[PATCH] D19271: Change CALL_ONCE_INITIALIZATION implementation to use static initialization

[Kostya Serebryany via llvm-commits]

kcc added a reviewer: rnk.
kcc added a comment.

Reid, are function scope statics thread-safe with VC++ 2013?


================
Comment at: include/llvm/PassSupport.h:35
@@ +34,3 @@
+#define CALL_ONCE_INITIALIZATION(function)                                     \
+  static bool init_once([&] {                                                  \
+    function(Registry);                                                        \
----------------
joker.eph wrote:
> kcc wrote:
> > kcc wrote:
> > > Mmm. I am probably missing something trivial (c++11-ish), but how is this going to work?
> > > For me it looks like you define a static bool function, but never call it. 
> > Aaaah, init_once is a bool variable, not a function. Oh, how much this is confusing... 
> > Will this work with the versions of Visual C++ we use for LLVM? 
> > I remember that static initialization was broken there (until recently?)
> > 
> > Why not std::call_once? 
> Yeah it would deserve a comment probably :)
> I don't know if static initialization is safe in MSVC 2013 (I know it was announced to be "fixed", but don't know which version).
> 
> `std::call_once` require `<mutex>`, and I was worried of portability (I had trouble last time I tried to include it), and because of that we have an abstraction of mutex in libSupport I believe). 
> 
If you can make it 
  static bool initialized_once = ... 
it will be much more clear and not require a comment. 


http://reviews.llvm.org/D19271