[llvm-dev] Loop unrolling pragma and OpenMP

[ardi via llvm-dev]

Thanks a lot for the clarification. This explains why I wasn't getting
any improvement by combining the two pragmas.

Thanks!
ardi



On Mon, May 24, 2021 at 10:45 PM Michael Kruse <llvmdev at meinersbur.de> wrote:
>
> Unfortunately, this does not work, by design. In the compiler pipeline
> "#pragma omp parallel for" is lowered in the front-end, while "#pragma
> clang loop unroll_count(N)" is in the LoopUnroll mid-end-pass.
> "#pragma omp parallel for" cannot be applied on a loop that has not
> yet been unrolled in the front-end. In practice, the #pragma clang
> loop will just be ignored.
>
> Michael
>
> Am Mo., 24. Mai 2021 um 03:02 Uhr schrieb ardi via llvm-dev
> <llvm-dev at lists.llvm.org>:
> >
> > Hi,
> >
> > Newer OpenMP versions include a construct for loop unrolling, but,
> > however, I was wondering if the regular clang pragma for loop
> > unrolling can be safely mixed with an OpenMP loop,without using the
> > new OpenMP feature.
> >
> > I mean, is this guaranteed to work?
> >
> > #pragma omp parallel for
> > #pragma clang loop unroll_count(N)
> > for(i=0;i<n;i++) {
> >    whatever();
> > }
> >
> > Thanks a lot!
> > _______________________________________________
> > LLVM Developers mailing list
> > llvm-dev at lists.llvm.org
> > https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev