Re: link error log2 exp2 gcc 3.4.0 and resolution

[Bill Allombert <allomber@math.u-bordeaux.fr>]

On Sun, Apr 25, 2004 at 04:36:21PM -0700, Carl Hansen wrote:
> a bug, and its resolution
> 
> just installed gcc 3.4.0 (on Solaris 9 sparc)
> 
> couldn't remake, link errors:
> Undefined  symbol   exp2   log2
> 
> examining pari-2.2.7.alpha/config/has_log2.c,
> (which is used by Configure) :
> 
> 
> #include <math.h>
> main(){ double x=log2(1.0); }
> 
> compiles cleanly,
> 
> however trying:
> 
> #include <math.h>
> main(){ double x=log2(0.666); }
> 
> get        Undefined  symbol   log2
> 
> 
> Having pondered, conclude the compiler compile-time-optimises
> log2(1.0)  to 0 , so log2 is never  called at runtime, as
> can be verified by looking at has_log2.s , so the Configure script
> will conclude has it HAS log2, whereas  in  the
> other case log2(0.666), Configure script will realise
> we  in fact DON'T have log2.
> 
> likewise has_exp2.c
> 
> So contemplate  assumptions made by compiler, compiler writer, writer
> of Configure, and become wise, and wonder what else could be wrong.

I have studied the problem on i386-linux, glibc 2.3.2.ds1-12 and
both gcc 3.3.3 and a snapshot of 3.5.0 (I don't have 3.4.0 at hand).

Apparently the way PARI check for log2 is broken.
PARI try to compile and link a simple program using log2 but:

1) It does not specify -lm, so it fails.

2) That does not check whether the prototype for log2 exists, only
   if the log2 symbol is in the library.

glibc has log2 but it provide an prototype for it only in c99 mode
(that might be a bug, I would expect c89 to have log2.)

So apparently this is working only by chance (problem 1 compensating for
problem 2).

On top of that, apparently gcc 3.4/3.5 precompute log2(1.0) so don't
need -lm to pass that test.

Cheers,
Bill.