The feed_event() function is called for every wasm binary instruction.
Let's optimize it by storing FunctionEnvironment object in
CraneliftFunctionCodeGenerator, rather than constructing it for every
feed_event() invocation.
This change reduces the time to run "ngix compile" benchmark by 68%:
Before:
nginx compile time: [1.4152 s 1.4186 s 1.4222 s]
Found 1 outliers among 10 measurements (10.00%)
1 (10.00%) high mild
After:
nginx compile time: [447.76 ms 448.32 ms 448.80 ms]
change: [-68.542% -68.440% -68.352%] (p = 0.00 < 0.05)
Performance has improved.
I assume some of the clone() calls are very expensive (Vec::clone(),
likely). I did see libc malloc()/free() high up in "perf top" profiles,
which are eliminted by this change.
I also looked into eliminating FunctionBuilder construction from
feed_event(). That turns out to be painful on lifetime rules because it
borrows bunch of other objects, so I am leaving it for someone who knows
the code better than I do.
