Regression in global_allocator when using prefer-dynamic on 1.71.0 and above

[alexkornitzer]

Since 1.71.0 a regression has been added into the compiler causing segmentation faults in applications compiled using prefer-dynamic, this issue is not present in 1.70.0. It appears that the system allocator is still being used in some cases causing segmentation faults.

Code

https://github.com/alexkornitzer/dylib-errors/tree/error/jemalloc

use jemallocator::Jemalloc;

#[global_allocator]
static GLOBAL: Jemalloc = Jemalloc;

fn main() {
    let a = std::thread::spawn(move || {});
    a.join().unwrap();
    println!("didn''t crash");
}

Version it worked on

It most recently worked on: 1.70.0

Version with regression

1.71.0

Have run git bisect and the regression was introduced with the following commit: a2b1646

[lqd]

cc @bjorn3 for that bisection to #86844

[bjorn3]

As far as I understand this shouldn't work on Windows ever before my PR as on Windows symbol preemption across dll's doesn't happen, so inside std.dll it would resolve __rust_alloc to the system allocator unconditionally. cc #100781 (comment)

[alexkornitzer]

This is happening for me on macOS and Linux. What is odd if that cc above is accurate is that global_allocator and -C prefer-dynamic has worked for years for me, its only with 1.71.0 and above that its now broken.

[bjorn3]

I mean it should always have been broken on Windows as far as I understand and now is broken on other platforms too.

[alexkornitzer]

Ah right gotcha, just did not get why Windows was being brought in scope but then again I was not clear that linux was my deployment target and macos my development target. Is there anything I can do to help you fix the regression?