// run time version of bitcast intrinsic
JL_DLLEXPORT jl_value_t *jl_bitcast(jl_value_t *ty, jl_value_t *v)
{
JL_TYPECHK(bitcast, datatype, ty);
if (!jl_is_concrete_type(ty) || !jl_is_primitivetype(ty))
jl_error("bitcast: target type not a leaf primitive type");
if (!jl_is_primitivetype(jl_typeof(v)))
jl_error("bitcast: value not a primitive type");
if (jl_datatype_size(jl_typeof(v)) != jl_datatype_size(ty))
jl_error("bitcast: argument size does not match size of target type");
if (ty == jl_typeof(v))
return v;
if (ty == (jl_value_t*)jl_bool_type)
return *(uint8_t*)jl_data_ptr(v) & 1 ? jl_true : jl_false;
return jl_new_bits(ty, jl_data_ptr(v));
}
// Determine if homogeneous tuple with fields of type t will have
// a special alignment beyond normal Julia rules.
// Return special alignment if one exists, 0 if normal alignment rules hold.
// A non-zero result *must* match the LLVM rules for a vector type <nfields x t>.
// For sake of Ahead-Of-Time (AOT) compilation, this routine has to work
// without LLVM being available.
unsigned jl_special_vector_alignment(size_t nfields, jl_value_t *t)
{
if (!jl_is_vecelement_type(t))
return 0;
// LLVM 3.7 and 3.8 either crash or generate wrong code for many
// SIMD vector sizes N. It seems the rule is that N can have at
// most 2 non-zero bits. (This is true at least for N<=100.) See
// also <https://llvm.org/bugs/show_bug.cgi?id=27708>.
size_t mask = nfields;
// See e.g.
// <https://graphics.stanford.edu/%7Eseander/bithacks.html> for an
// explanation of this bit-counting algorithm.
mask &= mask-1; // clear least-significant 1 if present
mask &= mask-1; // clear another 1
if (mask)
return 0; // nfields has more than two 1s
assert(jl_datatype_nfields(t)==1);
jl_value_t *ty = jl_field_type(t, 0);
if (!jl_is_primitivetype(ty))
// LLVM requires that a vector element be a primitive type.
// LLVM allows pointer types as vector elements, but until a
// motivating use case comes up for Julia, we reject pointers.
return 0;
size_t elsz = jl_datatype_size(ty);
if (elsz>8 || (1<<elsz & 0x116) == 0)
// Element size is not 1, 2, 4, or 8.
return 0;
size_t size = nfields*elsz;
// LLVM's alignment rule for vectors seems to be to round up to
// a power of two, even if that's overkill for the target hardware.
size_t alignment=1;
for( ; size>alignment; alignment*=2 )
continue;
return alignment;
}
bool needPassByRef(jl_datatype_t *dt, AttrBuilder &ab) override
{
size_t size = jl_datatype_size(dt);
if (is_complex64(dt) || is_complex128(dt) || (jl_is_primitivetype(dt) && size <= 8))
return false;
ab.addAttribute(Attribute::ByVal);
return true;
}
bool use_sret(jl_datatype_t *dt) override
{
size_t size = jl_datatype_size(dt);
if (size == 0)
return false;
if (is_complex64(dt) || (jl_is_primitivetype(dt) && size <= 8))
return false;
return true;
}
// 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;
}
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);
}
