Benchmark hang with lto = true and target-cpu=native

[TheIronBorn]

I first posted this in Cargo at rust-lang/cargo#5312

I'm benchmarking some code and often get hangs with $ RUSTFLAGS='-C target-cpu=native' cargo bench.

My Cargo.toml:

[package]
name = "test_hang"
version = "0.1.0"

[profile.bench]
lto = true

It only happens when both LTO and target-cpu=native are enabled.

A typical output looks like this, where it hangs for as long as I've left it:

$ RUST_BACKTRACE=1 RUSTFLAGS='-C target-cpu=native' cargo bench -vv
       Fresh <crates>
    Finished release [optimized] target(s) in 0.0 secs
     Running `<path to crate>target/release/deps/<crate name>-86f63a8b1891302e --bench`

running 19 tests
<ignored tests>

test result: ok. 0 passed; 0 failed; 19 ignored; 0 measured; 0 filtered out

     Running `<path to crate>/target/release/deps/bench-3dd5f8be084d72c9 --bench`

I'm using cargo 1.26.0-nightly (b70ab13b3 2018-04-04). The code requires nightly so I can't test whether this happens on stable.
$ rustc --print target-cpus says my processor is sandybridge.

Meta

$ rustc --version --verbose
rustc 1.27.0-nightly (eeea94c11 2018-04-06)
binary: rustc
commit-hash: eeea94c11d02ff62fb011d1afdda9301fdf9726b
commit-date: 2018-04-06
host: x86_64-apple-darwin
release: 1.27.0-nightly
LLVM version: 6.0

[TheIronBorn]

It doesn't seem to happen when using target-cpu=westmere, the preceding Intel CPU microarchitecture.

[TheIronBorn]

I'm trying to find a minimal code example.

[TheIronBorn]

Turns out it actually happens with a totally empty project:

├── Cargo.lock
├── Cargo.toml
├── benches
│   └── bench.rs
└── src
    └── lib.rs

[neachdainn]

I have also experienced this issue.

rustc 1.27.0-nightly (e5f80f2a4 2018-05-09)
binary: rustc
commit-hash: e5f80f2a4f016bf724a1cfb580619d71c8fd39ec
commit-date: 2018-05-09
host: x86_64-apple-darwin
release: 1.27.0-nightly
LLVM version: 6.0

[TheIronBorn]

@neachdainn what is your microarchitecture?

[neachdainn]

Core i7-4770HQ (Haswell)

[TheIronBorn]

Could you also test with a couple other target CPUs?

[TheIronBorn]

Specifically Sandy Bridge and Westmere.

[neachdainn]

Here's some tests on a different system:

rustc 1.27.0-nightly (acd3871ba 2018-05-10)
binary: rustc
commit-hash: acd3871ba17316419c644e17547887787628ec2f
commit-date: 2018-05-10
host: x86_64-unknown-linux-gnu
release: 1.27.0-nightly
LLVM version: 6.0

Core i5-2500k (Sandybridge)

I am not seeing the bug on this system, so it may be macOS specific. I'll be able to test this on my Mac on Monday.

[Mark-Simulacrum]

Can someone post a minimal example with code? Or at least some example?

This is probably a case of #50422.

[TheIronBorn]

As I mentioned here #49766 (comment), it happens with a totally empty project.
bench.rs is completely empty.
main.rs has:

fn main() {
    println!("Hello, world!");
}

[Mark-Simulacrum]

I've confirmed I cannot reproduce on linux. I believe the hang occurs here, but I've not tracked this down further.

* thread #1, queue = 'com.apple.main-thread', stop reason = signal SIGSTOP
    frame #0: 0x0000000103f80550 bench-d5493d9f2fa74606`std::sys::unix::stack_overflow::imp::signal_handler::h373dcc530442aa18 (.llvm.16322965540172889217) + 112
bench-d5493d9f2fa74606`std::sys::unix::stack_overflow::imp::signal_handler::h373dcc530442aa18:
->  0x103f80550 <+112>: vmovdqa (%rax), %ymm0
    0x103f80554 <+116>: vmovaps 0x45c24(%rip), %ymm1
    0x103f8055c <+124>: vmovaps %ymm1, (%rax)
    0x103f80560 <+128>: movq   0x20(%rax), %rax
Target 0: (bench-d5493d9f2fa74606) stopped.

[TheIronBorn]

Confirm it does not happen with -C lto=thin

$ RUSTFLAGS='-C target-cpu=native -C lto=thin' cargo bench

[TheIronBorn]

-C lto=fat: yes
-C lto: yes

[bluejekyll]

I'm seeing a more general problem as well, where a binary running in an benchmark almost always hangs with lto, but when run independently hangs ~50% of the time. lto-thin appears to not have this issue.

I haven't investigated yet.

[steveklabnik]

I'm on Windows, but I cannot reproduce this today.

[workingjubilee]

Compiling from x86_64-unknown-linux-gnu to

• x86_64-unknown-linux-gnu
• x86_64-apple-darwin (cross)

No repro with Ctarget-cpu=sandybridge or -Ctarget-cpu=znver3 (native).
Closing, but feel free to bring this back if it repros on another x86_64-apple-darwin host.