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_gimple_stmt (outf, DR_STMT (dr), 0, 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 /* Helper function for split_constant_offset. Expresses OP0 CODE OP1
504 (the type of the result is TYPE) as VAR + OFF, where OFF is a nonzero
505 constant of type ssizetype, and returns true. If we cannot do this
506 with OFF nonzero, OFF and VAR are set to NULL_TREE instead and false
510 split_constant_offset_1 (tree type, tree op0, enum tree_code code, tree op1,
511 tree *var, tree *off)
515 enum tree_code ocode = code;
523 *var = build_int_cst (type, 0);
524 *off = fold_convert (ssizetype, op0);
527 case POINTER_PLUS_EXPR:
532 split_constant_offset (op0, &var0, &off0);
533 split_constant_offset (op1, &var1, &off1);
534 *var = fold_build2 (code, type, var0, var1);
535 *off = size_binop (ocode, off0, off1);
539 if (TREE_CODE (op1) != INTEGER_CST)
542 split_constant_offset (op0, &var0, &off0);
543 *var = fold_build2 (MULT_EXPR, type, var0, op1);
544 *off = size_binop (MULT_EXPR, off0, fold_convert (ssizetype, op1));
550 HOST_WIDE_INT pbitsize, pbitpos;
551 enum machine_mode pmode;
552 int punsignedp, pvolatilep;
554 op0 = TREE_OPERAND (op0, 0);
555 if (!handled_component_p (op0))
558 base = get_inner_reference (op0, &pbitsize, &pbitpos, &poffset,
559 &pmode, &punsignedp, &pvolatilep, false);
561 if (pbitpos % BITS_PER_UNIT != 0)
563 base = build_fold_addr_expr (base);
564 off0 = ssize_int (pbitpos / BITS_PER_UNIT);
568 split_constant_offset (poffset, &poffset, &off1);
569 off0 = size_binop (PLUS_EXPR, off0, off1);
570 if (POINTER_TYPE_P (TREE_TYPE (base)))
571 base = fold_build2 (POINTER_PLUS_EXPR, TREE_TYPE (base),
572 base, fold_convert (sizetype, poffset));
574 base = fold_build2 (PLUS_EXPR, TREE_TYPE (base), base,
575 fold_convert (TREE_TYPE (base), poffset));
578 var0 = fold_convert (type, base);
580 /* If variable length types are involved, punt, otherwise casts
581 might be converted into ARRAY_REFs in gimplify_conversion.
582 To compute that ARRAY_REF's element size TYPE_SIZE_UNIT, which
583 possibly no longer appears in current GIMPLE, might resurface.
584 This perhaps could run
585 if (CONVERT_EXPR_P (var0))
587 gimplify_conversion (&var0);
588 // Attempt to fill in any within var0 found ARRAY_REF's
589 // element size from corresponding op embedded ARRAY_REF,
590 // if unsuccessful, just punt.
592 while (POINTER_TYPE_P (type))
593 type = TREE_TYPE (type);
594 if (int_size_in_bytes (type) < 0)
604 gimple def_stmt = SSA_NAME_DEF_STMT (op0);
605 enum tree_code subcode;
607 if (gimple_code (def_stmt) != GIMPLE_ASSIGN)
610 var0 = gimple_assign_rhs1 (def_stmt);
611 subcode = gimple_assign_rhs_code (def_stmt);
612 var1 = gimple_assign_rhs2 (def_stmt);
614 return split_constant_offset_1 (type, var0, subcode, var1, var, off);
622 /* Expresses EXP as VAR + OFF, where off is a constant. The type of OFF
623 will be ssizetype. */
626 split_constant_offset (tree exp, tree *var, tree *off)
628 tree type = TREE_TYPE (exp), otype, op0, op1, e, o;
632 *off = ssize_int (0);
635 if (automatically_generated_chrec_p (exp))
638 otype = TREE_TYPE (exp);
639 code = TREE_CODE (exp);
640 extract_ops_from_tree (exp, &code, &op0, &op1);
641 if (split_constant_offset_1 (otype, op0, code, op1, &e, &o))
643 *var = fold_convert (type, e);
648 /* Returns the address ADDR of an object in a canonical shape (without nop
649 casts, and with type of pointer to the object). */
652 canonicalize_base_object_address (tree addr)
658 /* The base address may be obtained by casting from integer, in that case
660 if (!POINTER_TYPE_P (TREE_TYPE (addr)))
663 if (TREE_CODE (addr) != ADDR_EXPR)
666 return build_fold_addr_expr (TREE_OPERAND (addr, 0));
669 /* Analyzes the behavior of the memory reference DR in the innermost loop that
670 contains it. Returns true if analysis succeed or false otherwise. */
673 dr_analyze_innermost (struct data_reference *dr)
675 gimple stmt = DR_STMT (dr);
676 struct loop *loop = loop_containing_stmt (stmt);
677 tree ref = DR_REF (dr);
678 HOST_WIDE_INT pbitsize, pbitpos;
680 enum machine_mode pmode;
681 int punsignedp, pvolatilep;
682 affine_iv base_iv, offset_iv;
683 tree init, dinit, step;
685 if (dump_file && (dump_flags & TDF_DETAILS))
686 fprintf (dump_file, "analyze_innermost: ");
688 base = get_inner_reference (ref, &pbitsize, &pbitpos, &poffset,
689 &pmode, &punsignedp, &pvolatilep, false);
690 gcc_assert (base != NULL_TREE);
692 if (pbitpos % BITS_PER_UNIT != 0)
694 if (dump_file && (dump_flags & TDF_DETAILS))
695 fprintf (dump_file, "failed: bit offset alignment.\n");
699 base = build_fold_addr_expr (base);
700 if (!simple_iv (loop, stmt, base, &base_iv, false))
702 if (dump_file && (dump_flags & TDF_DETAILS))
703 fprintf (dump_file, "failed: evolution of base is not affine.\n");
708 offset_iv.base = ssize_int (0);
709 offset_iv.step = ssize_int (0);
711 else if (!simple_iv (loop, stmt, poffset, &offset_iv, false))
713 if (dump_file && (dump_flags & TDF_DETAILS))
714 fprintf (dump_file, "failed: evolution of offset is not affine.\n");
718 init = ssize_int (pbitpos / BITS_PER_UNIT);
719 split_constant_offset (base_iv.base, &base_iv.base, &dinit);
720 init = size_binop (PLUS_EXPR, init, dinit);
721 split_constant_offset (offset_iv.base, &offset_iv.base, &dinit);
722 init = size_binop (PLUS_EXPR, init, dinit);
724 step = size_binop (PLUS_EXPR,
725 fold_convert (ssizetype, base_iv.step),
726 fold_convert (ssizetype, offset_iv.step));
728 DR_BASE_ADDRESS (dr) = canonicalize_base_object_address (base_iv.base);
730 DR_OFFSET (dr) = fold_convert (ssizetype, offset_iv.base);
734 DR_ALIGNED_TO (dr) = size_int (highest_pow2_factor (offset_iv.base));
736 if (dump_file && (dump_flags & TDF_DETAILS))
737 fprintf (dump_file, "success.\n");
742 /* Determines the base object and the list of indices of memory reference
743 DR, analyzed in loop nest NEST. */
746 dr_analyze_indices (struct data_reference *dr, struct loop *nest)
748 gimple stmt = DR_STMT (dr);
749 struct loop *loop = loop_containing_stmt (stmt);
750 VEC (tree, heap) *access_fns = NULL;
751 tree ref = unshare_expr (DR_REF (dr)), aref = ref, op;
752 tree base, off, access_fn;
753 basic_block before_loop = block_before_loop (nest);
755 while (handled_component_p (aref))
757 if (TREE_CODE (aref) == ARRAY_REF)
759 op = TREE_OPERAND (aref, 1);
760 access_fn = analyze_scalar_evolution (loop, op);
761 access_fn = instantiate_scev (before_loop, loop, access_fn);
762 VEC_safe_push (tree, heap, access_fns, access_fn);
764 TREE_OPERAND (aref, 1) = build_int_cst (TREE_TYPE (op), 0);
767 aref = TREE_OPERAND (aref, 0);
770 if (INDIRECT_REF_P (aref))
772 op = TREE_OPERAND (aref, 0);
773 access_fn = analyze_scalar_evolution (loop, op);
774 access_fn = instantiate_scev (before_loop, loop, access_fn);
775 base = initial_condition (access_fn);
776 split_constant_offset (base, &base, &off);
777 access_fn = chrec_replace_initial_condition (access_fn,
778 fold_convert (TREE_TYPE (base), off));
780 TREE_OPERAND (aref, 0) = base;
781 VEC_safe_push (tree, heap, access_fns, access_fn);
784 DR_BASE_OBJECT (dr) = ref;
785 DR_ACCESS_FNS (dr) = access_fns;
788 /* Extracts the alias analysis information from the memory reference DR. */
791 dr_analyze_alias (struct data_reference *dr)
793 gimple stmt = DR_STMT (dr);
794 tree ref = DR_REF (dr);
795 tree base = get_base_address (ref), addr, smt = NULL_TREE;
802 else if (INDIRECT_REF_P (base))
804 addr = TREE_OPERAND (base, 0);
805 if (TREE_CODE (addr) == SSA_NAME)
807 smt = symbol_mem_tag (SSA_NAME_VAR (addr));
808 DR_PTR_INFO (dr) = SSA_NAME_PTR_INFO (addr);
812 DR_SYMBOL_TAG (dr) = smt;
814 vops = BITMAP_ALLOC (NULL);
815 FOR_EACH_SSA_TREE_OPERAND (op, stmt, it, SSA_OP_VIRTUAL_USES)
817 bitmap_set_bit (vops, DECL_UID (SSA_NAME_VAR (op)));
823 /* Returns true if the address of DR is invariant. */
826 dr_address_invariant_p (struct data_reference *dr)
831 for (i = 0; VEC_iterate (tree, DR_ACCESS_FNS (dr), i, idx); i++)
832 if (tree_contains_chrecs (idx, NULL))
838 /* Frees data reference DR. */
841 free_data_ref (data_reference_p dr)
843 BITMAP_FREE (DR_VOPS (dr));
844 VEC_free (tree, heap, DR_ACCESS_FNS (dr));
848 /* Analyzes memory reference MEMREF accessed in STMT. The reference
849 is read if IS_READ is true, write otherwise. Returns the
850 data_reference description of MEMREF. NEST is the outermost loop of the
851 loop nest in that the reference should be analyzed. */
853 struct data_reference *
854 create_data_ref (struct loop *nest, tree memref, gimple stmt, bool is_read)
856 struct data_reference *dr;
858 if (dump_file && (dump_flags & TDF_DETAILS))
860 fprintf (dump_file, "Creating dr for ");
861 print_generic_expr (dump_file, memref, TDF_SLIM);
862 fprintf (dump_file, "\n");
865 dr = XCNEW (struct data_reference);
867 DR_REF (dr) = memref;
868 DR_IS_READ (dr) = is_read;
870 dr_analyze_innermost (dr);
871 dr_analyze_indices (dr, nest);
872 dr_analyze_alias (dr);
874 if (dump_file && (dump_flags & TDF_DETAILS))
876 fprintf (dump_file, "\tbase_address: ");
877 print_generic_expr (dump_file, DR_BASE_ADDRESS (dr), TDF_SLIM);
878 fprintf (dump_file, "\n\toffset from base address: ");
879 print_generic_expr (dump_file, DR_OFFSET (dr), TDF_SLIM);
880 fprintf (dump_file, "\n\tconstant offset from base address: ");
881 print_generic_expr (dump_file, DR_INIT (dr), TDF_SLIM);
882 fprintf (dump_file, "\n\tstep: ");
883 print_generic_expr (dump_file, DR_STEP (dr), TDF_SLIM);
884 fprintf (dump_file, "\n\taligned to: ");
885 print_generic_expr (dump_file, DR_ALIGNED_TO (dr), TDF_SLIM);
886 fprintf (dump_file, "\n\tbase_object: ");
887 print_generic_expr (dump_file, DR_BASE_OBJECT (dr), TDF_SLIM);
888 fprintf (dump_file, "\n\tsymbol tag: ");
889 print_generic_expr (dump_file, DR_SYMBOL_TAG (dr), TDF_SLIM);
890 fprintf (dump_file, "\n");
896 /* Returns true if FNA == FNB. */
899 affine_function_equal_p (affine_fn fna, affine_fn fnb)
901 unsigned i, n = VEC_length (tree, fna);
903 if (n != VEC_length (tree, fnb))
906 for (i = 0; i < n; i++)
907 if (!operand_equal_p (VEC_index (tree, fna, i),
908 VEC_index (tree, fnb, i), 0))
914 /* If all the functions in CF are the same, returns one of them,
915 otherwise returns NULL. */
918 common_affine_function (conflict_function *cf)
923 if (!CF_NONTRIVIAL_P (cf))
928 for (i = 1; i < cf->n; i++)
929 if (!affine_function_equal_p (comm, cf->fns[i]))
935 /* Returns the base of the affine function FN. */
938 affine_function_base (affine_fn fn)
940 return VEC_index (tree, fn, 0);
943 /* Returns true if FN is a constant. */
946 affine_function_constant_p (affine_fn fn)
951 for (i = 1; VEC_iterate (tree, fn, i, coef); i++)
952 if (!integer_zerop (coef))
958 /* Returns true if FN is the zero constant function. */
961 affine_function_zero_p (affine_fn fn)
963 return (integer_zerop (affine_function_base (fn))
964 && affine_function_constant_p (fn));
967 /* Returns a signed integer type with the largest precision from TA
971 signed_type_for_types (tree ta, tree tb)
973 if (TYPE_PRECISION (ta) > TYPE_PRECISION (tb))
974 return signed_type_for (ta);
976 return signed_type_for (tb);
979 /* Applies operation OP on affine functions FNA and FNB, and returns the
983 affine_fn_op (enum tree_code op, affine_fn fna, affine_fn fnb)
989 if (VEC_length (tree, fnb) > VEC_length (tree, fna))
991 n = VEC_length (tree, fna);
992 m = VEC_length (tree, fnb);
996 n = VEC_length (tree, fnb);
997 m = VEC_length (tree, fna);
1000 ret = VEC_alloc (tree, heap, m);
1001 for (i = 0; i < n; i++)
1003 tree type = signed_type_for_types (TREE_TYPE (VEC_index (tree, fna, i)),
1004 TREE_TYPE (VEC_index (tree, fnb, i)));
1006 VEC_quick_push (tree, ret,
1007 fold_build2 (op, type,
1008 VEC_index (tree, fna, i),
1009 VEC_index (tree, fnb, i)));
1012 for (; VEC_iterate (tree, fna, i, coef); i++)
1013 VEC_quick_push (tree, ret,
1014 fold_build2 (op, signed_type_for (TREE_TYPE (coef)),
1015 coef, integer_zero_node));
1016 for (; VEC_iterate (tree, fnb, i, coef); i++)
1017 VEC_quick_push (tree, ret,
1018 fold_build2 (op, signed_type_for (TREE_TYPE (coef)),
1019 integer_zero_node, coef));
1024 /* Returns the sum of affine functions FNA and FNB. */
1027 affine_fn_plus (affine_fn fna, affine_fn fnb)
1029 return affine_fn_op (PLUS_EXPR, fna, fnb);
1032 /* Returns the difference of affine functions FNA and FNB. */
1035 affine_fn_minus (affine_fn fna, affine_fn fnb)
1037 return affine_fn_op (MINUS_EXPR, fna, fnb);
1040 /* Frees affine function FN. */
1043 affine_fn_free (affine_fn fn)
1045 VEC_free (tree, heap, fn);
1048 /* Determine for each subscript in the data dependence relation DDR
1052 compute_subscript_distance (struct data_dependence_relation *ddr)
1054 conflict_function *cf_a, *cf_b;
1055 affine_fn fn_a, fn_b, diff;
1057 if (DDR_ARE_DEPENDENT (ddr) == NULL_TREE)
1061 for (i = 0; i < DDR_NUM_SUBSCRIPTS (ddr); i++)
1063 struct subscript *subscript;
1065 subscript = DDR_SUBSCRIPT (ddr, i);
1066 cf_a = SUB_CONFLICTS_IN_A (subscript);
1067 cf_b = SUB_CONFLICTS_IN_B (subscript);
1069 fn_a = common_affine_function (cf_a);
1070 fn_b = common_affine_function (cf_b);
1073 SUB_DISTANCE (subscript) = chrec_dont_know;
1076 diff = affine_fn_minus (fn_a, fn_b);
1078 if (affine_function_constant_p (diff))
1079 SUB_DISTANCE (subscript) = affine_function_base (diff);
1081 SUB_DISTANCE (subscript) = chrec_dont_know;
1083 affine_fn_free (diff);
1088 /* Returns the conflict function for "unknown". */
1090 static conflict_function *
1091 conflict_fn_not_known (void)
1093 conflict_function *fn = XCNEW (conflict_function);
1099 /* Returns the conflict function for "independent". */
1101 static conflict_function *
1102 conflict_fn_no_dependence (void)
1104 conflict_function *fn = XCNEW (conflict_function);
1105 fn->n = NO_DEPENDENCE;
1110 /* Returns true if the address of OBJ is invariant in LOOP. */
1113 object_address_invariant_in_loop_p (const struct loop *loop, const_tree obj)
1115 while (handled_component_p (obj))
1117 if (TREE_CODE (obj) == ARRAY_REF)
1119 /* Index of the ARRAY_REF was zeroed in analyze_indices, thus we only
1120 need to check the stride and the lower bound of the reference. */
1121 if (chrec_contains_symbols_defined_in_loop (TREE_OPERAND (obj, 2),
1123 || chrec_contains_symbols_defined_in_loop (TREE_OPERAND (obj, 3),
1127 else if (TREE_CODE (obj) == COMPONENT_REF)
1129 if (chrec_contains_symbols_defined_in_loop (TREE_OPERAND (obj, 2),
1133 obj = TREE_OPERAND (obj, 0);
1136 if (!INDIRECT_REF_P (obj))
1139 return !chrec_contains_symbols_defined_in_loop (TREE_OPERAND (obj, 0),
1143 /* Returns true if A and B are accesses to different objects, or to different
1144 fields of the same object. */
1147 disjoint_objects_p (tree a, tree b)
1149 tree base_a, base_b;
1150 VEC (tree, heap) *comp_a = NULL, *comp_b = NULL;
1153 base_a = get_base_address (a);
1154 base_b = get_base_address (b);
1158 && base_a != base_b)
1161 if (!operand_equal_p (base_a, base_b, 0))
1164 /* Compare the component references of A and B. We must start from the inner
1165 ones, so record them to the vector first. */
1166 while (handled_component_p (a))
1168 VEC_safe_push (tree, heap, comp_a, a);
1169 a = TREE_OPERAND (a, 0);
1171 while (handled_component_p (b))
1173 VEC_safe_push (tree, heap, comp_b, b);
1174 b = TREE_OPERAND (b, 0);
1180 if (VEC_length (tree, comp_a) == 0
1181 || VEC_length (tree, comp_b) == 0)
1184 a = VEC_pop (tree, comp_a);
1185 b = VEC_pop (tree, comp_b);
1187 /* Real and imaginary part of a variable do not alias. */
1188 if ((TREE_CODE (a) == REALPART_EXPR
1189 && TREE_CODE (b) == IMAGPART_EXPR)
1190 || (TREE_CODE (a) == IMAGPART_EXPR
1191 && TREE_CODE (b) == REALPART_EXPR))
1197 if (TREE_CODE (a) != TREE_CODE (b))
1200 /* Nothing to do for ARRAY_REFs, as the indices of array_refs in
1201 DR_BASE_OBJECT are always zero. */
1202 if (TREE_CODE (a) == ARRAY_REF)
1204 else if (TREE_CODE (a) == COMPONENT_REF)
1206 if (operand_equal_p (TREE_OPERAND (a, 1), TREE_OPERAND (b, 1), 0))
1209 /* Different fields of unions may overlap. */
1210 base_a = TREE_OPERAND (a, 0);
1211 if (TREE_CODE (TREE_TYPE (base_a)) == UNION_TYPE)
1214 /* Different fields of structures cannot. */
1222 VEC_free (tree, heap, comp_a);
1223 VEC_free (tree, heap, comp_b);
1228 /* Returns false if we can prove that data references A and B do not alias,
1232 dr_may_alias_p (const struct data_reference *a, const struct data_reference *b)
1234 const_tree addr_a = DR_BASE_ADDRESS (a);
1235 const_tree addr_b = DR_BASE_ADDRESS (b);
1236 const_tree type_a, type_b;
1237 const_tree decl_a = NULL_TREE, decl_b = NULL_TREE;
1239 /* If the sets of virtual operands are disjoint, the memory references do not
1241 if (!bitmap_intersect_p (DR_VOPS (a), DR_VOPS (b)))
1244 /* If the accessed objects are disjoint, the memory references do not
1246 if (disjoint_objects_p (DR_BASE_OBJECT (a), DR_BASE_OBJECT (b)))
1249 if (!addr_a || !addr_b)
1252 /* If the references are based on different static objects, they cannot alias
1253 (PTA should be able to disambiguate such accesses, but often it fails to,
1254 since currently we cannot distinguish between pointer and offset in pointer
1256 if (TREE_CODE (addr_a) == ADDR_EXPR
1257 && TREE_CODE (addr_b) == ADDR_EXPR)
1258 return TREE_OPERAND (addr_a, 0) == TREE_OPERAND (addr_b, 0);
1260 /* An instruction writing through a restricted pointer is "independent" of any
1261 instruction reading or writing through a different restricted pointer,
1262 in the same block/scope. */
1264 type_a = TREE_TYPE (addr_a);
1265 type_b = TREE_TYPE (addr_b);
1266 gcc_assert (POINTER_TYPE_P (type_a) && POINTER_TYPE_P (type_b));
1268 if (TREE_CODE (addr_a) == SSA_NAME)
1269 decl_a = SSA_NAME_VAR (addr_a);
1270 if (TREE_CODE (addr_b) == SSA_NAME)
1271 decl_b = SSA_NAME_VAR (addr_b);
1273 if (TYPE_RESTRICT (type_a) && TYPE_RESTRICT (type_b)
1274 && (!DR_IS_READ (a) || !DR_IS_READ (b))
1275 && decl_a && DECL_P (decl_a)
1276 && decl_b && DECL_P (decl_b)
1278 && TREE_CODE (DECL_CONTEXT (decl_a)) == FUNCTION_DECL
1279 && DECL_CONTEXT (decl_a) == DECL_CONTEXT (decl_b))
1285 static void compute_self_dependence (struct data_dependence_relation *);
1287 /* Initialize a data dependence relation between data accesses A and
1288 B. NB_LOOPS is the number of loops surrounding the references: the
1289 size of the classic distance/direction vectors. */
1291 static struct data_dependence_relation *
1292 initialize_data_dependence_relation (struct data_reference *a,
1293 struct data_reference *b,
1294 VEC (loop_p, heap) *loop_nest)
1296 struct data_dependence_relation *res;
1299 res = XNEW (struct data_dependence_relation);
1302 DDR_LOOP_NEST (res) = NULL;
1303 DDR_REVERSED_P (res) = false;
1304 DDR_SUBSCRIPTS (res) = NULL;
1305 DDR_DIR_VECTS (res) = NULL;
1306 DDR_DIST_VECTS (res) = NULL;
1308 if (a == NULL || b == NULL)
1310 DDR_ARE_DEPENDENT (res) = chrec_dont_know;
1314 /* If the data references do not alias, then they are independent. */
1315 if (!dr_may_alias_p (a, b))
1317 DDR_ARE_DEPENDENT (res) = chrec_known;
1321 /* When the references are exactly the same, don't spend time doing
1322 the data dependence tests, just initialize the ddr and return. */
1323 if (operand_equal_p (DR_REF (a), DR_REF (b), 0))
1325 DDR_AFFINE_P (res) = true;
1326 DDR_ARE_DEPENDENT (res) = NULL_TREE;
1327 DDR_SUBSCRIPTS (res) = VEC_alloc (subscript_p, heap, DR_NUM_DIMENSIONS (a));
1328 DDR_LOOP_NEST (res) = loop_nest;
1329 DDR_INNER_LOOP (res) = 0;
1330 DDR_SELF_REFERENCE (res) = true;
1331 compute_self_dependence (res);
1335 /* If the references do not access the same object, we do not know
1336 whether they alias or not. */
1337 if (!operand_equal_p (DR_BASE_OBJECT (a), DR_BASE_OBJECT (b), 0))
1339 DDR_ARE_DEPENDENT (res) = chrec_dont_know;
1343 /* If the base of the object is not invariant in the loop nest, we cannot
1344 analyze it. TODO -- in fact, it would suffice to record that there may
1345 be arbitrary dependences in the loops where the base object varies. */
1346 if (!object_address_invariant_in_loop_p (VEC_index (loop_p, loop_nest, 0),
1347 DR_BASE_OBJECT (a)))
1349 DDR_ARE_DEPENDENT (res) = chrec_dont_know;
1353 gcc_assert (DR_NUM_DIMENSIONS (a) == DR_NUM_DIMENSIONS (b));
1355 DDR_AFFINE_P (res) = true;
1356 DDR_ARE_DEPENDENT (res) = NULL_TREE;
1357 DDR_SUBSCRIPTS (res) = VEC_alloc (subscript_p, heap, DR_NUM_DIMENSIONS (a));
1358 DDR_LOOP_NEST (res) = loop_nest;
1359 DDR_INNER_LOOP (res) = 0;
1360 DDR_SELF_REFERENCE (res) = false;
1362 for (i = 0; i < DR_NUM_DIMENSIONS (a); i++)
1364 struct subscript *subscript;
1366 subscript = XNEW (struct subscript);
1367 SUB_CONFLICTS_IN_A (subscript) = conflict_fn_not_known ();
1368 SUB_CONFLICTS_IN_B (subscript) = conflict_fn_not_known ();
1369 SUB_LAST_CONFLICT (subscript) = chrec_dont_know;
1370 SUB_DISTANCE (subscript) = chrec_dont_know;
1371 VEC_safe_push (subscript_p, heap, DDR_SUBSCRIPTS (res), subscript);
1377 /* Frees memory used by the conflict function F. */
1380 free_conflict_function (conflict_function *f)
1384 if (CF_NONTRIVIAL_P (f))
1386 for (i = 0; i < f->n; i++)
1387 affine_fn_free (f->fns[i]);
1392 /* Frees memory used by SUBSCRIPTS. */
1395 free_subscripts (VEC (subscript_p, heap) *subscripts)
1400 for (i = 0; VEC_iterate (subscript_p, subscripts, i, s); i++)
1402 free_conflict_function (s->conflicting_iterations_in_a);
1403 free_conflict_function (s->conflicting_iterations_in_b);
1406 VEC_free (subscript_p, heap, subscripts);
1409 /* Set DDR_ARE_DEPENDENT to CHREC and finalize the subscript overlap
1413 finalize_ddr_dependent (struct data_dependence_relation *ddr,
1416 if (dump_file && (dump_flags & TDF_DETAILS))
1418 fprintf (dump_file, "(dependence classified: ");
1419 print_generic_expr (dump_file, chrec, 0);
1420 fprintf (dump_file, ")\n");
1423 DDR_ARE_DEPENDENT (ddr) = chrec;
1424 free_subscripts (DDR_SUBSCRIPTS (ddr));
1425 DDR_SUBSCRIPTS (ddr) = NULL;
1428 /* The dependence relation DDR cannot be represented by a distance
1432 non_affine_dependence_relation (struct data_dependence_relation *ddr)
1434 if (dump_file && (dump_flags & TDF_DETAILS))
1435 fprintf (dump_file, "(Dependence relation cannot be represented by distance vector.) \n");
1437 DDR_AFFINE_P (ddr) = false;
1442 /* This section contains the classic Banerjee tests. */
1444 /* Returns true iff CHREC_A and CHREC_B are not dependent on any index
1445 variables, i.e., if the ZIV (Zero Index Variable) test is true. */
1448 ziv_subscript_p (const_tree chrec_a, const_tree chrec_b)
1450 return (evolution_function_is_constant_p (chrec_a)
1451 && evolution_function_is_constant_p (chrec_b));
1454 /* Returns true iff CHREC_A and CHREC_B are dependent on an index
1455 variable, i.e., if the SIV (Single Index Variable) test is true. */
1458 siv_subscript_p (const_tree chrec_a, const_tree chrec_b)
1460 if ((evolution_function_is_constant_p (chrec_a)
1461 && evolution_function_is_univariate_p (chrec_b))
1462 || (evolution_function_is_constant_p (chrec_b)
1463 && evolution_function_is_univariate_p (chrec_a)))
1466 if (evolution_function_is_univariate_p (chrec_a)
1467 && evolution_function_is_univariate_p (chrec_b))
1469 switch (TREE_CODE (chrec_a))
1471 case POLYNOMIAL_CHREC:
1472 switch (TREE_CODE (chrec_b))
1474 case POLYNOMIAL_CHREC:
1475 if (CHREC_VARIABLE (chrec_a) != CHREC_VARIABLE (chrec_b))
1490 /* Creates a conflict function with N dimensions. The affine functions
1491 in each dimension follow. */
1493 static conflict_function *
1494 conflict_fn (unsigned n, ...)
1497 conflict_function *ret = XCNEW (conflict_function);
1500 gcc_assert (0 < n && n <= MAX_DIM);
1504 for (i = 0; i < n; i++)
1505 ret->fns[i] = va_arg (ap, affine_fn);
1511 /* Returns constant affine function with value CST. */
1514 affine_fn_cst (tree cst)
1516 affine_fn fn = VEC_alloc (tree, heap, 1);
1517 VEC_quick_push (tree, fn, cst);
1521 /* Returns affine function with single variable, CST + COEF * x_DIM. */
1524 affine_fn_univar (tree cst, unsigned dim, tree coef)
1526 affine_fn fn = VEC_alloc (tree, heap, dim + 1);
1529 gcc_assert (dim > 0);
1530 VEC_quick_push (tree, fn, cst);
1531 for (i = 1; i < dim; i++)
1532 VEC_quick_push (tree, fn, integer_zero_node);
1533 VEC_quick_push (tree, fn, coef);
1537 /* Analyze a ZIV (Zero Index Variable) subscript. *OVERLAPS_A and
1538 *OVERLAPS_B are initialized to the functions that describe the
1539 relation between the elements accessed twice by CHREC_A and
1540 CHREC_B. For k >= 0, the following property is verified:
1542 CHREC_A (*OVERLAPS_A (k)) = CHREC_B (*OVERLAPS_B (k)). */
1545 analyze_ziv_subscript (tree chrec_a,
1547 conflict_function **overlaps_a,
1548 conflict_function **overlaps_b,
1549 tree *last_conflicts)
1551 tree type, difference;
1552 dependence_stats.num_ziv++;
1554 if (dump_file && (dump_flags & TDF_DETAILS))
1555 fprintf (dump_file, "(analyze_ziv_subscript \n");
1557 type = signed_type_for_types (TREE_TYPE (chrec_a), TREE_TYPE (chrec_b));
1558 chrec_a = chrec_convert (type, chrec_a, NULL);
1559 chrec_b = chrec_convert (type, chrec_b, NULL);
1560 difference = chrec_fold_minus (type, chrec_a, chrec_b);
1562 switch (TREE_CODE (difference))
1565 if (integer_zerop (difference))
1567 /* The difference is equal to zero: the accessed index
1568 overlaps for each iteration in the loop. */
1569 *overlaps_a = conflict_fn (1, affine_fn_cst (integer_zero_node));
1570 *overlaps_b = conflict_fn (1, affine_fn_cst (integer_zero_node));
1571 *last_conflicts = chrec_dont_know;
1572 dependence_stats.num_ziv_dependent++;
1576 /* The accesses do not overlap. */
1577 *overlaps_a = conflict_fn_no_dependence ();
1578 *overlaps_b = conflict_fn_no_dependence ();
1579 *last_conflicts = integer_zero_node;
1580 dependence_stats.num_ziv_independent++;
1585 /* We're not sure whether the indexes overlap. For the moment,
1586 conservatively answer "don't know". */
1587 if (dump_file && (dump_flags & TDF_DETAILS))
1588 fprintf (dump_file, "ziv test failed: difference is non-integer.\n");
1590 *overlaps_a = conflict_fn_not_known ();
1591 *overlaps_b = conflict_fn_not_known ();
1592 *last_conflicts = chrec_dont_know;
1593 dependence_stats.num_ziv_unimplemented++;
1597 if (dump_file && (dump_flags & TDF_DETAILS))
1598 fprintf (dump_file, ")\n");
1601 /* Sets NIT to the estimated number of executions of the statements in
1602 LOOP. If CONSERVATIVE is true, we must be sure that NIT is at least as
1603 large as the number of iterations. If we have no reliable estimate,
1604 the function returns false, otherwise returns true. */
1607 estimated_loop_iterations (struct loop *loop, bool conservative,
1610 estimate_numbers_of_iterations_loop (loop);
1613 if (!loop->any_upper_bound)
1616 *nit = loop->nb_iterations_upper_bound;
1620 if (!loop->any_estimate)
1623 *nit = loop->nb_iterations_estimate;
1629 /* Similar to estimated_loop_iterations, but returns the estimate only
1630 if it fits to HOST_WIDE_INT. If this is not the case, or the estimate
1631 on the number of iterations of LOOP could not be derived, returns -1. */
1634 estimated_loop_iterations_int (struct loop *loop, bool conservative)
1637 HOST_WIDE_INT hwi_nit;
1639 if (!estimated_loop_iterations (loop, conservative, &nit))
1642 if (!double_int_fits_in_shwi_p (nit))
1644 hwi_nit = double_int_to_shwi (nit);
1646 return hwi_nit < 0 ? -1 : hwi_nit;
1649 /* Similar to estimated_loop_iterations, but returns the estimate as a tree,
1650 and only if it fits to the int type. If this is not the case, or the
1651 estimate on the number of iterations of LOOP could not be derived, returns
1655 estimated_loop_iterations_tree (struct loop *loop, bool conservative)
1660 if (!estimated_loop_iterations (loop, conservative, &nit))
1661 return chrec_dont_know;
1663 type = lang_hooks.types.type_for_size (INT_TYPE_SIZE, true);
1664 if (!double_int_fits_to_tree_p (type, nit))
1665 return chrec_dont_know;
1667 return double_int_to_tree (type, nit);
1670 /* Analyze a SIV (Single Index Variable) subscript where CHREC_A is a
1671 constant, and CHREC_B is an affine function. *OVERLAPS_A and
1672 *OVERLAPS_B are initialized to the functions that describe the
1673 relation between the elements accessed twice by CHREC_A and
1674 CHREC_B. For k >= 0, the following property is verified:
1676 CHREC_A (*OVERLAPS_A (k)) = CHREC_B (*OVERLAPS_B (k)). */
1679 analyze_siv_subscript_cst_affine (tree chrec_a,
1681 conflict_function **overlaps_a,
1682 conflict_function **overlaps_b,
1683 tree *last_conflicts)
1685 bool value0, value1, value2;
1686 tree type, difference, tmp;
1688 type = signed_type_for_types (TREE_TYPE (chrec_a), TREE_TYPE (chrec_b));
1689 chrec_a = chrec_convert (type, chrec_a, NULL);
1690 chrec_b = chrec_convert (type, chrec_b, NULL);
1691 difference = chrec_fold_minus (type, initial_condition (chrec_b), chrec_a);
1693 if (!chrec_is_positive (initial_condition (difference), &value0))
1695 if (dump_file && (dump_flags & TDF_DETAILS))
1696 fprintf (dump_file, "siv test failed: chrec is not positive.\n");
1698 dependence_stats.num_siv_unimplemented++;
1699 *overlaps_a = conflict_fn_not_known ();
1700 *overlaps_b = conflict_fn_not_known ();
1701 *last_conflicts = chrec_dont_know;
1706 if (value0 == false)
1708 if (!chrec_is_positive (CHREC_RIGHT (chrec_b), &value1))
1710 if (dump_file && (dump_flags & TDF_DETAILS))
1711 fprintf (dump_file, "siv test failed: chrec not positive.\n");
1713 *overlaps_a = conflict_fn_not_known ();
1714 *overlaps_b = conflict_fn_not_known ();
1715 *last_conflicts = chrec_dont_know;
1716 dependence_stats.num_siv_unimplemented++;
1725 chrec_b = {10, +, 1}
1728 if (tree_fold_divides_p (CHREC_RIGHT (chrec_b), difference))
1730 HOST_WIDE_INT numiter;
1731 struct loop *loop = get_chrec_loop (chrec_b);
1733 *overlaps_a = conflict_fn (1, affine_fn_cst (integer_zero_node));
1734 tmp = fold_build2 (EXACT_DIV_EXPR, type,
1735 fold_build1 (ABS_EXPR, type, difference),
1736 CHREC_RIGHT (chrec_b));
1737 *overlaps_b = conflict_fn (1, affine_fn_cst (tmp));
1738 *last_conflicts = integer_one_node;
1741 /* Perform weak-zero siv test to see if overlap is
1742 outside the loop bounds. */
1743 numiter = estimated_loop_iterations_int (loop, false);
1746 && compare_tree_int (tmp, numiter) > 0)
1748 free_conflict_function (*overlaps_a);
1749 free_conflict_function (*overlaps_b);
1750 *overlaps_a = conflict_fn_no_dependence ();
1751 *overlaps_b = conflict_fn_no_dependence ();
1752 *last_conflicts = integer_zero_node;
1753 dependence_stats.num_siv_independent++;
1756 dependence_stats.num_siv_dependent++;
1760 /* When the step does not divide the difference, there are
1764 *overlaps_a = conflict_fn_no_dependence ();
1765 *overlaps_b = conflict_fn_no_dependence ();
1766 *last_conflicts = integer_zero_node;
1767 dependence_stats.num_siv_independent++;
1776 chrec_b = {10, +, -1}
1778 In this case, chrec_a will not overlap with chrec_b. */
1779 *overlaps_a = conflict_fn_no_dependence ();
1780 *overlaps_b = conflict_fn_no_dependence ();
1781 *last_conflicts = integer_zero_node;
1782 dependence_stats.num_siv_independent++;
1789 if (!chrec_is_positive (CHREC_RIGHT (chrec_b), &value2))
1791 if (dump_file && (dump_flags & TDF_DETAILS))
1792 fprintf (dump_file, "siv test failed: chrec not positive.\n");
1794 *overlaps_a = conflict_fn_not_known ();
1795 *overlaps_b = conflict_fn_not_known ();
1796 *last_conflicts = chrec_dont_know;
1797 dependence_stats.num_siv_unimplemented++;
1802 if (value2 == false)
1806 chrec_b = {10, +, -1}
1808 if (tree_fold_divides_p (CHREC_RIGHT (chrec_b), difference))
1810 HOST_WIDE_INT numiter;
1811 struct loop *loop = get_chrec_loop (chrec_b);
1813 *overlaps_a = conflict_fn (1, affine_fn_cst (integer_zero_node));
1814 tmp = fold_build2 (EXACT_DIV_EXPR, type, difference,
1815 CHREC_RIGHT (chrec_b));
1816 *overlaps_b = conflict_fn (1, affine_fn_cst (tmp));
1817 *last_conflicts = integer_one_node;
1819 /* Perform weak-zero siv test to see if overlap is
1820 outside the loop bounds. */
1821 numiter = estimated_loop_iterations_int (loop, false);
1824 && compare_tree_int (tmp, numiter) > 0)
1826 free_conflict_function (*overlaps_a);
1827 free_conflict_function (*overlaps_b);
1828 *overlaps_a = conflict_fn_no_dependence ();
1829 *overlaps_b = conflict_fn_no_dependence ();
1830 *last_conflicts = integer_zero_node;
1831 dependence_stats.num_siv_independent++;
1834 dependence_stats.num_siv_dependent++;
1838 /* When the step does not divide the difference, there
1842 *overlaps_a = conflict_fn_no_dependence ();
1843 *overlaps_b = conflict_fn_no_dependence ();
1844 *last_conflicts = integer_zero_node;
1845 dependence_stats.num_siv_independent++;
1855 In this case, chrec_a will not overlap with chrec_b. */
1856 *overlaps_a = conflict_fn_no_dependence ();
1857 *overlaps_b = conflict_fn_no_dependence ();
1858 *last_conflicts = integer_zero_node;
1859 dependence_stats.num_siv_independent++;
1867 /* Helper recursive function for initializing the matrix A. Returns
1868 the initial value of CHREC. */
1871 initialize_matrix_A (lambda_matrix A, tree chrec, unsigned index, int mult)
1875 switch (TREE_CODE (chrec))
1877 case POLYNOMIAL_CHREC:
1878 gcc_assert (TREE_CODE (CHREC_RIGHT (chrec)) == INTEGER_CST);
1880 A[index][0] = mult * int_cst_value (CHREC_RIGHT (chrec));
1881 return initialize_matrix_A (A, CHREC_LEFT (chrec), index + 1, mult);
1887 tree op0 = initialize_matrix_A (A, TREE_OPERAND (chrec, 0), index, mult);
1888 tree op1 = initialize_matrix_A (A, TREE_OPERAND (chrec, 1), index, mult);
1890 return chrec_fold_op (TREE_CODE (chrec), chrec_type (chrec), op0, op1);
1895 tree op = initialize_matrix_A (A, TREE_OPERAND (chrec, 0), index, mult);
1896 return chrec_convert (chrec_type (chrec), op, NULL);
1901 /* Handle ~X as -1 - X. */
1902 tree op = initialize_matrix_A (A, TREE_OPERAND (chrec, 0), index, mult);
1903 return chrec_fold_op (MINUS_EXPR, chrec_type (chrec),
1904 build_int_cst (TREE_TYPE (chrec), -1), op);
1916 #define FLOOR_DIV(x,y) ((x) / (y))
1918 /* Solves the special case of the Diophantine equation:
1919 | {0, +, STEP_A}_x (OVERLAPS_A) = {0, +, STEP_B}_y (OVERLAPS_B)
1921 Computes the descriptions OVERLAPS_A and OVERLAPS_B. NITER is the
1922 number of iterations that loops X and Y run. The overlaps will be
1923 constructed as evolutions in dimension DIM. */
1926 compute_overlap_steps_for_affine_univar (int niter, int step_a, int step_b,
1927 affine_fn *overlaps_a,
1928 affine_fn *overlaps_b,
1929 tree *last_conflicts, int dim)
1931 if (((step_a > 0 && step_b > 0)
1932 || (step_a < 0 && step_b < 0)))
1934 int step_overlaps_a, step_overlaps_b;
1935 int gcd_steps_a_b, last_conflict, tau2;
1937 gcd_steps_a_b = gcd (step_a, step_b);
1938 step_overlaps_a = step_b / gcd_steps_a_b;
1939 step_overlaps_b = step_a / gcd_steps_a_b;
1943 tau2 = FLOOR_DIV (niter, step_overlaps_a);
1944 tau2 = MIN (tau2, FLOOR_DIV (niter, step_overlaps_b));
1945 last_conflict = tau2;
1946 *last_conflicts = build_int_cst (NULL_TREE, last_conflict);
1949 *last_conflicts = chrec_dont_know;
1951 *overlaps_a = affine_fn_univar (integer_zero_node, dim,
1952 build_int_cst (NULL_TREE,
1954 *overlaps_b = affine_fn_univar (integer_zero_node, dim,
1955 build_int_cst (NULL_TREE,
1961 *overlaps_a = affine_fn_cst (integer_zero_node);
1962 *overlaps_b = affine_fn_cst (integer_zero_node);
1963 *last_conflicts = integer_zero_node;
1967 /* Solves the special case of a Diophantine equation where CHREC_A is
1968 an affine bivariate function, and CHREC_B is an affine univariate
1969 function. For example,
1971 | {{0, +, 1}_x, +, 1335}_y = {0, +, 1336}_z
1973 has the following overlapping functions:
1975 | x (t, u, v) = {{0, +, 1336}_t, +, 1}_v
1976 | y (t, u, v) = {{0, +, 1336}_u, +, 1}_v
1977 | z (t, u, v) = {{{0, +, 1}_t, +, 1335}_u, +, 1}_v
1979 FORNOW: This is a specialized implementation for a case occurring in
1980 a common benchmark. Implement the general algorithm. */
1983 compute_overlap_steps_for_affine_1_2 (tree chrec_a, tree chrec_b,
1984 conflict_function **overlaps_a,
1985 conflict_function **overlaps_b,
1986 tree *last_conflicts)
1988 bool xz_p, yz_p, xyz_p;
1989 int step_x, step_y, step_z;
1990 HOST_WIDE_INT niter_x, niter_y, niter_z, niter;
1991 affine_fn overlaps_a_xz, overlaps_b_xz;
1992 affine_fn overlaps_a_yz, overlaps_b_yz;
1993 affine_fn overlaps_a_xyz, overlaps_b_xyz;
1994 affine_fn ova1, ova2, ovb;
1995 tree last_conflicts_xz, last_conflicts_yz, last_conflicts_xyz;
1997 step_x = int_cst_value (CHREC_RIGHT (CHREC_LEFT (chrec_a)));
1998 step_y = int_cst_value (CHREC_RIGHT (chrec_a));
1999 step_z = int_cst_value (CHREC_RIGHT (chrec_b));
2002 estimated_loop_iterations_int (get_chrec_loop (CHREC_LEFT (chrec_a)),
2004 niter_y = estimated_loop_iterations_int (get_chrec_loop (chrec_a), false);
2005 niter_z = estimated_loop_iterations_int (get_chrec_loop (chrec_b), false);
2007 if (niter_x < 0 || niter_y < 0 || niter_z < 0)
2009 if (dump_file && (dump_flags & TDF_DETAILS))
2010 fprintf (dump_file, "overlap steps test failed: no iteration counts.\n");
2012 *overlaps_a = conflict_fn_not_known ();
2013 *overlaps_b = conflict_fn_not_known ();
2014 *last_conflicts = chrec_dont_know;
2018 niter = MIN (niter_x, niter_z);
2019 compute_overlap_steps_for_affine_univar (niter, step_x, step_z,
2022 &last_conflicts_xz, 1);
2023 niter = MIN (niter_y, niter_z);
2024 compute_overlap_steps_for_affine_univar (niter, step_y, step_z,
2027 &last_conflicts_yz, 2);
2028 niter = MIN (niter_x, niter_z);
2029 niter = MIN (niter_y, niter);
2030 compute_overlap_steps_for_affine_univar (niter, step_x + step_y, step_z,
2033 &last_conflicts_xyz, 3);
2035 xz_p = !integer_zerop (last_conflicts_xz);
2036 yz_p = !integer_zerop (last_conflicts_yz);
2037 xyz_p = !integer_zerop (last_conflicts_xyz);
2039 if (xz_p || yz_p || xyz_p)
2041 ova1 = affine_fn_cst (integer_zero_node);
2042 ova2 = affine_fn_cst (integer_zero_node);
2043 ovb = affine_fn_cst (integer_zero_node);
2046 affine_fn t0 = ova1;
2049 ova1 = affine_fn_plus (ova1, overlaps_a_xz);
2050 ovb = affine_fn_plus (ovb, overlaps_b_xz);
2051 affine_fn_free (t0);
2052 affine_fn_free (t2);
2053 *last_conflicts = last_conflicts_xz;
2057 affine_fn t0 = ova2;
2060 ova2 = affine_fn_plus (ova2, overlaps_a_yz);
2061 ovb = affine_fn_plus (ovb, overlaps_b_yz);
2062 affine_fn_free (t0);
2063 affine_fn_free (t2);
2064 *last_conflicts = last_conflicts_yz;
2068 affine_fn t0 = ova1;
2069 affine_fn t2 = ova2;
2072 ova1 = affine_fn_plus (ova1, overlaps_a_xyz);
2073 ova2 = affine_fn_plus (ova2, overlaps_a_xyz);
2074 ovb = affine_fn_plus (ovb, overlaps_b_xyz);
2075 affine_fn_free (t0);
2076 affine_fn_free (t2);
2077 affine_fn_free (t4);
2078 *last_conflicts = last_conflicts_xyz;
2080 *overlaps_a = conflict_fn (2, ova1, ova2);
2081 *overlaps_b = conflict_fn (1, ovb);
2085 *overlaps_a = conflict_fn (1, affine_fn_cst (integer_zero_node));
2086 *overlaps_b = conflict_fn (1, affine_fn_cst (integer_zero_node));
2087 *last_conflicts = integer_zero_node;
2090 affine_fn_free (overlaps_a_xz);
2091 affine_fn_free (overlaps_b_xz);
2092 affine_fn_free (overlaps_a_yz);
2093 affine_fn_free (overlaps_b_yz);
2094 affine_fn_free (overlaps_a_xyz);
2095 affine_fn_free (overlaps_b_xyz);
2098 /* Determines the overlapping elements due to accesses CHREC_A and
2099 CHREC_B, that are affine functions. This function cannot handle
2100 symbolic evolution functions, ie. when initial conditions are
2101 parameters, because it uses lambda matrices of integers. */
2104 analyze_subscript_affine_affine (tree chrec_a,
2106 conflict_function **overlaps_a,
2107 conflict_function **overlaps_b,
2108 tree *last_conflicts)
2110 unsigned nb_vars_a, nb_vars_b, dim;
2111 HOST_WIDE_INT init_a, init_b, gamma, gcd_alpha_beta;
2112 lambda_matrix A, U, S;
2114 if (eq_evolutions_p (chrec_a, chrec_b))
2116 /* The accessed index overlaps for each iteration in the
2118 *overlaps_a = conflict_fn (1, affine_fn_cst (integer_zero_node));
2119 *overlaps_b = conflict_fn (1, affine_fn_cst (integer_zero_node));
2120 *last_conflicts = chrec_dont_know;
2123 if (dump_file && (dump_flags & TDF_DETAILS))
2124 fprintf (dump_file, "(analyze_subscript_affine_affine \n");
2126 /* For determining the initial intersection, we have to solve a
2127 Diophantine equation. This is the most time consuming part.
2129 For answering to the question: "Is there a dependence?" we have
2130 to prove that there exists a solution to the Diophantine
2131 equation, and that the solution is in the iteration domain,
2132 i.e. the solution is positive or zero, and that the solution
2133 happens before the upper bound loop.nb_iterations. Otherwise
2134 there is no dependence. This function outputs a description of
2135 the iterations that hold the intersections. */
2137 nb_vars_a = nb_vars_in_chrec (chrec_a);
2138 nb_vars_b = nb_vars_in_chrec (chrec_b);
2140 dim = nb_vars_a + nb_vars_b;
2141 U = lambda_matrix_new (dim, dim);
2142 A = lambda_matrix_new (dim, 1);
2143 S = lambda_matrix_new (dim, 1);
2145 init_a = int_cst_value (initialize_matrix_A (A, chrec_a, 0, 1));
2146 init_b = int_cst_value (initialize_matrix_A (A, chrec_b, nb_vars_a, -1));
2147 gamma = init_b - init_a;
2149 /* Don't do all the hard work of solving the Diophantine equation
2150 when we already know the solution: for example,
2153 | gamma = 3 - 3 = 0.
2154 Then the first overlap occurs during the first iterations:
2155 | {3, +, 1}_1 ({0, +, 4}_x) = {3, +, 4}_2 ({0, +, 1}_x)
2159 if (nb_vars_a == 1 && nb_vars_b == 1)
2161 HOST_WIDE_INT step_a, step_b;
2162 HOST_WIDE_INT niter, niter_a, niter_b;
2165 niter_a = estimated_loop_iterations_int (get_chrec_loop (chrec_a),
2167 niter_b = estimated_loop_iterations_int (get_chrec_loop (chrec_b),
2169 niter = MIN (niter_a, niter_b);
2170 step_a = int_cst_value (CHREC_RIGHT (chrec_a));
2171 step_b = int_cst_value (CHREC_RIGHT (chrec_b));
2173 compute_overlap_steps_for_affine_univar (niter, step_a, step_b,
2176 *overlaps_a = conflict_fn (1, ova);
2177 *overlaps_b = conflict_fn (1, ovb);
2180 else if (nb_vars_a == 2 && nb_vars_b == 1)
2181 compute_overlap_steps_for_affine_1_2
2182 (chrec_a, chrec_b, overlaps_a, overlaps_b, last_conflicts);
2184 else if (nb_vars_a == 1 && nb_vars_b == 2)
2185 compute_overlap_steps_for_affine_1_2
2186 (chrec_b, chrec_a, overlaps_b, overlaps_a, last_conflicts);
2190 if (dump_file && (dump_flags & TDF_DETAILS))
2191 fprintf (dump_file, "affine-affine test failed: too many variables.\n");
2192 *overlaps_a = conflict_fn_not_known ();
2193 *overlaps_b = conflict_fn_not_known ();
2194 *last_conflicts = chrec_dont_know;
2196 goto end_analyze_subs_aa;
2200 lambda_matrix_right_hermite (A, dim, 1, S, U);
2205 lambda_matrix_row_negate (U, dim, 0);
2207 gcd_alpha_beta = S[0][0];
2209 /* Something went wrong: for example in {1, +, 0}_5 vs. {0, +, 0}_5,
2210 but that is a quite strange case. Instead of ICEing, answer
2212 if (gcd_alpha_beta == 0)
2214 *overlaps_a = conflict_fn_not_known ();
2215 *overlaps_b = conflict_fn_not_known ();
2216 *last_conflicts = chrec_dont_know;
2217 goto end_analyze_subs_aa;
2220 /* The classic "gcd-test". */
2221 if (!int_divides_p (gcd_alpha_beta, gamma))
2223 /* The "gcd-test" has determined that there is no integer
2224 solution, i.e. there is no dependence. */
2225 *overlaps_a = conflict_fn_no_dependence ();
2226 *overlaps_b = conflict_fn_no_dependence ();
2227 *last_conflicts = integer_zero_node;
2230 /* Both access functions are univariate. This includes SIV and MIV cases. */
2231 else if (nb_vars_a == 1 && nb_vars_b == 1)
2233 /* Both functions should have the same evolution sign. */
2234 if (((A[0][0] > 0 && -A[1][0] > 0)
2235 || (A[0][0] < 0 && -A[1][0] < 0)))
2237 /* The solutions are given by:
2239 | [GAMMA/GCD_ALPHA_BETA t].[u11 u12] = [x0]
2242 For a given integer t. Using the following variables,
2244 | i0 = u11 * gamma / gcd_alpha_beta
2245 | j0 = u12 * gamma / gcd_alpha_beta
2252 | y0 = j0 + j1 * t. */
2253 HOST_WIDE_INT i0, j0, i1, j1;
2255 i0 = U[0][0] * gamma / gcd_alpha_beta;
2256 j0 = U[0][1] * gamma / gcd_alpha_beta;
2260 if ((i1 == 0 && i0 < 0)
2261 || (j1 == 0 && j0 < 0))
2263 /* There is no solution.
2264 FIXME: The case "i0 > nb_iterations, j0 > nb_iterations"
2265 falls in here, but for the moment we don't look at the
2266 upper bound of the iteration domain. */
2267 *overlaps_a = conflict_fn_no_dependence ();
2268 *overlaps_b = conflict_fn_no_dependence ();
2269 *last_conflicts = integer_zero_node;
2270 goto end_analyze_subs_aa;
2273 if (i1 > 0 && j1 > 0)
2275 HOST_WIDE_INT niter_a = estimated_loop_iterations_int
2276 (get_chrec_loop (chrec_a), false);
2277 HOST_WIDE_INT niter_b = estimated_loop_iterations_int
2278 (get_chrec_loop (chrec_b), false);
2279 HOST_WIDE_INT niter = MIN (niter_a, niter_b);
2281 /* (X0, Y0) is a solution of the Diophantine equation:
2282 "chrec_a (X0) = chrec_b (Y0)". */
2283 HOST_WIDE_INT tau1 = MAX (CEIL (-i0, i1),
2285 HOST_WIDE_INT x0 = i1 * tau1 + i0;
2286 HOST_WIDE_INT y0 = j1 * tau1 + j0;
2288 /* (X1, Y1) is the smallest positive solution of the eq
2289 "chrec_a (X1) = chrec_b (Y1)", i.e. this is where the
2290 first conflict occurs. */
2291 HOST_WIDE_INT min_multiple = MIN (x0 / i1, y0 / j1);
2292 HOST_WIDE_INT x1 = x0 - i1 * min_multiple;
2293 HOST_WIDE_INT y1 = y0 - j1 * min_multiple;
2297 HOST_WIDE_INT tau2 = MIN (FLOOR_DIV (niter - i0, i1),
2298 FLOOR_DIV (niter - j0, j1));
2299 HOST_WIDE_INT last_conflict = tau2 - (x1 - i0)/i1;
2301 /* If the overlap occurs outside of the bounds of the
2302 loop, there is no dependence. */
2303 if (x1 > niter || y1 > niter)
2305 *overlaps_a = conflict_fn_no_dependence ();
2306 *overlaps_b = conflict_fn_no_dependence ();
2307 *last_conflicts = integer_zero_node;
2308 goto end_analyze_subs_aa;
2311 *last_conflicts = build_int_cst (NULL_TREE, last_conflict);
2314 *last_conflicts = chrec_dont_know;
2318 affine_fn_univar (build_int_cst (NULL_TREE, x1),
2320 build_int_cst (NULL_TREE, i1)));
2323 affine_fn_univar (build_int_cst (NULL_TREE, y1),
2325 build_int_cst (NULL_TREE, j1)));
2329 /* FIXME: For the moment, the upper bound of the
2330 iteration domain for i and j is not checked. */
2331 if (dump_file && (dump_flags & TDF_DETAILS))
2332 fprintf (dump_file, "affine-affine test failed: unimplemented.\n");
2333 *overlaps_a = conflict_fn_not_known ();
2334 *overlaps_b = conflict_fn_not_known ();
2335 *last_conflicts = chrec_dont_know;
2340 if (dump_file && (dump_flags & TDF_DETAILS))
2341 fprintf (dump_file, "affine-affine test failed: unimplemented.\n");
2342 *overlaps_a = conflict_fn_not_known ();
2343 *overlaps_b = conflict_fn_not_known ();
2344 *last_conflicts = chrec_dont_know;
2349 if (dump_file && (dump_flags & TDF_DETAILS))
2350 fprintf (dump_file, "affine-affine test failed: unimplemented.\n");
2351 *overlaps_a = conflict_fn_not_known ();
2352 *overlaps_b = conflict_fn_not_known ();
2353 *last_conflicts = chrec_dont_know;
2356 end_analyze_subs_aa:
2357 if (dump_file && (dump_flags & TDF_DETAILS))
2359 fprintf (dump_file, " (overlaps_a = ");
2360 dump_conflict_function (dump_file, *overlaps_a);
2361 fprintf (dump_file, ")\n (overlaps_b = ");
2362 dump_conflict_function (dump_file, *overlaps_b);
2363 fprintf (dump_file, ")\n");
2364 fprintf (dump_file, ")\n");
2368 /* Returns true when analyze_subscript_affine_affine can be used for
2369 determining the dependence relation between chrec_a and chrec_b,
2370 that contain symbols. This function modifies chrec_a and chrec_b
2371 such that the analysis result is the same, and such that they don't
2372 contain symbols, and then can safely be passed to the analyzer.
2374 Example: The analysis of the following tuples of evolutions produce
2375 the same results: {x+1, +, 1}_1 vs. {x+3, +, 1}_1, and {-2, +, 1}_1
2378 {x+1, +, 1}_1 ({2, +, 1}_1) = {x+3, +, 1}_1 ({0, +, 1}_1)
2379 {-2, +, 1}_1 ({2, +, 1}_1) = {0, +, 1}_1 ({0, +, 1}_1)
2383 can_use_analyze_subscript_affine_affine (tree *chrec_a, tree *chrec_b)
2385 tree diff, type, left_a, left_b, right_b;
2387 if (chrec_contains_symbols (CHREC_RIGHT (*chrec_a))
2388 || chrec_contains_symbols (CHREC_RIGHT (*chrec_b)))
2389 /* FIXME: For the moment not handled. Might be refined later. */
2392 type = chrec_type (*chrec_a);
2393 left_a = CHREC_LEFT (*chrec_a);
2394 left_b = chrec_convert (type, CHREC_LEFT (*chrec_b), NULL);
2395 diff = chrec_fold_minus (type, left_a, left_b);
2397 if (!evolution_function_is_constant_p (diff))
2400 if (dump_file && (dump_flags & TDF_DETAILS))
2401 fprintf (dump_file, "can_use_subscript_aff_aff_for_symbolic \n");
2403 *chrec_a = build_polynomial_chrec (CHREC_VARIABLE (*chrec_a),
2404 diff, CHREC_RIGHT (*chrec_a));
2405 right_b = chrec_convert (type, CHREC_RIGHT (*chrec_b), NULL);
2406 *chrec_b = build_polynomial_chrec (CHREC_VARIABLE (*chrec_b),
2407 build_int_cst (type, 0),
2412 /* Analyze a SIV (Single Index Variable) subscript. *OVERLAPS_A and
2413 *OVERLAPS_B are initialized to the functions that describe the
2414 relation between the elements accessed twice by CHREC_A and
2415 CHREC_B. For k >= 0, the following property is verified:
2417 CHREC_A (*OVERLAPS_A (k)) = CHREC_B (*OVERLAPS_B (k)). */
2420 analyze_siv_subscript (tree chrec_a,
2422 conflict_function **overlaps_a,
2423 conflict_function **overlaps_b,
2424 tree *last_conflicts,
2427 dependence_stats.num_siv++;
2429 if (dump_file && (dump_flags & TDF_DETAILS))
2430 fprintf (dump_file, "(analyze_siv_subscript \n");
2432 if (evolution_function_is_constant_p (chrec_a)
2433 && evolution_function_is_affine_in_loop (chrec_b, loop_nest_num))
2434 analyze_siv_subscript_cst_affine (chrec_a, chrec_b,
2435 overlaps_a, overlaps_b, last_conflicts);
2437 else if (evolution_function_is_affine_in_loop (chrec_a, loop_nest_num)
2438 && evolution_function_is_constant_p (chrec_b))
2439 analyze_siv_subscript_cst_affine (chrec_b, chrec_a,
2440 overlaps_b, overlaps_a, last_conflicts);
2442 else if (evolution_function_is_affine_in_loop (chrec_a, loop_nest_num)
2443 && evolution_function_is_affine_in_loop (chrec_b, loop_nest_num))
2445 if (!chrec_contains_symbols (chrec_a)
2446 && !chrec_contains_symbols (chrec_b))
2448 analyze_subscript_affine_affine (chrec_a, chrec_b,
2449 overlaps_a, overlaps_b,
2452 if (CF_NOT_KNOWN_P (*overlaps_a)
2453 || CF_NOT_KNOWN_P (*overlaps_b))
2454 dependence_stats.num_siv_unimplemented++;
2455 else if (CF_NO_DEPENDENCE_P (*overlaps_a)
2456 || CF_NO_DEPENDENCE_P (*overlaps_b))
2457 dependence_stats.num_siv_independent++;
2459 dependence_stats.num_siv_dependent++;
2461 else if (can_use_analyze_subscript_affine_affine (&chrec_a,
2464 analyze_subscript_affine_affine (chrec_a, chrec_b,
2465 overlaps_a, overlaps_b,
2468 if (CF_NOT_KNOWN_P (*overlaps_a)
2469 || CF_NOT_KNOWN_P (*overlaps_b))
2470 dependence_stats.num_siv_unimplemented++;
2471 else if (CF_NO_DEPENDENCE_P (*overlaps_a)
2472 || CF_NO_DEPENDENCE_P (*overlaps_b))
2473 dependence_stats.num_siv_independent++;
2475 dependence_stats.num_siv_dependent++;
2478 goto siv_subscript_dontknow;
2483 siv_subscript_dontknow:;
2484 if (dump_file && (dump_flags & TDF_DETAILS))
2485 fprintf (dump_file, "siv test failed: unimplemented.\n");
2486 *overlaps_a = conflict_fn_not_known ();
2487 *overlaps_b = conflict_fn_not_known ();
2488 *last_conflicts = chrec_dont_know;
2489 dependence_stats.num_siv_unimplemented++;
2492 if (dump_file && (dump_flags & TDF_DETAILS))
2493 fprintf (dump_file, ")\n");
2496 /* Returns false if we can prove that the greatest common divisor of the steps
2497 of CHREC does not divide CST, false otherwise. */
2500 gcd_of_steps_may_divide_p (const_tree chrec, const_tree cst)
2502 HOST_WIDE_INT cd = 0, val;
2505 if (!host_integerp (cst, 0))
2507 val = tree_low_cst (cst, 0);
2509 while (TREE_CODE (chrec) == POLYNOMIAL_CHREC)
2511 step = CHREC_RIGHT (chrec);
2512 if (!host_integerp (step, 0))
2514 cd = gcd (cd, tree_low_cst (step, 0));
2515 chrec = CHREC_LEFT (chrec);
2518 return val % cd == 0;
2521 /* Analyze a MIV (Multiple Index Variable) subscript with respect to
2522 LOOP_NEST. *OVERLAPS_A and *OVERLAPS_B are initialized to the
2523 functions that describe the relation between the elements accessed
2524 twice by CHREC_A and CHREC_B. For k >= 0, the following property
2527 CHREC_A (*OVERLAPS_A (k)) = CHREC_B (*OVERLAPS_B (k)). */
2530 analyze_miv_subscript (tree chrec_a,
2532 conflict_function **overlaps_a,
2533 conflict_function **overlaps_b,
2534 tree *last_conflicts,
2535 struct loop *loop_nest)
2537 /* FIXME: This is a MIV subscript, not yet handled.
2538 Example: (A[{1, +, 1}_1] vs. A[{1, +, 1}_2]) that comes from
2541 In the SIV test we had to solve a Diophantine equation with two
2542 variables. In the MIV case we have to solve a Diophantine
2543 equation with 2*n variables (if the subscript uses n IVs).
2545 tree type, difference;
2547 dependence_stats.num_miv++;
2548 if (dump_file && (dump_flags & TDF_DETAILS))
2549 fprintf (dump_file, "(analyze_miv_subscript \n");
2551 type = signed_type_for_types (TREE_TYPE (chrec_a), TREE_TYPE (chrec_b));
2552 chrec_a = chrec_convert (type, chrec_a, NULL);
2553 chrec_b = chrec_convert (type, chrec_b, NULL);
2554 difference = chrec_fold_minus (type, chrec_a, chrec_b);
2556 if (eq_evolutions_p (chrec_a, chrec_b))
2558 /* Access functions are the same: all the elements are accessed
2559 in the same order. */
2560 *overlaps_a = conflict_fn (1, affine_fn_cst (integer_zero_node));
2561 *overlaps_b = conflict_fn (1, affine_fn_cst (integer_zero_node));
2562 *last_conflicts = estimated_loop_iterations_tree
2563 (get_chrec_loop (chrec_a), true);
2564 dependence_stats.num_miv_dependent++;
2567 else if (evolution_function_is_constant_p (difference)
2568 /* For the moment, the following is verified:
2569 evolution_function_is_affine_multivariate_p (chrec_a,
2571 && !gcd_of_steps_may_divide_p (chrec_a, difference))
2573 /* testsuite/.../ssa-chrec-33.c
2574 {{21, +, 2}_1, +, -2}_2 vs. {{20, +, 2}_1, +, -2}_2
2576 The difference is 1, and all the evolution steps are multiples
2577 of 2, consequently there are no overlapping elements. */
2578 *overlaps_a = conflict_fn_no_dependence ();
2579 *overlaps_b = conflict_fn_no_dependence ();
2580 *last_conflicts = integer_zero_node;
2581 dependence_stats.num_miv_independent++;
2584 else if (evolution_function_is_affine_multivariate_p (chrec_a, loop_nest->num)
2585 && !chrec_contains_symbols (chrec_a)
2586 && evolution_function_is_affine_multivariate_p (chrec_b, loop_nest->num)
2587 && !chrec_contains_symbols (chrec_b))
2589 /* testsuite/.../ssa-chrec-35.c
2590 {0, +, 1}_2 vs. {0, +, 1}_3
2591 the overlapping elements are respectively located at iterations:
2592 {0, +, 1}_x and {0, +, 1}_x,
2593 in other words, we have the equality:
2594 {0, +, 1}_2 ({0, +, 1}_x) = {0, +, 1}_3 ({0, +, 1}_x)
2597 {{0, +, 1}_1, +, 2}_2 ({0, +, 1}_x, {0, +, 1}_y) =
2598 {0, +, 1}_1 ({{0, +, 1}_x, +, 2}_y)
2600 {{0, +, 2}_1, +, 3}_2 ({0, +, 1}_y, {0, +, 1}_x) =
2601 {{0, +, 3}_1, +, 2}_2 ({0, +, 1}_x, {0, +, 1}_y)
2603 analyze_subscript_affine_affine (chrec_a, chrec_b,
2604 overlaps_a, overlaps_b, last_conflicts);
2606 if (CF_NOT_KNOWN_P (*overlaps_a)
2607 || CF_NOT_KNOWN_P (*overlaps_b))
2608 dependence_stats.num_miv_unimplemented++;
2609 else if (CF_NO_DEPENDENCE_P (*overlaps_a)
2610 || CF_NO_DEPENDENCE_P (*overlaps_b))
2611 dependence_stats.num_miv_independent++;
2613 dependence_stats.num_miv_dependent++;
2618 /* When the analysis is too difficult, answer "don't know". */
2619 if (dump_file && (dump_flags & TDF_DETAILS))
2620 fprintf (dump_file, "analyze_miv_subscript test failed: unimplemented.\n");
2622 *overlaps_a = conflict_fn_not_known ();
2623 *overlaps_b = conflict_fn_not_known ();
2624 *last_conflicts = chrec_dont_know;
2625 dependence_stats.num_miv_unimplemented++;
2628 if (dump_file && (dump_flags & TDF_DETAILS))
2629 fprintf (dump_file, ")\n");
2632 /* Determines the iterations for which CHREC_A is equal to CHREC_B in
2633 with respect to LOOP_NEST. OVERLAP_ITERATIONS_A and
2634 OVERLAP_ITERATIONS_B are initialized with two functions that
2635 describe the iterations that contain conflicting elements.
2637 Remark: For an integer k >= 0, the following equality is true:
2639 CHREC_A (OVERLAP_ITERATIONS_A (k)) == CHREC_B (OVERLAP_ITERATIONS_B (k)).
2643 analyze_overlapping_iterations (tree chrec_a,
2645 conflict_function **overlap_iterations_a,
2646 conflict_function **overlap_iterations_b,
2647 tree *last_conflicts, struct loop *loop_nest)
2649 unsigned int lnn = loop_nest->num;
2651 dependence_stats.num_subscript_tests++;
2653 if (dump_file && (dump_flags & TDF_DETAILS))
2655 fprintf (dump_file, "(analyze_overlapping_iterations \n");
2656 fprintf (dump_file, " (chrec_a = ");
2657 print_generic_expr (dump_file, chrec_a, 0);
2658 fprintf (dump_file, ")\n (chrec_b = ");
2659 print_generic_expr (dump_file, chrec_b, 0);
2660 fprintf (dump_file, ")\n");
2663 if (chrec_a == NULL_TREE
2664 || chrec_b == NULL_TREE
2665 || chrec_contains_undetermined (chrec_a)
2666 || chrec_contains_undetermined (chrec_b))
2668 dependence_stats.num_subscript_undetermined++;
2670 *overlap_iterations_a = conflict_fn_not_known ();
2671 *overlap_iterations_b = conflict_fn_not_known ();
2674 /* If they are the same chrec, and are affine, they overlap
2675 on every iteration. */
2676 else if (eq_evolutions_p (chrec_a, chrec_b)
2677 && evolution_function_is_affine_multivariate_p (chrec_a, lnn))
2679 dependence_stats.num_same_subscript_function++;
2680 *overlap_iterations_a = conflict_fn (1, affine_fn_cst (integer_zero_node));
2681 *overlap_iterations_b = conflict_fn (1, affine_fn_cst (integer_zero_node));
2682 *last_conflicts = chrec_dont_know;
2685 /* If they aren't the same, and aren't affine, we can't do anything
2687 else if ((chrec_contains_symbols (chrec_a)
2688 || chrec_contains_symbols (chrec_b))
2689 && (!evolution_function_is_affine_multivariate_p (chrec_a, lnn)
2690 || !evolution_function_is_affine_multivariate_p (chrec_b, lnn)))
2692 dependence_stats.num_subscript_undetermined++;
2693 *overlap_iterations_a = conflict_fn_not_known ();
2694 *overlap_iterations_b = conflict_fn_not_known ();
2697 else if (ziv_subscript_p (chrec_a, chrec_b))
2698 analyze_ziv_subscript (chrec_a, chrec_b,
2699 overlap_iterations_a, overlap_iterations_b,
2702 else if (siv_subscript_p (chrec_a, chrec_b))
2703 analyze_siv_subscript (chrec_a, chrec_b,
2704 overlap_iterations_a, overlap_iterations_b,
2705 last_conflicts, lnn);
2708 analyze_miv_subscript (chrec_a, chrec_b,
2709 overlap_iterations_a, overlap_iterations_b,
2710 last_conflicts, loop_nest);
2712 if (dump_file && (dump_flags & TDF_DETAILS))
2714 fprintf (dump_file, " (overlap_iterations_a = ");
2715 dump_conflict_function (dump_file, *overlap_iterations_a);
2716 fprintf (dump_file, ")\n (overlap_iterations_b = ");
2717 dump_conflict_function (dump_file, *overlap_iterations_b);
2718 fprintf (dump_file, ")\n");
2719 fprintf (dump_file, ")\n");
2723 /* Helper function for uniquely inserting distance vectors. */
2726 save_dist_v (struct data_dependence_relation *ddr, lambda_vector dist_v)
2731 for (i = 0; VEC_iterate (lambda_vector, DDR_DIST_VECTS (ddr), i, v); i++)
2732 if (lambda_vector_equal (v, dist_v, DDR_NB_LOOPS (ddr)))
2735 VEC_safe_push (lambda_vector, heap, DDR_DIST_VECTS (ddr), dist_v);
2738 /* Helper function for uniquely inserting direction vectors. */
2741 save_dir_v (struct data_dependence_relation *ddr, lambda_vector dir_v)
2746 for (i = 0; VEC_iterate (lambda_vector, DDR_DIR_VECTS (ddr), i, v); i++)
2747 if (lambda_vector_equal (v, dir_v, DDR_NB_LOOPS (ddr)))
2750 VEC_safe_push (lambda_vector, heap, DDR_DIR_VECTS (ddr), dir_v);
2753 /* Add a distance of 1 on all the loops outer than INDEX. If we
2754 haven't yet determined a distance for this outer loop, push a new
2755 distance vector composed of the previous distance, and a distance
2756 of 1 for this outer loop. Example:
2764 Saved vectors are of the form (dist_in_1, dist_in_2). First, we
2765 save (0, 1), then we have to save (1, 0). */
2768 add_outer_distances (struct data_dependence_relation *ddr,
2769 lambda_vector dist_v, int index)
2771 /* For each outer loop where init_v is not set, the accesses are
2772 in dependence of distance 1 in the loop. */
2773 while (--index >= 0)
2775 lambda_vector save_v = lambda_vector_new (DDR_NB_LOOPS (ddr));
2776 lambda_vector_copy (dist_v, save_v, DDR_NB_LOOPS (ddr));
2778 save_dist_v (ddr, save_v);
2782 /* Return false when fail to represent the data dependence as a
2783 distance vector. INIT_B is set to true when a component has been
2784 added to the distance vector DIST_V. INDEX_CARRY is then set to
2785 the index in DIST_V that carries the dependence. */
2788 build_classic_dist_vector_1 (struct data_dependence_relation *ddr,
2789 struct data_reference *ddr_a,
2790 struct data_reference *ddr_b,
2791 lambda_vector dist_v, bool *init_b,
2795 lambda_vector init_v = lambda_vector_new (DDR_NB_LOOPS (ddr));
2797 for (i = 0; i < DDR_NUM_SUBSCRIPTS (ddr); i++)
2799 tree access_fn_a, access_fn_b;
2800 struct subscript *subscript = DDR_SUBSCRIPT (ddr, i);
2802 if (chrec_contains_undetermined (SUB_DISTANCE (subscript)))
2804 non_affine_dependence_relation (ddr);
2808 access_fn_a = DR_ACCESS_FN (ddr_a, i);
2809 access_fn_b = DR_ACCESS_FN (ddr_b, i);
2811 if (TREE_CODE (access_fn_a) == POLYNOMIAL_CHREC
2812 && TREE_CODE (access_fn_b) == POLYNOMIAL_CHREC)
2815 int index_a = index_in_loop_nest (CHREC_VARIABLE (access_fn_a),
2816 DDR_LOOP_NEST (ddr));
2817 int index_b = index_in_loop_nest (CHREC_VARIABLE (access_fn_b),
2818 DDR_LOOP_NEST (ddr));
2820 /* The dependence is carried by the outermost loop. Example:
2827 In this case, the dependence is carried by loop_1. */
2828 index = index_a < index_b ? index_a : index_b;
2829 *index_carry = MIN (index, *index_carry);
2831 if (chrec_contains_undetermined (SUB_DISTANCE (subscript)))
2833 non_affine_dependence_relation (ddr);
2837 dist = int_cst_value (SUB_DISTANCE (subscript));
2839 /* This is the subscript coupling test. If we have already
2840 recorded a distance for this loop (a distance coming from
2841 another subscript), it should be the same. For example,
2842 in the following code, there is no dependence:
2849 if (init_v[index] != 0 && dist_v[index] != dist)
2851 finalize_ddr_dependent (ddr, chrec_known);
2855 dist_v[index] = dist;
2859 else if (!operand_equal_p (access_fn_a, access_fn_b, 0))
2861 /* This can be for example an affine vs. constant dependence
2862 (T[i] vs. T[3]) that is not an affine dependence and is
2863 not representable as a distance vector. */
2864 non_affine_dependence_relation (ddr);
2872 /* Return true when the DDR contains only constant access functions. */
2875 constant_access_functions (const struct data_dependence_relation *ddr)
2879 for (i = 0; i < DDR_NUM_SUBSCRIPTS (ddr); i++)
2880 if (!evolution_function_is_constant_p (DR_ACCESS_FN (DDR_A (ddr), i))
2881 || !evolution_function_is_constant_p (DR_ACCESS_FN (DDR_B (ddr), i)))
2887 /* Helper function for the case where DDR_A and DDR_B are the same
2888 multivariate access function with a constant step. For an example
2892 add_multivariate_self_dist (struct data_dependence_relation *ddr, tree c_2)
2895 tree c_1 = CHREC_LEFT (c_2);
2896 tree c_0 = CHREC_LEFT (c_1);
2897 lambda_vector dist_v;
2900 /* Polynomials with more than 2 variables are not handled yet. When
2901 the evolution steps are parameters, it is not possible to
2902 represent the dependence using classical distance vectors. */
2903 if (TREE_CODE (c_0) != INTEGER_CST
2904 || TREE_CODE (CHREC_RIGHT (c_1)) != INTEGER_CST
2905 || TREE_CODE (CHREC_RIGHT (c_2)) != INTEGER_CST)
2907 DDR_AFFINE_P (ddr) = false;
2911 x_2 = index_in_loop_nest (CHREC_VARIABLE (c_2), DDR_LOOP_NEST (ddr));
2912 x_1 = index_in_loop_nest (CHREC_VARIABLE (c_1), DDR_LOOP_NEST (ddr));
2914 /* For "{{0, +, 2}_1, +, 3}_2" the distance vector is (3, -2). */
2915 dist_v = lambda_vector_new (DDR_NB_LOOPS (ddr));
2916 v1 = int_cst_value (CHREC_RIGHT (c_1));
2917 v2 = int_cst_value (CHREC_RIGHT (c_2));
2930 save_dist_v (ddr, dist_v);
2932 add_outer_distances (ddr, dist_v, x_1);
2935 /* Helper function for the case where DDR_A and DDR_B are the same
2936 access functions. */
2939 add_other_self_distances (struct data_dependence_relation *ddr)
2941 lambda_vector dist_v;
2943 int index_carry = DDR_NB_LOOPS (ddr);
2945 for (i = 0; i < DDR_NUM_SUBSCRIPTS (ddr); i++)
2947 tree access_fun = DR_ACCESS_FN (DDR_A (ddr), i);
2949 if (TREE_CODE (access_fun) == POLYNOMIAL_CHREC)
2951 if (!evolution_function_is_univariate_p (access_fun))
2953 if (DDR_NUM_SUBSCRIPTS (ddr) != 1)
2955 DDR_ARE_DEPENDENT (ddr) = chrec_dont_know;
2959 access_fun = DR_ACCESS_FN (DDR_A (ddr), 0);
2961 if (TREE_CODE (CHREC_LEFT (access_fun)) == POLYNOMIAL_CHREC)
2962 add_multivariate_self_dist (ddr, access_fun);
2964 /* The evolution step is not constant: it varies in
2965 the outer loop, so this cannot be represented by a
2966 distance vector. For example in pr34635.c the
2967 evolution is {0, +, {0, +, 4}_1}_2. */
2968 DDR_AFFINE_P (ddr) = false;
2973 index_carry = MIN (index_carry,
2974 index_in_loop_nest (CHREC_VARIABLE (access_fun),
2975 DDR_LOOP_NEST (ddr)));
2979 dist_v = lambda_vector_new (DDR_NB_LOOPS (ddr));
2980 add_outer_distances (ddr, dist_v, index_carry);
2984 insert_innermost_unit_dist_vector (struct data_dependence_relation *ddr)
2986 lambda_vector dist_v = lambda_vector_new (DDR_NB_LOOPS (ddr));
2988 dist_v[DDR_INNER_LOOP (ddr)] = 1;
2989 save_dist_v (ddr, dist_v);
2992 /* Adds a unit distance vector to DDR when there is a 0 overlap. This
2993 is the case for example when access functions are the same and
2994 equal to a constant, as in:
3001 in which case the distance vectors are (0) and (1). */
3004 add_distance_for_zero_overlaps (struct data_dependence_relation *ddr)
3008 for (i = 0; i < DDR_NUM_SUBSCRIPTS (ddr); i++)
3010 subscript_p sub = DDR_SUBSCRIPT (ddr, i);
3011 conflict_function *ca = SUB_CONFLICTS_IN_A (sub);
3012 conflict_function *cb = SUB_CONFLICTS_IN_B (sub);
3014 for (j = 0; j < ca->n; j++)
3015 if (affine_function_zero_p (ca->fns[j]))
3017 insert_innermost_unit_dist_vector (ddr);
3021 for (j = 0; j < cb->n; j++)
3022 if (affine_function_zero_p (cb->fns[j]))
3024 insert_innermost_unit_dist_vector (ddr);
3030 /* Compute the classic per loop distance vector. DDR is the data
3031 dependence relation to build a vector from. Return false when fail
3032 to represent the data dependence as a distance vector. */
3035 build_classic_dist_vector (struct data_dependence_relation *ddr,
3036 struct loop *loop_nest)
3038 bool init_b = false;
3039 int index_carry = DDR_NB_LOOPS (ddr);
3040 lambda_vector dist_v;
3042 if (DDR_ARE_DEPENDENT (ddr) != NULL_TREE)
3045 if (same_access_functions (ddr))
3047 /* Save the 0 vector. */
3048 dist_v = lambda_vector_new (DDR_NB_LOOPS (ddr));
3049 save_dist_v (ddr, dist_v);
3051 if (constant_access_functions (ddr))
3052 add_distance_for_zero_overlaps (ddr);
3054 if (DDR_NB_LOOPS (ddr) > 1)
3055 add_other_self_distances (ddr);
3060 dist_v = lambda_vector_new (DDR_NB_LOOPS (ddr));
3061 if (!build_classic_dist_vector_1 (ddr, DDR_A (ddr), DDR_B (ddr),
3062 dist_v, &init_b, &index_carry))
3065 /* Save the distance vector if we initialized one. */
3068 /* Verify a basic constraint: classic distance vectors should
3069 always be lexicographically positive.
3071 Data references are collected in the order of execution of
3072 the program, thus for the following loop
3074 | for (i = 1; i < 100; i++)
3075 | for (j = 1; j < 100; j++)
3077 | t = T[j+1][i-1]; // A
3078 | T[j][i] = t + 2; // B
3081 references are collected following the direction of the wind:
3082 A then B. The data dependence tests are performed also
3083 following this order, such that we're looking at the distance
3084 separating the elements accessed by A from the elements later
3085 accessed by B. But in this example, the distance returned by
3086 test_dep (A, B) is lexicographically negative (-1, 1), that
3087 means that the access A occurs later than B with respect to
3088 the outer loop, ie. we're actually looking upwind. In this
3089 case we solve test_dep (B, A) looking downwind to the
3090 lexicographically positive solution, that returns the
3091 distance vector (1, -1). */
3092 if (!lambda_vector_lexico_pos (dist_v, DDR_NB_LOOPS (ddr)))
3094 lambda_vector save_v = lambda_vector_new (DDR_NB_LOOPS (ddr));
3095 if (!subscript_dependence_tester_1 (ddr, DDR_B (ddr), DDR_A (ddr),
3098 compute_subscript_distance (ddr);
3099 if (!build_classic_dist_vector_1 (ddr, DDR_B (ddr), DDR_A (ddr),
3100 save_v, &init_b, &index_carry))
3102 save_dist_v (ddr, save_v);
3103 DDR_REVERSED_P (ddr) = true;
3105 /* In this case there is a dependence forward for all the
3108 | for (k = 1; k < 100; k++)
3109 | for (i = 1; i < 100; i++)
3110 | for (j = 1; j < 100; j++)
3112 | t = T[j+1][i-1]; // A
3113 | T[j][i] = t + 2; // B
3121 if (DDR_NB_LOOPS (ddr) > 1)
3123 add_outer_distances (ddr, save_v, index_carry);
3124 add_outer_distances (ddr, dist_v, index_carry);
3129 lambda_vector save_v = lambda_vector_new (DDR_NB_LOOPS (ddr));
3130 lambda_vector_copy (dist_v, save_v, DDR_NB_LOOPS (ddr));
3132 if (DDR_NB_LOOPS (ddr) > 1)
3134 lambda_vector opposite_v = lambda_vector_new (DDR_NB_LOOPS (ddr));
3136 if (!subscript_dependence_tester_1 (ddr, DDR_B (ddr),
3137 DDR_A (ddr), loop_nest))
3139 compute_subscript_distance (ddr);
3140 if (!build_classic_dist_vector_1 (ddr, DDR_B (ddr), DDR_A (ddr),
3141 opposite_v, &init_b,
3145 save_dist_v (ddr, save_v);
3146 add_outer_distances (ddr, dist_v, index_carry);
3147 add_outer_distances (ddr, opposite_v, index_carry);
3150 save_dist_v (ddr, save_v);
3155 /* There is a distance of 1 on all the outer loops: Example:
3156 there is a dependence of distance 1 on loop_1 for the array A.
3162 add_outer_distances (ddr, dist_v,
3163 lambda_vector_first_nz (dist_v,
3164 DDR_NB_LOOPS (ddr), 0));
3167 if (dump_file && (dump_flags & TDF_DETAILS))
3171 fprintf (dump_file, "(build_classic_dist_vector\n");
3172 for (i = 0; i < DDR_NUM_DIST_VECTS (ddr); i++)
3174 fprintf (dump_file, " dist_vector = (");
3175 print_lambda_vector (dump_file, DDR_DIST_VECT (ddr, i),
3176 DDR_NB_LOOPS (ddr));
3177 fprintf (dump_file, " )\n");
3179 fprintf (dump_file, ")\n");
3185 /* Return the direction for a given distance.
3186 FIXME: Computing dir this way is suboptimal, since dir can catch
3187 cases that dist is unable to represent. */
3189 static inline enum data_dependence_direction
3190 dir_from_dist (int dist)
3193 return dir_positive;
3195 return dir_negative;
3200 /* Compute the classic per loop direction vector. DDR is the data
3201 dependence relation to build a vector from. */
3204 build_classic_dir_vector (struct data_dependence_relation *ddr)
3207 lambda_vector dist_v;
3209 for (i = 0; VEC_iterate (lambda_vector, DDR_DIST_VECTS (ddr), i, dist_v); i++)
3211 lambda_vector dir_v = lambda_vector_new (DDR_NB_LOOPS (ddr));
3213 for (j = 0; j < DDR_NB_LOOPS (ddr); j++)
3214 dir_v[j] = dir_from_dist (dist_v[j]);
3216 save_dir_v (ddr, dir_v);
3220 /* Helper function. Returns true when there is a dependence between
3221 data references DRA and DRB. */
3224 subscript_dependence_tester_1 (struct data_dependence_relation *ddr,
3225 struct data_reference *dra,
3226 struct data_reference *drb,
3227 struct loop *loop_nest)
3230 tree last_conflicts;
3231 struct subscript *subscript;
3233 for (i = 0; VEC_iterate (subscript_p, DDR_SUBSCRIPTS (ddr), i, subscript);
3236 conflict_function *overlaps_a, *overlaps_b;
3238 analyze_overlapping_iterations (DR_ACCESS_FN (dra, i),
3239 DR_ACCESS_FN (drb, i),
3240 &overlaps_a, &overlaps_b,
3241 &last_conflicts, loop_nest);
3243 if (CF_NOT_KNOWN_P (overlaps_a)
3244 || CF_NOT_KNOWN_P (overlaps_b))
3246 finalize_ddr_dependent (ddr, chrec_dont_know);
3247 dependence_stats.num_dependence_undetermined++;
3248 free_conflict_function (overlaps_a);
3249 free_conflict_function (overlaps_b);
3253 else if (CF_NO_DEPENDENCE_P (overlaps_a)
3254 || CF_NO_DEPENDENCE_P (overlaps_b))
3256 finalize_ddr_dependent (ddr, chrec_known);
3257 dependence_stats.num_dependence_independent++;
3258 free_conflict_function (overlaps_a);
3259 free_conflict_function (overlaps_b);
3265 if (SUB_CONFLICTS_IN_A (subscript))
3266 free_conflict_function (SUB_CONFLICTS_IN_A (subscript));
3267 if (SUB_CONFLICTS_IN_B (subscript))
3268 free_conflict_function (SUB_CONFLICTS_IN_B (subscript));
3270 SUB_CONFLICTS_IN_A (subscript) = overlaps_a;
3271 SUB_CONFLICTS_IN_B (subscript) = overlaps_b;
3272 SUB_LAST_CONFLICT (subscript) = last_conflicts;
3279 /* Computes the conflicting iterations in LOOP_NEST, and initialize DDR. */
3282 subscript_dependence_tester (struct data_dependence_relation *ddr,
3283 struct loop *loop_nest)
3286 if (dump_file && (dump_flags & TDF_DETAILS))
3287 fprintf (dump_file, "(subscript_dependence_tester \n");
3289 if (subscript_dependence_tester_1 (ddr, DDR_A (ddr), DDR_B (ddr), loop_nest))
3290 dependence_stats.num_dependence_dependent++;
3292 compute_subscript_distance (ddr);
3293 if (build_classic_dist_vector (ddr, loop_nest))
3294 build_classic_dir_vector (ddr);
3296 if (dump_file && (dump_flags & TDF_DETAILS))
3297 fprintf (dump_file, ")\n");
3300 /* Returns true when all the access functions of A are affine or
3301 constant with respect to LOOP_NEST. */
3304 access_functions_are_affine_or_constant_p (const struct data_reference *a,
3305 const struct loop *loop_nest)
3308 VEC(tree,heap) *fns = DR_ACCESS_FNS (a);
3311 for (i = 0; VEC_iterate (tree, fns, i, t); i++)
3312 if (!evolution_function_is_invariant_p (t, loop_nest->num)
3313 && !evolution_function_is_affine_multivariate_p (t, loop_nest->num))
3319 /* Return true if we can create an affine data-ref for OP in STMT. */
3322 stmt_simple_memref_p (struct loop *loop, gimple stmt, tree op)
3324 data_reference_p dr;
3327 dr = create_data_ref (loop, op, stmt, true);
3328 if (!access_functions_are_affine_or_constant_p (dr, loop))
3335 /* Initializes an equation for an OMEGA problem using the information
3336 contained in the ACCESS_FUN. Returns true when the operation
3339 PB is the omega constraint system.
3340 EQ is the number of the equation to be initialized.
3341 OFFSET is used for shifting the variables names in the constraints:
3342 a constrain is composed of 2 * the number of variables surrounding
3343 dependence accesses. OFFSET is set either to 0 for the first n variables,
3344 then it is set to n.
3345 ACCESS_FUN is expected to be an affine chrec. */
3348 init_omega_eq_with_af (omega_pb pb, unsigned eq,
3349 unsigned int offset, tree access_fun,
3350 struct data_dependence_relation *ddr)
3352 switch (TREE_CODE (access_fun))
3354 case POLYNOMIAL_CHREC:
3356 tree left = CHREC_LEFT (access_fun);
3357 tree right = CHREC_RIGHT (access_fun);
3358 int var = CHREC_VARIABLE (access_fun);
3361 if (TREE_CODE (right) != INTEGER_CST)
3364 var_idx = index_in_loop_nest (var, DDR_LOOP_NEST (ddr));
3365 pb->eqs[eq].coef[offset + var_idx + 1] = int_cst_value (right);
3367 /* Compute the innermost loop index. */
3368 DDR_INNER_LOOP (ddr) = MAX (DDR_INNER_LOOP (ddr), var_idx);
3371 pb->eqs[eq].coef[var_idx + DDR_NB_LOOPS (ddr) + 1]
3372 += int_cst_value (right);
3374 switch (TREE_CODE (left))
3376 case POLYNOMIAL_CHREC:
3377 return init_omega_eq_with_af (pb, eq, offset, left, ddr);
3380 pb->eqs[eq].coef[0] += int_cst_value (left);
3389 pb->eqs[eq].coef[0] += int_cst_value (access_fun);
3397 /* As explained in the comments preceding init_omega_for_ddr, we have
3398 to set up a system for each loop level, setting outer loops
3399 variation to zero, and current loop variation to positive or zero.
3400 Save each lexico positive distance vector. */
3403 omega_extract_distance_vectors (omega_pb pb,
3404 struct data_dependence_relation *ddr)
3408 struct loop *loopi, *loopj;
3409 enum omega_result res;
3411 /* Set a new problem for each loop in the nest. The basis is the
3412 problem that we have initialized until now. On top of this we
3413 add new constraints. */
3414 for (i = 0; i <= DDR_INNER_LOOP (ddr)
3415 && VEC_iterate (loop_p, DDR_LOOP_NEST (ddr), i, loopi); i++)
3418 omega_pb copy = omega_alloc_problem (2 * DDR_NB_LOOPS (ddr),
3419 DDR_NB_LOOPS (ddr));
3421 omega_copy_problem (copy, pb);
3423 /* For all the outer loops "loop_j", add "dj = 0". */
3425 j < i && VEC_iterate (loop_p, DDR_LOOP_NEST (ddr), j, loopj); j++)
3427 eq = omega_add_zero_eq (copy, omega_black);
3428 copy->eqs[eq].coef[j + 1] = 1;
3431 /* For "loop_i", add "0 <= di". */
3432 geq = omega_add_zero_geq (copy, omega_black);
3433 copy->geqs[geq].coef[i + 1] = 1;
3435 /* Reduce the constraint system, and test that the current
3436 problem is feasible. */
3437 res = omega_simplify_problem (copy);
3438 if (res == omega_false
3439 || res == omega_unknown
3440 || copy->num_geqs > (int) DDR_NB_LOOPS (ddr))
3443 for (eq = 0; eq < copy->num_subs; eq++)
3444 if (copy->subs[eq].key == (int) i + 1)
3446 dist = copy->subs[eq].coef[0];
3452 /* Reinitialize problem... */
3453 omega_copy_problem (copy, pb);
3455 j < i && VEC_iterate (loop_p, DDR_LOOP_NEST (ddr), j, loopj); j++)
3457 eq = omega_add_zero_eq (copy, omega_black);
3458 copy->eqs[eq].coef[j + 1] = 1;
3461 /* ..., but this time "di = 1". */
3462 eq = omega_add_zero_eq (copy, omega_black);
3463 copy->eqs[eq].coef[i + 1] = 1;
3464 copy->eqs[eq].coef[0] = -1;
3466 res = omega_simplify_problem (copy);
3467 if (res == omega_false
3468 || res == omega_unknown
3469 || copy->num_geqs > (int) DDR_NB_LOOPS (ddr))
3472 for (eq = 0; eq < copy->num_subs; eq++)
3473 if (copy->subs[eq].key == (int) i + 1)
3475 dist = copy->subs[eq].coef[0];
3481 /* Save the lexicographically positive distance vector. */
3484 lambda_vector dist_v = lambda_vector_new (DDR_NB_LOOPS (ddr));
3485 lambda_vector dir_v = lambda_vector_new (DDR_NB_LOOPS (ddr));
3489 for (eq = 0; eq < copy->num_subs; eq++)
3490 if (copy->subs[eq].key > 0)
3492 dist = copy->subs[eq].coef[0];
3493 dist_v[copy->subs[eq].key - 1] = dist;
3496 for (j = 0; j < DDR_NB_LOOPS (ddr); j++)
3497 dir_v[j] = dir_from_dist (dist_v[j]);
3499 save_dist_v (ddr, dist_v);
3500 save_dir_v (ddr, dir_v);
3504 omega_free_problem (copy);
3508 /* This is called for each subscript of a tuple of data references:
3509 insert an equality for representing the conflicts. */
3512 omega_setup_subscript (tree access_fun_a, tree access_fun_b,
3513 struct data_dependence_relation *ddr,
3514 omega_pb pb, bool *maybe_dependent)
3517 tree type = signed_type_for_types (TREE_TYPE (access_fun_a),
3518 TREE_TYPE (access_fun_b));
3519 tree fun_a = chrec_convert (type, access_fun_a, NULL);
3520 tree fun_b = chrec_convert (type, access_fun_b, NULL);
3521 tree difference = chrec_fold_minus (type, fun_a, fun_b);
3523 /* When the fun_a - fun_b is not constant, the dependence is not
3524 captured by the classic distance vector representation. */
3525 if (TREE_CODE (difference) != INTEGER_CST)
3529 if (ziv_subscript_p (fun_a, fun_b) && !integer_zerop (difference))
3531 /* There is no dependence. */
3532 *maybe_dependent = false;
3536 fun_b = chrec_fold_multiply (type, fun_b, integer_minus_one_node);
3538 eq = omega_add_zero_eq (pb, omega_black);
3539 if (!init_omega_eq_with_af (pb, eq, DDR_NB_LOOPS (ddr), fun_a, ddr)
3540 || !init_omega_eq_with_af (pb, eq, 0, fun_b, ddr))
3541 /* There is probably a dependence, but the system of
3542 constraints cannot be built: answer "don't know". */
3546 if (DDR_NB_LOOPS (ddr) != 0 && pb->eqs[eq].coef[0]
3547 && !int_divides_p (lambda_vector_gcd
3548 ((lambda_vector) &(pb->eqs[eq].coef[1]),
3549 2 * DDR_NB_LOOPS (ddr)),
3550 pb->eqs[eq].coef[0]))
3552 /* There is no dependence. */
3553 *maybe_dependent = false;
3560 /* Helper function, same as init_omega_for_ddr but specialized for
3561 data references A and B. */
3564 init_omega_for_ddr_1 (struct data_reference *dra, struct data_reference *drb,
3565 struct data_dependence_relation *ddr,
3566 omega_pb pb, bool *maybe_dependent)
3571 unsigned nb_loops = DDR_NB_LOOPS (ddr);
3573 /* Insert an equality per subscript. */
3574 for (i = 0; i < DDR_NUM_SUBSCRIPTS (ddr); i++)
3576 if (!omega_setup_subscript (DR_ACCESS_FN (dra, i), DR_ACCESS_FN (drb, i),
3577 ddr, pb, maybe_dependent))
3579 else if (*maybe_dependent == false)
3581 /* There is no dependence. */
3582 DDR_ARE_DEPENDENT (ddr) = chrec_known;
3587 /* Insert inequalities: constraints corresponding to the iteration
3588 domain, i.e. the loops surrounding the references "loop_x" and
3589 the distance variables "dx". The layout of the OMEGA
3590 representation is as follows:
3591 - coef[0] is the constant
3592 - coef[1..nb_loops] are the protected variables that will not be
3593 removed by the solver: the "dx"
3594 - coef[nb_loops + 1, 2*nb_loops] are the loop variables: "loop_x".
3596 for (i = 0; i <= DDR_INNER_LOOP (ddr)
3597 && VEC_iterate (loop_p, DDR_LOOP_NEST (ddr), i, loopi); i++)
3599 HOST_WIDE_INT nbi = estimated_loop_iterations_int (loopi, false);
3602 ineq = omega_add_zero_geq (pb, omega_black);
3603 pb->geqs[ineq].coef[i + nb_loops + 1] = 1;
3605 /* 0 <= loop_x + dx */
3606 ineq = omega_add_zero_geq (pb, omega_black);
3607 pb->geqs[ineq].coef[i + nb_loops + 1] = 1;
3608 pb->geqs[ineq].coef[i + 1] = 1;
3612 /* loop_x <= nb_iters */
3613 ineq = omega_add_zero_geq (pb, omega_black);
3614 pb->geqs[ineq].coef[i + nb_loops + 1] = -1;
3615 pb->geqs[ineq].coef[0] = nbi;
3617 /* loop_x + dx <= nb_iters */
3618 ineq = omega_add_zero_geq (pb, omega_black);
3619 pb->geqs[ineq].coef[i + nb_loops + 1] = -1;
3620 pb->geqs[ineq].coef[i + 1] = -1;
3621 pb->geqs[ineq].coef[0] = nbi;
3623 /* A step "dx" bigger than nb_iters is not feasible, so
3624 add "0 <= nb_iters + dx", */
3625 ineq = omega_add_zero_geq (pb, omega_black);
3626 pb->geqs[ineq].coef[i + 1] = 1;
3627 pb->geqs[ineq].coef[0] = nbi;
3628 /* and "dx <= nb_iters". */
3629 ineq = omega_add_zero_geq (pb, omega_black);
3630 pb->geqs[ineq].coef[i + 1] = -1;
3631 pb->geqs[ineq].coef[0] = nbi;
3635 omega_extract_distance_vectors (pb, ddr);
3640 /* Sets up the Omega dependence problem for the data dependence
3641 relation DDR. Returns false when the constraint system cannot be
3642 built, ie. when the test answers "don't know". Returns true
3643 otherwise, and when independence has been proved (using one of the
3644 trivial dependence test), set MAYBE_DEPENDENT to false, otherwise
3645 set MAYBE_DEPENDENT to true.
3647 Example: for setting up the dependence system corresponding to the
3648 conflicting accesses
3653 | ... A[2*j, 2*(i + j)]
3657 the following constraints come from the iteration domain:
3664 where di, dj are the distance variables. The constraints
3665 representing the conflicting elements are:
3668 i + 1 = 2 * (i + di + j + dj)
3670 For asking that the resulting distance vector (di, dj) be
3671 lexicographically positive, we insert the constraint "di >= 0". If
3672 "di = 0" in the solution, we fix that component to zero, and we
3673 look at the inner loops: we set a new problem where all the outer
3674 loop distances are zero, and fix this inner component to be
3675 positive. When one of the components is positive, we save that
3676 distance, and set a new problem where the distance on this loop is
3677 zero, searching for other distances in the inner loops. Here is
3678 the classic example that illustrates that we have to set for each
3679 inner loop a new problem:
3687 we have to save two distances (1, 0) and (0, 1).
3689 Given two array references, refA and refB, we have to set the
3690 dependence problem twice, refA vs. refB and refB vs. refA, and we
3691 cannot do a single test, as refB might occur before refA in the
3692 inner loops, and the contrary when considering outer loops: ex.
3697 | T[{1,+,1}_2][{1,+,1}_1] // refA
3698 | T[{2,+,1}_2][{0,+,1}_1] // refB
3703 refB touches the elements in T before refA, and thus for the same
3704 loop_0 refB precedes refA: ie. the distance vector (0, 1, -1)
3705 but for successive loop_0 iterations, we have (1, -1, 1)
3707 The Omega solver expects the distance variables ("di" in the
3708 previous example) to come first in the constraint system (as
3709 variables to be protected, or "safe" variables), the constraint
3710 system is built using the following layout:
3712 "cst | distance vars | index vars".
3716 init_omega_for_ddr (struct data_dependence_relation *ddr,
3717 bool *maybe_dependent)
3722 *maybe_dependent = true;
3724 if (same_access_functions (ddr))
3727 lambda_vector dir_v;
3729 /* Save the 0 vector. */
3730 save_dist_v (ddr, lambda_vector_new (DDR_NB_LOOPS (ddr)));
3731 dir_v = lambda_vector_new (DDR_NB_LOOPS (ddr));
3732 for (j = 0; j < DDR_NB_LOOPS (ddr); j++)
3733 dir_v[j] = dir_equal;
3734 save_dir_v (ddr, dir_v);
3736 /* Save the dependences carried by outer loops. */
3737 pb = omega_alloc_problem (2 * DDR_NB_LOOPS (ddr), DDR_NB_LOOPS (ddr));
3738 res = init_omega_for_ddr_1 (DDR_A (ddr), DDR_B (ddr), ddr, pb,
3740 omega_free_problem (pb);
3744 /* Omega expects the protected variables (those that have to be kept
3745 after elimination) to appear first in the constraint system.
3746 These variables are the distance variables. In the following
3747 initialization we declare NB_LOOPS safe variables, and the total
3748 number of variables for the constraint system is 2*NB_LOOPS. */
3749 pb = omega_alloc_problem (2 * DDR_NB_LOOPS (ddr), DDR_NB_LOOPS (ddr));
3750 res = init_omega_for_ddr_1 (DDR_A (ddr), DDR_B (ddr), ddr, pb,
3752 omega_free_problem (pb);
3754 /* Stop computation if not decidable, or no dependence. */
3755 if (res == false || *maybe_dependent == false)
3758 pb = omega_alloc_problem (2 * DDR_NB_LOOPS (ddr), DDR_NB_LOOPS (ddr));
3759 res = init_omega_for_ddr_1 (DDR_B (ddr), DDR_A (ddr), ddr, pb,
3761 omega_free_problem (pb);
3766 /* Return true when DDR contains the same information as that stored
3767 in DIR_VECTS and in DIST_VECTS, return false otherwise. */
3770 ddr_consistent_p (FILE *file,
3771 struct data_dependence_relation *ddr,
3772 VEC (lambda_vector, heap) *dist_vects,
3773 VEC (lambda_vector, heap) *dir_vects)
3777 /* If dump_file is set, output there. */
3778 if (dump_file && (dump_flags & TDF_DETAILS))
3781 if (VEC_length (lambda_vector, dist_vects) != DDR_NUM_DIST_VECTS (ddr))
3783 lambda_vector b_dist_v;
3784 fprintf (file, "\n(Number of distance vectors differ: Banerjee has %d, Omega has %d.\n",
3785 VEC_length (lambda_vector, dist_vects),
3786 DDR_NUM_DIST_VECTS (ddr));
3788 fprintf (file, "Banerjee dist vectors:\n");
3789 for (i = 0; VEC_iterate (lambda_vector, dist_vects, i, b_dist_v); i++)
3790 print_lambda_vector (file, b_dist_v, DDR_NB_LOOPS (ddr));
3792 fprintf (file, "Omega dist vectors:\n");
3793 for (i = 0; i < DDR_NUM_DIST_VECTS (ddr); i++)
3794 print_lambda_vector (file, DDR_DIST_VECT (ddr, i), DDR_NB_LOOPS (ddr));
3796 fprintf (file, "data dependence relation:\n");
3797 dump_data_dependence_relation (file, ddr);
3799 fprintf (file, ")\n");
3803 if (VEC_length (lambda_vector, dir_vects) != DDR_NUM_DIR_VECTS (ddr))
3805 fprintf (file, "\n(Number of direction vectors differ: Banerjee has %d, Omega has %d.)\n",
3806 VEC_length (lambda_vector, dir_vects),
3807 DDR_NUM_DIR_VECTS (ddr));
3811 for (i = 0; i < DDR_NUM_DIST_VECTS (ddr); i++)
3813 lambda_vector a_dist_v;
3814 lambda_vector b_dist_v = DDR_DIST_VECT (ddr, i);
3816 /* Distance vectors are not ordered in the same way in the DDR
3817 and in the DIST_VECTS: search for a matching vector. */
3818 for (j = 0; VEC_iterate (lambda_vector, dist_vects, j, a_dist_v); j++)
3819 if (lambda_vector_equal (a_dist_v, b_dist_v, DDR_NB_LOOPS (ddr)))
3822 if (j == VEC_length (lambda_vector, dist_vects))
3824 fprintf (file, "\n(Dist vectors from the first dependence analyzer:\n");
3825 print_dist_vectors (file, dist_vects, DDR_NB_LOOPS (ddr));
3826 fprintf (file, "not found in Omega dist vectors:\n");
3827 print_dist_vectors (file, DDR_DIST_VECTS (ddr), DDR_NB_LOOPS (ddr));
3828 fprintf (file, "data dependence relation:\n");
3829 dump_data_dependence_relation (file, ddr);
3830 fprintf (file, ")\n");
3834 for (i = 0; i < DDR_NUM_DIR_VECTS (ddr); i++)
3836 lambda_vector a_dir_v;
3837 lambda_vector b_dir_v = DDR_DIR_VECT (ddr, i);
3839 /* Direction vectors are not ordered in the same way in the DDR
3840 and in the DIR_VECTS: search for a matching vector. */
3841 for (j = 0; VEC_iterate (lambda_vector, dir_vects, j, a_dir_v); j++)
3842 if (lambda_vector_equal (a_dir_v, b_dir_v, DDR_NB_LOOPS (ddr)))
3845 if (j == VEC_length (lambda_vector, dist_vects))
3847 fprintf (file, "\n(Dir vectors from the first dependence analyzer:\n");
3848 print_dir_vectors (file, dir_vects, DDR_NB_LOOPS (ddr));
3849 fprintf (file, "not found in Omega dir vectors:\n");
3850 print_dir_vectors (file, DDR_DIR_VECTS (ddr), DDR_NB_LOOPS (ddr));
3851 fprintf (file, "data dependence relation:\n");
3852 dump_data_dependence_relation (file, ddr);
3853 fprintf (file, ")\n");
3860 /* This computes the affine dependence relation between A and B with
3861 respect to LOOP_NEST. CHREC_KNOWN is used for representing the
3862 independence between two accesses, while CHREC_DONT_KNOW is used
3863 for representing the unknown relation.
3865 Note that it is possible to stop the computation of the dependence
3866 relation the first time we detect a CHREC_KNOWN element for a given
3870 compute_affine_dependence (struct data_dependence_relation *ddr,
3871 struct loop *loop_nest)
3873 struct data_reference *dra = DDR_A (ddr);
3874 struct data_reference *drb = DDR_B (ddr);
3876 if (dump_file && (dump_flags & TDF_DETAILS))
3878 fprintf (dump_file, "(compute_affine_dependence\n");
3879 fprintf (dump_file, " (stmt_a = \n");
3880 print_gimple_stmt (dump_file, DR_STMT (dra), 0, 0);
3881 fprintf (dump_file, ")\n (stmt_b = \n");
3882 print_gimple_stmt (dump_file, DR_STMT (drb), 0, 0);
3883 fprintf (dump_file, ")\n");
3886 /* Analyze only when the dependence relation is not yet known. */
3887 if (DDR_ARE_DEPENDENT (ddr) == NULL_TREE
3888 && !DDR_SELF_REFERENCE (ddr))
3890 dependence_stats.num_dependence_tests++;
3892 if (access_functions_are_affine_or_constant_p (dra, loop_nest)
3893 && access_functions_are_affine_or_constant_p (drb, loop_nest))
3895 if (flag_check_data_deps)
3897 /* Compute the dependences using the first algorithm. */
3898 subscript_dependence_tester (ddr, loop_nest);
3900 if (dump_file && (dump_flags & TDF_DETAILS))
3902 fprintf (dump_file, "\n\nBanerjee Analyzer\n");
3903 dump_data_dependence_relation (dump_file, ddr);
3906 if (DDR_ARE_DEPENDENT (ddr) == NULL_TREE)
3908 bool maybe_dependent;
3909 VEC (lambda_vector, heap) *dir_vects, *dist_vects;
3911 /* Save the result of the first DD analyzer. */
3912 dist_vects = DDR_DIST_VECTS (ddr);
3913 dir_vects = DDR_DIR_VECTS (ddr);
3915 /* Reset the information. */
3916 DDR_DIST_VECTS (ddr) = NULL;
3917 DDR_DIR_VECTS (ddr) = NULL;
3919 /* Compute the same information using Omega. */
3920 if (!init_omega_for_ddr (ddr, &maybe_dependent))
3921 goto csys_dont_know;
3923 if (dump_file && (dump_flags & TDF_DETAILS))
3925 fprintf (dump_file, "Omega Analyzer\n");
3926 dump_data_dependence_relation (dump_file, ddr);
3929 /* Check that we get the same information. */
3930 if (maybe_dependent)
3931 gcc_assert (ddr_consistent_p (stderr, ddr, dist_vects,
3936 subscript_dependence_tester (ddr, loop_nest);
3939 /* As a last case, if the dependence cannot be determined, or if
3940 the dependence is considered too difficult to determine, answer
3945 dependence_stats.num_dependence_undetermined++;
3947 if (dump_file && (dump_flags & TDF_DETAILS))
3949 fprintf (dump_file, "Data ref a:\n");
3950 dump_data_reference (dump_file, dra);
3951 fprintf (dump_file, "Data ref b:\n");
3952 dump_data_reference (dump_file, drb);
3953 fprintf (dump_file, "affine dependence test not usable: access function not affine or constant.\n");
3955 finalize_ddr_dependent (ddr, chrec_dont_know);
3959 if (dump_file && (dump_flags & TDF_DETAILS))
3960 fprintf (dump_file, ")\n");
3963 /* This computes the dependence relation for the same data
3964 reference into DDR. */
3967 compute_self_dependence (struct data_dependence_relation *ddr)
3970 struct subscript *subscript;
3972 if (DDR_ARE_DEPENDENT (ddr) != NULL_TREE)
3975 for (i = 0; VEC_iterate (subscript_p, DDR_SUBSCRIPTS (ddr), i, subscript);
3978 if (SUB_CONFLICTS_IN_A (subscript))
3979 free_conflict_function (SUB_CONFLICTS_IN_A (subscript));
3980 if (SUB_CONFLICTS_IN_B (subscript))
3981 free_conflict_function (SUB_CONFLICTS_IN_B (subscript));
3983 /* The accessed index overlaps for each iteration. */
3984 SUB_CONFLICTS_IN_A (subscript)
3985 = conflict_fn (1, affine_fn_cst (integer_zero_node));
3986 SUB_CONFLICTS_IN_B (subscript)
3987 = conflict_fn (1, affine_fn_cst (integer_zero_node));
3988 SUB_LAST_CONFLICT (subscript) = chrec_dont_know;
3991 /* The distance vector is the zero vector. */
3992 save_dist_v (ddr, lambda_vector_new (DDR_NB_LOOPS (ddr)));
3993 save_dir_v (ddr, lambda_vector_new (DDR_NB_LOOPS (ddr)));
3996 /* Compute in DEPENDENCE_RELATIONS the data dependence graph for all
3997 the data references in DATAREFS, in the LOOP_NEST. When
3998 COMPUTE_SELF_AND_RR is FALSE, don't compute read-read and self
4002 compute_all_dependences (VEC (data_reference_p, heap) *datarefs,
4003 VEC (ddr_p, heap) **dependence_relations,
4004 VEC (loop_p, heap) *loop_nest,
4005 bool compute_self_and_rr)
4007 struct data_dependence_relation *ddr;
4008 struct data_reference *a, *b;
4011 for (i = 0; VEC_iterate (data_reference_p, datarefs, i, a); i++)
4012 for (j = i + 1; VEC_iterate (data_reference_p, datarefs, j, b); j++)
4013 if (!DR_IS_READ (a) || !DR_IS_READ (b) || compute_self_and_rr)
4015 ddr = initialize_data_dependence_relation (a, b, loop_nest);
4016 VEC_safe_push (ddr_p, heap, *dependence_relations, ddr);
4017 compute_affine_dependence (ddr, VEC_index (loop_p, loop_nest, 0));
4020 if (compute_self_and_rr)
4021 for (i = 0; VEC_iterate (data_reference_p, datarefs, i, a); i++)
4023 ddr = initialize_data_dependence_relation (a, a, loop_nest);
4024 VEC_safe_push (ddr_p, heap, *dependence_relations, ddr);
4025 compute_self_dependence (ddr);
4029 /* Stores the locations of memory references in STMT to REFERENCES. Returns
4030 true if STMT clobbers memory, false otherwise. */
4033 get_references_in_stmt (gimple stmt, VEC (data_ref_loc, heap) **references)
4035 bool clobbers_memory = false;
4038 enum gimple_code stmt_code = gimple_code (stmt);
4042 /* ASM_EXPR and CALL_EXPR may embed arbitrary side effects.
4043 Calls have side-effects, except those to const or pure
4045 if ((stmt_code == GIMPLE_CALL
4046 && !(gimple_call_flags (stmt) & (ECF_CONST | ECF_PURE)))
4047 || (stmt_code == GIMPLE_ASM
4048 && gimple_asm_volatile_p (stmt)))
4049 clobbers_memory = true;
4051 if (ZERO_SSA_OPERANDS (stmt, SSA_OP_ALL_VIRTUALS))
4052 return clobbers_memory;
4054 if (stmt_code == GIMPLE_ASSIGN)
4057 op0 = gimple_assign_lhs_ptr (stmt);
4058 op1 = gimple_assign_rhs1_ptr (stmt);
4061 || (REFERENCE_CLASS_P (*op1)
4062 && (base = get_base_address (*op1))
4063 && TREE_CODE (base) != SSA_NAME))
4065 ref = VEC_safe_push (data_ref_loc, heap, *references, NULL);
4067 ref->is_read = true;
4071 || (REFERENCE_CLASS_P (*op0) && get_base_address (*op0)))
4073 ref = VEC_safe_push (data_ref_loc, heap, *references, NULL);
4075 ref->is_read = false;
4078 else if (stmt_code == GIMPLE_CALL)
4080 unsigned i, n = gimple_call_num_args (stmt);
4082 for (i = 0; i < n; i++)
4084 op0 = gimple_call_arg_ptr (stmt, i);
4087 || (REFERENCE_CLASS_P (*op0) && get_base_address (*op0)))
4089 ref = VEC_safe_push (data_ref_loc, heap, *references, NULL);
4091 ref->is_read = true;
4096 return clobbers_memory;
4099 /* Stores the data references in STMT to DATAREFS. If there is an unanalyzable
4100 reference, returns false, otherwise returns true. NEST is the outermost
4101 loop of the loop nest in which the references should be analyzed. */
4104 find_data_references_in_stmt (struct loop *nest, gimple stmt,
4105 VEC (data_reference_p, heap) **datarefs)
4108 VEC (data_ref_loc, heap) *references;
4111 data_reference_p dr;
4113 if (get_references_in_stmt (stmt, &references))
4115 VEC_free (data_ref_loc, heap, references);
4119 for (i = 0; VEC_iterate (data_ref_loc, references, i, ref); i++)
4121 dr = create_data_ref (nest, *ref->pos, stmt, ref->is_read);
4122 gcc_assert (dr != NULL);
4124 /* FIXME -- data dependence analysis does not work correctly for objects with
4125 invariant addresses. Let us fail here until the problem is fixed. */
4126 if (dr_address_invariant_p (dr))
4129 if (dump_file && (dump_flags & TDF_DETAILS))
4130 fprintf (dump_file, "\tFAILED as dr address is invariant\n");
4135 VEC_safe_push (data_reference_p, heap, *datarefs, dr);
4137 VEC_free (data_ref_loc, heap, references);
4141 /* Search the data references in LOOP, and record the information into
4142 DATAREFS. Returns chrec_dont_know when failing to analyze a
4143 difficult case, returns NULL_TREE otherwise.
4145 TODO: This function should be made smarter so that it can handle address
4146 arithmetic as if they were array accesses, etc. */
4149 find_data_references_in_loop (struct loop *loop,
4150 VEC (data_reference_p, heap) **datarefs)
4152 basic_block bb, *bbs;
4154 gimple_stmt_iterator bsi;
4156 bbs = get_loop_body_in_dom_order (loop);
4158 for (i = 0; i < loop->num_nodes; i++)
4162 for (bsi = gsi_start_bb (bb); !gsi_end_p (bsi); gsi_next (&bsi))
4164 gimple stmt = gsi_stmt (bsi);
4166 if (!find_data_references_in_stmt (loop, stmt, datarefs))
4168 struct data_reference *res;
4169 res = XCNEW (struct data_reference);
4170 VEC_safe_push (data_reference_p, heap, *datarefs, res);
4173 return chrec_dont_know;
4182 /* Recursive helper function. */
4185 find_loop_nest_1 (struct loop *loop, VEC (loop_p, heap) **loop_nest)
4187 /* Inner loops of the nest should not contain siblings. Example:
4188 when there are two consecutive loops,
4199 the dependence relation cannot be captured by the distance
4204 VEC_safe_push (loop_p, heap, *loop_nest, loop);
4206 return find_loop_nest_1 (loop->inner, loop_nest);
4210 /* Return false when the LOOP is not well nested. Otherwise return
4211 true and insert in LOOP_NEST the loops of the nest. LOOP_NEST will
4212 contain the loops from the outermost to the innermost, as they will
4213 appear in the classic distance vector. */
4216 find_loop_nest (struct loop *loop, VEC (loop_p, heap) **loop_nest)
4218 VEC_safe_push (loop_p, heap, *loop_nest, loop);
4220 return find_loop_nest_1 (loop->inner, loop_nest);
4224 /* Returns true when the data dependences have been computed, false otherwise.
4225 Given a loop nest LOOP, the following vectors are returned:
4226 DATAREFS is initialized to all the array elements contained in this loop,
4227 DEPENDENCE_RELATIONS contains the relations between the data references.
4228 Compute read-read and self relations if
4229 COMPUTE_SELF_AND_READ_READ_DEPENDENCES is TRUE. */
4232 compute_data_dependences_for_loop (struct loop *loop,
4233 bool compute_self_and_read_read_dependences,
4234 VEC (data_reference_p, heap) **datarefs,
4235 VEC (ddr_p, heap) **dependence_relations)
4238 VEC (loop_p, heap) *vloops = VEC_alloc (loop_p, heap, 3);
4240 memset (&dependence_stats, 0, sizeof (dependence_stats));
4242 /* If the loop nest is not well formed, or one of the data references
4243 is not computable, give up without spending time to compute other
4246 || !find_loop_nest (loop, &vloops)
4247 || find_data_references_in_loop (loop, datarefs) == chrec_dont_know)
4249 struct data_dependence_relation *ddr;
4251 /* Insert a single relation into dependence_relations:
4253 ddr = initialize_data_dependence_relation (NULL, NULL, vloops);
4254 VEC_safe_push (ddr_p, heap, *dependence_relations, ddr);
4258 compute_all_dependences (*datarefs, dependence_relations, vloops,
4259 compute_self_and_read_read_dependences);
4261 if (dump_file && (dump_flags & TDF_STATS))
4263 fprintf (dump_file, "Dependence tester statistics:\n");
4265 fprintf (dump_file, "Number of dependence tests: %d\n",
4266 dependence_stats.num_dependence_tests);
4267 fprintf (dump_file, "Number of dependence tests classified dependent: %d\n",
4268 dependence_stats.num_dependence_dependent);
4269 fprintf (dump_file, "Number of dependence tests classified independent: %d\n",
4270 dependence_stats.num_dependence_independent);
4271 fprintf (dump_file, "Number of undetermined dependence tests: %d\n",
4272 dependence_stats.num_dependence_undetermined);
4274 fprintf (dump_file, "Number of subscript tests: %d\n",
4275 dependence_stats.num_subscript_tests);
4276 fprintf (dump_file, "Number of undetermined subscript tests: %d\n",
4277 dependence_stats.num_subscript_undetermined);
4278 fprintf (dump_file, "Number of same subscript function: %d\n",
4279 dependence_stats.num_same_subscript_function);
4281 fprintf (dump_file, "Number of ziv tests: %d\n",
4282 dependence_stats.num_ziv);
4283 fprintf (dump_file, "Number of ziv tests returning dependent: %d\n",
4284 dependence_stats.num_ziv_dependent);
4285 fprintf (dump_file, "Number of ziv tests returning independent: %d\n",
4286 dependence_stats.num_ziv_independent);
4287 fprintf (dump_file, "Number of ziv tests unimplemented: %d\n",
4288 dependence_stats.num_ziv_unimplemented);
4290 fprintf (dump_file, "Number of siv tests: %d\n",
4291 dependence_stats.num_siv);
4292 fprintf (dump_file, "Number of siv tests returning dependent: %d\n",
4293 dependence_stats.num_siv_dependent);
4294 fprintf (dump_file, "Number of siv tests returning independent: %d\n",
4295 dependence_stats.num_siv_independent);
4296 fprintf (dump_file, "Number of siv tests unimplemented: %d\n",
4297 dependence_stats.num_siv_unimplemented);
4299 fprintf (dump_file, "Number of miv tests: %d\n",
4300 dependence_stats.num_miv);
4301 fprintf (dump_file, "Number of miv tests returning dependent: %d\n",
4302 dependence_stats.num_miv_dependent);
4303 fprintf (dump_file, "Number of miv tests returning independent: %d\n",
4304 dependence_stats.num_miv_independent);
4305 fprintf (dump_file, "Number of miv tests unimplemented: %d\n",
4306 dependence_stats.num_miv_unimplemented);
4312 /* Entry point (for testing only). Analyze all the data references
4313 and the dependence relations in LOOP.
4315 The data references are computed first.
4317 A relation on these nodes is represented by a complete graph. Some
4318 of the relations could be of no interest, thus the relations can be
4321 In the following function we compute all the relations. This is
4322 just a first implementation that is here for:
4323 - for showing how to ask for the dependence relations,
4324 - for the debugging the whole dependence graph,
4325 - for the dejagnu testcases and maintenance.
4327 It is possible to ask only for a part of the graph, avoiding to
4328 compute the whole dependence graph. The computed dependences are
4329 stored in a knowledge base (KB) such that later queries don't
4330 recompute the same information. The implementation of this KB is
4331 transparent to the optimizer, and thus the KB can be changed with a
4332 more efficient implementation, or the KB could be disabled. */
4334 analyze_all_data_dependences (struct loop *loop)
4337 int nb_data_refs = 10;
4338 VEC (data_reference_p, heap) *datarefs =
4339 VEC_alloc (data_reference_p, heap, nb_data_refs);
4340 VEC (ddr_p, heap) *dependence_relations =
4341 VEC_alloc (ddr_p, heap, nb_data_refs * nb_data_refs);
4343 /* Compute DDs on the whole function. */
4344 compute_data_dependences_for_loop (loop, false, &datarefs,
4345 &dependence_relations);
4349 dump_data_dependence_relations (dump_file, dependence_relations);
4350 fprintf (dump_file, "\n\n");
4352 if (dump_flags & TDF_DETAILS)
4353 dump_dist_dir_vectors (dump_file, dependence_relations);
4355 if (dump_flags & TDF_STATS)
4357 unsigned nb_top_relations = 0;
4358 unsigned nb_bot_relations = 0;
4359 unsigned nb_basename_differ = 0;
4360 unsigned nb_chrec_relations = 0;
4361 struct data_dependence_relation *ddr;
4363 for (i = 0; VEC_iterate (ddr_p, dependence_relations, i, ddr); i++)
4365 if (chrec_contains_undetermined (DDR_ARE_DEPENDENT (ddr)))
4368 else if (DDR_ARE_DEPENDENT (ddr) == chrec_known)
4370 struct data_reference *a = DDR_A (ddr);
4371 struct data_reference *b = DDR_B (ddr);
4373 if (!bitmap_intersect_p (DR_VOPS (a), DR_VOPS (b)))
4374 nb_basename_differ++;
4380 nb_chrec_relations++;
4383 gather_stats_on_scev_database ();
4387 free_dependence_relations (dependence_relations);
4388 free_data_refs (datarefs);
4391 /* Computes all the data dependences and check that the results of
4392 several analyzers are the same. */
4395 tree_check_data_deps (void)
4398 struct loop *loop_nest;
4400 FOR_EACH_LOOP (li, loop_nest, 0)
4401 analyze_all_data_dependences (loop_nest);
4404 /* Free the memory used by a data dependence relation DDR. */
4407 free_dependence_relation (struct data_dependence_relation *ddr)
4412 if (DDR_SUBSCRIPTS (ddr))
4413 free_subscripts (DDR_SUBSCRIPTS (ddr));
4414 if (DDR_DIST_VECTS (ddr))
4415 VEC_free (lambda_vector, heap, DDR_DIST_VECTS (ddr));
4416 if (DDR_DIR_VECTS (ddr))
4417 VEC_free (lambda_vector, heap, DDR_DIR_VECTS (ddr));
4422 /* Free the memory used by the data dependence relations from
4423 DEPENDENCE_RELATIONS. */
4426 free_dependence_relations (VEC (ddr_p, heap) *dependence_relations)
4429 struct data_dependence_relation *ddr;
4430 VEC (loop_p, heap) *loop_nest = NULL;
4432 for (i = 0; VEC_iterate (ddr_p, dependence_relations, i, ddr); i++)
4436 if (loop_nest == NULL)
4437 loop_nest = DDR_LOOP_NEST (ddr);
4439 gcc_assert (DDR_LOOP_NEST (ddr) == NULL
4440 || DDR_LOOP_NEST (ddr) == loop_nest);
4441 free_dependence_relation (ddr);
4445 VEC_free (loop_p, heap, loop_nest);
4446 VEC_free (ddr_p, heap, dependence_relations);
4449 /* Free the memory used by the data references from DATAREFS. */
4452 free_data_refs (VEC (data_reference_p, heap) *datarefs)
4455 struct data_reference *dr;
4457 for (i = 0; VEC_iterate (data_reference_p, datarefs, i, dr); i++)
4459 VEC_free (data_reference_p, heap, datarefs);
4464 /* Dump vertex I in RDG to FILE. */
4467 dump_rdg_vertex (FILE *file, struct graph *rdg, int i)
4469 struct vertex *v = &(rdg->vertices[i]);
4470 struct graph_edge *e;
4472 fprintf (file, "(vertex %d: (%s%s) (in:", i,
4473 RDG_MEM_WRITE_STMT (rdg, i) ? "w" : "",
4474 RDG_MEM_READS_STMT (rdg, i) ? "r" : "");
4477 for (e = v->pred; e; e = e->pred_next)
4478 fprintf (file, " %d", e->src);
4480 fprintf (file, ") (out:");
4483 for (e = v->succ; e; e = e->succ_next)
4484 fprintf (file, " %d", e->dest);
4486 fprintf (file, ") \n");
4487 print_gimple_stmt (file, RDGV_STMT (v), 0, TDF_VOPS|TDF_MEMSYMS);
4488 fprintf (file, ")\n");
4491 /* Call dump_rdg_vertex on stderr. */
4494 debug_rdg_vertex (struct graph *rdg, int i)
4496 dump_rdg_vertex (stderr, rdg, i);
4499 /* Dump component C of RDG to FILE. If DUMPED is non-null, set the
4500 dumped vertices to that bitmap. */
4502 void dump_rdg_component (FILE *file, struct graph *rdg, int c, bitmap dumped)
4506 fprintf (file, "(%d\n", c);
4508 for (i = 0; i < rdg->n_vertices; i++)
4509 if (rdg->vertices[i].component == c)
4512 bitmap_set_bit (dumped, i);
4514 dump_rdg_vertex (file, rdg, i);
4517 fprintf (file, ")\n");
4520 /* Call dump_rdg_vertex on stderr. */
4523 debug_rdg_component (struct graph *rdg, int c)
4525 dump_rdg_component (stderr, rdg, c, NULL);
4528 /* Dump the reduced dependence graph RDG to FILE. */
4531 dump_rdg (FILE *file, struct graph *rdg)
4534 bitmap dumped = BITMAP_ALLOC (NULL);
4536 fprintf (file, "(rdg\n");
4538 for (i = 0; i < rdg->n_vertices; i++)
4539 if (!bitmap_bit_p (dumped, i))
4540 dump_rdg_component (file, rdg, rdg->vertices[i].component, dumped);
4542 fprintf (file, ")\n");
4543 BITMAP_FREE (dumped);
4546 /* Call dump_rdg on stderr. */
4549 debug_rdg (struct graph *rdg)
4551 dump_rdg (stderr, rdg);
4555 dot_rdg_1 (FILE *file, struct graph *rdg)
4559 fprintf (file, "digraph RDG {\n");
4561 for (i = 0; i < rdg->n_vertices; i++)
4563 struct vertex *v = &(rdg->vertices[i]);
4564 struct graph_edge *e;
4566 /* Highlight reads from memory. */
4567 if (RDG_MEM_READS_STMT (rdg, i))
4568 fprintf (file, "%d [style=filled, fillcolor=green]\n", i);
4570 /* Highlight stores to memory. */
4571 if (RDG_MEM_WRITE_STMT (rdg, i))
4572 fprintf (file, "%d [style=filled, fillcolor=red]\n", i);
4575 for (e = v->succ; e; e = e->succ_next)
4576 switch (RDGE_TYPE (e))
4579 fprintf (file, "%d -> %d [label=input] \n", i, e->dest);
4583 fprintf (file, "%d -> %d [label=output] \n", i, e->dest);
4587 /* These are the most common dependences: don't print these. */
4588 fprintf (file, "%d -> %d \n", i, e->dest);
4592 fprintf (file, "%d -> %d [label=anti] \n", i, e->dest);
4600 fprintf (file, "}\n\n");
4603 /* Display SCOP using dotty. */
4606 dot_rdg (struct graph *rdg)
4608 FILE *file = fopen ("/tmp/rdg.dot", "w");
4609 gcc_assert (file != NULL);
4611 dot_rdg_1 (file, rdg);
4614 system ("dotty /tmp/rdg.dot");
4618 /* This structure is used for recording the mapping statement index in
4621 struct rdg_vertex_info GTY(())
4627 /* Returns the index of STMT in RDG. */
4630 rdg_vertex_for_stmt (struct graph *rdg, gimple stmt)
4632 struct rdg_vertex_info rvi, *slot;
4635 slot = (struct rdg_vertex_info *) htab_find (rdg->indices, &rvi);
4643 /* Creates an edge in RDG for each distance vector from DDR. The
4644 order that we keep track of in the RDG is the order in which
4645 statements have to be executed. */
4648 create_rdg_edge_for_ddr (struct graph *rdg, ddr_p ddr)
4650 struct graph_edge *e;
4652 data_reference_p dra = DDR_A (ddr);
4653 data_reference_p drb = DDR_B (ddr);
4654 unsigned level = ddr_dependence_level (ddr);
4656 /* For non scalar dependences, when the dependence is REVERSED,
4657 statement B has to be executed before statement A. */
4659 && !DDR_REVERSED_P (ddr))
4661 data_reference_p tmp = dra;
4666 va = rdg_vertex_for_stmt (rdg, DR_STMT (dra));
4667 vb = rdg_vertex_for_stmt (rdg, DR_STMT (drb));
4669 if (va < 0 || vb < 0)
4672 e = add_edge (rdg, va, vb);
4673 e->data = XNEW (struct rdg_edge);
4675 RDGE_LEVEL (e) = level;
4676 RDGE_RELATION (e) = ddr;
4678 /* Determines the type of the data dependence. */
4679 if (DR_IS_READ (dra) && DR_IS_READ (drb))
4680 RDGE_TYPE (e) = input_dd;
4681 else if (!DR_IS_READ (dra) && !DR_IS_READ (drb))
4682 RDGE_TYPE (e) = output_dd;
4683 else if (!DR_IS_READ (dra) && DR_IS_READ (drb))
4684 RDGE_TYPE (e) = flow_dd;
4685 else if (DR_IS_READ (dra) && !DR_IS_READ (drb))
4686 RDGE_TYPE (e) = anti_dd;
4689 /* Creates dependence edges in RDG for all the uses of DEF. IDEF is
4690 the index of DEF in RDG. */
4693 create_rdg_edges_for_scalar (struct graph *rdg, tree def, int idef)
4695 use_operand_p imm_use_p;
4696 imm_use_iterator iterator;
4698 FOR_EACH_IMM_USE_FAST (imm_use_p, iterator, def)
4700 struct graph_edge *e;
4701 int use = rdg_vertex_for_stmt (rdg, USE_STMT (imm_use_p));
4706 e = add_edge (rdg, idef, use);
4707 e->data = XNEW (struct rdg_edge);
4708 RDGE_TYPE (e) = flow_dd;
4709 RDGE_RELATION (e) = NULL;
4713 /* Creates the edges of the reduced dependence graph RDG. */
4716 create_rdg_edges (struct graph *rdg, VEC (ddr_p, heap) *ddrs)
4719 struct data_dependence_relation *ddr;
4720 def_operand_p def_p;
4723 for (i = 0; VEC_iterate (ddr_p, ddrs, i, ddr); i++)
4724 if (DDR_ARE_DEPENDENT (ddr) == NULL_TREE)
4725 create_rdg_edge_for_ddr (rdg, ddr);
4727 for (i = 0; i < rdg->n_vertices; i++)
4728 FOR_EACH_PHI_OR_STMT_DEF (def_p, RDG_STMT (rdg, i),
4730 create_rdg_edges_for_scalar (rdg, DEF_FROM_PTR (def_p), i);
4733 /* Build the vertices of the reduced dependence graph RDG. */
4736 create_rdg_vertices (struct graph *rdg, VEC (gimple, heap) *stmts)
4741 for (i = 0; VEC_iterate (gimple, stmts, i, stmt); i++)
4743 VEC (data_ref_loc, heap) *references;
4745 struct vertex *v = &(rdg->vertices[i]);
4746 struct rdg_vertex_info *rvi = XNEW (struct rdg_vertex_info);
4747 struct rdg_vertex_info **slot;
4751 slot = (struct rdg_vertex_info **) htab_find_slot (rdg->indices, rvi, INSERT);
4758 v->data = XNEW (struct rdg_vertex);
4759 RDG_STMT (rdg, i) = stmt;
4761 RDG_MEM_WRITE_STMT (rdg, i) = false;
4762 RDG_MEM_READS_STMT (rdg, i) = false;
4763 if (gimple_code (stmt) == GIMPLE_PHI)
4766 get_references_in_stmt (stmt, &references);
4767 for (j = 0; VEC_iterate (data_ref_loc, references, j, ref); j++)
4769 RDG_MEM_WRITE_STMT (rdg, i) = true;
4771 RDG_MEM_READS_STMT (rdg, i) = true;
4773 VEC_free (data_ref_loc, heap, references);
4777 /* Initialize STMTS with all the statements of LOOP. When
4778 INCLUDE_PHIS is true, include also the PHI nodes. The order in
4779 which we discover statements is important as
4780 generate_loops_for_partition is using the same traversal for
4781 identifying statements. */
4784 stmts_from_loop (struct loop *loop, VEC (gimple, heap) **stmts)
4787 basic_block *bbs = get_loop_body_in_dom_order (loop);
4789 for (i = 0; i < loop->num_nodes; i++)
4791 basic_block bb = bbs[i];
4792 gimple_stmt_iterator bsi;
4795 for (bsi = gsi_start_phis (bb); !gsi_end_p (bsi); gsi_next (&bsi))
4796 VEC_safe_push (gimple, heap, *stmts, gsi_stmt (bsi));
4798 for (bsi = gsi_start_bb (bb); !gsi_end_p (bsi); gsi_next (&bsi))
4800 stmt = gsi_stmt (bsi);
4801 if (gimple_code (stmt) != GIMPLE_LABEL)
4802 VEC_safe_push (gimple, heap, *stmts, stmt);
4809 /* Returns true when all the dependences are computable. */
4812 known_dependences_p (VEC (ddr_p, heap) *dependence_relations)
4817 for (i = 0; VEC_iterate (ddr_p, dependence_relations, i, ddr); i++)
4818 if (DDR_ARE_DEPENDENT (ddr) == chrec_dont_know)
4824 /* Computes a hash function for element ELT. */
4827 hash_stmt_vertex_info (const void *elt)
4829 const struct rdg_vertex_info *const rvi =
4830 (const struct rdg_vertex_info *) elt;
4831 gimple stmt = rvi->stmt;
4833 return htab_hash_pointer (stmt);
4836 /* Compares database elements E1 and E2. */
4839 eq_stmt_vertex_info (const void *e1, const void *e2)
4841 const struct rdg_vertex_info *elt1 = (const struct rdg_vertex_info *) e1;
4842 const struct rdg_vertex_info *elt2 = (const struct rdg_vertex_info *) e2;
4844 return elt1->stmt == elt2->stmt;
4847 /* Free the element E. */
4850 hash_stmt_vertex_del (void *e)
4855 /* Build the Reduced Dependence Graph (RDG) with one vertex per
4856 statement of the loop nest, and one edge per data dependence or
4857 scalar dependence. */
4860 build_empty_rdg (int n_stmts)
4862 int nb_data_refs = 10;
4863 struct graph *rdg = new_graph (n_stmts);
4865 rdg->indices = htab_create (nb_data_refs, hash_stmt_vertex_info,
4866 eq_stmt_vertex_info, hash_stmt_vertex_del);
4870 /* Build the Reduced Dependence Graph (RDG) with one vertex per
4871 statement of the loop nest, and one edge per data dependence or
4872 scalar dependence. */
4875 build_rdg (struct loop *loop)
4877 int nb_data_refs = 10;
4878 struct graph *rdg = NULL;
4879 VEC (ddr_p, heap) *dependence_relations;
4880 VEC (data_reference_p, heap) *datarefs;
4881 VEC (gimple, heap) *stmts = VEC_alloc (gimple, heap, nb_data_refs);
4883 dependence_relations = VEC_alloc (ddr_p, heap, nb_data_refs * nb_data_refs) ;
4884 datarefs = VEC_alloc (data_reference_p, heap, nb_data_refs);
4885 compute_data_dependences_for_loop (loop,
4888 &dependence_relations);
4890 if (!known_dependences_p (dependence_relations))
4892 free_dependence_relations (dependence_relations);
4893 free_data_refs (datarefs);
4894 VEC_free (gimple, heap, stmts);
4899 stmts_from_loop (loop, &stmts);
4900 rdg = build_empty_rdg (VEC_length (gimple, stmts));
4902 rdg->indices = htab_create (nb_data_refs, hash_stmt_vertex_info,
4903 eq_stmt_vertex_info, hash_stmt_vertex_del);
4904 create_rdg_vertices (rdg, stmts);
4905 create_rdg_edges (rdg, dependence_relations);
4907 VEC_free (gimple, heap, stmts);
4911 /* Free the reduced dependence graph RDG. */
4914 free_rdg (struct graph *rdg)
4918 for (i = 0; i < rdg->n_vertices; i++)
4919 free (rdg->vertices[i].data);
4921 htab_delete (rdg->indices);
4925 /* Initialize STMTS with all the statements of LOOP that contain a
4929 stores_from_loop (struct loop *loop, VEC (gimple, heap) **stmts)
4932 basic_block *bbs = get_loop_body_in_dom_order (loop);
4934 for (i = 0; i < loop->num_nodes; i++)
4936 basic_block bb = bbs[i];
4937 gimple_stmt_iterator bsi;
4939 for (bsi = gsi_start_bb (bb); !gsi_end_p (bsi); gsi_next (&bsi))
4940 if (!ZERO_SSA_OPERANDS (gsi_stmt (bsi), SSA_OP_VDEF))
4941 VEC_safe_push (gimple, heap, *stmts, gsi_stmt (bsi));
4947 /* For a data reference REF, return the declaration of its base
4948 address or NULL_TREE if the base is not determined. */
4951 ref_base_address (gimple stmt, data_ref_loc *ref)
4953 tree base = NULL_TREE;
4955 struct data_reference *dr = XCNEW (struct data_reference);
4957 DR_STMT (dr) = stmt;
4958 DR_REF (dr) = *ref->pos;
4959 dr_analyze_innermost (dr);
4960 base_address = DR_BASE_ADDRESS (dr);
4965 switch (TREE_CODE (base_address))
4968 base = TREE_OPERAND (base_address, 0);
4972 base = base_address;
4981 /* Determines whether the statement from vertex V of the RDG has a
4982 definition used outside the loop that contains this statement. */
4985 rdg_defs_used_in_other_loops_p (struct graph *rdg, int v)
4987 gimple stmt = RDG_STMT (rdg, v);
4988 struct loop *loop = loop_containing_stmt (stmt);
4989 use_operand_p imm_use_p;
4990 imm_use_iterator iterator;
4992 def_operand_p def_p;
4997 FOR_EACH_PHI_OR_STMT_DEF (def_p, stmt, it, SSA_OP_DEF)
4999 FOR_EACH_IMM_USE_FAST (imm_use_p, iterator, DEF_FROM_PTR (def_p))
5001 if (loop_containing_stmt (USE_STMT (imm_use_p)) != loop)
5009 /* Determines whether statements S1 and S2 access to similar memory
5010 locations. Two memory accesses are considered similar when they
5011 have the same base address declaration, i.e. when their
5012 ref_base_address is the same. */
5015 have_similar_memory_accesses (gimple s1, gimple s2)
5019 VEC (data_ref_loc, heap) *refs1, *refs2;
5020 data_ref_loc *ref1, *ref2;
5022 get_references_in_stmt (s1, &refs1);
5023 get_references_in_stmt (s2, &refs2);
5025 for (i = 0; VEC_iterate (data_ref_loc, refs1, i, ref1); i++)
5027 tree base1 = ref_base_address (s1, ref1);
5030 for (j = 0; VEC_iterate (data_ref_loc, refs2, j, ref2); j++)
5031 if (base1 == ref_base_address (s2, ref2))
5039 VEC_free (data_ref_loc, heap, refs1);
5040 VEC_free (data_ref_loc, heap, refs2);
5044 /* Helper function for the hashtab. */
5047 have_similar_memory_accesses_1 (const void *s1, const void *s2)
5049 return have_similar_memory_accesses (CONST_CAST_GIMPLE ((const_gimple) s1),
5050 CONST_CAST_GIMPLE ((const_gimple) s2));
5053 /* Helper function for the hashtab. */
5056 ref_base_address_1 (const void *s)
5058 gimple stmt = CONST_CAST_GIMPLE ((const_gimple) s);
5060 VEC (data_ref_loc, heap) *refs;
5064 get_references_in_stmt (stmt, &refs);
5066 for (i = 0; VEC_iterate (data_ref_loc, refs, i, ref); i++)
5069 res = htab_hash_pointer (ref_base_address (stmt, ref));
5073 VEC_free (data_ref_loc, heap, refs);
5077 /* Try to remove duplicated write data references from STMTS. */
5080 remove_similar_memory_refs (VEC (gimple, heap) **stmts)
5084 htab_t seen = htab_create (VEC_length (gimple, *stmts), ref_base_address_1,
5085 have_similar_memory_accesses_1, NULL);
5087 for (i = 0; VEC_iterate (gimple, *stmts, i, stmt); )
5091 slot = htab_find_slot (seen, stmt, INSERT);
5094 VEC_ordered_remove (gimple, *stmts, i);
5097 *slot = (void *) stmt;
5105 /* Returns the index of PARAMETER in the parameters vector of the
5106 ACCESS_MATRIX. If PARAMETER does not exist return -1. */
5109 access_matrix_get_index_for_parameter (tree parameter,
5110 struct access_matrix *access_matrix)
5113 VEC (tree,heap) *lambda_parameters = AM_PARAMETERS (access_matrix);
5114 tree lambda_parameter;
5116 for (i = 0; VEC_iterate (tree, lambda_parameters, i, lambda_parameter); i++)
5117 if (lambda_parameter == parameter)
5118 return i + AM_NB_INDUCTION_VARS (access_matrix);