Miscompilation while working on PR #50882 #52694
Description
(Moving this to a separate issue because PR #50882 is already full of noise)
Reduced STR:
- Take current rust master
- Apply the following patch:
diff --git a/src/liballoc/alloc.rs b/src/liballoc/alloc.rs
index 84bd275..4df7730 100644
--- a/src/liballoc/alloc.rs
+++ b/src/liballoc/alloc.rs
@@ -172,7 +172,7 @@ pub(crate) unsafe fn box_free<T: ?Sized>(ptr: Unique<T>) {
// We do not allocate for Box<T> when T is ZST, so deallocation is also not necessary.
if size != 0 {
let layout = Layout::from_size_align_unchecked(size, align);
- dealloc(ptr as *mut u8, layout);
+ Global.dealloc(NonNull::new_unchecked(ptr).cast(), layout);
}
}
- Compile rust against the system llvm (5 or 6)
- Run libstd tests (
x.py test src/libstd --stage 1)
What happens then is that sync::once::tests::wait_for_force_to_finish fails with:
thread '<unnamed>' panicked at 'assertion failed: t1.join().is_ok()', libstd/sync/once.rs:582:9
The disassembly for wait_for_force_to_finish only contains one call to std::thread::JoinHandle::join, instead of the expected two when the code is not miscompiled. That call is followed by a test that jumps to a panic when the result of JoinHandle::join is ... Ok(()):
93371: e8 aa fd 07 00 callq 113120 <_ZN41_$LT$std..thread..JoinHandle$LT$T$GT$$GT
$4join17h0fb3b129ec2e38aeE>
93376: 48 89 c3 mov %rax,%rbx
93379: 49 89 d7 mov %rdx,%r15
9337c: 48 85 db test %rbx,%rbx
9337f: 74 1b je 9339c <_ZN3std4sync4once5tests24wait_for_force_to_fin
ish17h6199051fcaa3ff6aE+0x21c>That JoinHandle::join returns a Result<(), Box<Any>>, and Ok(()) is represented as (0, whatever). The destination of that jump is the panic code.
At some point, I'm not entirely sure with what state of the tree, it was even better. The error was:
thread '<unnamed>' panicked at 'assertion failed: t2.join().is_ok()', libstd/sync/once.rs:583:9
And there were two calls to std::thread::JoinHandle::join, as expected, but they didn't have the same result handling:
93c79: e8 32 c5 07 00 callq 1101b0 <_ZN41_$LT$std..thread..JoinHandle$LT$T$GT$$GT$4join17hc6e7f9bb7d72483aE>
93c7e: 48 89 c3 mov %rax,%rbx
93c81: 49 89 d7 mov %rdx,%r15
93c84: 48 85 db test %rbx,%rbx
93c87: 75 5e jne 93ce7 <_ZN3std4sync4once5tests24wait_for_force_to_finish17h13c0ef8dd5eb6a3aE+0x267>
(snip)
93c9f: e8 0c c5 07 00 callq 1101b0 <_ZN41_$LT$std..thread..JoinHandle$LT$T$GT$$GT$4join17hc6e7f9bb7d72483aE>
93ca4: 48 89 c3 mov %rax,%rbx
93ca7: 49 89 d7 mov %rdx,%r15
93caa: 48 85 db test %rbx,%rbx
93cad: 74 1b je 93cca <_ZN3std4sync4once5tests24wait_for_force_to_finish17h13c0ef8dd5eb6a3aE+0x24a>The destination of both jumps is panic code. The first jump, corresponding to t1.join().is_ok() is correct, and the second, corresponding to t2.join().is_ok() is broken, thus the test failure.
Even better: this doesn't happen when compiling with the bundled llvm. It also doesn't happen when extracting the test from libstd and compiling with a faulty compiler. It seems the fact that it's part of libstd, and that most of libstd is compiled along the test, plays a role.
I'm trying to get the corresponding mir and llvm-ir.