Diego Caballero

Generating an alloca shouldn't be a problem if it's not causing other issues, again, as an initial step. LLVM vectorizer does that as well. It could be that LLVM is...

An alloca is generated by bufferization but it could be doing the right thing. Providing an example so that @matthias-springer can take a look would be very helpful. Worse case,...

Thanks, Uday! Yes, it was a temporary (but very convenient) solution :). Agree it's very unfortunate that we didn't have testing coverage for that in MLIR. This has become a...

Thank you all for the feedback! I think we all agree on that introducing an alias model is the long-term solution, also regardless of this particular issue. Let me investigate...

I implemented my own LLVMTypeConverter with a custom lowering for MemRef type. However, that was not enough since Std-to-LLVM conversion uses MemRefDescriptor class all over the place, which assumes that...

Sorry, I saw this comment after replying to the PR. > Why are you not happy with it? Basically, what I said in the PR. Of course we can always...

This sounds pretty much the same kind of problem as the one discussed here https://github.com/tensorflow/mlir/pull/55#issuecomment-518134726. I think we should try to follow the same approach for both. I see the...

Or you could fold that into a select that is merging `%40` and `%cst_0`. This is pretty unfortunate as it's a limitation of `vector.transfer_read`. We can't say "I don't care...

@hanhanW, @MaheshRavishankar this is the `memref.subview` error we talked about. Also pinging @Max191. This is currently blocking our DT+UK enablement for i4.

Is there a reason not to fold when `in_bounds` is `false`? https://github.com/shark-infra/llvm-project/blob/6a22c340976abaa0a65f580f8e83fd8ea1593b95/mlir/lib/Dialect/MemRef/Transforms/FoldMemRefAliasOps.cpp#L351 I'm trying to understand what would be needed to support this case...