I don’t mind reverting if given specific concerns applicable to the patch in question. Otherwise, reverting it will not address those concerns and will incur even more churn on everybody.
In particular, I have seen the following:
- We don’t want
unrealized_conversion_castto “leak” across passes. This was already happening before this patch. The objection is to either splitting the std or unrealized cast itself. In the worst case, it could apply to the first split-out of std-to-llvm, probably memref-to-llvm (which incidentally also changed the same tests), but it landed a couple of months ago with no reaction whatsoever. - The conversion should run the folding hook differently that it does. The patch in question does not change how the conversion infrastructure works.
- There is another way achieve the same behavior. I am yet so see a concrete proposal how, preferably with code. I have repeatedly demonstrated that “it should be possible with local folding” is incorrect beyond a purely theoretical argument and is based on a misunderstanding of how conversions work.
The only concern that applies to the patch in question is that it landed too fast for somebody’s taste. I would actually be happy if all reviews were faster.