template<class _mm> void TimgFilterNoiseMplayer::Tprocess::lineNoise_simd(uint8_t *dst,const uint8_t *src, int8_t *noise, int len, int shift)
{
    if (_mm::align && (intptr_t(src)&15 || intptr_t(dst)&15)) {
        lineNoise_simd<typename _mm::T64>(dst,src,noise,len,shift);
        return;
    }
    int mmx_len=len&(~(_mm::size-1));
    noise+=shift;

    typename _mm::__m mm7;
    pcmpeqb(mm7,mm7);
    psllw(mm7,15);
    packsswb(mm7,mm7);
    for (int x=-mmx_len; x<0; x+=_mm::size) {
        //".balign 16            \n\t"
        typename _mm::__m mm0,mm1;
        movq   (mm0,src+mmx_len+x);
        movdqu (mm1,noise+mmx_len+x); //SSE3
        pxor   (mm0,mm7);
        paddsb (mm0,mm1);
        pxor   (mm0,mm7);
        _mm::movntq(dst+mmx_len+x,mm0);
    }

    if (mmx_len!=len) {
        lineNoise_C(dst+mmx_len,src+mmx_len,noise+mmx_len,len-mmx_len,0);
    }
}
template<class _mm> void TimgFilterNoiseMplayer::Tprocess::lineNoiseAvg_simd(uint8_t *dst,const uint8_t *src, int len, int8_t **shift_)
{
    if (_mm::align && (intptr_t(src)&15 || intptr_t(dst)&15)) {
        lineNoiseAvg_simd<typename _mm::T64>(dst,src,len,shift_);
        return;
    }
    const int mmx_len=len&(~(_mm::size-1));

    int8_t *shift2[3]= {shift_[0]+mmx_len, shift_[1]+mmx_len, shift_[2]+mmx_len};
    for (int x=-mmx_len; x<0; x+=_mm::size) {
        //".balign 16                   \n\t"
        typename _mm::__m mm0,mm1,mm2,mm3;
        movdqu (mm1,shift2[0]+mmx_len+x);
        movq (mm0,src+mmx_len+x);
        typename _mm::__m shift1_8;
        movVqu(shift1_8, shift2[1]+mmx_len+x);
        paddb (mm1,shift1_8);
        typename _mm::__m shift2_8;
        movVqu(shift2_8, shift2[2]+mmx_len+x);
        paddb (mm1,shift2_8);
        movq (mm2,mm0);
        movq (mm3,mm1);
        punpcklbw (mm0,mm0);
        punpckhbw (mm2,mm2);
        punpcklbw (mm1,mm1);
        punpckhbw (mm3,mm3);
        pmulhw (mm1,mm0);
        pmulhw (mm3,mm2);
        paddw (mm1,mm1);
        paddw (mm3,mm3);
        paddw (mm1,mm0);
        paddw (mm3,mm2);
        psrlw (mm1,8);
        psrlw (mm3,8);
        packuswb (mm1,mm3);
        movq (dst+mmx_len+x,mm1);
    }

    if (mmx_len!=len) {
        lineNoiseAvg_C(dst+mmx_len, src+mmx_len, len-mmx_len, shift2);
    }
}
void GPUDrawScanlineCodeGenerator::Init()
{
	mov(eax, dword[esp + _top]);

	// uint16* fb = (uint16*)m_global.vm + (top << (10 + sel.scalex)) + left;

	mov(edi, eax);
	shl(edi, 10 + m_sel.scalex);
	add(edi, edx);
	lea(edi, ptr[edi * 2 + (size_t)m_local.gd->vm]);

	// int steps = pixels - 8;

	sub(ecx, 8);

	if(m_sel.dtd)
	{
		// dither = GSVector4i::load<false>(&m_dither[top & 3][left & 3]);

		and(eax, 3);
		shl(eax, 5);
		and(edx, 3);
		shl(edx, 1);
		movdqu(xmm0, ptr[eax + edx + (size_t)m_dither]);
		movdqa(ptr[&m_local.temp.dither], xmm0);
	}

	mov(edx, dword[esp + _v]);

	if(m_sel.tme)
	{
		mov(esi, dword[&m_local.gd->tex]);

		// GSVector4i vt = GSVector4i(v.t).xxzzl();

		cvttps2dq(xmm4, ptr[edx + offsetof(GSVertexSW, t)]);
		pshuflw(xmm4, xmm4, _MM_SHUFFLE(2, 2, 0, 0));

		// s = vt.xxxx().add16(m_local.d.s);
		// t = vt.yyyy().add16(m_local.d.t);

		pshufd(xmm2, xmm4, _MM_SHUFFLE(0, 0, 0, 0));
		pshufd(xmm3, xmm4, _MM_SHUFFLE(1, 1, 1, 1));

		paddw(xmm2, ptr[&m_local.d.s]);

		if(!m_sel.sprite)
		{
			paddw(xmm3, ptr[&m_local.d.t]);
		}
		else
		{
			if(m_sel.ltf)
			{
				movdqa(xmm0, xmm3);
				psllw(xmm0, 8);
				psrlw(xmm0, 1);
				movdqa(ptr[&m_local.temp.vf], xmm0);
			}
		}

		movdqa(ptr[&m_local.temp.s], xmm2);
		movdqa(ptr[&m_local.temp.t], xmm3);
	}

	if(m_sel.tfx != 3) // != decal
	{
		// GSVector4i vc = GSVector4i(v.c).xxzzlh();

		cvttps2dq(xmm6, ptr[edx + offsetof(GSVertexSW, c)]);
		pshuflw(xmm6, xmm6, _MM_SHUFFLE(2, 2, 0, 0));
		pshufhw(xmm6, xmm6, _MM_SHUFFLE(2, 2, 0, 0));

		// r = vc.xxxx();
		// g = vc.yyyy();
		// b = vc.zzzz();

		pshufd(xmm4, xmm6, _MM_SHUFFLE(0, 0, 0, 0));
		pshufd(xmm5, xmm6, _MM_SHUFFLE(1, 1, 1, 1));
		pshufd(xmm6, xmm6, _MM_SHUFFLE(2, 2, 2, 2));

		if(m_sel.iip)
		{
			// r = r.add16(m_local.d.r);
			// g = g.add16(m_local.d.g);
			// b = b.add16(m_local.d.b);

			paddw(xmm4, ptr[&m_local.d.r]);
			paddw(xmm5, ptr[&m_local.d.g]);
			paddw(xmm6, ptr[&m_local.d.b]);
		}

		movdqa(ptr[&m_local.temp.r], xmm4);
		movdqa(ptr[&m_local.temp.g], xmm5);
		movdqa(ptr[&m_local.temp.b], xmm6);
	}
}
void MacroAssembler::fast_log(XMMRegister xmm0, XMMRegister xmm1, XMMRegister xmm2, XMMRegister xmm3, XMMRegister xmm4, XMMRegister xmm5, XMMRegister xmm6, XMMRegister xmm7, Register eax, Register ecx, Register edx, Register tmp1, Register tmp2) {
  Label L_2TAG_PACKET_0_0_2, L_2TAG_PACKET_1_0_2, L_2TAG_PACKET_2_0_2, L_2TAG_PACKET_3_0_2;
  Label L_2TAG_PACKET_4_0_2, L_2TAG_PACKET_5_0_2, L_2TAG_PACKET_6_0_2, L_2TAG_PACKET_7_0_2;
  Label L_2TAG_PACKET_8_0_2;
  Label L_2TAG_PACKET_12_0_2, L_2TAG_PACKET_13_0_2, B1_3, B1_5, start;

  assert_different_registers(tmp1, tmp2, eax, ecx, edx);
  jmp(start);
  address L_tbl = (address)_L_tbl;
  address log2 = (address)_log2;
  address coeff = (address)_coeff;

  bind(start);
  subq(rsp, 24);
  movsd(Address(rsp, 0), xmm0);
  mov64(rax, 0x3ff0000000000000);
  movdq(xmm2, rax);
  mov64(rdx, 0x77f0000000000000);
  movdq(xmm3, rdx);
  movl(ecx, 32768);
  movdl(xmm4, rcx);
  mov64(tmp1, 0xffffe00000000000);
  movdq(xmm5, tmp1);
  movdqu(xmm1, xmm0);
  pextrw(eax, xmm0, 3);
  por(xmm0, xmm2);
  movl(ecx, 16352);
  psrlq(xmm0, 27);
  lea(tmp2, ExternalAddress(L_tbl));
  psrld(xmm0, 2);
  rcpps(xmm0, xmm0);
  psllq(xmm1, 12);
  pshufd(xmm6, xmm5, 228);
  psrlq(xmm1, 12);
  subl(eax, 16);
  cmpl(eax, 32736);
  jcc(Assembler::aboveEqual, L_2TAG_PACKET_0_0_2);

  bind(L_2TAG_PACKET_1_0_2);
  paddd(xmm0, xmm4);
  por(xmm1, xmm3);
  movdl(edx, xmm0);
  psllq(xmm0, 29);
  pand(xmm5, xmm1);
  pand(xmm0, xmm6);
  subsd(xmm1, xmm5);
  mulpd(xmm5, xmm0);
  andl(eax, 32752);
  subl(eax, ecx);
  cvtsi2sdl(xmm7, eax);
  mulsd(xmm1, xmm0);
  movq(xmm6, ExternalAddress(log2));       // 0xfefa3800UL, 0x3fa62e42UL
  movdqu(xmm3, ExternalAddress(coeff));    // 0x92492492UL, 0x3fc24924UL, 0x00000000UL, 0xbfd00000UL
  subsd(xmm5, xmm2);
  andl(edx, 16711680);
  shrl(edx, 12);
  movdqu(xmm0, Address(tmp2, edx));
  movdqu(xmm4, ExternalAddress(16 + coeff)); // 0x3d6fb175UL, 0xbfc5555eUL, 0x55555555UL, 0x3fd55555UL
  addsd(xmm1, xmm5);
  movdqu(xmm2, ExternalAddress(32 + coeff)); // 0x9999999aUL, 0x3fc99999UL, 0x00000000UL, 0xbfe00000UL
  mulsd(xmm6, xmm7);
  movddup(xmm5, xmm1);
  mulsd(xmm7, ExternalAddress(8 + log2));    // 0x93c76730UL, 0x3ceef357UL
  mulsd(xmm3, xmm1);
  addsd(xmm0, xmm6);
  mulpd(xmm4, xmm5);
  mulpd(xmm5, xmm5);
  movddup(xmm6, xmm0);
  addsd(xmm0, xmm1);
  addpd(xmm4, xmm2);
  mulpd(xmm3, xmm5);
  subsd(xmm6, xmm0);
  mulsd(xmm4, xmm1);
  pshufd(xmm2, xmm0, 238);
  addsd(xmm1, xmm6);
  mulsd(xmm5, xmm5);
  addsd(xmm7, xmm2);
  addpd(xmm4, xmm3);
  addsd(xmm1, xmm7);
  mulpd(xmm4, xmm5);
  addsd(xmm1, xmm4);
  pshufd(xmm5, xmm4, 238);
  addsd(xmm1, xmm5);
  addsd(xmm0, xmm1);
  jmp(B1_5);

  bind(L_2TAG_PACKET_0_0_2);
  movq(xmm0, Address(rsp, 0));
  movq(xmm1, Address(rsp, 0));
  addl(eax, 16);
  cmpl(eax, 32768);
  jcc(Assembler::aboveEqual, L_2TAG_PACKET_2_0_2);
  cmpl(eax, 16);
  jcc(Assembler::below, L_2TAG_PACKET_3_0_2);

  bind(L_2TAG_PACKET_4_0_2);
  addsd(xmm0, xmm0);
  jmp(B1_5);

  bind(L_2TAG_PACKET_5_0_2);
  jcc(Assembler::above, L_2TAG_PACKET_4_0_2);
  cmpl(edx, 0);
  jcc(Assembler::above, L_2TAG_PACKET_4_0_2);
  jmp(L_2TAG_PACKET_6_0_2);

  bind(L_2TAG_PACKET_3_0_2);
  xorpd(xmm1, xmm1);
  addsd(xmm1, xmm0);
  movdl(edx, xmm1);
  psrlq(xmm1, 32);
  movdl(ecx, xmm1);
  orl(edx, ecx);
  cmpl(edx, 0);
  jcc(Assembler::equal, L_2TAG_PACKET_7_0_2);
  xorpd(xmm1, xmm1);
  movl(eax, 18416);
  pinsrw(xmm1, eax, 3);
  mulsd(xmm0, xmm1);
  movdqu(xmm1, xmm0);
  pextrw(eax, xmm0, 3);
  por(xmm0, xmm2);
  psrlq(xmm0, 27);
  movl(ecx, 18416);
  psrld(xmm0, 2);
  rcpps(xmm0, xmm0);
  psllq(xmm1, 12);
  pshufd(xmm6, xmm5, 228);
  psrlq(xmm1, 12);
  jmp(L_2TAG_PACKET_1_0_2);

  bind(L_2TAG_PACKET_2_0_2);
  movdl(edx, xmm1);
  psrlq(xmm1, 32);
  movdl(ecx, xmm1);
  addl(ecx, ecx);
  cmpl(ecx, -2097152);
  jcc(Assembler::aboveEqual, L_2TAG_PACKET_5_0_2);
  orl(edx, ecx);
  cmpl(edx, 0);
  jcc(Assembler::equal, L_2TAG_PACKET_7_0_2);

  bind(L_2TAG_PACKET_6_0_2);
  xorpd(xmm1, xmm1);
  xorpd(xmm0, xmm0);
  movl(eax, 32752);
  pinsrw(xmm1, eax, 3);
  mulsd(xmm0, xmm1);
  movl(Address(rsp, 16), 3);
  jmp(L_2TAG_PACKET_8_0_2);
  bind(L_2TAG_PACKET_7_0_2);
  xorpd(xmm1, xmm1);
  xorpd(xmm0, xmm0);
  movl(eax, 49136);
  pinsrw(xmm0, eax, 3);
  divsd(xmm0, xmm1);
  movl(Address(rsp, 16), 2);

  bind(L_2TAG_PACKET_8_0_2);
  movq(Address(rsp, 8), xmm0);

  bind(B1_3);
  movq(xmm0, Address(rsp, 8));

  bind(B1_5);
  addq(rsp, 24);
}
void MacroAssembler::fast_exp(XMMRegister xmm0, XMMRegister xmm1, XMMRegister xmm2, XMMRegister xmm3, XMMRegister xmm4, XMMRegister xmm5, XMMRegister xmm6, XMMRegister xmm7, Register eax, Register ecx, Register edx, Register tmp) {
  Label L_2TAG_PACKET_0_0_2, L_2TAG_PACKET_1_0_2, L_2TAG_PACKET_2_0_2, L_2TAG_PACKET_3_0_2;
  Label L_2TAG_PACKET_4_0_2, L_2TAG_PACKET_5_0_2, L_2TAG_PACKET_6_0_2, L_2TAG_PACKET_7_0_2;
  Label L_2TAG_PACKET_8_0_2, L_2TAG_PACKET_9_0_2, L_2TAG_PACKET_10_0_2, L_2TAG_PACKET_11_0_2;
  Label L_2TAG_PACKET_12_0_2, L_2TAG_PACKET_13_0_2, B1_3, B1_5, start;

  assert_different_registers(tmp, eax, ecx, edx);
  jmp(start);
  address static_const_table = (address)_static_const_table;

  bind(start);
  subl(rsp, 120);
  movl(Address(rsp, 64), tmp);
  lea(tmp, ExternalAddress(static_const_table));
  movdqu(xmm0, Address(rsp, 128));
  unpcklpd(xmm0, xmm0);
  movdqu(xmm1, Address(tmp, 64));          // 0x652b82feUL, 0x40571547UL, 0x652b82feUL, 0x40571547UL
  movdqu(xmm6, Address(tmp, 48));          // 0x00000000UL, 0x43380000UL, 0x00000000UL, 0x43380000UL
  movdqu(xmm2, Address(tmp, 80));          // 0xfefa0000UL, 0x3f862e42UL, 0xfefa0000UL, 0x3f862e42UL
  movdqu(xmm3, Address(tmp, 96));          // 0xbc9e3b3aUL, 0x3d1cf79aUL, 0xbc9e3b3aUL, 0x3d1cf79aUL
  pextrw(eax, xmm0, 3);
  andl(eax, 32767);
  movl(edx, 16527);
  subl(edx, eax);
  subl(eax, 15504);
  orl(edx, eax);
  cmpl(edx, INT_MIN);
  jcc(Assembler::aboveEqual, L_2TAG_PACKET_0_0_2);
  mulpd(xmm1, xmm0);
  addpd(xmm1, xmm6);
  movapd(xmm7, xmm1);
  subpd(xmm1, xmm6);
  mulpd(xmm2, xmm1);
  movdqu(xmm4, Address(tmp, 128));         // 0xe3289860UL, 0x3f56c15cUL, 0x555b9e25UL, 0x3fa55555UL
  mulpd(xmm3, xmm1);
  movdqu(xmm5, Address(tmp, 144));         // 0xc090cf0fUL, 0x3f811115UL, 0x55548ba1UL, 0x3fc55555UL
  subpd(xmm0, xmm2);
  movdl(eax, xmm7);
  movl(ecx, eax);
  andl(ecx, 63);
  shll(ecx, 4);
  sarl(eax, 6);
  movl(edx, eax);
  movdqu(xmm6, Address(tmp, 16));          // 0xffffffc0UL, 0x00000000UL, 0xffffffc0UL, 0x00000000UL
  pand(xmm7, xmm6);
  movdqu(xmm6, Address(tmp, 32));          // 0x0000ffc0UL, 0x00000000UL, 0x0000ffc0UL, 0x00000000UL
  paddq(xmm7, xmm6);
  psllq(xmm7, 46);
  subpd(xmm0, xmm3);
  movdqu(xmm2, Address(tmp, ecx, Address::times_1, 160));
  mulpd(xmm4, xmm0);
  movapd(xmm6, xmm0);
  movapd(xmm1, xmm0);
  mulpd(xmm6, xmm6);
  mulpd(xmm0, xmm6);
  addpd(xmm5, xmm4);
  mulsd(xmm0, xmm6);
  mulpd(xmm6, Address(tmp, 112));          // 0xfffffffeUL, 0x3fdfffffUL, 0xfffffffeUL, 0x3fdfffffUL
  addsd(xmm1, xmm2);
  unpckhpd(xmm2, xmm2);
  mulpd(xmm0, xmm5);
  addsd(xmm1, xmm0);
  por(xmm2, xmm7);
  unpckhpd(xmm0, xmm0);
  addsd(xmm0, xmm1);
  addsd(xmm0, xmm6);
  addl(edx, 894);
  cmpl(edx, 1916);
  jcc (Assembler::above, L_2TAG_PACKET_1_0_2);
  mulsd(xmm0, xmm2);
  addsd(xmm0, xmm2);
  jmp(L_2TAG_PACKET_2_0_2);

  bind(L_2TAG_PACKET_1_0_2);
  fnstcw(Address(rsp, 24));
  movzwl(edx, Address(rsp, 24));
  orl(edx, 768);
  movw(Address(rsp, 28), edx);
  fldcw(Address(rsp, 28));
  movl(edx, eax);
  sarl(eax, 1);
  subl(edx, eax);
  movdqu(xmm6, Address(tmp, 0));           // 0x00000000UL, 0xfff00000UL, 0x00000000UL, 0xfff00000UL
  pandn(xmm6, xmm2);
  addl(eax, 1023);
  movdl(xmm3, eax);
  psllq(xmm3, 52);
  por(xmm6, xmm3);
  addl(edx, 1023);
  movdl(xmm4, edx);
  psllq(xmm4, 52);
  movsd(Address(rsp, 8), xmm0);
  fld_d(Address(rsp, 8));
  movsd(Address(rsp, 16), xmm6);
  fld_d(Address(rsp, 16));
  fmula(1);
  faddp(1);
  movsd(Address(rsp, 8), xmm4);
  fld_d(Address(rsp, 8));
  fmulp(1);
  fstp_d(Address(rsp, 8));
  movsd(xmm0,Address(rsp, 8));
  fldcw(Address(rsp, 24));
  pextrw(ecx, xmm0, 3);
  andl(ecx, 32752);
  cmpl(ecx, 32752);
  jcc(Assembler::greaterEqual, L_2TAG_PACKET_3_0_2);
  cmpl(ecx, 0);
  jcc(Assembler::equal, L_2TAG_PACKET_4_0_2);
  jmp(L_2TAG_PACKET_2_0_2);
  cmpl(ecx, INT_MIN);
  jcc(Assembler::less, L_2TAG_PACKET_3_0_2);
  cmpl(ecx, -1064950997);
  jcc(Assembler::less, L_2TAG_PACKET_2_0_2);
  jcc(Assembler::greater, L_2TAG_PACKET_4_0_2);
  movl(edx, Address(rsp, 128));
  cmpl(edx ,-17155601);
  jcc(Assembler::less, L_2TAG_PACKET_2_0_2);
  jmp(L_2TAG_PACKET_4_0_2);

  bind(L_2TAG_PACKET_3_0_2);
  movl(edx, 14);
  jmp(L_2TAG_PACKET_5_0_2);

  bind(L_2TAG_PACKET_4_0_2);
  movl(edx, 15);

  bind(L_2TAG_PACKET_5_0_2);
  movsd(Address(rsp, 0), xmm0);
  movsd(xmm0, Address(rsp, 128));
  fld_d(Address(rsp, 0));
  jmp(L_2TAG_PACKET_6_0_2);

  bind(L_2TAG_PACKET_7_0_2);
  cmpl(eax, 2146435072);
  jcc(Assembler::greaterEqual, L_2TAG_PACKET_8_0_2);
  movl(eax, Address(rsp, 132));
  cmpl(eax, INT_MIN);
  jcc(Assembler::greaterEqual, L_2TAG_PACKET_9_0_2);
  movsd(xmm0, Address(tmp, 1208));         // 0xffffffffUL, 0x7fefffffUL
  mulsd(xmm0, xmm0);
  movl(edx, 14);
  jmp(L_2TAG_PACKET_5_0_2);

  bind(L_2TAG_PACKET_9_0_2);
  movsd(xmm0, Address(tmp, 1216));
  mulsd(xmm0, xmm0);
  movl(edx, 15);
  jmp(L_2TAG_PACKET_5_0_2);

  bind(L_2TAG_PACKET_8_0_2);
  movl(edx, Address(rsp, 128));
  cmpl(eax, 2146435072);
  jcc(Assembler::above, L_2TAG_PACKET_10_0_2);
  cmpl(edx, 0);
  jcc(Assembler::notEqual, L_2TAG_PACKET_10_0_2);
  movl(eax, Address(rsp, 132));
  cmpl(eax, 2146435072);
  jcc(Assembler::notEqual, L_2TAG_PACKET_11_0_2);
  movsd(xmm0, Address(tmp, 1192));         // 0x00000000UL, 0x7ff00000UL
  jmp(L_2TAG_PACKET_2_0_2);

  bind(L_2TAG_PACKET_11_0_2);
  movsd(xmm0, Address(tmp, 1200));         // 0x00000000UL, 0x00000000UL
  jmp(L_2TAG_PACKET_2_0_2);

  bind(L_2TAG_PACKET_10_0_2);
  movsd(xmm0, Address(rsp, 128));
  addsd(xmm0, xmm0);
  jmp(L_2TAG_PACKET_2_0_2);

  bind(L_2TAG_PACKET_0_0_2);
  movl(eax, Address(rsp, 132));
  andl(eax, 2147483647);
  cmpl(eax, 1083179008);
  jcc(Assembler::aboveEqual, L_2TAG_PACKET_7_0_2);
  movsd(xmm0, Address(rsp, 128));
  addsd(xmm0, Address(tmp, 1184));         // 0x00000000UL, 0x3ff00000UL
  jmp(L_2TAG_PACKET_2_0_2);

  bind(L_2TAG_PACKET_2_0_2);
  movsd(Address(rsp, 48), xmm0);
  fld_d(Address(rsp, 48));

  bind(L_2TAG_PACKET_6_0_2);
  movl(tmp, Address(rsp, 64));
}
void MacroAssembler::fast_exp(XMMRegister xmm0, XMMRegister xmm1, XMMRegister xmm2, XMMRegister xmm3, XMMRegister xmm4, XMMRegister xmm5, XMMRegister xmm6, XMMRegister xmm7, Register eax, Register ecx, Register edx, Register tmp) {
  Label L_2TAG_PACKET_0_0_2, L_2TAG_PACKET_1_0_2, L_2TAG_PACKET_2_0_2, L_2TAG_PACKET_3_0_2;
  Label L_2TAG_PACKET_4_0_2, L_2TAG_PACKET_5_0_2, L_2TAG_PACKET_6_0_2, L_2TAG_PACKET_7_0_2;
  Label L_2TAG_PACKET_8_0_2, L_2TAG_PACKET_9_0_2, L_2TAG_PACKET_10_0_2, L_2TAG_PACKET_11_0_2;
  Label L_2TAG_PACKET_12_0_2, B1_3, B1_5, start;

  assert_different_registers(tmp, eax, ecx, edx);
  jmp(start);
  address cv = (address)_cv;
  address Shifter = (address)_shifter;
  address mmask = (address)_mmask;
  address bias = (address)_bias;
  address Tbl_addr = (address)_Tbl_addr;
  address ALLONES = (address)_ALLONES;
  address ebias = (address)_ebias;
  address XMAX = (address)_XMAX;
  address XMIN = (address)_XMIN;
  address INF = (address)_INF;
  address ZERO = (address)_ZERO;
  address ONE_val = (address)_ONE_val;

  bind(start);
  subq(rsp, 24);
  movsd(Address(rsp, 8), xmm0);
  unpcklpd(xmm0, xmm0);
  movdqu(xmm1, ExternalAddress(cv));       // 0x652b82feUL, 0x40571547UL, 0x652b82feUL, 0x40571547UL
  movdqu(xmm6, ExternalAddress(Shifter));  // 0x00000000UL, 0x43380000UL, 0x00000000UL, 0x43380000UL
  movdqu(xmm2, ExternalAddress(16+cv));    // 0xfefa0000UL, 0x3f862e42UL, 0xfefa0000UL, 0x3f862e42UL
  movdqu(xmm3, ExternalAddress(32+cv));    // 0xbc9e3b3aUL, 0x3d1cf79aUL, 0xbc9e3b3aUL, 0x3d1cf79aUL
  pextrw(eax, xmm0, 3);
  andl(eax, 32767);
  movl(edx, 16527);
  subl(edx, eax);
  subl(eax, 15504);
  orl(edx, eax);
  cmpl(edx, INT_MIN);
  jcc(Assembler::aboveEqual, L_2TAG_PACKET_0_0_2);
  mulpd(xmm1, xmm0);
  addpd(xmm1, xmm6);
  movapd(xmm7, xmm1);
  subpd(xmm1, xmm6);
  mulpd(xmm2, xmm1);
  movdqu(xmm4, ExternalAddress(64+cv));    // 0xe3289860UL, 0x3f56c15cUL, 0x555b9e25UL, 0x3fa55555UL
  mulpd(xmm3, xmm1);
  movdqu(xmm5, ExternalAddress(80+cv));    // 0xc090cf0fUL, 0x3f811115UL, 0x55548ba1UL, 0x3fc55555UL
  subpd(xmm0, xmm2);
  movdl(eax, xmm7);
  movl(ecx, eax);
  andl(ecx, 63);
  shll(ecx, 4);
  sarl(eax, 6);
  movl(edx, eax);
  movdqu(xmm6, ExternalAddress(mmask));    // 0xffffffc0UL, 0x00000000UL, 0xffffffc0UL, 0x00000000UL
  pand(xmm7, xmm6);
  movdqu(xmm6, ExternalAddress(bias));     // 0x0000ffc0UL, 0x00000000UL, 0x0000ffc0UL, 0x00000000UL
  paddq(xmm7, xmm6);
  psllq(xmm7, 46);
  subpd(xmm0, xmm3);
  lea(tmp, ExternalAddress(Tbl_addr));
  movdqu(xmm2, Address(ecx,tmp));
  mulpd(xmm4, xmm0);
  movapd(xmm6, xmm0);
  movapd(xmm1, xmm0);
  mulpd(xmm6, xmm6);
  mulpd(xmm0, xmm6);
  addpd(xmm5, xmm4);
  mulsd(xmm0, xmm6);
  mulpd(xmm6, ExternalAddress(48+cv));     // 0xfffffffeUL, 0x3fdfffffUL, 0xfffffffeUL, 0x3fdfffffUL
  addsd(xmm1, xmm2);
  unpckhpd(xmm2, xmm2);
  mulpd(xmm0, xmm5);
  addsd(xmm1, xmm0);
  por(xmm2, xmm7);
  unpckhpd(xmm0, xmm0);
  addsd(xmm0, xmm1);
  addsd(xmm0, xmm6);
  addl(edx, 894);
  cmpl(edx, 1916);
  jcc (Assembler::above, L_2TAG_PACKET_1_0_2);
  mulsd(xmm0, xmm2);
  addsd(xmm0, xmm2);
  jmp (B1_5);

  bind(L_2TAG_PACKET_1_0_2);
  xorpd(xmm3, xmm3);
  movdqu(xmm4, ExternalAddress(ALLONES));  // 0xffffffffUL, 0xffffffffUL, 0xffffffffUL, 0xffffffffUL
  movl(edx, -1022);
  subl(edx, eax);
  movdl(xmm5, edx);
  psllq(xmm4, xmm5);
  movl(ecx, eax);
  sarl(eax, 1);
  pinsrw(xmm3, eax, 3);
  movdqu(xmm6, ExternalAddress(ebias));    // 0x00000000UL, 0x3ff00000UL, 0x00000000UL, 0x3ff00000UL
  psllq(xmm3, 4);
  psubd(xmm2, xmm3);
  mulsd(xmm0, xmm2);
  cmpl(edx, 52);
  jcc(Assembler::greater, L_2TAG_PACKET_2_0_2);
  pand(xmm4, xmm2);
  paddd(xmm3, xmm6);
  subsd(xmm2, xmm4);
  addsd(xmm0, xmm2);
  cmpl(ecx, 1023);
  jcc(Assembler::greaterEqual, L_2TAG_PACKET_3_0_2);
  pextrw(ecx, xmm0, 3);
  andl(ecx, 32768);
  orl(edx, ecx);
  cmpl(edx, 0);
  jcc(Assembler::equal, L_2TAG_PACKET_4_0_2);
  movapd(xmm6, xmm0);
  addsd(xmm0, xmm4);
  mulsd(xmm0, xmm3);
  pextrw(ecx, xmm0, 3);
  andl(ecx, 32752);
  cmpl(ecx, 0);
  jcc(Assembler::equal, L_2TAG_PACKET_5_0_2);
  jmp(B1_5);

  bind(L_2TAG_PACKET_5_0_2);
  mulsd(xmm6, xmm3);
  mulsd(xmm4, xmm3);
  movdqu(xmm0, xmm6);
  pxor(xmm6, xmm4);
  psrad(xmm6, 31);
  pshufd(xmm6, xmm6, 85);
  psllq(xmm0, 1);
  psrlq(xmm0, 1);
  pxor(xmm0, xmm6);
  psrlq(xmm6, 63);
  paddq(xmm0, xmm6);
  paddq(xmm0, xmm4);
  movl(Address(rsp,0), 15);
  jmp(L_2TAG_PACKET_6_0_2);

  bind(L_2TAG_PACKET_4_0_2);
  addsd(xmm0, xmm4);
  mulsd(xmm0, xmm3);
  jmp(B1_5);

  bind(L_2TAG_PACKET_3_0_2);
  addsd(xmm0, xmm4);
  mulsd(xmm0, xmm3);
  pextrw(ecx, xmm0, 3);
  andl(ecx, 32752);
  cmpl(ecx, 32752);
  jcc(Assembler::aboveEqual, L_2TAG_PACKET_7_0_2);
  jmp(B1_5);

  bind(L_2TAG_PACKET_2_0_2);
  paddd(xmm3, xmm6);
  addpd(xmm0, xmm2);
  mulsd(xmm0, xmm3);
  movl(Address(rsp,0), 15);
  jmp(L_2TAG_PACKET_6_0_2);

  bind(L_2TAG_PACKET_8_0_2);
  cmpl(eax, 2146435072);
  jcc(Assembler::aboveEqual, L_2TAG_PACKET_9_0_2);
  movl(eax, Address(rsp,12));
  cmpl(eax, INT_MIN);
  jcc(Assembler::aboveEqual, L_2TAG_PACKET_10_0_2);
  movsd(xmm0, ExternalAddress(XMAX));      // 0xffffffffUL, 0x7fefffffUL
  mulsd(xmm0, xmm0);

  bind(L_2TAG_PACKET_7_0_2);
  movl(Address(rsp,0), 14);
  jmp(L_2TAG_PACKET_6_0_2);

  bind(L_2TAG_PACKET_10_0_2);
  movsd(xmm0, ExternalAddress(XMIN));      // 0x00000000UL, 0x00100000UL
  mulsd(xmm0, xmm0);
  movl(Address(rsp,0), 15);
  jmp(L_2TAG_PACKET_6_0_2);

  bind(L_2TAG_PACKET_9_0_2);
  movl(edx, Address(rsp,8));
  cmpl(eax, 2146435072);
  jcc(Assembler::above, L_2TAG_PACKET_11_0_2);
  cmpl(edx, 0);
  jcc(Assembler::notEqual, L_2TAG_PACKET_11_0_2);
  movl(eax, Address(rsp,12));
  cmpl(eax, 2146435072);
  jcc(Assembler::notEqual, L_2TAG_PACKET_12_0_2);
  movsd(xmm0, ExternalAddress(INF));       // 0x00000000UL, 0x7ff00000UL
  jmp(B1_5);

  bind(L_2TAG_PACKET_12_0_2);
  movsd(xmm0, ExternalAddress(ZERO));      // 0x00000000UL, 0x00000000UL
  jmp(B1_5);

  bind(L_2TAG_PACKET_11_0_2);
  movsd(xmm0, Address(rsp, 8));
  addsd(xmm0, xmm0);
  jmp(B1_5);

  bind(L_2TAG_PACKET_0_0_2);
  movl(eax, Address(rsp, 12));
  andl(eax, 2147483647);
  cmpl(eax, 1083179008);
  jcc(Assembler::aboveEqual, L_2TAG_PACKET_8_0_2);
  movsd(Address(rsp, 8), xmm0);
  addsd(xmm0, ExternalAddress(ONE_val));   // 0x00000000UL, 0x3ff00000UL
  jmp(B1_5);

  bind(L_2TAG_PACKET_6_0_2);
  movq(Address(rsp, 16), xmm0);

  bind(B1_3);
  movq(xmm0, Address(rsp, 16));

  bind(B1_5);
  addq(rsp, 24);
}
void MacroAssembler::fast_log(XMMRegister xmm0, XMMRegister xmm1, XMMRegister xmm2, XMMRegister xmm3, XMMRegister xmm4, XMMRegister xmm5, XMMRegister xmm6, XMMRegister xmm7, Register eax, Register ecx, Register edx, Register tmp) {
  Label L_2TAG_PACKET_0_0_2, L_2TAG_PACKET_1_0_2, L_2TAG_PACKET_2_0_2, L_2TAG_PACKET_3_0_2;
  Label L_2TAG_PACKET_4_0_2, L_2TAG_PACKET_5_0_2, L_2TAG_PACKET_6_0_2, L_2TAG_PACKET_7_0_2;
  Label L_2TAG_PACKET_8_0_2, L_2TAG_PACKET_9_0_2;
  Label L_2TAG_PACKET_10_0_2, start;

  assert_different_registers(tmp, eax, ecx, edx);
  jmp(start);
  address static_const_table = (address)_static_const_table_log;

  bind(start);
  subl(rsp, 104);
  movl(Address(rsp, 40), tmp);
  lea(tmp, ExternalAddress(static_const_table));
  xorpd(xmm2, xmm2);
  movl(eax, 16368);
  pinsrw(xmm2, eax, 3);
  xorpd(xmm3, xmm3);
  movl(edx, 30704);
  pinsrw(xmm3, edx, 3);
  movsd(xmm0, Address(rsp, 112));
  movapd(xmm1, xmm0);
  movl(ecx, 32768);
  movdl(xmm4, ecx);
  movsd(xmm5, Address(tmp, 2128));         // 0x00000000UL, 0xffffe000UL
  pextrw(eax, xmm0, 3);
  por(xmm0, xmm2);
  psllq(xmm0, 5);
  movl(ecx, 16352);
  psrlq(xmm0, 34);
  rcpss(xmm0, xmm0);
  psllq(xmm1, 12);
  pshufd(xmm6, xmm5, 228);
  psrlq(xmm1, 12);
  subl(eax, 16);
  cmpl(eax, 32736);
  jcc(Assembler::aboveEqual, L_2TAG_PACKET_0_0_2);

  bind(L_2TAG_PACKET_1_0_2);
  paddd(xmm0, xmm4);
  por(xmm1, xmm3);
  movdl(edx, xmm0);
  psllq(xmm0, 29);
  pand(xmm5, xmm1);
  pand(xmm0, xmm6);
  subsd(xmm1, xmm5);
  mulpd(xmm5, xmm0);
  andl(eax, 32752);
  subl(eax, ecx);
  cvtsi2sdl(xmm7, eax);
  mulsd(xmm1, xmm0);
  movsd(xmm6, Address(tmp, 2064));         // 0xfefa3800UL, 0x3fa62e42UL
  movdqu(xmm3, Address(tmp, 2080));        // 0x92492492UL, 0x3fc24924UL, 0x00000000UL, 0xbfd00000UL
  subsd(xmm5, xmm2);
  andl(edx, 16711680);
  shrl(edx, 12);
  movdqu(xmm0, Address(tmp, edx));
  movdqu(xmm4, Address(tmp, 2096));        // 0x3d6fb175UL, 0xbfc5555eUL, 0x55555555UL, 0x3fd55555UL
  addsd(xmm1, xmm5);
  movdqu(xmm2, Address(tmp, 2112));        // 0x9999999aUL, 0x3fc99999UL, 0x00000000UL, 0xbfe00000UL
  mulsd(xmm6, xmm7);
  pshufd(xmm5, xmm1, 68);
  mulsd(xmm7, Address(tmp, 2072));         // 0x93c76730UL, 0x3ceef357UL, 0x92492492UL, 0x3fc24924UL
  mulsd(xmm3, xmm1);
  addsd(xmm0, xmm6);
  mulpd(xmm4, xmm5);
  mulpd(xmm5, xmm5);
  pshufd(xmm6, xmm0, 228);
  addsd(xmm0, xmm1);
  addpd(xmm4, xmm2);
  mulpd(xmm3, xmm5);
  subsd(xmm6, xmm0);
  mulsd(xmm4, xmm1);
  pshufd(xmm2, xmm0, 238);
  addsd(xmm1, xmm6);
  mulsd(xmm5, xmm5);
  addsd(xmm7, xmm2);
  addpd(xmm4, xmm3);
  addsd(xmm1, xmm7);
  mulpd(xmm4, xmm5);
  addsd(xmm1, xmm4);
  pshufd(xmm5, xmm4, 238);
  addsd(xmm1, xmm5);
  addsd(xmm0, xmm1);
  jmp(L_2TAG_PACKET_2_0_2);

  bind(L_2TAG_PACKET_0_0_2);
  movsd(xmm0, Address(rsp, 112));
  movdqu(xmm1, xmm0);
  addl(eax, 16);
  cmpl(eax, 32768);
  jcc(Assembler::aboveEqual, L_2TAG_PACKET_3_0_2);
  cmpl(eax, 16);
  jcc(Assembler::below, L_2TAG_PACKET_4_0_2);

  bind(L_2TAG_PACKET_5_0_2);
  addsd(xmm0, xmm0);
  jmp(L_2TAG_PACKET_2_0_2);

  bind(L_2TAG_PACKET_6_0_2);
  jcc(Assembler::above, L_2TAG_PACKET_5_0_2);
  cmpl(edx, 0);
  jcc(Assembler::above, L_2TAG_PACKET_5_0_2);
  jmp(L_2TAG_PACKET_7_0_2);

  bind(L_2TAG_PACKET_3_0_2);
  movdl(edx, xmm1);
  psrlq(xmm1, 32);
  movdl(ecx, xmm1);
  addl(ecx, ecx);
  cmpl(ecx, -2097152);
  jcc(Assembler::aboveEqual, L_2TAG_PACKET_6_0_2);
  orl(edx, ecx);
  cmpl(edx, 0);
  jcc(Assembler::equal, L_2TAG_PACKET_8_0_2);

  bind(L_2TAG_PACKET_7_0_2);
  xorpd(xmm1, xmm1);
  xorpd(xmm0, xmm0);
  movl(eax, 32752);
  pinsrw(xmm1, eax, 3);
  movl(edx, 3);
  mulsd(xmm0, xmm1);

  bind(L_2TAG_PACKET_9_0_2);
  movsd(Address(rsp, 0), xmm0);
  movsd(xmm0, Address(rsp, 112));
  fld_d(Address(rsp, 0));
  jmp(L_2TAG_PACKET_10_0_2);

  bind(L_2TAG_PACKET_8_0_2);
  xorpd(xmm1, xmm1);
  xorpd(xmm0, xmm0);
  movl(eax, 49136);
  pinsrw(xmm0, eax, 3);
  divsd(xmm0, xmm1);
  movl(edx, 2);
  jmp(L_2TAG_PACKET_9_0_2);

  bind(L_2TAG_PACKET_4_0_2);
  movdl(edx, xmm1);
  psrlq(xmm1, 32);
  movdl(ecx, xmm1);
  orl(edx, ecx);
  cmpl(edx, 0);
  jcc(Assembler::equal, L_2TAG_PACKET_8_0_2);
  xorpd(xmm1, xmm1);
  movl(eax, 18416);
  pinsrw(xmm1, eax, 3);
  mulsd(xmm0, xmm1);
  movapd(xmm1, xmm0);
  pextrw(eax, xmm0, 3);
  por(xmm0, xmm2);
  psllq(xmm0, 5);
  movl(ecx, 18416);
  psrlq(xmm0, 34);
  rcpss(xmm0, xmm0);
  psllq(xmm1, 12);
  pshufd(xmm6, xmm5, 228);
  psrlq(xmm1, 12);
  jmp(L_2TAG_PACKET_1_0_2);

  bind(L_2TAG_PACKET_2_0_2);
  movsd(Address(rsp, 24), xmm0);
  fld_d(Address(rsp, 24));

  bind(L_2TAG_PACKET_10_0_2);
  movl(tmp, Address(rsp, 40));
}