Ignore CpuProfiler::SetIdle call when not profiling.

src/api.cc

Index: src/api.cc
diff --git a/src/api.cc b/src/api.cc
index 8214028311bc1b7dcc60cd7cabb9801f86446438..084da50be8315e2a31302778157e846459222cee 100644
--- a/src/api.cc
+++ b/src/api.cc
@@ -8352,9 +8352,11 @@ CpuProfile* CpuProfiler::StopProfiling(Local<String> title) {
 
 
 void CpuProfiler::SetIdle(bool is_idle) {
-  i::Isolate* isolate = reinterpret_cast<i::CpuProfiler*>(this)->isolate();
+  i::CpuProfiler* profiler = reinterpret_cast<i::CpuProfiler*>(this);
+  if (!profiler->is_profiling()) return;

[alph, 2016/04/30 00:03:08]

Sorry, didn't get how it helps. If we're not profiling, but it's called from
within a compilation it'll get the state of v8::COMPILER and fail the DCHECK. Am
I missing something?

[dgozman, 2016/04/30 00:10:39]

This does not happen, since compilation cannot trigger idle state unless we run
nested message loop.
Do you think we should instead allow to call this at any time and just remove
the DCHECK? That's a possible solution.

[alph, 2016/04/30 00:29:20]

Looks more robust to me. The problem is that when you call setIdle(false) you
have to restore the previous state e.g. COMPILER, rather than set it to
EXTERNAL.

Relying on the fact that profiling disables nested messages loops looks fragile.
Despite it was done this way before.

+  i::Isolate* isolate = profiler->isolate();
   v8::StateTag state = isolate->current_vm_state();
-  DCHECK(state == v8::JS || state == v8::EXTERNAL || state == v8::IDLE);
+  DCHECK(state == v8::EXTERNAL || state == v8::IDLE);
   if (isolate->js_entry_sp() != NULL) return;
   if (is_idle) {
     isolate->set_current_vm_state(v8::IDLE);