https://gcc.gnu.org/bugzilla/show_bug.cgi?id=121207
Bug ID: 121207
Summary: __LDBL_EPSILON__ is wrongly defined on power
platforms.
Product: gcc
Version: 15.1.1
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: target
Assignee: unassigned at gcc dot gnu.org
Reporter: cassio.neri at gmail dot com
Target Milestone: ---
On power, power64 and power64le (where long double is the IBM128 format)
__LDBL_EPSILON__ is defined with the same value as __LDBL_DENORM_MIN__. See
https://godbolt.org/z/qTTaWsovq
Consequently, 1.0l + __LDBL_EPSILON__ isn't the minimum long double value
greater than 1.0l. Indeed, in the code below the assert fires when EPSILON is
undefined but, otherwise, it doesn't. The code also shows that __LDBL_EPSILON__
should be defined as 2^(-105) ~= 2.465190328815662e-32l.
#include <cassert>
#include <cmath>
#include <limits>
int main() {
#ifdef EPSILON
auto const epsilon = 2.465190328815662e-32l; // == 2^(-105)
#else
auto const epsilon = std::numeric_limits<long double>::epsilon();
#endif
auto const x1 = 1.0l + epsilon;
auto const x2 = std::nextafter(1.0l, 2.0l);
assert(x1 == x2);
}