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 = instantiate_scev (nest, loop, 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 static void compute_self_dependence (struct data_dependence_relation *);
1270 /* Initialize a data dependence relation between data accesses A and
1271 B. NB_LOOPS is the number of loops surrounding the references: the
1272 size of the classic distance/direction vectors. */
1274 static struct data_dependence_relation *
1275 initialize_data_dependence_relation (struct data_reference *a,
1276 struct data_reference *b,
1277 VEC (loop_p, heap) *loop_nest)
1279 struct data_dependence_relation *res;
1282 res = XNEW (struct data_dependence_relation);
1285 DDR_LOOP_NEST (res) = NULL;
1286 DDR_REVERSED_P (res) = false;
1287 DDR_SUBSCRIPTS (res) = NULL;
1288 DDR_DIR_VECTS (res) = NULL;
1289 DDR_DIST_VECTS (res) = NULL;
1291 if (a == NULL || b == NULL)
1293 DDR_ARE_DEPENDENT (res) = chrec_dont_know;
1297 /* If the data references do not alias, then they are independent. */
1298 if (!dr_may_alias_p (a, b))
1300 DDR_ARE_DEPENDENT (res) = chrec_known;
1304 /* When the references are exactly the same, don't spend time doing
1305 the data dependence tests, just initialize the ddr and return. */
1306 if (operand_equal_p (DR_REF (a), DR_REF (b), 0))
1308 DDR_AFFINE_P (res) = true;
1309 DDR_ARE_DEPENDENT (res) = NULL_TREE;
1310 DDR_SUBSCRIPTS (res) = VEC_alloc (subscript_p, heap, DR_NUM_DIMENSIONS (a));
1311 DDR_LOOP_NEST (res) = loop_nest;
1312 DDR_INNER_LOOP (res) = 0;
1313 DDR_SELF_REFERENCE (res) = true;
1314 compute_self_dependence (res);
1318 /* If the references do not access the same object, we do not know
1319 whether they alias or not. */
1320 if (!operand_equal_p (DR_BASE_OBJECT (a), DR_BASE_OBJECT (b), 0))
1322 DDR_ARE_DEPENDENT (res) = chrec_dont_know;
1326 /* If the base of the object is not invariant in the loop nest, we cannot
1327 analyze it. TODO -- in fact, it would suffice to record that there may
1328 be arbitrary dependences in the loops where the base object varies. */
1329 if (!object_address_invariant_in_loop_p (VEC_index (loop_p, loop_nest, 0),
1330 DR_BASE_OBJECT (a)))
1332 DDR_ARE_DEPENDENT (res) = chrec_dont_know;
1336 gcc_assert (DR_NUM_DIMENSIONS (a) == DR_NUM_DIMENSIONS (b));
1338 DDR_AFFINE_P (res) = true;
1339 DDR_ARE_DEPENDENT (res) = NULL_TREE;
1340 DDR_SUBSCRIPTS (res) = VEC_alloc (subscript_p, heap, DR_NUM_DIMENSIONS (a));
1341 DDR_LOOP_NEST (res) = loop_nest;
1342 DDR_INNER_LOOP (res) = 0;
1343 DDR_SELF_REFERENCE (res) = false;
1345 for (i = 0; i < DR_NUM_DIMENSIONS (a); i++)
1347 struct subscript *subscript;
1349 subscript = XNEW (struct subscript);
1350 SUB_CONFLICTS_IN_A (subscript) = conflict_fn_not_known ();
1351 SUB_CONFLICTS_IN_B (subscript) = conflict_fn_not_known ();
1352 SUB_LAST_CONFLICT (subscript) = chrec_dont_know;
1353 SUB_DISTANCE (subscript) = chrec_dont_know;
1354 VEC_safe_push (subscript_p, heap, DDR_SUBSCRIPTS (res), subscript);
1360 /* Frees memory used by the conflict function F. */
1363 free_conflict_function (conflict_function *f)
1367 if (CF_NONTRIVIAL_P (f))
1369 for (i = 0; i < f->n; i++)
1370 affine_fn_free (f->fns[i]);
1375 /* Frees memory used by SUBSCRIPTS. */
1378 free_subscripts (VEC (subscript_p, heap) *subscripts)
1383 for (i = 0; VEC_iterate (subscript_p, subscripts, i, s); i++)
1385 free_conflict_function (s->conflicting_iterations_in_a);
1386 free_conflict_function (s->conflicting_iterations_in_b);
1388 VEC_free (subscript_p, heap, subscripts);
1391 /* Set DDR_ARE_DEPENDENT to CHREC and finalize the subscript overlap
1395 finalize_ddr_dependent (struct data_dependence_relation *ddr,
1398 if (dump_file && (dump_flags & TDF_DETAILS))
1400 fprintf (dump_file, "(dependence classified: ");
1401 print_generic_expr (dump_file, chrec, 0);
1402 fprintf (dump_file, ")\n");
1405 DDR_ARE_DEPENDENT (ddr) = chrec;
1406 free_subscripts (DDR_SUBSCRIPTS (ddr));
1407 DDR_SUBSCRIPTS (ddr) = NULL;
1410 /* The dependence relation DDR cannot be represented by a distance
1414 non_affine_dependence_relation (struct data_dependence_relation *ddr)
1416 if (dump_file && (dump_flags & TDF_DETAILS))
1417 fprintf (dump_file, "(Dependence relation cannot be represented by distance vector.) \n");
1419 DDR_AFFINE_P (ddr) = false;
1424 /* This section contains the classic Banerjee tests. */
1426 /* Returns true iff CHREC_A and CHREC_B are not dependent on any index
1427 variables, i.e., if the ZIV (Zero Index Variable) test is true. */
1430 ziv_subscript_p (const_tree chrec_a, const_tree chrec_b)
1432 return (evolution_function_is_constant_p (chrec_a)
1433 && evolution_function_is_constant_p (chrec_b));
1436 /* Returns true iff CHREC_A and CHREC_B are dependent on an index
1437 variable, i.e., if the SIV (Single Index Variable) test is true. */
1440 siv_subscript_p (const_tree chrec_a, const_tree chrec_b)
1442 if ((evolution_function_is_constant_p (chrec_a)
1443 && evolution_function_is_univariate_p (chrec_b))
1444 || (evolution_function_is_constant_p (chrec_b)
1445 && evolution_function_is_univariate_p (chrec_a)))
1448 if (evolution_function_is_univariate_p (chrec_a)
1449 && evolution_function_is_univariate_p (chrec_b))
1451 switch (TREE_CODE (chrec_a))
1453 case POLYNOMIAL_CHREC:
1454 switch (TREE_CODE (chrec_b))
1456 case POLYNOMIAL_CHREC:
1457 if (CHREC_VARIABLE (chrec_a) != CHREC_VARIABLE (chrec_b))
1472 /* Creates a conflict function with N dimensions. The affine functions
1473 in each dimension follow. */
1475 static conflict_function *
1476 conflict_fn (unsigned n, ...)
1479 conflict_function *ret = XCNEW (conflict_function);
1482 gcc_assert (0 < n && n <= MAX_DIM);
1486 for (i = 0; i < n; i++)
1487 ret->fns[i] = va_arg (ap, affine_fn);
1493 /* Returns constant affine function with value CST. */
1496 affine_fn_cst (tree cst)
1498 affine_fn fn = VEC_alloc (tree, heap, 1);
1499 VEC_quick_push (tree, fn, cst);
1503 /* Returns affine function with single variable, CST + COEF * x_DIM. */
1506 affine_fn_univar (tree cst, unsigned dim, tree coef)
1508 affine_fn fn = VEC_alloc (tree, heap, dim + 1);
1511 gcc_assert (dim > 0);
1512 VEC_quick_push (tree, fn, cst);
1513 for (i = 1; i < dim; i++)
1514 VEC_quick_push (tree, fn, integer_zero_node);
1515 VEC_quick_push (tree, fn, coef);
1519 /* Analyze a ZIV (Zero Index Variable) subscript. *OVERLAPS_A and
1520 *OVERLAPS_B are initialized to the functions that describe the
1521 relation between the elements accessed twice by CHREC_A and
1522 CHREC_B. For k >= 0, the following property is verified:
1524 CHREC_A (*OVERLAPS_A (k)) = CHREC_B (*OVERLAPS_B (k)). */
1527 analyze_ziv_subscript (tree chrec_a,
1529 conflict_function **overlaps_a,
1530 conflict_function **overlaps_b,
1531 tree *last_conflicts)
1533 tree type, difference;
1534 dependence_stats.num_ziv++;
1536 if (dump_file && (dump_flags & TDF_DETAILS))
1537 fprintf (dump_file, "(analyze_ziv_subscript \n");
1539 type = signed_type_for_types (TREE_TYPE (chrec_a), TREE_TYPE (chrec_b));
1540 chrec_a = chrec_convert (type, chrec_a, NULL_TREE);
1541 chrec_b = chrec_convert (type, chrec_b, NULL_TREE);
1542 difference = chrec_fold_minus (type, chrec_a, chrec_b);
1544 switch (TREE_CODE (difference))
1547 if (integer_zerop (difference))
1549 /* The difference is equal to zero: the accessed index
1550 overlaps for each iteration in the loop. */
1551 *overlaps_a = conflict_fn (1, affine_fn_cst (integer_zero_node));
1552 *overlaps_b = conflict_fn (1, affine_fn_cst (integer_zero_node));
1553 *last_conflicts = chrec_dont_know;
1554 dependence_stats.num_ziv_dependent++;
1558 /* The accesses do not overlap. */
1559 *overlaps_a = conflict_fn_no_dependence ();
1560 *overlaps_b = conflict_fn_no_dependence ();
1561 *last_conflicts = integer_zero_node;
1562 dependence_stats.num_ziv_independent++;
1567 /* We're not sure whether the indexes overlap. For the moment,
1568 conservatively answer "don't know". */
1569 if (dump_file && (dump_flags & TDF_DETAILS))
1570 fprintf (dump_file, "ziv test failed: difference is non-integer.\n");
1572 *overlaps_a = conflict_fn_not_known ();
1573 *overlaps_b = conflict_fn_not_known ();
1574 *last_conflicts = chrec_dont_know;
1575 dependence_stats.num_ziv_unimplemented++;
1579 if (dump_file && (dump_flags & TDF_DETAILS))
1580 fprintf (dump_file, ")\n");
1583 /* Sets NIT to the estimated number of executions of the statements in
1584 LOOP. If CONSERVATIVE is true, we must be sure that NIT is at least as
1585 large as the number of iterations. If we have no reliable estimate,
1586 the function returns false, otherwise returns true. */
1589 estimated_loop_iterations (struct loop *loop, bool conservative,
1592 estimate_numbers_of_iterations_loop (loop);
1595 if (!loop->any_upper_bound)
1598 *nit = loop->nb_iterations_upper_bound;
1602 if (!loop->any_estimate)
1605 *nit = loop->nb_iterations_estimate;
1611 /* Similar to estimated_loop_iterations, but returns the estimate only
1612 if it fits to HOST_WIDE_INT. If this is not the case, or the estimate
1613 on the number of iterations of LOOP could not be derived, returns -1. */
1616 estimated_loop_iterations_int (struct loop *loop, bool conservative)
1619 HOST_WIDE_INT hwi_nit;
1621 if (!estimated_loop_iterations (loop, conservative, &nit))
1624 if (!double_int_fits_in_shwi_p (nit))
1626 hwi_nit = double_int_to_shwi (nit);
1628 return hwi_nit < 0 ? -1 : hwi_nit;
1631 /* Similar to estimated_loop_iterations, but returns the estimate as a tree,
1632 and only if it fits to the int type. If this is not the case, or the
1633 estimate on the number of iterations of LOOP could not be derived, returns
1637 estimated_loop_iterations_tree (struct loop *loop, bool conservative)
1642 if (!estimated_loop_iterations (loop, conservative, &nit))
1643 return chrec_dont_know;
1645 type = lang_hooks.types.type_for_size (INT_TYPE_SIZE, true);
1646 if (!double_int_fits_to_tree_p (type, nit))
1647 return chrec_dont_know;
1649 return double_int_to_tree (type, nit);
1652 /* Analyze a SIV (Single Index Variable) subscript where CHREC_A is a
1653 constant, and CHREC_B is an affine function. *OVERLAPS_A and
1654 *OVERLAPS_B are initialized to the functions that describe the
1655 relation between the elements accessed twice by CHREC_A and
1656 CHREC_B. For k >= 0, the following property is verified:
1658 CHREC_A (*OVERLAPS_A (k)) = CHREC_B (*OVERLAPS_B (k)). */
1661 analyze_siv_subscript_cst_affine (tree chrec_a,
1663 conflict_function **overlaps_a,
1664 conflict_function **overlaps_b,
1665 tree *last_conflicts)
1667 bool value0, value1, value2;
1668 tree type, difference, tmp;
1670 type = signed_type_for_types (TREE_TYPE (chrec_a), TREE_TYPE (chrec_b));
1671 chrec_a = chrec_convert (type, chrec_a, NULL_TREE);
1672 chrec_b = chrec_convert (type, chrec_b, NULL_TREE);
1673 difference = chrec_fold_minus (type, initial_condition (chrec_b), chrec_a);
1675 if (!chrec_is_positive (initial_condition (difference), &value0))
1677 if (dump_file && (dump_flags & TDF_DETAILS))
1678 fprintf (dump_file, "siv test failed: chrec is not positive.\n");
1680 dependence_stats.num_siv_unimplemented++;
1681 *overlaps_a = conflict_fn_not_known ();
1682 *overlaps_b = conflict_fn_not_known ();
1683 *last_conflicts = chrec_dont_know;
1688 if (value0 == false)
1690 if (!chrec_is_positive (CHREC_RIGHT (chrec_b), &value1))
1692 if (dump_file && (dump_flags & TDF_DETAILS))
1693 fprintf (dump_file, "siv test failed: chrec not positive.\n");
1695 *overlaps_a = conflict_fn_not_known ();
1696 *overlaps_b = conflict_fn_not_known ();
1697 *last_conflicts = chrec_dont_know;
1698 dependence_stats.num_siv_unimplemented++;
1707 chrec_b = {10, +, 1}
1710 if (tree_fold_divides_p (CHREC_RIGHT (chrec_b), difference))
1712 HOST_WIDE_INT numiter;
1713 struct loop *loop = get_chrec_loop (chrec_b);
1715 *overlaps_a = conflict_fn (1, affine_fn_cst (integer_zero_node));
1716 tmp = fold_build2 (EXACT_DIV_EXPR, type,
1717 fold_build1 (ABS_EXPR, type, difference),
1718 CHREC_RIGHT (chrec_b));
1719 *overlaps_b = conflict_fn (1, affine_fn_cst (tmp));
1720 *last_conflicts = integer_one_node;
1723 /* Perform weak-zero siv test to see if overlap is
1724 outside the loop bounds. */
1725 numiter = estimated_loop_iterations_int (loop, false);
1728 && compare_tree_int (tmp, numiter) > 0)
1730 free_conflict_function (*overlaps_a);
1731 free_conflict_function (*overlaps_b);
1732 *overlaps_a = conflict_fn_no_dependence ();
1733 *overlaps_b = conflict_fn_no_dependence ();
1734 *last_conflicts = integer_zero_node;
1735 dependence_stats.num_siv_independent++;
1738 dependence_stats.num_siv_dependent++;
1742 /* When the step does not divide the difference, there are
1746 *overlaps_a = conflict_fn_no_dependence ();
1747 *overlaps_b = conflict_fn_no_dependence ();
1748 *last_conflicts = integer_zero_node;
1749 dependence_stats.num_siv_independent++;
1758 chrec_b = {10, +, -1}
1760 In this case, chrec_a will not overlap with chrec_b. */
1761 *overlaps_a = conflict_fn_no_dependence ();
1762 *overlaps_b = conflict_fn_no_dependence ();
1763 *last_conflicts = integer_zero_node;
1764 dependence_stats.num_siv_independent++;
1771 if (!chrec_is_positive (CHREC_RIGHT (chrec_b), &value2))
1773 if (dump_file && (dump_flags & TDF_DETAILS))
1774 fprintf (dump_file, "siv test failed: chrec not positive.\n");
1776 *overlaps_a = conflict_fn_not_known ();
1777 *overlaps_b = conflict_fn_not_known ();
1778 *last_conflicts = chrec_dont_know;
1779 dependence_stats.num_siv_unimplemented++;
1784 if (value2 == false)
1788 chrec_b = {10, +, -1}
1790 if (tree_fold_divides_p (CHREC_RIGHT (chrec_b), difference))
1792 HOST_WIDE_INT numiter;
1793 struct loop *loop = get_chrec_loop (chrec_b);
1795 *overlaps_a = conflict_fn (1, affine_fn_cst (integer_zero_node));
1796 tmp = fold_build2 (EXACT_DIV_EXPR, type, difference,
1797 CHREC_RIGHT (chrec_b));
1798 *overlaps_b = conflict_fn (1, affine_fn_cst (tmp));
1799 *last_conflicts = integer_one_node;
1801 /* Perform weak-zero siv test to see if overlap is
1802 outside the loop bounds. */
1803 numiter = estimated_loop_iterations_int (loop, false);
1806 && compare_tree_int (tmp, numiter) > 0)
1808 free_conflict_function (*overlaps_a);
1809 free_conflict_function (*overlaps_b);
1810 *overlaps_a = conflict_fn_no_dependence ();
1811 *overlaps_b = conflict_fn_no_dependence ();
1812 *last_conflicts = integer_zero_node;
1813 dependence_stats.num_siv_independent++;
1816 dependence_stats.num_siv_dependent++;
1820 /* When the step does not divide the difference, there
1824 *overlaps_a = conflict_fn_no_dependence ();
1825 *overlaps_b = conflict_fn_no_dependence ();
1826 *last_conflicts = integer_zero_node;
1827 dependence_stats.num_siv_independent++;
1837 In this case, chrec_a will not overlap with chrec_b. */
1838 *overlaps_a = conflict_fn_no_dependence ();
1839 *overlaps_b = conflict_fn_no_dependence ();
1840 *last_conflicts = integer_zero_node;
1841 dependence_stats.num_siv_independent++;
1849 /* Helper recursive function for initializing the matrix A. Returns
1850 the initial value of CHREC. */
1852 static HOST_WIDE_INT
1853 initialize_matrix_A (lambda_matrix A, tree chrec, unsigned index, int mult)
1857 if (TREE_CODE (chrec) != POLYNOMIAL_CHREC)
1858 return int_cst_value (chrec);
1860 A[index][0] = mult * int_cst_value (CHREC_RIGHT (chrec));
1861 return initialize_matrix_A (A, CHREC_LEFT (chrec), index + 1, mult);
1864 #define FLOOR_DIV(x,y) ((x) / (y))
1866 /* Solves the special case of the Diophantine equation:
1867 | {0, +, STEP_A}_x (OVERLAPS_A) = {0, +, STEP_B}_y (OVERLAPS_B)
1869 Computes the descriptions OVERLAPS_A and OVERLAPS_B. NITER is the
1870 number of iterations that loops X and Y run. The overlaps will be
1871 constructed as evolutions in dimension DIM. */
1874 compute_overlap_steps_for_affine_univar (int niter, int step_a, int step_b,
1875 affine_fn *overlaps_a,
1876 affine_fn *overlaps_b,
1877 tree *last_conflicts, int dim)
1879 if (((step_a > 0 && step_b > 0)
1880 || (step_a < 0 && step_b < 0)))
1882 int step_overlaps_a, step_overlaps_b;
1883 int gcd_steps_a_b, last_conflict, tau2;
1885 gcd_steps_a_b = gcd (step_a, step_b);
1886 step_overlaps_a = step_b / gcd_steps_a_b;
1887 step_overlaps_b = step_a / gcd_steps_a_b;
1891 tau2 = FLOOR_DIV (niter, step_overlaps_a);
1892 tau2 = MIN (tau2, FLOOR_DIV (niter, step_overlaps_b));
1893 last_conflict = tau2;
1894 *last_conflicts = build_int_cst (NULL_TREE, last_conflict);
1897 *last_conflicts = chrec_dont_know;
1899 *overlaps_a = affine_fn_univar (integer_zero_node, dim,
1900 build_int_cst (NULL_TREE,
1902 *overlaps_b = affine_fn_univar (integer_zero_node, dim,
1903 build_int_cst (NULL_TREE,
1909 *overlaps_a = affine_fn_cst (integer_zero_node);
1910 *overlaps_b = affine_fn_cst (integer_zero_node);
1911 *last_conflicts = integer_zero_node;
1915 /* Solves the special case of a Diophantine equation where CHREC_A is
1916 an affine bivariate function, and CHREC_B is an affine univariate
1917 function. For example,
1919 | {{0, +, 1}_x, +, 1335}_y = {0, +, 1336}_z
1921 has the following overlapping functions:
1923 | x (t, u, v) = {{0, +, 1336}_t, +, 1}_v
1924 | y (t, u, v) = {{0, +, 1336}_u, +, 1}_v
1925 | z (t, u, v) = {{{0, +, 1}_t, +, 1335}_u, +, 1}_v
1927 FORNOW: This is a specialized implementation for a case occurring in
1928 a common benchmark. Implement the general algorithm. */
1931 compute_overlap_steps_for_affine_1_2 (tree chrec_a, tree chrec_b,
1932 conflict_function **overlaps_a,
1933 conflict_function **overlaps_b,
1934 tree *last_conflicts)
1936 bool xz_p, yz_p, xyz_p;
1937 int step_x, step_y, step_z;
1938 HOST_WIDE_INT niter_x, niter_y, niter_z, niter;
1939 affine_fn overlaps_a_xz, overlaps_b_xz;
1940 affine_fn overlaps_a_yz, overlaps_b_yz;
1941 affine_fn overlaps_a_xyz, overlaps_b_xyz;
1942 affine_fn ova1, ova2, ovb;
1943 tree last_conflicts_xz, last_conflicts_yz, last_conflicts_xyz;
1945 step_x = int_cst_value (CHREC_RIGHT (CHREC_LEFT (chrec_a)));
1946 step_y = int_cst_value (CHREC_RIGHT (chrec_a));
1947 step_z = int_cst_value (CHREC_RIGHT (chrec_b));
1950 estimated_loop_iterations_int (get_chrec_loop (CHREC_LEFT (chrec_a)),
1952 niter_y = estimated_loop_iterations_int (get_chrec_loop (chrec_a), false);
1953 niter_z = estimated_loop_iterations_int (get_chrec_loop (chrec_b), false);
1955 if (niter_x < 0 || niter_y < 0 || niter_z < 0)
1957 if (dump_file && (dump_flags & TDF_DETAILS))
1958 fprintf (dump_file, "overlap steps test failed: no iteration counts.\n");
1960 *overlaps_a = conflict_fn_not_known ();
1961 *overlaps_b = conflict_fn_not_known ();
1962 *last_conflicts = chrec_dont_know;
1966 niter = MIN (niter_x, niter_z);
1967 compute_overlap_steps_for_affine_univar (niter, step_x, step_z,
1970 &last_conflicts_xz, 1);
1971 niter = MIN (niter_y, niter_z);
1972 compute_overlap_steps_for_affine_univar (niter, step_y, step_z,
1975 &last_conflicts_yz, 2);
1976 niter = MIN (niter_x, niter_z);
1977 niter = MIN (niter_y, niter);
1978 compute_overlap_steps_for_affine_univar (niter, step_x + step_y, step_z,
1981 &last_conflicts_xyz, 3);
1983 xz_p = !integer_zerop (last_conflicts_xz);
1984 yz_p = !integer_zerop (last_conflicts_yz);
1985 xyz_p = !integer_zerop (last_conflicts_xyz);
1987 if (xz_p || yz_p || xyz_p)
1989 ova1 = affine_fn_cst (integer_zero_node);
1990 ova2 = affine_fn_cst (integer_zero_node);
1991 ovb = affine_fn_cst (integer_zero_node);
1994 affine_fn t0 = ova1;
1997 ova1 = affine_fn_plus (ova1, overlaps_a_xz);
1998 ovb = affine_fn_plus (ovb, overlaps_b_xz);
1999 affine_fn_free (t0);
2000 affine_fn_free (t2);
2001 *last_conflicts = last_conflicts_xz;
2005 affine_fn t0 = ova2;
2008 ova2 = affine_fn_plus (ova2, overlaps_a_yz);
2009 ovb = affine_fn_plus (ovb, overlaps_b_yz);
2010 affine_fn_free (t0);
2011 affine_fn_free (t2);
2012 *last_conflicts = last_conflicts_yz;
2016 affine_fn t0 = ova1;
2017 affine_fn t2 = ova2;
2020 ova1 = affine_fn_plus (ova1, overlaps_a_xyz);
2021 ova2 = affine_fn_plus (ova2, overlaps_a_xyz);
2022 ovb = affine_fn_plus (ovb, overlaps_b_xyz);
2023 affine_fn_free (t0);
2024 affine_fn_free (t2);
2025 affine_fn_free (t4);
2026 *last_conflicts = last_conflicts_xyz;
2028 *overlaps_a = conflict_fn (2, ova1, ova2);
2029 *overlaps_b = conflict_fn (1, ovb);
2033 *overlaps_a = conflict_fn (1, affine_fn_cst (integer_zero_node));
2034 *overlaps_b = conflict_fn (1, affine_fn_cst (integer_zero_node));
2035 *last_conflicts = integer_zero_node;
2038 affine_fn_free (overlaps_a_xz);
2039 affine_fn_free (overlaps_b_xz);
2040 affine_fn_free (overlaps_a_yz);
2041 affine_fn_free (overlaps_b_yz);
2042 affine_fn_free (overlaps_a_xyz);
2043 affine_fn_free (overlaps_b_xyz);
2046 /* Determines the overlapping elements due to accesses CHREC_A and
2047 CHREC_B, that are affine functions. This function cannot handle
2048 symbolic evolution functions, ie. when initial conditions are
2049 parameters, because it uses lambda matrices of integers. */
2052 analyze_subscript_affine_affine (tree chrec_a,
2054 conflict_function **overlaps_a,
2055 conflict_function **overlaps_b,
2056 tree *last_conflicts)
2058 unsigned nb_vars_a, nb_vars_b, dim;
2059 HOST_WIDE_INT init_a, init_b, gamma, gcd_alpha_beta;
2060 lambda_matrix A, U, S;
2062 if (eq_evolutions_p (chrec_a, chrec_b))
2064 /* The accessed index overlaps for each iteration in the
2066 *overlaps_a = conflict_fn (1, affine_fn_cst (integer_zero_node));
2067 *overlaps_b = conflict_fn (1, affine_fn_cst (integer_zero_node));
2068 *last_conflicts = chrec_dont_know;
2071 if (dump_file && (dump_flags & TDF_DETAILS))
2072 fprintf (dump_file, "(analyze_subscript_affine_affine \n");
2074 /* For determining the initial intersection, we have to solve a
2075 Diophantine equation. This is the most time consuming part.
2077 For answering to the question: "Is there a dependence?" we have
2078 to prove that there exists a solution to the Diophantine
2079 equation, and that the solution is in the iteration domain,
2080 i.e. the solution is positive or zero, and that the solution
2081 happens before the upper bound loop.nb_iterations. Otherwise
2082 there is no dependence. This function outputs a description of
2083 the iterations that hold the intersections. */
2085 nb_vars_a = nb_vars_in_chrec (chrec_a);
2086 nb_vars_b = nb_vars_in_chrec (chrec_b);
2088 dim = nb_vars_a + nb_vars_b;
2089 U = lambda_matrix_new (dim, dim);
2090 A = lambda_matrix_new (dim, 1);
2091 S = lambda_matrix_new (dim, 1);
2093 init_a = initialize_matrix_A (A, chrec_a, 0, 1);
2094 init_b = initialize_matrix_A (A, chrec_b, nb_vars_a, -1);
2095 gamma = init_b - init_a;
2097 /* Don't do all the hard work of solving the Diophantine equation
2098 when we already know the solution: for example,
2101 | gamma = 3 - 3 = 0.
2102 Then the first overlap occurs during the first iterations:
2103 | {3, +, 1}_1 ({0, +, 4}_x) = {3, +, 4}_2 ({0, +, 1}_x)
2107 if (nb_vars_a == 1 && nb_vars_b == 1)
2109 HOST_WIDE_INT step_a, step_b;
2110 HOST_WIDE_INT niter, niter_a, niter_b;
2113 niter_a = estimated_loop_iterations_int (get_chrec_loop (chrec_a),
2115 niter_b = estimated_loop_iterations_int (get_chrec_loop (chrec_b),
2117 niter = MIN (niter_a, niter_b);
2118 step_a = int_cst_value (CHREC_RIGHT (chrec_a));
2119 step_b = int_cst_value (CHREC_RIGHT (chrec_b));
2121 compute_overlap_steps_for_affine_univar (niter, step_a, step_b,
2124 *overlaps_a = conflict_fn (1, ova);
2125 *overlaps_b = conflict_fn (1, ovb);
2128 else if (nb_vars_a == 2 && nb_vars_b == 1)
2129 compute_overlap_steps_for_affine_1_2
2130 (chrec_a, chrec_b, overlaps_a, overlaps_b, last_conflicts);
2132 else if (nb_vars_a == 1 && nb_vars_b == 2)
2133 compute_overlap_steps_for_affine_1_2
2134 (chrec_b, chrec_a, overlaps_b, overlaps_a, last_conflicts);
2138 if (dump_file && (dump_flags & TDF_DETAILS))
2139 fprintf (dump_file, "affine-affine test failed: too many variables.\n");
2140 *overlaps_a = conflict_fn_not_known ();
2141 *overlaps_b = conflict_fn_not_known ();
2142 *last_conflicts = chrec_dont_know;
2144 goto end_analyze_subs_aa;
2148 lambda_matrix_right_hermite (A, dim, 1, S, U);
2153 lambda_matrix_row_negate (U, dim, 0);
2155 gcd_alpha_beta = S[0][0];
2157 /* Something went wrong: for example in {1, +, 0}_5 vs. {0, +, 0}_5,
2158 but that is a quite strange case. Instead of ICEing, answer
2160 if (gcd_alpha_beta == 0)
2162 *overlaps_a = conflict_fn_not_known ();
2163 *overlaps_b = conflict_fn_not_known ();
2164 *last_conflicts = chrec_dont_know;
2165 goto end_analyze_subs_aa;
2168 /* The classic "gcd-test". */
2169 if (!int_divides_p (gcd_alpha_beta, gamma))
2171 /* The "gcd-test" has determined that there is no integer
2172 solution, i.e. there is no dependence. */
2173 *overlaps_a = conflict_fn_no_dependence ();
2174 *overlaps_b = conflict_fn_no_dependence ();
2175 *last_conflicts = integer_zero_node;
2178 /* Both access functions are univariate. This includes SIV and MIV cases. */
2179 else if (nb_vars_a == 1 && nb_vars_b == 1)
2181 /* Both functions should have the same evolution sign. */
2182 if (((A[0][0] > 0 && -A[1][0] > 0)
2183 || (A[0][0] < 0 && -A[1][0] < 0)))
2185 /* The solutions are given by:
2187 | [GAMMA/GCD_ALPHA_BETA t].[u11 u12] = [x0]
2190 For a given integer t. Using the following variables,
2192 | i0 = u11 * gamma / gcd_alpha_beta
2193 | j0 = u12 * gamma / gcd_alpha_beta
2200 | y0 = j0 + j1 * t. */
2201 HOST_WIDE_INT i0, j0, i1, j1;
2203 i0 = U[0][0] * gamma / gcd_alpha_beta;
2204 j0 = U[0][1] * gamma / gcd_alpha_beta;
2208 if ((i1 == 0 && i0 < 0)
2209 || (j1 == 0 && j0 < 0))
2211 /* There is no solution.
2212 FIXME: The case "i0 > nb_iterations, j0 > nb_iterations"
2213 falls in here, but for the moment we don't look at the
2214 upper bound of the iteration domain. */
2215 *overlaps_a = conflict_fn_no_dependence ();
2216 *overlaps_b = conflict_fn_no_dependence ();
2217 *last_conflicts = integer_zero_node;
2218 goto end_analyze_subs_aa;
2221 if (i1 > 0 && j1 > 0)
2223 HOST_WIDE_INT niter_a = estimated_loop_iterations_int
2224 (get_chrec_loop (chrec_a), false);
2225 HOST_WIDE_INT niter_b = estimated_loop_iterations_int
2226 (get_chrec_loop (chrec_b), false);
2227 HOST_WIDE_INT niter = MIN (niter_a, niter_b);
2229 /* (X0, Y0) is a solution of the Diophantine equation:
2230 "chrec_a (X0) = chrec_b (Y0)". */
2231 HOST_WIDE_INT tau1 = MAX (CEIL (-i0, i1),
2233 HOST_WIDE_INT x0 = i1 * tau1 + i0;
2234 HOST_WIDE_INT y0 = j1 * tau1 + j0;
2236 /* (X1, Y1) is the smallest positive solution of the eq
2237 "chrec_a (X1) = chrec_b (Y1)", i.e. this is where the
2238 first conflict occurs. */
2239 HOST_WIDE_INT min_multiple = MIN (x0 / i1, y0 / j1);
2240 HOST_WIDE_INT x1 = x0 - i1 * min_multiple;
2241 HOST_WIDE_INT y1 = y0 - j1 * min_multiple;
2245 HOST_WIDE_INT tau2 = MIN (FLOOR_DIV (niter - i0, i1),
2246 FLOOR_DIV (niter - j0, j1));
2247 HOST_WIDE_INT last_conflict = tau2 - (x1 - i0)/i1;
2249 /* If the overlap occurs outside of the bounds of the
2250 loop, there is no dependence. */
2251 if (x1 > niter || y1 > niter)
2253 *overlaps_a = conflict_fn_no_dependence ();
2254 *overlaps_b = conflict_fn_no_dependence ();
2255 *last_conflicts = integer_zero_node;
2256 goto end_analyze_subs_aa;
2259 *last_conflicts = build_int_cst (NULL_TREE, last_conflict);
2262 *last_conflicts = chrec_dont_know;
2266 affine_fn_univar (build_int_cst (NULL_TREE, x1),
2268 build_int_cst (NULL_TREE, i1)));
2271 affine_fn_univar (build_int_cst (NULL_TREE, y1),
2273 build_int_cst (NULL_TREE, j1)));
2277 /* FIXME: For the moment, the upper bound of the
2278 iteration domain for i and j is not checked. */
2279 if (dump_file && (dump_flags & TDF_DETAILS))
2280 fprintf (dump_file, "affine-affine test failed: unimplemented.\n");
2281 *overlaps_a = conflict_fn_not_known ();
2282 *overlaps_b = conflict_fn_not_known ();
2283 *last_conflicts = chrec_dont_know;
2288 if (dump_file && (dump_flags & TDF_DETAILS))
2289 fprintf (dump_file, "affine-affine test failed: unimplemented.\n");
2290 *overlaps_a = conflict_fn_not_known ();
2291 *overlaps_b = conflict_fn_not_known ();
2292 *last_conflicts = chrec_dont_know;
2297 if (dump_file && (dump_flags & TDF_DETAILS))
2298 fprintf (dump_file, "affine-affine test failed: unimplemented.\n");
2299 *overlaps_a = conflict_fn_not_known ();
2300 *overlaps_b = conflict_fn_not_known ();
2301 *last_conflicts = chrec_dont_know;
2304 end_analyze_subs_aa:
2305 if (dump_file && (dump_flags & TDF_DETAILS))
2307 fprintf (dump_file, " (overlaps_a = ");
2308 dump_conflict_function (dump_file, *overlaps_a);
2309 fprintf (dump_file, ")\n (overlaps_b = ");
2310 dump_conflict_function (dump_file, *overlaps_b);
2311 fprintf (dump_file, ")\n");
2312 fprintf (dump_file, ")\n");
2316 /* Returns true when analyze_subscript_affine_affine can be used for
2317 determining the dependence relation between chrec_a and chrec_b,
2318 that contain symbols. This function modifies chrec_a and chrec_b
2319 such that the analysis result is the same, and such that they don't
2320 contain symbols, and then can safely be passed to the analyzer.
2322 Example: The analysis of the following tuples of evolutions produce
2323 the same results: {x+1, +, 1}_1 vs. {x+3, +, 1}_1, and {-2, +, 1}_1
2326 {x+1, +, 1}_1 ({2, +, 1}_1) = {x+3, +, 1}_1 ({0, +, 1}_1)
2327 {-2, +, 1}_1 ({2, +, 1}_1) = {0, +, 1}_1 ({0, +, 1}_1)
2331 can_use_analyze_subscript_affine_affine (tree *chrec_a, tree *chrec_b)
2333 tree diff, type, left_a, left_b, right_b;
2335 if (chrec_contains_symbols (CHREC_RIGHT (*chrec_a))
2336 || chrec_contains_symbols (CHREC_RIGHT (*chrec_b)))
2337 /* FIXME: For the moment not handled. Might be refined later. */
2340 type = chrec_type (*chrec_a);
2341 left_a = CHREC_LEFT (*chrec_a);
2342 left_b = chrec_convert (type, CHREC_LEFT (*chrec_b), NULL_TREE);
2343 diff = chrec_fold_minus (type, left_a, left_b);
2345 if (!evolution_function_is_constant_p (diff))
2348 if (dump_file && (dump_flags & TDF_DETAILS))
2349 fprintf (dump_file, "can_use_subscript_aff_aff_for_symbolic \n");
2351 *chrec_a = build_polynomial_chrec (CHREC_VARIABLE (*chrec_a),
2352 diff, CHREC_RIGHT (*chrec_a));
2353 right_b = chrec_convert (type, CHREC_RIGHT (*chrec_b), NULL_TREE);
2354 *chrec_b = build_polynomial_chrec (CHREC_VARIABLE (*chrec_b),
2355 build_int_cst (type, 0),
2360 /* Analyze a SIV (Single Index Variable) subscript. *OVERLAPS_A and
2361 *OVERLAPS_B are initialized to the functions that describe the
2362 relation between the elements accessed twice by CHREC_A and
2363 CHREC_B. For k >= 0, the following property is verified:
2365 CHREC_A (*OVERLAPS_A (k)) = CHREC_B (*OVERLAPS_B (k)). */
2368 analyze_siv_subscript (tree chrec_a,
2370 conflict_function **overlaps_a,
2371 conflict_function **overlaps_b,
2372 tree *last_conflicts)
2374 dependence_stats.num_siv++;
2376 if (dump_file && (dump_flags & TDF_DETAILS))
2377 fprintf (dump_file, "(analyze_siv_subscript \n");
2379 if (evolution_function_is_constant_p (chrec_a)
2380 && evolution_function_is_affine_p (chrec_b))
2381 analyze_siv_subscript_cst_affine (chrec_a, chrec_b,
2382 overlaps_a, overlaps_b, last_conflicts);
2384 else if (evolution_function_is_affine_p (chrec_a)
2385 && evolution_function_is_constant_p (chrec_b))
2386 analyze_siv_subscript_cst_affine (chrec_b, chrec_a,
2387 overlaps_b, overlaps_a, last_conflicts);
2389 else if (evolution_function_is_affine_p (chrec_a)
2390 && evolution_function_is_affine_p (chrec_b))
2392 if (!chrec_contains_symbols (chrec_a)
2393 && !chrec_contains_symbols (chrec_b))
2395 analyze_subscript_affine_affine (chrec_a, chrec_b,
2396 overlaps_a, overlaps_b,
2399 if (CF_NOT_KNOWN_P (*overlaps_a)
2400 || CF_NOT_KNOWN_P (*overlaps_b))
2401 dependence_stats.num_siv_unimplemented++;
2402 else if (CF_NO_DEPENDENCE_P (*overlaps_a)
2403 || CF_NO_DEPENDENCE_P (*overlaps_b))
2404 dependence_stats.num_siv_independent++;
2406 dependence_stats.num_siv_dependent++;
2408 else if (can_use_analyze_subscript_affine_affine (&chrec_a,
2411 analyze_subscript_affine_affine (chrec_a, chrec_b,
2412 overlaps_a, overlaps_b,
2415 if (CF_NOT_KNOWN_P (*overlaps_a)
2416 || CF_NOT_KNOWN_P (*overlaps_b))
2417 dependence_stats.num_siv_unimplemented++;
2418 else if (CF_NO_DEPENDENCE_P (*overlaps_a)
2419 || CF_NO_DEPENDENCE_P (*overlaps_b))
2420 dependence_stats.num_siv_independent++;
2422 dependence_stats.num_siv_dependent++;
2425 goto siv_subscript_dontknow;
2430 siv_subscript_dontknow:;
2431 if (dump_file && (dump_flags & TDF_DETAILS))
2432 fprintf (dump_file, "siv test failed: unimplemented.\n");
2433 *overlaps_a = conflict_fn_not_known ();
2434 *overlaps_b = conflict_fn_not_known ();
2435 *last_conflicts = chrec_dont_know;
2436 dependence_stats.num_siv_unimplemented++;
2439 if (dump_file && (dump_flags & TDF_DETAILS))
2440 fprintf (dump_file, ")\n");
2443 /* Returns false if we can prove that the greatest common divisor of the steps
2444 of CHREC does not divide CST, false otherwise. */
2447 gcd_of_steps_may_divide_p (const_tree chrec, const_tree cst)
2449 HOST_WIDE_INT cd = 0, val;
2452 if (!host_integerp (cst, 0))
2454 val = tree_low_cst (cst, 0);
2456 while (TREE_CODE (chrec) == POLYNOMIAL_CHREC)
2458 step = CHREC_RIGHT (chrec);
2459 if (!host_integerp (step, 0))
2461 cd = gcd (cd, tree_low_cst (step, 0));
2462 chrec = CHREC_LEFT (chrec);
2465 return val % cd == 0;
2468 /* Analyze a MIV (Multiple Index Variable) subscript with respect to
2469 LOOP_NEST. *OVERLAPS_A and *OVERLAPS_B are initialized to the
2470 functions that describe the relation between the elements accessed
2471 twice by CHREC_A and CHREC_B. For k >= 0, the following property
2474 CHREC_A (*OVERLAPS_A (k)) = CHREC_B (*OVERLAPS_B (k)). */
2477 analyze_miv_subscript (tree chrec_a,
2479 conflict_function **overlaps_a,
2480 conflict_function **overlaps_b,
2481 tree *last_conflicts,
2482 struct loop *loop_nest)
2484 /* FIXME: This is a MIV subscript, not yet handled.
2485 Example: (A[{1, +, 1}_1] vs. A[{1, +, 1}_2]) that comes from
2488 In the SIV test we had to solve a Diophantine equation with two
2489 variables. In the MIV case we have to solve a Diophantine
2490 equation with 2*n variables (if the subscript uses n IVs).
2492 tree type, difference;
2494 dependence_stats.num_miv++;
2495 if (dump_file && (dump_flags & TDF_DETAILS))
2496 fprintf (dump_file, "(analyze_miv_subscript \n");
2498 type = signed_type_for_types (TREE_TYPE (chrec_a), TREE_TYPE (chrec_b));
2499 chrec_a = chrec_convert (type, chrec_a, NULL_TREE);
2500 chrec_b = chrec_convert (type, chrec_b, NULL_TREE);
2501 difference = chrec_fold_minus (type, chrec_a, chrec_b);
2503 if (eq_evolutions_p (chrec_a, chrec_b))
2505 /* Access functions are the same: all the elements are accessed
2506 in the same order. */
2507 *overlaps_a = conflict_fn (1, affine_fn_cst (integer_zero_node));
2508 *overlaps_b = conflict_fn (1, affine_fn_cst (integer_zero_node));
2509 *last_conflicts = estimated_loop_iterations_tree
2510 (get_chrec_loop (chrec_a), true);
2511 dependence_stats.num_miv_dependent++;
2514 else if (evolution_function_is_constant_p (difference)
2515 /* For the moment, the following is verified:
2516 evolution_function_is_affine_multivariate_p (chrec_a,
2518 && !gcd_of_steps_may_divide_p (chrec_a, difference))
2520 /* testsuite/.../ssa-chrec-33.c
2521 {{21, +, 2}_1, +, -2}_2 vs. {{20, +, 2}_1, +, -2}_2
2523 The difference is 1, and all the evolution steps are multiples
2524 of 2, consequently there are no overlapping elements. */
2525 *overlaps_a = conflict_fn_no_dependence ();
2526 *overlaps_b = conflict_fn_no_dependence ();
2527 *last_conflicts = integer_zero_node;
2528 dependence_stats.num_miv_independent++;
2531 else if (evolution_function_is_affine_multivariate_p (chrec_a, loop_nest->num)
2532 && !chrec_contains_symbols (chrec_a)
2533 && evolution_function_is_affine_multivariate_p (chrec_b, loop_nest->num)
2534 && !chrec_contains_symbols (chrec_b))
2536 /* testsuite/.../ssa-chrec-35.c
2537 {0, +, 1}_2 vs. {0, +, 1}_3
2538 the overlapping elements are respectively located at iterations:
2539 {0, +, 1}_x and {0, +, 1}_x,
2540 in other words, we have the equality:
2541 {0, +, 1}_2 ({0, +, 1}_x) = {0, +, 1}_3 ({0, +, 1}_x)
2544 {{0, +, 1}_1, +, 2}_2 ({0, +, 1}_x, {0, +, 1}_y) =
2545 {0, +, 1}_1 ({{0, +, 1}_x, +, 2}_y)
2547 {{0, +, 2}_1, +, 3}_2 ({0, +, 1}_y, {0, +, 1}_x) =
2548 {{0, +, 3}_1, +, 2}_2 ({0, +, 1}_x, {0, +, 1}_y)
2550 analyze_subscript_affine_affine (chrec_a, chrec_b,
2551 overlaps_a, overlaps_b, last_conflicts);
2553 if (CF_NOT_KNOWN_P (*overlaps_a)
2554 || CF_NOT_KNOWN_P (*overlaps_b))
2555 dependence_stats.num_miv_unimplemented++;
2556 else if (CF_NO_DEPENDENCE_P (*overlaps_a)
2557 || CF_NO_DEPENDENCE_P (*overlaps_b))
2558 dependence_stats.num_miv_independent++;
2560 dependence_stats.num_miv_dependent++;
2565 /* When the analysis is too difficult, answer "don't know". */
2566 if (dump_file && (dump_flags & TDF_DETAILS))
2567 fprintf (dump_file, "analyze_miv_subscript test failed: unimplemented.\n");
2569 *overlaps_a = conflict_fn_not_known ();
2570 *overlaps_b = conflict_fn_not_known ();
2571 *last_conflicts = chrec_dont_know;
2572 dependence_stats.num_miv_unimplemented++;
2575 if (dump_file && (dump_flags & TDF_DETAILS))
2576 fprintf (dump_file, ")\n");
2579 /* Determines the iterations for which CHREC_A is equal to CHREC_B in
2580 with respect to LOOP_NEST. OVERLAP_ITERATIONS_A and
2581 OVERLAP_ITERATIONS_B are initialized with two functions that
2582 describe the iterations that contain conflicting elements.
2584 Remark: For an integer k >= 0, the following equality is true:
2586 CHREC_A (OVERLAP_ITERATIONS_A (k)) == CHREC_B (OVERLAP_ITERATIONS_B (k)).
2590 analyze_overlapping_iterations (tree chrec_a,
2592 conflict_function **overlap_iterations_a,
2593 conflict_function **overlap_iterations_b,
2594 tree *last_conflicts, struct loop *loop_nest)
2596 unsigned int lnn = loop_nest->num;
2598 dependence_stats.num_subscript_tests++;
2600 if (dump_file && (dump_flags & TDF_DETAILS))
2602 fprintf (dump_file, "(analyze_overlapping_iterations \n");
2603 fprintf (dump_file, " (chrec_a = ");
2604 print_generic_expr (dump_file, chrec_a, 0);
2605 fprintf (dump_file, ")\n (chrec_b = ");
2606 print_generic_expr (dump_file, chrec_b, 0);
2607 fprintf (dump_file, ")\n");
2610 if (chrec_a == NULL_TREE
2611 || chrec_b == NULL_TREE
2612 || chrec_contains_undetermined (chrec_a)
2613 || chrec_contains_undetermined (chrec_b))
2615 dependence_stats.num_subscript_undetermined++;
2617 *overlap_iterations_a = conflict_fn_not_known ();
2618 *overlap_iterations_b = conflict_fn_not_known ();
2621 /* If they are the same chrec, and are affine, they overlap
2622 on every iteration. */
2623 else if (eq_evolutions_p (chrec_a, chrec_b)
2624 && evolution_function_is_affine_multivariate_p (chrec_a, lnn))
2626 dependence_stats.num_same_subscript_function++;
2627 *overlap_iterations_a = conflict_fn (1, affine_fn_cst (integer_zero_node));
2628 *overlap_iterations_b = conflict_fn (1, affine_fn_cst (integer_zero_node));
2629 *last_conflicts = chrec_dont_know;
2632 /* If they aren't the same, and aren't affine, we can't do anything
2634 else if ((chrec_contains_symbols (chrec_a)
2635 || chrec_contains_symbols (chrec_b))
2636 && (!evolution_function_is_affine_multivariate_p (chrec_a, lnn)
2637 || !evolution_function_is_affine_multivariate_p (chrec_b, lnn)))
2639 dependence_stats.num_subscript_undetermined++;
2640 *overlap_iterations_a = conflict_fn_not_known ();
2641 *overlap_iterations_b = conflict_fn_not_known ();
2644 else if (ziv_subscript_p (chrec_a, chrec_b))
2645 analyze_ziv_subscript (chrec_a, chrec_b,
2646 overlap_iterations_a, overlap_iterations_b,
2649 else if (siv_subscript_p (chrec_a, chrec_b))
2650 analyze_siv_subscript (chrec_a, chrec_b,
2651 overlap_iterations_a, overlap_iterations_b,
2655 analyze_miv_subscript (chrec_a, chrec_b,
2656 overlap_iterations_a, overlap_iterations_b,
2657 last_conflicts, loop_nest);
2659 if (dump_file && (dump_flags & TDF_DETAILS))
2661 fprintf (dump_file, " (overlap_iterations_a = ");
2662 dump_conflict_function (dump_file, *overlap_iterations_a);
2663 fprintf (dump_file, ")\n (overlap_iterations_b = ");
2664 dump_conflict_function (dump_file, *overlap_iterations_b);
2665 fprintf (dump_file, ")\n");
2666 fprintf (dump_file, ")\n");
2670 /* Helper function for uniquely inserting distance vectors. */
2673 save_dist_v (struct data_dependence_relation *ddr, lambda_vector dist_v)
2678 for (i = 0; VEC_iterate (lambda_vector, DDR_DIST_VECTS (ddr), i, v); i++)
2679 if (lambda_vector_equal (v, dist_v, DDR_NB_LOOPS (ddr)))
2682 VEC_safe_push (lambda_vector, heap, DDR_DIST_VECTS (ddr), dist_v);
2685 /* Helper function for uniquely inserting direction vectors. */
2688 save_dir_v (struct data_dependence_relation *ddr, lambda_vector dir_v)
2693 for (i = 0; VEC_iterate (lambda_vector, DDR_DIR_VECTS (ddr), i, v); i++)
2694 if (lambda_vector_equal (v, dir_v, DDR_NB_LOOPS (ddr)))
2697 VEC_safe_push (lambda_vector, heap, DDR_DIR_VECTS (ddr), dir_v);
2700 /* Add a distance of 1 on all the loops outer than INDEX. If we
2701 haven't yet determined a distance for this outer loop, push a new
2702 distance vector composed of the previous distance, and a distance
2703 of 1 for this outer loop. Example:
2711 Saved vectors are of the form (dist_in_1, dist_in_2). First, we
2712 save (0, 1), then we have to save (1, 0). */
2715 add_outer_distances (struct data_dependence_relation *ddr,
2716 lambda_vector dist_v, int index)
2718 /* For each outer loop where init_v is not set, the accesses are
2719 in dependence of distance 1 in the loop. */
2720 while (--index >= 0)
2722 lambda_vector save_v = lambda_vector_new (DDR_NB_LOOPS (ddr));
2723 lambda_vector_copy (dist_v, save_v, DDR_NB_LOOPS (ddr));
2725 save_dist_v (ddr, save_v);
2729 /* Return false when fail to represent the data dependence as a
2730 distance vector. INIT_B is set to true when a component has been
2731 added to the distance vector DIST_V. INDEX_CARRY is then set to
2732 the index in DIST_V that carries the dependence. */
2735 build_classic_dist_vector_1 (struct data_dependence_relation *ddr,
2736 struct data_reference *ddr_a,
2737 struct data_reference *ddr_b,
2738 lambda_vector dist_v, bool *init_b,
2742 lambda_vector init_v = lambda_vector_new (DDR_NB_LOOPS (ddr));
2744 for (i = 0; i < DDR_NUM_SUBSCRIPTS (ddr); i++)
2746 tree access_fn_a, access_fn_b;
2747 struct subscript *subscript = DDR_SUBSCRIPT (ddr, i);
2749 if (chrec_contains_undetermined (SUB_DISTANCE (subscript)))
2751 non_affine_dependence_relation (ddr);
2755 access_fn_a = DR_ACCESS_FN (ddr_a, i);
2756 access_fn_b = DR_ACCESS_FN (ddr_b, i);
2758 if (TREE_CODE (access_fn_a) == POLYNOMIAL_CHREC
2759 && TREE_CODE (access_fn_b) == POLYNOMIAL_CHREC)
2762 int index_a = index_in_loop_nest (CHREC_VARIABLE (access_fn_a),
2763 DDR_LOOP_NEST (ddr));
2764 int index_b = index_in_loop_nest (CHREC_VARIABLE (access_fn_b),
2765 DDR_LOOP_NEST (ddr));
2767 /* The dependence is carried by the outermost loop. Example:
2774 In this case, the dependence is carried by loop_1. */
2775 index = index_a < index_b ? index_a : index_b;
2776 *index_carry = MIN (index, *index_carry);
2778 if (chrec_contains_undetermined (SUB_DISTANCE (subscript)))
2780 non_affine_dependence_relation (ddr);
2784 dist = int_cst_value (SUB_DISTANCE (subscript));
2786 /* This is the subscript coupling test. If we have already
2787 recorded a distance for this loop (a distance coming from
2788 another subscript), it should be the same. For example,
2789 in the following code, there is no dependence:
2796 if (init_v[index] != 0 && dist_v[index] != dist)
2798 finalize_ddr_dependent (ddr, chrec_known);
2802 dist_v[index] = dist;
2806 else if (!operand_equal_p (access_fn_a, access_fn_b, 0))
2808 /* This can be for example an affine vs. constant dependence
2809 (T[i] vs. T[3]) that is not an affine dependence and is
2810 not representable as a distance vector. */
2811 non_affine_dependence_relation (ddr);
2819 /* Return true when the DDR contains only constant access functions. */
2822 constant_access_functions (const struct data_dependence_relation *ddr)
2826 for (i = 0; i < DDR_NUM_SUBSCRIPTS (ddr); i++)
2827 if (!evolution_function_is_constant_p (DR_ACCESS_FN (DDR_A (ddr), i))
2828 || !evolution_function_is_constant_p (DR_ACCESS_FN (DDR_B (ddr), i)))
2834 /* Helper function for the case where DDR_A and DDR_B are the same
2835 multivariate access function with a constant step. For an example
2839 add_multivariate_self_dist (struct data_dependence_relation *ddr, tree c_2)
2842 tree c_1 = CHREC_LEFT (c_2);
2843 tree c_0 = CHREC_LEFT (c_1);
2844 lambda_vector dist_v;
2847 /* Polynomials with more than 2 variables are not handled yet. When
2848 the evolution steps are parameters, it is not possible to
2849 represent the dependence using classical distance vectors. */
2850 if (TREE_CODE (c_0) != INTEGER_CST
2851 || TREE_CODE (CHREC_RIGHT (c_1)) != INTEGER_CST
2852 || TREE_CODE (CHREC_RIGHT (c_2)) != INTEGER_CST)
2854 DDR_AFFINE_P (ddr) = false;
2858 x_2 = index_in_loop_nest (CHREC_VARIABLE (c_2), DDR_LOOP_NEST (ddr));
2859 x_1 = index_in_loop_nest (CHREC_VARIABLE (c_1), DDR_LOOP_NEST (ddr));
2861 /* For "{{0, +, 2}_1, +, 3}_2" the distance vector is (3, -2). */
2862 dist_v = lambda_vector_new (DDR_NB_LOOPS (ddr));
2863 v1 = int_cst_value (CHREC_RIGHT (c_1));
2864 v2 = int_cst_value (CHREC_RIGHT (c_2));
2877 save_dist_v (ddr, dist_v);
2879 add_outer_distances (ddr, dist_v, x_1);
2882 /* Helper function for the case where DDR_A and DDR_B are the same
2883 access functions. */
2886 add_other_self_distances (struct data_dependence_relation *ddr)
2888 lambda_vector dist_v;
2890 int index_carry = DDR_NB_LOOPS (ddr);
2892 for (i = 0; i < DDR_NUM_SUBSCRIPTS (ddr); i++)
2894 tree access_fun = DR_ACCESS_FN (DDR_A (ddr), i);
2896 if (TREE_CODE (access_fun) == POLYNOMIAL_CHREC)
2898 if (!evolution_function_is_univariate_p (access_fun))
2900 if (DDR_NUM_SUBSCRIPTS (ddr) != 1)
2902 DDR_ARE_DEPENDENT (ddr) = chrec_dont_know;
2906 access_fun = DR_ACCESS_FN (DDR_A (ddr), 0);
2908 if (TREE_CODE (CHREC_LEFT (access_fun)) == POLYNOMIAL_CHREC)
2909 add_multivariate_self_dist (ddr, access_fun);
2911 /* The evolution step is not constant: it varies in
2912 the outer loop, so this cannot be represented by a
2913 distance vector. For example in pr34635.c the
2914 evolution is {0, +, {0, +, 4}_1}_2. */
2915 DDR_AFFINE_P (ddr) = false;
2920 index_carry = MIN (index_carry,
2921 index_in_loop_nest (CHREC_VARIABLE (access_fun),
2922 DDR_LOOP_NEST (ddr)));
2926 dist_v = lambda_vector_new (DDR_NB_LOOPS (ddr));
2927 add_outer_distances (ddr, dist_v, index_carry);
2931 insert_innermost_unit_dist_vector (struct data_dependence_relation *ddr)
2933 lambda_vector dist_v = lambda_vector_new (DDR_NB_LOOPS (ddr));
2935 dist_v[DDR_INNER_LOOP (ddr)] = 1;
2936 save_dist_v (ddr, dist_v);
2939 /* Adds a unit distance vector to DDR when there is a 0 overlap. This
2940 is the case for example when access functions are the same and
2941 equal to a constant, as in:
2948 in which case the distance vectors are (0) and (1). */
2951 add_distance_for_zero_overlaps (struct data_dependence_relation *ddr)
2955 for (i = 0; i < DDR_NUM_SUBSCRIPTS (ddr); i++)
2957 subscript_p sub = DDR_SUBSCRIPT (ddr, i);
2958 conflict_function *ca = SUB_CONFLICTS_IN_A (sub);
2959 conflict_function *cb = SUB_CONFLICTS_IN_B (sub);
2961 for (j = 0; j < ca->n; j++)
2962 if (affine_function_zero_p (ca->fns[j]))
2964 insert_innermost_unit_dist_vector (ddr);
2968 for (j = 0; j < cb->n; j++)
2969 if (affine_function_zero_p (cb->fns[j]))
2971 insert_innermost_unit_dist_vector (ddr);
2977 /* Compute the classic per loop distance vector. DDR is the data
2978 dependence relation to build a vector from. Return false when fail
2979 to represent the data dependence as a distance vector. */
2982 build_classic_dist_vector (struct data_dependence_relation *ddr,
2983 struct loop *loop_nest)
2985 bool init_b = false;
2986 int index_carry = DDR_NB_LOOPS (ddr);
2987 lambda_vector dist_v;
2989 if (DDR_ARE_DEPENDENT (ddr) != NULL_TREE)
2992 if (same_access_functions (ddr))
2994 /* Save the 0 vector. */
2995 dist_v = lambda_vector_new (DDR_NB_LOOPS (ddr));
2996 save_dist_v (ddr, dist_v);
2998 if (constant_access_functions (ddr))
2999 add_distance_for_zero_overlaps (ddr);
3001 if (DDR_NB_LOOPS (ddr) > 1)
3002 add_other_self_distances (ddr);
3007 dist_v = lambda_vector_new (DDR_NB_LOOPS (ddr));
3008 if (!build_classic_dist_vector_1 (ddr, DDR_A (ddr), DDR_B (ddr),
3009 dist_v, &init_b, &index_carry))
3012 /* Save the distance vector if we initialized one. */
3015 /* Verify a basic constraint: classic distance vectors should
3016 always be lexicographically positive.
3018 Data references are collected in the order of execution of
3019 the program, thus for the following loop
3021 | for (i = 1; i < 100; i++)
3022 | for (j = 1; j < 100; j++)
3024 | t = T[j+1][i-1]; // A
3025 | T[j][i] = t + 2; // B
3028 references are collected following the direction of the wind:
3029 A then B. The data dependence tests are performed also
3030 following this order, such that we're looking at the distance
3031 separating the elements accessed by A from the elements later
3032 accessed by B. But in this example, the distance returned by
3033 test_dep (A, B) is lexicographically negative (-1, 1), that
3034 means that the access A occurs later than B with respect to
3035 the outer loop, ie. we're actually looking upwind. In this
3036 case we solve test_dep (B, A) looking downwind to the
3037 lexicographically positive solution, that returns the
3038 distance vector (1, -1). */
3039 if (!lambda_vector_lexico_pos (dist_v, DDR_NB_LOOPS (ddr)))
3041 lambda_vector save_v = lambda_vector_new (DDR_NB_LOOPS (ddr));
3042 if (!subscript_dependence_tester_1 (ddr, DDR_B (ddr), DDR_A (ddr),
3045 compute_subscript_distance (ddr);
3046 if (!build_classic_dist_vector_1 (ddr, DDR_B (ddr), DDR_A (ddr),
3047 save_v, &init_b, &index_carry))
3049 save_dist_v (ddr, save_v);
3050 DDR_REVERSED_P (ddr) = true;
3052 /* In this case there is a dependence forward for all the
3055 | for (k = 1; k < 100; k++)
3056 | for (i = 1; i < 100; i++)
3057 | for (j = 1; j < 100; j++)
3059 | t = T[j+1][i-1]; // A
3060 | T[j][i] = t + 2; // B
3068 if (DDR_NB_LOOPS (ddr) > 1)
3070 add_outer_distances (ddr, save_v, index_carry);
3071 add_outer_distances (ddr, dist_v, index_carry);
3076 lambda_vector save_v = lambda_vector_new (DDR_NB_LOOPS (ddr));
3077 lambda_vector_copy (dist_v, save_v, DDR_NB_LOOPS (ddr));
3079 if (DDR_NB_LOOPS (ddr) > 1)
3081 lambda_vector opposite_v = lambda_vector_new (DDR_NB_LOOPS (ddr));
3083 if (!subscript_dependence_tester_1 (ddr, DDR_B (ddr),
3084 DDR_A (ddr), loop_nest))
3086 compute_subscript_distance (ddr);
3087 if (!build_classic_dist_vector_1 (ddr, DDR_B (ddr), DDR_A (ddr),
3088 opposite_v, &init_b,
3092 save_dist_v (ddr, save_v);
3093 add_outer_distances (ddr, dist_v, index_carry);
3094 add_outer_distances (ddr, opposite_v, index_carry);
3097 save_dist_v (ddr, save_v);
3102 /* There is a distance of 1 on all the outer loops: Example:
3103 there is a dependence of distance 1 on loop_1 for the array A.
3109 add_outer_distances (ddr, dist_v,
3110 lambda_vector_first_nz (dist_v,
3111 DDR_NB_LOOPS (ddr), 0));
3114 if (dump_file && (dump_flags & TDF_DETAILS))
3118 fprintf (dump_file, "(build_classic_dist_vector\n");
3119 for (i = 0; i < DDR_NUM_DIST_VECTS (ddr); i++)
3121 fprintf (dump_file, " dist_vector = (");
3122 print_lambda_vector (dump_file, DDR_DIST_VECT (ddr, i),
3123 DDR_NB_LOOPS (ddr));
3124 fprintf (dump_file, " )\n");
3126 fprintf (dump_file, ")\n");
3132 /* Return the direction for a given distance.
3133 FIXME: Computing dir this way is suboptimal, since dir can catch
3134 cases that dist is unable to represent. */
3136 static inline enum data_dependence_direction
3137 dir_from_dist (int dist)
3140 return dir_positive;
3142 return dir_negative;
3147 /* Compute the classic per loop direction vector. DDR is the data
3148 dependence relation to build a vector from. */
3151 build_classic_dir_vector (struct data_dependence_relation *ddr)
3154 lambda_vector dist_v;
3156 for (i = 0; VEC_iterate (lambda_vector, DDR_DIST_VECTS (ddr), i, dist_v); i++)
3158 lambda_vector dir_v = lambda_vector_new (DDR_NB_LOOPS (ddr));
3160 for (j = 0; j < DDR_NB_LOOPS (ddr); j++)
3161 dir_v[j] = dir_from_dist (dist_v[j]);
3163 save_dir_v (ddr, dir_v);
3167 /* Helper function. Returns true when there is a dependence between
3168 data references DRA and DRB. */
3171 subscript_dependence_tester_1 (struct data_dependence_relation *ddr,
3172 struct data_reference *dra,
3173 struct data_reference *drb,
3174 struct loop *loop_nest)
3177 tree last_conflicts;
3178 struct subscript *subscript;
3180 for (i = 0; VEC_iterate (subscript_p, DDR_SUBSCRIPTS (ddr), i, subscript);
3183 conflict_function *overlaps_a, *overlaps_b;
3185 analyze_overlapping_iterations (DR_ACCESS_FN (dra, i),
3186 DR_ACCESS_FN (drb, i),
3187 &overlaps_a, &overlaps_b,
3188 &last_conflicts, loop_nest);
3190 if (CF_NOT_KNOWN_P (overlaps_a)
3191 || CF_NOT_KNOWN_P (overlaps_b))
3193 finalize_ddr_dependent (ddr, chrec_dont_know);
3194 dependence_stats.num_dependence_undetermined++;
3195 free_conflict_function (overlaps_a);
3196 free_conflict_function (overlaps_b);
3200 else if (CF_NO_DEPENDENCE_P (overlaps_a)
3201 || CF_NO_DEPENDENCE_P (overlaps_b))
3203 finalize_ddr_dependent (ddr, chrec_known);
3204 dependence_stats.num_dependence_independent++;
3205 free_conflict_function (overlaps_a);
3206 free_conflict_function (overlaps_b);
3212 if (SUB_CONFLICTS_IN_A (subscript))
3213 free_conflict_function (SUB_CONFLICTS_IN_A (subscript));
3214 if (SUB_CONFLICTS_IN_B (subscript))
3215 free_conflict_function (SUB_CONFLICTS_IN_B (subscript));
3217 SUB_CONFLICTS_IN_A (subscript) = overlaps_a;
3218 SUB_CONFLICTS_IN_B (subscript) = overlaps_b;
3219 SUB_LAST_CONFLICT (subscript) = last_conflicts;
3226 /* Computes the conflicting iterations in LOOP_NEST, and initialize DDR. */
3229 subscript_dependence_tester (struct data_dependence_relation *ddr,
3230 struct loop *loop_nest)
3233 if (dump_file && (dump_flags & TDF_DETAILS))
3234 fprintf (dump_file, "(subscript_dependence_tester \n");
3236 if (subscript_dependence_tester_1 (ddr, DDR_A (ddr), DDR_B (ddr), loop_nest))
3237 dependence_stats.num_dependence_dependent++;
3239 compute_subscript_distance (ddr);
3240 if (build_classic_dist_vector (ddr, loop_nest))
3241 build_classic_dir_vector (ddr);
3243 if (dump_file && (dump_flags & TDF_DETAILS))
3244 fprintf (dump_file, ")\n");
3247 /* Returns true when all the access functions of A are affine or
3248 constant with respect to LOOP_NEST. */
3251 access_functions_are_affine_or_constant_p (const struct data_reference *a,
3252 const struct loop *loop_nest)
3255 VEC(tree,heap) *fns = DR_ACCESS_FNS (a);
3258 for (i = 0; VEC_iterate (tree, fns, i, t); i++)
3259 if (!evolution_function_is_invariant_p (t, loop_nest->num)
3260 && !evolution_function_is_affine_multivariate_p (t, loop_nest->num))
3266 /* Initializes an equation for an OMEGA problem using the information
3267 contained in the ACCESS_FUN. Returns true when the operation
3270 PB is the omega constraint system.
3271 EQ is the number of the equation to be initialized.
3272 OFFSET is used for shifting the variables names in the constraints:
3273 a constrain is composed of 2 * the number of variables surrounding
3274 dependence accesses. OFFSET is set either to 0 for the first n variables,
3275 then it is set to n.
3276 ACCESS_FUN is expected to be an affine chrec. */
3279 init_omega_eq_with_af (omega_pb pb, unsigned eq,
3280 unsigned int offset, tree access_fun,
3281 struct data_dependence_relation *ddr)
3283 switch (TREE_CODE (access_fun))
3285 case POLYNOMIAL_CHREC:
3287 tree left = CHREC_LEFT (access_fun);
3288 tree right = CHREC_RIGHT (access_fun);
3289 int var = CHREC_VARIABLE (access_fun);
3292 if (TREE_CODE (right) != INTEGER_CST)
3295 var_idx = index_in_loop_nest (var, DDR_LOOP_NEST (ddr));
3296 pb->eqs[eq].coef[offset + var_idx + 1] = int_cst_value (right);
3298 /* Compute the innermost loop index. */
3299 DDR_INNER_LOOP (ddr) = MAX (DDR_INNER_LOOP (ddr), var_idx);
3302 pb->eqs[eq].coef[var_idx + DDR_NB_LOOPS (ddr) + 1]
3303 += int_cst_value (right);
3305 switch (TREE_CODE (left))
3307 case POLYNOMIAL_CHREC:
3308 return init_omega_eq_with_af (pb, eq, offset, left, ddr);
3311 pb->eqs[eq].coef[0] += int_cst_value (left);
3320 pb->eqs[eq].coef[0] += int_cst_value (access_fun);
3328 /* As explained in the comments preceding init_omega_for_ddr, we have
3329 to set up a system for each loop level, setting outer loops
3330 variation to zero, and current loop variation to positive or zero.
3331 Save each lexico positive distance vector. */
3334 omega_extract_distance_vectors (omega_pb pb,
3335 struct data_dependence_relation *ddr)
3339 struct loop *loopi, *loopj;
3340 enum omega_result res;
3342 /* Set a new problem for each loop in the nest. The basis is the
3343 problem that we have initialized until now. On top of this we
3344 add new constraints. */
3345 for (i = 0; i <= DDR_INNER_LOOP (ddr)
3346 && VEC_iterate (loop_p, DDR_LOOP_NEST (ddr), i, loopi); i++)
3349 omega_pb copy = omega_alloc_problem (2 * DDR_NB_LOOPS (ddr),
3350 DDR_NB_LOOPS (ddr));
3352 omega_copy_problem (copy, pb);
3354 /* For all the outer loops "loop_j", add "dj = 0". */
3356 j < i && VEC_iterate (loop_p, DDR_LOOP_NEST (ddr), j, loopj); j++)
3358 eq = omega_add_zero_eq (copy, omega_black);
3359 copy->eqs[eq].coef[j + 1] = 1;
3362 /* For "loop_i", add "0 <= di". */
3363 geq = omega_add_zero_geq (copy, omega_black);
3364 copy->geqs[geq].coef[i + 1] = 1;
3366 /* Reduce the constraint system, and test that the current
3367 problem is feasible. */
3368 res = omega_simplify_problem (copy);
3369 if (res == omega_false
3370 || res == omega_unknown
3371 || copy->num_geqs > (int) DDR_NB_LOOPS (ddr))
3374 for (eq = 0; eq < copy->num_subs; eq++)
3375 if (copy->subs[eq].key == (int) i + 1)
3377 dist = copy->subs[eq].coef[0];
3383 /* Reinitialize problem... */
3384 omega_copy_problem (copy, pb);
3386 j < i && VEC_iterate (loop_p, DDR_LOOP_NEST (ddr), j, loopj); j++)
3388 eq = omega_add_zero_eq (copy, omega_black);
3389 copy->eqs[eq].coef[j + 1] = 1;
3392 /* ..., but this time "di = 1". */
3393 eq = omega_add_zero_eq (copy, omega_black);
3394 copy->eqs[eq].coef[i + 1] = 1;
3395 copy->eqs[eq].coef[0] = -1;
3397 res = omega_simplify_problem (copy);
3398 if (res == omega_false
3399 || res == omega_unknown
3400 || copy->num_geqs > (int) DDR_NB_LOOPS (ddr))
3403 for (eq = 0; eq < copy->num_subs; eq++)
3404 if (copy->subs[eq].key == (int) i + 1)
3406 dist = copy->subs[eq].coef[0];
3412 /* Save the lexicographically positive distance vector. */
3415 lambda_vector dist_v = lambda_vector_new (DDR_NB_LOOPS (ddr));
3416 lambda_vector dir_v = lambda_vector_new (DDR_NB_LOOPS (ddr));
3420 for (eq = 0; eq < copy->num_subs; eq++)
3421 if (copy->subs[eq].key > 0)
3423 dist = copy->subs[eq].coef[0];
3424 dist_v[copy->subs[eq].key - 1] = dist;
3427 for (j = 0; j < DDR_NB_LOOPS (ddr); j++)
3428 dir_v[j] = dir_from_dist (dist_v[j]);
3430 save_dist_v (ddr, dist_v);
3431 save_dir_v (ddr, dir_v);
3435 omega_free_problem (copy);
3439 /* This is called for each subscript of a tuple of data references:
3440 insert an equality for representing the conflicts. */
3443 omega_setup_subscript (tree access_fun_a, tree access_fun_b,
3444 struct data_dependence_relation *ddr,
3445 omega_pb pb, bool *maybe_dependent)
3448 tree type = signed_type_for_types (TREE_TYPE (access_fun_a),
3449 TREE_TYPE (access_fun_b));
3450 tree fun_a = chrec_convert (type, access_fun_a, NULL_TREE);
3451 tree fun_b = chrec_convert (type, access_fun_b, NULL_TREE);
3452 tree difference = chrec_fold_minus (type, fun_a, fun_b);
3454 /* When the fun_a - fun_b is not constant, the dependence is not
3455 captured by the classic distance vector representation. */
3456 if (TREE_CODE (difference) != INTEGER_CST)
3460 if (ziv_subscript_p (fun_a, fun_b) && !integer_zerop (difference))
3462 /* There is no dependence. */
3463 *maybe_dependent = false;
3467 fun_b = chrec_fold_multiply (type, fun_b, integer_minus_one_node);
3469 eq = omega_add_zero_eq (pb, omega_black);
3470 if (!init_omega_eq_with_af (pb, eq, DDR_NB_LOOPS (ddr), fun_a, ddr)
3471 || !init_omega_eq_with_af (pb, eq, 0, fun_b, ddr))
3472 /* There is probably a dependence, but the system of
3473 constraints cannot be built: answer "don't know". */
3477 if (DDR_NB_LOOPS (ddr) != 0 && pb->eqs[eq].coef[0]
3478 && !int_divides_p (lambda_vector_gcd
3479 ((lambda_vector) &(pb->eqs[eq].coef[1]),
3480 2 * DDR_NB_LOOPS (ddr)),
3481 pb->eqs[eq].coef[0]))
3483 /* There is no dependence. */
3484 *maybe_dependent = false;
3491 /* Helper function, same as init_omega_for_ddr but specialized for
3492 data references A and B. */
3495 init_omega_for_ddr_1 (struct data_reference *dra, struct data_reference *drb,
3496 struct data_dependence_relation *ddr,
3497 omega_pb pb, bool *maybe_dependent)
3502 unsigned nb_loops = DDR_NB_LOOPS (ddr);
3504 /* Insert an equality per subscript. */
3505 for (i = 0; i < DDR_NUM_SUBSCRIPTS (ddr); i++)
3507 if (!omega_setup_subscript (DR_ACCESS_FN (dra, i), DR_ACCESS_FN (drb, i),
3508 ddr, pb, maybe_dependent))
3510 else if (*maybe_dependent == false)
3512 /* There is no dependence. */
3513 DDR_ARE_DEPENDENT (ddr) = chrec_known;
3518 /* Insert inequalities: constraints corresponding to the iteration
3519 domain, i.e. the loops surrounding the references "loop_x" and
3520 the distance variables "dx". The layout of the OMEGA
3521 representation is as follows:
3522 - coef[0] is the constant
3523 - coef[1..nb_loops] are the protected variables that will not be
3524 removed by the solver: the "dx"
3525 - coef[nb_loops + 1, 2*nb_loops] are the loop variables: "loop_x".
3527 for (i = 0; i <= DDR_INNER_LOOP (ddr)
3528 && VEC_iterate (loop_p, DDR_LOOP_NEST (ddr), i, loopi); i++)
3530 HOST_WIDE_INT nbi = estimated_loop_iterations_int (loopi, false);
3533 ineq = omega_add_zero_geq (pb, omega_black);
3534 pb->geqs[ineq].coef[i + nb_loops + 1] = 1;
3536 /* 0 <= loop_x + dx */
3537 ineq = omega_add_zero_geq (pb, omega_black);
3538 pb->geqs[ineq].coef[i + nb_loops + 1] = 1;
3539 pb->geqs[ineq].coef[i + 1] = 1;
3543 /* loop_x <= nb_iters */
3544 ineq = omega_add_zero_geq (pb, omega_black);
3545 pb->geqs[ineq].coef[i + nb_loops + 1] = -1;
3546 pb->geqs[ineq].coef[0] = nbi;
3548 /* loop_x + dx <= nb_iters */
3549 ineq = omega_add_zero_geq (pb, omega_black);
3550 pb->geqs[ineq].coef[i + nb_loops + 1] = -1;
3551 pb->geqs[ineq].coef[i + 1] = -1;
3552 pb->geqs[ineq].coef[0] = nbi;
3554 /* A step "dx" bigger than nb_iters is not feasible, so
3555 add "0 <= nb_iters + dx", */
3556 ineq = omega_add_zero_geq (pb, omega_black);
3557 pb->geqs[ineq].coef[i + 1] = 1;
3558 pb->geqs[ineq].coef[0] = nbi;
3559 /* and "dx <= nb_iters". */
3560 ineq = omega_add_zero_geq (pb, omega_black);
3561 pb->geqs[ineq].coef[i + 1] = -1;
3562 pb->geqs[ineq].coef[0] = nbi;
3566 omega_extract_distance_vectors (pb, ddr);
3571 /* Sets up the Omega dependence problem for the data dependence
3572 relation DDR. Returns false when the constraint system cannot be
3573 built, ie. when the test answers "don't know". Returns true
3574 otherwise, and when independence has been proved (using one of the
3575 trivial dependence test), set MAYBE_DEPENDENT to false, otherwise
3576 set MAYBE_DEPENDENT to true.
3578 Example: for setting up the dependence system corresponding to the
3579 conflicting accesses
3584 | ... A[2*j, 2*(i + j)]
3588 the following constraints come from the iteration domain:
3595 where di, dj are the distance variables. The constraints
3596 representing the conflicting elements are:
3599 i + 1 = 2 * (i + di + j + dj)
3601 For asking that the resulting distance vector (di, dj) be
3602 lexicographically positive, we insert the constraint "di >= 0". If
3603 "di = 0" in the solution, we fix that component to zero, and we
3604 look at the inner loops: we set a new problem where all the outer
3605 loop distances are zero, and fix this inner component to be
3606 positive. When one of the components is positive, we save that
3607 distance, and set a new problem where the distance on this loop is
3608 zero, searching for other distances in the inner loops. Here is
3609 the classic example that illustrates that we have to set for each
3610 inner loop a new problem:
3618 we have to save two distances (1, 0) and (0, 1).
3620 Given two array references, refA and refB, we have to set the
3621 dependence problem twice, refA vs. refB and refB vs. refA, and we
3622 cannot do a single test, as refB might occur before refA in the
3623 inner loops, and the contrary when considering outer loops: ex.
3628 | T[{1,+,1}_2][{1,+,1}_1] // refA
3629 | T[{2,+,1}_2][{0,+,1}_1] // refB
3634 refB touches the elements in T before refA, and thus for the same
3635 loop_0 refB precedes refA: ie. the distance vector (0, 1, -1)
3636 but for successive loop_0 iterations, we have (1, -1, 1)
3638 The Omega solver expects the distance variables ("di" in the
3639 previous example) to come first in the constraint system (as
3640 variables to be protected, or "safe" variables), the constraint
3641 system is built using the following layout:
3643 "cst | distance vars | index vars".
3647 init_omega_for_ddr (struct data_dependence_relation *ddr,
3648 bool *maybe_dependent)
3653 *maybe_dependent = true;
3655 if (same_access_functions (ddr))
3658 lambda_vector dir_v;
3660 /* Save the 0 vector. */
3661 save_dist_v (ddr, lambda_vector_new (DDR_NB_LOOPS (ddr)));
3662 dir_v = lambda_vector_new (DDR_NB_LOOPS (ddr));
3663 for (j = 0; j < DDR_NB_LOOPS (ddr); j++)
3664 dir_v[j] = dir_equal;
3665 save_dir_v (ddr, dir_v);
3667 /* Save the dependences carried by outer loops. */
3668 pb = omega_alloc_problem (2 * DDR_NB_LOOPS (ddr), DDR_NB_LOOPS (ddr));
3669 res = init_omega_for_ddr_1 (DDR_A (ddr), DDR_B (ddr), ddr, pb,
3671 omega_free_problem (pb);
3675 /* Omega expects the protected variables (those that have to be kept
3676 after elimination) to appear first in the constraint system.
3677 These variables are the distance variables. In the following
3678 initialization we declare NB_LOOPS safe variables, and the total
3679 number of variables for the constraint system is 2*NB_LOOPS. */
3680 pb = omega_alloc_problem (2 * DDR_NB_LOOPS (ddr), DDR_NB_LOOPS (ddr));
3681 res = init_omega_for_ddr_1 (DDR_A (ddr), DDR_B (ddr), ddr, pb,
3683 omega_free_problem (pb);
3685 /* Stop computation if not decidable, or no dependence. */
3686 if (res == false || *maybe_dependent == false)
3689 pb = omega_alloc_problem (2 * DDR_NB_LOOPS (ddr), DDR_NB_LOOPS (ddr));
3690 res = init_omega_for_ddr_1 (DDR_B (ddr), DDR_A (ddr), ddr, pb,
3692 omega_free_problem (pb);
3697 /* Return true when DDR contains the same information as that stored
3698 in DIR_VECTS and in DIST_VECTS, return false otherwise. */
3701 ddr_consistent_p (FILE *file,
3702 struct data_dependence_relation *ddr,
3703 VEC (lambda_vector, heap) *dist_vects,
3704 VEC (lambda_vector, heap) *dir_vects)
3708 /* If dump_file is set, output there. */
3709 if (dump_file && (dump_flags & TDF_DETAILS))
3712 if (VEC_length (lambda_vector, dist_vects) != DDR_NUM_DIST_VECTS (ddr))
3714 lambda_vector b_dist_v;
3715 fprintf (file, "\n(Number of distance vectors differ: Banerjee has %d, Omega has %d.\n",
3716 VEC_length (lambda_vector, dist_vects),
3717 DDR_NUM_DIST_VECTS (ddr));
3719 fprintf (file, "Banerjee dist vectors:\n");
3720 for (i = 0; VEC_iterate (lambda_vector, dist_vects, i, b_dist_v); i++)
3721 print_lambda_vector (file, b_dist_v, DDR_NB_LOOPS (ddr));
3723 fprintf (file, "Omega dist vectors:\n");
3724 for (i = 0; i < DDR_NUM_DIST_VECTS (ddr); i++)
3725 print_lambda_vector (file, DDR_DIST_VECT (ddr, i), DDR_NB_LOOPS (ddr));
3727 fprintf (file, "data dependence relation:\n");
3728 dump_data_dependence_relation (file, ddr);
3730 fprintf (file, ")\n");
3734 if (VEC_length (lambda_vector, dir_vects) != DDR_NUM_DIR_VECTS (ddr))
3736 fprintf (file, "\n(Number of direction vectors differ: Banerjee has %d, Omega has %d.)\n",
3737 VEC_length (lambda_vector, dir_vects),
3738 DDR_NUM_DIR_VECTS (ddr));
3742 for (i = 0; i < DDR_NUM_DIST_VECTS (ddr); i++)
3744 lambda_vector a_dist_v;
3745 lambda_vector b_dist_v = DDR_DIST_VECT (ddr, i);
3747 /* Distance vectors are not ordered in the same way in the DDR
3748 and in the DIST_VECTS: search for a matching vector. */
3749 for (j = 0; VEC_iterate (lambda_vector, dist_vects, j, a_dist_v); j++)
3750 if (lambda_vector_equal (a_dist_v, b_dist_v, DDR_NB_LOOPS (ddr)))
3753 if (j == VEC_length (lambda_vector, dist_vects))
3755 fprintf (file, "\n(Dist vectors from the first dependence analyzer:\n");
3756 print_dist_vectors (file, dist_vects, DDR_NB_LOOPS (ddr));
3757 fprintf (file, "not found in Omega dist vectors:\n");
3758 print_dist_vectors (file, DDR_DIST_VECTS (ddr), DDR_NB_LOOPS (ddr));
3759 fprintf (file, "data dependence relation:\n");
3760 dump_data_dependence_relation (file, ddr);
3761 fprintf (file, ")\n");
3765 for (i = 0; i < DDR_NUM_DIR_VECTS (ddr); i++)
3767 lambda_vector a_dir_v;
3768 lambda_vector b_dir_v = DDR_DIR_VECT (ddr, i);
3770 /* Direction vectors are not ordered in the same way in the DDR
3771 and in the DIR_VECTS: search for a matching vector. */
3772 for (j = 0; VEC_iterate (lambda_vector, dir_vects, j, a_dir_v); j++)
3773 if (lambda_vector_equal (a_dir_v, b_dir_v, DDR_NB_LOOPS (ddr)))
3776 if (j == VEC_length (lambda_vector, dist_vects))
3778 fprintf (file, "\n(Dir vectors from the first dependence analyzer:\n");
3779 print_dir_vectors (file, dir_vects, DDR_NB_LOOPS (ddr));
3780 fprintf (file, "not found in Omega dir vectors:\n");
3781 print_dir_vectors (file, DDR_DIR_VECTS (ddr), DDR_NB_LOOPS (ddr));
3782 fprintf (file, "data dependence relation:\n");
3783 dump_data_dependence_relation (file, ddr);
3784 fprintf (file, ")\n");
3791 /* This computes the affine dependence relation between A and B with
3792 respect to LOOP_NEST. CHREC_KNOWN is used for representing the
3793 independence between two accesses, while CHREC_DONT_KNOW is used
3794 for representing the unknown relation.
3796 Note that it is possible to stop the computation of the dependence
3797 relation the first time we detect a CHREC_KNOWN element for a given
3801 compute_affine_dependence (struct data_dependence_relation *ddr,
3802 struct loop *loop_nest)
3804 struct data_reference *dra = DDR_A (ddr);
3805 struct data_reference *drb = DDR_B (ddr);
3807 if (dump_file && (dump_flags & TDF_DETAILS))
3809 fprintf (dump_file, "(compute_affine_dependence\n");
3810 fprintf (dump_file, " (stmt_a = \n");
3811 print_generic_expr (dump_file, DR_STMT (dra), 0);
3812 fprintf (dump_file, ")\n (stmt_b = \n");
3813 print_generic_expr (dump_file, DR_STMT (drb), 0);
3814 fprintf (dump_file, ")\n");
3817 /* Analyze only when the dependence relation is not yet known. */
3818 if (DDR_ARE_DEPENDENT (ddr) == NULL_TREE
3819 && !DDR_SELF_REFERENCE (ddr))
3821 dependence_stats.num_dependence_tests++;
3823 if (access_functions_are_affine_or_constant_p (dra, loop_nest)
3824 && access_functions_are_affine_or_constant_p (drb, loop_nest))
3826 if (flag_check_data_deps)
3828 /* Compute the dependences using the first algorithm. */
3829 subscript_dependence_tester (ddr, loop_nest);
3831 if (dump_file && (dump_flags & TDF_DETAILS))
3833 fprintf (dump_file, "\n\nBanerjee Analyzer\n");
3834 dump_data_dependence_relation (dump_file, ddr);
3837 if (DDR_ARE_DEPENDENT (ddr) == NULL_TREE)
3839 bool maybe_dependent;
3840 VEC (lambda_vector, heap) *dir_vects, *dist_vects;
3842 /* Save the result of the first DD analyzer. */
3843 dist_vects = DDR_DIST_VECTS (ddr);
3844 dir_vects = DDR_DIR_VECTS (ddr);
3846 /* Reset the information. */
3847 DDR_DIST_VECTS (ddr) = NULL;
3848 DDR_DIR_VECTS (ddr) = NULL;
3850 /* Compute the same information using Omega. */
3851 if (!init_omega_for_ddr (ddr, &maybe_dependent))
3852 goto csys_dont_know;
3854 if (dump_file && (dump_flags & TDF_DETAILS))
3856 fprintf (dump_file, "Omega Analyzer\n");
3857 dump_data_dependence_relation (dump_file, ddr);
3860 /* Check that we get the same information. */
3861 if (maybe_dependent)
3862 gcc_assert (ddr_consistent_p (stderr, ddr, dist_vects,
3867 subscript_dependence_tester (ddr, loop_nest);
3870 /* As a last case, if the dependence cannot be determined, or if
3871 the dependence is considered too difficult to determine, answer
3876 dependence_stats.num_dependence_undetermined++;
3878 if (dump_file && (dump_flags & TDF_DETAILS))
3880 fprintf (dump_file, "Data ref a:\n");
3881 dump_data_reference (dump_file, dra);
3882 fprintf (dump_file, "Data ref b:\n");
3883 dump_data_reference (dump_file, drb);
3884 fprintf (dump_file, "affine dependence test not usable: access function not affine or constant.\n");
3886 finalize_ddr_dependent (ddr, chrec_dont_know);
3890 if (dump_file && (dump_flags & TDF_DETAILS))
3891 fprintf (dump_file, ")\n");
3894 /* This computes the dependence relation for the same data
3895 reference into DDR. */
3898 compute_self_dependence (struct data_dependence_relation *ddr)
3901 struct subscript *subscript;
3903 if (DDR_ARE_DEPENDENT (ddr) != NULL_TREE)
3906 for (i = 0; VEC_iterate (subscript_p, DDR_SUBSCRIPTS (ddr), i, subscript);
3909 if (SUB_CONFLICTS_IN_A (subscript))
3910 free_conflict_function (SUB_CONFLICTS_IN_A (subscript));
3911 if (SUB_CONFLICTS_IN_B (subscript))
3912 free_conflict_function (SUB_CONFLICTS_IN_B (subscript));
3914 /* The accessed index overlaps for each iteration. */
3915 SUB_CONFLICTS_IN_A (subscript)
3916 = conflict_fn (1, affine_fn_cst (integer_zero_node));
3917 SUB_CONFLICTS_IN_B (subscript)
3918 = conflict_fn (1, affine_fn_cst (integer_zero_node));
3919 SUB_LAST_CONFLICT (subscript) = chrec_dont_know;
3922 /* The distance vector is the zero vector. */
3923 save_dist_v (ddr, lambda_vector_new (DDR_NB_LOOPS (ddr)));
3924 save_dir_v (ddr, lambda_vector_new (DDR_NB_LOOPS (ddr)));
3927 /* Compute in DEPENDENCE_RELATIONS the data dependence graph for all
3928 the data references in DATAREFS, in the LOOP_NEST. When
3929 COMPUTE_SELF_AND_RR is FALSE, don't compute read-read and self
3933 compute_all_dependences (VEC (data_reference_p, heap) *datarefs,
3934 VEC (ddr_p, heap) **dependence_relations,
3935 VEC (loop_p, heap) *loop_nest,
3936 bool compute_self_and_rr)
3938 struct data_dependence_relation *ddr;
3939 struct data_reference *a, *b;
3942 for (i = 0; VEC_iterate (data_reference_p, datarefs, i, a); i++)
3943 for (j = i + 1; VEC_iterate (data_reference_p, datarefs, j, b); j++)
3944 if (!DR_IS_READ (a) || !DR_IS_READ (b) || compute_self_and_rr)
3946 ddr = initialize_data_dependence_relation (a, b, loop_nest);
3947 VEC_safe_push (ddr_p, heap, *dependence_relations, ddr);
3948 compute_affine_dependence (ddr, VEC_index (loop_p, loop_nest, 0));
3951 if (compute_self_and_rr)
3952 for (i = 0; VEC_iterate (data_reference_p, datarefs, i, a); i++)
3954 ddr = initialize_data_dependence_relation (a, a, loop_nest);
3955 VEC_safe_push (ddr_p, heap, *dependence_relations, ddr);
3956 compute_self_dependence (ddr);
3960 /* Stores the locations of memory references in STMT to REFERENCES. Returns
3961 true if STMT clobbers memory, false otherwise. */
3964 get_references_in_stmt (tree stmt, VEC (data_ref_loc, heap) **references)
3966 bool clobbers_memory = false;
3968 tree *op0, *op1, call;
3972 /* ASM_EXPR and CALL_EXPR may embed arbitrary side effects.
3973 Calls have side-effects, except those to const or pure
3975 call = get_call_expr_in (stmt);
3977 && !(call_expr_flags (call) & (ECF_CONST | ECF_PURE)))
3978 || (TREE_CODE (stmt) == ASM_EXPR
3979 && ASM_VOLATILE_P (stmt)))
3980 clobbers_memory = true;
3982 if (ZERO_SSA_OPERANDS (stmt, SSA_OP_ALL_VIRTUALS))
3983 return clobbers_memory;
3985 if (TREE_CODE (stmt) == GIMPLE_MODIFY_STMT)
3988 op0 = &GIMPLE_STMT_OPERAND (stmt, 0);
3989 op1 = &GIMPLE_STMT_OPERAND (stmt, 1);
3992 || (REFERENCE_CLASS_P (*op1)
3993 && (base = get_base_address (*op1))
3994 && TREE_CODE (base) != SSA_NAME))
3996 ref = VEC_safe_push (data_ref_loc, heap, *references, NULL);
3998 ref->is_read = true;
4002 || (REFERENCE_CLASS_P (*op0) && get_base_address (*op0)))
4004 ref = VEC_safe_push (data_ref_loc, heap, *references, NULL);
4006 ref->is_read = false;
4012 unsigned i, n = call_expr_nargs (call);
4014 for (i = 0; i < n; i++)
4016 op0 = &CALL_EXPR_ARG (call, i);
4019 || (REFERENCE_CLASS_P (*op0) && get_base_address (*op0)))
4021 ref = VEC_safe_push (data_ref_loc, heap, *references, NULL);
4023 ref->is_read = true;
4028 return clobbers_memory;
4031 /* Stores the data references in STMT to DATAREFS. If there is an unanalyzable
4032 reference, returns false, otherwise returns true. NEST is the outermost
4033 loop of the loop nest in that the references should be analyzed. */
4036 find_data_references_in_stmt (struct loop *nest, tree stmt,
4037 VEC (data_reference_p, heap) **datarefs)
4040 VEC (data_ref_loc, heap) *references;
4043 data_reference_p dr;
4045 if (get_references_in_stmt (stmt, &references))
4047 VEC_free (data_ref_loc, heap, references);
4051 for (i = 0; VEC_iterate (data_ref_loc, references, i, ref); i++)
4053 dr = create_data_ref (nest, *ref->pos, stmt, ref->is_read);
4054 gcc_assert (dr != NULL);
4056 /* FIXME -- data dependence analysis does not work correctly for objects with
4057 invariant addresses. Let us fail here until the problem is fixed. */
4058 if (dr_address_invariant_p (dr))
4061 if (dump_file && (dump_flags & TDF_DETAILS))
4062 fprintf (dump_file, "\tFAILED as dr address is invariant\n");
4067 VEC_safe_push (data_reference_p, heap, *datarefs, dr);
4069 VEC_free (data_ref_loc, heap, references);
4073 /* Search the data references in LOOP, and record the information into
4074 DATAREFS. Returns chrec_dont_know when failing to analyze a
4075 difficult case, returns NULL_TREE otherwise.
4077 TODO: This function should be made smarter so that it can handle address
4078 arithmetic as if they were array accesses, etc. */
4081 find_data_references_in_loop (struct loop *loop,
4082 VEC (data_reference_p, heap) **datarefs)
4084 basic_block bb, *bbs;
4086 block_stmt_iterator bsi;
4088 bbs = get_loop_body_in_dom_order (loop);
4090 for (i = 0; i < loop->num_nodes; i++)
4094 for (bsi = bsi_start (bb); !bsi_end_p (bsi); bsi_next (&bsi))
4096 tree stmt = bsi_stmt (bsi);
4098 if (!find_data_references_in_stmt (loop, stmt, datarefs))
4100 struct data_reference *res;
4101 res = XCNEW (struct data_reference);
4102 VEC_safe_push (data_reference_p, heap, *datarefs, res);
4105 return chrec_dont_know;
4114 /* Recursive helper function. */
4117 find_loop_nest_1 (struct loop *loop, VEC (loop_p, heap) **loop_nest)
4119 /* Inner loops of the nest should not contain siblings. Example:
4120 when there are two consecutive loops,
4131 the dependence relation cannot be captured by the distance
4136 VEC_safe_push (loop_p, heap, *loop_nest, loop);
4138 return find_loop_nest_1 (loop->inner, loop_nest);
4142 /* Return false when the LOOP is not well nested. Otherwise return
4143 true and insert in LOOP_NEST the loops of the nest. LOOP_NEST will
4144 contain the loops from the outermost to the innermost, as they will
4145 appear in the classic distance vector. */
4148 find_loop_nest (struct loop *loop, VEC (loop_p, heap) **loop_nest)
4150 VEC_safe_push (loop_p, heap, *loop_nest, loop);
4152 return find_loop_nest_1 (loop->inner, loop_nest);
4156 /* Given a loop nest LOOP, the following vectors are returned:
4157 DATAREFS is initialized to all the array elements contained in this loop,
4158 DEPENDENCE_RELATIONS contains the relations between the data references.
4159 Compute read-read and self relations if
4160 COMPUTE_SELF_AND_READ_READ_DEPENDENCES is TRUE. */
4163 compute_data_dependences_for_loop (struct loop *loop,
4164 bool compute_self_and_read_read_dependences,
4165 VEC (data_reference_p, heap) **datarefs,
4166 VEC (ddr_p, heap) **dependence_relations)
4168 VEC (loop_p, heap) *vloops = VEC_alloc (loop_p, heap, 3);
4170 memset (&dependence_stats, 0, sizeof (dependence_stats));
4172 /* If the loop nest is not well formed, or one of the data references
4173 is not computable, give up without spending time to compute other
4176 || !find_loop_nest (loop, &vloops)
4177 || find_data_references_in_loop (loop, datarefs) == chrec_dont_know)
4179 struct data_dependence_relation *ddr;
4181 /* Insert a single relation into dependence_relations:
4183 ddr = initialize_data_dependence_relation (NULL, NULL, vloops);
4184 VEC_safe_push (ddr_p, heap, *dependence_relations, ddr);
4187 compute_all_dependences (*datarefs, dependence_relations, vloops,
4188 compute_self_and_read_read_dependences);
4190 if (dump_file && (dump_flags & TDF_STATS))
4192 fprintf (dump_file, "Dependence tester statistics:\n");
4194 fprintf (dump_file, "Number of dependence tests: %d\n",
4195 dependence_stats.num_dependence_tests);
4196 fprintf (dump_file, "Number of dependence tests classified dependent: %d\n",
4197 dependence_stats.num_dependence_dependent);
4198 fprintf (dump_file, "Number of dependence tests classified independent: %d\n",
4199 dependence_stats.num_dependence_independent);
4200 fprintf (dump_file, "Number of undetermined dependence tests: %d\n",
4201 dependence_stats.num_dependence_undetermined);
4203 fprintf (dump_file, "Number of subscript tests: %d\n",
4204 dependence_stats.num_subscript_tests);
4205 fprintf (dump_file, "Number of undetermined subscript tests: %d\n",
4206 dependence_stats.num_subscript_undetermined);
4207 fprintf (dump_file, "Number of same subscript function: %d\n",
4208 dependence_stats.num_same_subscript_function);
4210 fprintf (dump_file, "Number of ziv tests: %d\n",
4211 dependence_stats.num_ziv);
4212 fprintf (dump_file, "Number of ziv tests returning dependent: %d\n",
4213 dependence_stats.num_ziv_dependent);
4214 fprintf (dump_file, "Number of ziv tests returning independent: %d\n",
4215 dependence_stats.num_ziv_independent);
4216 fprintf (dump_file, "Number of ziv tests unimplemented: %d\n",
4217 dependence_stats.num_ziv_unimplemented);
4219 fprintf (dump_file, "Number of siv tests: %d\n",
4220 dependence_stats.num_siv);
4221 fprintf (dump_file, "Number of siv tests returning dependent: %d\n",
4222 dependence_stats.num_siv_dependent);
4223 fprintf (dump_file, "Number of siv tests returning independent: %d\n",
4224 dependence_stats.num_siv_independent);
4225 fprintf (dump_file, "Number of siv tests unimplemented: %d\n",
4226 dependence_stats.num_siv_unimplemented);
4228 fprintf (dump_file, "Number of miv tests: %d\n",
4229 dependence_stats.num_miv);
4230 fprintf (dump_file, "Number of miv tests returning dependent: %d\n",
4231 dependence_stats.num_miv_dependent);
4232 fprintf (dump_file, "Number of miv tests returning independent: %d\n",
4233 dependence_stats.num_miv_independent);
4234 fprintf (dump_file, "Number of miv tests unimplemented: %d\n",
4235 dependence_stats.num_miv_unimplemented);
4239 /* Entry point (for testing only). Analyze all the data references
4240 and the dependence relations in LOOP.
4242 The data references are computed first.
4244 A relation on these nodes is represented by a complete graph. Some
4245 of the relations could be of no interest, thus the relations can be
4248 In the following function we compute all the relations. This is
4249 just a first implementation that is here for:
4250 - for showing how to ask for the dependence relations,
4251 - for the debugging the whole dependence graph,
4252 - for the dejagnu testcases and maintenance.
4254 It is possible to ask only for a part of the graph, avoiding to
4255 compute the whole dependence graph. The computed dependences are
4256 stored in a knowledge base (KB) such that later queries don't
4257 recompute the same information. The implementation of this KB is
4258 transparent to the optimizer, and thus the KB can be changed with a
4259 more efficient implementation, or the KB could be disabled. */
4261 analyze_all_data_dependences (struct loop *loop)
4264 int nb_data_refs = 10;
4265 VEC (data_reference_p, heap) *datarefs =
4266 VEC_alloc (data_reference_p, heap, nb_data_refs);
4267 VEC (ddr_p, heap) *dependence_relations =
4268 VEC_alloc (ddr_p, heap, nb_data_refs * nb_data_refs);
4270 /* Compute DDs on the whole function. */
4271 compute_data_dependences_for_loop (loop, false, &datarefs,
4272 &dependence_relations);
4276 dump_data_dependence_relations (dump_file, dependence_relations);
4277 fprintf (dump_file, "\n\n");
4279 if (dump_flags & TDF_DETAILS)
4280 dump_dist_dir_vectors (dump_file, dependence_relations);
4282 if (dump_flags & TDF_STATS)
4284 unsigned nb_top_relations = 0;
4285 unsigned nb_bot_relations = 0;
4286 unsigned nb_basename_differ = 0;
4287 unsigned nb_chrec_relations = 0;
4288 struct data_dependence_relation *ddr;
4290 for (i = 0; VEC_iterate (ddr_p, dependence_relations, i, ddr); i++)
4292 if (chrec_contains_undetermined (DDR_ARE_DEPENDENT (ddr)))
4295 else if (DDR_ARE_DEPENDENT (ddr) == chrec_known)
4297 struct data_reference *a = DDR_A (ddr);
4298 struct data_reference *b = DDR_B (ddr);
4300 if (!bitmap_intersect_p (DR_VOPS (a), DR_VOPS (b)))
4301 nb_basename_differ++;
4307 nb_chrec_relations++;
4310 gather_stats_on_scev_database ();
4314 free_dependence_relations (dependence_relations);
4315 free_data_refs (datarefs);
4318 /* Computes all the data dependences and check that the results of
4319 several analyzers are the same. */
4322 tree_check_data_deps (void)
4325 struct loop *loop_nest;
4327 FOR_EACH_LOOP (li, loop_nest, 0)
4328 analyze_all_data_dependences (loop_nest);
4331 /* Free the memory used by a data dependence relation DDR. */
4334 free_dependence_relation (struct data_dependence_relation *ddr)
4339 if (DDR_SUBSCRIPTS (ddr))
4340 free_subscripts (DDR_SUBSCRIPTS (ddr));
4341 if (DDR_DIST_VECTS (ddr))
4342 VEC_free (lambda_vector, heap, DDR_DIST_VECTS (ddr));
4343 if (DDR_DIR_VECTS (ddr))
4344 VEC_free (lambda_vector, heap, DDR_DIR_VECTS (ddr));
4349 /* Free the memory used by the data dependence relations from
4350 DEPENDENCE_RELATIONS. */
4353 free_dependence_relations (VEC (ddr_p, heap) *dependence_relations)
4356 struct data_dependence_relation *ddr;
4357 VEC (loop_p, heap) *loop_nest = NULL;
4359 for (i = 0; VEC_iterate (ddr_p, dependence_relations, i, ddr); i++)
4363 if (loop_nest == NULL)
4364 loop_nest = DDR_LOOP_NEST (ddr);
4366 gcc_assert (DDR_LOOP_NEST (ddr) == NULL
4367 || DDR_LOOP_NEST (ddr) == loop_nest);
4368 free_dependence_relation (ddr);
4372 VEC_free (loop_p, heap, loop_nest);
4373 VEC_free (ddr_p, heap, dependence_relations);
4376 /* Free the memory used by the data references from DATAREFS. */
4379 free_data_refs (VEC (data_reference_p, heap) *datarefs)
4382 struct data_reference *dr;
4384 for (i = 0; VEC_iterate (data_reference_p, datarefs, i, dr); i++)
4386 VEC_free (data_reference_p, heap, datarefs);
4391 /* Dump vertex I in RDG to FILE. */
4394 dump_rdg_vertex (FILE *file, struct graph *rdg, int i)
4396 struct vertex *v = &(rdg->vertices[i]);
4397 struct graph_edge *e;
4399 fprintf (file, "(vertex %d: (%s%s) (in:", i,
4400 RDG_MEM_WRITE_STMT (rdg, i) ? "w" : "",
4401 RDG_MEM_READS_STMT (rdg, i) ? "r" : "");
4404 for (e = v->pred; e; e = e->pred_next)
4405 fprintf (file, " %d", e->src);
4407 fprintf (file, ") (out:");
4410 for (e = v->succ; e; e = e->succ_next)
4411 fprintf (file, " %d", e->dest);
4413 fprintf (file, ") \n");
4414 print_generic_stmt (file, RDGV_STMT (v), TDF_VOPS|TDF_MEMSYMS);
4415 fprintf (file, ")\n");
4418 /* Call dump_rdg_vertex on stderr. */
4421 debug_rdg_vertex (struct graph *rdg, int i)
4423 dump_rdg_vertex (stderr, rdg, i);
4426 /* Dump component C of RDG to FILE. If DUMPED is non-null, set the
4427 dumped vertices to that bitmap. */
4429 void dump_rdg_component (FILE *file, struct graph *rdg, int c, bitmap dumped)
4433 fprintf (file, "(%d\n", c);
4435 for (i = 0; i < rdg->n_vertices; i++)
4436 if (rdg->vertices[i].component == c)
4439 bitmap_set_bit (dumped, i);
4441 dump_rdg_vertex (file, rdg, i);
4444 fprintf (file, ")\n");
4447 /* Call dump_rdg_vertex on stderr. */
4450 debug_rdg_component (struct graph *rdg, int c)
4452 dump_rdg_component (stderr, rdg, c, NULL);
4455 /* Dump the reduced dependence graph RDG to FILE. */
4458 dump_rdg (FILE *file, struct graph *rdg)
4461 bitmap dumped = BITMAP_ALLOC (NULL);
4463 fprintf (file, "(rdg\n");
4465 for (i = 0; i < rdg->n_vertices; i++)
4466 if (!bitmap_bit_p (dumped, i))
4467 dump_rdg_component (file, rdg, rdg->vertices[i].component, dumped);
4469 fprintf (file, ")\n");
4470 BITMAP_FREE (dumped);
4473 /* Call dump_rdg on stderr. */
4476 debug_rdg (struct graph *rdg)
4478 dump_rdg (stderr, rdg);
4482 dot_rdg_1 (FILE *file, struct graph *rdg)
4486 fprintf (file, "digraph RDG {\n");
4488 for (i = 0; i < rdg->n_vertices; i++)
4490 struct vertex *v = &(rdg->vertices[i]);
4491 struct graph_edge *e;
4493 /* Highlight reads from memory. */
4494 if (RDG_MEM_READS_STMT (rdg, i))
4495 fprintf (file, "%d [style=filled, fillcolor=green]\n", i);
4497 /* Highlight stores to memory. */
4498 if (RDG_MEM_WRITE_STMT (rdg, i))
4499 fprintf (file, "%d [style=filled, fillcolor=red]\n", i);
4502 for (e = v->succ; e; e = e->succ_next)
4503 switch (RDGE_TYPE (e))
4506 fprintf (file, "%d -> %d [label=input] \n", i, e->dest);
4510 fprintf (file, "%d -> %d [label=output] \n", i, e->dest);
4514 /* These are the most common dependences: don't print these. */
4515 fprintf (file, "%d -> %d \n", i, e->dest);
4519 fprintf (file, "%d -> %d [label=anti] \n", i, e->dest);
4527 fprintf (file, "}\n\n");
4530 /* Display SCOP using dotty. */
4533 dot_rdg (struct graph *rdg)
4535 FILE *file = fopen ("/tmp/rdg.dot", "w");
4536 gcc_assert (file != NULL);
4538 dot_rdg_1 (file, rdg);
4541 system ("dotty /tmp/rdg.dot");
4545 /* This structure is used for recording the mapping statement index in
4548 struct rdg_vertex_info GTY(())
4554 /* Returns the index of STMT in RDG. */
4557 rdg_vertex_for_stmt (struct graph *rdg, tree stmt)
4559 struct rdg_vertex_info rvi, *slot;
4562 slot = (struct rdg_vertex_info *) htab_find (rdg->indices, &rvi);
4570 /* Creates an edge in RDG for each distance vector from DDR. The
4571 order that we keep track of in the RDG is the order in which
4572 statements have to be executed. */
4575 create_rdg_edge_for_ddr (struct graph *rdg, ddr_p ddr)
4577 struct graph_edge *e;
4579 data_reference_p dra = DDR_A (ddr);
4580 data_reference_p drb = DDR_B (ddr);
4581 unsigned level = ddr_dependence_level (ddr);
4583 /* For non scalar dependences, when the dependence is REVERSED,
4584 statement B has to be executed before statement A. */
4586 && !DDR_REVERSED_P (ddr))
4588 data_reference_p tmp = dra;
4593 va = rdg_vertex_for_stmt (rdg, DR_STMT (dra));
4594 vb = rdg_vertex_for_stmt (rdg, DR_STMT (drb));
4596 if (va < 0 || vb < 0)
4599 e = add_edge (rdg, va, vb);
4600 e->data = XNEW (struct rdg_edge);
4602 RDGE_LEVEL (e) = level;
4604 /* Determines the type of the data dependence. */
4605 if (DR_IS_READ (dra) && DR_IS_READ (drb))
4606 RDGE_TYPE (e) = input_dd;
4607 else if (!DR_IS_READ (dra) && !DR_IS_READ (drb))
4608 RDGE_TYPE (e) = output_dd;
4609 else if (!DR_IS_READ (dra) && DR_IS_READ (drb))
4610 RDGE_TYPE (e) = flow_dd;
4611 else if (DR_IS_READ (dra) && !DR_IS_READ (drb))
4612 RDGE_TYPE (e) = anti_dd;
4615 /* Creates dependence edges in RDG for all the uses of DEF. IDEF is
4616 the index of DEF in RDG. */
4619 create_rdg_edges_for_scalar (struct graph *rdg, tree def, int idef)
4621 use_operand_p imm_use_p;
4622 imm_use_iterator iterator;
4624 FOR_EACH_IMM_USE_FAST (imm_use_p, iterator, def)
4626 struct graph_edge *e;
4627 int use = rdg_vertex_for_stmt (rdg, USE_STMT (imm_use_p));
4632 e = add_edge (rdg, idef, use);
4633 e->data = XNEW (struct rdg_edge);
4634 RDGE_TYPE (e) = flow_dd;
4638 /* Creates the edges of the reduced dependence graph RDG. */
4641 create_rdg_edges (struct graph *rdg, VEC (ddr_p, heap) *ddrs)
4644 struct data_dependence_relation *ddr;
4645 def_operand_p def_p;
4648 for (i = 0; VEC_iterate (ddr_p, ddrs, i, ddr); i++)
4649 if (DDR_ARE_DEPENDENT (ddr) == NULL_TREE)
4650 create_rdg_edge_for_ddr (rdg, ddr);
4652 for (i = 0; i < rdg->n_vertices; i++)
4653 FOR_EACH_PHI_OR_STMT_DEF (def_p, RDG_STMT (rdg, i),
4655 create_rdg_edges_for_scalar (rdg, DEF_FROM_PTR (def_p), i);
4658 /* Build the vertices of the reduced dependence graph RDG. */
4661 create_rdg_vertices (struct graph *rdg, VEC (tree, heap) *stmts)
4666 for (i = 0; VEC_iterate (tree, stmts, i, stmt); i++)
4668 VEC (data_ref_loc, heap) *references;
4670 struct vertex *v = &(rdg->vertices[i]);
4671 struct rdg_vertex_info *rvi = XNEW (struct rdg_vertex_info);
4672 struct rdg_vertex_info **slot;
4676 slot = (struct rdg_vertex_info **) htab_find_slot (rdg->indices, rvi, INSERT);
4683 v->data = XNEW (struct rdg_vertex);
4684 RDG_STMT (rdg, i) = stmt;
4686 RDG_MEM_WRITE_STMT (rdg, i) = false;
4687 RDG_MEM_READS_STMT (rdg, i) = false;
4688 if (TREE_CODE (stmt) == PHI_NODE)
4691 get_references_in_stmt (stmt, &references);
4692 for (j = 0; VEC_iterate (data_ref_loc, references, j, ref); j++)
4694 RDG_MEM_WRITE_STMT (rdg, i) = true;
4696 RDG_MEM_READS_STMT (rdg, i) = true;
4698 VEC_free (data_ref_loc, heap, references);
4702 /* Initialize STMTS with all the statements of LOOP. When
4703 INCLUDE_PHIS is true, include also the PHI nodes. The order in
4704 which we discover statements is important as
4705 generate_loops_for_partition is using the same traversal for
4706 identifying statements. */
4709 stmts_from_loop (struct loop *loop, VEC (tree, heap) **stmts)
4712 basic_block *bbs = get_loop_body_in_dom_order (loop);
4714 for (i = 0; i < loop->num_nodes; i++)
4717 basic_block bb = bbs[i];
4718 block_stmt_iterator bsi;
4720 for (phi = phi_nodes (bb); phi; phi = PHI_CHAIN (phi))
4721 VEC_safe_push (tree, heap, *stmts, phi);
4723 for (bsi = bsi_start (bb); !bsi_end_p (bsi); bsi_next (&bsi))
4724 if (TREE_CODE (stmt = bsi_stmt (bsi)) != LABEL_EXPR)
4725 VEC_safe_push (tree, heap, *stmts, stmt);
4731 /* Returns true when all the dependences are computable. */
4734 known_dependences_p (VEC (ddr_p, heap) *dependence_relations)
4739 for (i = 0; VEC_iterate (ddr_p, dependence_relations, i, ddr); i++)
4740 if (DDR_ARE_DEPENDENT (ddr) == chrec_dont_know)
4746 /* Computes a hash function for element ELT. */
4749 hash_stmt_vertex_info (const void *elt)
4751 struct rdg_vertex_info *rvi = (struct rdg_vertex_info *) elt;
4752 tree stmt = rvi->stmt;
4754 return htab_hash_pointer (stmt);
4757 /* Compares database elements E1 and E2. */
4760 eq_stmt_vertex_info (const void *e1, const void *e2)
4762 const struct rdg_vertex_info *elt1 = (const struct rdg_vertex_info *) e1;
4763 const struct rdg_vertex_info *elt2 = (const struct rdg_vertex_info *) e2;
4765 return elt1->stmt == elt2->stmt;
4768 /* Free the element E. */
4771 hash_stmt_vertex_del (void *e)
4776 /* Build the Reduced Dependence Graph (RDG) with one vertex per
4777 statement of the loop nest, and one edge per data dependence or
4778 scalar dependence. */
4781 build_rdg (struct loop *loop)
4783 int nb_data_refs = 10;
4784 struct graph *rdg = NULL;
4785 VEC (ddr_p, heap) *dependence_relations;
4786 VEC (data_reference_p, heap) *datarefs;
4787 VEC (tree, heap) *stmts = VEC_alloc (tree, heap, nb_data_refs);
4789 dependence_relations = VEC_alloc (ddr_p, heap, nb_data_refs * nb_data_refs) ;
4790 datarefs = VEC_alloc (data_reference_p, heap, nb_data_refs);
4791 compute_data_dependences_for_loop (loop,
4794 &dependence_relations);
4796 if (!known_dependences_p (dependence_relations))
4799 stmts_from_loop (loop, &stmts);
4800 rdg = new_graph (VEC_length (tree, stmts));
4802 rdg->indices = htab_create (nb_data_refs, hash_stmt_vertex_info,
4803 eq_stmt_vertex_info, hash_stmt_vertex_del);
4804 create_rdg_vertices (rdg, stmts);
4805 create_rdg_edges (rdg, dependence_relations);
4808 free_dependence_relations (dependence_relations);
4809 free_data_refs (datarefs);
4810 VEC_free (tree, heap, stmts);
4815 /* Free the reduced dependence graph RDG. */
4818 free_rdg (struct graph *rdg)
4822 for (i = 0; i < rdg->n_vertices; i++)
4823 free (rdg->vertices[i].data);
4825 htab_delete (rdg->indices);
4829 /* Initialize STMTS with all the statements of LOOP that contain a
4833 stores_from_loop (struct loop *loop, VEC (tree, heap) **stmts)
4836 basic_block *bbs = get_loop_body_in_dom_order (loop);
4838 for (i = 0; i < loop->num_nodes; i++)
4840 basic_block bb = bbs[i];
4841 block_stmt_iterator bsi;
4843 for (bsi = bsi_start (bb); !bsi_end_p (bsi); bsi_next (&bsi))
4844 if (!ZERO_SSA_OPERANDS (bsi_stmt (bsi), SSA_OP_VDEF))
4845 VEC_safe_push (tree, heap, *stmts, bsi_stmt (bsi));
4851 /* For a data reference REF, return the declaration of its base
4852 address or NULL_TREE if the base is not determined. */
4855 ref_base_address (tree stmt, data_ref_loc *ref)
4857 tree base = NULL_TREE;
4859 struct data_reference *dr = XCNEW (struct data_reference);
4861 DR_STMT (dr) = stmt;
4862 DR_REF (dr) = *ref->pos;
4863 dr_analyze_innermost (dr);
4864 base_address = DR_BASE_ADDRESS (dr);
4869 switch (TREE_CODE (base_address))
4872 base = TREE_OPERAND (base_address, 0);
4876 base = base_address;
4885 /* Determines whether the statement from vertex V of the RDG has a
4886 definition used outside the loop that contains this statement. */
4889 rdg_defs_used_in_other_loops_p (struct graph *rdg, int v)
4891 tree stmt = RDG_STMT (rdg, v);
4892 struct loop *loop = loop_containing_stmt (stmt);
4893 use_operand_p imm_use_p;
4894 imm_use_iterator iterator;
4896 def_operand_p def_p;
4901 FOR_EACH_PHI_OR_STMT_DEF (def_p, stmt, it, SSA_OP_DEF)
4903 FOR_EACH_IMM_USE_FAST (imm_use_p, iterator, DEF_FROM_PTR (def_p))
4905 if (loop_containing_stmt (USE_STMT (imm_use_p)) != loop)
4913 /* Determines whether statements S1 and S2 access to similar memory
4914 locations. Two memory accesses are considered similar when they
4915 have the same base address declaration, i.e. when their
4916 ref_base_address is the same. */
4919 have_similar_memory_accesses (tree s1, tree s2)
4923 VEC (data_ref_loc, heap) *refs1, *refs2;
4924 data_ref_loc *ref1, *ref2;
4926 get_references_in_stmt (s1, &refs1);
4927 get_references_in_stmt (s2, &refs2);
4929 for (i = 0; VEC_iterate (data_ref_loc, refs1, i, ref1); i++)
4931 tree base1 = ref_base_address (s1, ref1);
4934 for (j = 0; VEC_iterate (data_ref_loc, refs2, j, ref2); j++)
4935 if (base1 == ref_base_address (s2, ref2))
4943 VEC_free (data_ref_loc, heap, refs1);
4944 VEC_free (data_ref_loc, heap, refs2);
4948 /* Helper function for the hashtab. */
4951 have_similar_memory_accesses_1 (const void *s1, const void *s2)
4953 return have_similar_memory_accesses ((tree) s1, (tree) s2);
4956 /* Helper function for the hashtab. */
4959 ref_base_address_1 (const void *s)
4961 tree stmt = (tree) s;
4963 VEC (data_ref_loc, heap) *refs;
4967 get_references_in_stmt (stmt, &refs);
4969 for (i = 0; VEC_iterate (data_ref_loc, refs, i, ref); i++)
4972 res = htab_hash_pointer (ref_base_address (stmt, ref));
4976 VEC_free (data_ref_loc, heap, refs);
4980 /* Try to remove duplicated write data references from STMTS. */
4983 remove_similar_memory_refs (VEC (tree, heap) **stmts)
4987 htab_t seen = htab_create (VEC_length (tree, *stmts), ref_base_address_1,
4988 have_similar_memory_accesses_1, NULL);
4990 for (i = 0; VEC_iterate (tree, *stmts, i, stmt); )
4994 slot = htab_find_slot (seen, stmt, INSERT);
4997 VEC_ordered_remove (tree, *stmts, i);
5000 *slot = (void *) stmt;