GPU Kernel Conditional Statement Bug Fix

[ShreyasKhandekar]

Fixing a bug in the compiler where if a GPUizable loop has a conditional statement whose expression comes from outside the loop, we do not add it as a argument to the code generated for the kernel.

Added conditions in gpuTransforms.cpp so that it adds that expression as a kernel argument.

Also added a catch-all else block that causes an error if there happen to be any other such cases where we forget to add a compiler argument.

Added a test which fails on main but passes with this fix.

• Local test
• Paratests

[stonea]

Can you check that the INT_FATAL("Unexpected symbol expression"); you added doesn't get triggered by any of our existing tests? Just run start_tests under test/gpu/native on Horizon or Osprey and make sure nothing fails.

If that passes I'm good with merging this.

Note there's a couple of test failures we have from last night so these two are expected: https://chapel.discourse.group/t/cron-cray-cs-gpu-native/14597

[e-kayrakli]

Thanks Shreyas!

I hope the tests will pass. If it doesn't we probably revert that INT_FATAL into just a comment for now saying that we may need to special case some things down the road there, or we may need to generalize the implementation a little bit better. But we don't know what those special cases are yet and it is a bit too early to think about refactoring here. For now we are just directly copying nodes which contains symbols for which we don't have a special action.