[PATCH] D155982: Partially revert "clang/HIP: Remove __llvm_amdgcn_* wrapper hacks"

Yaxun Liu via Phabricator via cfe-commits cfe-commits at lists.llvm.org
Fri Jul 21 11:36:07 PDT 2023


yaxunl added inline comments.


================
Comment at: clang/lib/Headers/__clang_hip_libdevice_declares.h:319
+// relevant.
+__device__ inline _Float16 __llvm_amdgcn_rcp_f16(_Float16 x) {
+  return ((_Float16)1.0f) / x;
----------------
arsenm wrote:
> arsenm wrote:
> > yaxunl wrote:
> > > Can we add the deprecated attribute to urge people not to use them?
> > I initially added those, but thought it was a bit aggressive to put in always included headers. Is there an established practice for deprecating builtin header functions?
> Is there a hip header version macro I could guard this on?
> I initially added those, but thought it was a bit aggressive to put in always included headers. Is there an established practice for deprecating builtin header functions?

I saw it used by libstdc++ headers and pthread headers to deprecate APIs.

I think it is common practice to use deprecate attibute.

We'd also better mention it in clang release note and give a timeline to remove it, e.g. next clang major release.


================
Comment at: clang/lib/Headers/__clang_hip_libdevice_declares.h:319
+// relevant.
+__device__ inline _Float16 __llvm_amdgcn_rcp_f16(_Float16 x) {
+  return ((_Float16)1.0f) / x;
----------------
yaxunl wrote:
> arsenm wrote:
> > arsenm wrote:
> > > yaxunl wrote:
> > > > Can we add the deprecated attribute to urge people not to use them?
> > > I initially added those, but thought it was a bit aggressive to put in always included headers. Is there an established practice for deprecating builtin header functions?
> > Is there a hip header version macro I could guard this on?
> > I initially added those, but thought it was a bit aggressive to put in always included headers. Is there an established practice for deprecating builtin header functions?
> 
> I saw it used by libstdc++ headers and pthread headers to deprecate APIs.
> 
> I think it is common practice to use deprecate attibute.
> 
> We'd also better mention it in clang release note and give a timeline to remove it, e.g. next clang major release.
> Is there a hip header version macro I could guard this on?

Yes like https://github.com/llvm/llvm-project/blob/main/clang/lib/Headers/__clang_hip_runtime_wrapper.h#L80


CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D155982/new/

https://reviews.llvm.org/D155982



More information about the cfe-commits mailing list