ExpCode(const ExpVar<N> *self)
{
Xbyak::util::Cpu cpu;
bool useSSE41 = cpu.has(Xbyak::util::Cpu::tSSE41);
try {
makeExp(self, useSSE41);
exp_ = (float(*)(float))getCode();
align(16);
exp_ps_ = (__m128(*)(__m128))getCurr();
makeExpPs(self, useSSE41);
return;
} catch (Xbyak::Error err) {
fprintf(stderr, "ExpCode ERR:%s(%d)\n", Xbyak::ConvertErrorToString(err), err);
} catch (...) {
fprintf(stderr, "ExpCode ERR:unknown error\n");
}
::exit(1);
}
void makeExpPs(const ExpVar<N> *self, const Xbyak::util::Cpu& cpu)
{
typedef ExpVar<N> Self;
using namespace local;
using namespace Xbyak;
inLocalLabel();
#ifdef XBYAK64
const Reg64& base = rcx;
const Reg64& a = rax;
const Reg64& d = rdx;
#else
const Reg32& base = ecx;
const Reg32& a = eax;
const Reg32& d = edx;
#endif
/*
if abs(x) >= maxX then x = max(min(x, maxX), -maxX) and try
minps, maxps are very slow then avoid them
*/
const bool useSSE41 = cpu.has(Xbyak::util::Cpu::tSSE41);
#if defined(XBYAK64_WIN) && !defined(__INTEL_COMPILER)
movaps(xm0, ptr [rcx]);
#endif
mov(base, (size_t)self);
L(".retry");
movaps(xm5, xm0);
andps(xm5, ptr [base + offsetof(Self, i7fffffff)]);
movaps(xm3, ptr [base + offsetof(Self, a)]);
movaps(xm4, ptr [base + offsetof(Self, b)]);
pcmpgtd(xm5, ptr [base + offsetof(Self, maxX)]);
mulps(xm3, xm0);
movaps(xm1, ptr [base + offsetof(Self, i127s)]);
pmovmskb(eax, xm5);
movaps(xm5, ptr [base + offsetof(Self, mask_s)]);
cvtps2dq(xm2, xm3);
pand(xm5, xm2);
cvtdq2ps(xm3, xm2);
test(eax, eax);
jnz(".overflow");
paddd(xm1, xm2);
movd(eax, xm5);
mulps(xm4, xm3);
pextrw(edx, xm5, 2);
subps(xm0, xm4);
movd(xm4, ptr [base + a * 4 + offsetof(Self, tbl)]);
addps(xm0, ptr [base + offsetof(Self, f1)]);
pextrw(eax, xm5, 4);
if (useSSE41) {
pinsrd(xm4, ptr [base + d * 4 + offsetof(Self, tbl)], 1);
} else {
movd(xm3, ptr [base + d * 4 + offsetof(Self, tbl)]);
movlhps(xm4, xm3);
}
pextrw(edx, xm5, 6);
psrld(xm1, self->s);
pslld(xm1, 23);
if (useSSE41) {
pinsrd(xm4, ptr [base + a * 4 + offsetof(Self, tbl)], 2);
pinsrd(xm4, ptr [base + d * 4 + offsetof(Self, tbl)], 3);
} else {
movd(xm2, ptr [base + a * 4 + offsetof(Self, tbl)]);
movd(xm3, ptr [base + d * 4 + offsetof(Self, tbl)]);
movlhps(xm2, xm3);
shufps(xm4, xm2, MIE_PACK(2, 0, 2, 0));
}
por(xm1, xm4);
mulps(xm0, xm1);
ret();
L(".overflow");
minps(xm0, ptr [base + offsetof(Self, maxX)]);
maxps(xm0, ptr [base + offsetof(Self, minX)]);
jmp(".retry");
outLocalLabel();
}