Optimized for loop & overflow

[jvasileff]

Neither of these produce output:

    void spanTest(Integer min) {
        variable value x=0;
        for (i in min..runtime.maxIntegerValue) {
            print(i);
            if (x++ > 2) {
                break;
            }
        }
    }
    spanTest(-2);

    void measureTest(Integer min) {
        variable value x=0;
        for (i in min:runtime.maxIntegerValue) {
            print(i);
            if (x++ > 2) {
                break;
            }
        }
    }
    measureTest(1);

This relates to ceylon/ceylon.language/issues/656.

[tombentley]

Fixed for iteration over measures. Spans next.

[tombentley]

Optimizing span iteration turns out to be much harder. Specifically getting a C-style for loop to behave the same as the iterator would boils down to replicating the same logic as Span.iterator() and Span.By for increment and loop exit conditions. An additional problem is by() special cases step==1 and the overflow behaviour of Span.iterator() and Span.By.iterator() are different. This means that to generate correct code for for (i in (start..end).by(step)) the loop condition and increment depend on the step. I'm effectively teaching the compiler how to inline all the code used for the iterator.

[gavinking]

@tombentley Well, yes, but the difference is you don't need to box the longs.

[tombentley]

Still a couple of edge cases to fix here, but I need to work on something else today.

[gavinking]

@tombentley good ;)

[jvasileff]

I'm guessing related to this work, I'm getting:

    print(3 in 0:3); //true
    print((0:3).contains(3)); //false

[tombentley]

I think this is fixed with the exception of those Span affected by ceylon/ceylon.language#717, where the unoptimized loop never executes because the Span constructor throws, but the optimized loop is executing at least once. I suppose I can add a check for those cases, but I'll wait to see if people care about fixing ceylon/ceylon.language#717 first.

[tombentley]

Add immediately after I made that comment I find something else which is broken.

[tombentley]

I have at this point concluded we're flogging a dead horse in trying to squeeze the span iteration optimization into a for loop, and that a do/while loop is the much more natural construct, though it will complicate the transformation of continue directives with the iterated block.

[tombentley]

Fixed, finally.