This is the mail archive of the
mailing list for the glibc project.
Re: Dealing with target variance
* Joseph Myers:
> My preference is one file per property (or at most, group of extremely
> closely related properties), with sysdeps overrides in the usual fashion.
> That's what we do for a range of libm properties - consider
> nan-high-order-bit.h (overrides for hppa and mips),
> math-tests-trap-force.h (overrides for powerpc),
> fix-fp-int-convert-overflow.h (overrides for arm, csky, mips, s390), etc.,
> and it's easy to use a single "find" command to see what architectures
> override a given header.
The only concern I would have is something like generic → m68k →
coldfire, where coldfire goes back to the generic default. At that
point, it's no longer obvious in which direction the override works, you
need to look at the file contents, and there is risk that we end up with
the existing multitude of override mechanisms.
Can we make the default definition the one for future targets? Even if
it means scattering the tree with property headers for old targets? And
avoid using the Implies mechanism for such headers (adding further
copies which are technically unnecessary)?