[TIR] Simplify indices in layout transform

[masahi]

Without this, a sequence of blockize -> transform_layout -> tensorize would likely lead to a tensorize pattern match failure, since there is no simplification done between transform_layout -> tensorize but intrinsic descriptions are always written without taking extra factors like % 16 into account.

Co-authored-by: Yuanjing Shi [email protected]

@vinx13 @shingjan

[shingjan]

LGTM! Thanks for upstreaming this. Kudos to @vinx13 for helping troubleshooting this problem.

[Lunderberg]

Hmm, this is interesting. The IndexMapNode::MapIndices function already includes a call to simplify, which I would have expected to handle this case. For the test case, it looks like the first simplification didn't have the information that 0 <= vi < 16, and so it couldn't simplify floordiv(vi,16) to 0.

Since we could have similar issues in other cases, I think we could generalize it by allowing MapIndices to take an optional arith::Analyzer* to be used in making the simplifications. I'll see if that is a straightforward change to make.