Published binaries are kinda' slow

[lnicola]

The musl allocator is not one of the fastest ones out there. We should consider enabling jemalloc if possible.

$ target/release/rust-analyzer analysis-stats .                                                                                               Database loaded, 185 roots, 207.149744ms
Crates in this dir: 27
Total modules found: 391
Total declarations: 6479
Total functions: 4882
Item Collection: 12.928391044s, 0b allocated 0b resident
Total expressions: 134876                                                                                                               
Expressions of unknown type: 4420 (3%)
Expressions of partially unknown type: 4193 (3%)
Type mismatches: 271
Inference: 38.903396336s, 0b allocated 0b resident
Total: 51.83179408s, 0b allocated 0b resident

$ target/x86_64-unknown-linux-musl/release/rust-analyzer analysis-stats .                                                                                      Database loaded, 185 roots, 236.729051ms
Crates in this dir: 27
Total modules found: 391
Total declarations: 6479
Total functions: 4882
Item Collection: 14.239506998s, 0b allocated 0b resident
Total expressions: 134876                                                                                                               
Expressions of unknown type: 4438 (3%)
Expressions of partially unknown type: 4284 (3%)
Type mismatches: 271
Inference: 44.227304328s, 0b allocated 0b resident
Total: 58.466818431s, 0b allocated 0b resident

[matklad]

Does jemalloc work with musl target?

…

On Wed, 4 Mar 2020 at 18:37, Laurențiu Nicola ***@***.***> wrote: The musl allocator is not one of the fastest ones out there. We should consider enabling jemalloc if possible. $ target/release/rust-analyzer analysis-stats . Database loaded, 185 roots, 207.149744ms Crates in this dir: 27 Total modules found: 391 Total declarations: 6479 Total functions: 4882 Item Collection: 12.928391044s, 0b allocated 0b resident Total expressions: 134876 Expressions of unknown type: 4420 (3%) Expressions of partially unknown type: 4193 (3%) Type mismatches: 271 Inference: 38.903396336s, 0b allocated 0b resident Total: 51.83179408s, 0b allocated 0b resident $ target/x86_64-unknown-linux-musl/release/rust-analyzer analysis-stats . Database loaded, 185 roots, 236.729051ms Crates in this dir: 27 Total modules found: 391 Total declarations: 6479 Total functions: 4882 Item Collection: 14.239506998s, 0b allocated 0b resident Total expressions: 134876 Expressions of unknown type: 4438 (3%) Expressions of partially unknown type: 4284 (3%) Type mismatches: 271 Inference: 44.227304328s, 0b allocated 0b resident Total: 58.466818431s, 0b allocated 0b resident — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#3459>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AANB3MZ4Q4FEEAK334Z6INDRF2GV5ANCNFSM4LBOQTMQ> .

[lnicola]

It does, with a bit of prodding. I had to set my linker to musl-clang (remember that ~/.cargo/config tweak?). Looks like musl-jemalloc is on par with glibc.

$ target/release/rust-analyzer analysis-stats .                                                                                               Database loaded, 185 roots, 207.976195ms
Crates in this dir: 27
Total modules found: 391
Total declarations: 6479
Total functions: 4882
Item Collection: 12.205624894s, 534mb allocated 563mb resident
Total expressions: 134876                                                                                                               
Expressions of unknown type: 4429 (3%)
Expressions of partially unknown type: 4281 (3%)
Type mismatches: 271
Inference: 35.106950861s, 2107mb allocated 2185mb resident
Total: 47.312756967s, 2107mb allocated 2185mb resident

$ target/x86_64-unknown-linux-musl/release/rust-analyzer analysis-stats .                                                                                      Database loaded, 185 roots, 242.025021ms
Crates in this dir: 27
Total modules found: 391
Total declarations: 6479
Total functions: 4882
Item Collection: 13.722411216s, 534mb allocated 570mb resident
Total expressions: 134876                                                                                                               
Expressions of unknown type: 4448 (3%)
Expressions of partially unknown type: 4282 (3%)
Type mismatches: 271
Inference: 37.752254669s, 2104mb allocated 2183mb resident
Total: 51.474874324s, 2104mb allocated 2183mb resident

[memoryruins]

Noticed that ripgrep also enables jemalloc for 64-bit musl https://github.com/BurntSushi/ripgrep/blob/0874aa115c92f102a6ec474944f589667463fcd0/crates/core/main.rs#L29-L44

// However, when ripgrep is built with musl, this means ripgrep will use musl's
// allocator, which appears to be substantially worse. (musl's goal is not to
// have the fastest version of everything. Its goal is to be small and amenable
// to static compilation.) Even though ripgrep isn't particularly allocation
// heavy, musl's allocator appears to slow down ripgrep quite a bit. Therefore,
// when building with musl, we use jemalloc.
//
// We don't unconditionally use jemalloc because it can be nice to use the
// system's default allocator by default. Moreover, jemalloc seems to increase
// compilation times by a bit.
//
// Moreover, we only do this on 64-bit systems since jemalloc doesn't support
// i686.
#[cfg(all(target_env = "musl", target_pointer_width = "64"))]
#[global_allocator]
static ALLOC: jemallocator::Jemalloc = jemallocator::Jemalloc;

[lnicola]

Update: in #4157 we've switched to glibc binaries, so this is partly fixed, assuming it sticks. There's still the option of enabling jemalloc.

[matklad]

Yeah, I think we should maybe close this for now -- if the goal is better perf, the bigger impact would be in optimizing stuff inside rust-analyzer itself.