✨ Lower quantum-register functions to QIR Adaptive

[burgholzer]

🤖 AI text below 🤖

Problem Statement

QIR 2.1 Adaptive supports arrays and retained IR functions, but MQT Core assumes that helper calls have been inlined before QIR lowering. A retained qubit-register helper currently mixes the custom raw-pointer memref conversion with Func-to-LLVM descriptor promotion and can assert in MemRefDescriptor::unpack instead of producing QIR or a diagnostic.

The existing lowering also moves register releases to the entry point's output block, so preserving arbitrary allocating helpers would give invalid ownership and SSA behavior.

Proposed Solution

Lower a safe fixed-size borrowed-register subset to QIR Adaptive.

ABI and profile contract

• Represent a fixed memref<Nx!qc.qubit> formal as a borrowed opaque pointer to caller-owned contiguous [N x ptr] storage.
• Use MLIR's bare-pointer function calling convention consistently for function signatures and calls. Do not mix raw-pointer lowering with memref descriptor promotion.
• The caller owns allocation and release. A callee may load and operate on register elements but must not allocate, release, or return the borrowed storage.
• Preserve eligible helpers when existing no_inline policy requests it. Unmarked helpers may still inline.
• Emit arrays=true and ir_functions=true when a helper remains.
• Keep Base Profile behavior unchanged: inline helpers, or reject a retained no_inline helper with a diagnostic.

Initially reject dynamic register formals. They require a separate (ptr, i64 length) ABI and module-wide lifetime rules.

Pipeline and lowering

• Make QIR preparation profile-aware in the shared pass builder and every library/CLI entry point.
• Ensure fixed register allocations use a native LLVM array representation and type-correct GEP access.
• Extend metadata analysis across every retained function. ir_functions already scans the module; dynamic-resource, array, type, and backward-branch capabilities must not inspect only the entry point.
• Reject helper-local qubit/register/result allocation, measurement-result ownership, and deallocation until those lifetimes have a complete per-function lowering.
• Replace the current assertion path with success for the supported subset and diagnostics for everything else.

Likely implementation areas:

• mlir/lib/Support/Passes.cpp and its declaration/callers
• mlir/lib/Compiler/Pipeline.cpp
• mlir/tools/mqt-cc/mqt-cc.cpp
• mlir/lib/Conversion/QCToQIR/QIRCommon/QIRCommon.cpp
• mlir/lib/Conversion/QCToQIR/QIRAdaptive/QCToQIRAdaptive.cpp
• mlir/lib/Dialect/QIR/Transforms/AttachQIRAttributes.cpp
• Focused Adaptive/Base conversion and compiler-pipeline tests

Acceptance criteria

• Adaptive output retains a no_inline fixed-register helper and a matching LLVM call.
• The callee loads qubits from caller storage and never releases the register.
• The LLVM/QIR output verifies and records arrays=true and ir_functions=true.
• Capability metadata accounts for operations in retained helper bodies.
• Base output contains no retained helper call; an impossible retained helper fails cleanly.
• Dynamic register formals and unsafe helper-local resource operations receive diagnostics, never assertions or crashes.
• Direct pass use, the C++ library pipeline, and mqt-cc produce the same profile behavior.

Dependencies and related work

• Parent feature: ✨ Support qubit-register arguments in reusable functions #2427
• Depends on: ✨ Support quantum-register function arguments in QC and QCO #2428
• Related broader backlog: ✨ Improve QIR support #1734
• QIR array contract: https://github.com/qir-alliance/qir-spec/blob/main/specification/Memory_Management.md

Non-goals

Dynamic register formals; a new generic QIR ABI abstraction; Base Profile IR functions; helper-local quantum-resource ownership; register-return ABI; changing the default inlining policy for unrelated helpers.