emit_x64_floating_point: Introduce ICODE

This commit is contained in:
MerryMage 2020-06-25 19:33:12 +01:00
parent 7022281a0b
commit 4b8a781c04

View file

@ -61,19 +61,23 @@ constexpr u64 f64_max_s64_lim = 0x43e0000000000000u; // 2^63 as a double (actual
constexpr u64 f64_min_u64 = 0x0000000000000000u; // 0 as a double constexpr u64 f64_min_u64 = 0x0000000000000000u; // 0 as a double
constexpr u64 f64_max_u64_lim = 0x43f0000000000000u; // 2^64 as a double (actual maximum unrepresentable) constexpr u64 f64_max_u64_lim = 0x43f0000000000000u; // 2^64 as a double (actual maximum unrepresentable)
template<size_t fsize, typename T> #define FCODE(NAME) \
T ChooseOnFsize([[maybe_unused]] T f32, [[maybe_unused]] T f64) { [&code](auto... args){ \
static_assert(fsize == 32 || fsize == 64, "fsize must be either 32 or 64"); if constexpr (fsize == 32) { \
code.NAME##s(args...); \
if constexpr (fsize == 32) { } else { \
return f32; code.NAME##d(args...); \
} else { } \
return f64;
} }
#define ICODE(NAME) \
[&code](auto... args){ \
if constexpr (fsize == 32) { \
code.NAME##d(args...); \
} else { \
code.NAME##q(args...); \
} \
} }
#define FCODE(NAME) (code.*ChooseOnFsize<fsize>(&Xbyak::CodeGenerator::NAME##s, &Xbyak::CodeGenerator::NAME##d))
std::optional<int> ConvertRoundingModeToX64Immediate(FP::RoundingMode rounding_mode) { std::optional<int> ConvertRoundingModeToX64Immediate(FP::RoundingMode rounding_mode) {
switch (rounding_mode) { switch (rounding_mode) {
case FP::RoundingMode::ToNearest_TieEven: case FP::RoundingMode::ToNearest_TieEven:
@ -148,17 +152,10 @@ Xbyak::Label ProcessNaN(BlockOfCode& code, Xbyak::Xmm a) {
template<size_t fsize> template<size_t fsize>
void PostProcessNaN(BlockOfCode& code, Xbyak::Xmm result, Xbyak::Xmm tmp) { void PostProcessNaN(BlockOfCode& code, Xbyak::Xmm result, Xbyak::Xmm tmp) {
if constexpr (fsize == 32) {
code.movaps(tmp, result); code.movaps(tmp, result);
code.cmpunordps(tmp, tmp); FCODE(cmpunordp)(tmp, tmp);
code.pslld(tmp, 31); ICODE(psll)(tmp, int(fsize - 1));
code.xorps(result, tmp); code.xorps(result, tmp);
} else {
code.movaps(tmp, result);
code.cmpunordpd(tmp, tmp);
code.psllq(tmp, 63);
code.xorps(result, tmp);
}
} }
// This is necessary because x86 and ARM differ in they way they return NaNs from floating point operations // This is necessary because x86 and ARM differ in they way they return NaNs from floating point operations