Profile memory_usage is broken on Mac OS Monterey

[AlexMikhalev]

I noticed that profile memory_usage doesn't work on Mac, I extracted https://github.com/AlexMikhalev/profile and added simple basic example to illustrate.

use profile::memory_usage;
fn main() {

    let before = memory_usage().allocated;
    println!("Before {:?}",before.to_string());
    let result = 2 + 2;
    println!("After {:?}",memory_usage().allocated.to_string());
    assert_eq!(result, 4);
}

cargo run --example basic
   Compiling profile v0.0.0 (/Users/alexmikhalev/rust_code/profile)
    Finished dev [unoptimized + debuginfo] target(s) in 2.57s
     Running `target/debug/examples/basic`
Before "0b"
After "0b"

Works on Linux but doesn't work on Mac, which means caches are not cleaned up from here

rust-analyzer version: rust-analyzer version: 7d6fcbc 2021-12-06 stable

rustc version: (eg. output of rustc -V)

rustc -vV
rustc 1.60.0 (7737e0b5c 2022-04-04)
binary: rustc
commit-hash: 7737e0b5c4103216d6fd8cf941b7ab9bdbaace7c
commit-date: 2022-04-04
host: x86_64-apple-darwin
release: 1.60.0
LLVM version: 14.0.0

[lnicola]

As you can see in the implementation, this feature uses some special memory allocator APIs which are only available on Linux and jemalloc. Building with --feature jemalloc should make it work.

[lnicola]

Note that there are some, potentially obsolete, reasons why we don't build with jemalloc by default: rust-lang/rust#36963 (comment).

[AlexMikhalev]

Thank you, building with --features jemalloc works. I would suggest re-visiting building with jmallock support for rust-analyzer, I found cache cleaning in apply_change.rs quite clean and if mac os users don't have cache cleaned it may lead to performance degradation for rust-analyzer users.

[lnicola]

You're not supposed to use that command unless you're debugging the memory usage. What do you mean by that?

[AlexMikhalev]

Actually, while it runs it produces the same number on Mac - it's a different number between before and after on Linux how it should.

cargo run --example basic --features jemalloc
Before "35kb"
After "35kb"

with --features jemalloc it always returns "35kb" instead of 0b (without --features jemalloc) and is never near the correct number, even for large datasets.

[lnicola]

Adding two numbers doesn't really allocate anything. What exactly are you trying to do?

[AlexMikhalev]

    let before = memory_usage().allocated;
    println!("Before {:?}",before.to_string());
    let v: Vec<_> = (1..100000000).collect();
    println!("After {:?}",memory_usage().allocated.to_string());

still produces

Before "35kb"
After "35kb"

same as above two numbers.

[bjorn3]

It is possible that the memory allocation is optimized away by LLVM. Try adding unsafe { std::arch::asm!("", in(reg) &v); } to ensure LLVM can't remove the allocation.

[AlexMikhalev]

unsafe { std::arch::asm!("", in(reg) &v); } | ^^^^^^^^^^ argument never used
I am measuring memory usage of a larger dataset ~500Kb on linux and which still output 35Kb on mac.

[lnicola]

Did you set the global allocator as in https://github.com/rust-lang/rust-analyzer/blob/master/crates/rust-analyzer/src/bin/main.rs#L20=?

[AlexMikhalev]

Thank you @lnicola, adding global allocator and jemallocator into dependencies/default features helped.
If you guys would publish the profile crate separately it will be re-usable and very handy.
I close the issue - resolved.