Re: Implement C23 log2p1

[Joseph Myers <josmyers@redhat.com>]

On Mon, 15 Apr 2024, Carlos O'Donell wrote:

> This fails CI for glibc:
> https://patchwork.sourceware.org/project/glibc/patch/41a4a6c9-4942-a9be-9459-1aaeb44ffee8@redhat.com/

New ulps failures are expected for new libm functions.  I think it's 
appropriate for libm-test-ulps updates to be done by architecture 
maintainers or others testing on an architecture (not necessarily for 
every patch, sometimes such testers might batch updates and update 
libm-test-ulps for multiple architectures at a time), rather than 
expecting them all to be included in the initial patch or making such 
functions need a multi-stage patch process where the initial patch needs 
updating with more libm-test-ulps updates before commit.

(There is a reasonable argument for having a single libm-test-ulps file 
that lists expected error bounds by format (e.g. binary64, binary128) not 
type, notwithstanding that sometimes multiple implementations are used for 
a given function for a given floating-point format, and notwithstanding 
that sometimes the ulps for a given implementation may depend on e.g. 
whether the compiler contracts operations when building it.  But while we 
have the per-architecture files, I don't think we should expect 
new-function patches to update them for more than one architecture.)

-- 
Joseph S. Myers
josmyers@redhat.com