[Mlir-commits] [mlir] [mlir][transform] Don't modify the target in interpreter when loading library. (PR #67686)

[Ingo Müller]

ingomueller-net wrote:

Two questions about this PR:
* Should the loading/injection pass be part of this PR or can I do it in an independent one?
* Let us also reconsider the alternative of keeping loading and interpretation together. The [issue](https://github.com/llvm/llvm-project/pull/67560#issuecomment-1737611467) that led to this PR could also be solved by not expecting the possibility to run the interpreter twice *and supply the same library in the second invocation*. If the semantic is "load and interpreter", then doing "load and interpret" twice implies doing "load" twice, for which it is obvious that it leads to doubly defined symbols. Similarly, is the semantic is "interpret but don't load", then running the interpreter once with a library name and then once *without* (as one of the tests did) will break with the current PR.

   All in all, as I see it, the decision is between having more precise control (with two different passes) or having fewer passes. What's more desirable?

https://github.com/llvm/llvm-project/pull/67686