protected override void Compile(CodeContext c)
{
var dst = c.IntPtr("dst");
var src = c.IntPtr("src");
var i = c.IntPtr("i");
var j = c.IntPtr("j");
var t = c.IntPtr("t");
var cZero = c.Xmm("cZero");
var cMul255A = c.Xmm("cMul255A");
var cMul255M = c.Xmm("cMul255M");
var x0 = c.Xmm("x0");
var x1 = c.Xmm("x1");
var y0 = c.Xmm("y0");
var a0 = c.Xmm("a0");
var a1 = c.Xmm("a1");
var smallLoop = c.Label();
var smallEnd = c.Label();
var largeLoop = c.Label();
var largeEnd = c.Label();
var data = c.Label();
c.SetArgument(0, dst);
c.SetArgument(1, src);
c.SetArgument(2, i);
c.Allocate(dst);
c.Allocate(src);
c.Allocate(i);
// How many pixels have to be processed to make the loop aligned.
c.Lea(t, Memory.Ptr(data));
c.Xor(j, j);
c.Xorps(cZero, cZero);
c.Sub(j, dst);
c.Movaps(cMul255A, Memory.Ptr(t, 0));
c.And(j, 15);
c.Movaps(cMul255M, Memory.Ptr(t, 16));
c.Shr(j, 2);
c.Jz(smallEnd);
// j = min(i, j).
c.Cmp(j, i);
c.Cmovg(j, i);
// i -= j.
c.Sub(i, j);
// Small loop.
c.Bind(smallLoop);
c.Pcmpeqb(a0, a0);
c.Movd(y0, Memory.Ptr(src));
c.Pxor(a0, y0);
c.Movd(x0, Memory.Ptr(dst));
c.Psrlw(a0, 8);
c.Punpcklbw(x0, cZero);
c.Pshuflw(a0, a0, AsmJit.Common.Utils.Shuffle(1, 1, 1, 1));
c.Punpcklbw(y0, cZero);
c.Pmullw(x0, a0);
c.Paddsw(x0, cMul255A);
c.Pmulhuw(x0, cMul255M);
c.Paddw(x0, y0);
c.Packuswb(x0, x0);
c.Movd(Memory.Ptr(dst), x0);
c.Add(dst, 4);
c.Add(src, 4);
c.Dec(j);
c.Jnz(smallLoop);
// Second section, prepare for an aligned loop.
c.Bind(smallEnd);
c.Test(i, i);
c.Mov(j, i);
c.Jz(c.Exit);
c.And(j, 3);
c.Shr(i, 2);
c.Jz(largeEnd);
// Aligned loop.
c.Bind(largeLoop);
c.Movups(y0, Memory.Ptr(src));
c.Pcmpeqb(a0, a0);
c.Movaps(x0, Memory.Ptr(dst));
c.Xorps(a0, y0);
c.Movaps(x1, x0);
c.Psrlw(a0, 8);
c.Punpcklbw(x0, cZero);
c.Movaps(a1, a0);
c.Punpcklwd(a0, a0);
c.Punpckhbw(x1, cZero);
c.Punpckhwd(a1, a1);
c.Pshufd(a0, a0, AsmJit.Common.Utils.Shuffle(3, 3, 1, 1));
c.Pshufd(a1, a1, AsmJit.Common.Utils.Shuffle(3, 3, 1, 1));
c.Pmullw(x0, a0);
c.Pmullw(x1, a1);
c.Paddsw(x0, cMul255A);
c.Paddsw(x1, cMul255A);
c.Pmulhuw(x0, cMul255M);
c.Pmulhuw(x1, cMul255M);
c.Add(src, 16);
c.Packuswb(x0, x1);
c.Paddw(x0, y0);
c.Movaps(Memory.Ptr(dst), x0);
c.Add(dst, 16);
c.Dec(i);
c.Jnz(largeLoop);
c.Bind(largeEnd);
c.Test(j, j);
c.Jnz(smallLoop);
// Data
c.Data(data, 16,
Data.Of(0x0080008000800080, 0x0080008000800080),
Data.Of(0x0101010101010101, 0x0080008000800080));
}