Rust 1.45+ generates suboptimal control flow for ARMv7 atomic compare exchange

Code

I tried this code:

https://rust.godbolt.org/z/vP88v5GGc

I expected to see this happen:

A tight loop.

Instead, this happened:

In Rust 1.45 and later, LLVM generates a switch + unconditional branch to a conditional branch. This happens in opt-level=z, s, 1, 2, 3.

In some versions generates a dead branch, e.g.

mov r2, #0
cmp r2, #0
bne ... # If this branch is taken, hell has frozen over.

Version it worked on

It most recently worked on: 1.44

Version with regression

rustc --version --verbose:

rustc 1.79.0-nightly (0824b300e 2024-03-24)
binary: rustc
commit-hash: 0824b300eb0dae5d9ed59719d3f2732016683d66
commit-date: 2024-03-24
host: x86_64-unknown-linux-gnu
release: 1.79.0-nightly
LLVM version: 18.1.2

@rustbot modify labels: +regression-from-stable-to-stable -regression-untriaged

Start with the Rust Compiler Explorer reproduction linked in the issue, comparing ARMv7 output from Rust 1.44 with 1.45 and later at the listed optimization levels. Trace the rustc/LLVM code-generation path for atomic compare exchange and identify the change responsible; done means the generated control flow no longer includes the reported unnecessary switch, branch, or dead branch.