ROUND2(R22,R3,R4,R5,M01,0xf61e2562, 5);
	ROUND2(R5,R22,R3,R4,M06,0xc040b340, 9);
	ROUND2(R4,R5,R22,R3,M11,0x265e5a51,14);
	ROUND2(R3,R4,R5,R22,M00,0xe9b6c7aa,20);
	ROUND2(R22,R3,R4,R5,M05,0xd62f105d, 5);
	ROUND2(R5,R22,R3,R4,M10, 0x2441453, 9);
	ROUND2(R4,R5,R22,R3,M15,0xd8a1e681,14);
	ROUND2(R3,R4,R5,R22,M04,0xe7d3fbc8,20);
	ROUND2(R22,R3,R4,R5,M09,0x21e1cde6, 5);
	ROUND2(R5,R22,R3,R4,M14,0xc33707d6, 9);

	SUB  R21, R5, R21

	// Greater than 8 -- load the rightmost remaining bytes in msg
	// and put into R17 (h1)
	MOVD (R4)(R21), R17
	MOVD $16, R22

	// Find the offset to those bytes
	SUB R5, R22, R22
	SLD $3, R22

	// Shift to get only the bytes in msg
	SRD R22, R17, R17

	// Put 1 at high end
	MOVD $1, R23
	SLD  $3, R21
	SLD  R21, R23, R23
	OR   R23, R17, R17

	// Remainder is 8

	AND $0x7ffffffffffff, R4, R14

	ADD R0>>51, R11, R11
	ADD R1>>51, R12, R12
	ADD R2>>51, R13, R13
	ADD R3>>51, R14, R14
	// R4>>51 * 19 + R10 -> R10
	LSR $51, R4, R21
	MOVD $19, R22
	MADD R22, R10, R21, R10

	STP (R10, R11), 0(R20)
	STP (R12, R13), 16(R20)
	MOVD R14, 32(R20)