Segmentation fault cloning an empty ThinVec

[cuviper]

I encountered a segfault trying to initialize with vec![thinvec![]; N], which I managed to reproduce with a simple clone. In debug mode this runs fine, but it fails with optimizations on.

extern crate thincollections;

use thincollections::thin_vec::ThinVec;

fn main() {
    let v = ThinVec::<usize>::new();
    let _v = v.clone();
}

I added [profile.release] debug = true and gathered crash details with GDB:

Program received signal SIGSEGV, Segmentation fault.
<thincollections::thin_vec::ThinVec<T> as core::ops::deref::Deref>::deref (self=<optimized out>)
    at /home/jistone/.cargo/registry/src/github.com-1ecc6299db9ec823/thincollections-0.5.1/src/thin_vec.rs:1709
1709                    len = *len_ptr;
Missing separate debuginfos, use: dnf debuginfo-install libgcc-8.2.1-4.fc28.x86_64
(gdb) p len_ptr
$1 = (usize *) 0xffffffffffffffff
(gdb) bt
#0  <thincollections::thin_vec::ThinVec<T> as core::ops::deref::Deref>::deref (self=<optimized out>)
    at /home/jistone/.cargo/registry/src/github.com-1ecc6299db9ec823/thincollections-0.5.1/src/thin_vec.rs:1709
#1  <thincollections::thin_vec::ThinVec<T> as core::clone::Clone>::clone (self=0x7fffffffd178)
    at /home/jistone/.cargo/registry/src/github.com-1ecc6299db9ec823/thincollections-0.5.1/src/thin_vec.rs:2327
#2  0x000055555555a45e in thin::main () at src/main.rs:7

$ rustc -Vv
rustc 1.30.0 (da5f414c2 2018-10-24)
binary: rustc
commit-hash: da5f414c2c0bfe5198934493f04c676e2b23ff2e
commit-date: 2018-10-24
host: x86_64-unknown-linux-gnu
release: 1.30.0
LLVM version: 8.0

[cuviper]

Perhaps related, cargo test --release on this crate dies with SIGILL in stable_thinvec_tests.

[mohrezaei]

Thanks for the report! I'll have a look.

[cuviper]

I believe the problem is the use of ptr::null_mut() to construct the empty slice.

https://doc.rust-lang.org/stable/std/slice/fn.from_raw_parts.html

data must be non-null and aligned, even for zero-length slices.

The use of 1 as *const T and 1 as *mut T elsewhere also fails the alignment requirement. I think a better choice would be NonNull::dangling().as_ptr().

[mohrezaei]

I'll give that a try. It's odd, because that code is all over std::Vec.

[cuviper]

Maybe the alignment is not a big deal (yet), but non-null is definitely important for slices. Such references are given the nonnull attribute in LLVM IR, and optimizations can make heavy assumptions based on this.

[mohrezaei]

Fixed in 58739fd
Thanks for the report and the solution. I did change all the 1 as cases. They only occur for ZST types, so maybe that's why it just worked.

Published 0.5.2 with the fix.

[cuviper]

Thanks!

[cuviper]

FYI, your v0.5.1 and v0.5.2 git tags are pointing to the same commit 5376415 as v0.5.0. The actual published crate does appear to be correct though. (edited because I pasted the wrong commit ID too...)

[mohrezaei]

Yes, in all my excitement I forgot a step. I think I've pushed the right tag versions now. Thanks for the heads up.