MOVD	ptr+0(FP), R3
	MOVBZ	val+8(FP), R4
	LWSYNC
again:
	LBAR	(R3), R6
	OR	R4, R6
	STBCCC	R6, (R3)
	BNE	again
	RET

// void ·And8(byte volatile*, byte);
TEXT ·And8(SB), NOSPLIT, $0-9
	MOVD	ptr+0(FP), R3
	MOVBZ	val+8(FP), R4
	LWSYNC
again:
	LBAR	(R3), R6
	AND	R4, R6
	STBCCC	R6, (R3)
	BNE	again
	RET

// func Or(addr *uint32, v uint32)

len_8:
	// R5 contains 8-byte of sep
	MOVD	(R2), R5
loop_8:
	// R6 contains substring for comparison
	CMP	R4, R0
	BHI	not_found
	MOVD.P	1(R0), R6
	CMP	R5, R6
	BNE	loop_8
	B	found
len_2_7:
	TBZ	$2, R3, len_2_3
	TBZ	$1, R3, len_4_5
	TBZ	$0, R3, len_6
len_7:
	// R5 and R6 contain 7-byte of sep
	MOVWU	(R2), R5
	// 1-byte overlap with R5
	MOVWU	3(R2), R6
loop_7:
	CMP	R4, R0
	BHI	not_found
	MOVWU.P	1(R0), R3

	LDADDALW	R5, (R6), RSP                            // ERROR "illegal combination"
	LDADDALH	R5, (R6), RSP                            // ERROR "illegal combination"
	LDADDALB	R5, (R6), RSP                            // ERROR "illegal combination"
	LDADDD	R5, (R6), RSP                                    // ERROR "illegal combination"
	LDADDW	R5, (R6), RSP                                    // ERROR "illegal combination"

	SUB	$0x20, R4, R4
	NEG	R4<<1, R4
	LSL	R4, R6, R6
	LSR	R4, R6, R6

tail:
	// Check that we have found a character
	CBZ	R6, fail
	// Count the trailing zeros using bit reversing
	RBIT	R6, R6
	// Compensate the last post-increment
	SUB	$0x20, R3, R3
	// And count the leading zeros
	CLZ	R6, R6
	// R6 is twice the offset into the fragment
	ADD	R6>>1, R3, R0
	// Compute the offset result

	SUBVU	R6, R2, R4
	AND	$7, R4
	BNE	R4, out1

	// if less than 8 bytes, do byte copying
	SGTU	$8, R3, R4
	BNE	R4, out1

	// do one byte at a time until 8-aligned
	AND	$7, R6, R5
	BEQ	R5, words1
	ADDV	$-1, R2
	MOVB	(R2), R4
	ADDV	$-1, R6
	MOVB	R4, (R6)
	JMP	-6(PC)

words1:
	// do 8 bytes at a time if there is room
	ADDV	$7, R1, R3 // R3 is start pointer+7

	SGTU	R6, R3, R5
	BEQ	R5, out1

	MULU	R4, R5, R6		// a6101c00
	MULH	R4, R5	   		// a5901c00
	MULH	R4, R5, R6	   	// a6901c00
	MULHU	R4, R5			// a5101d00
	MULHU	R4, R5, R6		// a6101d00
	REM	R4, R5	  		// a5902000
	REM	R4, R5, R6	  	// a6902000
	REMU	R4, R5	   		// a5902100
	REMU	R4, R5, R6	   	// a6902100
	DIV	R4, R5	  		// a5102000
	DIV	R4, R5, R6	  	// a6102000
	DIVU	R4, R5	   		// a5102100
	DIVU	R4, R5, R6	   	// a6102100

	MOVV	ptr+0(FP), R4
	MOVW	delta+8(FP), R5
	DBAR
	LL	(R4), R6
	ADDU	R6, R5, R7
	MOVV	R7, R6
	SC	R7, (R4)
	BEQ	R7, -4(PC)
	MOVW	R6, ret+16(FP)
	DBAR
	RET

TEXT ·Xadd64(SB), NOSPLIT, $0-24
	MOVV	ptr+0(FP), R4
	MOVV	delta+8(FP), R5
	DBAR
	LLV	(R4), R6
	ADDVU	R6, R5, R7
	MOVV	R7, R6
	SCV	R7, (R4)
	BEQ	R7, -4(PC)
	MOVV	R6, ret+16(FP)
	DBAR
	RET

TEXT ·Xchg(SB), NOSPLIT, $0-20

	ADDV	$-7, R9, R6 // R6 is end pointer-7

	PCALIGN	$16
	SGTU	R6, R4, R8
	BEQ	R8, out
	MOVV	(R5), R7
	ADDV	$8, R5
	MOVV	R7, (R4)
	ADDV	$8, R4
	JMP	-6(PC)

out:
	BEQ	R4, R9, done
	MOVB	(R5), R7
	ADDV	$1, R5
	MOVB	R7, (R4)
	ADDV	$1, R4
	JMP	-5(PC)
done:
	RET

backward:
	ADDV	R6, R5 // from-end pointer
	ADDV	R4, R6, R9 // to-end pointer

// R7 points to return value (-1/0/1 will be written here)
//
// On exit:
// R4, R5, R6 and R8 are clobbered
TEXT cmpbody<>(SB),NOSPLIT|NOFRAME,$0-0
	CMP	R2, R3
	BEQ	samebytes
	CMP 	R0, R1
	MOVW 	R0, R6
	MOVW.LT	R1, R6		// R6 is min(R0, R1)

	CMP	$0, R6
	BEQ	samebytes
	CMP	$4, R6
	ADD	R2, R6		// R2 is current byte in a, R6 is the end of the range to compare
	BLT	byte_loop	// length < 4
	AND	$3, R2, R8
	CMP	$0, R8

	LDADDLD	R5, (R6), ZR                         // df0065f8
	LDADDLW	R5, (R6), ZR                         // df0065b8
	LDADDLH	R5, (R6), ZR                         // df006578
	LDADDLB	R5, (R6), ZR                         // df006538
	LDCLRD	R5, (R6), ZR                         // df1025f8
	LDCLRW	R5, (R6), ZR                         // df1025b8
	LDCLRH	R5, (R6), ZR                         // df102578