:sparkles: Add Quantum IPO Pass to `mqt-cc` by DRovara · Pull Request #1970 · munich-quantum-toolkit/core

DRovara · 2026-07-30T10:30:36Z

Description

This PR introduces the IPO passes from the corresponding paper to mqt-cc.
It includes:

Quantum Argument Promotion
Auxiliary Qubit Hoisting
Quantum Function Boundary Commutation
QuantumContextSensitiveSpecialization

AI Notice: Gemini 3.1 Pro and Claude Opus 5.0 were used for the generation of some code.

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.
I have added entries to the changelog for any noteworthy additions, changes, fixes, or removals.
I have added migration instructions to the upgrade guide (if needed).
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.

…toolkit/core into mlir/measurement-lifting

… first tests

…ontrols

Signed-off-by: Damian Rovara <93778306+DRovara@users.noreply.github.com>

`populateQuantumIPOPipeline` runs specialization, argument promotion, auxiliary qubit hoisting and boundary commutation in the order they are meant to compose, and registers them as the `quantum-ipo` pipeline. Commutation runs twice because it can expose further cancellations. The passes stay individually registered, so a caller assembling its own pipeline can pick only the ones it wants. `test_qco_quantum_ipo.cpp` keeps only the scenarios that need more than one pass. Everything about a single pass lives in that pass's own suite, scheduled on the pass alone. Extracted from #1970. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

…#2193) * ✨ Walk tensor chains across calls and compare untracked qubit tensors `TensorIterator` ends a tensor chain at `func.call` and `func.return` rather than aborting on them, so a chain reaching a call boundary can be walked. The tensor a call returns starts a life-chain of its own. `areModulesEquivalentWithPermutations` no longer assumes every qubit tensor belongs to an equivalence group. Tensors that do not descend from a `qtensor` allocation, such as function arguments, are compared through the regular SSA mapping, and the tensor threaded out of a `qtensor.extract` is now mapped alongside its qubit result. Both previously aborted inside the comparison instead of reporting a result. Extracted from #1970 as the first of a stack of smaller contributions. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> * 📝 Note the interprocedural optimization work in the changelog Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> * ✅ Cover walking a tensor chain backward from a call result The only branch this slice added that no existing test reached: a tensor a call hands back starts its own life-chain, so stepping back from it stops at the call rather than continuing into the tensor that was passed in. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 5 <noreply@anthropic.com>

`quantum-argument-promotion` replaces a qubit-tensor argument by one qubit argument and one qubit result per element the callee actually touches, so untouched elements no longer cross the call boundary. Call sites are rewritten to extract before the call and re-insert afterwards. The rewrite lifts every extraction in front of the call and sinks every insertion behind it, which is only sound when the accesses do not depend on each other's order. Two shapes are therefore rejected: a slot the callee writes before reading it again, which would otherwise be served from the caller's original tensor and silently compute the wrong thing; and an insertion that belongs to no promoted slot, which used to survive the rewrite still using the tensor argument erased right after and abort on MLIR's `use_empty()` assertion. Extracted from #1970. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

`populateQuantumIPOPipeline` runs specialization, argument promotion, auxiliary qubit hoisting and boundary commutation in the order they are meant to compose, and registers them as the `quantum-ipo` pipeline. Commutation runs twice because it can expose further cancellations. The passes stay individually registered, so a caller assembling its own pipeline can pick only the ones it wants. `test_qco_quantum_ipo.cpp` keeps only the scenarios that need more than one pass. Everything about a single pass lives in that pass's own suite, scheduled on the pass alone. Extracted from #1970. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

`WireIterator` traverses `func.call` in both directions instead of stopping there, and `atWireStart()` reports whether a wire begins at the current operation, which a call result does only when the callee creates the qubit. Which result continues which operand is derived by threading the callee's arguments through its body rather than assumed to be positional, so a callee that hands its qubits back in a different order is followed correctly. The derivation is cached per callee by `CallQubitMapping`, with `CallTensorMapping` providing the same for qubit tensors. External, recursive and multi-block callees fall back to the positional pairing. Also fixes two defects this surfaced: threading asked for a body's terminator before checking that one exists, which trips on a function still being built, and `qtensor.from_elements` was not treated as a wire sink even though it consumes qubits into a tensor exactly like `qtensor.insert`. Extracted from #1970. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

`startFunction`/`endFunction` open and close a function next to `main`, and `call` emits a `func.call` to one. Test programs spanning several functions can now be built with the builder instead of hand-written MLIR. Linear values are isolated per function: `startFunction` sets aside the caller's tracked qubits and tensors and restores them in `endFunction`, so an outer value used inside a callee is rejected rather than silently threaded, and anything the callee leaves behind is reported as a leak. At a call, each linear operand is paired with the result that continues it, derived from the callee through `CallQubitMapping` and `CallTensorMapping` rather than by position. An operand the callee keeps counts as consumed, and a result no operand flows into as freshly created. Extracted from #1970. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

`quantum-context-sensitive-specialization` redirects a `func.call` to a specialized copy of its callee whenever the call site pins something down: a qubit argument known to be in |0>, letting operations that fix |0> be dropped; one known to be in |+>, letting an `x` be dropped; or a compile-time constant rotation angle from a small distinguished set, which is folded into the copy. Copies are shared between call sites with the same context, and a callee left without callers is erased — but only when this pass created that situation, so unrelated unused functions are left alone. Specializations are private, so optimizing a public callee does not export the generated symbol. `IPOUtils.{h,cpp}` holds the two helpers the interprocedural passes share. The tests are scheduled on this pass alone rather than on a pipeline, so a regression here cannot be masked by a later pass. `IPOTestFixture.h` carries the fixture the sibling passes will reuse. Extracted from #1970. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

`quantum-argument-promotion` replaces a qubit-tensor argument by one qubit argument and one qubit result per element the callee actually touches, so untouched elements no longer cross the call boundary. Call sites are rewritten to extract before the call and re-insert afterwards. The rewrite lifts every extraction in front of the call and sinks every insertion behind it, which is only sound when the accesses do not depend on each other's order. Two shapes are therefore rejected: a slot the callee writes before reading it again, which would otherwise be served from the caller's original tensor and silently compute the wrong thing; and an insertion that belongs to no promoted slot, which used to survive the rewrite still using the tensor argument erased right after and abort on MLIR's `use_empty()` assertion. Extracted from #1970. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

`quantum-auxiliary-qubit-hoisting` turns a qubit a callee allocates and releases itself into an extra argument and an extra result, handed back reset, so the caller owns the allocation and can reuse one qubit across several calls. Externally visible functions, declarations and recursive functions are left alone. Candidates are visited callees first, in a post-order over the call graph. Hoisting out of a callee puts an allocation into each of its callers, which may itself be hoistable; processing in module order used to strand that allocation wherever the declarations happened to sit, so the same call graph gave different results depending on declaration order. Extracted from #1970. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

`quantum-function-boundary-commutation` drops a self-inverse gate applied right before a call when the callee starts by applying the same gate to that argument. The caller-side gate is erased and the call is redirected to a copy of the callee without the callee-side one. Copies are cached per callee *and* parameter index. The gate removed inside the copy belongs to one specific argument, so sharing a copy between parameters would drop a gate the second call site never had. Extracted from #1970. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

`populateQuantumIPOPipeline` runs specialization, argument promotion, auxiliary qubit hoisting and boundary commutation in the order they are meant to compose, and registers them as the `quantum-ipo` pipeline. Commutation runs twice because it can expose further cancellations. The passes stay individually registered, so a caller assembling its own pipeline can pick only the ones it wants. `test_qco_quantum_ipo.cpp` keeps only the scenarios that need more than one pass. Everything about a single pass lives in that pass's own suite, scheduled on the pass alone. Extracted from #1970. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

burgholzer · 2026-09-01T11:11:04Z

I will close this one for now as well, as individual PRs have been set up.

DRovara and others added 30 commits July 14, 2026 03:11

feat(mlir): ✨ add pass and patterns for measurement lifting

89f8ee3

test(mlir/hybrid-opt): 🚧 set up tests for measurement lifting

6d4332d

fix(mlir): ✅ fix tests

1a7664c

style(mlir): 🚨 fix linter issues

770e55c

style(mlir): ♻️ implement coderabbit suggestions

64b162c

fix(mlir): 🐛 fix all measurement lifting tests

137d93c

style(mlir): 🚨 fix linter issues

559c9ff

style(mlir): ♻️ minor code clean-up

679f454

style(mlir): 🚨 fix linter issue regarding unused include

5510fe8

🎨 pre-commit fixes

64cffab

chore(mlir): ✏️ clean up minor typo

df2301b

fix(mlir): ✏️ fix some further typos

8ca81d6

Merge branch 'mlir/measurement-lifting' of github.com:munich-quantum-…

c55987b

…toolkit/core into mlir/measurement-lifting

test(mlir): ✅ add further special cases to tests

653a8f4

chore(mlir): ♻️ add docstrings to new tests

1611c1b

test(mlir): ✅ add test for inverse gates

4a158fb

chore(mlir): ♻️ slight refactoring in tests

4d98606

chore(mlir): ♻️ further tests clean up

2deb2a4

fix(mlir): ♻️ address coderabbit issues

096075c

chore(mlir): ♻️ some further coderabbit suggestions

b1a24cd

feat(mlir): ✨ implement classical control replacement pattern and add…

b0d060d

… first tests

style(mlir): 🚨 fix includes

7042e3b

style(mlir): ♻️ apply coderabbit suggestions

58d10b9

fix(mlir): 🐛 fix issues due to new version

febb216

docs(mlir): 📝 add to changelog

6d4ce73

Merge branch 'mlir/measurement-lifting' into mlir/replace-classical-c…

38bbc15

…ontrols

docs(mlir): 📝 add to changelog

5e482a9

Merge branch 'main' into mlir/measurement-lifting

70de05f

Merge branch 'main' into mlir/measurement-lifting

102663b

Signed-off-by: Damian Rovara <93778306+DRovara@users.noreply.github.com>

chore(mlir): ♻️ refactor isDiagonal check to use getSoleBodyUnitary

91fe289

burgholzer closed this Sep 1, 2026

mergify Bot removed the conflict label Sep 1, 2026

Navigation Menu

Uh oh!

Uh oh!

✨ Add Quantum IPO Pass to `mqt-cc` - #1970

✨ Add Quantum IPO Pass to `mqt-cc`#1970
DRovara wants to merge 142 commits into
mainfrom
mlir/ipo-pass

DRovara commented Jul 30, 2026 •

edited

Loading

Uh oh!

burgholzer commented Sep 1, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Uh oh!

Uh oh!

Conversation

DRovara commented Jul 30, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Description

Checklist

Uh oh!

burgholzer commented Sep 1, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

DRovara commented Jul 30, 2026 •

edited

Loading