1
0
Fork 0
mirror of https://github.com/imjasonh/nescript synced 2026-07-08 08:55:38 +00:00
nescript/examples/war/COMPILER_BUGS.md
Claude 4e8e349d7c
ir: clear wide_hi between functions to fix 16-bit op aliasing
The IrLowerer's wide_hi map records "this u8 temp's high byte
lives at this other temp" pairs whenever a 16-bit value is
produced. Both lower_function and lower_handler reset next_temp
to 0 at the start of each function, but neither cleared wide_hi
— so stale (low_id -> high_id) entries from earlier functions
leaked into subsequent ones.

When a fresh function reused those temp IDs for unrelated u8
expressions, is_wide() returned spurious true and widen() handed
back stale (lo, hi) pairs whose hi happened to coincide with the
*next* temp ID fresh_temp() was about to allocate. The result
was 16-bit IR ops (CmpEq16 in particular) where the destination
temp aliased one of the source operand high bytes — for War this
made `match phase` arms past P_WIN_B impossible to enter and the
game would freeze with both face-up cards on the table forever.

Fix: clear wide_hi alongside the next_temp reset in both
lower_function and lower_handler. Adds a regression test
(ir::tests::wide_hi_does_not_leak_between_functions) that
constructs a function whose body has no u16 ops but follows a
function that does, and asserts no CmpEq16 op aliases its dest
with an operand high byte.

Also:
- Convert the war Playing state's phase machine from an
  if-chain to a `match`, which is what tripped this bug to the
  surface (it was lurking in earlier ROMs too but their layouts
  never produced the dest/source collision shape).
- Refactor begin_draw_a/b to set fly_card / fly_face_up via
  globals before calling arm_fly, since arm_fly only takes 4
  params (the v0.1 ABI limit, now diagnosed by E0506).
- Hoist the P_RESOLVE comparison result to the global pf_result
  to dodge the param-clobbering issue documented in
  examples/war/COMPILER_BUGS.md §2.
- Document the bug as item #6 in COMPILER_BUGS.md with a
  minimal repro and reproducer-test pointer.
- Refresh the war golden + audio hash to match the new ROM.

https://claude.ai/code/session_0143dTgh3UeRrtfHgQwzcv5z
2026-04-15 15:57:26 +00:00

17 KiB

NEScript v0.1 — Compiler Bugs and Limitations Found While Building War

This document captures bugs and limitations discovered while building examples/war.ne. Each entry includes a minimal reproduction, the symptom we observed, the root cause if known, and a workaround we used in examples/war/*.ne. The intent is to track these so they can be fixed in a future compiler pass — once they are, the corresponding workarounds in war/*.ne should be reverted to keep the example honest.


1. Functions with more than 4 parameters silently corrupt the 5th+

Symptom

Calling a function with 5 or 6 parameters compiles cleanly, with no warning or error, but at runtime the 5th and 6th parameter values are silently replaced by garbage (typically the value of parameter 3 or 4). Animations and state writes that depend on those parameters behave as if zero was passed.

Reproduction

fun arm_fly(sx: u8, sy: u8, dxsign: u8, dysign: u8, card: u8, fu: u8) {
    fly_x = sx
    fly_y = sy
    fly_dx_sign = dxsign
    fly_dy_sign = dysign
    fly_card = card        // gets the value of dxsign instead!
    fly_face_up = fu       // gets the value of dxsign instead!
}

fun caller() {
    arm_fly(32, 64, 0, 0, 147, 1)
    // After this call:
    //   fly_x = 32, fly_y = 64, fly_dx_sign = 0, fly_dy_sign = 0
    //   fly_card = 0   (NOT 147)
    //   fly_face_up = 0 (NOT 1)
}

Root cause

src/codegen/ir_codegen.rs (around line 240) iterates through func.locals and assigns the first 4 entries to zero-page parameter slots $04-$07:

for func in &ir.functions {
    for (i, local) in func.locals.iter().enumerate() {
        if i < func.param_count {
            if i < 4 {
                var_addrs.insert(local.var_id, 0x04 + i as u16);
                ...
            }
        } else {
            ...
        }
    }
}

The if i < 4 guard silently drops the mapping for params 5+ without inserting any RAM allocation for them. The corresponding caller-side codegen for Call writes only the first four arguments. Result: params 5 and 6 are never passed and the callee reads stale memory from $04-$07 in their place.

Workaround used in examples/war/

arm_fly is split: the four "arming" parameters stay in the function signature, and fly_card / fly_face_up are written to the global state directly at every call site instead. See war/play_state.ne (begin_draw_a / begin_draw_b).

Fix proposal

Two reasonable options:

  1. Diagnose-only: emit E05XX too many parameters when a fun declaration has more than 4 params. This is the smallest possible change and turns silent miscompiles into a loud compile-time error. Should ship immediately even if option 2 is also planned.

  2. Spill to RAM: extend the calling convention so params beyond the first four are passed via dedicated RAM slots in the callee's local frame. The caller-side Call codegen would write those slots before JSR, the callee-side prologue could leave them as-is. This grows the per-function RAM footprint but lets users write any signature they like.


1b. Function parameters with the same name in different functions share a VarId, which collides their zero-page slot mapping

Symptom

Two unrelated functions whose parameters happen to be named the same (e.g. both have a card: u8 parameter, or both have an x: u8 parameter) end up reading parameters from the wrong zero-page slot at runtime. One function reads $04, another reads $06, a third reads $05 — depending on the parameter's position in whichever function is processed last by the codegen.

This is a much sneakier sibling of bug #1: rather than dropping a parameter past the 4th slot, it silently reroutes parameter reads to slots that hold completely unrelated values from the caller.

Reproduction

// Function A: card is the 1st parameter, expected at $04
fun push_back_a(card: u8) {
    deck_a[deck_a_front] = card   // reads from $06, not $04!
    deck_a_count += 1
}

// Function B: card is the 3rd parameter, expected at $06
fun draw_card_face(x: u8, y: u8, card: u8) {
    // ... uses card normally ...
}

The IR lowering assigns card a single shared VarId because its var_map is global across all functions. The codegen then walks each function in turn, inserting (VarId(card), $0X) mappings into a single global var_addrs HashMap — and whichever function comes last in iteration order wins the mapping. If draw_card_face is processed after push_back_a, VarId(card) ends up mapped to $06, and push_back_a then reads its card parameter from $06 (which holds whatever the caller was using as a third argument — typically junk).

Root cause

src/ir/lowering.rs::get_or_create_var looks up names in self.var_map, which is shared across the whole program:

fn get_or_create_var(&mut self, name: &str) -> VarId {
    if let Some(&id) = self.var_map.get(name) {
        id
    } else {
        let id = VarId(self.next_var_id);
        self.next_var_id += 1;
        self.var_map.insert(name.to_string(), id);
        id
    }
}

lower_function calls get_or_create_var(&param.name) for each parameter, so two different functions both with a card parameter resolve to the same VarId. Once that single VarId flows into the codegen, the per-function "this is param index N of function F" relationship is lost — there's only one global mapping per VarId.

Workaround used in examples/war/

Every parameter name in the war source is unique across the entire program. Function-locals were already prefixed by function (see bug #3); we extended the same scheme to params: push_back_a(pba_arg_card: u8) instead of push_back_a(card: u8), etc. The wrapping pba_card / pbb_card / dcf_card snapshots from bug #2 stay because they also help with the bug-2 clobbering.

Fix proposal

Two layers to fix in:

  1. IR lowering: give every function its own var_map for parameters and locals. The global var_map should only hold top-level var / const / enum symbols.

  2. Codegen: even after the IR fix, the global var_addrs HashMap should grow a per-function dimension (one map per IrFunction) so two different functions can independently assign their own VarIds to overlapping zero-page slots.

Either fix alone is probably enough; both together is robust.


2. Function parameters share zero-page slots with nested calls — values clobbered across JSR

Symptom

A function that takes parameters and then calls another function sees its own parameters silently replaced by the inner call's arguments. Any code path that reads the original parameter after the inner call gets the wrong value.

Reproduction

fun draw_card_face(x: u8, y: u8, card: u8) {
    var rank: u8 = card_rank(card)   // x at $04 is now `card`
    var suit: u8 = card_suit(card)   // x at $04 is still `card`
    // x is supposed to be 120 here, but it's actually `card`
    var x1: u8 = x + 8               // computes card + 8, not 120 + 8
    draw Tileset at: (x, y) frame: ...   // draws at x = card, not 120
}

Concretely, calling draw_card_face(120, 128, 0x93) puts the card sprite at (0x93, 128) — completely wrong.

Root cause

Same allocator as bug #1: func.locals[0..param_count] are mapped to $04, $05, $06, $07. The caller writes its own arguments into the same zero-page slots before JSR, so the caller's parameters at those slots get clobbered by the callee's arguments. There is no save/restore wrapper around JSR and no spill/reload pass to refresh the caller's parameters from a backing copy.

Workaround used in examples/war/

Every helper that takes parameters AND makes any nested function call snapshots its parameters into fresh local variables at the top of the function, then references the locals exclusively throughout the body. See war/render.ne::draw_card_face, war/render.ne::draw_flying_card, war/deck.ne::push_back_a, war/deck.ne::push_back_b.

Fix proposal

  1. Spill on entry: at the top of every function body that makes a call, copy $04..$07 into per-function RAM slots and rewrite all parameter reads to load from the RAM copies. Equivalent to what users are doing manually today.

  2. Smarter scheduling: only spill a parameter slot if it's live across a call site (CFG-aware liveness pass on params). Same effect, less RAM cost for short helpers that never read their params after calling out.

Either fix would let users write straightforward function bodies without having to remember the snapshot dance.


3. Function-local variable names are in a flat global namespace

Symptom

Two different functions cannot declare locals with the same name. The compiler emits E0501 duplicate declaration of '<name>' even though the locals are in disjoint scopes.

Reproduction

fun foo() {
    var i: u8 = 0
    while i < 10 { i += 1 }
}

fun bar() {
    var i: u8 = 0   // E0501 duplicate declaration of 'i'
    while i < 5 { i += 1 }
}

Root cause

src/analyzer/mod.rs::register_var inserts every var declaration into a single self.symbols map keyed only on the variable's name, with no qualification by function or block:

fn register_var(&mut self, var: &VarDecl) {
    if self.symbols.contains_key(&var.name) {
        self.diagnostics.push(Diagnostic::error(
            ErrorCode::E0501,
            format!("duplicate declaration of '{}'", var.name),
            var.span,
        ));
        return;
    }
    ...
}

check_statement calls register_var for every Statement::VarDecl encountered while walking function bodies, so all locals across all functions and all nested blocks land in the same namespace.

Workaround used in examples/war/

Every function-local variable is prefixed with a short tag identifying its enclosing function (e.g. dfa_card in draw_front_a, pba_slot in push_back_a, dwp_px in draw_word_player). This makes long files harder to read but is fully mechanical.

Fix proposal

Rework register_var to maintain a stack of scopes (one per function body, one per nested block). Each Statement::VarDecl inserts into the current scope. Lookup walks the stack from innermost to outermost. The existing global symbol table is unchanged for top-level globals / consts / fun names; only function-locals shift to the scoped table.

A smaller intermediate fix: keep the flat table but qualify each local's stored name as <function>::<var> so the global table sees unique entries even when source names collide.


4. Per-frame sprite-per-scanline limit is invisible to user code

Symptom

Drawing more than 8 sprites whose Y rectangles intersect a single scanline causes the NES PPU to silently drop the excess sprites past the 8th in OAM order. There's no compile-time detection and no runtime warning — letters or tiles just don't render.

Reproduction

// 9 letters all on the same Y row:
draw_letter(0,   100, 0)
draw_letter(8,   100, 1)
draw_letter(16,  100, 2)
draw_letter(24,  100, 3)
draw_letter(32,  100, 4)
draw_letter(40,  100, 5)
draw_letter(48,  100, 6)
draw_letter(56,  100, 7)
draw_letter(64,  100, 8)   // this one will not render

Root cause

This is a real NES hardware constraint, not a compiler bug. However, because NEScript's draw allocator is purely sequential, the compiler cannot warn even when it has all the information needed to know the layout would overflow.

Workaround used in examples/war/

We staggered text rows. The title screen's "WAR / CARD GAME / 0 PLAYER / 1 PLAYER / 2 PLAYER" layout sits each row at a different y so no scanline carries more than 7 sprites; the victory screen's "PLAYER X / WINS" wraps after the player letter for the same reason.

Fix proposal

Two complementary improvements:

  1. Static analyzer pass: walk the IR for each frame handler, collect the set of (x, y) literal pairs feeding draw ops within the same basic block, and emit W01XX if any scanline (8-px row) would have > 8 sprites. Only catches the literal case but that's the most common.

  2. Sprite-cycling runtime helper: a cycle_sprites() intrinsic that rotates OAM order each frame so the same sprites get dropped on different frames, producing a flicker instead of a permanent dropout. Standard NES technique.


5. The inline keyword is a hint and is silently ignored for short functions

Symptom

Marking a tiny function inline fun does not always inline it. The compiler still emits a real JSR with full parameter passing through $04-$07, which means the inlining doesn't escape the bug-2 parameter clobbering.

Reproduction

inline fun card_rank(card: u8) -> u8 {
    return card >> 4
}

The asm dump shows JSR __ir_fn_card_rank at every call site — the function was not inlined.

Root cause

(Inferred — would need to confirm by reading the inliner pass.) The optimizer's inlining pass has a size threshold or a heuristic that prevents inlining in some contexts even when the function is marked inline. There's no diagnostic emitted when the hint is declined.

Workaround used in examples/war/

None — we just live with the JSR overhead and the bug-2 fallout.

Fix proposal

  1. Promote inline to a hard contract: when inline is present, always inline (or emit W01XX if it cannot be inlined for a structural reason like recursion).

  2. Optional dump: add --dump-inliner to print which inline fun declarations were inlined and which weren't, with the reason.


6. wide_hi IR-lowering map leaked between functions and corrupted 16-bit ops (FIXED)

Symptom

A function whose body had no 16-bit values whatsoever would nonetheless emit CmpEq16 (and other Op16 variants) where the destination temp aliased one of the source temps. The resulting comparison effectively became "is this byte equal to some uninitialised stack memory?", which in War caused the phase-machine match phase { ... } dispatcher to skip the P_WIN_B arm forever once the game first reached it — the game would freeze with both cards face-up and "PLAYER B WINS" never firing.

Reproduction (pre-fix)

A handful of u16 += 1 operations early in a state handler followed by a long match chain on a u8 was enough to trip it. The minimum repro is roughly:

var clock: u16 = 0
var phase: u8 = 0
on frame {
    clock += 1                    // wide op leaves wide_hi entries
    match phase {                 // u8 match — should be 8-bit
        0 => { phase = 1 }
        1 => { phase = 2 }
        2 => { phase = 3 }
        3 => { phase = 4 }
        4 => { phase = 5 }
        5 => { phase = 6 }
        6 => { phase = 7 }
        7 => { /* corrupt — never matched */ }
        _ => {}
    }
}

The IR for the phase == 7 arm came out as CmpEq16 { dest: T147, a_lo: T145, a_hi: T148, b_lo: T146, b_hi: T147 } — note dest == b_hi. The codegen happily emits the corresponding 16-bit asm, but reads garbage for the b_hi operand because it points at the same scratch slot the result will be written to.

Root cause

src/ir/lowering.rs::IrLowerer carries a wide_hi: HashMap<IrTemp, IrTemp> that records "this low temp's high byte lives at this other temp" pairs whenever a 16-bit value is produced. lower_function and lower_handler both reset next_temp = 0 at the start of each function — but they did not clear wide_hi. Stale entries from earlier functions stuck around and matched against fresh temp IDs in subsequent functions (which start counting from 0 again), causing is_wide(t) and widen(t) to return spurious "wide" results for what should have been narrow u8 values.

When that happens inside lower_binop's Eq path, widen(r) returns the stale (r, hi_r) pair where hi_r happens to be the next temp ID fresh_temp() will hand out a moment later — so the dest temp and b_hi end up identical.

Fix

src/ir/lowering.rs: in both lower_function and lower_handler, add self.wide_hi.clear(); immediately after self.next_temp = 0;. Done in this PR.

Why this didn't show up sooner

Every prior example either declared no u16 globals at all, or declared one and used it sparingly enough that the temp IDs the leaked entries claimed never collided with the rest of the function. War is the first example that combines a u16 free-running counter with a deep state machine that does many u8 comparisons in the same on frame body, which is exactly the shape the bug needs to manifest.

Regression test

src/ir/tests.rs::wide_hi_does_not_leak_between_functions (added in this PR) compiles a two-function program where function A uses a u16 += 1 (creating wide entries) and function B does u8 == const comparisons in a match. Pre-fix, the IR would emit CmpEq16 with aliased dest/source; post-fix it emits the expected 8-bit CmpEq.


Verification path after fixes

Once any of the bugs above are fixed in the compiler, the corresponding workarounds in examples/war/*.ne should be reverted in the same PR so:

  • The example demonstrates idiomatic code, not workaround code.
  • The PR's diff visibly proves the fix works end-to-end (the workaround removal would otherwise be a silent regression).
  • The committed examples/war.nes rebuilds byte-identically to the reverted source, which the pre-commit hook enforces.

The relevant workaround sites are catalogued in each bug's "Workaround used" section above; grep for the prefix tags (dcf_, dfa_, pba_, dwp_, …) to find them all.