Transforms: make layout-copy rewrites backend-configurable

[rascani]

Summary

The layout propagation passes currently assume every layout copy is aten.permute_copy. That prevents a backend from using the shared propagation algorithms when its layout copies must remain visible as a backend-specific operator.

This PR separates the shared transformation mechanics from that backend policy:

• permute_targets defines a closed set of operators representing layout copies.
• Synthesized copies preserve the source operator, and metadata refresh recognizes the configured targets.
• Mixed-dialect chains and regions are left unchanged.
• can_propagate lets a backend stop movement at barriers without adding backend-specific operators to the shared pass.
• Lower-rank constants are reshaped only when the permutation moves unit dimensions. Reordering non-unit dimensions still emits a real copy in the source dialect.

Existing callers retain the current defaults. In particular, this PR does not add multi-input upward distribution, boundary compensation, or output-layout policy, and it does not change the handling of aten.view.default. Those decisions are not required to support backend-defined layout-copy operators.

Review guide

1. canonicalize_view_copy_permute_pass.py and permute_view_meta.py
   • Recognizing configured copy targets and refreshing their metadata.
2. propagate_view_copy_permute_pass.py
   • Keeping the target family closed and preserving the source dialect.
3. remove_permutes_around_elementwise_ops.py
   • Backend barriers, constant handling, and conservative mixed-dialect behavior.
4. Arm tests
   • Verifying that existing Arm policy remains unchanged.

AI-assisted: Codex.

---

_{Stack created with GitHub Stacks CLI • Give Feedback 💬}

cc @digantdesai @freddan80 @per @zingo @oscarandersson8218 @mansnils @Sebastian-Larsson @robell

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/22375

• 📄 Preview Python docs built from this PR

Note: Links to docs will display an error until the docs builds have been completed.

❌ 1 New Failure, 15 Cancelled Jobs, 13 Pending, 1 Unrelated Failure

As of commit 57e4816 with merge base 684d4bd ():

NEW FAILURE - The following job has failed:

• trunk / test-cortex-m-e2e / run (mv2, cortex-m7) / mv2-cortex-m7 (gh)
  ##[error]An error occurred trying to start process '/usr/bin/bash' with working directory '/home/ec2-user/actions-runner/_work/executorch/executorch/pytorch/executorch'. No such file or directory

CANCELLED JOBS - The following jobs were cancelled. Please retry:

• Build Presets / apple (ios) / build (gh)
• Build Presets / linux (llm, linux.2xlarge, executorch-ubuntu-22.04-clang12) / build (gh)
• Build Presets / linux (llm, linux.arm64.2xlarge, executorch-ubuntu-22.04-gcc11-aarch64) / build (gh)
• Build Presets / linux (pybind, linux.arm64.2xlarge, executorch-ubuntu-22.04-gcc11-aarch64) / build (gh)
• Build Presets / windows (pybind) / build (gh)
• Lint / Get changed files / get-changed-files (gh)
• Lint / link-check / lint-file-size (gh)
• Lint / link-check / lint-urls (gh)
• Lint / link-check / lint-xrefs (gh)
• Lint / lintrunner (gh)
• Lint / lintrunner-mypy (gh)
• Test CoreML Backend / Get changed files / get-changed-files (gh)
• Test QNN Backend / test-qnn / package-golden-artifacts (gh)
• Test QNN Backend / test-qnn / test-backend-linux (qnn, models) / linux-job (gh)
• Test WebGPU Backend / test-webgpu / package-golden-artifacts (gh)

FLAKY - The following job failed but was likely due to flakiness present on trunk:

• pull / test-models-linux (linear, portable, linux.2xlarge) / linux-job (gh) (detected as infra flaky with no log or failing log classifier)

This comment was automatically generated by Dr. CI and updates every 15 minutes.

[AdrianLundell]

It seems like the patch is doing some things that are not that well advertised, why is test_mixed_reconvergence_fork_does_not_strand_a_permute removed and why is the new _maybe_distribute_upwards_permute_over_elementwise required just to support backend defined layout copies?

[rascani]

It seems like the patch is doing some things that are not that well advertised, why is test_mixed_reconvergence_fork_does_not_strand_a_permute removed and why is the new _maybe_distribute_upwards_permute_over_elementwise required just to support backend defined layout copies?

Agreed, sorry again about that. I have updated the summary to go in-depth and provided a review order suggestion. Upward distribution is not required to support backend-defined layout copies, and the unrestricted version can increase the copy count. I removed it from this PR and restored the mixed-reconvergence xfail. I also removed boundary compensation and the output-layout policy for the same reason.