* Fix stale benchmarks
* Data (de)serialization and execution benchmarks:
Two kind of benchmark: relatively short, but with huge call results, and
long trace of small call results. Moreover, there are two case for each:
with same data to be merged with comparison, and data from different
par branches merged w/o comparison.
`air-trace run` subcommand allows to run AquaVM on any data, it allows to define most AquaVM inputs, providing defaults for most of them, and sets up either human-readable or JSON tracing output, the latter can be later processed by `air-trace stats`.
Anomaly data input is also supported, that is useful for slow data investigation.
Native execution mode can be used for native profiling. Please note, however, that current version cannot be built natively on Apple Sillicon processor yet, as invariably depends on the `avm-server` because of leaking types that should be refactored or hidden. The `--repeat` option can repeat the execution several times for the execution to dominate on input data reading and output.
High-level or rare calls have "info" trace level, instructions are "debug", and sub-instruction calls are "tracing". Over-detailed tracing can induce overhead that spoils timing data.
The `log_instruction` macro unconditionally pre-formatted some string to
be logged, even if logging is disabled, making whole AquaVM very slow.
After getting it fixed, native execution time dropped from 5sec to
thousands of milliseconds.
Many internal structures refer to trace positions; this is an important type of values. In the code, it is sometimes `u32`, sometimes `usize`. While such variables usually have "_pos" suffix, having a dedicated type does improve code with type guarantees.
Intially, a fold instruction was designed based on idea that its iterable can't expand,
so AIR was considered as a subtree. But after introducing recursive streams in #225, it's
no more a subtree, but a subgraph.
Use pooled `avm::server::AVMRunner` instances of
air_iterpreter_server.wasm to reduce tests' running time. It avoids
repeated WASM loading and compilation.
On my hardware, `cargo test --release` execution time (precompiled)
decreases from almost 6 minutes to 1.5 minutes.
