Re: Unexpected interval comparison

Kyotaro HORIGUCHI <horiguchi(dot)kyotaro(at)lab(dot)ntt(dot)co(dot)jp> writes:
> The first attached is the revised patch and the second is
> temporary sanity check code for non-128bit environment code. (but
> works only on 128 bit environment)

This seemed to me to be probably even less correct, so I extracted
the addition and multiplication logic into a standalone test program
(attached), which compares the result of a multiplication to that
of native int128 arithmetic. I changed the order of the LinearInterval
fields to be LS-first so that I could overlay them onto an int128
result (on a little-endian machine); this is just for testing purposes
not something we must do in the finished code. I soon found cases
where it indeed fails, eg

$ ./a.out 0x7ffffffff 0x7ffffffff
7FFFFFFFF * 7FFFFFFFF
result = 62 18446744004990074881
result = 3E FFFFFFF000000001
MISMATCH!
result = 63 18446744004990074881
result = 3F FFFFFFF000000001

After fooling with it for awhile, I decided that the cause of the
problems was basically not thinking carefully about the lower half
of the value being unsigned: that affects when to do carries in
the addition macro, and we also have to be careful about whether
or not to sign-extend the partial product terms. The second
attached file is a version that I can't break anymore, though I'm
not quite sure it's bug-free.

regards, tom lane