9front - general discussion about 9front
 help / color / mirror / Atom feed
* [9front] Bug converting vlong to double on 32bit architectures
@ 2022-12-30 13:19 cosarara
  2022-12-30 16:57 ` Jaume Delclòs Coll
  2022-12-30 18:32 ` cinap_lenrek
  0 siblings, 2 replies; 3+ messages in thread
From: cosarara @ 2022-12-30 13:19 UTC (permalink / raw)
  To: 9front

Hi,

While trying to run kvik's lu9 on ARM, I found that when converting 
LLONG_MIN to a double on 32bit systems, the result is wrong (positive 
instead of negative).

Given the following test program:

#include <u.h>
#include <libc.h>

#define LLONG_MAX
#define LLONG_MIN    (-LLONG_MAX-1)

void main() {
     vlong min = LLONG_MIN;
     double dmin = min;
     print("minint %lld\n", min);
     print("minint as double %f\n", dmin);
     if (dmin > 0.0) {
         exits("int min as double turned positive");
     }
     exits(0);
}

The output on x86_64 will be:

minint -9223372036854775808
minint as double -9223372036854776400.000000

But on arm or 386 (and I expect also spim, 68000, mips, 68020, sparc, 
power, since they all use the same _v2d):

minint -9223372036854775808
minint as double 9223372036854776400.000000

And the value turned positive in the conversion.

The function used for the cast to double is (in /sys/src/libc/arm/vlrt.c):

#define    SIGN(n)    (1UL<<(n-1))

double
_v2d(Vlong x)
{
     if(x.hi & SIGN(32)) {
         if(x.lo) {
             x.lo = -x.lo;
             x.hi = ~x.hi;
         } else
             x.hi = -x.hi;
         return -((long)x.hi*4294967296. + x.lo);
     }
     return (long)x.hi*4294967296. + x.lo;
}

If I understand correctly, the issue is that where it tries to flip the 
sign for x.hi (x.hi = -x.hi), 0x80000000 has no positive, thus stays the 
same (it stays negative). Then when we get to the negative return, we 
get a positive out.

What came to my mind then, is that in the case that there is no x.lo, we 
can keep the x.hi sign and cast directly, thus:

double
_v2d(Vlong x)
{
     if(!x.lo) {
         return (long)x.hi*4294967296.;
     }
     if(x.hi & SIGN(32)) {
         x.lo = -x.lo;
         x.hi = ~x.hi;
         return -((long)x.hi*4294967296. + x.lo);
     }
     return (long)x.hi*4294967296. + x.lo;
}

This looks correct to me, but I don't trust myself to not make mistakes 
in such critical code, so I would like some feedback on the change.

Happy new year in advance,

cosa

^ permalink raw reply	[flat|nested] 3+ messages in thread

end of thread, other threads:[~2022-12-30 18:34 UTC | newest]

Thread overview: 3+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2022-12-30 13:19 [9front] Bug converting vlong to double on 32bit architectures cosarara
2022-12-30 16:57 ` Jaume Delclòs Coll
2022-12-30 18:32 ` cinap_lenrek

This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).