Re: [PATCH] mips: work-around R10k ll/sc errata

["Joseph S. Myers" <joseph at codesourcery dot com>]

On Tue, 28 Jun 2011, Matt Turner wrote:

> This was sent back in Nov 2008 by Joshua Kinard, but seems to have
> fallen through the cracks after a lot of bikeshedding.
> 
> I don't particularly care about implementing alternative work-arounds,
> like 28 nops. I only intend for this patch to be useful when the builder
> is using -march=r10000 and/or -mfix-r10000. All other situations are
> unaffected.

I think the discussion really illustrates that glibc isn't the right place 
to be making this decision; there are a lot of different compiler options 
for different MIPS variants, and I don't know if they all have relevant 
preprocessor macros, and the compiler already makes its own decisions 
based on those options about how to implement __sync_*.  There are also 
various out-of-tree patches to the MIPS atomics in glibc to optimize them 
for some processor variants.

The real fix is to use __sync_* for MIPS in glibc, as done for some other 
architectures.  But for this not to be an optimization regression you need 
versions of those functions with more specific barrier semantics.  Tom, 
you had such functions (MIPS-specific) and associated glibc changes.  A 
proper set of target-independent such functions has recently been 
discussed on gcc-patches by Aldy Hernandez, Andrew MacLeod et al; once the 
generic framework is in GCC trunk, perhaps you could contribute a version 
of your MIPS changes that builds on the generic framework, and then a 
version of the glibc patches that uses the new generic builtins?  The 
existing assembly implementations can then be left just for GCC <= 4.6, 
and not complicated by lots of processor conditionals.

-- 
Joseph S. Myers
joseph@codesourcery.com