Performance regression on pest in newest nightly

[dragostis]

Running the benchmark in pest_grammars on current nightly vs nightly 2017-12-04 yields 56% performance regression.

Current:

test data ... bench:      72,455 ns/iter (+/- 13,575)

2017-12-04:

test data ... bench:      46,380 ns/iter (+/- 1,352)

[alexcrichton]

I believe this is a result of #46910 where LLVM's making a different inlining decision in the ThinLTO phases than it previously did when everything was in one codegen unit. That PR was merged with prior knowledge that it would not yield a 100% across-the-board improvement in all cases.

That being said, it's not clear why an inlining decision wasn't made here! The difference can be verified by adding:

[profile.bench]
codegen-units = 1

to the Cargo.toml at the root, and cargo bench should go back to what it previously was.

[nikomatsakis]

OK, based on @alexcrichton's analysis I'm going to close this issue as being an expected byproduct of enabling ThinLTO. @dragostis you might want to consider adding #[inline] to some of your hot methods; hopefully profiling could help you find which ones. Thanks for the report!