[Mlir-commits] [mlir] [mlir] Add the ability to override attribute parsing/printing in attr-dicts (PR #103304)

Andrzej WarzyƄski llvmlistbot at llvm.org
Thu Aug 15 03:07:56 PDT 2024


banach-space wrote:

> I did propose moving the attribute out of the `attr-dict`, but people seemed generally against a change of syntax.

IMHO, we should avoid a situation where some named attributes are inside `attr-dict`, while others are outside:
```mlir
vector.transfer_read %arg0[%0, %1], %cst, in_bounds = [true] { permutation_map = #map3} : memref<12x16xf32>, vector<8xf32>
```

Unless there's some rationale for that other than pretty-printing? 

Btw, it sounds like we should take a step back here and look at the bigger picture:

> we have some cleanup to do in terms of syntax as with the introduction of properties we should really move all the inherent attributes outside of the attr-dict printing anyway

I'll need to get back to your presentation on properties before I have an opinion. Going back to #100336, I worry that we might be too attached to the current syntax. This would be perfectly fine with me:

```mlir
vector.transfer_read %arg0[%0, %1], %cst {in_bounds = array<i1: true>, permutation_map = #map3} : memref<12x16xf32>, vector<8xf32>
```

https://github.com/llvm/llvm-project/pull/103304


More information about the Mlir-commits mailing list