EBFE (x.delete@this.y.com) on November 28, 2012 7:54 pm wrote:
> I never said that "(avoiding insane optimizations)" is an overhead.
> "(without insane optimizations)" is the baseline [0 overhead].
> However you were talking as if "undefined behaviour" will bring an overhead, when replying Wilco's post:
> > > Wrong. Exceptions can only happen explicitly with throw in C++.
> > Or when the program hits an undefined behaviour...
> > And there are quite a few undefined behaviours in C++ spec.

So you mis-understood me, not, it is not the undefined behaviour that causes the overhead, but the fact that compilers writers choose to fill many of those undefined behaviours with more logical behaviours by avoiding some aggressive optimizations that Wilco as arguing managed language must avoid.

> Caller's try...catch introduces an additional (and nasty) execution path in callee.
> It splits basic block and gives trouble to a lot of optimizations.
> Some stores can no longer be eliminated; some operations can no longer be hoisted/sunk/reordered...

Oh, the optmization I tought, the caller doesn't need to access callee's locals, you can reorder, hoist, sunk those as mush as you want, only stores to globals must be ordered, I suspect many C/C++ compilers will try to not mess with global stores as well even when the specs allow it.

I would enjoy discussing others optimizations that are not possible.