Wasmtime GitHub notifications bot (Jul 23 2026 at 18:00): Wasmtime GitHub notifications bot (Jul 23 2026 at 18:00):fitzgen opened PR #13947 from fitzgen:dead-store-elimination-perf to bytecodealliance:main:
Follow up to #13806 with a few perf improvements; see each commit. This, combined with my other perf PRs that have landed recently, recovers the hit to compilation times from introducing a dead-store elimination pass.
Wasmtime GitHub notifications bot (Jul 23 2026 at 18:00):fitzgen requested cfallin for a review on PR #13947.
Wasmtime GitHub notifications bot (Jul 23 2026 at 18:00):fitzgen requested wasmtime-compiler-reviewers for a review on PR #13947.
Wasmtime GitHub notifications bot (Jul 23 2026 at 18:10)::thumbs_up: cfallin submitted PR review:
LGTM -- thanks!
I was idly wondering whether we could be "even more lazy" wrt postdom: for example a backward walk from the maybe-overwriting store to see if every path meets the maybe-overwritten store before the entrypoint (avoiding revisits with a visit-set, etc). More or less like the Braun SSA construction algorithm, marching backward to find defs. That might be marginally faster when we just have a few queries (i.e. when we only really need a sparse subset of postdom info) but
- that's a whole new algorithm, and
- more importantly, I suspect your change here cuts out the majority of the cost anyway, since the computation is only at the end of a series of earlier gates
so no need to do that now. Just thought I'd mention in case it spawns any ideas later.
Wasmtime GitHub notifications bot (Jul 23 2026 at 18:11):fitzgen added PR #13947 Cranelift: A few perf improvements for the dead-store elimination pass to the merge queue.
Wasmtime GitHub notifications bot (Jul 23 2026 at 18:35)::check: fitzgen merged PR #13947.
Wasmtime GitHub notifications bot (Jul 23 2026 at 18:35):fitzgen removed PR #13947 Cranelift: A few perf improvements for the dead-store elimination pass from the merge queue.
Last updated: Jul 29 2026 at 05:03 UTC