bug: i386/arm32 diverge on Int64 via fn-return / record-field / mixed ops

• Type: bug (Track A — 32-bit Int64 codegen)
• Status: done (2026-06-24)
• Track: A
• Opened: 2026-06-24 (exposed after the frozen-string-deref fix unmasked it)
• Severity: medium — blocks make test-i386 / make test-arm32 (the first wall both gates hit now is test_conformance_2). aarch64 is green.

Root cause / fix

Not a shl/div/mod/fn-return/record bug at all — the common thread was the explicit Int64(x) numeric reinterpret of a 32-bit operand. In ir.inc (AN_PTR_CAST, ASTIVal = -1) a widening numeric cast only re-tagged the inner node's IRTk to the 64-bit width. On 32-bit targets that tells consumers "this is 64-bit" but no sign/zero-extend ever runs — EmitNode64 widens only when the SOURCE tag is narrow, and the re-tag erased the narrowness — so the high word (edx / r1) held garbage. x86-64/aarch64 hid it because a 32-bit load is already full-width in one register.

Fix: when the cast widens 32->64 (castTk is Int64/UInt64 and the inner node is not already 64-bit), emit a real widen by adding a 64-bit 0 — the inner node keeps its narrow tag so the extend direction follows its signedness, and the sum carries the cast's 64-bit tag. No-op on 64-bit targets. The narrowing/same-width re-tag (the bug-shl-signed-integer-width path) is unchanged. Required a 1-gen reseed (make bootstrap); fixedpoint holds. All gates green: make test + test-i386 + test-arm32 + test-aarch64 + cross-bootstrap + float-determinism + esp-bare + esp-softfloat.

Symptom

test_conformance_2 run on i386/arm32 diverges from the x86-64 oracle:

i386 : q=7000000005 mix=222669149779   fact20=5729255460435132416  rec ... sum=-2427387009277703680
x86  : q=7000000005 mix=111000000083   fact20=2432902008176640000  rec ... sum=1000000000

Diverging constructs (all Int64 on a 32-bit target):

• Fact(20) — Fact := Int64(n) * Fact(n-1) (recursive Int64 fn-result operand).
• I64Mix(q,3) — (a shl 4) + Int64(b) - (a div 7) + (a mod 5) (mixed Int64 shl/div/mod).
• RecSum(r) — r.A + Int64(r.B) (Int64 record field add).

Not this

Plain Int64 multiply / add of locals is correct on i386/arm32 (verified: a*b=1e12, 100000*1000000=1e11, 20! literal all match). So it is specific to Int64 reached through a function return, a record field, or the shl/div/mod mix — likely the edx:eax / r0:r1 pairing of those operand sources.

Acceptance

make test-i386 and make test-arm32 match the x86-64 oracle on test_conformance_2; existing tests stay green.

Repro

./compiler/pascal26 --target=i386 test/test_conformance_2.pas /tmp/c2 && tools/run_target.sh i386 /tmp/c2 vs the x86-64 build. Narrow with Fact/I64Mix/RecSum in isolation.