Eric Blake <[email protected]> writes: > On 07/05/2017 08:48 AM, Max Reitz wrote: >>>> /** >>>> + * qnum_is_equal(): Test whether the two QNums are equal >>>> + */ >>>> +bool qnum_is_equal(const QObject *x, const QObject *y) >>>> +{ >>>> + QNum *num_x = qobject_to_qnum(x); >>>> + QNum *num_y = qobject_to_qnum(y); >>>> + >>>> + switch (num_x->kind) { >>>> + case QNUM_I64: >>>> + switch (num_y->kind) { >>>> + case QNUM_I64: >>>> + /* Comparison in native int64_t type */ >>>> + return num_x->u.i64 == num_y->u.i64; >>>> + case QNUM_U64: >>>> + /* Implicit conversion of x to uin64_t, so we have to >>>> + * check its sign before */ >>>> + return num_x->u.i64 >= 0 && num_x->u.i64 == num_y->u.u64; >>>> + case QNUM_DOUBLE: >>>> + /* Implicit conversion of x to double; no overflow >>>> + * possible */ >>>> + return num_x->u.i64 == num_y->u.dbl; >>> >>> Overflow is impossible, but loss of precision is possible: >>> >>> (double)9007199254740993ull == 9007199254740992.0 >>> >>> yields true. Is this what we want? >> >> I'd argue that yes, because the floating point value represents >> basically all of the values which are "equal" to it. > > But the problem is that we CAN represent the fully-precise number as an > integer, so having multiple distinct integers that compare differently > against each other, but equal to the same double, is awkward.
Yup. >> But I don't have a string opinion. I guess the alternative would be to >> convert the double to an integer instead and check for overflows before? > > That's the solution Markus gave, and I'm in favor of the tighter check: > [...] >>> I think there's more than one sane interpretations of "is equal", >>> including: >>> >>> * The mathematical numbers represented by @x and @y are equal. >>> >>> * @x and @y have the same contents, i.e. same kind and u. >>> >>> * @x and @y are the same object (listed for completeness; we don't need >>> a function to compare pointers). >>> >>> Your patch implements yet another one. Which one do we want, and why? >> >> Mine is the first one, except that I think that a floating point value >> does not represent a single number but just some number in a range. >> >>> The second is easier to implement than the first. >> >> It seems much less useful, though. Depends on what for. Common Lisp has * eq: same object * eql: eq, or both numbers with same type and value, or both characters that represent the same character * =: mathematically equal (arguments must be numbers) * equal: like eql, but it recursively descends into lists (I'm simplifying) Decades of use back the assertion that eq, eql and = are all useful. >>> If we really want the first, you need to fix the loss of precision bugs. >> >> I'm not sure, but I don't mind either, so... For what it's worth, your v2 had QInt compare not equal to QFloat. Makes me suspect that eql is good enough for the problem at hand. >>> I guess the obvious fix is >>> >>> return (double)x == x && x == y; >> >> Yes, that would do, too; and spares me of having to think about how well >> comparing an arbitrary double to UINT64_MAX actually works. :-) > > It basically says that we are unwilling to declare an integer equivalent > to the double if the double loses precision when trying to store the > integer. > >>> Note that this is what you do for mixed signedness: first check @x is >>> exactly representable in @y's type, then compare in @y's type. >>> >>> Regardless of which one we pick, the function comment needs to explain.
