#include "tree-pass.h"
#include "flags.h"
#include "ggc.h"
+#include "diagnostic.h"
/* Need to include rtl.h, expr.h, etc. for optabs. */
#include "expr.h"
do_binop (gimple_stmt_iterator *gsi, tree inner_type, tree a, tree b,
tree bitpos, tree bitsize, enum tree_code code)
{
- a = tree_vec_extract (gsi, inner_type, a, bitsize, bitpos);
- b = tree_vec_extract (gsi, inner_type, b, bitsize, bitpos);
+ if (TREE_CODE (TREE_TYPE (a)) == VECTOR_TYPE)
+ a = tree_vec_extract (gsi, inner_type, a, bitsize, bitpos);
+ if (TREE_CODE (TREE_TYPE (b)) == VECTOR_TYPE)
+ b = tree_vec_extract (gsi, inner_type, b, bitsize, bitpos);
return gimplify_build2 (gsi, code, inner_type, a, b);
}
-
/* Construct expression (A[BITPOS] code B[BITPOS]) ? -1 : 0
INNER_TYPE is the type of A and B elements
int delta = tree_low_cst (part_width, 1)
/ tree_low_cst (TYPE_SIZE (TREE_TYPE (type)), 1);
int i;
+ location_t loc = gimple_location (gsi_stmt (*gsi));
+
+ if (types_compatible_p (gimple_expr_type (gsi_stmt (*gsi)), type))
+ warning_at (loc, OPT_Wvector_operation_performance,
+ "vector operation will be expanded piecewise");
+ else
+ warning_at (loc, OPT_Wvector_operation_performance,
+ "vector operation will be expanded in parallel");
v = VEC_alloc(constructor_elt, gc, (nunits + delta - 1) / delta);
for (i = 0; i < nunits;
tree result, compute_type;
enum machine_mode mode;
int n_words = tree_low_cst (TYPE_SIZE_UNIT (type), 1) / UNITS_PER_WORD;
+ location_t loc = gimple_location (gsi_stmt (*gsi));
/* We have three strategies. If the type is already correct, just do
the operation an element at a time. Else, if the vector is wider than
mode = mode_for_size (tree_low_cst (TYPE_SIZE (type), 1), MODE_INT, 0);
compute_type = lang_hooks.types.type_for_mode (mode, 1);
result = f (gsi, compute_type, a, b, NULL_TREE, NULL_TREE, code);
+ warning_at (loc, OPT_Wvector_operation_performance,
+ "vector operation will be expanded with a "
+ "single scalar operation");
}
return result;
}
if (i != TYPE_VECTOR_SUBPARTS (TREE_TYPE (vec)))
return NULL_TREE;
-
+
return first;
}
-
+
return NULL_TREE;
}
case PLUS_EXPR:
case MINUS_EXPR:
if (!TYPE_OVERFLOW_TRAPS (type))
- return expand_vector_addition (gsi, do_binop, do_plus_minus, type,
- gimple_assign_rhs1 (assign),
+ return expand_vector_addition (gsi, do_binop, do_plus_minus, type,
+ gimple_assign_rhs1 (assign),
gimple_assign_rhs2 (assign), code);
break;
}
}
+
+/* Build a reference to the element of the vector VECT. Function
+ returns either the element itself, either BIT_FIELD_REF, or an
+ ARRAY_REF expression.
+
+ GSI is requred to insert temporary variables while building a
+ refernece to the element of the vector VECT.
+
+ PTMPVEC is a pointer to the temporary variable for caching
+ purposes. In case when PTMPVEC is NULL new temporary variable
+ will be created. */
+static tree
+vector_element (gimple_stmt_iterator *gsi, tree vect, tree idx, tree *ptmpvec)
+{
+ tree vect_type, vect_elt_type;
+ gimple asgn;
+ tree tmpvec;
+ tree arraytype;
+ bool need_asgn = true;
+ unsigned int elements;
+
+ vect_type = TREE_TYPE (vect);
+ vect_elt_type = TREE_TYPE (vect_type);
+ elements = TYPE_VECTOR_SUBPARTS (vect_type);
+
+ if (TREE_CODE (idx) == INTEGER_CST)
+ {
+ unsigned HOST_WIDE_INT index;
+
+ /* Given that we're about to compute a binary modulus,
+ we don't care about the high bits of the value. */
+ index = TREE_INT_CST_LOW (idx);
+ if (!host_integerp (idx, 1) || index >= elements)
+ {
+ index &= elements - 1;
+ idx = build_int_cst (TREE_TYPE (idx), index);
+ }
+
+ /* When lowering a vector statement sequence do some easy
+ simplification by looking through intermediate vector results. */
+ if (TREE_CODE (vect) == SSA_NAME)
+ {
+ gimple def_stmt = SSA_NAME_DEF_STMT (vect);
+ if (is_gimple_assign (def_stmt)
+ && (gimple_assign_rhs_code (def_stmt) == VECTOR_CST
+ || gimple_assign_rhs_code (def_stmt) == CONSTRUCTOR))
+ vect = gimple_assign_rhs1 (def_stmt);
+ }
+
+ if (TREE_CODE (vect) == VECTOR_CST)
+ {
+ unsigned i;
+ tree vals = TREE_VECTOR_CST_ELTS (vect);
+ for (i = 0; vals; vals = TREE_CHAIN (vals), ++i)
+ if (i == index)
+ return TREE_VALUE (vals);
+ return build_zero_cst (vect_elt_type);
+ }
+ else if (TREE_CODE (vect) == CONSTRUCTOR)
+ {
+ unsigned i;
+ tree elt_i, elt_v;
+
+ FOR_EACH_CONSTRUCTOR_ELT (CONSTRUCTOR_ELTS (vect), i, elt_i, elt_v)
+ if (operand_equal_p (elt_i, idx, 0))
+ return elt_v;
+ return build_zero_cst (vect_elt_type);
+ }
+ else
+ {
+ tree size = TYPE_SIZE (vect_elt_type);
+ tree pos = fold_build2 (MULT_EXPR, TREE_TYPE (idx), idx, size);
+ return fold_build3 (BIT_FIELD_REF, vect_elt_type, vect, size, pos);
+ }
+ }
+
+ if (!ptmpvec)
+ tmpvec = create_tmp_var (vect_type, "vectmp");
+ else if (!*ptmpvec)
+ tmpvec = *ptmpvec = create_tmp_var (vect_type, "vectmp");
+ else
+ {
+ tmpvec = *ptmpvec;
+ need_asgn = false;
+ }
+
+ if (need_asgn)
+ {
+ TREE_ADDRESSABLE (tmpvec) = 1;
+ asgn = gimple_build_assign (tmpvec, vect);
+ gsi_insert_before (gsi, asgn, GSI_SAME_STMT);
+ }
+
+ arraytype = build_array_type_nelts (vect_elt_type, elements);
+ return build4 (ARRAY_REF, vect_elt_type,
+ build1 (VIEW_CONVERT_EXPR, arraytype, tmpvec),
+ idx, NULL_TREE, NULL_TREE);
+}
+
+/* Check if VEC_PERM_EXPR within the given setting is supported
+ by hardware, or lower it piecewise.
+
+ When VEC_PERM_EXPR has the same first and second operands:
+ VEC_PERM_EXPR <v0, v0, mask> the lowered version would be
+ {v0[mask[0]], v0[mask[1]], ...}
+ MASK and V0 must have the same number of elements.
+
+ Otherwise VEC_PERM_EXPR <v0, v1, mask> is lowered to
+ {mask[0] < len(v0) ? v0[mask[0]] : v1[mask[0]], ...}
+ V0 and V1 must have the same type. MASK, V0, V1 must have the
+ same number of arguments. */
+
+static void
+lower_vec_perm (gimple_stmt_iterator *gsi)
+{
+ gimple stmt = gsi_stmt (*gsi);
+ tree mask = gimple_assign_rhs3 (stmt);
+ tree vec0 = gimple_assign_rhs1 (stmt);
+ tree vec1 = gimple_assign_rhs2 (stmt);
+ tree vect_type = TREE_TYPE (vec0);
+ tree mask_type = TREE_TYPE (mask);
+ tree vect_elt_type = TREE_TYPE (vect_type);
+ tree mask_elt_type = TREE_TYPE (mask_type);
+ unsigned int elements = TYPE_VECTOR_SUBPARTS (vect_type);
+ VEC(constructor_elt,gc) *v;
+ tree constr, t, si, i_val;
+ tree vec0tmp = NULL_TREE, vec1tmp = NULL_TREE, masktmp = NULL_TREE;
+ bool two_operand_p = !operand_equal_p (vec0, vec1, 0);
+ location_t loc = gimple_location (gsi_stmt (*gsi));
+ unsigned i;
+
+ if (TREE_CODE (mask) == VECTOR_CST)
+ {
+ unsigned char *sel_int = XALLOCAVEC (unsigned char, elements);
+ tree vals = TREE_VECTOR_CST_ELTS (mask);
+
+ for (i = 0; i < elements; ++i, vals = TREE_CHAIN (vals))
+ sel_int[i] = TREE_INT_CST_LOW (TREE_VALUE (vals)) & (2 * elements - 1);
+
+ if (can_vec_perm_p (TYPE_MODE (vect_type), false, sel_int))
+ return;
+ }
+ else if (can_vec_perm_p (TYPE_MODE (vect_type), true, NULL))
+ return;
+
+ warning_at (loc, OPT_Wvector_operation_performance,
+ "vector shuffling operation will be expanded piecewise");
+
+ v = VEC_alloc (constructor_elt, gc, elements);
+ for (i = 0; i < elements; i++)
+ {
+ si = size_int (i);
+ i_val = vector_element (gsi, mask, si, &masktmp);
+
+ if (TREE_CODE (i_val) == INTEGER_CST)
+ {
+ unsigned HOST_WIDE_INT index;
+
+ index = TREE_INT_CST_LOW (i_val);
+ if (!host_integerp (i_val, 1) || index >= elements)
+ i_val = build_int_cst (mask_elt_type, index & (elements - 1));
+
+ if (two_operand_p && (index & elements) != 0)
+ t = vector_element (gsi, vec1, i_val, &vec1tmp);
+ else
+ t = vector_element (gsi, vec0, i_val, &vec0tmp);
+
+ t = force_gimple_operand_gsi (gsi, t, true, NULL_TREE,
+ true, GSI_SAME_STMT);
+ }
+ else
+ {
+ tree cond = NULL_TREE, v0_val;
+
+ if (two_operand_p)
+ {
+ cond = fold_build2 (BIT_AND_EXPR, mask_elt_type, i_val,
+ build_int_cst (mask_elt_type, elements));
+ cond = force_gimple_operand_gsi (gsi, cond, true, NULL_TREE,
+ true, GSI_SAME_STMT);
+ }
+
+ i_val = fold_build2 (BIT_AND_EXPR, mask_elt_type, i_val,
+ build_int_cst (mask_elt_type, elements - 1));
+ i_val = force_gimple_operand_gsi (gsi, i_val, true, NULL_TREE,
+ true, GSI_SAME_STMT);
+
+ v0_val = vector_element (gsi, vec0, i_val, &vec0tmp);
+ v0_val = force_gimple_operand_gsi (gsi, v0_val, true, NULL_TREE,
+ true, GSI_SAME_STMT);
+
+ if (two_operand_p)
+ {
+ tree v1_val;
+
+ v1_val = vector_element (gsi, vec1, i_val, &vec1tmp);
+ v1_val = force_gimple_operand_gsi (gsi, v1_val, true, NULL_TREE,
+ true, GSI_SAME_STMT);
+
+ cond = fold_build2 (EQ_EXPR, boolean_type_node,
+ cond, build_zero_cst (mask_elt_type));
+ cond = fold_build3 (COND_EXPR, vect_elt_type,
+ cond, v0_val, v1_val);
+ t = force_gimple_operand_gsi (gsi, cond, true, NULL_TREE,
+ true, GSI_SAME_STMT);
+ }
+ else
+ t = v0_val;
+ }
+
+ CONSTRUCTOR_APPEND_ELT (v, si, t);
+ }
+
+ constr = build_constructor (vect_type, v);
+ gimple_assign_set_rhs_from_tree (gsi, constr);
+ update_stmt (gsi_stmt (*gsi));
+}
+
/* Process one statement. If we identify a vector operation, expand it. */
static void
rhs_class = get_gimple_rhs_class (code);
lhs = gimple_assign_lhs (stmt);
+ if (code == VEC_PERM_EXPR)
+ {
+ lower_vec_perm (gsi);
+ return;
+ }
+
if (rhs_class != GIMPLE_UNARY_RHS && rhs_class != GIMPLE_BINARY_RHS)
return;
|| code == VIEW_CONVERT_EXPR)
return;
+ /* These are only created by the vectorizer, after having queried
+ the target support. It's more than just looking at the optab,
+ and there's no need to do it again. */
+ if (code == VEC_INTERLEAVE_HIGH_EXPR
+ || code == VEC_INTERLEAVE_LOW_EXPR
+ || code == VEC_EXTRACT_EVEN_EXPR
+ || code == VEC_EXTRACT_ODD_EXPR)
+ return;
+
gcc_assert (code != CONVERT_EXPR);
/* The signedness is determined from input argument. */
|| code == LROTATE_EXPR
|| code == RROTATE_EXPR)
{
- bool vector_scalar_shift;
- op = optab_for_tree_code (code, type, optab_scalar);
-
- /* Vector/Scalar shift is supported. */
- vector_scalar_shift = (op && (optab_handler (op, TYPE_MODE (type))
- != CODE_FOR_nothing));
-
- /* If the 2nd argument is vector, we need a vector/vector shift.
- Except all the elements in the second vector are the same. */
+ /* Check whether we have vector <op> {x,x,x,x} where x
+ could be a scalar variable or a constant. Transform
+ vector <op> {x,x,x,x} ==> vector <op> scalar. */
if (VECTOR_MODE_P (TYPE_MODE (TREE_TYPE (rhs2))))
{
tree first;
gimple def_stmt;
- /* Check whether we have vector <op> {x,x,x,x} where x
- could be a scalar variable or a constant. Transform
- vector <op> {x,x,x,x} ==> vector <op> scalar. */
- if (vector_scalar_shift
- && ((TREE_CODE (rhs2) == VECTOR_CST
- && (first = uniform_vector_p (rhs2)) != NULL_TREE)
- || (TREE_CODE (rhs2) == SSA_NAME
- && (def_stmt = SSA_NAME_DEF_STMT (rhs2))
- && gimple_assign_single_p (def_stmt)
- && (first = uniform_vector_p
- (gimple_assign_rhs1 (def_stmt))) != NULL_TREE)))
+ if ((TREE_CODE (rhs2) == VECTOR_CST
+ && (first = uniform_vector_p (rhs2)) != NULL_TREE)
+ || (TREE_CODE (rhs2) == SSA_NAME
+ && (def_stmt = SSA_NAME_DEF_STMT (rhs2))
+ && gimple_assign_single_p (def_stmt)
+ && (first = uniform_vector_p
+ (gimple_assign_rhs1 (def_stmt))) != NULL_TREE))
{
gimple_assign_set_rhs2 (stmt, first);
update_stmt (stmt);
rhs2 = first;
}
- else
- op = optab_for_tree_code (code, type, optab_vector);
}
-
- /* Try for a vector/scalar shift, and if we don't have one, see if we
- have a vector/vector shift */
- else if (!vector_scalar_shift)
+
+ if (VECTOR_MODE_P (TYPE_MODE (TREE_TYPE (rhs2))))
+ op = optab_for_tree_code (code, type, optab_vector);
+ else
{
- op = optab_for_tree_code (code, type, optab_vector);
+ op = optab_for_tree_code (code, type, optab_scalar);
- if (op && (optab_handler (op, TYPE_MODE (type))
- != CODE_FOR_nothing))
- {
- /* Transform vector <op> scalar => vector <op> {x,x,x,x}. */
- int n_parts = TYPE_VECTOR_SUBPARTS (type);
- int part_size = tree_low_cst (TYPE_SIZE (TREE_TYPE (type)), 1);
- tree part_type = lang_hooks.types.type_for_size (part_size, 1);
- tree vect_type = build_vector_type (part_type, n_parts);
-
- rhs2 = fold_convert (part_type, rhs2);
- rhs2 = build_vector_from_val (vect_type, rhs2);
- gimple_assign_set_rhs2 (stmt, rhs2);
- update_stmt (stmt);
- }
+ /* The rtl expander will expand vector/scalar as vector/vector
+ if necessary. Don't bother converting the stmt here. */
+ if (op == NULL
+ || optab_handler (op, TYPE_MODE (type)) == CODE_FOR_nothing)
+ op = optab_for_tree_code (code, type, optab_vector);
}
}
else
|| code == VEC_UNPACK_LO_EXPR
|| code == VEC_PACK_TRUNC_EXPR
|| code == VEC_PACK_SAT_EXPR
- || code == VEC_PACK_FIX_TRUNC_EXPR)
+ || code == VEC_PACK_FIX_TRUNC_EXPR
+ || code == VEC_WIDEN_LSHIFT_HI_EXPR
+ || code == VEC_WIDEN_LSHIFT_LO_EXPR)
type = TREE_TYPE (rhs1);
/* Optabs will try converting a negation into a subtraction, so
if (compute_type == type)
{
compute_mode = TYPE_MODE (compute_type);
- if ((GET_MODE_CLASS (compute_mode) == MODE_VECTOR_INT
- || GET_MODE_CLASS (compute_mode) == MODE_VECTOR_FLOAT
- || GET_MODE_CLASS (compute_mode) == MODE_VECTOR_FRACT
- || GET_MODE_CLASS (compute_mode) == MODE_VECTOR_UFRACT
- || GET_MODE_CLASS (compute_mode) == MODE_VECTOR_ACCUM
- || GET_MODE_CLASS (compute_mode) == MODE_VECTOR_UACCUM)
+ if (VECTOR_MODE_P (compute_mode)
&& op != NULL
&& optab_handler (op, compute_mode) != CODE_FOR_nothing)
return;
if it may need the bit-twiddling tricks implemented in this file. */
static bool
-gate_expand_vector_operations (void)
+gate_expand_vector_operations_ssa (void)
{
- return flag_tree_vectorize != 0;
+ return optimize == 0;
}
static unsigned int
{
GIMPLE_PASS,
"veclower", /* name */
- 0, /* gate */
+ gate_expand_vector_operations_ssa, /* gate */
expand_vector_operations, /* execute */
NULL, /* sub */
NULL, /* next */
TODO_update_ssa /* todo_flags_finish */
| TODO_verify_ssa
| TODO_verify_stmts | TODO_verify_flow
+ | TODO_cleanup_cfg
}
};
{
GIMPLE_PASS,
"veclower2", /* name */
- gate_expand_vector_operations, /* gate */
+ 0, /* gate */
expand_vector_operations, /* execute */
NULL, /* sub */
NULL, /* next */
TODO_update_ssa /* todo_flags_finish */
| TODO_verify_ssa
| TODO_verify_stmts | TODO_verify_flow
+ | TODO_cleanup_cfg
}
};