[C++] Add more tests for control vs. broadcast

[arnavnagzirkar]

Summary

Root Cause

There were no dedicated tests documenting how cudaq-quake handles the full
range of broadcast vs. control semantics for quantum gates. The existing
ctrl_vector.cpp test covered the h(veq, qubit) / x(veq, qubit) implicit-
control pattern but did not cover:

• Broadcasting x over individually-indexed qubits (x(q[0], q[1], q[2]))
• Broadcasting x over an entire register (x(q))
• The surprising case where x<cudaq::ctrl>(q) with a single register still
  broadcasts (the compiler ignores the ctrl modifier for a single VeqType
  operand — see buildOp in ConvertExpr.cpp lines 283-299)
• Multiple individual-qubit controls (x<cudaq::ctrl>(q[0], q[1], q[2]))
• Simple swap, controlled swap with individual qubits, and controlled swap
  with a register of controls

Change Made

New file: cudaq/test/AST-Quake/ctrl_broadcast.cpp

Eight FileCheck tests covering the full broadcast/control matrix:

Kernel expression	Behavior	MLIR output
x(q[0], q[1], q[2])	Broadcast	3 separate quake.x ops
x(q)	Broadcast loop	cc.loop + quake.x
x<cudaq::ctrl>(q)	Still broadcasts (ctrl ignored for single veq)	cc.loop + quake.x
x<cudaq::ctrl>(q[0],q[1],q[2])	Controlled	quake.x [ctrl0,ctrl1] tgt
x(veq, qubit)	Implicit ctrl (default overload modifier)	quake.x [veq] qubit
swap(q[0], q[1])	Simple swap	quake.swap ref0, ref1
swap<cudaq::ctrl>(q[0],q[1],q[2])	Controlled swap	quake.swap [ctrl] tgt0, tgt1
swap(ctrl_reg, src, tgt)	Controlled swap via dedicated overload	quake.swap [veq] src, tgt

Key behavioral insight documented: the two-arg swap(QuantumRegister, qubit, qubit)
overload has no mod template parameter, so it cannot take <cudaq::ctrl> —
passing a register as the first argument implicitly makes it the control.

Issue

Fixes #913

Issue URL: #913

Changes

cudaq/test/AST-Quake/ctrl_broadcast.cpp | 142 ++++++++++++++++++++++++++++++++
 1 file changed, 142 insertions(+)

Testing

• Agent ran relevant tests during development

• Linting checks passed

• Changes are minimal and focused on the issue

AI Assistance Disclosure

This pull request was prepared with the assistance of AI coding tools (GitHub Copilot). The change has been read, understood, and is owned by the human contributor submitting it, who will respond to review feedback.

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[schweitzpgi]

These CHECK lines need to be expanded. They are too heavily pruned.

[schweitzpgi]

Ditto

[schweitzpgi]

And so on ... if we're going through the trouble to write tests for these cases, we should make sure that the tests themselves are robust enough to catch regressions. Dropping 90% of the operations and regex matching any possible string can cause false positives, rendering the test useless.

[arnavnagzirkar]

Thanks for the review. I have expanded the three pruned check blocks so they assert the full structure instead of matching arbitrary strings.

For the two broadcast loops (broadcast_register and ctrl_broadcast_register) the checks now bind the alloca and walk the whole invariant loop: quake.veq_size on the register, the cc.loop while header with its induction variable, the arith.cmpi slt bound test, cc.condition, the do block, quake.extract_ref with the concrete (!quake.veq<4>, i64) types, the quake.x on the extracted !quake.ref, cc.continue, the step block with arith.addi, and the {invariant} attribute. The ctrl_broadcast_register case keeps the CHECK-NOT to prove no controlled form is ever emitted.

For ctrl_individual the controls and target are now bound to the specific extracted refs (q[0] and q[1] as controls, q[2] as target) rather than wildcard operands, so a regression that swapped operands or changed the control set would fail.

These mirror the structure already used by the other cases in this file and by grover.cpp.