#include "config.h"
#include "system.h"
#include "coretypes.h"
-#include "tm.h"
-#include "ggc.h"
-#include "flags.h"
-#include "tree.h"
-#include "basic-block.h"
-#include "tree-pretty-print.h"
#include "gimple-pretty-print.h"
#include "tree-flow.h"
-#include "tree-dump.h"
-#include "timevar.h"
#include "cfgloop.h"
#include "tree-data-ref.h"
#include "tree-scalar-evolution.h"
{
unsigned int i;
- fprintf (outf, "#(Data Ref: \n# stmt: ");
+ 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, 0);
fprintf (outf, "# ref: ");
print_generic_stmt (outf, DR_REF (dr), 0);
print_direction_vector (outf, v, length);
}
+/* Print out a vector VEC of length N to OUTFILE. */
+
+static inline void
+print_lambda_vector (FILE * outfile, lambda_vector vector, int n)
+{
+ int i;
+
+ for (i = 0; i < n; i++)
+ fprintf (outfile, "%3d ", vector[i]);
+ fprintf (outfile, "\n");
+}
+
/* Print a vector of distance vectors. */
void
*off = ssize_int (0);
STRIP_NOPS (exp);
- if (automatically_generated_chrec_p (exp))
+ if (tree_is_chrec (exp))
return;
otype = TREE_TYPE (exp);
}
/* Determines the base object and the list of indices of memory reference
- DR, analyzed in loop nest NEST. */
+ DR, analyzed in LOOP and instantiated in loop nest NEST. */
static void
-dr_analyze_indices (struct data_reference *dr, struct loop *nest)
+dr_analyze_indices (struct data_reference *dr, loop_p nest, loop_p loop)
{
- 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 = NULL_TREE;
/* Analyzes memory reference MEMREF accessed in STMT. The reference
is read if IS_READ is true, write otherwise. Returns the
- data_reference description of MEMREF. NEST is the outermost loop of the
- loop nest in that the reference should be analyzed. */
+ 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 (struct loop *nest, tree memref, gimple stmt, bool is_read)
+create_data_ref (loop_p nest, loop_p loop, tree memref, gimple stmt,
+ bool is_read)
{
struct data_reference *dr;
DR_IS_READ (dr) = is_read;
dr_analyze_innermost (dr);
- dr_analyze_indices (dr, nest);
+ dr_analyze_indices (dr, nest, loop);
dr_analyze_alias (dr);
if (dump_file && (dump_flags & TDF_DETAILS))
return dr;
}
+/* 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_fn_free (overlaps_b_xyz);
}
+/* Copy the elements of vector VEC1 with length SIZE to VEC2. */
+
+static void
+lambda_vector_copy (lambda_vector vec1, lambda_vector vec2,
+ int size)
+{
+ memcpy (vec2, vec1, size * sizeof (*vec1));
+}
+
+/* Copy the elements of M x N matrix MAT1 to MAT2. */
+
+static void
+lambda_matrix_copy (lambda_matrix mat1, lambda_matrix mat2,
+ int m, int n)
+{
+ int i;
+
+ for (i = 0; i < m; i++)
+ lambda_vector_copy (mat1[i], mat2[i], n);
+}
+
+/* Store the N x N identity matrix in MAT. */
+
+static void
+lambda_matrix_id (lambda_matrix mat, int size)
+{
+ int i, j;
+
+ for (i = 0; i < size; i++)
+ for (j = 0; j < size; j++)
+ mat[i][j] = (i == j) ? 1 : 0;
+}
+
+/* Return the first nonzero element of vector VEC1 between START and N.
+ We must have START <= N. Returns N if VEC1 is the zero vector. */
+
+static int
+lambda_vector_first_nz (lambda_vector vec1, int n, int start)
+{
+ int j = start;
+ while (j < n && vec1[j] == 0)
+ j++;
+ return j;
+}
+
+/* Add a multiple of row R1 of matrix MAT with N columns to row R2:
+ R2 = R2 + CONST1 * R1. */
+
+static void
+lambda_matrix_row_add (lambda_matrix mat, int n, int r1, int r2, int const1)
+{
+ int i;
+
+ if (const1 == 0)
+ return;
+
+ for (i = 0; i < n; i++)
+ mat[r2][i] += const1 * mat[r1][i];
+}
+
+/* Swap rows R1 and R2 in matrix MAT. */
+
+static void
+lambda_matrix_row_exchange (lambda_matrix mat, int r1, int r2)
+{
+ lambda_vector row;
+
+ row = mat[r1];
+ mat[r1] = mat[r2];
+ mat[r2] = row;
+}
+
+/* Multiply vector VEC1 of length SIZE by a constant CONST1,
+ and store the result in VEC2. */
+
+static void
+lambda_vector_mult_const (lambda_vector vec1, lambda_vector vec2,
+ int size, int const1)
+{
+ int i;
+
+ if (const1 == 0)
+ lambda_vector_clear (vec2, size);
+ else
+ for (i = 0; i < size; i++)
+ vec2[i] = const1 * vec1[i];
+}
+
+/* Negate vector VEC1 with length SIZE and store it in VEC2. */
+
+static void
+lambda_vector_negate (lambda_vector vec1, lambda_vector vec2,
+ int size)
+{
+ lambda_vector_mult_const (vec1, vec2, size, -1);
+}
+
+/* Negate row R1 of matrix MAT which has N columns. */
+
+static void
+lambda_matrix_row_negate (lambda_matrix mat, int n, int r1)
+{
+ lambda_vector_negate (mat[r1], mat[r1], n);
+}
+
+/* Return true if two vectors are equal. */
+
+static bool
+lambda_vector_equal (lambda_vector vec1, lambda_vector vec2, int size)
+{
+ int i;
+ for (i = 0; i < size; i++)
+ if (vec1[i] != vec2[i])
+ return false;
+ return true;
+}
+
+/* Given an M x N integer matrix A, this function determines an M x
+ M unimodular matrix U, and an M x N echelon matrix S such that
+ "U.A = S". This decomposition is also known as "right Hermite".
+
+ Ref: Algorithm 2.1 page 33 in "Loop Transformations for
+ Restructuring Compilers" Utpal Banerjee. */
+
+static void
+lambda_matrix_right_hermite (lambda_matrix A, int m, int n,
+ lambda_matrix S, lambda_matrix U)
+{
+ int i, j, i0 = 0;
+
+ lambda_matrix_copy (A, S, m, n);
+ lambda_matrix_id (U, m);
+
+ for (j = 0; j < n; j++)
+ {
+ if (lambda_vector_first_nz (S[j], m, i0) < m)
+ {
+ ++i0;
+ for (i = m - 1; i >= i0; i--)
+ {
+ while (S[i][j] != 0)
+ {
+ int sigma, factor, a, b;
+
+ a = S[i-1][j];
+ b = S[i][j];
+ sigma = (a * b < 0) ? -1: 1;
+ a = abs (a);
+ b = abs (b);
+ factor = sigma * (a / b);
+
+ lambda_matrix_row_add (S, n, i, i-1, -factor);
+ lambda_matrix_row_exchange (S, i, i-1);
+
+ lambda_matrix_row_add (U, m, i, i-1, -factor);
+ lambda_matrix_row_exchange (U, i, i-1);
+ }
+ }
+ }
+ }
+}
+
/* Determines the overlapping elements due to accesses CHREC_A and
CHREC_B, that are affine functions. This function cannot handle
symbolic evolution functions, ie. when initial conditions are
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]).
-
- 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).
- */
tree type, difference;
dependence_stats.num_miv++;
/* 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))
+ && (evolution_function_is_affine_multivariate_p (chrec_a, lnn)
+ || operand_equal_p (chrec_a, chrec_b, 0)))
{
dependence_stats.num_same_subscript_function++;
*overlap_iterations_a = conflict_fn (1, affine_fn_cst (integer_zero_node));
}
/* If they aren't the same, and aren't affine, we can't do anything
- yet. */
+ yet. */
else if ((chrec_contains_symbols (chrec_a)
|| chrec_contains_symbols (chrec_b))
&& (!evolution_function_is_affine_multivariate_p (chrec_a, lnn)
&& TREE_CODE (access_fn_b) == POLYNOMIAL_CHREC)
{
int dist, index;
- int index_a = index_in_loop_nest (CHREC_VARIABLE (access_fn_a),
- DDR_LOOP_NEST (ddr));
- int index_b = index_in_loop_nest (CHREC_VARIABLE (access_fn_b),
- DDR_LOOP_NEST (ddr));
-
- /* The dependence is carried by the outermost loop. Example:
- | loop_1
- | A[{4, +, 1}_1]
- | loop_2
- | A[{5, +, 1}_2]
- | endloop_2
- | endloop_1
- In this case, the dependence is carried by loop_1. */
- index = index_a < index_b ? index_a : index_b;
- *index_carry = MIN (index, *index_carry);
+ int var_a = CHREC_VARIABLE (access_fn_a);
+ int var_b = CHREC_VARIABLE (access_fn_b);
- if (chrec_contains_undetermined (SUB_DISTANCE (subscript)))
+ if (var_a != var_b
+ || chrec_contains_undetermined (SUB_DISTANCE (subscript)))
{
non_affine_dependence_relation (ddr);
return false;
}
dist = int_cst_value (SUB_DISTANCE (subscript));
+ index = index_in_loop_nest (var_a, DDR_LOOP_NEST (ddr));
+ *index_carry = MIN (index, *index_carry);
/* This is the subscript coupling test. If we have already
recorded a distance for this loop (a distance coming from
FOR_EACH_VEC_ELT (data_ref_loc, references, i, ref)
{
- dr = create_data_ref (nest, *ref->pos, stmt, ref->is_read);
+ dr = create_data_ref (nest, loop_containing_stmt (stmt),
+ *ref->pos, stmt, ref->is_read);
gcc_assert (dr != NULL);
/* FIXME -- data dependence analysis does not work correctly for objects
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. */
+/* 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 instantiated, LOOP is the loop in which the references
+ should be analyzed. */
bool
-graphite_find_data_references_in_stmt (struct loop *nest, gimple stmt,
+graphite_find_data_references_in_stmt (loop_p nest, loop_p loop, gimple stmt,
VEC (data_reference_p, heap) **datarefs)
{
unsigned i;
FOR_EACH_VEC_ELT (data_ref_loc, references, i, ref)
{
- dr = create_data_ref (nest, *ref->pos, stmt, ref->is_read);
+ dr = create_data_ref (nest, loop, *ref->pos, stmt, ref->is_read);
gcc_assert (dr != NULL);
VEC_safe_push (data_reference_p, heap, *datarefs, dr);
}
DATAREFS. Returns chrec_dont_know when failing to analyze a
difficult case, returns NULL_TREE otherwise. */
-static tree
+tree
find_data_references_in_bb (struct loop *loop, basic_block bb,
VEC (data_reference_p, heap) **datarefs)
{
bool
compute_data_dependences_for_loop (struct loop *loop,
bool compute_self_and_read_read_dependences,
+ VEC (loop_p, heap) **loop_nest,
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));
is not computable, give up without spending time to compute other
dependences. */
if (!loop
- || !find_loop_nest (loop, &vloops)
+ || !find_loop_nest (loop, loop_nest)
|| find_data_references_in_loop (loop, datarefs) == chrec_dont_know)
{
struct data_dependence_relation *ddr;
/* Insert a single relation into dependence_relations:
chrec_dont_know. */
- ddr = initialize_data_dependence_relation (NULL, NULL, vloops);
+ ddr = initialize_data_dependence_relation (NULL, NULL, *loop_nest);
VEC_safe_push (ddr_p, heap, *dependence_relations, ddr);
res = false;
}
else
- compute_all_dependences (*datarefs, dependence_relations, vloops,
+ compute_all_dependences (*datarefs, dependence_relations, *loop_nest,
compute_self_and_read_read_dependences);
if (dump_file && (dump_flags & TDF_STATS))
VEC_alloc (data_reference_p, heap, nb_data_refs);
VEC (ddr_p, heap) *dependence_relations =
VEC_alloc (ddr_p, heap, nb_data_refs * nb_data_refs);
+ VEC (loop_p, heap) *loop_nest = VEC_alloc (loop_p, heap, 3);
/* Compute DDs on the whole function. */
- compute_data_dependences_for_loop (loop, false, &datarefs,
+ compute_data_dependences_for_loop (loop, false, &loop_nest, &datarefs,
&dependence_relations);
if (dump_file)
}
}
+ VEC_free (loop_p, heap, loop_nest);
free_dependence_relations (dependence_relations);
free_data_refs (datarefs);
}
{
unsigned int i;
struct data_dependence_relation *ddr;
- VEC (loop_p, heap) *loop_nest = NULL;
FOR_EACH_VEC_ELT (ddr_p, dependence_relations, i, ddr)
- {
- if (ddr == NULL)
- continue;
- if (loop_nest == NULL)
- loop_nest = DDR_LOOP_NEST (ddr);
- else
- gcc_assert (DDR_LOOP_NEST (ddr) == NULL
- || DDR_LOOP_NEST (ddr) == loop_nest);
+ if (ddr)
free_dependence_relation (ddr);
- }
- if (loop_nest)
- VEC_free (loop_p, heap, loop_nest);
VEC_free (ddr_p, heap, dependence_relations);
}
for (e = v->succ; e; e = e->succ_next)
fprintf (file, " %d", e->dest);
- fprintf (file, ") \n");
+ fprintf (file, ")\n");
print_gimple_stmt (file, RDGV_STMT (v), 0, TDF_VOPS|TDF_MEMSYMS);
fprintf (file, ")\n");
}
scalar dependence. */
struct graph *
-build_rdg (struct loop *loop)
+build_rdg (struct loop *loop,
+ VEC (loop_p, heap) **loop_nest,
+ VEC (ddr_p, heap) **dependence_relations,
+ VEC (data_reference_p, heap) **datarefs)
{
- int nb_data_refs = 10;
struct graph *rdg = NULL;
- VEC (ddr_p, heap) *dependence_relations;
- VEC (data_reference_p, heap) *datarefs;
- 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,
- false,
- &datarefs,
- &dependence_relations);
-
- if (!known_dependences_p (dependence_relations))
- {
- free_dependence_relations (dependence_relations);
- free_data_refs (datarefs);
- VEC_free (gimple, heap, stmts);
-
- return rdg;
- }
+ VEC (gimple, heap) *stmts = VEC_alloc (gimple, heap, 10);
- stmts_from_loop (loop, &stmts);
- rdg = build_empty_rdg (VEC_length (gimple, stmts));
+ compute_data_dependences_for_loop (loop, false, loop_nest, datarefs,
+ dependence_relations);
- 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);
+ if (known_dependences_p (*dependence_relations))
+ {
+ stmts_from_loop (loop, &stmts);
+ rdg = build_empty_rdg (VEC_length (gimple, stmts));
+ create_rdg_vertices (rdg, stmts);
+ create_rdg_edges (rdg, *dependence_relations);
+ }
VEC_free (gimple, heap, stmts);
return rdg;
int i;
for (i = 0; i < rdg->n_vertices; i++)
- free (rdg->vertices[i].data);
+ {
+ struct vertex *v = &(rdg->vertices[i]);
+ struct graph_edge *e;
+
+ for (e = v->succ; e; e = e->succ_next)
+ if (e->data)
+ free (e->data);
+
+ if (v->data)
+ free (v->data);
+ }
htab_delete (rdg->indices);
free_graph (rdg);
free (bbs);
}
-/* Returns true when STMT is an assignment that contains a data
- reference on its LHS with a stride of the same size as its unit
- type. */
+/* Returns true when the statement at STMT is of the form "A[i] = 0"
+ that contains a data reference on its LHS with a stride of the same
+ size as its unit type. */
-static bool
-mem_write_stride_of_same_size_as_unit_type_p (gimple stmt)
+bool
+stmt_with_adjacent_zero_store_dr_p (gimple stmt)
{
- struct data_reference *dr = XCNEW (struct data_reference);
- tree op0 = gimple_assign_lhs (stmt);
+ tree op0, op1;
bool res;
+ struct data_reference *dr;
+
+ if (!stmt
+ || !gimple_vdef (stmt)
+ || !is_gimple_assign (stmt)
+ || !gimple_assign_single_p (stmt)
+ || !(op1 = gimple_assign_rhs1 (stmt))
+ || !(integer_zerop (op1) || real_zerop (op1)))
+ return false;
+
+ dr = XCNEW (struct data_reference);
+ op0 = gimple_assign_lhs (stmt);
DR_STMT (dr) = stmt;
DR_REF (dr) = op0;
basic_block bb;
gimple_stmt_iterator si;
gimple stmt;
- tree op;
basic_block *bbs = get_loop_body_in_dom_order (loop);
for (i = 0; i < loop->num_nodes; i++)
for (bb = bbs[i], si = gsi_start_bb (bb); !gsi_end_p (si); gsi_next (&si))
if ((stmt = gsi_stmt (si))
- && gimple_vdef (stmt)
- && is_gimple_assign (stmt)
- && gimple_assign_rhs_code (stmt) == INTEGER_CST
- && (op = gimple_assign_rhs1 (stmt))
- && (integer_zerop (op) || real_zerop (op))
- && mem_write_stride_of_same_size_as_unit_type_p (stmt))
+ && stmt_with_adjacent_zero_store_dr_p (stmt))
VEC_safe_push (gimple, heap, *stmts, gsi_stmt (si));
free (bbs);