__builtin_isinf, __builtin_isnan and the various __builtin_is* functions
for unordered comparisons are declared as functions taking variable
arguments so they can be used directly to implement the corresponding C99
type-generic macros. However, __builtin_isinf does not work properly
as type-generic as it sometimes expands to a function call and that call
is always to isinf - which takes a double argument, not float or long double,
if it is defined as a function at all. For example, the following wrongly
returns 1 on x86_64-unknown-linux-gnu, where the double and long double
calling conventions are different.
#include <float.h>
double i = __builtin_inf();
int j;
int x(long double ld) { j = __builtin_isinf(i); return __builtin_isinf(ld); }
int main(void) { return x(LDBL_MAX); }
While __builtin_isinf may call a library function, that function should depend
on the argument type. But to be usable to implement the C99 macro, it should
not depend on a library function which may not be in the standard library;
it should be inlined based on examining the floating point number's bit
pattern or call a libgcc function.
[I don't think __builtin_isnan has this problem because I don't think it
will ever call a library function. __builtin_signbit is not currently
type-generic although it ought to be by analogy with the other functions
for type-generic macros; I'm not sure if it can ever call a library function
now. We don't have __builtin_isfinite or __builtin_isnormal functions.
libgcc2.c has its own macro versions of isfinite and isinf but the
bit-manipulation done e.g. in the glibc implementations may be more efficient.]
--
Summary: variadic __builtin_isinf broken
Product: gcc
Version: 4.1.0
Status: UNCONFIRMED
Severity: normal
Priority: P2
Component: middle-end
AssignedTo: unassigned at gcc dot gnu dot org
ReportedBy: jsm28 at gcc dot gnu dot org
CC: gcc-bugs at gcc dot gnu dot org
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=20558
