I did some investigation and found that FreeBSD and therefore DragonFly allows to set up rounding precision modes via fpgetprec(3), but the rounding precision that is supposed to be enough (FP_PE) is already set by default.

Also, quick search showed that IEEE only requires so called exact rounding only for a few basic operations: +,-,*,/ and sqrt, and doesn't seem to say anything about pow. But I would expect to get correct (as in exact rounding) result when rounding precision is set to FP_PE via fpgetprec, which means extended rounding precision, i.e. with 64-bit significand.

Aleksej Lebedev

On 22/09/16 19:33, Aleksej Lebedev wrote:
Hi, everyone!

I'm porting ChezScheme that was recently opensourced to DragonFly.

I was able to compile it, but lots of tests are failing. At least some
of them are due to (minor) bugs in DragonFly

Also I noticed that one of the tests is failing because function pow(3)
returns a different (wrong?) value compared to pow on other platforms.

On DragonFly:
a@kl:~/tmp$ uname -a
DragonFly kl.zta.lk 4.6-RELEASE DragonFly v4.6.0.10.g16fba-RELEASE #10:
Wed Aug 17 14:26:31 CEST 2016
r...@kl.zta.lk:/usr/obj/usr/src/sys/X86_64_GENERIC  x86_64
a@kl:~/tmp$ cat powtest.c
#include <stdio.h>
#include <math.h>

void printrep(double x)
  unsigned long r = *(unsigned long *)&x;
  for (int i = 0; i != 64; ++i)
    putchar((r & (1lu << (63 - i))) ? '1' : '0');

int main()
  double x, y;
  scanf("%lf%lf", &x, &y);
  printrep(pow(x, y));
  return 0;
a@kl:~/tmp$ cc powtest.c -lm
a@kl:~/tmp$ echo 10.0 -20.0 | ./a.out

On Mac OSX the same program:
a@zdev:~/tmp$ uname -a
Darwin zdev.local 15.6.0 Darwin Kernel Version 15.6.0: Mon Aug 29
20:21:34 PDT 2016; root:xnu-3248.60.11~1/RELEASE_X86_64 x86_64
a@zdev:~/tmp$ cc powtest.c
a@zdev:~/tmp$ echo 10.0 -20.0 | ./a.out

I am not an expert in floating point arithmetic, more importantly not an
expert in IEEE754. Maybe this is not a bug. Looks like a different
rounding, but I'm not sure if it is allowed by IEE754. Does anyone know
if it's OK? I can simply use powl instead of pow, this will make the
test pass, but I would like not to create a workaround if this is really
a bug.

Aleksej Lebedev

Reply via email to