investigate compile-time regressions from 2016-08-17

[nikomatsakis]

We don't really know what caused it, but http://perf.rust-lang.org/ reveals:

• 25% drop on issue-32278-big-array-of-strings

and smaller drops on a number of other tests. Merits investigation.

[eddyb]

We have more accurate stats for the switch to MIR, thanks to @Mark-Simulacrum.

EDIT: irrelevant for the observed range here.

[eddyb]

Looking at this graph:

• old trans -> MIR trans regression is a false positive
• anomalies on 8/12, 8/14, 8/17 and 8/18 are false positives
  • no commits explain both regressions and the reversal of those regressions
• minor increase on 8/14 (commits) is from [MIR] Add Storage{Live,Dead} statements to emit llvm.lifetime.{start,end}. #35409 (added Storage{Live,Dead})
• minor decrease on 8/17 (commits) is from Remove the ParamSpace separation from formal and actual generics in rustc. #35605 (more efficient Generics/Substs)

Everything else appears to be noise (similar on all benchmarks), nothing to do here!