double(q) / implicit qword→double converts as signed for values ≥ 2^63

• Type: bug (silent wrong value). Track A (IR float-conversion lowering, all backends).
• Status: done
• Opened: 2026-07-14
• Found by: tint642.pp burn-down (testqwordtypecast, do_error(20)) while resolving [[bug-pascal-record-cast-field-offset]].

Repro

var q: qword; d: double;
begin
  q := qword(1) shl 63;      { $8000000000000000 }
  d := q;
  writeln(d);                { FPC: 9.223372036854775808E+18 }
                             { pxx: -9.2233720368547758E+18 (signed cvt) }
  if q <> 2.0*double($80000000)*double($80000000) then writeln('BAD');
end.

cvtsi2sd (and its cross equivalents) is a SIGNED conversion; an unsigned 64-bit source with the top bit set needs the unsigned fixup (test top bit → halve, convert, double — the standard u64→f64 sequence) or a range-based branch. Mixed qword/double COMPARISONS convert the qword operand the same way, so they inherit the bug.

Impact

• tint642.pp (FPC testsuite) fails at error 20 — this is now the ONLY residual blocking it (record-cast field offsets and bitwise not (q or q) landed tonight; skip-list entry points here).
• Any program mixing qwords ≥ 2^63 with floating point gets silently wrong values on every backend.

Acceptance

• Repro matches FPC on x86-64 + cross targets.
• tint642.pp passes and drops from test/pascal-conformance/pxx.skip (check the rest of the test runs green past error 20 — it is a torture test, more walls may hide behind this one).

Log

• 2026-07-14 — resolved, commit 30eb98f0.