Refactor/optimisation : Reduction of LLVM Analysis Algorithmic Cost through Shared Traversals and Intermediate Artifacts

[SizzleUnrlsd]

Description

This plan aims to reduce the algorithmic cost of LLVM analyses by sharing traversals and intermediate artifacts, without changing the semantics of diagnostics. The approach prioritizes a pipeline engine with requires/provides dependencies, then a shared fact collector to reduce redundant traversals.

Decisions

• Primary objective: reduction of real time (wall-clock).
• Delivery strategy: migration in 2 phases.
• Analysis selection: review module by module in phase 1 to classify subscriber-compatible vs independent (dataflow/fixpoint).

Scope

In

• Evolution of analysis architecture (orchestration of modules, sharing of data between analyses, cache of derived artifacts, performance/functional validation).

Out

• Change of diagnostic business rules
• Modification of test files
• Complete migration of all analyses in a single iteration

Action items

• Phase 1 - Establish the target real performance baseline (timing + xctrace) and define primary/secondary KPIs (real priority, user/sys in support).
• Phase 1 - Add technical instrumentation of traversals (Module/Function/Instruction counters per analysis) to measure redundancy level before refactoring.
• Phase 1 - Review each analysis to classify: subscriber-compatible vs independent (dataflow/fixpoint).
• Phase 1 - Refactor AnalysisPipeline towards a requires/provides model to explicitly declare dependencies between analyses.
• Phase 1 - Introduce a minimal skeleton AnalysisArtifactStore type-safe (API + invalidation by module/config), without freezing all artifacts.
• Phase 1 - Add an IRFactCollector as the first pass of the pipeline (unique traversal Module/Function/BB/Instruction) exposing shared IRFacts.
• Phase 2 - Introduce a subscriber mechanism (Observer) for analyses classified as subscriber-compatible.
• Phase 2 - First migrate pilot analyses to shared facts (priority: StackBuffer, ResourceLifetime, Uninitialized; additional candidates: IntegerOverflow, GlobalReadBeforeWrite) and measure incremental gain on real time.
• Phase 2 - Extend PreparedModule/cache with stable derivable artifacts (debug index, symbol maps, type facts) with versioned key.
• Phase 2 - Add architectural safeguards: pipeline integration tests, cache invalidation checks, and non-regression diagnostic assertions.
• Phase 2 - Validate before/after and progressive rollout behind feature flag, driven by real time reduction.