[PATCH] D89525: [amdgpu] Enhance AMDGPU AA.

Mon Oct 19 23:54:01 PDT 2020

t-tye added inline comments.

================
Comment at: llvm/lib/Target/AMDGPU/AMDGPUAliasAnalysis.cpp:104-105
+    if (const LoadInst *LI = dyn_cast<LoadInst>(ObjA)) {
+      // If a generic pointer is loaded from the constant address space, it
+      // could only be a GLOBAL or CONSTANT one as that address space is soley
+      // prepared on the host side, where only GLOBAL or CONSATNT variables are
----------------
hliao wrote:
> arsenm wrote:
> > Is it true that we couldn't have a relocation for an LDS value?
> Do you mean GOT-like relocation for LDS? That's possible if we don't want to clone sub-function per kernel-function. But, as GOT is created in the backend, it needs a PseudoValue for IR instead of a real Value. We could easily tell them from the load from constant memory in user code.
> Is it true that we couldn't have a relocation for an LDS value?

Well it depends how they are defined. Graphics is using LDS relocations to stitch together pipeline parts to pass arguments through LDS.

Compute does not allow this. If LDS allocation is managed in a TLS like manner (which is not currently done) then maybe there is a place for LDS relocations. But it is not like a global memory relocation.

Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D89525/new/

https://reviews.llvm.org/D89525