Add thread pool class

Added a thread pool class loosely based on google3 code.

Modified the compiler to have a single thread pool instead of creating new threads in ForAll.

Moved barrier to be in top level directory as it is not GC specific code.

Performance Timings:

Reference:
boot.oat: 14.306596s
time mm oat-target:
real    2m33.748s
user    10m23.190s
sys 5m54.140s

Thread pool:
boot.oat: 13.111049s
time mm oat-target:
real    2m29.372s
user    10m3.130s
sys 5m46.290s

The speed increase is probably just noise.

Change-Id: If3c1280cbaa4c7e4361127d064ac744ea12cdf49
diff --git a/src/barrier.cc b/src/barrier.cc
new file mode 100644
index 0000000..9651828
--- /dev/null
+++ b/src/barrier.cc
@@ -0,0 +1,46 @@
+#include "barrier.h"
+#include "../src/mutex.h"
+#include "thread.h"
+
+namespace art {
+
+Barrier::Barrier()
+    : count_(0),
+      lock_("GC barrier lock"),
+      condition_("GC barrier condition", lock_) {
+}
+
+void Barrier::Pass(Thread* self) {
+  MutexLock mu(self, lock_);
+  SetCountLocked(self, count_ - 1);
+}
+
+void Barrier::Wait(Thread* self) {
+  Increment(self, -1);
+}
+
+void Barrier::Init(Thread* self, int count) {
+  MutexLock mu(self, lock_);
+  SetCountLocked(self, count);
+}
+
+void Barrier::Increment(Thread* self, int delta) {
+  MutexLock mu(self, lock_);
+  SetCountLocked(self, count_ + delta);
+  if (count_ != 0) {
+    condition_.Wait(self);
+  }
+}
+
+void Barrier::SetCountLocked(Thread* self, int count) {
+  count_ = count;
+  if (count_ == 0) {
+    condition_.Broadcast(self);
+  }
+}
+
+Barrier::~Barrier() {
+  CHECK(!count_) << "Attempted to destory barrier with non zero count";
+}
+
+}