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

emit_x64_floating_point: EmitFPToFixed: maxsd optimization

Browse source 
maxsd is not required when doing a signed conversion, because x64
produces a 0x80...00 value for out of range values.
This commit is contained in:
MerryMage 2018-09-06 15:44:09 +01:00
parent 1abf82ac4a
commit 27427595b7
1 changed files with 9 additions and 5 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

14
src/backend/x64/emit_x64_floating_point.cpp
View file

@ -49,11 +49,9 @@ constexpr u64 f64_non_sign_mask = 0x7fffffffffffffffu;
constexpr u64 f64_smallest_normal = 0x0010000000000000u; constexpr u64 f64_smallest_normal = 0x0010000000000000u;
constexpr u64 f64_penultimate_positive_denormal = 0x000ffffffffffffeu; constexpr u64 f64_penultimate_positive_denormal = 0x000ffffffffffffeu;
constexpr u64 f64_min_s32 = 0xc1e0000000000000u; // -2147483648 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_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_min_s64 = 0xc3e0000000000000u; // -2^63 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_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)
@ -1001,7 +999,9 @@ static void EmitFPToFixed(BlockOfCode& code, EmitContext& ctx, IR::Inst* inst, s
if (isize == 64) { if (isize == 64) {
Xbyak::Label saturate_max, end; Xbyak::Label saturate_max, end;
code.maxsd(src, code.MConst(xword, unsigned_ ? f64_min_u64 : f64_min_s64)); if (unsigned_) {
code.maxsd(src, code.MConst(xword, f64_min_u64));
}
code.movsd(scratch, code.MConst(xword, unsigned_ ? f64_max_u64_lim : f64_max_s64_lim)); code.movsd(scratch, code.MConst(xword, unsigned_ ? f64_max_u64_lim : 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);
@ -1027,8 +1027,12 @@ static void EmitFPToFixed(BlockOfCode& code, EmitContext& ctx, IR::Inst* inst, s
code.SwitchToNearCode(); code.SwitchToNearCode();
} else { } else {
code.minsd(src, code.MConst(xword, unsigned_ ? f64_max_u32 : f64_max_s32)); code.minsd(src, code.MConst(xword, unsigned_ ? f64_max_u32 : f64_max_s32));
code.maxsd(src, code.MConst(xword, unsigned_ ? f64_min_u32 : f64_min_s32)); if (unsigned_) {
code.cvttsd2si(result, src); // 64 bit gpr code.maxsd(src, code.MConst(xword, f64_min_u32));
code.cvttsd2si(result, src); // 64 bit gpr
} else {
code.cvttsd2si(result.cvt32(), src);
}
} }
ctx.reg_alloc.DefineValue(inst, result); ctx.reg_alloc.DefineValue(inst, result);