Re: [patch] A few things from 2.0.x (Was: Pre-2.1.80..)

Dr. Werner Fink (werner@suse.de)
Thu, 15 Jan 1998 15:40:11 +0100

>
> The "string-i486" part certainly still has bugs:
>
> > @@ -416,9 +416,11 @@
> >
> > #define __HAVE_ARCH_MEMCPY
> > #define memcpy(d,s,count) \
> > -(__builtin_constant_p(count) ? \
> > - __memcpy_c((d),(s),(count)) : \
> > - __memcpy_g((d),(s),(count)))
> > +(count == 0 \
> > + ? d \
> > + : __builtin_constant_p(count) \
> > + ? __memcpy_c((d),(s),(count)) \
> > + : __memcpy_g((d),(s),(count)))
> >
> > /*
> > * These ought to get tweaked to do some cache priming.
>
> The above makes the macro incorrectly use "count" twice. If count has side
> effects, that's a serious problem that is very hard to find. Same goes for
> the changes to memset().

Maybe something like:

------
extern inline void * __memcpy_c(void * to, const void * from, size_t n)
{
switch (n) {
case 0:
return to;

<maybe more code like in include/asm-i386/string.h>
}

switch (n % 4) {
case 0: return __memcpy_by4((d),(s),(count));
case 1: return __memcpy_g((d),(s),(count));
case 2: return __memcpy_by2((d),(s),(count));
case 3: return __memcpy_g((d),(s),(count));
}
}

#define __HAVE_ARCH_MEMCPY
#define memcpy(d,s,count)
(__builtin_constant_p(count) ? \
__memcpy_c((d),(s),(count)) : \
__memcpy_g((d),(s),(count)))
------

... and similar for memset. Remaining question is: does the compiler
optimize/remove unnecessary code?

>
> The rest looks ok, but I'll continue to consider string-i486 totally
> broken,

Werner