>> Inside of pg_atoi(), the value is read into a long. Comparing
>> a small positive long like 24 against INT_MIN returns TRUE -
>> dunno how. Putting INT_MIN into another long variable and
>> comparing the two returns the expected FALSE - so what's
>> going on here? long, int32 and int have all 4 bytes here.
I believe the problem is that the compiler is deciding that INT_MIN
is of type "unsigned int" or "unsigned long", whereupon the type
promotion rules will cause the comparison to be done in
unsigned-long arithmetic. And indeed, 24 < 0x80000000 in unsigned
arithmetic. When you compare two long variables,
you get the desired behavior of signed long comparison.
Do you have <limits.h>, and if so how does it define INT_MIN?
The default INT_MIN provided at the top of numutils.c is clearly
prone to cause this problem:#ifndef INT_MIN#define INT_MIN (-0x80000000L)#endif
If long is 32 bits then the constant 0x80000000L will be classified
as unsigned long by an ANSI-compliant compiler, whereupon the test
in pg_atoi fails.
The two systems I have here both avoid this problem in <limits.h>,
but I wonder whether you guys have different or no <limits.h>.
I recommend a two-pronged approach to dealing with this bug:
1. The default INT_MIN ought to read#define INT_MIN (-INT_MAX-1)
so that it is classified as a signed rather than unsigned long.
(This is how both of my systems define it in <limits.h>.)
2. The two tests in pg_atoi ought to readif (l < (long) INT_MIN)...if (l > (long) INT_MAX)
The second change should ensure we get a signed-long comparison
even if <limits.h> has provided a broken definition of INT_MIN.
The first change is not necessary to fix this particular bug
if we also apply the second change, but I think leaving INT_MIN
the way it is is trouble waiting to happen.
Bruce, I cannot check this since my system won't show the failure;
would you un-reverse-out the prior patch, add these changes, and
see if it works for you?
regards, tom lane