Summary

LLVM optimizes away a redundant checked integer conversion inside a loop on x86_64, but not on s390x.

This reproduces with the exact reduced shape from Rust’s codegen-llvm test tests/codegen-llvm/range-len-try-from.rs, compiled with optimization enabled. The call to unwrap_failed for the loop-local conversion u32::try_from(i).unwrap() is proven redundant on x86_64, but remains as an extra path on s390x.

The same Rust source produces the same high-level checked-conversion structure before optimization, and the divergence only appears after target-specific LLVM optimization (x86_64 removes the inner unwrap_failed path while s390x does not), which strongly suggests the missed optimization is in LLVM rather than Rust.

Targets

• affected: s390x-unknown-linux-gnu
• optimized as expected: x86_64-unknown-linux-gnu

Observed behavior

With -Copt-level=3:

• x86_64: the function contains only one unwrap_failed call, corresponding to the expected outer check:
  • u32::try_from(candidates.len()).unwrap()
• s390x: the function contains two unwrap_failed calls:
  • the expected outer length check
  • an extra path from the loop-local u32::try_from(i).unwrap()

This suggests LLVM is failing on s390x to prove that the loop index is bounded by the already-checked slice length.

Reproducer

use std::convert::TryFrom;
use std::hint::black_box;

#[no_mangle]
pub fn score_round_enumerate(candidates: &[bool]) {
    u32::try_from(candidates.len()).unwrap();

    for (i, _) in candidates.iter().enumerate() {
        u32::try_from(i).unwrap();
        black_box(42);
    }
}

Compile:

rustc -Copt-level=3 --emit=llvm-ir --crate-type=lib reproduce.rs --target s390x-unknown-linux-gnu
rustc -Copt-level=3 --emit=llvm-ir --crate-type=lib reproduce.rs --target x86_64-unknown-linux-gnu

Expected result

Both targets should retain only the single unwrap_failed associated with: u32::try_from(candidates.len()).unwrap();
The loop-local conversion unwrap_failed call (u32::try_from(i).unwrap();) should be optimized away on both targets.

Actual result

On s390x the function retain an extra unwrap_failed from the loop-local checked conversion. On x86_64, that extra unwrap_failed path is removed.

Attachments:

reproduce.rs.txt
reproduce_s390x_noopt.ll.txt
reproduce_s390x_o3.ll.txt
reproduce_x86_64_noopt.ll.txt
reproduce_x86_64_o3.ll.txt