Re: incorrect use of 'pure' attribute

Paul Eggert <[email protected]> writes:

> On 8/7/26 19:36, Sam James wrote:
>> __eloop_threshold isn't const and a comment
>> acknowledges this.
>
> Actually, the comment says that __eloop_threshold is validly marked
> 'const' even though it uses memory and calls a function, because its
> return value is unaffected by the memory use and the function call.
>
> If the comment is incorrect it'd be interesting to know why.

I'd attempted to explain that already: the attribute doesn't promise
that a function is idempotent or has 'call_once' semantics. It promises
beyond that, that it can't read global meomry in a way that affects what
it returns. In the case of Mesa, it was actually the return which
mattered, but the same thing as true anyway per the GCC bug.

The comment simply isn't right and it makes an assertion that doesn't
reflect reality. It is not validly const if it uses global memory to
affect what it returns, even if later calls are OK.

I gave some examples of two projects where this broke already and it may
well break in future too. Both of those had a fat comment saying exactly
the same thing as gnulib does for this.

thanks,
sam

-----BEGIN PGP SIGNATURE-----

iQEBBAEWCgCpFiEEJaa7iN2bdkxrVUHCc4QJ9SDfkZAFAmp3BzwbFIAAAAAABAAO
bWFudTIsMi41KzEuMTIsMiwyXxSAAAAAAC4AKGlzc3Vlci1mcHJAbm90YXRpb25z
Lm9wZW5wZ3AuZmlmdGhob3JzZW1hbi5uZXQyNUE2QkI4OEREOUI3NjRDNkI1NTQx
QzI3Mzg0MDlGNTIwREY5MTkwDxxzYW1AZ2VudG9vLm9yZwAKCRBzhAn1IN+RkJGU
AQCZyWjrr6Um+h8CzzzVFhhsTjc3uR+pZdQOTkkhEnuxuAD+LcqjOaDhdTbqXosx
w0+pkInAJkhY84AuNZtBgDfxZAc=
=a4Ll
-----END PGP SIGNATURE-----