The issue-0176 conformance gate was name-only, so an impl P for T with a mismatched return/param type (or arity) built a wrong-ABI thunk that silently miscompiled (exit 0, wrong value). firstUnimplementedMethod now validates arity (after self), each param type, and the return type against the protocol declaration, substituting protocol Self->concrete via resolveProtoTypeSubSelf (recurses through pointer/many-pointer/ optional/slice/array so []Self<->[]T match; conservative .unresolved for Self-in-generic-arg). Comparison is by structural formatTypeName (alias/module/spelling independent); typesClearlyDiffer skips when either side has an unresolved leaf at any depth, biasing against false-positives. Regressions: diagnostics/1201 (negative), protocols/0420 (positive, []Self param). Verified by 3+3 adversarial reviews (a mid-fix []Self false-positive was found and closed); suite 792/0.
3.4 KiB
0178 — protocol impl method with a mismatched return/param TYPE silently miscompiles
RESOLVED. The issue-0176 conformance gate was name-only, so an
impl P for Twith a mismatched return/param type (or arity) built a wrong-ABI thunk that silently miscompiled (exit 0, wrong value). Fix (src/ir/lower/protocol.zig):firstUnimplementedMethodnow validates each impl method's signature against the protocol declaration — arity (afterself), every param type, and the return type — substituting protocolSelf→concrete viaresolveProtoTypeSubSelf(recurses through pointer/many-pointer/optional/slice/array so[]Self↔[]Tetc. match; conservative.unresolvedforSelf-in-generic-arg). Comparison is by structuralformatTypeName(alias/module/spelling independent), andtypesClearlyDifferskips when either side has an unresolved leaf at any depth — biased against false-positives. Mismatch → located diagnostic. Verified by 3+3 adversarial reviews (a mid-fix[]Selffalse-positive was found and closed); suite 792/0. Regressions:examples/diagnostics/1201-diagnostics-protocol-impl-signature-mismatch.sx(negative),examples/protocols/0420-protocol-self-in-slice-param.sx(positive). Known gaps (pre-existing, loud not silent — out of scope): aSelf-through-generic-arg mismatch (Box(Self)) and by-value array protocol params ([2]Self) fail at LLVM verification, not silently.
Symptom
An impl P for T whose method has the right NAME but a mismatched return type or
parameter type is accepted (it satisfies the issue-0176 conformance gate, which
is name-based), and dispatch through the erased protocol silently produces the
WRONG result (exit 0). No diagnostic. (Arity mismatch and #builtin-body
mismatch fail loudly — exit 1 — and are not this bug; the TYPE-mismatch cases are
silent.)
Reproduction
#import "modules/std.sx";
P :: protocol { val :: (self: *Self) -> i64; }
T :: struct { n: i64 = 7; }
impl P for T { val :: (self: *T) -> bool { return true; } } // return type bool ≠ i64
main :: () {
t := T.{ n = 7 };
p : P = t;
print("{}\n", p.val()); // prints "1" (the bool), silently wrong — no diagnostic
}
A parameter-type mismatch (x: bool where the protocol declares x: i64)
similarly dispatches silently wrong.
Investigation prompt
The issue-0176 conformance gate (firstUnimplementedMethod in
src/ir/lower/protocol.zig) checks method PRESENCE (and rejects type_params > 0), but does NOT check that the impl method's SIGNATURE (parameter types,
arity, return type) matches the protocol method's declared signature. A
mismatched-type impl builds a thunk that calls the impl with the wrong ABI,
silently miscompiling. Add signature validation when registering / gating an
impl method against its protocol method: compare the impl method's params
(after the erased self) and return type against the protocol declaration, and
emit a located diagnostic on mismatch (arity, param type, or return type). The
protocol method declaration is in protocol_decl_map; the impl FnDecl is in
fn_ast_map. Decide whether this lives in the conformance gate or in
ProtocolResolver.registerImplBlock (src/ir/protocols.zig). Follow the
no-silent-fallback rule. Verify: the repro is now a clean diagnostic (exit 1);
a correctly-typed impl still works; add an examples/diagnostics/11xx-...
negative regression. (Found during adversarial review of issue 0176.)