[flang][OpenMP] Add initial pass to map do concurrent to OMP

[ergawy]

Adds a new pass to map do concurrent loops to OpenMP. So far the pass only maps to the host, i.e. parallel do. Later, it will be extended to map to the device/target as well.

This diff also adds the pass the code-gen pipeline flang. To invoke the new pass from flang, use -fdo-concurrent-parallel=<none|host|device>. The option is only active when OpenMP is enabled throught -fopenmp.

[ergawy]

Ping 🔔! As agreed with @mjklemm & @dpalermo, we will go ahead with fleshing out the transformation pass on ATD first and after that will work on upstreaming it and cater for the requirements of the community (i.e. have proper analyses to ensure the transformation is safe).

For now, please have a look at the pass and let me know if you have any comments. There is still a lot of work to do here so pushing this forward would be great.

[mjklemm]

LGTM

[skatrak]

Thanks Kareem, I think this is a reasonable approach. I have a couple of comments which are mainly to make this a bit more robust, hopefully without complicating the implementation too much. I'd be interested to know whether there are issues with these.

[skatrak]

This sounds to me like "we want to add the lower bound, upper bound and step variables to a firstprivate clause for the parallel region". So, if I understand it correctly, maybe we don't want to do much here, but rather rely on delayed privatization when that's implemented for omp.parallel. In the meantime, your implementation below at least addresses the constant bounds case. I'd suggest emitting an error if any of these variables is not a constant or type cast (otherwise it might silently produce incorrect code or crash) and also check that they indeed have a defining op before dereferencing it, or using function arguments would make this crash altogether.

If we want the first implementation to support something other than just constants and we can't wait for delayed privatization, maybe the approach with the least friction is to try to replicate here the "firstprivate" clause handling for parallel regions currently implemented in PFT to MLIR lowering (between OpenMP.cpp and the DataSharingProcessor).

[ergawy]

Added some checks to make sure defining ops meet the current restrictions of the pass.

Both delayed privatization and do concurrent mapping are in-flight at the moment. So indeed once delayed privatization is more complete and the switch in flang is flipped we should move this to use delayed privatization. I think that should happen soon 🤞. So once I get to beef up this pass a bit more, we indeed either use delayed privatization directly or replicate some of the logic in OMP lowering.

[skatrak]

Could this be replaced by a simpler check for live-ins into the loop region? That is, all values that are defined outside but used inside of it. Then, we could have some logic to decide whether each of these values should be private, firstprivate or shared and produce the corresponding omp.private ops and block args.

[ergawy]

I think that makes sense. But do you mind leaving that out for now and gradually work up the support for more complex loops as we go? For now, my immediate focus is to map a simple loop like the one in the test to the device instead of the host. So I am focusing on simple loops for both the device and host and will move up to more complex stuff after that.

[ergawy]

Thanks @mjklemm & @skatrak for your review.

[skatrak]

Thanks for addressing my comments. I think this a good first approach before we can use early privatization to support more complex uses.