Predefine __NO_MATH_ERRNO__ for -fno-math-errno.

This patch adds a predefined macro __NO_MATH_ERRNO__ for when
-fno-math-errno is passed or implied.  This allows math.h to provide a
more accurate definition of the C99 math_errhandling macro that takes
this option into account, and allows for choice of libm functions to
be optimized at compile time based on this option.

(There may be a case for such interfaces for -fno-rounding-math
(default) and -fno-trapping-math as well, but as C99 standard pragmas
would allow those to vary on a per-block basis, predefined macros
would be problematic as the interface; you can't select a
-fno-trapping-math or -fno-rounding-math version of a function in a
standard header if a conforming program could then use "#pragma STDC
FENV_ACCESS ON" to require a -ftrapping-math -frounding-math version
in a particular block.  So built-in functions might be a better way of
providing access to information about those options.)

Bootstrapped with no regressions on x86_64-unknown-linux-gnu.  OK to
commit?

	* cppbuiltin.c (define_builtin_macros_for_compilation_flags):
	Define __NO_MATH_ERRNO__ if -fno-math-errno.
	* doc/cpp.texi (__NO_MATH_ERRNO__): Document predefined macro.

c-family:
	* c-cppbuiltin.c (c_cpp_builtins_optimize_pragma): Define and
	undefine __NO_MATH_ERRNO__ based on changes to -fmath-errno state.

testsuite:
	* gcc.dg/no-math-errno-1.c, gcc.dg/no-math-errno-2.c,
	gcc.dg/no-math-errno-3.c, gcc.dg/no-math-errno-4.c: New tests.

From-SVN: r217420