[PPC] relax fastmath test

[vchuravy]

@shirodkara if I recall our discussion correctly the backend stance was there is no guarantee that
this operations yield Inf under fastmath?

[vtjnash]

I'm not sure of the purpose of any of these tests. Isn't this undefined behavior for "fast" math and the program is allowed to do anything when encountering a NaN or Inf value.

[vchuravy]

I'm not sure of the purpose of any of these tests. Isn't this undefined behavior for "fast" math and the program is allowed to do anything when encountering a NaN or Inf value.

I was wondering about that as well. These were originally added in #9759 by @eschnett.

[eschnett]

Yes, these tests are over-zealous. As @vtjnash suggests, they only test incidental behaviour. They could be replace by test that merely ensure that these tests do not throw an exception.

[eschnett]

You could say isa T instead of != one.

[vtjnash]

There’s no guarantee this doesn’t throw, or other much worse things.

[eschnett]

I see that LLVM's definition for the fast-math behaviour has changed over time. It used to guarantee that the resulting behaviour is an undefined value, but won't throw. This is now different – the fast-math flags seem to require that nans and infs are detected and produce a poison value? I'm not familiar enough with poison values to see how this could be implemented efficiently.

[vchuravy]

@eschnett a poison value is like undef but with propagation semantics. I collected a couple of references here #38735 (comment)

With regard to this PR. Should I remove the tests since the explicitly result in undef/poison?

[eschnett]

@vchuravy If you confirm that these tests should not be there (according to https://llvm.org/docs/LangRef.html#fast-math-flags), then yes, please remove them.

[vchuravy]

Yes that is my reading of the spec.

[eschnett]

@vchuravy I would just remove these lines.