WIP experiment with using isl to simplify affine exprs

[ivanradanov]

Work in progress, currently only dimps the isl expressions and their simplified versions. the changes need to be brought back to mlir

An example of a good simplification we achieved:

 Got aff
{ [i0, i1, i2, i3, i4, i5] -> [(7 + floor((i0)/256) + floor((96 + i0 + 240*floor((i0)/23) + 112*floor((i0)/184) - 256*floor((i0)/256) + 16*floor((7 + 17i0 - 16*floor((i0)/16) + 6*floor((i0)/23))/374))/2160))] }
Gisted aff
{ [i0, i1, i2, i3, i4, i5] -> [(7 + floor((i0)/184) + floor((i0)/256))] }

But we do not manage to simplify everything, for example:

Got aff
{ [i0, i1, i2, i3, i4, i5] -> [(7 + floor((i0)/16) + 15*floor((i0)/23) + 7*floor((i0)/184) - 16*floor((i0)/256) + floor((7 + 17i0 - 16*floor((i0)/16) + 6*floor((i0)/23))/374) - 135*floor((112 + i0 + 240*floor((i0)/23) + 112*floor((i0)/184) - 256*floor((i0)/256) + 16*floor((7 + 17i0 - 16*floor((i0)/16) + 6*floor((i0)/23))/374))/2160))] }
Gisted aff
{ [i0, i1, i2, i3, i4, i5] -> [(7 + floor((i0)/16) + 15*floor((i0)/23) + 7*floor((i0)/184) - 16*floor((i0)/256) + floor((7 + 17i0 - 16*floor((i0)/16) + 6*floor((i0)/23))/374) - 135*floor((112 + i0 + 240*floor((i0)/23) + 112*floor((i0)/184) - 256*floor((i0)/256) + 16*floor((7 + 17i0 - 16*floor((i0)/16) + 6*floor((i0)/23))/374))/2160))] }

[wsmoses]

can you mark the formatter as not running on external?

[ivanradanov]

@ftynse Do you have any insight on the AffineExpr floordiv/ceildiv/mod semantics and how they match the isl_ast_expr semantics regarding whether they floor to -inf or 0 or other edge cases I have not thought of?

[ivanradanov]

Merged manually.