Build the pinned compiler, then run acceptance, the state tests and the other checks against it.
Native hosts are AArch64 macOS and x86_64 Linux. Install Git, Python 3.10 or newer, Node/npm, Rust via rustup, CMake and the host C toolchain (the Xcode command-line tools on macOS; a C compiler and binutils on Linux). Install the Rust version recorded in the manifest. No third-party Python packages are needed.
toolchain.json pins Fe master
6ccb3ca8e
(argotorg/fe#1992), its
Cargo.lock's Sonatina revision and hash, and Rust 1.99.0.
python3 native/bootstrap.py --out native/out/toolchain --toolchain 1.99.0Bootstrap checks out the exact commits, verifies the lock hash, generates the
tree-sitter parser from npm's lockfile, builds the compiler and records its
identity, hash, host and build time in <out>/build.json. The compiler is
copied to <out>/bin/fe. Use a fresh output directory each time;
--target-dir reuses Cargo artifacts, and --prepare-only materializes the
sources without building. Cargo dependencies need network access or a
populated cache.
native/tools holds fevm's checks written in Fe, as subcommands of one program
that runs from the repository's root:
native/out/toolchain/bin/fe build . --ingot fevm_tools --backend native -O 2 \
--emit executable --out-dir native/out/toolsnative/out/tools/fevm_tools accept --build native/out/toolchain/build.json \
--out native/out/acceptance --check-onlyAcceptance checks the compiler hash and runs nine stages:
workspace-O0,workspace-O1,workspace-O2: every workspace test.cli-smoke: CLI arguments, parsing, execution, return data and exit statuses, plus allocation-failure fixtures that relink the interpreter object against a controlled allocator (fixtures/vm_allocator.c).osaka-frames: 1,845 frames against revm (differential).osaka-state: 1,149 transaction and checkpoint sequences against revm's journal.tools:fevm_toolsbuilds.session: the Rust client's tests againstfevm serve.arithmetic: the compiler arithmetic matrix (below).
It writes JSON and per-command logs, and exits nonzero if any stage fails.
--jobs runs test suites concurrently (the default is one, since optimized
interpreter tests use substantial memory); --build-timeout and
--stage-timeout override the deadlines. Use --check-only on shared
machines, which skips timing measurements. Run it on each host: an unexecuted
platform is not a passing result.
CI runs acceptance, the state tests and the
blockchain tests on both hosts for every pull request and push to main.
The conformance runner runs execution-spec-tests
Osaka and Amsterdam state tests through Evm::transact, and blockchain tests
of Osaka, Amsterdam, their BPO forks and the transitions between them
through Chain::import. It is not part of acceptance.
native/soak runs repeated Osaka transactions against one world in one
process. Each expands a caller's and a callee's memory, emits a log, sets and
clears a storage slot and returns data. Peak RSS must stay flat: the runner
fails if it grows by more than 1 MiB between 1,000 and 40,000 transactions.
native/out/tools/fevm_tools soak --fe native/out/toolchain/bin/fe --out native/out/soakpython3 native/arith/run.py --fe native/out/toolchain/bin/fe \
--out native/out/arithmeticThe runner checks the compiler's 256-bit arithmetic: one standalone Fe
executable per operation, representation and optimization level, against
Python's arbitrary-precision results. Operations are addition, wrapping
multiplication, unsigned division and remainder, shifts, comparison and
wrapping exponentiation, with boundary cases across every limb and
deterministic random cases. The representations are built-in u256 and a
four-u64 struct with limb arithmetic, a control for the compiler rather than
an optimal implementation.
Without --check-only it also times each kernel with
std::native::cpu_clock_ticks(), keeping every sample, build times and
artifact sizes. --operations and --levels select a subset, and
arith/compare.py remeasures a baseline and a candidate in alternating order:
python3 native/arith/compare.py --baseline native/out/baseline/results.json \
--candidate native/out/candidate/results.json --out native/out/comparisonMeasure on an idle machine. A failure keeps the generated source and compiler reports; fix the root cause rather than dropping a representation or level.