static void
generate_memset_zero (gimple stmt, tree op0, tree nb_iter,
gimple_stmt_iterator bsi)
{
tree addr_base, nb_bytes;
bool res = false;
gimple_seq stmt_list = NULL, stmts;
gimple fn_call;
tree mem, fn;
struct data_reference *dr = XCNEW (struct data_reference);
location_t loc = gimple_location (stmt);
DR_STMT (dr) = stmt;
DR_REF (dr) = op0;
res = dr_analyze_innermost (dr);
gcc_assert (res && stride_of_unit_type_p (DR_STEP (dr), TREE_TYPE (op0)));
nb_bytes = build_size_arg_loc (loc, nb_iter, op0, &stmt_list);
addr_base = size_binop_loc (loc, PLUS_EXPR, DR_OFFSET (dr), DR_INIT (dr));
addr_base = fold_convert_loc (loc, sizetype, addr_base);
/* Test for a negative stride, iterating over every element. */
if (integer_zerop (size_binop (PLUS_EXPR,
TYPE_SIZE_UNIT (TREE_TYPE (op0)),
fold_convert (sizetype, DR_STEP (dr)))))
{
addr_base = size_binop_loc (loc, MINUS_EXPR, addr_base,
fold_convert_loc (loc, sizetype, nb_bytes));
addr_base = size_binop_loc (loc, PLUS_EXPR, addr_base,
TYPE_SIZE_UNIT (TREE_TYPE (op0)));
}
addr_base = fold_build2_loc (loc, POINTER_PLUS_EXPR,
TREE_TYPE (DR_BASE_ADDRESS (dr)),
DR_BASE_ADDRESS (dr), addr_base);
mem = force_gimple_operand (addr_base, &stmts, true, NULL);
gimple_seq_add_seq (&stmt_list, stmts);
fn = build_fold_addr_expr (implicit_built_in_decls [BUILT_IN_MEMSET]);
fn_call = gimple_build_call (fn, 3, mem, integer_zero_node, nb_bytes);
gimple_seq_add_stmt (&stmt_list, fn_call);
gsi_insert_seq_after (&bsi, stmt_list, GSI_CONTINUE_LINKING);
if (dump_file && (dump_flags & TDF_DETAILS))
fprintf (dump_file, "generated memset zero\n");
free_data_ref (dr);
}
static bool
stmt_has_simple_data_refs_p (loop_p outermost_loop ATTRIBUTE_UNUSED,
gimple *stmt)
{
data_reference_p dr;
int j;
bool res = true;
vec<data_reference_p> drs = vNULL;
loop_p outer;
for (outer = loop_containing_stmt (stmt); outer; outer = loop_outer (outer))
{
graphite_find_data_references_in_stmt (outer,
loop_containing_stmt (stmt),
stmt, &drs);
FOR_EACH_VEC_ELT (drs, j, dr)
{
int nb_subscripts = DR_NUM_DIMENSIONS (dr);
tree ref = DR_REF (dr);
for (int i = nb_subscripts - 1; i >= 0; i--)
{
if (!graphite_can_represent_scev (DR_ACCESS_FN (dr, i))
|| (TREE_CODE (ref) != ARRAY_REF
&& TREE_CODE (ref) != MEM_REF
&& TREE_CODE (ref) != COMPONENT_REF))
{
free_data_refs (drs);
return false;
}
ref = TREE_OPERAND (ref, 0);
}
}
free_data_refs (drs);
drs.create (0);
}
static bool
generate_memset_zero (gimple stmt, tree op0, tree nb_iter,
gimple_stmt_iterator bsi)
{
tree addr_base;
tree nb_bytes = NULL;
bool res = false;
gimple_seq stmts = NULL, stmt_list = NULL;
gimple fn_call;
tree mem, fndecl, fntype, fn;
gimple_stmt_iterator i;
struct data_reference *dr = XCNEW (struct data_reference);
location_t loc = gimple_location (stmt);
DR_STMT (dr) = stmt;
DR_REF (dr) = op0;
if (!dr_analyze_innermost (dr))
goto end;
/* Test for a positive stride, iterating over every element. */
if (integer_zerop (fold_build2_loc (loc,
MINUS_EXPR, integer_type_node, DR_STEP (dr),
TYPE_SIZE_UNIT (TREE_TYPE (op0)))))
{
tree offset = fold_convert_loc (loc, sizetype,
size_binop_loc (loc, PLUS_EXPR,
DR_OFFSET (dr),
DR_INIT (dr)));
addr_base = fold_build2_loc (loc, POINTER_PLUS_EXPR,
TREE_TYPE (DR_BASE_ADDRESS (dr)),
DR_BASE_ADDRESS (dr), offset);
}
/* Test for a negative stride, iterating over every element. */
else if (integer_zerop (fold_build2_loc (loc, PLUS_EXPR, integer_type_node,
TYPE_SIZE_UNIT (TREE_TYPE (op0)),
DR_STEP (dr))))
{
nb_bytes = build_size_arg_loc (loc, nb_iter, op0, &stmt_list);
addr_base = size_binop_loc (loc, PLUS_EXPR, DR_OFFSET (dr), DR_INIT (dr));
addr_base = fold_build2_loc (loc, MINUS_EXPR, sizetype, addr_base,
fold_convert_loc (loc, sizetype, nb_bytes));
addr_base = force_gimple_operand (addr_base, &stmts, true, NULL);
gimple_seq_add_seq (&stmt_list, stmts);
addr_base = fold_build2_loc (loc, POINTER_PLUS_EXPR,
TREE_TYPE (DR_BASE_ADDRESS (dr)),
DR_BASE_ADDRESS (dr), addr_base);
}
else
goto end;
mem = force_gimple_operand (addr_base, &stmts, true, NULL);
gimple_seq_add_seq (&stmt_list, stmts);
fndecl = implicit_built_in_decls [BUILT_IN_MEMSET];
fntype = TREE_TYPE (fndecl);
fn = build1 (ADDR_EXPR, build_pointer_type (fntype), fndecl);
if (!nb_bytes)
nb_bytes = build_size_arg_loc (loc, nb_iter, op0, &stmt_list);
fn_call = gimple_build_call (fn, 3, mem, integer_zero_node, nb_bytes);
gimple_seq_add_stmt (&stmt_list, fn_call);
for (i = gsi_start (stmt_list); !gsi_end_p (i); gsi_next (&i))
{
gimple s = gsi_stmt (i);
update_stmt_if_modified (s);
}
gsi_insert_seq_after (&bsi, stmt_list, GSI_CONTINUE_LINKING);
res = true;
if (dump_file && (dump_flags & TDF_DETAILS))
fprintf (dump_file, "generated memset zero\n");
end:
free_data_ref (dr);
return res;
}
static bool
generate_memset_zero (gimple stmt, tree op0, tree nb_iter,
gimple_stmt_iterator bsi)
{
tree t, addr_base;
tree nb_bytes = NULL;
bool res = false;
gimple_seq stmts = NULL, stmt_list = NULL;
gimple fn_call;
tree mem, fndecl, fntype, fn;
gimple_stmt_iterator i;
ssa_op_iter iter;
struct data_reference *dr = XCNEW (struct data_reference);
DR_STMT (dr) = stmt;
DR_REF (dr) = op0;
if (!dr_analyze_innermost (dr))
goto end;
/* Test for a positive stride, iterating over every element. */
if (integer_zerop (fold_build2 (MINUS_EXPR, integer_type_node, DR_STEP (dr),
TYPE_SIZE_UNIT (TREE_TYPE (op0)))))
{
tree offset = fold_convert (sizetype,
size_binop (PLUS_EXPR,
DR_OFFSET (dr),
DR_INIT (dr)));
addr_base = fold_build2 (POINTER_PLUS_EXPR,
TREE_TYPE (DR_BASE_ADDRESS (dr)),
DR_BASE_ADDRESS (dr), offset);
}
/* Test for a negative stride, iterating over every element. */
else if (integer_zerop (fold_build2 (PLUS_EXPR, integer_type_node,
TYPE_SIZE_UNIT (TREE_TYPE (op0)),
DR_STEP (dr))))
{
nb_bytes = build_size_arg (nb_iter, op0, &stmt_list);
addr_base = size_binop (PLUS_EXPR, DR_OFFSET (dr), DR_INIT (dr));
addr_base = fold_build2 (MINUS_EXPR, sizetype, addr_base, nb_bytes);
addr_base = force_gimple_operand (addr_base, &stmts, true, NULL);
gimple_seq_add_seq (&stmt_list, stmts);
addr_base = fold_build2 (POINTER_PLUS_EXPR,
TREE_TYPE (DR_BASE_ADDRESS (dr)),
DR_BASE_ADDRESS (dr), addr_base);
}
else
goto end;
mem = force_gimple_operand (addr_base, &stmts, true, NULL);
gimple_seq_add_seq (&stmt_list, stmts);
fndecl = implicit_built_in_decls [BUILT_IN_MEMSET];
fntype = TREE_TYPE (fndecl);
fn = build1 (ADDR_EXPR, build_pointer_type (fntype), fndecl);
if (!nb_bytes)
nb_bytes = build_size_arg (nb_iter, op0, &stmt_list);
fn_call = gimple_build_call (fn, 3, mem, integer_zero_node, nb_bytes);
gimple_seq_add_stmt (&stmt_list, fn_call);
for (i = gsi_start (stmt_list); !gsi_end_p (i); gsi_next (&i))
{
gimple s = gsi_stmt (i);
update_stmt_if_modified (s);
FOR_EACH_SSA_TREE_OPERAND (t, s, iter, SSA_OP_VIRTUAL_DEFS)
{
if (TREE_CODE (t) == SSA_NAME)
t = SSA_NAME_VAR (t);
mark_sym_for_renaming (t);
}
}
/* Mark also the uses of the VDEFS of STMT to be renamed. */
FOR_EACH_SSA_TREE_OPERAND (t, stmt, iter, SSA_OP_VIRTUAL_DEFS)
{
if (TREE_CODE (t) == SSA_NAME)
{
gimple s;
imm_use_iterator imm_iter;
FOR_EACH_IMM_USE_STMT (s, imm_iter, t)
update_stmt (s);
t = SSA_NAME_VAR (t);
}
mark_sym_for_renaming (t);
}
gsi_insert_seq_after (&bsi, stmt_list, GSI_CONTINUE_LINKING);
res = true;
if (dump_file && (dump_flags & TDF_DETAILS))
fprintf (dump_file, "generated memset zero\n");
todo |= TODO_rebuild_alias;
end:
free_data_ref (dr);
return res;
}
static unsigned
self_reuse_distance (data_reference_p dr, unsigned *loop_sizes, unsigned n,
struct loop *loop)
{
tree stride, access_fn;
HOST_WIDE_INT *strides, astride;
VEC (tree, heap) *access_fns;
tree ref = DR_REF (dr);
unsigned i, ret = ~0u;
/* In the following example:
for (i = 0; i < N; i++)
for (j = 0; j < N; j++)
use (a[j][i]);
the same cache line is accessed each N steps (except if the change from
i to i + 1 crosses the boundary of the cache line). Thus, for self-reuse,
we cannot rely purely on the results of the data dependence analysis.
Instead, we compute the stride of the reference in each loop, and consider
the innermost loop in that the stride is less than cache size. */
strides = XCNEWVEC (HOST_WIDE_INT, n);
access_fns = DR_ACCESS_FNS (dr);
for (i = 0; VEC_iterate (tree, access_fns, i, access_fn); i++)
{
/* Keep track of the reference corresponding to the subscript, so that we
know its stride. */
while (handled_component_p (ref) && TREE_CODE (ref) != ARRAY_REF)
ref = TREE_OPERAND (ref, 0);
if (TREE_CODE (ref) == ARRAY_REF)
{
stride = TYPE_SIZE_UNIT (TREE_TYPE (ref));
if (host_integerp (stride, 1))
astride = tree_low_cst (stride, 1);
else
astride = L1_CACHE_LINE_SIZE;
ref = TREE_OPERAND (ref, 0);
}
else
astride = 1;
add_subscript_strides (access_fn, astride, strides, n, loop);
}
for (i = n; i-- > 0; )
{
unsigned HOST_WIDE_INT s;
s = strides[i] < 0 ? -strides[i] : strides[i];
if (s < (unsigned) L1_CACHE_LINE_SIZE
&& (loop_sizes[i]
> (unsigned) (L1_CACHE_SIZE_BYTES / NONTEMPORAL_FRACTION)))
{
ret = loop_sizes[i];
break;
}
}
free (strides);
return ret;
}
