Test code:
long double x = __builtin_infl();
int main() {
int r = x == __builtin_infl();
return r ? 0 : 1;
}
Commands to reproduce:
gcc -mlong-double-128 tc.c
./a.out && echo ok || echo fail
No problem with -mlong-double-64.
Cause:
GCC first transforms (x == infinity) to (x > long_double_max_val).
The maximum value returned by real_maxval is encoded as infinity due
to rounding in encode_ibm_extended. This does not happen with the
LDBL_MAX declared in system headers because one of the bits is
cleared.
Potential fix:
--- real.c.orig 2005-09-19 12:56:24.000000000 -0400
+++ real.c 2006-04-25 21:18:32.000000000 -0400
@@ -2205,6 +2205,8 @@ real_maxval (REAL_VALUE_TYPE *r, int sig
np2 = SIGNIFICAND_BITS - fmt->p * fmt->log2_b;
memset (r->sig, -1, SIGSZ * sizeof (unsigned long));
clear_significand_below (r, np2);
+ if (REAL_MODE_FORMAT_COMPOSITE_P (mode))
+ clear_significand_bit (r, SIGNIFICAND_BITS - fmt->pnan - 1);
}
--
Summary: Compare against constant infinity fails with IBM long
double format
Product: gcc
Version: 4.0.3
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: middle-end
AssignedTo: unassigned at gcc dot gnu dot org
ReportedBy: dvdfrdmn at users dot sf dot net
GCC target triplet: powerpc-apple-darwin8
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=27321
