Full Int64 codegen for i386

• Type: feature
• Status: done
• Owner: claude
• Unblocks: feature-cross-selfhost-i386
• Opened: 2026-06-13
• Resolved: 2026-06-13

Problem

The i386 backend still has a low-dword scalar model for many 8-byte values. That has now shown up in multiple unrelated-looking places during the cross-self-host burn-down:

• shr 32 on small Int64 values duplicated or lost the low dword.
• ELF/header/data/string-table 64-bit serialization needed local small-value guards.
• Float literal and float-writer constants lose their high dword under an i386-hosted compiler.
• Large integer constants, shifts, boolean comparisons over widened values, pointer-sized serialization, and memory writes are all at risk because the backend does not consistently keep (lo, hi) pairs for Int64/UInt64.

The latest i386 fixed-point probe terminates, but differs at byte 24194 because the i386-hosted compiler emits zero for the double bits of 1000000000000000.0 (0x430c6bf526340000) in generated x86-64 float-writer code.

Scope

• Define an explicit i386 value model for Int64/UInt64, likely EDX:EAX for expression results plus paired stack/local stores.
• Implement 64-bit load/store, constants, shl/shr/sar, arithmetic needed by the compiler, comparisons, and argument/return handling.
• Replace local serialization workarounds where the general model makes them unnecessary.
• Add focused i386 oracle tests for:
  • Int64 shifts by 0, 1, 31, 32, 40, 63.
  • high-dword constants and memory stores/loads.
  • comparisons and boolean results involving high dwords.
  • float literal bit patterns emitted by an i386-hosted compiler.

Acceptance

• make test-i386 includes focused 64-bit oracle coverage and passes.
• i386-hosted test/test_cross_float.pas --target=i386 matches native output.
• feature-cross-selfhost-i386 fixed-point probe reaches byte-identical compiler output without ad hoc shr 32 workarounds.

Log

• 2026-06-13 — opened from the i386 self-host burn-down. The repeated failures are no longer isolated bugs; they point to missing first-class 64-bit scalar support in the i386 backend.
• 2026-06-13 — DONE (commits 7756241 feat, 68bef67 fix). Replaced the low-dword model with an edx:eax value model: 64-bit const/load/store (sym + mem), add/sub/mul, and/or, full shl/shr (0..63), signed+unsigned div/mod (inline restoring long division, no RTL dep), neg/not, all six compares (signed and unsigned), 64-bit writeln, Int64 result return + param sign-extend home, syscall-result extension (errno sign- / address zero-extend), Trunc/Round sign-extend. Integer literals outside signed-32 range now type as Int64. New oracle test/test_i386_int64.pas in make test-i386; test / test-i386 / test-aarch64 / test-arm32 all green. Bugs fixed while landing: 64-bit eq/neq flag-clobber (add esp between or and setcc), signed divmod off-by-4 dividend abs, and (separate commit) the builtinheap PWord = ^Int64 machine-word landmine that on i386 wrote 8 bytes into a 4-byte handle slot. Acceptance #1 (oracle in make test-i386) and #2 (cross_float matches) met; #3 (byte-identical self-host fixed-point) belongs to feature-cross-selfhost -i386 and is unblocked but not yet reached — see that ticket for the next wall (lexer/unit-dispatch token miscompile in the i386-hosted compiler).