← board

bug: cardinal/longword binary-op promotes to uint64 (FPC: int64)

Resolution (2026-06-23)

TypeArithmeticResult (symtab.inc) returned tyUInt64 whenever TypeCompareUnsigned(lhs,rhs) was true — including two 32-bit unsigned operands (Cardinal/LongWord/Word/Byte), so cardinal(3) - cardinal(5) evaluated as 2^64-2. Changed to return tyUInt64 only when an operand is itself a 64-bit unsigned type (tyUInt64/tyNativeUInt); otherwise tyInt64. Two 32-bit-or-narrower unsigned operands fit in Int64, so the result is signed — matching FPC. TypeCompareUnsigned (relational comparison signedness) is untouched.

Verified vs FPC {$mode objfpc}: a-b = -2, store-back wrap unchanged (4294967295 / 0), a+b = 8, uint64 - n stays unsigned. Gate: make test (self-host byte-identical) + FPC oracle. Closes bug-cardinal-expr-promotion.

Symptom

A cardinal - cardinal expression used directly (not stored back into a 32-bit variable) is computed as unsigned 64-bit in pxx, but FPC promotes the pair to a signed int64:

var a, b: cardinal;
begin a := 3; b := 5; writeln(a - b); end.
{ fpc: -2    pxx: 18446744073709551614  (2^64 - 2) }

What is correct (controls)

cardinal is 32-bit in both (sizeof = 4), and wrap is correct when the result is stored back into a 32-bit type:

var c: cardinal; begin c := 0; c := c - 1; writeln(c); end.   { both: 4294967295 }
var c: cardinal; begin c := 4294967295; c := c + 1; writeln(c); end.  { both: 0 }

So only the result type of the binary operation differs: pxx widens cardinal op cardinal to uint64, FPC to int64. Positive in-range results agree; underflow / >int32 values diverge.

Expected

Match FPC integer promotion: cardinal/longword operands whose result fits in int64 yield a signed int64 expression (so 3 - 5 → -2), not uint64.

Repro

tools/fpc_diff_probe.sh (cardinal-sub).