LXSDX (R3)(R0), VS1             // 7c201c98
	LXSDX (R3), VS1                 // 7c201c98
	STXVD2X VS1, (R3)(R4)           // 7c241f98
	STXVD2X VS1, (R3)(R0)           // 7c201f98
	STXVD2X VS1, (R3)               // 7c201f98
	STXVW4X VS1, (R3)(R4)           // 7c241f18
	STXVW4X VS1, (R3)(R0)           // 7c201f18
	STXVW4X VS1, (R3)               // 7c201f18
	STXV VS1,16(R3)                 // f4230015

	ADD     $96, P1ptr
	BDNZ    loop_select

	// STXVD2X is used here so that alignment doesn't
	// need to be verified. Since values were loaded
	// using LXVD2X this is OK.
	STXVD2X X1H, (P3ptr+R0)
	STXVD2X X1L, (P3ptr+R16)
	STXVD2X Y1H, (P3ptr+R17)
	STXVD2X Y1L, (P3ptr+R18)
	STXVD2X Z1H, (P3ptr+R19)
	STXVD2X Z1L, (P3ptr+R20)
	RET

#undef P3ptr
#undef P1ptr
#undef COUNT

		// Set up loop counter in CTR, used by BC
		//       XXLXOR VS32,VS32,VS32
		//	 MOVD len/32,REG_TMP
		//	 MOVD REG_TMP,CTR
		//       MOVD $16,REG_TMP
		//	 loop:
		//	 STXVD2X VS32,(R0)(R20)
		//	 STXVD2X VS32,(R31)(R20)
		//	 ADD  $32,R20
		//	 BC   16, 0, loop
		//
		// any remainder is done as described below

		// for sizes < 64 bytes, first clear as many doublewords as possible,