http://gcc.gnu.org/bugzilla/show_bug.cgi?id=57408
--- Comment #3 from Daniel Krügler <daniel.kruegler at googlemail dot com> --- Further simplification down to a library-free test case: //------------------------------------------------------ template<typename Callable> struct Impl { Callable func; Impl(Callable f) : func(f) { } virtual void run() { func(); } }; template<typename Callable> void call(Callable f) { Impl<Callable>(f).run(); } extern "C" int printf(const char*, ...); int main(){ int y = 2; float fa[2][y]; // compiles fine if y were 2 hard-coded instead fa[0][0]=0.8; fa[0][1]=1.8; auto fx=[&](){ for(int c=0; c<2; c++){ // compiles fine if c<2 were c<1 instead printf("use me", fa[0][c]); } }; call(fx); //error (1*) } //------------------------------------------------------ It seems relevant, that there is a virtual function that invokes the lambda closure and that fa[0][c] is odr-used within the closure call expression.