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}

		setCarrierSym(t, s.Sym())
		return s.Sym()
	}
	var (
		symgostring = groupSym("go:string.*", sym.SGOSTRING)
		symgofunc   = groupSym("go:func.*", sym.SGOFUNC)
		symgcbits   = groupSym("runtime.gcbits.*", sym.SGCBITS)
	)

	symgofuncrel := symgofunc
	if ctxt.UseRelro() {
		symgofuncrel = groupSym("go:funcrel.*", sym.SGOFUNCRELRO)
	}

	symt := ldr.CreateSymForUpdate("runtime.symtab", 0)

	// n must always be a multiple of 8, since gcBits is
	// always rounded up 8 bytes.
	for _, n := range []int{8, 16, 32, 64, 128} {
		b.Run(fmt.Sprintf("bits=%d", n*8), func(b *testing.B) {
			// Initialize a new byte slice with pseduo-random data.
			bits := make([]byte, n)
			rand.Read(bits)

			b.ResetTimer()
			for i := 0; i < b.N; i++ {
				runtime.MSpanCountAlloc(s, bits)
			}
		})
	}
}

		lock(&mheap_.lock)
		mheap_.spanalloc.free(unsafe.Pointer(s))
		unlock(&mheap_.lock)
	})
}

func MSpanCountAlloc(ms *MSpan, bits []byte) int {
	s := (*mspan)(ms)
	s.nelems = uint16(len(bits) * 8)
	s.gcmarkBits = (*gcBits)(unsafe.Pointer(&bits[0]))
	result := s.countAlloc()
	s.gcmarkBits = nil
	return result
}

const (
	TimeHistSubBucketBits = timeHistSubBucketBits

		f := &instFormats[i]
		if xNoCond&(f.mask|condMask) != f.value || f.priority <= priority {
			continue
		}
		delta := uint32(0)
		deltaShift := uint(0)
		for opBits := f.opBits; opBits != 0; opBits >>= 16 {
			n := uint(opBits & 0xFF)
			off := uint((opBits >> 8) & 0xFF)
			delta |= (x >> off) & (1<<n - 1) << deltaShift
			deltaShift += n
		}
		op := f.op + Op(delta)

	mgf1XOR(db, hash, h)

	// 11. Set the leftmost 8 * emLen - emBits bits of the leftmost octet in
	//     maskedDB to zero.

	db[0] &= 0xff >> (8*emLen - emBits)

	// 12. Let EM = maskedDB || H || 0xbc.
	em[emLen-1] = 0xbc

	// 13. Output EM.
	return em, nil
}

func emsaPSSVerify(mHash, em []byte, emBits, sLen int, hash hash.Hash) error {
	// See RFC 8017, Section 9.1.2.

// 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).
//

	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 {

				nbits -= shift
			}

			// convert any partial leading digit and advance to next word
			if nbits == 0 {
				// no partial digit remaining, just advance
				w = x[k]
				nbits = _W
			} else {
				// partial digit in current word w (== x[k-1]) and next word x[k]
				w |= x[k] << nbits
				i--
				s[i] = digits[w&mask]

				// advance
				w = x[k] >> (shift - nbits)

// RUN: tf-opt %s -tfl-prepare-quantize="quantize-signed=true post-training-quantize=true activation-number-of-bits=16" -cse | FileCheck %s

// CHECK-LABEL: QuantizeUnidirectionalLstmFullPerTensor
func.func @QuantizeUnidirectionalLstmFullPerTensor(%arg0: tensor<1x2x3xf32>) -> (tensor<1x2x3xf32>) {
  %input = "quantfork.stats"(%arg0) {layerStats = dense<[0.0, 1.0]> : tensor<2xf32>} : (tensor<1x2x3xf32>) -> tensor<1x2x3xf32>