hand-optimize Buffer constructor

The Buffer constructor is used pervasively throughout io.js, yet it was
one of the most unwieldy functions in core.  This commit breaks up the
constructor into several small functions in a way that makes V8 happy.

About 8-10% CPU time was attributed to the constructor function before
in buffer-heavy benchmarks.  That pretty much drops to zero now because
V8 can now easily inline it at the call site.  It shortens the running
time of the following simple benchmark by about 15%:

    for (var i = 0; i < 25e6; ++i) new Buffer(1);

And about 8% from this benchmark:

    for (var i = 0; i < 1e7; ++i) new Buffer('x', 'ucs2');

From
https://github.com/iojs/io.js/commit/bbf54a554ababa91ced243acb0d2721bc19
eb1dc

diff --git a/index.js b/index.js
index 76269d183b246d3cf90fd7e69b741f33595d597f..e086c8f756f1e55548a7b472541324f0b211439f 100644 (file)
--- a/index.js
+++ b/index.js
@@ -64,69 +64,147 @@ Buffer.TYPED_ARRAY_SUPPORT = (function () {
  * By augmenting the instances, we can avoid modifying the `Uint8Array`
  * prototype.
  */
-function Buffer (subject, encoding) {
-  var self = this
-  if (!(self instanceof Buffer)) return new Buffer(subject, encoding)
-
-  var type = typeof subject
-  var length
-
-  if (type === 'number') {
-    length = +subject
-  } else if (type === 'string') {
-    length = Buffer.byteLength(subject, encoding)
-  } else if (type === 'object' && subject !== null) {
-    // assume object is array-like
-    if (subject.type === 'Buffer' && isArray(subject.data)) subject = subject.data
-    length = +subject.length
-  } else {
+function Buffer (arg) {
+  if (!(this instanceof Buffer)) {
+    // Avoid going through an ArgumentsAdaptorTrampoline in the common case.
+    if (arguments.length > 1) return new Buffer(arg, arguments[1])
+    return new Buffer(arg)
+  }
+
+  this.length = 0
+  this.parent = undefined
+
+  // Common case.
+  if (typeof arg === 'number') {
+    return fromNumber(this, arg)
+  }
+
+  // Slightly less common case.
+  if (typeof arg === 'string') {
+    return fromString(this, arg, arguments.length > 1 ? arguments[1] : 'utf8')
+  }
+
+  // Unusual.
+  return fromObject(this, arg)
+}
+
+function fromNumber (that, length) {
+  that = allocate(that, length < 0 ? 0 : checked(length) | 0)
+  if (!Buffer.TYPED_ARRAY_SUPPORT) {
+    for (var i = 0; i < length; i++) {
+      that[i] = 0
+    }
+  }
+  return that
+}
+
+function fromString (that, string, encoding) {
+  if (typeof encoding !== 'string' || encoding === '') encoding = 'utf8'
+
+  // Assumption: byteLength() return value is always < kMaxLength.
+  var length = byteLength(string, encoding) | 0
+  that = allocate(that, length)
+
+  that.write(string, encoding) | 0
+  return that
+}
+
+function fromObject (that, object) {
+  if (Buffer.isBuffer(object)) return fromBuffer(that, object)
+
+  if (isArray(object)) return fromArray(that, object)
+
+  if (object == null) {
     throw new TypeError('must start with number, buffer, array or string')
   }
 
-  if (length > kMaxLength) {
-    throw new RangeError('Attempt to allocate Buffer larger than maximum size: 0x' +
-      kMaxLength.toString(16) + ' bytes')
+  if (object.buffer instanceof ArrayBuffer) return fromTypedArray(that, object)
+
+  if (object.length) return fromArrayLike(that, object)
+
+  return fromJsonObject(that, object)
+}
+
+function fromBuffer (that, buffer) {
+  var length = checked(buffer.length) | 0
+  that = allocate(that, length)
+  buffer.copy(that, 0, 0, length)
+  return that
+}
+
+function fromArray (that, array) {
+  var length = checked(array.length) | 0
+  that = allocate(that, length)
+  for (var i = 0; i < length; i += 1) {
+    that[i] = array[i] & 255
   }
+  return that
+}
 
-  if (length < 0) length = 0
-  else length >>>= 0 // coerce to uint32
+// Duplicate of fromArray() to keep fromArray() monomorphic.
+function fromTypedArray (that, array) {
+  var length = checked(array.length) | 0
+  that = allocate(that, length)
+  // Truncating the elements is probably not what people expect from typed
+  // arrays with BYTES_PER_ELEMENT > 1 but it's compatible with the behavior
+  // of the old Buffer constructor.
+  for (var i = 0; i < length; i += 1) {
+    that[i] = array[i] & 255
+  }
+  return that
+}
 
-  if (Buffer.TYPED_ARRAY_SUPPORT) {
-    // Preferred: Return an augmented `Uint8Array` instance for best performance
-    self = Buffer._augment(new Uint8Array(length)) // eslint-disable-line consistent-this
-  } else {
-    // Fallback: Return THIS instance of Buffer (created by `new`)
-    self.length = length
-    self._isBuffer = true
+function fromArrayLike (that, array) {
+  var length = checked(array.length) | 0
+  that = allocate(that, length)
+  for (var i = 0; i < length; i += 1) {
+    that[i] = array[i] & 255
   }
+  return that
+}
 
-  var i
-  if (Buffer.TYPED_ARRAY_SUPPORT && typeof subject.byteLength === 'number') {
-    // Speed optimization -- use set if we're copying from a typed array
-    self._set(subject)
-  } else if (isArrayish(subject)) {
-    // Treat array-ish objects as a byte array
-    if (Buffer.isBuffer(subject)) {
-      for (i = 0; i < length; i++) {
-        self[i] = subject.readUInt8(i)
-      }
-    } else {
-      for (i = 0; i < length; i++) {
-        self[i] = ((subject[i] % 256) + 256) % 256
-      }
-    }
-  } else if (type === 'string') {
-    self.write(subject, 0, encoding)
-  } else if (type === 'number' && !Buffer.TYPED_ARRAY_SUPPORT) {
-    for (i = 0; i < length; i++) {
-      self[i] = 0
-    }
+// Deserialize { type: 'Buffer', data: [1,2,3,...] } into a Buffer object.
+// Returns a zero-length buffer for inputs that don't conform to the spec.
+function fromJsonObject (that, object) {
+  var array
+  var length = 0
+
+  if (object.type === 'Buffer' && isArray(object.data)) {
+    array = object.data
+    length = checked(array.length) | 0
+  }
+  that = allocate(that, length)
+
+  for (var i = 0; i < length; i += 1) {
+    that[i] = array[i] & 255
+  }
+  return that
+}
+
+function allocate (that, length) {
+  if (Buffer.TYPED_ARRAY_SUPPORT) {
+    // Return an augmented `Uint8Array` instance, for best performance
+    that = Buffer._augment(new Uint8Array(length))
+  } else {
+    // Fallback: Return an object instance of the Buffer class
+    that.length = length
+    that._isBuffer = true
   }
 
   var fromPool = length !== 0 && length <= Buffer.poolSize >>> 1
-  if (fromPool) self.parent = rootParent
+  if (fromPool) that.parent = rootParent
+
+  return that
+}
 
-  return self
+function checked (length) {
+  // Note: cannot use `length < kMaxLength` here because that fails when
+  // length is NaN (which is otherwise coerced to zero.)
+  if (length >= kMaxLength) {
+    throw new RangeError('Attempt to allocate Buffer larger than maximum ' +
+                         'size: 0x' + kMaxLength.toString(16) + ' bytes')
+  }
+  return length >>> 0
 }
 
 function SlowBuffer (subject, encoding) {
@@ -206,7 +284,7 @@ Buffer.concat = function concat (list, length) {
   return buf
 }
 
-Buffer.byteLength = function byteLength (string, encoding) {
+function byteLength (string, encoding) {
   if (typeof string !== 'string') string = String(string)
 
   if (string.length === 0) return 0
@@ -232,6 +310,7 @@ Buffer.byteLength = function byteLength (string, encoding) {
       return string.length
   }
 }
+Buffer.byteLength = byteLength
 
 // pre-set for values that may exist in the future
 Buffer.prototype.length = undefined
@@ -1196,12 +1275,6 @@ function stringtrim (str) {
   return str.replace(/^\s+|\s+$/g, '')
 }
 
-function isArrayish (subject) {
-  return isArray(subject) || Buffer.isBuffer(subject) ||
-      subject && typeof subject === 'object' &&
-      typeof subject.length === 'number'
-}
-
 function toHex (n) {
   if (n < 16) return '0' + n.toString(16)
   return n.toString(16)