On Wed, Sep 2, 2009 at 09:54, Bernd Schmidt wrote:
> The atomic op builtins we have only support word-sized operations.  Bug
> #4747 is a testsuite failure where the compiler generates a call to a
> nonexistant libgcc function for atomic ops on shorts.
>
> This is primarily a problem with the testsuite - a patch got misapplied
> due to two identical blocks of code in lib/target-supports.exp and thus
> we were running the wrong testcases.  I've committed a fix for this.
>
> However, even without additional kernel support we could implement
> atomic ops for short and char by emulating them with word-sized cmpxchg.
>  The ARM has an implementation which I've borrowed.  However, I'm not
> sure yet whether I want to apply the second patch or not, given that the
> functions have pretty high overhead and it's not clear to me that
> anything requires it.

boost and similar C++ packages use these sync functions, but a quick
check shows only long/int32_t usage.  the 4 byte version wasnt
implemented in the past, but testing my current svn trunk shows that
the situation has changed.  so perhaps we can close [#4119] now.
-mike
_______________________________________________
Toolchain-devel mailing list
[email protected]
https://blackfin.uclinux.org/mailman/listinfo/toolchain-devel

Reply via email to