Hi,
On 2019-07-17 17:29:58 +0900, Michael Paquier wrote:
> Actually, one thing which may be a problem is that we lack currently
> the equivalents of pg_mul_s16_overflow and such for unsigned
> integers.
It's much simpler to implement them for unsigned than for signed,
because unsigned overflow is well-defined. So I'd not be particularly
worried about just adding them. E.g. comparing the "slow" version of
pg_mul_s64_overflow() with an untested implementation of
pg_mul_u64_overflow():
pg_mul_s64_overflow:
/*
* Overflow can only happen if at least one value is outside the range
* sqrt(min)..sqrt(max) so check that first as the division can be quite a
* bit more expensive than the multiplication.
*
* Multiplying by 0 or 1 can't overflow of course and checking for 0
* separately avoids any risk of dividing by 0. Be careful about dividing
* INT_MIN by -1 also, note reversing the a and b to ensure we're always
* dividing it by a positive value.
*
*/
if ((a > PG_INT32_MAX || a < PG_INT32_MIN ||
b > PG_INT32_MAX || b < PG_INT32_MIN) &&
a != 0 && a != 1 && b != 0 && b != 1 &&
((a > 0 && b > 0 && a > PG_INT64_MAX / b) ||
(a > 0 && b < 0 && b < PG_INT64_MIN / a) ||
(a < 0 && b > 0 && a < PG_INT64_MIN / b) ||
(a < 0 && b < 0 && a < PG_INT64_MAX / b)))
{
*result = 0x5EED; /* to avoid spurious warnings */
return true;
}
*result = a * b;
return false;
pg_mul_s64_overflow:
/*
* Checking for unsigned overflow is simple, just check
* if reversing the multiplication indicates that the
* multiplication overflowed.
*/
int64 res = a * b;
if (a != 0 && b != res / a)
{
*result = 0x5EED; /* to avoid spurious warnings */
return true;
}
*result = res;
return false;
The cases for addition/subtraction are even easier:
addition:
res = a + b;
if (res < a)
/* overflow */
subtraction:
if (a < b)
/* underflow */
res = a - b;
Greetings,
Andres Freund
