Paul Eggert wrote:
> PƔdraig Brady <address@hidden> writes:
>
>> This patch makes `seq 0.1 0.1 0.9` output 0.1 to 0.9 inclusive, as expected.
>
> I see some problems with that patch.
>
> First, it continues to mishandle some similar cases. For example,
> 'seq 0.1 1e-1 0.9' outputs only 0.1 through 0.8, i.e. it behaves
> differently from 'seq 0.1 0.1 0.9', which is counterintutive. I can
> easily see getting some bug reports about that.
Agreed. The old behaviour is better than that.
Hmm, off the top of my head, I don't see any way to
approximate the precision of nums in the form 1e-1
>
> Second, it mishandles some cases that the current code handles
> correctly. For example, on my platform (Debian stable x86 with GCC
> 4.2) 'seq 922337203685477580.4 0.1 922337203685477580.5' currently
> outputs this:
>
> 922337203685477580.4
> 922337203685477580.5
>
> which is correct, but with the proposed patch it outputs this:
>
> 922337203685477580.4
> 922337203685477580.5
> 922337203685477580.6
>
> due to rounding errors in 'long double' arithmetic (which uses a 64
> bit fraction on my platform).
I don't think that is a normal use case for seq,
and I was basically sacrificing accuracy at this end.
> The "right" way to address this problem is to use arbitrary-precision
> decimal arithmetic. By comparison, fiddling with fixed-width IEEE
> binary formats is problematic; it may fix some cases but it'll
> probably introduce other problems and it'll make it harder to explain
> the inevitable quirks.
Agreed.
I'll think a bit more, whether it's possible to handle the first case better.
thanks for the review!
PƔdraig.