perf(codegen): bound TailCallElim's alloca walk on wide statepoint functions

changelog.d/8894-tre-alloca-walk-budget.md

@@ -0,0 +1,3 @@
+### Fixed
+
+- `PERRY_LL_TRE_MAX_ALLOCA_WALK` (default 2^26, #8883): before the optimization pipeline runs, a function whose `allocas × instructions` exceeds the budget is stamped `"disable-tail-calls"="true"`, the switch `TailCallElim` itself honours. LLVM's `markTails` walks the transitive SSA uses of every alloca and only loads, stores and `nocapture` arguments stop it; on a statepoint-rewritten body an alloca handed to any runtime call reaches the statepoint token, its `gc.relocate`s and, through their `gc-live` bundles, every later statepoint, so each walk covers the whole function and the pass is quadratic. Coop's Next.js route (jsonwebtoken's bundled entry: 400 allocas, 643k post-RS4GC instructions, 477k relocates, ~1.6M visited uses per alloca) held one LLVM worker for the better part of two CPU-hours in that walk on a unit whose remaining `-Os` passes take 16 s. The stamped function still runs every other pass at the requested level (#8421's contract); it gives up exactly tail-recursion-to-loop and sibling-call codegen, and it is never `optnone` (#8583). The trip is logged with the function's name, alloca and instruction counts; `<n>` raises or lowers the cap, `0`/`off` disables it; both caches key on it.