public static IntPtr Set(byte *dest, int c, long count)
{
Debug.Assert(count >= 0);
return(PlatformDetails.RunningOnPosix
? Syscall.Set(dest, c, count)
: Win32UnmanagedMemory.Set(dest, c, count));
}
public static int Move(byte *dest, byte *src, long count)
{
Debug.Assert(count >= 0);
return(PlatformDetails.RunningOnPosix
? Syscall.Move(dest, src, count)
: Win32UnmanagedMemory.Move(dest, src, count));
}
public static int Compare(byte *b1, byte *b2, long count)
{
Debug.Assert(count >= 0);
return(PlatformDetails.RunningOnPosix
? Syscall.Compare(b1, b2, count)
: Win32UnmanagedMemory.Compare(b1, b2, count));
}
public static void Move(byte *dest, byte *src, int n)
{
// if dest and src overlaps, we need to call specifically to memmove pinvoke supporting overlapping
if (dest + n >= src &&
src + n >= dest)
{
var _ = PlatformDetails.RunningOnPosix
? Syscall.Move(dest, src, n)
: Win32UnmanagedMemory.Move(dest, src, n);
return;
}
Copy(dest, src, n);
}
public static int CompareInline(void *p1, void *p2, int size)
{
// If we use an unmanaged bulk version with an inline compare the caller site does not get optimized properly.
// If you know you will be comparing big memory chunks do not use the inline version.
if (size > CompareInlineVsCallThreshold)
{
goto UnmanagedCompare;
}
byte *bpx = (byte *)p1;
byte *bpy = (byte *)p2;
// PERF: This allows us to do pointer arithmetics and use relative addressing using the
// hardware instructions without needed an extra register.
long offset = bpy - bpx;
if ((size & 7) == 0)
{
goto ProcessAligned;
}
// We process first the "unaligned" size.
ulong xor;
if ((size & 4) != 0)
{
xor = *((uint *)bpx) ^ *((uint *)(bpx + offset));
if (xor != 0)
{
goto Tail;
}
bpx += 4;
}
if ((size & 2) != 0)
{
xor = (ulong)(*((ushort *)bpx) ^ *((ushort *)(bpx + offset)));
if (xor != 0)
{
goto Tail;
}
bpx += 2;
}
if ((size & 1) != 0)
{
int value = *bpx - *(bpx + offset);
if (value != 0)
{
return(value);
}
bpx += 1;
}
ProcessAligned:
byte *end = (byte *)p1 + size;
byte *loopEnd = end - 16;
while (bpx <= loopEnd)
{
// PERF: JIT will emit: ```{op} {reg}, qword ptr [rdx+rax]```
if (*((ulong *)bpx) != *(ulong *)(bpx + offset))
{
goto XorTail;
}
if (*((ulong *)(bpx + 8)) != *(ulong *)(bpx + 8 + offset))
{
bpx += 8;
goto XorTail;
}
bpx += 16;
}
if (bpx < end)
{
goto XorTail;
}
return(0);
XorTail : xor = *((ulong *)bpx) ^ *(ulong *)(bpx + offset);
Tail:
// Fast-path for equals
if (xor == 0)
{
return(0);
}
// PERF: This is a bit twiddling hack. Given that bitwise xoring 2 values flag the bits difference,
// we can use that we know we are running on little endian hardware and the very first bit set
// will correspond to the first byte which is different.
bpx += Bits.TrailingZeroesInBytes(xor);
return(*bpx - *(bpx + offset));
UnmanagedCompare:
return(PlatformDetails.RunningOnPosix
? Syscall.Compare((byte *)p1, (byte *)p2, size)
: Win32UnmanagedMemory.Compare((byte *)p1, (byte *)p2, size));
}
public static IntPtr Copy(byte *dest, byte *src, int count)
{
return(Platform.RunningOnPosix
? PosixUnmanagedMemory.Copy(dest, src, count)
: Win32UnmanagedMemory.Copy(dest, src, count));
}
public static IntPtr Set(byte *dest, int c, int count)
{
return(Platform.RunningOnPosix
? PosixUnmanagedMemory.Set(dest, c, count)
: Win32UnmanagedMemory.Set(dest, c, count));
}
public static int Move(byte *b1, byte *b2, int count)
{
return(Platform.RunningOnPosix
? PosixUnmanagedMemory.Move(b1, b2, count)
: Win32UnmanagedMemory.Move(b1, b2, count));
}
