Bug in single/double operations?

Can anybody explain this behaviour?

If you run the following code, it behaves as expected. You finish up with approximately eps('single')=1e-7.

HOWEVER, if you comment out the disp(x) line, then you finish up with eps('double')=1e-16.

Why should a disp(x) change anything? Is it a sneaky invisible casting?

------------------------------------

format long

x = single(1)            % single precision 

while single(1+x) > 1    % keep going until no difference

   x= x/2; % if you can tell any difference, half it.

   disp(x) % comment this line out 
end % while
---------------------------------------

[Paulo Silva]

weird behavior, I have no idea why that happens

Answer 1

[James Tursa]

This exact issue has already been discussed on the newsgroup, with the JIT being the likely cause of the difference:

http://www.mathworks.com/matlabcentral/newsreader/view_thread/291577

Answer 2

[Mike Hosea]

It appears to be something to do with the MATLAB JIT-Accelerator. I'm going to say it's probably unintentional. :)

I have to add, however, that while one can usually get away with an algorithm like that in MATLAB (except for code generation), we had to give up on it in numerical software in general after Intel came out with 80-bit extended-precision floating-point registers. Usually you want the eps for the data type, not for the register.