Added compile-time define to gate "initial-exec" attribute.

The "initial-exec" attribute is added to the ThreadCache::threadlocal_heap_
variable under normal compilation. This causes the linker to emit a
R_X86_64_TPOFF64 relocation in the _pyautolib.so shared object. This relocation
can cause python to error with:

cannot allocate memory in static TLS block

when it calls dlopen() on _pyautolib.so (when the import statement is
interpreted).

This only happens when the TLS section is large enough. Building Chrome with
-fprofile-generate to add some instrumentation-related data to the TLS is enough
to trigger this condition.

This CL encloses the function attribute in #if ! defined(PGO_GENERATE). When
Chrome is built with -fprofile-generate, we will also pass in -DPGO_GENERATE and
-ftls-model=global-dynamic so profile data can be collected without any errors.

BUG=none
TEST=Rebuilt chromeos-chrome in the chroot with -fprofile-generate
-ftls-model=global-dynamic -DPGO_GENERATE && ran pyautoperf autotests.


Review URL: http://codereview.chromium.org/10035012

git-svn-id: svn://svn.chromium.org/chrome/trunk/src@133810 0039d316-1c4b-4281-b951-d872f2087c98

2 files changed, 13 insertions, 2 deletions

diff --git a/third_party/tcmalloc/chromium/src/thread_cache.cc b/third_party/tcmalloc/chromium/src/thread_cache.cc
index 765786ff..1ad0f6d 100644
--- a/third_party/tcmalloc/chromium/src/thread_cache.cc
+++ b/third_party/tcmalloc/chromium/src/thread_cache.cc
@@ -64,7 +64,9 @@ int ThreadCache::thread_heap_count_ = 0;
 ThreadCache* ThreadCache::next_memory_steal_ = NULL;
 #ifdef HAVE_TLS
 __thread ThreadCache* ThreadCache::threadlocal_heap_
-# ifdef HAVE___ATTRIBUTE__
+// See comments in thread_cache.h about this. Bug here:
+// http://code.google.com/p/chromium/issues/detail?id=124489
+#if defined(HAVE___ATTRIBUTE__) && !defined(PGO_GENERATE)
    __attribute__ ((tls_model ("initial-exec")))
 # endif
    ;
diff --git a/third_party/tcmalloc/chromium/src/thread_cache.h b/third_party/tcmalloc/chromium/src/thread_cache.h
index 5294029..221cacb 100644
--- a/third_party/tcmalloc/chromium/src/thread_cache.h
+++ b/third_party/tcmalloc/chromium/src/thread_cache.h
@@ -259,7 +259,16 @@ class ThreadCache {
   // a good tradeoff for us.
 #ifdef HAVE_TLS
   static __thread ThreadCache* threadlocal_heap_
-# ifdef HAVE___ATTRIBUTE__
+  // This code links against pyautolib.so, which causes dlopen() on that shared
+  // object to fail when -fprofile-generate is used with it. Ideally
+  // pyautolib.so should not link against this code. There is a bug filed for
+  // that:
+  // http://code.google.com/p/chromium/issues/detail?id=124489
+  // For now the workaround is to pass in -DPGO_GENERATE when building Chrome
+  // for instrumentation (-fprofile-generate).
+  // For all non-instrumentation builds, this define will not be set and the
+  // performance benefit of "intial-exec" will be achieved.
+#if defined(HAVE___ATTRIBUTE__) && !defined(PGO_GENERATE)
    __attribute__ ((tls_model ("initial-exec")))
 # endif
    ;