http://gcc.gnu.org/bugzilla/show_bug.cgi?id=50457
--- Comment #5 from Oleg Endo <olegendo at gcc dot gnu.org> 2012-09-25 10:53:36
UTC ---
(In reply to comment #4)
> There is no glibc port for SH2* and highly unlikely someone will
> add it. SH2* linux uses and will use totally different environment
> instead of glibc/nptl. Those CPUs have no user/privilege modes and
> folks tend to use disable/enable interrupts for the atomicity on them
> like as in #0.
> *-linux configuration of gcc assumes glibc/nptl everywhere. It means
> that only sh[34]*-linux configuration will work well. For this PR,
> the current atomic functions or builtins should be disabled for
> sh2*-linux configuration and add something that works on SH2*.
> You would have a free hand to the implementation in this case,
> since currently there is nothing correct for SH2*.
Thanks for the feedback.
I'd like to propose to do the following ...
- Add new option to select atomics model
-matomic=
With the following supported models:
soft-gusa (SH3*, SH4* only)
soft-tcb,gbr-offset=xxxx (SH1*,SH2*,SH3*,SH4*)
hard-llcs (SH4A only)
I think this is easier to handle than multiple individual options.
- Make current option -msoft-atomic an alias for -matomic=soft-gusa.
- Remove option -mhard-atomic. It's new in 4.8 and thus we don't need
an alias for that.
- Issue compiler error on unsupported atomic model for a given target
(e.g. -m2a -matomic=soft-gusa --> error)
- Add TCB (thread control block) soft atomics variations.
BTW, they would be compatible across all SH 1..4 types. E.g. running
precompiled SH2 libraries/binaries on SH4 would work -- if anyone wanted
to do so.
Does this make sense?
