PCMPEQB X1, X2
	PMOVMSKB X2, SI
	CMPQ  SI, $0xffff
	JE   success
	ADDQ $1,DI
	CMPQ DI,DX
	JB loop16
	JMP fail
_17_or_more:
	CMPQ AX, $31
	JA   _32_or_more
	LEAQ 1(DI)(DX*1), DX
	SUBQ AX, DX
	MOVOU -16(R8)(AX*1), X0
	MOVOU (R8), X1
loop17to31:
	MOVOU (DI), X2
	PCMPEQB X1,X2
	PMOVMSKB X2, SI
	CMPQ  SI, $0xffff
	JE   partial_success17to31
	ADDQ $1,DI
	CMPQ DI,DX

	for _, v := range x {
		if uint32(uint16(v)) != v {
			panic("bad 32->16 conversion")
		}
		y = append(y, uint16(v))
	}
	return y
}

func main() {
	flag.Parse()

	rang, except := scan(0, 0xFFFF)
	range16 = to16(rang)
	except16 = to16(except)
	range32, except32 = scan(0x10000, unicode.MaxRune)

	for i := rune(0); i <= unicode.MaxRune; i++ {
		if isPrint(i) != unicode.IsPrint(i) {

	// Get the number of bytes to consider in the last 16 bytes
	ANDQ $15, BX
	JZ end

	// Create mask to ignore overlap between previous 16 byte block
	// and the next.
	MOVQ $16,CX
	SUBQ BX, CX
	MOVQ $0xFFFF, R10
	SARQ CL, R10
	SALQ CL, R10

	// Process the last 16-byte chunk. This chunk may overlap with the
	// chunks we've already searched so we need to mask part of it.
	MOVOU	(AX), X1
	PCMPEQB	X0, X1

	}
	*d = digest(byteorder.BeUint32(b[len(magic):]))
	return nil
}

// Add p to the running checksum d.
func update(d digest, p []byte) digest {
	s1, s2 := uint32(d&0xffff), uint32(d>>16)
	for len(p) > 0 {
		var q []byte
		if len(p) > nmax {
			p, q = p[:nmax], p[nmax:]
		}
		for len(p) >= 4 {
			s1 += uint32(p[0])
			s2 += s1
			s1 += uint32(p[1])
			s2 += s1

	C_REG
	C_FREG
	C_FCSRREG
	C_FCCREG
	C_ZCON
	C_SCON // 12 bit signed
	C_UCON // 32 bit signed, low 12 bits 0
	C_ADD0CON
	C_AND0CON
	C_ADDCON // -0x800 <= v < 0
	C_ANDCON // 0 < v <= 0xFFF
	C_LCON   // other 32
	C_DCON   // other 64 (could subdivide further)
	C_SACON  // $n(REG) where n <= int12
	C_SECON
	C_LACON // $n(REG) where int12 < n <= int32
	C_LECON
	C_DACON // $n(REG) where int32 < n

    val version = fromSource.readByte() and 0xff
    if (version != VERSION_5) throw ProtocolException("unexpected version: $version")

    val command = fromSource.readByte() and 0xff

    val reserved = fromSource.readByte() and 0xff
    if (reserved != 0) throw ProtocolException("unexpected reserved: $reserved")

    val addressType = fromSource.readByte() and 0xff
    val toAddress =
      when (addressType) {

	C_WREG /* MSA registers */
	C_HI
	C_LO
	C_ZCON
	C_SCON /* 16 bit signed */
	C_UCON /* 32 bit signed, low 16 bits 0 */
	C_ADD0CON
	C_AND0CON
	C_ADDCON /* -0x8000 <= v < 0 */
	C_ANDCON /* 0 < v <= 0xFFFF */
	C_LCON   /* other 32 */
	C_DCON   /* other 64 (could subdivide further) */
	C_SACON  /* $n(REG) where n <= int16 */
	C_SECON
	C_LACON /* $n(REG) where int16 < n <= int32 */
	C_LECON

type huffmanNode struct {
	left, right           uint16
	leftValue, rightValue uint16
}

// invalidNodeValue is an invalid index which marks a leaf node in the tree.
const invalidNodeValue = 0xffff

// Decode reads bits from the given bitReader and navigates the tree until a
// symbol is found.
func (t *huffmanTree) Decode(br *bitReader) (v uint16) {
	nodeIndex := uint16(0) // node 0 is the root of the tree.

	procGlobalMemoryStatusEx    = modkernel32.NewProc("GlobalMemoryStatusEx")
	procGetActiveProcessorCount = modkernel32.NewProc("GetActiveProcessorCount")
)

const allProcessorGroups = 0xFFFF

// NewPerfCounterClient creates a client using perf counters
func NewPerfCounterClient() (Client, error) {
	// Initialize the cache
	initCache := cpuUsageCoreNanoSecondsCache{0, 0}

		if addr == nil {
			return nil, 0, &ParseError{Type: "IP address", Text: s}
		}
	}
	p, plen, ok := dtoi(s[i+1:])
	if !ok {
		return nil, 0, &ParseError{Type: "port", Text: s}
	}
	if p < 0 || p > 0xFFFF {
		return nil, 0, &AddrError{Err: "invalid port", Addr: s[i+1 : i+1+plen]}
	}
	return addr, p, nil
}

func readPlan9Addr(net, filename string) (addr Addr, err error) {
	var buf [128]byte