Turn constant folder and dce into passes

[justinchuby]

Turn constant folder and dce into passes to allow them to be used as individual passes in the future.

@gramalingam

[Copilot]

Pull Request Overview

This PR refactors the optimizer by turning the constant folding and dead code elimination logic into individual passes. Key changes include:

• Renaming and refactoring the unused nodes removal functionality into a new RemoveUnusedNodesPass class.
• Replacing the old ConstantFolder class with a new FoldConstantsPass class that conforms to the PassBase API.
• Updating related call sites and ensuring modification tracking is consistently applied.

Reviewed Changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated no comments.

File	Description
onnxscript/optimizer/_remove_unused.py	Renames an internal function and replaces a standalone function with a class.
onnxscript/optimizer/_constant_folding.py	Refactors constant folding into a pass-based API and updates modification logic.

Comments suppressed due to low confidence (2)

onnxscript/optimizer/_remove_unused.py:58

• [nitpick] The renaming of 'process_function_or_graph' to '_process_function_or_graph' clarifies its internal usage. Consider adding a brief docstring to summarize its behavior for future maintainers.

def _process_function_or_graph(function_or_graph: ir.Function | ir.Graph) -> int:

onnxscript/optimizer/_constant_folding.py:1016

• Ensuring the modification flag is set after node replacement is crucial; verify that similar modifications in other code paths consistently update this flag.

self.modified = True

[codecov]

❌ 11 Tests Failed:

Tests completed	Failed	Passed	Skipped
14767	11	14756	2221

View the top 3 failed test(s) by shortest run time

onnxscript.backend.onnx_export_test.TestOnnxBackEnd::test_export2python_produces_correct_onnx_script_model_0032_test_and_bcast3v2d

Stack Traces | 0.003s run time

onnxscript\backend\onnx_export_test.py:137: in extract_functions
    mod = importlib.import_module(import_name)
C:\hostedtoolcache\windows\Python\3.11.9\x64\Lib\importlib\__init__.py:126: in import_module
    return _bootstrap._gcd_import(name[level:], package, level)
E   ModuleNotFoundError: No module named 'tests.onnx_backend_test_code.test_and_bcast3v2d'

The above exception was the direct cause of the following exception:
.nox\test_onnx_weekly\Lib\site-packages\parameterized\parameterized.py:620: in standalone_func
    return func(*(a + p.args), **p.kwargs, **kw)
onnxscript\backend\onnx_export_test.py:271: in test_export2python_produces_correct_onnx_script_model
    functions = extract_functions(backend_test.name, code, self.test_folder)
onnxscript\backend\onnx_export_test.py:139: in extract_functions
    raise AssertionError(
E   AssertionError: Unable to import 'tests.onnx_backend_test_code.test_and_bcast3v2d' (e=No module named 'tests.onnx_backend_test_code.test_and_bcast3v2d') (file: 'D:\\a\\onnxscript\\onnxscript\\tests\\onnx_backend_test_code\\test_and_bcast3v2d.py', absolute path: 'D:\\a\\onnxscript\\onnxscript\\tests\\onnx_backend_test_code\\test_and_bcast3v2d.py', current folder: D:\a\onnxscript\onnxscript
E   ---- CONTENT --
E   import numpy
E   from onnx import TensorProto
E   from onnx.helper import make_tensor
E   from onnxscript import script, external_tensor
E   from onnxscript.values import Opset
E   from onnxscript.onnx_types import BOOL
E   from onnxscript.onnx_opset import opset7
E   
E   @script()
E   def bck_test_and_bcast3v2d(x: BOOL[3,4,5], y: BOOL[4,5]) -> (BOOL[3,4,5]):
E       r_and = opset7.And(x, y)
E       return r_and

onnxscript.backend.onnx_export_test.TestOnnxBackEnd::test_export2python_produces_correct_onnx_script_model_1032_test_scatter_without_axis

Stack Traces | 0.003s run time

onnxscript\backend\onnx_export_test.py:137: in extract_functions
    mod = importlib.import_module(import_name)
C:\hostedtoolcache\windows\Python\3.11.9\x64\Lib\importlib\__init__.py:126: in import_module
    return _bootstrap._gcd_import(name[level:], package, level)
E   ModuleNotFoundError: No module named 'tests.onnx_backend_test_code.test_scatter_without_axis'

The above exception was the direct cause of the following exception:
.nox\test_onnx_weekly\Lib\site-packages\parameterized\parameterized.py:620: in standalone_func
    return func(*(a + p.args), **p.kwargs, **kw)
onnxscript\backend\onnx_export_test.py:271: in test_export2python_produces_correct_onnx_script_model
    functions = extract_functions(backend_test.name, code, self.test_folder)
onnxscript\backend\onnx_export_test.py:139: in extract_functions
    raise AssertionError(
E   AssertionError: Unable to import 'tests.onnx_backend_test_code.test_scatter_without_axis' (e=No module named 'tests.onnx_backend_test_code.test_scatter_without_axis') (file: 'D:\\a\\onnxscript\\onnxscript\\tests\\onnx_backend_test_code\\test_scatter_without_axis.py', absolute path: 'D:\\a\\onnxscript\\onnxscript\\tests\\onnx_backend_test_code\\test_scatter_without_axis.py', current folder: D:\a\onnxscript\onnxscript
E   ---- CONTENT --
E   import numpy
E   from onnx import TensorProto
E   from onnx.helper import make_tensor
E   from onnxscript import script, external_tensor
E   from onnxscript.values import Opset
E   from onnxscript.onnx_types import FLOAT, INT64
E   from onnxscript.onnx_opset import opset10
E   
E   @script()
E   def bck_test_scatter_without_axis(data: FLOAT[3,3], indices: INT64[2,3], updates: FLOAT[2,3]) -> (FLOAT[3,3]):
E       y = opset10.Scatter(data, indices, updates)
E       return y

onnxscript.backend.onnx_export_test.TestOnnxBackEnd::test_export2python_produces_correct_onnx_script_model_1171_test_softmax_large_number

Stack Traces | 0.003s run time

onnxscript\backend\onnx_export_test.py:137: in extract_functions
    mod = importlib.import_module(import_name)
C:\hostedtoolcache\windows\Python\3.11.9\x64\Lib\importlib\__init__.py:126: in import_module
    return _bootstrap._gcd_import(name[level:], package, level)
E   ModuleNotFoundError: No module named 'tests.onnx_backend_test_code.test_softmax_large_number'

The above exception was the direct cause of the following exception:
.nox\test_onnx_weekly\Lib\site-packages\parameterized\parameterized.py:620: in standalone_func
    return func(*(a + p.args), **p.kwargs, **kw)
onnxscript\backend\onnx_export_test.py:271: in test_export2python_produces_correct_onnx_script_model
    functions = extract_functions(backend_test.name, code, self.test_folder)
onnxscript\backend\onnx_export_test.py:139: in extract_functions
    raise AssertionError(
E   AssertionError: Unable to import 'tests.onnx_backend_test_code.test_softmax_large_number' (e=No module named 'tests.onnx_backend_test_code.test_softmax_large_number') (file: 'D:\\a\\onnxscript\\onnxscript\\tests\\onnx_backend_test_code\\test_softmax_large_number.py', absolute path: 'D:\\a\\onnxscript\\onnxscript\\tests\\onnx_backend_test_code\\test_softmax_large_number.py', current folder: D:\a\onnxscript\onnxscript
E   ---- CONTENT --
E   import numpy
E   from onnx import TensorProto
E   from onnx.helper import make_tensor
E   from onnxscript import script, external_tensor
E   from onnxscript.values import Opset
E   from onnxscript.onnx_types import FLOAT
E   from onnxscript.onnx_opset import opset13
E   
E   @script()
E   def bck_test_softmax_large_number(x: FLOAT[2,4]) -> (FLOAT[2,4]):
E       y = opset13.Softmax(x)
E       return y

To view more test analytics, go to the Test Analytics Dashboard
_{📋 Got 3 mins? Take this short survey to help us improve Test Analytics.}

[gramalingam]

This looks fine, thanks.

What would be interesting (but is not clear yet to me) is how we use passes effectively and whether that requires other abstractions. For example, I see that the PassManager allows execution of a sequence of passes (iteratively). Would we need the ability to turn a sequence of passes (may be even an iteration N times) into a pass (which can itself then be used as one pass among another sequence of passes)? Essentially a "Sequence of Passes" as a "Pass"?

Of course, that leads to other ways of composing passes (eg. "pass1 OR pass2": in ORT fusion context, there is the possibility that we try GQA fusion only if MHA fusion fails). Of course, this may end up encoding a full programming-language within passes ... but that may be a bit too much right now, and it seems easier to use Python to express compositions for right now.

[justinchuby]

Essentially a "Sequence of Passes" as a "Pass"?

Sure! PassManager itself can be used as a pass. Something that is specialized for that will work too.