[flang-commits] [flang] [flang] Skip processing reductions for unstructured `do concurrent` loops (PR #150188)

[Kareem Ergawy via flang-commits]

ergawy wrote:

> This does seem like a reasonable approach.
> 
> What do you think about the following program:
> 
> ```
> program test
> integer i,r
> 
> do concurrent (i=1:10) reduce(+:r)
>   r = 1
> enddo
> 
> print *,r
> end
> ```
> 
> As I understand it, this should print the number of threads the loop actually used (because that will be the number of threadprivate values of `r` which were created). This would make something like what is done here user-visible. But I guess that is okay because the standard does not seem to require that do concurrent is actually run over multiple threads?

I probably missed what you are asking. This is a structured loop so reduction attributes will be emitted as usual. The changes in this PR only affect unstructured loops which are serialized so `r` will be `1` indeed.

https://github.com/llvm/llvm-project/pull/150188