[TIR][BugFix] Do not bind non-index type value of lets in CompactBufferAllocation

[wrongtest-intellif]

Fix a bug for let support, it should not put non-index typed values into dom_map.

[junrushao]

Hey thanks for the contribution! I don't think I fully comprehend this change (and the unittest) so would love to hear more about your insights :-) With this change, whether or not to add a tir::Var into dom_map is based on its dtype. Why is it the case?

[wrongtest-intellif]

Hey thanks for the contribution! I don't think I fully comprehend this change (and the unittest) so would love to hear more about your insights :-) With this change, whether or not to add a tir::Var into dom_map is based on its dtype. Why is it the case?

Emmmmm，my understand is that since the pass is only interested in buffer indices，record index typed vars could be safe. While arith utilities are written to handle integers，binding things like SinglePoint(int32x8) would just fail on sanity checks of dom analysis (eg, EvalNDSet).

[junrushao]

Agree that we shouldn’t bind vectorized data types. Besides that, is that anything else we shouldn’t bind? If not, we can instead only check lanes field

[junrushao]

Perhaps we can add an assertion in the Bind method where we always assume the value shouldn’t be vectorized

[wrongtest-intellif]

is that anything else we shouldn’t bind

And kVoid and kHandle should not bind.

• kVoid: lanes=0
• kHandle: lanes can be arbitrary

IsIndexType also excludes

• float, uint: bind them into dom_map or into analyzer seems cause no error, but maybe effectless in integer analysis about buffer indices.

[wrongtest-intellif]

Perhaps we can add an assertion in the Bind method where we always assume the value shouldn’t be vectorized

Thanks for the notes~ I'd like to add one ICHECK line to see if there are broken cases. For vectors, however, I'm not sure it should never be considered to get analyzed/simplified. For CompactBufferAllocation, just integer analysis part (IntSets) should not work with vectors.

[junrushao]

@wrongtest Thank you for your detailed explanation! That makes a lot of sense to me :-)

[junrushao]

Looks like some situations happened in the CI...

[junrushao]

@wrongtest would you mind retriggering the CI? looks like it's down

[junrushao]

Retriggering as CI is back to life