From 535fec83ea890775c31cbe041fc19db9c4b7ff1f Mon Sep 17 00:00:00 2001
From: Ben Noordhuis <info@bnoordhuis.nl>
Date: Tue, 9 Dec 2014 15:41:35 +0100
Subject: [PATCH] src: fix unaligned access in ucs2 string encoder

Seen with g++ 4.9.2 on x86_64 Linux: a SIGSEGV is generated when the
input to v8::String::NewFromTwoByte() is not suitably aligned.

g++ 4.9.2 emits SSE instructions for copy loops.  That requires aligned
input but that was something StringBytes::Encode() did not enforce until
now.  Make a properly aligned copy before handing off the input to V8.

We could, as an optimization, check that the pointer is aligned on a
two-byte boundary but that is technically still UB; pointers-to-char
are allowed to alias other pointers but the reverse is not true:
a pointer-to-uint16_t that aliases a pointer-to-char is in violation
of the pointer aliasing rules.

See https://code.google.com/p/v8/issues/detail?id=3694

Fixes segfaulting test simple/test-stream2-writable.

PR-URL: https://github.com/iojs/io.js/pull/127
Reviewed-by: Trevor Norris <trev.norris@gmail.com>
---
 src/string_bytes.cc | 38 +++++++++++++++++++-------------------
 1 file changed, 19 insertions(+), 19 deletions(-)

diff --git a/src/string_bytes.cc b/src/string_bytes.cc
index 3e79ce3990..f9892405f3 100644
--- a/src/string_bytes.cc
+++ b/src/string_bytes.cc
@@ -748,28 +748,28 @@ Local<Value> StringBytes::Encode(Isolate* isolate,
     }
 
     case UCS2: {
-      const uint16_t* out = reinterpret_cast<const uint16_t*>(buf);
-      uint16_t* dst = nullptr;
-      if (IsBigEndian()) {
-        // Node's "ucs2" encoding expects LE character data inside a
-        // Buffer, so we need to reorder on BE platforms.  See
-        // http://nodejs.org/api/buffer.html regarding Node's "ucs2"
-        // encoding specification
-        dst = new uint16_t[buflen / 2];
-        for (size_t i = 0; i < buflen / 2; i++) {
-          dst[i] = (out[i] << 8) | (out[i] >> 8);
-        }
-        out = dst;
+      // Node's "ucs2" encoding expects LE character data inside a
+      // Buffer, so we need to reorder on BE platforms.  See
+      // http://nodejs.org/api/buffer.html regarding Node's "ucs2"
+      // encoding specification.  Note that we have to make a copy
+      // to avoid pointer aliasing and unaligned access, something
+      // that we can't guarantee by just reinterpret_casting |buf|.
+      size_t srclen = buflen / 2;
+      uint16_t* src = new uint16_t[srclen];
+      for (size_t i = 0, k = 0; i < srclen; i += 1, k += 2) {
+        const uint8_t lo = static_cast<uint8_t>(buf[k + 0]);
+        const uint8_t hi = static_cast<uint8_t>(buf[k + 1]);
+        src[i] = lo | hi << 8;
       }
-      if (buflen < EXTERN_APEX)
+      if (buflen < EXTERN_APEX) {
         val = String::NewFromTwoByte(isolate,
-                                     out,
+                                     src,
                                      String::kNormalString,
-                                     buflen / 2);
-      else
-        val = ExternTwoByteString::NewFromCopy(isolate, out, buflen / 2);
-      if (dst)
-        delete[] dst;
+                                     srclen);
+      } else {
+        val = ExternTwoByteString::NewFromCopy(isolate, src, srclen);
+      }
+      delete[] src;
       break;
     }