[PATCH] D151994: [AMDGPU] Add amdgpu_cs_chain[_preserve] CCs to IR & verifier

[Nicolai Hähnle via Phabricator via llvm-commits]

nhaehnle added inline comments.


================
Comment at: llvm/test/Verifier/amdgpu-cc.ll:217
+; CHECK-NEXT: ptr @inalloca_cc_amdgpu_cs_chain_preserve
+define amdgpu_cs_chain_preserve void @inalloca_cc_amdgpu_cs_chain_preserve(ptr inalloca(i32) %ptr) {
+  ret void
----------------
arsenm wrote:
> Is a regular pointer to the stack allowed?
You mean a `ptr addrspace(5)`? Yes, it's allowed, and we plan to use it.

It's technically not going to be a *stack* pointer as far as LLVM is concerned, because the stack is "dormant"/"inexistent" between chain functions. The plan is to split the wave's scratch allocation into two parts, the bottom part under the control of LLVM and the top part under the control of the frontend, and so passing pointers to that top part can still be meaningful even though the stack "doesn't exist" at chain function boundaries as far as LLVM is concerned.

In principle, we also allow pointers to other address spaces, although I don't see us using those at least for the time being.


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D151994/new/

https://reviews.llvm.org/D151994