buffer: speed up swap16/32, add swap64

* Speed up buffer.swap16 and swap32 by using builtins. Up to ~6x gain.
  Drop transition point between JS and C++ implementations accordingly.
  Amount of performance improvement not only depends on buffer size but
  also memory alignment.
* Fix tests: C++ impl tests were testing 0-filled buffers so were
  always passing.
* Add similar buffer.swap64 method.
* Make buffer-swap benchmark mirror JS impl.

doc/api/buffer.markdown has an entry of "added: REPLACEME" that should
be changed to the correct release number before tagged.

Because node is currently using a very old version of cpplint.py it
doesn't know that std::swap() has moved from <algorithm> to <utility> in
c++11. So until cpplint.py is updated simply NOLINT the line.
Technically it should be NOLINT(build/include_what_you_use), but that
puts the line over 80 characters causing another lint error.

PR-URL: nodejs#7157
Reviewed-By: Trevor Norris <trev.norris@gmail.com>
Reviewed-By: James M Snell <jasnell@gmail.com>

benchmark/buffers/buffer-swap.js

@@ -1,61 +1,90 @@
 'use strict';
 
 const common = require('../common.js');
+const v8 = require('v8');
 
 const bench = common.createBenchmark(main, {
-  method: ['swap16', 'swap32', 'htons', 'htonl'],
-  len: [4, 64, 512, 768, 1024, 1536, 2056, 4096, 8192],
-  n: [1e6]
+  aligned: ['true', 'false'],
+  method: ['swap16', 'swap32', 'swap64'/*, 'htons', 'htonl', 'htonll'*/],
+  len: [8, 64, 128, 256, 512, 768, 1024, 1536, 2056, 4096, 8192],
+  n: [5e7]
 });
 
 // The htons and htonl methods below are used to benchmark the
 // performance difference between doing the byteswap in pure
 // javascript regardless of Buffer size as opposed to dropping
-// down to the native layer for larger Buffer sizes.
+// down to the native layer for larger Buffer sizes. Commented
+// out by default because they are slow for big buffers. If
+// re-evaluating the crossover point, uncomment those methods
+// and comment out their implementations in lib/buffer.js so
+// C++ version will always be used.
+
+function swap(b, n, m) {
+  const i = b[n];
+  b[n] = b[m];
+  b[m] = i;
+}
 
 Buffer.prototype.htons = function htons() {
   if (this.length % 2 !== 0)
     throw new RangeError();
-  for (var i = 0, n = 0; i < this.length; i += 2) {
-    n = this[i];
-    this[i] = this[i + 1];
-    this[i + 1] = n;
+  for (var i = 0; i < this.length; i += 2) {
+    swap(this, i, i + 1);
   }
   return this;
 };
 
 Buffer.prototype.htonl = function htonl() {
-  if (this.length % 2 !== 0)
+  if (this.length % 4 !== 0)
+    throw new RangeError();
+  for (var i = 0; i < this.length; i += 4) {
+    swap(this, i, i + 3);
+    swap(this, i + 1, i + 2);
+  }
+  return this;
+};
+
+Buffer.prototype.htonll = function htonl() {
+  if (this.length % 8 !== 0)
     throw new RangeError();
-  for (var i = 0, n = 0; i < this.length; i += 4) {
-    n = this[i];
-    this[i] = this[i + 3];
-    this[i + 3] = n;
-    n = this[i + 1];
-    this[i + 1] = this[i + 2];
-    this[i + 2] = n;
+  for (var i = 0; i < this.length; i += 8) {
+    swap(this, i, i + 7);
+    swap(this, i + 1, i + 6);
+    swap(this, i + 2, i + 5);
+    swap(this, i + 3, i + 4);
   }
   return this;
 };
 
-function createBuffer(len) {
+function createBuffer(len, aligned) {
+  len += aligned ? 0 : 1;
   const buf = Buffer.allocUnsafe(len);
   for (var i = 1; i <= len; i++)
     buf[i - 1] = i;
-  return buf;
+  return aligned ? buf : buf.slice(1);
 }
 
-function bufferSwap(n, buf, method) {
-  for (var i = 1; i <= n; i++)
-    buf[method]();
+function genMethod(method) {
+  const fnString =
+      'return function ' + method + '(n, buf) {' +
+      '  for (var i = 0; i <= n; i++)' +
+      '    buf.' + method + '();' +
+      '}';
+  return (new Function(fnString))();
 }
 
 function main(conf) {
   const method = conf.method;
   const len = conf.len | 0;
   const n = conf.n | 0;
-  const buf = createBuffer(len);
+  const aligned = conf.aligned || 'true';
+  const buf = createBuffer(len, aligned === 'true');
+  const bufferSwap = genMethod(method);
+
+  v8.setFlagsFromString('--allow_natives_syntax');
+  eval('%OptimizeFunctionOnNextCall(bufferSwap)');
+
   bench.start();
-  bufferSwap(n, buf, method);
+  bufferSwap(n, buf);
   bench.end(n);
 }

doc/api/buffer.md

@@ -1461,10 +1461,10 @@ calls can be chained.
 ```js
 const buf = Buffer.from([0x1, 0x2, 0x3, 0x4, 0x5, 0x6, 0x7, 0x8]);
 console.log(buf);
-  // Prints Buffer(0x1, 0x2, 0x3, 0x4, 0x5, 0x6, 0x7, 0x8)
+  // Prints <Buffer 01 02 03 04 05 06 07 08>
 buf.swap16();
 console.log(buf);
-  // Prints Buffer(0x2, 0x1, 0x4, 0x3, 0x6, 0x5, 0x8, 0x7)
+  // Prints <Buffer 02 01 04 03 06 05 08 07>
 ```
 
 ### buf.swap32()
@@ -1482,12 +1482,36 @@ calls can be chained.
 ```js
 const buf = Buffer.from([0x1, 0x2, 0x3, 0x4, 0x5, 0x6, 0x7, 0x8]);
 console.log(buf);
-  // Prints Buffer(0x1, 0x2, 0x3, 0x4, 0x5, 0x6, 0x7, 0x8)
+  // Prints <Buffer 01 02 03 04 05 06 07 08>
 buf.swap32();
 console.log(buf);
-  // Prints Buffer(0x4, 0x3, 0x2, 0x1, 0x8, 0x7, 0x6, 0x5)
+  // Prints <Buffer 04 03 02 01 08 07 06 05>
 ```
 
+### buf.swap64()
+<!-- YAML
+added: REPLACEME
+-->
+
+* Return: {Buffer}
+
+Interprets the `Buffer` as an array of 64-bit numbers and swaps
+the byte-order *in-place*. Throws a `RangeError` if the `Buffer` length is
+not a multiple of 64 bits. The method returns a reference to the Buffer, so
+calls can be chained.
+
+```js
+const buf = Buffer.from([0x1, 0x2, 0x3, 0x4, 0x5, 0x6, 0x7, 0x8]);
+console.log(buf);
+  // Prints <Buffer 01 02 03 04 05 06 07 08>
+buf.swap64();
+console.log(buf);
+  // Prints <Buffer 08 07 06 05 04 03 02 01>
+```
+
+Note that JavaScript cannot encode 64-bit integers. This method is intended
+for working with 64-bit floats.
+
 ### buf.toString([encoding[, start[, end]]])
 
 * `encoding` {String} Default: `'utf8'`

lib/buffer.js

@@ -24,47 +24,6 @@ var poolSize, poolOffset, allocPool;
 
 binding.setupBufferJS(Buffer.prototype, bindingObj);
 
-const swap16n = binding.swap16;
-const swap32n = binding.swap32;
-
-function swap(b, n, m) {
-  const i = b[n];
-  b[n] = b[m];
-  b[m] = i;
-}
-
-Buffer.prototype.swap16 = function swap16() {
-  // For Buffer.length < 512, it's generally faster to
-  // do the swap in javascript. For larger buffers,
-  // dropping down to the native code is faster.
-  const len = this.length;
-  if (len % 2 !== 0)
-    throw new RangeError('Buffer size must be a multiple of 16-bits');
-  if (len < 512) {
-    for (var i = 0; i < len; i += 2)
-      swap(this, i, i + 1);
-    return this;
-  }
-  return swap16n.apply(this);
-};
-
-Buffer.prototype.swap32 = function swap32() {
-  // For Buffer.length < 1024, it's generally faster to
-  // do the swap in javascript. For larger buffers,
-  // dropping down to the native code is faster.
-  const len = this.length;
-  if (len % 4 !== 0)
-    throw new RangeError('Buffer size must be a multiple of 32-bits');
-  if (len < 1024) {
-    for (var i = 0; i < len; i += 4) {
-      swap(this, i, i + 3);
-      swap(this, i + 1, i + 2);
-    }
-    return this;
-  }
-  return swap32n.apply(this);
-};
-
 const flags = bindingObj.flags;
 const kNoZeroFill = 0;
 
@@ -1320,3 +1279,67 @@ Buffer.prototype.writeDoubleBE = function writeDoubleBE(val, offset, noAssert) {
     binding.writeDoubleBE(this, val, offset, true);
   return offset + 8;
 };
+
+const swap16n = binding.swap16;
+const swap32n = binding.swap32;
+const swap64n = binding.swap64;
+
+function swap(b, n, m) {
+  const i = b[n];
+  b[n] = b[m];
+  b[m] = i;
+}
+
+
+Buffer.prototype.swap16 = function swap16() {
+  // For Buffer.length < 128, it's generally faster to
+  // do the swap in javascript. For larger buffers,
+  // dropping down to the native code is faster.
+  const len = this.length;
+  if (len % 2 !== 0)
+    throw new RangeError('Buffer size must be a multiple of 16-bits');
+  if (len < 128) {
+    for (var i = 0; i < len; i += 2)
+      swap(this, i, i + 1);
+    return this;
+  }
+  return swap16n(this);
+};
+
+
+Buffer.prototype.swap32 = function swap32() {
+  // For Buffer.length < 192, it's generally faster to
+  // do the swap in javascript. For larger buffers,
+  // dropping down to the native code is faster.
+  const len = this.length;
+  if (len % 4 !== 0)
+    throw new RangeError('Buffer size must be a multiple of 32-bits');
+  if (len < 192) {
+    for (var i = 0; i < len; i += 4) {
+      swap(this, i, i + 3);
+      swap(this, i + 1, i + 2);
+    }
+    return this;
+  }
+  return swap32n(this);
+};
+
+
+Buffer.prototype.swap64 = function swap64() {
+  // For Buffer.length < 192, it's generally faster to
+  // do the swap in javascript. For larger buffers,
+  // dropping down to the native code is faster.
+  const len = this.length;
+  if (len % 8 !== 0)
+    throw new RangeError('Buffer size must be a multiple of 64-bits');
+  if (len < 192) {
+    for (var i = 0; i < len; i += 8) {
+      swap(this, i, i + 7);
+      swap(this, i + 1, i + 6);
+      swap(this, i + 2, i + 5);
+      swap(this, i + 3, i + 4);
+    }
+    return this;
+  }
+  return swap64n(this);
+};