[9fans] Problem with native vs APE math

[9fans] Problem with native vs APE math

[Doug Henderson]

Here is my example program, running on a Celeron 400MHz. with the June 17
release. I have built it in native mode (mk, 8c, 8l) and in the APE
environment. It works when built in APE (and also, with minor mods in
linux). The native mode executable suicides during the log(0.0L) call:
edemo 355: sys: fp: invalid fppc=0x1391 pc=0x00001050

The library code looks ok. It should return -Inf for log(0) and NaN for
log(-1), which is what the APE does.

Any ideas?

#ifdef _POSIX_SOURCE
#include <stdio.h>
#include <math.h>
#else
#include <u.h>
#include <libc.h>
#include <stdio.h>
#endif
int main(int argc, char* argv[])
{
 printf("log(1.0) = %f\n", log(1.0L));
 printf("log(0.0) = %f\n", log(0.0L));
 printf("log(-1.0) = %f\n", log(-1.0L));
 return 0;
}
--  --
Doug Henderson, Calgary, Alberta, Canada

Re: [9fans] Problem with native vs APE math

[Doug Henderson]

On my Celeron, I get log(1)=0, log(0)=-HUGE_VAL and log(-1)=-HUGE_VAL, which
is what I expect from reading the code for log in
/sys/src/ape/lib/ap/math/log.c, and errno is set to ERANGE for log(0) and
EDOM for log(-1).

Clearing the FPINVAL bit in the pcr gets rid of the suicides.

There is an undocumented bit set in psr after log(0).
Also, there is a mismatch between some defines in ape/float.h and
ape/sys/fp.h

----- Original Message -----
From: "Russ Cox" <rsc@plan9.bell-labs.com>
To: <djhender@telusplanet.net>; <9fans@cse.psu.edu>
Sent: Monday, July 31, 2000 12:27 AM
Subject: Re: [9fans] Problem with native vs APE math


> When I compile that using APE I get 0.00000
> for all the answers, which is not what I expected
> but not unreasonable given the code.
> I don't think that the APE libraries use Inf and NaN.
>
> As for why it crashes natively, you need to
> turn off some exceptions related to use of
> Inf and NaN -- see getfcr(2).
>
> Russ
>

Re: [9fans] Problem with native vs APE math

[Russ Cox]

When I compile that using APE I get 0.00000
for all the answers, which is not what I expected
but not unreasonable given the code.
I don't think that the APE libraries use Inf and NaN.

As for why it crashes natively, you need to
turn off some exceptions related to use of
Inf and NaN -- see getfcr(2).

Russ