add condition to use emulated fma

[oscardssmith]

Testing has revealed that some old windows machines without fma were using broken emulated versions for Float32. This adds a test.

[oscardssmith]

This is kind of a regression from #42783 but only in the sense that we added a test that correctly found out that old machines were already doing the wrong thing.

[ViralBShah]

The old machine case we found here was a win64 ivybridge which does not have hardware fma, and the system libm is not quite doing the right thing.

The issue was actually discovered in #42299 (comment)

[oscardssmith]

@JeffBezanson can we please get fma multiversioning so I don't have to figure out why this doesn't work?

[ViralBShah]

The CI machines are AMD Epycs. Might that be a different case?

[oscardssmith]

If they are Epyc, than they have FMA which would imply that my emulated version is wrong which doesn't make any sense, because then the failures should happen on other platforms.

[ViralBShah]

And the tests are passing on epyc with Linux and win32. Is something off with llvm on win64?

[oscardssmith]

It really shouldn't be able to do anything for LLVM to mess up. It should just be using the processor instruction...

[ViralBShah]

Isn't it possible that if this is running in some kind of a VM, that the VM is not exhibiting a slightly different target processor? I mean we've seen these issues when running in VMs before. Perhaps @staticfloat may know more.

[oscardssmith]

Closing as this test can't do anything until we have fma multiversioning.