Performance regressions in nightly btwn nightly-2017-12-25 and nightly-2017-12-26

[lloydmeta]

I think this is the right place to report that I am seeing some performance regressions in Frunk, namely in the benchmark for LabelledGeneric-based conversions, where I compare hand-written conversions between structs vs auto LabelledGeneric-driven conversions.

I've narrowed it to changes made between nightly-2017-12-25 and nightly-2017-12-26

$ cargo +nightly-2017-12-25 bench big
running 4 tests
test big_from_24fields           ... bench:         128 ns/iter (+/- 10)
test big_from_25fields           ... bench:         159 ns/iter (+/- 27)
test big_transform_from_24fields ... bench:         136 ns/iter (+/- 68)
test big_transform_from_25fields ... bench:         186 ns/iter (+/- 36)

$ cargo +nightly-2017-12-26 bench big
test big_from_24fields           ... bench:         149 ns/iter (+/- 32)
test big_from_25fields           ... bench:         188 ns/iter (+/- 44)
test big_transform_from_24fields ... bench:       7,631 ns/iter (+/- 3,917)
test big_transform_from_25fields ... bench:       9,162 ns/iter (+/- 4,136)

As you can see above, the LabelledGeneric-based conversion (big_transform_from_*) went from being essentially the same as hand-written to ~70x the overhead 😢

Not really sure if it matters how LabelledGeneric struct-to-struct transformation works in Frunk, but it essentially gets the compiler to verify that for two given structs, their fields are "compatible" (target struct field names and fields types are a subset of source struct's) and then "just" aligns the generic representation of source to target and dumps one into the other. I say "just" because I still have no idea how Rust managed to make this a zero-overhead op 😆

The bloody details are available in multipart on my blog (part 1, part 2, part 3).

[alexcrichton]

Thanks for the report! I was able to "fix" the regession by setting codegen-units = 1, which means this is probably due to #47745

[Enselic]

Triage: Yes, let's close as duplicate of #47745.