Hi!

On Thu, Apr 20, 2023 at 07:23:39PM +0200, Alejandro Colomar wrote:
> On 4/20/23 17:35, наб wrote:
> > +.I errcode
> > +must be the latest error returned from an operation on
> > +.IR preg .
> > +If
> > +.I preg
> > +is a null pointer\(emthe latest error.
> I don't read that from the POSIX spec.
Whereas that's precisely where I got it from.

> If preg is NULL, then I think any
> error returned by a call to one of these APIs would be valid.
That's unspecified.

> In fact,
> since these functions are MT-Safe, they can't store any state,
Probably. OTOH, musl raw-dogs mbtowc() in regexec(), so.
(I'm pretty sure it's by accident since they do have a mbstate_t
 and juggle it a lot, but it's never actually used.)

> which leads
> me to think that they can't really distinguish between the latest error,
> and an error returned at a random point in the past, or even the result of
> csrand_interval(x, y)[1] with appropriate x and y.
Again, probably. But (line numbers from Issue 8 Draft 2.1):
57517  The regerror( ) function provides a mapping from error codes returned by regcomp( ) and
57518  regexec( ) to unspecified printable strings. It generates a string corresponding to the value of the
57519  errcode argument, which the application shall ensure is the last non-zero value returned by
57520  regcomp( ) or regexec( ) with the given value of preg. If errcode is not such a value, the content of
57521  the generated string is unspecified.

57522  If preg is a null pointer, but errcode is a value returned by a previous call to regexec( ) or regcomp( ),
57523  the regerror( ) still generates an error string corresponding to the value of errcode, but it might not
57524  be as detailed under some implementations.

57525  If the errbuf_size argument is not 0, regerror( ) shall place the generated string into the buffer of
57526  size errbuf_size bytes pointed to by errbuf. If the string (including the terminating null) cannot fit
57527  in the buffer, regerror( ) shall truncate the string and null-terminate the result.

57528  If errbuf_size is 0, regerror( ) shall ignore the errbuf argument, and return the size of the buffer
57529  needed to hold the generated string.

In these difficult times I tend to turn to what implementations do:
NetBSD, musl, illumos, and glibc, if you subtract REG_ATOI and REG_ITOA,
all essentially return lsearch(errors, errcode)->description
+ all sans NetBSD localise it.
None of them even use preg.

So yeah, I'll axe that.


And split out regfree() from this patch because I missed it.


Best,
наб