Use mimalloc as the global allocator on x86_64-pc-windows-msvc by Zoxc · Pull Request #138764 · rust-lang/rust (original) (raw)
This adds using mimalloc as the global allocator for the compiler as an option to bootstrap.toml and enables it for the x86_64-pc-windows-msvc dist builder.
try-job: dist-x86_64-msvc
rustbot has assigned @Mark-Simulacrum.
They will have a look at your PR within the next two weeks and either review your PR or reassign to another reviewer.
Use r? to explicitly pick a reviewer
Area: The testsuite used to check the correctness of rustc
Area: The tidy tool
Status: Awaiting review from the assignee but also interested parties.
Relevant to the bootstrap subteam: Rust's build system (x.py and src/bootstrap)
Relevant to the compiler team, which will review and decide on the PR/issue.
Relevant to the infrastructure team, which will review and decide on the PR/issue.
labels
Local benchmarks:
| Benchmark | Before | After | Before | After | Before | After | |||
|---|---|---|---|---|---|---|---|---|---|
| Time | Time | % | Physical Memory | Physical Memory | % | Committed Memory | Committed Memory | % | |
| 🟣 clap:check | 1.4133s | 1.2131s | 💚 -14.17% | 147.10 MiB | 147.52 MiB | 0.28% | 200.42 MiB | 261.79 MiB | 💔 30.62% |
| 🔵 clap:release | 5.7162s | 5.4820s | 💚 -4.10% | 285.77 MiB | 289.07 MiB | 💔 1.15% | 324.74 MiB | 384.72 MiB | 💔 18.47% |
| 🟠 clap:debug | 3.2341s | 2.9639s | 💚 -8.35% | 289.52 MiB | 305.29 MiB | 💔 5.44% | 331.62 MiB | 394.30 MiB | 💔 18.90% |
| 🟣 hyper:check | 0.2368s | 0.2082s | 💚 -12.07% | 80.00 MiB | 80.85 MiB | 💔 1.06% | 130.63 MiB | 195.31 MiB | 💔 49.51% |
| 🔵 hyper:release | 0.8620s | 0.8324s | 💚 -3.42% | 100.95 MiB | 103.28 MiB | 💔 2.31% | 144.99 MiB | 201.68 MiB | 💔 39.10% |
| 🟠 hyper:debug | 0.4746s | 0.4364s | 💚 -8.05% | 113.33 MiB | 115.87 MiB | 💔 2.24% | 152.39 MiB | 213.92 MiB | 💔 40.38% |
| 🟣 regex:check | 0.8054s | 0.6930s | 💚 -13.96% | 108.49 MiB | 108.13 MiB | -0.34% | 155.61 MiB | 223.32 MiB | 💔 43.51% |
| 🔵 regex:release | 5.2164s | 5.0395s | 💚 -3.39% | 311.05 MiB | 389.23 MiB | 💔 25.13% | 294.30 MiB | 428.30 MiB | 💔 45.53% |
| 🟠 regex:debug | 1.8545s | 1.7079s | 💚 -7.91% | 239.77 MiB | 242.05 MiB | 0.95% | 275.22 MiB | 324.94 MiB | 💔 18.07% |
| 🟣 syn:check | 1.3455s | 1.1368s | 💚 -15.51% | 142.38 MiB | 141.26 MiB | -0.78% | 190.93 MiB | 255.30 MiB | 💔 33.72% |
| 🔵 syn:release | 4.2406s | 3.9707s | 💚 -6.37% | 256.63 MiB | 296.60 MiB | 💔 15.58% | 295.81 MiB | 349.53 MiB | 💔 18.16% |
| 🟠 syn:debug | 2.5499s | 2.2981s | 💚 -9.87% | 236.51 MiB | 242.18 MiB | 💔 2.39% | 276.62 MiB | 358.63 MiB | 💔 29.65% |
| Total | 27.9492s | 25.9819s | 💚 -7.04% | 2.26 GiB | 2.40 GiB | 💔 6.48% | 2.71 GiB | 3.51 GiB | 💔 29.51% |
| Summary | 1.0000s | 0.9107s | 💚 -8.93% | 1 byte | 1.05 bytes | 💔 4.62% | 1 byte | 1.32 bytes | 💔 32.14% |
We do see the memory use regressions compared to Windows' allocator too.
just for my interest: does it work for windows-gnu too?
I haven't tested that, but probably.
bors added a commit to rust-lang-ci/rust that referenced this pull request
Use mimalloc as the global allocator on x86_64-pc-windows-msvc
This adds using mimalloc as the global allocator for the compiler as an option to bootstrap.toml and enables it for the x86_64-pc-windows-msvc dist builder.
try-job: dist-x86_64-msvc
These benchmarks also seem pretty small, what's the result on a bigger project like cargo?
☀️ Try build successful - checks-actions
Build commit: b200834 (b200834f64a4cc752eb39d638ca76d023af0e45f)
| @@ -748,6 +748,12 @@ |
|---|
| # [target.] section. |
| #jemalloc = false |
| # Use mimalloc as the global allocator for the compiler. Rust code will prefer `mimalloc` over |
| # `jemalloc` if that is also enabled. |
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Could we reject setting both? That feels needlessly confusing to me and probably not hugely needed?
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
We can probably just remove jemalloc if mimalloc v3 solves all regressions over it.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Maybe name this field override-allocators in case we'd want to swap allocators in the future?
| @@ -70,6 +70,10 @@ use time::OffsetDateTime; |
|---|
| use time::macros::format_description; |
| use tracing::trace; |
| #[cfg(feature = "mimalloc")] |
| #[global_allocator] |
| static GLOBAL: mimalloc::MiMalloc = mimalloc::MiMalloc; |
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Is this just temporary, or the long-term state? As-is this doesn't affect LLVM's malloc calls?
IOW, why is this not changign the rustc/src/main.rs malloc overrides?
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Long term. It's better than the linker hacks in main.rs.
Overriding the C allocator only in main.rs doesn't affect DLLs making it unsound without further work dealing with loadable backends and LLVM plugins.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
But isn't this still a regression for LLVM's allocations? It seems odd to measure perf for a halfway shift like this to mimalloc...
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
We'd still want the linker hacks on Linux / macOS for LLVM, using global_allocator just makes Rust allocation overriding less hacky.
As for perf on this PR, LLVM uses Windows' heap allocator both before and after.
Nit: if we modify the bootstrap.toml config, I would go for allocator = jemalloc/mimalloc/default, rather than introducing another boolean option.
Is there a noticeable difference with jemalloc on windows? Given, that it already exist but not enabled for windows.
I haven't tried to make jemalloc work on Windows. I focused on mimalloc since it outperformed jemalloc on our Linux tests.