[llvm] [AMDGPU] Compiler should synthesize private buffer resource descriptor from flat_scratch_init (PR #79586)

[Jon Chesterfield via llvm-commits]

JonChesterfield wrote:

> Literally, I need to track down the compilation of the code that caused incorrect memory access. Although, I know almost nothing about the openmp offload project and I would appreciate a reference on the docs that makes me understand better what went on.

It's quite obfuscated by the openmp toolchain providing a user friendly experience. Ultimately there's a single IR blob that gets fed into the backend to build a single code object containing some number of kernels. save-temps might give you that IR, or it might give you some unrelated IR - both have been true at various points in the past.

`clang -fopenmp -###`  should tell you the steps. Some are openmp-specific IR rewriting tools. Somewhere in that pipeline is the IR you seek.

https://github.com/llvm/llvm-project/pull/79586