gen-zig/gen-verilog: CSE hoists an expression over a reassigned local above its declaration, printed as ID:/LIT: debug text #7703

Part of the self-host epic #5980. Following AGENTS.md "Only t27", what gen cannot express is a compiler defect and is filed here.

Goal

A repeated expression over a reassigned local must not be hoisted by CSE in gen-zig and gen-verilog.

Where we are

Found while writing specs/fpga/eth_beacon.t27 (#6655 Ethernet bring-up). Minimal repro:

module CseRepro;
fn fold(x: u32) -> u32 {
    var s : u32 = x;
    s = (s & 65535) + (s >> 16);
    s = (s & 65535) + (s >> 16);
    return s;
}
test fold_carries_twice {
    assert_eq(fold(131071), 1);
}

t27c gen (t27c 0.5.0, tag t27c-v0.5.0, cb9df57) emits:

fn fold(x: u32) u32 {
    const _cse1 = s & 65535 + ID:s >> LIT:16;
    var s: u32 = x;
    _ = &s;
    s = (s & 65535) + (s >> 16);
    s = _cse1;
    return s;
}

gen-verilog emits the same: _cse1 = (s & 65535 + ID:s >> LIT:16);. gen-c is unaffected.

There are three defects in one:

1. The CSE temp is hoisted above the declaration of the variable it reads.
2. It treats the two textually equal right-hand sides as one value, although s is reassigned between them. The second fold would reuse the first's result: wrong semantics, silent if the text were valid.
3. The temp is printed in the internal debug form (ID:s, LIT:16), so today the output does not compile. That is the only reason the defect is loud.

Scope note

The CSE pass shared by gen-zig and gen-verilog, plus a regression test.

User Scenarios & Testing

• Given the repro, when t27c gen / gen-verilog run, then no _cse temp reads s, and zig test passes with fold(131071) == 1.

Requirements

• FR-1: an expression that reads a variable assigned between its occurrences is not a common subexpression.
• FR-2: a hoisted temp is never emitted before the declarations it reads.
• FR-3: no ID: / LIT: debug text reaches any backend's output.

Success Criteria

• t27c gen cse_repro.t27 > r.zig && zig test r.zig passes.
• t27c gen-verilog cse_repro.t27 | grep -c 'ID:' prints 0.

Workaround used in eth_beacon.t27: immutable intermediates (const a, const b, const c) instead of reassigning one var.

🤖 Generated with Claude Code

Boundary

• specs/fpga/eth_beacon.t27