M2compiler-V3/docs/summary_sigcheck.md

Step: strict implementation-vs-definition signature checks

Tag v3-sigcheck. Suite 129/129; fixpoint OK (bootstrap/fixpoint.sh, image 2,102,326 bytes).

What changed

An IMPLEMENTATION procedure heading must now match its DEFINITION (or exported FORWARD) beyond arity:

• result type, formal count, each formal type, and each formal's VAR-ness.

Mismatch → new semantic error 235 implementation does not match definition.

Before, only the arity was effectively covered and a differing result type or parameter type was silently accepted (a real separate-compilation hazard).

How it works

• ResumeProc (the implementation heading resuming a DEFINITION procedure) snapshots the definition signature into the symbol: snapRes (result), snapT[]/snapV[] (formal types / VAR-ness), nSnap, and sets fsnap.
• The implementation heading re-enters its own formals; as each formal's type becomes known (FixPending), the formal is compared against the snapshot. SetProcRes compares the result type (and is the backstop for a parameterless heading).
• Comparisons use FormalOk (value formals: same type, both integer-family, or string-compatible) and VarParamOk (VAR formals); a mismatched VAR-ness is caught directly. InvalidType is treated as "no evidence" so a single root error is not multiplied.
• Signatures are bounded to MaxSnap = 16 formals (self-hosting uses far fewer).

Deliberate leniency

• INTEGER/CARDINAL are accepted interchangeably for value formals and results: V3 models them as one integer family (a per- type split would reject the compiler's own valid-looking source). For VAR formals the types must match exactly.
• VAR open arrays vs fixed arrays still follow VarParamOk (a fixed array is an acceptable actual for an open formal).

Tests

tests/{sig.def,sig.mod,sig_prog.mod} + expect_fail_files sig: sig.mod declares Add(a,b):CHAR where the definition says :INTEGER, and Fill(buf:ARRAY OF CHAR;n:CARDINAL) where the definition has VAR buf. Both report 235. Exact-match sessions still compile (verified with the full m5 case).

Notes

• Error 231 is already "opaque type outside definition", so the signature mismatch got its own code, 235.
• qbe/.ssa output is suppressed when 235 fires (the session is rejected before codegen), matching other semantic errors.

Files

compiler/src/SymTab.def/.mod (ResumeProc snapshot, FixPending and SetProcRes returning the check result, FormalOk, MaxSnap), compiler/src/M2.atg (report 235 at SetProcRes/FixPending), compiler/src/compiler.frm (message 235), compiler/tests/{sig.def,sig.mod,sig_prog.mod}, compiler/run_tests.sh, docs/features.md.