http://gcc.gnu.org/bugzilla/show_bug.cgi?id=55771
--- Comment #4 from Mikael Pettersson <mikpe at it dot uu.se> 2012-12-21 10:06:50 UTC --- I'm beginning to think the test case is invalid. The operands of the multiplication in f1 are unsigned long and float, but they are not converted to double. Instead the unsigned long is converted to float (C1x, N1494, 6.3.1.8, 1st paragraph, 2nd "Otherwise" sub-paragraph), but -3UL is outside the range of a 32-bit float, so the result is undefined (C1x, N1494, 6.3.1.4, 2nd paragraph, last sentence).
