diff --git a/src/node_buffer.cc b/src/node_buffer.cc
index af53c98efbd3c8..21960f88528d3a 100644
--- a/src/node_buffer.cc
+++ b/src/node_buffer.cc
@@ -28,6 +28,7 @@
 
 #include <assert.h>
 #include <string.h> // memcpy
+#include <limits.h>
 
 #define MIN(a,b) ((a) < (b) ? (a) : (b))
 
@@ -247,7 +248,7 @@ Handle<Value> Buffer::BinarySlice(const Arguments &args) {
 }
 
 
-static bool contains_non_ascii(const char* buf, size_t len) {
+static bool contains_non_ascii_slow(const char* buf, size_t len) {
   for (size_t i = 0; i < len; ++i) {
     if (buf[i] & 0x80) return true;
   }
@@ -255,13 +256,100 @@ static bool contains_non_ascii(const char* buf, size_t len) {
 }
 
 
-static void force_ascii(const char* src, char* dst, size_t len) {
+static bool contains_non_ascii(const char* src, size_t len) {
+  if (len < 16) {
+    return contains_non_ascii_slow(src, len);
+  }
+
+  const unsigned bytes_per_word = BITS_PER_LONG / CHAR_BIT;
+  const unsigned align_mask = bytes_per_word - 1;
+  const unsigned unaligned = reinterpret_cast<uintptr_t>(src) & align_mask;
+
+  if (unaligned > 0) {
+    const unsigned n = bytes_per_word - unaligned;
+    if (contains_non_ascii_slow(src, n)) return true;
+    src += n;
+    len -= n;
+  }
+
+#if BITS_PER_LONG == 64
+  typedef uint64_t word;
+  const uint64_t mask = 0x8080808080808080ll;
+#else
+  typedef uint32_t word;
+  const uint32_t mask = 0x80808080l;
+#endif
+
+  const word* srcw = reinterpret_cast<const word*>(src);
+
+  for (size_t i = 0, n = len / bytes_per_word; i < n; ++i) {
+    if (srcw[i] & mask) return true;
+  }
+
+  const unsigned remainder = len & align_mask;
+  if (remainder > 0) {
+    const size_t offset = len - remainder;
+    if (contains_non_ascii_slow(src + offset, remainder)) return true;
+  }
+
+  return false;
+}
+
+
+static void force_ascii_slow(const char* src, char* dst, size_t len) {
   for (size_t i = 0; i < len; ++i) {
     dst[i] = src[i] & 0x7f;
   }
 }
 
 
+static void force_ascii(const char* src, char* dst, size_t len) {
+  if (len < 16) {
+    force_ascii_slow(src, dst, len);
+    return;
+  }
+
+  const unsigned bytes_per_word = BITS_PER_LONG / CHAR_BIT;
+  const unsigned align_mask = bytes_per_word - 1;
+  const unsigned src_unalign = reinterpret_cast<uintptr_t>(src) & align_mask;
+  const unsigned dst_unalign = reinterpret_cast<uintptr_t>(dst) & align_mask;
+
+  if (src_unalign > 0) {
+    if (src_unalign == dst_unalign) {
+      const unsigned unalign = bytes_per_word - src_unalign;
+      force_ascii_slow(src, dst, unalign);
+      src += unalign;
+      dst += unalign;
+      len -= src_unalign;
+    } else {
+      force_ascii_slow(src, dst, len);
+      return;
+    }
+  }
+
+#if BITS_PER_LONG == 64
+  typedef uint64_t word;
+  const uint64_t mask = ~0x8080808080808080ll;
+#else
+  typedef uint32_t word;
+  const uint32_t mask = ~0x80808080l;
+#endif
+
+  const word* srcw = reinterpret_cast<const word*>(src);
+  word* dstw = reinterpret_cast<word*>(dst);
+
+  for (size_t i = 0, n = len / bytes_per_word; i < n; ++i) {
+    dstw[i] = srcw[i] & mask;
+  }
+
+  const unsigned remainder = len & align_mask;
+  if (remainder > 0) {
+    const size_t offset = len - remainder;
+    force_ascii_slow(src + offset, dst + offset, remainder);
+  }
+}
+
+
 Handle<Value> Buffer::AsciiSlice(const Arguments &args) {
   HandleScope scope;
   Buffer *parent = ObjectWrap::Unwrap<Buffer>(args.This());
diff --git a/src/node_internals.h b/src/node_internals.h
index 0502ec90586e42..10b1742fb54f23 100644
--- a/src/node_internals.h
+++ b/src/node_internals.h
@@ -48,6 +48,12 @@ inline static int snprintf(char* buf, unsigned int len, const char* fmt, ...) {
 }
 #endif
 
+#if defined(__x86_64__)
+# define BITS_PER_LONG 64
+#else
+# define BITS_PER_LONG 32
+#endif
+
 #ifndef offset_of
 // g++ in strict mode complains loudly about the system offsetof() macro
 // because it uses NULL as the base address.
diff --git a/test/simple/test-buffer-ascii.js b/test/simple/test-buffer-ascii.js
index a741a3db1e5ac9..784597a48e40ed 100644
--- a/test/simple/test-buffer-ascii.js
+++ b/test/simple/test-buffer-ascii.js
@@ -25,3 +25,21 @@ var assert = require('assert');
 // ASCII conversion in node.js simply masks off the high bits,
 // it doesn't do transliteration.
 assert.equal(Buffer('hérité').toString('ascii'), 'hC)ritC)');
+
+// 71 characters, 78 bytes. The ’ character is a triple-byte sequence.
+var input = 'C’est, graphiquement, la réunion d’un accent aigu ' +
+            'et d’un accent grave.';
+
+var expected = 'Cb\u0000\u0019est, graphiquement, la rC)union ' +
+               'db\u0000\u0019un accent aigu et db\u0000\u0019un ' +
+               'accent grave.';
+
+var buf = Buffer(input);
+
+for (var i = 0; i < expected.length; ++i) {
+  assert.equal(buf.slice(i).toString('ascii'), expected.slice(i));
+
+  // Skip remainder of multi-byte sequence.
+  if (input.charCodeAt(i) > 65535) ++i;
+  if (input.charCodeAt(i) > 127) ++i;
+}