| /* Data references and dependences detectors. |
| Copyright (C) 2003-2019 Free Software Foundation, Inc. |
| Contributed by Sebastian Pop <pop@cri.ensmp.fr> |
| |
| This file is part of GCC. |
| |
| GCC is free software; you can redistribute it and/or modify it under |
| the terms of the GNU General Public License as published by the Free |
| Software Foundation; either version 3, or (at your option) any later |
| version. |
| |
| GCC is distributed in the hope that it will be useful, but WITHOUT ANY |
| WARRANTY; without even the implied warranty of MERCHANTABILITY or |
| FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License |
| for more details. |
| |
| You should have received a copy of the GNU General Public License |
| along with GCC; see the file COPYING3. If not see |
| <http://www.gnu.org/licenses/>. */ |
| |
| /* This pass walks a given loop structure searching for array |
| references. The information about the array accesses is recorded |
| in DATA_REFERENCE structures. |
| |
| The basic test for determining the dependences is: |
| given two access functions chrec1 and chrec2 to a same array, and |
| x and y two vectors from the iteration domain, the same element of |
| the array is accessed twice at iterations x and y if and only if: |
| | chrec1 (x) == chrec2 (y). |
| |
| The goals of this analysis are: |
| |
| - to determine the independence: the relation between two |
| independent accesses is qualified with the chrec_known (this |
| information allows a loop parallelization), |
| |
| - when two data references access the same data, to qualify the |
| dependence relation with classic dependence representations: |
| |
| - distance vectors |
| - direction vectors |
| - loop carried level dependence |
| - polyhedron dependence |
| or with the chains of recurrences based representation, |
| |
| - to define a knowledge base for storing the data dependence |
| information, |
| |
| - to define an interface to access this data. |
| |
| |
| Definitions: |
| |
| - subscript: given two array accesses a subscript is the tuple |
| composed of the access functions for a given dimension. Example: |
| Given A[f1][f2][f3] and B[g1][g2][g3], there are three subscripts: |
| (f1, g1), (f2, g2), (f3, g3). |
| |
| - Diophantine equation: an equation whose coefficients and |
| solutions are integer constants, for example the equation |
| | 3*x + 2*y = 1 |
| has an integer solution x = 1 and y = -1. |
| |
| References: |
| |
| - "Advanced Compilation for High Performance Computing" by Randy |
| Allen and Ken Kennedy. |
| http://citeseer.ist.psu.edu/goff91practical.html |
| |
| - "Loop Transformations for Restructuring Compilers - The Foundations" |
| by Utpal Banerjee. |
| |
| |
| */ |
| |
| #include "config.h" |
| #include "system.h" |
| #include "coretypes.h" |
| #include "backend.h" |
| #include "rtl.h" |
| #include "tree.h" |
| #include "gimple.h" |
| #include "gimple-pretty-print.h" |
| #include "alias.h" |
| #include "fold-const.h" |
| #include "expr.h" |
| #include "gimple-iterator.h" |
| #include "tree-ssa-loop-niter.h" |
| #include "tree-ssa-loop.h" |
| #include "tree-ssa.h" |
| #include "cfgloop.h" |
| #include "tree-data-ref.h" |
| #include "tree-scalar-evolution.h" |
| #include "dumpfile.h" |
| #include "tree-affine.h" |
| #include "params.h" |
| #include "builtins.h" |
| #include "tree-eh.h" |
| #include "ssa.h" |
| |
| static struct datadep_stats |
| { |
| int num_dependence_tests; |
| int num_dependence_dependent; |
| int num_dependence_independent; |
| int num_dependence_undetermined; |
| |
| int num_subscript_tests; |
| int num_subscript_undetermined; |
| int num_same_subscript_function; |
| |
| int num_ziv; |
| int num_ziv_independent; |
| int num_ziv_dependent; |
| int num_ziv_unimplemented; |
| |
| int num_siv; |
| int num_siv_independent; |
| int num_siv_dependent; |
| int num_siv_unimplemented; |
| |
| int num_miv; |
| int num_miv_independent; |
| int num_miv_dependent; |
| int num_miv_unimplemented; |
| } dependence_stats; |
| |
| static bool subscript_dependence_tester_1 (struct data_dependence_relation *, |
| unsigned int, unsigned int, |
| struct loop *); |
| /* Returns true iff A divides B. */ |
| |
| static inline bool |
| tree_fold_divides_p (const_tree a, const_tree b) |
| { |
| gcc_assert (TREE_CODE (a) == INTEGER_CST); |
| gcc_assert (TREE_CODE (b) == INTEGER_CST); |
| return integer_zerop (int_const_binop (TRUNC_MOD_EXPR, b, a)); |
| } |
| |
| /* Returns true iff A divides B. */ |
| |
| static inline bool |
| int_divides_p (lambda_int a, lambda_int b) |
| { |
| return ((b % a) == 0); |
| } |
| |
| /* Return true if reference REF contains a union access. */ |
| |
| static bool |
| ref_contains_union_access_p (tree ref) |
| { |
| while (handled_component_p (ref)) |
| { |
| ref = TREE_OPERAND (ref, 0); |
| if (TREE_CODE (TREE_TYPE (ref)) == UNION_TYPE |
| || TREE_CODE (TREE_TYPE (ref)) == QUAL_UNION_TYPE) |
| return true; |
| } |
| return false; |
| } |
| |
| |
| |
| /* Dump into FILE all the data references from DATAREFS. */ |
| |
| static void |
| dump_data_references (FILE *file, vec<data_reference_p> datarefs) |
| { |
| unsigned int i; |
| struct data_reference *dr; |
| |
| FOR_EACH_VEC_ELT (datarefs, i, dr) |
| dump_data_reference (file, dr); |
| } |
| |
| /* Unified dump into FILE all the data references from DATAREFS. */ |
| |
| DEBUG_FUNCTION void |
| debug (vec<data_reference_p> &ref) |
| { |
| dump_data_references (stderr, ref); |
| } |
| |
| DEBUG_FUNCTION void |
| debug (vec<data_reference_p> *ptr) |
| { |
| if (ptr) |
| debug (*ptr); |
| else |
| fprintf (stderr, "<nil>\n"); |
| } |
| |
| |
| /* Dump into STDERR all the data references from DATAREFS. */ |
| |
| DEBUG_FUNCTION void |
| debug_data_references (vec<data_reference_p> datarefs) |
| { |
| dump_data_references (stderr, datarefs); |
| } |
| |
| /* Print to STDERR the data_reference DR. */ |
| |
| DEBUG_FUNCTION void |
| debug_data_reference (struct data_reference *dr) |
| { |
| dump_data_reference (stderr, dr); |
| } |
| |
| /* Dump function for a DATA_REFERENCE structure. */ |
| |
| void |
| dump_data_reference (FILE *outf, |
| struct data_reference *dr) |
| { |
| unsigned int i; |
| |
| fprintf (outf, "#(Data Ref: \n"); |
| fprintf (outf, "# bb: %d \n", gimple_bb (DR_STMT (dr))->index); |
| fprintf (outf, "# stmt: "); |
| print_gimple_stmt (outf, DR_STMT (dr), 0); |
| fprintf (outf, "# ref: "); |
| print_generic_stmt (outf, DR_REF (dr)); |
| fprintf (outf, "# base_object: "); |
| print_generic_stmt (outf, DR_BASE_OBJECT (dr)); |
| |
| for (i = 0; i < DR_NUM_DIMENSIONS (dr); i++) |
| { |
| fprintf (outf, "# Access function %d: ", i); |
| print_generic_stmt (outf, DR_ACCESS_FN (dr, i)); |
| } |
| fprintf (outf, "#)\n"); |
| } |
| |
| /* Unified dump function for a DATA_REFERENCE structure. */ |
| |
| DEBUG_FUNCTION void |
| debug (data_reference &ref) |
| { |
| dump_data_reference (stderr, &ref); |
| } |
| |
| DEBUG_FUNCTION void |
| debug (data_reference *ptr) |
| { |
| if (ptr) |
| debug (*ptr); |
| else |
| fprintf (stderr, "<nil>\n"); |
| } |
| |
| |
| /* Dumps the affine function described by FN to the file OUTF. */ |
| |
| DEBUG_FUNCTION void |
| dump_affine_function (FILE *outf, affine_fn fn) |
| { |
| unsigned i; |
| tree coef; |
| |
| print_generic_expr (outf, fn[0], TDF_SLIM); |
| for (i = 1; fn.iterate (i, &coef); i++) |
| { |
| fprintf (outf, " + "); |
| print_generic_expr (outf, coef, TDF_SLIM); |
| fprintf (outf, " * x_%u", i); |
| } |
| } |
| |
| /* Dumps the conflict function CF to the file OUTF. */ |
| |
| DEBUG_FUNCTION void |
| dump_conflict_function (FILE *outf, conflict_function *cf) |
| { |
| unsigned i; |
| |
| if (cf->n == NO_DEPENDENCE) |
| fprintf (outf, "no dependence"); |
| else if (cf->n == NOT_KNOWN) |
| fprintf (outf, "not known"); |
| else |
| { |
| for (i = 0; i < cf->n; i++) |
| { |
| if (i != 0) |
| fprintf (outf, " "); |
| fprintf (outf, "["); |
| dump_affine_function (outf, cf->fns[i]); |
| fprintf (outf, "]"); |
| } |
| } |
| } |
| |
| /* Dump function for a SUBSCRIPT structure. */ |
| |
| DEBUG_FUNCTION void |
| dump_subscript (FILE *outf, struct subscript *subscript) |
| { |
| conflict_function *cf = SUB_CONFLICTS_IN_A (subscript); |
| |
| fprintf (outf, "\n (subscript \n"); |
| fprintf (outf, " iterations_that_access_an_element_twice_in_A: "); |
| dump_conflict_function (outf, cf); |
| if (CF_NONTRIVIAL_P (cf)) |
| { |
| tree last_iteration = SUB_LAST_CONFLICT (subscript); |
| fprintf (outf, "\n last_conflict: "); |
| print_generic_expr (outf, last_iteration); |
| } |
| |
| cf = SUB_CONFLICTS_IN_B (subscript); |
| fprintf (outf, "\n iterations_that_access_an_element_twice_in_B: "); |
| dump_conflict_function (outf, cf); |
| if (CF_NONTRIVIAL_P (cf)) |
| { |
| tree last_iteration = SUB_LAST_CONFLICT (subscript); |
| fprintf (outf, "\n last_conflict: "); |
| print_generic_expr (outf, last_iteration); |
| } |
| |
| fprintf (outf, "\n (Subscript distance: "); |
| print_generic_expr (outf, SUB_DISTANCE (subscript)); |
| fprintf (outf, " ))\n"); |
| } |
| |
| /* Print the classic direction vector DIRV to OUTF. */ |
| |
| DEBUG_FUNCTION void |
| print_direction_vector (FILE *outf, |
| lambda_vector dirv, |
| int length) |
| { |
| int eq; |
| |
| for (eq = 0; eq < length; eq++) |
| { |
| enum data_dependence_direction dir = ((enum data_dependence_direction) |
| dirv[eq]); |
| |
| switch (dir) |
| { |
| case dir_positive: |
| fprintf (outf, " +"); |
| break; |
| case dir_negative: |
| fprintf (outf, " -"); |
| break; |
| case dir_equal: |
| fprintf (outf, " ="); |
| break; |
| case dir_positive_or_equal: |
| fprintf (outf, " +="); |
| break; |
| case dir_positive_or_negative: |
| fprintf (outf, " +-"); |
| break; |
| case dir_negative_or_equal: |
| fprintf (outf, " -="); |
| break; |
| case dir_star: |
| fprintf (outf, " *"); |
| break; |
| default: |
| fprintf (outf, "indep"); |
| break; |
| } |
| } |
| fprintf (outf, "\n"); |
| } |
| |
| /* Print a vector of direction vectors. */ |
| |
| DEBUG_FUNCTION void |
| print_dir_vectors (FILE *outf, vec<lambda_vector> dir_vects, |
| int length) |
| { |
| unsigned j; |
| lambda_vector v; |
| |
| FOR_EACH_VEC_ELT (dir_vects, j, v) |
| print_direction_vector (outf, v, length); |
| } |
| |
| /* Print out a vector VEC of length N to OUTFILE. */ |
| |
| DEBUG_FUNCTION void |
| print_lambda_vector (FILE * outfile, lambda_vector vector, int n) |
| { |
| int i; |
| |
| for (i = 0; i < n; i++) |
| fprintf (outfile, HOST_WIDE_INT_PRINT_DEC " ", vector[i]); |
| fprintf (outfile, "\n"); |
| } |
| |
| /* Print a vector of distance vectors. */ |
| |
| DEBUG_FUNCTION void |
| print_dist_vectors (FILE *outf, vec<lambda_vector> dist_vects, |
| int length) |
| { |
| unsigned j; |
| lambda_vector v; |
| |
| FOR_EACH_VEC_ELT (dist_vects, j, v) |
| print_lambda_vector (outf, v, length); |
| } |
| |
| /* Dump function for a DATA_DEPENDENCE_RELATION structure. */ |
| |
| DEBUG_FUNCTION void |
| dump_data_dependence_relation (FILE *outf, |
| struct data_dependence_relation *ddr) |
| { |
| struct data_reference *dra, *drb; |
| |
| fprintf (outf, "(Data Dep: \n"); |
| |
| if (!ddr || DDR_ARE_DEPENDENT (ddr) == chrec_dont_know) |
| { |
| if (ddr) |
| { |
| dra = DDR_A (ddr); |
| drb = DDR_B (ddr); |
| if (dra) |
| dump_data_reference (outf, dra); |
| else |
| fprintf (outf, " (nil)\n"); |
| if (drb) |
| dump_data_reference (outf, drb); |
| else |
| fprintf (outf, " (nil)\n"); |
| } |
| fprintf (outf, " (don't know)\n)\n"); |
| return; |
| } |
| |
| dra = DDR_A (ddr); |
| drb = DDR_B (ddr); |
| dump_data_reference (outf, dra); |
| dump_data_reference (outf, drb); |
| |
| if (DDR_ARE_DEPENDENT (ddr) == chrec_known) |
| fprintf (outf, " (no dependence)\n"); |
| |
| else if (DDR_ARE_DEPENDENT (ddr) == NULL_TREE) |
| { |
| unsigned int i; |
| struct loop *loopi; |
| |
| subscript *sub; |
| FOR_EACH_VEC_ELT (DDR_SUBSCRIPTS (ddr), i, sub) |
| { |
| fprintf (outf, " access_fn_A: "); |
| print_generic_stmt (outf, SUB_ACCESS_FN (sub, 0)); |
| fprintf (outf, " access_fn_B: "); |
| print_generic_stmt (outf, SUB_ACCESS_FN (sub, 1)); |
| dump_subscript (outf, sub); |
| } |
| |
| fprintf (outf, " inner loop index: %d\n", DDR_INNER_LOOP (ddr)); |
| fprintf (outf, " loop nest: ("); |
| FOR_EACH_VEC_ELT (DDR_LOOP_NEST (ddr), i, loopi) |
| fprintf (outf, "%d ", loopi->num); |
| fprintf (outf, ")\n"); |
| |
| for (i = 0; i < DDR_NUM_DIST_VECTS (ddr); i++) |
| { |
| fprintf (outf, " distance_vector: "); |
| print_lambda_vector (outf, DDR_DIST_VECT (ddr, i), |
| DDR_NB_LOOPS (ddr)); |
| } |
| |
| for (i = 0; i < DDR_NUM_DIR_VECTS (ddr); i++) |
| { |
| fprintf (outf, " direction_vector: "); |
| print_direction_vector (outf, DDR_DIR_VECT (ddr, i), |
| DDR_NB_LOOPS (ddr)); |
| } |
| } |
| |
| fprintf (outf, ")\n"); |
| } |
| |
| /* Debug version. */ |
| |
| DEBUG_FUNCTION void |
| debug_data_dependence_relation (struct data_dependence_relation *ddr) |
| { |
| dump_data_dependence_relation (stderr, ddr); |
| } |
| |
| /* Dump into FILE all the dependence relations from DDRS. */ |
| |
| DEBUG_FUNCTION void |
| dump_data_dependence_relations (FILE *file, |
| vec<ddr_p> ddrs) |
| { |
| unsigned int i; |
| struct data_dependence_relation *ddr; |
| |
| FOR_EACH_VEC_ELT (ddrs, i, ddr) |
| dump_data_dependence_relation (file, ddr); |
| } |
| |
| DEBUG_FUNCTION void |
| debug (vec<ddr_p> &ref) |
| { |
| dump_data_dependence_relations (stderr, ref); |
| } |
| |
| DEBUG_FUNCTION void |
| debug (vec<ddr_p> *ptr) |
| { |
| if (ptr) |
| debug (*ptr); |
| else |
| fprintf (stderr, "<nil>\n"); |
| } |
| |
| |
| /* Dump to STDERR all the dependence relations from DDRS. */ |
| |
| DEBUG_FUNCTION void |
| debug_data_dependence_relations (vec<ddr_p> ddrs) |
| { |
| dump_data_dependence_relations (stderr, ddrs); |
| } |
| |
| /* Dumps the distance and direction vectors in FILE. DDRS contains |
| the dependence relations, and VECT_SIZE is the size of the |
| dependence vectors, or in other words the number of loops in the |
| considered nest. */ |
| |
| DEBUG_FUNCTION void |
| dump_dist_dir_vectors (FILE *file, vec<ddr_p> ddrs) |
| { |
| unsigned int i, j; |
| struct data_dependence_relation *ddr; |
| lambda_vector v; |
| |
| FOR_EACH_VEC_ELT (ddrs, i, ddr) |
| if (DDR_ARE_DEPENDENT (ddr) == NULL_TREE && DDR_AFFINE_P (ddr)) |
| { |
| FOR_EACH_VEC_ELT (DDR_DIST_VECTS (ddr), j, v) |
| { |
| fprintf (file, "DISTANCE_V ("); |
| print_lambda_vector (file, v, DDR_NB_LOOPS (ddr)); |
| fprintf (file, ")\n"); |
| } |
| |
| FOR_EACH_VEC_ELT (DDR_DIR_VECTS (ddr), j, v) |
| { |
| fprintf (file, "DIRECTION_V ("); |
| print_direction_vector (file, v, DDR_NB_LOOPS (ddr)); |
| fprintf (file, ")\n"); |
| } |
| } |
| |
| fprintf (file, "\n\n"); |
| } |
| |
| /* Dumps the data dependence relations DDRS in FILE. */ |
| |
| DEBUG_FUNCTION void |
| dump_ddrs (FILE *file, vec<ddr_p> ddrs) |
| { |
| unsigned int i; |
| struct data_dependence_relation *ddr; |
| |
| FOR_EACH_VEC_ELT (ddrs, i, ddr) |
| dump_data_dependence_relation (file, ddr); |
| |
| fprintf (file, "\n\n"); |
| } |
| |
| DEBUG_FUNCTION void |
| debug_ddrs (vec<ddr_p> ddrs) |
| { |
| dump_ddrs (stderr, ddrs); |
| } |
| |
| static void |
| split_constant_offset (tree exp, tree *var, tree *off, |
| hash_map<tree, std::pair<tree, tree> > &cache, |
| unsigned *limit); |
| |
| /* Helper function for split_constant_offset. Expresses OP0 CODE OP1 |
| (the type of the result is TYPE) as VAR + OFF, where OFF is a nonzero |
| constant of type ssizetype, and returns true. If we cannot do this |
| with OFF nonzero, OFF and VAR are set to NULL_TREE instead and false |
| is returned. */ |
| |
| static bool |
| split_constant_offset_1 (tree type, tree op0, enum tree_code code, tree op1, |
| tree *var, tree *off, |
| hash_map<tree, std::pair<tree, tree> > &cache, |
| unsigned *limit) |
| { |
| tree var0, var1; |
| tree off0, off1; |
| enum tree_code ocode = code; |
| |
| *var = NULL_TREE; |
| *off = NULL_TREE; |
| |
| switch (code) |
| { |
| case INTEGER_CST: |
| *var = build_int_cst (type, 0); |
| *off = fold_convert (ssizetype, op0); |
| return true; |
| |
| case POINTER_PLUS_EXPR: |
| ocode = PLUS_EXPR; |
| /* FALLTHROUGH */ |
| case PLUS_EXPR: |
| case MINUS_EXPR: |
| if (TREE_CODE (op1) == INTEGER_CST) |
| { |
| split_constant_offset (op0, &var0, &off0, cache, limit); |
| *var = var0; |
| *off = size_binop (ocode, off0, fold_convert (ssizetype, op1)); |
| return true; |
| } |
| split_constant_offset (op0, &var0, &off0, cache, limit); |
| split_constant_offset (op1, &var1, &off1, cache, limit); |
| *var = fold_build2 (code, type, var0, var1); |
| *off = size_binop (ocode, off0, off1); |
| return true; |
| |
| case MULT_EXPR: |
| if (TREE_CODE (op1) != INTEGER_CST) |
| return false; |
| |
| split_constant_offset (op0, &var0, &off0, cache, limit); |
| *var = fold_build2 (MULT_EXPR, type, var0, op1); |
| *off = size_binop (MULT_EXPR, off0, fold_convert (ssizetype, op1)); |
| return true; |
| |
| case ADDR_EXPR: |
| { |
| tree base, poffset; |
| poly_int64 pbitsize, pbitpos, pbytepos; |
| machine_mode pmode; |
| int punsignedp, preversep, pvolatilep; |
| |
| op0 = TREE_OPERAND (op0, 0); |
| base |
| = get_inner_reference (op0, &pbitsize, &pbitpos, &poffset, &pmode, |
| &punsignedp, &preversep, &pvolatilep); |
| |
| if (!multiple_p (pbitpos, BITS_PER_UNIT, &pbytepos)) |
| return false; |
| base = build_fold_addr_expr (base); |
| off0 = ssize_int (pbytepos); |
| |
| if (poffset) |
| { |
| split_constant_offset (poffset, &poffset, &off1, cache, limit); |
| off0 = size_binop (PLUS_EXPR, off0, off1); |
| if (POINTER_TYPE_P (TREE_TYPE (base))) |
| base = fold_build_pointer_plus (base, poffset); |
| else |
| base = fold_build2 (PLUS_EXPR, TREE_TYPE (base), base, |
| fold_convert (TREE_TYPE (base), poffset)); |
| } |
| |
| var0 = fold_convert (type, base); |
| |
| /* If variable length types are involved, punt, otherwise casts |
| might be converted into ARRAY_REFs in gimplify_conversion. |
| To compute that ARRAY_REF's element size TYPE_SIZE_UNIT, which |
| possibly no longer appears in current GIMPLE, might resurface. |
| This perhaps could run |
| if (CONVERT_EXPR_P (var0)) |
| { |
| gimplify_conversion (&var0); |
| // Attempt to fill in any within var0 found ARRAY_REF's |
| // element size from corresponding op embedded ARRAY_REF, |
| // if unsuccessful, just punt. |
| } */ |
| while (POINTER_TYPE_P (type)) |
| type = TREE_TYPE (type); |
| if (int_size_in_bytes (type) < 0) |
| return false; |
| |
| *var = var0; |
| *off = off0; |
| return true; |
| } |
| |
| case SSA_NAME: |
| { |
| if (SSA_NAME_OCCURS_IN_ABNORMAL_PHI (op0)) |
| return false; |
| |
| gimple *def_stmt = SSA_NAME_DEF_STMT (op0); |
| enum tree_code subcode; |
| |
| if (gimple_code (def_stmt) != GIMPLE_ASSIGN) |
| return false; |
| |
| subcode = gimple_assign_rhs_code (def_stmt); |
| |
| /* We are using a cache to avoid un-CSEing large amounts of code. */ |
| bool use_cache = false; |
| if (!has_single_use (op0) |
| && (subcode == POINTER_PLUS_EXPR |
| || subcode == PLUS_EXPR |
| || subcode == MINUS_EXPR |
| || subcode == MULT_EXPR |
| || subcode == ADDR_EXPR |
| || CONVERT_EXPR_CODE_P (subcode))) |
| { |
| use_cache = true; |
| bool existed; |
| std::pair<tree, tree> &e = cache.get_or_insert (op0, &existed); |
| if (existed) |
| { |
| if (integer_zerop (e.second)) |
| return false; |
| *var = e.first; |
| *off = e.second; |
| return true; |
| } |
| e = std::make_pair (op0, ssize_int (0)); |
| } |
| |
| if (*limit == 0) |
| return false; |
| --*limit; |
| |
| var0 = gimple_assign_rhs1 (def_stmt); |
| var1 = gimple_assign_rhs2 (def_stmt); |
| |
| bool res = split_constant_offset_1 (type, var0, subcode, var1, |
| var, off, cache, limit); |
| if (res && use_cache) |
| *cache.get (op0) = std::make_pair (*var, *off); |
| return res; |
| } |
| CASE_CONVERT: |
| { |
| /* We must not introduce undefined overflow, and we must not change |
| the value. Hence we're okay if the inner type doesn't overflow |
| to start with (pointer or signed), the outer type also is an |
| integer or pointer and the outer precision is at least as large |
| as the inner. */ |
| tree itype = TREE_TYPE (op0); |
| if ((POINTER_TYPE_P (itype) |
| || (INTEGRAL_TYPE_P (itype) && !TYPE_OVERFLOW_TRAPS (itype))) |
| && TYPE_PRECISION (type) >= TYPE_PRECISION (itype) |
| && (POINTER_TYPE_P (type) || INTEGRAL_TYPE_P (type))) |
| { |
| if (INTEGRAL_TYPE_P (itype) && TYPE_OVERFLOW_WRAPS (itype)) |
| { |
| /* Split the unconverted operand and try to prove that |
| wrapping isn't a problem. */ |
| tree tmp_var, tmp_off; |
| split_constant_offset (op0, &tmp_var, &tmp_off, cache, limit); |
| |
| /* See whether we have an SSA_NAME whose range is known |
| to be [A, B]. */ |
| if (TREE_CODE (tmp_var) != SSA_NAME) |
| return false; |
| wide_int var_min, var_max; |
| value_range_kind vr_type = get_range_info (tmp_var, &var_min, |
| &var_max); |
| wide_int var_nonzero = get_nonzero_bits (tmp_var); |
| signop sgn = TYPE_SIGN (itype); |
| if (intersect_range_with_nonzero_bits (vr_type, &var_min, |
| &var_max, var_nonzero, |
| sgn) != VR_RANGE) |
| return false; |
| |
| /* See whether the range of OP0 (i.e. TMP_VAR + TMP_OFF) |
| is known to be [A + TMP_OFF, B + TMP_OFF], with all |
| operations done in ITYPE. The addition must overflow |
| at both ends of the range or at neither. */ |
| wi::overflow_type overflow[2]; |
| unsigned int prec = TYPE_PRECISION (itype); |
| wide_int woff = wi::to_wide (tmp_off, prec); |
| wide_int op0_min = wi::add (var_min, woff, sgn, &overflow[0]); |
| wi::add (var_max, woff, sgn, &overflow[1]); |
| if ((overflow[0] != wi::OVF_NONE) != (overflow[1] != wi::OVF_NONE)) |
| return false; |
| |
| /* Calculate (ssizetype) OP0 - (ssizetype) TMP_VAR. */ |
| widest_int diff = (widest_int::from (op0_min, sgn) |
| - widest_int::from (var_min, sgn)); |
| var0 = tmp_var; |
| *off = wide_int_to_tree (ssizetype, diff); |
| } |
| else |
| split_constant_offset (op0, &var0, off, cache, limit); |
| *var = fold_convert (type, var0); |
| return true; |
| } |
| return false; |
| } |
| |
| default: |
| return false; |
| } |
| } |
| |
| /* Expresses EXP as VAR + OFF, where off is a constant. The type of OFF |
| will be ssizetype. */ |
| |
| static void |
| split_constant_offset (tree exp, tree *var, tree *off, |
| hash_map<tree, std::pair<tree, tree> > &cache, |
| unsigned *limit) |
| { |
| tree type = TREE_TYPE (exp), op0, op1, e, o; |
| enum tree_code code; |
| |
| *var = exp; |
| *off = ssize_int (0); |
| |
| if (tree_is_chrec (exp) |
| || get_gimple_rhs_class (TREE_CODE (exp)) == GIMPLE_TERNARY_RHS) |
| return; |
| |
| code = TREE_CODE (exp); |
| extract_ops_from_tree (exp, &code, &op0, &op1); |
| if (split_constant_offset_1 (type, op0, code, op1, &e, &o, cache, limit)) |
| { |
| *var = e; |
| *off = o; |
| } |
| } |
| |
| void |
| split_constant_offset (tree exp, tree *var, tree *off) |
| { |
| unsigned limit = PARAM_VALUE (PARAM_SSA_NAME_DEF_CHAIN_LIMIT); |
| static hash_map<tree, std::pair<tree, tree> > *cache; |
| if (!cache) |
| cache = new hash_map<tree, std::pair<tree, tree> > (37); |
| split_constant_offset (exp, var, off, *cache, &limit); |
| cache->empty (); |
| } |
| |
| /* Returns the address ADDR of an object in a canonical shape (without nop |
| casts, and with type of pointer to the object). */ |
| |
| static tree |
| canonicalize_base_object_address (tree addr) |
| { |
| tree orig = addr; |
| |
| STRIP_NOPS (addr); |
| |
| /* The base address may be obtained by casting from integer, in that case |
| keep the cast. */ |
| if (!POINTER_TYPE_P (TREE_TYPE (addr))) |
| return orig; |
| |
| if (TREE_CODE (addr) != ADDR_EXPR) |
| return addr; |
| |
| return build_fold_addr_expr (TREE_OPERAND (addr, 0)); |
| } |
| |
| /* Analyze the behavior of memory reference REF within STMT. |
| There are two modes: |
| |
| - BB analysis. In this case we simply split the address into base, |
| init and offset components, without reference to any containing loop. |
| The resulting base and offset are general expressions and they can |
| vary arbitrarily from one iteration of the containing loop to the next. |
| The step is always zero. |
| |
| - loop analysis. In this case we analyze the reference both wrt LOOP |
| and on the basis that the reference occurs (is "used") in LOOP; |
| see the comment above analyze_scalar_evolution_in_loop for more |
| information about this distinction. The base, init, offset and |
| step fields are all invariant in LOOP. |
| |
| Perform BB analysis if LOOP is null, or if LOOP is the function's |
| dummy outermost loop. In other cases perform loop analysis. |
| |
| Return true if the analysis succeeded and store the results in DRB if so. |
| BB analysis can only fail for bitfield or reversed-storage accesses. */ |
| |
| opt_result |
| dr_analyze_innermost (innermost_loop_behavior *drb, tree ref, |
| struct loop *loop, const gimple *stmt) |
| { |
| poly_int64 pbitsize, pbitpos; |
| tree base, poffset; |
| machine_mode pmode; |
| int punsignedp, preversep, pvolatilep; |
| affine_iv base_iv, offset_iv; |
| tree init, dinit, step; |
| bool in_loop = (loop && loop->num); |
| |
| if (dump_file && (dump_flags & TDF_DETAILS)) |
| fprintf (dump_file, "analyze_innermost: "); |
| |
| base = get_inner_reference (ref, &pbitsize, &pbitpos, &poffset, &pmode, |
| &punsignedp, &preversep, &pvolatilep); |
| gcc_assert (base != NULL_TREE); |
| |
| poly_int64 pbytepos; |
| if (!multiple_p (pbitpos, BITS_PER_UNIT, &pbytepos)) |
| return opt_result::failure_at (stmt, |
| "failed: bit offset alignment.\n"); |
| |
| if (preversep) |
| return opt_result::failure_at (stmt, |
| "failed: reverse storage order.\n"); |
| |
| /* Calculate the alignment and misalignment for the inner reference. */ |
| unsigned int HOST_WIDE_INT bit_base_misalignment; |
| unsigned int bit_base_alignment; |
| get_object_alignment_1 (base, &bit_base_alignment, &bit_base_misalignment); |
| |
| /* There are no bitfield references remaining in BASE, so the values |
| we got back must be whole bytes. */ |
| gcc_assert (bit_base_alignment % BITS_PER_UNIT == 0 |
| && bit_base_misalignment % BITS_PER_UNIT == 0); |
| unsigned int base_alignment = bit_base_alignment / BITS_PER_UNIT; |
| poly_int64 base_misalignment = bit_base_misalignment / BITS_PER_UNIT; |
| |
| if (TREE_CODE (base) == MEM_REF) |
| { |
| if (!integer_zerop (TREE_OPERAND (base, 1))) |
| { |
| /* Subtract MOFF from the base and add it to POFFSET instead. |
| Adjust the misalignment to reflect the amount we subtracted. */ |
| poly_offset_int moff = mem_ref_offset (base); |
| base_misalignment -= moff.force_shwi (); |
| tree mofft = wide_int_to_tree (sizetype, moff); |
| if (!poffset) |
| poffset = mofft; |
| else |
| poffset = size_binop (PLUS_EXPR, poffset, mofft); |
| } |
| base = TREE_OPERAND (base, 0); |
| } |
| else |
| base = build_fold_addr_expr (base); |
| |
| if (in_loop) |
| { |
| if (!simple_iv (loop, loop, base, &base_iv, true)) |
| return opt_result::failure_at |
| (stmt, "failed: evolution of base is not affine.\n"); |
| } |
| else |
| { |
| base_iv.base = base; |
| base_iv.step = ssize_int (0); |
| base_iv.no_overflow = true; |
| } |
| |
| if (!poffset) |
| { |
| offset_iv.base = ssize_int (0); |
| offset_iv.step = ssize_int (0); |
| } |
| else |
| { |
| if (!in_loop) |
| { |
| offset_iv.base = poffset; |
| offset_iv.step = ssize_int (0); |
| } |
| else if (!simple_iv (loop, loop, poffset, &offset_iv, true)) |
| return opt_result::failure_at |
| (stmt, "failed: evolution of offset is not affine.\n"); |
| } |
| |
| init = ssize_int (pbytepos); |
| |
| /* Subtract any constant component from the base and add it to INIT instead. |
| Adjust the misalignment to reflect the amount we subtracted. */ |
| split_constant_offset (base_iv.base, &base_iv.base, &dinit); |
| init = size_binop (PLUS_EXPR, init, dinit); |
| base_misalignment -= TREE_INT_CST_LOW (dinit); |
| |
| split_constant_offset (offset_iv.base, &offset_iv.base, &dinit); |
| init = size_binop (PLUS_EXPR, init, dinit); |
| |
| step = size_binop (PLUS_EXPR, |
| fold_convert (ssizetype, base_iv.step), |
| fold_convert (ssizetype, offset_iv.step)); |
| |
| base = canonicalize_base_object_address (base_iv.base); |
| |
| /* See if get_pointer_alignment can guarantee a higher alignment than |
| the one we calculated above. */ |
| unsigned int HOST_WIDE_INT alt_misalignment; |
| unsigned int alt_alignment; |
| get_pointer_alignment_1 (base, &alt_alignment, &alt_misalignment); |
| |
| /* As above, these values must be whole bytes. */ |
| gcc_assert (alt_alignment % BITS_PER_UNIT == 0 |
| && alt_misalignment % BITS_PER_UNIT == 0); |
| alt_alignment /= BITS_PER_UNIT; |
| alt_misalignment /= BITS_PER_UNIT; |
| |
| if (base_alignment < alt_alignment) |
| { |
| base_alignment = alt_alignment; |
| base_misalignment = alt_misalignment; |
| } |
| |
| drb->base_address = base; |
| drb->offset = fold_convert (ssizetype, offset_iv.base); |
| drb->init = init; |
| drb->step = step; |
| if (known_misalignment (base_misalignment, base_alignment, |
| &drb->base_misalignment)) |
| drb->base_alignment = base_alignment; |
| else |
| { |
| drb->base_alignment = known_alignment (base_misalignment); |
| drb->base_misalignment = 0; |
| } |
| drb->offset_alignment = highest_pow2_factor (offset_iv.base); |
| drb->step_alignment = highest_pow2_factor (step); |
| |
| if (dump_file && (dump_flags & TDF_DETAILS)) |
| fprintf (dump_file, "success.\n"); |
| |
| return opt_result::success (); |
| } |
| |
| /* Return true if OP is a valid component reference for a DR access |
| function. This accepts a subset of what handled_component_p accepts. */ |
| |
| static bool |
| access_fn_component_p (tree op) |
| { |
| switch (TREE_CODE (op)) |
| { |
| case REALPART_EXPR: |
| case IMAGPART_EXPR: |
| case ARRAY_REF: |
| return true; |
| |
| case COMPONENT_REF: |
| return TREE_CODE (TREE_TYPE (TREE_OPERAND (op, 0))) == RECORD_TYPE; |
| |
| default: |
| return false; |
| } |
| } |
| |
| /* Determines the base object and the list of indices of memory reference |
| DR, analyzed in LOOP and instantiated before NEST. */ |
| |
| static void |
| dr_analyze_indices (struct data_reference *dr, edge nest, loop_p loop) |
| { |
| vec<tree> access_fns = vNULL; |
| tree ref, op; |
| tree base, off, access_fn; |
| |
| /* If analyzing a basic-block there are no indices to analyze |
| and thus no access functions. */ |
| if (!nest) |
| { |
| DR_BASE_OBJECT (dr) = DR_REF (dr); |
| DR_ACCESS_FNS (dr).create (0); |
| return; |
| } |
| |
| ref = DR_REF (dr); |
| |
| /* REALPART_EXPR and IMAGPART_EXPR can be handled like accesses |
| into a two element array with a constant index. The base is |
| then just the immediate underlying object. */ |
| if (TREE_CODE (ref) == REALPART_EXPR) |
| { |
| ref = TREE_OPERAND (ref, 0); |
| access_fns.safe_push (integer_zero_node); |
| } |
| else if (TREE_CODE (ref) == IMAGPART_EXPR) |
| { |
| ref = TREE_OPERAND (ref, 0); |
| access_fns.safe_push (integer_one_node); |
| } |
| |
| /* Analyze access functions of dimensions we know to be independent. |
| The list of component references handled here should be kept in |
| sync with access_fn_component_p. */ |
| while (handled_component_p (ref)) |
| { |
| if (TREE_CODE (ref) == ARRAY_REF) |
| { |
| op = TREE_OPERAND (ref, 1); |
| access_fn = analyze_scalar_evolution (loop, op); |
| access_fn = instantiate_scev (nest, loop, access_fn); |
| access_fns.safe_push (access_fn); |
| } |
| else if (TREE_CODE (ref) == COMPONENT_REF |
| && TREE_CODE (TREE_TYPE (TREE_OPERAND (ref, 0))) == RECORD_TYPE) |
| { |
| /* For COMPONENT_REFs of records (but not unions!) use the |
| FIELD_DECL offset as constant access function so we can |
| disambiguate a[i].f1 and a[i].f2. */ |
| tree off = component_ref_field_offset (ref); |
| off = size_binop (PLUS_EXPR, |
| size_binop (MULT_EXPR, |
| fold_convert (bitsizetype, off), |
| bitsize_int (BITS_PER_UNIT)), |
| DECL_FIELD_BIT_OFFSET (TREE_OPERAND (ref, 1))); |
| access_fns.safe_push (off); |
| } |
| else |
| /* If we have an unhandled component we could not translate |
| to an access function stop analyzing. We have determined |
| our base object in this case. */ |
| break; |
| |
| ref = TREE_OPERAND (ref, 0); |
| } |
| |
| /* If the address operand of a MEM_REF base has an evolution in the |
| analyzed nest, add it as an additional independent access-function. */ |
| if (TREE_CODE (ref) == MEM_REF) |
| { |
| op = TREE_OPERAND (ref, 0); |
| access_fn = analyze_scalar_evolution (loop, op); |
| access_fn = instantiate_scev (nest, loop, access_fn); |
| if (TREE_CODE (access_fn) == POLYNOMIAL_CHREC) |
| { |
| tree orig_type; |
| tree memoff = TREE_OPERAND (ref, 1); |
| base = initial_condition (access_fn); |
| orig_type = TREE_TYPE (base); |
| STRIP_USELESS_TYPE_CONVERSION (base); |
| split_constant_offset (base, &base, &off); |
| STRIP_USELESS_TYPE_CONVERSION (base); |
| /* Fold the MEM_REF offset into the evolutions initial |
| value to make more bases comparable. */ |
| if (!integer_zerop (memoff)) |
| { |
| off = size_binop (PLUS_EXPR, off, |
| fold_convert (ssizetype, memoff)); |
| memoff = build_int_cst (TREE_TYPE (memoff), 0); |
| } |
| /* Adjust the offset so it is a multiple of the access type |
| size and thus we separate bases that can possibly be used |
| to produce partial overlaps (which the access_fn machinery |
| cannot handle). */ |
| wide_int rem; |
| if (TYPE_SIZE_UNIT (TREE_TYPE (ref)) |
| && TREE_CODE (TYPE_SIZE_UNIT (TREE_TYPE (ref))) == INTEGER_CST |
| && !integer_zerop (TYPE_SIZE_UNIT (TREE_TYPE (ref)))) |
| rem = wi::mod_trunc |
| (wi::to_wide (off), |
| wi::to_wide (TYPE_SIZE_UNIT (TREE_TYPE (ref))), |
| SIGNED); |
| else |
| /* If we can't compute the remainder simply force the initial |
| condition to zero. */ |
| rem = wi::to_wide (off); |
| off = wide_int_to_tree (ssizetype, wi::to_wide (off) - rem); |
| memoff = wide_int_to_tree (TREE_TYPE (memoff), rem); |
| /* And finally replace the initial condition. */ |
| access_fn = chrec_replace_initial_condition |
| (access_fn, fold_convert (orig_type, off)); |
| /* ??? This is still not a suitable base object for |
| dr_may_alias_p - the base object needs to be an |
| access that covers the object as whole. With |
| an evolution in the pointer this cannot be |
| guaranteed. |
| As a band-aid, mark the access so we can special-case |
| it in dr_may_alias_p. */ |
| tree old = ref; |
| ref = fold_build2_loc (EXPR_LOCATION (ref), |
| MEM_REF, TREE_TYPE (ref), |
| base, memoff); |
| MR_DEPENDENCE_CLIQUE (ref) = MR_DEPENDENCE_CLIQUE (old); |
| MR_DEPENDENCE_BASE (ref) = MR_DEPENDENCE_BASE (old); |
| DR_UNCONSTRAINED_BASE (dr) = true; |
| access_fns.safe_push (access_fn); |
| } |
| } |
| else if (DECL_P (ref)) |
| { |
| /* Canonicalize DR_BASE_OBJECT to MEM_REF form. */ |
| ref = build2 (MEM_REF, TREE_TYPE (ref), |
| build_fold_addr_expr (ref), |
| build_int_cst (reference_alias_ptr_type (ref), 0)); |
| } |
| |
| DR_BASE_OBJECT (dr) = ref; |
| DR_ACCESS_FNS (dr) = access_fns; |
| } |
| |
| /* Extracts the alias analysis information from the memory reference DR. */ |
| |
| static void |
| dr_analyze_alias (struct data_reference *dr) |
| { |
| tree ref = DR_REF (dr); |
| tree base = get_base_address (ref), addr; |
| |
| if (INDIRECT_REF_P (base) |
| || TREE_CODE (base) == MEM_REF) |
| { |
| addr = TREE_OPERAND (base, 0); |
| if (TREE_CODE (addr) == SSA_NAME) |
| DR_PTR_INFO (dr) = SSA_NAME_PTR_INFO (addr); |
| } |
| } |
| |
| /* Frees data reference DR. */ |
| |
| void |
| free_data_ref (data_reference_p dr) |
| { |
| DR_ACCESS_FNS (dr).release (); |
| free (dr); |
| } |
| |
| /* Analyze memory reference MEMREF, which is accessed in STMT. |
| The reference is a read if IS_READ is true, otherwise it is a write. |
| IS_CONDITIONAL_IN_STMT indicates that the reference is conditional |
| within STMT, i.e. that it might not occur even if STMT is executed |
| and runs to completion. |
| |
| Return the data_reference description of MEMREF. NEST is the outermost |
| loop in which the reference should be instantiated, LOOP is the loop |
| in which the data reference should be analyzed. */ |
| |
| struct data_reference * |
| create_data_ref (edge nest, loop_p loop, tree memref, gimple *stmt, |
| bool is_read, bool is_conditional_in_stmt) |
| { |
| struct data_reference *dr; |
| |
| if (dump_file && (dump_flags & TDF_DETAILS)) |
| { |
| fprintf (dump_file, "Creating dr for "); |
| print_generic_expr (dump_file, memref, TDF_SLIM); |
| fprintf (dump_file, "\n"); |
| } |
| |
| dr = XCNEW (struct data_reference); |
| DR_STMT (dr) = stmt; |
| DR_REF (dr) = memref; |
| DR_IS_READ (dr) = is_read; |
| DR_IS_CONDITIONAL_IN_STMT (dr) = is_conditional_in_stmt; |
| |
| dr_analyze_innermost (&DR_INNERMOST (dr), memref, |
| nest != NULL ? loop : NULL, stmt); |
| dr_analyze_indices (dr, nest, loop); |
| dr_analyze_alias (dr); |
| |
| if (dump_file && (dump_flags & TDF_DETAILS)) |
| { |
| unsigned i; |
| fprintf (dump_file, "\tbase_address: "); |
| print_generic_expr (dump_file, DR_BASE_ADDRESS (dr), TDF_SLIM); |
| fprintf (dump_file, "\n\toffset from base address: "); |
| print_generic_expr (dump_file, DR_OFFSET (dr), TDF_SLIM); |
| fprintf (dump_file, "\n\tconstant offset from base address: "); |
| print_generic_expr (dump_file, DR_INIT (dr), TDF_SLIM); |
| fprintf (dump_file, "\n\tstep: "); |
| print_generic_expr (dump_file, DR_STEP (dr), TDF_SLIM); |
| fprintf (dump_file, "\n\tbase alignment: %d", DR_BASE_ALIGNMENT (dr)); |
| fprintf (dump_file, "\n\tbase misalignment: %d", |
| DR_BASE_MISALIGNMENT (dr)); |
| fprintf (dump_file, "\n\toffset alignment: %d", |
| DR_OFFSET_ALIGNMENT (dr)); |
| fprintf (dump_file, "\n\tstep alignment: %d", DR_STEP_ALIGNMENT (dr)); |
| fprintf (dump_file, "\n\tbase_object: "); |
| print_generic_expr (dump_file, DR_BASE_OBJECT (dr), TDF_SLIM); |
| fprintf (dump_file, "\n"); |
| for (i = 0; i < DR_NUM_DIMENSIONS (dr); i++) |
| { |
| fprintf (dump_file, "\tAccess function %d: ", i); |
| print_generic_stmt (dump_file, DR_ACCESS_FN (dr, i), TDF_SLIM); |
| } |
| } |
| |
| return dr; |
| } |
| |
| /* A helper function computes order between two tree epxressions T1 and T2. |
| This is used in comparator functions sorting objects based on the order |
| of tree expressions. The function returns -1, 0, or 1. */ |
| |
| int |
| data_ref_compare_tree (tree t1, tree t2) |
| { |
| int i, cmp; |
| enum tree_code code; |
| char tclass; |
| |
| if (t1 == t2) |
| return 0; |
| if (t1 == NULL) |
| return -1; |
| if (t2 == NULL) |
| return 1; |
| |
| STRIP_USELESS_TYPE_CONVERSION (t1); |
| STRIP_USELESS_TYPE_CONVERSION (t2); |
| if (t1 == t2) |
| return 0; |
| |
| if (TREE_CODE (t1) != TREE_CODE (t2) |
| && ! (CONVERT_EXPR_P (t1) && CONVERT_EXPR_P (t2))) |
| return TREE_CODE (t1) < TREE_CODE (t2) ? -1 : 1; |
| |
| code = TREE_CODE (t1); |
| switch (code) |
| { |
| case INTEGER_CST: |
| return tree_int_cst_compare (t1, t2); |
| |
| case STRING_CST: |
| if (TREE_STRING_LENGTH (t1) != TREE_STRING_LENGTH (t2)) |
| return TREE_STRING_LENGTH (t1) < TREE_STRING_LENGTH (t2) ? -1 : 1; |
| return memcmp (TREE_STRING_POINTER (t1), TREE_STRING_POINTER (t2), |
| TREE_STRING_LENGTH (t1)); |
| |
| case SSA_NAME: |
| if (SSA_NAME_VERSION (t1) != SSA_NAME_VERSION (t2)) |
| return SSA_NAME_VERSION (t1) < SSA_NAME_VERSION (t2) ? -1 : 1; |
| break; |
| |
| default: |
| if (POLY_INT_CST_P (t1)) |
| return compare_sizes_for_sort (wi::to_poly_widest (t1), |
| wi::to_poly_widest (t2)); |
| |
| tclass = TREE_CODE_CLASS (code); |
| |
| /* For decls, compare their UIDs. */ |
| if (tclass == tcc_declaration) |
| { |
| if (DECL_UID (t1) != DECL_UID (t2)) |
| return DECL_UID (t1) < DECL_UID (t2) ? -1 : 1; |
| break; |
| } |
| /* For expressions, compare their operands recursively. */ |
| else if (IS_EXPR_CODE_CLASS (tclass)) |
| { |
| for (i = TREE_OPERAND_LENGTH (t1) - 1; i >= 0; --i) |
| { |
| cmp = data_ref_compare_tree (TREE_OPERAND (t1, i), |
| TREE_OPERAND (t2, i)); |
| if (cmp != 0) |
| return cmp; |
| } |
| } |
| else |
| gcc_unreachable (); |
| } |
| |
| return 0; |
| } |
| |
| /* Return TRUE it's possible to resolve data dependence DDR by runtime alias |
| check. */ |
| |
| opt_result |
| runtime_alias_check_p (ddr_p ddr, struct loop *loop, bool speed_p) |
| { |
| if (dump_enabled_p ()) |
| dump_printf (MSG_NOTE, |
| "consider run-time aliasing test between %T and %T\n", |
| DR_REF (DDR_A (ddr)), DR_REF (DDR_B (ddr))); |
| |
| if (!speed_p) |
| return opt_result::failure_at (DR_STMT (DDR_A (ddr)), |
| "runtime alias check not supported when" |
| " optimizing for size.\n"); |
| |
| /* FORNOW: We don't support versioning with outer-loop in either |
| vectorization or loop distribution. */ |
| if (loop != NULL && loop->inner != NULL) |
| return opt_result::failure_at (DR_STMT (DDR_A (ddr)), |
| "runtime alias check not supported for" |
| " outer loop.\n"); |
| |
| return opt_result::success (); |
| } |
| |
| /* Operator == between two dr_with_seg_len objects. |
| |
| This equality operator is used to make sure two data refs |
| are the same one so that we will consider to combine the |
| aliasing checks of those two pairs of data dependent data |
| refs. */ |
| |
| static bool |
| operator == (const dr_with_seg_len& d1, |
| const dr_with_seg_len& d2) |
| { |
| return (operand_equal_p (DR_BASE_ADDRESS (d1.dr), |
| DR_BASE_ADDRESS (d2.dr), 0) |
| && data_ref_compare_tree (DR_OFFSET (d1.dr), DR_OFFSET (d2.dr)) == 0 |
| && data_ref_compare_tree (DR_INIT (d1.dr), DR_INIT (d2.dr)) == 0 |
| && data_ref_compare_tree (d1.seg_len, d2.seg_len) == 0 |
| && known_eq (d1.access_size, d2.access_size) |
| && d1.align == d2.align); |
| } |
| |
| /* Comparison function for sorting objects of dr_with_seg_len_pair_t |
| so that we can combine aliasing checks in one scan. */ |
| |
| static int |
| comp_dr_with_seg_len_pair (const void *pa_, const void *pb_) |
| { |
| const dr_with_seg_len_pair_t* pa = (const dr_with_seg_len_pair_t *) pa_; |
| const dr_with_seg_len_pair_t* pb = (const dr_with_seg_len_pair_t *) pb_; |
| const dr_with_seg_len &a1 = pa->first, &a2 = pa->second; |
| const dr_with_seg_len &b1 = pb->first, &b2 = pb->second; |
| |
| /* For DR pairs (a, b) and (c, d), we only consider to merge the alias checks |
| if a and c have the same basic address snd step, and b and d have the same |
| address and step. Therefore, if any a&c or b&d don't have the same address |
| and step, we don't care the order of those two pairs after sorting. */ |
| int comp_res; |
| |
| if ((comp_res = data_ref_compare_tree (DR_BASE_ADDRESS (a1.dr), |
| DR_BASE_ADDRESS (b1.dr))) != 0) |
| return comp_res; |
| if ((comp_res = data_ref_compare_tree (DR_BASE_ADDRESS (a2.dr), |
| DR_BASE_ADDRESS (b2.dr))) != 0) |
| return comp_res; |
| if ((comp_res = data_ref_compare_tree (DR_STEP (a1.dr), |
| DR_STEP (b1.dr))) != 0) |
| return comp_res; |
| if ((comp_res = data_ref_compare_tree (DR_STEP (a2.dr), |
| DR_STEP (b2.dr))) != 0) |
| return comp_res; |
| if ((comp_res = data_ref_compare_tree (DR_OFFSET (a1.dr), |
| DR_OFFSET (b1.dr))) != 0) |
| return comp_res; |
| if ((comp_res = data_ref_compare_tree (DR_INIT (a1.dr), |
| DR_INIT (b1.dr))) != 0) |
| return comp_res; |
| if ((comp_res = data_ref_compare_tree (DR_OFFSET (a2.dr), |
| DR_OFFSET (b2.dr))) != 0) |
| return comp_res; |
| if ((comp_res = data_ref_compare_tree (DR_INIT (a2.dr), |
| DR_INIT (b2.dr))) != 0) |
| return comp_res; |
| |
| return 0; |
| } |
| |
| /* Merge alias checks recorded in ALIAS_PAIRS and remove redundant ones. |
| FACTOR is number of iterations that each data reference is accessed. |
| |
| Basically, for each pair of dependent data refs store_ptr_0 & load_ptr_0, |
| we create an expression: |
| |
| ((store_ptr_0 + store_segment_length_0) <= load_ptr_0) |
| || (load_ptr_0 + load_segment_length_0) <= store_ptr_0)) |
| |
| for aliasing checks. However, in some cases we can decrease the number |
| of checks by combining two checks into one. For example, suppose we have |
| another pair of data refs store_ptr_0 & load_ptr_1, and if the following |
| condition is satisfied: |
| |
| load_ptr_0 < load_ptr_1 && |
| load_ptr_1 - load_ptr_0 - load_segment_length_0 < store_segment_length_0 |
| |
| (this condition means, in each iteration of vectorized loop, the accessed |
| memory of store_ptr_0 cannot be between the memory of load_ptr_0 and |
| load_ptr_1.) |
| |
| we then can use only the following expression to finish the alising checks |
| between store_ptr_0 & load_ptr_0 and store_ptr_0 & load_ptr_1: |
| |
| ((store_ptr_0 + store_segment_length_0) <= load_ptr_0) |
| || (load_ptr_1 + load_segment_length_1 <= store_ptr_0)) |
| |
| Note that we only consider that load_ptr_0 and load_ptr_1 have the same |
| basic address. */ |
| |
| void |
| prune_runtime_alias_test_list (vec<dr_with_seg_len_pair_t> *alias_pairs, |
| poly_uint64) |
| { |
| /* Sort the collected data ref pairs so that we can scan them once to |
| combine all possible aliasing checks. */ |
| alias_pairs->qsort (comp_dr_with_seg_len_pair); |
| |
| /* Scan the sorted dr pairs and check if we can combine alias checks |
| of two neighboring dr pairs. */ |
| for (size_t i = 1; i < alias_pairs->length (); ++i) |
| { |
| /* Deal with two ddrs (dr_a1, dr_b1) and (dr_a2, dr_b2). */ |
| dr_with_seg_len *dr_a1 = &(*alias_pairs)[i-1].first, |
| *dr_b1 = &(*alias_pairs)[i-1].second, |
| *dr_a2 = &(*alias_pairs)[i].first, |
| *dr_b2 = &(*alias_pairs)[i].second; |
| |
| /* Remove duplicate data ref pairs. */ |
| if (*dr_a1 == *dr_a2 && *dr_b1 == *dr_b2) |
| { |
| if (dump_enabled_p ()) |
| dump_printf (MSG_NOTE, "found equal ranges %T, %T and %T, %T\n", |
| DR_REF (dr_a1->dr), DR_REF (dr_b1->dr), |
| DR_REF (dr_a2->dr), DR_REF (dr_b2->dr)); |
| alias_pairs->ordered_remove (i--); |
| continue; |
| } |
| |
| if (*dr_a1 == *dr_a2 || *dr_b1 == *dr_b2) |
| { |
| /* We consider the case that DR_B1 and DR_B2 are same memrefs, |
| and DR_A1 and DR_A2 are two consecutive memrefs. */ |
| if (*dr_a1 == *dr_a2) |
| { |
| std::swap (dr_a1, dr_b1); |
| std::swap (dr_a2, dr_b2); |
| } |
| |
| poly_int64 init_a1, init_a2; |
| /* Only consider cases in which the distance between the initial |
| DR_A1 and the initial DR_A2 is known at compile time. */ |
| if (!operand_equal_p (DR_BASE_ADDRESS (dr_a1->dr), |
| DR_BASE_ADDRESS (dr_a2->dr), 0) |
| || !operand_equal_p (DR_OFFSET (dr_a1->dr), |
| DR_OFFSET (dr_a2->dr), 0) |
| || !poly_int_tree_p (DR_INIT (dr_a1->dr), &init_a1) |
| || !poly_int_tree_p (DR_INIT (dr_a2->dr), &init_a2)) |
| continue; |
| |
| /* Don't combine if we can't tell which one comes first. */ |
| if (!ordered_p (init_a1, init_a2)) |
| continue; |
| |
| /* Make sure dr_a1 starts left of dr_a2. */ |
| if (maybe_gt (init_a1, init_a2)) |
| { |
| std::swap (*dr_a1, *dr_a2); |
| std::swap (init_a1, init_a2); |
| } |
| |
| /* Work out what the segment length would be if we did combine |
| DR_A1 and DR_A2: |
| |
| - If DR_A1 and DR_A2 have equal lengths, that length is |
| also the combined length. |
| |
| - If DR_A1 and DR_A2 both have negative "lengths", the combined |
| length is the lower bound on those lengths. |
| |
| - If DR_A1 and DR_A2 both have positive lengths, the combined |
| length is the upper bound on those lengths. |
| |
| Other cases are unlikely to give a useful combination. |
| |
| The lengths both have sizetype, so the sign is taken from |
| the step instead. */ |
| if (!operand_equal_p (dr_a1->seg_len, dr_a2->seg_len, 0)) |
| { |
| poly_uint64 seg_len_a1, seg_len_a2; |
| if (!poly_int_tree_p (dr_a1->seg_len, &seg_len_a1) |
| || !poly_int_tree_p (dr_a2->seg_len, &seg_len_a2)) |
| continue; |
| |
| tree indicator_a = dr_direction_indicator (dr_a1->dr); |
| if (TREE_CODE (indicator_a) != INTEGER_CST) |
| continue; |
| |
| tree indicator_b = dr_direction_indicator (dr_a2->dr); |
| if (TREE_CODE (indicator_b) != INTEGER_CST) |
| continue; |
| |
| int sign_a = tree_int_cst_sgn (indicator_a); |
| int sign_b = tree_int_cst_sgn (indicator_b); |
| |
| poly_uint64 new_seg_len; |
| if (sign_a <= 0 && sign_b <= 0) |
| new_seg_len = lower_bound (seg_len_a1, seg_len_a2); |
| else if (sign_a >= 0 && sign_b >= 0) |
| new_seg_len = upper_bound (seg_len_a1, seg_len_a2); |
| else |
| continue; |
| |
| dr_a1->seg_len = build_int_cst (TREE_TYPE (dr_a1->seg_len), |
| new_seg_len); |
| dr_a1->align = MIN (dr_a1->align, known_alignment (new_seg_len)); |
| } |
| |
| /* This is always positive due to the swap above. */ |
| poly_uint64 diff = init_a2 - init_a1; |
| |
| /* The new check will start at DR_A1. Make sure that its access |
| size encompasses the initial DR_A2. */ |
| if (maybe_lt (dr_a1->access_size, diff + dr_a2->access_size)) |
| { |
| dr_a1->access_size = upper_bound (dr_a1->access_size, |
| diff + dr_a2->access_size); |
| unsigned int new_align = known_alignment (dr_a1->access_size); |
| dr_a1->align = MIN (dr_a1->align, new_align); |
| } |
| if (dump_enabled_p ()) |
| dump_printf (MSG_NOTE, "merging ranges for %T, %T and %T, %T\n", |
| DR_REF (dr_a1->dr), DR_REF (dr_b1->dr), |
| DR_REF (dr_a2->dr), DR_REF (dr_b2->dr)); |
| alias_pairs->ordered_remove (i); |
| i--; |
| } |
| } |
| } |
| |
| /* Given LOOP's two data references and segment lengths described by DR_A |
| and DR_B, create expression checking if the two addresses ranges intersect |
| with each other based on index of the two addresses. This can only be |
| done if DR_A and DR_B referring to the same (array) object and the index |
| is the only difference. For example: |
| |
| DR_A DR_B |
| data-ref arr[i] arr[j] |
| base_object arr arr |
| index {i_0, +, 1}_loop {j_0, +, 1}_loop |
| |
| The addresses and their index are like: |
| |
| |<- ADDR_A ->| |<- ADDR_B ->| |
| -------------------------------------------------------> |
| | | | | | | | | | | |
| -------------------------------------------------------> |
| i_0 ... i_0+4 j_0 ... j_0+4 |
| |
| We can create expression based on index rather than address: |
| |
| (i_0 + 4 < j_0 || j_0 + 4 < i_0) |
| |
| Note evolution step of index needs to be considered in comparison. */ |
| |
| static bool |
| create_intersect_range_checks_index (struct loop *loop, tree *cond_expr, |
| const dr_with_seg_len& dr_a, |
| const dr_with_seg_len& dr_b) |
| { |
| if (integer_zerop (DR_STEP (dr_a.dr)) |
| || integer_zerop (DR_STEP (dr_b.dr)) |
| || DR_NUM_DIMENSIONS (dr_a.dr) != DR_NUM_DIMENSIONS (dr_b.dr)) |
| return false; |
| |
| poly_uint64 seg_len1, seg_len2; |
| if (!poly_int_tree_p (dr_a.seg_len, &seg_len1) |
| || !poly_int_tree_p (dr_b.seg_len, &seg_len2)) |
| return false; |
| |
| if (!tree_fits_shwi_p (DR_STEP (dr_a.dr))) |
| return false; |
| |
| if (!operand_equal_p (DR_BASE_OBJECT (dr_a.dr), DR_BASE_OBJECT (dr_b.dr), 0)) |
| return false; |
| |
| if (!operand_equal_p (DR_STEP (dr_a.dr), DR_STEP (dr_b.dr), 0)) |
| return false; |
| |
| gcc_assert (TREE_CODE (DR_STEP (dr_a.dr)) == INTEGER_CST); |
| |
| bool neg_step = tree_int_cst_compare (DR_STEP (dr_a.dr), size_zero_node) < 0; |
| unsigned HOST_WIDE_INT abs_step = tree_to_shwi (DR_STEP (dr_a.dr)); |
| if (neg_step) |
| { |
| abs_step = -abs_step; |
| seg_len1 = -seg_len1; |
| seg_len2 = -seg_len2; |
| } |
| else |
| { |
| /* Include the access size in the length, so that we only have one |
| tree addition below. */ |
| seg_len1 += dr_a.access_size; |
| seg_len2 += dr_b.access_size; |
| } |
| |
| /* Infer the number of iterations with which the memory segment is accessed |
| by DR. In other words, alias is checked if memory segment accessed by |
| DR_A in some iterations intersect with memory segment accessed by DR_B |
| in the same amount iterations. |
| Note segnment length is a linear function of number of iterations with |
| DR_STEP as the coefficient. */ |
| poly_uint64 niter_len1, niter_len2; |
| if (!can_div_trunc_p (seg_len1 + abs_step - 1, abs_step, &niter_len1) |
| || !can_div_trunc_p (seg_len2 + abs_step - 1, abs_step, &niter_len2)) |
| return false; |
| |
| poly_uint64 niter_access1 = 0, niter_access2 = 0; |
| if (neg_step) |
| { |
| /* Divide each access size by the byte step, rounding up. */ |
| if (!can_div_trunc_p (dr_a.access_size - abs_step - 1, |
| abs_step, &niter_access1) |
| || !can_div_trunc_p (dr_b.access_size + abs_step - 1, |
| abs_step, &niter_access2)) |
| return false; |
| } |
| |
| unsigned int i; |
| for (i = 0; i < DR_NUM_DIMENSIONS (dr_a.dr); i++) |
| { |
| tree access1 = DR_ACCESS_FN (dr_a.dr, i); |
| tree access2 = DR_ACCESS_FN (dr_b.dr, i); |
| /* Two indices must be the same if they are not scev, or not scev wrto |
| current loop being vecorized. */ |
| if (TREE_CODE (access1) != POLYNOMIAL_CHREC |
| || TREE_CODE (access2) != POLYNOMIAL_CHREC |
| || CHREC_VARIABLE (access1) != (unsigned)loop->num |
| || CHREC_VARIABLE (access2) != (unsigned)loop->num) |
| { |
| if (operand_equal_p (access1, access2, 0)) |
| continue; |
| |
| return false; |
| } |
| /* The two indices must have the same step. */ |
| if (!operand_equal_p (CHREC_RIGHT (access1), CHREC_RIGHT (access2), 0)) |
| return false; |
| |
| tree idx_step = CHREC_RIGHT (access1); |
| /* Index must have const step, otherwise DR_STEP won't be constant. */ |
| gcc_assert (TREE_CODE (idx_step) == INTEGER_CST); |
| /* Index must evaluate in the same direction as DR. */ |
| gcc_assert (!neg_step || tree_int_cst_sign_bit (idx_step) == 1); |
| |
| tree min1 = CHREC_LEFT (access1); |
| tree min2 = CHREC_LEFT (access2); |
| if (!types_compatible_p (TREE_TYPE (min1), TREE_TYPE (min2))) |
| return false; |
| |
| /* Ideally, alias can be checked against loop's control IV, but we |
| need to prove linear mapping between control IV and reference |
| index. Although that should be true, we check against (array) |
| index of data reference. Like segment length, index length is |
| linear function of the number of iterations with index_step as |
| the coefficient, i.e, niter_len * idx_step. */ |
| tree idx_len1 = fold_build2 (MULT_EXPR, TREE_TYPE (min1), idx_step, |
| build_int_cst (TREE_TYPE (min1), |
| niter_len1)); |
| tree idx_len2 = fold_build2 (MULT_EXPR, TREE_TYPE (min2), idx_step, |
| build_int_cst (TREE_TYPE (min2), |
| niter_len2)); |
| tree max1 = fold_build2 (PLUS_EXPR, TREE_TYPE (min1), min1, idx_len1); |
| tree max2 = fold_build2 (PLUS_EXPR, TREE_TYPE (min2), min2, idx_len2); |
| /* Adjust ranges for negative step. */ |
| if (neg_step) |
| { |
| /* IDX_LEN1 and IDX_LEN2 are negative in this case. */ |
| std::swap (min1, max1); |
| std::swap (min2, max2); |
| |
| /* As with the lengths just calculated, we've measured the access |
| sizes in iterations, so multiply them by the index step. */ |
| tree idx_access1 |
| = fold_build2 (MULT_EXPR, TREE_TYPE (min1), idx_step, |
| build_int_cst (TREE_TYPE (min1), niter_access1)); |
| tree idx_access2 |
| = fold_build2 (MULT_EXPR, TREE_TYPE (min2), idx_step, |
| build_int_cst (TREE_TYPE (min2), niter_access2)); |
| |
| /* MINUS_EXPR because the above values are negative. */ |
| max1 = fold_build2 (MINUS_EXPR, TREE_TYPE (max1), max1, idx_access1); |
| max2 = fold_build2 (MINUS_EXPR, TREE_TYPE (max2), max2, idx_access2); |
| } |
| tree part_cond_expr |
| = fold_build2 (TRUTH_OR_EXPR, boolean_type_node, |
| fold_build2 (LE_EXPR, boolean_type_node, max1, min2), |
| fold_build2 (LE_EXPR, boolean_type_node, max2, min1)); |
| if (*cond_expr) |
| *cond_expr = fold_build2 (TRUTH_AND_EXPR, boolean_type_node, |
| *cond_expr, part_cond_expr); |
| else |
| *cond_expr = part_cond_expr; |
| } |
| return true; |
| } |
| |
| /* If ALIGN is nonzero, set up *SEQ_MIN_OUT and *SEQ_MAX_OUT so that for |
| every address ADDR accessed by D: |
| |
| *SEQ_MIN_OUT <= ADDR (== ADDR & -ALIGN) <= *SEQ_MAX_OUT |
| |
| In this case, every element accessed by D is aligned to at least |
| ALIGN bytes. |
| |
| If ALIGN is zero then instead set *SEG_MAX_OUT so that: |
| |
| *SEQ_MIN_OUT <= ADDR < *SEQ_MAX_OUT. */ |
| |
| static void |
| get_segment_min_max (const dr_with_seg_len &d, tree *seg_min_out, |
| tree *seg_max_out, HOST_WIDE_INT align) |
| { |
| /* Each access has the following pattern: |
| |
| <- |seg_len| -> |
| <--- A: -ve step ---> |
| +-----+-------+-----+-------+-----+ |
| | n-1 | ,.... | 0 | ..... | n-1 | |
| +-----+-------+-----+-------+-----+ |
| <--- B: +ve step ---> |
| <- |seg_len| -> |
| | |
| base address |
| |
| where "n" is the number of scalar iterations covered by the segment. |
| (This should be VF for a particular pair if we know that both steps |
| are the same, otherwise it will be the full number of scalar loop |
| iterations.) |
| |
| A is the range of bytes accessed when the step is negative, |
| B is the range when the step is positive. |
| |
| If the access size is "access_size" bytes, the lowest addressed byte is: |
| |
| base + (step < 0 ? seg_len : 0) [LB] |
| |
| and the highest addressed byte is always below: |
| |
| base + (step < 0 ? 0 : seg_len) + access_size [UB] |
| |
| Thus: |
| |
| LB <= ADDR < UB |
| |
| If ALIGN is nonzero, all three values are aligned to at least ALIGN |
| bytes, so: |
| |
| LB <= ADDR <= UB - ALIGN |
| |
| where "- ALIGN" folds naturally with the "+ access_size" and often |
| cancels it out. |
| |
| We don't try to simplify LB and UB beyond this (e.g. by using |
| MIN and MAX based on whether seg_len rather than the stride is |
| negative) because it is possible for the absolute size of the |
| segment to overflow the range of a ssize_t. |
| |
| Keeping the pointer_plus outside of the cond_expr should allow |
| the cond_exprs to be shared with other alias checks. */ |
| tree indicator = dr_direction_indicator (d.dr); |
| tree neg_step = fold_build2 (LT_EXPR, boolean_type_node, |
| fold_convert (ssizetype, indicator), |
| ssize_int (0)); |
| tree addr_base = fold_build_pointer_plus (DR_BASE_ADDRESS (d.dr), |
| DR_OFFSET (d.dr)); |
| addr_base = fold_build_pointer_plus (addr_base, DR_INIT (d.dr)); |
| tree seg_len |
| = fold_convert (sizetype, rewrite_to_non_trapping_overflow (d.seg_len)); |
| |
| tree min_reach = fold_build3 (COND_EXPR, sizetype, neg_step, |
| seg_len, size_zero_node); |
| tree max_reach = fold_build3 (COND_EXPR, sizetype, neg_step, |
| size_zero_node, seg_len); |
| max_reach = fold_build2 (PLUS_EXPR, sizetype, max_reach, |
| size_int (d.access_size - align)); |
| |
| *seg_min_out = fold_build_pointer_plus (addr_base, min_reach); |
| *seg_max_out = fold_build_pointer_plus (addr_base, max_reach); |
| } |
| |
| /* Given two data references and segment lengths described by DR_A and DR_B, |
| create expression checking if the two addresses ranges intersect with |
| each other: |
| |
| ((DR_A_addr_0 + DR_A_segment_length_0) <= DR_B_addr_0) |
| || (DR_B_addr_0 + DER_B_segment_length_0) <= DR_A_addr_0)) */ |
| |
| static void |
| create_intersect_range_checks (struct loop *loop, tree *cond_expr, |
| const dr_with_seg_len& dr_a, |
| const dr_with_seg_len& dr_b) |
| { |
| *cond_expr = NULL_TREE; |
| if (create_intersect_range_checks_index (loop, cond_expr, dr_a, dr_b)) |
| return; |
| |
| unsigned HOST_WIDE_INT min_align; |
| tree_code cmp_code; |
| if (TREE_CODE (DR_STEP (dr_a.dr)) == INTEGER_CST |
| && TREE_CODE (DR_STEP (dr_b.dr)) == INTEGER_CST) |
| { |
| /* In this case adding access_size to seg_len is likely to give |
| a simple X * step, where X is either the number of scalar |
| iterations or the vectorization factor. We're better off |
| keeping that, rather than subtracting an alignment from it. |
| |
| In this case the maximum values are exclusive and so there is |
| no alias if the maximum of one segment equals the minimum |
| of another. */ |
| min_align = 0; |
| cmp_code = LE_EXPR; |
| } |
| else |
| { |
| /* Calculate the minimum alignment shared by all four pointers, |
| then arrange for this alignment to be subtracted from the |
| exclusive maximum values to get inclusive maximum values. |
| This "- min_align" is cumulative with a "+ access_size" |
| in the calculation of the maximum values. In the best |
| (and common) case, the two cancel each other out, leaving |
| us with an inclusive bound based only on seg_len. In the |
| worst case we're simply adding a smaller number than before. |
| |
| Because the maximum values are inclusive, there is an alias |
| if the maximum value of one segment is equal to the minimum |
| value of the other. */ |
| min_align = MIN (dr_a.align, dr_b.align); |
| cmp_code = LT_EXPR; |
| } |
| |
| tree seg_a_min, seg_a_max, seg_b_min, seg_b_max; |
| get_segment_min_max (dr_a, &seg_a_min, &seg_a_max, min_align); |
| get_segment_min_max (dr_b, &seg_b_min, &seg_b_max, min_align); |
| |
| *cond_expr |
| = fold_build2 (TRUTH_OR_EXPR, boolean_type_node, |
| fold_build2 (cmp_code, boolean_type_node, seg_a_max, seg_b_min), |
| fold_build2 (cmp_code, boolean_type_node, seg_b_max, seg_a_min)); |
| } |
| |
| /* Create a conditional expression that represents the run-time checks for |
| overlapping of address ranges represented by a list of data references |
| pairs passed in ALIAS_PAIRS. Data references are in LOOP. The returned |
| COND_EXPR is the conditional expression to be used in the if statement |
| that controls which version of the loop gets executed at runtime. */ |
| |
| void |
| create_runtime_alias_checks (struct loop *loop, |
| vec<dr_with_seg_len_pair_t> *alias_pairs, |
| tree * cond_expr) |
| { |
| tree part_cond_expr; |
| |
| fold_defer_overflow_warnings (); |
| for (size_t i = 0, s = alias_pairs->length (); i < s; ++i) |
| { |
| const dr_with_seg_len& dr_a = (*alias_pairs)[i].first; |
| const dr_with_seg_len& dr_b = (*alias_pairs)[i].second; |
| |
| if (dump_enabled_p ()) |
| dump_printf (MSG_NOTE, |
| "create runtime check for data references %T and %T\n", |
| DR_REF (dr_a.dr), DR_REF (dr_b.dr)); |
| |
| /* Create condition expression for each pair data references. */ |
| create_intersect_range_checks (loop, &part_cond_expr, dr_a, dr_b); |
| if (*cond_expr) |
| *cond_expr = fold_build2 (TRUTH_AND_EXPR, boolean_type_node, |
| *cond_expr, part_cond_expr); |
| else |
| *cond_expr = part_cond_expr; |
| } |
| fold_undefer_and_ignore_overflow_warnings (); |
| } |
| |
| /* Check if OFFSET1 and OFFSET2 (DR_OFFSETs of some data-refs) are identical |
| expressions. */ |
| static bool |
| dr_equal_offsets_p1 (tree offset1, tree offset2) |
| { |
| bool res; |
| |
| STRIP_NOPS (offset1); |
| STRIP_NOPS (offset2); |
| |
| if (offset1 == offset2) |
| return true; |
| |
| if (TREE_CODE (offset1) != TREE_CODE (offset2) |
| || (!BINARY_CLASS_P (offset1) && !UNARY_CLASS_P (offset1))) |
| return false; |
| |
| res = dr_equal_offsets_p1 (TREE_OPERAND (offset1, 0), |
| TREE_OPERAND (offset2, 0)); |
| |
| if (!res || !BINARY_CLASS_P (offset1)) |
| return res; |
| |
| res = dr_equal_offsets_p1 (TREE_OPERAND (offset1, 1), |
| TREE_OPERAND (offset2, 1)); |
| |
| return res; |
| } |
| |
| /* Check if DRA and DRB have equal offsets. */ |
| bool |
| dr_equal_offsets_p (struct data_reference *dra, |
| struct data_reference *drb) |
| { |
| tree offset1, offset2; |
| |
| offset1 = DR_OFFSET (dra); |
| offset2 = DR_OFFSET (drb); |
| |
| return dr_equal_offsets_p1 (offset1, offset2); |
| } |
| |
| /* Returns true if FNA == FNB. */ |
| |
| static bool |
| affine_function_equal_p (affine_fn fna, affine_fn fnb) |
| { |
| unsigned i, n = fna.length (); |
| |
| if (n != fnb.length ()) |
| return false; |
| |
| for (i = 0; i < n; i++) |
| if (!operand_equal_p (fna[i], fnb[i], 0)) |
| return false; |
| |
| return true; |
| } |
| |
| /* If all the functions in CF are the same, returns one of them, |
| otherwise returns NULL. */ |
| |
| static affine_fn |
| common_affine_function (conflict_function *cf) |
| { |
| unsigned i; |
| affine_fn comm; |
| |
| if (!CF_NONTRIVIAL_P (cf)) |
| return affine_fn (); |
| |
| comm = cf->fns[0]; |
| |
| for (i = 1; i < cf->n; i++) |
| if (!affine_function_equal_p (comm, cf->fns[i])) |
| return affine_fn (); |
| |
| return comm; |
| } |
| |
| /* Returns the base of the affine function FN. */ |
| |
| static tree |
| affine_function_base (affine_fn fn) |
| { |
| return fn[0]; |
| } |
| |
| /* Returns true if FN is a constant. */ |
| |
| static bool |
| affine_function_constant_p (affine_fn fn) |
| { |
| unsigned i; |
| tree coef; |
| |
| for (i = 1; fn.iterate (i, &coef); i++) |
| if (!integer_zerop (coef)) |
| return false; |
| |
| return true; |
| } |
| |
| /* Returns true if FN is the zero constant function. */ |
| |
| static bool |
| affine_function_zero_p (affine_fn fn) |
| { |
| return (integer_zerop (affine_function_base (fn)) |
| && affine_function_constant_p (fn)); |
| } |
| |
| /* Returns a signed integer type with the largest precision from TA |
| and TB. */ |
| |
| static tree |
| signed_type_for_types (tree ta, tree tb) |
| { |
| if (TYPE_PRECISION (ta) > TYPE_PRECISION (tb)) |
| return signed_type_for (ta); |
| else |
| return signed_type_for (tb); |
| } |
| |
| /* Applies operation OP on affine functions FNA and FNB, and returns the |
| result. */ |
| |
| static affine_fn |
| affine_fn_op (enum tree_code op, affine_fn fna, affine_fn fnb) |
| { |
| unsigned i, n, m; |
| affine_fn ret; |
| tree coef; |
| |
| if (fnb.length () > fna.length ()) |
| { |
| n = fna.length (); |
| m = fnb.length (); |
| } |
| else |
| { |
| n = fnb.length (); |
| m = fna.length (); |
| } |
| |
| ret.create (m); |
| for (i = 0; i < n; i++) |
| { |
| tree type = signed_type_for_types (TREE_TYPE (fna[i]), |
| TREE_TYPE (fnb[i])); |
| ret.quick_push (fold_build2 (op, type, fna[i], fnb[i])); |
| } |
| |
| for (; fna.iterate (i, &coef); i++) |
| ret.quick_push (fold_build2 (op, signed_type_for (TREE_TYPE (coef)), |
| coef, integer_zero_node)); |
| for (; fnb.iterate (i, &coef); i++) |
| ret.quick_push (fold_build2 (op, signed_type_for (TREE_TYPE (coef)), |
| integer_zero_node, coef)); |
| |
| return ret; |
| } |
| |
| /* Returns the sum of affine functions FNA and FNB. */ |
| |
| static affine_fn |
| affine_fn_plus (affine_fn fna, affine_fn fnb) |
| { |
| return affine_fn_op (PLUS_EXPR, fna, fnb); |
| } |
| |
| /* Returns the difference of affine functions FNA and FNB. */ |
| |
| static affine_fn |
| affine_fn_minus (affine_fn fna, affine_fn fnb) |
| { |
| return affine_fn_op (MINUS_EXPR, fna, fnb); |
| } |
| |
| /* Frees affine function FN. */ |
| |
| static void |
| affine_fn_free (affine_fn fn) |
| { |
| fn.release (); |
| } |
| |
| /* Determine for each subscript in the data dependence relation DDR |
| the distance. */ |
| |
| static void |
| compute_subscript_distance (struct data_dependence_relation *ddr) |
| { |
| conflict_function *cf_a, *cf_b; |
| affine_fn fn_a, fn_b, diff; |
| |
| if (DDR_ARE_DEPENDENT (ddr) == NULL_TREE) |
| { |
| unsigned int i; |
| |
| for (i = 0; i < DDR_NUM_SUBSCRIPTS (ddr); i++) |
| { |
| struct subscript *subscript; |
| |
| subscript = DDR_SUBSCRIPT (ddr, i); |
| cf_a = SUB_CONFLICTS_IN_A (subscript); |
| cf_b = SUB_CONFLICTS_IN_B (subscript); |
| |
| fn_a = common_affine_function (cf_a); |
| fn_b = common_affine_function (cf_b); |
| if (!fn_a.exists () || !fn_b.exists ()) |
| { |
| SUB_DISTANCE (subscript) = chrec_dont_know; |
| return; |
| } |
| diff = affine_fn_minus (fn_a, fn_b); |
| |
| if (affine_function_constant_p (diff)) |
| SUB_DISTANCE (subscript) = affine_function_base (diff); |
| else |
| SUB_DISTANCE (subscript) = chrec_dont_know; |
| |
| affine_fn_free (diff); |
| } |
| } |
| } |
| |
| /* Returns the conflict function for "unknown". */ |
| |
| static conflict_function * |
| conflict_fn_not_known (void) |
| { |
| conflict_function *fn = XCNEW (conflict_function); |
| fn->n = NOT_KNOWN; |
| |
| return fn; |
| } |
| |
| /* Returns the conflict function for "independent". */ |
| |
| static conflict_function * |
| conflict_fn_no_dependence (void) |
| { |
| conflict_function *fn = XCNEW (conflict_function); |
| fn->n = NO_DEPENDENCE; |
| |
| return fn; |
| } |
| |
| /* Returns true if the address of OBJ is invariant in LOOP. */ |
| |
| static bool |
| object_address_invariant_in_loop_p (const struct loop *loop, const_tree obj) |
| { |
| while (handled_component_p (obj)) |
| { |
| if (TREE_CODE (obj) == ARRAY_REF) |
| { |
| for (int i = 1; i < 4; ++i) |
| if (chrec_contains_symbols_defined_in_loop (TREE_OPERAND (obj, i), |
| loop->num)) |
| return false; |
| } |
| else if (TREE_CODE (obj) == COMPONENT_REF) |
| { |
| if (chrec_contains_symbols_defined_in_loop (TREE_OPERAND (obj, 2), |
| loop->num)) |
| return false; |
| } |
| obj = TREE_OPERAND (obj, 0); |
| } |
| |
| if (!INDIRECT_REF_P (obj) |
| && TREE_CODE (obj) != MEM_REF) |
| return true; |
| |
| return !chrec_contains_symbols_defined_in_loop (TREE_OPERAND (obj, 0), |
| loop->num); |
| } |
| |
| /* Returns false if we can prove that data references A and B do not alias, |
| true otherwise. If LOOP_NEST is false no cross-iteration aliases are |
| considered. */ |
| |
| bool |
| dr_may_alias_p (const struct data_reference *a, const struct data_reference *b, |
| struct loop *loop_nest) |
| { |
| tree addr_a = DR_BASE_OBJECT (a); |
| tree addr_b = DR_BASE_OBJECT (b); |
| |
| /* If we are not processing a loop nest but scalar code we |
| do not need to care about possible cross-iteration dependences |
| and thus can process the full original reference. Do so, |
| similar to how loop invariant motion applies extra offset-based |
| disambiguation. */ |
| if (!loop_nest) |
| { |
| aff_tree off1, off2; |
| poly_widest_int size1, size2; |
| get_inner_reference_aff (DR_REF (a), &off1, &size1); |
| get_inner_reference_aff (DR_REF (b), &off2, &size2); |
| aff_combination_scale (&off1, -1); |
| aff_combination_add (&off2, &off1); |
| if (aff_comb_cannot_overlap_p (&off2, size1, size2)) |
| return false; |
| } |
| |
| if ((TREE_CODE (addr_a) == MEM_REF || TREE_CODE (addr_a) == TARGET_MEM_REF) |
| && (TREE_CODE (addr_b) == MEM_REF || TREE_CODE (addr_b) == TARGET_MEM_REF) |
| /* For cross-iteration dependences the cliques must be valid for the |
| whole loop, not just individual iterations. */ |
| && (!loop_nest |
| || MR_DEPENDENCE_CLIQUE (addr_a) == 1 |
| || MR_DEPENDENCE_CLIQUE (addr_a) == loop_nest->owned_clique) |
| && MR_DEPENDENCE_CLIQUE (addr_a) == MR_DEPENDENCE_CLIQUE (addr_b) |
| && MR_DEPENDENCE_BASE (addr_a) != MR_DEPENDENCE_BASE (addr_b)) |
| return false; |
| |
| /* If we had an evolution in a pointer-based MEM_REF BASE_OBJECT we |
| do not know the size of the base-object. So we cannot do any |
| offset/overlap based analysis but have to rely on points-to |
| information only. */ |
| if (TREE_CODE (addr_a) == MEM_REF |
| && (DR_UNCONSTRAINED_BASE (a) |
| || TREE_CODE (TREE_OPERAND (addr_a, 0)) == SSA_NAME)) |
| { |
| /* For true dependences we can apply TBAA. */ |
| if (flag_strict_aliasing |
| && DR_IS_WRITE (a) && DR_IS_READ (b) |
| && !alias_sets_conflict_p (get_alias_set (DR_REF (a)), |
| get_alias_set (DR_REF (b)))) |
| return false; |
| if (TREE_CODE (addr_b) == MEM_REF) |
| return ptr_derefs_may_alias_p (TREE_OPERAND (addr_a, 0), |
| TREE_OPERAND (addr_b, 0)); |
| else |
| return ptr_derefs_may_alias_p (TREE_OPERAND (addr_a, 0), |
| build_fold_addr_expr (addr_b)); |
| } |
| else if (TREE_CODE (addr_b) == MEM_REF |
| && (DR_UNCONSTRAINED_BASE (b) |
| || TREE_CODE (TREE_OPERAND (addr_b, 0)) == SSA_NAME)) |
| { |
| /* For true dependences we can apply TBAA. */ |
| if (flag_strict_aliasing |
| && DR_IS_WRITE (a) && DR_IS_READ (b) |
| && !alias_sets_conflict_p (get_alias_set (DR_REF (a)), |
| get_alias_set (DR_REF (b)))) |
| return false; |
| if (TREE_CODE (addr_a) == MEM_REF) |
| return ptr_derefs_may_alias_p (TREE_OPERAND (addr_a, 0), |
| TREE_OPERAND (addr_b, 0)); |
| else |
| return ptr_derefs_may_alias_p (build_fold_addr_expr (addr_a), |
| TREE_OPERAND (addr_b, 0)); |
| } |
| |
| /* Otherwise DR_BASE_OBJECT is an access that covers the whole object |
| that is being subsetted in the loop nest. */ |
| if (DR_IS_WRITE (a) && DR_IS_WRITE (b)) |
| return refs_output_dependent_p (addr_a, addr_b); |
| else if (DR_IS_READ (a) && DR_IS_WRITE (b)) |
| return refs_anti_dependent_p (addr_a, addr_b); |
| return refs_may_alias_p (addr_a, addr_b); |
| } |
| |
| /* REF_A and REF_B both satisfy access_fn_component_p. Return true |
| if it is meaningful to compare their associated access functions |
| when checking for dependencies. */ |
| |
| static bool |
| access_fn_components_comparable_p (tree ref_a, tree ref_b) |
| { |
| /* Allow pairs of component refs from the following sets: |
| |
| { REALPART_EXPR, IMAGPART_EXPR } |
| { COMPONENT_REF } |
| { ARRAY_REF }. */ |
| tree_code code_a = TREE_CODE (ref_a); |
| tree_code code_b = TREE_CODE (ref_b); |
| if (code_a == IMAGPART_EXPR) |
| code_a = REALPART_EXPR; |
| if (code_b == IMAGPART_EXPR) |
| code_b = REALPART_EXPR; |
| if (code_a != code_b) |
| return false; |
| |
| if (TREE_CODE (ref_a) == COMPONENT_REF) |
| /* ??? We cannot simply use the type of operand #0 of the refs here as |
| the Fortran compiler smuggles type punning into COMPONENT_REFs. |
| Use the DECL_CONTEXT of the FIELD_DECLs instead. */ |
| return (DECL_CONTEXT (TREE_OPERAND (ref_a, 1)) |
| == DECL_CONTEXT (TREE_OPERAND (ref_b, 1))); |
| |
| return types_compatible_p (TREE_TYPE (TREE_OPERAND (ref_a, 0)), |
| TREE_TYPE (TREE_OPERAND (ref_b, 0))); |
| } |
| |
| /* Initialize a data dependence relation between data accesses A and |
| B. NB_LOOPS is the number of loops surrounding the references: the |
| size of the classic distance/direction vectors. */ |
| |
| struct data_dependence_relation * |
| initialize_data_dependence_relation (struct data_reference *a, |
| struct data_reference *b, |
| vec<loop_p> loop_nest) |
| { |
| struct data_dependence_relation *res; |
| unsigned int i; |
| |
| res = XCNEW (struct data_dependence_relation); |
| DDR_A (res) = a; |
| DDR_B (res) = b; |
| DDR_LOOP_NEST (res).create (0); |
| DDR_SUBSCRIPTS (res).create (0); |
| DDR_DIR_VECTS (res).create (0); |
| DDR_DIST_VECTS (res).create (0); |
| |
| if (a == NULL || b == NULL) |
| { |
| DDR_ARE_DEPENDENT (res) = chrec_dont_know; |
| return res; |
| } |
| |
| /* If the data references do not alias, then they are independent. */ |
| if (!dr_may_alias_p (a, b, loop_nest.exists () ? loop_nest[0] : NULL)) |
| { |
| DDR_ARE_DEPENDENT (res) = chrec_known; |
| return res; |
| } |
| |
| unsigned int num_dimensions_a = DR_NUM_DIMENSIONS (a); |
| unsigned int num_dimensions_b = DR_NUM_DIMENSIONS (b); |
| if (num_dimensions_a == 0 || num_dimensions_b == 0) |
| { |
| DDR_ARE_DEPENDENT (res) = chrec_dont_know; |
| return res; |
| } |
| |
| /* For unconstrained bases, the root (highest-indexed) subscript |
| describes a variation in the base of the original DR_REF rather |
| than a component access. We have no type that accurately describes |
| the new DR_BASE_OBJECT (whose TREE_TYPE describes the type *after* |
| applying this subscript) so limit the search to the last real |
| component access. |
| |
| E.g. for: |
| |
| void |
| f (int a[][8], int b[][8]) |
| { |
| for (int i = 0; i < 8; ++i) |
| a[i * 2][0] = b[i][0]; |
| } |
| |
| the a and b accesses have a single ARRAY_REF component reference [0] |
| but have two subscripts. */ |
| if (DR_UNCONSTRAINED_BASE (a)) |
| num_dimensions_a -= 1; |
| if (DR_UNCONSTRAINED_BASE (b)) |
| num_dimensions_b -= 1; |
| |
| /* These structures describe sequences of component references in |
| DR_REF (A) and DR_REF (B). Each component reference is tied to a |
| specific access function. */ |
| struct { |
| /* The sequence starts at DR_ACCESS_FN (A, START_A) of A and |
| DR_ACCESS_FN (B, START_B) of B (inclusive) and extends to higher |
| indices. In C notation, these are the indices of the rightmost |
| component references; e.g. for a sequence .b.c.d, the start |
| index is for .d. */ |
| unsigned int start_a; |
| unsigned int start_b; |
| |
| /* The sequence contains LENGTH consecutive access functions from |
| each DR. */ |
| unsigned int length; |
| |
| /* The enclosing objects for the A and B sequences respectively, |
| i.e. the objects to which DR_ACCESS_FN (A, START_A + LENGTH - 1) |
| and DR_ACCESS_FN (B, START_B + LENGTH - 1) are applied. */ |
| tree object_a; |
| tree object_b; |
| } full_seq = {}, struct_seq = {}; |
| |
| /* Before each iteration of the loop: |
| |
| - REF_A is what you get after applying DR_ACCESS_FN (A, INDEX_A) and |
| - REF_B is what you get after applying DR_ACCESS_FN (B, INDEX_B). */ |
| unsigned int index_a = 0; |
| unsigned int index_b = 0; |
| tree ref_a = DR_REF (a); |
| tree ref_b = DR_REF (b); |
| |
| /* Now walk the component references from the final DR_REFs back up to |
| the enclosing base objects. Each component reference corresponds |
| to one access function in the DR, with access function 0 being for |
| the final DR_REF and the highest-indexed access function being the |
| one that is applied to the base of the DR. |
| |
| Look for a sequence of component references whose access functions |
| are comparable (see access_fn_components_comparable_p). If more |
| than one such sequence exists, pick the one nearest the base |
| (which is the leftmost sequence in C notation). Store this sequence |
| in FULL_SEQ. |
| |
| For example, if we have: |
| |
| struct foo { struct bar s; ... } (*a)[10], (*b)[10]; |
| |
| A: a[0][i].s.c.d |
| B: __real b[0][i].s.e[i].f |
| |
| (where d is the same type as the real component of f) then the access |
| functions would be: |
| |
| 0 1 2 3 |
| A: .d .c .s [i] |
| |
| 0 1 2 3 4 5 |
| B: __real .f [i] .e .s [i] |
| |
| The A0/B2 column isn't comparable, since .d is a COMPONENT_REF |
| and [i] is an ARRAY_REF. However, the A1/B3 column contains two |
| COMPONENT_REF accesses for struct bar, so is comparable. Likewise |
| the A2/B4 column contains two COMPONENT_REF accesses for struct foo, |
| so is comparable. The A3/B5 column contains two ARRAY_REFs that |
| index foo[10] arrays, so is again comparable. The sequence is |
| therefore: |
| |
| A: [1, 3] (i.e. [i].s.c) |
| B: [3, 5] (i.e. [i].s.e) |
| |
| Also look for sequences of component references whose access |
| functions are comparable and whose enclosing objects have the same |
| RECORD_TYPE. Store this sequence in STRUCT_SEQ. In the above |
| example, STRUCT_SEQ would be: |
| |
| A: [1, 2] (i.e. s.c) |
| B: [3, 4] (i.e. s.e) */ |
| while (index_a < num_dimensions_a && index_b < num_dimensions_b) |
| { |
| /* REF_A and REF_B must be one of the component access types |
| allowed by dr_analyze_indices. */ |
| gcc_checking_assert (access_fn_component_p (ref_a)); |
| gcc_checking_assert (access_fn_component_p (ref_b)); |
| |
| /* Get the immediately-enclosing objects for REF_A and REF_B, |
| i.e. the references *before* applying DR_ACCESS_FN (A, INDEX_A) |
| and DR_ACCESS_FN (B, INDEX_B). */ |
| tree object_a = TREE_OPERAND (ref_a, 0); |
| tree object_b = TREE_OPERAND (ref_b, 0); |
| |
| tree type_a = TREE_TYPE (object_a); |
| tree type_b = TREE_TYPE (object_b); |
| if (access_fn_components_comparable_p (ref_a, ref_b)) |
| { |
| /* This pair of component accesses is comparable for dependence |
| analysis, so we can include DR_ACCESS_FN (A, INDEX_A) and |
| DR_ACCESS_FN (B, INDEX_B) in the sequence. */ |
| if (full_seq.start_a + full_seq.length != index_a |
| || full_seq.start_b + full_seq.length != index_b) |
| { |
| /* The accesses don't extend the current sequence, |
| so start a new one here. */ |
| full_seq.start_a = index_a; |
| full_seq.start_b = index_b; |
| full_seq.length = 0; |
| } |
| |
| /* Add this pair of references to the sequence. */ |
| full_seq.length += 1; |
| full_seq.object_a = object_a; |
| full_seq.object_b = object_b; |
| |
| /* If the enclosing objects are structures (and thus have the |
| same RECORD_TYPE), record the new sequence in STRUCT_SEQ. */ |
| if (TREE_CODE (type_a) == RECORD_TYPE) |
| struct_seq = full_seq; |
| |
| /* Move to the next containing reference for both A and B. */ |
| ref_a = object_a; |
| ref_b = object_b; |
| index_a += 1; |
| index_b += 1; |
| continue; |
| } |
| |
| /* Try to approach equal type sizes. */ |
| if (!COMPLETE_TYPE_P (type_a) |
| || !COMPLETE_TYPE_P (type_b) |
| || !tree_fits_uhwi_p (TYPE_SIZE_UNIT (type_a)) |
| || !tree_fits_uhwi_p (TYPE_SIZE_UNIT (type_b))) |
| break; |
| |
| unsigned HOST_WIDE_INT size_a = tree_to_uhwi (TYPE_SIZE_UNIT (type_a)); |
| unsigned HOST_WIDE_INT size_b = tree_to_uhwi (TYPE_SIZE_UNIT (type_b)); |
| if (size_a <= size_b) |
| { |
| index_a += 1; |
| ref_a = object_a; |
| } |
| if (size_b <= size_a) |
| { |
| index_b += 1; |
| ref_b = object_b; |
| } |
| } |
| |
| /* See whether FULL_SEQ ends at the base and whether the two bases |
| are equal. We do not care about TBAA or alignment info so we can |
| use OEP_ADDRESS_OF to avoid false negatives. */ |
| tree base_a = DR_BASE_OBJECT (a); |
| tree base_b = DR_BASE_OBJECT (b); |
| bool same_base_p = (full_seq.start_a + full_seq.length == num_dimensions_a |
| && full_seq.start_b + full_seq.length == num_dimensions_b |
| && DR_UNCONSTRAINED_BASE (a) == DR_UNCONSTRAINED_BASE (b) |
| && operand_equal_p (base_a, base_b, OEP_ADDRESS_OF) |
| && types_compatible_p (TREE_TYPE (base_a), |
| TREE_TYPE (base_b)) |
| && (!loop_nest.exists () |
| || (object_address_invariant_in_loop_p |
| (loop_nest[0], base_a)))); |
| |
| /* If the bases are the same, we can include the base variation too. |
| E.g. the b accesses in: |
| |
| for (int i = 0; i < n; ++i) |
| b[i + 4][0] = b[i][0]; |
| |
| have a definite dependence distance of 4, while for: |
| |
| for (int i = 0; i < n; ++i) |
| a[i + 4][0] = b[i][0]; |
| |
| the dependence distance depends on the gap between a and b. |
| |
| If the bases are different then we can only rely on the sequence |
| rooted at a structure access, since arrays are allowed to overlap |
| arbitrarily and change shape arbitrarily. E.g. we treat this as |
| valid code: |
| |
| int a[256]; |
| ... |
| ((int (*)[4][3]) &a[1])[i][0] += ((int (*)[4][3]) &a[2])[i][0]; |
| |
| where two lvalues with the same int[4][3] type overlap, and where |
| both lvalues are distinct from the object's declared type. */ |
| if (same_base_p) |
| { |
| if (DR_UNCONSTRAINED_BASE (a)) |
| full_seq.length += 1; |
| } |
| else |
| full_seq = struct_seq; |
| |
| /* Punt if we didn't find a suitable sequence. */ |
| if (full_seq.length == 0) |
| { |
| DDR_ARE_DEPENDENT (res) = chrec_dont_know; |
| return res; |
| } |
| |
| if (!same_base_p) |
| { |
| /* Partial overlap is possible for different bases when strict aliasing |
| is not in effect. It's also possible if either base involves a union |
| access; e.g. for: |
| |
| struct s1 { int a[2]; }; |
| struct s2 { struct s1 b; int c; }; |
| struct s3 { int d; struct s1 e; }; |
| union u { struct s2 f; struct s3 g; } *p, *q; |
| |
| the s1 at "p->f.b" (base "p->f") partially overlaps the s1 at |
| "p->g.e" (base "p->g") and might partially overlap the s1 at |
| "q->g.e" (base "q->g"). */ |
| if (!flag_strict_aliasing |
| || ref_contains_union_access_p (full_seq.object_a) |
| || ref_contains_union_access_p (full_seq.object_b)) |
| { |
| DDR_ARE_DEPENDENT (res) = chrec_dont_know; |
| return res; |
| } |
| |
| DDR_COULD_BE_INDEPENDENT_P (res) = true; |
| if (!loop_nest.exists () |
| || (object_address_invariant_in_loop_p (loop_nest[0], |
| full_seq.object_a) |
| && object_address_invariant_in_loop_p (loop_nest[0], |
| full_seq.object_b))) |
| { |
| DDR_OBJECT_A (res) = full_seq.object_a; |
| DDR_OBJECT_B (res) = full_seq.object_b; |
| } |
| } |
| |
| DDR_AFFINE_P (res) = true; |
| DDR_ARE_DEPENDENT (res) = NULL_TREE; |
| DDR_SUBSCRIPTS (res).create (full_seq.length); |
| DDR_LOOP_NEST (res) = loop_nest; |
| DDR_INNER_LOOP (res) = 0; |
| DDR_SELF_REFERENCE (res) = false; |
| |
| for (i = 0; i < full_seq.length; ++i) |
| { |
| struct subscript *subscript; |
| |
| subscript = XNEW (struct subscript); |
| SUB_ACCESS_FN (subscript, 0) = DR_ACCESS_FN (a, full_seq.start_a + i); |
| SUB_ACCESS_FN (subscript, 1) = DR_ACCESS_FN (b, full_seq.start_b + i); |
| SUB_CONFLICTS_IN_A (subscript) = conflict_fn_not_known (); |
| SUB_CONFLICTS_IN_B (subscript) = conflict_fn_not_known (); |
| SUB_LAST_CONFLICT (subscript) = chrec_dont_know; |
| SUB_DISTANCE (subscript) = chrec_dont_know; |
| DDR_SUBSCRIPTS (res).safe_push (subscript); |
| } |
| |
| return res; |
| } |
| |
| /* Frees memory used by the conflict function F. */ |
| |
| static void |
| free_conflict_function (conflict_function *f) |
| { |
| unsigned i; |
| |
| if (CF_NONTRIVIAL_P (f)) |
| { |
| for (i = 0; i < f->n; i++) |
| affine_fn_free (f->fns[i]); |
| } |
| free (f); |
| } |
| |
| /* Frees memory used by SUBSCRIPTS. */ |
| |
| static void |
| free_subscripts (vec<subscript_p> subscripts) |
| { |
| unsigned i; |
| subscript_p s; |
| |
| FOR_EACH_VEC_ELT (subscripts, i, s) |
| { |
| free_conflict_function (s->conflicting_iterations_in_a); |
| free_conflict_function (s->conflicting_iterations_in_b); |
| free (s); |
| } |
| subscripts.release (); |
| } |
| |
| /* Set DDR_ARE_DEPENDENT to CHREC and finalize the subscript overlap |
| description. */ |
| |
| static inline void |
| finalize_ddr_dependent (struct data_dependence_relation *ddr, |
| tree chrec) |
| { |
| DDR_ARE_DEPENDENT (ddr) = chrec; |
| free_subscripts (DDR_SUBSCRIPTS (ddr)); |
| DDR_SUBSCRIPTS (ddr).create (0); |
| } |
| |
| /* The dependence relation DDR cannot be represented by a distance |
| vector. */ |
| |
| static inline void |
| non_affine_dependence_relation (struct data_dependence_relation *ddr) |
| { |
| if (dump_file && (dump_flags & TDF_DETAILS)) |
| fprintf (dump_file, "(Dependence relation cannot be represented by distance vector.) \n"); |
| |
| DDR_AFFINE_P (ddr) = false; |
| } |
| |
| |
| |
| /* This section contains the classic Banerjee tests. */ |
| |
| /* Returns true iff CHREC_A and CHREC_B are not dependent on any index |
| variables, i.e., if the ZIV (Zero Index Variable) test is true. */ |
| |
| static inline bool |
| ziv_subscript_p (const_tree chrec_a, const_tree chrec_b) |
| { |
| return (evolution_function_is_constant_p (chrec_a) |
| && evolution_function_is_constant_p (chrec_b)); |
| } |
| |
| /* Returns true iff CHREC_A and CHREC_B are dependent on an index |
| variable, i.e., if the SIV (Single Index Variable) test is true. */ |
| |
| static bool |
| siv_subscript_p (const_tree chrec_a, const_tree chrec_b) |
| { |
| if ((evolution_function_is_constant_p (chrec_a) |
| && evolution_function_is_univariate_p (chrec_b)) |
| || (evolution_function_is_constant_p (chrec_b) |
| && evolution_function_is_univariate_p (chrec_a))) |
| return true; |
| |
| if (evolution_function_is_univariate_p (chrec_a) |
| && evolution_function_is_univariate_p (chrec_b)) |
| { |
| switch (TREE_CODE (chrec_a)) |
| { |
| case POLYNOMIAL_CHREC: |
| switch (TREE_CODE (chrec_b)) |
| { |
| case POLYNOMIAL_CHREC: |
| if (CHREC_VARIABLE (chrec_a) != CHREC_VARIABLE (chrec_b)) |
| return false; |
| /* FALLTHRU */ |
| |
| default: |
| return true; |
| } |
| |
| default: |
| return true; |
| } |
| } |
| |
| return false; |
| } |
| |
| /* Creates a conflict function with N dimensions. The affine functions |
| in each dimension follow. */ |
| |
| static conflict_function * |
| conflict_fn (unsigned n, ...) |
| { |
| unsigned i; |
| conflict_function *ret = XCNEW (conflict_function); |
| va_list ap; |
| |
| gcc_assert (n > 0 && n <= MAX_DIM); |
| va_start (ap, n); |
| |
| ret->n = n; |
| for (i = 0; i < n; i++) |
| ret->fns[i] = va_arg (ap, affine_fn); |
| va_end (ap); |
| |
| return ret; |
| } |
| |
| /* Returns constant affine function with value CST. */ |
| |
| static affine_fn |
| affine_fn_cst (tree cst) |
| { |
| affine_fn fn; |
| fn.create (1); |
| fn.quick_push (cst); |
| return fn; |
| } |
| |
| /* Returns affine function with single variable, CST + COEF * x_DIM. */ |
| |
| static affine_fn |
| affine_fn_univar (tree cst, unsigned dim, tree coef) |
| { |
| affine_fn fn; |
| fn.create (dim + 1); |
| unsigned i; |
| |
| gcc_assert (dim > 0); |
| fn.quick_push (cst); |
| for (i = 1; i < dim; i++) |
| fn.quick_push (integer_zero_node); |
| fn.quick_push (coef); |
| return fn; |
| } |
| |
| /* Analyze a ZIV (Zero Index Variable) subscript. *OVERLAPS_A and |
| *OVERLAPS_B are initialized to the functions that describe the |
| relation between the elements accessed twice by CHREC_A and |
| CHREC_B. For k >= 0, the following property is verified: |
| |
| CHREC_A (*OVERLAPS_A (k)) = CHREC_B (*OVERLAPS_B (k)). */ |
| |
| static void |
| analyze_ziv_subscript (tree chrec_a, |
| tree chrec_b, |
| conflict_function **overlaps_a, |
| conflict_function **overlaps_b, |
| tree *last_conflicts) |
| { |
| tree type, difference; |
| dependence_stats.num_ziv++; |
| |
| if (dump_file && (dump_flags & TDF_DETAILS)) |
| fprintf (dump_file, "(analyze_ziv_subscript \n"); |
| |
| type = signed_type_for_types (TREE_TYPE (chrec_a), TREE_TYPE (chrec_b)); |
| chrec_a = chrec_convert (type, chrec_a, NULL); |
| chrec_b = chrec_convert (type, chrec_b, NULL); |
| difference = chrec_fold_minus (type, chrec_a, chrec_b); |
| |
| switch (TREE_CODE (difference)) |
| { |
| case INTEGER_CST: |
| if (integer_zerop (difference)) |
| { |
| /* The difference is equal to zero: the accessed index |
| overlaps for each iteration in the loop. */ |
| *overlaps_a = conflict_fn (1, affine_fn_cst (integer_zero_node)); |
| *overlaps_b = conflict_fn (1, affine_fn_cst (integer_zero_node)); |
| *last_conflicts = chrec_dont_know; |
| dependence_stats.num_ziv_dependent++; |
| } |
| else |
| { |
| /* The accesses do not overlap. */ |
| *overlaps_a = conflict_fn_no_dependence (); |
| *overlaps_b = conflict_fn_no_dependence (); |
| *last_conflicts = integer_zero_node; |
| dependence_stats.num_ziv_independent++; |
| } |
| break; |
| |
| default: |
| /* We're not sure whether the indexes overlap. For the moment, |
| conservatively answer "don't know". */ |
| if (dump_file && (dump_flags & TDF_DETAILS)) |
| fprintf (dump_file, "ziv test failed: difference is non-integer.\n"); |
| |
| *overlaps_a = conflict_fn_not_known (); |
| *overlaps_b = conflict_fn_not_known (); |
| *last_conflicts = chrec_dont_know; |
| dependence_stats.num_ziv_unimplemented++; |
| break; |
| } |
| |
| if (dump_file && (dump_flags & TDF_DETAILS)) |
| fprintf (dump_file, ")\n"); |
| } |
| |
| /* Similar to max_stmt_executions_int, but returns the bound as a tree, |
| and only if it fits to the int type. If this is not the case, or the |
| bound on the number of iterations of LOOP could not be derived, returns |
| chrec_dont_know. */ |
| |
| static tree |
| max_stmt_executions_tree (struct loop *loop) |
| { |
| widest_int nit; |
| |
| if (!max_stmt_executions (loop, &nit)) |
| return chrec_dont_know; |
| |
| if (!wi::fits_to_tree_p (nit, unsigned_type_node)) |
| return chrec_dont_know; |
| |
| return wide_int_to_tree (unsigned_type_node, nit); |
| } |
| |
| /* Determine whether the CHREC is always positive/negative. If the expression |
| cannot be statically analyzed, return false, otherwise set the answer into |
| VALUE. */ |
| |
| static bool |
| chrec_is_positive (tree chrec, bool *value) |
| { |
| bool value0, value1, value2; |
| tree end_value, nb_iter; |
| |
| switch (TREE_CODE (chrec)) |
| { |
| case POLYNOMIAL_CHREC: |
| if (!chrec_is_positive (CHREC_LEFT (chrec), &value0) |
| || !chrec_is_positive (CHREC_RIGHT (chrec), &value1)) |
| return false; |
| |
| /* FIXME -- overflows. */ |
| if (value0 == value1) |
| { |
| *value = value0; |
| return true; |
| } |
| |
| /* Otherwise the chrec is under the form: "{-197, +, 2}_1", |
| and the proof consists in showing that the sign never |
| changes during the execution of the loop, from 0 to |
| loop->nb_iterations. */ |
| if (!evolution_function_is_affine_p (chrec)) |
| return false; |
| |
| nb_iter = number_of_latch_executions (get_chrec_loop (chrec)); |
| if (chrec_contains_undetermined (nb_iter)) |
| return false; |
| |
| #if 0 |
| /* TODO -- If the test is after the exit, we may decrease the number of |
| iterations by one. */ |
| if (after_exit) |
| nb_iter = chrec_fold_minus (type, nb_iter, build_int_cst (type, 1)); |
| #endif |
| |
| end_value = chrec_apply (CHREC_VARIABLE (chrec), chrec, nb_iter); |
| |
| if (!chrec_is_positive (end_value, &value2)) |
| return false; |
| |
| *value = value0; |
| return value0 == value1; |
| |
| case INTEGER_CST: |
| switch (tree_int_cst_sgn (chrec)) |
| { |
| case -1: |
| *value = false; |
| break; |
| case 1: |
| *value = true; |
| break; |
| default: |
| return false; |
| } |
| return true; |
| |
| default: |
| return false; |
| } |
| } |
| |
| |
| /* Analyze a SIV (Single Index Variable) subscript where CHREC_A is a |
| constant, and CHREC_B is an affine function. *OVERLAPS_A and |
| *OVERLAPS_B are initialized to the functions that describe the |
| relation between the elements accessed twice by CHREC_A and |
| CHREC_B. For k >= 0, the following property is verified: |
| |
| CHREC_A (*OVERLAPS_A (k)) = CHREC_B (*OVERLAPS_B (k)). */ |
| |
| static void |
| analyze_siv_subscript_cst_affine (tree chrec_a, |
| tree chrec_b, |
| conflict_function **overlaps_a, |
| conflict_function **overlaps_b, |
| tree *last_conflicts) |
| { |
| bool value0, value1, value2; |
| tree type, difference, tmp; |
| |
| type = signed_type_for_types (TREE_TYPE (chrec_a), TREE_TYPE (chrec_b)); |
| chrec_a = chrec_convert (type, chrec_a, NULL); |
| chrec_b = chrec_convert (type, chrec_b, NULL); |
| difference = chrec_fold_minus (type, initial_condition (chrec_b), chrec_a); |
| |
| /* Special case overlap in the first iteration. */ |
| if (integer_zerop (difference)) |
| { |
| *overlaps_a = conflict_fn (1, affine_fn_cst (integer_zero_node)); |
| *overlaps_b = conflict_fn (1, affine_fn_cst (integer_zero_node)); |
| *last_conflicts = integer_one_node; |
| return; |
| } |
| |
| if (!chrec_is_positive (initial_condition (difference), &value0)) |
| { |
| if (dump_file && (dump_flags & TDF_DETAILS)) |
| fprintf (dump_file, "siv test failed: chrec is not positive.\n"); |
| |
| dependence_stats.num_siv_unimplemented++; |
| *overlaps_a = conflict_fn_not_known (); |
| *overlaps_b = conflict_fn_not_known (); |
| *last_conflicts = chrec_dont_know; |
| return; |
| } |
| else |
| { |
| if (value0 == false) |
| { |
| if (TREE_CODE (chrec_b) != POLYNOMIAL_CHREC |
| || !chrec_is_positive (CHREC_RIGHT (chrec_b), &value1)) |
| { |
| if (dump_file && (dump_flags & TDF_DETAILS)) |
| fprintf (dump_file, "siv test failed: chrec not positive.\n"); |
| |
| *overlaps_a = conflict_fn_not_known (); |
| *overlaps_b = conflict_fn_not_known (); |
| *last_conflicts = chrec_dont_know; |
| dependence_stats.num_siv_unimplemented++; |
| return; |
| } |
| else |
| { |
| if (value1 == true) |
| { |
| /* Example: |
| chrec_a = 12 |
| chrec_b = {10, +, 1} |
| */ |
| |
| if (tree_fold_divides_p (CHREC_RIGHT (chrec_b), difference)) |
| { |
| HOST_WIDE_INT numiter; |
| struct loop *loop = get_chrec_loop (chrec_b); |
| |
| *overlaps_a = conflict_fn (1, affine_fn_cst (integer_zero_node)); |
| tmp = fold_build2 (EXACT_DIV_EXPR, type, |
| fold_build1 (ABS_EXPR, type, difference), |
| CHREC_RIGHT (chrec_b)); |
| *overlaps_b = conflict_fn (1, affine_fn_cst (tmp)); |
| *last_conflicts = integer_one_node; |
| |
| |
| /* Perform weak-zero siv test to see if overlap is |
| outside the loop bounds. */ |
| numiter = max_stmt_executions_int (loop); |
| |
| if (numiter >= 0 |
| && compare_tree_int (tmp, numiter) > 0) |
| { |
| free_conflict_function (*overlaps_a); |
| free_conflict_function (*overlaps_b); |
| *overlaps_a = conflict_fn_no_dependence (); |
| *overlaps_b = conflict_fn_no_dependence (); |
| *last_conflicts = integer_zero_node; |
| dependence_stats.num_siv_independent++; |
| return; |
| } |
| dependence_stats.num_siv_dependent++; |
| return; |
| } |
| |
| /* When the step does not divide the difference, there are |
| no overlaps. */ |
| else |
| { |
| *overlaps_a = conflict_fn_no_dependence (); |
| *overlaps_b = conflict_fn_no_dependence (); |
| *last_conflicts = integer_zero_node; |
| dependence_stats.num_siv_independent++; |
| return; |
| } |
| } |
| |
| else |
| { |
| /* Example: |
| chrec_a = 12 |
| chrec_b = {10, +, -1} |
| |
| In this case, chrec_a will not overlap with chrec_b. */ |
| *overlaps_a = conflict_fn_no_dependence (); |
| *overlaps_b = conflict_fn_no_dependence (); |
| *last_conflicts = integer_zero_node; |
| dependence_stats.num_siv_independent++; |
| return; |
| } |
| } |
| } |
| else |
| { |
| if (TREE_CODE (chrec_b) != POLYNOMIAL_CHREC |
| || !chrec_is_positive (CHREC_RIGHT (chrec_b), &value2)) |
| { |
| if (dump_file && (dump_flags & TDF_DETAILS)) |
| fprintf (dump_file, "siv test failed: chrec not positive.\n"); |
| |
| *overlaps_a = conflict_fn_not_known (); |
| *overlaps_b = conflict_fn_not_known (); |
| *last_conflicts = chrec_dont_know; |
| dependence_stats.num_siv_unimplemented++; |
| return; |
| } |
| else |
| { |
| if (value2 == false) |
| { |
| /* Example: |
| chrec_a = 3 |
| chrec_b = {10, +, -1} |
| */ |
| if (tree_fold_divides_p (CHREC_RIGHT (chrec_b), difference)) |
| { |
| HOST_WIDE_INT numiter; |
| struct loop *loop = get_chrec_loop (chrec_b); |
| |
| *overlaps_a = conflict_fn (1, affine_fn_cst (integer_zero_node)); |
| tmp = fold_build2 (EXACT_DIV_EXPR, type, difference, |
| CHREC_RIGHT (chrec_b)); |
| *overlaps_b = conflict_fn (1, affine_fn_cst (tmp)); |
| *last_conflicts = integer_one_node; |
| |
| /* Perform weak-zero siv test to see if overlap is |
| outside the loop bounds. */ |
| numiter = max_stmt_executions_int (loop); |
| |
| if (numiter >= 0 |
| && compare_tree_int (tmp, numiter) > 0) |
| { |
| free_conflict_function (*overlaps_a); |
| free_conflict_function (*overlaps_b); |
| *overlaps_a = conflict_fn_no_dependence (); |
| *overlaps_b = conflict_fn_no_dependence (); |
| *last_conflicts = integer_zero_node; |
| dependence_stats.num_siv_independent++; |
| return; |
| } |
| dependence_stats.num_siv_dependent++; |
| return; |
| } |
| |
| /* When the step does not divide the difference, there |
| are no overlaps. */ |
| else |
| { |
| *overlaps_a = conflict_fn_no_dependence (); |
| *overlaps_b = conflict_fn_no_dependence (); |
| *last_conflicts = integer_zero_node; |
| dependence_stats.num_siv_independent++; |
| return; |
| } |
| } |
| else |
| { |
| /* Example: |
| chrec_a = 3 |
| chrec_b = {4, +, 1} |
| |
| In this case, chrec_a will not overlap with chrec_b. */ |
| *overlaps_a = conflict_fn_no_dependence (); |
| *overlaps_b = conflict_fn_no_dependence (); |
| *last_conflicts = integer_zero_node; |
| dependence_stats.num_siv_independent++; |
| return; |
| } |
| } |
| } |
| } |
| } |
| |
| /* Helper recursive function for initializing the matrix A. Returns |
| the initial value of CHREC. */ |
| |
| static tree |
| initialize_matrix_A (lambda_matrix A, tree chrec, unsigned index, int mult) |
| { |
| gcc_assert (chrec); |
| |
| switch (TREE_CODE (chrec)) |
| { |
| case POLYNOMIAL_CHREC: |
| if (!cst_and_fits_in_hwi ( |