================
@@ -599,14 +599,6 @@ TARGET_BUILTIN(__nvvm_e4m3x2_to_f16x2_rn_relu, "V2hs", "",
AND(SM_89,PTX81))
TARGET_BUILTIN(__nvvm_e5m2x2_to_f16x2_rn, "V2hs", "", AND(SM_89,PTX81))
TARGET_BUILTIN(__nvvm_e5m2x2_to_f16x2_rn_relu, "V2hs", "", AND(SM_89,PTX81))
-// Bitcast
----------------
jlebar wrote:
(Thank you, past me, for the commit message.) The key in the commit message
from the previous commit is
> These builtins don't need to be backfilled in clang: They're not accessible
> to user code from nvcc.
If they were accessible from nvcc, we could have compatibility issues where
user code (or, just as likely, CUDA headers) expects the builtin to be there.
In which case we might want to keep the builtin (in the same way that we
auto-upgrade the llvm intrinsic).
I tried to check using cuda.godbolt.org whether these builtins exist in nvcc,
but at this very moment it seems to be down, it's not compiling anything.
Assuming that something like the following does not compile in nvcc, I think
this is fine.
```
__global__ void foo() {
__nvvm_bitcast_f2i(0.0);
}
```
https://github.com/llvm/llvm-project/pull/107936
_______________________________________________
cfe-commits mailing list
cfe-commits@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
