Re: Compiler errors

On Wednesday 12 October 2005 09:11 am, Steve Little wrote:
> I've run the examples (in batch mode) using examples/runall,
> and they look fine (with a minor difference in the last test,
> possibly the example output is out of date? i notice the
> version numbers are out of date on it...)

I missed this one ..

It is common for the results of regression testing to not match 
exactly.  Minor differences can be ignored.

This is an artifact of floating point arithmetic.  It is 
inherently non-exact.   Simple changes can cause changes in 
generated code, which can cause changes in the propagation of 
round off error.  The only way to get them to match exactly is 
to use the same compiler on the same type of system with the 
same options.  The "debug" version doesn't match the optimized 
version.  An Intel processor may not match an AMD.  It 
certainly doesn't match a power-pc or sparc.  In some cases, it 
is possible to take a different path through the code, which is 
probably a bug, but usually the difference is insignificant.