/* Data references and dependences detectors.
- Copyright (C) 2003, 2004, 2005, 2006, 2007 Free Software Foundation, Inc.
+ Copyright (C) 2003, 2004, 2005, 2006, 2007, 2008, 2009
+ Free Software Foundation, Inc.
Contributed by Sebastian Pop <pop@cri.ensmp.fr>
This file is part of GCC.
/* 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
+ 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
+
+ - 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
+ 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"
+ http://citeseer.ist.psu.edu/goff91practical.html
+
+ - "Loop Transformations for Restructuring Compilers - The Foundations"
by Utpal Banerjee.
-
+
*/
#include "config.h"
#include "coretypes.h"
#include "tm.h"
#include "ggc.h"
+#include "flags.h"
#include "tree.h"
/* These RTL headers are needed for basic-block.h. */
struct loop *);
/* Returns true iff A divides B. */
-static inline bool
+static inline bool
tree_fold_divides_p (const_tree a, const_tree b)
{
gcc_assert (TREE_CODE (a) == INTEGER_CST);
/* Returns true iff A divides B. */
-static inline bool
+static inline bool
int_divides_p (int a, int b)
{
return ((b % a) == 0);
\f
-/* Dump into FILE all the data references from DATAREFS. */
+/* Dump into FILE all the data references from DATAREFS. */
-void
+void
dump_data_references (FILE *file, VEC (data_reference_p, heap) *datarefs)
{
unsigned int i;
dump_data_reference (file, dr);
}
-/* Dump to STDERR all the dependence relations from DDRS. */
+/* Dump into STDERR all the data references from DATAREFS. */
+
+void
+debug_data_references (VEC (data_reference_p, heap) *datarefs)
+{
+ dump_data_references (stderr, datarefs);
+}
+
+/* Dump to STDERR all the dependence relations from DDRS. */
-void
+void
debug_data_dependence_relations (VEC (ddr_p, heap) *ddrs)
{
dump_data_dependence_relations (stderr, ddrs);
}
-/* Dump into FILE all the dependence relations from DDRS. */
+/* Dump into FILE all the dependence relations from DDRS. */
-void
-dump_data_dependence_relations (FILE *file,
+void
+dump_data_dependence_relations (FILE *file,
VEC (ddr_p, heap) *ddrs)
{
unsigned int i;
dump_data_dependence_relation (file, ddr);
}
+/* Print to STDERR the data_reference DR. */
+
+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,
+void
+dump_data_reference (FILE *outf,
struct data_reference *dr)
{
unsigned int i;
-
+
fprintf (outf, "(Data Ref: \n stmt: ");
- print_generic_stmt (outf, DR_STMT (dr), 0);
+ print_gimple_stmt (outf, DR_STMT (dr), 0, 0);
fprintf (outf, " ref: ");
print_generic_stmt (outf, DR_REF (dr), 0);
fprintf (outf, " base_object: ");
print_generic_stmt (outf, DR_BASE_OBJECT (dr), 0);
-
+
for (i = 0; i < DR_NUM_DIMENSIONS (dr); i++)
{
fprintf (outf, " Access function %d: ", i);
/* Dump function for a SUBSCRIPT structure. */
-void
+void
dump_subscript (FILE *outf, struct subscript *subscript)
{
conflict_function *cf = SUB_CONFLICTS_IN_A (subscript);
fprintf (outf, " last_conflict: ");
print_generic_stmt (outf, last_iteration, 0);
}
-
+
cf = SUB_CONFLICTS_IN_B (subscript);
fprintf (outf, " iterations_that_access_an_element_twice_in_B: ");
dump_conflict_function (outf, cf);
for (eq = 0; eq < length; eq++)
{
- enum data_dependence_direction dir = dirv[eq];
+ enum data_dependence_direction dir = ((enum data_dependence_direction)
+ dirv[eq]);
switch (dir)
{
/* Debug version. */
-void
+void
debug_data_dependence_relation (struct data_dependence_relation *ddr)
{
dump_data_dependence_relation (stderr, ddr);
/* Dump function for a DATA_DEPENDENCE_RELATION structure. */
-void
-dump_data_dependence_relation (FILE *outf,
+void
+dump_data_dependence_relation (FILE *outf,
struct data_dependence_relation *ddr)
{
struct data_reference *dra, *drb;
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;
}
if (DDR_ARE_DEPENDENT (ddr) == chrec_known)
fprintf (outf, " (no dependence)\n");
-
+
else if (DDR_ARE_DEPENDENT (ddr) == NULL_TREE)
{
unsigned int i;
/* Dump function for a DATA_DEPENDENCE_DIRECTION structure. */
void
-dump_data_dependence_direction (FILE *file,
+dump_data_dependence_direction (FILE *file,
enum data_dependence_direction dir)
{
switch (dir)
{
- case dir_positive:
+ case dir_positive:
fprintf (file, "+");
break;
-
+
case dir_negative:
fprintf (file, "-");
break;
-
+
case dir_equal:
fprintf (file, "=");
break;
-
+
case dir_positive_or_negative:
fprintf (file, "+-");
break;
-
- case dir_positive_or_equal:
+
+ case dir_positive_or_equal:
fprintf (file, "+=");
break;
-
- case dir_negative_or_equal:
+
+ case dir_negative_or_equal:
fprintf (file, "-=");
break;
-
- case dir_star:
- fprintf (file, "*");
+
+ case dir_star:
+ fprintf (file, "*");
break;
-
- default:
+
+ default:
break;
}
}
dependence vectors, or in other words the number of loops in the
considered nest. */
-void
+void
dump_dist_dir_vectors (FILE *file, VEC (ddr_p, heap) *ddrs)
{
unsigned int i, j;
/* Dumps the data dependence relations DDRS in FILE. */
-void
+void
dump_ddrs (FILE *file, VEC (ddr_p, heap) *ddrs)
{
unsigned int i;
fprintf (file, "\n\n");
}
-/* Expresses EXP as VAR + OFF, where off is a constant. The type of OFF
- will be ssizetype. */
+/* 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. */
-void
-split_constant_offset (tree exp, tree *var, tree *off)
+static bool
+split_constant_offset_1 (tree type, tree op0, enum tree_code code, tree op1,
+ tree *var, tree *off)
{
- tree type = TREE_TYPE (exp), otype;
tree var0, var1;
tree off0, off1;
- enum tree_code code;
+ enum tree_code ocode = code;
- *var = exp;
- STRIP_NOPS (exp);
- otype = TREE_TYPE (exp);
- code = TREE_CODE (exp);
+ *var = NULL_TREE;
+ *off = NULL_TREE;
switch (code)
{
case INTEGER_CST:
*var = build_int_cst (type, 0);
- *off = fold_convert (ssizetype, exp);
- return;
+ *off = fold_convert (ssizetype, op0);
+ return true;
case POINTER_PLUS_EXPR:
- code = PLUS_EXPR;
+ ocode = PLUS_EXPR;
/* FALLTHROUGH */
case PLUS_EXPR:
case MINUS_EXPR:
- split_constant_offset (TREE_OPERAND (exp, 0), &var0, &off0);
- split_constant_offset (TREE_OPERAND (exp, 1), &var1, &off1);
- *var = fold_convert (type, fold_build2 (TREE_CODE (exp), otype,
- var0, var1));
- *off = size_binop (code, off0, off1);
- return;
+ split_constant_offset (op0, &var0, &off0);
+ split_constant_offset (op1, &var1, &off1);
+ *var = fold_build2 (code, type, var0, var1);
+ *off = size_binop (ocode, off0, off1);
+ return true;
case MULT_EXPR:
- off1 = TREE_OPERAND (exp, 1);
- if (TREE_CODE (off1) != INTEGER_CST)
- break;
+ if (TREE_CODE (op1) != INTEGER_CST)
+ return false;
- split_constant_offset (TREE_OPERAND (exp, 0), &var0, &off0);
- *var = fold_convert (type, fold_build2 (MULT_EXPR, otype,
- var0, off1));
- *off = size_binop (MULT_EXPR, off0, fold_convert (ssizetype, off1));
- return;
+ split_constant_offset (op0, &var0, &off0);
+ *var = fold_build2 (MULT_EXPR, type, var0, op1);
+ *off = size_binop (MULT_EXPR, off0, fold_convert (ssizetype, op1));
+ return true;
case ADDR_EXPR:
{
- tree op, base, poffset;
+ tree base, poffset;
HOST_WIDE_INT pbitsize, pbitpos;
enum machine_mode pmode;
int punsignedp, pvolatilep;
- op = TREE_OPERAND (exp, 0);
- if (!handled_component_p (op))
- break;
+ op0 = TREE_OPERAND (op0, 0);
+ if (!handled_component_p (op0))
+ return false;
- base = get_inner_reference (op, &pbitsize, &pbitpos, &poffset,
+ base = get_inner_reference (op0, &pbitsize, &pbitpos, &poffset,
&pmode, &punsignedp, &pvolatilep, false);
if (pbitpos % BITS_PER_UNIT != 0)
- break;
+ return false;
base = build_fold_addr_expr (base);
off0 = ssize_int (pbitpos / BITS_PER_UNIT);
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 (TREE_CODE (var0) == NOP_EXPR
- || TREE_CODE (var0) == CONVERT_EXPR)
+ if (CONVERT_EXPR_P (var0))
{
gimplify_conversion (&var0);
// Attempt to fill in any within var0 found ARRAY_REF's
while (POINTER_TYPE_P (type))
type = TREE_TYPE (type);
if (int_size_in_bytes (type) < 0)
- break;
+ return false;
*var = var0;
*off = off0;
- return;
+ return true;
}
case SSA_NAME:
{
- tree def_stmt = SSA_NAME_DEF_STMT (exp);
- if (TREE_CODE (def_stmt) == GIMPLE_MODIFY_STMT)
- {
- tree def_stmt_rhs = GIMPLE_STMT_OPERAND (def_stmt, 1);
+ gimple def_stmt = SSA_NAME_DEF_STMT (op0);
+ enum tree_code subcode;
- if (!TREE_SIDE_EFFECTS (def_stmt_rhs)
- && EXPR_P (def_stmt_rhs)
- && !REFERENCE_CLASS_P (def_stmt_rhs)
- && !get_call_expr_in (def_stmt_rhs))
- {
- split_constant_offset (def_stmt_rhs, &var0, &off0);
- var0 = fold_convert (type, var0);
- *var = var0;
- *off = off0;
- return;
- }
+ if (gimple_code (def_stmt) != GIMPLE_ASSIGN)
+ return false;
+
+ var0 = gimple_assign_rhs1 (def_stmt);
+ subcode = gimple_assign_rhs_code (def_stmt);
+ var1 = gimple_assign_rhs2 (def_stmt);
+
+ return split_constant_offset_1 (type, var0, subcode, var1, var, off);
+ }
+ 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_UNDEFINED (itype)))
+ && TYPE_PRECISION (type) >= TYPE_PRECISION (itype)
+ && (POINTER_TYPE_P (type) || INTEGRAL_TYPE_P (type)))
+ {
+ split_constant_offset (op0, &var0, off);
+ *var = fold_convert (type, var0);
+ return true;
}
- break;
+ return false;
}
default:
- break;
+ return false;
}
+}
+
+/* Expresses EXP as VAR + OFF, where off is a constant. The type of OFF
+ will be ssizetype. */
+void
+split_constant_offset (tree exp, tree *var, tree *off)
+{
+ tree type = TREE_TYPE (exp), otype, op0, op1, e, o;
+ enum tree_code code;
+
+ *var = exp;
*off = ssize_int (0);
+ STRIP_NOPS (exp);
+
+ if (automatically_generated_chrec_p (exp))
+ return;
+
+ otype = TREE_TYPE (exp);
+ code = TREE_CODE (exp);
+ extract_ops_from_tree (exp, &code, &op0, &op1);
+ if (split_constant_offset_1 (otype, op0, code, op1, &e, &o))
+ {
+ *var = fold_convert (type, e);
+ *off = o;
+ }
}
/* Returns the address ADDR of an object in a canonical shape (without nop
return build_fold_addr_expr (TREE_OPERAND (addr, 0));
}
-/* Analyzes the behavior of the memory reference DR in the innermost loop that
- contains it. */
+/* Analyzes the behavior of the memory reference DR in the innermost loop or
+ basic block that contains it. Returns true if analysis succeed or false
+ otherwise. */
-void
+bool
dr_analyze_innermost (struct data_reference *dr)
{
- tree stmt = DR_STMT (dr);
+ gimple stmt = DR_STMT (dr);
struct loop *loop = loop_containing_stmt (stmt);
tree ref = DR_REF (dr);
HOST_WIDE_INT pbitsize, pbitpos;
int punsignedp, 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: ");
{
if (dump_file && (dump_flags & TDF_DETAILS))
fprintf (dump_file, "failed: bit offset alignment.\n");
- return;
+ return false;
}
base = build_fold_addr_expr (base);
- if (!simple_iv (loop, stmt, base, &base_iv, false))
+ if (in_loop)
{
- if (dump_file && (dump_flags & TDF_DETAILS))
- fprintf (dump_file, "failed: evolution of base is not affine.\n");
- return;
+ if (!simple_iv (loop, loop_containing_stmt (stmt), base, &base_iv,
+ false))
+ {
+ if (dump_file && (dump_flags & TDF_DETAILS))
+ fprintf (dump_file, "failed: evolution of base is not affine.\n");
+ return false;
+ }
+ }
+ 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 (!simple_iv (loop, stmt, poffset, &offset_iv, false))
+ else
{
- if (dump_file && (dump_flags & TDF_DETAILS))
- fprintf (dump_file, "failed: evolution of offset is not affine.\n");
- return;
+ if (!in_loop)
+ {
+ offset_iv.base = poffset;
+ offset_iv.step = ssize_int (0);
+ }
+ else if (!simple_iv (loop, loop_containing_stmt (stmt),
+ poffset, &offset_iv, false))
+ {
+ if (dump_file && (dump_flags & TDF_DETAILS))
+ fprintf (dump_file, "failed: evolution of offset is not"
+ " affine.\n");
+ return false;
+ }
}
init = ssize_int (pbitpos / BITS_PER_UNIT);
if (dump_file && (dump_flags & TDF_DETAILS))
fprintf (dump_file, "success.\n");
+
+ return true;
}
/* Determines the base object and the list of indices of memory reference
static void
dr_analyze_indices (struct data_reference *dr, struct loop *nest)
{
- tree stmt = DR_STMT (dr);
+ gimple stmt = DR_STMT (dr);
struct loop *loop = loop_containing_stmt (stmt);
VEC (tree, heap) *access_fns = NULL;
tree ref = unshare_expr (DR_REF (dr)), aref = ref, op;
- tree base, off, access_fn;
+ tree base, off, access_fn = NULL_TREE;
+ basic_block before_loop = NULL;
+
+ if (nest)
+ before_loop = block_before_loop (nest);
while (handled_component_p (aref))
{
if (TREE_CODE (aref) == ARRAY_REF)
{
op = TREE_OPERAND (aref, 1);
- access_fn = analyze_scalar_evolution (loop, op);
- access_fn = resolve_mixers (nest, access_fn);
- VEC_safe_push (tree, heap, access_fns, access_fn);
+ if (nest)
+ {
+ access_fn = analyze_scalar_evolution (loop, op);
+ access_fn = instantiate_scev (before_loop, loop, access_fn);
+ VEC_safe_push (tree, heap, access_fns, access_fn);
+ }
TREE_OPERAND (aref, 1) = build_int_cst (TREE_TYPE (op), 0);
}
-
+
aref = TREE_OPERAND (aref, 0);
}
- if (INDIRECT_REF_P (aref))
+ if (nest && INDIRECT_REF_P (aref))
{
op = TREE_OPERAND (aref, 0);
access_fn = analyze_scalar_evolution (loop, op);
- access_fn = resolve_mixers (nest, access_fn);
+ access_fn = instantiate_scev (before_loop, loop, access_fn);
base = initial_condition (access_fn);
split_constant_offset (base, &base, &off);
access_fn = chrec_replace_initial_condition (access_fn,
static void
dr_analyze_alias (struct data_reference *dr)
{
- tree stmt = DR_STMT (dr);
tree ref = DR_REF (dr);
- tree base = get_base_address (ref), addr, smt = NULL_TREE;
- ssa_op_iter it;
- tree op;
- bitmap vops;
+ tree base = get_base_address (ref), addr;
- if (DECL_P (base))
- smt = base;
- else if (INDIRECT_REF_P (base))
+ if (INDIRECT_REF_P (base))
{
addr = TREE_OPERAND (base, 0);
if (TREE_CODE (addr) == SSA_NAME)
- {
- smt = symbol_mem_tag (SSA_NAME_VAR (addr));
- DR_PTR_INFO (dr) = SSA_NAME_PTR_INFO (addr);
- }
+ DR_PTR_INFO (dr) = SSA_NAME_PTR_INFO (addr);
}
-
- DR_SYMBOL_TAG (dr) = smt;
- if (smt && var_can_have_subvars (smt))
- DR_SUBVARS (dr) = get_subvars_for_var (smt);
-
- vops = BITMAP_ALLOC (NULL);
- FOR_EACH_SSA_TREE_OPERAND (op, stmt, it, SSA_OP_VIRTUAL_USES)
- {
- bitmap_set_bit (vops, DECL_UID (SSA_NAME_VAR (op)));
- }
-
- DR_VOPS (dr) = vops;
}
/* Returns true if the address of DR is invariant. */
void
free_data_ref (data_reference_p dr)
{
- BITMAP_FREE (DR_VOPS (dr));
VEC_free (tree, heap, DR_ACCESS_FNS (dr));
free (dr);
}
loop nest in that the reference should be analyzed. */
struct data_reference *
-create_data_ref (struct loop *nest, tree memref, tree stmt, bool is_read)
+create_data_ref (struct loop *nest, tree memref, gimple stmt, bool is_read)
{
struct data_reference *dr;
print_generic_expr (dump_file, DR_ALIGNED_TO (dr), TDF_SLIM);
fprintf (dump_file, "\n\tbase_object: ");
print_generic_expr (dump_file, DR_BASE_OBJECT (dr), TDF_SLIM);
- fprintf (dump_file, "\n\tsymbol tag: ");
- print_generic_expr (dump_file, DR_SYMBOL_TAG (dr), TDF_SLIM);
fprintf (dump_file, "\n");
}
- return dr;
+ return dr;
}
/* Returns true if FNA == FNB. */
VEC_quick_push (tree, ret,
fold_build2 (op, type,
- VEC_index (tree, fna, i),
+ VEC_index (tree, fna, i),
VEC_index (tree, fnb, i)));
}
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);
return;
}
diff = affine_fn_minus (fn_a, fn_b);
-
+
if (affine_function_constant_p (diff))
SUB_DISTANCE (subscript) = affine_function_base (diff);
else
/* Returns false if we can prove that data references A and B do not alias,
true otherwise. */
-static bool
+bool
dr_may_alias_p (const struct data_reference *a, const struct data_reference *b)
{
const_tree addr_a = DR_BASE_ADDRESS (a);
const_tree type_a, type_b;
const_tree decl_a = NULL_TREE, decl_b = NULL_TREE;
- /* If the sets of virtual operands are disjoint, the memory references do not
- alias. */
- if (!bitmap_intersect_p (DR_VOPS (a), DR_VOPS (b)))
- return false;
-
/* If the accessed objects are disjoint, the memory references do not
alias. */
if (disjoint_objects_p (DR_BASE_OBJECT (a), DR_BASE_OBJECT (b)))
return false;
+ /* Query the alias oracle. */
+ if (!DR_IS_READ (a) && !DR_IS_READ (b))
+ {
+ if (!refs_output_dependent_p (DR_REF (a), DR_REF (b)))
+ return false;
+ }
+ else if (DR_IS_READ (a) && !DR_IS_READ (b))
+ {
+ if (!refs_anti_dependent_p (DR_REF (a), DR_REF (b)))
+ return false;
+ }
+ else if (!refs_may_alias_p (DR_REF (a), DR_REF (b)))
+ return false;
+
if (!addr_a || !addr_b)
return true;
- /* If the references are based on different static objects, they cannot alias
- (PTA should be able to disambiguate such accesses, but often it fails to,
- since currently we cannot distinguish between pointer and offset in pointer
- arithmetics). */
+ /* If the references are based on different static objects, they cannot
+ alias (PTA should be able to disambiguate such accesses, but often
+ it fails to). */
if (TREE_CODE (addr_a) == ADDR_EXPR
&& TREE_CODE (addr_b) == ADDR_EXPR)
return TREE_OPERAND (addr_a, 0) == TREE_OPERAND (addr_b, 0);
- /* An instruction writing through a restricted pointer is "independent" of any
- instruction reading or writing through a different restricted pointer,
+ /* An instruction writing through a restricted pointer is "independent" of any
+ instruction reading or writing through a different restricted pointer,
in the same block/scope. */
type_a = TREE_TYPE (addr_a);
if (TREE_CODE (addr_b) == SSA_NAME)
decl_b = SSA_NAME_VAR (addr_b);
- if (TYPE_RESTRICT (type_a) && TYPE_RESTRICT (type_b)
+ if (TYPE_RESTRICT (type_a) && TYPE_RESTRICT (type_b)
&& (!DR_IS_READ (a) || !DR_IS_READ (b))
&& decl_a && DECL_P (decl_a)
&& decl_b && DECL_P (decl_b)
return true;
}
+static void compute_self_dependence (struct data_dependence_relation *);
+
/* 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. */
static struct data_dependence_relation *
-initialize_data_dependence_relation (struct data_reference *a,
+initialize_data_dependence_relation (struct data_reference *a,
struct data_reference *b,
VEC (loop_p, heap) *loop_nest)
{
struct data_dependence_relation *res;
unsigned int i;
-
+
res = XNEW (struct data_dependence_relation);
DDR_A (res) = a;
DDR_B (res) = b;
if (a == NULL || b == NULL)
{
- DDR_ARE_DEPENDENT (res) = chrec_dont_know;
+ 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))
{
- DDR_ARE_DEPENDENT (res) = chrec_known;
+ DDR_ARE_DEPENDENT (res) = chrec_known;
+ return res;
+ }
+
+ /* When the references are exactly the same, don't spend time doing
+ the data dependence tests, just initialize the ddr and return. */
+ if (operand_equal_p (DR_REF (a), DR_REF (b), 0))
+ {
+ DDR_AFFINE_P (res) = true;
+ DDR_ARE_DEPENDENT (res) = NULL_TREE;
+ DDR_SUBSCRIPTS (res) = VEC_alloc (subscript_p, heap, DR_NUM_DIMENSIONS (a));
+ DDR_LOOP_NEST (res) = loop_nest;
+ DDR_INNER_LOOP (res) = 0;
+ DDR_SELF_REFERENCE (res) = true;
+ compute_self_dependence (res);
return res;
}
whether they alias or not. */
if (!operand_equal_p (DR_BASE_OBJECT (a), DR_BASE_OBJECT (b), 0))
{
- DDR_ARE_DEPENDENT (res) = chrec_dont_know;
+ DDR_ARE_DEPENDENT (res) = chrec_dont_know;
return res;
}
/* If the base of the object is not invariant in the loop nest, we cannot
analyze it. TODO -- in fact, it would suffice to record that there may
be arbitrary dependences in the loops where the base object varies. */
- if (!object_address_invariant_in_loop_p (VEC_index (loop_p, loop_nest, 0),
- DR_BASE_OBJECT (a)))
+ if (loop_nest
+ && !object_address_invariant_in_loop_p (VEC_index (loop_p, loop_nest, 0),
+ DR_BASE_OBJECT (a)))
{
- DDR_ARE_DEPENDENT (res) = chrec_dont_know;
+ DDR_ARE_DEPENDENT (res) = chrec_dont_know;
return res;
}
DDR_SUBSCRIPTS (res) = VEC_alloc (subscript_p, heap, DR_NUM_DIMENSIONS (a));
DDR_LOOP_NEST (res) = loop_nest;
DDR_INNER_LOOP (res) = 0;
+ DDR_SELF_REFERENCE (res) = false;
for (i = 0; i < DR_NUM_DIMENSIONS (a); i++)
{
struct subscript *subscript;
-
+
subscript = XNEW (struct subscript);
SUB_CONFLICTS_IN_A (subscript) = conflict_fn_not_known ();
SUB_CONFLICTS_IN_B (subscript) = conflict_fn_not_known ();
{
free_conflict_function (s->conflicting_iterations_in_a);
free_conflict_function (s->conflicting_iterations_in_b);
+ free (s);
}
VEC_free (subscript_p, heap, subscripts);
}
description. */
static inline void
-finalize_ddr_dependent (struct data_dependence_relation *ddr,
+finalize_ddr_dependent (struct data_dependence_relation *ddr,
tree chrec)
{
if (dump_file && (dump_flags & TDF_DETAILS))
fprintf (dump_file, ")\n");
}
- DDR_ARE_DEPENDENT (ddr) = chrec;
+ DDR_ARE_DEPENDENT (ddr) = chrec;
free_subscripts (DDR_SUBSCRIPTS (ddr));
DDR_SUBSCRIPTS (ddr) = NULL;
}
|| (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))
{
case POLYNOMIAL_CHREC:
if (CHREC_VARIABLE (chrec_a) != CHREC_VARIABLE (chrec_b))
return false;
-
+
default:
return true;
}
-
+
default:
return true;
}
}
-
+
return false;
}
gcc_assert (0 < n && n <= MAX_DIM);
va_start(ap, n);
-
+
ret->n = n;
for (i = 0; i < n; i++)
ret->fns[i] = va_arg (ap, affine_fn);
CHREC_A (*OVERLAPS_A (k)) = CHREC_B (*OVERLAPS_B (k)). */
-static void
-analyze_ziv_subscript (tree chrec_a,
- tree chrec_b,
+static void
+analyze_ziv_subscript (tree chrec_a,
+ tree chrec_b,
conflict_function **overlaps_a,
- conflict_function **overlaps_b,
+ 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_TREE);
- chrec_b = chrec_convert (type, chrec_b, NULL_TREE);
+ 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:
dependence_stats.num_ziv_independent++;
}
break;
-
+
default:
- /* We're not sure whether the indexes overlap. For the moment,
+ /* 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");
dependence_stats.num_ziv_unimplemented++;
break;
}
-
+
if (dump_file && (dump_flags & TDF_DETAILS))
fprintf (dump_file, ")\n");
}
return hwi_nit < 0 ? -1 : hwi_nit;
}
-
+
/* Similar to estimated_loop_iterations, but returns the estimate as a tree,
and only if it fits to the int type. If this is not the case, or the
estimate on the number of iterations of LOOP could not be derived, returns
CHREC_A (*OVERLAPS_A (k)) = CHREC_B (*OVERLAPS_B (k)). */
static void
-analyze_siv_subscript_cst_affine (tree chrec_a,
+analyze_siv_subscript_cst_affine (tree chrec_a,
tree chrec_b,
- conflict_function **overlaps_a,
- conflict_function **overlaps_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_TREE);
- chrec_b = chrec_convert (type, chrec_b, NULL_TREE);
+ 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);
-
+
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");
+ fprintf (dump_file, "siv test failed: chrec is not positive.\n");
dependence_stats.num_siv_unimplemented++;
*overlaps_a = conflict_fn_not_known ();
fprintf (dump_file, "siv test failed: chrec not positive.\n");
*overlaps_a = conflict_fn_not_known ();
- *overlaps_b = conflict_fn_not_known ();
+ *overlaps_b = conflict_fn_not_known ();
*last_conflicts = chrec_dont_know;
dependence_stats.num_siv_unimplemented++;
return;
{
if (value1 == true)
{
- /* Example:
+ /* Example:
chrec_a = 12
chrec_b = {10, +, 1}
*/
-
+
if (tree_fold_divides_p (CHREC_RIGHT (chrec_b), difference))
{
HOST_WIDE_INT numiter;
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. */
*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 ();
+ *overlaps_b = conflict_fn_no_dependence ();
*last_conflicts = integer_zero_node;
dependence_stats.num_siv_independent++;
return;
}
}
-
+
else
{
- /* Example:
+ /* 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 ();
}
}
}
- else
+ else
{
if (!chrec_is_positive (CHREC_RIGHT (chrec_b), &value2))
{
fprintf (dump_file, "siv test failed: chrec not positive.\n");
*overlaps_a = conflict_fn_not_known ();
- *overlaps_b = conflict_fn_not_known ();
+ *overlaps_b = conflict_fn_not_known ();
*last_conflicts = chrec_dont_know;
dependence_stats.num_siv_unimplemented++;
return;
{
if (value2 == false)
{
- /* Example:
+ /* Example:
chrec_a = 3
chrec_b = {10, +, -1}
*/
*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 ();
+ *overlaps_b = conflict_fn_no_dependence ();
*last_conflicts = integer_zero_node;
dependence_stats.num_siv_independent++;
return;
}
else
{
- /* Example:
- chrec_a = 3
+ /* 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 ();
/* Helper recursive function for initializing the matrix A. Returns
the initial value of CHREC. */
-static HOST_WIDE_INT
+static tree
initialize_matrix_A (lambda_matrix A, tree chrec, unsigned index, int mult)
{
gcc_assert (chrec);
- if (TREE_CODE (chrec) != POLYNOMIAL_CHREC)
- return int_cst_value (chrec);
+ switch (TREE_CODE (chrec))
+ {
+ case POLYNOMIAL_CHREC:
+ gcc_assert (TREE_CODE (CHREC_RIGHT (chrec)) == INTEGER_CST);
- A[index][0] = mult * int_cst_value (CHREC_RIGHT (chrec));
- return initialize_matrix_A (A, CHREC_LEFT (chrec), index + 1, mult);
+ A[index][0] = mult * int_cst_value (CHREC_RIGHT (chrec));
+ return initialize_matrix_A (A, CHREC_LEFT (chrec), index + 1, mult);
+
+ case PLUS_EXPR:
+ case MULT_EXPR:
+ case MINUS_EXPR:
+ {
+ tree op0 = initialize_matrix_A (A, TREE_OPERAND (chrec, 0), index, mult);
+ tree op1 = initialize_matrix_A (A, TREE_OPERAND (chrec, 1), index, mult);
+
+ return chrec_fold_op (TREE_CODE (chrec), chrec_type (chrec), op0, op1);
+ }
+
+ case NOP_EXPR:
+ {
+ tree op = initialize_matrix_A (A, TREE_OPERAND (chrec, 0), index, mult);
+ return chrec_convert (chrec_type (chrec), op, NULL);
+ }
+
+ case BIT_NOT_EXPR:
+ {
+ /* Handle ~X as -1 - X. */
+ tree op = initialize_matrix_A (A, TREE_OPERAND (chrec, 0), index, mult);
+ return chrec_fold_op (MINUS_EXPR, chrec_type (chrec),
+ build_int_cst (TREE_TYPE (chrec), -1), op);
+ }
+
+ case INTEGER_CST:
+ return chrec;
+
+ default:
+ gcc_unreachable ();
+ return NULL_TREE;
+ }
}
#define FLOOR_DIV(x,y) ((x) / (y))
-/* Solves the special case of the Diophantine equation:
+/* Solves the special case of the Diophantine equation:
| {0, +, STEP_A}_x (OVERLAPS_A) = {0, +, STEP_B}_y (OVERLAPS_B)
Computes the descriptions OVERLAPS_A and OVERLAPS_B. NITER is the
constructed as evolutions in dimension DIM. */
static void
-compute_overlap_steps_for_affine_univar (int niter, int step_a, int step_b,
+compute_overlap_steps_for_affine_univar (int niter, int step_a, int step_b,
affine_fn *overlaps_a,
- affine_fn *overlaps_b,
+ affine_fn *overlaps_b,
tree *last_conflicts, int dim)
{
if (((step_a > 0 && step_b > 0)
else
*last_conflicts = chrec_dont_know;
- *overlaps_a = affine_fn_univar (integer_zero_node, dim,
+ *overlaps_a = affine_fn_univar (integer_zero_node, dim,
build_int_cst (NULL_TREE,
step_overlaps_a));
- *overlaps_b = affine_fn_univar (integer_zero_node, dim,
- build_int_cst (NULL_TREE,
+ *overlaps_b = affine_fn_univar (integer_zero_node, dim,
+ build_int_cst (NULL_TREE,
step_overlaps_b));
}
/* Solves the special case of a Diophantine equation where CHREC_A is
an affine bivariate function, and CHREC_B is an affine univariate
- function. For example,
+ function. For example,
| {{0, +, 1}_x, +, 1335}_y = {0, +, 1336}_z
-
- has the following overlapping functions:
+
+ has the following overlapping functions:
| x (t, u, v) = {{0, +, 1336}_t, +, 1}_v
| y (t, u, v) = {{0, +, 1336}_u, +, 1}_v
a common benchmark. Implement the general algorithm. */
static void
-compute_overlap_steps_for_affine_1_2 (tree chrec_a, tree chrec_b,
+compute_overlap_steps_for_affine_1_2 (tree chrec_a, tree chrec_b,
conflict_function **overlaps_a,
- conflict_function **overlaps_b,
+ conflict_function **overlaps_b,
tree *last_conflicts)
{
bool xz_p, yz_p, xyz_p;
step_y = int_cst_value (CHREC_RIGHT (chrec_a));
step_z = int_cst_value (CHREC_RIGHT (chrec_b));
- niter_x =
+ niter_x =
estimated_loop_iterations_int (get_chrec_loop (CHREC_LEFT (chrec_a)),
false);
niter_y = estimated_loop_iterations_int (get_chrec_loop (chrec_a), false);
niter_z = estimated_loop_iterations_int (get_chrec_loop (chrec_b), false);
-
+
if (niter_x < 0 || niter_y < 0 || niter_z < 0)
{
if (dump_file && (dump_flags & TDF_DETAILS))
fprintf (dump_file, "overlap steps test failed: no iteration counts.\n");
-
+
*overlaps_a = conflict_fn_not_known ();
*overlaps_b = conflict_fn_not_known ();
*last_conflicts = chrec_dont_know;
parameters, because it uses lambda matrices of integers. */
static void
-analyze_subscript_affine_affine (tree chrec_a,
+analyze_subscript_affine_affine (tree chrec_a,
tree chrec_b,
- conflict_function **overlaps_a,
- conflict_function **overlaps_b,
+ conflict_function **overlaps_a,
+ conflict_function **overlaps_b,
tree *last_conflicts)
{
unsigned nb_vars_a, nb_vars_b, dim;
}
if (dump_file && (dump_flags & TDF_DETAILS))
fprintf (dump_file, "(analyze_subscript_affine_affine \n");
-
+
/* For determining the initial intersection, we have to solve a
Diophantine equation. This is the most time consuming part.
-
+
For answering to the question: "Is there a dependence?" we have
to prove that there exists a solution to the Diophantine
equation, and that the solution is in the iteration domain,
A = lambda_matrix_new (dim, 1);
S = lambda_matrix_new (dim, 1);
- init_a = initialize_matrix_A (A, chrec_a, 0, 1);
- init_b = initialize_matrix_A (A, chrec_b, nb_vars_a, -1);
+ init_a = int_cst_value (initialize_matrix_A (A, chrec_a, 0, 1));
+ init_b = int_cst_value (initialize_matrix_A (A, chrec_b, nb_vars_a, -1));
gamma = init_b - init_a;
/* Don't do all the hard work of solving the Diophantine equation
- when we already know the solution: for example,
+ when we already know the solution: for example,
| {3, +, 1}_1
| {3, +, 4}_2
| gamma = 3 - 3 = 0.
- Then the first overlap occurs during the first iterations:
+ Then the first overlap occurs during the first iterations:
| {3, +, 1}_1 ({0, +, 4}_x) = {3, +, 4}_2 ({0, +, 1}_x)
*/
if (gamma == 0)
step_a = int_cst_value (CHREC_RIGHT (chrec_a));
step_b = int_cst_value (CHREC_RIGHT (chrec_b));
- compute_overlap_steps_for_affine_univar (niter, step_a, step_b,
- &ova, &ovb,
+ compute_overlap_steps_for_affine_univar (niter, step_a, step_b,
+ &ova, &ovb,
last_conflicts, 1);
*overlaps_a = conflict_fn (1, ova);
*overlaps_b = conflict_fn (1, ovb);
|| (A[0][0] < 0 && -A[1][0] < 0)))
{
/* The solutions are given by:
- |
+ |
| [GAMMA/GCD_ALPHA_BETA t].[u11 u12] = [x0]
| [u21 u22] [y0]
-
+
For a given integer t. Using the following variables,
-
+
| i0 = u11 * gamma / gcd_alpha_beta
| j0 = u12 * gamma / gcd_alpha_beta
| i1 = u21
| j1 = u22
-
+
the solutions are:
-
- | x0 = i0 + i1 * t,
+
+ | x0 = i0 + i1 * t,
| y0 = j0 + j1 * t. */
HOST_WIDE_INT i0, j0, i1, j1;
if ((i1 == 0 && i0 < 0)
|| (j1 == 0 && j0 < 0))
{
- /* There is no solution.
- FIXME: The case "i0 > nb_iterations, j0 > nb_iterations"
- falls in here, but for the moment we don't look at the
+ /* There is no solution.
+ FIXME: The case "i0 > nb_iterations, j0 > nb_iterations"
+ falls in here, but for the moment we don't look at the
upper bound of the iteration domain. */
*overlaps_a = conflict_fn_no_dependence ();
*overlaps_b = conflict_fn_no_dependence ();
/* If the overlap occurs outside of the bounds of the
loop, there is no dependence. */
- if (x1 > niter || y1 > niter)
+ if (x1 >= niter || y1 >= niter)
{
*overlaps_a = conflict_fn_no_dependence ();
*overlaps_b = conflict_fn_no_dependence ();
*last_conflicts = chrec_dont_know;
}
-end_analyze_subs_aa:
+end_analyze_subs_aa:
if (dump_file && (dump_flags & TDF_DETAILS))
{
fprintf (dump_file, " (overlaps_a = ");
determining the dependence relation between chrec_a and chrec_b,
that contain symbols. This function modifies chrec_a and chrec_b
such that the analysis result is the same, and such that they don't
- contain symbols, and then can safely be passed to the analyzer.
+ contain symbols, and then can safely be passed to the analyzer.
Example: The analysis of the following tuples of evolutions produce
the same results: {x+1, +, 1}_1 vs. {x+3, +, 1}_1, and {-2, +, 1}_1
vs. {0, +, 1}_1
-
+
{x+1, +, 1}_1 ({2, +, 1}_1) = {x+3, +, 1}_1 ({0, +, 1}_1)
{-2, +, 1}_1 ({2, +, 1}_1) = {0, +, 1}_1 ({0, +, 1}_1)
*/
type = chrec_type (*chrec_a);
left_a = CHREC_LEFT (*chrec_a);
- left_b = chrec_convert (type, CHREC_LEFT (*chrec_b), NULL_TREE);
+ left_b = chrec_convert (type, CHREC_LEFT (*chrec_b), NULL);
diff = chrec_fold_minus (type, left_a, left_b);
if (!evolution_function_is_constant_p (diff))
if (dump_file && (dump_flags & TDF_DETAILS))
fprintf (dump_file, "can_use_subscript_aff_aff_for_symbolic \n");
- *chrec_a = build_polynomial_chrec (CHREC_VARIABLE (*chrec_a),
+ *chrec_a = build_polynomial_chrec (CHREC_VARIABLE (*chrec_a),
diff, CHREC_RIGHT (*chrec_a));
- right_b = chrec_convert (type, CHREC_RIGHT (*chrec_b), NULL_TREE);
+ right_b = chrec_convert (type, CHREC_RIGHT (*chrec_b), NULL);
*chrec_b = build_polynomial_chrec (CHREC_VARIABLE (*chrec_b),
build_int_cst (type, 0),
right_b);
CHREC_A (*OVERLAPS_A (k)) = CHREC_B (*OVERLAPS_B (k)). */
static void
-analyze_siv_subscript (tree chrec_a,
+analyze_siv_subscript (tree chrec_a,
tree chrec_b,
- conflict_function **overlaps_a,
- conflict_function **overlaps_b,
- tree *last_conflicts)
+ conflict_function **overlaps_a,
+ conflict_function **overlaps_b,
+ tree *last_conflicts,
+ int loop_nest_num)
{
dependence_stats.num_siv++;
-
+
if (dump_file && (dump_flags & TDF_DETAILS))
fprintf (dump_file, "(analyze_siv_subscript \n");
-
+
if (evolution_function_is_constant_p (chrec_a)
- && evolution_function_is_affine_p (chrec_b))
- analyze_siv_subscript_cst_affine (chrec_a, chrec_b,
+ && evolution_function_is_affine_in_loop (chrec_b, loop_nest_num))
+ analyze_siv_subscript_cst_affine (chrec_a, chrec_b,
overlaps_a, overlaps_b, last_conflicts);
-
- else if (evolution_function_is_affine_p (chrec_a)
+
+ else if (evolution_function_is_affine_in_loop (chrec_a, loop_nest_num)
&& evolution_function_is_constant_p (chrec_b))
- analyze_siv_subscript_cst_affine (chrec_b, chrec_a,
+ analyze_siv_subscript_cst_affine (chrec_b, chrec_a,
overlaps_b, overlaps_a, last_conflicts);
-
- else if (evolution_function_is_affine_p (chrec_a)
- && evolution_function_is_affine_p (chrec_b))
+
+ else if (evolution_function_is_affine_in_loop (chrec_a, loop_nest_num)
+ && evolution_function_is_affine_in_loop (chrec_b, loop_nest_num))
{
if (!chrec_contains_symbols (chrec_a)
&& !chrec_contains_symbols (chrec_b))
{
- analyze_subscript_affine_affine (chrec_a, chrec_b,
- overlaps_a, overlaps_b,
+ analyze_subscript_affine_affine (chrec_a, chrec_b,
+ overlaps_a, overlaps_b,
last_conflicts);
if (CF_NOT_KNOWN_P (*overlaps_a)
else
dependence_stats.num_siv_dependent++;
}
- else if (can_use_analyze_subscript_affine_affine (&chrec_a,
+ else if (can_use_analyze_subscript_affine_affine (&chrec_a,
&chrec_b))
{
- analyze_subscript_affine_affine (chrec_a, chrec_b,
- overlaps_a, overlaps_b,
+ analyze_subscript_affine_affine (chrec_a, chrec_b,
+ overlaps_a, overlaps_b,
last_conflicts);
if (CF_NOT_KNOWN_P (*overlaps_a)
*last_conflicts = chrec_dont_know;
dependence_stats.num_siv_unimplemented++;
}
-
+
if (dump_file && (dump_flags & TDF_DETAILS))
fprintf (dump_file, ")\n");
}
CHREC_A (*OVERLAPS_A (k)) = CHREC_B (*OVERLAPS_B (k)). */
static void
-analyze_miv_subscript (tree chrec_a,
- tree chrec_b,
- conflict_function **overlaps_a,
- conflict_function **overlaps_b,
+analyze_miv_subscript (tree chrec_a,
+ tree chrec_b,
+ conflict_function **overlaps_a,
+ conflict_function **overlaps_b,
tree *last_conflicts,
struct loop *loop_nest)
{
/* FIXME: This is a MIV subscript, not yet handled.
- Example: (A[{1, +, 1}_1] vs. A[{1, +, 1}_2]) that comes from
- (A[i] vs. A[j]).
-
+ Example: (A[{1, +, 1}_1] vs. A[{1, +, 1}_2]) that comes from
+ (A[i] vs. A[j]).
+
In the SIV test we had to solve a Diophantine equation with two
variables. In the MIV case we have to solve a Diophantine
equation with 2*n variables (if the subscript uses n IVs).
fprintf (dump_file, "(analyze_miv_subscript \n");
type = signed_type_for_types (TREE_TYPE (chrec_a), TREE_TYPE (chrec_b));
- chrec_a = chrec_convert (type, chrec_a, NULL_TREE);
- chrec_b = chrec_convert (type, chrec_b, NULL_TREE);
+ 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);
-
+
if (eq_evolutions_p (chrec_a, chrec_b))
{
/* Access functions are the same: all the elements are accessed
(get_chrec_loop (chrec_a), true);
dependence_stats.num_miv_dependent++;
}
-
+
else if (evolution_function_is_constant_p (difference)
/* For the moment, the following is verified:
evolution_function_is_affine_multivariate_p (chrec_a,
&& !gcd_of_steps_may_divide_p (chrec_a, difference))
{
/* testsuite/.../ssa-chrec-33.c
- {{21, +, 2}_1, +, -2}_2 vs. {{20, +, 2}_1, +, -2}_2
-
+ {{21, +, 2}_1, +, -2}_2 vs. {{20, +, 2}_1, +, -2}_2
+
The difference is 1, and all the evolution steps are multiples
of 2, consequently there are no overlapping elements. */
*overlaps_a = conflict_fn_no_dependence ();
*last_conflicts = integer_zero_node;
dependence_stats.num_miv_independent++;
}
-
+
else if (evolution_function_is_affine_multivariate_p (chrec_a, loop_nest->num)
&& !chrec_contains_symbols (chrec_a)
&& evolution_function_is_affine_multivariate_p (chrec_b, loop_nest->num)
/* testsuite/.../ssa-chrec-35.c
{0, +, 1}_2 vs. {0, +, 1}_3
the overlapping elements are respectively located at iterations:
- {0, +, 1}_x and {0, +, 1}_x,
- in other words, we have the equality:
+ {0, +, 1}_x and {0, +, 1}_x,
+ in other words, we have the equality:
{0, +, 1}_2 ({0, +, 1}_x) = {0, +, 1}_3 ({0, +, 1}_x)
-
- Other examples:
- {{0, +, 1}_1, +, 2}_2 ({0, +, 1}_x, {0, +, 1}_y) =
+
+ Other examples:
+ {{0, +, 1}_1, +, 2}_2 ({0, +, 1}_x, {0, +, 1}_y) =
{0, +, 1}_1 ({{0, +, 1}_x, +, 2}_y)
- {{0, +, 2}_1, +, 3}_2 ({0, +, 1}_y, {0, +, 1}_x) =
+ {{0, +, 2}_1, +, 3}_2 ({0, +, 1}_y, {0, +, 1}_x) =
{{0, +, 3}_1, +, 2}_2 ({0, +, 1}_x, {0, +, 1}_y)
*/
- analyze_subscript_affine_affine (chrec_a, chrec_b,
+ analyze_subscript_affine_affine (chrec_a, chrec_b,
overlaps_a, overlaps_b, last_conflicts);
if (CF_NOT_KNOWN_P (*overlaps_a)
else
dependence_stats.num_miv_dependent++;
}
-
+
else
{
/* When the analysis is too difficult, answer "don't know". */
*last_conflicts = chrec_dont_know;
dependence_stats.num_miv_unimplemented++;
}
-
+
if (dump_file && (dump_flags & TDF_DETAILS))
fprintf (dump_file, ")\n");
}
with respect to LOOP_NEST. OVERLAP_ITERATIONS_A and
OVERLAP_ITERATIONS_B are initialized with two functions that
describe the iterations that contain conflicting elements.
-
+
Remark: For an integer k >= 0, the following equality is true:
-
+
CHREC_A (OVERLAP_ITERATIONS_A (k)) == CHREC_B (OVERLAP_ITERATIONS_B (k)).
*/
-static void
-analyze_overlapping_iterations (tree chrec_a,
- tree chrec_b,
- conflict_function **overlap_iterations_a,
- conflict_function **overlap_iterations_b,
+static void
+analyze_overlapping_iterations (tree chrec_a,
+ tree chrec_b,
+ conflict_function **overlap_iterations_a,
+ conflict_function **overlap_iterations_b,
tree *last_conflicts, struct loop *loop_nest)
{
unsigned int lnn = loop_nest->num;
dependence_stats.num_subscript_tests++;
-
+
if (dump_file && (dump_flags & TDF_DETAILS))
{
fprintf (dump_file, "(analyze_overlapping_iterations \n");
|| chrec_contains_undetermined (chrec_b))
{
dependence_stats.num_subscript_undetermined++;
-
+
*overlap_iterations_a = conflict_fn_not_known ();
*overlap_iterations_b = conflict_fn_not_known ();
}
- /* If they are the same chrec, and are affine, they overlap
+ /* If they are the same chrec, and are affine, they overlap
on every iteration. */
else if (eq_evolutions_p (chrec_a, chrec_b)
&& evolution_function_is_affine_multivariate_p (chrec_a, lnn))
/* If they aren't the same, and aren't affine, we can't do anything
yet. */
- else if ((chrec_contains_symbols (chrec_a)
+ else if ((chrec_contains_symbols (chrec_a)
|| chrec_contains_symbols (chrec_b))
&& (!evolution_function_is_affine_multivariate_p (chrec_a, lnn)
|| !evolution_function_is_affine_multivariate_p (chrec_b, lnn)))
}
else if (ziv_subscript_p (chrec_a, chrec_b))
- analyze_ziv_subscript (chrec_a, chrec_b,
+ analyze_ziv_subscript (chrec_a, chrec_b,
overlap_iterations_a, overlap_iterations_b,
last_conflicts);
-
+
else if (siv_subscript_p (chrec_a, chrec_b))
- analyze_siv_subscript (chrec_a, chrec_b,
- overlap_iterations_a, overlap_iterations_b,
- last_conflicts);
-
+ analyze_siv_subscript (chrec_a, chrec_b,
+ overlap_iterations_a, overlap_iterations_b,
+ last_conflicts, lnn);
+
else
- analyze_miv_subscript (chrec_a, chrec_b,
+ analyze_miv_subscript (chrec_a, chrec_b,
overlap_iterations_a, overlap_iterations_b,
last_conflicts, loop_nest);
-
+
if (dump_file && (dump_flags & TDF_DETAILS))
{
fprintf (dump_file, " (overlap_iterations_a = ");
access_fn_a = DR_ACCESS_FN (ddr_a, i);
access_fn_b = DR_ACCESS_FN (ddr_b, i);
- if (TREE_CODE (access_fn_a) == POLYNOMIAL_CHREC
+ if (TREE_CODE (access_fn_a) == POLYNOMIAL_CHREC
&& TREE_CODE (access_fn_b) == POLYNOMIAL_CHREC)
{
int dist, index;
non_affine_dependence_relation (ddr);
return false;
}
-
+
dist = int_cst_value (SUB_DISTANCE (subscript));
/* This is the subscript coupling test. If we have already
| T[j][i] = t + 2; // B
| }
- the vectors are:
+ the vectors are:
(0, 1, -1)
(1, 1, -1)
(1, -1, 1)
{
conflict_function *overlaps_a, *overlaps_b;
- analyze_overlapping_iterations (DR_ACCESS_FN (dra, i),
+ analyze_overlapping_iterations (DR_ACCESS_FN (dra, i),
DR_ACCESS_FN (drb, i),
- &overlaps_a, &overlaps_b,
+ &overlaps_a, &overlaps_b,
&last_conflicts, loop_nest);
if (CF_NOT_KNOWN_P (overlaps_a)
subscript_dependence_tester (struct data_dependence_relation *ddr,
struct loop *loop_nest)
{
-
+
if (dump_file && (dump_flags & TDF_DETAILS))
fprintf (dump_file, "(subscript_dependence_tester \n");
-
+
if (subscript_dependence_tester_1 (ddr, DDR_A (ddr), DDR_B (ddr), loop_nest))
dependence_stats.num_dependence_dependent++;
/* Returns true when all the access functions of A are affine or
constant with respect to LOOP_NEST. */
-static bool
+static bool
access_functions_are_affine_or_constant_p (const struct data_reference *a,
const struct loop *loop_nest)
{
if (!evolution_function_is_invariant_p (t, loop_nest->num)
&& !evolution_function_is_affine_multivariate_p (t, loop_nest->num))
return false;
-
+
return true;
}
ACCESS_FUN is expected to be an affine chrec. */
static bool
-init_omega_eq_with_af (omega_pb pb, unsigned eq,
- unsigned int offset, tree access_fun,
+init_omega_eq_with_af (omega_pb pb, unsigned eq,
+ unsigned int offset, tree access_fun,
struct data_dependence_relation *ddr)
{
switch (TREE_CODE (access_fun))
DDR_INNER_LOOP (ddr) = MAX (DDR_INNER_LOOP (ddr), var_idx);
if (offset == 0)
- pb->eqs[eq].coef[var_idx + DDR_NB_LOOPS (ddr) + 1]
+ pb->eqs[eq].coef[var_idx + DDR_NB_LOOPS (ddr) + 1]
+= int_cst_value (right);
switch (TREE_CODE (left))
/* Set a new problem for each loop in the nest. The basis is the
problem that we have initialized until now. On top of this we
add new constraints. */
- for (i = 0; i <= DDR_INNER_LOOP (ddr)
+ for (i = 0; i <= DDR_INNER_LOOP (ddr)
&& VEC_iterate (loop_p, DDR_LOOP_NEST (ddr), i, loopi); i++)
{
int dist = 0;
/* Reduce the constraint system, and test that the current
problem is feasible. */
res = omega_simplify_problem (copy);
- if (res == omega_false
+ if (res == omega_false
|| res == omega_unknown
|| copy->num_geqs > (int) DDR_NB_LOOPS (ddr))
goto next_problem;
copy->eqs[eq].coef[0] = -1;
res = omega_simplify_problem (copy);
- if (res == omega_false
+ if (res == omega_false
|| res == omega_unknown
|| copy->num_geqs > (int) DDR_NB_LOOPS (ddr))
goto next_problem;
int eq;
tree type = signed_type_for_types (TREE_TYPE (access_fun_a),
TREE_TYPE (access_fun_b));
- tree fun_a = chrec_convert (type, access_fun_a, NULL_TREE);
- tree fun_b = chrec_convert (type, access_fun_b, NULL_TREE);
+ tree fun_a = chrec_convert (type, access_fun_a, NULL);
+ tree fun_b = chrec_convert (type, access_fun_b, NULL);
tree difference = chrec_fold_minus (type, fun_a, fun_b);
/* When the fun_a - fun_b is not constant, the dependence is not
/* GCD test. */
if (DDR_NB_LOOPS (ddr) != 0 && pb->eqs[eq].coef[0]
- && !int_divides_p (lambda_vector_gcd
+ && !int_divides_p (lambda_vector_gcd
((lambda_vector) &(pb->eqs[eq].coef[1]),
2 * DDR_NB_LOOPS (ddr)),
pb->eqs[eq].coef[0]))
removed by the solver: the "dx"
- coef[nb_loops + 1, 2*nb_loops] are the loop variables: "loop_x".
*/
- for (i = 0; i <= DDR_INNER_LOOP (ddr)
+ for (i = 0; i <= DDR_INNER_LOOP (ddr)
&& VEC_iterate (loop_p, DDR_LOOP_NEST (ddr), i, loopi); i++)
{
HOST_WIDE_INT nbi = estimated_loop_iterations_int (loopi, false);
set MAYBE_DEPENDENT to true.
Example: for setting up the dependence system corresponding to the
- conflicting accesses
+ conflicting accesses
| loop_i
| loop_j
| ... A[2*j, 2*(i + j)]
| endloop_j
| endloop_i
-
+
the following constraints come from the iteration domain:
0 <= i <= Ni
}
}
- return true;
+ return true;
}
/* This computes the affine dependence relation between A and B with
respect to LOOP_NEST. CHREC_KNOWN is used for representing the
independence between two accesses, while CHREC_DONT_KNOW is used
for representing the unknown relation.
-
+
Note that it is possible to stop the computation of the dependence
relation the first time we detect a CHREC_KNOWN element for a given
subscript. */
{
struct data_reference *dra = DDR_A (ddr);
struct data_reference *drb = DDR_B (ddr);
-
+
if (dump_file && (dump_flags & TDF_DETAILS))
{
fprintf (dump_file, "(compute_affine_dependence\n");
fprintf (dump_file, " (stmt_a = \n");
- print_generic_expr (dump_file, DR_STMT (dra), 0);
+ print_gimple_stmt (dump_file, DR_STMT (dra), 0, 0);
fprintf (dump_file, ")\n (stmt_b = \n");
- print_generic_expr (dump_file, DR_STMT (drb), 0);
+ print_gimple_stmt (dump_file, DR_STMT (drb), 0, 0);
fprintf (dump_file, ")\n");
}
/* Analyze only when the dependence relation is not yet known. */
- if (DDR_ARE_DEPENDENT (ddr) == NULL_TREE)
+ if (DDR_ARE_DEPENDENT (ddr) == NULL_TREE
+ && !DDR_SELF_REFERENCE (ddr))
{
dependence_stats.num_dependence_tests++;
else
subscript_dependence_tester (ddr, loop_nest);
}
-
+
/* As a last case, if the dependence cannot be determined, or if
the dependence is considered too difficult to determine, answer
"don't know". */
finalize_ddr_dependent (ddr, chrec_dont_know);
}
}
-
+
if (dump_file && (dump_flags & TDF_DETAILS))
fprintf (dump_file, ")\n");
}
COMPUTE_SELF_AND_RR is FALSE, don't compute read-read and self
relations. */
-void
+void
compute_all_dependences (VEC (data_reference_p, heap) *datarefs,
VEC (ddr_p, heap) **dependence_relations,
VEC (loop_p, heap) *loop_nest,
{
ddr = initialize_data_dependence_relation (a, b, loop_nest);
VEC_safe_push (ddr_p, heap, *dependence_relations, ddr);
- compute_affine_dependence (ddr, VEC_index (loop_p, loop_nest, 0));
+ if (loop_nest)
+ compute_affine_dependence (ddr, VEC_index (loop_p, loop_nest, 0));
}
if (compute_self_and_rr)
true if STMT clobbers memory, false otherwise. */
bool
-get_references_in_stmt (tree stmt, VEC (data_ref_loc, heap) **references)
+get_references_in_stmt (gimple stmt, VEC (data_ref_loc, heap) **references)
{
bool clobbers_memory = false;
data_ref_loc *ref;
- tree *op0, *op1, call;
+ tree *op0, *op1;
+ enum gimple_code stmt_code = gimple_code (stmt);
*references = NULL;
/* ASM_EXPR and CALL_EXPR may embed arbitrary side effects.
Calls have side-effects, except those to const or pure
functions. */
- call = get_call_expr_in (stmt);
- if ((call
- && !(call_expr_flags (call) & (ECF_CONST | ECF_PURE)))
- || (TREE_CODE (stmt) == ASM_EXPR
- && ASM_VOLATILE_P (stmt)))
+ if ((stmt_code == GIMPLE_CALL
+ && !(gimple_call_flags (stmt) & (ECF_CONST | ECF_PURE)))
+ || (stmt_code == GIMPLE_ASM
+ && gimple_asm_volatile_p (stmt)))
clobbers_memory = true;
- if (ZERO_SSA_OPERANDS (stmt, SSA_OP_ALL_VIRTUALS))
+ if (!gimple_vuse (stmt))
return clobbers_memory;
- if (TREE_CODE (stmt) == GIMPLE_MODIFY_STMT)
+ if (stmt_code == GIMPLE_ASSIGN)
{
tree base;
- op0 = &GIMPLE_STMT_OPERAND (stmt, 0);
- op1 = &GIMPLE_STMT_OPERAND (stmt, 1);
-
+ op0 = gimple_assign_lhs_ptr (stmt);
+ op1 = gimple_assign_rhs1_ptr (stmt);
+
if (DECL_P (*op1)
|| (REFERENCE_CLASS_P (*op1)
&& (base = get_base_address (*op1))
ref->is_read = false;
}
}
-
- if (call)
+ else if (stmt_code == GIMPLE_CALL)
{
- unsigned i, n = call_expr_nargs (call);
+ unsigned i, n = gimple_call_num_args (stmt);
for (i = 0; i < n; i++)
{
- op0 = &CALL_EXPR_ARG (call, i);
+ op0 = gimple_call_arg_ptr (stmt, i);
if (DECL_P (*op0)
|| (REFERENCE_CLASS_P (*op0) && get_base_address (*op0)))
/* Stores the data references in STMT to DATAREFS. If there is an unanalyzable
reference, returns false, otherwise returns true. NEST is the outermost
- loop of the loop nest in that the references should be analyzed. */
+ loop of the loop nest in which the references should be analyzed. */
-static bool
-find_data_references_in_stmt (struct loop *nest, tree stmt,
+bool
+find_data_references_in_stmt (struct loop *nest, gimple stmt,
VEC (data_reference_p, heap) **datarefs)
{
unsigned i;
{
dr = create_data_ref (nest, *ref->pos, stmt, ref->is_read);
gcc_assert (dr != NULL);
-
- /* FIXME -- data dependence analysis does not work correctly for objects with
- invariant addresses. Let us fail here until the problem is fixed. */
- if (dr_address_invariant_p (dr))
+
+ /* FIXME -- data dependence analysis does not work correctly for objects
+ with invariant addresses in loop nests. Let us fail here until the
+ problem is fixed. */
+ if (dr_address_invariant_p (dr) && nest)
{
free_data_ref (dr);
if (dump_file && (dump_flags & TDF_DETAILS))
return ret;
}
+/* Stores the data references in STMT to DATAREFS. If there is an unanalyzable
+ reference, returns false, otherwise returns true. NEST is the outermost
+ loop of the loop nest in which the references should be analyzed. */
+
+bool
+graphite_find_data_references_in_stmt (struct loop *nest, gimple stmt,
+ VEC (data_reference_p, heap) **datarefs)
+{
+ unsigned i;
+ VEC (data_ref_loc, heap) *references;
+ data_ref_loc *ref;
+ bool ret = true;
+ data_reference_p dr;
+
+ if (get_references_in_stmt (stmt, &references))
+ {
+ VEC_free (data_ref_loc, heap, references);
+ return false;
+ }
+
+ for (i = 0; VEC_iterate (data_ref_loc, references, i, ref); i++)
+ {
+ dr = create_data_ref (nest, *ref->pos, stmt, ref->is_read);
+ gcc_assert (dr != NULL);
+ VEC_safe_push (data_reference_p, heap, *datarefs, dr);
+ }
+
+ VEC_free (data_ref_loc, heap, references);
+ return ret;
+}
+
+/* Search the data references in LOOP, and record the information into
+ DATAREFS. Returns chrec_dont_know when failing to analyze a
+ difficult case, returns NULL_TREE otherwise. */
+
+static tree
+find_data_references_in_bb (struct loop *loop, basic_block bb,
+ VEC (data_reference_p, heap) **datarefs)
+{
+ gimple_stmt_iterator bsi;
+
+ for (bsi = gsi_start_bb (bb); !gsi_end_p (bsi); gsi_next (&bsi))
+ {
+ gimple stmt = gsi_stmt (bsi);
+
+ if (!find_data_references_in_stmt (loop, stmt, datarefs))
+ {
+ struct data_reference *res;
+ res = XCNEW (struct data_reference);
+ VEC_safe_push (data_reference_p, heap, *datarefs, res);
+
+ return chrec_dont_know;
+ }
+ }
+
+ return NULL_TREE;
+}
+
/* Search the data references in LOOP, and record the information into
DATAREFS. Returns chrec_dont_know when failing to analyze a
difficult case, returns NULL_TREE otherwise.
TODO: This function should be made smarter so that it can handle address
arithmetic as if they were array accesses, etc. */
-static tree
+tree
find_data_references_in_loop (struct loop *loop,
VEC (data_reference_p, heap) **datarefs)
{
basic_block bb, *bbs;
unsigned int i;
- block_stmt_iterator bsi;
bbs = get_loop_body_in_dom_order (loop);
{
bb = bbs[i];
- for (bsi = bsi_start (bb); !bsi_end_p (bsi); bsi_next (&bsi))
- {
- tree stmt = bsi_stmt (bsi);
-
- if (!find_data_references_in_stmt (loop, stmt, datarefs))
- {
- struct data_reference *res;
- res = XCNEW (struct data_reference);
- VEC_safe_push (data_reference_p, heap, *datarefs, res);
-
- free (bbs);
- return chrec_dont_know;
- }
- }
+ if (find_data_references_in_bb (loop, bb, datarefs) == chrec_dont_know)
+ {
+ free (bbs);
+ return chrec_dont_know;
+ }
}
free (bbs);
return true;
}
-/* Given a loop nest LOOP, the following vectors are returned:
- DATAREFS is initialized to all the array elements contained in this loop,
- DEPENDENCE_RELATIONS contains the relations between the data references.
- Compute read-read and self relations if
+/* Returns true when the data dependences have been computed, false otherwise.
+ Given a loop nest LOOP, the following vectors are returned:
+ DATAREFS is initialized to all the array elements contained in this loop,
+ DEPENDENCE_RELATIONS contains the relations between the data references.
+ Compute read-read and self relations if
COMPUTE_SELF_AND_READ_READ_DEPENDENCES is TRUE. */
-void
-compute_data_dependences_for_loop (struct loop *loop,
+bool
+compute_data_dependences_for_loop (struct loop *loop,
bool compute_self_and_read_read_dependences,
VEC (data_reference_p, heap) **datarefs,
VEC (ddr_p, heap) **dependence_relations)
{
+ bool res = true;
VEC (loop_p, heap) *vloops = VEC_alloc (loop_p, heap, 3);
memset (&dependence_stats, 0, sizeof (dependence_stats));
- /* If the loop nest is not well formed, or one of the data references
+ /* If the loop nest is not well formed, or one of the data references
is not computable, give up without spending time to compute other
dependences. */
if (!loop
chrec_dont_know. */
ddr = initialize_data_dependence_relation (NULL, NULL, vloops);
VEC_safe_push (ddr_p, heap, *dependence_relations, ddr);
+ res = false;
}
else
compute_all_dependences (*datarefs, dependence_relations, vloops,
{
fprintf (dump_file, "Dependence tester statistics:\n");
- fprintf (dump_file, "Number of dependence tests: %d\n",
+ fprintf (dump_file, "Number of dependence tests: %d\n",
dependence_stats.num_dependence_tests);
- fprintf (dump_file, "Number of dependence tests classified dependent: %d\n",
+ fprintf (dump_file, "Number of dependence tests classified dependent: %d\n",
dependence_stats.num_dependence_dependent);
- fprintf (dump_file, "Number of dependence tests classified independent: %d\n",
+ fprintf (dump_file, "Number of dependence tests classified independent: %d\n",
dependence_stats.num_dependence_independent);
- fprintf (dump_file, "Number of undetermined dependence tests: %d\n",
+ fprintf (dump_file, "Number of undetermined dependence tests: %d\n",
dependence_stats.num_dependence_undetermined);
- fprintf (dump_file, "Number of subscript tests: %d\n",
+ fprintf (dump_file, "Number of subscript tests: %d\n",
dependence_stats.num_subscript_tests);
- fprintf (dump_file, "Number of undetermined subscript tests: %d\n",
+ fprintf (dump_file, "Number of undetermined subscript tests: %d\n",
dependence_stats.num_subscript_undetermined);
- fprintf (dump_file, "Number of same subscript function: %d\n",
+ fprintf (dump_file, "Number of same subscript function: %d\n",
dependence_stats.num_same_subscript_function);
fprintf (dump_file, "Number of ziv tests: %d\n",
fprintf (dump_file, "Number of ziv tests returning independent: %d\n",
dependence_stats.num_ziv_independent);
fprintf (dump_file, "Number of ziv tests unimplemented: %d\n",
- dependence_stats.num_ziv_unimplemented);
+ dependence_stats.num_ziv_unimplemented);
- fprintf (dump_file, "Number of siv tests: %d\n",
+ fprintf (dump_file, "Number of siv tests: %d\n",
dependence_stats.num_siv);
fprintf (dump_file, "Number of siv tests returning dependent: %d\n",
dependence_stats.num_siv_dependent);
fprintf (dump_file, "Number of siv tests unimplemented: %d\n",
dependence_stats.num_siv_unimplemented);
- fprintf (dump_file, "Number of miv tests: %d\n",
+ fprintf (dump_file, "Number of miv tests: %d\n",
dependence_stats.num_miv);
fprintf (dump_file, "Number of miv tests returning dependent: %d\n",
dependence_stats.num_miv_dependent);
dependence_stats.num_miv_independent);
fprintf (dump_file, "Number of miv tests unimplemented: %d\n",
dependence_stats.num_miv_unimplemented);
- }
+ }
+
+ return res;
+}
+
+/* Returns true when the data dependences for the basic block BB have been
+ computed, false otherwise.
+ DATAREFS is initialized to all the array elements contained in this basic
+ block, DEPENDENCE_RELATIONS contains the relations between the data
+ references. Compute read-read and self relations if
+ COMPUTE_SELF_AND_READ_READ_DEPENDENCES is TRUE. */
+bool
+compute_data_dependences_for_bb (basic_block bb,
+ bool compute_self_and_read_read_dependences,
+ VEC (data_reference_p, heap) **datarefs,
+ VEC (ddr_p, heap) **dependence_relations)
+{
+ if (find_data_references_in_bb (NULL, bb, datarefs) == chrec_dont_know)
+ return false;
+
+ compute_all_dependences (*datarefs, dependence_relations, NULL,
+ compute_self_and_read_read_dependences);
+ return true;
}
/* Entry point (for testing only). Analyze all the data references
and the dependence relations in LOOP.
- The data references are computed first.
-
+ The data references are computed first.
+
A relation on these nodes is represented by a complete graph. Some
of the relations could be of no interest, thus the relations can be
computed on demand.
-
+
In the following function we compute all the relations. This is
just a first implementation that is here for:
- - for showing how to ask for the dependence relations,
+ - for showing how to ask for the dependence relations,
- for the debugging the whole dependence graph,
- for the dejagnu testcases and maintenance.
-
+
It is possible to ask only for a part of the graph, avoiding to
compute the whole dependence graph. The computed dependences are
stored in a knowledge base (KB) such that later queries don't
recompute the same information. The implementation of this KB is
transparent to the optimizer, and thus the KB can be changed with a
more efficient implementation, or the KB could be disabled. */
-static void
+static void
analyze_all_data_dependences (struct loop *loop)
{
unsigned int i;
int nb_data_refs = 10;
- VEC (data_reference_p, heap) *datarefs =
+ VEC (data_reference_p, heap) *datarefs =
VEC_alloc (data_reference_p, heap, nb_data_refs);
- VEC (ddr_p, heap) *dependence_relations =
+ VEC (ddr_p, heap) *dependence_relations =
VEC_alloc (ddr_p, heap, nb_data_refs * nb_data_refs);
/* Compute DDs on the whole function. */
{
unsigned nb_top_relations = 0;
unsigned nb_bot_relations = 0;
- unsigned nb_basename_differ = 0;
unsigned nb_chrec_relations = 0;
struct data_dependence_relation *ddr;
{
if (chrec_contains_undetermined (DDR_ARE_DEPENDENT (ddr)))
nb_top_relations++;
-
+
else if (DDR_ARE_DEPENDENT (ddr) == chrec_known)
- {
- struct data_reference *a = DDR_A (ddr);
- struct data_reference *b = DDR_B (ddr);
+ nb_bot_relations++;
- if (!bitmap_intersect_p (DR_VOPS (a), DR_VOPS (b)))
- nb_basename_differ++;
- else
- nb_bot_relations++;
- }
-
- else
+ else
nb_chrec_relations++;
}
-
+
gather_stats_on_scev_database ();
}
}
/* Free the memory used by the data dependence relations from
DEPENDENCE_RELATIONS. */
-void
+void
free_dependence_relations (VEC (ddr_p, heap) *dependence_relations)
{
unsigned int i;
struct vertex *v = &(rdg->vertices[i]);
struct graph_edge *e;
- fprintf (file, "(vertex %d: (%s%s) (in:", i,
+ fprintf (file, "(vertex %d: (%s%s) (in:", i,
RDG_MEM_WRITE_STMT (rdg, i) ? "w" : "",
RDG_MEM_READS_STMT (rdg, i) ? "r" : "");
fprintf (file, " %d", e->dest);
fprintf (file, ") \n");
- print_generic_stmt (file, RDGV_STMT (v), TDF_VOPS|TDF_MEMSYMS);
+ print_gimple_stmt (file, RDGV_STMT (v), 0, TDF_VOPS|TDF_MEMSYMS);
fprintf (file, ")\n");
}
dump_rdg (stderr, rdg);
}
-static void
-dot_rdg_1 (FILE *file, struct graph *rdg)
-{
- int i;
-
- fprintf (file, "digraph RDG {\n");
-
- for (i = 0; i < rdg->n_vertices; i++)
- {
- struct vertex *v = &(rdg->vertices[i]);
- struct graph_edge *e;
-
- /* Highlight reads from memory. */
- if (RDG_MEM_READS_STMT (rdg, i))
- fprintf (file, "%d [style=filled, fillcolor=green]\n", i);
-
- /* Highlight stores to memory. */
- if (RDG_MEM_WRITE_STMT (rdg, i))
- fprintf (file, "%d [style=filled, fillcolor=red]\n", i);
-
- if (v->succ)
- for (e = v->succ; e; e = e->succ_next)
- switch (RDGE_TYPE (e))
- {
- case input_dd:
- fprintf (file, "%d -> %d [label=input] \n", i, e->dest);
- break;
-
- case output_dd:
- fprintf (file, "%d -> %d [label=output] \n", i, e->dest);
- break;
-
- case flow_dd:
- /* These are the most common dependences: don't print these. */
- fprintf (file, "%d -> %d \n", i, e->dest);
- break;
-
- case anti_dd:
- fprintf (file, "%d -> %d [label=anti] \n", i, e->dest);
- break;
-
- default:
- gcc_unreachable ();
- }
- }
-
- fprintf (file, "}\n\n");
-}
-
-/* Display SCOP using dotty. */
-
-void
-dot_rdg (struct graph *rdg)
-{
- FILE *file = fopen ("/tmp/rdg.dot", "w");
- gcc_assert (file != NULL);
-
- dot_rdg_1 (file, rdg);
- fclose (file);
-
- system ("dotty /tmp/rdg.dot");
-}
-
-
/* This structure is used for recording the mapping statement index in
the RDG. */
-struct rdg_vertex_info GTY(())
+struct GTY(()) rdg_vertex_info
{
- tree stmt;
+ gimple stmt;
int index;
};
/* Returns the index of STMT in RDG. */
int
-rdg_vertex_for_stmt (struct graph *rdg, tree stmt)
+rdg_vertex_for_stmt (struct graph *rdg, gimple stmt)
{
struct rdg_vertex_info rvi, *slot;
e->data = XNEW (struct rdg_edge);
RDGE_LEVEL (e) = level;
+ RDGE_RELATION (e) = ddr;
/* Determines the type of the data dependence. */
if (DR_IS_READ (dra) && DR_IS_READ (drb))
{
use_operand_p imm_use_p;
imm_use_iterator iterator;
-
+
FOR_EACH_IMM_USE_FAST (imm_use_p, iterator, def)
{
struct graph_edge *e;
e = add_edge (rdg, idef, use);
e->data = XNEW (struct rdg_edge);
RDGE_TYPE (e) = flow_dd;
+ RDGE_RELATION (e) = NULL;
}
}
/* Build the vertices of the reduced dependence graph RDG. */
-static void
-create_rdg_vertices (struct graph *rdg, VEC (tree, heap) *stmts)
+void
+create_rdg_vertices (struct graph *rdg, VEC (gimple, heap) *stmts)
{
int i, j;
- tree stmt;
+ gimple stmt;
- for (i = 0; VEC_iterate (tree, stmts, i, stmt); i++)
+ for (i = 0; VEC_iterate (gimple, stmts, i, stmt); i++)
{
VEC (data_ref_loc, heap) *references;
data_ref_loc *ref;
RDG_MEM_WRITE_STMT (rdg, i) = false;
RDG_MEM_READS_STMT (rdg, i) = false;
- if (TREE_CODE (stmt) == PHI_NODE)
+ if (gimple_code (stmt) == GIMPLE_PHI)
continue;
get_references_in_stmt (stmt, &references);
identifying statements. */
static void
-stmts_from_loop (struct loop *loop, VEC (tree, heap) **stmts)
+stmts_from_loop (struct loop *loop, VEC (gimple, heap) **stmts)
{
unsigned int i;
basic_block *bbs = get_loop_body_in_dom_order (loop);
for (i = 0; i < loop->num_nodes; i++)
{
- tree phi, stmt;
basic_block bb = bbs[i];
- block_stmt_iterator bsi;
+ gimple_stmt_iterator bsi;
+ gimple stmt;
- for (phi = phi_nodes (bb); phi; phi = PHI_CHAIN (phi))
- VEC_safe_push (tree, heap, *stmts, phi);
+ for (bsi = gsi_start_phis (bb); !gsi_end_p (bsi); gsi_next (&bsi))
+ VEC_safe_push (gimple, heap, *stmts, gsi_stmt (bsi));
- for (bsi = bsi_start (bb); !bsi_end_p (bsi); bsi_next (&bsi))
- if (TREE_CODE (stmt = bsi_stmt (bsi)) != LABEL_EXPR)
- VEC_safe_push (tree, heap, *stmts, stmt);
+ for (bsi = gsi_start_bb (bb); !gsi_end_p (bsi); gsi_next (&bsi))
+ {
+ stmt = gsi_stmt (bsi);
+ if (gimple_code (stmt) != GIMPLE_LABEL)
+ VEC_safe_push (gimple, heap, *stmts, stmt);
+ }
}
free (bbs);
for (i = 0; VEC_iterate (ddr_p, dependence_relations, i, ddr); i++)
if (DDR_ARE_DEPENDENT (ddr) == chrec_dont_know)
return false;
-
+
return true;
}
static hashval_t
hash_stmt_vertex_info (const void *elt)
{
- struct rdg_vertex_info *rvi = (struct rdg_vertex_info *) elt;
- tree stmt = rvi->stmt;
+ const struct rdg_vertex_info *const rvi =
+ (const struct rdg_vertex_info *) elt;
+ gimple stmt = rvi->stmt;
return htab_hash_pointer (stmt);
}
scalar dependence. */
struct graph *
+build_empty_rdg (int n_stmts)
+{
+ int nb_data_refs = 10;
+ struct graph *rdg = new_graph (n_stmts);
+
+ rdg->indices = htab_create (nb_data_refs, hash_stmt_vertex_info,
+ eq_stmt_vertex_info, hash_stmt_vertex_del);
+ return rdg;
+}
+
+/* Build the Reduced Dependence Graph (RDG) with one vertex per
+ statement of the loop nest, and one edge per data dependence or
+ scalar dependence. */
+
+struct graph *
build_rdg (struct loop *loop)
{
int nb_data_refs = 10;
struct graph *rdg = NULL;
VEC (ddr_p, heap) *dependence_relations;
VEC (data_reference_p, heap) *datarefs;
- VEC (tree, heap) *stmts = VEC_alloc (tree, heap, nb_data_refs);
-
+ VEC (gimple, heap) *stmts = VEC_alloc (gimple, heap, nb_data_refs);
+
dependence_relations = VEC_alloc (ddr_p, heap, nb_data_refs * nb_data_refs) ;
datarefs = VEC_alloc (data_reference_p, heap, nb_data_refs);
- compute_data_dependences_for_loop (loop,
+ compute_data_dependences_for_loop (loop,
false,
&datarefs,
&dependence_relations);
if (!known_dependences_p (dependence_relations))
- goto end_rdg;
+ {
+ free_dependence_relations (dependence_relations);
+ free_data_refs (datarefs);
+ VEC_free (gimple, heap, stmts);
+
+ return rdg;
+ }
stmts_from_loop (loop, &stmts);
- rdg = new_graph (VEC_length (tree, stmts));
+ rdg = build_empty_rdg (VEC_length (gimple, stmts));
rdg->indices = htab_create (nb_data_refs, hash_stmt_vertex_info,
eq_stmt_vertex_info, hash_stmt_vertex_del);
create_rdg_vertices (rdg, stmts);
create_rdg_edges (rdg, dependence_relations);
- end_rdg:
- free_dependence_relations (dependence_relations);
- free_data_refs (datarefs);
- VEC_free (tree, heap, stmts);
-
+ VEC_free (gimple, heap, stmts);
return rdg;
}
store to memory. */
void
-stores_from_loop (struct loop *loop, VEC (tree, heap) **stmts)
+stores_from_loop (struct loop *loop, VEC (gimple, heap) **stmts)
{
unsigned int i;
basic_block *bbs = get_loop_body_in_dom_order (loop);
for (i = 0; i < loop->num_nodes; i++)
{
basic_block bb = bbs[i];
- block_stmt_iterator bsi;
+ gimple_stmt_iterator bsi;
- for (bsi = bsi_start (bb); !bsi_end_p (bsi); bsi_next (&bsi))
- if (!ZERO_SSA_OPERANDS (bsi_stmt (bsi), SSA_OP_VDEF))
- VEC_safe_push (tree, heap, *stmts, bsi_stmt (bsi));
+ for (bsi = gsi_start_bb (bb); !gsi_end_p (bsi); gsi_next (&bsi))
+ if (gimple_vdef (gsi_stmt (bsi)))
+ VEC_safe_push (gimple, heap, *stmts, gsi_stmt (bsi));
}
free (bbs);
address or NULL_TREE if the base is not determined. */
static inline tree
-ref_base_address (tree stmt, data_ref_loc *ref)
+ref_base_address (gimple stmt, data_ref_loc *ref)
{
tree base = NULL_TREE;
tree base_address;
bool
rdg_defs_used_in_other_loops_p (struct graph *rdg, int v)
{
- tree stmt = RDG_STMT (rdg, v);
+ gimple stmt = RDG_STMT (rdg, v);
struct loop *loop = loop_containing_stmt (stmt);
use_operand_p imm_use_p;
imm_use_iterator iterator;
ref_base_address is the same. */
bool
-have_similar_memory_accesses (tree s1, tree s2)
+have_similar_memory_accesses (gimple s1, gimple s2)
{
bool res = false;
unsigned i, j;
static int
have_similar_memory_accesses_1 (const void *s1, const void *s2)
{
- return have_similar_memory_accesses ((tree) s1, (tree) s2);
+ return have_similar_memory_accesses (CONST_CAST_GIMPLE ((const_gimple) s1),
+ CONST_CAST_GIMPLE ((const_gimple) s2));
}
/* Helper function for the hashtab. */
static hashval_t
ref_base_address_1 (const void *s)
{
- tree stmt = (tree) s;
+ gimple stmt = CONST_CAST_GIMPLE ((const_gimple) s);
unsigned i;
VEC (data_ref_loc, heap) *refs;
data_ref_loc *ref;
/* Try to remove duplicated write data references from STMTS. */
void
-remove_similar_memory_refs (VEC (tree, heap) **stmts)
+remove_similar_memory_refs (VEC (gimple, heap) **stmts)
{
unsigned i;
- tree stmt;
- htab_t seen = htab_create (VEC_length (tree, *stmts), ref_base_address_1,
+ gimple stmt;
+ htab_t seen = htab_create (VEC_length (gimple, *stmts), ref_base_address_1,
have_similar_memory_accesses_1, NULL);
- for (i = 0; VEC_iterate (tree, *stmts, i, stmt); )
+ for (i = 0; VEC_iterate (gimple, *stmts, i, stmt); )
{
void **slot;
slot = htab_find_slot (seen, stmt, INSERT);
if (*slot)
- VEC_ordered_remove (tree, *stmts, i);
+ VEC_ordered_remove (gimple, *stmts, i);
else
{
*slot = (void *) stmt;
htab_delete (seen);
}
+/* Returns the index of PARAMETER in the parameters vector of the
+ ACCESS_MATRIX. If PARAMETER does not exist return -1. */
+
+int
+access_matrix_get_index_for_parameter (tree parameter,
+ struct access_matrix *access_matrix)
+{
+ int i;
+ VEC (tree,heap) *lambda_parameters = AM_PARAMETERS (access_matrix);
+ tree lambda_parameter;
+
+ for (i = 0; VEC_iterate (tree, lambda_parameters, i, lambda_parameter); i++)
+ if (lambda_parameter == parameter)
+ return i + AM_NB_INDUCTION_VARS (access_matrix);
+
+ return -1;
+}