[Tracking issue][Performance]: Avoid pathological LLVM inputs #3198

Performance Issue

This issue is part of the performance workstream (Quantinuum/guppylang#2182).

Existing LLVM optimization passes were designed with some preconceptions on the shape of the IR they ingest, tuning them specifically for C-shaped code.
The rustc project has had to battle with these preconceptions before.

We have been bitten by this before when lowering guppy-generated Hugr code, and producing degenerate inputs to certain LLVM passes .

We need to profile LLVM runs on hugr-llvm output for real-world guppy programs, and see what can be improved.

Tasks:

• Profile LLVM optimization of real world guppy-generated LLVMIR, and add issues as needed.
• Local variable threading vs non-local edges: [Performance]: Local variable threading causes LLVM slow-downs #3169
  • May require [Feature]: type-scoped LocalizeEdges for linearised non-local edges tket2#1883
• Compile time escalates with the number of reachable function: [Performance]: Compile time scales with number of reachable functions, dominated by native backend (LLVM + tket passes) guppylang#2084

Older issues that may be relevant:

• Return array from array result operation: perf(hseries): return array from array result operations tket2#922

Related issues / PRs

• perf: Select required CFG predecessor outputs for successors through basic blocks instead of conditionals guppylang#2311