Bug 696211 - Align ChunkInfo by inserting padding in Chunk; r=billm

We get a 2% speedup on EarlyBoyer if ChunkInfo (hot all over the
GC allocator paths) is not split across a cache line.  An easy
and guaranteed effective way to do this is to just pad Chunk out
to the full 1MiB allocation.  This makes ChunkInfo abut the end
of the 1MiB allocation, rather than whereever the Arenas and
ChunkBitmap happen to end.  Since GC Chunks are aligned at 1MiB
address boundaries, this ensures that ChunkInfo is inside of a
cache line.

js/src/jsgc.h

@@ -615,12 +615,22 @@ struct ChunkBitmap {
 
 JS_STATIC_ASSERT(ArenaBitmapBytes * ArenasPerChunk == sizeof(ChunkBitmap));
 
+const size_t ChunkPadSize = GC_CHUNK_SIZE
+                            - (sizeof(Arena) * ArenasPerChunk)
+                            - sizeof(ChunkBitmap)
+                            - sizeof(ChunkInfo);
+JS_STATIC_ASSERT(ChunkPadSize < BytesPerArena);
+
 /*
  * Chunks contain arenas and associated data structures (mark bitmap, delayed
  * marking state).
  */
 struct Chunk {
     Arena           arenas[ArenasPerChunk];
+
+    /* Pad to full size to ensure cache alignment of ChunkInfo. */
+    uint8           padding[ChunkPadSize];
+
     ChunkBitmap     bitmap;
     ChunkInfo       info;
 
@@ -667,8 +677,7 @@ struct Chunk {
     inline void init();
 };
 
-JS_STATIC_ASSERT(sizeof(Chunk) <= GC_CHUNK_SIZE);
-JS_STATIC_ASSERT(sizeof(Chunk) + BytesPerArena > GC_CHUNK_SIZE);
+JS_STATIC_ASSERT(sizeof(Chunk) == GC_CHUNK_SIZE);
 
 class ChunkPool {
     Chunk   *emptyChunkListHead;