>   in this case, does someone know a fast memcpy that supports overlap ?
Maybe you could try this, I just hacked it up, haven't tried it yet, but it
is pretty straight forward:

*** locore.s-orig	Fri Oct 14 09:22:28 1994
--- locore.s	Fri Oct 14 09:26:13 1994
***************
*** 1794,1805 ****
--- 1794,1813 ----
   *
   * Works for counts up to 128K.
   */
+ ALTENTRY(ovmemcpy, _memcpy)
+ ENTRY(memcpy)
+ 	movl	sp@(12),d0		| get count
+ 	jeq	Lcpyexit		| if zero, return
+ 	movel	sp@(4),a1		| dest address
+ 	movel	sp@(8),a0		| src address
+ 	jra	Lcpyentry
  ALTENTRY(ovbcopy, _bcopy)
  ENTRY(bcopy)
  	movl	sp@(12),d0		| get count
  	jeq	Lcpyexit		| if zero, return
  	movl	sp@(4),a0		| src address
  	movl	sp@(8),a1		| dest address
+ Lcpyentry:
  	cmpl	a1,a0			| src before dest?
  	jlt	Lcpyback		| yes, copy backwards (avoids overlap)
  	movl	a0,d1

thus using the pre-existing and already linking bcopy() code.
>
>I am asking this because I feel that it makes no sense to stick to
>gcc 2.4.5 forever ...
>
Here! Here! While we're at it, I am working on getting gas-2.3 working.
How 'bout lets send fixes for gnu stuff into the gnu folks so they can
update it in future releases. That way updates stand a very good chance of
working, people that like to keep up with the latest release (which usually
means bug fixes and/or enhancements that are usually useful), only need to
get the stuff and compile, no re-porting hassles, no duplication of effort.

Errr... sorry about the soapbox bit, just gets old redoing what people have
already done, usually many times over.