Add more patterns to loop unroll pass

[annagrin]

Description

Add more patterns to loop unroll pass

To be able to optimize away loops with boundaries that are computed from array storage, other loops, conditions, and more, we need to run a bunch of optimizations together in the loop unroll pass.

Added following optimizations to cc-loop-unroll:

• LoopOpt pattern from cc-loop-normalize (to keep normalizing loops in loop unrolling iterations)
• RewriteIf from lower-to-cfg (to const-prop if statements)
• SimplifyWritesAnalysis from write-after-write-elimination (remove multiple writes to the same location on the stack)
• simplifyRegions (to remove dead code)
• AllocaPattern from lift-array-alloc (to lift constant arrays, helps const prop)

Added tests.

[github-actions]

CUDA Quantum Docs Bot: A preview of the documentation can be found here.

[github-actions]

CUDA Quantum Docs Bot: A preview of the documentation can be found here.

[github-actions]

CUDA Quantum Docs Bot: A preview of the documentation can be found here.

[github-actions]

CUDA Quantum Docs Bot: A preview of the documentation can be found here.

[schweitzpgi]

Removing loop normalization doesn't look correct.

[annagrin]

I added loop normalization pattern into the LoopUnroll pass itself so it is no longer needed here. I've run into a scenario where loop normalization was needed during the loop unroll fix point for it to fully succeed.

[annagrin]

I am thinking - given that I also needed to add something similar to the lift-array-alloc pass (in #2601) - instead of modifying the cc-loop-unroll and lift-array-alloc passes, would it be better to create a new classical-optimization-pipeline that would run the following:

classical-optimization-pipeline:

• memtoreg=quantum=0
• classical-optimization
• cse
• classical-optimization
  (Note: after we update the llvm version we could run cse inside the fixpoint of classical-optimization, so we would not need to run it twice here and would be able to compile more code)

Where classical-optimizations pass runs the following patterns as a part of single fix point (same as the cc-loop-unroll pass in current PR, but run them on code that does not contain loops as well):

• write-after-write-elimination
• lift-array-alloc
• canonicalize
• loop-normalize patterns
• simplifyRegions
• original loop unrolling patterns

We would use classical-optimization-pipeline instead of the initial set of opts in quantum config files:
func.func(const-prop-complex,canonicalize,cse,lift-array-alloc)

And later again after the following passes:
apply-op-specialization,aggressive-early-inlining,expand-measurements

[annagrin]

So, as an example, the new pipeline for Ionq would look like:

"func.func(classical-optimization-pipeline), globalize-array-values,func.func(state-prep),unitary-synthesis,canonicalize, apply-op-specialization,aggressive-early-inlining,expand-measurements, func.func(classical-optimization-pipeline), decomposition{enable-patterns=U3ToRotations}, func.func(lower-to-cfg,canonicalize,multicontrol-decomposition), ionq-gate-set-mapping"

[annagrin]

Replaced by #2605