[Mlir-commits] [mlir] [mlir] Speed up resolving ExtractValueOp source (PR #176478)

[llvmlistbot at llvm.org]

neildhar wrote:

I don't think so. The cases I considered are:

1. Each `InsertValueOp` has a huge number of users that are not `ExtractValueOp` or are `ExtractValueOp`s that cannot be resolved. In this case, the work is roughly proportional to the number of users. So for a single pass over the IR, it is bounded by the number of edges in the IR.
2. Each `InsertValueOp` has just a single `ExtractValueOp`. In this case, we have to separately do a traversal over the chain for each `ExtractValueOp`, but this is no worse than the current `ExtractValueOp` folder.

https://github.com/llvm/llvm-project/pull/176478