track: A prio: 55 type: bug blocked-by: [] resolved: 20e1ca641 summary: "FIXED for the four PXX-INTERNAL call kinds -- direct, constructor, indirect, virtual -- which now share one EmitCallArgRegsA64 and three Errors are deleted. NilPy builds and runs on aarch64: print(1+1) prints 2, and a class-heavy program matches CPython. THE TICKET'S PREMISE IS CORRECTED: it is TWO mechanisms, not one refused six times. The remaining three sites -- external, variadic external, cdecl indirect -- are the AAPCS64 C convention with INDEPENDENT int/fp register banks and a per-parameter classification, which the internal convention (every argument 8 bytes in an x-register, by position) does not have. Re-filed as bug-a-aarch64-has-no-stack-argument-passing-for-the-three-c-abi-call-kinds; nothing currently reaches it.

Fixed for the internal family — 2026-08-31 by frankA

EmitCallArgRegsA64(nArgs, nArgTotal, xtra1, xtra2): args 0..7 into x0..x7, args 8.. into an AAPCS64 outgoing block, extras (an indirect callee in x16, a hidden indirect-result pointer in x8) loaded from the temps pushed below arg0. Returns the bytes to drop after the call, 0 on the <=8 path which pops. Four call sites, one mechanism; the direct call was moved onto it too, so the arm that already worked cannot drift from the three that now do.

The premise was half right, and the correction is the useful part

The ticket said one mechanism, refused six times. Measured, it is two:

• the pxx internal convention — every argument is 8 bytes in an x-register, assigned by position. Four call kinds, and they really were one rule spelt four times. That is what this fix deletes.
• the AAPCS64 C convention — integer/pointer args in x0..x7 and FP args in v0..v7, allocated from two INDEPENDENT banks with a per-parameter classification (and a by-ref float counting as a pointer, which the x86-64 twin got wrong once already). Three call kinds. Sharing the internal helper there would be wrong, not merely incomplete.

Parking rather than microfixing the second half is deliberate (root-cause-over-microfix.md): nothing reaches those three today, the rule is different, and a C-interop marshalling bug is a segfault at a library boundary rather than a wrong number.

What caught the one real bug in it

test/test_a64_stack_args.pas — one row per call kind, because test_cross_many_params exercises only the DIRECT call and is therefore exactly the test that stayed green while five kinds refused. Every routine's answer depends on arguments 8 and 9, the two that travel on the stack.

The ctor row printed 1508537076 for 10936 while the other three were already right. The constructor is the only kind that does something between marshalling and the call — it saved Self with str x0, [sp,#-16]!, which on the stack-arg path moves sp 16 bytes BELOW the outgoing block, so the callee read Self as argument 8. Self is arg0 and its temp is still live, so it reloads from [sp, #sz-16] instead. A three-kind test would not have found it.

Gate

• 28 aarch64 binaries (14 programs x -O0/-O3) byte-identical to a control built with the change reverted — the <=8-argument path, which is every program that compiled before.
• The new test fails on the pre-fix compiler with the indirect-call refusal.
• make compiler/pascal26 converged; tools/gate.sh quick GREEN.
• NilPy: print(1+1) -> 2; a class + list + for-loop program prints 25 6 ok on aarch64, x86-64 and CPython alike.