[simon1hofmann]

🤖 AI text below 🤖

Description

QC-to-QCO conversion used a general dialect-conversion driver to maintain mutable quantum state, while builders and wire traversal repeated work or relied on implicit ownership and positional assumptions. This combines the source-order converter with the QC/QCO audit and its follow-up simplifications.

• Lower source operations in order, retain reference identity until users are converted, and replace function arguments in bulk. Use one structured-capture record, derive modifier state from preserved block arguments, and replace unused type-conversion machinery with exact type helpers.
• Fix builder region ownership, tensor reinsertion into the original slot, directly returned allocations, mixed classical/quantum call results, and wire correspondence through type-changing loops.
• Retain WireIterator and restore WireRange in three read-only consumers. Use operand positions for unitary correspondence, making traversal across all wires of a wide operation linear rather than quadratic. Keep tail-first deletion explicit and preserve generic-call boundaries.
• Remove redundant validation and origin collection only where the producer already establishes the contract. Preserve public-boundary checks and graceful rejection of unsupported input, including misplaced modifier yields.
• Consolidate shared conversion fixtures while retaining all 132 forward and 133 reverse cases. The complete change removes 219 net production lines and 409 net test lines. The follow-up adds one concrete rejection regression and extends an existing wire-call case.

API and supported inputs: QC/QCO builders own unfinished modules and are move-only. Borrowed contexts must outlive builders and IR. QC allocations remain restricted to function entry blocks; quantum region values preserve positional identity. QTensor scalars return to their original register slot, with equal dynamic extraction/insertion indices remaining a caller precondition. No new dependency is required.

Performance: Five paired rounds on shared DGX Spark hardware show 124×/45× faster forward traversal through an 8,192-wire barrier/unitary call. The isolated 32-wire call is 15% slower; this is not a universal speedup. Restoring ranges has no material cost in the measured QFT-18 and Ising-26 pipelines. Further converter cleanup saves 1.69 MiB (2.8%) of peak RSS for 2,000 sibling loops; flat and register-load conversion times are essentially unchanged. All 90 wire samples agree on traversal counts, and all 105 conversion/pipeline samples agree on IR hashes and gate counts. See the follow-up audit and its linked original comparison for pins, methods, and limits.

Validation: Rebased onto main 3f1ccfd69, preserving original authorship. Independent design and final Ponytail reviews are addressed. The full native build passes; 4,015 tests pass, including 2,910 MLIR tests, with one expected ScQDMIJobSpecificationTest.QueryJobId skip. Repository lint, whole-file C++ lint for 20 selected files, MLIR reference generation, and output checks pass. Python/wheels, sanitizers, and native Windows execution were not rerun locally. The portable death-test fix remains included; hosted CI must validate this published head separately.

Codex assisted with implementation, consolidation, measurement, review, and this description. Human review and acceptance remain outstanding.

Checklist

• The pull request only contains commits that are focused and relevant to this change.
• I have added appropriate tests that cover the new/changed functionality.
• I have updated the documentation to reflect these changes.
• The changes follow the project's style guidelines and introduce no new warnings.
• The changes are fully tested and pass the CI checks.
• I have reviewed my own code changes.

If PR contains AI-assisted content:

• Any agent that created, edited, or submitted GitHub content was explicitly authorized for that scope, as required by our AI Usage Guidelines.
• Every agent-authored or agent-edited public text body begins with the visible disclosure 🤖 *AI text below* 🤖 (titles are exempt).
• I have disclosed AI assistance in the PR description.
• I confirm that I have personally reviewed and understood all AI-generated content, and accept full responsibility for it.

[codecov]

Codecov Report

❌ Patch coverage is 98.74687% with 5 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
mlir/lib/Conversion/QCToQCO/QCToQCO.cpp	98.4%	4 Missing ⚠️
mlir/include/mqt/Dialect/QCO/QCOUtils.h	85.7%	1 Missing ⚠️

📢 Thoughts on this report? Let us know!

[burgholzer]

Thanks @simon1hofmann 🙏 I had codex combine the changes here with my local audit and optimized the resulting implementation a bit further.
Would you mind giving this another check?
This looks pretty good from my side.

Edit: Codex just had one more idea. I'll quickly explore that and push if it turns out to be useful.
Edit(2): ac30ce4 is the result. Should be good now.

[simon1hofmann]

Thanks @simon1hofmann 🙏 I had codex combine the changes here with my local audit and optimized the resulting implementation a bit further. Would you mind giving this another check? This looks pretty good from my side.

Edit: Codex just had one more idea. I'll quickly explore that and push if it turns out to be useful. Edit(2): ac30ce4 is the result. Should be good now.

This one also looks good to me and should speed up the conversions quite substantially ⚡

[burgholzer]

Thanks @simon1hofmann 🙏 I had codex combine the changes here with my local audit and optimized the resulting implementation a bit further. Would you mind giving this another check? This looks pretty good from my side.
Edit: Codex just had one more idea. I'll quickly explore that and push if it turns out to be useful. Edit(2): ac30ce4 is the result. Should be good now.

This one also looks good to me and should speed up the conversions quite substantially ⚡

Nice! Let's get it in! My local results across a couple of benchmark instances yielded 1.5-10x improvements for the conversion itself. ⚡