assume getfield is non-volatile if inferred to be constant

[JeffBezanson]

Hopefully fixes regression caused by #18017.

[JeffBezanson]

@nanosoldier runbenchmarks(ALL, vs="@b80dad40ab36bcd7e11859d36b51c6afcf008148")

[KristofferC]

@nanosoldier runbenchmarks(ALL, vs="@b80dad4")

[vtjnash]

lgtm. could we test with something like checking if the following folds down to jlapi 2?

type A; x::Void; end
f(a::A) = a.x

why did this cause such performance issues?

[JeffBezanson]

I believe the big regression was due to the getfield calls in Core.Inference.return_type not being removed, in turn causing promote_op not to be constant-folded.

[JeffBezanson]

32-bit AV timeout?

[nanosoldier]

Your benchmark job has completed - possible performance regressions were detected. A full report can be found here. cc @jrevels

[vtjnash]

looks like it fixed most of them, but all the sumelt are still regressed

[JeffBezanson]

It's possible that's a different issue. In my bisect I only looked at nbody_vec since it was the worst regression by a large margin.