vec_absd should build with -mpower9-vector as it does with -mcpu=power9

[lu-zero]

Right now with both clang-16 and clang-17 it triggers a:

fatal error: error in backend: Cannot select: intrinsic %llvm.ppc.altivec.vabsdub

[llvmbot]

@llvm/issue-subscribers-backend-powerpc

Author: Luca Barbato (lu-zero)

Right now with both clang-16 and clang-17 it triggers a: ``` fatal error: error in backend: Cannot select: intrinsic %llvm.ppc.altivec.vabsdub ```

[chenzheng1030]

hmm, actually -mpower9-vector is a switch for instructions that start with xv/xs/xx(vector-scalar instructions) on pwr9. Instruction starting with v like vabsdub is guarded by altivec feature bit. So this is some kind of by design. Unfortunately there is no -mpower9-altivec option in clang for now. So can only enable this builtin with -mcpu=power9.

[chenzheng1030]

There is doc about this builtin usage in https://www.ibm.com/docs/en/openxl-c-and-cpp-aix/17.1.2?topic=functions-vec-absd

[lu-zero]

gcc does allow vec_absd with -mpower9-vector and the intrinsic is guarded by __POWER9_VECTOR__ in altivec.h thus my confusion.

[chenzheng1030]

gcc does allow vec_absd with -mpower9-vector and the intrinsic is guarded by POWER9_VECTOR in altivec.h thus my confusion.

OK. That's a good reason we should check llvm. llvm should be compatible with gcc for this option and its behavior.

[bzEq]

Hi @lu-zero Which gcc version are you using? How is -mpower9-vector documented in GCC?

[lu-zero]

gcc-13, the manpage mentions it in passing as:

If you use the ISA 3.0 instruction set (-mpower9-vector or -mcpu=power9) ...

The part that is annoying with clang is that __POWER9_VECTOR__ is the only guard in the clang-provided altivec.h so one would be none-the-wiser when looking at the instructions there and papr also does not differentiate: ISA 3.0 the only discriminant documented.

[bzEq]

I've posted an RFC PR #86905. @lu-zero I've checked the manual of the latest GCC manual which doesn't list -mpower9-vector like -mcpu any more. I think clang should follow the latest gcc behavior, i.e., I suggest you use -mcpu=pwr9 rather than -mpower9-vector. Some GCC PPC port developers told me, this might be a debug option which is not for long-term uses.

[lu-zero]

pwr9 or power9 ? Right now I set the CFLAGS for the code using vec_absd to -mcpu=power9 in dav1d, and indeed would be nice if the -mcpu does not need to be changed again...

(and yes, it is already annoying to have -Ctarget-cpu to support pwr9 only in rustc).

[bzEq]

pwr9 or power9

I think clang should support both.