Thu Nov 29 07:59:43 2007 Google Inc. <opensource@google.com>

	* google-perftools: version 0.94 release
	* PORTING: MinGW/Msys support -- runs same code as MSVC does (csilvers)
	* PORTING: Add NumCPUs support for Mac OS X (csilvers)
	* Work around a sscanf bug in glibc(?) (waldemar)
	* Fix Windows MSVC bug triggered by thread deletion (csilvers)
	* Fix bug that triggers in MSVC /O2: missing volatile (gpike)
	* March-of-time support: quiet warnings/errors for gcc 4.2, OS X 10.5
	* Modify pprof so it works without nm: useful for windows (csilvers)
	* pprof: Support filtering for CPU profiles (cgd)
	* Bugfix: have realloc report to hooks in all situations (maxim)
	* Speed improvement: replace slow memcpy with std::copy (soren)
	* Speed: better iterator efficiency in RecordRegionRemoval (soren)
	* Speed: minor speed improvements via better bitfield alignment (gpike)
	* Documentation: add documentation of binary profile output (cgd)


git-svn-id: http://gperftools.googlecode.com/svn/trunk@40 6b5cf1ce-ec42-a296-1ba9-69fdba395a50

1 files changed, 49 insertions, 0 deletions

diff --git a/src/stacktrace_generic-inl.h b/src/stacktrace_generic-inl.h
index 3375ada..9392ffc 100644
--- a/src/stacktrace_generic-inl.h
+++ b/src/stacktrace_generic-inl.h
@@ -38,6 +38,7 @@
 #include <stdlib.h>
 #include "google/stacktrace.h"
 
+// If you change this function, also change GetStackFrames below.
 int GetStackTrace(void** result, int max_depth, int skip_count) {
   static const int kStackLength = 64;
   void * stack[kStackLength];
@@ -55,3 +56,51 @@ int GetStackTrace(void** result, int max_depth, int skip_count) {
 
   return result_count;
 }
+
+// If you change this function, also change GetStackTrace above:
+//
+// This GetStackFrames routine shares a lot of code with GetStackTrace
+// above. This code could have been refactored into a common routine,
+// and then both GetStackTrace/GetStackFrames could call that routine.
+// There are two problems with that:
+//
+// (1) The performance of the refactored-code suffers substantially - the
+//     refactored needs to be able to record the stack trace when called
+//     from GetStackTrace, and both the stack trace and stack frame sizes,
+//     when called from GetStackFrames - this introduces enough new
+//     conditionals that GetStackTrace performance can degrade by as much
+//     as 50%.
+//
+// (2) Whether the refactored routine gets inlined into GetStackTrace and
+//     GetStackFrames depends on the compiler, and we can't guarantee the
+//     behavior either-way, even with "__attribute__ ((always_inline))"
+//     or "__attribute__ ((noinline))". But we need this guarantee or the
+//     frame counts may be off by one.
+//
+// Both (1) and (2) can be addressed without this code duplication, by
+// clever use of template functions, and by defining GetStackTrace and
+// GetStackFrames as macros that expand to these template functions.
+// However, this approach comes with its own set of problems - namely,
+// macros and  preprocessor trouble - for example,  if GetStackTrace
+// and/or GetStackFrames is ever defined as a member functions in some
+// class, we are in trouble.
+int GetStackFrames(void** pcs, int* sizes, int max_depth, int skip_count) {
+  static const int kStackLength = 64;
+  void * stack[kStackLength];
+  int size;
+
+  size = backtrace(stack, kStackLength);
+  skip_count++;  // we want to skip the current frame as well
+  int result_count = size - skip_count;
+  if (result_count < 0)
+    result_count = 0;
+  if (result_count > max_depth)
+    result_count = max_depth;
+  for (int i = 0; i < result_count; i++)
+    pcs[i] = stack[i + skip_count];
+
+  // No implementation for finding out the stack frame sizes yet.
+  memset(sizes, 0, sizeof(*sizes) * result_count);
+
+  return result_count;
+}