Track adding control flow passes for optimization level 1

[ewinston]

What should we add?

This issue is to track the target set of passes to enable control flow handling for transpiler optimization_level=1.

• SetLayout (no change)
• TrivialLayout (no change)
• DenseLayout (add control flow handling to DenseLayout pass #8742)
• EnlargeWithAncilla (no change)
• FullAncillaAllocation (no change)
• ApplyLayout (no change)
• VF2Layout (Make VF2Layout and VF2PostLayout work with control flow #8804)
• VF2PostLayout (Make VF2Layout and VF2PostLayout work with control flow #8804)
• CheckMap (add controlflow handling to CheckMap transpiler pass. #8793 )
• CheckGateDirection (Update GateDirection and checker for control flow #8822)
• GateDirection (Update GateDirection and checker for control flow #8822)
• StochasticSwap (Add controlflow handling to stochastic swap pass #8418)
• ContainsInstruction (Add control-flow support to "properties" transpiler passes #8763)
• GatesInBasis (Add control-flow support to GatesInBasis #8823)
• UnitarySynthesis (add control flow support to UnitarySynthesis pass #8565 )
• Unroll3qOrMore (Lift trivial transpiler passes to handle control flow #8752 )
• UnrollCustomDefinitions (Lift trivial transpiler passes to handle control flow #8752)
• BasisTranslator(support control flow in BasisTranslator pass #7808)
• BarrierBeforeFinalMeasurements (no change)
• Depth (Add control-flow support to "properties" transpiler passes #8763)
• FixedPoint (no changes)
• Size (Add control-flow support to "properties" transpiler passes #8763)
• Optimize1qGatesDecomposition (add control flow support to Optimize1qDecomposition pass #8766)
• CXCancellation (Lift trivial transpiler passes to handle control flow #8752)
• RemoveResetInZeroState (no change required)

We also need to update the preset pass managers (esp level 1) to suitably modify the defaults / reject scheduling_method != None if a given circuit has control flow:

• Update preset pass managers (Modify preset pass managers for control-flow support #8830)

Time permitting:

• SabreSwap
• SabreLayout
• Layout2qDistance - we elected to swap the O1 PM to use CheckMap instead, to avoid adding the questionable "weighting" of control-flow blocks to another pass (add controlflow handling to Layout2qDistance transpiler pass #8733)

[mtreinish]

Should we add SabreLayout and SabreSwap to this list because of: #8552 ?

[jakelishman]

We can put it in the list. I don't think there's any chance of us supporting Sabre for 0.22 when there's control flow. Part of this is working out what (if any) chances to the preset pass-manager logic we need for the initial support, when there's control-flow in the circuit. The general idea being that we can have tweaked preset pass-manager logic if there's control flow or not, while we're building up to full support everywhere, but we want the literal call transpile(qc, optimization_level=1) to work for 0.22.

[ewinston]

For the timeline of 0.22 it might be better to prioritize a set of non-redundant passes. Since BasicSwap and StochasticSwap have already been started it may be faster to get some other items on the list complete before coming back to implement alternative (even if soon to be default) routing passes.

[kdk]

@ewinston Can you call out in the list above which passes are easily implementable as recursion over the blocks of each ControlFlowOp, and those which need some custom handling?

[ewinston]

@kdk Just saw this comment. Here's which passes might require special handling of blocks,

*** SetLayout
no
*** UnitarySynthesis
no
*** Unroll3qOrMore
no
*** TrivialLayout
no
*** Layout2qDistance
weight 2q distance for loop?
*** VF2Layout
possibly
*** DenseLayout
yes
*** FullAncillaAllocation
perhaps since controlflow ops may appear full width
*** EnlargeWithAncilla
no
*** ApplyLayout
no recursion
*** CheckMap
yes
*** BarrierBeforeFinalMeasurements
yes, simple
*** StochasticSwap
yes
*** UnrollCustomDefinitions
no
*** BasisTranslator
no
*** RemoveResetInZeroState
yes (reset occurs at beginning of block)
*** Depth
no change in pass but in DAGCircuit where this is calculated
*** FixedPoint
no
*** Size
no change in pass but in DAGCircuit where this is calculated
*** Optimize1qGatesDecomposition
not required
*** CXCancellation
not required
*** ContainsInstruction
no

[jakelishman]

I'll handle all the trivial passes in a PR that I'll post tomorrow.

Just a note: BarrierBeforeFinalMeasure doesn't need to recurse in, because any measurements in a control-flow block can't be final measurements by the original definition. ResetInZeroState can't trivially recurse, because the start of a block cannot be assumed to be zero.

[jakelishman]

I found some more transpiler passes while trying to update the preset pass managers, so I've added them to the list.