static bool
forward_propagate_addr_expr_1 (tree name, tree def_rhs,
gimple_stmt_iterator *use_stmt_gsi,
bool single_use_p)
{
tree lhs, rhs, rhs2, array_ref;
tree *rhsp, *lhsp;
gimple use_stmt = gsi_stmt (*use_stmt_gsi);
enum tree_code rhs_code;
gcc_assert (TREE_CODE (def_rhs) == ADDR_EXPR);
lhs = gimple_assign_lhs (use_stmt);
rhs_code = gimple_assign_rhs_code (use_stmt);
rhs = gimple_assign_rhs1 (use_stmt);
/* Trivial cases. The use statement could be a trivial copy or a
useless conversion. Recurse to the uses of the lhs as copyprop does
not copy through different variant pointers and FRE does not catch
all useless conversions. Treat the case of a single-use name and
a conversion to def_rhs type separate, though. */
if (TREE_CODE (lhs) == SSA_NAME
&& ((rhs_code == SSA_NAME && rhs == name)
|| CONVERT_EXPR_CODE_P (rhs_code)))
{
/* Only recurse if we don't deal with a single use or we cannot
do the propagation to the current statement. In particular
we can end up with a conversion needed for a non-invariant
address which we cannot do in a single statement. */
if (!single_use_p
|| (!useless_type_conversion_p (TREE_TYPE (lhs), TREE_TYPE (def_rhs))
&& !is_gimple_min_invariant (def_rhs)))
return forward_propagate_addr_expr (lhs, def_rhs);
gimple_assign_set_rhs1 (use_stmt, unshare_expr (def_rhs));
if (useless_type_conversion_p (TREE_TYPE (lhs), TREE_TYPE (def_rhs)))
gimple_assign_set_rhs_code (use_stmt, TREE_CODE (def_rhs));
else
gimple_assign_set_rhs_code (use_stmt, NOP_EXPR);
return true;
}
/* Now strip away any outer COMPONENT_REF/ARRAY_REF nodes from the LHS.
ADDR_EXPR will not appear on the LHS. */
lhsp = gimple_assign_lhs_ptr (use_stmt);
while (handled_component_p (*lhsp))
lhsp = &TREE_OPERAND (*lhsp, 0);
lhs = *lhsp;
/* Now see if the LHS node is an INDIRECT_REF using NAME. If so,
propagate the ADDR_EXPR into the use of NAME and fold the result. */
if (TREE_CODE (lhs) == INDIRECT_REF
&& TREE_OPERAND (lhs, 0) == name
&& may_propagate_address_into_dereference (def_rhs, lhs)
&& (lhsp != gimple_assign_lhs_ptr (use_stmt)
|| useless_type_conversion_p (TREE_TYPE (TREE_OPERAND (def_rhs, 0)),
TREE_TYPE (rhs))))
{
*lhsp = unshare_expr (TREE_OPERAND (def_rhs, 0));
fold_stmt_inplace (use_stmt);
tidy_after_forward_propagate_addr (use_stmt);
/* Continue propagating into the RHS if this was not the only use. */
if (single_use_p)
return true;
}
/* Strip away any outer COMPONENT_REF, ARRAY_REF or ADDR_EXPR
nodes from the RHS. */
rhsp = gimple_assign_rhs1_ptr (use_stmt);
while (handled_component_p (*rhsp)
|| TREE_CODE (*rhsp) == ADDR_EXPR)
rhsp = &TREE_OPERAND (*rhsp, 0);
rhs = *rhsp;
/* Now see if the RHS node is an INDIRECT_REF using NAME. If so,
propagate the ADDR_EXPR into the use of NAME and fold the result. */
if (TREE_CODE (rhs) == INDIRECT_REF
&& TREE_OPERAND (rhs, 0) == name
&& may_propagate_address_into_dereference (def_rhs, rhs))
{
*rhsp = unshare_expr (TREE_OPERAND (def_rhs, 0));
fold_stmt_inplace (use_stmt);
tidy_after_forward_propagate_addr (use_stmt);
return true;
}
/* Now see if the RHS node is an INDIRECT_REF using NAME. If so,
propagate the ADDR_EXPR into the use of NAME and try to
create a VCE and fold the result. */
if (TREE_CODE (rhs) == INDIRECT_REF
&& TREE_OPERAND (rhs, 0) == name
&& TYPE_SIZE (TREE_TYPE (rhs))
&& TYPE_SIZE (TREE_TYPE (TREE_OPERAND (def_rhs, 0)))
/* Function decls should not be used for VCE either as it could be a
function descriptor that we want and not the actual function code. */
&& TREE_CODE (TREE_OPERAND (def_rhs, 0)) != FUNCTION_DECL
/* We should not convert volatile loads to non volatile loads. */
&& !TYPE_VOLATILE (TREE_TYPE (rhs))
&& !TYPE_VOLATILE (TREE_TYPE (TREE_OPERAND (def_rhs, 0)))
&& operand_equal_p (TYPE_SIZE (TREE_TYPE (rhs)),
TYPE_SIZE (TREE_TYPE (TREE_OPERAND (def_rhs, 0))), 0))
{
tree def_rhs_base, new_rhs = unshare_expr (TREE_OPERAND (def_rhs, 0));
new_rhs = fold_build1 (VIEW_CONVERT_EXPR, TREE_TYPE (rhs), new_rhs);
if (TREE_CODE (new_rhs) != VIEW_CONVERT_EXPR)
{
/* If we have folded the VIEW_CONVERT_EXPR then the result is only
valid if we can replace the whole rhs of the use statement. */
if (rhs != gimple_assign_rhs1 (use_stmt))
return false;
new_rhs = force_gimple_operand_gsi (use_stmt_gsi, new_rhs, true, NULL,
true, GSI_NEW_STMT);
gimple_assign_set_rhs1 (use_stmt, new_rhs);
tidy_after_forward_propagate_addr (use_stmt);
return true;
}
/* If the defining rhs comes from an indirect reference, then do not
convert into a VIEW_CONVERT_EXPR. */
def_rhs_base = TREE_OPERAND (def_rhs, 0);
while (handled_component_p (def_rhs_base))
def_rhs_base = TREE_OPERAND (def_rhs_base, 0);
if (!INDIRECT_REF_P (def_rhs_base))
{
/* We may have arbitrary VIEW_CONVERT_EXPRs in a nested component
reference. Place it there and fold the thing. */
*rhsp = new_rhs;
fold_stmt_inplace (use_stmt);
tidy_after_forward_propagate_addr (use_stmt);
return true;
}
}
/* If the use of the ADDR_EXPR is not a POINTER_PLUS_EXPR, there
is nothing to do. */
if (gimple_assign_rhs_code (use_stmt) != POINTER_PLUS_EXPR
|| gimple_assign_rhs1 (use_stmt) != name)
return false;
/* The remaining cases are all for turning pointer arithmetic into
array indexing. They only apply when we have the address of
element zero in an array. If that is not the case then there
is nothing to do. */
array_ref = TREE_OPERAND (def_rhs, 0);
if (TREE_CODE (array_ref) != ARRAY_REF
|| TREE_CODE (TREE_TYPE (TREE_OPERAND (array_ref, 0))) != ARRAY_TYPE
|| !integer_zerop (TREE_OPERAND (array_ref, 1)))
return false;
rhs2 = gimple_assign_rhs2 (use_stmt);
/* Try to optimize &x[0] p+ C where C is a multiple of the size
of the elements in X into &x[C/element size]. */
if (TREE_CODE (rhs2) == INTEGER_CST)
{
tree new_rhs = maybe_fold_stmt_addition (gimple_expr_type (use_stmt),
array_ref, rhs2);
if (new_rhs)
{
gimple_assign_set_rhs_from_tree (use_stmt_gsi, new_rhs);
use_stmt = gsi_stmt (*use_stmt_gsi);
update_stmt (use_stmt);
tidy_after_forward_propagate_addr (use_stmt);
return true;
}
}
/* Try to optimize &x[0] p+ OFFSET where OFFSET is defined by
converting a multiplication of an index by the size of the
array elements, then the result is converted into the proper
type for the arithmetic. */
if (TREE_CODE (rhs2) == SSA_NAME
/* Avoid problems with IVopts creating PLUS_EXPRs with a
different type than their operands. */
&& useless_type_conversion_p (TREE_TYPE (lhs), TREE_TYPE (def_rhs)))
return forward_propagate_addr_into_variable_array_index (rhs2, def_rhs,
use_stmt_gsi);
return false;
}
static gimple
vect_recog_widen_sum_pattern (gimple last_stmt, tree *type_in, tree *type_out)
{
gimple stmt;
tree oprnd0, oprnd1;
stmt_vec_info stmt_vinfo = vinfo_for_stmt (last_stmt);
tree type, half_type;
gimple pattern_stmt;
loop_vec_info loop_info = STMT_VINFO_LOOP_VINFO (stmt_vinfo);
struct loop *loop = LOOP_VINFO_LOOP (loop_info);
tree var;
if (!is_gimple_assign (last_stmt))
return NULL;
type = gimple_expr_type (last_stmt);
/* Look for the following pattern
DX = (TYPE) X;
sum_1 = DX + sum_0;
In which DX is at least double the size of X, and sum_1 has been
recognized as a reduction variable.
*/
/* Starting from LAST_STMT, follow the defs of its uses in search
of the above pattern. */
if (gimple_assign_rhs_code (last_stmt) != PLUS_EXPR)
return NULL;
if (STMT_VINFO_DEF_TYPE (stmt_vinfo) != vect_reduction_def)
return NULL;
oprnd0 = gimple_assign_rhs1 (last_stmt);
oprnd1 = gimple_assign_rhs2 (last_stmt);
if (!types_compatible_p (TREE_TYPE (oprnd0), type)
|| !types_compatible_p (TREE_TYPE (oprnd1), type))
return NULL;
/* So far so good. Since last_stmt was detected as a (summation) reduction,
we know that oprnd1 is the reduction variable (defined by a loop-header
phi), and oprnd0 is an ssa-name defined by a stmt in the loop body.
Left to check that oprnd0 is defined by a cast from type 'type' to type
'TYPE'. */
if (!widened_name_p (oprnd0, last_stmt, &half_type, &stmt))
return NULL;
oprnd0 = gimple_assign_rhs1 (stmt);
*type_in = half_type;
*type_out = type;
/* Pattern detected. Create a stmt to be used to replace the pattern: */
var = vect_recog_temp_ssa_var (type, NULL);
pattern_stmt = gimple_build_assign_with_ops (WIDEN_SUM_EXPR, var,
oprnd0, oprnd1);
SSA_NAME_DEF_STMT (var) = pattern_stmt;
if (vect_print_dump_info (REPORT_DETAILS))
{
fprintf (vect_dump, "vect_recog_widen_sum_pattern: detected: ");
print_gimple_stmt (vect_dump, pattern_stmt, 0, TDF_SLIM);
}
/* We don't allow changing the order of the computation in the inner-loop
when doing outer-loop vectorization. */
gcc_assert (!nested_in_vect_loop_p (loop, last_stmt));
return pattern_stmt;
}
static gimple
vect_recog_widen_mult_pattern (gimple last_stmt,
tree *type_in,
tree *type_out)
{
gimple def_stmt0, def_stmt1;
tree oprnd0, oprnd1;
tree type, half_type0, half_type1;
gimple pattern_stmt;
tree vectype, vectype_out;
tree dummy;
tree var;
enum tree_code dummy_code;
int dummy_int;
VEC (tree, heap) *dummy_vec;
if (!is_gimple_assign (last_stmt))
return NULL;
type = gimple_expr_type (last_stmt);
/* Starting from LAST_STMT, follow the defs of its uses in search
of the above pattern. */
if (gimple_assign_rhs_code (last_stmt) != MULT_EXPR)
return NULL;
oprnd0 = gimple_assign_rhs1 (last_stmt);
oprnd1 = gimple_assign_rhs2 (last_stmt);
if (!types_compatible_p (TREE_TYPE (oprnd0), type)
|| !types_compatible_p (TREE_TYPE (oprnd1), type))
return NULL;
/* Check argument 0 */
if (!widened_name_p (oprnd0, last_stmt, &half_type0, &def_stmt0))
return NULL;
oprnd0 = gimple_assign_rhs1 (def_stmt0);
/* Check argument 1 */
if (!widened_name_p (oprnd1, last_stmt, &half_type1, &def_stmt1))
return NULL;
oprnd1 = gimple_assign_rhs1 (def_stmt1);
if (!types_compatible_p (half_type0, half_type1))
return NULL;
/* Pattern detected. */
if (vect_print_dump_info (REPORT_DETAILS))
fprintf (vect_dump, "vect_recog_widen_mult_pattern: detected: ");
/* Check target support */
vectype = get_vectype_for_scalar_type (half_type0);
vectype_out = get_vectype_for_scalar_type (type);
if (!vectype
|| !vectype_out
|| !supportable_widening_operation (WIDEN_MULT_EXPR, last_stmt,
vectype_out, vectype,
&dummy, &dummy, &dummy_code,
&dummy_code, &dummy_int, &dummy_vec))
return NULL;
*type_in = vectype;
*type_out = vectype_out;
/* Pattern supported. Create a stmt to be used to replace the pattern: */
var = vect_recog_temp_ssa_var (type, NULL);
pattern_stmt = gimple_build_assign_with_ops (WIDEN_MULT_EXPR, var, oprnd0,
oprnd1);
SSA_NAME_DEF_STMT (var) = pattern_stmt;
if (vect_print_dump_info (REPORT_DETAILS))
print_gimple_stmt (vect_dump, pattern_stmt, 0, TDF_SLIM);
return pattern_stmt;
}
static gimple
vect_recog_dot_prod_pattern (gimple last_stmt, tree *type_in, tree *type_out)
{
gimple stmt;
tree oprnd0, oprnd1;
tree oprnd00, oprnd01;
stmt_vec_info stmt_vinfo = vinfo_for_stmt (last_stmt);
tree type, half_type;
gimple pattern_stmt;
tree prod_type;
loop_vec_info loop_info = STMT_VINFO_LOOP_VINFO (stmt_vinfo);
struct loop *loop = LOOP_VINFO_LOOP (loop_info);
tree var;
if (!is_gimple_assign (last_stmt))
return NULL;
type = gimple_expr_type (last_stmt);
/* Look for the following pattern
DX = (TYPE1) X;
DY = (TYPE1) Y;
DPROD = DX * DY;
DDPROD = (TYPE2) DPROD;
sum_1 = DDPROD + sum_0;
In which
- DX is double the size of X
- DY is double the size of Y
- DX, DY, DPROD all have the same type
- sum is the same size of DPROD or bigger
- sum has been recognized as a reduction variable.
This is equivalent to:
DPROD = X w* Y; #widen mult
sum_1 = DPROD w+ sum_0; #widen summation
or
DPROD = X w* Y; #widen mult
sum_1 = DPROD + sum_0; #summation
*/
/* Starting from LAST_STMT, follow the defs of its uses in search
of the above pattern. */
if (gimple_assign_rhs_code (last_stmt) != PLUS_EXPR)
return NULL;
if (STMT_VINFO_IN_PATTERN_P (stmt_vinfo))
{
/* Has been detected as widening-summation? */
stmt = STMT_VINFO_RELATED_STMT (stmt_vinfo);
type = gimple_expr_type (stmt);
if (gimple_assign_rhs_code (stmt) != WIDEN_SUM_EXPR)
return NULL;
oprnd0 = gimple_assign_rhs1 (stmt);
oprnd1 = gimple_assign_rhs2 (stmt);
half_type = TREE_TYPE (oprnd0);
}
else
{
gimple def_stmt;
if (STMT_VINFO_DEF_TYPE (stmt_vinfo) != vect_reduction_def)
return NULL;
oprnd0 = gimple_assign_rhs1 (last_stmt);
oprnd1 = gimple_assign_rhs2 (last_stmt);
if (!types_compatible_p (TREE_TYPE (oprnd0), type)
|| !types_compatible_p (TREE_TYPE (oprnd1), type))
return NULL;
stmt = last_stmt;
if (widened_name_p (oprnd0, stmt, &half_type, &def_stmt))
{
stmt = def_stmt;
oprnd0 = gimple_assign_rhs1 (stmt);
}
else
half_type = type;
}
/* So far so good. Since last_stmt was detected as a (summation) reduction,
we know that oprnd1 is the reduction variable (defined by a loop-header
phi), and oprnd0 is an ssa-name defined by a stmt in the loop body.
Left to check that oprnd0 is defined by a (widen_)mult_expr */
prod_type = half_type;
stmt = SSA_NAME_DEF_STMT (oprnd0);
/* It could not be the dot_prod pattern if the stmt is outside the loop. */
if (!gimple_bb (stmt) || !flow_bb_inside_loop_p (loop, gimple_bb (stmt)))
return NULL;
/* FORNOW. Can continue analyzing the def-use chain when this stmt in a phi
inside the loop (in case we are analyzing an outer-loop). */
if (!is_gimple_assign (stmt))
return NULL;
stmt_vinfo = vinfo_for_stmt (stmt);
gcc_assert (stmt_vinfo);
if (STMT_VINFO_DEF_TYPE (stmt_vinfo) != vect_internal_def)
return NULL;
if (gimple_assign_rhs_code (stmt) != MULT_EXPR)
return NULL;
if (STMT_VINFO_IN_PATTERN_P (stmt_vinfo))
{
/* Has been detected as a widening multiplication? */
stmt = STMT_VINFO_RELATED_STMT (stmt_vinfo);
if (gimple_assign_rhs_code (stmt) != WIDEN_MULT_EXPR)
return NULL;
stmt_vinfo = vinfo_for_stmt (stmt);
gcc_assert (stmt_vinfo);
gcc_assert (STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_internal_def);
oprnd00 = gimple_assign_rhs1 (stmt);
oprnd01 = gimple_assign_rhs2 (stmt);
}
else
{
tree half_type0, half_type1;
gimple def_stmt;
tree oprnd0, oprnd1;
oprnd0 = gimple_assign_rhs1 (stmt);
oprnd1 = gimple_assign_rhs2 (stmt);
if (!types_compatible_p (TREE_TYPE (oprnd0), prod_type)
|| !types_compatible_p (TREE_TYPE (oprnd1), prod_type))
return NULL;
if (!widened_name_p (oprnd0, stmt, &half_type0, &def_stmt))
return NULL;
oprnd00 = gimple_assign_rhs1 (def_stmt);
if (!widened_name_p (oprnd1, stmt, &half_type1, &def_stmt))
return NULL;
oprnd01 = gimple_assign_rhs1 (def_stmt);
if (!types_compatible_p (half_type0, half_type1))
return NULL;
if (TYPE_PRECISION (prod_type) != TYPE_PRECISION (half_type0) * 2)
return NULL;
}
half_type = TREE_TYPE (oprnd00);
*type_in = half_type;
*type_out = type;
/* Pattern detected. Create a stmt to be used to replace the pattern: */
var = vect_recog_temp_ssa_var (type, NULL);
pattern_stmt = gimple_build_assign_with_ops3 (DOT_PROD_EXPR, var,
oprnd00, oprnd01, oprnd1);
if (vect_print_dump_info (REPORT_DETAILS))
{
fprintf (vect_dump, "vect_recog_dot_prod_pattern: detected: ");
print_gimple_stmt (vect_dump, pattern_stmt, 0, TDF_SLIM);
}
/* We don't allow changing the order of the computation in the inner-loop
when doing outer-loop vectorization. */
gcc_assert (!nested_in_vect_loop_p (loop, last_stmt));
return pattern_stmt;
}
static inline bool
is_replaceable_p (gimple stmt)
{
use_operand_p use_p;
tree def;
gimple use_stmt;
location_t locus1, locus2;
tree block1, block2;
/* Only consider modify stmts. */
if (!is_gimple_assign (stmt))
return false;
/* If the statement may throw an exception, it cannot be replaced. */
if (stmt_could_throw_p (stmt))
return false;
/* Punt if there is more than 1 def. */
def = SINGLE_SSA_TREE_OPERAND (stmt, SSA_OP_DEF);
if (!def)
return false;
/* Only consider definitions which have a single use. */
if (!single_imm_use (def, &use_p, &use_stmt))
return false;
/* If the use isn't in this block, it wont be replaced either. */
if (gimple_bb (use_stmt) != gimple_bb (stmt))
return false;
locus1 = gimple_location (stmt);
block1 = gimple_block (stmt);
if (gimple_code (use_stmt) == GIMPLE_PHI)
{
locus2 = 0;
block2 = NULL_TREE;
}
else
{
locus2 = gimple_location (use_stmt);
block2 = gimple_block (use_stmt);
}
if (!optimize
&& ((locus1 && locus1 != locus2) || (block1 && block1 != block2)))
return false;
/* Used in this block, but at the TOP of the block, not the end. */
if (gimple_code (use_stmt) == GIMPLE_PHI)
return false;
/* There must be no VDEFs. */
if (!(ZERO_SSA_OPERANDS (stmt, SSA_OP_VDEF)))
return false;
/* Without alias info we can't move around loads. */
if (gimple_references_memory_p (stmt) && !optimize)
return false;
/* Float expressions must go through memory if float-store is on. */
if (flag_float_store
&& FLOAT_TYPE_P (gimple_expr_type (stmt)))
return false;
/* An assignment with a register variable on the RHS is not
replaceable. */
if (gimple_assign_rhs_code (stmt) == VAR_DECL
&& DECL_HARD_REGISTER (gimple_assign_rhs1 (stmt)))
return false;
/* No function calls can be replaced. */
if (is_gimple_call (stmt))
return false;
/* Leave any stmt with volatile operands alone as well. */
if (gimple_has_volatile_ops (stmt))
return false;
return true;
}
static gimple
vect_recog_pow_pattern (gimple last_stmt, tree *type_in, tree *type_out)
{
tree type;
tree fn, base, exp = NULL;
gimple stmt;
tree var;
if (!is_gimple_call (last_stmt) || gimple_call_lhs (last_stmt) == NULL)
return NULL;
type = gimple_expr_type (last_stmt);
fn = gimple_call_fndecl (last_stmt);
if (fn == NULL_TREE || DECL_BUILT_IN_CLASS (fn) != BUILT_IN_NORMAL)
return NULL;
switch (DECL_FUNCTION_CODE (fn))
{
case BUILT_IN_POWIF:
case BUILT_IN_POWI:
case BUILT_IN_POWF:
case BUILT_IN_POW:
base = gimple_call_arg (last_stmt, 0);
exp = gimple_call_arg (last_stmt, 1);
if (TREE_CODE (exp) != REAL_CST
&& TREE_CODE (exp) != INTEGER_CST)
return NULL;
break;
default:
return NULL;
}
/* We now have a pow or powi builtin function call with a constant
exponent. */
*type_out = NULL_TREE;
/* Catch squaring. */
if ((host_integerp (exp, 0)
&& tree_low_cst (exp, 0) == 2)
|| (TREE_CODE (exp) == REAL_CST
&& REAL_VALUES_EQUAL (TREE_REAL_CST (exp), dconst2)))
{
*type_in = TREE_TYPE (base);
var = vect_recog_temp_ssa_var (TREE_TYPE (base), NULL);
stmt = gimple_build_assign_with_ops (MULT_EXPR, var, base, base);
SSA_NAME_DEF_STMT (var) = stmt;
return stmt;
}
/* Catch square root. */
if (TREE_CODE (exp) == REAL_CST
&& REAL_VALUES_EQUAL (TREE_REAL_CST (exp), dconsthalf))
{
tree newfn = mathfn_built_in (TREE_TYPE (base), BUILT_IN_SQRT);
*type_in = get_vectype_for_scalar_type (TREE_TYPE (base));
if (*type_in)
{
gimple stmt = gimple_build_call (newfn, 1, base);
if (vectorizable_function (stmt, *type_in, *type_in)
!= NULL_TREE)
{
var = vect_recog_temp_ssa_var (TREE_TYPE (base), stmt);
gimple_call_set_lhs (stmt, var);
return stmt;
}
}
}
return NULL;
}
