This is the mail archive of the gcc-patches@gcc.gnu.org mailing list for the GCC 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: [gomp] Tiny fix for #pragma omp atomic


On Thu, Sep 22, 2005 at 11:33:07AM -0400, Jakub Jelinek wrote:
> c_finish_omp_atomic has a bunch of bugs ... 

Yes, I have a large patch that went testing last night.

> also adds start for supporting float/complex types that are
> also allowed by OpenMP.

Float, yes.  Why do you believe that complex is allowed?
It's not a scalar...

I'm also not handling pointer types properly.

> 1) it doesn't handle computation in promoted types, say
>    unsigned short c; #pragma omp atomic c += 4; - will work on this next

Fixed.

> 2) probably it should also handle atomic stores, as e.g. fold will
>    turn _Bool b; b |= 1; into b = 1; and I guess say with a &= 0; etc.
>    will do the same

Eh, perhaps.

> 3) there seems to be a problem with the create_tmp_var_raw calls,
...
>    assertion - the create_tmp_var_raw temporary has none of those set

Fixed.

> 4) float/double/_Complex float atomics - once 3) is resolved,
>    this should be just a matter of making the pointer a union,
>    always using compare and swap and using VIEW_CONVERT_EXPRs or something

For float/double, yes this can work.  For long double I expect
to have to use an external mutex.

> 5) we should verify that rhs doesn't reference x

Yeah...

> Is the following ok for gomp for now?

No, it conflicts pretty heavily with my fixes.


r~


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