[llvm] [mlir] [OpenMP][MLIR][OMPIRBuilder] Add a small optional constant alloca raise function pass to finalize, utilised in convertTarget (PR #78818)

[Kiran Chandramohan via llvm-commits]

kiranchandramohan wrote:

> The relevant location is inside of the TargetOp, the rest is largely irrelevant for device but included for completeness. The HLFIR AssignOp will generate an allocation as part of it's lowering from my understanding (in this particular use case at least, there's different lowering's for it). Although, I wouldn't say the problem lies with the HLFIR operation necessarily, more just a side affect of how we lower to a target region for LLVM-IR being shown I think (maybe there's a way to make it more AMDGPU runtime friendly, or a different runtime function for AMDGPU, but I'm doubtful that's the ideal solution). Even if we make sure to raise all AllocaOp's (when we lower to the LLVM Dialect this is done currently, but TargetOp doesn't generate a new Blocks, so there is no isolated Entry Block as such) to the top this would still persist, as we inevitably embed the user code into a seperate block in-between some kernel entry code for the arguments that will branch off to a fail condition or the user code block containing the allocations, and then a later pass (not sure which yet unfortunately) will try to do some magic and end up breaking the generated executable.

Is your observation that the OpenMPIRBuilder creates a separate `Alloca` block and this `Alloca` block will become the entry block in LLVM IR. So even if we move all the allocas to the entry block in FIR/HLFIR/LLVM dialect MLIR, it still won't be the entry block in LLVM IR? 

Is it possible to merge the `Alloca` block to the succeeding block (which is hopefully the entry block at the MLIR dialects)? And if we have lifted in MLIR all to the entry block this will all work out fine?

https://github.com/llvm/llvm-project/pull/78818