MOVL	AX, CX; \
	RORL	$17, AX; \
	MOVL	CX, DX; \
	RORL	$19, CX; \
	SHRL	$10, DX; \
	MOVL	((index-15)*4)(BP), BX; \
	XORL	CX, AX; \
	MOVL	BX, CX; \
	XORL	DX, AX; \
	RORL	$7, BX; \
	MOVL	CX, DX; \
	SHRL	$3, DX; \
	RORL	$18, CX; \
	ADDL	((index-7)*4)(BP), AX; \
	XORL	CX, BX; \
	XORL	DX, BX; \
	ADDL	((index-16)*4)(BP), BX; \
	ADDL	BX, AX; \
	MOVL	AX, ((index)*4)(BP)

	LEAL	const(a)(DI*1),a; \
	MOVL	(index*4)(SI),DI; \
	XORL	d,		BP; \
	XORL	b,		BP; \
	ADDL	BP,		a; \
	ROLL	$shift,		a; \
	MOVL	b,		BP; \
	ADDL	b,		a

#define ROUND4(a, b, c, d, index, const, shift) \
	LEAL	const(a)(DI*1),a; \
	ORL	b,		BP; \
	XORL	c,		BP; \
	ADDL	BP,		a; \
	MOVL	(index*4)(SI),DI; \
	MOVL	$0xffffffff,	BP; \
	ROLL	$shift,		a; \
	XORL	c,		BP; \
	ADDL	b,		a

TEXT	·block(SB),NOSPLIT,$24-16

#define ROUND3FIRST(a, b, c, d, index, const, shift) \
	MOVL	d,		R9; \
	XORL	c,		R9; \
	XORL	b,		R9; \
	ADDL	$const,	a; \
	ADDL	R8,		a; \
	MOVL	(index*4)(SI),R8; \
	ADDL	R9,		a; \
	ROLL	$shift,		a; \
	ADDL	b,		a

#define ROUND3(a, b, c, d, index, const, shift) \
	XORL	a,		R9; \
	XORL	b,		R9; \
	ADDL	$const,	a; \
	ADDL	R8,		a; \
	MOVL	(index*4)(SI),R8; \
	ADDL	R9,		a; \

	BSWAPL	DI
	XORL	SI, DI	// find bit differences
	JEQ	allsame
	BSRL	DI, CX	// index of highest bit difference
	SHRL	CX, SI	// move a's bit to bottom
	ANDL	$1, SI	// mask bit
	LEAL	-1(SI*2), BX // 1/0 => +1/-1
	MOVL	BX, (AX)
	RET

	// all the bytes in common are the same, so we just need
	// to compare the lengths.
allsame:
	XORL	BX, BX
	XORL	CX, CX
	TESTL	DX, DX

	MOVL	AX, CX; \
	RORL	$17, AX; \
	MOVL	CX, DX; \
	RORL	$19, CX; \
	SHRL	$10, DX; \
	MOVL	((index-15)*4)(BP), BX; \
	XORL	CX, AX; \
	MOVL	BX, CX; \
	XORL	DX, AX; \
	RORL	$7, BX; \
	MOVL	CX, DX; \
	SHRL	$3, DX; \
	RORL	$18, CX; \
	ADDL	((index-7)*4)(BP), AX; \
	XORL	CX, BX; \
	XORL	DX, BX; \
	ADDL	((index-16)*4)(BP), BX; \
	ADDL	BX, AX; \
	MOVL	AX, ((index)*4)(BP)

	MOVL	(((index)&0xf)*4)(SP), DI; \
	XORL	(((index-3)&0xf)*4)(SP), DI; \
	XORL	(((index-8)&0xf)*4)(SP), DI; \
	XORL	(((index-14)&0xf)*4)(SP), DI; \
	ROLL	$1, DI; \
	MOVL	DI, (((index)&0xf)*4)(SP); \
	ADDL	DI, e

#define FUNC1(a, b, c, d, e) \
	MOVL	d, DI; \
	XORL	c, DI; \
	ANDL	b, DI; \
	XORL	d, DI

#define FUNC2(a, b, c, d, e) \
	MOVL	b, DI; \
	XORL	c, DI; \
	XORL	d, DI

#define FUNC3(a, b, c, d, e) \

	movq %rdi, %rax
	divq %r8

	# expand remainder to 128 for return
	movq %rdx, %rax
	xorl %edx, %edx
	ret

1:
	# crash - only want 64-bit divisor
	xorl %ecx, %ecx
	movl %ecx, 0(%ecx)
	jmp 1b

.section .note.GNU-stack,"",@progbits
EOF

cat >umod-arm64.c <<'EOF'
typedef unsigned int u128 __attribute__((mode(TI)));

	n += a &^ (1 << 0)

	return n
}

func bitcompl32(a, b uint32) (n uint32) {
	// amd64:"BTCL"
	n += b ^ (1 << (a & 31))

	// amd64:"XORL\t[$]-2147483648"
	n += a ^ (1 << 31)

	// amd64:"XORL\t[$]268435456"
	n += a ^ (1 << 28)

	// amd64:"XORL\t[$]1"
	n += a ^ (1 << 0)

	return n
}

// check direct operation on memory with constant and shifted constant sources

TEXT ·a2(SB), 0, $0-0
	CMPL runtime·writeBarrier(SB), $0
	MOVQ $0, R15
	RET
TEXT ·a3(SB), 0, $0-0
	CMPL runtime·writeBarrier(SB), $0
	XORL R15, R15
	RET
TEXT ·a4(SB), 0, $0-0
	CMPL runtime·writeBarrier(SB), $0
	XORQ R15, R15
	RET
TEXT ·a5(SB), 0, $0-0
	CMPL runtime·writeBarrier(SB), $0
	XORL R15, R15
	RET
TEXT ·a6(SB), 0, $0-0
	CMPL runtime·writeBarrier(SB), $0
	POPQ R15
	PUSHQ R15
	RET
TEXT ·a7(SB), 0, $0-0

// func memclrNoHeapPointers(ptr unsafe.Pointer, n uintptr)
TEXT runtime·memclrNoHeapPointers(SB), NOSPLIT, $0-8
	MOVL	ptr+0(FP), DI
	MOVL	n+4(FP), BX
	XORL	AX, AX

tail:
	TESTL	BX, BX
	JEQ	_0
	CMPL	BX, $2
	JBE	_1or2
	CMPL	BX, $4
	JB	_3
	JE	_4
	CMPL	BX, $8
	JBE	_5through8
	CMPL	BX, $16
	JBE	_9through16
	MOVL	BX, CX
	SHRL	$2, CX
	REP