// R4 = length
	// R6 = byte to count
	MTVRD	R6, V1		// move compare byte
	MOVD	R6, R5
	VSPLTB	$7, V1, V1	// replicate byte across V1
	BR	countbytebody<>(SB)

TEXT ·CountString<ABIInternal>(SB), NOSPLIT|NOFRAME, $0-32
	// R3 = byte array pointer
	// R4 = length
	// R5 = byte to count
	MTVRD	R5, V1		// move compare byte
	VSPLTB	$7, V1, V1	// replicate byte across V1
	BR	countbytebody<>(SB)

// R3: addr of string

	"VCMPNEB",
	"VCMPNEBCC",
	"VCMPNEH",
	"VCMPNEHCC",
	"VCMPNEW",
	"VCMPNEWCC",
	"VPERM",
	"VPERMXOR",
	"VPERMR",
	"VBPERMQ",
	"VBPERMD",
	"VSEL",
	"VSPLTB",
	"VSPLTH",
	"VSPLTW",
	"VSPLTISB",
	"VSPLTISH",
	"VSPLTISW",
	"VCIPH",
	"VCIPHER",
	"VCIPHERLAST",
	"VNCIPH",
	"VNCIPHER",
	"VNCIPHERLAST",
	"VSBOX",
	"VSHASIGMA",
	"VSHASIGMAW",
	"VSHASIGMAD",

	CMPU	R4,$32

#ifndef GOPPC64_power9
	// Load VBPERMQ constant to reduce compare into an ordered bit mask.
	MOVD	$indexbytevbperm<>+00(SB),R16
	LXVD2X	(R16),V0	// Set up swap string
#endif

	MTVRD	R5,V1
	VSPLTB	$7,V1,V1	// Replicate byte across V1

	BLT	cmp16		// Jump to the small string case if it's <32 bytes.

	CMP	R4,$64,CR1
	MOVD	$16,R11
	MOVD	R3,R8
	BLT	CR1,cmp32	// Special case for length 32 - 63
	MOVD	$32,R12

	// splat values from V17, V18 into V4-V11
	VSPLTW $0, V17, V4
	VSPLTW $1, V17, V5
	VSPLTW $2, V17, V6
	VSPLTW $3, V17, V7
	VSPLTW $0, V18, V8
	VSPLTW $1, V18, V9
	VSPLTW $2, V18, V10
	VSPLTW $3, V18, V11

	// VOR
	VOR V26, V26, V12

	// splat values from V19 -> V13, V14, V15
	VSPLTW $1, V19, V13
	VSPLTW $2, V19, V14
	VSPLTW $3, V19, V15

	// splat   const values

	"VSUMQ",
	"VSUMQF",
	"VSUMQG",
	"VSUM",
	"VSUMB",
	"VSUMH",
	"VTM",
	"VUPH",
	"VUPHB",
	"VUPHH",
	"VUPHF",
	"VUPLH",
	"VUPLHB",
	"VUPLHH",
	"VUPLHF",
	"VUPLL",
	"VUPLLB",
	"VUPLLH",
	"VUPLLF",
	"VUPL",
	"VUPLB",
	"VUPLHW",
	"VUPLF",
	"VMSLG",
	"VMSLEG",
	"VMSLOG",
	"VMSLEOG",
	"NOPH",
	"BYTE",
	"WORD",
	"DWORD",
	"BRRK",
	"LAST",

        return new HashMap<>();
    }

    ArchitectureSpecificVisualCpp buildDescriptor(VersionNumber compilerVersion, File basePath, File vsPath) {
        File commonTools = new File(vsPath, PATH_COMMONTOOLS);
        File commonIde = new File(vsPath, PATH_COMMONIDE);
        List<File> paths = Lists.newArrayList(commonTools, commonIde);
        File crossCompilePath = getCrossCompilePath(basePath);

	VONE	V16
	VREPF	$0, V0, V1
index4loop:
	VL	(R7), V2           // load 16-bytes into V2
	VLL	R0, 16(R7), V3     // load 3-bytes into V3
	VSLDB	$1, V2, V3, V4     // V4=(V2:V3)<<1
	VSLDB	$2, V2, V3, V9     // V9=(V2:V3)<<1
	VSLDB	$3, V2, V3, V10    // V10=(V2:V3)<<1
	VCEQF	V1, V2, V5         // compare index 0, 4, ...
	VCEQF	V1, V4, V6         // compare index 1, 5, ...

	// 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

	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

chacha: