I thought this was a pretty funny thing but unfortunately when I tried this on ICC it seemed to compile just fine.
Though I am amused by one thing: the VLA version generates worse code on all compilers I've tried. Seems to validate the common refrain that VLAs tend to break optimizations. (Surely it's worse when you have an on-stack VLA though.)