memcpy madness

[SMurf]

Hello,

I'm copying some memory from one region to another. As I'm dealing with 32-bit colour data, the area of each memory region is a multiple of 4, so I was wondering whether memcpy(...), which copies a single byte at a time, is best optimised for the task at hand.

May I be better off using a bit of inline asm involving MOVSD (Move a doubleword string)? I gather there's been a bit of argument about this on the net, as some people say that the some CPUs doesn't always do it faster.

[Prelude]

Try it with memcpy first, then once you have that working and want it to be faster you can try to improve the speed. Simplicity and clarity come first, then performance if it's truly needed.

[anonytmouse]

which copies a single byte at a time

Why? I think memcpy, memset, etc would use native word size as much as possible and unroll the loop.

Rough example:

nCount = nBytes >>  2;  // = /4 maybe.
while (nCount--) {
     *destint = *fromint;
}

nCount = nBytes & 0x3; // = % 4 maybe.
while (nCount--) {
     *destchar = *fromchar;
}

This could be optimised in C by unrolling the loop with a Duff construct.