This is the mail archive of the libc-alpha@sourceware.org mailing list for the glibc project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: Disable 64-bit atomics for MIPS n32 [committed]


On Wed, 28 Jan 2015, Torvald Riegel wrote:

> On Wed, 2015-01-28 at 18:45 +0000, Joseph Myers wrote:
> > This patch disables use of 64-bit atomics for MIPS n32 to fix the
> > problems with unaligned semaphores.
> 
> That works.  I would probably preferred to just do this in the semaphore
> bits (ie, checking for __HAVE_64B_ATOMICS and _LP64, but both have the
> same effect for 2.21.

Carlos requested the minimal change in 
<https://sourceware.org/ml/libc-alpha/2015-01/msg00597.html>.

> I guess I'll have to add another flag for whether 64B atomics need to be
> naturally aligned or not.  I'm not yet sure how to express this best, so
> suggestions are welcome.  If we can say (for now) that we only support
> ILP32 and LP64, this might clarify the options we have -- I suspect that
> in most cases where 64B atomics will be useful, this will be for legacy
> code.
> 
> OTOH, I'm not sure whether non-naturally-aligned 64B atomics actually
> really help us anywhere.  So we could as well just require them to be
> naturally aligned, and use _LP64 to check whether alignment for legacy
> code is sufficient.

My view is: only use atomic operations on properly aligned values, define 
SEM_USE_64B_ATOMICS or similar to 1 if __HAVE_64B_ATOMICS && (defined 
_LP64 || !SHLIB_COMPAT (libpthread, GLIBC_2_0, GLIBC_2_22)) to reflect 
that new ILP32 ports added in 2.22 and later (such as AArch64 ILP32) 
should define their semaphores to be 8-byte-aligned.  (Presuming that such 
ports should indeed prefer to use the 64-bit atomics here.)  And maybe 
then allow ports to override that choice.

-- 
Joseph S. Myers
joseph@codesourcery.com


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]