child_tid  unsafe.Pointer // *int64
	parent_tid *int64
	flags      int32
	pad_cgo_0  [4]byte
	rtp        *rtprio
	spare      [3]uintptr
}

type thread int64 // long

type sigset struct {
	__bits [4]uint32
}

type stackt struct {
	ss_sp     uintptr
	ss_size   uintptr
	ss_flags  int32
	pad_cgo_0 [4]byte
}

type siginfo struct {
	si_signo  int32
	si_errno  int32

	child_tid  unsafe.Pointer // *int64
	parent_tid *int64
	flags      int32
	pad_cgo_0  [4]byte
	rtp        *rtprio
	spare      [3]uintptr
}

type thread int64 // long

type sigset struct {
	__bits [4]uint32
}

type stackt struct {
	ss_sp     uintptr
	ss_size   uintptr
	ss_flags  int32
	pad_cgo_0 [4]byte
}

type siginfo struct {
	si_signo  int32
	si_errno  int32

	// cur is the bit buffer that has not been fed into n.
	// cbits is the number of low order bits in cur that are valid.
	// sbits is the number of bits of the symbol prefix being decoded.
	cur, cbits, sbits := uint(0), uint8(0), uint8(0)
	for _, b := range v {
		cur = cur<<8 | uint(b)
		cbits += 8
		sbits += 8
		for cbits >= 8 {
			idx := byte(cur >> (cbits - 8))
			n = n.children[idx]
			if n == nil {

func test32(a, b float32) {
	abits := math.Float32bits(a)
	bbits := math.Float32bits(b)
	if abits != bbits {
		panic(fmt.Sprintf("%08x != %08x\n", abits, bbits))
	}
}

func test64(a, b float64) {
	abits := math.Float64bits(a)
	bbits := math.Float64bits(b)
	if abits != bbits {
		panic(fmt.Sprintf("%016x != %016x\n", abits, bbits))
	}

	for i, offs := range offset {
		rodata := sections[i]
		for _, off := range offs {
			typ := (*rtype)(resolveTypeOff(rodata, off))
			r = append(r, typ.String())
		}
	}
	return r
}

var GCBits = gcbits

func gcbits(any) []byte // provided by runtime

func MapBucketOf(x, y Type) Type {
	return toType(bucketOf(x.common(), y.common()))
}

func CachedBucketOf(m Type) Type {
	t := m.(*rtype)

	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 {

Russ Cox <******@****.***> 1643915146 -0500

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
}

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>