irinterp: improve semi-concrete interpretation accuracy

[aviatesk]

By enforcing re-inference on calls with all constant arguments.

While it's debatable whether this approach is the most efficient, it was the easiest choice given that used_ssas based on IncrementaCompact wasn't an option for irinterp.

• fixes Aggressive constant propagation elides certain branches but not others #52202
• fixes type inference regression in 1.9 due to semi-concrete interpretation #50037

[aviatesk]

@nanosoldier runbenchmarks("inference", vs=":master")

[nanosoldier]

Your benchmark job has completed - no performance regressions were detected. A full report can be found here.

[vtjnash]

LGTM. It does seem to suggest to me that we should consider how to record a bit of information on the calls whether a more precise result is possible from constprop (either because of some widening vs the return value being obviously unchangable like just return nothing)

[aviatesk]

It does seem to suggest to me that we should consider how to record a bit of information on the calls whether a more precise result is possible from constprop (either because of some widening vs the return value being obviously unchangable like just return nothing)

That sounds to be a great idea. I would like to give that a try in a separate PR.

[ViralBShah]

@KristofferC Can we backport this to 1.10 to resolve JuliaSparse/SparseArrays.jl#469 (assuming this is the correct fix for that regression)?