The Problem

v0.8.15 introduced for-range loop emission — converting While i is at most n to for i in 1..=n. Benchmarks revealed bubble sort regressed by 41.4% at N=2000.

Root cause: Rust's RangeInclusive (..=) carries per-iteration overhead. The internal ExactSizeIterator implementation needs bookkeeping to handle the edge case where the range includes T::MAX. In an O(n^2) inner loop, this overhead compounds.

The test that caught it (tier1a_simple_counting_loop in phase_optimize.rs):

let rust = compile_to_rust(source).unwrap();
assert!(rust.contains("for i in 1..6"));

The assertion checks that the generated Rust uses an exclusive range (1..6) rather than an inclusive one (1..=5).

The Fix

All inclusive ranges now emit the exclusive form with limit + 1:

Before (v0.8.15)After (v0.8.16)
for i in 1..=5for i in 1..6
for i in 1..=nfor i in 1..(n + 1)
for j in 1..=((n - 1) - i)for j in 1..(((n - 1) - i) + 1)

For literal limits, the addition is computed at compile time — 1..6 not 1..(5 + 1).

Exclusive < bounds are unchanged: for i in 0..5 was already correct.

Bubble Sort Codegen (After Fix)

The inner loop of the bubble sort benchmark now generates:

for j in 1..(((n - 1) - i) + 1) {
    if arr[(j - 1) as usize] > arr[((j + 1) - 1) as usize] {
        arr.swap((j - 1) as usize, ((j + 1) - 1) as usize);
    }
}

No ..= anywhere in generated code. Range (exclusive) uses a simple counter increment with no bookkeeping.