[simon1hofmann]

🤖 AI text below 🤖

Description

Reduce repeated verification in pipelines created by runWithPassManager. With NDEBUG, disable per-pass verification and verify the complete output module after the pipeline succeeds. Assertions-enabled builds retain per-pass verification and the final module check.

The helper requires valid input IR. Parsers and program imports validate input; callers that mutate a program's module must preserve its validity. In release builds, invalid intermediate IR can reach a later pass before detection, so assertions-enabled builds remain useful for locating compiler defects.

The change applies to API pipelines using this helper, including target compilation, cleanup, and QIR conversion. The normal mqt-cc pass-manager path, QCProgram::intoQCO, caller-owned managers, and custom textual pipelines keep their existing verification settings. QCO linearity and target-conformance checks remain in place. No new options or dependencies.

Performance

Compare main at ddee68053bb523ccfe71e01703120df9ed96796b against this PR at ea8a87f77a810eb836775b5bf3d0e0fd2986cdba. Both include the conversion improvements from #2695 and the assertion-free release SDK configuration from #2678, isolating this PR's verification policy.

Times are medians. Reductions use the geometric mean of within-pair candidate/baseline ratios, so they need not equal the ratio of the displayed medians.

Workload / topology	Pairs	Main	PR	Elapsed-time reduction
bwt_n37 / linear †	4	242.317 s	212.906 s	10.5%
factor247_n15 / square †	6	85.236 s	64.568 s	23.3%
cat_n260 / heavy-hex	12	104.322 ms	96.943 ms	6.6%
wstate_n380 / square	12	109.602 ms	86.866 ms	20.6%
qugan_n395 / heavy-hex	12	5.984 s	5.889 s	No clear change
BWT → QIR Adaptive	6	15.187 s	12.721 s	16.0%
2M H gates, one QCO → QC pass	4	2.467 s	2.464 s	No clear change

† BWT target compilation and Factor247 ran with recorded background work on a shared host; their percentages are less precise. For the other workloads, whole pairs were repeated when compiler overlap or average external CPU load of at least half a core was observed. This screening does not prove exclusive machine use. Excluded attempts were retained. QuGAN and the single-pass control have paired 95% bootstrap intervals spanning no improvement.

Separate verifier profiling supports the end-to-end results: W-state's accumulated verifier time fell from 27.8 ms to 5.1 ms, and BWT QIR's from 3.77 s to 1.21 s. QuGAN removed only about 0.08 s of verifier work from roughly 6 s of compilation, explaining its neutral result. The one-pass control has no intermediate verification to remove. Profiling used separate binaries; accumulated verifier durations can include concurrent calls and are not necessarily additive wall-time shares.

Measurement setup

• DGX Spark ARM64; LLVM/MLIR 23.1.2 with assertions off; Clang 23, Release, ThinLTO, mold; affinity restricted to ten Cortex-X925 cores.
• Native C++ API harness, serial fresh processes, balanced AB/BA pairs. Seed 42, ten mapping trials, one iteration, lookahead 20.
• Original QASM inputs from Benchpress revision b695f30e83a32bac05b9b4d8e98d37ba9aae5236. Targets: a 37-site line, 4×4 and 20×20 grids, and heavy-hex graphs of distance 11 and 13. Basis: id, sx, x, rz, cz, measure, reset.
• Target timing includes QC input copy, QC → QCO, target compilation, and QCO → QC. Production pipeline verification is included. Input loading, target setup, additional harness verification, serialization, and hashing are excluded.
• QIR timing includes QC copy and intoQIR(Adaptive), including preparation, conversion, and cleanup.
• All primary and profiling variants produced matching serialized output hashes and operation counts; QC outputs also had matching gate counts. Output IR was verified after timing.

Validation

• All 272 compiler tests pass with the actual assertion-free SDK. The assertions-enabled compiler suite also passes.
• One focused test covers invalid final output and the assertion-dependent intermediate verification policy. Rebuilt mutations that omit final verification or restore intermediate verification under NDEBUG fail the test.
• Whole-file C++ lint, repository lint, and git diff --check pass. Hosted checks for the updated commit remain pending.

Codex assisted with implementation, review, tests, benchmarks, and this description.

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

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[burgholzer]

Moving on to this one now.

[burgholzer]

Let's get this in!