void CodeGen_X86::visit(const Select *op) {
if (op->condition.type().is_vector()) {
// LLVM handles selects on vector conditions much better at native width
Value *cond = codegen(op->condition);
Value *true_val = codegen(op->true_value);
Value *false_val = codegen(op->false_value);
Type t = op->true_value.type();
int slice_size = 128 / t.bits();
if (slice_size < t.lanes()) {
slice_size = target.natural_vector_size(t);
}
vector<Value *> result;
for (int i = 0; i < t.lanes(); i += slice_size) {
Value *st = slice_vector(true_val, i, slice_size);
Value *sf = slice_vector(false_val, i, slice_size);
Value *sc = slice_vector(cond, i, slice_size);
Value *slice_value = builder->CreateSelect(sc, st, sf);
result.push_back(slice_value);
}
value = concat_vectors(result);
value = slice_vector(value, 0, t.lanes());
} else {
CodeGen_Posix::visit(op);
}
}
void CodeGen_X86::visit(const GT *op) {
if (op->type.is_vector()) {
// Non-native vector widths get legalized poorly by llvm. We
// split it up ourselves.
Type t = op->a.type();
int slice_size = 128 / t.bits();
if (slice_size < t.lanes()) {
slice_size = target.natural_vector_size(t);
}
Value *a = codegen(op->a), *b = codegen(op->b);
vector<Value *> result;
for (int i = 0; i < op->type.lanes(); i += slice_size) {
Value *sa = slice_vector(a, i, slice_size);
Value *sb = slice_vector(b, i, slice_size);
Value *slice_value;
if (t.is_float()) {
slice_value = builder->CreateFCmpOGT(sa, sb);
} else if (t.is_int()) {
slice_value = builder->CreateICmpSGT(sa, sb);
} else {
slice_value = builder->CreateICmpUGT(sa, sb);
}
result.push_back(slice_value);
}
value = concat_vectors(result);
value = slice_vector(value, 0, t.lanes());
} else {
CodeGen_Posix::visit(op);
}
}
void CodeGen_X86::visit(const EQ *op) {
Type t = op->a.type();
int bits = t.lanes() * t.bits();
if (t.lanes() == 1 || bits % 128 == 0) {
// LLVM is fine for native vector widths or scalars
CodeGen_Posix::visit(op);
} else {
// Non-native vector widths get legalized poorly by llvm. We
// split it up ourselves.
Value *a = codegen(op->a), *b = codegen(op->b);
int slice_size = 128 / t.bits();
if (target.has_feature(Target::AVX) && bits > 128) {
slice_size = 256 / t.bits();
}
vector<Value *> result;
for (int i = 0; i < op->type.lanes(); i += slice_size) {
Value *sa = slice_vector(a, i, slice_size);
Value *sb = slice_vector(b, i, slice_size);
Value *slice_value;
if (t.is_float()) {
slice_value = builder->CreateFCmpOEQ(sa, sb);
} else {
slice_value = builder->CreateICmpEQ(sa, sb);
}
result.push_back(slice_value);
}
value = concat_vectors(result);
value = slice_vector(value, 0, t.lanes());
}
}
