Performance regression between cc 1.0.90 and 1.0.91, regression is also present in 1.0.96

[AlexanderSchuetz97]

Hello, I am noticing a massive decrease in performance of the CC dependency when updating to version 1.0.91.

When used with mingw to build a windows shared library the decrease in performance is about factor 2. It used to take 30 seconds to build my project now it takes 60 seconds.

To build a linux shared library I use cargo zigbuild. The difference is about factor 7 with zigbuild, It used to take about 50 seconds, now it takes well above 7 minutes. Nothing else has changed in my project other than the cc version. Between each test I ran cargo clean to ensure that it does a clean build. I think something is up here.

My project features a large percentage of C code. (About 80% is C code, the rest is rust) I dont have any C++ in my project.

I would not have reported this issue here if it was just zigbuild that was affected, even the normal mingw toolchain is affected.

Does anyone have any idea on significant changes betwen those versions that could affect performance? The only observable difference in terms of log output appears to be that warnings about gcc version mismatch that used to be there are now missing when building the windows library. (Which is positive I guess... but not worth the horrible performance)

These are my dependencies and cargo configuration:

[package]
name = "XXXXXXXXXXXX"
version = "0.1.0"
edition = "2021"

[lib]
name = "XXXXXXXXXXXXXXXXX"
crate-type = ["cdylib"]

[dependencies]
jni = "0.21.1"
half = "2.4.0"
bytemuck = "1.15.0"
region = "3.0.1"
lazy_static = "1.4"

[build-dependencies]
cc = "1.0.90"

[profile.release]
strip = true
opt-level = 3
codegen-units = 1
lto = "fat"

This is my build.rs, the only thing configured here is the cc crate

use cc;

fn main() {
    let src = &[
        //lots of stuff Stuff goes here
    ];
    cc::Build::new()
        .files(src)
        .include("stuff goes here")
        .define("__ANSI__", None)
        .flag_if_supported("-Wno-constant-conversion")
        .flag_if_supported("-Wno-unused-const-variable")
        .flag_if_supported("-Wno-deprecated-declarations")
        .flag_if_supported("-Wno-comment")
        .flag_if_supported("-Wno-unused-value")
        .flag_if_supported("-Wno-unused-function")
        .flag_if_supported("-Wno-unknown-pragmas")
        .flag_if_supported("-Wno-extra-tokens")
        .flag_if_supported("-Wno-missing-field-initializers")
        .flag_if_supported("-Wno-shift-negative-value")
        .flag_if_supported("-Wno-dangling-else")
        .flag_if_supported("-Wno-sign-compare")
        .flag_if_supported("-Wno-strict-aliasing")
        .flag_if_supported("-Wno-implicit-fallthrough")
        .flag_if_supported("-Wno-old-style-declaration")
        .flag_if_supported("-Wno-endif-labels")
        .flag_if_supported("-Wno-parentheses")
        .flag_if_supported("-Wno-misleading-indentation")
        .flag_if_supported("-Wno-unused-but-set-variable")
        .opt_level(3)
        .compile("mylib");
}

As mentioned the only change needed to tank into oblivion the compilation speed is to update the cc crate. This entire build is done on windows 11.

[dpaoliello]

Given the number of calls to flag_if_supported, it would guess that https://github.com/rust-lang/cc-rs/commit/be62f4a9027a30ccbfffa494ce8ee3487de8cb9b is the culprit.

[dpaoliello]

@AlexanderSchuetz97 can you please try https://github.com/dpaoliello/cc-rs/tree/perf?

[patch.crates-io]
cc = { git = "https://github.com/dpaoliello/cc-rs.git", branch = "perf" }

[NobodyXu]

@dpaoliello Thanks for the quick response!

Though I'd have to say that adding a new parameter to a public function would break many crates.

I'd suggest to

• create a new function is_flag_supported_inner which takes a compiler_path
• use Build::get_base_compiler() directly to retrieve the compiler in is_flag_supported, which might improve performance

[dpaoliello]

Ah, didn't realize that is_flag_supported was exported - I've applied your suggestions (I can also create a PR if you think it's worthwhile to make this change anyway)

[NobodyXu]

Yeah I think it's worth a PR regardless of if it fixes the regression.

[AlexanderSchuetz97]

I will only be able to access this project on monday. I will perform the perf suggestion then and report my findings.

[dpaoliello]

Done #1050

[AlexanderSchuetz97]

Hello again, this fix appears to significantly improve performance. It is still 5-10% slower than using version 1.0.90, however the time to compile is acceptable for me now. When can I expect v 1.0.97 (Which will presumably contain the fix from dpoliello?)

[NobodyXu]

I was waiting for one last nit to be resolved, but I can edit the PR directly and merge it.

Would cut a release soon after merging it.

[NobodyXu]

It also looks like we need a performance benchmark suite in cc

[AlexanderSchuetz97]

It also looks like we need a performance benchmark suite in cc

This would probably be a good idea.

Would cut a release soon after merging it.

Thank you.

Since the PR is merged I will close this issue. You should track the benchmark implementation in issue #1065