[R-pkg-devel] Package submission issue - OMP reduction (flang-new)

[Tomas Kalibera]

On 11/13/23 19:18, Ivan Krylov wrote:
> On Mon, 13 Nov 2023 16:39:45 +0100 (CET)
> Romain Pierlot <romain.pierlot using u-bordeaux.fr> wrote:
>
>> Here is the error message, and the adequate code part is joint in the
>> mail :
>>
>> error:
>> loc("/data/gannet/ripley/R/packages/incoming/frailtypack.Rcheck/00_pkg_src/frailtypack/src/Integrale_mult_scl.f90":1811:17):
>> 'omp.reduction' op must be used within an operation supporting
>> reduction clause interface error: verification of lowering to FIR
>> failed
> Have you tried explaining this in the package submission comment? What
> was the response? Does the reviewer need more evidence? Linking to the
> LLVM issue on GitHub and/or the previous R-package-devel thread may
> help.
>
> (LLVM giveth and LLVM taketh away. While I'm sure that their
> diagnostics helped improve many packages, I have encountered a few
> places in the R code where clang suggested to replace boolean
> expressions && and || with bitwise operations & and |. The code still
> works, but now relies on boolean constants having certain exact bits
> set. Replace a 1 with a 2 and things will break.)
>
> It might be the case that CRAN cannot afford to make an exception for
> this compiler bug. I can imagine a number of workarounds:
>
> 1. In your ./configure script, check for the Fortran compiler being
> flang-new version 17. If the check succeeds, disable OpenMP altogether.
>
> 2. Rewrite your Fortran code so that a variable is never touched before
> it is used in an OpenMP reduction. You may have to declare a new
> variable for every OpenMP reduction loop inside a function.
>
> If you don't succeed in convincing CRAN regarding the compiler bug,
> which one of the two workarounds would you prefer?

I would definitely look for a work-around. In the end we want to give 
users software that works, rather than something that doesn't, but 
telling them, possibly with good evidence, than it is not our fault.

Tomas