VESLG $26, H_1, H_1
	VESLG $26, H_3, H_3
	VO    H_0, H_1, H_0
	VO    H_2, H_3, H_2
	VESLG $4, H_2, H_2
	VLEIB $7, $48, H_1
	VSLB  H_1, H_2, H_2
	VO    H_0, H_2, H_0
	VLEIB $7, $104, H_1
	VSLB  H_1, H_4, H_3
	VO    H_3, H_0, H_0
	VLEIB $7, $24, H_1
	VSRLB H_1, H_4, H_1

	// update state
	VSTEG $1, H_0, 0(R1)
	VSTEG $0, H_0, 8(R1)
	VSTEG $1, H_1, 16(R1)
	RET

	ADD     $16, R3
	ADD     $-16, R4

	// Process remaining data chunks
	CMP     R4 ,$16
	BGE     fold_16bytes_loop

final_fold:
	VLEIB   $7, $0x40, V9
	VSRLB   V9, CONST_R4R3, V0
	VLEIG   $0, $1, V0

	VGFMG   V0, V1, V1

	VLEIB   $7, $0x20, V9         // Shift by words
	VSRLB   V9, V1, V2            // Store remaining bits in V2
	VUPLLF  V1, V1                // Split rightmost doubleword

	"VLREPB",
	"VLREPH",
	"VLREPF",
	"VLREPG",
	"VLC",
	"VLCB",
	"VLCH",
	"VLCF",
	"VLCG",
	"VLEH",
	"VLEF",
	"VLEG",
	"VLEB",
	"VLEIH",
	"VLEIF",
	"VLEIG",
	"VLEIB",
	"VFI",
	"VFIDB",
	"WFIDB",
	"VLGV",
	"VLGVB",
	"VLGVH",
	"VLGVF",
	"VLGVG",
	"VLLEZ",
	"VLLEZB",
	"VLLEZH",
	"VLLEZF",
	"VLLEZG",
	"VLM",
	"VLP",
	"VLPB",

	MOVD counter+64(FP), R7    // R7=counter

	// load BSWAP and J0
	VLM (R1), BSWAP, J0

	// setup
	MOVD  $95, R0
	VLM   (R5), KEY0, KEY1
	VLL   R0, (R6), NONCE
	VZERO M0
	VLEIB $7, $32, M0
	VSRLB M0, NONCE, NONCE

	// initialize counter values
	VLREPF (R7), CTR
	VZERO  INC
	VLEIF  $1, $1, INC
	VLEIF  $2, $2, INC
	VLEIF  $3, $3, INC
	VAF    INC, CTR, CTR
	VREPIF $4, INC

	VSCEF	$0, V2, 4095(R15)(V1*1) // e721ffff001b
	VLL	R0, (R15), V1           // e710f0000037
	VSTL	R0, V16, (R15)          // e700f000083f
	VGMH	$8, $16, V12            // e7c008101046
	VLEIB	$15, $255, V0           // e70000fff040
	VLEIH	$7, $-32768, V15        // e7f080007041
	VLEIF	$2, $-43, V16           // e700ffd52843
	VLEIG	$1, $32767, V31         // e7f07fff1842