new LLVM optimization pass list

[JeffBezanson]

This is a rough cut of some possible changes to our pass list, that should run a bit faster and possibly even produce better code. There is some inline commentary that will be removed, and just reflects some notes from a discussion with an LLVM developer. Several variations are possible, but I'm curious to try some benchmarks first. This does make the build a bit faster, and reduces time-to-first-plot by about 4-5%.

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

[nanosoldier]

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

[JeffBezanson]

So some of the regressions were caused by moving IndVarSimplify and removing the second LoopDeletion. I put those back. I'll also try replacing the internal InstCombines with InstSimplify, which is much faster. Passing false to InstCombine to disable "expensive" combines does not provide any real speedup.

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

[JeffBezanson]

With this pass list time-to-first-plot is ~8% faster instead of ~4% (due to using InstSimplify). Let's see how it does.

[nanosoldier]

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

[JeffBezanson]

Ok, one of those InstCombines was important for vectorizing loops over Union arrays. Let's put that back and try again.

[JeffBezanson]

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

[nanosoldier]

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

[JeffBezanson]

That's starting to look pretty good. We may have a solid candidate here.

[JeffBezanson]

Ok, @Keno @vtjnash and I went over this and we think it's ready to go.