Add "test-transform-dialect-interpreter" to pass manager in MLIR python binding

MLIR
1

I’m trying to add test-transform-dialect-interpreter to a pass manager using MLIR Python binding:

passes="builtin.module(test-transform-dialect-interpreter)"
pm = PassManager.parse(passes)
pm.run(mod.operation)

However, test-transform-dialect-interpreter is not a registered pass, the above Python file comes with the following error:

error: 'test-transform-dialect-interpreter' does not refer to a registered pass or pass pipeline.

The above pipeline can work using the mlir-opt:

./build/bin/mlir-opt -pass-pipeline="builtin.module(test-transform-dialect-interpreter)"

Does anyone know how to add this in MLIR Python binding? Thanks!

2

Test passes are intended for testing and therefore (mainly) not exposed to the bindings. You can use the non-test transform-interpreter pass instead. The only difference is that the entry point is an explicit symbol, typically a transform.named_sequence called __transform_main by default, which can be overridden through pass options.

3

Putting on my “I’m a pedant” hat, it’s not that those passes aren’t exposed to the bindings (we don’t have a way to selectively filter passes that the bindings can see), it’s that those passes aren’t by default linked into anything (not mlir-opt, not AggregateCAPI, etc). If you do decide you want them, you need to rebuild LLVM+MLIR accordingly mlir-wheels/patches/register_test_pass.patch at main · makslevental/mlir-wheels · GitHub. But I agree with Alex, just use transform-interpreter (roughly like this).