Add ind_to_const to enable fewer equivalence classes.

[DrTodd13]

Resolves #4963

[DrTodd13]

@stuartarchibald This patch seems to work for your test case and passes array analysis and parfor testing for me locally. This is a moderately scary change. Let me know what you think and how you want to progress.

[stuartarchibald]

Thanks for the patch @DrTodd13, much appreciated.

This is a moderately scary change. Let me know what you think and how you want to progress.

What are you most worried about? Getting the const var scopes right (local vs global)?

[DrTodd13]

Thanks for the patch @DrTodd13, much appreciated.

This is a moderately scary change. Let me know what you think and how you want to progress.

What are you most worried about? Getting the const var scopes right (local vs global)?

I'm not worried about scoping. It is just a feeling about the degree of the change in combination with what triggered it which is new kinds of nodes being exposed in ways which previously weren't. Anytime you add some new data structure to this alias analysis code I think that is a significant change. I thought that might be a more isolated change than forcing both vars and consts into the ind_to_var dict.

[stuartarchibald]

Thanks for fixing this, much appreciated. I understand your concerns over the addition of handling new node kinds, but unfortunately I don't think there's a equivalent/cheaper alternative available. I checked out and tested this locally, fix seems good. Only comment on the code is that it'd be good to add a test for freevars in the same circumstances (this triggers the same bug). Thanks again!

[stuartarchibald]

I think it'd be good to also add a freevar variant:

    _FREEVAR_INT_FOR_TESTING1 = 17
    _FREEVAR_INT_FOR_TESTING2 = 5

    @skip_unsupported
    def test_issue4963_freevars(self):
        def test_impl():
            buf = np.zeros((_FREEVAR_INT_FOR_TESTING1, _FREEVAR_INT_FOR_TESTING2))
            return buf
        self.check(test_impl)

[DrTodd13]

Merged.

[stuartarchibald]

Thanks!

[stuartarchibald]

@DrTodd13 Added IntelLabs#65 which implements the above, if you are happy with it please merge and this PR can be considered complete! Thanks.