https://gcc.gnu.org/bugzilla/show_bug.cgi?id=126464
Bug ID: 126464
Summary: [16/17 Regression] wrong code with FP ranges
Product: gcc
Version: 17.0
Status: UNCONFIRMED
Keywords: wrong-code
Severity: normal
Priority: P3
Component: tree-optimization
Assignee: unassigned at gcc dot gnu.org
Reporter: ktkachov at gcc dot gnu.org
Target Milestone: ---
__attribute__((noipa)) double
scale (double x)
{
float y = (float) x;
if (y == -__builtin_inf ())
return x * 0.5; /* retry at half scale */
return y;
}
/* Same shape, widening the other way, to show it is not _Float16 or float
specific. */
__attribute__((noipa)) long double
scale2 (long double x)
{
double y = (double) x;
if (y == __builtin_inf ())
return x * 0.5L;
return y;
}
int
main (void)
{
if (scale (-1e300) != -5e299)
__builtin_abort ();
if (sizeof (long double) > sizeof (double)
&& scale2 (1e4000L) != 5e3999L)
__builtin_abort ();
return 0;
}
runs fine at -O0 and aborts at -O1 and above. Also with GCC 15 it works fine.
I think something to do with fp ranges
