reduce op count for isinteger, add tests

[simonbyrne]

Okay, not the biggest optimisation in the world, but we reduce

  %1 = call double @llvm.trunc.f64(double %0)
  %2 = fcmp oeq double %1, %0
  %3 = fsub double %0, %0
  %4 = fcmp oeq double %3, 0.000000e+00
  %5 = and i1 %2, %4
  ret i1 %5

to

  %1 = call double @llvm.trunc.f64(double %0)
  %2 = fsub double %1, %0
  %3 = fcmp oeq double %2, 0.000000e+00
  ret i1 %3

[StefanKarpinski]

No, this is great. I'd love to audit all of the simple numeric functions in base and make sure that they generate sane code and see if each one can be made any more efficient.

[simonbyrne]

There probably won't be any real performance benefit until we enable new LLVM, as that should hopefully inline the trunc call using SSE instructions.

[yuyichao]

This seems to have 12-13% performance improvement with LLVM 3.7 =)

[jakebolewski]

It would be great if we could begin to add benchmarks for this stuff. cc @jrevels

[StefanKarpinski]

I wonder if benchmarking is really the best way to test for this kind of thing. Maybe checking for specific LLVM assembly? Of course, that doesn't address regressions in LLVM codegen, so maybe both.

[tkelman]

LGTM, I like the test additions.

[hayd]

@jrevels I guess the benchmark label was added for the idea of measuring the llvm (i.e. by LOC/instructions?)

[jrevels]

My guess is that @jakebolewski added it because he thought we should perform actual timing measurements, so I added benchmarks that check various scalar predicates.

Adding tests that check the LLVM bitcode might be useful, but I'm not sure I'm knowledgable enough to write such a test effectively, and it's not currently something that gets covered by BaseBenchmarks.jl.