ROUND4(R7,R4,R5,R6,10,0x8f0ccc92,10);
	ROUND4(R6,R7,R4,R5, 1,0xffeff47d,15);
	ROUND4(R5,R6,R7,R4, 8,0x85845dd1,21);
	ROUND4(R4,R5,R6,R7,15,0x6fa87e4f, 6);
	ROUND4(R7,R4,R5,R6, 6,0xfe2ce6e0,10);
	ROUND4(R6,R7,R4,R5,13,0xa3014314,15);
	ROUND4(R5,R6,R7,R4, 4,0x4e0811a1,21);
	ROUND4(R4,R5,R6,R7,11,0xf7537e82, 6);
	ROUND4(R7,R4,R5,R6, 2,0xbd3af235,10);
	ROUND4(R6,R7,R4,R5, 9,0x2ad7d2bb,15);

	MOVL	$0xffffffff,	BP
	XORL	DX,		BP

	ROUND4(AX,BX,CX,DX, 7,0xf4292244, 6);
	ROUND4(DX,AX,BX,CX,14,0x432aff97,10);
	ROUND4(CX,DX,AX,BX, 5,0xab9423a7,15);
	ROUND4(BX,CX,DX,AX,12,0xfc93a039,21);
	ROUND4(AX,BX,CX,DX, 3,0x655b59c3, 6);
	ROUND4(DX,AX,BX,CX,10,0x8f0ccc92,10);
	ROUND4(CX,DX,AX,BX, 1,0xffeff47d,15);
	ROUND4(BX,CX,DX,AX, 8,0x85845dd1,21);
	ROUND4(AX,BX,CX,DX,15,0x6fa87e4f, 6);

	ROUND4(R2,R3,R4,R5, 3,0x655b59c3, 6);
	ROUND4(R5,R2,R3,R4,10,0x8f0ccc92,10);
	ROUND4(R4,R5,R2,R3, 1,0xffeff47d,15);
	ROUND4(R3,R4,R5,R2, 8,0x85845dd1,21);
	ROUND4(R2,R3,R4,R5,15,0x6fa87e4f, 6);
	ROUND4(R5,R2,R3,R4, 6,0xfe2ce6e0,10);
	ROUND4(R4,R5,R2,R3,13,0xa3014314,15);
	ROUND4(R3,R4,R5,R2, 4,0x4e0811a1,21);
	ROUND4(R2,R3,R4,R5,11,0xf7537e82, 6);
	ROUND4(R5,R2,R3,R4, 2,0xbd3af235,10);
	ROUND4(R4,R5,R2,R3, 9,0x2ad7d2bb,15);

	ROUND3(R3,R4,R5,R22,M02,0xc4ac5665,23);

	ROUND4(R22,R3,R4,R5,M00,0xf4292244, 6);
	ROUND4(R5,R22,R3,R4,M07,0x432aff97,10);
	ROUND4(R4,R5,R22,R3,M14,0xab9423a7,15);
	ROUND4(R3,R4,R5,R22,M05,0xfc93a039,21);
	ROUND4(R22,R3,R4,R5,M12,0x655b59c3, 6);
	ROUND4(R5,R22,R3,R4,M03,0x8f0ccc92,10);
	ROUND4(R4,R5,R22,R3,M10,0xffeff47d,15);
	ROUND4(R3,R4,R5,R22,M01,0x85845dd1,21);
	ROUND4(R22,R3,R4,R5,M08,0x6fa87e4f, 6);

	ROUND4(AX,BX,CX,DX, 3,0x655b59c3, 6);
	ROUND4(DX,AX,BX,CX,10,0x8f0ccc92,10);
	ROUND4(CX,DX,AX,BX, 1,0xffeff47d,15);
	ROUND4(BX,CX,DX,AX, 8,0x85845dd1,21);
	ROUND4(AX,BX,CX,DX,15,0x6fa87e4f, 6);
	ROUND4(DX,AX,BX,CX, 6,0xfe2ce6e0,10);
	ROUND4(CX,DX,AX,BX,13,0xa3014314,15);
	ROUND4(BX,CX,DX,AX, 4,0x4e0811a1,21);
	ROUND4(AX,BX,CX,DX,11,0xf7537e82, 6);
	ROUND4(DX,AX,BX,CX, 2,0xbd3af235,10);
	ROUND4(CX,DX,AX,BX, 9,0x2ad7d2bb,15);

	ROUND3(BX, CX, DX, BP, AX, 59)

	ROUND4(AX, BX, CX, DX, BP, 60)
	ROUND4(BP, AX, BX, CX, DX, 61)
	ROUND4(DX, BP, AX, BX, CX, 62)
	ROUND4(CX, DX, BP, AX, BX, 63)
	ROUND4(BX, CX, DX, BP, AX, 64)
	ROUND4(AX, BX, CX, DX, BP, 65)
	ROUND4(BP, AX, BX, CX, DX, 66)
	ROUND4(DX, BP, AX, BX, CX, 67)
	ROUND4(CX, DX, BP, AX, BX, 68)
	ROUND4(BX, CX, DX, BP, AX, 69)
	ROUND4(AX, BX, CX, DX, BP, 70)
	ROUND4(BP, AX, BX, CX, DX, 71)

	ROUND4(Rb, Rc, Rd, Ra,  5, 21, Rc3)

	MOVM.IA.W (Rtable), [Rc0,Rc1,Rc2,Rc3]
	ROUND4(Ra, Rb, Rc, Rd, 12,	6, Rc0)
	ROUND4(Rd, Ra, Rb, Rc,  3, 10, Rc1)
	ROUND4(Rc, Rd, Ra, Rb, 10, 15, Rc2)
	ROUND4(Rb, Rc, Rd, Ra,  1, 21, Rc3)

	MOVM.IA.W (Rtable), [Rc0,Rc1,Rc2,Rc3]
	ROUND4(Ra, Rb, Rc, Rd,  8,	6, Rc0)
	ROUND4(Rd, Ra, Rb, Rc, 15, 10, Rc1)
	ROUND4(Rc, Rd, Ra, Rb,  6, 15, Rc2)
	ROUND4(Rb, Rc, Rd, Ra, 13, 21, Rc3)

	ROUND3(Rc, Rd, Re, Ra, Rb)
	ROUND3(Rb, Rc, Rd, Re, Ra)
	SUB.S	$1, Rctr
	BNE	loop3

	MOVW	$0xCA62C1D6, Rconst
	MOVW	$4, Rctr
loop4:	ROUND4(Ra, Rb, Rc, Rd, Re)
	ROUND4(Re, Ra, Rb, Rc, Rd)
	ROUND4(Rd, Re, Ra, Rb, Rc)
	ROUND4(Rc, Rd, Re, Ra, Rb)
	ROUND4(Rb, Rc, Rd, Re, Ra)
	SUB.S	$1, Rctr
	BNE	loop4

	// Accumulate - restoring registers from SP+4
	MOVM.IB (R13), [Rt0,Rt1,Rt2,Rctr,Rw]
	ADD	Rt0, Ra

	ROUND3(BX, CX, DX, BP, AX, 59)

	ROUND4(AX, BX, CX, DX, BP, 60)
	ROUND4(BP, AX, BX, CX, DX, 61)
	ROUND4(DX, BP, AX, BX, CX, 62)
	ROUND4(CX, DX, BP, AX, BX, 63)
	ROUND4(BX, CX, DX, BP, AX, 64)
	ROUND4(AX, BX, CX, DX, BP, 65)
	ROUND4(BP, AX, BX, CX, DX, 66)
	ROUND4(DX, BP, AX, BX, CX, 67)
	ROUND4(CX, DX, BP, AX, BX, 68)
	ROUND4(BX, CX, DX, BP, AX, 69)
	ROUND4(AX, BX, CX, DX, BP, 70)
	ROUND4(BP, AX, BX, CX, DX, 71)

	VREPF $3, KEY1, X11
	VLR   CTR, X12
	VREPF $1, NONCE, X13
	VREPF $2, NONCE, X14
	VREPF $3, NONCE, X15

	MOVD $(NUM_ROUNDS/2), R1

loop:
	ROUND4(X0, X4, X12,  X8, X1, X5, X13,  X9, X2, X6, X14, X10, X3, X7, X15, X11)
	ROUND4(X0, X5, X15, X10, X1, X6, X12, X11, X2, X7, X13, X8,  X3, X4, X14, X9)

	ADD $-1, R1
	BNE loop

	// decrement length
	ADD $-256, R4

	// rearrange vectors