cnt  uint32  // number of valid bits in the bits field
}

// makeBitReader makes a bit reader starting at off.
func (r *Reader) makeBitReader(data block, off int) bitReader {
	return bitReader{
		r:    r,
		data: data,
		off:  uint32(off),
	}
}

// moreBits is called to read more bits.
// This ensures that at least 16 bits are available.
func (br *bitReader) moreBits() error {
	for br.cnt < 16 {

// To rotate x right by k bits, call RotateLeft8(x, -k).
//
// This function's execution time does not depend on the inputs.
func RotateLeft8(x uint8, k int) uint8 {
	const n = 8
	s := uint(k) & (n - 1)
	return x<<s | x>>(n-s)
}

// RotateLeft16 returns the value of x rotated left by (k mod 16) bits.
// To rotate x right by k bits, call RotateLeft16(x, -k).
//

func bitSetTest(x int) bool {
	// amd64:"ANDL\t[$]9, AX"
	// amd64:"CMPQ\tAX, [$]9"
	return x&9 == 9
}

// mask contiguous one bits
func cont1Mask64U(x uint64) uint64 {
	// s390x:"RISBGZ\t[$]16, [$]47, [$]0,"
	return x & 0x0000ffffffff0000
}

// mask contiguous zero bits
func cont0Mask64U(x uint64) uint64 {
	// s390x:"RISBGZ\t[$]48, [$]15, [$]0,"
	return x & 0xffff00000000ffff
}

  %dq_weight = "quantfork.dcast"(%q_weight) : (tensor<2x3x3x2x!quant.uniform<i8:f32, 0.074855112561992565:-1>>) -> tensor<2x3x3x2xf32>
  %q_bias = "quantfork.qcast"(%bias) : (tensor<2xf32>) -> tensor<2x!quant.uniform<i32:f32, 0.044022349891595126>>
  %dq_bias = "quantfork.dcast"(%q_bias) : (tensor<2x!quant.uniform<i32:f32, 0.044022349891595126>>) -> tensor<2xf32>

namespace mlir {
namespace quant {

constexpr int k8Bits = 8;
constexpr int k32Bits = 32;
constexpr unsigned kSigned = QuantizationFlags::Signed;

DeviceTarget::DeviceTarget(MLIRContext* ctx) : ctx_(ctx) {
  f32_ = FloatType::getF32(ctx_);
  i8_ = IntegerType::get(ctx_, k8Bits);
  i8_min_ = QuantizedType::getDefaultMinimumForInteger(kSigned, k8Bits);
  i8_max_ = QuantizedType::getDefaultMaximumForInteger(kSigned, k8Bits);

  %dq_weight = "quantfork.dcast"(%q_weight) : (tensor<2x3x3x2x!quant.uniform<i8:f32, 0.074855112561992565:-1>>) -> tensor<2x3x3x2xf32>
  %q_bias = "quantfork.qcast"(%bias) : (tensor<2xf32>) -> tensor<2x!quant.uniform<i32:f32, 0.044022349891595126>>
  %dq_bias = "quantfork.dcast"(%q_bias) : (tensor<2x!quant.uniform<i32:f32, 0.044022349891595126>>) -> tensor<2xf32>

	if w.err != nil {
		return
	}
	w.bits |= uint64(b) << w.nbits
	w.nbits += nb
	if w.nbits >= 48 {
		bits := w.bits
		w.bits >>= 48
		w.nbits -= 48
		n := w.nbytes
		bytes := w.bytes[n : n+6]
		bytes[0] = byte(bits)
		bytes[1] = byte(bits >> 8)
		bytes[2] = byte(bits >> 16)
		bytes[3] = byte(bits >> 24)
		bytes[4] = byte(bits >> 32)
		bytes[5] = byte(bits >> 40)
		n += 6
		if n >= bufferFlushSize {

}

// writeLSB writes the code c for "Least Significant Bits first" data.
func (w *Writer) writeLSB(c uint32) error {
	w.bits |= c << w.nBits
	w.nBits += w.width
	for w.nBits >= 8 {
		if err := w.w.WriteByte(uint8(w.bits)); err != nil {
			return err
		}
		w.bits >>= 8
		w.nBits -= 8
	}
	return nil
}

// writeMSB writes the code c for "Most Significant Bits first" data.
func (w *Writer) writeMSB(c uint32) error {

		atomic.And8(v.bytep, ^mask)
	}
}

// pinnerBits is the same type as gcBits but has different methods.
type pinnerBits gcBits

// ofObject returns the pinState of the n'th object.
// nosplit, because it's called by isPinned, which is nosplit
//
//go:nosplit
func (p *pinnerBits) ofObject(n uintptr) pinState {
	bytep, mask := (*gcBits)(p).bitp(n * 2)
	byteVal := atomic.Load8(bytep)
	return pinState{bytep, byteVal, mask}

}

// readLSB returns the next code for "Least Significant Bits first" data.
func (r *Reader) readLSB() (uint16, error) {
	for r.nBits < r.width {
		x, err := r.r.ReadByte()
		if err != nil {
			return 0, err
		}
		r.bits |= uint32(x) << r.nBits
		r.nBits += 8
	}
	code := uint16(r.bits & (1<<r.width - 1))
	r.bits >>= r.width
	r.nBits -= r.width
	return code, nil
}