[Bug math/21930] C-only gcc builtins used in <math.h> isinf

["joseph at codesourcery dot com" <sourceware-bugzilla at sourceware dot org>]

https://sourceware.org/bugzilla/show_bug.cgi?id=21930

--- Comment #3 from joseph at codesourcery dot com <joseph at codesourcery dot com> ---
On Wed, 9 Aug 2017, gftg at linux dot vnet.ibm.com wrote:

> > (c) Use some suitable C++ magic in place of __builtin_types_compatible_p
> > (remembering that __builtin_types_compatible_p explicitly removes type
> > qualifiers, and alternative C++ magic needs to do likewise).
> 
> Since typeid also removes qualifiers implicitly
> (http://en.cppreference.com/w/cpp/language/typeid), would something in the
> lines of the following be acceptable for c++98:
> 
> if (typeid(f) == typeid(float)) isinff (f)

No, because that involves a runtime comparison of type_info objects.  We 
want something (to go in a __MATH_TG macro definition for the "C++ and 
distinct float128" case - and, similarly, an isinf definition for when C++ 
is combined with the present conditions for handling GCC < 7 with 
_Float128) that's optimized away at compile time - something where the 
conditionals are always folded to true or false.  (And preferably not 
depending on including any C++ headers from <math.h>; I don't know what 
the rules are for such header inclusion in the C++ standard library.)

Maybe we need to understand further what cases are the problem.  Various 
macros may be used in that configure test, but most of them are using 
type-generic built-in functions which work fine for C++ (given that 
libstdc++ is not built with sNaNs enabled).  The problem for the configure 
test might only be the "__builtin_isinf_sign is broken for float128 only 
before GCC 7.0." definition of isinf, which uses 
__builtin_types_compatible_p.  Separately, with the installed compiler, 
most of the macros should be undefined in the libstdc++ headers, leaving 
issignaling as the likely problem case using __MATH_TG (issubnormal is 
defined using fpclassify; iszero has a separate C++ template definition, 
to avoid problems with macros).

Possible solutions for isinf include:

* Do not use the special-case isinf definition with 
__builtin_types_compatible_p for C++.  It's possible this suffices to make 
the configure test work (and exactly what the isinf macro does doesn't 
matter when using an installed C++ compiler because the libstdc++ headers 
undefine that macro, and the configure test doesn't use it with float128).

* Add a C++ variant of that definition, which would only be used for the 
configure test but which still has the right semantics.

And for issignaling:

* Change __MATH_TG to have a fully functional C++ variant.  float and 
double can be distinguished with sizeof; it's only long double / float128 
where you need a replacement for __builtin_types_compatible_p (__typeof 
(TG_ARG), long double) using C++ features.

* Add template versions of issignaling for C++, like iszero.

-- 
You are receiving this mail because:
You are on the CC list for the bug.