Re: please keep 'memmem' module simple

[Simon Josefsson]

Eric Blake <address@hidden> writes:

> Simon Josefsson <simon <at> josefsson.org> writes:
>
>> >     return !result || !memmem ("a", 1, 0, 0);]])],
>> >
>> Ah, thanks.  This was not part of the old memmem.m4 checks.  Still, the
>> glibc documentation (which I regard as the memmem specification) doesn't
>> say what should be returned for empty needles.  My conclusion is that
>> empty needles should lead to undefined behaviour.
>
> Fair enough for the memmem-simple module.  But since POSIX requires that 0-
> length needles in strstr(ptr,"") return ptr, and I consider that the 
> similarity 
> between memmem and strstr is essential, (in part so someone could do a naive 
> implementation of strstr as:
>  return memmem (haystack, strlen (haystack), needle, strlen (needle));
> and in part because if the haystack length is known in advance, memmem can be 
> more efficient than strstr).  Therefore, the full memmem module should 
> enforce 
> an empty needle as well-defined behavior.  (Or, put another way, I 
> want 'echo "index(abc,)"|m4' to work without me having to special case 
> whether 
> the second argument to index is empty).

I hadn't considered the strstr similarity here.  I think that is a
fairly strong argument that empty needles should return pointers too
even for memmem.  But until the documentation actually states it, I
think people need to be careful about empty needles anyway.

>> Unfortunately, cross-compilation is in practice the typical scenario
>> where people care about size issues, so I would prefer to rely on
>> AC_CHECK_FUNC instead.
>
> OK then, based on this email, the gl_FUNC_MEMMEM_SIMPLE macro uses 
> AC_CHECK_FUNC when cross-compiling, while gl_FUNC_MEMMEM is pessimistic and 
> always uses the replacement (unless you did primed the cache with 
> gl_cv_func_memmem_works=yes first).  Clients of memmem-simple must avoid 
> empty 
> needles, and should avoid arbitrary needles, and as a result get smaller code 
> size on glibc and cygwin.

Right.

> Here's what I'm committing (I reversed your proposed macro dependency - the 
> full module should depend on and augment the simple check, not the other way 
> around):

Excellent.  Thank you very much for discussing and addressing this.
I'll update gnutls and will report back any problems.

Regards,
/Simon