Therefore I don't really see the point with the change in 7.8. Note also that a mapping cannot carry UNDEFINED as a value for natural reasons, so in Sql.query() those UNDEFINED's still becomes normal zeroes.
Aaargh. Didn't know that.
That's the raison d'être for UNDEFINED, to be able to see if a mapping lookup failed to find the key. Unfortunately that semantic has gotten severely muddled since many are very anxious to use it as a sort of NULL value (the name "UNDEFINED" is actually rather well chosen).
Well, actually, a cursory check amongst database backends in the
current (and older) Pike versions shows:
Returns
- Mysql string/float/int
Your check appears to be a little too cursory; this case is string only (since OLD_SQL_COMPAT is typically not defined). Although it's only one backend, it's in my world the by far most important one.
- Oracle string/float/int
Probably true. The Oracle driver has always been a bit odd. It's unfortunate that this semantic isn't uniform across all drivers. But still, string only is the standard. Much of the code I'm concerned with (e.g. Roxen) would break quite a bit if it's anything else. The Oracle module ought to be fixed, but it requires compat of course.
- pgsql string/float/int
This is your new one, right? Then, I'm sorry to say, it has little relevance in this discussion.
Well, in light of learning that mappings cannot contain UNDEFINEDs, I propose to introduce the following SqlNULL() type which should be returned by any mappingvalues from Sql.Sql()->query() which are actually an SQL NULL type.
An Sql.NULL value would be good, but this would also need new API calls, since I stubbornly maintain that query and big_query should return values as either string or 0. This has been up for discussion earlier, actually, and iirc the proposed names for those new functions were typed_query and big_typed_query, respectively.
+class SqlNULL +{
- constant is_SqlNULL = 1;
- // Treat these objects as indistinguishable from each other. We
- // ought to ensure that there's only one in the pike process
- // instead, but that's tricky to solve in the PCode codec.
- protected int `== (mixed other) {
- return zero_type(other)
|| objectp(other) && other->is_SqlNULL;- }
Thought I recognized that comment. ;) Actually, in a pike module that uniqueness problem shouldn't exist, so the kludge isn't necessary.
- protected int(0..1) _is_type(mixed type) {
- return 1;
- }
I'm not sure this is a good idea. Is there a good palpable reason to "lie" in type checks like this? Because if there isn't, then the default, I think, is "don't".
- protected mixed cast(string to) {
- mixed ret=UNDEFINED;
- switch(to) {
case "string":- ret="";
- break;
case "float":- ret=Math.nan;
- break;
case "mixed":case "object":- ret=this;
- }
- return ret;
- }
Casting to unsupported types ought to throw errors, not silently return UNDEFINED.
- protected int __hash() {
- return 17;
- }
This is the second part of the kludge above. It can be removed too.