LLVM 6 upgrade caused ~25% runtime regression for inflate 0.3.4 decoding

[anp]

After bisecting, 39abcc0 introduced a roughly 25% increase in the runtime of lolbench's inflate_0_3_4::decode benchmark.

Summary:

rust-lang/rust commit	est. runtime	95% ci
3bcda48	1.5933 ms	1.5778,1.6081
39abcc0	1.9997 ms	1.9914,2.0083
change	~ 25 %

The regression is (hopefully) reproducible by cloning:

$ git clone https://github.com/anp/inflate-regression
$ cd inflate-regression

(note: I haven't reduced this beyond the inflate crate's public API -- I need to try to find the actual section of code that's affected)

Here's my output (2016 MBP15). I ran these on my laptop after seeing the change on a x86_64-unknown-linux-gnu machine first.

(rustup targets downloaded by cargo-bisect-rustc)

3bcda48, bors merge before LLVM 6:

$ cargo +ci-3bcda48a30b21e46b81a7989deb30a3ba85fb918-x86_64-apple-darwin run --release
...
decode                  time:   [1.5778 ms 1.5933 ms 1.6081 ms]
...

39abcc0:

$ cargo +ci-39abcc04139a0fd24422f422271849dc91e39a88-x86_64-apple-darwin run --release
...
decode                  time:   [1.9914 ms 1.9997 ms 2.0083 ms]

Current nightly:

$ cargo +nightly run --release
...
decode                  time:   [1.9383 ms 1.9466 ms 1.9555 ms]
...

[anp]

cc @rust-lang/wg-codegen

[upsuper]

Might be related to #49873.

[eddyb]

Please always compare inflate performance changes before and after the debloating image-rs/inflate#35 (which greatly reduces the code passed to LLVM).

[anp]

Haven't had time to reduce this further yet, but a fix for one potential culprit has been backported: rust-lang/llvm#115. Will see what results for this are like when that's in nightly.

[nikic]

Current nightly did not improve this, so this is unrelated to #49873.

[anp]

@nikic thanks for checking! I'll try to reduce this soon then.

[steveklabnik]

Triage: it's been over a year. We're on LLVM... 8? now? Is this ticket still relevant?

[nox]

Closing it.