fix effect_free computation over getfield

[vtjnash]

fix #18015

@JeffBezanson this also fixes the bug you noted where inlineable was passing the wrong sv to effect_free. It would be appreciated if you could verify I've picked the right linfo to pass in each case.

[tkelman]

osx travis got stuck, log backed up to https://gist.github.com/a62ec9f360ea4b4c8a8cdaec465b52dd and restarted

[JeffBezanson]

I think we might need || typ.abstract here.

[JeffBezanson]

if you could verify I've picked the right linfo to pass in each case

Yes that part looks right to me.

[tkelman]

It looks like this caused some pretty major performance regressions.

https://github.com/JuliaCI/BaseBenchmarkReports/blob/1ea0d8c99a047f5d72c9d2aa70af49ccc8ab5463/c48d203_vs_b80dad4/report.md is mostly fine at c48d203,

https://github.com/JuliaCI/BaseBenchmarkReports/blob/ec71a5c4cb488dde420d1908f81b7b66524630c8/a0d75a4_vs_b80dad4/report.md is bad at a0d75a4

[KristofferC]

Can we try to use this super awesome benchmark tool a bit more frequently. A run doesn't take that long. Saves @tkelman from manually having to bisect performance regressions.

[JeffBezanson]

I agree. Running benchmarks on every commit is starting to get urgent. If it takes too long or we don't have enough resources, we should run a reduced version of the suite. Running on every commit is much more important than having perfect coverage.