paradigmxyz/revmc - CodSpeed

revmc

Blog Docs Changelog

Performance History

Latest Results

perf: share single JIT compiler across all benchmarks Add load_with that takes an external compiler, and clear_ir() between benchmarks to free IR while keeping JIT'd machine code alive. This avoids rebuilding the LLVM compiler per benchmark, which is especially costly under valgrind (codspeed).

dani/unified-bench

23 hours ago

perf: separate EVM setup from execution in benchmarks Use iter_batched so DB + EVM construction happens in the setup closure and only transact/handler.run is measured. Introduce EvmWithDb to safely co-locate the CacheDB and the EVM that borrows it.

dani/unified-bench

23 hours ago

fix: finalize JIT module for IR dumps and relax sanity check Call jit_function after translate_inner in JIT mode so the module gets finalized and opt.ll/opt.s dumps are written. Without this the codegen FileCheck tests failed because opt.ll was never produced. Change sanity_check to compare interpreter vs JIT result parity instead of asserting both succeed. Benchmarks like fiat_token called with no calldata legitimately revert.

dani/unified-bench

24 hours ago

fix(aarch64): save all callee-saved registers in entry/exit trampoline Save x27/x28 and NEON d8-d15 in addition to x19-x26. These are all callee-saved per AAPCS64 and must be preserved across the revmc_exit longjump back to the caller of revmc_entry.

dani/exit-trampoline

1 day ago

fix: use saturating i256-to-i64 convert at jump site

dani/truncate-i64-dynamic-jump-table

1 day ago

perf: truncate to i64 before jumping to dynamic jump table Truncate the i256 jump target to i64 at each dynamic jump site instead of doing a saturating convert in the dynamic jump table block. This removes the zext+icmp+select overhead from the jump table phi node, simplifying codegen.

dani/truncate-i64-dynamic-jump-table

1 day ago

Merge remote-tracking branch 'origin/main' into dani/exit-trampoline

dani/exit-trampoline

1 day ago

refactor: use EvmCompiler::new_llvm where backend is concrete (#315) Add `EvmCompiler::new_llvm(aot)` convenience constructor that combines `EvmLlvmBackend::new` and `EvmCompiler::new`. Replace all call sites that were using the two-step pattern with the concrete LLVM backend.

main

1 day ago

Latest Branches

-100%

feat: add burntpix benchmark and unify execution paths#316

23 hours ago

91ff075

dani/unified-bench

+17%

perf: exit trampoline for builtin errors#314

1 day ago

62dc978

dani/exit-trampoline

-1%

perf: truncate to i64 before jumping to dynamic jump table#317

1 day ago

275c49d

dani/truncate-i64-dynamic-jump-table

© 2026 CodSpeed Technology

Home Terms Privacy Docs