optimisation -O1 changes floating-point results

optimisation -O1 changes floating-point results

[Diggory Hardy]

Dear list,

I am using GCC to compile some modeling software (C++; 
http://code.google.com/p/openmalaria/). This uses quite a lot of "double"s and 
results can be quite sensitive to the exact outcomes of floating-point 
operations (this is due to stochastic influences, and only really a concern 
since we want to be able to reproduce results independent of the hardware/OS 
used).

Anyway, on 32-bit Linux (Ubuntu 12.04 with the default GCC 4.6.3) some of my 
test cases generated different results when optimisation is enabled. (This 
doesn't happen on 64-bit Linux (identical Ubuntu/GCC versions) or our Windows 
or Mac builds.)

On 32-bit linux results are as expected with no optimisation, but different 
when -O1 (or 2 or 3) is used, _however_, enabling all of the same optimisation 
flags (either as listed in the man page or as given by `gcc -O -Q --
help=optimizers`) (without listing -O) still produces different results.

Is this a known issue?

Is it the same with later GCC versions?

Is there a way I can turn on most optimisations while still getting exactly 
the same results as without optimisation?

Cheers,

Diggory Hardy

Re: optimisation -O1 changes floating-point results

[Florian Weimer]

On 12/18/2013 03:23 PM, Diggory Hardy wrote:

> Is this a known issue?

Yes, it's the excess precision problem on i386:

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

> Is it the same with later GCC versions?

GCC 4.5 has some fixes.

> Is there a way I can turn on most optimisations while still getting exactly
> the same results as without optimisation?

You could compile with -mfpmath=sse -msse2, or you could set the 
floating point control word to restrict precision to 53 bits.  The 
latter will cover doubles only, the issue can still surface with floats.

-- 
Florian Weimer / Red Hat Product Security Team