On Friday, December 19, 2003, at 11:17 AM, Thomas Fischbacher wrote:

> Please note that this bug concerns the "old" Num implementation that 
> was
> removed in a recent update.

This bug is in fact in the ML part of the Num library, which is common
to both the old and new Num implementations.  I think it dates back to
Caml Light.

> The attachment will reproduce a condition where doing the same 
> calculation
> (which is purely functional to the outside and does not use global 
> state
> or side effects, but involves Num) many times in sequence will give
> different results.

This "nondeterminism" comes from using a Nat.nat with an uninitialised
digit, using whatever was already in memory at that address.

> arithmetics I am using? Yet to me it seems plausible that some serious
> issue may have been overlooked here, since I suppose very few people 
> use
> ratio arith (in contrast to bignum-only arith, which has applications 
> for,
> e.g., RSA).

Indeed, the bug is in the big-int part of the library (arbitrary 
precision
signed integers).


Thank you for your well-delimited bug report with a reproducible test
case.  It was rather easy to find the bug with it.

I enclose a patch, which should work on 3.06 and on 3.07pl2.

-- Damien