[PATCH] D130096: [Clang][AMDGPU] Emit AMDGPU library control constants in clang

[Yaxun Liu via Phabricator via cfe-commits]

yaxunl added inline comments.


================
Comment at: clang/lib/CodeGen/TargetInfo.cpp:9436
+        CGM.getModule(), Type, true,
+        llvm::GlobalValue::LinkageTypes::LinkOnceODRLinkage,
+        llvm::ConstantInt::get(Type, Value), Name, nullptr,
----------------
jhuber6 wrote:
> yaxunl wrote:
> > This does not support per-TU control variables. Probably should use internal linkage.
> The AMDGPU device libraries use `linkone_odr` so I figured it was the most appropriate here. It should mean that we can have multiple identical definitions and they don't clash. There's also no requirement for these to be emitted as symbols AFAIK.
> The AMDGPU device libraries use `linkone_odr` so I figured it was the most appropriate here. It should mean that we can have multiple identical definitions and they don't clash. There's also no requirement for these to be emitted as symbols AFAIK.

clang uses  -mlink-builtin-bitcode to link these device libraries for HIP and OpenCL. Then the linkage of these variables becomes internal linkage. That's why it works for per-TU control.


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D130096/new/

https://reviews.llvm.org/D130096