Unrealized_conversion_cast failing bufferization passes

[gcmn]

@herhut

This bounded amount of time can be relatively long sometimes and we should not require those ops to not exist. For instance, one might have done the signed/signless conversion on the linalg op but not on the compute operations in its body.

But it does not matter for this bug anyway. linalg.generic bufferization just should support ops from other dialects.

Fair. that makes sense. I think bounding that time ties in with PSA: run -reconcile-unrealized-casts after all -convert-*-to-llvm from now on

@youben

The way we solved it was by introducing a pass that removes all those unrealized_conversion_cast and simplify the types for the bufferization one

Just because it’s not clear from your examples, I hope you don’t mean a pass that operates on unrealized_conversion_cast directly and does something other than folding away cast A → B followed by cast B → A.

%1 = builtin.unrealized_conversion_cast %0 : i32 to index
%2 = builtin.unrealized_conversion_cast %1 : index to i32
%3 = "foo.op"(%2) : (i32) -> i32

→

%3 = "foo.op"(%0) : (i32) -> i32

because those are the only semantics that unrealized_conversion_cast has