Anything that is a @CoreMethod can use lowerFixnumParameters. In the case of MatchDataNodes, the lowerFixnumParameters is present, but is pointing at the wrong parameter index. Note that the index is 0-based and does not include self. I messed this up originally and it looks like you found a couple lingering cases there.

-Xtruffle.randomize_storage.array=true can help you find these problems - it randomly choose a way to represent arrays whenever they are modified - including the rare-in-practice long[].

I hate complicating this, but it looks on MRI, Integer#downto and Integer#upto work fine on values > 32-bit. So casting here would be wrong. @chrisseaton should we just effectively duplicate that code for (long, long), (long, int), and (int, long) permutations?

Yes casting long to int without checking the range is absolutely wrong - at least the first specialisation in this PR does that. It will silently truncate, which is the worst behaviour we could have.

Casting int to long is of course fine and often what we want to do.

Note that lowerFixnumParameters only lowers a long to an int if it is in that range. Otherwise it passes it through as a long.

@chrisseaton what would be the error to raise if the long is outside of the int range for a methods that only currently support int?

could the the same raise exception be thrown from a primitive node?

I think the key one is array indexing - as that is restricted to int by limitations of the JVM. JRuby seems to allow reads from an array with any index, and returns nil. Writes it gives an error that I'm not sure if MRI ever gives.

$ ~/.rbenv/versions/jruby-1.7.19/bin/ruby -e '[1, 2][0xffffffffffffff]'
$ ~/.rbenv/versions/jruby-1.7.19/bin/ruby -e '[1, 2][0xffffffffffffff] = 14'
ArgumentError: index too big
     []= at org/jruby/RubyArray.java:1408
  (root) at -e:1

For a Bignum MRI does give an error message.

$ ruby -e '[1, 2][10000000000000000000000000]'
-e:1:in `[]': bignum too big to convert into `long' (RangeError)
    from -e:1:in `<main>'

Note that you can index an array with some Bignums, as there are some values of the C type long, which is what MRI indexes arrays with, that do not fit into a Fixnum due to tagging.

Short answer - I think it depends on the method! I would concentrate on the specs for now - if the specs cover a case then fix it. Otherwise I am working on some stress test utilities that will help us explore this problem space more in the future.

Yeah primitives could have the lower annotation as well.

I've updated the PR to just fix String#byteslice #2843. This is used a lot and fixes a stackoverflow.