multiply:
	POLY1305_MUL(R8, R9, R10, R11, R12, BX, CX, R13, R14)
	SUBQ $16, R15
	CMPQ R15, $16
	JAE  loop

bytes_between_0_and_15:
	TESTQ R15, R15
	JZ    done
	MOVQ  $1, BX
	XORQ  CX, CX
	XORQ  R13, R13
	ADDQ  R15, SI

flush_buffer:
	SHLQ $8, BX, CX
	SHLQ $8, BX
	MOVB -1(SI), R13
	XORQ R13, BX
	DECQ SI
	DECQ R15
	JNZ  flush_buffer

	ADDQ BX, R8
	ADCQ CX, R9

	ADDV	$(1024*8), R3

	// Ask signaller to setgid
	MOVW	$1, R12
	DBAR
	MOVW	R12, ·Baton(SB)
	DBAR

	// Wait for setgid completion
loop:
	DBAR
	MOVW	·Baton(SB), R12
	OR	R13, R13, R13	// hint that we're in a spin loop
	BNE	R12, loop
	DBAR

	// Restore stack
	ADDV	$(-1024*8), R3

	// regular loop body unrolled 4x
	ADDQ CX, CX		// restore CF
	MOVQ 0(R8)(SI*8), R11
	MOVQ 8(R8)(SI*8), R12
	MOVQ 16(R8)(SI*8), R13
	MOVQ 24(R8)(SI*8), R14
	ADCQ 0(R9)(SI*8), R11
	ADCQ 8(R9)(SI*8), R12
	ADCQ 16(R9)(SI*8), R13
	ADCQ 24(R9)(SI*8), R14
	MOVQ R11, 0(R10)(SI*8)
	MOVQ R12, 8(R10)(SI*8)
	MOVQ R13, 16(R10)(SI*8)
	MOVQ R14, 24(R10)(SI*8)
	SBBQ CX, CX		// save CF

	ADDQ $4, SI		// i += 4
	SUBQ $4, DI		// n -= 4

	MOVD	0(R1), R3 // argc
	ADD	$8, R1, R4 // argv

	// Did the TLS pointer get set? If so, don't change it (e.g musl).
	CMP	R13, $0
	BNE	tls_and_argcv_in_reg

	MOVD	$runtime·m0+m_tls(SB), R13 // TLS
	ADD	$0x7000, R13

tls_and_argcv_in_reg:
	BR	main(SB)

TEXT main(SB),NOSPLIT,$-8
	MOVD	$runtime·rt0_go(SB), R12
	MOVD	R12, CTR

	MOVW	$1, R13
	TBZ	$1, R8, ks192
	TBNZ	$2, R8, ks256
	LDPW	(R9), (R4, R5)
	LDPW	8(R9), (R6, R7)
	STPW.P	(R4, R5), 8(R10)
	STPW.P	(R6, R7), 8(R10)
	MOVW	$0x1b, R14
ks128Loop:
		VMOV	R7, V2.S[0]
		WORD	$0x4E030042       // TBL V3.B16, [V2.B16], V2.B16
		AESE	V0.B16, V2.B16    // Use AES to compute the SBOX
		EORW	R13, R4
		LSLW	$1, R13           // Compute next Rcon

	//TODO LDTR -0x1e(R3), R4                  // 64285eb8
	//TODO LDTR -0xe5(R3), R10                 // 6ab851f8
	//TODO LDTRB 0xf0(R13), R10                // aa094f38
	//TODO LDTRH 0xe8(R13), R23                // b7894e78
	//TODO LDTRSB -0x24(R20), R5               // 85cadd38
	//TODO LDTRSB -0x75(R9), R13               // 2db99838
	//TODO LDTRSH 0xef(R3), LR                 // 7ef8ce78
	//TODO LDTRSH 0x96(R19), R24               // 786a8978

	// Execute call on m->g0.
	get_tls(R15)
	CMPQ	R15, $0
	JE	noswitch

	MOVQ	g(R15), R13
	CMPQ	R13, $0
	JE	noswitch
	MOVQ	g_m(R13), R13
	CMPQ	R13, $0
	JE	noswitch
	// TODO(aram): do something about the cpu profiler here.

	MOVQ	m_g0(R13), R14
	CMPQ	g(R15), R14
	JNE	switch
	// executing on m->g0 already
	CALL	AX
	RET

switch:

	SAVE_R19_TO_R28(8*9)

	// Create a struct callbackArgs on our stack.
	MOVD	$cbargs-(18*8+callbackArgs__size)(SP), R13
	MOVD	R12, callbackArgs_index(R13)	// callback index
	MOVD	R14, R0
	MOVD	R0, callbackArgs_args(R13)		// address of args vector
	MOVD	$0, R0
	MOVD	R0, callbackArgs_result(R13)	// result

	// Call cgocallback, which will call callbackWrap(frame).

	*R10
	*R11
	*R12
}

type R6 R5
type R7 R5
type R8 R5

type R9 struct {
	*R13
	*R14
	*R15
	*R16
}

type R10 R9
type R11 R9
type R12 R9

type R13 struct {
	*R17
	*R18
	*R19
	*R20
}

type R14 R13
type R15 R13
type R16 R13

type R17 struct {
	*R21
	*R22
	*R23
	*R24
}

type R18 R17
type R19 R17

	MOVV	(g_sched+gobuf_sp)(g), R3

	// Now on a scheduling stack (a pthread-created stack).
g0:
	// Save room for two of our pointers.
	ADDV	$-16, R3
	MOVV	R13, 0(R3)	// save old g on stack
	MOVV	(g_stack+stack_hi)(R13), R13
	SUBVU	R12, R13
	MOVV	R13, 8(R3)	// save depth in old g stack (can't just save SP, as stack might be copied during a callback)
	JAL	(R25)

	// Restore g, stack pointer. R4 is return value.