MOVD msg_base+8(FP), R4
	MOVD msg_len+16(FP), R5

	MOVD 0(R3), R8   // h0
	MOVD 8(R3), R9   // h1
	MOVD 16(R3), R10 // h2
	MOVD 24(R3), R11 // r0
	MOVD 32(R3), R12 // r1

	CMP R5, $16
	BLT bytes_between_0_and_15

loop:
	POLY1305_ADD(R4, R8, R9, R10, R20, R21, R22)

	PCALIGN $16
multiply:
	POLY1305_MUL(R8, R9, R10, R11, R12, R16, R17, R18, R14, R20, R21)
	ADD $-16, R5
	CMP R5, $16
	BGE loop

			"\tMOVBLZX	(BP)(DX*4),	R8 \\",
			"\tload((off+1), r2) \\",
			"\tMOVB	R8,		(off*4)(R12) \\",
			"\tPINSRW	$index, (BP)(R8*4), xmm",
			"#define LOAD(off, reg) \\",
			"\tMOVBLZX	(off*4)(R12),	reg \\",
			"\tADDB	reg,		DX",
			"KEYROUND(X0, LOAD, 8, AX, BX, 0)",
		),

TEXT ·Syscall6(SB),NOSPLIT,$0-80
	JMP	syscall·Syscall6(SB)

TEXT ·SyscallNoError(SB),NOSPLIT,$0-48
	JAL	runtime·entersyscall(SB)
	MOVV	a1+8(FP), R4
	MOVV	a2+16(FP), R5
	MOVV	a3+24(FP), R6
	MOVV	R0, R7
	MOVV	R0, R8
	MOVV	R0, R9
	MOVV	trap+0(FP), R2	// syscall entry
	SYSCALL
	MOVV	R2, r1+32(FP)
	MOVV	R3, r2+40(FP)
	JAL	runtime·exitsyscall(SB)
	RET

TEXT ·RawSyscall(SB),NOSPLIT,$0-56
	JMP	syscall·RawSyscall(SB)

	CASALW	R5, (RSP), R7                        // e7ffe588
	CASALH	ZR, (R5), R8                         // a8fcff48
	CASALB	R8, (R9), ZR                         // 3ffde808
	CASPD	(R30, ZR), (RSP), (R8, R9)           // e87f3e48
	CASPW	(R6, R7), (R8), (R4, R5)             // 047d2608
	CASPD	(R2, R3), (R2), (R8, R9)             // 487c2248

// RET
	RET
	RET	foo(SB)

TEXT ·Syscall6(SB),NOSPLIT,$0-80
	JMP	syscall·Syscall6(SB)

TEXT ·SyscallNoError(SB),NOSPLIT,$0-48
	CALL	runtime·entersyscall(SB)
	MOVQ	a1+8(FP), DI
	MOVQ	a2+16(FP), SI
	MOVQ	a3+24(FP), DX
	MOVQ	$0, R10
	MOVQ	$0, R8
	MOVQ	$0, R9
	MOVQ	trap+0(FP), AX	// syscall entry
	SYSCALL
	MOVQ	AX, r1+32(FP)
	MOVQ	DX, r2+40(FP)
	CALL	runtime·exitsyscall(SB)
	RET

TEXT ·RawSyscall(SB),NOSPLIT,$0-56
	JMP	syscall·RawSyscall(SB)

	B	syscall·Syscall6(SB)

TEXT ·SyscallNoError(SB),NOSPLIT,$0-48
	BL	runtime·entersyscall(SB)
	MOVD	a1+8(FP), R0
	MOVD	a2+16(FP), R1
	MOVD	a3+24(FP), R2
	MOVD	$0, R3
	MOVD	$0, R4
	MOVD	$0, R5
	MOVD	trap+0(FP), R8	// syscall entry
	SVC
	MOVD	R0, r1+32(FP)	// r1
	MOVD	R1, r2+40(FP)	// r2
	BL	runtime·exitsyscall(SB)
	RET

TEXT ·RawSyscall(SB),NOSPLIT,$0-56
	B	syscall·RawSyscall(SB)

TEXT ·RawSyscall6(SB),NOSPLIT,$0-80

TEXT ·Syscall6(SB),NOSPLIT,$0-80
	JMP	syscall·Syscall6(SB)

TEXT ·SyscallNoError(SB),NOSPLIT,$0-48
	JAL	runtime·entersyscall(SB)
	MOVV	a1+8(FP), R4
	MOVV	a2+16(FP), R5
	MOVV	a3+24(FP), R6
	MOVV	R0, R7
	MOVV	R0, R8
	MOVV	R0, R9
	MOVV	trap+0(FP), R11	// syscall entry
	SYSCALL
	MOVV	R4, r1+32(FP)
	MOVV	R0, r2+40(FP)	// r2 is not used. Always set to 0
	JAL	runtime·exitsyscall(SB)
	RET

TEXT ·RawSyscall(SB),NOSPLIT,$0-56

	MOVQ	m_curg(AX), CX
	TESTQ	CX, CX
	JZ	sigtramp        // g.m.curg == nil
	MOVQ	g_syscallsp(CX), CX
	TESTQ	CX, CX
	JZ	sigtramp        // g.m.curg.syscallsp == 0
	MOVQ	m_cgoCallers(AX), R8
	TESTQ	R8, R8
	JZ	sigtramp        // g.m.cgoCallers == nil
	MOVL	m_cgoCallersUse(AX), CX
	TESTL	CX, CX
	JNZ	sigtramp	// g.m.cgoCallersUse != 0

	// Jump to a function in runtime/cgo.

	MOVL	b, R9; \
	XORL	c, R9; \
	XORL	d, R9

#define FUNC3(a, b, c, d, e) \
	MOVL	b, R8; \
	ORL	c, R8; \
	ANDL	d, R8; \
	MOVL	b, R9; \
	ANDL	c, R9; \
	ORL	R8, R9

#define FUNC4 FUNC2

#define MIX(a, b, c, d, e, const) \
	ROLL	$30, b; \
	ADDL	R9, e; \
	MOVL	a, R8; \
	ROLL	$5, R8; \
	LEAL	const(e)(R10*1), e; \
	ADDL	R8, e

#define ROUND1(a, b, c, d, e, index) \
	LOAD(index); \
	FUNC1(a, b, c, d, e); \

//

// func rawVforkSyscall(trap, a1, a2, a3 uintptr) (r1, err uintptr)
TEXT ·rawVforkSyscall(SB),NOSPLIT,$0-48
	MOVV	a1+8(FP), R4
	MOVV	a2+16(FP), R5
	MOVV	a3+24(FP), R6
	MOVV	$0, R7
	MOVV	$0, R8
	MOVV	$0, R9
	MOVV	trap+0(FP), R11	// syscall entry
	SYSCALL
	MOVW	$-4096, R12
	BGEU	R12, R4, ok
	MOVV	$-1, R12
	MOVV	R12, r1+32(FP)	// r1
	SUBVU	R4, R0, R4
	MOVV	R4, err+40(FP)	// errno
	RET
ok: