[PATCH] D120566: [OpenCL][AMDGPU]: Do not allow a call to kernel

[Christudasan Devadasan via Phabricator via cfe-commits]

cdevadas created this revision.
cdevadas added reviewers: rjmccall, Anastasia, yaxunl, arsenm.
Herald added subscribers: Naghasan, ldrumm, kerbowa, t-tye, tpr, dstuttard, jvesely, kzhuravl.
cdevadas requested review of this revision.
Herald added subscribers: cfe-commits, wdng.
Herald added a project: clang.

In OpenCL, a kernel is allowed to call other kernels as if
they are regular functions. To support it, clang emits
amdgpu_kernel calling convention for both caller and callee.
A backend pass in our downstream compiler alters such calls
by introducing regular function bodies which are clones of
the callee kernels. This implementation currently limits us
in certain ways. For instance, the restriction to not use
byref attribute for callee kernels.

To avoid such limitations, this patch brings in those
cloned functions early on and prevents clang from generating
amdgpu_kernel call sites. A new function body will be added
for each kernel in the compilation unit expecting that the
unused clones will get removed at link time.


Repository:
  rG LLVM Github Monorepo

https://reviews.llvm.org/D120566

Files:
  clang/lib/CodeGen/CodeGenModule.cpp
  clang/lib/CodeGen/TargetInfo.cpp
  clang/lib/CodeGen/TargetInfo.h
  clang/test/CodeGenOpenCL/amdgpu-kernel-calls.cl
  clang/test/CodeGenOpenCL/visibility.cl

-------------- next part --------------
A non-text attachment was scrubbed...
Name: D120566.411402.patch
Type: text/x-patch
Size: 10257 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/cfe-commits/attachments/20220225/db968c6e/attachment-0001.bin>