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

	"WFLNDB",
	"VFLPDB",
	"WFLPDB",
	"VFSQ",
	"VFSQDB",
	"WFSQDB",
	"VFS",
	"VFSDB",
	"WFSDB",
	"VFTCI",
	"VFTCIDB",
	"WFTCIDB",
	"VGFM",
	"VGFMB",
	"VGFMH",
	"VGFMF",
	"VGFMG",
	"VGFMA",
	"VGFMAB",
	"VGFMAH",
	"VGFMAF",
	"VGFMAG",
	"VGEF",
	"VGEG",
	"VGBM",
	"VZERO",
	"VONE",
	"VGM",
	"VGMB",
	"VGMH",
	"VGMF",
	"VGMG",
	"VISTR",