Clang's intrinsic header currently tries to negate the third operand of a vfmadd mask3 in order to create vfmsub, but this fails isel. This patch add scalar vfmsub mask3 that we can use instead to avoid the negate. This is consistent with the packed instructions.
Details
Details
Diff Detail
Diff Detail
- Repository
- rL LLVM
Event Timeline
Comment Actions
Turns out we also need vfnmsub too. At least until we teach combineFMA to be called for FMSUB/FMSUB_RND.