MOVW	16(R1), R2
	MOVW	(R1), R2
	LL	(R1), R2

	//	LMOVB rreg ',' rreg
	//	{
	//		outcode(int($1), &$2, 0, &$4);
	//	}
	MOVB	R1, R2

	//	LMOVB addr ',' rreg
	//	{
	//		outcode(int($1), &$2, 0, &$4);
	//	}
	MOVB	foo<>+3(SB), R2
	MOVB	16(R1), R2
	MOVB	(R1), R2

	//
	// load floats
	//
	//	LFMOV addr ',' freg
	//	{
	//		outcode(int($1), &$2, 0, &$4);
	//	}
	MOVF	foo<>+3(SB), F2

	MOVV	8(R3), R13
	MOVV	8(R4), R14
	ADDV	$16, R3
	ADDV	$16, R4
	SUBVU	$1, R10
	BEQ	R13, R14, chunk16_loop
	SUBV	$8, R3
	SUBV	$8, R4

byte_loop:
	BEQ	R3, R8, samebytes
	MOVBU	(R3), R6
	ADDVU	$1, R3
	MOVBU	(R4), R7
	ADDVU	$1, R4
	BEQ	R6, R7, byte_loop

byte_cmp:
	SGTU	R6, R7, R8 // R8 = 1 if (R6 > R7)
	BNE	R0, R8, ret
	MOVV	$-1, R8
	JMP	ret

samebytes:
	SGTU	R1, R2, R6

	MOVHZ	(R5)(R8), R15
	XOR	R14, R15, R16
	MOVH	R16, (R3)(R8)
	ADD	$2,R8
	ADD	$-2,R6
xor1:
	CMP	R6, $0
	BC	12,2,LR		// BEQLR
	MOVBZ	(R4)(R8), R14	// R14 = a[i]
	MOVBZ	(R5)(R8), R15	// R15 = b[i]
	XOR	R14, R15, R16	// R16 = a[i] ^ b[i]
	MOVB	R16, (R3)(R8)	// Store to dst
#endif
done:

chunk4_loop:
	MOVW.P	4(R2), R4
	MOVW.P	4(R3), R5
	CMP	R4, R5
	BNE	cmp
	CMP	R2, R8
	BNE	chunk4_loop
	CMP	R2, R6
	BEQ	samebytes	// all compared bytes were the same; compare lengths
byte_loop:
	MOVBU.P	1(R2), R4
	MOVBU.P	1(R3), R5
	CMP	R4, R5
	BNE	ret
	CMP	R2, R6
	BNE	byte_loop
samebytes:
	CMP	R0, R1
	MOVW.LT	$1, R0
	MOVW.GT	$-1, R0
	MOVW.EQ	$0, R0
	MOVW	R0, (R7)
	RET
ret:
	// bytes differed

// Total size +4 for saved LR is 352

	// w[i] = p[j]<<24 | p[j+1]<<16 | p[j+2]<<8 | p[j+3]
	// e += w[i]
#define LOAD(Re) \
	MOVBU	2(Rdata), Rt0 ; \
	MOVBU	3(Rdata), Rt1 ; \
	MOVBU	1(Rdata), Rt2 ; \
	ORR	Rt0<<8, Rt1, Rt0	    ; \
	MOVBU.P	4(Rdata), Rt1 ; \
	ORR	Rt2<<16, Rt0, Rt0	    ; \
	ORR	Rt1<<24, Rt0, Rt0	    ; \
	MOVW.P	Rt0, 4(Rw)		    ; \
	ADD	Rt0, Re, Re

	// R3 = byte array pointer
	// R4 = length
	// R6 = byte to count
	MTVRD	R6, V1		// move compare byte
	MOVD	R6, R5
	VSPLTB	$7, V1, V1	// replicate byte across V1
	BR	countbytebody<>(SB)

TEXT ·CountString<ABIInternal>(SB), NOSPLIT|NOFRAME, $0-32
	// R3 = byte array pointer
	// R4 = length
	// R5 = byte to count
	MTVRD	R5, V1		// move compare byte
	VSPLTB	$7, V1, V1	// replicate byte across V1
	BR	countbytebody<>(SB)

	ADD	$4, R0
	ADD	$4, R2
lt_4:
	TBZ	$1, R6, lt_2
	MOVHU	(R0), R4
	MOVHU	(R2), R5
	CMPW	R4, R5
	BNE	cmp
	ADD	$2, R0
	ADD	$2, R2
lt_2:
	TBZ	$0, R6, samebytes
one:
	MOVBU	(R0), R4
	MOVBU	(R2), R5
	CMPW	R4, R5
	BNE	ret
samebytes:
	CMP	R3, R1
	CSET	NE, R0
	CNEG	LO, R0, R0
	RET
cmpnext:
	REV	R8, R4
	REV	R9, R5
	CMP	R4, R5

	MOVW	n+4(FP), N
	MOVW	$0, R0

	ADD	N, TO, TOE	/* to end pointer */

	CMP	$4, N		/* need at least 4 bytes to copy */
	BLT	_1tail

_4align:				/* align on 4 */
	AND.S	$3, TO, TMP
	BEQ	_4aligned

	MOVBU.P	R0, 1(TO)		/* implicit write back */
	B	_4align

_4aligned:
	SUB	$31, TOE, TMP	/* do 32-byte chunks if possible */
	CMP	TMP, TO
	BHS	_4tail

	MOVW	R0, R1			/* replicate */
	MOVW	R0, R2
	MOVW	R0, R3

// No argsize here, gc generates argsize info at call site.
TEXT ·makeFuncStub(SB),(NOSPLIT|WRAPPER),$20
	NO_LOCAL_POINTERS
	MOVW	R7, 4(R13)
	MOVW	$argframe+0(FP), R1
	MOVW	R1, 8(R13)
	MOVW	$0, R1
	MOVB	R1, 20(R13)
	ADD	$20, R13, R1
	MOVW	R1, 12(R13)
	MOVW	$0, R1
	MOVW	R1, 16(R13)
	BL	·callReflect(SB)
	RET

// methodValueCall is the code half of the function returned by makeMethodValue.

	MOVW R3, 16(R13)
	MOVW R4, 20(R13)
	MOVW R5, 24(R13)
	MOVW R6, 28(R13)
	MOVW R7, 32(R13)
	MOVW R8, 36(R13)
	MOVW R9, 40(R13)
	MOVW R11, 44(R13)
	MOVW R12, 48(R13)
	MOVW CPSR, R0
	MOVW R0, 52(R13)
	MOVB ·goarmsoftfp(SB), R0
	CMP $0, R0
	BNE nofp
	MOVW FPCR, R0
	MOVW R0, 56(R13)
	MOVD F0, 60(R13)
	MOVD F1, 68(R13)
	MOVD F2, 76(R13)
	MOVD F3, 84(R13)
	MOVD F4, 92(R13)
	MOVD F5, 100(R13)