fn decode_3bytes(src: &[u8], dst: &mut [u8]) -> i16 {
        debug_assert_eq!(src.len(), 4);
        debug_assert!(dst.len() >= 3, "dst too short: {}", dst.len());

        let c0 = Self::decode_6bits(src[0]);
        let c1 = Self::decode_6bits(src[1]);
        let c2 = Self::decode_6bits(src[2]);
        let c3 = Self::decode_6bits(src[3]);

        dst[0] = ((c0 << 2) | (c1 >> 4)) as u8;
        dst[1] = ((c1 << 4) | (c2 >> 2)) as u8;
        dst[2] = ((c2 << 6) | c3) as u8;

        ((c0 | c1 | c2 | c3) >> 8) & 1
    }

    /// Decode 6-bits of a Base64 message.
    fn decode_6bits(src: u8) -> i16 {
        let mut ret: i16 = -1;

        for step in Self::DECODER {
            ret += match step {
                DecodeStep::Range(range, offset) => {
                    // Compute exclusive range from inclusive one
                    let start = *range.start() as i16 - 1;
                    let end = *range.end() as i16 + 1;
                    (((start - src as i16) & (src as i16 - end)) >> 8) & (src as i16 + *offset)
                }
                DecodeStep::Eq(value, offset) => {
                    let start = *value as i16 - 1;
                    let end = *value as i16 + 1;
                    (((start - src as i16) & (src as i16 - end)) >> 8) & *offset
                }
            };
        }

        ret
    }

    /// Encode 3-bytes of a Base64 message.
    #[inline(always)]
    fn encode_3bytes(src: &[u8], dst: &mut [u8]) {
        debug_assert_eq!(src.len(), 3);
        debug_assert!(dst.len() >= 4, "dst too short: {}", dst.len());

        let b0 = src[0] as i16;
        let b1 = src[1] as i16;
        let b2 = src[2] as i16;

        dst[0] = Self::encode_6bits(b0 >> 2);
        dst[1] = Self::encode_6bits(((b0 << 4) | (b1 >> 4)) & 63);
        dst[2] = Self::encode_6bits(((b1 << 2) | (b2 >> 6)) & 63);
        dst[3] = Self::encode_6bits(b2 & 63);
    }

    /// Encode 6-bits of a Base64 message.
    #[inline(always)]
    fn encode_6bits(src: i16) -> u8 {
        let mut diff = src + Self::BASE as i16;

        for &step in Self::ENCODER {
            diff += match step {
                EncodeStep::Apply(threshold, offset) => ((threshold as i16 - diff) >> 8) & offset,
                EncodeStep::Diff(threshold, offset) => ((threshold as i16 - src) >> 8) & offset,
            };
        }

        diff as u8
    }

        for chunk in 0..full_chunks {
            // SAFETY: `p3` and `p4` point inside `buf`, while they may overlap,
            // read and write are clearly separated from each other and done via
            // raw pointers.
            #[allow(unsafe_code)]
            unsafe {
                debug_assert!(3 * chunk + 3 <= buf.len());
                debug_assert!(4 * chunk + 4 <= buf.len());

                let p3 = buf.as_mut_ptr().add(3 * chunk) as *mut [u8; 3];
                let p4 = buf.as_ptr().add(4 * chunk) as *const [u8; 4];

                let mut tmp_out = [0u8; 3];
                err |= Self::decode_3bytes(&*p4, &mut tmp_out);
                *p3 = tmp_out;
            }
        }

        let src_rem_pos = 4 * full_chunks;
        let src_rem_len = buf.len() - src_rem_pos;
        let dst_rem_pos = 3 * full_chunks;
        let dst_rem_len = dlen - dst_rem_pos;

        err |= !(src_rem_len == 0 || src_rem_len >= 2) as i16;
        let mut tmp_in = [b'A'; 4];
        tmp_in[..src_rem_len].copy_from_slice(&buf[src_rem_pos..]);
        let mut tmp_out = [0u8; 3];

        err |= Self::decode_3bytes(&tmp_in, &mut tmp_out);

        if err == 0 {
            // SAFETY: `dst_rem_len` is always smaller than 4, so we don't
            // read outside of `tmp_out`, write and the final slicing never go
            // outside of `buf`.
            #[allow(unsafe_code)]
            unsafe {
                debug_assert!(dst_rem_pos + dst_rem_len <= buf.len());
                debug_assert!(dst_rem_len <= tmp_out.len());
                debug_assert!(dlen <= buf.len());

                core::ptr::copy_nonoverlapping(
                    tmp_out.as_ptr(),
                    buf.as_mut_ptr().add(dst_rem_pos),
                    dst_rem_len,
                );
                Ok(buf.get_unchecked(..dlen))
            }
        } else {
            Err(InvalidEncodingError)
        }
    }

        debug_assert!(str::from_utf8(dst).is_ok());

        // SAFETY: values written by `encode_3bytes` are valid one-byte UTF-8 chars
        #[allow(unsafe_code)]
        Ok(unsafe { str::from_utf8_unchecked(dst) })
    }

    #[cfg(feature = "alloc")]
    fn encode_string(input: &[u8]) -> String {
        let elen = encoded_len_inner(input.len(), T::PADDED).expect("input is too big");
        let mut dst = vec![0u8; elen];
        let res = Self::encode(input, &mut dst).expect("encoding error");

        debug_assert_eq!(elen, res.len());
        debug_assert!(str::from_utf8(&dst).is_ok());

        // SAFETY: `dst` is fully written and contains only valid one-byte UTF-8 chars
        #[allow(unsafe_code)]
        unsafe {
            String::from_utf8_unchecked(dst)
        }
    }

#[inline(always)]
pub(crate) fn decode_padding(input: &[u8]) -> Result<(usize, i16), InvalidEncodingError> {
    if input.len() % 4 != 0 {
        return Err(InvalidEncodingError);
    }

    let unpadded_len = match *input {
        [.., b0, b1] => is_pad_ct(b0)
            .checked_add(is_pad_ct(b1))
            .and_then(|len| len.try_into().ok())
            .and_then(|len| input.len().checked_sub(len))
            .ok_or(InvalidEncodingError)?,
        _ => input.len(),
    };

    let padding_len = input
        .len()
        .checked_sub(unpadded_len)
        .ok_or(InvalidEncodingError)?;

    let err = match *input {
        [.., b0] if padding_len == 1 => is_pad_ct(b0) ^ 1,
        [.., b0, b1] if padding_len == 2 => (is_pad_ct(b0) & is_pad_ct(b1)) ^ 1,
        _ => {
            if padding_len == 0 {
                0
            } else {
                return Err(InvalidEncodingError);
            }
        }
    };

    Ok((unpadded_len, err))
}

    fn decode_in_place(mut buf: &mut [u8]) -> Result<&[u8], InvalidEncodingError> {
        // TODO: eliminate unsafe code when LLVM12 is stable
        // See: https://github.com/rust-lang/rust/issues/80963
        let mut err = if T::PADDED {
            let (unpadded_len, e) = decode_padding(buf)?;
            buf = &mut buf[..unpadded_len];
            e
        } else {
            0
        };

        let dlen = decoded_len(buf.len());
        let full_chunks = buf.len() / 4;

        for chunk in 0..full_chunks {
            // SAFETY: `p3` and `p4` point inside `buf`, while they may overlap,
            // read and write are clearly separated from each other and done via
            // raw pointers.
            #[allow(unsafe_code)]
            unsafe {
                debug_assert!(3 * chunk + 3 <= buf.len());
                debug_assert!(4 * chunk + 4 <= buf.len());

                let p3 = buf.as_mut_ptr().add(3 * chunk) as *mut [u8; 3];
                let p4 = buf.as_ptr().add(4 * chunk) as *const [u8; 4];

                let mut tmp_out = [0u8; 3];
                err |= Self::decode_3bytes(&*p4, &mut tmp_out);
                *p3 = tmp_out;
            }
        }

        let src_rem_pos = 4 * full_chunks;
        let src_rem_len = buf.len() - src_rem_pos;
        let dst_rem_pos = 3 * full_chunks;
        let dst_rem_len = dlen - dst_rem_pos;

        err |= !(src_rem_len == 0 || src_rem_len >= 2) as i16;
        let mut tmp_in = [b'A'; 4];
        tmp_in[..src_rem_len].copy_from_slice(&buf[src_rem_pos..]);
        let mut tmp_out = [0u8; 3];

        err |= Self::decode_3bytes(&tmp_in, &mut tmp_out);

        if err == 0 {
            // SAFETY: `dst_rem_len` is always smaller than 4, so we don't
            // read outside of `tmp_out`, write and the final slicing never go
            // outside of `buf`.
            #[allow(unsafe_code)]
            unsafe {
                debug_assert!(dst_rem_pos + dst_rem_len <= buf.len());
                debug_assert!(dst_rem_len <= tmp_out.len());
                debug_assert!(dlen <= buf.len());

                core::ptr::copy_nonoverlapping(
                    tmp_out.as_ptr(),
                    buf.as_mut_ptr().add(dst_rem_pos),
                    dst_rem_len,
                );
                Ok(buf.get_unchecked(..dlen))
            }
        } else {
            Err(InvalidEncodingError)
        }
    }

    fn encoded_len(bytes: &[u8]) -> usize {
        encoded_len_inner(bytes.len(), T::PADDED).unwrap_or(0)
    }