tree 4d019b84039dda05892fe623c933adb7ff0634d3
parent 77ac727c7f6cbe32ca8972b8c69bd32cba17b690
author bjacob <benoitjacob@google.com> 1685614213 -0400
committer GitHub <noreply@github.com> 1685614213 -0400
gpgsig -----BEGIN PGP SIGNATURE-----
 
 wsBcBAABCAAQBQJkeG6FCRBK7hj4Ov3rIwAAiA4IACNpzOZGQWZNCSqkL2RK4r6v
 aegAK+hLaXWTlSNKORL2G9yusZ04VF0VfPMCqzz9JU1SPHwmsb405a+XmSz+PfsY
 8Z4AwCIw+EE2K2CxHuza62hJK9xVZHLGhUzJXg1lQFpaHqdxxMQrVJAnnbNAsWx4
 rorIvQSdwHqB+QN8LLtMnHVT1b8x8Cow06GxK/5meDZQKFz7lFWZzpDZCs6pr2fJ
 ScJBcnvscjUrzr20zxRTivSng8z2R9F0ikEAM7PVdOu4wH2n3x9Z5Y0ONWJfOqRy
 EkcJMqTjgdke4ZGmxFXP8HjC4xE9om7/e8YPgZCVL+Arlc6c7XNwLOKnv9Jd0Pk=
 =Linx
 -----END PGP SIGNATURE-----
 

Drop unused `_none` values in ukernel enums (#13877)

This is a minor optimization and arguably cleaner separation of
validation vs implementation code, but the primary motivation was to
avoid a miscompile in LLVM/riscv32, explained in that comment in
`mmt4d_internal.h`:

```
      // This unreachable statement is not just an optimization, it also works
      // around a LLVM/riscv32 miscompile.

      // When we used to have a iree_uk_mmt4d_type_none value equal to 0 and
      // were returning it here, that caused this whole switch statement to be
      // miscompiled by LLVM/riscv32 as if it were UB. That value was passed to
      // `iree_uk_type_bit_count(x)`, which evaluates to `1<<(x - 3)`, which is
      // UB if x<3. So it was fair to treat that default: clause as UB, but
      // LLVM/riscv32 was incorrectly treating the whole switch as UB.
```