CLD
	REP; MOVSL

docall:
	// Call stdcall or cdecl function.
	// DI SI BP BX are preserved, SP is not
	CALL	libcall_fn(BX)
	MOVL	BP, SP

	// Return result.
	MOVL	fn+0(FP), BX
	MOVL	AX, libcall_r1(BX)
	MOVL	DX, libcall_r2(BX)

	// GetLastError().
	MOVL	0x34(FS), AX
	MOVL	AX, libcall_err(BX)

	RET

// faster get/set last error
TEXT runtime·getlasterror(SB),NOSPLIT,$0
	MOVL	0x34(FS), AX

_3args:
	MOVD	(2*8)(R12), R2
_2args:
	MOVD	(1*8)(R12), R1
_1args:
	MOVD	(0*8)(R12), R0
_0args:

	MOVD	libcall_fn(R19), R12	// branch to libcall->fn
	BL	(R12)

	MOVD	R20, RSP			// free stack space
	MOVD	R0, libcall_r1(R19)		// save return value to libcall->r1
	// TODO(rsc) floating point like amd64 in libcall->r2?

	// GetLastError
	MOVD	TEB_error(R18_PLATFORM), R0
	MOVD	R0, libcall_err(R19)

TEXT runtime·asmstdcall(SB),NOSPLIT,$16
	MOVQ	SP, AX
	ANDQ	$~15, SP	// alignment as per Windows requirement
	MOVQ	AX, 8(SP)
	MOVQ	CX, 0(SP)	// asmcgocall will put first argument into CX.

	MOVQ	libcall_fn(CX), AX
	MOVQ	libcall_args(CX), SI
	MOVQ	libcall_n(CX), CX

	// SetLastError(0).
	MOVQ	0x30(GS), DI
	MOVL	$0, 0x68(DI)

	SUBQ	$(const_maxArgs*8), SP	// room for args

	// Fast version, do not store args on the stack.

	MOVM.DB.W [R4, R5, R14], (R13)	// push {r4, r5, lr}
	MOVW	R0, R4			// put libcall * in r4
	MOVW	R13, R5			// save stack pointer in r5

	// SetLastError(0)
	MOVW	$0, R0
	MRC	15, 0, R1, C13, C0, 2
	MOVW	R0, 0x34(R1)

	MOVW	8(R4), R12	// libcall->args

	// Do we have more than 4 arguments?
	MOVW	4(R4), R0	// libcall->n
	SUB.S	$4, R0, R2
	BLE	loadregs

	// Reserve stack space for remaining args

//
//go:nosplit
//go:cgo_unsafe_args
//go:linkname syscall_syscall6
func syscall_syscall6(fn, nargs, a1, a2, a3, a4, a5, a6 uintptr) (r1, r2, err uintptr) {
	c := libcall{
		fn:   fn,
		n:    nargs,
		args: uintptr(unsafe.Pointer(&a1)),
	}

	entersyscallblock()
	asmcgocall(unsafe.Pointer(&asmsyscall6), unsafe.Pointer(&c))
	exitsyscall()
	return c.r1, 0, c.err
}