The IBM long double implementation of erfl() and erfcl() both give inaccurate results when given a negative argument such that the component doubles have differing sign.  Also, erfl() with a large negative argument returns +1.0 rather than -1.0.

The following compiled with -O0 shows one such bad result.  (Compare with results from -O2 where gcc evalates erfl and erfl.)

#include <stdio.h>
#include <math.h>

int
main (void)
{
  long double x, z;
  union { long double ld; double d[2]; } u;

  x = -0x1.fffffffffffff8p-2L;
  z = erfl (x);

  u.ld = x;
  printf ("%24.21Lf %36La %21a %21a\n", u.ld, u.ld, u.d[0], u.d[1]);
  u.ld = z;
  printf ("%24.21Lf %36La %21a %21a\n", u.ld, u.ld, u.d[0], u.d[1]);

  z = erfcl (x);
  u.ld = z;
  printf ("%24.21Lf %36La %21a %21a\n", u.ld, u.ld, u.d[0], u.d[1]);

  return 0;
}