[Mlir-commits] [llvm] [mlir] [mlir] split transform interfaces into a separate library (PR #85221)

[Stella Laurenzo]

stellaraccident wrote:

I have a mental model of what seems to work out when using TD, and I have a fair bit of practical observations of what is off label usage that we end up having to back out and write real pass pipelines for. 

Right now, it is effectively a tool for writing a specific "granularity" of pass in more of a meta-programmed way. Most of the issues I see with it come from it having no guard rails on recommended use (and then people use it for things it will be bad at, someone has to come along and redo months of work, and then bad feelings emerge). It's positioning as a piece of core infra adds to that because, probably unintentionally, it gets assumed to be more of "the way" than it is really positioned to serve.

My point around encapsulation was that such a thing can succeed very well if it is a self contained entity that is really good at what it does and people reach for it when they know they need that thing. I need to look more closely vs just making a fly by comment, but the amount of touch points that implementing this thing has on everything else tells us something. And I think that the answer is not necessarily to move it closer to the core infra. I'm not sure what it is but we're missing something here if that is the way that such a thing must exist.

https://github.com/llvm/llvm-project/pull/85221