ext/node/ops/buffer.rs - main - githubmirror/deno

ext/node/ops/buffer.rs
483 строки14 KB
半岛的蒟蒻bddjr
fix(ext/node): fix asciiSlice & ucs2Slice & hexSlice being too slow (#34525)
09 июн 2026, 11:01
09 июн 2026, 11:015fe5b77
// Copyright 2018-2026 the Deno authors. MIT license.

use deno_core::convert::Uint8Array;
use deno_core::op2;
use deno_core::v8;
use deno_core::v8_static_strings;
use deno_error::JsErrorBox;

#[op2(fast)]
pub fn op_mark_as_untransferable(
  scope: &mut v8::PinScope<'_, '_>,
  ab: v8::Local<v8::ArrayBuffer>,
) {
  v8_static_strings! {
      UNTRANSFERABLE = "untransferable",
  }

  let key = UNTRANSFERABLE.v8_string(scope).unwrap();
  ab.set_detach_key(key.into());
}

#[op2(fast)]
pub fn op_is_ascii(#[buffer] buf: &[u8]) -> bool {
  buf.is_ascii()
}

#[op2(fast)]
pub fn op_is_utf8(#[buffer] buf: &[u8]) -> bool {
  std::str::from_utf8(buf).is_ok()
}

#[op2]
pub fn op_transcode(
  #[buffer] source: &[u8],
  #[string] from_encoding: &str,
  #[string] to_encoding: &str,
) -> Result<Uint8Array, JsErrorBox> {
  match (from_encoding, to_encoding) {
    ("utf8", "ascii") => Ok(utf8_to_ascii(source)),
    ("utf8", "latin1") => Ok(utf8_to_latin1(source)),
    ("utf8", "utf16le") => utf8_to_utf16le(source),
    ("utf16le", "utf8") => utf16le_to_utf8(source),
    ("latin1", "utf16le") | ("ascii", "utf16le") => {
      Ok(latin1_ascii_to_utf16le(source))
    }
    (from, to) => Err(JsErrorBox::generic(format!(
      "Unable to transcode Buffer {from}->{to}"
    ))),
  }
}

fn latin1_ascii_to_utf16le(source: &[u8]) -> Uint8Array {
  let mut result = Vec::with_capacity(source.len() * 2);
  for &byte in source {
    result.push(byte);
    result.push(0);
  }
  result.into()
}

fn utf16le_to_utf8(source: &[u8]) -> Result<Uint8Array, JsErrorBox> {
  let ucs2_vec: Vec<u16> = source
    .chunks_exact(2)
    .map(|chunk| u16::from_le_bytes([chunk[0], chunk[1]]))
    .collect();
  String::from_utf16(&ucs2_vec)
    .map(|utf8_string| utf8_string.into_bytes().into())
    .map_err(|e| JsErrorBox::generic(format!("Invalid UTF-16 sequence: {}", e)))
}

fn utf8_to_utf16le(source: &[u8]) -> Result<Uint8Array, JsErrorBox> {
  let utf8_string =
    std::str::from_utf8(source).map_err(JsErrorBox::from_err)?;
  let ucs2_vec: Vec<u16> = utf8_string.encode_utf16().collect();
  let bytes: Vec<u8> = ucs2_vec.iter().flat_map(|&x| x.to_le_bytes()).collect();
  Ok(bytes.into())
}

fn utf8_to_latin1(source: &[u8]) -> Uint8Array {
  let mut latin1_bytes = Vec::with_capacity(source.len());
  let mut i = 0;
  while i < source.len() {
    match source[i] {
      byte if byte <= 0x7F => {
        // ASCII character
        latin1_bytes.push(byte);
        i += 1;
      }
      byte if (0xC2..=0xDF).contains(&byte) && i + 1 < source.len() => {
        // 2-byte UTF-8 sequence
        let codepoint =
          ((byte as u16 & 0x1F) << 6) | (source[i + 1] as u16 & 0x3F);
        latin1_bytes.push(if codepoint <= 0xFF {
          codepoint as u8
        } else {
          b'?'
        });
        i += 2;
      }
      _ => {
        // 3-byte or 4-byte UTF-8 sequence, or invalid UTF-8
        latin1_bytes.push(b'?');
        // Skip to the next valid UTF-8 start byte
        i += 1;
        while i < source.len() && (source[i] & 0xC0) == 0x80 {
          i += 1;
        }
      }
    }
  }
  latin1_bytes.into()
}

fn utf8_to_ascii(source: &[u8]) -> Uint8Array {
  let mut ascii_bytes = Vec::with_capacity(source.len());
  let mut i = 0;
  while i < source.len() {
    match source[i] {
      byte if byte <= 0x7F => {
        // ASCII character
        ascii_bytes.push(byte);
        i += 1;
      }
      _ => {
        // Non-ASCII character
        ascii_bytes.push(b'?');
        // Skip to the next valid UTF-8 start byte
        i += 1;
        while i < source.len() && (source[i] & 0xC0) == 0x80 {
          i += 1;
        }
      }
    }
  }
  ascii_bytes.into()
}

#[op2(fast)]
#[smi]
pub fn op_node_buffer_compare(
  #[buffer] buf1: &[u8],
  #[buffer] buf2: &[u8],
) -> i32 {
  buf1.cmp(buf2) as i32
}

#[op2(fast)]
#[smi]
pub fn op_node_buffer_compare_offset(
  #[buffer] source: &[u8],
  #[buffer] target: &[u8],
  #[smi] source_start: usize,
  #[smi] target_start: usize,
  #[smi] source_end: usize,
  #[smi] target_end: usize,
) -> Result<i32, JsErrorBox> {
  if source_start > source.len() {
    return Err(JsErrorBox::from_err(BufferError::OutOfRangeNamed(
      "sourceStart".to_string(),
    )));
  }
  if target_start > target.len() {
    return Err(JsErrorBox::from_err(BufferError::OutOfRangeNamed(
      "targetStart".to_string(),
    )));
  }

  if source_start > source_end {
    panic!("source_start > source_end");
  }
  if target_start > target_end {
    panic!("target_start > target_end");
  }

  Ok(
    source[source_start..source_end].cmp(&target[target_start..target_end])
      as i32,
  )
}

// Threshold for falling back to V8's internal string copy allocation
// instead of creating an ExternalString to reduce GC finalizer overhead.
const ZERO_COPY_THRESHOLD: usize = 1024;

#[op2(reentrant)]
pub fn op_node_encoding_slice<'a>(
  scope: &mut v8::PinScope<'a, '_>,
  buf: v8::Local<v8::ArrayBufferView>,
  start: v8::Local<v8::Value>,
  end: v8::Local<v8::Value>,
  encoding: u8,
) -> Result<v8::Local<'a, v8::String>, JsErrorBox> {
  let buf_len = buf.byte_length();

  let start =
    parse_array_index(scope, start, 0).map_err(JsErrorBox::from_err)?;
  let mut end =
    parse_array_index(scope, end, buf_len).map_err(JsErrorBox::from_err)?;

  let mut storage = [0; v8::TYPED_ARRAY_MAX_SIZE_IN_HEAP];
  let buf = buf.get_contents(&mut storage);

  if end < start {
    end = start;
  }

  if end > buf.len() {
    return Err(JsErrorBox::from_err(BufferError::OutOfRange));
  }

  if end == start {
    return Ok(v8::String::empty(scope));
  }

  let buffer = &buf[start..end];

  match encoding {
    0 => {
      // utf8Slice
      if buffer.len() <= 256 && buffer.is_ascii() {
        v8::String::new_from_one_byte(scope, buffer, v8::NewStringType::Normal)
      } else {
        // `v8::String::new_from_utf8` does not replace ill-formed UTF-8
        // sequences the same way Node.js does, so decode lossily ourselves.
        // Rust's `from_utf8_lossy` follows the WHATWG "maximal subpart"
        // replacement (one U+FFFD per ill-formed subsequence), matching Node's
        // `Buffer.prototype.toString('utf8')` and `string_decoder`. For valid
        // UTF-8 it borrows the input without allocating.
        let decoded = String::from_utf8_lossy(buffer);
        v8::String::new_from_utf8(
          scope,
          decoded.as_bytes(),
          v8::NewStringType::Normal,
        )
      }
    }
    1 => {
      // latin1Slice
      v8::String::new_from_one_byte(scope, buffer, v8::NewStringType::Normal)
    }
    2 => {
      // asciiSlice
      if buffer.len() > v8::String::MAX_LENGTH {
        // String too long
        None
      } else if buffer.len() > ZERO_COPY_THRESHOLD {
        let ascii_bytes = mask_ascii_fast(buffer);
        // `ascii_bytes` is already a copied clone
        // (not a zero‑copy reference to the ArrayBufferView),
        // so we can zero‑copy create a V8 string from it.
        v8::String::new_external_onebyte(scope, ascii_bytes.into_boxed_slice())
      } else if buffer.is_ascii() {
        // A copy is required to prevent subsequent ArrayBufferView modifications
        // from altering the immutable string.
        // Cannot zero-copy create a V8 string here.
        v8::String::new_from_one_byte(scope, buffer, v8::NewStringType::Normal)
      } else {
        let ascii_bytes = mask_ascii_fast(buffer);
        // Copy bytes to a string
        v8::String::new_from_one_byte(
          scope,
          &ascii_bytes,
          v8::NewStringType::Normal,
        )
      }
    }
    3 => {
      // ucs2Slice
      decode_utf16le_from_bytes(scope, buffer)
    }
    4 => {
      // hexSlice
      if buffer.len() > (v8::String::MAX_LENGTH / 2) {
        // String too long
        None
      } else {
        let target_len = buffer.len() * 2;
        let mut hex_bytes = vec![0u8; target_len];
        // infallible: output is exactly 2x input
        faster_hex::hex_encode(buffer, &mut hex_bytes).unwrap();
        if target_len <= ZERO_COPY_THRESHOLD {
          // Copy bytes to a string
          v8::String::new_from_one_byte(
            scope,
            &hex_bytes,
            v8::NewStringType::Normal,
          )
        } else {
          // Create a V8 string with zero-copy
          v8::String::new_external_onebyte(scope, hex_bytes.into_boxed_slice())
        }
      }
    }
    _ => return Err(JsErrorBox::from_err(BufferError::InvalidType)),
  }
  .ok_or_else(|| JsErrorBox::from_err(BufferError::StringTooLong))
}

#[inline(always)]
fn mask_ascii_fast(bytes: &[u8]) -> Vec<u8> {
  const CHUNK_SIZE: usize = std::mem::size_of::<usize>();
  const MASK: usize = usize::from_ne_bytes([0x7F; CHUNK_SIZE]);

  let len = bytes.len();
  let mut ascii_bytes = Vec::<u8>::with_capacity(len);

  let src = bytes.as_ptr();
  let dst = ascii_bytes.as_mut_ptr();

  // SAFETY:
  // 1. Bounds & Capacity:
  //    - `src` is valid for `len` bytes.
  //    - `dst` has an allocated capacity of `len` bytes.
  //    - If `len >= CHUNK_SIZE`: `i < limit` implies
  //      `i + CHUNK_SIZE < len`. The out-of-loop block at `limit`
  //      accesses exactly the last `CHUNK_SIZE` bytes.
  //    - If `len < CHUNK_SIZE`: The `for` loop bounds are `0..len`.
  //    Therefore, all pointer arithmetic stays within valid bounds.
  // 2. Alignment:
  //    `read_unaligned` and `write_unaligned` are used for `usize`
  //    accesses, preventing UB from potentially unaligned pointers.
  // 3. Initialization:
  //    Every byte from `0` to `len` in `dst` is guaranteed to be
  //    written before `set_len(len)` is called. Overlapping writes
  //    are idempotent and safe.
  unsafe {
    if len >= CHUNK_SIZE {
      let limit = len - CHUNK_SIZE;
      let mut i: usize = 0;
      while i < limit {
        let tmp = src.add(i).cast::<usize>().read_unaligned();
        dst.add(i).cast::<usize>().write_unaligned(tmp & MASK);
        i += CHUNK_SIZE;
      }
      let tmp = src.add(limit).cast::<usize>().read_unaligned();
      dst.add(limit).cast::<usize>().write_unaligned(tmp & MASK);
    } else {
      for i in 0..len {
        dst.add(i).write(src.add(i).read() & 0x7F);
      }
    }

    ascii_bytes.set_len(len);
  }

  ascii_bytes
}

#[inline(always)]
fn decode_utf16le_from_bytes<'a>(
  scope: &mut v8::PinScope<'a, '_>,
  bytes: &[u8],
) -> Option<v8::Local<'a, v8::String>> {
  // UTF-16 must be a multiple of 2 bytes. Discard any trailing odd byte.
  let len = bytes.len() & !1;
  let target_len = len / 2;

  if target_len > v8::String::MAX_LENGTH {
    // String too long
    return None;
  }

  let buf = &bytes[..len];

  #[cfg(target_endian = "little")]
  {
    // Attempt a zero-copy cast to &[u16]
    // SAFETY:
    // `u16` has no invalid bit patterns. Reinterpreting
    // any initialized `u8` pairs as `u16` is safe.
    let (prefix, u16_slice, suffix) = unsafe { buf.align_to::<u16>() };

    if prefix.is_empty() && suffix.is_empty() {
      // Fast path: Memory is perfectly 2-byte aligned.
      // A copy is required to prevent subsequent ArrayBufferView modifications
      // from altering the immutable string.
      // Cannot zero-copy create a V8 string here.
      v8::String::new_from_two_byte(scope, u16_slice, v8::NewStringType::Normal)
    } else {
      // Slow path: Unaligned memory (rare in V8, but must be handled).
      // Use uninitialized memory to avoid Vec's memset(0) overhead.
      let mut u16_data = Vec::<u16>::with_capacity(target_len);

      // SAFETY:
      // 1. `buf` is valid for reads of `len` bytes.
      // 2. `u16_data` has a capacity of `target_len`
      //    `u16`s (exactly `len` bytes), so writing
      //    `len` bytes is within bounds.
      // 3. Source and destination do not overlap.
      // 4. `copy_nonoverlapping` fully initializes
      //    the memory, making `set_len` safe.
      unsafe {
        // Memcpy the data byte-by-byte into the newly allocated Vec memory.
        std::ptr::copy_nonoverlapping(
          buf.as_ptr(),
          u16_data.as_mut_ptr().cast::<u8>(),
          len,
        );
        // Manually set the length.
        u16_data.set_len(target_len);
      }

      // `u16_data` is already a copied clone
      // (not a zero‑copy reference to the ArrayBufferView),
      // so we can zero‑copy create a V8 string from it.
      if len <= ZERO_COPY_THRESHOLD {
        // Copy bytes to a string
        v8::String::new_from_two_byte(
          scope,
          &u16_data,
          v8::NewStringType::Normal,
        )
      } else {
        // Create a V8 string with zero-copy
        v8::String::new_external_twobyte(scope, u16_data.into_boxed_slice())
      }
    }
  }

  // Fallback for big-endian architectures (uncommon environments).
  #[cfg(target_endian = "big")]
  {
    let u16_data = buf
      .chunks_exact(2)
      .map(|chunk| u16::from_le_bytes([chunk[0], chunk[1]]))
      .collect();

    // `u16_data` is already a copied clone
    // (not a zero‑copy reference to the ArrayBufferView),
    // so we can zero‑copy create a V8 string from it.
    if len <= ZERO_COPY_THRESHOLD {
      // Copy bytes to a string
      v8::String::new_from_two_byte(scope, &u16_data, v8::NewStringType::Normal)
    } else {
      // Create a V8 string with zero-copy
      v8::String::new_external_twobyte(scope, u16_data.into_boxed_slice())
    }
  }
}

#[derive(Debug, thiserror::Error, deno_error::JsError)]
enum BufferError {
  #[error(
    "Cannot create a string longer than 0x{:x} characters",
    v8::String::MAX_LENGTH
  )]
  #[class(generic)]
  #[property("code" = "ERR_STRING_TOO_LONG")]
  StringTooLong,
  #[error("Invalid type")]
  #[class(generic)]
  InvalidType,
  #[error("Index out of range")]
  #[class(range)]
  #[property("code" = "ERR_OUT_OF_RANGE")]
  OutOfRange,
  #[error("The value of \"{0}\" is out of range.")]
  #[class(range)]
  #[property("code" = "ERR_OUT_OF_RANGE")]
  OutOfRangeNamed(String),
}

#[inline(always)]
fn parse_array_index(
  scope: &mut v8::PinScope<'_, '_>,
  arg: v8::Local<v8::Value>,
  default: usize,
) -> Result<usize, BufferError> {
  if arg.is_undefined() {
    return Ok(default);
  }

  let Some(arg) = arg.integer_value(scope) else {
    return Err(BufferError::InvalidType);
  };
  if arg < 0 {
    return Err(BufferError::OutOfRange);
  }
  if arg > isize::MAX as i64 {
    return Err(BufferError::OutOfRange);
  }
  Ok(arg as usize)
}