U++ forum: Welcome to the forum

Re: Optimized memcmp for x86 [message #14335 is a reply to message #14308]

Fri, 22 February 2008 19:08

mr_ped
Messages: 825
Registered: November 2005
Location: Czech Republic - Praha

Experienced Contributor

luzr wrote on Fri, 22 February 2008 11:07

	if(len & 2)
		if(*(word *)x != *(word *)y)
			return int(_byteswap_ushort(*(word *)x) - _byteswap_ushort(*(word *)y));
	if(len & 1)
		return int(*((byte *)x + 2)) - int(*((byte *)y + 2));

I don't get this end.

switch (len & 3)
0: it looks ok to me.
1: the return int(*((byte *)x)) - int(*((byte *)y)); should be returned?
2: looks ok
3: looks ok

I would maybe try masking out unused bytes, but that would lead to read out of buffer boundary. Is it safe?
I mean something like this

   ...
   const static dword masks[4] = { 0x00000000, 0x000000FF, 0x0000FFFF, 0x00FFFFFF };  //Intel-like endian only!
   return int(_byteswap_ulong(*x & masks[len&3]) - _byteswap_ulong(*y & masks[len&3]));

I'm not sure I got the byteswap purpose correctly, but I think I got, so my code is probably ok (but I didn't test it).

Of course it reads beyond buffer end, so you need to know it will not raise exception or crash the application on target platform.

[Updated on: Fri, 22 February 2008 19:11]

Report message to a moderator