sdardis added a comment.
Checking the generated IR for the incorrect uses of v8f16 shows that clang is
silently generating bitcasts from <8 x i16> to <4 x float>. There are some
other cases where the type of the operands is incorrect w.r.t. the builtin used.
I'm not familiar with the type checking machinery, so I'm unsure if the silent
conversion is a bug or feature.
cfe-commits mailing list