http://gcc.gnu.org/bugzilla/show_bug.cgi?id=57974
--- Comment #4 from Uroš Bizjak <ubizjak at gmail dot com> ---
(In reply to Paolo Carlini from comment #3)
> I suspect the back end could be also involved because it happens only for
> long double and I didn't see anything special for long double in builtins.c
-funsafe-math-optimizations are used here:
'-funsafe-math-optimizations'
Allow optimizations for floating-point arithmetic that (a) assume
that arguments and results are valid and (b) may violate IEEE or
ANSI standards. When used at link-time, it may include libraries
or startup files that change the default FPU control word or other
similar optimizations.
The __builtin_expl is expanded to:
flds .LC0(%rip)
fld %st(0)
frndint
(*) fsubr %st, %st(1)
fxch %st(1)
f2xm1
fadds .LC2(%rip)
fscale
fstp %st(1)
fstpt 16(%rsp)
and with fsubr, we hit Inf - Inf, which is by definition NaN.
Obviously, this testcase exposes the unsafe part of
-funsafe-math-optimizations. Without this option everything works OK.