On Thu, Aug 11, 2022 at 10:40 AM Tom Lane <tgl@sss.pgh.pa.us> wrote:
Zhihong Yu <zyu@yugabyte.com> writes: > In cash_out(), we have the following code: > if (value < 0) > { > /* make the amount positive for digit-reconstruction loop */ > value = -value;
> The negation cannot be represented in type long when the value is LONG_MIN.
Possibly not good, but it seems like the subsequent loop works anyway:
Note that this exact test case appears in money.sql, so we know that it works everywhere, not only my machine.
> It seems we can error out when LONG_MIN is detected instead of continuing > with computation.
How could you think that that's an acceptable solution? Once the value is stored, we'd better be able to print it.
regards, tom lane
Thanks for taking a look.
I raise this thread due to the following assertion :
src/backend/utils/adt/cash.c:356:11: runtime error: negation of -9223372036854775808 cannot be represented in type 'Cash' (aka 'long'); cast to an unsigned type to negate this value to itself
Though '-92233720368547758.085'::money displays correct error message in other builds, this statement wouldn't pass the build where UndefinedBehaviorSanitizer is active.
I think we should fix this otherwise when there is new assertion triggered due to future changes around Cash (or other types covered by money.sql), we wouldn't see it.
I am open to other ways of bypassing the above assertion.