doubly-linked free-lists for thread caches and page heaps

third_party/tcmalloc/chromium/src/common.h

Index: third_party/tcmalloc/chromium/src/common.h
diff --git a/third_party/tcmalloc/chromium/src/common.h b/third_party/tcmalloc/chromium/src/common.h
index a3df8de02fd0056b750e096d0d74eb5477cde6bb..48afd9fa444db857efe1a3613bed3d11cfc01f32 100644
--- a/third_party/tcmalloc/chromium/src/common.h
+++ b/third_party/tcmalloc/chromium/src/common.h
@@ -1,4 +1,4 @@
-// Copyright (c) 2008, Google Inc.
+// Copyright (c) 2011, Google Inc.
 // All rights reserved.
 // 
 // Redistribution and use in source and binary forms, with or without
@@ -31,15 +31,18 @@
 // Author: Sanjay Ghemawat <opensource@google.com>
 //
 // Common definitions for tcmalloc code.
-
+ 

[jar (doing other things), 2011/08/20 02:40:30]

nit: remove trailing whitespace.

[bxx, 2011/08/24 00:19:24]

Done.

 #ifndef TCMALLOC_COMMON_H_
 #define TCMALLOC_COMMON_H_
 
 #include "config.h"
+

[jar (doing other things), 2011/08/20 02:40:30]

nit: remove extra newline.

[bxx, 2011/08/24 00:19:24]

Done.

 #include <stddef.h>                     // for size_t
 #ifdef HAVE_STDINT_H
 #include <stdint.h>                     // for uintptr_t, uint64_t
 #endif
+
+#include "free_list.h"  // for SIZE_CLASS macros
 #include "internal_logging.h"  // for ASSERT, etc
 
 // Type that can hold a page number
@@ -175,7 +178,8 @@ class SizeMap {
     const bool big = (s > kMaxSmallSize);
     const int add_amount = big ? (127 + (120<<7)) : 7;
     const int shift_amount = big ? 7 : 3;
-    return (s + add_amount) >> shift_amount;
+    const int size_class = (s + add_amount) >> shift_amount;
+    return  (size_class < MIN_SIZE_CLASS) ? MIN_SIZE_CLASS : size_class;

[jar (doing other things), 2011/08/20 02:40:30]

I assume this was added to preclude 8 byte allocations on a 64bit architecture. 


As written, it adds a perf(?) hit near what looks to be highly optimized code. 
My guess is that this was written expecting (hoping) to be inlined and they
wanted to stay really short and sweet.

I like that you avoided broader changes (minimalist!), but perhaps it could be
written so that an optimizer could remove it when not active.  Note that s and
the return value are both signed ints, so it is (I think) hard to tell when the
compiler needn't bother.  

Suggest you change line 182 to something like:
if (!AVOID_8_BYTE_CLASSES || size > 8)
  return size_class;
return 2;  // The smallest class that needs more than 8 bytes.

That would allow the compiler to discard the comparison unless the macro was
set, and would have the fast-path(?) for the common case of returning size class
directly.

I don't really know if this is perf critical... but it sure (to me) looks a bit
hand-tuned... and it is nice to have zero impact (guaranteed) when we are not
doing this twiddling.  We know we have some perf impact when we are active (more
broadly)... and we can tune that case better if need be in the future.

[bxx, 2011/08/24 00:19:24]

Done.

   }
 
   int NumMoveSize(size_t size);