Float + Variant codegen on cross targets

• Type: feature
• Status: done
• Owner: claude
• Unblocks: feature-cross-bootstrap-selfhost
• Opened: 2026-06-11 (user request)

Motivation

The cross backends are integer/pointer only. No floating-point (tySingle/ tyDouble/tyExtended) and no Variant (16-byte tag+payload) codegen. This blocks compiling the full builtin unit on cross targets — FloatToStr, VariantToStr, and Str/float paths use float arithmetic and variant ops — and therefore blocks the cross self-host (uses SysUtils pulls these in).

Scope

Floats

• IR_BINOP float arith + compare, float IR_WRITE (the EmitWriteFloat* family), float load/store, int↔float conversion, float literals.
• i386: x87 or SSE2 (SSE2 is the simpler, modern choice; 32-bit SSE2 is baseline on any real i686). ARM32: VFP. AArch64: native FP/SIMD (v0..v31).
• Float parameter ABI per target (ties to feature-cross-param-abi): x86-64 SSE class, AArch64 v-regs, ARM32 VFP/soft-float, i386 stack/x87.

Variants

• IR_VAR_STORE / IR_VAR_BINOP / IR_VAR_BOX / EmitVariantClear / EmitVariantRetain / EmitWriteVariant on cross targets. The 16-byte slot is the same layout everywhere; the managed payload (string) release already has portable helpers.
• Consider moving variant box/clear/compare bodies into portable Pascal helpers (as with the string/record runtime) to minimise per-arch emission.

Acceptance

Float and Variant test programs compile and run on i386, ARM32, AArch64 with output identical to x86-64. The full builtin unit compiles on every target. New test/test_cross_float.pas and test/test_cross_variant.pas in the suites.

Staging

Floats and Variants can land as two separate slices; floats first (broader value, and builtin's FloatToStr needs them).

Log

• 2026-06-11 — claimed; aarch64 float slice landed (per-arch emitters: EmitWriteFloatNat/Fixed/SciA64, variant emitters, PXXVarBinOp runtime, test_cross_float in test-aarch64). Commits 2a98d4d, cf05aff.
• 2026-06-12 — RESOLVED. i386 (SSE2, xmm0 channel) and ARM32 (VFP, d0 channel) float + Variant codegen landed; float→text and variant write/clear/retain moved to portable builtinheap helpers (PXXWriteFloat*/PXXVar*) — double-only arithmetic with the 2^52 round-even trick, so no per-arch formatting emitters and no 64-bit integer registers needed. aarch64 variant write path fixed (its hand-emitted tag compares were mis-encoded and untested) and switched to the same helper. test_cross_float + new test_cross_variant pass on i386/ARM32/AArch64 with output identical to x86-64; make bootstrap and make test green. Commits 683fbc5, fc299c3, f23f4da, 592132f, 1e3f992. Scope note: "full builtin unit compiles on every target" moved to feature-cross-codegen-gaps item 6 — it needs float params/results in the 32-bit call ABI (feature-cross-param-abi territory), not float codegen. Variant locals (16-byte frame zero-init) also noted there.