Full parameter/result ABI on cross targets

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

Motivation

The cross param-copy and call paths only handle ordinal + pointer-sized arguments and results. compiler.pas uses richer signatures — the first wall when cross-compiling it is target i386: only ordinal/pointer parameters supported yet. Needed for the cross self-host.

Scope

• Record / aggregate by value — pass and return records larger than a register (hidden-pointer / by-reference-copy ABI, mirroring the x86-64 TypeIsAggregate + ProcAggregateDestSym path) on i386, ARM32, AArch64.
• Float parameters/results — per-target FP arg registers (ties to feature-cross-float-variant): AArch64 v0..v7, ARM32 VFP/soft-float, i386 stack/x87, and the SSE class on the SysV side already done for x86-64.
• >N register args — stack-arg spill on AArch64 (>8) and ARM32 (>4) for internal calls (currently a hard error).
• Open-array / array of T params — the (ptr,len) pair convention on cross targets.
• Audit EmitProcPrologue param-copy (parser.inc) and the internal-call arg marshalling in each ir_codegen_* for these cases.

Acceptance

Programs exercising record-by-value, float, open-array, and >N-arg calls compile and run on i386, ARM32, AArch64 identically to x86-64. New test/test_cross_params.pas in the suites.

Log

• 2026-06-13 — arm32 string-result ABI slice landed. The epilogue result guard (symtab.inc) rejected a tyString result (legacy inline [len][data] struct, whose Result slot is skGlobal) — the first arm32 wall when cross-compiling compiler.pas (StrInt). Fixed: tyString results now return the struct's address via EmitLoadVarAddrArm32 (BSS for the global slot), mirroring the x86-64 path. That exposed a second, entangled gap — legacy tyString variable assignment on arm32 stored a bare pointer instead of copying the struct, so strings read back empty. Fixed the arm32 IR_STORE_SYM tyString path to copy [len:8][data] with EmitArm32CopyBytes (mirrors the x86-64 len-field + rep-movsb store). Legacy strings now round-trip on arm32. New oracle test test/test_cross_strresult.pas wired into make test-arm32; arm32 + core + self-host/threadsafe fixedpoints green. Frozen compiler.pas cross-compile advanced from the StrInt result wall (parser line 88) to the shared SetLength wall (line 1201), reaching parity with the managed path — next wall (SetLength expects an array variable) is feature-cross-codegen-gaps territory, not this ticket.
• 2026-06-13 — claimed. i386 by-ref / Variant params slice landed: the i386 EmitProcPrologue param-copy guard (parser.inc) rejected any non-ordinal / non-pointer-sized param. A by-ref param (var/const aggregate, incl. const v: Variant) is a pointer-sized handle whatever its declared type — the caller already pushes the argument's address and the 4-byte copy moves that pointer into the slot. Guard now admits IsRef and tyVariant. This was the first wall when cross-compiling compiler.pas to i386 (VariantToStr(const v: Variant)). New oracle test test/test_cross_byref_params.pas wired into make test-i386; core + i386 + self-host/threadsafe fixedpoints green. Next i386 wall is an aggregate local variable (only ordinal/pointer/string variables supported yet) — that is feature-cross-codegen-gaps territory, not this ticket. Record-by-value (true

  register payload), float params/results, and >N-arg spill on AArch64/ARM32 still pending here.

• 2026-06-14 — i386 Int64 by-value param passing landed (commit e3b8866), completing the i386 side of this ticket: Int64/UInt64 args now pass as a full 8-byte slot (caller pushes both dwords; prologue copies both; displacement counts them as 8). Runtime helper size/len params moved to NativeInt so the hand-emitted 4-byte pushes stay valid. This was the last param-ABI gap for the i386 self-host, which is now byte-identical (feature-cross-selfhost-i386 done). Record-by-value, float params/results, and >N-arg spill on AArch64/ARM32 still pending here.

Closure (2026-06-16)

feature-cross-bootstrap-selfhost is DONE — byte-identical self-fixedpoint on i386/aarch64/arm32 (and x86-64). This ticket existed to unblock that gate, so its blocking purpose is met: every code path compiler.pas itself exercises now works byte-identically on all cross targets. Residual gaps are only in language features the compiler does NOT self-use (e.g. classes, interfaces, some param/ ABI shapes user code hits) — those move to the language-surface hardening effort driven by the synthetic conformance harness ([[feature-synthetic-feature-matrix-test]]). Closed.