed before statement A. */
4568 && !DDR_REVERSED_P (ddr))
4570 data_reference_p tmp = dra;
4575 va = rdg_vertex_for_stmt (rdg, DR_STMT (dra));
4576 vb = rdg_vertex_for_stmt (rdg, DR_STMT (drb));
4578 if (va < 0 || vb < 0)
4581 e = add_edge (rdg, va, vb);
4582 e->data = XNEW (struct rdg_edge);
4584 RDGE_LEVEL (e) = level;
4586 /* Determines the type of the data dependence. */
4587 if (DR_IS_READ (dra) && DR_IS_READ (drb))
4588 RDGE_TYPE (e) = input_dd;
4589 else if (!DR_IS_READ (dra) && !DR_IS_READ (drb))
4590 RDGE_TYPE (e) = output_dd;
4591 else if (!DR_IS_READ (dra) && DR_IS_READ (drb))
4592 RDGE_TYPE (e) = flow_dd;
4593 else if (DR_IS_READ (dra) && !DR_IS_READ (drb))
4594 RDGE_TYPE (e) = anti_dd;
4597 /* Creates dependence edges in RDG for all the uses of DEF. IDEF is
4598 the index of DEF in RDG. */
4601 create_rdg_edges_for_scalar (struct graph *rdg, tree def, int idef)
4603 use_operand_p imm_use_p;
4604 imm_use_iterator iterator;
4606 FOR_EACH_IMM_USE_FAST (imm_use_p, iterator, def)
4608 struct graph_edge *e;
4609 int use = rdg_vertex_for_stmt (rdg, USE_STMT (imm_use_p));
4614 e = add_edge (rdg, idef, use);
4615 e->data = XNEW (struct rdg_edge);
4616 RDGE_TYPE (e) = flow_dd;
4620 /* Creates the edges of the reduced dependence graph RDG. */
4623 create_rdg_edges (struct graph *rdg, VEC (ddr_p, heap) *ddrs)
4626 struct data_dependence_relation *ddr;
4627 def_operand_p def_p;
4630 for (i = 0; VEC_iterate (ddr_p, ddrs, i, ddr); i++)
4631 if (DDR_ARE_DEPENDENT (ddr) == NULL_TREE)
4632 create_rdg_edge_for_ddr (rdg, ddr);
4634 for (i = 0; i < rdg->n_vertices; i++)
4635 FOR_EACH_PHI_OR_STMT_DEF (def_p, RDG_STMT (rdg, i),
4637 create_rdg_edges_for_scalar (rdg, DEF_FROM_PTR (def_p), i);
4640 /* Build the vertices of the reduced dependence graph RDG. */
4643 create_rdg_vertices (struct graph *rdg, VEC (tree, heap) *stmts)
4648 for (i = 0; VEC_iterate (tree, stmts, i, stmt); i++)
4650 VEC (data_ref_loc, heap) *references;
4652 struct vertex *v = &(rdg->vertices[i]);
4653 struct rdg_vertex_info *rvi = XNEW (struct rdg_vertex_info);
4654 struct rdg_vertex_info **slot;
4658 slot = (struct rdg_vertex_info **) htab_find_slot (rdg->indices, rvi, INSERT);
4665 v->data = XNEW (struct rdg_vertex);
4666 RDG_STMT (rdg, i) = stmt;
4668 RDG_MEM_WRITE_STMT (rdg, i) = false;
4669 RDG_MEM_READS_STMT (rdg, i) = false;
4670 if (TREE_CODE (stmt) == PHI_NODE)
4673 get_references_in_stmt (stmt, &references);
4674 for (j = 0; VEC_iterate (data_ref_loc, references, j, ref); j++)
4676 RDG_MEM_WRITE_STMT (rdg, i) = true;
4678 RDG_MEM_READS_STMT (rdg, i) = true;
4680 VEC_free (data_ref_loc, heap, references);
4684 /* Initialize STMTS with all the statements of LOOP. When
4685 INCLUDE_PHIS is true, include also the PHI nodes. The order in
4686 which we discover statements is important as
4687 generate_loops_for_partition is using the same traversal for
4688 identifying statements. */
4691 stmts_from_loop (struct loop *loop, VEC (tree, heap) **stmts)
4694 basic_block *bbs = get_loop_body_in_dom_order (loop);
4696 for (i = 0; i < loop->num_nodes; i++)
4699 basic_block bb = bbs[i];
4700 block_stmt_iterator bsi;
4702 for (phi = phi_nodes (bb); phi; phi = PHI_CHAIN (phi))
4703 VEC_safe_push (tree, heap, *stmts, phi);
4705 for (bsi = bsi_start (bb); !bsi_end_p (bsi); bsi_next (&bsi))
4706 if (TREE_CODE (stmt = bsi_stmt (bsi)) != LABEL_EXPR)
4707 VEC_safe_push (tree, heap, *stmts, stmt);
4713 /* Returns true when all the dependences are computable. */
4716 known_dependences_p (VEC (ddr_p, heap) *dependence_relations)
4721 for (i = 0; VEC_iterate (ddr_p, dependence_relations, i, ddr); i++)
4722 if (DDR_ARE_DEPENDENT (ddr) == chrec_dont_know)
4728 /* Computes a hash function for element ELT. */
4731 hash_stmt_vertex_info (const void *elt)
4733 struct rdg_vertex_info *rvi = (struct rdg_vertex_info *) elt;
4734 tree stmt = rvi->stmt;
4736 return htab_hash_pointer (stmt);
4739 /* Compares database elements E1 and E2. */
4742 eq_stmt_vertex_info (const void *e1, const void *e2)
4744 const struct rdg_vertex_info *elt1 = (const struct rdg_vertex_info *) e1;
4745 const struct rdg_vertex_info *elt2 = (const struct rdg_vertex_info *) e2;
4747 return elt1->stmt == elt2->stmt;
4750 /* Free the element E. */
4753 hash_stmt_vertex_del (void *e)
4758 /* Build the Reduced Dependence Graph (RDG) with one vertex per
4759 statement of the loop nest, and one edge per data dependence or
4760 scalar dependence. */
4763 build_rdg (struct loop *loop)
4765 int nb_data_refs = 10;
4766 struct graph *rdg = NULL;
4767 VEC (ddr_p, heap) *dependence_relations;
4768 VEC (data_reference_p, heap) *datarefs;
4769 VEC (tree, heap) *stmts = VEC_alloc (tree, heap, nb_data_refs);
4771 dependence_relations = VEC_alloc (ddr_p, heap, nb_data_refs * nb_data_refs) ;
4772 datarefs = VEC_alloc (data_reference_p, heap, nb_data_refs);
4773 compute_data_dependences_for_loop (loop,
4776 &dependence_relations);
4778 if (!known_dependences_p (dependence_relations))
4781 stmts_from_loop (loop, &stmts);
4782 rdg = new_graph (VEC_length (tree, stmts));
4784 rdg->indices = htab_create (nb_data_refs, hash_stmt_vertex_info,
4785 eq_stmt_vertex_info, hash_stmt_vertex_del);
4786 create_rdg_vertices (rdg, stmts);
4787 create_rdg_edges (rdg, dependence_relations);
4790 free_dependence_relations (dependence_relations);
4791 free_data_refs (datarefs);
4792 VEC_free (tree, heap, stmts);
4797 /* Free the reduced dependence graph RDG. */
4800 free_rdg (struct graph *rdg)
4804 for (i = 0; i < rdg->n_vertices; i++)
4805 free (rdg->vertices[i].data);
4807 htab_delete (rdg->indices);
4811 /* Initialize STMTS with all the statements of LOOP that contain a
4815 stores_from_loop (struct loop *loop, VEC (tree, heap) **stmts)
4818 basic_block *bbs = get_loop_body_in_dom_order (loop);
4820 for (i = 0; i < loop->num_nodes; i++)
4822 basic_block bb = bbs[i];
4823 block_stmt_iterator bsi;
4825 for (bsi = bsi_start (bb); !bsi_end_p (bsi); bsi_next (&bsi))
4826 if (!ZERO_SSA_OPERANDS (bsi_stmt (bsi), SSA_OP_VDEF))
4827 VEC_safe_push (tree, heap, *stmts, bsi_stmt (bsi));
4833 /* For a data reference REF, return the declaration of its base
4834 address or NULL_TREE if the base is not determined. */
4837 ref_base_address (tree stmt, data_ref_loc *ref)
4839 tree base = NULL_TREE;
4841 struct data_reference *dr = XCNEW (struct data_reference);
4843 DR_STMT (dr) = stmt;
4844 DR_REF (dr) = *ref->pos;
4845 dr_analyze_innermost (dr);
4846 base_address = DR_BASE_ADDRESS (dr);
4851 switch (TREE_CODE (base_address))
4854 base = TREE_OPERAND (base_address, 0);
4858 base = base_address;
4867 /* Determines whether the statement from vertex V of the RDG has a
4868 definition used outside the loop that contains this statement. */
4871 rdg_defs_used_in_other_loops_p (struct graph *rdg, int v)
4873 tree stmt = RDG_STMT (rdg, v);
4874 struct loop *loop = loop_containing_stmt (stmt);
4875 use_operand_p imm_use_p;
4876 imm_use_iterator iterator;
4878 def_operand_p def_p;
4883 FOR_EACH_PHI_OR_STMT_DEF (def_p, stmt, it, SSA_OP_DEF)
4885 FOR_EACH_IMM_USE_FAST (imm_use_p, iterator, DEF_FROM_PTR (def_p))
4887 if (loop_containing_stmt (USE_STMT (imm_use_p)) != loop)
4895 /* Determines whether statements S1 and S2 access to similar memory
4896 locations. Two memory accesses are considered similar when they
4897 have the same base address declaration, i.e. when their
4898 ref_base_address is the same. */
4901 have_similar_memory_accesses (tree s1, tree s2)
4905 VEC (data_ref_loc, heap) *refs1, *refs2;
4906 data_ref_loc *ref1, *ref2;
4908 get_references_in_stmt (s1, &refs1);
4909 get_references_in_stmt (s2, &refs2);
4911 for (i = 0; VEC_iterate (data_ref_loc, refs1, i, ref1); i++)
4913 tree base1 = ref_base_address (s1, ref1);
4916 for (j = 0; VEC_iterate (data_ref_loc, refs2, j, ref2); j++)
4917 if (base1 == ref_base_address (s2, ref2))
4925 VEC_free (data_ref_loc, heap, refs1);
4926 VEC_free (data_ref_loc, heap, refs2);
4930 /* Helper function for the hashtab. */
4933 have_similar_memory_accesses_1 (const void *s1, const void *s2)
4935 return have_similar_memory_accesses ((tree) s1, (tree) s2);
4938 /* Helper function for the hashtab. */
4941 ref_base_address_1 (const void *s)
4943 tree stmt = (tree) s;
4945 VEC (data_ref_loc, heap) *refs;
4949 get_references_in_stmt (stmt, &refs);
4951 for (i = 0; VEC_iterate (data_ref_loc, refs, i, ref); i++)
4954 res = htab_hash_pointer (ref_base_address (stmt, ref));
4958 VEC_free (data_ref_loc, heap, refs);
4962 /* Try to remove duplicated write data references from STMTS. */
4965 remove_similar_memory_refs (VEC (tree, heap) **stmts)
4969 htab_t seen = htab_create (VEC_length (tree, *stmts), ref_base_address_1,
4970 have_similar_memory_accesses_1, NULL);
4972 for (i = 0; VEC_iterate (tree, *stmts, i, stmt); )
4976 slot = htab_find_slot (seen, stmt, INSERT);
4979 VEC_ordered_remove (tree, *stmts, i);
4982 *slot = (void *) stmt;