[Optimization] Make CSE per offload rather than whole kernel#790

Draft

hughperkins wants to merge 21 commits into

mainfrom

cse/fix-licm-cache

Commits on Jun 30, 2026

cse: per-offloaded-task CSE in parallel codegen workers + disable LICM cache
hughperkins
committed

Commits on Jul 15, 2026

Merge remote-tracking branch 'origin/main' into cse/per-task-parallel
hughperkins
committed

Commits on Jul 16, 2026

cse/fix-licm-cache: address-key the loop-invariant global cache (fix stale break-flag)
hughperkins
committed
cse/fix-licm-cache: two-phase soundness - exclude snodes with uncacheable (conditional) accesses
hughperkins
committed
cse/fix-licm-cache: add cheap pre-offload pointer-only merge (merge_global_ptrs)
hughperkins
committed
merge_global_ptrs: also merge integer addressing arithmetic + PTRMERGE log
hughperkins
committed
merge_global_ptrs: move to AFTER offload (before flag_access #2)
hughperkins
committed
merge_global_ptrs: run inside full_simplify fixpoint (pre-offload dedup), drop standalone
hughperkins
committed
cache_loop_invariant: QD_LICM_NO_EXCLUDE to disable two-phase exclusion
hughperkins
committed
cache_loop_invariant: exclusion OFF by default (merge_global_ptrs is the root-cause fix)
hughperkins
committed
merge_global_ptrs: QD_NO_PTR_MERGE=1 disables it (compile-cost A/B on one build)
hughperkins
committed