buildRustCrate: do not forcibly disable rustc parallelism #274135
Conversation
buildRustCrate currently sets `codegen-units` to `1` if unspecified. This option controls the amount of process-level parallelism that LLVM employs during the build. The [default is supposed to be 16](https://doc.rust-lang.org/rustc/codegen-options/index.html#codegen-units). By forcing this to `1` we are disabling all intra-derivation parallelism. This makes `buildRustCrate` and `crate2nix` much slower than `cargo`; we shouldn't do that!
ghost
requested review from
ofborg
bot
added
10.rebuild-darwin: 0
delroth
added
the
12.approvals: 1
There was a problem hiding this comment.
I followed git blame a bit - this has been set to a static value due to nondeterminism if using NIX_BUILD_CORES in #170981.
The discussion in that same PR mentions the following too:
the difference between codegen-units=1 and codegen-units=16 was a bit less than a factor of 2 of build time on my 24-core machine
It also links to #130309 (comment), which prefers 1 due to performance of the resulting binary.
Your PR description "[…] cripples all parallelism […]" suggests a much higher compilation performance penalty.
I'm also not sure about the implications on runtime performance in real-world scenarios. Do you have any numbers to back both things up?
ghost
closed this
|
What is the reason for closing this PR ? |
|
I think this PR was closed in in the light of #50105 (comment). |
|
This pull request has been mentioned on NixOS Discourse. There might be relevant details there: https://discourse.nixos.org/t/crate2nix-setting-codegenunits-for-all-crates/53014/2 |
|
This pull request has been mentioned on NixOS Discourse. There might be relevant details there: https://discourse.nixos.org/t/crate2nix-setting-codegenunits-for-all-crates/53014/3 |
Our
buildRustCratecurrently cripples all parallelism, unlike ourbuildRustPackage. We should not do this.Our
buildRustCratefor setscodegen-unitsto1if unspecified. Thecodegen-unitsoption controls the amount of process-level parallelism that LLVM employs during the build.The default is supposed to be 16, but we should simply leave the option unspecified in order to track any future changes to the default. By forcing this to
1we are disabling all intra-derivation parallelism. This makesbuildRustCrateandcrate2nixmuch slower thancargo, for no good reason.