A Boolean Variant writes as 1/0 on every target except x86-64

Found 2026-08-24 while gating [[bug-a-not-on-an-integer-variant-answers-a-boolean]] — it is what was left in the diff once that fix landed, and it is not part of it.

var a: Variant;
begin
  a := True;  writeln('T: ', a);
  a := False; writeln('F: ', a);
end.

Reproduces with the PINNED binary, so it predates today's variant work.

The tag is right on every target — VarType(c) answers the Boolean code everywhere. Only the rendering differs, which puts the defect in whatever writes a variant, not in what produced it.

Where to look

PXXWriteVariant (compiler/builtin/builtinheap.pas) is the runtime renderer the cross targets use; x86-64 has its own inline path. That is the same split that produced the operator-table bug filed alongside this one — one concept, two implementations, and the one nobody runs locally is the one that drifted. So check whether the runtime one has a VT_BOOL arm at all before assuming a subtle difference.

Worth a grep for the sibling while there: any other tag the two renderers disagree about (VT_CHAR, VT_EMPTY, VT_DOUBLE formatting) is the same bug and should be fixed in the same pass rather than filed three more times.

Gate

Track A's, plus the four rows above matching fpc on x86-64, i386, arm32 and aarch64. test/test_variant_bitwise_and_not.pas deliberately reads tags and values instead of rendering Booleans, because of this; once this is fixed that test can assert the rendered form directly.

FIXED 2026-08-24 (claude-A)

PXXWriteVariant (compiler/builtin/builtinheap.pas) folded VT_BOOL in with the integer tags:

if (tag = 1) or (tag = 2) or (tag = 4) then  { VT_INT / VT_INT64 / VT_BOOL }
begin
  iv := PWord(Int64(v) + 8)^;
  write(iv);
end

and its own header claimed it "mirrors EmitWriteVariant (x86-64)", which has a True/False arm. The comment described the intent; the code did something else, and nobody diffed them because x86-64 was right. Split out as its own arm — the Python spelling is the same word, so one arm serves both frontends.

What the sibling grep found, which is the part worth recording

The ticket said to grep for the other tags the two renderers might disagree about rather than fix one and file three more. Measured across bool, integer, negative integer, double, char and string on x86-64 / i386 / aarch64 / arm32 against fpc 3.2.2 -Mobjfpc -O1: Boolean was the only one. After this fix all four targets produce output identical to FPC's, line for line.

Two tags are deliberately still unmirrored, and that is not drift:

• VT_EMPTY — x86-64 writes None, the runtime writes nothing. FPC writes nothing for a cleared Variant and raises for Null, so x86-64's spelling is the contested one. Copying it here would have spread a rendering that is probably wrong to three more targets. That is [[bug-a-a-null-variant-renders-as-none-in-pascal]], and it now settles in one place for every target when it is answered.
• VT_OBJECT — x86-64 writes <object> as a debugging placeholder; FPC raises. Same reasoning.

Both are stated in PXXWriteVariant's header so the next reader does not "finish the mirror" without noticing they are open questions.

Found in passing, not folded in

VarClear is undefined — VarClear(v) is error: undefined variable, though FPC accepts it and prints an empty rendering afterwards. Filed as [[bug-a-varclear-is-undefined]]; it is a missing library entry point, not a renderer defect, and it is why the empty-slot row could not be measured from Pascal at all.

Verified

test/test_variant_writes_every_tag.pas, wired into test-core. It asserts the whole OUTPUT STREAM rather than printing ALL OK, because the rendering is the thing under test: True|False|42|-7|1.25|q|hey|. Identical under fpc 3.2.2 and under pxx on x86-64, i386, aarch64 and arm32. NilPy re-measured against CPython (print(True), print(False), a Boolean expression, a comparison result): unchanged, all four rows.

Self-host fixedpoint converged in one round; tools/gate.sh quick GREEN.

Gate

make compiler/pascal26 converged + the seven-row rendering diff on four targets against FPC + the NilPy re-measure + tools/gate.sh quick GREEN.

Log

• 2026-08-24 — resolved, commit 3bf667e92.