Summary

Several passes in exir/passes/ return PassResult(graph_module, True) unconditionally — even when they walk the graph and rewrite nothing.

# exir/program/_program.py:271
if transformed_gm is self.graph_module and not res.modified:
    return self
return _update_exported_program_graph_module(self, transformed_gm, override_verifiers)

A false-positive modified=True forces the consumer to rebuild the ExportedProgram (deep-copy module_call_graph, recompute the graph signature, re-run verifiers) on every pass run, even when nothing changed.

PR #16231 fixed this for ScalarToTensorPass. This PR applies the same pattern to the remaining passes.

Passes touched

Each fix follows the pattern used by CSEPass, ScalarToTensorPass, and others: a local modified boolean is set at the point of actual graph mutation and returned via PassResult(gm, modified). Cleanup operations (eliminate_dead_code, lint, recompile) are only called when modified=True.

As a minor cleanup, RemoveGraphAssertsPass and RemoveNonCoreAtenOpGraphAssertsPass are refactored to share a private helper _erase_asserts_from_modules — the two classes were near-identical, differing only in the target op set.

flowchart TD
    A["EdgeProgramManager.transform(MyPass)"] --> B[_transform_with_pass_manager]
    B --> C{res.modified}
    C -- True --> D[rebuild ExportedProgram]
    C -- False --> E[return self\nintended fast path]

Test results

pytest exir/tests/test_passes.py::TestPassModifiedFlag -q
# 0 failed

lintrunner exir/passes/remove_noop_pass.py \
           exir/passes/prune_empty_tensors_pass.py \
           exir/passes/remove_graph_asserts_pass.py \
           exir/passes/replace_sym_size_op_pass.py \
           exir/passes/to_device_pass.py \
           exir/tests/test_passes.py
# No lint issues.