Inconsistent results casting double to int on x86

[Andrew Haley]

Perry Smith writes:
 > On Mar 22, 2006, at 3:05 PM, dups41@gmail.com wrote:
 > 
 > > On x86 the following code intended to calculate log2(4096) gives an
 > > unexpected result of 11 when the value is cast directly to an int. If
 > > the result is stored in a double and then cast to an int the expected
 > > value (12) is given.
 > >
 > >
 > > double r0;
 > > double d_msb;
 > > int i_msb;
 > >
 > > r0=4096;
 > >
 > > i_msb = (int)(log10(r0)/log10(2));
 > > printf("%d\n",i_msb);
 > >
 > > d_msb = (log10(r0)/log10(2));
 > > i_msb = (int)d_msb;
 > > printf("%d\n",i_msb);
 > >
 > >
 > > # uname -srm
 > > Linux 2.4.9-34smp i686
 > > # g++ -o log log.c
 > > # ./log
 > > 11
 > > 12
 > >
 > >
 > > I have tried several versions of gcc on x86 and all give the same
 > > behaviour. (3.0, 3.0.1,  3.0.4,  3.2,  3.2.2,  3.3.1)
 > >
 > > On Solaris 8 this code gives the expected output (12 12). AMD64 with
 > > -m64 gives the expected results but with -m32 the results differ (11
 > > 12).
 > >
 > > Why do the two values differ? If it is the case that a
 > > rounding/precision error is causing an off by one result, should it
 > > not be consistent between the two forms of the code?
 > 
 > This seems odd to me too.  I realize that floating point is not exact  
 > but that inexactness should be the same in both methods of your  
 > algorithms.
 > 
 > My suggestion is to dump out the listing of the code and look at it.   
 > The instruction sequence must be different and that may clue you in  
 > on why the results are different.  I'm curious what you find.

This is because of excess precision in the x87.

This is http://gcc.gnu.org/bugzilla/show_bug.cgi?id=323.

Andrew.