mirrors/dynarmic
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity

emit_x64_floating_point: Improve EmitFPToFixed codegen

Browse source
This commit is contained in:
merry 2022-02-23 17:51:54 +00:00 committed by Merry
parent 5fe274f510
commit b34214f953
1 changed files with 35 additions and 30 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

65
src/dynarmic/backend/x64/emit_x64_floating_point.cpp
View file

@ -54,11 +54,8 @@ constexpr u64 f64_max_s16 = 0x40dfffc000000000u; // 32767 as a double
constexpr u64 f64_min_u16 = 0x0000000000000000u; // 0 as a double constexpr u64 f64_min_u16 = 0x0000000000000000u; // 0 as a double
constexpr u64 f64_max_u16 = 0x40efffe000000000u; // 65535 as a double constexpr u64 f64_max_u16 = 0x40efffe000000000u; // 65535 as a double
constexpr u64 f64_max_s32 = 0x41dfffffffc00000u; // 2147483647 as a double constexpr u64 f64_max_s32 = 0x41dfffffffc00000u; // 2147483647 as a double
constexpr u64 f64_min_u32 = 0x0000000000000000u; // 0 as a double
constexpr u64 f64_max_u32 = 0x41efffffffe00000u; // 4294967295 as a double constexpr u64 f64_max_u32 = 0x41efffffffe00000u; // 4294967295 as a double
constexpr u64 f64_max_s64_lim = 0x43e0000000000000u; // 2^63 as a double (actual maximum unrepresentable) constexpr u64 f64_max_s64_lim = 0x43e0000000000000u; // 2^63 as a double (actual maximum unrepresentable)
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)
#define FCODE(NAME) \ #define FCODE(NAME) \
[&code](auto... args) { \ [&code](auto... args) { \
@ -1475,7 +1472,6 @@ static void EmitFPToFixed(BlockOfCode& code, EmitContext& ctx, IR::Inst* inst) {
if (code.HasHostFeature(HostFeature::SSE41) && round_imm) { if (code.HasHostFeature(HostFeature::SSE41) && round_imm) {
const Xbyak::Xmm src = ctx.reg_alloc.UseScratchXmm(args[0]); const Xbyak::Xmm src = ctx.reg_alloc.UseScratchXmm(args[0]);
const Xbyak::Xmm scratch = ctx.reg_alloc.ScratchXmm();
const Xbyak::Reg64 result = ctx.reg_alloc.ScratchGpr().cvt64(); const Xbyak::Reg64 result = ctx.reg_alloc.ScratchGpr().cvt64();
if constexpr (fsize == 64) { if constexpr (fsize == 64) {
@ -1495,12 +1491,14 @@ static void EmitFPToFixed(BlockOfCode& code, EmitContext& ctx, IR::Inst* inst) {
code.cvtss2sd(src, src); code.cvtss2sd(src, src);
} }
ZeroIfNaN<64>(code, src, scratch);
if constexpr (isize == 64) { if constexpr (isize == 64) {
const Xbyak::Xmm scratch = ctx.reg_alloc.ScratchXmm();
Xbyak::Label saturate_max, end; Xbyak::Label saturate_max, end;
if (!unsigned_) { if (!unsigned_) {
ZeroIfNaN<64>(code, src, scratch);
code.movsd(scratch, code.MConst(xword, f64_max_s64_lim)); code.movsd(scratch, code.MConst(xword, f64_max_s64_lim));
code.comisd(scratch, src); code.comisd(scratch, src);
code.jna(saturate_max, code.T_NEAR); code.jna(saturate_max, code.T_NEAR);
@ -1515,41 +1513,48 @@ static void EmitFPToFixed(BlockOfCode& code, EmitContext& ctx, IR::Inst* inst) {
} else { } else {
Xbyak::Label below_max; Xbyak::Label below_max;
code.maxsd(src, code.MConst(xword, f64_min_u64)); const Xbyak::Reg64 result2 = ctx.reg_alloc.ScratchGpr().cvt64();
code.movsd(scratch, code.MConst(xword, f64_max_u64_lim));
code.comisd(scratch, src);
code.jna(saturate_max, code.T_NEAR);
code.movsd(scratch, code.MConst(xword, f64_max_s64_lim)); code.pxor(xmm0, xmm0);
code.comisd(src, scratch);
code.jb(below_max); code.movaps(scratch, src);
code.subsd(scratch, code.MConst(xword, f64_max_s64_lim));
// these both result in zero if src/scratch are NaN
code.maxsd(src, xmm0);
code.maxsd(scratch, xmm0);
code.subsd(src, scratch);
code.cvttsd2si(result, src); code.cvttsd2si(result, src);
code.btc(result, 63); code.cvttsd2si(result2, scratch);
code.jmp(end); code.or_(result, result2);
code.L(below_max); // when src < 2^63, result2 == 0, and result contains the final result
code.cvttsd2si(result, src); // 64 bit gpr // when src >= 2^63, result contains 0x800.... and result2 contains the non-MSB bits
code.L(end); // MSB if result2 is 1 when src >= 2^64
code.SwitchToFarCode(); code.sar(result2, 63);
code.L(saturate_max); code.or_(result, result2);
code.mov(result, 0xFFFF'FFFF'FFFF'FFFF);
code.jmp(end, code.T_NEAR);
code.SwitchToNearCode();
} }
} else if constexpr (isize == 32) { } else if constexpr (isize == 32) {
code.minsd(src, code.MConst(xword, unsigned_ ? f64_max_u32 : f64_max_s32)); if (!unsigned_) {
if (unsigned_) { const Xbyak::Xmm scratch = ctx.reg_alloc.ScratchXmm();
code.maxsd(src, code.MConst(xword, f64_min_u32));
code.cvttsd2si(result, src); // 64 bit gpr ZeroIfNaN<64>(code, src, scratch);
code.minsd(src, code.MConst(xword, f64_max_s32));
// maxsd not required as cvttsd2si results in 0x8000'0000 when out of range
code.cvttsd2si(result.cvt32(), src); // 32 bit gpr
} else { } else {
code.cvttsd2si(result.cvt32(), src); code.pxor(xmm0, xmm0);
code.maxsd(src, xmm0); // results in a zero if src is NaN
code.minsd(src, code.MConst(xword, f64_max_u32));
code.cvttsd2si(result, src); // 64 bit gpr
} }
} else { } else {
code.minsd(src, code.MConst(xword, unsigned_ ? f64_max_u16 : f64_max_s16)); const Xbyak::Xmm scratch = ctx.reg_alloc.ScratchXmm();
ZeroIfNaN<64>(code, src, scratch);
code.maxsd(src, code.MConst(xword, unsigned_ ? f64_min_u16 : f64_min_s16)); code.maxsd(src, code.MConst(xword, unsigned_ ? f64_min_u16 : f64_min_s16));
code.minsd(src, code.MConst(xword, unsigned_ ? f64_max_u16 : f64_max_s16));
code.cvttsd2si(result, src); // 64 bit gpr code.cvttsd2si(result, src); // 64 bit gpr
} }