[llvm] [AMDGPU] Don't DEALLOC_VGPRS from callable functions (PR #72245)

[Jay Foad via llvm-commits]

jayfoad wrote:

> Callable functions should not send the DEALLOC_VGPRS message, because that might release the VGPRs and scratch allocation before potential scratch stores in the caller have completed.

I am still confused about this. I thought callable functions waited for all counters to be zero on entry (see "Wait for any outstanding memory operations that the input registers may depend on" in `SIInsertWaitcnts::runOnMachineFunction`) so how can there still be any outstanding scratch stores from a caller?

https://github.com/llvm/llvm-project/pull/72245