What's the appropriate approach to #1157?

[SainoNamkho]

Suppose I'm writing an insertion sort

void insertSort(span<int> sp)
{
    for (int i = 0; i < sp.size(); ++i) {
        auto insert_pos = ranges::upper_bound(sp.first(i), sp[i]);
        if (insert_pos == sp.first(i).end()) {
            int temp = sp[i];
            ranges::copy(span{insert_pos, sp.begin()+i}, insert_pos+1);
            *insert_pos = temp;
        }
    }
}

It's wrong since insert_pos, sp.first(i).end() and sp.begin() are from three different span objects.
#1157 suggest to do this with ranges.

void insertSort(span<int> sp)
{
    for (auto it = sp.begin(); it != sp.end(); ++it) {
        auto insert_pos = ranges::upper_bound(ranges::subrange{sp.begin(), it}, *it);
        if (insert_pos == it) {
            int temp = *it;
            ranges::copy(ranges::subrange{insert_pos, it}, insert_pos+1);
            *insert_pos = temp;
        }
    }
}

But four years since that issue, clang still cannot compile it. llvm/llvm-project#52696 says it's fixed, while actually it just do not compile.

So what are we supposed to do? Just sit down and wait for clang 15 to fix it?

[jwakely]

Yes, or use a different compiler.

The future is already here, it's just not evenly distributed.

[jwakely]

You can also convert the span iterators to pointers and compare those, so that any debug checking on the iterators doesn't complain about iterators to different ranges.

[SainoNamkho]

You can also convert the span iterators to pointers and compare those, so that any debug checking on the iterators doesn't complain about iterators to different ranges.

Thanks a lot, maybe I'll go back to the original unconstrained algorithms, since such convertion needs care of subspan's past the end iterator.