Re: [musl] strcmp() guarantees and assumptions

[Pedro Falcato <pedro.falcato@gmail.com>]

On Sun, Jul 16, 2023 at 7:00 PM Robert Clausecker <fuz@fuz.su> wrote:
>
> Hi NRK,
>
> Thank you for your response.
>
> Am Sun, Jul 16, 2023 at 11:49:45PM +0600 schrieb NRK:
> > Hi Robert,
> >
> > > Or to phrase it differently, is the following a legal implementation of
> > > strcmp()?
> > >
> > >     int strcmp(char *a, char *b) {
> > >             size_t la = strlen(a), lb = strlen(b);
> > >
> > >             if (la != lb)
> > >                     return ((la > lb) - (lb > la));
> >                       ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
> >
> > I don't see how this can ever be a valid strcmp implementation. The
> > return value of the comparison functions must be about the first
> > mismatching byte, not about the string lengths.
> >
> > | The sign of a nonzero value returned by the comparison functions is
> > | determined by the sign of the difference between the values of the
> > | first pair of characters that differ in the objects being compared.
> >   ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
>
> Yes, sorry.  The code would have to be extended to call memcmp() on the
> common prefix in case there is a mismatch in length.  E.g.
>
>     if (la != lb)
>         return (memcmp(la, lb, la > lb ? lb + 1 : la + 1));
>
> > ref: https://port70.net/~nsz/c/c11/n1570.html#7.24.4p1
> >
> > > Or is it generally agreed upon that libc implementations support
> > > strcmp() calls on unterminated strings?
> >
> > memchr (since C11) has the following requirement:
> >
> > | The implementation shall behave as if it reads the characters
> > | sequentially and stops as soon as a matching character is found.
> >
> > I don't believe any such requirement exists for strcmp, so unless
> > someone proves otherwise, I'd say it's fair game for libc to assume that
> > the strings are nul-terminated.
>
> That's good to hear.  Any idea on the “what do existing libc
> implementations permit” bit?

Looks like it's permissive.
At the moment, musl does (non-SIMD, obviously) unsigned long loads *as
long as they're aligned* (you don't want to page fault! and reads
don't have side effects unless it's MMIO or something, and that's
non-standard) and does standard(tm) bit tricks to find null bytes in
that same word.

-- 
Pedro