[Mlir-commits] [mlir] [mlir][DLTI] Introduce an operation interface for scoped DLTI queries (PR #232733)

[Mehdi Amini]

================

----------------
joker-eph wrote:

What do you mean with "downstream users are broken"? We serialize a newer version of builtin so older readers are broken? If so yes but that's not a bug I believe.

Looking at the test more in depth. This file is used by `mlir/test/Dialect/Builtin/Bytecode/builtin_fixed.mlir`:

```
// RUN: mlir-opt -allow-unregistered-dialect -emit-bytecode -emit-bytecode-producer=fixed %s -o %t
// RUN: cmp -s %t %p/builtin_fixed_0.mlirbc
```

The issue here is that the emission does not fix a version of the bytecode, neither does it fix a version of the builtin dialect. So this test is just badly designed right now I think.

We could fix the bytecode version emitted here with mlir-opt, but we don't have a command line option to set the version of a dialect (here builtin).

This PR bumped the version of the builtin dialect from 0 to 1, and while I think you can programmatically emit version 0, mlir-opt lacks such an option.

https://github.com/llvm/llvm-project/pull/232733