[clang] Revert "[Clang][AMDGPU] Change __fp16 to _Float16 in builtin definitions" (PR #185861)

[Shilei Tian via cfe-commits]

shiltian wrote:

I think what we need is a full CQE against a version of LLVM with all __fp16 to _Float16, then we will have a good idea of what will be broken in the next bulk promotion. Like I mentioned earlier, this revert will not resolve the issue that downstream code is relying on implicit type conversion. We fix CK build now but everything replying on implicit type conversion, such as RCCL, will fail the moment we do bulk promotion, if they don't fix them ahead of time. They will not know this if they don't use the latest upstream LLVM or our ROCm's new main fork.

https://github.com/llvm/llvm-project/pull/185861