[ICE]: rustc nightly SIGSEGV in llvm::LiveRange::join with -C opt-level=2

[466774321-code]

Code

pub static mut LIMIT: u16 = 16;
pub static mut TOTAL: u64 = 64;
pub static mut IS_GE: bool = false;
pub static mut IS_NE: bool = false;
pub static mut RUN: bool = false;

pub unsafe fn repro(x: u16) {
    if RUN {
        IS_GE = (x as i16) >= (LIMIT as i16);
        IS_NE = LIMIT != x;
        TOTAL = x as u64 + 630_600_177;
        panic!();
    }
}

Command used:

rustc +nightly --crate-type=lib -C opt-level=2 repro.rs

Meta

rustc --version --verbose:

rustc 1.100.0-nightly (8fa1c96cf 2026-08-17)
binary: rustc
commit-hash: 8fa1c96cfd489e4c27654c144ae871ce2c4db6c6
commit-date: 2026-08-17
host: x86_64-unknown-linux-gnu
release: 1.100.0-nightly
LLVM version: 23.1.0

Error output

error: rustc interrupted by SIGSEGV, printing backtrace

Backtrace:

librustc_driver-fb6ae9a3e0ae9f81.so
libc.so.6
libLLVM.so.23.1-rust-1.100.0-nightly
llvm::LiveRange::join(...)
libLLVM.so.23.1-rust-1.100.0-nightly
llvm::FPPassManager::runOnFunction(...)
llvm::FPPassManager::runOnModule(...)
llvm::legacy::PassManagerImpl::run(...)
librustc_driver-fb6ae9a3e0ae9f81.so
libc.so.6

Running with machine instruction verification:

rustc +nightly --crate-type=lib -C opt-level=2 \
  -Cllvm-args=-verify-machineinstrs repro.rs

produces:

# After Peephole Optimizations

*** Bad machine code: Virtual register defs don't dominate all uses. ***
- v. register: %9
rustc-LLVM ERROR: Found 1 machine code errors.

Disabling the peephole optimizer allows the code to compile:

rustc +nightly --crate-type=lib -C opt-level=2 \
  -Cllvm-args=-disable-peephole repro.rs

After finding this bug, I searched for related issues and found llvm/llvm-project#208746, which may have the same underlying cause.

[maxdexh]

Also happens on 1.99.0-beta.1 and/or opt-level=1; probably LLVM 23

[maxdexh]

Slightly simplified

pub static mut LIMIT: u16 = 16;
pub static mut TOTAL: u64 = 64;
pub static mut IS_GE: bool = false;
pub static mut IS_NE: bool = false;

pub unsafe fn repro(x: u16) {
    IS_GE = x >= LIMIT;
    IS_NE = LIMIT != x;
    TOTAL = x as u64;
    std::process::abort(); // any non-trivial call, e.g. `thread::spawn(|| {})`
}

[theemathas]

cc @nikic

[nikic]

Yes, this does seem to be the same as llvm/llvm-project#208746, which already has an open PR.

[maxdexh]

Is it?
The generated IR looks nothing alike and I'm getting a different LLVM error:

• LLVM Issue: https://godbolt.org/z/T7qb68oEf
• IR generated by from issue: https://godbolt.org/z/s38nrjMns (generated from the simplfied code above)

[nikic]

@maxdexh It's more obvious if you look at the reduced test case from the associated PR (llvm/llvm-project#211479).

The actually relevant part of the reproducer are the load and the icmps.