[OpenMP] Add clang option to change device RTL stack size

[jhuber6]

Summary

The OpenMP RTL uses a stack of shared-memory to handle data-sharing
that's necessary for communicating data between threads on the device.
This patch add an option in clang to set this at compile time by using a
global varialbe in the new device runtime.

The shared memory used is set by -fopenmp-target-stack-size=. This
will configure a global <kernel>_stack_size used by the libomptarget
plugin to set the amount of dynamic shared memory allocated. The size
will also be added to the __kmpc_target_init runtime call so the
device runtime can check the size and if the stack must be initialized.

[jdoerfert]

The runtime code is missing, right?

[jhuber6]

In D110108#3010861, @jdoerfert wrote:

The runtime code is missing, right?

Yes, was going to make that a separate patch, also need to fix the LLVM tests.

[jhuber6]

Fixing OpenMPOpt tests.

[tianshilei1992]

clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp

1057

Do we want to check somewhere if it's a valid value?

[jhuber6]

clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp

1057

It's an unsigned value that corresponds to how much memory gets allocated, every value should be equally valid since we don't really know how much memory there is. I was considering making it a 64-bit unsigned as well, but I figured no one would need that much shared memory.

[tianshilei1992]

LGTM

clang/lib/CodeGen/CGOpenMPRuntimeGPU.cpp

1057

That makes sense.

[jhuber6]

Rebase