Re: [PATCH v2][malloc] Use relaxed atomics for malloc have_fastchunks

[Joseph Myers <joseph at codesourcery dot com>]

On Wed, 18 Oct 2017, Chris Metcalf wrote:

> To be fair, we could also synthesize subword atomics for tilepro with compare
> and swap on the surrounding word, so it's not technically impossible to add
> this support. And (as Wilco already demonstrated) it's trivial to fix the
> caller to work around this problem.

I'm pretty sure other architectures in glibc would also require subword 
atomics to be synthesized, and some certainly have errors for them at 
present - what's unique about tilepro is that the checks trigger on 
subword relaxed atomic load/store (whereas more architectures would have 
problems with non-relaxed atomics, compre/exchange, etc.).

> However, it's also true that tilepro is no longer an actively-maintained 
> platform; the tilegx platform has pretty well supplanted it in practice. 
> Perhaps it is time to look at obsoleting tilepro so that its presence in 
> the tree doesn't block forward progress for other architectures.  I'm 
> not aware of any existing tilepro customers that are looking to use 
> newer versions of glibc.

Certainly architecture maintainers can propose removal / obsoletion of 
their architectures (or subarchitecture variants, etc.), and in the 
absence of other people expressing interest in using that architecture I'd 
expect such removal / obsoletion proposals to achieve consensus.

-- 
Joseph S. Myers
joseph@codesourcery.com