Re: Building current...

[Johnny Billquist <bqt%update.uu.se@localhost>]

On 2015-12-02 04:14, Christos Zoulas wrote:
> In article <565E5BE9.5090607%update.uu.se@localhost>,
> Johnny Billquist  <bqt%update.uu.se@localhost> wrote:
> > On 2015-12-02 03:05, Christos Zoulas wrote:
> > > In article <3F70DD86-0204-466E-8082-CBAD923F8B84%comcast.net@localhost>,
> > > Paul Koning  <paulkoning%comcast.net@localhost> wrote:
> > > > > On Dec 1, 2015, at 6:29 PM, Johnny Billquist <bqt%softjar.se@localhost> wrote:
> > > > >
> > > > > After some hiatus, "my" VAX 8650 is back among the living, so I
> > > > decided to try and build current again.
> > > > > ...
> > > > > Now the "analysis". It appears that programs call log10() with an
> > > > argument of 0, and consider this to be normal.
> > > > > Unfortunately, this is one of those places where the VAX not doing
> > > > IEEE FP bites us. The IEEE log10() of 0 will return -inf, and happily
> > > > chug on. The VAX log10() of 0 will cause an illegal instruction trap.
> > > > > I suspect the other programs crashing might be because of the same
> > > > reason. The question is - what should we do? I'm tempted to just change
> > > > the code for log() to return 0, or possible -MAX_whatever and not trap.
> > > > Opinions?
> > > >
> > > > Why not simply fix the bug?  log(0) is invalid.  It's probably an
> > > > initialization error, where log() is used for some magic timekeeping
> > > > math that is only valid the second time around.
> > >
> > > I fixed it.
> >
> > Care to share some details of how you fixed it?
>
> The code really want ilog2() it is an accident that it works anyway
> because it clamps the entries. I made it not use the log if the number
> of keys is <= 4. Check out this test program:
>
> #include <stdio.h>
> #include <math.h>
> #include <sys/types.h>
>
> #include <sys/bitops.h>
>
> static inline u_short
> auth_log2(double x)
> {
>          return (u_short)(log10(x) / log10(2));
> }
>
> int
> main(void)
> {
> 	unsigned long i;
> 	for (i = 0; i < 100; i++)
> 		printf("%lu %ld %ld\n", i, (long)auth_log2(i / 4.0) + 1,
> 		    (long)ilog2(i) - 1);
> 	return 0;
> }

Yeah. Cool. Thanks. If you think that it truly was a bug then I guess 
I'd go back and hunt why gcc barfs on me as well... Maybe it is 
something else after all...

	Johnny

--
Johnny Billquist                  || "I'm on a bus
                                  ||  on a psychedelic trip
email: bqt%softjar.se@localhost             ||  Reading murder books
pdp is alive!                     ||  tryin' to stay hip" - B. Idol