// 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
	MOVL	AX, ret+0(FP)
	RET

_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)

	// Restore callee-saved registers.
	LDP	16(RSP), (R19, R20)
	RET

TEXT runtime·getlasterror(SB),NOSPLIT,$0

	MOVQ	0(SP), CX
	MOVQ	8(SP), SP
	MOVQ	AX, libcall_r1(CX)
	// Floating point return values are returned in XMM0. Setting r2 to this
	// value in case this call returned a floating point value. For details,
	// see https://docs.microsoft.com/en-us/cpp/build/x64-calling-convention
	MOVQ    X0, libcall_r2(CX)

	// GetLastError().
	MOVQ	0x30(GS), DI
	MOVL	0x68(DI), AX
	MOVQ	AX, libcall_err(CX)

	RET

// faster get/set last error