While looking into the regression tests' compatibility with C++14, many of the failures were because that specification defined UDLs for imaginary constants, specifically ones ending in 'i' and 'il'. This conflicted with the previous GNU extension that allowed them purely as a builtin literal syntax so they'd been disabled in C++14 mode.
GCC's behaviour is to use the builtin constants in all "-std=gnu++N" modes but disable them for "-std=c++11" and "-std=c++14" (they are still permitted in "-std=c++98"). This seemed inconsistent (and likely an implementation detail rather than intended behaviour) so my patch makes the decision purely on whether we're in GNU mode.
Does it look reasonable?