void rewriteFunctionType(TypeFunction *tf, IrFuncTy &fty) override {
Type *retTy = fty.ret->type->toBasetype();
if (!fty.ret->byref && retTy->ty == Tstruct) {
// Rewrite HFAs only because union HFAs are turned into IR types that are
// non-HFA and messes up register selection
if (isHFA((TypeStruct *)retTy, &fty.ret->ltype)) {
fty.ret->rewrite = &hfaToArray;
fty.ret->ltype = hfaToArray.type(fty.ret->type);
}
else {
fty.ret->rewrite = &integerRewrite;
fty.ret->ltype = integerRewrite.type(fty.ret->type);
}
}
for (auto arg : fty.args) {
if (!arg->byref)
rewriteArgument(fty, *arg);
}
// extern(D): reverse parameter order for non variadics, for DMD-compliance
if (tf->linkage == LINKd && tf->varargs != 1 && fty.args.size() > 1) {
fty.reverseParams = true;
}
}
void rewriteArgument(IrFuncTy &fty, IrFuncTyArg &arg) override {
// structs and arrays need rewrite as i32 arrays. This keeps data layout
// unchanged when passed in registers r0-r3 and is necessary to match C ABI
// for struct passing. Without out this rewrite, each field or array
// element is passed in own register. For example: char[4] now all fits in
// r0, where before it consumed r0-r3.
Type *ty = arg.type->toBasetype();
// TODO: want to also rewrite Tsarray as i32 arrays, but sometimes
// llvm selects an aligned ldrd instruction even though the ptr is
// unaligned (e.g. walking through members of array char[5][]).
// if (ty->ty == Tstruct || ty->ty == Tsarray)
if (ty->ty == Tstruct) {
// Rewrite HFAs only because union HFAs are turned into IR types that are
// non-HFA and messes up register selection
if (isHFA((TypeStruct *)ty, &arg.ltype)) {
arg.rewrite = &hfaToArray;
} else if (DtoAlignment(ty) <= 4) {
arg.rewrite = &compositeToArray32;
arg.ltype = compositeToArray32.type(arg.type);
} else {
arg.rewrite = &compositeToArray64;
arg.ltype = compositeToArray64.type(arg.type);
}
}
}
bool use_sret(jl_datatype_t *dt) override
{
jl_datatype_t *ty0 = NULL;
bool hva = false;
if (jl_datatype_size(dt) > 16 && isHFA(dt, &ty0, &hva) > 8)
return true;
return false;
}
bool needPassByRef(jl_datatype_t *dt, AttrBuilder &ab) override
{
jl_datatype_t *ty0 = NULL;
bool hva = false;
if (jl_datatype_size(dt) > 64 && isHFA(dt, &ty0, &hva) > 8) {
ab.addAttribute(Attribute::ByVal);
return true;
}
return false;
}
void rewriteArgument(IrFuncTy &fty, IrFuncTyArg &arg) override {
// FIXME
Type *ty = arg.type->toBasetype();
if (ty->ty == Tstruct || ty->ty == Tsarray) {
// Rewrite HFAs only because union HFAs are turned into IR types that are
// non-HFA and messes up register selection
if (ty->ty == Tstruct && isHFA((TypeStruct *)ty, &arg.ltype)) {
hfaToArray.applyTo(arg, arg.ltype);
} else {
compositeToArray64.applyTo(arg);
}
}
}
bool returnInArg(TypeFunction *tf) override {
if (tf->isref) {
return false;
}
Type *rt = tf->next->toBasetype();
// FIXME
if (tf->linkage == LINKd)
return rt->ty == Tsarray || rt->ty == Tstruct;
return rt->ty == Tsarray ||
(rt->ty == Tstruct && rt->size() > 16 && !isHFA((TypeStruct *)rt));
}
// count the homogeneous floating agregate size (saturating at max count of 8)
unsigned isHFA(jl_datatype_t *ty, jl_datatype_t **ty0, bool *hva) const
{
size_t i, l = ty->layout->nfields;
// handle homogeneous float aggregates
if (l == 0) {
if (ty != jl_float64_type && ty != jl_float32_type)
return 9;
*hva = false;
if (*ty0 == NULL)
*ty0 = ty;
else if (*hva || ty->size != (*ty0)->size)
return 9;
return 1;
}
// handle homogeneous vector aggregates
jl_datatype_t *fld0 = (jl_datatype_t*)jl_field_type(ty, 0);
if (!jl_is_datatype(fld0) || ty->name == jl_vecelement_typename)
return 9;
if (fld0->name == jl_vecelement_typename) {
if (!jl_is_primitivetype(jl_tparam0(fld0)) || jl_datatype_size(ty) > 16)
return 9;
if (l != 1 && l != 2 && l != 4 && l != 8 && l != 16)
return 9;
*hva = true;
if (*ty0 == NULL)
*ty0 = ty;
else if (!*hva || ty->size != (*ty0)->size)
return 9;
for (i = 1; i < l; i++) {
jl_datatype_t *fld = (jl_datatype_t*)jl_field_type(ty, i);
if (fld != fld0)
return 9;
}
return 1;
}
// recurse through other struct types
int n = 0;
for (i = 0; i < l; i++) {
jl_datatype_t *fld = (jl_datatype_t*)jl_field_type(ty, i);
if (!jl_is_datatype(fld) || ((jl_datatype_t*)fld)->layout == NULL)
return 9;
n += isHFA((jl_datatype_t*)fld, ty0, hva);
if (n > 8)
return 9;
}
return n;
}
bool returnInArg(TypeFunction *tf) override {
// AAPCS 5.4 wants composites > 4-bytes returned by arg except for
// Homogeneous Aggregates of up-to 4 float types (6.1.2.1) - an HFA.
// TODO: see if Tsarray should be candidate for HFA.
if (tf->isref)
return false;
Type *rt = tf->next->toBasetype();
if (!isPOD(rt))
return true;
return rt->ty == Tsarray ||
(rt->ty == Tstruct && rt->size() > 4 &&
(gTargetMachine->Options.FloatABIType == llvm::FloatABI::Soft ||
!isHFA((TypeStruct *)rt)));
}
void rewriteFunctionType(IrFuncTy &fty) override {
Type *retTy = fty.ret->type->toBasetype();
if (!fty.ret->byref && retTy->ty == Tstruct) {
// Rewrite HFAs only because union HFAs are turned into IR types that are
// non-HFA and messes up register selection
if (isHFA((TypeStruct *)retTy, &fty.ret->ltype)) {
hfaToArray.applyTo(*fty.ret, fty.ret->ltype);
} else {
integerRewrite.applyTo(*fty.ret);
}
}
for (auto arg : fty.args) {
if (!arg->byref)
rewriteArgument(fty, *arg);
}
}
void rewriteArgument(IrFuncTy &fty, IrFuncTyArg &arg) override {
Type *ty = arg.type->toBasetype();
if (ty->ty == Tstruct || ty->ty == Tsarray) {
if (ty->ty == Tstruct && isHFA((TypeStruct *)ty, &arg.ltype, 8)) {
arg.rewrite = &hfaToArray;
arg.ltype = hfaToArray.type(arg.type);
} else if (canRewriteAsInt(ty, true)) {
arg.rewrite = &integerRewrite;
arg.ltype = integerRewrite.type(arg.type);
} else {
arg.rewrite = &compositeToArray64;
arg.ltype = compositeToArray64.type(arg.type);
}
} else if (ty->isintegral()) {
arg.attrs.add(ty->isunsigned() ? LLAttribute::ZExt : LLAttribute::SExt);
}
}
Type *preferred_llvm_type(jl_datatype_t *dt, bool isret) const override
{
// Arguments are either scalar or passed by value
size_t size = jl_datatype_size(dt);
// don't need to change bitstypes
if (!jl_datatype_nfields(dt))
return NULL;
// legalize this into [n x f32/f64]
jl_datatype_t *ty0 = NULL;
bool hva = false;
int hfa = isHFA(dt, &ty0, &hva);
if (hfa <= 8) {
if (ty0 == jl_float32_type) {
return ArrayType::get(T_float32, hfa);
}
else if (ty0 == jl_float64_type) {
return ArrayType::get(T_float64, hfa);
}
else {
jl_datatype_t *vecty = (jl_datatype_t*)jl_field_type(ty0, 0);
assert(jl_is_datatype(vecty) && vecty->name == jl_vecelement_typename);
jl_value_t *elemty = jl_tparam0(vecty);
assert(jl_is_primitivetype(elemty));
Type *ety = julia_type_to_llvm(elemty);
Type *vty = VectorType::get(ety, jl_datatype_nfields(ty0));
return ArrayType::get(vty, hfa);
}
}
// rewrite integer-sized (non-HFA) struct to an array
// the bitsize of the integer gives the desired alignment
if (size > 8) {
if (jl_datatype_align(dt) <= 8) {
return ArrayType::get(T_int64, (size + 7) / 8);
}
else {
Type *T_int128 = Type::getIntNTy(jl_LLVMContext, 128);
return ArrayType::get(T_int128, (size + 15) / 16);
}
}
return Type::getIntNTy(jl_LLVMContext, size * 8);
}
void rewriteFunctionType(TypeFunction *tf, IrFuncTy &fty) override {
Type *retTy = fty.ret->type->toBasetype();
if (!fty.ret->byref && retTy->ty == Tstruct) {
// Rewrite HFAs only because union HFAs are turned into IR types that are
// non-HFA and messes up register selection
if (isHFA((TypeStruct *)retTy, &fty.ret->ltype)) {
fty.ret->rewrite = &hfaToArray;
fty.ret->ltype = hfaToArray.type(fty.ret->type);
}
else {
fty.ret->rewrite = &integerRewrite;
fty.ret->ltype = integerRewrite.type(fty.ret->type);
}
}
for (auto arg : fty.args) {
if (!arg->byref)
rewriteArgument(fty, *arg);
else if (passByVal(arg->type))
arg->attrs.remove(LLAttribute::ByVal);
}
}
bool passByVal(TypeFunction *, Type *t) override {
t = t->toBasetype();
return t->ty == Tsarray ||
(t->ty == Tstruct && t->size() > 16 && !isHFA((TypeStruct *)t));
}