[mlir][LLVM] Allow types other than ptr to have alignstack

[frabert]

No description provided.

[Jezurko]

Hm, I don't understand why is this necessary? Is this a bug in upstream? Or an issue specific to us?

[frabert]

Hrmm I guess it would make sense to try and upstream this 🤔

Basically I've encountered things like

"llvm.func"() <{CConv = #llvm.cconv<ccc>, arg_attrs = [{llvm.alignstack = 8 : i64, llvm.noundef}], dso_local, frame_pointer = #llvm.framePointerKind<"non-leaf">, function_type = !llvm.func<void (array<2 x f64>)>, linkage = #llvm.linkage<external>, no_inline, no_unwind, optimize_none, passthrough = [["uwtable", "2"], ["no-trapping-math", "true"], ["stack-protector-buffer-size", "8"], ["target-cpu", "generic"]], sym_name = "emit", target_cpu = "generic", target_features = #llvm.target_features<["+fp-armv8", "+neon", "+outline-atomics", "+v8a", "-fmv"]>, visibility_ = 0 : i64}> ({

where the attr is applied to an array, which is apparently fine in LLVM but considered invalid in the LLVM MLIR dialect

[Jezurko]

Oh, yeah. In that case I would probably vote for making an upstream PR and then backporting it to our repo? I think we should avoid diverging from upstream in these cases, to make our life simpler (and at least we get a little bit of street cred for contributing)